From: Hyung-Kyu Choi <hk0110.choi@samsung.com>
Date: Thu, 29 Mar 2018 07:33:12 +0000 (+0900)
Subject: Introduce CPU operation of NN runtime
X-Git-Tag: 0.1~515
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=804f8a002d19b65abab23ed623a68cb5bb82fbb5;p=platform%2Fcore%2Fml%2Fnnfw.git

Introduce CPU operation of NN runtime

- Introduce CPU operation of NN runtime
- Also introduce eigen and gemmlowp in depend/external

Signed-off-by: Hyung-Kyu Choi <hk0110.choi@samsung.com>
---

diff --git a/src/runtime/ref/nn/common/CMakeLists.txt b/src/runtime/ref/nn/common/CMakeLists.txt
index 4223f7f..5a97f04 100644
--- a/src/runtime/ref/nn/common/CMakeLists.txt
+++ b/src/runtime/ref/nn/common/CMakeLists.txt
@@ -1,5 +1,6 @@
 SET (CUR_INCS
      ${CMAKE_CURRENT_SOURCE_DIR}/include
+     ${CMAKE_CURRENT_SOURCE_DIR}/operations
      ${CMAKE_CURRENT_SOURCE_DIR}
 )
 SET (INC_DIRS
@@ -12,6 +13,7 @@ SET (CUR_SRCS
      ${CMAKE_CURRENT_SOURCE_DIR}/CpuExecutor.cpp
      ${CMAKE_CURRENT_SOURCE_DIR}/OperationsUtils.cpp
      ${CMAKE_CURRENT_SOURCE_DIR}/Utils.cpp
+     ${CMAKE_CURRENT_SOURCE_DIR}/operations/Activation.cpp
 )
 SET (SRCS
      ${SRCS}
diff --git a/src/runtime/ref/nn/common/include/ActivationFunctor.h b/src/runtime/ref/nn/common/include/ActivationFunctor.h
new file mode 100644
index 0000000..7e6600e
--- /dev/null
+++ b/src/runtime/ref/nn/common/include/ActivationFunctor.h
@@ -0,0 +1,69 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_ML_NN_ACTIVATION_FUNCTOR_H
+#define ANDROID_ML_NN_ACTIVATION_FUNCTOR_H
+
+#if 0 // REF-ANN
+#include "android/log.h"
+#endif
+
+#include <algorithm>
+#include <cmath>
+#include <vector>
+#include <cstdint>
+
+enum ActivationFn {
+    kActivationNone = 0,
+    kActivationRelu,
+    kActivationRelu1,
+    kActivationRelu6,
+    kActivationTanh,
+    kActivationSignBit,
+    kActivationSigmoid,
+};
+
+class ActivationFunctor {
+ public:
+  explicit ActivationFunctor(ActivationFn act) : act_(act) {}
+
+  float operator()(float a) const {
+    switch (act_) {
+      case kActivationNone:
+        return a;
+      case kActivationRelu:
+        return a < 0.f ? 0.f : a;
+      case kActivationRelu6:
+        return std::max(0.f, std::min(a, 6.f));
+      case kActivationTanh:
+        return std::tanh(a);
+      case kActivationSigmoid:
+        return 1.0f / (1.0f + std::exp(-a));
+      default:
+#if 0 // REF-ANN
+        __android_log_print(ANDROID_LOG_ERROR, "NN API",
+                            "Invalid enum value for activation function: 0x%0X",
+                            act_);
+#endif
+        exit(1);
+    }
+  }
+
+ private:
+  ActivationFn act_;
+};
+
+#endif  // ANDROID_ML_NN_ACTIVATION_FUNCTOR_H
diff --git a/src/runtime/ref/nn/common/operations/Activation.cpp b/src/runtime/ref/nn/common/operations/Activation.cpp
new file mode 100644
index 0000000..b40b34b
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/Activation.cpp
@@ -0,0 +1,211 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "Operations.h"
+#include "OperationsUtils.h"
+
+#include "internal/optimized/optimized_ops.h"
+
+// TODO-NNRT: There was no inlcude "ActivationFunctor.h" in Android NN.
+//            This may be included from some other header files.
+#include "ActivationFunctor.h"
+
+namespace android {
+namespace nn {
+
+bool reluFloat32(const float* inputData, const Shape& inputShape,
+                 float* outputData, const Shape& outputShape) {
+    int numElements = getNumberOfElements(inputShape);
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {
+        *outputData = std::max(0.f, *inputData);
+    }
+    return true;
+}
+
+bool relu1Float32(const float* inputData, const Shape& inputShape,
+                  float* outputData, const Shape& outputShape) {
+    int numElements = getNumberOfElements(inputShape);
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {
+        *outputData = std::min(std::max(-1.f, *inputData), 1.f);
+    }
+    return true;
+}
+
+bool relu6Float32(const float* inputData, const Shape& inputShape,
+                  float* outputData, const Shape& outputShape) {
+    int numElements = getNumberOfElements(inputShape);
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {
+        *outputData = std::min(std::max(0.f, *inputData), 6.f);
+    }
+    return true;
+}
+
+bool tanhFloat32(const float* inputData, const Shape& inputShape,
+                 float* outputData, const Shape& outputShape) {
+    int numElements = getNumberOfElements(inputShape);
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {
+        *outputData = std::tanh(*inputData);
+    }
+    return true;
+}
+
+bool logisticFloat32(const float* inputData, const Shape& inputShape,
+                     float* outputData, const Shape& outputShape) {
+    int numElements = getNumberOfElements(inputShape);
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {
+        *outputData = 1.f / (1.f + std::exp(-*inputData));
+    }
+    return true;
+}
+
+bool softmaxFloat32(const float* inputData, const Shape& inputShape,
+                    const float beta,
+                    float* outputData, const Shape& outputShape) {
+    Dims<4> dim;
+    if (getNumberOfDimensions(inputShape) == 2) {
+        uint32_t batch_size = getSizeOfDimension(inputShape, 0);
+        uint32_t input_size = getNumberOfElements(inputShape) / batch_size;
+
+        Shape shapeIn4D;
+        shapeIn4D.dimensions = {batch_size, 1, 1, input_size};
+        dim = convertShapeToDims(shapeIn4D);
+    } else if (getNumberOfDimensions(inputShape) == 4) {
+        dim = convertShapeToDims(inputShape);
+    } else {
+        LOG(ERROR) << "only 2D and 4D tensors supported";
+        return false;
+    }
+
+    optimized_ops::Softmax(inputData, dim, beta,
+                           outputData, dim);
+    return true;
+}
+
+#define ANDROID_NN_RELUX_QUANT8(activation)                             \
+    int numElements = getNumberOfElements(inputShape);                  \
+    int32_t output_activation_min = 0;                                  \
+    int32_t output_activation_max = 0;                                  \
+                                                                        \
+    CalculateActivationRangeUint8(activation, inputShape,               \
+                                  &output_activation_min,               \
+                                  &output_activation_max);              \
+                                                                        \
+    for (int i=0; i<numElements; i++, inputData++, outputData++) {      \
+        *outputData = std::min((uint8_t)output_activation_max,          \
+                std::max((uint8_t)output_activation_min, *inputData));  \
+    }
+
+
+bool reluQuant8(const uint8_t* inputData, const Shape& inputShape,
+                uint8_t* outputData, const Shape& outputShape) {
+    ANDROID_NN_RELUX_QUANT8(kActivationRelu)
+    return true;
+}
+
+bool relu1Quant8(const uint8_t* inputData, const Shape& inputShape,
+                 uint8_t* outputData, const Shape& outputShape) {
+    ANDROID_NN_RELUX_QUANT8(kActivationRelu1)
+    return true;
+}
+
+bool relu6Quant8(const uint8_t* inputData, const Shape& inputShape,
+                 uint8_t* outputData, const Shape& outputShape) {
+    ANDROID_NN_RELUX_QUANT8(kActivationRelu6)
+    return true;
+}
+
+#undef ANDROID_NN_RELUX_QUANT8
+
+bool logisticQuant8(const uint8_t* inputData, const Shape& inputShape,
+                    uint8_t* outputData, const Shape& outputShape) {
+    if (outputShape.offset != 0 || outputShape.scale != 1.f / 256) {
+        LOG(ERROR) << "incorrect scale / offset for output";
+        return false;
+    }
+
+    int numElements = getNumberOfElements(inputShape);
+    static constexpr int kInputIntegerBits = 4;
+
+    const double input_real_multiplier =
+            inputShape.scale *
+            static_cast<double>(1 << (31 - kInputIntegerBits));
+
+    int32_t input_multiplier = 0;
+    int32_t input_left_shift = 0;
+    if (!QuantizeMultiplierGreaterThanOne(input_real_multiplier,
+                                          &input_multiplier,
+                                          &input_left_shift)) {
+        return false;
+    }
+    int32_t input_range_radius =
+            CalculateInputRadius(kInputIntegerBits, input_left_shift);
+
+    optimized_ops::Logistic(
+            inputData, convertShapeToDims(inputShape),
+            inputShape.offset, input_range_radius,
+            input_multiplier, input_left_shift,
+            outputData, convertShapeToDims(outputShape));
+
+    return true;
+}
+
+bool softmaxQuant8(const uint8_t* inputData, const Shape& inputShape,
+                   const float beta,
+                   uint8_t* outputData, const Shape& outputShape) {
+    Dims<4> dim;
+    if (getNumberOfDimensions(inputShape) == 2) {
+        uint32_t batch_size = getSizeOfDimension(inputShape, 0);
+        uint32_t input_size = getNumberOfElements(inputShape) / batch_size;
+
+        Shape shapeIn4D;
+        shapeIn4D.dimensions = {batch_size, 1, 1, input_size};
+        dim = convertShapeToDims(shapeIn4D);
+    } else if (getNumberOfDimensions(inputShape) == 4) {
+        dim = convertShapeToDims(inputShape);
+    } else {
+        LOG(ERROR) << "only 2D and 4D tensors supported";
+        return false;
+    }
+
+    if (outputShape.offset != 0 || outputShape.scale != 1.f / 256) {
+        LOG(ERROR) << "incorrect scale / offset for output";
+        return false;
+    }
+
+    static const int32_t kScaledDiffIntegerBits = 5;
+    const double input_beta_real_multiplier = std::min(
+            1.0 * beta * inputShape.scale * (1 << (31 - kScaledDiffIntegerBits)),
+            (1ll << 31) - 1.0);
+
+    int32_t input_multiplier = 0;
+    int32_t input_left_shift = 0;
+    if (!QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier,
+                                          &input_multiplier,
+                                          &input_left_shift)) {
+        return false;
+    }
+    float diff_min = -1.0f * CalculateInputRadius(kScaledDiffIntegerBits,
+                                                  input_left_shift);
+
+    optimized_ops::Softmax(inputData, dim, input_multiplier,
+                           input_left_shift, diff_min,
+                           outputData, dim);
+    return true;
+}
+
+
+}  // namespace nn
+}  // namespace android
diff --git a/src/runtime/ref/nn/common/operations/internal/common.h b/src/runtime/ref/nn/common/operations/internal/common.h
new file mode 100644
index 0000000..c54ee6e
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/common.h
@@ -0,0 +1,79 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMMON_H_
+#define ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMMON_H_
+
+#ifndef USE_NEON
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define USE_NEON
+#include <arm_neon.h>
+#endif
+#endif
+
+#include "gemmlowp.h"
+#include "types.h"
+
+namespace android {
+namespace nn {
+
+template <FusedActivationFunctionType Ac>
+struct ActivationFunctionImpl {};
+
+template <>
+struct ActivationFunctionImpl<FusedActivationFunctionType::kNone> {
+  static float Eval(float x) { return x; }
+};
+
+template <>
+struct ActivationFunctionImpl<FusedActivationFunctionType::kRelu> {
+  static float Eval(float x) { return x < 0.f ? 0.f : x; }
+};
+
+template <>
+struct ActivationFunctionImpl<FusedActivationFunctionType::kRelu1> {
+  static float Eval(float x) { return x > 1.f ? 1.f : x < -1.f ? -1.f : x; }
+};
+
+template <>
+struct ActivationFunctionImpl<FusedActivationFunctionType::kRelu6> {
+  static float Eval(float x) { return x > 6.f ? 6.f : x < 0.f ? 0.f : x; }
+};
+
+template <FusedActivationFunctionType Ac>
+float ActivationFunction(float x) {
+  return ActivationFunctionImpl<Ac>::Eval(x);
+}
+
+inline int32 MultiplyByQuantizedMultiplierSmallerThanOne(
+    int32 x, int32 quantized_multiplier, int right_shift) {
+  using gemmlowp::RoundingDivideByPOT;
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  return RoundingDivideByPOT(
+      SaturatingRoundingDoublingHighMul(x, quantized_multiplier), right_shift);
+}
+
+inline int32 MultiplyByQuantizedMultiplierGreaterThanOne(
+    int32 x, int32 quantized_multiplier, int left_shift) {
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  return SaturatingRoundingDoublingHighMul(x * (1 << left_shift),
+                                           quantized_multiplier);
+}
+
+}  // namespace nn
+}  // namespace android
+
+#endif  // ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMMON_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/compatibility.h b/src/runtime/ref/nn/common/operations/internal/compatibility.h
new file mode 100644
index 0000000..9d9cdb7
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/compatibility.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMPATIBILITY_H_
+#define ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMPATIBILITY_H_
+
+#ifndef ANDROID_ML_NN_COMPATIBILITY
+#define ANDROID_ML_NN_COMPATIBILITY
+
+#include <cassert>
+#include <cstdint>
+
+#ifndef DCHECK
+#define DCHECK(condition) (condition) ? (void)0 : assert(false)
+#endif
+
+#ifndef DCHECK_EQ
+#define DCHECK_EQ(x, y) ((x) == (y)) ? (void)0 : assert(false)
+#endif
+
+#ifndef DCHECK_GE
+#define DCHECK_GE(x, y) ((x) >= (y)) ? (void)0 : assert(false)
+#endif
+
+#ifndef DCHECK_GT
+#define DCHECK_GT(x, y) ((x) > (y)) ? (void)0 : assert(false)
+#endif
+
+#ifndef DCHECK_LE
+#define DCHECK_LE(x, y) ((x) <= (y)) ? (void)0 : assert(false)
+#endif
+
+#ifndef DCHECK_LT
+#define DCHECK_LT(x, y) ((x) < (y)) ? (void)0 : assert(false)
+#endif
+
+#ifndef CHECK_EQ
+#define CHECK_EQ(x, y) ((x) == (y)) ? (void)0 : assert(false)
+#endif
+
+using uint8 = std::uint8_t;
+using int16 = std::int16_t;
+using uint16 = std::uint16_t;
+using int32 = std::int32_t;
+using uint32 = std::uint32_t;
+
+#endif
+
+#endif  // ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_COMPATIBILITY_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/optimized/cpu_check.h b/src/runtime/ref/nn/common/operations/internal/optimized/cpu_check.h
new file mode 100644
index 0000000..02f42fd
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/optimized/cpu_check.h
@@ -0,0 +1,28 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_CPU_CHECK_
+#define FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_CPU_CHECK_
+
+// NEON_OR_PORTABLE(SomeFunc, arcs) calls NeonSomeFunc(args) if NEON is
+// enabled at build time, or PortableSomeFunc(args) otherwise.
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define NEON_OR_PORTABLE(funcname, ...) Neon##funcname(__VA_ARGS__)
+#else
+#define NEON_OR_PORTABLE(funcname, ...) Portable##funcname(__VA_ARGS__)
+#endif
+
+#endif  // FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_CPU_CHECK_
diff --git a/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.cc b/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.cc
new file mode 100644
index 0000000..f17730f
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.cc
@@ -0,0 +1,216 @@
+/*
+ * Copyright 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <string.h>
+
+#include "ActivationFunctor.h"
+#include "tensor_utils_impl.h"
+
+#ifdef USE_NEON
+
+#include <arm_neon.h>
+#define kFloatWeightsPerNeonLane 4
+
+namespace android {
+namespace nn {
+namespace tensor_utils {
+
+void NeonMatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows,
+                                             int m_cols, const float* vector,
+                                             int n_batch, float* result,
+                                             int result_stride) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      m_cols - (m_cols & (kFloatWeightsPerNeonLane - 1));
+
+  // The arrays used to cache the vector.
+  float32x4_t* vector_cache_float32x4 =
+      new float32x4_t[(m_cols / kFloatWeightsPerNeonLane) *
+                      sizeof(float32x4_t)];
+
+  for (int b = 0; b < n_batch; b++) {
+    float* result_in_batch = result + b * m_rows;
+    const float* vector_in_batch = vector + b * m_cols;
+    const float* matrix_ptr = matrix;
+    for (int c = 0; c < postamble_start; c += kFloatWeightsPerNeonLane) {
+      vector_cache_float32x4[c >> 2] = vld1q_f32(vector_in_batch + c);
+    }
+    for (int r = 0; r < m_rows; r++) {
+      float32x4_t acc_32x4 = vmovq_n_f32(0.0);
+      for (int c = 0; c < postamble_start; c += kFloatWeightsPerNeonLane) {
+        float32x4_t temp = vector_cache_float32x4[c >> 2];
+        // Load 4 float values from vector1 and vector2 and accumulator.
+        float32x4_t v1_f32x4 = vld1q_f32(matrix_ptr + c);
+        // Vector multiply-accumulate 4 float
+        acc_32x4 = vmlaq_f32(acc_32x4, v1_f32x4, temp);
+      }
+      // Add the 4 intermediate sum values to get the final dot-prod value for
+      // this column.
+      *result_in_batch +=
+          (vgetq_lane_f32(acc_32x4, 0) + vgetq_lane_f32(acc_32x4, 1) +
+           vgetq_lane_f32(acc_32x4, 2) + vgetq_lane_f32(acc_32x4, 3));
+      for (int c = postamble_start; c < m_cols; c++) {
+        *result_in_batch += matrix_ptr[c] * vector_in_batch[c];
+      }
+      matrix_ptr += m_cols;
+      result_in_batch += result_stride;
+    }
+  }
+  delete[] vector_cache_float32x4;
+}
+
+void NeonVectorVectorCwiseProduct(const float* vector1, const float* vector2,
+                                  int v_size, float* result) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      v_size - (v_size & (kFloatWeightsPerNeonLane - 1));
+  for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+    // Load 4 float values from vector1 and vector2.
+    float32x4_t v1_f32x4 = vld1q_f32(vector1 + v);
+    float32x4_t v2_f32x4 = vld1q_f32(vector2 + v);
+    // Vector multiply 4 float
+    float32x4_t mul_32x4 = vmulq_f32(v1_f32x4, v2_f32x4);
+    // Save to result array.
+    vst1q_f32(&result[v], mul_32x4);
+  }
+  for (int v = postamble_start; v < v_size; v++) {
+    result[v] = vector1[v] * vector2[v];
+  }
+}
+
+void NeonVectorVectorCwiseProductAccumulate(const float* vector1,
+                                            const float* vector2, int v_size,
+                                            float* result) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      v_size - (v_size & (kFloatWeightsPerNeonLane - 1));
+  for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+    // Load 4 float values from vector1 and vector2 and accumulator.
+    float32x4_t v1_f32x4 = vld1q_f32(vector1 + v);
+    float32x4_t v2_f32x4 = vld1q_f32(vector2 + v);
+    float32x4_t acc_32x4 = vld1q_f32(result + v);
+    // Vector multiply-accumulate 4 float
+    acc_32x4 = vmlaq_f32(acc_32x4, v1_f32x4, v2_f32x4);
+    // Save to result array.
+    vst1q_f32(&result[v], acc_32x4);
+  }
+  for (int v = postamble_start; v < v_size; v++) {
+    result[v] += vector1[v] * vector2[v];
+  }
+}
+
+void NeonVectorBatchVectorCwiseProductAccumulate(const float* vector,
+                                                 int v_size,
+                                                 const float* batch_vector,
+                                                 int n_batch, float* result) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      v_size - (v_size & (kFloatWeightsPerNeonLane - 1));
+
+  // The arrays used to cache the vector.
+  float32x4_t* vector_cache_float32x4 =
+      new float32x4_t[(v_size / kFloatWeightsPerNeonLane) *
+                      sizeof(float32x4_t)];
+  for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+    vector_cache_float32x4[v >> 2] = vld1q_f32(vector + v);
+  }
+
+  float* result_ptr = result;
+  const float* batch_vector_ptr = batch_vector;
+  for (int b = 0; b < n_batch; b++) {
+    for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+      // Load from memory to vectors.
+      float32x4_t result_f32x4 = vld1q_f32(result_ptr + v);
+      float32x4_t batch_vector_f32x4 = vld1q_f32(batch_vector_ptr + v);
+      // Multiply-accumulate.
+      result_f32x4 = vmlaq_f32(result_f32x4, batch_vector_f32x4,
+                               vector_cache_float32x4[v >> 2]);
+      // Store.
+      vst1q_f32(result_ptr + v, result_f32x4);
+    }
+    // Postamble loop
+    for (int v = postamble_start; v < v_size; v++) {
+      result_ptr[v] += vector[v] * batch_vector_ptr[v];
+    }
+    // Update the pointers.
+    result_ptr += v_size;
+    batch_vector_ptr += v_size;
+  }
+  delete[] vector_cache_float32x4;
+}
+
+void NeonSub1Vector(const float* vector, int v_size, float* result) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      v_size - (v_size & (kFloatWeightsPerNeonLane - 1));
+
+  float32x4_t one_f32x4 = vmovq_n_f32(1.0);
+  for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+    // Load 4 float values from the current pointers of the input column and
+    // subtract from 1.
+    float32x4_t v_f32x4 = vld1q_f32(vector + v);
+    float32x4_t result_f32x4 = vsubq_f32(one_f32x4, v_f32x4);
+    // Save to output.
+    vst1q_f32(result + v, result_f32x4);
+  }
+  for (int v = postamble_start; v < v_size; v++) {
+    result[v] = 1.0f - vector[v];
+  }
+}
+
+void NeonClipVector(const float* vector, int v_size, float abs_limit,
+                    float* result) {
+  // If v_size is not divisible by kWeightsPerNeonLane, we cannot use the main
+  // vectorized loop, and we need to process sequentially. postamble_start shows
+  // the start index where this should happen.
+  const int postamble_start =
+      v_size - (v_size & (kFloatWeightsPerNeonLane - 1));
+
+  // Replicate abs_limit and -abs_limit in two vectors.
+  const float32x4_t abs_limit_f32x4 = vmovq_n_f32(abs_limit);
+  const float32x4_t neg_abs_limit_f32x4 = vmovq_n_f32(-abs_limit);
+
+  for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) {
+    // Load from memory to vector.
+    float32x4_t v_f32x4 = vld1q_f32(vector + v);
+    // Clip between abs_limit and -abs_limit.
+    float32x4_t result_f32x4 = vminq_f32(abs_limit_f32x4, v_f32x4);
+    result_f32x4 = vmaxq_f32(neg_abs_limit_f32x4, result_f32x4);
+    // Save to output.
+    vst1q_f32(result + v, result_f32x4);
+  }
+  // Postamble loop.
+  for (int v = postamble_start; v < v_size; v++) {
+    result[v] = (abs_limit < vector[v]) ? abs_limit : vector[v];
+    result[v] = (-abs_limit > result[v]) ? -abs_limit : result[v];
+  }
+}
+
+}  // namespace tensor_utils
+}  // namespace nn
+}  // namespace android
+
+#endif  // USE_NEON
diff --git a/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.h b/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.h
new file mode 100644
index 0000000..92a151c
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/optimized/neon_tensor_utils.h
@@ -0,0 +1,118 @@
+/*
+ * Copyright 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_NEON_TENSOR_UTILS_H_
+#define FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_NEON_TENSOR_UTILS_H_
+
+#include "ActivationFunctor.h"
+#include "cpu_check.h"
+#include "tensor_utils_impl.h"
+
+namespace android {
+namespace nn {
+namespace tensor_utils {
+
+void MatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows,
+                                         int m_cols, const float* vector,
+                                         int n_batch, float* result,
+                                         int result_stride) {
+  NEON_OR_PORTABLE(MatrixBatchVectorMultiplyAccumulate, matrix, m_rows, m_cols,
+                   vector, n_batch, result, result_stride);
+}
+
+void VectorVectorCwiseProduct(const float* vector1, const float* vector2,
+                              int v_size, float* result) {
+  NEON_OR_PORTABLE(VectorVectorCwiseProduct, vector1, vector2, v_size, result);
+}
+
+void VectorVectorCwiseProductAccumulate(const float* vector1,
+                                        const float* vector2, int v_size,
+                                        float* result) {
+  NEON_OR_PORTABLE(VectorVectorCwiseProductAccumulate, vector1, vector2, v_size,
+                   result);
+}
+
+void VectorBatchVectorCwiseProductAccumulate(const float* vector, int v_size,
+                                             const float* batch_vector,
+                                             int n_batch, float* result) {
+  NEON_OR_PORTABLE(VectorBatchVectorCwiseProductAccumulate, vector, v_size,
+                   batch_vector, n_batch, result);
+}
+
+float VectorVectorDotProduct(const float* vector1, const float* vector2,
+                             int v_size) {
+  return PortableVectorVectorDotProduct(vector1, vector2, v_size);
+}
+
+void BatchVectorBatchVectorDotProduct(const float* vector1,
+                                      const float* vector2, int v_size,
+                                      int n_batch, float* result,
+                                      int result_stride) {
+  PortableBatchVectorBatchVectorDotProduct(vector1, vector2, v_size, n_batch,
+                                           result, result_stride);
+}
+
+void VectorBatchVectorAssign(const float* vector, int v_size, int n_batch,
+                             float* batch_vector) {
+  PortableVectorBatchVectorAssign(vector, v_size, n_batch, batch_vector);
+}
+
+void ApplySigmoidToVector(const float* vector, int v_size, float* result) {
+  PortableApplySigmoidToVector(vector, v_size, result);
+}
+
+void ApplyActivationToVector(const float* vector, int v_size,
+                             ActivationFn activation, float* result) {
+  PortableApplyActivationToVector(vector, v_size, activation, result);
+}
+
+void CopyVector(const float* vector, int v_size, float* result) {
+  PortableCopyVector(vector, v_size, result);
+}
+
+void Sub1Vector(const float* vector, int v_size, float* result) {
+  NEON_OR_PORTABLE(Sub1Vector, vector, v_size, result);
+}
+
+void ZeroVector(float* vector, int v_size) {
+  PortableZeroVector(vector, v_size);
+}
+
+float Clip(float f, float abs_limit) { return PortableClip(f, abs_limit); }
+
+void ClipVector(const float* vector, int v_size, float abs_limit,
+                float* result) {
+  NEON_OR_PORTABLE(ClipVector, vector, v_size, abs_limit, result);
+}
+
+// TODO(ghodrat): Implement Neon version.
+void VectorShiftLeft(float* vector, int v_size, float shift_value) {
+  PortableVectorShiftLeft(vector, v_size, shift_value);
+}
+
+// TODO(ghodrat): Implement Neon version.
+void ReductionSumVector(const float* input_vector, int input_stride,
+                        float* output_vector, int output_size,
+                        int reduction_size) {
+  PortableReductionSumVector(input_vector, input_stride, output_vector,
+                             output_size, reduction_size);
+}
+
+}  // namespace tensor_utils
+}  // namespace nn
+}  // namespace android
+
+#endif  // FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_NEON_TENSOR_UTILS_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/optimized/optimized_ops.h b/src/runtime/ref/nn/common/operations/internal/optimized/optimized_ops.h
new file mode 100644
index 0000000..5bbec3d
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/optimized/optimized_ops.h
@@ -0,0 +1,2712 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_OPS_H_
+#define ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_OPS_H_
+
+#include <assert.h>
+#include <stdint.h>
+#include <sys/types.h>
+#include <algorithm>
+#include <cmath>
+#include <limits>
+#include <memory>
+#include <tuple>
+#include <type_traits>
+
+#include "Eigen/Core"
+#include "fixedpoint.h"
+#include "gemmlowp.h"
+#include "../common.h"
+#include "../types.h"
+
+namespace android {
+namespace nn {
+namespace optimized_ops {
+
+// Make a local VectorMap typedef allowing to map a float array
+// as a Eigen vector expression. The std::conditional here is to
+// construct the suitable Eigen type for the constness of the
+// data. Indeed, for const data, we need to produce
+//    Eigen::Map<const Eigen::Matrix<float, ...>>
+// and not the more straightforward
+//    Eigen::Map<Eigen::Matrix<const float, ...>>
+template <typename Scalar>
+using VectorMap = typename std::conditional<
+    std::is_const<Scalar>::value,
+    Eigen::Map<const Eigen::Matrix<typename std::remove_const<Scalar>::type,
+                                   Eigen::Dynamic, 1>>,
+    Eigen::Map<Eigen::Matrix<Scalar, Eigen::Dynamic, 1>>>::type;
+
+template <typename Scalar, int N>
+VectorMap<Scalar> MapAsVector(Scalar* data, const Dims<N>& dims) {
+  const int size = RequiredBufferSizeForDims(dims);
+  return VectorMap<Scalar>(data, size, 1);
+}
+
+// Make a local VectorMap typedef allowing to map a float array
+// as a Eigen matrix expression. The same explanation as for VectorMap
+// above also applies here.
+template <typename Scalar>
+using MatrixMap = typename std::conditional<
+    std::is_const<Scalar>::value,
+    Eigen::Map<const Eigen::Matrix<typename std::remove_const<Scalar>::type,
+                                   Eigen::Dynamic, Eigen::Dynamic>>,
+    Eigen::Map<Eigen::Matrix<Scalar, Eigen::Dynamic, Eigen::Dynamic>>>::type;
+
+template <typename Scalar, int N>
+MatrixMap<Scalar> MapAsMatrixWithFirstDimAsRows(Scalar* data,
+                                                const Dims<N>& dims) {
+  const int rows = dims.sizes[0];
+  int cols = 1;
+  for (int d = 1; d < N; d++) {
+    cols *= dims.sizes[d];
+  }
+  return MatrixMap<Scalar>(data, rows, cols);
+}
+
+template <typename Scalar, int N>
+MatrixMap<Scalar> MapAsMatrixWithLastDimAsCols(Scalar* data,
+                                               const Dims<N>& dims) {
+  const int cols = dims.sizes[N - 1];
+  int rows = 1;
+  for (int d = 0; d < N - 1; d++) {
+    rows *= dims.sizes[d];
+  }
+  return MatrixMap<Scalar>(data, rows, cols);
+}
+
+template <typename Scalar>
+using ArrayMap = typename std::conditional<
+    std::is_const<Scalar>::value,
+    Eigen::Map<const Eigen::Array<typename std::remove_const<Scalar>::type,
+                                  Eigen::Dynamic, Eigen::Dynamic>>,
+    Eigen::Map<Eigen::Array<Scalar, Eigen::Dynamic, Eigen::Dynamic>>>::type;
+
+template <typename Scalar, int N>
+ArrayMap<Scalar> MapAsArrayWithFirstDimAsRows(Scalar* data,
+                                              const Dims<N>& dims) {
+  const int rows = dims.sizes[0];
+  int cols = 1;
+  for (int d = 1; d < N; d++) {
+    cols *= dims.sizes[d];
+  }
+  return ArrayMap<Scalar>(data, rows, cols);
+}
+
+// TODO(b/62193649): this function is only needed as long
+// as we have the --variable_batch hack.
+template <typename Scalar, int N>
+MatrixMap<Scalar> MapAsMatrixWithGivenNumberOfRows(Scalar* data,
+                                                   const Dims<N>& dims,
+                                                   int rows) {
+  int cols = 1;
+  bool matched_rows = false;
+  for (int d = 0; d < N; d++) {
+    cols *= dims.sizes[d];
+    if (cols == rows) {
+      matched_rows = true;
+      cols = 1;
+    }
+  }
+  DCHECK(matched_rows);
+  return MatrixMap<Scalar>(data, rows, cols);
+}
+
+// DO NOT USE THIS STRUCT FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING ELEMENT-WISE
+// BROADCASTING.
+//
+// NdArrayDesc<N> describes the shape and memory layout of an N-dimensional
+// rectangular array of numbers.
+//
+// NdArrayDesc<N> is basically identical to Dims<N> defined in types.h.
+// However, as Dims<N> is to be deprecated, this class exists as an adaptor
+// to enable simple unoptimized implementations of element-wise broadcasting
+// operations.
+template<int N>
+struct NdArrayDesc {
+  // The "extent" of each dimension. Indices along dimension d must be in the
+  // half-open interval [0, extents[d]).
+  int extents[N];
+
+  // The number of *elements* (not bytes) between consecutive indices of each
+  // dimension.
+  int strides[N];
+};
+
+// DO NOT USE THIS FUNCTION FOR NEW FUNCTIONALITY BEYOND IMPLEMENTING
+// ELEMENT-WISE BROADCASTING.
+//
+// Same as Offset(), except takes as NdArrayDesc<N> instead of Dims<N>.
+inline int SubscriptToIndex(const NdArrayDesc<4>& desc, int i0, int i1, int i2,
+                            int i3) {
+  DCHECK(i0 >= 0 && i0 < desc.extents[0]);
+  DCHECK(i1 >= 0 && i1 < desc.extents[1]);
+  DCHECK(i2 >= 0 && i2 < desc.extents[2]);
+  DCHECK(i3 >= 0 && i3 < desc.extents[3]);
+  return i0 * desc.strides[0] + i1 * desc.strides[1] + i2 * desc.strides[2] +
+         i3 * desc.strides[3];
+}
+
+// Given the dimensions of the operands for an element-wise binary broadcast,
+// adjusts them so that they can be directly iterated over with simple loops.
+// Returns the adjusted dims as instances of NdArrayDesc in 'desc0_out' and
+// 'desc1_out'. 'desc0_out' and 'desc1_out' cannot be nullptr.
+//
+// This function assumes that the two input shapes are compatible up to
+// broadcasting and the shorter one has already been prepended with 1s to be the
+// same length. E.g., if shape0 is (1, 16, 16, 64) and shape1 is (1, 64),
+// shape1 must already have been prepended to be (1, 1, 1, 64). Recall that
+// Dims<N> refer to shapes in reverse order. In this case, input0_dims will be
+// (64, 16, 16, 1) and input1_dims will be (64, 1, 1, 1).
+//
+// When two shapes are compatible up to broadcasting, for each dimension d,
+// the input extents are either equal, or one of them is 1.
+//
+// This function performs the following for each dimension d:
+// - If the extents are equal, then do nothing since the loop that walks over
+//   both of the input arrays is correct.
+// - Otherwise, one (and only one) of the extents must be 1. Say extent0 is 1
+//   and extent1 is e1. Then set extent0 to e1 and stride0 *to 0*. This allows
+//   array0 to be referenced *at any index* in dimension d and still access the
+//   same slice.
+template <int N>
+inline void NdArrayDescsForElementwiseBroadcast(const Dims<N>& input0_dims,
+                                                const Dims<N>& input1_dims,
+                                                NdArrayDesc<N>* desc0_out,
+                                                NdArrayDesc<N>* desc1_out) {
+  DCHECK(desc0_out != nullptr);
+  DCHECK(desc1_out != nullptr);
+
+  // Copy dims to desc.
+  for (int i = 0; i < N; ++i) {
+    desc0_out->extents[i] = input0_dims.sizes[i];
+    desc0_out->strides[i] = input0_dims.strides[i];
+    desc1_out->extents[i] = input1_dims.sizes[i];
+    desc1_out->strides[i] = input1_dims.strides[i];
+  }
+
+  // Walk over each dimension. If the extents are equal do nothing.
+  // Otherwise, set the desc with extent 1 to have extent equal to the other and
+  // stride 0.
+  for (int i = 0; i < N; ++i) {
+    const int extent0 = ArraySize(input0_dims, i);
+    const int extent1 = ArraySize(input1_dims, i);
+    if (extent0 != extent1) {
+      if (extent0 == 1) {
+        desc0_out->strides[i] = 0;
+        desc0_out->extents[i] = extent1;
+      } else {
+        DCHECK_EQ(extent1, 1);
+        desc1_out->strides[i] = 0;
+        desc1_out->extents[i] = extent0;
+      }
+    }
+  }
+}
+
+#ifdef USE_NEON
+template <FusedActivationFunctionType Ac>
+void AddBiasAndEvalActivationFunction(const float* bias_data,
+                                      const Dims<4>& bias_dims,
+                                      float* array_data,
+                                      const Dims<4>& array_dims) {
+  gemmlowp::ScopedProfilingLabel label("AddBiasAndEvalActivationFunction");
+  const int bias_size = bias_dims.sizes[3] * bias_dims.strides[3];
+  const int array_size = array_dims.sizes[3] * array_dims.strides[3];
+  DCHECK_EQ((array_size % bias_size), 0);
+  float* array_ptr = array_data;
+  float* array_end_ptr = array_ptr + array_size;
+  const auto zero = vdupq_n_f32(0);
+  const auto six = vdupq_n_f32(6);
+  const auto neg_one = vdupq_n_f32(-1);
+  const auto one = vdupq_n_f32(1);
+  for (; array_ptr != array_end_ptr; array_ptr += bias_size) {
+    int i = 0;
+    for (; i <= bias_size - 16; i += 16) {
+      auto b0 = vld1q_f32(bias_data + i);
+      auto b1 = vld1q_f32(bias_data + i + 4);
+      auto b2 = vld1q_f32(bias_data + i + 8);
+      auto b3 = vld1q_f32(bias_data + i + 12);
+      auto a0 = vld1q_f32(array_ptr + i);
+      auto a1 = vld1q_f32(array_ptr + i + 4);
+      auto a2 = vld1q_f32(array_ptr + i + 8);
+      auto a3 = vld1q_f32(array_ptr + i + 12);
+      auto x0 = vaddq_f32(a0, b0);
+      auto x1 = vaddq_f32(a1, b1);
+      auto x2 = vaddq_f32(a2, b2);
+      auto x3 = vaddq_f32(a3, b3);
+      if (Ac == FusedActivationFunctionType::kRelu ||
+          Ac == FusedActivationFunctionType::kRelu6) {
+        x0 = vmaxq_f32(zero, x0);
+        x1 = vmaxq_f32(zero, x1);
+        x2 = vmaxq_f32(zero, x2);
+        x3 = vmaxq_f32(zero, x3);
+        if (Ac == FusedActivationFunctionType::kRelu6) {
+          x0 = vminq_f32(six, x0);
+          x1 = vminq_f32(six, x1);
+          x2 = vminq_f32(six, x2);
+          x3 = vminq_f32(six, x3);
+        }
+      } else if (Ac == FusedActivationFunctionType::kRelu1) {
+        x0 = vmaxq_f32(neg_one, x0);
+        x1 = vmaxq_f32(neg_one, x1);
+        x2 = vmaxq_f32(neg_one, x2);
+        x3 = vmaxq_f32(neg_one, x3);
+        x0 = vminq_f32(one, x0);
+        x1 = vminq_f32(one, x1);
+        x2 = vminq_f32(one, x2);
+        x3 = vminq_f32(one, x3);
+      }
+      vst1q_f32(array_ptr + i, x0);
+      vst1q_f32(array_ptr + i + 4, x1);
+      vst1q_f32(array_ptr + i + 8, x2);
+      vst1q_f32(array_ptr + i + 12, x3);
+    }
+    for (; i <= bias_size - 4; i += 4) {
+      auto b = vld1q_f32(bias_data + i);
+      auto a = vld1q_f32(array_ptr + i);
+      auto x = vaddq_f32(a, b);
+      if (Ac == FusedActivationFunctionType::kRelu ||
+          Ac == FusedActivationFunctionType::kRelu6) {
+        x = vmaxq_f32(zero, x);
+        if (Ac == FusedActivationFunctionType::kRelu6) {
+          x = vminq_f32(six, x);
+        }
+      } else if (Ac == FusedActivationFunctionType::kRelu1) {
+        x = vmaxq_f32(neg_one, x);
+        x = vminq_f32(one, x);
+      }
+      vst1q_f32(array_ptr + i, x);
+    }
+    for (; i < bias_size; i++) {
+      array_ptr[i] = ActivationFunction<Ac>(array_ptr[i] + bias_data[i]);
+    }
+  }
+}
+#else  // not NEON
+template <FusedActivationFunctionType Ac>
+void AddBiasAndEvalActivationFunction(const float* bias_data,
+                                      const Dims<4>& bias_dims,
+                                      float* array_data,
+                                      const Dims<4>& array_dims) {
+  gemmlowp::ScopedProfilingLabel label("AddBiasAndEvalActivationFunction");
+  const int bias_size = bias_dims.sizes[3] * bias_dims.strides[3];
+  const int array_size = array_dims.sizes[3] * array_dims.strides[3];
+  DCHECK_EQ((array_size % bias_size), 0);
+  for (int array_offset = 0; array_offset < array_size;
+       array_offset += bias_size) {
+    for (int i = 0; i < bias_size; i++) {
+      array_data[array_offset + i] =
+          ActivationFunction<Ac>(array_data[array_offset + i] + bias_data[i]);
+    }
+  }
+}
+#endif
+
+template <typename Lhs, typename Rhs, typename Result>
+void Gemm(const Eigen::MatrixBase<Lhs>& lhs, const Eigen::MatrixBase<Rhs>& rhs,
+          Eigen::MatrixBase<Result>* result) {
+  if (rhs.cols() == 1) {
+    gemmlowp::ScopedProfilingLabel label("GEMV");
+    result->col(0).noalias() = lhs * rhs.col(0);
+  } else {
+    gemmlowp::ScopedProfilingLabel label("GEMM");
+    result->noalias() = lhs * rhs;
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void FullyConnected(const float* input_data, const Dims<4>& input_dims,
+                    const float* weights_data, const Dims<4>& weights_dims,
+                    const float* bias_data, const Dims<4>& bias_dims,
+                    float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("FullyConnected");
+  // TODO(b/62193649): this convoluted shape computation (determining
+  // input_rows from the weights_dims, then MapAsMatrixWithGivenNumberOfRows)
+  // is because the current --variable_batch hack consists in overwriting the
+  // 3rd dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  // When that is fixed, this should become:
+  // const auto input_matrix_map =
+  //     MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  const int input_rows = ArraySize(weights_dims, 0);
+  const auto input_matrix_map =
+      MapAsMatrixWithGivenNumberOfRows(input_data, input_dims, input_rows);
+  const auto filter_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(weights_data, weights_dims);
+  auto output_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+
+  Gemm(filter_matrix_map.transpose(), input_matrix_map, &output_matrix_map);
+  AddBiasAndEvalActivationFunction<Ac>(bias_data, bias_dims, output_data,
+                                       output_dims);
+}
+
+inline void preload_l1_stream(const uint8* ptr) {
+#ifdef GEMMLOWP_ARM_64
+  asm volatile("prfm pldl1strm, [%[ptr]]\n" ::[ptr] "r"(ptr) :);
+#else
+  gemmlowp::Prefetch(ptr);
+#endif
+}
+
+#ifdef USE_NEON
+template <FusedActivationFunctionType Ac>
+void FullyConnectedAsGEMV(const uint8* input_data, const Dims<4>& input_dims,
+                          int32 input_offset, const uint8* filter_data,
+                          const Dims<4>& filter_dims, int32 filter_offset,
+                          const int32* bias_data, const Dims<4>& bias_dims,
+                          int32 output_offset, int32 output_multiplier,
+                          int output_shift, int32 output_activation_min,
+                          int32 output_activation_max, uint8* output_data,
+                          const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("FullyConnectedAsGEMV/8bit");
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK(IsPackedWithoutStrides(input_dims));
+  DCHECK(IsPackedWithoutStrides(filter_dims));
+  DCHECK(IsPackedWithoutStrides(bias_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+  DCHECK_EQ(ArraySize(output_dims, 1) * ArraySize(output_dims, 2) *
+                ArraySize(output_dims, 3),
+            1);
+  const int input_size = input_dims.strides[3];
+  const int output_size = MatchingArraySize(filter_dims, 1, output_dims, 0);
+  static constexpr int kPeel = 4;
+  for (int k = 0; k < input_size; k += 64) {
+    preload_l1_stream(input_data + k);
+  }
+  for (int k = 0; k < kPeel * input_size; k += 64) {
+    preload_l1_stream(filter_data + k);
+  }
+  DCHECK(!(output_size % kPeel));
+  const int32* bias_ptr = bias_data;
+  uint8* output_ptr = output_data;
+  for (int out = 0; out < output_size; out += kPeel) {
+    int32x4_t acc[kPeel];
+    for (int k = 0; k < kPeel; k++) {
+      acc[k] = vdupq_n_s32(0);
+    }
+    const int16x8_t input_offset_vec = vdupq_n_s16(input_offset);
+    const int16x8_t filter_offset_vec = vdupq_n_s16(filter_offset);
+    int in = 0;
+    for (; in <= input_size - 16; in += 16) {
+      const uint8x16_t input_val_u8 = vld1q_u8(input_data + in);
+      uint8x16_t filter_val_u8[kPeel];
+      for (int k = 0; k < kPeel; k++) {
+        const uint8* filter_ptr = filter_data + in + (out + k) * input_size;
+        filter_val_u8[k] = vld1q_u8(filter_ptr);
+        preload_l1_stream(filter_ptr + 64);
+      }
+      int16x8_t input_val[2];
+      const uint8x8_t low = vget_low_u8(input_val_u8);
+      const uint8x8_t high = vget_high_u8(input_val_u8);
+      input_val[0] = vreinterpretq_s16_u16(vmovl_u8(low));
+      input_val[1] = vreinterpretq_s16_u16(vmovl_u8(high));
+      input_val[0] = vaddq_s16(input_val[0], input_offset_vec);
+      input_val[1] = vaddq_s16(input_val[1], input_offset_vec);
+      int16x8_t filter_val[kPeel][2];
+      for (int k = 0; k < kPeel; k++) {
+        const uint8x8_t low = vget_low_u8(filter_val_u8[k]);
+        const uint8x8_t high = vget_high_u8(filter_val_u8[k]);
+        filter_val[k][0] = vreinterpretq_s16_u16(vmovl_u8(low));
+        filter_val[k][1] = vreinterpretq_s16_u16(vmovl_u8(high));
+        filter_val[k][0] = vaddq_s16(filter_val[k][0], filter_offset_vec);
+        filter_val[k][1] = vaddq_s16(filter_val[k][1], filter_offset_vec);
+      }
+      for (int p = 0; p < 2; p++) {
+        for (int k = 0; k < kPeel; k++) {
+          acc[k] = vmlal_s16(acc[k], vget_low_s16(filter_val[k][p]),
+                             vget_low_s16(input_val[p]));
+        }
+        for (int k = 0; k < kPeel; k++) {
+          acc[k] = vmlal_s16(acc[k], vget_high_s16(filter_val[k][p]),
+                             vget_high_s16(input_val[p]));
+        }
+      }
+    }
+    for (; in <= input_size - 8; in += 8) {
+      const uint8x8_t input_val_u8 = vld1_u8(input_data + in);
+      uint8x8_t filter_val_u8[kPeel];
+      for (int k = 0; k < kPeel; k++) {
+        const uint8* filter_ptr = filter_data + in + (out + k) * input_size;
+        filter_val_u8[k] = vld1_u8(filter_ptr);
+      }
+      int16x8_t input_val;
+      input_val = vreinterpretq_s16_u16(vmovl_u8(input_val_u8));
+      input_val = vaddq_s16(input_val, input_offset_vec);
+      int16x8_t filter_val[kPeel];
+      for (int k = 0; k < kPeel; k++) {
+        filter_val[k] = vreinterpretq_s16_u16(vmovl_u8(filter_val_u8[k]));
+        filter_val[k] = vaddq_s16(filter_val[k], filter_offset_vec);
+      }
+      for (int k = 0; k < kPeel; k++) {
+        acc[k] = vmlal_s16(acc[k], vget_low_s16(filter_val[k]),
+                           vget_low_s16(input_val));
+      }
+      for (int k = 0; k < kPeel; k++) {
+        acc[k] = vmlal_s16(acc[k], vget_high_s16(filter_val[k]),
+                           vget_high_s16(input_val));
+      }
+    }
+    if (in < input_size) {
+      int32 buf[4 * kPeel];
+      for (int k = 0; k < 4; k++) {
+        vst1q_s32(buf + 4 * k, acc[k]);
+      }
+      for (; in < input_size; in++) {
+        int lane = (in + 8 - input_size) % 4;
+        const int32 input_val = input_data[in] + input_offset;
+        for (int k = 0; k < kPeel; k++) {
+          int32 filter_val =
+              filter_data[in + (out + k) * input_size] + filter_offset;
+          buf[lane + 4 * k] += filter_val * input_val;
+        }
+      }
+      for (int k = 0; k < 4; k++) {
+        acc[k] = vld1q_s32(buf + 4 * k);
+      }
+    }
+
+    // Horizontally reduce accumulators
+    int32x2_t pairwise_reduced_acc[kPeel];
+    for (int k = 0; k < kPeel; k++) {
+      pairwise_reduced_acc[k] =
+          vpadd_s32(vget_low_s32(acc[k]), vget_high_s32(acc[k]));
+    }
+    static_assert(kPeel == 4, "the code below currently assumes kPeel = 4");
+    const int32x2_t reduced_lo =
+        vpadd_s32(pairwise_reduced_acc[0], pairwise_reduced_acc[1]);
+    const int32x2_t reduced_hi =
+        vpadd_s32(pairwise_reduced_acc[2], pairwise_reduced_acc[3]);
+    int32x4_t reduced = vcombine_s32(reduced_lo, reduced_hi);
+    // Add bias values.
+    int32x4_t bias_vec = vld1q_s32(bias_ptr);
+    bias_ptr += 4;
+    reduced = vaddq_s32(reduced, bias_vec);
+    // Multiply by the fixed-point multiplier.
+    reduced = vqrdmulhq_n_s32(reduced, output_multiplier);
+    // Rounding-shift-right.
+    using gemmlowp::RoundingDivideByPOT;
+    reduced = RoundingDivideByPOT(reduced, output_shift);
+    // Add the output offset.
+    const int32x4_t output_offset_vec = vdupq_n_s32(output_offset);
+    reduced = vaddq_s32(reduced, output_offset_vec);
+    // Narrow values down to 16 bit signed.
+    const int16x4_t res16 = vqmovn_s32(reduced);
+    // Narrow values down to 8 bit unsigned, saturating.
+    uint8x8_t res8 = vqmovun_s16(vcombine_s16(res16, res16));
+    if (Ac != FusedActivationFunctionType::kNone) {
+      // Apply the clamping from the activation function
+      res8 = vmax_u8(res8, vdup_n_u8(output_activation_min));
+      res8 = vmin_u8(res8, vdup_n_u8(output_activation_max));
+    }
+    // Store results to destination. Assumes 32bit alignment.
+    vst1_lane_u32(reinterpret_cast<uint32*>(output_ptr),
+                  vreinterpret_u32_u8(res8), 0);
+    output_ptr += kPeel;
+  }
+}
+#endif  // USE_NEON
+
+template <FusedActivationFunctionType Ac>
+struct GemmlowpOutputPipeline {
+  typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
+      ColVectorMap;
+  typedef std::tuple<
+      gemmlowp::OutputStageBiasAddition<ColVectorMap>,
+      gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint,
+      gemmlowp::OutputStageClamp, gemmlowp::OutputStageSaturatingCastToUint8>
+      Pipeline;
+  static Pipeline Make(const int32* bias_data, int output_rows,
+                       int32 output_offset, int32 output_multiplier,
+                       int output_shift, int32 output_activation_min,
+                       int32 output_activation_max) {
+    ColVectorMap bias_vector(bias_data, output_rows);
+    gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
+    bias_addition_stage.bias_vector = bias_vector;
+    gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint
+        quantize_down_stage;
+    quantize_down_stage.result_offset_after_shift = output_offset;
+    quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
+    quantize_down_stage.result_shift = output_shift;
+    gemmlowp::OutputStageClamp clamp_stage;
+    clamp_stage.min = output_activation_min;
+    clamp_stage.max = output_activation_max;
+    gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage;
+    return std::make_tuple(bias_addition_stage, quantize_down_stage,
+                           clamp_stage, saturating_cast_stage);
+  }
+};
+
+template <>
+struct GemmlowpOutputPipeline<FusedActivationFunctionType::kNone> {
+  typedef gemmlowp::VectorMap<const int32, gemmlowp::VectorShape::Col>
+      ColVectorMap;
+  typedef std::tuple<
+      gemmlowp::OutputStageBiasAddition<ColVectorMap>,
+      gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint,
+      gemmlowp::OutputStageSaturatingCastToUint8>
+      Pipeline;
+  static Pipeline Make(const int32* bias_data, int output_rows,
+                       int32 output_offset, int32 output_multiplier,
+                       int output_shift, int32 output_activation_min,
+                       int32 output_activation_max) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+    ColVectorMap bias_vector(bias_data, output_rows);
+    gemmlowp::OutputStageBiasAddition<ColVectorMap> bias_addition_stage;
+    bias_addition_stage.bias_vector = bias_vector;
+    gemmlowp::OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint
+        quantize_down_stage;
+    quantize_down_stage.result_offset_after_shift = output_offset;
+    quantize_down_stage.result_fixedpoint_multiplier = output_multiplier;
+    quantize_down_stage.result_shift = output_shift;
+    gemmlowp::OutputStageSaturatingCastToUint8 saturating_cast_stage;
+    return std::make_tuple(bias_addition_stage, quantize_down_stage,
+                           saturating_cast_stage);
+  }
+};
+
+template <FusedActivationFunctionType Ac>
+void FullyConnected(const uint8* input_data, const Dims<4>& input_dims,
+                    int32 input_offset, const uint8* filter_data,
+                    const Dims<4>& filter_dims, int32 filter_offset,
+                    const int32* bias_data, const Dims<4>& bias_dims,
+                    int32 output_offset, int32 output_multiplier,
+                    int output_shift, int32 output_activation_min,
+                    int32 output_activation_max, uint8* output_data,
+                    const Dims<4>& output_dims,
+                    gemmlowp::GemmContext* gemm_context) {
+  gemmlowp::ScopedProfilingLabel label("FullyConnected/8bit");
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  // TODO: This really should be:
+  //     const int batches = ArraySize(output_dims, 1);
+  // but the current --variable_batch hack consists in overwriting the 3rd
+  // dimension with the runtime batch size, as we don't keep track for each
+  // array of which dimension is the batch dimension in it.
+  const int batches = ArraySize(output_dims, 1) * ArraySize(output_dims, 2) *
+                      ArraySize(output_dims, 3);
+#ifdef USE_NEON
+  const int output_size = MatchingArraySize(filter_dims, 1, output_dims, 0);
+  if (batches == 1 && !(output_size % 4)) {
+    return FullyConnectedAsGEMV<Ac>(
+        input_data, input_dims, input_offset, filter_data, filter_dims,
+        filter_offset, bias_data, bias_dims, output_offset, output_multiplier,
+        output_shift, output_activation_min, output_activation_max, output_data,
+        output_dims);
+  }
+#endif  // USE_NEON
+  const int filter_rows = filter_dims.sizes[1];
+  const int filter_cols = filter_dims.sizes[0];
+  DCHECK_EQ(filter_dims.sizes[2], 1);
+  DCHECK_EQ(filter_dims.sizes[3], 1);
+  const int output_rows = output_dims.sizes[0];
+  DCHECK_EQ(output_rows, filter_rows);
+  DCHECK_EQ(bias_dims.sizes[0], output_rows);
+  DCHECK_EQ(bias_dims.sizes[1], 1);
+  DCHECK_EQ(bias_dims.sizes[2], 1);
+  DCHECK_EQ(bias_dims.sizes[3], 1);
+
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::RowMajor> filter_matrix(
+      filter_data, output_rows, filter_cols, filter_cols);
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::ColMajor> input_matrix(
+      input_data, filter_cols, batches, filter_cols);
+  gemmlowp::MatrixMap<uint8, gemmlowp::MapOrder::ColMajor> output_matrix(
+      output_data, output_rows, batches, output_rows);
+  const auto& output_pipeline = GemmlowpOutputPipeline<Ac>::Make(
+      bias_data, output_rows, output_offset, output_multiplier, output_shift,
+      output_activation_min, output_activation_max);
+  gemmlowp::GemmWithOutputPipeline<uint8, uint8,
+                                   gemmlowp::L8R8WithLhsNonzeroBitDepthParams>(
+      gemm_context, filter_matrix, input_matrix, &output_matrix, filter_offset,
+      input_offset, output_pipeline);
+}
+
+template <typename T>
+inline void ExtractPatchIntoBufferColumn(
+    const Dims<4>& input_dims, int w, int h, int b, int kheight, int kwidth,
+    int stride_width, int stride_height, int pad_width, int pad_height,
+    int in_width, int in_height, int in_depth, int single_buffer_length,
+    int buffer_id, const T* in_data, T* conv_buffer_data, uint8 byte_zero) {
+  gemmlowp::ScopedProfilingLabel label("ExtractPatchIntoBufferColumn");
+  // This chunk of code reshapes all the inputs corresponding to
+  // output (b, h, w) to a column vector in conv_buffer(:, buffer_id).
+  const int kwidth_times_indepth = kwidth * in_depth;
+  const int inwidth_times_indepth = in_width * in_depth;
+  const int ih_ungated_start = h * stride_height - pad_height;
+  const int ih_ungated_end = (ih_ungated_start + kheight);
+  const int ih_end = std::min(ih_ungated_end, in_height);
+  const int iw_ungated_start = w * stride_width - pad_width;
+  const int iw_ungated_end = (iw_ungated_start + kwidth);
+  const int iw_end = std::min(iw_ungated_end, in_width);
+  // If the patch is off the edge of the input image, skip writing those rows
+  // and columns from the patch into the output array.
+  const int h_offset = std::max(0, -ih_ungated_start);
+  const int w_offset = std::max(0, -iw_ungated_start);
+  const int ih_start = std::max(0, ih_ungated_start);
+  const int iw_start = std::max(0, iw_ungated_start);
+  const int single_row_num =
+      std::min(kwidth - w_offset, in_width - iw_start) * in_depth;
+  const int output_row_offset = (buffer_id * single_buffer_length);
+  int out_offset =
+      output_row_offset + (h_offset * kwidth + w_offset) * in_depth;
+  int in_offset = Offset(input_dims, 0, iw_start, ih_start, b);
+
+  // Express all of the calculations as padding around the input patch.
+  const int top_padding = h_offset;
+  const int bottom_padding = (ih_ungated_end - ih_end);
+  const int left_padding = w_offset;
+  const int right_padding = (iw_ungated_end - iw_end);
+  assert(single_row_num ==
+         ((kwidth - (left_padding + right_padding)) * in_depth));
+
+  // Write out zeroes to the elements representing the top rows of the input
+  // patch that are off the edge of the input image.
+  if (top_padding > 0) {
+    const int top_row_elements = (top_padding * kwidth * in_depth);
+    memset(conv_buffer_data + output_row_offset, byte_zero,
+           (top_row_elements * sizeof(T)));
+  }
+
+  // If the patch is on the interior of the input image horizontally, just copy
+  // over the rows sequentially, otherwise add zero padding at the start or end.
+  if ((left_padding == 0) && (right_padding == 0)) {
+    for (int ih = ih_start; ih < ih_end; ++ih) {
+      memcpy(conv_buffer_data + out_offset, in_data + in_offset,
+             single_row_num * sizeof(T));
+      out_offset += kwidth_times_indepth;
+      in_offset += inwidth_times_indepth;
+    }
+  } else {
+    for (int ih = ih_start; ih < ih_end; ++ih) {
+      if (left_padding > 0) {
+        const int left_start = (out_offset - (left_padding * in_depth));
+        memset(conv_buffer_data + left_start, byte_zero,
+               (left_padding * in_depth * sizeof(T)));
+      }
+      memcpy(conv_buffer_data + out_offset, in_data + in_offset,
+             single_row_num * sizeof(T));
+      if (right_padding > 0) {
+        const int right_start = (out_offset + single_row_num);
+        memset(conv_buffer_data + right_start, byte_zero,
+               (right_padding * in_depth * sizeof(T)));
+      }
+      out_offset += kwidth_times_indepth;
+      in_offset += inwidth_times_indepth;
+    }
+  }
+
+  // If the bottom of the patch falls off the input image, pad the values
+  // representing those input rows with zeroes.
+  if (bottom_padding > 0) {
+    const int bottom_row_elements = (bottom_padding * kwidth * in_depth);
+    const int bottom_start =
+        output_row_offset +
+        ((top_padding + (ih_end - ih_start)) * kwidth * in_depth);
+    memset(conv_buffer_data + bottom_start, byte_zero,
+           (bottom_row_elements * sizeof(T)));
+  }
+}
+
+template <typename T>
+void Im2col(const T* input_data, const Dims<4>& input_dims, int stride_width,
+            int stride_height, int pad_width, int pad_height, int kheight,
+            int kwidth, uint8 byte_zero, T* output_data,
+            const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Im2col");
+  DCHECK(IsPackedWithoutStrides(input_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int input_depth = ArraySize(input_dims, 0);
+  const int input_width = ArraySize(input_dims, 1);
+  const int input_height = ArraySize(input_dims, 2);
+  const int output_depth = ArraySize(output_dims, 0);
+  const int output_width = ArraySize(output_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+
+  int buffer_id = 0;
+  // Loop over the output nodes.
+  for (int b = 0; b < batches; ++b) {
+    for (int h = 0; h < output_height; ++h) {
+      for (int w = 0; w < output_width; ++w) {
+        ExtractPatchIntoBufferColumn(
+            input_dims, w, h, b, kheight, kwidth, stride_width, stride_height,
+            pad_width, pad_height, input_width, input_height, input_depth,
+            output_depth, buffer_id, input_data, output_data, byte_zero);
+        ++buffer_id;
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void Conv(const float* input_data, const Dims<4>& input_dims,
+          const float* filter_data, const Dims<4>& filter_dims,
+          const float* bias_data, const Dims<4>& bias_dims, int stride_width,
+          int stride_height, int pad_width, int pad_height, float* output_data,
+          const Dims<4>& output_dims, float* im2col_data,
+          const Dims<4>& im2col_dims) {
+  (void)im2col_data;
+  (void)im2col_dims;
+  gemmlowp::ScopedProfilingLabel label("Conv");
+
+  const float* gemm_input_data = nullptr;
+  const Dims<4>* gemm_input_dims = nullptr;
+  const int filter_width = ArraySize(filter_dims, 1);
+  const int filter_height = ArraySize(filter_dims, 2);
+  const bool need_im2col = stride_width != 1 || stride_height != 1 ||
+                           filter_width != 1 || filter_height != 1;
+  if (need_im2col) {
+    DCHECK(im2col_data);
+    Im2col(input_data, input_dims, stride_width, stride_height, pad_width,
+           pad_height, filter_height, filter_width, 0, im2col_data,
+           im2col_dims);
+    gemm_input_data = im2col_data;
+    gemm_input_dims = &im2col_dims;
+  } else {
+    DCHECK(!im2col_data);
+    gemm_input_data = input_data;
+    gemm_input_dims = &input_dims;
+  }
+
+  const auto im2col_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(gemm_input_data, *gemm_input_dims);
+  const auto filter_matrix_map =
+      MapAsMatrixWithLastDimAsCols(filter_data, filter_dims);
+  auto output_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+
+  Gemm(filter_matrix_map.transpose(), im2col_matrix_map, &output_matrix_map);
+
+  AddBiasAndEvalActivationFunction<Ac>(bias_data, bias_dims, output_data,
+                                       output_dims);
+}
+
+template <FusedActivationFunctionType Ac>
+void Conv(const uint8* input_data, const Dims<4>& input_dims,
+          int32 input_offset, const uint8* filter_data,
+          const Dims<4>& filter_dims, int32 filter_offset,
+          const int32* bias_data, const Dims<4>& bias_dims, int stride_width,
+          int stride_height, int pad_width, int pad_height, int32 output_offset,
+          int32 output_multiplier, int output_shift,
+          int32 output_activation_min, int32 output_activation_max,
+          uint8* output_data, const Dims<4>& output_dims, uint8* im2col_data,
+          const Dims<4>& im2col_dims, gemmlowp::GemmContext* gemm_context) {
+  gemmlowp::ScopedProfilingLabel label("Conv/8bit");
+
+  DCHECK(IsPackedWithoutStrides(input_dims));
+  DCHECK(IsPackedWithoutStrides(filter_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+
+  const uint8* gemm_input_data = nullptr;
+  const Dims<4>* gemm_input_dims = nullptr;
+  const int filter_width = ArraySize(filter_dims, 1);
+  const int filter_height = ArraySize(filter_dims, 2);
+  const bool need_im2col = stride_width != 1 || stride_height != 1 ||
+                           filter_width != 1 || filter_height != 1;
+  if (need_im2col) {
+    DCHECK(im2col_data);
+    const int input_zero_point = -input_offset;
+    DCHECK_GE(input_zero_point, 0);
+    DCHECK_LE(input_zero_point, 255);
+    Im2col(input_data, input_dims, stride_width, stride_height, pad_width,
+           pad_height, filter_height, filter_width, input_zero_point,
+           im2col_data, im2col_dims);
+    gemm_input_data = im2col_data;
+    gemm_input_dims = &im2col_dims;
+  } else {
+    DCHECK(!im2col_data);
+    gemm_input_data = input_data;
+    gemm_input_dims = &input_dims;
+  }
+
+  const int gemm_input_rows = gemm_input_dims->sizes[0];
+  const int gemm_input_cols = gemm_input_dims->sizes[1] *
+                              gemm_input_dims->sizes[2] *
+                              gemm_input_dims->sizes[3];
+  const int filter_rows = filter_dims.sizes[3];
+  const int filter_cols =
+      filter_dims.sizes[0] * filter_dims.sizes[1] * filter_dims.sizes[2];
+  const int output_rows = output_dims.sizes[0];
+  const int output_cols =
+      output_dims.sizes[1] * output_dims.sizes[2] * output_dims.sizes[3];
+  DCHECK_EQ(output_rows, filter_rows);
+  DCHECK_EQ(output_cols, gemm_input_cols);
+  DCHECK_EQ(filter_cols, gemm_input_rows);
+  DCHECK_EQ(bias_dims.sizes[0], output_rows);
+  DCHECK_EQ(bias_dims.sizes[1], 1);
+  DCHECK_EQ(bias_dims.sizes[2], 1);
+  DCHECK_EQ(bias_dims.sizes[3], 1);
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::RowMajor> filter_matrix(
+      filter_data, filter_rows, filter_cols);
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::ColMajor> input_matrix(
+      gemm_input_data, gemm_input_rows, gemm_input_cols);
+  gemmlowp::MatrixMap<uint8, gemmlowp::MapOrder::ColMajor> output_matrix(
+      output_data, output_rows, output_cols);
+  const auto& output_pipeline = GemmlowpOutputPipeline<Ac>::Make(
+      bias_data, output_rows, output_offset, output_multiplier, output_shift,
+      output_activation_min, output_activation_max);
+  gemmlowp::GemmWithOutputPipeline<uint8, uint8,
+                                   gemmlowp::L8R8WithLhsNonzeroBitDepthParams>(
+      gemm_context, filter_matrix, input_matrix, &output_matrix, filter_offset,
+      input_offset, output_pipeline);
+}
+
+template <typename T>
+inline void DepthToSpace(const T* input_data, const Dims<4>& input_dims,
+                         int block_size, T* output_data,
+                         const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("DepthToSpace");
+
+  const int input_depth = ArraySize(input_dims, 0);
+  const int input_width = ArraySize(input_dims, 1);
+  const int input_height = ArraySize(input_dims, 2);
+
+  const int output_depth = ArraySize(output_dims, 0);
+  const int batch_size = ArraySize(output_dims, 3);
+
+  // Number of continuous values that we can copy in one interation.
+  const int stride = block_size * output_depth;
+
+  for (int batch = 0; batch < batch_size; ++batch) {
+    for (int in_h = 0; in_h < input_height; ++in_h) {
+      const T* input_ptr = input_data + Offset(input_dims, 0, 0, in_h, batch);
+      for (int offset_h = 0; offset_h < block_size; ++offset_h) {
+        const T* src = input_ptr;
+        for (int in_w = 0; in_w < input_width; ++in_w) {
+          memcpy(output_data, src, stride * sizeof(T));
+          output_data += stride;
+          src += input_depth;
+        }
+        input_ptr += stride;
+      }
+    }
+  }
+}
+
+// legacy, for compatibility with old checked-in code
+template <FusedActivationFunctionType Ac, typename T>
+void Im2col(const T* input_data, const Dims<4>& input_dims, int stride,
+            int pad_width, int pad_height, int kheight, int kwidth,
+            uint8 byte_zero, T* output_data, const Dims<4>& output_dims) {
+  Im2col(input_data, input_dims, stride, stride, pad_width, pad_height, kheight,
+         kwidth, byte_zero, output_data, output_dims);
+}
+
+// legacy, for compatibility with old checked-in code
+template <FusedActivationFunctionType Ac>
+void ConvAsGemm(const float* input_data, const Dims<4>& input_dims,
+                const float* filter_data, const Dims<4>& filter_dims,
+                const float* bias_data, const Dims<4>& bias_dims,
+                float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("ConvAsGemm");
+
+  const auto input_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  const auto filter_matrix_map =
+      MapAsMatrixWithLastDimAsCols(filter_data, filter_dims);
+  auto output_matrix_map =
+      MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+
+  Gemm(filter_matrix_map.transpose(), input_matrix_map, &output_matrix_map);
+
+  AddBiasAndEvalActivationFunction<Ac>(bias_data, bias_dims, output_data,
+                                       output_dims);
+}
+
+// legacy, for compatibility with old checked-in code
+template <FusedActivationFunctionType Ac>
+void ConvAsGemm(const uint8* input_data, const Dims<4>& input_dims,
+                int32 input_offset, const uint8* filter_data,
+                const Dims<4>& filter_dims, int32 filter_offset,
+                const int32* bias_data, const Dims<4>& bias_dims,
+                int32 output_offset, int32 output_multiplier, int output_shift,
+                int32 output_activation_min, int32 output_activation_max,
+                uint8* output_data, const Dims<4>& output_dims,
+                gemmlowp::GemmContext* gemm_context) {
+  gemmlowp::ScopedProfilingLabel label("ConvAsGemm/8bit");
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  const int input_rows = input_dims.sizes[0];
+  const int input_cols =
+      input_dims.sizes[1] * input_dims.sizes[2] * input_dims.sizes[3];
+  const int filter_rows = filter_dims.sizes[3];
+  const int filter_cols =
+      filter_dims.sizes[0] * filter_dims.sizes[1] * filter_dims.sizes[2];
+  const int output_rows = output_dims.sizes[0];
+  const int output_cols =
+      output_dims.sizes[1] * output_dims.sizes[2] * output_dims.sizes[3];
+  DCHECK_EQ(output_rows, filter_rows);
+  DCHECK_EQ(output_cols, input_cols);
+  DCHECK_EQ(filter_cols, input_rows);
+  DCHECK_EQ(bias_dims.sizes[0], output_rows);
+  DCHECK_EQ(bias_dims.sizes[1], 1);
+  DCHECK_EQ(bias_dims.sizes[2], 1);
+  DCHECK_EQ(bias_dims.sizes[3], 1);
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::RowMajor> filter_matrix(
+      filter_data, output_rows, filter_cols, filter_cols);
+  gemmlowp::MatrixMap<const uint8, gemmlowp::MapOrder::ColMajor> input_matrix(
+      input_data, filter_cols, output_cols, filter_cols);
+  gemmlowp::MatrixMap<uint8, gemmlowp::MapOrder::ColMajor> output_matrix(
+      output_data, output_rows, output_cols, output_rows);
+  const auto& output_pipeline = GemmlowpOutputPipeline<Ac>::Make(
+      bias_data, output_rows, output_offset, output_multiplier, output_shift,
+      output_activation_min, output_activation_max);
+  gemmlowp::GemmWithOutputPipeline<uint8, uint8,
+                                   gemmlowp::L8R8WithLhsNonzeroBitDepthParams>(
+      gemm_context, filter_matrix, input_matrix, &output_matrix, filter_offset,
+      input_offset, output_pipeline);
+}
+
+template <typename T>
+inline void SpaceToDepth(const T* input_data, const Dims<4>& input_dims,
+                         int block_size, T* output_data,
+                         const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("SpaceToDepth");
+
+  const int output_depth = ArraySize(output_dims, 0);
+  const int output_width = ArraySize(output_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+
+  const int input_depth = ArraySize(input_dims, 0);
+  const int batch_size = ArraySize(input_dims, 3);
+
+  // Number of continuous values that we can copy in one interation.
+  const int stride = block_size * input_depth;
+
+  for (int batch = 0; batch < batch_size; ++batch) {
+    for (int out_h = 0; out_h < output_height; ++out_h) {
+      T* output_ptr = output_data + Offset(output_dims, 0, 0, out_h, batch);
+      for (int offset_h = 0; offset_h < block_size; ++offset_h) {
+        T* dst = output_ptr;
+        for (int out_w = 0; out_w < output_width; ++out_w) {
+          memcpy(dst, input_data, stride * sizeof(T));
+          input_data += stride;
+          dst += output_depth;
+        }
+        output_ptr += stride;
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void NonGlobalBatchNormalization(
+    const float* input_data, const Dims<4>& input_dims, const float* mean_data,
+    const Dims<4>& mean_dims, const float* multiplier_data,
+    const Dims<4>& multiplier_dims, const float* offset_data,
+    const Dims<4>& offset_dims, float* output_data,
+    const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("NonGlobalBatchNormalization");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height =
+      MatchingArraySize(input_dims, 2, mean_dims, 2, multiplier_dims, 2,
+                        offset_dims, 2, output_dims, 2);
+  const int width =
+      MatchingArraySize(input_dims, 1, mean_dims, 1, multiplier_dims, 1,
+                        offset_dims, 1, output_dims, 1);
+  const int depth =
+      MatchingArraySize(input_dims, 0, mean_dims, 0, multiplier_dims, 0,
+                        offset_dims, 0, output_dims, 0);
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              (input_data[Offset(input_dims, c, x, y, b)] -
+               mean_data[Offset(mean_dims, c, x, y, 0)]) *
+                  multiplier_data[Offset(multiplier_dims, c, x, y, 0)] +
+              offset_data[Offset(offset_dims, c, x, y, 0)]);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void GlobalBatchNormalization(const float* input_data,
+                              const Dims<4>& input_dims, const float* mean_data,
+                              const Dims<4>& mean_dims,
+                              const float* multiplier_data,
+                              const Dims<4>& multiplier_dims,
+                              const float* offset_data,
+                              const Dims<4>& offset_dims, float* output_data,
+                              const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("GlobalBatchNormalization");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth =
+      MatchingArraySize(input_dims, 0, mean_dims, 0, multiplier_dims, 0,
+                        offset_dims, 0, output_dims, 0);
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              (input_data[Offset(input_dims, c, x, y, b)] -
+               mean_data[Offset(mean_dims, c, 0, 0, 0)]) *
+                  multiplier_data[Offset(multiplier_dims, c, 0, 0, 0)] +
+              offset_data[Offset(offset_dims, c, 0, 0, 0)]);
+        }
+      }
+    }
+  }
+}
+
+inline void Relu(const float* input_data, const Dims<4>& input_dims,
+                 float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Relu (not fused)");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          float val = input_data[Offset(input_dims, c, x, y, b)];
+          const float lower = 0;
+          float clamped = val < lower ? lower : val;
+          output_data[Offset(output_dims, c, x, y, b)] = clamped;
+        }
+      }
+    }
+  }
+}
+
+inline void Relu1(const float* input_data, const Dims<4>& input_dims,
+                  float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Relu1 (not fused)");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          float val = input_data[Offset(input_dims, c, x, y, b)];
+          const float upper = 1;
+          const float lower = -1;
+          float clamped = val > upper ? upper : val < lower ? lower : val;
+          output_data[Offset(output_dims, c, x, y, b)] = clamped;
+        }
+      }
+    }
+  }
+}
+
+inline void Relu6(const float* input_data, const Dims<4>& input_dims,
+                  float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Relu6 (not fused)");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          float val = input_data[Offset(input_dims, c, x, y, b)];
+          const float upper = 6;
+          const float lower = 0;
+          float clamped = val > upper ? upper : val < lower ? lower : val;
+          output_data[Offset(output_dims, c, x, y, b)] = clamped;
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void L2Normalization(const float* input_data, const Dims<4>& input_dims,
+                     float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("L2Normalization");
+  static_assert(Ac == FusedActivationFunctionType::kNone, "");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        float squared_l2_norm = 0;
+        for (int c = 0; c < depth; ++c) {
+          float val = input_data[Offset(input_dims, c, x, y, b)];
+          squared_l2_norm += val * val;
+        }
+        float inverse_l2_norm = 1.0f / std::sqrt(squared_l2_norm);
+        for (int c = 0; c < depth; ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] =
+              input_data[Offset(input_dims, c, x, y, b)] * inverse_l2_norm;
+        }
+      }
+    }
+  }
+}
+
+inline void GetInvSqrtQuantizedMultiplier(int32 input, int32* output_inv_sqrt,
+                                          int* output_shift) {
+  *output_shift = 11;
+  while (input >= (1 << 29)) {
+    input /= 4;
+    ++*output_shift;
+  }
+  DCHECK_GT(input, 0);
+  const unsigned max_left_shift_bits = __builtin_clz(input) - 1;
+  const unsigned max_left_shift_bit_pairs = max_left_shift_bits / 2;
+  const unsigned left_shift_bit_pairs = max_left_shift_bit_pairs - 1;
+  *output_shift -= left_shift_bit_pairs;
+  input <<= 2 * left_shift_bit_pairs;
+  DCHECK_GE(input, (1 << 27));
+  DCHECK_LT(input, (1 << 29));
+  using gemmlowp::FixedPoint;
+  using gemmlowp::Rescale;
+  using gemmlowp::SaturatingRoundingMultiplyByPOT;
+  // Using 3 integer bits gives us enough room for the internal arithmetic in
+  // this Newton-Raphson iteration.
+  using F3 = FixedPoint<int32, 3>;
+  using F0 = FixedPoint<int32, 0>;
+  const F3 fixedpoint_input = F3::FromRaw(input >> 1);
+  const F3 fixedpoint_half_input =
+      SaturatingRoundingMultiplyByPOT<-1>(fixedpoint_input);
+  const F3 fixedpoint_half_three =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F3, (1 << 28) + (1 << 27), 1.5);
+  // Newton-Raphson iteration
+  // Naive unoptimized starting guess: x = 1
+  F3 x = F3::One();
+  // Naive unoptimized number of iterations: 5
+  for (int i = 0; i < 5; i++) {
+    const F3 x3 = Rescale<3>(x * x * x);
+    x = Rescale<3>(fixedpoint_half_three * x - fixedpoint_half_input * x3);
+  }
+  const F0 fixedpoint_half_sqrt_2 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F0, 1518500250, std::sqrt(2.) / 2.);
+  x = x * fixedpoint_half_sqrt_2;
+  *output_inv_sqrt = x.raw();
+  if (*output_shift < 0) {
+    *output_inv_sqrt <<= -*output_shift;
+    *output_shift = 0;
+  }
+}
+
+inline void L2Normalization(const uint8* input_data, const Dims<4>& input_dims,
+                            int32 input_zero_point, uint8* output_data,
+                            const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("L2Normalization/8bit");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  DCHECK(IsPackedWithoutStrides(input_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+  DCHECK_EQ(batches, 1);
+  DCHECK_EQ(height, 1);
+  DCHECK_EQ(width, 1);
+  int32 square_l2_norm = 0;
+  for (int i = 0; i < depth; i++) {
+    int32 diff = input_data[i] - input_zero_point;
+    square_l2_norm += diff * diff;
+  }
+  int32 inv_l2norm_multiplier;
+  int inv_l2norm_shift;
+  GetInvSqrtQuantizedMultiplier(square_l2_norm, &inv_l2norm_multiplier,
+                                &inv_l2norm_shift);
+
+  for (int i = 0; i < depth; i++) {
+    int32 diff = input_data[i] - input_zero_point;
+    int32 rescaled_diff = MultiplyByQuantizedMultiplierSmallerThanOne(
+        128 * diff, inv_l2norm_multiplier, inv_l2norm_shift);
+    int32 unclamped_output_val = 128 + rescaled_diff;
+    int32 output_val = std::min(255, std::max(0, unclamped_output_val));
+    output_data[i] = static_cast<uint8>(output_val);
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void Add(const float* input1_data, const Dims<4>& input1_dims,
+         const float* input2_data, const Dims<4>& input2_dims,
+         float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Add");
+  /* const int batches = */ MatchingArraySize(input1_dims, 3, input2_dims, 3,
+                                              output_dims, 3);
+  /* const int height = */ MatchingArraySize(input1_dims, 2, input2_dims, 2,
+                                             output_dims, 2);
+  /* const int width = */ MatchingArraySize(input1_dims, 1, input2_dims, 1,
+                                            output_dims, 1);
+  /* const int depth = */ MatchingArraySize(input1_dims, 0, input2_dims, 0,
+                                            output_dims, 0);
+  DCHECK(IsPackedWithoutStrides(input1_dims));
+  DCHECK(IsPackedWithoutStrides(input2_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+
+  int i = 0;
+  const int size = input1_dims.sizes[3] * input1_dims.strides[3];
+#ifdef USE_NEON
+  const auto zero = vdupq_n_f32(0);
+  const auto six = vdupq_n_f32(6);
+  const auto neg_one = vdupq_n_f32(-1);
+  const auto one = vdupq_n_f32(1);
+  for (; i <= size - 16; i += 16) {
+    auto a10 = vld1q_f32(input1_data + i);
+    auto a11 = vld1q_f32(input1_data + i + 4);
+    auto a12 = vld1q_f32(input1_data + i + 8);
+    auto a13 = vld1q_f32(input1_data + i + 12);
+    auto a20 = vld1q_f32(input2_data + i);
+    auto a21 = vld1q_f32(input2_data + i + 4);
+    auto a22 = vld1q_f32(input2_data + i + 8);
+    auto a23 = vld1q_f32(input2_data + i + 12);
+    auto x0 = vaddq_f32(a10, a20);
+    auto x1 = vaddq_f32(a11, a21);
+    auto x2 = vaddq_f32(a12, a22);
+    auto x3 = vaddq_f32(a13, a23);
+    if (Ac == FusedActivationFunctionType::kRelu ||
+        Ac == FusedActivationFunctionType::kRelu6) {
+      x0 = vmaxq_f32(zero, x0);
+      x1 = vmaxq_f32(zero, x1);
+      x2 = vmaxq_f32(zero, x2);
+      x3 = vmaxq_f32(zero, x3);
+      if (Ac == FusedActivationFunctionType::kRelu6) {
+        x0 = vminq_f32(six, x0);
+        x1 = vminq_f32(six, x1);
+        x2 = vminq_f32(six, x2);
+        x3 = vminq_f32(six, x3);
+      }
+    } else if (Ac == FusedActivationFunctionType::kRelu1) {
+      x0 = vmaxq_f32(neg_one, x0);
+      x1 = vmaxq_f32(neg_one, x1);
+      x2 = vmaxq_f32(neg_one, x2);
+      x3 = vmaxq_f32(neg_one, x3);
+      x0 = vminq_f32(one, x0);
+      x1 = vminq_f32(one, x1);
+      x2 = vminq_f32(one, x2);
+      x3 = vminq_f32(one, x3);
+    }
+    vst1q_f32(output_data + i, x0);
+    vst1q_f32(output_data + i + 4, x1);
+    vst1q_f32(output_data + i + 8, x2);
+    vst1q_f32(output_data + i + 12, x3);
+  }
+  for (; i <= size - 4; i += 4) {
+    auto a1 = vld1q_f32(input1_data + i);
+    auto a2 = vld1q_f32(input2_data + i);
+    auto x = vaddq_f32(a1, a2);
+    if (Ac == FusedActivationFunctionType::kRelu ||
+        Ac == FusedActivationFunctionType::kRelu6) {
+      x = vmaxq_f32(zero, x);
+      if (Ac == FusedActivationFunctionType::kRelu6) {
+        x = vminq_f32(six, x);
+      }
+    } else if (Ac == FusedActivationFunctionType::kRelu1) {
+      x = vmaxq_f32(neg_one, x);
+      x = vminq_f32(one, x);
+    }
+    vst1q_f32(output_data + i, x);
+  }
+#endif  // NEON
+
+  for (; i < size; i++) {
+    auto x = input1_data[i] + input2_data[i];
+    output_data[i] = ActivationFunction<Ac>(x);
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+inline void Add(int left_shift, const uint8* input1_data,
+                const Dims<4>& input1_dims, int32 input1_offset,
+                int32 input1_multiplier, int input1_shift,
+                const uint8* input2_data, const Dims<4>& input2_dims,
+                int32 input2_offset, int32 input2_multiplier, int input2_shift,
+                int32 output_offset, int32 output_multiplier, int output_shift,
+                int32 output_activation_min, int32 output_activation_max,
+                uint8* output_data, const Dims<4>& output_dims) {
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK_LE(output_activation_min, output_activation_max);
+  if (Ac == FusedActivationFunctionType::kNone) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+  }
+  gemmlowp::ScopedProfilingLabel label("Add/8bit");
+  /* const int batches = */ MatchingArraySize(input1_dims, 3, input2_dims, 3,
+                                              output_dims, 3);
+  /* const int height = */ MatchingArraySize(input1_dims, 2, input2_dims, 2,
+                                             output_dims, 2);
+  /* const int width = */ MatchingArraySize(input1_dims, 1, input2_dims, 1,
+                                            output_dims, 1);
+  /* const int depth = */ MatchingArraySize(input1_dims, 0, input2_dims, 0,
+                                            output_dims, 0);
+  DCHECK(IsPackedWithoutStrides(input1_dims));
+  DCHECK(IsPackedWithoutStrides(input2_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+
+  int i = 0;
+  const int size = input1_dims.sizes[3] * input1_dims.strides[3];
+  DCHECK_GT(input1_offset, -256);
+  DCHECK_GT(input2_offset, -256);
+  DCHECK_LT(input1_offset, 256);
+  DCHECK_LT(input2_offset, 256);
+#ifdef USE_NEON
+  for (; i <= size - 8; i += 8) {
+    const auto input1_val_original = vld1_u8(input1_data + i);
+    const auto input2_val_original = vld1_u8(input2_data + i);
+    const auto input1_val_s16 =
+        vreinterpretq_s16_u16(vmovl_u8(input1_val_original));
+    const auto input2_val_s16 =
+        vreinterpretq_s16_u16(vmovl_u8(input2_val_original));
+    const auto input1_val =
+        vaddq_s16(input1_val_s16, vdupq_n_s16(input1_offset));
+    const auto input2_val =
+        vaddq_s16(input2_val_s16, vdupq_n_s16(input2_offset));
+    const auto input1_val_high = vget_high_s16(input1_val);
+    const auto input1_val_low = vget_low_s16(input1_val);
+    const auto input2_val_high = vget_high_s16(input2_val);
+    const auto input2_val_low = vget_low_s16(input2_val);
+    auto x11 = vmovl_s16(input1_val_low);
+    auto x12 = vmovl_s16(input1_val_high);
+    auto x21 = vmovl_s16(input2_val_low);
+    auto x22 = vmovl_s16(input2_val_high);
+    const auto left_shift_dup = vdupq_n_s32(left_shift);
+    x11 = vshlq_s32(x11, left_shift_dup);
+    x12 = vshlq_s32(x12, left_shift_dup);
+    x21 = vshlq_s32(x21, left_shift_dup);
+    x22 = vshlq_s32(x22, left_shift_dup);
+    x11 = vqrdmulhq_n_s32(x11, input1_multiplier);
+    x12 = vqrdmulhq_n_s32(x12, input1_multiplier);
+    x21 = vqrdmulhq_n_s32(x21, input2_multiplier);
+    x22 = vqrdmulhq_n_s32(x22, input2_multiplier);
+    const auto input1_shift_dup = vdupq_n_s32(-input1_shift);
+    const auto input2_shift_dup = vdupq_n_s32(-input2_shift);
+    x11 = vshlq_s32(x11, input1_shift_dup);
+    x12 = vshlq_s32(x12, input1_shift_dup);
+    x21 = vshlq_s32(x21, input2_shift_dup);
+    x22 = vshlq_s32(x22, input2_shift_dup);
+    auto s1 = vaddq_s32(x11, x21);
+    auto s2 = vaddq_s32(x12, x22);
+    s1 = vqrdmulhq_n_s32(s1, output_multiplier);
+    s2 = vqrdmulhq_n_s32(s2, output_multiplier);
+    using gemmlowp::RoundingDivideByPOT;
+    s1 = RoundingDivideByPOT(s1, output_shift);
+    s2 = RoundingDivideByPOT(s2, output_shift);
+    const auto s1_narrowed = vmovn_s32(s1);
+    const auto s2_narrowed = vmovn_s32(s2);
+    const auto s = vaddq_s16(vcombine_s16(s1_narrowed, s2_narrowed),
+                             vdupq_n_s16(output_offset));
+    vst1_u8(output_data + i, vqmovun_s16(s));
+  }
+#endif  // NEON
+
+  for (; i < size; i++) {
+    const int32 input1_val = input1_offset + input1_data[i];
+    const int32 input2_val = input2_offset + input2_data[i];
+    const int32 shifted_input1_val = input1_val * (1 << left_shift);
+    const int32 shifted_input2_val = input2_val * (1 << left_shift);
+    const int32 scaled_input1_val = MultiplyByQuantizedMultiplierSmallerThanOne(
+        shifted_input1_val, input1_multiplier, input1_shift);
+    const int32 scaled_input2_val = MultiplyByQuantizedMultiplierSmallerThanOne(
+        shifted_input2_val, input2_multiplier, input2_shift);
+    const int32 raw_sum = scaled_input1_val + scaled_input2_val;
+    const int32 raw_output = MultiplyByQuantizedMultiplierSmallerThanOne(
+                                 raw_sum, output_multiplier, output_shift) +
+                             output_offset;
+    const int32 clamped_output = std::min(
+        output_activation_max, std::max(output_activation_min, raw_output));
+    output_data[i] = static_cast<uint8>(clamped_output);
+  }
+}
+
+
+// TODO: We can implement BroadcastAdd on buffers of arbitrary
+// dimensionality if the runtime code does a single loop over one dimension
+// that handles broadcasting as the base case. The code generator would then
+// generate max(D1, D2) nested for loops.
+// TODO: BroadcastAdd is intentionally duplicated from
+// reference_ops.h. Once an optimized version is implemented and NdArrayDesc<T>
+// is no longer referenced in this file, move NdArrayDesc<T> from types.h to
+// reference_ops.h.
+template <FusedActivationFunctionType Ac>
+void BroadcastAdd(const float* input1_data, const Dims<4>& input1_dims,
+                  const float* input2_data, const Dims<4>& input2_dims,
+                  float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("BroadcastAdd");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_dims, input2_dims, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < ArraySize(output_dims, 3); ++b) {
+    for (int y = 0; y < ArraySize(output_dims, 2); ++y) {
+      for (int x = 0; x < ArraySize(output_dims, 1); ++x) {
+        for (int c = 0; c < ArraySize(output_dims, 0); ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              input1_data[SubscriptToIndex(desc1, c, x, y, b)] +
+              input2_data[SubscriptToIndex(desc2, c, x, y, b)]);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+inline void BroadcastAdd(int left_shift, const uint8* input1_data,
+                         const Dims<4>& input1_dims, int32 input1_offset,
+                         int32 input1_multiplier, int input1_shift,
+                         const uint8* input2_data, const Dims<4>& input2_dims,
+                         int32 input2_offset, int32 input2_multiplier,
+                         int input2_shift, int32 output_offset,
+                         int32 output_multiplier, int output_shift,
+                         int32 output_activation_min,
+                         int32 output_activation_max, uint8* output_data,
+                         const Dims<4>& output_dims) {
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK_LE(output_activation_min, output_activation_max);
+  if (Ac == FusedActivationFunctionType::kNone) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+  }
+  gemmlowp::ScopedProfilingLabel label("BroadcastAdd/8bit");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_dims, input2_dims, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < ArraySize(output_dims, 3); ++b) {
+    for (int y = 0; y < ArraySize(output_dims, 2); ++y) {
+      for (int x = 0; x < ArraySize(output_dims, 1); ++x) {
+        for (int c = 0; c < ArraySize(output_dims, 0); ++c) {
+          const int32 input1_val =
+              input1_offset + input1_data[SubscriptToIndex(desc1, c, x, y, b)];
+          const int32 input2_val =
+              input2_offset + input2_data[SubscriptToIndex(desc2, c, x, y, b)];
+          const int32 shifted_input1_val = input1_val * (1 << left_shift);
+          const int32 shifted_input2_val = input2_val * (1 << left_shift);
+          const int32 scaled_input1_val =
+              MultiplyByQuantizedMultiplierSmallerThanOne(
+                  shifted_input1_val, input1_multiplier, input1_shift);
+          const int32 scaled_input2_val =
+              MultiplyByQuantizedMultiplierSmallerThanOne(
+                  shifted_input2_val, input2_multiplier, input2_shift);
+          const int32 raw_sum = scaled_input1_val + scaled_input2_val;
+          const int32 raw_output =
+              MultiplyByQuantizedMultiplierSmallerThanOne(
+                  raw_sum, output_multiplier, output_shift) +
+              output_offset;
+          const int32 clamped_output =
+              std::min(output_activation_max,
+                       std::max(output_activation_min, raw_output));
+          output_data[Offset(output_dims, c, x, y, b)] =
+              static_cast<uint8>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void Mul(const float* input1_data, const Dims<4>& input1_dims,
+         const float* input2_data, const Dims<4>& input2_dims,
+         float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Mul");
+  /* const int batches = */ MatchingArraySize(input1_dims, 3, input2_dims, 3,
+                                              output_dims, 3);
+  /* const int height = */ MatchingArraySize(input1_dims, 2, input2_dims, 2,
+                                             output_dims, 2);
+  /* const int width = */ MatchingArraySize(input1_dims, 1, input2_dims, 1,
+                                            output_dims, 1);
+  /* const int depth = */ MatchingArraySize(input1_dims, 0, input2_dims, 0,
+                                            output_dims, 0);
+  DCHECK(IsPackedWithoutStrides(input1_dims));
+  DCHECK(IsPackedWithoutStrides(input2_dims));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+
+  int i = 0;
+  const int size = input1_dims.sizes[3] * input1_dims.strides[3];
+#ifdef USE_NEON
+  const auto zero = vdupq_n_f32(0);
+  const auto six = vdupq_n_f32(6);
+  const auto neg_one = vdupq_n_f32(-1);
+  const auto one = vdupq_n_f32(1);
+  for (; i <= size - 16; i += 16) {
+    auto a10 = vld1q_f32(input1_data + i);
+    auto a11 = vld1q_f32(input1_data + i + 4);
+    auto a12 = vld1q_f32(input1_data + i + 8);
+    auto a13 = vld1q_f32(input1_data + i + 12);
+    auto a20 = vld1q_f32(input2_data + i);
+    auto a21 = vld1q_f32(input2_data + i + 4);
+    auto a22 = vld1q_f32(input2_data + i + 8);
+    auto a23 = vld1q_f32(input2_data + i + 12);
+    auto x0 = vmulq_f32(a10, a20);
+    auto x1 = vmulq_f32(a11, a21);
+    auto x2 = vmulq_f32(a12, a22);
+    auto x3 = vmulq_f32(a13, a23);
+    if (Ac == FusedActivationFunctionType::kRelu ||
+        Ac == FusedActivationFunctionType::kRelu6) {
+      x0 = vmaxq_f32(zero, x0);
+      x1 = vmaxq_f32(zero, x1);
+      x2 = vmaxq_f32(zero, x2);
+      x3 = vmaxq_f32(zero, x3);
+      if (Ac == FusedActivationFunctionType::kRelu6) {
+        x0 = vminq_f32(six, x0);
+        x1 = vminq_f32(six, x1);
+        x2 = vminq_f32(six, x2);
+        x3 = vminq_f32(six, x3);
+      }
+    } else if (Ac == FusedActivationFunctionType::kRelu1) {
+      x0 = vmaxq_f32(neg_one, x0);
+      x1 = vmaxq_f32(neg_one, x1);
+      x2 = vmaxq_f32(neg_one, x2);
+      x3 = vmaxq_f32(neg_one, x3);
+      x0 = vminq_f32(one, x0);
+      x1 = vminq_f32(one, x1);
+      x2 = vminq_f32(one, x2);
+      x3 = vminq_f32(one, x3);
+    }
+    vst1q_f32(output_data + i, x0);
+    vst1q_f32(output_data + i + 4, x1);
+    vst1q_f32(output_data + i + 8, x2);
+    vst1q_f32(output_data + i + 12, x3);
+  }
+  for (; i <= size - 4; i += 4) {
+    auto a1 = vld1q_f32(input1_data + i);
+    auto a2 = vld1q_f32(input2_data + i);
+    auto x = vmulq_f32(a1, a2);
+    if (Ac == FusedActivationFunctionType::kRelu ||
+        Ac == FusedActivationFunctionType::kRelu6) {
+      x = vmaxq_f32(zero, x);
+      if (Ac == FusedActivationFunctionType::kRelu6) {
+        x = vminq_f32(six, x);
+      }
+    } else if (Ac == FusedActivationFunctionType::kRelu1) {
+      x = vmaxq_f32(neg_one, x);
+      x = vminq_f32(one, x);
+    }
+    vst1q_f32(output_data + i, x);
+  }
+#endif  // NEON
+
+  for (; i < size; i++) {
+    auto x = input1_data[i] * input2_data[i];
+    output_data[i] = ActivationFunction<Ac>(x);
+  }
+}
+
+// TODO: We can implement BroadcastMul on buffers of arbitrary
+// dimensionality if the runtime code does a single loop over one dimension
+// that handles broadcasting as the base case. The code generator would then
+// generate max(D1, D2) nested for loops.
+// TODO: BroadcastMul is intentionally duplicated from
+// reference_ops.h. Once an optimized version is implemented and NdArrayDesc<T>
+// is no longer referenced in this file, move NdArrayDesc<T> from types.h to
+// reference_ops.h.
+template <FusedActivationFunctionType Ac>
+void BroadcastMul(const float* input1_data, const Dims<4>& input1_dims,
+                  const float* input2_data, const Dims<4>& input2_dims,
+                  float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("BroadcastMul");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_dims, input2_dims, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < ArraySize(output_dims, 3); ++b) {
+    for (int y = 0; y < ArraySize(output_dims, 2); ++y) {
+      for (int x = 0; x < ArraySize(output_dims, 1); ++x) {
+        for (int c = 0; c < ArraySize(output_dims, 0); ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              input1_data[SubscriptToIndex(desc1, c, x, y, b)] *
+              input2_data[SubscriptToIndex(desc2, c, x, y, b)]);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+inline void BroadcastMul(const uint8* input1_data, const Dims<4>& input1_dims,
+                         int32 input1_offset, const uint8* input2_data,
+                         const Dims<4>& input2_dims, int32 input2_offset,
+                         int32 output_offset, int32 output_multiplier,
+                         int output_shift, int32 output_activation_min,
+                         int32 output_activation_max, uint8* output_data,
+                         const Dims<4>& output_dims) {
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK_LE(output_activation_min, output_activation_max);
+  if (Ac == FusedActivationFunctionType::kNone) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+  }
+  gemmlowp::ScopedProfilingLabel label("BroadcastMul/8bit");
+
+  NdArrayDesc<4> desc1;
+  NdArrayDesc<4> desc2;
+  NdArrayDescsForElementwiseBroadcast(input1_dims, input2_dims, &desc1, &desc2);
+
+  // In Tensorflow, the dimensions are canonically named (batch_number, row,
+  // col, channel), with extents (batches, height, width, depth), with the
+  // trailing dimension changing most rapidly (channels has the smallest stride,
+  // typically 1 element).
+  //
+  // In generated C code, we store arrays with the dimensions reversed. The
+  // first dimension has smallest stride.
+  //
+  // We name our variables by their Tensorflow convention, but generate C code
+  // nesting loops such that the innermost loop has the smallest stride for the
+  // best cache behavior.
+  for (int b = 0; b < ArraySize(output_dims, 3); ++b) {
+    for (int y = 0; y < ArraySize(output_dims, 2); ++y) {
+      for (int x = 0; x < ArraySize(output_dims, 1); ++x) {
+        for (int c = 0; c < ArraySize(output_dims, 0); ++c) {
+          const int32 input1_val =
+              input1_offset + input1_data[SubscriptToIndex(desc1, c, x, y, b)];
+          const int32 input2_val =
+              input2_offset + input2_data[SubscriptToIndex(desc2, c, x, y, b)];
+          const int32 unclamped_result =
+              output_offset +
+              MultiplyByQuantizedMultiplierSmallerThanOne(
+                  input1_val * input2_val, output_multiplier, output_shift);
+          const int32 clamped_output =
+              std::min(output_activation_max,
+                       std::max(output_activation_min, unclamped_result));
+          output_data[Offset(output_dims, c, x, y, b)] =
+              static_cast<uint8>(clamped_output);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac, typename Scalar>
+void Concatenation(int concat_dim, const Scalar* const* input_data,
+                   const Dims<4>* const* input_dims, int inputs_count,
+                   Scalar* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Concatenation");
+  DCHECK_GT(inputs_count, 1);
+  int concat_size = 0;
+  for (int i = 0; i < inputs_count; i++) {
+    for (int j = 0; j < 4; j++) {
+      if (j != concat_dim) {
+        MatchingArraySize(*input_dims[i], j, output_dims, j);
+      }
+    }
+    concat_size += ArraySize(*input_dims[i], concat_dim);
+  }
+  DCHECK_EQ(concat_size, ArraySize(output_dims, concat_dim));
+  DCHECK(IsPackedWithoutStrides(output_dims));
+  // for now we dont have a model with a Concatenation
+  // with fused activation function.
+  DCHECK(Ac == FusedActivationFunctionType::kNone);
+  int outer_size = 1;
+  for (int i = concat_dim + 1; i < 4; i++) {
+    outer_size *= output_dims.sizes[i];
+  }
+  Scalar* output_ptr = output_data;
+  for (int k = 0; k < outer_size; k++) {
+    for (int i = 0; i < inputs_count; ++i) {
+      const int copy_size =
+          input_dims[i]->sizes[concat_dim] * input_dims[i]->strides[concat_dim];
+      memcpy(output_ptr, input_data[i] + k * copy_size,
+             copy_size * sizeof(Scalar));
+      output_ptr += copy_size;
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac, typename Scalar>
+void DepthConcatenation(const Scalar* const* input_data,
+                        const Dims<4>* const* input_dims, int inputs_count,
+                        Scalar* output_data, const Dims<4>& output_dims) {
+  Concatenation<Ac, Scalar>(0, input_data, input_dims, inputs_count,
+                            output_data, output_dims);
+}
+
+inline void LstmCell(const float* input_data, const Dims<4>& input_dims,
+                     const float* prev_activ_data,
+                     const Dims<4>& prev_activ_dims, const float* weights_data,
+                     const Dims<4>& weights_dims, const float* bias_data,
+                     const Dims<4>& bias_dims, const float* prev_state_data,
+                     const Dims<4>& prev_state_dims, float* output_state_data,
+                     const Dims<4>& output_state_dims, float* output_activ_data,
+                     const Dims<4>& output_activ_dims, float* concat_temp_data,
+                     const Dims<4>& concat_temp_dims, float* activ_temp_data,
+                     const Dims<4>& activ_temp_dims) {
+  gemmlowp::ScopedProfilingLabel label("LstmCell");
+  MatchingArraySize(  // batches
+      input_dims, 3, prev_activ_dims, 3, prev_state_dims, 3, output_state_dims,
+      3, output_activ_dims, 3);
+  MatchingArraySize(  // height
+      input_dims, 2, prev_activ_dims, 2, prev_state_dims, 2, output_state_dims,
+      2, output_activ_dims, 2);
+  MatchingArraySize(  // width
+      input_dims, 1, prev_activ_dims, 1, prev_state_dims, 1, output_state_dims,
+      1, output_activ_dims, 1);
+  CHECK_EQ(ArraySize(weights_dims, 2), 1);
+  CHECK_EQ(ArraySize(weights_dims, 3), 1);
+  const int input_depth = ArraySize(input_dims, 0);
+  const int prev_activ_depth = ArraySize(prev_activ_dims, 0);
+  const int total_input_depth = prev_activ_depth + input_depth;
+  CHECK_EQ(ArraySize(weights_dims, 0), total_input_depth);
+  CHECK_EQ(MatchingArraySize(bias_dims, 1, bias_dims, 2, bias_dims, 3), 1);
+  const int intern_activ_depth = MatchingArraySize(
+      weights_dims, 1,
+      bias_dims,    0);
+  CHECK_EQ(intern_activ_depth % 4, 0);
+  const int output_depth = MatchingArraySize(
+      prev_state_dims,   0,
+      prev_activ_dims,   0,
+      output_state_dims, 0,
+      output_activ_dims, 0);
+  CHECK_EQ(output_depth, intern_activ_depth / 4);
+
+  // Concatenate prev_activ and input data together
+  std::vector<float const*> concat_input_arrays_data;
+  std::vector<Dims<4> const*> concat_input_arrays_dims;
+  concat_input_arrays_data.push_back(input_data);
+  concat_input_arrays_data.push_back(prev_activ_data);
+  concat_input_arrays_dims.push_back(&input_dims);
+  concat_input_arrays_dims.push_back(&prev_activ_dims);
+  Concatenation<FusedActivationFunctionType::kNone, float>(
+      0, &(concat_input_arrays_data[0]), &(concat_input_arrays_dims[0]),
+      concat_input_arrays_data.size(), concat_temp_data, concat_temp_dims);
+
+  // Fully connected
+  FullyConnected<FusedActivationFunctionType::kNone>(
+      concat_temp_data, concat_temp_dims, weights_data, weights_dims, bias_data,
+      bias_dims, activ_temp_data, activ_temp_dims);
+
+  // Map raw arrays to Eigen arrays so we can use Eigen's optimized array
+  // operations.
+  ArrayMap<float> activ_temp_map =
+      MapAsArrayWithFirstDimAsRows(activ_temp_data, activ_temp_dims);
+  auto input_gate_sm = activ_temp_map.block(0 * output_depth, 0, output_depth,
+                                            activ_temp_map.cols());
+  auto new_input_sm = activ_temp_map.block(1 * output_depth, 0, output_depth,
+                                           activ_temp_map.cols());
+  auto forget_gate_sm = activ_temp_map.block(2 * output_depth, 0, output_depth,
+                                             activ_temp_map.cols());
+  auto output_gate_sm = activ_temp_map.block(3 * output_depth, 0, output_depth,
+                                             activ_temp_map.cols());
+  ArrayMap<const float> prev_state_map =
+      MapAsArrayWithFirstDimAsRows(prev_state_data, prev_state_dims);
+  ArrayMap<float> output_state_map =
+      MapAsArrayWithFirstDimAsRows(output_state_data, output_state_dims);
+  ArrayMap<float> output_activ_map =
+      MapAsArrayWithFirstDimAsRows(output_activ_data, output_activ_dims);
+
+  // Combined memory state and final output calculation
+  gemmlowp::ScopedProfilingLabel label2("MemoryStateAndFinalOutput");
+  output_state_map =
+      input_gate_sm.unaryExpr(Eigen::internal::scalar_sigmoid_op<float>()) *
+          new_input_sm.tanh() +
+      forget_gate_sm.unaryExpr(Eigen::internal::scalar_sigmoid_op<float>()) *
+          prev_state_map;
+  output_activ_map =
+      output_gate_sm.unaryExpr(Eigen::internal::scalar_sigmoid_op<float>()) *
+      output_state_map.tanh();
+}
+
+template <FusedActivationFunctionType Ac, typename Scalar>
+void TensorFlowSplit(const Scalar* input_data, const Dims<4>& input_dims,
+                     int outputs_count, Scalar* const* output_data,
+                     const Dims<4>* const* output_dims) {
+  gemmlowp::ScopedProfilingLabel label("TensorFlowSplit");
+  DCHECK_GE(outputs_count, 1);
+  for (int i = 0; i < outputs_count; i++) {
+    /* batches = */ MatchingArraySize(*output_dims[i], 3, input_dims, 3);
+    /* height = */ MatchingArraySize(*output_dims[i], 2, input_dims, 2);
+    /* width = */ MatchingArraySize(*output_dims[i], 1, input_dims, 1);
+  }
+  const int batches = MatchingArraySize(*output_dims[0], 3, input_dims, 3);
+  const int height = MatchingArraySize(*output_dims[0], 2, input_dims, 2);
+  const int width = MatchingArraySize(*output_dims[0], 1, input_dims, 1);
+  DCHECK(IsPackedWithoutStrides(input_dims));
+  // for now we dont have a model with a TensorFlowSplit
+  // with fused activation function.
+  DCHECK(Ac == FusedActivationFunctionType::kNone);
+  const int whb = width * height * batches;
+  const Scalar* input_ptr = input_data;
+  for (int k = 0; k < whb; k++) {
+    for (int i = 0; i < outputs_count; ++i) {
+      memcpy(output_data[i] + k * output_dims[i]->sizes[0], input_ptr,
+             output_dims[i]->sizes[0] * sizeof(Scalar));
+      input_ptr += output_dims[i]->sizes[0];
+    }
+  }
+}
+
+inline int NodeOffset(int b, int h, int w, int height, int width) {
+  return (b * height + h) * width + w;
+}
+
+template <FusedActivationFunctionType Ac>
+void AveragePool(const float* input_data, const Dims<4>& input_dims,
+                 int stride_width, int stride_height,
+                 int pad_width, int pad_height, int kwidth, int kheight,
+                 float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("AveragePool");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int input_height = ArraySize(input_dims, 2);
+  const int input_width = ArraySize(input_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+  const int output_width = ArraySize(output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  const auto in_mat = MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  auto out_mat = MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+  // TODO: get rid of the dynamic memory allocation here!
+  Eigen::VectorXf out_count(out_mat.cols());
+  out_count.setZero();
+  // Prefill the output to 0.
+  out_mat.setZero();
+  for (int b = 0; b < batches; ++b) {
+    for (int h = 0; h < input_height; ++h) {
+      for (int w = 0; w < input_width; ++w) {
+        // (h_start, h_end) * (w_start, w_end) is the range that the input
+        // vector projects to.
+        int hpad = h + pad_height;
+        int wpad = w + pad_width;
+        int h_start = (hpad < kheight) ? 0 : (hpad - kheight) / stride_height + 1;
+        int h_end = std::min(hpad / stride_height + 1, output_height);
+        int w_start = (wpad < kwidth) ? 0 : (wpad - kwidth) / stride_width + 1;
+        int w_end = std::min(wpad / stride_width + 1, output_width);
+        // compute elementwise sum
+        for (int ph = h_start; ph < h_end; ++ph) {
+          for (int pw = w_start; pw < w_end; ++pw) {
+            int out_offset = NodeOffset(b, ph, pw, output_height, output_width);
+            out_mat.col(out_offset) +=
+                in_mat.col(NodeOffset(b, h, w, input_height, input_width));
+            out_count(out_offset)++;
+          }
+        }
+      }
+    }
+  }
+  // Divide the output by the actual number of elements being averaged over
+  DCHECK_GT(out_count.minCoeff(), 0);
+  out_mat.array().rowwise() /= out_count.transpose().array();
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      for (int x = 0; x < output_width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              output_data[Offset(output_dims, c, x, y, b)]);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void AveragePool(const uint8* input_data, const Dims<4>& input_dims,
+                 int stride_width, int stride_height,
+                 int pad_width, int pad_height, int filter_width,
+                 int filter_height, int32 output_activation_min,
+                 int32 output_activation_max, uint8* output_data,
+                 const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("AveragePool/8bit");
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK_LE(output_activation_min, output_activation_max);
+  if (Ac == FusedActivationFunctionType::kNone) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+  }
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  const int input_height = ArraySize(input_dims, 2);
+  const int input_width = ArraySize(input_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+  const int output_width = ArraySize(output_dims, 1);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        const int in_y_origin = (out_y * stride_height) - pad_height;
+        const int filter_x_start = std::max(0, -in_x_origin);
+        const int filter_x_end =
+            std::min(filter_width, input_width - in_x_origin);
+        const int filter_y_start = std::max(0, -in_y_origin);
+        const int filter_y_end =
+            std::min(filter_height, input_height - in_y_origin);
+        const int filter_count =
+            (filter_x_end - filter_x_start) * (filter_y_end - filter_y_start);
+        // TODO: Add a dynamic buffer allocation path instead of hardcoded size.
+        static constexpr int kAccBufferMaxSize = 2048;
+        DCHECK_LE(depth, kAccBufferMaxSize);
+        uint16 acc[kAccBufferMaxSize];
+        memset(acc, 0, depth * sizeof(acc[0]));
+        const uint8* input_ptr =
+            input_data + input_dims.strides[1] * in_x_origin +
+            input_dims.strides[2] * in_y_origin + input_dims.strides[3] * batch;
+        for (int fy = filter_y_start; fy < filter_y_end; fy++) {
+          const uint8* input_row_ptr = input_ptr + fy * input_dims.strides[2] +
+                                       filter_x_start * input_dims.strides[1];
+          for (int fx = filter_x_start; fx < filter_x_end; fx++) {
+            int channel = 0;
+#ifdef USE_NEON
+            for (; channel <= depth - 16; channel += 16) {
+              uint16x8_t acc_reg[2];
+              for (int i = 0; i < 2; i++) {
+                acc_reg[i] = vld1q_u16(acc + channel + 8 * i);
+              }
+              uint8x16_t input_reg = vld1q_u8(input_row_ptr);
+              input_row_ptr += 16;
+              acc_reg[0] = vaddw_u8(acc_reg[0], vget_low_u8(input_reg));
+              acc_reg[1] = vaddw_u8(acc_reg[1], vget_high_u8(input_reg));
+              for (int i = 0; i < 2; i++) {
+                vst1q_u16(acc + channel + 8 * i, acc_reg[i]);
+              }
+            }
+            for (; channel <= depth - 8; channel += 8) {
+              uint16x8_t acc_reg = vld1q_u16(acc + channel);
+              uint8x8_t input_reg = vld1_u8(input_row_ptr);
+              input_row_ptr += 8;
+              acc_reg = vaddw_u8(acc_reg, input_reg);
+              vst1q_u16(acc + channel, acc_reg);
+            }
+#endif
+            for (; channel < depth; ++channel) {
+              acc[channel] += *input_row_ptr++;
+            }
+          }
+        }
+        uint8* output_ptr =
+            output_data + Offset(output_dims, 0, out_x, out_y, batch);
+        int channel = 0;
+#ifdef USE_NEON
+#define AVGPOOL_DIVIDING_BY(FILTER_COUNT)                              \
+  if (filter_count == FILTER_COUNT) {                                  \
+    for (; channel <= depth - 8; channel += 8) {                       \
+      uint16 buf[8];                                                   \
+      for (int i = 0; i < 8; i++) {                                    \
+        buf[i] = (acc[channel + i] + FILTER_COUNT / 2) / FILTER_COUNT; \
+      }                                                                \
+      uint8x8_t buf8 = vqmovn_u16(vld1q_u16(buf));                     \
+      buf8 = vmin_u8(buf8, vdup_n_u8(output_activation_max));          \
+      buf8 = vmax_u8(buf8, vdup_n_u8(output_activation_min));          \
+      vst1_u8(output_ptr + channel, buf8);                             \
+    }                                                                  \
+  }
+        AVGPOOL_DIVIDING_BY(9)
+        AVGPOOL_DIVIDING_BY(15)
+#undef AVGPOOL_DIVIDING_BY
+        for (; channel <= depth - 8; channel += 8) {
+          uint16 buf[8];
+          for (int i = 0; i < 8; i++) {
+            buf[i] = (acc[channel + i] + filter_count / 2) / filter_count;
+          }
+          uint8x8_t buf8 = vqmovn_u16(vld1q_u16(buf));
+          buf8 = vmin_u8(buf8, vdup_n_u8(output_activation_max));
+          buf8 = vmax_u8(buf8, vdup_n_u8(output_activation_min));
+          vst1_u8(output_ptr + channel, buf8);
+        }
+#endif
+        for (; channel < depth; ++channel) {
+          uint16 a = (acc[channel] + filter_count / 2) / filter_count;
+          a = std::max<uint16>(a, output_activation_min);
+          a = std::min<uint16>(a, output_activation_max);
+          output_ptr[channel] = static_cast<uint8>(a);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void MaxPool(const float* input_data, const Dims<4>& input_dims,
+             int stride_width, int stride_height,
+             int pad_width, int pad_height, int kwidth, int kheight,
+             float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("MaxPool");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int input_height = ArraySize(input_dims, 2);
+  const int input_width = ArraySize(input_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+  const int output_width = ArraySize(output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  const auto in_mat = MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  auto out_mat = MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+  // Prefill the output to minimum representable float value
+  out_mat.setConstant(std::numeric_limits<float>::lowest());
+  for (int b = 0; b < batches; ++b) {
+    for (int h = 0; h < input_height; ++h) {
+      for (int w = 0; w < input_width; ++w) {
+        // (h_start, h_end) * (w_start, w_end) is the range that the input
+        // vector projects to.
+        int hpad = h + pad_height;
+        int wpad = w + pad_width;
+        int h_start = (hpad < kheight) ? 0 : (hpad - kheight) / stride_height + 1;
+        int h_end = std::min(hpad / stride_height + 1, output_height);
+        int w_start = (wpad < kwidth) ? 0 : (wpad - kwidth) / stride_width + 1;
+        int w_end = std::min(wpad / stride_width + 1, output_width);
+        // compute elementwise sum
+        for (int ph = h_start; ph < h_end; ++ph) {
+          for (int pw = w_start; pw < w_end; ++pw) {
+            int out_offset = NodeOffset(b, ph, pw, output_height, output_width);
+            out_mat.col(out_offset) =
+                out_mat.col(out_offset)
+                    .cwiseMax(in_mat.col(
+                        NodeOffset(b, h, w, input_height, input_width)));
+          }
+        }
+      }
+    }
+  }
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      for (int x = 0; x < output_width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          output_data[Offset(output_dims, c, x, y, b)] = ActivationFunction<Ac>(
+              output_data[Offset(output_dims, c, x, y, b)]);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void MaxPool(const uint8* input_data, const Dims<4>& input_dims,
+             int stride_width, int stride_height,
+             int pad_width, int pad_height, int filter_width, int filter_height,
+             int32 output_activation_min, int32 output_activation_max,
+             uint8* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("MaxPool/8bit");
+  static_assert(Ac == FusedActivationFunctionType::kNone ||
+                    Ac == FusedActivationFunctionType::kRelu ||
+                    Ac == FusedActivationFunctionType::kRelu6 ||
+                    Ac == FusedActivationFunctionType::kRelu1,
+                "");
+  DCHECK_LE(output_activation_min, output_activation_max);
+  if (Ac == FusedActivationFunctionType::kNone) {
+    DCHECK_EQ(output_activation_min, 0);
+    DCHECK_EQ(output_activation_max, 255);
+  }
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  const int input_height = ArraySize(input_dims, 2);
+  const int input_width = ArraySize(input_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+  const int output_width = ArraySize(output_dims, 1);
+  for (int batch = 0; batch < batches; ++batch) {
+    for (int out_y = 0; out_y < output_height; ++out_y) {
+      for (int out_x = 0; out_x < output_width; ++out_x) {
+        const int in_x_origin = (out_x * stride_width) - pad_width;
+        const int in_y_origin = (out_y * stride_height) - pad_height;
+        const int filter_x_start = std::max(0, -in_x_origin);
+        const int filter_x_end =
+            std::min(filter_width, input_width - in_x_origin);
+        const int filter_y_start = std::max(0, -in_y_origin);
+        const int filter_y_end =
+            std::min(filter_height, input_height - in_y_origin);
+        // TODO: Add a dynamic buffer allocation path instead of hardcoded size.
+        static constexpr int kAccBufferMaxSize = 2048;
+        DCHECK_LE(depth, kAccBufferMaxSize);
+        uint8 acc[kAccBufferMaxSize];
+        memset(acc, 0, depth * sizeof(acc[0]));
+        const uint8* input_ptr =
+            input_data + input_dims.strides[1] * in_x_origin +
+            input_dims.strides[2] * in_y_origin + input_dims.strides[3] * batch;
+        for (int fy = filter_y_start; fy < filter_y_end; fy++) {
+          const uint8* input_row_ptr = input_ptr + fy * input_dims.strides[2] +
+                                       filter_x_start * input_dims.strides[1];
+          for (int fx = filter_x_start; fx < filter_x_end; fx++) {
+            int channel = 0;
+#ifdef USE_NEON
+            for (; channel <= depth - 16; channel += 16) {
+              uint8x16_t acc_reg = vld1q_u8(acc + channel);
+              uint8x16_t input_reg = vld1q_u8(input_row_ptr);
+              input_row_ptr += 16;
+              acc_reg = vmaxq_u8(acc_reg, input_reg);
+              vst1q_u8(acc + channel, acc_reg);
+            }
+
+            for (; channel <= depth - 8; channel += 8) {
+              uint8x8_t acc_reg = vld1_u8(acc + channel);
+              uint8x8_t input_reg = vld1_u8(input_row_ptr);
+              input_row_ptr += 8;
+              acc_reg = vmax_u8(acc_reg, input_reg);
+              vst1_u8(acc + channel, acc_reg);
+            }
+#endif
+            for (; channel < depth; ++channel) {
+              acc[channel] = std::max(acc[channel], *input_row_ptr++);
+            }
+          }
+        }
+        uint8* output_ptr =
+            output_data + Offset(output_dims, 0, out_x, out_y, batch);
+        int channel = 0;
+#ifdef USE_NEON
+        for (; channel <= depth - 16; channel += 16) {
+          uint8x16_t a = vld1q_u8(acc + channel);
+          a = vminq_u8(a, vdupq_n_u8(output_activation_max));
+          a = vmaxq_u8(a, vdupq_n_u8(output_activation_min));
+          vst1q_u8(output_ptr + channel, a);
+        }
+        for (; channel <= depth - 8; channel += 8) {
+          uint8x8_t a = vld1_u8(acc + channel);
+          a = vmin_u8(a, vdup_n_u8(output_activation_max));
+          a = vmax_u8(a, vdup_n_u8(output_activation_min));
+          vst1_u8(output_ptr + channel, a);
+        }
+#endif
+        for (; channel < depth; ++channel) {
+          uint8 a = acc[channel];
+          a = std::max<uint8>(a, output_activation_min);
+          a = std::min<uint8>(a, output_activation_max);
+          output_ptr[channel] = static_cast<uint8>(a);
+        }
+      }
+    }
+  }
+}
+
+template <FusedActivationFunctionType Ac>
+void L2Pool(const float* input_data, const Dims<4>& input_dims,
+            int stride_width, int stride_height,
+            int pad_width, int pad_height, int filter_width, int filter_height,
+            float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("L2Pool");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int input_height = ArraySize(input_dims, 2);
+  const int input_width = ArraySize(input_dims, 1);
+  const int output_height = ArraySize(output_dims, 2);
+  const int output_width = ArraySize(output_dims, 1);
+  // Actually carry out L2 Pool. Code is written in forward mode: we go through
+  // the input values once, and write to all the pooled regions that it maps to.
+  const auto in_mat = MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  auto out_mat = MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+  Eigen::VectorXf in_square(in_mat.rows());
+  Eigen::VectorXf out_count(out_mat.cols());
+  out_count.setZero();
+  // Prefill the output to 0.
+  out_mat.setZero();
+  for (int b = 0; b < batches; ++b) {
+    for (int h = 0; h < input_height; ++h) {
+      for (int w = 0; w < input_width; ++w) {
+        // (h_start, h_end) * (w_start, w_end) is the range that the input
+        // vector projects to.
+        const int hpad = h + pad_height;
+        const int wpad = w + pad_width;
+        const int h_start =
+            (hpad < filter_height) ? 0 : (hpad - filter_height) / stride_height + 1;
+        const int h_end = std::min(hpad / stride_height + 1, output_height);
+        const int w_start =
+            (wpad < filter_width) ? 0 : (wpad - filter_width) / stride_width + 1;
+        const int w_end = std::min(wpad / stride_width + 1, output_width);
+        // pre-compute square
+        const int in_offset = w + input_width * (h + input_height * b);
+        in_square =
+            in_mat.col(in_offset).array() * in_mat.col(in_offset).array();
+        // compute elementwise sum of squares
+        for (int ph = h_start; ph < h_end; ++ph) {
+          for (int pw = w_start; pw < w_end; ++pw) {
+            const int out_offset = pw + output_width * (ph + output_height * b);
+            out_mat.col(out_offset) += in_square;
+            out_count(out_offset)++;
+          }
+        }
+      }
+    }
+  }
+
+  out_count = out_count.array().inverse();
+  out_mat =
+      (out_mat.array().rowwise() * out_count.transpose().array()).cwiseSqrt();
+}
+
+inline void LocalResponseNormalization(const float* input_data,
+                                       const Dims<4>& input_dims, int range,
+                                       float bias, float alpha, float beta,
+                                       float* output_data,
+                                       const Dims<4>& output_dims) {
+  /* const int batches = */ MatchingArraySize(input_dims, 3, output_dims, 3);
+  /* const int height = */ MatchingArraySize(input_dims, 2, output_dims, 2);
+  /* const int width = */ MatchingArraySize(input_dims, 1, output_dims, 1);
+  /* const int depth = */ MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  const auto data_in = MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  auto data_out = MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+
+  // Carry out local response normalization, vector by vector.
+  // Since the data are stored column major, making row-wise operation
+  // probably not memory efficient anyway, we do an explicit for loop over
+  // the columns.
+  const int double_range = range * 2;
+  Eigen::VectorXf padded_square(data_in.rows() + double_range);
+  padded_square.setZero();
+  for (int r = 0; r < data_in.cols(); ++r) {
+    // Do local response normalization for data_in(:, r)
+    // first, compute the square and store them in buffer for repeated use
+    padded_square.block(range, 0, data_in.rows(), 1) =
+        data_in.col(r).cwiseProduct(data_in.col(r)) * alpha;
+    // Then, compute the scale and writes them to data_out
+    float accumulated_scale = 0;
+    for (int i = 0; i < double_range; ++i) {
+      accumulated_scale += padded_square(i);
+    }
+    for (int i = 0; i < data_in.rows(); ++i) {
+      accumulated_scale += padded_square(i + double_range);
+      data_out(i, r) = bias + accumulated_scale;
+      accumulated_scale -= padded_square(i);
+    }
+  }
+
+  // In a few cases, the pow computation could benefit from speedups.
+  if (beta == 1) {
+    data_out.array() = data_in.array() * data_out.array().inverse();
+  } else if (beta == 0.5) {
+    data_out.array() = data_in.array() * data_out.array().sqrt().inverse();
+  } else {
+    data_out.array() = data_in.array() * data_out.array().pow(-beta);
+  }
+}
+
+inline void Softmax(const float* input_data, const Dims<4>& input_dims,
+                    float beta, float* output_data,
+                    const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Softmax");
+  /* const int batches = */ MatchingArraySize(input_dims, 3, output_dims, 3);
+  /* const int height = */ MatchingArraySize(input_dims, 2, output_dims, 2);
+  /* const int width = */ MatchingArraySize(input_dims, 1, output_dims, 1);
+  /* const int depth = */ MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  const auto in_mat = MapAsMatrixWithFirstDimAsRows(input_data, input_dims);
+  auto out_mat = MapAsMatrixWithFirstDimAsRows(output_data, output_dims);
+  // Compute the exponential first, removing the max coefficient for numerical
+  // stability.
+  out_mat = (in_mat.rowwise() - in_mat.colwise().maxCoeff()).array() * beta;
+  // We are separating out the exp function so that exp can be vectorized.
+  out_mat = out_mat.array().exp();
+  // Normalize to get the activations.
+  Eigen::Array<float, 1, Eigen::Dynamic> scale =
+      out_mat.array().colwise().sum().inverse();
+  out_mat.array().rowwise() *= scale;
+}
+
+inline void Softmax(const uint8* input_data, const Dims<4>& input_dims,
+                    int32 input_beta_multiplier, int32 input_beta_left_shift,
+                    int diff_min, uint8* output_data,
+                    const Dims<4>& output_dims) {
+  // The representation chosen for the input to the exp() function is Q5.26.
+  // We need to leave extra space since values that we skip might be as large as
+  // -32 before multiplying by input_beta_multiplier, and therefore as large as
+  // -16 afterwards.  Note that exp(-8) is definitely not insignificant to
+  // accumulation, but exp(-16) definitely is.
+  static const int kScaledDiffIntegerBits = 5;
+  static const int kAccumulationIntegerBits = 12;
+  using FixedPointScaledDiff =
+      gemmlowp::FixedPoint<int32, kScaledDiffIntegerBits>;
+  using FixedPointAccum = gemmlowp::FixedPoint<int32, kAccumulationIntegerBits>;
+  using FixedPoint0 = gemmlowp::FixedPoint<int32, 0>;
+
+  gemmlowp::ScopedProfilingLabel label("Softmax");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  for (int b = 0; b < batches; ++b) {
+    for (int x = 0; x < width; ++x) {
+      for (int y = 0; y < height; ++y) {
+        uint8 max_in_row = 0;
+        for (int c = 0; c < depth; ++c) {
+          max_in_row =
+              std::max(max_in_row, input_data[Offset(input_dims, c, x, y, b)]);
+        }
+
+        FixedPointAccum sum_of_exps = FixedPointAccum::Zero();
+        for (int c = 0; c < depth; ++c) {
+          int32 input_diff =
+              static_cast<int32>(input_data[Offset(input_dims, c, x, y, b)]) -
+              max_in_row;
+          if (input_diff >= diff_min) {
+            const int32 input_diff_rescaled =
+                MultiplyByQuantizedMultiplierGreaterThanOne(
+                    input_diff, input_beta_multiplier, input_beta_left_shift);
+            const FixedPointScaledDiff scaled_diff_f8 =
+                FixedPointScaledDiff::FromRaw(input_diff_rescaled);
+            sum_of_exps =
+                sum_of_exps + gemmlowp::Rescale<kAccumulationIntegerBits>(
+                                  exp_on_negative_values(scaled_diff_f8));
+          }
+        }
+
+        int32 fixed_sum_of_exps = sum_of_exps.raw();
+        // TODO: Use a NEON intrinsic like vclzq_u32 instead.
+        int headroom_plus_one =
+            __builtin_clz(static_cast<uint32>(fixed_sum_of_exps));
+        // This is the number of bits to the left of the binary point above 1.0.
+        // Consider fixed_sum_of_exps=1.25.  In that case shifted_scale=0.8 and
+        // no later adjustment will be needed.
+        int num_bits_over_unit = kAccumulationIntegerBits - headroom_plus_one;
+        int32 shifted_sum_minus_one = static_cast<int32>(
+            (static_cast<uint32>(fixed_sum_of_exps) << headroom_plus_one) -
+            (static_cast<uint32>(1) << 31));
+
+        FixedPoint0 shifted_scale = gemmlowp::one_over_one_plus_x_for_x_in_0_1(
+            FixedPoint0::FromRaw(shifted_sum_minus_one));
+
+        for (int c = 0; c < depth; ++c) {
+          int32 input_diff =
+              static_cast<int32>(input_data[Offset(input_dims, c, x, y, b)]) -
+              max_in_row;
+          if (input_diff >= diff_min) {
+            const int32 input_diff_rescaled =
+                MultiplyByQuantizedMultiplierGreaterThanOne(
+                    input_diff, input_beta_multiplier, input_beta_left_shift);
+            const FixedPointScaledDiff scaled_diff_f8 =
+                FixedPointScaledDiff::FromRaw(input_diff_rescaled);
+
+            FixedPoint0 exp_in_0 = exp_on_negative_values(scaled_diff_f8);
+            int32 unsat_output = gemmlowp::RoundingDivideByPOT(
+                (shifted_scale * exp_in_0).raw(), num_bits_over_unit + 31 - 8);
+
+            output_data[Offset(output_dims, c, x, y, b)] =
+                std::max(std::min(unsat_output, 255), 0);
+
+          } else {
+            output_data[Offset(output_dims, c, x, y, b)] = 0;
+          }
+        }
+      }
+    }
+  }
+}
+
+inline void Logistic(const float* input_data, const Dims<4>& input_dims,
+                     float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Logistic");
+  auto input_map = MapAsVector(input_data, input_dims);
+  auto output_map = MapAsVector(output_data, output_dims);
+  output_map.array() =
+      input_map.array().unaryExpr(Eigen::internal::scalar_sigmoid_op<float>());
+}
+
+inline void Logistic(const uint8* input_data, const Dims<4>& input_dims,
+                     int32 input_zero_point, int32 input_range_radius,
+                     int32 input_multiplier, int input_left_shift,
+                     uint8* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Logistic");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          const uint8 input_val_u8 = input_data[Offset(input_dims, c, x, y, b)];
+          const int32 input_val_centered =
+              static_cast<int32>(input_val_u8) - input_zero_point;
+          uint8 output_val;
+          if (input_val_centered < -input_range_radius) {
+            output_val = 0;
+          } else if (input_val_centered > input_range_radius) {
+            output_val = 255;
+          } else {
+            const int32 input_val_rescaled =
+                MultiplyByQuantizedMultiplierGreaterThanOne(
+                    input_val_centered, input_multiplier, input_left_shift);
+            using FixedPoint4 = gemmlowp::FixedPoint<int32, 4>;
+            using FixedPoint0 = gemmlowp::FixedPoint<int32, 0>;
+            const FixedPoint4 input_val_f4 =
+                FixedPoint4::FromRaw(input_val_rescaled);
+            const FixedPoint0 output_val_f0 = gemmlowp::logistic(input_val_f4);
+            using gemmlowp::RoundingDivideByPOT;
+            int32 output_val_s32 = RoundingDivideByPOT(output_val_f0.raw(), 23);
+            if (output_val_s32 == 256) {
+              output_val_s32 = 255;
+            }
+            DCHECK_GE(output_val_s32, 0);
+            DCHECK_LE(output_val_s32, 255);
+            output_val = static_cast<uint8>(output_val_s32);
+          }
+          output_data[Offset(output_dims, c, x, y, b)] = output_val;
+        }
+      }
+    }
+  }
+}
+
+inline void Tanh(const float* input_data, const Dims<4>& input_dims,
+                 float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Tanh");
+  auto input_map = MapAsVector(input_data, input_dims);
+  auto output_map = MapAsVector(output_data, output_dims);
+  output_map.array() = input_map.array().tanh();
+}
+
+inline void Dequantize(const uint8* input_data, const Dims<4>& input_dims,
+                       int32 zero_point, double scale, float* output_data,
+                       const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Dequantize");
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          int32 val = input_data[Offset(input_dims, c, x, y, b)];
+          float result = static_cast<float>(scale * (val - zero_point));
+          output_data[Offset(output_dims, c, x, y, b)] = result;
+        }
+      }
+    }
+  }
+}
+
+inline void FakeQuant(const float* input_data, const Dims<4>& input_dims,
+                      float rmin, float rmax, float* output_data,
+                      const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("FakeQuant");
+
+  // 0 should always be a representable value. Let's assume that the initial
+  // min,max range contains 0.
+  DCHECK_LE(rmin, 0.);
+  DCHECK_GE(rmax, 0.);
+
+  // Determine quantization parameters: zero_point, scale.
+  using Integer = uint8;
+  const Integer qmin = std::numeric_limits<Integer>::min();
+  const Integer qmax = std::numeric_limits<Integer>::max();
+  const float qmin_float = qmin;
+  const float qmax_float = qmax;
+  int32 zero_point = 0;
+  float scale = 0.f;
+  // If rmin==rmax, both must be zero per the above assertion,
+  // so we are done.
+  if (rmin != rmax) {
+    // First determine the scale.
+    scale = (rmax - rmin) / (qmax_float - qmin_float);
+
+    // Zero-point computation.
+    // First the initial floating-point computation. The zero-point can be
+    // determined from solving an affine equation for any known pair
+    // (real value, corresponding quantized value).
+    // We know two such pairs: (rmin, qmin) and (rmax, qmax).
+    // The arithmetic error on the zero point computed from either pair
+    // will be roughly machine_epsilon * (sum of absolute values of terms)
+    // so we want to use the variant that adds the smaller terms.
+    const float zero_point_from_min = qmin_float - rmin / scale;
+    const float zero_point_from_max = qmax_float - rmax / scale;
+    const float zero_point_from_min_error =
+        std::abs(qmin_float) + std::abs(rmin / scale);
+    const float zero_point_from_max_error =
+        std::abs(qmax_float) + std::abs(rmax / scale);
+
+    const float zero_point_float =
+        zero_point_from_min_error < zero_point_from_max_error
+            ? zero_point_from_min
+            : zero_point_from_max;
+
+    // Now we need to nudge the zero point to be an integer
+    // (our zero points are integer, and this is motivated by the requirement
+    // to be able to represent the real value "0" exactly as a quantized value,
+    // which is required in multiple places, for example in Im2col with SAME
+    // padding).
+    if (zero_point_float < qmin_float) {
+      zero_point = qmin;
+    } else if (zero_point_float > qmax_float) {
+      zero_point = qmax;
+    } else {
+      zero_point = static_cast<int32>(std::round(zero_point_float));
+    }
+    // The zero point should always be in the range of quantized value,
+    // [qmin, qmax].
+    DCHECK_GE(zero_point, qmin);
+    DCHECK_LE(zero_point, qmax);
+  }
+
+  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  const int height = MatchingArraySize(input_dims, 2, output_dims, 2);
+  const int width = MatchingArraySize(input_dims, 1, output_dims, 1);
+  const int depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < height; ++y) {
+      for (int x = 0; x < width; ++x) {
+        for (int c = 0; c < depth; ++c) {
+          const float src_val = input_data[Offset(input_dims, c, x, y, b)];
+          const float unclamped_quantized_val =
+              std::round(zero_point + src_val / scale);
+          const float quantized_val = std::min(
+              qmax_float, std::max(qmin_float, unclamped_quantized_val));
+          const float dst_val = scale * (quantized_val - zero_point);
+          output_data[Offset(output_dims, c, x, y, b)] = dst_val;
+        }
+      }
+    }
+  }
+}
+
+template <typename SrcT, typename DstT>
+inline void Cast(const SrcT* input_data, const Dims<4>& input_dims,
+                 DstT* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Cast");
+  auto input_map = MapAsVector(input_data, input_dims);
+  auto output_map = MapAsVector(output_data, output_dims);
+  output_map.array() = input_map.array().template cast<DstT>();
+}
+
+inline void Floor(const float* input_data, const Dims<4>& input_dims,
+                  float* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Floor");
+  auto input_map = MapAsVector(input_data, input_dims);
+  auto output_map = MapAsVector(output_data, output_dims);
+  output_map.array() = Eigen::floor(input_map.array());
+}
+
+template <typename T>
+inline void Gather(const T* input_data, const Dims<4>& input_dims,
+                   const int32* coords_data, const Dims<4>& coords_dims,
+                   T* output_data, const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("Gather");
+  DCHECK_EQ(RequiredBufferSizeForDims(output_dims),
+            RequiredBufferSizeForDims(coords_dims));
+  for (int i = 0; i < RequiredBufferSizeForDims(coords_dims); i++) {
+    DCHECK_GE(coords_data[i], 0);
+    DCHECK_LT(coords_data[i], RequiredBufferSizeForDims(input_dims));
+    output_data[i] = input_data[coords_data[i]];
+  }
+}
+
+inline void ResizeBilinear(const float* input_data, const Dims<4>& input_dims,
+                           const int32* output_size_data,
+                           const Dims<4>& output_size_dims, float* output_data,
+                           const Dims<4>& output_dims) {
+  gemmlowp::ScopedProfilingLabel label("ResizeBilinear");
+  int32 batches = MatchingArraySize(input_dims, 3, output_dims, 3);
+  int32 input_height = ArraySize(input_dims, 2);
+  int32 input_width = ArraySize(input_dims, 1);
+  int32 depth = MatchingArraySize(input_dims, 0, output_dims, 0);
+
+  DCHECK_EQ(ArraySize(output_size_dims, 3), 1);
+  DCHECK_EQ(ArraySize(output_size_dims, 2), 1);
+  DCHECK_EQ(ArraySize(output_size_dims, 1), 1);
+  DCHECK_EQ(ArraySize(output_size_dims, 0), 2);
+  int32 output_height = output_size_data[Offset(output_size_dims, 0, 0, 0, 0)];
+  int32 output_width = output_size_data[Offset(output_size_dims, 1, 0, 0, 0)];
+  float height_scale = static_cast<float>(input_height) / output_height;
+  float width_scale = static_cast<float>(input_width) / output_width;
+
+  for (int b = 0; b < batches; ++b) {
+    for (int y = 0; y < output_height; ++y) {
+      float input_y = y * height_scale;
+      int32 y0 = static_cast<int32>(input_y);
+      int32 y1 = std::min(y0 + 1, input_height - 1);
+      for (int x = 0; x < output_width; ++x) {
+        float input_x = x * width_scale;
+        int32 x0 = static_cast<int32>(input_x);
+        int32 x1 = std::min(x0 + 1, input_width - 1);
+        for (int c = 0; c < depth; ++c) {
+          float interpolation = input_data[Offset(input_dims, c, x0, y0, b)] *
+                                    (1 - (input_y - y0)) *
+                                    (1 - (input_x - x0)) +
+                                input_data[Offset(input_dims, c, x0, y1, b)] *
+                                    (input_y - y0) * (1 - (input_x - x0)) +
+                                input_data[Offset(input_dims, c, x1, y0, b)] *
+                                    (1 - (input_y - y0)) * (input_x - x0) +
+                                input_data[Offset(input_dims, c, x1, y1, b)] *
+                                    (input_y - y0) * (input_x - x0);
+          output_data[Offset(output_dims, c, x, y, b)] = interpolation;
+        }
+      }
+    }
+  }
+}
+
+}  // namespace optimized_ops
+}  // namespace nn
+}  // namespace android
+
+#if defined OPTIMIZED_OPS_H__IGNORE_DEPRECATED_DECLARATIONS
+#undef OPTIMIZED_OPS_H__IGNORE_DEPRECATED_DECLARATIONS
+#pragma GCC diagnostic pop
+#endif
+
+#endif  // ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_OPS_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/optimized/tensor_utils_impl.h b/src/runtime/ref/nn/common/operations/internal/optimized/tensor_utils_impl.h
new file mode 100644
index 0000000..249a672
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/optimized/tensor_utils_impl.h
@@ -0,0 +1,132 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_TENSOR_UTILS_IMPL_H_
+#define FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_TENSOR_UTILS_IMPL_H_
+
+#include "ActivationFunctor.h"
+
+#ifndef USE_NEON
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define USE_NEON
+#endif  //  defined(__ARM_NEON__) || defined(__ARM_NEON)
+#endif  //  USE_NEON
+
+namespace android {
+namespace nn {
+namespace tensor_utils {
+
+// Multiply a matrix by a batch vector, and store results in a batch-size
+// vector.
+void PortableMatrixBatchVectorMultiplyAccumulate(const float* matrix,
+                                                 int m_rows, int m_cols,
+                                                 const float* vector,
+                                                 int n_batch, float* result,
+                                                 int result_stride);
+void NeonMatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows,
+                                             int m_cols, const float* vector,
+                                             int n_batch, float* result,
+                                             int result_stride);
+
+// Cwise product of two vectors.
+void PortableVectorVectorCwiseProduct(const float* vector1,
+                                      const float* vector2, int v_size,
+                                      float* result);
+void NeonVectorVectorCwiseProduct(const float* vector1, const float* vector2,
+                                  int v_size, float* result);
+
+// Cwise product and accumulate of two vectors. Since it's a MAC operation, the
+// assumption here is that result array is initialized to valid values.
+void PortableVectorVectorCwiseProductAccumulate(const float* vector1,
+                                                const float* vector2,
+                                                int v_size, float* result);
+void NeonVectorVectorCwiseProductAccumulate(const float* vector1,
+                                            const float* vector2, int v_size,
+                                            float* result);
+
+// Dot product of two vectors.
+float PortableVectorVectorDotProduct(const float* vector1, const float* vector2,
+                                     int v_size);
+
+// Dot product of two batch vectors.
+void PortableBatchVectorBatchVectorDotProduct(const float* vector1,
+                                              const float* vector2, int v_size,
+                                              int n_batch, float* result,
+                                              int result_stride);
+
+// Cwise product and accumulate of a vector and a batch-vector. Since it's a MAC
+// operation, the assumption here is that result array is initialized to valid
+// values.
+void PortableVectorBatchVectorCwiseProductAccumulate(const float* vector,
+                                                     int v_size,
+                                                     const float* batch_vector,
+                                                     int n_batch,
+                                                     float* result);
+void NeonVectorBatchVectorCwiseProductAccumulate(const float* vector,
+                                                 int v_size,
+                                                 const float* batch_vector,
+                                                 int n_batch, float* result);
+
+// Compute "1.0f - elements of vector" (used in CIFG).
+void PortableSub1Vector(const float* vector, int v_size, float* result);
+void NeonSub1Vector(const float* vector, int v_size, float* result);
+
+// Clip elements of a vector using a abs_limit value.
+void PortableClipVector(const float* vector, int v_size, float abs_limit,
+                        float* result);
+void NeonClipVector(const float* vector, int v_size, float abs_limit,
+                    float* result);
+
+// Batch vector initialization with another vector.
+void PortableVectorBatchVectorAssign(const float* vector, int v_size,
+                                     int n_batch, float* batch_vector);
+
+// Apply sigmoid to elements of a vector.
+void PortableApplySigmoidToVector(const float* vector, int v_size,
+                                  float* result);
+
+// Apply activation function to elements of a vector.
+void PortableApplyActivationToVector(const float* vector, int v_size,
+                                     ActivationFn activation,
+                                     float* result);
+
+// Copy vector to another vector.
+void PortableCopyVector(const float* vector, int v_size, float* result);
+
+// Fill vector with 0.f.
+void PortableZeroVector(float* vector, int v_size);
+
+// Limit a float input f between +abs_limit and -abs_limit.
+float PortableClip(float f, float abs_limit);
+
+// Shift left a vector in place with v_size size.
+void PortableVectorShiftLeft(float* vector, int v_size, float shift_value);
+
+// Reduce-sum on a float input vector:
+// input_vector: float pointer to input vector.
+// input_stride: input vector stride.
+// output_vector: float pointer to vector.
+// output_size: output vector size.
+// reduction_size: number of consecutive elements from input vector which are
+// added to get one element of output.
+void PortableReductionSumVector(const float* input_vector, int input_stride,
+                                float* output_vector, int output_size,
+                                int reduction_size);
+}  // namespace tensor_utils
+}  // namespace nn
+}  // namespace android
+
+#endif  //  FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_TENSOR_UTILS_IMPL_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/tensor_utils.cc b/src/runtime/ref/nn/common/operations/internal/tensor_utils.cc
new file mode 100644
index 0000000..78275bb
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/tensor_utils.cc
@@ -0,0 +1,29 @@
+/*
+ * Copyright 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "tensor_utils.h"
+
+#ifndef USE_NEON
+#if defined(__ARM_NEON__) || defined(__ARM_NEON)
+#define USE_NEON
+#endif  //  defined(__ARM_NEON__) || defined(__ARM_NEON)
+#endif  //  USE_NEON
+
+#ifdef USE_NEON
+#include "optimized/neon_tensor_utils.h"
+#else
+#include "reference/portable_tensor_utils.h"
+#endif  // USE_NEON
diff --git a/src/runtime/ref/nn/common/operations/internal/tensor_utils.h b/src/runtime/ref/nn/common/operations/internal/tensor_utils.h
new file mode 100644
index 0000000..56167d2
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/tensor_utils.h
@@ -0,0 +1,122 @@
+/*
+ * Copyright 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_TENSOR_UTILS_H_
+#define FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_TENSOR_UTILS_H_
+
+#include "ActivationFunctor.h"
+
+namespace android {
+namespace nn {
+namespace tensor_utils {
+
+// Limit a float input f betweeen +abs_limit and -abs_limit.
+float Clip(float f, float abs_limit);
+
+// Multiply a matrix by a batch vector, and store results in a batch-size
+// vector using a stride value provided in result_stride. 'result_stride' shows
+// how the number of elements between consecutive result values. For example
+// result_stride = 1, will cause the output to look like this:
+// [O_1, 0_2, ... O_rows] in memory, but result_stride = 3, will cause it to be
+// arranged like this in memory: [O_1, x, x, 0_2, x, x, ..., O_rows]
+void MatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows,
+                                         int m_cols, const float* vector,
+                                         int n_batch, float* result,
+                                         int result_stride);
+
+// Cwise product of two vectors.
+void VectorVectorCwiseProduct(const float* vector1, const float* vector2,
+                              int v_size, float* result);
+
+// Cwise product and accumulate of two vectors. Since it's a MAC opertation, the
+// assumption here is that result array is initialized to valid values.
+void VectorVectorCwiseProductAccumulate(const float* vector1,
+                                        const float* vector2, int v_size,
+                                        float* result);
+
+// Dot product of two vectors.
+float VectorVectorDotProduct(const float* vector1, const float* vector2,
+                             int v_size);
+
+// Dot product of two batch vectors of size n_batch * v_size:
+// vector1 = [x_1_1, x_1_2, ..., x_1_vsize,
+//            x_2_1, x_2_2, ..., x_2_vsize,
+//            ...
+//            x_nbatch_1,..., x_nbatch_vsize]
+// vector2 = [y_1_1, y_1_2, ..., y_1_vsize,
+//            y_2_1, y_2_2, ..., y_2_vsize,
+//            ...
+//            y_nbatch_1,..., y_nbatch_vsize]
+// Then result will be a vector of n_batch size which will be saved with a
+// stride of result_stride in memory starting from 'result':
+// [x_1_1 * y_1_1 + x_1_2 * y_1_2 + ... + x_1_vsize * y_1_vsize,
+//  x_2_1 * y_2_1 + x_2_2 * y_2_2 + ... + x_2_vsize * y_2_vsize,
+//  ...
+//  x_nbatch_1 * y_nbatch_1 + ... + x_nbatch_vsize * y_nbatch_vsize]
+void BatchVectorBatchVectorDotProduct(const float* vector1,
+                                      const float* vector2, int v_size,
+                                      int n_batch, float* result,
+                                      int result_stride);
+
+// Cwise product and accumulate of a vector and a batch-vector. Since it's a MAC
+// operation, the assumption here is that result array is initialized to valid
+// values.
+void VectorBatchVectorCwiseProductAccumulate(const float* vector, int v_size,
+                                             const float* batch_vector,
+                                             int n_batch, float* result);
+
+// Batch vector initialization with another vector.
+void VectorBatchVectorAssign(const float* vector, int v_size, int n_batch,
+                             float* batch_vector);
+
+// Apply sigmoid to elements of a vector.
+void ApplySigmoidToVector(const float* vector, int v_size, float* result);
+
+// Apply activation function to elements of a vector.
+void ApplyActivationToVector(const float* vector, int v_size,
+                             ActivationFn activation, float* result);
+
+// Copy vector to another vector.
+void CopyVector(const float* vector, int v_size, float* result);
+
+// Compute "1.0f - elements of vector" (used in CIFG).
+void Sub1Vector(const float* vector, int v_size, float* result);
+
+// Fill vector with 0.f.
+void ZeroVector(float* vector, int v_size);
+
+// Clip elements of a vector using a abs_limit value.
+void ClipVector(const float* vector, int v_size, float abs_limit,
+                float* result);
+
+// Shift left a vector in place with v_size size.
+void VectorShiftLeft(float* vector, int v_size, float shift_value);
+
+// Reduce-sum on a float input vector:
+// input_vector: float pointer to input vector.
+// input_stride: input vector stride.
+// output_vector: float pointer to vector.
+// output_size: output vector size.
+// reduction_size: number of consecutive elements from input vector which are
+// added to get one element of output.
+void ReductionSumVector(const float* input_vector, int input_stride,
+                        float* output_vector, int output_size,
+                        int reduction_size);
+}  // namespace tensor_utils
+}  // namespace nn
+}  // namespace android
+
+#endif  // FRAMEWORKS_ML_NN_COMMON_OPERATIONS_INTERNAL_TENSOR_UTILS_H_
diff --git a/src/runtime/ref/nn/common/operations/internal/tensor_utils_test.cc b/src/runtime/ref/nn/common/operations/internal/tensor_utils_test.cc
new file mode 100644
index 0000000..ad5204d
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/tensor_utils_test.cc
@@ -0,0 +1,197 @@
+/*
+ * Copyright 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "gmock/gmock-matchers.h"
+#include "gtest/gtest.h"
+#include "tensor_utils.h"
+
+namespace android {
+namespace nn {
+namespace tensor_utils {
+
+namespace {
+
+using ::testing::FloatNear;
+using ::testing::Matcher;
+
+std::vector<Matcher<float>> ArrayFloatNear(const std::vector<float>& values,
+                                           float max_abs_error=1.e-6) {
+  std::vector<Matcher<float>> matchers;
+  matchers.reserve(values.size());
+  for (const float& v : values) {
+    matchers.emplace_back(FloatNear(v, max_abs_error));
+  }
+  return matchers;
+}
+
+}  // anonymous namespace
+
+TEST(uKernels, ClipTest) {
+  constexpr int kVectorSize = 10;
+  constexpr float kAbsLimit = 2.0;
+  static float input[kVectorSize] = {0.0,  -0.5, 1.0,  -1.5, 2.0,
+                                     -2.5, 3.0,  -3.5, 4.0,  -4.5};
+  std::vector<float> output(kVectorSize);
+  ClipVector(input, kVectorSize, kAbsLimit, output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear(
+                  {0.0, -0.5, 1.0, -1.5, 2.0, -2.0, 2.0, -2.0, 2.0, -2.0})));
+}
+
+TEST(uKernels, MatrixBatchVectorMultiplyAccumulateTest) {
+  constexpr int kRow = 3;
+  constexpr int kCol = 4;
+  constexpr int kBatch = 2;
+  static float matrix[kRow * kCol] = {1.0,  2.0,  3.0,  4.0,   //
+                                      -1.0, -2.0, -3.0, -4.0,  //
+                                      1.0,  -2.0, 3.0,  -4.0};
+  static float vector[kCol * kBatch] = {1.0, -1.0, 1.0, -1.0,  //
+                                        2.0, -2.0, 2.0, -2.0};
+  std::vector<float> output(kRow * kBatch);
+  std::fill(output.begin(), output.end(), 3.0);
+  MatrixBatchVectorMultiplyAccumulate(matrix, kRow, kCol, vector, kBatch,
+                                      output.data(), /*result_stride=*/1);
+  EXPECT_THAT(output, ElementsAreArray(ArrayFloatNear({1., 5., 13.,  //
+                                                       -1., 7., 23.})));
+}
+
+TEST(uKernels, VectorVectorCwiseProductTest) {
+  constexpr int kVectorSize = 10;
+  static float input1[kVectorSize] = {0.0,  -0.5, 1.0,  -1.5, 2.0,
+                                      -2.5, 3.0,  -3.5, 4.0,  -4.5};
+  static float input2[kVectorSize] = {0.1,  -0.1, 0.1,  -0.1, 0.1,
+                                      -0.1, 0.1,  -0.1, 0.1,  -0.1};
+  std::vector<float> output(kVectorSize);
+  VectorVectorCwiseProduct(input1, input2, kVectorSize, output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear(
+                  {0.0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45})));
+}
+
+TEST(uKernels, VectorVectorCwiseProductAccumulateTest) {
+  constexpr int kVectorSize = 10;
+  static float input1[kVectorSize] = {0.0,  -0.5, 1.0,  -1.5, 2.0,
+                                      -2.5, 3.0,  -3.5, 4.0,  -4.5};
+  static float input2[kVectorSize] = {0.1,  -0.1, 0.1,  -0.1, 0.1,
+                                      -0.1, 0.1,  -0.1, 0.1,  -0.1};
+  std::vector<float> output(kVectorSize);
+  std::fill(output.begin(), output.end(), 1.0);
+  VectorVectorCwiseProductAccumulate(input1, input2, kVectorSize,
+                                     output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear(
+                  {1.0, 1.05, 1.1, 1.15, 1.2, 1.25, 1.3, 1.35, 1.4, 1.45})));
+}
+
+TEST(uKernels, VectorBatchVectorAssignTest) {
+  constexpr int kVectorSize = 5;
+  constexpr int kBatchSize = 3;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> output(kVectorSize * kBatchSize);
+  VectorBatchVectorAssign(input, kVectorSize, kBatchSize, output.data());
+  EXPECT_THAT(output, ElementsAreArray(ArrayFloatNear(
+                          {0.0, -0.5, 1.0, -1.5, 2.0, 0.0, -0.5, 1.0, -1.5, 2.0,
+                           0.0, -0.5, 1.0, -1.5, 2.0})));
+}
+
+TEST(uKernels, ApplySigmoidToVectorTest) {
+  constexpr int kVectorSize = 5;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> output(kVectorSize);
+  ApplySigmoidToVector(input, kVectorSize, output.data());
+  EXPECT_THAT(output, ElementsAreArray(ArrayFloatNear(
+                          {0.5, 0.377541, 0.731059, 0.182426, 0.880797})));
+}
+
+TEST(uKernels, ApplyActivationToVectorTest) {
+  constexpr int kVectorSize = 5;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> output(kVectorSize);
+  ApplyActivationToVector(input, kVectorSize, kActivationRelu, output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear({0.0, 0.0, 1.0, 0.0, 2.0})));
+
+  ApplyActivationToVector(input, kVectorSize, kActivationTanh, output.data());
+  EXPECT_THAT(output, ElementsAreArray(ArrayFloatNear(
+                          {0.0, -0.462117, 0.761594, -0.905148, 0.964028})));
+}
+
+TEST(uKernels, CopyVectorTest) {
+  constexpr int kVectorSize = 5;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> output(kVectorSize);
+  CopyVector(input, kVectorSize, output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear({0.0, -0.5, 1.0, -1.5, 2.0})));
+}
+
+TEST(uKernels, Sub1VectorTest) {
+  constexpr int kVectorSize = 5;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> output(kVectorSize);
+  Sub1Vector(input, kVectorSize, output.data());
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear({1.0, 1.5, 0.0, 2.5, -1.0})));
+}
+
+TEST(uKernels, ZeroVectorTest) {
+  constexpr int kVectorSize = 5;
+  std::vector<float> output(kVectorSize);
+  ZeroVector(output.data(), kVectorSize);
+  EXPECT_THAT(output,
+              ElementsAreArray(ArrayFloatNear({0.0, 0.0, 0.0, 0.0, 0.0})));
+}
+
+TEST(uKernels, BatchVectorBatchVectorDotProductTest) {
+  constexpr int kVectorSize = 5;
+  constexpr int kBatch = 2;
+  static float input1[kVectorSize * kBatch] = {0.0,  -0.5, 1.0,  -1.5, 2.0,
+                                               -2.5, 3.0,  -3.5, 4.0,  -4.5};
+  static float input2[kVectorSize * kBatch] = {0.1,  -0.1, 0.1,  -0.1, 0.1,
+                                               -0.1, 0.1,  -0.1, 0.1,  -0.1};
+  std::vector<float> output(kBatch);
+  BatchVectorBatchVectorDotProduct(input1, input2, kVectorSize, kBatch,
+                                   output.data(), /*result_stride=*/1);
+  EXPECT_THAT(output, ElementsAreArray(ArrayFloatNear({0.5, 1.75})));
+}
+
+TEST(uKernels, VectorShiftLeftTest) {
+  constexpr int kVectorSize = 5;
+  static float input[kVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0};
+  std::vector<float> result(kVectorSize);
+  VectorShiftLeft(input, kVectorSize, 3.0);
+  result.assign(input, input + kVectorSize);
+  EXPECT_THAT(result,
+              ElementsAreArray(ArrayFloatNear({-0.5, 1.0, -1.5, 2.0, 3.0})));
+}
+
+TEST(uKernels, ReductionSumVectorTest) {
+  constexpr int kInputVectorSize = 10;
+  constexpr int kOutputVectorSize = 5;
+  constexpr int kReductionSize = 2;
+  static float input[kInputVectorSize] = {0.0, -0.5, 1.0, -1.5, 2.0,
+                                          0.0, -0.5, 1.0, 1.0,  2.0};
+  std::vector<float> result(kOutputVectorSize);
+  ReductionSumVector(input,
+                     /*input_stride=*/1, result.data(), kOutputVectorSize,
+                     kReductionSize);
+  EXPECT_THAT(result,
+              ElementsAreArray(ArrayFloatNear({-0.5, -0.5, 2.0, 0.5, 3.0})));
+}
+
+}  // namespace tensor_utils
+}  // namespace nn
+}  // namespace android
diff --git a/src/runtime/ref/nn/common/operations/internal/types.h b/src/runtime/ref/nn/common/operations/internal/types.h
new file mode 100644
index 0000000..2000f73
--- /dev/null
+++ b/src/runtime/ref/nn/common/operations/internal/types.h
@@ -0,0 +1,111 @@
+/*
+ * Copyright (C) 2017 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_TYPES_H_
+#define ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_TYPES_H_
+
+#include "compatibility.h"
+
+namespace android {
+namespace nn {
+
+enum class FusedActivationFunctionType { kNone, kRelu6, kRelu1, kRelu };
+
+template <int N>
+struct Dims {
+  int sizes[N];
+  int strides[N];
+};
+
+struct Shape;
+
+inline Dims<4> convertShapeToDims(const Shape& shape) {
+  Dims<4> dims;
+  for (int i=0; i<4; i++) {
+    dims.sizes[i] = 1;
+  }
+
+  if (shape.dimensions.size() == 1) {
+    dims.sizes[0] = (int)getSizeOfDimension(shape, 0);
+  } else {
+    for (int i=0; i<4; i++) {
+      int src = (int)shape.dimensions.size()-i-1;
+      if (src >= 0) {
+        dims.sizes[i] = (int)getSizeOfDimension(shape, src);
+      }
+    }
+  }
+
+  dims.strides[0] = 1;
+  for (int i = 1; i<4; i++) {
+    dims.strides[i] = dims.strides[i-1] * dims.sizes[i-1];
+  }
+  return dims;
+}
+
+inline int Offset(const Dims<4>& dims, int i0, int i1, int i2, int i3) {
+  DCHECK(i0 >= 0 && i0 < dims.sizes[0]);
+  DCHECK(i1 >= 0 && i1 < dims.sizes[1]);
+  DCHECK(i2 >= 0 && i2 < dims.sizes[2]);
+  DCHECK(i3 >= 0 && i3 < dims.sizes[3]);
+  return i0 * dims.strides[0] + i1 * dims.strides[1] + i2 * dims.strides[2] +
+         i3 * dims.strides[3];
+}
+
+// Get array size, DCHECKing that the dim index is in range.
+template <int N>
+int ArraySize(const Dims<N>& array, int index) {
+  DCHECK(index >= 0 && index < N);
+  return array.sizes[index];
+}
+
+// Get common array size, DCHECKing that they all agree.
+template <typename ArrayType1, typename ArrayType2>
+int MatchingArraySize(const ArrayType1& array1, int index1,
+                      const ArrayType2& array2, int index2) {
+  DCHECK_EQ(ArraySize(array1, index1), ArraySize(array2, index2));
+  return ArraySize(array1, index1);
+}
+
+template <typename ArrayType1, typename ArrayType2, typename... Args>
+int MatchingArraySize(const ArrayType1& array1, int index1,
+                      const ArrayType2& array2, int index2, Args... args) {
+  DCHECK_EQ(ArraySize(array1, index1), ArraySize(array2, index2));
+  return MatchingArraySize(array1, index1, args...);
+}
+
+inline int RequiredBufferSizeForDims(const Dims<4>& dims) {
+  int max_offset = 0;
+  for (int i = 0; i < 4; i++) {
+    max_offset += (dims.sizes[i] - 1) * dims.strides[i];
+  }
+  return max_offset + 1;
+}
+
+template <int N>
+bool IsPackedWithoutStrides(const Dims<N>& dims) {
+  int expected_stride = 1;
+  for (int d = 0; d < N; d++) {
+    if (dims.strides[d] != expected_stride) return false;
+    expected_stride *= dims.sizes[d];
+  }
+  return true;
+}
+
+}  // namespace nn
+}  // namespace android
+
+#endif  // ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_TYPES_H_
diff --git a/src/runtime/ref/nn/depend/CMakeLists.txt b/src/runtime/ref/nn/depend/CMakeLists.txt
index 9ac4286..04debbc 100644
--- a/src/runtime/ref/nn/depend/CMakeLists.txt
+++ b/src/runtime/ref/nn/depend/CMakeLists.txt
@@ -1,5 +1,5 @@
 ADD_SUBDIRECTORY(android-base)
-#ADD_SUBDIRECTORY(external)
+ADD_SUBDIRECTORY(external)
 ADD_SUBDIRECTORY(hal)
 #ADD_SUBDIRECTORY(libc)
 ADD_SUBDIRECTORY(libcutils)
diff --git a/src/runtime/ref/nn/depend/external/CMakeLists.txt b/src/runtime/ref/nn/depend/external/CMakeLists.txt
new file mode 100644
index 0000000..3b679b0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/CMakeLists.txt
@@ -0,0 +1,13 @@
+ADD_SUBDIRECTORY(eigen)
+ADD_SUBDIRECTORY(gemmlowp)
+
+SET(INC_DIRS
+  ${INC_DIRS}
+  PARENT_SCOPE
+)
+
+SET(SRCS
+  ${SRCS}
+  PARENT_SCOPE
+)
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/CMakeLists.txt b/src/runtime/ref/nn/depend/external/eigen/CMakeLists.txt
new file mode 100644
index 0000000..b76ac5a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/CMakeLists.txt
@@ -0,0 +1,10 @@
+
+SET(CUR_INCS
+   ${CMAKE_CURRENT_SOURCE_DIR}
+)
+
+SET(INC_DIRS
+  ${INC_DIRS}
+  ${CUR_INCS}
+  PARENT_SCOPE
+)
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/CMakeLists.txt b/src/runtime/ref/nn/depend/external/eigen/Eigen/CMakeLists.txt
new file mode 100644
index 0000000..9eb502b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/CMakeLists.txt
@@ -0,0 +1,19 @@
+include(RegexUtils)
+test_escape_string_as_regex()
+
+file(GLOB Eigen_directory_files "*")
+
+escape_string_as_regex(ESCAPED_CMAKE_CURRENT_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}")
+
+foreach(f ${Eigen_directory_files})
+  if(NOT f MATCHES "\\.txt" AND NOT f MATCHES "${ESCAPED_CMAKE_CURRENT_SOURCE_DIR}/[.].+" AND NOT f MATCHES "${ESCAPED_CMAKE_CURRENT_SOURCE_DIR}/src")
+    list(APPEND Eigen_directory_files_to_install ${f})
+  endif()
+endforeach(f ${Eigen_directory_files})
+
+install(FILES
+  ${Eigen_directory_files_to_install}
+  DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel
+  )
+
+install(DIRECTORY src DESTINATION ${INCLUDE_INSTALL_DIR}/Eigen COMPONENT Devel FILES_MATCHING PATTERN "*.h")
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Cholesky b/src/runtime/ref/nn/depend/external/eigen/Eigen/Cholesky
new file mode 100644
index 0000000..369d1f5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Cholesky
@@ -0,0 +1,41 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CHOLESKY_MODULE_H
+#define EIGEN_CHOLESKY_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup Cholesky_Module Cholesky module
+  *
+  *
+  *
+  * This module provides two variants of the Cholesky decomposition for selfadjoint (hermitian) matrices.
+  * Those decompositions are also accessible via the following methods:
+  *  - MatrixBase::llt()
+  *  - MatrixBase::ldlt()
+  *  - SelfAdjointView::llt()
+  *  - SelfAdjointView::ldlt()
+  *
+  * \code
+  * #include <Eigen/Cholesky>
+  * \endcode
+  */
+
+#include "src/Cholesky/LLT.h"
+#include "src/Cholesky/LDLT.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/misc/lapacke.h"
+#include "src/Cholesky/LLT_LAPACKE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CHOLESKY_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/CholmodSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/CholmodSupport
new file mode 100644
index 0000000..bed8924
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/CholmodSupport
@@ -0,0 +1,48 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CHOLMODSUPPORT_MODULE_H
+#define EIGEN_CHOLMODSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+  #include <cholmod.h>
+}
+
+/** \ingroup Support_modules
+  * \defgroup CholmodSupport_Module CholmodSupport module
+  *
+  * This module provides an interface to the Cholmod library which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
+  * It provides the two following main factorization classes:
+  * - class CholmodSupernodalLLT: a supernodal LLT Cholesky factorization.
+  * - class CholmodDecomposiiton: a general L(D)LT Cholesky factorization with automatic or explicit runtime selection of the underlying factorization method (supernodal or simplicial).
+  *
+  * For the sake of completeness, this module also propose the two following classes:
+  * - class CholmodSimplicialLLT
+  * - class CholmodSimplicialLDLT
+  * Note that these classes does not bring any particular advantage compared to the built-in
+  * SimplicialLLT and SimplicialLDLT factorization classes.
+  *
+  * \code
+  * #include <Eigen/CholmodSupport>
+  * \endcode
+  *
+  * In order to use this module, the cholmod headers must be accessible from the include paths, and your binary must be linked to the cholmod library and its dependencies.
+  * The dependencies depend on how cholmod has been compiled.
+  * For a cmake based project, you can use our FindCholmod.cmake module to help you in this task.
+  *
+  */
+
+#include "src/CholmodSupport/CholmodSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CHOLMODSUPPORT_MODULE_H
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Core b/src/runtime/ref/nn/depend/external/eigen/Eigen/Core
new file mode 100644
index 0000000..0f7fa63
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Core
@@ -0,0 +1,516 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2007-2011 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CORE_H
+#define EIGEN_CORE_H
+
+// first thing Eigen does: stop the compiler from committing suicide
+#include "src/Core/util/DisableStupidWarnings.h"
+
+// Handle NVCC/CUDA/SYCL
+#if defined(__CUDACC__) || defined(__SYCL_DEVICE_ONLY__)
+  // Do not try asserts on CUDA and SYCL!
+  #ifndef EIGEN_NO_DEBUG
+  #define EIGEN_NO_DEBUG
+  #endif
+
+  #ifdef EIGEN_INTERNAL_DEBUGGING
+  #undef EIGEN_INTERNAL_DEBUGGING
+  #endif
+
+  #ifdef EIGEN_EXCEPTIONS
+  #undef EIGEN_EXCEPTIONS
+  #endif
+
+  // All functions callable from CUDA code must be qualified with __device__
+  #ifdef __CUDACC__
+    // Do not try to vectorize on CUDA and SYCL!
+    #ifndef EIGEN_DONT_VECTORIZE
+    #define EIGEN_DONT_VECTORIZE
+    #endif
+
+    #define EIGEN_DEVICE_FUNC __host__ __device__
+    // We need math_functions.hpp to ensure that that EIGEN_USING_STD_MATH macro
+    // works properly on the device side
+    #include <math_functions.hpp>
+  #else
+    #define EIGEN_DEVICE_FUNC
+  #endif
+
+#else
+  #define EIGEN_DEVICE_FUNC
+
+#endif
+
+// When compiling CUDA device code with NVCC, pull in math functions from the
+// global namespace.  In host mode, and when device doee with clang, use the
+// std versions.
+#if defined(__CUDA_ARCH__) && defined(__NVCC__)
+  #define EIGEN_USING_STD_MATH(FUNC) using ::FUNC;
+#else
+  #define EIGEN_USING_STD_MATH(FUNC) using std::FUNC;
+#endif
+
+#if (defined(_CPPUNWIND) || defined(__EXCEPTIONS)) && !defined(__CUDA_ARCH__) && !defined(EIGEN_EXCEPTIONS) && !defined(EIGEN_USE_SYCL)
+  #define EIGEN_EXCEPTIONS
+#endif
+
+#ifdef EIGEN_EXCEPTIONS
+  #include <new>
+#endif
+
+// then include this file where all our macros are defined. It's really important to do it first because
+// it's where we do all the alignment settings (platform detection and honoring the user's will if he
+// defined e.g. EIGEN_DONT_ALIGN) so it needs to be done before we do anything with vectorization.
+#include "src/Core/util/Macros.h"
+
+// Disable the ipa-cp-clone optimization flag with MinGW 6.x or newer (enabled by default with -O3)
+// See http://eigen.tuxfamily.org/bz/show_bug.cgi?id=556 for details.
+#if EIGEN_COMP_MINGW && EIGEN_GNUC_AT_LEAST(4,6)
+  #pragma GCC optimize ("-fno-ipa-cp-clone")
+#endif
+
+#include <complex>
+
+// this include file manages BLAS and MKL related macros
+// and inclusion of their respective header files
+#include "src/Core/util/MKL_support.h"
+
+// if alignment is disabled, then disable vectorization. Note: EIGEN_MAX_ALIGN_BYTES is the proper check, it takes into
+// account both the user's will (EIGEN_MAX_ALIGN_BYTES,EIGEN_DONT_ALIGN) and our own platform checks
+#if EIGEN_MAX_ALIGN_BYTES==0
+  #ifndef EIGEN_DONT_VECTORIZE
+    #define EIGEN_DONT_VECTORIZE
+  #endif
+#endif
+
+#if EIGEN_COMP_MSVC
+  #include <malloc.h> // for _aligned_malloc -- need it regardless of whether vectorization is enabled
+  #if (EIGEN_COMP_MSVC >= 1500) // 2008 or later
+    // Remember that usage of defined() in a #define is undefined by the standard.
+    // a user reported that in 64-bit mode, MSVC doesn't care to define _M_IX86_FP.
+    #if (defined(_M_IX86_FP) && (_M_IX86_FP >= 2)) || EIGEN_ARCH_x86_64
+      #define EIGEN_SSE2_ON_MSVC_2008_OR_LATER
+    #endif
+  #endif
+#else
+  // Remember that usage of defined() in a #define is undefined by the standard
+  #if (defined __SSE2__) && ( (!EIGEN_COMP_GNUC) || EIGEN_COMP_ICC || EIGEN_GNUC_AT_LEAST(4,2) )
+    #define EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC
+  #endif
+#endif
+
+#ifndef EIGEN_DONT_VECTORIZE
+
+  #if defined (EIGEN_SSE2_ON_NON_MSVC_BUT_NOT_OLD_GCC) || defined(EIGEN_SSE2_ON_MSVC_2008_OR_LATER)
+
+    // Defines symbols for compile-time detection of which instructions are
+    // used.
+    // EIGEN_VECTORIZE_YY is defined if and only if the instruction set YY is used
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_SSE
+    #define EIGEN_VECTORIZE_SSE2
+
+    // Detect sse3/ssse3/sse4:
+    // gcc and icc defines __SSE3__, ...
+    // there is no way to know about this on msvc. You can define EIGEN_VECTORIZE_SSE* if you
+    // want to force the use of those instructions with msvc.
+    #ifdef __SSE3__
+      #define EIGEN_VECTORIZE_SSE3
+    #endif
+    #ifdef __SSSE3__
+      #define EIGEN_VECTORIZE_SSSE3
+    #endif
+    #ifdef __SSE4_1__
+      #define EIGEN_VECTORIZE_SSE4_1
+    #endif
+    #ifdef __SSE4_2__
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __AVX__
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_SSE3
+      #define EIGEN_VECTORIZE_SSSE3
+      #define EIGEN_VECTORIZE_SSE4_1
+      #define EIGEN_VECTORIZE_SSE4_2
+    #endif
+    #ifdef __AVX2__
+      #define EIGEN_VECTORIZE_AVX2
+    #endif
+    #ifdef __FMA__
+      #define EIGEN_VECTORIZE_FMA
+    #endif
+    #if defined(__AVX512F__) && defined(EIGEN_ENABLE_AVX512)
+      #define EIGEN_VECTORIZE_AVX512
+      #define EIGEN_VECTORIZE_AVX2
+      #define EIGEN_VECTORIZE_AVX
+      #define EIGEN_VECTORIZE_FMA
+      #ifdef __AVX512DQ__
+        #define EIGEN_VECTORIZE_AVX512DQ
+      #endif
+    #endif
+
+    // include files
+
+    // This extern "C" works around a MINGW-w64 compilation issue
+    // https://sourceforge.net/tracker/index.php?func=detail&aid=3018394&group_id=202880&atid=983354
+    // In essence, intrin.h is included by windows.h and also declares intrinsics (just as emmintrin.h etc. below do).
+    // However, intrin.h uses an extern "C" declaration, and g++ thus complains of duplicate declarations
+    // with conflicting linkage.  The linkage for intrinsics doesn't matter, but at that stage the compiler doesn't know;
+    // so, to avoid compile errors when windows.h is included after Eigen/Core, ensure intrinsics are extern "C" here too.
+    // notice that since these are C headers, the extern "C" is theoretically needed anyways.
+    extern "C" {
+      // In theory we should only include immintrin.h and not the other *mmintrin.h header files directly.
+      // Doing so triggers some issues with ICC. However old gcc versions seems to not have this file, thus:
+      #if EIGEN_COMP_ICC >= 1110
+        #include <immintrin.h>
+      #else
+        #include <mmintrin.h>
+        #include <emmintrin.h>
+        #include <xmmintrin.h>
+        #ifdef  EIGEN_VECTORIZE_SSE3
+        #include <pmmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSSE3
+        #include <tmmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSE4_1
+        #include <smmintrin.h>
+        #endif
+        #ifdef EIGEN_VECTORIZE_SSE4_2
+        #include <nmmintrin.h>
+        #endif
+        #if defined(EIGEN_VECTORIZE_AVX) || defined(EIGEN_VECTORIZE_AVX512)
+        #include <immintrin.h>
+        #endif
+      #endif
+    } // end extern "C"
+  #elif defined __VSX__
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_VSX
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
+  #elif defined __ALTIVEC__
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_ALTIVEC
+    #include <altivec.h>
+    // We need to #undef all these ugly tokens defined in <altivec.h>
+    // => use __vector instead of vector
+    #undef bool
+    #undef vector
+    #undef pixel
+  #elif (defined  __ARM_NEON) || (defined __ARM_NEON__)
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_NEON
+    #include <arm_neon.h>
+  #elif (defined __s390x__ && defined __VEC__)
+    #define EIGEN_VECTORIZE
+    #define EIGEN_VECTORIZE_ZVECTOR
+    #include <vecintrin.h>
+  #endif
+#endif
+
+#if defined(__F16C__) && !defined(EIGEN_COMP_CLANG)
+  // We can use the optimized fp16 to float and float to fp16 conversion routines
+  #define EIGEN_HAS_FP16_C
+#endif
+
+#if defined __CUDACC__
+  #define EIGEN_VECTORIZE_CUDA
+  #include <vector_types.h>
+  #if defined __CUDACC_VER__ && __CUDACC_VER__ >= 70500
+    #define EIGEN_HAS_CUDA_FP16
+  #endif
+#endif
+
+#if defined EIGEN_HAS_CUDA_FP16
+  #include <host_defines.h>
+  #include <cuda_fp16.h>
+#endif
+
+#if (defined _OPENMP) && (!defined EIGEN_DONT_PARALLELIZE)
+  #define EIGEN_HAS_OPENMP
+#endif
+
+#ifdef EIGEN_HAS_OPENMP
+#include <omp.h>
+#endif
+
+// MSVC for windows mobile does not have the errno.h file
+#if !(EIGEN_COMP_MSVC && EIGEN_OS_WINCE) && !EIGEN_COMP_ARM
+#define EIGEN_HAS_ERRNO
+#endif
+
+#ifdef EIGEN_HAS_ERRNO
+#include <cerrno>
+#endif
+#include <cstddef>
+#include <cstdlib>
+#include <cmath>
+#include <cassert>
+#include <functional>
+#include <iosfwd>
+#include <cstring>
+#include <string>
+#include <limits>
+#include <climits> // for CHAR_BIT
+// for min/max:
+#include <algorithm>
+
+// for std::is_nothrow_move_assignable
+#ifdef EIGEN_INCLUDE_TYPE_TRAITS
+#include <type_traits>
+#endif
+
+// for outputting debug info
+#ifdef EIGEN_DEBUG_ASSIGN
+#include <iostream>
+#endif
+
+// required for __cpuid, needs to be included after cmath
+#if EIGEN_COMP_MSVC && EIGEN_ARCH_i386_OR_x86_64 && !EIGEN_OS_WINCE
+  #include <intrin.h>
+#endif
+
+/** \brief Namespace containing all symbols from the %Eigen library. */
+namespace Eigen {
+
+inline static const char *SimdInstructionSetsInUse(void) {
+#if defined(EIGEN_VECTORIZE_AVX512)
+  return "AVX512, FMA, AVX2, AVX, SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_AVX)
+  return "AVX SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_SSE4_2)
+  return "SSE, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2";
+#elif defined(EIGEN_VECTORIZE_SSE4_1)
+  return "SSE, SSE2, SSE3, SSSE3, SSE4.1";
+#elif defined(EIGEN_VECTORIZE_SSSE3)
+  return "SSE, SSE2, SSE3, SSSE3";
+#elif defined(EIGEN_VECTORIZE_SSE3)
+  return "SSE, SSE2, SSE3";
+#elif defined(EIGEN_VECTORIZE_SSE2)
+  return "SSE, SSE2";
+#elif defined(EIGEN_VECTORIZE_ALTIVEC)
+  return "AltiVec";
+#elif defined(EIGEN_VECTORIZE_VSX)
+  return "VSX";
+#elif defined(EIGEN_VECTORIZE_NEON)
+  return "ARM NEON";
+#elif defined(EIGEN_VECTORIZE_ZVECTOR)
+  return "S390X ZVECTOR";
+#else
+  return "None";
+#endif
+}
+
+} // end namespace Eigen
+
+#if defined EIGEN2_SUPPORT_STAGE40_FULL_EIGEN3_STRICTNESS || defined EIGEN2_SUPPORT_STAGE30_FULL_EIGEN3_API || defined EIGEN2_SUPPORT_STAGE20_RESOLVE_API_CONFLICTS || defined EIGEN2_SUPPORT_STAGE10_FULL_EIGEN2_API || defined EIGEN2_SUPPORT
+// This will generate an error message:
+#error Eigen2-support is only available up to version 3.2. Please go to "http://eigen.tuxfamily.org/index.php?title=Eigen2" for further information
+#endif
+
+namespace Eigen {
+
+// we use size_t frequently and we'll never remember to prepend it with std:: everytime just to
+// ensure QNX/QCC support
+using std::size_t;
+// gcc 4.6.0 wants std:: for ptrdiff_t
+using std::ptrdiff_t;
+
+}
+
+/** \defgroup Core_Module Core module
+  * This is the main module of Eigen providing dense matrix and vector support
+  * (both fixed and dynamic size) with all the features corresponding to a BLAS library
+  * and much more...
+  *
+  * \code
+  * #include <Eigen/Core>
+  * \endcode
+  */
+
+#include "src/Core/util/Constants.h"
+#include "src/Core/util/Meta.h"
+#include "src/Core/util/ForwardDeclarations.h"
+#include "src/Core/util/StaticAssert.h"
+#include "src/Core/util/XprHelper.h"
+#include "src/Core/util/Memory.h"
+
+#include "src/Core/NumTraits.h"
+#include "src/Core/MathFunctions.h"
+#include "src/Core/GenericPacketMath.h"
+#include "src/Core/MathFunctionsImpl.h"
+
+#if defined EIGEN_VECTORIZE_AVX512
+  #include "src/Core/arch/SSE/PacketMath.h"
+  #include "src/Core/arch/AVX/PacketMath.h"
+  #include "src/Core/arch/AVX512/PacketMath.h"
+  #include "src/Core/arch/AVX512/MathFunctions.h"
+#elif defined EIGEN_VECTORIZE_AVX
+  // Use AVX for floats and doubles, SSE for integers
+  #include "src/Core/arch/SSE/PacketMath.h"
+  #include "src/Core/arch/SSE/Complex.h"
+  #include "src/Core/arch/SSE/MathFunctions.h"
+  #include "src/Core/arch/AVX/PacketMath.h"
+  #include "src/Core/arch/AVX/MathFunctions.h"
+  #include "src/Core/arch/AVX/Complex.h"
+  #include "src/Core/arch/AVX/TypeCasting.h"
+#elif defined EIGEN_VECTORIZE_SSE
+  #include "src/Core/arch/SSE/PacketMath.h"
+  #include "src/Core/arch/SSE/MathFunctions.h"
+  #include "src/Core/arch/SSE/Complex.h"
+  #include "src/Core/arch/SSE/TypeCasting.h"
+#elif defined(EIGEN_VECTORIZE_ALTIVEC) || defined(EIGEN_VECTORIZE_VSX)
+  #include "src/Core/arch/AltiVec/PacketMath.h"
+  #include "src/Core/arch/AltiVec/MathFunctions.h"
+  #include "src/Core/arch/AltiVec/Complex.h"
+#elif defined EIGEN_VECTORIZE_NEON
+  #include "src/Core/arch/NEON/PacketMath.h"
+  #include "src/Core/arch/NEON/MathFunctions.h"
+  #include "src/Core/arch/NEON/Complex.h"
+#elif defined EIGEN_VECTORIZE_ZVECTOR
+  #include "src/Core/arch/ZVector/PacketMath.h"
+  #include "src/Core/arch/ZVector/MathFunctions.h"
+  #include "src/Core/arch/ZVector/Complex.h"
+#endif
+
+// Half float support
+#include "src/Core/arch/CUDA/Half.h"
+#include "src/Core/arch/CUDA/PacketMathHalf.h"
+#include "src/Core/arch/CUDA/TypeCasting.h"
+
+#if defined EIGEN_VECTORIZE_CUDA
+  #include "src/Core/arch/CUDA/PacketMath.h"
+  #include "src/Core/arch/CUDA/MathFunctions.h"
+#endif
+
+#include "src/Core/arch/Default/Settings.h"
+
+#include "src/Core/functors/TernaryFunctors.h"
+#include "src/Core/functors/BinaryFunctors.h"
+#include "src/Core/functors/UnaryFunctors.h"
+#include "src/Core/functors/NullaryFunctors.h"
+#include "src/Core/functors/StlFunctors.h"
+#include "src/Core/functors/AssignmentFunctors.h"
+
+// Specialized functors to enable the processing of complex numbers
+// on CUDA devices
+#include "src/Core/arch/CUDA/Complex.h"
+
+#include "src/Core/IO.h"
+#include "src/Core/DenseCoeffsBase.h"
+#include "src/Core/DenseBase.h"
+#include "src/Core/MatrixBase.h"
+#include "src/Core/EigenBase.h"
+
+#include "src/Core/Product.h"
+#include "src/Core/CoreEvaluators.h"
+#include "src/Core/AssignEvaluator.h"
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN // work around Doxygen bug triggered by Assign.h r814874
+                                // at least confirmed with Doxygen 1.5.5 and 1.5.6
+  #include "src/Core/Assign.h"
+#endif
+
+#include "src/Core/ArrayBase.h"
+#include "src/Core/util/BlasUtil.h"
+#include "src/Core/DenseStorage.h"
+#include "src/Core/NestByValue.h"
+
+// #include "src/Core/ForceAlignedAccess.h"
+
+#include "src/Core/ReturnByValue.h"
+#include "src/Core/NoAlias.h"
+#include "src/Core/PlainObjectBase.h"
+#include "src/Core/Matrix.h"
+#include "src/Core/Array.h"
+#include "src/Core/CwiseTernaryOp.h"
+#include "src/Core/CwiseBinaryOp.h"
+#include "src/Core/CwiseUnaryOp.h"
+#include "src/Core/CwiseNullaryOp.h"
+#include "src/Core/CwiseUnaryView.h"
+#include "src/Core/SelfCwiseBinaryOp.h"
+#include "src/Core/Dot.h"
+#include "src/Core/StableNorm.h"
+#include "src/Core/Stride.h"
+#include "src/Core/MapBase.h"
+#include "src/Core/Map.h"
+#include "src/Core/Ref.h"
+#include "src/Core/Block.h"
+#include "src/Core/VectorBlock.h"
+#include "src/Core/Transpose.h"
+#include "src/Core/DiagonalMatrix.h"
+#include "src/Core/Diagonal.h"
+#include "src/Core/DiagonalProduct.h"
+#include "src/Core/Redux.h"
+#include "src/Core/Visitor.h"
+#include "src/Core/Fuzzy.h"
+#include "src/Core/Swap.h"
+#include "src/Core/CommaInitializer.h"
+#include "src/Core/GeneralProduct.h"
+#include "src/Core/Solve.h"
+#include "src/Core/Inverse.h"
+#include "src/Core/SolverBase.h"
+#include "src/Core/PermutationMatrix.h"
+#include "src/Core/Transpositions.h"
+#include "src/Core/TriangularMatrix.h"
+#include "src/Core/SelfAdjointView.h"
+#include "src/Core/products/GeneralBlockPanelKernel.h"
+#include "src/Core/products/Parallelizer.h"
+#include "src/Core/ProductEvaluators.h"
+#include "src/Core/products/GeneralMatrixVector.h"
+#include "src/Core/products/GeneralMatrixMatrix.h"
+#include "src/Core/SolveTriangular.h"
+#include "src/Core/products/GeneralMatrixMatrixTriangular.h"
+#include "src/Core/products/SelfadjointMatrixVector.h"
+#include "src/Core/products/SelfadjointMatrixMatrix.h"
+#include "src/Core/products/SelfadjointProduct.h"
+#include "src/Core/products/SelfadjointRank2Update.h"
+#include "src/Core/products/TriangularMatrixVector.h"
+#include "src/Core/products/TriangularMatrixMatrix.h"
+#include "src/Core/products/TriangularSolverMatrix.h"
+#include "src/Core/products/TriangularSolverVector.h"
+#include "src/Core/BandMatrix.h"
+#include "src/Core/CoreIterators.h"
+#include "src/Core/ConditionEstimator.h"
+
+#include "src/Core/BooleanRedux.h"
+#include "src/Core/Select.h"
+#include "src/Core/VectorwiseOp.h"
+#include "src/Core/Random.h"
+#include "src/Core/Replicate.h"
+#include "src/Core/Reverse.h"
+#include "src/Core/ArrayWrapper.h"
+
+#ifdef EIGEN_USE_BLAS
+#include "src/Core/products/GeneralMatrixMatrix_BLAS.h"
+#include "src/Core/products/GeneralMatrixVector_BLAS.h"
+#include "src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h"
+#include "src/Core/products/SelfadjointMatrixMatrix_BLAS.h"
+#include "src/Core/products/SelfadjointMatrixVector_BLAS.h"
+#include "src/Core/products/TriangularMatrixMatrix_BLAS.h"
+#include "src/Core/products/TriangularMatrixVector_BLAS.h"
+#include "src/Core/products/TriangularSolverMatrix_BLAS.h"
+#endif // EIGEN_USE_BLAS
+
+#ifdef EIGEN_USE_MKL_VML
+#include "src/Core/Assign_MKL.h"
+#endif
+
+#include "src/Core/GlobalFunctions.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_CORE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Dense b/src/runtime/ref/nn/depend/external/eigen/Eigen/Dense
new file mode 100644
index 0000000..5768910
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Dense
@@ -0,0 +1,7 @@
+#include "Core"
+#include "LU"
+#include "Cholesky"
+#include "QR"
+#include "SVD"
+#include "Geometry"
+#include "Eigenvalues"
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigen b/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigen
new file mode 100644
index 0000000..654c8dc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigen
@@ -0,0 +1,2 @@
+#include "Dense"
+#include "Sparse"
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigenvalues b/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigenvalues
new file mode 100644
index 0000000..009e529
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Eigenvalues
@@ -0,0 +1,57 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EIGENVALUES_MODULE_H
+#define EIGEN_EIGENVALUES_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include "Cholesky"
+#include "Jacobi"
+#include "Householder"
+#include "LU"
+#include "Geometry"
+
+/** \defgroup Eigenvalues_Module Eigenvalues module
+  *
+  *
+  *
+  * This module mainly provides various eigenvalue solvers.
+  * This module also provides some MatrixBase methods, including:
+  *  - MatrixBase::eigenvalues(),
+  *  - MatrixBase::operatorNorm()
+  *
+  * \code
+  * #include <Eigen/Eigenvalues>
+  * \endcode
+  */
+
+#include "src/misc/RealSvd2x2.h"
+#include "src/Eigenvalues/Tridiagonalization.h"
+#include "src/Eigenvalues/RealSchur.h"
+#include "src/Eigenvalues/EigenSolver.h"
+#include "src/Eigenvalues/SelfAdjointEigenSolver.h"
+#include "src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h"
+#include "src/Eigenvalues/HessenbergDecomposition.h"
+#include "src/Eigenvalues/ComplexSchur.h"
+#include "src/Eigenvalues/ComplexEigenSolver.h"
+#include "src/Eigenvalues/RealQZ.h"
+#include "src/Eigenvalues/GeneralizedEigenSolver.h"
+#include "src/Eigenvalues/MatrixBaseEigenvalues.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/misc/lapacke.h"
+#include "src/Eigenvalues/RealSchur_LAPACKE.h"
+#include "src/Eigenvalues/ComplexSchur_LAPACKE.h"
+#include "src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_EIGENVALUES_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Geometry b/src/runtime/ref/nn/depend/external/eigen/Eigen/Geometry
new file mode 100644
index 0000000..716d529
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Geometry
@@ -0,0 +1,62 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GEOMETRY_MODULE_H
+#define EIGEN_GEOMETRY_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include "SVD"
+#include "LU"
+#include <limits>
+
+/** \defgroup Geometry_Module Geometry module
+  *
+  * This module provides support for:
+  *  - fixed-size homogeneous transformations
+  *  - translation, scaling, 2D and 3D rotations
+  *  - \link Quaternion quaternions \endlink
+  *  - cross products (\ref MatrixBase::cross, \ref MatrixBase::cross3)
+  *  - orthognal vector generation (\ref MatrixBase::unitOrthogonal)
+  *  - some linear components: \link ParametrizedLine parametrized-lines \endlink and \link Hyperplane hyperplanes \endlink
+  *  - \link AlignedBox axis aligned bounding boxes \endlink
+  *  - \link umeyama least-square transformation fitting \endlink
+  *
+  * \code
+  * #include <Eigen/Geometry>
+  * \endcode
+  */
+
+#include "src/Geometry/OrthoMethods.h"
+#include "src/Geometry/EulerAngles.h"
+
+#include "src/Geometry/Homogeneous.h"
+#include "src/Geometry/RotationBase.h"
+#include "src/Geometry/Rotation2D.h"
+#include "src/Geometry/Quaternion.h"
+#include "src/Geometry/AngleAxis.h"
+#include "src/Geometry/Transform.h"
+#include "src/Geometry/Translation.h"
+#include "src/Geometry/Scaling.h"
+#include "src/Geometry/Hyperplane.h"
+#include "src/Geometry/ParametrizedLine.h"
+#include "src/Geometry/AlignedBox.h"
+#include "src/Geometry/Umeyama.h"
+
+// Use the SSE optimized version whenever possible. At the moment the
+// SSE version doesn't compile when AVX is enabled
+#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX
+#include "src/Geometry/arch/Geometry_SSE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_GEOMETRY_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Householder b/src/runtime/ref/nn/depend/external/eigen/Eigen/Householder
new file mode 100644
index 0000000..89cd81b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Householder
@@ -0,0 +1,30 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HOUSEHOLDER_MODULE_H
+#define EIGEN_HOUSEHOLDER_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup Householder_Module Householder module
+  * This module provides Householder transformations.
+  *
+  * \code
+  * #include <Eigen/Householder>
+  * \endcode
+  */
+
+#include "src/Householder/Householder.h"
+#include "src/Householder/HouseholderSequence.h"
+#include "src/Householder/BlockHouseholder.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_HOUSEHOLDER_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/IterativeLinearSolvers b/src/runtime/ref/nn/depend/external/eigen/Eigen/IterativeLinearSolvers
new file mode 100644
index 0000000..957d575
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/IterativeLinearSolvers
@@ -0,0 +1,48 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
+#define EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
+
+#include "SparseCore"
+#include "OrderingMethods"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** 
+  * \defgroup IterativeLinearSolvers_Module IterativeLinearSolvers module
+  *
+  * This module currently provides iterative methods to solve problems of the form \c A \c x = \c b, where \c A is a squared matrix, usually very large and sparse.
+  * Those solvers are accessible via the following classes:
+  *  - ConjugateGradient for selfadjoint (hermitian) matrices,
+  *  - LeastSquaresConjugateGradient for rectangular least-square problems,
+  *  - BiCGSTAB for general square matrices.
+  *
+  * These iterative solvers are associated with some preconditioners:
+  *  - IdentityPreconditioner - not really useful
+  *  - DiagonalPreconditioner - also called Jacobi preconditioner, work very well on diagonal dominant matrices.
+  *  - IncompleteLUT - incomplete LU factorization with dual thresholding
+  *
+  * Such problems can also be solved using the direct sparse decomposition modules: SparseCholesky, CholmodSupport, UmfPackSupport, SuperLUSupport.
+  *
+    \code
+    #include <Eigen/IterativeLinearSolvers>
+    \endcode
+  */
+
+#include "src/IterativeLinearSolvers/SolveWithGuess.h"
+#include "src/IterativeLinearSolvers/IterativeSolverBase.h"
+#include "src/IterativeLinearSolvers/BasicPreconditioners.h"
+#include "src/IterativeLinearSolvers/ConjugateGradient.h"
+#include "src/IterativeLinearSolvers/LeastSquareConjugateGradient.h"
+#include "src/IterativeLinearSolvers/BiCGSTAB.h"
+#include "src/IterativeLinearSolvers/IncompleteLUT.h"
+#include "src/IterativeLinearSolvers/IncompleteCholesky.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ITERATIVELINEARSOLVERS_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Jacobi b/src/runtime/ref/nn/depend/external/eigen/Eigen/Jacobi
new file mode 100644
index 0000000..17c1d78
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Jacobi
@@ -0,0 +1,33 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_JACOBI_MODULE_H
+#define EIGEN_JACOBI_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup Jacobi_Module Jacobi module
+  * This module provides Jacobi and Givens rotations.
+  *
+  * \code
+  * #include <Eigen/Jacobi>
+  * \endcode
+  *
+  * In addition to listed classes, it defines the two following MatrixBase methods to apply a Jacobi or Givens rotation:
+  *  - MatrixBase::applyOnTheLeft()
+  *  - MatrixBase::applyOnTheRight().
+  */
+
+#include "src/Jacobi/Jacobi.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_JACOBI_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/LU b/src/runtime/ref/nn/depend/external/eigen/Eigen/LU
new file mode 100644
index 0000000..6f6c556
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/LU
@@ -0,0 +1,46 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LU_MODULE_H
+#define EIGEN_LU_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup LU_Module LU module
+  * This module includes %LU decomposition and related notions such as matrix inversion and determinant.
+  * This module defines the following MatrixBase methods:
+  *  - MatrixBase::inverse()
+  *  - MatrixBase::determinant()
+  *
+  * \code
+  * #include <Eigen/LU>
+  * \endcode
+  */
+
+#include "src/misc/Kernel.h"
+#include "src/misc/Image.h"
+#include "src/LU/FullPivLU.h"
+#include "src/LU/PartialPivLU.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/misc/lapacke.h"
+#include "src/LU/PartialPivLU_LAPACKE.h"
+#endif
+#include "src/LU/Determinant.h"
+#include "src/LU/InverseImpl.h"
+
+// Use the SSE optimized version whenever possible. At the moment the
+// SSE version doesn't compile when AVX is enabled
+#if defined EIGEN_VECTORIZE_SSE && !defined EIGEN_VECTORIZE_AVX
+  #include "src/LU/arch/Inverse_SSE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_LU_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/MetisSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/MetisSupport
new file mode 100644
index 0000000..85c41bf
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/MetisSupport
@@ -0,0 +1,35 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_METISSUPPORT_MODULE_H
+#define EIGEN_METISSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+#include <metis.h>
+}
+
+
+/** \ingroup Support_modules
+  * \defgroup MetisSupport_Module MetisSupport module
+  *
+  * \code
+  * #include <Eigen/MetisSupport>
+  * \endcode
+  * This module defines an interface to the METIS reordering package (http://glaros.dtc.umn.edu/gkhome/views/metis). 
+  * It can be used just as any other built-in method as explained in \link OrderingMethods_Module here. \endlink
+  */
+
+
+#include "src/MetisSupport/MetisSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_METISSUPPORT_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/OrderingMethods b/src/runtime/ref/nn/depend/external/eigen/Eigen/OrderingMethods
new file mode 100644
index 0000000..d8ea361
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/OrderingMethods
@@ -0,0 +1,73 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ORDERINGMETHODS_MODULE_H
+#define EIGEN_ORDERINGMETHODS_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** 
+  * \defgroup OrderingMethods_Module OrderingMethods module
+  *
+  * This module is currently for internal use only
+  * 
+  * It defines various built-in and external ordering methods for sparse matrices. 
+  * They are typically used to reduce the number of elements during 
+  * the sparse matrix decomposition (LLT, LU, QR).
+  * Precisely, in a preprocessing step, a permutation matrix P is computed using 
+  * those ordering methods and applied to the columns of the matrix. 
+  * Using for instance the sparse Cholesky decomposition, it is expected that 
+  * the nonzeros elements in LLT(A*P) will be much smaller than that in LLT(A).
+  * 
+  * 
+  * Usage : 
+  * \code
+  * #include <Eigen/OrderingMethods>
+  * \endcode
+  * 
+  * A simple usage is as a template parameter in the sparse decomposition classes : 
+  * 
+  * \code 
+  * SparseLU<MatrixType, COLAMDOrdering<int> > solver;
+  * \endcode 
+  * 
+  * \code 
+  * SparseQR<MatrixType, COLAMDOrdering<int> > solver;
+  * \endcode
+  * 
+  * It is possible as well to call directly a particular ordering method for your own purpose, 
+  * \code 
+  * AMDOrdering<int> ordering;
+  * PermutationMatrix<Dynamic, Dynamic, int> perm;
+  * SparseMatrix<double> A; 
+  * //Fill the matrix ...
+  * 
+  * ordering(A, perm); // Call AMD
+  * \endcode
+  * 
+  * \note Some of these methods (like AMD or METIS), need the sparsity pattern 
+  * of the input matrix to be symmetric. When the matrix is structurally unsymmetric, 
+  * Eigen computes internally the pattern of \f$A^T*A\f$ before calling the method.
+  * If your matrix is already symmetric (at leat in structure), you can avoid that
+  * by calling the method with a SelfAdjointView type.
+  * 
+  * \code
+  *  // Call the ordering on the pattern of the lower triangular matrix A
+  * ordering(A.selfadjointView<Lower>(), perm);
+  * \endcode
+  */
+
+#ifndef EIGEN_MPL2_ONLY
+#include "src/OrderingMethods/Amd.h"
+#endif
+
+#include "src/OrderingMethods/Ordering.h"
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_ORDERINGMETHODS_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/PaStiXSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/PaStiXSupport
new file mode 100644
index 0000000..de3a63b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/PaStiXSupport
@@ -0,0 +1,48 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PASTIXSUPPORT_MODULE_H
+#define EIGEN_PASTIXSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+#include <pastix_nompi.h>
+#include <pastix.h>
+}
+
+#ifdef complex
+#undef complex
+#endif
+
+/** \ingroup Support_modules
+  * \defgroup PaStiXSupport_Module PaStiXSupport module
+  * 
+  * This module provides an interface to the <a href="http://pastix.gforge.inria.fr/">PaSTiX</a> library.
+  * PaSTiX is a general \b supernodal, \b parallel and \b opensource sparse solver.
+  * It provides the two following main factorization classes:
+  * - class PastixLLT : a supernodal, parallel LLt Cholesky factorization.
+  * - class PastixLDLT: a supernodal, parallel LDLt Cholesky factorization.
+  * - class PastixLU : a supernodal, parallel LU factorization (optimized for a symmetric pattern).
+  * 
+  * \code
+  * #include <Eigen/PaStiXSupport>
+  * \endcode
+  *
+  * In order to use this module, the PaSTiX headers must be accessible from the include paths, and your binary must be linked to the PaSTiX library and its dependencies.
+  * The dependencies depend on how PaSTiX has been compiled.
+  * For a cmake based project, you can use our FindPaSTiX.cmake module to help you in this task.
+  *
+  */
+
+#include "src/PaStiXSupport/PaStiXSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_PASTIXSUPPORT_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/PardisoSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/PardisoSupport
new file mode 100755
index 0000000..340edf5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/PardisoSupport
@@ -0,0 +1,35 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARDISOSUPPORT_MODULE_H
+#define EIGEN_PARDISOSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include <mkl_pardiso.h>
+
+/** \ingroup Support_modules
+  * \defgroup PardisoSupport_Module PardisoSupport module
+  *
+  * This module brings support for the Intel(R) MKL PARDISO direct sparse solvers.
+  *
+  * \code
+  * #include <Eigen/PardisoSupport>
+  * \endcode
+  *
+  * In order to use this module, the MKL headers must be accessible from the include paths, and your binary must be linked to the MKL library and its dependencies.
+  * See this \ref TopicUsingIntelMKL "page" for more information on MKL-Eigen integration.
+  * 
+  */
+
+#include "src/PardisoSupport/PardisoSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_PARDISOSUPPORT_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/QR b/src/runtime/ref/nn/depend/external/eigen/Eigen/QR
new file mode 100644
index 0000000..80838e3
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/QR
@@ -0,0 +1,47 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_QR_MODULE_H
+#define EIGEN_QR_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include "Cholesky"
+#include "Jacobi"
+#include "Householder"
+
+/** \defgroup QR_Module QR module
+  *
+  *
+  *
+  * This module provides various QR decompositions
+  * This module also provides some MatrixBase methods, including:
+  *  - MatrixBase::householderQr()
+  *  - MatrixBase::colPivHouseholderQr()
+  *  - MatrixBase::fullPivHouseholderQr()
+  *
+  * \code
+  * #include <Eigen/QR>
+  * \endcode
+  */
+
+#include "src/QR/HouseholderQR.h"
+#include "src/QR/FullPivHouseholderQR.h"
+#include "src/QR/ColPivHouseholderQR.h"
+#include "src/QR/CompleteOrthogonalDecomposition.h"
+#ifdef EIGEN_USE_LAPACKE
+#include "src/misc/lapacke.h"
+#include "src/QR/HouseholderQR_LAPACKE.h"
+#include "src/QR/ColPivHouseholderQR_LAPACKE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_QR_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/QtAlignedMalloc b/src/runtime/ref/nn/depend/external/eigen/Eigen/QtAlignedMalloc
new file mode 100644
index 0000000..c6571f1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/QtAlignedMalloc
@@ -0,0 +1,40 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_QTMALLOC_MODULE_H
+#define EIGEN_QTMALLOC_MODULE_H
+
+#include "Core"
+
+#if (!EIGEN_MALLOC_ALREADY_ALIGNED)
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+void *qMalloc(std::size_t size)
+{
+  return Eigen::internal::aligned_malloc(size);
+}
+
+void qFree(void *ptr)
+{
+  Eigen::internal::aligned_free(ptr);
+}
+
+void *qRealloc(void *ptr, std::size_t size)
+{
+  void* newPtr = Eigen::internal::aligned_malloc(size);
+  memcpy(newPtr, ptr, size);
+  Eigen::internal::aligned_free(ptr);
+  return newPtr;
+}
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif
+
+#endif // EIGEN_QTMALLOC_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SPQRSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/SPQRSupport
new file mode 100644
index 0000000..f70390c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SPQRSupport
@@ -0,0 +1,34 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPQRSUPPORT_MODULE_H
+#define EIGEN_SPQRSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include "SuiteSparseQR.hpp"
+
+/** \ingroup Support_modules
+  * \defgroup SPQRSupport_Module SuiteSparseQR module
+  * 
+  * This module provides an interface to the SPQR library, which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
+  *
+  * \code
+  * #include <Eigen/SPQRSupport>
+  * \endcode
+  *
+  * In order to use this module, the SPQR headers must be accessible from the include paths, and your binary must be linked to the SPQR library and its dependencies (Cholmod, AMD, COLAMD,...).
+  * For a cmake based project, you can use our FindSPQR.cmake and FindCholmod.Cmake modules
+  *
+  */
+
+#include "src/CholmodSupport/CholmodSupport.h"
+#include "src/SPQRSupport/SuiteSparseQRSupport.h"
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SVD b/src/runtime/ref/nn/depend/external/eigen/Eigen/SVD
new file mode 100644
index 0000000..86143c2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SVD
@@ -0,0 +1,47 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SVD_MODULE_H
+#define EIGEN_SVD_MODULE_H
+
+#include "QR"
+#include "Householder"
+#include "Jacobi"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup SVD_Module SVD module
+  *
+  *
+  *
+  * This module provides SVD decomposition for matrices (both real and complex).
+  * Two decomposition algorithms are provided:
+  *  - JacobiSVD implementing two-sided Jacobi iterations is numerically very accurate, fast for small matrices, but very slow for larger ones.
+  *  - BDCSVD implementing a recursive divide & conquer strategy on top of an upper-bidiagonalization which remains fast for large problems.
+  * These decompositions are accessible via the respective classes and following MatrixBase methods:
+  *  - MatrixBase::jacobiSvd()
+  *  - MatrixBase::bdcSvd()
+  *
+  * \code
+  * #include <Eigen/SVD>
+  * \endcode
+  */
+
+#include "src/misc/RealSvd2x2.h"
+#include "src/SVD/UpperBidiagonalization.h"
+#include "src/SVD/SVDBase.h"
+#include "src/SVD/JacobiSVD.h"
+#include "src/SVD/BDCSVD.h"
+#if defined(EIGEN_USE_LAPACKE) && !defined(EIGEN_USE_LAPACKE_STRICT)
+#include "src/misc/lapacke.h"
+#include "src/SVD/JacobiSVD_LAPACKE.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SVD_MODULE_H
+/* vim: set filetype=cpp et sw=2 ts=2 ai: */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/Sparse b/src/runtime/ref/nn/depend/external/eigen/Eigen/Sparse
new file mode 100644
index 0000000..136e681
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/Sparse
@@ -0,0 +1,36 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_MODULE_H
+#define EIGEN_SPARSE_MODULE_H
+
+/** \defgroup Sparse_Module Sparse meta-module
+  *
+  * Meta-module including all related modules:
+  * - \ref SparseCore_Module
+  * - \ref OrderingMethods_Module
+  * - \ref SparseCholesky_Module
+  * - \ref SparseLU_Module
+  * - \ref SparseQR_Module
+  * - \ref IterativeLinearSolvers_Module
+  *
+    \code
+    #include <Eigen/Sparse>
+    \endcode
+  */
+
+#include "SparseCore"
+#include "OrderingMethods"
+#ifndef EIGEN_MPL2_ONLY
+#include "SparseCholesky"
+#endif
+#include "SparseLU"
+#include "SparseQR"
+#include "IterativeLinearSolvers"
+
+#endif // EIGEN_SPARSE_MODULE_H
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCholesky b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCholesky
new file mode 100644
index 0000000..b6a320c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCholesky
@@ -0,0 +1,45 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2013 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSECHOLESKY_MODULE_H
+#define EIGEN_SPARSECHOLESKY_MODULE_H
+
+#include "SparseCore"
+#include "OrderingMethods"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** 
+  * \defgroup SparseCholesky_Module SparseCholesky module
+  *
+  * This module currently provides two variants of the direct sparse Cholesky decomposition for selfadjoint (hermitian) matrices.
+  * Those decompositions are accessible via the following classes:
+  *  - SimplicialLLt,
+  *  - SimplicialLDLt
+  *
+  * Such problems can also be solved using the ConjugateGradient solver from the IterativeLinearSolvers module.
+  *
+  * \code
+  * #include <Eigen/SparseCholesky>
+  * \endcode
+  */
+
+#ifdef EIGEN_MPL2_ONLY
+#error The SparseCholesky module has nothing to offer in MPL2 only mode
+#endif
+
+#include "src/SparseCholesky/SimplicialCholesky.h"
+
+#ifndef EIGEN_MPL2_ONLY
+#include "src/SparseCholesky/SimplicialCholesky_impl.h"
+#endif
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPARSECHOLESKY_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCore b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCore
new file mode 100644
index 0000000..76966c4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseCore
@@ -0,0 +1,69 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSECORE_MODULE_H
+#define EIGEN_SPARSECORE_MODULE_H
+
+#include "Core"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#include <vector>
+#include <map>
+#include <cstdlib>
+#include <cstring>
+#include <algorithm>
+
+/** 
+  * \defgroup SparseCore_Module SparseCore module
+  *
+  * This module provides a sparse matrix representation, and basic associated matrix manipulations
+  * and operations.
+  *
+  * See the \ref TutorialSparse "Sparse tutorial"
+  *
+  * \code
+  * #include <Eigen/SparseCore>
+  * \endcode
+  *
+  * This module depends on: Core.
+  */
+
+#include "src/SparseCore/SparseUtil.h"
+#include "src/SparseCore/SparseMatrixBase.h"
+#include "src/SparseCore/SparseAssign.h"
+#include "src/SparseCore/CompressedStorage.h"
+#include "src/SparseCore/AmbiVector.h"
+#include "src/SparseCore/SparseCompressedBase.h"
+#include "src/SparseCore/SparseMatrix.h"
+#include "src/SparseCore/SparseMap.h"
+#include "src/SparseCore/MappedSparseMatrix.h"
+#include "src/SparseCore/SparseVector.h"
+#include "src/SparseCore/SparseRef.h"
+#include "src/SparseCore/SparseCwiseUnaryOp.h"
+#include "src/SparseCore/SparseCwiseBinaryOp.h"
+#include "src/SparseCore/SparseTranspose.h"
+#include "src/SparseCore/SparseBlock.h"
+#include "src/SparseCore/SparseDot.h"
+#include "src/SparseCore/SparseRedux.h"
+#include "src/SparseCore/SparseView.h"
+#include "src/SparseCore/SparseDiagonalProduct.h"
+#include "src/SparseCore/ConservativeSparseSparseProduct.h"
+#include "src/SparseCore/SparseSparseProductWithPruning.h"
+#include "src/SparseCore/SparseProduct.h"
+#include "src/SparseCore/SparseDenseProduct.h"
+#include "src/SparseCore/SparseSelfAdjointView.h"
+#include "src/SparseCore/SparseTriangularView.h"
+#include "src/SparseCore/TriangularSolver.h"
+#include "src/SparseCore/SparsePermutation.h"
+#include "src/SparseCore/SparseFuzzy.h"
+#include "src/SparseCore/SparseSolverBase.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SPARSECORE_MODULE_H
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseLU b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseLU
new file mode 100644
index 0000000..38b38b5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseLU
@@ -0,0 +1,46 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSELU_MODULE_H
+#define EIGEN_SPARSELU_MODULE_H
+
+#include "SparseCore"
+
+/** 
+  * \defgroup SparseLU_Module SparseLU module
+  * This module defines a supernodal factorization of general sparse matrices.
+  * The code is fully optimized for supernode-panel updates with specialized kernels.
+  * Please, see the documentation of the SparseLU class for more details.
+  */
+
+// Ordering interface
+#include "OrderingMethods"
+
+#include "src/SparseLU/SparseLU_gemm_kernel.h"
+
+#include "src/SparseLU/SparseLU_Structs.h"
+#include "src/SparseLU/SparseLU_SupernodalMatrix.h"
+#include "src/SparseLU/SparseLUImpl.h"
+#include "src/SparseCore/SparseColEtree.h"
+#include "src/SparseLU/SparseLU_Memory.h"
+#include "src/SparseLU/SparseLU_heap_relax_snode.h"
+#include "src/SparseLU/SparseLU_relax_snode.h"
+#include "src/SparseLU/SparseLU_pivotL.h"
+#include "src/SparseLU/SparseLU_panel_dfs.h"
+#include "src/SparseLU/SparseLU_kernel_bmod.h"
+#include "src/SparseLU/SparseLU_panel_bmod.h"
+#include "src/SparseLU/SparseLU_column_dfs.h"
+#include "src/SparseLU/SparseLU_column_bmod.h"
+#include "src/SparseLU/SparseLU_copy_to_ucol.h"
+#include "src/SparseLU/SparseLU_pruneL.h"
+#include "src/SparseLU/SparseLU_Utils.h"
+#include "src/SparseLU/SparseLU.h"
+
+#endif // EIGEN_SPARSELU_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseQR b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseQR
new file mode 100644
index 0000000..a6f3b7f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SparseQR
@@ -0,0 +1,37 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEQR_MODULE_H
+#define EIGEN_SPARSEQR_MODULE_H
+
+#include "SparseCore"
+#include "OrderingMethods"
+#include "src/Core/util/DisableStupidWarnings.h"
+
+/** \defgroup SparseQR_Module SparseQR module
+  * \brief Provides QR decomposition for sparse matrices
+  * 
+  * This module provides a simplicial version of the left-looking Sparse QR decomposition. 
+  * The columns of the input matrix should be reordered to limit the fill-in during the 
+  * decomposition. Built-in methods (COLAMD, AMD) or external  methods (METIS) can be used to this end.
+  * See the \link OrderingMethods_Module OrderingMethods\endlink module for the list 
+  * of built-in and external ordering methods.
+  * 
+  * \code
+  * #include <Eigen/SparseQR>
+  * \endcode
+  * 
+  * 
+  */
+
+#include "OrderingMethods"
+#include "src/SparseCore/SparseColEtree.h"
+#include "src/SparseQR/SparseQR.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/StdDeque b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdDeque
new file mode 100644
index 0000000..bc68397
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdDeque
@@ -0,0 +1,27 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDDEQUE_MODULE_H
+#define EIGEN_STDDEQUE_MODULE_H
+
+#include "Core"
+#include <deque>
+
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
+
+#define EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(...)
+
+#else
+
+#include "src/StlSupport/StdDeque.h"
+
+#endif
+
+#endif // EIGEN_STDDEQUE_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/StdList b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdList
new file mode 100644
index 0000000..4c6262c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdList
@@ -0,0 +1,26 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDLIST_MODULE_H
+#define EIGEN_STDLIST_MODULE_H
+
+#include "Core"
+#include <list>
+
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
+
+#define EIGEN_DEFINE_STL_LIST_SPECIALIZATION(...)
+
+#else
+
+#include "src/StlSupport/StdList.h"
+
+#endif
+
+#endif // EIGEN_STDLIST_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/StdVector b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdVector
new file mode 100644
index 0000000..0c4697a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/StdVector
@@ -0,0 +1,27 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDVECTOR_MODULE_H
+#define EIGEN_STDVECTOR_MODULE_H
+
+#include "Core"
+#include <vector>
+
+#if EIGEN_COMP_MSVC && EIGEN_OS_WIN64 && (EIGEN_MAX_STATIC_ALIGN_BYTES<=16) /* MSVC auto aligns up to 16 bytes in 64 bit builds */
+
+#define EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(...)
+
+#else
+
+#include "src/StlSupport/StdVector.h"
+
+#endif
+
+#endif // EIGEN_STDVECTOR_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/SuperLUSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/SuperLUSupport
new file mode 100644
index 0000000..59312a8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/SuperLUSupport
@@ -0,0 +1,64 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SUPERLUSUPPORT_MODULE_H
+#define EIGEN_SUPERLUSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+#ifdef EMPTY
+#define EIGEN_EMPTY_WAS_ALREADY_DEFINED
+#endif
+
+typedef int int_t;
+#include <slu_Cnames.h>
+#include <supermatrix.h>
+#include <slu_util.h>
+
+// slu_util.h defines a preprocessor token named EMPTY which is really polluting,
+// so we remove it in favor of a SUPERLU_EMPTY token.
+// If EMPTY was already defined then we don't undef it.
+
+#if defined(EIGEN_EMPTY_WAS_ALREADY_DEFINED)
+# undef EIGEN_EMPTY_WAS_ALREADY_DEFINED
+#elif defined(EMPTY)
+# undef EMPTY
+#endif
+
+#define SUPERLU_EMPTY (-1)
+
+namespace Eigen { struct SluMatrix; }
+
+/** \ingroup Support_modules
+  * \defgroup SuperLUSupport_Module SuperLUSupport module
+  *
+  * This module provides an interface to the <a href="http://crd-legacy.lbl.gov/~xiaoye/SuperLU/">SuperLU</a> library.
+  * It provides the following factorization class:
+  * - class SuperLU: a supernodal sequential LU factorization.
+  * - class SuperILU: a supernodal sequential incomplete LU factorization (to be used as a preconditioner for iterative methods).
+  *
+  * \warning This wrapper requires at least versions 4.0 of SuperLU. The 3.x versions are not supported.
+  *
+  * \warning When including this module, you have to use SUPERLU_EMPTY instead of EMPTY which is no longer defined because it is too polluting.
+  *
+  * \code
+  * #include <Eigen/SuperLUSupport>
+  * \endcode
+  *
+  * In order to use this module, the superlu headers must be accessible from the include paths, and your binary must be linked to the superlu library and its dependencies.
+  * The dependencies depend on how superlu has been compiled.
+  * For a cmake based project, you can use our FindSuperLU.cmake module to help you in this task.
+  *
+  */
+
+#include "src/SuperLUSupport/SuperLUSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_SUPERLUSUPPORT_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/UmfPackSupport b/src/runtime/ref/nn/depend/external/eigen/Eigen/UmfPackSupport
new file mode 100644
index 0000000..00eec80
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/UmfPackSupport
@@ -0,0 +1,40 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_UMFPACKSUPPORT_MODULE_H
+#define EIGEN_UMFPACKSUPPORT_MODULE_H
+
+#include "SparseCore"
+
+#include "src/Core/util/DisableStupidWarnings.h"
+
+extern "C" {
+#include <umfpack.h>
+}
+
+/** \ingroup Support_modules
+  * \defgroup UmfPackSupport_Module UmfPackSupport module
+  *
+  * This module provides an interface to the UmfPack library which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
+  * It provides the following factorization class:
+  * - class UmfPackLU: a multifrontal sequential LU factorization.
+  *
+  * \code
+  * #include <Eigen/UmfPackSupport>
+  * \endcode
+  *
+  * In order to use this module, the umfpack headers must be accessible from the include paths, and your binary must be linked to the umfpack library and its dependencies.
+  * The dependencies depend on how umfpack has been compiled.
+  * For a cmake based project, you can use our FindUmfPack.cmake module to help you in this task.
+  *
+  */
+
+#include "src/UmfPackSupport/UmfPackSupport.h"
+
+#include "src/Core/util/ReenableStupidWarnings.h"
+
+#endif // EIGEN_UMFPACKSUPPORT_MODULE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LDLT.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LDLT.h
new file mode 100644
index 0000000..fcee7b2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LDLT.h
@@ -0,0 +1,669 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Keir Mierle <mierle@gmail.com>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011 Timothy E. Holy <tim.holy@gmail.com >
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LDLT_H
+#define EIGEN_LDLT_H
+
+namespace Eigen {
+
+namespace internal {
+  template<typename MatrixType, int UpLo> struct LDLT_Traits;
+
+  // PositiveSemiDef means positive semi-definite and non-zero; same for NegativeSemiDef
+  enum SignMatrix { PositiveSemiDef, NegativeSemiDef, ZeroSign, Indefinite };
+}
+
+/** \ingroup Cholesky_Module
+  *
+  * \class LDLT
+  *
+  * \brief Robust Cholesky decomposition of a matrix with pivoting
+  *
+  * \tparam _MatrixType the type of the matrix of which to compute the LDL^T Cholesky decomposition
+  * \tparam _UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
+  *             The other triangular part won't be read.
+  *
+  * Perform a robust Cholesky decomposition of a positive semidefinite or negative semidefinite
+  * matrix \f$ A \f$ such that \f$ A =  P^TLDL^*P \f$, where P is a permutation matrix, L
+  * is lower triangular with a unit diagonal and D is a diagonal matrix.
+  *
+  * The decomposition uses pivoting to ensure stability, so that L will have
+  * zeros in the bottom right rank(A) - n submatrix. Avoiding the square root
+  * on D also stabilizes the computation.
+  *
+  * Remember that Cholesky decompositions are not rank-revealing. Also, do not use a Cholesky
+  * decomposition to determine whether a system of equations has a solution.
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt(), class LLT
+  */
+template<typename _MatrixType, int _UpLo> class LDLT
+{
+  public:
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+      UpLo = _UpLo
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar, RowsAtCompileTime, 1, 0, MaxRowsAtCompileTime, 1> TmpMatrixType;
+
+    typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
+    typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationType;
+
+    typedef internal::LDLT_Traits<MatrixType,UpLo> Traits;
+
+    /** \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via LDLT::compute(const MatrixType&).
+      */
+    LDLT()
+      : m_matrix(),
+        m_transpositions(),
+        m_sign(internal::ZeroSign),
+        m_isInitialized(false)
+    {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa LDLT()
+      */
+    explicit LDLT(Index size)
+      : m_matrix(size, size),
+        m_transpositions(size),
+        m_temporary(size),
+        m_sign(internal::ZeroSign),
+        m_isInitialized(false)
+    {}
+
+    /** \brief Constructor with decomposition
+      *
+      * This calculates the decomposition for the input \a matrix.
+      *
+      * \sa LDLT(Index size)
+      */
+    template<typename InputType>
+    explicit LDLT(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.rows(), matrix.cols()),
+        m_transpositions(matrix.rows()),
+        m_temporary(matrix.rows()),
+        m_sign(internal::ZeroSign),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** \brief Constructs a LDLT factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
+      *
+      * \sa LDLT(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit LDLT(EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
+        m_transpositions(matrix.rows()),
+        m_temporary(matrix.rows()),
+        m_sign(internal::ZeroSign),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** Clear any existing decomposition
+     * \sa rankUpdate(w,sigma)
+     */
+    void setZero()
+    {
+      m_isInitialized = false;
+    }
+
+    /** \returns a view of the upper triangular matrix U */
+    inline typename Traits::MatrixU matrixU() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return Traits::getU(m_matrix);
+    }
+
+    /** \returns a view of the lower triangular matrix L */
+    inline typename Traits::MatrixL matrixL() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return Traits::getL(m_matrix);
+    }
+
+    /** \returns the permutation matrix P as a transposition sequence.
+      */
+    inline const TranspositionType& transpositionsP() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_transpositions;
+    }
+
+    /** \returns the coefficients of the diagonal matrix D */
+    inline Diagonal<const MatrixType> vectorD() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_matrix.diagonal();
+    }
+
+    /** \returns true if the matrix is positive (semidefinite) */
+    inline bool isPositive() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_sign == internal::PositiveSemiDef || m_sign == internal::ZeroSign;
+    }
+
+    /** \returns true if the matrix is negative (semidefinite) */
+    inline bool isNegative(void) const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_sign == internal::NegativeSemiDef || m_sign == internal::ZeroSign;
+    }
+
+    /** \returns a solution x of \f$ A x = b \f$ using the current decomposition of A.
+      *
+      * This function also supports in-place solves using the syntax <tt>x = decompositionObject.solve(x)</tt> .
+      *
+      * \note_about_checking_solutions
+      *
+      * More precisely, this method solves \f$ A x = b \f$ using the decomposition \f$ A = P^T L D L^* P \f$
+      * by solving the systems \f$ P^T y_1 = b \f$, \f$ L y_2 = y_1 \f$, \f$ D y_3 = y_2 \f$,
+      * \f$ L^* y_4 = y_3 \f$ and \f$ P x = y_4 \f$ in succession. If the matrix \f$ A \f$ is singular, then
+      * \f$ D \f$ will also be singular (all the other matrices are invertible). In that case, the
+      * least-square solution of \f$ D y_3 = y_2 \f$ is computed. This does not mean that this function
+      * computes the least-square solution of \f$ A x = b \f$ is \f$ A \f$ is singular.
+      *
+      * \sa MatrixBase::ldlt(), SelfAdjointView::ldlt()
+      */
+    template<typename Rhs>
+    inline const Solve<LDLT, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      eigen_assert(m_matrix.rows()==b.rows()
+                && "LDLT::solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<LDLT, Rhs>(*this, b.derived());
+    }
+
+    template<typename Derived>
+    bool solveInPlace(MatrixBase<Derived> &bAndX) const;
+
+    template<typename InputType>
+    LDLT& compute(const EigenBase<InputType>& matrix);
+
+    /** \returns an estimate of the reciprocal condition number of the matrix of
+     *  which \c *this is the LDLT decomposition.
+     */
+    RealScalar rcond() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return internal::rcond_estimate_helper(m_l1_norm, *this);
+    }
+
+    template <typename Derived>
+    LDLT& rankUpdate(const MatrixBase<Derived>& w, const RealScalar& alpha=1);
+
+    /** \returns the internal LDLT decomposition matrix
+      *
+      * TODO: document the storage layout
+      */
+    inline const MatrixType& matrixLDLT() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_matrix;
+    }
+
+    MatrixType reconstructedMatrix() const;
+
+    /** \returns the adjoint of \c *this, that is, a const reference to the decomposition itself as the underlying matrix is self-adjoint.
+      *
+      * This method is provided for compatibility with other matrix decompositions, thus enabling generic code such as:
+      * \code x = decomposition.adjoint().solve(b) \endcode
+      */
+    const LDLT& adjoint() const { return *this; };
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "LDLT is not initialized.");
+      return m_info;
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    /** \internal
+      * Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
+      * The strict upper part is used during the decomposition, the strict lower
+      * part correspond to the coefficients of L (its diagonal is equal to 1 and
+      * is not stored), and the diagonal entries correspond to D.
+      */
+    MatrixType m_matrix;
+    RealScalar m_l1_norm;
+    TranspositionType m_transpositions;
+    TmpMatrixType m_temporary;
+    internal::SignMatrix m_sign;
+    bool m_isInitialized;
+    ComputationInfo m_info;
+};
+
+namespace internal {
+
+template<int UpLo> struct ldlt_inplace;
+
+template<> struct ldlt_inplace<Lower>
+{
+  template<typename MatrixType, typename TranspositionType, typename Workspace>
+  static bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign)
+  {
+    using std::abs;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename TranspositionType::StorageIndex IndexType;
+    eigen_assert(mat.rows()==mat.cols());
+    const Index size = mat.rows();
+    bool found_zero_pivot = false;
+    bool ret = true;
+
+    if (size <= 1)
+    {
+      transpositions.setIdentity();
+      if (numext::real(mat.coeff(0,0)) > static_cast<RealScalar>(0) ) sign = PositiveSemiDef;
+      else if (numext::real(mat.coeff(0,0)) < static_cast<RealScalar>(0)) sign = NegativeSemiDef;
+      else sign = ZeroSign;
+      return true;
+    }
+
+    for (Index k = 0; k < size; ++k)
+    {
+      // Find largest diagonal element
+      Index index_of_biggest_in_corner;
+      mat.diagonal().tail(size-k).cwiseAbs().maxCoeff(&index_of_biggest_in_corner);
+      index_of_biggest_in_corner += k;
+
+      transpositions.coeffRef(k) = IndexType(index_of_biggest_in_corner);
+      if(k != index_of_biggest_in_corner)
+      {
+        // apply the transposition while taking care to consider only
+        // the lower triangular part
+        Index s = size-index_of_biggest_in_corner-1; // trailing size after the biggest element
+        mat.row(k).head(k).swap(mat.row(index_of_biggest_in_corner).head(k));
+        mat.col(k).tail(s).swap(mat.col(index_of_biggest_in_corner).tail(s));
+        std::swap(mat.coeffRef(k,k),mat.coeffRef(index_of_biggest_in_corner,index_of_biggest_in_corner));
+        for(Index i=k+1;i<index_of_biggest_in_corner;++i)
+        {
+          Scalar tmp = mat.coeffRef(i,k);
+          mat.coeffRef(i,k) = numext::conj(mat.coeffRef(index_of_biggest_in_corner,i));
+          mat.coeffRef(index_of_biggest_in_corner,i) = numext::conj(tmp);
+        }
+        if(NumTraits<Scalar>::IsComplex)
+          mat.coeffRef(index_of_biggest_in_corner,k) = numext::conj(mat.coeff(index_of_biggest_in_corner,k));
+      }
+
+      // partition the matrix:
+      //       A00 |  -  |  -
+      // lu  = A10 | A11 |  -
+      //       A20 | A21 | A22
+      Index rs = size - k - 1;
+      Block<MatrixType,Dynamic,1> A21(mat,k+1,k,rs,1);
+      Block<MatrixType,1,Dynamic> A10(mat,k,0,1,k);
+      Block<MatrixType,Dynamic,Dynamic> A20(mat,k+1,0,rs,k);
+
+      if(k>0)
+      {
+        temp.head(k) = mat.diagonal().real().head(k).asDiagonal() * A10.adjoint();
+        mat.coeffRef(k,k) -= (A10 * temp.head(k)).value();
+        if(rs>0)
+          A21.noalias() -= A20 * temp.head(k);
+      }
+
+      // In some previous versions of Eigen (e.g., 3.2.1), the scaling was omitted if the pivot
+      // was smaller than the cutoff value. However, since LDLT is not rank-revealing
+      // we should only make sure that we do not introduce INF or NaN values.
+      // Remark that LAPACK also uses 0 as the cutoff value.
+      RealScalar realAkk = numext::real(mat.coeffRef(k,k));
+      bool pivot_is_valid = (abs(realAkk) > RealScalar(0));
+
+      if(k==0 && !pivot_is_valid)
+      {
+        // The entire diagonal is zero, there is nothing more to do
+        // except filling the transpositions, and checking whether the matrix is zero.
+        sign = ZeroSign;
+        for(Index j = 0; j<size; ++j)
+        {
+          transpositions.coeffRef(j) = IndexType(j);
+          ret = ret && (mat.col(j).tail(size-j-1).array()==Scalar(0)).all();
+        }
+        return ret;
+      }
+
+      if((rs>0) && pivot_is_valid)
+        A21 /= realAkk;
+
+      if(found_zero_pivot && pivot_is_valid) ret = false; // factorization failed
+      else if(!pivot_is_valid) found_zero_pivot = true;
+
+      if (sign == PositiveSemiDef) {
+        if (realAkk < static_cast<RealScalar>(0)) sign = Indefinite;
+      } else if (sign == NegativeSemiDef) {
+        if (realAkk > static_cast<RealScalar>(0)) sign = Indefinite;
+      } else if (sign == ZeroSign) {
+        if (realAkk > static_cast<RealScalar>(0)) sign = PositiveSemiDef;
+        else if (realAkk < static_cast<RealScalar>(0)) sign = NegativeSemiDef;
+      }
+    }
+
+    return ret;
+  }
+
+  // Reference for the algorithm: Davis and Hager, "Multiple Rank
+  // Modifications of a Sparse Cholesky Factorization" (Algorithm 1)
+  // Trivial rearrangements of their computations (Timothy E. Holy)
+  // allow their algorithm to work for rank-1 updates even if the
+  // original matrix is not of full rank.
+  // Here only rank-1 updates are implemented, to reduce the
+  // requirement for intermediate storage and improve accuracy
+  template<typename MatrixType, typename WDerived>
+  static bool updateInPlace(MatrixType& mat, MatrixBase<WDerived>& w, const typename MatrixType::RealScalar& sigma=1)
+  {
+    using numext::isfinite;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+
+    const Index size = mat.rows();
+    eigen_assert(mat.cols() == size && w.size()==size);
+
+    RealScalar alpha = 1;
+
+    // Apply the update
+    for (Index j = 0; j < size; j++)
+    {
+      // Check for termination due to an original decomposition of low-rank
+      if (!(isfinite)(alpha))
+        break;
+
+      // Update the diagonal terms
+      RealScalar dj = numext::real(mat.coeff(j,j));
+      Scalar wj = w.coeff(j);
+      RealScalar swj2 = sigma*numext::abs2(wj);
+      RealScalar gamma = dj*alpha + swj2;
+
+      mat.coeffRef(j,j) += swj2/alpha;
+      alpha += swj2/dj;
+
+
+      // Update the terms of L
+      Index rs = size-j-1;
+      w.tail(rs) -= wj * mat.col(j).tail(rs);
+      if(gamma != 0)
+        mat.col(j).tail(rs) += (sigma*numext::conj(wj)/gamma)*w.tail(rs);
+    }
+    return true;
+  }
+
+  template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
+  static bool update(MatrixType& mat, const TranspositionType& transpositions, Workspace& tmp, const WType& w, const typename MatrixType::RealScalar& sigma=1)
+  {
+    // Apply the permutation to the input w
+    tmp = transpositions * w;
+
+    return ldlt_inplace<Lower>::updateInPlace(mat,tmp,sigma);
+  }
+};
+
+template<> struct ldlt_inplace<Upper>
+{
+  template<typename MatrixType, typename TranspositionType, typename Workspace>
+  static EIGEN_STRONG_INLINE bool unblocked(MatrixType& mat, TranspositionType& transpositions, Workspace& temp, SignMatrix& sign)
+  {
+    Transpose<MatrixType> matt(mat);
+    return ldlt_inplace<Lower>::unblocked(matt, transpositions, temp, sign);
+  }
+
+  template<typename MatrixType, typename TranspositionType, typename Workspace, typename WType>
+  static EIGEN_STRONG_INLINE bool update(MatrixType& mat, TranspositionType& transpositions, Workspace& tmp, WType& w, const typename MatrixType::RealScalar& sigma=1)
+  {
+    Transpose<MatrixType> matt(mat);
+    return ldlt_inplace<Lower>::update(matt, transpositions, tmp, w.conjugate(), sigma);
+  }
+};
+
+template<typename MatrixType> struct LDLT_Traits<MatrixType,Lower>
+{
+  typedef const TriangularView<const MatrixType, UnitLower> MatrixL;
+  typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitUpper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
+};
+
+template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
+{
+  typedef const TriangularView<const typename MatrixType::AdjointReturnType, UnitLower> MatrixL;
+  typedef const TriangularView<const MatrixType, UnitUpper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
+};
+
+} // end namespace internal
+
+/** Compute / recompute the LDLT decomposition A = L D L^* = U^* D U of \a matrix
+  */
+template<typename MatrixType, int _UpLo>
+template<typename InputType>
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>& a)
+{
+  check_template_parameters();
+
+  eigen_assert(a.rows()==a.cols());
+  const Index size = a.rows();
+
+  m_matrix = a.derived();
+
+  // Compute matrix L1 norm = max abs column sum.
+  m_l1_norm = RealScalar(0);
+  // TODO move this code to SelfAdjointView
+  for (Index col = 0; col < size; ++col) {
+    RealScalar abs_col_sum;
+    if (_UpLo == Lower)
+      abs_col_sum = m_matrix.col(col).tail(size - col).template lpNorm<1>() + m_matrix.row(col).head(col).template lpNorm<1>();
+    else
+      abs_col_sum = m_matrix.col(col).head(col).template lpNorm<1>() + m_matrix.row(col).tail(size - col).template lpNorm<1>();
+    if (abs_col_sum > m_l1_norm)
+      m_l1_norm = abs_col_sum;
+  }
+
+  m_transpositions.resize(size);
+  m_isInitialized = false;
+  m_temporary.resize(size);
+  m_sign = internal::ZeroSign;
+
+  m_info = internal::ldlt_inplace<UpLo>::unblocked(m_matrix, m_transpositions, m_temporary, m_sign) ? Success : NumericalIssue;
+
+  m_isInitialized = true;
+  return *this;
+}
+
+/** Update the LDLT decomposition:  given A = L D L^T, efficiently compute the decomposition of A + sigma w w^T.
+ * \param w a vector to be incorporated into the decomposition.
+ * \param sigma a scalar, +1 for updates and -1 for "downdates," which correspond to removing previously-added column vectors. Optional; default value is +1.
+ * \sa setZero()
+  */
+template<typename MatrixType, int _UpLo>
+template<typename Derived>
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename LDLT<MatrixType,_UpLo>::RealScalar& sigma)
+{
+  typedef typename TranspositionType::StorageIndex IndexType;
+  const Index size = w.rows();
+  if (m_isInitialized)
+  {
+    eigen_assert(m_matrix.rows()==size);
+  }
+  else
+  {
+    m_matrix.resize(size,size);
+    m_matrix.setZero();
+    m_transpositions.resize(size);
+    for (Index i = 0; i < size; i++)
+      m_transpositions.coeffRef(i) = IndexType(i);
+    m_temporary.resize(size);
+    m_sign = sigma>=0 ? internal::PositiveSemiDef : internal::NegativeSemiDef;
+    m_isInitialized = true;
+  }
+
+  internal::ldlt_inplace<UpLo>::update(m_matrix, m_transpositions, m_temporary, w, sigma);
+
+  return *this;
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType, int _UpLo>
+template<typename RhsType, typename DstType>
+void LDLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  eigen_assert(rhs.rows() == rows());
+  // dst = P b
+  dst = m_transpositions * rhs;
+
+  // dst = L^-1 (P b)
+  matrixL().solveInPlace(dst);
+
+  // dst = D^-1 (L^-1 P b)
+  // more precisely, use pseudo-inverse of D (see bug 241)
+  using std::abs;
+  const typename Diagonal<const MatrixType>::RealReturnType vecD(vectorD());
+  // In some previous versions, tolerance was set to the max of 1/highest and the maximal diagonal entry * epsilon
+  // as motivated by LAPACK's xGELSS:
+  // RealScalar tolerance = numext::maxi(vecD.array().abs().maxCoeff() * NumTraits<RealScalar>::epsilon(),RealScalar(1) / NumTraits<RealScalar>::highest());
+  // However, LDLT is not rank revealing, and so adjusting the tolerance wrt to the highest
+  // diagonal element is not well justified and leads to numerical issues in some cases.
+  // Moreover, Lapack's xSYTRS routines use 0 for the tolerance.
+  RealScalar tolerance = RealScalar(1) / NumTraits<RealScalar>::highest();
+
+  for (Index i = 0; i < vecD.size(); ++i)
+  {
+    if(abs(vecD(i)) > tolerance)
+      dst.row(i) /= vecD(i);
+    else
+      dst.row(i).setZero();
+  }
+
+  // dst = L^-T (D^-1 L^-1 P b)
+  matrixU().solveInPlace(dst);
+
+  // dst = P^-1 (L^-T D^-1 L^-1 P b) = A^-1 b
+  dst = m_transpositions.transpose() * dst;
+}
+#endif
+
+/** \internal use x = ldlt_object.solve(x);
+  *
+  * This is the \em in-place version of solve().
+  *
+  * \param bAndX represents both the right-hand side matrix b and result x.
+  *
+  * \returns true always! If you need to check for existence of solutions, use another decomposition like LU, QR, or SVD.
+  *
+  * This version avoids a copy when the right hand side matrix b is not
+  * needed anymore.
+  *
+  * \sa LDLT::solve(), MatrixBase::ldlt()
+  */
+template<typename MatrixType,int _UpLo>
+template<typename Derived>
+bool LDLT<MatrixType,_UpLo>::solveInPlace(MatrixBase<Derived> &bAndX) const
+{
+  eigen_assert(m_isInitialized && "LDLT is not initialized.");
+  eigen_assert(m_matrix.rows() == bAndX.rows());
+
+  bAndX = this->solve(bAndX);
+
+  return true;
+}
+
+/** \returns the matrix represented by the decomposition,
+ * i.e., it returns the product: P^T L D L^* P.
+ * This function is provided for debug purpose. */
+template<typename MatrixType, int _UpLo>
+MatrixType LDLT<MatrixType,_UpLo>::reconstructedMatrix() const
+{
+  eigen_assert(m_isInitialized && "LDLT is not initialized.");
+  const Index size = m_matrix.rows();
+  MatrixType res(size,size);
+
+  // P
+  res.setIdentity();
+  res = transpositionsP() * res;
+  // L^* P
+  res = matrixU() * res;
+  // D(L^*P)
+  res = vectorD().real().asDiagonal() * res;
+  // L(DL^*P)
+  res = matrixL() * res;
+  // P^T (LDL^*P)
+  res = transpositionsP().transpose() * res;
+
+  return res;
+}
+
+/** \cholesky_module
+  * \returns the Cholesky decomposition with full pivoting without square root of \c *this
+  * \sa MatrixBase::ldlt()
+  */
+template<typename MatrixType, unsigned int UpLo>
+inline const LDLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
+SelfAdjointView<MatrixType, UpLo>::ldlt() const
+{
+  return LDLT<PlainObject,UpLo>(m_matrix);
+}
+
+/** \cholesky_module
+  * \returns the Cholesky decomposition with full pivoting without square root of \c *this
+  * \sa SelfAdjointView::ldlt()
+  */
+template<typename Derived>
+inline const LDLT<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::ldlt() const
+{
+  return LDLT<PlainObject>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_LDLT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT.h
new file mode 100644
index 0000000..87ca8d4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT.h
@@ -0,0 +1,534 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LLT_H
+#define EIGEN_LLT_H
+
+namespace Eigen {
+
+namespace internal{
+template<typename MatrixType, int UpLo> struct LLT_Traits;
+}
+
+/** \ingroup Cholesky_Module
+  *
+  * \class LLT
+  *
+  * \brief Standard Cholesky decomposition (LL^T) of a matrix and associated features
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the LL^T Cholesky decomposition
+  * \tparam _UpLo the triangular part that will be used for the decompositon: Lower (default) or Upper.
+  *             The other triangular part won't be read.
+  *
+  * This class performs a LL^T Cholesky decomposition of a symmetric, positive definite
+  * matrix A such that A = LL^* = U^*U, where L is lower triangular.
+  *
+  * While the Cholesky decomposition is particularly useful to solve selfadjoint problems like  D^*D x = b,
+  * for that purpose, we recommend the Cholesky decomposition without square root which is more stable
+  * and even faster. Nevertheless, this standard Cholesky decomposition remains useful in many other
+  * situations like generalised eigen problems with hermitian matrices.
+  *
+  * Remember that Cholesky decompositions are not rank-revealing. This LLT decomposition is only stable on positive definite matrices,
+  * use LDLT instead for the semidefinite case. Also, do not use a Cholesky decomposition to determine whether a system of equations
+  * has a solution.
+  *
+  * Example: \include LLT_example.cpp
+  * Output: \verbinclude LLT_example.out
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  *
+  * \sa MatrixBase::llt(), SelfAdjointView::llt(), class LDLT
+  */
+ /* HEY THIS DOX IS DISABLED BECAUSE THERE's A BUG EITHER HERE OR IN LDLT ABOUT THAT (OR BOTH)
+  * Note that during the decomposition, only the upper triangular part of A is considered. Therefore,
+  * the strict lower part does not have to store correct values.
+  */
+template<typename _MatrixType, int _UpLo> class LLT
+{
+  public:
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename MatrixType::StorageIndex StorageIndex;
+
+    enum {
+      PacketSize = internal::packet_traits<Scalar>::size,
+      AlignmentMask = int(PacketSize)-1,
+      UpLo = _UpLo
+    };
+
+    typedef internal::LLT_Traits<MatrixType,UpLo> Traits;
+
+    /**
+      * \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via LLT::compute(const MatrixType&).
+      */
+    LLT() : m_matrix(), m_isInitialized(false) {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa LLT()
+      */
+    explicit LLT(Index size) : m_matrix(size, size),
+                    m_isInitialized(false) {}
+
+    template<typename InputType>
+    explicit LLT(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.rows(), matrix.cols()),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** \brief Constructs a LDLT factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when
+      * \c MatrixType is a Eigen::Ref.
+      *
+      * \sa LLT(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit LLT(EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** \returns a view of the upper triangular matrix U */
+    inline typename Traits::MatrixU matrixU() const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      return Traits::getU(m_matrix);
+    }
+
+    /** \returns a view of the lower triangular matrix L */
+    inline typename Traits::MatrixL matrixL() const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      return Traits::getL(m_matrix);
+    }
+
+    /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      *
+      * Since this LLT class assumes anyway that the matrix A is invertible, the solution
+      * theoretically exists and is unique regardless of b.
+      *
+      * Example: \include LLT_solve.cpp
+      * Output: \verbinclude LLT_solve.out
+      *
+      * \sa solveInPlace(), MatrixBase::llt(), SelfAdjointView::llt()
+      */
+    template<typename Rhs>
+    inline const Solve<LLT, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      eigen_assert(m_matrix.rows()==b.rows()
+                && "LLT::solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<LLT, Rhs>(*this, b.derived());
+    }
+
+    template<typename Derived>
+    void solveInPlace(MatrixBase<Derived> &bAndX) const;
+
+    template<typename InputType>
+    LLT& compute(const EigenBase<InputType>& matrix);
+
+    /** \returns an estimate of the reciprocal condition number of the matrix of
+      *  which \c *this is the Cholesky decomposition.
+      */
+    RealScalar rcond() const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      eigen_assert(m_info == Success && "LLT failed because matrix appears to be negative");
+      return internal::rcond_estimate_helper(m_l1_norm, *this);
+    }
+
+    /** \returns the LLT decomposition matrix
+      *
+      * TODO: document the storage layout
+      */
+    inline const MatrixType& matrixLLT() const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      return m_matrix;
+    }
+
+    MatrixType reconstructedMatrix() const;
+
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "LLT is not initialized.");
+      return m_info;
+    }
+
+    /** \returns the adjoint of \c *this, that is, a const reference to the decomposition itself as the underlying matrix is self-adjoint.
+      *
+      * This method is provided for compatibility with other matrix decompositions, thus enabling generic code such as:
+      * \code x = decomposition.adjoint().solve(b) \endcode
+      */
+    const LLT& adjoint() const { return *this; };
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+
+    template<typename VectorType>
+    LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    /** \internal
+      * Used to compute and store L
+      * The strict upper part is not used and even not initialized.
+      */
+    MatrixType m_matrix;
+    RealScalar m_l1_norm;
+    bool m_isInitialized;
+    ComputationInfo m_info;
+};
+
+namespace internal {
+
+template<typename Scalar, int UpLo> struct llt_inplace;
+
+template<typename MatrixType, typename VectorType>
+static Index llt_rank_update_lower(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma)
+{
+  using std::sqrt;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef typename MatrixType::ColXpr ColXpr;
+  typedef typename internal::remove_all<ColXpr>::type ColXprCleaned;
+  typedef typename ColXprCleaned::SegmentReturnType ColXprSegment;
+  typedef Matrix<Scalar,Dynamic,1> TempVectorType;
+  typedef typename TempVectorType::SegmentReturnType TempVecSegment;
+
+  Index n = mat.cols();
+  eigen_assert(mat.rows()==n && vec.size()==n);
+
+  TempVectorType temp;
+
+  if(sigma>0)
+  {
+    // This version is based on Givens rotations.
+    // It is faster than the other one below, but only works for updates,
+    // i.e., for sigma > 0
+    temp = sqrt(sigma) * vec;
+
+    for(Index i=0; i<n; ++i)
+    {
+      JacobiRotation<Scalar> g;
+      g.makeGivens(mat(i,i), -temp(i), &mat(i,i));
+
+      Index rs = n-i-1;
+      if(rs>0)
+      {
+        ColXprSegment x(mat.col(i).tail(rs));
+        TempVecSegment y(temp.tail(rs));
+        apply_rotation_in_the_plane(x, y, g);
+      }
+    }
+  }
+  else
+  {
+    temp = vec;
+    RealScalar beta = 1;
+    for(Index j=0; j<n; ++j)
+    {
+      RealScalar Ljj = numext::real(mat.coeff(j,j));
+      RealScalar dj = numext::abs2(Ljj);
+      Scalar wj = temp.coeff(j);
+      RealScalar swj2 = sigma*numext::abs2(wj);
+      RealScalar gamma = dj*beta + swj2;
+
+      RealScalar x = dj + swj2/beta;
+      if (x<=RealScalar(0))
+        return j;
+      RealScalar nLjj = sqrt(x);
+      mat.coeffRef(j,j) = nLjj;
+      beta += swj2/dj;
+
+      // Update the terms of L
+      Index rs = n-j-1;
+      if(rs)
+      {
+        temp.tail(rs) -= (wj/Ljj) * mat.col(j).tail(rs);
+        if(gamma != 0)
+          mat.col(j).tail(rs) = (nLjj/Ljj) * mat.col(j).tail(rs) + (nLjj * sigma*numext::conj(wj)/gamma)*temp.tail(rs);
+      }
+    }
+  }
+  return -1;
+}
+
+template<typename Scalar> struct llt_inplace<Scalar, Lower>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  template<typename MatrixType>
+  static Index unblocked(MatrixType& mat)
+  {
+    using std::sqrt;
+
+    eigen_assert(mat.rows()==mat.cols());
+    const Index size = mat.rows();
+    for(Index k = 0; k < size; ++k)
+    {
+      Index rs = size-k-1; // remaining size
+
+      Block<MatrixType,Dynamic,1> A21(mat,k+1,k,rs,1);
+      Block<MatrixType,1,Dynamic> A10(mat,k,0,1,k);
+      Block<MatrixType,Dynamic,Dynamic> A20(mat,k+1,0,rs,k);
+
+      RealScalar x = numext::real(mat.coeff(k,k));
+      if (k>0) x -= A10.squaredNorm();
+      if (x<=RealScalar(0))
+        return k;
+      mat.coeffRef(k,k) = x = sqrt(x);
+      if (k>0 && rs>0) A21.noalias() -= A20 * A10.adjoint();
+      if (rs>0) A21 /= x;
+    }
+    return -1;
+  }
+
+  template<typename MatrixType>
+  static Index blocked(MatrixType& m)
+  {
+    eigen_assert(m.rows()==m.cols());
+    Index size = m.rows();
+    if(size<32)
+      return unblocked(m);
+
+    Index blockSize = size/8;
+    blockSize = (blockSize/16)*16;
+    blockSize = (std::min)((std::max)(blockSize,Index(8)), Index(128));
+
+    for (Index k=0; k<size; k+=blockSize)
+    {
+      // partition the matrix:
+      //       A00 |  -  |  -
+      // lu  = A10 | A11 |  -
+      //       A20 | A21 | A22
+      Index bs = (std::min)(blockSize, size-k);
+      Index rs = size - k - bs;
+      Block<MatrixType,Dynamic,Dynamic> A11(m,k,   k,   bs,bs);
+      Block<MatrixType,Dynamic,Dynamic> A21(m,k+bs,k,   rs,bs);
+      Block<MatrixType,Dynamic,Dynamic> A22(m,k+bs,k+bs,rs,rs);
+
+      Index ret;
+      if((ret=unblocked(A11))>=0) return k+ret;
+      if(rs>0) A11.adjoint().template triangularView<Upper>().template solveInPlace<OnTheRight>(A21);
+      if(rs>0) A22.template selfadjointView<Lower>().rankUpdate(A21,typename NumTraits<RealScalar>::Literal(-1)); // bottleneck
+    }
+    return -1;
+  }
+
+  template<typename MatrixType, typename VectorType>
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  {
+    return Eigen::internal::llt_rank_update_lower(mat, vec, sigma);
+  }
+};
+
+template<typename Scalar> struct llt_inplace<Scalar, Upper>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+
+  template<typename MatrixType>
+  static EIGEN_STRONG_INLINE Index unblocked(MatrixType& mat)
+  {
+    Transpose<MatrixType> matt(mat);
+    return llt_inplace<Scalar, Lower>::unblocked(matt);
+  }
+  template<typename MatrixType>
+  static EIGEN_STRONG_INLINE Index blocked(MatrixType& mat)
+  {
+    Transpose<MatrixType> matt(mat);
+    return llt_inplace<Scalar, Lower>::blocked(matt);
+  }
+  template<typename MatrixType, typename VectorType>
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const RealScalar& sigma)
+  {
+    Transpose<MatrixType> matt(mat);
+    return llt_inplace<Scalar, Lower>::rankUpdate(matt, vec.conjugate(), sigma);
+  }
+};
+
+template<typename MatrixType> struct LLT_Traits<MatrixType,Lower>
+{
+  typedef const TriangularView<const MatrixType, Lower> MatrixL;
+  typedef const TriangularView<const typename MatrixType::AdjointReturnType, Upper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
+  static bool inplace_decomposition(MatrixType& m)
+  { return llt_inplace<typename MatrixType::Scalar, Lower>::blocked(m)==-1; }
+};
+
+template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
+{
+  typedef const TriangularView<const typename MatrixType::AdjointReturnType, Lower> MatrixL;
+  typedef const TriangularView<const MatrixType, Upper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m.adjoint()); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m); }
+  static bool inplace_decomposition(MatrixType& m)
+  { return llt_inplace<typename MatrixType::Scalar, Upper>::blocked(m)==-1; }
+};
+
+} // end namespace internal
+
+/** Computes / recomputes the Cholesky decomposition A = LL^* = U^*U of \a matrix
+  *
+  * \returns a reference to *this
+  *
+  * Example: \include TutorialLinAlgComputeTwice.cpp
+  * Output: \verbinclude TutorialLinAlgComputeTwice.out
+  */
+template<typename MatrixType, int _UpLo>
+template<typename InputType>
+LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const EigenBase<InputType>& a)
+{
+  check_template_parameters();
+
+  eigen_assert(a.rows()==a.cols());
+  const Index size = a.rows();
+  m_matrix.resize(size, size);
+  m_matrix = a.derived();
+
+  // Compute matrix L1 norm = max abs column sum.
+  m_l1_norm = RealScalar(0);
+  // TODO move this code to SelfAdjointView
+  for (Index col = 0; col < size; ++col) {
+    RealScalar abs_col_sum;
+    if (_UpLo == Lower)
+      abs_col_sum = m_matrix.col(col).tail(size - col).template lpNorm<1>() + m_matrix.row(col).head(col).template lpNorm<1>();
+    else
+      abs_col_sum = m_matrix.col(col).head(col).template lpNorm<1>() + m_matrix.row(col).tail(size - col).template lpNorm<1>();
+    if (abs_col_sum > m_l1_norm)
+      m_l1_norm = abs_col_sum;
+  }
+
+  m_isInitialized = true;
+  bool ok = Traits::inplace_decomposition(m_matrix);
+  m_info = ok ? Success : NumericalIssue;
+
+  return *this;
+}
+
+/** Performs a rank one update (or dowdate) of the current decomposition.
+  * If A = LL^* before the rank one update,
+  * then after it we have LL^* = A + sigma * v v^* where \a v must be a vector
+  * of same dimension.
+  */
+template<typename _MatrixType, int _UpLo>
+template<typename VectorType>
+LLT<_MatrixType,_UpLo> LLT<_MatrixType,_UpLo>::rankUpdate(const VectorType& v, const RealScalar& sigma)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorType);
+  eigen_assert(v.size()==m_matrix.cols());
+  eigen_assert(m_isInitialized);
+  if(internal::llt_inplace<typename MatrixType::Scalar, UpLo>::rankUpdate(m_matrix,v,sigma)>=0)
+    m_info = NumericalIssue;
+  else
+    m_info = Success;
+
+  return *this;
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType,int _UpLo>
+template<typename RhsType, typename DstType>
+void LLT<_MatrixType,_UpLo>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  dst = rhs;
+  solveInPlace(dst);
+}
+#endif
+
+/** \internal use x = llt_object.solve(x);
+  *
+  * This is the \em in-place version of solve().
+  *
+  * \param bAndX represents both the right-hand side matrix b and result x.
+  *
+  * This version avoids a copy when the right hand side matrix b is not needed anymore.
+  *
+  * \sa LLT::solve(), MatrixBase::llt()
+  */
+template<typename MatrixType, int _UpLo>
+template<typename Derived>
+void LLT<MatrixType,_UpLo>::solveInPlace(MatrixBase<Derived> &bAndX) const
+{
+  eigen_assert(m_isInitialized && "LLT is not initialized.");
+  eigen_assert(m_matrix.rows()==bAndX.rows());
+  matrixL().solveInPlace(bAndX);
+  matrixU().solveInPlace(bAndX);
+}
+
+/** \returns the matrix represented by the decomposition,
+ * i.e., it returns the product: L L^*.
+ * This function is provided for debug purpose. */
+template<typename MatrixType, int _UpLo>
+MatrixType LLT<MatrixType,_UpLo>::reconstructedMatrix() const
+{
+  eigen_assert(m_isInitialized && "LLT is not initialized.");
+  return matrixL() * matrixL().adjoint().toDenseMatrix();
+}
+
+/** \cholesky_module
+  * \returns the LLT decomposition of \c *this
+  * \sa SelfAdjointView::llt()
+  */
+template<typename Derived>
+inline const LLT<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::llt() const
+{
+  return LLT<PlainObject>(derived());
+}
+
+/** \cholesky_module
+  * \returns the LLT decomposition of \c *this
+  * \sa SelfAdjointView::llt()
+  */
+template<typename MatrixType, unsigned int UpLo>
+inline const LLT<typename SelfAdjointView<MatrixType, UpLo>::PlainObject, UpLo>
+SelfAdjointView<MatrixType, UpLo>::llt() const
+{
+  return LLT<PlainObject,UpLo>(m_matrix);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_LLT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT_LAPACKE.h
new file mode 100644
index 0000000..bc6489e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Cholesky/LLT_LAPACKE.h
@@ -0,0 +1,99 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *     LLt decomposition based on LAPACKE_?potrf function.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_LLT_LAPACKE_H
+#define EIGEN_LLT_LAPACKE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Scalar> struct lapacke_llt;
+
+#define EIGEN_LAPACKE_LLT(EIGTYPE, BLASTYPE, LAPACKE_PREFIX) \
+template<> struct lapacke_llt<EIGTYPE> \
+{ \
+  template<typename MatrixType> \
+  static inline Index potrf(MatrixType& m, char uplo) \
+  { \
+    lapack_int matrix_order; \
+    lapack_int size, lda, info, StorageOrder; \
+    EIGTYPE* a; \
+    eigen_assert(m.rows()==m.cols()); \
+    /* Set up parameters for ?potrf */ \
+    size = convert_index<lapack_int>(m.rows()); \
+    StorageOrder = MatrixType::Flags&RowMajorBit?RowMajor:ColMajor; \
+    matrix_order = StorageOrder==RowMajor ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+    a = &(m.coeffRef(0,0)); \
+    lda = convert_index<lapack_int>(m.outerStride()); \
+\
+    info = LAPACKE_##LAPACKE_PREFIX##potrf( matrix_order, uplo, size, (BLASTYPE*)a, lda ); \
+    info = (info==0) ? -1 : info>0 ? info-1 : size; \
+    return info; \
+  } \
+}; \
+template<> struct llt_inplace<EIGTYPE, Lower> \
+{ \
+  template<typename MatrixType> \
+  static Index blocked(MatrixType& m) \
+  { \
+    return lapacke_llt<EIGTYPE>::potrf(m, 'L'); \
+  } \
+  template<typename MatrixType, typename VectorType> \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  { return Eigen::internal::llt_rank_update_lower(mat, vec, sigma); } \
+}; \
+template<> struct llt_inplace<EIGTYPE, Upper> \
+{ \
+  template<typename MatrixType> \
+  static Index blocked(MatrixType& m) \
+  { \
+    return lapacke_llt<EIGTYPE>::potrf(m, 'U'); \
+  } \
+  template<typename MatrixType, typename VectorType> \
+  static Index rankUpdate(MatrixType& mat, const VectorType& vec, const typename MatrixType::RealScalar& sigma) \
+  { \
+    Transpose<MatrixType> matt(mat); \
+    return llt_inplace<EIGTYPE, Lower>::rankUpdate(matt, vec.conjugate(), sigma); \
+  } \
+};
+
+EIGEN_LAPACKE_LLT(double, double, d)
+EIGEN_LAPACKE_LLT(float, float, s)
+EIGEN_LAPACKE_LLT(dcomplex, lapack_complex_double, z)
+EIGEN_LAPACKE_LLT(scomplex, lapack_complex_float, c)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_LLT_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/CholmodSupport/CholmodSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/CholmodSupport/CholmodSupport.h
new file mode 100644
index 0000000..5719720
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/CholmodSupport/CholmodSupport.h
@@ -0,0 +1,639 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CHOLMODSUPPORT_H
+#define EIGEN_CHOLMODSUPPORT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Scalar> struct cholmod_configure_matrix;
+
+template<> struct cholmod_configure_matrix<double> {
+  template<typename CholmodType>
+  static void run(CholmodType& mat) {
+    mat.xtype = CHOLMOD_REAL;
+    mat.dtype = CHOLMOD_DOUBLE;
+  }
+};
+
+template<> struct cholmod_configure_matrix<std::complex<double> > {
+  template<typename CholmodType>
+  static void run(CholmodType& mat) {
+    mat.xtype = CHOLMOD_COMPLEX;
+    mat.dtype = CHOLMOD_DOUBLE;
+  }
+};
+
+// Other scalar types are not yet suppotred by Cholmod
+// template<> struct cholmod_configure_matrix<float> {
+//   template<typename CholmodType>
+//   static void run(CholmodType& mat) {
+//     mat.xtype = CHOLMOD_REAL;
+//     mat.dtype = CHOLMOD_SINGLE;
+//   }
+// };
+//
+// template<> struct cholmod_configure_matrix<std::complex<float> > {
+//   template<typename CholmodType>
+//   static void run(CholmodType& mat) {
+//     mat.xtype = CHOLMOD_COMPLEX;
+//     mat.dtype = CHOLMOD_SINGLE;
+//   }
+// };
+
+} // namespace internal
+
+/** Wraps the Eigen sparse matrix \a mat into a Cholmod sparse matrix object.
+  * Note that the data are shared.
+  */
+template<typename _Scalar, int _Options, typename _StorageIndex>
+cholmod_sparse viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_StorageIndex> > mat)
+{
+  cholmod_sparse res;
+  res.nzmax   = mat.nonZeros();
+  res.nrow    = mat.rows();
+  res.ncol    = mat.cols();
+  res.p       = mat.outerIndexPtr();
+  res.i       = mat.innerIndexPtr();
+  res.x       = mat.valuePtr();
+  res.z       = 0;
+  res.sorted  = 1;
+  if(mat.isCompressed())
+  {
+    res.packed  = 1;
+    res.nz = 0;
+  }
+  else
+  {
+    res.packed  = 0;
+    res.nz = mat.innerNonZeroPtr();
+  }
+
+  res.dtype   = 0;
+  res.stype   = -1;
+  
+  if (internal::is_same<_StorageIndex,int>::value)
+  {
+    res.itype = CHOLMOD_INT;
+  }
+  else if (internal::is_same<_StorageIndex,long>::value)
+  {
+    res.itype = CHOLMOD_LONG;
+  }
+  else
+  {
+    eigen_assert(false && "Index type not supported yet");
+  }
+
+  // setup res.xtype
+  internal::cholmod_configure_matrix<_Scalar>::run(res);
+  
+  res.stype = 0;
+  
+  return res;
+}
+
+template<typename _Scalar, int _Options, typename _Index>
+const cholmod_sparse viewAsCholmod(const SparseMatrix<_Scalar,_Options,_Index>& mat)
+{
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.const_cast_derived()));
+  return res;
+}
+
+template<typename _Scalar, int _Options, typename _Index>
+const cholmod_sparse viewAsCholmod(const SparseVector<_Scalar,_Options,_Index>& mat)
+{
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.const_cast_derived()));
+  return res;
+}
+
+/** Returns a view of the Eigen sparse matrix \a mat as Cholmod sparse matrix.
+  * The data are not copied but shared. */
+template<typename _Scalar, int _Options, typename _Index, unsigned int UpLo>
+cholmod_sparse viewAsCholmod(const SparseSelfAdjointView<const SparseMatrix<_Scalar,_Options,_Index>, UpLo>& mat)
+{
+  cholmod_sparse res = viewAsCholmod(Ref<SparseMatrix<_Scalar,_Options,_Index> >(mat.matrix().const_cast_derived()));
+  
+  if(UpLo==Upper) res.stype =  1;
+  if(UpLo==Lower) res.stype = -1;
+
+  return res;
+}
+
+/** Returns a view of the Eigen \b dense matrix \a mat as Cholmod dense matrix.
+  * The data are not copied but shared. */
+template<typename Derived>
+cholmod_dense viewAsCholmod(MatrixBase<Derived>& mat)
+{
+  EIGEN_STATIC_ASSERT((internal::traits<Derived>::Flags&RowMajorBit)==0,THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+  typedef typename Derived::Scalar Scalar;
+
+  cholmod_dense res;
+  res.nrow   = mat.rows();
+  res.ncol   = mat.cols();
+  res.nzmax  = res.nrow * res.ncol;
+  res.d      = Derived::IsVectorAtCompileTime ? mat.derived().size() : mat.derived().outerStride();
+  res.x      = (void*)(mat.derived().data());
+  res.z      = 0;
+
+  internal::cholmod_configure_matrix<Scalar>::run(res);
+
+  return res;
+}
+
+/** Returns a view of the Cholmod sparse matrix \a cm as an Eigen sparse matrix.
+  * The data are not copied but shared. */
+template<typename Scalar, int Flags, typename StorageIndex>
+MappedSparseMatrix<Scalar,Flags,StorageIndex> viewAsEigen(cholmod_sparse& cm)
+{
+  return MappedSparseMatrix<Scalar,Flags,StorageIndex>
+         (cm.nrow, cm.ncol, static_cast<StorageIndex*>(cm.p)[cm.ncol],
+          static_cast<StorageIndex*>(cm.p), static_cast<StorageIndex*>(cm.i),static_cast<Scalar*>(cm.x) );
+}
+
+enum CholmodMode {
+  CholmodAuto, CholmodSimplicialLLt, CholmodSupernodalLLt, CholmodLDLt
+};
+
+
+/** \ingroup CholmodSupport_Module
+  * \class CholmodBase
+  * \brief The base class for the direct Cholesky factorization of Cholmod
+  * \sa class CholmodSupernodalLLT, class CholmodSimplicialLDLT, class CholmodSimplicialLLT
+  */
+template<typename _MatrixType, int _UpLo, typename Derived>
+class CholmodBase : public SparseSolverBase<Derived>
+{
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
+  public:
+    typedef _MatrixType MatrixType;
+    enum { UpLo = _UpLo };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef MatrixType CholMatrixType;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+  public:
+
+    CholmodBase()
+      : m_cholmodFactor(0), m_info(Success), m_factorizationIsOk(false), m_analysisIsOk(false)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<double,RealScalar>::value), CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY);
+      m_shiftOffset[0] = m_shiftOffset[1] = 0.0;
+      cholmod_start(&m_cholmod);
+    }
+
+    explicit CholmodBase(const MatrixType& matrix)
+      : m_cholmodFactor(0), m_info(Success), m_factorizationIsOk(false), m_analysisIsOk(false)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<double,RealScalar>::value), CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY);
+      m_shiftOffset[0] = m_shiftOffset[1] = 0.0;
+      cholmod_start(&m_cholmod);
+      compute(matrix);
+    }
+
+    ~CholmodBase()
+    {
+      if(m_cholmodFactor)
+        cholmod_free_factor(&m_cholmodFactor, &m_cholmod);
+      cholmod_finish(&m_cholmod);
+    }
+    
+    inline StorageIndex cols() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
+    inline StorageIndex rows() const { return internal::convert_index<StorageIndex, Index>(m_cholmodFactor->n); }
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    Derived& compute(const MatrixType& matrix)
+    {
+      analyzePattern(matrix);
+      factorize(matrix);
+      return derived();
+    }
+    
+    /** Performs a symbolic decomposition on the sparsity pattern of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      * 
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    {
+      if(m_cholmodFactor)
+      {
+        cholmod_free_factor(&m_cholmodFactor, &m_cholmod);
+        m_cholmodFactor = 0;
+      }
+      cholmod_sparse A = viewAsCholmod(matrix.template selfadjointView<UpLo>());
+      m_cholmodFactor = cholmod_analyze(&A, &m_cholmod);
+      
+      this->m_isInitialized = true;
+      this->m_info = Success;
+      m_analysisIsOk = true;
+      m_factorizationIsOk = false;
+    }
+    
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must have the same sparsity pattern as the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& matrix)
+    {
+      eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+      cholmod_sparse A = viewAsCholmod(matrix.template selfadjointView<UpLo>());
+      cholmod_factorize_p(&A, m_shiftOffset, 0, 0, m_cholmodFactor, &m_cholmod);
+
+      // If the factorization failed, minor is the column at which it did. On success minor == n.
+      this->m_info = (m_cholmodFactor->minor == m_cholmodFactor->n ? Success : NumericalIssue);
+      m_factorizationIsOk = true;
+    }
+    
+    /** Returns a reference to the Cholmod's configuration structure to get a full control over the performed operations.
+     *  See the Cholmod user guide for details. */
+    cholmod_common& cholmod() { return m_cholmod; }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    {
+      eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
+      const Index size = m_cholmodFactor->n;
+      EIGEN_UNUSED_VARIABLE(size);
+      eigen_assert(size==b.rows());
+      
+      // Cholmod needs column-major stoarge without inner-stride, which corresponds to the default behavior of Ref.
+      Ref<const Matrix<typename Rhs::Scalar,Dynamic,Dynamic,ColMajor> > b_ref(b.derived());
+
+      cholmod_dense b_cd = viewAsCholmod(b_ref);
+      cholmod_dense* x_cd = cholmod_solve(CHOLMOD_A, m_cholmodFactor, &b_cd, &m_cholmod);
+      if(!x_cd)
+      {
+        this->m_info = NumericalIssue;
+        return;
+      }
+      // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
+      dest = Matrix<Scalar,Dest::RowsAtCompileTime,Dest::ColsAtCompileTime>::Map(reinterpret_cast<Scalar*>(x_cd->x),b.rows(),b.cols());
+      cholmod_free_dense(&x_cd, &m_cholmod);
+    }
+    
+    /** \internal */
+    template<typename RhsDerived, typename DestDerived>
+    void _solve_impl(const SparseMatrixBase<RhsDerived> &b, SparseMatrixBase<DestDerived> &dest) const
+    {
+      eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
+      const Index size = m_cholmodFactor->n;
+      EIGEN_UNUSED_VARIABLE(size);
+      eigen_assert(size==b.rows());
+
+      // note: cs stands for Cholmod Sparse
+      Ref<SparseMatrix<typename RhsDerived::Scalar,ColMajor,typename RhsDerived::StorageIndex> > b_ref(b.const_cast_derived());
+      cholmod_sparse b_cs = viewAsCholmod(b_ref);
+      cholmod_sparse* x_cs = cholmod_spsolve(CHOLMOD_A, m_cholmodFactor, &b_cs, &m_cholmod);
+      if(!x_cs)
+      {
+        this->m_info = NumericalIssue;
+        return;
+      }
+      // TODO optimize this copy by swapping when possible (be careful with alignment, etc.)
+      dest.derived() = viewAsEigen<typename DestDerived::Scalar,ColMajor,typename DestDerived::StorageIndex>(*x_cs);
+      cholmod_free_sparse(&x_cs, &m_cholmod);
+    }
+    #endif // EIGEN_PARSED_BY_DOXYGEN
+    
+    
+    /** Sets the shift parameter that will be used to adjust the diagonal coefficients during the numerical factorization.
+      *
+      * During the numerical factorization, an offset term is added to the diagonal coefficients:\n
+      * \c d_ii = \a offset + \c d_ii
+      *
+      * The default is \a offset=0.
+      *
+      * \returns a reference to \c *this.
+      */
+    Derived& setShift(const RealScalar& offset)
+    {
+      m_shiftOffset[0] = double(offset);
+      return derived();
+    }
+    
+    /** \returns the determinant of the underlying matrix from the current factorization */
+    Scalar determinant() const
+    {
+      using std::exp;
+      return exp(logDeterminant());
+    }
+
+    /** \returns the log determinant of the underlying matrix from the current factorization */
+    Scalar logDeterminant() const
+    {
+      using std::log;
+      using numext::real;
+      eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
+
+      RealScalar logDet = 0;
+      Scalar *x = static_cast<Scalar*>(m_cholmodFactor->x);
+      if (m_cholmodFactor->is_super)
+      {
+        // Supernodal factorization stored as a packed list of dense column-major blocs,
+        // as described by the following structure:
+
+        // super[k] == index of the first column of the j-th super node
+        StorageIndex *super = static_cast<StorageIndex*>(m_cholmodFactor->super);
+        // pi[k] == offset to the description of row indices
+        StorageIndex *pi = static_cast<StorageIndex*>(m_cholmodFactor->pi);
+        // px[k] == offset to the respective dense block
+        StorageIndex *px = static_cast<StorageIndex*>(m_cholmodFactor->px);
+
+        Index nb_super_nodes = m_cholmodFactor->nsuper;
+        for (Index k=0; k < nb_super_nodes; ++k)
+        {
+          StorageIndex ncols = super[k + 1] - super[k];
+          StorageIndex nrows = pi[k + 1] - pi[k];
+
+          Map<const Array<Scalar,1,Dynamic>, 0, InnerStride<> > sk(x + px[k], ncols, InnerStride<>(nrows+1));
+          logDet += sk.real().log().sum();
+        }
+      }
+      else
+      {
+        // Simplicial factorization stored as standard CSC matrix.
+        StorageIndex *p = static_cast<StorageIndex*>(m_cholmodFactor->p);
+        Index size = m_cholmodFactor->n;
+        for (Index k=0; k<size; ++k)
+          logDet += log(real( x[p[k]] ));
+      }
+      if (m_cholmodFactor->is_ll)
+        logDet *= 2.0;
+      return logDet;
+    };
+
+    template<typename Stream>
+    void dumpMemory(Stream& /*s*/)
+    {}
+    
+  protected:
+    mutable cholmod_common m_cholmod;
+    cholmod_factor* m_cholmodFactor;
+    double m_shiftOffset[2];
+    mutable ComputationInfo m_info;
+    int m_factorizationIsOk;
+    int m_analysisIsOk;
+};
+
+/** \ingroup CholmodSupport_Module
+  * \class CholmodSimplicialLLT
+  * \brief A simplicial direct Cholesky (LLT) factorization and solver based on Cholmod
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a simplicial LL^T Cholesky factorization
+  * using the Cholmod library.
+  * This simplicial variant is equivalent to Eigen's built-in SimplicialLLT class. Therefore, it has little practical interest.
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
+  * X and B can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  *
+  * \implsparsesolverconcept
+  *
+  * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
+  *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
+  * \sa \ref TutorialSparseSolverConcept, class CholmodSupernodalLLT, class SimplicialLLT
+  */
+template<typename _MatrixType, int _UpLo = Lower>
+class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLLT<_MatrixType, _UpLo> >
+{
+    typedef CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLLT> Base;
+    using Base::m_cholmod;
+    
+  public:
+    
+    typedef _MatrixType MatrixType;
+    
+    CholmodSimplicialLLT() : Base() { init(); }
+
+    CholmodSimplicialLLT(const MatrixType& matrix) : Base()
+    {
+      init();
+      this->compute(matrix);
+    }
+
+    ~CholmodSimplicialLLT() {}
+  protected:
+    void init()
+    {
+      m_cholmod.final_asis = 0;
+      m_cholmod.supernodal = CHOLMOD_SIMPLICIAL;
+      m_cholmod.final_ll = 1;
+    }
+};
+
+
+/** \ingroup CholmodSupport_Module
+  * \class CholmodSimplicialLDLT
+  * \brief A simplicial direct Cholesky (LDLT) factorization and solver based on Cholmod
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a simplicial LDL^T Cholesky factorization
+  * using the Cholmod library.
+  * This simplicial variant is equivalent to Eigen's built-in SimplicialLDLT class. Therefore, it has little practical interest.
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
+  * X and B can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  *
+  * \implsparsesolverconcept
+  *
+  * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
+  *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
+  * \sa \ref TutorialSparseSolverConcept, class CholmodSupernodalLLT, class SimplicialLDLT
+  */
+template<typename _MatrixType, int _UpLo = Lower>
+class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLDLT<_MatrixType, _UpLo> >
+{
+    typedef CholmodBase<_MatrixType, _UpLo, CholmodSimplicialLDLT> Base;
+    using Base::m_cholmod;
+    
+  public:
+    
+    typedef _MatrixType MatrixType;
+    
+    CholmodSimplicialLDLT() : Base() { init(); }
+
+    CholmodSimplicialLDLT(const MatrixType& matrix) : Base()
+    {
+      init();
+      this->compute(matrix);
+    }
+
+    ~CholmodSimplicialLDLT() {}
+  protected:
+    void init()
+    {
+      m_cholmod.final_asis = 1;
+      m_cholmod.supernodal = CHOLMOD_SIMPLICIAL;
+    }
+};
+
+/** \ingroup CholmodSupport_Module
+  * \class CholmodSupernodalLLT
+  * \brief A supernodal Cholesky (LLT) factorization and solver based on Cholmod
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a supernodal LL^T Cholesky factorization
+  * using the Cholmod library.
+  * This supernodal variant performs best on dense enough problems, e.g., 3D FEM, or very high order 2D FEM.
+  * The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
+  * X and B can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  *
+  * \implsparsesolverconcept
+  *
+  * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
+  *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
+  * \sa \ref TutorialSparseSolverConcept
+  */
+template<typename _MatrixType, int _UpLo = Lower>
+class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSupernodalLLT<_MatrixType, _UpLo> >
+{
+    typedef CholmodBase<_MatrixType, _UpLo, CholmodSupernodalLLT> Base;
+    using Base::m_cholmod;
+    
+  public:
+    
+    typedef _MatrixType MatrixType;
+    
+    CholmodSupernodalLLT() : Base() { init(); }
+
+    CholmodSupernodalLLT(const MatrixType& matrix) : Base()
+    {
+      init();
+      this->compute(matrix);
+    }
+
+    ~CholmodSupernodalLLT() {}
+  protected:
+    void init()
+    {
+      m_cholmod.final_asis = 1;
+      m_cholmod.supernodal = CHOLMOD_SUPERNODAL;
+    }
+};
+
+/** \ingroup CholmodSupport_Module
+  * \class CholmodDecomposition
+  * \brief A general Cholesky factorization and solver based on Cholmod
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a LL^T or LDL^T Cholesky factorization
+  * using the Cholmod library. The sparse matrix A must be selfadjoint and positive definite. The vectors or matrices
+  * X and B can be either dense or sparse.
+  *
+  * This variant permits to change the underlying Cholesky method at runtime.
+  * On the other hand, it does not provide access to the result of the factorization.
+  * The default is to let Cholmod automatically choose between a simplicial and supernodal factorization.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  *
+  * \implsparsesolverconcept
+  *
+  * This class supports all kind of SparseMatrix<>: row or column major; upper, lower, or both; compressed or non compressed.
+  *
+  * \warning Only double precision real and complex scalar types are supported by Cholmod.
+  *
+  * \sa \ref TutorialSparseSolverConcept
+  */
+template<typename _MatrixType, int _UpLo = Lower>
+class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecomposition<_MatrixType, _UpLo> >
+{
+    typedef CholmodBase<_MatrixType, _UpLo, CholmodDecomposition> Base;
+    using Base::m_cholmod;
+    
+  public:
+    
+    typedef _MatrixType MatrixType;
+    
+    CholmodDecomposition() : Base() { init(); }
+
+    CholmodDecomposition(const MatrixType& matrix) : Base()
+    {
+      init();
+      this->compute(matrix);
+    }
+
+    ~CholmodDecomposition() {}
+    
+    void setMode(CholmodMode mode)
+    {
+      switch(mode)
+      {
+        case CholmodAuto:
+          m_cholmod.final_asis = 1;
+          m_cholmod.supernodal = CHOLMOD_AUTO;
+          break;
+        case CholmodSimplicialLLt:
+          m_cholmod.final_asis = 0;
+          m_cholmod.supernodal = CHOLMOD_SIMPLICIAL;
+          m_cholmod.final_ll = 1;
+          break;
+        case CholmodSupernodalLLt:
+          m_cholmod.final_asis = 1;
+          m_cholmod.supernodal = CHOLMOD_SUPERNODAL;
+          break;
+        case CholmodLDLt:
+          m_cholmod.final_asis = 1;
+          m_cholmod.supernodal = CHOLMOD_SIMPLICIAL;
+          break;
+        default:
+          break;
+      }
+    }
+  protected:
+    void init()
+    {
+      m_cholmod.final_asis = 1;
+      m_cholmod.supernodal = CHOLMOD_AUTO;
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CHOLMODSUPPORT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Array.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Array.h
new file mode 100644
index 0000000..e10020d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Array.h
@@ -0,0 +1,331 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARRAY_H
+#define EIGEN_ARRAY_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct traits<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > : traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+  typedef ArrayXpr XprKind;
+  typedef ArrayBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> > XprBase;
+};
+}
+
+/** \class Array
+  * \ingroup Core_Module
+  *
+  * \brief General-purpose arrays with easy API for coefficient-wise operations
+  *
+  * The %Array class is very similar to the Matrix class. It provides
+  * general-purpose one- and two-dimensional arrays. The difference between the
+  * %Array and the %Matrix class is primarily in the API: the API for the
+  * %Array class provides easy access to coefficient-wise operations, while the
+  * API for the %Matrix class provides easy access to linear-algebra
+  * operations.
+  *
+  * See documentation of class Matrix for detailed information on the template parameters
+  * storage layout.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAY_PLUGIN.
+  *
+  * \sa \blank \ref TutorialArrayClass, \ref TopicClassHierarchy
+  */
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+class Array
+  : public PlainObjectBase<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+  public:
+
+    typedef PlainObjectBase<Array> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Array)
+
+    enum { Options = _Options };
+    typedef typename Base::PlainObject PlainObject;
+
+  protected:
+    template <typename Derived, typename OtherDerived, bool IsVector>
+    friend struct internal::conservative_resize_like_impl;
+
+    using Base::m_storage;
+
+  public:
+
+    using Base::base;
+    using Base::coeff;
+    using Base::coeffRef;
+
+    /**
+      * The usage of
+      *   using Base::operator=;
+      * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped
+      * the usage of 'using'. This should be done only for operator=.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const EigenBase<OtherDerived> &other)
+    {
+      return Base::operator=(other);
+    }
+
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill()
+      */
+    /* This overload is needed because the usage of
+      *   using Base::operator=;
+      * fails on MSVC. Since the code below is working with GCC and MSVC, we skipped
+      * the usage of 'using'. This should be done only for operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const Scalar &value)
+    {
+      Base::setConstant(value);
+      return *this;
+    }
+
+    /** Copies the value of the expression \a other into \c *this with automatic resizing.
+      *
+      * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
+      * it will be initialized.
+      *
+      * Note that copying a row-vector into a vector (and conversely) is allowed.
+      * The resizing, if any, is then done in the appropriate way so that row-vectors
+      * remain row-vectors and vectors remain vectors.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const DenseBase<OtherDerived>& other)
+    {
+      return Base::_set(other);
+    }
+
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array& operator=(const Array& other)
+    {
+      return Base::_set(other);
+    }
+    
+    /** Default constructor.
+      *
+      * For fixed-size matrices, does nothing.
+      *
+      * For dynamic-size matrices, creates an empty matrix of size 0. Does not allocate any array. Such a matrix
+      * is called a null matrix. This constructor is the unique way to create null matrices: resizing
+      * a matrix to 0 is not supported.
+      *
+      * \sa resize(Index,Index)
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array() : Base()
+    {
+      Base::_check_template_params();
+      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    // FIXME is it still needed ??
+    /** \internal */
+    EIGEN_DEVICE_FUNC
+    Array(internal::constructor_without_unaligned_array_assert)
+      : Base(internal::constructor_without_unaligned_array_assert())
+    {
+      Base::_check_template_params();
+      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+#endif
+
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Array(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
+      : Base(std::move(other))
+    {
+      Base::_check_template_params();
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
+    }
+    EIGEN_DEVICE_FUNC
+    Array& operator=(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(const T& x)
+    {
+      Base::_check_template_params();
+      Base::template _init1<T>(x);
+    }
+
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array(const T0& val0, const T1& val1)
+    {
+      Base::_check_template_params();
+      this->template _init2<T0,T1>(val0, val1);
+    }
+    #else
+    /** \brief Constructs a fixed-sized array initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC explicit Array(const Scalar *data);
+    /** Constructs a vector or row-vector with given dimension. \only_for_vectors
+      *
+      * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass the dimension here, so it makes more sense to use the default
+      * constructor Array() instead.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Array(Index dim);
+    /** constructs an initialized 1x1 Array with the given coefficient */
+    Array(const Scalar& value);
+    /** constructs an uninitialized array with \a rows rows and \a cols columns.
+      *
+      * This is useful for dynamic-size arrays. For fixed-size arrays,
+      * it is redundant to pass these parameters, so one should use the default constructor
+      * Array() instead. */
+    Array(Index rows, Index cols);
+    /** constructs an initialized 2D vector with given coefficients */
+    Array(const Scalar& val0, const Scalar& val1);
+    #endif
+
+    /** constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2)
+    {
+      Base::_check_template_params();
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Array, 3)
+      m_storage.data()[0] = val0;
+      m_storage.data()[1] = val1;
+      m_storage.data()[2] = val2;
+    }
+    /** constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array(const Scalar& val0, const Scalar& val1, const Scalar& val2, const Scalar& val3)
+    {
+      Base::_check_template_params();
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Array, 4)
+      m_storage.data()[0] = val0;
+      m_storage.data()[1] = val1;
+      m_storage.data()[2] = val2;
+      m_storage.data()[3] = val3;
+    }
+
+    /** Copy constructor */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array(const Array& other)
+            : Base(other)
+    { }
+
+  private:
+    struct PrivateType {};
+  public:
+
+    /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Array(const EigenBase<OtherDerived> &other,
+                              typename internal::enable_if<internal::is_convertible<typename OtherDerived::Scalar,Scalar>::value,
+                                                           PrivateType>::type = PrivateType())
+      : Base(other.derived())
+    { }
+
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
+
+    #ifdef EIGEN_ARRAY_PLUGIN
+    #include EIGEN_ARRAY_PLUGIN
+    #endif
+
+  private:
+
+    template<typename MatrixType, typename OtherDerived, bool SwapPointers>
+    friend struct internal::matrix_swap_impl;
+};
+
+/** \defgroup arraytypedefs Global array typedefs
+  * \ingroup Core_Module
+  *
+  * Eigen defines several typedef shortcuts for most common 1D and 2D array types.
+  *
+  * The general patterns are the following:
+  *
+  * \c ArrayRowsColsType where \c Rows and \c Cols can be \c 2,\c 3,\c 4 for fixed size square matrices or \c X for dynamic size,
+  * and where \c Type can be \c i for integer, \c f for float, \c d for double, \c cf for complex float, \c cd
+  * for complex double.
+  *
+  * For example, \c Array33d is a fixed-size 3x3 array type of doubles, and \c ArrayXXf is a dynamic-size matrix of floats.
+  *
+  * There are also \c ArraySizeType which are self-explanatory. For example, \c Array4cf is
+  * a fixed-size 1D array of 4 complex floats.
+  *
+  * \sa class Array
+  */
+
+#define EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, Size, SizeSuffix)   \
+/** \ingroup arraytypedefs */                                    \
+typedef Array<Type, Size, Size> Array##SizeSuffix##SizeSuffix##TypeSuffix;  \
+/** \ingroup arraytypedefs */                                    \
+typedef Array<Type, Size, 1>    Array##SizeSuffix##TypeSuffix;
+
+#define EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, Size)         \
+/** \ingroup arraytypedefs */                                    \
+typedef Array<Type, Size, Dynamic> Array##Size##X##TypeSuffix;  \
+/** \ingroup arraytypedefs */                                    \
+typedef Array<Type, Dynamic, Size> Array##X##Size##TypeSuffix;
+
+#define EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(Type, TypeSuffix) \
+EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 2, 2) \
+EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 3, 3) \
+EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, 4, 4) \
+EIGEN_MAKE_ARRAY_TYPEDEFS(Type, TypeSuffix, Dynamic, X) \
+EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 2) \
+EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 3) \
+EIGEN_MAKE_ARRAY_FIXED_TYPEDEFS(Type, TypeSuffix, 4)
+
+EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(int,                  i)
+EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(float,                f)
+EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(double,               d)
+EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(std::complex<float>,  cf)
+EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES(std::complex<double>, cd)
+
+#undef EIGEN_MAKE_ARRAY_TYPEDEFS_ALL_SIZES
+#undef EIGEN_MAKE_ARRAY_TYPEDEFS
+
+#undef EIGEN_MAKE_ARRAY_TYPEDEFS_LARGE
+
+#define EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, SizeSuffix) \
+using Eigen::Matrix##SizeSuffix##TypeSuffix; \
+using Eigen::Vector##SizeSuffix##TypeSuffix; \
+using Eigen::RowVector##SizeSuffix##TypeSuffix;
+
+#define EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(TypeSuffix) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 2) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 3) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, 4) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE_AND_SIZE(TypeSuffix, X) \
+
+#define EIGEN_USING_ARRAY_TYPEDEFS \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(i) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(f) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(d) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(cf) \
+EIGEN_USING_ARRAY_TYPEDEFS_FOR_TYPE(cd)
+
+} // end namespace Eigen
+
+#endif // EIGEN_ARRAY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h
new file mode 100644
index 0000000..3dbc708
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayBase.h
@@ -0,0 +1,226 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARRAYBASE_H
+#define EIGEN_ARRAYBASE_H
+
+namespace Eigen { 
+
+template<typename ExpressionType> class MatrixWrapper;
+
+/** \class ArrayBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for all 1D and 2D array, and related expressions
+  *
+  * An array is similar to a dense vector or matrix. While matrices are mathematical
+  * objects with well defined linear algebra operators, an array is just a collection
+  * of scalar values arranged in a one or two dimensionnal fashion. As the main consequence,
+  * all operations applied to an array are performed coefficient wise. Furthermore,
+  * arrays support scalar math functions of the c++ standard library (e.g., std::sin(x)), and convenient
+  * constructors allowing to easily write generic code working for both scalar values
+  * and arrays.
+  *
+  * This class is the base that is inherited by all array expression types.
+  *
+  * \tparam Derived is the derived type, e.g., an array or an expression type.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_ARRAYBASE_PLUGIN.
+  *
+  * \sa class MatrixBase, \ref TopicClassHierarchy
+  */
+template<typename Derived> class ArrayBase
+  : public DenseBase<Derived>
+{
+  public:
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** The base class for a given storage type. */
+    typedef ArrayBase StorageBaseType;
+
+    typedef ArrayBase Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl;
+
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    typedef DenseBase<Derived> Base;
+    using Base::RowsAtCompileTime;
+    using Base::ColsAtCompileTime;
+    using Base::SizeAtCompileTime;
+    using Base::MaxRowsAtCompileTime;
+    using Base::MaxColsAtCompileTime;
+    using Base::MaxSizeAtCompileTime;
+    using Base::IsVectorAtCompileTime;
+    using Base::Flags;
+    
+    using Base::derived;
+    using Base::const_cast_derived;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::coeff;
+    using Base::coeffRef;
+    using Base::lazyAssign;
+    using Base::operator=;
+    using Base::operator+=;
+    using Base::operator-=;
+    using Base::operator*=;
+    using Base::operator/=;
+
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename Base::PlainObject PlainObject;
+
+    /** \internal Represents a matrix with all coefficients equal to one another*/
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::ArrayBase
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
+#   include "../plugins/CommonCwiseUnaryOps.h"
+#   include "../plugins/MatrixCwiseUnaryOps.h"
+#   include "../plugins/ArrayCwiseUnaryOps.h"
+#   include "../plugins/CommonCwiseBinaryOps.h"
+#   include "../plugins/MatrixCwiseBinaryOps.h"
+#   include "../plugins/ArrayCwiseBinaryOps.h"
+#   ifdef EIGEN_ARRAYBASE_PLUGIN
+#     include EIGEN_ARRAYBASE_PLUGIN
+#   endif
+#undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
+
+    /** Special case of the template operator=, in order to prevent the compiler
+      * from generating a default operator= (issue hit with g++ 4.1)
+      */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const ArrayBase& other)
+    {
+      internal::call_assignment(derived(), other.derived());
+      return derived();
+    }
+    
+    /** Set all the entries to \a value.
+      * \sa DenseBase::setConstant(), DenseBase::fill() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const Scalar &value)
+    { Base::setConstant(value); return derived(); }
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator+=(const Scalar& scalar);
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator-=(const Scalar& scalar);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator+=(const ArrayBase<OtherDerived>& other);
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator-=(const ArrayBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator*=(const ArrayBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator/=(const ArrayBase<OtherDerived>& other);
+
+  public:
+    EIGEN_DEVICE_FUNC
+    ArrayBase<Derived>& array() { return *this; }
+    EIGEN_DEVICE_FUNC
+    const ArrayBase<Derived>& array() const { return *this; }
+
+    /** \returns an \link Eigen::MatrixBase Matrix \endlink expression of this array
+      * \sa MatrixBase::array() */
+    EIGEN_DEVICE_FUNC
+    MatrixWrapper<Derived> matrix() { return MatrixWrapper<Derived>(derived()); }
+    EIGEN_DEVICE_FUNC
+    const MatrixWrapper<const Derived> matrix() const { return MatrixWrapper<const Derived>(derived()); }
+
+//     template<typename Dest>
+//     inline void evalTo(Dest& dst) const { dst = matrix(); }
+
+  protected:
+    EIGEN_DEVICE_FUNC
+    ArrayBase() : Base() {}
+
+  private:
+    explicit ArrayBase(Index);
+    ArrayBase(Index,Index);
+    template<typename OtherDerived> explicit ArrayBase(const ArrayBase<OtherDerived>&);
+  protected:
+    // mixing arrays and matrices is not legal
+    template<typename OtherDerived> Derived& operator+=(const MatrixBase<OtherDerived>& )
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
+    // mixing arrays and matrices is not legal
+    template<typename OtherDerived> Derived& operator-=(const MatrixBase<OtherDerived>& )
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
+};
+
+/** replaces \c *this by \c *this - \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
+ArrayBase<Derived>::operator-=(const ArrayBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+/** replaces \c *this by \c *this + \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
+ArrayBase<Derived>::operator+=(const ArrayBase<OtherDerived>& other)
+{
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+/** replaces \c *this by \c *this * \a other coefficient wise.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
+ArrayBase<Derived>::operator*=(const ArrayBase<OtherDerived>& other)
+{
+  call_assignment(derived(), other.derived(), internal::mul_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+/** replaces \c *this by \c *this / \a other coefficient wise.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived &
+ArrayBase<Derived>::operator/=(const ArrayBase<OtherDerived>& other)
+{
+  call_assignment(derived(), other.derived(), internal::div_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_ARRAYBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h
new file mode 100644
index 0000000..688aadd
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ArrayWrapper.h
@@ -0,0 +1,209 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ARRAYWRAPPER_H
+#define EIGEN_ARRAYWRAPPER_H
+
+namespace Eigen { 
+
+/** \class ArrayWrapper
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a mathematical vector or matrix as an array object
+  *
+  * This class is the return type of MatrixBase::array(), and most of the time
+  * this is the only way it is use.
+  *
+  * \sa MatrixBase::array(), class MatrixWrapper
+  */
+
+namespace internal {
+template<typename ExpressionType>
+struct traits<ArrayWrapper<ExpressionType> >
+  : public traits<typename remove_all<typename ExpressionType::Nested>::type >
+{
+  typedef ArrayXpr XprKind;
+  // Let's remove NestByRefBit
+  enum {
+    Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
+    LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0,
+    Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag
+  };
+};
+}
+
+template<typename ExpressionType>
+class ArrayWrapper : public ArrayBase<ArrayWrapper<ExpressionType> >
+{
+  public:
+    typedef ArrayBase<ArrayWrapper> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(ArrayWrapper)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(ArrayWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
+
+    typedef typename internal::conditional<
+                       internal::is_lvalue<ExpressionType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
+
+    using Base::coeffRef;
+
+    EIGEN_DEVICE_FUNC
+    explicit EIGEN_STRONG_INLINE ArrayWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return m_expression.innerStride(); }
+
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    EIGEN_DEVICE_FUNC
+    inline const Scalar* data() const { return m_expression.data(); }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      return m_expression.coeffRef(rowId, colId);
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return m_expression.coeffRef(index);
+    }
+
+    template<typename Dest>
+    EIGEN_DEVICE_FUNC
+    inline void evalTo(Dest& dst) const { dst = m_expression; }
+
+    const typename internal::remove_all<NestedExpressionType>::type& 
+    EIGEN_DEVICE_FUNC
+    nestedExpression() const 
+    {
+      return m_expression;
+    }
+
+    /** Forwards the resizing request to the nested expression
+      * \sa DenseBase::resize(Index)  */
+    EIGEN_DEVICE_FUNC
+    void resize(Index newSize) { m_expression.resize(newSize); }
+    /** Forwards the resizing request to the nested expression
+      * \sa DenseBase::resize(Index,Index)*/
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.resize(rows,cols); }
+
+  protected:
+    NestedExpressionType m_expression;
+};
+
+/** \class MatrixWrapper
+  * \ingroup Core_Module
+  *
+  * \brief Expression of an array as a mathematical vector or matrix
+  *
+  * This class is the return type of ArrayBase::matrix(), and most of the time
+  * this is the only way it is use.
+  *
+  * \sa MatrixBase::matrix(), class ArrayWrapper
+  */
+
+namespace internal {
+template<typename ExpressionType>
+struct traits<MatrixWrapper<ExpressionType> >
+ : public traits<typename remove_all<typename ExpressionType::Nested>::type >
+{
+  typedef MatrixXpr XprKind;
+  // Let's remove NestByRefBit
+  enum {
+    Flags0 = traits<typename remove_all<typename ExpressionType::Nested>::type >::Flags,
+    LvalueBitFlag = is_lvalue<ExpressionType>::value ? LvalueBit : 0,
+    Flags = (Flags0 & ~(NestByRefBit | LvalueBit)) | LvalueBitFlag
+  };
+};
+}
+
+template<typename ExpressionType>
+class MatrixWrapper : public MatrixBase<MatrixWrapper<ExpressionType> >
+{
+  public:
+    typedef MatrixBase<MatrixWrapper<ExpressionType> > Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(MatrixWrapper)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(MatrixWrapper)
+    typedef typename internal::remove_all<ExpressionType>::type NestedExpression;
+
+    typedef typename internal::conditional<
+                       internal::is_lvalue<ExpressionType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type NestedExpressionType;
+
+    using Base::coeffRef;
+
+    EIGEN_DEVICE_FUNC
+    explicit inline MatrixWrapper(ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return m_expression.innerStride(); }
+
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return m_expression.data(); }
+    EIGEN_DEVICE_FUNC
+    inline const Scalar* data() const { return m_expression.data(); }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      return m_expression.derived().coeffRef(rowId, colId);
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return m_expression.coeffRef(index);
+    }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<NestedExpressionType>::type& 
+    nestedExpression() const 
+    {
+      return m_expression;
+    }
+
+    /** Forwards the resizing request to the nested expression
+      * \sa DenseBase::resize(Index)  */
+    EIGEN_DEVICE_FUNC
+    void resize(Index newSize) { m_expression.resize(newSize); }
+    /** Forwards the resizing request to the nested expression
+      * \sa DenseBase::resize(Index,Index)*/
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols) { m_expression.resize(rows,cols); }
+
+  protected:
+    NestedExpressionType m_expression;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_ARRAYWRAPPER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign.h
new file mode 100644
index 0000000..53806ba
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign.h
@@ -0,0 +1,90 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Michael Olbrich <michael.olbrich@gmx.net>
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ASSIGN_H
+#define EIGEN_ASSIGN_H
+
+namespace Eigen {
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE Derived& DenseBase<Derived>
+  ::lazyAssign(const DenseBase<OtherDerived>& other)
+{
+  enum{
+    SameType = internal::is_same<typename Derived::Scalar,typename OtherDerived::Scalar>::value
+  };
+
+  EIGEN_STATIC_ASSERT_LVALUE(Derived)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Derived,OtherDerived)
+  EIGEN_STATIC_ASSERT(SameType,YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+  eigen_assert(rows() == other.rows() && cols() == other.cols());
+  internal::call_assignment_no_alias(derived(),other.derived());
+  
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
+{
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator=(const DenseBase& other)
+{
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const MatrixBase& other)
+{
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const DenseBase<OtherDerived>& other)
+{
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+template <typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const EigenBase<OtherDerived>& other)
+{
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
+{
+  other.derived().evalTo(derived());
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_ASSIGN_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h
new file mode 100644
index 0000000..b0ec7b7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/AssignEvaluator.h
@@ -0,0 +1,935 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ASSIGN_EVALUATOR_H
+#define EIGEN_ASSIGN_EVALUATOR_H
+
+namespace Eigen {
+
+// This implementation is based on Assign.h
+
+namespace internal {
+  
+/***************************************************************************
+* Part 1 : the logic deciding a strategy for traversal and unrolling       *
+***************************************************************************/
+
+// copy_using_evaluator_traits is based on assign_traits
+
+template <typename DstEvaluator, typename SrcEvaluator, typename AssignFunc>
+struct copy_using_evaluator_traits
+{
+  typedef typename DstEvaluator::XprType Dst;
+  typedef typename Dst::Scalar DstScalar;
+  
+  enum {
+    DstFlags = DstEvaluator::Flags,
+    SrcFlags = SrcEvaluator::Flags
+  };
+  
+public:
+  enum {
+    DstAlignment = DstEvaluator::Alignment,
+    SrcAlignment = SrcEvaluator::Alignment,
+    DstHasDirectAccess = DstFlags & DirectAccessBit,
+    JointAlignment = EIGEN_PLAIN_ENUM_MIN(DstAlignment,SrcAlignment)
+  };
+
+private:
+  enum {
+    InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
+              : int(Dst::RowsAtCompileTime),
+    InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
+              : int(DstFlags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
+              : int(Dst::MaxRowsAtCompileTime),
+    OuterStride = int(outer_stride_at_compile_time<Dst>::ret),
+    MaxSizeAtCompileTime = Dst::SizeAtCompileTime
+  };
+
+  // TODO distinguish between linear traversal and inner-traversals
+  typedef typename find_best_packet<DstScalar,Dst::SizeAtCompileTime>::type LinearPacketType;
+  typedef typename find_best_packet<DstScalar,InnerSize>::type InnerPacketType;
+
+  enum {
+    LinearPacketSize = unpacket_traits<LinearPacketType>::size,
+    InnerPacketSize = unpacket_traits<InnerPacketType>::size
+  };
+
+public:
+  enum {
+    LinearRequiredAlignment = unpacket_traits<LinearPacketType>::alignment,
+    InnerRequiredAlignment = unpacket_traits<InnerPacketType>::alignment
+  };
+
+private:
+  enum {
+    DstIsRowMajor = DstFlags&RowMajorBit,
+    SrcIsRowMajor = SrcFlags&RowMajorBit,
+    StorageOrdersAgree = (int(DstIsRowMajor) == int(SrcIsRowMajor)),
+    MightVectorize = bool(StorageOrdersAgree)
+                  && (int(DstFlags) & int(SrcFlags) & ActualPacketAccessBit)
+                  && bool(functor_traits<AssignFunc>::PacketAccess),
+    MayInnerVectorize  = MightVectorize
+                       && int(InnerSize)!=Dynamic && int(InnerSize)%int(InnerPacketSize)==0
+                       && int(OuterStride)!=Dynamic && int(OuterStride)%int(InnerPacketSize)==0
+                       && (EIGEN_UNALIGNED_VECTORIZE  || int(JointAlignment)>=int(InnerRequiredAlignment)),
+    MayLinearize = bool(StorageOrdersAgree) && (int(DstFlags) & int(SrcFlags) & LinearAccessBit),
+    MayLinearVectorize = bool(MightVectorize) && MayLinearize && DstHasDirectAccess
+                       && (EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)) || MaxSizeAtCompileTime == Dynamic),
+      /* If the destination isn't aligned, we have to do runtime checks and we don't unroll,
+         so it's only good for large enough sizes. */
+    MaySliceVectorize  = bool(MightVectorize) && bool(DstHasDirectAccess)
+                       && (int(InnerMaxSize)==Dynamic || int(InnerMaxSize)>=(EIGEN_UNALIGNED_VECTORIZE?InnerPacketSize:(3*InnerPacketSize)))
+      /* slice vectorization can be slow, so we only want it if the slices are big, which is
+         indicated by InnerMaxSize rather than InnerSize, think of the case of a dynamic block
+         in a fixed-size matrix
+         However, with EIGEN_UNALIGNED_VECTORIZE and unrolling, slice vectorization is still worth it */
+  };
+
+public:
+  enum {
+    Traversal = int(MayLinearVectorize) && (LinearPacketSize>InnerPacketSize) ? int(LinearVectorizedTraversal)
+              : int(MayInnerVectorize)   ? int(InnerVectorizedTraversal)
+              : int(MayLinearVectorize)  ? int(LinearVectorizedTraversal)
+              : int(MaySliceVectorize)   ? int(SliceVectorizedTraversal)
+              : int(MayLinearize)        ? int(LinearTraversal)
+                                         : int(DefaultTraversal),
+    Vectorized = int(Traversal) == InnerVectorizedTraversal
+              || int(Traversal) == LinearVectorizedTraversal
+              || int(Traversal) == SliceVectorizedTraversal
+  };
+
+  typedef typename conditional<int(Traversal)==LinearVectorizedTraversal, LinearPacketType, InnerPacketType>::type PacketType;
+
+private:
+  enum {
+    ActualPacketSize    = int(Traversal)==LinearVectorizedTraversal ? LinearPacketSize
+                        : Vectorized ? InnerPacketSize
+                        : 1,
+    UnrollingLimit      = EIGEN_UNROLLING_LIMIT * ActualPacketSize,
+    MayUnrollCompletely = int(Dst::SizeAtCompileTime) != Dynamic
+                       && int(Dst::SizeAtCompileTime) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit),
+    MayUnrollInner      = int(InnerSize) != Dynamic
+                       && int(InnerSize) * (int(DstEvaluator::CoeffReadCost)+int(SrcEvaluator::CoeffReadCost)) <= int(UnrollingLimit)
+  };
+
+public:
+  enum {
+    Unrolling = (int(Traversal) == int(InnerVectorizedTraversal) || int(Traversal) == int(DefaultTraversal))
+                ? (
+                    int(MayUnrollCompletely) ? int(CompleteUnrolling)
+                  : int(MayUnrollInner)      ? int(InnerUnrolling)
+                                             : int(NoUnrolling)
+                  )
+              : int(Traversal) == int(LinearVectorizedTraversal)
+                ? ( bool(MayUnrollCompletely) && ( EIGEN_UNALIGNED_VECTORIZE || (int(DstAlignment)>=int(LinearRequiredAlignment)))
+                          ? int(CompleteUnrolling)
+                          : int(NoUnrolling) )
+              : int(Traversal) == int(LinearTraversal)
+                ? ( bool(MayUnrollCompletely) ? int(CompleteUnrolling) 
+                                              : int(NoUnrolling) )
+#if EIGEN_UNALIGNED_VECTORIZE
+              : int(Traversal) == int(SliceVectorizedTraversal)
+                ? ( bool(MayUnrollInner) ? int(InnerUnrolling)
+                                         : int(NoUnrolling) )
+#endif
+              : int(NoUnrolling)
+  };
+
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "DstXpr: " << typeid(typename DstEvaluator::XprType).name() << std::endl;
+    std::cerr << "SrcXpr: " << typeid(typename SrcEvaluator::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    std::cerr << "DstFlags" << " = " << DstFlags << " (" << demangle_flags(DstFlags) << " )" << std::endl;
+    std::cerr << "SrcFlags" << " = " << SrcFlags << " (" << demangle_flags(SrcFlags) << " )" << std::endl;
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(DstAlignment)
+    EIGEN_DEBUG_VAR(SrcAlignment)
+    EIGEN_DEBUG_VAR(LinearRequiredAlignment)
+    EIGEN_DEBUG_VAR(InnerRequiredAlignment)
+    EIGEN_DEBUG_VAR(JointAlignment)
+    EIGEN_DEBUG_VAR(InnerSize)
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(LinearPacketSize)
+    EIGEN_DEBUG_VAR(InnerPacketSize)
+    EIGEN_DEBUG_VAR(ActualPacketSize)
+    EIGEN_DEBUG_VAR(StorageOrdersAgree)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearize)
+    EIGEN_DEBUG_VAR(MayInnerVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    std::cerr << "Traversal" << " = " << Traversal << " (" << demangle_traversal(Traversal) << ")" << std::endl;
+    EIGEN_DEBUG_VAR(SrcEvaluator::CoeffReadCost)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(MayUnrollCompletely)
+    EIGEN_DEBUG_VAR(MayUnrollInner)
+    std::cerr << "Unrolling" << " = " << Unrolling << " (" << demangle_unrolling(Unrolling) << ")" << std::endl;
+    std::cerr << std::endl;
+  }
+#endif
+};
+
+/***************************************************************************
+* Part 2 : meta-unrollers
+***************************************************************************/
+
+/************************
+*** Default traversal ***
+************************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  
+  enum {
+    outer = Index / DstXprType::InnerSizeAtCompileTime,
+    inner = Index % DstXprType::InnerSizeAtCompileTime
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    kernel.assignCoeffByOuterInner(outer, inner);
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+template<typename Kernel, int Index_, int Stop>
+struct copy_using_evaluator_DefaultTraversal_InnerUnrolling
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
+  {
+    kernel.assignCoeffByOuterInner(outer, Index_);
+    copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Index_+1, Stop>::run(kernel, outer);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index) { }
+};
+
+/***********************
+*** Linear traversal ***
+***********************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_LinearTraversal_CompleteUnrolling
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel& kernel)
+  {
+    kernel.assignCoeff(Index);
+    copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Index+1, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+/**************************
+*** Inner vectorization ***
+**************************/
+
+template<typename Kernel, int Index, int Stop>
+struct copy_using_evaluator_innervec_CompleteUnrolling
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  typedef typename Kernel::PacketType PacketType;
+  
+  enum {
+    outer = Index / DstXprType::InnerSizeAtCompileTime,
+    inner = Index % DstXprType::InnerSizeAtCompileTime,
+    SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+    DstAlignment = Kernel::AssignmentTraits::DstAlignment
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner);
+    enum { NextIndex = Index + unpacket_traits<PacketType>::size };
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, NextIndex, Stop>::run(kernel);
+  }
+};
+
+template<typename Kernel, int Stop>
+struct copy_using_evaluator_innervec_CompleteUnrolling<Kernel, Stop, Stop>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&) { }
+};
+
+template<typename Kernel, int Index_, int Stop, int SrcAlignment, int DstAlignment>
+struct copy_using_evaluator_innervec_InnerUnrolling
+{
+  typedef typename Kernel::PacketType PacketType;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel, Index outer)
+  {
+    kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, Index_);
+    enum { NextIndex = Index_ + unpacket_traits<PacketType>::size };
+    copy_using_evaluator_innervec_InnerUnrolling<Kernel, NextIndex, Stop, SrcAlignment, DstAlignment>::run(kernel, outer);
+  }
+};
+
+template<typename Kernel, int Stop, int SrcAlignment, int DstAlignment>
+struct copy_using_evaluator_innervec_InnerUnrolling<Kernel, Stop, Stop, SrcAlignment, DstAlignment>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &, Index) { }
+};
+
+/***************************************************************************
+* Part 3 : implementation of all cases
+***************************************************************************/
+
+// dense_assignment_loop is based on assign_impl
+
+template<typename Kernel,
+         int Traversal = Kernel::AssignmentTraits::Traversal,
+         int Unrolling = Kernel::AssignmentTraits::Unrolling>
+struct dense_assignment_loop;
+
+/************************
+*** Default traversal ***
+************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static void EIGEN_STRONG_INLINE run(Kernel &kernel)
+  {
+    for(Index outer = 0; outer < kernel.outerSize(); ++outer) {
+      for(Index inner = 0; inner < kernel.innerSize(); ++inner) {
+        kernel.assignCoeffByOuterInner(outer, inner);
+      }
+    }
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, DefaultTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+
+    const Index outerSize = kernel.outerSize();
+    for(Index outer = 0; outer < outerSize; ++outer)
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime>::run(kernel, outer);
+  }
+};
+
+/***************************
+*** Linear vectorization ***
+***************************/
+
+
+// The goal of unaligned_dense_assignment_loop is simply to factorize the handling
+// of the non vectorizable beginning and ending parts
+
+template <bool IsAligned = false>
+struct unaligned_dense_assignment_loop
+{
+  // if IsAligned = true, then do nothing
+  template <typename Kernel>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel&, Index, Index) {}
+};
+
+template <>
+struct unaligned_dense_assignment_loop<false>
+{
+  // MSVC must not inline this functions. If it does, it fails to optimize the
+  // packet access path.
+  // FIXME check which version exhibits this issue
+#if EIGEN_COMP_MSVC
+  template <typename Kernel>
+  static EIGEN_DONT_INLINE void run(Kernel &kernel,
+                                    Index start,
+                                    Index end)
+#else
+  template <typename Kernel>
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel,
+                                      Index start,
+                                      Index end)
+#endif
+  {
+    for (Index index = start; index < end; ++index)
+      kernel.assignCoeff(index);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index size = kernel.size();
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
+    enum {
+      requestedAlignment = Kernel::AssignmentTraits::LinearRequiredAlignment,
+      packetSize = unpacket_traits<PacketType>::size,
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = packet_traits<Scalar>::AlignedOnScalar ? int(requestedAlignment)
+                                                            : int(Kernel::AssignmentTraits::DstAlignment),
+      srcAlignment = Kernel::AssignmentTraits::JointAlignment
+    };
+    const Index alignedStart = dstIsAligned ? 0 : internal::first_aligned<requestedAlignment>(kernel.dstDataPtr(), size);
+    const Index alignedEnd = alignedStart + ((size-alignedStart)/packetSize)*packetSize;
+
+    unaligned_dense_assignment_loop<dstIsAligned!=0>::run(kernel, 0, alignedStart);
+
+    for(Index index = alignedStart; index < alignedEnd; index += packetSize)
+      kernel.template assignPacket<dstAlignment, srcAlignment, PacketType>(index);
+
+    unaligned_dense_assignment_loop<>::run(kernel, alignedEnd, size);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearVectorizedTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
+    
+    enum { size = DstXprType::SizeAtCompileTime,
+           packetSize =unpacket_traits<PacketType>::size,
+           alignedSize = (size/packetSize)*packetSize };
+
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, alignedSize>::run(kernel);
+    copy_using_evaluator_DefaultTraversal_CompleteUnrolling<Kernel, alignedSize, size>::run(kernel);
+  }
+};
+
+/**************************
+*** Inner vectorization ***
+**************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, NoUnrolling>
+{
+  typedef typename Kernel::PacketType PacketType;
+  enum {
+    SrcAlignment = Kernel::AssignmentTraits::SrcAlignment,
+    DstAlignment = Kernel::AssignmentTraits::DstAlignment
+  };
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    const Index packetSize = unpacket_traits<PacketType>::size;
+    for(Index outer = 0; outer < outerSize; ++outer)
+      for(Index inner = 0; inner < innerSize; inner+=packetSize)
+        kernel.template assignPacketByOuterInner<DstAlignment, SrcAlignment, PacketType>(outer, inner);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_innervec_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, InnerVectorizedTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::AssignmentTraits Traits;
+    const Index outerSize = kernel.outerSize();
+    for(Index outer = 0; outer < outerSize; ++outer)
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, DstXprType::InnerSizeAtCompileTime,
+                                                   Traits::SrcAlignment, Traits::DstAlignment>::run(kernel, outer);
+  }
+};
+
+/***********************
+*** Linear traversal ***
+***********************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    const Index size = kernel.size();
+    for(Index i = 0; i < size; ++i)
+      kernel.assignCoeff(i);
+  }
+};
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, LinearTraversal, CompleteUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    copy_using_evaluator_LinearTraversal_CompleteUnrolling<Kernel, 0, DstXprType::SizeAtCompileTime>::run(kernel);
+  }
+};
+
+/**************************
+*** Slice vectorization ***
+***************************/
+
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, NoUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::Scalar Scalar;
+    typedef typename Kernel::PacketType PacketType;
+    enum {
+      packetSize = unpacket_traits<PacketType>::size,
+      requestedAlignment = int(Kernel::AssignmentTraits::InnerRequiredAlignment),
+      alignable = packet_traits<Scalar>::AlignedOnScalar || int(Kernel::AssignmentTraits::DstAlignment)>=sizeof(Scalar),
+      dstIsAligned = int(Kernel::AssignmentTraits::DstAlignment)>=int(requestedAlignment),
+      dstAlignment = alignable ? int(requestedAlignment)
+                               : int(Kernel::AssignmentTraits::DstAlignment)
+    };
+    const Scalar *dst_ptr = kernel.dstDataPtr();
+    if((!bool(dstIsAligned)) && (UIntPtr(dst_ptr) % sizeof(Scalar))>0)
+    {
+      // the pointer is not aligend-on scalar, so alignment is not possible
+      return dense_assignment_loop<Kernel,DefaultTraversal,NoUnrolling>::run(kernel);
+    }
+    const Index packetAlignedMask = packetSize - 1;
+    const Index innerSize = kernel.innerSize();
+    const Index outerSize = kernel.outerSize();
+    const Index alignedStep = alignable ? (packetSize - kernel.outerStride() % packetSize) & packetAlignedMask : 0;
+    Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned<requestedAlignment>(dst_ptr, innerSize);
+
+    for(Index outer = 0; outer < outerSize; ++outer)
+    {
+      const Index alignedEnd = alignedStart + ((innerSize-alignedStart) & ~packetAlignedMask);
+      // do the non-vectorizable part of the assignment
+      for(Index inner = 0; inner<alignedStart ; ++inner)
+        kernel.assignCoeffByOuterInner(outer, inner);
+
+      // do the vectorizable part of the assignment
+      for(Index inner = alignedStart; inner<alignedEnd; inner+=packetSize)
+        kernel.template assignPacketByOuterInner<dstAlignment, Unaligned, PacketType>(outer, inner);
+
+      // do the non-vectorizable part of the assignment
+      for(Index inner = alignedEnd; inner<innerSize ; ++inner)
+        kernel.assignCoeffByOuterInner(outer, inner);
+
+      alignedStart = numext::mini((alignedStart+alignedStep)%packetSize, innerSize);
+    }
+  }
+};
+
+#if EIGEN_UNALIGNED_VECTORIZE
+template<typename Kernel>
+struct dense_assignment_loop<Kernel, SliceVectorizedTraversal, InnerUnrolling>
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void run(Kernel &kernel)
+  {
+    typedef typename Kernel::DstEvaluatorType::XprType DstXprType;
+    typedef typename Kernel::PacketType PacketType;
+
+    enum { size = DstXprType::InnerSizeAtCompileTime,
+           packetSize =unpacket_traits<PacketType>::size,
+           vectorizableSize = (size/packetSize)*packetSize };
+
+    for(Index outer = 0; outer < kernel.outerSize(); ++outer)
+    {
+      copy_using_evaluator_innervec_InnerUnrolling<Kernel, 0, vectorizableSize, 0, 0>::run(kernel, outer);
+      copy_using_evaluator_DefaultTraversal_InnerUnrolling<Kernel, vectorizableSize, size>::run(kernel, outer);
+    }
+  }
+};
+#endif
+
+
+/***************************************************************************
+* Part 4 : Generic dense assignment kernel
+***************************************************************************/
+
+// This class generalize the assignment of a coefficient (or packet) from one dense evaluator
+// to another dense writable evaluator.
+// It is parametrized by the two evaluators, and the actual assignment functor.
+// This abstraction level permits to keep the evaluation loops as simple and as generic as possible.
+// One can customize the assignment using this generic dense_assignment_kernel with different
+// functors, or by completely overloading it, by-passing a functor.
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized>
+class generic_dense_assignment_kernel
+{
+protected:
+  typedef typename DstEvaluatorTypeT::XprType DstXprType;
+  typedef typename SrcEvaluatorTypeT::XprType SrcXprType;
+public:
+  
+  typedef DstEvaluatorTypeT DstEvaluatorType;
+  typedef SrcEvaluatorTypeT SrcEvaluatorType;
+  typedef typename DstEvaluatorType::Scalar Scalar;
+  typedef copy_using_evaluator_traits<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor> AssignmentTraits;
+  typedef typename AssignmentTraits::PacketType PacketType;
+  
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : m_dst(dst), m_src(src), m_functor(func), m_dstExpr(dstExpr)
+  {
+    #ifdef EIGEN_DEBUG_ASSIGN
+    AssignmentTraits::debug();
+    #endif
+  }
+  
+  EIGEN_DEVICE_FUNC Index size() const        { return m_dstExpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const   { return m_dstExpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const   { return m_dstExpr.outerSize(); }
+  EIGEN_DEVICE_FUNC Index rows() const        { return m_dstExpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const        { return m_dstExpr.cols(); }
+  EIGEN_DEVICE_FUNC Index outerStride() const { return m_dstExpr.outerStride(); }
+  
+  EIGEN_DEVICE_FUNC DstEvaluatorType& dstEvaluator() { return m_dst; }
+  EIGEN_DEVICE_FUNC const SrcEvaluatorType& srcEvaluator() const { return m_src; }
+  
+  /// Assign src(row,col) to dst(row,col) through the assignment functor.
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index row, Index col)
+  {
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), m_src.coeff(row,col));
+  }
+  
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Index index)
+  {
+    m_functor.assignCoeff(m_dst.coeffRef(index), m_src.coeff(index));
+  }
+  
+  /// \sa assignCoeff(Index,Index)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeffByOuterInner(Index outer, Index inner)
+  {
+    Index row = rowIndexByOuterInner(outer, inner); 
+    Index col = colIndexByOuterInner(outer, inner); 
+    assignCoeff(row, col);
+  }
+  
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index row, Index col)
+  {
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(row,col), m_src.template packet<LoadMode,PacketType>(row,col));
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacket(Index index)
+  {
+    m_functor.template assignPacket<StoreMode>(&m_dst.coeffRef(index), m_src.template packet<LoadMode,PacketType>(index));
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = rowIndexByOuterInner(outer, inner); 
+    Index col = colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
+  
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner)
+  {
+    typedef typename DstEvaluatorType::ExpressionTraits Traits;
+    return int(Traits::RowsAtCompileTime) == 1 ? 0
+      : int(Traits::ColsAtCompileTime) == 1 ? inner
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? outer
+      : inner;
+  }
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner)
+  {
+    typedef typename DstEvaluatorType::ExpressionTraits Traits;
+    return int(Traits::ColsAtCompileTime) == 1 ? 0
+      : int(Traits::RowsAtCompileTime) == 1 ? inner
+      : int(DstEvaluatorType::Flags)&RowMajorBit ? inner
+      : outer;
+  }
+
+  EIGEN_DEVICE_FUNC const Scalar* dstDataPtr() const
+  {
+    return m_dstExpr.data();
+  }
+  
+protected:
+  DstEvaluatorType& m_dst;
+  const SrcEvaluatorType& m_src;
+  const Functor &m_functor;
+  // TODO find a way to avoid the needs of the original expression
+  DstXprType& m_dstExpr;
+};
+
+/***************************************************************************
+* Part 5 : Entry point for dense rectangular assignment
+***************************************************************************/
+
+template<typename DstXprType,typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const Functor &/*func*/)
+{
+  EIGEN_ONLY_USED_FOR_DEBUG(dst);
+  EIGEN_ONLY_USED_FOR_DEBUG(src);
+  eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+}
+
+template<typename DstXprType,typename SrcXprType, typename T1, typename T2>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void resize_if_allowed(DstXprType &dst, const SrcXprType& src, const internal::assign_op<T1,T2> &/*func*/)
+{
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if(((dst.rows()!=dstRows) || (dst.cols()!=dstCols)))
+    dst.resize(dstRows, dstCols);
+  eigen_assert(dst.rows() == dstRows && dst.cols() == dstCols);
+}
+
+template<typename DstXprType, typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
+{
+  typedef evaluator<DstXprType> DstEvaluatorType;
+  typedef evaluator<SrcXprType> SrcEvaluatorType;
+
+  SrcEvaluatorType srcEvaluator(src);
+
+  // NOTE To properly handle A = (A*A.transpose())/s with A rectangular,
+  // we need to resize the destination after the source evaluator has been created.
+  resize_if_allowed(dst, src, func);
+
+  DstEvaluatorType dstEvaluator(dst);
+    
+  typedef generic_dense_assignment_kernel<DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
+  Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived());
+
+  dense_assignment_loop<Kernel>::run(kernel);
+}
+
+template<typename DstXprType, typename SrcXprType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void call_dense_assignment_loop(DstXprType& dst, const SrcXprType& src)
+{
+  call_dense_assignment_loop(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
+}
+
+/***************************************************************************
+* Part 6 : Generic assignment
+***************************************************************************/
+
+// Based on the respective shapes of the destination and source,
+// the class AssignmentKind determine the kind of assignment mechanism.
+// AssignmentKind must define a Kind typedef.
+template<typename DstShape, typename SrcShape> struct AssignmentKind;
+
+// Assignement kind defined in this file:
+struct Dense2Dense {};
+struct EigenBase2EigenBase {};
+
+template<typename,typename> struct AssignmentKind { typedef EigenBase2EigenBase Kind; };
+template<> struct AssignmentKind<DenseShape,DenseShape> { typedef Dense2Dense Kind; };
+    
+// This is the main assignment class
+template< typename DstXprType, typename SrcXprType, typename Functor,
+          typename Kind = typename AssignmentKind< typename evaluator_traits<DstXprType>::Shape , typename evaluator_traits<SrcXprType>::Shape >::Kind,
+          typename EnableIf = void>
+struct Assignment;
+
+
+// The only purpose of this call_assignment() function is to deal with noalias() / "assume-aliasing" and automatic transposition.
+// Indeed, I (Gael) think that this concept of "assume-aliasing" was a mistake, and it makes thing quite complicated.
+// So this intermediate function removes everything related to "assume-aliasing" such that Assignment
+// does not has to bother about these annoying details.
+
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(const Dst& dst, const Src& src)
+{
+  call_assignment(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+                     
+// Deal with "assume-aliasing"
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if< evaluator_assume_aliasing<Src>::value, void*>::type = 0)
+{
+  typename plain_matrix_type<Src>::type tmp(src);
+  call_assignment_no_alias(dst, tmp, func);
+}
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(Dst& dst, const Src& src, const Func& func, typename enable_if<!evaluator_assume_aliasing<Src>::value, void*>::type = 0)
+{
+  call_assignment_no_alias(dst, src, func);
+}
+
+// by-pass "assume-aliasing"
+// When there is no aliasing, we require that 'dst' has been properly resized
+template<typename Dst, template <typename> class StorageBase, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment(NoAlias<Dst,StorageBase>& dst, const Src& src, const Func& func)
+{
+  call_assignment_no_alias(dst.expression(), src, func);
+}
+
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias(Dst& dst, const Src& src, const Func& func)
+{
+  enum {
+    NeedToTranspose = (    (int(Dst::RowsAtCompileTime) == 1 && int(Src::ColsAtCompileTime) == 1)
+                        || (int(Dst::ColsAtCompileTime) == 1 && int(Src::RowsAtCompileTime) == 1)
+                      ) && int(Dst::SizeAtCompileTime) != 1
+  };
+
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst>::type ActualDstTypeCleaned;
+  typedef typename internal::conditional<NeedToTranspose, Transpose<Dst>, Dst&>::type ActualDstType;
+  ActualDstType actualDst(dst);
+  
+  // TODO check whether this is the right place to perform these checks:
+  EIGEN_STATIC_ASSERT_LVALUE(Dst)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(ActualDstTypeCleaned,Src)
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename ActualDstTypeCleaned::Scalar,typename Src::Scalar);
+  
+  Assignment<ActualDstTypeCleaned,Src,Func>::run(actualDst, src, func);
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias(Dst& dst, const Src& src)
+{
+  call_assignment_no_alias(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+
+template<typename Dst, typename Src, typename Func>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src, const Func& func)
+{
+  // TODO check whether this is the right place to perform these checks:
+  EIGEN_STATIC_ASSERT_LVALUE(Dst)
+  EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Dst,Src)
+  EIGEN_CHECK_BINARY_COMPATIBILIY(Func,typename Dst::Scalar,typename Src::Scalar);
+
+  Assignment<Dst,Src,Func>::run(dst, src, func);
+}
+template<typename Dst, typename Src>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_assignment_no_alias_no_transpose(Dst& dst, const Src& src)
+{
+  call_assignment_no_alias_no_transpose(dst, src, internal::assign_op<typename Dst::Scalar,typename Src::Scalar>());
+}
+
+// forward declaration
+template<typename Dst, typename Src> void check_for_aliasing(const Dst &dst, const Src &src);
+
+// Generic Dense to Dense assignment
+// Note that the last template argument "Weak" is needed to make it possible to perform
+// both partial specialization+SFINAE without ambiguous specialization
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak>
+struct Assignment<DstXprType, SrcXprType, Functor, Dense2Dense, Weak>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+#ifndef EIGEN_NO_DEBUG
+    internal::check_for_aliasing(dst, src);
+#endif
+    
+    call_dense_assignment_loop(dst, src, func);
+  }
+};
+
+// Generic assignment through evalTo.
+// TODO: not sure we have to keep that one, but it helps porting current code to new evaluator mechanism.
+// Note that the last template argument "Weak" is needed to make it possible to perform
+// both partial specialization+SFINAE without ambiguous specialization
+template< typename DstXprType, typename SrcXprType, typename Functor, typename Weak>
+struct Assignment<DstXprType, SrcXprType, Functor, EigenBase2EigenBase, Weak>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.evalTo(dst);
+  }
+
+  // NOTE The following two functions are templated to avoid their instanciation if not needed
+  //      This is needed because some expressions supports evalTo only and/or have 'void' as scalar type.
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.addTo(dst);
+  }
+
+  template<typename SrcScalarType>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,SrcScalarType> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    src.subTo(dst);
+  }
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ASSIGN_EVALUATOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h
new file mode 100755
index 0000000..6c2ab92
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Assign_MKL.h
@@ -0,0 +1,176 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+ Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+ 
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   MKL VML support for coefficient-wise unary Eigen expressions like a=b.sin()
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_ASSIGN_VML_H
+#define EIGEN_ASSIGN_VML_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Dst, typename Src>
+class vml_assign_traits
+{
+  private:
+    enum {
+      DstHasDirectAccess = Dst::Flags & DirectAccessBit,
+      SrcHasDirectAccess = Src::Flags & DirectAccessBit,
+      StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)),
+      InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime)
+                : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime)
+                : int(Dst::RowsAtCompileTime),
+      InnerMaxSize  = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime)
+                    : int(Dst::Flags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime)
+                    : int(Dst::MaxRowsAtCompileTime),
+      MaxSizeAtCompileTime = Dst::SizeAtCompileTime,
+
+      MightEnableVml = StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1,
+      MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit),
+      VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize,
+      LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD
+    };
+  public:
+    enum {
+      EnableVml = MightEnableVml && LargeEnough,
+      Traversal = MightLinearize ? LinearTraversal : DefaultTraversal
+    };
+};
+
+#define EIGEN_PP_EXPAND(ARG) ARG
+#if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1)
+#define EIGEN_VMLMODE_EXPAND_LA , VML_HA
+#else
+#define EIGEN_VMLMODE_EXPAND_LA , VML_LA
+#endif
+
+#define EIGEN_VMLMODE_EXPAND__ 
+
+#define EIGEN_VMLMODE_PREFIX_LA vm
+#define EIGEN_VMLMODE_PREFIX__  v
+#define EIGEN_VMLMODE_PREFIX(VMLMODE) EIGEN_CAT(EIGEN_VMLMODE_PREFIX_,VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested>                                                                         \
+  struct Assignment<DstXprType, CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested>, assign_op<EIGENTYPE,EIGENTYPE>,   \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {              \
+    typedef CwiseUnaryOp<scalar_##EIGENOP##_op<EIGENTYPE>, SrcXprNested> SrcXprType;                                            \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                   \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal) {                                              \
+        VMLOP(dst.size(), (const VMLTYPE*)src.nestedExpression().data(),                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                           \
+      } else {                                                                                                                  \
+        const Index outerSize = dst.outerSize();                                                                                \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                      \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) :                             \
+                                                      &(src.nestedExpression().coeffRef(0, outer));                             \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                           \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr,                                                                      \
+                (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                             \
+        }                                                                                                                       \
+      }                                                                                                                         \
+    }                                                                                                                           \
+  };                                                                                                                            \
+
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),s##VMLOP), float, float, VMLMODE)           \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),d##VMLOP), double, double, VMLMODE)
+
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)                                                         \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),c##VMLOP), scomplex, MKL_Complex8, VMLMODE) \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, EIGEN_CAT(EIGEN_VMLMODE_PREFIX(VMLMODE),z##VMLOP), dcomplex, MKL_Complex16, VMLMODE)
+  
+#define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP, VMLMODE)                                                              \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP, VMLMODE)                                                               \
+  EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(EIGENOP, VMLOP, VMLMODE)
+
+  
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sin,   Sin,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(asin,  Asin,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sinh,  Sinh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cos,   Cos,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(acos,  Acos,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(cosh,  Cosh,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tan,   Tan,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(atan,  Atan,  LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(tanh,  Tanh,  LA)
+// EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs,   Abs,    _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(exp,   Exp,   LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log,   Ln,    LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(log10, Log10, LA)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS(sqrt,  Sqrt,  _)
+
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr,   _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_CPLX(arg, Arg,      _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(round, Round,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(floor, Floor,  _)
+EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(ceil,  Ceil,   _)
+
+#define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE, VMLMODE)                                           \
+  template< typename DstXprType, typename SrcXprNested, typename Plain>                                                       \
+  struct Assignment<DstXprType, CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested,                       \
+                    const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> >, assign_op<EIGENTYPE,EIGENTYPE>,    \
+                   Dense2Dense, typename enable_if<vml_assign_traits<DstXprType,SrcXprNested>::EnableVml>::type> {            \
+    typedef CwiseBinaryOp<scalar_##EIGENOP##_op<EIGENTYPE,EIGENTYPE>, SrcXprNested,                                           \
+                    const CwiseNullaryOp<internal::scalar_constant_op<EIGENTYPE>,Plain> > SrcXprType;                         \
+    static void run(DstXprType &dst, const SrcXprType &src, const assign_op<EIGENTYPE,EIGENTYPE> &/*func*/) {                 \
+      eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());                                                     \
+      VMLTYPE exponent = reinterpret_cast<const VMLTYPE&>(src.rhs().functor().m_other);                                       \
+      if(vml_assign_traits<DstXprType,SrcXprNested>::Traversal==LinearTraversal)                                              \
+      {                                                                                                                       \
+        VMLOP( dst.size(), (const VMLTYPE*)src.lhs().data(), exponent,                                                        \
+              (VMLTYPE*)dst.data() EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE) );                                         \
+      } else {                                                                                                                \
+        const Index outerSize = dst.outerSize();                                                                              \
+        for(Index outer = 0; outer < outerSize; ++outer) {                                                                    \
+          const EIGENTYPE *src_ptr = src.IsRowMajor ? &(src.lhs().coeffRef(outer,0)) :                                        \
+                                                      &(src.lhs().coeffRef(0, outer));                                        \
+          EIGENTYPE *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer));                         \
+          VMLOP( dst.innerSize(), (const VMLTYPE*)src_ptr, exponent,                                                          \
+                 (VMLTYPE*)dst_ptr EIGEN_PP_EXPAND(EIGEN_VMLMODE_EXPAND_##VMLMODE));                                          \
+        }                                                                                                                     \
+      }                                                                                                                       \
+    }                                                                                                                         \
+  };
+  
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmsPowx, float,    float,         LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdPowx, double,   double,        LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcPowx, scomplex, MKL_Complex8,  LA)
+EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzPowx, dcomplex, MKL_Complex16, LA)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ASSIGN_VML_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h
new file mode 100644
index 0000000..4978c91
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BandMatrix.h
@@ -0,0 +1,353 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BANDMATRIX_H
+#define EIGEN_BANDMATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Derived>
+class BandMatrixBase : public EigenBase<Derived>
+{
+  public:
+
+    enum {
+      Flags = internal::traits<Derived>::Flags,
+      CoeffReadCost = internal::traits<Derived>::CoeffReadCost,
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      Supers = internal::traits<Derived>::Supers,
+      Subs   = internal::traits<Derived>::Subs,
+      Options = internal::traits<Derived>::Options
+    };
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType;
+    typedef typename DenseMatrixType::StorageIndex StorageIndex;
+    typedef typename internal::traits<Derived>::CoefficientsType CoefficientsType;
+    typedef EigenBase<Derived> Base;
+
+  protected:
+    enum {
+      DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic))
+                            ? 1 + Supers + Subs
+                            : Dynamic,
+      SizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime)
+    };
+
+  public:
+    
+    using Base::derived;
+    using Base::rows;
+    using Base::cols;
+
+    /** \returns the number of super diagonals */
+    inline Index supers() const { return derived().supers(); }
+
+    /** \returns the number of sub diagonals */
+    inline Index subs() const { return derived().subs(); }
+    
+    /** \returns an expression of the underlying coefficient matrix */
+    inline const CoefficientsType& coeffs() const { return derived().coeffs(); }
+    
+    /** \returns an expression of the underlying coefficient matrix */
+    inline CoefficientsType& coeffs() { return derived().coeffs(); }
+
+    /** \returns a vector expression of the \a i -th column,
+      * only the meaningful part is returned.
+      * \warning the internal storage must be column major. */
+    inline Block<CoefficientsType,Dynamic,1> col(Index i)
+    {
+      EIGEN_STATIC_ASSERT((Options&RowMajor)==0,THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+      Index start = 0;
+      Index len = coeffs().rows();
+      if (i<=supers())
+      {
+        start = supers()-i;
+        len = (std::min)(rows(),std::max<Index>(0,coeffs().rows() - (supers()-i)));
+      }
+      else if (i>=rows()-subs())
+        len = std::max<Index>(0,coeffs().rows() - (i + 1 - rows() + subs()));
+      return Block<CoefficientsType,Dynamic,1>(coeffs(), start, i, len, 1);
+    }
+
+    /** \returns a vector expression of the main diagonal */
+    inline Block<CoefficientsType,1,SizeAtCompileTime> diagonal()
+    { return Block<CoefficientsType,1,SizeAtCompileTime>(coeffs(),supers(),0,1,(std::min)(rows(),cols())); }
+
+    /** \returns a vector expression of the main diagonal (const version) */
+    inline const Block<const CoefficientsType,1,SizeAtCompileTime> diagonal() const
+    { return Block<const CoefficientsType,1,SizeAtCompileTime>(coeffs(),supers(),0,1,(std::min)(rows(),cols())); }
+
+    template<int Index> struct DiagonalIntReturnType {
+      enum {
+        ReturnOpposite = (Options&SelfAdjoint) && (((Index)>0 && Supers==0) || ((Index)<0 && Subs==0)),
+        Conjugate = ReturnOpposite && NumTraits<Scalar>::IsComplex,
+        ActualIndex = ReturnOpposite ? -Index : Index,
+        DiagonalSize = (RowsAtCompileTime==Dynamic || ColsAtCompileTime==Dynamic)
+                     ? Dynamic
+                     : (ActualIndex<0
+                     ? EIGEN_SIZE_MIN_PREFER_DYNAMIC(ColsAtCompileTime, RowsAtCompileTime + ActualIndex)
+                     : EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime, ColsAtCompileTime - ActualIndex))
+      };
+      typedef Block<CoefficientsType,1, DiagonalSize> BuildType;
+      typedef typename internal::conditional<Conjugate,
+                 CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>,BuildType >,
+                 BuildType>::type Type;
+    };
+
+    /** \returns a vector expression of the \a N -th sub or super diagonal */
+    template<int N> inline typename DiagonalIntReturnType<N>::Type diagonal()
+    {
+      return typename DiagonalIntReturnType<N>::BuildType(coeffs(), supers()-N, (std::max)(0,N), 1, diagonalLength(N));
+    }
+
+    /** \returns a vector expression of the \a N -th sub or super diagonal */
+    template<int N> inline const typename DiagonalIntReturnType<N>::Type diagonal() const
+    {
+      return typename DiagonalIntReturnType<N>::BuildType(coeffs(), supers()-N, (std::max)(0,N), 1, diagonalLength(N));
+    }
+
+    /** \returns a vector expression of the \a i -th sub or super diagonal */
+    inline Block<CoefficientsType,1,Dynamic> diagonal(Index i)
+    {
+      eigen_assert((i<0 && -i<=subs()) || (i>=0 && i<=supers()));
+      return Block<CoefficientsType,1,Dynamic>(coeffs(), supers()-i, std::max<Index>(0,i), 1, diagonalLength(i));
+    }
+
+    /** \returns a vector expression of the \a i -th sub or super diagonal */
+    inline const Block<const CoefficientsType,1,Dynamic> diagonal(Index i) const
+    {
+      eigen_assert((i<0 && -i<=subs()) || (i>=0 && i<=supers()));
+      return Block<const CoefficientsType,1,Dynamic>(coeffs(), supers()-i, std::max<Index>(0,i), 1, diagonalLength(i));
+    }
+    
+    template<typename Dest> inline void evalTo(Dest& dst) const
+    {
+      dst.resize(rows(),cols());
+      dst.setZero();
+      dst.diagonal() = diagonal();
+      for (Index i=1; i<=supers();++i)
+        dst.diagonal(i) = diagonal(i);
+      for (Index i=1; i<=subs();++i)
+        dst.diagonal(-i) = diagonal(-i);
+    }
+
+    DenseMatrixType toDenseMatrix() const
+    {
+      DenseMatrixType res(rows(),cols());
+      evalTo(res);
+      return res;
+    }
+
+  protected:
+
+    inline Index diagonalLength(Index i) const
+    { return i<0 ? (std::min)(cols(),rows()+i) : (std::min)(rows(),cols()-i); }
+};
+
+/**
+  * \class BandMatrix
+  * \ingroup Core_Module
+  *
+  * \brief Represents a rectangular matrix with a banded storage
+  *
+  * \tparam _Scalar Numeric type, i.e. float, double, int
+  * \tparam _Rows Number of rows, or \b Dynamic
+  * \tparam _Cols Number of columns, or \b Dynamic
+  * \tparam _Supers Number of super diagonal
+  * \tparam _Subs Number of sub diagonal
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of \b #SelfAdjoint
+  *                  The former controls \ref TopicStorageOrders "storage order", and defaults to
+  *                  column-major. The latter controls whether the matrix represents a selfadjoint
+  *                  matrix in which case either Supers of Subs have to be null.
+  *
+  * \sa class TridiagonalMatrix
+  */
+
+template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options>
+struct traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef _Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef Eigen::Index StorageIndex;
+  enum {
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    RowsAtCompileTime = _Rows,
+    ColsAtCompileTime = _Cols,
+    MaxRowsAtCompileTime = _Rows,
+    MaxColsAtCompileTime = _Cols,
+    Flags = LvalueBit,
+    Supers = _Supers,
+    Subs = _Subs,
+    Options = _Options,
+    DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) ? 1 + Supers + Subs : Dynamic
+  };
+  typedef Matrix<Scalar,DataRowsAtCompileTime,ColsAtCompileTime,Options&RowMajor?RowMajor:ColMajor> CoefficientsType;
+};
+
+template<typename _Scalar, int Rows, int Cols, int Supers, int Subs, int Options>
+class BandMatrix : public BandMatrixBase<BandMatrix<_Scalar,Rows,Cols,Supers,Subs,Options> >
+{
+  public:
+
+    typedef typename internal::traits<BandMatrix>::Scalar Scalar;
+    typedef typename internal::traits<BandMatrix>::StorageIndex StorageIndex;
+    typedef typename internal::traits<BandMatrix>::CoefficientsType CoefficientsType;
+
+    explicit inline BandMatrix(Index rows=Rows, Index cols=Cols, Index supers=Supers, Index subs=Subs)
+      : m_coeffs(1+supers+subs,cols),
+        m_rows(rows), m_supers(supers), m_subs(subs)
+    {
+    }
+
+    /** \returns the number of columns */
+    inline Index rows() const { return m_rows.value(); }
+
+    /** \returns the number of rows */
+    inline Index cols() const { return m_coeffs.cols(); }
+
+    /** \returns the number of super diagonals */
+    inline Index supers() const { return m_supers.value(); }
+
+    /** \returns the number of sub diagonals */
+    inline Index subs() const { return m_subs.value(); }
+
+    inline const CoefficientsType& coeffs() const { return m_coeffs; }
+    inline CoefficientsType& coeffs() { return m_coeffs; }
+
+  protected:
+
+    CoefficientsType m_coeffs;
+    internal::variable_if_dynamic<Index, Rows>   m_rows;
+    internal::variable_if_dynamic<Index, Supers> m_supers;
+    internal::variable_if_dynamic<Index, Subs>   m_subs;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+class BandMatrixWrapper;
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+struct traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef typename _CoefficientsType::Scalar Scalar;
+  typedef typename _CoefficientsType::StorageKind StorageKind;
+  typedef typename _CoefficientsType::StorageIndex StorageIndex;
+  enum {
+    CoeffReadCost = internal::traits<_CoefficientsType>::CoeffReadCost,
+    RowsAtCompileTime = _Rows,
+    ColsAtCompileTime = _Cols,
+    MaxRowsAtCompileTime = _Rows,
+    MaxColsAtCompileTime = _Cols,
+    Flags = LvalueBit,
+    Supers = _Supers,
+    Subs = _Subs,
+    Options = _Options,
+    DataRowsAtCompileTime = ((Supers!=Dynamic) && (Subs!=Dynamic)) ? 1 + Supers + Subs : Dynamic
+  };
+  typedef _CoefficientsType CoefficientsType;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+class BandMatrixWrapper : public BandMatrixBase<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  public:
+
+    typedef typename internal::traits<BandMatrixWrapper>::Scalar Scalar;
+    typedef typename internal::traits<BandMatrixWrapper>::CoefficientsType CoefficientsType;
+    typedef typename internal::traits<BandMatrixWrapper>::StorageIndex StorageIndex;
+
+    explicit inline BandMatrixWrapper(const CoefficientsType& coeffs, Index rows=_Rows, Index cols=_Cols, Index supers=_Supers, Index subs=_Subs)
+      : m_coeffs(coeffs),
+        m_rows(rows), m_supers(supers), m_subs(subs)
+    {
+      EIGEN_UNUSED_VARIABLE(cols);
+      //internal::assert(coeffs.cols()==cols() && (supers()+subs()+1)==coeffs.rows());
+    }
+
+    /** \returns the number of columns */
+    inline Index rows() const { return m_rows.value(); }
+
+    /** \returns the number of rows */
+    inline Index cols() const { return m_coeffs.cols(); }
+
+    /** \returns the number of super diagonals */
+    inline Index supers() const { return m_supers.value(); }
+
+    /** \returns the number of sub diagonals */
+    inline Index subs() const { return m_subs.value(); }
+
+    inline const CoefficientsType& coeffs() const { return m_coeffs; }
+
+  protected:
+
+    const CoefficientsType& m_coeffs;
+    internal::variable_if_dynamic<Index, _Rows>   m_rows;
+    internal::variable_if_dynamic<Index, _Supers> m_supers;
+    internal::variable_if_dynamic<Index, _Subs>   m_subs;
+};
+
+/**
+  * \class TridiagonalMatrix
+  * \ingroup Core_Module
+  *
+  * \brief Represents a tridiagonal matrix with a compact banded storage
+  *
+  * \tparam Scalar Numeric type, i.e. float, double, int
+  * \tparam Size Number of rows and cols, or \b Dynamic
+  * \tparam Options Can be 0 or \b SelfAdjoint
+  *
+  * \sa class BandMatrix
+  */
+template<typename Scalar, int Size, int Options>
+class TridiagonalMatrix : public BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor>
+{
+    typedef BandMatrix<Scalar,Size,Size,Options&SelfAdjoint?0:1,1,Options|RowMajor> Base;
+    typedef typename Base::StorageIndex StorageIndex;
+  public:
+    explicit TridiagonalMatrix(Index size = Size) : Base(size,size,Options&SelfAdjoint?0:1,1) {}
+
+    inline typename Base::template DiagonalIntReturnType<1>::Type super()
+    { return Base::template diagonal<1>(); }
+    inline const typename Base::template DiagonalIntReturnType<1>::Type super() const
+    { return Base::template diagonal<1>(); }
+    inline typename Base::template DiagonalIntReturnType<-1>::Type sub()
+    { return Base::template diagonal<-1>(); }
+    inline const typename Base::template DiagonalIntReturnType<-1>::Type sub() const
+    { return Base::template diagonal<-1>(); }
+  protected:
+};
+
+
+struct BandShape {};
+
+template<typename _Scalar, int _Rows, int _Cols, int _Supers, int _Subs, int _Options>
+struct evaluator_traits<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrix<_Scalar,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<typename _CoefficientsType,int _Rows, int _Cols, int _Supers, int _Subs,int _Options>
+struct evaluator_traits<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+  : public evaluator_traits_base<BandMatrixWrapper<_CoefficientsType,_Rows,_Cols,_Supers,_Subs,_Options> >
+{
+  typedef BandShape Shape;
+};
+
+template<> struct AssignmentKind<DenseShape,BandShape> { typedef EigenBase2EigenBase Kind; };
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BANDMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Block.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Block.h
new file mode 100644
index 0000000..11de45c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Block.h
@@ -0,0 +1,452 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BLOCK_H
+#define EIGEN_BLOCK_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
+struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprType>
+{
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef typename traits<XprType>::StorageKind StorageKind;
+  typedef typename traits<XprType>::XprKind XprKind;
+  typedef typename ref_selector<XprType>::type XprTypeNested;
+  typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+  enum{
+    MatrixRows = traits<XprType>::RowsAtCompileTime,
+    MatrixCols = traits<XprType>::ColsAtCompileTime,
+    RowsAtCompileTime = MatrixRows == 0 ? 0 : BlockRows,
+    ColsAtCompileTime = MatrixCols == 0 ? 0 : BlockCols,
+    MaxRowsAtCompileTime = BlockRows==0 ? 0
+                         : RowsAtCompileTime != Dynamic ? int(RowsAtCompileTime)
+                         : int(traits<XprType>::MaxRowsAtCompileTime),
+    MaxColsAtCompileTime = BlockCols==0 ? 0
+                         : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime)
+                         : int(traits<XprType>::MaxColsAtCompileTime),
+
+    XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
+    IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+               : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+               : XprTypeIsRowMajor,
+    HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor),
+    InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
+    InnerStrideAtCompileTime = HasSameStorageOrderAsXprType
+                             ? int(inner_stride_at_compile_time<XprType>::ret)
+                             : int(outer_stride_at_compile_time<XprType>::ret),
+    OuterStrideAtCompileTime = HasSameStorageOrderAsXprType
+                             ? int(outer_stride_at_compile_time<XprType>::ret)
+                             : int(inner_stride_at_compile_time<XprType>::ret),
+
+    // FIXME, this traits is rather specialized for dense object and it needs to be cleaned further
+    FlagsLvalueBit = is_lvalue<XprType>::value ? LvalueBit : 0,
+    FlagsRowMajorBit = IsRowMajor ? RowMajorBit : 0,
+    Flags = (traits<XprType>::Flags & (DirectAccessBit | (InnerPanel?CompressedAccessBit:0))) | FlagsLvalueBit | FlagsRowMajorBit,
+    // FIXME DirectAccessBit should not be handled by expressions
+    // 
+    // Alignment is needed by MapBase's assertions
+    // We can sefely set it to false here. Internal alignment errors will be detected by an eigen_internal_assert in the respective evaluator
+    Alignment = 0
+  };
+};
+
+template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false,
+         bool HasDirectAccess = internal::has_direct_access<XprType>::ret> class BlockImpl_dense;
+         
+} // end namespace internal
+
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, typename StorageKind> class BlockImpl;
+
+/** \class Block
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a fixed-size or dynamic-size block
+  *
+  * \tparam XprType the type of the expression in which we are taking a block
+  * \tparam BlockRows the number of rows of the block we are taking at compile time (optional)
+  * \tparam BlockCols the number of columns of the block we are taking at compile time (optional)
+  * \tparam InnerPanel is true, if the block maps to a set of rows of a row major matrix or
+  *         to set of columns of a column major matrix (optional). The parameter allows to determine
+  *         at compile time whether aligned access is possible on the block expression.
+  *
+  * This class represents an expression of either a fixed-size or dynamic-size block. It is the return
+  * type of DenseBase::block(Index,Index,Index,Index) and DenseBase::block<int,int>(Index,Index) and
+  * most of the time this is the only way it is used.
+  *
+  * However, if you want to directly maniputate block expressions,
+  * for instance if you want to write a function returning such an expression, you
+  * will need to use this class.
+  *
+  * Here is an example illustrating the dynamic case:
+  * \include class_Block.cpp
+  * Output: \verbinclude class_Block.out
+  *
+  * \note Even though this expression has dynamic size, in the case where \a XprType
+  * has fixed size, this expression inherits a fixed maximal size which means that evaluating
+  * it does not cause a dynamic memory allocation.
+  *
+  * Here is an example illustrating the fixed-size case:
+  * \include class_FixedBlock.cpp
+  * Output: \verbinclude class_FixedBlock.out
+  *
+  * \sa DenseBase::block(Index,Index,Index,Index), DenseBase::block(Index,Index), class VectorBlock
+  */
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel> class Block
+  : public BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, typename internal::traits<XprType>::StorageKind>
+{
+    typedef BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, typename internal::traits<XprType>::StorageKind> Impl;
+  public:
+    //typedef typename Impl::Base Base;
+    typedef Impl Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Block)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Block)
+    
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
+  
+    /** Column or Row constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline Block(XprType& xpr, Index i) : Impl(xpr,i)
+    {
+      eigen_assert( (i>=0) && (
+          ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && i<xpr.rows())
+        ||((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && i<xpr.cols())));
+    }
+
+    /** Fixed-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline Block(XprType& xpr, Index startRow, Index startCol)
+      : Impl(xpr, startRow, startCol)
+    {
+      EIGEN_STATIC_ASSERT(RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic,THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE)
+      eigen_assert(startRow >= 0 && BlockRows >= 0 && startRow + BlockRows <= xpr.rows()
+             && startCol >= 0 && BlockCols >= 0 && startCol + BlockCols <= xpr.cols());
+    }
+
+    /** Dynamic-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline Block(XprType& xpr,
+          Index startRow, Index startCol,
+          Index blockRows, Index blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols)
+    {
+      eigen_assert((RowsAtCompileTime==Dynamic || RowsAtCompileTime==blockRows)
+          && (ColsAtCompileTime==Dynamic || ColsAtCompileTime==blockCols));
+      eigen_assert(startRow >= 0 && blockRows >= 0 && startRow  <= xpr.rows() - blockRows
+          && startCol >= 0 && blockCols >= 0 && startCol <= xpr.cols() - blockCols);
+    }
+};
+         
+// The generic default implementation for dense block simplu forward to the internal::BlockImpl_dense
+// that must be specialized for direct and non-direct access...
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
+class BlockImpl<XprType, BlockRows, BlockCols, InnerPanel, Dense>
+  : public internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel>
+{
+    typedef internal::BlockImpl_dense<XprType, BlockRows, BlockCols, InnerPanel> Impl;
+    typedef typename XprType::StorageIndex StorageIndex;
+  public:
+    typedef Impl Base;
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl)
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index i) : Impl(xpr,i) {}
+    EIGEN_DEVICE_FUNC inline BlockImpl(XprType& xpr, Index startRow, Index startCol) : Impl(xpr, startRow, startCol) {}
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : Impl(xpr, startRow, startCol, blockRows, blockCols) {}
+};
+
+namespace internal {
+
+/** \internal Internal implementation of dense Blocks in the general case. */
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool HasDirectAccess> class BlockImpl_dense
+  : public internal::dense_xpr_base<Block<XprType, BlockRows, BlockCols, InnerPanel> >::type
+{
+    typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested;
+  public:
+
+    typedef typename internal::dense_xpr_base<BlockType>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(BlockType)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense)
+
+    // class InnerIterator; // FIXME apparently never used
+
+    /** Column or Row constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index i)
+      : m_xpr(xpr),
+        // It is a row if and only if BlockRows==1 and BlockCols==XprType::ColsAtCompileTime,
+        // and it is a column if and only if BlockRows==XprType::RowsAtCompileTime and BlockCols==1,
+        // all other cases are invalid.
+        // The case a 1x1 matrix seems ambiguous, but the result is the same anyway.
+        m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0),
+        m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0),
+        m_blockRows(BlockRows==1 ? 1 : xpr.rows()),
+        m_blockCols(BlockCols==1 ? 1 : xpr.cols())
+    {}
+
+    /** Fixed-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
+                    m_blockRows(BlockRows), m_blockCols(BlockCols)
+    {}
+
+    /** Dynamic-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr,
+          Index startRow, Index startCol,
+          Index blockRows, Index blockCols)
+      : m_xpr(xpr), m_startRow(startRow), m_startCol(startCol),
+                    m_blockRows(blockRows), m_blockCols(blockCols)
+    {}
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_blockRows.value(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_blockCols.value(); }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index rowId, Index colId)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(XprType)
+      return m_xpr.coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      return m_xpr.derived().coeffRef(rowId + m_startRow.value(), colId + m_startCol.value());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index rowId, Index colId) const
+    {
+      return m_xpr.coeff(rowId + m_startRow.value(), colId + m_startCol.value());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index index)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(XprType)
+      return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                            m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return m_xpr.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                            m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const CoeffReturnType coeff(Index index) const
+    {
+      return m_xpr.coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                         m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    template<int LoadMode>
+    inline PacketScalar packet(Index rowId, Index colId) const
+    {
+      return m_xpr.template packet<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value());
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index rowId, Index colId, const PacketScalar& val)
+    {
+      m_xpr.template writePacket<Unaligned>(rowId + m_startRow.value(), colId + m_startCol.value(), val);
+    }
+
+    template<int LoadMode>
+    inline PacketScalar packet(Index index) const
+    {
+      return m_xpr.template packet<Unaligned>
+              (m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+               m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index index, const PacketScalar& val)
+    {
+      m_xpr.template writePacket<Unaligned>
+         (m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+          m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0), val);
+    }
+
+    #ifdef EIGEN_PARSED_BY_DOXYGEN
+    /** \sa MapBase::data() */
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const;
+    EIGEN_DEVICE_FUNC inline Index innerStride() const;
+    EIGEN_DEVICE_FUNC inline Index outerStride() const;
+    #endif
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const
+    { 
+      return m_xpr; 
+    }
+
+    EIGEN_DEVICE_FUNC
+    XprType& nestedExpression() { return m_xpr; }
+      
+    EIGEN_DEVICE_FUNC
+    StorageIndex startRow() const
+    { 
+      return m_startRow.value(); 
+    }
+      
+    EIGEN_DEVICE_FUNC
+    StorageIndex startCol() const
+    { 
+      return m_startCol.value(); 
+    }
+
+  protected:
+
+    XprTypeNested m_xpr;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+    const internal::variable_if_dynamic<StorageIndex, RowsAtCompileTime> m_blockRows;
+    const internal::variable_if_dynamic<StorageIndex, ColsAtCompileTime> m_blockCols;
+};
+
+/** \internal Internal implementation of dense Blocks in the direct access case.*/
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
+class BlockImpl_dense<XprType,BlockRows,BlockCols, InnerPanel,true>
+  : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel> >
+{
+    typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    typedef typename internal::ref_selector<XprType>::non_const_type XprTypeNested;
+    enum {
+      XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0
+    };
+  public:
+
+    typedef MapBase<BlockType> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(BlockType)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(BlockImpl_dense)
+
+    /** Column or Row constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index i)
+      : Base(xpr.data() + i * (    ((BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) && (!XprTypeIsRowMajor)) 
+                                || ((BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) && ( XprTypeIsRowMajor)) ? xpr.innerStride() : xpr.outerStride()),
+             BlockRows==1 ? 1 : xpr.rows(),
+             BlockCols==1 ? 1 : xpr.cols()),
+        m_xpr(xpr),
+        m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? i : 0),
+        m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? i : 0)
+    {
+      init();
+    }
+
+    /** Fixed-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, Index startRow, Index startCol)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol)),
+        m_xpr(xpr), m_startRow(startRow), m_startCol(startCol)
+    {
+      init();
+    }
+
+    /** Dynamic-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr,
+          Index startRow, Index startCol,
+          Index blockRows, Index blockCols)
+      : Base(xpr.data()+xpr.innerStride()*(XprTypeIsRowMajor?startCol:startRow) + xpr.outerStride()*(XprTypeIsRowMajor?startRow:startCol), blockRows, blockCols),
+        m_xpr(xpr), m_startRow(startRow), m_startCol(startCol)
+    {
+      init();
+    }
+
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<XprTypeNested>::type& nestedExpression() const
+    { 
+      return m_xpr; 
+    }
+
+    EIGEN_DEVICE_FUNC
+    XprType& nestedExpression() { return m_xpr; }
+      
+    /** \sa MapBase::innerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const
+    {
+      return internal::traits<BlockType>::HasSameStorageOrderAsXprType
+             ? m_xpr.innerStride()
+             : m_xpr.outerStride();
+    }
+
+    /** \sa MapBase::outerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const
+    {
+      return m_outerStride;
+    }
+
+    EIGEN_DEVICE_FUNC
+    StorageIndex startRow() const
+    {
+      return m_startRow.value();
+    }
+
+    EIGEN_DEVICE_FUNC
+    StorageIndex startCol() const
+    {
+      return m_startCol.value();
+    }
+
+  #ifndef __SUNPRO_CC
+  // FIXME sunstudio is not friendly with the above friend...
+  // META-FIXME there is no 'friend' keyword around here. Is this obsolete?
+  protected:
+  #endif
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal used by allowAligned() */
+    EIGEN_DEVICE_FUNC
+    inline BlockImpl_dense(XprType& xpr, const Scalar* data, Index blockRows, Index blockCols)
+      : Base(data, blockRows, blockCols), m_xpr(xpr)
+    {
+      init();
+    }
+    #endif
+
+  protected:
+    EIGEN_DEVICE_FUNC
+    void init()
+    {
+      m_outerStride = internal::traits<BlockType>::HasSameStorageOrderAsXprType
+                    ? m_xpr.outerStride()
+                    : m_xpr.innerStride();
+    }
+
+    XprTypeNested m_xpr;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<StorageIndex, (XprType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+    Index m_outerStride;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BLOCK_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h
new file mode 100644
index 0000000..8409d87
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/BooleanRedux.h
@@ -0,0 +1,164 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ALLANDANY_H
+#define EIGEN_ALLANDANY_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Derived, int UnrollCount>
+struct all_unroller
+{
+  typedef typename Derived::ExpressionTraits Traits;
+  enum {
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
+  };
+
+  static inline bool run(const Derived &mat)
+  {
+    return all_unroller<Derived, UnrollCount-1>::run(mat) && mat.coeff(row, col);
+  }
+};
+
+template<typename Derived>
+struct all_unroller<Derived, 0>
+{
+  static inline bool run(const Derived &/*mat*/) { return true; }
+};
+
+template<typename Derived>
+struct all_unroller<Derived, Dynamic>
+{
+  static inline bool run(const Derived &) { return false; }
+};
+
+template<typename Derived, int UnrollCount>
+struct any_unroller
+{
+  typedef typename Derived::ExpressionTraits Traits;
+  enum {
+    col = (UnrollCount-1) / Traits::RowsAtCompileTime,
+    row = (UnrollCount-1) % Traits::RowsAtCompileTime
+  };
+  
+  static inline bool run(const Derived &mat)
+  {
+    return any_unroller<Derived, UnrollCount-1>::run(mat) || mat.coeff(row, col);
+  }
+};
+
+template<typename Derived>
+struct any_unroller<Derived, 0>
+{
+  static inline bool run(const Derived & /*mat*/) { return false; }
+};
+
+template<typename Derived>
+struct any_unroller<Derived, Dynamic>
+{
+  static inline bool run(const Derived &) { return false; }
+};
+
+} // end namespace internal
+
+/** \returns true if all coefficients are true
+  *
+  * Example: \include MatrixBase_all.cpp
+  * Output: \verbinclude MatrixBase_all.out
+  *
+  * \sa any(), Cwise::operator<()
+  */
+template<typename Derived>
+inline bool DenseBase<Derived>::all() const
+{
+  typedef internal::evaluator<Derived> Evaluator;
+  enum {
+    unroll = SizeAtCompileTime != Dynamic
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+  };
+  Evaluator evaluator(derived());
+  if(unroll)
+    return internal::all_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
+  else
+  {
+    for(Index j = 0; j < cols(); ++j)
+      for(Index i = 0; i < rows(); ++i)
+        if (!evaluator.coeff(i, j)) return false;
+    return true;
+  }
+}
+
+/** \returns true if at least one coefficient is true
+  *
+  * \sa all()
+  */
+template<typename Derived>
+inline bool DenseBase<Derived>::any() const
+{
+  typedef internal::evaluator<Derived> Evaluator;
+  enum {
+    unroll = SizeAtCompileTime != Dynamic
+          && SizeAtCompileTime * (Evaluator::CoeffReadCost + NumTraits<Scalar>::AddCost) <= EIGEN_UNROLLING_LIMIT
+  };
+  Evaluator evaluator(derived());
+  if(unroll)
+    return internal::any_unroller<Evaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(evaluator);
+  else
+  {
+    for(Index j = 0; j < cols(); ++j)
+      for(Index i = 0; i < rows(); ++i)
+        if (evaluator.coeff(i, j)) return true;
+    return false;
+  }
+}
+
+/** \returns the number of coefficients which evaluate to true
+  *
+  * \sa all(), any()
+  */
+template<typename Derived>
+inline Eigen::Index DenseBase<Derived>::count() const
+{
+  return derived().template cast<bool>().template cast<Index>().sum();
+}
+
+/** \returns true is \c *this contains at least one Not A Number (NaN).
+  *
+  * \sa allFinite()
+  */
+template<typename Derived>
+inline bool DenseBase<Derived>::hasNaN() const
+{
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isNaN().any();
+#else
+  return !((derived().array()==derived().array()).all());
+#endif
+}
+
+/** \returns true if \c *this contains only finite numbers, i.e., no NaN and no +/-INF values.
+  *
+  * \sa hasNaN()
+  */
+template<typename Derived>
+inline bool DenseBase<Derived>::allFinite() const
+{
+#if EIGEN_COMP_MSVC || (defined __FAST_MATH__)
+  return derived().array().isFinite().all();
+#else
+  return !((derived()-derived()).hasNaN());
+#endif
+}
+    
+} // end namespace Eigen
+
+#endif // EIGEN_ALLANDANY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h
new file mode 100644
index 0000000..d218e98
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CommaInitializer.h
@@ -0,0 +1,160 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMMAINITIALIZER_H
+#define EIGEN_COMMAINITIALIZER_H
+
+namespace Eigen { 
+
+/** \class CommaInitializer
+  * \ingroup Core_Module
+  *
+  * \brief Helper class used by the comma initializer operator
+  *
+  * This class is internally used to implement the comma initializer feature. It is
+  * the return type of MatrixBase::operator<<, and most of the time this is the only
+  * way it is used.
+  *
+  * \sa \blank \ref MatrixBaseCommaInitRef "MatrixBase::operator<<", CommaInitializer::finished()
+  */
+template<typename XprType>
+struct CommaInitializer
+{
+  typedef typename XprType::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC
+  inline CommaInitializer(XprType& xpr, const Scalar& s)
+    : m_xpr(xpr), m_row(0), m_col(1), m_currentBlockRows(1)
+  {
+    m_xpr.coeffRef(0,0) = s;
+  }
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
+  inline CommaInitializer(XprType& xpr, const DenseBase<OtherDerived>& other)
+    : m_xpr(xpr), m_row(0), m_col(other.cols()), m_currentBlockRows(other.rows())
+  {
+    m_xpr.block(0, 0, other.rows(), other.cols()) = other;
+  }
+
+  /* Copy/Move constructor which transfers ownership. This is crucial in 
+   * absence of return value optimization to avoid assertions during destruction. */
+  // FIXME in C++11 mode this could be replaced by a proper RValue constructor
+  EIGEN_DEVICE_FUNC
+  inline CommaInitializer(const CommaInitializer& o)
+  : m_xpr(o.m_xpr), m_row(o.m_row), m_col(o.m_col), m_currentBlockRows(o.m_currentBlockRows) {
+    // Mark original object as finished. In absence of R-value references we need to const_cast:
+    const_cast<CommaInitializer&>(o).m_row = m_xpr.rows();
+    const_cast<CommaInitializer&>(o).m_col = m_xpr.cols();
+    const_cast<CommaInitializer&>(o).m_currentBlockRows = 0;
+  }
+
+  /* inserts a scalar value in the target matrix */
+  EIGEN_DEVICE_FUNC
+  CommaInitializer& operator,(const Scalar& s)
+  {
+    if (m_col==m_xpr.cols())
+    {
+      m_row+=m_currentBlockRows;
+      m_col = 0;
+      m_currentBlockRows = 1;
+      eigen_assert(m_row<m_xpr.rows()
+        && "Too many rows passed to comma initializer (operator<<)");
+    }
+    eigen_assert(m_col<m_xpr.cols()
+      && "Too many coefficients passed to comma initializer (operator<<)");
+    eigen_assert(m_currentBlockRows==1);
+    m_xpr.coeffRef(m_row, m_col++) = s;
+    return *this;
+  }
+
+  /* inserts a matrix expression in the target matrix */
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
+  CommaInitializer& operator,(const DenseBase<OtherDerived>& other)
+  {
+    if (m_col==m_xpr.cols() && (other.cols()!=0 || other.rows()!=m_currentBlockRows))
+    {
+      m_row+=m_currentBlockRows;
+      m_col = 0;
+      m_currentBlockRows = other.rows();
+      eigen_assert(m_row+m_currentBlockRows<=m_xpr.rows()
+        && "Too many rows passed to comma initializer (operator<<)");
+    }
+    eigen_assert((m_col + other.cols() <= m_xpr.cols())
+      && "Too many coefficients passed to comma initializer (operator<<)");
+    eigen_assert(m_currentBlockRows==other.rows());
+    m_xpr.template block<OtherDerived::RowsAtCompileTime, OtherDerived::ColsAtCompileTime>
+                    (m_row, m_col, other.rows(), other.cols()) = other;
+    m_col += other.cols();
+    return *this;
+  }
+
+  EIGEN_DEVICE_FUNC
+  inline ~CommaInitializer()
+#if defined VERIFY_RAISES_ASSERT && (!defined EIGEN_NO_ASSERTION_CHECKING) && defined EIGEN_EXCEPTIONS
+  EIGEN_EXCEPTION_SPEC(Eigen::eigen_assert_exception)
+#endif
+  {
+      finished();
+  }
+
+  /** \returns the built matrix once all its coefficients have been set.
+    * Calling finished is 100% optional. Its purpose is to write expressions
+    * like this:
+    * \code
+    * quaternion.fromRotationMatrix((Matrix3f() << axis0, axis1, axis2).finished());
+    * \endcode
+    */
+  EIGEN_DEVICE_FUNC
+  inline XprType& finished() {
+      eigen_assert(((m_row+m_currentBlockRows) == m_xpr.rows() || m_xpr.cols() == 0)
+           && m_col == m_xpr.cols()
+           && "Too few coefficients passed to comma initializer (operator<<)");
+      return m_xpr;
+  }
+
+  XprType& m_xpr;           // target expression
+  Index m_row;              // current row id
+  Index m_col;              // current col id
+  Index m_currentBlockRows; // current block height
+};
+
+/** \anchor MatrixBaseCommaInitRef
+  * Convenient operator to set the coefficients of a matrix.
+  *
+  * The coefficients must be provided in a row major order and exactly match
+  * the size of the matrix. Otherwise an assertion is raised.
+  *
+  * Example: \include MatrixBase_set.cpp
+  * Output: \verbinclude MatrixBase_set.out
+  * 
+  * \note According the c++ standard, the argument expressions of this comma initializer are evaluated in arbitrary order.
+  *
+  * \sa CommaInitializer::finished(), class CommaInitializer
+  */
+template<typename Derived>
+inline CommaInitializer<Derived> DenseBase<Derived>::operator<< (const Scalar& s)
+{
+  return CommaInitializer<Derived>(*static_cast<Derived*>(this), s);
+}
+
+/** \sa operator<<(const Scalar&) */
+template<typename Derived>
+template<typename OtherDerived>
+inline CommaInitializer<Derived>
+DenseBase<Derived>::operator<<(const DenseBase<OtherDerived>& other)
+{
+  return CommaInitializer<Derived>(*static_cast<Derived *>(this), other);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMMAINITIALIZER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h
new file mode 100644
index 0000000..aa7efdc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ConditionEstimator.h
@@ -0,0 +1,175 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Rasmus Munk Larsen (rmlarsen@google.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONDITIONESTIMATOR_H
+#define EIGEN_CONDITIONESTIMATOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <typename Vector, typename RealVector, bool IsComplex>
+struct rcond_compute_sign {
+  static inline Vector run(const Vector& v) {
+    const RealVector v_abs = v.cwiseAbs();
+    return (v_abs.array() == static_cast<typename Vector::RealScalar>(0))
+            .select(Vector::Ones(v.size()), v.cwiseQuotient(v_abs));
+  }
+};
+
+// Partial specialization to avoid elementwise division for real vectors.
+template <typename Vector>
+struct rcond_compute_sign<Vector, Vector, false> {
+  static inline Vector run(const Vector& v) {
+    return (v.array() < static_cast<typename Vector::RealScalar>(0))
+           .select(-Vector::Ones(v.size()), Vector::Ones(v.size()));
+  }
+};
+
+/**
+  * \returns an estimate of ||inv(matrix)||_1 given a decomposition of
+  * \a matrix that implements .solve() and .adjoint().solve() methods.
+  *
+  * This function implements Algorithms 4.1 and 5.1 from
+  *   http://www.maths.manchester.ac.uk/~higham/narep/narep135.pdf
+  * which also forms the basis for the condition number estimators in
+  * LAPACK. Since at most 10 calls to the solve method of dec are
+  * performed, the total cost is O(dims^2), as opposed to O(dims^3)
+  * needed to compute the inverse matrix explicitly.
+  *
+  * The most common usage is in estimating the condition number
+  * ||matrix||_1 * ||inv(matrix)||_1. The first term ||matrix||_1 can be
+  * computed directly in O(n^2) operations.
+  *
+  * Supports the following decompositions: FullPivLU, PartialPivLU, LDLT, and
+  * LLT.
+  *
+  * \sa FullPivLU, PartialPivLU, LDLT, LLT.
+  */
+template <typename Decomposition>
+typename Decomposition::RealScalar rcond_invmatrix_L1_norm_estimate(const Decomposition& dec)
+{
+  typedef typename Decomposition::MatrixType MatrixType;
+  typedef typename Decomposition::Scalar Scalar;
+  typedef typename Decomposition::RealScalar RealScalar;
+  typedef typename internal::plain_col_type<MatrixType>::type Vector;
+  typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVector;
+  const bool is_complex = (NumTraits<Scalar>::IsComplex != 0);
+
+  eigen_assert(dec.rows() == dec.cols());
+  const Index n = dec.rows();
+  if (n == 0)
+    return 0;
+
+  // Disable Index to float conversion warning
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  Vector v = dec.solve(Vector::Ones(n) / Scalar(n));
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
+
+  // lower_bound is a lower bound on
+  //   ||inv(matrix)||_1  = sup_v ||inv(matrix) v||_1 / ||v||_1
+  // and is the objective maximized by the ("super-") gradient ascent
+  // algorithm below.
+  RealScalar lower_bound = v.template lpNorm<1>();
+  if (n == 1)
+    return lower_bound;
+
+  // Gradient ascent algorithm follows: We know that the optimum is achieved at
+  // one of the simplices v = e_i, so in each iteration we follow a
+  // super-gradient to move towards the optimal one.
+  RealScalar old_lower_bound = lower_bound;
+  Vector sign_vector(n);
+  Vector old_sign_vector;
+  Index v_max_abs_index = -1;
+  Index old_v_max_abs_index = v_max_abs_index;
+  for (int k = 0; k < 4; ++k)
+  {
+    sign_vector = internal::rcond_compute_sign<Vector, RealVector, is_complex>::run(v);
+    if (k > 0 && !is_complex && sign_vector == old_sign_vector) {
+      // Break if the solution stagnated.
+      break;
+    }
+    // v_max_abs_index = argmax |real( inv(matrix)^T * sign_vector )|
+    v = dec.adjoint().solve(sign_vector);
+    v.real().cwiseAbs().maxCoeff(&v_max_abs_index);
+    if (v_max_abs_index == old_v_max_abs_index) {
+      // Break if the solution stagnated.
+      break;
+    }
+    // Move to the new simplex e_j, where j = v_max_abs_index.
+    v = dec.solve(Vector::Unit(n, v_max_abs_index));  // v = inv(matrix) * e_j.
+    lower_bound = v.template lpNorm<1>();
+    if (lower_bound <= old_lower_bound) {
+      // Break if the gradient step did not increase the lower_bound.
+      break;
+    }
+    if (!is_complex) {
+      old_sign_vector = sign_vector;
+    }
+    old_v_max_abs_index = v_max_abs_index;
+    old_lower_bound = lower_bound;
+  }
+  // The following calculates an independent estimate of ||matrix||_1 by
+  // multiplying matrix by a vector with entries of slowly increasing
+  // magnitude and alternating sign:
+  //   v_i = (-1)^{i} (1 + (i / (dim-1))), i = 0,...,dim-1.
+  // This improvement to Hager's algorithm above is due to Higham. It was
+  // added to make the algorithm more robust in certain corner cases where
+  // large elements in the matrix might otherwise escape detection due to
+  // exact cancellation (especially when op and op_adjoint correspond to a
+  // sequence of backsubstitutions and permutations), which could cause
+  // Hager's algorithm to vastly underestimate ||matrix||_1.
+  Scalar alternating_sign(RealScalar(1));
+  for (Index i = 0; i < n; ++i) {
+    // The static_cast is needed when Scalar is a complex and RealScalar implements expression templates
+    v[i] = alternating_sign * static_cast<RealScalar>(RealScalar(1) + (RealScalar(i) / (RealScalar(n - 1))));
+    alternating_sign = -alternating_sign;
+  }
+  v = dec.solve(v);
+  const RealScalar alternate_lower_bound = (2 * v.template lpNorm<1>()) / (3 * RealScalar(n));
+  return numext::maxi(lower_bound, alternate_lower_bound);
+}
+
+/** \brief Reciprocal condition number estimator.
+  *
+  * Computing a decomposition of a dense matrix takes O(n^3) operations, while
+  * this method estimates the condition number quickly and reliably in O(n^2)
+  * operations.
+  *
+  * \returns an estimate of the reciprocal condition number
+  * (1 / (||matrix||_1 * ||inv(matrix)||_1)) of matrix, given ||matrix||_1 and
+  * its decomposition. Supports the following decompositions: FullPivLU,
+  * PartialPivLU, LDLT, and LLT.
+  *
+  * \sa FullPivLU, PartialPivLU, LDLT, LLT.
+  */
+template <typename Decomposition>
+typename Decomposition::RealScalar
+rcond_estimate_helper(typename Decomposition::RealScalar matrix_norm, const Decomposition& dec)
+{
+  typedef typename Decomposition::RealScalar RealScalar;
+  eigen_assert(dec.rows() == dec.cols());
+  if (dec.rows() == 0)              return RealScalar(1);
+  if (matrix_norm == RealScalar(0)) return RealScalar(0);
+  if (dec.rows() == 1)              return RealScalar(1);
+  const RealScalar inverse_matrix_norm = rcond_invmatrix_L1_norm_estimate(dec);
+  return (inverse_matrix_norm == RealScalar(0) ? RealScalar(0)
+                                               : (RealScalar(1) / inverse_matrix_norm) / matrix_norm);
+}
+
+}  // namespace internal
+
+}  // namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h
new file mode 100644
index 0000000..f7c1eff
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreEvaluators.h
@@ -0,0 +1,1671 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011-2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_COREEVALUATORS_H
+#define EIGEN_COREEVALUATORS_H
+
+namespace Eigen {
+  
+namespace internal {
+
+// This class returns the evaluator kind from the expression storage kind.
+// Default assumes index based accessors
+template<typename StorageKind>
+struct storage_kind_to_evaluator_kind {
+  typedef IndexBased Kind;
+};
+
+// This class returns the evaluator shape from the expression storage kind.
+// It can be Dense, Sparse, Triangular, Diagonal, SelfAdjoint, Band, etc.
+template<typename StorageKind> struct storage_kind_to_shape;
+
+template<> struct storage_kind_to_shape<Dense>                  { typedef DenseShape Shape;           };
+template<> struct storage_kind_to_shape<SolverStorage>          { typedef SolverShape Shape;           };
+template<> struct storage_kind_to_shape<PermutationStorage>     { typedef PermutationShape Shape;     };
+template<> struct storage_kind_to_shape<TranspositionsStorage>  { typedef TranspositionsShape Shape;  };
+
+// Evaluators have to be specialized with respect to various criteria such as:
+//  - storage/structure/shape
+//  - scalar type
+//  - etc.
+// Therefore, we need specialization of evaluator providing additional template arguments for each kind of evaluators.
+// We currently distinguish the following kind of evaluators:
+// - unary_evaluator    for expressions taking only one arguments (CwiseUnaryOp, CwiseUnaryView, Transpose, MatrixWrapper, ArrayWrapper, Reverse, Replicate)
+// - binary_evaluator   for expression taking two arguments (CwiseBinaryOp)
+// - ternary_evaluator   for expression taking three arguments (CwiseTernaryOp)
+// - product_evaluator  for linear algebra products (Product); special case of binary_evaluator because it requires additional tags for dispatching.
+// - mapbase_evaluator  for Map, Block, Ref
+// - block_evaluator    for Block (special dispatching to a mapbase_evaluator or unary_evaluator)
+
+template< typename T,
+          typename Arg1Kind   = typename evaluator_traits<typename T::Arg1>::Kind,
+          typename Arg2Kind   = typename evaluator_traits<typename T::Arg2>::Kind,
+          typename Arg3Kind   = typename evaluator_traits<typename T::Arg3>::Kind,
+          typename Arg1Scalar = typename traits<typename T::Arg1>::Scalar,
+          typename Arg2Scalar = typename traits<typename T::Arg2>::Scalar,
+          typename Arg3Scalar = typename traits<typename T::Arg3>::Scalar> struct ternary_evaluator;
+
+template< typename T,
+          typename LhsKind   = typename evaluator_traits<typename T::Lhs>::Kind,
+          typename RhsKind   = typename evaluator_traits<typename T::Rhs>::Kind,
+          typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
+          typename RhsScalar = typename traits<typename T::Rhs>::Scalar> struct binary_evaluator;
+
+template< typename T,
+          typename Kind   = typename evaluator_traits<typename T::NestedExpression>::Kind,
+          typename Scalar = typename T::Scalar> struct unary_evaluator;
+          
+// evaluator_traits<T> contains traits for evaluator<T> 
+
+template<typename T>
+struct evaluator_traits_base
+{
+  // by default, get evaluator kind and shape from storage
+  typedef typename storage_kind_to_evaluator_kind<typename traits<T>::StorageKind>::Kind Kind;
+  typedef typename storage_kind_to_shape<typename traits<T>::StorageKind>::Shape Shape;
+};
+
+// Default evaluator traits
+template<typename T>
+struct evaluator_traits : public evaluator_traits_base<T>
+{
+};
+
+template<typename T, typename Shape = typename evaluator_traits<T>::Shape >
+struct evaluator_assume_aliasing {
+  static const bool value = false;
+};
+
+// By default, we assume a unary expression:
+template<typename T>
+struct evaluator : public unary_evaluator<T>
+{
+  typedef unary_evaluator<T> Base;
+  EIGEN_DEVICE_FUNC explicit evaluator(const T& xpr) : Base(xpr) {}
+};
+
+
+// TODO: Think about const-correctness
+template<typename T>
+struct evaluator<const T>
+  : evaluator<T>
+{
+  EIGEN_DEVICE_FUNC
+  explicit evaluator(const T& xpr) : evaluator<T>(xpr) {}
+};
+
+// ---------- base class for all evaluators ----------
+
+template<typename ExpressionType>
+struct evaluator_base : public noncopyable
+{
+  // TODO that's not very nice to have to propagate all these traits. They are currently only needed to handle outer,inner indices.
+  typedef traits<ExpressionType> ExpressionTraits;
+  
+  enum {
+    Alignment = 0
+  };
+};
+
+// -------------------- Matrix and Array --------------------
+//
+// evaluator<PlainObjectBase> is a common base class for the
+// Matrix and Array evaluators.
+// Here we directly specialize evaluator. This is not really a unary expression, and it is, by definition, dense,
+// so no need for more sophisticated dispatching.
+
+template<typename Derived>
+struct evaluator<PlainObjectBase<Derived> >
+  : evaluator_base<Derived>
+{
+  typedef PlainObjectBase<Derived> PlainObjectType;
+  typedef typename PlainObjectType::Scalar Scalar;
+  typedef typename PlainObjectType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    IsRowMajor = PlainObjectType::IsRowMajor,
+    IsVectorAtCompileTime = PlainObjectType::IsVectorAtCompileTime,
+    RowsAtCompileTime = PlainObjectType::RowsAtCompileTime,
+    ColsAtCompileTime = PlainObjectType::ColsAtCompileTime,
+    
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    Flags = traits<Derived>::EvaluatorFlags,
+    Alignment = traits<Derived>::Alignment
+  };
+  
+  EIGEN_DEVICE_FUNC evaluator()
+    : m_data(0),
+      m_outerStride(IsVectorAtCompileTime  ? 0 
+                                           : int(IsRowMajor) ? ColsAtCompileTime 
+                                           : RowsAtCompileTime)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const PlainObjectType& m)
+    : m_data(m.data()), m_outerStride(IsVectorAtCompileTime ? 0 : m.outerStride()) 
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    if (IsRowMajor)
+      return m_data[row * m_outerStride.value() + col];
+    else
+      return m_data[row + col * m_outerStride.value()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_data[index];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    if (IsRowMajor)
+      return const_cast<Scalar*>(m_data)[row * m_outerStride.value() + col];
+    else
+      return const_cast<Scalar*>(m_data)[row + col * m_outerStride.value()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return const_cast<Scalar*>(m_data)[index];
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    if (IsRowMajor)
+      return ploadt<PacketType, LoadMode>(m_data + row * m_outerStride.value() + col);
+    else
+      return ploadt<PacketType, LoadMode>(m_data + row + col * m_outerStride.value());
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return ploadt<PacketType, LoadMode>(m_data + index);
+  }
+
+  template<int StoreMode,typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    if (IsRowMajor)
+      return pstoret<Scalar, PacketType, StoreMode>
+	            (const_cast<Scalar*>(m_data) + row * m_outerStride.value() + col, x);
+    else
+      return pstoret<Scalar, PacketType, StoreMode>
+                    (const_cast<Scalar*>(m_data) + row + col * m_outerStride.value(), x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    return pstoret<Scalar, PacketType, StoreMode>(const_cast<Scalar*>(m_data) + index, x);
+  }
+
+protected:
+  const Scalar *m_data;
+
+  // We do not need to know the outer stride for vectors
+  variable_if_dynamic<Index, IsVectorAtCompileTime  ? 0 
+                                                    : int(IsRowMajor) ? ColsAtCompileTime 
+                                                    : RowsAtCompileTime> m_outerStride;
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct evaluator<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+  : evaluator<PlainObjectBase<Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > >
+{
+  typedef Matrix<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType;
+  
+  EIGEN_DEVICE_FUNC evaluator() {}
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m)
+    : evaluator<PlainObjectBase<XprType> >(m) 
+  { }
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct evaluator<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+  : evaluator<PlainObjectBase<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> > >
+{
+  typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> XprType;
+
+  EIGEN_DEVICE_FUNC evaluator() {}
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& m)
+    : evaluator<PlainObjectBase<XprType> >(m) 
+  { }
+};
+
+// -------------------- Transpose --------------------
+
+template<typename ArgType>
+struct unary_evaluator<Transpose<ArgType>, IndexBased>
+  : evaluator_base<Transpose<ArgType> >
+{
+  typedef Transpose<ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,    
+    Flags = evaluator<ArgType>::Flags ^ RowMajorBit,
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& t) : m_argImpl(t.nestedExpression()) {}
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(col, row);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(col, row);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  typename XprType::Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(col, row);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(index);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode,PacketType>(col, row, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode,PacketType>(index, x);
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+};
+
+// -------------------- CwiseNullaryOp --------------------
+// Like Matrix and Array, this is not really a unary expression, so we directly specialize evaluator.
+// Likewise, there is not need to more sophisticated dispatching here.
+
+template<typename Scalar,typename NullaryOp,
+         bool has_nullary = has_nullary_operator<NullaryOp>::value,
+         bool has_unary   = has_unary_operator<NullaryOp>::value,
+         bool has_binary  = has_binary_operator<NullaryOp>::value>
+struct nullary_wrapper
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const { return op(i,j); }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const { return op.template packetOp<T>(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,false,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType=0, IndexType=0) const { return op(); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType=0, IndexType=0) const { return op.template packetOp<T>(); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j=0) const { return op(i,j); }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j=0) const { return op.template packetOp<T>(i,j); }
+};
+
+// We need the following specialization for vector-only functors assigned to a runtime vector,
+// for instance, using linspace and assigning a RowVectorXd to a MatrixXd or even a row of a MatrixXd.
+// In this case, i==0 and j is used for the actual iteration.
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,true,false>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op(i+j);
+  }
+  template <typename T, typename IndexType> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    eigen_assert(i==0 || j==0);
+    return op.template packetOp<T>(i+j);
+  }
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const { return op(i); }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const { return op.template packetOp<T>(i); }
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,false,false,false> {};
+
+#if 0 && EIGEN_COMP_MSVC>0
+// Disable this ugly workaround. This is now handled in traits<Ref>::match,
+// but this piece of code might still become handly if some other weird compilation
+// erros pop up again.
+
+// MSVC exhibits a weird compilation error when
+// compiling:
+//    Eigen::MatrixXf A = MatrixXf::Random(3,3);
+//    Ref<const MatrixXf> R = 2.f*A;
+// and that has_*ary_operator<scalar_constant_op<float>> have not been instantiated yet.
+// The "problem" is that evaluator<2.f*A> is instantiated by traits<Ref>::match<2.f*A>
+// and at that time has_*ary_operator<T> returns true regardless of T.
+// Then nullary_wrapper is badly instantiated as nullary_wrapper<.,.,true,true,true>.
+// The trick is thus to defer the proper instantiation of nullary_wrapper when coeff(),
+// and packet() are really instantiated as implemented below:
+
+// This is a simple wrapper around Index to enforce the re-instantiation of
+// has_*ary_operator when needed.
+template<typename T> struct nullary_wrapper_workaround_msvc {
+  nullary_wrapper_workaround_msvc(const T&);
+  operator T()const;
+};
+
+template<typename Scalar,typename NullaryOp>
+struct nullary_wrapper<Scalar,NullaryOp,true,true,true>
+{
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i,j);
+  }
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar operator()(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().operator()(op,i);
+  }
+
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i, IndexType j) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i,j);
+  }
+  template <typename T, typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T packetOp(const NullaryOp& op, IndexType i) const {
+    return nullary_wrapper<Scalar,NullaryOp,
+    has_nullary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_unary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value,
+    has_binary_operator<NullaryOp,nullary_wrapper_workaround_msvc<IndexType> >::value>().template packetOp<T>(op,i);
+  }
+};
+#endif // MSVC workaround
+
+template<typename NullaryOp, typename PlainObjectType>
+struct evaluator<CwiseNullaryOp<NullaryOp,PlainObjectType> >
+  : evaluator_base<CwiseNullaryOp<NullaryOp,PlainObjectType> >
+{
+  typedef CwiseNullaryOp<NullaryOp,PlainObjectType> XprType;
+  typedef typename internal::remove_all<PlainObjectType>::type PlainObjectTypeCleaned;
+  
+  enum {
+    CoeffReadCost = internal::functor_traits<NullaryOp>::Cost,
+    
+    Flags = (evaluator<PlainObjectTypeCleaned>::Flags
+          &  (  HereditaryBits
+              | (functor_has_linear_access<NullaryOp>::ret  ? LinearAccessBit : 0)
+              | (functor_traits<NullaryOp>::PacketAccess    ? PacketAccessBit : 0)))
+          | (functor_traits<NullaryOp>::IsRepeatable ? 0 : EvalBeforeNestingBit),
+    Alignment = AlignedMax
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& n)
+    : m_functor(n.functor()), m_wrapper()
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(IndexType row, IndexType col) const
+  {
+    return m_wrapper(m_functor, row, col);
+  }
+
+  template <typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(IndexType index) const
+  {
+    return m_wrapper(m_functor,index);
+  }
+
+  template<int LoadMode, typename PacketType, typename IndexType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(IndexType row, IndexType col) const
+  {
+    return m_wrapper.template packetOp<PacketType>(m_functor, row, col);
+  }
+
+  template<int LoadMode, typename PacketType, typename IndexType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(IndexType index) const
+  {
+    return m_wrapper.template packetOp<PacketType>(m_functor, index);
+  }
+
+protected:
+  const NullaryOp m_functor;
+  const internal::nullary_wrapper<CoeffReturnType,NullaryOp> m_wrapper;
+};
+
+// -------------------- CwiseUnaryOp --------------------
+
+template<typename UnaryOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryOp<UnaryOp, ArgType>, IndexBased >
+  : evaluator_base<CwiseUnaryOp<UnaryOp, ArgType> >
+{
+  typedef CwiseUnaryOp<UnaryOp, ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+    
+    Flags = evaluator<ArgType>::Flags
+          & (HereditaryBits | LinearAccessBit | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit unary_evaluator(const XprType& op)
+    : m_functor(op.functor()), 
+      m_argImpl(op.nestedExpression()) 
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_functor(m_argImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_argImpl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_functor.packetOp(m_argImpl.template packet<LoadMode, PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_functor.packetOp(m_argImpl.template packet<LoadMode, PacketType>(index));
+  }
+
+protected:
+  const UnaryOp m_functor;
+  evaluator<ArgType> m_argImpl;
+};
+
+// -------------------- CwiseTernaryOp --------------------
+
+// this is a ternary expression
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+  : public ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+{
+  typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType;
+  typedef ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct ternary_evaluator<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3>, IndexBased, IndexBased>
+  : evaluator_base<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >
+{
+  typedef CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<Arg1>::CoeffReadCost + evaluator<Arg2>::CoeffReadCost + evaluator<Arg3>::CoeffReadCost + functor_traits<TernaryOp>::Cost,
+    
+    Arg1Flags = evaluator<Arg1>::Flags,
+    Arg2Flags = evaluator<Arg2>::Flags,
+    Arg3Flags = evaluator<Arg3>::Flags,
+    SameType = is_same<typename Arg1::Scalar,typename Arg2::Scalar>::value && is_same<typename Arg1::Scalar,typename Arg3::Scalar>::value,
+    StorageOrdersAgree = (int(Arg1Flags)&RowMajorBit)==(int(Arg2Flags)&RowMajorBit) && (int(Arg1Flags)&RowMajorBit)==(int(Arg3Flags)&RowMajorBit),
+    Flags0 = (int(Arg1Flags) | int(Arg2Flags) | int(Arg3Flags)) & (
+        HereditaryBits
+        | (int(Arg1Flags) & int(Arg2Flags) & int(Arg3Flags) &
+           ( (StorageOrdersAgree ? LinearAccessBit : 0)
+           | (functor_traits<TernaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
+           )
+        )
+     ),
+    Flags = (Flags0 & ~RowMajorBit) | (Arg1Flags & RowMajorBit),
+    Alignment = EIGEN_PLAIN_ENUM_MIN(
+        EIGEN_PLAIN_ENUM_MIN(evaluator<Arg1>::Alignment, evaluator<Arg2>::Alignment),
+        evaluator<Arg3>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit ternary_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_arg1Impl(xpr.arg1()), 
+      m_arg2Impl(xpr.arg2()), 
+      m_arg3Impl(xpr.arg3())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<TernaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_functor(m_arg1Impl.coeff(row, col), m_arg2Impl.coeff(row, col), m_arg3Impl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_arg1Impl.coeff(index), m_arg2Impl.coeff(index), m_arg3Impl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_functor.packetOp(m_arg1Impl.template packet<LoadMode,PacketType>(row, col),
+                              m_arg2Impl.template packet<LoadMode,PacketType>(row, col),
+                              m_arg3Impl.template packet<LoadMode,PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_functor.packetOp(m_arg1Impl.template packet<LoadMode,PacketType>(index),
+                              m_arg2Impl.template packet<LoadMode,PacketType>(index),
+                              m_arg3Impl.template packet<LoadMode,PacketType>(index));
+  }
+
+protected:
+  const TernaryOp m_functor;
+  evaluator<Arg1> m_arg1Impl;
+  evaluator<Arg2> m_arg2Impl;
+  evaluator<Arg3> m_arg3Impl;
+};
+
+// -------------------- CwiseBinaryOp --------------------
+
+// this is a binary expression
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+  : public binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  typedef binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IndexBased>
+  : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    
+    LhsFlags = evaluator<Lhs>::Flags,
+    RhsFlags = evaluator<Rhs>::Flags,
+    SameType = is_same<typename Lhs::Scalar,typename Rhs::Scalar>::value,
+    StorageOrdersAgree = (int(LhsFlags)&RowMajorBit)==(int(RhsFlags)&RowMajorBit),
+    Flags0 = (int(LhsFlags) | int(RhsFlags)) & (
+        HereditaryBits
+      | (int(LhsFlags) & int(RhsFlags) &
+           ( (StorageOrdersAgree ? LinearAccessBit : 0)
+           | (functor_traits<BinaryOp>::PacketAccess && StorageOrdersAgree && SameType ? PacketAccessBit : 0)
+           )
+        )
+     ),
+    Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<Lhs>::Alignment,evaluator<Rhs>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit binary_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()), 
+      m_rhsImpl(xpr.rhs())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_functor(m_lhsImpl.coeff(row, col), m_rhsImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_functor(m_lhsImpl.coeff(index), m_rhsImpl.coeff(index));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_functor.packetOp(m_lhsImpl.template packet<LoadMode,PacketType>(row, col),
+                              m_rhsImpl.template packet<LoadMode,PacketType>(row, col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_functor.packetOp(m_lhsImpl.template packet<LoadMode,PacketType>(index),
+                              m_rhsImpl.template packet<LoadMode,PacketType>(index));
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<Lhs> m_lhsImpl;
+  evaluator<Rhs> m_rhsImpl;
+};
+
+// -------------------- CwiseUnaryView --------------------
+
+template<typename UnaryOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryView<UnaryOp, ArgType>, IndexBased>
+  : evaluator_base<CwiseUnaryView<UnaryOp, ArgType> >
+{
+  typedef CwiseUnaryView<UnaryOp, ArgType> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+    
+    Flags = (evaluator<ArgType>::Flags & (HereditaryBits | LinearAccessBit | DirectAccessBit)),
+    
+    Alignment = 0 // FIXME it is not very clear why alignment is necessarily lost...
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op)
+    : m_unaryOp(op.functor()), 
+      m_argImpl(op.nestedExpression()) 
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_unaryOp(m_argImpl.coeff(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_unaryOp(m_argImpl.coeff(index));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_unaryOp(m_argImpl.coeffRef(row, col));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_unaryOp(m_argImpl.coeffRef(index));
+  }
+
+protected:
+  const UnaryOp m_unaryOp;
+  evaluator<ArgType> m_argImpl;
+};
+
+// -------------------- Map --------------------
+
+// FIXME perhaps the PlainObjectType could be provided by Derived::PlainObject ?
+// but that might complicate template specialization
+template<typename Derived, typename PlainObjectType>
+struct mapbase_evaluator;
+
+template<typename Derived, typename PlainObjectType>
+struct mapbase_evaluator : evaluator_base<Derived>
+{
+  typedef Derived  XprType;
+  typedef typename XprType::PointerType PointerType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    IsRowMajor = XprType::RowsAtCompileTime,
+    ColsAtCompileTime = XprType::ColsAtCompileTime,
+    CoeffReadCost = NumTraits<Scalar>::ReadCost
+  };
+
+  EIGEN_DEVICE_FUNC explicit mapbase_evaluator(const XprType& map)
+    : m_data(const_cast<PointerType>(map.data())),
+      m_innerStride(map.innerStride()),
+      m_outerStride(map.outerStride())
+  {
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(evaluator<Derived>::Flags&PacketAccessBit, internal::inner_stride_at_compile_time<Derived>::ret==1),
+                        PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_data[col * colStride() + row * rowStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_data[index * m_innerStride.value()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_data[col * colStride() + row * rowStride()];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_data[index * m_innerStride.value()];
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
+    return internal::ploadt<PacketType, LoadMode>(ptr);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return internal::ploadt<PacketType, LoadMode>(m_data + index * m_innerStride.value());
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    PointerType ptr = m_data + row * rowStride() + col * colStride();
+    return internal::pstoret<Scalar, PacketType, StoreMode>(ptr, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    internal::pstoret<Scalar, PacketType, StoreMode>(m_data + index * m_innerStride.value(), x);
+  }
+protected:
+  EIGEN_DEVICE_FUNC
+  inline Index rowStride() const { return XprType::IsRowMajor ? m_outerStride.value() : m_innerStride.value(); }
+  EIGEN_DEVICE_FUNC
+  inline Index colStride() const { return XprType::IsRowMajor ? m_innerStride.value() : m_outerStride.value(); }
+
+  PointerType m_data;
+  const internal::variable_if_dynamic<Index, XprType::InnerStrideAtCompileTime> m_innerStride;
+  const internal::variable_if_dynamic<Index, XprType::OuterStrideAtCompileTime> m_outerStride;
+};
+
+template<typename PlainObjectType, int MapOptions, typename StrideType> 
+struct evaluator<Map<PlainObjectType, MapOptions, StrideType> >
+  : public mapbase_evaluator<Map<PlainObjectType, MapOptions, StrideType>, PlainObjectType>
+{
+  typedef Map<PlainObjectType, MapOptions, StrideType> XprType;
+  typedef typename XprType::Scalar Scalar;
+  // TODO: should check for smaller packet types once we can handle multi-sized packet types
+  typedef typename packet_traits<Scalar>::type PacketScalar;
+  
+  enum {
+    InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
+                             ? int(PlainObjectType::InnerStrideAtCompileTime)
+                             : int(StrideType::InnerStrideAtCompileTime),
+    OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
+                             ? int(PlainObjectType::OuterStrideAtCompileTime)
+                             : int(StrideType::OuterStrideAtCompileTime),
+    HasNoInnerStride = InnerStrideAtCompileTime == 1,
+    HasNoOuterStride = StrideType::OuterStrideAtCompileTime == 0,
+    HasNoStride = HasNoInnerStride && HasNoOuterStride,
+    IsDynamicSize = PlainObjectType::SizeAtCompileTime==Dynamic,
+    
+    PacketAccessMask = bool(HasNoInnerStride) ? ~int(0) : ~int(PacketAccessBit),
+    LinearAccessMask = bool(HasNoStride) || bool(PlainObjectType::IsVectorAtCompileTime) ? ~int(0) : ~int(LinearAccessBit),
+    Flags = int( evaluator<PlainObjectType>::Flags) & (LinearAccessMask&PacketAccessMask),
+    
+    Alignment = int(MapOptions)&int(AlignedMask)
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& map)
+    : mapbase_evaluator<XprType, PlainObjectType>(map) 
+  { }
+};
+
+// -------------------- Ref --------------------
+
+template<typename PlainObjectType, int RefOptions, typename StrideType> 
+struct evaluator<Ref<PlainObjectType, RefOptions, StrideType> >
+  : public mapbase_evaluator<Ref<PlainObjectType, RefOptions, StrideType>, PlainObjectType>
+{
+  typedef Ref<PlainObjectType, RefOptions, StrideType> XprType;
+  
+  enum {
+    Flags = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Flags,
+    Alignment = evaluator<Map<PlainObjectType, RefOptions, StrideType> >::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& ref)
+    : mapbase_evaluator<XprType, PlainObjectType>(ref) 
+  { }
+};
+
+// -------------------- Block --------------------
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel,
+         bool HasDirectAccess = internal::has_direct_access<ArgType>::ret> struct block_evaluator;
+         
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> 
+struct evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+  : block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel>
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+  typedef typename XprType::Scalar Scalar;
+  // TODO: should check for smaller packet types once we can handle multi-sized packet types
+  typedef typename packet_traits<Scalar>::type PacketScalar;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    RowsAtCompileTime = traits<XprType>::RowsAtCompileTime,
+    ColsAtCompileTime = traits<XprType>::ColsAtCompileTime,
+    MaxRowsAtCompileTime = traits<XprType>::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = traits<XprType>::MaxColsAtCompileTime,
+    
+    ArgTypeIsRowMajor = (int(evaluator<ArgType>::Flags)&RowMajorBit) != 0,
+    IsRowMajor = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? 1
+               : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+               : ArgTypeIsRowMajor,
+    HasSameStorageOrderAsArgType = (IsRowMajor == ArgTypeIsRowMajor),
+    InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
+    InnerStrideAtCompileTime = HasSameStorageOrderAsArgType
+                             ? int(inner_stride_at_compile_time<ArgType>::ret)
+                             : int(outer_stride_at_compile_time<ArgType>::ret),
+    OuterStrideAtCompileTime = HasSameStorageOrderAsArgType
+                             ? int(outer_stride_at_compile_time<ArgType>::ret)
+                             : int(inner_stride_at_compile_time<ArgType>::ret),
+    MaskPacketAccessBit = (InnerStrideAtCompileTime == 1) ? PacketAccessBit : 0,
+    
+    FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1 || (InnerPanel && (evaluator<ArgType>::Flags&LinearAccessBit))) ? LinearAccessBit : 0,    
+    FlagsRowMajorBit = XprType::Flags&RowMajorBit,
+    Flags0 = evaluator<ArgType>::Flags & ( (HereditaryBits & ~RowMajorBit) |
+                                           DirectAccessBit |
+                                           MaskPacketAccessBit),
+    Flags = Flags0 | FlagsLinearAccessBit | FlagsRowMajorBit,
+    
+    PacketAlignment = unpacket_traits<PacketScalar>::alignment,
+    Alignment0 = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && (((OuterStrideAtCompileTime * int(sizeof(Scalar))) % int(PacketAlignment)) == 0)) ? int(PacketAlignment) : 0,
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ArgType>::Alignment, Alignment0)
+  };
+  typedef block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel> block_evaluator_type;
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& block) : block_evaluator_type(block)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+};
+
+// no direct-access => dispatch to a unary evaluator
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /*HasDirectAccess*/ false>
+  : unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+
+  EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
+    : unary_evaluator<XprType>(block) 
+  {}
+};
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+struct unary_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>, IndexBased>
+  : evaluator_base<Block<ArgType, BlockRows, BlockCols, InnerPanel> >
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& block)
+    : m_argImpl(block.nestedExpression()), 
+      m_startRow(block.startRow()), 
+      m_startCol(block.startCol()) 
+  { }
+ 
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime
+  };
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  { 
+    return m_argImpl.coeff(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  { 
+    return coeff(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  { 
+    return m_argImpl.coeffRef(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  { 
+    return coeffRef(RowsAtCompileTime == 1 ? 0 : index, RowsAtCompileTime == 1 ? index : 0);
+  }
+ 
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const 
+  { 
+    return m_argImpl.template packet<LoadMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col); 
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const 
+  { 
+    return packet<LoadMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                       RowsAtCompileTime == 1 ? index : 0);
+  }
+  
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x) 
+  {
+    return m_argImpl.template writePacket<StoreMode,PacketType>(m_startRow.value() + row, m_startCol.value() + col, x); 
+  }
+  
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x) 
+  {
+    return writePacket<StoreMode,PacketType>(RowsAtCompileTime == 1 ? 0 : index,
+                                             RowsAtCompileTime == 1 ? index : 0,
+                                             x);
+  }
+ 
+protected:
+  evaluator<ArgType> m_argImpl;
+  const variable_if_dynamic<Index, (ArgType::RowsAtCompileTime == 1 && BlockRows==1) ? 0 : Dynamic> m_startRow;
+  const variable_if_dynamic<Index, (ArgType::ColsAtCompileTime == 1 && BlockCols==1) ? 0 : Dynamic> m_startCol;
+};
+
+// TODO: This evaluator does not actually use the child evaluator; 
+// all action is via the data() as returned by the Block expression.
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel> 
+struct block_evaluator<ArgType, BlockRows, BlockCols, InnerPanel, /* HasDirectAccess */ true>
+  : mapbase_evaluator<Block<ArgType, BlockRows, BlockCols, InnerPanel>,
+                      typename Block<ArgType, BlockRows, BlockCols, InnerPanel>::PlainObject>
+{
+  typedef Block<ArgType, BlockRows, BlockCols, InnerPanel> XprType;
+  typedef typename XprType::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC explicit block_evaluator(const XprType& block)
+    : mapbase_evaluator<XprType, typename XprType::PlainObject>(block) 
+  {
+    // TODO: for the 3.3 release, this should be turned to an internal assertion, but let's keep it as is for the beta lifetime
+    eigen_assert(((internal::UIntPtr(block.data()) % EIGEN_PLAIN_ENUM_MAX(1,evaluator<XprType>::Alignment)) == 0) && "data is not aligned");
+  }
+};
+
+
+// -------------------- Select --------------------
+// NOTE shall we introduce a ternary_evaluator?
+
+// TODO enable vectorization for Select
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+struct evaluator<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+  : evaluator_base<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+{
+  typedef Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> XprType;
+  enum {
+    CoeffReadCost = evaluator<ConditionMatrixType>::CoeffReadCost
+                  + EIGEN_PLAIN_ENUM_MAX(evaluator<ThenMatrixType>::CoeffReadCost,
+                                         evaluator<ElseMatrixType>::CoeffReadCost),
+
+    Flags = (unsigned int)evaluator<ThenMatrixType>::Flags & evaluator<ElseMatrixType>::Flags & HereditaryBits,
+    
+    Alignment = EIGEN_PLAIN_ENUM_MIN(evaluator<ThenMatrixType>::Alignment, evaluator<ElseMatrixType>::Alignment)
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& select)
+    : m_conditionImpl(select.conditionMatrix()),
+      m_thenImpl(select.thenMatrix()),
+      m_elseImpl(select.elseMatrix())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+ 
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    if (m_conditionImpl.coeff(row, col))
+      return m_thenImpl.coeff(row, col);
+    else
+      return m_elseImpl.coeff(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    if (m_conditionImpl.coeff(index))
+      return m_thenImpl.coeff(index);
+    else
+      return m_elseImpl.coeff(index);
+  }
+ 
+protected:
+  evaluator<ConditionMatrixType> m_conditionImpl;
+  evaluator<ThenMatrixType> m_thenImpl;
+  evaluator<ElseMatrixType> m_elseImpl;
+};
+
+
+// -------------------- Replicate --------------------
+
+template<typename ArgType, int RowFactor, int ColFactor> 
+struct unary_evaluator<Replicate<ArgType, RowFactor, ColFactor> >
+  : evaluator_base<Replicate<ArgType, RowFactor, ColFactor> >
+{
+  typedef Replicate<ArgType, RowFactor, ColFactor> XprType;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  enum {
+    Factor = (RowFactor==Dynamic || ColFactor==Dynamic) ? Dynamic : RowFactor*ColFactor
+  };
+  typedef typename internal::nested_eval<ArgType,Factor>::type ArgTypeNested;
+  typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgTypeNestedCleaned>::CoeffReadCost,
+    LinearAccessMask = XprType::IsVectorAtCompileTime ? LinearAccessBit : 0,
+    Flags = (evaluator<ArgTypeNestedCleaned>::Flags & (HereditaryBits|LinearAccessMask) & ~RowMajorBit) | (traits<XprType>::Flags & RowMajorBit),
+    
+    Alignment = evaluator<ArgTypeNestedCleaned>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& replicate)
+    : m_arg(replicate.nestedExpression()),
+      m_argImpl(m_arg),
+      m_rows(replicate.nestedExpression().rows()),
+      m_cols(replicate.nestedExpression().cols())
+  {}
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    // try to avoid using modulo; this is a pure optimization strategy
+    const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0
+                           : RowFactor==1 ? row
+                           : row % m_rows.value();
+    const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0
+                           : ColFactor==1 ? col
+                           : col % m_cols.value();
+    
+    return m_argImpl.coeff(actual_row, actual_col);
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    // try to avoid using modulo; this is a pure optimization strategy
+    const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1
+                                  ? (ColFactor==1 ?  index : index%m_cols.value())
+                                  : (RowFactor==1 ?  index : index%m_rows.value());
+    
+    return m_argImpl.coeff(actual_index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    const Index actual_row = internal::traits<XprType>::RowsAtCompileTime==1 ? 0
+                           : RowFactor==1 ? row
+                           : row % m_rows.value();
+    const Index actual_col = internal::traits<XprType>::ColsAtCompileTime==1 ? 0
+                           : ColFactor==1 ? col
+                           : col % m_cols.value();
+
+    return m_argImpl.template packet<LoadMode,PacketType>(actual_row, actual_col);
+  }
+  
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    const Index actual_index = internal::traits<XprType>::RowsAtCompileTime==1
+                                  ? (ColFactor==1 ?  index : index%m_cols.value())
+                                  : (RowFactor==1 ?  index : index%m_rows.value());
+
+    return m_argImpl.template packet<LoadMode,PacketType>(actual_index);
+  }
+ 
+protected:
+  const ArgTypeNested m_arg;
+  evaluator<ArgTypeNestedCleaned> m_argImpl;
+  const variable_if_dynamic<Index, ArgType::RowsAtCompileTime> m_rows;
+  const variable_if_dynamic<Index, ArgType::ColsAtCompileTime> m_cols;
+};
+
+
+// -------------------- PartialReduxExpr --------------------
+
+template< typename ArgType, typename MemberOp, int Direction>
+struct evaluator<PartialReduxExpr<ArgType, MemberOp, Direction> >
+  : evaluator_base<PartialReduxExpr<ArgType, MemberOp, Direction> >
+{
+  typedef PartialReduxExpr<ArgType, MemberOp, Direction> XprType;
+  typedef typename internal::nested_eval<ArgType,1>::type ArgTypeNested;
+  typedef typename internal::remove_all<ArgTypeNested>::type ArgTypeNestedCleaned;
+  typedef typename ArgType::Scalar InputScalar;
+  typedef typename XprType::Scalar Scalar;
+  enum {
+    TraversalSize = Direction==int(Vertical) ? int(ArgType::RowsAtCompileTime) :  int(ArgType::ColsAtCompileTime)
+  };
+  typedef typename MemberOp::template Cost<InputScalar,int(TraversalSize)> CostOpType;
+  enum {
+    CoeffReadCost = TraversalSize==Dynamic ? HugeCost
+                  : TraversalSize * evaluator<ArgType>::CoeffReadCost + int(CostOpType::value),
+    
+    Flags = (traits<XprType>::Flags&RowMajorBit) | (evaluator<ArgType>::Flags&(HereditaryBits&(~RowMajorBit))) | LinearAccessBit,
+    
+    Alignment = 0 // FIXME this will need to be improved once PartialReduxExpr is vectorized
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType xpr)
+    : m_arg(xpr.nestedExpression()), m_functor(xpr.functor())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(TraversalSize==Dynamic ? HugeCost : int(CostOpType::value));
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar coeff(Index i, Index j) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(j));
+    else
+      return m_functor(m_arg.row(i));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar coeff(Index index) const
+  {
+    if (Direction==Vertical)
+      return m_functor(m_arg.col(index));
+    else
+      return m_functor(m_arg.row(index));
+  }
+
+protected:
+  typename internal::add_const_on_value_type<ArgTypeNested>::type m_arg;
+  const MemberOp m_functor;
+};
+
+
+// -------------------- MatrixWrapper and ArrayWrapper --------------------
+//
+// evaluator_wrapper_base<T> is a common base class for the
+// MatrixWrapper and ArrayWrapper evaluators.
+
+template<typename XprType>
+struct evaluator_wrapper_base
+  : evaluator_base<XprType>
+{
+  typedef typename remove_all<typename XprType::NestedExpressionType>::type ArgType;
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    Flags = evaluator<ArgType>::Flags,
+    Alignment = evaluator<ArgType>::Alignment
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator_wrapper_base(const ArgType& arg) : m_argImpl(arg) {}
+
+  typedef typename ArgType::Scalar Scalar;
+  typedef typename ArgType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(row, col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(row, col);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    return m_argImpl.template packet<LoadMode,PacketType>(index);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode>(row, col, x);
+  }
+
+  template<int StoreMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    m_argImpl.template writePacket<StoreMode>(index, x);
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+};
+
+template<typename TArgType>
+struct unary_evaluator<MatrixWrapper<TArgType> >
+  : evaluator_wrapper_base<MatrixWrapper<TArgType> >
+{
+  typedef MatrixWrapper<TArgType> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper)
+    : evaluator_wrapper_base<MatrixWrapper<TArgType> >(wrapper.nestedExpression())
+  { }
+};
+
+template<typename TArgType>
+struct unary_evaluator<ArrayWrapper<TArgType> >
+  : evaluator_wrapper_base<ArrayWrapper<TArgType> >
+{
+  typedef ArrayWrapper<TArgType> XprType;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& wrapper)
+    : evaluator_wrapper_base<ArrayWrapper<TArgType> >(wrapper.nestedExpression())
+  { }
+};
+
+
+// -------------------- Reverse --------------------
+
+// defined in Reverse.h:
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond;
+
+template<typename ArgType, int Direction>
+struct unary_evaluator<Reverse<ArgType, Direction> >
+  : evaluator_base<Reverse<ArgType, Direction> >
+{
+  typedef Reverse<ArgType, Direction> XprType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  enum {
+    IsRowMajor = XprType::IsRowMajor,
+    IsColMajor = !IsRowMajor,
+    ReverseRow = (Direction == Vertical)   || (Direction == BothDirections),
+    ReverseCol = (Direction == Horizontal) || (Direction == BothDirections),
+    ReversePacket = (Direction == BothDirections)
+                    || ((Direction == Vertical)   && IsColMajor)
+                    || ((Direction == Horizontal) && IsRowMajor),
+                    
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    // let's enable LinearAccess only with vectorization because of the product overhead
+    // FIXME enable DirectAccess with negative strides?
+    Flags0 = evaluator<ArgType>::Flags,
+    LinearAccess = ( (Direction==BothDirections) && (int(Flags0)&PacketAccessBit) )
+                  || ((ReverseRow && XprType::ColsAtCompileTime==1) || (ReverseCol && XprType::RowsAtCompileTime==1))
+                 ? LinearAccessBit : 0,
+
+    Flags = int(Flags0) & (HereditaryBits | PacketAccessBit | LinearAccess),
+    
+    Alignment = 0 // FIXME in some rare cases, Alignment could be preserved, like a Vector4f.
+  };
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& reverse)
+    : m_argImpl(reverse.nestedExpression()),
+      m_rows(ReverseRow ? reverse.nestedExpression().rows() : 1),
+      m_cols(ReverseCol ? reverse.nestedExpression().cols() : 1)
+  { }
+ 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index col) const
+  {
+    return m_argImpl.coeff(ReverseRow ? m_rows.value() - row - 1 : row,
+                           ReverseCol ? m_cols.value() - col - 1 : col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(m_rows.value() * m_cols.value() - index - 1);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index col)
+  {
+    return m_argImpl.coeffRef(ReverseRow ? m_rows.value() - row - 1 : row,
+                              ReverseCol ? m_cols.value() - col - 1 : col);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(m_rows.value() * m_cols.value() - index - 1);
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index row, Index col) const
+  {
+    enum {
+      PacketSize = unpacket_traits<PacketType>::size,
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1
+    };
+    typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet;
+    return reverse_packet::run(m_argImpl.template packet<LoadMode,PacketType>(
+                                  ReverseRow ? m_rows.value() - row - OffsetRow : row,
+                                  ReverseCol ? m_cols.value() - col - OffsetCol : col));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  PacketType packet(Index index) const
+  {
+    enum { PacketSize = unpacket_traits<PacketType>::size };
+    return preverse(m_argImpl.template packet<LoadMode,PacketType>(m_rows.value() * m_cols.value() - index - PacketSize));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index row, Index col, const PacketType& x)
+  {
+    // FIXME we could factorize some code with packet(i,j)
+    enum {
+      PacketSize = unpacket_traits<PacketType>::size,
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1
+    };
+    typedef internal::reverse_packet_cond<PacketType,ReversePacket> reverse_packet;
+    m_argImpl.template writePacket<LoadMode>(
+                                  ReverseRow ? m_rows.value() - row - OffsetRow : row,
+                                  ReverseCol ? m_cols.value() - col - OffsetCol : col,
+                                  reverse_packet::run(x));
+  }
+
+  template<int LoadMode, typename PacketType>
+  EIGEN_STRONG_INLINE
+  void writePacket(Index index, const PacketType& x)
+  {
+    enum { PacketSize = unpacket_traits<PacketType>::size };
+    m_argImpl.template writePacket<LoadMode>
+      (m_rows.value() * m_cols.value() - index - PacketSize, preverse(x));
+  }
+ 
+protected:
+  evaluator<ArgType> m_argImpl;
+
+  // If we do not reverse rows, then we do not need to know the number of rows; same for columns
+  // Nonetheless, in this case it is important to set to 1 such that the coeff(index) method works fine for vectors.
+  const variable_if_dynamic<Index, ReverseRow ? ArgType::RowsAtCompileTime : 1> m_rows;
+  const variable_if_dynamic<Index, ReverseCol ? ArgType::ColsAtCompileTime : 1> m_cols;
+};
+
+
+// -------------------- Diagonal --------------------
+
+template<typename ArgType, int DiagIndex>
+struct evaluator<Diagonal<ArgType, DiagIndex> >
+  : evaluator_base<Diagonal<ArgType, DiagIndex> >
+{
+  typedef Diagonal<ArgType, DiagIndex> XprType;
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    
+    Flags = (unsigned int)(evaluator<ArgType>::Flags & (HereditaryBits | DirectAccessBit) & ~RowMajorBit) | LinearAccessBit,
+    
+    Alignment = 0
+  };
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& diagonal)
+    : m_argImpl(diagonal.nestedExpression()),
+      m_index(diagonal.index())
+  { }
+ 
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index row, Index) const
+  {
+    return m_argImpl.coeff(row + rowOffset(), row + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  CoeffReturnType coeff(Index index) const
+  {
+    return m_argImpl.coeff(index + rowOffset(), index + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index row, Index)
+  {
+    return m_argImpl.coeffRef(row + rowOffset(), row + colOffset());
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Scalar& coeffRef(Index index)
+  {
+    return m_argImpl.coeffRef(index + rowOffset(), index + colOffset());
+  }
+
+protected:
+  evaluator<ArgType> m_argImpl;
+  const internal::variable_if_dynamicindex<Index, XprType::DiagIndex> m_index;
+
+private:
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value() > 0 ? 0 : -m_index.value(); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value() > 0 ? m_index.value() : 0; }
+};
+
+
+//----------------------------------------------------------------------
+// deprecated code
+//----------------------------------------------------------------------
+
+// -------------------- EvalToTemp --------------------
+
+// expression class for evaluating nested expression to a temporary
+
+template<typename ArgType> class EvalToTemp;
+
+template<typename ArgType>
+struct traits<EvalToTemp<ArgType> >
+  : public traits<ArgType>
+{ };
+
+template<typename ArgType>
+class EvalToTemp
+  : public dense_xpr_base<EvalToTemp<ArgType> >::type
+{
+ public:
+ 
+  typedef typename dense_xpr_base<EvalToTemp>::type Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(EvalToTemp)
+ 
+  explicit EvalToTemp(const ArgType& arg)
+    : m_arg(arg)
+  { }
+ 
+  const ArgType& arg() const
+  {
+    return m_arg;
+  }
+
+  Index rows() const 
+  {
+    return m_arg.rows();
+  }
+
+  Index cols() const 
+  {
+    return m_arg.cols();
+  }
+
+ private:
+  const ArgType& m_arg;
+};
+ 
+template<typename ArgType>
+struct evaluator<EvalToTemp<ArgType> >
+  : public evaluator<typename ArgType::PlainObject>
+{
+  typedef EvalToTemp<ArgType>                   XprType;
+  typedef typename ArgType::PlainObject         PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.arg())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+  }
+
+  // This constructor is used when nesting an EvalTo evaluator in another evaluator
+  EIGEN_DEVICE_FUNC evaluator(const ArgType& arg)
+    : m_result(arg)
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COREEVALUATORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h
new file mode 100644
index 0000000..4eb42b9
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CoreIterators.h
@@ -0,0 +1,127 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COREITERATORS_H
+#define EIGEN_COREITERATORS_H
+
+namespace Eigen { 
+
+/* This file contains the respective InnerIterator definition of the expressions defined in Eigen/Core
+ */
+
+namespace internal {
+
+template<typename XprType, typename EvaluatorKind>
+class inner_iterator_selector;
+
+}
+
+/** \class InnerIterator
+  * \brief An InnerIterator allows to loop over the element of any matrix expression.
+  * 
+  * \warning To be used with care because an evaluator is constructed every time an InnerIterator iterator is constructed.
+  * 
+  * TODO: add a usage example
+  */
+template<typename XprType>
+class InnerIterator
+{
+protected:
+  typedef internal::inner_iterator_selector<XprType, typename internal::evaluator_traits<XprType>::Kind> IteratorType;
+  typedef internal::evaluator<XprType> EvaluatorType;
+  typedef typename internal::traits<XprType>::Scalar Scalar;
+public:
+  /** Construct an iterator over the \a outerId -th row or column of \a xpr */
+  InnerIterator(const XprType &xpr, const Index &outerId)
+    : m_eval(xpr), m_iter(m_eval, outerId, xpr.innerSize())
+  {}
+  
+  /// \returns the value of the current coefficient.
+  EIGEN_STRONG_INLINE Scalar value() const          { return m_iter.value(); }
+  /** Increment the iterator \c *this to the next non-zero coefficient.
+    * Explicit zeros are not skipped over. To skip explicit zeros, see class SparseView
+    */
+  EIGEN_STRONG_INLINE InnerIterator& operator++()   { m_iter.operator++(); return *this; }
+  /// \returns the column or row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index index() const           { return m_iter.index(); }
+  /// \returns the row index of the current coefficient.
+  EIGEN_STRONG_INLINE Index row() const             { return m_iter.row(); }
+  /// \returns the column index of the current coefficient.
+  EIGEN_STRONG_INLINE Index col() const             { return m_iter.col(); }
+  /// \returns \c true if the iterator \c *this still references a valid coefficient.
+  EIGEN_STRONG_INLINE operator bool() const         { return m_iter; }
+  
+protected:
+  EvaluatorType m_eval;
+  IteratorType m_iter;
+private:
+  // If you get here, then you're not using the right InnerIterator type, e.g.:
+  //   SparseMatrix<double,RowMajor> A;
+  //   SparseMatrix<double>::InnerIterator it(A,0);
+  template<typename T> InnerIterator(const EigenBase<T>&,Index outer);
+};
+
+namespace internal {
+
+// Generic inner iterator implementation for dense objects
+template<typename XprType>
+class inner_iterator_selector<XprType, IndexBased>
+{
+protected:
+  typedef evaluator<XprType> EvaluatorType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit };
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &innerSize)
+    : m_eval(eval), m_inner(0), m_outer(outerId), m_end(innerSize)
+  {}
+
+  EIGEN_STRONG_INLINE Scalar value() const
+  {
+    return (IsRowMajor) ? m_eval.coeff(m_outer, m_inner)
+                        : m_eval.coeff(m_inner, m_outer);
+  }
+
+  EIGEN_STRONG_INLINE inner_iterator_selector& operator++() { m_inner++; return *this; }
+
+  EIGEN_STRONG_INLINE Index index() const { return m_inner; }
+  inline Index row() const { return IsRowMajor ? m_outer : index(); }
+  inline Index col() const { return IsRowMajor ? index() : m_outer; }
+
+  EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
+
+protected:
+  const EvaluatorType& m_eval;
+  Index m_inner;
+  const Index m_outer;
+  const Index m_end;
+};
+
+// For iterator-based evaluator, inner-iterator is already implemented as
+// evaluator<>::InnerIterator
+template<typename XprType>
+class inner_iterator_selector<XprType, IteratorBased>
+ : public evaluator<XprType>::InnerIterator
+{
+protected:
+  typedef typename evaluator<XprType>::InnerIterator Base;
+  typedef evaluator<XprType> EvaluatorType;
+  
+public:
+  EIGEN_STRONG_INLINE inner_iterator_selector(const EvaluatorType &eval, const Index &outerId, const Index &/*innerSize*/)
+    : Base(eval, outerId)
+  {}  
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COREITERATORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h
new file mode 100644
index 0000000..a36765e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseBinaryOp.h
@@ -0,0 +1,184 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_BINARY_OP_H
+#define EIGEN_CWISE_BINARY_OP_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  // we must not inherit from traits<Lhs> since it has
+  // the potential to cause problems with MSVC
+  typedef typename remove_all<Lhs>::type Ancestor;
+  typedef typename traits<Ancestor>::XprKind XprKind;
+  enum {
+    RowsAtCompileTime = traits<Ancestor>::RowsAtCompileTime,
+    ColsAtCompileTime = traits<Ancestor>::ColsAtCompileTime,
+    MaxRowsAtCompileTime = traits<Ancestor>::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = traits<Ancestor>::MaxColsAtCompileTime
+  };
+
+  // even though we require Lhs and Rhs to have the same scalar type (see CwiseBinaryOp constructor),
+  // we still want to handle the case when the result type is different.
+  typedef typename result_of<
+                     BinaryOp(
+                       const typename Lhs::Scalar&,
+                       const typename Rhs::Scalar&
+                     )
+                   >::type Scalar;
+  typedef typename cwise_promote_storage_type<typename traits<Lhs>::StorageKind,
+                                              typename traits<Rhs>::StorageKind,
+                                              BinaryOp>::ret StorageKind;
+  typedef typename promote_index_type<typename traits<Lhs>::StorageIndex,
+                                      typename traits<Rhs>::StorageIndex>::type StorageIndex;
+  typedef typename Lhs::Nested LhsNested;
+  typedef typename Rhs::Nested RhsNested;
+  typedef typename remove_reference<LhsNested>::type _LhsNested;
+  typedef typename remove_reference<RhsNested>::type _RhsNested;
+  enum {
+    Flags = cwise_promote_storage_order<typename traits<Lhs>::StorageKind,typename traits<Rhs>::StorageKind,_LhsNested::Flags & RowMajorBit,_RhsNested::Flags & RowMajorBit>::value
+  };
+};
+} // end namespace internal
+
+template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
+class CwiseBinaryOpImpl;
+
+/** \class CwiseBinaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise binary operator is applied to two expressions
+  *
+  * \tparam BinaryOp template functor implementing the operator
+  * \tparam LhsType the type of the left-hand side
+  * \tparam RhsType the type of the right-hand side
+  *
+  * This class represents an expression  where a coefficient-wise binary operator is applied to two expressions.
+  * It is the return type of binary operators, by which we mean only those binary operators where
+  * both the left-hand side and the right-hand side are Eigen expressions.
+  * For example, the return type of matrix1+matrix2 is a CwiseBinaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically don't have to name
+  * CwiseBinaryOp types explicitly.
+  *
+  * \sa MatrixBase::binaryExpr(const MatrixBase<OtherDerived> &,const CustomBinaryOp &) const, class CwiseUnaryOp, class CwiseNullaryOp
+  */
+template<typename BinaryOp, typename LhsType, typename RhsType>
+class CwiseBinaryOp : 
+  public CwiseBinaryOpImpl<
+          BinaryOp, LhsType, RhsType,
+          typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                        typename internal::traits<RhsType>::StorageKind,
+                                                        BinaryOp>::ret>,
+  internal::no_assignment_operator
+{
+  public:
+    
+    typedef typename internal::remove_all<BinaryOp>::type Functor;
+    typedef typename internal::remove_all<LhsType>::type Lhs;
+    typedef typename internal::remove_all<RhsType>::type Rhs;
+
+    typedef typename CwiseBinaryOpImpl<
+        BinaryOp, LhsType, RhsType,
+        typename internal::cwise_promote_storage_type<typename internal::traits<LhsType>::StorageKind,
+                                                      typename internal::traits<Rhs>::StorageKind,
+                                                      BinaryOp>::ret>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseBinaryOp)
+
+    typedef typename internal::ref_selector<LhsType>::type LhsNested;
+    typedef typename internal::ref_selector<RhsType>::type RhsNested;
+    typedef typename internal::remove_reference<LhsNested>::type _LhsNested;
+    typedef typename internal::remove_reference<RhsNested>::type _RhsNested;
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CwiseBinaryOp(const Lhs& aLhs, const Rhs& aRhs, const BinaryOp& func = BinaryOp())
+      : m_lhs(aLhs), m_rhs(aRhs), m_functor(func)
+    {
+      EIGEN_CHECK_BINARY_COMPATIBILIY(BinaryOp,typename Lhs::Scalar,typename Rhs::Scalar);
+      // require the sizes to match
+      EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Lhs, Rhs)
+      eigen_assert(aLhs.rows() == aRhs.rows() && aLhs.cols() == aRhs.cols());
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rows() const {
+      // return the fixed size type if available to enable compile time optimizations
+      if (internal::traits<typename internal::remove_all<LhsNested>::type>::RowsAtCompileTime==Dynamic)
+        return m_rhs.rows();
+      else
+        return m_lhs.rows();
+    }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index cols() const {
+      // return the fixed size type if available to enable compile time optimizations
+      if (internal::traits<typename internal::remove_all<LhsNested>::type>::ColsAtCompileTime==Dynamic)
+        return m_rhs.cols();
+      else
+        return m_lhs.cols();
+    }
+
+    /** \returns the left hand side nested expression */
+    EIGEN_DEVICE_FUNC
+    const _LhsNested& lhs() const { return m_lhs; }
+    /** \returns the right hand side nested expression */
+    EIGEN_DEVICE_FUNC
+    const _RhsNested& rhs() const { return m_rhs; }
+    /** \returns the functor representing the binary operation */
+    EIGEN_DEVICE_FUNC
+    const BinaryOp& functor() const { return m_functor; }
+
+  protected:
+    LhsNested m_lhs;
+    RhsNested m_rhs;
+    const BinaryOp m_functor;
+};
+
+// Generic API dispatcher
+template<typename BinaryOp, typename Lhs, typename Rhs, typename StorageKind>
+class CwiseBinaryOpImpl
+  : public internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >::type Base;
+};
+
+/** replaces \c *this by \c *this - \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE Derived &
+MatrixBase<Derived>::operator-=(const MatrixBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+/** replaces \c *this by \c *this + \a other.
+  *
+  * \returns a reference to \c *this
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE Derived &
+MatrixBase<Derived>::operator+=(const MatrixBase<OtherDerived>& other)
+{
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_CWISE_BINARY_OP_H
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h
new file mode 100644
index 0000000..ddd607e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseNullaryOp.h
@@ -0,0 +1,866 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_NULLARY_OP_H
+#define EIGEN_CWISE_NULLARY_OP_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename NullaryOp, typename PlainObjectType>
+struct traits<CwiseNullaryOp<NullaryOp, PlainObjectType> > : traits<PlainObjectType>
+{
+  enum {
+    Flags = traits<PlainObjectType>::Flags & RowMajorBit
+  };
+};
+
+} // namespace internal
+
+/** \class CwiseNullaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression of a matrix where all coefficients are defined by a functor
+  *
+  * \tparam NullaryOp template functor implementing the operator
+  * \tparam PlainObjectType the underlying plain matrix/array type
+  *
+  * This class represents an expression of a generic nullary operator.
+  * It is the return type of the Ones(), Zero(), Constant(), Identity() and Random() methods,
+  * and most of the time this is the only way it is used.
+  *
+  * However, if you want to write a function returning such an expression, you
+  * will need to use this class.
+  *
+  * The functor NullaryOp must expose one of the following method:
+    <table class="manual">
+    <tr            ><td>\c operator()() </td><td>if the procedural generation does not depend on the coefficient entries (e.g., random numbers)</td></tr>
+    <tr class="alt"><td>\c operator()(Index i)</td><td>if the procedural generation makes sense for vectors only and that it depends on the coefficient index \c i (e.g., linspace) </td></tr>
+    <tr            ><td>\c operator()(Index i,Index j)</td><td>if the procedural generation depends on the matrix coordinates \c i, \c j (e.g., to generate a checkerboard with 0 and 1)</td></tr>
+    </table>
+  * It is also possible to expose the last two operators if the generation makes sense for matrices but can be optimized for vectors.
+  *
+  * See DenseBase::NullaryExpr(Index,const CustomNullaryOp&) for an example binding
+  * C++11 random number generators.
+  *
+  * A nullary expression can also be used to implement custom sophisticated matrix manipulations
+  * that cannot be covered by the existing set of natively supported matrix manipulations.
+  * See this \ref TopicCustomizing_NullaryExpr "page" for some examples and additional explanations
+  * on the behavior of CwiseNullaryOp.
+  *
+  * \sa class CwiseUnaryOp, class CwiseBinaryOp, DenseBase::NullaryExpr
+  */
+template<typename NullaryOp, typename PlainObjectType>
+class CwiseNullaryOp : public internal::dense_xpr_base< CwiseNullaryOp<NullaryOp, PlainObjectType> >::type, internal::no_assignment_operator
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<CwiseNullaryOp>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(CwiseNullaryOp)
+
+    EIGEN_DEVICE_FUNC
+    CwiseNullaryOp(Index rows, Index cols, const NullaryOp& func = NullaryOp())
+      : m_rows(rows), m_cols(cols), m_functor(func)
+    {
+      eigen_assert(rows >= 0
+            && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+            &&  cols >= 0
+            && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols));
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rows() const { return m_rows.value(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index cols() const { return m_cols.value(); }
+
+    /** \returns the functor representing the nullary operation */
+    EIGEN_DEVICE_FUNC
+    const NullaryOp& functor() const { return m_functor; }
+
+  protected:
+    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows;
+    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols;
+    const NullaryOp m_functor;
+};
+
+
+/** \returns an expression of a matrix defined by a custom functor \a func
+  *
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this MatrixBase type.
+  *
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
+  * instead.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+template<typename CustomNullaryOp>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+DenseBase<Derived>::NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func)
+{
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(rows, cols, func);
+}
+
+/** \returns an expression of a matrix defined by a custom functor \a func
+  *
+  * The parameter \a size is the size of the returned vector.
+  * Must be compatible with this MatrixBase type.
+  *
+  * \only_for_vectors
+  *
+  * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
+  * it is redundant to pass \a size as argument, so Zero() should be used
+  * instead.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * Here is an example with C++11 random generators: \include random_cpp11.cpp
+  * Output: \verbinclude random_cpp11.out
+  * 
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+template<typename CustomNullaryOp>
+EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+DenseBase<Derived>::NullaryExpr(Index size, const CustomNullaryOp& func)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  if(RowsAtCompileTime == 1) return CwiseNullaryOp<CustomNullaryOp, PlainObject>(1, size, func);
+  else return CwiseNullaryOp<CustomNullaryOp, PlainObject>(size, 1, func);
+}
+
+/** \returns an expression of a matrix defined by a custom functor \a func
+  *
+  * This variant is only for fixed-size DenseBase types. For dynamic-size types, you
+  * need to use the variants taking size arguments.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+template<typename CustomNullaryOp>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseNullaryOp<CustomNullaryOp, typename DenseBase<Derived>::PlainObject>
+DenseBase<Derived>::NullaryExpr(const CustomNullaryOp& func)
+{
+  return CwiseNullaryOp<CustomNullaryOp, PlainObject>(RowsAtCompileTime, ColsAtCompileTime, func);
+}
+
+/** \returns an expression of a constant matrix of value \a value
+  *
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this DenseBase type.
+  *
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
+  * instead.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Constant(Index rows, Index cols, const Scalar& value)
+{
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_constant_op<Scalar>(value));
+}
+
+/** \returns an expression of a constant matrix of value \a value
+  *
+  * The parameter \a size is the size of the returned vector.
+  * Must be compatible with this DenseBase type.
+  *
+  * \only_for_vectors
+  *
+  * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
+  * it is redundant to pass \a size as argument, so Zero() should be used
+  * instead.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Constant(Index size, const Scalar& value)
+{
+  return DenseBase<Derived>::NullaryExpr(size, internal::scalar_constant_op<Scalar>(value));
+}
+
+/** \returns an expression of a constant matrix of value \a value
+  *
+  * This variant is only for fixed-size DenseBase types. For dynamic-size types, you
+  * need to use the variants taking size arguments.
+  *
+  * The template parameter \a CustomNullaryOp is the type of the functor.
+  *
+  * \sa class CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Constant(const Scalar& value)
+{
+  EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
+  return DenseBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_constant_op<Scalar>(value));
+}
+
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(Index,const Scalar&,const Scalar&)
+  *
+  * \sa LinSpaced(Index,Scalar,Scalar), setLinSpaced(Index,const Scalar&,const Scalar&)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+DenseBase<Derived>::LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
+}
+
+/** \deprecated because of accuracy loss. In Eigen 3.3, it is an alias for LinSpaced(const Scalar&,const Scalar&)
+  *
+  * \sa LinSpaced(Scalar,Scalar)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+DenseBase<Derived>::LinSpaced(Sequential_t, const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
+}
+
+/**
+  * \brief Sets a linearly spaced vector.
+  *
+  * The function generates 'size' equally spaced values in the closed interval [low,high].
+  * When size is set to 1, a vector of length 1 containing 'high' is returned.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include DenseBase_LinSpaced.cpp
+  * Output: \verbinclude DenseBase_LinSpaced.out
+  *
+  * For integer scalar types, an even spacing is possible if and only if the length of the range,
+  * i.e., \c high-low is a scalar multiple of \c size-1, or if \c size is a scalar multiple of the
+  * number of values \c high-low+1 (meaning each value can be repeated the same number of time).
+  * If one of these two considions is not satisfied, then \c high is lowered to the largest value
+  * satisfying one of this constraint.
+  * Here are some examples:
+  *
+  * Example: \include DenseBase_LinSpacedInt.cpp
+  * Output: \verbinclude DenseBase_LinSpacedInt.out
+  *
+  * \sa setLinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+DenseBase<Derived>::LinSpaced(Index size, const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return DenseBase<Derived>::NullaryExpr(size, internal::linspaced_op<Scalar,PacketScalar>(low,high,size));
+}
+
+/**
+  * \copydoc DenseBase::LinSpaced(Index, const Scalar&, const Scalar&)
+  * Special version for fixed size types which does not require the size parameter.
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::RandomAccessLinSpacedReturnType
+DenseBase<Derived>::LinSpaced(const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
+  return DenseBase<Derived>::NullaryExpr(Derived::SizeAtCompileTime, internal::linspaced_op<Scalar,PacketScalar>(low,high,Derived::SizeAtCompileTime));
+}
+
+/** \returns true if all coefficients in this matrix are approximately equal to \a val, to within precision \a prec */
+template<typename Derived>
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isApproxToConstant
+(const Scalar& val, const RealScalar& prec) const
+{
+  typename internal::nested_eval<Derived,1>::type self(derived());
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = 0; i < rows(); ++i)
+      if(!internal::isApprox(self.coeff(i, j), val, prec))
+        return false;
+  return true;
+}
+
+/** This is just an alias for isApproxToConstant().
+  *
+  * \returns true if all coefficients in this matrix are approximately equal to \a value, to within precision \a prec */
+template<typename Derived>
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isConstant
+(const Scalar& val, const RealScalar& prec) const
+{
+  return isApproxToConstant(val, prec);
+}
+
+/** Alias for setConstant(): sets all coefficients in this expression to \a val.
+  *
+  * \sa setConstant(), Constant(), class CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void DenseBase<Derived>::fill(const Scalar& val)
+{
+  setConstant(val);
+}
+
+/** Sets all coefficients in this expression to value \a val.
+  *
+  * \sa fill(), setConstant(Index,const Scalar&), setConstant(Index,Index,const Scalar&), setZero(), setOnes(), Constant(), class CwiseNullaryOp, setZero(), setOnes()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setConstant(const Scalar& val)
+{
+  return derived() = Constant(rows(), cols(), val);
+}
+
+/** Resizes to the given \a size, and sets all coefficients in this expression to the given value \a val.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include Matrix_setConstant_int.cpp
+  * Output: \verbinclude Matrix_setConstant_int.out
+  *
+  * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setConstant(Index size, const Scalar& val)
+{
+  resize(size);
+  return setConstant(val);
+}
+
+/** Resizes to the given size, and sets all coefficients in this expression to the given value \a val.
+  *
+  * \param rows the new number of rows
+  * \param cols the new number of columns
+  * \param val the value to which all coefficients are set
+  *
+  * Example: \include Matrix_setConstant_int_int.cpp
+  * Output: \verbinclude Matrix_setConstant_int_int.out
+  *
+  * \sa MatrixBase::setConstant(const Scalar&), setConstant(Index,const Scalar&), class CwiseNullaryOp, MatrixBase::Constant(const Scalar&)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setConstant(Index rows, Index cols, const Scalar& val)
+{
+  resize(rows, cols);
+  return setConstant(val);
+}
+
+/**
+  * \brief Sets a linearly spaced vector.
+  *
+  * The function generates 'size' equally spaced values in the closed interval [low,high].
+  * When size is set to 1, a vector of length 1 containing 'high' is returned.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include DenseBase_setLinSpaced.cpp
+  * Output: \verbinclude DenseBase_setLinSpaced.out
+  *
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(Index newSize, const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return derived() = Derived::NullaryExpr(newSize, internal::linspaced_op<Scalar,PacketScalar>(low,high,newSize));
+}
+
+/**
+  * \brief Sets a linearly spaced vector.
+  *
+  * The function fills \c *this with equally spaced values in the closed interval [low,high].
+  * When size is set to 1, a vector of length 1 containing 'high' is returned.
+  *
+  * \only_for_vectors
+  *
+  * For integer scalar types, do not miss the explanations on the definition
+  * of \link LinSpaced(Index,const Scalar&,const Scalar&) even spacing \endlink.
+  *
+  * \sa LinSpaced(Index,const Scalar&,const Scalar&), setLinSpaced(Index, const Scalar&, const Scalar&), CwiseNullaryOp
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setLinSpaced(const Scalar& low, const Scalar& high)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return setLinSpaced(size(), low, high);
+}
+
+// zero:
+
+/** \returns an expression of a zero matrix.
+  *
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this MatrixBase type.
+  *
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Zero() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_zero_int_int.cpp
+  * Output: \verbinclude MatrixBase_zero_int_int.out
+  *
+  * \sa Zero(), Zero(Index)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Zero(Index rows, Index cols)
+{
+  return Constant(rows, cols, Scalar(0));
+}
+
+/** \returns an expression of a zero vector.
+  *
+  * The parameter \a size is the size of the returned vector.
+  * Must be compatible with this MatrixBase type.
+  *
+  * \only_for_vectors
+  *
+  * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
+  * it is redundant to pass \a size as argument, so Zero() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_zero_int.cpp
+  * Output: \verbinclude MatrixBase_zero_int.out
+  *
+  * \sa Zero(), Zero(Index,Index)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Zero(Index size)
+{
+  return Constant(size, Scalar(0));
+}
+
+/** \returns an expression of a fixed-size zero matrix or vector.
+  *
+  * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
+  * need to use the variants taking size arguments.
+  *
+  * Example: \include MatrixBase_zero.cpp
+  * Output: \verbinclude MatrixBase_zero.out
+  *
+  * \sa Zero(Index), Zero(Index,Index)
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Zero()
+{
+  return Constant(Scalar(0));
+}
+
+/** \returns true if *this is approximately equal to the zero matrix,
+  *          within the precision given by \a prec.
+  *
+  * Example: \include MatrixBase_isZero.cpp
+  * Output: \verbinclude MatrixBase_isZero.out
+  *
+  * \sa class CwiseNullaryOp, Zero()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isZero(const RealScalar& prec) const
+{
+  typename internal::nested_eval<Derived,1>::type self(derived());
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = 0; i < rows(); ++i)
+      if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<Scalar>(1), prec))
+        return false;
+  return true;
+}
+
+/** Sets all coefficients in this expression to zero.
+  *
+  * Example: \include MatrixBase_setZero.cpp
+  * Output: \verbinclude MatrixBase_setZero.out
+  *
+  * \sa class CwiseNullaryOp, Zero()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setZero()
+{
+  return setConstant(Scalar(0));
+}
+
+/** Resizes to the given \a size, and sets all coefficients in this expression to zero.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include Matrix_setZero_int.cpp
+  * Output: \verbinclude Matrix_setZero_int.out
+  *
+  * \sa DenseBase::setZero(), setZero(Index,Index), class CwiseNullaryOp, DenseBase::Zero()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setZero(Index newSize)
+{
+  resize(newSize);
+  return setConstant(Scalar(0));
+}
+
+/** Resizes to the given size, and sets all coefficients in this expression to zero.
+  *
+  * \param rows the new number of rows
+  * \param cols the new number of columns
+  *
+  * Example: \include Matrix_setZero_int_int.cpp
+  * Output: \verbinclude Matrix_setZero_int_int.out
+  *
+  * \sa DenseBase::setZero(), setZero(Index), class CwiseNullaryOp, DenseBase::Zero()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setZero(Index rows, Index cols)
+{
+  resize(rows, cols);
+  return setConstant(Scalar(0));
+}
+
+// ones:
+
+/** \returns an expression of a matrix where all coefficients equal one.
+  *
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this MatrixBase type.
+  *
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Ones() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_ones_int_int.cpp
+  * Output: \verbinclude MatrixBase_ones_int_int.out
+  *
+  * \sa Ones(), Ones(Index), isOnes(), class Ones
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Ones(Index rows, Index cols)
+{
+  return Constant(rows, cols, Scalar(1));
+}
+
+/** \returns an expression of a vector where all coefficients equal one.
+  *
+  * The parameter \a newSize is the size of the returned vector.
+  * Must be compatible with this MatrixBase type.
+  *
+  * \only_for_vectors
+  *
+  * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
+  * it is redundant to pass \a size as argument, so Ones() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_ones_int.cpp
+  * Output: \verbinclude MatrixBase_ones_int.out
+  *
+  * \sa Ones(), Ones(Index,Index), isOnes(), class Ones
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Ones(Index newSize)
+{
+  return Constant(newSize, Scalar(1));
+}
+
+/** \returns an expression of a fixed-size matrix or vector where all coefficients equal one.
+  *
+  * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
+  * need to use the variants taking size arguments.
+  *
+  * Example: \include MatrixBase_ones.cpp
+  * Output: \verbinclude MatrixBase_ones.out
+  *
+  * \sa Ones(Index), Ones(Index,Index), isOnes(), class Ones
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename DenseBase<Derived>::ConstantReturnType
+DenseBase<Derived>::Ones()
+{
+  return Constant(Scalar(1));
+}
+
+/** \returns true if *this is approximately equal to the matrix where all coefficients
+  *          are equal to 1, within the precision given by \a prec.
+  *
+  * Example: \include MatrixBase_isOnes.cpp
+  * Output: \verbinclude MatrixBase_isOnes.out
+  *
+  * \sa class CwiseNullaryOp, Ones()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC bool DenseBase<Derived>::isOnes
+(const RealScalar& prec) const
+{
+  return isApproxToConstant(Scalar(1), prec);
+}
+
+/** Sets all coefficients in this expression to one.
+  *
+  * Example: \include MatrixBase_setOnes.cpp
+  * Output: \verbinclude MatrixBase_setOnes.out
+  *
+  * \sa class CwiseNullaryOp, Ones()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::setOnes()
+{
+  return setConstant(Scalar(1));
+}
+
+/** Resizes to the given \a newSize, and sets all coefficients in this expression to one.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include Matrix_setOnes_int.cpp
+  * Output: \verbinclude Matrix_setOnes_int.out
+  *
+  * \sa MatrixBase::setOnes(), setOnes(Index,Index), class CwiseNullaryOp, MatrixBase::Ones()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setOnes(Index newSize)
+{
+  resize(newSize);
+  return setConstant(Scalar(1));
+}
+
+/** Resizes to the given size, and sets all coefficients in this expression to one.
+  *
+  * \param rows the new number of rows
+  * \param cols the new number of columns
+  *
+  * Example: \include Matrix_setOnes_int_int.cpp
+  * Output: \verbinclude Matrix_setOnes_int_int.out
+  *
+  * \sa MatrixBase::setOnes(), setOnes(Index), class CwiseNullaryOp, MatrixBase::Ones()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setOnes(Index rows, Index cols)
+{
+  resize(rows, cols);
+  return setConstant(Scalar(1));
+}
+
+// Identity:
+
+/** \returns an expression of the identity matrix (not necessarily square).
+  *
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this MatrixBase type.
+  *
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Identity() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_identity_int_int.cpp
+  * Output: \verbinclude MatrixBase_identity_int_int.out
+  *
+  * \sa Identity(), setIdentity(), isIdentity()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+MatrixBase<Derived>::Identity(Index rows, Index cols)
+{
+  return DenseBase<Derived>::NullaryExpr(rows, cols, internal::scalar_identity_op<Scalar>());
+}
+
+/** \returns an expression of the identity matrix (not necessarily square).
+  *
+  * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
+  * need to use the variant taking size arguments.
+  *
+  * Example: \include MatrixBase_identity.cpp
+  * Output: \verbinclude MatrixBase_identity.out
+  *
+  * \sa Identity(Index,Index), setIdentity(), isIdentity()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::IdentityReturnType
+MatrixBase<Derived>::Identity()
+{
+  EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
+  return MatrixBase<Derived>::NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_identity_op<Scalar>());
+}
+
+/** \returns true if *this is approximately equal to the identity matrix
+  *          (not necessarily square),
+  *          within the precision given by \a prec.
+  *
+  * Example: \include MatrixBase_isIdentity.cpp
+  * Output: \verbinclude MatrixBase_isIdentity.out
+  *
+  * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), setIdentity()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isIdentity
+(const RealScalar& prec) const
+{
+  typename internal::nested_eval<Derived,1>::type self(derived());
+  for(Index j = 0; j < cols(); ++j)
+  {
+    for(Index i = 0; i < rows(); ++i)
+    {
+      if(i == j)
+      {
+        if(!internal::isApprox(self.coeff(i, j), static_cast<Scalar>(1), prec))
+          return false;
+      }
+      else
+      {
+        if(!internal::isMuchSmallerThan(self.coeff(i, j), static_cast<RealScalar>(1), prec))
+          return false;
+      }
+    }
+  }
+  return true;
+}
+
+namespace internal {
+
+template<typename Derived, bool Big = (Derived::SizeAtCompileTime>=16)>
+struct setIdentity_impl
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Derived& run(Derived& m)
+  {
+    return m = Derived::Identity(m.rows(), m.cols());
+  }
+};
+
+template<typename Derived>
+struct setIdentity_impl<Derived, true>
+{
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Derived& run(Derived& m)
+  {
+    m.setZero();
+    const Index size = numext::mini(m.rows(), m.cols());
+    for(Index i = 0; i < size; ++i) m.coeffRef(i,i) = typename Derived::Scalar(1);
+    return m;
+  }
+};
+
+} // end namespace internal
+
+/** Writes the identity expression (not necessarily square) into *this.
+  *
+  * Example: \include MatrixBase_setIdentity.cpp
+  * Output: \verbinclude MatrixBase_setIdentity.out
+  *
+  * \sa class CwiseNullaryOp, Identity(), Identity(Index,Index), isIdentity()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity()
+{
+  return internal::setIdentity_impl<Derived>::run(derived());
+}
+
+/** \brief Resizes to the given size, and writes the identity expression (not necessarily square) into *this.
+  *
+  * \param rows the new number of rows
+  * \param cols the new number of columns
+  *
+  * Example: \include Matrix_setIdentity_int_int.cpp
+  * Output: \verbinclude Matrix_setIdentity_int_int.out
+  *
+  * \sa MatrixBase::setIdentity(), class CwiseNullaryOp, MatrixBase::Identity()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& MatrixBase<Derived>::setIdentity(Index rows, Index cols)
+{
+  derived().resize(rows, cols);
+  return setIdentity();
+}
+
+/** \returns an expression of the i-th unit (basis) vector.
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::Unit(Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index newSize, Index i)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return BasisReturnType(SquareMatrixType::Identity(newSize,newSize), i);
+}
+
+/** \returns an expression of the i-th unit (basis) vector.
+  *
+  * \only_for_vectors
+  *
+  * This variant is for fixed-size vector only.
+  *
+  * \sa MatrixBase::Unit(Index,Index), MatrixBase::UnitX(), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::Unit(Index i)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return BasisReturnType(SquareMatrixType::Identity(),i);
+}
+
+/** \returns an expression of the X axis unit vector (1{,0}^*)
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitX()
+{ return Derived::Unit(0); }
+
+/** \returns an expression of the Y axis unit vector (0,1{,0}^*)
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitY()
+{ return Derived::Unit(1); }
+
+/** \returns an expression of the Z axis unit vector (0,0,1{,0}^*)
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitZ()
+{ return Derived::Unit(2); }
+
+/** \returns an expression of the W axis unit vector (0,0,0,1)
+  *
+  * \only_for_vectors
+  *
+  * \sa MatrixBase::Unit(Index,Index), MatrixBase::Unit(Index), MatrixBase::UnitY(), MatrixBase::UnitZ(), MatrixBase::UnitW()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename MatrixBase<Derived>::BasisReturnType MatrixBase<Derived>::UnitW()
+{ return Derived::Unit(3); }
+
+} // end namespace Eigen
+
+#endif // EIGEN_CWISE_NULLARY_OP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h
new file mode 100644
index 0000000..9f3576f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseTernaryOp.h
@@ -0,0 +1,197 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_TERNARY_OP_H
+#define EIGEN_CWISE_TERNARY_OP_H
+
+namespace Eigen {
+
+namespace internal {
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>
+struct traits<CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> > {
+  // we must not inherit from traits<Arg1> since it has
+  // the potential to cause problems with MSVC
+  typedef typename remove_all<Arg1>::type Ancestor;
+  typedef typename traits<Ancestor>::XprKind XprKind;
+  enum {
+    RowsAtCompileTime = traits<Ancestor>::RowsAtCompileTime,
+    ColsAtCompileTime = traits<Ancestor>::ColsAtCompileTime,
+    MaxRowsAtCompileTime = traits<Ancestor>::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = traits<Ancestor>::MaxColsAtCompileTime
+  };
+
+  // even though we require Arg1, Arg2, and Arg3 to have the same scalar type
+  // (see CwiseTernaryOp constructor),
+  // we still want to handle the case when the result type is different.
+  typedef typename result_of<TernaryOp(
+      const typename Arg1::Scalar&, const typename Arg2::Scalar&,
+      const typename Arg3::Scalar&)>::type Scalar;
+
+  typedef typename internal::traits<Arg1>::StorageKind StorageKind;
+  typedef typename internal::traits<Arg1>::StorageIndex StorageIndex;
+
+  typedef typename Arg1::Nested Arg1Nested;
+  typedef typename Arg2::Nested Arg2Nested;
+  typedef typename Arg3::Nested Arg3Nested;
+  typedef typename remove_reference<Arg1Nested>::type _Arg1Nested;
+  typedef typename remove_reference<Arg2Nested>::type _Arg2Nested;
+  typedef typename remove_reference<Arg3Nested>::type _Arg3Nested;
+  enum { Flags = _Arg1Nested::Flags & RowMajorBit };
+};
+}  // end namespace internal
+
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3,
+          typename StorageKind>
+class CwiseTernaryOpImpl;
+
+/** \class CwiseTernaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise ternary operator is
+ * applied to two expressions
+  *
+  * \tparam TernaryOp template functor implementing the operator
+  * \tparam Arg1Type the type of the first argument
+  * \tparam Arg2Type the type of the second argument
+  * \tparam Arg3Type the type of the third argument
+  *
+  * This class represents an expression where a coefficient-wise ternary
+ * operator is applied to three expressions.
+  * It is the return type of ternary operators, by which we mean only those
+ * ternary operators where
+  * all three arguments are Eigen expressions.
+  * For example, the return type of betainc(matrix1, matrix2, matrix3) is a
+ * CwiseTernaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically
+ * don't have to name
+  * CwiseTernaryOp types explicitly.
+  *
+  * \sa MatrixBase::ternaryExpr(const MatrixBase<Argument2> &, const
+ * MatrixBase<Argument3> &, const CustomTernaryOp &) const, class CwiseBinaryOp,
+ * class CwiseUnaryOp, class CwiseNullaryOp
+  */
+template <typename TernaryOp, typename Arg1Type, typename Arg2Type,
+          typename Arg3Type>
+class CwiseTernaryOp : public CwiseTernaryOpImpl<
+                           TernaryOp, Arg1Type, Arg2Type, Arg3Type,
+                           typename internal::traits<Arg1Type>::StorageKind>,
+                       internal::no_assignment_operator
+{
+ public:
+  typedef typename internal::remove_all<Arg1Type>::type Arg1;
+  typedef typename internal::remove_all<Arg2Type>::type Arg2;
+  typedef typename internal::remove_all<Arg3Type>::type Arg3;
+
+  typedef typename CwiseTernaryOpImpl<
+      TernaryOp, Arg1Type, Arg2Type, Arg3Type,
+      typename internal::traits<Arg1Type>::StorageKind>::Base Base;
+  EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseTernaryOp)
+
+  typedef typename internal::ref_selector<Arg1Type>::type Arg1Nested;
+  typedef typename internal::ref_selector<Arg2Type>::type Arg2Nested;
+  typedef typename internal::ref_selector<Arg3Type>::type Arg3Nested;
+  typedef typename internal::remove_reference<Arg1Nested>::type _Arg1Nested;
+  typedef typename internal::remove_reference<Arg2Nested>::type _Arg2Nested;
+  typedef typename internal::remove_reference<Arg3Nested>::type _Arg3Nested;
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE CwiseTernaryOp(const Arg1& a1, const Arg2& a2,
+                                     const Arg3& a3,
+                                     const TernaryOp& func = TernaryOp())
+      : m_arg1(a1), m_arg2(a2), m_arg3(a3), m_functor(func) {
+    // require the sizes to match
+    EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg2)
+    EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(Arg1, Arg3)
+
+    // The index types should match
+    EIGEN_STATIC_ASSERT((internal::is_same<
+                         typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg2Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+    EIGEN_STATIC_ASSERT((internal::is_same<
+                         typename internal::traits<Arg1Type>::StorageKind,
+                         typename internal::traits<Arg3Type>::StorageKind>::value),
+                        STORAGE_KIND_MUST_MATCH)
+
+    eigen_assert(a1.rows() == a2.rows() && a1.cols() == a2.cols() &&
+                 a1.rows() == a3.rows() && a1.cols() == a3.cols());
+  }
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Index rows() const {
+    // return the fixed size type if available to enable compile time
+    // optimizations
+    if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                RowsAtCompileTime == Dynamic &&
+        internal::traits<typename internal::remove_all<Arg2Nested>::type>::
+                RowsAtCompileTime == Dynamic)
+      return m_arg3.rows();
+    else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                     RowsAtCompileTime == Dynamic &&
+             internal::traits<typename internal::remove_all<Arg3Nested>::type>::
+                     RowsAtCompileTime == Dynamic)
+      return m_arg2.rows();
+    else
+      return m_arg1.rows();
+  }
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE Index cols() const {
+    // return the fixed size type if available to enable compile time
+    // optimizations
+    if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                ColsAtCompileTime == Dynamic &&
+        internal::traits<typename internal::remove_all<Arg2Nested>::type>::
+                ColsAtCompileTime == Dynamic)
+      return m_arg3.cols();
+    else if (internal::traits<typename internal::remove_all<Arg1Nested>::type>::
+                     ColsAtCompileTime == Dynamic &&
+             internal::traits<typename internal::remove_all<Arg3Nested>::type>::
+                     ColsAtCompileTime == Dynamic)
+      return m_arg2.cols();
+    else
+      return m_arg1.cols();
+  }
+
+  /** \returns the first argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg1Nested& arg1() const { return m_arg1; }
+  /** \returns the first argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg2Nested& arg2() const { return m_arg2; }
+  /** \returns the third argument nested expression */
+  EIGEN_DEVICE_FUNC
+  const _Arg3Nested& arg3() const { return m_arg3; }
+  /** \returns the functor representing the ternary operation */
+  EIGEN_DEVICE_FUNC
+  const TernaryOp& functor() const { return m_functor; }
+
+ protected:
+  Arg1Nested m_arg1;
+  Arg2Nested m_arg2;
+  Arg3Nested m_arg3;
+  const TernaryOp m_functor;
+};
+
+// Generic API dispatcher
+template <typename TernaryOp, typename Arg1, typename Arg2, typename Arg3,
+          typename StorageKind>
+class CwiseTernaryOpImpl
+    : public internal::generic_xpr_base<
+          CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type {
+ public:
+  typedef typename internal::generic_xpr_base<
+      CwiseTernaryOp<TernaryOp, Arg1, Arg2, Arg3> >::type Base;
+};
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_CWISE_TERNARY_OP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h
new file mode 100644
index 0000000..1d2dd19
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryOp.h
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_UNARY_OP_H
+#define EIGEN_CWISE_UNARY_OP_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename UnaryOp, typename XprType>
+struct traits<CwiseUnaryOp<UnaryOp, XprType> >
+ : traits<XprType>
+{
+  typedef typename result_of<
+                     UnaryOp(const typename XprType::Scalar&)
+                   >::type Scalar;
+  typedef typename XprType::Nested XprTypeNested;
+  typedef typename remove_reference<XprTypeNested>::type _XprTypeNested;
+  enum {
+    Flags = _XprTypeNested::Flags & RowMajorBit 
+  };
+};
+}
+
+template<typename UnaryOp, typename XprType, typename StorageKind>
+class CwiseUnaryOpImpl;
+
+/** \class CwiseUnaryOp
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression where a coefficient-wise unary operator is applied to an expression
+  *
+  * \tparam UnaryOp template functor implementing the operator
+  * \tparam XprType the type of the expression to which we are applying the unary operator
+  *
+  * This class represents an expression where a unary operator is applied to an expression.
+  * It is the return type of all operations taking exactly 1 input expression, regardless of the
+  * presence of other inputs such as scalars. For example, the operator* in the expression 3*matrix
+  * is considered unary, because only the right-hand side is an expression, and its
+  * return type is a specialization of CwiseUnaryOp.
+  *
+  * Most of the time, this is the only way that it is used, so you typically don't have to name
+  * CwiseUnaryOp types explicitly.
+  *
+  * \sa MatrixBase::unaryExpr(const CustomUnaryOp &) const, class CwiseBinaryOp, class CwiseNullaryOp
+  */
+template<typename UnaryOp, typename XprType>
+class CwiseUnaryOp : public CwiseUnaryOpImpl<UnaryOp, XprType, typename internal::traits<XprType>::StorageKind>, internal::no_assignment_operator
+{
+  public:
+
+    typedef typename CwiseUnaryOpImpl<UnaryOp, XprType,typename internal::traits<XprType>::StorageKind>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryOp)
+    typedef typename internal::ref_selector<XprType>::type XprTypeNested;
+    typedef typename internal::remove_all<XprType>::type NestedExpression;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    explicit CwiseUnaryOp(const XprType& xpr, const UnaryOp& func = UnaryOp())
+      : m_xpr(xpr), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Index rows() const { return m_xpr.rows(); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Index cols() const { return m_xpr.cols(); }
+
+    /** \returns the functor representing the unary operation */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const UnaryOp& functor() const { return m_functor; }
+
+    /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    const typename internal::remove_all<XprTypeNested>::type&
+    nestedExpression() const { return m_xpr; }
+
+    /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    typename internal::remove_all<XprTypeNested>::type&
+    nestedExpression() { return m_xpr; }
+
+  protected:
+    XprTypeNested m_xpr;
+    const UnaryOp m_functor;
+};
+
+// Generic API dispatcher
+template<typename UnaryOp, typename XprType, typename StorageKind>
+class CwiseUnaryOpImpl
+  : public internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryOp<UnaryOp, XprType> >::type Base;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CWISE_UNARY_OP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h
new file mode 100644
index 0000000..2710330
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/CwiseUnaryView.h
@@ -0,0 +1,128 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CWISE_UNARY_VIEW_H
+#define EIGEN_CWISE_UNARY_VIEW_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename ViewOp, typename MatrixType>
+struct traits<CwiseUnaryView<ViewOp, MatrixType> >
+ : traits<MatrixType>
+{
+  typedef typename result_of<
+                     ViewOp(const typename traits<MatrixType>::Scalar&)
+                   >::type Scalar;
+  typedef typename MatrixType::Nested MatrixTypeNested;
+  typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = traits<_MatrixTypeNested>::Flags & (RowMajorBit | FlagsLvalueBit | DirectAccessBit), // FIXME DirectAccessBit should not be handled by expressions
+    MatrixTypeInnerStride =  inner_stride_at_compile_time<MatrixType>::ret,
+    // need to cast the sizeof's from size_t to int explicitly, otherwise:
+    // "error: no integral type can represent all of the enumerator values
+    InnerStrideAtCompileTime = MatrixTypeInnerStride == Dynamic
+                             ? int(Dynamic)
+                             : int(MatrixTypeInnerStride) * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar)),
+    OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret == Dynamic
+                             ? int(Dynamic)
+                             : outer_stride_at_compile_time<MatrixType>::ret * int(sizeof(typename traits<MatrixType>::Scalar) / sizeof(Scalar))
+  };
+};
+}
+
+template<typename ViewOp, typename MatrixType, typename StorageKind>
+class CwiseUnaryViewImpl;
+
+/** \class CwiseUnaryView
+  * \ingroup Core_Module
+  *
+  * \brief Generic lvalue expression of a coefficient-wise unary operator of a matrix or a vector
+  *
+  * \tparam ViewOp template functor implementing the view
+  * \tparam MatrixType the type of the matrix we are applying the unary operator
+  *
+  * This class represents a lvalue expression of a generic unary view operator of a matrix or a vector.
+  * It is the return type of real() and imag(), and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::unaryViewExpr(const CustomUnaryOp &) const, class CwiseUnaryOp
+  */
+template<typename ViewOp, typename MatrixType>
+class CwiseUnaryView : public CwiseUnaryViewImpl<ViewOp, MatrixType, typename internal::traits<MatrixType>::StorageKind>
+{
+  public:
+
+    typedef typename CwiseUnaryViewImpl<ViewOp, MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(CwiseUnaryView)
+    typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
+
+    explicit inline CwiseUnaryView(MatrixType& mat, const ViewOp& func = ViewOp())
+      : m_matrix(mat), m_functor(func) {}
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryView)
+
+    EIGEN_STRONG_INLINE Index rows() const { return m_matrix.rows(); }
+    EIGEN_STRONG_INLINE Index cols() const { return m_matrix.cols(); }
+
+    /** \returns the functor representing unary operation */
+    const ViewOp& functor() const { return m_functor; }
+
+    /** \returns the nested expression */
+    const typename internal::remove_all<MatrixTypeNested>::type&
+    nestedExpression() const { return m_matrix; }
+
+    /** \returns the nested expression */
+    typename internal::remove_reference<MatrixTypeNested>::type&
+    nestedExpression() { return m_matrix.const_cast_derived(); }
+
+  protected:
+    MatrixTypeNested m_matrix;
+    ViewOp m_functor;
+};
+
+// Generic API dispatcher
+template<typename ViewOp, typename XprType, typename StorageKind>
+class CwiseUnaryViewImpl
+  : public internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<CwiseUnaryView<ViewOp, XprType> >::type Base;
+};
+
+template<typename ViewOp, typename MatrixType>
+class CwiseUnaryViewImpl<ViewOp,MatrixType,Dense>
+  : public internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type
+{
+  public:
+
+    typedef CwiseUnaryView<ViewOp, MatrixType> Derived;
+    typedef typename internal::dense_xpr_base< CwiseUnaryView<ViewOp, MatrixType> >::type Base;
+
+    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(CwiseUnaryViewImpl)
+    
+    EIGEN_DEVICE_FUNC inline Scalar* data() { return &(this->coeffRef(0)); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return &(this->coeff(0)); }
+
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
+    {
+      return derived().nestedExpression().innerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
+    }
+
+    EIGEN_DEVICE_FUNC inline Index outerStride() const
+    {
+      return derived().nestedExpression().outerStride() * sizeof(typename internal::traits<MatrixType>::Scalar) / sizeof(Scalar);
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CWISE_UNARY_VIEW_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h
new file mode 100644
index 0000000..90066ae
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseBase.h
@@ -0,0 +1,611 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DENSEBASE_H
+#define EIGEN_DENSEBASE_H
+
+namespace Eigen {
+
+namespace internal {
+  
+// The index type defined by EIGEN_DEFAULT_DENSE_INDEX_TYPE must be a signed type.
+// This dummy function simply aims at checking that at compile time.
+static inline void check_DenseIndex_is_signed() {
+  EIGEN_STATIC_ASSERT(NumTraits<DenseIndex>::IsSigned,THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE); 
+}
+
+} // end namespace internal
+  
+/** \class DenseBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for all dense matrices, vectors, and arrays
+  *
+  * This class is the base that is inherited by all dense objects (matrix, vector, arrays,
+  * and related expression types). The common Eigen API for dense objects is contained in this class.
+  *
+  * \tparam Derived is the derived type, e.g., a matrix type or an expression.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_DENSEBASE_PLUGIN.
+  *
+  * \sa \blank \ref TopicClassHierarchy
+  */
+template<typename Derived> class DenseBase
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  : public DenseCoeffsBase<Derived>
+#else
+  : public DenseCoeffsBase<Derived,DirectWriteAccessors>
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+{
+  public:
+
+    /** Inner iterator type to iterate over the coefficients of a row or column.
+      * \sa class InnerIterator
+      */
+    typedef Eigen::InnerIterator<Derived> InnerIterator;
+
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+
+    /**
+      * \brief The type used to store indices
+      * \details This typedef is relevant for types that store multiple indices such as
+      *          PermutationMatrix or Transpositions, otherwise it defaults to Eigen::Index
+      * \sa \blank \ref TopicPreprocessorDirectives, Eigen::Index, SparseMatrixBase.
+     */
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc. */
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc.
+      *
+      * It is an alias for the Scalar type */
+    typedef Scalar value_type;
+    
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef DenseCoeffsBase<Derived> Base;
+
+    using Base::derived;
+    using Base::const_cast_derived;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::rowIndexByOuterInner;
+    using Base::colIndexByOuterInner;
+    using Base::coeff;
+    using Base::coeffByOuterInner;
+    using Base::operator();
+    using Base::operator[];
+    using Base::x;
+    using Base::y;
+    using Base::z;
+    using Base::w;
+    using Base::stride;
+    using Base::innerStride;
+    using Base::outerStride;
+    using Base::rowStride;
+    using Base::colStride;
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+
+    enum {
+
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+        /**< The number of rows at compile-time. This is just a copy of the value provided
+          * by the \a Derived type. If a value is not known at compile-time,
+          * it is set to the \a Dynamic constant.
+          * \sa MatrixBase::rows(), MatrixBase::cols(), ColsAtCompileTime, SizeAtCompileTime */
+
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+        /**< The number of columns at compile-time. This is just a copy of the value provided
+          * by the \a Derived type. If a value is not known at compile-time,
+          * it is set to the \a Dynamic constant.
+          * \sa MatrixBase::rows(), MatrixBase::cols(), RowsAtCompileTime, SizeAtCompileTime */
+
+
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                   internal::traits<Derived>::ColsAtCompileTime>::ret),
+        /**< This is equal to the number of coefficients, i.e. the number of
+          * rows times the number of columns, or to \a Dynamic if this is not
+          * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */
+
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+        /**< This value is equal to the maximum possible number of rows that this expression
+          * might have. If this expression might have an arbitrarily high number of rows,
+          * this value is set to \a Dynamic.
+          *
+          * This value is useful to know when evaluating an expression, in order to determine
+          * whether it is possible to avoid doing a dynamic memory allocation.
+          *
+          * \sa RowsAtCompileTime, MaxColsAtCompileTime, MaxSizeAtCompileTime
+          */
+
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+        /**< This value is equal to the maximum possible number of columns that this expression
+          * might have. If this expression might have an arbitrarily high number of columns,
+          * this value is set to \a Dynamic.
+          *
+          * This value is useful to know when evaluating an expression, in order to determine
+          * whether it is possible to avoid doing a dynamic memory allocation.
+          *
+          * \sa ColsAtCompileTime, MaxRowsAtCompileTime, MaxSizeAtCompileTime
+          */
+
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                      internal::traits<Derived>::MaxColsAtCompileTime>::ret),
+        /**< This value is equal to the maximum possible number of coefficients that this expression
+          * might have. If this expression might have an arbitrarily high number of coefficients,
+          * this value is set to \a Dynamic.
+          *
+          * This value is useful to know when evaluating an expression, in order to determine
+          * whether it is possible to avoid doing a dynamic memory allocation.
+          *
+          * \sa SizeAtCompileTime, MaxRowsAtCompileTime, MaxColsAtCompileTime
+          */
+
+      IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1
+                           || internal::traits<Derived>::MaxColsAtCompileTime == 1,
+        /**< This is set to true if either the number of rows or the number of
+          * columns is known at compile-time to be equal to 1. Indeed, in that case,
+          * we are dealing with a column-vector (if there is only one column) or with
+          * a row-vector (if there is only one row). */
+
+      Flags = internal::traits<Derived>::Flags,
+        /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
+          * constructed from this one. See the \ref flags "list of flags".
+          */
+
+      IsRowMajor = int(Flags) & RowMajorBit, /**< True if this expression has row-major storage order. */
+
+      InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime)
+                             : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
+
+      InnerStrideAtCompileTime = internal::inner_stride_at_compile_time<Derived>::ret,
+      OuterStrideAtCompileTime = internal::outer_stride_at_compile_time<Derived>::ret
+    };
+    
+    typedef typename internal::find_best_packet<Scalar,SizeAtCompileTime>::type PacketScalar;
+
+    enum { IsPlainObjectBase = 0 };
+    
+    /** The plain matrix type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Matrix<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainMatrix;
+    
+    /** The plain array type corresponding to this expression.
+      * \sa PlainObject */
+    typedef Array<typename internal::traits<Derived>::Scalar,
+                internal::traits<Derived>::RowsAtCompileTime,
+                internal::traits<Derived>::ColsAtCompileTime,
+                AutoAlign | (internal::traits<Derived>::Flags&RowMajorBit ? RowMajor : ColMajor),
+                internal::traits<Derived>::MaxRowsAtCompileTime,
+                internal::traits<Derived>::MaxColsAtCompileTime
+          > PlainArray;
+
+    /** \brief The plain matrix or array type corresponding to this expression.
+      *
+      * This is not necessarily exactly the return type of eval(). In the case of plain matrices,
+      * the return type of eval() is a const reference to a matrix, not a matrix! It is however guaranteed
+      * that the return type of eval() is either PlainObject or const PlainObject&.
+      */
+    typedef typename internal::conditional<internal::is_same<typename internal::traits<Derived>::XprKind,MatrixXpr >::value,
+                                 PlainMatrix, PlainArray>::type PlainObject;
+
+    /** \returns the number of nonzero coefficients which is in practice the number
+      * of stored coefficients. */
+    EIGEN_DEVICE_FUNC
+    inline Index nonZeros() const { return size(); }
+
+    /** \returns the outer size.
+      *
+      * \note For a vector, this returns just 1. For a matrix (non-vector), this is the major dimension
+      * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of columns for a
+      * column-major matrix, and the number of rows for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
+    Index outerSize() const
+    {
+      return IsVectorAtCompileTime ? 1
+           : int(IsRowMajor) ? this->rows() : this->cols();
+    }
+
+    /** \returns the inner size.
+      *
+      * \note For a vector, this is just the size. For a matrix (non-vector), this is the minor dimension
+      * with respect to the \ref TopicStorageOrders "storage order", i.e., the number of rows for a 
+      * column-major matrix, and the number of columns for a row-major matrix. */
+    EIGEN_DEVICE_FUNC
+    Index innerSize() const
+    {
+      return IsVectorAtCompileTime ? this->size()
+           : int(IsRowMajor) ? this->cols() : this->rows();
+    }
+
+    /** Only plain matrices/arrays, not expressions, may be resized; therefore the only useful resize methods are
+      * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
+      * nothing else.
+      */
+    EIGEN_DEVICE_FUNC
+    void resize(Index newSize)
+    {
+      EIGEN_ONLY_USED_FOR_DEBUG(newSize);
+      eigen_assert(newSize == this->size()
+                && "DenseBase::resize() does not actually allow to resize.");
+    }
+    /** Only plain matrices/arrays, not expressions, may be resized; therefore the only useful resize methods are
+      * Matrix::resize() and Array::resize(). The present method only asserts that the new size equals the old size, and does
+      * nothing else.
+      */
+    EIGEN_DEVICE_FUNC
+    void resize(Index rows, Index cols)
+    {
+      EIGEN_ONLY_USED_FOR_DEBUG(rows);
+      EIGEN_ONLY_USED_FOR_DEBUG(cols);
+      eigen_assert(rows == this->rows() && cols == this->cols()
+                && "DenseBase::resize() does not actually allow to resize.");
+    }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal Represents a matrix with all coefficients equal to one another*/
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
+    /** \internal \deprecated Represents a vector with linearly spaced coefficients that allows sequential access only. */
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> SequentialLinSpacedReturnType;
+    /** \internal Represents a vector with linearly spaced coefficients that allows random access. */
+    typedef CwiseNullaryOp<internal::linspaced_op<Scalar,PacketScalar>,PlainObject> RandomAccessLinSpacedReturnType;
+    /** \internal the return type of MatrixBase::eigenvalues() */
+    typedef Matrix<typename NumTraits<typename internal::traits<Derived>::Scalar>::Real, internal::traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType;
+
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    /** Copies \a other into *this. \returns a reference to *this. */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const DenseBase<OtherDerived>& other);
+
+    /** Special case of the template operator=, in order to prevent the compiler
+      * from generating a default operator= (issue hit with g++ 4.1)
+      */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const DenseBase& other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const EigenBase<OtherDerived> &other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator+=(const EigenBase<OtherDerived> &other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator-=(const EigenBase<OtherDerived> &other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const ReturnByValue<OtherDerived>& func);
+
+    /** \internal
+      * Copies \a other into *this without evaluating other. \returns a reference to *this.
+      * \deprecated */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& lazyAssign(const DenseBase<OtherDerived>& other);
+
+    EIGEN_DEVICE_FUNC
+    CommaInitializer<Derived> operator<< (const Scalar& s);
+
+    /** \deprecated it now returns \c *this */
+    template<unsigned int Added,unsigned int Removed>
+    EIGEN_DEPRECATED
+    const Derived& flagged() const
+    { return derived(); }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    CommaInitializer<Derived> operator<< (const DenseBase<OtherDerived>& other);
+
+    typedef Transpose<Derived> TransposeReturnType;
+    EIGEN_DEVICE_FUNC
+    TransposeReturnType transpose();
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    EIGEN_DEVICE_FUNC
+    ConstTransposeReturnType transpose() const;
+    EIGEN_DEVICE_FUNC
+    void transposeInPlace();
+
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
+    Constant(Index rows, Index cols, const Scalar& value);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
+    Constant(Index size, const Scalar& value);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType
+    Constant(const Scalar& value);
+
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
+    LinSpaced(Sequential_t, Index size, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
+    LinSpaced(Index size, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC static const SequentialLinSpacedReturnType
+    LinSpaced(Sequential_t, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC static const RandomAccessLinSpacedReturnType
+    LinSpaced(const Scalar& low, const Scalar& high);
+
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
+    NullaryExpr(Index rows, Index cols, const CustomNullaryOp& func);
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
+    NullaryExpr(Index size, const CustomNullaryOp& func);
+    template<typename CustomNullaryOp> EIGEN_DEVICE_FUNC
+    static const CwiseNullaryOp<CustomNullaryOp, PlainObject>
+    NullaryExpr(const CustomNullaryOp& func);
+
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Zero();
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones(Index size);
+    EIGEN_DEVICE_FUNC static const ConstantReturnType Ones();
+
+    EIGEN_DEVICE_FUNC void fill(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setConstant(const Scalar& value);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(Index size, const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setLinSpaced(const Scalar& low, const Scalar& high);
+    EIGEN_DEVICE_FUNC Derived& setZero();
+    EIGEN_DEVICE_FUNC Derived& setOnes();
+    EIGEN_DEVICE_FUNC Derived& setRandom();
+
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
+    bool isApprox(const DenseBase<OtherDerived>& other,
+                  const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC 
+    bool isMuchSmallerThan(const RealScalar& other,
+                           const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC
+    bool isMuchSmallerThan(const DenseBase<OtherDerived>& other,
+                           const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+
+    EIGEN_DEVICE_FUNC bool isApproxToConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isConstant(const Scalar& value, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isZero(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC bool isOnes(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    
+    inline bool hasNaN() const;
+    inline bool allFinite() const;
+
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator*=(const Scalar& other);
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator/=(const Scalar& other);
+
+    typedef typename internal::add_const_on_value_type<typename internal::eval<Derived>::type>::type EvalReturnType;
+    /** \returns the matrix or vector obtained by evaluating this expression.
+      *
+      * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
+      * a const reference, in order to avoid a useless copy.
+      * 
+      * \warning Be carefull with eval() and the auto C++ keyword, as detailed in this \link TopicPitfalls_auto_keyword page \endlink.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE EvalReturnType eval() const
+    {
+      // Even though MSVC does not honor strong inlining when the return type
+      // is a dynamic matrix, we desperately need strong inlining for fixed
+      // size types on MSVC.
+      return typename internal::eval<Derived>::type(derived());
+    }
+    
+    /** swaps *this with the expression \a other.
+      *
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT(!OtherDerived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
+    }
+
+    /** swaps *this with the matrix or array \a other.
+      *
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(PlainObjectBase<OtherDerived>& other)
+    {
+      eigen_assert(rows()==other.rows() && cols()==other.cols());
+      call_assignment(derived(), other.derived(), internal::swap_assign_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC inline const NestByValue<Derived> nestByValue() const;
+    EIGEN_DEVICE_FUNC inline const ForceAlignedAccess<Derived> forceAlignedAccess() const;
+    EIGEN_DEVICE_FUNC inline ForceAlignedAccess<Derived> forceAlignedAccess();
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline const typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf() const;
+    template<bool Enable> EIGEN_DEVICE_FUNC
+    inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type forceAlignedAccessIf();
+
+    EIGEN_DEVICE_FUNC Scalar sum() const;
+    EIGEN_DEVICE_FUNC Scalar mean() const;
+    EIGEN_DEVICE_FUNC Scalar trace() const;
+
+    EIGEN_DEVICE_FUNC Scalar prod() const;
+
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar minCoeff() const;
+    EIGEN_DEVICE_FUNC typename internal::traits<Derived>::Scalar maxCoeff() const;
+
+    template<typename IndexType> EIGEN_DEVICE_FUNC
+    typename internal::traits<Derived>::Scalar minCoeff(IndexType* row, IndexType* col) const;
+    template<typename IndexType> EIGEN_DEVICE_FUNC
+    typename internal::traits<Derived>::Scalar maxCoeff(IndexType* row, IndexType* col) const;
+    template<typename IndexType> EIGEN_DEVICE_FUNC
+    typename internal::traits<Derived>::Scalar minCoeff(IndexType* index) const;
+    template<typename IndexType> EIGEN_DEVICE_FUNC
+    typename internal::traits<Derived>::Scalar maxCoeff(IndexType* index) const;
+
+    template<typename BinaryOp>
+    EIGEN_DEVICE_FUNC
+    Scalar redux(const BinaryOp& func) const;
+
+    template<typename Visitor>
+    EIGEN_DEVICE_FUNC
+    void visit(Visitor& func) const;
+
+    /** \returns a WithFormat proxy object allowing to print a matrix the with given
+      * format \a fmt.
+      *
+      * See class IOFormat for some examples.
+      *
+      * \sa class IOFormat, class WithFormat
+      */
+    inline const WithFormat<Derived> format(const IOFormat& fmt) const
+    {
+      return WithFormat<Derived>(derived(), fmt);
+    }
+
+    /** \returns the unique coefficient of a 1x1 expression */
+    EIGEN_DEVICE_FUNC
+    CoeffReturnType value() const
+    {
+      EIGEN_STATIC_ASSERT_SIZE_1x1(Derived)
+      eigen_assert(this->rows() == 1 && this->cols() == 1);
+      return derived().coeff(0,0);
+    }
+
+    EIGEN_DEVICE_FUNC bool all() const;
+    EIGEN_DEVICE_FUNC bool any() const;
+    EIGEN_DEVICE_FUNC Index count() const;
+
+    typedef VectorwiseOp<Derived, Horizontal> RowwiseReturnType;
+    typedef const VectorwiseOp<const Derived, Horizontal> ConstRowwiseReturnType;
+    typedef VectorwiseOp<Derived, Vertical> ColwiseReturnType;
+    typedef const VectorwiseOp<const Derived, Vertical> ConstColwiseReturnType;
+
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_rowwise.cpp
+    * Output: \verbinclude MatrixBase_rowwise.out
+    *
+    * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC inline ConstRowwiseReturnType rowwise() const {
+      return ConstRowwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC RowwiseReturnType rowwise();
+
+    /** \returns a VectorwiseOp wrapper of *this providing additional partial reduction operations
+    *
+    * Example: \include MatrixBase_colwise.cpp
+    * Output: \verbinclude MatrixBase_colwise.out
+    *
+    * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+    */
+    EIGEN_DEVICE_FUNC inline ConstColwiseReturnType colwise() const {
+      return ConstColwiseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC ColwiseReturnType colwise();
+
+    typedef CwiseNullaryOp<internal::scalar_random_op<Scalar>,PlainObject> RandomReturnType;
+    static const RandomReturnType Random(Index rows, Index cols);
+    static const RandomReturnType Random(Index size);
+    static const RandomReturnType Random();
+
+    template<typename ThenDerived,typename ElseDerived>
+    const Select<Derived,ThenDerived,ElseDerived>
+    select(const DenseBase<ThenDerived>& thenMatrix,
+           const DenseBase<ElseDerived>& elseMatrix) const;
+
+    template<typename ThenDerived>
+    inline const Select<Derived,ThenDerived, typename ThenDerived::ConstantReturnType>
+    select(const DenseBase<ThenDerived>& thenMatrix, const typename ThenDerived::Scalar& elseScalar) const;
+
+    template<typename ElseDerived>
+    inline const Select<Derived, typename ElseDerived::ConstantReturnType, ElseDerived >
+    select(const typename ElseDerived::Scalar& thenScalar, const DenseBase<ElseDerived>& elseMatrix) const;
+
+    template<int p> RealScalar lpNorm() const;
+
+    template<int RowFactor, int ColFactor>
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived,RowFactor,ColFactor> replicate() const;
+    /**
+    * \return an expression of the replication of \c *this
+    *
+    * Example: \include MatrixBase_replicate_int_int.cpp
+    * Output: \verbinclude MatrixBase_replicate_int_int.out
+    *
+    * \sa VectorwiseOp::replicate(), DenseBase::replicate<int,int>(), class Replicate
+    */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC
+    const Replicate<Derived, Dynamic, Dynamic> replicate(Index rowFactor, Index colFactor) const
+    {
+      return Replicate<Derived, Dynamic, Dynamic>(derived(), rowFactor, colFactor);
+    }
+
+    typedef Reverse<Derived, BothDirections> ReverseReturnType;
+    typedef const Reverse<const Derived, BothDirections> ConstReverseReturnType;
+    EIGEN_DEVICE_FUNC ReverseReturnType reverse();
+    /** This is the const version of reverse(). */
+    //Code moved here due to a CUDA compiler bug
+    EIGEN_DEVICE_FUNC ConstReverseReturnType reverse() const
+    {
+      return ConstReverseReturnType(derived());
+    }
+    EIGEN_DEVICE_FUNC void reverseInPlace();
+
+#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::DenseBase
+#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND)
+#   include "../plugins/BlockMethods.h"
+#   ifdef EIGEN_DENSEBASE_PLUGIN
+#     include EIGEN_DENSEBASE_PLUGIN
+#   endif
+#undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#undef EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF
+
+    // disable the use of evalTo for dense objects with a nice compilation error
+    template<typename Dest>
+    EIGEN_DEVICE_FUNC
+    inline void evalTo(Dest& ) const
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<Dest,void>::value),THE_EVAL_EVALTO_FUNCTION_SHOULD_NEVER_BE_CALLED_FOR_DENSE_OBJECTS);
+    }
+
+  protected:
+    /** Default constructor. Do nothing. */
+    EIGEN_DEVICE_FUNC DenseBase()
+    {
+      /* Just checks for self-consistency of the flags.
+       * Only do it when debugging Eigen, as this borders on paranoiac and could slow compilation down
+       */
+#ifdef EIGEN_INTERNAL_DEBUGGING
+      EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, int(IsRowMajor))
+                        && EIGEN_IMPLIES(MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1, int(!IsRowMajor))),
+                          INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION)
+#endif
+    }
+
+  private:
+    EIGEN_DEVICE_FUNC explicit DenseBase(int);
+    EIGEN_DEVICE_FUNC DenseBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit DenseBase(const DenseBase<OtherDerived>&);
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_DENSEBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h
new file mode 100644
index 0000000..c4af48a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseCoeffsBase.h
@@ -0,0 +1,681 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DENSECOEFFSBASE_H
+#define EIGEN_DENSECOEFFSBASE_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename T> struct add_const_on_value_type_if_arithmetic
+{
+  typedef typename conditional<is_arithmetic<T>::value, T, typename add_const_on_value_type<T>::type>::type type;
+};
+}
+
+/** \brief Base class providing read-only coefficient access to matrices and arrays.
+  * \ingroup Core_Module
+  * \tparam Derived Type of the derived class
+  * \tparam #ReadOnlyAccessors Constant indicating read-only access
+  *
+  * This class defines the \c operator() \c const function and friends, which can be used to read specific
+  * entries of a matrix or array.
+  * 
+  * \sa DenseCoeffsBase<Derived, WriteAccessors>, DenseCoeffsBase<Derived, DirectAccessors>,
+  *     \ref TopicClassHierarchy
+  */
+template<typename Derived>
+class DenseCoeffsBase<Derived,ReadOnlyAccessors> : public EigenBase<Derived>
+{
+  public:
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+
+    // Explanation for this CoeffReturnType typedef.
+    // - This is the return type of the coeff() method.
+    // - The LvalueBit means exactly that we can offer a coeffRef() method, which means exactly that we can get references
+    // to coeffs, which means exactly that we can have coeff() return a const reference (as opposed to returning a value).
+    // - The is_artihmetic check is required since "const int", "const double", etc. will cause warnings on some systems
+    // while the declaration of "const T", where T is a non arithmetic type does not. Always returning "const Scalar&" is
+    // not possible, since the underlying expressions might not offer a valid address the reference could be referring to.
+    typedef typename internal::conditional<bool(internal::traits<Derived>::Flags&LvalueBit),
+                         const Scalar&,
+                         typename internal::conditional<internal::is_arithmetic<Scalar>::value, Scalar, const Scalar>::type
+                     >::type CoeffReturnType;
+
+    typedef typename internal::add_const_on_value_type_if_arithmetic<
+                         typename internal::packet_traits<Scalar>::type
+                     >::type PacketReturnType;
+
+    typedef EigenBase<Derived> Base;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::derived;
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rowIndexByOuterInner(Index outer, Index inner) const
+    {
+      return int(Derived::RowsAtCompileTime) == 1 ? 0
+          : int(Derived::ColsAtCompileTime) == 1 ? inner
+          : int(Derived::Flags)&RowMajorBit ? outer
+          : inner;
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index colIndexByOuterInner(Index outer, Index inner) const
+    {
+      return int(Derived::ColsAtCompileTime) == 1 ? 0
+          : int(Derived::RowsAtCompileTime) == 1 ? inner
+          : int(Derived::Flags)&RowMajorBit ? inner
+          : outer;
+    }
+
+    /** Short version: don't use this function, use
+      * \link operator()(Index,Index) const \endlink instead.
+      *
+      * Long version: this function is similar to
+      * \link operator()(Index,Index) const \endlink, but without the assertion.
+      * Use this for limiting the performance cost of debugging code when doing
+      * repeated coefficient access. Only use this when it is guaranteed that the
+      * parameters \a row and \a col are in range.
+      *
+      * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this
+      * function equivalent to \link operator()(Index,Index) const \endlink.
+      *
+      * \sa operator()(Index,Index) const, coeffRef(Index,Index), coeff(Index) const
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType coeff(Index row, Index col) const
+    {
+      eigen_internal_assert(row >= 0 && row < rows()
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
+    {
+      return coeff(rowIndexByOuterInner(outer, inner),
+                   colIndexByOuterInner(outer, inner));
+    }
+
+    /** \returns the coefficient at given the given row and column.
+      *
+      * \sa operator()(Index,Index), operator[](Index)
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType operator()(Index row, Index col) const
+    {
+      eigen_assert(row >= 0 && row < rows()
+          && col >= 0 && col < cols());
+      return coeff(row, col);
+    }
+
+    /** Short version: don't use this function, use
+      * \link operator[](Index) const \endlink instead.
+      *
+      * Long version: this function is similar to
+      * \link operator[](Index) const \endlink, but without the assertion.
+      * Use this for limiting the performance cost of debugging code when doing
+      * repeated coefficient access. Only use this when it is guaranteed that the
+      * parameter \a index is in range.
+      *
+      * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this
+      * function equivalent to \link operator[](Index) const \endlink.
+      *
+      * \sa operator[](Index) const, coeffRef(Index), coeff(Index,Index) const
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    coeff(Index index) const
+    {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
+      eigen_internal_assert(index >= 0 && index < size());
+      return internal::evaluator<Derived>(derived()).coeff(index);
+    }
+
+
+    /** \returns the coefficient at given index.
+      *
+      * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit.
+      *
+      * \sa operator[](Index), operator()(Index,Index) const, x() const, y() const,
+      * z() const, w() const
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    operator[](Index index) const
+    {
+      EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
+                          THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
+      eigen_assert(index >= 0 && index < size());
+      return coeff(index);
+    }
+
+    /** \returns the coefficient at given index.
+      *
+      * This is synonymous to operator[](Index) const.
+      *
+      * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit.
+      *
+      * \sa operator[](Index), operator()(Index,Index) const, x() const, y() const,
+      * z() const, w() const
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    operator()(Index index) const
+    {
+      eigen_assert(index >= 0 && index < size());
+      return coeff(index);
+    }
+
+    /** equivalent to operator[](0).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    x() const { return (*this)[0]; }
+
+    /** equivalent to operator[](1).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    y() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS);
+      return (*this)[1];
+    }
+
+    /** equivalent to operator[](2).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    z() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS);
+      return (*this)[2];
+    }
+
+    /** equivalent to operator[](3).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE CoeffReturnType
+    w() const
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS);
+      return (*this)[3];
+    }
+
+    /** \internal
+      * \returns the packet of coefficients starting at the given row and column. It is your responsibility
+      * to ensure that a packet really starts there. This method is only available on expressions having the
+      * PacketAccessBit.
+      *
+      * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select
+      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
+      * starting at an address which is a multiple of the packet size.
+      */
+
+    template<int LoadMode>
+    EIGEN_STRONG_INLINE PacketReturnType packet(Index row, Index col) const
+    {
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
+      eigen_internal_assert(row >= 0 && row < rows() && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(row,col);
+    }
+
+
+    /** \internal */
+    template<int LoadMode>
+    EIGEN_STRONG_INLINE PacketReturnType packetByOuterInner(Index outer, Index inner) const
+    {
+      return packet<LoadMode>(rowIndexByOuterInner(outer, inner),
+                              colIndexByOuterInner(outer, inner));
+    }
+
+    /** \internal
+      * \returns the packet of coefficients starting at the given index. It is your responsibility
+      * to ensure that a packet really starts there. This method is only available on expressions having the
+      * PacketAccessBit and the LinearAccessBit.
+      *
+      * The \a LoadMode parameter may have the value \a #Aligned or \a #Unaligned. Its effect is to select
+      * the appropriate vectorization instruction. Aligned access is faster, but is only possible for packets
+      * starting at an address which is a multiple of the packet size.
+      */
+
+    template<int LoadMode>
+    EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+    {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
+      typedef typename internal::packet_traits<Scalar>::type DefaultPacketType;
+      eigen_internal_assert(index >= 0 && index < size());
+      return internal::evaluator<Derived>(derived()).template packet<LoadMode,DefaultPacketType>(index);
+    }
+
+  protected:
+    // explanation: DenseBase is doing "using ..." on the methods from DenseCoeffsBase.
+    // But some methods are only available in the DirectAccess case.
+    // So we add dummy methods here with these names, so that "using... " doesn't fail.
+    // It's not private so that the child class DenseBase can access them, and it's not public
+    // either since it's an implementation detail, so has to be protected.
+    void coeffRef();
+    void coeffRefByOuterInner();
+    void writePacket();
+    void writePacketByOuterInner();
+    void copyCoeff();
+    void copyCoeffByOuterInner();
+    void copyPacket();
+    void copyPacketByOuterInner();
+    void stride();
+    void innerStride();
+    void outerStride();
+    void rowStride();
+    void colStride();
+};
+
+/** \brief Base class providing read/write coefficient access to matrices and arrays.
+  * \ingroup Core_Module
+  * \tparam Derived Type of the derived class
+  * \tparam #WriteAccessors Constant indicating read/write access
+  *
+  * This class defines the non-const \c operator() function and friends, which can be used to write specific
+  * entries of a matrix or array. This class inherits DenseCoeffsBase<Derived, ReadOnlyAccessors> which
+  * defines the const variant for reading specific entries.
+  * 
+  * \sa DenseCoeffsBase<Derived, DirectAccessors>, \ref TopicClassHierarchy
+  */
+template<typename Derived>
+class DenseCoeffsBase<Derived, WriteAccessors> : public DenseCoeffsBase<Derived, ReadOnlyAccessors>
+{
+  public:
+
+    typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
+
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    using Base::coeff;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::derived;
+    using Base::rowIndexByOuterInner;
+    using Base::colIndexByOuterInner;
+    using Base::operator[];
+    using Base::operator();
+    using Base::x;
+    using Base::y;
+    using Base::z;
+    using Base::w;
+
+    /** Short version: don't use this function, use
+      * \link operator()(Index,Index) \endlink instead.
+      *
+      * Long version: this function is similar to
+      * \link operator()(Index,Index) \endlink, but without the assertion.
+      * Use this for limiting the performance cost of debugging code when doing
+      * repeated coefficient access. Only use this when it is guaranteed that the
+      * parameters \a row and \a col are in range.
+      *
+      * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this
+      * function equivalent to \link operator()(Index,Index) \endlink.
+      *
+      * \sa operator()(Index,Index), coeff(Index, Index) const, coeffRef(Index)
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index row, Index col)
+    {
+      eigen_internal_assert(row >= 0 && row < rows()
+                         && col >= 0 && col < cols());
+      return internal::evaluator<Derived>(derived()).coeffRef(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    coeffRefByOuterInner(Index outer, Index inner)
+    {
+      return coeffRef(rowIndexByOuterInner(outer, inner),
+                      colIndexByOuterInner(outer, inner));
+    }
+
+    /** \returns a reference to the coefficient at given the given row and column.
+      *
+      * \sa operator[](Index)
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    operator()(Index row, Index col)
+    {
+      eigen_assert(row >= 0 && row < rows()
+          && col >= 0 && col < cols());
+      return coeffRef(row, col);
+    }
+
+
+    /** Short version: don't use this function, use
+      * \link operator[](Index) \endlink instead.
+      *
+      * Long version: this function is similar to
+      * \link operator[](Index) \endlink, but without the assertion.
+      * Use this for limiting the performance cost of debugging code when doing
+      * repeated coefficient access. Only use this when it is guaranteed that the
+      * parameters \a row and \a col are in range.
+      *
+      * If EIGEN_INTERNAL_DEBUGGING is defined, an assertion will be made, making this
+      * function equivalent to \link operator[](Index) \endlink.
+      *
+      * \sa operator[](Index), coeff(Index) const, coeffRef(Index,Index)
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    coeffRef(Index index)
+    {
+      EIGEN_STATIC_ASSERT(internal::evaluator<Derived>::Flags & LinearAccessBit,
+                          THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS)
+      eigen_internal_assert(index >= 0 && index < size());
+      return internal::evaluator<Derived>(derived()).coeffRef(index);
+    }
+
+    /** \returns a reference to the coefficient at given index.
+      *
+      * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit.
+      *
+      * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    operator[](Index index)
+    {
+      EIGEN_STATIC_ASSERT(Derived::IsVectorAtCompileTime,
+                          THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD)
+      eigen_assert(index >= 0 && index < size());
+      return coeffRef(index);
+    }
+
+    /** \returns a reference to the coefficient at given index.
+      *
+      * This is synonymous to operator[](Index).
+      *
+      * This method is allowed only for vector expressions, and for matrix expressions having the LinearAccessBit.
+      *
+      * \sa operator[](Index) const, operator()(Index,Index), x(), y(), z(), w()
+      */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    operator()(Index index)
+    {
+      eigen_assert(index >= 0 && index < size());
+      return coeffRef(index);
+    }
+
+    /** equivalent to operator[](0).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    x() { return (*this)[0]; }
+
+    /** equivalent to operator[](1).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    y()
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=2, OUT_OF_RANGE_ACCESS);
+      return (*this)[1];
+    }
+
+    /** equivalent to operator[](2).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    z()
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=3, OUT_OF_RANGE_ACCESS);
+      return (*this)[2];
+    }
+
+    /** equivalent to operator[](3).  */
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar&
+    w()
+    {
+      EIGEN_STATIC_ASSERT(Derived::SizeAtCompileTime==-1 || Derived::SizeAtCompileTime>=4, OUT_OF_RANGE_ACCESS);
+      return (*this)[3];
+    }
+};
+
+/** \brief Base class providing direct read-only coefficient access to matrices and arrays.
+  * \ingroup Core_Module
+  * \tparam Derived Type of the derived class
+  * \tparam #DirectAccessors Constant indicating direct access
+  *
+  * This class defines functions to work with strides which can be used to access entries directly. This class
+  * inherits DenseCoeffsBase<Derived, ReadOnlyAccessors> which defines functions to access entries read-only using
+  * \c operator() .
+  *
+  * \sa \blank \ref TopicClassHierarchy
+  */
+template<typename Derived>
+class DenseCoeffsBase<Derived, DirectAccessors> : public DenseCoeffsBase<Derived, ReadOnlyAccessors>
+{
+  public:
+
+    typedef DenseCoeffsBase<Derived, ReadOnlyAccessors> Base;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::derived;
+
+    /** \returns the pointer increment between two consecutive elements within a slice in the inner direction.
+      *
+      * \sa outerStride(), rowStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const
+    {
+      return derived().innerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive inner slices (for example, between two consecutive columns
+      *          in a column-major matrix).
+      *
+      * \sa innerStride(), rowStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const
+    {
+      return derived().outerStride();
+    }
+
+    // FIXME shall we remove it ?
+    inline Index stride() const
+    {
+      return Derived::IsVectorAtCompileTime ? innerStride() : outerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive rows.
+      *
+      * \sa innerStride(), outerStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index rowStride() const
+    {
+      return Derived::IsRowMajor ? outerStride() : innerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive columns.
+      *
+      * \sa innerStride(), outerStride(), rowStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index colStride() const
+    {
+      return Derived::IsRowMajor ? innerStride() : outerStride();
+    }
+};
+
+/** \brief Base class providing direct read/write coefficient access to matrices and arrays.
+  * \ingroup Core_Module
+  * \tparam Derived Type of the derived class
+  * \tparam #DirectWriteAccessors Constant indicating direct access
+  *
+  * This class defines functions to work with strides which can be used to access entries directly. This class
+  * inherits DenseCoeffsBase<Derived, WriteAccessors> which defines functions to access entries read/write using
+  * \c operator().
+  *
+  * \sa \blank \ref TopicClassHierarchy
+  */
+template<typename Derived>
+class DenseCoeffsBase<Derived, DirectWriteAccessors>
+  : public DenseCoeffsBase<Derived, WriteAccessors>
+{
+  public:
+
+    typedef DenseCoeffsBase<Derived, WriteAccessors> Base;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::derived;
+
+    /** \returns the pointer increment between two consecutive elements within a slice in the inner direction.
+      *
+      * \sa outerStride(), rowStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const
+    {
+      return derived().innerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive inner slices (for example, between two consecutive columns
+      *          in a column-major matrix).
+      *
+      * \sa innerStride(), rowStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const
+    {
+      return derived().outerStride();
+    }
+
+    // FIXME shall we remove it ?
+    inline Index stride() const
+    {
+      return Derived::IsVectorAtCompileTime ? innerStride() : outerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive rows.
+      *
+      * \sa innerStride(), outerStride(), colStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index rowStride() const
+    {
+      return Derived::IsRowMajor ? outerStride() : innerStride();
+    }
+
+    /** \returns the pointer increment between two consecutive columns.
+      *
+      * \sa innerStride(), outerStride(), rowStride()
+      */
+    EIGEN_DEVICE_FUNC
+    inline Index colStride() const
+    {
+      return Derived::IsRowMajor ? innerStride() : outerStride();
+    }
+};
+
+namespace internal {
+
+template<int Alignment, typename Derived, bool JustReturnZero>
+struct first_aligned_impl
+{
+  static inline Index run(const Derived&)
+  { return 0; }
+};
+
+template<int Alignment, typename Derived>
+struct first_aligned_impl<Alignment, Derived, false>
+{
+  static inline Index run(const Derived& m)
+  {
+    return internal::first_aligned<Alignment>(m.data(), m.size());
+  }
+};
+
+/** \internal \returns the index of the first element of the array stored by \a m that is properly aligned with respect to \a Alignment for vectorization.
+  *
+  * \tparam Alignment requested alignment in Bytes.
+  *
+  * There is also the variant first_aligned(const Scalar*, Integer) defined in Memory.h. See it for more
+  * documentation.
+  */
+template<int Alignment, typename Derived>
+static inline Index first_aligned(const DenseBase<Derived>& m)
+{
+  enum { ReturnZero = (int(evaluator<Derived>::Alignment) >= Alignment) || !(Derived::Flags & DirectAccessBit) };
+  return first_aligned_impl<Alignment, Derived, ReturnZero>::run(m.derived());
+}
+
+template<typename Derived>
+static inline Index first_default_aligned(const DenseBase<Derived>& m)
+{
+  typedef typename Derived::Scalar Scalar;
+  typedef typename packet_traits<Scalar>::type DefaultPacketType;
+  return internal::first_aligned<int(unpacket_traits<DefaultPacketType>::alignment),Derived>(m);
+}
+
+template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>
+struct inner_stride_at_compile_time
+{
+  enum { ret = traits<Derived>::InnerStrideAtCompileTime };
+};
+
+template<typename Derived>
+struct inner_stride_at_compile_time<Derived, false>
+{
+  enum { ret = 0 };
+};
+
+template<typename Derived, bool HasDirectAccess = has_direct_access<Derived>::ret>
+struct outer_stride_at_compile_time
+{
+  enum { ret = traits<Derived>::OuterStrideAtCompileTime };
+};
+
+template<typename Derived>
+struct outer_stride_at_compile_time<Derived, false>
+{
+  enum { ret = 0 };
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_DENSECOEFFSBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h
new file mode 100644
index 0000000..7958fee
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DenseStorage.h
@@ -0,0 +1,570 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2010-2013 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIXSTORAGE_H
+#define EIGEN_MATRIXSTORAGE_H
+
+#ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X) X; EIGEN_DENSE_STORAGE_CTOR_PLUGIN;
+#else
+  #define EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(X)
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+struct constructor_without_unaligned_array_assert {};
+
+template<typename T, int Size>
+EIGEN_DEVICE_FUNC
+void check_static_allocation_size()
+{
+  // if EIGEN_STACK_ALLOCATION_LIMIT is defined to 0, then no limit
+  #if EIGEN_STACK_ALLOCATION_LIMIT
+  EIGEN_STATIC_ASSERT(Size * sizeof(T) <= EIGEN_STACK_ALLOCATION_LIMIT, OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG);
+  #endif
+}
+
+/** \internal
+  * Static array. If the MatrixOrArrayOptions require auto-alignment, the array will be automatically aligned:
+  * to 16 bytes boundary if the total size is a multiple of 16 bytes.
+  */
+template <typename T, int Size, int MatrixOrArrayOptions,
+          int Alignment = (MatrixOrArrayOptions&DontAlign) ? 0
+                        : compute_default_alignment<T,Size>::value >
+struct plain_array
+{
+  T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array()
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert)
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+#if defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
+  #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask)
+#elif EIGEN_GNUC_AT_LEAST(4,7) 
+  // GCC 4.7 is too aggressive in its optimizations and remove the alignement test based on the fact the array is declared to be aligned.
+  // See this bug report: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=53900
+  // Hiding the origin of the array pointer behind a function argument seems to do the trick even if the function is inlined:
+  template<typename PtrType>
+  EIGEN_ALWAYS_INLINE PtrType eigen_unaligned_array_assert_workaround_gcc47(PtrType array) { return array; }
+  #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
+    eigen_assert((internal::UIntPtr(eigen_unaligned_array_assert_workaround_gcc47(array)) & (sizemask)) == 0 \
+              && "this assertion is explained here: " \
+              "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
+              " **** READ THIS WEB PAGE !!! ****");
+#else
+  #define EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(sizemask) \
+    eigen_assert((internal::UIntPtr(array) & (sizemask)) == 0 \
+              && "this assertion is explained here: " \
+              "http://eigen.tuxfamily.org/dox-devel/group__TopicUnalignedArrayAssert.html" \
+              " **** READ THIS WEB PAGE !!! ****");
+#endif
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 8>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(8) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(7);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 16>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(16) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(15);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 32>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(32) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  {
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(31);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int Size, int MatrixOrArrayOptions>
+struct plain_array<T, Size, MatrixOrArrayOptions, 64>
+{
+  EIGEN_ALIGN_TO_BOUNDARY(64) T array[Size];
+
+  EIGEN_DEVICE_FUNC
+  plain_array() 
+  { 
+    EIGEN_MAKE_UNALIGNED_ARRAY_ASSERT(63);
+    check_static_allocation_size<T,Size>();
+  }
+
+  EIGEN_DEVICE_FUNC
+  plain_array(constructor_without_unaligned_array_assert) 
+  { 
+    check_static_allocation_size<T,Size>();
+  }
+};
+
+template <typename T, int MatrixOrArrayOptions, int Alignment>
+struct plain_array<T, 0, MatrixOrArrayOptions, Alignment>
+{
+  T array[1];
+  EIGEN_DEVICE_FUNC plain_array() {}
+  EIGEN_DEVICE_FUNC plain_array(constructor_without_unaligned_array_assert) {}
+};
+
+} // end namespace internal
+
+/** \internal
+  *
+  * \class DenseStorage
+  * \ingroup Core_Module
+  *
+  * \brief Stores the data of a matrix
+  *
+  * This class stores the data of fixed-size, dynamic-size or mixed matrices
+  * in a way as compact as possible.
+  *
+  * \sa Matrix
+  */
+template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseStorage;
+
+// purely fixed-size matrix
+template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseStorage
+{
+    internal::plain_array<T,Size,_Options> m_data;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
+    EIGEN_DEVICE_FUNC
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
+      : m_data(internal::constructor_without_unaligned_array_assert()) {}
+    EIGEN_DEVICE_FUNC 
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = Size)
+    }
+    EIGEN_DEVICE_FUNC 
+    DenseStorage& operator=(const DenseStorage& other)
+    { 
+      if (this != &other) m_data = other.m_data;
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols==_Cols);
+      EIGEN_UNUSED_VARIABLE(size);
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
+};
+
+// null matrix
+template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0, _Rows, _Cols, _Options>
+{
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage&) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage&) { return *this; }
+    EIGEN_DEVICE_FUNC DenseStorage(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& ) {}
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC void resize(Index,Index,Index) {}
+    EIGEN_DEVICE_FUNC const T *data() const { return 0; }
+    EIGEN_DEVICE_FUNC T *data() { return 0; }
+};
+
+// more specializations for null matrices; these are necessary to resolve ambiguities
+template<typename T, int _Options> class DenseStorage<T, 0, Dynamic, Dynamic, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
+template<typename T, int _Rows, int _Options> class DenseStorage<T, 0, _Rows, Dynamic, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
+template<typename T, int _Cols, int _Options> class DenseStorage<T, 0, Dynamic, _Cols, _Options>
+: public DenseStorage<T, 0, 0, 0, _Options> { };
+
+// dynamic-size matrix with fixed-size storage
+template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic, Dynamic, _Options>
+{
+    internal::plain_array<T,Size,_Options> m_data;
+    Index m_rows;
+    Index m_cols;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
+      : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
+    { 
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_rows = other.m_rows;
+        m_cols = other.m_cols;
+      }
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index cols) : m_rows(rows), m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
+    { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows() const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols() const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index cols) { m_rows = rows; m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
+};
+
+// dynamic-size matrix with fixed-size storage and fixed width
+template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Size, Dynamic, _Cols, _Options>
+{
+    internal::plain_array<T,Size,_Options> m_data;
+    Index m_rows;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_rows(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
+      : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other) 
+    {
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_rows = other.m_rows;
+      }
+      return *this; 
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index rows, Index) : m_rows(rows) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return _Cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC void resize(Index, Index rows, Index) { m_rows = rows; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
+};
+
+// dynamic-size matrix with fixed-size storage and fixed height
+template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Size, _Rows, Dynamic, _Options>
+{
+    internal::plain_array<T,Size,_Options> m_data;
+    Index m_cols;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
+      : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_cols = other.m_cols;
+      }
+      return *this;
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(Index, Index, Index cols) : m_cols(cols) {}
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index, Index, Index cols) { m_cols = cols; }
+    void resize(Index, Index, Index cols) { m_cols = cols; }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data.array; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data.array; }
+};
+
+// purely dynamic matrix.
+template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynamic, _Options>
+{
+    T *m_data;
+    Index m_rows;
+    Index m_cols;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC explicit DenseStorage(internal::constructor_without_unaligned_array_assert)
+       : m_data(0), m_rows(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols >=0);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*other.m_cols))
+      , m_rows(other.m_rows)
+      , m_cols(other.m_cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*m_cols)
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*other.m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
+      : m_data(std::move(other.m_data))
+      , m_rows(std::move(other.m_rows))
+      , m_cols(std::move(other.m_cols))
+    {
+      other.m_data = nullptr;
+      other.m_rows = 0;
+      other.m_cols = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other)
+    { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    void conservativeResize(Index size, Index rows, Index cols)
+    {
+      m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols);
+      m_rows = rows;
+      m_cols = cols;
+    }
+    EIGEN_DEVICE_FUNC void resize(Index size, Index rows, Index cols)
+    {
+      if(size != m_rows*m_cols)
+      {
+        internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols);
+        if (size)
+          m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
+        else
+          m_data = 0;
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      }
+      m_rows = rows;
+      m_cols = cols;
+    }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
+};
+
+// matrix with dynamic width and fixed height (so that matrix has dynamic size).
+template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Rows, Dynamic, _Options>
+{
+    T *m_data;
+    Index m_cols;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_cols(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows==_Rows && cols >=0);
+      EIGEN_UNUSED_VARIABLE(rows);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(_Rows*other.m_cols))
+      , m_cols(other.m_cols)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_cols*_Rows)
+      internal::smart_copy(other.m_data, other.m_data+_Rows*m_cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
+      : m_data(std::move(other.m_data))
+      , m_cols(std::move(other.m_cols))
+    {
+      other.m_data = nullptr;
+      other.m_cols = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    EIGEN_DEVICE_FUNC static Index rows(void) {return _Rows;}
+    EIGEN_DEVICE_FUNC Index cols(void) const {return m_cols;}
+    EIGEN_DEVICE_FUNC void conservativeResize(Index size, Index, Index cols)
+    {
+      m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols);
+      m_cols = cols;
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index, Index cols)
+    {
+      if(size != _Rows*m_cols)
+      {
+        internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols);
+        if (size)
+          m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
+        else
+          m_data = 0;
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      }
+      m_cols = cols;
+    }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
+};
+
+// matrix with dynamic height and fixed width (so that matrix has dynamic size).
+template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dynamic, _Cols, _Options>
+{
+    T *m_data;
+    Index m_rows;
+  public:
+    EIGEN_DEVICE_FUNC DenseStorage() : m_data(0), m_rows(0) {}
+    explicit DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
+    EIGEN_DEVICE_FUNC DenseStorage(Index size, Index rows, Index cols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      eigen_internal_assert(size==rows*cols && rows>=0 && cols == _Cols);
+      EIGEN_UNUSED_VARIABLE(cols);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage(const DenseStorage& other)
+      : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(other.m_rows*_Cols))
+      , m_rows(other.m_rows)
+    {
+      EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN(Index size = m_rows*_Cols)
+      internal::smart_copy(other.m_data, other.m_data+other.m_rows*_Cols, m_data);
+    }
+    EIGEN_DEVICE_FUNC DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        DenseStorage tmp(other);
+        this->swap(tmp);
+      }
+      return *this;
+    }    
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    DenseStorage(DenseStorage&& other) EIGEN_NOEXCEPT
+      : m_data(std::move(other.m_data))
+      , m_rows(std::move(other.m_rows))
+    {
+      other.m_data = nullptr;
+      other.m_rows = 0;
+    }
+    EIGEN_DEVICE_FUNC
+    DenseStorage& operator=(DenseStorage&& other) EIGEN_NOEXCEPT
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      return *this;
+    }
+#endif
+    EIGEN_DEVICE_FUNC ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+    EIGEN_DEVICE_FUNC void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    EIGEN_DEVICE_FUNC Index rows(void) const {return m_rows;}
+    EIGEN_DEVICE_FUNC static Index cols(void) {return _Cols;}
+    void conservativeResize(Index size, Index rows, Index)
+    {
+      m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);
+      m_rows = rows;
+    }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void resize(Index size, Index rows, Index)
+    {
+      if(size != m_rows*_Cols)
+      {
+        internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows);
+        if (size)
+          m_data = internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size);
+        else
+          m_data = 0;
+        EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN({})
+      }
+      m_rows = rows;
+    }
+    EIGEN_DEVICE_FUNC const T *data() const { return m_data; }
+    EIGEN_DEVICE_FUNC T *data() { return m_data; }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h
new file mode 100644
index 0000000..49e7112
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Diagonal.h
@@ -0,0 +1,257 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DIAGONAL_H
+#define EIGEN_DIAGONAL_H
+
+namespace Eigen { 
+
+/** \class Diagonal
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a diagonal/subdiagonal/superdiagonal in a matrix
+  *
+  * \param MatrixType the type of the object in which we are taking a sub/main/super diagonal
+  * \param DiagIndex the index of the sub/super diagonal. The default is 0 and it means the main diagonal.
+  *              A positive value means a superdiagonal, a negative value means a subdiagonal.
+  *              You can also use DynamicIndex so the index can be set at runtime.
+  *
+  * The matrix is not required to be square.
+  *
+  * This class represents an expression of the main diagonal, or any sub/super diagonal
+  * of a square matrix. It is the return type of MatrixBase::diagonal() and MatrixBase::diagonal(Index) and most of the
+  * time this is the only way it is used.
+  *
+  * \sa MatrixBase::diagonal(), MatrixBase::diagonal(Index)
+  */
+
+namespace internal {
+template<typename MatrixType, int DiagIndex>
+struct traits<Diagonal<MatrixType,DiagIndex> >
+ : traits<MatrixType>
+{
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+  typedef typename MatrixType::StorageKind StorageKind;
+  enum {
+    RowsAtCompileTime = (int(DiagIndex) == DynamicIndex || int(MatrixType::SizeAtCompileTime) == Dynamic) ? Dynamic
+                      : (EIGEN_PLAIN_ENUM_MIN(MatrixType::RowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0),
+                                              MatrixType::ColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))),
+    ColsAtCompileTime = 1,
+    MaxRowsAtCompileTime = int(MatrixType::MaxSizeAtCompileTime) == Dynamic ? Dynamic
+                         : DiagIndex == DynamicIndex ? EIGEN_SIZE_MIN_PREFER_FIXED(MatrixType::MaxRowsAtCompileTime,
+                                                                              MatrixType::MaxColsAtCompileTime)
+                         : (EIGEN_PLAIN_ENUM_MIN(MatrixType::MaxRowsAtCompileTime - EIGEN_PLAIN_ENUM_MAX(-DiagIndex, 0),
+                                                 MatrixType::MaxColsAtCompileTime - EIGEN_PLAIN_ENUM_MAX( DiagIndex, 0))),
+    MaxColsAtCompileTime = 1,
+    MaskLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = (unsigned int)_MatrixTypeNested::Flags & (RowMajorBit | MaskLvalueBit | DirectAccessBit) & ~RowMajorBit, // FIXME DirectAccessBit should not be handled by expressions
+    MatrixTypeOuterStride = outer_stride_at_compile_time<MatrixType>::ret,
+    InnerStrideAtCompileTime = MatrixTypeOuterStride == Dynamic ? Dynamic : MatrixTypeOuterStride+1,
+    OuterStrideAtCompileTime = 0
+  };
+};
+}
+
+template<typename MatrixType, int _DiagIndex> class Diagonal
+   : public internal::dense_xpr_base< Diagonal<MatrixType,_DiagIndex> >::type
+{
+  public:
+
+    enum { DiagIndex = _DiagIndex };
+    typedef typename internal::dense_xpr_base<Diagonal>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Diagonal)
+
+    EIGEN_DEVICE_FUNC
+    explicit inline Diagonal(MatrixType& matrix, Index a_index = DiagIndex) : m_matrix(matrix), m_index(a_index) {}
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Diagonal)
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const
+    {
+      return m_index.value()<0 ? numext::mini<Index>(m_matrix.cols(),m_matrix.rows()+m_index.value())
+                               : numext::mini<Index>(m_matrix.rows(),m_matrix.cols()-m_index.value());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return 1; }
+
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const
+    {
+      return m_matrix.outerStride() + 1;
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const
+    {
+      return 0;
+    }
+
+    typedef typename internal::conditional<
+                       internal::is_lvalue<MatrixType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return &(m_matrix.coeffRef(rowOffset(), colOffset())); }
+    EIGEN_DEVICE_FUNC
+    inline const Scalar* data() const { return &(m_matrix.coeffRef(rowOffset(), colOffset())); }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index row, Index)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      return m_matrix.coeffRef(row+rowOffset(), row+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index row, Index) const
+    {
+      return m_matrix.coeffRef(row+rowOffset(), row+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline CoeffReturnType coeff(Index row, Index) const
+    {
+      return m_matrix.coeff(row+rowOffset(), row+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index idx)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index idx) const
+    {
+      return m_matrix.coeffRef(idx+rowOffset(), idx+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline CoeffReturnType coeff(Index idx) const
+    {
+      return m_matrix.coeff(idx+rowOffset(), idx+colOffset());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const typename internal::remove_all<typename MatrixType::Nested>::type& 
+    nestedExpression() const 
+    {
+      return m_matrix;
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Index index() const
+    {
+      return m_index.value();
+    }
+
+  protected:
+    typename internal::ref_selector<MatrixType>::non_const_type m_matrix;
+    const internal::variable_if_dynamicindex<Index, DiagIndex> m_index;
+
+  private:
+    // some compilers may fail to optimize std::max etc in case of compile-time constants...
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index absDiagIndex() const { return m_index.value()>0 ? m_index.value() : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rowOffset() const { return m_index.value()>0 ? 0 : -m_index.value(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index colOffset() const { return m_index.value()>0 ? m_index.value() : 0; }
+    // trigger a compile-time error if someone try to call packet
+    template<int LoadMode> typename MatrixType::PacketReturnType packet(Index) const;
+    template<int LoadMode> typename MatrixType::PacketReturnType packet(Index,Index) const;
+};
+
+/** \returns an expression of the main diagonal of the matrix \c *this
+  *
+  * \c *this is not required to be square.
+  *
+  * Example: \include MatrixBase_diagonal.cpp
+  * Output: \verbinclude MatrixBase_diagonal.out
+  *
+  * \sa class Diagonal */
+template<typename Derived>
+inline typename MatrixBase<Derived>::DiagonalReturnType
+MatrixBase<Derived>::diagonal()
+{
+  return DiagonalReturnType(derived());
+}
+
+/** This is the const version of diagonal(). */
+template<typename Derived>
+inline typename MatrixBase<Derived>::ConstDiagonalReturnType
+MatrixBase<Derived>::diagonal() const
+{
+  return ConstDiagonalReturnType(derived());
+}
+
+/** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this
+  *
+  * \c *this is not required to be square.
+  *
+  * The template parameter \a DiagIndex represent a super diagonal if \a DiagIndex > 0
+  * and a sub diagonal otherwise. \a DiagIndex == 0 is equivalent to the main diagonal.
+  *
+  * Example: \include MatrixBase_diagonal_int.cpp
+  * Output: \verbinclude MatrixBase_diagonal_int.out
+  *
+  * \sa MatrixBase::diagonal(), class Diagonal */
+template<typename Derived>
+inline typename MatrixBase<Derived>::DiagonalDynamicIndexReturnType
+MatrixBase<Derived>::diagonal(Index index)
+{
+  return DiagonalDynamicIndexReturnType(derived(), index);
+}
+
+/** This is the const version of diagonal(Index). */
+template<typename Derived>
+inline typename MatrixBase<Derived>::ConstDiagonalDynamicIndexReturnType
+MatrixBase<Derived>::diagonal(Index index) const
+{
+  return ConstDiagonalDynamicIndexReturnType(derived(), index);
+}
+
+/** \returns an expression of the \a DiagIndex-th sub or super diagonal of the matrix \c *this
+  *
+  * \c *this is not required to be square.
+  *
+  * The template parameter \a DiagIndex represent a super diagonal if \a DiagIndex > 0
+  * and a sub diagonal otherwise. \a DiagIndex == 0 is equivalent to the main diagonal.
+  *
+  * Example: \include MatrixBase_diagonal_template_int.cpp
+  * Output: \verbinclude MatrixBase_diagonal_template_int.out
+  *
+  * \sa MatrixBase::diagonal(), class Diagonal */
+template<typename Derived>
+template<int Index_>
+inline typename MatrixBase<Derived>::template DiagonalIndexReturnType<Index_>::Type
+MatrixBase<Derived>::diagonal()
+{
+  return typename DiagonalIndexReturnType<Index_>::Type(derived());
+}
+
+/** This is the const version of diagonal<int>(). */
+template<typename Derived>
+template<int Index_>
+inline typename MatrixBase<Derived>::template ConstDiagonalIndexReturnType<Index_>::Type
+MatrixBase<Derived>::diagonal() const
+{
+  return typename ConstDiagonalIndexReturnType<Index_>::Type(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_DIAGONAL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h
new file mode 100644
index 0000000..ecfdce8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalMatrix.h
@@ -0,0 +1,343 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DIAGONALMATRIX_H
+#define EIGEN_DIAGONALMATRIX_H
+
+namespace Eigen { 
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename Derived>
+class DiagonalBase : public EigenBase<Derived>
+{
+  public:
+    typedef typename internal::traits<Derived>::DiagonalVectorType DiagonalVectorType;
+    typedef typename DiagonalVectorType::Scalar Scalar;
+    typedef typename DiagonalVectorType::RealScalar RealScalar;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+
+    enum {
+      RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
+      ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
+      MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+      MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+      IsVectorAtCompileTime = 0,
+      Flags = NoPreferredStorageOrderBit
+    };
+
+    typedef Matrix<Scalar, RowsAtCompileTime, ColsAtCompileTime, 0, MaxRowsAtCompileTime, MaxColsAtCompileTime> DenseMatrixType;
+    typedef DenseMatrixType DenseType;
+    typedef DiagonalMatrix<Scalar,DiagonalVectorType::SizeAtCompileTime,DiagonalVectorType::MaxSizeAtCompileTime> PlainObject;
+
+    EIGEN_DEVICE_FUNC
+    inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
+    inline Derived& derived() { return *static_cast<Derived*>(this); }
+
+    EIGEN_DEVICE_FUNC
+    DenseMatrixType toDenseMatrix() const { return derived(); }
+    
+    EIGEN_DEVICE_FUNC
+    inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
+    EIGEN_DEVICE_FUNC
+    inline DiagonalVectorType& diagonal() { return derived().diagonal(); }
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return diagonal().size(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return diagonal().size(); }
+
+    template<typename MatrixDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,MatrixDerived,LazyProduct>
+    operator*(const MatrixBase<MatrixDerived> &matrix) const
+    {
+      return Product<Derived, MatrixDerived, LazyProduct>(derived(),matrix.derived());
+    }
+
+    typedef DiagonalWrapper<const CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const DiagonalVectorType> > InverseReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const InverseReturnType
+    inverse() const
+    {
+      return InverseReturnType(diagonal().cwiseInverse());
+    }
+    
+    EIGEN_DEVICE_FUNC
+    inline const DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) >
+    operator*(const Scalar& scalar) const
+    {
+      return DiagonalWrapper<const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(DiagonalVectorType,Scalar,product) >(diagonal() * scalar);
+    }
+    EIGEN_DEVICE_FUNC
+    friend inline const DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) >
+    operator*(const Scalar& scalar, const DiagonalBase& other)
+    {
+      return DiagonalWrapper<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,DiagonalVectorType,product) >(scalar * other.diagonal());
+    }
+};
+
+#endif
+
+/** \class DiagonalMatrix
+  * \ingroup Core_Module
+  *
+  * \brief Represents a diagonal matrix with its storage
+  *
+  * \param _Scalar the type of coefficients
+  * \param SizeAtCompileTime the dimension of the matrix, or Dynamic
+  * \param MaxSizeAtCompileTime the dimension of the matrix, or Dynamic. This parameter is optional and defaults
+  *        to SizeAtCompileTime. Most of the time, you do not need to specify it.
+  *
+  * \sa class DiagonalWrapper
+  */
+
+namespace internal {
+template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime>
+struct traits<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> >
+ : traits<Matrix<_Scalar,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+{
+  typedef Matrix<_Scalar,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1> DiagonalVectorType;
+  typedef DiagonalShape StorageKind;
+  enum {
+    Flags = LvalueBit | NoPreferredStorageOrderBit
+  };
+};
+}
+template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime>
+class DiagonalMatrix
+  : public DiagonalBase<DiagonalMatrix<_Scalar,SizeAtCompileTime,MaxSizeAtCompileTime> >
+{
+  public:
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename internal::traits<DiagonalMatrix>::DiagonalVectorType DiagonalVectorType;
+    typedef const DiagonalMatrix& Nested;
+    typedef _Scalar Scalar;
+    typedef typename internal::traits<DiagonalMatrix>::StorageKind StorageKind;
+    typedef typename internal::traits<DiagonalMatrix>::StorageIndex StorageIndex;
+    #endif
+
+  protected:
+
+    DiagonalVectorType m_diagonal;
+
+  public:
+
+    /** const version of diagonal(). */
+    EIGEN_DEVICE_FUNC
+    inline const DiagonalVectorType& diagonal() const { return m_diagonal; }
+    /** \returns a reference to the stored vector of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
+    inline DiagonalVectorType& diagonal() { return m_diagonal; }
+
+    /** Default constructor without initialization */
+    EIGEN_DEVICE_FUNC
+    inline DiagonalMatrix() {}
+
+    /** Constructs a diagonal matrix with given dimension  */
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalMatrix(Index dim) : m_diagonal(dim) {}
+
+    /** 2D constructor. */
+    EIGEN_DEVICE_FUNC
+    inline DiagonalMatrix(const Scalar& x, const Scalar& y) : m_diagonal(x,y) {}
+
+    /** 3D constructor. */
+    EIGEN_DEVICE_FUNC
+    inline DiagonalMatrix(const Scalar& x, const Scalar& y, const Scalar& z) : m_diagonal(x,y,z) {}
+
+    /** Copy constructor. */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    inline DiagonalMatrix(const DiagonalBase<OtherDerived>& other) : m_diagonal(other.diagonal()) {}
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** copy constructor. prevent a default copy constructor from hiding the other templated constructor */
+    inline DiagonalMatrix(const DiagonalMatrix& other) : m_diagonal(other.diagonal()) {}
+    #endif
+
+    /** generic constructor from expression of the diagonal coefficients */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalMatrix(const MatrixBase<OtherDerived>& other) : m_diagonal(other)
+    {}
+
+    /** Copy operator. */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    DiagonalMatrix& operator=(const DiagonalBase<OtherDerived>& other)
+    {
+      m_diagonal = other.diagonal();
+      return *this;
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    DiagonalMatrix& operator=(const DiagonalMatrix& other)
+    {
+      m_diagonal = other.diagonal();
+      return *this;
+    }
+    #endif
+
+    /** Resizes to given size. */
+    EIGEN_DEVICE_FUNC
+    inline void resize(Index size) { m_diagonal.resize(size); }
+    /** Sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
+    inline void setZero() { m_diagonal.setZero(); }
+    /** Resizes and sets all coefficients to zero. */
+    EIGEN_DEVICE_FUNC
+    inline void setZero(Index size) { m_diagonal.setZero(size); }
+    /** Sets this matrix to be the identity matrix of the current size. */
+    EIGEN_DEVICE_FUNC
+    inline void setIdentity() { m_diagonal.setOnes(); }
+    /** Sets this matrix to be the identity matrix of the given size. */
+    EIGEN_DEVICE_FUNC
+    inline void setIdentity(Index size) { m_diagonal.setOnes(size); }
+};
+
+/** \class DiagonalWrapper
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a diagonal matrix
+  *
+  * \param _DiagonalVectorType the type of the vector of diagonal coefficients
+  *
+  * This class is an expression of a diagonal matrix, but not storing its own vector of diagonal coefficients,
+  * instead wrapping an existing vector expression. It is the return type of MatrixBase::asDiagonal()
+  * and most of the time this is the only way that it is used.
+  *
+  * \sa class DiagonalMatrix, class DiagonalBase, MatrixBase::asDiagonal()
+  */
+
+namespace internal {
+template<typename _DiagonalVectorType>
+struct traits<DiagonalWrapper<_DiagonalVectorType> >
+{
+  typedef _DiagonalVectorType DiagonalVectorType;
+  typedef typename DiagonalVectorType::Scalar Scalar;
+  typedef typename DiagonalVectorType::StorageIndex StorageIndex;
+  typedef DiagonalShape StorageKind;
+  typedef typename traits<DiagonalVectorType>::XprKind XprKind;
+  enum {
+    RowsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
+    ColsAtCompileTime = DiagonalVectorType::SizeAtCompileTime,
+    MaxRowsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = DiagonalVectorType::MaxSizeAtCompileTime,
+    Flags =  (traits<DiagonalVectorType>::Flags & LvalueBit) | NoPreferredStorageOrderBit
+  };
+};
+}
+
+template<typename _DiagonalVectorType>
+class DiagonalWrapper
+  : public DiagonalBase<DiagonalWrapper<_DiagonalVectorType> >, internal::no_assignment_operator
+{
+  public:
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef _DiagonalVectorType DiagonalVectorType;
+    typedef DiagonalWrapper Nested;
+    #endif
+
+    /** Constructor from expression of diagonal coefficients to wrap. */
+    EIGEN_DEVICE_FUNC
+    explicit inline DiagonalWrapper(DiagonalVectorType& a_diagonal) : m_diagonal(a_diagonal) {}
+
+    /** \returns a const reference to the wrapped expression of diagonal coefficients. */
+    EIGEN_DEVICE_FUNC
+    const DiagonalVectorType& diagonal() const { return m_diagonal; }
+
+  protected:
+    typename DiagonalVectorType::Nested m_diagonal;
+};
+
+/** \returns a pseudo-expression of a diagonal matrix with *this as vector of diagonal coefficients
+  *
+  * \only_for_vectors
+  *
+  * Example: \include MatrixBase_asDiagonal.cpp
+  * Output: \verbinclude MatrixBase_asDiagonal.out
+  *
+  * \sa class DiagonalWrapper, class DiagonalMatrix, diagonal(), isDiagonal()
+  **/
+template<typename Derived>
+inline const DiagonalWrapper<const Derived>
+MatrixBase<Derived>::asDiagonal() const
+{
+  return DiagonalWrapper<const Derived>(derived());
+}
+
+/** \returns true if *this is approximately equal to a diagonal matrix,
+  *          within the precision given by \a prec.
+  *
+  * Example: \include MatrixBase_isDiagonal.cpp
+  * Output: \verbinclude MatrixBase_isDiagonal.out
+  *
+  * \sa asDiagonal()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isDiagonal(const RealScalar& prec) const
+{
+  if(cols() != rows()) return false;
+  RealScalar maxAbsOnDiagonal = static_cast<RealScalar>(-1);
+  for(Index j = 0; j < cols(); ++j)
+  {
+    RealScalar absOnDiagonal = numext::abs(coeff(j,j));
+    if(absOnDiagonal > maxAbsOnDiagonal) maxAbsOnDiagonal = absOnDiagonal;
+  }
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = 0; i < j; ++i)
+    {
+      if(!internal::isMuchSmallerThan(coeff(i, j), maxAbsOnDiagonal, prec)) return false;
+      if(!internal::isMuchSmallerThan(coeff(j, i), maxAbsOnDiagonal, prec)) return false;
+    }
+  return true;
+}
+
+namespace internal {
+
+template<> struct storage_kind_to_shape<DiagonalShape> { typedef DiagonalShape Shape; };
+
+struct Diagonal2Dense {};
+
+template<> struct AssignmentKind<DenseShape,DiagonalShape> { typedef Diagonal2Dense Kind; };
+
+// Diagonal matrix to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Dense>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    dst.setZero();
+    dst.diagonal() = src.diagonal();
+  }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() += src.diagonal(); }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() -= src.diagonal(); }
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_DIAGONALMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h
new file mode 100644
index 0000000..d372b93
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/DiagonalProduct.h
@@ -0,0 +1,28 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DIAGONALPRODUCT_H
+#define EIGEN_DIAGONALPRODUCT_H
+
+namespace Eigen { 
+
+/** \returns the diagonal matrix product of \c *this by the diagonal matrix \a diagonal.
+  */
+template<typename Derived>
+template<typename DiagonalDerived>
+inline const Product<Derived, DiagonalDerived, LazyProduct>
+MatrixBase<Derived>::operator*(const DiagonalBase<DiagonalDerived> &a_diagonal) const
+{
+  return Product<Derived, DiagonalDerived, LazyProduct>(derived(),a_diagonal.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_DIAGONALPRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Dot.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Dot.h
new file mode 100644
index 0000000..06ef18b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Dot.h
@@ -0,0 +1,315 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008, 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DOT_H
+#define EIGEN_DOT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// helper function for dot(). The problem is that if we put that in the body of dot(), then upon calling dot
+// with mismatched types, the compiler emits errors about failing to instantiate cwiseProduct BEFORE
+// looking at the static assertions. Thus this is a trick to get better compile errors.
+template<typename T, typename U,
+// the NeedToTranspose condition here is taken straight from Assign.h
+         bool NeedToTranspose = T::IsVectorAtCompileTime
+                && U::IsVectorAtCompileTime
+                && ((int(T::RowsAtCompileTime) == 1 && int(U::ColsAtCompileTime) == 1)
+                      |  // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
+                         // revert to || as soon as not needed anymore.
+                    (int(T::ColsAtCompileTime) == 1 && int(U::RowsAtCompileTime) == 1))
+>
+struct dot_nocheck
+{
+  typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
+  typedef typename conj_prod::result_type ResScalar;
+  EIGEN_DEVICE_FUNC
+  static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
+  {
+    return a.template binaryExpr<conj_prod>(b).sum();
+  }
+};
+
+template<typename T, typename U>
+struct dot_nocheck<T, U, true>
+{
+  typedef scalar_conj_product_op<typename traits<T>::Scalar,typename traits<U>::Scalar> conj_prod;
+  typedef typename conj_prod::result_type ResScalar;
+  EIGEN_DEVICE_FUNC
+  static inline ResScalar run(const MatrixBase<T>& a, const MatrixBase<U>& b)
+  {
+    return a.transpose().template binaryExpr<conj_prod>(b).sum();
+  }
+};
+
+} // end namespace internal
+
+/** \fn MatrixBase::dot
+  * \returns the dot product of *this with other.
+  *
+  * \only_for_vectors
+  *
+  * \note If the scalar type is complex numbers, then this function returns the hermitian
+  * (sesquilinear) dot product, conjugate-linear in the first variable and linear in the
+  * second variable.
+  *
+  * \sa squaredNorm(), norm()
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
+MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+  EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
+#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG))
+  typedef internal::scalar_conj_product_op<Scalar,typename OtherDerived::Scalar> func;
+  EIGEN_CHECK_BINARY_COMPATIBILIY(func,Scalar,typename OtherDerived::Scalar);
+#endif
+  
+  eigen_assert(size() == other.size());
+
+  return internal::dot_nocheck<Derived,OtherDerived>::run(*this, other);
+}
+
+//---------- implementation of L2 norm and related functions ----------
+
+/** \returns, for vectors, the squared \em l2 norm of \c *this, and for matrices the Frobenius norm.
+  * In both cases, it consists in the sum of the square of all the matrix entries.
+  * For vectors, this is also equals to the dot product of \c *this with itself.
+  *
+  * \sa dot(), norm(), lpNorm()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::squaredNorm() const
+{
+  return numext::real((*this).cwiseAbs2().sum());
+}
+
+/** \returns, for vectors, the \em l2 norm of \c *this, and for matrices the Frobenius norm.
+  * In both cases, it consists in the square root of the sum of the square of all the matrix entries.
+  * For vectors, this is also equals to the square root of the dot product of \c *this with itself.
+  *
+  * \sa lpNorm(), dot(), squaredNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real MatrixBase<Derived>::norm() const
+{
+  return numext::sqrt(squaredNorm());
+}
+
+/** \returns an expression of the quotient of \c *this by its own norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0),
+  *          then this function returns a copy of the input.
+  *
+  * \only_for_vectors
+  *
+  * \sa norm(), normalize()
+  */
+template<typename Derived>
+inline const typename MatrixBase<Derived>::PlainObject
+MatrixBase<Derived>::normalized() const
+{
+  typedef typename internal::nested_eval<Derived,2>::type _Nested;
+  _Nested n(derived());
+  RealScalar z = n.squaredNorm();
+  // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
+  if(z>RealScalar(0))
+    return n / numext::sqrt(z);
+  else
+    return n;
+}
+
+/** Normalizes the vector, i.e. divides it by its own norm.
+  *
+  * \only_for_vectors
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged.
+  *
+  * \sa norm(), normalized()
+  */
+template<typename Derived>
+inline void MatrixBase<Derived>::normalize()
+{
+  RealScalar z = squaredNorm();
+  // NOTE: after extensive benchmarking, this conditional does not impact performance, at least on recent x86 CPU
+  if(z>RealScalar(0))
+    derived() /= numext::sqrt(z);
+}
+
+/** \returns an expression of the quotient of \c *this by its own norm while avoiding underflow and overflow.
+  *
+  * \only_for_vectors
+  *
+  * This method is analogue to the normalized() method, but it reduces the risk of
+  * underflow and overflow when computing the norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0),
+  *          then this function returns a copy of the input.
+  *
+  * \sa stableNorm(), stableNormalize(), normalized()
+  */
+template<typename Derived>
+inline const typename MatrixBase<Derived>::PlainObject
+MatrixBase<Derived>::stableNormalized() const
+{
+  typedef typename internal::nested_eval<Derived,3>::type _Nested;
+  _Nested n(derived());
+  RealScalar w = n.cwiseAbs().maxCoeff();
+  RealScalar z = (n/w).squaredNorm();
+  if(z>RealScalar(0))
+    return n / (numext::sqrt(z)*w);
+  else
+    return n;
+}
+
+/** Normalizes the vector while avoid underflow and overflow
+  *
+  * \only_for_vectors
+  *
+  * This method is analogue to the normalize() method, but it reduces the risk of
+  * underflow and overflow when computing the norm.
+  *
+  * \warning If the input vector is too small (i.e., this->norm()==0), then \c *this is left unchanged.
+  *
+  * \sa stableNorm(), stableNormalized(), normalize()
+  */
+template<typename Derived>
+inline void MatrixBase<Derived>::stableNormalize()
+{
+  RealScalar w = cwiseAbs().maxCoeff();
+  RealScalar z = (derived()/w).squaredNorm();
+  if(z>RealScalar(0))
+    derived() /= numext::sqrt(z)*w;
+}
+
+//---------- implementation of other norms ----------
+
+namespace internal {
+
+template<typename Derived, int p>
+struct lpNorm_selector
+{
+  typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const MatrixBase<Derived>& m)
+  {
+    EIGEN_USING_STD_MATH(pow)
+    return pow(m.cwiseAbs().array().pow(p).sum(), RealScalar(1)/p);
+  }
+};
+
+template<typename Derived>
+struct lpNorm_selector<Derived, 1>
+{
+  EIGEN_DEVICE_FUNC
+  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  {
+    return m.cwiseAbs().sum();
+  }
+};
+
+template<typename Derived>
+struct lpNorm_selector<Derived, 2>
+{
+  EIGEN_DEVICE_FUNC
+  static inline typename NumTraits<typename traits<Derived>::Scalar>::Real run(const MatrixBase<Derived>& m)
+  {
+    return m.norm();
+  }
+};
+
+template<typename Derived>
+struct lpNorm_selector<Derived, Infinity>
+{
+  typedef typename NumTraits<typename traits<Derived>::Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const MatrixBase<Derived>& m)
+  {
+    if(Derived::SizeAtCompileTime==0 || (Derived::SizeAtCompileTime==Dynamic && m.size()==0))
+      return RealScalar(0);
+    return m.cwiseAbs().maxCoeff();
+  }
+};
+
+} // end namespace internal
+
+/** \returns the \b coefficient-wise \f$ \ell^p \f$ norm of \c *this, that is, returns the p-th root of the sum of the p-th powers of the absolute values
+  *          of the coefficients of \c *this. If \a p is the special value \a Eigen::Infinity, this function returns the \f$ \ell^\infty \f$
+  *          norm, that is the maximum of the absolute values of the coefficients of \c *this.
+  *
+  * In all cases, if \c *this is empty, then the value 0 is returned.
+  *
+  * \note For matrices, this function does not compute the <a href="https://en.wikipedia.org/wiki/Operator_norm">operator-norm</a>. That is, if \c *this is a matrix, then its coefficients are interpreted as a 1D vector. Nonetheless, you can easily compute the 1-norm and \f$\infty\f$-norm matrix operator norms using \link TutorialReductionsVisitorsBroadcastingReductionsNorm partial reductions \endlink.
+  *
+  * \sa norm()
+  */
+template<typename Derived>
+template<int p>
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+#else
+MatrixBase<Derived>::RealScalar
+#endif
+MatrixBase<Derived>::lpNorm() const
+{
+  return internal::lpNorm_selector<Derived, p>::run(*this);
+}
+
+//---------- implementation of isOrthogonal / isUnitary ----------
+
+/** \returns true if *this is approximately orthogonal to \a other,
+  *          within the precision given by \a prec.
+  *
+  * Example: \include MatrixBase_isOrthogonal.cpp
+  * Output: \verbinclude MatrixBase_isOrthogonal.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+bool MatrixBase<Derived>::isOrthogonal
+(const MatrixBase<OtherDerived>& other, const RealScalar& prec) const
+{
+  typename internal::nested_eval<Derived,2>::type nested(derived());
+  typename internal::nested_eval<OtherDerived,2>::type otherNested(other.derived());
+  return numext::abs2(nested.dot(otherNested)) <= prec * prec * nested.squaredNorm() * otherNested.squaredNorm();
+}
+
+/** \returns true if *this is approximately an unitary matrix,
+  *          within the precision given by \a prec. In the case where the \a Scalar
+  *          type is real numbers, a unitary matrix is an orthogonal matrix, whence the name.
+  *
+  * \note This can be used to check whether a family of vectors forms an orthonormal basis.
+  *       Indeed, \c m.isUnitary() returns true if and only if the columns (equivalently, the rows) of m form an
+  *       orthonormal basis.
+  *
+  * Example: \include MatrixBase_isUnitary.cpp
+  * Output: \verbinclude MatrixBase_isUnitary.out
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isUnitary(const RealScalar& prec) const
+{
+  typename internal::nested_eval<Derived,1>::type self(derived());
+  for(Index i = 0; i < cols(); ++i)
+  {
+    if(!internal::isApprox(self.col(i).squaredNorm(), static_cast<RealScalar>(1), prec))
+      return false;
+    for(Index j = 0; j < i; ++j)
+      if(!internal::isMuchSmallerThan(self.col(i).dot(self.col(j)), static_cast<Scalar>(1), prec))
+        return false;
+  }
+  return true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_DOT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h
new file mode 100644
index 0000000..b195506
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/EigenBase.h
@@ -0,0 +1,159 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EIGENBASE_H
+#define EIGEN_EIGENBASE_H
+
+namespace Eigen {
+
+/** \class EigenBase
+  * \ingroup Core_Module
+  * 
+  * Common base class for all classes T such that MatrixBase has an operator=(T) and a constructor MatrixBase(T).
+  *
+  * In other words, an EigenBase object is an object that can be copied into a MatrixBase.
+  *
+  * Besides MatrixBase-derived classes, this also includes special matrix classes such as diagonal matrices, etc.
+  *
+  * Notice that this class is trivial, it is only used to disambiguate overloaded functions.
+  *
+  * \sa \blank \ref TopicClassHierarchy
+  */
+template<typename Derived> struct EigenBase
+{
+//   typedef typename internal::plain_matrix_type<Derived>::type PlainObject;
+  
+  /** \brief The interface type of indices
+    * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+    * \deprecated Since Eigen 3.3, its usage is deprecated. Use Eigen::Index instead.
+    * \sa StorageIndex, \ref TopicPreprocessorDirectives.
+    */
+  typedef Eigen::Index Index;
+
+  // FIXME is it needed?
+  typedef typename internal::traits<Derived>::StorageKind StorageKind;
+
+  /** \returns a reference to the derived object */
+  EIGEN_DEVICE_FUNC
+  Derived& derived() { return *static_cast<Derived*>(this); }
+  /** \returns a const reference to the derived object */
+  EIGEN_DEVICE_FUNC
+  const Derived& derived() const { return *static_cast<const Derived*>(this); }
+
+  EIGEN_DEVICE_FUNC
+  inline Derived& const_cast_derived() const
+  { return *static_cast<Derived*>(const_cast<EigenBase*>(this)); }
+  EIGEN_DEVICE_FUNC
+  inline const Derived& const_derived() const
+  { return *static_cast<const Derived*>(this); }
+
+  /** \returns the number of rows. \sa cols(), RowsAtCompileTime */
+  EIGEN_DEVICE_FUNC
+  inline Index rows() const { return derived().rows(); }
+  /** \returns the number of columns. \sa rows(), ColsAtCompileTime*/
+  EIGEN_DEVICE_FUNC
+  inline Index cols() const { return derived().cols(); }
+  /** \returns the number of coefficients, which is rows()*cols().
+    * \sa rows(), cols(), SizeAtCompileTime. */
+  EIGEN_DEVICE_FUNC
+  inline Index size() const { return rows() * cols(); }
+
+  /** \internal Don't use it, but do the equivalent: \code dst = *this; \endcode */
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void evalTo(Dest& dst) const
+  { derived().evalTo(dst); }
+
+  /** \internal Don't use it, but do the equivalent: \code dst += *this; \endcode */
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void addTo(Dest& dst) const
+  {
+    // This is the default implementation,
+    // derived class can reimplement it in a more optimized way.
+    typename Dest::PlainObject res(rows(),cols());
+    evalTo(res);
+    dst += res;
+  }
+
+  /** \internal Don't use it, but do the equivalent: \code dst -= *this; \endcode */
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC
+  inline void subTo(Dest& dst) const
+  {
+    // This is the default implementation,
+    // derived class can reimplement it in a more optimized way.
+    typename Dest::PlainObject res(rows(),cols());
+    evalTo(res);
+    dst -= res;
+  }
+
+  /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheRight(*this); \endcode */
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheRight(Dest& dst) const
+  {
+    // This is the default implementation,
+    // derived class can reimplement it in a more optimized way.
+    dst = dst * this->derived();
+  }
+
+  /** \internal Don't use it, but do the equivalent: \code dst.applyOnTheLeft(*this); \endcode */
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC inline void applyThisOnTheLeft(Dest& dst) const
+  {
+    // This is the default implementation,
+    // derived class can reimplement it in a more optimized way.
+    dst = this->derived() * dst;
+  }
+
+};
+
+/***************************************************************************
+* Implementation of matrix base methods
+***************************************************************************/
+
+/** \brief Copies the generic expression \a other into *this.
+  *
+  * \details The expression must provide a (templated) evalTo(Derived& dst) const
+  * function which does the actual job. In practice, this allows any user to write
+  * its own special matrix without having to modify MatrixBase
+  *
+  * \returns a reference to *this.
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+Derived& DenseBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+Derived& DenseBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+Derived& DenseBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_EIGENBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h
new file mode 100644
index 0000000..7b08b45
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ForceAlignedAccess.h
@@ -0,0 +1,146 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FORCEALIGNEDACCESS_H
+#define EIGEN_FORCEALIGNEDACCESS_H
+
+namespace Eigen {
+
+/** \class ForceAlignedAccess
+  * \ingroup Core_Module
+  *
+  * \brief Enforce aligned packet loads and stores regardless of what is requested
+  *
+  * \param ExpressionType the type of the object of which we are forcing aligned packet access
+  *
+  * This class is the return type of MatrixBase::forceAlignedAccess()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::forceAlignedAccess()
+  */
+
+namespace internal {
+template<typename ExpressionType>
+struct traits<ForceAlignedAccess<ExpressionType> > : public traits<ExpressionType>
+{};
+}
+
+template<typename ExpressionType> class ForceAlignedAccess
+  : public internal::dense_xpr_base< ForceAlignedAccess<ExpressionType> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<ForceAlignedAccess>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(ForceAlignedAccess)
+
+    EIGEN_DEVICE_FUNC explicit inline ForceAlignedAccess(const ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
+
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
+    {
+      return m_expression.coeff(row, col);
+    }
+
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_expression.const_cast_derived().coeffRef(row, col);
+    }
+
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
+    {
+      return m_expression.coeff(index);
+    }
+
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
+    {
+      return m_expression.const_cast_derived().coeffRef(index);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index row, Index col) const
+    {
+      return m_expression.template packet<Aligned>(row, col);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    {
+      m_expression.const_cast_derived().template writePacket<Aligned>(row, col, x);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index index) const
+    {
+      return m_expression.template packet<Aligned>(index);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index index, const PacketScalar& x)
+    {
+      m_expression.const_cast_derived().template writePacket<Aligned>(index, x);
+    }
+
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
+
+  protected:
+    const ExpressionType& m_expression;
+
+  private:
+    ForceAlignedAccess& operator=(const ForceAlignedAccess&);
+};
+
+/** \returns an expression of *this with forced aligned access
+  * \sa forceAlignedAccessIf(),class ForceAlignedAccess
+  */
+template<typename Derived>
+inline const ForceAlignedAccess<Derived>
+MatrixBase<Derived>::forceAlignedAccess() const
+{
+  return ForceAlignedAccess<Derived>(derived());
+}
+
+/** \returns an expression of *this with forced aligned access
+  * \sa forceAlignedAccessIf(), class ForceAlignedAccess
+  */
+template<typename Derived>
+inline ForceAlignedAccess<Derived>
+MatrixBase<Derived>::forceAlignedAccess()
+{
+  return ForceAlignedAccess<Derived>(derived());
+}
+
+/** \returns an expression of *this with forced aligned access if \a Enable is true.
+  * \sa forceAlignedAccess(), class ForceAlignedAccess
+  */
+template<typename Derived>
+template<bool Enable>
+inline typename internal::add_const_on_value_type<typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type>::type
+MatrixBase<Derived>::forceAlignedAccessIf() const
+{
+  return derived();  // FIXME This should not work but apparently is never used
+}
+
+/** \returns an expression of *this with forced aligned access if \a Enable is true.
+  * \sa forceAlignedAccess(), class ForceAlignedAccess
+  */
+template<typename Derived>
+template<bool Enable>
+inline typename internal::conditional<Enable,ForceAlignedAccess<Derived>,Derived&>::type
+MatrixBase<Derived>::forceAlignedAccessIf()
+{
+  return derived();  // FIXME This should not work but apparently is never used
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_FORCEALIGNEDACCESS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h
new file mode 100644
index 0000000..3e403a0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Fuzzy.h
@@ -0,0 +1,155 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FUZZY_H
+#define EIGEN_FUZZY_H
+
+namespace Eigen { 
+
+namespace internal
+{
+
+template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
+struct isApprox_selector
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
+  {
+    typename internal::nested_eval<Derived,2>::type nested(x);
+    typename internal::nested_eval<OtherDerived,2>::type otherNested(y);
+    return (nested - otherNested).cwiseAbs2().sum() <= prec * prec * numext::mini(nested.cwiseAbs2().sum(), otherNested.cwiseAbs2().sum());
+  }
+};
+
+template<typename Derived, typename OtherDerived>
+struct isApprox_selector<Derived, OtherDerived, true>
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar&)
+  {
+    return x.matrix() == y.matrix();
+  }
+};
+
+template<typename Derived, typename OtherDerived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
+struct isMuchSmallerThan_object_selector
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const OtherDerived& y, const typename Derived::RealScalar& prec)
+  {
+    return x.cwiseAbs2().sum() <= numext::abs2(prec) * y.cwiseAbs2().sum();
+  }
+};
+
+template<typename Derived, typename OtherDerived>
+struct isMuchSmallerThan_object_selector<Derived, OtherDerived, true>
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const OtherDerived&, const typename Derived::RealScalar&)
+  {
+    return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();
+  }
+};
+
+template<typename Derived, bool is_integer = NumTraits<typename Derived::Scalar>::IsInteger>
+struct isMuchSmallerThan_scalar_selector
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const typename Derived::RealScalar& y, const typename Derived::RealScalar& prec)
+  {
+    return x.cwiseAbs2().sum() <= numext::abs2(prec * y);
+  }
+};
+
+template<typename Derived>
+struct isMuchSmallerThan_scalar_selector<Derived, true>
+{
+  EIGEN_DEVICE_FUNC
+  static bool run(const Derived& x, const typename Derived::RealScalar&, const typename Derived::RealScalar&)
+  {
+    return x.matrix() == Derived::Zero(x.rows(), x.cols()).matrix();
+  }
+};
+
+} // end namespace internal
+
+
+/** \returns \c true if \c *this is approximately equal to \a other, within the precision
+  * determined by \a prec.
+  *
+  * \note The fuzzy compares are done multiplicatively. Two vectors \f$ v \f$ and \f$ w \f$
+  * are considered to be approximately equal within precision \f$ p \f$ if
+  * \f[ \Vert v - w \Vert \leqslant p\,\min(\Vert v\Vert, \Vert w\Vert). \f]
+  * For matrices, the comparison is done using the Hilbert-Schmidt norm (aka Frobenius norm
+  * L2 norm).
+  *
+  * \note Because of the multiplicativeness of this comparison, one can't use this function
+  * to check whether \c *this is approximately equal to the zero matrix or vector.
+  * Indeed, \c isApprox(zero) returns false unless \c *this itself is exactly the zero matrix
+  * or vector. If you want to test whether \c *this is zero, use internal::isMuchSmallerThan(const
+  * RealScalar&, RealScalar) instead.
+  *
+  * \sa internal::isMuchSmallerThan(const RealScalar&, RealScalar) const
+  */
+template<typename Derived>
+template<typename OtherDerived>
+bool DenseBase<Derived>::isApprox(
+  const DenseBase<OtherDerived>& other,
+  const RealScalar& prec
+) const
+{
+  return internal::isApprox_selector<Derived, OtherDerived>::run(derived(), other.derived(), prec);
+}
+
+/** \returns \c true if the norm of \c *this is much smaller than \a other,
+  * within the precision determined by \a prec.
+  *
+  * \note The fuzzy compares are done multiplicatively. A vector \f$ v \f$ is
+  * considered to be much smaller than \f$ x \f$ within precision \f$ p \f$ if
+  * \f[ \Vert v \Vert \leqslant p\,\vert x\vert. \f]
+  *
+  * For matrices, the comparison is done using the Hilbert-Schmidt norm. For this reason,
+  * the value of the reference scalar \a other should come from the Hilbert-Schmidt norm
+  * of a reference matrix of same dimensions.
+  *
+  * \sa isApprox(), isMuchSmallerThan(const DenseBase<OtherDerived>&, RealScalar) const
+  */
+template<typename Derived>
+bool DenseBase<Derived>::isMuchSmallerThan(
+  const typename NumTraits<Scalar>::Real& other,
+  const RealScalar& prec
+) const
+{
+  return internal::isMuchSmallerThan_scalar_selector<Derived>::run(derived(), other, prec);
+}
+
+/** \returns \c true if the norm of \c *this is much smaller than the norm of \a other,
+  * within the precision determined by \a prec.
+  *
+  * \note The fuzzy compares are done multiplicatively. A vector \f$ v \f$ is
+  * considered to be much smaller than a vector \f$ w \f$ within precision \f$ p \f$ if
+  * \f[ \Vert v \Vert \leqslant p\,\Vert w\Vert. \f]
+  * For matrices, the comparison is done using the Hilbert-Schmidt norm.
+  *
+  * \sa isApprox(), isMuchSmallerThan(const RealScalar&, RealScalar) const
+  */
+template<typename Derived>
+template<typename OtherDerived>
+bool DenseBase<Derived>::isMuchSmallerThan(
+  const DenseBase<OtherDerived>& other,
+  const RealScalar& prec
+) const
+{
+  return internal::isMuchSmallerThan_object_selector<Derived, OtherDerived>::run(derived(), other.derived(), prec);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_FUZZY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h
new file mode 100644
index 0000000..0f16cd8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GeneralProduct.h
@@ -0,0 +1,454 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_PRODUCT_H
+#define EIGEN_GENERAL_PRODUCT_H
+
+namespace Eigen {
+
+enum {
+  Large = 2,
+  Small = 3
+};
+
+namespace internal {
+
+template<int Rows, int Cols, int Depth> struct product_type_selector;
+
+template<int Size, int MaxSize> struct product_size_category
+{
+  enum { is_large = MaxSize == Dynamic ||
+                    Size >= EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD ||
+                    (Size==Dynamic && MaxSize>=EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD),
+         value = is_large  ? Large
+               : Size == 1 ? 1
+                           : Small
+  };
+};
+
+template<typename Lhs, typename Rhs> struct product_type
+{
+  typedef typename remove_all<Lhs>::type _Lhs;
+  typedef typename remove_all<Rhs>::type _Rhs;
+  enum {
+    MaxRows = traits<_Lhs>::MaxRowsAtCompileTime,
+    Rows    = traits<_Lhs>::RowsAtCompileTime,
+    MaxCols = traits<_Rhs>::MaxColsAtCompileTime,
+    Cols    = traits<_Rhs>::ColsAtCompileTime,
+    MaxDepth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::MaxColsAtCompileTime,
+                                           traits<_Rhs>::MaxRowsAtCompileTime),
+    Depth = EIGEN_SIZE_MIN_PREFER_FIXED(traits<_Lhs>::ColsAtCompileTime,
+                                        traits<_Rhs>::RowsAtCompileTime)
+  };
+
+  // the splitting into different lines of code here, introducing the _select enums and the typedef below,
+  // is to work around an internal compiler error with gcc 4.1 and 4.2.
+private:
+  enum {
+    rows_select = product_size_category<Rows,MaxRows>::value,
+    cols_select = product_size_category<Cols,MaxCols>::value,
+    depth_select = product_size_category<Depth,MaxDepth>::value
+  };
+  typedef product_type_selector<rows_select, cols_select, depth_select> selector;
+
+public:
+  enum {
+    value = selector::ret,
+    ret = selector::ret
+  };
+#ifdef EIGEN_DEBUG_PRODUCT
+  static void debug()
+  {
+      EIGEN_DEBUG_VAR(Rows);
+      EIGEN_DEBUG_VAR(Cols);
+      EIGEN_DEBUG_VAR(Depth);
+      EIGEN_DEBUG_VAR(rows_select);
+      EIGEN_DEBUG_VAR(cols_select);
+      EIGEN_DEBUG_VAR(depth_select);
+      EIGEN_DEBUG_VAR(value);
+  }
+#endif
+};
+
+/* The following allows to select the kind of product at compile time
+ * based on the three dimensions of the product.
+ * This is a compile time mapping from {1,Small,Large}^3 -> {product types} */
+// FIXME I'm not sure the current mapping is the ideal one.
+template<int M, int N>  struct product_type_selector<M,N,1>              { enum { ret = OuterProduct }; };
+template<int M>         struct product_type_selector<M, 1, 1>            { enum { ret = LazyCoeffBasedProductMode }; };
+template<int N>         struct product_type_selector<1, N, 1>            { enum { ret = LazyCoeffBasedProductMode }; };
+template<int Depth>     struct product_type_selector<1,    1,    Depth>  { enum { ret = InnerProduct }; };
+template<>              struct product_type_selector<1,    1,    1>      { enum { ret = InnerProduct }; };
+template<>              struct product_type_selector<Small,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small, Large, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large, Small, 1>    { enum { ret = LazyCoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Large,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<1,    Large,Large>  { enum { ret = GemvProduct }; };
+template<>              struct product_type_selector<1,    Small,Large>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large,1,    Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large,1,    Large>  { enum { ret = GemvProduct }; };
+template<>              struct product_type_selector<Small,1,    Large>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Small,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Small,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Small,Large,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Large,Large>  { enum { ret = GemmProduct }; };
+template<>              struct product_type_selector<Large,Small,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Small,Large,Small>  { enum { ret = CoeffBasedProductMode }; };
+template<>              struct product_type_selector<Large,Large,Small>  { enum { ret = GemmProduct }; };
+
+} // end namespace internal
+
+/***********************************************************************
+*  Implementation of Inner Vector Vector Product
+***********************************************************************/
+
+// FIXME : maybe the "inner product" could return a Scalar
+// instead of a 1x1 matrix ??
+// Pro: more natural for the user
+// Cons: this could be a problem if in a meta unrolled algorithm a matrix-matrix
+// product ends up to a row-vector times col-vector product... To tackle this use
+// case, we could have a specialization for Block<MatrixType,1,1> with: operator=(Scalar x);
+
+/***********************************************************************
+*  Implementation of Outer Vector Vector Product
+***********************************************************************/
+
+/***********************************************************************
+*  Implementation of General Matrix Vector Product
+***********************************************************************/
+
+/*  According to the shape/flags of the matrix we have to distinghish 3 different cases:
+ *   1 - the matrix is col-major, BLAS compatible and M is large => call fast BLAS-like colmajor routine
+ *   2 - the matrix is row-major, BLAS compatible and N is large => call fast BLAS-like rowmajor routine
+ *   3 - all other cases are handled using a simple loop along the outer-storage direction.
+ *  Therefore we need a lower level meta selector.
+ *  Furthermore, if the matrix is the rhs, then the product has to be transposed.
+ */
+namespace internal {
+
+template<int Side, int StorageOrder, bool BlasCompatible>
+struct gemv_dense_selector;
+
+} // end namespace internal
+
+namespace internal {
+
+template<typename Scalar,int Size,int MaxSize,bool Cond> struct gemv_static_vector_if;
+
+template<typename Scalar,int Size,int MaxSize>
+struct gemv_static_vector_if<Scalar,Size,MaxSize,false>
+{
+  EIGEN_STRONG_INLINE  Scalar* data() { eigen_internal_assert(false && "should never be called"); return 0; }
+};
+
+template<typename Scalar,int Size>
+struct gemv_static_vector_if<Scalar,Size,Dynamic,true>
+{
+  EIGEN_STRONG_INLINE Scalar* data() { return 0; }
+};
+
+template<typename Scalar,int Size,int MaxSize>
+struct gemv_static_vector_if<Scalar,Size,MaxSize,true>
+{
+  enum {
+    ForceAlignment  = internal::packet_traits<Scalar>::Vectorizable,
+    PacketSize      = internal::packet_traits<Scalar>::size
+  };
+  #if EIGEN_MAX_STATIC_ALIGN_BYTES!=0
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize),0,EIGEN_PLAIN_ENUM_MIN(AlignedMax,PacketSize)> m_data;
+  EIGEN_STRONG_INLINE Scalar* data() { return m_data.array; }
+  #else
+  // Some architectures cannot align on the stack,
+  // => let's manually enforce alignment by allocating more data and return the address of the first aligned element.
+  internal::plain_array<Scalar,EIGEN_SIZE_MIN_PREFER_FIXED(Size,MaxSize)+(ForceAlignment?EIGEN_MAX_ALIGN_BYTES:0),0> m_data;
+  EIGEN_STRONG_INLINE Scalar* data() {
+    return ForceAlignment
+            ? reinterpret_cast<Scalar*>((internal::UIntPtr(m_data.array) & ~(std::size_t(EIGEN_MAX_ALIGN_BYTES-1))) + EIGEN_MAX_ALIGN_BYTES)
+            : m_data.array;
+  }
+  #endif
+};
+
+// The vector is on the left => transposition
+template<int StorageOrder, bool BlasCompatible>
+struct gemv_dense_selector<OnTheLeft,StorageOrder,BlasCompatible>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    Transpose<Dest> destT(dest);
+    enum { OtherStorageOrder = StorageOrder == RowMajor ? ColMajor : RowMajor };
+    gemv_dense_selector<OnTheRight,OtherStorageOrder,BlasCompatible>
+      ::run(rhs.transpose(), lhs.transpose(), destT, alpha);
+  }
+};
+
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,true>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static inline void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    typedef typename Dest::RealScalar  RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+
+    ActualLhsType actualLhs = LhsBlasTraits::extract(lhs);
+    ActualRhsType actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
+
+    // make sure Dest is a compile-time vector type (bug 1166)
+    typedef typename conditional<Dest::IsVectorAtCompileTime, Dest, typename Dest::ColXpr>::type ActualDest;
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1),
+      ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
+      MightCannotUseDest = (!EvalToDestAtCompileTime) || ComplexByReal
+    };
+
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
+
+    if(!MightCannotUseDest)
+    {
+      // shortcut if we are sure to be able to use dest directly,
+      // this ease the compiler to generate cleaner and more optimzized code for most common cases
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          dest.data(), 1,
+          compatibleAlpha);
+    }
+    else
+    {
+      gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+      const bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+      const bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
+      ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                    evalToDest ? dest.data() : static_dest.data());
+
+      if(!evalToDest)
+      {
+        #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        Index size = dest.size();
+        EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+        #endif
+        if(!alphaIsCompatible)
+        {
+          MappedDest(actualDestPtr, dest.size()).setZero();
+          compatibleAlpha = RhsScalar(1);
+        }
+        else
+          MappedDest(actualDestPtr, dest.size()) = dest;
+      }
+
+      general_matrix_vector_product
+          <Index,LhsScalar,LhsMapper,ColMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+          actualLhs.rows(), actualLhs.cols(),
+          LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+          RhsMapper(actualRhs.data(), actualRhs.innerStride()),
+          actualDestPtr, 1,
+          compatibleAlpha);
+
+      if (!evalToDest)
+      {
+        if(!alphaIsCompatible)
+          dest.matrix() += actualAlpha * MappedDest(actualDestPtr, dest.size());
+        else
+          dest = MappedDest(actualDestPtr, dest.size());
+      }
+    }
+  }
+};
+
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,true>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar   LhsScalar;
+    typedef typename Rhs::Scalar   RhsScalar;
+    typedef typename Dest::Scalar  ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
+    };
+
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
+        DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
+
+    if(!DirectlyUseRhs)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = actualRhs.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    }
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
+    general_matrix_vector_product
+        <Index,LhsScalar,LhsMapper,RowMajor,LhsBlasTraits::NeedToConjugate,RhsScalar,RhsMapper,RhsBlasTraits::NeedToConjugate>::run(
+        actualLhs.rows(), actualLhs.cols(),
+        LhsMapper(actualLhs.data(), actualLhs.outerStride()),
+        RhsMapper(actualRhsPtr, 1),
+        dest.data(), dest.col(0).innerStride(), //NOTE  if dest is not a vector at compile-time, then dest.innerStride() might be wrong. (bug 1166)
+        actualAlpha);
+  }
+};
+
+template<> struct gemv_dense_selector<OnTheRight,ColMajor,false>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
+    // TODO if rhs is large enough it might be beneficial to make sure that dest is sequentially stored in memory, otherwise use a temp
+    typename nested_eval<Rhs,1>::type actual_rhs(rhs);
+    const Index size = rhs.rows();
+    for(Index k=0; k<size; ++k)
+      dest += (alpha*actual_rhs.coeff(k)) * lhs.col(k);
+  }
+};
+
+template<> struct gemv_dense_selector<OnTheRight,RowMajor,false>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    EIGEN_STATIC_ASSERT((!nested_eval<Lhs,1>::Evaluate),EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE);
+    typename nested_eval<Rhs,Lhs::RowsAtCompileTime>::type actual_rhs(rhs);
+    const Index rows = dest.rows();
+    for(Index i=0; i<rows; ++i)
+      dest.coeffRef(i) += alpha * (lhs.row(i).cwiseProduct(actual_rhs.transpose())).sum();
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Implementation of matrix base methods
+***************************************************************************/
+
+/** \returns the matrix product of \c *this and \a other.
+  *
+  * \note If instead of the matrix product you want the coefficient-wise product, see Cwise::operator*().
+  *
+  * \sa lazyProduct(), operator*=(const MatrixBase&), Cwise::operator*()
+  */
+#ifndef __CUDACC__
+
+template<typename Derived>
+template<typename OtherDerived>
+inline const Product<Derived, OtherDerived>
+MatrixBase<Derived>::operator*(const MatrixBase<OtherDerived> &other) const
+{
+  // A note regarding the function declaration: In MSVC, this function will sometimes
+  // not be inlined since DenseStorage is an unwindable object for dynamic
+  // matrices and product types are holding a member to store the result.
+  // Thus it does not help tagging this function with EIGEN_STRONG_INLINE.
+  enum {
+    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
+                   || OtherDerived::RowsAtCompileTime==Dynamic
+                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
+    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
+    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
+  };
+  // note to the lost user:
+  //    * for a dot product use: v1.dot(v2)
+  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
+    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
+    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
+  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
+#ifdef EIGEN_DEBUG_PRODUCT
+  internal::product_type<Derived,OtherDerived>::debug();
+#endif
+
+  return Product<Derived, OtherDerived>(derived(), other.derived());
+}
+
+#endif // __CUDACC__
+
+/** \returns an expression of the matrix product of \c *this and \a other without implicit evaluation.
+  *
+  * The returned product will behave like any other expressions: the coefficients of the product will be
+  * computed once at a time as requested. This might be useful in some extremely rare cases when only
+  * a small and no coherent fraction of the result's coefficients have to be computed.
+  *
+  * \warning This version of the matrix product can be much much slower. So use it only if you know
+  * what you are doing and that you measured a true speed improvement.
+  *
+  * \sa operator*(const MatrixBase&)
+  */
+template<typename Derived>
+template<typename OtherDerived>
+const Product<Derived,OtherDerived,LazyProduct>
+MatrixBase<Derived>::lazyProduct(const MatrixBase<OtherDerived> &other) const
+{
+  enum {
+    ProductIsValid =  Derived::ColsAtCompileTime==Dynamic
+                   || OtherDerived::RowsAtCompileTime==Dynamic
+                   || int(Derived::ColsAtCompileTime)==int(OtherDerived::RowsAtCompileTime),
+    AreVectors = Derived::IsVectorAtCompileTime && OtherDerived::IsVectorAtCompileTime,
+    SameSizes = EIGEN_PREDICATE_SAME_MATRIX_SIZE(Derived,OtherDerived)
+  };
+  // note to the lost user:
+  //    * for a dot product use: v1.dot(v2)
+  //    * for a coeff-wise product use: v1.cwiseProduct(v2)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(AreVectors && SameSizes),
+    INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS)
+  EIGEN_STATIC_ASSERT(ProductIsValid || !(SameSizes && !AreVectors),
+    INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION)
+  EIGEN_STATIC_ASSERT(ProductIsValid || SameSizes, INVALID_MATRIX_PRODUCT)
+
+  return Product<Derived,OtherDerived,LazyProduct>(derived(), other.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_PRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h
new file mode 100644
index 0000000..029f8ac
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GenericPacketMath.h
@@ -0,0 +1,593 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERIC_PACKET_MATH_H
+#define EIGEN_GENERIC_PACKET_MATH_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal
+  * \file GenericPacketMath.h
+  *
+  * Default implementation for types not supported by the vectorization.
+  * In practice these functions are provided to make easier the writing
+  * of generic vectorized code.
+  */
+
+#ifndef EIGEN_DEBUG_ALIGNED_LOAD
+#define EIGEN_DEBUG_ALIGNED_LOAD
+#endif
+
+#ifndef EIGEN_DEBUG_UNALIGNED_LOAD
+#define EIGEN_DEBUG_UNALIGNED_LOAD
+#endif
+
+#ifndef EIGEN_DEBUG_ALIGNED_STORE
+#define EIGEN_DEBUG_ALIGNED_STORE
+#endif
+
+#ifndef EIGEN_DEBUG_UNALIGNED_STORE
+#define EIGEN_DEBUG_UNALIGNED_STORE
+#endif
+
+struct default_packet_traits
+{
+  enum {
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasNegate = 1,
+    HasAbs    = 1,
+    HasArg    = 0,
+    HasAbs2   = 1,
+    HasMin    = 1,
+    HasMax    = 1,
+    HasConj   = 1,
+    HasSetLinear = 1,
+    HasBlend  = 0,
+
+    HasDiv    = 0,
+    HasSqrt   = 0,
+    HasRsqrt  = 0,
+    HasExp    = 0,
+    HasLog    = 0,
+    HasLog1p  = 0,
+    HasLog10  = 0,
+    HasPow    = 0,
+
+    HasSin    = 0,
+    HasCos    = 0,
+    HasTan    = 0,
+    HasASin   = 0,
+    HasACos   = 0,
+    HasATan   = 0,
+    HasSinh   = 0,
+    HasCosh   = 0,
+    HasTanh   = 0,
+    HasLGamma = 0,
+    HasDiGamma = 0,
+    HasZeta = 0,
+    HasPolygamma = 0,
+    HasErf = 0,
+    HasErfc = 0,
+    HasIGamma = 0,
+    HasIGammac = 0,
+    HasBetaInc = 0,
+
+    HasRound  = 0,
+    HasFloor  = 0,
+    HasCeil   = 0,
+
+    HasSign   = 0
+  };
+};
+
+template<typename T> struct packet_traits : default_packet_traits
+{
+  typedef T type;
+  typedef T half;
+  enum {
+    Vectorizable = 0,
+    size = 1,
+    AlignedOnScalar = 0,
+    HasHalfPacket = 0
+  };
+  enum {
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<typename T> struct packet_traits<const T> : packet_traits<T> { };
+
+template <typename Src, typename Tgt> struct type_casting_traits {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+/** \internal \returns static_cast<TgtType>(a) (coeff-wise) */
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a) {
+  return static_cast<TgtPacket>(a);
+}
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/) {
+  return static_cast<TgtPacket>(a);
+}
+
+template <typename SrcPacket, typename TgtPacket>
+EIGEN_DEVICE_FUNC inline TgtPacket
+pcast(const SrcPacket& a, const SrcPacket& /*b*/, const SrcPacket& /*c*/, const SrcPacket& /*d*/) {
+  return static_cast<TgtPacket>(a);
+}
+
+/** \internal \returns a + b (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+padd(const Packet& a,
+        const Packet& b) { return a+b; }
+
+/** \internal \returns a - b (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+psub(const Packet& a,
+        const Packet& b) { return a-b; }
+
+/** \internal \returns -a (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pnegate(const Packet& a) { return -a; }
+
+/** \internal \returns conj(a) (coeff-wise) */
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pconj(const Packet& a) { return numext::conj(a); }
+
+/** \internal \returns a * b (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pmul(const Packet& a,
+        const Packet& b) { return a*b; }
+
+/** \internal \returns a / b (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pdiv(const Packet& a,
+        const Packet& b) { return a/b; }
+
+/** \internal \returns the min of \a a and \a b  (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pmin(const Packet& a,
+        const Packet& b) { return numext::mini(a, b); }
+
+/** \internal \returns the max of \a a and \a b  (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pmax(const Packet& a,
+        const Packet& b) { return numext::maxi(a, b); }
+
+/** \internal \returns the absolute value of \a a */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pabs(const Packet& a) { using std::abs; return abs(a); }
+
+/** \internal \returns the phase angle of \a a */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+parg(const Packet& a) { using numext::arg; return arg(a); }
+
+/** \internal \returns the bitwise and of \a a and \a b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pand(const Packet& a, const Packet& b) { return a & b; }
+
+/** \internal \returns the bitwise or of \a a and \a b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+por(const Packet& a, const Packet& b) { return a | b; }
+
+/** \internal \returns the bitwise xor of \a a and \a b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pxor(const Packet& a, const Packet& b) { return a ^ b; }
+
+/** \internal \returns the bitwise andnot of \a a and \a b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pandnot(const Packet& a, const Packet& b) { return a & (!b); }
+
+/** \internal \returns a packet version of \a *from, from must be 16 bytes aligned */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pload(const typename unpacket_traits<Packet>::type* from) { return *from; }
+
+/** \internal \returns a packet version of \a *from, (un-aligned load) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+ploadu(const typename unpacket_traits<Packet>::type* from) { return *from; }
+
+/** \internal \returns a packet with constant coefficients \a a, e.g.: (a,a,a,a) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pset1(const typename unpacket_traits<Packet>::type& a) { return a; }
+
+/** \internal \returns a packet with constant coefficients \a a[0], e.g.: (a[0],a[0],a[0],a[0]) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pload1(const typename unpacket_traits<Packet>::type  *a) { return pset1<Packet>(*a); }
+
+/** \internal \returns a packet with elements of \a *from duplicated.
+  * For instance, for a packet of 8 elements, 4 scalars will be read from \a *from and
+  * duplicated to form: {from[0],from[0],from[1],from[1],from[2],from[2],from[3],from[3]}
+  * Currently, this function is only used for scalar * complex products.
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
+ploaddup(const typename unpacket_traits<Packet>::type* from) { return *from; }
+
+/** \internal \returns a packet with elements of \a *from quadrupled.
+  * For instance, for a packet of 8 elements, 2 scalars will be read from \a *from and
+  * replicated to form: {from[0],from[0],from[0],from[0],from[1],from[1],from[1],from[1]}
+  * Currently, this function is only used in matrix products.
+  * For packet-size smaller or equal to 4, this function is equivalent to pload1 
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+ploadquad(const typename unpacket_traits<Packet>::type* from)
+{ return pload1<Packet>(from); }
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * a2 = pload1(a+2);
+  * a3 = pload1(a+3);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast2
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast4(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1, Packet& a2, Packet& a3)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+  a2 = pload1<Packet>(a+2);
+  a3 = pload1<Packet>(a+3);
+}
+
+/** \internal equivalent to
+  * \code
+  * a0 = pload1(a+0);
+  * a1 = pload1(a+1);
+  * \endcode
+  * \sa pset1, pload1, ploaddup, pbroadcast4
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC
+inline void pbroadcast2(const typename unpacket_traits<Packet>::type *a,
+                        Packet& a0, Packet& a1)
+{
+  a0 = pload1<Packet>(a+0);
+  a1 = pload1<Packet>(a+1);
+}
+
+/** \internal \brief Returns a packet with coefficients (a,a+1,...,a+packet_size-1). */
+template<typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Packet
+plset(const typename unpacket_traits<Packet>::type& a) { return a; }
+
+/** \internal copy the packet \a from to \a *to, \a to must be 16 bytes aligned */
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstore(Scalar* to, const Packet& from)
+{ (*to) = from; }
+
+/** \internal copy the packet \a from to \a *to, (un-aligned store) */
+template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pstoreu(Scalar* to, const Packet& from)
+{  (*to) = from; }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline Packet pgather(const Scalar* from, Index /*stride*/)
+ { return ploadu<Packet>(from); }
+
+ template<typename Scalar, typename Packet> EIGEN_DEVICE_FUNC inline void pscatter(Scalar* to, const Packet& from, Index /*stride*/)
+ { pstore(to, from); }
+
+/** \internal tries to do cache prefetching of \a addr */
+template<typename Scalar> EIGEN_DEVICE_FUNC inline void prefetch(const Scalar* addr)
+{
+#ifdef __CUDA_ARCH__
+#if defined(__LP64__)
+  // 64-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=l"(addr) : "l"(addr));
+#else
+  // 32-bit pointer operand constraint for inlined asm
+  asm(" prefetch.L1 [ %1 ];" : "=r"(addr) : "r"(addr));
+#endif
+#elif (!EIGEN_COMP_MSVC) && (EIGEN_COMP_GNUC || EIGEN_COMP_CLANG || EIGEN_COMP_ICC)
+  __builtin_prefetch(addr);
+#endif
+}
+
+/** \internal \returns the first element of a packet */
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type pfirst(const Packet& a)
+{ return a; }
+
+/** \internal \returns a packet where the element i contains the sum of the packet of \a vec[i] */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+preduxp(const Packet* vecs) { return vecs[0]; }
+
+/** \internal \returns the sum of the elements of \a a*/
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux(const Packet& a)
+{ return a; }
+
+/** \internal \returns the sum of the elements of \a a by block of 4 elements.
+  * For a packet {a0, a1, a2, a3, a4, a5, a6, a7}, it returns a half packet {a0+a4, a1+a5, a2+a6, a3+a7}
+  * For packet-size smaller or equal to 4, this boils down to a noop.
+  */
+template<typename Packet> EIGEN_DEVICE_FUNC inline
+typename conditional<(unpacket_traits<Packet>::size%8)==0,typename unpacket_traits<Packet>::half,Packet>::type
+predux_downto4(const Packet& a)
+{ return a; }
+
+/** \internal \returns the product of the elements of \a a*/
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_mul(const Packet& a)
+{ return a; }
+
+/** \internal \returns the min of the elements of \a a*/
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_min(const Packet& a)
+{ return a; }
+
+/** \internal \returns the max of the elements of \a a*/
+template<typename Packet> EIGEN_DEVICE_FUNC inline typename unpacket_traits<Packet>::type predux_max(const Packet& a)
+{ return a; }
+
+/** \internal \returns the reversed elements of \a a*/
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet preverse(const Packet& a)
+{ return a; }
+
+/** \internal \returns \a a with real and imaginary part flipped (for complex type only) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet pcplxflip(const Packet& a)
+{
+  // FIXME: uncomment the following in case we drop the internal imag and real functions.
+//   using std::imag;
+//   using std::real;
+  return Packet(imag(a),real(a));
+}
+
+/**************************
+* Special math functions
+***************************/
+
+/** \internal \returns the sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psin(const Packet& a) { using std::sin; return sin(a); }
+
+/** \internal \returns the cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pcos(const Packet& a) { using std::cos; return cos(a); }
+
+/** \internal \returns the tan of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ptan(const Packet& a) { using std::tan; return tan(a); }
+
+/** \internal \returns the arc sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pasin(const Packet& a) { using std::asin; return asin(a); }
+
+/** \internal \returns the arc cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pacos(const Packet& a) { using std::acos; return acos(a); }
+
+/** \internal \returns the arc tangent of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet patan(const Packet& a) { using std::atan; return atan(a); }
+
+/** \internal \returns the hyperbolic sine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psinh(const Packet& a) { using std::sinh; return sinh(a); }
+
+/** \internal \returns the hyperbolic cosine of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pcosh(const Packet& a) { using std::cosh; return cosh(a); }
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet ptanh(const Packet& a) { using std::tanh; return tanh(a); }
+
+/** \internal \returns the exp of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pexp(const Packet& a) { using std::exp; return exp(a); }
+
+/** \internal \returns the log of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog(const Packet& a) { using std::log; return log(a); }
+
+/** \internal \returns the log1p of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog1p(const Packet& a) { return numext::log1p(a); }
+
+/** \internal \returns the log10 of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet plog10(const Packet& a) { using std::log10; return log10(a); }
+
+/** \internal \returns the square-root of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet psqrt(const Packet& a) { using std::sqrt; return sqrt(a); }
+
+/** \internal \returns the reciprocal square-root of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet prsqrt(const Packet& a) {
+  return pdiv(pset1<Packet>(1), psqrt(a));
+}
+
+/** \internal \returns the rounded value of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pround(const Packet& a) { using numext::round; return round(a); }
+
+/** \internal \returns the floor of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pfloor(const Packet& a) { using numext::floor; return floor(a); }
+
+/** \internal \returns the ceil of \a a (coeff-wise) */
+template<typename Packet> EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet pceil(const Packet& a) { using numext::ceil; return ceil(a); }
+
+/***************************************************************************
+* The following functions might not have to be overwritten for vectorized types
+***************************************************************************/
+
+/** \internal copy a packet with constant coeficient \a a (e.g., [a,a,a,a]) to \a *to. \a to must be 16 bytes aligned */
+// NOTE: this function must really be templated on the packet type (think about different packet types for the same scalar type)
+template<typename Packet>
+inline void pstore1(typename unpacket_traits<Packet>::type* to, const typename unpacket_traits<Packet>::type& a)
+{
+  pstore(to, pset1<Packet>(a));
+}
+
+/** \internal \returns a * b + c (coeff-wise) */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pmadd(const Packet&  a,
+         const Packet&  b,
+         const Packet&  c)
+{ return padd(pmul(a, b),c); }
+
+/** \internal \returns a packet version of \a *from.
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt(const typename unpacket_traits<Packet>::type* from)
+{
+  if(Alignment >= unpacket_traits<Packet>::alignment)
+    return pload<Packet>(from);
+  else
+    return ploadu<Packet>(from);
+}
+
+/** \internal copy the packet \a from to \a *to.
+  * The pointer \a from must be aligned on a \a Alignment bytes boundary. */
+template<typename Scalar, typename Packet, int Alignment>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void pstoret(Scalar* to, const Packet& from)
+{
+  if(Alignment >= unpacket_traits<Packet>::alignment)
+    pstore(to, from);
+  else
+    pstoreu(to, from);
+}
+
+/** \internal \returns a packet version of \a *from.
+  * Unlike ploadt, ploadt_ro takes advantage of the read-only memory path on the
+  * hardware if available to speedup the loading of data that won't be modified
+  * by the current computation.
+  */
+template<typename Packet, int LoadMode>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet ploadt_ro(const typename unpacket_traits<Packet>::type* from)
+{
+  return ploadt<Packet, LoadMode>(from);
+}
+
+/** \internal default implementation of palign() allowing partial specialization */
+template<int Offset,typename PacketType>
+struct palign_impl
+{
+  // by default data are aligned, so there is nothing to be done :)
+  static inline void run(PacketType&, const PacketType&) {}
+};
+
+/** \internal update \a first using the concatenation of the packet_size minus \a Offset last elements
+  * of \a first and \a Offset first elements of \a second.
+  * 
+  * This function is currently only used to optimize matrix-vector products on unligned matrices.
+  * It takes 2 packets that represent a contiguous memory array, and returns a packet starting
+  * at the position \a Offset. For instance, for packets of 4 elements, we have:
+  *  Input:
+  *  - first = {f0,f1,f2,f3}
+  *  - second = {s0,s1,s2,s3}
+  * Output: 
+  *   - if Offset==0 then {f0,f1,f2,f3}
+  *   - if Offset==1 then {f1,f2,f3,s0}
+  *   - if Offset==2 then {f2,f3,s0,s1}
+  *   - if Offset==3 then {f3,s0,s1,s3}
+  */
+template<int Offset,typename PacketType>
+inline void palign(PacketType& first, const PacketType& second)
+{
+  palign_impl<Offset,PacketType>::run(first,second);
+}
+
+/***************************************************************************
+* Fast complex products (GCC generates a function call which is very slow)
+***************************************************************************/
+
+// Eigen+CUDA does not support complexes.
+#ifndef __CUDACC__
+
+template<> inline std::complex<float> pmul(const std::complex<float>& a, const std::complex<float>& b)
+{ return std::complex<float>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
+
+template<> inline std::complex<double> pmul(const std::complex<double>& a, const std::complex<double>& b)
+{ return std::complex<double>(real(a)*real(b) - imag(a)*imag(b), imag(a)*real(b) + real(a)*imag(b)); }
+
+#endif
+
+
+/***************************************************************************
+ * PacketBlock, that is a collection of N packets where the number of words
+ * in the packet is a multiple of N.
+***************************************************************************/
+template <typename Packet,int N=unpacket_traits<Packet>::size> struct PacketBlock {
+  Packet packet[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet,1>& /*kernel*/) {
+  // Nothing to do in the scalar case, i.e. a 1x1 matrix.
+}
+
+/***************************************************************************
+ * Selector, i.e. vector of N boolean values used to select (i.e. blend)
+ * words from 2 packets.
+***************************************************************************/
+template <size_t N> struct Selector {
+  bool select[N];
+};
+
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pblend(const Selector<unpacket_traits<Packet>::size>& ifPacket, const Packet& thenPacket, const Packet& elsePacket) {
+  return ifPacket.select[0] ? thenPacket : elsePacket;
+}
+
+/** \internal \returns \a a with the first coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertfirst(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  mask.select[0] = true;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=1; i<unpacket_traits<Packet>::size; ++i)
+    mask.select[i] = false;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
+/** \internal \returns \a a with the last coefficient replaced by the scalar b */
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
+pinsertlast(const Packet& a, typename unpacket_traits<Packet>::type b)
+{
+  // Default implementation based on pblend.
+  // It must be specialized for higher performance.
+  Selector<unpacket_traits<Packet>::size> mask;
+  // This for loop should be optimized away by the compiler.
+  for(Index i=0; i<unpacket_traits<Packet>::size-1; ++i)
+    mask.select[i] = false;
+  mask.select[unpacket_traits<Packet>::size-1] = true;
+  return pblend(mask, pset1<Packet>(b), a);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERIC_PACKET_MATH_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h
new file mode 100644
index 0000000..769dc25
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/GlobalFunctions.h
@@ -0,0 +1,187 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GLOBAL_FUNCTIONS_H
+#define EIGEN_GLOBAL_FUNCTIONS_H
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \
+  /** \returns an expression of the coefficient-wise DOC_OP of \a x
+
+    DOC_DETAILS
+
+    \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_##NAME">Math functions</a>, class CwiseUnaryOp
+    */ \
+  template<typename Derived> \
+  inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
+  NAME(const Eigen::ArrayBase<Derived>& x);
+
+#else
+
+#define EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(NAME,FUNCTOR,DOC_OP,DOC_DETAILS) \
+  template<typename Derived> \
+  inline const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> \
+  (NAME)(const Eigen::ArrayBase<Derived>& x) { \
+    return Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived>(x.derived()); \
+  }
+
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(NAME,FUNCTOR) \
+  \
+  template<typename Derived> \
+  struct NAME##_retval<ArrayBase<Derived> > \
+  { \
+    typedef const Eigen::CwiseUnaryOp<Eigen::internal::FUNCTOR<typename Derived::Scalar>, const Derived> type; \
+  }; \
+  template<typename Derived> \
+  struct NAME##_impl<ArrayBase<Derived> > \
+  { \
+    static inline typename NAME##_retval<ArrayBase<Derived> >::type run(const Eigen::ArrayBase<Derived>& x) \
+    { \
+      return typename NAME##_retval<ArrayBase<Derived> >::type(x.derived()); \
+    } \
+  };
+
+namespace Eigen
+{
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(real,scalar_real_op,real part,\sa ArrayBase::real)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(imag,scalar_imag_op,imaginary part,\sa ArrayBase::imag)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(conj,scalar_conjugate_op,complex conjugate,\sa ArrayBase::conjugate)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(inverse,scalar_inverse_op,inverse,\sa ArrayBase::inverse)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sin,scalar_sin_op,sine,\sa ArrayBase::sin)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cos,scalar_cos_op,cosine,\sa ArrayBase::cos)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tan,scalar_tan_op,tangent,\sa ArrayBase::tan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(atan,scalar_atan_op,arc-tangent,\sa ArrayBase::atan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(asin,scalar_asin_op,arc-sine,\sa ArrayBase::asin)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(acos,scalar_acos_op,arc-consine,\sa ArrayBase::acos)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sinh,scalar_sinh_op,hyperbolic sine,\sa ArrayBase::sinh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cosh,scalar_cosh_op,hyperbolic cosine,\sa ArrayBase::cosh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(tanh,scalar_tanh_op,hyperbolic tangent,\sa ArrayBase::tanh)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(lgamma,scalar_lgamma_op,natural logarithm of the gamma function,\sa ArrayBase::lgamma)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(digamma,scalar_digamma_op,derivative of lgamma,\sa ArrayBase::digamma)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erf,scalar_erf_op,error function,\sa ArrayBase::erf)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(erfc,scalar_erfc_op,complement error function,\sa ArrayBase::erfc)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(exp,scalar_exp_op,exponential,\sa ArrayBase::exp)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log,scalar_log_op,natural logarithm,\sa Eigen::log10 DOXCOMMA ArrayBase::log)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log1p,scalar_log1p_op,natural logarithm of 1 plus the value,\sa ArrayBase::log1p)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(log10,scalar_log10_op,base 10 logarithm,\sa Eigen::log DOXCOMMA ArrayBase::log)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs,scalar_abs_op,absolute value,\sa ArrayBase::abs DOXCOMMA MatrixBase::cwiseAbs)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(abs2,scalar_abs2_op,squared absolute value,\sa ArrayBase::abs2 DOXCOMMA MatrixBase::cwiseAbs2)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(arg,scalar_arg_op,complex argument,\sa ArrayBase::arg)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sqrt,scalar_sqrt_op,square root,\sa ArrayBase::sqrt DOXCOMMA MatrixBase::cwiseSqrt)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(rsqrt,scalar_rsqrt_op,reciprocal square root,\sa ArrayBase::rsqrt)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(square,scalar_square_op,square (power 2),\sa Eigen::abs2 DOXCOMMA Eigen::pow DOXCOMMA ArrayBase::square)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(cube,scalar_cube_op,cube (power 3),\sa Eigen::pow DOXCOMMA ArrayBase::cube)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(round,scalar_round_op,nearest integer,\sa Eigen::floor DOXCOMMA Eigen::ceil DOXCOMMA ArrayBase::round)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(floor,scalar_floor_op,nearest integer not greater than the giben value,\sa Eigen::ceil DOXCOMMA ArrayBase::floor)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(ceil,scalar_ceil_op,nearest integer not less than the giben value,\sa Eigen::floor DOXCOMMA ArrayBase::ceil)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isnan,scalar_isnan_op,not-a-number test,\sa Eigen::isinf DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isnan)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isinf,scalar_isinf_op,infinite value test,\sa Eigen::isnan DOXCOMMA Eigen::isfinite DOXCOMMA ArrayBase::isinf)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(isfinite,scalar_isfinite_op,finite value test,\sa Eigen::isinf DOXCOMMA Eigen::isnan DOXCOMMA ArrayBase::isfinite)
+  EIGEN_ARRAY_DECLARE_GLOBAL_UNARY(sign,scalar_sign_op,sign (or 0),\sa ArrayBase::sign)
+  
+  /** \returns an expression of the coefficient-wise power of \a x to the given constant \a exponent.
+    *
+    * \tparam ScalarExponent is the scalar type of \a exponent. It must be compatible with the scalar type of the given expression (\c Derived::Scalar).
+    *
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+  template<typename Derived,typename ScalarExponent>
+  inline const CwiseBinaryOp<internal::scalar_pow_op<Derived::Scalar,ScalarExponent>,Derived,Constant<ScalarExponent> >
+  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent);
+#else
+  template<typename Derived,typename ScalarExponent>
+  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,ScalarExponent>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,typename Derived::Scalar,ScalarExponent),
+          const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,ScalarExponent,pow) >::type
+  pow(const Eigen::ArrayBase<Derived>& x, const ScalarExponent& exponent) {
+    return x.derived().pow(exponent);
+  }
+
+  template<typename Derived>
+  inline const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename Derived::Scalar,pow)
+  pow(const Eigen::ArrayBase<Derived>& x, const typename Derived::Scalar& exponent) {
+    return x.derived().pow(exponent);
+  }
+#endif
+
+  /** \returns an expression of the coefficient-wise power of \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power.
+    *
+    * Example: \include Cwise_array_power_array.cpp
+    * Output: \verbinclude Cwise_array_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+  template<typename Derived,typename ExponentDerived>
+  inline const Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>
+  pow(const Eigen::ArrayBase<Derived>& x, const Eigen::ArrayBase<ExponentDerived>& exponents) 
+  {
+    return Eigen::CwiseBinaryOp<Eigen::internal::scalar_pow_op<typename Derived::Scalar, typename ExponentDerived::Scalar>, const Derived, const ExponentDerived>(
+      x.derived(),
+      exponents.derived()
+    );
+  }
+  
+  /** \returns an expression of the coefficient-wise power of the scalar \a x to the given array of \a exponents.
+    *
+    * This function computes the coefficient-wise power between a scalar and an array of exponents.
+    *
+    * \tparam Scalar is the scalar type of \a x. It must be compatible with the scalar type of the given array expression (\c Derived::Scalar).
+    *
+    * Example: \include Cwise_scalar_power_array.cpp
+    * Output: \verbinclude Cwise_scalar_power_array.out
+    * 
+    * \sa ArrayBase::pow()
+    *
+    * \relates ArrayBase
+    */
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+  template<typename Scalar,typename Derived>
+  inline const CwiseBinaryOp<internal::scalar_pow_op<Scalar,Derived::Scalar>,Constant<Scalar>,Derived>
+  pow(const Scalar& x,const Eigen::ArrayBase<Derived>& x);
+#else
+  template<typename Scalar, typename Derived>
+  inline typename internal::enable_if<   !(internal::is_same<typename Derived::Scalar,Scalar>::value) && EIGEN_SCALAR_BINARY_SUPPORTED(pow,Scalar,typename Derived::Scalar),
+          const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow) >::type
+  pow(const Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
+  {
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,Derived,pow)(
+            typename internal::plain_constant_type<Derived,Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
+  }
+
+  template<typename Derived>
+  inline const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)
+  pow(const typename Derived::Scalar& x, const Eigen::ArrayBase<Derived>& exponents)
+  {
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename Derived::Scalar,Derived,pow)(
+      typename internal::plain_constant_type<Derived,typename Derived::Scalar>::type(exponents.rows(), exponents.cols(), x), exponents.derived() );
+  }
+#endif
+
+
+  namespace internal
+  {
+    EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(real,scalar_real_op)
+    EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(imag,scalar_imag_op)
+    EIGEN_ARRAY_DECLARE_GLOBAL_EIGEN_UNARY(abs2,scalar_abs2_op)
+  }
+}
+
+// TODO: cleanly disable those functions that are not supported on Array (numext::real_ref, internal::random, internal::isApprox...)
+
+#endif // EIGEN_GLOBAL_FUNCTIONS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/IO.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/IO.h
new file mode 100644
index 0000000..da7fd6c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/IO.h
@@ -0,0 +1,225 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_IO_H
+#define EIGEN_IO_H
+
+namespace Eigen { 
+
+enum { DontAlignCols = 1 };
+enum { StreamPrecision = -1,
+       FullPrecision = -2 };
+
+namespace internal {
+template<typename Derived>
+std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& fmt);
+}
+
+/** \class IOFormat
+  * \ingroup Core_Module
+  *
+  * \brief Stores a set of parameters controlling the way matrices are printed
+  *
+  * List of available parameters:
+  *  - \b precision number of digits for floating point values, or one of the special constants \c StreamPrecision and \c FullPrecision.
+  *                 The default is the special value \c StreamPrecision which means to use the
+  *                 stream's own precision setting, as set for instance using \c cout.precision(3). The other special value
+  *                 \c FullPrecision means that the number of digits will be computed to match the full precision of each floating-point
+  *                 type.
+  *  - \b flags an OR-ed combination of flags, the default value is 0, the only currently available flag is \c DontAlignCols which
+  *             allows to disable the alignment of columns, resulting in faster code.
+  *  - \b coeffSeparator string printed between two coefficients of the same row
+  *  - \b rowSeparator string printed between two rows
+  *  - \b rowPrefix string printed at the beginning of each row
+  *  - \b rowSuffix string printed at the end of each row
+  *  - \b matPrefix string printed at the beginning of the matrix
+  *  - \b matSuffix string printed at the end of the matrix
+  *
+  * Example: \include IOFormat.cpp
+  * Output: \verbinclude IOFormat.out
+  *
+  * \sa DenseBase::format(), class WithFormat
+  */
+struct IOFormat
+{
+  /** Default constructor, see class IOFormat for the meaning of the parameters */
+  IOFormat(int _precision = StreamPrecision, int _flags = 0,
+    const std::string& _coeffSeparator = " ",
+    const std::string& _rowSeparator = "\n", const std::string& _rowPrefix="", const std::string& _rowSuffix="",
+    const std::string& _matPrefix="", const std::string& _matSuffix="")
+  : matPrefix(_matPrefix), matSuffix(_matSuffix), rowPrefix(_rowPrefix), rowSuffix(_rowSuffix), rowSeparator(_rowSeparator),
+    rowSpacer(""), coeffSeparator(_coeffSeparator), precision(_precision), flags(_flags)
+  {
+    // TODO check if rowPrefix, rowSuffix or rowSeparator contains a newline
+    // don't add rowSpacer if columns are not to be aligned
+    if((flags & DontAlignCols))
+      return;
+    int i = int(matSuffix.length())-1;
+    while (i>=0 && matSuffix[i]!='\n')
+    {
+      rowSpacer += ' ';
+      i--;
+    }
+  }
+  std::string matPrefix, matSuffix;
+  std::string rowPrefix, rowSuffix, rowSeparator, rowSpacer;
+  std::string coeffSeparator;
+  int precision;
+  int flags;
+};
+
+/** \class WithFormat
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression providing matrix output with given format
+  *
+  * \tparam ExpressionType the type of the object on which IO stream operations are performed
+  *
+  * This class represents an expression with stream operators controlled by a given IOFormat.
+  * It is the return type of DenseBase::format()
+  * and most of the time this is the only way it is used.
+  *
+  * See class IOFormat for some examples.
+  *
+  * \sa DenseBase::format(), class IOFormat
+  */
+template<typename ExpressionType>
+class WithFormat
+{
+  public:
+
+    WithFormat(const ExpressionType& matrix, const IOFormat& format)
+      : m_matrix(matrix), m_format(format)
+    {}
+
+    friend std::ostream & operator << (std::ostream & s, const WithFormat& wf)
+    {
+      return internal::print_matrix(s, wf.m_matrix.eval(), wf.m_format);
+    }
+
+  protected:
+    typename ExpressionType::Nested m_matrix;
+    IOFormat m_format;
+};
+
+namespace internal {
+
+// NOTE: This helper is kept for backward compatibility with previous code specializing
+//       this internal::significant_decimals_impl structure. In the future we should directly
+//       call digits10() which has been introduced in July 2016 in 3.3.
+template<typename Scalar>
+struct significant_decimals_impl
+{
+  static inline int run()
+  {
+    return NumTraits<Scalar>::digits10();
+  }
+};
+
+/** \internal
+  * print the matrix \a _m to the output stream \a s using the output format \a fmt */
+template<typename Derived>
+std::ostream & print_matrix(std::ostream & s, const Derived& _m, const IOFormat& fmt)
+{
+  if(_m.size() == 0)
+  {
+    s << fmt.matPrefix << fmt.matSuffix;
+    return s;
+  }
+  
+  typename Derived::Nested m = _m;
+  typedef typename Derived::Scalar Scalar;
+
+  Index width = 0;
+
+  std::streamsize explicit_precision;
+  if(fmt.precision == StreamPrecision)
+  {
+    explicit_precision = 0;
+  }
+  else if(fmt.precision == FullPrecision)
+  {
+    if (NumTraits<Scalar>::IsInteger)
+    {
+      explicit_precision = 0;
+    }
+    else
+    {
+      explicit_precision = significant_decimals_impl<Scalar>::run();
+    }
+  }
+  else
+  {
+    explicit_precision = fmt.precision;
+  }
+
+  std::streamsize old_precision = 0;
+  if(explicit_precision) old_precision = s.precision(explicit_precision);
+
+  bool align_cols = !(fmt.flags & DontAlignCols);
+  if(align_cols)
+  {
+    // compute the largest width
+    for(Index j = 0; j < m.cols(); ++j)
+      for(Index i = 0; i < m.rows(); ++i)
+      {
+        std::stringstream sstr;
+        sstr.copyfmt(s);
+        sstr << m.coeff(i,j);
+        width = std::max<Index>(width, Index(sstr.str().length()));
+      }
+  }
+  s << fmt.matPrefix;
+  for(Index i = 0; i < m.rows(); ++i)
+  {
+    if (i)
+      s << fmt.rowSpacer;
+    s << fmt.rowPrefix;
+    if(width) s.width(width);
+    s << m.coeff(i, 0);
+    for(Index j = 1; j < m.cols(); ++j)
+    {
+      s << fmt.coeffSeparator;
+      if (width) s.width(width);
+      s << m.coeff(i, j);
+    }
+    s << fmt.rowSuffix;
+    if( i < m.rows() - 1)
+      s << fmt.rowSeparator;
+  }
+  s << fmt.matSuffix;
+  if(explicit_precision) s.precision(old_precision);
+  return s;
+}
+
+} // end namespace internal
+
+/** \relates DenseBase
+  *
+  * Outputs the matrix, to the given stream.
+  *
+  * If you wish to print the matrix with a format different than the default, use DenseBase::format().
+  *
+  * It is also possible to change the default format by defining EIGEN_DEFAULT_IO_FORMAT before including Eigen headers.
+  * If not defined, this will automatically be defined to Eigen::IOFormat(), that is the Eigen::IOFormat with default parameters.
+  *
+  * \sa DenseBase::format()
+  */
+template<typename Derived>
+std::ostream & operator <<
+(std::ostream & s,
+ const DenseBase<Derived> & m)
+{
+  return internal::print_matrix(s, m.eval(), EIGEN_DEFAULT_IO_FORMAT);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_IO_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Inverse.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Inverse.h
new file mode 100644
index 0000000..b76f043
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Inverse.h
@@ -0,0 +1,118 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INVERSE_H
+#define EIGEN_INVERSE_H
+
+namespace Eigen { 
+
+template<typename XprType,typename StorageKind> class InverseImpl;
+
+namespace internal {
+
+template<typename XprType>
+struct traits<Inverse<XprType> >
+  : traits<typename XprType::PlainObject>
+{
+  typedef typename XprType::PlainObject PlainObject;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Inverse
+  *
+  * \brief Expression of the inverse of another expression
+  *
+  * \tparam XprType the type of the expression we are taking the inverse
+  *
+  * This class represents an abstract expression of A.inverse()
+  * and most of the time this is the only way it is used.
+  *
+  */
+template<typename XprType>
+class Inverse : public InverseImpl<XprType,typename internal::traits<XprType>::StorageKind>
+{
+public:
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename XprType::PlainObject                       PlainObject;
+  typedef typename XprType::Scalar                            Scalar;
+  typedef typename internal::ref_selector<XprType>::type      XprTypeNested;
+  typedef typename internal::remove_all<XprTypeNested>::type  XprTypeNestedCleaned;
+  typedef typename internal::ref_selector<Inverse>::type Nested;
+  typedef typename internal::remove_all<XprType>::type NestedExpression;
+  
+  explicit EIGEN_DEVICE_FUNC Inverse(const XprType &xpr)
+    : m_xpr(xpr)
+  {}
+
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+
+  EIGEN_DEVICE_FUNC const XprTypeNestedCleaned& nestedExpression() const { return m_xpr; }
+
+protected:
+  XprTypeNested m_xpr;
+};
+
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class InverseImpl
+  : public internal::generic_xpr_base<Inverse<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Inverse<XprType> >::type Base;
+  typedef typename XprType::Scalar Scalar;
+private:
+
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+namespace internal {
+
+/** \internal
+  * \brief Default evaluator for Inverse expression.
+  * 
+  * This default evaluator for Inverse expression simply evaluate the inverse into a temporary
+  * by a call to internal::call_assignment_no_alias.
+  * Therefore, inverse implementers only have to specialize Assignment<Dst,Inverse<...>, ...> for
+  * there own nested expression.
+  *
+  * \sa class Inverse
+  */
+template<typename ArgType>
+struct unary_evaluator<Inverse<ArgType> >
+  : public evaluator<typename Inverse<ArgType>::PlainObject>
+{
+  typedef Inverse<ArgType> InverseType;
+  typedef typename InverseType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+
+  unary_evaluator(const InverseType& inv_xpr)
+    : m_result(inv_xpr.rows(), inv_xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    internal::call_assignment_no_alias(m_result, inv_xpr);
+  }
+  
+protected:
+  PlainObject m_result;
+};
+  
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Map.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Map.h
new file mode 100644
index 0000000..06d1967
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Map.h
@@ -0,0 +1,164 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MAP_H
+#define EIGEN_MAP_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename PlainObjectType, int MapOptions, typename StrideType>
+struct traits<Map<PlainObjectType, MapOptions, StrideType> >
+  : public traits<PlainObjectType>
+{
+  typedef traits<PlainObjectType> TraitsBase;
+  enum {
+    InnerStrideAtCompileTime = StrideType::InnerStrideAtCompileTime == 0
+                             ? int(PlainObjectType::InnerStrideAtCompileTime)
+                             : int(StrideType::InnerStrideAtCompileTime),
+    OuterStrideAtCompileTime = StrideType::OuterStrideAtCompileTime == 0
+                             ? int(PlainObjectType::OuterStrideAtCompileTime)
+                             : int(StrideType::OuterStrideAtCompileTime),
+    Alignment = int(MapOptions)&int(AlignedMask),
+    Flags0 = TraitsBase::Flags & (~NestByRefBit),
+    Flags = is_lvalue<PlainObjectType>::value ? int(Flags0) : (int(Flags0) & ~LvalueBit)
+  };
+private:
+  enum { Options }; // Expressions don't have Options
+};
+}
+
+/** \class Map
+  * \ingroup Core_Module
+  *
+  * \brief A matrix or vector expression mapping an existing array of data.
+  *
+  * \tparam PlainObjectType the equivalent matrix type of the mapped data
+  * \tparam MapOptions specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
+  *                The default is \c #Unaligned.
+  * \tparam StrideType optionally specifies strides. By default, Map assumes the memory layout
+  *                   of an ordinary, contiguous array. This can be overridden by specifying strides.
+  *                   The type passed here must be a specialization of the Stride template, see examples below.
+  *
+  * This class represents a matrix or vector expression mapping an existing array of data.
+  * It can be used to let Eigen interface without any overhead with non-Eigen data structures,
+  * such as plain C arrays or structures from other libraries. By default, it assumes that the
+  * data is laid out contiguously in memory. You can however override this by explicitly specifying
+  * inner and outer strides.
+  *
+  * Here's an example of simply mapping a contiguous array as a \ref TopicStorageOrders "column-major" matrix:
+  * \include Map_simple.cpp
+  * Output: \verbinclude Map_simple.out
+  *
+  * If you need to map non-contiguous arrays, you can do so by specifying strides:
+  *
+  * Here's an example of mapping an array as a vector, specifying an inner stride, that is, the pointer
+  * increment between two consecutive coefficients. Here, we're specifying the inner stride as a compile-time
+  * fixed value.
+  * \include Map_inner_stride.cpp
+  * Output: \verbinclude Map_inner_stride.out
+  *
+  * Here's an example of mapping an array while specifying an outer stride. Here, since we're mapping
+  * as a column-major matrix, 'outer stride' means the pointer increment between two consecutive columns.
+  * Here, we're specifying the outer stride as a runtime parameter. Note that here \c OuterStride<> is
+  * a short version of \c OuterStride<Dynamic> because the default template parameter of OuterStride
+  * is  \c Dynamic
+  * \include Map_outer_stride.cpp
+  * Output: \verbinclude Map_outer_stride.out
+  *
+  * For more details and for an example of specifying both an inner and an outer stride, see class Stride.
+  *
+  * \b Tip: to change the array of data mapped by a Map object, you can use the C++
+  * placement new syntax:
+  *
+  * Example: \include Map_placement_new.cpp
+  * Output: \verbinclude Map_placement_new.out
+  *
+  * This class is the return type of PlainObjectBase::Map() but can also be used directly.
+  *
+  * \sa PlainObjectBase::Map(), \ref TopicStorageOrders
+  */
+template<typename PlainObjectType, int MapOptions, typename StrideType> class Map
+  : public MapBase<Map<PlainObjectType, MapOptions, StrideType> >
+{
+  public:
+
+    typedef MapBase<Map> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Map)
+
+    typedef typename Base::PointerType PointerType;
+    typedef PointerType PointerArgType;
+    EIGEN_DEVICE_FUNC
+    inline PointerType cast_to_pointer_type(PointerArgType ptr) { return ptr; }
+
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const
+    {
+      return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const
+    {
+      return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
+           : IsVectorAtCompileTime ? this->size()
+           : int(Flags)&RowMajorBit ? this->cols()
+           : this->rows();
+    }
+
+    /** Constructor in the fixed-size case.
+      *
+      * \param dataPtr pointer to the array to map
+      * \param stride optional Stride object, passing the strides.
+      */
+    EIGEN_DEVICE_FUNC
+    explicit inline Map(PointerArgType dataPtr, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr)), m_stride(stride)
+    {
+      PlainObjectType::Base::_check_template_params();
+    }
+
+    /** Constructor in the dynamic-size vector case.
+      *
+      * \param dataPtr pointer to the array to map
+      * \param size the size of the vector expression
+      * \param stride optional Stride object, passing the strides.
+      */
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index size, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), size), m_stride(stride)
+    {
+      PlainObjectType::Base::_check_template_params();
+    }
+
+    /** Constructor in the dynamic-size matrix case.
+      *
+      * \param dataPtr pointer to the array to map
+      * \param rows the number of rows of the matrix expression
+      * \param cols the number of columns of the matrix expression
+      * \param stride optional Stride object, passing the strides.
+      */
+    EIGEN_DEVICE_FUNC
+    inline Map(PointerArgType dataPtr, Index rows, Index cols, const StrideType& stride = StrideType())
+      : Base(cast_to_pointer_type(dataPtr), rows, cols), m_stride(stride)
+    {
+      PlainObjectType::Base::_check_template_params();
+    }
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
+
+  protected:
+    StrideType m_stride;
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_MAP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MapBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MapBase.h
new file mode 100644
index 0000000..020f939
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MapBase.h
@@ -0,0 +1,299 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MAPBASE_H
+#define EIGEN_MAPBASE_H
+
+#define EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived) \
+      EIGEN_STATIC_ASSERT((int(internal::evaluator<Derived>::Flags) & LinearAccessBit) || Derived::IsVectorAtCompileTime, \
+                          YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT)
+
+namespace Eigen { 
+
+/** \ingroup Core_Module
+  *
+  * \brief Base class for dense Map and Block expression with direct access
+  *
+  * This base class provides the const low-level accessors (e.g. coeff, coeffRef) of dense
+  * Map and Block objects with direct access.
+  * Typical users do not have to directly deal with this class.
+  *
+  * This class can be extended by through the macro plugin \c EIGEN_MAPBASE_PLUGIN.
+  * See \link TopicCustomizing_Plugins customizing Eigen \endlink for details.
+  *
+  * The \c Derived class has to provide the following two methods describing the memory layout:
+  *  \code Index innerStride() const; \endcode
+  *  \code Index outerStride() const; \endcode
+  *
+  * \sa class Map, class Block
+  */
+template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
+  : public internal::dense_xpr_base<Derived>::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<Derived>::type Base;
+    enum {
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      SizeAtCompileTime = Base::SizeAtCompileTime
+    };
+
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef typename internal::conditional<
+                         bool(internal::is_lvalue<Derived>::value),
+                         Scalar *,
+                         const Scalar *>::type
+                     PointerType;
+
+    using Base::derived;
+//    using Base::RowsAtCompileTime;
+//    using Base::ColsAtCompileTime;
+//    using Base::SizeAtCompileTime;
+    using Base::MaxRowsAtCompileTime;
+    using Base::MaxColsAtCompileTime;
+    using Base::MaxSizeAtCompileTime;
+    using Base::IsVectorAtCompileTime;
+    using Base::Flags;
+    using Base::IsRowMajor;
+
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::coeff;
+    using Base::coeffRef;
+    using Base::lazyAssign;
+    using Base::eval;
+
+    using Base::innerStride;
+    using Base::outerStride;
+    using Base::rowStride;
+    using Base::colStride;
+
+    // bug 217 - compile error on ICC 11.1
+    using Base::operator=;
+
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+
+    /** \copydoc DenseBase::rows() */
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_rows.value(); }
+    /** \copydoc DenseBase::cols() */
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_cols.value(); }
+
+    /** Returns a pointer to the first coefficient of the matrix or vector.
+      *
+      * \note When addressing this data, make sure to honor the strides returned by innerStride() and outerStride().
+      *
+      * \sa innerStride(), outerStride()
+      */
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return m_data; }
+
+    /** \copydoc PlainObjectBase::coeff(Index,Index) const */
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeff(Index rowId, Index colId) const
+    {
+      return m_data[colId * colStride() + rowId * rowStride()];
+    }
+
+    /** \copydoc PlainObjectBase::coeff(Index) const */
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeff(Index index) const
+    {
+      EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
+      return m_data[index * innerStride()];
+    }
+
+    /** \copydoc PlainObjectBase::coeffRef(Index,Index) const */
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      return this->m_data[colId * colStride() + rowId * rowStride()];
+    }
+
+    /** \copydoc PlainObjectBase::coeffRef(Index) const */
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index index) const
+    {
+      EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
+      return this->m_data[index * innerStride()];
+    }
+
+    /** \internal */
+    template<int LoadMode>
+    inline PacketScalar packet(Index rowId, Index colId) const
+    {
+      return internal::ploadt<PacketScalar, LoadMode>
+               (m_data + (colId * colStride() + rowId * rowStride()));
+    }
+
+    /** \internal */
+    template<int LoadMode>
+    inline PacketScalar packet(Index index) const
+    {
+      EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
+      return internal::ploadt<PacketScalar, LoadMode>(m_data + index * innerStride());
+    }
+
+    /** \internal Constructor for fixed size matrices or vectors */
+    EIGEN_DEVICE_FUNC
+    explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
+    {
+      EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
+      checkSanity<Derived>();
+    }
+
+    /** \internal Constructor for dynamically sized vectors */
+    EIGEN_DEVICE_FUNC
+    inline MapBase(PointerType dataPtr, Index vecSize)
+            : m_data(dataPtr),
+              m_rows(RowsAtCompileTime == Dynamic ? vecSize : Index(RowsAtCompileTime)),
+              m_cols(ColsAtCompileTime == Dynamic ? vecSize : Index(ColsAtCompileTime))
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+      eigen_assert(vecSize >= 0);
+      eigen_assert(dataPtr == 0 || SizeAtCompileTime == Dynamic || SizeAtCompileTime == vecSize);
+      checkSanity<Derived>();
+    }
+
+    /** \internal Constructor for dynamically sized matrices */
+    EIGEN_DEVICE_FUNC
+    inline MapBase(PointerType dataPtr, Index rows, Index cols)
+            : m_data(dataPtr), m_rows(rows), m_cols(cols)
+    {
+      eigen_assert( (dataPtr == 0)
+              || (   rows >= 0 && (RowsAtCompileTime == Dynamic || RowsAtCompileTime == rows)
+                  && cols >= 0 && (ColsAtCompileTime == Dynamic || ColsAtCompileTime == cols)));
+      checkSanity<Derived>();
+    }
+
+    #ifdef EIGEN_MAPBASE_PLUGIN
+    #include EIGEN_MAPBASE_PLUGIN
+    #endif
+
+  protected:
+
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    void checkSanity(typename internal::enable_if<(internal::traits<T>::Alignment>0),void*>::type = 0) const
+    {
+#if EIGEN_MAX_ALIGN_BYTES>0
+      eigen_assert((   ((internal::UIntPtr(m_data) % internal::traits<Derived>::Alignment) == 0)
+                    || (cols() * rows() * innerStride() * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
+#endif
+    }
+
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    void checkSanity(typename internal::enable_if<internal::traits<T>::Alignment==0,void*>::type = 0) const
+    {}
+
+    PointerType m_data;
+    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_rows;
+    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_cols;
+};
+
+/** \ingroup Core_Module
+  *
+  * \brief Base class for non-const dense Map and Block expression with direct access
+  *
+  * This base class provides the non-const low-level accessors (e.g. coeff and coeffRef) of
+  * dense Map and Block objects with direct access.
+  * It inherits MapBase<Derived, ReadOnlyAccessors> which defines the const variant for reading specific entries.
+  *
+  * \sa class Map, class Block
+  */
+template<typename Derived> class MapBase<Derived, WriteAccessors>
+  : public MapBase<Derived, ReadOnlyAccessors>
+{
+    typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase;
+  public:
+
+    typedef MapBase<Derived, ReadOnlyAccessors> Base;
+
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::PacketScalar PacketScalar;
+    typedef typename Base::StorageIndex StorageIndex;
+    typedef typename Base::PointerType PointerType;
+
+    using Base::derived;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::coeff;
+    using Base::coeffRef;
+
+    using Base::innerStride;
+    using Base::outerStride;
+    using Base::rowStride;
+    using Base::colStride;
+
+    typedef typename internal::conditional<
+                    internal::is_lvalue<Derived>::value,
+                    Scalar,
+                    const Scalar
+                  >::type ScalarWithConstIfNotLvalue;
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar* data() const { return this->m_data; }
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue* data() { return this->m_data; } // no const-cast here so non-const-correct code will give a compile error
+
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue& coeffRef(Index row, Index col)
+    {
+      return this->m_data[col * colStride() + row * rowStride()];
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline ScalarWithConstIfNotLvalue& coeffRef(Index index)
+    {
+      EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
+      return this->m_data[index * innerStride()];
+    }
+
+    template<int StoreMode>
+    inline void writePacket(Index row, Index col, const PacketScalar& val)
+    {
+      internal::pstoret<Scalar, PacketScalar, StoreMode>
+               (this->m_data + (col * colStride() + row * rowStride()), val);
+    }
+
+    template<int StoreMode>
+    inline void writePacket(Index index, const PacketScalar& val)
+    {
+      EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS(Derived)
+      internal::pstoret<Scalar, PacketScalar, StoreMode>
+                (this->m_data + index * innerStride(), val);
+    }
+
+    EIGEN_DEVICE_FUNC explicit inline MapBase(PointerType dataPtr) : Base(dataPtr) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index vecSize) : Base(dataPtr, vecSize) {}
+    EIGEN_DEVICE_FUNC inline MapBase(PointerType dataPtr, Index rows, Index cols) : Base(dataPtr, rows, cols) {}
+
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const MapBase& other)
+    {
+      ReadOnlyMapBase::Base::operator=(other);
+      return derived();
+    }
+
+    // In theory we could simply refer to Base:Base::operator=, but MSVC does not like Base::Base,
+    // see bugs 821 and 920.
+    using ReadOnlyMapBase::Base::operator=;
+};
+
+#undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS
+
+} // end namespace Eigen
+
+#endif // EIGEN_MAPBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h
new file mode 100644
index 0000000..a648aa0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctions.h
@@ -0,0 +1,1431 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATHFUNCTIONS_H
+#define EIGEN_MATHFUNCTIONS_H
+
+// source: http://www.geom.uiuc.edu/~huberty/math5337/groupe/digits.html
+// TODO this should better be moved to NumTraits
+#define EIGEN_PI 3.141592653589793238462643383279502884197169399375105820974944592307816406L
+
+
+namespace Eigen {
+
+// On WINCE, std::abs is defined for int only, so let's defined our own overloads:
+// This issue has been confirmed with MSVC 2008 only, but the issue might exist for more recent versions too.
+#if EIGEN_OS_WINCE && EIGEN_COMP_MSVC && EIGEN_COMP_MSVC<=1500
+long        abs(long        x) { return (labs(x));  }
+double      abs(double      x) { return (fabs(x));  }
+float       abs(float       x) { return (fabsf(x)); }
+long double abs(long double x) { return (fabsl(x)); }
+#endif
+
+namespace internal {
+
+/** \internal \class global_math_functions_filtering_base
+  *
+  * What it does:
+  * Defines a typedef 'type' as follows:
+  * - if type T has a member typedef Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl, then
+  *   global_math_functions_filtering_base<T>::type is a typedef for it.
+  * - otherwise, global_math_functions_filtering_base<T>::type is a typedef for T.
+  *
+  * How it's used:
+  * To allow to defined the global math functions (like sin...) in certain cases, like the Array expressions.
+  * When you do sin(array1+array2), the object array1+array2 has a complicated expression type, all what you want to know
+  * is that it inherits ArrayBase. So we implement a partial specialization of sin_impl for ArrayBase<Derived>.
+  * So we must make sure to use sin_impl<ArrayBase<Derived> > and not sin_impl<Derived>, otherwise our partial specialization
+  * won't be used. How does sin know that? That's exactly what global_math_functions_filtering_base tells it.
+  *
+  * How it's implemented:
+  * SFINAE in the style of enable_if. Highly susceptible of breaking compilers. With GCC, it sure does work, but if you replace
+  * the typename dummy by an integer template parameter, it doesn't work anymore!
+  */
+
+template<typename T, typename dummy = void>
+struct global_math_functions_filtering_base
+{
+  typedef T type;
+};
+
+template<typename T> struct always_void { typedef void type; };
+
+template<typename T>
+struct global_math_functions_filtering_base
+  <T,
+   typename always_void<typename T::Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl>::type
+  >
+{
+  typedef typename T::Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl type;
+};
+
+#define EIGEN_MATHFUNC_IMPL(func, scalar) Eigen::internal::func##_impl<typename Eigen::internal::global_math_functions_filtering_base<scalar>::type>
+#define EIGEN_MATHFUNC_RETVAL(func, scalar) typename Eigen::internal::func##_retval<typename Eigen::internal::global_math_functions_filtering_base<scalar>::type>::type
+
+/****************************************************************************
+* Implementation of real                                                 *
+****************************************************************************/
+
+template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+struct real_default_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    return x;
+  }
+};
+
+template<typename Scalar>
+struct real_default_impl<Scalar,true>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    using std::real;
+    return real(x);
+  }
+};
+
+template<typename Scalar> struct real_impl : real_default_impl<Scalar> {};
+
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct real_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.real();
+  }
+};
+#endif
+
+template<typename Scalar>
+struct real_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of imag                                                 *
+****************************************************************************/
+
+template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+struct imag_default_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar&)
+  {
+    return RealScalar(0);
+  }
+};
+
+template<typename Scalar>
+struct imag_default_impl<Scalar,true>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    using std::imag;
+    return imag(x);
+  }
+};
+
+template<typename Scalar> struct imag_impl : imag_default_impl<Scalar> {};
+
+#ifdef __CUDA_ARCH__
+template<typename T>
+struct imag_impl<std::complex<T> >
+{
+  typedef T RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline T run(const std::complex<T>& x)
+  {
+    return x.imag();
+  }
+};
+#endif
+
+template<typename Scalar>
+struct imag_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of real_ref                                             *
+****************************************************************************/
+
+template<typename Scalar>
+struct real_ref_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar& run(Scalar& x)
+  {
+    return reinterpret_cast<RealScalar*>(&x)[0];
+  }
+  EIGEN_DEVICE_FUNC
+  static inline const RealScalar& run(const Scalar& x)
+  {
+    return reinterpret_cast<const RealScalar*>(&x)[0];
+  }
+};
+
+template<typename Scalar>
+struct real_ref_retval
+{
+  typedef typename NumTraits<Scalar>::Real & type;
+};
+
+/****************************************************************************
+* Implementation of imag_ref                                             *
+****************************************************************************/
+
+template<typename Scalar, bool IsComplex>
+struct imag_ref_default_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar& run(Scalar& x)
+  {
+    return reinterpret_cast<RealScalar*>(&x)[1];
+  }
+  EIGEN_DEVICE_FUNC
+  static inline const RealScalar& run(const Scalar& x)
+  {
+    return reinterpret_cast<RealScalar*>(&x)[1];
+  }
+};
+
+template<typename Scalar>
+struct imag_ref_default_impl<Scalar, false>
+{
+  EIGEN_DEVICE_FUNC
+  static inline Scalar run(Scalar&)
+  {
+    return Scalar(0);
+  }
+  EIGEN_DEVICE_FUNC
+  static inline const Scalar run(const Scalar&)
+  {
+    return Scalar(0);
+  }
+};
+
+template<typename Scalar>
+struct imag_ref_impl : imag_ref_default_impl<Scalar, NumTraits<Scalar>::IsComplex> {};
+
+template<typename Scalar>
+struct imag_ref_retval
+{
+  typedef typename NumTraits<Scalar>::Real & type;
+};
+
+/****************************************************************************
+* Implementation of conj                                                 *
+****************************************************************************/
+
+template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+struct conj_impl
+{
+  EIGEN_DEVICE_FUNC
+  static inline Scalar run(const Scalar& x)
+  {
+    return x;
+  }
+};
+
+template<typename Scalar>
+struct conj_impl<Scalar,true>
+{
+  EIGEN_DEVICE_FUNC
+  static inline Scalar run(const Scalar& x)
+  {
+    using std::conj;
+    return conj(x);
+  }
+};
+
+template<typename Scalar>
+struct conj_retval
+{
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of abs2                                                 *
+****************************************************************************/
+
+template<typename Scalar,bool IsComplex>
+struct abs2_impl_default
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    return x*x;
+  }
+};
+
+template<typename Scalar>
+struct abs2_impl_default<Scalar, true> // IsComplex
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    return real(x)*real(x) + imag(x)*imag(x);
+  }
+};
+
+template<typename Scalar>
+struct abs2_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    return abs2_impl_default<Scalar,NumTraits<Scalar>::IsComplex>::run(x);
+  }
+};
+
+template<typename Scalar>
+struct abs2_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of norm1                                                *
+****************************************************************************/
+
+template<typename Scalar, bool IsComplex>
+struct norm1_default_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar run(const Scalar& x)
+  {
+    EIGEN_USING_STD_MATH(abs);
+    return abs(real(x)) + abs(imag(x));
+  }
+};
+
+template<typename Scalar>
+struct norm1_default_impl<Scalar, false>
+{
+  EIGEN_DEVICE_FUNC
+  static inline Scalar run(const Scalar& x)
+  {
+    EIGEN_USING_STD_MATH(abs);
+    return abs(x);
+  }
+};
+
+template<typename Scalar>
+struct norm1_impl : norm1_default_impl<Scalar, NumTraits<Scalar>::IsComplex> {};
+
+template<typename Scalar>
+struct norm1_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of hypot                                                *
+****************************************************************************/
+
+template<typename Scalar>
+struct hypot_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  static inline RealScalar run(const Scalar& x, const Scalar& y)
+  {
+    EIGEN_USING_STD_MATH(abs);
+    EIGEN_USING_STD_MATH(sqrt);
+    RealScalar _x = abs(x);
+    RealScalar _y = abs(y);
+    Scalar p, qp;
+    if(_x>_y)
+    {
+      p = _x;
+      qp = _y / p;
+    }
+    else
+    {
+      p = _y;
+      qp = _x / p;
+    }
+    if(p==RealScalar(0)) return RealScalar(0);
+    return p * sqrt(RealScalar(1) + qp*qp);
+  }
+};
+
+template<typename Scalar>
+struct hypot_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of cast                                                 *
+****************************************************************************/
+
+template<typename OldType, typename NewType>
+struct cast_impl
+{
+  EIGEN_DEVICE_FUNC
+  static inline NewType run(const OldType& x)
+  {
+    return static_cast<NewType>(x);
+  }
+};
+
+// here, for once, we're plainly returning NewType: we don't want cast to do weird things.
+
+template<typename OldType, typename NewType>
+EIGEN_DEVICE_FUNC
+inline NewType cast(const OldType& x)
+{
+  return cast_impl<OldType, NewType>::run(x);
+}
+
+/****************************************************************************
+* Implementation of round                                                   *
+****************************************************************************/
+
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct round_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      using std::round;
+      return round(x);
+    }
+  };
+#else
+  template<typename Scalar>
+  struct round_impl
+  {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_STATIC_ASSERT((!NumTraits<Scalar>::IsComplex), NUMERIC_TYPE_MUST_BE_REAL)
+      EIGEN_USING_STD_MATH(floor);
+      EIGEN_USING_STD_MATH(ceil);
+      return (x > Scalar(0)) ? floor(x + Scalar(0.5)) : ceil(x - Scalar(0.5));
+    }
+  };
+#endif
+
+template<typename Scalar>
+struct round_retval
+{
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of arg                                                     *
+****************************************************************************/
+
+#if EIGEN_HAS_CXX11_MATH
+  template<typename Scalar>
+  struct arg_impl {
+    static inline Scalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+#else
+  template<typename Scalar, bool IsComplex = NumTraits<Scalar>::IsComplex>
+  struct arg_default_impl
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      return (x < Scalar(0)) ? Scalar(EIGEN_PI) : Scalar(0); }
+  };
+
+  template<typename Scalar>
+  struct arg_default_impl<Scalar,true>
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_DEVICE_FUNC
+    static inline RealScalar run(const Scalar& x)
+    {
+      EIGEN_USING_STD_MATH(arg);
+      return arg(x);
+    }
+  };
+
+  template<typename Scalar> struct arg_impl : arg_default_impl<Scalar> {};
+#endif
+
+template<typename Scalar>
+struct arg_retval
+{
+  typedef typename NumTraits<Scalar>::Real type;
+};
+
+/****************************************************************************
+* Implementation of log1p                                                   *
+****************************************************************************/
+
+namespace std_fallback {
+  // fallback log1p implementation in case there is no log1p(Scalar) function in namespace of Scalar,
+  // or that there is no suitable std::log1p function available
+  template<typename Scalar>
+  EIGEN_DEVICE_FUNC inline Scalar log1p(const Scalar& x) {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    EIGEN_USING_STD_MATH(log);
+    Scalar x1p = RealScalar(1) + x;
+    return ( x1p == Scalar(1) ) ? x : x * ( log(x1p) / (x1p - RealScalar(1)) );
+  }
+}
+
+template<typename Scalar>
+struct log1p_impl {
+  static inline Scalar run(const Scalar& x)
+  {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar)
+    #if EIGEN_HAS_CXX11_MATH
+    using std::log1p;
+    #endif
+    using std_fallback::log1p;
+    return log1p(x);
+  }
+};
+
+
+template<typename Scalar>
+struct log1p_retval
+{
+  typedef Scalar type;
+};
+
+/****************************************************************************
+* Implementation of pow                                                  *
+****************************************************************************/
+
+template<typename ScalarX,typename ScalarY, bool IsInteger = NumTraits<ScalarX>::IsInteger&&NumTraits<ScalarY>::IsInteger>
+struct pow_impl
+{
+  //typedef Scalar retval;
+  typedef typename ScalarBinaryOpTraits<ScalarX,ScalarY,internal::scalar_pow_op<ScalarX,ScalarY> >::ReturnType result_type;
+  static EIGEN_DEVICE_FUNC inline result_type run(const ScalarX& x, const ScalarY& y)
+  {
+    EIGEN_USING_STD_MATH(pow);
+    return pow(x, y);
+  }
+};
+
+template<typename ScalarX,typename ScalarY>
+struct pow_impl<ScalarX,ScalarY, true>
+{
+  typedef ScalarX result_type;
+  static EIGEN_DEVICE_FUNC inline ScalarX run(ScalarX x, ScalarY y)
+  {
+    ScalarX res(1);
+    eigen_assert(!NumTraits<ScalarY>::IsSigned || y >= 0);
+    if(y & 1) res *= x;
+    y >>= 1;
+    while(y)
+    {
+      x *= x;
+      if(y&1) res *= x;
+      y >>= 1;
+    }
+    return res;
+  }
+};
+
+/****************************************************************************
+* Implementation of random                                               *
+****************************************************************************/
+
+template<typename Scalar,
+         bool IsComplex,
+         bool IsInteger>
+struct random_default_impl {};
+
+template<typename Scalar>
+struct random_impl : random_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {};
+
+template<typename Scalar>
+struct random_retval
+{
+  typedef Scalar type;
+};
+
+template<typename Scalar> inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random(const Scalar& x, const Scalar& y);
+template<typename Scalar> inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random();
+
+template<typename Scalar>
+struct random_default_impl<Scalar, false, false>
+{
+  static inline Scalar run(const Scalar& x, const Scalar& y)
+  {
+    return x + (y-x) * Scalar(std::rand()) / Scalar(RAND_MAX);
+  }
+  static inline Scalar run()
+  {
+    return run(Scalar(NumTraits<Scalar>::IsSigned ? -1 : 0), Scalar(1));
+  }
+};
+
+enum {
+  meta_floor_log2_terminate,
+  meta_floor_log2_move_up,
+  meta_floor_log2_move_down,
+  meta_floor_log2_bogus
+};
+
+template<unsigned int n, int lower, int upper> struct meta_floor_log2_selector
+{
+  enum { middle = (lower + upper) / 2,
+         value = (upper <= lower + 1) ? int(meta_floor_log2_terminate)
+               : (n < (1 << middle)) ? int(meta_floor_log2_move_down)
+               : (n==0) ? int(meta_floor_log2_bogus)
+               : int(meta_floor_log2_move_up)
+  };
+};
+
+template<unsigned int n,
+         int lower = 0,
+         int upper = sizeof(unsigned int) * CHAR_BIT - 1,
+         int selector = meta_floor_log2_selector<n, lower, upper>::value>
+struct meta_floor_log2 {};
+
+template<unsigned int n, int lower, int upper>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_down>
+{
+  enum { value = meta_floor_log2<n, lower, meta_floor_log2_selector<n, lower, upper>::middle>::value };
+};
+
+template<unsigned int n, int lower, int upper>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_move_up>
+{
+  enum { value = meta_floor_log2<n, meta_floor_log2_selector<n, lower, upper>::middle, upper>::value };
+};
+
+template<unsigned int n, int lower, int upper>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_terminate>
+{
+  enum { value = (n >= ((unsigned int)(1) << (lower+1))) ? lower+1 : lower };
+};
+
+template<unsigned int n, int lower, int upper>
+struct meta_floor_log2<n, lower, upper, meta_floor_log2_bogus>
+{
+  // no value, error at compile time
+};
+
+template<typename Scalar>
+struct random_default_impl<Scalar, false, true>
+{
+  static inline Scalar run(const Scalar& x, const Scalar& y)
+  { 
+    typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
+    if(y<x)
+      return x;
+    // the following difference might overflow on a 32 bits system,
+    // but since y>=x the result converted to an unsigned long is still correct.
+    std::size_t range = ScalarX(y)-ScalarX(x);
+    std::size_t offset = 0;
+    // rejection sampling
+    std::size_t divisor = 1;
+    std::size_t multiplier = 1;
+    if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1);
+    else               multiplier = 1 + range/(std::size_t(RAND_MAX)+1);
+    do {
+      offset = (std::size_t(std::rand()) * multiplier) / divisor;
+    } while (offset > range);
+    return Scalar(ScalarX(x) + offset);
+  }
+
+  static inline Scalar run()
+  {
+#ifdef EIGEN_MAKING_DOCS
+    return run(Scalar(NumTraits<Scalar>::IsSigned ? -10 : 0), Scalar(10));
+#else
+    enum { rand_bits = meta_floor_log2<(unsigned int)(RAND_MAX)+1>::value,
+           scalar_bits = sizeof(Scalar) * CHAR_BIT,
+           shift = EIGEN_PLAIN_ENUM_MAX(0, int(rand_bits) - int(scalar_bits)),
+           offset = NumTraits<Scalar>::IsSigned ? (1 << (EIGEN_PLAIN_ENUM_MIN(rand_bits,scalar_bits)-1)) : 0
+    };
+    return Scalar((std::rand() >> shift) - offset);
+#endif
+  }
+};
+
+template<typename Scalar>
+struct random_default_impl<Scalar, true, false>
+{
+  static inline Scalar run(const Scalar& x, const Scalar& y)
+  {
+    return Scalar(random(real(x), real(y)),
+                  random(imag(x), imag(y)));
+  }
+  static inline Scalar run()
+  {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    return Scalar(random<RealScalar>(), random<RealScalar>());
+  }
+};
+
+template<typename Scalar>
+inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random(const Scalar& x, const Scalar& y)
+{
+  return EIGEN_MATHFUNC_IMPL(random, Scalar)::run(x, y);
+}
+
+template<typename Scalar>
+inline EIGEN_MATHFUNC_RETVAL(random, Scalar) random()
+{
+  return EIGEN_MATHFUNC_IMPL(random, Scalar)::run();
+}
+
+// Implementatin of is* functions
+
+// std::is* do not work with fast-math and gcc, std::is* are available on MSVC 2013 and newer, as well as in clang.
+#if (EIGEN_HAS_CXX11_MATH && !(EIGEN_COMP_GNUC_STRICT && __FINITE_MATH_ONLY__)) || (EIGEN_COMP_MSVC>=1800) || (EIGEN_COMP_CLANG)
+#define EIGEN_USE_STD_FPCLASSIFY 1
+#else
+#define EIGEN_USE_STD_FPCLASSIFY 0
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isnan_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isinf_impl(const T&) { return false; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<internal::is_integral<T>::value,bool>::type
+isfinite_impl(const T&) { return true; }
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isfinite_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isfinite)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isfinite;
+    return isfinite EIGEN_NOT_A_MACRO (x);
+  #else
+    return x<=NumTraits<T>::highest() && x>=NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isinf_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isinf)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isinf;
+    return isinf EIGEN_NOT_A_MACRO (x);
+  #else
+    return x>NumTraits<T>::highest() || x<NumTraits<T>::lowest();
+  #endif
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+typename internal::enable_if<(!internal::is_integral<T>::value)&&(!NumTraits<T>::IsComplex),bool>::type
+isnan_impl(const T& x)
+{
+  #ifdef __CUDA_ARCH__
+    return (::isnan)(x);
+  #elif EIGEN_USE_STD_FPCLASSIFY
+    using std::isnan;
+    return isnan EIGEN_NOT_A_MACRO (x);
+  #else
+    return x != x;
+  #endif
+}
+
+#if (!EIGEN_USE_STD_FPCLASSIFY)
+
+#if EIGEN_COMP_MSVC
+
+template<typename T> EIGEN_DEVICE_FUNC bool isinf_msvc_helper(T x)
+{
+  return _fpclass(x)==_FPCLASS_NINF || _fpclass(x)==_FPCLASS_PINF;
+}
+
+//MSVC defines a _isnan builtin function, but for double only
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const long double& x) { return _isnan(x)!=0; }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const double& x)      { return _isnan(x)!=0; }
+EIGEN_DEVICE_FUNC inline bool isnan_impl(const float& x)       { return _isnan(x)!=0; }
+
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const long double& x) { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const double& x)      { return isinf_msvc_helper(x); }
+EIGEN_DEVICE_FUNC inline bool isinf_impl(const float& x)       { return isinf_msvc_helper(x); }
+
+#elif (defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ && EIGEN_COMP_GNUC)
+
+#if EIGEN_GNUC_AT_LEAST(5,0)
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((optimize("no-finite-math-only")))
+#else
+  // NOTE the inline qualifier and noinline attribute are both needed: the former is to avoid linking issue (duplicate symbol),
+  //      while the second prevent too aggressive optimizations in fast-math mode:
+  #define EIGEN_TMP_NOOPT_ATTRIB EIGEN_DEVICE_FUNC inline __attribute__((noinline,optimize("no-finite-math-only")))
+#endif
+
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const long double& x) { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const double& x)      { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isnan_impl(const float& x)       { return __builtin_isnan(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const double& x)      { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const float& x)       { return __builtin_isinf(x); }
+template<> EIGEN_TMP_NOOPT_ATTRIB bool isinf_impl(const long double& x) { return __builtin_isinf(x); }
+
+#undef EIGEN_TMP_NOOPT_ATTRIB
+
+#endif
+
+#endif
+
+// The following overload are defined at the end of this file
+template<typename T> EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x);
+template<typename T> EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x);
+template<typename T> EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x);
+
+template<typename T> T generic_fast_tanh_float(const T& a_x);
+
+} // end namespace internal
+
+/****************************************************************************
+* Generic math functions                                                    *
+****************************************************************************/
+
+namespace numext {
+
+#ifndef __CUDA_ARCH__
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(min);
+  return min EIGEN_NOT_A_MACRO (x,y);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  EIGEN_USING_STD_MATH(max);
+  return max EIGEN_NOT_A_MACRO (x,y);
+}
+#else
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T mini(const T& x, const T& y)
+{
+  return y < x ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float mini(const float& x, const float& y)
+{
+  return fminf(x, y);
+}
+template<typename T>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE T maxi(const T& x, const T& y)
+{
+  return x < y ? y : x;
+}
+template<>
+EIGEN_DEVICE_FUNC
+EIGEN_ALWAYS_INLINE float maxi(const float& x, const float& y)
+{
+  return fmaxf(x, y);
+}
+#endif
+
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(real, Scalar) real(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(real, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) >::type real_ref(const Scalar& x)
+{
+  return internal::real_ref_impl<Scalar>::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(real_ref, Scalar) real_ref(Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(real_ref, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(imag, Scalar) imag(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(imag, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(arg, Scalar) arg(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(arg, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline typename internal::add_const_on_value_type< EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) >::type imag_ref(const Scalar& x)
+{
+  return internal::imag_ref_impl<Scalar>::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(imag_ref, Scalar) imag_ref(Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(imag_ref, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(conj, Scalar) conj(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(conj, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(abs2, Scalar) abs2(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(abs2, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(norm1, Scalar) norm1(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(norm1, Scalar)::run(x);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(hypot, Scalar) hypot(const Scalar& x, const Scalar& y)
+{
+  return EIGEN_MATHFUNC_IMPL(hypot, Scalar)::run(x, y);
+}
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(log1p, Scalar) log1p(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(log1p, Scalar)::run(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float log1p(const float &x) { return ::log1pf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double log1p(const double &x) { return ::log1p(x); }
+#endif
+
+template<typename ScalarX,typename ScalarY>
+EIGEN_DEVICE_FUNC
+inline typename internal::pow_impl<ScalarX,ScalarY>::result_type pow(const ScalarX& x, const ScalarY& y)
+{
+  return internal::pow_impl<ScalarX,ScalarY>::run(x, y);
+}
+
+template<typename T> EIGEN_DEVICE_FUNC bool (isnan)   (const T &x) { return internal::isnan_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isinf)   (const T &x) { return internal::isinf_impl(x); }
+template<typename T> EIGEN_DEVICE_FUNC bool (isfinite)(const T &x) { return internal::isfinite_impl(x); }
+
+template<typename Scalar>
+EIGEN_DEVICE_FUNC
+inline EIGEN_MATHFUNC_RETVAL(round, Scalar) round(const Scalar& x)
+{
+  return EIGEN_MATHFUNC_IMPL(round, Scalar)::run(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (floor)(const T& x)
+{
+  EIGEN_USING_STD_MATH(floor);
+  return floor(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float floor(const float &x) { return ::floorf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double floor(const double &x) { return ::floor(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+T (ceil)(const T& x)
+{
+  EIGEN_USING_STD_MATH(ceil);
+  return ceil(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float ceil(const float &x) { return ::ceilf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double ceil(const double &x) { return ::ceil(x); }
+#endif
+
+
+/** Log base 2 for 32 bits positive integers.
+  * Conveniently returns 0 for x==0. */
+inline int log2(int x)
+{
+  eigen_assert(x>=0);
+  unsigned int v(x);
+  static const int table[32] = { 0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30, 8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31 };
+  v |= v >> 1;
+  v |= v >> 2;
+  v |= v >> 4;
+  v |= v >> 8;
+  v |= v >> 16;
+  return table[(v * 0x07C4ACDDU) >> 27];
+}
+
+/** \returns the square root of \a x.
+  *
+  * It is essentially equivalent to \code using std::sqrt; return sqrt(x); \endcode,
+  * but slightly faster for float/double and some compilers (e.g., gcc), thanks to
+  * specializations when SSE is enabled.
+  *
+  * It's usage is justified in performance critical functions, like norm/normalize.
+  */
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sqrt(const T &x)
+{
+  EIGEN_USING_STD_MATH(sqrt);
+  return sqrt(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T log(const T &x) {
+  EIGEN_USING_STD_MATH(log);
+  return log(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float log(const float &x) { return ::logf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double log(const double &x) { return ::log(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+typename internal::enable_if<NumTraits<T>::IsSigned || NumTraits<T>::IsComplex,typename NumTraits<T>::Real>::type
+abs(const T &x) {
+  EIGEN_USING_STD_MATH(abs);
+  return abs(x);
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+typename internal::enable_if<!(NumTraits<T>::IsSigned || NumTraits<T>::IsComplex),typename NumTraits<T>::Real>::type
+abs(const T &x) {
+  return x;
+}
+
+#if defined(__SYCL_DEVICE_ONLY__)
+EIGEN_ALWAYS_INLINE float   abs(float x) { return cl::sycl::fabs(x); }
+EIGEN_ALWAYS_INLINE double  abs(double x) { return cl::sycl::fabs(x); }
+#endif // defined(__SYCL_DEVICE_ONLY__)
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float abs(const float &x) { return ::fabsf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double abs(const double &x) { return ::fabs(x); }
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float abs(const std::complex<float>& x) {
+  return ::hypotf(x.real(), x.imag());
+}
+
+template <> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double abs(const std::complex<double>& x) {
+  return ::hypot(x.real(), x.imag());
+}
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T exp(const T &x) {
+  EIGEN_USING_STD_MATH(exp);
+  return exp(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float exp(const float &x) { return ::expf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double exp(const double &x) { return ::exp(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T cos(const T &x) {
+  EIGEN_USING_STD_MATH(cos);
+  return cos(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float cos(const float &x) { return ::cosf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double cos(const double &x) { return ::cos(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sin(const T &x) {
+  EIGEN_USING_STD_MATH(sin);
+  return sin(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sin(const float &x) { return ::sinf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sin(const double &x) { return ::sin(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T tan(const T &x) {
+  EIGEN_USING_STD_MATH(tan);
+  return tan(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tan(const float &x) { return ::tanf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double tan(const double &x) { return ::tan(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T acos(const T &x) {
+  EIGEN_USING_STD_MATH(acos);
+  return acos(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float acos(const float &x) { return ::acosf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double acos(const double &x) { return ::acos(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T asin(const T &x) {
+  EIGEN_USING_STD_MATH(asin);
+  return asin(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float asin(const float &x) { return ::asinf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double asin(const double &x) { return ::asin(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T atan(const T &x) {
+  EIGEN_USING_STD_MATH(atan);
+  return atan(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float atan(const float &x) { return ::atanf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double atan(const double &x) { return ::atan(x); }
+#endif
+
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T cosh(const T &x) {
+  EIGEN_USING_STD_MATH(cosh);
+  return cosh(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float cosh(const float &x) { return ::coshf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double cosh(const double &x) { return ::cosh(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T sinh(const T &x) {
+  EIGEN_USING_STD_MATH(sinh);
+  return sinh(x);
+}
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sinh(const float &x) { return ::sinhf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sinh(const double &x) { return ::sinh(x); }
+#endif
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T tanh(const T &x) {
+  EIGEN_USING_STD_MATH(tanh);
+  return tanh(x);
+}
+
+#if (!defined(__CUDACC__)) && EIGEN_FAST_MATH
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tanh(float x) { return internal::generic_fast_tanh_float(x); }
+#endif
+
+#ifdef __CUDACC__
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float tanh(const float &x) { return ::tanhf(x); }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double tanh(const double &x) { return ::tanh(x); }
+#endif
+
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+T fmod(const T& a, const T& b) {
+  EIGEN_USING_STD_MATH(fmod);
+  return fmod(a, b);
+}
+
+#ifdef __CUDACC__
+template <>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float fmod(const float& a, const float& b) {
+  return ::fmodf(a, b);
+}
+
+template <>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double fmod(const double& a, const double& b) {
+  return ::fmod(a, b);
+}
+#endif
+
+} // end namespace numext
+
+namespace internal {
+
+template<typename T>
+EIGEN_DEVICE_FUNC bool isfinite_impl(const std::complex<T>& x)
+{
+  return (numext::isfinite)(numext::real(x)) && (numext::isfinite)(numext::imag(x));
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC bool isnan_impl(const std::complex<T>& x)
+{
+  return (numext::isnan)(numext::real(x)) || (numext::isnan)(numext::imag(x));
+}
+
+template<typename T>
+EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x)
+{
+  return ((numext::isinf)(numext::real(x)) || (numext::isinf)(numext::imag(x))) && (!(numext::isnan)(x));
+}
+
+/****************************************************************************
+* Implementation of fuzzy comparisons                                       *
+****************************************************************************/
+
+template<typename Scalar,
+         bool IsComplex,
+         bool IsInteger>
+struct scalar_fuzzy_default_impl {};
+
+template<typename Scalar>
+struct scalar_fuzzy_default_impl<Scalar, false, false>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
+  static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
+  {
+    return numext::abs(x) <= numext::abs(y) * prec;
+  }
+  EIGEN_DEVICE_FUNC
+  static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
+  {
+    return numext::abs(x - y) <= numext::mini(numext::abs(x), numext::abs(y)) * prec;
+  }
+  EIGEN_DEVICE_FUNC
+  static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar& prec)
+  {
+    return x <= y || isApprox(x, y, prec);
+  }
+};
+
+template<typename Scalar>
+struct scalar_fuzzy_default_impl<Scalar, false, true>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
+  static inline bool isMuchSmallerThan(const Scalar& x, const Scalar&, const RealScalar&)
+  {
+    return x == Scalar(0);
+  }
+  EIGEN_DEVICE_FUNC
+  static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar&)
+  {
+    return x == y;
+  }
+  EIGEN_DEVICE_FUNC
+  static inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y, const RealScalar&)
+  {
+    return x <= y;
+  }
+};
+
+template<typename Scalar>
+struct scalar_fuzzy_default_impl<Scalar, true, false>
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
+  static inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y, const RealScalar& prec)
+  {
+    return numext::abs2(x) <= numext::abs2(y) * prec * prec;
+  }
+  EIGEN_DEVICE_FUNC
+  static inline bool isApprox(const Scalar& x, const Scalar& y, const RealScalar& prec)
+  {
+    return numext::abs2(x - y) <= numext::mini(numext::abs2(x), numext::abs2(y)) * prec * prec;
+  }
+};
+
+template<typename Scalar>
+struct scalar_fuzzy_impl : scalar_fuzzy_default_impl<Scalar, NumTraits<Scalar>::IsComplex, NumTraits<Scalar>::IsInteger> {};
+
+template<typename Scalar, typename OtherScalar> EIGEN_DEVICE_FUNC
+inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y,
+                              const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
+{
+  return scalar_fuzzy_impl<Scalar>::template isMuchSmallerThan<OtherScalar>(x, y, precision);
+}
+
+template<typename Scalar> EIGEN_DEVICE_FUNC
+inline bool isApprox(const Scalar& x, const Scalar& y,
+                     const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
+{
+  return scalar_fuzzy_impl<Scalar>::isApprox(x, y, precision);
+}
+
+template<typename Scalar> EIGEN_DEVICE_FUNC
+inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y,
+                               const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
+{
+  return scalar_fuzzy_impl<Scalar>::isApproxOrLessThan(x, y, precision);
+}
+
+/******************************************
+***  The special case of the  bool type ***
+******************************************/
+
+template<> struct random_impl<bool>
+{
+  static inline bool run()
+  {
+    return random<int>(0,1)==0 ? false : true;
+  }
+};
+
+template<> struct scalar_fuzzy_impl<bool>
+{
+  typedef bool RealScalar;
+  
+  template<typename OtherScalar> EIGEN_DEVICE_FUNC
+  static inline bool isMuchSmallerThan(const bool& x, const bool&, const bool&)
+  {
+    return !x;
+  }
+  
+  EIGEN_DEVICE_FUNC
+  static inline bool isApprox(bool x, bool y, bool)
+  {
+    return x == y;
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline bool isApproxOrLessThan(const bool& x, const bool& y, const bool&)
+  {
+    return (!x) || y;
+  }
+  
+};
+
+  
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATHFUNCTIONS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h
new file mode 100644
index 0000000..3c9ef22
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MathFunctionsImpl.h
@@ -0,0 +1,78 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+// Copyright (C) 2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATHFUNCTIONSIMPL_H
+#define EIGEN_MATHFUNCTIONSIMPL_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal \returns the hyperbolic tan of \a a (coeff-wise)
+    Doesn't do anything fancy, just a 13/6-degree rational interpolant which
+    is accurate up to a couple of ulp in the range [-9, 9], outside of which
+    the tanh(x) = +/-1.
+
+    This implementation works on both scalars and packets.
+*/
+template<typename T>
+T generic_fast_tanh_float(const T& a_x)
+{
+  // Clamp the inputs to the range [-9, 9] since anything outside
+  // this range is +/-1.0f in single-precision.
+  const T plus_9 = pset1<T>(9.f);
+  const T minus_9 = pset1<T>(-9.f);
+  // NOTE GCC prior to 6.3 might improperly optimize this max/min
+  //      step such that if a_x is nan, x will be either 9 or -9,
+  //      and tanh will return 1 or -1 instead of nan.
+  //      This is supposed to be fixed in gcc6.3,
+  //      see: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=72867
+  const T x = pmax(minus_9,pmin(plus_9,a_x));
+  // The monomial coefficients of the numerator polynomial (odd).
+  const T alpha_1 = pset1<T>(4.89352455891786e-03f);
+  const T alpha_3 = pset1<T>(6.37261928875436e-04f);
+  const T alpha_5 = pset1<T>(1.48572235717979e-05f);
+  const T alpha_7 = pset1<T>(5.12229709037114e-08f);
+  const T alpha_9 = pset1<T>(-8.60467152213735e-11f);
+  const T alpha_11 = pset1<T>(2.00018790482477e-13f);
+  const T alpha_13 = pset1<T>(-2.76076847742355e-16f);
+
+  // The monomial coefficients of the denominator polynomial (even).
+  const T beta_0 = pset1<T>(4.89352518554385e-03f);
+  const T beta_2 = pset1<T>(2.26843463243900e-03f);
+  const T beta_4 = pset1<T>(1.18534705686654e-04f);
+  const T beta_6 = pset1<T>(1.19825839466702e-06f);
+
+  // Since the polynomials are odd/even, we need x^2.
+  const T x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial p.
+  T p = pmadd(x2, alpha_13, alpha_11);
+  p = pmadd(x2, p, alpha_9);
+  p = pmadd(x2, p, alpha_7);
+  p = pmadd(x2, p, alpha_5);
+  p = pmadd(x2, p, alpha_3);
+  p = pmadd(x2, p, alpha_1);
+  p = pmul(x, p);
+
+  // Evaluate the denominator polynomial p.
+  T q = pmadd(x2, beta_6, beta_4);
+  q = pmadd(x2, q, beta_2);
+  q = pmadd(x2, q, beta_0);
+
+  // Divide the numerator by the denominator.
+  return pdiv(p, q);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATHFUNCTIONSIMPL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Matrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Matrix.h
new file mode 100644
index 0000000..90c336d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Matrix.h
@@ -0,0 +1,461 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIX_H
+#define EIGEN_MATRIX_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct traits<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+private:
+  enum { size = internal::size_at_compile_time<_Rows,_Cols>::ret };
+  typedef typename find_best_packet<_Scalar,size>::type PacketScalar;
+  enum {
+      row_major_bit = _Options&RowMajor ? RowMajorBit : 0,
+      is_dynamic_size_storage = _MaxRows==Dynamic || _MaxCols==Dynamic,
+      max_size = is_dynamic_size_storage ? Dynamic : _MaxRows*_MaxCols,
+      default_alignment = compute_default_alignment<_Scalar,max_size>::value,
+      actual_alignment = ((_Options&DontAlign)==0) ? default_alignment : 0,
+      required_alignment = unpacket_traits<PacketScalar>::alignment,
+      packet_access_bit = (packet_traits<_Scalar>::Vectorizable && (EIGEN_UNALIGNED_VECTORIZE || (actual_alignment>=required_alignment))) ? PacketAccessBit : 0
+    };
+    
+public:
+  typedef _Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef Eigen::Index StorageIndex;
+  typedef MatrixXpr XprKind;
+  enum {
+    RowsAtCompileTime = _Rows,
+    ColsAtCompileTime = _Cols,
+    MaxRowsAtCompileTime = _MaxRows,
+    MaxColsAtCompileTime = _MaxCols,
+    Flags = compute_matrix_flags<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>::ret,
+    Options = _Options,
+    InnerStrideAtCompileTime = 1,
+    OuterStrideAtCompileTime = (Options&RowMajor) ? ColsAtCompileTime : RowsAtCompileTime,
+    
+    // FIXME, the following flag in only used to define NeedsToAlign in PlainObjectBase
+    EvaluatorFlags = LinearAccessBit | DirectAccessBit | packet_access_bit | row_major_bit,
+    Alignment = actual_alignment
+  };
+};
+}
+
+/** \class Matrix
+  * \ingroup Core_Module
+  *
+  * \brief The matrix class, also used for vectors and row-vectors
+  *
+  * The %Matrix class is the work-horse for all \em dense (\ref dense "note") matrices and vectors within Eigen.
+  * Vectors are matrices with one column, and row-vectors are matrices with one row.
+  *
+  * The %Matrix class encompasses \em both fixed-size and dynamic-size objects (\ref fixedsize "note").
+  *
+  * The first three template parameters are required:
+  * \tparam _Scalar Numeric type, e.g. float, double, int or std::complex<float>.
+  *                 User defined scalar types are supported as well (see \ref user_defined_scalars "here").
+  * \tparam _Rows Number of rows, or \b Dynamic
+  * \tparam _Cols Number of columns, or \b Dynamic
+  *
+  * The remaining template parameters are optional -- in most cases you don't have to worry about them.
+  * \tparam _Options A combination of either \b #RowMajor or \b #ColMajor, and of either
+  *                 \b #AutoAlign or \b #DontAlign.
+  *                 The former controls \ref TopicStorageOrders "storage order", and defaults to column-major. The latter controls alignment, which is required
+  *                 for vectorization. It defaults to aligning matrices except for fixed sizes that aren't a multiple of the packet size.
+  * \tparam _MaxRows Maximum number of rows. Defaults to \a _Rows (\ref maxrows "note").
+  * \tparam _MaxCols Maximum number of columns. Defaults to \a _Cols (\ref maxrows "note").
+  *
+  * Eigen provides a number of typedefs covering the usual cases. Here are some examples:
+  *
+  * \li \c Matrix2d is a 2x2 square matrix of doubles (\c Matrix<double, 2, 2>)
+  * \li \c Vector4f is a vector of 4 floats (\c Matrix<float, 4, 1>)
+  * \li \c RowVector3i is a row-vector of 3 ints (\c Matrix<int, 1, 3>)
+  *
+  * \li \c MatrixXf is a dynamic-size matrix of floats (\c Matrix<float, Dynamic, Dynamic>)
+  * \li \c VectorXf is a dynamic-size vector of floats (\c Matrix<float, Dynamic, 1>)
+  *
+  * \li \c Matrix2Xf is a partially fixed-size (dynamic-size) matrix of floats (\c Matrix<float, 2, Dynamic>)
+  * \li \c MatrixX3d is a partially dynamic-size (fixed-size) matrix of double (\c Matrix<double, Dynamic, 3>)
+  *
+  * See \link matrixtypedefs this page \endlink for a complete list of predefined \em %Matrix and \em Vector typedefs.
+  *
+  * You can access elements of vectors and matrices using normal subscripting:
+  *
+  * \code
+  * Eigen::VectorXd v(10);
+  * v[0] = 0.1;
+  * v[1] = 0.2;
+  * v(0) = 0.3;
+  * v(1) = 0.4;
+  *
+  * Eigen::MatrixXi m(10, 10);
+  * m(0, 1) = 1;
+  * m(0, 2) = 2;
+  * m(0, 3) = 3;
+  * \endcode
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIX_PLUGIN.
+  *
+  * <i><b>Some notes:</b></i>
+  *
+  * <dl>
+  * <dt><b>\anchor dense Dense versus sparse:</b></dt>
+  * <dd>This %Matrix class handles dense, not sparse matrices and vectors. For sparse matrices and vectors, see the Sparse module.
+  *
+  * Dense matrices and vectors are plain usual arrays of coefficients. All the coefficients are stored, in an ordinary contiguous array.
+  * This is unlike Sparse matrices and vectors where the coefficients are stored as a list of nonzero coefficients.</dd>
+  *
+  * <dt><b>\anchor fixedsize Fixed-size versus dynamic-size:</b></dt>
+  * <dd>Fixed-size means that the numbers of rows and columns are known are compile-time. In this case, Eigen allocates the array
+  * of coefficients as a fixed-size array, as a class member. This makes sense for very small matrices, typically up to 4x4, sometimes up
+  * to 16x16. Larger matrices should be declared as dynamic-size even if one happens to know their size at compile-time.
+  *
+  * Dynamic-size means that the numbers of rows or columns are not necessarily known at compile-time. In this case they are runtime
+  * variables, and the array of coefficients is allocated dynamically on the heap.
+  *
+  * Note that \em dense matrices, be they Fixed-size or Dynamic-size, <em>do not</em> expand dynamically in the sense of a std::map.
+  * If you want this behavior, see the Sparse module.</dd>
+  *
+  * <dt><b>\anchor maxrows _MaxRows and _MaxCols:</b></dt>
+  * <dd>In most cases, one just leaves these parameters to the default values.
+  * These parameters mean the maximum size of rows and columns that the matrix may have. They are useful in cases
+  * when the exact numbers of rows and columns are not known are compile-time, but it is known at compile-time that they cannot
+  * exceed a certain value. This happens when taking dynamic-size blocks inside fixed-size matrices: in this case _MaxRows and _MaxCols
+  * are the dimensions of the original matrix, while _Rows and _Cols are Dynamic.</dd>
+  * </dl>
+  *
+  * <i><b>ABI and storage layout</b></i>
+  *
+  * The table below summarizes the ABI of some possible Matrix instances which is fixed thorough the lifetime of Eigen 3.
+  * <table  class="manual">
+  * <tr><th>Matrix type</th><th>Equivalent C structure</th></tr>
+  * <tr><td>\code Matrix<T,Dynamic,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code
+  * Matrix<T,Dynamic,1>
+  * Matrix<T,1,Dynamic> \endcode</td><td>\code
+  * struct {
+  *   T *data;                  // with (size_t(data)%EIGEN_MAX_ALIGN_BYTES)==0
+  *   Eigen::Index size;
+  *  };
+  * \endcode</td></tr>
+  * <tr><td>\code Matrix<T,Rows,Cols> \endcode</td><td>\code
+  * struct {
+  *   T data[Rows*Cols];        // with (size_t(data)%A(Rows*Cols*sizeof(T)))==0
+  *  };
+  * \endcode</td></tr>
+  * <tr class="alt"><td>\code Matrix<T,Dynamic,Dynamic,0,MaxRows,MaxCols> \endcode</td><td>\code
+  * struct {
+  *   T data[MaxRows*MaxCols];  // with (size_t(data)%A(MaxRows*MaxCols*sizeof(T)))==0
+  *   Eigen::Index rows, cols;
+  *  };
+  * \endcode</td></tr>
+  * </table>
+  * Note that in this table Rows, Cols, MaxRows and MaxCols are all positive integers. A(S) is defined to the largest possible power-of-two
+  * smaller to EIGEN_MAX_STATIC_ALIGN_BYTES.
+  *
+  * \see MatrixBase for the majority of the API methods for matrices, \ref TopicClassHierarchy,
+  * \ref TopicStorageOrders
+  */
+
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+class Matrix
+  : public PlainObjectBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+{
+  public:
+
+    /** \brief Base class typedef.
+      * \sa PlainObjectBase
+      */
+    typedef PlainObjectBase<Matrix> Base;
+
+    enum { Options = _Options };
+
+    EIGEN_DENSE_PUBLIC_INTERFACE(Matrix)
+
+    typedef typename Base::PlainObject PlainObject;
+
+    using Base::base;
+    using Base::coeffRef;
+
+    /**
+      * \brief Assigns matrices to each other.
+      *
+      * \note This is a special case of the templated operator=. Its purpose is
+      * to prevent a default operator= from hiding the templated operator=.
+      *
+      * \callgraph
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const Matrix& other)
+    {
+      return Base::_set(other);
+    }
+
+    /** \internal
+      * \brief Copies the value of the expression \a other into \c *this with automatic resizing.
+      *
+      * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
+      * it will be initialized.
+      *
+      * Note that copying a row-vector into a vector (and conversely) is allowed.
+      * The resizing, if any, is then done in the appropriate way so that row-vectors
+      * remain row-vectors and vectors remain vectors.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const DenseBase<OtherDerived>& other)
+    {
+      return Base::_set(other);
+    }
+
+    /* Here, doxygen failed to copy the brief information when using \copydoc */
+
+    /**
+      * \brief Copies the generic expression \a other into *this.
+      * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const EigenBase<OtherDerived> &other)
+    {
+      return Base::operator=(other);
+    }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix& operator=(const ReturnByValue<OtherDerived>& func)
+    {
+      return Base::operator=(func);
+    }
+
+    /** \brief Default constructor.
+      *
+      * For fixed-size matrices, does nothing.
+      *
+      * For dynamic-size matrices, creates an empty matrix of size 0. Does not allocate any array. Such a matrix
+      * is called a null matrix. This constructor is the unique way to create null matrices: resizing
+      * a matrix to 0 is not supported.
+      *
+      * \sa resize(Index,Index)
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix() : Base()
+    {
+      Base::_check_template_params();
+      EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+
+    // FIXME is it still needed
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(internal::constructor_without_unaligned_array_assert)
+      : Base(internal::constructor_without_unaligned_array_assert())
+    { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED }
+
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    Matrix(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_constructible<Scalar>::value)
+      : Base(std::move(other))
+    {
+      Base::_check_template_params();
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
+    }
+    EIGEN_DEVICE_FUNC
+    Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+
+    // This constructor is for both 1x1 matrices and dynamic vectors
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE explicit Matrix(const T& x)
+    {
+      Base::_check_template_params();
+      Base::template _init1<T>(x);
+    }
+
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const T0& x, const T1& y)
+    {
+      Base::_check_template_params();
+      Base::template _init2<T0,T1>(x, y);
+    }
+    #else
+    /** \brief Constructs a fixed-sized matrix initialized with coefficients starting at \a data */
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(const Scalar *data);
+
+    /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
+      *
+      * This is useful for dynamic-size vectors. For fixed-size vectors,
+      * it is redundant to pass these parameters, so one should use the default constructor
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x1 matrices. For instance,
+      * calling Matrix<double,1,1>(1) will call the initialization constructor: Matrix(const Scalar&).
+      * For fixed-size \c 1x1 matrices it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_STRONG_INLINE explicit Matrix(Index dim);
+    /** \brief Constructs an initialized 1x1 matrix with the given coefficient */
+    Matrix(const Scalar& x);
+    /** \brief Constructs an uninitialized matrix with \a rows rows and \a cols columns.
+      *
+      * This is useful for dynamic-size matrices. For fixed-size matrices,
+      * it is redundant to pass these parameters, so one should use the default constructor
+      * Matrix() instead.
+      * 
+      * \warning This constructor is disabled for fixed-size \c 1x2 and \c 2x1 vectors. For instance,
+      * calling Matrix2f(2,1) will call the initialization constructor: Matrix(const Scalar& x, const Scalar& y).
+      * For fixed-size \c 1x2 or \c 2x1 vectors it is therefore recommended to use the default
+      * constructor Matrix() instead, especially when using one of the non standard
+      * \c EIGEN_INITIALIZE_MATRICES_BY_{ZERO,\c NAN} macros (see \ref TopicPreprocessorDirectives).
+      */
+    EIGEN_DEVICE_FUNC
+    Matrix(Index rows, Index cols);
+    
+    /** \brief Constructs an initialized 2D vector with given coefficients */
+    Matrix(const Scalar& x, const Scalar& y);
+    #endif
+
+    /** \brief Constructs an initialized 3D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z)
+    {
+      Base::_check_template_params();
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Matrix, 3)
+      m_storage.data()[0] = x;
+      m_storage.data()[1] = y;
+      m_storage.data()[2] = z;
+    }
+    /** \brief Constructs an initialized 4D vector with given coefficients */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const Scalar& x, const Scalar& y, const Scalar& z, const Scalar& w)
+    {
+      Base::_check_template_params();
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Matrix, 4)
+      m_storage.data()[0] = x;
+      m_storage.data()[1] = y;
+      m_storage.data()[2] = z;
+      m_storage.data()[3] = w;
+    }
+
+
+    /** \brief Copy constructor */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const Matrix& other) : Base(other)
+    { }
+
+    /** \brief Copy constructor for generic expressions.
+      * \sa MatrixBase::operator=(const EigenBase<OtherDerived>&)
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Matrix(const EigenBase<OtherDerived> &other)
+      : Base(other.derived())
+    { }
+
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return 1; }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return this->innerSize(); }
+
+    /////////// Geometry module ///////////
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    explicit Matrix(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Matrix& operator=(const RotationBase<OtherDerived,ColsAtCompileTime>& r);
+
+    // allow to extend Matrix outside Eigen
+    #ifdef EIGEN_MATRIX_PLUGIN
+    #include EIGEN_MATRIX_PLUGIN
+    #endif
+
+  protected:
+    template <typename Derived, typename OtherDerived, bool IsVector>
+    friend struct internal::conservative_resize_like_impl;
+
+    using Base::m_storage;
+};
+
+/** \defgroup matrixtypedefs Global matrix typedefs
+  *
+  * \ingroup Core_Module
+  *
+  * Eigen defines several typedef shortcuts for most common matrix and vector types.
+  *
+  * The general patterns are the following:
+  *
+  * \c MatrixSizeType where \c Size can be \c 2,\c 3,\c 4 for fixed size square matrices or \c X for dynamic size,
+  * and where \c Type can be \c i for integer, \c f for float, \c d for double, \c cf for complex float, \c cd
+  * for complex double.
+  *
+  * For example, \c Matrix3d is a fixed-size 3x3 matrix type of doubles, and \c MatrixXf is a dynamic-size matrix of floats.
+  *
+  * There are also \c VectorSizeType and \c RowVectorSizeType which are self-explanatory. For example, \c Vector4cf is
+  * a fixed-size vector of 4 complex floats.
+  *
+  * \sa class Matrix
+  */
+
+#define EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Size, SizeSuffix)   \
+/** \ingroup matrixtypedefs */                                    \
+typedef Matrix<Type, Size, Size> Matrix##SizeSuffix##TypeSuffix;  \
+/** \ingroup matrixtypedefs */                                    \
+typedef Matrix<Type, Size, 1>    Vector##SizeSuffix##TypeSuffix;  \
+/** \ingroup matrixtypedefs */                                    \
+typedef Matrix<Type, 1, Size>    RowVector##SizeSuffix##TypeSuffix;
+
+#define EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, Size)         \
+/** \ingroup matrixtypedefs */                                    \
+typedef Matrix<Type, Size, Dynamic> Matrix##Size##X##TypeSuffix;  \
+/** \ingroup matrixtypedefs */                                    \
+typedef Matrix<Type, Dynamic, Size> Matrix##X##Size##TypeSuffix;
+
+#define EIGEN_MAKE_TYPEDEFS_ALL_SIZES(Type, TypeSuffix) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 2, 2) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 3, 3) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 4, 4) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Dynamic, X) \
+EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 2) \
+EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 3) \
+EIGEN_MAKE_FIXED_TYPEDEFS(Type, TypeSuffix, 4)
+
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(int,                  i)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(float,                f)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(double,               d)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<float>,  cf)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(std::complex<double>, cd)
+
+#undef EIGEN_MAKE_TYPEDEFS_ALL_SIZES
+#undef EIGEN_MAKE_TYPEDEFS
+#undef EIGEN_MAKE_FIXED_TYPEDEFS
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h
new file mode 100644
index 0000000..ce41218
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/MatrixBase.h
@@ -0,0 +1,530 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIXBASE_H
+#define EIGEN_MATRIXBASE_H
+
+namespace Eigen {
+
+/** \class MatrixBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for all dense matrices, vectors, and expressions
+  *
+  * This class is the base that is inherited by all matrix, vector, and related expression
+  * types. Most of the Eigen API is contained in this class, and its base classes. Other important
+  * classes for the Eigen API are Matrix, and VectorwiseOp.
+  *
+  * Note that some methods are defined in other modules such as the \ref LU_Module LU module
+  * for all functions related to matrix inversions.
+  *
+  * \tparam Derived is the derived type, e.g. a matrix type, or an expression, etc.
+  *
+  * When writing a function taking Eigen objects as argument, if you want your function
+  * to take as argument any matrix, vector, or expression, just let it take a
+  * MatrixBase argument. As an example, here is a function printFirstRow which, given
+  * a matrix, vector, or expression \a x, prints the first row of \a x.
+  *
+  * \code
+    template<typename Derived>
+    void printFirstRow(const Eigen::MatrixBase<Derived>& x)
+    {
+      cout << x.row(0) << endl;
+    }
+  * \endcode
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_MATRIXBASE_PLUGIN.
+  *
+  * \sa \blank \ref TopicClassHierarchy
+  */
+template<typename Derived> class MatrixBase
+  : public DenseBase<Derived>
+{
+  public:
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef MatrixBase StorageBaseType;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    typedef DenseBase<Derived> Base;
+    using Base::RowsAtCompileTime;
+    using Base::ColsAtCompileTime;
+    using Base::SizeAtCompileTime;
+    using Base::MaxRowsAtCompileTime;
+    using Base::MaxColsAtCompileTime;
+    using Base::MaxSizeAtCompileTime;
+    using Base::IsVectorAtCompileTime;
+    using Base::Flags;
+
+    using Base::derived;
+    using Base::const_cast_derived;
+    using Base::rows;
+    using Base::cols;
+    using Base::size;
+    using Base::coeff;
+    using Base::coeffRef;
+    using Base::lazyAssign;
+    using Base::eval;
+    using Base::operator+=;
+    using Base::operator-=;
+    using Base::operator*=;
+    using Base::operator/=;
+
+    typedef typename Base::CoeffReturnType CoeffReturnType;
+    typedef typename Base::ConstTransposeReturnType ConstTransposeReturnType;
+    typedef typename Base::RowXpr RowXpr;
+    typedef typename Base::ColXpr ColXpr;
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** type of the equivalent square matrix */
+    typedef Matrix<Scalar,EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime),
+                          EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime)> SquareMatrixType;
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    /** \returns the size of the main diagonal, which is min(rows(),cols()).
+      * \sa rows(), cols(), SizeAtCompileTime. */
+    EIGEN_DEVICE_FUNC
+    inline Index diagonalSize() const { return (numext::mini)(rows(),cols()); }
+
+    typedef typename Base::PlainObject PlainObject;
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal Represents a matrix with all coefficients equal to one another*/
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,PlainObject> ConstantReturnType;
+    /** \internal the return type of MatrixBase::adjoint() */
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                        CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
+                        ConstTransposeReturnType
+                     >::type AdjointReturnType;
+    /** \internal Return type of eigenvalues() */
+    typedef Matrix<std::complex<RealScalar>, internal::traits<Derived>::ColsAtCompileTime, 1, ColMajor> EigenvaluesReturnType;
+    /** \internal the return type of identity */
+    typedef CwiseNullaryOp<internal::scalar_identity_op<Scalar>,PlainObject> IdentityReturnType;
+    /** \internal the return type of unit vectors */
+    typedef Block<const CwiseNullaryOp<internal::scalar_identity_op<Scalar>, SquareMatrixType>,
+                  internal::traits<Derived>::RowsAtCompileTime,
+                  internal::traits<Derived>::ColsAtCompileTime> BasisReturnType;
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::MatrixBase
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
+#   include "../plugins/CommonCwiseUnaryOps.h"
+#   include "../plugins/CommonCwiseBinaryOps.h"
+#   include "../plugins/MatrixCwiseUnaryOps.h"
+#   include "../plugins/MatrixCwiseBinaryOps.h"
+#   ifdef EIGEN_MATRIXBASE_PLUGIN
+#     include EIGEN_MATRIXBASE_PLUGIN
+#   endif
+#undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
+
+    /** Special case of the template operator=, in order to prevent the compiler
+      * from generating a default operator= (issue hit with g++ 4.1)
+      */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const MatrixBase& other);
+
+    // We cannot inherit here via Base::operator= since it is causing
+    // trouble with MSVC.
+
+    template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator=(const DenseBase<OtherDerived>& other);
+
+    template <typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const EigenBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    Derived& operator=(const ReturnByValue<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator+=(const MatrixBase<OtherDerived>& other);
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+    Derived& operator-=(const MatrixBase<OtherDerived>& other);
+
+#ifdef __CUDACC__
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,OtherDerived,LazyProduct>
+    operator*(const MatrixBase<OtherDerived> &other) const
+    { return this->lazyProduct(other); }
+#else
+
+    template<typename OtherDerived>
+    const Product<Derived,OtherDerived>
+    operator*(const MatrixBase<OtherDerived> &other) const;
+
+#endif
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived,OtherDerived,LazyProduct>
+    lazyProduct(const MatrixBase<OtherDerived> &other) const;
+
+    template<typename OtherDerived>
+    Derived& operator*=(const EigenBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    void applyOnTheLeft(const EigenBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    void applyOnTheRight(const EigenBase<OtherDerived>& other);
+
+    template<typename DiagonalDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<Derived, DiagonalDerived, LazyProduct>
+    operator*(const DiagonalBase<DiagonalDerived> &diagonal) const;
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
+    dot(const MatrixBase<OtherDerived>& other) const;
+
+    EIGEN_DEVICE_FUNC RealScalar squaredNorm() const;
+    EIGEN_DEVICE_FUNC RealScalar norm() const;
+    RealScalar stableNorm() const;
+    RealScalar blueNorm() const;
+    RealScalar hypotNorm() const;
+    EIGEN_DEVICE_FUNC const PlainObject normalized() const;
+    EIGEN_DEVICE_FUNC const PlainObject stableNormalized() const;
+    EIGEN_DEVICE_FUNC void normalize();
+    EIGEN_DEVICE_FUNC void stableNormalize();
+
+    EIGEN_DEVICE_FUNC const AdjointReturnType adjoint() const;
+    EIGEN_DEVICE_FUNC void adjointInPlace();
+
+    typedef Diagonal<Derived> DiagonalReturnType;
+    EIGEN_DEVICE_FUNC
+    DiagonalReturnType diagonal();
+
+    typedef typename internal::add_const<Diagonal<const Derived> >::type ConstDiagonalReturnType;
+    EIGEN_DEVICE_FUNC
+    ConstDiagonalReturnType diagonal() const;
+
+    template<int Index> struct DiagonalIndexReturnType { typedef Diagonal<Derived,Index> Type; };
+    template<int Index> struct ConstDiagonalIndexReturnType { typedef const Diagonal<const Derived,Index> Type; };
+
+    template<int Index>
+    EIGEN_DEVICE_FUNC
+    typename DiagonalIndexReturnType<Index>::Type diagonal();
+
+    template<int Index>
+    EIGEN_DEVICE_FUNC
+    typename ConstDiagonalIndexReturnType<Index>::Type diagonal() const;
+
+    typedef Diagonal<Derived,DynamicIndex> DiagonalDynamicIndexReturnType;
+    typedef typename internal::add_const<Diagonal<const Derived,DynamicIndex> >::type ConstDiagonalDynamicIndexReturnType;
+
+    EIGEN_DEVICE_FUNC
+    DiagonalDynamicIndexReturnType diagonal(Index index);
+    EIGEN_DEVICE_FUNC
+    ConstDiagonalDynamicIndexReturnType diagonal(Index index) const;
+
+    template<unsigned int Mode> struct TriangularViewReturnType { typedef TriangularView<Derived, Mode> Type; };
+    template<unsigned int Mode> struct ConstTriangularViewReturnType { typedef const TriangularView<const Derived, Mode> Type; };
+
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename TriangularViewReturnType<Mode>::Type triangularView();
+    template<unsigned int Mode>
+    EIGEN_DEVICE_FUNC
+    typename ConstTriangularViewReturnType<Mode>::Type triangularView() const;
+
+    template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SelfAdjointView<Derived, UpLo> Type; };
+    template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SelfAdjointView<const Derived, UpLo> Type; };
+
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
+    template<unsigned int UpLo>
+    EIGEN_DEVICE_FUNC
+    typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
+
+    const SparseView<Derived> sparseView(const Scalar& m_reference = Scalar(0),
+                                         const typename NumTraits<Scalar>::Real& m_epsilon = NumTraits<Scalar>::dummy_precision()) const;
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity();
+    EIGEN_DEVICE_FUNC static const IdentityReturnType Identity(Index rows, Index cols);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index size, Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType Unit(Index i);
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitX();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitY();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitZ();
+    EIGEN_DEVICE_FUNC static const BasisReturnType UnitW();
+
+    EIGEN_DEVICE_FUNC
+    const DiagonalWrapper<const Derived> asDiagonal() const;
+    const PermutationWrapper<const Derived> asPermutation() const;
+
+    EIGEN_DEVICE_FUNC
+    Derived& setIdentity();
+    EIGEN_DEVICE_FUNC
+    Derived& setIdentity(Index rows, Index cols);
+
+    bool isIdentity(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    bool isDiagonal(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+
+    bool isUpperTriangular(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    bool isLowerTriangular(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+
+    template<typename OtherDerived>
+    bool isOrthogonal(const MatrixBase<OtherDerived>& other,
+                      const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+    bool isUnitary(const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+
+    /** \returns true if each coefficients of \c *this and \a other are all exactly equal.
+      * \warning When using floating point scalar values you probably should rather use a
+      *          fuzzy comparison such as isApprox()
+      * \sa isApprox(), operator!= */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC inline bool operator==(const MatrixBase<OtherDerived>& other) const
+    { return cwiseEqual(other).all(); }
+
+    /** \returns true if at least one pair of coefficients of \c *this and \a other are not exactly equal to each other.
+      * \warning When using floating point scalar values you probably should rather use a
+      *          fuzzy comparison such as isApprox()
+      * \sa isApprox(), operator== */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC inline bool operator!=(const MatrixBase<OtherDerived>& other) const
+    { return cwiseNotEqual(other).any(); }
+
+    NoAlias<Derived,Eigen::MatrixBase > noalias();
+
+    // TODO forceAlignedAccess is temporarily disabled
+    // Need to find a nicer workaround.
+    inline const Derived& forceAlignedAccess() const { return derived(); }
+    inline Derived& forceAlignedAccess() { return derived(); }
+    template<bool Enable> inline const Derived& forceAlignedAccessIf() const { return derived(); }
+    template<bool Enable> inline Derived& forceAlignedAccessIf() { return derived(); }
+
+    EIGEN_DEVICE_FUNC Scalar trace() const;
+
+    template<int p> EIGEN_DEVICE_FUNC RealScalar lpNorm() const;
+
+    EIGEN_DEVICE_FUNC MatrixBase<Derived>& matrix() { return *this; }
+    EIGEN_DEVICE_FUNC const MatrixBase<Derived>& matrix() const { return *this; }
+
+    /** \returns an \link Eigen::ArrayBase Array \endlink expression of this matrix
+      * \sa ArrayBase::matrix() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE ArrayWrapper<Derived> array() { return ArrayWrapper<Derived>(derived()); }
+    /** \returns a const \link Eigen::ArrayBase Array \endlink expression of this matrix
+      * \sa ArrayBase::matrix() */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const ArrayWrapper<const Derived> array() const { return ArrayWrapper<const Derived>(derived()); }
+
+/////////// LU module ///////////
+
+    inline const FullPivLU<PlainObject> fullPivLu() const;
+    inline const PartialPivLU<PlainObject> partialPivLu() const;
+
+    inline const PartialPivLU<PlainObject> lu() const;
+
+    inline const Inverse<Derived> inverse() const;
+
+    template<typename ResultType>
+    inline void computeInverseAndDetWithCheck(
+      ResultType& inverse,
+      typename ResultType::Scalar& determinant,
+      bool& invertible,
+      const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
+    ) const;
+    template<typename ResultType>
+    inline void computeInverseWithCheck(
+      ResultType& inverse,
+      bool& invertible,
+      const RealScalar& absDeterminantThreshold = NumTraits<Scalar>::dummy_precision()
+    ) const;
+    Scalar determinant() const;
+
+/////////// Cholesky module ///////////
+
+    inline const LLT<PlainObject>  llt() const;
+    inline const LDLT<PlainObject> ldlt() const;
+
+/////////// QR module ///////////
+
+    inline const HouseholderQR<PlainObject> householderQr() const;
+    inline const ColPivHouseholderQR<PlainObject> colPivHouseholderQr() const;
+    inline const FullPivHouseholderQR<PlainObject> fullPivHouseholderQr() const;
+    inline const CompleteOrthogonalDecomposition<PlainObject> completeOrthogonalDecomposition() const;
+
+/////////// Eigenvalues module ///////////
+
+    inline EigenvaluesReturnType eigenvalues() const;
+    inline RealScalar operatorNorm() const;
+
+/////////// SVD module ///////////
+
+    inline JacobiSVD<PlainObject> jacobiSvd(unsigned int computationOptions = 0) const;
+    inline BDCSVD<PlainObject>    bdcSvd(unsigned int computationOptions = 0) const;
+
+/////////// Geometry module ///////////
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /// \internal helper struct to form the return type of the cross product
+    template<typename OtherDerived> struct cross_product_return_type {
+      typedef typename ScalarBinaryOpTraits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType Scalar;
+      typedef Matrix<Scalar,MatrixBase::RowsAtCompileTime,MatrixBase::ColsAtCompileTime> type;
+    };
+    #endif // EIGEN_PARSED_BY_DOXYGEN
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    inline typename cross_product_return_type<OtherDerived>::type
+#else
+    inline PlainObject
+#endif
+    cross(const MatrixBase<OtherDerived>& other) const;
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    inline PlainObject cross3(const MatrixBase<OtherDerived>& other) const;
+
+    EIGEN_DEVICE_FUNC
+    inline PlainObject unitOrthogonal(void) const;
+
+    EIGEN_DEVICE_FUNC
+    inline Matrix<Scalar,3,1> eulerAngles(Index a0, Index a1, Index a2) const;
+
+    // put this as separate enum value to work around possible GCC 4.3 bug (?)
+    enum { HomogeneousReturnTypeDirection = ColsAtCompileTime==1&&RowsAtCompileTime==1 ? ((internal::traits<Derived>::Flags&RowMajorBit)==RowMajorBit ? Horizontal : Vertical)
+                                          : ColsAtCompileTime==1 ? Vertical : Horizontal };
+    typedef Homogeneous<Derived, HomogeneousReturnTypeDirection> HomogeneousReturnType;
+    EIGEN_DEVICE_FUNC
+    inline HomogeneousReturnType homogeneous() const;
+
+    enum {
+      SizeMinusOne = SizeAtCompileTime==Dynamic ? Dynamic : SizeAtCompileTime-1
+    };
+    typedef Block<const Derived,
+                  internal::traits<Derived>::ColsAtCompileTime==1 ? SizeMinusOne : 1,
+                  internal::traits<Derived>::ColsAtCompileTime==1 ? 1 : SizeMinusOne> ConstStartMinusOne;
+    typedef EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(ConstStartMinusOne,Scalar,quotient) HNormalizedReturnType;
+    EIGEN_DEVICE_FUNC
+    inline const HNormalizedReturnType hnormalized() const;
+
+////////// Householder module ///////////
+
+    void makeHouseholderInPlace(Scalar& tau, RealScalar& beta);
+    template<typename EssentialPart>
+    void makeHouseholder(EssentialPart& essential,
+                         Scalar& tau, RealScalar& beta) const;
+    template<typename EssentialPart>
+    void applyHouseholderOnTheLeft(const EssentialPart& essential,
+                                   const Scalar& tau,
+                                   Scalar* workspace);
+    template<typename EssentialPart>
+    void applyHouseholderOnTheRight(const EssentialPart& essential,
+                                    const Scalar& tau,
+                                    Scalar* workspace);
+
+///////// Jacobi module /////////
+
+    template<typename OtherScalar>
+    void applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j);
+    template<typename OtherScalar>
+    void applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j);
+
+///////// SparseCore module /////////
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE const typename SparseMatrixBase<OtherDerived>::template CwiseProductDenseReturnType<Derived>::Type
+    cwiseProduct(const SparseMatrixBase<OtherDerived> &other) const
+    {
+      return other.cwiseProduct(derived());
+    }
+
+///////// MatrixFunctions module /////////
+
+    typedef typename internal::stem_function<Scalar>::type StemFunction;
+    const MatrixExponentialReturnValue<Derived> exp() const;
+    const MatrixFunctionReturnValue<Derived> matrixFunction(StemFunction f) const;
+    const MatrixFunctionReturnValue<Derived> cosh() const;
+    const MatrixFunctionReturnValue<Derived> sinh() const;
+    const MatrixFunctionReturnValue<Derived> cos() const;
+    const MatrixFunctionReturnValue<Derived> sin() const;
+    const MatrixSquareRootReturnValue<Derived> sqrt() const;
+    const MatrixLogarithmReturnValue<Derived> log() const;
+    const MatrixPowerReturnValue<Derived> pow(const RealScalar& p) const;
+    const MatrixComplexPowerReturnValue<Derived> pow(const std::complex<RealScalar>& p) const;
+
+  protected:
+    EIGEN_DEVICE_FUNC MatrixBase() : Base() {}
+
+  private:
+    EIGEN_DEVICE_FUNC explicit MatrixBase(int);
+    EIGEN_DEVICE_FUNC MatrixBase(int,int);
+    template<typename OtherDerived> EIGEN_DEVICE_FUNC explicit MatrixBase(const MatrixBase<OtherDerived>&);
+  protected:
+    // mixing arrays and matrices is not legal
+    template<typename OtherDerived> Derived& operator+=(const ArrayBase<OtherDerived>& )
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
+    // mixing arrays and matrices is not legal
+    template<typename OtherDerived> Derived& operator-=(const ArrayBase<OtherDerived>& )
+    {EIGEN_STATIC_ASSERT(std::ptrdiff_t(sizeof(typename OtherDerived::Scalar))==-1,YOU_CANNOT_MIX_ARRAYS_AND_MATRICES); return *this;}
+};
+
+
+/***************************************************************************
+* Implementation of matrix base methods
+***************************************************************************/
+
+/** replaces \c *this by \c *this * \a other.
+  *
+  * \returns a reference to \c *this
+  *
+  * Example: \include MatrixBase_applyOnTheRight.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheRight.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline Derived&
+MatrixBase<Derived>::operator*=(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheRight(derived());
+  return derived();
+}
+
+/** replaces \c *this by \c *this * \a other. It is equivalent to MatrixBase::operator*=().
+  *
+  * Example: \include MatrixBase_applyOnTheRight.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheRight.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline void MatrixBase<Derived>::applyOnTheRight(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheRight(derived());
+}
+
+/** replaces \c *this by \a other * \c *this.
+  *
+  * Example: \include MatrixBase_applyOnTheLeft.cpp
+  * Output: \verbinclude MatrixBase_applyOnTheLeft.out
+  */
+template<typename Derived>
+template<typename OtherDerived>
+inline void MatrixBase<Derived>::applyOnTheLeft(const EigenBase<OtherDerived> &other)
+{
+  other.derived().applyThisOnTheLeft(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATRIXBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h
new file mode 100644
index 0000000..13adf07
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NestByValue.h
@@ -0,0 +1,110 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NESTBYVALUE_H
+#define EIGEN_NESTBYVALUE_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename ExpressionType>
+struct traits<NestByValue<ExpressionType> > : public traits<ExpressionType>
+{};
+}
+
+/** \class NestByValue
+  * \ingroup Core_Module
+  *
+  * \brief Expression which must be nested by value
+  *
+  * \tparam ExpressionType the type of the object of which we are requiring nesting-by-value
+  *
+  * This class is the return type of MatrixBase::nestByValue()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::nestByValue()
+  */
+template<typename ExpressionType> class NestByValue
+  : public internal::dense_xpr_base< NestByValue<ExpressionType> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<NestByValue>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(NestByValue)
+
+    EIGEN_DEVICE_FUNC explicit inline NestByValue(const ExpressionType& matrix) : m_expression(matrix) {}
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_expression.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_expression.cols(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return m_expression.outerStride(); }
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return m_expression.innerStride(); }
+
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index row, Index col) const
+    {
+      return m_expression.coeff(row, col);
+    }
+
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_expression.const_cast_derived().coeffRef(row, col);
+    }
+
+    EIGEN_DEVICE_FUNC inline const CoeffReturnType coeff(Index index) const
+    {
+      return m_expression.coeff(index);
+    }
+
+    EIGEN_DEVICE_FUNC inline Scalar& coeffRef(Index index)
+    {
+      return m_expression.const_cast_derived().coeffRef(index);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index row, Index col) const
+    {
+      return m_expression.template packet<LoadMode>(row, col);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index row, Index col, const PacketScalar& x)
+    {
+      m_expression.const_cast_derived().template writePacket<LoadMode>(row, col, x);
+    }
+
+    template<int LoadMode>
+    inline const PacketScalar packet(Index index) const
+    {
+      return m_expression.template packet<LoadMode>(index);
+    }
+
+    template<int LoadMode>
+    inline void writePacket(Index index, const PacketScalar& x)
+    {
+      m_expression.const_cast_derived().template writePacket<LoadMode>(index, x);
+    }
+
+    EIGEN_DEVICE_FUNC operator const ExpressionType&() const { return m_expression; }
+
+  protected:
+    const ExpressionType m_expression;
+};
+
+/** \returns an expression of the temporary version of *this.
+  */
+template<typename Derived>
+inline const NestByValue<Derived>
+DenseBase<Derived>::nestByValue() const
+{
+  return NestByValue<Derived>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_NESTBYVALUE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h
new file mode 100644
index 0000000..3390801
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NoAlias.h
@@ -0,0 +1,108 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NOALIAS_H
+#define EIGEN_NOALIAS_H
+
+namespace Eigen {
+
+/** \class NoAlias
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression providing an operator = assuming no aliasing
+  *
+  * \tparam ExpressionType the type of the object on which to do the lazy assignment
+  *
+  * This class represents an expression with special assignment operators
+  * assuming no aliasing between the target expression and the source expression.
+  * More precisely it alloas to bypass the EvalBeforeAssignBit flag of the source expression.
+  * It is the return type of MatrixBase::noalias()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::noalias()
+  */
+template<typename ExpressionType, template <typename> class StorageBase>
+class NoAlias
+{
+  public:
+    typedef typename ExpressionType::Scalar Scalar;
+    
+    explicit NoAlias(ExpressionType& expression) : m_expression(expression) {}
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator+=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+    
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE ExpressionType& operator-=(const StorageBase<OtherDerived>& other)
+    {
+      call_assignment_no_alias(m_expression, other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+      return m_expression;
+    }
+
+    EIGEN_DEVICE_FUNC
+    ExpressionType& expression() const
+    {
+      return m_expression;
+    }
+
+  protected:
+    ExpressionType& m_expression;
+};
+
+/** \returns a pseudo expression of \c *this with an operator= assuming
+  * no aliasing between \c *this and the source expression.
+  *
+  * More precisely, noalias() allows to bypass the EvalBeforeAssignBit flag.
+  * Currently, even though several expressions may alias, only product
+  * expressions have this flag. Therefore, noalias() is only usefull when
+  * the source expression contains a matrix product.
+  *
+  * Here are some examples where noalias is usefull:
+  * \code
+  * D.noalias()  = A * B;
+  * D.noalias() += A.transpose() * B;
+  * D.noalias() -= 2 * A * B.adjoint();
+  * \endcode
+  *
+  * On the other hand the following example will lead to a \b wrong result:
+  * \code
+  * A.noalias() = A * B;
+  * \endcode
+  * because the result matrix A is also an operand of the matrix product. Therefore,
+  * there is no alternative than evaluating A * B in a temporary, that is the default
+  * behavior when you write:
+  * \code
+  * A = A * B;
+  * \endcode
+  *
+  * \sa class NoAlias
+  */
+template<typename Derived>
+NoAlias<Derived,MatrixBase> MatrixBase<Derived>::noalias()
+{
+  return NoAlias<Derived, Eigen::MatrixBase >(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_NOALIAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h
new file mode 100644
index 0000000..daf4898
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/NumTraits.h
@@ -0,0 +1,248 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NUMTRAITS_H
+#define EIGEN_NUMTRAITS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// default implementation of digits10(), based on numeric_limits if specialized,
+// 0 for integer types, and log10(epsilon()) otherwise.
+template< typename T,
+          bool use_numeric_limits = std::numeric_limits<T>::is_specialized,
+          bool is_integer = NumTraits<T>::IsInteger>
+struct default_digits10_impl
+{
+  static int run() { return std::numeric_limits<T>::digits10; }
+};
+
+template<typename T>
+struct default_digits10_impl<T,false,false> // Floating point
+{
+  static int run() {
+    using std::log10;
+    using std::ceil;
+    typedef typename NumTraits<T>::Real Real;
+    return int(ceil(-log10(NumTraits<Real>::epsilon())));
+  }
+};
+
+template<typename T>
+struct default_digits10_impl<T,false,true> // Integer
+{
+  static int run() { return 0; }
+};
+
+} // end namespace internal
+
+/** \class NumTraits
+  * \ingroup Core_Module
+  *
+  * \brief Holds information about the various numeric (i.e. scalar) types allowed by Eigen.
+  *
+  * \tparam T the numeric type at hand
+  *
+  * This class stores enums, typedefs and static methods giving information about a numeric type.
+  *
+  * The provided data consists of:
+  * \li A typedef \c Real, giving the "real part" type of \a T. If \a T is already real,
+  *     then \c Real is just a typedef to \a T. If \a T is \c std::complex<U> then \c Real
+  *     is a typedef to \a U.
+  * \li A typedef \c NonInteger, giving the type that should be used for operations producing non-integral values,
+  *     such as quotients, square roots, etc. If \a T is a floating-point type, then this typedef just gives
+  *     \a T again. Note however that many Eigen functions such as internal::sqrt simply refuse to
+  *     take integers. Outside of a few cases, Eigen doesn't do automatic type promotion. Thus, this typedef is
+  *     only intended as a helper for code that needs to explicitly promote types.
+  * \li A typedef \c Literal giving the type to use for numeric literals such as "2" or "0.5". For instance, for \c std::complex<U>, Literal is defined as \c U.
+  *     Of course, this type must be fully compatible with \a T. In doubt, just use \a T here.
+  * \li A typedef \a Nested giving the type to use to nest a value inside of the expression tree. If you don't know what
+  *     this means, just use \a T here.
+  * \li An enum value \a IsComplex. It is equal to 1 if \a T is a \c std::complex
+  *     type, and to 0 otherwise.
+  * \li An enum value \a IsInteger. It is equal to \c 1 if \a T is an integer type such as \c int,
+  *     and to \c 0 otherwise.
+  * \li Enum values ReadCost, AddCost and MulCost representing a rough estimate of the number of CPU cycles needed
+  *     to by move / add / mul instructions respectively, assuming the data is already stored in CPU registers.
+  *     Stay vague here. No need to do architecture-specific stuff.
+  * \li An enum value \a IsSigned. It is equal to \c 1 if \a T is a signed type and to 0 if \a T is unsigned.
+  * \li An enum value \a RequireInitialization. It is equal to \c 1 if the constructor of the numeric type \a T must
+  *     be called, and to 0 if it is safe not to call it. Default is 0 if \a T is an arithmetic type, and 1 otherwise.
+  * \li An epsilon() function which, unlike <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/epsilon">std::numeric_limits::epsilon()</a>,
+  *     it returns a \a Real instead of a \a T.
+  * \li A dummy_precision() function returning a weak epsilon value. It is mainly used as a default
+  *     value by the fuzzy comparison operators.
+  * \li highest() and lowest() functions returning the highest and lowest possible values respectively.
+  * \li digits10() function returning the number of decimal digits that can be represented without change. This is
+  *     the analogue of <a href="http://en.cppreference.com/w/cpp/types/numeric_limits/digits10">std::numeric_limits<T>::digits10</a>
+  *     which is used as the default implementation if specialized.
+  */
+
+template<typename T> struct GenericNumTraits
+{
+  enum {
+    IsInteger = std::numeric_limits<T>::is_integer,
+    IsSigned = std::numeric_limits<T>::is_signed,
+    IsComplex = 0,
+    RequireInitialization = internal::is_arithmetic<T>::value ? 0 : 1,
+    ReadCost = 1,
+    AddCost = 1,
+    MulCost = 1
+  };
+
+  typedef T Real;
+  typedef typename internal::conditional<
+                     IsInteger,
+                     typename internal::conditional<sizeof(T)<=2, float, double>::type,
+                     T
+                   >::type NonInteger;
+  typedef T Nested;
+  typedef T Literal;
+
+  EIGEN_DEVICE_FUNC
+  static inline Real epsilon()
+  {
+    return numext::numeric_limits<T>::epsilon();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline int digits10()
+  {
+    return internal::default_digits10_impl<T>::run();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline Real dummy_precision()
+  {
+    // make sure to override this for floating-point types
+    return Real(0);
+  }
+
+
+  EIGEN_DEVICE_FUNC
+  static inline T highest() {
+    return (numext::numeric_limits<T>::max)();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T lowest()  {
+    return IsInteger ? (numext::numeric_limits<T>::min)() : (-(numext::numeric_limits<T>::max)());
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T infinity() {
+    return numext::numeric_limits<T>::infinity();
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline T quiet_NaN() {
+    return numext::numeric_limits<T>::quiet_NaN();
+  }
+};
+
+template<typename T> struct NumTraits : GenericNumTraits<T>
+{};
+
+template<> struct NumTraits<float>
+  : GenericNumTraits<float>
+{
+  EIGEN_DEVICE_FUNC
+  static inline float dummy_precision() { return 1e-5f; }
+};
+
+template<> struct NumTraits<double> : GenericNumTraits<double>
+{
+  EIGEN_DEVICE_FUNC
+  static inline double dummy_precision() { return 1e-12; }
+};
+
+template<> struct NumTraits<long double>
+  : GenericNumTraits<long double>
+{
+  static inline long double dummy_precision() { return 1e-15l; }
+};
+
+template<typename _Real> struct NumTraits<std::complex<_Real> >
+  : GenericNumTraits<std::complex<_Real> >
+{
+  typedef _Real Real;
+  typedef typename NumTraits<_Real>::Literal Literal;
+  enum {
+    IsComplex = 1,
+    RequireInitialization = NumTraits<_Real>::RequireInitialization,
+    ReadCost = 2 * NumTraits<_Real>::ReadCost,
+    AddCost = 2 * NumTraits<Real>::AddCost,
+    MulCost = 4 * NumTraits<Real>::MulCost + 2 * NumTraits<Real>::AddCost
+  };
+
+  EIGEN_DEVICE_FUNC
+  static inline Real epsilon() { return NumTraits<Real>::epsilon(); }
+  EIGEN_DEVICE_FUNC
+  static inline Real dummy_precision() { return NumTraits<Real>::dummy_precision(); }
+  EIGEN_DEVICE_FUNC
+  static inline int digits10() { return NumTraits<Real>::digits10(); }
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+struct NumTraits<Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> >
+{
+  typedef Array<Scalar, Rows, Cols, Options, MaxRows, MaxCols> ArrayType;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Array<RealScalar, Rows, Cols, Options, MaxRows, MaxCols> Real;
+  typedef typename NumTraits<Scalar>::NonInteger NonIntegerScalar;
+  typedef Array<NonIntegerScalar, Rows, Cols, Options, MaxRows, MaxCols> NonInteger;
+  typedef ArrayType & Nested;
+  typedef typename NumTraits<Scalar>::Literal Literal;
+
+  enum {
+    IsComplex = NumTraits<Scalar>::IsComplex,
+    IsInteger = NumTraits<Scalar>::IsInteger,
+    IsSigned  = NumTraits<Scalar>::IsSigned,
+    RequireInitialization = 1,
+    ReadCost = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::ReadCost,
+    AddCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::AddCost,
+    MulCost  = ArrayType::SizeAtCompileTime==Dynamic ? HugeCost : ArrayType::SizeAtCompileTime * NumTraits<Scalar>::MulCost
+  };
+
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar epsilon() { return NumTraits<RealScalar>::epsilon(); }
+  EIGEN_DEVICE_FUNC
+  static inline RealScalar dummy_precision() { return NumTraits<RealScalar>::dummy_precision(); }
+
+  static inline int digits10() { return NumTraits<Scalar>::digits10(); }
+};
+
+template<> struct NumTraits<std::string>
+  : GenericNumTraits<std::string>
+{
+  enum {
+    RequireInitialization = 1,
+    ReadCost = HugeCost,
+    AddCost  = HugeCost,
+    MulCost  = HugeCost
+  };
+
+  static inline int digits10() { return 0; }
+
+private:
+  static inline std::string epsilon();
+  static inline std::string dummy_precision();
+  static inline std::string lowest();
+  static inline std::string highest();
+  static inline std::string infinity();
+  static inline std::string quiet_NaN();
+};
+
+// Empty specialization for void to allow template specialization based on NumTraits<T>::Real with T==void and SFINAE.
+template<> struct NumTraits<void> {};
+
+} // end namespace Eigen
+
+#endif // EIGEN_NUMTRAITS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h
new file mode 100644
index 0000000..b1fb455
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PermutationMatrix.h
@@ -0,0 +1,633 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PERMUTATIONMATRIX_H
+#define EIGEN_PERMUTATIONMATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+enum PermPermProduct_t {PermPermProduct};
+
+} // end namespace internal
+
+/** \class PermutationBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for permutations
+  *
+  * \tparam Derived the derived class
+  *
+  * This class is the base class for all expressions representing a permutation matrix,
+  * internally stored as a vector of integers.
+  * The convention followed here is that if \f$ \sigma \f$ is a permutation, the corresponding permutation matrix
+  * \f$ P_\sigma \f$ is such that if \f$ (e_1,\ldots,e_p) \f$ is the canonical basis, we have:
+  *  \f[ P_\sigma(e_i) = e_{\sigma(i)}. \f]
+  * This convention ensures that for any two permutations \f$ \sigma, \tau \f$, we have:
+  *  \f[ P_{\sigma\circ\tau} = P_\sigma P_\tau. \f]
+  *
+  * Permutation matrices are square and invertible.
+  *
+  * Notice that in addition to the member functions and operators listed here, there also are non-member
+  * operator* to multiply any kind of permutation object with any kind of matrix expression (MatrixBase)
+  * on either side.
+  *
+  * \sa class PermutationMatrix, class PermutationWrapper
+  */
+template<typename Derived>
+class PermutationBase : public EigenBase<Derived>
+{
+    typedef internal::traits<Derived> Traits;
+    typedef EigenBase<Derived> Base;
+  public:
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename Traits::IndicesType IndicesType;
+    enum {
+      Flags = Traits::Flags,
+      RowsAtCompileTime = Traits::RowsAtCompileTime,
+      ColsAtCompileTime = Traits::ColsAtCompileTime,
+      MaxRowsAtCompileTime = Traits::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = Traits::MaxColsAtCompileTime
+    };
+    typedef typename Traits::StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,RowsAtCompileTime,ColsAtCompileTime,0,MaxRowsAtCompileTime,MaxColsAtCompileTime>
+            DenseMatrixType;
+    typedef PermutationMatrix<IndicesType::SizeAtCompileTime,IndicesType::MaxSizeAtCompileTime,StorageIndex>
+            PlainPermutationType;
+    typedef PlainPermutationType PlainObject;
+    using Base::derived;
+    typedef Inverse<Derived> InverseReturnType;
+    typedef void Scalar;
+    #endif
+
+    /** Copies the other permutation into *this */
+    template<typename OtherDerived>
+    Derived& operator=(const PermutationBase<OtherDerived>& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+
+    /** Assignment from the Transpositions \a tr */
+    template<typename OtherDerived>
+    Derived& operator=(const TranspositionsBase<OtherDerived>& tr)
+    {
+      setIdentity(tr.size());
+      for(Index k=size()-1; k>=0; --k)
+        applyTranspositionOnTheRight(k,tr.coeff(k));
+      return derived();
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Derived& operator=(const PermutationBase& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+    #endif
+
+    /** \returns the number of rows */
+    inline Index rows() const { return Index(indices().size()); }
+
+    /** \returns the number of columns */
+    inline Index cols() const { return Index(indices().size()); }
+
+    /** \returns the size of a side of the respective square matrix, i.e., the number of indices */
+    inline Index size() const { return Index(indices().size()); }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename DenseDerived>
+    void evalTo(MatrixBase<DenseDerived>& other) const
+    {
+      other.setZero();
+      for (Index i=0; i<rows(); ++i)
+        other.coeffRef(indices().coeff(i),i) = typename DenseDerived::Scalar(1);
+    }
+    #endif
+
+    /** \returns a Matrix object initialized from this permutation matrix. Notice that it
+      * is inefficient to return this Matrix object by value. For efficiency, favor using
+      * the Matrix constructor taking EigenBase objects.
+      */
+    DenseMatrixType toDenseMatrix() const
+    {
+      return derived();
+    }
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return derived().indices(); }
+    /** \returns a reference to the stored array representing the permutation. */
+    IndicesType& indices() { return derived().indices(); }
+
+    /** Resizes to given size.
+      */
+    inline void resize(Index newSize)
+    {
+      indices().resize(newSize);
+    }
+
+    /** Sets *this to be the identity permutation matrix */
+    void setIdentity()
+    {
+      StorageIndex n = StorageIndex(size());
+      for(StorageIndex i = 0; i < n; ++i)
+        indices().coeffRef(i) = i;
+    }
+
+    /** Sets *this to be the identity permutation matrix of given size.
+      */
+    void setIdentity(Index newSize)
+    {
+      resize(newSize);
+      setIdentity();
+    }
+
+    /** Multiplies *this by the transposition \f$(ij)\f$ on the left.
+      *
+      * \returns a reference to *this.
+      *
+      * \warning This is much slower than applyTranspositionOnTheRight(Index,Index):
+      * this has linear complexity and requires a lot of branching.
+      *
+      * \sa applyTranspositionOnTheRight(Index,Index)
+      */
+    Derived& applyTranspositionOnTheLeft(Index i, Index j)
+    {
+      eigen_assert(i>=0 && j>=0 && i<size() && j<size());
+      for(Index k = 0; k < size(); ++k)
+      {
+        if(indices().coeff(k) == i) indices().coeffRef(k) = StorageIndex(j);
+        else if(indices().coeff(k) == j) indices().coeffRef(k) = StorageIndex(i);
+      }
+      return derived();
+    }
+
+    /** Multiplies *this by the transposition \f$(ij)\f$ on the right.
+      *
+      * \returns a reference to *this.
+      *
+      * This is a fast operation, it only consists in swapping two indices.
+      *
+      * \sa applyTranspositionOnTheLeft(Index,Index)
+      */
+    Derived& applyTranspositionOnTheRight(Index i, Index j)
+    {
+      eigen_assert(i>=0 && j>=0 && i<size() && j<size());
+      std::swap(indices().coeffRef(i), indices().coeffRef(j));
+      return derived();
+    }
+
+    /** \returns the inverse permutation matrix.
+      *
+      * \note \blank \note_try_to_help_rvo
+      */
+    inline InverseReturnType inverse() const
+    { return InverseReturnType(derived()); }
+    /** \returns the tranpose permutation matrix.
+      *
+      * \note \blank \note_try_to_help_rvo
+      */
+    inline InverseReturnType transpose() const
+    { return InverseReturnType(derived()); }
+
+    /**** multiplication helpers to hopefully get RVO ****/
+
+  
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  protected:
+    template<typename OtherDerived>
+    void assignTranspose(const PermutationBase<OtherDerived>& other)
+    {
+      for (Index i=0; i<rows();++i) indices().coeffRef(other.indices().coeff(i)) = i;
+    }
+    template<typename Lhs,typename Rhs>
+    void assignProduct(const Lhs& lhs, const Rhs& rhs)
+    {
+      eigen_assert(lhs.cols() == rhs.rows());
+      for (Index i=0; i<rows();++i) indices().coeffRef(i) = lhs.indices().coeff(rhs.indices().coeff(i));
+    }
+#endif
+
+  public:
+
+    /** \returns the product permutation matrix.
+      *
+      * \note \blank \note_try_to_help_rvo
+      */
+    template<typename Other>
+    inline PlainPermutationType operator*(const PermutationBase<Other>& other) const
+    { return PlainPermutationType(internal::PermPermProduct, derived(), other.derived()); }
+
+    /** \returns the product of a permutation with another inverse permutation.
+      *
+      * \note \blank \note_try_to_help_rvo
+      */
+    template<typename Other>
+    inline PlainPermutationType operator*(const InverseImpl<Other,PermutationStorage>& other) const
+    { return PlainPermutationType(internal::PermPermProduct, *this, other.eval()); }
+
+    /** \returns the product of an inverse permutation with another permutation.
+      *
+      * \note \blank \note_try_to_help_rvo
+      */
+    template<typename Other> friend
+    inline PlainPermutationType operator*(const InverseImpl<Other, PermutationStorage>& other, const PermutationBase& perm)
+    { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); }
+    
+    /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation.
+      *
+      * This function is O(\c n) procedure allocating a buffer of \c n booleans.
+      */
+    Index determinant() const
+    {
+      Index res = 1;
+      Index n = size();
+      Matrix<bool,RowsAtCompileTime,1,0,MaxRowsAtCompileTime> mask(n);
+      mask.fill(false);
+      Index r = 0;
+      while(r < n)
+      {
+        // search for the next seed
+        while(r<n && mask[r]) r++;
+        if(r>=n)
+          break;
+        // we got one, let's follow it until we are back to the seed
+        Index k0 = r++;
+        mask.coeffRef(k0) = true;
+        for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k))
+        {
+          mask.coeffRef(k) = true;
+          res = -res;
+        }
+      }
+      return res;
+    }
+
+  protected:
+
+};
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+{
+  typedef PermutationStorage StorageKind;
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
+};
+}
+
+/** \class PermutationMatrix
+  * \ingroup Core_Module
+  *
+  * \brief Permutation matrix
+  *
+  * \tparam SizeAtCompileTime the number of rows/cols, or Dynamic
+  * \tparam MaxSizeAtCompileTime the maximum number of rows/cols, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
+  * \tparam _StorageIndex the integer type of the indices
+  *
+  * This class represents a permutation matrix, internally stored as a vector of integers.
+  *
+  * \sa class PermutationBase, class PermutationWrapper, class DiagonalMatrix
+  */
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class PermutationMatrix : public PermutationBase<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex> >
+{
+    typedef PermutationBase<PermutationMatrix> Base;
+    typedef internal::traits<PermutationMatrix> Traits;
+  public:
+
+    typedef const PermutationMatrix& Nested;
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename Traits::StorageIndex StorageIndex;
+    #endif
+
+    inline PermutationMatrix()
+    {}
+
+    /** Constructs an uninitialized permutation matrix of given size.
+      */
+    explicit inline PermutationMatrix(Index size) : m_indices(size)
+    {
+      eigen_internal_assert(size <= NumTraits<StorageIndex>::highest());
+    }
+
+    /** Copy constructor. */
+    template<typename OtherDerived>
+    inline PermutationMatrix(const PermutationBase<OtherDerived>& other)
+      : m_indices(other.indices()) {}
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** Standard copy constructor. Defined only to prevent a default copy constructor
+      * from hiding the other templated constructor */
+    inline PermutationMatrix(const PermutationMatrix& other) : m_indices(other.indices()) {}
+    #endif
+
+    /** Generic constructor from expression of the indices. The indices
+      * array has the meaning that the permutations sends each integer i to indices[i].
+      *
+      * \warning It is your responsibility to check that the indices array that you passes actually
+      * describes a permutation, i.e., each value between 0 and n-1 occurs exactly once, where n is the
+      * array's size.
+      */
+    template<typename Other>
+    explicit inline PermutationMatrix(const MatrixBase<Other>& indices) : m_indices(indices)
+    {}
+
+    /** Convert the Transpositions \a tr to a permutation matrix */
+    template<typename Other>
+    explicit PermutationMatrix(const TranspositionsBase<Other>& tr)
+      : m_indices(tr.size())
+    {
+      *this = tr;
+    }
+
+    /** Copies the other permutation into *this */
+    template<typename Other>
+    PermutationMatrix& operator=(const PermutationBase<Other>& other)
+    {
+      m_indices = other.indices();
+      return *this;
+    }
+
+    /** Assignment from the Transpositions \a tr */
+    template<typename Other>
+    PermutationMatrix& operator=(const TranspositionsBase<Other>& tr)
+    {
+      return Base::operator=(tr.derived());
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    PermutationMatrix& operator=(const PermutationMatrix& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    /** \returns a reference to the stored array representing the permutation. */
+    IndicesType& indices() { return m_indices; }
+
+
+    /**** multiplication helpers to hopefully get RVO ****/
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename Other>
+    PermutationMatrix(const InverseImpl<Other,PermutationStorage>& other)
+      : m_indices(other.derived().nestedExpression().size())
+    {
+      eigen_internal_assert(m_indices.size() <= NumTraits<StorageIndex>::highest());
+      StorageIndex end = StorageIndex(m_indices.size());
+      for (StorageIndex i=0; i<end;++i)
+        m_indices.coeffRef(other.derived().nestedExpression().indices().coeff(i)) = i;
+    }
+    template<typename Lhs,typename Rhs>
+    PermutationMatrix(internal::PermPermProduct_t, const Lhs& lhs, const Rhs& rhs)
+      : m_indices(lhs.indices().size())
+    {
+      Base::assignProduct(lhs,rhs);
+    }
+#endif
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
+ : traits<Matrix<_StorageIndex,SizeAtCompileTime,SizeAtCompileTime,0,MaxSizeAtCompileTime,MaxSizeAtCompileTime> >
+{
+  typedef PermutationStorage StorageKind;
+  typedef Map<const Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef void Scalar;
+};
+}
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+class Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess>
+  : public PermutationBase<Map<PermutationMatrix<SizeAtCompileTime, MaxSizeAtCompileTime, _StorageIndex>,_PacketAccess> >
+{
+    typedef PermutationBase<Map> Base;
+    typedef internal::traits<Map> Traits;
+  public:
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar StorageIndex;
+    #endif
+
+    inline Map(const StorageIndex* indicesPtr)
+      : m_indices(indicesPtr)
+    {}
+
+    inline Map(const StorageIndex* indicesPtr, Index size)
+      : m_indices(indicesPtr,size)
+    {}
+
+    /** Copies the other permutation into *this */
+    template<typename Other>
+    Map& operator=(const PermutationBase<Other>& other)
+    { return Base::operator=(other.derived()); }
+
+    /** Assignment from the Transpositions \a tr */
+    template<typename Other>
+    Map& operator=(const TranspositionsBase<Other>& tr)
+    { return Base::operator=(tr.derived()); }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Map& operator=(const Map& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    /** \returns a reference to the stored array representing the permutation. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+template<typename _IndicesType> class TranspositionsWrapper;
+namespace internal {
+template<typename _IndicesType>
+struct traits<PermutationWrapper<_IndicesType> >
+{
+  typedef PermutationStorage StorageKind;
+  typedef void Scalar;
+  typedef typename _IndicesType::Scalar StorageIndex;
+  typedef _IndicesType IndicesType;
+  enum {
+    RowsAtCompileTime = _IndicesType::SizeAtCompileTime,
+    ColsAtCompileTime = _IndicesType::SizeAtCompileTime,
+    MaxRowsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    MaxColsAtCompileTime = IndicesType::MaxSizeAtCompileTime,
+    Flags = 0
+  };
+};
+}
+
+/** \class PermutationWrapper
+  * \ingroup Core_Module
+  *
+  * \brief Class to view a vector of integers as a permutation matrix
+  *
+  * \tparam _IndicesType the type of the vector of integer (can be any compatible expression)
+  *
+  * This class allows to view any vector expression of integers as a permutation matrix.
+  *
+  * \sa class PermutationBase, class PermutationMatrix
+  */
+template<typename _IndicesType>
+class PermutationWrapper : public PermutationBase<PermutationWrapper<_IndicesType> >
+{
+    typedef PermutationBase<PermutationWrapper> Base;
+    typedef internal::traits<PermutationWrapper> Traits;
+  public:
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename Traits::IndicesType IndicesType;
+    #endif
+
+    inline PermutationWrapper(const IndicesType& indices)
+      : m_indices(indices)
+    {}
+
+    /** const version of indices(). */
+    const typename internal::remove_all<typename IndicesType::Nested>::type&
+    indices() const { return m_indices; }
+
+  protected:
+
+    typename IndicesType::Nested m_indices;
+};
+
+
+/** \returns the matrix with the permutation applied to the columns.
+  */
+template<typename MatrixDerived, typename PermutationDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const PermutationBase<PermutationDerived>& permutation)
+{
+  return Product<MatrixDerived, PermutationDerived, AliasFreeProduct>
+            (matrix.derived(), permutation.derived());
+}
+
+/** \returns the matrix with the permutation applied to the rows.
+  */
+template<typename PermutationDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
+operator*(const PermutationBase<PermutationDerived> &permutation,
+          const MatrixBase<MatrixDerived>& matrix)
+{
+  return Product<PermutationDerived, MatrixDerived, AliasFreeProduct>
+            (permutation.derived(), matrix.derived());
+}
+
+
+template<typename PermutationType>
+class InverseImpl<PermutationType, PermutationStorage>
+  : public EigenBase<Inverse<PermutationType> >
+{
+    typedef typename PermutationType::PlainPermutationType PlainPermutationType;
+    typedef internal::traits<PermutationType> PermTraits;
+  protected:
+    InverseImpl() {}
+  public:
+    typedef Inverse<PermutationType> InverseType;
+    using EigenBase<Inverse<PermutationType> >::derived;
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    typedef typename PermutationType::DenseMatrixType DenseMatrixType;
+    enum {
+      RowsAtCompileTime = PermTraits::RowsAtCompileTime,
+      ColsAtCompileTime = PermTraits::ColsAtCompileTime,
+      MaxRowsAtCompileTime = PermTraits::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = PermTraits::MaxColsAtCompileTime
+    };
+    #endif
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename DenseDerived>
+    void evalTo(MatrixBase<DenseDerived>& other) const
+    {
+      other.setZero();
+      for (Index i=0; i<derived().rows();++i)
+        other.coeffRef(i, derived().nestedExpression().indices().coeff(i)) = typename DenseDerived::Scalar(1);
+    }
+    #endif
+
+    /** \return the equivalent permutation matrix */
+    PlainPermutationType eval() const { return derived(); }
+
+    DenseMatrixType toDenseMatrix() const { return derived(); }
+
+    /** \returns the matrix with the inverse permutation applied to the columns.
+      */
+    template<typename OtherDerived> friend
+    const Product<OtherDerived, InverseType, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const InverseType& trPerm)
+    {
+      return Product<OtherDerived, InverseType, AliasFreeProduct>(matrix.derived(), trPerm.derived());
+    }
+
+    /** \returns the matrix with the inverse permutation applied to the rows.
+      */
+    template<typename OtherDerived>
+    const Product<InverseType, OtherDerived, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix) const
+    {
+      return Product<InverseType, OtherDerived, AliasFreeProduct>(derived(), matrix.derived());
+    }
+};
+
+template<typename Derived>
+const PermutationWrapper<const Derived> MatrixBase<Derived>::asPermutation() const
+{
+  return derived();
+}
+
+namespace internal {
+
+template<> struct AssignmentKind<DenseShape,PermutationShape> { typedef EigenBase2EigenBase Kind; };
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PERMUTATIONMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h
new file mode 100644
index 0000000..77f4f60
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/PlainObjectBase.h
@@ -0,0 +1,1031 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DENSESTORAGEBASE_H
+#define EIGEN_DENSESTORAGEBASE_H
+
+#if defined(EIGEN_INITIALIZE_MATRICES_BY_ZERO)
+# define EIGEN_INITIALIZE_COEFFS
+# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED for(int i=0;i<base().size();++i) coeffRef(i)=Scalar(0);
+#elif defined(EIGEN_INITIALIZE_MATRICES_BY_NAN)
+# define EIGEN_INITIALIZE_COEFFS
+# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED for(int i=0;i<base().size();++i) coeffRef(i)=std::numeric_limits<Scalar>::quiet_NaN();
+#else
+# undef EIGEN_INITIALIZE_COEFFS
+# define EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+template<int MaxSizeAtCompileTime> struct check_rows_cols_for_overflow {
+  template<typename Index>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_ALWAYS_INLINE void run(Index, Index)
+  {
+  }
+};
+
+template<> struct check_rows_cols_for_overflow<Dynamic> {
+  template<typename Index>
+  EIGEN_DEVICE_FUNC
+  static EIGEN_ALWAYS_INLINE void run(Index rows, Index cols)
+  {
+    // http://hg.mozilla.org/mozilla-central/file/6c8a909977d3/xpcom/ds/CheckedInt.h#l242
+    // we assume Index is signed
+    Index max_index = (std::size_t(1) << (8 * sizeof(Index) - 1)) - 1; // assume Index is signed
+    bool error = (rows == 0 || cols == 0) ? false
+               : (rows > max_index / cols);
+    if (error)
+      throw_std_bad_alloc();
+  }
+};
+
+template <typename Derived,
+          typename OtherDerived = Derived,
+          bool IsVector = bool(Derived::IsVectorAtCompileTime) && bool(OtherDerived::IsVectorAtCompileTime)>
+struct conservative_resize_like_impl;
+
+template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers> struct matrix_swap_impl;
+
+} // end namespace internal
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+namespace doxygen {
+
+// This is a workaround to doxygen not being able to understand the inheritance logic
+// when it is hidden by the dense_xpr_base helper struct.
+// Moreover, doxygen fails to include members that are not documented in the declaration body of
+// MatrixBase if we inherits MatrixBase<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >,
+// this is why we simply inherits MatrixBase, though this does not make sense.
+
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
+template<typename Derived> struct dense_xpr_base_dispatcher;
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct dense_xpr_base_dispatcher<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public MatrixBase {};
+/** This class is just a workaround for Doxygen and it does not not actually exist. */
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct dense_xpr_base_dispatcher<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols> >
+    : public ArrayBase {};
+
+} // namespace doxygen
+
+/** \class PlainObjectBase
+  * \ingroup Core_Module
+  * \brief %Dense storage base class for matrices and arrays.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_PLAINOBJECTBASE_PLUGIN.
+  *
+  * \tparam Derived is the derived type, e.g., a Matrix or Array
+  *
+  * \sa \ref TopicClassHierarchy
+  */
+template<typename Derived>
+class PlainObjectBase : public doxygen::dense_xpr_base_dispatcher<Derived>
+#else
+template<typename Derived>
+class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
+#endif
+{
+  public:
+    enum { Options = internal::traits<Derived>::Options };
+    typedef typename internal::dense_xpr_base<Derived>::type Base;
+
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Derived DenseType;
+
+    using Base::RowsAtCompileTime;
+    using Base::ColsAtCompileTime;
+    using Base::SizeAtCompileTime;
+    using Base::MaxRowsAtCompileTime;
+    using Base::MaxColsAtCompileTime;
+    using Base::MaxSizeAtCompileTime;
+    using Base::IsVectorAtCompileTime;
+    using Base::Flags;
+
+    template<typename PlainObjectType, int MapOptions, typename StrideType> friend class Eigen::Map;
+    friend  class Eigen::Map<Derived, Unaligned>;
+    typedef Eigen::Map<Derived, Unaligned>  MapType;
+    friend  class Eigen::Map<const Derived, Unaligned>;
+    typedef const Eigen::Map<const Derived, Unaligned> ConstMapType;
+#if EIGEN_MAX_ALIGN_BYTES>0
+    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
+    friend  class Eigen::Map<Derived, AlignedMax>;
+    friend  class Eigen::Map<const Derived, AlignedMax>;
+#endif
+    typedef Eigen::Map<Derived, AlignedMax> AlignedMapType;
+    typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType;
+    template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; };
+    template<typename StrideType> struct StridedConstMapType { typedef Eigen::Map<const Derived, Unaligned, StrideType> type; };
+    template<typename StrideType> struct StridedAlignedMapType { typedef Eigen::Map<Derived, AlignedMax, StrideType> type; };
+    template<typename StrideType> struct StridedConstAlignedMapType { typedef Eigen::Map<const Derived, AlignedMax, StrideType> type; };
+
+  protected:
+    DenseStorage<Scalar, Base::MaxSizeAtCompileTime, Base::RowsAtCompileTime, Base::ColsAtCompileTime, Options> m_storage;
+
+  public:
+    enum { NeedsToAlign = (SizeAtCompileTime != Dynamic) && (internal::traits<Derived>::Alignment>0) };
+    EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
+
+    EIGEN_DEVICE_FUNC
+    Base& base() { return *static_cast<Base*>(this); }
+    EIGEN_DEVICE_FUNC
+    const Base& base() const { return *static_cast<const Base*>(this); }
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index rows() const { return m_storage.rows(); }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Index cols() const { return m_storage.cols(); }
+
+    /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index,Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeff(Index rowId, Index colId) const
+    {
+      if(Flags & RowMajorBit)
+        return m_storage.data()[colId + rowId * m_storage.cols()];
+      else // column-major
+        return m_storage.data()[rowId + colId * m_storage.rows()];
+    }
+
+    /** This is an overloaded version of DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,ReadOnlyAccessors>::coeff(Index) const for details. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeff(Index index) const
+    {
+      return m_storage.data()[index];
+    }
+
+    /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index,Index) const for details. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index rowId, Index colId)
+    {
+      if(Flags & RowMajorBit)
+        return m_storage.data()[colId + rowId * m_storage.cols()];
+      else // column-major
+        return m_storage.data()[rowId + colId * m_storage.rows()];
+    }
+
+    /** This is an overloaded version of DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const
+      * provided to by-pass the creation of an evaluator of the expression, thus saving compilation efforts.
+      *
+      * See DenseCoeffsBase<Derived,WriteAccessors>::coeffRef(Index) const for details. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Scalar& coeffRef(Index index)
+    {
+      return m_storage.data()[index];
+    }
+
+    /** This is the const version of coeffRef(Index,Index) which is thus synonym of coeff(Index,Index).
+      * It is provided for convenience. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      if(Flags & RowMajorBit)
+        return m_storage.data()[colId + rowId * m_storage.cols()];
+      else // column-major
+        return m_storage.data()[rowId + colId * m_storage.rows()];
+    }
+
+    /** This is the const version of coeffRef(Index) which is thus synonym of coeff(Index).
+      * It is provided for convenience. */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const Scalar& coeffRef(Index index) const
+    {
+      return m_storage.data()[index];
+    }
+
+    /** \internal */
+    template<int LoadMode>
+    EIGEN_STRONG_INLINE PacketScalar packet(Index rowId, Index colId) const
+    {
+      return internal::ploadt<PacketScalar, LoadMode>
+               (m_storage.data() + (Flags & RowMajorBit
+                                   ? colId + rowId * m_storage.cols()
+                                   : rowId + colId * m_storage.rows()));
+    }
+
+    /** \internal */
+    template<int LoadMode>
+    EIGEN_STRONG_INLINE PacketScalar packet(Index index) const
+    {
+      return internal::ploadt<PacketScalar, LoadMode>(m_storage.data() + index);
+    }
+
+    /** \internal */
+    template<int StoreMode>
+    EIGEN_STRONG_INLINE void writePacket(Index rowId, Index colId, const PacketScalar& val)
+    {
+      internal::pstoret<Scalar, PacketScalar, StoreMode>
+              (m_storage.data() + (Flags & RowMajorBit
+                                   ? colId + rowId * m_storage.cols()
+                                   : rowId + colId * m_storage.rows()), val);
+    }
+
+    /** \internal */
+    template<int StoreMode>
+    EIGEN_STRONG_INLINE void writePacket(Index index, const PacketScalar& val)
+    {
+      internal::pstoret<Scalar, PacketScalar, StoreMode>(m_storage.data() + index, val);
+    }
+
+    /** \returns a const pointer to the data array of this matrix */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar *data() const
+    { return m_storage.data(); }
+
+    /** \returns a pointer to the data array of this matrix */
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar *data()
+    { return m_storage.data(); }
+
+    /** Resizes \c *this to a \a rows x \a cols matrix.
+      *
+      * This method is intended for dynamic-size matrices, although it is legal to call it on any
+      * matrix as long as fixed dimensions are left unchanged. If you only want to change the number
+      * of rows and/or of columns, you can use resize(NoChange_t, Index), resize(Index, NoChange_t).
+      *
+      * If the current number of coefficients of \c *this exactly matches the
+      * product \a rows * \a cols, then no memory allocation is performed and
+      * the current values are left unchanged. In all other cases, including
+      * shrinking, the data is reallocated and all previous values are lost.
+      *
+      * Example: \include Matrix_resize_int_int.cpp
+      * Output: \verbinclude Matrix_resize_int_int.out
+      *
+      * \sa resize(Index) for vectors, resize(NoChange_t, Index), resize(Index, NoChange_t)
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void resize(Index rows, Index cols)
+    {
+      eigen_assert(   EIGEN_IMPLIES(RowsAtCompileTime!=Dynamic,rows==RowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime!=Dynamic,cols==ColsAtCompileTime)
+                   && EIGEN_IMPLIES(RowsAtCompileTime==Dynamic && MaxRowsAtCompileTime!=Dynamic,rows<=MaxRowsAtCompileTime)
+                   && EIGEN_IMPLIES(ColsAtCompileTime==Dynamic && MaxColsAtCompileTime!=Dynamic,cols<=MaxColsAtCompileTime)
+                   && rows>=0 && cols>=0 && "Invalid sizes when resizing a matrix or array.");
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(rows, cols);
+      #ifdef EIGEN_INITIALIZE_COEFFS
+        Index size = rows*cols;
+        bool size_changed = size != this->size();
+        m_storage.resize(size, rows, cols);
+        if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      #else
+        m_storage.resize(rows*cols, rows, cols);
+      #endif
+    }
+
+    /** Resizes \c *this to a vector of length \a size
+      *
+      * \only_for_vectors. This method does not work for
+      * partially dynamic matrices when the static dimension is anything other
+      * than 1. For example it will not work with Matrix<double, 2, Dynamic>.
+      *
+      * Example: \include Matrix_resize_int.cpp
+      * Output: \verbinclude Matrix_resize_int.out
+      *
+      * \sa resize(Index,Index), resize(NoChange_t, Index), resize(Index, NoChange_t)
+      */
+    EIGEN_DEVICE_FUNC
+    inline void resize(Index size)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(PlainObjectBase)
+      eigen_assert(((SizeAtCompileTime == Dynamic && (MaxSizeAtCompileTime==Dynamic || size<=MaxSizeAtCompileTime)) || SizeAtCompileTime == size) && size>=0);
+      #ifdef EIGEN_INITIALIZE_COEFFS
+        bool size_changed = size != this->size();
+      #endif
+      if(RowsAtCompileTime == 1)
+        m_storage.resize(size, 1, size);
+      else
+        m_storage.resize(size, size, 1);
+      #ifdef EIGEN_INITIALIZE_COEFFS
+        if(size_changed) EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+      #endif
+    }
+
+    /** Resizes the matrix, changing only the number of columns. For the parameter of type NoChange_t, just pass the special value \c NoChange
+      * as in the example below.
+      *
+      * Example: \include Matrix_resize_NoChange_int.cpp
+      * Output: \verbinclude Matrix_resize_NoChange_int.out
+      *
+      * \sa resize(Index,Index)
+      */
+    EIGEN_DEVICE_FUNC
+    inline void resize(NoChange_t, Index cols)
+    {
+      resize(rows(), cols);
+    }
+
+    /** Resizes the matrix, changing only the number of rows. For the parameter of type NoChange_t, just pass the special value \c NoChange
+      * as in the example below.
+      *
+      * Example: \include Matrix_resize_int_NoChange.cpp
+      * Output: \verbinclude Matrix_resize_int_NoChange.out
+      *
+      * \sa resize(Index,Index)
+      */
+    EIGEN_DEVICE_FUNC
+    inline void resize(Index rows, NoChange_t)
+    {
+      resize(rows, cols());
+    }
+
+    /** Resizes \c *this to have the same dimensions as \a other.
+      * Takes care of doing all the checking that's needed.
+      *
+      * Note that copying a row-vector into a vector (and conversely) is allowed.
+      * The resizing, if any, is then done in the appropriate way so that row-vectors
+      * remain row-vectors and vectors remain vectors.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void resizeLike(const EigenBase<OtherDerived>& _other)
+    {
+      const OtherDerived& other = _other.derived();
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.rows(), other.cols());
+      const Index othersize = other.rows()*other.cols();
+      if(RowsAtCompileTime == 1)
+      {
+        eigen_assert(other.rows() == 1 || other.cols() == 1);
+        resize(1, othersize);
+      }
+      else if(ColsAtCompileTime == 1)
+      {
+        eigen_assert(other.rows() == 1 || other.cols() == 1);
+        resize(othersize, 1);
+      }
+      else resize(other.rows(), other.cols());
+    }
+
+    /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
+      *
+      * The method is intended for matrices of dynamic size. If you only want to change the number
+      * of rows and/or of columns, you can use conservativeResize(NoChange_t, Index) or
+      * conservativeResize(Index, NoChange_t).
+      *
+      * Matrices are resized relative to the top-left element. In case values need to be 
+      * appended to the matrix they will be uninitialized.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, Index cols)
+    {
+      internal::conservative_resize_like_impl<Derived>::run(*this, rows, cols);
+    }
+
+    /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
+      *
+      * As opposed to conservativeResize(Index rows, Index cols), this version leaves
+      * the number of columns unchanged.
+      *
+      * In case the matrix is growing, new rows will be uninitialized.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index rows, NoChange_t)
+    {
+      // Note: see the comment in conservativeResize(Index,Index)
+      conservativeResize(rows, cols());
+    }
+
+    /** Resizes the matrix to \a rows x \a cols while leaving old values untouched.
+      *
+      * As opposed to conservativeResize(Index rows, Index cols), this version leaves
+      * the number of rows unchanged.
+      *
+      * In case the matrix is growing, new columns will be uninitialized.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(NoChange_t, Index cols)
+    {
+      // Note: see the comment in conservativeResize(Index,Index)
+      conservativeResize(rows(), cols);
+    }
+
+    /** Resizes the vector to \a size while retaining old values.
+      *
+      * \only_for_vectors. This method does not work for
+      * partially dynamic matrices when the static dimension is anything other
+      * than 1. For example it will not work with Matrix<double, 2, Dynamic>.
+      *
+      * When values are appended, they will be uninitialized.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResize(Index size)
+    {
+      internal::conservative_resize_like_impl<Derived>::run(*this, size);
+    }
+
+    /** Resizes the matrix to \a rows x \a cols of \c other, while leaving old values untouched.
+      *
+      * The method is intended for matrices of dynamic size. If you only want to change the number
+      * of rows and/or of columns, you can use conservativeResize(NoChange_t, Index) or
+      * conservativeResize(Index, NoChange_t).
+      *
+      * Matrices are resized relative to the top-left element. In case values need to be 
+      * appended to the matrix they will copied from \c other.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void conservativeResizeLike(const DenseBase<OtherDerived>& other)
+    {
+      internal::conservative_resize_like_impl<Derived,OtherDerived>::run(*this, other);
+    }
+
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& operator=(const PlainObjectBase& other)
+    {
+      return _set(other);
+    }
+
+    /** \sa MatrixBase::lazyAssign() */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& lazyAssign(const DenseBase<OtherDerived>& other)
+    {
+      _resize_to_match(other);
+      return Base::lazyAssign(other.derived());
+    }
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE Derived& operator=(const ReturnByValue<OtherDerived>& func)
+    {
+      resize(func.rows(), func.cols());
+      return Base::operator=(func);
+    }
+
+    // Prevent user from trying to instantiate PlainObjectBase objects
+    // by making all its constructor protected. See bug 1074.
+  protected:
+
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase() : m_storage()
+    {
+//       _check_template_params();
+//       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    // FIXME is it still needed ?
+    /** \internal */
+    EIGEN_DEVICE_FUNC
+    explicit PlainObjectBase(internal::constructor_without_unaligned_array_assert)
+      : m_storage(internal::constructor_without_unaligned_array_assert())
+    {
+//       _check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+#endif
+
+#if EIGEN_HAS_RVALUE_REFERENCES
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase(PlainObjectBase&& other) EIGEN_NOEXCEPT
+      : m_storage( std::move(other.m_storage) )
+    {
+    }
+
+    EIGEN_DEVICE_FUNC
+    PlainObjectBase& operator=(PlainObjectBase&& other) EIGEN_NOEXCEPT
+    {
+      using std::swap;
+      swap(m_storage, other.m_storage);
+      return *this;
+    }
+#endif
+
+    /** Copy constructor */
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
+      : Base(), m_storage(other.m_storage) { }
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(Index size, Index rows, Index cols)
+      : m_storage(size, rows, cols)
+    {
+//       _check_template_params();
+//       EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED
+    }
+
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other)
+      : m_storage()
+    {
+      _check_template_params();
+      resizeLike(other);
+      _set_noalias(other);
+    }
+
+    /** \sa PlainObjectBase::operator=(const EigenBase<OtherDerived>&) */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
+      : m_storage()
+    {
+      _check_template_params();
+      resizeLike(other);
+      *this = other.derived();
+    }
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE PlainObjectBase(const ReturnByValue<OtherDerived>& other)
+    {
+      _check_template_params();
+      // FIXME this does not automatically transpose vectors if necessary
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
+    }
+
+  public:
+
+    /** \brief Copies the generic expression \a other into *this.
+      * \copydetails DenseBase::operator=(const EigenBase<OtherDerived> &other)
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE Derived& operator=(const EigenBase<OtherDerived> &other)
+    {
+      _resize_to_match(other);
+      Base::operator=(other.derived());
+      return this->derived();
+    }
+
+    /** \name Map
+      * These are convenience functions returning Map objects. The Map() static functions return unaligned Map objects,
+      * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned
+      * \a data pointers.
+      *
+      * \see class Map
+      */
+    //@{
+    static inline ConstMapType Map(const Scalar* data)
+    { return ConstMapType(data); }
+    static inline MapType Map(Scalar* data)
+    { return MapType(data); }
+    static inline ConstMapType Map(const Scalar* data, Index size)
+    { return ConstMapType(data, size); }
+    static inline MapType Map(Scalar* data, Index size)
+    { return MapType(data, size); }
+    static inline ConstMapType Map(const Scalar* data, Index rows, Index cols)
+    { return ConstMapType(data, rows, cols); }
+    static inline MapType Map(Scalar* data, Index rows, Index cols)
+    { return MapType(data, rows, cols); }
+
+    static inline ConstAlignedMapType MapAligned(const Scalar* data)
+    { return ConstAlignedMapType(data); }
+    static inline AlignedMapType MapAligned(Scalar* data)
+    { return AlignedMapType(data); }
+    static inline ConstAlignedMapType MapAligned(const Scalar* data, Index size)
+    { return ConstAlignedMapType(data, size); }
+    static inline AlignedMapType MapAligned(Scalar* data, Index size)
+    { return AlignedMapType(data, size); }
+    static inline ConstAlignedMapType MapAligned(const Scalar* data, Index rows, Index cols)
+    { return ConstAlignedMapType(data, rows, cols); }
+    static inline AlignedMapType MapAligned(Scalar* data, Index rows, Index cols)
+    { return AlignedMapType(data, rows, cols); }
+
+    template<int Outer, int Inner>
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, const Stride<Outer, Inner>& stride)
+    { return typename StridedMapType<Stride<Outer, Inner> >::type(data, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, size, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    { return typename StridedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedConstMapType<Stride<Outer, Inner> >::type Map(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedMapType<Stride<Outer, Inner> >::type Map(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    { return typename StridedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
+
+    template<int Outer, int Inner>
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, const Stride<Outer, Inner>& stride)
+    { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index size, const Stride<Outer, Inner>& stride)
+    { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, size, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type MapAligned(const Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    { return typename StridedConstAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
+    template<int Outer, int Inner>
+    static inline typename StridedAlignedMapType<Stride<Outer, Inner> >::type MapAligned(Scalar* data, Index rows, Index cols, const Stride<Outer, Inner>& stride)
+    { return typename StridedAlignedMapType<Stride<Outer, Inner> >::type(data, rows, cols, stride); }
+    //@}
+
+    using Base::setConstant;
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index size, const Scalar& val);
+    EIGEN_DEVICE_FUNC Derived& setConstant(Index rows, Index cols, const Scalar& val);
+
+    using Base::setZero;
+    EIGEN_DEVICE_FUNC Derived& setZero(Index size);
+    EIGEN_DEVICE_FUNC Derived& setZero(Index rows, Index cols);
+
+    using Base::setOnes;
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index size);
+    EIGEN_DEVICE_FUNC Derived& setOnes(Index rows, Index cols);
+
+    using Base::setRandom;
+    Derived& setRandom(Index size);
+    Derived& setRandom(Index rows, Index cols);
+
+    #ifdef EIGEN_PLAINOBJECTBASE_PLUGIN
+    #include EIGEN_PLAINOBJECTBASE_PLUGIN
+    #endif
+
+  protected:
+    /** \internal Resizes *this in preparation for assigning \a other to it.
+      * Takes care of doing all the checking that's needed.
+      *
+      * Note that copying a row-vector into a vector (and conversely) is allowed.
+      * The resizing, if any, is then done in the appropriate way so that row-vectors
+      * remain row-vectors and vectors remain vectors.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _resize_to_match(const EigenBase<OtherDerived>& other)
+    {
+      #ifdef EIGEN_NO_AUTOMATIC_RESIZING
+      eigen_assert((this->size()==0 || (IsVectorAtCompileTime ? (this->size() == other.size())
+                 : (rows() == other.rows() && cols() == other.cols())))
+        && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined");
+      EIGEN_ONLY_USED_FOR_DEBUG(other);
+      #else
+      resizeLike(other);
+      #endif
+    }
+
+    /**
+      * \brief Copies the value of the expression \a other into \c *this with automatic resizing.
+      *
+      * *this might be resized to match the dimensions of \a other. If *this was a null matrix (not already initialized),
+      * it will be initialized.
+      *
+      * Note that copying a row-vector into a vector (and conversely) is allowed.
+      * The resizing, if any, is then done in the appropriate way so that row-vectors
+      * remain row-vectors and vectors remain vectors.
+      *
+      * \sa operator=(const MatrixBase<OtherDerived>&), _set_noalias()
+      *
+      * \internal
+      */
+    // aliasing is dealt once in internall::call_assignment
+    // so at this stage we have to assume aliasing... and resising has to be done later.
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE Derived& _set(const DenseBase<OtherDerived>& other)
+    {
+      internal::call_assignment(this->derived(), other.derived());
+      return this->derived();
+    }
+
+    /** \internal Like _set() but additionally makes the assumption that no aliasing effect can happen (which
+      * is the case when creating a new matrix) so one can enforce lazy evaluation.
+      *
+      * \sa operator=(const MatrixBase<OtherDerived>&), _set()
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE Derived& _set_noalias(const DenseBase<OtherDerived>& other)
+    {
+      // I don't think we need this resize call since the lazyAssign will anyways resize
+      // and lazyAssign will be called by the assign selector.
+      //_resize_to_match(other);
+      // the 'false' below means to enforce lazy evaluation. We don't use lazyAssign() because
+      // it wouldn't allow to copy a row-vector into a column-vector.
+      internal::call_assignment_no_alias(this->derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+      return this->derived();
+    }
+
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init2(Index rows, Index cols, typename internal::enable_if<Base::SizeAtCompileTime!=2,T0>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT(bool(NumTraits<T0>::IsInteger) &&
+                          bool(NumTraits<T1>::IsInteger),
+                          FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
+      resize(rows,cols);
+    }
+    
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _init2(const T0& val0, const T1& val1, typename internal::enable_if<Base::SizeAtCompileTime==2,T0>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
+      m_storage.data()[0] = Scalar(val0);
+      m_storage.data()[1] = Scalar(val1);
+    }
+    
+    template<typename T0, typename T1>
+    EIGEN_DEVICE_FUNC 
+    EIGEN_STRONG_INLINE void _init2(const Index& val0, const Index& val1,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<T0,Index>::value)
+                                                                  && (internal::is_same<T1,Index>::value)
+                                                                  && Base::SizeAtCompileTime==2,T1>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 2)
+      m_storage.data()[0] = Scalar(val0);
+      m_storage.data()[1] = Scalar(val1);
+    }
+
+    // The argument is convertible to the Index type and we either have a non 1x1 Matrix, or a dynamic-sized Array,
+    // then the argument is meant to be the size of the object.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(Index size, typename internal::enable_if<    (Base::SizeAtCompileTime!=1 || !internal::is_convertible<T, Scalar>::value)
+                                                                              && ((!internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value || Base::SizeAtCompileTime==Dynamic)),T>::type* = 0)
+    {
+      // NOTE MSVC 2008 complains if we directly put bool(NumTraits<T>::IsInteger) as the EIGEN_STATIC_ASSERT argument.
+      const bool is_integer = NumTraits<T>::IsInteger;
+      EIGEN_UNUSED_VARIABLE(is_integer);
+      EIGEN_STATIC_ASSERT(is_integer,
+                          FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED)
+      resize(size);
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type can be implicitely converted)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0, typename internal::enable_if<Base::SizeAtCompileTime==1 && internal::is_convertible<T, Scalar>::value,T>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
+      m_storage.data()[0] = val0;
+    }
+    
+    // We have a 1x1 matrix/array => the argument is interpreted as the value of the unique coefficient (case where scalar type match the index type)
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime==1
+                                                                  && internal::is_convertible<T, Scalar>::value,T*>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(PlainObjectBase, 1)
+      m_storage.data()[0] = Scalar(val0);
+    }
+
+    // Initialize a fixed size matrix from a pointer to raw data
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar* data){
+      this->_set_noalias(ConstMapType(data));
+    }
+
+    // Initialize an arbitrary matrix from a dense expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const DenseBase<OtherDerived>& other){
+      this->_set_noalias(other);
+    }
+
+    // Initialize an arbitrary matrix from an object convertible to the Derived type.
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Derived& other){
+      this->_set_noalias(other);
+    }
+
+    // Initialize an arbitrary matrix from a generic Eigen expression
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const EigenBase<OtherDerived>& other){
+      this->derived() = other;
+    }
+
+    template<typename T, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const ReturnByValue<OtherDerived>& other)
+    {
+      resize(other.rows(), other.cols());
+      other.evalTo(this->derived());
+    }
+
+    template<typename T, typename OtherDerived, int ColsAtCompileTime>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const RotationBase<OtherDerived,ColsAtCompileTime>& r)
+    {
+      this->derived() = r;
+    }
+    
+    // For fixed-size Array<Scalar,...>
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Scalar& val0,
+                                    typename internal::enable_if<    Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
+    // For fixed-size Array<Index,...>
+    template<typename T>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _init1(const Index& val0,
+                                    typename internal::enable_if<    (!internal::is_same<Index,Scalar>::value)
+                                                                  && (internal::is_same<Index,T>::value)
+                                                                  && Base::SizeAtCompileTime!=Dynamic
+                                                                  && Base::SizeAtCompileTime!=1
+                                                                  && internal::is_convertible<T, Scalar>::value
+                                                                  && internal::is_same<typename internal::traits<Derived>::XprKind,ArrayXpr>::value,T*>::type* = 0)
+    {
+      Base::setConstant(val0);
+    }
+    
+    template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
+    friend struct internal::matrix_swap_impl;
+
+  public:
+    
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal
+      * \brief Override DenseBase::swap() since for dynamic-sized matrices
+      * of same type it is enough to swap the data pointers.
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> & other)
+    {
+      enum { SwapPointers = internal::is_same<Derived, OtherDerived>::value && Base::SizeAtCompileTime==Dynamic };
+      internal::matrix_swap_impl<Derived, OtherDerived, bool(SwapPointers)>::run(this->derived(), other.derived());
+    }
+    
+    /** \internal
+      * \brief const version forwarded to DenseBase::swap
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(DenseBase<OtherDerived> const & other)
+    { Base::swap(other.derived()); }
+    
+    EIGEN_DEVICE_FUNC 
+    static EIGEN_STRONG_INLINE void _check_template_params()
+    {
+      EIGEN_STATIC_ASSERT((EIGEN_IMPLIES(MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1, (Options&RowMajor)==RowMajor)
+                        && EIGEN_IMPLIES(MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1, (Options&RowMajor)==0)
+                        && ((RowsAtCompileTime == Dynamic) || (RowsAtCompileTime >= 0))
+                        && ((ColsAtCompileTime == Dynamic) || (ColsAtCompileTime >= 0))
+                        && ((MaxRowsAtCompileTime == Dynamic) || (MaxRowsAtCompileTime >= 0))
+                        && ((MaxColsAtCompileTime == Dynamic) || (MaxColsAtCompileTime >= 0))
+                        && (MaxRowsAtCompileTime == RowsAtCompileTime || RowsAtCompileTime==Dynamic)
+                        && (MaxColsAtCompileTime == ColsAtCompileTime || ColsAtCompileTime==Dynamic)
+                        && (Options & (DontAlign|RowMajor)) == Options),
+        INVALID_MATRIX_TEMPLATE_PARAMETERS)
+    }
+
+    enum { IsPlainObjectBase = 1 };
+#endif
+};
+
+namespace internal {
+
+template <typename Derived, typename OtherDerived, bool IsVector>
+struct conservative_resize_like_impl
+{
+  static void run(DenseBase<Derived>& _this, Index rows, Index cols)
+  {
+    if (_this.rows() == rows && _this.cols() == cols) return;
+    EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived)
+
+    if ( ( Derived::IsRowMajor && _this.cols() == cols) || // row-major and we change only the number of rows
+         (!Derived::IsRowMajor && _this.rows() == rows) )  // column-major and we change only the number of columns
+    {
+      internal::check_rows_cols_for_overflow<Derived::MaxSizeAtCompileTime>::run(rows, cols);
+      _this.derived().m_storage.conservativeResize(rows*cols,rows,cols);
+    }
+    else
+    {
+      // The storage order does not allow us to use reallocation.
+      typename Derived::PlainObject tmp(rows,cols);
+      const Index common_rows = numext::mini(rows, _this.rows());
+      const Index common_cols = numext::mini(cols, _this.cols());
+      tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
+      _this.derived().swap(tmp);
+    }
+  }
+
+  static void run(DenseBase<Derived>& _this, const DenseBase<OtherDerived>& other)
+  {
+    if (_this.rows() == other.rows() && _this.cols() == other.cols()) return;
+
+    // Note: Here is space for improvement. Basically, for conservativeResize(Index,Index),
+    // neither RowsAtCompileTime or ColsAtCompileTime must be Dynamic. If only one of the
+    // dimensions is dynamic, one could use either conservativeResize(Index rows, NoChange_t) or
+    // conservativeResize(NoChange_t, Index cols). For these methods new static asserts like
+    // EIGEN_STATIC_ASSERT_DYNAMIC_ROWS and EIGEN_STATIC_ASSERT_DYNAMIC_COLS would be good.
+    EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(Derived)
+    EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(OtherDerived)
+
+    if ( ( Derived::IsRowMajor && _this.cols() == other.cols()) || // row-major and we change only the number of rows
+         (!Derived::IsRowMajor && _this.rows() == other.rows()) )  // column-major and we change only the number of columns
+    {
+      const Index new_rows = other.rows() - _this.rows();
+      const Index new_cols = other.cols() - _this.cols();
+      _this.derived().m_storage.conservativeResize(other.size(),other.rows(),other.cols());
+      if (new_rows>0)
+        _this.bottomRightCorner(new_rows, other.cols()) = other.bottomRows(new_rows);
+      else if (new_cols>0)
+        _this.bottomRightCorner(other.rows(), new_cols) = other.rightCols(new_cols);
+    }
+    else
+    {
+      // The storage order does not allow us to use reallocation.
+      typename Derived::PlainObject tmp(other);
+      const Index common_rows = numext::mini(tmp.rows(), _this.rows());
+      const Index common_cols = numext::mini(tmp.cols(), _this.cols());
+      tmp.block(0,0,common_rows,common_cols) = _this.block(0,0,common_rows,common_cols);
+      _this.derived().swap(tmp);
+    }
+  }
+};
+
+// Here, the specialization for vectors inherits from the general matrix case
+// to allow calling .conservativeResize(rows,cols) on vectors.
+template <typename Derived, typename OtherDerived>
+struct conservative_resize_like_impl<Derived,OtherDerived,true>
+  : conservative_resize_like_impl<Derived,OtherDerived,false>
+{
+  using conservative_resize_like_impl<Derived,OtherDerived,false>::run;
+  
+  static void run(DenseBase<Derived>& _this, Index size)
+  {
+    const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : size;
+    const Index new_cols = Derived::RowsAtCompileTime==1 ? size : 1;
+    _this.derived().m_storage.conservativeResize(size,new_rows,new_cols);
+  }
+
+  static void run(DenseBase<Derived>& _this, const DenseBase<OtherDerived>& other)
+  {
+    if (_this.rows() == other.rows() && _this.cols() == other.cols()) return;
+
+    const Index num_new_elements = other.size() - _this.size();
+
+    const Index new_rows = Derived::RowsAtCompileTime==1 ? 1 : other.rows();
+    const Index new_cols = Derived::RowsAtCompileTime==1 ? other.cols() : 1;
+    _this.derived().m_storage.conservativeResize(other.size(),new_rows,new_cols);
+
+    if (num_new_elements > 0)
+      _this.tail(num_new_elements) = other.tail(num_new_elements);
+  }
+};
+
+template<typename MatrixTypeA, typename MatrixTypeB, bool SwapPointers>
+struct matrix_swap_impl
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(MatrixTypeA& a, MatrixTypeB& b)
+  {
+    a.base().swap(b);
+  }
+};
+
+template<typename MatrixTypeA, typename MatrixTypeB>
+struct matrix_swap_impl<MatrixTypeA, MatrixTypeB, true>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(MatrixTypeA& a, MatrixTypeB& b)
+  {
+    static_cast<typename MatrixTypeA::Base&>(a).m_storage.swap(static_cast<typename MatrixTypeB::Base&>(b).m_storage);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_DENSESTORAGEBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Product.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Product.h
new file mode 100644
index 0000000..ae0c94b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Product.h
@@ -0,0 +1,186 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PRODUCT_H
+#define EIGEN_PRODUCT_H
+
+namespace Eigen {
+
+template<typename Lhs, typename Rhs, int Option, typename StorageKind> class ProductImpl;
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, int Option>
+struct traits<Product<Lhs, Rhs, Option> >
+{
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename remove_all<Rhs>::type RhsCleaned;
+  typedef traits<LhsCleaned> LhsTraits;
+  typedef traits<RhsCleaned> RhsTraits;
+  
+  typedef MatrixXpr XprKind;
+  
+  typedef typename ScalarBinaryOpTraits<typename traits<LhsCleaned>::Scalar, typename traits<RhsCleaned>::Scalar>::ReturnType Scalar;
+  typedef typename product_promote_storage_type<typename LhsTraits::StorageKind,
+                                                typename RhsTraits::StorageKind,
+                                                internal::product_type<Lhs,Rhs>::ret>::ret StorageKind;
+  typedef typename promote_index_type<typename LhsTraits::StorageIndex,
+                                      typename RhsTraits::StorageIndex>::type StorageIndex;
+  
+  enum {
+    RowsAtCompileTime    = LhsTraits::RowsAtCompileTime,
+    ColsAtCompileTime    = RhsTraits::ColsAtCompileTime,
+    MaxRowsAtCompileTime = LhsTraits::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsTraits::MaxColsAtCompileTime,
+    
+    // FIXME: only needed by GeneralMatrixMatrixTriangular
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsTraits::ColsAtCompileTime, RhsTraits::RowsAtCompileTime),
+    
+    // The storage order is somewhat arbitrary here. The correct one will be determined through the evaluator.
+    Flags = (MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1) ? RowMajorBit
+          : (MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1) ? 0
+          : (   ((LhsTraits::Flags&NoPreferredStorageOrderBit) && (RhsTraits::Flags&RowMajorBit))
+             || ((RhsTraits::Flags&NoPreferredStorageOrderBit) && (LhsTraits::Flags&RowMajorBit)) ) ? RowMajorBit
+          : NoPreferredStorageOrderBit
+  };
+};
+
+} // end namespace internal
+
+/** \class Product
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the product of two arbitrary matrices or vectors
+  *
+  * \tparam _Lhs the type of the left-hand side expression
+  * \tparam _Rhs the type of the right-hand side expression
+  *
+  * This class represents an expression of the product of two arbitrary matrices.
+  *
+  * The other template parameters are:
+  * \tparam Option     can be DefaultProduct, AliasFreeProduct, or LazyProduct
+  *
+  */
+template<typename _Lhs, typename _Rhs, int Option>
+class Product : public ProductImpl<_Lhs,_Rhs,Option,
+                                   typename internal::product_promote_storage_type<typename internal::traits<_Lhs>::StorageKind,
+                                                                                   typename internal::traits<_Rhs>::StorageKind,
+                                                                                   internal::product_type<_Lhs,_Rhs>::ret>::ret>
+{
+  public:
+    
+    typedef _Lhs Lhs;
+    typedef _Rhs Rhs;
+    
+    typedef typename ProductImpl<
+        Lhs, Rhs, Option,
+        typename internal::product_promote_storage_type<typename internal::traits<Lhs>::StorageKind,
+                                                        typename internal::traits<Rhs>::StorageKind,
+                                                        internal::product_type<Lhs,Rhs>::ret>::ret>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
+
+    typedef typename internal::ref_selector<Lhs>::type LhsNested;
+    typedef typename internal::ref_selector<Rhs>::type RhsNested;
+    typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+    typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+
+    EIGEN_DEVICE_FUNC Product(const Lhs& lhs, const Rhs& rhs) : m_lhs(lhs), m_rhs(rhs)
+    {
+      eigen_assert(lhs.cols() == rhs.rows()
+        && "invalid matrix product"
+        && "if you wanted a coeff-wise or a dot product use the respective explicit functions");
+    }
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
+
+    EIGEN_DEVICE_FUNC const LhsNestedCleaned& lhs() const { return m_lhs; }
+    EIGEN_DEVICE_FUNC const RhsNestedCleaned& rhs() const { return m_rhs; }
+
+  protected:
+
+    LhsNested m_lhs;
+    RhsNested m_rhs;
+};
+
+namespace internal {
+  
+template<typename Lhs, typename Rhs, int Option, int ProductTag = internal::product_type<Lhs,Rhs>::ret>
+class dense_product_base
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{};
+
+/** Convertion to scalar for inner-products */
+template<typename Lhs, typename Rhs, int Option>
+class dense_product_base<Lhs, Rhs, Option, InnerProduct>
+ : public internal::dense_xpr_base<Product<Lhs,Rhs,Option> >::type
+{
+  typedef Product<Lhs,Rhs,Option> ProductXpr;
+  typedef typename internal::dense_xpr_base<ProductXpr>::type Base;
+public:
+  using Base::derived;
+  typedef typename Base::Scalar Scalar;
+  
+  operator const Scalar() const
+  {
+    return internal::evaluator<ProductXpr>(derived()).coeff(0,0);
+  }
+};
+
+} // namespace internal
+
+// Generic API dispatcher
+template<typename Lhs, typename Rhs, int Option, typename StorageKind>
+class ProductImpl : public internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Product<Lhs,Rhs,Option>, MatrixXpr, StorageKind>::type Base;
+};
+
+template<typename Lhs, typename Rhs, int Option>
+class ProductImpl<Lhs,Rhs,Option,Dense>
+  : public internal::dense_product_base<Lhs,Rhs,Option>
+{
+    typedef Product<Lhs, Rhs, Option> Derived;
+    
+  public:
+    
+    typedef typename internal::dense_product_base<Lhs, Rhs, Option> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+  protected:
+    enum {
+      IsOneByOne = (RowsAtCompileTime == 1 || RowsAtCompileTime == Dynamic) && 
+                   (ColsAtCompileTime == 1 || ColsAtCompileTime == Dynamic),
+      EnableCoeff = IsOneByOne || Option==LazyProduct
+    };
+    
+  public:
+  
+    EIGEN_DEVICE_FUNC Scalar coeff(Index row, Index col) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(row,col);
+    }
+
+    EIGEN_DEVICE_FUNC Scalar coeff(Index i) const
+    {
+      EIGEN_STATIC_ASSERT(EnableCoeff, THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS);
+      eigen_assert( (Option==LazyProduct) || (this->rows() == 1 && this->cols() == 1) );
+      
+      return internal::evaluator<Derived>(derived()).coeff(i);
+    }
+    
+  
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_PRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h
new file mode 100644
index 0000000..c42725d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ProductEvaluators.h
@@ -0,0 +1,1105 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2011 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_PRODUCTEVALUATORS_H
+#define EIGEN_PRODUCTEVALUATORS_H
+
+namespace Eigen {
+  
+namespace internal {
+
+/** \internal
+  * Evaluator of a product expression.
+  * Since products require special treatments to handle all possible cases,
+  * we simply deffer the evaluation logic to a product_evaluator class
+  * which offers more partial specialization possibilities.
+  * 
+  * \sa class product_evaluator
+  */
+template<typename Lhs, typename Rhs, int Options>
+struct evaluator<Product<Lhs, Rhs, Options> > 
+ : public product_evaluator<Product<Lhs, Rhs, Options> >
+{
+  typedef Product<Lhs, Rhs, Options> XprType;
+  typedef product_evaluator<XprType> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr) : Base(xpr) {}
+};
+ 
+// Catch "scalar * ( A * B )" and transform it to "(A*scalar) * B"
+// TODO we should apply that rule only if that's really helpful
+template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                                               const Product<Lhs, Rhs, DefaultProduct> > >
+{
+  static const bool value = true;
+};
+template<typename Lhs, typename Rhs, typename Scalar1, typename Scalar2, typename Plain1>
+struct evaluator<CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                               const Product<Lhs, Rhs, DefaultProduct> > >
+ : public evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> >
+{
+  typedef CwiseBinaryOp<internal::scalar_product_op<Scalar1,Scalar2>,
+                               const CwiseNullaryOp<internal::scalar_constant_op<Scalar1>, Plain1>,
+                               const Product<Lhs, Rhs, DefaultProduct> > XprType;
+  typedef evaluator<Product<EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar1,Lhs,product), Rhs, DefaultProduct> > Base;
+
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : Base(xpr.lhs().functor().m_other * xpr.rhs().lhs() * xpr.rhs().rhs())
+  {}
+};
+
+
+template<typename Lhs, typename Rhs, int DiagIndex>
+struct evaluator<Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> > 
+ : public evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> >
+{
+  typedef Diagonal<const Product<Lhs, Rhs, DefaultProduct>, DiagIndex> XprType;
+  typedef evaluator<Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex> > Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : Base(Diagonal<const Product<Lhs, Rhs, LazyProduct>, DiagIndex>(
+        Product<Lhs, Rhs, LazyProduct>(xpr.nestedExpression().lhs(), xpr.nestedExpression().rhs()),
+        xpr.index() ))
+  {}
+};
+
+
+// Helper class to perform a matrix product with the destination at hand.
+// Depending on the sizes of the factors, there are different evaluation strategies
+// as controlled by internal::product_type.
+template< typename Lhs, typename Rhs,
+          typename LhsShape = typename evaluator_traits<Lhs>::Shape,
+          typename RhsShape = typename evaluator_traits<Rhs>::Shape,
+          int ProductType = internal::product_type<Lhs,Rhs>::value>
+struct generic_product_impl;
+
+template<typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<Product<Lhs, Rhs, DefaultProduct> > {
+  static const bool value = true;
+};
+
+// This is the default evaluator implementation for products:
+// It creates a temporary and call generic_product_impl
+template<typename Lhs, typename Rhs, int Options, int ProductTag, typename LhsShape, typename RhsShape>
+struct product_evaluator<Product<Lhs, Rhs, Options>, ProductTag, LhsShape, RhsShape>
+  : public evaluator<typename Product<Lhs, Rhs, Options>::PlainObject>
+{
+  typedef Product<Lhs, Rhs, Options> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit product_evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    
+// FIXME shall we handle nested_eval here?,
+// if so, then we must take care at removing the call to nested_eval in the specializations (e.g., in permutation_matrix_product, transposition_matrix_product, etc.)
+//     typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+//     typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+//     typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+//     typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+//     
+//     const LhsNested lhs(xpr.lhs());
+//     const RhsNested rhs(xpr.rhs());
+//   
+//     generic_product_impl<LhsNestedCleaned, RhsNestedCleaned>::evalTo(m_result, lhs, rhs);
+
+    generic_product_impl<Lhs, Rhs, LhsShape, RhsShape, ProductTag>::evalTo(m_result, xpr.lhs(), xpr.rhs());
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// The following three shortcuts are enabled only if the scalar types match excatly.
+// TODO: we could enable them for different scalar types when the product is not vectorized.
+
+// Dense = Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::evalTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+// Dense += Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::add_assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,Scalar> &)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::addTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+// Dense -= Product
+template< typename DstXprType, typename Lhs, typename Rhs, int Options, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,Options>, internal::sub_assign_op<Scalar,Scalar>, Dense2Dense,
+  typename enable_if<(Options==DefaultProduct || Options==AliasFreeProduct)>::type>
+{
+  typedef Product<Lhs,Rhs,Options> SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,Scalar> &)
+  {
+    eigen_assert(dst.rows() == src.rows() && dst.cols() == src.cols());
+    // FIXME shall we handle nested_eval here?
+    generic_product_impl<Lhs, Rhs>::subTo(dst, src.lhs(), src.rhs());
+  }
+};
+
+
+// Dense ?= scalar * Product
+// TODO we should apply that rule if that's really helpful
+// for instance, this is not good for inner products
+template< typename DstXprType, typename Lhs, typename Rhs, typename AssignFunc, typename Scalar, typename ScalarBis, typename Plain>
+struct Assignment<DstXprType, CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>, const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>,
+                                           const Product<Lhs,Rhs,DefaultProduct> >, AssignFunc, Dense2Dense>
+{
+  typedef CwiseBinaryOp<internal::scalar_product_op<ScalarBis,Scalar>,
+                        const CwiseNullaryOp<internal::scalar_constant_op<ScalarBis>,Plain>,
+                        const Product<Lhs,Rhs,DefaultProduct> > SrcXprType;
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const AssignFunc& func)
+  {
+    call_assignment_no_alias(dst, (src.lhs().functor().m_other * src.rhs().lhs())*src.rhs().rhs(), func);
+  }
+};
+
+//----------------------------------------
+// Catch "Dense ?= xpr + Product<>" expression to save one temporary
+// FIXME we could probably enable these rules for any product, i.e., not only Dense and DefaultProduct
+
+template<typename OtherXpr, typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_sum_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr,
+                                               const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > {
+  static const bool value = true;
+};
+
+template<typename OtherXpr, typename Lhs, typename Rhs>
+struct evaluator_assume_aliasing<CwiseBinaryOp<internal::scalar_difference_op<typename OtherXpr::Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, const OtherXpr,
+                                               const Product<Lhs,Rhs,DefaultProduct> >, DenseShape > {
+  static const bool value = true;
+};
+
+template<typename DstXprType, typename OtherXpr, typename ProductType, typename Func1, typename Func2>
+struct assignment_from_xpr_op_product
+{
+  template<typename SrcXprType, typename InitialFunc>
+  static EIGEN_STRONG_INLINE
+  void run(DstXprType &dst, const SrcXprType &src, const InitialFunc& /*func*/)
+  {
+    call_assignment_no_alias(dst, src.lhs(), Func1());
+    call_assignment_no_alias(dst, src.rhs(), Func2());
+  }
+};
+
+#define EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(ASSIGN_OP,BINOP,ASSIGN_OP2) \
+  template< typename DstXprType, typename OtherXpr, typename Lhs, typename Rhs, typename DstScalar, typename SrcScalar, typename OtherScalar,typename ProdScalar> \
+  struct Assignment<DstXprType, CwiseBinaryOp<internal::BINOP<OtherScalar,ProdScalar>, const OtherXpr, \
+                                            const Product<Lhs,Rhs,DefaultProduct> >, internal::ASSIGN_OP<DstScalar,SrcScalar>, Dense2Dense> \
+    : assignment_from_xpr_op_product<DstXprType, OtherXpr, Product<Lhs,Rhs,DefaultProduct>, internal::ASSIGN_OP<DstScalar,OtherScalar>, internal::ASSIGN_OP2<DstScalar,ProdScalar> > \
+  {}
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_sum_op,add_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_sum_op,sub_assign_op);
+
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(assign_op,    scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(add_assign_op,scalar_difference_op,sub_assign_op);
+EIGEN_CATCH_ASSIGN_XPR_OP_PRODUCT(sub_assign_op,scalar_difference_op,add_assign_op);
+
+//----------------------------------------
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,InnerProduct>
+{
+  template<typename Dst>
+  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    dst.coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
+  }
+  
+  template<typename Dst>
+  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    dst.coeffRef(0,0) += (lhs.transpose().cwiseProduct(rhs)).sum();
+  }
+  
+  template<typename Dst>
+  static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { dst.coeffRef(0,0) -= (lhs.transpose().cwiseProduct(rhs)).sum(); }
+};
+
+
+/***********************************************************************
+*  Implementation of outer dense * dense vector product
+***********************************************************************/
+
+// Column major result
+template<typename Dst, typename Lhs, typename Rhs, typename Func>
+void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const false_type&)
+{
+  evaluator<Rhs> rhsEval(rhs);
+  typename nested_eval<Lhs,Rhs::SizeAtCompileTime>::type actual_lhs(lhs);
+  // FIXME if cols is large enough, then it might be useful to make sure that lhs is sequentially stored
+  // FIXME not very good if rhs is real and lhs complex while alpha is real too
+  const Index cols = dst.cols();
+  for (Index j=0; j<cols; ++j)
+    func(dst.col(j), rhsEval.coeff(Index(0),j) * actual_lhs);
+}
+
+// Row major result
+template<typename Dst, typename Lhs, typename Rhs, typename Func>
+void outer_product_selector_run(Dst& dst, const Lhs &lhs, const Rhs &rhs, const Func& func, const true_type&)
+{
+  evaluator<Lhs> lhsEval(lhs);
+  typename nested_eval<Rhs,Lhs::SizeAtCompileTime>::type actual_rhs(rhs);
+  // FIXME if rows is large enough, then it might be useful to make sure that rhs is sequentially stored
+  // FIXME not very good if lhs is real and rhs complex while alpha is real too
+  const Index rows = dst.rows();
+  for (Index i=0; i<rows; ++i)
+    func(dst.row(i), lhsEval.coeff(i,Index(0)) * actual_rhs);
+}
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,OuterProduct>
+{
+  template<typename T> struct is_row_major : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  // TODO it would be nice to be able to exploit our *_assign_op functors for that purpose
+  struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
+  struct add  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() += src; } };
+  struct sub  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived() -= src; } };
+  struct adds {
+    Scalar m_scale;
+    explicit adds(const Scalar& s) : m_scale(s) {}
+    template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const {
+      dst.const_cast_derived() += m_scale * src;
+    }
+  };
+  
+  template<typename Dst>
+  static inline void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, set(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, add(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, sub(), is_row_major<Dst>());
+  }
+  
+  template<typename Dst>
+  static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    internal::outer_product_selector_run(dst, lhs, rhs, adds(alpha), is_row_major<Dst>());
+  }
+  
+};
+
+
+// This base class provides default implementations for evalTo, addTo, subTo, in terms of scaleAndAddTo
+template<typename Lhs, typename Rhs, typename Derived>
+struct generic_product_impl_base
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { dst.setZero(); scaleAndAddTo(dst, lhs, rhs, Scalar(1)); }
+
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { scaleAndAddTo(dst,lhs, rhs, Scalar(1)); }
+
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  { scaleAndAddTo(dst, lhs, rhs, Scalar(-1)); }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  { Derived::scaleAndAddTo(dst,lhs,rhs,alpha); }
+
+};
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemvProduct> >
+{
+  typedef typename nested_eval<Lhs,1>::type LhsNested;
+  typedef typename nested_eval<Rhs,1>::type RhsNested;
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  enum { Side = Lhs::IsVectorAtCompileTime ? OnTheLeft : OnTheRight };
+  typedef typename internal::remove_all<typename internal::conditional<int(Side)==OnTheRight,LhsNested,RhsNested>::type>::type MatrixType;
+
+  template<typename Dest>
+  static EIGEN_STRONG_INLINE void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    LhsNested actual_lhs(lhs);
+    RhsNested actual_rhs(rhs);
+    internal::gemv_dense_selector<Side,
+                            (int(MatrixType::Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                            bool(internal::blas_traits<MatrixType>::HasUsableDirectAccess)
+                           >::run(actual_lhs, actual_rhs, dst, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> 
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // Same as: dst.noalias() = lhs.lazyProduct(rhs);
+    // but easier on the compiler side
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // dst.noalias() += lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::add_assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+  template<typename Dst>
+  static EIGEN_STRONG_INLINE void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    // dst.noalias() -= lhs.lazyProduct(rhs);
+    call_assignment_no_alias(dst, lhs.lazyProduct(rhs), internal::sub_assign_op<typename Dst::Scalar,Scalar>());
+  }
+  
+//   template<typename Dst>
+//   static inline void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+//   { dst.noalias() += alpha * lhs.lazyProduct(rhs); }
+};
+
+// This specialization enforces the use of a coefficient-based evaluation strategy
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,LazyCoeffBasedProductMode>
+  : generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> {};
+
+// Case 2: Evaluate coeff by coeff
+//
+// This is mostly taken from CoeffBasedProduct.h
+// The main difference is that we add an extra argument to the etor_product_*_impl::run() function
+// for the inner dimension of the product, because evaluator object do not know their size.
+
+template<int Traversal, int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
+struct etor_product_coeff_impl;
+
+template<int StorageOrder, int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, DenseShape>
+    : evaluator_base<Product<Lhs, Rhs, LazyProduct> >
+{
+  typedef Product<Lhs, Rhs, LazyProduct> XprType;
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  explicit product_evaluator(const XprType& xpr)
+    : m_lhs(xpr.lhs()),
+      m_rhs(xpr.rhs()),
+      m_lhsImpl(m_lhs),     // FIXME the creation of the evaluator objects should result in a no-op, but check that!
+      m_rhsImpl(m_rhs),     //       Moreover, they are only useful for the packet path, so we could completely disable them when not needed,
+                            //       or perhaps declare them on the fly on the packet method... We have experiment to check what's best.
+      m_innerDim(xpr.lhs().cols())
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::AddCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+#if 0
+    std::cerr << "LhsOuterStrideBytes=  " << LhsOuterStrideBytes << "\n";
+    std::cerr << "RhsOuterStrideBytes=  " << RhsOuterStrideBytes << "\n";
+    std::cerr << "LhsAlignment=         " << LhsAlignment << "\n";
+    std::cerr << "RhsAlignment=         " << RhsAlignment << "\n";
+    std::cerr << "CanVectorizeLhs=      " << CanVectorizeLhs << "\n";
+    std::cerr << "CanVectorizeRhs=      " << CanVectorizeRhs << "\n";
+    std::cerr << "CanVectorizeInner=    " << CanVectorizeInner << "\n";
+    std::cerr << "EvalToRowMajor=       " << EvalToRowMajor << "\n";
+    std::cerr << "Alignment=            " << Alignment << "\n";
+    std::cerr << "Flags=                " << Flags << "\n";
+#endif
+  }
+
+  // Everything below here is taken from CoeffBasedProduct.h
+
+  typedef typename internal::nested_eval<Lhs,Rhs::ColsAtCompileTime>::type LhsNested;
+  typedef typename internal::nested_eval<Rhs,Lhs::RowsAtCompileTime>::type RhsNested;
+  
+  typedef typename internal::remove_all<LhsNested>::type LhsNestedCleaned;
+  typedef typename internal::remove_all<RhsNested>::type RhsNestedCleaned;
+
+  typedef evaluator<LhsNestedCleaned> LhsEtorType;
+  typedef evaluator<RhsNestedCleaned> RhsEtorType;
+
+  enum {
+    RowsAtCompileTime = LhsNestedCleaned::RowsAtCompileTime,
+    ColsAtCompileTime = RhsNestedCleaned::ColsAtCompileTime,
+    InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(LhsNestedCleaned::ColsAtCompileTime, RhsNestedCleaned::RowsAtCompileTime),
+    MaxRowsAtCompileTime = LhsNestedCleaned::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = RhsNestedCleaned::MaxColsAtCompileTime
+  };
+
+  typedef typename find_best_packet<Scalar,RowsAtCompileTime>::type LhsVecPacketType;
+  typedef typename find_best_packet<Scalar,ColsAtCompileTime>::type RhsVecPacketType;
+
+  enum {
+      
+    LhsCoeffReadCost = LhsEtorType::CoeffReadCost,
+    RhsCoeffReadCost = RhsEtorType::CoeffReadCost,
+    CoeffReadCost = InnerSize==0 ? NumTraits<Scalar>::ReadCost
+                  : InnerSize == Dynamic ? HugeCost
+                  : InnerSize * (NumTraits<Scalar>::MulCost + LhsCoeffReadCost + RhsCoeffReadCost)
+                    + (InnerSize - 1) * NumTraits<Scalar>::AddCost,
+
+    Unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT,
+    
+    LhsFlags = LhsEtorType::Flags,
+    RhsFlags = RhsEtorType::Flags,
+    
+    LhsRowMajor = LhsFlags & RowMajorBit,
+    RhsRowMajor = RhsFlags & RowMajorBit,
+
+    LhsVecPacketSize = unpacket_traits<LhsVecPacketType>::size,
+    RhsVecPacketSize = unpacket_traits<RhsVecPacketType>::size,
+
+    // Here, we don't care about alignment larger than the usable packet size.
+    LhsAlignment = EIGEN_PLAIN_ENUM_MIN(LhsEtorType::Alignment,LhsVecPacketSize*int(sizeof(typename LhsNestedCleaned::Scalar))),
+    RhsAlignment = EIGEN_PLAIN_ENUM_MIN(RhsEtorType::Alignment,RhsVecPacketSize*int(sizeof(typename RhsNestedCleaned::Scalar))),
+      
+    SameType = is_same<typename LhsNestedCleaned::Scalar,typename RhsNestedCleaned::Scalar>::value,
+
+    CanVectorizeRhs = bool(RhsRowMajor) && (RhsFlags & PacketAccessBit) && (ColsAtCompileTime!=1),
+    CanVectorizeLhs = (!LhsRowMajor) && (LhsFlags & PacketAccessBit) && (RowsAtCompileTime!=1),
+
+    EvalToRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+                    : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+                    : (bool(RhsRowMajor) && !CanVectorizeLhs),
+
+    Flags = ((unsigned int)(LhsFlags | RhsFlags) & HereditaryBits & ~RowMajorBit)
+          | (EvalToRowMajor ? RowMajorBit : 0)
+          // TODO enable vectorization for mixed types
+          | (SameType && (CanVectorizeLhs || CanVectorizeRhs) ? PacketAccessBit : 0)
+          | (XprType::IsVectorAtCompileTime ? LinearAccessBit : 0),
+          
+    LhsOuterStrideBytes = int(LhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename LhsNestedCleaned::Scalar)),
+    RhsOuterStrideBytes = int(RhsNestedCleaned::OuterStrideAtCompileTime) * int(sizeof(typename RhsNestedCleaned::Scalar)),
+
+    Alignment = bool(CanVectorizeLhs) ? (LhsOuterStrideBytes<=0 || (int(LhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,LhsAlignment))!=0 ? 0 : LhsAlignment)
+              : bool(CanVectorizeRhs) ? (RhsOuterStrideBytes<=0 || (int(RhsOuterStrideBytes) % EIGEN_PLAIN_ENUM_MAX(1,RhsAlignment))!=0 ? 0 : RhsAlignment)
+              : 0,
+
+    /* CanVectorizeInner deserves special explanation. It does not affect the product flags. It is not used outside
+     * of Product. If the Product itself is not a packet-access expression, there is still a chance that the inner
+     * loop of the product might be vectorized. This is the meaning of CanVectorizeInner. Since it doesn't affect
+     * the Flags, it is safe to make this value depend on ActualPacketAccessBit, that doesn't affect the ABI.
+     */
+    CanVectorizeInner =    SameType
+                        && LhsRowMajor
+                        && (!RhsRowMajor)
+                        && (LhsFlags & RhsFlags & ActualPacketAccessBit)
+                        && (InnerSize % packet_traits<Scalar>::size == 0)
+  };
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CoeffReturnType coeff(Index row, Index col) const
+  {
+    return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
+  }
+
+  /* Allow index-based non-packet access. It is impossible though to allow index-based packed access,
+   * which is why we don't set the LinearAccessBit.
+   * TODO: this seems possible when the result is a vector
+   */
+  EIGEN_DEVICE_FUNC const CoeffReturnType coeff(Index index) const
+  {
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
+    return (m_lhs.row(row).transpose().cwiseProduct( m_rhs.col(col) )).sum();
+  }
+
+  template<int LoadMode, typename PacketType>
+  const PacketType packet(Index row, Index col) const
+  {
+    PacketType res;
+    typedef etor_product_packet_impl<bool(int(Flags)&RowMajorBit) ? RowMajor : ColMajor,
+                                     Unroll ? int(InnerSize) : Dynamic,
+                                     LhsEtorType, RhsEtorType, PacketType, LoadMode> PacketImpl;
+    PacketImpl::run(row, col, m_lhsImpl, m_rhsImpl, m_innerDim, res);
+    return res;
+  }
+
+  template<int LoadMode, typename PacketType>
+  const PacketType packet(Index index) const
+  {
+    const Index row = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? 0 : index;
+    const Index col = (RowsAtCompileTime == 1 || MaxRowsAtCompileTime==1) ? index : 0;
+    return packet<LoadMode,PacketType>(row,col);
+  }
+
+protected:
+  typename internal::add_const_on_value_type<LhsNested>::type m_lhs;
+  typename internal::add_const_on_value_type<RhsNested>::type m_rhs;
+  
+  LhsEtorType m_lhsImpl;
+  RhsEtorType m_rhsImpl;
+
+  // TODO: Get rid of m_innerDim if known at compile time
+  Index m_innerDim;
+};
+
+template<typename Lhs, typename Rhs>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, LazyCoeffBasedProductMode, DenseShape, DenseShape>
+  : product_evaluator<Product<Lhs, Rhs, LazyProduct>, CoeffBasedProductMode, DenseShape, DenseShape>
+{
+  typedef Product<Lhs, Rhs, DefaultProduct> XprType;
+  typedef Product<Lhs, Rhs, LazyProduct> BaseProduct;
+  typedef product_evaluator<BaseProduct, CoeffBasedProductMode, DenseShape, DenseShape> Base;
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(BaseProduct(xpr.lhs(),xpr.rhs()))
+  {}
+};
+
+/****************************************
+*** Coeff based product, Packet path  ***
+****************************************/
+
+template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
+  {
+    etor_product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
+    res =  pmadd(pset1<Packet>(lhs.coeff(row, Index(UnrollingIndex-1))), rhs.template packet<LoadMode,Packet>(Index(UnrollingIndex-1), col), res);
+  }
+};
+
+template<int UnrollingIndex, typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet &res)
+  {
+    etor_product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, innerDim, res);
+    res =  pmadd(lhs.template packet<LoadMode,Packet>(row, Index(UnrollingIndex-1)), pset1<Packet>(rhs.coeff(Index(UnrollingIndex-1), col)), res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
+  {
+    res = pmul(pset1<Packet>(lhs.coeff(row, Index(0))),rhs.template packet<LoadMode,Packet>(Index(0), col));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index /*innerDim*/, Packet &res)
+  {
+    res = pmul(lhs.template packet<LoadMode,Packet>(row, Index(0)), pset1<Packet>(rhs.coeff(Index(0), col)));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Index /*innerDim*/, Packet &res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+    for(Index i = 0; i < innerDim; ++i)
+      res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode,Packet>(i, col), res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct etor_product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
+{
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Index innerDim, Packet& res)
+  {
+    res = pset1<Packet>(typename unpacket_traits<Packet>::type(0));
+    for(Index i = 0; i < innerDim; ++i)
+      res =  pmadd(lhs.template packet<LoadMode,Packet>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
+  }
+};
+
+
+/***************************************************************************
+* Triangular products
+***************************************************************************/
+template<int Mode, bool LhsIsTriangular,
+         typename Lhs, bool LhsIsVector,
+         typename Rhs, bool RhsIsVector>
+struct triangular_product_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,TriangularShape,DenseShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    triangular_product_impl<Lhs::Mode,true,typename Lhs::MatrixType,false,Rhs, Rhs::ColsAtCompileTime==1>
+        ::run(dst, lhs.nestedExpression(), rhs, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag>
+: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,TriangularShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    triangular_product_impl<Rhs::Mode,false,Lhs,Lhs::RowsAtCompileTime==1, typename Rhs::MatrixType, false>::run(dst, lhs, rhs.nestedExpression(), alpha);
+  }
+};
+
+
+/***************************************************************************
+* SelfAdjoint products
+***************************************************************************/
+template <typename Lhs, int LhsMode, bool LhsIsVector,
+          typename Rhs, int RhsMode, bool RhsIsVector>
+struct selfadjoint_product_impl;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,SelfAdjointShape,DenseShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    selfadjoint_product_impl<typename Lhs::MatrixType,Lhs::Mode,false,Rhs,0,Rhs::IsVectorAtCompileTime>::run(dst, lhs.nestedExpression(), rhs, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag>
+: generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,SelfAdjointShape,ProductTag> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    selfadjoint_product_impl<Lhs,0,Lhs::IsVectorAtCompileTime,typename Rhs::MatrixType,Rhs::Mode,false>::run(dst, lhs, rhs.nestedExpression(), alpha);
+  }
+};
+
+
+/***************************************************************************
+* Diagonal products
+***************************************************************************/
+  
+template<typename MatrixType, typename DiagonalType, typename Derived, int ProductOrder>
+struct diagonal_product_evaluator_base
+  : evaluator_base<Derived>
+{
+   typedef typename ScalarBinaryOpTraits<typename MatrixType::Scalar, typename DiagonalType::Scalar>::ReturnType Scalar;
+public:
+  enum {
+    CoeffReadCost = NumTraits<Scalar>::MulCost + evaluator<MatrixType>::CoeffReadCost + evaluator<DiagonalType>::CoeffReadCost,
+    
+    MatrixFlags = evaluator<MatrixType>::Flags,
+    DiagFlags = evaluator<DiagonalType>::Flags,
+    _StorageOrder = MatrixFlags & RowMajorBit ? RowMajor : ColMajor,
+    _ScalarAccessOnDiag =  !((int(_StorageOrder) == ColMajor && int(ProductOrder) == OnTheLeft)
+                           ||(int(_StorageOrder) == RowMajor && int(ProductOrder) == OnTheRight)),
+    _SameTypes = is_same<typename MatrixType::Scalar, typename DiagonalType::Scalar>::value,
+    // FIXME currently we need same types, but in the future the next rule should be the one
+    //_Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && ((!_PacketOnDiag) || (_SameTypes && bool(int(DiagFlags)&PacketAccessBit))),
+    _Vectorizable = bool(int(MatrixFlags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagFlags)&PacketAccessBit))),
+    _LinearAccessMask = (MatrixType::RowsAtCompileTime==1 || MatrixType::ColsAtCompileTime==1) ? LinearAccessBit : 0,
+    Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixFlags)) | (_Vectorizable ? PacketAccessBit : 0),
+    Alignment = evaluator<MatrixType>::Alignment
+  };
+  
+  diagonal_product_evaluator_base(const MatrixType &mat, const DiagonalType &diag)
+    : m_diagImpl(diag), m_matImpl(mat)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(NumTraits<Scalar>::MulCost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index idx) const
+  {
+    return m_diagImpl.coeff(idx) * m_matImpl.coeff(idx);
+  }
+  
+protected:
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::true_type) const
+  {
+    return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col),
+                          internal::pset1<PacketType>(m_diagImpl.coeff(id)));
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet_impl(Index row, Index col, Index id, internal::false_type) const
+  {
+    enum {
+      InnerSize = (MatrixType::Flags & RowMajorBit) ? MatrixType::ColsAtCompileTime : MatrixType::RowsAtCompileTime,
+      DiagonalPacketLoadMode = EIGEN_PLAIN_ENUM_MIN(LoadMode,((InnerSize%16) == 0) ? int(Aligned16) : int(evaluator<DiagonalType>::Alignment)) // FIXME hardcoded 16!!
+    };
+    return internal::pmul(m_matImpl.template packet<LoadMode,PacketType>(row, col),
+                          m_diagImpl.template packet<DiagonalPacketLoadMode,PacketType>(id));
+  }
+  
+  evaluator<DiagonalType> m_diagImpl;
+  evaluator<MatrixType>   m_matImpl;
+};
+
+// diagonal * dense
+template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DiagonalShape, DenseShape>
+  : diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft>
+{
+  typedef diagonal_product_evaluator_base<Rhs, typename Lhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheLeft> Base;
+  using Base::m_diagImpl;
+  using Base::m_matImpl;
+  using Base::coeff;
+  typedef typename Base::Scalar Scalar;
+  
+  typedef Product<Lhs, Rhs, ProductKind> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  
+  enum {
+    StorageOrder = int(Rhs::Flags) & RowMajorBit ? RowMajor : ColMajor
+  };
+
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.rhs(), xpr.lhs().diagonal())
+  {
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
+  {
+    return m_diagImpl.coeff(row) * m_matImpl.coeff(row, col);
+  }
+  
+#ifndef __CUDACC__
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
+  {
+    // FIXME: NVCC used to complain about the template keyword, but we have to check whether this is still the case.
+    // See also similar calls below.
+    return this->template packet_impl<LoadMode,PacketType>(row,col, row,
+                                 typename internal::conditional<int(StorageOrder)==RowMajor, internal::true_type, internal::false_type>::type());
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index idx) const
+  {
+    return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
+  }
+#endif
+};
+
+// dense * diagonal
+template<typename Lhs, typename Rhs, int ProductKind, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, ProductKind>, ProductTag, DenseShape, DiagonalShape>
+  : diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight>
+{
+  typedef diagonal_product_evaluator_base<Lhs, typename Rhs::DiagonalVectorType, Product<Lhs, Rhs, LazyProduct>, OnTheRight> Base;
+  using Base::m_diagImpl;
+  using Base::m_matImpl;
+  using Base::coeff;
+  typedef typename Base::Scalar Scalar;
+  
+  typedef Product<Lhs, Rhs, ProductKind> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  
+  enum { StorageOrder = int(Lhs::Flags) & RowMajorBit ? RowMajor : ColMajor };
+
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.lhs(), xpr.rhs().diagonal())
+  {
+  }
+  
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar coeff(Index row, Index col) const
+  {
+    return m_matImpl.coeff(row, col) * m_diagImpl.coeff(col);
+  }
+  
+#ifndef __CUDACC__
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index row, Index col) const
+  {
+    return this->template packet_impl<LoadMode,PacketType>(row,col, col,
+                                 typename internal::conditional<int(StorageOrder)==ColMajor, internal::true_type, internal::false_type>::type());
+  }
+  
+  template<int LoadMode,typename PacketType>
+  EIGEN_STRONG_INLINE PacketType packet(Index idx) const
+  {
+    return packet<LoadMode,PacketType>(int(StorageOrder)==ColMajor?idx:0,int(StorageOrder)==ColMajor?0:idx);
+  }
+#endif
+};
+
+/***************************************************************************
+* Products with permutation matrices
+***************************************************************************/
+
+/** \internal
+  * \class permutation_matrix_product
+  * Internal helper class implementing the product between a permutation matrix and a matrix.
+  * This class is specialized for DenseShape below and for SparseShape in SparseCore/SparsePermutation.h
+  */
+template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape>
+struct permutation_matrix_product;
+
+template<typename ExpressionType, int Side, bool Transposed>
+struct permutation_matrix_product<ExpressionType, Side, Transposed, DenseShape>
+{
+    typedef typename nested_eval<ExpressionType, 1>::type MatrixType;
+    typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+
+    template<typename Dest, typename PermutationType>
+    static inline void run(Dest& dst, const PermutationType& perm, const ExpressionType& xpr)
+    {
+      MatrixType mat(xpr);
+      const Index n = Side==OnTheLeft ? mat.rows() : mat.cols();
+      // FIXME we need an is_same for expression that is not sensitive to constness. For instance
+      // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true.
+      //if(is_same<MatrixTypeCleaned,Dest>::value && extract_data(dst) == extract_data(mat))
+      if(is_same_dense(dst, mat))
+      {
+        // apply the permutation inplace
+        Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(perm.size());
+        mask.fill(false);
+        Index r = 0;
+        while(r < perm.size())
+        {
+          // search for the next seed
+          while(r<perm.size() && mask[r]) r++;
+          if(r>=perm.size())
+            break;
+          // we got one, let's follow it until we are back to the seed
+          Index k0 = r++;
+          Index kPrev = k0;
+          mask.coeffRef(k0) = true;
+          for(Index k=perm.indices().coeff(k0); k!=k0; k=perm.indices().coeff(k))
+          {
+                  Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>(dst, k)
+            .swap(Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
+                       (dst,((Side==OnTheLeft) ^ Transposed) ? k0 : kPrev));
+
+            mask.coeffRef(k) = true;
+            kPrev = k;
+          }
+        }
+      }
+      else
+      {
+        for(Index i = 0; i < n; ++i)
+        {
+          Block<Dest, Side==OnTheLeft ? 1 : Dest::RowsAtCompileTime, Side==OnTheRight ? 1 : Dest::ColsAtCompileTime>
+               (dst, ((Side==OnTheLeft) ^ Transposed) ? perm.indices().coeff(i) : i)
+
+          =
+
+          Block<const MatrixTypeCleaned,Side==OnTheLeft ? 1 : MatrixTypeCleaned::RowsAtCompileTime,Side==OnTheRight ? 1 : MatrixTypeCleaned::ColsAtCompileTime>
+               (mat, ((Side==OnTheRight) ^ Transposed) ? perm.indices().coeff(i) : i);
+        }
+      }
+    }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, PermutationShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, MatrixShape, PermutationShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Inverse<Lhs>, Rhs, PermutationShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Inverse<Lhs>& lhs, const Rhs& rhs)
+  {
+    permutation_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Inverse<Rhs>, MatrixShape, PermutationShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Inverse<Rhs>& rhs)
+  {
+    permutation_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs);
+  }
+};
+
+
+/***************************************************************************
+* Products with transpositions matrices
+***************************************************************************/
+
+// FIXME could we unify Transpositions and Permutation into a single "shape"??
+
+/** \internal
+  * \class transposition_matrix_product
+  * Internal helper class implementing the product between a permutation matrix and a matrix.
+  */
+template<typename ExpressionType, int Side, bool Transposed, typename ExpressionShape>
+struct transposition_matrix_product
+{
+  typedef typename nested_eval<ExpressionType, 1>::type MatrixType;
+  typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+  
+  template<typename Dest, typename TranspositionType>
+  static inline void run(Dest& dst, const TranspositionType& tr, const ExpressionType& xpr)
+  {
+    MatrixType mat(xpr);
+    typedef typename TranspositionType::StorageIndex StorageIndex;
+    const Index size = tr.size();
+    StorageIndex j = 0;
+
+    if(!is_same_dense(dst,mat))
+      dst = mat;
+
+    for(Index k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)
+      if(Index(j=tr.coeff(k))!=k)
+      {
+        if(Side==OnTheLeft)        dst.row(k).swap(dst.row(j));
+        else if(Side==OnTheRight)  dst.col(k).swap(dst.col(j));
+      }
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, TranspositionsShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Rhs, OnTheLeft, false, MatrixShape>::run(dst, lhs, rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Rhs, MatrixShape, TranspositionsShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Lhs, OnTheRight, false, MatrixShape>::run(dst, rhs, lhs);
+  }
+};
+
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Transpose<Lhs>, Rhs, TranspositionsShape, MatrixShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Transpose<Lhs>& lhs, const Rhs& rhs)
+  {
+    transposition_matrix_product<Rhs, OnTheLeft, true, MatrixShape>::run(dst, lhs.nestedExpression(), rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductTag, typename MatrixShape>
+struct generic_product_impl<Lhs, Transpose<Rhs>, MatrixShape, TranspositionsShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Transpose<Rhs>& rhs)
+  {
+    transposition_matrix_product<Lhs, OnTheRight, true, MatrixShape>::run(dst, rhs.nestedExpression(), lhs);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PRODUCT_EVALUATORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Random.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Random.h
new file mode 100644
index 0000000..6faf789
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Random.h
@@ -0,0 +1,182 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_RANDOM_H
+#define EIGEN_RANDOM_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Scalar> struct scalar_random_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_random_op)
+  inline const Scalar operator() () const { return random<Scalar>(); }
+};
+
+template<typename Scalar>
+struct functor_traits<scalar_random_op<Scalar> >
+{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false, IsRepeatable = false }; };
+
+} // end namespace internal
+
+/** \returns a random matrix expression
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * The parameters \a rows and \a cols are the number of rows and of columns of
+  * the returned matrix. Must be compatible with this MatrixBase type.
+  *
+  * \not_reentrant
+  * 
+  * This variant is meant to be used for dynamic-size matrix types. For fixed-size types,
+  * it is redundant to pass \a rows and \a cols as arguments, so Random() should be used
+  * instead.
+  * 
+  *
+  * Example: \include MatrixBase_random_int_int.cpp
+  * Output: \verbinclude MatrixBase_random_int_int.out
+  *
+  * This expression has the "evaluate before nesting" flag so that it will be evaluated into
+  * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
+  * behavior with expressions involving random matrices.
+  * 
+  * See DenseBase::NullaryExpr(Index, const CustomNullaryOp&) for an example using C++11 random generators.
+  *
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index), DenseBase::Random()
+  */
+template<typename Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
+DenseBase<Derived>::Random(Index rows, Index cols)
+{
+  return NullaryExpr(rows, cols, internal::scalar_random_op<Scalar>());
+}
+
+/** \returns a random vector expression
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  *
+  * The parameter \a size is the size of the returned vector.
+  * Must be compatible with this MatrixBase type.
+  *
+  * \only_for_vectors
+  * \not_reentrant
+  *
+  * This variant is meant to be used for dynamic-size vector types. For fixed-size types,
+  * it is redundant to pass \a size as argument, so Random() should be used
+  * instead.
+  *
+  * Example: \include MatrixBase_random_int.cpp
+  * Output: \verbinclude MatrixBase_random_int.out
+  *
+  * This expression has the "evaluate before nesting" flag so that it will be evaluated into
+  * a temporary vector whenever it is nested in a larger expression. This prevents unexpected
+  * behavior with expressions involving random matrices.
+  *
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random()
+  */
+template<typename Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
+DenseBase<Derived>::Random(Index size)
+{
+  return NullaryExpr(size, internal::scalar_random_op<Scalar>());
+}
+
+/** \returns a fixed-size random matrix or vector expression
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * This variant is only for fixed-size MatrixBase types. For dynamic-size types, you
+  * need to use the variants taking size arguments.
+  *
+  * Example: \include MatrixBase_random.cpp
+  * Output: \verbinclude MatrixBase_random.out
+  *
+  * This expression has the "evaluate before nesting" flag so that it will be evaluated into
+  * a temporary matrix whenever it is nested in a larger expression. This prevents unexpected
+  * behavior with expressions involving random matrices.
+  * 
+  * \not_reentrant
+  *
+  * \sa DenseBase::setRandom(), DenseBase::Random(Index,Index), DenseBase::Random(Index)
+  */
+template<typename Derived>
+inline const typename DenseBase<Derived>::RandomReturnType
+DenseBase<Derived>::Random()
+{
+  return NullaryExpr(RowsAtCompileTime, ColsAtCompileTime, internal::scalar_random_op<Scalar>());
+}
+
+/** Sets all coefficients in this expression to random values.
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * \not_reentrant
+  * 
+  * Example: \include MatrixBase_setRandom.cpp
+  * Output: \verbinclude MatrixBase_setRandom.out
+  *
+  * \sa class CwiseNullaryOp, setRandom(Index), setRandom(Index,Index)
+  */
+template<typename Derived>
+inline Derived& DenseBase<Derived>::setRandom()
+{
+  return *this = Random(rows(), cols());
+}
+
+/** Resizes to the given \a newSize, and sets all coefficients in this expression to random values.
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  * 
+  * \only_for_vectors
+  * \not_reentrant
+  *
+  * Example: \include Matrix_setRandom_int.cpp
+  * Output: \verbinclude Matrix_setRandom_int.out
+  *
+  * \sa DenseBase::setRandom(), setRandom(Index,Index), class CwiseNullaryOp, DenseBase::Random()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setRandom(Index newSize)
+{
+  resize(newSize);
+  return setRandom();
+}
+
+/** Resizes to the given size, and sets all coefficients in this expression to random values.
+  *
+  * Numbers are uniformly spread through their whole definition range for integer types,
+  * and in the [-1:1] range for floating point scalar types.
+  *
+  * \not_reentrant
+  * 
+  * \param rows the new number of rows
+  * \param cols the new number of columns
+  *
+  * Example: \include Matrix_setRandom_int_int.cpp
+  * Output: \verbinclude Matrix_setRandom_int_int.out
+  *
+  * \sa DenseBase::setRandom(), setRandom(Index), class CwiseNullaryOp, DenseBase::Random()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE Derived&
+PlainObjectBase<Derived>::setRandom(Index rows, Index cols)
+{
+  resize(rows, cols);
+  return setRandom();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_RANDOM_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Redux.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Redux.h
new file mode 100644
index 0000000..b6e8f88
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Redux.h
@@ -0,0 +1,505 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REDUX_H
+#define EIGEN_REDUX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// TODO
+//  * implement other kind of vectorization
+//  * factorize code
+
+/***************************************************************************
+* Part 1 : the logic deciding a strategy for vectorization and unrolling
+***************************************************************************/
+
+template<typename Func, typename Derived>
+struct redux_traits
+{
+public:
+    typedef typename find_best_packet<typename Derived::Scalar,Derived::SizeAtCompileTime>::type PacketType;
+  enum {
+    PacketSize = unpacket_traits<PacketType>::size,
+    InnerMaxSize = int(Derived::IsRowMajor)
+                 ? Derived::MaxColsAtCompileTime
+                 : Derived::MaxRowsAtCompileTime
+  };
+
+  enum {
+    MightVectorize = (int(Derived::Flags)&ActualPacketAccessBit)
+                  && (functor_traits<Func>::PacketAccess),
+    MayLinearVectorize = bool(MightVectorize) && (int(Derived::Flags)&LinearAccessBit),
+    MaySliceVectorize  = bool(MightVectorize) && int(InnerMaxSize)>=3*PacketSize
+  };
+
+public:
+  enum {
+    Traversal = int(MayLinearVectorize) ? int(LinearVectorizedTraversal)
+              : int(MaySliceVectorize)  ? int(SliceVectorizedTraversal)
+                                        : int(DefaultTraversal)
+  };
+
+public:
+  enum {
+    Cost = Derived::SizeAtCompileTime == Dynamic ? HugeCost
+         : Derived::SizeAtCompileTime * Derived::CoeffReadCost + (Derived::SizeAtCompileTime-1) * functor_traits<Func>::Cost,
+    UnrollingLimit = EIGEN_UNROLLING_LIMIT * (int(Traversal) == int(DefaultTraversal) ? 1 : int(PacketSize))
+  };
+
+public:
+  enum {
+    Unrolling = Cost <= UnrollingLimit ? CompleteUnrolling : NoUnrolling
+  };
+  
+#ifdef EIGEN_DEBUG_ASSIGN
+  static void debug()
+  {
+    std::cerr << "Xpr: " << typeid(typename Derived::XprType).name() << std::endl;
+    std::cerr.setf(std::ios::hex, std::ios::basefield);
+    EIGEN_DEBUG_VAR(Derived::Flags)
+    std::cerr.unsetf(std::ios::hex);
+    EIGEN_DEBUG_VAR(InnerMaxSize)
+    EIGEN_DEBUG_VAR(PacketSize)
+    EIGEN_DEBUG_VAR(MightVectorize)
+    EIGEN_DEBUG_VAR(MayLinearVectorize)
+    EIGEN_DEBUG_VAR(MaySliceVectorize)
+    EIGEN_DEBUG_VAR(Traversal)
+    EIGEN_DEBUG_VAR(UnrollingLimit)
+    EIGEN_DEBUG_VAR(Unrolling)
+    std::cerr << std::endl;
+  }
+#endif
+};
+
+/***************************************************************************
+* Part 2 : unrollers
+***************************************************************************/
+
+/*** no vectorization ***/
+
+template<typename Func, typename Derived, int Start, int Length>
+struct redux_novec_unroller
+{
+  enum {
+    HalfLength = Length/2
+  };
+
+  typedef typename Derived::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+  {
+    return func(redux_novec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
+                redux_novec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func));
+  }
+};
+
+template<typename Func, typename Derived, int Start>
+struct redux_novec_unroller<Func, Derived, Start, 1>
+{
+  enum {
+    outer = Start / Derived::InnerSizeAtCompileTime,
+    inner = Start % Derived::InnerSizeAtCompileTime
+  };
+
+  typedef typename Derived::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func&)
+  {
+    return mat.coeffByOuterInner(outer, inner);
+  }
+};
+
+// This is actually dead code and will never be called. It is required
+// to prevent false warnings regarding failed inlining though
+// for 0 length run() will never be called at all.
+template<typename Func, typename Derived, int Start>
+struct redux_novec_unroller<Func, Derived, Start, 0>
+{
+  typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC 
+  static EIGEN_STRONG_INLINE Scalar run(const Derived&, const Func&) { return Scalar(); }
+};
+
+/*** vectorization ***/
+
+template<typename Func, typename Derived, int Start, int Length>
+struct redux_vec_unroller
+{
+  enum {
+    PacketSize = redux_traits<Func, Derived>::PacketSize,
+    HalfLength = Length/2
+  };
+
+  typedef typename Derived::Scalar Scalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+
+  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func& func)
+  {
+    return func.packetOp(
+            redux_vec_unroller<Func, Derived, Start, HalfLength>::run(mat,func),
+            redux_vec_unroller<Func, Derived, Start+HalfLength, Length-HalfLength>::run(mat,func) );
+  }
+};
+
+template<typename Func, typename Derived, int Start>
+struct redux_vec_unroller<Func, Derived, Start, 1>
+{
+  enum {
+    index = Start * redux_traits<Func, Derived>::PacketSize,
+    outer = index / int(Derived::InnerSizeAtCompileTime),
+    inner = index % int(Derived::InnerSizeAtCompileTime),
+    alignment = Derived::Alignment
+  };
+
+  typedef typename Derived::Scalar Scalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+
+  static EIGEN_STRONG_INLINE PacketScalar run(const Derived &mat, const Func&)
+  {
+    return mat.template packetByOuterInner<alignment,PacketScalar>(outer, inner);
+  }
+};
+
+/***************************************************************************
+* Part 3 : implementation of all cases
+***************************************************************************/
+
+template<typename Func, typename Derived,
+         int Traversal = redux_traits<Func, Derived>::Traversal,
+         int Unrolling = redux_traits<Func, Derived>::Unrolling
+>
+struct redux_impl;
+
+template<typename Func, typename Derived>
+struct redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>
+{
+  typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+  {
+    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
+    Scalar res;
+    res = mat.coeffByOuterInner(0, 0);
+    for(Index i = 1; i < mat.innerSize(); ++i)
+      res = func(res, mat.coeffByOuterInner(0, i));
+    for(Index i = 1; i < mat.outerSize(); ++i)
+      for(Index j = 0; j < mat.innerSize(); ++j)
+        res = func(res, mat.coeffByOuterInner(i, j));
+    return res;
+  }
+};
+
+template<typename Func, typename Derived>
+struct redux_impl<Func,Derived, DefaultTraversal, CompleteUnrolling>
+  : public redux_novec_unroller<Func,Derived, 0, Derived::SizeAtCompileTime>
+{};
+
+template<typename Func, typename Derived>
+struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
+{
+  typedef typename Derived::Scalar Scalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+
+  static Scalar run(const Derived &mat, const Func& func)
+  {
+    const Index size = mat.size();
+    
+    const Index packetSize = redux_traits<Func, Derived>::PacketSize;
+    const int packetAlignment = unpacket_traits<PacketScalar>::alignment;
+    enum {
+      alignment0 = (bool(Derived::Flags & DirectAccessBit) && bool(packet_traits<Scalar>::AlignedOnScalar)) ? int(packetAlignment) : int(Unaligned),
+      alignment = EIGEN_PLAIN_ENUM_MAX(alignment0, Derived::Alignment)
+    };
+    const Index alignedStart = internal::first_default_aligned(mat.nestedExpression());
+    const Index alignedSize2 = ((size-alignedStart)/(2*packetSize))*(2*packetSize);
+    const Index alignedSize = ((size-alignedStart)/(packetSize))*(packetSize);
+    const Index alignedEnd2 = alignedStart + alignedSize2;
+    const Index alignedEnd  = alignedStart + alignedSize;
+    Scalar res;
+    if(alignedSize)
+    {
+      PacketScalar packet_res0 = mat.template packet<alignment,PacketScalar>(alignedStart);
+      if(alignedSize>packetSize) // we have at least two packets to partly unroll the loop
+      {
+        PacketScalar packet_res1 = mat.template packet<alignment,PacketScalar>(alignedStart+packetSize);
+        for(Index index = alignedStart + 2*packetSize; index < alignedEnd2; index += 2*packetSize)
+        {
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(index));
+          packet_res1 = func.packetOp(packet_res1, mat.template packet<alignment,PacketScalar>(index+packetSize));
+        }
+
+        packet_res0 = func.packetOp(packet_res0,packet_res1);
+        if(alignedEnd>alignedEnd2)
+          packet_res0 = func.packetOp(packet_res0, mat.template packet<alignment,PacketScalar>(alignedEnd2));
+      }
+      res = func.predux(packet_res0);
+
+      for(Index index = 0; index < alignedStart; ++index)
+        res = func(res,mat.coeff(index));
+
+      for(Index index = alignedEnd; index < size; ++index)
+        res = func(res,mat.coeff(index));
+    }
+    else // too small to vectorize anything.
+         // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
+    {
+      res = mat.coeff(0);
+      for(Index index = 1; index < size; ++index)
+        res = func(res,mat.coeff(index));
+    }
+
+    return res;
+  }
+};
+
+// NOTE: for SliceVectorizedTraversal we simply bypass unrolling
+template<typename Func, typename Derived, int Unrolling>
+struct redux_impl<Func, Derived, SliceVectorizedTraversal, Unrolling>
+{
+  typedef typename Derived::Scalar Scalar;
+  typedef typename redux_traits<Func, Derived>::PacketType PacketType;
+
+  EIGEN_DEVICE_FUNC static Scalar run(const Derived &mat, const Func& func)
+  {
+    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
+    const Index innerSize = mat.innerSize();
+    const Index outerSize = mat.outerSize();
+    enum {
+      packetSize = redux_traits<Func, Derived>::PacketSize
+    };
+    const Index packetedInnerSize = ((innerSize)/packetSize)*packetSize;
+    Scalar res;
+    if(packetedInnerSize)
+    {
+      PacketType packet_res = mat.template packet<Unaligned,PacketType>(0,0);
+      for(Index j=0; j<outerSize; ++j)
+        for(Index i=(j==0?packetSize:0); i<packetedInnerSize; i+=Index(packetSize))
+          packet_res = func.packetOp(packet_res, mat.template packetByOuterInner<Unaligned,PacketType>(j,i));
+
+      res = func.predux(packet_res);
+      for(Index j=0; j<outerSize; ++j)
+        for(Index i=packetedInnerSize; i<innerSize; ++i)
+          res = func(res, mat.coeffByOuterInner(j,i));
+    }
+    else // too small to vectorize anything.
+         // since this is dynamic-size hence inefficient anyway for such small sizes, don't try to optimize.
+    {
+      res = redux_impl<Func, Derived, DefaultTraversal, NoUnrolling>::run(mat, func);
+    }
+
+    return res;
+  }
+};
+
+template<typename Func, typename Derived>
+struct redux_impl<Func, Derived, LinearVectorizedTraversal, CompleteUnrolling>
+{
+  typedef typename Derived::Scalar Scalar;
+
+  typedef typename redux_traits<Func, Derived>::PacketType PacketScalar;
+  enum {
+    PacketSize = redux_traits<Func, Derived>::PacketSize,
+    Size = Derived::SizeAtCompileTime,
+    VectorizedSize = (Size / PacketSize) * PacketSize
+  };
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Scalar run(const Derived &mat, const Func& func)
+  {
+    eigen_assert(mat.rows()>0 && mat.cols()>0 && "you are using an empty matrix");
+    if (VectorizedSize > 0) {
+      Scalar res = func.predux(redux_vec_unroller<Func, Derived, 0, Size / PacketSize>::run(mat,func));
+      if (VectorizedSize != Size)
+        res = func(res,redux_novec_unroller<Func, Derived, VectorizedSize, Size-VectorizedSize>::run(mat,func));
+      return res;
+    }
+    else {
+      return redux_novec_unroller<Func, Derived, 0, Size>::run(mat,func);
+    }
+  }
+};
+
+// evaluator adaptor
+template<typename _XprType>
+class redux_evaluator
+{
+public:
+  typedef _XprType XprType;
+  EIGEN_DEVICE_FUNC explicit redux_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  typedef typename XprType::PacketScalar PacketScalar;
+  typedef typename XprType::PacketReturnType PacketReturnType;
+  
+  enum {
+    MaxRowsAtCompileTime = XprType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = XprType::MaxColsAtCompileTime,
+    // TODO we should not remove DirectAccessBit and rather find an elegant way to query the alignment offset at runtime from the evaluator
+    Flags = evaluator<XprType>::Flags & ~DirectAccessBit,
+    IsRowMajor = XprType::IsRowMajor,
+    SizeAtCompileTime = XprType::SizeAtCompileTime,
+    InnerSizeAtCompileTime = XprType::InnerSizeAtCompileTime,
+    CoeffReadCost = evaluator<XprType>::CoeffReadCost,
+    Alignment = evaluator<XprType>::Alignment
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+  EIGEN_DEVICE_FUNC Index innerSize() const { return m_xpr.innerSize(); }
+  EIGEN_DEVICE_FUNC Index outerSize() const { return m_xpr.outerSize(); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeff(Index index) const
+  { return m_evaluator.coeff(index); }
+
+  template<int LoadMode, typename PacketType>
+  PacketType packet(Index row, Index col) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(row, col); }
+
+  template<int LoadMode, typename PacketType>
+  PacketType packet(Index index) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(index); }
+  
+  EIGEN_DEVICE_FUNC
+  CoeffReturnType coeffByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.coeff(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  template<int LoadMode, typename PacketType>
+  PacketType packetByOuterInner(Index outer, Index inner) const
+  { return m_evaluator.template packet<LoadMode,PacketType>(IsRowMajor ? outer : inner, IsRowMajor ? inner : outer); }
+  
+  const XprType & nestedExpression() const { return m_xpr; }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Part 4 : public API
+***************************************************************************/
+
+
+/** \returns the result of a full redux operation on the whole matrix or vector using \a func
+  *
+  * The template parameter \a BinaryOp is the type of the functor \a func which must be
+  * an associative operator. Both current C++98 and C++11 functor styles are handled.
+  *
+  * \sa DenseBase::sum(), DenseBase::minCoeff(), DenseBase::maxCoeff(), MatrixBase::colwise(), MatrixBase::rowwise()
+  */
+template<typename Derived>
+template<typename Func>
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::redux(const Func& func) const
+{
+  eigen_assert(this->rows()>0 && this->cols()>0 && "you are using an empty matrix");
+
+  typedef typename internal::redux_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  return internal::redux_impl<Func, ThisEvaluator>::run(thisEval, func);
+}
+
+/** \returns the minimum of all coefficients of \c *this.
+  * \warning the result is undefined if \c *this contains NaN.
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::minCoeff() const
+{
+  return derived().redux(Eigen::internal::scalar_min_op<Scalar,Scalar>());
+}
+
+/** \returns the maximum of all coefficients of \c *this.
+  * \warning the result is undefined if \c *this contains NaN.
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::maxCoeff() const
+{
+  return derived().redux(Eigen::internal::scalar_max_op<Scalar,Scalar>());
+}
+
+/** \returns the sum of all coefficients of \c *this
+  *
+  * If \c *this is empty, then the value 0 is returned.
+  *
+  * \sa trace(), prod(), mean()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::sum() const
+{
+  if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
+    return Scalar(0);
+  return derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>());
+}
+
+/** \returns the mean of all coefficients of *this
+*
+* \sa trace(), prod(), sum()
+*/
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::mean() const
+{
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  return Scalar(derived().redux(Eigen::internal::scalar_sum_op<Scalar,Scalar>())) / Scalar(this->size());
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
+}
+
+/** \returns the product of all coefficients of *this
+  *
+  * Example: \include MatrixBase_prod.cpp
+  * Output: \verbinclude MatrixBase_prod.out
+  *
+  * \sa sum(), mean(), trace()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::prod() const
+{
+  if(SizeAtCompileTime==0 || (SizeAtCompileTime==Dynamic && size()==0))
+    return Scalar(1);
+  return derived().redux(Eigen::internal::scalar_product_op<Scalar>());
+}
+
+/** \returns the trace of \c *this, i.e. the sum of the coefficients on the main diagonal.
+  *
+  * \c *this can be any matrix, not necessarily square.
+  *
+  * \sa diagonal(), sum()
+  */
+template<typename Derived>
+EIGEN_STRONG_INLINE typename internal::traits<Derived>::Scalar
+MatrixBase<Derived>::trace() const
+{
+  return derived().diagonal().sum();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_REDUX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Ref.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Ref.h
new file mode 100644
index 0000000..bdf24f5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Ref.h
@@ -0,0 +1,281 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REF_H
+#define EIGEN_REF_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename _PlainObjectType, int _Options, typename _StrideType>
+struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
+  : public traits<Map<_PlainObjectType, _Options, _StrideType> >
+{
+  typedef _PlainObjectType PlainObjectType;
+  typedef _StrideType StrideType;
+  enum {
+    Options = _Options,
+    Flags = traits<Map<_PlainObjectType, _Options, _StrideType> >::Flags | NestByRefBit,
+    Alignment = traits<Map<_PlainObjectType, _Options, _StrideType> >::Alignment
+  };
+
+  template<typename Derived> struct match {
+    enum {
+      HasDirectAccess = internal::has_direct_access<Derived>::ret,
+      StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)),
+      InnerStrideMatch = int(StrideType::InnerStrideAtCompileTime)==int(Dynamic)
+                      || int(StrideType::InnerStrideAtCompileTime)==int(Derived::InnerStrideAtCompileTime)
+                      || (int(StrideType::InnerStrideAtCompileTime)==0 && int(Derived::InnerStrideAtCompileTime)==1),
+      OuterStrideMatch = Derived::IsVectorAtCompileTime
+                      || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
+      // NOTE, this indirection of evaluator<Derived>::Alignment is needed
+      // to workaround a very strange bug in MSVC related to the instantiation
+      // of has_*ary_operator in evaluator<CwiseNullaryOp>.
+      // This line is surprisingly very sensitive. For instance, simply adding parenthesis
+      // as "DerivedAlignment = (int(evaluator<Derived>::Alignment))," will make MSVC fail...
+      DerivedAlignment = int(evaluator<Derived>::Alignment),
+      AlignmentMatch = (int(traits<PlainObjectType>::Alignment)==int(Unaligned)) || (DerivedAlignment >= int(Alignment)), // FIXME the first condition is not very clear, it should be replaced by the required alignment
+      ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value,
+      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch
+    };
+    typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
+  };
+  
+};
+
+template<typename Derived>
+struct traits<RefBase<Derived> > : public traits<Derived> {};
+
+}
+
+template<typename Derived> class RefBase
+ : public MapBase<Derived>
+{
+  typedef typename internal::traits<Derived>::PlainObjectType PlainObjectType;
+  typedef typename internal::traits<Derived>::StrideType StrideType;
+
+public:
+
+  typedef MapBase<Derived> Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(RefBase)
+
+  EIGEN_DEVICE_FUNC inline Index innerStride() const
+  {
+    return StrideType::InnerStrideAtCompileTime != 0 ? m_stride.inner() : 1;
+  }
+
+  EIGEN_DEVICE_FUNC inline Index outerStride() const
+  {
+    return StrideType::OuterStrideAtCompileTime != 0 ? m_stride.outer()
+         : IsVectorAtCompileTime ? this->size()
+         : int(Flags)&RowMajorBit ? this->cols()
+         : this->rows();
+  }
+
+  EIGEN_DEVICE_FUNC RefBase()
+    : Base(0,RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime),
+      // Stride<> does not allow default ctor for Dynamic strides, so let' initialize it with dummy values:
+      m_stride(StrideType::OuterStrideAtCompileTime==Dynamic?0:StrideType::OuterStrideAtCompileTime,
+               StrideType::InnerStrideAtCompileTime==Dynamic?0:StrideType::InnerStrideAtCompileTime)
+  {}
+  
+  EIGEN_INHERIT_ASSIGNMENT_OPERATORS(RefBase)
+
+protected:
+
+  typedef Stride<StrideType::OuterStrideAtCompileTime,StrideType::InnerStrideAtCompileTime> StrideBase;
+
+  template<typename Expression>
+  EIGEN_DEVICE_FUNC void construct(Expression& expr)
+  {
+    if(PlainObjectType::RowsAtCompileTime==1)
+    {
+      eigen_assert(expr.rows()==1 || expr.cols()==1);
+      ::new (static_cast<Base*>(this)) Base(expr.data(), 1, expr.size());
+    }
+    else if(PlainObjectType::ColsAtCompileTime==1)
+    {
+      eigen_assert(expr.rows()==1 || expr.cols()==1);
+      ::new (static_cast<Base*>(this)) Base(expr.data(), expr.size(), 1);
+    }
+    else
+      ::new (static_cast<Base*>(this)) Base(expr.data(), expr.rows(), expr.cols());
+    
+    if(Expression::IsVectorAtCompileTime && (!PlainObjectType::IsVectorAtCompileTime) && ((Expression::Flags&RowMajorBit)!=(PlainObjectType::Flags&RowMajorBit)))
+      ::new (&m_stride) StrideBase(expr.innerStride(), StrideType::InnerStrideAtCompileTime==0?0:1);
+    else
+      ::new (&m_stride) StrideBase(StrideType::OuterStrideAtCompileTime==0?0:expr.outerStride(),
+                                   StrideType::InnerStrideAtCompileTime==0?0:expr.innerStride());    
+  }
+
+  StrideBase m_stride;
+};
+
+/** \class Ref
+  * \ingroup Core_Module
+  *
+  * \brief A matrix or vector expression mapping an existing expression
+  *
+  * \tparam PlainObjectType the equivalent matrix type of the mapped data
+  * \tparam Options specifies the pointer alignment in bytes. It can be: \c #Aligned128, , \c #Aligned64, \c #Aligned32, \c #Aligned16, \c #Aligned8 or \c #Unaligned.
+  *                 The default is \c #Unaligned.
+  * \tparam StrideType optionally specifies strides. By default, Ref implies a contiguous storage along the inner dimension (inner stride==1),
+  *                   but accepts a variable outer stride (leading dimension).
+  *                   This can be overridden by specifying strides.
+  *                   The type passed here must be a specialization of the Stride template, see examples below.
+  *
+  * This class provides a way to write non-template functions taking Eigen objects as parameters while limiting the number of copies.
+  * A Ref<> object can represent either a const expression or a l-value:
+  * \code
+  * // in-out argument:
+  * void foo1(Ref<VectorXf> x);
+  *
+  * // read-only const argument:
+  * void foo2(const Ref<const VectorXf>& x);
+  * \endcode
+  *
+  * In the in-out case, the input argument must satisfy the constraints of the actual Ref<> type, otherwise a compilation issue will be triggered.
+  * By default, a Ref<VectorXf> can reference any dense vector expression of float having a contiguous memory layout.
+  * Likewise, a Ref<MatrixXf> can reference any column-major dense matrix expression of float whose column's elements are contiguously stored with
+  * the possibility to have a constant space in-between each column, i.e. the inner stride must be equal to 1, but the outer stride (or leading dimension)
+  * can be greater than the number of rows.
+  *
+  * In the const case, if the input expression does not match the above requirement, then it is evaluated into a temporary before being passed to the function.
+  * Here are some examples:
+  * \code
+  * MatrixXf A;
+  * VectorXf a;
+  * foo1(a.head());             // OK
+  * foo1(A.col());              // OK
+  * foo1(A.row());              // Compilation error because here innerstride!=1
+  * foo2(A.row());              // Compilation error because A.row() is a 1xN object while foo2 is expecting a Nx1 object
+  * foo2(A.row().transpose());  // The row is copied into a contiguous temporary
+  * foo2(2*a);                  // The expression is evaluated into a temporary
+  * foo2(A.col().segment(2,4)); // No temporary
+  * \endcode
+  *
+  * The range of inputs that can be referenced without temporary can be enlarged using the last two template parameters.
+  * Here is an example accepting an innerstride!=1:
+  * \code
+  * // in-out argument:
+  * void foo3(Ref<VectorXf,0,InnerStride<> > x);
+  * foo3(A.row());              // OK
+  * \endcode
+  * The downside here is that the function foo3 might be significantly slower than foo1 because it won't be able to exploit vectorization, and will involve more
+  * expensive address computations even if the input is contiguously stored in memory. To overcome this issue, one might propose to overload internally calling a
+  * template function, e.g.:
+  * \code
+  * // in the .h:
+  * void foo(const Ref<MatrixXf>& A);
+  * void foo(const Ref<MatrixXf,0,Stride<> >& A);
+  *
+  * // in the .cpp:
+  * template<typename TypeOfA> void foo_impl(const TypeOfA& A) {
+  *     ... // crazy code goes here
+  * }
+  * void foo(const Ref<MatrixXf>& A) { foo_impl(A); }
+  * void foo(const Ref<MatrixXf,0,Stride<> >& A) { foo_impl(A); }
+  * \endcode
+  *
+  *
+  * \sa PlainObjectBase::Map(), \ref TopicStorageOrders
+  */
+template<typename PlainObjectType, int Options, typename StrideType> class Ref
+  : public RefBase<Ref<PlainObjectType, Options, StrideType> >
+{
+  private:
+    typedef internal::traits<Ref> Traits;
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(const PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
+  public:
+
+    typedef RefBase<Ref> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
+
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(PlainObjectBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    {
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      Base::construct(expr.derived());
+    }
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
+    #else
+    /** Implicit constructor from any dense expression */
+    template<typename Derived>
+    inline Ref(DenseBase<Derived>& expr)
+    #endif
+    {
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      EIGEN_STATIC_ASSERT(!Derived::IsPlainObjectBase,THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      Base::construct(expr.const_cast_derived());
+    }
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Ref)
+
+};
+
+// this is the const ref version
+template<typename TPlainObjectType, int Options, typename StrideType> class Ref<const TPlainObjectType, Options, StrideType>
+  : public RefBase<Ref<const TPlainObjectType, Options, StrideType> >
+{
+    typedef internal::traits<Ref> Traits;
+  public:
+
+    typedef RefBase<Ref> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
+
+    template<typename Derived>
+    EIGEN_DEVICE_FUNC inline Ref(const DenseBase<Derived>& expr,
+                                 typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
+    {
+//      std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n";
+//      std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n";
+//      std::cout << int(StrideType::InnerStrideAtCompileTime) << " - " << int(Derived::InnerStrideAtCompileTime) << "\n";
+      construct(expr.derived(), typename Traits::template match<Derived>::type());
+    }
+
+    EIGEN_DEVICE_FUNC inline Ref(const Ref& other) : Base(other) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    EIGEN_DEVICE_FUNC inline Ref(const RefBase<OtherRef>& other) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
+
+  protected:
+
+    template<typename Expression>
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr,internal::true_type)
+    {
+      Base::construct(expr);
+    }
+
+    template<typename Expression>
+    EIGEN_DEVICE_FUNC void construct(const Expression& expr, internal::false_type)
+    {
+      internal::call_assignment_no_alias(m_object,expr,internal::assign_op<Scalar,Scalar>());
+      Base::construct(m_object);
+    }
+
+  protected:
+    TPlainObjectType m_object;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_REF_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Replicate.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Replicate.h
new file mode 100644
index 0000000..9960ef8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Replicate.h
@@ -0,0 +1,142 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REPLICATE_H
+#define EIGEN_REPLICATE_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename MatrixType,int RowFactor,int ColFactor>
+struct traits<Replicate<MatrixType,RowFactor,ColFactor> >
+ : traits<MatrixType>
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    RowsAtCompileTime = RowFactor==Dynamic || int(MatrixType::RowsAtCompileTime)==Dynamic
+                      ? Dynamic
+                      : RowFactor * MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = ColFactor==Dynamic || int(MatrixType::ColsAtCompileTime)==Dynamic
+                      ? Dynamic
+                      : ColFactor * MatrixType::ColsAtCompileTime,
+   //FIXME we don't propagate the max sizes !!!
+    MaxRowsAtCompileTime = RowsAtCompileTime,
+    MaxColsAtCompileTime = ColsAtCompileTime,
+    IsRowMajor = MaxRowsAtCompileTime==1 && MaxColsAtCompileTime!=1 ? 1
+               : MaxColsAtCompileTime==1 && MaxRowsAtCompileTime!=1 ? 0
+               : (MatrixType::Flags & RowMajorBit) ? 1 : 0,
+    
+    // FIXME enable DirectAccess with negative strides?
+    Flags = IsRowMajor ? RowMajorBit : 0
+  };
+};
+}
+
+/**
+  * \class Replicate
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the multiple replication of a matrix or vector
+  *
+  * \tparam MatrixType the type of the object we are replicating
+  * \tparam RowFactor number of repetitions at compile time along the vertical direction, can be Dynamic.
+  * \tparam ColFactor number of repetitions at compile time along the horizontal direction, can be Dynamic.
+  *
+  * This class represents an expression of the multiple replication of a matrix or vector.
+  * It is the return type of DenseBase::replicate() and most of the time
+  * this is the only way it is used.
+  *
+  * \sa DenseBase::replicate()
+  */
+template<typename MatrixType,int RowFactor,int ColFactor> class Replicate
+  : public internal::dense_xpr_base< Replicate<MatrixType,RowFactor,ColFactor> >::type
+{
+    typedef typename internal::traits<Replicate>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<Replicate>::_MatrixTypeNested _MatrixTypeNested;
+  public:
+
+    typedef typename internal::dense_xpr_base<Replicate>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Replicate)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
+
+    template<typename OriginalMatrixType>
+    EIGEN_DEVICE_FUNC
+    inline explicit Replicate(const OriginalMatrixType& matrix)
+      : m_matrix(matrix), m_rowFactor(RowFactor), m_colFactor(ColFactor)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
+                          THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
+      eigen_assert(RowFactor!=Dynamic && ColFactor!=Dynamic);
+    }
+
+    template<typename OriginalMatrixType>
+    EIGEN_DEVICE_FUNC
+    inline Replicate(const OriginalMatrixType& matrix, Index rowFactor, Index colFactor)
+      : m_matrix(matrix), m_rowFactor(rowFactor), m_colFactor(colFactor)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<typename internal::remove_const<MatrixType>::type,OriginalMatrixType>::value),
+                          THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE)
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return m_matrix.rows() * m_rowFactor.value(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return m_matrix.cols() * m_colFactor.value(); }
+
+    EIGEN_DEVICE_FUNC
+    const _MatrixTypeNested& nestedExpression() const
+    { 
+      return m_matrix; 
+    }
+
+  protected:
+    MatrixTypeNested m_matrix;
+    const internal::variable_if_dynamic<Index, RowFactor> m_rowFactor;
+    const internal::variable_if_dynamic<Index, ColFactor> m_colFactor;
+};
+
+/**
+  * \return an expression of the replication of \c *this
+  *
+  * Example: \include MatrixBase_replicate.cpp
+  * Output: \verbinclude MatrixBase_replicate.out
+  *
+  * \sa VectorwiseOp::replicate(), DenseBase::replicate(Index,Index), class Replicate
+  */
+template<typename Derived>
+template<int RowFactor, int ColFactor>
+const Replicate<Derived,RowFactor,ColFactor>
+DenseBase<Derived>::replicate() const
+{
+  return Replicate<Derived,RowFactor,ColFactor>(derived());
+}
+
+/**
+  * \return an expression of the replication of each column (or row) of \c *this
+  *
+  * Example: \include DirectionWise_replicate_int.cpp
+  * Output: \verbinclude DirectionWise_replicate_int.out
+  *
+  * \sa VectorwiseOp::replicate(), DenseBase::replicate(), class Replicate
+  */
+template<typename ExpressionType, int Direction>
+const typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
+VectorwiseOp<ExpressionType,Direction>::replicate(Index factor) const
+{
+  return typename VectorwiseOp<ExpressionType,Direction>::ReplicateReturnType
+          (_expression(),Direction==Vertical?factor:1,Direction==Horizontal?factor:1);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_REPLICATE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h
new file mode 100644
index 0000000..c44b767
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/ReturnByValue.h
@@ -0,0 +1,117 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_RETURNBYVALUE_H
+#define EIGEN_RETURNBYVALUE_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename Derived>
+struct traits<ReturnByValue<Derived> >
+  : public traits<typename traits<Derived>::ReturnType>
+{
+  enum {
+    // We're disabling the DirectAccess because e.g. the constructor of
+    // the Block-with-DirectAccess expression requires to have a coeffRef method.
+    // Also, we don't want to have to implement the stride stuff.
+    Flags = (traits<typename traits<Derived>::ReturnType>::Flags
+             | EvalBeforeNestingBit) & ~DirectAccessBit
+  };
+};
+
+/* The ReturnByValue object doesn't even have a coeff() method.
+ * So the only way that nesting it in an expression can work, is by evaluating it into a plain matrix.
+ * So internal::nested always gives the plain return matrix type.
+ *
+ * FIXME: I don't understand why we need this specialization: isn't this taken care of by the EvalBeforeNestingBit ??
+ * Answer: EvalBeforeNestingBit should be deprecated since we have the evaluators
+ */
+template<typename Derived,int n,typename PlainObject>
+struct nested_eval<ReturnByValue<Derived>, n, PlainObject>
+{
+  typedef typename traits<Derived>::ReturnType type;
+};
+
+} // end namespace internal
+
+/** \class ReturnByValue
+  * \ingroup Core_Module
+  *
+  */
+template<typename Derived> class ReturnByValue
+  : public internal::dense_xpr_base< ReturnByValue<Derived> >::type, internal::no_assignment_operator
+{
+  public:
+    typedef typename internal::traits<Derived>::ReturnType ReturnType;
+
+    typedef typename internal::dense_xpr_base<ReturnByValue>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(ReturnByValue)
+
+    template<typename Dest>
+    EIGEN_DEVICE_FUNC
+    inline void evalTo(Dest& dst) const
+    { static_cast<const Derived*>(this)->evalTo(dst); }
+    EIGEN_DEVICE_FUNC inline Index rows() const { return static_cast<const Derived*>(this)->rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return static_cast<const Derived*>(this)->cols(); }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+#define Unusable YOU_ARE_TRYING_TO_ACCESS_A_SINGLE_COEFFICIENT_IN_A_SPECIAL_EXPRESSION_WHERE_THAT_IS_NOT_ALLOWED_BECAUSE_THAT_WOULD_BE_INEFFICIENT
+    class Unusable{
+      Unusable(const Unusable&) {}
+      Unusable& operator=(const Unusable&) {return *this;}
+    };
+    const Unusable& coeff(Index) const { return *reinterpret_cast<const Unusable*>(this); }
+    const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); }
+    Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); }
+    Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); }
+#undef Unusable
+#endif
+};
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
+{
+  other.evalTo(derived());
+  return derived();
+}
+
+namespace internal {
+
+// Expression is evaluated in a temporary; default implementation of Assignment is bypassed so that
+// when a ReturnByValue expression is assigned, the evaluator is not constructed.
+// TODO: Finalize port to new regime; ReturnByValue should not exist in the expression world
+  
+template<typename Derived>
+struct evaluator<ReturnByValue<Derived> >
+  : public evaluator<typename internal::traits<Derived>::ReturnType>
+{
+  typedef ReturnByValue<Derived> XprType;
+  typedef typename internal::traits<Derived>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    xpr.evalTo(m_result);
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_RETURNBYVALUE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Reverse.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Reverse.h
new file mode 100644
index 0000000..0640cda
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Reverse.h
@@ -0,0 +1,211 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Ricard Marxer <email@ricardmarxer.com>
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REVERSE_H
+#define EIGEN_REVERSE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename MatrixType, int Direction>
+struct traits<Reverse<MatrixType, Direction> >
+ : traits<MatrixType>
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    Flags = _MatrixTypeNested::Flags & (RowMajorBit | LvalueBit)
+  };
+};
+
+template<typename PacketType, bool ReversePacket> struct reverse_packet_cond
+{
+  static inline PacketType run(const PacketType& x) { return preverse(x); }
+};
+
+template<typename PacketType> struct reverse_packet_cond<PacketType,false>
+{
+  static inline PacketType run(const PacketType& x) { return x; }
+};
+
+} // end namespace internal 
+
+/** \class Reverse
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the reverse of a vector or matrix
+  *
+  * \tparam MatrixType the type of the object of which we are taking the reverse
+  * \tparam Direction defines the direction of the reverse operation, can be Vertical, Horizontal, or BothDirections
+  *
+  * This class represents an expression of the reverse of a vector.
+  * It is the return type of MatrixBase::reverse() and VectorwiseOp::reverse()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::reverse(), VectorwiseOp::reverse()
+  */
+template<typename MatrixType, int Direction> class Reverse
+  : public internal::dense_xpr_base< Reverse<MatrixType, Direction> >::type
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<Reverse>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Reverse)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
+    using Base::IsRowMajor;
+
+  protected:
+    enum {
+      PacketSize = internal::packet_traits<Scalar>::size,
+      IsColMajor = !IsRowMajor,
+      ReverseRow = (Direction == Vertical)   || (Direction == BothDirections),
+      ReverseCol = (Direction == Horizontal) || (Direction == BothDirections),
+      OffsetRow  = ReverseRow && IsColMajor ? PacketSize : 1,
+      OffsetCol  = ReverseCol && IsRowMajor ? PacketSize : 1,
+      ReversePacket = (Direction == BothDirections)
+                    || ((Direction == Vertical)   && IsColMajor)
+                    || ((Direction == Horizontal) && IsRowMajor)
+    };
+    typedef internal::reverse_packet_cond<PacketScalar,ReversePacket> reverse_packet;
+  public:
+
+    EIGEN_DEVICE_FUNC explicit inline Reverse(const MatrixType& matrix) : m_matrix(matrix) { }
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Reverse)
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols(); }
+
+    EIGEN_DEVICE_FUNC inline Index innerStride() const
+    {
+      return -m_matrix.innerStride();
+    }
+
+    EIGEN_DEVICE_FUNC const typename internal::remove_all<typename MatrixType::Nested>::type&
+    nestedExpression() const 
+    {
+      return m_matrix;
+    }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+};
+
+/** \returns an expression of the reverse of *this.
+  *
+  * Example: \include MatrixBase_reverse.cpp
+  * Output: \verbinclude MatrixBase_reverse.out
+  *
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::ReverseReturnType
+DenseBase<Derived>::reverse()
+{
+  return ReverseReturnType(derived());
+}
+
+
+//reverse const overload moved DenseBase.h due to a CUDA compiler bug
+
+/** This is the "in place" version of reverse: it reverses \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional benefits:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API enables reverse operations without the need for a temporary
+  *  - it allows future optimizations (cache friendliness, etc.)
+  *
+  * \sa VectorwiseOp::reverseInPlace(), reverse() */
+template<typename Derived>
+inline void DenseBase<Derived>::reverseInPlace()
+{
+  if(cols()>rows())
+  {
+    Index half = cols()/2;
+    leftCols(half).swap(rightCols(half).reverse());
+    if((cols()%2)==1)
+    {
+      Index half2 = rows()/2;
+      col(half).head(half2).swap(col(half).tail(half2).reverse());
+    }
+  }
+  else
+  {
+    Index half = rows()/2;
+    topRows(half).swap(bottomRows(half).reverse());
+    if((rows()%2)==1)
+    {
+      Index half2 = cols()/2;
+      row(half).head(half2).swap(row(half).tail(half2).reverse());
+    }
+  }
+}
+
+namespace internal {
+  
+template<int Direction>
+struct vectorwise_reverse_inplace_impl;
+
+template<>
+struct vectorwise_reverse_inplace_impl<Vertical>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.rows()/2;
+    xpr.topRows(half).swap(xpr.bottomRows(half).colwise().reverse());
+  }
+};
+
+template<>
+struct vectorwise_reverse_inplace_impl<Horizontal>
+{
+  template<typename ExpressionType>
+  static void run(ExpressionType &xpr)
+  {
+    Index half = xpr.cols()/2;
+    xpr.leftCols(half).swap(xpr.rightCols(half).rowwise().reverse());
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of VectorwiseOp::reverse: it reverses each column or row of \c *this.
+  *
+  * In most cases it is probably better to simply use the reversed expression
+  * of a matrix. However, when reversing the matrix data itself is really needed,
+  * then this "in-place" version is probably the right choice because it provides
+  * the following additional benefits:
+  *  - less error prone: doing the same operation with .reverse() requires special care:
+  *    \code m = m.reverse().eval(); \endcode
+  *  - this API enables reverse operations without the need for a temporary
+  *
+  * \sa DenseBase::reverseInPlace(), reverse() */
+template<typename ExpressionType, int Direction>
+void VectorwiseOp<ExpressionType,Direction>::reverseInPlace()
+{
+  internal::vectorwise_reverse_inplace_impl<Direction>::run(_expression().const_cast_derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_REVERSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Select.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Select.h
new file mode 100644
index 0000000..79eec1b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Select.h
@@ -0,0 +1,162 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELECT_H
+#define EIGEN_SELECT_H
+
+namespace Eigen { 
+
+/** \class Select
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a coefficient wise version of the C++ ternary operator ?:
+  *
+  * \param ConditionMatrixType the type of the \em condition expression which must be a boolean matrix
+  * \param ThenMatrixType the type of the \em then expression
+  * \param ElseMatrixType the type of the \em else expression
+  *
+  * This class represents an expression of a coefficient wise version of the C++ ternary operator ?:.
+  * It is the return type of DenseBase::select() and most of the time this is the only way it is used.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const
+  */
+
+namespace internal {
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+struct traits<Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >
+ : traits<ThenMatrixType>
+{
+  typedef typename traits<ThenMatrixType>::Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef typename traits<ThenMatrixType>::XprKind XprKind;
+  typedef typename ConditionMatrixType::Nested ConditionMatrixNested;
+  typedef typename ThenMatrixType::Nested ThenMatrixNested;
+  typedef typename ElseMatrixType::Nested ElseMatrixNested;
+  enum {
+    RowsAtCompileTime = ConditionMatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = ConditionMatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = ConditionMatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = ConditionMatrixType::MaxColsAtCompileTime,
+    Flags = (unsigned int)ThenMatrixType::Flags & ElseMatrixType::Flags & RowMajorBit
+  };
+};
+}
+
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType>
+class Select : public internal::dense_xpr_base< Select<ConditionMatrixType, ThenMatrixType, ElseMatrixType> >::type,
+               internal::no_assignment_operator
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<Select>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Select)
+
+    inline EIGEN_DEVICE_FUNC
+    Select(const ConditionMatrixType& a_conditionMatrix,
+           const ThenMatrixType& a_thenMatrix,
+           const ElseMatrixType& a_elseMatrix)
+      : m_condition(a_conditionMatrix), m_then(a_thenMatrix), m_else(a_elseMatrix)
+    {
+      eigen_assert(m_condition.rows() == m_then.rows() && m_condition.rows() == m_else.rows());
+      eigen_assert(m_condition.cols() == m_then.cols() && m_condition.cols() == m_else.cols());
+    }
+
+    inline EIGEN_DEVICE_FUNC Index rows() const { return m_condition.rows(); }
+    inline EIGEN_DEVICE_FUNC Index cols() const { return m_condition.cols(); }
+
+    inline EIGEN_DEVICE_FUNC
+    const Scalar coeff(Index i, Index j) const
+    {
+      if (m_condition.coeff(i,j))
+        return m_then.coeff(i,j);
+      else
+        return m_else.coeff(i,j);
+    }
+
+    inline EIGEN_DEVICE_FUNC
+    const Scalar coeff(Index i) const
+    {
+      if (m_condition.coeff(i))
+        return m_then.coeff(i);
+      else
+        return m_else.coeff(i);
+    }
+
+    inline EIGEN_DEVICE_FUNC const ConditionMatrixType& conditionMatrix() const
+    {
+      return m_condition;
+    }
+
+    inline EIGEN_DEVICE_FUNC const ThenMatrixType& thenMatrix() const
+    {
+      return m_then;
+    }
+
+    inline EIGEN_DEVICE_FUNC const ElseMatrixType& elseMatrix() const
+    {
+      return m_else;
+    }
+
+  protected:
+    typename ConditionMatrixType::Nested m_condition;
+    typename ThenMatrixType::Nested m_then;
+    typename ElseMatrixType::Nested m_else;
+};
+
+
+/** \returns a matrix where each coefficient (i,j) is equal to \a thenMatrix(i,j)
+  * if \c *this(i,j), and \a elseMatrix(i,j) otherwise.
+  *
+  * Example: \include MatrixBase_select.cpp
+  * Output: \verbinclude MatrixBase_select.out
+  *
+  * \sa class Select
+  */
+template<typename Derived>
+template<typename ThenDerived,typename ElseDerived>
+inline const Select<Derived,ThenDerived,ElseDerived>
+DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix,
+                            const DenseBase<ElseDerived>& elseMatrix) const
+{
+  return Select<Derived,ThenDerived,ElseDerived>(derived(), thenMatrix.derived(), elseMatrix.derived());
+}
+
+/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with
+  * the \em else expression being a scalar value.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select
+  */
+template<typename Derived>
+template<typename ThenDerived>
+inline const Select<Derived,ThenDerived, typename ThenDerived::ConstantReturnType>
+DenseBase<Derived>::select(const DenseBase<ThenDerived>& thenMatrix,
+                           const typename ThenDerived::Scalar& elseScalar) const
+{
+  return Select<Derived,ThenDerived,typename ThenDerived::ConstantReturnType>(
+    derived(), thenMatrix.derived(), ThenDerived::Constant(rows(),cols(),elseScalar));
+}
+
+/** Version of DenseBase::select(const DenseBase&, const DenseBase&) with
+  * the \em then expression being a scalar value.
+  *
+  * \sa DenseBase::select(const DenseBase<ThenDerived>&, const DenseBase<ElseDerived>&) const, class Select
+  */
+template<typename Derived>
+template<typename ElseDerived>
+inline const Select<Derived, typename ElseDerived::ConstantReturnType, ElseDerived >
+DenseBase<Derived>::select(const typename ElseDerived::Scalar& thenScalar,
+                           const DenseBase<ElseDerived>& elseMatrix) const
+{
+  return Select<Derived,typename ElseDerived::ConstantReturnType,ElseDerived>(
+    derived(), ElseDerived::Constant(rows(),cols(),thenScalar), elseMatrix.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELECT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h
new file mode 100644
index 0000000..504c98f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfAdjointView.h
@@ -0,0 +1,350 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINTMATRIX_H
+#define EIGEN_SELFADJOINTMATRIX_H
+
+namespace Eigen { 
+
+/** \class SelfAdjointView
+  * \ingroup Core_Module
+  *
+  *
+  * \brief Expression of a selfadjoint matrix from a triangular part of a dense matrix
+  *
+  * \param MatrixType the type of the dense matrix storing the coefficients
+  * \param TriangularPart can be either \c #Lower or \c #Upper
+  *
+  * This class is an expression of a sefladjoint matrix from a triangular part of a matrix
+  * with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView()
+  * and most of the time this is the only way that it is used.
+  *
+  * \sa class TriangularBase, MatrixBase::selfadjointView()
+  */
+
+namespace internal {
+template<typename MatrixType, unsigned int UpLo>
+struct traits<SelfAdjointView<MatrixType, UpLo> > : traits<MatrixType>
+{
+  typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+  typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+  typedef MatrixType ExpressionType;
+  typedef typename MatrixType::PlainObject FullMatrixType;
+  enum {
+    Mode = UpLo | SelfAdjoint,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags =  MatrixTypeNestedCleaned::Flags & (HereditaryBits|FlagsLvalueBit)
+           & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)) // FIXME these flags should be preserved
+  };
+};
+}
+
+
+template<typename _MatrixType, unsigned int UpLo> class SelfAdjointView
+  : public TriangularBase<SelfAdjointView<_MatrixType, UpLo> >
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    typedef TriangularBase<SelfAdjointView> Base;
+    typedef typename internal::traits<SelfAdjointView>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<SelfAdjointView>::MatrixTypeNestedCleaned MatrixTypeNestedCleaned;
+    typedef MatrixTypeNestedCleaned NestedExpression;
+
+    /** \brief The type of coefficients in this matrix */
+    typedef typename internal::traits<SelfAdjointView>::Scalar Scalar; 
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
+
+    enum {
+      Mode = internal::traits<SelfAdjointView>::Mode,
+      Flags = internal::traits<SelfAdjointView>::Flags,
+      TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0)
+    };
+    typedef typename MatrixType::PlainObject PlainObject;
+
+    EIGEN_DEVICE_FUNC
+    explicit inline SelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    {}
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return m_matrix.rows(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return m_matrix.outerStride(); }
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return m_matrix.innerStride(); }
+
+    /** \sa MatrixBase::coeff()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    EIGEN_DEVICE_FUNC
+    inline Scalar coeff(Index row, Index col) const
+    {
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.coeff(row, col);
+    }
+
+    /** \sa MatrixBase::coeffRef()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(SelfAdjointView);
+      Base::check_coordinates_internal(row, col);
+      return m_matrix.coeffRef(row, col);
+    }
+
+    /** \internal */
+    EIGEN_DEVICE_FUNC
+    const MatrixTypeNestedCleaned& _expression() const { return m_matrix; }
+
+    EIGEN_DEVICE_FUNC
+    const MatrixTypeNestedCleaned& nestedExpression() const { return m_matrix; }
+    EIGEN_DEVICE_FUNC
+    MatrixTypeNestedCleaned& nestedExpression() { return m_matrix; }
+
+    /** Efficient triangular matrix times vector/matrix product */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<SelfAdjointView,OtherDerived>
+    operator*(const MatrixBase<OtherDerived>& rhs) const
+    {
+      return Product<SelfAdjointView,OtherDerived>(*this, rhs.derived());
+    }
+
+    /** Efficient vector/matrix times triangular matrix product */
+    template<typename OtherDerived> friend
+    EIGEN_DEVICE_FUNC
+    const Product<OtherDerived,SelfAdjointView>
+    operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView& rhs)
+    {
+      return Product<OtherDerived,SelfAdjointView>(lhs.derived(),rhs);
+    }
+    
+    friend EIGEN_DEVICE_FUNC
+    const SelfAdjointView<const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(Scalar,MatrixType,product),UpLo>
+    operator*(const Scalar& s, const SelfAdjointView& mat)
+    {
+      return (s*mat.nestedExpression()).template selfadjointView<UpLo>();
+    }
+
+    /** Perform a symmetric rank 2 update of the selfadjoint matrix \c *this:
+      * \f$ this = this + \alpha u v^* + conj(\alpha) v u^* \f$
+      * \returns a reference to \c *this
+      *
+      * The vectors \a u and \c v \b must be column vectors, however they can be
+      * a adjoint expression without any overhead. Only the meaningful triangular
+      * part of the matrix is updated, the rest is left unchanged.
+      *
+      * \sa rankUpdate(const MatrixBase<DerivedU>&, Scalar)
+      */
+    template<typename DerivedU, typename DerivedV>
+    EIGEN_DEVICE_FUNC
+    SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha = Scalar(1));
+
+    /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
+      * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix.
+      *
+      * \returns a reference to \c *this
+      *
+      * Note that to perform \f$ this = this + \alpha ( u^* u ) \f$ you can simply
+      * call this function with u.adjoint().
+      *
+      * \sa rankUpdate(const MatrixBase<DerivedU>&, const MatrixBase<DerivedV>&, Scalar)
+      */
+    template<typename DerivedU>
+    EIGEN_DEVICE_FUNC
+    SelfAdjointView& rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1));
+
+    /** \returns an expression of a triangular view extracted from the current selfadjoint view of a given triangular part
+      *
+      * The parameter \a TriMode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper,
+      * \c #Lower, \c #StrictlyLower, \c #UnitLower.
+      *
+      * If \c TriMode references the same triangular part than \c *this, then this method simply return a \c TriangularView of the nested expression,
+      * otherwise, the nested expression is first transposed, thus returning a \c TriangularView<Transpose<MatrixType>> object.
+      *
+      * \sa MatrixBase::triangularView(), class TriangularView
+      */
+    template<unsigned int TriMode>
+    EIGEN_DEVICE_FUNC
+    typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)),
+                                   TriangularView<MatrixType,TriMode>,
+                                   TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type
+    triangularView() const
+    {
+      typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::ConstTransposeReturnType>::type tmp1(m_matrix);
+      typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)), MatrixType&, typename MatrixType::AdjointReturnType>::type tmp2(tmp1);
+      return typename internal::conditional<(TriMode&(Upper|Lower))==(UpLo&(Upper|Lower)),
+                                   TriangularView<MatrixType,TriMode>,
+                                   TriangularView<typename MatrixType::AdjointReturnType,TriMode> >::type(tmp2);
+    }
+
+    typedef SelfAdjointView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConjugateReturnType conjugate() const
+    { return ConjugateReturnType(m_matrix.conjugate()); }
+
+    typedef SelfAdjointView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType;
+    /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
+    inline const AdjointReturnType adjoint() const
+    { return AdjointReturnType(m_matrix.adjoint()); }
+
+    typedef SelfAdjointView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType;
+     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
+    inline TransposeReturnType transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      typename MatrixType::TransposeReturnType tmp(m_matrix);
+      return TransposeReturnType(tmp);
+    }
+
+    typedef SelfAdjointView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType;
+    /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(m_matrix.transpose());
+    }
+
+    /** \returns a const expression of the main diagonal of the matrix \c *this
+      *
+      * This method simply returns the diagonal of the nested expression, thus by-passing the SelfAdjointView decorator.
+      *
+      * \sa MatrixBase::diagonal(), class Diagonal */
+    EIGEN_DEVICE_FUNC
+    typename MatrixType::ConstDiagonalReturnType diagonal() const
+    {
+      return typename MatrixType::ConstDiagonalReturnType(m_matrix);
+    }
+
+/////////// Cholesky module ///////////
+
+    const LLT<PlainObject, UpLo> llt() const;
+    const LDLT<PlainObject, UpLo> ldlt() const;
+
+/////////// Eigenvalue module ///////////
+
+    /** Real part of #Scalar */
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    /** Return type of eigenvalues() */
+    typedef Matrix<RealScalar, internal::traits<MatrixType>::ColsAtCompileTime, 1> EigenvaluesReturnType;
+
+    EIGEN_DEVICE_FUNC
+    EigenvaluesReturnType eigenvalues() const;
+    EIGEN_DEVICE_FUNC
+    RealScalar operatorNorm() const;
+
+  protected:
+    MatrixTypeNested m_matrix;
+};
+
+
+// template<typename OtherDerived, typename MatrixType, unsigned int UpLo>
+// internal::selfadjoint_matrix_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> >
+// operator*(const MatrixBase<OtherDerived>& lhs, const SelfAdjointView<MatrixType,UpLo>& rhs)
+// {
+//   return internal::matrix_selfadjoint_product_returntype<OtherDerived,SelfAdjointView<MatrixType,UpLo> >(lhs.derived(),rhs);
+// }
+
+// selfadjoint to dense matrix
+
+namespace internal {
+
+// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.>
+//      in the future selfadjoint-ness should be defined by the expression traits
+//      such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<SelfAdjointView<MatrixType,Mode> >
+{
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SelfAdjointShape Shape;
+};
+
+template<int UpLo, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version>
+class triangular_dense_assignment_kernel<UpLo,SelfAdjoint,SetOpposite,DstEvaluatorTypeT,SrcEvaluatorTypeT,Functor,Version>
+  : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
+  {
+    eigen_internal_assert(row!=col);
+    Scalar tmp = m_src.coeff(row,col);
+    m_functor.assignCoeff(m_dst.coeffRef(row,col), tmp);
+    m_functor.assignCoeff(m_dst.coeffRef(col,row), numext::conj(tmp));
+  }
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
+  {
+    Base::assignCoeff(id,id);
+  }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index, Index)
+  { eigen_internal_assert(false && "should never be called"); }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Implementation of MatrixBase methods
+***************************************************************************/
+
+/** This is the const version of MatrixBase::selfadjointView() */
+template<typename Derived>
+template<unsigned int UpLo>
+typename MatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type
+MatrixBase<Derived>::selfadjointView() const
+{
+  return typename ConstSelfAdjointViewReturnType<UpLo>::Type(derived());
+}
+
+/** \returns an expression of a symmetric/self-adjoint view extracted from the upper or lower triangular part of the current matrix
+  *
+  * The parameter \a UpLo can be either \c #Upper or \c #Lower
+  *
+  * Example: \include MatrixBase_selfadjointView.cpp
+  * Output: \verbinclude MatrixBase_selfadjointView.out
+  *
+  * \sa class SelfAdjointView
+  */
+template<typename Derived>
+template<unsigned int UpLo>
+typename MatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type
+MatrixBase<Derived>::selfadjointView()
+{
+  return typename SelfAdjointViewReturnType<UpLo>::Type(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINTMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h
new file mode 100644
index 0000000..50099df
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SelfCwiseBinaryOp.h
@@ -0,0 +1,51 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFCWISEBINARYOP_H
+#define EIGEN_SELFCWISEBINARYOP_H
+
+namespace Eigen { 
+
+// TODO generalize the scalar type of 'other'
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator*=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::mul_assign_op<Scalar,Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator+=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::add_assign_op<Scalar,Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& ArrayBase<Derived>::operator-=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::sub_assign_op<Scalar,Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& DenseBase<Derived>::operator/=(const Scalar& other)
+{
+  typedef typename Derived::PlainObject PlainObject;
+  internal::call_assignment(this->derived(), PlainObject::Constant(rows(),cols(),other), internal::div_assign_op<Scalar,Scalar>());
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFCWISEBINARYOP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Solve.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Solve.h
new file mode 100644
index 0000000..a8daea5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Solve.h
@@ -0,0 +1,188 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVE_H
+#define EIGEN_SOLVE_H
+
+namespace Eigen {
+
+template<typename Decomposition, typename RhsType, typename StorageKind> class SolveImpl;
+  
+/** \class Solve
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression representing a solving operation
+  *
+  * \tparam Decomposition the type of the matrix or decomposion object
+  * \tparam Rhstype the type of the right-hand side
+  *
+  * This class represents an expression of A.solve(B)
+  * and most of the time this is the only way it is used.
+  *
+  */
+namespace internal {
+
+// this solve_traits class permits to determine the evaluation type with respect to storage kind (Dense vs Sparse)
+template<typename Decomposition, typename RhsType,typename StorageKind> struct solve_traits;
+
+template<typename Decomposition, typename RhsType>
+struct solve_traits<Decomposition,RhsType,Dense>
+{
+  typedef typename make_proper_matrix_type<typename RhsType::Scalar,
+                 Decomposition::ColsAtCompileTime,
+                 RhsType::ColsAtCompileTime,
+                 RhsType::PlainObject::Options,
+                 Decomposition::MaxColsAtCompileTime,
+                 RhsType::MaxColsAtCompileTime>::type PlainObject;
+};
+
+template<typename Decomposition, typename RhsType>
+struct traits<Solve<Decomposition, RhsType> >
+  : traits<typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject>
+{
+  typedef typename solve_traits<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>::PlainObject PlainObject;
+  typedef typename promote_index_type<typename Decomposition::StorageIndex, typename RhsType::StorageIndex>::type StorageIndex;
+  typedef traits<PlainObject> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit,
+    CoeffReadCost = HugeCost
+  };
+};
+
+}
+
+
+template<typename Decomposition, typename RhsType>
+class Solve : public SolveImpl<Decomposition,RhsType,typename internal::traits<RhsType>::StorageKind>
+{
+public:
+  typedef typename internal::traits<Solve>::PlainObject PlainObject;
+  typedef typename internal::traits<Solve>::StorageIndex StorageIndex;
+  
+  Solve(const Decomposition &dec, const RhsType &rhs)
+    : m_dec(dec), m_rhs(rhs)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); }
+
+  EIGEN_DEVICE_FUNC const Decomposition& dec() const { return m_dec; }
+  EIGEN_DEVICE_FUNC const RhsType&       rhs() const { return m_rhs; }
+
+protected:
+  const Decomposition &m_dec;
+  const RhsType       &m_rhs;
+};
+
+
+// Specialization of the Solve expression for dense results
+template<typename Decomposition, typename RhsType>
+class SolveImpl<Decomposition,RhsType,Dense>
+  : public MatrixBase<Solve<Decomposition,RhsType> >
+{
+  typedef Solve<Decomposition,RhsType> Derived;
+  
+public:
+  
+  typedef MatrixBase<Solve<Decomposition,RhsType> > Base;
+  EIGEN_DENSE_PUBLIC_INTERFACE(Derived)
+
+private:
+  
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+// Generic API dispatcher
+template<typename Decomposition, typename RhsType, typename StorageKind>
+class SolveImpl : public internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type
+{
+  public:
+    typedef typename internal::generic_xpr_base<Solve<Decomposition,RhsType>, MatrixXpr, StorageKind>::type Base;
+};
+
+namespace internal {
+
+// Evaluator of Solve -> eval into a temporary
+template<typename Decomposition, typename RhsType>
+struct evaluator<Solve<Decomposition,RhsType> >
+  : public evaluator<typename Solve<Decomposition,RhsType>::PlainObject>
+{
+  typedef Solve<Decomposition,RhsType> SolveType;
+  typedef typename SolveType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  enum { Flags = Base::Flags | EvalBeforeNestingBit };
+  
+  EIGEN_DEVICE_FUNC explicit evaluator(const SolveType& solve)
+    : m_result(solve.rows(), solve.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    solve.dec()._solve_impl(solve.rhs(), m_result);
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// Specialization for "dst = dec.solve(rhs)"
+// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<DecType,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    src.dec()._solve_impl(src.rhs(), dst);
+  }
+};
+
+// Specialization for "dst = dec.transpose().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<Transpose<const DecType>,RhsType>, internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<Transpose<const DecType>,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    src.dec().nestedExpression().template _solve_impl_transposed<false>(src.rhs(), dst);
+  }
+};
+
+// Specialization for "dst = dec.adjoint().solve(rhs)"
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType>,
+                  internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef Solve<CwiseUnaryOp<internal::scalar_conjugate_op<typename DecType::Scalar>, const Transpose<const DecType> >,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    src.dec().nestedExpression().nestedExpression().template _solve_impl_transposed<true>(src.rhs(), dst);
+  }
+};
+
+} // end namepsace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h
new file mode 100644
index 0000000..049890b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolveTriangular.h
@@ -0,0 +1,232 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVETRIANGULAR_H
+#define EIGEN_SOLVETRIANGULAR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// Forward declarations:
+// The following two routines are implemented in the products/TriangularSolver*.h files
+template<typename LhsScalar, typename RhsScalar, typename Index, int Side, int Mode, bool Conjugate, int StorageOrder>
+struct triangular_solve_vector;
+
+template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder, int OtherStorageOrder>
+struct triangular_solve_matrix;
+
+// small helper struct extracting some traits on the underlying solver operation
+template<typename Lhs, typename Rhs, int Side>
+class trsolve_traits
+{
+  private:
+    enum {
+      RhsIsVectorAtCompileTime = (Side==OnTheLeft ? Rhs::ColsAtCompileTime : Rhs::RowsAtCompileTime)==1
+    };
+  public:
+    enum {
+      Unrolling   = (RhsIsVectorAtCompileTime && Rhs::SizeAtCompileTime != Dynamic && Rhs::SizeAtCompileTime <= 8)
+                  ? CompleteUnrolling : NoUnrolling,
+      RhsVectors  = RhsIsVectorAtCompileTime ? 1 : Dynamic
+    };
+};
+
+template<typename Lhs, typename Rhs,
+  int Side, // can be OnTheLeft/OnTheRight
+  int Mode, // can be Upper/Lower | UnitDiag
+  int Unrolling = trsolve_traits<Lhs,Rhs,Side>::Unrolling,
+  int RhsVectors = trsolve_traits<Lhs,Rhs,Side>::RhsVectors
+  >
+struct triangular_solver_selector;
+
+template<typename Lhs, typename Rhs, int Side, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,1>
+{
+  typedef typename Lhs::Scalar LhsScalar;
+  typedef typename Rhs::Scalar RhsScalar;
+  typedef blas_traits<Lhs> LhsProductTraits;
+  typedef typename LhsProductTraits::ExtractType ActualLhsType;
+  typedef Map<Matrix<RhsScalar,Dynamic,1>, Aligned> MappedRhs;
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    ActualLhsType actualLhs = LhsProductTraits::extract(lhs);
+
+    // FIXME find a way to allow an inner stride if packet_traits<Scalar>::size==1
+
+    bool useRhsDirectly = Rhs::InnerStrideAtCompileTime==1 || rhs.innerStride()==1;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhs,rhs.size(),
+                                                  (useRhsDirectly ? rhs.data() : 0));
+                                                  
+    if(!useRhsDirectly)
+      MappedRhs(actualRhs,rhs.size()) = rhs;
+
+    triangular_solve_vector<LhsScalar, RhsScalar, Index, Side, Mode, LhsProductTraits::NeedToConjugate,
+                            (int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor>
+      ::run(actualLhs.cols(), actualLhs.data(), actualLhs.outerStride(), actualRhs);
+
+    if(!useRhsDirectly)
+      rhs = MappedRhs(actualRhs, rhs.size());
+  }
+};
+
+// the rhs is a matrix
+template<typename Lhs, typename Rhs, int Side, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,Side,Mode,NoUnrolling,Dynamic>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef blas_traits<Lhs> LhsProductTraits;
+  typedef typename LhsProductTraits::DirectLinearAccessType ActualLhsType;
+
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsProductTraits::extract(lhs);
+
+    const Index size = lhs.rows();
+    const Index othersize = Side==OnTheLeft? rhs.cols() : rhs.rows();
+
+    typedef internal::gemm_blocking_space<(Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Rhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxRowsAtCompileTime,4> BlockingType;
+
+    BlockingType blocking(rhs.rows(), rhs.cols(), size, 1, false);
+
+    triangular_solve_matrix<Scalar,Index,Side,Mode,LhsProductTraits::NeedToConjugate,(int(Lhs::Flags) & RowMajorBit) ? RowMajor : ColMajor,
+                               (Rhs::Flags&RowMajorBit) ? RowMajor : ColMajor>
+      ::run(size, othersize, &actualLhs.coeffRef(0,0), actualLhs.outerStride(), &rhs.coeffRef(0,0), rhs.outerStride(), blocking);
+  }
+};
+
+/***************************************************************************
+* meta-unrolling implementation
+***************************************************************************/
+
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size,
+         bool Stop = LoopIndex==Size>
+struct triangular_solver_unroller;
+
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,false> {
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    DiagIndex  = IsLower ? LoopIndex : Size - LoopIndex - 1,
+    StartIndex = IsLower ? 0         : DiagIndex+1
+  };
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    if (LoopIndex>0)
+      rhs.coeffRef(DiagIndex) -= lhs.row(DiagIndex).template segment<LoopIndex>(StartIndex).transpose()
+                                .cwiseProduct(rhs.template segment<LoopIndex>(StartIndex)).sum();
+
+    if(!(Mode & UnitDiag))
+      rhs.coeffRef(DiagIndex) /= lhs.coeff(DiagIndex,DiagIndex);
+
+    triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex+1,Size>::run(lhs,rhs);
+  }
+};
+
+template<typename Lhs, typename Rhs, int Mode, int LoopIndex, int Size>
+struct triangular_solver_unroller<Lhs,Rhs,Mode,LoopIndex,Size,true> {
+  static void run(const Lhs&, Rhs&) {}
+};
+
+template<typename Lhs, typename Rhs, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,OnTheLeft,Mode,CompleteUnrolling,1> {
+  static void run(const Lhs& lhs, Rhs& rhs)
+  { triangular_solver_unroller<Lhs,Rhs,Mode,0,Rhs::SizeAtCompileTime>::run(lhs,rhs); }
+};
+
+template<typename Lhs, typename Rhs, int Mode>
+struct triangular_solver_selector<Lhs,Rhs,OnTheRight,Mode,CompleteUnrolling,1> {
+  static void run(const Lhs& lhs, Rhs& rhs)
+  {
+    Transpose<const Lhs> trLhs(lhs);
+    Transpose<Rhs> trRhs(rhs);
+    
+    triangular_solver_unroller<Transpose<const Lhs>,Transpose<Rhs>,
+                              ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag),
+                              0,Rhs::SizeAtCompileTime>::run(trLhs,trRhs);
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* TriangularView methods
+***************************************************************************/
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename MatrixType, unsigned int Mode>
+template<int Side, typename OtherDerived>
+void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<OtherDerived>& _other) const
+{
+  OtherDerived& other = _other.const_cast_derived();
+  eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
+  eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
+
+  enum { copy = (internal::traits<OtherDerived>::Flags & RowMajorBit)  && OtherDerived::IsVectorAtCompileTime && OtherDerived::SizeAtCompileTime!=1};
+  typedef typename internal::conditional<copy,
+    typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy;
+  OtherCopy otherCopy(other);
+
+  internal::triangular_solver_selector<MatrixType, typename internal::remove_reference<OtherCopy>::type,
+    Side, Mode>::run(derived().nestedExpression(), otherCopy);
+
+  if (copy)
+    other = otherCopy;
+}
+
+template<typename Derived, unsigned int Mode>
+template<int Side, typename Other>
+const internal::triangular_solve_retval<Side,TriangularView<Derived,Mode>,Other>
+TriangularViewImpl<Derived,Mode,Dense>::solve(const MatrixBase<Other>& other) const
+{
+  return internal::triangular_solve_retval<Side,TriangularViewType,Other>(derived(), other.derived());
+}
+#endif
+
+namespace internal {
+
+
+template<int Side, typename TriangularType, typename Rhs>
+struct traits<triangular_solve_retval<Side, TriangularType, Rhs> >
+{
+  typedef typename internal::plain_matrix_type_column_major<Rhs>::type ReturnType;
+};
+
+template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval
+ : public ReturnByValue<triangular_solve_retval<Side, TriangularType, Rhs> >
+{
+  typedef typename remove_all<typename Rhs::Nested>::type RhsNestedCleaned;
+  typedef ReturnByValue<triangular_solve_retval> Base;
+
+  triangular_solve_retval(const TriangularType& tri, const Rhs& rhs)
+    : m_triangularMatrix(tri), m_rhs(rhs)
+  {}
+
+  inline Index rows() const { return m_rhs.rows(); }
+  inline Index cols() const { return m_rhs.cols(); }
+
+  template<typename Dest> inline void evalTo(Dest& dst) const
+  {
+    if(!is_same_dense(dst,m_rhs))
+      dst = m_rhs;
+    m_triangularMatrix.template solveInPlace<Side>(dst);
+  }
+
+  protected:
+    const TriangularType& m_triangularMatrix;
+    typename Rhs::Nested m_rhs;
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVETRIANGULAR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h
new file mode 100644
index 0000000..8a4adc2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/SolverBase.h
@@ -0,0 +1,130 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVERBASE_H
+#define EIGEN_SOLVERBASE_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+
+} // end namespace internal
+
+/** \class SolverBase
+  * \brief A base class for matrix decomposition and solvers
+  *
+  * \tparam Derived the actual type of the decomposition/solver.
+  *
+  * Any matrix decomposition inheriting this base class provide the following API:
+  *
+  * \code
+  * MatrixType A, b, x;
+  * DecompositionType dec(A);
+  * x = dec.solve(b);             // solve A   * x = b
+  * x = dec.transpose().solve(b); // solve A^T * x = b
+  * x = dec.adjoint().solve(b);   // solve A'  * x = b
+  * \endcode
+  *
+  * \warning Currently, any other usage of transpose() and adjoint() are not supported and will produce compilation errors.
+  *
+  * \sa class PartialPivLU, class FullPivLU
+  */
+template<typename Derived>
+class SolverBase : public EigenBase<Derived>
+{
+  public:
+
+    typedef EigenBase<Derived> Base;
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef Scalar CoeffReturnType;
+
+    enum {
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                          internal::traits<Derived>::ColsAtCompileTime>::ret),
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                             internal::traits<Derived>::MaxColsAtCompileTime>::ret),
+      IsVectorAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime == 1
+                           || internal::traits<Derived>::MaxColsAtCompileTime == 1
+    };
+
+    /** Default constructor */
+    SolverBase()
+    {}
+
+    ~SolverBase()
+    {}
+
+    using Base::derived;
+
+    /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      */
+    template<typename Rhs>
+    inline const Solve<Derived, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<Derived, Rhs>(derived(), b.derived());
+    }
+
+    /** \internal the return type of transpose() */
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+    /** \returns an expression of the transposed of the factored matrix.
+      *
+      * A typical usage is to solve for the transposed problem A^T x = b:
+      * \code x = dec.transpose().solve(b); \endcode
+      *
+      * \sa adjoint(), solve()
+      */
+    inline ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(derived());
+    }
+
+    /** \internal the return type of adjoint() */
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                        CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, ConstTransposeReturnType>,
+                        ConstTransposeReturnType
+                     >::type AdjointReturnType;
+    /** \returns an expression of the adjoint of the factored matrix
+      *
+      * A typical usage is to solve for the adjoint problem A' x = b:
+      * \code x = dec.adjoint().solve(b); \endcode
+      *
+      * For real scalar types, this function is equivalent to transpose().
+      *
+      * \sa transpose(), solve()
+      */
+    inline AdjointReturnType adjoint() const
+    {
+      return AdjointReturnType(derived().transpose());
+    }
+
+  protected:
+};
+
+namespace internal {
+
+template<typename Derived>
+struct generic_xpr_base<Derived, MatrixXpr, SolverStorage>
+{
+  typedef SolverBase<Derived> type;
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVERBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h
new file mode 100644
index 0000000..be04ed4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/StableNorm.h
@@ -0,0 +1,221 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STABLENORM_H
+#define EIGEN_STABLENORM_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename ExpressionType, typename Scalar>
+inline void stable_norm_kernel(const ExpressionType& bl, Scalar& ssq, Scalar& scale, Scalar& invScale)
+{
+  Scalar maxCoeff = bl.cwiseAbs().maxCoeff();
+  
+  if(maxCoeff>scale)
+  {
+    ssq = ssq * numext::abs2(scale/maxCoeff);
+    Scalar tmp = Scalar(1)/maxCoeff;
+    if(tmp > NumTraits<Scalar>::highest())
+    {
+      invScale = NumTraits<Scalar>::highest();
+      scale = Scalar(1)/invScale;
+    }
+    else if(maxCoeff>NumTraits<Scalar>::highest()) // we got a INF
+    {
+      invScale = Scalar(1);
+      scale = maxCoeff;
+    }
+    else
+    {
+      scale = maxCoeff;
+      invScale = tmp;
+    }
+  }
+  else if(maxCoeff!=maxCoeff) // we got a NaN
+  {
+    scale = maxCoeff;
+  }
+  
+  // TODO if the maxCoeff is much much smaller than the current scale,
+  // then we can neglect this sub vector
+  if(scale>Scalar(0)) // if scale==0, then bl is 0 
+    ssq += (bl*invScale).squaredNorm();
+}
+
+template<typename Derived>
+inline typename NumTraits<typename traits<Derived>::Scalar>::Real
+blueNorm_impl(const EigenBase<Derived>& _vec)
+{
+  typedef typename Derived::RealScalar RealScalar;  
+  using std::pow;
+  using std::sqrt;
+  using std::abs;
+  const Derived& vec(_vec.derived());
+  static bool initialized = false;
+  static RealScalar b1, b2, s1m, s2m, rbig, relerr;
+  if(!initialized)
+  {
+    int ibeta, it, iemin, iemax, iexp;
+    RealScalar eps;
+    // This program calculates the machine-dependent constants
+    // bl, b2, slm, s2m, relerr overfl
+    // from the "basic" machine-dependent numbers
+    // nbig, ibeta, it, iemin, iemax, rbig.
+    // The following define the basic machine-dependent constants.
+    // For portability, the PORT subprograms "ilmaeh" and "rlmach"
+    // are used. For any specific computer, each of the assignment
+    // statements can be replaced
+    ibeta = std::numeric_limits<RealScalar>::radix;                 // base for floating-point numbers
+    it    = std::numeric_limits<RealScalar>::digits;                // number of base-beta digits in mantissa
+    iemin = std::numeric_limits<RealScalar>::min_exponent;          // minimum exponent
+    iemax = std::numeric_limits<RealScalar>::max_exponent;          // maximum exponent
+    rbig  = (std::numeric_limits<RealScalar>::max)();               // largest floating-point number
+
+    iexp  = -((1-iemin)/2);
+    b1    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // lower boundary of midrange
+    iexp  = (iemax + 1 - it)/2;
+    b2    = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // upper boundary of midrange
+
+    iexp  = (2-iemin)/2;
+    s1m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for lower range
+    iexp  = - ((iemax+it)/2);
+    s2m   = RealScalar(pow(RealScalar(ibeta),RealScalar(iexp)));    // scaling factor for upper range
+
+    eps     = RealScalar(pow(double(ibeta), 1-it));
+    relerr  = sqrt(eps);                                            // tolerance for neglecting asml
+    initialized = true;
+  }
+  Index n = vec.size();
+  RealScalar ab2 = b2 / RealScalar(n);
+  RealScalar asml = RealScalar(0);
+  RealScalar amed = RealScalar(0);
+  RealScalar abig = RealScalar(0);
+  for(typename Derived::InnerIterator it(vec, 0); it; ++it)
+  {
+    RealScalar ax = abs(it.value());
+    if(ax > ab2)     abig += numext::abs2(ax*s2m);
+    else if(ax < b1) asml += numext::abs2(ax*s1m);
+    else             amed += numext::abs2(ax);
+  }
+  if(amed!=amed)
+    return amed;  // we got a NaN
+  if(abig > RealScalar(0))
+  {
+    abig = sqrt(abig);
+    if(abig > rbig) // overflow, or *this contains INF values
+      return abig;  // return INF
+    if(amed > RealScalar(0))
+    {
+      abig = abig/s2m;
+      amed = sqrt(amed);
+    }
+    else
+      return abig/s2m;
+  }
+  else if(asml > RealScalar(0))
+  {
+    if (amed > RealScalar(0))
+    {
+      abig = sqrt(amed);
+      amed = sqrt(asml) / s1m;
+    }
+    else
+      return sqrt(asml)/s1m;
+  }
+  else
+    return sqrt(amed);
+  asml = numext::mini(abig, amed);
+  abig = numext::maxi(abig, amed);
+  if(asml <= abig*relerr)
+    return abig;
+  else
+    return abig * sqrt(RealScalar(1) + numext::abs2(asml/abig));
+}
+
+} // end namespace internal
+
+/** \returns the \em l2 norm of \c *this avoiding underflow and overflow.
+  * This version use a blockwise two passes algorithm:
+  *  1 - find the absolute largest coefficient \c s
+  *  2 - compute \f$ s \Vert \frac{*this}{s} \Vert \f$ in a standard way
+  *
+  * For architecture/scalar types supporting vectorization, this version
+  * is faster than blueNorm(). Otherwise the blueNorm() is much faster.
+  *
+  * \sa norm(), blueNorm(), hypotNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::stableNorm() const
+{
+  using std::sqrt;
+  using std::abs;
+  const Index blockSize = 4096;
+  RealScalar scale(0);
+  RealScalar invScale(1);
+  RealScalar ssq(0); // sum of square
+  
+  typedef typename internal::nested_eval<Derived,2>::type DerivedCopy;
+  typedef typename internal::remove_all<DerivedCopy>::type DerivedCopyClean;
+  DerivedCopy copy(derived());
+  
+  enum {
+    CanAlign = (   (int(DerivedCopyClean::Flags)&DirectAccessBit)
+                || (int(internal::evaluator<DerivedCopyClean>::Alignment)>0) // FIXME Alignment)>0 might not be enough
+               ) && (blockSize*sizeof(Scalar)*2<EIGEN_STACK_ALLOCATION_LIMIT)
+                 && (EIGEN_MAX_STATIC_ALIGN_BYTES>0) // if we cannot allocate on the stack, then let's not bother about this optimization
+  };
+  typedef typename internal::conditional<CanAlign, Ref<const Matrix<Scalar,Dynamic,1,0,blockSize,1>, internal::evaluator<DerivedCopyClean>::Alignment>,
+                                                   typename DerivedCopyClean::ConstSegmentReturnType>::type SegmentWrapper;
+  Index n = size();
+  
+  if(n==1)
+    return abs(this->coeff(0));
+  
+  Index bi = internal::first_default_aligned(copy);
+  if (bi>0)
+    internal::stable_norm_kernel(copy.head(bi), ssq, scale, invScale);
+  for (; bi<n; bi+=blockSize)
+    internal::stable_norm_kernel(SegmentWrapper(copy.segment(bi,numext::mini(blockSize, n - bi))), ssq, scale, invScale);
+  return scale * sqrt(ssq);
+}
+
+/** \returns the \em l2 norm of \c *this using the Blue's algorithm.
+  * A Portable Fortran Program to Find the Euclidean Norm of a Vector,
+  * ACM TOMS, Vol 4, Issue 1, 1978.
+  *
+  * For architecture/scalar types without vectorization, this version
+  * is much faster than stableNorm(). Otherwise the stableNorm() is faster.
+  *
+  * \sa norm(), stableNorm(), hypotNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::blueNorm() const
+{
+  return internal::blueNorm_impl(*this);
+}
+
+/** \returns the \em l2 norm of \c *this avoiding undeflow and overflow.
+  * This version use a concatenation of hypot() calls, and it is very slow.
+  *
+  * \sa norm(), stableNorm()
+  */
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+MatrixBase<Derived>::hypotNorm() const
+{
+  return this->cwiseAbs().redux(internal::scalar_hypot_op<RealScalar>());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_STABLENORM_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Stride.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Stride.h
new file mode 100644
index 0000000..513742f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Stride.h
@@ -0,0 +1,111 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STRIDE_H
+#define EIGEN_STRIDE_H
+
+namespace Eigen { 
+
+/** \class Stride
+  * \ingroup Core_Module
+  *
+  * \brief Holds strides information for Map
+  *
+  * This class holds the strides information for mapping arrays with strides with class Map.
+  *
+  * It holds two values: the inner stride and the outer stride.
+  *
+  * The inner stride is the pointer increment between two consecutive entries within a given row of a
+  * row-major matrix or within a given column of a column-major matrix.
+  *
+  * The outer stride is the pointer increment between two consecutive rows of a row-major matrix or
+  * between two consecutive columns of a column-major matrix.
+  *
+  * These two values can be passed either at compile-time as template parameters, or at runtime as
+  * arguments to the constructor.
+  *
+  * Indeed, this class takes two template parameters:
+  *  \tparam _OuterStrideAtCompileTime the outer stride, or Dynamic if you want to specify it at runtime.
+  *  \tparam _InnerStrideAtCompileTime the inner stride, or Dynamic if you want to specify it at runtime.
+  *
+  * Here is an example:
+  * \include Map_general_stride.cpp
+  * Output: \verbinclude Map_general_stride.out
+  *
+  * \sa class InnerStride, class OuterStride, \ref TopicStorageOrders
+  */
+template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
+class Stride
+{
+  public:
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    enum {
+      InnerStrideAtCompileTime = _InnerStrideAtCompileTime,
+      OuterStrideAtCompileTime = _OuterStrideAtCompileTime
+    };
+
+    /** Default constructor, for use when strides are fixed at compile time */
+    EIGEN_DEVICE_FUNC
+    Stride()
+      : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime)
+    {
+      eigen_assert(InnerStrideAtCompileTime != Dynamic && OuterStrideAtCompileTime != Dynamic);
+    }
+
+    /** Constructor allowing to pass the strides at runtime */
+    EIGEN_DEVICE_FUNC
+    Stride(Index outerStride, Index innerStride)
+      : m_outer(outerStride), m_inner(innerStride)
+    {
+      eigen_assert(innerStride>=0 && outerStride>=0);
+    }
+
+    /** Copy constructor */
+    EIGEN_DEVICE_FUNC
+    Stride(const Stride& other)
+      : m_outer(other.outer()), m_inner(other.inner())
+    {}
+
+    /** \returns the outer stride */
+    EIGEN_DEVICE_FUNC
+    inline Index outer() const { return m_outer.value(); }
+    /** \returns the inner stride */
+    EIGEN_DEVICE_FUNC
+    inline Index inner() const { return m_inner.value(); }
+
+  protected:
+    internal::variable_if_dynamic<Index, OuterStrideAtCompileTime> m_outer;
+    internal::variable_if_dynamic<Index, InnerStrideAtCompileTime> m_inner;
+};
+
+/** \brief Convenience specialization of Stride to specify only an inner stride
+  * See class Map for some examples */
+template<int Value>
+class InnerStride : public Stride<0, Value>
+{
+    typedef Stride<0, Value> Base;
+  public:
+    EIGEN_DEVICE_FUNC InnerStride() : Base() {}
+    EIGEN_DEVICE_FUNC InnerStride(Index v) : Base(0, v) {} // FIXME making this explicit could break valid code
+};
+
+/** \brief Convenience specialization of Stride to specify only an outer stride
+  * See class Map for some examples */
+template<int Value>
+class OuterStride : public Stride<Value, 0>
+{
+    typedef Stride<Value, 0> Base;
+  public:
+    EIGEN_DEVICE_FUNC OuterStride() : Base() {}
+    EIGEN_DEVICE_FUNC OuterStride(Index v) : Base(v,0) {} // FIXME making this explicit could break valid code
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_STRIDE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Swap.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Swap.h
new file mode 100644
index 0000000..d702009
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Swap.h
@@ -0,0 +1,67 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SWAP_H
+#define EIGEN_SWAP_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// Overload default assignPacket behavior for swapping them
+template<typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT>
+class generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, Specialized>
+ : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, swap_assign_op<typename DstEvaluatorTypeT::Scalar>, BuiltIn> Base;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+  
+public:
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::DstXprType DstXprType;
+  typedef swap_assign_op<Scalar> Functor;
+  
+  EIGEN_DEVICE_FUNC generic_dense_assignment_kernel(DstEvaluatorTypeT &dst, const SrcEvaluatorTypeT &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index row, Index col)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(row,col);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(row,col, m_dst.template packet<StoreMode,PacketType>(row,col));
+    m_dst.template writePacket<StoreMode>(row,col,tmp);
+  }
+  
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacket(Index index)
+  {
+    PacketType tmp = m_src.template packet<LoadMode,PacketType>(index);
+    const_cast<SrcEvaluatorTypeT&>(m_src).template writePacket<LoadMode>(index, m_dst.template packet<StoreMode,PacketType>(index));
+    m_dst.template writePacket<StoreMode>(index,tmp);
+  }
+  
+  // TODO find a simple way not to have to copy/paste this function from generic_dense_assignment_kernel, by simple I mean no CRTP (Gael)
+  template<int StoreMode, int LoadMode, typename PacketType>
+  void assignPacketByOuterInner(Index outer, Index inner)
+  {
+    Index row = Base::rowIndexByOuterInner(outer, inner); 
+    Index col = Base::colIndexByOuterInner(outer, inner);
+    assignPacket<StoreMode,LoadMode,PacketType>(row, col);
+  }
+};
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SWAP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpose.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpose.h
new file mode 100644
index 0000000..79b767b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpose.h
@@ -0,0 +1,403 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRANSPOSE_H
+#define EIGEN_TRANSPOSE_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename MatrixType>
+struct traits<Transpose<MatrixType> > : public traits<MatrixType>
+{
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedPlain;
+  enum {
+    RowsAtCompileTime = MatrixType::ColsAtCompileTime,
+    ColsAtCompileTime = MatrixType::RowsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags0 = traits<MatrixTypeNestedPlain>::Flags & ~(LvalueBit | NestByRefBit),
+    Flags1 = Flags0 | FlagsLvalueBit,
+    Flags = Flags1 ^ RowMajorBit,
+    InnerStrideAtCompileTime = inner_stride_at_compile_time<MatrixType>::ret,
+    OuterStrideAtCompileTime = outer_stride_at_compile_time<MatrixType>::ret
+  };
+};
+}
+
+template<typename MatrixType, typename StorageKind> class TransposeImpl;
+
+/** \class Transpose
+  * \ingroup Core_Module
+  *
+  * \brief Expression of the transpose of a matrix
+  *
+  * \tparam MatrixType the type of the object of which we are taking the transpose
+  *
+  * This class represents an expression of the transpose of a matrix.
+  * It is the return type of MatrixBase::transpose() and MatrixBase::adjoint()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::transpose(), MatrixBase::adjoint()
+  */
+template<typename MatrixType> class Transpose
+  : public TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>
+{
+  public:
+
+    typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+
+    typedef typename TransposeImpl<MatrixType,typename internal::traits<MatrixType>::StorageKind>::Base Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(Transpose)
+    typedef typename internal::remove_all<MatrixType>::type NestedExpression;
+
+    EIGEN_DEVICE_FUNC
+    explicit inline Transpose(MatrixType& matrix) : m_matrix(matrix) {}
+
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Transpose)
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.cols(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.rows(); }
+
+    /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
+    const typename internal::remove_all<MatrixTypeNested>::type&
+    nestedExpression() const { return m_matrix; }
+
+    /** \returns the nested expression */
+    EIGEN_DEVICE_FUNC
+    typename internal::remove_reference<MatrixTypeNested>::type&
+    nestedExpression() { return m_matrix; }
+
+    /** \internal */
+    void resize(Index nrows, Index ncols) {
+      m_matrix.resize(ncols,nrows);
+    }
+
+  protected:
+    typename internal::ref_selector<MatrixType>::non_const_type m_matrix;
+};
+
+namespace internal {
+
+template<typename MatrixType, bool HasDirectAccess = has_direct_access<MatrixType>::ret>
+struct TransposeImpl_base
+{
+  typedef typename dense_xpr_base<Transpose<MatrixType> >::type type;
+};
+
+template<typename MatrixType>
+struct TransposeImpl_base<MatrixType, false>
+{
+  typedef typename dense_xpr_base<Transpose<MatrixType> >::type type;
+};
+
+} // end namespace internal
+
+// Generic API dispatcher
+template<typename XprType, typename StorageKind>
+class TransposeImpl
+  : public internal::generic_xpr_base<Transpose<XprType> >::type
+{
+public:
+  typedef typename internal::generic_xpr_base<Transpose<XprType> >::type Base;
+};
+
+template<typename MatrixType> class TransposeImpl<MatrixType,Dense>
+  : public internal::TransposeImpl_base<MatrixType>::type
+{
+  public:
+
+    typedef typename internal::TransposeImpl_base<MatrixType>::type Base;
+    using Base::coeffRef;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Transpose<MatrixType>)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(TransposeImpl)
+
+    EIGEN_DEVICE_FUNC inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
+    EIGEN_DEVICE_FUNC inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+
+    typedef typename internal::conditional<
+                       internal::is_lvalue<MatrixType>::value,
+                       Scalar,
+                       const Scalar
+                     >::type ScalarWithConstIfNotLvalue;
+
+    EIGEN_DEVICE_FUNC inline ScalarWithConstIfNotLvalue* data() { return derived().nestedExpression().data(); }
+    EIGEN_DEVICE_FUNC inline const Scalar* data() const { return derived().nestedExpression().data(); }
+
+    // FIXME: shall we keep the const version of coeffRef?
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index rowId, Index colId) const
+    {
+      return derived().nestedExpression().coeffRef(colId, rowId);
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline const Scalar& coeffRef(Index index) const
+    {
+      return derived().nestedExpression().coeffRef(index);
+    }
+};
+
+/** \returns an expression of the transpose of *this.
+  *
+  * Example: \include MatrixBase_transpose.cpp
+  * Output: \verbinclude MatrixBase_transpose.out
+  *
+  * \warning If you want to replace a matrix by its own transpose, do \b NOT do this:
+  * \code
+  * m = m.transpose(); // bug!!! caused by aliasing effect
+  * \endcode
+  * Instead, use the transposeInPlace() method:
+  * \code
+  * m.transposeInPlace();
+  * \endcode
+  * which gives Eigen good opportunities for optimization, or alternatively you can also do:
+  * \code
+  * m = m.transpose().eval();
+  * \endcode
+  *
+  * \sa transposeInPlace(), adjoint() */
+template<typename Derived>
+inline Transpose<Derived>
+DenseBase<Derived>::transpose()
+{
+  return TransposeReturnType(derived());
+}
+
+/** This is the const version of transpose().
+  *
+  * Make sure you read the warning for transpose() !
+  *
+  * \sa transposeInPlace(), adjoint() */
+template<typename Derived>
+inline typename DenseBase<Derived>::ConstTransposeReturnType
+DenseBase<Derived>::transpose() const
+{
+  return ConstTransposeReturnType(derived());
+}
+
+/** \returns an expression of the adjoint (i.e. conjugate transpose) of *this.
+  *
+  * Example: \include MatrixBase_adjoint.cpp
+  * Output: \verbinclude MatrixBase_adjoint.out
+  *
+  * \warning If you want to replace a matrix by its own adjoint, do \b NOT do this:
+  * \code
+  * m = m.adjoint(); // bug!!! caused by aliasing effect
+  * \endcode
+  * Instead, use the adjointInPlace() method:
+  * \code
+  * m.adjointInPlace();
+  * \endcode
+  * which gives Eigen good opportunities for optimization, or alternatively you can also do:
+  * \code
+  * m = m.adjoint().eval();
+  * \endcode
+  *
+  * \sa adjointInPlace(), transpose(), conjugate(), class Transpose, class internal::scalar_conjugate_op */
+template<typename Derived>
+inline const typename MatrixBase<Derived>::AdjointReturnType
+MatrixBase<Derived>::adjoint() const
+{
+  return AdjointReturnType(this->transpose());
+}
+
+/***************************************************************************
+* "in place" transpose implementation
+***************************************************************************/
+
+namespace internal {
+
+template<typename MatrixType,
+  bool IsSquare = (MatrixType::RowsAtCompileTime == MatrixType::ColsAtCompileTime) && MatrixType::RowsAtCompileTime!=Dynamic,
+  bool MatchPacketSize =
+        (int(MatrixType::RowsAtCompileTime) == int(internal::packet_traits<typename MatrixType::Scalar>::size))
+    &&  (internal::evaluator<MatrixType>::Flags&PacketAccessBit) >
+struct inplace_transpose_selector;
+
+template<typename MatrixType>
+struct inplace_transpose_selector<MatrixType,true,false> { // square matrix
+  static void run(MatrixType& m) {
+    m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
+  }
+};
+
+// TODO: vectorized path is currently limited to LargestPacketSize x LargestPacketSize cases only.
+template<typename MatrixType>
+struct inplace_transpose_selector<MatrixType,true,true> { // PacketSize x PacketSize
+  static void run(MatrixType& m) {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename internal::packet_traits<typename MatrixType::Scalar>::type Packet;
+    const Index PacketSize = internal::packet_traits<Scalar>::size;
+    const Index Alignment = internal::evaluator<MatrixType>::Alignment;
+    PacketBlock<Packet> A;
+    for (Index i=0; i<PacketSize; ++i)
+      A.packet[i] = m.template packetByOuterInner<Alignment>(i,0);
+    internal::ptranspose(A);
+    for (Index i=0; i<PacketSize; ++i)
+      m.template writePacket<Alignment>(m.rowIndexByOuterInner(i,0), m.colIndexByOuterInner(i,0), A.packet[i]);
+  }
+};
+
+template<typename MatrixType,bool MatchPacketSize>
+struct inplace_transpose_selector<MatrixType,false,MatchPacketSize> { // non square matrix
+  static void run(MatrixType& m) {
+    if (m.rows()==m.cols())
+      m.matrix().template triangularView<StrictlyUpper>().swap(m.matrix().transpose());
+    else
+      m = m.transpose().eval();
+  }
+};
+
+} // end namespace internal
+
+/** This is the "in place" version of transpose(): it replaces \c *this by its own transpose.
+  * Thus, doing
+  * \code
+  * m.transposeInPlace();
+  * \endcode
+  * has the same effect on m as doing
+  * \code
+  * m = m.transpose().eval();
+  * \endcode
+  * and is faster and also safer because in the latter line of code, forgetting the eval() results
+  * in a bug caused by \ref TopicAliasing "aliasing".
+  *
+  * Notice however that this method is only useful if you want to replace a matrix by its own transpose.
+  * If you just need the transpose of a matrix, use transpose().
+  *
+  * \note if the matrix is not square, then \c *this must be a resizable matrix. 
+  * This excludes (non-square) fixed-size matrices, block-expressions and maps.
+  *
+  * \sa transpose(), adjoint(), adjointInPlace() */
+template<typename Derived>
+inline void DenseBase<Derived>::transposeInPlace()
+{
+  eigen_assert((rows() == cols() || (RowsAtCompileTime == Dynamic && ColsAtCompileTime == Dynamic))
+               && "transposeInPlace() called on a non-square non-resizable matrix");
+  internal::inplace_transpose_selector<Derived>::run(derived());
+}
+
+/***************************************************************************
+* "in place" adjoint implementation
+***************************************************************************/
+
+/** This is the "in place" version of adjoint(): it replaces \c *this by its own transpose.
+  * Thus, doing
+  * \code
+  * m.adjointInPlace();
+  * \endcode
+  * has the same effect on m as doing
+  * \code
+  * m = m.adjoint().eval();
+  * \endcode
+  * and is faster and also safer because in the latter line of code, forgetting the eval() results
+  * in a bug caused by aliasing.
+  *
+  * Notice however that this method is only useful if you want to replace a matrix by its own adjoint.
+  * If you just need the adjoint of a matrix, use adjoint().
+  *
+  * \note if the matrix is not square, then \c *this must be a resizable matrix.
+  * This excludes (non-square) fixed-size matrices, block-expressions and maps.
+  *
+  * \sa transpose(), adjoint(), transposeInPlace() */
+template<typename Derived>
+inline void MatrixBase<Derived>::adjointInPlace()
+{
+  derived() = adjoint().eval();
+}
+
+#ifndef EIGEN_NO_DEBUG
+
+// The following is to detect aliasing problems in most common cases.
+
+namespace internal {
+
+template<bool DestIsTransposed, typename OtherDerived>
+struct check_transpose_aliasing_compile_time_selector
+{
+  enum { ret = bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed };
+};
+
+template<bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB>
+struct check_transpose_aliasing_compile_time_selector<DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> >
+{
+  enum { ret =    bool(blas_traits<DerivedA>::IsTransposed) != DestIsTransposed
+               || bool(blas_traits<DerivedB>::IsTransposed) != DestIsTransposed
+  };
+};
+
+template<typename Scalar, bool DestIsTransposed, typename OtherDerived>
+struct check_transpose_aliasing_run_time_selector
+{
+  static bool run(const Scalar* dest, const OtherDerived& src)
+  {
+    return (bool(blas_traits<OtherDerived>::IsTransposed) != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src));
+  }
+};
+
+template<typename Scalar, bool DestIsTransposed, typename BinOp, typename DerivedA, typename DerivedB>
+struct check_transpose_aliasing_run_time_selector<Scalar,DestIsTransposed,CwiseBinaryOp<BinOp,DerivedA,DerivedB> >
+{
+  static bool run(const Scalar* dest, const CwiseBinaryOp<BinOp,DerivedA,DerivedB>& src)
+  {
+    return ((blas_traits<DerivedA>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.lhs())))
+        || ((blas_traits<DerivedB>::IsTransposed != DestIsTransposed) && (dest!=0 && dest==(const Scalar*)extract_data(src.rhs())));
+  }
+};
+
+// the following selector, checkTransposeAliasing_impl, based on MightHaveTransposeAliasing,
+// is because when the condition controlling the assert is known at compile time, ICC emits a warning.
+// This is actually a good warning: in expressions that don't have any transposing, the condition is
+// known at compile time to be false, and using that, we can avoid generating the code of the assert again
+// and again for all these expressions that don't need it.
+
+template<typename Derived, typename OtherDerived,
+         bool MightHaveTransposeAliasing
+                 = check_transpose_aliasing_compile_time_selector
+                     <blas_traits<Derived>::IsTransposed,OtherDerived>::ret
+        >
+struct checkTransposeAliasing_impl
+{
+    static void run(const Derived& dst, const OtherDerived& other)
+    {
+        eigen_assert((!check_transpose_aliasing_run_time_selector
+                      <typename Derived::Scalar,blas_traits<Derived>::IsTransposed,OtherDerived>
+                      ::run(extract_data(dst), other))
+          && "aliasing detected during transposition, use transposeInPlace() "
+             "or evaluate the rhs into a temporary using .eval()");
+
+    }
+};
+
+template<typename Derived, typename OtherDerived>
+struct checkTransposeAliasing_impl<Derived, OtherDerived, false>
+{
+    static void run(const Derived&, const OtherDerived&)
+    {
+    }
+};
+
+template<typename Dst, typename Src>
+void check_for_aliasing(const Dst &dst, const Src &src)
+{
+  internal::checkTransposeAliasing_impl<Dst, Src>::run(dst, src);
+}
+
+} // end namespace internal
+
+#endif // EIGEN_NO_DEBUG
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSPOSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h
new file mode 100644
index 0000000..19c17bb
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Transpositions.h
@@ -0,0 +1,407 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRANSPOSITIONS_H
+#define EIGEN_TRANSPOSITIONS_H
+
+namespace Eigen { 
+
+template<typename Derived>
+class TranspositionsBase
+{
+    typedef internal::traits<Derived> Traits;
+    
+  public:
+
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar StorageIndex;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    Derived& derived() { return *static_cast<Derived*>(this); }
+    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Derived& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Derived& operator=(const TranspositionsBase& other)
+    {
+      indices() = other.indices();
+      return derived();
+    }
+    #endif
+
+    /** \returns the number of transpositions */
+    Index size() const { return indices().size(); }
+    /** \returns the number of rows of the equivalent permutation matrix */
+    Index rows() const { return indices().size(); }
+    /** \returns the number of columns of the equivalent permutation matrix */
+    Index cols() const { return indices().size(); }
+
+    /** Direct access to the underlying index vector */
+    inline const StorageIndex& coeff(Index i) const { return indices().coeff(i); }
+    /** Direct access to the underlying index vector */
+    inline StorageIndex& coeffRef(Index i) { return indices().coeffRef(i); }
+    /** Direct access to the underlying index vector */
+    inline const StorageIndex& operator()(Index i) const { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline StorageIndex& operator()(Index i) { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline const StorageIndex& operator[](Index i) const { return indices()(i); }
+    /** Direct access to the underlying index vector */
+    inline StorageIndex& operator[](Index i) { return indices()(i); }
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return derived().indices(); }
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return derived().indices(); }
+
+    /** Resizes to given size. */
+    inline void resize(Index newSize)
+    {
+      indices().resize(newSize);
+    }
+
+    /** Sets \c *this to represents an identity transformation */
+    void setIdentity()
+    {
+      for(StorageIndex i = 0; i < indices().size(); ++i)
+        coeffRef(i) = i;
+    }
+
+    // FIXME: do we want such methods ?
+    // might be usefull when the target matrix expression is complex, e.g.:
+    // object.matrix().block(..,..,..,..) = trans * object.matrix().block(..,..,..,..);
+    /*
+    template<typename MatrixType>
+    void applyForwardToRows(MatrixType& mat) const
+    {
+      for(Index k=0 ; k<size() ; ++k)
+        if(m_indices(k)!=k)
+          mat.row(k).swap(mat.row(m_indices(k)));
+    }
+
+    template<typename MatrixType>
+    void applyBackwardToRows(MatrixType& mat) const
+    {
+      for(Index k=size()-1 ; k>=0 ; --k)
+        if(m_indices(k)!=k)
+          mat.row(k).swap(mat.row(m_indices(k)));
+    }
+    */
+
+    /** \returns the inverse transformation */
+    inline Transpose<TranspositionsBase> inverse() const
+    { return Transpose<TranspositionsBase>(derived()); }
+
+    /** \returns the tranpose transformation */
+    inline Transpose<TranspositionsBase> transpose() const
+    { return Transpose<TranspositionsBase>(derived()); }
+
+  protected:
+};
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+struct traits<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+{
+  typedef Matrix<_StorageIndex, SizeAtCompileTime, 1, 0, MaxSizeAtCompileTime, 1> IndicesType;
+  typedef TranspositionsStorage StorageKind;
+};
+}
+
+/** \class Transpositions
+  * \ingroup Core_Module
+  *
+  * \brief Represents a sequence of transpositions (row/column interchange)
+  *
+  * \tparam SizeAtCompileTime the number of transpositions, or Dynamic
+  * \tparam MaxSizeAtCompileTime the maximum number of transpositions, or Dynamic. This optional parameter defaults to SizeAtCompileTime. Most of the time, you should not have to specify it.
+  *
+  * This class represents a permutation transformation as a sequence of \em n transpositions
+  * \f$[T_{n-1} \ldots T_{i} \ldots T_{0}]\f$. It is internally stored as a vector of integers \c indices.
+  * Each transposition \f$ T_{i} \f$ applied on the left of a matrix (\f$ T_{i} M\f$) interchanges
+  * the rows \c i and \c indices[i] of the matrix \c M.
+  * A transposition applied on the right (e.g., \f$ M T_{i}\f$) yields a column interchange.
+  *
+  * Compared to the class PermutationMatrix, such a sequence of transpositions is what is
+  * computed during a decomposition with pivoting, and it is faster when applying the permutation in-place.
+  *
+  * To apply a sequence of transpositions to a matrix, simply use the operator * as in the following example:
+  * \code
+  * Transpositions tr;
+  * MatrixXf mat;
+  * mat = tr * mat;
+  * \endcode
+  * In this example, we detect that the matrix appears on both side, and so the transpositions
+  * are applied in-place without any temporary or extra copy.
+  *
+  * \sa class PermutationMatrix
+  */
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex>
+class Transpositions : public TranspositionsBase<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+{
+    typedef internal::traits<Transpositions> Traits;
+  public:
+
+    typedef TranspositionsBase<Transpositions> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar StorageIndex;
+
+    inline Transpositions() {}
+
+    /** Copy constructor. */
+    template<typename OtherDerived>
+    inline Transpositions(const TranspositionsBase<OtherDerived>& other)
+      : m_indices(other.indices()) {}
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** Standard copy constructor. Defined only to prevent a default copy constructor
+      * from hiding the other templated constructor */
+    inline Transpositions(const Transpositions& other) : m_indices(other.indices()) {}
+    #endif
+
+    /** Generic constructor from expression of the transposition indices. */
+    template<typename Other>
+    explicit inline Transpositions(const MatrixBase<Other>& indices) : m_indices(indices)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Transpositions& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Transpositions& operator=(const Transpositions& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** Constructs an uninitialized permutation matrix of given size.
+      */
+    inline Transpositions(Index size) : m_indices(size)
+    {}
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+
+namespace internal {
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int _PacketAccess>
+struct traits<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,_PacketAccess> >
+ : traits<PermutationMatrix<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex> >
+{
+  typedef Map<const Matrix<_StorageIndex,SizeAtCompileTime,1,0,MaxSizeAtCompileTime,1>, _PacketAccess> IndicesType;
+  typedef _StorageIndex StorageIndex;
+  typedef TranspositionsStorage StorageKind;
+};
+}
+
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime, typename _StorageIndex, int PacketAccess>
+class Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess>
+ : public TranspositionsBase<Map<Transpositions<SizeAtCompileTime,MaxSizeAtCompileTime,_StorageIndex>,PacketAccess> >
+{
+    typedef internal::traits<Map> Traits;
+  public:
+
+    typedef TranspositionsBase<Map> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar StorageIndex;
+
+    explicit inline Map(const StorageIndex* indicesPtr)
+      : m_indices(indicesPtr)
+    {}
+
+    inline Map(const StorageIndex* indicesPtr, Index size)
+      : m_indices(indicesPtr,size)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    Map& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    Map& operator=(const Map& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+    
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    IndicesType m_indices;
+};
+
+namespace internal {
+template<typename _IndicesType>
+struct traits<TranspositionsWrapper<_IndicesType> >
+ : traits<PermutationWrapper<_IndicesType> >
+{
+  typedef TranspositionsStorage StorageKind;
+};
+}
+
+template<typename _IndicesType>
+class TranspositionsWrapper
+ : public TranspositionsBase<TranspositionsWrapper<_IndicesType> >
+{
+    typedef internal::traits<TranspositionsWrapper> Traits;
+  public:
+
+    typedef TranspositionsBase<TranspositionsWrapper> Base;
+    typedef typename Traits::IndicesType IndicesType;
+    typedef typename IndicesType::Scalar StorageIndex;
+
+    explicit inline TranspositionsWrapper(IndicesType& indices)
+      : m_indices(indices)
+    {}
+
+    /** Copies the \a other transpositions into \c *this */
+    template<typename OtherDerived>
+    TranspositionsWrapper& operator=(const TranspositionsBase<OtherDerived>& other)
+    {
+      return Base::operator=(other);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is a special case of the templated operator=. Its purpose is to
+      * prevent a default operator= from hiding the templated operator=.
+      */
+    TranspositionsWrapper& operator=(const TranspositionsWrapper& other)
+    {
+      m_indices = other.m_indices;
+      return *this;
+    }
+    #endif
+
+    /** const version of indices(). */
+    const IndicesType& indices() const { return m_indices; }
+
+    /** \returns a reference to the stored array representing the transpositions. */
+    IndicesType& indices() { return m_indices; }
+
+  protected:
+
+    typename IndicesType::Nested m_indices;
+};
+
+
+
+/** \returns the \a matrix with the \a transpositions applied to the columns.
+  */
+template<typename MatrixDerived, typename TranspositionsDerived>
+EIGEN_DEVICE_FUNC
+const Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+operator*(const MatrixBase<MatrixDerived> &matrix,
+          const TranspositionsBase<TranspositionsDerived>& transpositions)
+{
+  return Product<MatrixDerived, TranspositionsDerived, AliasFreeProduct>
+            (matrix.derived(), transpositions.derived());
+}
+
+/** \returns the \a matrix with the \a transpositions applied to the rows.
+  */
+template<typename TranspositionsDerived, typename MatrixDerived>
+EIGEN_DEVICE_FUNC
+const Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+operator*(const TranspositionsBase<TranspositionsDerived> &transpositions,
+          const MatrixBase<MatrixDerived>& matrix)
+{
+  return Product<TranspositionsDerived, MatrixDerived, AliasFreeProduct>
+            (transpositions.derived(), matrix.derived());
+}
+
+// Template partial specialization for transposed/inverse transpositions
+
+namespace internal {
+
+template<typename Derived>
+struct traits<Transpose<TranspositionsBase<Derived> > >
+ : traits<Derived>
+{};
+
+} // end namespace internal
+
+template<typename TranspositionsDerived>
+class Transpose<TranspositionsBase<TranspositionsDerived> >
+{
+    typedef TranspositionsDerived TranspositionType;
+    typedef typename TranspositionType::IndicesType IndicesType;
+  public:
+
+    explicit Transpose(const TranspositionType& t) : m_transpositions(t) {}
+
+    Index size() const { return m_transpositions.size(); }
+    Index rows() const { return m_transpositions.size(); }
+    Index cols() const { return m_transpositions.size(); }
+
+    /** \returns the \a matrix with the inverse transpositions applied to the columns.
+      */
+    template<typename OtherDerived> friend
+    const Product<OtherDerived, Transpose, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix, const Transpose& trt)
+    {
+      return Product<OtherDerived, Transpose, AliasFreeProduct>(matrix.derived(), trt.derived());
+    }
+
+    /** \returns the \a matrix with the inverse transpositions applied to the rows.
+      */
+    template<typename OtherDerived>
+    const Product<Transpose, OtherDerived, AliasFreeProduct>
+    operator*(const MatrixBase<OtherDerived>& matrix) const
+    {
+      return Product<Transpose, OtherDerived, AliasFreeProduct>(*this, matrix.derived());
+    }
+    
+    const TranspositionType& nestedExpression() const { return m_transpositions; }
+
+  protected:
+    const TranspositionType& m_transpositions;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSPOSITIONS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h
new file mode 100644
index 0000000..667ef09
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/TriangularMatrix.h
@@ -0,0 +1,983 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIANGULARMATRIX_H
+#define EIGEN_TRIANGULARMATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<int Side, typename TriangularType, typename Rhs> struct triangular_solve_retval;
+  
+}
+
+/** \class TriangularBase
+  * \ingroup Core_Module
+  *
+  * \brief Base class for triangular part in a matrix
+  */
+template<typename Derived> class TriangularBase : public EigenBase<Derived>
+{
+  public:
+
+    enum {
+      Mode = internal::traits<Derived>::Mode,
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      MaxRowsAtCompileTime = internal::traits<Derived>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<Derived>::MaxColsAtCompileTime,
+      
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                   internal::traits<Derived>::ColsAtCompileTime>::ret),
+      /**< This is equal to the number of coefficients, i.e. the number of
+          * rows times the number of columns, or to \a Dynamic if this is not
+          * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */
+      
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::MaxRowsAtCompileTime,
+                                                   internal::traits<Derived>::MaxColsAtCompileTime>::ret)
+        
+    };
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+    typedef typename internal::traits<Derived>::FullMatrixType DenseMatrixType;
+    typedef DenseMatrixType DenseType;
+    typedef Derived const& Nested;
+
+    EIGEN_DEVICE_FUNC
+    inline TriangularBase() { eigen_assert(!((Mode&UnitDiag) && (Mode&ZeroDiag))); }
+
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return derived().rows(); }
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return derived().cols(); }
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return derived().outerStride(); }
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return derived().innerStride(); }
+    
+    // dummy resize function
+    void resize(Index rows, Index cols)
+    {
+      EIGEN_UNUSED_VARIABLE(rows);
+      EIGEN_UNUSED_VARIABLE(cols);
+      eigen_assert(rows==this->rows() && cols==this->cols());
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar coeff(Index row, Index col) const  { return derived().coeff(row,col); }
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index row, Index col) { return derived().coeffRef(row,col); }
+
+    /** \see MatrixBase::copyCoeff(row,col)
+      */
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void copyCoeff(Index row, Index col, Other& other)
+    {
+      derived().coeffRef(row, col) = other.coeff(row, col);
+    }
+
+    EIGEN_DEVICE_FUNC
+    inline Scalar operator()(Index row, Index col) const
+    {
+      check_coordinates(row, col);
+      return coeff(row,col);
+    }
+    EIGEN_DEVICE_FUNC
+    inline Scalar& operator()(Index row, Index col)
+    {
+      check_coordinates(row, col);
+      return coeffRef(row,col);
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC
+    inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC
+    inline Derived& derived() { return *static_cast<Derived*>(this); }
+    #endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
+    void evalTo(MatrixBase<DenseDerived> &other) const;
+    template<typename DenseDerived>
+    EIGEN_DEVICE_FUNC
+    void evalToLazy(MatrixBase<DenseDerived> &other) const;
+
+    EIGEN_DEVICE_FUNC
+    DenseMatrixType toDenseMatrix() const
+    {
+      DenseMatrixType res(rows(), cols());
+      evalToLazy(res);
+      return res;
+    }
+
+  protected:
+
+    void check_coordinates(Index row, Index col) const
+    {
+      EIGEN_ONLY_USED_FOR_DEBUG(row);
+      EIGEN_ONLY_USED_FOR_DEBUG(col);
+      eigen_assert(col>=0 && col<cols() && row>=0 && row<rows());
+      const int mode = int(Mode) & ~SelfAdjoint;
+      EIGEN_ONLY_USED_FOR_DEBUG(mode);
+      eigen_assert((mode==Upper && col>=row)
+                || (mode==Lower && col<=row)
+                || ((mode==StrictlyUpper || mode==UnitUpper) && col>row)
+                || ((mode==StrictlyLower || mode==UnitLower) && col<row));
+    }
+
+    #ifdef EIGEN_INTERNAL_DEBUGGING
+    void check_coordinates_internal(Index row, Index col) const
+    {
+      check_coordinates(row, col);
+    }
+    #else
+    void check_coordinates_internal(Index , Index ) const {}
+    #endif
+
+};
+
+/** \class TriangularView
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a triangular part in a matrix
+  *
+  * \param MatrixType the type of the object in which we are taking the triangular part
+  * \param Mode the kind of triangular matrix expression to construct. Can be #Upper,
+  *             #Lower, #UnitUpper, #UnitLower, #StrictlyUpper, or #StrictlyLower.
+  *             This is in fact a bit field; it must have either #Upper or #Lower, 
+  *             and additionally it may have #UnitDiag or #ZeroDiag or neither.
+  *
+  * This class represents a triangular part of a matrix, not necessarily square. Strictly speaking, for rectangular
+  * matrices one should speak of "trapezoid" parts. This class is the return type
+  * of MatrixBase::triangularView() and SparseMatrixBase::triangularView(), and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::triangularView()
+  */
+namespace internal {
+template<typename MatrixType, unsigned int _Mode>
+struct traits<TriangularView<MatrixType, _Mode> > : traits<MatrixType>
+{
+  typedef typename ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type MatrixTypeNestedNonRef;
+  typedef typename remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+  typedef typename MatrixType::PlainObject FullMatrixType;
+  typedef MatrixType ExpressionType;
+  enum {
+    Mode = _Mode,
+    FlagsLvalueBit = is_lvalue<MatrixType>::value ? LvalueBit : 0,
+    Flags = (MatrixTypeNestedCleaned::Flags & (HereditaryBits | FlagsLvalueBit) & (~(PacketAccessBit | DirectAccessBit | LinearAccessBit)))
+  };
+};
+}
+
+template<typename _MatrixType, unsigned int _Mode, typename StorageKind> class TriangularViewImpl;
+
+template<typename _MatrixType, unsigned int _Mode> class TriangularView
+  : public TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind >
+{
+  public:
+
+    typedef TriangularViewImpl<_MatrixType, _Mode, typename internal::traits<_MatrixType>::StorageKind > Base;
+    typedef typename internal::traits<TriangularView>::Scalar Scalar;
+    typedef _MatrixType MatrixType;
+
+  protected:
+    typedef typename internal::traits<TriangularView>::MatrixTypeNested MatrixTypeNested;
+    typedef typename internal::traits<TriangularView>::MatrixTypeNestedNonRef MatrixTypeNestedNonRef;
+
+    typedef typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type MatrixConjugateReturnType;
+    
+  public:
+
+    typedef typename internal::traits<TriangularView>::StorageKind StorageKind;
+    typedef typename internal::traits<TriangularView>::MatrixTypeNestedCleaned NestedExpression;
+
+    enum {
+      Mode = _Mode,
+      Flags = internal::traits<TriangularView>::Flags,
+      TransposeMode = (Mode & Upper ? Lower : 0)
+                    | (Mode & Lower ? Upper : 0)
+                    | (Mode & (UnitDiag))
+                    | (Mode & (ZeroDiag)),
+      IsVectorAtCompileTime = false
+    };
+
+    EIGEN_DEVICE_FUNC
+    explicit inline TriangularView(MatrixType& matrix) : m_matrix(matrix)
+    {}
+    
+    using Base::operator=;
+    TriangularView& operator=(const TriangularView &other)
+    { return Base::operator=(other); }
+
+    /** \copydoc EigenBase::rows() */
+    EIGEN_DEVICE_FUNC
+    inline Index rows() const { return m_matrix.rows(); }
+    /** \copydoc EigenBase::cols() */
+    EIGEN_DEVICE_FUNC
+    inline Index cols() const { return m_matrix.cols(); }
+
+    /** \returns a const reference to the nested expression */
+    EIGEN_DEVICE_FUNC
+    const NestedExpression& nestedExpression() const { return m_matrix; }
+
+    /** \returns a reference to the nested expression */
+    EIGEN_DEVICE_FUNC
+    NestedExpression& nestedExpression() { return m_matrix; }
+    
+    typedef TriangularView<const MatrixConjugateReturnType,Mode> ConjugateReturnType;
+    /** \sa MatrixBase::conjugate() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConjugateReturnType conjugate() const
+    { return ConjugateReturnType(m_matrix.conjugate()); }
+
+    typedef TriangularView<const typename MatrixType::AdjointReturnType,TransposeMode> AdjointReturnType;
+    /** \sa MatrixBase::adjoint() const */
+    EIGEN_DEVICE_FUNC
+    inline const AdjointReturnType adjoint() const
+    { return AdjointReturnType(m_matrix.adjoint()); }
+
+    typedef TriangularView<typename MatrixType::TransposeReturnType,TransposeMode> TransposeReturnType;
+     /** \sa MatrixBase::transpose() */
+    EIGEN_DEVICE_FUNC
+    inline TransposeReturnType transpose()
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(MatrixType)
+      typename MatrixType::TransposeReturnType tmp(m_matrix);
+      return TransposeReturnType(tmp);
+    }
+    
+    typedef TriangularView<const typename MatrixType::ConstTransposeReturnType,TransposeMode> ConstTransposeReturnType;
+    /** \sa MatrixBase::transpose() const */
+    EIGEN_DEVICE_FUNC
+    inline const ConstTransposeReturnType transpose() const
+    {
+      return ConstTransposeReturnType(m_matrix.transpose());
+    }
+
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    inline const Solve<TriangularView, Other> 
+    solve(const MatrixBase<Other>& other) const
+    { return Solve<TriangularView, Other>(*this, other.derived()); }
+    
+  // workaround MSVC ICE
+  #if EIGEN_COMP_MSVC
+    template<int Side, typename Other>
+    EIGEN_DEVICE_FUNC
+    inline const internal::triangular_solve_retval<Side,TriangularView, Other>
+    solve(const MatrixBase<Other>& other) const
+    { return Base::template solve<Side>(other); }
+  #else
+    using Base::solve;
+  #endif
+
+    /** \returns a selfadjoint view of the referenced triangular part which must be either \c #Upper or \c #Lower.
+      *
+      * This is a shortcut for \code this->nestedExpression().selfadjointView<(*this)::Mode>() \endcode
+      * \sa MatrixBase::selfadjointView() */
+    EIGEN_DEVICE_FUNC
+    SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView()
+    {
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+    /** This is the const version of selfadjointView() */
+    EIGEN_DEVICE_FUNC
+    const SelfAdjointView<MatrixTypeNestedNonRef,Mode> selfadjointView() const
+    {
+      EIGEN_STATIC_ASSERT((Mode&(UnitDiag|ZeroDiag))==0,PROGRAMMING_ERROR);
+      return SelfAdjointView<MatrixTypeNestedNonRef,Mode>(m_matrix);
+    }
+
+
+    /** \returns the determinant of the triangular matrix
+      * \sa MatrixBase::determinant() */
+    EIGEN_DEVICE_FUNC
+    Scalar determinant() const
+    {
+      if (Mode & UnitDiag)
+        return 1;
+      else if (Mode & ZeroDiag)
+        return 0;
+      else
+        return m_matrix.diagonal().prod();
+    }
+      
+  protected:
+
+    MatrixTypeNested m_matrix;
+};
+
+/** \ingroup Core_Module
+  *
+  * \brief Base class for a triangular part in a \b dense matrix
+  *
+  * This class is an abstract base class of class TriangularView, and objects of type TriangularViewImpl cannot be instantiated.
+  * It extends class TriangularView with additional methods which available for dense expressions only.
+  *
+  * \sa class TriangularView, MatrixBase::triangularView()
+  */
+template<typename _MatrixType, unsigned int _Mode> class TriangularViewImpl<_MatrixType,_Mode,Dense>
+  : public TriangularBase<TriangularView<_MatrixType, _Mode> >
+{
+  public:
+
+    typedef TriangularView<_MatrixType, _Mode> TriangularViewType;
+    typedef TriangularBase<TriangularViewType> Base;
+    typedef typename internal::traits<TriangularViewType>::Scalar Scalar;
+
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::PlainObject DenseMatrixType;
+    typedef DenseMatrixType PlainObject;
+
+  public:
+    using Base::evalToLazy;
+    using Base::derived;
+
+    typedef typename internal::traits<TriangularViewType>::StorageKind StorageKind;
+
+    enum {
+      Mode = _Mode,
+      Flags = internal::traits<TriangularViewType>::Flags
+    };
+
+    /** \returns the outer-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::outerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index outerStride() const { return derived().nestedExpression().outerStride(); }
+    /** \returns the inner-stride of the underlying dense matrix
+      * \sa DenseCoeffsBase::innerStride() */
+    EIGEN_DEVICE_FUNC
+    inline Index innerStride() const { return derived().nestedExpression().innerStride(); }
+
+    /** \sa MatrixBase::operator+=() */
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator+=(const DenseBase<Other>& other) {
+      internal::call_assignment_no_alias(derived(), other.derived(), internal::add_assign_op<Scalar,typename Other::Scalar>());
+      return derived();
+    }
+    /** \sa MatrixBase::operator-=() */
+    template<typename Other>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator-=(const DenseBase<Other>& other) {
+      internal::call_assignment_no_alias(derived(), other.derived(), internal::sub_assign_op<Scalar,typename Other::Scalar>());
+      return derived();
+    }
+    
+    /** \sa MatrixBase::operator*=() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator*=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() * other; }
+    /** \sa DenseBase::operator/=() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType&  operator/=(const typename internal::traits<MatrixType>::Scalar& other) { return *this = derived().nestedExpression() / other; }
+
+    /** \sa MatrixBase::fill() */
+    EIGEN_DEVICE_FUNC
+    void fill(const Scalar& value) { setConstant(value); }
+    /** \sa MatrixBase::setConstant() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setConstant(const Scalar& value)
+    { return *this = MatrixType::Constant(derived().rows(), derived().cols(), value); }
+    /** \sa MatrixBase::setZero() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setZero() { return setConstant(Scalar(0)); }
+    /** \sa MatrixBase::setOnes() */
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& setOnes() { return setConstant(Scalar(1)); }
+
+    /** \sa MatrixBase::coeff()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    EIGEN_DEVICE_FUNC
+    inline Scalar coeff(Index row, Index col) const
+    {
+      Base::check_coordinates_internal(row, col);
+      return derived().nestedExpression().coeff(row, col);
+    }
+
+    /** \sa MatrixBase::coeffRef()
+      * \warning the coordinates must fit into the referenced triangular part
+      */
+    EIGEN_DEVICE_FUNC
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(TriangularViewType);
+      Base::check_coordinates_internal(row, col);
+      return derived().nestedExpression().coeffRef(row, col);
+    }
+
+    /** Assigns a triangular matrix to a triangular part of a dense matrix */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const TriangularBase<OtherDerived>& other);
+
+    /** Shortcut for\code *this = other.other.triangularView<(*this)::Mode>() \endcode */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const MatrixBase<OtherDerived>& other);
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC
+    TriangularViewType& operator=(const TriangularViewImpl& other)
+    { return *this = other.derived().nestedExpression(); }
+
+    /** \deprecated */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void lazyAssign(const TriangularBase<OtherDerived>& other);
+
+    /** \deprecated */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void lazyAssign(const MatrixBase<OtherDerived>& other);
+#endif
+
+    /** Efficient triangular matrix times vector/matrix product */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const Product<TriangularViewType,OtherDerived>
+    operator*(const MatrixBase<OtherDerived>& rhs) const
+    {
+      return Product<TriangularViewType,OtherDerived>(derived(), rhs.derived());
+    }
+
+    /** Efficient vector/matrix times triangular matrix product */
+    template<typename OtherDerived> friend
+    EIGEN_DEVICE_FUNC
+    const Product<OtherDerived,TriangularViewType>
+    operator*(const MatrixBase<OtherDerived>& lhs, const TriangularViewImpl& rhs)
+    {
+      return Product<OtherDerived,TriangularViewType>(lhs.derived(),rhs.derived());
+    }
+
+    /** \returns the product of the inverse of \c *this with \a other, \a *this being triangular.
+      *
+      * This function computes the inverse-matrix matrix product inverse(\c *this) * \a other if
+      * \a Side==OnTheLeft (the default), or the right-inverse-multiply  \a other * inverse(\c *this) if
+      * \a Side==OnTheRight.
+      *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
+      * The matrix \c *this must be triangular and invertible (i.e., all the coefficients of the
+      * diagonal must be non zero). It works as a forward (resp. backward) substitution if \c *this
+      * is an upper (resp. lower) triangular matrix.
+      *
+      * Example: \include Triangular_solve.cpp
+      * Output: \verbinclude Triangular_solve.out
+      *
+      * This function returns an expression of the inverse-multiply and can works in-place if it is assigned
+      * to the same matrix or vector \a other.
+      *
+      * For users coming from BLAS, this function (and more specifically solveInPlace()) offer
+      * all the operations supported by the \c *TRSV and \c *TRSM BLAS routines.
+      *
+      * \sa TriangularView::solveInPlace()
+      */
+    template<int Side, typename Other>
+    EIGEN_DEVICE_FUNC
+    inline const internal::triangular_solve_retval<Side,TriangularViewType, Other>
+    solve(const MatrixBase<Other>& other) const;
+
+    /** "in-place" version of TriangularView::solve() where the result is written in \a other
+      *
+      * \warning The parameter is only marked 'const' to make the C++ compiler accept a temporary expression here.
+      * This function will const_cast it, so constness isn't honored here.
+      *
+      * Note that the template parameter \c Side can be ommitted, in which case \c Side==OnTheLeft
+      *
+      * See TriangularView:solve() for the details.
+      */
+    template<int Side, typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void solveInPlace(const MatrixBase<OtherDerived>& other) const;
+
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void solveInPlace(const MatrixBase<OtherDerived>& other) const
+    { return solveInPlace<OnTheLeft>(other); }
+
+    /** Swaps the coefficients of the common triangular parts of two matrices */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+    void swap(TriangularBase<OtherDerived> &other)
+#else
+    void swap(TriangularBase<OtherDerived> const & other)
+#endif
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
+    }
+
+    /** \deprecated
+      * Shortcut for \code (*this).swap(other.triangularView<(*this)::Mode>()) \endcode */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    void swap(MatrixBase<OtherDerived> const & other)
+    {
+      EIGEN_STATIC_ASSERT_LVALUE(OtherDerived);
+      call_assignment(derived(), other.const_cast_derived(), internal::swap_assign_op<Scalar>());
+    }
+
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const {
+      if(!internal::is_same_dense(dst,rhs))
+        dst = rhs;
+      this->solveInPlace(dst);
+    }
+
+    template<typename ProductType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE TriangularViewType& _assignProduct(const ProductType& prod, const Scalar& alpha, bool beta);
+};
+
+/***************************************************************************
+* Implementation of triangular evaluation/assignment
+***************************************************************************/
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+// FIXME should we keep that possibility
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+inline TriangularView<MatrixType, Mode>&
+TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const MatrixBase<OtherDerived>& other)
+{
+  internal::call_assignment_no_alias(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+// FIXME should we keep that possibility
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const MatrixBase<OtherDerived>& other)
+{
+  internal::call_assignment_no_alias(derived(), other.template triangularView<Mode>());
+}
+
+
+
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+inline TriangularView<MatrixType, Mode>&
+TriangularViewImpl<MatrixType, Mode, Dense>::operator=(const TriangularBase<OtherDerived>& other)
+{
+  eigen_assert(Mode == int(OtherDerived::Mode));
+  internal::call_assignment(derived(), other.derived());
+  return derived();
+}
+
+template<typename MatrixType, unsigned int Mode>
+template<typename OtherDerived>
+void TriangularViewImpl<MatrixType, Mode, Dense>::lazyAssign(const TriangularBase<OtherDerived>& other)
+{
+  eigen_assert(Mode == int(OtherDerived::Mode));
+  internal::call_assignment_no_alias(derived(), other.derived());
+}
+#endif
+
+/***************************************************************************
+* Implementation of TriangularBase methods
+***************************************************************************/
+
+/** Assigns a triangular or selfadjoint matrix to a dense matrix.
+  * If the matrix is triangular, the opposite part is set to zero. */
+template<typename Derived>
+template<typename DenseDerived>
+void TriangularBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
+{
+  evalToLazy(other.derived());
+}
+
+/***************************************************************************
+* Implementation of TriangularView methods
+***************************************************************************/
+
+/***************************************************************************
+* Implementation of MatrixBase methods
+***************************************************************************/
+
+/**
+  * \returns an expression of a triangular view extracted from the current matrix
+  *
+  * The parameter \a Mode can have the following values: \c #Upper, \c #StrictlyUpper, \c #UnitUpper,
+  * \c #Lower, \c #StrictlyLower, \c #UnitLower.
+  *
+  * Example: \include MatrixBase_triangularView.cpp
+  * Output: \verbinclude MatrixBase_triangularView.out
+  *
+  * \sa class TriangularView
+  */
+template<typename Derived>
+template<unsigned int Mode>
+typename MatrixBase<Derived>::template TriangularViewReturnType<Mode>::Type
+MatrixBase<Derived>::triangularView()
+{
+  return typename TriangularViewReturnType<Mode>::Type(derived());
+}
+
+/** This is the const version of MatrixBase::triangularView() */
+template<typename Derived>
+template<unsigned int Mode>
+typename MatrixBase<Derived>::template ConstTriangularViewReturnType<Mode>::Type
+MatrixBase<Derived>::triangularView() const
+{
+  return typename ConstTriangularViewReturnType<Mode>::Type(derived());
+}
+
+/** \returns true if *this is approximately equal to an upper triangular matrix,
+  *          within the precision given by \a prec.
+  *
+  * \sa isLowerTriangular()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isUpperTriangular(const RealScalar& prec) const
+{
+  RealScalar maxAbsOnUpperPart = static_cast<RealScalar>(-1);
+  for(Index j = 0; j < cols(); ++j)
+  {
+    Index maxi = numext::mini(j, rows()-1);
+    for(Index i = 0; i <= maxi; ++i)
+    {
+      RealScalar absValue = numext::abs(coeff(i,j));
+      if(absValue > maxAbsOnUpperPart) maxAbsOnUpperPart = absValue;
+    }
+  }
+  RealScalar threshold = maxAbsOnUpperPart * prec;
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = j+1; i < rows(); ++i)
+      if(numext::abs(coeff(i, j)) > threshold) return false;
+  return true;
+}
+
+/** \returns true if *this is approximately equal to a lower triangular matrix,
+  *          within the precision given by \a prec.
+  *
+  * \sa isUpperTriangular()
+  */
+template<typename Derived>
+bool MatrixBase<Derived>::isLowerTriangular(const RealScalar& prec) const
+{
+  RealScalar maxAbsOnLowerPart = static_cast<RealScalar>(-1);
+  for(Index j = 0; j < cols(); ++j)
+    for(Index i = j; i < rows(); ++i)
+    {
+      RealScalar absValue = numext::abs(coeff(i,j));
+      if(absValue > maxAbsOnLowerPart) maxAbsOnLowerPart = absValue;
+    }
+  RealScalar threshold = maxAbsOnLowerPart * prec;
+  for(Index j = 1; j < cols(); ++j)
+  {
+    Index maxi = numext::mini(j, rows()-1);
+    for(Index i = 0; i < maxi; ++i)
+      if(numext::abs(coeff(i, j)) > threshold) return false;
+  }
+  return true;
+}
+
+
+/***************************************************************************
+****************************************************************************
+* Evaluators and Assignment of triangular expressions
+***************************************************************************
+***************************************************************************/
+
+namespace internal {
+
+  
+// TODO currently a triangular expression has the form TriangularView<.,.>
+//      in the future triangular-ness should be defined by the expression traits
+//      such that Transpose<TriangularView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<TriangularView<MatrixType,Mode> >
+{
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef typename glue_shapes<typename evaluator_traits<MatrixType>::Shape, TriangularShape>::type Shape;
+};
+
+template<typename MatrixType, unsigned int Mode>
+struct unary_evaluator<TriangularView<MatrixType,Mode>, IndexBased>
+ : evaluator<typename internal::remove_all<MatrixType>::type>
+{
+  typedef TriangularView<MatrixType,Mode> XprType;
+  typedef evaluator<typename internal::remove_all<MatrixType>::type> Base;
+  unary_evaluator(const XprType &xpr) : Base(xpr.nestedExpression()) {}
+};
+
+// Additional assignment kinds:
+struct Triangular2Triangular    {};
+struct Triangular2Dense         {};
+struct Dense2Triangular         {};
+
+
+template<typename Kernel, unsigned int Mode, int UnrollCount, bool ClearOpposite> struct triangular_assignment_loop;
+
+ 
+/** \internal Specialization of the dense assignment kernel for triangular matrices.
+  * The main difference is that the triangular, diagonal, and opposite parts are processed through three different functions.
+  * \tparam UpLo must be either Lower or Upper
+  * \tparam Mode must be either 0, UnitDiag, ZeroDiag, or SelfAdjoint
+  */
+template<int UpLo, int Mode, int SetOpposite, typename DstEvaluatorTypeT, typename SrcEvaluatorTypeT, typename Functor, int Version = Specialized>
+class triangular_dense_assignment_kernel : public generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version>
+{
+protected:
+  typedef generic_dense_assignment_kernel<DstEvaluatorTypeT, SrcEvaluatorTypeT, Functor, Version> Base;
+  typedef typename Base::DstXprType DstXprType;
+  typedef typename Base::SrcXprType SrcXprType;
+  using Base::m_dst;
+  using Base::m_src;
+  using Base::m_functor;
+public:
+  
+  typedef typename Base::DstEvaluatorType DstEvaluatorType;
+  typedef typename Base::SrcEvaluatorType SrcEvaluatorType;
+  typedef typename Base::Scalar Scalar;
+  typedef typename Base::AssignmentTraits AssignmentTraits;
+  
+  
+  EIGEN_DEVICE_FUNC triangular_dense_assignment_kernel(DstEvaluatorType &dst, const SrcEvaluatorType &src, const Functor &func, DstXprType& dstExpr)
+    : Base(dst, src, func, dstExpr)
+  {}
+  
+#ifdef EIGEN_INTERNAL_DEBUGGING
+  EIGEN_DEVICE_FUNC void assignCoeff(Index row, Index col)
+  {
+    eigen_internal_assert(row!=col);
+    Base::assignCoeff(row,col);
+  }
+#else
+  using Base::assignCoeff;
+#endif
+  
+  EIGEN_DEVICE_FUNC void assignDiagonalCoeff(Index id)
+  {
+         if(Mode==UnitDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(1));
+    else if(Mode==ZeroDiag && SetOpposite) m_functor.assignCoeff(m_dst.coeffRef(id,id), Scalar(0));
+    else if(Mode==0)                       Base::assignCoeff(id,id);
+  }
+  
+  EIGEN_DEVICE_FUNC void assignOppositeCoeff(Index row, Index col)
+  { 
+    eigen_internal_assert(row!=col);
+    if(SetOpposite)
+      m_functor.assignCoeff(m_dst.coeffRef(row,col), Scalar(0));
+  }
+};
+
+template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType, typename Functor>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src, const Functor &func)
+{
+  typedef evaluator<DstXprType> DstEvaluatorType;
+  typedef evaluator<SrcXprType> SrcEvaluatorType;
+
+  SrcEvaluatorType srcEvaluator(src);
+
+  Index dstRows = src.rows();
+  Index dstCols = src.cols();
+  if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+    dst.resize(dstRows, dstCols);
+  DstEvaluatorType dstEvaluator(dst);
+    
+  typedef triangular_dense_assignment_kernel< Mode&(Lower|Upper),Mode&(UnitDiag|ZeroDiag|SelfAdjoint),SetOpposite,
+                                              DstEvaluatorType,SrcEvaluatorType,Functor> Kernel;
+  Kernel kernel(dstEvaluator, srcEvaluator, func, dst.const_cast_derived());
+  
+  enum {
+      unroll = DstXprType::SizeAtCompileTime != Dynamic
+            && SrcEvaluatorType::CoeffReadCost < HugeCost
+            && DstXprType::SizeAtCompileTime * (DstEvaluatorType::CoeffReadCost+SrcEvaluatorType::CoeffReadCost) / 2 <= EIGEN_UNROLLING_LIMIT
+    };
+  
+  triangular_assignment_loop<Kernel, Mode, unroll ? int(DstXprType::SizeAtCompileTime) : Dynamic, SetOpposite>::run(kernel);
+}
+
+template<int Mode, bool SetOpposite, typename DstXprType, typename SrcXprType>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+void call_triangular_assignment_loop(DstXprType& dst, const SrcXprType& src)
+{
+  call_triangular_assignment_loop<Mode,SetOpposite>(dst, src, internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>());
+}
+
+template<> struct AssignmentKind<TriangularShape,TriangularShape> { typedef Triangular2Triangular Kind; };
+template<> struct AssignmentKind<DenseShape,TriangularShape>      { typedef Triangular2Dense      Kind; };
+template<> struct AssignmentKind<TriangularShape,DenseShape>      { typedef Dense2Triangular      Kind; };
+
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Triangular>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    eigen_assert(int(DstXprType::Mode) == int(SrcXprType::Mode));
+    
+    call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func);  
+  }
+};
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Triangular2Dense>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    call_triangular_assignment_loop<SrcXprType::Mode, (SrcXprType::Mode&SelfAdjoint)==0>(dst, src, func);  
+  }
+};
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Dense2Triangular>
+{
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    call_triangular_assignment_loop<DstXprType::Mode, false>(dst, src, func);  
+  }
+};
+
+
+template<typename Kernel, unsigned int Mode, int UnrollCount, bool SetOpposite>
+struct triangular_assignment_loop
+{
+  // FIXME: this is not very clean, perhaps this information should be provided by the kernel?
+  typedef typename Kernel::DstEvaluatorType DstEvaluatorType;
+  typedef typename DstEvaluatorType::XprType DstXprType;
+  
+  enum {
+    col = (UnrollCount-1) / DstXprType::RowsAtCompileTime,
+    row = (UnrollCount-1) % DstXprType::RowsAtCompileTime
+  };
+  
+  typedef typename Kernel::Scalar Scalar;
+
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &kernel)
+  {
+    triangular_assignment_loop<Kernel, Mode, UnrollCount-1, SetOpposite>::run(kernel);
+    
+    if(row==col)
+      kernel.assignDiagonalCoeff(row);
+    else if( ((Mode&Lower) && row>col) || ((Mode&Upper) && row<col) )
+      kernel.assignCoeff(row,col);
+    else if(SetOpposite)
+      kernel.assignOppositeCoeff(row,col);
+  }
+};
+
+// prevent buggy user code from causing an infinite recursion
+template<typename Kernel, unsigned int Mode, bool SetOpposite>
+struct triangular_assignment_loop<Kernel, Mode, 0, SetOpposite>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &) {}
+};
+
+
+
+// TODO: experiment with a recursive assignment procedure splitting the current
+//       triangular part into one rectangular and two triangular parts.
+
+
+template<typename Kernel, unsigned int Mode, bool SetOpposite>
+struct triangular_assignment_loop<Kernel, Mode, Dynamic, SetOpposite>
+{
+  typedef typename Kernel::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
+  static inline void run(Kernel &kernel)
+  {
+    for(Index j = 0; j < kernel.cols(); ++j)
+    {
+      Index maxi = numext::mini(j, kernel.rows());
+      Index i = 0;
+      if (((Mode&Lower) && SetOpposite) || (Mode&Upper))
+      {
+        for(; i < maxi; ++i)
+          if(Mode&Upper) kernel.assignCoeff(i, j);
+          else           kernel.assignOppositeCoeff(i, j);
+      }
+      else
+        i = maxi;
+      
+      if(i<kernel.rows()) // then i==j
+        kernel.assignDiagonalCoeff(i++);
+      
+      if (((Mode&Upper) && SetOpposite) || (Mode&Lower))
+      {
+        for(; i < kernel.rows(); ++i)
+          if(Mode&Lower) kernel.assignCoeff(i, j);
+          else           kernel.assignOppositeCoeff(i, j);
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+/** Assigns a triangular or selfadjoint matrix to a dense matrix.
+  * If the matrix is triangular, the opposite part is set to zero. */
+template<typename Derived>
+template<typename DenseDerived>
+void TriangularBase<Derived>::evalToLazy(MatrixBase<DenseDerived> &other) const
+{
+  other.derived().resize(this->rows(), this->cols());
+  internal::call_triangular_assignment_loop<Derived::Mode,(Derived::Mode&SelfAdjoint)==0 /* SetOpposite */>(other.derived(), derived().nestedExpression());
+}
+
+namespace internal {
+  
+// Triangular = Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst._assignProduct(src, 1, 0);
+  }
+};
+
+// Triangular += Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::add_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    dst._assignProduct(src, 1, 1);
+  }
+};
+
+// Triangular -= Product
+template< typename DstXprType, typename Lhs, typename Rhs, typename Scalar>
+struct Assignment<DstXprType, Product<Lhs,Rhs,DefaultProduct>, internal::sub_assign_op<Scalar,typename Product<Lhs,Rhs,DefaultProduct>::Scalar>, Dense2Triangular>
+{
+  typedef Product<Lhs,Rhs,DefaultProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<Scalar,typename SrcXprType::Scalar> &)
+  {
+    dst._assignProduct(src, -1, 1);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULARMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h
new file mode 100644
index 0000000..d72fbf7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorBlock.h
@@ -0,0 +1,96 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_VECTORBLOCK_H
+#define EIGEN_VECTORBLOCK_H
+
+namespace Eigen { 
+
+namespace internal {
+template<typename VectorType, int Size>
+struct traits<VectorBlock<VectorType, Size> >
+  : public traits<Block<VectorType,
+                     traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     traits<VectorType>::Flags & RowMajorBit ? Size : 1> >
+{
+};
+}
+
+/** \class VectorBlock
+  * \ingroup Core_Module
+  *
+  * \brief Expression of a fixed-size or dynamic-size sub-vector
+  *
+  * \tparam VectorType the type of the object in which we are taking a sub-vector
+  * \tparam Size size of the sub-vector we are taking at compile time (optional)
+  *
+  * This class represents an expression of either a fixed-size or dynamic-size sub-vector.
+  * It is the return type of DenseBase::segment(Index,Index) and DenseBase::segment<int>(Index) and
+  * most of the time this is the only way it is used.
+  *
+  * However, if you want to directly maniputate sub-vector expressions,
+  * for instance if you want to write a function returning such an expression, you
+  * will need to use this class.
+  *
+  * Here is an example illustrating the dynamic case:
+  * \include class_VectorBlock.cpp
+  * Output: \verbinclude class_VectorBlock.out
+  *
+  * \note Even though this expression has dynamic size, in the case where \a VectorType
+  * has fixed size, this expression inherits a fixed maximal size which means that evaluating
+  * it does not cause a dynamic memory allocation.
+  *
+  * Here is an example illustrating the fixed-size case:
+  * \include class_FixedVectorBlock.cpp
+  * Output: \verbinclude class_FixedVectorBlock.out
+  *
+  * \sa class Block, DenseBase::segment(Index,Index,Index,Index), DenseBase::segment(Index,Index)
+  */
+template<typename VectorType, int Size> class VectorBlock
+  : public Block<VectorType,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1>
+{
+    typedef Block<VectorType,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? 1 : Size,
+                     internal::traits<VectorType>::Flags & RowMajorBit ? Size : 1> Base;
+    enum {
+      IsColVector = !(internal::traits<VectorType>::Flags & RowMajorBit)
+    };
+  public:
+    EIGEN_DENSE_PUBLIC_INTERFACE(VectorBlock)
+
+    using Base::operator=;
+
+    /** Dynamic-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline VectorBlock(VectorType& vector, Index start, Index size)
+      : Base(vector,
+             IsColVector ? start : 0, IsColVector ? 0 : start,
+             IsColVector ? size  : 1, IsColVector ? 1 : size)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock);
+    }
+
+    /** Fixed-size constructor
+      */
+    EIGEN_DEVICE_FUNC
+    inline VectorBlock(VectorType& vector, Index start)
+      : Base(vector, IsColVector ? start : 0, IsColVector ? 0 : start)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(VectorBlock);
+    }
+};
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_VECTORBLOCK_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h
new file mode 100644
index 0000000..4fe267e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/VectorwiseOp.h
@@ -0,0 +1,695 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARTIAL_REDUX_H
+#define EIGEN_PARTIAL_REDUX_H
+
+namespace Eigen {
+
+/** \class PartialReduxExpr
+  * \ingroup Core_Module
+  *
+  * \brief Generic expression of a partially reduxed matrix
+  *
+  * \tparam MatrixType the type of the matrix we are applying the redux operation
+  * \tparam MemberOp type of the member functor
+  * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal)
+  *
+  * This class represents an expression of a partial redux operator of a matrix.
+  * It is the return type of some VectorwiseOp functions,
+  * and most of the time this is the only way it is used.
+  *
+  * \sa class VectorwiseOp
+  */
+
+template< typename MatrixType, typename MemberOp, int Direction>
+class PartialReduxExpr;
+
+namespace internal {
+template<typename MatrixType, typename MemberOp, int Direction>
+struct traits<PartialReduxExpr<MatrixType, MemberOp, Direction> >
+ : traits<MatrixType>
+{
+  typedef typename MemberOp::result_type Scalar;
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename traits<MatrixType>::XprKind XprKind;
+  typedef typename MatrixType::Scalar InputScalar;
+  enum {
+    RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
+    Flags = RowsAtCompileTime == 1 ? RowMajorBit : 0,
+    TraversalSize = Direction==Vertical ? MatrixType::RowsAtCompileTime :  MatrixType::ColsAtCompileTime
+  };
+};
+}
+
+template< typename MatrixType, typename MemberOp, int Direction>
+class PartialReduxExpr : public internal::dense_xpr_base< PartialReduxExpr<MatrixType, MemberOp, Direction> >::type,
+                         internal::no_assignment_operator
+{
+  public:
+
+    typedef typename internal::dense_xpr_base<PartialReduxExpr>::type Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(PartialReduxExpr)
+
+    EIGEN_DEVICE_FUNC
+    explicit PartialReduxExpr(const MatrixType& mat, const MemberOp& func = MemberOp())
+      : m_matrix(mat), m_functor(func) {}
+
+    EIGEN_DEVICE_FUNC
+    Index rows() const { return (Direction==Vertical   ? 1 : m_matrix.rows()); }
+    EIGEN_DEVICE_FUNC
+    Index cols() const { return (Direction==Horizontal ? 1 : m_matrix.cols()); }
+
+    EIGEN_DEVICE_FUNC
+    typename MatrixType::Nested nestedExpression() const { return m_matrix; }
+
+    EIGEN_DEVICE_FUNC
+    const MemberOp& functor() const { return m_functor; }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+    const MemberOp m_functor;
+};
+
+#define EIGEN_MEMBER_FUNCTOR(MEMBER,COST)                               \
+  template <typename ResultType>                                        \
+  struct member_##MEMBER {                                              \
+    EIGEN_EMPTY_STRUCT_CTOR(member_##MEMBER)                            \
+    typedef ResultType result_type;                                     \
+    template<typename Scalar, int Size> struct Cost                     \
+    { enum { value = COST }; };                                         \
+    template<typename XprType>                                          \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE                               \
+    ResultType operator()(const XprType& mat) const                     \
+    { return mat.MEMBER(); } \
+  }
+
+namespace internal {
+
+EIGEN_MEMBER_FUNCTOR(squaredNorm, Size * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(norm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(stableNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(blueNorm, (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(hypotNorm, (Size-1) * functor_traits<scalar_hypot_op<Scalar> >::Cost );
+EIGEN_MEMBER_FUNCTOR(sum, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(mean, (Size-1)*NumTraits<Scalar>::AddCost + NumTraits<Scalar>::MulCost);
+EIGEN_MEMBER_FUNCTOR(minCoeff, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(maxCoeff, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(all, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(any, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(count, (Size-1)*NumTraits<Scalar>::AddCost);
+EIGEN_MEMBER_FUNCTOR(prod, (Size-1)*NumTraits<Scalar>::MulCost);
+
+template <int p, typename ResultType>
+struct member_lpnorm {
+  typedef ResultType result_type;
+  template<typename Scalar, int Size> struct Cost
+  { enum { value = (Size+5) * NumTraits<Scalar>::MulCost + (Size-1)*NumTraits<Scalar>::AddCost }; };
+  EIGEN_DEVICE_FUNC member_lpnorm() {}
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC inline ResultType operator()(const XprType& mat) const
+  { return mat.template lpNorm<p>(); }
+};
+
+template <typename BinaryOp, typename Scalar>
+struct member_redux {
+  typedef typename result_of<
+                     BinaryOp(const Scalar&,const Scalar&)
+                   >::type  result_type;
+  template<typename _Scalar, int Size> struct Cost
+  { enum { value = (Size-1) * functor_traits<BinaryOp>::Cost }; };
+  EIGEN_DEVICE_FUNC explicit member_redux(const BinaryOp func) : m_functor(func) {}
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline result_type operator()(const DenseBase<Derived>& mat) const
+  { return mat.redux(m_functor); }
+  const BinaryOp m_functor;
+};
+}
+
+/** \class VectorwiseOp
+  * \ingroup Core_Module
+  *
+  * \brief Pseudo expression providing partial reduction operations
+  *
+  * \tparam ExpressionType the type of the object on which to do partial reductions
+  * \tparam Direction indicates the direction of the redux (#Vertical or #Horizontal)
+  *
+  * This class represents a pseudo expression with partial reduction features.
+  * It is the return type of DenseBase::colwise() and DenseBase::rowwise()
+  * and most of the time this is the only way it is used.
+  *
+  * Example: \include MatrixBase_colwise.cpp
+  * Output: \verbinclude MatrixBase_colwise.out
+  *
+  * \sa DenseBase::colwise(), DenseBase::rowwise(), class PartialReduxExpr
+  */
+template<typename ExpressionType, int Direction> class VectorwiseOp
+{
+  public:
+
+    typedef typename ExpressionType::Scalar Scalar;
+    typedef typename ExpressionType::RealScalar RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef typename internal::ref_selector<ExpressionType>::non_const_type ExpressionTypeNested;
+    typedef typename internal::remove_all<ExpressionTypeNested>::type ExpressionTypeNestedCleaned;
+
+    template<template<typename _Scalar> class Functor,
+                      typename Scalar_=Scalar> struct ReturnType
+    {
+      typedef PartialReduxExpr<ExpressionType,
+                               Functor<Scalar_>,
+                               Direction
+                              > Type;
+    };
+
+    template<typename BinaryOp> struct ReduxReturnType
+    {
+      typedef PartialReduxExpr<ExpressionType,
+                               internal::member_redux<BinaryOp,Scalar>,
+                               Direction
+                              > Type;
+    };
+
+    enum {
+      isVertical   = (Direction==Vertical) ? 1 : 0,
+      isHorizontal = (Direction==Horizontal) ? 1 : 0
+    };
+
+  protected:
+
+    typedef typename internal::conditional<isVertical,
+                               typename ExpressionType::ColXpr,
+                               typename ExpressionType::RowXpr>::type SubVector;
+    /** \internal
+      * \returns the i-th subvector according to the \c Direction */
+    EIGEN_DEVICE_FUNC
+    SubVector subVector(Index i)
+    {
+      return SubVector(m_matrix.derived(),i);
+    }
+
+    /** \internal
+      * \returns the number of subvectors in the direction \c Direction */
+    EIGEN_DEVICE_FUNC
+    Index subVectors() const
+    { return isVertical?m_matrix.cols():m_matrix.rows(); }
+
+    template<typename OtherDerived> struct ExtendedType {
+      typedef Replicate<OtherDerived,
+                        isVertical   ? 1 : ExpressionType::RowsAtCompileTime,
+                        isHorizontal ? 1 : ExpressionType::ColsAtCompileTime> Type;
+    };
+
+    /** \internal
+      * Replicates a vector to match the size of \c *this */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    typename ExtendedType<OtherDerived>::Type
+    extendedTo(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxColsAtCompileTime==1),
+                          YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxRowsAtCompileTime==1),
+                          YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
+      return typename ExtendedType<OtherDerived>::Type
+                      (other.derived(),
+                       isVertical   ? 1 : m_matrix.rows(),
+                       isHorizontal ? 1 : m_matrix.cols());
+    }
+
+    template<typename OtherDerived> struct OppositeExtendedType {
+      typedef Replicate<OtherDerived,
+                        isHorizontal ? 1 : ExpressionType::RowsAtCompileTime,
+                        isVertical   ? 1 : ExpressionType::ColsAtCompileTime> Type;
+    };
+
+    /** \internal
+      * Replicates a vector in the opposite direction to match the size of \c *this */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    typename OppositeExtendedType<OtherDerived>::Type
+    extendedToOpposite(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isHorizontal, OtherDerived::MaxColsAtCompileTime==1),
+                          YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED)
+      EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(isVertical, OtherDerived::MaxRowsAtCompileTime==1),
+                          YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED)
+      return typename OppositeExtendedType<OtherDerived>::Type
+                      (other.derived(),
+                       isHorizontal  ? 1 : m_matrix.rows(),
+                       isVertical    ? 1 : m_matrix.cols());
+    }
+
+  public:
+    EIGEN_DEVICE_FUNC
+    explicit inline VectorwiseOp(ExpressionType& matrix) : m_matrix(matrix) {}
+
+    /** \internal */
+    EIGEN_DEVICE_FUNC
+    inline const ExpressionType& _expression() const { return m_matrix; }
+
+    /** \returns a row or column vector expression of \c *this reduxed by \a func
+      *
+      * The template parameter \a BinaryOp is the type of the functor
+      * of the custom redux operator. Note that func must be an associative operator.
+      *
+      * \sa class VectorwiseOp, DenseBase::colwise(), DenseBase::rowwise()
+      */
+    template<typename BinaryOp>
+    EIGEN_DEVICE_FUNC
+    const typename ReduxReturnType<BinaryOp>::Type
+    redux(const BinaryOp& func = BinaryOp()) const
+    { return typename ReduxReturnType<BinaryOp>::Type(_expression(), internal::member_redux<BinaryOp,Scalar>(func)); }
+
+    typedef typename ReturnType<internal::member_minCoeff>::Type MinCoeffReturnType;
+    typedef typename ReturnType<internal::member_maxCoeff>::Type MaxCoeffReturnType;
+    typedef typename ReturnType<internal::member_squaredNorm,RealScalar>::Type SquaredNormReturnType;
+    typedef typename ReturnType<internal::member_norm,RealScalar>::Type NormReturnType;
+    typedef typename ReturnType<internal::member_blueNorm,RealScalar>::Type BlueNormReturnType;
+    typedef typename ReturnType<internal::member_stableNorm,RealScalar>::Type StableNormReturnType;
+    typedef typename ReturnType<internal::member_hypotNorm,RealScalar>::Type HypotNormReturnType;
+    typedef typename ReturnType<internal::member_sum>::Type SumReturnType;
+    typedef typename ReturnType<internal::member_mean>::Type MeanReturnType;
+    typedef typename ReturnType<internal::member_all>::Type AllReturnType;
+    typedef typename ReturnType<internal::member_any>::Type AnyReturnType;
+    typedef PartialReduxExpr<ExpressionType, internal::member_count<Index>, Direction> CountReturnType;
+    typedef typename ReturnType<internal::member_prod>::Type ProdReturnType;
+    typedef Reverse<const ExpressionType, Direction> ConstReverseReturnType;
+    typedef Reverse<ExpressionType, Direction> ReverseReturnType;
+
+    template<int p> struct LpNormReturnType {
+      typedef PartialReduxExpr<ExpressionType, internal::member_lpnorm<p,RealScalar>,Direction> Type;
+    };
+
+    /** \returns a row (or column) vector expression of the smallest coefficient
+      * of each column (or row) of the referenced expression.
+      *
+      * \warning the result is undefined if \c *this contains NaN.
+      *
+      * Example: \include PartialRedux_minCoeff.cpp
+      * Output: \verbinclude PartialRedux_minCoeff.out
+      *
+      * \sa DenseBase::minCoeff() */
+    EIGEN_DEVICE_FUNC
+    const MinCoeffReturnType minCoeff() const
+    { return MinCoeffReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the largest coefficient
+      * of each column (or row) of the referenced expression.
+      *
+      * \warning the result is undefined if \c *this contains NaN.
+      *
+      * Example: \include PartialRedux_maxCoeff.cpp
+      * Output: \verbinclude PartialRedux_maxCoeff.out
+      *
+      * \sa DenseBase::maxCoeff() */
+    EIGEN_DEVICE_FUNC
+    const MaxCoeffReturnType maxCoeff() const
+    { return MaxCoeffReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the squared norm
+      * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * Example: \include PartialRedux_squaredNorm.cpp
+      * Output: \verbinclude PartialRedux_squaredNorm.out
+      *
+      * \sa DenseBase::squaredNorm() */
+    EIGEN_DEVICE_FUNC
+    const SquaredNormReturnType squaredNorm() const
+    { return SquaredNormReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * Example: \include PartialRedux_norm.cpp
+      * Output: \verbinclude PartialRedux_norm.out
+      *
+      * \sa DenseBase::norm() */
+    EIGEN_DEVICE_FUNC
+    const NormReturnType norm() const
+    { return NormReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * Example: \include PartialRedux_norm.cpp
+      * Output: \verbinclude PartialRedux_norm.out
+      *
+      * \sa DenseBase::norm() */
+    template<int p>
+    EIGEN_DEVICE_FUNC
+    const typename LpNormReturnType<p>::Type lpNorm() const
+    { return typename LpNormReturnType<p>::Type(_expression()); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, using
+      * Blue's algorithm.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * \sa DenseBase::blueNorm() */
+    EIGEN_DEVICE_FUNC
+    const BlueNormReturnType blueNorm() const
+    { return BlueNormReturnType(_expression()); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, avoiding
+      * underflow and overflow.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * \sa DenseBase::stableNorm() */
+    EIGEN_DEVICE_FUNC
+    const StableNormReturnType stableNorm() const
+    { return StableNormReturnType(_expression()); }
+
+
+    /** \returns a row (or column) vector expression of the norm
+      * of each column (or row) of the referenced expression, avoiding
+      * underflow and overflow using a concatenation of hypot() calls.
+      * This is a vector with real entries, even if the original matrix has complex entries.
+      *
+      * \sa DenseBase::hypotNorm() */
+    EIGEN_DEVICE_FUNC
+    const HypotNormReturnType hypotNorm() const
+    { return HypotNormReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the sum
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_sum.cpp
+      * Output: \verbinclude PartialRedux_sum.out
+      *
+      * \sa DenseBase::sum() */
+    EIGEN_DEVICE_FUNC
+    const SumReturnType sum() const
+    { return SumReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the mean
+    * of each column (or row) of the referenced expression.
+    *
+    * \sa DenseBase::mean() */
+    EIGEN_DEVICE_FUNC
+    const MeanReturnType mean() const
+    { return MeanReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression representing
+      * whether \b all coefficients of each respective column (or row) are \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
+      *
+      * \sa DenseBase::all() */
+    EIGEN_DEVICE_FUNC
+    const AllReturnType all() const
+    { return AllReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression representing
+      * whether \b at \b least one coefficient of each respective column (or row) is \c true.
+      * This expression can be assigned to a vector with entries of type \c bool.
+      *
+      * \sa DenseBase::any() */
+    EIGEN_DEVICE_FUNC
+    const AnyReturnType any() const
+    { return AnyReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression representing
+      * the number of \c true coefficients of each respective column (or row).
+      * This expression can be assigned to a vector whose entries have the same type as is used to
+      * index entries of the original matrix; for dense matrices, this is \c std::ptrdiff_t .
+      *
+      * Example: \include PartialRedux_count.cpp
+      * Output: \verbinclude PartialRedux_count.out
+      *
+      * \sa DenseBase::count() */
+    EIGEN_DEVICE_FUNC
+    const CountReturnType count() const
+    { return CountReturnType(_expression()); }
+
+    /** \returns a row (or column) vector expression of the product
+      * of each column (or row) of the referenced expression.
+      *
+      * Example: \include PartialRedux_prod.cpp
+      * Output: \verbinclude PartialRedux_prod.out
+      *
+      * \sa DenseBase::prod() */
+    EIGEN_DEVICE_FUNC
+    const ProdReturnType prod() const
+    { return ProdReturnType(_expression()); }
+
+
+    /** \returns a matrix expression
+      * where each column (or row) are reversed.
+      *
+      * Example: \include Vectorwise_reverse.cpp
+      * Output: \verbinclude Vectorwise_reverse.out
+      *
+      * \sa DenseBase::reverse() */
+    EIGEN_DEVICE_FUNC
+    const ConstReverseReturnType reverse() const
+    { return ConstReverseReturnType( _expression() ); }
+
+    /** \returns a writable matrix expression
+      * where each column (or row) are reversed.
+      *
+      * \sa reverse() const */
+    EIGEN_DEVICE_FUNC
+    ReverseReturnType reverse()
+    { return ReverseReturnType( _expression() ); }
+
+    typedef Replicate<ExpressionType,(isVertical?Dynamic:1),(isHorizontal?Dynamic:1)> ReplicateReturnType;
+    EIGEN_DEVICE_FUNC
+    const ReplicateReturnType replicate(Index factor) const;
+
+    /**
+      * \return an expression of the replication of each column (or row) of \c *this
+      *
+      * Example: \include DirectionWise_replicate.cpp
+      * Output: \verbinclude DirectionWise_replicate.out
+      *
+      * \sa VectorwiseOp::replicate(Index), DenseBase::replicate(), class Replicate
+      */
+    // NOTE implemented here because of sunstudio's compilation errors
+    // isVertical*Factor+isHorizontal instead of (isVertical?Factor:1) to handle CUDA bug with ternary operator
+    template<int Factor> const Replicate<ExpressionType,isVertical*Factor+isHorizontal,isHorizontal*Factor+isVertical>
+    EIGEN_DEVICE_FUNC
+    replicate(Index factor = Factor) const
+    {
+      return Replicate<ExpressionType,(isVertical?Factor:1),(isHorizontal?Factor:1)>
+          (_expression(),isVertical?factor:1,isHorizontal?factor:1);
+    }
+
+/////////// Artithmetic operators ///////////
+
+    /** Copies the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    ExpressionType& operator=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      //eigen_assert((m_matrix.isNull()) == (other.isNull())); FIXME
+      return const_cast<ExpressionType&>(m_matrix = extendedTo(other.derived()));
+    }
+
+    /** Adds the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    ExpressionType& operator+=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix += extendedTo(other.derived()));
+    }
+
+    /** Substracts the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    ExpressionType& operator-=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return const_cast<ExpressionType&>(m_matrix -= extendedTo(other.derived()));
+    }
+
+    /** Multiples each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    ExpressionType& operator*=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix *= extendedTo(other.derived());
+      return const_cast<ExpressionType&>(m_matrix);
+    }
+
+    /** Divides each subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    ExpressionType& operator/=(const DenseBase<OtherDerived>& other)
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      m_matrix /= extendedTo(other.derived());
+      return const_cast<ExpressionType&>(m_matrix);
+    }
+
+    /** Returns the expression of the sum of the vector \a other to each subvector of \c *this */
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_sum_op<Scalar,typename OtherDerived::Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator+(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix + extendedTo(other.derived());
+    }
+
+    /** Returns the expression of the difference between each subvector of \c *this and the vector \a other */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_difference_op<Scalar,typename OtherDerived::Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator-(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix - extendedTo(other.derived());
+    }
+
+    /** Returns the expression where each subvector is the product of the vector \a other
+      * by the corresponding subvector of \c *this */
+    template<typename OtherDerived> EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_product_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    EIGEN_DEVICE_FUNC
+    operator*(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix * extendedTo(other.derived());
+    }
+
+    /** Returns the expression where each subvector is the quotient of the corresponding
+      * subvector of \c *this by the vector \a other */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename ExtendedType<OtherDerived>::Type>
+    operator/(const DenseBase<OtherDerived>& other) const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+      EIGEN_STATIC_ASSERT_ARRAYXPR(ExpressionType)
+      EIGEN_STATIC_ASSERT_SAME_XPR_KIND(ExpressionType, OtherDerived)
+      return m_matrix / extendedTo(other.derived());
+    }
+
+    /** \returns an expression where each column (or row) of the referenced matrix are normalized.
+      * The referenced matrix is \b not modified.
+      * \sa MatrixBase::normalized(), normalize()
+      */
+    EIGEN_DEVICE_FUNC
+    CwiseBinaryOp<internal::scalar_quotient_op<Scalar>,
+                  const ExpressionTypeNestedCleaned,
+                  const typename OppositeExtendedType<typename ReturnType<internal::member_norm,RealScalar>::Type>::Type>
+    normalized() const { return m_matrix.cwiseQuotient(extendedToOpposite(this->norm())); }
+
+
+    /** Normalize in-place each row or columns of the referenced matrix.
+      * \sa MatrixBase::normalize(), normalized()
+      */
+    EIGEN_DEVICE_FUNC void normalize() {
+      m_matrix = this->normalized();
+    }
+
+    EIGEN_DEVICE_FUNC inline void reverseInPlace();
+
+/////////// Geometry module ///////////
+
+    typedef Homogeneous<ExpressionType,Direction> HomogeneousReturnType;
+    EIGEN_DEVICE_FUNC
+    HomogeneousReturnType homogeneous() const;
+
+    typedef typename ExpressionType::PlainObject CrossReturnType;
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC
+    const CrossReturnType cross(const MatrixBase<OtherDerived>& other) const;
+
+    enum {
+      HNormalized_Size = Direction==Vertical ? internal::traits<ExpressionType>::RowsAtCompileTime
+                                             : internal::traits<ExpressionType>::ColsAtCompileTime,
+      HNormalized_SizeMinusOne = HNormalized_Size==Dynamic ? Dynamic : HNormalized_Size-1
+    };
+    typedef Block<const ExpressionType,
+                  Direction==Vertical   ? int(HNormalized_SizeMinusOne)
+                                        : int(internal::traits<ExpressionType>::RowsAtCompileTime),
+                  Direction==Horizontal ? int(HNormalized_SizeMinusOne)
+                                        : int(internal::traits<ExpressionType>::ColsAtCompileTime)>
+            HNormalized_Block;
+    typedef Block<const ExpressionType,
+                  Direction==Vertical   ? 1 : int(internal::traits<ExpressionType>::RowsAtCompileTime),
+                  Direction==Horizontal ? 1 : int(internal::traits<ExpressionType>::ColsAtCompileTime)>
+            HNormalized_Factors;
+    typedef CwiseBinaryOp<internal::scalar_quotient_op<typename internal::traits<ExpressionType>::Scalar>,
+                const HNormalized_Block,
+                const Replicate<HNormalized_Factors,
+                  Direction==Vertical   ? HNormalized_SizeMinusOne : 1,
+                  Direction==Horizontal ? HNormalized_SizeMinusOne : 1> >
+            HNormalizedReturnType;
+
+    EIGEN_DEVICE_FUNC
+    const HNormalizedReturnType hnormalized() const;
+
+  protected:
+    ExpressionTypeNested m_matrix;
+};
+
+//const colwise moved to DenseBase.h due to CUDA compiler bug
+
+
+/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * \sa rowwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::ColwiseReturnType
+DenseBase<Derived>::colwise()
+{
+  return ColwiseReturnType(derived());
+}
+
+//const rowwise moved to DenseBase.h due to CUDA compiler bug
+
+
+/** \returns a writable VectorwiseOp wrapper of *this providing additional partial reduction operations
+  *
+  * \sa colwise(), class VectorwiseOp, \ref TutorialReductionsVisitorsBroadcasting
+  */
+template<typename Derived>
+inline typename DenseBase<Derived>::RowwiseReturnType
+DenseBase<Derived>::rowwise()
+{
+  return RowwiseReturnType(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARTIAL_REDUX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Visitor.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Visitor.h
new file mode 100644
index 0000000..54c1883
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/Visitor.h
@@ -0,0 +1,273 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_VISITOR_H
+#define EIGEN_VISITOR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Visitor, typename Derived, int UnrollCount>
+struct visitor_impl
+{
+  enum {
+    col = (UnrollCount-1) / Derived::RowsAtCompileTime,
+    row = (UnrollCount-1) % Derived::RowsAtCompileTime
+  };
+
+  EIGEN_DEVICE_FUNC
+  static inline void run(const Derived &mat, Visitor& visitor)
+  {
+    visitor_impl<Visitor, Derived, UnrollCount-1>::run(mat, visitor);
+    visitor(mat.coeff(row, col), row, col);
+  }
+};
+
+template<typename Visitor, typename Derived>
+struct visitor_impl<Visitor, Derived, 1>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const Derived &mat, Visitor& visitor)
+  {
+    return visitor.init(mat.coeff(0, 0), 0, 0);
+  }
+};
+
+template<typename Visitor, typename Derived>
+struct visitor_impl<Visitor, Derived, Dynamic>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const Derived& mat, Visitor& visitor)
+  {
+    visitor.init(mat.coeff(0,0), 0, 0);
+    for(Index i = 1; i < mat.rows(); ++i)
+      visitor(mat.coeff(i, 0), i, 0);
+    for(Index j = 1; j < mat.cols(); ++j)
+      for(Index i = 0; i < mat.rows(); ++i)
+        visitor(mat.coeff(i, j), i, j);
+  }
+};
+
+// evaluator adaptor
+template<typename XprType>
+class visitor_evaluator
+{
+public:
+  EIGEN_DEVICE_FUNC
+  explicit visitor_evaluator(const XprType &xpr) : m_evaluator(xpr), m_xpr(xpr) {}
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::CoeffReturnType CoeffReturnType;
+  
+  enum {
+    RowsAtCompileTime = XprType::RowsAtCompileTime,
+    CoeffReadCost = internal::evaluator<XprType>::CoeffReadCost
+  };
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_xpr.rows(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_xpr.cols(); }
+  EIGEN_DEVICE_FUNC Index size() const { return m_xpr.size(); }
+
+  EIGEN_DEVICE_FUNC CoeffReturnType coeff(Index row, Index col) const
+  { return m_evaluator.coeff(row, col); }
+  
+protected:
+  internal::evaluator<XprType> m_evaluator;
+  const XprType &m_xpr;
+};
+} // end namespace internal
+
+/** Applies the visitor \a visitor to the whole coefficients of the matrix or vector.
+  *
+  * The template parameter \a Visitor is the type of the visitor and provides the following interface:
+  * \code
+  * struct MyVisitor {
+  *   // called for the first coefficient
+  *   void init(const Scalar& value, Index i, Index j);
+  *   // called for all other coefficients
+  *   void operator() (const Scalar& value, Index i, Index j);
+  * };
+  * \endcode
+  *
+  * \note compared to one or two \em for \em loops, visitors offer automatic
+  * unrolling for small fixed size matrix.
+  *
+  * \sa minCoeff(Index*,Index*), maxCoeff(Index*,Index*), DenseBase::redux()
+  */
+template<typename Derived>
+template<typename Visitor>
+EIGEN_DEVICE_FUNC
+void DenseBase<Derived>::visit(Visitor& visitor) const
+{
+  typedef typename internal::visitor_evaluator<Derived> ThisEvaluator;
+  ThisEvaluator thisEval(derived());
+  
+  enum {
+    unroll =  SizeAtCompileTime != Dynamic
+           && SizeAtCompileTime * ThisEvaluator::CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost <= EIGEN_UNROLLING_LIMIT
+  };
+  return internal::visitor_impl<Visitor, ThisEvaluator, unroll ? int(SizeAtCompileTime) : Dynamic>::run(thisEval, visitor);
+}
+
+namespace internal {
+
+/** \internal
+  * \brief Base class to implement min and max visitors
+  */
+template <typename Derived>
+struct coeff_visitor
+{
+  typedef typename Derived::Scalar Scalar;
+  Index row, col;
+  Scalar res;
+  EIGEN_DEVICE_FUNC
+  inline void init(const Scalar& value, Index i, Index j)
+  {
+    res = value;
+    row = i;
+    col = j;
+  }
+};
+
+/** \internal
+  * \brief Visitor computing the min coefficient with its value and coordinates
+  *
+  * \sa DenseBase::minCoeff(Index*, Index*)
+  */
+template <typename Derived>
+struct min_coeff_visitor : coeff_visitor<Derived>
+{
+  typedef typename Derived::Scalar Scalar;
+  EIGEN_DEVICE_FUNC
+  void operator() (const Scalar& value, Index i, Index j)
+  {
+    if(value < this->res)
+    {
+      this->res = value;
+      this->row = i;
+      this->col = j;
+    }
+  }
+};
+
+template<typename Scalar>
+struct functor_traits<min_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
+/** \internal
+  * \brief Visitor computing the max coefficient with its value and coordinates
+  *
+  * \sa DenseBase::maxCoeff(Index*, Index*)
+  */
+template <typename Derived>
+struct max_coeff_visitor : coeff_visitor<Derived>
+{
+  typedef typename Derived::Scalar Scalar; 
+  EIGEN_DEVICE_FUNC
+  void operator() (const Scalar& value, Index i, Index j)
+  {
+    if(value > this->res)
+    {
+      this->res = value;
+      this->row = i;
+      this->col = j;
+    }
+  }
+};
+
+template<typename Scalar>
+struct functor_traits<max_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
+} // end namespace internal
+
+/** \fn DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the minimum of all coefficients of *this and puts in *row and *col its location.
+  * \warning the result is undefined if \c *this contains NaN.
+  *
+  * \sa DenseBase::minCoeff(Index*), DenseBase::maxCoeff(Index*,Index*), DenseBase::visit(), DenseBase::minCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+EIGEN_DEVICE_FUNC
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::minCoeff(IndexType* rowId, IndexType* colId) const
+{
+  internal::min_coeff_visitor<Derived> minVisitor;
+  this->visit(minVisitor);
+  *rowId = minVisitor.row;
+  if (colId) *colId = minVisitor.col;
+  return minVisitor.res;
+}
+
+/** \returns the minimum of all coefficients of *this and puts in *index its location.
+  * \warning the result is undefined if \c *this contains NaN. 
+  *
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::minCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+EIGEN_DEVICE_FUNC
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::minCoeff(IndexType* index) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  internal::min_coeff_visitor<Derived> minVisitor;
+  this->visit(minVisitor);
+  *index = IndexType((RowsAtCompileTime==1) ? minVisitor.col : minVisitor.row);
+  return minVisitor.res;
+}
+
+/** \fn DenseBase<Derived>::maxCoeff(IndexType* rowId, IndexType* colId) const
+  * \returns the maximum of all coefficients of *this and puts in *row and *col its location.
+  * \warning the result is undefined if \c *this contains NaN. 
+  *
+  * \sa DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visit(), DenseBase::maxCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+EIGEN_DEVICE_FUNC
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::maxCoeff(IndexType* rowPtr, IndexType* colPtr) const
+{
+  internal::max_coeff_visitor<Derived> maxVisitor;
+  this->visit(maxVisitor);
+  *rowPtr = maxVisitor.row;
+  if (colPtr) *colPtr = maxVisitor.col;
+  return maxVisitor.res;
+}
+
+/** \returns the maximum of all coefficients of *this and puts in *index its location.
+  * \warning the result is undefined if \c *this contains NaN.
+  *
+  * \sa DenseBase::maxCoeff(IndexType*,IndexType*), DenseBase::minCoeff(IndexType*,IndexType*), DenseBase::visitor(), DenseBase::maxCoeff()
+  */
+template<typename Derived>
+template<typename IndexType>
+EIGEN_DEVICE_FUNC
+typename internal::traits<Derived>::Scalar
+DenseBase<Derived>::maxCoeff(IndexType* index) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  internal::max_coeff_visitor<Derived> maxVisitor;
+  this->visit(maxVisitor);
+  *index = (RowsAtCompileTime==1) ? maxVisitor.col : maxVisitor.row;
+  return maxVisitor.res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_VISITOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h
new file mode 100644
index 0000000..99439c8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/Complex.h
@@ -0,0 +1,483 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_AVX_H
+#define EIGEN_COMPLEX_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- float ----------
+struct Packet4cf
+{
+  EIGEN_STRONG_INLINE Packet4cf() {}
+  EIGEN_STRONG_INLINE explicit Packet4cf(const __m256& a) : v(a) {}
+  __m256  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet4cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet4cf> { typedef std::complex<float> type; enum {size=4, alignment=Aligned32}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet4cf padd<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_add_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf psub<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_sub_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pnegate(const Packet4cf& a)
+{
+  return Packet4cf(pnegate(a.v));
+}
+template<> EIGEN_STRONG_INLINE Packet4cf pconj(const Packet4cf& a)
+{
+  const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000,0x00000000,0x80000000));
+  return Packet4cf(_mm256_xor_ps(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pmul<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  __m256 tmp1 = _mm256_mul_ps(_mm256_moveldup_ps(a.v), b.v);
+  __m256 tmp2 = _mm256_mul_ps(_mm256_movehdup_ps(a.v), _mm256_permute_ps(b.v, _MM_SHUFFLE(2,3,0,1)));
+  __m256 result = _mm256_addsub_ps(tmp1, tmp2);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pand   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_and_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf por    <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_or_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pxor   <Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_xor_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet4cf pandnot<Packet4cf>(const Packet4cf& a, const Packet4cf& b) { return Packet4cf(_mm256_andnot_ps(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet4cf pload <Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet4cf(pload<Packet8f>(&numext::real_ref(*from))); }
+template<> EIGEN_STRONG_INLINE Packet4cf ploadu<Packet4cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet4cf(ploadu<Packet8f>(&numext::real_ref(*from))); }
+
+
+template<> EIGEN_STRONG_INLINE Packet4cf pset1<Packet4cf>(const std::complex<float>& from)
+{
+  return Packet4cf(_mm256_castpd_ps(_mm256_broadcast_sd((const double*)(const void*)&from)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf ploaddup<Packet4cf>(const std::complex<float>* from)
+{
+  // FIXME The following might be optimized using _mm256_movedup_pd
+  Packet2cf a = ploaddup<Packet2cf>(from);
+  Packet2cf b = ploaddup<Packet2cf>(from+1);
+  return  Packet4cf(_mm256_insertf128_ps(_mm256_castps128_ps256(a.v), b.v, 1));
+}
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float>* to, const Packet4cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4cf pgather<std::complex<float>, Packet4cf>(const std::complex<float>* from, Index stride)
+{
+  return Packet4cf(_mm256_set_ps(std::imag(from[3*stride]), std::real(from[3*stride]),
+                                 std::imag(from[2*stride]), std::real(from[2*stride]),
+                                 std::imag(from[1*stride]), std::real(from[1*stride]),
+                                 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet4cf>(std::complex<float>* to, const Packet4cf& from, Index stride)
+{
+  __m128 low = _mm256_extractf128_ps(from.v, 0);
+  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1)));
+  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(low, low, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3)));
+
+  __m128 high = _mm256_extractf128_ps(from.v, 1);
+  to[stride*2] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1)));
+  to[stride*3] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(high, high, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3)));
+
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet4cf>(const Packet4cf& a)
+{
+  return pfirst(Packet2cf(_mm256_castps256_ps128(a.v)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preverse(const Packet4cf& a) {
+  __m128 low  = _mm256_extractf128_ps(a.v, 0);
+  __m128 high = _mm256_extractf128_ps(a.v, 1);
+  __m128d lowd  = _mm_castps_pd(low);
+  __m128d highd = _mm_castps_pd(high);
+  low  = _mm_castpd_ps(_mm_shuffle_pd(lowd,lowd,0x1));
+  high = _mm_castpd_ps(_mm_shuffle_pd(highd,highd,0x1));
+  __m256 result = _mm256_setzero_ps();
+  result = _mm256_insertf128_ps(result, low, 1);
+  result = _mm256_insertf128_ps(result, high, 0);
+  return Packet4cf(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet4cf>(const Packet4cf& a)
+{
+  return predux(padd(Packet2cf(_mm256_extractf128_ps(a.v,0)),
+                     Packet2cf(_mm256_extractf128_ps(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf preduxp<Packet4cf>(const Packet4cf* vecs)
+{
+  Packet8f t0 = _mm256_shuffle_ps(vecs[0].v, vecs[0].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t1 = _mm256_shuffle_ps(vecs[1].v, vecs[1].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t0 = _mm256_hadd_ps(t0,t1);
+  Packet8f t2 = _mm256_shuffle_ps(vecs[2].v, vecs[2].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  Packet8f t3 = _mm256_shuffle_ps(vecs[3].v, vecs[3].v, _MM_SHUFFLE(3, 1, 2 ,0));
+  t2 = _mm256_hadd_ps(t2,t3);
+  
+  t1 = _mm256_permute2f128_ps(t0,t2, 0 + (2<<4));
+  t3 = _mm256_permute2f128_ps(t0,t2, 1 + (3<<4));
+
+  return Packet4cf(_mm256_add_ps(t1,t3));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet4cf>(const Packet4cf& a)
+{
+  return predux_mul(pmul(Packet2cf(_mm256_extractf128_ps(a.v, 0)),
+                         Packet2cf(_mm256_extractf128_ps(a.v, 1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4cf& first, const Packet4cf& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet8f>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet4cf, Packet4cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& a, const Packet4cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet8f, Packet4cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet8f& x, const Packet4cf& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet8f& x, const Packet4cf& y) const
+  { return Packet4cf(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet4cf, Packet8f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet4cf pmadd(const Packet4cf& x, const Packet8f& y, const Packet4cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet4cf pmul(const Packet4cf& x, const Packet8f& y) const
+  { return Packet4cf(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4cf pdiv<Packet4cf>(const Packet4cf& a, const Packet4cf& b)
+{
+  Packet4cf num = pmul(a, pconj(b));
+  __m256 tmp = _mm256_mul_ps(b.v, b.v);
+  __m256 tmp2    = _mm256_shuffle_ps(tmp,tmp,0xB1);
+  __m256 denom = _mm256_add_ps(tmp, tmp2);
+  return Packet4cf(_mm256_div_ps(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pcplxflip<Packet4cf>(const Packet4cf& x)
+{
+  return Packet4cf(_mm256_shuffle_ps(x.v, x.v, _MM_SHUFFLE(2, 3, 0 ,1)));
+}
+
+//---------- double ----------
+struct Packet2cd
+{
+  EIGEN_STRONG_INLINE Packet2cd() {}
+  EIGEN_STRONG_INLINE explicit Packet2cd(const __m256d& a) : v(a) {}
+  __m256d  v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet2cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 2,
+    HasHalfPacket = 1,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cd> { typedef std::complex<double> type; enum {size=2, alignment=Aligned32}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cd padd<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_add_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd psub<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_sub_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pnegate(const Packet2cd& a) { return Packet2cd(pnegate(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pconj(const Packet2cd& a)
+{
+  const __m256d mask = _mm256_castsi256_pd(_mm256_set_epi32(0x80000000,0x0,0x0,0x0,0x80000000,0x0,0x0,0x0));
+  return Packet2cd(_mm256_xor_pd(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pmul<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  __m256d tmp1 = _mm256_shuffle_pd(a.v,a.v,0x0);
+  __m256d even = _mm256_mul_pd(tmp1, b.v);
+  __m256d tmp2 = _mm256_shuffle_pd(a.v,a.v,0xF);
+  __m256d tmp3 = _mm256_shuffle_pd(b.v,b.v,0x5);
+  __m256d odd  = _mm256_mul_pd(tmp2, tmp3);
+  return Packet2cd(_mm256_addsub_pd(even, odd));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pand   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_and_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd por    <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_or_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pxor   <Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_xor_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cd pandnot<Packet2cd>(const Packet2cd& a, const Packet2cd& b) { return Packet2cd(_mm256_andnot_pd(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pload <Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_ALIGNED_LOAD return Packet2cd(pload<Packet4d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cd ploadu<Packet2cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cd(ploadu<Packet4d>((const double*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cd pset1<Packet2cd>(const std::complex<double>& from)
+{
+  // in case casting to a __m128d* is really not safe, then we can still fallback to this version: (much slower though)
+//   return Packet2cd(_mm256_loadu2_m128d((const double*)&from,(const double*)&from));
+    return Packet2cd(_mm256_broadcast_pd((const __m128d*)(const void*)&from));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd ploaddup<Packet2cd>(const std::complex<double>* from) { return pset1<Packet2cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet2cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cd pgather<std::complex<double>, Packet2cd>(const std::complex<double>* from, Index stride)
+{
+  return Packet2cd(_mm256_set_pd(std::imag(from[1*stride]), std::real(from[1*stride]),
+				 std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet2cd>(std::complex<double>* to, const Packet2cd& from, Index stride)
+{
+  __m128d low = _mm256_extractf128_pd(from.v, 0);
+  to[stride*0] = std::complex<double>(_mm_cvtsd_f64(low), _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1)));
+  __m128d high = _mm256_extractf128_pd(from.v, 1);
+  to[stride*1] = std::complex<double>(_mm_cvtsd_f64(high), _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1)));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> pfirst<Packet2cd>(const Packet2cd& a)
+{
+  __m128d low = _mm256_extractf128_pd(a.v, 0);
+  EIGEN_ALIGN16 double res[2];
+  _mm_store_pd(res, low);
+  return std::complex<double>(res[0],res[1]);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preverse(const Packet2cd& a) {
+  __m256d result = _mm256_permute2f128_pd(a.v, a.v, 1);
+  return Packet2cd(result);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet2cd>(const Packet2cd& a)
+{
+  return predux(padd(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd preduxp<Packet2cd>(const Packet2cd* vecs)
+{
+  Packet4d t0 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 0 + (2<<4));
+  Packet4d t1 = _mm256_permute2f128_pd(vecs[0].v,vecs[1].v, 1 + (3<<4));
+
+  return Packet2cd(_mm256_add_pd(t0,t1));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet2cd>(const Packet2cd& a)
+{
+  return predux(pmul(Packet1cd(_mm256_extractf128_pd(a.v,0)),
+                     Packet1cd(_mm256_extractf128_pd(a.v,1))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cd& first, const Packet2cd& second)
+  {
+    if (Offset==0) return;
+    palign_impl<Offset*2,Packet4d>::run(first.v, second.v);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cd, Packet2cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& a, const Packet2cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet4d, Packet2cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet4d& x, const Packet2cd& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet4d& x, const Packet2cd& y) const
+  { return Packet2cd(Eigen::internal::pmul(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cd, Packet4d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cd pmadd(const Packet2cd& x, const Packet4d& y, const Packet2cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cd pmul(const Packet2cd& x, const Packet4d& y) const
+  { return Packet2cd(Eigen::internal::pmul(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cd pdiv<Packet2cd>(const Packet2cd& a, const Packet2cd& b)
+{
+  Packet2cd num = pmul(a, pconj(b));
+  __m256d tmp = _mm256_mul_pd(b.v, b.v);
+  __m256d denom = _mm256_hadd_pd(tmp, tmp);
+  return Packet2cd(_mm256_div_pd(num.v, denom));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pcplxflip<Packet2cd>(const Packet2cd& x)
+{
+  return Packet2cd(_mm256_shuffle_pd(x.v, x.v, 0x5));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4cf,4>& kernel) {
+  __m256d P0 = _mm256_castps_pd(kernel.packet[0].v);
+  __m256d P1 = _mm256_castps_pd(kernel.packet[1].v);
+  __m256d P2 = _mm256_castps_pd(kernel.packet[2].v);
+  __m256d P3 = _mm256_castps_pd(kernel.packet[3].v);
+
+  __m256d T0 = _mm256_shuffle_pd(P0, P1, 15);
+  __m256d T1 = _mm256_shuffle_pd(P0, P1, 0);
+  __m256d T2 = _mm256_shuffle_pd(P2, P3, 15);
+  __m256d T3 = _mm256_shuffle_pd(P2, P3, 0);
+
+  kernel.packet[1].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 32));
+  kernel.packet[3].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T0, T2, 49));
+  kernel.packet[0].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 32));
+  kernel.packet[2].v = _mm256_castpd_ps(_mm256_permute2f128_pd(T1, T3, 49));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cd,2>& kernel) {
+  __m256d tmp = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 0+(2<<4));
+  kernel.packet[1].v = _mm256_permute2f128_pd(kernel.packet[0].v, kernel.packet[1].v, 1+(3<<4));
+ kernel.packet[0].v = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertfirst(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertfirst(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,1|2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4cf pinsertlast(const Packet4cf& a, std::complex<float> b)
+{
+  return Packet4cf(_mm256_blend_ps(a.v,pset1<Packet4cf>(b).v,(1<<7)|(1<<6)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cd pinsertlast(const Packet2cd& a, std::complex<double> b)
+{
+  return Packet2cd(_mm256_blend_pd(a.v,pset1<Packet2cd>(b).v,(1<<3)|(1<<2)));
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_AVX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h
new file mode 100644
index 0000000..6af67ce
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/MathFunctions.h
@@ -0,0 +1,439 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Pedro Gonnet (pedro.gonnet@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATH_FUNCTIONS_AVX_H
+#define EIGEN_MATH_FUNCTIONS_AVX_H
+
+/* The sin, cos, exp, and log functions of this file are loosely derived from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+namespace Eigen {
+
+namespace internal {
+
+inline Packet8i pshiftleft(Packet8i v, int n)
+{
+#ifdef EIGEN_VECTORIZE_AVX2
+  return _mm256_slli_epi32(v, n);
+#else
+  __m128i lo = _mm_slli_epi32(_mm256_extractf128_si256(v, 0), n);
+  __m128i hi = _mm_slli_epi32(_mm256_extractf128_si256(v, 1), n);
+  return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1);
+#endif
+}
+
+inline Packet8f pshiftright(Packet8f v, int n)
+{
+#ifdef EIGEN_VECTORIZE_AVX2
+  return _mm256_cvtepi32_ps(_mm256_srli_epi32(_mm256_castps_si256(v), n));
+#else
+  __m128i lo = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 0), n);
+  __m128i hi = _mm_srli_epi32(_mm256_extractf128_si256(_mm256_castps_si256(v), 1), n);
+  return _mm256_cvtepi32_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(lo), (hi), 1));
+#endif
+}
+
+// Sine function
+// Computes sin(x) by wrapping x to the interval [-Pi/4,3*Pi/4] and
+// evaluating interpolants in [-Pi/4,Pi/4] or [Pi/4,3*Pi/4]. The interpolants
+// are (anti-)symmetric and thus have only odd/even coefficients
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psin<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+
+  // Some useful values.
+  _EIGEN_DECLARE_CONST_Packet8i(one, 1);
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(two, 2.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(one_over_four, 0.25f);
+  _EIGEN_DECLARE_CONST_Packet8f(one_over_pi, 3.183098861837907e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_first, -3.140625000000000e+00f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_second, -9.670257568359375e-04f);
+  _EIGEN_DECLARE_CONST_Packet8f(neg_pi_third, -6.278329571784980e-07f);
+  _EIGEN_DECLARE_CONST_Packet8f(four_over_pi, 1.273239544735163e+00f);
+
+  // Map x from [-Pi/4,3*Pi/4] to z in [-1,3] and subtract the shifted period.
+  Packet8f z = pmul(x, p8f_one_over_pi);
+  Packet8f shift = _mm256_floor_ps(padd(z, p8f_one_over_four));
+  x = pmadd(shift, p8f_neg_pi_first, x);
+  x = pmadd(shift, p8f_neg_pi_second, x);
+  x = pmadd(shift, p8f_neg_pi_third, x);
+  z = pmul(x, p8f_four_over_pi);
+
+  // Make a mask for the entries that need flipping, i.e. wherever the shift
+  // is odd.
+  Packet8i shift_ints = _mm256_cvtps_epi32(shift);
+  Packet8i shift_isodd = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(shift_ints), _mm256_castsi256_ps(p8i_one)));
+  Packet8i sign_flip_mask = pshiftleft(shift_isodd, 31);
+
+  // Create a mask for which interpolant to use, i.e. if z > 1, then the mask
+  // is set to ones for that entry.
+  Packet8f ival_mask = _mm256_cmp_ps(z, p8f_one, _CMP_GT_OQ);
+
+  // Evaluate the polynomial for the interval [1,3] in z.
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_0, 9.999999724233232e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_2, -3.084242535619928e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_4, 1.584991525700324e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_right_6, -3.188805084631342e-04f);
+  Packet8f z_minus_two = psub(z, p8f_two);
+  Packet8f z_minus_two2 = pmul(z_minus_two, z_minus_two);
+  Packet8f right = pmadd(p8f_coeff_right_6, z_minus_two2, p8f_coeff_right_4);
+  right = pmadd(right, z_minus_two2, p8f_coeff_right_2);
+  right = pmadd(right, z_minus_two2, p8f_coeff_right_0);
+
+  // Evaluate the polynomial for the interval [-1,1] in z.
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_1, 7.853981525427295e-01f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_3, -8.074536727092352e-02f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_5, 2.489871967827018e-03f);
+  _EIGEN_DECLARE_CONST_Packet8f(coeff_left_7, -3.587725841214251e-05f);
+  Packet8f z2 = pmul(z, z);
+  Packet8f left = pmadd(p8f_coeff_left_7, z2, p8f_coeff_left_5);
+  left = pmadd(left, z2, p8f_coeff_left_3);
+  left = pmadd(left, z2, p8f_coeff_left_1);
+  left = pmul(left, z);
+
+  // Assemble the results, i.e. select the left and right polynomials.
+  left = _mm256_andnot_ps(ival_mask, left);
+  right = _mm256_and_ps(ival_mask, right);
+  Packet8f res = _mm256_or_ps(left, right);
+
+  // Flip the sign on the odd intervals and return the result.
+  res = _mm256_xor_ps(res, _mm256_castsi256_ps(sign_flip_mask));
+  return res;
+}
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+// TODO(gonnet): Further reduce the interval allowing for lower-degree
+//               polynomial interpolants -> ... -> profit!
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+plog<Packet8f>(const Packet8f& _x) {
+  Packet8f x = _x;
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(minus_inf, 0xff800000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_log_q2, 0.693359375f);
+
+  Packet8f invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_NGE_UQ); // not greater equal is true if x is NaN
+  Packet8f iszero_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_EQ_OQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p8f_min_norm_pos);
+
+  Packet8f emm0 = pshiftright(x,23);
+  Packet8f e = _mm256_sub_ps(emm0, p8f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm256_and_ps(x, p8f_inv_mant_mask);
+  x = _mm256_or_ps(x, p8f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  Packet8f mask = _mm256_cmp_ps(x, p8f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet8f tmp = _mm256_and_ps(x, mask);
+  x = psub(x, p8f_1);
+  e = psub(e, _mm256_and_ps(p8f_1, mask));
+  x = padd(x, tmp);
+
+  Packet8f x2 = pmul(x, x);
+  Packet8f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet8f y, y1, y2;
+  y = pmadd(p8f_cephes_log_p0, x, p8f_cephes_log_p1);
+  y1 = pmadd(p8f_cephes_log_p3, x, p8f_cephes_log_p4);
+  y2 = pmadd(p8f_cephes_log_p6, x, p8f_cephes_log_p7);
+  y = pmadd(y, x, p8f_cephes_log_p2);
+  y1 = pmadd(y1, x, p8f_cephes_log_p5);
+  y2 = pmadd(y2, x, p8f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p8f_cephes_log_q1);
+  tmp = pmul(x2, p8f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p8f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm256_or_ps(
+      _mm256_andnot_ps(iszero_mask, _mm256_or_ps(x, invalid_mask)),
+      _mm256_and_ps(iszero_mask, p8f_minus_inf));
+}
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+pexp<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet8f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet8f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet8f x = pmax(pmin(_x, p8f_exp_hi), p8f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet8f m = _mm256_floor_ps(pmadd(x, p8f_cephes_LOG2EF, p8f_half));
+
+// Get r = x - m*ln(2). If no FMA instructions are available, m*ln(2) is
+// subtracted out in two parts, m*C1+m*C2 = m*ln(2), to avoid accumulating
+// truncation errors. Note that we don't use the "pmadd" function here to
+// ensure that a precision-preserving FMA instruction is used.
+#ifdef EIGEN_VECTORIZE_FMA
+  _EIGEN_DECLARE_CONST_Packet8f(nln2, -0.6931471805599453f);
+  Packet8f r = _mm256_fmadd_ps(m, p8f_nln2, x);
+#else
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet8f(cephes_exp_C2, -2.12194440e-4f);
+  Packet8f r = psub(x, pmul(m, p8f_cephes_exp_C1));
+  r = psub(r, pmul(m, p8f_cephes_exp_C2));
+#endif
+
+  Packet8f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet8f y = p8f_cephes_exp_p0;
+  y = pmadd(y, r, p8f_cephes_exp_p1);
+  y = pmadd(y, r, p8f_cephes_exp_p2);
+  y = pmadd(y, r, p8f_cephes_exp_p3);
+  y = pmadd(y, r, p8f_cephes_exp_p4);
+  y = pmadd(y, r, p8f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p8f_1);
+
+  // Build emm0 = 2^m.
+  Packet8i emm0 = _mm256_cvttps_epi32(padd(m, p8f_127));
+  emm0 = pshiftleft(emm0, 23);
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm256_castsi256_ps(emm0)), _x);
+}
+
+// Hyperbolic Tangent function.
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+ptanh<Packet8f>(const Packet8f& x) {
+  return internal::generic_fast_tanh_float(x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4d
+pexp<Packet4d>(const Packet4d& _x) {
+  Packet4d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet4d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet4d(2, 2.0);
+  _EIGEN_DECLARE_CONST_Packet4d(half, 0.5);
+
+  _EIGEN_DECLARE_CONST_Packet4d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet4d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet4d(cephes_exp_C2, 1.42860682030941723212e-6);
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+
+  Packet4d tmp, fx;
+
+  // clamp x
+  x = pmax(pmin(x, p4d_exp_hi), p4d_exp_lo);
+  // Express exp(x) as exp(g + n*log(2)).
+  fx = pmadd(p4d_cephes_LOG2EF, x, p4d_half);
+
+  // Get the integer modulus of log(2), i.e. the "n" described above.
+  fx = _mm256_floor_pd(fx);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  tmp = pmul(fx, p4d_cephes_exp_C1);
+  Packet4d z = pmul(fx, p4d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet4d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet4d px = p4d_cephes_exp_p0;
+  px = pmadd(px, x2, p4d_cephes_exp_p1);
+  px = pmadd(px, x2, p4d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet4d qx = p4d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p4d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p4d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm256_div_pd(px, psub(qx, px));
+  x = pmadd(p4d_2, x, p4d_1);
+
+  // Build e=2^n by constructing the exponents in a 128-bit vector and
+  // shifting them to where they belong in double-precision values.
+  __m128i emm0 = _mm256_cvtpd_epi32(fx);
+  emm0 = _mm_add_epi32(emm0, p4i_1023);
+  emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(3, 1, 2, 0));
+  __m128i lo = _mm_slli_epi64(emm0, 52);
+  __m128i hi = _mm_slli_epi64(_mm_srli_epi64(emm0, 32), 52);
+  __m256i e = _mm256_insertf128_si256(_mm256_setzero_si256(), lo, 0);
+  e = _mm256_insertf128_si256(e, hi, 1);
+
+  // Construct the result 2^n * exp(g) = e * x. The max is used to catch
+  // non-finite values in the input.
+  return pmax(pmul(x, _mm256_castsi256_pd(e)), _x);
+}
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
+// For detail see here: http://www.beyond3d.com/content/articles/8/
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8f
+psqrt<Packet8f>(const Packet8f& _x) {
+  Packet8f half = pmul(_x, pset1<Packet8f>(.5f));
+  Packet8f denormal_mask = _mm256_and_ps(
+      _mm256_cmp_ps(_x, pset1<Packet8f>((std::numeric_limits<float>::min)()),
+                    _CMP_LT_OQ),
+      _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_GE_OQ));
+
+  // Compute approximate reciprocal sqrt.
+  Packet8f x = _mm256_rsqrt_ps(_x);
+  // Do a single step of Newton's iteration.
+  x = pmul(x, psub(pset1<Packet8f>(1.5f), pmul(half, pmul(x,x))));
+  // Flush results for denormals to zero.
+  return _mm256_andnot_ps(denormal_mask, pmul(_x,x));
+}
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f psqrt<Packet8f>(const Packet8f& x) {
+  return _mm256_sqrt_ps(x);
+}
+#endif
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d psqrt<Packet4d>(const Packet4d& x) {
+  return _mm256_sqrt_pd(x);
+}
+#if EIGEN_FAST_MATH
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f prsqrt<Packet8f>(const Packet8f& _x) {
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet8f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet8f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(flt_min, 0x00800000);
+
+  Packet8f neg_half = pmul(_x, p8f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet8f le_zero_mask = _mm256_cmp_ps(_x, p8f_flt_min, _CMP_LT_OQ);
+  Packet8f x = _mm256_andnot_ps(le_zero_mask, _mm256_rsqrt_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  Packet8f neg_mask = _mm256_cmp_ps(_x, _mm256_setzero_ps(), _CMP_LT_OQ);
+  Packet8f zero_mask = _mm256_andnot_ps(neg_mask, le_zero_mask);
+  Packet8f infs_and_nans = _mm256_or_ps(_mm256_and_ps(neg_mask, p8f_nan),
+                                        _mm256_and_ps(zero_mask, p8f_inf));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm256_or_ps(x, infs_and_nans);
+}
+
+#else
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet8f prsqrt<Packet8f>(const Packet8f& x) {
+  _EIGEN_DECLARE_CONST_Packet8f(one, 1.0f);
+  return _mm256_div_ps(p8f_one, _mm256_sqrt_ps(x));
+}
+#endif
+
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4d prsqrt<Packet4d>(const Packet4d& x) {
+  _EIGEN_DECLARE_CONST_Packet4d(one, 1.0);
+  return _mm256_div_pd(p4d_one, _mm256_sqrt_pd(x));
+}
+
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_AVX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h
new file mode 100644
index 0000000..195d40f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/PacketMath.h
@@ -0,0 +1,633 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_AVX_H
+#define EIGEN_PACKET_MATH_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
+#endif
+
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+#endif
+
+typedef __m256  Packet8f;
+typedef __m256i Packet8i;
+typedef __m256d Packet4d;
+
+template<> struct is_arithmetic<__m256>  { enum { value = true }; };
+template<> struct is_arithmetic<__m256i> { enum { value = true }; };
+template<> struct is_arithmetic<__m256d> { enum { value = true }; };
+
+#define _EIGEN_DECLARE_CONST_Packet8f(NAME,X) \
+  const Packet8f p8f_##NAME = pset1<Packet8f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4d(NAME,X) \
+  const Packet4d p4d_##NAME = pset1<Packet4d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8f_FROM_INT(NAME,X) \
+  const Packet8f p8f_##NAME = _mm256_castsi256_ps(pset1<Packet8i>(X))
+
+#define _EIGEN_DECLARE_CONST_Packet8i(NAME,X) \
+  const Packet8i p8i_##NAME = pset1<Packet8i>(X)
+
+// Use the packet_traits defined in AVX512/PacketMath.h instead if we're going
+// to leverage AVX512 instructions.
+#ifndef EIGEN_VECTORIZE_AVX512
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet8f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8,
+    HasHalfPacket = 1,
+
+    HasDiv  = 1,
+    HasSin  = EIGEN_FAST_MATH,
+    HasCos  = 0,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasTanh  = EIGEN_FAST_MATH,
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+  };
+};
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet4d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 1,
+
+    HasDiv  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+  };
+};
+#endif
+
+template<> struct scalar_div_cost<float,true> { enum { value = 14 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 16 }; };
+
+/* Proper support for integers is only provided by AVX2. In the meantime, we'll
+   use SSE instructions and packets to deal with integers.
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet8i type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8
+  };
+};
+*/
+
+template<> struct unpacket_traits<Packet8f> { typedef float  type; typedef Packet4f half; enum {size=8, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet4d> { typedef double type; typedef Packet2d half; enum {size=4, alignment=Aligned32}; };
+template<> struct unpacket_traits<Packet8i> { typedef int    type; typedef Packet4i half; enum {size=8, alignment=Aligned32}; };
+
+template<> EIGEN_STRONG_INLINE Packet8f pset1<Packet8f>(const float&  from) { return _mm256_set1_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pset1<Packet4d>(const double& from) { return _mm256_set1_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i pset1<Packet8i>(const int&    from) { return _mm256_set1_epi32(from); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pload1<Packet8f>(const float*  from) { return _mm256_broadcast_ss(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pload1<Packet4d>(const double* from) { return _mm256_broadcast_sd(from); }
+
+template<> EIGEN_STRONG_INLINE Packet8f plset<Packet8f>(const float& a) { return _mm256_add_ps(_mm256_set1_ps(a), _mm256_set_ps(7.0,6.0,5.0,4.0,3.0,2.0,1.0,0.0)); }
+template<> EIGEN_STRONG_INLINE Packet4d plset<Packet4d>(const double& a) { return _mm256_add_pd(_mm256_set1_pd(a), _mm256_set_pd(3.0,2.0,1.0,0.0)); }
+
+template<> EIGEN_STRONG_INLINE Packet8f padd<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_add_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d padd<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_add_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f psub<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_sub_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d psub<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_sub_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pnegate(const Packet8f& a)
+{
+  return _mm256_sub_ps(_mm256_set1_ps(0.0),a);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pnegate(const Packet4d& a)
+{
+  return _mm256_sub_pd(_mm256_set1_pd(0.0),a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pconj(const Packet8f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4d pconj(const Packet4d& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet8i pconj(const Packet8i& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet8f pmul<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_mul_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pmul<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_mul_pd(a,b); }
+
+
+template<> EIGEN_STRONG_INLINE Packet8f pdiv<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_div_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pdiv<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_div_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, const Packet8i& /*b*/)
+{ eigen_assert(false && "packet integer division are not supported by AVX");
+  return pset1<Packet8i>(0);
+}
+
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
+  // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
+  // and gcc stupidly generates a vfmadd132ps instruction,
+  // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
+  // the result of the product.
+  Packet8f res = c;
+  __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
+  return res;
+#else
+  return _mm256_fmadd_ps(a,b,c);
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
+#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
+  // see above
+  Packet4d res = c;
+  __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
+  return res;
+#else
+  return _mm256_fmadd_pd(a,b,c);
+#endif
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet8f pmin<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_min_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pmin<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_min_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pmax<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_max_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pmax<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_max_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pround<Packet8f>(const Packet8f& a) { return _mm256_round_ps(a, _MM_FROUND_CUR_DIRECTION); }
+template<> EIGEN_STRONG_INLINE Packet4d pround<Packet4d>(const Packet4d& a) { return _mm256_round_pd(a, _MM_FROUND_CUR_DIRECTION); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pceil<Packet8f>(const Packet8f& a) { return _mm256_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pceil<Packet4d>(const Packet4d& a) { return _mm256_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pfloor<Packet8f>(const Packet8f& a) { return _mm256_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet4d pfloor<Packet4d>(const Packet4d& a) { return _mm256_floor_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pand<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_and_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pand<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_and_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f por<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_or_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d por<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_or_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pxor<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_xor_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pxor<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_xor_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pandnot<Packet8f>(const Packet8f& a, const Packet8f& b) { return _mm256_andnot_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4d pandnot<Packet4d>(const Packet4d& a, const Packet4d& b) { return _mm256_andnot_pd(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet8f pload<Packet8f>(const float*   from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d pload<Packet4d>(const double*  from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i pload<Packet8i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm256_load_si256(reinterpret_cast<const __m256i*>(from)); }
+
+template<> EIGEN_STRONG_INLINE Packet8f ploadu<Packet8f>(const float* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet4d ploadu<Packet4d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet8i ploadu<Packet8i>(const int* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from)); }
+
+// Loads 4 floats from memory a returns the packet {a0, a0  a1, a1, a2, a2, a3, a3}
+template<> EIGEN_STRONG_INLINE Packet8f ploaddup<Packet8f>(const float* from)
+{
+  // TODO try to find a way to avoid the need of a temporary register
+//   Packet8f tmp  = _mm256_castps128_ps256(_mm_loadu_ps(from));
+//   tmp = _mm256_insertf128_ps(tmp, _mm_movehl_ps(_mm256_castps256_ps128(tmp),_mm256_castps256_ps128(tmp)), 1);
+//   return _mm256_unpacklo_ps(tmp,tmp);
+  
+  // _mm256_insertf128_ps is very slow on Haswell, thus:
+  Packet8f tmp = _mm256_broadcast_ps((const __m128*)(const void*)from);
+  // mimic an "inplace" permutation of the lower 128bits using a blend
+  tmp = _mm256_blend_ps(tmp,_mm256_castps128_ps256(_mm_permute_ps( _mm256_castps256_ps128(tmp), _MM_SHUFFLE(1,0,1,0))), 15);
+  // then we can perform a consistent permutation on the global register to get everything in shape:
+  return  _mm256_permute_ps(tmp, _MM_SHUFFLE(3,3,2,2));
+}
+// Loads 2 doubles from memory a returns the packet {a0, a0  a1, a1}
+template<> EIGEN_STRONG_INLINE Packet4d ploaddup<Packet4d>(const double* from)
+{
+  Packet4d tmp = _mm256_broadcast_pd((const __m128d*)(const void*)from);
+  return  _mm256_permute_pd(tmp, 3<<2);
+}
+
+// Loads 2 floats from memory a returns the packet {a0, a0  a0, a0, a1, a1, a1, a1}
+template<> EIGEN_STRONG_INLINE Packet8f ploadquad<Packet8f>(const float* from)
+{
+  Packet8f tmp = _mm256_castps128_ps256(_mm_broadcast_ss(from));
+  return _mm256_insertf128_ps(tmp, _mm_broadcast_ss(from+1), 1);
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet8f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_store_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet8i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet8f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet4d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet8i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm256_storeu_si256(reinterpret_cast<__m256i*>(to), from); }
+
+// NOTE: leverage _mm256_i32gather_ps and _mm256_i32gather_pd if AVX2 instructions are available
+// NOTE: for the record the following seems to be slower: return _mm256_i32gather_ps(from, _mm256_set1_epi32(stride), 4);
+template<> EIGEN_DEVICE_FUNC inline Packet8f pgather<float, Packet8f>(const float* from, Index stride)
+{
+  return _mm256_set_ps(from[7*stride], from[6*stride], from[5*stride], from[4*stride],
+                       from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4d pgather<double, Packet4d>(const double* from, Index stride)
+{
+  return _mm256_set_pd(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet8f>(float* to, const Packet8f& from, Index stride)
+{
+  __m128 low = _mm256_extractf128_ps(from, 0);
+  to[stride*0] = _mm_cvtss_f32(low);
+  to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 1));
+  to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 2));
+  to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(low, low, 3));
+
+  __m128 high = _mm256_extractf128_ps(from, 1);
+  to[stride*4] = _mm_cvtss_f32(high);
+  to[stride*5] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 1));
+  to[stride*6] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 2));
+  to[stride*7] = _mm_cvtss_f32(_mm_shuffle_ps(high, high, 3));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet4d>(double* to, const Packet4d& from, Index stride)
+{
+  __m128d low = _mm256_extractf128_pd(from, 0);
+  to[stride*0] = _mm_cvtsd_f64(low);
+  to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(low, low, 1));
+  __m128d high = _mm256_extractf128_pd(from, 1);
+  to[stride*2] = _mm_cvtsd_f64(high);
+  to[stride*3] = _mm_cvtsd_f64(_mm_shuffle_pd(high, high, 1));
+}
+
+template<> EIGEN_STRONG_INLINE void pstore1<Packet8f>(float* to, const float& a)
+{
+  Packet8f pa = pset1<Packet8f>(a);
+  pstore(to, pa);
+}
+template<> EIGEN_STRONG_INLINE void pstore1<Packet4d>(double* to, const double& a)
+{
+  Packet4d pa = pset1<Packet4d>(a);
+  pstore(to, pa);
+}
+template<> EIGEN_STRONG_INLINE void pstore1<Packet8i>(int* to, const int& a)
+{
+  Packet8i pa = pset1<Packet8i>(a);
+  pstore(to, pa);
+}
+
+#ifndef EIGEN_VECTORIZE_AVX512
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+#endif
+
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet8f>(const Packet8f& a) {
+  return _mm_cvtss_f32(_mm256_castps256_ps128(a));
+}
+template<> EIGEN_STRONG_INLINE double pfirst<Packet4d>(const Packet4d& a) {
+  return _mm_cvtsd_f64(_mm256_castpd256_pd128(a));
+}
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet8i>(const Packet8i& a) {
+  return _mm_cvtsi128_si32(_mm256_castsi256_si128(a));
+}
+
+
+template<> EIGEN_STRONG_INLINE Packet8f preverse(const Packet8f& a)
+{
+  __m256 tmp = _mm256_shuffle_ps(a,a,0x1b);
+  return _mm256_permute2f128_ps(tmp, tmp, 1);
+}
+template<> EIGEN_STRONG_INLINE Packet4d preverse(const Packet4d& a)
+{
+   __m256d tmp = _mm256_shuffle_pd(a,a,5);
+  return _mm256_permute2f128_pd(tmp, tmp, 1);
+
+  __m256d swap_halves = _mm256_permute2f128_pd(a,a,1);
+    return _mm256_permute_pd(swap_halves,5);
+}
+
+// pabs should be ok
+template<> EIGEN_STRONG_INLINE Packet8f pabs(const Packet8f& a)
+{
+  const Packet8f mask = _mm256_castsi256_ps(_mm256_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
+  return _mm256_and_ps(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pabs(const Packet4d& a)
+{
+  const Packet4d mask = _mm256_castsi256_pd(_mm256_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
+  return _mm256_and_pd(a,mask);
+}
+
+// preduxp should be ok
+// FIXME: why is this ok? why isn't the simply implementation working as expected?
+template<> EIGEN_STRONG_INLINE Packet8f preduxp<Packet8f>(const Packet8f* vecs)
+{
+    __m256 hsum1 = _mm256_hadd_ps(vecs[0], vecs[1]);
+    __m256 hsum2 = _mm256_hadd_ps(vecs[2], vecs[3]);
+    __m256 hsum3 = _mm256_hadd_ps(vecs[4], vecs[5]);
+    __m256 hsum4 = _mm256_hadd_ps(vecs[6], vecs[7]);
+
+    __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+    __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+    __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+    __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+    __m256 perm1 =  _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+    __m256 perm2 =  _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+    __m256 perm3 =  _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+    __m256 perm4 =  _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+    __m256 sum1 = _mm256_add_ps(perm1, hsum5);
+    __m256 sum2 = _mm256_add_ps(perm2, hsum6);
+    __m256 sum3 = _mm256_add_ps(perm3, hsum7);
+    __m256 sum4 = _mm256_add_ps(perm4, hsum8);
+
+    __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+    __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+    __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
+    return final;
+}
+template<> EIGEN_STRONG_INLINE Packet4d preduxp<Packet4d>(const Packet4d* vecs)
+{
+ Packet4d tmp0, tmp1;
+
+  tmp0 = _mm256_hadd_pd(vecs[0], vecs[1]);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs[2], vecs[3]);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  return _mm256_blend_pd(tmp0, tmp1, 0xC);
+}
+
+template<> EIGEN_STRONG_INLINE float predux<Packet8f>(const Packet8f& a)
+{
+  return predux(Packet4f(_mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1))));
+}
+template<> EIGEN_STRONG_INLINE double predux<Packet4d>(const Packet4d& a)
+{
+  return predux(Packet2d(_mm_add_pd(_mm256_castpd256_pd128(a),_mm256_extractf128_pd(a,1))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f predux_downto4<Packet8f>(const Packet8f& a)
+{
+  return _mm_add_ps(_mm256_castps256_ps128(a),_mm256_extractf128_ps(a,1));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp;
+  tmp = _mm256_mul_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_mul_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp;
+  tmp = _mm256_mul_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_mul_pd(tmp, _mm256_shuffle_pd(tmp,tmp,1)));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_min<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp = _mm256_min_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_min_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_min<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp = _mm256_min_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_min_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1)));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_max<Packet8f>(const Packet8f& a)
+{
+  Packet8f tmp = _mm256_max_ps(a, _mm256_permute2f128_ps(a,a,1));
+  tmp = _mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,_MM_SHUFFLE(1,0,3,2)));
+  return pfirst(_mm256_max_ps(tmp, _mm256_shuffle_ps(tmp,tmp,1)));
+}
+
+template<> EIGEN_STRONG_INLINE double predux_max<Packet4d>(const Packet4d& a)
+{
+  Packet4d tmp = _mm256_max_pd(a, _mm256_permute2f128_pd(a,a,1));
+  return pfirst(_mm256_max_pd(tmp, _mm256_shuffle_pd(tmp, tmp, 1)));
+}
+
+
+template<int Offset>
+struct palign_impl<Offset,Packet8f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet8f& first, const Packet8f& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm256_blend_ps(first, second, 1);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0x88);
+    }
+    else if (Offset==2)
+    {
+      first = _mm256_blend_ps(first, second, 3);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xcc);
+    }
+    else if (Offset==3)
+    {
+      first = _mm256_blend_ps(first, second, 7);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_blend_ps(tmp1, tmp2, 0xee);
+    }
+    else if (Offset==4)
+    {
+      first = _mm256_blend_ps(first, second, 15);
+      Packet8f tmp1 = _mm256_permute_ps (first, _MM_SHUFFLE(3,2,1,0));
+      Packet8f tmp2 = _mm256_permute2f128_ps (tmp1, tmp1, 1);
+      first = _mm256_permute_ps(tmp2, _MM_SHUFFLE(3,2,1,0));
+    }
+    else if (Offset==5)
+    {
+      first = _mm256_blend_ps(first, second, 31);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(0,3,2,1));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0x88);
+    }
+    else if (Offset==6)
+    {
+      first = _mm256_blend_ps(first, second, 63);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(1,0,3,2));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0xcc);
+    }
+    else if (Offset==7)
+    {
+      first = _mm256_blend_ps(first, second, 127);
+      first = _mm256_permute2f128_ps(first, first, 1);
+      Packet8f tmp = _mm256_permute_ps (first, _MM_SHUFFLE(2,1,0,3));
+      first = _mm256_permute2f128_ps(tmp, tmp, 1);
+      first = _mm256_blend_ps(tmp, first, 0xee);
+    }
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet4d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4d& first, const Packet4d& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm256_blend_pd(first, second, 1);
+      __m256d tmp = _mm256_permute_pd(first, 5);
+      first = _mm256_permute2f128_pd(tmp, tmp, 1);
+      first = _mm256_blend_pd(tmp, first, 0xA);
+    }
+    else if (Offset==2)
+    {
+      first = _mm256_blend_pd(first, second, 3);
+      first = _mm256_permute2f128_pd(first, first, 1);
+    }
+    else if (Offset==3)
+    {
+      first = _mm256_blend_pd(first, second, 7);
+      __m256d tmp = _mm256_permute_pd(first, 5);
+      first = _mm256_permute2f128_pd(tmp, tmp, 1);
+      first = _mm256_blend_pd(tmp, first, 5);
+    }
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet8f,8>& kernel) {
+  __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T4 = _mm256_unpacklo_ps(kernel.packet[4], kernel.packet[5]);
+  __m256 T5 = _mm256_unpackhi_ps(kernel.packet[4], kernel.packet[5]);
+  __m256 T6 = _mm256_unpacklo_ps(kernel.packet[6], kernel.packet[7]);
+  __m256 T7 = _mm256_unpackhi_ps(kernel.packet[6], kernel.packet[7]);
+  __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0));
+  __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2));
+  __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0));
+  __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2));
+  __m256 S4 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(1,0,1,0));
+  __m256 S5 = _mm256_shuffle_ps(T4,T6,_MM_SHUFFLE(3,2,3,2));
+  __m256 S6 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(1,0,1,0));
+  __m256 S7 = _mm256_shuffle_ps(T5,T7,_MM_SHUFFLE(3,2,3,2));
+  kernel.packet[0] = _mm256_permute2f128_ps(S0, S4, 0x20);
+  kernel.packet[1] = _mm256_permute2f128_ps(S1, S5, 0x20);
+  kernel.packet[2] = _mm256_permute2f128_ps(S2, S6, 0x20);
+  kernel.packet[3] = _mm256_permute2f128_ps(S3, S7, 0x20);
+  kernel.packet[4] = _mm256_permute2f128_ps(S0, S4, 0x31);
+  kernel.packet[5] = _mm256_permute2f128_ps(S1, S5, 0x31);
+  kernel.packet[6] = _mm256_permute2f128_ps(S2, S6, 0x31);
+  kernel.packet[7] = _mm256_permute2f128_ps(S3, S7, 0x31);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet8f,4>& kernel) {
+  __m256 T0 = _mm256_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T1 = _mm256_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m256 T2 = _mm256_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m256 T3 = _mm256_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+
+  __m256 S0 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(1,0,1,0));
+  __m256 S1 = _mm256_shuffle_ps(T0,T2,_MM_SHUFFLE(3,2,3,2));
+  __m256 S2 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(1,0,1,0));
+  __m256 S3 = _mm256_shuffle_ps(T1,T3,_MM_SHUFFLE(3,2,3,2));
+
+  kernel.packet[0] = _mm256_permute2f128_ps(S0, S1, 0x20);
+  kernel.packet[1] = _mm256_permute2f128_ps(S2, S3, 0x20);
+  kernel.packet[2] = _mm256_permute2f128_ps(S0, S1, 0x31);
+  kernel.packet[3] = _mm256_permute2f128_ps(S2, S3, 0x31);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4d,4>& kernel) {
+  __m256d T0 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 15);
+  __m256d T1 = _mm256_shuffle_pd(kernel.packet[0], kernel.packet[1], 0);
+  __m256d T2 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 15);
+  __m256d T3 = _mm256_shuffle_pd(kernel.packet[2], kernel.packet[3], 0);
+
+  kernel.packet[1] = _mm256_permute2f128_pd(T0, T2, 32);
+  kernel.packet[3] = _mm256_permute2f128_pd(T0, T2, 49);
+  kernel.packet[0] = _mm256_permute2f128_pd(T1, T3, 32);
+  kernel.packet[2] = _mm256_permute2f128_pd(T1, T3, 49);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pblend(const Selector<8>& ifPacket, const Packet8f& thenPacket, const Packet8f& elsePacket) {
+  const __m256 zero = _mm256_setzero_ps();
+  const __m256 select = _mm256_set_ps(ifPacket.select[7], ifPacket.select[6], ifPacket.select[5], ifPacket.select[4], ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256 false_mask = _mm256_cmp_ps(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_ps(thenPacket, elsePacket, false_mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4d pblend(const Selector<4>& ifPacket, const Packet4d& thenPacket, const Packet4d& elsePacket) {
+  const __m256d zero = _mm256_setzero_pd();
+  const __m256d select = _mm256_set_pd(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m256d false_mask = _mm256_cmp_pd(select, zero, _CMP_EQ_UQ);
+  return _mm256_blendv_pd(thenPacket, elsePacket, false_mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pinsertfirst(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertfirst(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),1);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pinsertlast(const Packet8f& a, float b)
+{
+  return _mm256_blend_ps(a,pset1<Packet8f>(b),(1<<7));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4d pinsertlast(const Packet4d& a, double b)
+{
+  return _mm256_blend_pd(a,pset1<Packet4d>(b),(1<<3));
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_AVX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h
new file mode 100644
index 0000000..83bfdc6
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX/TypeCasting.h
@@ -0,0 +1,51 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_AVX_H
+#define EIGEN_TYPE_CASTING_AVX_H
+
+namespace Eigen {
+
+namespace internal {
+
+// For now we use SSE to handle integers, so we can't use AVX instructions to cast
+// from int to float
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 0,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+
+
+template<> EIGEN_STRONG_INLINE Packet8i pcast<Packet8f, Packet8i>(const Packet8f& a) {
+  return _mm256_cvtps_epi32(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8i, Packet8f>(const Packet8i& a) {
+  return _mm256_cvtepi32_ps(a);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_AVX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h
new file mode 100644
index 0000000..399be0e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/MathFunctions.h
@@ -0,0 +1,396 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Pedro Gonnet (pedro.gonnet@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+#define THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
+
+namespace Eigen {
+
+namespace internal {
+
+// Disable the code for older versions of gcc that don't support many of the required avx512 instrinsics.
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+
+#define _EIGEN_DECLARE_CONST_Packet16f(NAME, X) \
+  const Packet16f p16f_##NAME = pset1<Packet16f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(NAME, X) \
+  const Packet16f p16f_##NAME = (__m512)pset1<Packet16i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d(NAME, X) \
+  const Packet8d p8d_##NAME = pset1<Packet8d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(NAME, X) \
+  const Packet8d p8d_##NAME = _mm512_castsi512_pd(_mm512_set1_epi64(X))
+
+// Natural logarithm
+// Computes log(x) as log(2^e * m) = C*e + log(m), where the constant C =log(2)
+// and m is in the range [sqrt(1/2),sqrt(2)). In this range, the logarithm can
+// be easily approximated by a polynomial centered on m=1 for stability.
+#if defined(EIGEN_VECTORIZE_AVX512DQ)
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+plog<Packet16f>(const Packet16f& _x) {
+  Packet16f x = _x;
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(126f, 126.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  // The smallest non denormalized float number.
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(min_norm_pos, 0x00800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(minus_inf, 0xff800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+
+  // Polynomial coefficients.
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p1, -1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p3, -1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p4, +1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p5, -1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p6, +2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p7, -2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_p8, +3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_log_q2, 0.693359375f);
+
+  // invalid_mask is set to true when x is NaN
+  __mmask16 invalid_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_NGE_UQ);
+  __mmask16 iszero_mask =
+      _mm512_cmp_ps_mask(x, _mm512_setzero_ps(), _CMP_EQ_UQ);
+
+  // Truncate input values to the minimum positive normal.
+  x = pmax(x, p16f_min_norm_pos);
+
+  // Extract the shifted exponents.
+  Packet16f emm0 = _mm512_cvtepi32_ps(_mm512_srli_epi32((__m512i)x, 23));
+  Packet16f e = _mm512_sub_ps(emm0, p16f_126f);
+
+  // Set the exponents to -1, i.e. x are in the range [0.5,1).
+  x = _mm512_and_ps(x, p16f_inv_mant_mask);
+  x = _mm512_or_ps(x, p16f_half);
+
+  // part2: Shift the inputs from the range [0.5,1) to [sqrt(1/2),sqrt(2))
+  // and shift by -1. The values are then centered around 0, which improves
+  // the stability of the polynomial evaluation.
+  //   if( x < SQRTHF ) {
+  //     e -= 1;
+  //     x = x + x - 1.0;
+  //   } else { x = x - 1.0; }
+  __mmask16 mask = _mm512_cmp_ps_mask(x, p16f_cephes_SQRTHF, _CMP_LT_OQ);
+  Packet16f tmp = _mm512_mask_blend_ps(mask, x, _mm512_setzero_ps());
+  x = psub(x, p16f_1);
+  e = psub(e, _mm512_mask_blend_ps(mask, p16f_1, _mm512_setzero_ps()));
+  x = padd(x, tmp);
+
+  Packet16f x2 = pmul(x, x);
+  Packet16f x3 = pmul(x2, x);
+
+  // Evaluate the polynomial approximant of degree 8 in three parts, probably
+  // to improve instruction-level parallelism.
+  Packet16f y, y1, y2;
+  y = pmadd(p16f_cephes_log_p0, x, p16f_cephes_log_p1);
+  y1 = pmadd(p16f_cephes_log_p3, x, p16f_cephes_log_p4);
+  y2 = pmadd(p16f_cephes_log_p6, x, p16f_cephes_log_p7);
+  y = pmadd(y, x, p16f_cephes_log_p2);
+  y1 = pmadd(y1, x, p16f_cephes_log_p5);
+  y2 = pmadd(y2, x, p16f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  // Add the logarithm of the exponent back to the result of the interpolation.
+  y1 = pmul(e, p16f_cephes_log_q1);
+  tmp = pmul(x2, p16f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p16f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+
+  // Filter out invalid inputs, i.e. negative arg will be NAN, 0 will be -INF.
+  return _mm512_mask_blend_ps(iszero_mask, p16f_minus_inf,
+                              _mm512_mask_blend_ps(invalid_mask, p16f_nan, x));
+}
+#endif
+
+// Exponential function. Works by writing "x = m*log(2) + r" where
+// "m = floor(x/log(2)+1/2)" and "r" is the remainder. The result is then
+// "exp(x) = 2^m*exp(r)" where exp(r) is in the range [-1,1).
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+pexp<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(1, 1.0f);
+  _EIGEN_DECLARE_CONST_Packet16f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(127, 127.0f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(exp_hi, 88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet16f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_LOG2EF, 1.44269504088896341f);
+
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet16f(cephes_exp_p5, 5.0000001201E-1f);
+
+  // Clamp x.
+  Packet16f x = pmax(pmin(_x, p16f_exp_hi), p16f_exp_lo);
+
+  // Express exp(x) as exp(m*ln(2) + r), start by extracting
+  // m = floor(x/ln(2) + 0.5).
+  Packet16f m = _mm512_floor_ps(pmadd(x, p16f_cephes_LOG2EF, p16f_half));
+
+  // Get r = x - m*ln(2). Note that we can do this without losing more than one
+  // ulp precision due to the FMA instruction.
+  _EIGEN_DECLARE_CONST_Packet16f(nln2, -0.6931471805599453f);
+  Packet16f r = _mm512_fmadd_ps(m, p16f_nln2, x);
+  Packet16f r2 = pmul(r, r);
+
+  // TODO(gonnet): Split into odd/even polynomials and try to exploit
+  //               instruction-level parallelism.
+  Packet16f y = p16f_cephes_exp_p0;
+  y = pmadd(y, r, p16f_cephes_exp_p1);
+  y = pmadd(y, r, p16f_cephes_exp_p2);
+  y = pmadd(y, r, p16f_cephes_exp_p3);
+  y = pmadd(y, r, p16f_cephes_exp_p4);
+  y = pmadd(y, r, p16f_cephes_exp_p5);
+  y = pmadd(y, r2, r);
+  y = padd(y, p16f_1);
+
+  // Build emm0 = 2^m.
+  Packet16i emm0 = _mm512_cvttps_epi32(padd(m, p16f_127));
+  emm0 = _mm512_slli_epi32(emm0, 23);
+
+  // Return 2^m * exp(r).
+  return pmax(pmul(y, _mm512_castsi512_ps(emm0)), _x);
+}
+
+/*template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+pexp<Packet8d>(const Packet8d& _x) {
+  Packet8d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet8d(1, 1.0);
+  _EIGEN_DECLARE_CONST_Packet8d(2, 2.0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(exp_hi, 709.437);
+  _EIGEN_DECLARE_CONST_Packet8d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet8d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+  // clamp x
+  x = pmax(pmin(x, p8d_exp_hi), p8d_exp_lo);
+
+  // Express exp(x) as exp(g + n*log(2)).
+  const Packet8d n =
+      _mm512_mul_round_pd(p8d_cephes_LOG2EF, x, _MM_FROUND_TO_NEAREST_INT);
+
+  // Get the remainder modulo log(2), i.e. the "g" described above. Subtract
+  // n*log(2) out in two steps, i.e. n*C1 + n*C2, C1+C2=log2 to get the last
+  // digits right.
+  const Packet8d nC1 = pmul(n, p8d_cephes_exp_C1);
+  const Packet8d nC2 = pmul(n, p8d_cephes_exp_C2);
+  x = psub(x, nC1);
+  x = psub(x, nC2);
+
+  const Packet8d x2 = pmul(x, x);
+
+  // Evaluate the numerator polynomial of the rational interpolant.
+  Packet8d px = p8d_cephes_exp_p0;
+  px = pmadd(px, x2, p8d_cephes_exp_p1);
+  px = pmadd(px, x2, p8d_cephes_exp_p2);
+  px = pmul(px, x);
+
+  // Evaluate the denominator polynomial of the rational interpolant.
+  Packet8d qx = p8d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p8d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p8d_cephes_exp_q3);
+
+  // I don't really get this bit, copied from the SSE2 routines, so...
+  // TODO(gonnet): Figure out what is going on here, perhaps find a better
+  // rational interpolant?
+  x = _mm512_div_pd(px, psub(qx, px));
+  x = pmadd(p8d_2, x, p8d_1);
+
+  // Build e=2^n.
+  const Packet8d e = _mm512_castsi512_pd(_mm512_slli_epi64(
+      _mm512_add_epi64(_mm512_cvtpd_epi64(n), _mm512_set1_epi64(1023)), 52));
+
+  // Construct the result 2^n * exp(g) = e * x. The max is used to catch
+  // non-finite values in the input.
+  return pmax(pmul(x, e), _x);
+  }*/
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. The main advantage of this approach is not just speed, but
+// also the fact that it can be inlined and pipelined with other computations,
+// further reducing its effective latency.
+#if EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+psqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 non_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_GE_OQ);
+  Packet16f x = _mm512_mask_blend_ps(non_zero_mask, _mm512_rsqrt14_ps(_x),
+                                     _mm512_setzero_ps());
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+psqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 non_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_GE_OQ);
+  Packet8d x = _mm512_mask_blend_pd(non_zero_mask, _mm512_rsqrt14_pd(_x),
+                                    _mm512_setzero_pd());
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Multiply the original _x by it's reciprocal square root to extract the
+  // square root.
+  return pmul(_x, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f psqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_sqrt_ps(x);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d psqrt<Packet8d>(const Packet8d& x) {
+  return _mm512_sqrt_pd(x);
+}
+#endif
+
+// Functions for rsqrt.
+// Almost identical to the sqrt routine, just leave out the last multiplication
+// and fill in NaN/Inf where needed. Note that this function only exists as an
+// iterative version for doubles since there is no instruction for diretly
+// computing the reciprocal square root in AVX-512.
+#ifdef EIGEN_FAST_MATH
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet16f
+prsqrt<Packet16f>(const Packet16f& _x) {
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet16f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet16f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet16f_FROM_INT(flt_min, 0x00800000);
+
+  Packet16f neg_half = pmul(_x, p16f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask16 le_zero_mask = _mm512_cmp_ps_mask(_x, p16f_flt_min, _CMP_LT_OQ);
+  Packet16f x = _mm512_mask_blend_ps(le_zero_mask, _mm512_setzero_ps(),
+                                     _mm512_rsqrt14_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask16 neg_mask = _mm512_cmp_ps_mask(_x, _mm512_setzero_ps(), _CMP_LT_OQ);
+  Packet16f infs_and_nans = _mm512_mask_blend_ps(
+      neg_mask, p16f_nan,
+      _mm512_mask_blend_ps(le_zero_mask, p16f_inf, _mm512_setzero_ps()));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p16f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_ps(le_zero_mask, infs_and_nans, x);
+}
+
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet8d
+prsqrt<Packet8d>(const Packet8d& _x) {
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(inf, 0x7ff0000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(nan, 0x7ff1000000000000LL);
+  _EIGEN_DECLARE_CONST_Packet8d(one_point_five, 1.5);
+  _EIGEN_DECLARE_CONST_Packet8d(minus_half, -0.5);
+  _EIGEN_DECLARE_CONST_Packet8d_FROM_INT64(dbl_min, 0x0010000000000000LL);
+
+  Packet8d neg_half = pmul(_x, p8d_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  __mmask8 le_zero_mask = _mm512_cmp_pd_mask(_x, p8d_dbl_min, _CMP_LT_OQ);
+  Packet8d x = _mm512_mask_blend_pd(le_zero_mask, _mm512_setzero_pd(),
+                                    _mm512_rsqrt14_pd(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  __mmask8 neg_mask = _mm512_cmp_pd_mask(_x, _mm512_setzero_pd(), _CMP_LT_OQ);
+  Packet8d infs_and_nans = _mm512_mask_blend_pd(
+      neg_mask, p8d_nan,
+      _mm512_mask_blend_pd(le_zero_mask, p8d_inf, _mm512_setzero_pd()));
+
+  // Do a first step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Do a second step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p8d_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm512_mask_blend_pd(le_zero_mask, infs_and_nans, x);
+}
+#else
+template <>
+EIGEN_STRONG_INLINE Packet16f prsqrt<Packet16f>(const Packet16f& x) {
+  return _mm512_rsqrt28_ps(x);
+}
+#endif
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // THIRD_PARTY_EIGEN3_EIGEN_SRC_CORE_ARCH_AVX512_MATHFUNCTIONS_H_
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h
new file mode 100644
index 0000000..f6500a1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AVX512/PacketMath.h
@@ -0,0 +1,1316 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner (benoit.steiner.goog@gmail.com)
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_AVX512_H
+#define EIGEN_PACKET_MATH_AVX512_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
+#endif
+
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+#endif
+
+typedef __m512 Packet16f;
+typedef __m512i Packet16i;
+typedef __m512d Packet8d;
+
+template <>
+struct is_arithmetic<__m512> {
+  enum { value = true };
+};
+template <>
+struct is_arithmetic<__m512i> {
+  enum { value = true };
+};
+template <>
+struct is_arithmetic<__m512d> {
+  enum { value = true };
+};
+
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet16f type;
+  typedef Packet8f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 1,
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+    HasLog = 1,
+#endif
+    HasExp = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+#endif
+    HasDiv = 1
+  };
+ };
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet8d type;
+  typedef Packet4d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 8,
+    HasHalfPacket = 1,
+#if EIGEN_GNUC_AT_LEAST(5, 3)
+    HasSqrt = 1,
+    HasRsqrt = EIGEN_FAST_MATH,
+#endif
+    HasDiv = 1
+  };
+};
+
+/* TODO Implement AVX512 for integers
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet16i type;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=8
+  };
+};
+*/
+
+template <>
+struct unpacket_traits<Packet16f> {
+  typedef float type;
+  typedef Packet8f half;
+  enum { size = 16, alignment=Aligned64 };
+};
+template <>
+struct unpacket_traits<Packet8d> {
+  typedef double type;
+  typedef Packet4d half;
+  enum { size = 8, alignment=Aligned64 };
+};
+template <>
+struct unpacket_traits<Packet16i> {
+  typedef int type;
+  typedef Packet8i half;
+  enum { size = 16, alignment=Aligned64 };
+};
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pset1<Packet16f>(const float& from) {
+  return _mm512_set1_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pset1<Packet8d>(const double& from) {
+  return _mm512_set1_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pset1<Packet16i>(const int& from) {
+  return _mm512_set1_epi32(from);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pload1<Packet16f>(const float* from) {
+  return _mm512_broadcastss_ps(_mm_load_ps1(from));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pload1<Packet8d>(const double* from) {
+  return _mm512_broadcastsd_pd(_mm_load_pd1(from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f plset<Packet16f>(const float& a) {
+  return _mm512_add_ps(
+      _mm512_set1_ps(a),
+      _mm512_set_ps(15.0f, 14.0f, 13.0f, 12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f,
+                    4.0f, 3.0f, 2.0f, 1.0f, 0.0f));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d plset<Packet8d>(const double& a) {
+  return _mm512_add_pd(_mm512_set1_pd(a),
+                       _mm512_set_pd(7.0, 6.0, 5.0, 4.0, 3.0, 2.0, 1.0, 0.0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f padd<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_add_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d padd<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_add_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f psub<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_sub_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d psub<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_sub_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pnegate(const Packet16f& a) {
+  return _mm512_sub_ps(_mm512_set1_ps(0.0), a);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pnegate(const Packet8d& a) {
+  return _mm512_sub_pd(_mm512_set1_pd(0.0), a);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pconj(const Packet16f& a) {
+  return a;
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pconj(const Packet8d& a) {
+  return a;
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pconj(const Packet16i& a) {
+  return a;
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmul<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_mul_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmul<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_mul_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pdiv<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_div_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pdiv<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_div_pd(a, b);
+}
+
+#ifdef __FMA__
+template <>
+EIGEN_STRONG_INLINE Packet16f pmadd(const Packet16f& a, const Packet16f& b,
+                                    const Packet16f& c) {
+  return _mm512_fmadd_ps(a, b, c);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmadd(const Packet8d& a, const Packet8d& b,
+                                   const Packet8d& c) {
+  return _mm512_fmadd_pd(a, b, c);
+}
+#endif
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmin<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_min_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmin<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_min_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pmax<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+  return _mm512_max_ps(a, b);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pmax<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+  return _mm512_max_pd(a, b);
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pand<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_and_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_and_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pand<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_and_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_and_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_and_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet16f por<Packet16f>(const Packet16f& a,
+                                             const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_or_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_or_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet8d por<Packet8d>(const Packet8d& a,
+                                           const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_or_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_or_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_or_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pxor<Packet16f>(const Packet16f& a,
+                                              const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_xor_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_xor_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pxor<Packet8d>(const Packet8d& a,
+                                            const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_xor_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_xor_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_xor_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pandnot<Packet16f>(const Packet16f& a,
+                                                 const Packet16f& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_andnot_ps(a, b);
+#else
+  Packet16f res = _mm512_undefined_ps();
+  Packet4f lane0_a = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane0_b = _mm512_extractf32x4_ps(b, 0);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane0_a, lane0_b), 0);
+
+  Packet4f lane1_a = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane1_b = _mm512_extractf32x4_ps(b, 1);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane1_a, lane1_b), 1);
+
+  Packet4f lane2_a = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane2_b = _mm512_extractf32x4_ps(b, 2);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane2_a, lane2_b), 2);
+
+  Packet4f lane3_a = _mm512_extractf32x4_ps(a, 3);
+  Packet4f lane3_b = _mm512_extractf32x4_ps(b, 3);
+  res = _mm512_insertf32x4(res, _mm_andnot_ps(lane3_a, lane3_b), 3);
+
+  return res;
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pandnot<Packet8d>(const Packet8d& a,
+                                               const Packet8d& b) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  return _mm512_andnot_pd(a, b);
+#else
+  Packet8d res = _mm512_undefined_pd();
+  Packet4d lane0_a = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane0_b = _mm512_extractf64x4_pd(b, 0);
+  res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane0_a, lane0_b), 0);
+
+  Packet4d lane1_a = _mm512_extractf64x4_pd(a, 1);
+  Packet4d lane1_b = _mm512_extractf64x4_pd(b, 1);
+  res = _mm512_insertf64x4(res, _mm256_andnot_pd(lane1_a, lane1_b), 1);
+
+  return res;
+#endif
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f pload<Packet16f>(const float* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pload<Packet8d>(const double* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i pload<Packet16i>(const int* from) {
+  EIGEN_DEBUG_ALIGNED_LOAD return _mm512_load_si512(
+      reinterpret_cast<const __m512i*>(from));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet16f ploadu<Packet16f>(const float* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_ps(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploadu<Packet8d>(const double* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_pd(from);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16i ploadu<Packet16i>(const int* from) {
+  EIGEN_DEBUG_UNALIGNED_LOAD return _mm512_loadu_si512(
+      reinterpret_cast<const __m512i*>(from));
+}
+
+// Loads 8 floats from memory a returns the packet
+// {a0, a0  a1, a1, a2, a2, a3, a3, a4, a4, a5, a5, a6, a6, a7, a7}
+template <>
+EIGEN_STRONG_INLINE Packet16f ploaddup<Packet16f>(const float* from) {
+  Packet8f lane0 = _mm256_broadcast_ps((const __m128*)(const void*)from);
+  // mimic an "inplace" permutation of the lower 128bits using a blend
+  lane0 = _mm256_blend_ps(
+      lane0, _mm256_castps128_ps256(_mm_permute_ps(
+                 _mm256_castps256_ps128(lane0), _MM_SHUFFLE(1, 0, 1, 0))),
+      15);
+  // then we can perform a consistent permutation on the global register to get
+  // everything in shape:
+  lane0 = _mm256_permute_ps(lane0, _MM_SHUFFLE(3, 3, 2, 2));
+
+  Packet8f lane1 = _mm256_broadcast_ps((const __m128*)(const void*)(from + 4));
+  // mimic an "inplace" permutation of the lower 128bits using a blend
+  lane1 = _mm256_blend_ps(
+      lane1, _mm256_castps128_ps256(_mm_permute_ps(
+                 _mm256_castps256_ps128(lane1), _MM_SHUFFLE(1, 0, 1, 0))),
+      15);
+  // then we can perform a consistent permutation on the global register to get
+  // everything in shape:
+  lane1 = _mm256_permute_ps(lane1, _MM_SHUFFLE(3, 3, 2, 2));
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  Packet16f res = _mm512_undefined_ps();
+  return _mm512_insertf32x8(res, lane0, 0);
+  return _mm512_insertf32x8(res, lane1, 1);
+  return res;
+#else
+  Packet16f res = _mm512_undefined_ps();
+  res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane0, 0), 0);
+  res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane0, 1), 1);
+  res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane1, 0), 2);
+  res = _mm512_insertf32x4(res, _mm256_extractf128_ps(lane1, 1), 3);
+  return res;
+#endif
+}
+// Loads 4 doubles from memory a returns the packet {a0, a0  a1, a1, a2, a2, a3,
+// a3}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploaddup<Packet8d>(const double* from) {
+  Packet4d lane0 = _mm256_broadcast_pd((const __m128d*)(const void*)from);
+  lane0 = _mm256_permute_pd(lane0, 3 << 2);
+
+  Packet4d lane1 = _mm256_broadcast_pd((const __m128d*)(const void*)(from + 2));
+  lane1 = _mm256_permute_pd(lane1, 3 << 2);
+
+  Packet8d res = _mm512_undefined_pd();
+  res = _mm512_insertf64x4(res, lane0, 0);
+  return _mm512_insertf64x4(res, lane1, 1);
+}
+
+// Loads 4 floats from memory a returns the packet
+// {a0, a0  a0, a0, a1, a1, a1, a1, a2, a2, a2, a2, a3, a3, a3, a3}
+template <>
+EIGEN_STRONG_INLINE Packet16f ploadquad<Packet16f>(const float* from) {
+  Packet16f tmp = _mm512_undefined_ps();
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from), 0);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 1), 1);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 2), 2);
+  tmp = _mm512_insertf32x4(tmp, _mm_load_ps1(from + 3), 3);
+  return tmp;
+}
+// Loads 2 doubles from memory a returns the packet
+// {a0, a0  a0, a0, a1, a1, a1, a1}
+template <>
+EIGEN_STRONG_INLINE Packet8d ploadquad<Packet8d>(const double* from) {
+  Packet8d tmp = _mm512_undefined_pd();
+  Packet2d tmp0 = _mm_load_pd1(from);
+  Packet2d tmp1 = _mm_load_pd1(from + 1);
+  Packet4d lane0 = _mm256_broadcastsd_pd(tmp0);
+  Packet4d lane1 = _mm256_broadcastsd_pd(tmp1);
+  tmp = _mm512_insertf64x4(tmp, lane0, 0);
+  return _mm512_insertf64x4(tmp, lane1, 1);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore<float>(float* to, const Packet16f& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_store_ps(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet8d& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_store_pd(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore<int>(int* to, const Packet16i& from) {
+  EIGEN_DEBUG_ALIGNED_STORE _mm512_storeu_si512(reinterpret_cast<__m512i*>(to),
+                                                from);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstoreu<float>(float* to, const Packet16f& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_ps(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet8d& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_pd(to, from);
+}
+template <>
+EIGEN_STRONG_INLINE void pstoreu<int>(int* to, const Packet16i& from) {
+  EIGEN_DEBUG_UNALIGNED_STORE _mm512_storeu_si512(
+      reinterpret_cast<__m512i*>(to), from);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline Packet16f pgather<float, Packet16f>(const float* from,
+                                                             Index stride) {
+  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_multiplier =
+      _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+  Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
+
+  return _mm512_i32gather_ps(indices, from, 4);
+}
+template <>
+EIGEN_DEVICE_FUNC inline Packet8d pgather<double, Packet8d>(const double* from,
+                                                            Index stride) {
+  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+  Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
+
+  return _mm512_i32gather_pd(indices, from, 8);
+}
+
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<float, Packet16f>(float* to,
+                                                         const Packet16f& from,
+                                                         Index stride) {
+  Packet16i stride_vector = _mm512_set1_epi32(stride);
+  Packet16i stride_multiplier =
+      _mm512_set_epi32(15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
+  Packet16i indices = _mm512_mullo_epi32(stride_vector, stride_multiplier);
+  _mm512_i32scatter_ps(to, indices, from, 4);
+}
+template <>
+EIGEN_DEVICE_FUNC inline void pscatter<double, Packet8d>(double* to,
+                                                         const Packet8d& from,
+                                                         Index stride) {
+  Packet8i stride_vector = _mm256_set1_epi32(stride);
+  Packet8i stride_multiplier = _mm256_set_epi32(7, 6, 5, 4, 3, 2, 1, 0);
+  Packet8i indices = _mm256_mullo_epi32(stride_vector, stride_multiplier);
+  _mm512_i32scatter_pd(to, indices, from, 8);
+}
+
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet16f>(float* to, const float& a) {
+  Packet16f pa = pset1<Packet16f>(a);
+  pstore(to, pa);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet8d>(double* to, const double& a) {
+  Packet8d pa = pset1<Packet8d>(a);
+  pstore(to, pa);
+}
+template <>
+EIGEN_STRONG_INLINE void pstore1<Packet16i>(int* to, const int& a) {
+  Packet16i pa = pset1<Packet16i>(a);
+  pstore(to, pa);
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+
+template <>
+EIGEN_STRONG_INLINE float pfirst<Packet16f>(const Packet16f& a) {
+  return _mm_cvtss_f32(_mm512_extractf32x4_ps(a, 0));
+}
+template <>
+EIGEN_STRONG_INLINE double pfirst<Packet8d>(const Packet8d& a) {
+  return _mm_cvtsd_f64(_mm256_extractf128_pd(_mm512_extractf64x4_pd(a, 0), 0));
+}
+template <>
+EIGEN_STRONG_INLINE int pfirst<Packet16i>(const Packet16i& a) {
+  return _mm_extract_epi32(_mm512_extracti32x4_epi32(a, 0), 0);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f preverse(const Packet16f& a)
+{
+  return _mm512_permutexvar_ps(_mm512_set_epi32(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15), a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d preverse(const Packet8d& a)
+{
+  return _mm512_permutexvar_pd(_mm512_set_epi32(0, 0, 0, 1, 0, 2, 0, 3, 0, 4, 0, 5, 0, 6, 0, 7), a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16f pabs(const Packet16f& a)
+{
+  // _mm512_abs_ps intrinsic not found, so hack around it
+  return (__m512)_mm512_and_si512((__m512i)a, _mm512_set1_epi32(0x7fffffff));
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pabs(const Packet8d& a) {
+  // _mm512_abs_ps intrinsic not found, so hack around it
+  return (__m512d)_mm512_and_si512((__m512i)a,
+                                   _mm512_set1_epi64(0x7fffffffffffffff));
+}
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+// AVX512F does not define _mm512_extractf32x8_ps to extract _m256 from _m512
+#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT)                           \
+  __m256 OUTPUT##_0 = _mm512_extractf32x8_ps(INPUT, 0) __m256 OUTPUT##_1 = \
+      _mm512_extractf32x8_ps(INPUT, 1)
+#else
+#define EIGEN_EXTRACT_8f_FROM_16f(INPUT, OUTPUT)                \
+  __m256 OUTPUT##_0 = _mm256_insertf128_ps(                     \
+      _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 0)), \
+      _mm512_extractf32x4_ps(INPUT, 1), 1);                     \
+  __m256 OUTPUT##_1 = _mm256_insertf128_ps(                     \
+      _mm256_castps128_ps256(_mm512_extractf32x4_ps(INPUT, 2)), \
+      _mm512_extractf32x4_ps(INPUT, 3), 1);
+#endif
+
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB) \
+  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTA, 0);        \
+  OUTPUT = _mm512_insertf32x8(OUTPUT, INPUTB, 1);
+#else
+#define EIGEN_INSERT_8f_INTO_16f(OUTPUT, INPUTA, INPUTB)                    \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 0), 0); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTA, 1), 1); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 0), 2); \
+  OUTPUT = _mm512_insertf32x4(OUTPUT, _mm256_extractf128_ps(INPUTB, 1), 3);
+#endif
+template<> EIGEN_STRONG_INLINE Packet16f preduxp<Packet16f>(const Packet16f*
+vecs)
+{
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[0], vecs0);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[1], vecs1);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[2], vecs2);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[3], vecs3);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[4], vecs4);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[5], vecs5);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[6], vecs6);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[7], vecs7);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[8], vecs8);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[9], vecs9);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[10], vecs10);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[11], vecs11);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[12], vecs12);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[13], vecs13);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[14], vecs14);
+  EIGEN_EXTRACT_8f_FROM_16f(vecs[15], vecs15);
+
+  __m256 hsum1 = _mm256_hadd_ps(vecs0_0, vecs1_0);
+  __m256 hsum2 = _mm256_hadd_ps(vecs2_0, vecs3_0);
+  __m256 hsum3 = _mm256_hadd_ps(vecs4_0, vecs5_0);
+  __m256 hsum4 = _mm256_hadd_ps(vecs6_0, vecs7_0);
+
+  __m256 hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  __m256 hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  __m256 hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  __m256 hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  __m256 perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  __m256 perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  __m256 perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  __m256 perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  __m256 sum1 = _mm256_add_ps(perm1, hsum5);
+  __m256 sum2 = _mm256_add_ps(perm2, hsum6);
+  __m256 sum3 = _mm256_add_ps(perm3, hsum7);
+  __m256 sum4 = _mm256_add_ps(perm4, hsum8);
+
+  __m256 blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  __m256 blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  __m256 final = _mm256_blend_ps(blend1, blend2, 0xf0);
+
+  hsum1 = _mm256_hadd_ps(vecs0_1, vecs1_1);
+  hsum2 = _mm256_hadd_ps(vecs2_1, vecs3_1);
+  hsum3 = _mm256_hadd_ps(vecs4_1, vecs5_1);
+  hsum4 = _mm256_hadd_ps(vecs6_1, vecs7_1);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  final = padd(final, _mm256_blend_ps(blend1, blend2, 0xf0));
+
+  hsum1 = _mm256_hadd_ps(vecs8_0, vecs9_0);
+  hsum2 = _mm256_hadd_ps(vecs10_0, vecs11_0);
+  hsum3 = _mm256_hadd_ps(vecs12_0, vecs13_0);
+  hsum4 = _mm256_hadd_ps(vecs14_0, vecs15_0);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  __m256 final_1 = _mm256_blend_ps(blend1, blend2, 0xf0);
+
+  hsum1 = _mm256_hadd_ps(vecs8_1, vecs9_1);
+  hsum2 = _mm256_hadd_ps(vecs10_1, vecs11_1);
+  hsum3 = _mm256_hadd_ps(vecs12_1, vecs13_1);
+  hsum4 = _mm256_hadd_ps(vecs14_1, vecs15_1);
+
+  hsum5 = _mm256_hadd_ps(hsum1, hsum1);
+  hsum6 = _mm256_hadd_ps(hsum2, hsum2);
+  hsum7 = _mm256_hadd_ps(hsum3, hsum3);
+  hsum8 = _mm256_hadd_ps(hsum4, hsum4);
+
+  perm1 = _mm256_permute2f128_ps(hsum5, hsum5, 0x23);
+  perm2 = _mm256_permute2f128_ps(hsum6, hsum6, 0x23);
+  perm3 = _mm256_permute2f128_ps(hsum7, hsum7, 0x23);
+  perm4 = _mm256_permute2f128_ps(hsum8, hsum8, 0x23);
+
+  sum1 = _mm256_add_ps(perm1, hsum5);
+  sum2 = _mm256_add_ps(perm2, hsum6);
+  sum3 = _mm256_add_ps(perm3, hsum7);
+  sum4 = _mm256_add_ps(perm4, hsum8);
+
+  blend1 = _mm256_blend_ps(sum1, sum2, 0xcc);
+  blend2 = _mm256_blend_ps(sum3, sum4, 0xcc);
+
+  final_1 = padd(final_1, _mm256_blend_ps(blend1, blend2, 0xf0));
+
+  __m512 final_output;
+
+  EIGEN_INSERT_8f_INTO_16f(final_output, final, final_1);
+  return final_output;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8d preduxp<Packet8d>(const Packet8d* vecs)
+{
+  Packet4d vecs0_0 = _mm512_extractf64x4_pd(vecs[0], 0);
+  Packet4d vecs0_1 = _mm512_extractf64x4_pd(vecs[0], 1);
+
+  Packet4d vecs1_0 = _mm512_extractf64x4_pd(vecs[1], 0);
+  Packet4d vecs1_1 = _mm512_extractf64x4_pd(vecs[1], 1);
+
+  Packet4d vecs2_0 = _mm512_extractf64x4_pd(vecs[2], 0);
+  Packet4d vecs2_1 = _mm512_extractf64x4_pd(vecs[2], 1);
+
+  Packet4d vecs3_0 = _mm512_extractf64x4_pd(vecs[3], 0);
+  Packet4d vecs3_1 = _mm512_extractf64x4_pd(vecs[3], 1);
+
+  Packet4d vecs4_0 = _mm512_extractf64x4_pd(vecs[4], 0);
+  Packet4d vecs4_1 = _mm512_extractf64x4_pd(vecs[4], 1);
+
+  Packet4d vecs5_0 = _mm512_extractf64x4_pd(vecs[5], 0);
+  Packet4d vecs5_1 = _mm512_extractf64x4_pd(vecs[5], 1);
+
+  Packet4d vecs6_0 = _mm512_extractf64x4_pd(vecs[6], 0);
+  Packet4d vecs6_1 = _mm512_extractf64x4_pd(vecs[6], 1);
+
+  Packet4d vecs7_0 = _mm512_extractf64x4_pd(vecs[7], 0);
+  Packet4d vecs7_1 = _mm512_extractf64x4_pd(vecs[7], 1);
+
+  Packet4d tmp0, tmp1;
+
+  tmp0 = _mm256_hadd_pd(vecs0_0, vecs1_0);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs2_0, vecs3_0);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  __m256d final_0 = _mm256_blend_pd(tmp0, tmp1, 0xC);
+
+  tmp0 = _mm256_hadd_pd(vecs0_1, vecs1_1);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs2_1, vecs3_1);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  final_0 = padd(final_0, _mm256_blend_pd(tmp0, tmp1, 0xC));
+
+  tmp0 = _mm256_hadd_pd(vecs4_0, vecs5_0);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs6_0, vecs7_0);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  __m256d final_1 = _mm256_blend_pd(tmp0, tmp1, 0xC);
+
+  tmp0 = _mm256_hadd_pd(vecs4_1, vecs5_1);
+  tmp0 = _mm256_add_pd(tmp0, _mm256_permute2f128_pd(tmp0, tmp0, 1));
+
+  tmp1 = _mm256_hadd_pd(vecs6_1, vecs7_1);
+  tmp1 = _mm256_add_pd(tmp1, _mm256_permute2f128_pd(tmp1, tmp1, 1));
+
+  final_1 = padd(final_1, _mm256_blend_pd(tmp0, tmp1, 0xC));
+
+  __m512d final_output = _mm512_insertf64x4(final_output, final_0, 0);
+
+  return _mm512_insertf64x4(final_output, final_1, 1);
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux<Packet16f>(const Packet16f& a) {
+  //#ifdef EIGEN_VECTORIZE_AVX512DQ
+#if 0
+  Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
+  Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
+  Packet8f sum = padd(lane0, lane1);
+  Packet8f tmp0 = _mm256_hadd_ps(sum, _mm256_permute2f128_ps(a, a, 1));
+  tmp0 = _mm256_hadd_ps(tmp0, tmp0);
+  return pfirst(_mm256_hadd_ps(tmp0, tmp0));
+#else
+  Packet4f lane0 = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane1 = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane2 = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane3 = _mm512_extractf32x4_ps(a, 3);
+  Packet4f sum = padd(padd(lane0, lane1), padd(lane2, lane3));
+  sum = _mm_hadd_ps(sum, sum);
+  sum = _mm_hadd_ps(sum, _mm_permute_ps(sum, 1));
+  return pfirst(sum);
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE double predux<Packet8d>(const Packet8d& a) {
+  Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
+  Packet4d sum = padd(lane0, lane1);
+  Packet4d tmp0 = _mm256_hadd_pd(sum, _mm256_permute2f128_pd(sum, sum, 1));
+  return pfirst(_mm256_hadd_pd(tmp0, tmp0));
+}
+
+template <>
+EIGEN_STRONG_INLINE Packet8f predux_downto4<Packet16f>(const Packet16f& a) {
+#ifdef EIGEN_VECTORIZE_AVX512DQ
+  Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
+  Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
+  return padd(lane0, lane1);
+#else
+  Packet4f lane0 = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane1 = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane2 = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane3 = _mm512_extractf32x4_ps(a, 3);
+  Packet4f sum0 = padd(lane0, lane2);
+  Packet4f sum1 = padd(lane1, lane3);
+  return _mm256_insertf128_ps(_mm256_castps128_ps256(sum0), sum1, 1);
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE Packet4d predux_downto4<Packet8d>(const Packet8d& a) {
+  Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
+  Packet4d res = padd(lane0, lane1);
+  return res;
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_mul<Packet16f>(const Packet16f& a) {
+//#ifdef EIGEN_VECTORIZE_AVX512DQ
+#if 0
+  Packet8f lane0 = _mm512_extractf32x8_ps(a, 0);
+  Packet8f lane1 = _mm512_extractf32x8_ps(a, 1);
+  Packet8f res = pmul(lane0, lane1);
+  res = pmul(res, _mm256_permute2f128_ps(res, res, 1));
+  res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+#else
+  Packet4f lane0 = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane1 = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane2 = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane3 = _mm512_extractf32x4_ps(a, 3);
+  Packet4f res = pmul(pmul(lane0, lane1), pmul(lane2, lane3));
+  res = pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(pmul(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+#endif
+}
+template <>
+EIGEN_STRONG_INLINE double predux_mul<Packet8d>(const Packet8d& a) {
+  Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
+  Packet4d res = pmul(lane0, lane1);
+  res = pmul(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(pmul(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_min<Packet16f>(const Packet16f& a) {
+  Packet4f lane0 = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane1 = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane2 = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane3 = _mm512_extractf32x4_ps(a, 3);
+  Packet4f res = _mm_min_ps(_mm_min_ps(lane0, lane1), _mm_min_ps(lane2, lane3));
+  res = _mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(_mm_min_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+}
+template <>
+EIGEN_STRONG_INLINE double predux_min<Packet8d>(const Packet8d& a) {
+  Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
+  Packet4d res = _mm256_min_pd(lane0, lane1);
+  res = _mm256_min_pd(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(_mm256_min_pd(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <>
+EIGEN_STRONG_INLINE float predux_max<Packet16f>(const Packet16f& a) {
+  Packet4f lane0 = _mm512_extractf32x4_ps(a, 0);
+  Packet4f lane1 = _mm512_extractf32x4_ps(a, 1);
+  Packet4f lane2 = _mm512_extractf32x4_ps(a, 2);
+  Packet4f lane3 = _mm512_extractf32x4_ps(a, 3);
+  Packet4f res = _mm_max_ps(_mm_max_ps(lane0, lane1), _mm_max_ps(lane2, lane3));
+  res = _mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 3, 2)));
+  return pfirst(_mm_max_ps(res, _mm_permute_ps(res, _MM_SHUFFLE(0, 0, 0, 1))));
+}
+template <>
+EIGEN_STRONG_INLINE double predux_max<Packet8d>(const Packet8d& a) {
+  Packet4d lane0 = _mm512_extractf64x4_pd(a, 0);
+  Packet4d lane1 = _mm512_extractf64x4_pd(a, 1);
+  Packet4d res = _mm256_max_pd(lane0, lane1);
+  res = _mm256_max_pd(res, _mm256_permute2f128_pd(res, res, 1));
+  return pfirst(_mm256_max_pd(res, _mm256_shuffle_pd(res, res, 1)));
+}
+
+template <int Offset>
+struct palign_impl<Offset, Packet16f> {
+  static EIGEN_STRONG_INLINE void run(Packet16f& first,
+                                      const Packet16f& second) {
+    if (Offset != 0) {
+      __m512i first_idx = _mm512_set_epi32(
+          Offset + 15, Offset + 14, Offset + 13, Offset + 12, Offset + 11,
+          Offset + 10, Offset + 9, Offset + 8, Offset + 7, Offset + 6,
+          Offset + 5, Offset + 4, Offset + 3, Offset + 2, Offset + 1, Offset);
+
+      __m512i second_idx =
+          _mm512_set_epi32(Offset - 1, Offset - 2, Offset - 3, Offset - 4,
+                           Offset - 5, Offset - 6, Offset - 7, Offset - 8,
+                           Offset - 9, Offset - 10, Offset - 11, Offset - 12,
+                           Offset - 13, Offset - 14, Offset - 15, Offset - 16);
+
+      unsigned short mask = 0xFFFF;
+      mask <<= (16 - Offset);
+
+      first = _mm512_permutexvar_ps(first_idx, first);
+      Packet16f tmp = _mm512_permutexvar_ps(second_idx, second);
+      first = _mm512_mask_blend_ps(mask, first, tmp);
+    }
+  }
+};
+template <int Offset>
+struct palign_impl<Offset, Packet8d> {
+  static EIGEN_STRONG_INLINE void run(Packet8d& first, const Packet8d& second) {
+    if (Offset != 0) {
+      __m512i first_idx = _mm512_set_epi32(
+          0, Offset + 7, 0, Offset + 6, 0, Offset + 5, 0, Offset + 4, 0,
+          Offset + 3, 0, Offset + 2, 0, Offset + 1, 0, Offset);
+
+      __m512i second_idx = _mm512_set_epi32(
+          0, Offset - 1, 0, Offset - 2, 0, Offset - 3, 0, Offset - 4, 0,
+          Offset - 5, 0, Offset - 6, 0, Offset - 7, 0, Offset - 8);
+
+      unsigned char mask = 0xFF;
+      mask <<= (8 - Offset);
+
+      first = _mm512_permutexvar_pd(first_idx, first);
+      Packet8d tmp = _mm512_permutexvar_pd(second_idx, second);
+      first = _mm512_mask_blend_pd(mask, first, tmp);
+    }
+  }
+};
+
+
+#define PACK_OUTPUT(OUTPUT, INPUT, INDEX, STRIDE) \
+  EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[INDEX], INPUT[INDEX + STRIDE]);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 16>& kernel) {
+  __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T4 = _mm512_unpacklo_ps(kernel.packet[4], kernel.packet[5]);
+  __m512 T5 = _mm512_unpackhi_ps(kernel.packet[4], kernel.packet[5]);
+  __m512 T6 = _mm512_unpacklo_ps(kernel.packet[6], kernel.packet[7]);
+  __m512 T7 = _mm512_unpackhi_ps(kernel.packet[6], kernel.packet[7]);
+  __m512 T8 = _mm512_unpacklo_ps(kernel.packet[8], kernel.packet[9]);
+  __m512 T9 = _mm512_unpackhi_ps(kernel.packet[8], kernel.packet[9]);
+  __m512 T10 = _mm512_unpacklo_ps(kernel.packet[10], kernel.packet[11]);
+  __m512 T11 = _mm512_unpackhi_ps(kernel.packet[10], kernel.packet[11]);
+  __m512 T12 = _mm512_unpacklo_ps(kernel.packet[12], kernel.packet[13]);
+  __m512 T13 = _mm512_unpackhi_ps(kernel.packet[12], kernel.packet[13]);
+  __m512 T14 = _mm512_unpacklo_ps(kernel.packet[14], kernel.packet[15]);
+  __m512 T15 = _mm512_unpackhi_ps(kernel.packet[14], kernel.packet[15]);
+  __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S4 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S5 = _mm512_shuffle_ps(T4, T6, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S6 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S7 = _mm512_shuffle_ps(T5, T7, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S8 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S9 = _mm512_shuffle_ps(T8, T10, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S10 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S11 = _mm512_shuffle_ps(T9, T11, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S12 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S13 = _mm512_shuffle_ps(T12, T14, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S14 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S15 = _mm512_shuffle_ps(T13, T15, _MM_SHUFFLE(3, 2, 3, 2));
+
+  EIGEN_EXTRACT_8f_FROM_16f(S0, S0);
+  EIGEN_EXTRACT_8f_FROM_16f(S1, S1);
+  EIGEN_EXTRACT_8f_FROM_16f(S2, S2);
+  EIGEN_EXTRACT_8f_FROM_16f(S3, S3);
+  EIGEN_EXTRACT_8f_FROM_16f(S4, S4);
+  EIGEN_EXTRACT_8f_FROM_16f(S5, S5);
+  EIGEN_EXTRACT_8f_FROM_16f(S6, S6);
+  EIGEN_EXTRACT_8f_FROM_16f(S7, S7);
+  EIGEN_EXTRACT_8f_FROM_16f(S8, S8);
+  EIGEN_EXTRACT_8f_FROM_16f(S9, S9);
+  EIGEN_EXTRACT_8f_FROM_16f(S10, S10);
+  EIGEN_EXTRACT_8f_FROM_16f(S11, S11);
+  EIGEN_EXTRACT_8f_FROM_16f(S12, S12);
+  EIGEN_EXTRACT_8f_FROM_16f(S13, S13);
+  EIGEN_EXTRACT_8f_FROM_16f(S14, S14);
+  EIGEN_EXTRACT_8f_FROM_16f(S15, S15);
+
+  PacketBlock<Packet8f, 32> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S4_0, 0x20);
+  tmp.packet[1] = _mm256_permute2f128_ps(S1_0, S5_0, 0x20);
+  tmp.packet[2] = _mm256_permute2f128_ps(S2_0, S6_0, 0x20);
+  tmp.packet[3] = _mm256_permute2f128_ps(S3_0, S7_0, 0x20);
+  tmp.packet[4] = _mm256_permute2f128_ps(S0_0, S4_0, 0x31);
+  tmp.packet[5] = _mm256_permute2f128_ps(S1_0, S5_0, 0x31);
+  tmp.packet[6] = _mm256_permute2f128_ps(S2_0, S6_0, 0x31);
+  tmp.packet[7] = _mm256_permute2f128_ps(S3_0, S7_0, 0x31);
+
+  tmp.packet[8] = _mm256_permute2f128_ps(S0_1, S4_1, 0x20);
+  tmp.packet[9] = _mm256_permute2f128_ps(S1_1, S5_1, 0x20);
+  tmp.packet[10] = _mm256_permute2f128_ps(S2_1, S6_1, 0x20);
+  tmp.packet[11] = _mm256_permute2f128_ps(S3_1, S7_1, 0x20);
+  tmp.packet[12] = _mm256_permute2f128_ps(S0_1, S4_1, 0x31);
+  tmp.packet[13] = _mm256_permute2f128_ps(S1_1, S5_1, 0x31);
+  tmp.packet[14] = _mm256_permute2f128_ps(S2_1, S6_1, 0x31);
+  tmp.packet[15] = _mm256_permute2f128_ps(S3_1, S7_1, 0x31);
+
+  // Second set of _m256 outputs
+  tmp.packet[16] = _mm256_permute2f128_ps(S8_0, S12_0, 0x20);
+  tmp.packet[17] = _mm256_permute2f128_ps(S9_0, S13_0, 0x20);
+  tmp.packet[18] = _mm256_permute2f128_ps(S10_0, S14_0, 0x20);
+  tmp.packet[19] = _mm256_permute2f128_ps(S11_0, S15_0, 0x20);
+  tmp.packet[20] = _mm256_permute2f128_ps(S8_0, S12_0, 0x31);
+  tmp.packet[21] = _mm256_permute2f128_ps(S9_0, S13_0, 0x31);
+  tmp.packet[22] = _mm256_permute2f128_ps(S10_0, S14_0, 0x31);
+  tmp.packet[23] = _mm256_permute2f128_ps(S11_0, S15_0, 0x31);
+
+  tmp.packet[24] = _mm256_permute2f128_ps(S8_1, S12_1, 0x20);
+  tmp.packet[25] = _mm256_permute2f128_ps(S9_1, S13_1, 0x20);
+  tmp.packet[26] = _mm256_permute2f128_ps(S10_1, S14_1, 0x20);
+  tmp.packet[27] = _mm256_permute2f128_ps(S11_1, S15_1, 0x20);
+  tmp.packet[28] = _mm256_permute2f128_ps(S8_1, S12_1, 0x31);
+  tmp.packet[29] = _mm256_permute2f128_ps(S9_1, S13_1, 0x31);
+  tmp.packet[30] = _mm256_permute2f128_ps(S10_1, S14_1, 0x31);
+  tmp.packet[31] = _mm256_permute2f128_ps(S11_1, S15_1, 0x31);
+
+  // Pack them into the output
+  PACK_OUTPUT(kernel.packet, tmp.packet, 0, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 1, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 2, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 3, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 4, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 5, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 6, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 7, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 8, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 9, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 10, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 11, 16);
+
+  PACK_OUTPUT(kernel.packet, tmp.packet, 12, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 13, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 14, 16);
+  PACK_OUTPUT(kernel.packet, tmp.packet, 15, 16);
+}
+#define PACK_OUTPUT_2(OUTPUT, INPUT, INDEX, STRIDE)         \
+  EIGEN_INSERT_8f_INTO_16f(OUTPUT[INDEX], INPUT[2 * INDEX], \
+                           INPUT[2 * INDEX + STRIDE]);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16f, 4>& kernel) {
+  __m512 T0 = _mm512_unpacklo_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T1 = _mm512_unpackhi_ps(kernel.packet[0], kernel.packet[1]);
+  __m512 T2 = _mm512_unpacklo_ps(kernel.packet[2], kernel.packet[3]);
+  __m512 T3 = _mm512_unpackhi_ps(kernel.packet[2], kernel.packet[3]);
+
+  __m512 S0 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S1 = _mm512_shuffle_ps(T0, T2, _MM_SHUFFLE(3, 2, 3, 2));
+  __m512 S2 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(1, 0, 1, 0));
+  __m512 S3 = _mm512_shuffle_ps(T1, T3, _MM_SHUFFLE(3, 2, 3, 2));
+
+  EIGEN_EXTRACT_8f_FROM_16f(S0, S0);
+  EIGEN_EXTRACT_8f_FROM_16f(S1, S1);
+  EIGEN_EXTRACT_8f_FROM_16f(S2, S2);
+  EIGEN_EXTRACT_8f_FROM_16f(S3, S3);
+
+  PacketBlock<Packet8f, 8> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_ps(S0_0, S1_0, 0x20);
+  tmp.packet[1] = _mm256_permute2f128_ps(S2_0, S3_0, 0x20);
+  tmp.packet[2] = _mm256_permute2f128_ps(S0_0, S1_0, 0x31);
+  tmp.packet[3] = _mm256_permute2f128_ps(S2_0, S3_0, 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_ps(S0_1, S1_1, 0x20);
+  tmp.packet[5] = _mm256_permute2f128_ps(S2_1, S3_1, 0x20);
+  tmp.packet[6] = _mm256_permute2f128_ps(S0_1, S1_1, 0x31);
+  tmp.packet[7] = _mm256_permute2f128_ps(S2_1, S3_1, 0x31);
+
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 0, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 1, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 2, 1);
+  PACK_OUTPUT_2(kernel.packet, tmp.packet, 3, 1);
+}
+
+#define PACK_OUTPUT_SQ_D(OUTPUT, INPUT, INDEX, STRIDE)                \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX], 0); \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[INDEX + STRIDE], 1);
+
+#define PACK_OUTPUT_D(OUTPUT, INPUT, INDEX, STRIDE)                         \
+  OUTPUT[INDEX] = _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX)], 0); \
+  OUTPUT[INDEX] =                                                           \
+      _mm512_insertf64x4(OUTPUT[INDEX], INPUT[(2 * INDEX) + STRIDE], 1);
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 4>& kernel) {
+  __m512d T0 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0);
+  __m512d T1 = _mm512_shuffle_pd(kernel.packet[0], kernel.packet[1], 0xff);
+  __m512d T2 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0);
+  __m512d T3 = _mm512_shuffle_pd(kernel.packet[2], kernel.packet[3], 0xff);
+
+  PacketBlock<Packet4d, 8> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x20);
+  tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x20);
+  tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x31);
+  tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x20);
+  tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x20);
+  tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x31);
+  tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x31);
+
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 0, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 1, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 2, 1);
+  PACK_OUTPUT_D(kernel.packet, tmp.packet, 3, 1);
+}
+
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8d, 8>& kernel) {
+  __m512d T0 = _mm512_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
+  __m512d T1 = _mm512_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
+  __m512d T2 = _mm512_unpacklo_pd(kernel.packet[2], kernel.packet[3]);
+  __m512d T3 = _mm512_unpackhi_pd(kernel.packet[2], kernel.packet[3]);
+  __m512d T4 = _mm512_unpacklo_pd(kernel.packet[4], kernel.packet[5]);
+  __m512d T5 = _mm512_unpackhi_pd(kernel.packet[4], kernel.packet[5]);
+  __m512d T6 = _mm512_unpacklo_pd(kernel.packet[6], kernel.packet[7]);
+  __m512d T7 = _mm512_unpackhi_pd(kernel.packet[6], kernel.packet[7]);
+
+  PacketBlock<Packet4d, 16> tmp;
+
+  tmp.packet[0] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x20);
+  tmp.packet[1] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x20);
+  tmp.packet[2] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 0),
+                                         _mm512_extractf64x4_pd(T2, 0), 0x31);
+  tmp.packet[3] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 0),
+                                         _mm512_extractf64x4_pd(T3, 0), 0x31);
+
+  tmp.packet[4] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x20);
+  tmp.packet[5] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x20);
+  tmp.packet[6] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T0, 1),
+                                         _mm512_extractf64x4_pd(T2, 1), 0x31);
+  tmp.packet[7] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T1, 1),
+                                         _mm512_extractf64x4_pd(T3, 1), 0x31);
+
+  tmp.packet[8] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0),
+                                         _mm512_extractf64x4_pd(T6, 0), 0x20);
+  tmp.packet[9] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0),
+                                         _mm512_extractf64x4_pd(T7, 0), 0x20);
+  tmp.packet[10] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 0),
+                                          _mm512_extractf64x4_pd(T6, 0), 0x31);
+  tmp.packet[11] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 0),
+                                          _mm512_extractf64x4_pd(T7, 0), 0x31);
+
+  tmp.packet[12] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1),
+                                          _mm512_extractf64x4_pd(T6, 1), 0x20);
+  tmp.packet[13] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1),
+                                          _mm512_extractf64x4_pd(T7, 1), 0x20);
+  tmp.packet[14] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T4, 1),
+                                          _mm512_extractf64x4_pd(T6, 1), 0x31);
+  tmp.packet[15] = _mm256_permute2f128_pd(_mm512_extractf64x4_pd(T5, 1),
+                                          _mm512_extractf64x4_pd(T7, 1), 0x31);
+
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 0, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 1, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 2, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 3, 8);
+
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 4, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 5, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 6, 8);
+  PACK_OUTPUT_SQ_D(kernel.packet, tmp.packet, 7, 8);
+}
+template <>
+EIGEN_STRONG_INLINE Packet16f pblend(const Selector<16>& /*ifPacket*/,
+                                     const Packet16f& /*thenPacket*/,
+                                     const Packet16f& /*elsePacket*/) {
+  assert(false && "To be implemented");
+  return Packet16f();
+}
+template <>
+EIGEN_STRONG_INLINE Packet8d pblend(const Selector<8>& /*ifPacket*/,
+                                    const Packet8d& /*thenPacket*/,
+                                    const Packet8d& /*elsePacket*/) {
+  assert(false && "To be implemented");
+  return Packet8d();
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_AVX512_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h
new file mode 100644
index 0000000..67db2f8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/Complex.h
@@ -0,0 +1,461 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010-2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX32_ALTIVEC_H
+#define EIGEN_COMPLEX32_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static Packet4ui  p4ui_CONJ_XOR = vec_mergeh((Packet4ui)p4i_ZERO, (Packet4ui)p4f_MZERO);//{ 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+#ifdef __VSX__
+#if defined(_BIG_ENDIAN)
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#else
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_MZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2d_MZERO, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+#endif
+#endif
+
+//---------- float ----------
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE explicit Packet2cf() : v(p4f_ZERO) {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  Packet4f  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+#ifdef __VSX__
+    HasBlend  = 1,
+#endif
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  if((std::ptrdiff_t(&from) % 16) == 0)
+    res.v = pload<Packet4f>((const float *)&from);
+  else
+    res.v = ploadu<Packet4f>((const float *)&from);
+  res.v = vec_perm(res.v, res.v, p16uc_PSET64_HI);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>*        from) { return Packet2cf(pload<Packet4f>((const float *) from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>*       from) { return Packet2cf(ploadu<Packet4f>((const float*) from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*     from) { return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { pstoreu((float*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(a.v - b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a) { return Packet2cf(pxor<Packet4f>(a.v, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR))); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet4f v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  v1 = vec_perm(a.v, a.v, p16uc_PSET32_WODD);
+  // Get the imaginary parts of a
+  v2 = vec_perm(a.v, a.v, p16uc_PSET32_WEVEN);
+  // multiply a_re * b 
+  v1 = vec_madd(v1, b.v, p4f_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(v2, b.v, p4f_ZERO);
+  v2 = reinterpret_cast<Packet4f>(pxor(v2, reinterpret_cast<Packet4f>(p4ui_CONJ_XOR)));
+  // permute back to a proper order
+  v2 = vec_perm(v2, v2, p16uc_COMPLEX32_REV);
+  
+  return Packet2cf(padd<Packet4f>(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v, b.v)); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> * addr)    { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore((float *)&res, a.v);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet4f rev_a;
+  rev_a = vec_perm(a.v, a.v, p16uc_COMPLEX32_REV2);
+  return Packet2cf(rev_a);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  b = vec_sld(a.v, a.v, 8);
+  b = padd<Packet4f>(a.v, b);
+  return pfirst<Packet2cf>(Packet2cf(b));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  Packet4f b1, b2;
+#ifdef _BIG_ENDIAN  
+  b1 = vec_sld(vecs[0].v, vecs[1].v, 8);
+  b2 = vec_sld(vecs[1].v, vecs[0].v, 8);
+#else
+  b1 = vec_sld(vecs[1].v, vecs[0].v, 8);
+  b2 = vec_sld(vecs[0].v, vecs[1].v, 8);
+#endif
+  b2 = vec_sld(b2, b2, 8);
+  b2 = padd<Packet4f>(b1, b2);
+
+  return Packet2cf(b2);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  Packet4f b;
+  Packet2cf prod;
+  b = vec_sld(a.v, a.v, 8);
+  prod = pmul<Packet2cf>(a, Packet2cf(b));
+
+  return pfirst<Packet2cf>(prod);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset==1)
+    {
+#ifdef _BIG_ENDIAN
+      first.v = vec_sld(first.v, second.v, 8);
+#else
+      first.v = vec_sld(second.v, first.v, 8);
+#endif
+    }
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet4f, Packet2cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
+  { return Packet2cf(internal::pmul<Packet4f>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cf, Packet4f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
+  { return Packet2cf(internal::pmul<Packet4f>(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a, b);
+  Packet4f s = pmul<Packet4f>(b.v, b.v);
+  return Packet2cf(pdiv(res.v, padd<Packet4f>(s, vec_perm(s, s, p16uc_COMPLEX32_REV))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& x)
+{
+  return Packet2cf(vec_perm(x.v, x.v, p16uc_COMPLEX32_REV));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet4f tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+#ifdef __VSX__
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  result.v = reinterpret_cast<Packet4f>(pblend<Packet2d>(ifPacket, reinterpret_cast<Packet2d>(thenPacket.v), reinterpret_cast<Packet2d>(elsePacket.v)));
+  return result;
+}
+#endif
+
+//---------- double ----------
+#ifdef __VSX__
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet1cd>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride)
+{
+  std::complex<double> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<double> >(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(pxor(a.v, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR2))); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d a_re, a_im, v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
+  // Get the imaginary parts of a
+  a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
+  // multiply a_re * b
+  v1 = vec_madd(a_re, b.v, p2d_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(a_im, b.v, p2d_ZERO);
+  v2 = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v2), reinterpret_cast<Packet4ui>(v2), 8));
+  v2 = pxor(v2, reinterpret_cast<Packet2d>(p2ul_CONJ_XOR1));
+
+  return Packet1cd(padd<Packet2d>(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pand(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(por(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pxor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(pandnot(a.v, b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from)  { return pset1<Packet1cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> * addr)    { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<double> >(res, a);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)        { return vecs[0]; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+template<> struct conj_helper<Packet2d, Packet1cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
+  { return Packet1cd(internal::pmul<Packet2d>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet1cd, Packet2d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
+  { return Packet1cd(internal::pmul<Packet2d>(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = pmul<Packet2d>(b.v, b.v);
+  return Packet1cd(pdiv(res.v, padd<Packet2d>(s, vec_perm(s, s, p16uc_REVERSE64))));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+#endif // __VSX__
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX32_ALTIVEC_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h
new file mode 100644
index 0000000..c5e4bed
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/MathFunctions.h
@@ -0,0 +1,322 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+static _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+static _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+static _EIGEN_DECLARE_CONST_Packet4i(23, 23);
+
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+/* the smallest non denormalized float number */
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000); // -1.f/0.f
+static _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_nan,     0xffffffff);
+  
+/* natural logarithm computed for 4 simultaneous float
+  return NaN for x <= 0
+*/
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+static _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+static _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+#ifdef __VSX__
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+#ifdef __POWER8_VECTOR__
+static Packet2l p2l_1023 = { 1023, 1023 };
+static Packet2ul p2ul_52 = { 52, 52 };
+#endif
+
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f plog<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+
+  Packet4i emm0;
+
+  /* isvalid_mask is 0 if x < 0 or x is NaN. */
+  Packet4ui isvalid_mask = reinterpret_cast<Packet4ui>(vec_cmpge(x, p4f_ZERO));
+  Packet4ui iszero_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(x, p4f_ZERO));
+
+  x = pmax(x, p4f_min_norm_pos);  /* cut off denormalized stuff */
+  emm0 = vec_sr(reinterpret_cast<Packet4i>(x),
+                reinterpret_cast<Packet4ui>(p4i_23));
+
+  /* keep only the fractional part */
+  x = pand(x, p4f_inv_mant_mask);
+  x = por(x, p4f_half);
+
+  emm0 = psub(emm0, p4i_0x7f);
+  Packet4f e = padd(vec_ctf(emm0, 0), p4f_1);
+
+  /* part2:
+     if( x < SQRTHF ) {
+       e -= 1;
+       x = x + x - 1.0;
+     } else { x = x - 1.0; }
+  */
+  Packet4f mask = reinterpret_cast<Packet4f>(vec_cmplt(x, p4f_cephes_SQRTHF));
+  Packet4f tmp = pand(x, mask);
+  x = psub(x, p4f_1);
+  e = psub(e, pand(p4f_1, mask));
+  x = padd(x, tmp);
+
+  Packet4f x2 = pmul(x,x);
+  Packet4f x3 = pmul(x2,x);
+
+  Packet4f y, y1, y2;
+  y  = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
+  y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
+  y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
+  y  = pmadd(y , x, p4f_cephes_log_p2);
+  y1 = pmadd(y1, x, p4f_cephes_log_p5);
+  y2 = pmadd(y2, x, p4f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  y1 = pmul(e, p4f_cephes_log_q1);
+  tmp = pmul(x2, p4f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p4f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+  // negative arg will be NAN, 0 will be -INF
+  x = vec_sel(x, p4f_minus_inf, iszero_mask);
+  x = vec_sel(p4f_minus_nan, x, isvalid_mask);
+  return x;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+
+  Packet4f tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
+
+  // express exp(x) as exp(g + n*log(2))
+  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
+
+  fx = pfloor(fx);
+
+  tmp = pmul(fx, p4f_cephes_exp_C1);
+  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  z = pmul(x,x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // build 2^n
+  emm0 = vec_cts(fx, 0);
+  emm0 = vec_add(emm0, p4i_0x7f);
+  emm0 = vec_sl(emm0, reinterpret_cast<Packet4ui>(p4i_23));
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet4ui isnumber_mask = reinterpret_cast<Packet4ui>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(y, reinterpret_cast<Packet4f>(emm0)), _x),
+                 isnumber_mask);
+}
+
+#ifndef EIGEN_COMP_CLANG
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x)
+{
+  return  vec_rsqrt(x);
+}
+#endif
+
+#ifdef __VSX__
+#ifndef EIGEN_COMP_CLANG
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x)
+{
+  return  vec_rsqrt(x);
+}
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  return  vec_sqrt(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x)
+{
+  return  vec_sqrt(x);
+}
+
+// VSX support varies between different compilers and even different
+// versions of the same compiler.  For gcc version >= 4.9.3, we can use
+// vec_cts to efficiently convert Packet2d to Packet2l.  Otherwise, use
+// a slow version that works with older compilers. 
+// Update: apparently vec_cts/vec_ctf intrinsics for 64-bit doubles
+// are buggy, https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70963
+static inline Packet2l ConvertToPacket2l(const Packet2d& x) {
+#if EIGEN_GNUC_AT_LEAST(5, 4) || \
+    (EIGEN_GNUC_AT(6, 1) && __GNUC_PATCHLEVEL__ >= 1)
+  return vec_cts(x, 0);    // TODO: check clang version.
+#else
+  double tmp[2];
+  memcpy(tmp, &x, sizeof(tmp));
+  Packet2l l = { static_cast<long long>(tmp[0]),
+                 static_cast<long long>(tmp[1]) };
+  return l;
+#endif
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  Packet2d tmp, fx;
+  Packet2l emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(x, p2d_cephes_LOG2EF, p2d_half);
+
+  fx = pfloor(fx);
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = ConvertToPacket2l(fx);
+
+#ifdef __POWER8_VECTOR__ 
+  emm0 = vec_add(emm0, p2l_1023);
+  emm0 = vec_sl(emm0, p2ul_52);
+#else
+  // Code is a bit complex for POWER7.  There is actually a
+  // vec_xxsldi intrinsic but it is not supported by some gcc versions.
+  // So we shift (52-32) bits and do a word swap with zeros.
+  _EIGEN_DECLARE_CONST_Packet4i(1023, 1023);
+  _EIGEN_DECLARE_CONST_Packet4i(20, 20);    // 52 - 32
+
+  Packet4i emm04i = reinterpret_cast<Packet4i>(emm0);
+  emm04i = vec_add(emm04i, p4i_1023);
+  emm04i = vec_sl(emm04i, reinterpret_cast<Packet4ui>(p4i_20));
+  static const Packet16uc perm = {
+    0x14, 0x15, 0x16, 0x17, 0x00, 0x01, 0x02, 0x03, 
+    0x1c, 0x1d, 0x1e, 0x1f, 0x08, 0x09, 0x0a, 0x0b };
+#ifdef  _BIG_ENDIAN
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(p4i_ZERO, emm04i, perm));
+#else
+  emm0 = reinterpret_cast<Packet2l>(vec_perm(emm04i, p4i_ZERO, perm));
+#endif
+
+#endif
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x),
+                 isnumber_mask);
+}
+#endif
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_ALTIVEC_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h
new file mode 100755
index 0000000..b3f1ea1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/AltiVec/PacketMath.h
@@ -0,0 +1,1033 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_ALTIVEC_H
+#define EIGEN_PACKET_MATH_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
+// NOTE Altivec has 32 registers, but Eigen only accepts a value of 8 or 16
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  32
+#endif
+
+typedef __vector float          Packet4f;
+typedef __vector int            Packet4i;
+typedef __vector unsigned int   Packet4ui;
+typedef __vector __bool int     Packet4bi;
+typedef __vector short int      Packet8i;
+typedef __vector unsigned char  Packet16uc;
+
+// We don't want to write the same code all the time, but we need to reuse the constants
+// and it doesn't really work to declare them global, so we define macros instead
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4f(NAME,X) \
+  Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(vec_splat_s32(X))
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = vec_splat_s32(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+  Packet4f p4f_##NAME = pset1<Packet4f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = pset1<Packet4i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = pset1<Packet2l>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = reinterpret_cast<Packet4f>(pset1<Packet4i>(X))
+
+#define DST_CHAN 1
+#define DST_CTRL(size, count, stride) (((size) << 24) | ((count) << 16) | (stride))
+
+
+// These constants are endian-agnostic
+static _EIGEN_DECLARE_CONST_FAST_Packet4f(ZERO, 0); //{ 0.0, 0.0, 0.0, 0.0}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE,1); //{ 1, 1, 1, 1}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS16,-16); //{ -16, -16, -16, -16}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(MINUS1,-1); //{ -1, -1, -1, -1}
+static Packet4f p4f_MZERO = (Packet4f) vec_sl((Packet4ui)p4i_MINUS1, (Packet4ui)p4i_MINUS1); //{ 0x80000000, 0x80000000, 0x80000000, 0x80000000}
+#ifndef __VSX__
+static Packet4f p4f_ONE = vec_ctf(p4i_ONE, 0); //{ 1.0, 1.0, 1.0, 1.0}
+#endif
+
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
+static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+
+static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
+static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
+
+// Mask alignment
+#ifdef __PPC64__
+#define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
+#else
+#define _EIGEN_MASK_ALIGNMENT	0xfffffff0
+#endif
+
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+
+// Handle endianness properly while loading constants
+// Define global static constants:
+#ifdef _BIG_ENDIAN
+static Packet16uc p16uc_FORWARD = vec_lvsl(0, (float*)0);
+#ifdef __VSX__
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#endif
+static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#else
+static Packet16uc p16uc_FORWARD = p16uc_REVERSE32; 
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET32_WODD = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 1), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+static Packet16uc p16uc_HALF64_0_16 = vec_sld(vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 0), (Packet16uc)p4i_ZERO, 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+#endif // _BIG_ENDIAN
+
+static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+static Packet16uc p16uc_TRANSPOSE64_HI = p16uc_PSET64_HI + p16uc_HALF64_0_16;                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = p16uc_PSET64_LO + p16uc_HALF64_0_16;                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+
+static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+
+#ifdef _BIG_ENDIAN
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#else
+static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_PSET64_HI, p16uc_PSET64_LO, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+#endif // _BIG_ENDIAN
+
+#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+  #define EIGEN_PPC_PREFETCH(ADDR) __builtin_prefetch(ADDR);
+#else
+  #define EIGEN_PPC_PREFETCH(ADDR) asm( "   dcbt [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#endif
+
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+#ifdef __VSX__
+    HasSqrt = 1,
+#if !EIGEN_COMP_CLANG
+    HasRsqrt = 1,
+#else
+    HasRsqrt = 0,
+#endif
+#else
+    HasSqrt = 0,
+    HasRsqrt = 0,
+#endif
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet4i type;
+  typedef Packet4i half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 0,
+    HasBlend = 1
+  };
+};
+
+
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+inline std::ostream & operator <<(std::ostream & s, const Packet16uc & v)
+{
+  union {
+    Packet16uc   v;
+    unsigned char n[16];
+  } vt;
+  vt.v = v;
+  for (int i=0; i< 16; i++)
+    s << (int)vt.n[i] << ", ";
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4f & v)
+{
+  union {
+    Packet4f   v;
+    float n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
+{
+  union {
+    Packet4i   v;
+    int n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
+{
+  union {
+    Packet4ui   v;
+    unsigned int n[4];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1] << ", " << vt.n[2] << ", " << vt.n[3];
+  return s;
+}
+
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) {
+  Packet4f v = {from, from, from, from};
+  return v;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)   {
+  Packet4i v = {from, from, from, from};
+  return v;
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4i>(const int *a,
+                      Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
+{
+  a3 = pload<Packet4i>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 af[4];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  af[2] = from[2*stride];
+  af[3] = from[3*stride];
+ return pload<Packet4f>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4i>(ai);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 af[4];
+  pstore<float>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+  to[2*stride] = af[2];
+  to[3*stride] = af[3];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  pstore<int>((int *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return pset1<Packet4f>(a) + p4f_COUNTDOWN; }
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)   { return pset1<Packet4i>(a) + p4i_COUNTDOWN; }
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return a + b; }
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return a + b; }
+
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return a - b; }
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return a - b; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return p4f_ZERO - a; }
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return p4i_ZERO - a; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_madd(a,b, p4f_MZERO); }
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return a * b; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+#ifndef __VSX__  // VSX actually provides a div instruction
+  Packet4f t, y_0, y_1;
+
+  // Altivec does not offer a divide instruction, we have to do a reciprocal approximation
+  y_0 = vec_re(b);
+
+  // Do one Newton-Raphson iteration to get the needed accuracy
+  t   = vec_nmsub(y_0, b, p4f_ONE);
+  y_1 = vec_madd(y_0, t, y_0);
+
+  return vec_madd(a, y_1, p4f_MZERO);
+#else
+  return vec_div(a, b);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
+{ eigen_assert(false && "packet integer division are not supported by AltiVec");
+  return pset1<Packet4i>(0);
+}
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vec_madd(a,b,c); }
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return a*b + c; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, vec_nor(b, b)); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return vec_floor(a); }
+
+#ifdef _BIG_ENDIAN
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet16uc MSQ, LSQ;
+  Packet16uc mask;
+  MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
+  mask = vec_lvsl(0, from);                        // create the permute mask
+  return (Packet4f) vec_perm(MSQ, LSQ, mask);           // align the data
+
+}
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
+  Packet16uc MSQ, LSQ;
+  Packet16uc mask;
+  MSQ = vec_ld(0, (unsigned char *)from);          // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)from);         // least significant quadword
+  mask = vec_lvsl(0, from);                        // create the permute mask
+  return (Packet4i) vec_perm(MSQ, LSQ, mask);    // align the data
+}
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return (Packet4i) vec_vsx_ld((long)from & 15, (const int*) _EIGEN_ALIGNED_PTR(from));
+}
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return (Packet4f) vec_vsx_ld((long)from & 15, (const float*) _EIGEN_ALIGNED_PTR(from));
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
+{
+  Packet4f p;
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4f>(from);
+  else                                  p = ploadu<Packet4f>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
+}
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+{
+  Packet4i p;
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet4i>(from);
+  else                                  p = ploadu<Packet4i>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
+}
+
+#ifdef _BIG_ENDIAN
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from)
+{
+  EIGEN_DEBUG_UNALIGNED_STORE
+  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
+  // Warning: not thread safe!
+  Packet16uc MSQ, LSQ, edges;
+  Packet16uc edgeAlign, align;
+
+  MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
+  edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
+  edges=vec_perm(LSQ,MSQ,edgeAlign);                        // extract the edges
+  align = vec_lvsr( 0, to );                                // permute map to misalign data
+  MSQ = vec_perm(edges,(Packet16uc)from,align);             // misalign the data (MSQ)
+  LSQ = vec_perm((Packet16uc)from,edges,align);             // misalign the data (LSQ)
+  vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
+  vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
+}
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from)
+{
+  EIGEN_DEBUG_UNALIGNED_STORE
+  // Taken from http://developer.apple.com/hardwaredrivers/ve/alignment.html
+  // Warning: not thread safe!
+  Packet16uc MSQ, LSQ, edges;
+  Packet16uc edgeAlign, align;
+
+  MSQ = vec_ld(0, (unsigned char *)to);                     // most significant quadword
+  LSQ = vec_ld(15, (unsigned char *)to);                    // least significant quadword
+  edgeAlign = vec_lvsl(0, to);                              // permute map to extract edges
+  edges=vec_perm(LSQ, MSQ, edgeAlign);                      // extract the edges
+  align = vec_lvsr( 0, to );                                // permute map to misalign data
+  MSQ = vec_perm(edges, (Packet16uc) from, align);          // misalign the data (MSQ)
+  LSQ = vec_perm((Packet16uc) from, edges, align);          // misalign the data (LSQ)
+  vec_st( LSQ, 15, (unsigned char *)to );                   // Store the LSQ part first
+  vec_st( MSQ, 0, (unsigned char *)to );                    // Store the MSQ part
+}
+#else
+// We also need ot redefine little endian loading of Packet4i/Packet4f using VSX
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (int*) _EIGEN_ALIGNED_PTR(to));
+}
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st(from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr)    { EIGEN_PPC_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr)    { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int   EIGEN_ALIGN16 x; vec_ste(a, 0, &x); return x; }
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  return reinterpret_cast<Packet4f>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
+}
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
+{
+  return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32)); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vec_abs(a); }
+
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, sum;
+  b   = vec_sld(a, a, 8);
+  sum = a + b;
+  b   = vec_sld(sum, sum, 4);
+  sum += b;
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  Packet4f v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = sum[0] + sum[1];
+  // Lines 2+3
+  sum[1] = sum[2] + sum[3];
+  // Add the results
+  sum[0] = sum[0] + sum[1];
+
+  return sum[0];
+}
+
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i sum;
+  sum = vec_sums(a, p4i_ZERO);
+#ifdef _BIG_ENDIAN
+  sum = vec_sld(sum, p4i_ZERO, 12);
+#else
+  sum = vec_sld(p4i_ZERO, sum, 4);
+#endif
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  Packet4i v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = sum[0] + sum[1];
+  // Lines 2+3
+  sum[1] = sum[2] + sum[3];
+  // Add the results
+  sum[0] = sum[0] + sum[1];
+
+  return sum[0];
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  Packet4f prod;
+  prod = pmul(a, vec_sld(a, a, 8));
+  return pfirst(pmul(prod, vec_sld(prod, prod, 4)));
+}
+
+template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+{
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  return aux[0] * aux[1] * aux[2] * aux[3];
+}
+
+// min
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b = vec_min(a, vec_sld(a, a, 8));
+  res = vec_min(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = vec_min(a, vec_sld(a, a, 8));
+  res = vec_min(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet4f b, res;
+  b = vec_max(a, vec_sld(a, a, 8));
+  res = vec_max(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = vec_max(a, vec_sld(a, a, 8));
+  res = vec_max(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet4i>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
+  {
+#ifdef _BIG_ENDIAN
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+#else
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(second, first, 12); break;
+    case 2:
+      first = vec_sld(second, first, 8); break;
+    case 3:
+      first = vec_sld(second, first, 4); break;
+    }
+#endif
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  Packet4f t0, t1, t2, t3;
+  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  Packet4i t0, t1, t2, t3;
+  t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = reinterpret_cast<Packet4ui>(vec_cmpeq(reinterpret_cast<Packet4ui>(select), reinterpret_cast<Packet4ui>(p4i_ONE)));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+
+//---------- double ----------
+#ifdef __VSX__
+typedef __vector double              Packet2d;
+typedef __vector unsigned long long  Packet2ul;
+typedef __vector long long           Packet2l;
+#if EIGEN_COMP_CLANG
+typedef Packet2ul                    Packet2bl;
+#else
+typedef __vector __bool long         Packet2bl;
+#endif
+
+static Packet2l  p2l_ONE  = { 1, 1 };
+static Packet2l  p2l_ZERO = reinterpret_cast<Packet2l>(p4i_ZERO);
+static Packet2d  p2d_ONE  = { 1.0, 1.0 }; 
+static Packet2d  p2d_ZERO = reinterpret_cast<Packet2d>(p4f_ZERO);
+static Packet2d  p2d_MZERO = { -0.0, -0.0 };
+
+#ifdef _BIG_ENDIAN
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ZERO), reinterpret_cast<Packet4f>(p2d_ONE), 8));
+#else
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(p2d_ONE), reinterpret_cast<Packet4f>(p2d_ZERO), 8));
+#endif
+
+template<int index> Packet2d vec_splat_dbl(Packet2d& a);
+
+template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<0>(Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_HI));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d vec_splat_dbl<1>(Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(a, a, p16uc_PSET64_LO));
+}
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2l & v)
+{
+  union {
+    Packet2l   v;
+    int64_t n[2];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
+{
+  union {
+    Packet2d   v;
+    double n[2];
+  } vt;
+  vt.v = v;
+  s << vt.n[0] << ", " << vt.n[1];
+  return s;
+}
+
+// Need to define them first or we get specialization after instantiation errors
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+#ifdef __VSX__
+  return vec_vsx_ld(0, from);
+#else
+  return vec_ld(0, from);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+#ifdef __VSX__
+  vec_vsx_st(from, 0, to);
+#else
+  vec_st(from, 0, to);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) {
+  Packet2d v = {from, from};
+  return v;
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+  a1 = pload<Packet2d>(a);
+  a0 = vec_splat_dbl<0>(a1);
+  a1 = vec_splat_dbl<1>(a1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec_splat_dbl<0>(a3);
+  a3 = vec_splat_dbl<1>(a3);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+ return pload<Packet2d>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  pstore<double>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return pset1<Packet2d>(a) + p2d_COUNTDOWN; }
+
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return a + b; }
+
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return a - b; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return p2d_ZERO - a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_madd(a,b,p2d_MZERO); }
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_div(a,b); }
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_ALIGNED_LOAD
+  return (Packet2d) vec_vsx_ld((long)from & 15, (const double*) _EIGEN_ALIGNED_PTR(from));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  Packet2d p;
+  if((std::ptrdiff_t(from) % 16) == 0)  p = pload<Packet2d>(from);
+  else                                  p = ploadu<Packet2d>(from);
+  return vec_splat_dbl<0>(p);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from)
+{
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_vsx_st((Packet4f)from, (long)to & 15, (float*) _EIGEN_ALIGNED_PTR(to));
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_PPC_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE double  pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore<double>(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
+}
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vec_abs(a); }
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  Packet2d b, sum;
+  b   = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(a), reinterpret_cast<Packet4f>(a), 8));
+  sum = a + b;
+  return pfirst<Packet2d>(sum);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  Packet2d v[2], sum;
+  v[0] = vecs[0] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[0]), reinterpret_cast<Packet4f>(vecs[0]), 8));
+  v[1] = vecs[1] + reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(vecs[1]), reinterpret_cast<Packet4f>(vecs[1]), 8));
+ 
+#ifdef _BIG_ENDIAN
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[0]), reinterpret_cast<Packet4f>(v[1]), 8));
+#else
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4f>(v[1]), reinterpret_cast<Packet4f>(v[0]), 8));
+#endif
+
+  return sum;
+}
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmin(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmax(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(a), reinterpret_cast<Packet4ui>(a), 8))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset == 1)
+#ifdef _BIG_ENDIAN
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(first), reinterpret_cast<Packet4ui>(second), 8));
+#else
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(second), reinterpret_cast<Packet4ui>(first), 8));
+#endif
+  }
+};
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  Packet2d t0, t1;
+  t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
+  t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
+  kernel.packet[0] = t0;
+  kernel.packet[1] = t1;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  Packet2l select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2bl mask = vec_cmpeq(reinterpret_cast<Packet2d>(select), reinterpret_cast<Packet2d>(p2l_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+#endif // __VSX__
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_ALTIVEC_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h
new file mode 100644
index 0000000..9c25365
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Complex.h
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_CUDA_H
+#define EIGEN_COMPLEX_CUDA_H
+
+// clang-format off
+
+namespace Eigen {
+
+namespace internal {
+
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+
+// Many std::complex methods such as operator+, operator-, operator* and
+// operator/ are not constexpr. Due to this, clang does not treat them as device
+// functions and thus Eigen functors making use of these operators fail to
+// compile. Here, we manually specialize these functors for complex types when
+// building for CUDA to avoid non-constexpr methods.
+
+// Sum
+template<typename T> struct scalar_sum_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) + numext::real(b),
+                           numext::imag(a) + numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_sum_op<std::complex<T>, std::complex<T> > : scalar_sum_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Difference
+template<typename T> struct scalar_difference_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    return std::complex<T>(numext::real(a) - numext::real(b),
+                           numext::imag(a) - numext::imag(b));
+  }
+};
+
+template<typename T> struct scalar_difference_op<std::complex<T>, std::complex<T> > : scalar_difference_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Product
+template<typename T> struct scalar_product_op<const std::complex<T>, const std::complex<T> >  : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasMul
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    return std::complex<T>(a_real * b_real - a_imag * b_imag,
+                           a_real * b_imag + a_imag * b_real);
+  }
+};
+
+template<typename T> struct scalar_product_op<std::complex<T>, std::complex<T> > : scalar_product_op<const std::complex<T>, const std::complex<T> > {};
+
+
+// Quotient
+template<typename T> struct scalar_quotient_op<const std::complex<T>, const std::complex<T> > : binary_op_base<const std::complex<T>, const std::complex<T> > {
+  enum {
+    Vectorizable = packet_traits<std::complex<T>>::HasDiv
+  };
+  typedef typename std::complex<T> result_type;
+
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::complex<T> operator() (const std::complex<T>& a, const std::complex<T>& b) const {
+    const T a_real = numext::real(a);
+    const T a_imag = numext::imag(a);
+    const T b_real = numext::real(b);
+    const T b_imag = numext::imag(b);
+    const T norm = T(1) / (b_real * b_real + b_imag * b_imag);
+    return std::complex<T>((a_real * b_real + a_imag * b_imag) * norm,
+                           (a_imag * b_real - a_real * b_imag) * norm);
+  }
+};
+
+template<typename T> struct scalar_quotient_op<std::complex<T>, std::complex<T> > : scalar_quotient_op<const std::complex<T>, const std::complex<T> > {};
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h
new file mode 100644
index 0000000..294c517
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/Half.h
@@ -0,0 +1,635 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+//
+// The conversion routines are Copyright (c) Fabian Giesen, 2016.
+// The original license follows:
+//
+// Copyright (c) Fabian Giesen, 2016
+// All rights reserved.
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted.
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+
+// Standard 16-bit float type, mostly useful for GPUs. Defines a new
+// type Eigen::half (inheriting from CUDA's __half struct) with
+// operator overloads such that it behaves basically as an arithmetic
+// type. It will be quite slow on CPUs (so it is recommended to stay
+// in fp32 for CPUs, except for simple parameter conversions, I/O
+// to disk and the likes), but fast on GPUs.
+
+
+#ifndef EIGEN_HALF_CUDA_H
+#define EIGEN_HALF_CUDA_H
+
+#if __cplusplus > 199711L
+#define EIGEN_EXPLICIT_CAST(tgt_type) explicit operator tgt_type()
+#else
+#define EIGEN_EXPLICIT_CAST(tgt_type) operator tgt_type()
+#endif
+
+
+namespace Eigen {
+
+struct half;
+
+namespace half_impl {
+
+#if !defined(EIGEN_HAS_CUDA_FP16)
+
+// Make our own __half definition that is similar to CUDA's.
+struct __half {
+  EIGEN_DEVICE_FUNC __half() {}
+  explicit EIGEN_DEVICE_FUNC __half(unsigned short raw) : x(raw) {}
+  unsigned short x;
+};
+
+#endif
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff);
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h);
+
+struct half_base : public __half {
+  EIGEN_DEVICE_FUNC half_base() {}
+  EIGEN_DEVICE_FUNC half_base(const half_base& h) : __half(h) {}
+  EIGEN_DEVICE_FUNC half_base(const __half& h) : __half(h) {}
+};
+
+} // namespace half_impl
+
+// Class definition.
+struct half : public half_impl::half_base {
+  #if !defined(EIGEN_HAS_CUDA_FP16)
+    typedef half_impl::__half __half;
+  #endif
+
+  EIGEN_DEVICE_FUNC half() {}
+
+  EIGEN_DEVICE_FUNC half(const __half& h) : half_impl::half_base(h) {}
+  EIGEN_DEVICE_FUNC half(const half& h) : half_impl::half_base(h) {}
+
+  explicit EIGEN_DEVICE_FUNC half(bool b)
+      : half_impl::half_base(half_impl::raw_uint16_to_half(b ? 0x3c00 : 0)) {}
+  template<class T>
+  explicit EIGEN_DEVICE_FUNC half(const T& val)
+      : half_impl::half_base(half_impl::float_to_half_rtne(static_cast<float>(val))) {}
+  explicit EIGEN_DEVICE_FUNC half(float f)
+      : half_impl::half_base(half_impl::float_to_half_rtne(f)) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(bool) const {
+    // +0.0 and -0.0 become false, everything else becomes true.
+    return (x & 0x7fff) != 0;
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(signed char) const {
+    return static_cast<signed char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned char) const {
+    return static_cast<unsigned char>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(short) const {
+    return static_cast<short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned short) const {
+    return static_cast<unsigned short>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(int) const {
+    return static_cast<int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned int) const {
+    return static_cast<unsigned int>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long) const {
+    return static_cast<long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long) const {
+    return static_cast<unsigned long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(long long) const {
+    return static_cast<long long>(half_impl::half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(unsigned long long) const {
+    return static_cast<unsigned long long>(half_to_float(*this));
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(float) const {
+    return half_impl::half_to_float(*this);
+  }
+  EIGEN_DEVICE_FUNC EIGEN_EXPLICIT_CAST(double) const {
+    return static_cast<double>(half_impl::half_to_float(*this));
+  }
+
+  EIGEN_DEVICE_FUNC half& operator=(const half& other) {
+    x = other.x;
+    return *this;
+  }
+};
+
+namespace half_impl {
+
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+
+// Intrinsics for native fp16 support. Note that on current hardware,
+// these are no faster than fp32 arithmetic (you need to use the half2
+// versions to get the ALU speed increased), but you do save the
+// conversion steps back and forth.
+
+__device__ half operator + (const half& a, const half& b) {
+  return __hadd(a, b);
+}
+__device__ half operator * (const half& a, const half& b) {
+  return __hmul(a, b);
+}
+__device__ half operator - (const half& a, const half& b) {
+  return __hsub(a, b);
+}
+__device__ half operator / (const half& a, const half& b) {
+  float num = __half2float(a);
+  float denom = __half2float(b);
+  return __float2half(num / denom);
+}
+__device__ half operator - (const half& a) {
+  return __hneg(a);
+}
+__device__ half& operator += (half& a, const half& b) {
+  a = a + b;
+  return a;
+}
+__device__ half& operator *= (half& a, const half& b) {
+  a = a * b;
+  return a;
+}
+__device__ half& operator -= (half& a, const half& b) {
+  a = a - b;
+  return a;
+}
+__device__ half& operator /= (half& a, const half& b) {
+  a = a / b;
+  return a;
+}
+__device__ bool operator == (const half& a, const half& b) {
+  return __heq(a, b);
+}
+__device__ bool operator != (const half& a, const half& b) {
+  return __hne(a, b);
+}
+__device__ bool operator < (const half& a, const half& b) {
+  return __hlt(a, b);
+}
+__device__ bool operator <= (const half& a, const half& b) {
+  return __hle(a, b);
+}
+__device__ bool operator > (const half& a, const half& b) {
+  return __hgt(a, b);
+}
+__device__ bool operator >= (const half& a, const half& b) {
+  return __hge(a, b);
+}
+
+#else  // Emulate support for half floats
+
+// Definitions for CPUs and older CUDA, mostly working through conversion
+// to/from fp32.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator + (const half& a, const half& b) {
+  return half(float(a) + float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator * (const half& a, const half& b) {
+  return half(float(a) * float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a, const half& b) {
+  return half(float(a) - float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, const half& b) {
+  return half(float(a) / float(b));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator - (const half& a) {
+  half result;
+  result.x = a.x ^ 0x8000;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator += (half& a, const half& b) {
+  a = half(float(a) + float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator *= (half& a, const half& b) {
+  a = half(float(a) * float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator -= (half& a, const half& b) {
+  a = half(float(a) - float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half& operator /= (half& a, const half& b) {
+  a = half(float(a) / float(b));
+  return a;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator == (const half& a, const half& b) {
+  return float(a) == float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator != (const half& a, const half& b) {
+  return float(a) != float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator < (const half& a, const half& b) {
+  return float(a) < float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator <= (const half& a, const half& b) {
+  return float(a) <= float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator > (const half& a, const half& b) {
+  return float(a) > float(b);
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool operator >= (const half& a, const half& b) {
+  return float(a) >= float(b);
+}
+
+#endif  // Emulate support for half floats
+
+// Division by an index. Do it in full float precision to avoid accuracy
+// issues in converting the denominator to half.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half operator / (const half& a, Index b) {
+  return half(static_cast<float>(a) / static_cast<float>(b));
+}
+
+// Conversion routines, including fallbacks for the host or older CUDA.
+// Note that newer Intel CPUs (Haswell or newer) have vectorized versions of
+// these in hardware. If we need more performance on older/other CPUs, they are
+// also possible to vectorize directly.
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half raw_uint16_to_half(unsigned short x) {
+  __half h;
+  h.x = x;
+  return h;
+}
+
+union FP32 {
+  unsigned int u;
+  float f;
+};
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC __half float_to_half_rtne(float ff) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+  return __float2half(ff);
+
+#elif defined(EIGEN_HAS_FP16_C)
+  __half h;
+  h.x = _cvtss_sh(ff, 0);
+  return h;
+
+#else
+  FP32 f; f.f = ff;
+
+  const FP32 f32infty = { 255 << 23 };
+  const FP32 f16max = { (127 + 16) << 23 };
+  const FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
+  unsigned int sign_mask = 0x80000000u;
+  __half o;
+  o.x = static_cast<unsigned short>(0x0u);
+
+  unsigned int sign = f.u & sign_mask;
+  f.u ^= sign;
+
+  // NOTE all the integer compares in this function can be safely
+  // compiled into signed compares since all operands are below
+  // 0x80000000. Important if you want fast straight SSE2 code
+  // (since there's no unsigned PCMPGTD).
+
+  if (f.u >= f16max.u) {  // result is Inf or NaN (all exponent bits set)
+    o.x = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
+  } else {  // (De)normalized number or zero
+    if (f.u < (113 << 23)) {  // resulting FP16 is subnormal or zero
+      // use a magic value to align our 10 mantissa bits at the bottom of
+      // the float. as long as FP addition is round-to-nearest-even this
+      // just works.
+      f.f += denorm_magic.f;
+
+      // and one integer subtract of the bias later, we have our final float!
+      o.x = static_cast<unsigned short>(f.u - denorm_magic.u);
+    } else {
+      unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
+
+      // update exponent, rounding bias part 1
+      f.u += ((unsigned int)(15 - 127) << 23) + 0xfff;
+      // rounding bias part 2
+      f.u += mant_odd;
+      // take the bits!
+      o.x = static_cast<unsigned short>(f.u >> 13);
+    }
+  }
+
+  o.x |= static_cast<unsigned short>(sign >> 16);
+  return o;
+#endif
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC float half_to_float(__half h) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+  return __half2float(h);
+
+#elif defined(EIGEN_HAS_FP16_C)
+  return _cvtsh_ss(h.x);
+
+#else
+  const FP32 magic = { 113 << 23 };
+  const unsigned int shifted_exp = 0x7c00 << 13; // exponent mask after shift
+  FP32 o;
+
+  o.u = (h.x & 0x7fff) << 13;             // exponent/mantissa bits
+  unsigned int exp = shifted_exp & o.u;   // just the exponent
+  o.u += (127 - 15) << 23;                // exponent adjust
+
+  // handle exponent special cases
+  if (exp == shifted_exp) {     // Inf/NaN?
+    o.u += (128 - 16) << 23;    // extra exp adjust
+  } else if (exp == 0) {        // Zero/Denormal?
+    o.u += 1 << 23;             // extra exp adjust
+    o.f -= magic.f;             // renormalize
+  }
+
+  o.u |= (h.x & 0x8000) << 16;    // sign bit
+  return o.f;
+#endif
+}
+
+// --- standard functions ---
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isinf)(const half& a) {
+  return (a.x & 0x7fff) == 0x7c00;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isnan)(const half& a) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hisnan(a);
+#else
+  return (a.x & 0x7fff) > 0x7c00;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool (isfinite)(const half& a) {
+  return !(isinf EIGEN_NOT_A_MACRO (a)) && !(isnan EIGEN_NOT_A_MACRO (a));
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half abs(const half& a) {
+  half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half exp(const half& a) {
+  return half(::expf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log(const half& a) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return Eigen::half(::hlog(a));
+#else
+  return half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log1p(const half& a) {
+  return half(numext::log1p(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half log10(const half& a) {
+  return half(::log10f(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sqrt(const half& a) {
+  return half(::sqrtf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half pow(const half& a, const half& b) {
+  return half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half sin(const half& a) {
+  return half(::sinf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half cos(const half& a) {
+  return half(::cosf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tan(const half& a) {
+  return half(::tanf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half tanh(const half& a) {
+  return half(::tanhf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half floor(const half& a) {
+  return half(::floorf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half ceil(const half& a) {
+  return half(::ceilf(float(a)));
+}
+
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (min)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hlt(b, a) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f2 < f1 ? b : a;
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC half (max)(const half& a, const half& b) {
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return __hlt(a, b) ? b : a;
+#else
+  const float f1 = static_cast<float>(a);
+  const float f2 = static_cast<float>(b);
+  return f1 < f2 ? b : a;
+#endif
+}
+
+EIGEN_ALWAYS_INLINE std::ostream& operator << (std::ostream& os, const half& v) {
+  os << static_cast<float>(v);
+  return os;
+}
+
+} // end namespace half_impl
+
+// import Eigen::half_impl::half into Eigen namespace
+// using half_impl::half;
+
+namespace internal {
+
+template<>
+struct random_default_impl<half, false, false>
+{
+  static inline half run(const half& x, const half& y)
+  {
+    return x + (y-x) * half(float(std::rand()) / float(RAND_MAX));
+  }
+  static inline half run()
+  {
+    return run(half(-1.f), half(1.f));
+  }
+};
+
+template<> struct is_arithmetic<half> { enum { value = true }; };
+
+} // end namespace internal
+
+}  // end namespace Eigen
+
+namespace std {
+template<>
+struct numeric_limits<Eigen::half> {
+  static const bool is_specialized = true;
+  static const bool is_signed = true;
+  static const bool is_integer = false;
+  static const bool is_exact = false;
+  static const bool has_infinity = true;
+  static const bool has_quiet_NaN = true;
+  static const bool has_signaling_NaN = true;
+  static const float_denorm_style has_denorm = denorm_present;
+  static const bool has_denorm_loss = false;
+  static const std::float_round_style round_style = std::round_to_nearest;
+  static const bool is_iec559 = false;
+  static const bool is_bounded = false;
+  static const bool is_modulo = false;
+  static const int digits = 11;
+  static const int digits10 = 2;
+  //static const int max_digits10 = ;
+  static const int radix = 2;
+  static const int min_exponent = -13;
+  static const int min_exponent10 = -4;
+  static const int max_exponent = 16;
+  static const int max_exponent10 = 4;
+  static const bool traps = true;
+  static const bool tinyness_before = false;
+
+  static Eigen::half (min)() { return Eigen::half_impl::raw_uint16_to_half(0x400); }
+  static Eigen::half lowest() { return Eigen::half_impl::raw_uint16_to_half(0xfbff); }
+  static Eigen::half (max)() { return Eigen::half_impl::raw_uint16_to_half(0x7bff); }
+  static Eigen::half epsilon() { return Eigen::half_impl::raw_uint16_to_half(0x0800); }
+  static Eigen::half round_error() { return Eigen::half(0.5); }
+  static Eigen::half infinity() { return Eigen::half_impl::raw_uint16_to_half(0x7c00); }
+  static Eigen::half quiet_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half signaling_NaN() { return Eigen::half_impl::raw_uint16_to_half(0x7e00); }
+  static Eigen::half denorm_min() { return Eigen::half_impl::raw_uint16_to_half(0x1); }
+};
+}
+
+namespace Eigen {
+
+template<> struct NumTraits<Eigen::half>
+    : GenericNumTraits<Eigen::half>
+{
+  enum {
+    IsSigned = true,
+    IsInteger = false,
+    IsComplex = false,
+    RequireInitialization = false
+  };
+
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half epsilon() {
+    return half_impl::raw_uint16_to_half(0x0800);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half dummy_precision() { return Eigen::half(1e-2f); }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half highest() {
+    return half_impl::raw_uint16_to_half(0x7bff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half lowest() {
+    return half_impl::raw_uint16_to_half(0xfbff);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half infinity() {
+    return half_impl::raw_uint16_to_half(0x7c00);
+  }
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Eigen::half quiet_NaN() {
+    return half_impl::raw_uint16_to_half(0x7c01);
+  }
+};
+
+} // end namespace Eigen
+
+// C-like standard mathematical functions and trancendentals.
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half fabsh(const Eigen::half& a) {
+  Eigen::half result;
+  result.x = a.x & 0x7FFF;
+  return result;
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half exph(const Eigen::half& a) {
+  return Eigen::half(::expf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half logh(const Eigen::half& a) {
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 530
+  return Eigen::half(::hlog(a));
+#else
+  return Eigen::half(::logf(float(a)));
+#endif
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half sqrth(const Eigen::half& a) {
+  return Eigen::half(::sqrtf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half powh(const Eigen::half& a, const Eigen::half& b) {
+  return Eigen::half(::powf(float(a), float(b)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half floorh(const Eigen::half& a) {
+  return Eigen::half(::floorf(float(a)));
+}
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half ceilh(const Eigen::half& a) {
+  return Eigen::half(::ceilf(float(a)));
+}
+
+namespace std {
+
+#if __cplusplus > 199711L
+template <>
+struct hash<Eigen::half> {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE std::size_t operator()(const Eigen::half& a) const {
+    return static_cast<std::size_t>(a.x);
+  }
+};
+#endif
+
+} // end namespace std
+
+
+// Add the missing shfl_xor intrinsic
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+__device__ EIGEN_STRONG_INLINE Eigen::half __shfl_xor(Eigen::half var, int laneMask, int width=warpSize) {
+  return static_cast<Eigen::half>(__shfl_xor(static_cast<float>(var), laneMask, width));
+}
+#endif
+
+// ldg() has an overload for __half, but we also need one for Eigen::half.
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC Eigen::half __ldg(const Eigen::half* ptr) {
+  return Eigen::half_impl::raw_uint16_to_half(
+      __ldg(reinterpret_cast<const unsigned short*>(ptr)));
+}
+#endif
+
+
+#if defined(__CUDA_ARCH__)
+namespace Eigen {
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isnan)(const Eigen::half& h) {
+  return (half_impl::isnan)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isinf)(const Eigen::half& h) {
+  return (half_impl::isinf)(h);
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+bool (isfinite)(const Eigen::half& h) {
+  return (half_impl::isfinite)(h);
+}
+
+} // namespace Eigen
+}  // namespace numext
+#endif
+
+#endif // EIGEN_HALF_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h
new file mode 100644
index 0000000..0348b41
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/MathFunctions.h
@@ -0,0 +1,91 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATH_FUNCTIONS_CUDA_H
+#define EIGEN_MATH_FUNCTIONS_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog<float4>(const float4& a)
+{
+  return make_float4(logf(a.x), logf(a.y), logf(a.z), logf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog<double2>(const double2& a)
+{
+  using ::log;
+  return make_double2(log(a.x), log(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 plog1p<float4>(const float4& a)
+{
+  return make_float4(log1pf(a.x), log1pf(a.y), log1pf(a.z), log1pf(a.w));
+}
+
+template<>  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 plog1p<double2>(const double2& a)
+{
+  return make_double2(log1p(a.x), log1p(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pexp<float4>(const float4& a)
+{
+  return make_float4(expf(a.x), expf(a.y), expf(a.z), expf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pexp<double2>(const double2& a)
+{
+  using ::exp;
+  return make_double2(exp(a.x), exp(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 psqrt<float4>(const float4& a)
+{
+  return make_float4(sqrtf(a.x), sqrtf(a.y), sqrtf(a.z), sqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 psqrt<double2>(const double2& a)
+{
+  using ::sqrt;
+  return make_double2(sqrt(a.x), sqrt(a.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 prsqrt<float4>(const float4& a)
+{
+  return make_float4(rsqrtf(a.x), rsqrtf(a.y), rsqrtf(a.z), rsqrtf(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 prsqrt<double2>(const double2& a)
+{
+  return make_double2(rsqrt(a.x), rsqrt(a.y));
+}
+
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h
new file mode 100644
index 0000000..4dda631
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMath.h
@@ -0,0 +1,333 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_CUDA_H
+#define EIGEN_PACKET_MATH_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Make sure this is only available when targeting a GPU: we don't want to
+// introduce conflicts between these packet_traits definitions and the ones
+// we'll use on the host side (SSE, AVX, ...)
+#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+template<> struct is_arithmetic<float4>  { enum { value = true }; };
+template<> struct is_arithmetic<double2> { enum { value = true }; };
+
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef float4 type;
+  typedef float4 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasLGamma = 1,
+    HasDiGamma = 1,
+    HasZeta = 1,
+    HasPolygamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
+    HasIGamma = 1,
+    HasIGammac = 1,
+    HasBetaInc = 1,
+
+    HasBlend = 0,
+  };
+};
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef double2 type;
+  typedef double2 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasLGamma = 1,
+    HasDiGamma = 1,
+    HasZeta = 1,
+    HasPolygamma = 1,
+    HasErf = 1,
+    HasErfc = 1,
+    HasIGamma = 1,
+    HasIGammac = 1,
+    HasBetaInc = 1,
+
+    HasBlend = 0,
+  };
+};
+
+
+template<> struct unpacket_traits<float4>  { typedef float  type; enum {size=4, alignment=Aligned16}; typedef float4 half; };
+template<> struct unpacket_traits<double2> { typedef double type; enum {size=2, alignment=Aligned16}; typedef double2 half; };
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pset1<float4>(const float&  from) {
+  return make_float4(from, from, from, from);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pset1<double2>(const double& from) {
+  return make_double2(from, from);
+}
+
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 plset<float4>(const float& a) {
+  return make_float4(a, a+1, a+2, a+3);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 plset<double2>(const double& a) {
+  return make_double2(a, a+1);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 padd<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x+b.x, a.y+b.y, a.z+b.z, a.w+b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 padd<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x+b.x, a.y+b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 psub<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x-b.x, a.y-b.y, a.z-b.z, a.w-b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 psub<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x-b.x, a.y-b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pnegate(const float4& a) {
+  return make_float4(-a.x, -a.y, -a.z, -a.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pnegate(const double2& a) {
+  return make_double2(-a.x, -a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pconj(const float4& a) { return a; }
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pconj(const double2& a) { return a; }
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmul<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x*b.x, a.y*b.y, a.z*b.z, a.w*b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmul<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x*b.x, a.y*b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pdiv<float4>(const float4& a, const float4& b) {
+  return make_float4(a.x/b.x, a.y/b.y, a.z/b.z, a.w/b.w);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pdiv<double2>(const double2& a, const double2& b) {
+  return make_double2(a.x/b.x, a.y/b.y);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmin<float4>(const float4& a, const float4& b) {
+  return make_float4(fminf(a.x, b.x), fminf(a.y, b.y), fminf(a.z, b.z), fminf(a.w, b.w));
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmin<double2>(const double2& a, const double2& b) {
+  return make_double2(fmin(a.x, b.x), fmin(a.y, b.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pmax<float4>(const float4& a, const float4& b) {
+  return make_float4(fmaxf(a.x, b.x), fmaxf(a.y, b.y), fmaxf(a.z, b.z), fmaxf(a.w, b.w));
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pmax<double2>(const double2& a, const double2& b) {
+  return make_double2(fmax(a.x, b.x), fmax(a.y, b.y));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pload<float4>(const float* from) {
+  return *reinterpret_cast<const float4*>(from);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 pload<double2>(const double* from) {
+  return *reinterpret_cast<const double2*>(from);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 ploadu<float4>(const float* from) {
+  return make_float4(from[0], from[1], from[2], from[3]);
+}
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE double2 ploadu<double2>(const double* from) {
+  return make_double2(from[0], from[1]);
+}
+
+template<> EIGEN_STRONG_INLINE float4 ploaddup<float4>(const float*   from) {
+  return make_float4(from[0], from[0], from[1], from[1]);
+}
+template<> EIGEN_STRONG_INLINE double2 ploaddup<double2>(const double*  from) {
+  return make_double2(from[0], from[0]);
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<float>(float*   to, const float4& from) {
+  *reinterpret_cast<float4*>(to) = from;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstore<double>(double* to, const double2& from) {
+  *reinterpret_cast<double2*>(to) = from;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const float4& from) {
+  to[0] = from.x;
+  to[1] = from.y;
+  to[2] = from.z;
+  to[3] = from.w;
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const double2& from) {
+  to[0] = from.x;
+  to[1] = from.y;
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Aligned>(const float* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return __ldg((const float4*)from);
+#else
+  return make_float4(from[0], from[1], from[2], from[3]);
+#endif
+}
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Aligned>(const double* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return __ldg((const double2*)from);
+#else
+  return make_double2(from[0], from[1]);
+#endif
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE float4 ploadt_ro<float4, Unaligned>(const float* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return make_float4(__ldg(from+0), __ldg(from+1), __ldg(from+2), __ldg(from+3));
+#else
+  return make_float4(from[0], from[1], from[2], from[3]);
+#endif
+}
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE double2 ploadt_ro<double2, Unaligned>(const double* from) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 350
+  return make_double2(__ldg(from+0), __ldg(from+1));
+#else
+  return make_double2(from[0], from[1]);
+#endif
+}
+
+template<> EIGEN_DEVICE_FUNC inline float4 pgather<float, float4>(const float* from, Index stride) {
+  return make_float4(from[0*stride], from[1*stride], from[2*stride], from[3*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline double2 pgather<double, double2>(const double* from, Index stride) {
+  return make_double2(from[0*stride], from[1*stride]);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, float4>(float* to, const float4& from, Index stride) {
+  to[stride*0] = from.x;
+  to[stride*1] = from.y;
+  to[stride*2] = from.z;
+  to[stride*3] = from.w;
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, double2>(double* to, const double2& from, Index stride) {
+  to[stride*0] = from.x;
+  to[stride*1] = from.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  pfirst<float4>(const float4& a) {
+  return a.x;
+}
+template<> EIGEN_DEVICE_FUNC inline double pfirst<double2>(const double2& a) {
+  return a.x;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux<float4>(const float4& a) {
+  return a.x + a.y + a.z + a.w;
+}
+template<> EIGEN_DEVICE_FUNC inline double predux<double2>(const double2& a) {
+  return a.x + a.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_max<float4>(const float4& a) {
+  return fmaxf(fmaxf(a.x, a.y), fmaxf(a.z, a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_max<double2>(const double2& a) {
+  return fmax(a.x, a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_min<float4>(const float4& a) {
+  return fminf(fminf(a.x, a.y), fminf(a.z, a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_min<double2>(const double2& a) {
+  return fmin(a.x, a.y);
+}
+
+template<> EIGEN_DEVICE_FUNC inline float  predux_mul<float4>(const float4& a) {
+  return a.x * a.y * a.z * a.w;
+}
+template<> EIGEN_DEVICE_FUNC inline double predux_mul<double2>(const double2& a) {
+  return a.x * a.y;
+}
+
+template<> EIGEN_DEVICE_FUNC inline float4  pabs<float4>(const float4& a) {
+  return make_float4(fabsf(a.x), fabsf(a.y), fabsf(a.z), fabsf(a.w));
+}
+template<> EIGEN_DEVICE_FUNC inline double2 pabs<double2>(const double2& a) {
+  return make_double2(fabs(a.x), fabs(a.y));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<float4,4>& kernel) {
+  float tmp = kernel.packet[0].y;
+  kernel.packet[0].y = kernel.packet[1].x;
+  kernel.packet[1].x = tmp;
+
+  tmp = kernel.packet[0].z;
+  kernel.packet[0].z = kernel.packet[2].x;
+  kernel.packet[2].x = tmp;
+
+  tmp = kernel.packet[0].w;
+  kernel.packet[0].w = kernel.packet[3].x;
+  kernel.packet[3].x = tmp;
+
+  tmp = kernel.packet[1].z;
+  kernel.packet[1].z = kernel.packet[2].y;
+  kernel.packet[2].y = tmp;
+
+  tmp = kernel.packet[1].w;
+  kernel.packet[1].w = kernel.packet[3].y;
+  kernel.packet[3].y = tmp;
+
+  tmp = kernel.packet[2].w;
+  kernel.packet[2].w = kernel.packet[3].z;
+  kernel.packet[3].z = tmp;
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<double2,2>& kernel) {
+  double tmp = kernel.packet[0].y;
+  kernel.packet[0].y = kernel.packet[1].x;
+  kernel.packet[1].x = tmp;
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+
+#endif // EIGEN_PACKET_MATH_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
new file mode 100644
index 0000000..ae54225
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/PacketMathHalf.h
@@ -0,0 +1,1123 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_HALF_CUDA_H
+#define EIGEN_PACKET_MATH_HALF_CUDA_H
+
+
+namespace Eigen {
+namespace internal {
+
+// Most of the following operations require arch >= 3.0
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDACC__) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+
+template<> struct is_arithmetic<half2> { enum { value = true }; };
+
+template<> struct packet_traits<Eigen::half> : default_packet_traits
+{
+  typedef half2 type;
+  typedef half2 half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+    HasAdd    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasSqrt   = 1,
+    HasRsqrt  = 1,
+    HasExp    = 1,
+    HasLog    = 1,
+    HasLog1p  = 1
+  };
+};
+
+template<> struct unpacket_traits<half2> { typedef Eigen::half type; enum {size=2, alignment=Aligned16}; typedef half2 half; };
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pset1<half2>(const Eigen::half& from) {
+  return __half2half2(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pload<half2>(const Eigen::half* from) {
+  return *reinterpret_cast<const half2*>(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 ploadu<half2>(const Eigen::half* from) {
+  return __halves2half2(from[0], from[1]);
+}
+
+template<> EIGEN_STRONG_INLINE half2 ploaddup<half2>(const Eigen::half*  from) {
+  return __halves2half2(from[0], from[0]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const half2& from) {
+  *reinterpret_cast<half2*>(to) = from;
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const half2& from) {
+  to[0] = __low2half(from);
+  to[1] = __high2half(from);
+}
+
+template<>
+ __device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Aligned>(const Eigen::half* from) {
+#if __CUDA_ARCH__ >= 350
+   return __ldg((const half2*)from);
+#else
+  return __halves2half2(*(from+0), *(from+1));
+#endif
+}
+
+template<>
+__device__ EIGEN_ALWAYS_INLINE half2 ploadt_ro<half2, Unaligned>(const Eigen::half* from) {
+#if __CUDA_ARCH__ >= 350
+   return __halves2half2(__ldg(from+0), __ldg(from+1));
+#else
+  return __halves2half2(*(from+0), *(from+1));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pgather<Eigen::half, half2>(const Eigen::half* from, Index stride) {
+  return __halves2half2(from[0*stride], from[1*stride]);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE void pscatter<Eigen::half, half2>(Eigen::half* to, const half2& from, Index stride) {
+  to[stride*0] = __low2half(from);
+  to[stride*1] = __high2half(from);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half pfirst<half2>(const half2& a) {
+  return __low2half(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pabs<half2>(const half2& a) {
+  half2 result;
+  result.x = a.x & 0x7FFF7FFF;
+  return result;
+}
+
+
+__device__ EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<half2,2>& kernel) {
+  __half a1 = __low2half(kernel.packet[0]);
+  __half a2 = __high2half(kernel.packet[0]);
+  __half b1 = __low2half(kernel.packet[1]);
+  __half b2 = __high2half(kernel.packet[1]);
+  kernel.packet[0] = __halves2half2(a1, b1);
+  kernel.packet[1] = __halves2half2(a2, b2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plset<half2>(const Eigen::half& a) {
+#if __CUDA_ARCH__ >= 530
+  return __halves2half2(a, __hadd(a, __float2half(1.0f)));
+#else
+  float f = __half2float(a) + 1.0f;
+  return __halves2half2(a, __float2half(f));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 padd<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hadd2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 + b1;
+  float r2 = a2 + b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 psub<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hsub2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 - b1;
+  float r2 = a2 - b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pnegate(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hneg2(a);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return __floats2half2_rn(-a1, -a2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pconj(const half2& a) { return a; }
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmul<half2>(const half2& a, const half2& b) {
+#if __CUDA_ARCH__ >= 530
+  return __hmul2(a, b);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 * b1;
+  float r2 = a2 * b2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmadd<half2>(const half2& a, const half2& b, const half2& c) {
+#if __CUDA_ARCH__ >= 530
+   return __hfma2(a, b, c);
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float c1 = __low2float(c);
+  float c2 = __high2float(c);
+  float r1 = a1 * b1 + c1;
+  float r2 = a2 * b2 + c2;
+  return __floats2half2_rn(r1, r2);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pdiv<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  float r1 = a1 / b1;
+  float r2 = a2 / b2;
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmin<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 < b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 < b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pmax<half2>(const half2& a, const half2& b) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float b1 = __low2float(b);
+  float b2 = __high2float(b);
+  __half r1 = a1 > b1 ? __low2half(a) : __low2half(b);
+  __half r2 = a2 > b2 ? __high2half(a) : __high2half(b);
+  return __halves2half2(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hadd(__low2half(a), __high2half(a));
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 + a2)));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_max<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hgt(first, second) ? first : second;
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return a1 > a2 ? __low2half(a) : __high2half(a);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_min<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  __half first = __low2half(a);
+  __half second = __high2half(a);
+  return __hlt(first, second) ? first : second;
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return a1 < a2 ? __low2half(a) : __high2half(a);
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE Eigen::half predux_mul<half2>(const half2& a) {
+#if __CUDA_ARCH__ >= 530
+  return __hmul(__low2half(a), __high2half(a));
+#else
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  return Eigen::half(half_impl::raw_uint16_to_half(__float2half_rn(a1 * a2)));
+#endif
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plog1p<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = log1pf(a1);
+  float r2 = log1pf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+#if defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000 && defined __CUDA_ARCH__ && __CUDA_ARCH__ >= 530
+
+template<>  __device__ EIGEN_STRONG_INLINE
+half2 plog<half2>(const half2& a) {
+  return h2log(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 pexp<half2>(const half2& a) {
+  return h2exp(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 psqrt<half2>(const half2& a) {
+  return h2sqrt(a);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE
+half2 prsqrt<half2>(const half2& a) {
+  return h2rsqrt(a);
+}
+
+#else
+
+template<> __device__ EIGEN_STRONG_INLINE half2 plog<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = logf(a1);
+  float r2 = logf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 pexp<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = expf(a1);
+  float r2 = expf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 psqrt<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = sqrtf(a1);
+  float r2 = sqrtf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+template<> __device__ EIGEN_STRONG_INLINE half2 prsqrt<half2>(const half2& a) {
+  float a1 = __low2float(a);
+  float a2 = __high2float(a);
+  float r1 = rsqrtf(a1);
+  float r2 = rsqrtf(a2);
+  return __floats2half2_rn(r1, r2);
+}
+
+#endif
+
+#elif defined EIGEN_VECTORIZE_AVX512
+
+typedef struct {
+  __m256i x;
+} Packet16h;
+
+
+template<> struct is_arithmetic<Packet16h> { enum { value = true }; };
+
+template <>
+struct packet_traits<half> : default_packet_traits {
+  typedef Packet16h type;
+  // There is no half-size packet for Packet16h.
+  typedef Packet16h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 16,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet16h> { typedef Eigen::half type; enum {size=16, alignment=Aligned32}; typedef Packet16h half; };
+
+template<> EIGEN_STRONG_INLINE Packet16h pset1<Packet16h>(const Eigen::half& from) {
+  Packet16h result;
+  result.x = _mm256_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet16h>(const Packet16h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm256_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pload<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_load_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h ploadu<Packet16h>(const Eigen::half* from) {
+  Packet16h result;
+  result.x = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_store_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<half>(Eigen::half* to, const Packet16h& from) {
+  _mm256_storeu_si256((__m256i*)to, from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h
+ploadquad(const Eigen::half* from) {
+  Packet16h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  unsigned short c = from[2].x;
+  unsigned short d = from[3].x;
+  result.x = _mm256_set_epi16(d, d, d, d, c, c, c, c, b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet16f half2float(const Packet16h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm512_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+  float f8(aux[8]);
+  float f9(aux[9]);
+  float fa(aux[10]);
+  float fb(aux[11]);
+  float fc(aux[12]);
+  float fd(aux[13]);
+  float fe(aux[14]);
+  float ff(aux[15]);
+
+  return _mm512_set_ps(
+      ff, fe, fd, fc, fb, fa, f9, f8, f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet16h float2half(const Packet16f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet16h result;
+  result.x = _mm512_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN64 float aux[16];
+  pstore(aux, a);
+  half h0(aux[0]);
+  half h1(aux[1]);
+  half h2(aux[2]);
+  half h3(aux[3]);
+  half h4(aux[4]);
+  half h5(aux[5]);
+  half h6(aux[6]);
+  half h7(aux[7]);
+  half h8(aux[8]);
+  half h9(aux[9]);
+  half ha(aux[10]);
+  half hb(aux[11]);
+  half hc(aux[12]);
+  half hd(aux[13]);
+  half he(aux[14]);
+  half hf(aux[15]);
+
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      hf.x, he.x, hd.x, hc.x, hb.x, ha.x, h9.x, h8.x,
+      h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h padd<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pmul<Packet16h>(const Packet16h& a, const Packet16h& b) {
+  Packet16f af = half2float(a);
+  Packet16f bf = half2float(b);
+  Packet16f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE half predux<Packet16h>(const Packet16h& from) {
+  Packet16f from_float = half2float(from);
+  return half(predux(from_float));
+}
+
+template<> EIGEN_STRONG_INLINE Packet16h pgather<Eigen::half, Packet16h>(const Eigen::half* from, Index stride)
+{
+  Packet16h result;
+  result.x = _mm256_set_epi16(
+      from[15*stride].x, from[14*stride].x, from[13*stride].x, from[12*stride].x,
+      from[11*stride].x, from[10*stride].x, from[9*stride].x, from[8*stride].x,
+      from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x,
+      from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<half, Packet16h>(half* to, const Packet16h& from, Index stride)
+{
+  EIGEN_ALIGN64 half aux[16];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+  to[stride*8].x = aux[8].x;
+  to[stride*9].x = aux[9].x;
+  to[stride*10].x = aux[10].x;
+  to[stride*11].x = aux[11].x;
+  to[stride*12].x = aux[12].x;
+  to[stride*13].x = aux[13].x;
+  to[stride*14].x = aux[14].x;
+  to[stride*15].x = aux[15].x;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,16>& kernel) {
+  __m256i a = kernel.packet[0].x;
+  __m256i b = kernel.packet[1].x;
+  __m256i c = kernel.packet[2].x;
+  __m256i d = kernel.packet[3].x;
+  __m256i e = kernel.packet[4].x;
+  __m256i f = kernel.packet[5].x;
+  __m256i g = kernel.packet[6].x;
+  __m256i h = kernel.packet[7].x;
+  __m256i i = kernel.packet[8].x;
+  __m256i j = kernel.packet[9].x;
+  __m256i k = kernel.packet[10].x;
+  __m256i l = kernel.packet[11].x;
+  __m256i m = kernel.packet[12].x;
+  __m256i n = kernel.packet[13].x;
+  __m256i o = kernel.packet[14].x;
+  __m256i p = kernel.packet[15].x;
+
+  __m256i ab_07 = _mm256_unpacklo_epi16(a, b);
+  __m256i cd_07 = _mm256_unpacklo_epi16(c, d);
+  __m256i ef_07 = _mm256_unpacklo_epi16(e, f);
+  __m256i gh_07 = _mm256_unpacklo_epi16(g, h);
+  __m256i ij_07 = _mm256_unpacklo_epi16(i, j);
+  __m256i kl_07 = _mm256_unpacklo_epi16(k, l);
+  __m256i mn_07 = _mm256_unpacklo_epi16(m, n);
+  __m256i op_07 = _mm256_unpacklo_epi16(o, p);
+
+  __m256i ab_8f = _mm256_unpackhi_epi16(a, b);
+  __m256i cd_8f = _mm256_unpackhi_epi16(c, d);
+  __m256i ef_8f = _mm256_unpackhi_epi16(e, f);
+  __m256i gh_8f = _mm256_unpackhi_epi16(g, h);
+  __m256i ij_8f = _mm256_unpackhi_epi16(i, j);
+  __m256i kl_8f = _mm256_unpackhi_epi16(k, l);
+  __m256i mn_8f = _mm256_unpackhi_epi16(m, n);
+  __m256i op_8f = _mm256_unpackhi_epi16(o, p);
+
+  __m256i abcd_03 = _mm256_unpacklo_epi32(ab_07, cd_07);
+  __m256i abcd_47 = _mm256_unpackhi_epi32(ab_07, cd_07);
+  __m256i efgh_03 = _mm256_unpacklo_epi32(ef_07, gh_07);
+  __m256i efgh_47 = _mm256_unpackhi_epi32(ef_07, gh_07);
+  __m256i ijkl_03 = _mm256_unpacklo_epi32(ij_07, kl_07);
+  __m256i ijkl_47 = _mm256_unpackhi_epi32(ij_07, kl_07);
+  __m256i mnop_03 = _mm256_unpacklo_epi32(mn_07, op_07);
+  __m256i mnop_47 = _mm256_unpackhi_epi32(mn_07, op_07);
+
+  __m256i abcd_8b = _mm256_unpacklo_epi32(ab_8f, cd_8f);
+  __m256i abcd_cf = _mm256_unpackhi_epi32(ab_8f, cd_8f);
+  __m256i efgh_8b = _mm256_unpacklo_epi32(ef_8f, gh_8f);
+  __m256i efgh_cf = _mm256_unpackhi_epi32(ef_8f, gh_8f);
+  __m256i ijkl_8b = _mm256_unpacklo_epi32(ij_8f, kl_8f);
+  __m256i ijkl_cf = _mm256_unpackhi_epi32(ij_8f, kl_8f);
+  __m256i mnop_8b = _mm256_unpacklo_epi32(mn_8f, op_8f);
+  __m256i mnop_cf = _mm256_unpackhi_epi32(mn_8f, op_8f);
+
+  __m256i abcdefgh_01 = _mm256_unpacklo_epi64(abcd_03, efgh_03);
+  __m256i abcdefgh_23 = _mm256_unpackhi_epi64(abcd_03, efgh_03);
+  __m256i ijklmnop_01 = _mm256_unpacklo_epi64(ijkl_03, mnop_03);
+  __m256i ijklmnop_23 = _mm256_unpackhi_epi64(ijkl_03, mnop_03);
+  __m256i abcdefgh_45 = _mm256_unpacklo_epi64(abcd_47, efgh_47);
+  __m256i abcdefgh_67 = _mm256_unpackhi_epi64(abcd_47, efgh_47);
+  __m256i ijklmnop_45 = _mm256_unpacklo_epi64(ijkl_47, mnop_47);
+  __m256i ijklmnop_67 = _mm256_unpackhi_epi64(ijkl_47, mnop_47);
+  __m256i abcdefgh_89 = _mm256_unpacklo_epi64(abcd_8b, efgh_8b);
+  __m256i abcdefgh_ab = _mm256_unpackhi_epi64(abcd_8b, efgh_8b);
+  __m256i ijklmnop_89 = _mm256_unpacklo_epi64(ijkl_8b, mnop_8b);
+  __m256i ijklmnop_ab = _mm256_unpackhi_epi64(ijkl_8b, mnop_8b);
+  __m256i abcdefgh_cd = _mm256_unpacklo_epi64(abcd_cf, efgh_cf);
+  __m256i abcdefgh_ef = _mm256_unpackhi_epi64(abcd_cf, efgh_cf);
+  __m256i ijklmnop_cd = _mm256_unpacklo_epi64(ijkl_cf, mnop_cf);
+  __m256i ijklmnop_ef = _mm256_unpackhi_epi64(ijkl_cf, mnop_cf);
+
+  // NOTE: no unpacklo/hi instr in this case, so using permute instr.
+  __m256i a_p_0 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x20);
+  __m256i a_p_1 = _mm256_permute2x128_si256(abcdefgh_01, ijklmnop_01, 0x31);
+  __m256i a_p_2 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x20);
+  __m256i a_p_3 = _mm256_permute2x128_si256(abcdefgh_23, ijklmnop_23, 0x31);
+  __m256i a_p_4 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x20);
+  __m256i a_p_5 = _mm256_permute2x128_si256(abcdefgh_45, ijklmnop_45, 0x31);
+  __m256i a_p_6 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x20);
+  __m256i a_p_7 = _mm256_permute2x128_si256(abcdefgh_67, ijklmnop_67, 0x31);
+  __m256i a_p_8 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x20);
+  __m256i a_p_9 = _mm256_permute2x128_si256(abcdefgh_89, ijklmnop_89, 0x31);
+  __m256i a_p_a = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x20);
+  __m256i a_p_b = _mm256_permute2x128_si256(abcdefgh_ab, ijklmnop_ab, 0x31);
+  __m256i a_p_c = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x20);
+  __m256i a_p_d = _mm256_permute2x128_si256(abcdefgh_cd, ijklmnop_cd, 0x31);
+  __m256i a_p_e = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x20);
+  __m256i a_p_f = _mm256_permute2x128_si256(abcdefgh_ef, ijklmnop_ef, 0x31);
+
+  kernel.packet[0].x = a_p_0;
+  kernel.packet[1].x = a_p_1;
+  kernel.packet[2].x = a_p_2;
+  kernel.packet[3].x = a_p_3;
+  kernel.packet[4].x = a_p_4;
+  kernel.packet[5].x = a_p_5;
+  kernel.packet[6].x = a_p_6;
+  kernel.packet[7].x = a_p_7;
+  kernel.packet[8].x = a_p_8;
+  kernel.packet[9].x = a_p_9;
+  kernel.packet[10].x = a_p_a;
+  kernel.packet[11].x = a_p_b;
+  kernel.packet[12].x = a_p_c;
+  kernel.packet[13].x = a_p_d;
+  kernel.packet[14].x = a_p_e;
+  kernel.packet[15].x = a_p_f;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,8>& kernel) {
+  EIGEN_ALIGN64 half in[8][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+  pstore<half>(in[4], kernel.packet[4]);
+  pstore<half>(in[5], kernel.packet[5]);
+  pstore<half>(in[6], kernel.packet[6]);
+  pstore<half>(in[7], kernel.packet[7]);
+
+  EIGEN_ALIGN64 half out[8][16];
+
+  for (int i = 0; i < 8; ++i) {
+    for (int j = 0; j < 8; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 8; ++j) {
+      out[i][j+8] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+  kernel.packet[4] = pload<Packet16h>(out[4]);
+  kernel.packet[5] = pload<Packet16h>(out[5]);
+  kernel.packet[6] = pload<Packet16h>(out[6]);
+  kernel.packet[7] = pload<Packet16h>(out[7]);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet16h,4>& kernel) {
+  EIGEN_ALIGN64 half in[4][16];
+  pstore<half>(in[0], kernel.packet[0]);
+  pstore<half>(in[1], kernel.packet[1]);
+  pstore<half>(in[2], kernel.packet[2]);
+  pstore<half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN64 half out[4][16];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][4*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][4*i+1];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+8] = in[j][4*i+2];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+12] = in[j][4*i+3];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet16h>(out[0]);
+  kernel.packet[1] = pload<Packet16h>(out[1]);
+  kernel.packet[2] = pload<Packet16h>(out[2]);
+  kernel.packet[3] = pload<Packet16h>(out[3]);
+}
+
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+typedef struct {
+  __m128i x;
+} Packet8h;
+
+
+template<> struct is_arithmetic<Packet8h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet8h type;
+  // There is no half-size packet for Packet8h.
+  typedef Packet8h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 8,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet8h> { typedef Eigen::half type; enum {size=8, alignment=Aligned16}; typedef Packet8h half; };
+
+template<> EIGEN_STRONG_INLINE Packet8h pset1<Packet8h>(const Eigen::half& from) {
+  Packet8h result;
+  result.x = _mm_set1_epi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet8h>(const Packet8h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_extract_epi16(from.x, 0)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pload<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_load_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h ploadu<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  result.x = _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_store_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet8h& from) {
+  _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from.x);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h
+ploadquad<Packet8h>(const Eigen::half* from) {
+  Packet8h result;
+  unsigned short a = from[0].x;
+  unsigned short b = from[1].x;
+  result.x = _mm_set_epi16(b, b, b, b, a, a, a, a);
+  return result;
+}
+
+EIGEN_STRONG_INLINE Packet8f half2float(const Packet8h& a) {
+#ifdef EIGEN_HAS_FP16_C
+  return _mm256_cvtph_ps(a.x);
+#else
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, a);
+  float f0(aux[0]);
+  float f1(aux[1]);
+  float f2(aux[2]);
+  float f3(aux[3]);
+  float f4(aux[4]);
+  float f5(aux[5]);
+  float f6(aux[6]);
+  float f7(aux[7]);
+
+  return _mm256_set_ps(f7, f6, f5, f4, f3, f2, f1, f0);
+#endif
+}
+
+EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
+#ifdef EIGEN_HAS_FP16_C
+  Packet8h result;
+  result.x = _mm256_cvtps_ph(a, _MM_FROUND_TO_NEAREST_INT|_MM_FROUND_NO_EXC);
+  return result;
+#else
+  EIGEN_ALIGN32 float aux[8];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+  Eigen::half h4(aux[4]);
+  Eigen::half h5(aux[5]);
+  Eigen::half h6(aux[6]);
+  Eigen::half h7(aux[7]);
+
+  Packet8h result;
+  result.x = _mm_set_epi16(h7.x, h6.x, h5.x, h4.x, h3.x, h2.x, h1.x, h0.x);
+  return result;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pconj(const Packet8h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet8h padd<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = padd(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pmul<Packet8h>(const Packet8h& a, const Packet8h& b) {
+  Packet8f af = half2float(a);
+  Packet8f bf = half2float(b);
+  Packet8f rf = pmul(af, bf);
+  return float2half(rf);
+}
+
+template<> EIGEN_STRONG_INLINE Packet8h pgather<Eigen::half, Packet8h>(const Eigen::half* from, Index stride)
+{
+  Packet8h result;
+  result.x = _mm_set_epi16(from[7*stride].x, from[6*stride].x, from[5*stride].x, from[4*stride].x, from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet8h>(Eigen::half* to, const Packet8h& from, Index stride)
+{
+  EIGEN_ALIGN32 Eigen::half aux[8];
+  pstore(aux, from);
+  to[stride*0].x = aux[0].x;
+  to[stride*1].x = aux[1].x;
+  to[stride*2].x = aux[2].x;
+  to[stride*3].x = aux[3].x;
+  to[stride*4].x = aux[4].x;
+  to[stride*5].x = aux[5].x;
+  to[stride*6].x = aux[6].x;
+  to[stride*7].x = aux[7].x;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_max<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_max<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_min<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_min<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half predux_mul<Packet8h>(const Packet8h& a) {
+  Packet8f af = half2float(a);
+  float reduced = predux_mul<Packet8f>(af);
+  return Eigen::half(reduced);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,8>& kernel) {
+  __m128i a = kernel.packet[0].x;
+  __m128i b = kernel.packet[1].x;
+  __m128i c = kernel.packet[2].x;
+  __m128i d = kernel.packet[3].x;
+  __m128i e = kernel.packet[4].x;
+  __m128i f = kernel.packet[5].x;
+  __m128i g = kernel.packet[6].x;
+  __m128i h = kernel.packet[7].x;
+
+  __m128i a03b03 = _mm_unpacklo_epi16(a, b);
+  __m128i c03d03 = _mm_unpacklo_epi16(c, d);
+  __m128i e03f03 = _mm_unpacklo_epi16(e, f);
+  __m128i g03h03 = _mm_unpacklo_epi16(g, h);
+  __m128i a47b47 = _mm_unpackhi_epi16(a, b);
+  __m128i c47d47 = _mm_unpackhi_epi16(c, d);
+  __m128i e47f47 = _mm_unpackhi_epi16(e, f);
+  __m128i g47h47 = _mm_unpackhi_epi16(g, h);
+
+  __m128i a01b01c01d01 = _mm_unpacklo_epi32(a03b03, c03d03);
+  __m128i a23b23c23d23 = _mm_unpackhi_epi32(a03b03, c03d03);
+  __m128i e01f01g01h01 = _mm_unpacklo_epi32(e03f03, g03h03);
+  __m128i e23f23g23h23 = _mm_unpackhi_epi32(e03f03, g03h03);
+  __m128i a45b45c45d45 = _mm_unpacklo_epi32(a47b47, c47d47);
+  __m128i a67b67c67d67 = _mm_unpackhi_epi32(a47b47, c47d47);
+  __m128i e45f45g45h45 = _mm_unpacklo_epi32(e47f47, g47h47);
+  __m128i e67f67g67h67 = _mm_unpackhi_epi32(e47f47, g47h47);
+
+  __m128i a0b0c0d0e0f0g0h0 = _mm_unpacklo_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a1b1c1d1e1f1g1h1 = _mm_unpackhi_epi64(a01b01c01d01, e01f01g01h01);
+  __m128i a2b2c2d2e2f2g2h2 = _mm_unpacklo_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a3b3c3d3e3f3g3h3 = _mm_unpackhi_epi64(a23b23c23d23, e23f23g23h23);
+  __m128i a4b4c4d4e4f4g4h4 = _mm_unpacklo_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a5b5c5d5e5f5g5h5 = _mm_unpackhi_epi64(a45b45c45d45, e45f45g45h45);
+  __m128i a6b6c6d6e6f6g6h6 = _mm_unpacklo_epi64(a67b67c67d67, e67f67g67h67);
+  __m128i a7b7c7d7e7f7g7h7 = _mm_unpackhi_epi64(a67b67c67d67, e67f67g67h67);
+
+  kernel.packet[0].x = a0b0c0d0e0f0g0h0;
+  kernel.packet[1].x = a1b1c1d1e1f1g1h1;
+  kernel.packet[2].x = a2b2c2d2e2f2g2h2;
+  kernel.packet[3].x = a3b3c3d3e3f3g3h3;
+  kernel.packet[4].x = a4b4c4d4e4f4g4h4;
+  kernel.packet[5].x = a5b5c5d5e5f5g5h5;
+  kernel.packet[6].x = a6b6c6d6e6f6g6h6;
+  kernel.packet[7].x = a7b7c7d7e7f7g7h7;
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet8h,4>& kernel) {
+  EIGEN_ALIGN32 Eigen::half in[4][8];
+  pstore<Eigen::half>(in[0], kernel.packet[0]);
+  pstore<Eigen::half>(in[1], kernel.packet[1]);
+  pstore<Eigen::half>(in[2], kernel.packet[2]);
+  pstore<Eigen::half>(in[3], kernel.packet[3]);
+
+  EIGEN_ALIGN32 Eigen::half out[4][8];
+
+  for (int i = 0; i < 4; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      out[i][j] = in[j][2*i];
+    }
+    for (int j = 0; j < 4; ++j) {
+      out[i][j+4] = in[j][2*i+1];
+    }
+  }
+
+  kernel.packet[0] = pload<Packet8h>(out[0]);
+  kernel.packet[1] = pload<Packet8h>(out[1]);
+  kernel.packet[2] = pload<Packet8h>(out[2]);
+  kernel.packet[3] = pload<Packet8h>(out[3]);
+}
+
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+typedef struct {
+  __m64 x;
+} Packet4h;
+
+
+template<> struct is_arithmetic<Packet4h> { enum { value = true }; };
+
+template <>
+struct packet_traits<Eigen::half> : default_packet_traits {
+  typedef Packet4h type;
+  // There is no half-size packet for Packet4h.
+  typedef Packet4h half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+    HasAdd    = 0,
+    HasSub    = 0,
+    HasMul    = 0,
+    HasNegate = 0,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasConj   = 0,
+    HasSetLinear = 0,
+    HasDiv = 0,
+    HasSqrt = 0,
+    HasRsqrt = 0,
+    HasExp = 0,
+    HasLog = 0,
+    HasBlend = 0
+  };
+};
+
+
+template<> struct unpacket_traits<Packet4h> { typedef Eigen::half type; enum {size=4, alignment=Aligned16}; typedef Packet4h half; };
+
+template<> EIGEN_STRONG_INLINE Packet4h pset1<Packet4h>(const Eigen::half& from) {
+  Packet4h result;
+  result.x = _mm_set1_pi16(from.x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Eigen::half pfirst<Packet4h>(const Packet4h& from) {
+  return half_impl::raw_uint16_to_half(static_cast<unsigned short>(_mm_cvtsi64_si32(from.x)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pconj(const Packet4h& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4h padd<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha + hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha + hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pmul<Packet4h>(const Packet4h& a, const Packet4h& b) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  __int64_t b64 = _mm_cvtm64_si64(b.x);
+
+  Eigen::half h[4];
+
+  Eigen::half ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64));
+  Eigen::half hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64));
+  h[0] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 16));
+  h[1] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 32));
+  h[2] = ha * hb;
+  ha = half_impl::raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  hb = half_impl::raw_uint16_to_half(static_cast<unsigned short>(b64 >> 48));
+  h[3] = ha * hb;
+  Packet4h result;
+  result.x = _mm_set_pi16(h[3].x, h[2].x, h[1].x, h[0].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pload<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h ploadu<Packet4h>(const Eigen::half* from) {
+  Packet4h result;
+  result.x = _mm_cvtsi64_m64(*reinterpret_cast<const __int64_t*>(from));
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<Eigen::half>(Eigen::half* to, const Packet4h& from) {
+  __int64_t r = _mm_cvtm64_si64(from.x);
+  *(reinterpret_cast<__int64_t*>(to)) = r;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h
+ploadquad<Packet4h>(const Eigen::half* from) {
+  return pset1<Packet4h>(*from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4h pgather<Eigen::half, Packet4h>(const Eigen::half* from, Index stride)
+{
+  Packet4h result;
+  result.x = _mm_set_pi16(from[3*stride].x, from[2*stride].x, from[1*stride].x, from[0*stride].x);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE void pscatter<Eigen::half, Packet4h>(Eigen::half* to, const Packet4h& from, Index stride)
+{
+  __int64_t a = _mm_cvtm64_si64(from.x);
+  to[stride*0].x = static_cast<unsigned short>(a);
+  to[stride*1].x = static_cast<unsigned short>(a >> 16);
+  to[stride*2].x = static_cast<unsigned short>(a >> 32);
+  to[stride*3].x = static_cast<unsigned short>(a >> 48);
+}
+
+EIGEN_STRONG_INLINE void
+ptranspose(PacketBlock<Packet4h,4>& kernel) {
+  __m64 T0 = _mm_unpacklo_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T1 = _mm_unpacklo_pi16(kernel.packet[2].x, kernel.packet[3].x);
+  __m64 T2 = _mm_unpackhi_pi16(kernel.packet[0].x, kernel.packet[1].x);
+  __m64 T3 = _mm_unpackhi_pi16(kernel.packet[2].x, kernel.packet[3].x);
+
+  kernel.packet[0].x = _mm_unpacklo_pi32(T0, T1);
+  kernel.packet[1].x = _mm_unpackhi_pi32(T0, T1);
+  kernel.packet[2].x = _mm_unpacklo_pi32(T2, T3);
+  kernel.packet[3].x = _mm_unpackhi_pi32(T2, T3);
+}
+
+#endif
+
+}
+}
+
+#endif // EIGEN_PACKET_MATH_HALF_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h
new file mode 100644
index 0000000..aa5fbce
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/CUDA/TypeCasting.h
@@ -0,0 +1,212 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_CUDA_H
+#define EIGEN_TYPE_CASTING_CUDA_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<>
+struct scalar_cast_op<float, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __float2half(a);
+    #else
+      return Eigen::half(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<float, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<int, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __float2half(static_cast<float>(a));
+    #else
+      return Eigen::half(static_cast<float>(a));
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<int, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<Eigen::half, float> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef float result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
+    #if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+      return __half2float(a);
+    #else
+      return static_cast<float>(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<Eigen::half, float> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+
+#if defined(EIGEN_HAS_CUDA_FP16) && defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 300
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float4 pcast<half2, float4>(const half2& a, const half2& b) {
+  float2 r1 = __half22float2(a);
+  float2 r2 = __half22float2(b);
+  return make_float4(r1.x, r1.y, r2.x, r2.y);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pcast<float4, half2>(const float4& a) {
+  // Simply discard the second half of the input
+  return __floats2half2_rn(a.x, a.y);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX512
+template <>
+struct type_casting_traits<half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16f pcast<Packet16h, Packet16f>(const Packet16h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet16h pcast<Packet16f, Packet16h>(const Packet16f& a) {
+  return float2half(a);
+}
+
+#elif defined EIGEN_VECTORIZE_AVX
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8f pcast<Packet8h, Packet8f>(const Packet8h& a) {
+  return half2float(a);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet8h pcast<Packet8f, Packet8h>(const Packet8f& a) {
+  return float2half(a);
+}
+
+// Disable the following code since it's broken on too many platforms / compilers.
+//#elif defined(EIGEN_VECTORIZE_SSE) && (!EIGEN_ARCH_x86_64) && (!EIGEN_COMP_MSVC)
+#elif 0
+
+template <>
+struct type_casting_traits<Eigen::half, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4h, Packet4f>(const Packet4h& a) {
+  __int64_t a64 = _mm_cvtm64_si64(a.x);
+  Eigen::half h = raw_uint16_to_half(static_cast<unsigned short>(a64));
+  float f1 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 16));
+  float f2 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 32));
+  float f3 = static_cast<float>(h);
+  h = raw_uint16_to_half(static_cast<unsigned short>(a64 >> 48));
+  float f4 = static_cast<float>(h);
+  return _mm_set_ps(f4, f3, f2, f1);
+}
+
+template <>
+struct type_casting_traits<float, Eigen::half> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4h pcast<Packet4f, Packet4h>(const Packet4f& a) {
+  EIGEN_ALIGN16 float aux[4];
+  pstore(aux, a);
+  Eigen::half h0(aux[0]);
+  Eigen::half h1(aux[1]);
+  Eigen::half h2(aux[2]);
+  Eigen::half h3(aux[3]);
+
+  Packet4h result;
+  result.x = _mm_set_pi16(h3.x, h2.x, h1.x, h0.x);
+  return result;
+}
+
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_CUDA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h
new file mode 100644
index 0000000..097373c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/Default/Settings.h
@@ -0,0 +1,49 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+/* All the parameters defined in this file can be specialized in the
+ * architecture specific files, and/or by the user.
+ * More to come... */
+
+#ifndef EIGEN_DEFAULT_SETTINGS_H
+#define EIGEN_DEFAULT_SETTINGS_H
+
+/** Defines the maximal loop size to enable meta unrolling of loops.
+  * Note that the value here is expressed in Eigen's own notion of "number of FLOPS",
+  * it does not correspond to the number of iterations or the number of instructions
+  */
+#ifndef EIGEN_UNROLLING_LIMIT
+#define EIGEN_UNROLLING_LIMIT 100
+#endif
+
+/** Defines the threshold between a "small" and a "large" matrix.
+  * This threshold is mainly used to select the proper product implementation.
+  */
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+/** Defines the maximal width of the blocks used in the triangular product and solver
+  * for vectors (level 2 blas xTRMV and xTRSV). The default is 8.
+  */
+#ifndef EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH
+#define EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH 8
+#endif
+
+
+/** Defines the default number of registers available for that architecture.
+  * Currently it must be 8 or 16. Other values will fail.
+  */
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 8
+#endif
+
+#endif // EIGEN_DEFAULT_SETTINGS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h
new file mode 100644
index 0000000..57e9b43
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/Complex.h
@@ -0,0 +1,486 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_NEON_H
+#define EIGEN_COMPLEX_NEON_H
+
+namespace Eigen {
+
+namespace internal {
+
+inline uint32x4_t p4ui_CONJ_XOR() {
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  uint32x4_t ret = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+  return ret;
+#else
+  static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000, 0x00000000, 0x80000000 };
+  return vld1q_u32( conj_XOR_DATA );
+#endif
+}
+
+inline uint32x2_t p2ui_CONJ_XOR() {
+  static const uint32_t conj_XOR_DATA[] = { 0x00000000, 0x80000000 };
+  return vld1_u32( conj_XOR_DATA );
+}
+
+//---------- float ----------
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  Packet4f  v;
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  float32x2_t r64;
+  r64 = vld1_f32((float *)&from);
+
+  return Packet2cf(vcombine_f32(r64, r64));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate<Packet4f>(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet4ui b = vreinterpretq_u32_f32(a.v);
+  return Packet2cf(vreinterpretq_f32_u32(veorq_u32(b, p4ui_CONJ_XOR())));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet4f v1, v2;
+
+  // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
+  v1 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 0), vdup_lane_f32(vget_high_f32(a.v), 0));
+  // Get the imag values of a | a1_im | a1_im | a2_im | a2_im |
+  v2 = vcombine_f32(vdup_lane_f32(vget_low_f32(a.v), 1), vdup_lane_f32(vget_high_f32(a.v), 1));
+  // Multiply the real a with b
+  v1 = vmulq_f32(v1, b.v);
+  // Multiply the imag a with b
+  v2 = vmulq_f32(v2, b.v);
+  // Conjugate v2 
+  v2 = vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(v2), p4ui_CONJ_XOR()));
+  // Swap real/imag elements in v2.
+  v2 = vrev64q_f32(v2);
+  // Add and return the result
+  return Packet2cf(vaddq_f32(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  return Packet2cf(vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  return Packet2cf(vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  return Packet2cf(vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  return Packet2cf(vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a.v),vreinterpretq_u32_f32(b.v))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  Packet4f res = pset1<Packet4f>(0.f);
+  res = vsetq_lane_f32(std::real(from[0*stride]), res, 0);
+  res = vsetq_lane_f32(std::imag(from[0*stride]), res, 1);
+  res = vsetq_lane_f32(std::real(from[1*stride]), res, 2);
+  res = vsetq_lane_f32(std::imag(from[1*stride]), res, 3);
+  return Packet2cf(res);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  to[stride*0] = std::complex<float>(vgetq_lane_f32(from.v, 0), vgetq_lane_f32(from.v, 1));
+  to[stride*1] = std::complex<float>(vgetq_lane_f32(from.v, 2), vgetq_lane_f32(from.v, 3));
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { EIGEN_ARM_PREFETCH((float *)addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 x[2];
+  vst1q_f32((float *)x, a.v);
+  return x[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  float32x2_t a_lo, a_hi;
+  Packet4f a_r128;
+
+  a_lo = vget_low_f32(a.v);
+  a_hi = vget_high_f32(a.v);
+  a_r128 = vcombine_f32(a_hi, a_lo);
+
+  return Packet2cf(a_r128);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pcplxflip<Packet2cf>(const Packet2cf& a)
+{
+  return Packet2cf(vrev64q_f32(a.v));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  float32x2_t a1, a2;
+  std::complex<float> s;
+
+  a1 = vget_low_f32(a.v);
+  a2 = vget_high_f32(a.v);
+  a2 = vadd_f32(a1, a2);
+  vst1_f32((float *)&s, a2);
+
+  return s;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  Packet4f sum1, sum2, sum;
+
+  // Add the first two 64-bit float32x2_t of vecs[0]
+  sum1 = vcombine_f32(vget_low_f32(vecs[0].v), vget_low_f32(vecs[1].v));
+  sum2 = vcombine_f32(vget_high_f32(vecs[0].v), vget_high_f32(vecs[1].v));
+  sum = vaddq_f32(sum1, sum2);
+
+  return Packet2cf(sum);
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  float32x2_t a1, a2, v1, v2, prod;
+  std::complex<float> s;
+
+  a1 = vget_low_f32(a.v);
+  a2 = vget_high_f32(a.v);
+   // Get the real values of a | a1_re | a1_re | a2_re | a2_re |
+  v1 = vdup_lane_f32(a1, 0);
+  // Get the real values of a | a1_im | a1_im | a2_im | a2_im |
+  v2 = vdup_lane_f32(a1, 1);
+  // Multiply the real a with b
+  v1 = vmul_f32(v1, a2);
+  // Multiply the imag a with b
+  v2 = vmul_f32(v2, a2);
+  // Conjugate v2 
+  v2 = vreinterpret_f32_u32(veor_u32(vreinterpret_u32_f32(v2), p2ui_CONJ_XOR()));
+  // Swap real/imag elements in v2.
+  v2 = vrev64_f32(v2);
+  // Add v1, v2
+  prod = vadd_f32(v1, v2);
+
+  vst1_f32((float *)&s, prod);
+
+  return s;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  EIGEN_STRONG_INLINE static void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset==1)
+    {
+      first.v = vextq_f32(first.v, second.v, 2);
+    }
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for NEON
+  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
+  Packet4f s, rev_s;
+
+  // this computes the norm
+  s = vmulq_f32(b.v, b.v);
+  rev_s = vrev64q_f32(s);
+
+  return Packet2cf(pdiv(res.v, vaddq_f32(s,rev_s)));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cf,2>& kernel) {
+  Packet4f tmp = vcombine_f32(vget_high_f32(kernel.packet[0].v), vget_high_f32(kernel.packet[1].v));
+  kernel.packet[0].v = vcombine_f32(vget_low_f32(kernel.packet[0].v), vget_low_f32(kernel.packet[1].v));
+  kernel.packet[1].v = tmp;
+}
+
+//---------- double ----------
+#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
+
+// See bug 1325, clang fails to call vld1q_u64.
+#if EIGEN_COMP_CLANG
+  static uint64x2_t p2ul_CONJ_XOR = {0x0, 0x8000000000000000};
+#else
+  const uint64_t  p2ul_conj_XOR_DATA[] = { 0x0, 0x8000000000000000 };
+  static uint64x2_t p2ul_CONJ_XOR = vld1q_u64( p2ul_conj_XOR_DATA );
+#endif
+
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd pload<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(padd<Packet2d>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(psub<Packet2d>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate<Packet2d>(a.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v), p2ul_CONJ_XOR))); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d v1, v2;
+
+  // Get the real values of a 
+  v1 = vdupq_lane_f64(vget_low_f64(a.v), 0);
+  // Get the imag values of a
+  v2 = vdupq_lane_f64(vget_high_f64(a.v), 0);
+  // Multiply the real a with b
+  v1 = vmulq_f64(v1, b.v);
+  // Multiply the imag a with b
+  v2 = vmulq_f64(v2, b.v);
+  // Conjugate v2 
+  v2 = vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(v2), p2ul_CONJ_XOR));
+  // Swap real/imag elements in v2.
+  v2 = preverse<Packet2d>(v2);
+  // Add and return the result
+  return Packet1cd(vaddq_f64(v1, v2));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  return Packet1cd(vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a.v),vreinterpretq_u64_f64(b.v))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ARM_PREFETCH((double *)addr); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride)
+{
+  Packet2d res = pset1<Packet2d>(0.0);
+  res = vsetq_lane_f64(std::real(from[0*stride]), res, 0);
+  res = vsetq_lane_f64(std::imag(from[0*stride]), res, 1);
+  return Packet1cd(res);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride)
+{
+  to[stride*0] = std::complex<double>(vgetq_lane_f64(from.v, 0), vgetq_lane_f64(from.v, 1));
+}
+
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs) { return vecs[0]; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a) { return pfirst(a); }
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for NEON
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = pmul<Packet2d>(b.v, b.v);
+  Packet2d rev_s = preverse<Packet2d>(s);
+
+  return Packet1cd(pdiv(res.v, padd<Packet2d>(s,rev_s)));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vcombine_f64(vget_high_f64(kernel.packet[0].v), vget_high_f64(kernel.packet[1].v));
+  kernel.packet[0].v = vcombine_f64(vget_low_f64(kernel.packet[0].v), vget_low_f64(kernel.packet[1].v));
+  kernel.packet[1].v = tmp;
+}
+#endif // EIGEN_ARCH_ARM64
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_NEON_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h
new file mode 100644
index 0000000..6bb05bb
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/MathFunctions.h
@@ -0,0 +1,91 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_NEON_H
+#define EIGEN_MATH_FUNCTIONS_NEON_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  Packet4f tmp, fx;
+
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+  x = vminq_f32(x, p4f_exp_hi);
+  x = vmaxq_f32(x, p4f_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = vmlaq_f32(p4f_half, x, p4f_cephes_LOG2EF);
+
+  /* perform a floorf */
+  tmp = vcvtq_f32_s32(vcvtq_s32_f32(fx));
+
+  /* if greater, substract 1 */
+  Packet4ui mask = vcgtq_f32(tmp, fx);
+  mask = vandq_u32(mask, vreinterpretq_u32_f32(p4f_1));
+
+  fx = vsubq_f32(tmp, vreinterpretq_f32_u32(mask));
+
+  tmp = vmulq_f32(fx, p4f_cephes_exp_C1);
+  Packet4f z = vmulq_f32(fx, p4f_cephes_exp_C2);
+  x = vsubq_f32(x, tmp);
+  x = vsubq_f32(x, z);
+
+  Packet4f y = vmulq_f32(p4f_cephes_exp_p0, x);
+  z = vmulq_f32(x, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p1);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p2);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p3);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p4);
+  y = vmulq_f32(y, x);
+  y = vaddq_f32(y, p4f_cephes_exp_p5);
+
+  y = vmulq_f32(y, z);
+  y = vaddq_f32(y, x);
+  y = vaddq_f32(y, p4f_1);
+
+  /* build 2^n */
+  int32x4_t mm;
+  mm = vcvtq_s32_f32(fx);
+  mm = vaddq_s32(mm, p4i_0x7f);
+  mm = vshlq_n_s32(mm, 23);
+  Packet4f pow2n = vreinterpretq_f32_s32(mm);
+
+  y = vmulq_f32(y, pow2n);
+  return y;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_NEON_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h
new file mode 100644
index 0000000..836fbc0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/NEON/PacketMath.h
@@ -0,0 +1,729 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Konstantinos Margaritis <markos@freevec.org>
+// Heavily based on Gael's SSE version.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_NEON_H
+#define EIGEN_PACKET_MATH_NEON_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#if EIGEN_ARCH_ARM64
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 32
+#else
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS 16 
+#endif
+#endif
+
+typedef float32x2_t Packet2f;
+typedef float32x4_t Packet4f;
+typedef int32x4_t   Packet4i;
+typedef int32x2_t   Packet2i;
+typedef uint32x4_t  Packet4ui;
+
+#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+  const Packet4f p4f_##NAME = pset1<Packet4f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = vreinterpretq_f32_u32(pset1<int32_t>(X))
+
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+  const Packet4i p4i_##NAME = pset1<Packet4i>(X)
+
+#if EIGEN_ARCH_ARM64
+  // __builtin_prefetch tends to do nothing on ARM64 compilers because the
+  // prefetch instructions there are too detailed for __builtin_prefetch to map
+  // meaningfully to them.
+  #define EIGEN_ARM_PREFETCH(ADDR)  __asm__ __volatile__("prfm pldl1keep, [%[addr]]\n" ::[addr] "r"(ADDR) : );
+#elif EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+  #define EIGEN_ARM_PREFETCH(ADDR) __builtin_prefetch(ADDR);
+#elif defined __pld
+  #define EIGEN_ARM_PREFETCH(ADDR) __pld(ADDR)
+#elif EIGEN_ARCH_ARM32
+  #define EIGEN_ARM_PREFETCH(ADDR) __asm__ __volatile__ ("pld [%[addr]]\n" :: [addr] "r" (ADDR) : );
+#else
+  // by default no explicit prefetching
+  #define EIGEN_ARM_PREFETCH(ADDR)
+#endif
+
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half; // Packet2f intrinsics not implemented yet
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket=0, // Packet2f intrinsics not implemented yet
+   
+    HasDiv  = 1,
+    // FIXME check the Has*
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 0
+  };
+};
+template<> struct packet_traits<int32_t>    : default_packet_traits
+{
+  typedef Packet4i type;
+  typedef Packet4i half; // Packet2i intrinsics not implemented yet
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket=0 // Packet2i intrinsics not implemented yet
+    // FIXME check the Has*
+  };
+};
+
+#if EIGEN_GNUC_AT_MOST(4,4) && !EIGEN_COMP_LLVM
+// workaround gcc 4.2, 4.3 and 4.4 compilatin issue
+EIGEN_STRONG_INLINE float32x4_t vld1q_f32(const float* x) { return ::vld1q_f32((const float32_t*)x); }
+EIGEN_STRONG_INLINE float32x2_t vld1_f32 (const float* x) { return ::vld1_f32 ((const float32_t*)x); }
+EIGEN_STRONG_INLINE float32x2_t vld1_dup_f32 (const float* x) { return ::vld1_dup_f32 ((const float32_t*)x); }
+EIGEN_STRONG_INLINE void        vst1q_f32(float* to, float32x4_t from) { ::vst1q_f32((float32_t*)to,from); }
+EIGEN_STRONG_INLINE void        vst1_f32 (float* to, float32x2_t from) { ::vst1_f32 ((float32_t*)to,from); }
+#endif
+
+template<> struct unpacket_traits<Packet4f> { typedef float   type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet4i> { typedef int32_t type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return vdupq_n_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int32_t&    from)   { return vdupq_n_s32(from); }
+
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)
+{
+  const float f[] = {0, 1, 2, 3};
+  Packet4f countdown = vld1q_f32(f);
+  return vaddq_f32(pset1<Packet4f>(a), countdown);
+}
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int32_t& a)
+{
+  const int32_t i[] = {0, 1, 2, 3};
+  Packet4i countdown = vld1q_s32(i);
+  return vaddq_s32(pset1<Packet4i>(a), countdown);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return vaddq_f32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return vaddq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return vsubq_f32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return vsubq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a) { return vnegq_f32(a); }
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return vnegq_s32(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmulq_f32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmulq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+#if EIGEN_ARCH_ARM64
+  return vdivq_f32(a,b);
+#else
+  Packet4f inv, restep, div;
+
+  // NEON does not offer a divide instruction, we have to do a reciprocal approximation
+  // However NEON in contrast to other SIMD engines (AltiVec/SSE), offers
+  // a reciprocal estimate AND a reciprocal step -which saves a few instructions
+  // vrecpeq_f32() returns an estimate to 1/b, which we will finetune with
+  // Newton-Raphson and vrecpsq_f32()
+  inv = vrecpeq_f32(b);
+
+  // This returns a differential, by which we will have to multiply inv to get a better
+  // approximation of 1/b.
+  restep = vrecpsq_f32(b, inv);
+  inv = vmulq_f32(restep, inv);
+
+  // Finally, multiply a by 1/b and get the wanted result of the division.
+  div = vmulq_f32(a, inv);
+
+  return div;
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& /*a*/, const Packet4i& /*b*/)
+{ eigen_assert(false && "packet integer division are not supported by NEON");
+  return pset1<Packet4i>(0);
+}
+
+// Clang/ARM wrongly advertises __ARM_FEATURE_FMA even when it's not available,
+// then implements a slow software scalar fallback calling fmaf()!
+// Filed LLVM bug:
+//     https://llvm.org/bugs/show_bug.cgi?id=27216
+#if (defined __ARM_FEATURE_FMA) && !(EIGEN_COMP_CLANG && EIGEN_ARCH_ARM)
+// See bug 936.
+// FMA is available on VFPv4 i.e. when compiling with -mfpu=neon-vfpv4.
+// FMA is a true fused multiply-add i.e. only 1 rounding at the end, no intermediate rounding.
+// MLA is not fused i.e. does 2 roundings.
+// In addition to giving better accuracy, FMA also gives better performance here on a Krait (Nexus 4):
+// MLA: 10 GFlop/s ; FMA: 12 GFlops/s.
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return vfmaq_f32(c,a,b); }
+#else
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) {
+#if EIGEN_COMP_CLANG && EIGEN_ARCH_ARM
+  // Clang/ARM will replace VMLA by VMUL+VADD at least for some values of -mcpu,
+  // at least -mcpu=cortex-a8 and -mcpu=cortex-a7. Since the former is the default on
+  // -march=armv7-a, that is a very common case.
+  // See e.g. this thread:
+  //     http://lists.llvm.org/pipermail/llvm-dev/2013-December/068806.html
+  // Filed LLVM bug:
+  //     https://llvm.org/bugs/show_bug.cgi?id=27219
+  Packet4f r = c;
+  asm volatile(
+    "vmla.f32 %q[r], %q[a], %q[b]"
+    : [r] "+w" (r)
+    : [a] "w" (a),
+      [b] "w" (b)
+    : );
+  return r;
+#else
+  return vmlaq_f32(c,a,b);
+#endif
+}
+#endif
+
+// No FMA instruction for int, so use MLA unconditionally.
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return vmlaq_s32(c,a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return vminq_f32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vminq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return vmaxq_f32(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vmaxq_s32(a,b); }
+
+// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  return vreinterpretq_f32_u32(vandq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
+}
+template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vandq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  return vreinterpretq_f32_u32(vorrq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
+}
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vorrq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  return vreinterpretq_f32_u32(veorq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
+}
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return veorq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  return vreinterpretq_f32_u32(vbicq_u32(vreinterpretq_u32_f32(a),vreinterpretq_u32_f32(b)));
+}
+template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return vbicq_s32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*    from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int32_t*  from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_s32(from); }
+
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*   from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f32(from); }
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int32_t* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_s32(from); }
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float* from)
+{
+  float32x2_t lo, hi;
+  lo = vld1_dup_f32(from);
+  hi = vld1_dup_f32(from+1);
+  return vcombine_f32(lo, hi);
+}
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int32_t* from)
+{
+  int32x2_t lo, hi;
+  lo = vld1_dup_s32(from);
+  hi = vld1_dup_s32(from+1);
+  return vcombine_s32(lo, hi);
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>  (float*    to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int32_t>(int32_t*  to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_s32(to, from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<float>  (float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int32_t>(int32_t* to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  Packet4f res = pset1<Packet4f>(0.f);
+  res = vsetq_lane_f32(from[0*stride], res, 0);
+  res = vsetq_lane_f32(from[1*stride], res, 1);
+  res = vsetq_lane_f32(from[2*stride], res, 2);
+  res = vsetq_lane_f32(from[3*stride], res, 3);
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int32_t, Packet4i>(const int32_t* from, Index stride)
+{
+  Packet4i res = pset1<Packet4i>(0);
+  res = vsetq_lane_s32(from[0*stride], res, 0);
+  res = vsetq_lane_s32(from[1*stride], res, 1);
+  res = vsetq_lane_s32(from[2*stride], res, 2);
+  res = vsetq_lane_s32(from[3*stride], res, 3);
+  return res;
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_f32(from, 0);
+  to[stride*1] = vgetq_lane_f32(from, 1);
+  to[stride*2] = vgetq_lane_f32(from, 2);
+  to[stride*3] = vgetq_lane_f32(from, 3);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int32_t, Packet4i>(int32_t* to, const Packet4i& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_s32(from, 0);
+  to[stride*1] = vgetq_lane_s32(from, 1);
+  to[stride*2] = vgetq_lane_s32(from, 2);
+  to[stride*3] = vgetq_lane_s32(from, 3);
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<float>  (const float*    addr) { EIGEN_ARM_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int32_t>(const int32_t*  addr) { EIGEN_ARM_PREFETCH(addr); }
+
+// FIXME only store the 2 first elements ?
+template<> EIGEN_STRONG_INLINE float   pfirst<Packet4f>(const Packet4f& a) { float   EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE int32_t pfirst<Packet4i>(const Packet4i& a) { int32_t EIGEN_ALIGN16 x[4]; vst1q_s32(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a) {
+  float32x2_t a_lo, a_hi;
+  Packet4f a_r64;
+
+  a_r64 = vrev64q_f32(a);
+  a_lo = vget_low_f32(a_r64);
+  a_hi = vget_high_f32(a_r64);
+  return vcombine_f32(a_hi, a_lo);
+}
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a) {
+  int32x2_t a_lo, a_hi;
+  Packet4i a_r64;
+
+  a_r64 = vrev64q_s32(a);
+  a_lo = vget_low_s32(a_r64);
+  a_hi = vget_high_s32(a_r64);
+  return vcombine_s32(a_hi, a_lo);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a) { return vabsq_f32(a); }
+template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a) { return vabsq_s32(a); }
+
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  float32x2_t a_lo, a_hi, sum;
+
+  a_lo = vget_low_f32(a);
+  a_hi = vget_high_f32(a);
+  sum = vpadd_f32(a_lo, a_hi);
+  sum = vpadd_f32(sum, sum);
+  return vget_lane_f32(sum, 0);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  float32x4x2_t vtrn1, vtrn2, res1, res2;
+  Packet4f sum1, sum2, sum;
+
+  // NEON zip performs interleaving of the supplied vectors.
+  // We perform two interleaves in a row to acquire the transposed vector
+  vtrn1 = vzipq_f32(vecs[0], vecs[2]);
+  vtrn2 = vzipq_f32(vecs[1], vecs[3]);
+  res1 = vzipq_f32(vtrn1.val[0], vtrn2.val[0]);
+  res2 = vzipq_f32(vtrn1.val[1], vtrn2.val[1]);
+
+  // Do the addition of the resulting vectors
+  sum1 = vaddq_f32(res1.val[0], res1.val[1]);
+  sum2 = vaddq_f32(res2.val[0], res2.val[1]);
+  sum = vaddq_f32(sum1, sum2);
+
+  return sum;
+}
+
+template<> EIGEN_STRONG_INLINE int32_t predux<Packet4i>(const Packet4i& a)
+{
+  int32x2_t a_lo, a_hi, sum;
+
+  a_lo = vget_low_s32(a);
+  a_hi = vget_high_s32(a);
+  sum = vpadd_s32(a_lo, a_hi);
+  sum = vpadd_s32(sum, sum);
+  return vget_lane_s32(sum, 0);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  int32x4x2_t vtrn1, vtrn2, res1, res2;
+  Packet4i sum1, sum2, sum;
+
+  // NEON zip performs interleaving of the supplied vectors.
+  // We perform two interleaves in a row to acquire the transposed vector
+  vtrn1 = vzipq_s32(vecs[0], vecs[2]);
+  vtrn2 = vzipq_s32(vecs[1], vecs[3]);
+  res1 = vzipq_s32(vtrn1.val[0], vtrn2.val[0]);
+  res2 = vzipq_s32(vtrn1.val[1], vtrn2.val[1]);
+
+  // Do the addition of the resulting vectors
+  sum1 = vaddq_s32(res1.val[0], res1.val[1]);
+  sum2 = vaddq_s32(res2.val[0], res2.val[1]);
+  sum = vaddq_s32(sum1, sum2);
+
+  return sum;
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  float32x2_t a_lo, a_hi, prod;
+
+  // Get a_lo = |a1|a2| and a_hi = |a3|a4|
+  a_lo = vget_low_f32(a);
+  a_hi = vget_high_f32(a);
+  // Get the product of a_lo * a_hi -> |a1*a3|a2*a4|
+  prod = vmul_f32(a_lo, a_hi);
+  // Multiply prod with its swapped value |a2*a4|a1*a3|
+  prod = vmul_f32(prod, vrev64_f32(prod));
+
+  return vget_lane_f32(prod, 0);
+}
+template<> EIGEN_STRONG_INLINE int32_t predux_mul<Packet4i>(const Packet4i& a)
+{
+  int32x2_t a_lo, a_hi, prod;
+
+  // Get a_lo = |a1|a2| and a_hi = |a3|a4|
+  a_lo = vget_low_s32(a);
+  a_hi = vget_high_s32(a);
+  // Get the product of a_lo * a_hi -> |a1*a3|a2*a4|
+  prod = vmul_s32(a_lo, a_hi);
+  // Multiply prod with its swapped value |a2*a4|a1*a3|
+  prod = vmul_s32(prod, vrev64_s32(prod));
+
+  return vget_lane_s32(prod, 0);
+}
+
+// min
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  float32x2_t a_lo, a_hi, min;
+
+  a_lo = vget_low_f32(a);
+  a_hi = vget_high_f32(a);
+  min = vpmin_f32(a_lo, a_hi);
+  min = vpmin_f32(min, min);
+
+  return vget_lane_f32(min, 0);
+}
+
+template<> EIGEN_STRONG_INLINE int32_t predux_min<Packet4i>(const Packet4i& a)
+{
+  int32x2_t a_lo, a_hi, min;
+
+  a_lo = vget_low_s32(a);
+  a_hi = vget_high_s32(a);
+  min = vpmin_s32(a_lo, a_hi);
+  min = vpmin_s32(min, min);
+  
+  return vget_lane_s32(min, 0);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  float32x2_t a_lo, a_hi, max;
+
+  a_lo = vget_low_f32(a);
+  a_hi = vget_high_f32(a);
+  max = vpmax_f32(a_lo, a_hi);
+  max = vpmax_f32(max, max);
+
+  return vget_lane_f32(max, 0);
+}
+
+template<> EIGEN_STRONG_INLINE int32_t predux_max<Packet4i>(const Packet4i& a)
+{
+  int32x2_t a_lo, a_hi, max;
+
+  a_lo = vget_low_s32(a);
+  a_hi = vget_high_s32(a);
+  max = vpmax_s32(a_lo, a_hi);
+  max = vpmax_s32(max, max);
+
+  return vget_lane_s32(max, 0);
+}
+
+// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors,
+// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074
+#define PALIGN_NEON(Offset,Type,Command) \
+template<>\
+struct palign_impl<Offset,Type>\
+{\
+    EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\
+    {\
+        if (Offset!=0)\
+            first = Command(first, second, Offset);\
+    }\
+};\
+
+PALIGN_NEON(0,Packet4f,vextq_f32)
+PALIGN_NEON(1,Packet4f,vextq_f32)
+PALIGN_NEON(2,Packet4f,vextq_f32)
+PALIGN_NEON(3,Packet4f,vextq_f32)
+PALIGN_NEON(0,Packet4i,vextq_s32)
+PALIGN_NEON(1,Packet4i,vextq_s32)
+PALIGN_NEON(2,Packet4i,vextq_s32)
+PALIGN_NEON(3,Packet4i,vextq_s32)
+
+#undef PALIGN_NEON
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  float32x4x2_t tmp1 = vzipq_f32(kernel.packet[0], kernel.packet[1]);
+  float32x4x2_t tmp2 = vzipq_f32(kernel.packet[2], kernel.packet[3]);
+
+  kernel.packet[0] = vcombine_f32(vget_low_f32(tmp1.val[0]), vget_low_f32(tmp2.val[0]));
+  kernel.packet[1] = vcombine_f32(vget_high_f32(tmp1.val[0]), vget_high_f32(tmp2.val[0]));
+  kernel.packet[2] = vcombine_f32(vget_low_f32(tmp1.val[1]), vget_low_f32(tmp2.val[1]));
+  kernel.packet[3] = vcombine_f32(vget_high_f32(tmp1.val[1]), vget_high_f32(tmp2.val[1]));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  int32x4x2_t tmp1 = vzipq_s32(kernel.packet[0], kernel.packet[1]);
+  int32x4x2_t tmp2 = vzipq_s32(kernel.packet[2], kernel.packet[3]);
+  kernel.packet[0] = vcombine_s32(vget_low_s32(tmp1.val[0]), vget_low_s32(tmp2.val[0]));
+  kernel.packet[1] = vcombine_s32(vget_high_s32(tmp1.val[0]), vget_high_s32(tmp2.val[0]));
+  kernel.packet[2] = vcombine_s32(vget_low_s32(tmp1.val[1]), vget_low_s32(tmp2.val[1]));
+  kernel.packet[3] = vcombine_s32(vget_high_s32(tmp1.val[1]), vget_high_s32(tmp2.val[1]));
+}
+
+//---------- double ----------
+
+// Clang 3.5 in the iOS toolchain has an ICE triggered by NEON intrisics for double.
+// Confirmed at least with __apple_build_version__ = 6000054.
+#ifdef __apple_build_version__
+// Let's hope that by the time __apple_build_version__ hits the 601* range, the bug will be fixed.
+// https://gist.github.com/yamaya/2924292 suggests that the 3 first digits are only updated with
+// major toolchain updates.
+#define EIGEN_APPLE_DOUBLE_NEON_BUG (__apple_build_version__ < 6010000)
+#else
+#define EIGEN_APPLE_DOUBLE_NEON_BUG 0
+#endif
+
+#if EIGEN_ARCH_ARM64 && !EIGEN_APPLE_DOUBLE_NEON_BUG
+
+// Bug 907: workaround missing declarations of the following two functions in the ADK
+// Defining these functions as templates ensures that if these intrinsics are
+// already defined in arm_neon.h, then our workaround doesn't cause a conflict
+// and has lower priority in overload resolution.
+template <typename T>
+uint64x2_t vreinterpretq_u64_f64(T a)
+{
+  return (uint64x2_t) a;
+}
+
+template <typename T>
+float64x2_t vreinterpretq_f64_u64(T a)
+{
+  return (float64x2_t) a;
+}
+
+typedef float64x2_t Packet2d;
+typedef float64x1_t Packet1d;
+
+template<> struct packet_traits<double>  : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket=0,
+   
+    HasDiv  = 1,
+    // FIXME check the Has*
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 0,
+    HasSqrt = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2d> { typedef double  type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double&  from) { return vdupq_n_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a)
+{
+  const double countdown_raw[] = {0.0,1.0};
+  const Packet2d countdown = vld1q_f64(countdown_raw);
+  return vaddq_f64(pset1<Packet2d>(a), countdown);
+}
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return vaddq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return vsubq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return vnegq_f64(a); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmulq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return vdivq_f64(a,b); }
+
+#ifdef __ARM_FEATURE_FMA
+// See bug 936. See above comment about FMA for float.
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vfmaq_f64(c,a,b); }
+#else
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vmlaq_f64(c,a,b); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vminq_f64(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vmaxq_f64(a,b); }
+
+// Logical Operations are not supported for float, so we have to reinterpret casts using NEON intrinsics
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vandq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vorrq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(veorq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b)
+{
+  return vreinterpretq_f64_u64(vbicq_u64(vreinterpretq_u64_f64(a),vreinterpretq_u64_f64(b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from) { EIGEN_DEBUG_ALIGNED_LOAD return vld1q_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return vld1q_f64(from); }
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  return vld1q_dup_f64(from);
+}
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE vst1q_f64(to, from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f64(to, from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  Packet2d res = pset1<Packet2d>(0.0);
+  res = vsetq_lane_f64(from[0*stride], res, 0);
+  res = vsetq_lane_f64(from[1*stride], res, 1);
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  to[stride*0] = vgetq_lane_f64(from, 0);
+  to[stride*1] = vgetq_lane_f64(from, 1);
+}
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ARM_PREFETCH(addr); }
+
+// FIXME only store the 2 first elements ?
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(a, 0); }
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a) { return vcombine_f64(vget_high_f64(a), vget_low_f64(a)); }
+
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a) { return vabsq_f64(a); }
+
+#if EIGEN_COMP_CLANG && defined(__apple_build_version__)
+// workaround ICE, see bug 907
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) + vget_high_f64(a))[0]; }
+#else
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) + vget_high_f64(a), 0); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  float64x2_t trn1, trn2;
+
+  // NEON zip performs interleaving of the supplied vectors.
+  // We perform two interleaves in a row to acquire the transposed vector
+  trn1 = vzip1q_f64(vecs[0], vecs[1]);
+  trn2 = vzip2q_f64(vecs[0], vecs[1]);
+
+  // Do the addition of the resulting vectors
+  return vaddq_f64(trn1, trn2);
+}
+// Other reduction functions:
+// mul
+#if EIGEN_COMP_CLANG && defined(__apple_build_version__)
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return (vget_low_f64(a) * vget_high_f64(a))[0]; }
+#else
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a) { return vget_lane_f64(vget_low_f64(a) * vget_high_f64(a), 0); }
+#endif
+
+// min
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpminq_f64(a, a), 0); }
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a) { return vgetq_lane_f64(vpmaxq_f64(a, a), 0); }
+
+// this PALIGN_NEON business is to work around a bug in LLVM Clang 3.0 causing incorrect compilation errors,
+// see bug 347 and this LLVM bug: http://llvm.org/bugs/show_bug.cgi?id=11074
+#define PALIGN_NEON(Offset,Type,Command) \
+template<>\
+struct palign_impl<Offset,Type>\
+{\
+    EIGEN_STRONG_INLINE static void run(Type& first, const Type& second)\
+    {\
+        if (Offset!=0)\
+            first = Command(first, second, Offset);\
+    }\
+};\
+
+PALIGN_NEON(0,Packet2d,vextq_f64)
+PALIGN_NEON(1,Packet2d,vextq_f64)
+#undef PALIGN_NEON
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  float64x2_t trn1 = vzip1q_f64(kernel.packet[0], kernel.packet[1]);
+  float64x2_t trn2 = vzip2q_f64(kernel.packet[0], kernel.packet[1]);
+
+  kernel.packet[0] = trn1;
+  kernel.packet[1] = trn2;
+}
+#endif // EIGEN_ARCH_ARM64 
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_NEON_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h
new file mode 100644
index 0000000..5607fe0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/Complex.h
@@ -0,0 +1,503 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_SSE_H
+#define EIGEN_COMPLEX_SSE_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- float ----------
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const __m128& a) : v(a) {}
+  __m128  v;
+};
+
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0,
+    HasBlend = 1
+  };
+};
+#endif
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float> type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_add_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_sub_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a)
+{
+  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
+  return Packet2cf(_mm_xor_ps(a.v,mask));
+}
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
+  return Packet2cf(_mm_xor_ps(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  #ifdef EIGEN_VECTORIZE_SSE3
+  return Packet2cf(_mm_addsub_ps(_mm_mul_ps(_mm_moveldup_ps(a.v), b.v),
+                                 _mm_mul_ps(_mm_movehdup_ps(a.v),
+                                            vec4f_swizzle1(b.v, 1, 0, 3, 2))));
+//   return Packet2cf(_mm_addsub_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
+//                                  _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
+//                                             vec4f_swizzle1(b.v, 1, 0, 3, 2))));
+  #else
+  const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x00000000,0x80000000,0x00000000));
+  return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
+                              _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
+                                                    vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
+  #endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_and_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_or_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_xor_ps(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(_mm_andnot_ps(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>(&numext::real_ref(*from))); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>(&numext::real_ref(*from))); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+#if EIGEN_GNUC_AT_MOST(4,2)
+  // Workaround annoying "may be used uninitialized in this function" warning with gcc 4.2
+  res.v = _mm_loadl_pi(_mm_set1_ps(0.0f), reinterpret_cast<const __m64*>(&from));
+#elif EIGEN_GNUC_AT_LEAST(4,6)
+  // Suppress annoying "may be used uninitialized in this function" warning with gcc >= 4.6
+  #pragma GCC diagnostic push
+  #pragma GCC diagnostic ignored "-Wuninitialized"
+  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
+  #pragma GCC diagnostic pop
+#else
+  res.v = _mm_loadl_pi(res.v, (const __m64*)&from);
+#endif
+  return Packet2cf(_mm_movelh_ps(res.v,res.v));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>* from) { return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore(&numext::real_ref(*to), Packet4f(from.v)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu(&numext::real_ref(*to), Packet4f(from.v)); }
+
+
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  return Packet2cf(_mm_set_ps(std::imag(from[1*stride]), std::real(from[1*stride]),
+                              std::imag(from[0*stride]), std::real(from[0*stride])));
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  to[stride*0] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 0)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 1)));
+  to[stride*1] = std::complex<float>(_mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 2)),
+                                     _mm_cvtss_f32(_mm_shuffle_ps(from.v, from.v, 3)));
+}
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  #if EIGEN_GNUC_AT_MOST(4,3)
+  // Workaround gcc 4.2 ICE - this is not performance wise ideal, but who cares...
+  // This workaround also fix invalid code generation with gcc 4.3
+  EIGEN_ALIGN16 std::complex<float> res[2];
+  _mm_store_ps((float*)res, a.v);
+  return res[0];
+  #else
+  std::complex<float> res;
+  _mm_storel_pi((__m64*)&res, a.v);
+  return res;
+  #endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a) { return Packet2cf(_mm_castpd_ps(preverse(Packet2d(_mm_castps_pd(a.v))))); }
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  return pfirst(Packet2cf(_mm_add_ps(a.v, _mm_movehl_ps(a.v,a.v))));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  return Packet2cf(_mm_add_ps(_mm_movelh_ps(vecs[0].v,vecs[1].v), _mm_movehl_ps(vecs[1].v,vecs[0].v)));
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  return pfirst(pmul(a, Packet2cf(_mm_movehl_ps(a.v,a.v))));
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset==1)
+    {
+      first.v = _mm_movehl_ps(first.v, first.v);
+      first.v = _mm_movelh_ps(first.v, second.v);
+    }
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return internal::pmul(a, pconj(b));
+    #else
+    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
+    return Packet2cf(_mm_add_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
+                                _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
+                                           vec4f_swizzle1(b.v, 1, 0, 3, 2))));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return internal::pmul(pconj(a), b);
+    #else
+    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
+    return Packet2cf(_mm_add_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v),
+                                _mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
+                                                      vec4f_swizzle1(b.v, 1, 0, 3, 2)), mask)));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return pconj(internal::pmul(a, b));
+    #else
+    const __m128 mask = _mm_castsi128_ps(_mm_setr_epi32(0x00000000,0x80000000,0x00000000,0x80000000));
+    return Packet2cf(_mm_sub_ps(_mm_xor_ps(_mm_mul_ps(vec4f_swizzle1(a.v, 0, 0, 2, 2), b.v), mask),
+                                _mm_mul_ps(vec4f_swizzle1(a.v, 1, 1, 3, 3),
+                                           vec4f_swizzle1(b.v, 1, 0, 3, 2))));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet4f, Packet2cf, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet4f& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet4f& x, const Packet2cf& y) const
+  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet2cf, Packet4f, false,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet4f& y, const Packet2cf& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& x, const Packet4f& y) const
+  { return Packet2cf(Eigen::internal::pmul<Packet4f>(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for SSE3 and 4
+  Packet2cf res = conj_helper<Packet2cf,Packet2cf,false,true>().pmul(a,b);
+  __m128 s = _mm_mul_ps(b.v,b.v);
+  return Packet2cf(_mm_div_ps(res.v,_mm_add_ps(s,_mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(s), 0xb1)))));
+}
+
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/* <Packet2cf> */(const Packet2cf& x)
+{
+  return Packet2cf(vec4f_swizzle1(x.v, 1, 0, 3, 2));
+}
+
+
+//---------- double ----------
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const __m128d& a) : v(a) {}
+  __m128d  v;
+};
+
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 0,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+#endif
+
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_add_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_sub_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a)
+{
+  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
+  return Packet1cd(_mm_xor_pd(a.v,mask));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  #ifdef EIGEN_VECTORIZE_SSE3
+  return Packet1cd(_mm_addsub_pd(_mm_mul_pd(_mm_movedup_pd(a.v), b.v),
+                                 _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
+                                            vec2d_swizzle1(b.v, 1, 0))));
+  #else
+  const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
+  return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
+                              _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
+                                                    vec2d_swizzle1(b.v, 1, 0)), mask)));
+  #endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_and_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_or_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_xor_pd(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(_mm_andnot_pd(a.v,b.v)); }
+
+// FIXME force unaligned load, this is a temporary fix
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from)
+{ EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>* from) { return pset1<Packet1cd>(*from); }
+
+// FIXME force unaligned store, this is a temporary fix
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, Packet2d(from.v)); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, Packet2d(from.v)); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  EIGEN_ALIGN16 double res[2];
+  _mm_store_pd(res, a.v);
+  return std::complex<double>(res[0],res[1]);
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
+{
+  return vecs[0];
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return internal::pmul(a, pconj(b));
+    #else
+    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
+    return Packet1cd(_mm_add_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
+                                _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
+                                           vec2d_swizzle1(b.v, 1, 0))));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return internal::pmul(pconj(a), b);
+    #else
+    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
+    return Packet1cd(_mm_add_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v),
+                                _mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
+                                                      vec2d_swizzle1(b.v, 1, 0)), mask)));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    #ifdef EIGEN_VECTORIZE_SSE3
+    return pconj(internal::pmul(a, b));
+    #else
+    const __m128d mask = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
+    return Packet1cd(_mm_sub_pd(_mm_xor_pd(_mm_mul_pd(vec2d_swizzle1(a.v, 0, 0), b.v), mask),
+                                _mm_mul_pd(vec2d_swizzle1(a.v, 1, 1),
+                                           vec2d_swizzle1(b.v, 1, 0))));
+    #endif
+  }
+};
+
+template<> struct conj_helper<Packet2d, Packet1cd, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet2d& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet2d& x, const Packet1cd& y) const
+  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x, y.v)); }
+};
+
+template<> struct conj_helper<Packet1cd, Packet2d, false,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet2d& y, const Packet1cd& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& x, const Packet2d& y) const
+  { return Packet1cd(Eigen::internal::pmul<Packet2d>(x.v, y)); }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for SSE3 and 4
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  __m128d s = _mm_mul_pd(b.v,b.v);
+  return Packet1cd(_mm_div_pd(res.v, _mm_add_pd(s,_mm_shuffle_pd(s, s, 0x1))));
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/* <Packet1cd> */(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2cf,2>& kernel) {
+  __m128d w1 = _mm_castps_pd(kernel.packet[0].v);
+  __m128d w2 = _mm_castps_pd(kernel.packet[1].v);
+
+  __m128 tmp = _mm_castpd_ps(_mm_unpackhi_pd(w1, w2));
+  kernel.packet[0].v = _mm_castpd_ps(_mm_unpacklo_pd(w1, w2));
+  kernel.packet[1].v = tmp;
+}
+
+template<>  EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  __m128d result = pblend<Packet2d>(ifPacket, _mm_castps_pd(thenPacket.v), _mm_castps_pd(elsePacket.v));
+  return Packet2cf(_mm_castpd_ps(result));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertfirst(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadl_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertfirst(const Packet1cd&, std::complex<double> b)
+{
+  return pset1<Packet1cd>(b);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pinsertlast(const Packet2cf& a, std::complex<float> b)
+{
+  return Packet2cf(_mm_loadh_pi(a.v, reinterpret_cast<const __m64*>(&b)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pinsertlast(const Packet1cd&, std::complex<double> b)
+{
+  return pset1<Packet1cd>(b);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h
new file mode 100644
index 0000000..7b5f948
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/MathFunctions.h
@@ -0,0 +1,562 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_SSE_H
+#define EIGEN_MATH_FUNCTIONS_SSE_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f plog<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inv_mant_mask, ~0x7f800000);
+
+  /* the smallest non denormalized float number */
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(min_norm_pos,  0x00800000);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(minus_inf,     0xff800000);//-1.f/0.f);
+
+  /* natural logarithm computed for 4 simultaneous float
+    return NaN for x <= 0
+  */
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_SQRTHF, 0.707106781186547524f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p0, 7.0376836292E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p1, - 1.1514610310E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p2, 1.1676998740E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p3, - 1.2420140846E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p4, + 1.4249322787E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p5, - 1.6668057665E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p6, + 2.0000714765E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p7, - 2.4999993993E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_p8, + 3.3333331174E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q1, -2.12194440e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_log_q2, 0.693359375f);
+
+
+  Packet4i emm0;
+
+  Packet4f invalid_mask = _mm_cmpnge_ps(x, _mm_setzero_ps()); // not greater equal is true if x is NaN
+  Packet4f iszero_mask = _mm_cmpeq_ps(x, _mm_setzero_ps());
+
+  x = pmax(x, p4f_min_norm_pos);  /* cut off denormalized stuff */
+  emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23);
+
+  /* keep only the fractional part */
+  x = _mm_and_ps(x, p4f_inv_mant_mask);
+  x = _mm_or_ps(x, p4f_half);
+
+  emm0 = _mm_sub_epi32(emm0, p4i_0x7f);
+  Packet4f e = padd(Packet4f(_mm_cvtepi32_ps(emm0)), p4f_1);
+
+  /* part2:
+     if( x < SQRTHF ) {
+       e -= 1;
+       x = x + x - 1.0;
+     } else { x = x - 1.0; }
+  */
+  Packet4f mask = _mm_cmplt_ps(x, p4f_cephes_SQRTHF);
+  Packet4f tmp = pand(x, mask);
+  x = psub(x, p4f_1);
+  e = psub(e, pand(p4f_1, mask));
+  x = padd(x, tmp);
+
+  Packet4f x2 = pmul(x,x);
+  Packet4f x3 = pmul(x2,x);
+
+  Packet4f y, y1, y2;
+  y  = pmadd(p4f_cephes_log_p0, x, p4f_cephes_log_p1);
+  y1 = pmadd(p4f_cephes_log_p3, x, p4f_cephes_log_p4);
+  y2 = pmadd(p4f_cephes_log_p6, x, p4f_cephes_log_p7);
+  y  = pmadd(y , x, p4f_cephes_log_p2);
+  y1 = pmadd(y1, x, p4f_cephes_log_p5);
+  y2 = pmadd(y2, x, p4f_cephes_log_p8);
+  y = pmadd(y, x3, y1);
+  y = pmadd(y, x3, y2);
+  y = pmul(y, x3);
+
+  y1 = pmul(e, p4f_cephes_log_q1);
+  tmp = pmul(x2, p4f_half);
+  y = padd(y, y1);
+  x = psub(x, tmp);
+  y2 = pmul(e, p4f_cephes_log_q2);
+  x = padd(x, y);
+  x = padd(x, y2);
+  // negative arg will be NAN, 0 will be -INF
+  return _mm_or_ps(_mm_andnot_ps(iszero_mask, _mm_or_ps(x, invalid_mask)),
+                   _mm_and_ps(iszero_mask, p4f_minus_inf));
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+  _EIGEN_DECLARE_CONST_Packet4i(0x7f, 0x7f);
+
+
+  _EIGEN_DECLARE_CONST_Packet4f(exp_hi,  88.3762626647950f);
+  _EIGEN_DECLARE_CONST_Packet4f(exp_lo, -88.3762626647949f);
+
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_LOG2EF, 1.44269504088896341f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C1, 0.693359375f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_C2, -2.12194440e-4f);
+
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p0, 1.9875691500E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p1, 1.3981999507E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p2, 8.3334519073E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p3, 4.1665795894E-2f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
+
+  Packet4f tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p4f_exp_hi), p4f_exp_lo);
+
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(x, p4f_cephes_LOG2EF, p4f_half);
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  fx = _mm_floor_ps(fx);
+#else
+  emm0 = _mm_cvttps_epi32(fx);
+  tmp  = _mm_cvtepi32_ps(emm0);
+  /* if greater, substract 1 */
+  Packet4f mask = _mm_cmpgt_ps(tmp, fx);
+  mask = _mm_and_ps(mask, p4f_1);
+  fx = psub(tmp, mask);
+#endif
+
+  tmp = pmul(fx, p4f_cephes_exp_C1);
+  Packet4f z = pmul(fx, p4f_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  z = pmul(x,x);
+
+  Packet4f y = p4f_cephes_exp_p0;
+  y = pmadd(y, x, p4f_cephes_exp_p1);
+  y = pmadd(y, x, p4f_cephes_exp_p2);
+  y = pmadd(y, x, p4f_cephes_exp_p3);
+  y = pmadd(y, x, p4f_cephes_exp_p4);
+  y = pmadd(y, x, p4f_cephes_exp_p5);
+  y = pmadd(y, z, x);
+  y = padd(y, p4f_1);
+
+  // build 2^n
+  emm0 = _mm_cvttps_epi32(fx);
+  emm0 = _mm_add_epi32(emm0, p4i_0x7f);
+  emm0 = _mm_slli_epi32(emm0, 23);
+  return pmax(pmul(y, Packet4f(_mm_castsi128_ps(emm0))), _x);
+}
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+  _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+  _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+  _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+  _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+  _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+  static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0);
+
+  Packet2d tmp, fx;
+  Packet4i emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  fx = _mm_floor_pd(fx);
+#else
+  emm0 = _mm_cvttpd_epi32(fx);
+  tmp  = _mm_cvtepi32_pd(emm0);
+  /* if greater, substract 1 */
+  Packet2d mask = _mm_cmpgt_pd(tmp, fx);
+  mask = _mm_and_pd(mask, p2d_1);
+  fx = psub(tmp, mask);
+#endif
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = _mm_cvttpd_epi32(fx);
+  emm0 = _mm_add_epi32(emm0, p4i_1023_0);
+  emm0 = _mm_slli_epi32(emm0, 20);
+  emm0 = _mm_shuffle_epi32(emm0, _MM_SHUFFLE(1,2,0,3));
+  return pmax(pmul(x, Packet2d(_mm_castsi128_pd(emm0))), _x);
+}
+
+/* evaluation of 4 sines at onces, using SSE2 intrinsics.
+
+   The code is the exact rewriting of the cephes sinf function.
+   Precision is excellent as long as x < 8192 (I did not bother to
+   take into account the special handling they have for greater values
+   -- it does not return garbage for arguments over 8192, though, but
+   the extra precision is missing).
+
+   Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
+   surprising but correct result.
+*/
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psin<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+
+  _EIGEN_DECLARE_CONST_Packet4i(1, 1);
+  _EIGEN_DECLARE_CONST_Packet4i(not1, ~1);
+  _EIGEN_DECLARE_CONST_Packet4i(2, 2);
+  _EIGEN_DECLARE_CONST_Packet4i(4, 4);
+
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(sign_mask, 0x80000000);
+
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p1,  8.3321608736E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p0,  2.443315711809948E-005f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
+
+  Packet4f xmm1, xmm2, xmm3, sign_bit, y;
+
+  Packet4i emm0, emm2;
+  sign_bit = x;
+  /* take the absolute value */
+  x = pabs(x);
+
+  /* take the modulo */
+
+  /* extract the sign bit (upper one) */
+  sign_bit = _mm_and_ps(sign_bit, p4f_sign_mask);
+
+  /* scale by 4/Pi */
+  y = pmul(x, p4f_cephes_FOPI);
+
+  /* store the integer part of y in mm0 */
+  emm2 = _mm_cvttps_epi32(y);
+  /* j=(j+1) & (~1) (see the cephes sources) */
+  emm2 = _mm_add_epi32(emm2, p4i_1);
+  emm2 = _mm_and_si128(emm2, p4i_not1);
+  y = _mm_cvtepi32_ps(emm2);
+  /* get the swap sign flag */
+  emm0 = _mm_and_si128(emm2, p4i_4);
+  emm0 = _mm_slli_epi32(emm0, 29);
+  /* get the polynom selection mask
+     there is one polynom for 0 <= x <= Pi/4
+     and another one for Pi/4<x<=Pi/2
+
+     Both branches will be computed.
+  */
+  emm2 = _mm_and_si128(emm2, p4i_2);
+  emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
+
+  Packet4f swap_sign_bit = _mm_castsi128_ps(emm0);
+  Packet4f poly_mask = _mm_castsi128_ps(emm2);
+  sign_bit = _mm_xor_ps(sign_bit, swap_sign_bit);
+
+  /* The magic pass: "Extended precision modular arithmetic"
+     x = ((x - y * DP1) - y * DP2) - y * DP3; */
+  xmm1 = pmul(y, p4f_minus_cephes_DP1);
+  xmm2 = pmul(y, p4f_minus_cephes_DP2);
+  xmm3 = pmul(y, p4f_minus_cephes_DP3);
+  x = padd(x, xmm1);
+  x = padd(x, xmm2);
+  x = padd(x, xmm3);
+
+  /* Evaluate the first polynom  (0 <= x <= Pi/4) */
+  y = p4f_coscof_p0;
+  Packet4f z = _mm_mul_ps(x,x);
+
+  y = pmadd(y, z, p4f_coscof_p1);
+  y = pmadd(y, z, p4f_coscof_p2);
+  y = pmul(y, z);
+  y = pmul(y, z);
+  Packet4f tmp = pmul(z, p4f_half);
+  y = psub(y, tmp);
+  y = padd(y, p4f_1);
+
+  /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
+
+  Packet4f y2 = p4f_sincof_p0;
+  y2 = pmadd(y2, z, p4f_sincof_p1);
+  y2 = pmadd(y2, z, p4f_sincof_p2);
+  y2 = pmul(y2, z);
+  y2 = pmul(y2, x);
+  y2 = padd(y2, x);
+
+  /* select the correct result from the two polynoms */
+  y2 = _mm_and_ps(poly_mask, y2);
+  y = _mm_andnot_ps(poly_mask, y);
+  y = _mm_or_ps(y,y2);
+  /* update the sign */
+  return _mm_xor_ps(y, sign_bit);
+}
+
+/* almost the same as psin */
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pcos<Packet4f>(const Packet4f& _x)
+{
+  Packet4f x = _x;
+  _EIGEN_DECLARE_CONST_Packet4f(1 , 1.0f);
+  _EIGEN_DECLARE_CONST_Packet4f(half, 0.5f);
+
+  _EIGEN_DECLARE_CONST_Packet4i(1, 1);
+  _EIGEN_DECLARE_CONST_Packet4i(not1, ~1);
+  _EIGEN_DECLARE_CONST_Packet4i(2, 2);
+  _EIGEN_DECLARE_CONST_Packet4i(4, 4);
+
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP1,-0.78515625f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP2, -2.4187564849853515625e-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_cephes_DP3, -3.77489497744594108e-8f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p0, -1.9515295891E-4f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p1,  8.3321608736E-3f);
+  _EIGEN_DECLARE_CONST_Packet4f(sincof_p2, -1.6666654611E-1f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p0,  2.443315711809948E-005f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p1, -1.388731625493765E-003f);
+  _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
+  _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
+
+  Packet4f xmm1, xmm2, xmm3, y;
+  Packet4i emm0, emm2;
+
+  x = pabs(x);
+
+  /* scale by 4/Pi */
+  y = pmul(x, p4f_cephes_FOPI);
+
+  /* get the integer part of y */
+  emm2 = _mm_cvttps_epi32(y);
+  /* j=(j+1) & (~1) (see the cephes sources) */
+  emm2 = _mm_add_epi32(emm2, p4i_1);
+  emm2 = _mm_and_si128(emm2, p4i_not1);
+  y = _mm_cvtepi32_ps(emm2);
+
+  emm2 = _mm_sub_epi32(emm2, p4i_2);
+
+  /* get the swap sign flag */
+  emm0 = _mm_andnot_si128(emm2, p4i_4);
+  emm0 = _mm_slli_epi32(emm0, 29);
+  /* get the polynom selection mask */
+  emm2 = _mm_and_si128(emm2, p4i_2);
+  emm2 = _mm_cmpeq_epi32(emm2, _mm_setzero_si128());
+
+  Packet4f sign_bit = _mm_castsi128_ps(emm0);
+  Packet4f poly_mask = _mm_castsi128_ps(emm2);
+
+  /* The magic pass: "Extended precision modular arithmetic"
+     x = ((x - y * DP1) - y * DP2) - y * DP3; */
+  xmm1 = pmul(y, p4f_minus_cephes_DP1);
+  xmm2 = pmul(y, p4f_minus_cephes_DP2);
+  xmm3 = pmul(y, p4f_minus_cephes_DP3);
+  x = padd(x, xmm1);
+  x = padd(x, xmm2);
+  x = padd(x, xmm3);
+
+  /* Evaluate the first polynom  (0 <= x <= Pi/4) */
+  y = p4f_coscof_p0;
+  Packet4f z = pmul(x,x);
+
+  y = pmadd(y,z,p4f_coscof_p1);
+  y = pmadd(y,z,p4f_coscof_p2);
+  y = pmul(y, z);
+  y = pmul(y, z);
+  Packet4f tmp = _mm_mul_ps(z, p4f_half);
+  y = psub(y, tmp);
+  y = padd(y, p4f_1);
+
+  /* Evaluate the second polynom  (Pi/4 <= x <= 0) */
+  Packet4f y2 = p4f_sincof_p0;
+  y2 = pmadd(y2, z, p4f_sincof_p1);
+  y2 = pmadd(y2, z, p4f_sincof_p2);
+  y2 = pmul(y2, z);
+  y2 = pmadd(y2, x, x);
+
+  /* select the correct result from the two polynoms */
+  y2 = _mm_and_ps(poly_mask, y2);
+  y  = _mm_andnot_ps(poly_mask, y);
+  y  = _mm_or_ps(y,y2);
+
+  /* update the sign */
+  return _mm_xor_ps(y, sign_bit);
+}
+
+#if EIGEN_FAST_MATH
+
+// Functions for sqrt.
+// The EIGEN_FAST_MATH version uses the _mm_rsqrt_ps approximation and one step
+// of Newton's method, at a cost of 1-2 bits of precision as opposed to the
+// exact solution. It does not handle +inf, or denormalized numbers correctly.
+// The main advantage of this approach is not just speed, but also the fact that
+// it can be inlined and pipelined with other computations, further reducing its
+// effective latency. This is similar to Quake3's fast inverse square root.
+// For detail see here: http://www.beyond3d.com/content/articles/8/
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& _x)
+{
+  Packet4f half = pmul(_x, pset1<Packet4f>(.5f));
+  Packet4f denormal_mask = _mm_and_ps(
+      _mm_cmpge_ps(_x, _mm_setzero_ps()),
+      _mm_cmplt_ps(_x, pset1<Packet4f>((std::numeric_limits<float>::min)())));
+
+  // Compute approximate reciprocal sqrt.
+  Packet4f x = _mm_rsqrt_ps(_x);
+  // Do a single step of Newton's iteration.
+  x = pmul(x, psub(pset1<Packet4f>(1.5f), pmul(half, pmul(x,x))));
+  // Flush results for denormals to zero.
+  return _mm_andnot_ps(denormal_mask, pmul(_x,x));
+}
+
+#else
+
+template<>EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x) { return _mm_sqrt_ps(x); }
+
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x) { return _mm_sqrt_pd(x); }
+
+#if EIGEN_FAST_MATH
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& _x) {
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(inf, 0x7f800000);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(nan, 0x7fc00000);
+  _EIGEN_DECLARE_CONST_Packet4f(one_point_five, 1.5f);
+  _EIGEN_DECLARE_CONST_Packet4f(minus_half, -0.5f);
+  _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(flt_min, 0x00800000);
+
+  Packet4f neg_half = pmul(_x, p4f_minus_half);
+
+  // select only the inverse sqrt of positive normal inputs (denormals are
+  // flushed to zero and cause infs as well).
+  Packet4f le_zero_mask = _mm_cmple_ps(_x, p4f_flt_min);
+  Packet4f x = _mm_andnot_ps(le_zero_mask, _mm_rsqrt_ps(_x));
+
+  // Fill in NaNs and Infs for the negative/zero entries.
+  Packet4f neg_mask = _mm_cmplt_ps(_x, _mm_setzero_ps());
+  Packet4f zero_mask = _mm_andnot_ps(neg_mask, le_zero_mask);
+  Packet4f infs_and_nans = _mm_or_ps(_mm_and_ps(neg_mask, p4f_nan),
+                                     _mm_and_ps(zero_mask, p4f_inf));
+
+  // Do a single step of Newton's iteration.
+  x = pmul(x, pmadd(neg_half, pmul(x, x), p4f_one_point_five));
+
+  // Insert NaNs and Infs in all the right places.
+  return _mm_or_ps(x, infs_and_nans);
+}
+
+#else
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_ps since it only provides an approximation.
+  return _mm_div_ps(pset1<Packet4f>(1.0f), _mm_sqrt_ps(x));
+}
+
+#endif
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
+  return _mm_div_pd(pset1<Packet2d>(1.0), _mm_sqrt_pd(x));
+}
+
+// Hyperbolic Tangent function.
+template <>
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED Packet4f
+ptanh<Packet4f>(const Packet4f& x) {
+  return internal::generic_fast_tanh_float(x);
+}
+
+} // end namespace internal
+
+namespace numext {
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+float sqrt(const float &x)
+{
+  return internal::pfirst(internal::Packet4f(_mm_sqrt_ss(_mm_set_ss(x))));
+}
+
+template<>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
+double sqrt(const double &x)
+{
+#if EIGEN_COMP_GNUC_STRICT
+  // This works around a GCC bug generating poor code for _mm_sqrt_pd
+  // See https://bitbucket.org/eigen/eigen/commits/14f468dba4d350d7c19c9b93072e19f7b3df563b
+  return internal::pfirst(internal::Packet2d(__builtin_ia32_sqrtsd(_mm_set_sd(x))));
+#else
+  return internal::pfirst(internal::Packet2d(_mm_sqrt_pd(_mm_set_sd(x))));
+#endif
+}
+
+} // end namespace numex
+
+} // end namespace Eigen
+
+#endif // EIGEN_MATH_FUNCTIONS_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h
new file mode 100755
index 0000000..3832de1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/PacketMath.h
@@ -0,0 +1,879 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_SSE_H
+#define EIGEN_PACKET_MATH_SSE_H
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 8
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS (2*sizeof(void*))
+#endif
+
+#ifdef __FMA__
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD 1
+#endif
+#endif
+
+#if (defined EIGEN_VECTORIZE_AVX) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_MINGW) && (__GXX_ABI_VERSION < 1004)
+// With GCC's default ABI version, a __m128 or __m256 are the same types and therefore we cannot
+// have overloads for both types without linking error.
+// One solution is to increase ABI version using -fabi-version=4 (or greater).
+// Otherwise, we workaround this inconvenience by wrapping 128bit types into the following helper
+// structure:
+template<typename T>
+struct eigen_packet_wrapper
+{
+  EIGEN_ALWAYS_INLINE operator T&() { return m_val; }
+  EIGEN_ALWAYS_INLINE operator const T&() const { return m_val; }
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper() {}
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper(const T &v) : m_val(v) {}
+  EIGEN_ALWAYS_INLINE eigen_packet_wrapper& operator=(const T &v) {
+    m_val = v;
+    return *this;
+  }
+  
+  T m_val;
+};
+typedef eigen_packet_wrapper<__m128>  Packet4f;
+typedef eigen_packet_wrapper<__m128i> Packet4i;
+typedef eigen_packet_wrapper<__m128d> Packet2d;
+#else
+typedef __m128  Packet4f;
+typedef __m128i Packet4i;
+typedef __m128d Packet2d;
+#endif
+
+template<> struct is_arithmetic<__m128>  { enum { value = true }; };
+template<> struct is_arithmetic<__m128i> { enum { value = true }; };
+template<> struct is_arithmetic<__m128d> { enum { value = true }; };
+
+#define vec4f_swizzle1(v,p,q,r,s) \
+  (_mm_castsi128_ps(_mm_shuffle_epi32( _mm_castps_si128(v), ((s)<<6|(r)<<4|(q)<<2|(p)))))
+
+#define vec4i_swizzle1(v,p,q,r,s) \
+  (_mm_shuffle_epi32( v, ((s)<<6|(r)<<4|(q)<<2|(p))))
+
+#define vec2d_swizzle1(v,p,q) \
+  (_mm_castsi128_pd(_mm_shuffle_epi32( _mm_castpd_si128(v), ((q*2+1)<<6|(q*2)<<4|(p*2+1)<<2|(p*2)))))
+  
+#define vec4f_swizzle2(a,b,p,q,r,s) \
+  (_mm_shuffle_ps( (a), (b), ((s)<<6|(r)<<4|(q)<<2|(p))))
+
+#define vec4i_swizzle2(a,b,p,q,r,s) \
+  (_mm_castps_si128( (_mm_shuffle_ps( _mm_castsi128_ps(a), _mm_castsi128_ps(b), ((s)<<6|(r)<<4|(q)<<2|(p))))))
+
+#define _EIGEN_DECLARE_CONST_Packet4f(NAME,X) \
+  const Packet4f p4f_##NAME = pset1<Packet4f>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  const Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet4f_FROM_INT(NAME,X) \
+  const Packet4f p4f_##NAME = _mm_castsi128_ps(pset1<Packet4i>(X))
+
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+  const Packet4i p4i_##NAME = pset1<Packet4i>(X)
+
+
+// Use the packet_traits defined in AVX/PacketMath.h instead if we're going
+// to leverage AVX instructions.
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct packet_traits<float>  : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasSin  = EIGEN_FAST_MATH,
+    HasCos  = EIGEN_FAST_MATH,
+    HasLog  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasTanh  = EIGEN_FAST_MATH,
+    HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
+  };
+};
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 0,
+
+    HasDiv  = 1,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasBlend = 1
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+    ,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1
+#endif
+  };
+};
+#endif
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet4i type;
+  typedef Packet4i half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+
+    HasBlend = 1
+  };
+};
+
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+
+#ifndef EIGEN_VECTORIZE_AVX
+template<> struct scalar_div_cost<float,true> { enum { value = 7 }; };
+template<> struct scalar_div_cost<double,true> { enum { value = 8 }; };
+#endif
+
+#if EIGEN_COMP_MSVC==1500
+// Workaround MSVC 9 internal compiler error.
+// TODO: It has been detected with win64 builds (amd64), so let's check whether it also happens in 32bits+SSE mode
+// TODO: let's check whether there does not exist a better fix, like adding a pset0() function. (it crashed on pset1(0)).
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set_ps(from,from,from,from); }
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set_pd(from,from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set_epi32(from,from,from,from); }
+#else
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&  from) { return _mm_set_ps1(from); }
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) { return _mm_set1_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from) { return _mm_set1_epi32(from); }
+#endif
+
+// GCC generates a shufps instruction for _mm_set1_ps/_mm_load1_ps instead of the more efficient pshufd instruction.
+// However, using inrinsics for pset1 makes gcc to generate crappy code in some cases (see bug 203)
+// Using inline assembly is also not an option because then gcc fails to reorder properly the instructions.
+// Therefore, we introduced the pload1 functions to be used in product kernels for which bug 203 does not apply.
+// Also note that with AVX, we want it to generate a vbroadcastss.
+#if EIGEN_COMP_GNUC_STRICT && (!defined __AVX__)
+template<> EIGEN_STRONG_INLINE Packet4f pload1<Packet4f>(const float *from) {
+  return vec4f_swizzle1(_mm_load_ss(from),0,0,0,0);
+}
+#endif
+  
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a) { return _mm_add_ps(pset1<Packet4f>(a), _mm_set_ps(3,2,1,0)); }
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return _mm_add_pd(pset1<Packet2d>(a),_mm_set_pd(1,0)); }
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a) { return _mm_add_epi32(pset1<Packet4i>(a),_mm_set_epi32(3,2,1,0)); }
+
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_add_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_add_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_add_epi32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_sub_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_sub_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_sub_epi32(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x80000000,0x80000000,0x80000000,0x80000000));
+  return _mm_xor_ps(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a)
+{
+  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x80000000,0x0,0x80000000));
+  return _mm_xor_pd(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a)
+{
+  return psub(Packet4i(_mm_setr_epi32(0,0,0,0)), a);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_mul_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_mul_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_mullo_epi32(a,b);
+#else
+  // this version is slightly faster than 4 scalar products
+  return vec4i_swizzle1(
+            vec4i_swizzle2(
+              _mm_mul_epu32(a,b),
+              _mm_mul_epu32(vec4i_swizzle1(a,1,0,3,2),
+                            vec4i_swizzle1(b,1,0,3,2)),
+              0,2,0,2),
+            0,2,1,3);
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_div_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_div_pd(a,b); }
+
+// for some weird raisons, it has to be overloaded for packet of integers
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd(pmul(a,b), c); }
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c) { return _mm_fmadd_ps(a,b,c); }
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return _mm_fmadd_pd(a,b,c); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_min_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_min_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_min_epi32(a,b);
+#else
+  // after some bench, this version *is* faster than a scalar implementation
+  Packet4i mask = _mm_cmplt_epi32(a,b);
+  return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_max_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_max_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_max_epi32(a,b);
+#else
+  // after some bench, this version *is* faster than a scalar implementation
+  Packet4i mask = _mm_cmpgt_epi32(a,b);
+  return _mm_or_si128(_mm_and_si128(mask,a),_mm_andnot_si128(mask,b));
+#endif
+}
+
+#ifdef EIGEN_VECTORIZE_SSE4_1
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a) { return _mm_round_ps(a, 0); }
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return _mm_round_pd(a, 0); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a) { return _mm_ceil_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const Packet2d& a) { return _mm_ceil_pd(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) { return _mm_floor_ps(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return _mm_floor_pd(a); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_and_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_and_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_and_si128(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_or_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_or_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_or_si128(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_xor_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_xor_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_xor_si128(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b) { return _mm_andnot_ps(a,b); }
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return _mm_andnot_pd(a,b); }
+template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return _mm_andnot_si128(a,b); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_ps(from); }
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double*  from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_pd(from); }
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { EIGEN_DEBUG_ALIGNED_LOAD return _mm_load_si128(reinterpret_cast<const __m128i*>(from)); }
+
+#if EIGEN_COMP_MSVC
+  template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*  from) {
+    EIGEN_DEBUG_UNALIGNED_LOAD
+    #if (EIGEN_COMP_MSVC==1600)
+    // NOTE Some version of MSVC10 generates bad code when using _mm_loadu_ps
+    // (i.e., it does not generate an unaligned load!!
+    __m128 res = _mm_loadl_pi(_mm_set1_ps(0.0f), (const __m64*)(from));
+    res = _mm_loadh_pi(res, (const __m64*)(from+2));
+    return res;
+    #else
+    return _mm_loadu_ps(from);
+    #endif
+  }
+#else
+// NOTE: with the code below, MSVC's compiler crashes!
+
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return _mm_loadu_ps(from);
+}
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return _mm_loadu_pd(from);
+}
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
+{
+  EIGEN_DEBUG_UNALIGNED_LOAD
+  return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
+}
+
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
+{
+  return vec4f_swizzle1(_mm_castpd_ps(_mm_load_sd(reinterpret_cast<const double*>(from))), 0, 0, 1, 1);
+}
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*  from)
+{ return pset1<Packet2d>(from[0]); }
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+{
+  Packet4i tmp;
+  tmp = _mm_loadl_epi64(reinterpret_cast<const __m128i*>(from));
+  return vec4i_swizzle1(tmp, 0, 0, 1, 1);
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_ALIGNED_STORE _mm_store_si128(reinterpret_cast<__m128i*>(to), from); }
+
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double* to, const Packet2d& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_pd(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*   to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_ps(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*       to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE _mm_storeu_si128(reinterpret_cast<__m128i*>(to), from); }
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+ return _mm_set_ps(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+ return _mm_set_pd(from[1*stride], from[0*stride]);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+ return _mm_set_epi32(from[3*stride], from[2*stride], from[1*stride], from[0*stride]);
+ }
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  to[stride*0] = _mm_cvtss_f32(from);
+  to[stride*1] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 1));
+  to[stride*2] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 2));
+  to[stride*3] = _mm_cvtss_f32(_mm_shuffle_ps(from, from, 3));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  to[stride*0] = _mm_cvtsd_f64(from);
+  to[stride*1] = _mm_cvtsd_f64(_mm_shuffle_pd(from, from, 1));
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  to[stride*0] = _mm_cvtsi128_si32(from);
+  to[stride*1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1));
+  to[stride*2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2));
+  to[stride*3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3));
+}
+
+// some compilers might be tempted to perform multiple moves instead of using a vector path.
+template<> EIGEN_STRONG_INLINE void pstore1<Packet4f>(float* to, const float& a)
+{
+  Packet4f pa = _mm_set_ss(a);
+  pstore(to, Packet4f(vec4f_swizzle1(pa,0,0,0,0)));
+}
+// some compilers might be tempted to perform multiple moves instead of using a vector path.
+template<> EIGEN_STRONG_INLINE void pstore1<Packet2d>(double* to, const double& a)
+{
+  Packet2d pa = _mm_set_sd(a);
+  pstore(to, Packet2d(vec2d_swizzle1(pa,0,0)));
+}
+
+#ifndef EIGEN_VECTORIZE_AVX
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { _mm_prefetch((const char*)(addr), _MM_HINT_T0); }
+#endif
+
+#if EIGEN_COMP_MSVC_STRICT && EIGEN_OS_WIN64
+// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
+// Direct of the struct members fixed bug #62.
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { return a.m128_f32[0]; }
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return a.m128d_f64[0]; }
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
+#elif EIGEN_COMP_MSVC_STRICT
+// The temporary variable fixes an internal compilation error in vs <= 2008 and a wrong-result bug in vs 2010
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float x = _mm_cvtss_f32(a); return x; }
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double x = _mm_cvtsd_f64(a); return x; }
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int x = _mm_cvtsi128_si32(a); return x; }
+#else
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { return _mm_cvtss_f32(a); }
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { return _mm_cvtsd_f64(a); }
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { return _mm_cvtsi128_si32(a); }
+#endif
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{ return _mm_shuffle_ps(a,a,0x1B); }
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
+{ return _mm_shuffle_pd(a,a,0x1); }
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
+{ return _mm_shuffle_epi32(a,0x1B); }
+
+template<> EIGEN_STRONG_INLINE Packet4f pabs(const Packet4f& a)
+{
+  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF,0x7FFFFFFF));
+  return _mm_and_ps(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet2d pabs(const Packet2d& a)
+{
+  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0xFFFFFFFF,0x7FFFFFFF,0xFFFFFFFF,0x7FFFFFFF));
+  return _mm_and_pd(a,mask);
+}
+template<> EIGEN_STRONG_INLINE Packet4i pabs(const Packet4i& a)
+{
+  #ifdef EIGEN_VECTORIZE_SSSE3
+  return _mm_abs_epi32(a);
+  #else
+  Packet4i aux = _mm_srai_epi32(a,31);
+  return _mm_sub_epi32(_mm_xor_si128(a,aux),aux);
+  #endif
+}
+
+// with AVX, the default implementations based on pload1 are faster
+#ifndef __AVX__
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec4f_swizzle1(a3, 0,0,0,0);
+  a1 = vec4f_swizzle1(a3, 1,1,1,1);
+  a2 = vec4f_swizzle1(a3, 2,2,2,2);
+  a3 = vec4f_swizzle1(a3, 3,3,3,3);
+}
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+#ifdef EIGEN_VECTORIZE_SSE3
+  a0 = _mm_loaddup_pd(a+0);
+  a1 = _mm_loaddup_pd(a+1);
+  a2 = _mm_loaddup_pd(a+2);
+  a3 = _mm_loaddup_pd(a+3);
+#else
+  a1 = pload<Packet2d>(a);
+  a0 = vec2d_swizzle1(a1, 0,0);
+  a1 = vec2d_swizzle1(a1, 1,1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec2d_swizzle1(a3, 0,0);
+  a3 = vec2d_swizzle1(a3, 1,1);
+#endif
+}
+#endif
+
+EIGEN_STRONG_INLINE void punpackp(Packet4f* vecs)
+{
+  vecs[1] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x55));
+  vecs[2] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xAA));
+  vecs[3] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0xFF));
+  vecs[0] = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(vecs[0]), 0x00));
+}
+
+#ifdef EIGEN_VECTORIZE_SSE3
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  return _mm_hadd_ps(_mm_hadd_ps(vecs[0], vecs[1]),_mm_hadd_ps(vecs[2], vecs[3]));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  return _mm_hadd_pd(vecs[0], vecs[1]);
+}
+
+#else
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  Packet4f tmp0, tmp1, tmp2;
+  tmp0 = _mm_unpacklo_ps(vecs[0], vecs[1]);
+  tmp1 = _mm_unpackhi_ps(vecs[0], vecs[1]);
+  tmp2 = _mm_unpackhi_ps(vecs[2], vecs[3]);
+  tmp0 = _mm_add_ps(tmp0, tmp1);
+  tmp1 = _mm_unpacklo_ps(vecs[2], vecs[3]);
+  tmp1 = _mm_add_ps(tmp1, tmp2);
+  tmp2 = _mm_movehl_ps(tmp1, tmp0);
+  tmp0 = _mm_movelh_ps(tmp0, tmp1);
+  return _mm_add_ps(tmp0, tmp2);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  return _mm_add_pd(_mm_unpacklo_pd(vecs[0], vecs[1]), _mm_unpackhi_pd(vecs[0], vecs[1]));
+}
+#endif  // SSE3
+
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   Packet4f tmp = _mm_add_ps(a, vec4f_swizzle1(a,2,3,2,3));
+//   return pfirst<Packet4f>(_mm_hadd_ps(tmp, tmp));
+// #else
+  Packet4f tmp = _mm_add_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_add_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+// #endif
+}
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  // Disable SSE3 _mm_hadd_pd that is extremely slow on all existing Intel's architectures
+  // (from Nehalem to Haswell)
+// #ifdef EIGEN_VECTORIZE_SSE3
+//   return pfirst<Packet2d>(_mm_hadd_pd(a, a));
+// #else
+  return pfirst<Packet2d>(_mm_add_sd(a, _mm_unpackhi_pd(a,a)));
+// #endif
+}
+
+#ifdef EIGEN_VECTORIZE_SSSE3
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  return _mm_hadd_epi32(_mm_hadd_epi32(vecs[0], vecs[1]),_mm_hadd_epi32(vecs[2], vecs[3]));
+}
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i tmp0 = _mm_hadd_epi32(a,a);
+  return pfirst<Packet4i>(_mm_hadd_epi32(tmp0,tmp0));
+}
+#else
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i tmp = _mm_add_epi32(a, _mm_unpackhi_epi64(a,a));
+  return pfirst(tmp) + pfirst<Packet4i>(_mm_shuffle_epi32(tmp, 1));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  Packet4i tmp0, tmp1, tmp2;
+  tmp0 = _mm_unpacklo_epi32(vecs[0], vecs[1]);
+  tmp1 = _mm_unpackhi_epi32(vecs[0], vecs[1]);
+  tmp2 = _mm_unpackhi_epi32(vecs[2], vecs[3]);
+  tmp0 = _mm_add_epi32(tmp0, tmp1);
+  tmp1 = _mm_unpacklo_epi32(vecs[2], vecs[3]);
+  tmp1 = _mm_add_epi32(tmp1, tmp2);
+  tmp2 = _mm_unpacklo_epi64(tmp0, tmp1);
+  tmp0 = _mm_unpackhi_epi64(tmp0, tmp1);
+  return _mm_add_epi32(tmp0, tmp2);
+}
+#endif
+// Other reduction functions:
+
+// mul
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  Packet4f tmp = _mm_mul_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_mul_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst<Packet2d>(_mm_mul_sd(a, _mm_unpackhi_pd(a,a)));
+}
+template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+{
+  // after some experiments, it is seems this is the fastest way to implement it
+  // for GCC (eg., reusing pmul is very slow !)
+  // TODO try to call _mm_mul_epu32 directly
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  return  (aux[0] * aux[1]) * (aux[2] * aux[3]);;
+}
+
+// min
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet4f tmp = _mm_min_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_min_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst<Packet2d>(_mm_min_sd(a, _mm_unpackhi_pd(a,a)));
+}
+template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  Packet4i tmp = _mm_min_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
+  return pfirst<Packet4i>(_mm_min_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
+#else
+  // after some experiments, it is seems this is the fastest way to implement it
+  // for GCC (eg., it does not like using std::min after the pstore !!)
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  int aux0 = aux[0]<aux[1] ? aux[0] : aux[1];
+  int aux2 = aux[2]<aux[3] ? aux[2] : aux[3];
+  return aux0<aux2 ? aux0 : aux2;
+#endif // EIGEN_VECTORIZE_SSE4_1
+}
+
+// max
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet4f tmp = _mm_max_ps(a, _mm_movehl_ps(a,a));
+  return pfirst<Packet4f>(_mm_max_ss(tmp, _mm_shuffle_ps(tmp,tmp, 1)));
+}
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst<Packet2d>(_mm_max_sd(a, _mm_unpackhi_pd(a,a)));
+}
+template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  Packet4i tmp = _mm_max_epi32(a, _mm_shuffle_epi32(a, _MM_SHUFFLE(0,0,3,2)));
+  return pfirst<Packet4i>(_mm_max_epi32(tmp,_mm_shuffle_epi32(tmp, 1)));
+#else
+  // after some experiments, it is seems this is the fastest way to implement it
+  // for GCC (eg., it does not like using std::min after the pstore !!)
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  int aux0 = aux[0]>aux[1] ? aux[0] : aux[1];
+  int aux2 = aux[2]>aux[3] ? aux[2] : aux[3];
+  return aux0>aux2 ? aux0 : aux2;
+#endif // EIGEN_VECTORIZE_SSE4_1
+}
+
+#if EIGEN_COMP_GNUC
+// template <> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f&  a, const Packet4f&  b, const Packet4f&  c)
+// {
+//   Packet4f res = b;
+//   asm("mulps %[a], %[b] \n\taddps %[c], %[b]" : [b] "+x" (res) : [a] "x" (a), [c] "x" (c));
+//   return res;
+// }
+// EIGEN_STRONG_INLINE Packet4i _mm_alignr_epi8(const Packet4i&  a, const Packet4i&  b, const int i)
+// {
+//   Packet4i res = a;
+//   asm("palignr %[i], %[a], %[b] " : [b] "+x" (res) : [a] "x" (a), [i] "i" (i));
+//   return res;
+// }
+#endif
+
+#ifdef EIGEN_VECTORIZE_SSSE3
+// SSSE3 versions
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    if (Offset!=0)
+      first = _mm_castsi128_ps(_mm_alignr_epi8(_mm_castps_si128(second), _mm_castps_si128(first), Offset*4));
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet4i>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
+  {
+    if (Offset!=0)
+      first = _mm_alignr_epi8(second,first, Offset*4);
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset==1)
+      first = _mm_castsi128_pd(_mm_alignr_epi8(_mm_castpd_si128(second), _mm_castpd_si128(first), 8));
+  }
+};
+#else
+// SSE2 versions
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm_move_ss(first,second);
+      first = _mm_castsi128_ps(_mm_shuffle_epi32(_mm_castps_si128(first),0x39));
+    }
+    else if (Offset==2)
+    {
+      first = _mm_movehl_ps(first,first);
+      first = _mm_movelh_ps(first,second);
+    }
+    else if (Offset==3)
+    {
+      first = _mm_move_ss(first,second);
+      first = _mm_shuffle_ps(first,second,0x93);
+    }
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet4i>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
+      first = _mm_shuffle_epi32(first,0x39);
+    }
+    else if (Offset==2)
+    {
+      first = _mm_castps_si128(_mm_movehl_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(first)));
+      first = _mm_castps_si128(_mm_movelh_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
+    }
+    else if (Offset==3)
+    {
+      first = _mm_castps_si128(_mm_move_ss(_mm_castsi128_ps(first),_mm_castsi128_ps(second)));
+      first = _mm_castps_si128(_mm_shuffle_ps(_mm_castsi128_ps(first),_mm_castsi128_ps(second),0x93));
+    }
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset==1)
+    {
+      first = _mm_castps_pd(_mm_movehl_ps(_mm_castpd_ps(first),_mm_castpd_ps(first)));
+      first = _mm_castps_pd(_mm_movelh_ps(_mm_castpd_ps(first),_mm_castpd_ps(second)));
+    }
+  }
+};
+#endif
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  _MM_TRANSPOSE4_PS(kernel.packet[0], kernel.packet[1], kernel.packet[2], kernel.packet[3]);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  __m128d tmp = _mm_unpackhi_pd(kernel.packet[0], kernel.packet[1]);
+  kernel.packet[0] = _mm_unpacklo_pd(kernel.packet[0], kernel.packet[1]);
+  kernel.packet[1] = tmp;
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  __m128i T0 = _mm_unpacklo_epi32(kernel.packet[0], kernel.packet[1]);
+  __m128i T1 = _mm_unpacklo_epi32(kernel.packet[2], kernel.packet[3]);
+  __m128i T2 = _mm_unpackhi_epi32(kernel.packet[0], kernel.packet[1]);
+  __m128i T3 = _mm_unpackhi_epi32(kernel.packet[2], kernel.packet[3]);
+
+  kernel.packet[0] = _mm_unpacklo_epi64(T0, T1);
+  kernel.packet[1] = _mm_unpackhi_epi64(T0, T1);
+  kernel.packet[2] = _mm_unpacklo_epi64(T2, T3);
+  kernel.packet[3] = _mm_unpackhi_epi64(T2, T3);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i select = _mm_set_epi32(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m128i false_mask = _mm_cmpeq_epi32(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_epi8(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_si128(_mm_andnot_si128(false_mask, thenPacket), _mm_and_si128(false_mask, elsePacket));
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  const __m128 zero = _mm_setzero_ps();
+  const __m128 select = _mm_set_ps(ifPacket.select[3], ifPacket.select[2], ifPacket.select[1], ifPacket.select[0]);
+  __m128 false_mask = _mm_cmpeq_ps(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_ps(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_ps(_mm_andnot_ps(false_mask, thenPacket), _mm_and_ps(false_mask, elsePacket));
+#endif
+}
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  const __m128d zero = _mm_setzero_pd();
+  const __m128d select = _mm_set_pd(ifPacket.select[1], ifPacket.select[0]);
+  __m128d false_mask = _mm_cmpeq_pd(select, zero);
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blendv_pd(thenPacket, elsePacket, false_mask);
+#else
+  return _mm_or_pd(_mm_andnot_pd(false_mask, thenPacket), _mm_and_pd(false_mask, elsePacket));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pinsertfirst(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),1);
+#else
+  return _mm_move_ss(a, _mm_load_ss(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertfirst(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),1);
+#else
+  return _mm_move_sd(a, _mm_load_sd(&b));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pinsertlast(const Packet4f& a, float b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_ps(a,pset1<Packet4f>(b),(1<<3));
+#else
+  const Packet4f mask = _mm_castsi128_ps(_mm_setr_epi32(0x0,0x0,0x0,0xFFFFFFFF));
+  return _mm_or_ps(_mm_andnot_ps(mask, a), _mm_and_ps(mask, pset1<Packet4f>(b)));
+#endif
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pinsertlast(const Packet2d& a, double b)
+{
+#ifdef EIGEN_VECTORIZE_SSE4_1
+  return _mm_blend_pd(a,pset1<Packet2d>(b),(1<<1));
+#else
+  const Packet2d mask = _mm_castsi128_pd(_mm_setr_epi32(0x0,0x0,0xFFFFFFFF,0xFFFFFFFF));
+  return _mm_or_pd(_mm_andnot_pd(mask, a), _mm_and_pd(mask, pset1<Packet2d>(b)));
+#endif
+}
+
+// Scalar path for pmadd with FMA to ensure consistency with vectorized path.
+#ifdef __FMA__
+template<> EIGEN_STRONG_INLINE float pmadd(const float& a, const float& b, const float& c) {
+  return ::fmaf(a,b,c);
+}
+template<> EIGEN_STRONG_INLINE double pmadd(const double& a, const double& b, const double& c) {
+  return ::fma(a,b,c);
+}
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h
new file mode 100644
index 0000000..c848932
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/SSE/TypeCasting.h
@@ -0,0 +1,77 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Benoit Steiner <benoit.steiner.goog@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TYPE_CASTING_SSE_H
+#define EIGEN_TYPE_CASTING_SSE_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <>
+struct type_casting_traits<float, int> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4i pcast<Packet4f, Packet4i>(const Packet4f& a) {
+  return _mm_cvttps_epi32(a);
+}
+
+
+template <>
+struct type_casting_traits<int, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet4i, Packet4f>(const Packet4i& a) {
+  return _mm_cvtepi32_ps(a);
+}
+
+
+template <>
+struct type_casting_traits<double, float> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 2,
+    TgtCoeffRatio = 1
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet4f pcast<Packet2d, Packet4f>(const Packet2d& a, const Packet2d& b) {
+  return _mm_shuffle_ps(_mm_cvtpd_ps(a), _mm_cvtpd_ps(b), (1 << 2) | (1 << 6));
+}
+
+template <>
+struct type_casting_traits<float, double> {
+  enum {
+    VectorizedCast = 1,
+    SrcCoeffRatio = 1,
+    TgtCoeffRatio = 2
+  };
+};
+
+template<> EIGEN_STRONG_INLINE Packet2d pcast<Packet4f, Packet2d>(const Packet4f& a) {
+  // Simply discard the second half of the input
+  return _mm_cvtps_pd(a);
+}
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TYPE_CASTING_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h
new file mode 100644
index 0000000..d39d2d1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/Complex.h
@@ -0,0 +1,394 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX32_ALTIVEC_H
+#define EIGEN_COMPLEX32_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static Packet2ul  p2ul_CONJ_XOR1 = (Packet2ul) vec_sld((Packet4ui) p2d_ZERO_, (Packet4ui) p2l_ZERO, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+static Packet2ul  p2ul_CONJ_XOR2 = (Packet2ul) vec_sld((Packet4ui) p2l_ZERO,  (Packet4ui) p2d_ZERO_, 8);//{ 0x8000000000000000, 0x0000000000000000 };
+
+struct Packet1cd
+{
+  EIGEN_STRONG_INLINE Packet1cd() {}
+  EIGEN_STRONG_INLINE explicit Packet1cd(const Packet2d& a) : v(a) {}
+  Packet2d v;
+};
+
+struct Packet2cf
+{
+  EIGEN_STRONG_INLINE Packet2cf() {}
+  EIGEN_STRONG_INLINE explicit Packet2cf(const Packet4f& a) : v(a) {}
+  union {
+    Packet4f v;
+    Packet1cd cd[2];
+  };
+};
+
+template<> struct packet_traits<std::complex<float> >  : default_packet_traits
+{
+  typedef Packet2cf type;
+  typedef Packet2cf half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 2,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasBlend  = 1,
+    HasSetLinear = 0
+  };
+};
+
+
+template<> struct packet_traits<std::complex<double> >  : default_packet_traits
+{
+  typedef Packet1cd type;
+  typedef Packet1cd half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 1,
+    HasHalfPacket = 0,
+
+    HasAdd    = 1,
+    HasSub    = 1,
+    HasMul    = 1,
+    HasDiv    = 1,
+    HasNegate = 1,
+    HasAbs    = 0,
+    HasAbs2   = 0,
+    HasMin    = 0,
+    HasMax    = 0,
+    HasSetLinear = 0
+  };
+};
+
+template<> struct unpacket_traits<Packet2cf> { typedef std::complex<float>  type; enum {size=2, alignment=Aligned16}; typedef Packet2cf half; };
+template<> struct unpacket_traits<Packet1cd> { typedef std::complex<double> type; enum {size=1, alignment=Aligned16}; typedef Packet1cd half; };
+
+/* Forward declaration */
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel);
+
+template<> EIGEN_STRONG_INLINE Packet2cf pload <Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_ALIGNED_LOAD return Packet2cf(pload<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pload <Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_ALIGNED_LOAD return Packet1cd(pload<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploadu<Packet2cf>(const std::complex<float>* from)  { EIGEN_DEBUG_UNALIGNED_LOAD return Packet2cf(ploadu<Packet4f>((const float*)from)); }
+template<> EIGEN_STRONG_INLINE Packet1cd ploadu<Packet1cd>(const std::complex<double>* from) { EIGEN_DEBUG_UNALIGNED_LOAD return Packet1cd(ploadu<Packet2d>((const double*)from)); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstore <std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((double*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *     to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
+template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<double> >(std::complex<double> *   to, const Packet1cd& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((double*)to, from.v); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd pset1<Packet1cd>(const std::complex<double>&  from)
+{ /* here we really have to use unaligned loads :( */ return ploadu<Packet1cd>(&from); }
+
+template<> EIGEN_STRONG_INLINE Packet2cf pset1<Packet2cf>(const std::complex<float>&  from)
+{
+  Packet2cf res;
+  res.cd[0] = Packet1cd(vec_ld2f((const float *)&from));
+  res.cd[1] = res.cd[0];
+  return res;
+}
+template<> EIGEN_DEVICE_FUNC inline Packet2cf pgather<std::complex<float>, Packet2cf>(const std::complex<float>* from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+  return pload<Packet2cf>(af);
+}
+template<> EIGEN_DEVICE_FUNC inline Packet1cd pgather<std::complex<double>, Packet1cd>(const std::complex<double>* from, Index stride EIGEN_UNUSED)
+{
+  return pload<Packet1cd>(from);
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<float>, Packet2cf>(std::complex<float>* to, const Packet2cf& from, Index stride)
+{
+  std::complex<float> EIGEN_ALIGN16 af[2];
+  pstore<std::complex<float> >((std::complex<float> *) af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+template<> EIGEN_DEVICE_FUNC inline void pscatter<std::complex<double>, Packet1cd>(std::complex<double>* to, const Packet1cd& from, Index stride EIGEN_UNUSED)
+{
+  pstore<std::complex<double> >(to, from);
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf padd<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(padd<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd padd<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v + b.v); }
+template<> EIGEN_STRONG_INLINE Packet2cf psub<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(psub<Packet4f>(a.v, b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd psub<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(a.v - b.v); }
+template<> EIGEN_STRONG_INLINE Packet1cd pnegate(const Packet1cd& a) { return Packet1cd(pnegate(Packet2d(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pnegate(const Packet2cf& a) { return Packet2cf(pnegate(Packet4f(a.v))); }
+template<> EIGEN_STRONG_INLINE Packet1cd pconj(const Packet1cd& a) { return Packet1cd((Packet2d)vec_xor((Packet2d)a.v, (Packet2d)p2ul_CONJ_XOR2)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pconj(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0]))).v;
+  res.v.v4f[1] = pconj(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pmul<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  Packet2d a_re, a_im, v1, v2;
+
+  // Permute and multiply the real parts of a and b
+  a_re = vec_perm(a.v, a.v, p16uc_PSET64_HI);
+  // Get the imaginary parts of a
+  a_im = vec_perm(a.v, a.v, p16uc_PSET64_LO);
+  // multiply a_re * b
+  v1 = vec_madd(a_re, b.v, p2d_ZERO);
+  // multiply a_im * b and get the conjugate result
+  v2 = vec_madd(a_im, b.v, p2d_ZERO);
+  v2 = (Packet2d) vec_sld((Packet4ui)v2, (Packet4ui)v2, 8);
+  v2 = (Packet2d) vec_xor((Packet2d)v2, (Packet2d) p2ul_CONJ_XOR1);
+
+  return Packet1cd(v1 + v2);
+}
+template<> EIGEN_STRONG_INLINE Packet2cf pmul<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  Packet2cf res;
+  res.v.v4f[0] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[0])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[0]))).v;
+  res.v.v4f[1] = pmul(Packet1cd(reinterpret_cast<Packet2d>(a.v.v4f[1])), Packet1cd(reinterpret_cast<Packet2d>(b.v.v4f[1]))).v;
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd pand   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pand   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pand<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd por    <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_or(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf por    <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(por<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pxor   <Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_xor(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet2cf pxor   <Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pxor<Packet4f>(a.v,b.v)); }
+template<> EIGEN_STRONG_INLINE Packet1cd pandnot<Packet1cd>(const Packet1cd& a, const Packet1cd& b) { return Packet1cd(vec_and(a.v, vec_nor(b.v,b.v))); }
+template<> EIGEN_STRONG_INLINE Packet2cf pandnot<Packet2cf>(const Packet2cf& a, const Packet2cf& b) { return Packet2cf(pandnot<Packet4f>(a.v,b.v)); }
+
+template<> EIGEN_STRONG_INLINE Packet1cd ploaddup<Packet1cd>(const std::complex<double>*     from) {  return pset1<Packet1cd>(*from); }
+template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<float>*      from) {  return pset1<Packet2cf>(*from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *     addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<double> >(const std::complex<double> *   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE std::complex<double>  pfirst<Packet1cd>(const Packet1cd& a)
+{
+  std::complex<double> EIGEN_ALIGN16 res;
+  pstore<std::complex<double> >(&res, a);
+
+  return res;
+}
+template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> EIGEN_ALIGN16 res[2];
+  pstore<std::complex<float> >(res, a);
+
+  return res[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preverse(const Packet1cd& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2cf preverse(const Packet2cf& a)
+{
+  Packet2cf res;
+  res.cd[0] = a.cd[1];
+  res.cd[1] = a.cd[0];
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+template<> EIGEN_STRONG_INLINE std::complex<float> predux<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = padd<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet1cd preduxp<Packet1cd>(const Packet1cd* vecs)
+{
+  return vecs[0];
+}
+template<> EIGEN_STRONG_INLINE Packet2cf preduxp<Packet2cf>(const Packet2cf* vecs)
+{
+  PacketBlock<Packet2cf,2> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  ptranspose(transpose);
+
+  return padd<Packet2cf>(transpose.packet[0], transpose.packet[1]);
+} 
+
+template<> EIGEN_STRONG_INLINE std::complex<double> predux_mul<Packet1cd>(const Packet1cd& a)
+{
+  return pfirst(a);
+}
+template<> EIGEN_STRONG_INLINE std::complex<float> predux_mul<Packet2cf>(const Packet2cf& a)
+{
+  std::complex<float> res;
+  Packet1cd b = pmul<Packet1cd>(a.cd[0], a.cd[1]);
+  vec_st2f(b.v, (float*)&res);
+  return res;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet1cd>
+{
+  static EIGEN_STRONG_INLINE void run(Packet1cd& /*first*/, const Packet1cd& /*second*/)
+  {
+    // FIXME is it sure we never have to align a Packet1cd?
+    // Even though a std::complex<double> has 16 bytes, it is not necessarily aligned on a 16 bytes boundary...
+  }
+};
+
+template<int Offset>
+struct palign_impl<Offset,Packet2cf>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2cf& first, const Packet2cf& second)
+  {
+    if (Offset == 1) {
+      first.cd[0] = first.cd[1];
+      first.cd[1] = second.cd[0];
+    }
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, false,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,false>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet1cd, Packet1cd, true,true>
+{
+  EIGEN_STRONG_INLINE Packet1cd pmadd(const Packet1cd& x, const Packet1cd& y, const Packet1cd& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet1cd pmul(const Packet1cd& a, const Packet1cd& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, false,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(a, pconj(b));
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,false>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return internal::pmul(pconj(a), b);
+  }
+};
+
+template<> struct conj_helper<Packet2cf, Packet2cf, true,true>
+{
+  EIGEN_STRONG_INLINE Packet2cf pmadd(const Packet2cf& x, const Packet2cf& y, const Packet2cf& c) const
+  { return padd(pmul(x,y),c); }
+
+  EIGEN_STRONG_INLINE Packet2cf pmul(const Packet2cf& a, const Packet2cf& b) const
+  {
+    return pconj(internal::pmul(a, b));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet1cd pdiv<Packet1cd>(const Packet1cd& a, const Packet1cd& b)
+{
+  // TODO optimize it for AltiVec
+  Packet1cd res = conj_helper<Packet1cd,Packet1cd,false,true>().pmul(a,b);
+  Packet2d s = vec_madd(b.v, b.v, p2d_ZERO_);
+  return Packet1cd(pdiv(res.v, s + vec_perm(s, s, p16uc_REVERSE64)));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pdiv<Packet2cf>(const Packet2cf& a, const Packet2cf& b)
+{
+  // TODO optimize it for AltiVec
+  Packet2cf res;
+  res.cd[0] = pdiv<Packet1cd>(a.cd[0], b.cd[0]);
+  res.cd[1] = pdiv<Packet1cd>(a.cd[1], b.cd[1]);
+  return res;
+}
+
+EIGEN_STRONG_INLINE Packet1cd pcplxflip/*<Packet1cd>*/(const Packet1cd& x)
+{
+  return Packet1cd(preverse(Packet2d(x.v)));
+}
+
+EIGEN_STRONG_INLINE Packet2cf pcplxflip/*<Packet2cf>*/(const Packet2cf& x)
+{
+  Packet2cf res;
+  res.cd[0] = pcplxflip(x.cd[0]);
+  res.cd[1] = pcplxflip(x.cd[1]);
+  return res;
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet1cd,2>& kernel)
+{
+  Packet2d tmp = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_HI);
+  kernel.packet[1].v = vec_perm(kernel.packet[0].v, kernel.packet[1].v, p16uc_TRANSPOSE64_LO);
+  kernel.packet[0].v = tmp;
+}
+
+EIGEN_STRONG_INLINE void ptranspose(PacketBlock<Packet2cf,2>& kernel)
+{
+  Packet1cd tmp = kernel.packet[0].cd[1];
+  kernel.packet[0].cd[1] = kernel.packet[1].cd[0];
+  kernel.packet[1].cd[0] = tmp;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2cf pblend(const Selector<2>& ifPacket, const Packet2cf& thenPacket, const Packet2cf& elsePacket) {
+  Packet2cf result;
+  const Selector<4> ifPacket4 = { ifPacket.select[0], ifPacket.select[0], ifPacket.select[1], ifPacket.select[1] };
+  result.v = pblend<Packet4f>(ifPacket4, thenPacket.v, elsePacket.v);
+  return result;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX32_ALTIVEC_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h
new file mode 100644
index 0000000..5c7aa72
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/MathFunctions.h
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007 Julien Pommier
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* The sin, cos, exp, and log functions of this file come from
+ * Julien Pommier's sse math library: http://gruntthepeon.free.fr/ssemath/
+ */
+
+#ifndef EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+#define EIGEN_MATH_FUNCTIONS_ALTIVEC_H
+
+namespace Eigen {
+
+namespace internal {
+
+static _EIGEN_DECLARE_CONST_Packet2d(1 , 1.0);
+static _EIGEN_DECLARE_CONST_Packet2d(2 , 2.0);
+static _EIGEN_DECLARE_CONST_Packet2d(half, 0.5);
+
+static _EIGEN_DECLARE_CONST_Packet2d(exp_hi,  709.437);
+static _EIGEN_DECLARE_CONST_Packet2d(exp_lo, -709.436139303);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_LOG2EF, 1.4426950408889634073599);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p0, 1.26177193074810590878e-4);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p1, 3.02994407707441961300e-2);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_p2, 9.99999999999999999910e-1);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q0, 3.00198505138664455042e-6);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q1, 2.52448340349684104192e-3);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q2, 2.27265548208155028766e-1);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_q3, 2.00000000000000000009e0);
+
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C1, 0.693145751953125);
+static _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d pexp<Packet2d>(const Packet2d& _x)
+{
+  Packet2d x = _x;
+
+  Packet2d tmp, fx;
+  Packet2l emm0;
+
+  // clamp x
+  x = pmax(pmin(x, p2d_exp_hi), p2d_exp_lo);
+  /* express exp(x) as exp(g + n*log(2)) */
+  fx = pmadd(p2d_cephes_LOG2EF, x, p2d_half);
+
+  fx = vec_floor(fx);
+
+  tmp = pmul(fx, p2d_cephes_exp_C1);
+  Packet2d z = pmul(fx, p2d_cephes_exp_C2);
+  x = psub(x, tmp);
+  x = psub(x, z);
+
+  Packet2d x2 = pmul(x,x);
+
+  Packet2d px = p2d_cephes_exp_p0;
+  px = pmadd(px, x2, p2d_cephes_exp_p1);
+  px = pmadd(px, x2, p2d_cephes_exp_p2);
+  px = pmul (px, x);
+
+  Packet2d qx = p2d_cephes_exp_q0;
+  qx = pmadd(qx, x2, p2d_cephes_exp_q1);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q2);
+  qx = pmadd(qx, x2, p2d_cephes_exp_q3);
+
+  x = pdiv(px,psub(qx,px));
+  x = pmadd(p2d_2,x,p2d_1);
+
+  // build 2^n
+  emm0 = vec_ctsl(fx, 0);
+
+  static const Packet2l p2l_1023 = { 1023, 1023 };
+  static const Packet2ul p2ul_52 = { 52, 52 };
+
+  emm0 = emm0 + p2l_1023;
+  emm0 = emm0 << reinterpret_cast<Packet2l>(p2ul_52);
+
+  // Altivec's max & min operators just drop silent NaNs. Check NaNs in 
+  // inputs and return them unmodified.
+  Packet2ul isnumber_mask = reinterpret_cast<Packet2ul>(vec_cmpeq(_x, _x));
+  return vec_sel(_x, pmax(pmul(x, reinterpret_cast<Packet2d>(emm0)), _x),
+                 isnumber_mask);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f pexp<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = pexp<Packet2d>(x.v4f[0]);
+  res.v4f[1] = pexp<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d psqrt<Packet2d>(const Packet2d& x)
+{
+  return  __builtin_s390_vfsqdb(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f psqrt<Packet4f>(const Packet4f& x)
+{
+  Packet4f res;
+  res.v4f[0] = psqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = psqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet2d prsqrt<Packet2d>(const Packet2d& x) {
+  // Unfortunately we can't use the much faster mm_rqsrt_pd since it only provides an approximation.
+  return pset1<Packet2d>(1.0) / psqrt<Packet2d>(x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
+Packet4f prsqrt<Packet4f>(const Packet4f& x) {
+  Packet4f res;
+  res.v4f[0] = prsqrt<Packet2d>(x.v4f[0]);
+  res.v4f[1] = prsqrt<Packet2d>(x.v4f[1]);
+  return res;
+}
+
+}  // end namespace internal
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_MATH_FUNCTIONS_ALTIVEC_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h
new file mode 100755
index 0000000..57b01fc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/arch/ZVector/PacketMath.h
@@ -0,0 +1,945 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Konstantinos Margaritis <markos@freevec.org>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PACKET_MATH_ZVECTOR_H
+#define EIGEN_PACKET_MATH_ZVECTOR_H
+
+#include <stdint.h>
+
+namespace Eigen {
+
+namespace internal {
+
+#ifndef EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD
+#define EIGEN_CACHEFRIENDLY_PRODUCT_THRESHOLD 4
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+#endif
+
+#ifndef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#define EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+#endif
+
+#ifndef EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS
+#define EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS  16
+#endif
+
+typedef __vector int                 Packet4i;
+typedef __vector unsigned int        Packet4ui;
+typedef __vector __bool int          Packet4bi;
+typedef __vector short int           Packet8i;
+typedef __vector unsigned char       Packet16uc;
+typedef __vector double              Packet2d;
+typedef __vector unsigned long long  Packet2ul;
+typedef __vector long long           Packet2l;
+
+typedef struct {
+	Packet2d  v4f[2];
+} Packet4f;
+
+typedef union {
+  int32_t   i[4];
+  uint32_t ui[4];
+  int64_t   l[2];
+  uint64_t ul[2];
+  double    d[2];
+  Packet4i  v4i;
+  Packet4ui v4ui;
+  Packet2l  v2l;
+  Packet2ul v2ul;
+  Packet2d  v2d;
+} Packet;
+
+// We don't want to write the same code all the time, but we need to reuse the constants
+// and it doesn't really work to declare them global, so we define macros instead
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = reinterpret_cast<Packet4i>(vec_splat_s32(X))
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = reinterpret_cast<Packet2d>(vec_splat_s64(X))
+
+#define _EIGEN_DECLARE_CONST_FAST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = reinterpret_cast<Packet2l>(vec_splat_s64(X))
+
+#define _EIGEN_DECLARE_CONST_Packet4i(NAME,X) \
+  Packet4i p4i_##NAME = pset1<Packet4i>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2d(NAME,X) \
+  Packet2d p2d_##NAME = pset1<Packet2d>(X)
+
+#define _EIGEN_DECLARE_CONST_Packet2l(NAME,X) \
+  Packet2l p2l_##NAME = pset1<Packet2l>(X)
+
+// These constants are endian-agnostic
+//static _EIGEN_DECLARE_CONST_FAST_Packet4i(ZERO, 0); //{ 0, 0, 0, 0,}
+static _EIGEN_DECLARE_CONST_FAST_Packet4i(ONE, 1); //{ 1, 1, 1, 1}
+
+static _EIGEN_DECLARE_CONST_FAST_Packet2d(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet2l(ZERO, 0);
+static _EIGEN_DECLARE_CONST_FAST_Packet2l(ONE, 1);
+
+static Packet2d p2d_ONE = { 1.0, 1.0 }; 
+static Packet2d p2d_ZERO_ = { -0.0, -0.0 };
+
+static Packet4i p4i_COUNTDOWN = { 0, 1, 2, 3 };
+static Packet4f p4f_COUNTDOWN = { 0.0, 1.0, 2.0, 3.0 };
+static Packet2d p2d_COUNTDOWN = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet16uc>(p2d_ZERO), reinterpret_cast<Packet16uc>(p2d_ONE), 8));
+
+static Packet16uc p16uc_PSET64_HI = { 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };
+static Packet16uc p16uc_DUPLICATE32_HI = { 0,1,2,3, 0,1,2,3, 4,5,6,7, 4,5,6,7 };
+
+// Mask alignment
+#define _EIGEN_MASK_ALIGNMENT	0xfffffffffffffff0
+
+#define _EIGEN_ALIGNED_PTR(x)	((std::ptrdiff_t)(x) & _EIGEN_MASK_ALIGNMENT)
+
+// Handle endianness properly while loading constants
+// Define global static constants:
+
+static Packet16uc p16uc_FORWARD =   { 0,1,2,3, 4,5,6,7, 8,9,10,11, 12,13,14,15 };
+static Packet16uc p16uc_REVERSE32 = { 12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3 };
+static Packet16uc p16uc_REVERSE64 = { 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+
+static Packet16uc p16uc_PSET32_WODD   = vec_sld((Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 0), (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 2), 8);//{ 0,1,2,3, 0,1,2,3, 8,9,10,11, 8,9,10,11 };
+static Packet16uc p16uc_PSET32_WEVEN  = vec_sld(p16uc_DUPLICATE32_HI, (Packet16uc) vec_splat((Packet4ui)p16uc_FORWARD, 3), 8);//{ 4,5,6,7, 4,5,6,7, 12,13,14,15, 12,13,14,15 };
+/*static Packet16uc p16uc_HALF64_0_16 = vec_sld((Packet16uc)p4i_ZERO, vec_splat((Packet16uc) vec_abs(p4i_MINUS16), 3), 8);      //{ 0,0,0,0, 0,0,0,0, 16,16,16,16, 16,16,16,16};
+
+static Packet16uc p16uc_PSET64_HI = (Packet16uc) vec_mergeh((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 0,1,2,3, 4,5,6,7, 0,1,2,3, 4,5,6,7 };*/
+static Packet16uc p16uc_PSET64_LO = (Packet16uc) vec_mergel((Packet4ui)p16uc_PSET32_WODD, (Packet4ui)p16uc_PSET32_WEVEN);     //{ 8,9,10,11, 12,13,14,15, 8,9,10,11, 12,13,14,15 };
+/*static Packet16uc p16uc_TRANSPOSE64_HI = vec_add(p16uc_PSET64_HI, p16uc_HALF64_0_16);                                         //{ 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = vec_add(p16uc_PSET64_LO, p16uc_HALF64_0_16);                                         //{ 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};*/
+static Packet16uc p16uc_TRANSPOSE64_HI = { 0,1,2,3, 4,5,6,7, 16,17,18,19, 20,21,22,23};
+static Packet16uc p16uc_TRANSPOSE64_LO = { 8,9,10,11, 12,13,14,15, 24,25,26,27, 28,29,30,31};
+
+//static Packet16uc p16uc_COMPLEX32_REV = vec_sld(p16uc_REVERSE32, p16uc_REVERSE32, 8);                                         //{ 4,5,6,7, 0,1,2,3, 12,13,14,15, 8,9,10,11 };
+
+//static Packet16uc p16uc_COMPLEX32_REV2 = vec_sld(p16uc_FORWARD, p16uc_FORWARD, 8);                                            //{ 8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7 };
+
+
+#if EIGEN_HAS_BUILTIN(__builtin_prefetch) || EIGEN_COMP_GNUC
+  #define EIGEN_ZVECTOR_PREFETCH(ADDR) __builtin_prefetch(ADDR);
+#else
+  #define EIGEN_ZVECTOR_PREFETCH(ADDR) asm( "   pfd [%[addr]]\n" :: [addr] "r" (ADDR) : "cc" );
+#endif
+
+template<> struct packet_traits<int>    : default_packet_traits
+{
+  typedef Packet4i type;
+  typedef Packet4i half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size = 4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct packet_traits<float> : default_packet_traits
+{
+  typedef Packet4f type;
+  typedef Packet4f half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=4,
+    HasHalfPacket = 0,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct packet_traits<double> : default_packet_traits
+{
+  typedef Packet2d type;
+  typedef Packet2d half;
+  enum {
+    Vectorizable = 1,
+    AlignedOnScalar = 1,
+    size=2,
+    HasHalfPacket = 1,
+
+    HasAdd  = 1,
+    HasSub  = 1,
+    HasMul  = 1,
+    HasDiv  = 1,
+    HasMin  = 1,
+    HasMax  = 1,
+    HasAbs  = 1,
+    HasSin  = 0,
+    HasCos  = 0,
+    HasLog  = 0,
+    HasExp  = 1,
+    HasSqrt = 1,
+    HasRsqrt = 1,
+    HasRound = 1,
+    HasFloor = 1,
+    HasCeil = 1,
+    HasNegate = 1,
+    HasBlend = 1
+  };
+};
+
+template<> struct unpacket_traits<Packet4i> { typedef int    type; enum {size=4, alignment=Aligned16}; typedef Packet4i half; };
+template<> struct unpacket_traits<Packet4f> { typedef float  type; enum {size=4, alignment=Aligned16}; typedef Packet4f half; };
+template<> struct unpacket_traits<Packet2d> { typedef double type; enum {size=2, alignment=Aligned16}; typedef Packet2d half; };
+
+/* Forward declaration */
+EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4f,4>& kernel);
+ 
+inline std::ostream & operator <<(std::ostream & s, const Packet4i & v)
+{
+  Packet vt;
+  vt.v4i = v;
+  s << vt.i[0] << ", " << vt.i[1] << ", " << vt.i[2] << ", " << vt.i[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet4ui & v)
+{
+  Packet vt;
+  vt.v4ui = v;
+  s << vt.ui[0] << ", " << vt.ui[1] << ", " << vt.ui[2] << ", " << vt.ui[3];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2l & v)
+{
+  Packet vt;
+  vt.v2l = v;
+  s << vt.l[0] << ", " << vt.l[1];
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2ul & v)
+{
+  Packet vt;
+  vt.v2ul = v;
+  s << vt.ul[0] << ", " << vt.ul[1] ;
+  return s;
+}
+
+inline std::ostream & operator <<(std::ostream & s, const Packet2d & v)
+{
+  Packet vt;
+  vt.v2d = v;
+  s << vt.d[0] << ", " << vt.d[1];
+  return s;
+}
+
+/* Helper function to simulate a vec_splat_packet4f
+ */
+template<int element> EIGEN_STRONG_INLINE Packet4f vec_splat_packet4f(const Packet4f&   from)
+{
+  Packet4f splat;
+  switch (element) {
+  case 0:
+    splat.v4f[0] = vec_splat(from.v4f[0], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 1:
+    splat.v4f[0] = vec_splat(from.v4f[0], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 2:
+    splat.v4f[0] = vec_splat(from.v4f[1], 0);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  case 3:
+    splat.v4f[0] = vec_splat(from.v4f[1], 1);
+    splat.v4f[1] = splat.v4f[0];
+    break;
+  }
+  return splat;
+}
+
+template<int Offset>
+struct palign_impl<Offset,Packet4i>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4i& first, const Packet4i& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first = vec_sld(first, second, 4); break;
+    case 2:
+      first = vec_sld(first, second, 8); break;
+    case 3:
+      first = vec_sld(first, second, 12); break;
+    }
+  }
+};
+
+/* This is a tricky one, we have to translate float alignment to vector elements of sizeof double
+ */
+template<int Offset>
+struct palign_impl<Offset,Packet4f>
+{
+  static EIGEN_STRONG_INLINE void run(Packet4f& first, const Packet4f& second)
+  {
+    switch (Offset % 4) {
+    case 1:
+      first.v4f[0] = vec_sld(first.v4f[0], first.v4f[1], 8);
+      first.v4f[1] = vec_sld(first.v4f[1], second.v4f[0], 8);
+      break;
+    case 2:
+      first.v4f[0] = first.v4f[1];
+      first.v4f[1] = second.v4f[0];
+      break;
+    case 3:
+      first.v4f[0] = vec_sld(first.v4f[1],  second.v4f[0], 8);
+      first.v4f[1] = vec_sld(second.v4f[0], second.v4f[1], 8);
+      break;
+    }
+  }
+};
+
+
+template<int Offset>
+struct palign_impl<Offset,Packet2d>
+{
+  static EIGEN_STRONG_INLINE void run(Packet2d& first, const Packet2d& second)
+  {
+    if (Offset == 1)
+      first = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(first), reinterpret_cast<Packet4i>(second), 8));
+  }
+};
+
+template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet *vfrom;
+  vfrom = (Packet *) from;
+  return vfrom->v4i;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pload<Packet4f>(const float*   from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet4f vfrom;
+  vfrom.v4f[0] = vec_ld2f(&from[0]);
+  vfrom.v4f[1] = vec_ld2f(&from[2]);
+  return vfrom;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pload<Packet2d>(const double* from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_LOAD
+  Packet *vfrom;
+  vfrom = (Packet *) from;
+  return vfrom->v2d;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  Packet *vto;
+  vto = (Packet *) to;
+  vto->v4i = from;
+}
+
+template<> EIGEN_STRONG_INLINE void pstore<float>(float*   to, const Packet4f& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  vec_st2f(from.v4f[0], &to[0]);
+  vec_st2f(from.v4f[1], &to[2]);
+}
+
+
+template<> EIGEN_STRONG_INLINE void pstore<double>(double*   to, const Packet2d& from)
+{
+  // FIXME: No intrinsic yet
+  EIGEN_DEBUG_ALIGNED_STORE
+  Packet *vto;
+  vto = (Packet *) to;
+  vto->v2d = from;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pset1<Packet4i>(const int&    from)
+{
+  return vec_splats(from);
+}
+template<> EIGEN_STRONG_INLINE Packet2d pset1<Packet2d>(const double& from) {
+  return vec_splats(from);
+}
+template<> EIGEN_STRONG_INLINE Packet4f pset1<Packet4f>(const float&    from)
+{
+  Packet4f to;
+  to.v4f[0] = pset1<Packet2d>(static_cast<const double&>(from));
+  to.v4f[1] = to.v4f[0];
+  return to;
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4i>(const int *a,
+                      Packet4i& a0, Packet4i& a1, Packet4i& a2, Packet4i& a3)
+{
+  a3 = pload<Packet4i>(a);
+  a0 = vec_splat(a3, 0);
+  a1 = vec_splat(a3, 1);
+  a2 = vec_splat(a3, 2);
+  a3 = vec_splat(a3, 3);
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet4f>(const float *a,
+                      Packet4f& a0, Packet4f& a1, Packet4f& a2, Packet4f& a3)
+{
+  a3 = pload<Packet4f>(a);
+  a0 = vec_splat_packet4f<0>(a3);
+  a1 = vec_splat_packet4f<1>(a3);
+  a2 = vec_splat_packet4f<2>(a3);
+  a3 = vec_splat_packet4f<3>(a3);
+}
+
+template<> EIGEN_STRONG_INLINE void
+pbroadcast4<Packet2d>(const double *a,
+                      Packet2d& a0, Packet2d& a1, Packet2d& a2, Packet2d& a3)
+{
+  a1 = pload<Packet2d>(a);
+  a0 = vec_splat(a1, 0);
+  a1 = vec_splat(a1, 1);
+  a3 = pload<Packet2d>(a+2);
+  a2 = vec_splat(a3, 0);
+  a3 = vec_splat(a3, 1);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4i pgather<int, Packet4i>(const int* from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4i>(ai);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet4f pgather<float, Packet4f>(const float* from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  ai[0] = from[0*stride];
+  ai[1] = from[1*stride];
+  ai[2] = from[2*stride];
+  ai[3] = from[3*stride];
+ return pload<Packet4f>(ai);
+}
+
+template<> EIGEN_DEVICE_FUNC inline Packet2d pgather<double, Packet2d>(const double* from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  af[0] = from[0*stride];
+  af[1] = from[1*stride];
+ return pload<Packet2d>(af);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<int, Packet4i>(int* to, const Packet4i& from, Index stride)
+{
+  int EIGEN_ALIGN16 ai[4];
+  pstore<int>((int *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<float, Packet4f>(float* to, const Packet4f& from, Index stride)
+{
+  float EIGEN_ALIGN16 ai[4];
+  pstore<float>((float *)ai, from);
+  to[0*stride] = ai[0];
+  to[1*stride] = ai[1];
+  to[2*stride] = ai[2];
+  to[3*stride] = ai[3];
+}
+
+template<> EIGEN_DEVICE_FUNC inline void pscatter<double, Packet2d>(double* to, const Packet2d& from, Index stride)
+{
+  double EIGEN_ALIGN16 af[2];
+  pstore<double>(af, from);
+  to[0*stride] = af[0];
+  to[1*stride] = af[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i padd<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a + b); }
+template<> EIGEN_STRONG_INLINE Packet4f padd<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] + b.v4f[0];
+  c.v4f[1] = a.v4f[1] + b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d padd<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a + b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i psub<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a - b); }
+template<> EIGEN_STRONG_INLINE Packet4f psub<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] - b.v4f[0];
+  c.v4f[1] = a.v4f[1] - b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d psub<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a - b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmul<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a * b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmul<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] * b.v4f[0];
+  c.v4f[1] = a.v4f[1] * b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmul<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a * b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pdiv<Packet4i>(const Packet4i& a, const Packet4i& b) { return (a / b); }
+template<> EIGEN_STRONG_INLINE Packet4f pdiv<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f c;
+  c.v4f[0] = a.v4f[0] / b.v4f[0];
+  c.v4f[1] = a.v4f[1] / b.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pdiv<Packet2d>(const Packet2d& a, const Packet2d& b) { return (a / b); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pnegate(const Packet4i& a) { return (-a); }
+template<> EIGEN_STRONG_INLINE Packet4f pnegate(const Packet4f& a)
+{
+  Packet4f c;
+  c.v4f[0] = -a.v4f[0];
+  c.v4f[1] = -a.v4f[1];
+  return c;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pnegate(const Packet2d& a) { return (-a); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pconj(const Packet4i& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet4f pconj(const Packet4f& a) { return a; }
+template<> EIGEN_STRONG_INLINE Packet2d pconj(const Packet2d& a) { return a; }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmadd(const Packet4i& a, const Packet4i& b, const Packet4i& c) { return padd<Packet4i>(pmul<Packet4i>(a, b), c); }
+template<> EIGEN_STRONG_INLINE Packet4f pmadd(const Packet4f& a, const Packet4f& b, const Packet4f& c)
+{
+  Packet4f res;
+  res.v4f[0] = vec_madd(a.v4f[0], b.v4f[0], c.v4f[0]);
+  res.v4f[1] = vec_madd(a.v4f[1], b.v4f[1], c.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pmadd(const Packet2d& a, const Packet2d& b, const Packet2d& c) { return vec_madd(a, b, c); }
+
+template<> EIGEN_STRONG_INLINE Packet4i plset<Packet4i>(const int& a)    { return padd<Packet4i>(pset1<Packet4i>(a), p4i_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet4f plset<Packet4f>(const float& a)  { return padd<Packet4f>(pset1<Packet4f>(a), p4f_COUNTDOWN); }
+template<> EIGEN_STRONG_INLINE Packet2d plset<Packet2d>(const double& a) { return padd<Packet2d>(pset1<Packet2d>(a), p2d_COUNTDOWN); }
+
+template<> EIGEN_STRONG_INLINE Packet4i pmin<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmin<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_min(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmin<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmin(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmin(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pmax<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pmax<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_max(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pmax<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pmax(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pmax(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pand<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pand<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pand<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i por<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d por<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_or(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f por<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pxor<Packet4i>(const Packet4i& a, const Packet4i& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet2d pxor<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_xor(a, b); }
+template<> EIGEN_STRONG_INLINE Packet4f pxor<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pand(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pand(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pandnot<Packet4i>(const Packet4i& a, const Packet4i& b) { return pand<Packet4i>(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet2d pandnot<Packet2d>(const Packet2d& a, const Packet2d& b) { return vec_and(a, vec_nor(b, b)); }
+template<> EIGEN_STRONG_INLINE Packet4f pandnot<Packet4f>(const Packet4f& a, const Packet4f& b)
+{
+  Packet4f res;
+  res.v4f[0] = pandnot(a.v4f[0], b.v4f[0]);
+  res.v4f[1] = pandnot(a.v4f[1], b.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pround<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_round(a.v4f[0]);
+  res.v4f[1] = vec_round(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pround<Packet2d>(const Packet2d& a) { return vec_round(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const  Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_ceil(a.v4f[0]);
+  res.v4f[1] = vec_ceil(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pceil<Packet2d>(const  Packet2d& a) { return vec_ceil(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = vec_floor(a.v4f[0]);
+  res.v4f[1] = vec_floor(a.v4f[1]);
+  return res;
+}
+template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { return vec_floor(a); }
+
+template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*       from) { return pload<Packet4i>(from); }
+template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float*     from) { return pload<Packet4f>(from); }
+template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double*    from) { return pload<Packet2d>(from); }
+
+
+template<> EIGEN_STRONG_INLINE Packet4i ploaddup<Packet4i>(const int*     from)
+{
+  Packet4i p = pload<Packet4i>(from);
+  return vec_perm(p, p, p16uc_DUPLICATE32_HI);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*    from)
+{
+  Packet4f p = pload<Packet4f>(from);
+  p.v4f[1] = vec_splat(p.v4f[0], 1);
+  p.v4f[0] = vec_splat(p.v4f[0], 0);
+  return p;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d ploaddup<Packet2d>(const double*   from)
+{
+  Packet2d p = pload<Packet2d>(from);
+  return vec_perm(p, p, p16uc_PSET64_HI);
+}
+
+template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*        to, const Packet4i& from) { pstore<int>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*    to, const Packet4f& from) { pstore<float>(to, from); }
+template<> EIGEN_STRONG_INLINE void pstoreu<double>(double*  to, const Packet2d& from) { pstore<double>(to, from); }
+
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*       addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float*   addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<double>(const double* addr) { EIGEN_ZVECTOR_PREFETCH(addr); }
+
+template<> EIGEN_STRONG_INLINE int    pfirst<Packet4i>(const Packet4i& a) { int    EIGEN_ALIGN16 x[4]; pstore(x, a); return x[0]; }
+template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float  EIGEN_ALIGN16 x[2]; vec_st2f(a.v4f[0], &x[0]); return x[0]; }
+template<> EIGEN_STRONG_INLINE double pfirst<Packet2d>(const Packet2d& a) { double EIGEN_ALIGN16 x[2]; pstore(x, a); return x[0]; }
+
+template<> EIGEN_STRONG_INLINE Packet4i preverse(const Packet4i& a)
+{
+  return reinterpret_cast<Packet4i>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE32));
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preverse(const Packet2d& a)
+{
+  return reinterpret_cast<Packet2d>(vec_perm(reinterpret_cast<Packet16uc>(a), reinterpret_cast<Packet16uc>(a), p16uc_REVERSE64));
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preverse(const Packet4f& a)
+{
+  Packet4f rev;
+  rev.v4f[0] = preverse<Packet2d>(a.v4f[1]);
+  rev.v4f[1] = preverse<Packet2d>(a.v4f[0]);
+  return rev;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pabs<Packet4i>(const Packet4i& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet2d pabs<Packet2d>(const Packet2d& a) { return vec_abs(a); }
+template<> EIGEN_STRONG_INLINE Packet4f pabs<Packet4f>(const Packet4f& a)
+{
+  Packet4f res;
+  res.v4f[0] = pabs(a.v4f[0]);
+  res.v4f[1] = pabs(a.v4f[1]);
+  return res;
+}
+
+template<> EIGEN_STRONG_INLINE int predux<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, sum;
+  b   = vec_sld(a, a, 8);
+  sum = padd<Packet4i>(a, b);
+  b   = vec_sld(sum, sum, 4);
+  sum = padd<Packet4i>(sum, b);
+  return pfirst(sum);
+}
+
+template<> EIGEN_STRONG_INLINE double predux<Packet2d>(const Packet2d& a)
+{
+  Packet2d b, sum;
+  b   = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8));
+  sum = padd<Packet2d>(a, b);
+  return pfirst(sum);
+}
+template<> EIGEN_STRONG_INLINE float predux<Packet4f>(const Packet4f& a)
+{
+  Packet2d sum;
+  sum = padd<Packet2d>(a.v4f[0], a.v4f[1]);
+  double first = predux<Packet2d>(sum);
+  return static_cast<float>(first);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i preduxp<Packet4i>(const Packet4i* vecs)
+{
+  Packet4i v[4], sum[4];
+
+  // It's easier and faster to transpose then add as columns
+  // Check: http://www.freevec.org/function/matrix_4x4_transpose_floats for explanation
+  // Do the transpose, first set of moves
+  v[0] = vec_mergeh(vecs[0], vecs[2]);
+  v[1] = vec_mergel(vecs[0], vecs[2]);
+  v[2] = vec_mergeh(vecs[1], vecs[3]);
+  v[3] = vec_mergel(vecs[1], vecs[3]);
+  // Get the resulting vectors
+  sum[0] = vec_mergeh(v[0], v[2]);
+  sum[1] = vec_mergel(v[0], v[2]);
+  sum[2] = vec_mergeh(v[1], v[3]);
+  sum[3] = vec_mergel(v[1], v[3]);
+
+  // Now do the summation:
+  // Lines 0+1
+  sum[0] = padd<Packet4i>(sum[0], sum[1]);
+  // Lines 2+3
+  sum[1] = padd<Packet4i>(sum[2], sum[3]);
+  // Add the results
+  sum[0] = padd<Packet4i>(sum[0], sum[1]);
+
+  return sum[0];
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d preduxp<Packet2d>(const Packet2d* vecs)
+{
+  Packet2d v[2], sum;
+  v[0] = padd<Packet2d>(vecs[0], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[0]), reinterpret_cast<Packet4ui>(vecs[0]), 8)));
+  v[1] = padd<Packet2d>(vecs[1], reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(vecs[1]), reinterpret_cast<Packet4ui>(vecs[1]), 8)));
+ 
+  sum = reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4ui>(v[0]), reinterpret_cast<Packet4ui>(v[1]), 8));
+
+  return sum;
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f preduxp<Packet4f>(const Packet4f* vecs)
+{
+  PacketBlock<Packet4f,4> transpose;
+  transpose.packet[0] = vecs[0];
+  transpose.packet[1] = vecs[1];
+  transpose.packet[2] = vecs[2];
+  transpose.packet[3] = vecs[3];
+  ptranspose(transpose);
+
+  Packet4f sum = padd(transpose.packet[0], transpose.packet[1]);
+  sum = padd(sum, transpose.packet[2]);
+  sum = padd(sum, transpose.packet[3]);
+  return sum;
+}
+
+// Other reduction functions:
+// mul
+template<> EIGEN_STRONG_INLINE int predux_mul<Packet4i>(const Packet4i& a)
+{
+  EIGEN_ALIGN16 int aux[4];
+  pstore(aux, a);
+  return aux[0] * aux[1] * aux[2] * aux[3];
+}
+
+template<> EIGEN_STRONG_INLINE double predux_mul<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmul(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_mul<Packet4f>(const Packet4f& a)
+{
+  // Return predux_mul<Packet2d> of the subvectors product
+  return static_cast<float>(pfirst(predux_mul(pmul(a.v4f[0], a.v4f[1]))));
+}
+
+// min
+template<> EIGEN_STRONG_INLINE int predux_min<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b   = pmin<Packet4i>(a, vec_sld(a, a, 8));
+  res = pmin<Packet4i>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+template<> EIGEN_STRONG_INLINE double predux_min<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmin<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_min<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmin<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmin<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
+// max
+template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
+{
+  Packet4i b, res;
+  b = pmax<Packet4i>(a, vec_sld(a, a, 8));
+  res = pmax<Packet4i>(b, vec_sld(b, b, 4));
+  return pfirst(res);
+}
+
+// max
+template<> EIGEN_STRONG_INLINE double predux_max<Packet2d>(const Packet2d& a)
+{
+  return pfirst(pmax<Packet2d>(a, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(a), reinterpret_cast<Packet4i>(a), 8))));
+}
+
+template<> EIGEN_STRONG_INLINE float predux_max<Packet4f>(const Packet4f& a)
+{
+  Packet2d b, res;
+  b   = pmax<Packet2d>(a.v4f[0], a.v4f[1]);
+  res = pmax<Packet2d>(b, reinterpret_cast<Packet2d>(vec_sld(reinterpret_cast<Packet4i>(b), reinterpret_cast<Packet4i>(b), 8)));
+  return static_cast<float>(pfirst(res));
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4i,4>& kernel) {
+  Packet4i t0 = vec_mergeh(kernel.packet[0], kernel.packet[2]);
+  Packet4i t1 = vec_mergel(kernel.packet[0], kernel.packet[2]);
+  Packet4i t2 = vec_mergeh(kernel.packet[1], kernel.packet[3]);
+  Packet4i t3 = vec_mergel(kernel.packet[1], kernel.packet[3]);
+  kernel.packet[0] = vec_mergeh(t0, t2);
+  kernel.packet[1] = vec_mergel(t0, t2);
+  kernel.packet[2] = vec_mergeh(t1, t3);
+  kernel.packet[3] = vec_mergel(t1, t3);
+}
+
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet2d,2>& kernel) {
+  Packet2d t0 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_HI);
+  Packet2d t1 = vec_perm(kernel.packet[0], kernel.packet[1], p16uc_TRANSPOSE64_LO);
+  kernel.packet[0] = t0;
+  kernel.packet[1] = t1;
+}
+
+/* Split the Packet4f PacketBlock into 4 Packet2d PacketBlocks and transpose each one
+ */
+EIGEN_DEVICE_FUNC inline void
+ptranspose(PacketBlock<Packet4f,4>& kernel) {
+  PacketBlock<Packet2d,2> t0,t1,t2,t3;
+  // copy top-left 2x2 Packet2d block
+  t0.packet[0] = kernel.packet[0].v4f[0];
+  t0.packet[1] = kernel.packet[1].v4f[0];
+
+  // copy top-right 2x2 Packet2d block
+  t1.packet[0] = kernel.packet[0].v4f[1];
+  t1.packet[1] = kernel.packet[1].v4f[1];
+
+  // copy bottom-left 2x2 Packet2d block
+  t2.packet[0] = kernel.packet[2].v4f[0];
+  t2.packet[1] = kernel.packet[3].v4f[0];
+
+  // copy bottom-right 2x2 Packet2d block
+  t3.packet[0] = kernel.packet[2].v4f[1];
+  t3.packet[1] = kernel.packet[3].v4f[1];
+
+  // Transpose all 2x2 blocks
+  ptranspose(t0);
+  ptranspose(t1);
+  ptranspose(t2);
+  ptranspose(t3);
+
+  // Copy back transposed blocks, but exchange t1 and t2 due to transposition
+  kernel.packet[0].v4f[0] = t0.packet[0];
+  kernel.packet[0].v4f[1] = t2.packet[0];
+  kernel.packet[1].v4f[0] = t0.packet[1];
+  kernel.packet[1].v4f[1] = t2.packet[1];
+  kernel.packet[2].v4f[0] = t1.packet[0];
+  kernel.packet[2].v4f[1] = t3.packet[0];
+  kernel.packet[3].v4f[0] = t1.packet[1];
+  kernel.packet[3].v4f[1] = t3.packet[1];
+}
+
+template<> EIGEN_STRONG_INLINE Packet4i pblend(const Selector<4>& ifPacket, const Packet4i& thenPacket, const Packet4i& elsePacket) {
+  Packet4ui select = { ifPacket.select[0], ifPacket.select[1], ifPacket.select[2], ifPacket.select[3] };
+  Packet4ui mask = vec_cmpeq(select, reinterpret_cast<Packet4ui>(p4i_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+template<> EIGEN_STRONG_INLINE Packet4f pblend(const Selector<4>& ifPacket, const Packet4f& thenPacket, const Packet4f& elsePacket) {
+  Packet2ul select_hi = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul select_lo = { ifPacket.select[2], ifPacket.select[3] };
+  Packet2ul mask_hi = vec_cmpeq(select_hi, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet2ul mask_lo = vec_cmpeq(select_lo, reinterpret_cast<Packet2ul>(p2l_ONE));
+  Packet4f result;
+  result.v4f[0] = vec_sel(elsePacket.v4f[0], thenPacket.v4f[0], mask_hi);
+  result.v4f[1] = vec_sel(elsePacket.v4f[1], thenPacket.v4f[1], mask_lo);
+  return result;
+}
+
+template<> EIGEN_STRONG_INLINE Packet2d pblend(const Selector<2>& ifPacket, const Packet2d& thenPacket, const Packet2d& elsePacket) {
+  Packet2ul select = { ifPacket.select[0], ifPacket.select[1] };
+  Packet2ul mask = vec_cmpeq(select, reinterpret_cast<Packet2ul>(p2l_ONE));
+  return vec_sel(elsePacket, thenPacket, mask);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PACKET_MATH_ZVECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h
new file mode 100644
index 0000000..4153b87
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/AssignmentFunctors.h
@@ -0,0 +1,168 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ASSIGNMENT_FUNCTORS_H
+#define EIGEN_ASSIGNMENT_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+  
+/** \internal
+  * \brief Template functor for scalar/packet assignment
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a = b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,b); }
+};
+
+// Empty overload for void type (used by PermutationMatrix)
+template<typename DstScalar> struct assign_op<DstScalar,void> {};
+
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::Vectorizable && packet_traits<SrcScalar>::Vectorizable
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with addition
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct add_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(add_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a += b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::padd(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<add_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasAdd
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with subtraction
+  *
+  */
+template<typename DstScalar,typename SrcScalar> struct sub_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(sub_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a -= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::psub(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar,typename SrcScalar>
+struct functor_traits<sub_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::AddCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasSub
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with multiplication
+  *
+  */
+template<typename DstScalar, typename SrcScalar=DstScalar>
+struct mul_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(mul_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a *= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pmul(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar, typename SrcScalar>
+struct functor_traits<mul_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasMul
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with diviving
+  *
+  */
+template<typename DstScalar, typename SrcScalar=DstScalar> struct div_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(div_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(DstScalar& a, const SrcScalar& b) const { a /= b; }
+  
+  template<int Alignment, typename Packet>
+  EIGEN_STRONG_INLINE void assignPacket(DstScalar* a, const Packet& b) const
+  { internal::pstoret<DstScalar,Packet,Alignment>(a,internal::pdiv(internal::ploadt<Packet,Alignment>(a),b)); }
+};
+template<typename DstScalar, typename SrcScalar>
+struct functor_traits<div_assign_op<DstScalar,SrcScalar> > {
+  enum {
+    Cost = NumTraits<DstScalar>::ReadCost + NumTraits<DstScalar>::MulCost,
+    PacketAccess = is_same<DstScalar,SrcScalar>::value && packet_traits<DstScalar>::HasDiv
+  };
+};
+
+/** \internal
+  * \brief Template functor for scalar/packet assignment with swapping
+  *
+  * It works as follow. For a non-vectorized evaluation loop, we have:
+  *   for(i) func(A.coeffRef(i), B.coeff(i));
+  * where B is a SwapWrapper expression. The trick is to make SwapWrapper::coeff behaves like a non-const coeffRef.
+  * Actually, SwapWrapper might not even be needed since even if B is a plain expression, since it has to be writable
+  * B.coeff already returns a const reference to the underlying scalar value.
+  * 
+  * The case of a vectorized loop is more tricky:
+  *   for(i,j) func.assignPacket<A_Align>(&A.coeffRef(i,j), B.packet<B_Align>(i,j));
+  * Here, B must be a SwapWrapper whose packet function actually returns a proxy object holding a Scalar*,
+  * the actual alignment and Packet type.
+  *
+  */
+template<typename Scalar> struct swap_assign_op {
+
+  EIGEN_EMPTY_STRUCT_CTOR(swap_assign_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void assignCoeff(Scalar& a, const Scalar& b) const
+  {
+#ifdef __CUDACC__
+    // FIXME is there some kind of cuda::swap?
+    Scalar t=b; const_cast<Scalar&>(b)=a; a=t;
+#else
+    using std::swap;
+    swap(a,const_cast<Scalar&>(b));
+#endif
+  }
+};
+template<typename Scalar>
+struct functor_traits<swap_assign_op<Scalar> > {
+  enum {
+    Cost = 3 * NumTraits<Scalar>::ReadCost,
+    PacketAccess = packet_traits<Scalar>::Vectorizable
+  };
+};
+
+} // namespace internal
+
+} // namespace Eigen
+
+#endif // EIGEN_ASSIGNMENT_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h
new file mode 100644
index 0000000..96747ba
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/BinaryFunctors.h
@@ -0,0 +1,482 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BINARY_FUNCTORS_H
+#define EIGEN_BINARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- associative binary functors ----------
+
+template<typename Arg1, typename Arg2>
+struct binary_op_base
+{
+  typedef Arg1 first_argument_type;
+  typedef Arg2 second_argument_type;
+};
+
+/** \internal
+  * \brief Template functor to compute the sum of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::operator+, class VectorwiseOp, DenseBase::sum()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_sum_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_sum_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sum_op)
+#else
+  scalar_sum_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a + b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::padd(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_sum_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2, // rough estimate!
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasAdd && packet_traits<RhsScalar>::HasAdd
+    // TODO vectorize mixed sum
+  };
+};
+
+/** \internal
+  * \brief Template specialization to deprecate the summation of boolean expressions.
+  * This is required to solve Bug 426.
+  * \sa DenseBase::count(), DenseBase::any(), ArrayBase::cast(), MatrixBase::cast()
+  */
+template<> struct scalar_sum_op<bool,bool> : scalar_sum_op<int,int> {
+  EIGEN_DEPRECATED
+  scalar_sum_op() {}
+};
+
+
+/** \internal
+  * \brief Template functor to compute the product of two scalars
+  *
+  * \sa class CwiseBinaryOp, Cwise::operator*(), class VectorwiseOp, MatrixBase::redux()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_product_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_product_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_product_op)
+#else
+  scalar_product_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmul(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_mul(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_product_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::MulCost + NumTraits<RhsScalar>::MulCost)/2, // rough estimate!
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasMul && packet_traits<RhsScalar>::HasMul
+    // TODO vectorize mixed product
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the conjugate product of two scalars
+  *
+  * This is a short cut for conj(x) * y which is needed for optimization purpose; in Eigen2 support mode, this becomes x * conj(y)
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_conj_product_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+
+  enum {
+    Conj = NumTraits<LhsScalar>::IsComplex
+  };
+  
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_conj_product_op>::ReturnType result_type;
+  
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_conj_product_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const
+  { return conj_helper<LhsScalar,RhsScalar,Conj,false>().pmul(a,b); }
+  
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return conj_helper<Packet,Packet,Conj,false>().pmul(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_conj_product_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = NumTraits<LhsScalar>::MulCost,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMul
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the min of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::cwiseMin, class VectorwiseOp, MatrixBase::minCoeff()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_min_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_min_op>::ReturnType result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_min_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::mini(a, b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmin(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_min(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_min_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMin
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the max of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::cwiseMax, class VectorwiseOp, MatrixBase::maxCoeff()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_max_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_max_op>::ReturnType result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_max_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return numext::maxi(a, b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pmax(a,b); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type predux(const Packet& a) const
+  { return internal::predux_max(a); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_max_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = internal::is_same<LhsScalar, RhsScalar>::value && packet_traits<LhsScalar>::HasMax
+  };
+};
+
+/** \internal
+  * \brief Template functors for comparison of two scalars
+  * \todo Implement packet-comparisons
+  */
+template<typename LhsScalar, typename RhsScalar, ComparisonName cmp> struct scalar_cmp_op;
+
+template<typename LhsScalar, typename RhsScalar, ComparisonName cmp>
+struct functor_traits<scalar_cmp_op<LhsScalar,RhsScalar, cmp> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = false
+  };
+};
+
+template<ComparisonName Cmp, typename LhsScalar, typename RhsScalar>
+struct result_of<scalar_cmp_op<LhsScalar, RhsScalar, Cmp>(LhsScalar,RhsScalar)> {
+  typedef bool type;
+};
+
+
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_EQ> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a==b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LT> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_LE> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a<=b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GT> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_GE> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a>=b;}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_UNORD> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return !(a<=b || b<=a);}
+};
+template<typename LhsScalar, typename RhsScalar>
+struct scalar_cmp_op<LhsScalar,RhsScalar, cmp_NEQ> : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef bool result_type;
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator()(const LhsScalar& a, const RhsScalar& b) const {return a!=b;}
+};
+
+
+/** \internal
+  * \brief Template functor to compute the hypot of two scalars
+  *
+  * \sa MatrixBase::stableNorm(), class Redux
+  */
+template<typename Scalar>
+struct scalar_hypot_op<Scalar,Scalar> : binary_op_base<Scalar,Scalar>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_hypot_op)
+//   typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& _x, const Scalar& _y) const
+  {
+    EIGEN_USING_STD_MATH(sqrt)
+    Scalar p, qp;
+    if(_x>_y)
+    {
+      p = _x;
+      qp = _y / p;
+    }
+    else
+    {
+      p = _y;
+      qp = _x / p;
+    }
+    return p * sqrt(Scalar(1) + qp*qp);
+  }
+};
+template<typename Scalar>
+struct functor_traits<scalar_hypot_op<Scalar,Scalar> > {
+  enum
+  {
+    Cost = 3 * NumTraits<Scalar>::AddCost +
+           2 * NumTraits<Scalar>::MulCost +
+           2 * scalar_div_cost<Scalar,false>::value,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the pow of two scalars
+  */
+template<typename Scalar, typename Exponent>
+struct scalar_pow_op  : binary_op_base<Scalar,Exponent>
+{
+  typedef typename ScalarBinaryOpTraits<Scalar,Exponent,scalar_pow_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_pow_op)
+#else
+  scalar_pow_op() {
+    typedef Scalar LhsScalar;
+    typedef Exponent RhsScalar;
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC
+  inline result_type operator() (const Scalar& a, const Exponent& b) const { return numext::pow(a, b); }
+};
+template<typename Scalar, typename Exponent>
+struct functor_traits<scalar_pow_op<Scalar,Exponent> > {
+  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = false };
+};
+
+
+
+//---------- non associative binary functors ----------
+
+/** \internal
+  * \brief Template functor to compute the difference of two scalars
+  *
+  * \sa class CwiseBinaryOp, MatrixBase::operator-
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_difference_op : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_difference_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_difference_op)
+#else
+  scalar_difference_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a - b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::psub(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_difference_op<LhsScalar,RhsScalar> > {
+  enum {
+    Cost = (NumTraits<LhsScalar>::AddCost+NumTraits<RhsScalar>::AddCost)/2,
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasSub && packet_traits<RhsScalar>::HasSub
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the quotient of two scalars
+  *
+  * \sa class CwiseBinaryOp, Cwise::operator/()
+  */
+template<typename LhsScalar,typename RhsScalar>
+struct scalar_quotient_op  : binary_op_base<LhsScalar,RhsScalar>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar,scalar_quotient_op>::ReturnType result_type;
+#ifndef EIGEN_SCALAR_BINARY_OP_PLUGIN
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_quotient_op)
+#else
+  scalar_quotient_op() {
+    EIGEN_SCALAR_BINARY_OP_PLUGIN
+  }
+#endif
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a / b; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a, const Packet& b) const
+  { return internal::pdiv(a,b); }
+};
+template<typename LhsScalar,typename RhsScalar>
+struct functor_traits<scalar_quotient_op<LhsScalar,RhsScalar> > {
+  typedef typename scalar_quotient_op<LhsScalar,RhsScalar>::result_type result_type;
+  enum {
+    PacketAccess = is_same<LhsScalar,RhsScalar>::value && packet_traits<LhsScalar>::HasDiv && packet_traits<RhsScalar>::HasDiv,
+    Cost = scalar_div_cost<result_type,PacketAccess>::value
+  };
+};
+
+
+
+/** \internal
+  * \brief Template functor to compute the and of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator&&
+  */
+struct scalar_boolean_and_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_and_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a && b; }
+};
+template<> struct functor_traits<scalar_boolean_and_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the or of two booleans
+  *
+  * \sa class CwiseBinaryOp, ArrayBase::operator||
+  */
+struct scalar_boolean_or_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_or_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a || b; }
+};
+template<> struct functor_traits<scalar_boolean_or_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+ * \brief Template functor to compute the xor of two booleans
+ *
+ * \sa class CwiseBinaryOp, ArrayBase::operator^
+ */
+struct scalar_boolean_xor_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_xor_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a, const bool& b) const { return a ^ b; }
+};
+template<> struct functor_traits<scalar_boolean_xor_op> {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+
+
+//---------- binary functors bound to a constant, thus appearing as a unary functor ----------
+
+// The following two classes permits to turn any binary functor into a unary one with one argument bound to a constant value.
+// They are analogues to std::binder1st/binder2nd but with the following differences:
+//  - they are compatible with packetOp
+//  - they are portable across C++ versions (the std::binder* are deprecated in C++11)
+template<typename BinaryOp> struct bind1st_op : BinaryOp {
+
+  typedef typename BinaryOp::first_argument_type  first_argument_type;
+  typedef typename BinaryOp::second_argument_type second_argument_type;
+  typedef typename BinaryOp::result_type          result_type;
+
+  bind1st_op(const first_argument_type &val) : m_value(val) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const second_argument_type& b) const { return BinaryOp::operator()(m_value,b); }
+
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& b) const
+  { return BinaryOp::packetOp(internal::pset1<Packet>(m_value), b); }
+
+  first_argument_type m_value;
+};
+template<typename BinaryOp> struct functor_traits<bind1st_op<BinaryOp> > : functor_traits<BinaryOp> {};
+
+
+template<typename BinaryOp> struct bind2nd_op : BinaryOp {
+
+  typedef typename BinaryOp::first_argument_type  first_argument_type;
+  typedef typename BinaryOp::second_argument_type second_argument_type;
+  typedef typename BinaryOp::result_type          result_type;
+
+  bind2nd_op(const second_argument_type &val) : m_value(val) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const first_argument_type& a) const { return BinaryOp::operator()(a,m_value); }
+
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return BinaryOp::packetOp(a,internal::pset1<Packet>(m_value)); }
+
+  second_argument_type m_value;
+};
+template<typename BinaryOp> struct functor_traits<bind2nd_op<BinaryOp> > : functor_traits<BinaryOp> {};
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BINARY_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h
new file mode 100644
index 0000000..b03be02
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/NullaryFunctors.h
@@ -0,0 +1,188 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_NULLARY_FUNCTORS_H
+#define EIGEN_NULLARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename Scalar>
+struct scalar_constant_op {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const scalar_constant_op& other) : m_other(other.m_other) { }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE scalar_constant_op(const Scalar& other) : m_other(other) { }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() () const { return m_other; }
+  template<typename PacketType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PacketType packetOp() const { return internal::pset1<PacketType>(m_other); }
+  const Scalar m_other;
+};
+template<typename Scalar>
+struct functor_traits<scalar_constant_op<Scalar> >
+{ enum { Cost = 0 /* as the constant value should be loaded in register only once for the whole expression */,
+         PacketAccess = packet_traits<Scalar>::Vectorizable, IsRepeatable = true }; };
+
+template<typename Scalar> struct scalar_identity_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_identity_op)
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType row, IndexType col) const { return row==col ? Scalar(1) : Scalar(0); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_identity_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::AddCost, PacketAccess = false, IsRepeatable = true }; };
+
+template <typename Scalar, typename Packet, bool IsInteger> struct linspaced_op_impl;
+
+template <typename Scalar, typename Packet>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/false>
+{
+  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
+    m_low(low), m_high(high), m_size1(num_steps==1 ? 1 : num_steps-1), m_step(num_steps==1 ? Scalar() : (high-low)/Scalar(num_steps-1)),
+    m_flip(numext::abs(high)<numext::abs(low))
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const {
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    if(m_flip)
+      return (i==0)? m_low : (m_high - RealScalar(m_size1-i)*m_step);
+    else
+      return (i==m_size1)? m_high : (m_low + RealScalar(i)*m_step);
+  }
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const
+  {
+    // Principle:
+    // [low, ..., low] + ( [step, ..., step] * ( [i, ..., i] + [0, ..., size] ) )
+    if(m_flip)
+    {
+      Packet pi = plset<Packet>(Scalar(i-m_size1));
+      Packet res = padd(pset1<Packet>(m_high), pmul(pset1<Packet>(m_step), pi));
+      if(i==0)
+        res = pinsertfirst(res, m_low);
+      return res;
+    }
+    else
+    {
+      Packet pi = plset<Packet>(Scalar(i));
+      Packet res = padd(pset1<Packet>(m_low), pmul(pset1<Packet>(m_step), pi));
+      if(i==m_size1-unpacket_traits<Packet>::size+1)
+        res = pinsertlast(res, m_high);
+      return res;
+    }
+  }
+
+  const Scalar m_low;
+  const Scalar m_high;
+  const Index m_size1;
+  const Scalar m_step;
+  const bool m_flip;
+};
+
+template <typename Scalar, typename Packet>
+struct linspaced_op_impl<Scalar,Packet,/*IsInteger*/true>
+{
+  linspaced_op_impl(const Scalar& low, const Scalar& high, Index num_steps) :
+    m_low(low),
+    m_multiplier((high-low)/convert_index<Scalar>(num_steps<=1 ? 1 : num_steps-1)),
+    m_divisor(convert_index<Scalar>((high>=low?num_steps:-num_steps)+(high-low))/((numext::abs(high-low)+1)==0?1:(numext::abs(high-low)+1))),
+    m_use_divisor(num_steps>1 && (numext::abs(high-low)+1)<num_steps)
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  const Scalar operator() (IndexType i) const
+  {
+    if(m_use_divisor) return m_low + convert_index<Scalar>(i)/m_divisor;
+    else              return m_low + convert_index<Scalar>(i)*m_multiplier;
+  }
+
+  const Scalar m_low;
+  const Scalar m_multiplier;
+  const Scalar m_divisor;
+  const bool m_use_divisor;
+};
+
+// ----- Linspace functor ----------------------------------------------------------------
+
+// Forward declaration (we default to random access which does not really give
+// us a speed gain when using packet access but it allows to use the functor in
+// nested expressions).
+template <typename Scalar, typename PacketType> struct linspaced_op;
+template <typename Scalar, typename PacketType> struct functor_traits< linspaced_op<Scalar,PacketType> >
+{
+  enum
+  {
+    Cost = 1,
+    PacketAccess =   (!NumTraits<Scalar>::IsInteger) && packet_traits<Scalar>::HasSetLinear && packet_traits<Scalar>::HasBlend,
+                  /*&& ((!NumTraits<Scalar>::IsInteger) || packet_traits<Scalar>::HasDiv),*/ // <- vectorization for integer is currently disabled
+    IsRepeatable = true
+  };
+};
+template <typename Scalar, typename PacketType> struct linspaced_op
+{
+  linspaced_op(const Scalar& low, const Scalar& high, Index num_steps)
+    : impl((num_steps==1 ? high : low),high,num_steps)
+  {}
+
+  template<typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (IndexType i) const { return impl(i); }
+
+  template<typename Packet,typename IndexType>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(IndexType i) const { return impl.packetOp(i); }
+
+  // This proxy object handles the actual required temporaries and the different
+  // implementations (integer vs. floating point).
+  const linspaced_op_impl<Scalar,PacketType,NumTraits<Scalar>::IsInteger> impl;
+};
+
+// Linear access is automatically determined from the operator() prototypes available for the given functor.
+// If it exposes an operator()(i,j), then we assume the i and j coefficients are required independently
+// and linear access is not possible. In all other cases, linear access is enabled.
+// Users should not have to deal with this structure.
+template<typename Functor> struct functor_has_linear_access { enum { ret = !has_binary_operator<Functor>::value }; };
+
+// For unreliable compilers, let's specialize the has_*ary_operator
+// helpers so that at least built-in nullary functors work fine.
+#if !( (EIGEN_COMP_MSVC>1600) || (EIGEN_GNUC_AT_LEAST(4,8)) || (EIGEN_COMP_ICC>=1600))
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_constant_op<Scalar>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_identity_op<Scalar>,IndexType> { enum { value = 1}; };
+
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_nullary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_unary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 1}; };
+template<typename Scalar, typename PacketType,typename IndexType>
+struct has_binary_operator<linspaced_op<Scalar,PacketType>,IndexType> { enum { value = 0}; };
+
+template<typename Scalar,typename IndexType>
+struct has_nullary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 1}; };
+template<typename Scalar,typename IndexType>
+struct has_unary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+template<typename Scalar,typename IndexType>
+struct has_binary_operator<scalar_random_op<Scalar>,IndexType> { enum { value = 0}; };
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_NULLARY_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h
new file mode 100644
index 0000000..6df3fa5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/StlFunctors.h
@@ -0,0 +1,132 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STL_FUNCTORS_H
+#define EIGEN_STL_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// default functor traits for STL functors:
+
+template<typename T>
+struct functor_traits<std::multiplies<T> >
+{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::divides<T> >
+{ enum { Cost = NumTraits<T>::MulCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::plus<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::minus<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::negate<T> >
+{ enum { Cost = NumTraits<T>::AddCost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_or<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_and<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::logical_not<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::greater<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::less<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::greater_equal<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::less_equal<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::equal_to<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::not_equal_to<T> >
+{ enum { Cost = 1, PacketAccess = false }; };
+
+#if (__cplusplus < 201103L) && (EIGEN_COMP_MSVC <= 1900)
+// std::binder* are deprecated since c++11 and will be removed in c++17
+template<typename T>
+struct functor_traits<std::binder2nd<T> >
+{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::binder1st<T> >
+{ enum { Cost = functor_traits<T>::Cost, PacketAccess = false }; };
+#endif
+
+template<typename T>
+struct functor_traits<std::unary_negate<T> >
+{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
+
+template<typename T>
+struct functor_traits<std::binary_negate<T> >
+{ enum { Cost = 1 + functor_traits<T>::Cost, PacketAccess = false }; };
+
+#ifdef EIGEN_STDEXT_SUPPORT
+
+template<typename T0,typename T1>
+struct functor_traits<std::project1st<T0,T1> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::project2nd<T0,T1> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::select2nd<std::pair<T0,T1> > >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::select1st<std::pair<T0,T1> > >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+template<typename T0,typename T1>
+struct functor_traits<std::unary_compose<T0,T1> >
+{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost, PacketAccess = false }; };
+
+template<typename T0,typename T1,typename T2>
+struct functor_traits<std::binary_compose<T0,T1,T2> >
+{ enum { Cost = functor_traits<T0>::Cost + functor_traits<T1>::Cost + functor_traits<T2>::Cost, PacketAccess = false }; };
+
+#endif // EIGEN_STDEXT_SUPPORT
+
+// allow to add new functors and specializations of functor_traits from outside Eigen.
+// this macro is really needed because functor_traits must be specialized after it is declared but before it is used...
+#ifdef EIGEN_FUNCTORS_PLUGIN
+#include EIGEN_FUNCTORS_PLUGIN
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_STL_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h
new file mode 100644
index 0000000..b254e96
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/TernaryFunctors.h
@@ -0,0 +1,25 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Eugene Brevdo <ebrevdo@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TERNARY_FUNCTORS_H
+#define EIGEN_TERNARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+//---------- associative ternary functors ----------
+
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TERNARY_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h
new file mode 100644
index 0000000..581a3c9
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/functors/UnaryFunctors.h
@@ -0,0 +1,823 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_UNARY_FUNCTORS_H
+#define EIGEN_UNARY_FUNCTORS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// Copied from unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+// TODO: remove or update after upstream
+/** \internal
+  * \brief Template functor to compute the sigmoid of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sigmoid()
+  */
+template <typename T>
+struct scalar_sigmoid_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sigmoid_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T operator()(const T& x) const {
+    const T one = T(1);
+    return one / (one + numext::exp(-x));
+  }
+
+  template <typename Packet> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+  Packet packetOp(const Packet& x) const {
+    const Packet one = pset1<Packet>(T(1));
+    return pdiv(one, padd(one, pexp(pnegate(x))));
+  }
+};
+
+template <typename T>
+struct functor_traits<scalar_sigmoid_op<T> > {
+  enum {
+    Cost = NumTraits<T>::AddCost * 2 + NumTraits<T>::MulCost * 6,
+    PacketAccess = packet_traits<T>::HasAdd && packet_traits<T>::HasDiv &&
+                   packet_traits<T>::HasNegate && packet_traits<T>::HasExp
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the opposite of a scalar
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::operator-
+  */
+template<typename Scalar> struct scalar_opposite_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_opposite_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return -a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pnegate(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_opposite_op<Scalar> >
+{ enum {
+    Cost = NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasNegate };
+};
+
+/** \internal
+  * \brief Template functor to compute the absolute value of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::abs
+  */
+template<typename Scalar> struct scalar_abs_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pabs(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_abs_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasAbs
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the score of a scalar, to chose a pivot
+  *
+  * \sa class CwiseUnaryOp
+  */
+template<typename Scalar> struct scalar_score_coeff_op : scalar_abs_op<Scalar>
+{
+  typedef void Score_is_abs;
+};
+template<typename Scalar>
+struct functor_traits<scalar_score_coeff_op<Scalar> > : functor_traits<scalar_abs_op<Scalar> > {};
+
+/* Avoid recomputing abs when we know the score and they are the same. Not a true Eigen functor.  */
+template<typename Scalar, typename=void> struct abs_knowing_score
+{
+  EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  template<typename Score>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a, const Score&) const { return numext::abs(a); }
+};
+template<typename Scalar> struct abs_knowing_score<Scalar, typename scalar_score_coeff_op<Scalar>::Score_is_abs>
+{
+  EIGEN_EMPTY_STRUCT_CTOR(abs_knowing_score)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  template<typename Scal>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scal&, const result_type& a) const { return a; }
+};
+
+/** \internal
+  * \brief Template functor to compute the squared absolute value of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::abs2
+  */
+template<typename Scalar> struct scalar_abs2_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_abs2_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { return numext::abs2(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_abs2_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasAbs2 }; };
+
+/** \internal
+  * \brief Template functor to compute the conjugate of a complex value
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::conjugate()
+  */
+template<typename Scalar> struct scalar_conjugate_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_conjugate_op)
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { using numext::conj; return conj(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const { return internal::pconj(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_conjugate_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::IsComplex ? NumTraits<Scalar>::AddCost : 0,
+    PacketAccess = packet_traits<Scalar>::HasConj
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the phase angle of a complex
+  *
+  * \sa class CwiseUnaryOp, Cwise::arg
+  */
+template<typename Scalar> struct scalar_arg_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_arg_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const result_type operator() (const Scalar& a) const { using numext::arg; return arg(a); }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Packet packetOp(const Packet& a) const
+  { return internal::parg(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_arg_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::IsComplex ? 5 * NumTraits<Scalar>::MulCost : NumTraits<Scalar>::AddCost,
+    PacketAccess = packet_traits<Scalar>::HasArg
+  };
+};
+/** \internal
+  * \brief Template functor to cast a scalar to another type
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::cast()
+  */
+template<typename Scalar, typename NewType>
+struct scalar_cast_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef NewType result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const NewType operator() (const Scalar& a) const { return cast<Scalar, NewType>(a); }
+};
+template<typename Scalar, typename NewType>
+struct functor_traits<scalar_cast_op<Scalar,NewType> >
+{ enum { Cost = is_same<Scalar, NewType>::value ? 0 : NumTraits<NewType>::AddCost, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the real part of a complex
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::real()
+  */
+template<typename Scalar>
+struct scalar_real_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::real(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_real_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the imaginary part of a complex
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::imag()
+  */
+template<typename Scalar>
+struct scalar_imag_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return numext::imag(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_imag_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the real part of a complex as a reference
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::real()
+  */
+template<typename Scalar>
+struct scalar_real_ref_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_real_ref_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::real_ref(*const_cast<Scalar*>(&a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_real_ref_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  * \brief Template functor to extract the imaginary part of a complex as a reference
+  *
+  * \sa class CwiseUnaryOp, MatrixBase::imag()
+  */
+template<typename Scalar>
+struct scalar_imag_ref_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_imag_ref_op)
+  typedef typename NumTraits<Scalar>::Real result_type;
+  EIGEN_DEVICE_FUNC
+  EIGEN_STRONG_INLINE result_type& operator() (const Scalar& a) const { return numext::imag_ref(*const_cast<Scalar*>(&a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_imag_ref_op<Scalar> >
+{ enum { Cost = 0, PacketAccess = false }; };
+
+/** \internal
+  *
+  * \brief Template functor to compute the exponential of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::exp()
+  */
+template<typename Scalar> struct scalar_exp_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_exp_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::exp(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pexp(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_exp_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasExp,
+    // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+    // Haswell can issue 2 add/mul/madd per cycle.
+    Cost =
+    (sizeof(Scalar) == 4
+     // float: 8 pmadd, 4 pmul, 2 padd/psub, 6 other
+     ? (8 * NumTraits<Scalar>::AddCost + 6 * NumTraits<Scalar>::MulCost)
+     // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
+     : (14 * NumTraits<Scalar>::AddCost +
+        6 * NumTraits<Scalar>::MulCost +
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
+#else
+    Cost =
+    (sizeof(Scalar) == 4
+     // float: 7 pmadd, 6 pmul, 4 padd/psub, 10 other
+     ? (21 * NumTraits<Scalar>::AddCost + 13 * NumTraits<Scalar>::MulCost)
+     // double: 7 pmadd, 5 pmul, 3 padd/psub, 1 div,  13 other
+     : (23 * NumTraits<Scalar>::AddCost +
+        12 * NumTraits<Scalar>::MulCost +
+        scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value))
+#endif
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the logarithm of a scalar
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::log()
+  */
+template<typename Scalar> struct scalar_log_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_log_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasLog,
+    Cost =
+    (PacketAccess
+     // The following numbers are based on the AVX implementation.
+#ifdef EIGEN_VECTORIZE_FMA
+     // 8 pmadd, 6 pmul, 8 padd/psub, 16 other, can issue 2 add/mul/madd per cycle.
+     ? (20 * NumTraits<Scalar>::AddCost + 7 * NumTraits<Scalar>::MulCost)
+#else
+     // 8 pmadd, 6 pmul, 8 padd/psub, 20 other
+     ? (36 * NumTraits<Scalar>::AddCost + 14 * NumTraits<Scalar>::MulCost)
+#endif
+     // Measured cost of std::log.
+     : sizeof(Scalar)==4 ? 40 : 85)
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the logarithm of 1 plus a scalar value
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::log1p()
+  */
+template<typename Scalar> struct scalar_log1p_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log1p_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::log1p(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog1p(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_log1p_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasLog1p,
+    Cost = functor_traits<scalar_log_op<Scalar> >::Cost // TODO measure cost of log1p
+  };
+};
+
+/** \internal
+  *
+  * \brief Template functor to compute the base-10 logarithm of a scalar
+  *
+  * \sa class CwiseUnaryOp, Cwise::log10()
+  */
+template<typename Scalar> struct scalar_log10_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_log10_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { EIGEN_USING_STD_MATH(log10) return log10(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::plog10(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_log10_op<Scalar> >
+{ enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasLog10 }; };
+
+/** \internal
+  * \brief Template functor to compute the square root of a scalar
+  * \sa class CwiseUnaryOp, Cwise::sqrt()
+  */
+template<typename Scalar> struct scalar_sqrt_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sqrt_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sqrt(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psqrt(a); }
+};
+template <typename Scalar>
+struct functor_traits<scalar_sqrt_op<Scalar> > {
+  enum {
+#if EIGEN_FAST_MATH
+    // The following numbers are based on the AVX implementation.
+    Cost = (sizeof(Scalar) == 8 ? 28
+                                // 4 pmul, 1 pmadd, 3 other
+                                : (3 * NumTraits<Scalar>::AddCost +
+                                   5 * NumTraits<Scalar>::MulCost)),
+#else
+    // The following numbers are based on min VSQRT throughput on Haswell.
+    Cost = (sizeof(Scalar) == 8 ? 28 : 14),
+#endif
+    PacketAccess = packet_traits<Scalar>::HasSqrt
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the reciprocal square root of a scalar
+  * \sa class CwiseUnaryOp, Cwise::rsqrt()
+  */
+template<typename Scalar> struct scalar_rsqrt_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_rsqrt_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return Scalar(1)/numext::sqrt(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::prsqrt(a); }
+};
+
+template<typename Scalar>
+struct functor_traits<scalar_rsqrt_op<Scalar> >
+{ enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasRsqrt
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the cosine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::cos()
+  */
+template<typename Scalar> struct scalar_cos_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cos_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return numext::cos(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcos(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cos_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCos
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the sine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sin()
+  */
+template<typename Scalar> struct scalar_sin_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sin_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sin(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psin(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sin_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasSin
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the tan of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::tan()
+  */
+template<typename Scalar> struct scalar_tan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_tan_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::tan(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::ptan(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_tan_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasTan
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the arc cosine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::acos()
+  */
+template<typename Scalar> struct scalar_acos_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_acos_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::acos(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pacos(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_acos_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasACos
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the arc sine of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::asin()
+  */
+template<typename Scalar> struct scalar_asin_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_asin_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::asin(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pasin(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_asin_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasASin
+  };
+};
+
+
+/** \internal
+  * \brief Template functor to compute the atan of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::atan()
+  */
+template<typename Scalar> struct scalar_atan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_atan_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::atan(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::patan(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_atan_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasATan
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the tanh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::tanh()
+  */
+template <typename Scalar>
+struct scalar_tanh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_tanh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator()(const Scalar& a) const { return numext::tanh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& x) const { return ptanh(x); }
+};
+
+template <typename Scalar>
+struct functor_traits<scalar_tanh_op<Scalar> > {
+  enum {
+    PacketAccess = packet_traits<Scalar>::HasTanh,
+    Cost = ( (EIGEN_FAST_MATH && is_same<Scalar,float>::value)
+// The following numbers are based on the AVX implementation,
+#ifdef EIGEN_VECTORIZE_FMA
+                // Haswell can issue 2 add/mul/madd per cycle.
+                // 9 pmadd, 2 pmul, 1 div, 2 other
+                ? (2 * NumTraits<Scalar>::AddCost +
+                   6 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
+#else
+                ? (11 * NumTraits<Scalar>::AddCost +
+                   11 * NumTraits<Scalar>::MulCost +
+                   scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value)
+#endif
+                // This number assumes a naive implementation of tanh
+                : (6 * NumTraits<Scalar>::AddCost +
+                   3 * NumTraits<Scalar>::MulCost +
+                   2 * scalar_div_cost<Scalar,packet_traits<Scalar>::HasDiv>::value +
+                   functor_traits<scalar_exp_op<Scalar> >::Cost))
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the sinh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::sinh()
+  */
+template<typename Scalar> struct scalar_sinh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sinh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::sinh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psinh(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sinh_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasSinh
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the cosh of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::cosh()
+  */
+template<typename Scalar> struct scalar_cosh_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cosh_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const { return numext::cosh(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pcosh(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cosh_op<Scalar> >
+{
+  enum {
+    Cost = 5 * NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCosh
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the inverse of a scalar
+  * \sa class CwiseUnaryOp, Cwise::inverse()
+  */
+template<typename Scalar>
+struct scalar_inverse_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_inverse_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return Scalar(1)/a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pdiv(pset1<Packet>(Scalar(1)),a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_inverse_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasDiv }; };
+
+/** \internal
+  * \brief Template functor to compute the square of a scalar
+  * \sa class CwiseUnaryOp, Cwise::square()
+  */
+template<typename Scalar>
+struct scalar_square_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_square_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_square_op<Scalar> >
+{ enum { Cost = NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
+
+/** \internal
+  * \brief Template functor to compute the cube of a scalar
+  * \sa class CwiseUnaryOp, Cwise::cube()
+  */
+template<typename Scalar>
+struct scalar_cube_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cube_op)
+  EIGEN_DEVICE_FUNC inline Scalar operator() (const Scalar& a) const { return a*a*a; }
+  template<typename Packet>
+  EIGEN_DEVICE_FUNC inline const Packet packetOp(const Packet& a) const
+  { return internal::pmul(a,pmul(a,a)); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_cube_op<Scalar> >
+{ enum { Cost = 2*NumTraits<Scalar>::MulCost, PacketAccess = packet_traits<Scalar>::HasMul }; };
+
+/** \internal
+  * \brief Template functor to compute the rounded value of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::round()
+  */
+template<typename Scalar> struct scalar_round_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_round_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::round(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pround(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_round_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasRound
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the floor of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::floor()
+  */
+template<typename Scalar> struct scalar_floor_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_floor_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::floor(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pfloor(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_floor_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasFloor
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the ceil of a scalar
+  * \sa class CwiseUnaryOp, ArrayBase::ceil()
+  */
+template<typename Scalar> struct scalar_ceil_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_ceil_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Scalar operator() (const Scalar& a) const { return numext::ceil(a); }
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::pceil(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_ceil_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = packet_traits<Scalar>::HasCeil
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute whether a scalar is NaN
+  * \sa class CwiseUnaryOp, ArrayBase::isnan()
+  */
+template<typename Scalar> struct scalar_isnan_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isnan_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isnan)(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isnan_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to check whether a scalar is +/-inf
+  * \sa class CwiseUnaryOp, ArrayBase::isinf()
+  */
+template<typename Scalar> struct scalar_isinf_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isinf_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isinf)(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isinf_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to check whether a scalar has a finite value
+  * \sa class CwiseUnaryOp, ArrayBase::isfinite()
+  */
+template<typename Scalar> struct scalar_isfinite_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_isfinite_op)
+  typedef bool result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const Scalar& a) const { return (numext::isfinite)(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_isfinite_op<Scalar> >
+{
+  enum {
+    Cost = NumTraits<Scalar>::MulCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the logical not of a boolean
+  *
+  * \sa class CwiseUnaryOp, ArrayBase::operator!
+  */
+template<typename Scalar> struct scalar_boolean_not_op {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_boolean_not_op)
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool operator() (const bool& a) const { return !a; }
+};
+template<typename Scalar>
+struct functor_traits<scalar_boolean_not_op<Scalar> > {
+  enum {
+    Cost = NumTraits<bool>::AddCost,
+    PacketAccess = false
+  };
+};
+
+/** \internal
+  * \brief Template functor to compute the signum of a scalar
+  * \sa class CwiseUnaryOp, Cwise::sign()
+  */
+template<typename Scalar,bool iscpx=(NumTraits<Scalar>::IsComplex!=0) > struct scalar_sign_op;
+template<typename Scalar>
+struct scalar_sign_op<Scalar,false> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const
+  {
+      return Scalar( (a>Scalar(0)) - (a<Scalar(0)) );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar>
+struct scalar_sign_op<Scalar,true> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_sign_op)
+  EIGEN_DEVICE_FUNC inline const Scalar operator() (const Scalar& a) const
+  {
+    typedef typename NumTraits<Scalar>::Real real_type;
+    real_type aa = numext::abs(a);
+    if (aa==real_type(0))
+      return Scalar(0);
+    aa = real_type(1)/aa;
+    return Scalar(real(a)*aa, imag(a)*aa );
+  }
+  //TODO
+  //template <typename Packet>
+  //EIGEN_DEVICE_FUNC inline Packet packetOp(const Packet& a) const { return internal::psign(a); }
+};
+template<typename Scalar>
+struct functor_traits<scalar_sign_op<Scalar> >
+{ enum {
+    Cost = 
+        NumTraits<Scalar>::IsComplex
+        ? ( 8*NumTraits<Scalar>::MulCost  ) // roughly
+        : ( 3*NumTraits<Scalar>::AddCost),
+    PacketAccess = packet_traits<Scalar>::HasSign
+  };
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_FUNCTORS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h
new file mode 100644
index 0000000..45230bc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralBlockPanelKernel.h
@@ -0,0 +1,2149 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_BLOCK_PANEL_H
+#define EIGEN_GENERAL_BLOCK_PANEL_H
+
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs=false, bool _ConjRhs=false>
+class gebp_traits;
+
+
+/** \internal \returns b if a<=0, and returns a otherwise. */
+inline std::ptrdiff_t manage_caching_sizes_helper(std::ptrdiff_t a, std::ptrdiff_t b)
+{
+  return a<=0 ? b : a;
+}
+
+#if EIGEN_ARCH_i386_OR_x86_64
+const std::ptrdiff_t defaultL1CacheSize = 32*1024;
+const std::ptrdiff_t defaultL2CacheSize = 256*1024;
+const std::ptrdiff_t defaultL3CacheSize = 2*1024*1024;
+#else
+const std::ptrdiff_t defaultL1CacheSize = 16*1024;
+const std::ptrdiff_t defaultL2CacheSize = 512*1024;
+const std::ptrdiff_t defaultL3CacheSize = 512*1024;
+#endif
+
+/** \internal */
+struct CacheSizes {
+  CacheSizes(): m_l1(-1),m_l2(-1),m_l3(-1) {
+    int l1CacheSize, l2CacheSize, l3CacheSize;
+    queryCacheSizes(l1CacheSize, l2CacheSize, l3CacheSize);
+    m_l1 = manage_caching_sizes_helper(l1CacheSize, defaultL1CacheSize);
+    m_l2 = manage_caching_sizes_helper(l2CacheSize, defaultL2CacheSize);
+    m_l3 = manage_caching_sizes_helper(l3CacheSize, defaultL3CacheSize);
+  }
+
+  std::ptrdiff_t m_l1;
+  std::ptrdiff_t m_l2;
+  std::ptrdiff_t m_l3;
+};
+
+
+/** \internal */
+inline void manage_caching_sizes(Action action, std::ptrdiff_t* l1, std::ptrdiff_t* l2, std::ptrdiff_t* l3)
+{
+  static CacheSizes m_cacheSizes;
+
+  if(action==SetAction)
+  {
+    // set the cpu cache size and cache all block sizes from a global cache size in byte
+    eigen_internal_assert(l1!=0 && l2!=0);
+    m_cacheSizes.m_l1 = *l1;
+    m_cacheSizes.m_l2 = *l2;
+    m_cacheSizes.m_l3 = *l3;
+  }
+  else if(action==GetAction)
+  {
+    eigen_internal_assert(l1!=0 && l2!=0);
+    *l1 = m_cacheSizes.m_l1;
+    *l2 = m_cacheSizes.m_l2;
+    *l3 = m_cacheSizes.m_l3;
+  }
+  else
+  {
+    eigen_internal_assert(false);
+  }
+}
+
+/* Helper for computeProductBlockingSizes.
+ *
+ * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
+ * this function computes the blocking size parameters along the respective dimensions
+ * for matrix products and related algorithms. The blocking sizes depends on various
+ * parameters:
+ * - the L1 and L2 cache sizes,
+ * - the register level blocking sizes defined by gebp_traits,
+ * - the number of scalars that fit into a packet (when vectorization is enabled).
+ *
+ * \sa setCpuCacheSizes */
+
+template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index>
+void evaluateProductBlockingSizesHeuristic(Index& k, Index& m, Index& n, Index num_threads = 1)
+{
+  typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+  // Explanations:
+  // Let's recall that the product algorithms form mc x kc vertical panels A' on the lhs and
+  // kc x nc blocks B' on the rhs. B' has to fit into L2/L3 cache. Moreover, A' is processed
+  // per mr x kc horizontal small panels where mr is the blocking size along the m dimension
+  // at the register level. This small horizontal panel has to stay within L1 cache.
+  std::ptrdiff_t l1, l2, l3;
+  manage_caching_sizes(GetAction, &l1, &l2, &l3);
+
+  if (num_threads > 1) {
+    typedef typename Traits::ResScalar ResScalar;
+    enum {
+      kdiv = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)),
+      ksub = Traits::mr * Traits::nr * sizeof(ResScalar),
+      kr = 8,
+      mr = Traits::mr,
+      nr = Traits::nr
+    };
+    // Increasing k gives us more time to prefetch the content of the "C"
+    // registers. However once the latency is hidden there is no point in
+    // increasing the value of k, so we'll cap it at 320 (value determined
+    // experimentally).
+    const Index k_cache = (numext::mini<Index>)((l1-ksub)/kdiv, 320);
+    if (k_cache < k) {
+      k = k_cache - (k_cache % kr);
+      eigen_internal_assert(k > 0);
+    }
+
+    const Index n_cache = (l2-l1) / (nr * sizeof(RhsScalar) * k);
+    const Index n_per_thread = numext::div_ceil(n, num_threads);
+    if (n_cache <= n_per_thread) {
+      // Don't exceed the capacity of the l2 cache.
+      eigen_internal_assert(n_cache >= static_cast<Index>(nr));
+      n = n_cache - (n_cache % nr);
+      eigen_internal_assert(n > 0);
+    } else {
+      n = (numext::mini<Index>)(n, (n_per_thread + nr - 1) - ((n_per_thread + nr - 1) % nr));
+    }
+
+    if (l3 > l2) {
+      // l3 is shared between all cores, so we'll give each thread its own chunk of l3.
+      const Index m_cache = (l3-l2) / (sizeof(LhsScalar) * k * num_threads);
+      const Index m_per_thread = numext::div_ceil(m, num_threads);
+      if(m_cache < m_per_thread && m_cache >= static_cast<Index>(mr)) {
+        m = m_cache - (m_cache % mr);
+        eigen_internal_assert(m > 0);
+      } else {
+        m = (numext::mini<Index>)(m, (m_per_thread + mr - 1) - ((m_per_thread + mr - 1) % mr));
+      }
+    }
+  }
+  else {
+    // In unit tests we do not want to use extra large matrices,
+    // so we reduce the cache size to check the blocking strategy is not flawed
+#ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
+    l1 = 9*1024;
+    l2 = 32*1024;
+    l3 = 512*1024;
+#endif
+
+    // Early return for small problems because the computation below are time consuming for small problems.
+    // Perhaps it would make more sense to consider k*n*m??
+    // Note that for very tiny problem, this function should be bypassed anyway
+    // because we use the coefficient-based implementation for them.
+    if((numext::maxi)(k,(numext::maxi)(m,n))<48)
+      return;
+
+    typedef typename Traits::ResScalar ResScalar;
+    enum {
+      k_peeling = 8,
+      k_div = KcFactor * (Traits::mr * sizeof(LhsScalar) + Traits::nr * sizeof(RhsScalar)),
+      k_sub = Traits::mr * Traits::nr * sizeof(ResScalar)
+    };
+
+    // ---- 1st level of blocking on L1, yields kc ----
+
+    // Blocking on the third dimension (i.e., k) is chosen so that an horizontal panel
+    // of size mr x kc of the lhs plus a vertical panel of kc x nr of the rhs both fits within L1 cache.
+    // We also include a register-level block of the result (mx x nr).
+    // (In an ideal world only the lhs panel would stay in L1)
+    // Moreover, kc has to be a multiple of 8 to be compatible with loop peeling, leading to a maximum blocking size of:
+    const Index max_kc = numext::maxi<Index>(((l1-k_sub)/k_div) & (~(k_peeling-1)),1);
+    const Index old_k = k;
+    if(k>max_kc)
+    {
+      // We are really blocking on the third dimension:
+      // -> reduce blocking size to make sure the last block is as large as possible
+      //    while keeping the same number of sweeps over the result.
+      k = (k%max_kc)==0 ? max_kc
+                        : max_kc - k_peeling * ((max_kc-1-(k%max_kc))/(k_peeling*(k/max_kc+1)));
+
+      eigen_internal_assert(((old_k/k) == (old_k/max_kc)) && "the number of sweeps has to remain the same");
+    }
+
+    // ---- 2nd level of blocking on max(L2,L3), yields nc ----
+
+    // TODO find a reliable way to get the actual amount of cache per core to use for 2nd level blocking, that is:
+    //      actual_l2 = max(l2, l3/nb_core_sharing_l3)
+    // The number below is quite conservative: it is better to underestimate the cache size rather than overestimating it)
+    // For instance, it corresponds to 6MB of L3 shared among 4 cores.
+    #ifdef EIGEN_DEBUG_SMALL_PRODUCT_BLOCKS
+    const Index actual_l2 = l3;
+    #else
+    const Index actual_l2 = 1572864; // == 1.5 MB
+    #endif
+
+    // Here, nc is chosen such that a block of kc x nc of the rhs fit within half of L2.
+    // The second half is implicitly reserved to access the result and lhs coefficients.
+    // When k<max_kc, then nc can arbitrarily growth. In practice, it seems to be fruitful
+    // to limit this growth: we bound nc to growth by a factor x1.5.
+    // However, if the entire lhs block fit within L1, then we are not going to block on the rows at all,
+    // and it becomes fruitful to keep the packed rhs blocks in L1 if there is enough remaining space.
+    Index max_nc;
+    const Index lhs_bytes = m * k * sizeof(LhsScalar);
+    const Index remaining_l1 = l1- k_sub - lhs_bytes;
+    if(remaining_l1 >= Index(Traits::nr*sizeof(RhsScalar))*k)
+    {
+      // L1 blocking
+      max_nc = remaining_l1 / (k*sizeof(RhsScalar));
+    }
+    else
+    {
+      // L2 blocking
+      max_nc = (3*actual_l2)/(2*2*max_kc*sizeof(RhsScalar));
+    }
+    // WARNING Below, we assume that Traits::nr is a power of two.
+    Index nc = numext::mini<Index>(actual_l2/(2*k*sizeof(RhsScalar)), max_nc) & (~(Traits::nr-1));
+    if(n>nc)
+    {
+      // We are really blocking over the columns:
+      // -> reduce blocking size to make sure the last block is as large as possible
+      //    while keeping the same number of sweeps over the packed lhs.
+      //    Here we allow one more sweep if this gives us a perfect match, thus the commented "-1"
+      n = (n%nc)==0 ? nc
+                    : (nc - Traits::nr * ((nc/*-1*/-(n%nc))/(Traits::nr*(n/nc+1))));
+    }
+    else if(old_k==k)
+    {
+      // So far, no blocking at all, i.e., kc==k, and nc==n.
+      // In this case, let's perform a blocking over the rows such that the packed lhs data is kept in cache L1/L2
+      // TODO: part of this blocking strategy is now implemented within the kernel itself, so the L1-based heuristic here should be obsolete.
+      Index problem_size = k*n*sizeof(LhsScalar);
+      Index actual_lm = actual_l2;
+      Index max_mc = m;
+      if(problem_size<=1024)
+      {
+        // problem is small enough to keep in L1
+        // Let's choose m such that lhs's block fit in 1/3 of L1
+        actual_lm = l1;
+      }
+      else if(l3!=0 && problem_size<=32768)
+      {
+        // we have both L2 and L3, and problem is small enough to be kept in L2
+        // Let's choose m such that lhs's block fit in 1/3 of L2
+        actual_lm = l2;
+        max_mc = (numext::mini<Index>)(576,max_mc);
+      }
+      Index mc = (numext::mini<Index>)(actual_lm/(3*k*sizeof(LhsScalar)), max_mc);
+      if (mc > Traits::mr) mc -= mc % Traits::mr;
+      else if (mc==0) return;
+      m = (m%mc)==0 ? mc
+                    : (mc - Traits::mr * ((mc/*-1*/-(m%mc))/(Traits::mr*(m/mc+1))));
+    }
+  }
+}
+
+template <typename Index>
+inline bool useSpecificBlockingSizes(Index& k, Index& m, Index& n)
+{
+#ifdef EIGEN_TEST_SPECIFIC_BLOCKING_SIZES
+  if (EIGEN_TEST_SPECIFIC_BLOCKING_SIZES) {
+    k = numext::mini<Index>(k, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_K);
+    m = numext::mini<Index>(m, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_M);
+    n = numext::mini<Index>(n, EIGEN_TEST_SPECIFIC_BLOCKING_SIZE_N);
+    return true;
+  }
+#else
+  EIGEN_UNUSED_VARIABLE(k)
+  EIGEN_UNUSED_VARIABLE(m)
+  EIGEN_UNUSED_VARIABLE(n)
+#endif
+  return false;
+}
+
+/** \brief Computes the blocking parameters for a m x k times k x n matrix product
+  *
+  * \param[in,out] k Input: the third dimension of the product. Output: the blocking size along the same dimension.
+  * \param[in,out] m Input: the number of rows of the left hand side. Output: the blocking size along the same dimension.
+  * \param[in,out] n Input: the number of columns of the right hand side. Output: the blocking size along the same dimension.
+  *
+  * Given a m x k times k x n matrix product of scalar types \c LhsScalar and \c RhsScalar,
+  * this function computes the blocking size parameters along the respective dimensions
+  * for matrix products and related algorithms.
+  *
+  * The blocking size parameters may be evaluated:
+  *   - either by a heuristic based on cache sizes;
+  *   - or using fixed prescribed values (for testing purposes).
+  *
+  * \sa setCpuCacheSizes */
+
+template<typename LhsScalar, typename RhsScalar, int KcFactor, typename Index>
+void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
+{
+  if (!useSpecificBlockingSizes(k, m, n)) {
+    evaluateProductBlockingSizesHeuristic<LhsScalar, RhsScalar, KcFactor, Index>(k, m, n, num_threads);
+  }
+}
+
+template<typename LhsScalar, typename RhsScalar, typename Index>
+inline void computeProductBlockingSizes(Index& k, Index& m, Index& n, Index num_threads = 1)
+{
+  computeProductBlockingSizes<LhsScalar,RhsScalar,1,Index>(k, m, n, num_threads);
+}
+
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_CJMADD
+  #define CJMADD(CJ,A,B,C,T)  C = CJ.pmadd(A,B,C);
+#else
+
+  // FIXME (a bit overkill maybe ?)
+
+  template<typename CJ, typename A, typename B, typename C, typename T> struct gebp_madd_selector {
+    EIGEN_ALWAYS_INLINE static void run(const CJ& cj, A& a, B& b, C& c, T& /*t*/)
+    {
+      c = cj.pmadd(a,b,c);
+    }
+  };
+
+  template<typename CJ, typename T> struct gebp_madd_selector<CJ,T,T,T,T> {
+    EIGEN_ALWAYS_INLINE static void run(const CJ& cj, T& a, T& b, T& c, T& t)
+    {
+      t = b; t = cj.pmul(a,t); c = padd(c,t);
+    }
+  };
+
+  template<typename CJ, typename A, typename B, typename C, typename T>
+  EIGEN_STRONG_INLINE void gebp_madd(const CJ& cj, A& a, B& b, C& c, T& t)
+  {
+    gebp_madd_selector<CJ,A,B,C,T>::run(cj,a,b,c,t);
+  }
+
+  #define CJMADD(CJ,A,B,C,T)  gebp_madd(CJ,A,B,C,T);
+//   #define CJMADD(CJ,A,B,C,T)  T = B; T = CJ.pmul(A,T); C = padd(C,T);
+#endif
+
+/* Vectorization logic
+ *  real*real: unpack rhs to constant packets, ...
+ * 
+ *  cd*cd : unpack rhs to (b_r,b_r), (b_i,b_i), mul to get (a_r b_r,a_i b_r) (a_r b_i,a_i b_i),
+ *          storing each res packet into two packets (2x2),
+ *          at the end combine them: swap the second and addsub them 
+ *  cf*cf : same but with 2x4 blocks
+ *  cplx*real : unpack rhs to constant packets, ...
+ *  real*cplx : load lhs as (a0,a0,a1,a1), and mul as usual
+ */
+template<typename _LhsScalar, typename _RhsScalar, bool _ConjLhs, bool _ConjRhs>
+class gebp_traits
+{
+public:
+  typedef _LhsScalar LhsScalar;
+  typedef _RhsScalar RhsScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+
+  enum {
+    ConjLhs = _ConjLhs,
+    ConjRhs = _ConjRhs,
+    Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable,
+    LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+    RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+    ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1,
+    
+    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+
+    // register block size along the N direction must be 1 or 4
+    nr = 4,
+
+    // register block size along the M direction (currently, this one cannot be modified)
+    default_mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
+#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
+    // we assume 16 registers
+    // See bug 992, if the scalar type is not vectorizable but that EIGEN_HAS_SINGLE_INSTRUCTION_MADD is defined,
+    // then using 3*LhsPacketSize triggers non-implemented paths in syrk.
+    mr = Vectorizable ? 3*LhsPacketSize : default_mr,
+#else
+    mr = default_mr,
+#endif
+    
+    LhsProgress = LhsPacketSize,
+    RhsProgress = 1
+  };
+
+  typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
+  typedef typename packet_traits<RhsScalar>::type  _RhsPacket;
+  typedef typename packet_traits<ResScalar>::type  _ResPacket;
+
+  typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
+  typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
+  typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+
+  typedef ResPacket AccPacket;
+  
+  EIGEN_STRONG_INLINE void initAcc(AccPacket& p)
+  {
+    p = pset1<ResPacket>(ResScalar(0));
+  }
+  
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    pbroadcast4(b, b0, b1, b2, b3);
+  }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     pbroadcast2(b, b0, b1);
+//   }
+  
+  template<typename RhsPacketType>
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacketType& dest) const
+  {
+    dest = pset1<RhsPacketType>(*b);
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = ploadquad<RhsPacket>(b);
+  }
+
+  template<typename LhsPacketType>
+  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacketType& dest) const
+  {
+    dest = pload<LhsPacketType>(a);
+  }
+
+  template<typename LhsPacketType>
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacketType& dest) const
+  {
+    dest = ploadu<LhsPacketType>(a);
+  }
+
+  template<typename LhsPacketType, typename RhsPacketType, typename AccPacketType>
+  EIGEN_STRONG_INLINE void madd(const LhsPacketType& a, const RhsPacketType& b, AccPacketType& c, AccPacketType& tmp) const
+  {
+    conj_helper<LhsPacketType,RhsPacketType,ConjLhs,ConjRhs> cj;
+    // It would be a lot cleaner to call pmadd all the time. Unfortunately if we
+    // let gcc allocate the register in which to store the result of the pmul
+    // (in the case where there is no FMA) gcc fails to figure out how to avoid
+    // spilling register.
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c = cj.pmadd(a,b,c);
+#else
+    tmp = b; tmp = cj.pmul(a,tmp); c = padd(c,tmp);
+#endif
+  }
+
+  EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const
+  {
+    r = pmadd(c,alpha,r);
+  }
+  
+  template<typename ResPacketHalf>
+  EIGEN_STRONG_INLINE void acc(const ResPacketHalf& c, const ResPacketHalf& alpha, ResPacketHalf& r) const
+  {
+    r = pmadd(c,alpha,r);
+  }
+
+};
+
+template<typename RealScalar, bool _ConjLhs>
+class gebp_traits<std::complex<RealScalar>, RealScalar, _ConjLhs, false>
+{
+public:
+  typedef std::complex<RealScalar> LhsScalar;
+  typedef RealScalar RhsScalar;
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+
+  enum {
+    ConjLhs = _ConjLhs,
+    ConjRhs = false,
+    Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable,
+    LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+    RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+    ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1,
+    
+    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+    nr = 4,
+#if defined(EIGEN_HAS_SINGLE_INSTRUCTION_MADD) && !defined(EIGEN_VECTORIZE_ALTIVEC) && !defined(EIGEN_VECTORIZE_VSX)
+    // we assume 16 registers
+    mr = 3*LhsPacketSize,
+#else
+    mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*LhsPacketSize,
+#endif
+
+    LhsProgress = LhsPacketSize,
+    RhsProgress = 1
+  };
+
+  typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
+  typedef typename packet_traits<RhsScalar>::type  _RhsPacket;
+  typedef typename packet_traits<ResScalar>::type  _ResPacket;
+
+  typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
+  typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
+  typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+
+  typedef ResPacket AccPacket;
+
+  EIGEN_STRONG_INLINE void initAcc(AccPacket& p)
+  {
+    p = pset1<ResPacket>(ResScalar(0));
+  }
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = pset1<RhsPacket>(*b);
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = pset1<RhsPacket>(*b);
+  }
+
+  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = pload<LhsPacket>(a);
+  }
+
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploadu<LhsPacket>(a);
+  }
+
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    pbroadcast4(b, b0, b1, b2, b3);
+  }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     pbroadcast2(b, b0, b1);
+//   }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const
+  {
+    madd_impl(a, b, c, tmp, typename conditional<Vectorizable,true_type,false_type>::type());
+  }
+
+  EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const
+  {
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c.v = pmadd(a.v,b,c.v);
+#else
+    tmp = b; tmp = pmul(a.v,tmp); c.v = padd(c.v,tmp);
+#endif
+  }
+
+  EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const
+  {
+    c += a * b;
+  }
+
+  EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const
+  {
+    r = cj.pmadd(c,alpha,r);
+  }
+
+protected:
+  conj_helper<ResPacket,ResPacket,ConjLhs,false> cj;
+};
+
+template<typename Packet>
+struct DoublePacket
+{
+  Packet first;
+  Packet second;
+};
+
+template<typename Packet>
+DoublePacket<Packet> padd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
+{
+  DoublePacket<Packet> res;
+  res.first  = padd(a.first, b.first);
+  res.second = padd(a.second,b.second);
+  return res;
+}
+
+template<typename Packet>
+const DoublePacket<Packet>& predux_downto4(const DoublePacket<Packet> &a)
+{
+  return a;
+}
+
+template<typename Packet> struct unpacket_traits<DoublePacket<Packet> > { typedef DoublePacket<Packet> half; };
+// template<typename Packet>
+// DoublePacket<Packet> pmadd(const DoublePacket<Packet> &a, const DoublePacket<Packet> &b)
+// {
+//   DoublePacket<Packet> res;
+//   res.first  = padd(a.first, b.first);
+//   res.second = padd(a.second,b.second);
+//   return res;
+// }
+
+template<typename RealScalar, bool _ConjLhs, bool _ConjRhs>
+class gebp_traits<std::complex<RealScalar>, std::complex<RealScalar>, _ConjLhs, _ConjRhs >
+{
+public:
+  typedef std::complex<RealScalar>  Scalar;
+  typedef std::complex<RealScalar>  LhsScalar;
+  typedef std::complex<RealScalar>  RhsScalar;
+  typedef std::complex<RealScalar>  ResScalar;
+  
+  enum {
+    ConjLhs = _ConjLhs,
+    ConjRhs = _ConjRhs,
+    Vectorizable = packet_traits<RealScalar>::Vectorizable
+                && packet_traits<Scalar>::Vectorizable,
+    RealPacketSize  = Vectorizable ? packet_traits<RealScalar>::size : 1,
+    ResPacketSize   = Vectorizable ? packet_traits<ResScalar>::size : 1,
+    LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+    RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+
+    // FIXME: should depend on NumberOfRegisters
+    nr = 4,
+    mr = ResPacketSize,
+
+    LhsProgress = ResPacketSize,
+    RhsProgress = 1
+  };
+  
+  typedef typename packet_traits<RealScalar>::type RealPacket;
+  typedef typename packet_traits<Scalar>::type     ScalarPacket;
+  typedef DoublePacket<RealPacket> DoublePacketType;
+
+  typedef typename conditional<Vectorizable,RealPacket,  Scalar>::type LhsPacket;
+  typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type RhsPacket;
+  typedef typename conditional<Vectorizable,ScalarPacket,Scalar>::type ResPacket;
+  typedef typename conditional<Vectorizable,DoublePacketType,Scalar>::type AccPacket;
+  
+  EIGEN_STRONG_INLINE void initAcc(Scalar& p) { p = Scalar(0); }
+
+  EIGEN_STRONG_INLINE void initAcc(DoublePacketType& p)
+  {
+    p.first   = pset1<RealPacket>(RealScalar(0));
+    p.second  = pset1<RealPacket>(RealScalar(0));
+  }
+
+  // Scalar path
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, ResPacket& dest) const
+  {
+    dest = pset1<ResPacket>(*b);
+  }
+
+  // Vectorized path
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, DoublePacketType& dest) const
+  {
+    dest.first  = pset1<RealPacket>(real(*b));
+    dest.second = pset1<RealPacket>(imag(*b));
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, ResPacket& dest) const
+  {
+    loadRhs(b,dest);
+  }
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, DoublePacketType& dest) const
+  {
+    eigen_internal_assert(unpacket_traits<ScalarPacket>::size<=4);
+    loadRhs(b,dest);
+  }
+  
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
+    loadRhs(b+2, b2);
+    loadRhs(b+3, b3);
+  }
+  
+  // Vectorized path
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, DoublePacketType& b0, DoublePacketType& b1)
+  {
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
+  }
+  
+  // Scalar path
+  EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsScalar& b0, RhsScalar& b1)
+  {
+    // FIXME not sure that's the best way to implement it!
+    loadRhs(b+0, b0);
+    loadRhs(b+1, b1);
+  }
+
+  // nothing special here
+  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = pload<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a));
+  }
+
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploadu<LhsPacket>((const typename unpacket_traits<LhsPacket>::type*)(a));
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, DoublePacketType& c, RhsPacket& /*tmp*/) const
+  {
+    c.first   = padd(pmul(a,b.first), c.first);
+    c.second  = padd(pmul(a,b.second),c.second);
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, ResPacket& c, RhsPacket& /*tmp*/) const
+  {
+    c = cj.pmadd(a,b,c);
+  }
+  
+  EIGEN_STRONG_INLINE void acc(const Scalar& c, const Scalar& alpha, Scalar& r) const { r += alpha * c; }
+  
+  EIGEN_STRONG_INLINE void acc(const DoublePacketType& c, const ResPacket& alpha, ResPacket& r) const
+  {
+    // assemble c
+    ResPacket tmp;
+    if((!ConjLhs)&&(!ConjRhs))
+    {
+      tmp = pcplxflip(pconj(ResPacket(c.second)));
+      tmp = padd(ResPacket(c.first),tmp);
+    }
+    else if((!ConjLhs)&&(ConjRhs))
+    {
+      tmp = pconj(pcplxflip(ResPacket(c.second)));
+      tmp = padd(ResPacket(c.first),tmp);
+    }
+    else if((ConjLhs)&&(!ConjRhs))
+    {
+      tmp = pcplxflip(ResPacket(c.second));
+      tmp = padd(pconj(ResPacket(c.first)),tmp);
+    }
+    else if((ConjLhs)&&(ConjRhs))
+    {
+      tmp = pcplxflip(ResPacket(c.second));
+      tmp = psub(pconj(ResPacket(c.first)),tmp);
+    }
+    
+    r = pmadd(tmp,alpha,r);
+  }
+
+protected:
+  conj_helper<LhsScalar,RhsScalar,ConjLhs,ConjRhs> cj;
+};
+
+template<typename RealScalar, bool _ConjRhs>
+class gebp_traits<RealScalar, std::complex<RealScalar>, false, _ConjRhs >
+{
+public:
+  typedef std::complex<RealScalar>  Scalar;
+  typedef RealScalar  LhsScalar;
+  typedef Scalar      RhsScalar;
+  typedef Scalar      ResScalar;
+
+  enum {
+    ConjLhs = false,
+    ConjRhs = _ConjRhs,
+    Vectorizable = packet_traits<RealScalar>::Vectorizable
+                && packet_traits<Scalar>::Vectorizable,
+    LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+    RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+    ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1,
+    
+    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+    // FIXME: should depend on NumberOfRegisters
+    nr = 4,
+    mr = (EIGEN_PLAIN_ENUM_MIN(16,NumberOfRegisters)/2/nr)*ResPacketSize,
+
+    LhsProgress = ResPacketSize,
+    RhsProgress = 1
+  };
+
+  typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
+  typedef typename packet_traits<RhsScalar>::type  _RhsPacket;
+  typedef typename packet_traits<ResScalar>::type  _ResPacket;
+
+  typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
+  typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
+  typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+
+  typedef ResPacket AccPacket;
+
+  EIGEN_STRONG_INLINE void initAcc(AccPacket& p)
+  {
+    p = pset1<ResPacket>(ResScalar(0));
+  }
+
+  EIGEN_STRONG_INLINE void loadRhs(const RhsScalar* b, RhsPacket& dest) const
+  {
+    dest = pset1<RhsPacket>(*b);
+  }
+  
+  void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1, RhsPacket& b2, RhsPacket& b3)
+  {
+    pbroadcast4(b, b0, b1, b2, b3);
+  }
+  
+//   EIGEN_STRONG_INLINE void broadcastRhs(const RhsScalar* b, RhsPacket& b0, RhsPacket& b1)
+//   {
+//     // FIXME not sure that's the best way to implement it!
+//     b0 = pload1<RhsPacket>(b+0);
+//     b1 = pload1<RhsPacket>(b+1);
+//   }
+
+  EIGEN_STRONG_INLINE void loadLhs(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploaddup<LhsPacket>(a);
+  }
+  
+  EIGEN_STRONG_INLINE void loadRhsQuad(const RhsScalar* b, RhsPacket& dest) const
+  {
+    eigen_internal_assert(unpacket_traits<RhsPacket>::size<=4);
+    loadRhs(b,dest);
+  }
+
+  EIGEN_STRONG_INLINE void loadLhsUnaligned(const LhsScalar* a, LhsPacket& dest) const
+  {
+    dest = ploaddup<LhsPacket>(a);
+  }
+
+  EIGEN_STRONG_INLINE void madd(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp) const
+  {
+    madd_impl(a, b, c, tmp, typename conditional<Vectorizable,true_type,false_type>::type());
+  }
+
+  EIGEN_STRONG_INLINE void madd_impl(const LhsPacket& a, const RhsPacket& b, AccPacket& c, RhsPacket& tmp, const true_type&) const
+  {
+#ifdef EIGEN_HAS_SINGLE_INSTRUCTION_MADD
+    EIGEN_UNUSED_VARIABLE(tmp);
+    c.v = pmadd(a,b.v,c.v);
+#else
+    tmp = b; tmp.v = pmul(a,tmp.v); c = padd(c,tmp);
+#endif
+    
+  }
+
+  EIGEN_STRONG_INLINE void madd_impl(const LhsScalar& a, const RhsScalar& b, ResScalar& c, RhsScalar& /*tmp*/, const false_type&) const
+  {
+    c += a * b;
+  }
+
+  EIGEN_STRONG_INLINE void acc(const AccPacket& c, const ResPacket& alpha, ResPacket& r) const
+  {
+    r = cj.pmadd(alpha,c,r);
+  }
+
+protected:
+  conj_helper<ResPacket,ResPacket,false,ConjRhs> cj;
+};
+
+/* optimized GEneral packed Block * packed Panel product kernel
+ *
+ * Mixing type logic: C += A * B
+ *  |  A  |  B  | comments
+ *  |real |cplx | no vectorization yet, would require to pack A with duplication
+ *  |cplx |real | easy vectorization
+ */
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
+struct gebp_kernel
+{
+  typedef gebp_traits<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> Traits;
+  typedef typename Traits::ResScalar ResScalar;
+  typedef typename Traits::LhsPacket LhsPacket;
+  typedef typename Traits::RhsPacket RhsPacket;
+  typedef typename Traits::ResPacket ResPacket;
+  typedef typename Traits::AccPacket AccPacket;
+
+  typedef gebp_traits<RhsScalar,LhsScalar,ConjugateRhs,ConjugateLhs> SwappedTraits;
+  typedef typename SwappedTraits::ResScalar SResScalar;
+  typedef typename SwappedTraits::LhsPacket SLhsPacket;
+  typedef typename SwappedTraits::RhsPacket SRhsPacket;
+  typedef typename SwappedTraits::ResPacket SResPacket;
+  typedef typename SwappedTraits::AccPacket SAccPacket;
+
+  typedef typename DataMapper::LinearMapper LinearMapper;
+
+  enum {
+    Vectorizable  = Traits::Vectorizable,
+    LhsProgress   = Traits::LhsProgress,
+    RhsProgress   = Traits::RhsProgress,
+    ResPacketSize = Traits::ResPacketSize
+  };
+
+  EIGEN_DONT_INLINE
+  void operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB,
+                  Index rows, Index depth, Index cols, ResScalar alpha,
+                  Index strideA=-1, Index strideB=-1, Index offsetA=0, Index offsetB=0);
+};
+
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs, bool ConjugateRhs>
+EIGEN_DONT_INLINE
+void gebp_kernel<LhsScalar,RhsScalar,Index,DataMapper,mr,nr,ConjugateLhs,ConjugateRhs>
+  ::operator()(const DataMapper& res, const LhsScalar* blockA, const RhsScalar* blockB,
+               Index rows, Index depth, Index cols, ResScalar alpha,
+               Index strideA, Index strideB, Index offsetA, Index offsetB)
+  {
+    Traits traits;
+    SwappedTraits straits;
+    
+    if(strideA==-1) strideA = depth;
+    if(strideB==-1) strideB = depth;
+    conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
+    Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+    const Index peeled_mc3 = mr>=3*Traits::LhsProgress ? (rows/(3*LhsProgress))*(3*LhsProgress) : 0;
+    const Index peeled_mc2 = mr>=2*Traits::LhsProgress ? peeled_mc3+((rows-peeled_mc3)/(2*LhsProgress))*(2*LhsProgress) : 0;
+    const Index peeled_mc1 = mr>=1*Traits::LhsProgress ? (rows/(1*LhsProgress))*(1*LhsProgress) : 0;
+    enum { pk = 8 }; // NOTE Such a large peeling factor is important for large matrices (~ +5% when >1000 on Haswell)
+    const Index peeled_kc  = depth & ~(pk-1);
+    const Index prefetch_res_offset = 32/sizeof(ResScalar);    
+//     const Index depth2     = depth & ~1;
+
+    //---------- Process 3 * LhsProgress rows at once ----------
+    // This corresponds to 3*LhsProgress x nr register blocks.
+    // Usually, make sense only with FMA
+    if(mr>=3*Traits::LhsProgress)
+    {
+      // Here, the general idea is to loop on each largest micro horizontal panel of the lhs (3*Traits::LhsProgress x depth)
+      // and on each largest micro vertical panel of the rhs (depth * nr).
+      // Blocking sizes, i.e., 'depth' has been computed so that the micro horizontal panel of the lhs fit in L1.
+      // However, if depth is too small, we can extend the number of rows of these horizontal panels.
+      // This actual number of rows is computed as follow:
+      const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
+      // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
+      // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
+      // or because we are testing specific blocking sizes.
+      const Index actual_panel_rows = (3*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 3*LhsProgress) ));
+      for(Index i1=0; i1<peeled_mc3; i1+=actual_panel_rows)
+      {
+        const Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc3);
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
+        {
+          for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress)
+          {
+          
+          // We selected a 3*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 3 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2,  C3,
+                    C4, C5, C6,  C7,
+                    C8, C9, C10, C11;
+          traits.initAcc(C0);  traits.initAcc(C1);  traits.initAcc(C2);  traits.initAcc(C3);
+          traits.initAcc(C4);  traits.initAcc(C5);  traits.initAcc(C6);  traits.initAcc(C7);
+          traits.initAcc(C8);  traits.initAcc(C9);  traits.initAcc(C10); traits.initAcc(C11);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(0);
+          r1.prefetch(0);
+          r2.prefetch(0);
+          r3.prefetch(0);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0, A1;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX4");
+            RhsPacket B_0, T0;
+            LhsPacket A2;
+
+#define EIGEN_GEBP_ONESTEP(K) \
+            do { \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX4"); \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              internal::prefetch(blA+(3*K+16)*LhsProgress); \
+              if (EIGEN_ARCH_ARM) { internal::prefetch(blB+(4*K+16)*RhsProgress); } /* Bug 953 */ \
+              traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
+              traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
+              traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \
+              traits.loadRhs(blB + (0+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C0, T0); \
+              traits.madd(A1, B_0, C4, T0); \
+              traits.madd(A2, B_0, C8, B_0); \
+              traits.loadRhs(blB + (1+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C1, T0); \
+              traits.madd(A1, B_0, C5, T0); \
+              traits.madd(A2, B_0, C9, B_0); \
+              traits.loadRhs(blB + (2+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C2,  T0); \
+              traits.madd(A1, B_0, C6,  T0); \
+              traits.madd(A2, B_0, C10, B_0); \
+              traits.loadRhs(blB + (3+4*K)*Traits::RhsProgress, B_0); \
+              traits.madd(A0, B_0, C3 , T0); \
+              traits.madd(A1, B_0, C7,  T0); \
+              traits.madd(A2, B_0, C11, B_0); \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX4"); \
+            } while(false)
+
+            internal::prefetch(blB);
+            EIGEN_GEBP_ONESTEP(0);
+            EIGEN_GEBP_ONESTEP(1);
+            EIGEN_GEBP_ONESTEP(2);
+            EIGEN_GEBP_ONESTEP(3);
+            EIGEN_GEBP_ONESTEP(4);
+            EIGEN_GEBP_ONESTEP(5);
+            EIGEN_GEBP_ONESTEP(6);
+            EIGEN_GEBP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*3*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 3pX4");
+          }
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0, T0;
+            LhsPacket A2;
+            EIGEN_GEBP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 3*Traits::LhsProgress;
+          }
+
+#undef EIGEN_GEBP_ONESTEP
+
+          ResPacket R0, R1, R2;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C8, alphav, R2);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r0.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r1.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r1.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C1, alphav, R0);
+          traits.acc(C5, alphav, R1);
+          traits.acc(C9, alphav, R2);
+          r1.storePacket(0 * Traits::ResPacketSize, R0);
+          r1.storePacket(1 * Traits::ResPacketSize, R1);
+          r1.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r2.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C2, alphav, R0);
+          traits.acc(C6, alphav, R1);
+          traits.acc(C10, alphav, R2);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r2.storePacket(1 * Traits::ResPacketSize, R1);
+          r2.storePacket(2 * Traits::ResPacketSize, R2);
+
+          R0 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r3.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r3.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C3, alphav, R0);
+          traits.acc(C7, alphav, R1);
+          traits.acc(C11, alphav, R2);
+          r3.storePacket(0 * Traits::ResPacketSize, R0);
+          r3.storePacket(1 * Traits::ResPacketSize, R1);
+          r3.storePacket(2 * Traits::ResPacketSize, R2);          
+          }
+        }
+
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
+        {
+          for(Index i=i1; i<actual_panel_end; i+=3*LhsProgress)
+          {
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(3*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C4, C8;
+          traits.initAcc(C0);
+          traits.initAcc(C4);
+          traits.initAcc(C8);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2);
+          r0.prefetch(0);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0, A1, A2;
+          
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 3pX1");
+            RhsPacket B_0;
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do { \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 3pX1"); \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+3*K)*LhsProgress], A0);  \
+              traits.loadLhs(&blA[(1+3*K)*LhsProgress], A1);  \
+              traits.loadLhs(&blA[(2+3*K)*LhsProgress], A2);  \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);   \
+              traits.madd(A0, B_0, C0, B_0); \
+              traits.madd(A1, B_0, C4, B_0); \
+              traits.madd(A2, B_0, C8, B_0); \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 3pX1"); \
+            } while(false)
+        
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*3*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 3pX1");
+          }
+
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 3*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0, R1, R2;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r0.loadPacket(2 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C8, alphav, R2);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r0.storePacket(2 * Traits::ResPacketSize, R2);          
+          }
+        }
+      }
+    }
+
+    //---------- Process 2 * LhsProgress rows at once ----------
+    if(mr>=2*Traits::LhsProgress)
+    {
+      const Index l1 = defaultL1CacheSize; // in Bytes, TODO, l1 should be passed to this function.
+      // The max(1, ...) here is needed because we may be using blocking params larger than what our known l1 cache size
+      // suggests we should be using: either because our known l1 cache size is inaccurate (e.g. on Android, we can only guess),
+      // or because we are testing specific blocking sizes.
+      Index actual_panel_rows = (2*LhsProgress) * std::max<Index>(1,( (l1 - sizeof(ResScalar)*mr*nr - depth*nr*sizeof(RhsScalar)) / (depth * sizeof(LhsScalar) * 2*LhsProgress) ));
+
+      for(Index i1=peeled_mc3; i1<peeled_mc2; i1+=actual_panel_rows)
+      {
+        Index actual_panel_end = (std::min)(i1+actual_panel_rows, peeled_mc2);
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
+        {
+          for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress)
+          {
+          
+          // We selected a 2*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 2 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2, C3,
+                    C4, C5, C6, C7;
+          traits.initAcc(C0); traits.initAcc(C1); traits.initAcc(C2); traits.initAcc(C3);
+          traits.initAcc(C4); traits.initAcc(C5); traits.initAcc(C6); traits.initAcc(C7);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(prefetch_res_offset);
+          r1.prefetch(prefetch_res_offset);
+          r2.prefetch(prefetch_res_offset);
+          r3.prefetch(prefetch_res_offset);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0, A1;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX4");
+            RhsPacket B_0, B1, B2, B3, T0;
+
+   #define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX4");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0);                    \
+              traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1);                    \
+              traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3);  \
+              traits.madd(A0, B_0, C0, T0);                                     \
+              traits.madd(A1, B_0, C4, B_0);                                    \
+              traits.madd(A0, B1,  C1, T0);                                     \
+              traits.madd(A1, B1,  C5, B1);                                     \
+              traits.madd(A0, B2,  C2, T0);                                     \
+              traits.madd(A1, B2,  C6, B2);                                     \
+              traits.madd(A0, B3,  C3, T0);                                     \
+              traits.madd(A1, B3,  C7, B3);                                     \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX4");          \
+            } while(false)
+            
+            internal::prefetch(blB+(48+0));
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            internal::prefetch(blB+(48+16));
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*(2*Traits::LhsProgress);
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 2pX4");
+          }
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0, B1, B2, B3, T0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 2*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+
+          ResPacket R0, R1, R2, R3;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r1.loadPacket(0 * Traits::ResPacketSize);
+          R3 = r1.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          traits.acc(C1, alphav, R2);
+          traits.acc(C5, alphav, R3);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          r1.storePacket(0 * Traits::ResPacketSize, R2);
+          r1.storePacket(1 * Traits::ResPacketSize, R3);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r2.loadPacket(1 * Traits::ResPacketSize);
+          R2 = r3.loadPacket(0 * Traits::ResPacketSize);
+          R3 = r3.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C2,  alphav, R0);
+          traits.acc(C6,  alphav, R1);
+          traits.acc(C3,  alphav, R2);
+          traits.acc(C7,  alphav, R3);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r2.storePacket(1 * Traits::ResPacketSize, R1);
+          r3.storePacket(0 * Traits::ResPacketSize, R2);
+          r3.storePacket(1 * Traits::ResPacketSize, R3);
+          }
+        }
+      
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
+        {
+          for(Index i=i1; i<actual_panel_end; i+=2*LhsProgress)
+          {
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(2*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C4;
+          traits.initAcc(C0);
+          traits.initAcc(C4);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2);
+          r0.prefetch(prefetch_res_offset);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0, A1;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 2pX1");
+            RhsPacket B_0, B1;
+        
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                  \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 2pX1");          \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+2*K)*LhsProgress], A0);                      \
+              traits.loadLhs(&blA[(1+2*K)*LhsProgress], A1);                      \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);                       \
+              traits.madd(A0, B_0, C0, B1);                                       \
+              traits.madd(A1, B_0, C4, B_0);                                      \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 2pX1");            \
+            } while(false)
+        
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*2*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 2pX1");
+          }
+
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0, B1;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 2*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0, R1;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r0.loadPacket(1 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C4, alphav, R1);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r0.storePacket(1 * Traits::ResPacketSize, R1);
+          }
+        }
+      }
+    }
+    //---------- Process 1 * LhsProgress rows at once ----------
+    if(mr>=1*Traits::LhsProgress)
+    {
+      // loops on each largest micro horizontal panel of lhs (1*LhsProgress x depth)
+      for(Index i=peeled_mc2; i<peeled_mc1; i+=1*LhsProgress)
+      {
+        // loops on each largest micro vertical panel of rhs (depth * nr)
+        for(Index j2=0; j2<packet_cols4; j2+=nr)
+        {
+          // We select a 1*Traits::LhsProgress x nr micro block of res which is entirely
+          // stored into 1 x nr registers.
+          
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0, C1, C2, C3;
+          traits.initAcc(C0);
+          traits.initAcc(C1);
+          traits.initAcc(C2);
+          traits.initAcc(C3);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2 + 0);
+          LinearMapper r1 = res.getLinearMapper(i, j2 + 1);
+          LinearMapper r2 = res.getLinearMapper(i, j2 + 2);
+          LinearMapper r3 = res.getLinearMapper(i, j2 + 3);
+
+          r0.prefetch(prefetch_res_offset);
+          r1.prefetch(prefetch_res_offset);
+          r2.prefetch(prefetch_res_offset);
+          r3.prefetch(prefetch_res_offset);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+          prefetch(&blB[0]);
+          LhsPacket A0;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX4");
+            RhsPacket B_0, B1, B2, B3;
+               
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX4");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0);                    \
+              traits.broadcastRhs(&blB[(0+4*K)*RhsProgress], B_0, B1, B2, B3);  \
+              traits.madd(A0, B_0, C0, B_0);                                    \
+              traits.madd(A0, B1,  C1, B1);                                     \
+              traits.madd(A0, B2,  C2, B2);                                     \
+              traits.madd(A0, B3,  C3, B3);                                     \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX4");          \
+            } while(false)
+            
+            internal::prefetch(blB+(48+0));
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            internal::prefetch(blB+(48+16));
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*4*RhsProgress;
+            blA += pk*1*LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 1pX4");
+          }
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0, B1, B2, B3;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += 4*RhsProgress;
+            blA += 1*LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+
+          ResPacket R0, R1;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r1.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          traits.acc(C1,  alphav, R1);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+          r1.storePacket(0 * Traits::ResPacketSize, R1);
+
+          R0 = r2.loadPacket(0 * Traits::ResPacketSize);
+          R1 = r3.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C2,  alphav, R0);
+          traits.acc(C3,  alphav, R1);
+          r2.storePacket(0 * Traits::ResPacketSize, R0);
+          r3.storePacket(0 * Traits::ResPacketSize, R1);
+        }
+
+        // Deal with remaining columns of the rhs
+        for(Index j2=packet_cols4; j2<cols; j2++)
+        {
+          // One column at a time
+          const LhsScalar* blA = &blockA[i*strideA+offsetA*(1*Traits::LhsProgress)];
+          prefetch(&blA[0]);
+
+          // gets res block as register
+          AccPacket C0;
+          traits.initAcc(C0);
+
+          LinearMapper r0 = res.getLinearMapper(i, j2);
+
+          // performs "inner" products
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          LhsPacket A0;
+
+          for(Index k=0; k<peeled_kc; k+=pk)
+          {
+            EIGEN_ASM_COMMENT("begin gebp micro kernel 1pX1");
+            RhsPacket B_0;
+        
+#define EIGEN_GEBGP_ONESTEP(K) \
+            do {                                                                \
+              EIGEN_ASM_COMMENT("begin step of gebp micro kernel 1pX1");        \
+              EIGEN_ASM_COMMENT("Note: these asm comments work around bug 935!"); \
+              traits.loadLhs(&blA[(0+1*K)*LhsProgress], A0);                    \
+              traits.loadRhs(&blB[(0+K)*RhsProgress], B_0);                     \
+              traits.madd(A0, B_0, C0, B_0);                                    \
+              EIGEN_ASM_COMMENT("end step of gebp micro kernel 1pX1");          \
+            } while(false);
+
+            EIGEN_GEBGP_ONESTEP(0);
+            EIGEN_GEBGP_ONESTEP(1);
+            EIGEN_GEBGP_ONESTEP(2);
+            EIGEN_GEBGP_ONESTEP(3);
+            EIGEN_GEBGP_ONESTEP(4);
+            EIGEN_GEBGP_ONESTEP(5);
+            EIGEN_GEBGP_ONESTEP(6);
+            EIGEN_GEBGP_ONESTEP(7);
+
+            blB += pk*RhsProgress;
+            blA += pk*1*Traits::LhsProgress;
+
+            EIGEN_ASM_COMMENT("end gebp micro kernel 1pX1");
+          }
+
+          // process remaining peeled loop
+          for(Index k=peeled_kc; k<depth; k++)
+          {
+            RhsPacket B_0;
+            EIGEN_GEBGP_ONESTEP(0);
+            blB += RhsProgress;
+            blA += 1*Traits::LhsProgress;
+          }
+#undef EIGEN_GEBGP_ONESTEP
+          ResPacket R0;
+          ResPacket alphav = pset1<ResPacket>(alpha);
+          R0 = r0.loadPacket(0 * Traits::ResPacketSize);
+          traits.acc(C0, alphav, R0);
+          r0.storePacket(0 * Traits::ResPacketSize, R0);
+        }
+      }
+    }
+    //---------- Process remaining rows, 1 at once ----------
+    if(peeled_mc1<rows)
+    {
+      // loop on each panel of the rhs
+      for(Index j2=0; j2<packet_cols4; j2+=nr)
+      {
+        // loop on each row of the lhs (1*LhsProgress x depth)
+        for(Index i=peeled_mc1; i<rows; i+=1)
+        {
+          const LhsScalar* blA = &blockA[i*strideA+offsetA];
+          prefetch(&blA[0]);
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB*nr];
+
+          // The following piece of code wont work for 512 bit registers
+          // Moreover, if LhsProgress==8 it assumes that there is a half packet of the same size
+          // as nr (which is currently 4) for the return type.
+          typedef typename unpacket_traits<SResPacket>::half SResPacketHalf;
+          if ((SwappedTraits::LhsProgress % 4) == 0 &&
+              (SwappedTraits::LhsProgress <= 8) &&
+              (SwappedTraits::LhsProgress!=8 || unpacket_traits<SResPacketHalf>::size==nr))
+          {
+            SAccPacket C0, C1, C2, C3;
+            straits.initAcc(C0);
+            straits.initAcc(C1);
+            straits.initAcc(C2);
+            straits.initAcc(C3);
+
+            const Index spk   = (std::max)(1,SwappedTraits::LhsProgress/4);
+            const Index endk  = (depth/spk)*spk;
+            const Index endk4 = (depth/(spk*4))*(spk*4);
+
+            Index k=0;
+            for(; k<endk4; k+=4*spk)
+            {
+              SLhsPacket A0,A1;
+              SRhsPacket B_0,B_1;
+
+              straits.loadLhsUnaligned(blB+0*SwappedTraits::LhsProgress, A0);
+              straits.loadLhsUnaligned(blB+1*SwappedTraits::LhsProgress, A1);
+
+              straits.loadRhsQuad(blA+0*spk, B_0);
+              straits.loadRhsQuad(blA+1*spk, B_1);
+              straits.madd(A0,B_0,C0,B_0);
+              straits.madd(A1,B_1,C1,B_1);
+
+              straits.loadLhsUnaligned(blB+2*SwappedTraits::LhsProgress, A0);
+              straits.loadLhsUnaligned(blB+3*SwappedTraits::LhsProgress, A1);
+              straits.loadRhsQuad(blA+2*spk, B_0);
+              straits.loadRhsQuad(blA+3*spk, B_1);
+              straits.madd(A0,B_0,C2,B_0);
+              straits.madd(A1,B_1,C3,B_1);
+
+              blB += 4*SwappedTraits::LhsProgress;
+              blA += 4*spk;
+            }
+            C0 = padd(padd(C0,C1),padd(C2,C3));
+            for(; k<endk; k+=spk)
+            {
+              SLhsPacket A0;
+              SRhsPacket B_0;
+
+              straits.loadLhsUnaligned(blB, A0);
+              straits.loadRhsQuad(blA, B_0);
+              straits.madd(A0,B_0,C0,B_0);
+
+              blB += SwappedTraits::LhsProgress;
+              blA += spk;
+            }
+            if(SwappedTraits::LhsProgress==8)
+            {
+              // Special case where we have to first reduce the accumulation register C0
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SResPacket>::half,SResPacket>::type SResPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SLhsPacket>::type SLhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SLhsPacket>::half,SRhsPacket>::type SRhsPacketHalf;
+              typedef typename conditional<SwappedTraits::LhsProgress>=8,typename unpacket_traits<SAccPacket>::half,SAccPacket>::type SAccPacketHalf;
+
+              SResPacketHalf R = res.template gatherPacket<SResPacketHalf>(i, j2);
+              SResPacketHalf alphav = pset1<SResPacketHalf>(alpha);
+
+              if(depth-endk>0)
+              {
+                // We have to handle the last row of the rhs which corresponds to a half-packet
+                SLhsPacketHalf a0;
+                SRhsPacketHalf b0;
+                straits.loadLhsUnaligned(blB, a0);
+                straits.loadRhs(blA, b0);
+                SAccPacketHalf c0 = predux_downto4(C0);
+                straits.madd(a0,b0,c0,b0);
+                straits.acc(c0, alphav, R);
+              }
+              else
+              {
+                straits.acc(predux_downto4(C0), alphav, R);
+              }
+              res.scatterPacket(i, j2, R);
+            }
+            else
+            {
+              SResPacket R = res.template gatherPacket<SResPacket>(i, j2);
+              SResPacket alphav = pset1<SResPacket>(alpha);
+              straits.acc(C0, alphav, R);
+              res.scatterPacket(i, j2, R);
+            }
+          }
+          else // scalar path
+          {
+            // get a 1 x 4 res block as registers
+            ResScalar C0(0), C1(0), C2(0), C3(0);
+
+            for(Index k=0; k<depth; k++)
+            {
+              LhsScalar A0;
+              RhsScalar B_0, B_1;
+
+              A0 = blA[k];
+
+              B_0 = blB[0];
+              B_1 = blB[1];
+              CJMADD(cj,A0,B_0,C0,  B_0);
+              CJMADD(cj,A0,B_1,C1,  B_1);
+              
+              B_0 = blB[2];
+              B_1 = blB[3];
+              CJMADD(cj,A0,B_0,C2,  B_0);
+              CJMADD(cj,A0,B_1,C3,  B_1);
+              
+              blB += 4;
+            }
+            res(i, j2 + 0) += alpha * C0;
+            res(i, j2 + 1) += alpha * C1;
+            res(i, j2 + 2) += alpha * C2;
+            res(i, j2 + 3) += alpha * C3;
+          }
+        }
+      }
+      // remaining columns
+      for(Index j2=packet_cols4; j2<cols; j2++)
+      {
+        // loop on each row of the lhs (1*LhsProgress x depth)
+        for(Index i=peeled_mc1; i<rows; i+=1)
+        {
+          const LhsScalar* blA = &blockA[i*strideA+offsetA];
+          prefetch(&blA[0]);
+          // gets a 1 x 1 res block as registers
+          ResScalar C0(0);
+          const RhsScalar* blB = &blockB[j2*strideB+offsetB];
+          for(Index k=0; k<depth; k++)
+          {
+            LhsScalar A0 = blA[k];
+            RhsScalar B_0 = blB[k];
+            CJMADD(cj, A0, B_0, C0, B_0);
+          }
+          res(i, j2) += alpha * C0;
+        }
+      }
+    }
+  }
+
+
+#undef CJMADD
+
+// pack a block of the lhs
+// The traversal is as follow (mr==4):
+//   0  4  8 12 ...
+//   1  5  9 13 ...
+//   2  6 10 14 ...
+//   3  7 11 15 ...
+//
+//  16 20 24 28 ...
+//  17 21 25 29 ...
+//  18 22 26 30 ...
+//  19 23 27 31 ...
+//
+//  32 33 34 35 ...
+//  36 36 38 39 ...
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+{
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
+};
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, ColMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum { PacketSize = packet_traits<Scalar>::size };
+
+  EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
+  eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+  eigen_assert( ((Pack1%PacketSize)==0 && Pack1<=4*PacketSize) || (Pack1<=4) );
+  conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+  Index count = 0;
+
+  const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+  const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+  const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+  const Index peeled_mc0 = Pack2>=1*PacketSize ? peeled_mc1
+                         : Pack2>1             ? (rows/Pack2)*Pack2 : 0;
+
+  Index i=0;
+
+  // Pack 3 packets
+  if(Pack1>=3*PacketSize)
+  {
+    for(; i<peeled_mc3; i+=3*PacketSize)
+    {
+      if(PanelMode) count += (3*PacketSize) * offset;
+
+      for(Index k=0; k<depth; k++)
+      {
+        Packet A, B, C;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        B = lhs.loadPacket(i+1*PacketSize, k);
+        C = lhs.loadPacket(i+2*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(C)); count+=PacketSize;
+      }
+      if(PanelMode) count += (3*PacketSize) * (stride-offset-depth);
+    }
+  }
+  // Pack 2 packets
+  if(Pack1>=2*PacketSize)
+  {
+    for(; i<peeled_mc2; i+=2*PacketSize)
+    {
+      if(PanelMode) count += (2*PacketSize) * offset;
+
+      for(Index k=0; k<depth; k++)
+      {
+        Packet A, B;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        B = lhs.loadPacket(i+1*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A)); count+=PacketSize;
+        pstore(blockA+count, cj.pconj(B)); count+=PacketSize;
+      }
+      if(PanelMode) count += (2*PacketSize) * (stride-offset-depth);
+    }
+  }
+  // Pack 1 packets
+  if(Pack1>=1*PacketSize)
+  {
+    for(; i<peeled_mc1; i+=1*PacketSize)
+    {
+      if(PanelMode) count += (1*PacketSize) * offset;
+
+      for(Index k=0; k<depth; k++)
+      {
+        Packet A;
+        A = lhs.loadPacket(i+0*PacketSize, k);
+        pstore(blockA+count, cj.pconj(A));
+        count+=PacketSize;
+      }
+      if(PanelMode) count += (1*PacketSize) * (stride-offset-depth);
+    }
+  }
+  // Pack scalars
+  if(Pack2<PacketSize && Pack2>1)
+  {
+    for(; i<peeled_mc0; i+=Pack2)
+    {
+      if(PanelMode) count += Pack2 * offset;
+
+      for(Index k=0; k<depth; k++)
+        for(Index w=0; w<Pack2; w++)
+          blockA[count++] = cj(lhs(i+w, k));
+
+      if(PanelMode) count += Pack2 * (stride-offset-depth);
+    }
+  }
+  for(; i<rows; i++)
+  {
+    if(PanelMode) count += offset;
+    for(Index k=0; k<depth; k++)
+      blockA[count++] = cj(lhs(i, k));
+    if(PanelMode) count += (stride-offset-depth);
+  }
+}
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+struct gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+{
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  EIGEN_DONT_INLINE void operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride=0, Index offset=0);
+};
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_lhs<Scalar, Index, DataMapper, Pack1, Pack2, RowMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockA, const DataMapper& lhs, Index depth, Index rows, Index stride, Index offset)
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum { PacketSize = packet_traits<Scalar>::size };
+
+  EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK LHS");
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
+  eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+  conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+  Index count = 0;
+
+//   const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+//   const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+//   const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+
+  int pack = Pack1;
+  Index i = 0;
+  while(pack>0)
+  {
+    Index remaining_rows = rows-i;
+    Index peeled_mc = i+(remaining_rows/pack)*pack;
+    for(; i<peeled_mc; i+=pack)
+    {
+      if(PanelMode) count += pack * offset;
+
+      const Index peeled_k = (depth/PacketSize)*PacketSize;
+      Index k=0;
+      if(pack>=PacketSize)
+      {
+        for(; k<peeled_k; k+=PacketSize)
+        {
+          for (Index m = 0; m < pack; m += PacketSize)
+          {
+            PacketBlock<Packet> kernel;
+            for (int p = 0; p < PacketSize; ++p) kernel.packet[p] = lhs.loadPacket(i+p+m, k);
+            ptranspose(kernel);
+            for (int p = 0; p < PacketSize; ++p) pstore(blockA+count+m+(pack)*p, cj.pconj(kernel.packet[p]));
+          }
+          count += PacketSize*pack;
+        }
+      }
+      for(; k<depth; k++)
+      {
+        Index w=0;
+        for(; w<pack-3; w+=4)
+        {
+          Scalar a(cj(lhs(i+w+0, k))),
+                 b(cj(lhs(i+w+1, k))),
+                 c(cj(lhs(i+w+2, k))),
+                 d(cj(lhs(i+w+3, k)));
+          blockA[count++] = a;
+          blockA[count++] = b;
+          blockA[count++] = c;
+          blockA[count++] = d;
+        }
+        if(pack%4)
+          for(;w<pack;++w)
+            blockA[count++] = cj(lhs(i+w, k));
+      }
+
+      if(PanelMode) count += pack * (stride-offset-depth);
+    }
+
+    pack -= PacketSize;
+    if(pack<Pack2 && (pack+PacketSize)!=Pack2)
+      pack = Pack2;
+  }
+
+  for(; i<rows; i++)
+  {
+    if(PanelMode) count += offset;
+    for(Index k=0; k<depth; k++)
+      blockA[count++] = cj(lhs(i, k));
+    if(PanelMode) count += (stride-offset-depth);
+  }
+}
+
+// copy a complete panel of the rhs
+// this version is optimized for column major matrices
+// The traversal order is as follow: (nr==4):
+//  0  1  2  3   12 13 14 15   24 27
+//  4  5  6  7   16 17 18 19   25 28
+//  8  9 10 11   20 21 22 23   26 29
+//  .  .  .  .    .  .  .  .    .  .
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode>
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  enum { PacketSize = packet_traits<Scalar>::size };
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0);
+};
+
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, ColMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset)
+{
+  EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS COLMAJOR");
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
+  eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+  conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+  Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+  Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+  Index count = 0;
+  const Index peeled_k = (depth/PacketSize)*PacketSize;
+//   if(nr>=8)
+//   {
+//     for(Index j2=0; j2<packet_cols8; j2+=8)
+//     {
+//       // skip what we have before
+//       if(PanelMode) count += 8 * offset;
+//       const Scalar* b0 = &rhs[(j2+0)*rhsStride];
+//       const Scalar* b1 = &rhs[(j2+1)*rhsStride];
+//       const Scalar* b2 = &rhs[(j2+2)*rhsStride];
+//       const Scalar* b3 = &rhs[(j2+3)*rhsStride];
+//       const Scalar* b4 = &rhs[(j2+4)*rhsStride];
+//       const Scalar* b5 = &rhs[(j2+5)*rhsStride];
+//       const Scalar* b6 = &rhs[(j2+6)*rhsStride];
+//       const Scalar* b7 = &rhs[(j2+7)*rhsStride];
+//       Index k=0;
+//       if(PacketSize==8) // TODO enbale vectorized transposition for PacketSize==4
+//       {
+//         for(; k<peeled_k; k+=PacketSize) {
+//           PacketBlock<Packet> kernel;
+//           for (int p = 0; p < PacketSize; ++p) {
+//             kernel.packet[p] = ploadu<Packet>(&rhs[(j2+p)*rhsStride+k]);
+//           }
+//           ptranspose(kernel);
+//           for (int p = 0; p < PacketSize; ++p) {
+//             pstoreu(blockB+count, cj.pconj(kernel.packet[p]));
+//             count+=PacketSize;
+//           }
+//         }
+//       }
+//       for(; k<depth; k++)
+//       {
+//         blockB[count+0] = cj(b0[k]);
+//         blockB[count+1] = cj(b1[k]);
+//         blockB[count+2] = cj(b2[k]);
+//         blockB[count+3] = cj(b3[k]);
+//         blockB[count+4] = cj(b4[k]);
+//         blockB[count+5] = cj(b5[k]);
+//         blockB[count+6] = cj(b6[k]);
+//         blockB[count+7] = cj(b7[k]);
+//         count += 8;
+//       }
+//       // skip what we have after
+//       if(PanelMode) count += 8 * (stride-offset-depth);
+//     }
+//   }
+
+  if(nr>=4)
+  {
+    for(Index j2=packet_cols8; j2<packet_cols4; j2+=4)
+    {
+      // skip what we have before
+      if(PanelMode) count += 4 * offset;
+      const LinearMapper dm0 = rhs.getLinearMapper(0, j2 + 0);
+      const LinearMapper dm1 = rhs.getLinearMapper(0, j2 + 1);
+      const LinearMapper dm2 = rhs.getLinearMapper(0, j2 + 2);
+      const LinearMapper dm3 = rhs.getLinearMapper(0, j2 + 3);
+
+      Index k=0;
+      if((PacketSize%4)==0) // TODO enable vectorized transposition for PacketSize==2 ??
+      {
+        for(; k<peeled_k; k+=PacketSize) {
+          PacketBlock<Packet,(PacketSize%4)==0?4:PacketSize> kernel;
+          kernel.packet[0] = dm0.loadPacket(k);
+          kernel.packet[1%PacketSize] = dm1.loadPacket(k);
+          kernel.packet[2%PacketSize] = dm2.loadPacket(k);
+          kernel.packet[3%PacketSize] = dm3.loadPacket(k);
+          ptranspose(kernel);
+          pstoreu(blockB+count+0*PacketSize, cj.pconj(kernel.packet[0]));
+          pstoreu(blockB+count+1*PacketSize, cj.pconj(kernel.packet[1%PacketSize]));
+          pstoreu(blockB+count+2*PacketSize, cj.pconj(kernel.packet[2%PacketSize]));
+          pstoreu(blockB+count+3*PacketSize, cj.pconj(kernel.packet[3%PacketSize]));
+          count+=4*PacketSize;
+        }
+      }
+      for(; k<depth; k++)
+      {
+        blockB[count+0] = cj(dm0(k));
+        blockB[count+1] = cj(dm1(k));
+        blockB[count+2] = cj(dm2(k));
+        blockB[count+3] = cj(dm3(k));
+        count += 4;
+      }
+      // skip what we have after
+      if(PanelMode) count += 4 * (stride-offset-depth);
+    }
+  }
+
+  // copy the remaining columns one at a time (nr==1)
+  for(Index j2=packet_cols4; j2<cols; ++j2)
+  {
+    if(PanelMode) count += offset;
+    const LinearMapper dm0 = rhs.getLinearMapper(0, j2);
+    for(Index k=0; k<depth; k++)
+    {
+      blockB[count] = cj(dm0(k));
+      count += 1;
+    }
+    if(PanelMode) count += (stride-offset-depth);
+  }
+}
+
+// this version is optimized for row major matrices
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+struct gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode>
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename DataMapper::LinearMapper LinearMapper;
+  enum { PacketSize = packet_traits<Scalar>::size };
+  EIGEN_DONT_INLINE void operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride=0, Index offset=0);
+};
+
+template<typename Scalar, typename Index, typename DataMapper, int nr, bool Conjugate, bool PanelMode>
+EIGEN_DONT_INLINE void gemm_pack_rhs<Scalar, Index, DataMapper, nr, RowMajor, Conjugate, PanelMode>
+  ::operator()(Scalar* blockB, const DataMapper& rhs, Index depth, Index cols, Index stride, Index offset)
+{
+  EIGEN_ASM_COMMENT("EIGEN PRODUCT PACK RHS ROWMAJOR");
+  EIGEN_UNUSED_VARIABLE(stride);
+  EIGEN_UNUSED_VARIABLE(offset);
+  eigen_assert(((!PanelMode) && stride==0 && offset==0) || (PanelMode && stride>=depth && offset<=stride));
+  conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+  Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+  Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+  Index count = 0;
+
+//   if(nr>=8)
+//   {
+//     for(Index j2=0; j2<packet_cols8; j2+=8)
+//     {
+//       // skip what we have before
+//       if(PanelMode) count += 8 * offset;
+//       for(Index k=0; k<depth; k++)
+//       {
+//         if (PacketSize==8) {
+//           Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
+//           pstoreu(blockB+count, cj.pconj(A));
+//         } else if (PacketSize==4) {
+//           Packet A = ploadu<Packet>(&rhs[k*rhsStride + j2]);
+//           Packet B = ploadu<Packet>(&rhs[k*rhsStride + j2 + PacketSize]);
+//           pstoreu(blockB+count, cj.pconj(A));
+//           pstoreu(blockB+count+PacketSize, cj.pconj(B));
+//         } else {
+//           const Scalar* b0 = &rhs[k*rhsStride + j2];
+//           blockB[count+0] = cj(b0[0]);
+//           blockB[count+1] = cj(b0[1]);
+//           blockB[count+2] = cj(b0[2]);
+//           blockB[count+3] = cj(b0[3]);
+//           blockB[count+4] = cj(b0[4]);
+//           blockB[count+5] = cj(b0[5]);
+//           blockB[count+6] = cj(b0[6]);
+//           blockB[count+7] = cj(b0[7]);
+//         }
+//         count += 8;
+//       }
+//       // skip what we have after
+//       if(PanelMode) count += 8 * (stride-offset-depth);
+//     }
+//   }
+  if(nr>=4)
+  {
+    for(Index j2=packet_cols8; j2<packet_cols4; j2+=4)
+    {
+      // skip what we have before
+      if(PanelMode) count += 4 * offset;
+      for(Index k=0; k<depth; k++)
+      {
+        if (PacketSize==4) {
+          Packet A = rhs.loadPacket(k, j2);
+          pstoreu(blockB+count, cj.pconj(A));
+          count += PacketSize;
+        } else {
+          const LinearMapper dm0 = rhs.getLinearMapper(k, j2);
+          blockB[count+0] = cj(dm0(0));
+          blockB[count+1] = cj(dm0(1));
+          blockB[count+2] = cj(dm0(2));
+          blockB[count+3] = cj(dm0(3));
+          count += 4;
+        }
+      }
+      // skip what we have after
+      if(PanelMode) count += 4 * (stride-offset-depth);
+    }
+  }
+  // copy the remaining columns one at a time (nr==1)
+  for(Index j2=packet_cols4; j2<cols; ++j2)
+  {
+    if(PanelMode) count += offset;
+    for(Index k=0; k<depth; k++)
+    {
+      blockB[count] = cj(rhs(k, j2));
+      count += 1;
+    }
+    if(PanelMode) count += stride-offset-depth;
+  }
+}
+
+} // end namespace internal
+
+/** \returns the currently set level 1 cpu cache size (in bytes) used to estimate the ideal blocking size parameters.
+  * \sa setCpuCacheSize */
+inline std::ptrdiff_t l1CacheSize()
+{
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+  return l1;
+}
+
+/** \returns the currently set level 2 cpu cache size (in bytes) used to estimate the ideal blocking size parameters.
+  * \sa setCpuCacheSize */
+inline std::ptrdiff_t l2CacheSize()
+{
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+  return l2;
+}
+
+/** \returns the currently set level 3 cpu cache size (in bytes) used to estimate the ideal blocking size paramete\
+rs.                                                                                                                
+* \sa setCpuCacheSize */
+inline std::ptrdiff_t l3CacheSize()
+{
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+  return l3;
+}
+
+/** Set the cpu L1 and L2 cache sizes (in bytes).
+  * These values are use to adjust the size of the blocks
+  * for the algorithms working per blocks.
+  *
+  * \sa computeProductBlockingSizes */
+inline void setCpuCacheSizes(std::ptrdiff_t l1, std::ptrdiff_t l2, std::ptrdiff_t l3)
+{
+  internal::manage_caching_sizes(SetAction, &l1, &l2, &l3);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_BLOCK_PANEL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h
new file mode 100644
index 0000000..6440e1d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix.h
@@ -0,0 +1,492 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename _LhsScalar, typename _RhsScalar> class level3_blocking;
+
+/* Specialization for a row-major destination matrix => simple transposition of the product */
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor>
+{
+  typedef gebp_traits<RhsScalar,LhsScalar> Traits;
+
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols, Index depth,
+    const LhsScalar* lhs, Index lhsStride,
+    const RhsScalar* rhs, Index rhsStride,
+    ResScalar* res, Index resStride,
+    ResScalar alpha,
+    level3_blocking<RhsScalar,LhsScalar>& blocking,
+    GemmParallelInfo<Index>* info = 0)
+  {
+    // transpose the product such that the result is column major
+    general_matrix_matrix_product<Index,
+      RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
+      LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
+      ColMajor>
+    ::run(cols,rows,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking,info);
+  }
+};
+
+/*  Specialization for a col-major destination matrix
+ *    => Blocking algorithm following Goto's paper */
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs>
+struct general_matrix_matrix_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor>
+{
+
+typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+static void run(Index rows, Index cols, Index depth,
+  const LhsScalar* _lhs, Index lhsStride,
+  const RhsScalar* _rhs, Index rhsStride,
+  ResScalar* _res, Index resStride,
+  ResScalar alpha,
+  level3_blocking<LhsScalar,RhsScalar>& blocking,
+  GemmParallelInfo<Index>* info = 0)
+{
+  typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
+  typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
+  typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+  LhsMapper lhs(_lhs,lhsStride);
+  RhsMapper rhs(_rhs,rhsStride);
+  ResMapper res(_res, resStride);
+
+  Index kc = blocking.kc();                   // cache block size along the K direction
+  Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+  Index nc = (std::min)(cols,blocking.nc());  // cache block size along the N direction
+
+  gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+  gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+  gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+
+#ifdef EIGEN_HAS_OPENMP
+  if(info)
+  {
+    // this is the parallel version!
+    int tid = omp_get_thread_num();
+    int threads = omp_get_num_threads();
+
+    LhsScalar* blockA = blocking.blockA();
+    eigen_internal_assert(blockA!=0);
+
+    std::size_t sizeB = kc*nc;
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, 0);
+
+    // For each horizontal panel of the rhs, and corresponding vertical panel of the lhs...
+    for(Index k=0; k<depth; k+=kc)
+    {
+      const Index actual_kc = (std::min)(k+kc,depth)-k; // => rows of B', and cols of the A'
+
+      // In order to reduce the chance that a thread has to wait for the other,
+      // let's start by packing B'.
+      pack_rhs(blockB, rhs.getSubMapper(k,0), actual_kc, nc);
+
+      // Pack A_k to A' in a parallel fashion:
+      // each thread packs the sub block A_k,i to A'_i where i is the thread id.
+
+      // However, before copying to A'_i, we have to make sure that no other thread is still using it,
+      // i.e., we test that info[tid].users equals 0.
+      // Then, we set info[tid].users to the number of threads to mark that all other threads are going to use it.
+      while(info[tid].users!=0) {}
+      info[tid].users += threads;
+
+      pack_lhs(blockA+info[tid].lhs_start*actual_kc, lhs.getSubMapper(info[tid].lhs_start,k), actual_kc, info[tid].lhs_length);
+
+      // Notify the other threads that the part A'_i is ready to go.
+      info[tid].sync = k;
+
+      // Computes C_i += A' * B' per A'_i
+      for(int shift=0; shift<threads; ++shift)
+      {
+        int i = (tid+shift)%threads;
+
+        // At this point we have to make sure that A'_i has been updated by the thread i,
+        // we use testAndSetOrdered to mimic a volatile access.
+        // However, no need to wait for the B' part which has been updated by the current thread!
+        if (shift>0) {
+          while(info[i].sync!=k) {
+          }
+        }
+
+        gebp(res.getSubMapper(info[i].lhs_start, 0), blockA+info[i].lhs_start*actual_kc, blockB, info[i].lhs_length, actual_kc, nc, alpha);
+      }
+
+      // Then keep going as usual with the remaining B'
+      for(Index j=nc; j<cols; j+=nc)
+      {
+        const Index actual_nc = (std::min)(j+nc,cols)-j;
+
+        // pack B_k,j to B'
+        pack_rhs(blockB, rhs.getSubMapper(k,j), actual_kc, actual_nc);
+
+        // C_j += A' * B'
+        gebp(res.getSubMapper(0, j), blockA, blockB, rows, actual_kc, actual_nc, alpha);
+      }
+
+      // Release all the sub blocks A'_i of A' for the current thread,
+      // i.e., we simply decrement the number of users by 1
+      for(Index i=0; i<threads; ++i)
+        #pragma omp atomic
+        info[i].users -= 1;
+    }
+  }
+  else
+#endif // EIGEN_HAS_OPENMP
+  {
+    EIGEN_UNUSED_VARIABLE(info);
+
+    // this is the sequential version!
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*nc;
+
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
+
+    const bool pack_rhs_once = mc!=rows && kc==depth && nc==cols;
+
+    // For each horizontal panel of the rhs, and corresponding panel of the lhs...
+    for(Index i2=0; i2<rows; i2+=mc)
+    {
+      const Index actual_mc = (std::min)(i2+mc,rows)-i2;
+
+      for(Index k2=0; k2<depth; k2+=kc)
+      {
+        const Index actual_kc = (std::min)(k2+kc,depth)-k2;
+
+        // OK, here we have selected one horizontal panel of rhs and one vertical panel of lhs.
+        // => Pack lhs's panel into a sequential chunk of memory (L2/L3 caching)
+        // Note that this panel will be read as many times as the number of blocks in the rhs's
+        // horizontal panel which is, in practice, a very low number.
+        pack_lhs(blockA, lhs.getSubMapper(i2,k2), actual_kc, actual_mc);
+
+        // For each kc x nc block of the rhs's horizontal panel...
+        for(Index j2=0; j2<cols; j2+=nc)
+        {
+          const Index actual_nc = (std::min)(j2+nc,cols)-j2;
+
+          // We pack the rhs's block into a sequential chunk of memory (L2 caching)
+          // Note that this block will be read a very high number of times, which is equal to the number of
+          // micro horizontal panel of the large rhs's panel (e.g., rows/12 times).
+          if((!pack_rhs_once) || i2==0)
+            pack_rhs(blockB, rhs.getSubMapper(k2,j2), actual_kc, actual_nc);
+
+          // Everything is packed, we can now call the panel * block kernel:
+          gebp(res.getSubMapper(i2, j2), blockA, blockB, actual_mc, actual_kc, actual_nc, alpha);
+        }
+      }
+    }
+  }
+}
+
+};
+
+/*********************************************************************************
+*  Specialization of generic_product_impl for "large" GEMM, i.e.,
+*  implementation of the high level wrapper to general_matrix_matrix_product
+**********************************************************************************/
+
+template<typename Scalar, typename Index, typename Gemm, typename Lhs, typename Rhs, typename Dest, typename BlockingType>
+struct gemm_functor
+{
+  gemm_functor(const Lhs& lhs, const Rhs& rhs, Dest& dest, const Scalar& actualAlpha, BlockingType& blocking)
+    : m_lhs(lhs), m_rhs(rhs), m_dest(dest), m_actualAlpha(actualAlpha), m_blocking(blocking)
+  {}
+
+  void initParallelSession(Index num_threads) const
+  {
+    m_blocking.initParallel(m_lhs.rows(), m_rhs.cols(), m_lhs.cols(), num_threads);
+    m_blocking.allocateA();
+  }
+
+  void operator() (Index row, Index rows, Index col=0, Index cols=-1, GemmParallelInfo<Index>* info=0) const
+  {
+    if(cols==-1)
+      cols = m_rhs.cols();
+
+    Gemm::run(rows, cols, m_lhs.cols(),
+              &m_lhs.coeffRef(row,0), m_lhs.outerStride(),
+              &m_rhs.coeffRef(0,col), m_rhs.outerStride(),
+              (Scalar*)&(m_dest.coeffRef(row,col)), m_dest.outerStride(),
+              m_actualAlpha, m_blocking, info);
+  }
+
+  typedef typename Gemm::Traits Traits;
+
+  protected:
+    const Lhs& m_lhs;
+    const Rhs& m_rhs;
+    Dest& m_dest;
+    Scalar m_actualAlpha;
+    BlockingType& m_blocking;
+};
+
+template<int StorageOrder, typename LhsScalar, typename RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor=1,
+bool FiniteAtCompileTime = MaxRows!=Dynamic && MaxCols!=Dynamic && MaxDepth != Dynamic> class gemm_blocking_space;
+
+template<typename _LhsScalar, typename _RhsScalar>
+class level3_blocking
+{
+    typedef _LhsScalar LhsScalar;
+    typedef _RhsScalar RhsScalar;
+
+  protected:
+    LhsScalar* m_blockA;
+    RhsScalar* m_blockB;
+
+    Index m_mc;
+    Index m_nc;
+    Index m_kc;
+
+  public:
+
+    level3_blocking()
+      : m_blockA(0), m_blockB(0), m_mc(0), m_nc(0), m_kc(0)
+    {}
+
+    inline Index mc() const { return m_mc; }
+    inline Index nc() const { return m_nc; }
+    inline Index kc() const { return m_kc; }
+
+    inline LhsScalar* blockA() { return m_blockA; }
+    inline RhsScalar* blockB() { return m_blockB; }
+};
+
+template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
+class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, true /* == FiniteAtCompileTime */>
+  : public level3_blocking<
+      typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
+      typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
+{
+    enum {
+      Transpose = StorageOrder==RowMajor,
+      ActualRows = Transpose ? MaxCols : MaxRows,
+      ActualCols = Transpose ? MaxRows : MaxCols
+    };
+    typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
+    typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+    enum {
+      SizeA = ActualRows * MaxDepth,
+      SizeB = ActualCols * MaxDepth
+    };
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
+    EIGEN_ALIGN_MAX LhsScalar m_staticA[SizeA];
+    EIGEN_ALIGN_MAX RhsScalar m_staticB[SizeB];
+#else
+    EIGEN_ALIGN_MAX char m_staticA[SizeA * sizeof(LhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
+    EIGEN_ALIGN_MAX char m_staticB[SizeB * sizeof(RhsScalar) + EIGEN_DEFAULT_ALIGN_BYTES-1];
+#endif
+
+  public:
+
+    gemm_blocking_space(Index /*rows*/, Index /*cols*/, Index /*depth*/, Index /*num_threads*/, bool /*full_rows = false*/)
+    {
+      this->m_mc = ActualRows;
+      this->m_nc = ActualCols;
+      this->m_kc = MaxDepth;
+#if EIGEN_MAX_STATIC_ALIGN_BYTES >= EIGEN_DEFAULT_ALIGN_BYTES
+      this->m_blockA = m_staticA;
+      this->m_blockB = m_staticB;
+#else
+      this->m_blockA = reinterpret_cast<LhsScalar*>((internal::UIntPtr(m_staticA) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
+      this->m_blockB = reinterpret_cast<RhsScalar*>((internal::UIntPtr(m_staticB) + (EIGEN_DEFAULT_ALIGN_BYTES-1)) & ~std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1));
+#endif
+    }
+
+    void initParallel(Index, Index, Index, Index)
+    {}
+
+    inline void allocateA() {}
+    inline void allocateB() {}
+    inline void allocateAll() {}
+};
+
+template<int StorageOrder, typename _LhsScalar, typename _RhsScalar, int MaxRows, int MaxCols, int MaxDepth, int KcFactor>
+class gemm_blocking_space<StorageOrder,_LhsScalar,_RhsScalar,MaxRows, MaxCols, MaxDepth, KcFactor, false>
+  : public level3_blocking<
+      typename conditional<StorageOrder==RowMajor,_RhsScalar,_LhsScalar>::type,
+      typename conditional<StorageOrder==RowMajor,_LhsScalar,_RhsScalar>::type>
+{
+    enum {
+      Transpose = StorageOrder==RowMajor
+    };
+    typedef typename conditional<Transpose,_RhsScalar,_LhsScalar>::type LhsScalar;
+    typedef typename conditional<Transpose,_LhsScalar,_RhsScalar>::type RhsScalar;
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+    Index m_sizeA;
+    Index m_sizeB;
+
+  public:
+
+    gemm_blocking_space(Index rows, Index cols, Index depth, Index num_threads, bool l3_blocking)
+    {
+      this->m_mc = Transpose ? cols : rows;
+      this->m_nc = Transpose ? rows : cols;
+      this->m_kc = depth;
+
+      if(l3_blocking)
+      {
+        computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, this->m_nc, num_threads);
+      }
+      else  // no l3 blocking
+      {
+        Index n = this->m_nc;
+        computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, this->m_mc, n, num_threads);
+      }
+
+      m_sizeA = this->m_mc * this->m_kc;
+      m_sizeB = this->m_kc * this->m_nc;
+    }
+
+    void initParallel(Index rows, Index cols, Index depth, Index num_threads)
+    {
+      this->m_mc = Transpose ? cols : rows;
+      this->m_nc = Transpose ? rows : cols;
+      this->m_kc = depth;
+
+      eigen_internal_assert(this->m_blockA==0 && this->m_blockB==0);
+      Index m = this->m_mc;
+      computeProductBlockingSizes<LhsScalar,RhsScalar,KcFactor>(this->m_kc, m, this->m_nc, num_threads);
+      m_sizeA = this->m_mc * this->m_kc;
+      m_sizeB = this->m_kc * this->m_nc;
+    }
+
+    void allocateA()
+    {
+      if(this->m_blockA==0)
+        this->m_blockA = aligned_new<LhsScalar>(m_sizeA);
+    }
+
+    void allocateB()
+    {
+      if(this->m_blockB==0)
+        this->m_blockB = aligned_new<RhsScalar>(m_sizeB);
+    }
+
+    void allocateAll()
+    {
+      allocateA();
+      allocateB();
+    }
+
+    ~gemm_blocking_space()
+    {
+      aligned_delete(this->m_blockA, m_sizeA);
+      aligned_delete(this->m_blockB, m_sizeB);
+    }
+};
+
+} // end namespace internal
+
+namespace internal {
+
+template<typename Lhs, typename Rhs>
+struct generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,GemmProduct> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  typedef typename Lhs::Scalar LhsScalar;
+  typedef typename Rhs::Scalar RhsScalar;
+
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+  enum {
+    MaxDepthAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(Lhs::MaxColsAtCompileTime,Rhs::MaxRowsAtCompileTime)
+  };
+
+  typedef generic_product_impl<Lhs,Rhs,DenseShape,DenseShape,CoeffBasedProductMode> lazyproduct;
+
+  template<typename Dst>
+  static void evalTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::evalTo(dst, lhs, rhs);
+    else
+    {
+      dst.setZero();
+      scaleAndAddTo(dst, lhs, rhs, Scalar(1));
+    }
+  }
+
+  template<typename Dst>
+  static void addTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::addTo(dst, lhs, rhs);
+    else
+      scaleAndAddTo(dst,lhs, rhs, Scalar(1));
+  }
+
+  template<typename Dst>
+  static void subTo(Dst& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    if((rhs.rows()+dst.rows()+dst.cols())<20 && rhs.rows()>0)
+      lazyproduct::subTo(dst, lhs, rhs);
+    else
+      scaleAndAddTo(dst, lhs, rhs, Scalar(-1));
+  }
+
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& a_lhs, const Rhs& a_rhs, const Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
+    if(a_lhs.cols()==0 || a_lhs.rows()==0 || a_rhs.cols()==0)
+      return;
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,LhsScalar,RhsScalar,
+            Dest::MaxRowsAtCompileTime,Dest::MaxColsAtCompileTime,MaxDepthAtCompileTime> BlockingType;
+
+    typedef internal::gemm_functor<
+      Scalar, Index,
+      internal::general_matrix_matrix_product<
+        Index,
+        LhsScalar, (ActualLhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(LhsBlasTraits::NeedToConjugate),
+        RhsScalar, (ActualRhsTypeCleaned::Flags&RowMajorBit) ? RowMajor : ColMajor, bool(RhsBlasTraits::NeedToConjugate),
+        (Dest::Flags&RowMajorBit) ? RowMajor : ColMajor>,
+      ActualLhsTypeCleaned, ActualRhsTypeCleaned, Dest, BlockingType> GemmFunctor;
+
+    BlockingType blocking(dst.rows(), dst.cols(), lhs.cols(), 1, true);
+    internal::parallelize_gemm<(Dest::MaxRowsAtCompileTime>32 || Dest::MaxRowsAtCompileTime==Dynamic)>
+        (GemmFunctor(lhs, rhs, dst, actualAlpha, blocking), a_lhs.rows(), a_rhs.cols(), a_lhs.cols(), Dest::Flags&RowMajorBit);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
new file mode 100644
index 0000000..e844e37
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular.h
@@ -0,0 +1,311 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H
+
+namespace Eigen { 
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update;
+
+namespace internal {
+
+/**********************************************************************
+* This file implements a general A * B product while
+* evaluating only one triangular part of the product.
+* This is a more general version of self adjoint product (C += A A^T)
+* as the level 3 SYRK Blas routine.
+**********************************************************************/
+
+// forward declarations (defined at the end of this file)
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+struct tribb_kernel;
+  
+/* Optimized matrix-matrix product evaluating only one triangular half */
+template <typename Index,
+          typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+                              int ResStorageOrder, int  UpLo, int Version = Specialized>
+struct general_matrix_matrix_triangular_product;
+
+// as usual if the result is row major => we transpose the product
+template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,RowMajor,UpLo,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* lhs, Index lhsStride,
+                                      const RhsScalar* rhs, Index rhsStride, ResScalar* res, Index resStride,
+                                      const ResScalar& alpha, level3_blocking<RhsScalar,LhsScalar>& blocking)
+  {
+    general_matrix_matrix_triangular_product<Index,
+        RhsScalar, RhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateRhs,
+        LhsScalar, LhsStorageOrder==RowMajor ? ColMajor : RowMajor, ConjugateLhs,
+        ColMajor, UpLo==Lower?Upper:Lower>
+      ::run(size,depth,rhs,rhsStride,lhs,lhsStride,res,resStride,alpha,blocking);
+  }
+};
+
+template <typename Index, typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+                          typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs, int  UpLo, int Version>
+struct general_matrix_matrix_triangular_product<Index,LhsScalar,LhsStorageOrder,ConjugateLhs,RhsScalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const LhsScalar* _lhs, Index lhsStride,
+                                      const RhsScalar* _rhs, Index rhsStride, ResScalar* _res, Index resStride,
+                                      const ResScalar& alpha, level3_blocking<LhsScalar,RhsScalar>& blocking)
+  {
+    typedef gebp_traits<LhsScalar,RhsScalar> Traits;
+
+    typedef const_blas_data_mapper<LhsScalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();
+    Index mc = (std::min)(size,blocking.mc());
+
+    // !!! mc must be a multiple of nr:
+    if(mc > Traits::nr)
+      mc = (mc/Traits::nr)*Traits::nr;
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*size;
+
+    ei_declare_aligned_stack_constructed_variable(LhsScalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(RhsScalar, blockB, sizeB, blocking.blockB());
+
+    gemm_pack_lhs<LhsScalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<RhsScalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp;
+    tribb_kernel<LhsScalar, RhsScalar, Index, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs, UpLo> sybb;
+
+    for(Index k2=0; k2<depth; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,depth)-k2;
+
+      // note that the actual rhs is the transpose/adjoint of mat
+      pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, size);
+
+      for(Index i2=0; i2<size; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,size)-i2;
+
+        pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        // the selected actual_mc * size panel of res is split into three different part:
+        //  1 - before the diagonal => processed with gebp or skipped
+        //  2 - the actual_mc x actual_mc symmetric block => processed with a special kernel
+        //  3 - after the diagonal => processed with gebp or skipped
+        if (UpLo==Lower)
+          gebp(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc,
+               (std::min)(size,i2), alpha, -1, -1, 0, 0);
+
+
+        sybb(_res+resStride*i2 + i2, resStride, blockA, blockB + actual_kc*i2, actual_mc, actual_kc, alpha);
+
+        if (UpLo==Upper)
+        {
+          Index j2 = i2+actual_mc;
+          gebp(res.getSubMapper(i2, j2), blockA, blockB+actual_kc*j2, actual_mc,
+               actual_kc, (std::max)(Index(0), size-j2), alpha, -1, -1, 0, 0);
+        }
+      }
+    }
+  }
+};
+
+// Optimized packed Block * packed Block product kernel evaluating only one given triangular part
+// This kernel is built on top of the gebp kernel:
+// - the current destination block is processed per panel of actual_mc x BlockSize
+//   where BlockSize is set to the minimal value allowing gebp to be as fast as possible
+// - then, as usual, each panel is split into three parts along the diagonal,
+//   the sub blocks above and below the diagonal are processed as usual,
+//   while the triangular block overlapping the diagonal is evaluated into a
+//   small temporary buffer which is then accumulated into the result using a
+//   triangular traversal.
+template<typename LhsScalar, typename RhsScalar, typename Index, int mr, int nr, bool ConjLhs, bool ConjRhs, int UpLo>
+struct tribb_kernel
+{
+  typedef gebp_traits<LhsScalar,RhsScalar,ConjLhs,ConjRhs> Traits;
+  typedef typename Traits::ResScalar ResScalar;
+
+  enum {
+    BlockSize  = meta_least_common_multiple<EIGEN_PLAIN_ENUM_MAX(mr,nr),EIGEN_PLAIN_ENUM_MIN(mr,nr)>::ret
+  };
+  void operator()(ResScalar* _res, Index resStride, const LhsScalar* blockA, const RhsScalar* blockB, Index size, Index depth, const ResScalar& alpha)
+  {
+    typedef blas_data_mapper<ResScalar, Index, ColMajor> ResMapper;
+    ResMapper res(_res, resStride);
+    gebp_kernel<LhsScalar, RhsScalar, Index, ResMapper, mr, nr, ConjLhs, ConjRhs> gebp_kernel;
+
+    Matrix<ResScalar,BlockSize,BlockSize,ColMajor> buffer((internal::constructor_without_unaligned_array_assert()));
+
+    // let's process the block per panel of actual_mc x BlockSize,
+    // again, each is split into three parts, etc.
+    for (Index j=0; j<size; j+=BlockSize)
+    {
+      Index actualBlockSize = std::min<Index>(BlockSize,size - j);
+      const RhsScalar* actual_b = blockB+j*depth;
+
+      if(UpLo==Upper)
+        gebp_kernel(res.getSubMapper(0, j), blockA, actual_b, j, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0);
+
+      // selfadjoint micro block
+      {
+        Index i = j;
+        buffer.setZero();
+        // 1 - apply the kernel on the temporary buffer
+        gebp_kernel(ResMapper(buffer.data(), BlockSize), blockA+depth*i, actual_b, actualBlockSize, depth, actualBlockSize, alpha,
+                    -1, -1, 0, 0);
+        // 2 - triangular accumulation
+        for(Index j1=0; j1<actualBlockSize; ++j1)
+        {
+          ResScalar* r = &res(i, j + j1);
+          for(Index i1=UpLo==Lower ? j1 : 0;
+              UpLo==Lower ? i1<actualBlockSize : i1<=j1; ++i1)
+            r[i1] += buffer(i1,j1);
+        }
+      }
+
+      if(UpLo==Lower)
+      {
+        Index i = j+actualBlockSize;
+        gebp_kernel(res.getSubMapper(i, j), blockA+depth*i, actual_b, size-i, 
+                    depth, actualBlockSize, alpha, -1, -1, 0, 0);
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+// high level API
+
+template<typename MatrixType, typename ProductType, int UpLo, bool IsOuterProduct>
+struct general_product_to_triangular_selector;
+
+
+template<typename MatrixType, typename ProductType, int UpLo>
+struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,true>
+{
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    
+    typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
+    typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+    typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    
+    typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
+    typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+    typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
+
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
+    enum {
+      StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+      UseLhsDirectly = _ActualLhs::InnerStrideAtCompileTime==1,
+      UseRhsDirectly = _ActualRhs::InnerStrideAtCompileTime==1
+    };
+    
+    internal::gemv_static_vector_if<Scalar,Lhs::SizeAtCompileTime,Lhs::MaxSizeAtCompileTime,!UseLhsDirectly> static_lhs;
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualLhsPtr, actualLhs.size(),
+      (UseLhsDirectly ? const_cast<Scalar*>(actualLhs.data()) : static_lhs.data()));
+    if(!UseLhsDirectly) Map<typename _ActualLhs::PlainObject>(actualLhsPtr, actualLhs.size()) = actualLhs;
+    
+    internal::gemv_static_vector_if<Scalar,Rhs::SizeAtCompileTime,Rhs::MaxSizeAtCompileTime,!UseRhsDirectly> static_rhs;
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualRhsPtr, actualRhs.size(),
+      (UseRhsDirectly ? const_cast<Scalar*>(actualRhs.data()) : static_rhs.data()));
+    if(!UseRhsDirectly) Map<typename _ActualRhs::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    
+    
+    selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
+                              LhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex,
+                              RhsBlasTraits::NeedToConjugate && NumTraits<Scalar>::IsComplex>
+          ::run(actualLhs.size(), mat.data(), mat.outerStride(), actualLhsPtr, actualRhsPtr, actualAlpha);
+  }
+};
+
+template<typename MatrixType, typename ProductType, int UpLo>
+struct general_product_to_triangular_selector<MatrixType,ProductType,UpLo,false>
+{
+  static void run(MatrixType& mat, const ProductType& prod, const typename MatrixType::Scalar& alpha, bool beta)
+  {
+    typedef typename internal::remove_all<typename ProductType::LhsNested>::type Lhs;
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhs;
+    typedef typename internal::remove_all<ActualLhs>::type _ActualLhs;
+    typename internal::add_const_on_value_type<ActualLhs>::type actualLhs = LhsBlasTraits::extract(prod.lhs());
+    
+    typedef typename internal::remove_all<typename ProductType::RhsNested>::type Rhs;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhs;
+    typedef typename internal::remove_all<ActualRhs>::type _ActualRhs;
+    typename internal::add_const_on_value_type<ActualRhs>::type actualRhs = RhsBlasTraits::extract(prod.rhs());
+
+    typename ProductType::Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs().derived()) * RhsBlasTraits::extractScalarFactor(prod.rhs().derived());
+
+    if(!beta)
+      mat.template triangularView<UpLo>().setZero();
+
+    enum {
+      IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
+      LhsIsRowMajor = _ActualLhs::Flags&RowMajorBit ? 1 : 0,
+      RhsIsRowMajor = _ActualRhs::Flags&RowMajorBit ? 1 : 0,
+      SkipDiag = (UpLo&(UnitDiag|ZeroDiag))!=0
+    };
+
+    Index size = mat.cols();
+    if(SkipDiag)
+      size--;
+    Index depth = actualLhs.cols();
+
+    typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,typename Lhs::Scalar,typename Rhs::Scalar,
+          MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualRhs::MaxColsAtCompileTime> BlockingType;
+
+    BlockingType blocking(size, size, depth, 1, false);
+
+    internal::general_matrix_matrix_triangular_product<Index,
+      typename Lhs::Scalar, LhsIsRowMajor ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      typename Rhs::Scalar, RhsIsRowMajor ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      IsRowMajor ? RowMajor : ColMajor, UpLo&(Lower|Upper)>
+      ::run(size, depth,
+            &actualLhs.coeffRef(SkipDiag&&(UpLo&Lower)==Lower ? 1 : 0,0), actualLhs.outerStride(),
+            &actualRhs.coeffRef(0,SkipDiag&&(UpLo&Upper)==Upper ? 1 : 0), actualRhs.outerStride(),
+            mat.data() + (SkipDiag ? (bool(IsRowMajor) != ((UpLo&Lower)==Lower) ? 1 : mat.outerStride() ) : 0), mat.outerStride(), actualAlpha, blocking);
+  }
+};
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename ProductType>
+TriangularView<MatrixType,UpLo>& TriangularViewImpl<MatrixType,UpLo,Dense>::_assignProduct(const ProductType& prod, const Scalar& alpha, bool beta)
+{
+  EIGEN_STATIC_ASSERT((UpLo&UnitDiag)==0, WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED);
+  eigen_assert(derived().nestedExpression().rows() == prod.rows() && derived().cols() == prod.cols());
+  
+  general_product_to_triangular_selector<MatrixType, ProductType, UpLo, internal::traits<ProductType>::InnerSize==1>::run(derived().nestedExpression().const_cast_derived(), prod, alpha, beta);
+  
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
new file mode 100644
index 0000000..41e18ff
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrixTriangular_BLAS.h
@@ -0,0 +1,141 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Level 3 BLAS SYRK/HERK implementation.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
+
+namespace Eigen {
+
+namespace internal {
+
+template <typename Index, typename Scalar, int AStorageOrder, bool ConjugateA, int ResStorageOrder, int  UpLo>
+struct general_matrix_matrix_rankupdate :
+       general_matrix_matrix_triangular_product<
+         Index,Scalar,AStorageOrder,ConjugateA,Scalar,AStorageOrder,ConjugateA,ResStorageOrder,UpLo,BuiltIn> {};
+
+
+// try to go to BLAS specialization
+#define EIGEN_BLAS_RANKUPDATE_SPECIALIZE(Scalar) \
+template <typename Index, int LhsStorageOrder, bool ConjugateLhs, \
+                          int RhsStorageOrder, bool ConjugateRhs, int  UpLo> \
+struct general_matrix_matrix_triangular_product<Index,Scalar,LhsStorageOrder,ConjugateLhs, \
+               Scalar,RhsStorageOrder,ConjugateRhs,ColMajor,UpLo,Specialized> { \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const Scalar* lhs, Index lhsStride, \
+                          const Scalar* rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar, Scalar>& blocking) \
+  { \
+    if ( lhs==rhs && ((UpLo&(Lower|Upper)==UpLo)) ) { \
+      general_matrix_matrix_rankupdate<Index,Scalar,LhsStorageOrder,ConjugateLhs,ColMajor,UpLo> \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
+    } else { \
+      general_matrix_matrix_triangular_product<Index, \
+        Scalar, LhsStorageOrder, ConjugateLhs, \
+        Scalar, RhsStorageOrder, ConjugateRhs, \
+        ColMajor, UpLo, BuiltIn> \
+      ::run(size,depth,lhs,lhsStride,rhs,rhsStride,res,resStride,alpha,blocking); \
+    } \
+  } \
+};
+
+EIGEN_BLAS_RANKUPDATE_SPECIALIZE(double)
+EIGEN_BLAS_RANKUPDATE_SPECIALIZE(float)
+// TODO handle complex cases
+// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(dcomplex)
+// EIGEN_BLAS_RANKUPDATE_SPECIALIZE(scomplex)
+
+// SYRK for float/double
+#define EIGEN_BLAS_RANKUPDATE_R(EIGTYPE, BLASTYPE, BLASFUNC) \
+template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
+struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
+  enum { \
+    IsLower = (UpLo&Lower) == Lower, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((AStorageOrder==ColMajor) && ConjugateA) ? 1 : 0 \
+  }; \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
+                          const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+  /* typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs;*/ \
+\
+   BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'T':'N'); \
+   EIGTYPE beta(1); \
+   BLASFUNC(&uplo, &trans, &n, &k, &numext::real_ref(alpha), lhs, &lda, &numext::real_ref(beta), res, &ldc); \
+  } \
+};
+
+// HERK for complex data
+#define EIGEN_BLAS_RANKUPDATE_C(EIGTYPE, BLASTYPE, RTYPE, BLASFUNC) \
+template <typename Index, int AStorageOrder, bool ConjugateA, int  UpLo> \
+struct general_matrix_matrix_rankupdate<Index,EIGTYPE,AStorageOrder,ConjugateA,ColMajor,UpLo> { \
+  enum { \
+    IsLower = (UpLo&Lower) == Lower, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = (((AStorageOrder==ColMajor) && ConjugateA) || ((AStorageOrder==RowMajor) && !ConjugateA)) ? 1 : 0 \
+  }; \
+  static EIGEN_STRONG_INLINE void run(Index size, Index depth,const EIGTYPE* lhs, Index lhsStride, \
+                          const EIGTYPE* /*rhs*/, Index /*rhsStride*/, EIGTYPE* res, Index resStride, EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, AStorageOrder> MatrixType; \
+\
+   BlasIndex lda=convert_index<BlasIndex>(lhsStride), ldc=convert_index<BlasIndex>(resStride), n=convert_index<BlasIndex>(size), k=convert_index<BlasIndex>(depth); \
+   char uplo=((IsLower) ? 'L' : 'U'), trans=((AStorageOrder==RowMajor) ? 'C':'N'); \
+   RTYPE alpha_, beta_; \
+   const EIGTYPE* a_ptr; \
+\
+   alpha_ = alpha.real(); \
+   beta_ = 1.0; \
+/* Copy with conjugation in some cases*/ \
+   MatrixType a; \
+   if (conjA) { \
+     Map<const MatrixType, 0, OuterStride<> > mapA(lhs,n,k,OuterStride<>(lhsStride)); \
+     a = mapA.conjugate(); \
+     lda = a.outerStride(); \
+     a_ptr = a.data(); \
+   } else a_ptr=lhs; \
+   BLASFUNC(&uplo, &trans, &n, &k, &alpha_, (BLASTYPE*)a_ptr, &lda, &beta_, (BLASTYPE*)res, &ldc); \
+  } \
+};
+
+
+EIGEN_BLAS_RANKUPDATE_R(double, double, dsyrk_)
+EIGEN_BLAS_RANKUPDATE_R(float,  float,  ssyrk_)
+
+// TODO hanlde complex cases
+// EIGEN_BLAS_RANKUPDATE_C(dcomplex, double, double, zherk_)
+// EIGEN_BLAS_RANKUPDATE_C(scomplex, float,  float, cherk_)
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_TRIANGULAR_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
new file mode 100644
index 0000000..7a3bdbf
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixMatrix_BLAS.h
@@ -0,0 +1,115 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   General matrix-matrix product functionality based on ?GEMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
+#define EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements general matrix-matrix multiplication using BLAS
+* gemm function via partial specialization of
+* general_matrix_matrix_product::run(..) method for float, double,
+* std::complex<float> and std::complex<double> types
+**********************************************************************/
+
+// gemm specialization
+
+#define GEMM_SPECIALIZATION(EIGTYPE, EIGPREFIX, BLASTYPE, BLASPREFIX) \
+template< \
+  typename Index, \
+  int LhsStorageOrder, bool ConjugateLhs, \
+  int RhsStorageOrder, bool ConjugateRhs> \
+struct general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+typedef gebp_traits<EIGTYPE,EIGTYPE> Traits; \
+\
+static void run(Index rows, Index cols, Index depth, \
+  const EIGTYPE* _lhs, Index lhsStride, \
+  const EIGTYPE* _rhs, Index rhsStride, \
+  EIGTYPE* res, Index resStride, \
+  EIGTYPE alpha, \
+  level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/, \
+  GemmParallelInfo<Index>* /*info = 0*/) \
+{ \
+  using std::conj; \
+\
+  char transa, transb; \
+  BlasIndex m, n, k, lda, ldb, ldc; \
+  const EIGTYPE *a, *b; \
+  EIGTYPE beta(1); \
+  MatrixX##EIGPREFIX a_tmp, b_tmp; \
+\
+/* Set transpose options */ \
+  transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
+  transb = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set m, n, k */ \
+  m = convert_index<BlasIndex>(rows);  \
+  n = convert_index<BlasIndex>(cols);  \
+  k = convert_index<BlasIndex>(depth); \
+\
+/* Set lda, ldb, ldc */ \
+  lda = convert_index<BlasIndex>(lhsStride); \
+  ldb = convert_index<BlasIndex>(rhsStride); \
+  ldc = convert_index<BlasIndex>(resStride); \
+\
+/* Set a, b, c */ \
+  if ((LhsStorageOrder==ColMajor) && (ConjugateLhs)) { \
+    Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,k,OuterStride<>(lhsStride)); \
+    a_tmp = lhs.conjugate(); \
+    a = a_tmp.data(); \
+    lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+  } else a = _lhs; \
+\
+  if ((RhsStorageOrder==ColMajor) && (ConjugateRhs)) { \
+    Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,k,n,OuterStride<>(rhsStride)); \
+    b_tmp = rhs.conjugate(); \
+    b = b_tmp.data(); \
+    ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+  } else b = _rhs; \
+\
+  BLASPREFIX##gemm_(&transa, &transb, &m, &n, &k, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+}};
+
+GEMM_SPECIALIZATION(double,   d,  double, d)
+GEMM_SPECIALIZATION(float,    f,  float,  s)
+GEMM_SPECIALIZATION(dcomplex, cd, double, z)
+GEMM_SPECIALIZATION(scomplex, cf, float,  c)
+
+} // end namespase internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_MATRIX_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h
new file mode 100644
index 0000000..3c1a7fc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector.h
@@ -0,0 +1,619 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERAL_MATRIX_VECTOR_H
+#define EIGEN_GENERAL_MATRIX_VECTOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+/* Optimized col-major matrix * vector product:
+ * This algorithm processes 4 columns at onces that allows to both reduce
+ * the number of load/stores of the result by a factor 4 and to reduce
+ * the instruction dependency. Moreover, we know that all bands have the
+ * same alignment pattern.
+ *
+ * Mixing type logic: C += alpha * A * B
+ *  |  A  |  B  |alpha| comments
+ *  |real |cplx |cplx | no vectorization
+ *  |real |cplx |real | alpha is converted to a cplx when calling the run function, no vectorization
+ *  |cplx |real |cplx | invalid, the caller has to do tmp: = A * B; C += alpha*tmp
+ *  |cplx |real |real | optimal case, vectorization possible via real-cplx mul
+ *
+ * Accesses to the matrix coefficients follow the following logic:
+ *
+ * - if all columns have the same alignment then
+ *   - if the columns have the same alignment as the result vector, then easy! (-> AllAligned case)
+ *   - otherwise perform unaligned loads only (-> NoneAligned case)
+ * - otherwise
+ *   - if even columns have the same alignment then
+ *     // odd columns are guaranteed to have the same alignment too
+ *     - if even or odd columns have the same alignment as the result, then
+ *       // for a register size of 2 scalars, this is guarantee to be the case (e.g., SSE with double)
+ *       - perform half aligned and half unaligned loads (-> EvenAligned case)
+ *     - otherwise perform unaligned loads only (-> NoneAligned case)
+ *   - otherwise, if the register size is 4 scalars (e.g., SSE with float) then
+ *     - one over 4 consecutive columns is guaranteed to be aligned with the result vector,
+ *       perform simple aligned loads for this column and aligned loads plus re-alignment for the other. (-> FirstAligned case)
+ *       // this re-alignment is done by the palign function implemented for SSE in Eigen/src/Core/arch/SSE/PacketMath.h
+ *   - otherwise,
+ *     // if we get here, this means the register size is greater than 4 (e.g., AVX with floats),
+ *     // we currently fall back to the NoneAligned case
+ *
+ * The same reasoning apply for the transposed case.
+ *
+ * The last case (PacketSize>4) could probably be improved by generalizing the FirstAligned case, but since we do not support AVX yet...
+ * One might also wonder why in the EvenAligned case we perform unaligned loads instead of using the aligned-loads plus re-alignment
+ * strategy as in the FirstAligned case. The reason is that we observed that unaligned loads on a 8 byte boundary are not too slow
+ * compared to unaligned loads on a 4 byte boundary.
+ *
+ */
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+
+enum {
+  Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable
+              && int(packet_traits<LhsScalar>::size)==int(packet_traits<RhsScalar>::size),
+  LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+  RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+  ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1
+};
+
+typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
+typedef typename packet_traits<RhsScalar>::type  _RhsPacket;
+typedef typename packet_traits<ResScalar>::type  _ResPacket;
+
+typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
+typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
+typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+
+EIGEN_DONT_INLINE static void run(
+  Index rows, Index cols,
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
+  RhsScalar alpha);
+};
+
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
+  Index rows, Index cols,
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
+  RhsScalar alpha)
+{
+  EIGEN_UNUSED_VARIABLE(resIncr);
+  eigen_internal_assert(resIncr==1);
+  #ifdef _EIGEN_ACCUMULATE_PACKETS
+  #error _EIGEN_ACCUMULATE_PACKETS has already been defined
+  #endif
+  #define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) \
+    pstore(&res[j], \
+      padd(pload<ResPacket>(&res[j]), \
+        padd( \
+      padd(pcj.pmul(lhs0.template load<LhsPacket, Alignment0>(j),    ptmp0), \
+      pcj.pmul(lhs1.template load<LhsPacket, Alignment13>(j),   ptmp1)),   \
+      padd(pcj.pmul(lhs2.template load<LhsPacket, Alignment2>(j),    ptmp2), \
+      pcj.pmul(lhs3.template load<LhsPacket, Alignment13>(j),   ptmp3)) )))
+
+  typedef typename LhsMapper::VectorMapper LhsScalars;
+
+  conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
+  conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
+  if(ConjugateRhs)
+    alpha = numext::conj(alpha);
+
+  enum { AllAligned = 0, EvenAligned, FirstAligned, NoneAligned };
+  const Index columnsAtOnce = 4;
+  const Index peels = 2;
+  const Index LhsPacketAlignedMask = LhsPacketSize-1;
+  const Index ResPacketAlignedMask = ResPacketSize-1;
+//  const Index PeelAlignedMask = ResPacketSize*peels-1;
+  const Index size = rows;
+
+  const Index lhsStride = lhs.stride();
+
+  // How many coeffs of the result do we have to skip to be aligned.
+  // Here we assume data are at least aligned on the base scalar type.
+  Index alignedStart = internal::first_default_aligned(res,size);
+  Index alignedSize = ResPacketSize>1 ? alignedStart + ((size-alignedStart) & ~ResPacketAlignedMask) : 0;
+  const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1;
+
+  const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0;
+  Index alignmentPattern = alignmentStep==0 ? AllAligned
+                       : alignmentStep==(LhsPacketSize/2) ? EvenAligned
+                       : FirstAligned;
+
+  // we cannot assume the first element is aligned because of sub-matrices
+  const Index lhsAlignmentOffset = lhs.firstAligned(size);
+
+  // find how many columns do we have to skip to be aligned with the result (if possible)
+  Index skipColumns = 0;
+  // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
+  if( (lhsAlignmentOffset < 0) || (lhsAlignmentOffset == size) || (UIntPtr(res)%sizeof(ResScalar)) )
+  {
+    alignedSize = 0;
+    alignedStart = 0;
+    alignmentPattern = NoneAligned;
+  }
+  else if(LhsPacketSize > 4)
+  {
+    // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
+    // Currently, it seems to be better to perform unaligned loads anyway
+    alignmentPattern = NoneAligned;
+  }
+  else if (LhsPacketSize>1)
+  {
+  //    eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0 || size<LhsPacketSize);
+
+    while (skipColumns<LhsPacketSize &&
+          alignedStart != ((lhsAlignmentOffset + alignmentStep*skipColumns)%LhsPacketSize))
+      ++skipColumns;
+    if (skipColumns==LhsPacketSize)
+    {
+      // nothing can be aligned, no need to skip any column
+      alignmentPattern = NoneAligned;
+      skipColumns = 0;
+    }
+    else
+    {
+      skipColumns = (std::min)(skipColumns,cols);
+      // note that the skiped columns are processed later.
+    }
+
+    /*    eigen_internal_assert(  (alignmentPattern==NoneAligned)
+                      || (skipColumns + columnsAtOnce >= cols)
+                      || LhsPacketSize > size
+                      || (size_t(firstLhs+alignedStart+lhsStride*skipColumns)%sizeof(LhsPacket))==0);*/
+  }
+  else if(Vectorizable)
+  {
+    alignedStart = 0;
+    alignedSize = size;
+    alignmentPattern = AllAligned;
+  }
+
+  const Index offset1 = (FirstAligned && alignmentStep==1)?3:1;
+  const Index offset3 = (FirstAligned && alignmentStep==1)?1:3;
+
+  Index columnBound = ((cols-skipColumns)/columnsAtOnce)*columnsAtOnce + skipColumns;
+  for (Index i=skipColumns; i<columnBound; i+=columnsAtOnce)
+  {
+    RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(i, 0)),
+              ptmp1 = pset1<RhsPacket>(alpha*rhs(i+offset1, 0)),
+              ptmp2 = pset1<RhsPacket>(alpha*rhs(i+2, 0)),
+              ptmp3 = pset1<RhsPacket>(alpha*rhs(i+offset3, 0));
+
+    // this helps a lot generating better binary code
+    const LhsScalars lhs0 = lhs.getVectorMapper(0, i+0),   lhs1 = lhs.getVectorMapper(0, i+offset1),
+                     lhs2 = lhs.getVectorMapper(0, i+2),   lhs3 = lhs.getVectorMapper(0, i+offset3);
+
+    if (Vectorizable)
+    {
+      /* explicit vectorization */
+      // process initial unaligned coeffs
+      for (Index j=0; j<alignedStart; ++j)
+      {
+        res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]);
+        res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]);
+        res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]);
+        res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]);
+      }
+
+      if (alignedSize>alignedStart)
+      {
+        switch(alignmentPattern)
+        {
+          case AllAligned:
+            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned);
+            break;
+          case EvenAligned:
+            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
+            break;
+          case FirstAligned:
+          {
+            Index j = alignedStart;
+            if(peels>1)
+            {
+              LhsPacket A00, A01, A02, A03, A10, A11, A12, A13;
+              ResPacket T0, T1;
+
+              A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1);
+              A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2);
+              A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3);
+
+              for (; j<peeledSize; j+=peels*ResPacketSize)
+              {
+                A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize);  palign<1>(A01,A11);
+                A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize);  palign<2>(A02,A12);
+                A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize);  palign<3>(A03,A13);
+
+                A00 = lhs0.template load<LhsPacket, Aligned>(j);
+                A10 = lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize);
+                T0  = pcj.pmadd(A00, ptmp0, pload<ResPacket>(&res[j]));
+                T1  = pcj.pmadd(A10, ptmp0, pload<ResPacket>(&res[j+ResPacketSize]));
+
+                T0  = pcj.pmadd(A01, ptmp1, T0);
+                A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize);  palign<1>(A11,A01);
+                T0  = pcj.pmadd(A02, ptmp2, T0);
+                A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize);  palign<2>(A12,A02);
+                T0  = pcj.pmadd(A03, ptmp3, T0);
+                pstore(&res[j],T0);
+                A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize);  palign<3>(A13,A03);
+                T1  = pcj.pmadd(A11, ptmp1, T1);
+                T1  = pcj.pmadd(A12, ptmp2, T1);
+                T1  = pcj.pmadd(A13, ptmp3, T1);
+                pstore(&res[j+ResPacketSize],T1);
+              }
+            }
+            for (; j<alignedSize; j+=ResPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
+            break;
+          }
+          default:
+            for (Index j = alignedStart; j<alignedSize; j+=ResPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
+            break;
+        }
+      }
+    } // end explicit vectorization
+
+    /* process remaining coeffs (or all if there is no explicit vectorization) */
+    for (Index j=alignedSize; j<size; ++j)
+    {
+      res[j] = cj.pmadd(lhs0(j), pfirst(ptmp0), res[j]);
+      res[j] = cj.pmadd(lhs1(j), pfirst(ptmp1), res[j]);
+      res[j] = cj.pmadd(lhs2(j), pfirst(ptmp2), res[j]);
+      res[j] = cj.pmadd(lhs3(j), pfirst(ptmp3), res[j]);
+    }
+  }
+
+  // process remaining first and last columns (at most columnsAtOnce-1)
+  Index end = cols;
+  Index start = columnBound;
+  do
+  {
+    for (Index k=start; k<end; ++k)
+    {
+      RhsPacket ptmp0 = pset1<RhsPacket>(alpha*rhs(k, 0));
+      const LhsScalars lhs0 = lhs.getVectorMapper(0, k);
+
+      if (Vectorizable)
+      {
+        /* explicit vectorization */
+        // process first unaligned result's coeffs
+        for (Index j=0; j<alignedStart; ++j)
+          res[j] += cj.pmul(lhs0(j), pfirst(ptmp0));
+        // process aligned result's coeffs
+        if (lhs0.template aligned<LhsPacket>(alignedStart))
+          for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
+            pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(i), ptmp0, pload<ResPacket>(&res[i])));
+        else
+          for (Index i = alignedStart;i<alignedSize;i+=ResPacketSize)
+            pstore(&res[i], pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(i), ptmp0, pload<ResPacket>(&res[i])));
+      }
+
+      // process remaining scalars (or all if no explicit vectorization)
+      for (Index i=alignedSize; i<size; ++i)
+        res[i] += cj.pmul(lhs0(i), pfirst(ptmp0));
+    }
+    if (skipColumns)
+    {
+      start = 0;
+      end = skipColumns;
+      skipColumns = 0;
+    }
+    else
+      break;
+  } while(Vectorizable);
+  #undef _EIGEN_ACCUMULATE_PACKETS
+}
+
+/* Optimized row-major matrix * vector product:
+ * This algorithm processes 4 rows at onces that allows to both reduce
+ * the number of load/stores of the result by a factor 4 and to reduce
+ * the instruction dependency. Moreover, we know that all bands have the
+ * same alignment pattern.
+ *
+ * Mixing type logic:
+ *  - alpha is always a complex (or converted to a complex)
+ *  - no vectorization
+ */
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+struct general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>
+{
+typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+
+enum {
+  Vectorizable = packet_traits<LhsScalar>::Vectorizable && packet_traits<RhsScalar>::Vectorizable
+              && int(packet_traits<LhsScalar>::size)==int(packet_traits<RhsScalar>::size),
+  LhsPacketSize = Vectorizable ? packet_traits<LhsScalar>::size : 1,
+  RhsPacketSize = Vectorizable ? packet_traits<RhsScalar>::size : 1,
+  ResPacketSize = Vectorizable ? packet_traits<ResScalar>::size : 1
+};
+
+typedef typename packet_traits<LhsScalar>::type  _LhsPacket;
+typedef typename packet_traits<RhsScalar>::type  _RhsPacket;
+typedef typename packet_traits<ResScalar>::type  _ResPacket;
+
+typedef typename conditional<Vectorizable,_LhsPacket,LhsScalar>::type LhsPacket;
+typedef typename conditional<Vectorizable,_RhsPacket,RhsScalar>::type RhsPacket;
+typedef typename conditional<Vectorizable,_ResPacket,ResScalar>::type ResPacket;
+
+EIGEN_DONT_INLINE static void run(
+  Index rows, Index cols,
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+        ResScalar* res, Index resIncr,
+  ResScalar alpha);
+};
+
+template<typename Index, typename LhsScalar, typename LhsMapper, bool ConjugateLhs, typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjugateLhs,RhsScalar,RhsMapper,ConjugateRhs,Version>::run(
+  Index rows, Index cols,
+  const LhsMapper& lhs,
+  const RhsMapper& rhs,
+  ResScalar* res, Index resIncr,
+  ResScalar alpha)
+{
+  eigen_internal_assert(rhs.stride()==1);
+
+  #ifdef _EIGEN_ACCUMULATE_PACKETS
+  #error _EIGEN_ACCUMULATE_PACKETS has already been defined
+  #endif
+
+  #define _EIGEN_ACCUMULATE_PACKETS(Alignment0,Alignment13,Alignment2) {\
+    RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0);  \
+    ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Alignment0>(j), b, ptmp0); \
+    ptmp1 = pcj.pmadd(lhs1.template load<LhsPacket, Alignment13>(j), b, ptmp1); \
+    ptmp2 = pcj.pmadd(lhs2.template load<LhsPacket, Alignment2>(j), b, ptmp2); \
+    ptmp3 = pcj.pmadd(lhs3.template load<LhsPacket, Alignment13>(j), b, ptmp3); }
+
+  conj_helper<LhsScalar,RhsScalar,ConjugateLhs,ConjugateRhs> cj;
+  conj_helper<LhsPacket,RhsPacket,ConjugateLhs,ConjugateRhs> pcj;
+
+  typedef typename LhsMapper::VectorMapper LhsScalars;
+
+  enum { AllAligned=0, EvenAligned=1, FirstAligned=2, NoneAligned=3 };
+  const Index rowsAtOnce = 4;
+  const Index peels = 2;
+  const Index RhsPacketAlignedMask = RhsPacketSize-1;
+  const Index LhsPacketAlignedMask = LhsPacketSize-1;
+  const Index depth = cols;
+  const Index lhsStride = lhs.stride();
+
+  // How many coeffs of the result do we have to skip to be aligned.
+  // Here we assume data are at least aligned on the base scalar type
+  // if that's not the case then vectorization is discarded, see below.
+  Index alignedStart = rhs.firstAligned(depth);
+  Index alignedSize = RhsPacketSize>1 ? alignedStart + ((depth-alignedStart) & ~RhsPacketAlignedMask) : 0;
+  const Index peeledSize = alignedSize - RhsPacketSize*peels - RhsPacketSize + 1;
+
+  const Index alignmentStep = LhsPacketSize>1 ? (LhsPacketSize - lhsStride % LhsPacketSize) & LhsPacketAlignedMask : 0;
+  Index alignmentPattern = alignmentStep==0 ? AllAligned
+                           : alignmentStep==(LhsPacketSize/2) ? EvenAligned
+                           : FirstAligned;
+
+  // we cannot assume the first element is aligned because of sub-matrices
+  const Index lhsAlignmentOffset = lhs.firstAligned(depth);
+  const Index rhsAlignmentOffset = rhs.firstAligned(rows);
+
+  // find how many rows do we have to skip to be aligned with rhs (if possible)
+  Index skipRows = 0;
+  // if the data cannot be aligned (TODO add some compile time tests when possible, e.g. for floats)
+  if( (sizeof(LhsScalar)!=sizeof(RhsScalar)) ||
+      (lhsAlignmentOffset < 0) || (lhsAlignmentOffset == depth) ||
+      (rhsAlignmentOffset < 0) || (rhsAlignmentOffset == rows) )
+  {
+    alignedSize = 0;
+    alignedStart = 0;
+    alignmentPattern = NoneAligned;
+  }
+  else if(LhsPacketSize > 4)
+  {
+    // TODO: extend the code to support aligned loads whenever possible when LhsPacketSize > 4.
+    alignmentPattern = NoneAligned;
+  }
+  else if (LhsPacketSize>1)
+  {
+  //    eigen_internal_assert(size_t(firstLhs+lhsAlignmentOffset)%sizeof(LhsPacket)==0  || depth<LhsPacketSize);
+
+    while (skipRows<LhsPacketSize &&
+           alignedStart != ((lhsAlignmentOffset + alignmentStep*skipRows)%LhsPacketSize))
+      ++skipRows;
+    if (skipRows==LhsPacketSize)
+    {
+      // nothing can be aligned, no need to skip any column
+      alignmentPattern = NoneAligned;
+      skipRows = 0;
+    }
+    else
+    {
+      skipRows = (std::min)(skipRows,Index(rows));
+      // note that the skiped columns are processed later.
+    }
+    /*    eigen_internal_assert(  alignmentPattern==NoneAligned
+                      || LhsPacketSize==1
+                      || (skipRows + rowsAtOnce >= rows)
+                      || LhsPacketSize > depth
+                      || (size_t(firstLhs+alignedStart+lhsStride*skipRows)%sizeof(LhsPacket))==0);*/
+  }
+  else if(Vectorizable)
+  {
+    alignedStart = 0;
+    alignedSize = depth;
+    alignmentPattern = AllAligned;
+  }
+
+  const Index offset1 = (FirstAligned && alignmentStep==1)?3:1;
+  const Index offset3 = (FirstAligned && alignmentStep==1)?1:3;
+
+  Index rowBound = ((rows-skipRows)/rowsAtOnce)*rowsAtOnce + skipRows;
+  for (Index i=skipRows; i<rowBound; i+=rowsAtOnce)
+  {
+    // FIXME: what is the purpose of this EIGEN_ALIGN_DEFAULT ??
+    EIGEN_ALIGN_MAX ResScalar tmp0 = ResScalar(0);
+    ResScalar tmp1 = ResScalar(0), tmp2 = ResScalar(0), tmp3 = ResScalar(0);
+
+    // this helps the compiler generating good binary code
+    const LhsScalars lhs0 = lhs.getVectorMapper(i+0, 0),    lhs1 = lhs.getVectorMapper(i+offset1, 0),
+                     lhs2 = lhs.getVectorMapper(i+2, 0),    lhs3 = lhs.getVectorMapper(i+offset3, 0);
+
+    if (Vectorizable)
+    {
+      /* explicit vectorization */
+      ResPacket ptmp0 = pset1<ResPacket>(ResScalar(0)), ptmp1 = pset1<ResPacket>(ResScalar(0)),
+                ptmp2 = pset1<ResPacket>(ResScalar(0)), ptmp3 = pset1<ResPacket>(ResScalar(0));
+
+      // process initial unaligned coeffs
+      // FIXME this loop get vectorized by the compiler !
+      for (Index j=0; j<alignedStart; ++j)
+      {
+        RhsScalar b = rhs(j, 0);
+        tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b);
+        tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b);
+      }
+
+      if (alignedSize>alignedStart)
+      {
+        switch(alignmentPattern)
+        {
+          case AllAligned:
+            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Aligned,Aligned);
+            break;
+          case EvenAligned:
+            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Aligned);
+            break;
+          case FirstAligned:
+          {
+            Index j = alignedStart;
+            if (peels>1)
+            {
+              /* Here we proccess 4 rows with with two peeled iterations to hide
+               * the overhead of unaligned loads. Moreover unaligned loads are handled
+               * using special shift/move operations between the two aligned packets
+               * overlaping the desired unaligned packet. This is *much* more efficient
+               * than basic unaligned loads.
+               */
+              LhsPacket A01, A02, A03, A11, A12, A13;
+              A01 = lhs1.template load<LhsPacket, Aligned>(alignedStart-1);
+              A02 = lhs2.template load<LhsPacket, Aligned>(alignedStart-2);
+              A03 = lhs3.template load<LhsPacket, Aligned>(alignedStart-3);
+
+              for (; j<peeledSize; j+=peels*RhsPacketSize)
+              {
+                RhsPacket b = rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0);
+                A11 = lhs1.template load<LhsPacket, Aligned>(j-1+LhsPacketSize);  palign<1>(A01,A11);
+                A12 = lhs2.template load<LhsPacket, Aligned>(j-2+LhsPacketSize);  palign<2>(A02,A12);
+                A13 = lhs3.template load<LhsPacket, Aligned>(j-3+LhsPacketSize);  palign<3>(A03,A13);
+
+                ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), b, ptmp0);
+                ptmp1 = pcj.pmadd(A01, b, ptmp1);
+                A01 = lhs1.template load<LhsPacket, Aligned>(j-1+2*LhsPacketSize);  palign<1>(A11,A01);
+                ptmp2 = pcj.pmadd(A02, b, ptmp2);
+                A02 = lhs2.template load<LhsPacket, Aligned>(j-2+2*LhsPacketSize);  palign<2>(A12,A02);
+                ptmp3 = pcj.pmadd(A03, b, ptmp3);
+                A03 = lhs3.template load<LhsPacket, Aligned>(j-3+2*LhsPacketSize);  palign<3>(A13,A03);
+
+                b = rhs.getVectorMapper(j+RhsPacketSize, 0).template load<RhsPacket, Aligned>(0);
+                ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j+LhsPacketSize), b, ptmp0);
+                ptmp1 = pcj.pmadd(A11, b, ptmp1);
+                ptmp2 = pcj.pmadd(A12, b, ptmp2);
+                ptmp3 = pcj.pmadd(A13, b, ptmp3);
+              }
+            }
+            for (; j<alignedSize; j+=RhsPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Aligned,Unaligned,Unaligned);
+            break;
+          }
+          default:
+            for (Index j = alignedStart; j<alignedSize; j+=RhsPacketSize)
+              _EIGEN_ACCUMULATE_PACKETS(Unaligned,Unaligned,Unaligned);
+            break;
+        }
+        tmp0 += predux(ptmp0);
+        tmp1 += predux(ptmp1);
+        tmp2 += predux(ptmp2);
+        tmp3 += predux(ptmp3);
+      }
+    } // end explicit vectorization
+
+    // process remaining coeffs (or all if no explicit vectorization)
+    // FIXME this loop get vectorized by the compiler !
+    for (Index j=alignedSize; j<depth; ++j)
+    {
+      RhsScalar b = rhs(j, 0);
+      tmp0 += cj.pmul(lhs0(j),b); tmp1 += cj.pmul(lhs1(j),b);
+      tmp2 += cj.pmul(lhs2(j),b); tmp3 += cj.pmul(lhs3(j),b);
+    }
+    res[i*resIncr]            += alpha*tmp0;
+    res[(i+offset1)*resIncr]  += alpha*tmp1;
+    res[(i+2)*resIncr]        += alpha*tmp2;
+    res[(i+offset3)*resIncr]  += alpha*tmp3;
+  }
+
+  // process remaining first and last rows (at most columnsAtOnce-1)
+  Index end = rows;
+  Index start = rowBound;
+  do
+  {
+    for (Index i=start; i<end; ++i)
+    {
+      EIGEN_ALIGN_MAX ResScalar tmp0 = ResScalar(0);
+      ResPacket ptmp0 = pset1<ResPacket>(tmp0);
+      const LhsScalars lhs0 = lhs.getVectorMapper(i, 0);
+      // process first unaligned result's coeffs
+      // FIXME this loop get vectorized by the compiler !
+      for (Index j=0; j<alignedStart; ++j)
+        tmp0 += cj.pmul(lhs0(j), rhs(j, 0));
+
+      if (alignedSize>alignedStart)
+      {
+        // process aligned rhs coeffs
+        if (lhs0.template aligned<LhsPacket>(alignedStart))
+          for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
+            ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Aligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0);
+        else
+          for (Index j = alignedStart;j<alignedSize;j+=RhsPacketSize)
+            ptmp0 = pcj.pmadd(lhs0.template load<LhsPacket, Unaligned>(j), rhs.getVectorMapper(j, 0).template load<RhsPacket, Aligned>(0), ptmp0);
+        tmp0 += predux(ptmp0);
+      }
+
+      // process remaining scalars
+      // FIXME this loop get vectorized by the compiler !
+      for (Index j=alignedSize; j<depth; ++j)
+        tmp0 += cj.pmul(lhs0(j), rhs(j, 0));
+      res[i*resIncr] += alpha*tmp0;
+    }
+    if (skipRows)
+    {
+      start = 0;
+      end = skipRows;
+      skipRows = 0;
+    }
+    else
+      break;
+  } while(Vectorizable);
+
+  #undef _EIGEN_ACCUMULATE_PACKETS
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_VECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
new file mode 100644
index 0000000..e3a5d58
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/GeneralMatrixVector_BLAS.h
@@ -0,0 +1,129 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   General matrix-vector product functionality based on ?GEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
+#define EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements general matrix-vector multiplication using BLAS
+* gemv function via partial specialization of
+* general_matrix_vector_product::run(..) method for float, double,
+* std::complex<float> and std::complex<double> types
+**********************************************************************/
+
+// gemv specialization
+
+template<typename Index, typename LhsScalar, int StorageOrder, bool ConjugateLhs, typename RhsScalar, bool ConjugateRhs>
+struct general_matrix_vector_product_gemv;
+
+#define EIGEN_BLAS_GEMV_SPECIALIZE(Scalar) \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,Specialized> { \
+static void run( \
+  Index rows, Index cols, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &rhs, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+  if (ConjugateLhs) { \
+    general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ColMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,ConjugateRhs,BuiltIn>::run( \
+      rows, cols, lhs, rhs, res, resIncr, alpha); \
+  } else { \
+    general_matrix_vector_product_gemv<Index,Scalar,ColMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
+  } \
+} \
+}; \
+template<typename Index, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product<Index,Scalar,const_blas_data_mapper<Scalar,Index,RowMajor>,RowMajor,ConjugateLhs,Scalar,const_blas_data_mapper<Scalar,Index,ColMajor>,ConjugateRhs,Specialized> { \
+static void run( \
+  Index rows, Index cols, \
+  const const_blas_data_mapper<Scalar,Index,RowMajor> &lhs, \
+  const const_blas_data_mapper<Scalar,Index,ColMajor> &rhs, \
+  Scalar* res, Index resIncr, Scalar alpha) \
+{ \
+    general_matrix_vector_product_gemv<Index,Scalar,RowMajor,ConjugateLhs,Scalar,ConjugateRhs>::run( \
+      rows, cols, lhs.data(), lhs.stride(), rhs.data(), rhs.stride(), res, resIncr, alpha); \
+} \
+}; \
+
+EIGEN_BLAS_GEMV_SPECIALIZE(double)
+EIGEN_BLAS_GEMV_SPECIALIZE(float)
+EIGEN_BLAS_GEMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_GEMV_SPECIALIZE(scomplex)
+
+#define EIGEN_BLAS_GEMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASPREFIX) \
+template<typename Index, int LhsStorageOrder, bool ConjugateLhs, bool ConjugateRhs> \
+struct general_matrix_vector_product_gemv<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> GEMVVector;\
+\
+static void run( \
+  Index rows, Index cols, \
+  const EIGTYPE* lhs, Index lhsStride, \
+  const EIGTYPE* rhs, Index rhsIncr, \
+  EIGTYPE* res, Index resIncr, EIGTYPE alpha) \
+{ \
+  BlasIndex m=convert_index<BlasIndex>(rows), n=convert_index<BlasIndex>(cols), \
+            lda=convert_index<BlasIndex>(lhsStride), incx=convert_index<BlasIndex>(rhsIncr), incy=convert_index<BlasIndex>(resIncr); \
+  const EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
+  char trans=(LhsStorageOrder==ColMajor) ? 'N' : (ConjugateLhs) ? 'C' : 'T'; \
+  if (LhsStorageOrder==RowMajor) { \
+    m = convert_index<BlasIndex>(cols); \
+    n = convert_index<BlasIndex>(rows); \
+  }\
+  GEMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const GEMVVector, 0, InnerStride<> > map_x(rhs,cols,1,InnerStride<>(incx)); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+    incx=1; \
+  } else x_ptr=rhs; \
+  BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
+}\
+};
+
+EIGEN_BLAS_GEMV_SPECIALIZATION(double,   double, d)
+EIGEN_BLAS_GEMV_SPECIALIZATION(float,    float,  s)
+EIGEN_BLAS_GEMV_SPECIALIZATION(dcomplex, double, z)
+EIGEN_BLAS_GEMV_SPECIALIZATION(scomplex, float,  c)
+
+} // end namespase internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERAL_MATRIX_VECTOR_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h
new file mode 100644
index 0000000..c2f084c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/Parallelizer.h
@@ -0,0 +1,163 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARALLELIZER_H
+#define EIGEN_PARALLELIZER_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** \internal */
+inline void manage_multi_threading(Action action, int* v)
+{
+  static EIGEN_UNUSED int m_maxThreads = -1;
+
+  if(action==SetAction)
+  {
+    eigen_internal_assert(v!=0);
+    m_maxThreads = *v;
+  }
+  else if(action==GetAction)
+  {
+    eigen_internal_assert(v!=0);
+    #ifdef EIGEN_HAS_OPENMP
+    if(m_maxThreads>0)
+      *v = m_maxThreads;
+    else
+      *v = omp_get_max_threads();
+    #else
+    *v = 1;
+    #endif
+  }
+  else
+  {
+    eigen_internal_assert(false);
+  }
+}
+
+}
+
+/** Must be call first when calling Eigen from multiple threads */
+inline void initParallel()
+{
+  int nbt;
+  internal::manage_multi_threading(GetAction, &nbt);
+  std::ptrdiff_t l1, l2, l3;
+  internal::manage_caching_sizes(GetAction, &l1, &l2, &l3);
+}
+
+/** \returns the max number of threads reserved for Eigen
+  * \sa setNbThreads */
+inline int nbThreads()
+{
+  int ret;
+  internal::manage_multi_threading(GetAction, &ret);
+  return ret;
+}
+
+/** Sets the max number of threads reserved for Eigen
+  * \sa nbThreads */
+inline void setNbThreads(int v)
+{
+  internal::manage_multi_threading(SetAction, &v);
+}
+
+namespace internal {
+
+template<typename Index> struct GemmParallelInfo
+{
+  GemmParallelInfo() : sync(-1), users(0), lhs_start(0), lhs_length(0) {}
+
+  Index volatile sync;
+  int volatile users;
+
+  Index lhs_start;
+  Index lhs_length;
+};
+
+template<bool Condition, typename Functor, typename Index>
+void parallelize_gemm(const Functor& func, Index rows, Index cols, Index depth, bool transpose)
+{
+  // TODO when EIGEN_USE_BLAS is defined,
+  // we should still enable OMP for other scalar types
+#if !(defined (EIGEN_HAS_OPENMP)) || defined (EIGEN_USE_BLAS)
+  // FIXME the transpose variable is only needed to properly split
+  // the matrix product when multithreading is enabled. This is a temporary
+  // fix to support row-major destination matrices. This whole
+  // parallelizer mechanism has to be redisigned anyway.
+  EIGEN_UNUSED_VARIABLE(depth);
+  EIGEN_UNUSED_VARIABLE(transpose);
+  func(0,rows, 0,cols);
+#else
+
+  // Dynamically check whether we should enable or disable OpenMP.
+  // The conditions are:
+  // - the max number of threads we can create is greater than 1
+  // - we are not already in a parallel code
+  // - the sizes are large enough
+
+  // compute the maximal number of threads from the size of the product:
+  // This first heuristic takes into account that the product kernel is fully optimized when working with nr columns at once.
+  Index size = transpose ? rows : cols;
+  Index pb_max_threads = std::max<Index>(1,size / Functor::Traits::nr);
+
+  // compute the maximal number of threads from the total amount of work:
+  double work = static_cast<double>(rows) * static_cast<double>(cols) *
+      static_cast<double>(depth);
+  double kMinTaskSize = 50000;  // FIXME improve this heuristic.
+  pb_max_threads = std::max<Index>(1, std::min<Index>(pb_max_threads, work / kMinTaskSize));
+
+  // compute the number of threads we are going to use
+  Index threads = std::min<Index>(nbThreads(), pb_max_threads);
+
+  // if multi-threading is explicitely disabled, not useful, or if we already are in a parallel session,
+  // then abort multi-threading
+  // FIXME omp_get_num_threads()>1 only works for openmp, what if the user does not use openmp?
+  if((!Condition) || (threads==1) || (omp_get_num_threads()>1))
+    return func(0,rows, 0,cols);
+
+  Eigen::initParallel();
+  func.initParallelSession(threads);
+
+  if(transpose)
+    std::swap(rows,cols);
+
+  ei_declare_aligned_stack_constructed_variable(GemmParallelInfo<Index>,info,threads,0);
+
+  #pragma omp parallel num_threads(threads)
+  {
+    Index i = omp_get_thread_num();
+    // Note that the actual number of threads might be lower than the number of request ones.
+    Index actual_threads = omp_get_num_threads();
+
+    Index blockCols = (cols / actual_threads) & ~Index(0x3);
+    Index blockRows = (rows / actual_threads);
+    blockRows = (blockRows/Functor::Traits::mr)*Functor::Traits::mr;
+
+    Index r0 = i*blockRows;
+    Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
+
+    Index c0 = i*blockCols;
+    Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
+
+    info[i].lhs_start = r0;
+    info[i].lhs_length = actualBlockRows;
+
+    if(transpose) func(c0, actualBlockCols, 0, rows, info);
+    else          func(0, rows, c0, actualBlockCols, info);
+  }
+#endif
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARALLELIZER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
new file mode 100644
index 0000000..da6f82a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix.h
@@ -0,0 +1,521 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_H
+#define EIGEN_SELFADJOINT_MATRIX_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// pack a selfadjoint block diagonal for use with the gebp_kernel
+template<typename Scalar, typename Index, int Pack1, int Pack2_dummy, int StorageOrder>
+struct symm_pack_lhs
+{
+  template<int BlockRows> inline
+  void pack(Scalar* blockA, const const_blas_data_mapper<Scalar,Index,StorageOrder>& lhs, Index cols, Index i, Index& count)
+  {
+    // normal copy
+    for(Index k=0; k<i; k++)
+      for(Index w=0; w<BlockRows; w++)
+        blockA[count++] = lhs(i+w,k);           // normal
+    // symmetric copy
+    Index h = 0;
+    for(Index k=i; k<i+BlockRows; k++)
+    {
+      for(Index w=0; w<h; w++)
+        blockA[count++] = numext::conj(lhs(k, i+w)); // transposed
+
+      blockA[count++] = numext::real(lhs(k,k));   // real (diagonal)
+
+      for(Index w=h+1; w<BlockRows; w++)
+        blockA[count++] = lhs(i+w, k);          // normal
+      ++h;
+    }
+    // transposed copy
+    for(Index k=i+BlockRows; k<cols; k++)
+      for(Index w=0; w<BlockRows; w++)
+        blockA[count++] = numext::conj(lhs(k, i+w)); // transposed
+  }
+  void operator()(Scalar* blockA, const Scalar* _lhs, Index lhsStride, Index cols, Index rows)
+  {
+    enum { PacketSize = packet_traits<Scalar>::size };
+    const_blas_data_mapper<Scalar,Index,StorageOrder> lhs(_lhs,lhsStride);
+    Index count = 0;
+    //Index peeled_mc3 = (rows/Pack1)*Pack1;
+    
+    const Index peeled_mc3 = Pack1>=3*PacketSize ? (rows/(3*PacketSize))*(3*PacketSize) : 0;
+    const Index peeled_mc2 = Pack1>=2*PacketSize ? peeled_mc3+((rows-peeled_mc3)/(2*PacketSize))*(2*PacketSize) : 0;
+    const Index peeled_mc1 = Pack1>=1*PacketSize ? (rows/(1*PacketSize))*(1*PacketSize) : 0;
+    
+    if(Pack1>=3*PacketSize)
+      for(Index i=0; i<peeled_mc3; i+=3*PacketSize)
+        pack<3*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=2*PacketSize)
+      for(Index i=peeled_mc3; i<peeled_mc2; i+=2*PacketSize)
+        pack<2*PacketSize>(blockA, lhs, cols, i, count);
+    
+    if(Pack1>=1*PacketSize)
+      for(Index i=peeled_mc2; i<peeled_mc1; i+=1*PacketSize)
+        pack<1*PacketSize>(blockA, lhs, cols, i, count);
+
+    // do the same with mr==1
+    for(Index i=peeled_mc1; i<rows; i++)
+    {
+      for(Index k=0; k<i; k++)
+        blockA[count++] = lhs(i, k);                   // normal
+
+      blockA[count++] = numext::real(lhs(i, i));       // real (diagonal)
+
+      for(Index k=i+1; k<cols; k++)
+        blockA[count++] = numext::conj(lhs(k, i));     // transposed
+    }
+  }
+};
+
+template<typename Scalar, typename Index, int nr, int StorageOrder>
+struct symm_pack_rhs
+{
+  enum { PacketSize = packet_traits<Scalar>::size };
+  void operator()(Scalar* blockB, const Scalar* _rhs, Index rhsStride, Index rows, Index cols, Index k2)
+  {
+    Index end_k = k2 + rows;
+    Index count = 0;
+    const_blas_data_mapper<Scalar,Index,StorageOrder> rhs(_rhs,rhsStride);
+    Index packet_cols8 = nr>=8 ? (cols/8) * 8 : 0;
+    Index packet_cols4 = nr>=4 ? (cols/4) * 4 : 0;
+
+    // first part: normal case
+    for(Index j2=0; j2<k2; j2+=nr)
+    {
+      for(Index k=k2; k<end_k; k++)
+      {
+        blockB[count+0] = rhs(k,j2+0);
+        blockB[count+1] = rhs(k,j2+1);
+        if (nr>=4)
+        {
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+        }
+        if (nr>=8)
+        {
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+        }
+        count += nr;
+      }
+    }
+
+    // second part: diagonal block
+    Index end8 = nr>=8 ? (std::min)(k2+rows,packet_cols8) : k2;
+    if(nr>=8)
+    {
+      for(Index j2=k2; j2<end8; j2+=8)
+      {
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
+        }
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+8; k++)
+        {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
+
+          blockB[count+h] = numext::real(rhs(k,k));
+
+          // transpose
+          for (Index w=h+1 ; w<8; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 8;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+8; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+          blockB[count+4] = rhs(k,j2+4);
+          blockB[count+5] = rhs(k,j2+5);
+          blockB[count+6] = rhs(k,j2+6);
+          blockB[count+7] = rhs(k,j2+7);
+          count += 8;
+        }
+      }
+    }
+    if(nr>=4)
+    {
+      for(Index j2=end8; j2<(std::min)(k2+rows,packet_cols4); j2+=4)
+      {
+        // again we can split vertically in three different parts (transpose, symmetric, normal)
+        // transpose
+        for(Index k=k2; k<j2; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
+        }
+        // symmetric
+        Index h = 0;
+        for(Index k=j2; k<j2+4; k++)
+        {
+          // normal
+          for (Index w=0 ; w<h; ++w)
+            blockB[count+w] = rhs(k,j2+w);
+
+          blockB[count+h] = numext::real(rhs(k,k));
+
+          // transpose
+          for (Index w=h+1 ; w<4; ++w)
+            blockB[count+w] = numext::conj(rhs(j2+w,k));
+          count += 4;
+          ++h;
+        }
+        // normal
+        for(Index k=j2+4; k<end_k; k++)
+        {
+          blockB[count+0] = rhs(k,j2+0);
+          blockB[count+1] = rhs(k,j2+1);
+          blockB[count+2] = rhs(k,j2+2);
+          blockB[count+3] = rhs(k,j2+3);
+          count += 4;
+        }
+      }
+    }
+
+    // third part: transposed
+    if(nr>=8)
+    {
+      for(Index j2=k2+rows; j2<packet_cols8; j2+=8)
+      {
+        for(Index k=k2; k<end_k; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          blockB[count+4] = numext::conj(rhs(j2+4,k));
+          blockB[count+5] = numext::conj(rhs(j2+5,k));
+          blockB[count+6] = numext::conj(rhs(j2+6,k));
+          blockB[count+7] = numext::conj(rhs(j2+7,k));
+          count += 8;
+        }
+      }
+    }
+    if(nr>=4)
+    {
+      for(Index j2=(std::max)(packet_cols8,k2+rows); j2<packet_cols4; j2+=4)
+      {
+        for(Index k=k2; k<end_k; k++)
+        {
+          blockB[count+0] = numext::conj(rhs(j2+0,k));
+          blockB[count+1] = numext::conj(rhs(j2+1,k));
+          blockB[count+2] = numext::conj(rhs(j2+2,k));
+          blockB[count+3] = numext::conj(rhs(j2+3,k));
+          count += 4;
+        }
+      }
+    }
+
+    // copy the remaining columns one at a time (=> the same with nr==1)
+    for(Index j2=packet_cols4; j2<cols; ++j2)
+    {
+      // transpose
+      Index half = (std::min)(end_k,j2);
+      for(Index k=k2; k<half; k++)
+      {
+        blockB[count] = numext::conj(rhs(j2,k));
+        count += 1;
+      }
+
+      if(half==j2 && half<k2+rows)
+      {
+        blockB[count] = numext::real(rhs(j2,j2));
+        count += 1;
+      }
+      else
+        half--;
+
+      // normal
+      for(Index k=half+1; k<k2+rows; k++)
+      {
+        blockB[count] = rhs(k,j2);
+        count += 1;
+      }
+    }
+  }
+};
+
+/* Optimized selfadjoint matrix * matrix (_SYMM) product built on top of
+ * the general matrix matrix product.
+ */
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs,
+          int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs,
+          int ResStorageOrder>
+struct product_selfadjoint_matrix;
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool LhsSelfAdjoint, bool ConjugateLhs,
+          int RhsStorageOrder, bool RhsSelfAdjoint, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,LhsSelfAdjoint,ConjugateLhs, RhsStorageOrder,RhsSelfAdjoint,ConjugateRhs,RowMajor>
+{
+
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols,
+    const Scalar* lhs, Index lhsStride,
+    const Scalar* rhs, Index rhsStride,
+    Scalar* res,       Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    product_selfadjoint_matrix<Scalar, Index,
+      EIGEN_LOGICAL_XOR(RhsSelfAdjoint,RhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
+      RhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsSelfAdjoint,ConjugateRhs),
+      EIGEN_LOGICAL_XOR(LhsSelfAdjoint,LhsStorageOrder==RowMajor) ? ColMajor : RowMajor,
+      LhsSelfAdjoint, NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsSelfAdjoint,ConjugateLhs),
+      ColMajor>
+      ::run(cols, rows,  rhs, rhsStride,  lhs, lhsStride,  res, resStride,  alpha, blocking);
+  }
+};
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor>
+{
+
+  static EIGEN_DONT_INLINE void run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,true,ConjugateLhs, RhsStorageOrder,false,ConjugateRhs,ColMajor>::run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index size = rows;
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, (LhsStorageOrder == RowMajor) ? ColMajor : RowMajor> LhsTransposeMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    LhsTransposeMapper lhs_transpose(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    // kc must be smaller than mc
+    kc = (std::min)(kc,mc);
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    symm_pack_lhs<Scalar, Index, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_lhs<Scalar, Index, LhsTransposeMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder==RowMajor?ColMajor:RowMajor, true> pack_lhs_transposed;
+
+    for(Index k2=0; k2<size; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,size)-k2;
+
+      // we have selected one row panel of rhs and one column panel of lhs
+      // pack rhs's panel into a sequential chunk of memory
+      // and expand each coeff to a constant packet for further reuse
+      pack_rhs(blockB, rhs.getSubMapper(k2,0), actual_kc, cols);
+
+      // the select lhs's panel has to be split in three different parts:
+      //  1 - the transposed panel above the diagonal block => transposed packed copy
+      //  2 - the diagonal block => special packed copy
+      //  3 - the panel below the diagonal block => generic packed copy
+      for(Index i2=0; i2<k2; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,k2)-i2;
+        // transposed packed copy
+        pack_lhs_transposed(blockA, lhs_transpose.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+      // the block diagonal
+      {
+        const Index actual_mc = (std::min)(k2+kc,size)-k2;
+        // symmetric packed copy
+        pack_lhs(blockA, &lhs(k2,k2), lhsStride, actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(k2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+
+      for(Index i2=k2+kc; i2<size; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,size)-i2;
+        gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder,false>()
+          (blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+    }
+  }
+
+// matrix * selfadjoint product
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+struct product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor>
+{
+
+  static EIGEN_DONT_INLINE void run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs>
+EIGEN_DONT_INLINE void product_selfadjoint_matrix<Scalar,Index,LhsStorageOrder,false,ConjugateLhs, RhsStorageOrder,true,ConjugateRhs,ColMajor>::run(
+    Index rows, Index cols,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index size = cols;
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    ResMapper res(_res,resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    symm_pack_rhs<Scalar, Index, Traits::nr,RhsStorageOrder> pack_rhs;
+
+    for(Index k2=0; k2<size; k2+=kc)
+    {
+      const Index actual_kc = (std::min)(k2+kc,size)-k2;
+
+      pack_rhs(blockB, _rhs, rhsStride, actual_kc, cols, k2);
+
+      // => GEPP
+      for(Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(i2+mc,rows)-i2;
+        pack_lhs(blockA, lhs.getSubMapper(i2, k2), actual_kc, actual_mc);
+
+        gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, alpha);
+      }
+    }
+  }
+
+} // end namespace internal
+
+/***************************************************************************
+* Wrapper to product_selfadjoint_matrix
+***************************************************************************/
+
+namespace internal {
+  
+template<typename Lhs, int LhsMode, typename Rhs, int RhsMode>
+struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,RhsMode,false>
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  
+  enum {
+    LhsIsUpper = (LhsMode&(Upper|Lower))==Upper,
+    LhsIsSelfAdjoint = (LhsMode&SelfAdjoint)==SelfAdjoint,
+    RhsIsUpper = (RhsMode&(Upper|Lower))==Upper,
+    RhsIsSelfAdjoint = (RhsMode&SelfAdjoint)==SelfAdjoint
+  };
+  
+  template<typename Dest>
+  static void run(Dest &dst, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==a_lhs.rows() && dst.cols()==a_rhs.cols());
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,1> BlockingType;
+
+    BlockingType blocking(lhs.rows(), rhs.cols(), lhs.cols(), 1, false);
+
+    internal::product_selfadjoint_matrix<Scalar, Index,
+      EIGEN_LOGICAL_XOR(LhsIsUpper,internal::traits<Lhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, LhsIsSelfAdjoint,
+      NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(LhsIsUpper,bool(LhsBlasTraits::NeedToConjugate)),
+      EIGEN_LOGICAL_XOR(RhsIsUpper,internal::traits<Rhs>::Flags &RowMajorBit) ? RowMajor : ColMajor, RhsIsSelfAdjoint,
+      NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(RhsIsUpper,bool(RhsBlasTraits::NeedToConjugate)),
+      internal::traits<Dest>::Flags&RowMajorBit  ? RowMajor : ColMajor>
+      ::run(
+        lhs.rows(), rhs.cols(),                 // sizes
+        &lhs.coeffRef(0,0), lhs.outerStride(),  // lhs info
+        &rhs.coeffRef(0,0), rhs.outerStride(),  // rhs info
+        &dst.coeffRef(0,0), dst.outerStride(),  // result info
+        actualAlpha, blocking                   // alpha
+      );
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
new file mode 100644
index 0000000..a45238d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixMatrix_BLAS.h
@@ -0,0 +1,275 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+//
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Self adjoint matrix * matrix product functionality based on ?SYMM/?HEMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
+#define EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+
+/* Optimized selfadjoint matrix * matrix (?SYMM/?HEMM) product */
+
+#define EIGEN_BLAS_SYMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \
+{\
+\
+  static void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+    char side='L', uplo='L'; \
+    BlasIndex m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    EIGTYPE beta(1); \
+    MatrixX##EIGPREFIX b_tmp; \
+\
+/* Set transpose options */ \
+/* Set m, n, k */ \
+    m = convert_index<BlasIndex>(rows);  \
+    n = convert_index<BlasIndex>(cols);  \
+\
+/* Set lda, ldb, ldc */ \
+    lda = convert_index<BlasIndex>(lhsStride); \
+    ldb = convert_index<BlasIndex>(rhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
+\
+/* Set a, b, c */ \
+    if (LhsStorageOrder==RowMajor) uplo='U'; \
+    a = _lhs; \
+\
+    if (RhsStorageOrder==RowMajor) { \
+      Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+      b_tmp = rhs.adjoint(); \
+      b = b_tmp.data(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+    } else b = _rhs; \
+\
+    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+\
+  } \
+};
+
+
+#define EIGEN_BLAS_HEMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,true,ConjugateLhs,RhsStorageOrder,false,ConjugateRhs,ColMajor> \
+{\
+  static void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+    char side='L', uplo='L'; \
+    BlasIndex m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    EIGTYPE beta(1); \
+    MatrixX##EIGPREFIX b_tmp; \
+    Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> a_tmp; \
+\
+/* Set transpose options */ \
+/* Set m, n, k */ \
+    m = convert_index<BlasIndex>(rows); \
+    n = convert_index<BlasIndex>(cols); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = convert_index<BlasIndex>(lhsStride); \
+    ldb = convert_index<BlasIndex>(rhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
+\
+/* Set a, b, c */ \
+    if (((LhsStorageOrder==ColMajor) && ConjugateLhs) || ((LhsStorageOrder==RowMajor) && (!ConjugateLhs))) { \
+      Map<const Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder>, 0, OuterStride<> > lhs(_lhs,m,m,OuterStride<>(lhsStride)); \
+      a_tmp = lhs.conjugate(); \
+      a = a_tmp.data(); \
+      lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+    } else a = _lhs; \
+    if (LhsStorageOrder==RowMajor) uplo='U'; \
+\
+    if (RhsStorageOrder==ColMajor && (!ConjugateRhs)) { \
+       b = _rhs; } \
+    else { \
+      if (RhsStorageOrder==ColMajor && ConjugateRhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,m,n,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.conjugate(); \
+      } else \
+      if (ConjugateRhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.adjoint(); \
+      } else { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > rhs(_rhs,n,m,OuterStride<>(rhsStride)); \
+        b_tmp = rhs.transpose(); \
+      } \
+      b = b_tmp.data(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+    } \
+\
+    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+\
+  } \
+};
+
+EIGEN_BLAS_SYMM_L(double, double, d, d)
+EIGEN_BLAS_SYMM_L(float, float, f, s)
+EIGEN_BLAS_HEMM_L(dcomplex, double, cd, z)
+EIGEN_BLAS_HEMM_L(scomplex, float, cf, c)
+
+
+/* Optimized matrix * selfadjoint matrix (?SYMM/?HEMM) product */
+
+#define EIGEN_BLAS_SYMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \
+{\
+\
+  static void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+    char side='R', uplo='L'; \
+    BlasIndex m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    EIGTYPE beta(1); \
+    MatrixX##EIGPREFIX b_tmp; \
+\
+/* Set m, n, k */ \
+    m = convert_index<BlasIndex>(rows);  \
+    n = convert_index<BlasIndex>(cols);  \
+\
+/* Set lda, ldb, ldc */ \
+    lda = convert_index<BlasIndex>(rhsStride); \
+    ldb = convert_index<BlasIndex>(lhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
+\
+/* Set a, b, c */ \
+    if (RhsStorageOrder==RowMajor) uplo='U'; \
+    a = _rhs; \
+\
+    if (LhsStorageOrder==RowMajor) { \
+      Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(rhsStride)); \
+      b_tmp = lhs.adjoint(); \
+      b = b_tmp.data(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+    } else b = _lhs; \
+\
+    BLASPREFIX##symm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+\
+  } \
+};
+
+
+#define EIGEN_BLAS_HEMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_selfadjoint_matrix<EIGTYPE,Index,LhsStorageOrder,false,ConjugateLhs,RhsStorageOrder,true,ConjugateRhs,ColMajor> \
+{\
+  static void run( \
+    Index rows, Index cols, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE, EIGTYPE>& /*blocking*/) \
+  { \
+    char side='R', uplo='L'; \
+    BlasIndex m, n, lda, ldb, ldc; \
+    const EIGTYPE *a, *b; \
+    EIGTYPE beta(1); \
+    MatrixX##EIGPREFIX b_tmp; \
+    Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> a_tmp; \
+\
+/* Set m, n, k */ \
+    m = convert_index<BlasIndex>(rows); \
+    n = convert_index<BlasIndex>(cols); \
+\
+/* Set lda, ldb, ldc */ \
+    lda = convert_index<BlasIndex>(rhsStride); \
+    ldb = convert_index<BlasIndex>(lhsStride); \
+    ldc = convert_index<BlasIndex>(resStride); \
+\
+/* Set a, b, c */ \
+    if (((RhsStorageOrder==ColMajor) && ConjugateRhs) || ((RhsStorageOrder==RowMajor) && (!ConjugateRhs))) { \
+      Map<const Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder>, 0, OuterStride<> > rhs(_rhs,n,n,OuterStride<>(rhsStride)); \
+      a_tmp = rhs.conjugate(); \
+      a = a_tmp.data(); \
+      lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+    } else a = _rhs; \
+    if (RhsStorageOrder==RowMajor) uplo='U'; \
+\
+    if (LhsStorageOrder==ColMajor && (!ConjugateLhs)) { \
+       b = _lhs; } \
+    else { \
+      if (LhsStorageOrder==ColMajor && ConjugateLhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,m,n,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.conjugate(); \
+      } else \
+      if (ConjugateLhs) { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.adjoint(); \
+      } else { \
+        Map<const MatrixX##EIGPREFIX, 0, OuterStride<> > lhs(_lhs,n,m,OuterStride<>(lhsStride)); \
+        b_tmp = lhs.transpose(); \
+      } \
+      b = b_tmp.data(); \
+      ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+    } \
+\
+    BLASPREFIX##hemm_(&side, &uplo, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)b, &ldb, &numext::real_ref(beta), (BLASTYPE*)res, &ldc); \
+  } \
+};
+
+EIGEN_BLAS_SYMM_R(double, double, d, d)
+EIGEN_BLAS_SYMM_R(float, float, f, s)
+EIGEN_BLAS_HEMM_R(dcomplex, double, cd, z)
+EIGEN_BLAS_HEMM_R(scomplex, float, cf, c)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_MATRIX_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h
new file mode 100644
index 0000000..3fd180e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector.h
@@ -0,0 +1,260 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_H
+#define EIGEN_SELFADJOINT_MATRIX_VECTOR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/* Optimized selfadjoint matrix * vector product:
+ * This algorithm processes 2 columns at onces that allows to both reduce
+ * the number of load/stores of the result by a factor 2 and to reduce
+ * the instruction dependency.
+ */
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version=Specialized>
+struct selfadjoint_matrix_vector_product;
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
+struct selfadjoint_matrix_vector_product
+
+{
+static EIGEN_DONT_INLINE void run(
+  Index size,
+  const Scalar*  lhs, Index lhsStride,
+  const Scalar*  rhs,
+  Scalar* res,
+  Scalar alpha);
+};
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Version>::run(
+  Index size,
+  const Scalar*  lhs, Index lhsStride,
+  const Scalar*  rhs,
+  Scalar* res,
+  Scalar alpha)
+{
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  const Index PacketSize = sizeof(Packet)/sizeof(Scalar);
+
+  enum {
+    IsRowMajor = StorageOrder==RowMajor ? 1 : 0,
+    IsLower = UpLo == Lower ? 1 : 0,
+    FirstTriangular = IsRowMajor == IsLower
+  };
+
+  conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> cj0;
+  conj_helper<Scalar,Scalar,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> cj1;
+  conj_helper<RealScalar,Scalar,false, ConjugateRhs> cjd;
+
+  conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs,  IsRowMajor), ConjugateRhs> pcj0;
+  conj_helper<Packet,Packet,NumTraits<Scalar>::IsComplex && EIGEN_LOGICAL_XOR(ConjugateLhs, !IsRowMajor), ConjugateRhs> pcj1;
+
+  Scalar cjAlpha = ConjugateRhs ? numext::conj(alpha) : alpha;
+
+
+  Index bound = (std::max)(Index(0),size-8) & 0xfffffffe;
+  if (FirstTriangular)
+    bound = size - bound;
+
+  for (Index j=FirstTriangular ? bound : 0;
+       j<(FirstTriangular ? size : bound);j+=2)
+  {
+    const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
+    const Scalar* EIGEN_RESTRICT A1 = lhs + (j+1)*lhsStride;
+
+    Scalar t0 = cjAlpha * rhs[j];
+    Packet ptmp0 = pset1<Packet>(t0);
+    Scalar t1 = cjAlpha * rhs[j+1];
+    Packet ptmp1 = pset1<Packet>(t1);
+
+    Scalar t2(0);
+    Packet ptmp2 = pset1<Packet>(t2);
+    Scalar t3(0);
+    Packet ptmp3 = pset1<Packet>(t3);
+
+    Index starti = FirstTriangular ? 0 : j+2;
+    Index endi   = FirstTriangular ? j : size;
+    Index alignedStart = (starti) + internal::first_default_aligned(&res[starti], endi-starti);
+    Index alignedEnd = alignedStart + ((endi-alignedStart)/(PacketSize))*(PacketSize);
+
+    res[j]   += cjd.pmul(numext::real(A0[j]), t0);
+    res[j+1] += cjd.pmul(numext::real(A1[j+1]), t1);
+    if(FirstTriangular)
+    {
+      res[j]   += cj0.pmul(A1[j],   t1);
+      t3       += cj1.pmul(A1[j],   rhs[j]);
+    }
+    else
+    {
+      res[j+1] += cj0.pmul(A0[j+1],t0);
+      t2 += cj1.pmul(A0[j+1], rhs[j+1]);
+    }
+
+    for (Index i=starti; i<alignedStart; ++i)
+    {
+      res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
+      t2 += cj1.pmul(A0[i], rhs[i]);
+      t3 += cj1.pmul(A1[i], rhs[i]);
+    }
+    // Yes this an optimization for gcc 4.3 and 4.4 (=> huge speed up)
+    // gcc 4.2 does this optimization automatically.
+    const Scalar* EIGEN_RESTRICT a0It  = A0  + alignedStart;
+    const Scalar* EIGEN_RESTRICT a1It  = A1  + alignedStart;
+    const Scalar* EIGEN_RESTRICT rhsIt = rhs + alignedStart;
+          Scalar* EIGEN_RESTRICT resIt = res + alignedStart;
+    for (Index i=alignedStart; i<alignedEnd; i+=PacketSize)
+    {
+      Packet A0i = ploadu<Packet>(a0It);  a0It  += PacketSize;
+      Packet A1i = ploadu<Packet>(a1It);  a1It  += PacketSize;
+      Packet Bi  = ploadu<Packet>(rhsIt); rhsIt += PacketSize; // FIXME should be aligned in most cases
+      Packet Xi  = pload <Packet>(resIt);
+
+      Xi    = pcj0.pmadd(A0i,ptmp0, pcj0.pmadd(A1i,ptmp1,Xi));
+      ptmp2 = pcj1.pmadd(A0i,  Bi, ptmp2);
+      ptmp3 = pcj1.pmadd(A1i,  Bi, ptmp3);
+      pstore(resIt,Xi); resIt += PacketSize;
+    }
+    for (Index i=alignedEnd; i<endi; i++)
+    {
+      res[i] += cj0.pmul(A0[i], t0) + cj0.pmul(A1[i],t1);
+      t2 += cj1.pmul(A0[i], rhs[i]);
+      t3 += cj1.pmul(A1[i], rhs[i]);
+    }
+
+    res[j]   += alpha * (t2 + predux(ptmp2));
+    res[j+1] += alpha * (t3 + predux(ptmp3));
+  }
+  for (Index j=FirstTriangular ? 0 : bound;j<(FirstTriangular ? bound : size);j++)
+  {
+    const Scalar* EIGEN_RESTRICT A0 = lhs + j*lhsStride;
+
+    Scalar t1 = cjAlpha * rhs[j];
+    Scalar t2(0);
+    res[j] += cjd.pmul(numext::real(A0[j]), t1);
+    for (Index i=FirstTriangular ? 0 : j+1; i<(FirstTriangular ? j : size); i++)
+    {
+      res[i] += cj0.pmul(A0[i], t1);
+      t2 += cj1.pmul(A0[i], rhs[i]);
+    }
+    res[j] += alpha * t2;
+  }
+}
+
+} // end namespace internal 
+
+/***************************************************************************
+* Wrapper to product_selfadjoint_vector
+***************************************************************************/
+
+namespace internal {
+
+template<typename Lhs, int LhsMode, typename Rhs>
+struct selfadjoint_product_impl<Lhs,LhsMode,false,Rhs,0,true>
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  typedef internal::blas_traits<Lhs> LhsBlasTraits;
+  typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+  typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+  
+  typedef internal::blas_traits<Rhs> RhsBlasTraits;
+  typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+  typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+  enum { LhsUpLo = LhsMode&(Upper|Lower) };
+
+  template<typename Dest>
+  static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
+  {
+    typedef typename Dest::Scalar ResScalar;
+    typedef typename Rhs::Scalar RhsScalar;
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+    
+    eigen_assert(dest.rows()==a_lhs.rows() && dest.cols()==a_rhs.cols());
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    enum {
+      EvalToDest = (Dest::InnerStrideAtCompileTime==1),
+      UseRhs = (ActualRhsTypeCleaned::InnerStrideAtCompileTime==1)
+    };
+    
+    internal::gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,!EvalToDest> static_dest;
+    internal::gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!UseRhs> static_rhs;
+
+    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                  EvalToDest ? dest.data() : static_dest.data());
+                                                  
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,rhs.size(),
+        UseRhs ? const_cast<RhsScalar*>(rhs.data()) : static_rhs.data());
+    
+    if(!EvalToDest)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = dest.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      MappedDest(actualDestPtr, dest.size()) = dest;
+    }
+      
+    if(!UseRhs)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = rhs.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, rhs.size()) = rhs;
+    }
+      
+      
+    internal::selfadjoint_matrix_vector_product<Scalar, Index, (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+                                                int(LhsUpLo), bool(LhsBlasTraits::NeedToConjugate), bool(RhsBlasTraits::NeedToConjugate)>::run
+      (
+        lhs.rows(),                             // size
+        &lhs.coeffRef(0,0),  lhs.outerStride(), // lhs info
+        actualRhsPtr,                           // rhs info
+        actualDestPtr,                          // result info
+        actualAlpha                             // scale factor
+      );
+    
+    if(!EvalToDest)
+      dest = MappedDest(actualDestPtr, dest.size());
+  }
+};
+
+template<typename Lhs, typename Rhs, int RhsMode>
+struct selfadjoint_product_impl<Lhs,0,true,Rhs,RhsMode,false>
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  enum { RhsUpLo = RhsMode&(Upper|Lower)  };
+
+  template<typename Dest>
+  static void run(Dest& dest, const Lhs &a_lhs, const Rhs &a_rhs, const Scalar& alpha)
+  {
+    // let's simply transpose the product
+    Transpose<Dest> destT(dest);
+    selfadjoint_product_impl<Transpose<const Rhs>, int(RhsUpLo)==Upper ? Lower : Upper, false,
+                             Transpose<const Lhs>, 0, true>::run(destT, a_rhs.transpose(), a_lhs.transpose(), alpha);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
new file mode 100644
index 0000000..38f23ac
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointMatrixVector_BLAS.h
@@ -0,0 +1,111 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Selfadjoint matrix-vector product functionality based on ?SYMV/HEMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
+#define EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements selfadjoint matrix-vector multiplication using BLAS
+**********************************************************************/
+
+// symv/hemv specialization
+
+template<typename Scalar, typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs>
+struct selfadjoint_matrix_vector_product_symv :
+  selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn> {};
+
+#define EIGEN_BLAS_SYMV_SPECIALIZE(Scalar) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,Specialized> { \
+static void run( \
+  Index size, const Scalar*  lhs, Index lhsStride, \
+  const Scalar* _rhs, Scalar* res, Scalar alpha) { \
+    enum {\
+      IsColMajor = StorageOrder==ColMajor \
+    }; \
+    if (IsColMajor == ConjugateLhs) {\
+      selfadjoint_matrix_vector_product<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs,BuiltIn>::run( \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
+    } else {\
+      selfadjoint_matrix_vector_product_symv<Scalar,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs>::run( \
+        size, lhs, lhsStride, _rhs, res, alpha);  \
+    }\
+  } \
+}; \
+
+EIGEN_BLAS_SYMV_SPECIALIZE(double)
+EIGEN_BLAS_SYMV_SPECIALIZE(float)
+EIGEN_BLAS_SYMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_SYMV_SPECIALIZE(scomplex)
+
+#define EIGEN_BLAS_SYMV_SPECIALIZATION(EIGTYPE,BLASTYPE,BLASFUNC) \
+template<typename Index, int StorageOrder, int UpLo, bool ConjugateLhs, bool ConjugateRhs> \
+struct selfadjoint_matrix_vector_product_symv<EIGTYPE,Index,StorageOrder,UpLo,ConjugateLhs,ConjugateRhs> \
+{ \
+typedef Matrix<EIGTYPE,Dynamic,1,ColMajor> SYMVVector;\
+\
+static void run( \
+Index size, const EIGTYPE*  lhs, Index lhsStride, \
+const EIGTYPE* _rhs, EIGTYPE* res, EIGTYPE alpha) \
+{ \
+  enum {\
+    IsRowMajor = StorageOrder==RowMajor ? 1 : 0, \
+    IsLower = UpLo == Lower ? 1 : 0 \
+  }; \
+  BlasIndex n=convert_index<BlasIndex>(size), lda=convert_index<BlasIndex>(lhsStride), incx=1, incy=1; \
+  EIGTYPE beta(1); \
+  const EIGTYPE *x_ptr; \
+  char uplo=(IsRowMajor) ? (IsLower ? 'U' : 'L') : (IsLower ? 'L' : 'U'); \
+  SYMVVector x_tmp; \
+  if (ConjugateRhs) { \
+    Map<const SYMVVector, 0 > map_x(_rhs,size,1); \
+    x_tmp=map_x.conjugate(); \
+    x_ptr=x_tmp.data(); \
+  } else x_ptr=_rhs; \
+  BLASFUNC(&uplo, &n, &numext::real_ref(alpha), (const BLASTYPE*)lhs, &lda, (const BLASTYPE*)x_ptr, &incx, &numext::real_ref(beta), (BLASTYPE*)res, &incy); \
+}\
+};
+
+EIGEN_BLAS_SYMV_SPECIALIZATION(double,   double, dsymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(float,    float,  ssymv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(dcomplex, double, zhemv_)
+EIGEN_BLAS_SYMV_SPECIALIZATION(scomplex, float,  chemv_)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_MATRIX_VECTOR_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h
new file mode 100644
index 0000000..f038d68
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointProduct.h
@@ -0,0 +1,133 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINT_PRODUCT_H
+#define EIGEN_SELFADJOINT_PRODUCT_H
+
+/**********************************************************************
+* This file implements a self adjoint product: C += A A^T updating only
+* half of the selfadjoint matrix C.
+* It corresponds to the level 3 SYRK and level 2 SYR Blas routines.
+**********************************************************************/
+
+namespace Eigen { 
+
+
+template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo,ConjLhs,ConjRhs>
+{
+  static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha)
+  {
+    internal::conj_if<ConjRhs> cj;
+    typedef Map<const Matrix<Scalar,Dynamic,1> > OtherMap;
+    typedef typename internal::conditional<ConjLhs,typename OtherMap::ConjugateReturnType,const OtherMap&>::type ConjLhsType;
+    for (Index i=0; i<size; ++i)
+    {
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+(UpLo==Lower ? i : 0), (UpLo==Lower ? size-i : (i+1)))
+          += (alpha * cj(vecY[i])) * ConjLhsType(OtherMap(vecX+(UpLo==Lower ? i : 0),UpLo==Lower ? size-i : (i+1)));
+    }
+  }
+};
+
+template<typename Scalar, typename Index, int UpLo, bool ConjLhs, bool ConjRhs>
+struct selfadjoint_rank1_update<Scalar,Index,RowMajor,UpLo,ConjLhs,ConjRhs>
+{
+  static void run(Index size, Scalar* mat, Index stride, const Scalar* vecX, const Scalar* vecY, const Scalar& alpha)
+  {
+    selfadjoint_rank1_update<Scalar,Index,ColMajor,UpLo==Lower?Upper:Lower,ConjRhs,ConjLhs>::run(size,mat,stride,vecY,vecX,alpha);
+  }
+};
+
+template<typename MatrixType, typename OtherType, int UpLo, bool OtherIsVector = OtherType::IsVectorAtCompileTime>
+struct selfadjoint_product_selector;
+
+template<typename MatrixType, typename OtherType, int UpLo>
+struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,true>
+{
+  static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef internal::blas_traits<OtherType> OtherBlasTraits;
+    typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
+    typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
+
+    Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
+
+    enum {
+      StorageOrder = (internal::traits<MatrixType>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+      UseOtherDirectly = _ActualOtherType::InnerStrideAtCompileTime==1
+    };
+    internal::gemv_static_vector_if<Scalar,OtherType::SizeAtCompileTime,OtherType::MaxSizeAtCompileTime,!UseOtherDirectly> static_other;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, actualOtherPtr, other.size(),
+      (UseOtherDirectly ? const_cast<Scalar*>(actualOther.data()) : static_other.data()));
+      
+    if(!UseOtherDirectly)
+      Map<typename _ActualOtherType::PlainObject>(actualOtherPtr, actualOther.size()) = actualOther;
+    
+    selfadjoint_rank1_update<Scalar,Index,StorageOrder,UpLo,
+                              OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
+                            (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex>
+          ::run(other.size(), mat.data(), mat.outerStride(), actualOtherPtr, actualOtherPtr, actualAlpha);
+  }
+};
+
+template<typename MatrixType, typename OtherType, int UpLo>
+struct selfadjoint_product_selector<MatrixType,OtherType,UpLo,false>
+{
+  static void run(MatrixType& mat, const OtherType& other, const typename MatrixType::Scalar& alpha)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    typedef internal::blas_traits<OtherType> OtherBlasTraits;
+    typedef typename OtherBlasTraits::DirectLinearAccessType ActualOtherType;
+    typedef typename internal::remove_all<ActualOtherType>::type _ActualOtherType;
+    typename internal::add_const_on_value_type<ActualOtherType>::type actualOther = OtherBlasTraits::extract(other.derived());
+
+    Scalar actualAlpha = alpha * OtherBlasTraits::extractScalarFactor(other.derived());
+
+    enum {
+      IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0,
+      OtherIsRowMajor = _ActualOtherType::Flags&RowMajorBit ? 1 : 0
+    };
+
+    Index size = mat.cols();
+    Index depth = actualOther.cols();
+
+    typedef internal::gemm_blocking_space<IsRowMajor ? RowMajor : ColMajor,Scalar,Scalar,
+              MatrixType::MaxColsAtCompileTime, MatrixType::MaxColsAtCompileTime, _ActualOtherType::MaxColsAtCompileTime> BlockingType;
+
+    BlockingType blocking(size, size, depth, 1, false);
+
+
+    internal::general_matrix_matrix_triangular_product<Index,
+      Scalar, OtherIsRowMajor ? RowMajor : ColMajor,   OtherBlasTraits::NeedToConjugate  && NumTraits<Scalar>::IsComplex,
+      Scalar, OtherIsRowMajor ? ColMajor : RowMajor, (!OtherBlasTraits::NeedToConjugate) && NumTraits<Scalar>::IsComplex,
+      IsRowMajor ? RowMajor : ColMajor, UpLo>
+      ::run(size, depth,
+            &actualOther.coeffRef(0,0), actualOther.outerStride(), &actualOther.coeffRef(0,0), actualOther.outerStride(),
+            mat.data(), mat.outerStride(), actualAlpha, blocking);
+  }
+};
+
+// high level API
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename DerivedU>
+SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+::rankUpdate(const MatrixBase<DerivedU>& u, const Scalar& alpha)
+{
+  selfadjoint_product_selector<MatrixType,DerivedU,UpLo>::run(_expression().const_cast_derived(), u.derived(), alpha);
+
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINT_PRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h
new file mode 100644
index 0000000..2ae3641
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/SelfadjointRank2Update.h
@@ -0,0 +1,93 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINTRANK2UPTADE_H
+#define EIGEN_SELFADJOINTRANK2UPTADE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/* Optimized selfadjoint matrix += alpha * uv' + conj(alpha)*vu'
+ * It corresponds to the Level2 syr2 BLAS routine
+ */
+
+template<typename Scalar, typename Index, typename UType, typename VType, int UpLo>
+struct selfadjoint_rank2_update_selector;
+
+template<typename Scalar, typename Index, typename UType, typename VType>
+struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Lower>
+{
+  static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
+  {
+    const Index size = u.size();
+    for (Index i=0; i<size; ++i)
+    {
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i+i, size-i) +=
+                        (numext::conj(alpha) * numext::conj(u.coeff(i))) * v.tail(size-i)
+                      + (alpha * numext::conj(v.coeff(i))) * u.tail(size-i);
+    }
+  }
+};
+
+template<typename Scalar, typename Index, typename UType, typename VType>
+struct selfadjoint_rank2_update_selector<Scalar,Index,UType,VType,Upper>
+{
+  static void run(Scalar* mat, Index stride, const UType& u, const VType& v, const Scalar& alpha)
+  {
+    const Index size = u.size();
+    for (Index i=0; i<size; ++i)
+      Map<Matrix<Scalar,Dynamic,1> >(mat+stride*i, i+1) +=
+                        (numext::conj(alpha)  * numext::conj(u.coeff(i))) * v.head(i+1)
+                      + (alpha * numext::conj(v.coeff(i))) * u.head(i+1);
+  }
+};
+
+template<bool Cond, typename T> struct conj_expr_if
+  : conditional<!Cond, const T&,
+      CwiseUnaryOp<scalar_conjugate_op<typename traits<T>::Scalar>,T> > {};
+
+} // end namespace internal
+
+template<typename MatrixType, unsigned int UpLo>
+template<typename DerivedU, typename DerivedV>
+SelfAdjointView<MatrixType,UpLo>& SelfAdjointView<MatrixType,UpLo>
+::rankUpdate(const MatrixBase<DerivedU>& u, const MatrixBase<DerivedV>& v, const Scalar& alpha)
+{
+  typedef internal::blas_traits<DerivedU> UBlasTraits;
+  typedef typename UBlasTraits::DirectLinearAccessType ActualUType;
+  typedef typename internal::remove_all<ActualUType>::type _ActualUType;
+  typename internal::add_const_on_value_type<ActualUType>::type actualU = UBlasTraits::extract(u.derived());
+
+  typedef internal::blas_traits<DerivedV> VBlasTraits;
+  typedef typename VBlasTraits::DirectLinearAccessType ActualVType;
+  typedef typename internal::remove_all<ActualVType>::type _ActualVType;
+  typename internal::add_const_on_value_type<ActualVType>::type actualV = VBlasTraits::extract(v.derived());
+
+  // If MatrixType is row major, then we use the routine for lower triangular in the upper triangular case and
+  // vice versa, and take the complex conjugate of all coefficients and vector entries.
+
+  enum { IsRowMajor = (internal::traits<MatrixType>::Flags&RowMajorBit) ? 1 : 0 };
+  Scalar actualAlpha = alpha * UBlasTraits::extractScalarFactor(u.derived())
+                             * numext::conj(VBlasTraits::extractScalarFactor(v.derived()));
+  if (IsRowMajor)
+    actualAlpha = numext::conj(actualAlpha);
+
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ UBlasTraits::NeedToConjugate,_ActualUType>::type>::type UType;
+  typedef typename internal::remove_all<typename internal::conj_expr_if<IsRowMajor ^ VBlasTraits::NeedToConjugate,_ActualVType>::type>::type VType;
+  internal::selfadjoint_rank2_update_selector<Scalar, Index, UType, VType,
+    (IsRowMajor ? int(UpLo==Upper ? Lower : Upper) : UpLo)>
+    ::run(_expression().const_cast_derived().data(),_expression().outerStride(),UType(actualU),VType(actualV),actualAlpha);
+
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINTRANK2UPTADE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h
new file mode 100644
index 0000000..6ec5a8a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix.h
@@ -0,0 +1,441 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// template<typename Scalar, int mr, int StorageOrder, bool Conjugate, int Mode>
+// struct gemm_pack_lhs_triangular
+// {
+//   Matrix<Scalar,mr,mr,
+//   void operator()(Scalar* blockA, const EIGEN_RESTRICT Scalar* _lhs, int lhsStride, int depth, int rows)
+//   {
+//     conj_if<NumTraits<Scalar>::IsComplex && Conjugate> cj;
+//     const_blas_data_mapper<Scalar, StorageOrder> lhs(_lhs,lhsStride);
+//     int count = 0;
+//     const int peeled_mc = (rows/mr)*mr;
+//     for(int i=0; i<peeled_mc; i+=mr)
+//     {
+//       for(int k=0; k<depth; k++)
+//         for(int w=0; w<mr; w++)
+//           blockA[count++] = cj(lhs(i+w, k));
+//     }
+//     for(int i=peeled_mc; i<rows; i++)
+//     {
+//       for(int k=0; k<depth; k++)
+//         blockA[count++] = cj(lhs(i, k));
+//     }
+//   }
+// };
+
+/* Optimized triangular matrix * matrix (_TRMM++) product built on top of
+ * the general matrix matrix product.
+ */
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResStorageOrder, int Version = Specialized>
+struct product_triangular_matrix_matrix;
+
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,LhsIsTriangular,
+                                           LhsStorageOrder,ConjugateLhs,
+                                           RhsStorageOrder,ConjugateRhs,RowMajor,Version>
+{
+  static EIGEN_STRONG_INLINE void run(
+    Index rows, Index cols, Index depth,
+    const Scalar* lhs, Index lhsStride,
+    const Scalar* rhs, Index rhsStride,
+    Scalar* res,       Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    product_triangular_matrix_matrix<Scalar, Index,
+      (Mode&(UnitDiag|ZeroDiag)) | ((Mode&Upper) ? Lower : Upper),
+      (!LhsIsTriangular),
+      RhsStorageOrder==RowMajor ? ColMajor : RowMajor,
+      ConjugateRhs,
+      LhsStorageOrder==RowMajor ? ColMajor : RowMajor,
+      ConjugateLhs,
+      ColMajor>
+      ::run(cols, rows, depth, rhs, rhsStride, lhs, lhsStride, res, resStride, alpha, blocking);
+  }
+};
+
+// implements col-major += alpha * op(triangular) * op(general)
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,true,
+                                           LhsStorageOrder,ConjugateLhs,
+                                           RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+{
+  
+  typedef gebp_traits<Scalar,Scalar> Traits;
+  enum {
+    SmallPanelWidth   = 2 * EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+    IsLower = (Mode&Lower) == Lower,
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1
+  };
+
+  static EIGEN_DONT_INLINE void run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,true,
+                                                        LhsStorageOrder,ConjugateLhs,
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    // strip zeros
+    Index diagSize  = (std::min)(_rows,_depth);
+    Index rows      = IsLower ? _rows : diagSize;
+    Index depth     = IsLower ? diagSize : _depth;
+    Index cols      = _cols;
+    
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+    // The small panel size must not be larger than blocking size.
+    // Usually this should never be the case because SmallPanelWidth^2 is very small
+    // compared to L2 cache size, but let's be safe:
+    Index panelWidth = (std::min)(Index(SmallPanelWidth),(std::min)(kc,mc));
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,LhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
+    triangularBuffer.setZero();
+    if((Mode&ZeroDiag)==ZeroDiag)
+      triangularBuffer.diagonal().setZero();
+    else
+      triangularBuffer.diagonal().setOnes();
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+
+    for(Index k2=IsLower ? depth : 0;
+        IsLower ? k2>0 : k2<depth;
+        IsLower ? k2-=kc : k2+=kc)
+    {
+      Index actual_kc = (std::min)(IsLower ? k2 : depth-k2, kc);
+      Index actual_k2 = IsLower ? k2-actual_kc : k2;
+
+      // align blocks with the end of the triangular part for trapezoidal lhs
+      if((!IsLower)&&(k2<rows)&&(k2+actual_kc>rows))
+      {
+        actual_kc = rows-k2;
+        k2 = k2+actual_kc-kc;
+      }
+
+      pack_rhs(blockB, rhs.getSubMapper(actual_k2,0), actual_kc, cols);
+
+      // the selected lhs's panel has to be split in three different parts:
+      //  1 - the part which is zero => skip it
+      //  2 - the diagonal block => special kernel
+      //  3 - the dense panel below (lower case) or above (upper case) the diagonal block => GEPP
+
+      // the block diagonal, if any:
+      if(IsLower || actual_k2<rows)
+      {
+        // for each small vertical panels of lhs
+        for (Index k1=0; k1<actual_kc; k1+=panelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-k1, panelWidth);
+          Index lengthTarget = IsLower ? actual_kc-k1-actualPanelWidth : k1;
+          Index startBlock   = actual_k2+k1;
+          Index blockBOffset = k1;
+
+          // => GEBP with the micro triangular block
+          // The trick is to pack this micro block while filling the opposite triangular part with zeros.
+          // To this end we do an extra triangular copy to a small temporary buffer
+          for (Index k=0;k<actualPanelWidth;++k)
+          {
+            if (SetDiag)
+              triangularBuffer.coeffRef(k,k) = lhs(startBlock+k,startBlock+k);
+            for (Index i=IsLower ? k+1 : 0; IsLower ? i<actualPanelWidth : i<k; ++i)
+              triangularBuffer.coeffRef(i,k) = lhs(startBlock+i,startBlock+k);
+          }
+          pack_lhs(blockA, LhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()), actualPanelWidth, actualPanelWidth);
+
+          gebp_kernel(res.getSubMapper(startBlock, 0), blockA, blockB,
+                      actualPanelWidth, actualPanelWidth, cols, alpha,
+                      actualPanelWidth, actual_kc, 0, blockBOffset);
+
+          // GEBP with remaining micro panel
+          if (lengthTarget>0)
+          {
+            Index startTarget  = IsLower ? actual_k2+k1+actualPanelWidth : actual_k2;
+
+            pack_lhs(blockA, lhs.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
+
+            gebp_kernel(res.getSubMapper(startTarget, 0), blockA, blockB,
+                        lengthTarget, actualPanelWidth, cols, alpha,
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
+          }
+        }
+      }
+      // the part below (lower case) or above (upper case) the diagonal => GEPP
+      {
+        Index start = IsLower ? k2 : 0;
+        Index end   = IsLower ? rows : (std::min)(actual_k2,rows);
+        for(Index i2=start; i2<end; i2+=mc)
+        {
+          const Index actual_mc = (std::min)(i2+mc,end)-i2;
+          gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr,Traits::LhsProgress, LhsStorageOrder,false>()
+            (blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
+
+          gebp_kernel(res.getSubMapper(i2, 0), blockA, blockB, actual_mc,
+                      actual_kc, cols, alpha, -1, -1, 0, 0);
+        }
+      }
+    }
+  }
+
+// implements col-major += alpha * op(general) * op(triangular)
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+struct product_triangular_matrix_matrix<Scalar,Index,Mode,false,
+                                        LhsStorageOrder,ConjugateLhs,
+                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>
+{
+  typedef gebp_traits<Scalar,Scalar> Traits;
+  enum {
+    SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+    IsLower = (Mode&Lower) == Lower,
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1
+  };
+
+  static EIGEN_DONT_INLINE void run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking);
+};
+
+template <typename Scalar, typename Index, int Mode,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs, int Version>
+EIGEN_DONT_INLINE void product_triangular_matrix_matrix<Scalar,Index,Mode,false,
+                                                        LhsStorageOrder,ConjugateLhs,
+                                                        RhsStorageOrder,ConjugateRhs,ColMajor,Version>::run(
+    Index _rows, Index _cols, Index _depth,
+    const Scalar* _lhs, Index lhsStride,
+    const Scalar* _rhs, Index rhsStride,
+    Scalar* _res,        Index resStride,
+    const Scalar& alpha, level3_blocking<Scalar,Scalar>& blocking)
+  {
+    const Index PacketBytes = packet_traits<Scalar>::size*sizeof(Scalar);
+    // strip zeros
+    Index diagSize  = (std::min)(_cols,_depth);
+    Index rows      = _rows;
+    Index depth     = IsLower ? _depth : diagSize;
+    Index cols      = IsLower ? diagSize : _cols;
+    
+    typedef const_blas_data_mapper<Scalar, Index, LhsStorageOrder> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, RhsStorageOrder> RhsMapper;
+    typedef blas_data_mapper<typename Traits::ResScalar, Index, ColMajor> ResMapper;
+    LhsMapper lhs(_lhs,lhsStride);
+    RhsMapper rhs(_rhs,rhsStride);
+    ResMapper res(_res, resStride);
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols+EIGEN_MAX_ALIGN_BYTES/sizeof(Scalar);
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    Matrix<Scalar,SmallPanelWidth,SmallPanelWidth,RhsStorageOrder> triangularBuffer((internal::constructor_without_unaligned_array_assert()));
+    triangularBuffer.setZero();
+    if((Mode&ZeroDiag)==ZeroDiag)
+      triangularBuffer.diagonal().setZero();
+    else
+      triangularBuffer.diagonal().setOnes();
+
+    gebp_kernel<Scalar, Scalar, Index, ResMapper, Traits::mr, Traits::nr, ConjugateLhs, ConjugateRhs> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, LhsStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder> pack_rhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr,RhsStorageOrder,false,true> pack_rhs_panel;
+
+    for(Index k2=IsLower ? 0 : depth;
+        IsLower ? k2<depth  : k2>0;
+        IsLower ? k2+=kc   : k2-=kc)
+    {
+      Index actual_kc = (std::min)(IsLower ? depth-k2 : k2, kc);
+      Index actual_k2 = IsLower ? k2 : k2-actual_kc;
+
+      // align blocks with the end of the triangular part for trapezoidal rhs
+      if(IsLower && (k2<cols) && (actual_k2+actual_kc>cols))
+      {
+        actual_kc = cols-k2;
+        k2 = actual_k2 + actual_kc - kc;
+      }
+
+      // remaining size
+      Index rs = IsLower ? (std::min)(cols,actual_k2) : cols - k2;
+      // size of the triangular part
+      Index ts = (IsLower && actual_k2>=cols) ? 0 : actual_kc;
+
+      Scalar* geb = blockB+ts*ts;
+      geb = geb + internal::first_aligned<PacketBytes>(geb,PacketBytes/sizeof(Scalar));
+
+      pack_rhs(geb, rhs.getSubMapper(actual_k2,IsLower ? 0 : k2), actual_kc, rs);
+
+      // pack the triangular part of the rhs padding the unrolled blocks with zeros
+      if(ts>0)
+      {
+        for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+          Index actual_j2 = actual_k2 + j2;
+          Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
+          Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
+          // general part
+          pack_rhs_panel(blockB+j2*actual_kc,
+                         rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
+                         panelLength, actualPanelWidth,
+                         actual_kc, panelOffset);
+
+          // append the triangular part via a temporary buffer
+          for (Index j=0;j<actualPanelWidth;++j)
+          {
+            if (SetDiag)
+              triangularBuffer.coeffRef(j,j) = rhs(actual_j2+j,actual_j2+j);
+            for (Index k=IsLower ? j+1 : 0; IsLower ? k<actualPanelWidth : k<j; ++k)
+              triangularBuffer.coeffRef(k,j) = rhs(actual_j2+k,actual_j2+j);
+          }
+
+          pack_rhs_panel(blockB+j2*actual_kc,
+                         RhsMapper(triangularBuffer.data(), triangularBuffer.outerStride()),
+                         actualPanelWidth, actualPanelWidth,
+                         actual_kc, j2);
+        }
+      }
+
+      for (Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(mc,rows-i2);
+        pack_lhs(blockA, lhs.getSubMapper(i2, actual_k2), actual_kc, actual_mc);
+
+        // triangular kernel
+        if(ts>0)
+        {
+          for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
+          {
+            Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+            Index panelLength = IsLower ? actual_kc-j2 : j2+actualPanelWidth;
+            Index blockOffset = IsLower ? j2 : 0;
+
+            gebp_kernel(res.getSubMapper(i2, actual_k2 + j2),
+                        blockA, blockB+j2*actual_kc,
+                        actual_mc, panelLength, actualPanelWidth,
+                        alpha,
+                        actual_kc, actual_kc,  // strides
+                        blockOffset, blockOffset);// offsets
+          }
+        }
+        gebp_kernel(res.getSubMapper(i2, IsLower ? 0 : k2),
+                    blockA, geb, actual_mc, actual_kc, rs,
+                    alpha,
+                    -1, -1, 0, 0);
+      }
+    }
+  }
+
+/***************************************************************************
+* Wrapper to product_triangular_matrix_matrix
+***************************************************************************/
+
+} // end namespace internal
+
+namespace internal {
+template<int Mode, bool LhsIsTriangular, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,LhsIsTriangular,Lhs,false,Rhs,false>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &a_lhs, const Rhs &a_rhs, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Dest::Scalar     Scalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef typename internal::remove_all<ActualLhsType>::type ActualLhsTypeCleaned;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+    
+    typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(a_lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(a_rhs);
+
+    Scalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(a_lhs)
+                               * RhsBlasTraits::extractScalarFactor(a_rhs);
+
+    typedef internal::gemm_blocking_space<(Dest::Flags&RowMajorBit) ? RowMajor : ColMajor,Scalar,Scalar,
+              Lhs::MaxRowsAtCompileTime, Rhs::MaxColsAtCompileTime, Lhs::MaxColsAtCompileTime,4> BlockingType;
+
+    enum { IsLower = (Mode&Lower) == Lower };
+    Index stripedRows  = ((!LhsIsTriangular) || (IsLower))  ? lhs.rows() : (std::min)(lhs.rows(),lhs.cols());
+    Index stripedCols  = ((LhsIsTriangular)  || (!IsLower)) ? rhs.cols() : (std::min)(rhs.cols(),rhs.rows());
+    Index stripedDepth = LhsIsTriangular ? ((!IsLower) ? lhs.cols() : (std::min)(lhs.cols(),lhs.rows()))
+                                         : ((IsLower)  ? rhs.rows() : (std::min)(rhs.rows(),rhs.cols()));
+
+    BlockingType blocking(stripedRows, stripedCols, stripedDepth, 1, false);
+
+    internal::product_triangular_matrix_matrix<Scalar, Index,
+      Mode, LhsIsTriangular,
+      (internal::traits<ActualLhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, LhsBlasTraits::NeedToConjugate,
+      (internal::traits<ActualRhsTypeCleaned>::Flags&RowMajorBit) ? RowMajor : ColMajor, RhsBlasTraits::NeedToConjugate,
+      (internal::traits<Dest          >::Flags&RowMajorBit) ? RowMajor : ColMajor>
+      ::run(
+        stripedRows, stripedCols, stripedDepth,   // sizes
+        &lhs.coeffRef(0,0), lhs.outerStride(),    // lhs info
+        &rhs.coeffRef(0,0), rhs.outerStride(),    // rhs info
+        &dst.coeffRef(0,0), dst.outerStride(),    // result info
+        actualAlpha, blocking
+      );
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
new file mode 100644
index 0000000..aecded6
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixMatrix_BLAS.h
@@ -0,0 +1,302 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
+#define EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+
+template <typename Scalar, typename Index,
+          int Mode, bool LhsIsTriangular,
+          int LhsStorageOrder, bool ConjugateLhs,
+          int RhsStorageOrder, bool ConjugateRhs,
+          int ResStorageOrder>
+struct product_triangular_matrix_matrix_trmm :
+       product_triangular_matrix_matrix<Scalar,Index,Mode,
+          LhsIsTriangular,LhsStorageOrder,ConjugateLhs,
+          RhsStorageOrder, ConjugateRhs, ResStorageOrder, BuiltIn> {};
+
+
+// try to go to BLAS specialization
+#define EIGEN_BLAS_TRMM_SPECIALIZE(Scalar, LhsIsTriangular) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix<Scalar,Index, Mode, LhsIsTriangular, \
+           LhsStorageOrder,ConjugateLhs, RhsStorageOrder,ConjugateRhs,ColMajor,Specialized> { \
+  static inline void run(Index _rows, Index _cols, Index _depth, const Scalar* _lhs, Index lhsStride,\
+    const Scalar* _rhs, Index rhsStride, Scalar* res, Index resStride, Scalar alpha, level3_blocking<Scalar,Scalar>& blocking) { \
+      product_triangular_matrix_matrix_trmm<Scalar,Index,Mode, \
+        LhsIsTriangular,LhsStorageOrder,ConjugateLhs, \
+        RhsStorageOrder, ConjugateRhs, ColMajor>::run( \
+        _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+  } \
+};
+
+EIGEN_BLAS_TRMM_SPECIALIZE(double, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(double, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(dcomplex, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(float, false)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, true)
+EIGEN_BLAS_TRMM_SPECIALIZE(scomplex, false)
+
+// implements col-major += alpha * op(triangular) * op(general)
+#define EIGEN_BLAS_TRMM_L(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((LhsStorageOrder==ColMajor) && ConjugateLhs) ? 1 : 0 \
+  }; \
+\
+  static void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
+  { \
+   Index diagSize  = (std::min)(_rows,_depth); \
+   Index rows      = IsLower ? _rows : diagSize; \
+   Index depth     = IsLower ? diagSize : _depth; \
+   Index cols      = _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
+   if (rows != depth) { \
+\
+     /* FIXME handle mkl_domain_get_max_threads */ \
+     /*int nthr = mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS);*/ int nthr = 1;\
+\
+     if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS */ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,true, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+     /*std::cout << "TRMM_L: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixLhs, 0, OuterStride<> > lhsMap(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+       MatrixLhs aa_tmp=lhsMap.template triangularView<Mode>(); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, aa_tmp.data(), aStride, _rhs, rhsStride, res, resStride, alpha, gemm_blocking, 0); \
+\
+     /*std::cout << "TRMM_L: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'L', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   BlasIndex m, n, lda, ldb; \
+\
+/* Set m, n */ \
+   m = convert_index<BlasIndex>(diagSize); \
+   n = convert_index<BlasIndex>(cols); \
+\
+/* Set trans */ \
+   transa = (LhsStorageOrder==RowMajor) ? ((ConjugateLhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols,OuterStride<>(rhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateRhs) b_tmp = rhs.conjugate(); else b_tmp = rhs; \
+   b = b_tmp.data(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (LhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixLhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = lhs.conjugate(); else a_tmp = lhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _lhs; \
+     lda = convert_index<BlasIndex>(lhsStride); \
+   } \
+   /*std::cout << "TRMM_L: A is square! Go to BLAS TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+EIGEN_BLAS_TRMM_L(double, double, d, d)
+EIGEN_BLAS_TRMM_L(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMM_L(float, float, f, s)
+EIGEN_BLAS_TRMM_L(scomplex, float, cf, c)
+
+// implements col-major += alpha * op(general) * op(triangular)
+#define EIGEN_BLAS_TRMM_R(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template <typename Index, int Mode, \
+          int LhsStorageOrder, bool ConjugateLhs, \
+          int RhsStorageOrder, bool ConjugateRhs> \
+struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
+         LhsStorageOrder,ConjugateLhs,RhsStorageOrder,ConjugateRhs,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper, \
+    conjA = ((RhsStorageOrder==ColMajor) && ConjugateRhs) ? 1 : 0 \
+  }; \
+\
+  static void run( \
+    Index _rows, Index _cols, Index _depth, \
+    const EIGTYPE* _lhs, Index lhsStride, \
+    const EIGTYPE* _rhs, Index rhsStride, \
+    EIGTYPE* res,        Index resStride, \
+    EIGTYPE alpha, level3_blocking<EIGTYPE,EIGTYPE>& blocking) \
+  { \
+   Index diagSize  = (std::min)(_cols,_depth); \
+   Index rows      = _rows; \
+   Index depth     = IsLower ? _depth : diagSize; \
+   Index cols      = IsLower ? diagSize : _cols; \
+\
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, LhsStorageOrder> MatrixLhs; \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, RhsStorageOrder> MatrixRhs; \
+\
+/* Non-square case - doesn't fit to BLAS ?TRMM. Fall to default triangular product or call BLAS ?GEMM*/ \
+   if (cols != depth) { \
+\
+     int nthr = 1 /*mkl_domain_get_max_threads(EIGEN_BLAS_DOMAIN_BLAS)*/; \
+\
+     if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
+     /* Most likely no benefit to call TRMM or GEMM from BLAS*/ \
+       product_triangular_matrix_matrix<EIGTYPE,Index,Mode,false, \
+       LhsStorageOrder,ConjugateLhs, RhsStorageOrder, ConjugateRhs, ColMajor, BuiltIn>::run( \
+           _rows, _cols, _depth, _lhs, lhsStride, _rhs, rhsStride, res, resStride, alpha, blocking); \
+       /*std::cout << "TRMM_R: A is not square! Go to Eigen TRMM implementation!\n";*/ \
+     } else { \
+     /* Make sense to call GEMM */ \
+       Map<const MatrixRhs, 0, OuterStride<> > rhsMap(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+       MatrixRhs aa_tmp=rhsMap.template triangularView<Mode>(); \
+       BlasIndex aStride = convert_index<BlasIndex>(aa_tmp.outerStride()); \
+       gemm_blocking_space<ColMajor,EIGTYPE,EIGTYPE,Dynamic,Dynamic,Dynamic> gemm_blocking(_rows,_cols,_depth, 1, true); \
+       general_matrix_matrix_product<Index,EIGTYPE,LhsStorageOrder,ConjugateLhs,EIGTYPE,RhsStorageOrder,ConjugateRhs,ColMajor>::run( \
+       rows, cols, depth, _lhs, lhsStride, aa_tmp.data(), aStride, res, resStride, alpha, gemm_blocking, 0); \
+\
+     /*std::cout << "TRMM_R: A is not square! Go to BLAS GEMM implementation! " << nthr<<" \n";*/ \
+     } \
+     return; \
+   } \
+   char side = 'R', transa, uplo, diag = 'N'; \
+   EIGTYPE *b; \
+   const EIGTYPE *a; \
+   BlasIndex m, n, lda, ldb; \
+\
+/* Set m, n */ \
+   m = convert_index<BlasIndex>(rows); \
+   n = convert_index<BlasIndex>(diagSize); \
+\
+/* Set trans */ \
+   transa = (RhsStorageOrder==RowMajor) ? ((ConjugateRhs) ? 'C' : 'T') : 'N'; \
+\
+/* Set b, ldb */ \
+   Map<const MatrixLhs, 0, OuterStride<> > lhs(_lhs,rows,depth,OuterStride<>(lhsStride)); \
+   MatrixX##EIGPREFIX b_tmp; \
+\
+   if (ConjugateLhs) b_tmp = lhs.conjugate(); else b_tmp = lhs; \
+   b = b_tmp.data(); \
+   ldb = convert_index<BlasIndex>(b_tmp.outerStride()); \
+\
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (RhsStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   Map<const MatrixRhs, 0, OuterStride<> > rhs(_rhs,depth,cols, OuterStride<>(rhsStride)); \
+   MatrixRhs a_tmp; \
+\
+   if ((conjA!=0) || (SetDiag==0)) { \
+     if (conjA) a_tmp = rhs.conjugate(); else a_tmp = rhs; \
+     if (IsZeroDiag) \
+       a_tmp.diagonal().setZero(); \
+     else if (IsUnitDiag) \
+       a_tmp.diagonal().setOnes();\
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _rhs; \
+     lda = convert_index<BlasIndex>(rhsStride); \
+   } \
+   /*std::cout << "TRMM_R: A is square! Go to BLAS TRMM implementation! \n";*/ \
+/* call ?trmm*/ \
+   BLASPREFIX##trmm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)b, &ldb); \
+\
+/* Add op(a_triangular)*b into res*/ \
+   Map<MatrixX##EIGPREFIX, 0, OuterStride<> > res_tmp(res,rows,cols,OuterStride<>(resStride)); \
+   res_tmp=res_tmp+b_tmp; \
+  } \
+};
+
+EIGEN_BLAS_TRMM_R(double, double, d, d)
+EIGEN_BLAS_TRMM_R(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMM_R(float, float, f, s)
+EIGEN_BLAS_TRMM_R(scomplex, float, cf, c)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_MATRIX_MATRIX_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h
new file mode 100644
index 0000000..4b292e7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector.h
@@ -0,0 +1,336 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIANGULARMATRIXVECTOR_H
+#define EIGEN_TRIANGULARMATRIXVECTOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder, int Version=Specialized>
+struct triangular_matrix_vector_product;
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
+  };
+  static EIGEN_DONT_INLINE  void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+                                     const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha);
+};
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int Version>
+EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,ColMajor,Version>
+  ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const RhsScalar& alpha)
+  {
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index size = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : (std::min)(_rows,_cols);
+    Index cols = IsLower ? (std::min)(_rows,_cols) : _cols;
+
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
+    typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs);
+
+    typedef Map<const Matrix<RhsScalar,Dynamic,1>, 0, InnerStride<> > RhsMap;
+    const RhsMap rhs(_rhs,cols,InnerStride<>(rhsIncr));
+    typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs);
+
+    typedef Map<Matrix<ResScalar,Dynamic,1> > ResMap;
+    ResMap res(_res,rows);
+
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
+    for (Index pi=0; pi<size; pi+=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(PanelWidth, size-pi);
+      for (Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = pi + k;
+        Index s = IsLower ? ((HasUnitDiag||HasZeroDiag) ? i+1 : i ) : pi;
+        Index r = IsLower ? actualPanelWidth-k : k+1;
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
+          res.segment(s,r) += (alpha * cjRhs.coeff(i)) * cjLhs.col(i).segment(s,r);
+        if (HasUnitDiag)
+          res.coeffRef(i) += alpha * cjRhs.coeff(i);
+      }
+      Index r = IsLower ? rows - pi - actualPanelWidth : pi;
+      if (r>0)
+      {
+        Index s = IsLower ? pi+actualPanelWidth : 0;
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
+            r, actualPanelWidth,
+            LhsMapper(&lhs.coeffRef(s,pi), lhsStride),
+            RhsMapper(&rhs.coeffRef(pi), rhsIncr),
+            &res.coeffRef(s), resIncr, alpha);
+      }
+    }
+    if((!IsLower) && cols>size)
+    {
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
+          rows, cols-size,
+          LhsMapper(&lhs.coeffRef(0,size), lhsStride),
+          RhsMapper(&rhs.coeffRef(size), rhsIncr),
+          _res, resIncr, alpha);
+    }
+  }
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
+struct triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar, RhsScalar>::ReturnType ResScalar;
+  enum {
+    IsLower = ((Mode&Lower)==Lower),
+    HasUnitDiag = (Mode & UnitDiag)==UnitDiag,
+    HasZeroDiag = (Mode & ZeroDiag)==ZeroDiag
+  };
+  static EIGEN_DONT_INLINE void run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+                                    const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha);
+};
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs,int Version>
+EIGEN_DONT_INLINE void triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,RowMajor,Version>
+  ::run(Index _rows, Index _cols, const LhsScalar* _lhs, Index lhsStride,
+        const RhsScalar* _rhs, Index rhsIncr, ResScalar* _res, Index resIncr, const ResScalar& alpha)
+  {
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    Index diagSize = (std::min)(_rows,_cols);
+    Index rows = IsLower ? _rows : diagSize;
+    Index cols = IsLower ? diagSize : _cols;
+
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,rows,cols,OuterStride<>(lhsStride));
+    typename conj_expr_if<ConjLhs,LhsMap>::type cjLhs(lhs);
+
+    typedef Map<const Matrix<RhsScalar,Dynamic,1> > RhsMap;
+    const RhsMap rhs(_rhs,cols);
+    typename conj_expr_if<ConjRhs,RhsMap>::type cjRhs(rhs);
+
+    typedef Map<Matrix<ResScalar,Dynamic,1>, 0, InnerStride<> > ResMap;
+    ResMap res(_res,rows,InnerStride<>(resIncr));
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,RowMajor> RhsMapper;
+
+    for (Index pi=0; pi<diagSize; pi+=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(PanelWidth, diagSize-pi);
+      for (Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = pi + k;
+        Index s = IsLower ? pi  : ((HasUnitDiag||HasZeroDiag) ? i+1 : i);
+        Index r = IsLower ? k+1 : actualPanelWidth-k;
+        if ((!(HasUnitDiag||HasZeroDiag)) || (--r)>0)
+          res.coeffRef(i) += alpha * (cjLhs.row(i).segment(s,r).cwiseProduct(cjRhs.segment(s,r).transpose())).sum();
+        if (HasUnitDiag)
+          res.coeffRef(i) += alpha * cjRhs.coeff(i);
+      }
+      Index r = IsLower ? pi : cols - pi - actualPanelWidth;
+      if (r>0)
+      {
+        Index s = IsLower ? 0 : pi + actualPanelWidth;
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs,BuiltIn>::run(
+            actualPanelWidth, r,
+            LhsMapper(&lhs.coeffRef(pi,s), lhsStride),
+            RhsMapper(&rhs.coeffRef(s), rhsIncr),
+            &res.coeffRef(pi), resIncr, alpha);
+      }
+    }
+    if(IsLower && rows>diagSize)
+    {
+      general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,ConjLhs,RhsScalar,RhsMapper,ConjRhs>::run(
+            rows-diagSize, cols,
+            LhsMapper(&lhs.coeffRef(diagSize,0), lhsStride),
+            RhsMapper(&rhs.coeffRef(0), rhsIncr),
+            &res.coeffRef(diagSize), resIncr, alpha);
+    }
+  }
+
+/***************************************************************************
+* Wrapper to product_triangular_vector
+***************************************************************************/
+
+template<int Mode,int StorageOrder>
+struct trmv_selector;
+
+} // end namespace internal
+
+namespace internal {
+
+template<int Mode, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,true,Lhs,false,Rhs,true>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
+  
+    internal::trmv_selector<Mode,(int(internal::traits<Lhs>::Flags)&RowMajorBit) ? RowMajor : ColMajor>::run(lhs, rhs, dst, alpha);
+  }
+};
+
+template<int Mode, typename Lhs, typename Rhs>
+struct triangular_product_impl<Mode,false,Lhs,true,Rhs,false>
+{
+  template<typename Dest> static void run(Dest& dst, const Lhs &lhs, const Rhs &rhs, const typename Dest::Scalar& alpha)
+  {
+    eigen_assert(dst.rows()==lhs.rows() && dst.cols()==rhs.cols());
+
+    Transpose<Dest> dstT(dst);
+    internal::trmv_selector<(Mode & (UnitDiag|ZeroDiag)) | ((Mode & Lower) ? Upper : Lower),
+                            (int(internal::traits<Rhs>::Flags)&RowMajorBit) ? ColMajor : RowMajor>
+            ::run(rhs.transpose(),lhs.transpose(), dstT, alpha);
+  }
+};
+
+} // end namespace internal
+
+namespace internal {
+
+// TODO: find a way to factorize this piece of code with gemv_selector since the logic is exactly the same.
+  
+template<int Mode> struct trmv_selector<Mode,ColMajor>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    typedef typename Dest::RealScalar RealScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    
+    typedef Map<Matrix<ResScalar,Dynamic,1>, EIGEN_PLAIN_ENUM_MIN(AlignedMax,internal::packet_traits<ResScalar>::size)> MappedDest;
+
+    typename internal::add_const_on_value_type<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename internal::add_const_on_value_type<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
+
+    enum {
+      // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
+      // on, the other hand it is good for the cache to pack the vector anyways...
+      EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1,
+      ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
+      MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal
+    };
+
+    gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+
+    bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
+    bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
+
+    RhsScalar compatibleAlpha = get_factor<ResScalar,RhsScalar>::run(actualAlpha);
+
+    ei_declare_aligned_stack_constructed_variable(ResScalar,actualDestPtr,dest.size(),
+                                                  evalToDest ? dest.data() : static_dest.data());
+
+    if(!evalToDest)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = dest.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      if(!alphaIsCompatible)
+      {
+        MappedDest(actualDestPtr, dest.size()).setZero();
+        compatibleAlpha = RhsScalar(1);
+      }
+      else
+        MappedDest(actualDestPtr, dest.size()) = dest;
+    }
+
+    internal::triangular_matrix_vector_product
+      <Index,Mode,
+       LhsScalar, LhsBlasTraits::NeedToConjugate,
+       RhsScalar, RhsBlasTraits::NeedToConjugate,
+       ColMajor>
+      ::run(actualLhs.rows(),actualLhs.cols(),
+            actualLhs.data(),actualLhs.outerStride(),
+            actualRhs.data(),actualRhs.innerStride(),
+            actualDestPtr,1,compatibleAlpha);
+
+    if (!evalToDest)
+    {
+      if(!alphaIsCompatible)
+        dest += actualAlpha * MappedDest(actualDestPtr, dest.size());
+      else
+        dest = MappedDest(actualDestPtr, dest.size());
+    }
+  }
+};
+
+template<int Mode> struct trmv_selector<Mode,RowMajor>
+{
+  template<typename Lhs, typename Rhs, typename Dest>
+  static void run(const Lhs &lhs, const Rhs &rhs, Dest& dest, const typename Dest::Scalar& alpha)
+  {
+    typedef typename Lhs::Scalar      LhsScalar;
+    typedef typename Rhs::Scalar      RhsScalar;
+    typedef typename Dest::Scalar     ResScalar;
+    
+    typedef internal::blas_traits<Lhs> LhsBlasTraits;
+    typedef typename LhsBlasTraits::DirectLinearAccessType ActualLhsType;
+    typedef internal::blas_traits<Rhs> RhsBlasTraits;
+    typedef typename RhsBlasTraits::DirectLinearAccessType ActualRhsType;
+    typedef typename internal::remove_all<ActualRhsType>::type ActualRhsTypeCleaned;
+
+    typename add_const<ActualLhsType>::type actualLhs = LhsBlasTraits::extract(lhs);
+    typename add_const<ActualRhsType>::type actualRhs = RhsBlasTraits::extract(rhs);
+
+    ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(lhs)
+                                  * RhsBlasTraits::extractScalarFactor(rhs);
+
+    enum {
+      DirectlyUseRhs = ActualRhsTypeCleaned::InnerStrideAtCompileTime==1
+    };
+
+    gemv_static_vector_if<RhsScalar,ActualRhsTypeCleaned::SizeAtCompileTime,ActualRhsTypeCleaned::MaxSizeAtCompileTime,!DirectlyUseRhs> static_rhs;
+
+    ei_declare_aligned_stack_constructed_variable(RhsScalar,actualRhsPtr,actualRhs.size(),
+        DirectlyUseRhs ? const_cast<RhsScalar*>(actualRhs.data()) : static_rhs.data());
+
+    if(!DirectlyUseRhs)
+    {
+      #ifdef EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      Index size = actualRhs.size();
+      EIGEN_DENSE_STORAGE_CTOR_PLUGIN
+      #endif
+      Map<typename ActualRhsTypeCleaned::PlainObject>(actualRhsPtr, actualRhs.size()) = actualRhs;
+    }
+
+    internal::triangular_matrix_vector_product
+      <Index,Mode,
+       LhsScalar, LhsBlasTraits::NeedToConjugate,
+       RhsScalar, RhsBlasTraits::NeedToConjugate,
+       RowMajor>
+      ::run(actualLhs.rows(),actualLhs.cols(),
+            actualLhs.data(),actualLhs.outerStride(),
+            actualRhsPtr,1,
+            dest.data(),dest.innerStride(),
+            actualAlpha);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULARMATRIXVECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
new file mode 100644
index 0000000..07bf26c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularMatrixVector_BLAS.h
@@ -0,0 +1,241 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Triangular matrix-vector product functionality based on ?TRMV.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
+#define EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************************************************
+* This file implements triangular matrix-vector multiplication using BLAS
+**********************************************************************/
+
+// trmv/hemv specialization
+
+template<typename Index, int Mode, typename LhsScalar, bool ConjLhs, typename RhsScalar, bool ConjRhs, int StorageOrder>
+struct triangular_matrix_vector_product_trmv :
+  triangular_matrix_vector_product<Index,Mode,LhsScalar,ConjLhs,RhsScalar,ConjRhs,StorageOrder,BuiltIn> {};
+
+#define EIGEN_BLAS_TRMV_SPECIALIZE(Scalar) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor,Specialized> { \
+ static void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \
+                                     const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \
+      triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,ColMajor>::run( \
+        _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+  } \
+}; \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor,Specialized> { \
+ static void run(Index _rows, Index _cols, const Scalar* _lhs, Index lhsStride, \
+                                     const Scalar* _rhs, Index rhsIncr, Scalar* _res, Index resIncr, Scalar alpha) { \
+      triangular_matrix_vector_product_trmv<Index,Mode,Scalar,ConjLhs,Scalar,ConjRhs,RowMajor>::run( \
+        _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+  } \
+};
+
+EIGEN_BLAS_TRMV_SPECIALIZE(double)
+EIGEN_BLAS_TRMV_SPECIALIZE(float)
+EIGEN_BLAS_TRMV_SPECIALIZE(dcomplex)
+EIGEN_BLAS_TRMV_SPECIALIZE(scomplex)
+
+// implements col-major: res += alpha * op(triangular) * vector
+#define EIGEN_BLAS_TRMV_CM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor> { \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper \
+  }; \
+ static void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
+                 const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
+ { \
+   if (ConjLhs || IsZeroDiag) { \
+     triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,ColMajor,BuiltIn>::run( \
+       _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+     return; \
+   }\
+   Index size = (std::min)(_rows,_cols); \
+   Index rows = IsLower ? _rows : size; \
+   Index cols = IsLower ? size : _cols; \
+\
+   typedef VectorX##EIGPREFIX VectorRhs; \
+   EIGTYPE *x, *y;\
+\
+/* Set x*/ \
+   Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \
+   VectorRhs x_tmp; \
+   if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+   x = x_tmp.data(); \
+\
+/* Square part handling */\
+\
+   char trans, uplo, diag; \
+   BlasIndex m, n, lda, incx, incy; \
+   EIGTYPE const *a; \
+   EIGTYPE beta(1); \
+\
+/* Set m, n */ \
+   n = convert_index<BlasIndex>(size); \
+   lda = convert_index<BlasIndex>(lhsStride); \
+   incx = 1; \
+   incy = convert_index<BlasIndex>(resIncr); \
+\
+/* Set uplo, trans and diag*/ \
+   trans = 'N'; \
+   uplo = IsLower ? 'L' : 'U'; \
+   diag = IsUnitDiag ? 'U' : 'N'; \
+\
+/* call ?TRMV*/ \
+   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
+\
+/* Add op(a_tr)rhs into res*/ \
+   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
+   if (size<(std::max)(rows,cols)) { \
+     if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+     x = x_tmp.data(); \
+     if (size<rows) { \
+       y = _res + size*resIncr; \
+       a = _lhs + size; \
+       m = convert_index<BlasIndex>(rows-size); \
+       n = convert_index<BlasIndex>(size); \
+     } \
+     else { \
+       x += size; \
+       y = _res; \
+       a = _lhs + size*lda; \
+       m = convert_index<BlasIndex>(size); \
+       n = convert_index<BlasIndex>(cols-size); \
+     } \
+     BLASPREFIX##gemv_(&trans, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
+   } \
+  } \
+};
+
+EIGEN_BLAS_TRMV_CM(double,   double, d,  d)
+EIGEN_BLAS_TRMV_CM(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMV_CM(float,    float,  f,  s)
+EIGEN_BLAS_TRMV_CM(scomplex, float,  cf, c)
+
+// implements row-major: res += alpha * op(triangular) * vector
+#define EIGEN_BLAS_TRMV_RM(EIGTYPE, BLASTYPE, EIGPREFIX, BLASPREFIX) \
+template<typename Index, int Mode, bool ConjLhs, bool ConjRhs> \
+struct triangular_matrix_vector_product_trmv<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor> { \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    SetDiag = (Mode&(ZeroDiag|UnitDiag)) ? 0 : 1, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    LowUp = IsLower ? Lower : Upper \
+  }; \
+ static void run(Index _rows, Index _cols, const EIGTYPE* _lhs, Index lhsStride, \
+                 const EIGTYPE* _rhs, Index rhsIncr, EIGTYPE* _res, Index resIncr, EIGTYPE alpha) \
+ { \
+   if (IsZeroDiag) { \
+     triangular_matrix_vector_product<Index,Mode,EIGTYPE,ConjLhs,EIGTYPE,ConjRhs,RowMajor,BuiltIn>::run( \
+       _rows, _cols, _lhs, lhsStride, _rhs, rhsIncr, _res, resIncr, alpha); \
+     return; \
+   }\
+   Index size = (std::min)(_rows,_cols); \
+   Index rows = IsLower ? _rows : size; \
+   Index cols = IsLower ? size : _cols; \
+\
+   typedef VectorX##EIGPREFIX VectorRhs; \
+   EIGTYPE *x, *y;\
+\
+/* Set x*/ \
+   Map<const VectorRhs, 0, InnerStride<> > rhs(_rhs,cols,InnerStride<>(rhsIncr)); \
+   VectorRhs x_tmp; \
+   if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+   x = x_tmp.data(); \
+\
+/* Square part handling */\
+\
+   char trans, uplo, diag; \
+   BlasIndex m, n, lda, incx, incy; \
+   EIGTYPE const *a; \
+   EIGTYPE beta(1); \
+\
+/* Set m, n */ \
+   n = convert_index<BlasIndex>(size); \
+   lda = convert_index<BlasIndex>(lhsStride); \
+   incx = 1; \
+   incy = convert_index<BlasIndex>(resIncr); \
+\
+/* Set uplo, trans and diag*/ \
+   trans = ConjLhs ? 'C' : 'T'; \
+   uplo = IsLower ? 'U' : 'L'; \
+   diag = IsUnitDiag ? 'U' : 'N'; \
+\
+/* call ?TRMV*/ \
+   BLASPREFIX##trmv_(&uplo, &trans, &diag, &n, (const BLASTYPE*)_lhs, &lda, (BLASTYPE*)x, &incx); \
+\
+/* Add op(a_tr)rhs into res*/ \
+   BLASPREFIX##axpy_(&n, &numext::real_ref(alpha),(const BLASTYPE*)x, &incx, (BLASTYPE*)_res, &incy); \
+/* Non-square case - doesn't fit to BLAS ?TRMV. Fall to default triangular product*/ \
+   if (size<(std::max)(rows,cols)) { \
+     if (ConjRhs) x_tmp = rhs.conjugate(); else x_tmp = rhs; \
+     x = x_tmp.data(); \
+     if (size<rows) { \
+       y = _res + size*resIncr; \
+       a = _lhs + size*lda; \
+       m = convert_index<BlasIndex>(rows-size); \
+       n = convert_index<BlasIndex>(size); \
+     } \
+     else { \
+       x += size; \
+       y = _res; \
+       a = _lhs + size; \
+       m = convert_index<BlasIndex>(size); \
+       n = convert_index<BlasIndex>(cols-size); \
+     } \
+     BLASPREFIX##gemv_(&trans, &n, &m, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (const BLASTYPE*)x, &incx, &numext::real_ref(beta), (BLASTYPE*)y, &incy); \
+   } \
+  } \
+};
+
+EIGEN_BLAS_TRMV_RM(double,   double, d,  d)
+EIGEN_BLAS_TRMV_RM(dcomplex, double, cd, z)
+EIGEN_BLAS_TRMV_RM(float,    float,  f,  s)
+EIGEN_BLAS_TRMV_RM(scomplex, float,  cf, c)
+
+} // end namespase internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_MATRIX_VECTOR_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h
new file mode 100644
index 0000000..223c38b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix.h
@@ -0,0 +1,335 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_H
+#define EIGEN_TRIANGULAR_SOLVER_MATRIX_H
+
+namespace Eigen { 
+
+namespace internal {
+
+// if the rhs is row major, let's transpose the product
+template <typename Scalar, typename Index, int Side, int Mode, bool Conjugate, int TriStorageOrder>
+struct triangular_solve_matrix<Scalar,Index,Side,Mode,Conjugate,TriStorageOrder,RowMajor>
+{
+  static void run(
+    Index size, Index cols,
+    const Scalar*  tri, Index triStride,
+    Scalar* _other, Index otherStride,
+    level3_blocking<Scalar,Scalar>& blocking)
+  {
+    triangular_solve_matrix<
+      Scalar, Index, Side==OnTheLeft?OnTheRight:OnTheLeft,
+      (Mode&UnitDiag) | ((Mode&Upper) ? Lower : Upper),
+      NumTraits<Scalar>::IsComplex && Conjugate,
+      TriStorageOrder==RowMajor ? ColMajor : RowMajor, ColMajor>
+      ::run(size, cols, tri, triStride, _other, otherStride, blocking);
+  }
+};
+
+/* Optimized triangular solver with multiple right hand side and the triangular matrix on the left
+ */
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
+struct triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor>
+{
+  static EIGEN_DONT_INLINE void run(
+    Index size, Index otherSize,
+    const Scalar* _tri, Index triStride,
+    Scalar* _other, Index otherStride,
+    level3_blocking<Scalar,Scalar>& blocking);
+};
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
+EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor>::run(
+    Index size, Index otherSize,
+    const Scalar* _tri, Index triStride,
+    Scalar* _other, Index otherStride,
+    level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index cols = otherSize;
+
+    typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> TriMapper;
+    typedef blas_data_mapper<Scalar, Index, ColMajor> OtherMapper;
+    TriMapper tri(_tri, triStride);
+    OtherMapper other(_other, otherStride);
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+
+    enum {
+      SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+      IsLower = (Mode&Lower) == Lower
+    };
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(size,blocking.mc());  // cache block size along the M direction
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*cols;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    conj_if<Conjugate> conj;
+    gebp_kernel<Scalar, Scalar, Index, OtherMapper, Traits::mr, Traits::nr, Conjugate, false> gebp_kernel;
+    gemm_pack_lhs<Scalar, Index, TriMapper, Traits::mr, Traits::LhsProgress, TriStorageOrder> pack_lhs;
+    gemm_pack_rhs<Scalar, Index, OtherMapper, Traits::nr, ColMajor, false, true> pack_rhs;
+
+    // the goal here is to subdivise the Rhs panels such that we keep some cache
+    // coherence when accessing the rhs elements
+    std::ptrdiff_t l1, l2, l3;
+    manage_caching_sizes(GetAction, &l1, &l2, &l3);
+    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * std::max<Index>(otherStride,size)) : 0;
+    subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr);
+
+    for(Index k2=IsLower ? 0 : size;
+        IsLower ? k2<size : k2>0;
+        IsLower ? k2+=kc : k2-=kc)
+    {
+      const Index actual_kc = (std::min)(IsLower ? size-k2 : k2, kc);
+
+      // We have selected and packed a big horizontal panel R1 of rhs. Let B be the packed copy of this panel,
+      // and R2 the remaining part of rhs. The corresponding vertical panel of lhs is split into
+      // A11 (the triangular part) and A21 the remaining rectangular part.
+      // Then the high level algorithm is:
+      //  - B = R1                    => general block copy (done during the next step)
+      //  - R1 = A11^-1 B             => tricky part
+      //  - update B from the new R1  => actually this has to be performed continuously during the above step
+      //  - R2 -= A21 * B             => GEPP
+
+      // The tricky part: compute R1 = A11^-1 B while updating B from R1
+      // The idea is to split A11 into multiple small vertical panels.
+      // Each panel can be split into a small triangular part T1k which is processed without optimization,
+      // and the remaining small part T2k which is processed using gebp with appropriate block strides
+      for(Index j2=0; j2<cols; j2+=subcols)
+      {
+        Index actual_cols = (std::min)(cols-j2,subcols);
+        // for each small vertical panels [T1k^T, T2k^T]^T of lhs
+        for (Index k1=0; k1<actual_kc; k1+=SmallPanelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-k1, SmallPanelWidth);
+          // tr solve
+          for (Index k=0; k<actualPanelWidth; ++k)
+          {
+            // TODO write a small kernel handling this (can be shared with trsv)
+            Index i  = IsLower ? k2+k1+k : k2-k1-k-1;
+            Index rs = actualPanelWidth - k - 1; // remaining size
+            Index s  = TriStorageOrder==RowMajor ? (IsLower ? k2+k1 : i+1)
+                                                 :  IsLower ? i+1 : i-rs;
+
+            Scalar a = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(tri(i,i));
+            for (Index j=j2; j<j2+actual_cols; ++j)
+            {
+              if (TriStorageOrder==RowMajor)
+              {
+                Scalar b(0);
+                const Scalar* l = &tri(i,s);
+                Scalar* r = &other(s,j);
+                for (Index i3=0; i3<k; ++i3)
+                  b += conj(l[i3]) * r[i3];
+
+                other(i,j) = (other(i,j) - b)*a;
+              }
+              else
+              {
+                Scalar b = (other(i,j) *= a);
+                Scalar* r = &other(s,j);
+                const Scalar* l = &tri(s,i);
+                for (Index i3=0;i3<rs;++i3)
+                  r[i3] -= b * conj(l[i3]);
+              }
+            }
+          }
+
+          Index lengthTarget = actual_kc-k1-actualPanelWidth;
+          Index startBlock   = IsLower ? k2+k1 : k2-k1-actualPanelWidth;
+          Index blockBOffset = IsLower ? k1 : lengthTarget;
+
+          // update the respective rows of B from other
+          pack_rhs(blockB+actual_kc*j2, other.getSubMapper(startBlock,j2), actualPanelWidth, actual_cols, actual_kc, blockBOffset);
+
+          // GEBP
+          if (lengthTarget>0)
+          {
+            Index startTarget  = IsLower ? k2+k1+actualPanelWidth : k2-actual_kc;
+
+            pack_lhs(blockA, tri.getSubMapper(startTarget,startBlock), actualPanelWidth, lengthTarget);
+
+            gebp_kernel(other.getSubMapper(startTarget,j2), blockA, blockB+actual_kc*j2, lengthTarget, actualPanelWidth, actual_cols, Scalar(-1),
+                        actualPanelWidth, actual_kc, 0, blockBOffset);
+          }
+        }
+      }
+      
+      // R2 -= A21 * B => GEPP
+      {
+        Index start = IsLower ? k2+kc : 0;
+        Index end   = IsLower ? size : k2-kc;
+        for(Index i2=start; i2<end; i2+=mc)
+        {
+          const Index actual_mc = (std::min)(mc,end-i2);
+          if (actual_mc>0)
+          {
+            pack_lhs(blockA, tri.getSubMapper(i2, IsLower ? k2 : k2-kc), actual_kc, actual_mc);
+
+            gebp_kernel(other.getSubMapper(i2, 0), blockA, blockB, actual_mc, actual_kc, cols, Scalar(-1), -1, -1, 0, 0);
+          }
+        }
+      }
+    }
+  }
+
+/* Optimized triangular solver with multiple left hand sides and the triangular matrix on the right
+ */
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
+struct triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>
+{
+  static EIGEN_DONT_INLINE void run(
+    Index size, Index otherSize,
+    const Scalar* _tri, Index triStride,
+    Scalar* _other, Index otherStride,
+    level3_blocking<Scalar,Scalar>& blocking);
+};
+template <typename Scalar, typename Index, int Mode, bool Conjugate, int TriStorageOrder>
+EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor>::run(
+    Index size, Index otherSize,
+    const Scalar* _tri, Index triStride,
+    Scalar* _other, Index otherStride,
+    level3_blocking<Scalar,Scalar>& blocking)
+  {
+    Index rows = otherSize;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    typedef blas_data_mapper<Scalar, Index, ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<Scalar, Index, TriStorageOrder> RhsMapper;
+    LhsMapper lhs(_other, otherStride);
+    RhsMapper rhs(_tri, triStride);
+
+    typedef gebp_traits<Scalar,Scalar> Traits;
+    enum {
+      RhsStorageOrder   = TriStorageOrder,
+      SmallPanelWidth   = EIGEN_PLAIN_ENUM_MAX(Traits::mr,Traits::nr),
+      IsLower = (Mode&Lower) == Lower
+    };
+
+    Index kc = blocking.kc();                   // cache block size along the K direction
+    Index mc = (std::min)(rows,blocking.mc());  // cache block size along the M direction
+
+    std::size_t sizeA = kc*mc;
+    std::size_t sizeB = kc*size;
+
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockA, sizeA, blocking.blockA());
+    ei_declare_aligned_stack_constructed_variable(Scalar, blockB, sizeB, blocking.blockB());
+
+    conj_if<Conjugate> conj;
+    gebp_kernel<Scalar, Scalar, Index, LhsMapper, Traits::mr, Traits::nr, false, Conjugate> gebp_kernel;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder> pack_rhs;
+    gemm_pack_rhs<Scalar, Index, RhsMapper, Traits::nr, RhsStorageOrder,false,true> pack_rhs_panel;
+    gemm_pack_lhs<Scalar, Index, LhsMapper, Traits::mr, Traits::LhsProgress, ColMajor, false, true> pack_lhs_panel;
+
+    for(Index k2=IsLower ? size : 0;
+        IsLower ? k2>0 : k2<size;
+        IsLower ? k2-=kc : k2+=kc)
+    {
+      const Index actual_kc = (std::min)(IsLower ? k2 : size-k2, kc);
+      Index actual_k2 = IsLower ? k2-actual_kc : k2 ;
+
+      Index startPanel = IsLower ? 0 : k2+actual_kc;
+      Index rs = IsLower ? actual_k2 : size - actual_k2 - actual_kc;
+      Scalar* geb = blockB+actual_kc*actual_kc;
+
+      if (rs>0) pack_rhs(geb, rhs.getSubMapper(actual_k2,startPanel), actual_kc, rs);
+
+      // triangular packing (we only pack the panels off the diagonal,
+      // neglecting the blocks overlapping the diagonal
+      {
+        for (Index j2=0; j2<actual_kc; j2+=SmallPanelWidth)
+        {
+          Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+          Index actual_j2 = actual_k2 + j2;
+          Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
+          Index panelLength = IsLower ? actual_kc-j2-actualPanelWidth : j2;
+
+          if (panelLength>0)
+          pack_rhs_panel(blockB+j2*actual_kc,
+                         rhs.getSubMapper(actual_k2+panelOffset, actual_j2),
+                         panelLength, actualPanelWidth,
+                         actual_kc, panelOffset);
+        }
+      }
+
+      for(Index i2=0; i2<rows; i2+=mc)
+      {
+        const Index actual_mc = (std::min)(mc,rows-i2);
+
+        // triangular solver kernel
+        {
+          // for each small block of the diagonal (=> vertical panels of rhs)
+          for (Index j2 = IsLower
+                      ? (actual_kc - ((actual_kc%SmallPanelWidth) ? Index(actual_kc%SmallPanelWidth)
+                                                                  : Index(SmallPanelWidth)))
+                      : 0;
+               IsLower ? j2>=0 : j2<actual_kc;
+               IsLower ? j2-=SmallPanelWidth : j2+=SmallPanelWidth)
+          {
+            Index actualPanelWidth = std::min<Index>(actual_kc-j2, SmallPanelWidth);
+            Index absolute_j2 = actual_k2 + j2;
+            Index panelOffset = IsLower ? j2+actualPanelWidth : 0;
+            Index panelLength = IsLower ? actual_kc - j2 - actualPanelWidth : j2;
+
+            // GEBP
+            if(panelLength>0)
+            {
+              gebp_kernel(lhs.getSubMapper(i2,absolute_j2),
+                          blockA, blockB+j2*actual_kc,
+                          actual_mc, panelLength, actualPanelWidth,
+                          Scalar(-1),
+                          actual_kc, actual_kc, // strides
+                          panelOffset, panelOffset); // offsets
+            }
+
+            // unblocked triangular solve
+            for (Index k=0; k<actualPanelWidth; ++k)
+            {
+              Index j = IsLower ? absolute_j2+actualPanelWidth-k-1 : absolute_j2+k;
+
+              Scalar* r = &lhs(i2,j);
+              for (Index k3=0; k3<k; ++k3)
+              {
+                Scalar b = conj(rhs(IsLower ? j+1+k3 : absolute_j2+k3,j));
+                Scalar* a = &lhs(i2,IsLower ? j+1+k3 : absolute_j2+k3);
+                for (Index i=0; i<actual_mc; ++i)
+                  r[i] -= a[i] * b;
+              }
+              if((Mode & UnitDiag)==0)
+              {
+                Scalar inv_rjj = RealScalar(1)/conj(rhs(j,j));
+                for (Index i=0; i<actual_mc; ++i)
+                  r[i] *= inv_rjj;
+              }
+            }
+
+            // pack the just computed part of lhs to A
+            pack_lhs_panel(blockA, LhsMapper(_other+absolute_j2*otherStride+i2, otherStride),
+                           actualPanelWidth, actual_mc,
+                           actual_kc, j2);
+          }
+        }
+
+        if (rs>0)
+          gebp_kernel(lhs.getSubMapper(i2, startPanel), blockA, geb,
+                      actual_mc, actual_kc, rs, Scalar(-1),
+                      -1, -1, 0, 0);
+      }
+    }
+  }
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
new file mode 100644
index 0000000..88c0fb7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverMatrix_BLAS.h
@@ -0,0 +1,151 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to BLAS F77
+ *   Triangular matrix * matrix product functionality based on ?TRMM.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
+#define EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
+
+namespace Eigen {
+
+namespace internal {
+
+// implements LeftSide op(triangular)^-1 * general
+#define EIGEN_BLAS_TRSM_L(EIGTYPE, BLASTYPE, BLASPREFIX) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheLeft,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+  { \
+   BlasIndex m = convert_index<BlasIndex>(size), n = convert_index<BlasIndex>(otherSize), lda, ldb; \
+   char side = 'L', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _tri; \
+     lda = convert_index<BlasIndex>(triStride); \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+ } \
+};
+
+EIGEN_BLAS_TRSM_L(double,   double, d)
+EIGEN_BLAS_TRSM_L(dcomplex, double, z)
+EIGEN_BLAS_TRSM_L(float,    float,  s)
+EIGEN_BLAS_TRSM_L(scomplex, float,  c)
+
+
+// implements RightSide general * op(triangular)^-1
+#define EIGEN_BLAS_TRSM_R(EIGTYPE, BLASTYPE, BLASPREFIX) \
+template <typename Index, int Mode, bool Conjugate, int TriStorageOrder> \
+struct triangular_solve_matrix<EIGTYPE,Index,OnTheRight,Mode,Conjugate,TriStorageOrder,ColMajor> \
+{ \
+  enum { \
+    IsLower = (Mode&Lower) == Lower, \
+    IsUnitDiag  = (Mode&UnitDiag) ? 1 : 0, \
+    IsZeroDiag  = (Mode&ZeroDiag) ? 1 : 0, \
+    conjA = ((TriStorageOrder==ColMajor) && Conjugate) ? 1 : 0 \
+  }; \
+  static void run( \
+      Index size, Index otherSize, \
+      const EIGTYPE* _tri, Index triStride, \
+      EIGTYPE* _other, Index otherStride, level3_blocking<EIGTYPE,EIGTYPE>& /*blocking*/) \
+  { \
+   BlasIndex m = convert_index<BlasIndex>(otherSize), n = convert_index<BlasIndex>(size), lda, ldb; \
+   char side = 'R', uplo, diag='N', transa; \
+   /* Set alpha_ */ \
+   EIGTYPE alpha(1); \
+   ldb = convert_index<BlasIndex>(otherStride);\
+\
+   const EIGTYPE *a; \
+/* Set trans */ \
+   transa = (TriStorageOrder==RowMajor) ? ((Conjugate) ? 'C' : 'T') : 'N'; \
+/* Set uplo */ \
+   uplo = IsLower ? 'L' : 'U'; \
+   if (TriStorageOrder==RowMajor) uplo = (uplo == 'L') ? 'U' : 'L'; \
+/* Set a, lda */ \
+   typedef Matrix<EIGTYPE, Dynamic, Dynamic, TriStorageOrder> MatrixTri; \
+   Map<const MatrixTri, 0, OuterStride<> > tri(_tri,size,size,OuterStride<>(triStride)); \
+   MatrixTri a_tmp; \
+\
+   if (conjA) { \
+     a_tmp = tri.conjugate(); \
+     a = a_tmp.data(); \
+     lda = convert_index<BlasIndex>(a_tmp.outerStride()); \
+   } else { \
+     a = _tri; \
+     lda = convert_index<BlasIndex>(triStride); \
+   } \
+   if (IsUnitDiag) diag='U'; \
+/* call ?trsm*/ \
+   BLASPREFIX##trsm_(&side, &uplo, &transa, &diag, &m, &n, &numext::real_ref(alpha), (const BLASTYPE*)a, &lda, (BLASTYPE*)_other, &ldb); \
+   /*std::cout << "TRMS_L specialization!\n";*/ \
+ } \
+};
+
+EIGEN_BLAS_TRSM_R(double,   double, d)
+EIGEN_BLAS_TRSM_R(dcomplex, double, z)
+EIGEN_BLAS_TRSM_R(float,    float,  s)
+EIGEN_BLAS_TRSM_R(scomplex, float,  c)
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_MATRIX_BLAS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h
new file mode 100644
index 0000000..b994759
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/products/TriangularSolverVector.h
@@ -0,0 +1,145 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIANGULAR_SOLVER_VECTOR_H
+#define EIGEN_TRIANGULAR_SOLVER_VECTOR_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate, int StorageOrder>
+struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheRight, Mode, Conjugate, StorageOrder>
+{
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    triangular_solve_vector<LhsScalar,RhsScalar,Index,OnTheLeft,
+        ((Mode&Upper)==Upper ? Lower : Upper) | (Mode&UnitDiag),
+        Conjugate,StorageOrder==RowMajor?ColMajor:RowMajor
+      >::run(size, _lhs, lhsStride, rhs);
+  }
+};
+
+// forward and backward substitution, row-major, rhs is a vector
+template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate>
+struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Conjugate, RowMajor>
+{
+  enum {
+    IsLower = ((Mode&Lower)==Lower)
+  };
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,RowMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
+
+    typedef const_blas_data_mapper<LhsScalar,Index,RowMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
+
+    typename internal::conditional<
+                          Conjugate,
+                          const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
+                          const LhsMap&>
+                        ::type cjLhs(lhs);
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+    for(Index pi=IsLower ? 0 : size;
+        IsLower ? pi<size : pi>0;
+        IsLower ? pi+=PanelWidth : pi-=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth);
+
+      Index r = IsLower ? pi : size - pi; // remaining size
+      if (r > 0)
+      {
+        // let's directly call the low level product function because:
+        // 1 - it is faster to compile
+        // 2 - it is slighlty faster at runtime
+        Index startRow = IsLower ? pi : pi-actualPanelWidth;
+        Index startCol = IsLower ? 0 : pi;
+
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,RowMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
+          actualPanelWidth, r,
+          LhsMapper(&lhs.coeffRef(startRow,startCol), lhsStride),
+          RhsMapper(rhs + startCol, 1),
+          rhs + startRow, 1,
+          RhsScalar(-1));
+      }
+
+      for(Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = IsLower ? pi+k : pi-k-1;
+        Index s = IsLower ? pi   : i+1;
+        if (k>0)
+          rhs[i] -= (cjLhs.row(i).segment(s,k).transpose().cwiseProduct(Map<const Matrix<RhsScalar,Dynamic,1> >(rhs+s,k))).sum();
+
+        if(!(Mode & UnitDiag))
+          rhs[i] /= cjLhs(i,i);
+      }
+    }
+  }
+};
+
+// forward and backward substitution, column-major, rhs is a vector
+template<typename LhsScalar, typename RhsScalar, typename Index, int Mode, bool Conjugate>
+struct triangular_solve_vector<LhsScalar, RhsScalar, Index, OnTheLeft, Mode, Conjugate, ColMajor>
+{
+  enum {
+    IsLower = ((Mode&Lower)==Lower)
+  };
+  static void run(Index size, const LhsScalar* _lhs, Index lhsStride, RhsScalar* rhs)
+  {
+    typedef Map<const Matrix<LhsScalar,Dynamic,Dynamic,ColMajor>, 0, OuterStride<> > LhsMap;
+    const LhsMap lhs(_lhs,size,size,OuterStride<>(lhsStride));
+    typedef const_blas_data_mapper<LhsScalar,Index,ColMajor> LhsMapper;
+    typedef const_blas_data_mapper<RhsScalar,Index,ColMajor> RhsMapper;
+    typename internal::conditional<Conjugate,
+                                   const CwiseUnaryOp<typename internal::scalar_conjugate_op<LhsScalar>,LhsMap>,
+                                   const LhsMap&
+                                  >::type cjLhs(lhs);
+    static const Index PanelWidth = EIGEN_TUNE_TRIANGULAR_PANEL_WIDTH;
+
+    for(Index pi=IsLower ? 0 : size;
+        IsLower ? pi<size : pi>0;
+        IsLower ? pi+=PanelWidth : pi-=PanelWidth)
+    {
+      Index actualPanelWidth = (std::min)(IsLower ? size - pi : pi, PanelWidth);
+      Index startBlock = IsLower ? pi : pi-actualPanelWidth;
+      Index endBlock = IsLower ? pi + actualPanelWidth : 0;
+
+      for(Index k=0; k<actualPanelWidth; ++k)
+      {
+        Index i = IsLower ? pi+k : pi-k-1;
+        if(!(Mode & UnitDiag))
+          rhs[i] /= cjLhs.coeff(i,i);
+
+        Index r = actualPanelWidth - k - 1; // remaining size
+        Index s = IsLower ? i+1 : i-r;
+        if (r>0)
+          Map<Matrix<RhsScalar,Dynamic,1> >(rhs+s,r) -= rhs[i] * cjLhs.col(i).segment(s,r);
+      }
+      Index r = IsLower ? size - endBlock : startBlock; // remaining size
+      if (r > 0)
+      {
+        // let's directly call the low level product function because:
+        // 1 - it is faster to compile
+        // 2 - it is slighlty faster at runtime
+        general_matrix_vector_product<Index,LhsScalar,LhsMapper,ColMajor,Conjugate,RhsScalar,RhsMapper,false>::run(
+            r, actualPanelWidth,
+            LhsMapper(&lhs.coeffRef(endBlock,startBlock), lhsStride),
+            RhsMapper(rhs+startBlock, 1),
+            rhs+endBlock, 1, RhsScalar(-1));
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIANGULAR_SOLVER_VECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h
new file mode 100755
index 0000000..6e6ee11
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/BlasUtil.h
@@ -0,0 +1,398 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BLASUTIL_H
+#define EIGEN_BLASUTIL_H
+
+// This file contains many lightweight helper classes used to
+// implement and control fast level 2 and level 3 BLAS-like routines.
+
+namespace Eigen {
+
+namespace internal {
+
+// forward declarations
+template<typename LhsScalar, typename RhsScalar, typename Index, typename DataMapper, int mr, int nr, bool ConjugateLhs=false, bool ConjugateRhs=false>
+struct gebp_kernel;
+
+template<typename Scalar, typename Index, typename DataMapper, int nr, int StorageOrder, bool Conjugate = false, bool PanelMode=false>
+struct gemm_pack_rhs;
+
+template<typename Scalar, typename Index, typename DataMapper, int Pack1, int Pack2, int StorageOrder, bool Conjugate = false, bool PanelMode = false>
+struct gemm_pack_lhs;
+
+template<
+  typename Index,
+  typename LhsScalar, int LhsStorageOrder, bool ConjugateLhs,
+  typename RhsScalar, int RhsStorageOrder, bool ConjugateRhs,
+  int ResStorageOrder>
+struct general_matrix_matrix_product;
+
+template<typename Index,
+         typename LhsScalar, typename LhsMapper, int LhsStorageOrder, bool ConjugateLhs,
+         typename RhsScalar, typename RhsMapper, bool ConjugateRhs, int Version=Specialized>
+struct general_matrix_vector_product;
+
+
+template<bool Conjugate> struct conj_if;
+
+template<> struct conj_if<true> {
+  template<typename T>
+  inline T operator()(const T& x) const { return numext::conj(x); }
+  template<typename T>
+  inline T pconj(const T& x) const { return internal::pconj(x); }
+};
+
+template<> struct conj_if<false> {
+  template<typename T>
+  inline const T& operator()(const T& x) const { return x; }
+  template<typename T>
+  inline const T& pconj(const T& x) const { return x; }
+};
+
+// Generic implementation for custom complex types.
+template<typename LhsScalar, typename RhsScalar, bool ConjLhs, bool ConjRhs>
+struct conj_helper
+{
+  typedef typename ScalarBinaryOpTraits<LhsScalar,RhsScalar>::ReturnType Scalar;
+
+  EIGEN_STRONG_INLINE Scalar pmadd(const LhsScalar& x, const RhsScalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const LhsScalar& x, const RhsScalar& y) const
+  { return conj_if<ConjLhs>()(x) *  conj_if<ConjRhs>()(y); }
+};
+
+template<typename Scalar> struct conj_helper<Scalar,Scalar,false,false>
+{
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const { return internal::pmadd(x,y,c); }
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const { return internal::pmul(x,y); }
+};
+
+template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, false,true>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::imag(x)*numext::real(y) - numext::real(x)*numext::imag(y)); }
+};
+
+template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,false>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) + numext::imag(x)*numext::imag(y), numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); }
+};
+
+template<typename RealScalar> struct conj_helper<std::complex<RealScalar>, std::complex<RealScalar>, true,true>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const Scalar& y, const Scalar& c) const
+  { return c + pmul(x,y); }
+
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const Scalar& y) const
+  { return Scalar(numext::real(x)*numext::real(y) - numext::imag(x)*numext::imag(y), - numext::real(x)*numext::imag(y) - numext::imag(x)*numext::real(y)); }
+};
+
+template<typename RealScalar,bool Conj> struct conj_helper<std::complex<RealScalar>, RealScalar, Conj,false>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const Scalar& x, const RealScalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+  EIGEN_STRONG_INLINE Scalar pmul(const Scalar& x, const RealScalar& y) const
+  { return conj_if<Conj>()(x)*y; }
+};
+
+template<typename RealScalar,bool Conj> struct conj_helper<RealScalar, std::complex<RealScalar>, false,Conj>
+{
+  typedef std::complex<RealScalar> Scalar;
+  EIGEN_STRONG_INLINE Scalar pmadd(const RealScalar& x, const Scalar& y, const Scalar& c) const
+  { return padd(c, pmul(x,y)); }
+  EIGEN_STRONG_INLINE Scalar pmul(const RealScalar& x, const Scalar& y) const
+  { return x*conj_if<Conj>()(y); }
+};
+
+template<typename From,typename To> struct get_factor {
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE To run(const From& x) { return To(x); }
+};
+
+template<typename Scalar> struct get_factor<Scalar,typename NumTraits<Scalar>::Real> {
+  EIGEN_DEVICE_FUNC
+  static EIGEN_STRONG_INLINE typename NumTraits<Scalar>::Real run(const Scalar& x) { return numext::real(x); }
+};
+
+
+template<typename Scalar, typename Index>
+class BlasVectorMapper {
+  public:
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasVectorMapper(Scalar *data) : m_data(data) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar operator()(Index i) const {
+    return m_data[i];
+  }
+  template <typename Packet, int AlignmentType>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet load(Index i) const {
+    return ploadt<Packet, AlignmentType>(m_data + i);
+  }
+
+  template <typename Packet>
+  EIGEN_DEVICE_FUNC bool aligned(Index i) const {
+    return (UIntPtr(m_data+i)%sizeof(Packet))==0;
+  }
+
+  protected:
+  Scalar* m_data;
+};
+
+template<typename Scalar, typename Index, int AlignmentType>
+class BlasLinearMapper {
+  public:
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<Scalar>::half HalfPacket;
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE BlasLinearMapper(Scalar *data) : m_data(data) {}
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void prefetch(int i) const {
+    internal::prefetch(&operator()(i));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Scalar& operator()(Index i) const {
+    return m_data[i];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i) const {
+    return ploadt<Packet, AlignmentType>(m_data + i);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i) const {
+    return ploadt<HalfPacket, AlignmentType>(m_data + i);
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void storePacket(Index i, const Packet &p) const {
+    pstoret<Scalar, Packet, AlignmentType>(m_data + i, p);
+  }
+
+  protected:
+  Scalar *m_data;
+};
+
+// Lightweight helper class to access matrix coefficients.
+template<typename Scalar, typename Index, int StorageOrder, int AlignmentType = Unaligned>
+class blas_data_mapper {
+  public:
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<Scalar>::half HalfPacket;
+
+  typedef BlasLinearMapper<Scalar, Index, AlignmentType> LinearMapper;
+  typedef BlasVectorMapper<Scalar, Index> VectorMapper;
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE blas_data_mapper(Scalar* data, Index stride) : m_data(data), m_stride(stride) {}
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType>
+  getSubMapper(Index i, Index j) const {
+    return blas_data_mapper<Scalar, Index, StorageOrder, AlignmentType>(&operator()(i, j), m_stride);
+  }
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE LinearMapper getLinearMapper(Index i, Index j) const {
+    return LinearMapper(&operator()(i, j));
+  }
+
+  EIGEN_DEVICE_FUNC  EIGEN_ALWAYS_INLINE VectorMapper getVectorMapper(Index i, Index j) const {
+    return VectorMapper(&operator()(i, j));
+  }
+
+
+  EIGEN_DEVICE_FUNC
+  EIGEN_ALWAYS_INLINE Scalar& operator()(Index i, Index j) const {
+    return m_data[StorageOrder==RowMajor ? j + i*m_stride : i + j*m_stride];
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Packet loadPacket(Index i, Index j) const {
+    return ploadt<Packet, AlignmentType>(&operator()(i, j));
+  }
+
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE HalfPacket loadHalfPacket(Index i, Index j) const {
+    return ploadt<HalfPacket, AlignmentType>(&operator()(i, j));
+  }
+
+  template<typename SubPacket>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void scatterPacket(Index i, Index j, const SubPacket &p) const {
+    pscatter<Scalar, SubPacket>(&operator()(i, j), p, m_stride);
+  }
+
+  template<typename SubPacket>
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE SubPacket gatherPacket(Index i, Index j) const {
+    return pgather<Scalar, SubPacket>(&operator()(i, j), m_stride);
+  }
+
+  EIGEN_DEVICE_FUNC const Index stride() const { return m_stride; }
+  EIGEN_DEVICE_FUNC const Scalar* data() const { return m_data; }
+
+  EIGEN_DEVICE_FUNC Index firstAligned(Index size) const {
+    if (UIntPtr(m_data)%sizeof(Scalar)) {
+      return -1;
+    }
+    return internal::first_default_aligned(m_data, size);
+  }
+
+  protected:
+  Scalar* EIGEN_RESTRICT m_data;
+  const Index m_stride;
+};
+
+// lightweight helper class to access matrix coefficients (const version)
+template<typename Scalar, typename Index, int StorageOrder>
+class const_blas_data_mapper : public blas_data_mapper<const Scalar, Index, StorageOrder> {
+  public:
+  EIGEN_ALWAYS_INLINE const_blas_data_mapper(const Scalar *data, Index stride) : blas_data_mapper<const Scalar, Index, StorageOrder>(data, stride) {}
+
+  EIGEN_ALWAYS_INLINE const_blas_data_mapper<Scalar, Index, StorageOrder> getSubMapper(Index i, Index j) const {
+    return const_blas_data_mapper<Scalar, Index, StorageOrder>(&(this->operator()(i, j)), this->m_stride);
+  }
+};
+
+
+/* Helper class to analyze the factors of a Product expression.
+ * In particular it allows to pop out operator-, scalar multiples,
+ * and conjugate */
+template<typename XprType> struct blas_traits
+{
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef const XprType& ExtractType;
+  typedef XprType _ExtractType;
+  enum {
+    IsComplex = NumTraits<Scalar>::IsComplex,
+    IsTransposed = false,
+    NeedToConjugate = false,
+    HasUsableDirectAccess = (    (int(XprType::Flags)&DirectAccessBit)
+                              && (   bool(XprType::IsVectorAtCompileTime)
+                                  || int(inner_stride_at_compile_time<XprType>::ret) == 1)
+                             ) ?  1 : 0
+  };
+  typedef typename conditional<bool(HasUsableDirectAccess),
+    ExtractType,
+    typename _ExtractType::PlainObject
+    >::type DirectLinearAccessType;
+  static inline ExtractType extract(const XprType& x) { return x; }
+  static inline const Scalar extractScalarFactor(const XprType&) { return Scalar(1); }
+};
+
+// pop conjugate
+template<typename Scalar, typename NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> XprType;
+  typedef typename Base::ExtractType ExtractType;
+
+  enum {
+    IsComplex = NumTraits<Scalar>::IsComplex,
+    NeedToConjugate = Base::NeedToConjugate ? 0 : IsComplex
+  };
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline Scalar extractScalarFactor(const XprType& x) { return conj(Base::extractScalarFactor(x.nestedExpression())); }
+};
+
+// pop scalar multiple
+template<typename Scalar, typename NestedXpr, typename Plain>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain>, NestedXpr> XprType;
+  typedef typename Base::ExtractType ExtractType;
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.rhs()); }
+  static inline Scalar extractScalarFactor(const XprType& x)
+  { return x.lhs().functor().m_other * Base::extractScalarFactor(x.rhs()); }
+};
+template<typename Scalar, typename NestedXpr, typename Plain>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseBinaryOp<scalar_product_op<Scalar>, NestedXpr, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain> > XprType;
+  typedef typename Base::ExtractType ExtractType;
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.lhs()); }
+  static inline Scalar extractScalarFactor(const XprType& x)
+  { return Base::extractScalarFactor(x.lhs()) * x.rhs().functor().m_other; }
+};
+template<typename Scalar, typename Plain1, typename Plain2>
+struct blas_traits<CwiseBinaryOp<scalar_product_op<Scalar>, const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1>,
+                                                            const CwiseNullaryOp<scalar_constant_op<Scalar>,Plain2> > >
+ : blas_traits<CwiseNullaryOp<scalar_constant_op<Scalar>,Plain1> >
+{};
+
+// pop opposite
+template<typename Scalar, typename NestedXpr>
+struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef blas_traits<NestedXpr> Base;
+  typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType;
+  typedef typename Base::ExtractType ExtractType;
+  static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
+  static inline Scalar extractScalarFactor(const XprType& x)
+  { return - Base::extractScalarFactor(x.nestedExpression()); }
+};
+
+// pop/push transpose
+template<typename NestedXpr>
+struct blas_traits<Transpose<NestedXpr> >
+ : blas_traits<NestedXpr>
+{
+  typedef typename NestedXpr::Scalar Scalar;
+  typedef blas_traits<NestedXpr> Base;
+  typedef Transpose<NestedXpr> XprType;
+  typedef Transpose<const typename Base::_ExtractType>  ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS
+  typedef Transpose<const typename Base::_ExtractType> _ExtractType;
+  typedef typename conditional<bool(Base::HasUsableDirectAccess),
+    ExtractType,
+    typename ExtractType::PlainObject
+    >::type DirectLinearAccessType;
+  enum {
+    IsTransposed = Base::IsTransposed ? 0 : 1
+  };
+  static inline ExtractType extract(const XprType& x) { return ExtractType(Base::extract(x.nestedExpression())); }
+  static inline Scalar extractScalarFactor(const XprType& x) { return Base::extractScalarFactor(x.nestedExpression()); }
+};
+
+template<typename T>
+struct blas_traits<const T>
+     : blas_traits<T>
+{};
+
+template<typename T, bool HasUsableDirectAccess=blas_traits<T>::HasUsableDirectAccess>
+struct extract_data_selector {
+  static const typename T::Scalar* run(const T& m)
+  {
+    return blas_traits<T>::extract(m).data();
+  }
+};
+
+template<typename T>
+struct extract_data_selector<T,false> {
+  static typename T::Scalar* run(const T&) { return 0; }
+};
+
+template<typename T> const typename T::Scalar* extract_data(const T& m)
+{
+  return extract_data_selector<T>::run(m);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BLASUTIL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h
new file mode 100644
index 0000000..7587d68
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Constants.h
@@ -0,0 +1,547 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2007-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONSTANTS_H
+#define EIGEN_CONSTANTS_H
+
+namespace Eigen {
+
+/** This value means that a positive quantity (e.g., a size) is not known at compile-time, and that instead the value is
+  * stored in some runtime variable.
+  *
+  * Changing the value of Dynamic breaks the ABI, as Dynamic is often used as a template parameter for Matrix.
+  */
+const int Dynamic = -1;
+
+/** This value means that a signed quantity (e.g., a signed index) is not known at compile-time, and that instead its value
+  * has to be specified at runtime.
+  */
+const int DynamicIndex = 0xffffff;
+
+/** This value means +Infinity; it is currently used only as the p parameter to MatrixBase::lpNorm<int>().
+  * The value Infinity there means the L-infinity norm.
+  */
+const int Infinity = -1;
+
+/** This value means that the cost to evaluate an expression coefficient is either very expensive or
+  * cannot be known at compile time.
+  *
+  * This value has to be positive to (1) simplify cost computation, and (2) allow to distinguish between a very expensive and very very expensive expressions.
+  * It thus must also be large enough to make sure unrolling won't happen and that sub expressions will be evaluated, but not too large to avoid overflow.
+  */
+const int HugeCost = 10000;
+
+/** \defgroup flags Flags
+  * \ingroup Core_Module
+  *
+  * These are the possible bits which can be OR'ed to constitute the flags of a matrix or
+  * expression.
+  *
+  * It is important to note that these flags are a purely compile-time notion. They are a compile-time property of
+  * an expression type, implemented as enum's. They are not stored in memory at runtime, and they do not incur any
+  * runtime overhead.
+  *
+  * \sa MatrixBase::Flags
+  */
+
+/** \ingroup flags
+  *
+  * for a matrix, this means that the storage order is row-major.
+  * If this bit is not set, the storage order is column-major.
+  * For an expression, this determines the storage order of
+  * the matrix created by evaluation of that expression.
+  * \sa \blank  \ref TopicStorageOrders */
+const unsigned int RowMajorBit = 0x1;
+
+/** \ingroup flags
+  * means the expression should be evaluated by the calling expression */
+const unsigned int EvalBeforeNestingBit = 0x2;
+
+/** \ingroup flags
+  * \deprecated
+  * means the expression should be evaluated before any assignment */
+EIGEN_DEPRECATED
+const unsigned int EvalBeforeAssigningBit = 0x4; // FIXME deprecated
+
+/** \ingroup flags
+  *
+  * Short version: means the expression might be vectorized
+  *
+  * Long version: means that the coefficients can be handled by packets
+  * and start at a memory location whose alignment meets the requirements
+  * of the present CPU architecture for optimized packet access. In the fixed-size
+  * case, there is the additional condition that it be possible to access all the
+  * coefficients by packets (this implies the requirement that the size be a multiple of 16 bytes,
+  * and that any nontrivial strides don't break the alignment). In the dynamic-size case,
+  * there is no such condition on the total size and strides, so it might not be possible to access
+  * all coeffs by packets.
+  *
+  * \note This bit can be set regardless of whether vectorization is actually enabled.
+  *       To check for actual vectorizability, see \a ActualPacketAccessBit.
+  */
+const unsigned int PacketAccessBit = 0x8;
+
+#ifdef EIGEN_VECTORIZE
+/** \ingroup flags
+  *
+  * If vectorization is enabled (EIGEN_VECTORIZE is defined) this constant
+  * is set to the value \a PacketAccessBit.
+  *
+  * If vectorization is not enabled (EIGEN_VECTORIZE is not defined) this constant
+  * is set to the value 0.
+  */
+const unsigned int ActualPacketAccessBit = PacketAccessBit;
+#else
+const unsigned int ActualPacketAccessBit = 0x0;
+#endif
+
+/** \ingroup flags
+  *
+  * Short version: means the expression can be seen as 1D vector.
+  *
+  * Long version: means that one can access the coefficients
+  * of this expression by coeff(int), and coeffRef(int) in the case of a lvalue expression. These
+  * index-based access methods are guaranteed
+  * to not have to do any runtime computation of a (row, col)-pair from the index, so that it
+  * is guaranteed that whenever it is available, index-based access is at least as fast as
+  * (row,col)-based access. Expressions for which that isn't possible don't have the LinearAccessBit.
+  *
+  * If both PacketAccessBit and LinearAccessBit are set, then the
+  * packets of this expression can be accessed by packet(int), and writePacket(int) in the case of a
+  * lvalue expression.
+  *
+  * Typically, all vector expressions have the LinearAccessBit, but there is one exception:
+  * Product expressions don't have it, because it would be troublesome for vectorization, even when the
+  * Product is a vector expression. Thus, vector Product expressions allow index-based coefficient access but
+  * not index-based packet access, so they don't have the LinearAccessBit.
+  */
+const unsigned int LinearAccessBit = 0x10;
+
+/** \ingroup flags
+  *
+  * Means the expression has a coeffRef() method, i.e. is writable as its individual coefficients are directly addressable.
+  * This rules out read-only expressions.
+  *
+  * Note that DirectAccessBit and LvalueBit are mutually orthogonal, as there are examples of expression having one but note
+  * the other:
+  *   \li writable expressions that don't have a very simple memory layout as a strided array, have LvalueBit but not DirectAccessBit
+  *   \li Map-to-const expressions, for example Map<const Matrix>, have DirectAccessBit but not LvalueBit
+  *
+  * Expressions having LvalueBit also have their coeff() method returning a const reference instead of returning a new value.
+  */
+const unsigned int LvalueBit = 0x20;
+
+/** \ingroup flags
+  *
+  * Means that the underlying array of coefficients can be directly accessed as a plain strided array. The memory layout
+  * of the array of coefficients must be exactly the natural one suggested by rows(), cols(),
+  * outerStride(), innerStride(), and the RowMajorBit. This rules out expressions such as Diagonal, whose coefficients,
+  * though referencable, do not have such a regular memory layout.
+  *
+  * See the comment on LvalueBit for an explanation of how LvalueBit and DirectAccessBit are mutually orthogonal.
+  */
+const unsigned int DirectAccessBit = 0x40;
+
+/** \deprecated \ingroup flags
+  *
+  * means the first coefficient packet is guaranteed to be aligned.
+  * An expression cannot has the AlignedBit without the PacketAccessBit flag.
+  * In other words, this means we are allow to perform an aligned packet access to the first element regardless
+  * of the expression kind:
+  * \code
+  * expression.packet<Aligned>(0);
+  * \endcode
+  */
+EIGEN_DEPRECATED const unsigned int AlignedBit = 0x80;
+
+const unsigned int NestByRefBit = 0x100;
+
+/** \ingroup flags
+  *
+  * for an expression, this means that the storage order
+  * can be either row-major or column-major.
+  * The precise choice will be decided at evaluation time or when
+  * combined with other expressions.
+  * \sa \blank  \ref RowMajorBit, \ref TopicStorageOrders */
+const unsigned int NoPreferredStorageOrderBit = 0x200;
+
+/** \ingroup flags
+  *
+  * Means that the underlying coefficients can be accessed through pointers to the sparse (un)compressed storage format,
+  * that is, the expression provides:
+  * \code
+    inline const Scalar* valuePtr() const;
+    inline const Index* innerIndexPtr() const;
+    inline const Index* outerIndexPtr() const;
+    inline const Index* innerNonZeroPtr() const;
+    \endcode
+  */
+const unsigned int CompressedAccessBit = 0x400;
+
+
+// list of flags that are inherited by default
+const unsigned int HereditaryBits = RowMajorBit
+                                  | EvalBeforeNestingBit;
+
+/** \defgroup enums Enumerations
+  * \ingroup Core_Module
+  *
+  * Various enumerations used in %Eigen. Many of these are used as template parameters.
+  */
+
+/** \ingroup enums
+  * Enum containing possible values for the \c Mode or \c UpLo parameter of
+  * MatrixBase::selfadjointView() and MatrixBase::triangularView(), and selfadjoint solvers. */
+enum UpLoType {
+  /** View matrix as a lower triangular matrix. */
+  Lower=0x1,                      
+  /** View matrix as an upper triangular matrix. */
+  Upper=0x2,                      
+  /** %Matrix has ones on the diagonal; to be used in combination with #Lower or #Upper. */
+  UnitDiag=0x4, 
+  /** %Matrix has zeros on the diagonal; to be used in combination with #Lower or #Upper. */
+  ZeroDiag=0x8,
+  /** View matrix as a lower triangular matrix with ones on the diagonal. */
+  UnitLower=UnitDiag|Lower, 
+  /** View matrix as an upper triangular matrix with ones on the diagonal. */
+  UnitUpper=UnitDiag|Upper,
+  /** View matrix as a lower triangular matrix with zeros on the diagonal. */
+  StrictlyLower=ZeroDiag|Lower, 
+  /** View matrix as an upper triangular matrix with zeros on the diagonal. */
+  StrictlyUpper=ZeroDiag|Upper,
+  /** Used in BandMatrix and SelfAdjointView to indicate that the matrix is self-adjoint. */
+  SelfAdjoint=0x10,
+  /** Used to support symmetric, non-selfadjoint, complex matrices. */
+  Symmetric=0x20
+};
+
+/** \ingroup enums
+  * Enum for indicating whether a buffer is aligned or not. */
+enum AlignmentType {
+  Unaligned=0,        /**< Data pointer has no specific alignment. */
+  Aligned8=8,         /**< Data pointer is aligned on a 8 bytes boundary. */
+  Aligned16=16,       /**< Data pointer is aligned on a 16 bytes boundary. */
+  Aligned32=32,       /**< Data pointer is aligned on a 32 bytes boundary. */
+  Aligned64=64,       /**< Data pointer is aligned on a 64 bytes boundary. */
+  Aligned128=128,     /**< Data pointer is aligned on a 128 bytes boundary. */
+  AlignedMask=255,
+  Aligned=16,         /**< \deprecated Synonym for Aligned16. */
+#if EIGEN_MAX_ALIGN_BYTES==128
+  AlignedMax = Aligned128
+#elif EIGEN_MAX_ALIGN_BYTES==64
+  AlignedMax = Aligned64
+#elif EIGEN_MAX_ALIGN_BYTES==32
+  AlignedMax = Aligned32
+#elif EIGEN_MAX_ALIGN_BYTES==16
+  AlignedMax = Aligned16
+#elif EIGEN_MAX_ALIGN_BYTES==8
+  AlignedMax = Aligned8
+#elif EIGEN_MAX_ALIGN_BYTES==0
+  AlignedMax = Unaligned
+#else
+#error Invalid value for EIGEN_MAX_ALIGN_BYTES
+#endif
+};
+
+/** \ingroup enums
+ * Enum used by DenseBase::corner() in Eigen2 compatibility mode. */
+// FIXME after the corner() API change, this was not needed anymore, except by AlignedBox
+// TODO: find out what to do with that. Adapt the AlignedBox API ?
+enum CornerType { TopLeft, TopRight, BottomLeft, BottomRight };
+
+/** \ingroup enums
+  * Enum containing possible values for the \p Direction parameter of
+  * Reverse, PartialReduxExpr and VectorwiseOp. */
+enum DirectionType { 
+  /** For Reverse, all columns are reversed; 
+    * for PartialReduxExpr and VectorwiseOp, act on columns. */
+  Vertical, 
+  /** For Reverse, all rows are reversed; 
+    * for PartialReduxExpr and VectorwiseOp, act on rows. */
+  Horizontal, 
+  /** For Reverse, both rows and columns are reversed; 
+    * not used for PartialReduxExpr and VectorwiseOp. */
+  BothDirections 
+};
+
+/** \internal \ingroup enums
+  * Enum to specify how to traverse the entries of a matrix. */
+enum TraversalType {
+  /** \internal Default traversal, no vectorization, no index-based access */
+  DefaultTraversal,
+  /** \internal No vectorization, use index-based access to have only one for loop instead of 2 nested loops */
+  LinearTraversal,
+  /** \internal Equivalent to a slice vectorization for fixed-size matrices having good alignment
+    * and good size */
+  InnerVectorizedTraversal,
+  /** \internal Vectorization path using a single loop plus scalar loops for the
+    * unaligned boundaries */
+  LinearVectorizedTraversal,
+  /** \internal Generic vectorization path using one vectorized loop per row/column with some
+    * scalar loops to handle the unaligned boundaries */
+  SliceVectorizedTraversal,
+  /** \internal Special case to properly handle incompatible scalar types or other defecting cases*/
+  InvalidTraversal,
+  /** \internal Evaluate all entries at once */
+  AllAtOnceTraversal
+};
+
+/** \internal \ingroup enums
+  * Enum to specify whether to unroll loops when traversing over the entries of a matrix. */
+enum UnrollingType {
+  /** \internal Do not unroll loops. */
+  NoUnrolling,
+  /** \internal Unroll only the inner loop, but not the outer loop. */
+  InnerUnrolling,
+  /** \internal Unroll both the inner and the outer loop. If there is only one loop, 
+    * because linear traversal is used, then unroll that loop. */
+  CompleteUnrolling
+};
+
+/** \internal \ingroup enums
+  * Enum to specify whether to use the default (built-in) implementation or the specialization. */
+enum SpecializedType {
+  Specialized,
+  BuiltIn
+};
+
+/** \ingroup enums
+  * Enum containing possible values for the \p _Options template parameter of
+  * Matrix, Array and BandMatrix. */
+enum StorageOptions {
+  /** Storage order is column major (see \ref TopicStorageOrders). */
+  ColMajor = 0,
+  /** Storage order is row major (see \ref TopicStorageOrders). */
+  RowMajor = 0x1,  // it is only a coincidence that this is equal to RowMajorBit -- don't rely on that
+  /** Align the matrix itself if it is vectorizable fixed-size */
+  AutoAlign = 0,
+  /** Don't require alignment for the matrix itself (the array of coefficients, if dynamically allocated, may still be requested to be aligned) */ // FIXME --- clarify the situation
+  DontAlign = 0x2
+};
+
+/** \ingroup enums
+  * Enum for specifying whether to apply or solve on the left or right. */
+enum SideType {
+  /** Apply transformation on the left. */
+  OnTheLeft = 1,  
+  /** Apply transformation on the right. */
+  OnTheRight = 2  
+};
+
+/* the following used to be written as:
+ *
+ *   struct NoChange_t {};
+ *   namespace {
+ *     EIGEN_UNUSED NoChange_t NoChange;
+ *   }
+ *
+ * on the ground that it feels dangerous to disambiguate overloaded functions on enum/integer types.  
+ * However, this leads to "variable declared but never referenced" warnings on Intel Composer XE,
+ * and we do not know how to get rid of them (bug 450).
+ */
+
+enum NoChange_t   { NoChange };
+enum Sequential_t { Sequential };
+enum Default_t    { Default };
+
+/** \internal \ingroup enums
+  * Used in AmbiVector. */
+enum AmbiVectorMode {
+  IsDense         = 0,
+  IsSparse
+};
+
+/** \ingroup enums
+  * Used as template parameter in DenseCoeffBase and MapBase to indicate 
+  * which accessors should be provided. */
+enum AccessorLevels {
+  /** Read-only access via a member function. */
+  ReadOnlyAccessors, 
+  /** Read/write access via member functions. */
+  WriteAccessors, 
+  /** Direct read-only access to the coefficients. */
+  DirectAccessors, 
+  /** Direct read/write access to the coefficients. */
+  DirectWriteAccessors
+};
+
+/** \ingroup enums
+  * Enum with options to give to various decompositions. */
+enum DecompositionOptions {
+  /** \internal Not used (meant for LDLT?). */
+  Pivoting            = 0x01, 
+  /** \internal Not used (meant for LDLT?). */
+  NoPivoting          = 0x02, 
+  /** Used in JacobiSVD to indicate that the square matrix U is to be computed. */
+  ComputeFullU        = 0x04,
+  /** Used in JacobiSVD to indicate that the thin matrix U is to be computed. */
+  ComputeThinU        = 0x08,
+  /** Used in JacobiSVD to indicate that the square matrix V is to be computed. */
+  ComputeFullV        = 0x10,
+  /** Used in JacobiSVD to indicate that the thin matrix V is to be computed. */
+  ComputeThinV        = 0x20,
+  /** Used in SelfAdjointEigenSolver and GeneralizedSelfAdjointEigenSolver to specify
+    * that only the eigenvalues are to be computed and not the eigenvectors. */
+  EigenvaluesOnly     = 0x40,
+  /** Used in SelfAdjointEigenSolver and GeneralizedSelfAdjointEigenSolver to specify
+    * that both the eigenvalues and the eigenvectors are to be computed. */
+  ComputeEigenvectors = 0x80,
+  /** \internal */
+  EigVecMask = EigenvaluesOnly | ComputeEigenvectors,
+  /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should
+    * solve the generalized eigenproblem \f$ Ax = \lambda B x \f$. */
+  Ax_lBx              = 0x100,
+  /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should
+    * solve the generalized eigenproblem \f$ ABx = \lambda x \f$. */
+  ABx_lx              = 0x200,
+  /** Used in GeneralizedSelfAdjointEigenSolver to indicate that it should
+    * solve the generalized eigenproblem \f$ BAx = \lambda x \f$. */
+  BAx_lx              = 0x400,
+  /** \internal */
+  GenEigMask = Ax_lBx | ABx_lx | BAx_lx
+};
+
+/** \ingroup enums
+  * Possible values for the \p QRPreconditioner template parameter of JacobiSVD. */
+enum QRPreconditioners {
+  /** Do not specify what is to be done if the SVD of a non-square matrix is asked for. */
+  NoQRPreconditioner,
+  /** Use a QR decomposition without pivoting as the first step. */
+  HouseholderQRPreconditioner,
+  /** Use a QR decomposition with column pivoting as the first step. */
+  ColPivHouseholderQRPreconditioner,
+  /** Use a QR decomposition with full pivoting as the first step. */
+  FullPivHouseholderQRPreconditioner
+};
+
+#ifdef Success
+#error The preprocessor symbol 'Success' is defined, possibly by the X11 header file X.h
+#endif
+
+/** \ingroup enums
+  * Enum for reporting the status of a computation. */
+enum ComputationInfo {
+  /** Computation was successful. */
+  Success = 0,        
+  /** The provided data did not satisfy the prerequisites. */
+  NumericalIssue = 1, 
+  /** Iterative procedure did not converge. */
+  NoConvergence = 2,
+  /** The inputs are invalid, or the algorithm has been improperly called.
+    * When assertions are enabled, such errors trigger an assert. */
+  InvalidInput = 3
+};
+
+/** \ingroup enums
+  * Enum used to specify how a particular transformation is stored in a matrix.
+  * \sa Transform, Hyperplane::transform(). */
+enum TransformTraits {
+  /** Transformation is an isometry. */
+  Isometry      = 0x1,
+  /** Transformation is an affine transformation stored as a (Dim+1)^2 matrix whose last row is 
+    * assumed to be [0 ... 0 1]. */
+  Affine        = 0x2,
+  /** Transformation is an affine transformation stored as a (Dim) x (Dim+1) matrix. */
+  AffineCompact = 0x10 | Affine,
+  /** Transformation is a general projective transformation stored as a (Dim+1)^2 matrix. */
+  Projective    = 0x20
+};
+
+/** \internal \ingroup enums
+  * Enum used to choose between implementation depending on the computer architecture. */
+namespace Architecture
+{
+  enum Type {
+    Generic = 0x0,
+    SSE = 0x1,
+    AltiVec = 0x2,
+    VSX = 0x3,
+    NEON = 0x4,
+#if defined EIGEN_VECTORIZE_SSE
+    Target = SSE
+#elif defined EIGEN_VECTORIZE_ALTIVEC
+    Target = AltiVec
+#elif defined EIGEN_VECTORIZE_VSX
+    Target = VSX
+#elif defined EIGEN_VECTORIZE_NEON
+    Target = NEON
+#else
+    Target = Generic
+#endif
+  };
+}
+
+/** \internal \ingroup enums
+  * Enum used as template parameter in Product and product evaluators. */
+enum ProductImplType
+{ DefaultProduct=0, LazyProduct, AliasFreeProduct, CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct };
+
+/** \internal \ingroup enums
+  * Enum used in experimental parallel implementation. */
+enum Action {GetAction, SetAction};
+
+/** The type used to identify a dense storage. */
+struct Dense {};
+
+/** The type used to identify a general sparse storage. */
+struct Sparse {};
+
+/** The type used to identify a general solver (factored) storage. */
+struct SolverStorage {};
+
+/** The type used to identify a permutation storage. */
+struct PermutationStorage {};
+
+/** The type used to identify a permutation storage. */
+struct TranspositionsStorage {};
+
+/** The type used to identify a matrix expression */
+struct MatrixXpr {};
+
+/** The type used to identify an array expression */
+struct ArrayXpr {};
+
+// An evaluator must define its shape. By default, it can be one of the following:
+struct DenseShape             { static std::string debugName() { return "DenseShape"; } };
+struct SolverShape            { static std::string debugName() { return "SolverShape"; } };
+struct HomogeneousShape       { static std::string debugName() { return "HomogeneousShape"; } };
+struct DiagonalShape          { static std::string debugName() { return "DiagonalShape"; } };
+struct BandShape              { static std::string debugName() { return "BandShape"; } };
+struct TriangularShape        { static std::string debugName() { return "TriangularShape"; } };
+struct SelfAdjointShape       { static std::string debugName() { return "SelfAdjointShape"; } };
+struct PermutationShape       { static std::string debugName() { return "PermutationShape"; } };
+struct TranspositionsShape    { static std::string debugName() { return "TranspositionsShape"; } };
+struct SparseShape            { static std::string debugName() { return "SparseShape"; } };
+
+namespace internal {
+
+  // random access iterators based on coeff*() accessors.
+struct IndexBased {};
+
+// evaluator based on iterators to access coefficients. 
+struct IteratorBased {};
+
+/** \internal
+ * Constants for comparison functors
+ */
+enum ComparisonName {
+  cmp_EQ = 0,
+  cmp_LT = 1,
+  cmp_LE = 2,
+  cmp_UNORD = 3,
+  cmp_NEQ = 4,
+  cmp_GT = 5,
+  cmp_GE = 6
+};
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CONSTANTS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h
new file mode 100755
index 0000000..7559e12
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/DisableStupidWarnings.h
@@ -0,0 +1,75 @@
+#ifndef EIGEN_WARNINGS_DISABLED
+#define EIGEN_WARNINGS_DISABLED
+
+#ifdef _MSC_VER
+  // 4100 - unreferenced formal parameter (occurred e.g. in aligned_allocator::destroy(pointer p))
+  // 4101 - unreferenced local variable
+  // 4127 - conditional expression is constant
+  // 4181 - qualifier applied to reference type ignored
+  // 4211 - nonstandard extension used : redefined extern to static
+  // 4244 - 'argument' : conversion from 'type1' to 'type2', possible loss of data
+  // 4273 - QtAlignedMalloc, inconsistent DLL linkage
+  // 4324 - structure was padded due to declspec(align())
+  // 4503 - decorated name length exceeded, name was truncated
+  // 4512 - assignment operator could not be generated
+  // 4522 - 'class' : multiple assignment operators specified
+  // 4700 - uninitialized local variable 'xyz' used
+  // 4714 - function marked as __forceinline not inlined
+  // 4717 - 'function' : recursive on all control paths, function will cause runtime stack overflow
+  // 4800 - 'type' : forcing value to bool 'true' or 'false' (performance warning)
+  #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+    #pragma warning( push )
+  #endif
+  #pragma warning( disable : 4100 4101 4127 4181 4211 4244 4273 4324 4503 4512 4522 4700 4714 4717 4800)
+
+#elif defined __INTEL_COMPILER
+  // 2196 - routine is both "inline" and "noinline" ("noinline" assumed)
+  //        ICC 12 generates this warning even without any inline keyword, when defining class methods 'inline' i.e. inside of class body
+  //        typedef that may be a reference type.
+  // 279  - controlling expression is constant
+  //        ICC 12 generates this warning on assert(constant_expression_depending_on_template_params) and frankly this is a legitimate use case.
+  // 1684 - conversion from pointer to same-sized integral type (potential portability problem)
+  // 2259 - non-pointer conversion from "Eigen::Index={ptrdiff_t={long}}" to "int" may lose significant bits
+  #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+    #pragma warning push
+  #endif
+  #pragma warning disable 2196 279 1684 2259
+
+#elif defined __clang__
+  // -Wconstant-logical-operand - warning: use of logical && with constant operand; switch to bitwise & or remove constant
+  //     this is really a stupid warning as it warns on compile-time expressions involving enums
+  #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+    #pragma clang diagnostic push
+  #endif
+  #pragma clang diagnostic ignored "-Wconstant-logical-operand"
+
+#elif defined __GNUC__ && __GNUC__>=6
+
+  #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+    #pragma GCC diagnostic push
+  #endif
+  #pragma GCC diagnostic ignored "-Wignored-attributes"
+
+#endif
+
+#if defined __NVCC__
+  // Disable the "statement is unreachable" message
+  #pragma diag_suppress code_is_unreachable
+  // Disable the "dynamic initialization in unreachable code" message
+  #pragma diag_suppress initialization_not_reachable
+  // Disable the "invalid error number" message that we get with older versions of nvcc
+  #pragma diag_suppress 1222
+  // Disable the "calling a __host__ function from a __host__ __device__ function is not allowed" messages (yes, there are many of them and they seem to change with every version of the compiler)
+  #pragma diag_suppress 2527
+  #pragma diag_suppress 2529
+  #pragma diag_suppress 2651
+  #pragma diag_suppress 2653
+  #pragma diag_suppress 2668
+  #pragma diag_suppress 2669
+  #pragma diag_suppress 2670
+  #pragma diag_suppress 2671
+  #pragma diag_suppress 2735
+  #pragma diag_suppress 2737
+#endif
+
+#endif // not EIGEN_WARNINGS_DISABLED
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h
new file mode 100644
index 0000000..ea10739
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ForwardDeclarations.h
@@ -0,0 +1,302 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2007-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FORWARDDECLARATIONS_H
+#define EIGEN_FORWARDDECLARATIONS_H
+
+namespace Eigen {
+namespace internal {
+
+template<typename T> struct traits;
+
+// here we say once and for all that traits<const T> == traits<T>
+// When constness must affect traits, it has to be constness on template parameters on which T itself depends.
+// For example, traits<Map<const T> > != traits<Map<T> >, but
+//              traits<const Map<T> > == traits<Map<T> >
+template<typename T> struct traits<const T> : traits<T> {};
+
+template<typename Derived> struct has_direct_access
+{
+  enum { ret = (traits<Derived>::Flags & DirectAccessBit) ? 1 : 0 };
+};
+
+template<typename Derived> struct accessors_level
+{
+  enum { has_direct_access = (traits<Derived>::Flags & DirectAccessBit) ? 1 : 0,
+         has_write_access = (traits<Derived>::Flags & LvalueBit) ? 1 : 0,
+         value = has_direct_access ? (has_write_access ? DirectWriteAccessors : DirectAccessors)
+                                   : (has_write_access ? WriteAccessors       : ReadOnlyAccessors)
+  };
+};
+
+template<typename T> struct evaluator_traits;
+
+template< typename T> struct evaluator;
+
+} // end namespace internal
+
+template<typename T> struct NumTraits;
+
+template<typename Derived> struct EigenBase;
+template<typename Derived> class DenseBase;
+template<typename Derived> class PlainObjectBase;
+
+
+template<typename Derived,
+         int Level = internal::accessors_level<Derived>::value >
+class DenseCoeffsBase;
+
+template<typename _Scalar, int _Rows, int _Cols,
+         int _Options = AutoAlign |
+#if EIGEN_GNUC_AT(3,4)
+    // workaround a bug in at least gcc 3.4.6
+    // the innermost ?: ternary operator is misparsed. We write it slightly
+    // differently and this makes gcc 3.4.6 happy, but it's ugly.
+    // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined
+    // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor)
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : !(_Cols==1 && _Rows!=1) ?  EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION
+                          : Eigen::ColMajor ),
+#else
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor
+                          : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ),
+#endif
+         int _MaxRows = _Rows,
+         int _MaxCols = _Cols
+> class Matrix;
+
+template<typename Derived> class MatrixBase;
+template<typename Derived> class ArrayBase;
+
+template<typename ExpressionType, unsigned int Added, unsigned int Removed> class Flagged;
+template<typename ExpressionType, template <typename> class StorageBase > class NoAlias;
+template<typename ExpressionType> class NestByValue;
+template<typename ExpressionType> class ForceAlignedAccess;
+template<typename ExpressionType> class SwapWrapper;
+
+template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false> class Block;
+
+template<typename MatrixType, int Size=Dynamic> class VectorBlock;
+template<typename MatrixType> class Transpose;
+template<typename MatrixType> class Conjugate;
+template<typename NullaryOp, typename MatrixType>         class CwiseNullaryOp;
+template<typename UnaryOp,   typename MatrixType>         class CwiseUnaryOp;
+template<typename ViewOp,    typename MatrixType>         class CwiseUnaryView;
+template<typename BinaryOp,  typename Lhs, typename Rhs>  class CwiseBinaryOp;
+template<typename TernaryOp, typename Arg1, typename Arg2, typename Arg3>  class CwiseTernaryOp;
+template<typename Decomposition, typename Rhstype>        class Solve;
+template<typename XprType>                                class Inverse;
+
+template<typename Lhs, typename Rhs, int Option = DefaultProduct> class Product;
+
+template<typename Derived> class DiagonalBase;
+template<typename _DiagonalVectorType> class DiagonalWrapper;
+template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime=SizeAtCompileTime> class DiagonalMatrix;
+template<typename MatrixType, typename DiagonalType, int ProductOrder> class DiagonalProduct;
+template<typename MatrixType, int Index = 0> class Diagonal;
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime = SizeAtCompileTime, typename IndexType=int> class PermutationMatrix;
+template<int SizeAtCompileTime, int MaxSizeAtCompileTime = SizeAtCompileTime, typename IndexType=int> class Transpositions;
+template<typename Derived> class PermutationBase;
+template<typename Derived> class TranspositionsBase;
+template<typename _IndicesType> class PermutationWrapper;
+template<typename _IndicesType> class TranspositionsWrapper;
+
+template<typename Derived,
+         int Level = internal::accessors_level<Derived>::has_write_access ? WriteAccessors : ReadOnlyAccessors
+> class MapBase;
+template<int InnerStrideAtCompileTime, int OuterStrideAtCompileTime> class Stride;
+template<int Value = Dynamic> class InnerStride;
+template<int Value = Dynamic> class OuterStride;
+template<typename MatrixType, int MapOptions=Unaligned, typename StrideType = Stride<0,0> > class Map;
+template<typename Derived> class RefBase;
+template<typename PlainObjectType, int Options = 0,
+         typename StrideType = typename internal::conditional<PlainObjectType::IsVectorAtCompileTime,InnerStride<1>,OuterStride<> >::type > class Ref;
+
+template<typename Derived> class TriangularBase;
+template<typename MatrixType, unsigned int Mode> class TriangularView;
+template<typename MatrixType, unsigned int Mode> class SelfAdjointView;
+template<typename MatrixType> class SparseView;
+template<typename ExpressionType> class WithFormat;
+template<typename MatrixType> struct CommaInitializer;
+template<typename Derived> class ReturnByValue;
+template<typename ExpressionType> class ArrayWrapper;
+template<typename ExpressionType> class MatrixWrapper;
+template<typename Derived> class SolverBase;
+template<typename XprType> class InnerIterator;
+
+namespace internal {
+template<typename DecompositionType> struct kernel_retval_base;
+template<typename DecompositionType> struct kernel_retval;
+template<typename DecompositionType> struct image_retval_base;
+template<typename DecompositionType> struct image_retval;
+} // end namespace internal
+
+namespace internal {
+template<typename _Scalar, int Rows=Dynamic, int Cols=Dynamic, int Supers=Dynamic, int Subs=Dynamic, int Options=0> class BandMatrix;
+}
+
+namespace internal {
+template<typename Lhs, typename Rhs> struct product_type;
+
+template<bool> struct EnableIf;
+
+/** \internal
+  * \class product_evaluator
+  * Products need their own evaluator with more template arguments allowing for
+  * easier partial template specializations.
+  */
+template< typename T,
+          int ProductTag = internal::product_type<typename T::Lhs,typename T::Rhs>::ret,
+          typename LhsShape = typename evaluator_traits<typename T::Lhs>::Shape,
+          typename RhsShape = typename evaluator_traits<typename T::Rhs>::Shape,
+          typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
+          typename RhsScalar = typename traits<typename T::Rhs>::Scalar
+        > struct product_evaluator;
+}
+
+template<typename Lhs, typename Rhs,
+         int ProductType = internal::product_type<Lhs,Rhs>::value>
+struct ProductReturnType;
+
+// this is a workaround for sun CC
+template<typename Lhs, typename Rhs> struct LazyProductReturnType;
+
+namespace internal {
+
+// Provides scalar/packet-wise product and product with accumulation
+// with optional conjugation of the arguments.
+template<typename LhsScalar, typename RhsScalar, bool ConjLhs=false, bool ConjRhs=false> struct conj_helper;
+
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_sum_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_difference_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_conj_product_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_min_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_max_op;
+template<typename Scalar> struct scalar_opposite_op;
+template<typename Scalar> struct scalar_conjugate_op;
+template<typename Scalar> struct scalar_real_op;
+template<typename Scalar> struct scalar_imag_op;
+template<typename Scalar> struct scalar_abs_op;
+template<typename Scalar> struct scalar_abs2_op;
+template<typename Scalar> struct scalar_sqrt_op;
+template<typename Scalar> struct scalar_rsqrt_op;
+template<typename Scalar> struct scalar_exp_op;
+template<typename Scalar> struct scalar_log_op;
+template<typename Scalar> struct scalar_cos_op;
+template<typename Scalar> struct scalar_sin_op;
+template<typename Scalar> struct scalar_acos_op;
+template<typename Scalar> struct scalar_asin_op;
+template<typename Scalar> struct scalar_tan_op;
+template<typename Scalar> struct scalar_inverse_op;
+template<typename Scalar> struct scalar_square_op;
+template<typename Scalar> struct scalar_cube_op;
+template<typename Scalar, typename NewType> struct scalar_cast_op;
+template<typename Scalar> struct scalar_random_op;
+template<typename Scalar> struct scalar_constant_op;
+template<typename Scalar> struct scalar_identity_op;
+template<typename Scalar,bool iscpx> struct scalar_sign_op;
+template<typename Scalar,typename ScalarExponent> struct scalar_pow_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_hypot_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_product_op;
+template<typename LhsScalar,typename RhsScalar=LhsScalar> struct scalar_quotient_op;
+
+// SpecialFunctions module
+template<typename Scalar> struct scalar_lgamma_op;
+template<typename Scalar> struct scalar_digamma_op;
+template<typename Scalar> struct scalar_erf_op;
+template<typename Scalar> struct scalar_erfc_op;
+template<typename Scalar> struct scalar_igamma_op;
+template<typename Scalar> struct scalar_igammac_op;
+template<typename Scalar> struct scalar_zeta_op;
+template<typename Scalar> struct scalar_betainc_op;
+
+} // end namespace internal
+
+struct IOFormat;
+
+// Array module
+template<typename _Scalar, int _Rows, int _Cols,
+         int _Options = AutoAlign |
+#if EIGEN_GNUC_AT(3,4)
+    // workaround a bug in at least gcc 3.4.6
+    // the innermost ?: ternary operator is misparsed. We write it slightly
+    // differently and this makes gcc 3.4.6 happy, but it's ugly.
+    // The error would only show up with EIGEN_DEFAULT_TO_ROW_MAJOR is defined
+    // (when EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION is RowMajor)
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : !(_Cols==1 && _Rows!=1) ?  EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION
+                          : Eigen::ColMajor ),
+#else
+                          ( (_Rows==1 && _Cols!=1) ? Eigen::RowMajor
+                          : (_Cols==1 && _Rows!=1) ? Eigen::ColMajor
+                          : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ),
+#endif
+         int _MaxRows = _Rows, int _MaxCols = _Cols> class Array;
+template<typename ConditionMatrixType, typename ThenMatrixType, typename ElseMatrixType> class Select;
+template<typename MatrixType, typename BinaryOp, int Direction> class PartialReduxExpr;
+template<typename ExpressionType, int Direction> class VectorwiseOp;
+template<typename MatrixType,int RowFactor,int ColFactor> class Replicate;
+template<typename MatrixType, int Direction = BothDirections> class Reverse;
+
+template<typename MatrixType> class FullPivLU;
+template<typename MatrixType> class PartialPivLU;
+namespace internal {
+template<typename MatrixType> struct inverse_impl;
+}
+template<typename MatrixType> class HouseholderQR;
+template<typename MatrixType> class ColPivHouseholderQR;
+template<typename MatrixType> class FullPivHouseholderQR;
+template<typename MatrixType> class CompleteOrthogonalDecomposition;
+template<typename MatrixType, int QRPreconditioner = ColPivHouseholderQRPreconditioner> class JacobiSVD;
+template<typename MatrixType> class BDCSVD;
+template<typename MatrixType, int UpLo = Lower> class LLT;
+template<typename MatrixType, int UpLo = Lower> class LDLT;
+template<typename VectorsType, typename CoeffsType, int Side=OnTheLeft> class HouseholderSequence;
+template<typename Scalar>     class JacobiRotation;
+
+// Geometry module:
+template<typename Derived, int _Dim> class RotationBase;
+template<typename Lhs, typename Rhs> class Cross;
+template<typename Derived> class QuaternionBase;
+template<typename Scalar> class Rotation2D;
+template<typename Scalar> class AngleAxis;
+template<typename Scalar,int Dim> class Translation;
+template<typename Scalar,int Dim> class AlignedBox;
+template<typename Scalar, int Options = AutoAlign> class Quaternion;
+template<typename Scalar,int Dim,int Mode,int _Options=AutoAlign> class Transform;
+template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class ParametrizedLine;
+template <typename _Scalar, int _AmbientDim, int Options=AutoAlign> class Hyperplane;
+template<typename Scalar> class UniformScaling;
+template<typename MatrixType,int Direction> class Homogeneous;
+
+// Sparse module:
+template<typename Derived> class SparseMatrixBase;
+
+// MatrixFunctions module
+template<typename Derived> struct MatrixExponentialReturnValue;
+template<typename Derived> class MatrixFunctionReturnValue;
+template<typename Derived> class MatrixSquareRootReturnValue;
+template<typename Derived> class MatrixLogarithmReturnValue;
+template<typename Derived> class MatrixPowerReturnValue;
+template<typename Derived> class MatrixComplexPowerReturnValue;
+
+namespace internal {
+template <typename Scalar>
+struct stem_function
+{
+  typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
+  typedef ComplexScalar type(ComplexScalar, int);
+};
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_FORWARDDECLARATIONS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h
new file mode 100755
index 0000000..26b5966
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/MKL_support.h
@@ -0,0 +1,128 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL
+ *   Include file with common MKL declarations
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_MKL_SUPPORT_H
+#define EIGEN_MKL_SUPPORT_H
+
+#ifdef EIGEN_USE_MKL_ALL
+  #ifndef EIGEN_USE_BLAS
+    #define EIGEN_USE_BLAS
+  #endif
+  #ifndef EIGEN_USE_LAPACKE
+    #define EIGEN_USE_LAPACKE
+  #endif
+  #ifndef EIGEN_USE_MKL_VML
+    #define EIGEN_USE_MKL_VML
+  #endif
+#endif
+
+#ifdef EIGEN_USE_LAPACKE_STRICT
+  #define EIGEN_USE_LAPACKE
+#endif
+
+#if defined(EIGEN_USE_MKL_VML)
+  #define EIGEN_USE_MKL
+#endif
+
+#if defined EIGEN_USE_MKL
+#   include <mkl.h> 
+/*Check IMKL version for compatibility: < 10.3 is not usable with Eigen*/
+#   ifndef INTEL_MKL_VERSION
+#       undef EIGEN_USE_MKL /* INTEL_MKL_VERSION is not even defined on older versions */
+#   elif INTEL_MKL_VERSION < 100305    /* the intel-mkl-103-release-notes say this was when the lapacke.h interface was added*/
+#       undef EIGEN_USE_MKL
+#   endif
+#   ifndef EIGEN_USE_MKL
+    /*If the MKL version is too old, undef everything*/
+#       undef   EIGEN_USE_MKL_ALL
+#       undef   EIGEN_USE_LAPACKE
+#       undef   EIGEN_USE_MKL_VML
+#       undef   EIGEN_USE_LAPACKE_STRICT
+#       undef   EIGEN_USE_LAPACKE
+#   endif
+#endif
+
+#if defined EIGEN_USE_MKL
+
+#define EIGEN_MKL_VML_THRESHOLD 128
+
+/* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */
+/* MKL_BLAS, etc are not defined in 11.2 */
+#ifdef MKL_DOMAIN_ALL
+#define EIGEN_MKL_DOMAIN_ALL MKL_DOMAIN_ALL
+#else
+#define EIGEN_MKL_DOMAIN_ALL MKL_ALL
+#endif
+
+#ifdef MKL_DOMAIN_BLAS
+#define EIGEN_MKL_DOMAIN_BLAS MKL_DOMAIN_BLAS
+#else
+#define EIGEN_MKL_DOMAIN_BLAS MKL_BLAS
+#endif
+
+#ifdef MKL_DOMAIN_FFT
+#define EIGEN_MKL_DOMAIN_FFT MKL_DOMAIN_FFT
+#else
+#define EIGEN_MKL_DOMAIN_FFT MKL_FFT
+#endif
+
+#ifdef MKL_DOMAIN_VML
+#define EIGEN_MKL_DOMAIN_VML MKL_DOMAIN_VML
+#else
+#define EIGEN_MKL_DOMAIN_VML MKL_VML
+#endif
+
+#ifdef MKL_DOMAIN_PARDISO
+#define EIGEN_MKL_DOMAIN_PARDISO MKL_DOMAIN_PARDISO
+#else
+#define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO
+#endif
+#endif
+
+namespace Eigen {
+
+typedef std::complex<double> dcomplex;
+typedef std::complex<float>  scomplex;
+
+#if defined(EIGEN_USE_MKL)
+typedef MKL_INT BlasIndex;
+#else
+typedef int BlasIndex;
+#endif
+
+} // end namespace Eigen
+
+#if defined(EIGEN_USE_BLAS)
+#include "../../misc/blas.h"
+#endif
+
+#endif // EIGEN_MKL_SUPPORT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h
new file mode 100644
index 0000000..38d6ddb
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Macros.h
@@ -0,0 +1,992 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MACROS_H
+#define EIGEN_MACROS_H
+
+#define EIGEN_WORLD_VERSION 3
+#define EIGEN_MAJOR_VERSION 3
+#define EIGEN_MINOR_VERSION 4
+
+#define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
+                                      (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
+                                                                 EIGEN_MINOR_VERSION>=z))))
+
+// Compiler identification, EIGEN_COMP_*
+
+/// \internal EIGEN_COMP_GNUC set to 1 for all compilers compatible with GCC
+#ifdef __GNUC__
+  #define EIGEN_COMP_GNUC 1
+#else
+  #define EIGEN_COMP_GNUC 0
+#endif
+
+/// \internal EIGEN_COMP_CLANG set to major+minor version (e.g., 307 for clang 3.7) if the compiler is clang
+#if defined(__clang__)
+  #define EIGEN_COMP_CLANG (__clang_major__*100+__clang_minor__)
+#else
+  #define EIGEN_COMP_CLANG 0
+#endif
+
+
+/// \internal EIGEN_COMP_LLVM set to 1 if the compiler backend is llvm
+#if defined(__llvm__)
+  #define EIGEN_COMP_LLVM 1
+#else
+  #define EIGEN_COMP_LLVM 0
+#endif
+
+/// \internal EIGEN_COMP_ICC set to __INTEL_COMPILER if the compiler is Intel compiler, 0 otherwise
+#if defined(__INTEL_COMPILER)
+  #define EIGEN_COMP_ICC __INTEL_COMPILER
+#else
+  #define EIGEN_COMP_ICC 0
+#endif
+
+/// \internal EIGEN_COMP_MINGW set to 1 if the compiler is mingw
+#if defined(__MINGW32__)
+  #define EIGEN_COMP_MINGW 1
+#else
+  #define EIGEN_COMP_MINGW 0
+#endif
+
+/// \internal EIGEN_COMP_SUNCC set to 1 if the compiler is Solaris Studio
+#if defined(__SUNPRO_CC)
+  #define EIGEN_COMP_SUNCC 1
+#else
+  #define EIGEN_COMP_SUNCC 0
+#endif
+
+/// \internal EIGEN_COMP_MSVC set to _MSC_VER if the compiler is Microsoft Visual C++, 0 otherwise.
+#if defined(_MSC_VER)
+  #define EIGEN_COMP_MSVC _MSC_VER
+#else
+  #define EIGEN_COMP_MSVC 0
+#endif
+
+// For the record, here is a table summarizing the possible values for EIGEN_COMP_MSVC:
+//  name  ver   MSC_VER
+//  2008    9      1500
+//  2010   10      1600
+//  2012   11      1700
+//  2013   12      1800
+//  2015   14      1900
+//  "15"   15      1900
+
+/// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC or clang-cl
+#if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC || EIGEN_COMP_LLVM || EIGEN_COMP_CLANG)
+  #define EIGEN_COMP_MSVC_STRICT _MSC_VER
+#else
+  #define EIGEN_COMP_MSVC_STRICT 0
+#endif
+
+/// \internal EIGEN_COMP_IBM set to 1 if the compiler is IBM XL C++
+#if defined(__IBMCPP__) || defined(__xlc__)
+  #define EIGEN_COMP_IBM 1
+#else
+  #define EIGEN_COMP_IBM 0
+#endif
+
+/// \internal EIGEN_COMP_PGI set to 1 if the compiler is Portland Group Compiler
+#if defined(__PGI)
+  #define EIGEN_COMP_PGI 1
+#else
+  #define EIGEN_COMP_PGI 0
+#endif
+
+/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler
+#if defined(__CC_ARM) || defined(__ARMCC_VERSION)
+  #define EIGEN_COMP_ARM 1
+#else
+  #define EIGEN_COMP_ARM 0
+#endif
+
+/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler
+#if defined(__EMSCRIPTEN__)
+  #define EIGEN_COMP_EMSCRIPTEN 1
+#else
+  #define EIGEN_COMP_EMSCRIPTEN 0
+#endif
+
+
+/// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.)
+#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM || EIGEN_COMP_EMSCRIPTEN)
+  #define EIGEN_COMP_GNUC_STRICT 1
+#else
+  #define EIGEN_COMP_GNUC_STRICT 0
+#endif
+
+
+#if EIGEN_COMP_GNUC
+  #define EIGEN_GNUC_AT_LEAST(x,y) ((__GNUC__==x && __GNUC_MINOR__>=y) || __GNUC__>x)
+  #define EIGEN_GNUC_AT_MOST(x,y)  ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x)
+  #define EIGEN_GNUC_AT(x,y)       ( __GNUC__==x && __GNUC_MINOR__==y )
+#else
+  #define EIGEN_GNUC_AT_LEAST(x,y) 0
+  #define EIGEN_GNUC_AT_MOST(x,y)  0
+  #define EIGEN_GNUC_AT(x,y)       0
+#endif
+
+// FIXME: could probably be removed as we do not support gcc 3.x anymore
+#if EIGEN_COMP_GNUC && (__GNUC__ <= 3)
+#define EIGEN_GCC3_OR_OLDER 1
+#else
+#define EIGEN_GCC3_OR_OLDER 0
+#endif
+
+
+// Architecture identification, EIGEN_ARCH_*
+
+#if defined(__x86_64__) || defined(_M_X64) || defined(__amd64)
+  #define EIGEN_ARCH_x86_64 1
+#else
+  #define EIGEN_ARCH_x86_64 0
+#endif
+
+#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) || defined(__i386)
+  #define EIGEN_ARCH_i386 1
+#else
+  #define EIGEN_ARCH_i386 0
+#endif
+
+#if EIGEN_ARCH_x86_64 || EIGEN_ARCH_i386
+  #define EIGEN_ARCH_i386_OR_x86_64 1
+#else
+  #define EIGEN_ARCH_i386_OR_x86_64 0
+#endif
+
+/// \internal EIGEN_ARCH_ARM set to 1 if the architecture is ARM
+#if defined(__arm__)
+  #define EIGEN_ARCH_ARM 1
+#else
+  #define EIGEN_ARCH_ARM 0
+#endif
+
+/// \internal EIGEN_ARCH_ARM64 set to 1 if the architecture is ARM64
+#if defined(__aarch64__)
+  #define EIGEN_ARCH_ARM64 1
+#else
+  #define EIGEN_ARCH_ARM64 0
+#endif
+
+#if EIGEN_ARCH_ARM || EIGEN_ARCH_ARM64
+  #define EIGEN_ARCH_ARM_OR_ARM64 1
+#else
+  #define EIGEN_ARCH_ARM_OR_ARM64 0
+#endif
+
+/// \internal EIGEN_ARCH_MIPS set to 1 if the architecture is MIPS
+#if defined(__mips__) || defined(__mips)
+  #define EIGEN_ARCH_MIPS 1
+#else
+  #define EIGEN_ARCH_MIPS 0
+#endif
+
+/// \internal EIGEN_ARCH_SPARC set to 1 if the architecture is SPARC
+#if defined(__sparc__) || defined(__sparc)
+  #define EIGEN_ARCH_SPARC 1
+#else
+  #define EIGEN_ARCH_SPARC 0
+#endif
+
+/// \internal EIGEN_ARCH_IA64 set to 1 if the architecture is Intel Itanium
+#if defined(__ia64__)
+  #define EIGEN_ARCH_IA64 1
+#else
+  #define EIGEN_ARCH_IA64 0
+#endif
+
+/// \internal EIGEN_ARCH_PPC set to 1 if the architecture is PowerPC
+#if defined(__powerpc__) || defined(__ppc__) || defined(_M_PPC)
+  #define EIGEN_ARCH_PPC 1
+#else
+  #define EIGEN_ARCH_PPC 0
+#endif
+
+
+
+// Operating system identification, EIGEN_OS_*
+
+/// \internal EIGEN_OS_UNIX set to 1 if the OS is a unix variant
+#if defined(__unix__) || defined(__unix)
+  #define EIGEN_OS_UNIX 1
+#else
+  #define EIGEN_OS_UNIX 0
+#endif
+
+/// \internal EIGEN_OS_LINUX set to 1 if the OS is based on Linux kernel
+#if defined(__linux__)
+  #define EIGEN_OS_LINUX 1
+#else
+  #define EIGEN_OS_LINUX 0
+#endif
+
+/// \internal EIGEN_OS_ANDROID set to 1 if the OS is Android
+// note: ANDROID is defined when using ndk_build, __ANDROID__ is defined when using a standalone toolchain.
+#if defined(__ANDROID__) || defined(ANDROID)
+  #define EIGEN_OS_ANDROID 1
+#else
+  #define EIGEN_OS_ANDROID 0
+#endif
+
+/// \internal EIGEN_OS_GNULINUX set to 1 if the OS is GNU Linux and not Linux-based OS (e.g., not android)
+#if defined(__gnu_linux__) && !(EIGEN_OS_ANDROID)
+  #define EIGEN_OS_GNULINUX 1
+#else
+  #define EIGEN_OS_GNULINUX 0
+#endif
+
+/// \internal EIGEN_OS_BSD set to 1 if the OS is a BSD variant
+#if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__bsdi__) || defined(__DragonFly__)
+  #define EIGEN_OS_BSD 1
+#else
+  #define EIGEN_OS_BSD 0
+#endif
+
+/// \internal EIGEN_OS_MAC set to 1 if the OS is MacOS
+#if defined(__APPLE__)
+  #define EIGEN_OS_MAC 1
+#else
+  #define EIGEN_OS_MAC 0
+#endif
+
+/// \internal EIGEN_OS_QNX set to 1 if the OS is QNX
+#if defined(__QNX__)
+  #define EIGEN_OS_QNX 1
+#else
+  #define EIGEN_OS_QNX 0
+#endif
+
+/// \internal EIGEN_OS_WIN set to 1 if the OS is Windows based
+#if defined(_WIN32)
+  #define EIGEN_OS_WIN 1
+#else
+  #define EIGEN_OS_WIN 0
+#endif
+
+/// \internal EIGEN_OS_WIN64 set to 1 if the OS is Windows 64bits
+#if defined(_WIN64)
+  #define EIGEN_OS_WIN64 1
+#else
+  #define EIGEN_OS_WIN64 0
+#endif
+
+/// \internal EIGEN_OS_WINCE set to 1 if the OS is Windows CE
+#if defined(_WIN32_WCE)
+  #define EIGEN_OS_WINCE 1
+#else
+  #define EIGEN_OS_WINCE 0
+#endif
+
+/// \internal EIGEN_OS_CYGWIN set to 1 if the OS is Windows/Cygwin
+#if defined(__CYGWIN__)
+  #define EIGEN_OS_CYGWIN 1
+#else
+  #define EIGEN_OS_CYGWIN 0
+#endif
+
+/// \internal EIGEN_OS_WIN_STRICT set to 1 if the OS is really Windows and not some variants
+#if EIGEN_OS_WIN && !( EIGEN_OS_WINCE || EIGEN_OS_CYGWIN )
+  #define EIGEN_OS_WIN_STRICT 1
+#else
+  #define EIGEN_OS_WIN_STRICT 0
+#endif
+
+/// \internal EIGEN_OS_SUN set to 1 if the OS is SUN
+#if (defined(sun) || defined(__sun)) && !(defined(__SVR4) || defined(__svr4__))
+  #define EIGEN_OS_SUN 1
+#else
+  #define EIGEN_OS_SUN 0
+#endif
+
+/// \internal EIGEN_OS_SOLARIS set to 1 if the OS is Solaris
+#if (defined(sun) || defined(__sun)) && (defined(__SVR4) || defined(__svr4__))
+  #define EIGEN_OS_SOLARIS 1
+#else
+  #define EIGEN_OS_SOLARIS 0
+#endif
+
+
+
+#if EIGEN_GNUC_AT_MOST(4,3) && !EIGEN_COMP_CLANG
+  // see bug 89
+  #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 0
+#else
+  #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1
+#endif
+
+// This macro can be used to prevent from macro expansion, e.g.:
+//   std::max EIGEN_NOT_A_MACRO(a,b)
+#define EIGEN_NOT_A_MACRO
+
+#ifdef EIGEN_DEFAULT_TO_ROW_MAJOR
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::RowMajor
+#else
+#define EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION Eigen::ColMajor
+#endif
+
+#ifndef EIGEN_DEFAULT_DENSE_INDEX_TYPE
+#define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
+#endif
+
+// Cross compiler wrapper around LLVM's __has_builtin
+#ifdef __has_builtin
+#  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
+#else
+#  define EIGEN_HAS_BUILTIN(x) 0
+#endif
+
+// A Clang feature extension to determine compiler features.
+// We use it to determine 'cxx_rvalue_references'
+#ifndef __has_feature
+# define __has_feature(x) 0
+#endif
+
+// Upperbound on the C++ version to use.
+// Expected values are 03, 11, 14, 17, etc.
+// By default, let's use an arbitrarily large C++ version.
+#ifndef EIGEN_MAX_CPP_VER
+#define EIGEN_MAX_CPP_VER 99
+#endif
+
+#if EIGEN_MAX_CPP_VER>=11 && (defined(__cplusplus) && (__cplusplus >= 201103L) || EIGEN_COMP_MSVC >= 1900)
+#define EIGEN_HAS_CXX11 1
+#else
+#define EIGEN_HAS_CXX11 0
+#endif
+
+
+// Do we support r-value references?
+#ifndef EIGEN_HAS_RVALUE_REFERENCES
+#if EIGEN_MAX_CPP_VER>=11 && \
+    (__has_feature(cxx_rvalue_references) || \
+    (defined(__cplusplus) && __cplusplus >= 201103L) || \
+    (EIGEN_COMP_MSVC >= 1600))
+  #define EIGEN_HAS_RVALUE_REFERENCES 1
+#else
+  #define EIGEN_HAS_RVALUE_REFERENCES 0
+#endif
+#endif
+
+// Does the compiler support C99?
+#ifndef EIGEN_HAS_C99_MATH
+#if EIGEN_MAX_CPP_VER>=11 && \
+    ((defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901))       \
+  || (defined(__GNUC__) && defined(_GLIBCXX_USE_C99)) \
+  || (defined(_LIBCPP_VERSION) && !defined(_MSC_VER)))
+  #define EIGEN_HAS_C99_MATH 1
+#else
+  #define EIGEN_HAS_C99_MATH 0
+#endif
+#endif
+
+// Does the compiler support result_of?
+#ifndef EIGEN_HAS_STD_RESULT_OF
+#if EIGEN_MAX_CPP_VER>=11 && ((__has_feature(cxx_lambdas) || (defined(__cplusplus) && __cplusplus >= 201103L)))
+#define EIGEN_HAS_STD_RESULT_OF 1
+#else
+#define EIGEN_HAS_STD_RESULT_OF 0
+#endif
+#endif
+
+// Does the compiler support variadic templates?
+#ifndef EIGEN_HAS_VARIADIC_TEMPLATES
+#if EIGEN_MAX_CPP_VER>=11 && (__cplusplus > 199711L || EIGEN_COMP_MSVC >= 1900) \
+  && ( !defined(__NVCC__) || !EIGEN_ARCH_ARM_OR_ARM64 || (defined __CUDACC_VER__ && __CUDACC_VER__ >= 80000) )
+    // ^^ Disable the use of variadic templates when compiling with versions of nvcc older than 8.0 on ARM devices:
+    //    this prevents nvcc from crashing when compiling Eigen on Tegra X1
+#define EIGEN_HAS_VARIADIC_TEMPLATES 1
+#else
+#define EIGEN_HAS_VARIADIC_TEMPLATES 0
+#endif
+#endif
+
+// Does the compiler fully support const expressions? (as in c++14)
+#ifndef EIGEN_HAS_CONSTEXPR
+
+#ifdef __CUDACC__
+// Const expressions are supported provided that c++11 is enabled and we're using either clang or nvcc 7.5 or above
+#if EIGEN_MAX_CPP_VER>=14 && (__cplusplus > 199711L && defined(__CUDACC_VER__) && (EIGEN_COMP_CLANG || __CUDACC_VER__ >= 70500))
+  #define EIGEN_HAS_CONSTEXPR 1
+#endif
+#elif EIGEN_MAX_CPP_VER>=14 && (__has_feature(cxx_relaxed_constexpr) || (defined(__cplusplus) && __cplusplus >= 201402L) || \
+  (EIGEN_GNUC_AT_LEAST(4,8) && (__cplusplus > 199711L)))
+#define EIGEN_HAS_CONSTEXPR 1
+#endif
+
+#ifndef EIGEN_HAS_CONSTEXPR
+#define EIGEN_HAS_CONSTEXPR 0
+#endif
+
+#endif
+
+// Does the compiler support C++11 math?
+// Let's be conservative and enable the default C++11 implementation only if we are sure it exists
+#ifndef EIGEN_HAS_CXX11_MATH
+  #if EIGEN_MAX_CPP_VER>=11 && ((__cplusplus > 201103L) || (__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_MSVC || EIGEN_COMP_ICC)  \
+      && (EIGEN_ARCH_i386_OR_x86_64) && (EIGEN_OS_GNULINUX || EIGEN_OS_WIN_STRICT || EIGEN_OS_MAC))
+    #define EIGEN_HAS_CXX11_MATH 1
+  #else
+    #define EIGEN_HAS_CXX11_MATH 0
+  #endif
+#endif
+
+// Does the compiler support proper C++11 containers?
+#ifndef EIGEN_HAS_CXX11_CONTAINERS
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         ((__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900)
+    #define EIGEN_HAS_CXX11_CONTAINERS 1
+  #else
+    #define EIGEN_HAS_CXX11_CONTAINERS 0
+  #endif
+#endif
+
+// Does the compiler support C++11 noexcept?
+#ifndef EIGEN_HAS_CXX11_NOEXCEPT
+  #if    EIGEN_MAX_CPP_VER>=11 && \
+         (__has_feature(cxx_noexcept) \
+      || (__cplusplus > 201103L) \
+      || ((__cplusplus >= 201103L) && (EIGEN_COMP_GNUC_STRICT || EIGEN_COMP_CLANG || EIGEN_COMP_ICC>=1400)) \
+      || EIGEN_COMP_MSVC >= 1900)
+    #define EIGEN_HAS_CXX11_NOEXCEPT 1
+  #else
+    #define EIGEN_HAS_CXX11_NOEXCEPT 0
+  #endif
+#endif
+
+/** Allows to disable some optimizations which might affect the accuracy of the result.
+  * Such optimization are enabled by default, and set EIGEN_FAST_MATH to 0 to disable them.
+  * They currently include:
+  *   - single precision ArrayBase::sin() and ArrayBase::cos() for SSE and AVX vectorization.
+  */
+#ifndef EIGEN_FAST_MATH
+#define EIGEN_FAST_MATH 1
+#endif
+
+#define EIGEN_DEBUG_VAR(x) std::cerr << #x << " = " << x << std::endl;
+
+// concatenate two tokens
+#define EIGEN_CAT2(a,b) a ## b
+#define EIGEN_CAT(a,b) EIGEN_CAT2(a,b)
+
+#define EIGEN_COMMA ,
+
+// convert a token to a string
+#define EIGEN_MAKESTRING2(a) #a
+#define EIGEN_MAKESTRING(a) EIGEN_MAKESTRING2(a)
+
+// EIGEN_STRONG_INLINE is a stronger version of the inline, using __forceinline on MSVC,
+// but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline
+// but GCC is still doing fine with just inline.
+#if EIGEN_COMP_MSVC || EIGEN_COMP_ICC
+#define EIGEN_STRONG_INLINE __forceinline
+#else
+#define EIGEN_STRONG_INLINE inline
+#endif
+
+// EIGEN_ALWAYS_INLINE is the stronget, it has the effect of making the function inline and adding every possible
+// attribute to maximize inlining. This should only be used when really necessary: in particular,
+// it uses __attribute__((always_inline)) on GCC, which most of the time is useless and can severely harm compile times.
+// FIXME with the always_inline attribute,
+// gcc 3.4.x and 4.1 reports the following compilation error:
+//   Eval.h:91: sorry, unimplemented: inlining failed in call to 'const Eigen::Eval<Derived> Eigen::MatrixBase<Scalar, Derived>::eval() const'
+//    : function body not available
+//   See also bug 1367
+#if EIGEN_GNUC_AT_LEAST(4,2)
+#define EIGEN_ALWAYS_INLINE __attribute__((always_inline)) inline
+#else
+#define EIGEN_ALWAYS_INLINE EIGEN_STRONG_INLINE
+#endif
+
+#if EIGEN_COMP_GNUC
+#define EIGEN_DONT_INLINE __attribute__((noinline))
+#elif EIGEN_COMP_MSVC
+#define EIGEN_DONT_INLINE __declspec(noinline)
+#else
+#define EIGEN_DONT_INLINE
+#endif
+
+#if EIGEN_COMP_GNUC
+#define EIGEN_PERMISSIVE_EXPR __extension__
+#else
+#define EIGEN_PERMISSIVE_EXPR
+#endif
+
+// this macro allows to get rid of linking errors about multiply defined functions.
+//  - static is not very good because it prevents definitions from different object files to be merged.
+//           So static causes the resulting linked executable to be bloated with multiple copies of the same function.
+//  - inline is not perfect either as it unwantedly hints the compiler toward inlining the function.
+#define EIGEN_DECLARE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+#define EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS inline
+
+#ifdef NDEBUG
+# ifndef EIGEN_NO_DEBUG
+#  define EIGEN_NO_DEBUG
+# endif
+#endif
+
+// eigen_plain_assert is where we implement the workaround for the assert() bug in GCC <= 4.3, see bug 89
+#ifdef EIGEN_NO_DEBUG
+  #define eigen_plain_assert(x)
+#else
+  #if EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO
+    namespace Eigen {
+    namespace internal {
+    inline bool copy_bool(bool b) { return b; }
+    }
+    }
+    #define eigen_plain_assert(x) assert(x)
+  #else
+    // work around bug 89
+    #include <cstdlib>   // for abort
+    #include <iostream>  // for std::cerr
+
+    namespace Eigen {
+    namespace internal {
+    // trivial function copying a bool. Must be EIGEN_DONT_INLINE, so we implement it after including Eigen headers.
+    // see bug 89.
+    namespace {
+    EIGEN_DONT_INLINE bool copy_bool(bool b) { return b; }
+    }
+    inline void assert_fail(const char *condition, const char *function, const char *file, int line)
+    {
+      std::cerr << "assertion failed: " << condition << " in function " << function << " at " << file << ":" << line << std::endl;
+      abort();
+    }
+    }
+    }
+    #define eigen_plain_assert(x) \
+      do { \
+        if(!Eigen::internal::copy_bool(x)) \
+          Eigen::internal::assert_fail(EIGEN_MAKESTRING(x), __PRETTY_FUNCTION__, __FILE__, __LINE__); \
+      } while(false)
+  #endif
+#endif
+
+// eigen_assert can be overridden
+#ifndef eigen_assert
+#define eigen_assert(x) eigen_plain_assert(x)
+#endif
+
+#ifdef EIGEN_INTERNAL_DEBUGGING
+#define eigen_internal_assert(x) eigen_assert(x)
+#else
+#define eigen_internal_assert(x)
+#endif
+
+#ifdef EIGEN_NO_DEBUG
+#define EIGEN_ONLY_USED_FOR_DEBUG(x) EIGEN_UNUSED_VARIABLE(x)
+#else
+#define EIGEN_ONLY_USED_FOR_DEBUG(x)
+#endif
+
+#ifndef EIGEN_NO_DEPRECATED_WARNING
+  #if EIGEN_COMP_GNUC
+    #define EIGEN_DEPRECATED __attribute__((deprecated))
+  #elif EIGEN_COMP_MSVC
+    #define EIGEN_DEPRECATED __declspec(deprecated)
+  #else
+    #define EIGEN_DEPRECATED
+  #endif
+#else
+  #define EIGEN_DEPRECATED
+#endif
+
+#if EIGEN_COMP_GNUC
+#define EIGEN_UNUSED __attribute__((unused))
+#else
+#define EIGEN_UNUSED
+#endif
+
+// Suppresses 'unused variable' warnings.
+namespace Eigen {
+  namespace internal {
+    template<typename T> EIGEN_DEVICE_FUNC void ignore_unused_variable(const T&) {}
+  }
+}
+#define EIGEN_UNUSED_VARIABLE(var) Eigen::internal::ignore_unused_variable(var);
+
+#if !defined(EIGEN_ASM_COMMENT)
+  #if EIGEN_COMP_GNUC && (EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64)
+    #define EIGEN_ASM_COMMENT(X)  __asm__("#" X)
+  #else
+    #define EIGEN_ASM_COMMENT(X)
+  #endif
+#endif
+
+
+//------------------------------------------------------------------------------------------
+// Static and dynamic alignment control
+//
+// The main purpose of this section is to define EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES
+// as the maximal boundary in bytes on which dynamically and statically allocated data may be alignment respectively.
+// The values of EIGEN_MAX_ALIGN_BYTES and EIGEN_MAX_STATIC_ALIGN_BYTES can be specified by the user. If not,
+// a default value is automatically computed based on architecture, compiler, and OS.
+//
+// This section also defines macros EIGEN_ALIGN_TO_BOUNDARY(N) and the shortcuts EIGEN_ALIGN{8,16,32,_MAX}
+// to be used to declare statically aligned buffers.
+//------------------------------------------------------------------------------------------
+
+
+/* EIGEN_ALIGN_TO_BOUNDARY(n) forces data to be n-byte aligned. This is used to satisfy SIMD requirements.
+ * However, we do that EVEN if vectorization (EIGEN_VECTORIZE) is disabled,
+ * so that vectorization doesn't affect binary compatibility.
+ *
+ * If we made alignment depend on whether or not EIGEN_VECTORIZE is defined, it would be impossible to link
+ * vectorized and non-vectorized code.
+ */
+#if (defined __CUDACC__)
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __align__(n)
+#elif EIGEN_COMP_GNUC || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#elif EIGEN_COMP_MSVC
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __declspec(align(n))
+#elif EIGEN_COMP_SUNCC
+  // FIXME not sure about this one:
+  #define EIGEN_ALIGN_TO_BOUNDARY(n) __attribute__((aligned(n)))
+#else
+  #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
+#endif
+
+// If the user explicitly disable vectorization, then we also disable alignment
+#if defined(EIGEN_DONT_VECTORIZE)
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 0
+#elif defined(EIGEN_VECTORIZE_AVX512)
+  // 64 bytes static alignmeent is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 64
+#elif defined(__AVX__)
+  // 32 bytes static alignmeent is preferred only if really required
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 32
+#else
+  #define EIGEN_IDEAL_MAX_ALIGN_BYTES 16
+#endif
+
+
+// EIGEN_MIN_ALIGN_BYTES defines the minimal value for which the notion of explicit alignment makes sense
+#define EIGEN_MIN_ALIGN_BYTES 16
+
+// Defined the boundary (in bytes) on which the data needs to be aligned. Note
+// that unless EIGEN_ALIGN is defined and not equal to 0, the data may not be
+// aligned at all regardless of the value of this #define.
+
+#if (defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN))  && defined(EIGEN_MAX_STATIC_ALIGN_BYTES) && EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#error EIGEN_MAX_STATIC_ALIGN_BYTES and EIGEN_DONT_ALIGN[_STATICALLY] are both defined with EIGEN_MAX_STATIC_ALIGN_BYTES!=0. Use EIGEN_MAX_STATIC_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN_STATICALLY.
+#endif
+
+// EIGEN_DONT_ALIGN_STATICALLY and EIGEN_DONT_ALIGN are deprectated
+// They imply EIGEN_MAX_STATIC_ALIGN_BYTES=0
+#if defined(EIGEN_DONT_ALIGN_STATICALLY) || defined(EIGEN_DONT_ALIGN)
+  #ifdef EIGEN_MAX_STATIC_ALIGN_BYTES
+    #undef EIGEN_MAX_STATIC_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+#endif
+
+#ifndef EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // Try to automatically guess what is the best default value for EIGEN_MAX_STATIC_ALIGN_BYTES
+
+  // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
+  // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
+  // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
+  // certain common platform (compiler+architecture combinations) to avoid these problems.
+  // Only static alignment is really problematic (relies on nonstandard compiler extensions),
+  // try to keep heap alignment even when we have to disable static alignment.
+  #if EIGEN_COMP_GNUC && !(EIGEN_ARCH_i386_OR_x86_64 || EIGEN_ARCH_ARM_OR_ARM64 || EIGEN_ARCH_PPC || EIGEN_ARCH_IA64)
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #elif EIGEN_ARCH_ARM_OR_ARM64 && EIGEN_COMP_GNUC_STRICT && EIGEN_GNUC_AT_MOST(4, 6)
+  // Old versions of GCC on ARM, at least 4.4, were once seen to have buggy static alignment support.
+  // Not sure which version fixed it, hopefully it doesn't affect 4.7, which is still somewhat in use.
+  // 4.8 and newer seem definitely unaffected.
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 1
+  #else
+  #define EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT 0
+  #endif
+
+  // static alignment is completely disabled with GCC 3, Sun Studio, and QCC/QNX
+  #if !EIGEN_GCC_AND_ARCH_DOESNT_WANT_STACK_ALIGNMENT \
+  && !EIGEN_GCC3_OR_OLDER \
+  && !EIGEN_COMP_SUNCC \
+  && !EIGEN_OS_QNX
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 1
+  #else
+    #define EIGEN_ARCH_WANTS_STACK_ALIGNMENT 0
+  #endif
+
+  #if EIGEN_ARCH_WANTS_STACK_ALIGNMENT
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+  #else
+    #define EIGEN_MAX_STATIC_ALIGN_BYTES 0
+  #endif
+
+#endif
+
+// If EIGEN_MAX_ALIGN_BYTES is defined, then it is considered as an upper bound for EIGEN_MAX_ALIGN_BYTES
+#if defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES<EIGEN_MAX_STATIC_ALIGN_BYTES
+#undef EIGEN_MAX_STATIC_ALIGN_BYTES
+#define EIGEN_MAX_STATIC_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES==0 && !defined(EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT)
+  #define EIGEN_DISABLE_UNALIGNED_ARRAY_ASSERT
+#endif
+
+// At this stage, EIGEN_MAX_STATIC_ALIGN_BYTES>0 is the true test whether we want to align arrays on the stack or not.
+// It takes into account both the user choice to explicitly enable/disable alignment (by settting EIGEN_MAX_STATIC_ALIGN_BYTES)
+// and the architecture config (EIGEN_ARCH_WANTS_STACK_ALIGNMENT).
+// Henceforth, only EIGEN_MAX_STATIC_ALIGN_BYTES should be used.
+
+
+// Shortcuts to EIGEN_ALIGN_TO_BOUNDARY
+#define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
+#define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
+#define EIGEN_ALIGN32 EIGEN_ALIGN_TO_BOUNDARY(32)
+#define EIGEN_ALIGN64 EIGEN_ALIGN_TO_BOUNDARY(64)
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+#define EIGEN_ALIGN_MAX EIGEN_ALIGN_TO_BOUNDARY(EIGEN_MAX_STATIC_ALIGN_BYTES)
+#else
+#define EIGEN_ALIGN_MAX
+#endif
+
+
+// Dynamic alignment control
+
+#if defined(EIGEN_DONT_ALIGN) && defined(EIGEN_MAX_ALIGN_BYTES) && EIGEN_MAX_ALIGN_BYTES>0
+#error EIGEN_MAX_ALIGN_BYTES and EIGEN_DONT_ALIGN are both defined with EIGEN_MAX_ALIGN_BYTES!=0. Use EIGEN_MAX_ALIGN_BYTES=0 as a synonym of EIGEN_DONT_ALIGN.
+#endif
+
+#ifdef EIGEN_DONT_ALIGN
+  #ifdef EIGEN_MAX_ALIGN_BYTES
+    #undef EIGEN_MAX_ALIGN_BYTES
+  #endif
+  #define EIGEN_MAX_ALIGN_BYTES 0
+#elif !defined(EIGEN_MAX_ALIGN_BYTES)
+  #define EIGEN_MAX_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+#endif
+
+#if EIGEN_IDEAL_MAX_ALIGN_BYTES > EIGEN_MAX_ALIGN_BYTES
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_IDEAL_MAX_ALIGN_BYTES
+#else
+#define EIGEN_DEFAULT_ALIGN_BYTES EIGEN_MAX_ALIGN_BYTES
+#endif
+
+
+#ifndef EIGEN_UNALIGNED_VECTORIZE
+#define EIGEN_UNALIGNED_VECTORIZE 1
+#endif
+
+//----------------------------------------------------------------------
+
+
+#ifdef EIGEN_DONT_USE_RESTRICT_KEYWORD
+  #define EIGEN_RESTRICT
+#endif
+#ifndef EIGEN_RESTRICT
+  #define EIGEN_RESTRICT __restrict
+#endif
+
+#ifndef EIGEN_STACK_ALLOCATION_LIMIT
+// 131072 == 128 KB
+#define EIGEN_STACK_ALLOCATION_LIMIT 131072
+#endif
+
+#ifndef EIGEN_DEFAULT_IO_FORMAT
+#ifdef EIGEN_MAKING_DOCS
+// format used in Eigen's documentation
+// needed to define it here as escaping characters in CMake add_definition's argument seems very problematic.
+#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat(3, 0, " ", "\n", "", "")
+#else
+#define EIGEN_DEFAULT_IO_FORMAT Eigen::IOFormat()
+#endif
+#endif
+
+// just an empty macro !
+#define EIGEN_EMPTY
+
+#if EIGEN_COMP_MSVC_STRICT && (EIGEN_COMP_MSVC < 1900 ||  defined(__CUDACC_VER__)) // for older MSVC versions, as well as 1900 && CUDA 8, using the base operator is sufficient (cf Bugs 1000, 1324)
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =;
+#elif EIGEN_COMP_CLANG // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =; \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) { Base::operator=(other); return *this; } \
+    template <typename OtherDerived> \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const DenseBase<OtherDerived>& other) { Base::operator=(other.derived()); return *this; }
+#else
+  #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
+    using Base::operator =; \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const Derived& other) \
+    { \
+      Base::operator=(other); \
+      return *this; \
+    }
+#endif
+
+
+/** \internal
+ * \brief Macro to manually inherit assignment operators.
+ * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined.
+ */
+#define EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Derived) EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived)
+
+/**
+* Just a side note. Commenting within defines works only by documenting
+* behind the object (via '!<'). Comments cannot be multi-line and thus
+* we have these extra long lines. What is confusing doxygen over here is
+* that we use '\' and basically have a bunch of typedefs with their
+* documentation in a single line.
+**/
+
+#define EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
+  typedef typename Eigen::internal::traits<Derived>::Scalar Scalar; /*!< \brief Numeric type, e.g. float, double, int or std::complex<float>. */ \
+  typedef typename Eigen::NumTraits<Scalar>::Real RealScalar; /*!< \brief The underlying numeric type for composed scalar types. \details In cases where Scalar is e.g. std::complex<T>, T were corresponding to RealScalar. */ \
+  typedef typename Base::CoeffReturnType CoeffReturnType; /*!< \brief The return type for coefficient access. \details Depending on whether the object allows direct coefficient access (e.g. for a MatrixXd), this type is either 'const Scalar&' or simply 'Scalar' for objects that do not allow direct coefficient access. */ \
+  typedef typename Eigen::internal::ref_selector<Derived>::type Nested; \
+  typedef typename Eigen::internal::traits<Derived>::StorageKind StorageKind; \
+  typedef typename Eigen::internal::traits<Derived>::StorageIndex StorageIndex; \
+  enum { RowsAtCompileTime = Eigen::internal::traits<Derived>::RowsAtCompileTime, \
+        ColsAtCompileTime = Eigen::internal::traits<Derived>::ColsAtCompileTime, \
+        Flags = Eigen::internal::traits<Derived>::Flags, \
+        SizeAtCompileTime = Base::SizeAtCompileTime, \
+        MaxSizeAtCompileTime = Base::MaxSizeAtCompileTime, \
+        IsVectorAtCompileTime = Base::IsVectorAtCompileTime }; \
+  using Base::derived; \
+  using Base::const_cast_derived;
+
+
+// FIXME Maybe the EIGEN_DENSE_PUBLIC_INTERFACE could be removed as importing PacketScalar is rarely needed
+#define EIGEN_DENSE_PUBLIC_INTERFACE(Derived) \
+  EIGEN_GENERIC_PUBLIC_INTERFACE(Derived) \
+  typedef typename Base::PacketScalar PacketScalar;
+
+
+#define EIGEN_PLAIN_ENUM_MIN(a,b) (((int)a <= (int)b) ? (int)a : (int)b)
+#define EIGEN_PLAIN_ENUM_MAX(a,b) (((int)a >= (int)b) ? (int)a : (int)b)
+
+// EIGEN_SIZE_MIN_PREFER_DYNAMIC gives the min between compile-time sizes. 0 has absolute priority, followed by 1,
+// followed by Dynamic, followed by other finite values. The reason for giving Dynamic the priority over
+// finite values is that min(3, Dynamic) should be Dynamic, since that could be anything between 0 and 3.
+#define EIGEN_SIZE_MIN_PREFER_DYNAMIC(a,b) (((int)a == 0 || (int)b == 0) ? 0 \
+                           : ((int)a == 1 || (int)b == 1) ? 1 \
+                           : ((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
+                           : ((int)a <= (int)b) ? (int)a : (int)b)
+
+// EIGEN_SIZE_MIN_PREFER_FIXED is a variant of EIGEN_SIZE_MIN_PREFER_DYNAMIC comparing MaxSizes. The difference is that finite values
+// now have priority over Dynamic, so that min(3, Dynamic) gives 3. Indeed, whatever the actual value is
+// (between 0 and 3), it is not more than 3.
+#define EIGEN_SIZE_MIN_PREFER_FIXED(a,b)  (((int)a == 0 || (int)b == 0) ? 0 \
+                           : ((int)a == 1 || (int)b == 1) ? 1 \
+                           : ((int)a == Dynamic && (int)b == Dynamic) ? Dynamic \
+                           : ((int)a == Dynamic) ? (int)b \
+                           : ((int)b == Dynamic) ? (int)a \
+                           : ((int)a <= (int)b) ? (int)a : (int)b)
+
+// see EIGEN_SIZE_MIN_PREFER_DYNAMIC. No need for a separate variant for MaxSizes here.
+#define EIGEN_SIZE_MAX(a,b) (((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
+                           : ((int)a >= (int)b) ? (int)a : (int)b)
+
+#define EIGEN_LOGICAL_XOR(a,b) (((a) || (b)) && !((a) && (b)))
+
+#define EIGEN_IMPLIES(a,b) (!(a) || (b))
+
+// the expression type of a standard coefficient wise binary operation
+#define EIGEN_CWISE_BINARY_RETURN_TYPE(LHS,RHS,OPNAME) \
+    CwiseBinaryOp< \
+      EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)< \
+          typename internal::traits<LHS>::Scalar, \
+          typename internal::traits<RHS>::Scalar \
+      >, \
+      const LHS, \
+      const RHS \
+    >
+
+#define EIGEN_MAKE_CWISE_BINARY_OP(METHOD,OPNAME) \
+  template<typename OtherDerived> \
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME) \
+  (METHOD)(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
+  { \
+    return EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,OPNAME)(derived(), other.derived()); \
+  }
+
+#define EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,TYPEA,TYPEB) \
+  (Eigen::internal::has_ReturnType<Eigen::ScalarBinaryOpTraits<TYPEA,TYPEB,EIGEN_CAT(EIGEN_CAT(Eigen::internal::scalar_,OPNAME),_op)<TYPEA,TYPEB>  > >::value)
+
+#define EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(EXPR,SCALAR,OPNAME) \
+  CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<typename internal::traits<EXPR>::Scalar,SCALAR>, const EXPR, \
+                const typename internal::plain_constant_type<EXPR,SCALAR>::type>
+
+#define EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(SCALAR,EXPR,OPNAME) \
+  CwiseBinaryOp<EIGEN_CAT(EIGEN_CAT(internal::scalar_,OPNAME),_op)<SCALAR,typename internal::traits<EXPR>::Scalar>, \
+                const typename internal::plain_constant_type<EXPR,SCALAR>::type, const EXPR>
+
+// Workaround for MSVC 2010 (see ML thread "patch with compile for for MSVC 2010")
+#if EIGEN_COMP_MSVC_STRICT<=1600
+#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) typename internal::enable_if<true,X>::type
+#else
+#define EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(X) X
+#endif
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME) \
+  template <typename T> EIGEN_DEVICE_FUNC inline \
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type,OPNAME))\
+  (METHOD)(const T& scalar) const { \
+    typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,Scalar,T)>::type PromotedT; \
+    return EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,PromotedT,OPNAME)(derived(), \
+           typename internal::plain_constant_type<Derived,PromotedT>::type(derived().rows(), derived().cols(), internal::scalar_constant_op<PromotedT>(scalar))); \
+  }
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \
+  template <typename T> EIGEN_DEVICE_FUNC inline friend \
+  EIGEN_MSVC10_WORKAROUND_BINARYOP_RETURN_TYPE(const EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(typename internal::promote_scalar_arg<Scalar EIGEN_COMMA T EIGEN_COMMA EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type,Derived,OPNAME)) \
+  (METHOD)(const T& scalar, const StorageBaseType& matrix) { \
+    typedef typename internal::promote_scalar_arg<Scalar,T,EIGEN_SCALAR_BINARY_SUPPORTED(OPNAME,T,Scalar)>::type PromotedT; \
+    return EIGEN_SCALAR_BINARYOP_EXPR_RETURN_TYPE(PromotedT,Derived,OPNAME)( \
+           typename internal::plain_constant_type<Derived,PromotedT>::type(matrix.derived().rows(), matrix.derived().cols(), internal::scalar_constant_op<PromotedT>(scalar)), matrix.derived()); \
+  }
+
+#define EIGEN_MAKE_SCALAR_BINARY_OP(METHOD,OPNAME) \
+  EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(METHOD,OPNAME) \
+  EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(METHOD,OPNAME)
+
+
+#ifdef EIGEN_EXCEPTIONS
+#  define EIGEN_THROW_X(X) throw X
+#  define EIGEN_THROW throw
+#  define EIGEN_TRY try
+#  define EIGEN_CATCH(X) catch (X)
+#else
+#  ifdef __CUDA_ARCH__
+#    define EIGEN_THROW_X(X) asm("trap;")
+#    define EIGEN_THROW asm("trap;")
+#  else
+#    define EIGEN_THROW_X(X) std::abort()
+#    define EIGEN_THROW std::abort()
+#  endif
+#  define EIGEN_TRY if (true)
+#  define EIGEN_CATCH(X) else
+#endif
+
+
+#if EIGEN_HAS_CXX11_NOEXCEPT
+#   define EIGEN_INCLUDE_TYPE_TRAITS
+#   define EIGEN_NOEXCEPT noexcept
+#   define EIGEN_NOEXCEPT_IF(x) noexcept(x)
+#   define EIGEN_NO_THROW noexcept(true)
+#   define EIGEN_EXCEPTION_SPEC(X) noexcept(false)
+#else
+#   define EIGEN_NOEXCEPT
+#   define EIGEN_NOEXCEPT_IF(x)
+#   define EIGEN_NO_THROW throw()
+#   define EIGEN_EXCEPTION_SPEC(X) throw(X)
+#endif
+
+#endif // EIGEN_MACROS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h
new file mode 100644
index 0000000..c634d7e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Memory.h
@@ -0,0 +1,977 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Kenneth Riddile <kfriddile@yahoo.com>
+// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
+// Copyright (C) 2010 Thomas Capricelli <orzel@freehackers.org>
+// Copyright (C) 2013 Pavel Holoborodko <pavel@holoborodko.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+/*****************************************************************************
+*** Platform checks for aligned malloc functions                           ***
+*****************************************************************************/
+
+#ifndef EIGEN_MEMORY_H
+#define EIGEN_MEMORY_H
+
+#ifndef EIGEN_MALLOC_ALREADY_ALIGNED
+
+// Try to determine automatically if malloc is already aligned.
+
+// On 64-bit systems, glibc's malloc returns 16-byte-aligned pointers, see:
+//   http://www.gnu.org/s/libc/manual/html_node/Aligned-Memory-Blocks.html
+// This is true at least since glibc 2.8.
+// This leaves the question how to detect 64-bit. According to this document,
+//   http://gcc.fyxm.net/summit/2003/Porting%20to%2064%20bit.pdf
+// page 114, "[The] LP64 model [...] is used by all 64-bit UNIX ports" so it's indeed
+// quite safe, at least within the context of glibc, to equate 64-bit with LP64.
+#if defined(__GLIBC__) && ((__GLIBC__>=2 && __GLIBC_MINOR__ >= 8) || __GLIBC__>2) \
+ && defined(__LP64__) && ! defined( __SANITIZE_ADDRESS__ ) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
+  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+// FreeBSD 6 seems to have 16-byte aligned malloc
+//   See http://svn.freebsd.org/viewvc/base/stable/6/lib/libc/stdlib/malloc.c?view=markup
+// FreeBSD 7 seems to have 16-byte aligned malloc except on ARM and MIPS architectures
+//   See http://svn.freebsd.org/viewvc/base/stable/7/lib/libc/stdlib/malloc.c?view=markup
+#if defined(__FreeBSD__) && !(EIGEN_ARCH_ARM || EIGEN_ARCH_MIPS) && (EIGEN_DEFAULT_ALIGN_BYTES == 16)
+  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+#if (EIGEN_OS_MAC && (EIGEN_DEFAULT_ALIGN_BYTES == 16))     \
+ || (EIGEN_OS_WIN64 && (EIGEN_DEFAULT_ALIGN_BYTES == 16))   \
+ || EIGEN_GLIBC_MALLOC_ALREADY_ALIGNED              \
+ || EIGEN_FREEBSD_MALLOC_ALREADY_ALIGNED
+  #define EIGEN_MALLOC_ALREADY_ALIGNED 1
+#else
+  #define EIGEN_MALLOC_ALREADY_ALIGNED 0
+#endif
+
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+EIGEN_DEVICE_FUNC 
+inline void throw_std_bad_alloc()
+{
+  #ifdef EIGEN_EXCEPTIONS
+    throw std::bad_alloc();
+  #else
+    std::size_t huge = static_cast<std::size_t>(-1);
+    new int[huge];
+  #endif
+}
+
+/*****************************************************************************
+*** Implementation of handmade aligned functions                           ***
+*****************************************************************************/
+
+/* ----- Hand made implementations of aligned malloc/free and realloc ----- */
+
+/** \internal Like malloc, but the returned pointer is guaranteed to be 16-byte aligned.
+  * Fast, but wastes 16 additional bytes of memory. Does not throw any exception.
+  */
+inline void* handmade_aligned_malloc(std::size_t size)
+{
+  void *original = std::malloc(size+EIGEN_DEFAULT_ALIGN_BYTES);
+  if (original == 0) return 0;
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
+  *(reinterpret_cast<void**>(aligned) - 1) = original;
+  return aligned;
+}
+
+/** \internal Frees memory allocated with handmade_aligned_malloc */
+inline void handmade_aligned_free(void *ptr)
+{
+  if (ptr) std::free(*(reinterpret_cast<void**>(ptr) - 1));
+}
+
+/** \internal
+  * \brief Reallocates aligned memory.
+  * Since we know that our handmade version is based on std::malloc
+  * we can use std::realloc to implement efficient reallocation.
+  */
+inline void* handmade_aligned_realloc(void* ptr, std::size_t size, std::size_t = 0)
+{
+  if (ptr == 0) return handmade_aligned_malloc(size);
+  void *original = *(reinterpret_cast<void**>(ptr) - 1);
+  std::ptrdiff_t previous_offset = static_cast<char *>(ptr)-static_cast<char *>(original);
+  original = std::realloc(original,size+EIGEN_DEFAULT_ALIGN_BYTES);
+  if (original == 0) return 0;
+  void *aligned = reinterpret_cast<void*>((reinterpret_cast<std::size_t>(original) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1))) + EIGEN_DEFAULT_ALIGN_BYTES);
+  void *previous_aligned = static_cast<char *>(original)+previous_offset;
+  if(aligned!=previous_aligned)
+    std::memmove(aligned, previous_aligned, size);
+  
+  *(reinterpret_cast<void**>(aligned) - 1) = original;
+  return aligned;
+}
+
+/*****************************************************************************
+*** Implementation of portable aligned versions of malloc/free/realloc     ***
+*****************************************************************************/
+
+#ifdef EIGEN_NO_MALLOC
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
+{
+  eigen_assert(false && "heap allocation is forbidden (EIGEN_NO_MALLOC is defined)");
+}
+#elif defined EIGEN_RUNTIME_NO_MALLOC
+EIGEN_DEVICE_FUNC inline bool is_malloc_allowed_impl(bool update, bool new_value = false)
+{
+  static bool value = true;
+  if (update == 1)
+    value = new_value;
+  return value;
+}
+EIGEN_DEVICE_FUNC inline bool is_malloc_allowed() { return is_malloc_allowed_impl(false); }
+EIGEN_DEVICE_FUNC inline bool set_is_malloc_allowed(bool new_value) { return is_malloc_allowed_impl(true, new_value); }
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
+{
+  eigen_assert(is_malloc_allowed() && "heap allocation is forbidden (EIGEN_RUNTIME_NO_MALLOC is defined and g_is_malloc_allowed is false)");
+}
+#else 
+EIGEN_DEVICE_FUNC inline void check_that_malloc_is_allowed()
+{}
+#endif
+
+/** \internal Allocates \a size bytes. The returned pointer is guaranteed to have 16 or 32 bytes alignment depending on the requirements.
+  * On allocation error, the returned pointer is null, and std::bad_alloc is thrown.
+  */
+EIGEN_DEVICE_FUNC inline void* aligned_malloc(std::size_t size)
+{
+  check_that_malloc_is_allowed();
+
+  void *result;
+  #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
+    result = std::malloc(size);
+    #if EIGEN_DEFAULT_ALIGN_BYTES==16
+    eigen_assert((size<16 || (std::size_t(result)%16)==0) && "System's malloc returned an unaligned pointer. Compile with EIGEN_MALLOC_ALREADY_ALIGNED=0 to fallback to handmade alignd memory allocator.");
+    #endif
+  #else
+    result = handmade_aligned_malloc(size);
+  #endif
+
+  if(!result && size)
+    throw_std_bad_alloc();
+
+  return result;
+}
+
+/** \internal Frees memory allocated with aligned_malloc. */
+EIGEN_DEVICE_FUNC inline void aligned_free(void *ptr)
+{
+  #if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
+    std::free(ptr);
+  #else
+    handmade_aligned_free(ptr);
+  #endif
+}
+
+/**
+  * \internal
+  * \brief Reallocates an aligned block of memory.
+  * \throws std::bad_alloc on allocation failure
+  */
+inline void* aligned_realloc(void *ptr, std::size_t new_size, std::size_t old_size)
+{
+  EIGEN_UNUSED_VARIABLE(old_size);
+
+  void *result;
+#if (EIGEN_DEFAULT_ALIGN_BYTES==0) || EIGEN_MALLOC_ALREADY_ALIGNED
+  result = std::realloc(ptr,new_size);
+#else
+  result = handmade_aligned_realloc(ptr,new_size,old_size);
+#endif
+
+  if (!result && new_size)
+    throw_std_bad_alloc();
+
+  return result;
+}
+
+/*****************************************************************************
+*** Implementation of conditionally aligned functions                      ***
+*****************************************************************************/
+
+/** \internal Allocates \a size bytes. If Align is true, then the returned ptr is 16-byte-aligned.
+  * On allocation error, the returned pointer is null, and a std::bad_alloc is thrown.
+  */
+template<bool Align> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc(std::size_t size)
+{
+  return aligned_malloc(size);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void* conditional_aligned_malloc<false>(std::size_t size)
+{
+  check_that_malloc_is_allowed();
+
+  void *result = std::malloc(size);
+  if(!result && size)
+    throw_std_bad_alloc();
+  return result;
+}
+
+/** \internal Frees memory allocated with conditional_aligned_malloc */
+template<bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_free(void *ptr)
+{
+  aligned_free(ptr);
+}
+
+template<> EIGEN_DEVICE_FUNC inline void conditional_aligned_free<false>(void *ptr)
+{
+  std::free(ptr);
+}
+
+template<bool Align> inline void* conditional_aligned_realloc(void* ptr, std::size_t new_size, std::size_t old_size)
+{
+  return aligned_realloc(ptr, new_size, old_size);
+}
+
+template<> inline void* conditional_aligned_realloc<false>(void* ptr, std::size_t new_size, std::size_t)
+{
+  return std::realloc(ptr, new_size);
+}
+
+/*****************************************************************************
+*** Construction/destruction of array elements                             ***
+*****************************************************************************/
+
+/** \internal Destructs the elements of an array.
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T> EIGEN_DEVICE_FUNC inline void destruct_elements_of_array(T *ptr, std::size_t size)
+{
+  // always destruct an array starting from the end.
+  if(ptr)
+    while(size) ptr[--size].~T();
+}
+
+/** \internal Constructs the elements of an array.
+  * The \a size parameter tells on how many objects to call the constructor of T.
+  */
+template<typename T> EIGEN_DEVICE_FUNC inline T* construct_elements_of_array(T *ptr, std::size_t size)
+{
+  std::size_t i;
+  EIGEN_TRY
+  {
+      for (i = 0; i < size; ++i) ::new (ptr + i) T;
+      return ptr;
+  }
+  EIGEN_CATCH(...)
+  {
+    destruct_elements_of_array(ptr, i);
+    EIGEN_THROW;
+  }
+  return NULL;
+}
+
+/*****************************************************************************
+*** Implementation of aligned new/delete-like functions                    ***
+*****************************************************************************/
+
+template<typename T>
+EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE void check_size_for_overflow(std::size_t size)
+{
+  if(size > std::size_t(-1) / sizeof(T))
+    throw_std_bad_alloc();
+}
+
+/** \internal Allocates \a size objects of type T. The returned pointer is guaranteed to have 16 bytes alignment.
+  * On allocation error, the returned pointer is undefined, but a std::bad_alloc is thrown.
+  * The default constructor of T is called.
+  */
+template<typename T> EIGEN_DEVICE_FUNC inline T* aligned_new(std::size_t size)
+{
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(aligned_malloc(sizeof(T)*size));
+  EIGEN_TRY
+  {
+    return construct_elements_of_array(result, size);
+  }
+  EIGEN_CATCH(...)
+  {
+    aligned_free(result);
+    EIGEN_THROW;
+  }
+  return result;
+}
+
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new(std::size_t size)
+{
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
+  EIGEN_TRY
+  {
+    return construct_elements_of_array(result, size);
+  }
+  EIGEN_CATCH(...)
+  {
+    conditional_aligned_free<Align>(result);
+    EIGEN_THROW;
+  }
+  return result;
+}
+
+/** \internal Deletes objects constructed with aligned_new
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T> EIGEN_DEVICE_FUNC inline void aligned_delete(T *ptr, std::size_t size)
+{
+  destruct_elements_of_array<T>(ptr, size);
+  aligned_free(ptr);
+}
+
+/** \internal Deletes objects constructed with conditional_aligned_new
+  * The \a size parameters tells on how many objects to call the destructor of T.
+  */
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete(T *ptr, std::size_t size)
+{
+  destruct_elements_of_array<T>(ptr, size);
+  conditional_aligned_free<Align>(ptr);
+}
+
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_realloc_new(T* pts, std::size_t new_size, std::size_t old_size)
+{
+  check_size_for_overflow<T>(new_size);
+  check_size_for_overflow<T>(old_size);
+  if(new_size < old_size)
+    destruct_elements_of_array(pts+new_size, old_size-new_size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
+  if(new_size > old_size)
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result+old_size, new_size-old_size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
+  return result;
+}
+
+
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline T* conditional_aligned_new_auto(std::size_t size)
+{
+  if(size==0)
+    return 0; // short-cut. Also fixes Bug 884
+  check_size_for_overflow<T>(size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_malloc<Align>(sizeof(T)*size));
+  if(NumTraits<T>::RequireInitialization)
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result, size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
+  return result;
+}
+
+template<typename T, bool Align> inline T* conditional_aligned_realloc_new_auto(T* pts, std::size_t new_size, std::size_t old_size)
+{
+  check_size_for_overflow<T>(new_size);
+  check_size_for_overflow<T>(old_size);
+  if(NumTraits<T>::RequireInitialization && (new_size < old_size))
+    destruct_elements_of_array(pts+new_size, old_size-new_size);
+  T *result = reinterpret_cast<T*>(conditional_aligned_realloc<Align>(reinterpret_cast<void*>(pts), sizeof(T)*new_size, sizeof(T)*old_size));
+  if(NumTraits<T>::RequireInitialization && (new_size > old_size))
+  {
+    EIGEN_TRY
+    {
+      construct_elements_of_array(result+old_size, new_size-old_size);
+    }
+    EIGEN_CATCH(...)
+    {
+      conditional_aligned_free<Align>(result);
+      EIGEN_THROW;
+    }
+  }
+  return result;
+}
+
+template<typename T, bool Align> EIGEN_DEVICE_FUNC inline void conditional_aligned_delete_auto(T *ptr, std::size_t size)
+{
+  if(NumTraits<T>::RequireInitialization)
+    destruct_elements_of_array<T>(ptr, size);
+  conditional_aligned_free<Align>(ptr);
+}
+
+/****************************************************************************/
+
+/** \internal Returns the index of the first element of the array that is well aligned with respect to the requested \a Alignment.
+  *
+  * \tparam Alignment requested alignment in Bytes.
+  * \param array the address of the start of the array
+  * \param size the size of the array
+  *
+  * \note If no element of the array is well aligned or the requested alignment is not a multiple of a scalar,
+  * the size of the array is returned. For example with SSE, the requested alignment is typically 16-bytes. If
+  * packet size for the given scalar type is 1, then everything is considered well-aligned.
+  *
+  * \note Otherwise, if the Alignment is larger that the scalar size, we rely on the assumptions that sizeof(Scalar) is a
+  * power of 2. On the other hand, we do not assume that the array address is a multiple of sizeof(Scalar), as that fails for
+  * example with Scalar=double on certain 32-bit platforms, see bug #79.
+  *
+  * There is also the variant first_aligned(const MatrixBase&) defined in DenseCoeffsBase.h.
+  * \sa first_default_aligned()
+  */
+template<int Alignment, typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC inline Index first_aligned(const Scalar* array, Index size)
+{
+  const Index ScalarSize = sizeof(Scalar);
+  const Index AlignmentSize = Alignment / ScalarSize;
+  const Index AlignmentMask = AlignmentSize-1;
+
+  if(AlignmentSize<=1)
+  {
+    // Either the requested alignment if smaller than a scalar, or it exactly match a 1 scalar
+    // so that all elements of the array have the same alignment.
+    return 0;
+  }
+  else if( (UIntPtr(array) & (sizeof(Scalar)-1)) || (Alignment%ScalarSize)!=0)
+  {
+    // The array is not aligned to the size of a single scalar, or the requested alignment is not a multiple of the scalar size.
+    // Consequently, no element of the array is well aligned.
+    return size;
+  }
+  else
+  {
+    Index first = (AlignmentSize - (Index((UIntPtr(array)/sizeof(Scalar))) & AlignmentMask)) & AlignmentMask;
+    return (first < size) ? first : size;
+  }
+}
+
+/** \internal Returns the index of the first element of the array that is well aligned with respect the largest packet requirement.
+   * \sa first_aligned(Scalar*,Index) and first_default_aligned(DenseBase<Derived>) */
+template<typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC inline Index first_default_aligned(const Scalar* array, Index size)
+{
+  typedef typename packet_traits<Scalar>::type DefaultPacketType;
+  return first_aligned<unpacket_traits<DefaultPacketType>::alignment>(array, size);
+}
+
+/** \internal Returns the smallest integer multiple of \a base and greater or equal to \a size
+  */ 
+template<typename Index> 
+inline Index first_multiple(Index size, Index base)
+{
+  return ((size+base-1)/base)*base;
+}
+
+// std::copy is much slower than memcpy, so let's introduce a smart_copy which
+// use memcpy on trivial types, i.e., on types that does not require an initialization ctor.
+template<typename T, bool UseMemcpy> struct smart_copy_helper;
+
+template<typename T> EIGEN_DEVICE_FUNC void smart_copy(const T* start, const T* end, T* target)
+{
+  smart_copy_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_copy_helper<T,true> {
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
+  {
+    IntPtr size = IntPtr(end)-IntPtr(start);
+    if(size==0) return;
+    eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    memcpy(target, start, size);
+  }
+};
+
+template<typename T> struct smart_copy_helper<T,false> {
+  EIGEN_DEVICE_FUNC static inline void run(const T* start, const T* end, T* target)
+  { std::copy(start, end, target); }
+};
+
+// intelligent memmove. falls back to std::memmove for POD types, uses std::copy otherwise. 
+template<typename T, bool UseMemmove> struct smart_memmove_helper;
+
+template<typename T> void smart_memmove(const T* start, const T* end, T* target)
+{
+  smart_memmove_helper<T,!NumTraits<T>::RequireInitialization>::run(start, end, target);
+}
+
+template<typename T> struct smart_memmove_helper<T,true> {
+  static inline void run(const T* start, const T* end, T* target)
+  {
+    IntPtr size = IntPtr(end)-IntPtr(start);
+    if(size==0) return;
+    eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    std::memmove(target, start, size);
+  }
+};
+
+template<typename T> struct smart_memmove_helper<T,false> {
+  static inline void run(const T* start, const T* end, T* target)
+  { 
+    if (UIntPtr(target) < UIntPtr(start))
+    {
+      std::copy(start, end, target);
+    }
+    else                                 
+    {
+      std::ptrdiff_t count = (std::ptrdiff_t(end)-std::ptrdiff_t(start)) / sizeof(T);
+      std::copy_backward(start, end, target + count); 
+    }
+  }
+};
+
+
+/*****************************************************************************
+*** Implementation of runtime stack allocation (falling back to malloc)    ***
+*****************************************************************************/
+
+// you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
+// to the appropriate stack allocation function
+#ifndef EIGEN_ALLOCA
+  #if EIGEN_OS_LINUX || EIGEN_OS_MAC || (defined alloca)
+    #define EIGEN_ALLOCA alloca
+  #elif EIGEN_COMP_MSVC
+    #define EIGEN_ALLOCA _alloca
+  #endif
+#endif
+
+// This helper class construct the allocated memory, and takes care of destructing and freeing the handled data
+// at destruction time. In practice this helper class is mainly useful to avoid memory leak in case of exceptions.
+template<typename T> class aligned_stack_memory_handler : noncopyable
+{
+  public:
+    /* Creates a stack_memory_handler responsible for the buffer \a ptr of size \a size.
+     * Note that \a ptr can be 0 regardless of the other parameters.
+     * This constructor takes care of constructing/initializing the elements of the buffer if required by the scalar type T (see NumTraits<T>::RequireInitialization).
+     * In this case, the buffer elements will also be destructed when this handler will be destructed.
+     * Finally, if \a dealloc is true, then the pointer \a ptr is freed.
+     **/
+    aligned_stack_memory_handler(T* ptr, std::size_t size, bool dealloc)
+      : m_ptr(ptr), m_size(size), m_deallocate(dealloc)
+    {
+      if(NumTraits<T>::RequireInitialization && m_ptr)
+        Eigen::internal::construct_elements_of_array(m_ptr, size);
+    }
+    ~aligned_stack_memory_handler()
+    {
+      if(NumTraits<T>::RequireInitialization && m_ptr)
+        Eigen::internal::destruct_elements_of_array<T>(m_ptr, m_size);
+      if(m_deallocate)
+        Eigen::internal::aligned_free(m_ptr);
+    }
+  protected:
+    T* m_ptr;
+    std::size_t m_size;
+    bool m_deallocate;
+};
+
+template<typename T> class scoped_array : noncopyable
+{
+  T* m_ptr;
+public:
+  explicit scoped_array(std::ptrdiff_t size)
+  {
+    m_ptr = new T[size];
+  }
+  ~scoped_array()
+  {
+    delete[] m_ptr;
+  }
+  T& operator[](std::ptrdiff_t i) { return m_ptr[i]; }
+  const T& operator[](std::ptrdiff_t i) const { return m_ptr[i]; }
+  T* &ptr() { return m_ptr; }
+  const T* ptr() const { return m_ptr; }
+  operator const T*() const { return m_ptr; }
+};
+
+template<typename T> void swap(scoped_array<T> &a,scoped_array<T> &b)
+{
+  std::swap(a.ptr(),b.ptr());
+}
+    
+} // end namespace internal
+
+/** \internal
+  * Declares, allocates and construct an aligned buffer named NAME of SIZE elements of type TYPE on the stack
+  * if SIZE is smaller than EIGEN_STACK_ALLOCATION_LIMIT, and if stack allocation is supported by the platform
+  * (currently, this is Linux and Visual Studio only). Otherwise the memory is allocated on the heap.
+  * The allocated buffer is automatically deleted when exiting the scope of this declaration.
+  * If BUFFER is non null, then the declared variable is simply an alias for BUFFER, and no allocation/deletion occurs.
+  * Here is an example:
+  * \code
+  * {
+  *   ei_declare_aligned_stack_constructed_variable(float,data,size,0);
+  *   // use data[0] to data[size-1]
+  * }
+  * \endcode
+  * The underlying stack allocation function can controlled with the EIGEN_ALLOCA preprocessor token.
+  */
+#ifdef EIGEN_ALLOCA
+  
+  #if EIGEN_DEFAULT_ALIGN_BYTES>0
+    // We always manually re-align the result of EIGEN_ALLOCA.
+    // If alloca is already aligned, the compiler should be smart enough to optimize away the re-alignment.
+    #define EIGEN_ALIGNED_ALLOCA(SIZE) reinterpret_cast<void*>((internal::UIntPtr(EIGEN_ALLOCA(SIZE+EIGEN_DEFAULT_ALIGN_BYTES-1)) + EIGEN_DEFAULT_ALIGN_BYTES-1) & ~(std::size_t(EIGEN_DEFAULT_ALIGN_BYTES-1)))
+  #else
+    #define EIGEN_ALIGNED_ALLOCA(SIZE) EIGEN_ALLOCA(SIZE)
+  #endif
+
+  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
+    Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
+    TYPE* NAME = (BUFFER)!=0 ? (BUFFER) \
+               : reinterpret_cast<TYPE*>( \
+                      (sizeof(TYPE)*SIZE<=EIGEN_STACK_ALLOCATION_LIMIT) ? EIGEN_ALIGNED_ALLOCA(sizeof(TYPE)*SIZE) \
+                    : Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE) );  \
+    Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,sizeof(TYPE)*SIZE>EIGEN_STACK_ALLOCATION_LIMIT)
+
+#else
+
+  #define ei_declare_aligned_stack_constructed_variable(TYPE,NAME,SIZE,BUFFER) \
+    Eigen::internal::check_size_for_overflow<TYPE>(SIZE); \
+    TYPE* NAME = (BUFFER)!=0 ? BUFFER : reinterpret_cast<TYPE*>(Eigen::internal::aligned_malloc(sizeof(TYPE)*SIZE));    \
+    Eigen::internal::aligned_stack_memory_handler<TYPE> EIGEN_CAT(NAME,_stack_memory_destructor)((BUFFER)==0 ? NAME : 0,SIZE,true)
+    
+#endif
+
+
+/*****************************************************************************
+*** Implementation of EIGEN_MAKE_ALIGNED_OPERATOR_NEW [_IF]                ***
+*****************************************************************************/
+
+#if EIGEN_MAX_ALIGN_BYTES!=0
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void* operator new(std::size_t size, const std::nothrow_t&) EIGEN_NO_THROW { \
+        EIGEN_TRY { return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); } \
+        EIGEN_CATCH (...) { return 0; } \
+      }
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign) \
+      void *operator new(std::size_t size) { \
+        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
+      } \
+      void *operator new[](std::size_t size) { \
+        return Eigen::internal::conditional_aligned_malloc<NeedsToAlign>(size); \
+      } \
+      void operator delete(void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete(void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr, std::size_t /* sz */) EIGEN_NO_THROW { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      /* in-place new and delete. since (at least afaik) there is no actual   */ \
+      /* memory allocated we can safely let the default implementation handle */ \
+      /* this particular case. */ \
+      static void *operator new(std::size_t size, void *ptr) { return ::operator new(size,ptr); } \
+      static void *operator new[](std::size_t size, void* ptr) { return ::operator new[](size,ptr); } \
+      void operator delete(void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete(memory,ptr); } \
+      void operator delete[](void * memory, void *ptr) EIGEN_NO_THROW { return ::operator delete[](memory,ptr); } \
+      /* nothrow-new (returns zero instead of std::bad_alloc) */ \
+      EIGEN_MAKE_ALIGNED_OPERATOR_NEW_NOTHROW(NeedsToAlign) \
+      void operator delete(void *ptr, const std::nothrow_t&) EIGEN_NO_THROW { \
+        Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); \
+      } \
+      typedef void eigen_aligned_operator_new_marker_type;
+#else
+  #define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(NeedsToAlign)
+#endif
+
+#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(true)
+#define EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(Scalar,Size) \
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(((Size)!=Eigen::Dynamic) && ((sizeof(Scalar)*(Size))%EIGEN_MAX_ALIGN_BYTES==0)))
+
+/****************************************************************************/
+
+/** \class aligned_allocator
+* \ingroup Core_Module
+*
+* \brief STL compatible allocator to use with with 16 byte aligned types
+*
+* Example:
+* \code
+* // Matrix4f requires 16 bytes alignment:
+* std::map< int, Matrix4f, std::less<int>, 
+*           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
+* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
+* std::map< int, Vector3f > my_map_vec3;
+* \endcode
+*
+* \sa \blank \ref TopicStlContainers.
+*/
+template<class T>
+class aligned_allocator : public std::allocator<T>
+{
+public:
+  typedef std::size_t     size_type;
+  typedef std::ptrdiff_t  difference_type;
+  typedef T*              pointer;
+  typedef const T*        const_pointer;
+  typedef T&              reference;
+  typedef const T&        const_reference;
+  typedef T               value_type;
+
+  template<class U>
+  struct rebind
+  {
+    typedef aligned_allocator<U> other;
+  };
+
+  aligned_allocator() : std::allocator<T>() {}
+
+  aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
+
+  template<class U>
+  aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
+
+  ~aligned_allocator() {}
+
+  pointer allocate(size_type num, const void* /*hint*/ = 0)
+  {
+    internal::check_size_for_overflow<T>(num);
+    return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
+  }
+
+  void deallocate(pointer p, size_type /*num*/)
+  {
+    internal::aligned_free(p);
+  }
+};
+
+//---------- Cache sizes ----------
+
+#if !defined(EIGEN_NO_CPUID)
+#  if EIGEN_COMP_GNUC && EIGEN_ARCH_i386_OR_x86_64
+#    if defined(__PIC__) && EIGEN_ARCH_i386
+       // Case for x86 with PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("xchgl %%ebx, %k1;cpuid; xchgl %%ebx,%k1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "a" (func), "c" (id));
+#    elif defined(__PIC__) && EIGEN_ARCH_x86_64
+       // Case for x64 with PIC. In theory this is only a problem with recent gcc and with medium or large code model, not with the default small code model.
+       // However, we cannot detect which code model is used, and the xchg overhead is negligible anyway.
+#      define EIGEN_CPUID(abcd,func,id) \
+        __asm__ __volatile__ ("xchg{q}\t{%%}rbx, %q1; cpuid; xchg{q}\t{%%}rbx, %q1": "=a" (abcd[0]), "=&r" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id));
+#    else
+       // Case for x86_64 or x86 w/o PIC
+#      define EIGEN_CPUID(abcd,func,id) \
+         __asm__ __volatile__ ("cpuid": "=a" (abcd[0]), "=b" (abcd[1]), "=c" (abcd[2]), "=d" (abcd[3]) : "0" (func), "2" (id) );
+#    endif
+#  elif EIGEN_COMP_MSVC
+#    if (EIGEN_COMP_MSVC > 1500) && EIGEN_ARCH_i386_OR_x86_64
+#      define EIGEN_CPUID(abcd,func,id) __cpuidex((int*)abcd,func,id)
+#    endif
+#  endif
+#endif
+
+namespace internal {
+
+#ifdef EIGEN_CPUID
+
+inline bool cpuid_is_vendor(int abcd[4], const int vendor[3])
+{
+  return abcd[1]==vendor[0] && abcd[3]==vendor[1] && abcd[2]==vendor[2];
+}
+
+inline void queryCacheSizes_intel_direct(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  l1 = l2 = l3 = 0;
+  int cache_id = 0;
+  int cache_type = 0;
+  do {
+    abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+    EIGEN_CPUID(abcd,0x4,cache_id);
+    cache_type  = (abcd[0] & 0x0F) >> 0;
+    if(cache_type==1||cache_type==3) // data or unified cache
+    {
+      int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
+      int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
+      int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
+      int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
+      int sets        = (abcd[2]);                    // C[31:0]
+
+      int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
+
+      switch(cache_level)
+      {
+        case 1: l1 = cache_size; break;
+        case 2: l2 = cache_size; break;
+        case 3: l3 = cache_size; break;
+        default: break;
+      }
+    }
+    cache_id++;
+  } while(cache_type>0 && cache_id<16);
+}
+
+inline void queryCacheSizes_intel_codes(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  l1 = l2 = l3 = 0;
+  EIGEN_CPUID(abcd,0x00000002,0);
+  unsigned char * bytes = reinterpret_cast<unsigned char *>(abcd)+2;
+  bool check_for_p2_core2 = false;
+  for(int i=0; i<14; ++i)
+  {
+    switch(bytes[i])
+    {
+      case 0x0A: l1 = 8; break;   // 0Ah   data L1 cache, 8 KB, 2 ways, 32 byte lines
+      case 0x0C: l1 = 16; break;  // 0Ch   data L1 cache, 16 KB, 4 ways, 32 byte lines
+      case 0x0E: l1 = 24; break;  // 0Eh   data L1 cache, 24 KB, 6 ways, 64 byte lines
+      case 0x10: l1 = 16; break;  // 10h   data L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
+      case 0x15: l1 = 16; break;  // 15h   code L1 cache, 16 KB, 4 ways, 32 byte lines (IA-64)
+      case 0x2C: l1 = 32; break;  // 2Ch   data L1 cache, 32 KB, 8 ways, 64 byte lines
+      case 0x30: l1 = 32; break;  // 30h   code L1 cache, 32 KB, 8 ways, 64 byte lines
+      case 0x60: l1 = 16; break;  // 60h   data L1 cache, 16 KB, 8 ways, 64 byte lines, sectored
+      case 0x66: l1 = 8; break;   // 66h   data L1 cache, 8 KB, 4 ways, 64 byte lines, sectored
+      case 0x67: l1 = 16; break;  // 67h   data L1 cache, 16 KB, 4 ways, 64 byte lines, sectored
+      case 0x68: l1 = 32; break;  // 68h   data L1 cache, 32 KB, 4 ways, 64 byte lines, sectored
+      case 0x1A: l2 = 96; break;   // code and data L2 cache, 96 KB, 6 ways, 64 byte lines (IA-64)
+      case 0x22: l3 = 512; break;   // code and data L3 cache, 512 KB, 4 ways (!), 64 byte lines, dual-sectored
+      case 0x23: l3 = 1024; break;   // code and data L3 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x25: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x29: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x39: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 64 byte lines, sectored
+      case 0x3A: l2 = 192; break;   // code and data L2 cache, 192 KB, 6 ways, 64 byte lines, sectored
+      case 0x3B: l2 = 128; break;   // code and data L2 cache, 128 KB, 2 ways, 64 byte lines, sectored
+      case 0x3C: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 64 byte lines, sectored
+      case 0x3D: l2 = 384; break;   // code and data L2 cache, 384 KB, 6 ways, 64 byte lines, sectored
+      case 0x3E: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines, sectored
+      case 0x40: l2 = 0; break;   // no integrated L2 cache (P6 core) or L3 cache (P4 core)
+      case 0x41: l2 = 128; break;   // code and data L2 cache, 128 KB, 4 ways, 32 byte lines
+      case 0x42: l2 = 256; break;   // code and data L2 cache, 256 KB, 4 ways, 32 byte lines
+      case 0x43: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 32 byte lines
+      case 0x44: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 32 byte lines
+      case 0x45: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 4 ways, 32 byte lines
+      case 0x46: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines
+      case 0x47: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 8 ways, 64 byte lines
+      case 0x48: l2 = 3072; break;   // code and data L2 cache, 3072 KB, 12 ways, 64 byte lines
+      case 0x49: if(l2!=0) l3 = 4096; else {check_for_p2_core2=true; l3 = l2 = 4096;} break;// code and data L3 cache, 4096 KB, 16 ways, 64 byte lines (P4) or L2 for core2
+      case 0x4A: l3 = 6144; break;   // code and data L3 cache, 6144 KB, 12 ways, 64 byte lines
+      case 0x4B: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 16 ways, 64 byte lines
+      case 0x4C: l3 = 12288; break;   // code and data L3 cache, 12288 KB, 12 ways, 64 byte lines
+      case 0x4D: l3 = 16384; break;   // code and data L3 cache, 16384 KB, 16 ways, 64 byte lines
+      case 0x4E: l2 = 6144; break;   // code and data L2 cache, 6144 KB, 24 ways, 64 byte lines
+      case 0x78: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 4 ways, 64 byte lines
+      case 0x79: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7A: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7B: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7C: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines, dual-sectored
+      case 0x7D: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 64 byte lines
+      case 0x7E: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 128 byte lines, sect. (IA-64)
+      case 0x7F: l2 = 512; break;   // code and data L2 cache, 512 KB, 2 ways, 64 byte lines
+      case 0x80: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 64 byte lines
+      case 0x81: l2 = 128; break;   // code and data L2 cache, 128 KB, 8 ways, 32 byte lines
+      case 0x82: l2 = 256; break;   // code and data L2 cache, 256 KB, 8 ways, 32 byte lines
+      case 0x83: l2 = 512; break;   // code and data L2 cache, 512 KB, 8 ways, 32 byte lines
+      case 0x84: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 32 byte lines
+      case 0x85: l2 = 2048; break;   // code and data L2 cache, 2048 KB, 8 ways, 32 byte lines
+      case 0x86: l2 = 512; break;   // code and data L2 cache, 512 KB, 4 ways, 64 byte lines
+      case 0x87: l2 = 1024; break;   // code and data L2 cache, 1024 KB, 8 ways, 64 byte lines
+      case 0x88: l3 = 2048; break;   // code and data L3 cache, 2048 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x89: l3 = 4096; break;   // code and data L3 cache, 4096 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x8A: l3 = 8192; break;   // code and data L3 cache, 8192 KB, 4 ways, 64 byte lines (IA-64)
+      case 0x8D: l3 = 3072; break;   // code and data L3 cache, 3072 KB, 12 ways, 128 byte lines (IA-64)
+
+      default: break;
+    }
+  }
+  if(check_for_p2_core2 && l2 == l3)
+    l3 = 0;
+  l1 *= 1024;
+  l2 *= 1024;
+  l3 *= 1024;
+}
+
+inline void queryCacheSizes_intel(int& l1, int& l2, int& l3, int max_std_funcs)
+{
+  if(max_std_funcs>=4)
+    queryCacheSizes_intel_direct(l1,l2,l3);
+  else
+    queryCacheSizes_intel_codes(l1,l2,l3);
+}
+
+inline void queryCacheSizes_amd(int& l1, int& l2, int& l3)
+{
+  int abcd[4];
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  EIGEN_CPUID(abcd,0x80000005,0);
+  l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
+  abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
+  EIGEN_CPUID(abcd,0x80000006,0);
+  l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
+  l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
+}
+#endif
+
+/** \internal
+ * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
+inline void queryCacheSizes(int& l1, int& l2, int& l3)
+{
+  #ifdef EIGEN_CPUID
+  int abcd[4];
+  const int GenuineIntel[] = {0x756e6547, 0x49656e69, 0x6c65746e};
+  const int AuthenticAMD[] = {0x68747541, 0x69746e65, 0x444d4163};
+  const int AMDisbetter_[] = {0x69444d41, 0x74656273, 0x21726574}; // "AMDisbetter!"
+
+  // identify the CPU vendor
+  EIGEN_CPUID(abcd,0x0,0);
+  int max_std_funcs = abcd[1];
+  if(cpuid_is_vendor(abcd,GenuineIntel))
+    queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
+  else if(cpuid_is_vendor(abcd,AuthenticAMD) || cpuid_is_vendor(abcd,AMDisbetter_))
+    queryCacheSizes_amd(l1,l2,l3);
+  else
+    // by default let's use Intel's API
+    queryCacheSizes_intel(l1,l2,l3,max_std_funcs);
+
+  // here is the list of other vendors:
+//   ||cpuid_is_vendor(abcd,"VIA VIA VIA ")
+//   ||cpuid_is_vendor(abcd,"CyrixInstead")
+//   ||cpuid_is_vendor(abcd,"CentaurHauls")
+//   ||cpuid_is_vendor(abcd,"GenuineTMx86")
+//   ||cpuid_is_vendor(abcd,"TransmetaCPU")
+//   ||cpuid_is_vendor(abcd,"RiseRiseRise")
+//   ||cpuid_is_vendor(abcd,"Geode by NSC")
+//   ||cpuid_is_vendor(abcd,"SiS SiS SiS ")
+//   ||cpuid_is_vendor(abcd,"UMC UMC UMC ")
+//   ||cpuid_is_vendor(abcd,"NexGenDriven")
+  #else
+  l1 = l2 = l3 = -1;
+  #endif
+}
+
+/** \internal
+ * \returns the size in Bytes of the L1 data cache */
+inline int queryL1CacheSize()
+{
+  int l1(-1), l2, l3;
+  queryCacheSizes(l1,l2,l3);
+  return l1;
+}
+
+/** \internal
+ * \returns the size in Bytes of the L2 or L3 cache if this later is present */
+inline int queryTopLevelCacheSize()
+{
+  int l1, l2(-1), l3(-1);
+  queryCacheSizes(l1,l2,l3);
+  return (std::max)(l2,l3);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_MEMORY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h
new file mode 100755
index 0000000..7f63707
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/Meta.h
@@ -0,0 +1,492 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_META_H
+#define EIGEN_META_H
+
+#if defined(__CUDA_ARCH__)
+#include <cfloat>
+#include <math_constants.h>
+#endif
+
+#if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
+#include <cstdint>
+#endif
+
+namespace Eigen {
+
+typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE DenseIndex;
+
+/**
+ * \brief The Index type as used for the API.
+ * \details To change this, \c \#define the preprocessor symbol \c EIGEN_DEFAULT_DENSE_INDEX_TYPE.
+ * \sa \blank \ref TopicPreprocessorDirectives, StorageIndex.
+ */
+
+typedef EIGEN_DEFAULT_DENSE_INDEX_TYPE Index;
+
+namespace internal {
+
+/** \internal
+  * \file Meta.h
+  * This file contains generic metaprogramming classes which are not specifically related to Eigen.
+  * \note In case you wonder, yes we're aware that Boost already provides all these features,
+  * we however don't want to add a dependency to Boost.
+  */
+
+// Only recent versions of ICC complain about using ptrdiff_t to hold pointers,
+// and older versions do not provide *intptr_t types.
+#if EIGEN_COMP_ICC>=1600 &&  __cplusplus >= 201103L
+typedef std::intptr_t  IntPtr;
+typedef std::uintptr_t UIntPtr;
+#else
+typedef std::ptrdiff_t IntPtr;
+typedef std::size_t UIntPtr;
+#endif
+
+struct true_type {  enum { value = 1 }; };
+struct false_type { enum { value = 0 }; };
+
+template<bool Condition, typename Then, typename Else>
+struct conditional { typedef Then type; };
+
+template<typename Then, typename Else>
+struct conditional <false, Then, Else> { typedef Else type; };
+
+template<typename T, typename U> struct is_same { enum { value = 0 }; };
+template<typename T> struct is_same<T,T> { enum { value = 1 }; };
+
+template<typename T> struct remove_reference { typedef T type; };
+template<typename T> struct remove_reference<T&> { typedef T type; };
+
+template<typename T> struct remove_pointer { typedef T type; };
+template<typename T> struct remove_pointer<T*> { typedef T type; };
+template<typename T> struct remove_pointer<T*const> { typedef T type; };
+
+template <class T> struct remove_const { typedef T type; };
+template <class T> struct remove_const<const T> { typedef T type; };
+template <class T> struct remove_const<const T[]> { typedef T type[]; };
+template <class T, unsigned int Size> struct remove_const<const T[Size]> { typedef T type[Size]; };
+
+template<typename T> struct remove_all { typedef T type; };
+template<typename T> struct remove_all<const T>   { typedef typename remove_all<T>::type type; };
+template<typename T> struct remove_all<T const&>  { typedef typename remove_all<T>::type type; };
+template<typename T> struct remove_all<T&>        { typedef typename remove_all<T>::type type; };
+template<typename T> struct remove_all<T const*>  { typedef typename remove_all<T>::type type; };
+template<typename T> struct remove_all<T*>        { typedef typename remove_all<T>::type type; };
+
+template<typename T> struct is_arithmetic      { enum { value = false }; };
+template<> struct is_arithmetic<float>         { enum { value = true }; };
+template<> struct is_arithmetic<double>        { enum { value = true }; };
+template<> struct is_arithmetic<long double>   { enum { value = true }; };
+template<> struct is_arithmetic<bool>          { enum { value = true }; };
+template<> struct is_arithmetic<char>          { enum { value = true }; };
+template<> struct is_arithmetic<signed char>   { enum { value = true }; };
+template<> struct is_arithmetic<unsigned char> { enum { value = true }; };
+template<> struct is_arithmetic<signed short>  { enum { value = true }; };
+template<> struct is_arithmetic<unsigned short>{ enum { value = true }; };
+template<> struct is_arithmetic<signed int>    { enum { value = true }; };
+template<> struct is_arithmetic<unsigned int>  { enum { value = true }; };
+template<> struct is_arithmetic<signed long>   { enum { value = true }; };
+template<> struct is_arithmetic<unsigned long> { enum { value = true }; };
+
+template<typename T> struct is_integral        { enum { value = false }; };
+template<> struct is_integral<bool>            { enum { value = true }; };
+template<> struct is_integral<char>            { enum { value = true }; };
+template<> struct is_integral<signed char>     { enum { value = true }; };
+template<> struct is_integral<unsigned char>   { enum { value = true }; };
+template<> struct is_integral<signed short>    { enum { value = true }; };
+template<> struct is_integral<unsigned short>  { enum { value = true }; };
+template<> struct is_integral<signed int>      { enum { value = true }; };
+template<> struct is_integral<unsigned int>    { enum { value = true }; };
+template<> struct is_integral<signed long>     { enum { value = true }; };
+template<> struct is_integral<unsigned long>   { enum { value = true }; };
+
+template <typename T> struct add_const { typedef const T type; };
+template <typename T> struct add_const<T&> { typedef T& type; };
+
+template <typename T> struct is_const { enum { value = 0 }; };
+template <typename T> struct is_const<T const> { enum { value = 1 }; };
+
+template<typename T> struct add_const_on_value_type            { typedef const T type;  };
+template<typename T> struct add_const_on_value_type<T&>        { typedef T const& type; };
+template<typename T> struct add_const_on_value_type<T*>        { typedef T const* type; };
+template<typename T> struct add_const_on_value_type<T* const>  { typedef T const* const type; };
+template<typename T> struct add_const_on_value_type<T const* const>  { typedef T const* const type; };
+
+
+template<typename From, typename To>
+struct is_convertible_impl
+{
+private:
+  struct any_conversion
+  {
+    template <typename T> any_conversion(const volatile T&);
+    template <typename T> any_conversion(T&);
+  };
+  struct yes {int a[1];};
+  struct no  {int a[2];};
+
+  static yes test(const To&, int);
+  static no  test(any_conversion, ...);
+
+public:
+  static From ms_from;
+#ifdef __INTEL_COMPILER
+  #pragma warning push
+  #pragma warning ( disable : 2259 )
+#endif
+  enum { value = sizeof(test(ms_from, 0))==sizeof(yes) };
+#ifdef __INTEL_COMPILER
+  #pragma warning pop
+#endif
+};
+
+template<typename From, typename To>
+struct is_convertible
+{
+  enum { value = is_convertible_impl<typename remove_all<From>::type,
+                                     typename remove_all<To  >::type>::value };
+};
+
+/** \internal Allows to enable/disable an overload
+  * according to a compile time condition.
+  */
+template<bool Condition, typename T=void> struct enable_if;
+
+template<typename T> struct enable_if<true,T>
+{ typedef T type; };
+
+#if defined(__CUDA_ARCH__)
+#if !defined(__FLT_EPSILON__)
+#define __FLT_EPSILON__ FLT_EPSILON
+#define __DBL_EPSILON__ DBL_EPSILON
+#endif
+
+namespace device {
+
+template<typename T> struct numeric_limits
+{
+  EIGEN_DEVICE_FUNC
+  static T epsilon() { return 0; }
+  static T (max)() { assert(false && "Highest not supported for this type"); }
+  static T (min)() { assert(false && "Lowest not supported for this type"); }
+  static T infinity() { assert(false && "Infinity not supported for this type"); }
+  static T quiet_NaN() { assert(false && "quiet_NaN not supported for this type"); }
+};
+template<> struct numeric_limits<float>
+{
+  EIGEN_DEVICE_FUNC
+  static float epsilon() { return __FLT_EPSILON__; }
+  EIGEN_DEVICE_FUNC
+  static float (max)() { return CUDART_MAX_NORMAL_F; }
+  EIGEN_DEVICE_FUNC
+  static float (min)() { return FLT_MIN; }
+  EIGEN_DEVICE_FUNC
+  static float infinity() { return CUDART_INF_F; }
+  EIGEN_DEVICE_FUNC
+  static float quiet_NaN() { return CUDART_NAN_F; }
+};
+template<> struct numeric_limits<double>
+{
+  EIGEN_DEVICE_FUNC
+  static double epsilon() { return __DBL_EPSILON__; }
+  EIGEN_DEVICE_FUNC
+  static double (max)() { return DBL_MAX; }
+  EIGEN_DEVICE_FUNC
+  static double (min)() { return DBL_MIN; }
+  EIGEN_DEVICE_FUNC
+  static double infinity() { return CUDART_INF; }
+  EIGEN_DEVICE_FUNC
+  static double quiet_NaN() { return CUDART_NAN; }
+};
+template<> struct numeric_limits<int>
+{
+  EIGEN_DEVICE_FUNC
+  static int epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static int (max)() { return INT_MAX; }
+  EIGEN_DEVICE_FUNC
+  static int (min)() { return INT_MIN; }
+};
+template<> struct numeric_limits<unsigned int>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned int epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (max)() { return UINT_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned int (min)() { return 0; }
+};
+template<> struct numeric_limits<long>
+{
+  EIGEN_DEVICE_FUNC
+  static long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static long (max)() { return LONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static long (min)() { return LONG_MIN; }
+};
+template<> struct numeric_limits<unsigned long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (max)() { return ULONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long (min)() { return 0; }
+};
+template<> struct numeric_limits<long long>
+{
+  EIGEN_DEVICE_FUNC
+  static long long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static long long (max)() { return LLONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static long long (min)() { return LLONG_MIN; }
+};
+template<> struct numeric_limits<unsigned long long>
+{
+  EIGEN_DEVICE_FUNC
+  static unsigned long long epsilon() { return 0; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (max)() { return ULLONG_MAX; }
+  EIGEN_DEVICE_FUNC
+  static unsigned long long (min)() { return 0; }
+};
+
+}
+
+#endif
+
+/** \internal
+  * A base class do disable default copy ctor and copy assignement operator.
+  */
+class noncopyable
+{
+  EIGEN_DEVICE_FUNC noncopyable(const noncopyable&);
+  EIGEN_DEVICE_FUNC const noncopyable& operator=(const noncopyable&);
+protected:
+  EIGEN_DEVICE_FUNC noncopyable() {}
+  EIGEN_DEVICE_FUNC ~noncopyable() {}
+};
+
+/** \internal
+  * Convenient struct to get the result type of a unary or binary functor.
+  *
+  * It supports both the current STL mechanism (using the result_type member) as well as
+  * upcoming next STL generation (using a templated result member).
+  * If none of these members is provided, then the type of the first argument is returned. FIXME, that behavior is a pretty bad hack.
+  */
+#if EIGEN_HAS_STD_RESULT_OF
+template<typename T> struct result_of {
+  typedef typename std::result_of<T>::type type1;
+  typedef typename remove_all<type1>::type type;
+};
+#else
+template<typename T> struct result_of { };
+
+struct has_none {int a[1];};
+struct has_std_result_type {int a[2];};
+struct has_tr1_result {int a[3];};
+
+template<typename Func, typename ArgType, int SizeOf=sizeof(has_none)>
+struct unary_result_of_select {typedef typename internal::remove_all<ArgType>::type type;};
+
+template<typename Func, typename ArgType>
+struct unary_result_of_select<Func, ArgType, sizeof(has_std_result_type)> {typedef typename Func::result_type type;};
+
+template<typename Func, typename ArgType>
+struct unary_result_of_select<Func, ArgType, sizeof(has_tr1_result)> {typedef typename Func::template result<Func(ArgType)>::type type;};
+
+template<typename Func, typename ArgType>
+struct result_of<Func(ArgType)> {
+    template<typename T>
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
+    template<typename T>
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType)>::type const * = 0);
+    static has_none               testFunctor(...);
+
+    // note that the following indirection is needed for gcc-3.3
+    enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
+    typedef typename unary_result_of_select<Func, ArgType, FunctorType>::type type;
+};
+
+template<typename Func, typename ArgType0, typename ArgType1, int SizeOf=sizeof(has_none)>
+struct binary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1>
+struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_std_result_type)>
+{typedef typename Func::result_type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1>
+struct binary_result_of_select<Func, ArgType0, ArgType1, sizeof(has_tr1_result)>
+{typedef typename Func::template result<Func(ArgType0,ArgType1)>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1>
+struct result_of<Func(ArgType0,ArgType1)> {
+    template<typename T>
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
+    template<typename T>
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1)>::type const * = 0);
+    static has_none               testFunctor(...);
+
+    // note that the following indirection is needed for gcc-3.3
+    enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
+    typedef typename binary_result_of_select<Func, ArgType0, ArgType1, FunctorType>::type type;
+};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2, int SizeOf=sizeof(has_none)>
+struct ternary_result_of_select {typedef typename internal::remove_all<ArgType0>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_std_result_type)>
+{typedef typename Func::result_type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, sizeof(has_tr1_result)>
+{typedef typename Func::template result<Func(ArgType0,ArgType1,ArgType2)>::type type;};
+
+template<typename Func, typename ArgType0, typename ArgType1, typename ArgType2>
+struct result_of<Func(ArgType0,ArgType1,ArgType2)> {
+    template<typename T>
+    static has_std_result_type    testFunctor(T const *, typename T::result_type const * = 0);
+    template<typename T>
+    static has_tr1_result         testFunctor(T const *, typename T::template result<T(ArgType0,ArgType1,ArgType2)>::type const * = 0);
+    static has_none               testFunctor(...);
+
+    // note that the following indirection is needed for gcc-3.3
+    enum {FunctorType = sizeof(testFunctor(static_cast<Func*>(0)))};
+    typedef typename ternary_result_of_select<Func, ArgType0, ArgType1, ArgType2, FunctorType>::type type;
+};
+#endif
+
+struct meta_yes { char a[1]; };
+struct meta_no  { char a[2]; };
+
+// Check whether T::ReturnType does exist
+template <typename T>
+struct has_ReturnType
+{
+  template <typename C> static meta_yes testFunctor(typename C::ReturnType const *);
+  template <typename C> static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor<T>(0)) == sizeof(meta_yes) };
+};
+
+template<typename T> const T* return_ptr();
+
+template <typename T, typename IndexType=Index>
+struct has_nullary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()())>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+template <typename T, typename IndexType=Index>
+struct has_unary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0)))>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+template <typename T, typename IndexType=Index>
+struct has_binary_operator
+{
+  template <typename C> static meta_yes testFunctor(C const *,typename enable_if<(sizeof(return_ptr<C>()->operator()(IndexType(0),IndexType(0)))>0)>::type * = 0);
+  static meta_no testFunctor(...);
+
+  enum { value = sizeof(testFunctor(static_cast<T*>(0))) == sizeof(meta_yes) };
+};
+
+/** \internal In short, it computes int(sqrt(\a Y)) with \a Y an integer.
+  * Usage example: \code meta_sqrt<1023>::ret \endcode
+  */
+template<int Y,
+         int InfX = 0,
+         int SupX = ((Y==1) ? 1 : Y/2),
+         bool Done = ((SupX-InfX)<=1 ? true : ((SupX*SupX <= Y) && ((SupX+1)*(SupX+1) > Y))) >
+                                // use ?: instead of || just to shut up a stupid gcc 4.3 warning
+class meta_sqrt
+{
+    enum {
+      MidX = (InfX+SupX)/2,
+      TakeInf = MidX*MidX > Y ? 1 : 0,
+      NewInf = int(TakeInf) ? InfX : int(MidX),
+      NewSup = int(TakeInf) ? int(MidX) : SupX
+    };
+  public:
+    enum { ret = meta_sqrt<Y,NewInf,NewSup>::ret };
+};
+
+template<int Y, int InfX, int SupX>
+class meta_sqrt<Y, InfX, SupX, true> { public:  enum { ret = (SupX*SupX <= Y) ? SupX : InfX }; };
+
+
+/** \internal Computes the least common multiple of two positive integer A and B
+  * at compile-time. It implements a naive algorithm testing all multiples of A.
+  * It thus works better if A>=B.
+  */
+template<int A, int B, int K=1, bool Done = ((A*K)%B)==0>
+struct meta_least_common_multiple
+{
+  enum { ret = meta_least_common_multiple<A,B,K+1>::ret };
+};
+template<int A, int B, int K>
+struct meta_least_common_multiple<A,B,K,true>
+{
+  enum { ret = A*K };
+};
+
+/** \internal determines whether the product of two numeric types is allowed and what the return type is */
+template<typename T, typename U> struct scalar_product_traits
+{
+  enum { Defined = 0 };
+};
+
+// FIXME quick workaround around current limitation of result_of
+// template<typename Scalar, typename ArgType0, typename ArgType1>
+// struct result_of<scalar_product_op<Scalar>(ArgType0,ArgType1)> {
+// typedef typename scalar_product_traits<typename remove_all<ArgType0>::type, typename remove_all<ArgType1>::type>::ReturnType type;
+// };
+
+} // end namespace internal
+
+namespace numext {
+  
+#if defined(__CUDA_ARCH__)
+template<typename T> EIGEN_DEVICE_FUNC   void swap(T &a, T &b) { T tmp = b; b = a; a = tmp; }
+#else
+template<typename T> EIGEN_STRONG_INLINE void swap(T &a, T &b) { std::swap(a,b); }
+#endif
+
+#if defined(__CUDA_ARCH__)
+using internal::device::numeric_limits;
+#else
+using std::numeric_limits;
+#endif
+
+// Integer division with rounding up.
+// T is assumed to be an integer type with a>=0, and b>0
+template<typename T>
+T div_ceil(const T &a, const T &b)
+{
+  return (a+b-1) / b;
+}
+
+} // end namespace numext
+
+} // end namespace Eigen
+
+#endif // EIGEN_META_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h
new file mode 100644
index 0000000..1af67cf
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/NonMPL2.h
@@ -0,0 +1,3 @@
+#ifdef EIGEN_MPL2_ONLY
+#error Including non-MPL2 code in EIGEN_MPL2_ONLY mode
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h
new file mode 100644
index 0000000..86b60f5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/ReenableStupidWarnings.h
@@ -0,0 +1,27 @@
+#ifdef EIGEN_WARNINGS_DISABLED
+#undef EIGEN_WARNINGS_DISABLED
+
+#ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+  #ifdef _MSC_VER
+    #pragma warning( pop )
+  #elif defined __INTEL_COMPILER
+    #pragma warning pop
+  #elif defined __clang__
+    #pragma clang diagnostic pop
+  #elif defined __GNUC__ && __GNUC__>=6
+    #pragma GCC diagnostic pop
+  #endif
+
+  #if defined __NVCC__
+//    Don't reenable the diagnostic messages, as it turns out these messages need
+//    to be disabled at the point of the template instantiation (i.e the user code)
+//    otherwise they'll be triggered by nvcc.
+//    #pragma diag_default code_is_unreachable
+//    #pragma diag_default initialization_not_reachable
+//    #pragma diag_default 2651
+//    #pragma diag_default 2653
+  #endif
+
+#endif
+
+#endif // EIGEN_WARNINGS_DISABLED
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h
new file mode 100644
index 0000000..983361a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/StaticAssert.h
@@ -0,0 +1,216 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STATIC_ASSERT_H
+#define EIGEN_STATIC_ASSERT_H
+
+/* Some notes on Eigen's static assertion mechanism:
+ *
+ *  - in EIGEN_STATIC_ASSERT(CONDITION,MSG) the parameter CONDITION must be a compile time boolean
+ *    expression, and MSG an enum listed in struct internal::static_assertion<true>
+ *
+ *  - define EIGEN_NO_STATIC_ASSERT to disable them (and save compilation time)
+ *    in that case, the static assertion is converted to the following runtime assert:
+ *      eigen_assert(CONDITION && "MSG")
+ *
+ *  - currently EIGEN_STATIC_ASSERT can only be used in function scope
+ *
+ */
+
+#ifndef EIGEN_NO_STATIC_ASSERT
+
+  #if EIGEN_MAX_CPP_VER>=11 && (__has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600))
+
+    // if native static_assert is enabled, let's use it
+    #define EIGEN_STATIC_ASSERT(X,MSG) static_assert(X,#MSG);
+
+  #else // not CXX0X
+
+    namespace Eigen {
+
+    namespace internal {
+
+    template<bool condition>
+    struct static_assertion {};
+
+    template<>
+    struct static_assertion<true>
+    {
+      enum {
+        YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX,
+        YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES,
+        YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES,
+        THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE,
+        THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE,
+        THIS_METHOD_IS_ONLY_FOR_OBJECTS_OF_A_SPECIFIC_SIZE,
+        OUT_OF_RANGE_ACCESS,
+        YOU_MADE_A_PROGRAMMING_MISTAKE,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT,
+        EIGEN_INTERNAL_COMPILATION_ERROR_OR_YOU_MADE_A_PROGRAMMING_MISTAKE,
+        YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR,
+        YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR,
+        UNALIGNED_LOAD_AND_STORE_OPERATIONS_UNIMPLEMENTED_ON_ALTIVEC,
+        THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES,
+        FLOATING_POINT_ARGUMENT_PASSED__INTEGER_WAS_EXPECTED,
+        NUMERIC_TYPE_MUST_BE_REAL,
+        COEFFICIENT_WRITE_ACCESS_TO_SELFADJOINT_NOT_SUPPORTED,
+        WRITING_TO_TRIANGULAR_PART_WITH_UNIT_DIAGONAL_IS_NOT_SUPPORTED,
+        THIS_METHOD_IS_ONLY_FOR_FIXED_SIZE,
+        INVALID_MATRIX_PRODUCT,
+        INVALID_VECTOR_VECTOR_PRODUCT__IF_YOU_WANTED_A_DOT_OR_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTIONS,
+        INVALID_MATRIX_PRODUCT__IF_YOU_WANTED_A_COEFF_WISE_PRODUCT_YOU_MUST_USE_THE_EXPLICIT_FUNCTION,
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY,
+        THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES,
+        THIS_METHOD_IS_ONLY_FOR_ROW_MAJOR_MATRICES,
+        INVALID_MATRIX_TEMPLATE_PARAMETERS,
+        INVALID_MATRIXBASE_TEMPLATE_PARAMETERS,
+        BOTH_MATRICES_MUST_HAVE_THE_SAME_STORAGE_ORDER,
+        THIS_METHOD_IS_ONLY_FOR_DIAGONAL_MATRIX,
+        THE_MATRIX_OR_EXPRESSION_THAT_YOU_PASSED_DOES_NOT_HAVE_THE_EXPECTED_TYPE,
+        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_WITH_DIRECT_MEMORY_ACCESS_SUCH_AS_MAP_OR_PLAIN_MATRICES,
+        YOU_ALREADY_SPECIFIED_THIS_STRIDE,
+        INVALID_STORAGE_ORDER_FOR_THIS_VECTOR_EXPRESSION,
+        THE_BRACKET_OPERATOR_IS_ONLY_FOR_VECTORS__USE_THE_PARENTHESIS_OPERATOR_INSTEAD,
+        PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1,
+        THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS,
+        YOU_CANNOT_MIX_ARRAYS_AND_MATRICES,
+        YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION,
+        THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY,
+        YOU_ARE_TRYING_TO_USE_AN_INDEX_BASED_ACCESSOR_ON_AN_EXPRESSION_THAT_DOES_NOT_SUPPORT_THAT,
+        THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS,
+        THIS_METHOD_IS_ONLY_FOR_INNER_OR_LAZY_PRODUCTS,
+        THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL,
+        THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES,
+        YOU_PASSED_A_ROW_VECTOR_BUT_A_COLUMN_VECTOR_WAS_EXPECTED,
+        YOU_PASSED_A_COLUMN_VECTOR_BUT_A_ROW_VECTOR_WAS_EXPECTED,
+        THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE,
+        THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH,
+        OBJECT_ALLOCATED_ON_STACK_IS_TOO_BIG,
+        IMPLICIT_CONVERSION_TO_SCALAR_IS_FOR_INNER_PRODUCT_ONLY,
+        STORAGE_LAYOUT_DOES_NOT_MATCH,
+        EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE,
+        THIS_COEFFICIENT_ACCESSOR_TAKING_ONE_ACCESS_IS_ONLY_FOR_EXPRESSIONS_ALLOWING_LINEAR_ACCESS,
+        MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY,
+        THIS_TYPE_IS_NOT_SUPPORTED,
+        STORAGE_KIND_MUST_MATCH,
+        STORAGE_INDEX_MUST_MATCH,
+        CHOLMOD_SUPPORTS_DOUBLE_PRECISION_ONLY
+      };
+    };
+
+    } // end namespace internal
+
+    } // end namespace Eigen
+
+    // Specialized implementation for MSVC to avoid "conditional
+    // expression is constant" warnings.  This implementation doesn't
+    // appear to work under GCC, hence the multiple implementations.
+    #if EIGEN_COMP_MSVC
+
+      #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \
+        {Eigen::internal::static_assertion<bool(CONDITION)>::MSG;}
+
+    #else
+      // In some cases clang interprets bool(CONDITION) as function declaration
+      #define EIGEN_STATIC_ASSERT(CONDITION,MSG) \
+        if (Eigen::internal::static_assertion<static_cast<bool>(CONDITION)>::MSG) {}
+
+    #endif
+
+  #endif // not CXX0X
+
+#else // EIGEN_NO_STATIC_ASSERT
+
+  #define EIGEN_STATIC_ASSERT(CONDITION,MSG) eigen_assert((CONDITION) && #MSG);
+
+#endif // EIGEN_NO_STATIC_ASSERT
+
+
+// static assertion failing if the type \a TYPE is not a vector type
+#define EIGEN_STATIC_ASSERT_VECTOR_ONLY(TYPE) \
+  EIGEN_STATIC_ASSERT(TYPE::IsVectorAtCompileTime, \
+                      YOU_TRIED_CALLING_A_VECTOR_METHOD_ON_A_MATRIX)
+
+// static assertion failing if the type \a TYPE is not fixed-size
+#define EIGEN_STATIC_ASSERT_FIXED_SIZE(TYPE) \
+  EIGEN_STATIC_ASSERT(TYPE::SizeAtCompileTime!=Eigen::Dynamic, \
+                      YOU_CALLED_A_FIXED_SIZE_METHOD_ON_A_DYNAMIC_SIZE_MATRIX_OR_VECTOR)
+
+// static assertion failing if the type \a TYPE is not dynamic-size
+#define EIGEN_STATIC_ASSERT_DYNAMIC_SIZE(TYPE) \
+  EIGEN_STATIC_ASSERT(TYPE::SizeAtCompileTime==Eigen::Dynamic, \
+                      YOU_CALLED_A_DYNAMIC_SIZE_METHOD_ON_A_FIXED_SIZE_MATRIX_OR_VECTOR)
+
+// static assertion failing if the type \a TYPE is not a vector type of the given size
+#define EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(TYPE, SIZE) \
+  EIGEN_STATIC_ASSERT(TYPE::IsVectorAtCompileTime && TYPE::SizeAtCompileTime==SIZE, \
+                      THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE)
+
+// static assertion failing if the type \a TYPE is not a vector type of the given size
+#define EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(TYPE, ROWS, COLS) \
+  EIGEN_STATIC_ASSERT(TYPE::RowsAtCompileTime==ROWS && TYPE::ColsAtCompileTime==COLS, \
+                      THIS_METHOD_IS_ONLY_FOR_MATRICES_OF_A_SPECIFIC_SIZE)
+
+// static assertion failing if the two vector expression types are not compatible (same fixed-size or dynamic size)
+#define EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(TYPE0,TYPE1) \
+  EIGEN_STATIC_ASSERT( \
+      (int(TYPE0::SizeAtCompileTime)==Eigen::Dynamic \
+    || int(TYPE1::SizeAtCompileTime)==Eigen::Dynamic \
+    || int(TYPE0::SizeAtCompileTime)==int(TYPE1::SizeAtCompileTime)),\
+    YOU_MIXED_VECTORS_OF_DIFFERENT_SIZES)
+
+#define EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1) \
+     ( \
+        (int(Eigen::internal::size_of_xpr_at_compile_time<TYPE0>::ret)==0 && int(Eigen::internal::size_of_xpr_at_compile_time<TYPE1>::ret)==0) \
+    || (\
+          (int(TYPE0::RowsAtCompileTime)==Eigen::Dynamic \
+        || int(TYPE1::RowsAtCompileTime)==Eigen::Dynamic \
+        || int(TYPE0::RowsAtCompileTime)==int(TYPE1::RowsAtCompileTime)) \
+      &&  (int(TYPE0::ColsAtCompileTime)==Eigen::Dynamic \
+        || int(TYPE1::ColsAtCompileTime)==Eigen::Dynamic \
+        || int(TYPE0::ColsAtCompileTime)==int(TYPE1::ColsAtCompileTime))\
+       ) \
+     )
+
+#define EIGEN_STATIC_ASSERT_NON_INTEGER(TYPE) \
+    EIGEN_STATIC_ASSERT(!NumTraits<TYPE>::IsInteger, THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES)
+
+
+// static assertion failing if it is guaranteed at compile-time that the two matrix expression types have different sizes
+#define EIGEN_STATIC_ASSERT_SAME_MATRIX_SIZE(TYPE0,TYPE1) \
+  EIGEN_STATIC_ASSERT( \
+     EIGEN_PREDICATE_SAME_MATRIX_SIZE(TYPE0,TYPE1),\
+    YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES)
+
+#define EIGEN_STATIC_ASSERT_SIZE_1x1(TYPE) \
+      EIGEN_STATIC_ASSERT((TYPE::RowsAtCompileTime == 1 || TYPE::RowsAtCompileTime == Dynamic) && \
+                          (TYPE::ColsAtCompileTime == 1 || TYPE::ColsAtCompileTime == Dynamic), \
+                          THIS_METHOD_IS_ONLY_FOR_1x1_EXPRESSIONS)
+
+#define EIGEN_STATIC_ASSERT_LVALUE(Derived) \
+      EIGEN_STATIC_ASSERT(Eigen::internal::is_lvalue<Derived>::value, \
+                          THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY)
+
+#define EIGEN_STATIC_ASSERT_ARRAYXPR(Derived) \
+      EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived>::XprKind, ArrayXpr>::value), \
+                          THIS_METHOD_IS_ONLY_FOR_ARRAYS_NOT_MATRICES)
+
+#define EIGEN_STATIC_ASSERT_SAME_XPR_KIND(Derived1, Derived2) \
+      EIGEN_STATIC_ASSERT((Eigen::internal::is_same<typename Eigen::internal::traits<Derived1>::XprKind, \
+                                             typename Eigen::internal::traits<Derived2>::XprKind \
+                                            >::value), \
+                          YOU_CANNOT_MIX_ARRAYS_AND_MATRICES)
+
+// Check that a cost value is positive, and that is stay within a reasonable range
+// TODO this check could be enabled for internal debugging only
+#define EIGEN_INTERNAL_CHECK_COST_VALUE(C) \
+      EIGEN_STATIC_ASSERT((C)>=0 && (C)<=HugeCost*HugeCost, EIGEN_INTERNAL_ERROR_PLEASE_FILE_A_BUG_REPORT__INVALID_COST_VALUE);
+
+#endif // EIGEN_STATIC_ASSERT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h
new file mode 100644
index 0000000..ba5bd18
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Core/util/XprHelper.h
@@ -0,0 +1,821 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_XPRHELPER_H
+#define EIGEN_XPRHELPER_H
+
+// just a workaround because GCC seems to not really like empty structs
+// FIXME: gcc 4.3 generates bad code when strict-aliasing is enabled
+// so currently we simply disable this optimization for gcc 4.3
+#if EIGEN_COMP_GNUC && !EIGEN_GNUC_AT(4,3)
+  #define EIGEN_EMPTY_STRUCT_CTOR(X) \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X() {} \
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE X(const X& ) {}
+#else
+  #define EIGEN_EMPTY_STRUCT_CTOR(X)
+#endif
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename IndexDest, typename IndexSrc>
+EIGEN_DEVICE_FUNC
+inline IndexDest convert_index(const IndexSrc& idx) {
+  // for sizeof(IndexDest)>=sizeof(IndexSrc) compilers should be able to optimize this away:
+  eigen_internal_assert(idx <= NumTraits<IndexDest>::highest() && "Index value to big for target type");
+  return IndexDest(idx);
+}
+
+
+// promote_scalar_arg is an helper used in operation between an expression and a scalar, like:
+//    expression * scalar
+// Its role is to determine how the type T of the scalar operand should be promoted given the scalar type ExprScalar of the given expression.
+// The IsSupported template parameter must be provided by the caller as: internal::has_ReturnType<ScalarBinaryOpTraits<ExprScalar,T,op> >::value using the proper order for ExprScalar and T.
+// Then the logic is as follows:
+//  - if the operation is natively supported as defined by IsSupported, then the scalar type is not promoted, and T is returned.
+//  - otherwise, NumTraits<ExprScalar>::Literal is returned if T is implicitly convertible to NumTraits<ExprScalar>::Literal AND that this does not imply a float to integer conversion.
+//  - otherwise, ExprScalar is returned if T is implicitly convertible to ExprScalar AND that this does not imply a float to integer conversion.
+//  - In all other cases, the promoted type is not defined, and the respective operation is thus invalid and not available (SFINAE).
+template<typename ExprScalar,typename T, bool IsSupported>
+struct promote_scalar_arg;
+
+template<typename S,typename T>
+struct promote_scalar_arg<S,T,true>
+{
+  typedef T type;
+};
+
+// Recursively check safe conversion to PromotedType, and then ExprScalar if they are different.
+template<typename ExprScalar,typename T,typename PromotedType,
+  bool ConvertibleToLiteral = internal::is_convertible<T,PromotedType>::value,
+  bool IsSafe = NumTraits<T>::IsInteger || !NumTraits<PromotedType>::IsInteger>
+struct promote_scalar_arg_unsupported;
+
+// Start recursion with NumTraits<ExprScalar>::Literal
+template<typename S,typename T>
+struct promote_scalar_arg<S,T,false> : promote_scalar_arg_unsupported<S,T,typename NumTraits<S>::Literal> {};
+
+// We found a match!
+template<typename S,typename T, typename PromotedType>
+struct promote_scalar_arg_unsupported<S,T,PromotedType,true,true>
+{
+  typedef PromotedType type;
+};
+
+// No match, but no real-to-integer issues, and ExprScalar and current PromotedType are different,
+// so let's try to promote to ExprScalar
+template<typename ExprScalar,typename T, typename PromotedType>
+struct promote_scalar_arg_unsupported<ExprScalar,T,PromotedType,false,true>
+   : promote_scalar_arg_unsupported<ExprScalar,T,ExprScalar>
+{};
+
+// Unsafe real-to-integer, let's stop.
+template<typename S,typename T, typename PromotedType, bool ConvertibleToLiteral>
+struct promote_scalar_arg_unsupported<S,T,PromotedType,ConvertibleToLiteral,false> {};
+
+// T is not even convertible to ExprScalar, let's stop.
+template<typename S,typename T>
+struct promote_scalar_arg_unsupported<S,T,S,false,true> {};
+
+//classes inheriting no_assignment_operator don't generate a default operator=.
+class no_assignment_operator
+{
+  private:
+    no_assignment_operator& operator=(const no_assignment_operator&);
+};
+
+/** \internal return the index type with the largest number of bits */
+template<typename I1, typename I2>
+struct promote_index_type
+{
+  typedef typename conditional<(sizeof(I1)<sizeof(I2)), I2, I1>::type type;
+};
+
+/** \internal If the template parameter Value is Dynamic, this class is just a wrapper around a T variable that
+  * can be accessed using value() and setValue().
+  * Otherwise, this class is an empty structure and value() just returns the template parameter Value.
+  */
+template<typename T, int Value> class variable_if_dynamic
+{
+  public:
+    EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamic)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); }
+    EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {}
+};
+
+template<typename T> class variable_if_dynamic<T, Dynamic>
+{
+    T m_value;
+    EIGEN_DEVICE_FUNC variable_if_dynamic() { eigen_assert(false); }
+  public:
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamic(T value) : m_value(value) {}
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T value() const { return m_value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; }
+};
+
+/** \internal like variable_if_dynamic but for DynamicIndex
+  */
+template<typename T, int Value> class variable_if_dynamicindex
+{
+  public:
+    EIGEN_EMPTY_STRUCT_CTOR(variable_if_dynamicindex)
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T v) { EIGEN_ONLY_USED_FOR_DEBUG(v); eigen_assert(v == T(Value)); }
+    EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE T value() { return T(Value); }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T) {}
+};
+
+template<typename T> class variable_if_dynamicindex<T, DynamicIndex>
+{
+    T m_value;
+    EIGEN_DEVICE_FUNC variable_if_dynamicindex() { eigen_assert(false); }
+  public:
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE explicit variable_if_dynamicindex(T value) : m_value(value) {}
+    EIGEN_DEVICE_FUNC T EIGEN_STRONG_INLINE value() const { return m_value; }
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void setValue(T value) { m_value = value; }
+};
+
+template<typename T> struct functor_traits
+{
+  enum
+  {
+    Cost = 10,
+    PacketAccess = false,
+    IsRepeatable = false
+  };
+};
+
+template<typename T> struct packet_traits;
+
+template<typename T> struct unpacket_traits
+{
+  typedef T type;
+  typedef T half;
+  enum
+  {
+    size = 1,
+    alignment = 1
+  };
+};
+
+template<int Size, typename PacketType,
+         bool Stop = Size==Dynamic || (Size%unpacket_traits<PacketType>::size)==0 || is_same<PacketType,typename unpacket_traits<PacketType>::half>::value>
+struct find_best_packet_helper;
+
+template< int Size, typename PacketType>
+struct find_best_packet_helper<Size,PacketType,true>
+{
+  typedef PacketType type;
+};
+
+template<int Size, typename PacketType>
+struct find_best_packet_helper<Size,PacketType,false>
+{
+  typedef typename find_best_packet_helper<Size,typename unpacket_traits<PacketType>::half>::type type;
+};
+
+template<typename T, int Size>
+struct find_best_packet
+{
+  typedef typename find_best_packet_helper<Size,typename packet_traits<T>::type>::type type;
+};
+
+#if EIGEN_MAX_STATIC_ALIGN_BYTES>0
+template<int ArrayBytes, int AlignmentBytes,
+         bool Match     =  bool((ArrayBytes%AlignmentBytes)==0),
+         bool TryHalf   =  bool(EIGEN_MIN_ALIGN_BYTES<AlignmentBytes) >
+struct compute_default_alignment_helper
+{
+  enum { value = 0 };
+};
+
+template<int ArrayBytes, int AlignmentBytes, bool TryHalf>
+struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, true, TryHalf> // Match
+{
+  enum { value = AlignmentBytes };
+};
+
+template<int ArrayBytes, int AlignmentBytes>
+struct compute_default_alignment_helper<ArrayBytes, AlignmentBytes, false, true> // Try-half
+{
+  // current packet too large, try with an half-packet
+  enum { value = compute_default_alignment_helper<ArrayBytes, AlignmentBytes/2>::value };
+};
+#else
+// If static alignment is disabled, no need to bother.
+// This also avoids a division by zero in "bool Match =  bool((ArrayBytes%AlignmentBytes)==0)"
+template<int ArrayBytes, int AlignmentBytes>
+struct compute_default_alignment_helper
+{
+  enum { value = 0 };
+};
+#endif
+
+template<typename T, int Size> struct compute_default_alignment {
+  enum { value = compute_default_alignment_helper<Size*sizeof(T),EIGEN_MAX_STATIC_ALIGN_BYTES>::value };
+};
+
+template<typename T> struct compute_default_alignment<T,Dynamic> {
+  enum { value = EIGEN_MAX_ALIGN_BYTES };
+};
+
+template<typename _Scalar, int _Rows, int _Cols,
+         int _Options = AutoAlign |
+                          ( (_Rows==1 && _Cols!=1) ? RowMajor
+                          : (_Cols==1 && _Rows!=1) ? ColMajor
+                          : EIGEN_DEFAULT_MATRIX_STORAGE_ORDER_OPTION ),
+         int _MaxRows = _Rows,
+         int _MaxCols = _Cols
+> class make_proper_matrix_type
+{
+    enum {
+      IsColVector = _Cols==1 && _Rows!=1,
+      IsRowVector = _Rows==1 && _Cols!=1,
+      Options = IsColVector ? (_Options | ColMajor) & ~RowMajor
+              : IsRowVector ? (_Options | RowMajor) & ~ColMajor
+              : _Options
+    };
+  public:
+    typedef Matrix<_Scalar, _Rows, _Cols, Options, _MaxRows, _MaxCols> type;
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MaxRows, int MaxCols>
+class compute_matrix_flags
+{
+    enum { row_major_bit = Options&RowMajor ? RowMajorBit : 0 };
+  public:
+    // FIXME currently we still have to handle DirectAccessBit at the expression level to handle DenseCoeffsBase<>
+    // and then propagate this information to the evaluator's flags.
+    // However, I (Gael) think that DirectAccessBit should only matter at the evaluation stage.
+    enum { ret = DirectAccessBit | LvalueBit | NestByRefBit | row_major_bit };
+};
+
+template<int _Rows, int _Cols> struct size_at_compile_time
+{
+  enum { ret = (_Rows==Dynamic || _Cols==Dynamic) ? Dynamic : _Rows * _Cols };
+};
+
+template<typename XprType> struct size_of_xpr_at_compile_time
+{
+  enum { ret = size_at_compile_time<traits<XprType>::RowsAtCompileTime,traits<XprType>::ColsAtCompileTime>::ret };
+};
+
+/* plain_matrix_type : the difference from eval is that plain_matrix_type is always a plain matrix type,
+ * whereas eval is a const reference in the case of a matrix
+ */
+
+template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_matrix_type;
+template<typename T, typename BaseClassType, int Flags> struct plain_matrix_type_dense;
+template<typename T> struct plain_matrix_type<T,Dense>
+{
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, traits<T>::Flags>::type type;
+};
+template<typename T> struct plain_matrix_type<T,DiagonalShape>
+{
+  typedef typename T::PlainObject type;
+};
+
+template<typename T, int Flags> struct plain_matrix_type_dense<T,MatrixXpr,Flags>
+{
+  typedef Matrix<typename traits<T>::Scalar,
+                traits<T>::RowsAtCompileTime,
+                traits<T>::ColsAtCompileTime,
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
+                traits<T>::MaxRowsAtCompileTime,
+                traits<T>::MaxColsAtCompileTime
+          > type;
+};
+
+template<typename T, int Flags> struct plain_matrix_type_dense<T,ArrayXpr,Flags>
+{
+  typedef Array<typename traits<T>::Scalar,
+                traits<T>::RowsAtCompileTime,
+                traits<T>::ColsAtCompileTime,
+                AutoAlign | (Flags&RowMajorBit ? RowMajor : ColMajor),
+                traits<T>::MaxRowsAtCompileTime,
+                traits<T>::MaxColsAtCompileTime
+          > type;
+};
+
+/* eval : the return type of eval(). For matrices, this is just a const reference
+ * in order to avoid a useless copy
+ */
+
+template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct eval;
+
+template<typename T> struct eval<T,Dense>
+{
+  typedef typename plain_matrix_type<T>::type type;
+//   typedef typename T::PlainObject type;
+//   typedef T::Matrix<typename traits<T>::Scalar,
+//                 traits<T>::RowsAtCompileTime,
+//                 traits<T>::ColsAtCompileTime,
+//                 AutoAlign | (traits<T>::Flags&RowMajorBit ? RowMajor : ColMajor),
+//                 traits<T>::MaxRowsAtCompileTime,
+//                 traits<T>::MaxColsAtCompileTime
+//           > type;
+};
+
+template<typename T> struct eval<T,DiagonalShape>
+{
+  typedef typename plain_matrix_type<T>::type type;
+};
+
+// for matrices, no need to evaluate, just use a const reference to avoid a useless copy
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct eval<Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense>
+{
+  typedef const Matrix<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>& type;
+};
+
+template<typename _Scalar, int _Rows, int _Cols, int _Options, int _MaxRows, int _MaxCols>
+struct eval<Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>, Dense>
+{
+  typedef const Array<_Scalar, _Rows, _Cols, _Options, _MaxRows, _MaxCols>& type;
+};
+
+
+/* similar to plain_matrix_type, but using the evaluator's Flags */
+template<typename T, typename StorageKind = typename traits<T>::StorageKind> struct plain_object_eval;
+
+template<typename T>
+struct plain_object_eval<T,Dense>
+{
+  typedef typename plain_matrix_type_dense<T,typename traits<T>::XprKind, evaluator<T>::Flags>::type type;
+};
+
+
+/* plain_matrix_type_column_major : same as plain_matrix_type but guaranteed to be column-major
+ */
+template<typename T> struct plain_matrix_type_column_major
+{
+  enum { Rows = traits<T>::RowsAtCompileTime,
+         Cols = traits<T>::ColsAtCompileTime,
+         MaxRows = traits<T>::MaxRowsAtCompileTime,
+         MaxCols = traits<T>::MaxColsAtCompileTime
+  };
+  typedef Matrix<typename traits<T>::Scalar,
+                Rows,
+                Cols,
+                (MaxRows==1&&MaxCols!=1) ? RowMajor : ColMajor,
+                MaxRows,
+                MaxCols
+          > type;
+};
+
+/* plain_matrix_type_row_major : same as plain_matrix_type but guaranteed to be row-major
+ */
+template<typename T> struct plain_matrix_type_row_major
+{
+  enum { Rows = traits<T>::RowsAtCompileTime,
+         Cols = traits<T>::ColsAtCompileTime,
+         MaxRows = traits<T>::MaxRowsAtCompileTime,
+         MaxCols = traits<T>::MaxColsAtCompileTime
+  };
+  typedef Matrix<typename traits<T>::Scalar,
+                Rows,
+                Cols,
+                (MaxCols==1&&MaxRows!=1) ? RowMajor : ColMajor,
+                MaxRows,
+                MaxCols
+          > type;
+};
+
+/** \internal The reference selector for template expressions. The idea is that we don't
+  * need to use references for expressions since they are light weight proxy
+  * objects which should generate no copying overhead. */
+template <typename T>
+struct ref_selector
+{
+  typedef typename conditional<
+    bool(traits<T>::Flags & NestByRefBit),
+    T const&,
+    const T
+  >::type type;
+  
+  typedef typename conditional<
+    bool(traits<T>::Flags & NestByRefBit),
+    T &,
+    T
+  >::type non_const_type;
+};
+
+/** \internal Adds the const qualifier on the value-type of T2 if and only if T1 is a const type */
+template<typename T1, typename T2>
+struct transfer_constness
+{
+  typedef typename conditional<
+    bool(internal::is_const<T1>::value),
+    typename internal::add_const_on_value_type<T2>::type,
+    T2
+  >::type type;
+};
+
+
+// However, we still need a mechanism to detect whether an expression which is evaluated multiple time
+// has to be evaluated into a temporary.
+// That's the purpose of this new nested_eval helper:
+/** \internal Determines how a given expression should be nested when evaluated multiple times.
+  * For example, when you do a * (b+c), Eigen will determine how the expression b+c should be
+  * evaluated into the bigger product expression. The choice is between nesting the expression b+c as-is, or
+  * evaluating that expression b+c into a temporary variable d, and nest d so that the resulting expression is
+  * a*d. Evaluating can be beneficial for example if every coefficient access in the resulting expression causes
+  * many coefficient accesses in the nested expressions -- as is the case with matrix product for example.
+  *
+  * \tparam T the type of the expression being nested.
+  * \tparam n the number of coefficient accesses in the nested expression for each coefficient access in the bigger expression.
+  * \tparam PlainObject the type of the temporary if needed.
+  */
+template<typename T, int n, typename PlainObject = typename plain_object_eval<T>::type> struct nested_eval
+{
+  enum {
+    ScalarReadCost = NumTraits<typename traits<T>::Scalar>::ReadCost,
+    CoeffReadCost = evaluator<T>::CoeffReadCost,  // NOTE What if an evaluator evaluate itself into a tempory?
+                                                  //      Then CoeffReadCost will be small (e.g., 1) but we still have to evaluate, especially if n>1.
+                                                  //      This situation is already taken care by the EvalBeforeNestingBit flag, which is turned ON
+                                                  //      for all evaluator creating a temporary. This flag is then propagated by the parent evaluators.
+                                                  //      Another solution could be to count the number of temps?
+    NAsInteger = n == Dynamic ? HugeCost : n,
+    CostEval   = (NAsInteger+1) * ScalarReadCost + CoeffReadCost,
+    CostNoEval = NAsInteger * CoeffReadCost,
+    Evaluate = (int(evaluator<T>::Flags) & EvalBeforeNestingBit) || (int(CostEval) < int(CostNoEval))
+  };
+
+  typedef typename conditional<Evaluate, PlainObject, typename ref_selector<T>::type>::type type;
+};
+
+template<typename T>
+EIGEN_DEVICE_FUNC
+inline T* const_cast_ptr(const T* ptr)
+{
+  return const_cast<T*>(ptr);
+}
+
+template<typename Derived, typename XprKind = typename traits<Derived>::XprKind>
+struct dense_xpr_base
+{
+  /* dense_xpr_base should only ever be used on dense expressions, thus falling either into the MatrixXpr or into the ArrayXpr cases */
+};
+
+template<typename Derived>
+struct dense_xpr_base<Derived, MatrixXpr>
+{
+  typedef MatrixBase<Derived> type;
+};
+
+template<typename Derived>
+struct dense_xpr_base<Derived, ArrayXpr>
+{
+  typedef ArrayBase<Derived> type;
+};
+
+template<typename Derived, typename XprKind = typename traits<Derived>::XprKind, typename StorageKind = typename traits<Derived>::StorageKind>
+struct generic_xpr_base;
+
+template<typename Derived, typename XprKind>
+struct generic_xpr_base<Derived, XprKind, Dense>
+{
+  typedef typename dense_xpr_base<Derived,XprKind>::type type;
+};
+
+template<typename XprType, typename CastType> struct cast_return_type
+{
+  typedef typename XprType::Scalar CurrentScalarType;
+  typedef typename remove_all<CastType>::type _CastType;
+  typedef typename _CastType::Scalar NewScalarType;
+  typedef typename conditional<is_same<CurrentScalarType,NewScalarType>::value,
+                              const XprType&,CastType>::type type;
+};
+
+template <typename A, typename B> struct promote_storage_type;
+
+template <typename A> struct promote_storage_type<A,A>
+{
+  typedef A ret;
+};
+template <typename A> struct promote_storage_type<A, const A>
+{
+  typedef A ret;
+};
+template <typename A> struct promote_storage_type<const A, A>
+{
+  typedef A ret;
+};
+
+/** \internal Specify the "storage kind" of applying a coefficient-wise
+  * binary operations between two expressions of kinds A and B respectively.
+  * The template parameter Functor permits to specialize the resulting storage kind wrt to
+  * the functor.
+  * The default rules are as follows:
+  * \code
+  * A      op A      -> A
+  * A      op dense  -> dense
+  * dense  op B      -> dense
+  * sparse op dense  -> sparse
+  * dense  op sparse -> sparse
+  * \endcode
+  */
+template <typename A, typename B, typename Functor> struct cwise_promote_storage_type;
+
+template <typename A, typename Functor>                   struct cwise_promote_storage_type<A,A,Functor>                                      { typedef A      ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Dense,Dense,Functor>                              { typedef Dense  ret; };
+template <typename A, typename Functor>                   struct cwise_promote_storage_type<A,Dense,Functor>                                  { typedef Dense  ret; };
+template <typename B, typename Functor>                   struct cwise_promote_storage_type<Dense,B,Functor>                                  { typedef Dense  ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Sparse,Dense,Functor>                             { typedef Sparse ret; };
+template <typename Functor>                               struct cwise_promote_storage_type<Dense,Sparse,Functor>                             { typedef Sparse ret; };
+
+template <typename LhsKind, typename RhsKind, int LhsOrder, int RhsOrder> struct cwise_promote_storage_order {
+  enum { value = LhsOrder };
+};
+
+template <typename LhsKind, int LhsOrder, int RhsOrder>   struct cwise_promote_storage_order<LhsKind,Sparse,LhsOrder,RhsOrder>                { enum { value = RhsOrder }; };
+template <typename RhsKind, int LhsOrder, int RhsOrder>   struct cwise_promote_storage_order<Sparse,RhsKind,LhsOrder,RhsOrder>                { enum { value = LhsOrder }; };
+template <int Order>                                      struct cwise_promote_storage_order<Sparse,Sparse,Order,Order>                       { enum { value = Order }; };
+
+
+/** \internal Specify the "storage kind" of multiplying an expression of kind A with kind B.
+  * The template parameter ProductTag permits to specialize the resulting storage kind wrt to
+  * some compile-time properties of the product: GemmProduct, GemvProduct, OuterProduct, InnerProduct.
+  * The default rules are as follows:
+  * \code
+  *  K * K            -> K
+  *  dense * K        -> dense
+  *  K * dense        -> dense
+  *  diag * K         -> K
+  *  K * diag         -> K
+  *  Perm * K         -> K
+  * K * Perm          -> K
+  * \endcode
+  */
+template <typename A, typename B, int ProductTag> struct product_promote_storage_type;
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  A,                  ProductTag> { typedef A     ret;};
+template <int ProductTag>             struct product_promote_storage_type<Dense,              Dense,              ProductTag> { typedef Dense ret;};
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  Dense,              ProductTag> { typedef Dense ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<Dense,              B,                  ProductTag> { typedef Dense ret; };
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  DiagonalShape,      ProductTag> { typedef A ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<DiagonalShape,      B,                  ProductTag> { typedef B ret; };
+template <int ProductTag>             struct product_promote_storage_type<Dense,              DiagonalShape,      ProductTag> { typedef Dense ret; };
+template <int ProductTag>             struct product_promote_storage_type<DiagonalShape,      Dense,              ProductTag> { typedef Dense ret; };
+
+template <typename A, int ProductTag> struct product_promote_storage_type<A,                  PermutationStorage, ProductTag> { typedef A ret; };
+template <typename B, int ProductTag> struct product_promote_storage_type<PermutationStorage, B,                  ProductTag> { typedef B ret; };
+template <int ProductTag>             struct product_promote_storage_type<Dense,              PermutationStorage, ProductTag> { typedef Dense ret; };
+template <int ProductTag>             struct product_promote_storage_type<PermutationStorage, Dense,              ProductTag> { typedef Dense ret; };
+
+/** \internal gives the plain matrix or array type to store a row/column/diagonal of a matrix type.
+  * \tparam Scalar optional parameter allowing to pass a different scalar type than the one of the MatrixType.
+  */
+template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar>
+struct plain_row_type
+{
+  typedef Matrix<Scalar, 1, ExpressionType::ColsAtCompileTime,
+                 ExpressionType::PlainObject::Options | RowMajor, 1, ExpressionType::MaxColsAtCompileTime> MatrixRowType;
+  typedef Array<Scalar, 1, ExpressionType::ColsAtCompileTime,
+                 ExpressionType::PlainObject::Options | RowMajor, 1, ExpressionType::MaxColsAtCompileTime> ArrayRowType;
+
+  typedef typename conditional<
+    is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value,
+    MatrixRowType,
+    ArrayRowType 
+  >::type type;
+};
+
+template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar>
+struct plain_col_type
+{
+  typedef Matrix<Scalar, ExpressionType::RowsAtCompileTime, 1,
+                 ExpressionType::PlainObject::Options & ~RowMajor, ExpressionType::MaxRowsAtCompileTime, 1> MatrixColType;
+  typedef Array<Scalar, ExpressionType::RowsAtCompileTime, 1,
+                 ExpressionType::PlainObject::Options & ~RowMajor, ExpressionType::MaxRowsAtCompileTime, 1> ArrayColType;
+
+  typedef typename conditional<
+    is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value,
+    MatrixColType,
+    ArrayColType 
+  >::type type;
+};
+
+template<typename ExpressionType, typename Scalar = typename ExpressionType::Scalar>
+struct plain_diag_type
+{
+  enum { diag_size = EIGEN_SIZE_MIN_PREFER_DYNAMIC(ExpressionType::RowsAtCompileTime, ExpressionType::ColsAtCompileTime),
+         max_diag_size = EIGEN_SIZE_MIN_PREFER_FIXED(ExpressionType::MaxRowsAtCompileTime, ExpressionType::MaxColsAtCompileTime)
+  };
+  typedef Matrix<Scalar, diag_size, 1, ExpressionType::PlainObject::Options & ~RowMajor, max_diag_size, 1> MatrixDiagType;
+  typedef Array<Scalar, diag_size, 1, ExpressionType::PlainObject::Options & ~RowMajor, max_diag_size, 1> ArrayDiagType;
+
+  typedef typename conditional<
+    is_same< typename traits<ExpressionType>::XprKind, MatrixXpr >::value,
+    MatrixDiagType,
+    ArrayDiagType 
+  >::type type;
+};
+
+template<typename Expr,typename Scalar = typename Expr::Scalar>
+struct plain_constant_type
+{
+  enum { Options = (traits<Expr>::Flags&RowMajorBit)?RowMajor:0 };
+
+  typedef Array<Scalar,  traits<Expr>::RowsAtCompileTime,   traits<Expr>::ColsAtCompileTime,
+                Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> array_type;
+
+  typedef Matrix<Scalar,  traits<Expr>::RowsAtCompileTime,   traits<Expr>::ColsAtCompileTime,
+                 Options, traits<Expr>::MaxRowsAtCompileTime,traits<Expr>::MaxColsAtCompileTime> matrix_type;
+
+  typedef CwiseNullaryOp<scalar_constant_op<Scalar>, const typename conditional<is_same< typename traits<Expr>::XprKind, MatrixXpr >::value, matrix_type, array_type>::type > type;
+};
+
+template<typename ExpressionType>
+struct is_lvalue
+{
+  enum { value = (!bool(is_const<ExpressionType>::value)) &&
+                 bool(traits<ExpressionType>::Flags & LvalueBit) };
+};
+
+template<typename T> struct is_diagonal
+{ enum { ret = false }; };
+
+template<typename T> struct is_diagonal<DiagonalBase<T> >
+{ enum { ret = true }; };
+
+template<typename T> struct is_diagonal<DiagonalWrapper<T> >
+{ enum { ret = true }; };
+
+template<typename T, int S> struct is_diagonal<DiagonalMatrix<T,S> >
+{ enum { ret = true }; };
+
+template<typename S1, typename S2> struct glue_shapes;
+template<> struct glue_shapes<DenseShape,TriangularShape> { typedef TriangularShape type;  };
+
+template<typename T1, typename T2>
+bool is_same_dense(const T1 &mat1, const T2 &mat2, typename enable_if<has_direct_access<T1>::ret&&has_direct_access<T2>::ret, T1>::type * = 0)
+{
+  return (mat1.data()==mat2.data()) && (mat1.innerStride()==mat2.innerStride()) && (mat1.outerStride()==mat2.outerStride());
+}
+
+template<typename T1, typename T2>
+bool is_same_dense(const T1 &, const T2 &, typename enable_if<!(has_direct_access<T1>::ret&&has_direct_access<T2>::ret), T1>::type * = 0)
+{
+  return false;
+}
+
+// Internal helper defining the cost of a scalar division for the type T.
+// The default heuristic can be specialized for each scalar type and architecture.
+template<typename T,bool Vectorized=false,typename EnaleIf = void>
+struct scalar_div_cost {
+  enum { value = 8*NumTraits<T>::MulCost };
+};
+
+template<typename T,bool Vectorized>
+struct scalar_div_cost<std::complex<T>, Vectorized> {
+  enum { value = 2*scalar_div_cost<T>::value
+               + 6*NumTraits<T>::MulCost
+               + 3*NumTraits<T>::AddCost
+  };
+};
+
+
+template<bool Vectorized>
+struct scalar_div_cost<signed long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 24 }; };
+template<bool Vectorized>
+struct scalar_div_cost<unsigned long,Vectorized,typename conditional<sizeof(long)==8,void,false_type>::type> { enum { value = 21 }; };
+
+
+#ifdef EIGEN_DEBUG_ASSIGN
+std::string demangle_traversal(int t)
+{
+  if(t==DefaultTraversal) return "DefaultTraversal";
+  if(t==LinearTraversal) return "LinearTraversal";
+  if(t==InnerVectorizedTraversal) return "InnerVectorizedTraversal";
+  if(t==LinearVectorizedTraversal) return "LinearVectorizedTraversal";
+  if(t==SliceVectorizedTraversal) return "SliceVectorizedTraversal";
+  return "?";
+}
+std::string demangle_unrolling(int t)
+{
+  if(t==NoUnrolling) return "NoUnrolling";
+  if(t==InnerUnrolling) return "InnerUnrolling";
+  if(t==CompleteUnrolling) return "CompleteUnrolling";
+  return "?";
+}
+std::string demangle_flags(int f)
+{
+  std::string res;
+  if(f&RowMajorBit)                 res += " | RowMajor";
+  if(f&PacketAccessBit)             res += " | Packet";
+  if(f&LinearAccessBit)             res += " | Linear";
+  if(f&LvalueBit)                   res += " | Lvalue";
+  if(f&DirectAccessBit)             res += " | Direct";
+  if(f&NestByRefBit)                res += " | NestByRef";
+  if(f&NoPreferredStorageOrderBit)  res += " | NoPreferredStorageOrderBit";
+  
+  return res;
+}
+#endif
+
+} // end namespace internal
+
+
+/** \class ScalarBinaryOpTraits
+  * \ingroup Core_Module
+  *
+  * \brief Determines whether the given binary operation of two numeric types is allowed and what the scalar return type is.
+  *
+  * This class permits to control the scalar return type of any binary operation performed on two different scalar types through (partial) template specializations.
+  *
+  * For instance, let \c U1, \c U2 and \c U3 be three user defined scalar types for which most operations between instances of \c U1 and \c U2 returns an \c U3.
+  * You can let %Eigen knows that by defining:
+    \code
+    template<typename BinaryOp>
+    struct ScalarBinaryOpTraits<U1,U2,BinaryOp> { typedef U3 ReturnType;  };
+    template<typename BinaryOp>
+    struct ScalarBinaryOpTraits<U2,U1,BinaryOp> { typedef U3 ReturnType;  };
+    \endcode
+  * You can then explicitly disable some particular operations to get more explicit error messages:
+    \code
+    template<>
+    struct ScalarBinaryOpTraits<U1,U2,internal::scalar_max_op<U1,U2> > {};
+    \endcode
+  * Or customize the return type for individual operation:
+    \code
+    template<>
+    struct ScalarBinaryOpTraits<U1,U2,internal::scalar_sum_op<U1,U2> > { typedef U1 ReturnType; };
+    \endcode
+  *
+  * By default, the following generic combinations are supported:
+  <table class="manual">
+  <tr><th>ScalarA</th><th>ScalarB</th><th>BinaryOp</th><th>ReturnType</th><th>Note</th></tr>
+  <tr            ><td>\c T </td><td>\c T </td><td>\c * </td><td>\c T </td><td></td></tr>
+  <tr class="alt"><td>\c NumTraits<T>::Real </td><td>\c T </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr>
+  <tr            ><td>\c T </td><td>\c NumTraits<T>::Real </td><td>\c * </td><td>\c T </td><td>Only if \c NumTraits<T>::IsComplex </td></tr>
+  </table>
+  *
+  * \sa CwiseBinaryOp
+  */
+template<typename ScalarA, typename ScalarB, typename BinaryOp=internal::scalar_product_op<ScalarA,ScalarB> >
+struct ScalarBinaryOpTraits
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  // for backward compatibility, use the hints given by the (deprecated) internal::scalar_product_traits class.
+  : internal::scalar_product_traits<ScalarA,ScalarB>
+#endif // EIGEN_PARSED_BY_DOXYGEN
+{};
+
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T,T,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+template <typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T, typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, BinaryOp>
+{
+  typedef T ReturnType;
+};
+template <typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<typename NumTraits<typename internal::enable_if<NumTraits<T>::IsComplex,T>::type>::Real, T, BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// For Matrix * Permutation
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<T,void,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// For Permutation * Matrix
+template<typename T, typename BinaryOp>
+struct ScalarBinaryOpTraits<void,T,BinaryOp>
+{
+  typedef T ReturnType;
+};
+
+// for Permutation*Permutation
+template<typename BinaryOp>
+struct ScalarBinaryOpTraits<void,void,BinaryOp>
+{
+  typedef void ReturnType;
+};
+
+// We require Lhs and Rhs to have "compatible" scalar types.
+// It is tempting to always allow mixing different types but remember that this is often impossible in the vectorized paths.
+// So allowing mixing different types gives very unexpected errors when enabling vectorization, when the user tries to
+// add together a float matrix and a double matrix.
+#define EIGEN_CHECK_BINARY_COMPATIBILIY(BINOP,LHS,RHS) \
+  EIGEN_STATIC_ASSERT((Eigen::internal::has_ReturnType<ScalarBinaryOpTraits<LHS, RHS,BINOP> >::value), \
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+    
+} // end namespace Eigen
+
+#endif // EIGEN_XPRHELPER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexEigenSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexEigenSolver.h
new file mode 100644
index 0000000..dc5fae0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexEigenSolver.h
@@ -0,0 +1,346 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Claire Maurice
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_EIGEN_SOLVER_H
+#define EIGEN_COMPLEX_EIGEN_SOLVER_H
+
+#include "./ComplexSchur.h"
+
+namespace Eigen { 
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class ComplexEigenSolver
+  *
+  * \brief Computes eigenvalues and eigenvectors of general complex matrices
+  *
+  * \tparam _MatrixType the type of the matrix of which we are
+  * computing the eigendecomposition; this is expected to be an
+  * instantiation of the Matrix class template.
+  *
+  * The eigenvalues and eigenvectors of a matrix \f$ A \f$ are scalars
+  * \f$ \lambda \f$ and vectors \f$ v \f$ such that \f$ Av = \lambda v
+  * \f$.  If \f$ D \f$ is a diagonal matrix with the eigenvalues on
+  * the diagonal, and \f$ V \f$ is a matrix with the eigenvectors as
+  * its columns, then \f$ A V = V D \f$. The matrix \f$ V \f$ is
+  * almost always invertible, in which case we have \f$ A = V D V^{-1}
+  * \f$. This is called the eigendecomposition.
+  *
+  * The main function in this class is compute(), which computes the
+  * eigenvalues and eigenvectors of a given function. The
+  * documentation for that function contains an example showing the
+  * main features of the class.
+  *
+  * \sa class EigenSolver, class SelfAdjointEigenSolver
+  */
+template<typename _MatrixType> class ComplexEigenSolver
+{
+  public:
+
+    /** \brief Synonym for the template parameter \p _MatrixType. */
+    typedef _MatrixType MatrixType;
+
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+    /** \brief Scalar type for matrices of type #MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    /** \brief Complex scalar type for #MatrixType.
+      *
+      * This is \c std::complex<Scalar> if #Scalar is real (e.g.,
+      * \c float or \c double) and just \c Scalar if #Scalar is
+      * complex.
+      */
+    typedef std::complex<RealScalar> ComplexScalar;
+
+    /** \brief Type for vector of eigenvalues as returned by eigenvalues().
+      *
+      * This is a column vector with entries of type #ComplexScalar.
+      * The length of the vector is the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options&(~RowMajor), MaxColsAtCompileTime, 1> EigenvalueType;
+
+    /** \brief Type for matrix of eigenvectors as returned by eigenvectors().
+      *
+      * This is a square matrix with entries of type #ComplexScalar.
+      * The size is the same as the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, RowsAtCompileTime, ColsAtCompileTime, Options, MaxRowsAtCompileTime, MaxColsAtCompileTime> EigenvectorType;
+
+    /** \brief Default constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute().
+      */
+    ComplexEigenSolver()
+            : m_eivec(),
+              m_eivalues(),
+              m_schur(),
+              m_isInitialized(false),
+              m_eigenvectorsOk(false),
+              m_matX()
+    {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa ComplexEigenSolver()
+      */
+    explicit ComplexEigenSolver(Index size)
+            : m_eivec(size, size),
+              m_eivalues(size),
+              m_schur(size),
+              m_isInitialized(false),
+              m_eigenvectorsOk(false),
+              m_matX(size, size)
+    {}
+
+    /** \brief Constructor; computes eigendecomposition of given matrix.
+      *
+      * \param[in]  matrix  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are
+      *    computed.
+      *
+      * This constructor calls compute() to compute the eigendecomposition.
+      */
+    template<typename InputType>
+    explicit ComplexEigenSolver(const EigenBase<InputType>& matrix, bool computeEigenvectors = true)
+            : m_eivec(matrix.rows(),matrix.cols()),
+              m_eivalues(matrix.cols()),
+              m_schur(matrix.rows()),
+              m_isInitialized(false),
+              m_eigenvectorsOk(false),
+              m_matX(matrix.rows(),matrix.cols())
+    {
+      compute(matrix.derived(), computeEigenvectors);
+    }
+
+    /** \brief Returns the eigenvectors of given matrix.
+      *
+      * \returns  A const reference to the matrix whose columns are the eigenvectors.
+      *
+      * \pre Either the constructor
+      * ComplexEigenSolver(const MatrixType& matrix, bool) or the member
+      * function compute(const MatrixType& matrix, bool) has been called before
+      * to compute the eigendecomposition of a matrix, and
+      * \p computeEigenvectors was set to true (the default).
+      *
+      * This function returns a matrix whose columns are the eigenvectors. Column
+      * \f$ k \f$ is an eigenvector corresponding to eigenvalue number \f$ k
+      * \f$ as returned by eigenvalues().  The eigenvectors are normalized to
+      * have (Euclidean) norm equal to one. The matrix returned by this
+      * function is the matrix \f$ V \f$ in the eigendecomposition \f$ A = V D
+      * V^{-1} \f$, if it exists.
+      *
+      * Example: \include ComplexEigenSolver_eigenvectors.cpp
+      * Output: \verbinclude ComplexEigenSolver_eigenvectors.out
+      */
+    const EigenvectorType& eigenvectors() const
+    {
+      eigen_assert(m_isInitialized && "ComplexEigenSolver is not initialized.");
+      eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+      return m_eivec;
+    }
+
+    /** \brief Returns the eigenvalues of given matrix.
+      *
+      * \returns A const reference to the column vector containing the eigenvalues.
+      *
+      * \pre Either the constructor
+      * ComplexEigenSolver(const MatrixType& matrix, bool) or the member
+      * function compute(const MatrixType& matrix, bool) has been called before
+      * to compute the eigendecomposition of a matrix.
+      *
+      * This function returns a column vector containing the
+      * eigenvalues. Eigenvalues are repeated according to their
+      * algebraic multiplicity, so there are as many eigenvalues as
+      * rows in the matrix. The eigenvalues are not sorted in any particular
+      * order.
+      *
+      * Example: \include ComplexEigenSolver_eigenvalues.cpp
+      * Output: \verbinclude ComplexEigenSolver_eigenvalues.out
+      */
+    const EigenvalueType& eigenvalues() const
+    {
+      eigen_assert(m_isInitialized && "ComplexEigenSolver is not initialized.");
+      return m_eivalues;
+    }
+
+    /** \brief Computes eigendecomposition of given matrix.
+      *
+      * \param[in]  matrix  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are
+      *    computed.
+      * \returns    Reference to \c *this
+      *
+      * This function computes the eigenvalues of the complex matrix \p matrix.
+      * The eigenvalues() function can be used to retrieve them.  If
+      * \p computeEigenvectors is true, then the eigenvectors are also computed
+      * and can be retrieved by calling eigenvectors().
+      *
+      * The matrix is first reduced to Schur form using the
+      * ComplexSchur class. The Schur decomposition is then used to
+      * compute the eigenvalues and eigenvectors.
+      *
+      * The cost of the computation is dominated by the cost of the
+      * Schur decomposition, which is \f$ O(n^3) \f$ where \f$ n \f$
+      * is the size of the matrix.
+      *
+      * Example: \include ComplexEigenSolver_compute.cpp
+      * Output: \verbinclude ComplexEigenSolver_compute.out
+      */
+    template<typename InputType>
+    ComplexEigenSolver& compute(const EigenBase<InputType>& matrix, bool computeEigenvectors = true);
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "ComplexEigenSolver is not initialized.");
+      return m_schur.info();
+    }
+
+    /** \brief Sets the maximum number of iterations allowed. */
+    ComplexEigenSolver& setMaxIterations(Index maxIters)
+    {
+      m_schur.setMaxIterations(maxIters);
+      return *this;
+    }
+
+    /** \brief Returns the maximum number of iterations. */
+    Index getMaxIterations()
+    {
+      return m_schur.getMaxIterations();
+    }
+
+  protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
+    EigenvectorType m_eivec;
+    EigenvalueType m_eivalues;
+    ComplexSchur<MatrixType> m_schur;
+    bool m_isInitialized;
+    bool m_eigenvectorsOk;
+    EigenvectorType m_matX;
+
+  private:
+    void doComputeEigenvectors(RealScalar matrixnorm);
+    void sortEigenvalues(bool computeEigenvectors);
+};
+
+
+template<typename MatrixType>
+template<typename InputType>
+ComplexEigenSolver<MatrixType>& 
+ComplexEigenSolver<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeEigenvectors)
+{
+  check_template_parameters();
+  
+  // this code is inspired from Jampack
+  eigen_assert(matrix.cols() == matrix.rows());
+
+  // Do a complex Schur decomposition, A = U T U^*
+  // The eigenvalues are on the diagonal of T.
+  m_schur.compute(matrix.derived(), computeEigenvectors);
+
+  if(m_schur.info() == Success)
+  {
+    m_eivalues = m_schur.matrixT().diagonal();
+    if(computeEigenvectors)
+      doComputeEigenvectors(m_schur.matrixT().norm());
+    sortEigenvalues(computeEigenvectors);
+  }
+
+  m_isInitialized = true;
+  m_eigenvectorsOk = computeEigenvectors;
+  return *this;
+}
+
+
+template<typename MatrixType>
+void ComplexEigenSolver<MatrixType>::doComputeEigenvectors(RealScalar matrixnorm)
+{
+  const Index n = m_eivalues.size();
+
+  matrixnorm = numext::maxi(matrixnorm,(std::numeric_limits<RealScalar>::min)());
+
+  // Compute X such that T = X D X^(-1), where D is the diagonal of T.
+  // The matrix X is unit triangular.
+  m_matX = EigenvectorType::Zero(n, n);
+  for(Index k=n-1 ; k>=0 ; k--)
+  {
+    m_matX.coeffRef(k,k) = ComplexScalar(1.0,0.0);
+    // Compute X(i,k) using the (i,k) entry of the equation X T = D X
+    for(Index i=k-1 ; i>=0 ; i--)
+    {
+      m_matX.coeffRef(i,k) = -m_schur.matrixT().coeff(i,k);
+      if(k-i-1>0)
+        m_matX.coeffRef(i,k) -= (m_schur.matrixT().row(i).segment(i+1,k-i-1) * m_matX.col(k).segment(i+1,k-i-1)).value();
+      ComplexScalar z = m_schur.matrixT().coeff(i,i) - m_schur.matrixT().coeff(k,k);
+      if(z==ComplexScalar(0))
+      {
+        // If the i-th and k-th eigenvalue are equal, then z equals 0.
+        // Use a small value instead, to prevent division by zero.
+        numext::real_ref(z) = NumTraits<RealScalar>::epsilon() * matrixnorm;
+      }
+      m_matX.coeffRef(i,k) = m_matX.coeff(i,k) / z;
+    }
+  }
+
+  // Compute V as V = U X; now A = U T U^* = U X D X^(-1) U^* = V D V^(-1)
+  m_eivec.noalias() = m_schur.matrixU() * m_matX;
+  // .. and normalize the eigenvectors
+  for(Index k=0 ; k<n ; k++)
+  {
+    m_eivec.col(k).normalize();
+  }
+}
+
+
+template<typename MatrixType>
+void ComplexEigenSolver<MatrixType>::sortEigenvalues(bool computeEigenvectors)
+{
+  const Index n =  m_eivalues.size();
+  for (Index i=0; i<n; i++)
+  {
+    Index k;
+    m_eivalues.cwiseAbs().tail(n-i).minCoeff(&k);
+    if (k != 0)
+    {
+      k += i;
+      std::swap(m_eivalues[k],m_eivalues[i]);
+      if(computeEigenvectors)
+	m_eivec.col(i).swap(m_eivec.col(k));
+    }
+  }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_EIGEN_SOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur.h
new file mode 100644
index 0000000..7f38919
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur.h
@@ -0,0 +1,459 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Claire Maurice
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLEX_SCHUR_H
+#define EIGEN_COMPLEX_SCHUR_H
+
+#include "./HessenbergDecomposition.h"
+
+namespace Eigen { 
+
+namespace internal {
+template<typename MatrixType, bool IsComplex> struct complex_schur_reduce_to_hessenberg;
+}
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class ComplexSchur
+  *
+  * \brief Performs a complex Schur decomposition of a real or complex square matrix
+  *
+  * \tparam _MatrixType the type of the matrix of which we are
+  * computing the Schur decomposition; this is expected to be an
+  * instantiation of the Matrix class template.
+  *
+  * Given a real or complex square matrix A, this class computes the
+  * Schur decomposition: \f$ A = U T U^*\f$ where U is a unitary
+  * complex matrix, and T is a complex upper triangular matrix.  The
+  * diagonal of the matrix T corresponds to the eigenvalues of the
+  * matrix A.
+  *
+  * Call the function compute() to compute the Schur decomposition of
+  * a given matrix. Alternatively, you can use the 
+  * ComplexSchur(const MatrixType&, bool) constructor which computes
+  * the Schur decomposition at construction time. Once the
+  * decomposition is computed, you can use the matrixU() and matrixT()
+  * functions to retrieve the matrices U and V in the decomposition.
+  *
+  * \note This code is inspired from Jampack
+  *
+  * \sa class RealSchur, class EigenSolver, class ComplexEigenSolver
+  */
+template<typename _MatrixType> class ComplexSchur
+{
+  public:
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+    /** \brief Scalar type for matrices of type \p _MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    /** \brief Complex scalar type for \p _MatrixType. 
+      *
+      * This is \c std::complex<Scalar> if #Scalar is real (e.g.,
+      * \c float or \c double) and just \c Scalar if #Scalar is
+      * complex.
+      */
+    typedef std::complex<RealScalar> ComplexScalar;
+
+    /** \brief Type for the matrices in the Schur decomposition.
+      *
+      * This is a square matrix with entries of type #ComplexScalar. 
+      * The size is the same as the size of \p _MatrixType.
+      */
+    typedef Matrix<ComplexScalar, RowsAtCompileTime, ColsAtCompileTime, Options, MaxRowsAtCompileTime, MaxColsAtCompileTime> ComplexMatrixType;
+
+    /** \brief Default constructor.
+      *
+      * \param [in] size  Positive integer, size of the matrix whose Schur decomposition will be computed.
+      *
+      * The default constructor is useful in cases in which the user
+      * intends to perform decompositions via compute().  The \p size
+      * parameter is only used as a hint. It is not an error to give a
+      * wrong \p size, but it may impair performance.
+      *
+      * \sa compute() for an example.
+      */
+    explicit ComplexSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime)
+      : m_matT(size,size),
+        m_matU(size,size),
+        m_hess(size),
+        m_isInitialized(false),
+        m_matUisUptodate(false),
+        m_maxIters(-1)
+    {}
+
+    /** \brief Constructor; computes Schur decomposition of given matrix. 
+      * 
+      * \param[in]  matrix    Square matrix whose Schur decomposition is to be computed.
+      * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
+      *
+      * This constructor calls compute() to compute the Schur decomposition.
+      *
+      * \sa matrixT() and matrixU() for examples.
+      */
+    template<typename InputType>
+    explicit ComplexSchur(const EigenBase<InputType>& matrix, bool computeU = true)
+      : m_matT(matrix.rows(),matrix.cols()),
+        m_matU(matrix.rows(),matrix.cols()),
+        m_hess(matrix.rows()),
+        m_isInitialized(false),
+        m_matUisUptodate(false),
+        m_maxIters(-1)
+    {
+      compute(matrix.derived(), computeU);
+    }
+
+    /** \brief Returns the unitary matrix in the Schur decomposition. 
+      *
+      * \returns A const reference to the matrix U.
+      *
+      * It is assumed that either the constructor
+      * ComplexSchur(const MatrixType& matrix, bool computeU) or the
+      * member function compute(const MatrixType& matrix, bool computeU)
+      * has been called before to compute the Schur decomposition of a
+      * matrix, and that \p computeU was set to true (the default
+      * value).
+      *
+      * Example: \include ComplexSchur_matrixU.cpp
+      * Output: \verbinclude ComplexSchur_matrixU.out
+      */
+    const ComplexMatrixType& matrixU() const
+    {
+      eigen_assert(m_isInitialized && "ComplexSchur is not initialized.");
+      eigen_assert(m_matUisUptodate && "The matrix U has not been computed during the ComplexSchur decomposition.");
+      return m_matU;
+    }
+
+    /** \brief Returns the triangular matrix in the Schur decomposition. 
+      *
+      * \returns A const reference to the matrix T.
+      *
+      * It is assumed that either the constructor
+      * ComplexSchur(const MatrixType& matrix, bool computeU) or the
+      * member function compute(const MatrixType& matrix, bool computeU)
+      * has been called before to compute the Schur decomposition of a
+      * matrix.
+      *
+      * Note that this function returns a plain square matrix. If you want to reference
+      * only the upper triangular part, use:
+      * \code schur.matrixT().triangularView<Upper>() \endcode 
+      *
+      * Example: \include ComplexSchur_matrixT.cpp
+      * Output: \verbinclude ComplexSchur_matrixT.out
+      */
+    const ComplexMatrixType& matrixT() const
+    {
+      eigen_assert(m_isInitialized && "ComplexSchur is not initialized.");
+      return m_matT;
+    }
+
+    /** \brief Computes Schur decomposition of given matrix. 
+      * 
+      * \param[in]  matrix  Square matrix whose Schur decomposition is to be computed.
+      * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
+
+      * \returns    Reference to \c *this
+      *
+      * The Schur decomposition is computed by first reducing the
+      * matrix to Hessenberg form using the class
+      * HessenbergDecomposition. The Hessenberg matrix is then reduced
+      * to triangular form by performing QR iterations with a single
+      * shift. The cost of computing the Schur decomposition depends
+      * on the number of iterations; as a rough guide, it may be taken
+      * on the number of iterations; as a rough guide, it may be taken
+      * to be \f$25n^3\f$ complex flops, or \f$10n^3\f$ complex flops
+      * if \a computeU is false.
+      *
+      * Example: \include ComplexSchur_compute.cpp
+      * Output: \verbinclude ComplexSchur_compute.out
+      *
+      * \sa compute(const MatrixType&, bool, Index)
+      */
+    template<typename InputType>
+    ComplexSchur& compute(const EigenBase<InputType>& matrix, bool computeU = true);
+    
+    /** \brief Compute Schur decomposition from a given Hessenberg matrix
+     *  \param[in] matrixH Matrix in Hessenberg form H
+     *  \param[in] matrixQ orthogonal matrix Q that transform a matrix A to H : A = Q H Q^T
+     *  \param computeU Computes the matriX U of the Schur vectors
+     * \return Reference to \c *this
+     * 
+     *  This routine assumes that the matrix is already reduced in Hessenberg form matrixH
+     *  using either the class HessenbergDecomposition or another mean. 
+     *  It computes the upper quasi-triangular matrix T of the Schur decomposition of H
+     *  When computeU is true, this routine computes the matrix U such that 
+     *  A = U T U^T =  (QZ) T (QZ)^T = Q H Q^T where A is the initial matrix
+     * 
+     * NOTE Q is referenced if computeU is true; so, if the initial orthogonal matrix
+     * is not available, the user should give an identity matrix (Q.setIdentity())
+     * 
+     * \sa compute(const MatrixType&, bool)
+     */
+    template<typename HessMatrixType, typename OrthMatrixType>
+    ComplexSchur& computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ,  bool computeU=true);
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "ComplexSchur is not initialized.");
+      return m_info;
+    }
+
+    /** \brief Sets the maximum number of iterations allowed. 
+      *
+      * If not specified by the user, the maximum number of iterations is m_maxIterationsPerRow times the size
+      * of the matrix.
+      */
+    ComplexSchur& setMaxIterations(Index maxIters)
+    {
+      m_maxIters = maxIters;
+      return *this;
+    }
+
+    /** \brief Returns the maximum number of iterations. */
+    Index getMaxIterations()
+    {
+      return m_maxIters;
+    }
+
+    /** \brief Maximum number of iterations per row.
+      *
+      * If not otherwise specified, the maximum number of iterations is this number times the size of the
+      * matrix. It is currently set to 30.
+      */
+    static const int m_maxIterationsPerRow = 30;
+
+  protected:
+    ComplexMatrixType m_matT, m_matU;
+    HessenbergDecomposition<MatrixType> m_hess;
+    ComputationInfo m_info;
+    bool m_isInitialized;
+    bool m_matUisUptodate;
+    Index m_maxIters;
+
+  private:  
+    bool subdiagonalEntryIsNeglegible(Index i);
+    ComplexScalar computeShift(Index iu, Index iter);
+    void reduceToTriangularForm(bool computeU);
+    friend struct internal::complex_schur_reduce_to_hessenberg<MatrixType, NumTraits<Scalar>::IsComplex>;
+};
+
+/** If m_matT(i+1,i) is neglegible in floating point arithmetic
+  * compared to m_matT(i,i) and m_matT(j,j), then set it to zero and
+  * return true, else return false. */
+template<typename MatrixType>
+inline bool ComplexSchur<MatrixType>::subdiagonalEntryIsNeglegible(Index i)
+{
+  RealScalar d = numext::norm1(m_matT.coeff(i,i)) + numext::norm1(m_matT.coeff(i+1,i+1));
+  RealScalar sd = numext::norm1(m_matT.coeff(i+1,i));
+  if (internal::isMuchSmallerThan(sd, d, NumTraits<RealScalar>::epsilon()))
+  {
+    m_matT.coeffRef(i+1,i) = ComplexScalar(0);
+    return true;
+  }
+  return false;
+}
+
+
+/** Compute the shift in the current QR iteration. */
+template<typename MatrixType>
+typename ComplexSchur<MatrixType>::ComplexScalar ComplexSchur<MatrixType>::computeShift(Index iu, Index iter)
+{
+  using std::abs;
+  if (iter == 10 || iter == 20) 
+  {
+    // exceptional shift, taken from http://www.netlib.org/eispack/comqr.f
+    return abs(numext::real(m_matT.coeff(iu,iu-1))) + abs(numext::real(m_matT.coeff(iu-1,iu-2)));
+  }
+
+  // compute the shift as one of the eigenvalues of t, the 2x2
+  // diagonal block on the bottom of the active submatrix
+  Matrix<ComplexScalar,2,2> t = m_matT.template block<2,2>(iu-1,iu-1);
+  RealScalar normt = t.cwiseAbs().sum();
+  t /= normt;     // the normalization by sf is to avoid under/overflow
+
+  ComplexScalar b = t.coeff(0,1) * t.coeff(1,0);
+  ComplexScalar c = t.coeff(0,0) - t.coeff(1,1);
+  ComplexScalar disc = sqrt(c*c + RealScalar(4)*b);
+  ComplexScalar det = t.coeff(0,0) * t.coeff(1,1) - b;
+  ComplexScalar trace = t.coeff(0,0) + t.coeff(1,1);
+  ComplexScalar eival1 = (trace + disc) / RealScalar(2);
+  ComplexScalar eival2 = (trace - disc) / RealScalar(2);
+
+  if(numext::norm1(eival1) > numext::norm1(eival2))
+    eival2 = det / eival1;
+  else
+    eival1 = det / eival2;
+
+  // choose the eigenvalue closest to the bottom entry of the diagonal
+  if(numext::norm1(eival1-t.coeff(1,1)) < numext::norm1(eival2-t.coeff(1,1)))
+    return normt * eival1;
+  else
+    return normt * eival2;
+}
+
+
+template<typename MatrixType>
+template<typename InputType>
+ComplexSchur<MatrixType>& ComplexSchur<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeU)
+{
+  m_matUisUptodate = false;
+  eigen_assert(matrix.cols() == matrix.rows());
+
+  if(matrix.cols() == 1)
+  {
+    m_matT = matrix.derived().template cast<ComplexScalar>();
+    if(computeU)  m_matU = ComplexMatrixType::Identity(1,1);
+    m_info = Success;
+    m_isInitialized = true;
+    m_matUisUptodate = computeU;
+    return *this;
+  }
+
+  internal::complex_schur_reduce_to_hessenberg<MatrixType, NumTraits<Scalar>::IsComplex>::run(*this, matrix.derived(), computeU);
+  computeFromHessenberg(m_matT, m_matU, computeU);
+  return *this;
+}
+
+template<typename MatrixType>
+template<typename HessMatrixType, typename OrthMatrixType>
+ComplexSchur<MatrixType>& ComplexSchur<MatrixType>::computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ, bool computeU)
+{
+  m_matT = matrixH;
+  if(computeU)
+    m_matU = matrixQ;
+  reduceToTriangularForm(computeU);
+  return *this;
+}
+namespace internal {
+
+/* Reduce given matrix to Hessenberg form */
+template<typename MatrixType, bool IsComplex>
+struct complex_schur_reduce_to_hessenberg
+{
+  // this is the implementation for the case IsComplex = true
+  static void run(ComplexSchur<MatrixType>& _this, const MatrixType& matrix, bool computeU)
+  {
+    _this.m_hess.compute(matrix);
+    _this.m_matT = _this.m_hess.matrixH();
+    if(computeU)  _this.m_matU = _this.m_hess.matrixQ();
+  }
+};
+
+template<typename MatrixType>
+struct complex_schur_reduce_to_hessenberg<MatrixType, false>
+{
+  static void run(ComplexSchur<MatrixType>& _this, const MatrixType& matrix, bool computeU)
+  {
+    typedef typename ComplexSchur<MatrixType>::ComplexScalar ComplexScalar;
+
+    // Note: m_hess is over RealScalar; m_matT and m_matU is over ComplexScalar
+    _this.m_hess.compute(matrix);
+    _this.m_matT = _this.m_hess.matrixH().template cast<ComplexScalar>();
+    if(computeU)  
+    {
+      // This may cause an allocation which seems to be avoidable
+      MatrixType Q = _this.m_hess.matrixQ(); 
+      _this.m_matU = Q.template cast<ComplexScalar>();
+    }
+  }
+};
+
+} // end namespace internal
+
+// Reduce the Hessenberg matrix m_matT to triangular form by QR iteration.
+template<typename MatrixType>
+void ComplexSchur<MatrixType>::reduceToTriangularForm(bool computeU)
+{  
+  Index maxIters = m_maxIters;
+  if (maxIters == -1)
+    maxIters = m_maxIterationsPerRow * m_matT.rows();
+
+  // The matrix m_matT is divided in three parts. 
+  // Rows 0,...,il-1 are decoupled from the rest because m_matT(il,il-1) is zero. 
+  // Rows il,...,iu is the part we are working on (the active submatrix).
+  // Rows iu+1,...,end are already brought in triangular form.
+  Index iu = m_matT.cols() - 1;
+  Index il;
+  Index iter = 0; // number of iterations we are working on the (iu,iu) element
+  Index totalIter = 0; // number of iterations for whole matrix
+
+  while(true)
+  {
+    // find iu, the bottom row of the active submatrix
+    while(iu > 0)
+    {
+      if(!subdiagonalEntryIsNeglegible(iu-1)) break;
+      iter = 0;
+      --iu;
+    }
+
+    // if iu is zero then we are done; the whole matrix is triangularized
+    if(iu==0) break;
+
+    // if we spent too many iterations, we give up
+    iter++;
+    totalIter++;
+    if(totalIter > maxIters) break;
+
+    // find il, the top row of the active submatrix
+    il = iu-1;
+    while(il > 0 && !subdiagonalEntryIsNeglegible(il-1))
+    {
+      --il;
+    }
+
+    /* perform the QR step using Givens rotations. The first rotation
+       creates a bulge; the (il+2,il) element becomes nonzero. This
+       bulge is chased down to the bottom of the active submatrix. */
+
+    ComplexScalar shift = computeShift(iu, iter);
+    JacobiRotation<ComplexScalar> rot;
+    rot.makeGivens(m_matT.coeff(il,il) - shift, m_matT.coeff(il+1,il));
+    m_matT.rightCols(m_matT.cols()-il).applyOnTheLeft(il, il+1, rot.adjoint());
+    m_matT.topRows((std::min)(il+2,iu)+1).applyOnTheRight(il, il+1, rot);
+    if(computeU) m_matU.applyOnTheRight(il, il+1, rot);
+
+    for(Index i=il+1 ; i<iu ; i++)
+    {
+      rot.makeGivens(m_matT.coeffRef(i,i-1), m_matT.coeffRef(i+1,i-1), &m_matT.coeffRef(i,i-1));
+      m_matT.coeffRef(i+1,i-1) = ComplexScalar(0);
+      m_matT.rightCols(m_matT.cols()-i).applyOnTheLeft(i, i+1, rot.adjoint());
+      m_matT.topRows((std::min)(i+2,iu)+1).applyOnTheRight(i, i+1, rot);
+      if(computeU) m_matU.applyOnTheRight(i, i+1, rot);
+    }
+  }
+
+  if(totalIter <= maxIters)
+    m_info = Success;
+  else
+    m_info = NoConvergence;
+
+  m_isInitialized = true;
+  m_matUisUptodate = computeU;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_SCHUR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur_LAPACKE.h
new file mode 100644
index 0000000..4980a3e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/ComplexSchur_LAPACKE.h
@@ -0,0 +1,91 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Complex Schur needed to complex unsymmetrical eigenvalues/eigenvectors.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_COMPLEX_SCHUR_LAPACKE_H
+#define EIGEN_COMPLEX_SCHUR_LAPACKE_H
+
+namespace Eigen { 
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_SCHUR_COMPLEX(EIGTYPE, LAPACKE_TYPE, LAPACKE_PREFIX, LAPACKE_PREFIX_U, EIGCOLROW, LAPACKE_COLROW) \
+template<> template<typename InputType> inline \
+ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
+ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, bool computeU) \
+{ \
+  typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> MatrixType; \
+  typedef MatrixType::RealScalar RealScalar; \
+  typedef std::complex<RealScalar> ComplexScalar; \
+\
+  eigen_assert(matrix.cols() == matrix.rows()); \
+\
+  m_matUisUptodate = false; \
+  if(matrix.cols() == 1) \
+  { \
+    m_matT = matrix.derived().template cast<ComplexScalar>(); \
+    if(computeU)  m_matU = ComplexMatrixType::Identity(1,1); \
+      m_info = Success; \
+      m_isInitialized = true; \
+      m_matUisUptodate = computeU; \
+      return *this; \
+  } \
+  lapack_int n = internal::convert_index<lapack_int>(matrix.cols()), sdim, info; \
+  lapack_int matrix_order = LAPACKE_COLROW; \
+  char jobvs, sort='N'; \
+  LAPACK_##LAPACKE_PREFIX_U##_SELECT1 select = 0; \
+  jobvs = (computeU) ? 'V' : 'N'; \
+  m_matU.resize(n, n); \
+  lapack_int ldvs  = internal::convert_index<lapack_int>(m_matU.outerStride()); \
+  m_matT = matrix; \
+  lapack_int lda = internal::convert_index<lapack_int>(m_matT.outerStride()); \
+  Matrix<EIGTYPE, Dynamic, Dynamic> w; \
+  w.resize(n, 1);\
+  info = LAPACKE_##LAPACKE_PREFIX##gees( matrix_order, jobvs, sort, select, n, (LAPACKE_TYPE*)m_matT.data(), lda, &sdim, (LAPACKE_TYPE*)w.data(), (LAPACKE_TYPE*)m_matU.data(), ldvs ); \
+  if(info == 0) \
+    m_info = Success; \
+  else \
+    m_info = NoConvergence; \
+\
+  m_isInitialized = true; \
+  m_matUisUptodate = computeU; \
+  return *this; \
+\
+}
+
+EIGEN_LAPACKE_SCHUR_COMPLEX(dcomplex, lapack_complex_double, z, Z, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SCHUR_COMPLEX(scomplex, lapack_complex_float,  c, C, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SCHUR_COMPLEX(dcomplex, lapack_complex_double, z, Z, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_SCHUR_COMPLEX(scomplex, lapack_complex_float,  c, C, RowMajor, LAPACK_ROW_MAJOR)
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPLEX_SCHUR_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/EigenSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/EigenSolver.h
new file mode 100644
index 0000000..f205b18
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/EigenSolver.h
@@ -0,0 +1,622 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EIGENSOLVER_H
+#define EIGEN_EIGENSOLVER_H
+
+#include "./RealSchur.h"
+
+namespace Eigen { 
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class EigenSolver
+  *
+  * \brief Computes eigenvalues and eigenvectors of general matrices
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the
+  * eigendecomposition; this is expected to be an instantiation of the Matrix
+  * class template. Currently, only real matrices are supported.
+  *
+  * The eigenvalues and eigenvectors of a matrix \f$ A \f$ are scalars
+  * \f$ \lambda \f$ and vectors \f$ v \f$ such that \f$ Av = \lambda v \f$.  If
+  * \f$ D \f$ is a diagonal matrix with the eigenvalues on the diagonal, and
+  * \f$ V \f$ is a matrix with the eigenvectors as its columns, then \f$ A V =
+  * V D \f$. The matrix \f$ V \f$ is almost always invertible, in which case we
+  * have \f$ A = V D V^{-1} \f$. This is called the eigendecomposition.
+  *
+  * The eigenvalues and eigenvectors of a matrix may be complex, even when the
+  * matrix is real. However, we can choose real matrices \f$ V \f$ and \f$ D
+  * \f$ satisfying \f$ A V = V D \f$, just like the eigendecomposition, if the
+  * matrix \f$ D \f$ is not required to be diagonal, but if it is allowed to
+  * have blocks of the form
+  * \f[ \begin{bmatrix} u & v \\ -v & u \end{bmatrix} \f]
+  * (where \f$ u \f$ and \f$ v \f$ are real numbers) on the diagonal.  These
+  * blocks correspond to complex eigenvalue pairs \f$ u \pm iv \f$. We call
+  * this variant of the eigendecomposition the pseudo-eigendecomposition.
+  *
+  * Call the function compute() to compute the eigenvalues and eigenvectors of
+  * a given matrix. Alternatively, you can use the 
+  * EigenSolver(const MatrixType&, bool) constructor which computes the
+  * eigenvalues and eigenvectors at construction time. Once the eigenvalue and
+  * eigenvectors are computed, they can be retrieved with the eigenvalues() and
+  * eigenvectors() functions. The pseudoEigenvalueMatrix() and
+  * pseudoEigenvectors() methods allow the construction of the
+  * pseudo-eigendecomposition.
+  *
+  * The documentation for EigenSolver(const MatrixType&, bool) contains an
+  * example of the typical use of this class.
+  *
+  * \note The implementation is adapted from
+  * <a href="http://math.nist.gov/javanumerics/jama/">JAMA</a> (public domain).
+  * Their code is based on EISPACK.
+  *
+  * \sa MatrixBase::eigenvalues(), class ComplexEigenSolver, class SelfAdjointEigenSolver
+  */
+template<typename _MatrixType> class EigenSolver
+{
+  public:
+
+    /** \brief Synonym for the template parameter \p _MatrixType. */
+    typedef _MatrixType MatrixType;
+
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+    /** \brief Scalar type for matrices of type #MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    /** \brief Complex scalar type for #MatrixType. 
+      *
+      * This is \c std::complex<Scalar> if #Scalar is real (e.g.,
+      * \c float or \c double) and just \c Scalar if #Scalar is
+      * complex.
+      */
+    typedef std::complex<RealScalar> ComplexScalar;
+
+    /** \brief Type for vector of eigenvalues as returned by eigenvalues(). 
+      *
+      * This is a column vector with entries of type #ComplexScalar.
+      * The length of the vector is the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType;
+
+    /** \brief Type for matrix of eigenvectors as returned by eigenvectors(). 
+      *
+      * This is a square matrix with entries of type #ComplexScalar. 
+      * The size is the same as the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, RowsAtCompileTime, ColsAtCompileTime, Options, MaxRowsAtCompileTime, MaxColsAtCompileTime> EigenvectorsType;
+
+    /** \brief Default constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via EigenSolver::compute(const MatrixType&, bool).
+      *
+      * \sa compute() for an example.
+      */
+    EigenSolver() : m_eivec(), m_eivalues(), m_isInitialized(false), m_realSchur(), m_matT(), m_tmp() {}
+
+    /** \brief Default constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa EigenSolver()
+      */
+    explicit EigenSolver(Index size)
+      : m_eivec(size, size),
+        m_eivalues(size),
+        m_isInitialized(false),
+        m_eigenvectorsOk(false),
+        m_realSchur(size),
+        m_matT(size, size), 
+        m_tmp(size)
+    {}
+
+    /** \brief Constructor; computes eigendecomposition of given matrix. 
+      * 
+      * \param[in]  matrix  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are
+      *    computed. 
+      *
+      * This constructor calls compute() to compute the eigenvalues
+      * and eigenvectors.
+      *
+      * Example: \include EigenSolver_EigenSolver_MatrixType.cpp
+      * Output: \verbinclude EigenSolver_EigenSolver_MatrixType.out
+      *
+      * \sa compute()
+      */
+    template<typename InputType>
+    explicit EigenSolver(const EigenBase<InputType>& matrix, bool computeEigenvectors = true)
+      : m_eivec(matrix.rows(), matrix.cols()),
+        m_eivalues(matrix.cols()),
+        m_isInitialized(false),
+        m_eigenvectorsOk(false),
+        m_realSchur(matrix.cols()),
+        m_matT(matrix.rows(), matrix.cols()), 
+        m_tmp(matrix.cols())
+    {
+      compute(matrix.derived(), computeEigenvectors);
+    }
+
+    /** \brief Returns the eigenvectors of given matrix. 
+      *
+      * \returns  %Matrix whose columns are the (possibly complex) eigenvectors.
+      *
+      * \pre Either the constructor 
+      * EigenSolver(const MatrixType&,bool) or the member function
+      * compute(const MatrixType&, bool) has been called before, and
+      * \p computeEigenvectors was set to true (the default).
+      *
+      * Column \f$ k \f$ of the returned matrix is an eigenvector corresponding
+      * to eigenvalue number \f$ k \f$ as returned by eigenvalues().  The
+      * eigenvectors are normalized to have (Euclidean) norm equal to one. The
+      * matrix returned by this function is the matrix \f$ V \f$ in the
+      * eigendecomposition \f$ A = V D V^{-1} \f$, if it exists.
+      *
+      * Example: \include EigenSolver_eigenvectors.cpp
+      * Output: \verbinclude EigenSolver_eigenvectors.out
+      *
+      * \sa eigenvalues(), pseudoEigenvectors()
+      */
+    EigenvectorsType eigenvectors() const;
+
+    /** \brief Returns the pseudo-eigenvectors of given matrix. 
+      *
+      * \returns  Const reference to matrix whose columns are the pseudo-eigenvectors.
+      *
+      * \pre Either the constructor 
+      * EigenSolver(const MatrixType&,bool) or the member function
+      * compute(const MatrixType&, bool) has been called before, and
+      * \p computeEigenvectors was set to true (the default).
+      *
+      * The real matrix \f$ V \f$ returned by this function and the
+      * block-diagonal matrix \f$ D \f$ returned by pseudoEigenvalueMatrix()
+      * satisfy \f$ AV = VD \f$.
+      *
+      * Example: \include EigenSolver_pseudoEigenvectors.cpp
+      * Output: \verbinclude EigenSolver_pseudoEigenvectors.out
+      *
+      * \sa pseudoEigenvalueMatrix(), eigenvectors()
+      */
+    const MatrixType& pseudoEigenvectors() const
+    {
+      eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
+      eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+      return m_eivec;
+    }
+
+    /** \brief Returns the block-diagonal matrix in the pseudo-eigendecomposition.
+      *
+      * \returns  A block-diagonal matrix.
+      *
+      * \pre Either the constructor 
+      * EigenSolver(const MatrixType&,bool) or the member function
+      * compute(const MatrixType&, bool) has been called before.
+      *
+      * The matrix \f$ D \f$ returned by this function is real and
+      * block-diagonal. The blocks on the diagonal are either 1-by-1 or 2-by-2
+      * blocks of the form
+      * \f$ \begin{bmatrix} u & v \\ -v & u \end{bmatrix} \f$.
+      * These blocks are not sorted in any particular order.
+      * The matrix \f$ D \f$ and the matrix \f$ V \f$ returned by
+      * pseudoEigenvectors() satisfy \f$ AV = VD \f$.
+      *
+      * \sa pseudoEigenvectors() for an example, eigenvalues()
+      */
+    MatrixType pseudoEigenvalueMatrix() const;
+
+    /** \brief Returns the eigenvalues of given matrix. 
+      *
+      * \returns A const reference to the column vector containing the eigenvalues.
+      *
+      * \pre Either the constructor 
+      * EigenSolver(const MatrixType&,bool) or the member function
+      * compute(const MatrixType&, bool) has been called before.
+      *
+      * The eigenvalues are repeated according to their algebraic multiplicity,
+      * so there are as many eigenvalues as rows in the matrix. The eigenvalues 
+      * are not sorted in any particular order.
+      *
+      * Example: \include EigenSolver_eigenvalues.cpp
+      * Output: \verbinclude EigenSolver_eigenvalues.out
+      *
+      * \sa eigenvectors(), pseudoEigenvalueMatrix(),
+      *     MatrixBase::eigenvalues()
+      */
+    const EigenvalueType& eigenvalues() const
+    {
+      eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
+      return m_eivalues;
+    }
+
+    /** \brief Computes eigendecomposition of given matrix. 
+      * 
+      * \param[in]  matrix  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are
+      *    computed. 
+      * \returns    Reference to \c *this
+      *
+      * This function computes the eigenvalues of the real matrix \p matrix.
+      * The eigenvalues() function can be used to retrieve them.  If 
+      * \p computeEigenvectors is true, then the eigenvectors are also computed
+      * and can be retrieved by calling eigenvectors().
+      *
+      * The matrix is first reduced to real Schur form using the RealSchur
+      * class. The Schur decomposition is then used to compute the eigenvalues
+      * and eigenvectors.
+      *
+      * The cost of the computation is dominated by the cost of the
+      * Schur decomposition, which is very approximately \f$ 25n^3 \f$
+      * (where \f$ n \f$ is the size of the matrix) if \p computeEigenvectors 
+      * is true, and \f$ 10n^3 \f$ if \p computeEigenvectors is false.
+      *
+      * This method reuses of the allocated data in the EigenSolver object.
+      *
+      * Example: \include EigenSolver_compute.cpp
+      * Output: \verbinclude EigenSolver_compute.out
+      */
+    template<typename InputType>
+    EigenSolver& compute(const EigenBase<InputType>& matrix, bool computeEigenvectors = true);
+
+    /** \returns NumericalIssue if the input contains INF or NaN values or overflow occured. Returns Success otherwise. */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
+      return m_info;
+    }
+
+    /** \brief Sets the maximum number of iterations allowed. */
+    EigenSolver& setMaxIterations(Index maxIters)
+    {
+      m_realSchur.setMaxIterations(maxIters);
+      return *this;
+    }
+
+    /** \brief Returns the maximum number of iterations. */
+    Index getMaxIterations()
+    {
+      return m_realSchur.getMaxIterations();
+    }
+
+  private:
+    void doComputeEigenvectors();
+
+  protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+      EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
+    }
+    
+    MatrixType m_eivec;
+    EigenvalueType m_eivalues;
+    bool m_isInitialized;
+    bool m_eigenvectorsOk;
+    ComputationInfo m_info;
+    RealSchur<MatrixType> m_realSchur;
+    MatrixType m_matT;
+
+    typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType;
+    ColumnVectorType m_tmp;
+};
+
+template<typename MatrixType>
+MatrixType EigenSolver<MatrixType>::pseudoEigenvalueMatrix() const
+{
+  eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
+  const RealScalar precision = RealScalar(2)*NumTraits<RealScalar>::epsilon();
+  Index n = m_eivalues.rows();
+  MatrixType matD = MatrixType::Zero(n,n);
+  for (Index i=0; i<n; ++i)
+  {
+    if (internal::isMuchSmallerThan(numext::imag(m_eivalues.coeff(i)), numext::real(m_eivalues.coeff(i)), precision))
+      matD.coeffRef(i,i) = numext::real(m_eivalues.coeff(i));
+    else
+    {
+      matD.template block<2,2>(i,i) <<  numext::real(m_eivalues.coeff(i)), numext::imag(m_eivalues.coeff(i)),
+                                       -numext::imag(m_eivalues.coeff(i)), numext::real(m_eivalues.coeff(i));
+      ++i;
+    }
+  }
+  return matD;
+}
+
+template<typename MatrixType>
+typename EigenSolver<MatrixType>::EigenvectorsType EigenSolver<MatrixType>::eigenvectors() const
+{
+  eigen_assert(m_isInitialized && "EigenSolver is not initialized.");
+  eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+  const RealScalar precision = RealScalar(2)*NumTraits<RealScalar>::epsilon();
+  Index n = m_eivec.cols();
+  EigenvectorsType matV(n,n);
+  for (Index j=0; j<n; ++j)
+  {
+    if (internal::isMuchSmallerThan(numext::imag(m_eivalues.coeff(j)), numext::real(m_eivalues.coeff(j)), precision) || j+1==n)
+    {
+      // we have a real eigen value
+      matV.col(j) = m_eivec.col(j).template cast<ComplexScalar>();
+      matV.col(j).normalize();
+    }
+    else
+    {
+      // we have a pair of complex eigen values
+      for (Index i=0; i<n; ++i)
+      {
+        matV.coeffRef(i,j)   = ComplexScalar(m_eivec.coeff(i,j),  m_eivec.coeff(i,j+1));
+        matV.coeffRef(i,j+1) = ComplexScalar(m_eivec.coeff(i,j), -m_eivec.coeff(i,j+1));
+      }
+      matV.col(j).normalize();
+      matV.col(j+1).normalize();
+      ++j;
+    }
+  }
+  return matV;
+}
+
+template<typename MatrixType>
+template<typename InputType>
+EigenSolver<MatrixType>& 
+EigenSolver<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeEigenvectors)
+{
+  check_template_parameters();
+  
+  using std::sqrt;
+  using std::abs;
+  using numext::isfinite;
+  eigen_assert(matrix.cols() == matrix.rows());
+
+  // Reduce to real Schur form.
+  m_realSchur.compute(matrix.derived(), computeEigenvectors);
+  
+  m_info = m_realSchur.info();
+
+  if (m_info == Success)
+  {
+    m_matT = m_realSchur.matrixT();
+    if (computeEigenvectors)
+      m_eivec = m_realSchur.matrixU();
+  
+    // Compute eigenvalues from matT
+    m_eivalues.resize(matrix.cols());
+    Index i = 0;
+    while (i < matrix.cols()) 
+    {
+      if (i == matrix.cols() - 1 || m_matT.coeff(i+1, i) == Scalar(0)) 
+      {
+        m_eivalues.coeffRef(i) = m_matT.coeff(i, i);
+        if(!(isfinite)(m_eivalues.coeffRef(i)))
+        {
+          m_isInitialized = true;
+          m_eigenvectorsOk = false;
+          m_info = NumericalIssue;
+          return *this;
+        }
+        ++i;
+      }
+      else
+      {
+        Scalar p = Scalar(0.5) * (m_matT.coeff(i, i) - m_matT.coeff(i+1, i+1));
+        Scalar z;
+        // Compute z = sqrt(abs(p * p + m_matT.coeff(i+1, i) * m_matT.coeff(i, i+1)));
+        // without overflow
+        {
+          Scalar t0 = m_matT.coeff(i+1, i);
+          Scalar t1 = m_matT.coeff(i, i+1);
+          Scalar maxval = numext::maxi<Scalar>(abs(p),numext::maxi<Scalar>(abs(t0),abs(t1)));
+          t0 /= maxval;
+          t1 /= maxval;
+          Scalar p0 = p/maxval;
+          z = maxval * sqrt(abs(p0 * p0 + t0 * t1));
+        }
+        
+        m_eivalues.coeffRef(i)   = ComplexScalar(m_matT.coeff(i+1, i+1) + p, z);
+        m_eivalues.coeffRef(i+1) = ComplexScalar(m_matT.coeff(i+1, i+1) + p, -z);
+        if(!((isfinite)(m_eivalues.coeffRef(i)) && (isfinite)(m_eivalues.coeffRef(i+1))))
+        {
+          m_isInitialized = true;
+          m_eigenvectorsOk = false;
+          m_info = NumericalIssue;
+          return *this;
+        }
+        i += 2;
+      }
+    }
+    
+    // Compute eigenvectors.
+    if (computeEigenvectors)
+      doComputeEigenvectors();
+  }
+
+  m_isInitialized = true;
+  m_eigenvectorsOk = computeEigenvectors;
+
+  return *this;
+}
+
+
+template<typename MatrixType>
+void EigenSolver<MatrixType>::doComputeEigenvectors()
+{
+  using std::abs;
+  const Index size = m_eivec.cols();
+  const Scalar eps = NumTraits<Scalar>::epsilon();
+
+  // inefficient! this is already computed in RealSchur
+  Scalar norm(0);
+  for (Index j = 0; j < size; ++j)
+  {
+    norm += m_matT.row(j).segment((std::max)(j-1,Index(0)), size-(std::max)(j-1,Index(0))).cwiseAbs().sum();
+  }
+  
+  // Backsubstitute to find vectors of upper triangular form
+  if (norm == Scalar(0))
+  {
+    return;
+  }
+
+  for (Index n = size-1; n >= 0; n--)
+  {
+    Scalar p = m_eivalues.coeff(n).real();
+    Scalar q = m_eivalues.coeff(n).imag();
+
+    // Scalar vector
+    if (q == Scalar(0))
+    {
+      Scalar lastr(0), lastw(0);
+      Index l = n;
+
+      m_matT.coeffRef(n,n) = Scalar(1);
+      for (Index i = n-1; i >= 0; i--)
+      {
+        Scalar w = m_matT.coeff(i,i) - p;
+        Scalar r = m_matT.row(i).segment(l,n-l+1).dot(m_matT.col(n).segment(l, n-l+1));
+
+        if (m_eivalues.coeff(i).imag() < Scalar(0))
+        {
+          lastw = w;
+          lastr = r;
+        }
+        else
+        {
+          l = i;
+          if (m_eivalues.coeff(i).imag() == Scalar(0))
+          {
+            if (w != Scalar(0))
+              m_matT.coeffRef(i,n) = -r / w;
+            else
+              m_matT.coeffRef(i,n) = -r / (eps * norm);
+          }
+          else // Solve real equations
+          {
+            Scalar x = m_matT.coeff(i,i+1);
+            Scalar y = m_matT.coeff(i+1,i);
+            Scalar denom = (m_eivalues.coeff(i).real() - p) * (m_eivalues.coeff(i).real() - p) + m_eivalues.coeff(i).imag() * m_eivalues.coeff(i).imag();
+            Scalar t = (x * lastr - lastw * r) / denom;
+            m_matT.coeffRef(i,n) = t;
+            if (abs(x) > abs(lastw))
+              m_matT.coeffRef(i+1,n) = (-r - w * t) / x;
+            else
+              m_matT.coeffRef(i+1,n) = (-lastr - y * t) / lastw;
+          }
+
+          // Overflow control
+          Scalar t = abs(m_matT.coeff(i,n));
+          if ((eps * t) * t > Scalar(1))
+            m_matT.col(n).tail(size-i) /= t;
+        }
+      }
+    }
+    else if (q < Scalar(0) && n > 0) // Complex vector
+    {
+      Scalar lastra(0), lastsa(0), lastw(0);
+      Index l = n-1;
+
+      // Last vector component imaginary so matrix is triangular
+      if (abs(m_matT.coeff(n,n-1)) > abs(m_matT.coeff(n-1,n)))
+      {
+        m_matT.coeffRef(n-1,n-1) = q / m_matT.coeff(n,n-1);
+        m_matT.coeffRef(n-1,n) = -(m_matT.coeff(n,n) - p) / m_matT.coeff(n,n-1);
+      }
+      else
+      {
+        ComplexScalar cc = ComplexScalar(Scalar(0),-m_matT.coeff(n-1,n)) / ComplexScalar(m_matT.coeff(n-1,n-1)-p,q);
+        m_matT.coeffRef(n-1,n-1) = numext::real(cc);
+        m_matT.coeffRef(n-1,n) = numext::imag(cc);
+      }
+      m_matT.coeffRef(n,n-1) = Scalar(0);
+      m_matT.coeffRef(n,n) = Scalar(1);
+      for (Index i = n-2; i >= 0; i--)
+      {
+        Scalar ra = m_matT.row(i).segment(l, n-l+1).dot(m_matT.col(n-1).segment(l, n-l+1));
+        Scalar sa = m_matT.row(i).segment(l, n-l+1).dot(m_matT.col(n).segment(l, n-l+1));
+        Scalar w = m_matT.coeff(i,i) - p;
+
+        if (m_eivalues.coeff(i).imag() < Scalar(0))
+        {
+          lastw = w;
+          lastra = ra;
+          lastsa = sa;
+        }
+        else
+        {
+          l = i;
+          if (m_eivalues.coeff(i).imag() == RealScalar(0))
+          {
+            ComplexScalar cc = ComplexScalar(-ra,-sa) / ComplexScalar(w,q);
+            m_matT.coeffRef(i,n-1) = numext::real(cc);
+            m_matT.coeffRef(i,n) = numext::imag(cc);
+          }
+          else
+          {
+            // Solve complex equations
+            Scalar x = m_matT.coeff(i,i+1);
+            Scalar y = m_matT.coeff(i+1,i);
+            Scalar vr = (m_eivalues.coeff(i).real() - p) * (m_eivalues.coeff(i).real() - p) + m_eivalues.coeff(i).imag() * m_eivalues.coeff(i).imag() - q * q;
+            Scalar vi = (m_eivalues.coeff(i).real() - p) * Scalar(2) * q;
+            if ((vr == Scalar(0)) && (vi == Scalar(0)))
+              vr = eps * norm * (abs(w) + abs(q) + abs(x) + abs(y) + abs(lastw));
+
+            ComplexScalar cc = ComplexScalar(x*lastra-lastw*ra+q*sa,x*lastsa-lastw*sa-q*ra) / ComplexScalar(vr,vi);
+            m_matT.coeffRef(i,n-1) = numext::real(cc);
+            m_matT.coeffRef(i,n) = numext::imag(cc);
+            if (abs(x) > (abs(lastw) + abs(q)))
+            {
+              m_matT.coeffRef(i+1,n-1) = (-ra - w * m_matT.coeff(i,n-1) + q * m_matT.coeff(i,n)) / x;
+              m_matT.coeffRef(i+1,n) = (-sa - w * m_matT.coeff(i,n) - q * m_matT.coeff(i,n-1)) / x;
+            }
+            else
+            {
+              cc = ComplexScalar(-lastra-y*m_matT.coeff(i,n-1),-lastsa-y*m_matT.coeff(i,n)) / ComplexScalar(lastw,q);
+              m_matT.coeffRef(i+1,n-1) = numext::real(cc);
+              m_matT.coeffRef(i+1,n) = numext::imag(cc);
+            }
+          }
+
+          // Overflow control
+          Scalar t = numext::maxi<Scalar>(abs(m_matT.coeff(i,n-1)),abs(m_matT.coeff(i,n)));
+          if ((eps * t) * t > Scalar(1))
+            m_matT.block(i, n-1, size-i, 2) /= t;
+
+        }
+      }
+      
+      // We handled a pair of complex conjugate eigenvalues, so need to skip them both
+      n--;
+    }
+    else
+    {
+      eigen_assert(0 && "Internal bug in EigenSolver (INF or NaN has not been detected)"); // this should not happen
+    }
+  }
+
+  // Back transformation to get eigenvectors of original matrix
+  for (Index j = size-1; j >= 0; j--)
+  {
+    m_tmp.noalias() = m_eivec.leftCols(j+1) * m_matT.col(j).segment(0, j+1);
+    m_eivec.col(j) = m_tmp;
+  }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_EIGENSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
new file mode 100644
index 0000000..36a91df
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedEigenSolver.h
@@ -0,0 +1,419 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+// Copyright (C) 2016 Tobias Wood <tobias@spinicist.org.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERALIZEDEIGENSOLVER_H
+#define EIGEN_GENERALIZEDEIGENSOLVER_H
+
+#include "./RealQZ.h"
+
+namespace Eigen { 
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class GeneralizedEigenSolver
+  *
+  * \brief Computes the generalized eigenvalues and eigenvectors of a pair of general matrices
+  *
+  * \tparam _MatrixType the type of the matrices of which we are computing the
+  * eigen-decomposition; this is expected to be an instantiation of the Matrix
+  * class template. Currently, only real matrices are supported.
+  *
+  * The generalized eigenvalues and eigenvectors of a matrix pair \f$ A \f$ and \f$ B \f$ are scalars
+  * \f$ \lambda \f$ and vectors \f$ v \f$ such that \f$ Av = \lambda Bv \f$.  If
+  * \f$ D \f$ is a diagonal matrix with the eigenvalues on the diagonal, and
+  * \f$ V \f$ is a matrix with the eigenvectors as its columns, then \f$ A V =
+  * B V D \f$. The matrix \f$ V \f$ is almost always invertible, in which case we
+  * have \f$ A = B V D V^{-1} \f$. This is called the generalized eigen-decomposition.
+  *
+  * The generalized eigenvalues and eigenvectors of a matrix pair may be complex, even when the
+  * matrices are real. Moreover, the generalized eigenvalue might be infinite if the matrix B is
+  * singular. To workaround this difficulty, the eigenvalues are provided as a pair of complex \f$ \alpha \f$
+  * and real \f$ \beta \f$ such that: \f$ \lambda_i = \alpha_i / \beta_i \f$. If \f$ \beta_i \f$ is (nearly) zero,
+  * then one can consider the well defined left eigenvalue \f$ \mu = \beta_i / \alpha_i\f$ such that:
+  * \f$ \mu_i A v_i = B v_i \f$, or even \f$ \mu_i u_i^T A  = u_i^T B \f$ where \f$ u_i \f$ is
+  * called the left eigenvector.
+  *
+  * Call the function compute() to compute the generalized eigenvalues and eigenvectors of
+  * a given matrix pair. Alternatively, you can use the
+  * GeneralizedEigenSolver(const MatrixType&, const MatrixType&, bool) constructor which computes the
+  * eigenvalues and eigenvectors at construction time. Once the eigenvalue and
+  * eigenvectors are computed, they can be retrieved with the eigenvalues() and
+  * eigenvectors() functions.
+  *
+  * Here is an usage example of this class:
+  * Example: \include GeneralizedEigenSolver.cpp
+  * Output: \verbinclude GeneralizedEigenSolver.out
+  *
+  * \sa MatrixBase::eigenvalues(), class ComplexEigenSolver, class SelfAdjointEigenSolver
+  */
+template<typename _MatrixType> class GeneralizedEigenSolver
+{
+  public:
+
+    /** \brief Synonym for the template parameter \p _MatrixType. */
+    typedef _MatrixType MatrixType;
+
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+    /** \brief Scalar type for matrices of type #MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    /** \brief Complex scalar type for #MatrixType. 
+      *
+      * This is \c std::complex<Scalar> if #Scalar is real (e.g.,
+      * \c float or \c double) and just \c Scalar if #Scalar is
+      * complex.
+      */
+    typedef std::complex<RealScalar> ComplexScalar;
+
+    /** \brief Type for vector of real scalar values eigenvalues as returned by betas().
+      *
+      * This is a column vector with entries of type #Scalar.
+      * The length of the vector is the size of #MatrixType.
+      */
+    typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> VectorType;
+
+    /** \brief Type for vector of complex scalar values eigenvalues as returned by alphas().
+      *
+      * This is a column vector with entries of type #ComplexScalar.
+      * The length of the vector is the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ComplexVectorType;
+
+    /** \brief Expression type for the eigenvalues as returned by eigenvalues().
+      */
+    typedef CwiseBinaryOp<internal::scalar_quotient_op<ComplexScalar,Scalar>,ComplexVectorType,VectorType> EigenvalueType;
+
+    /** \brief Type for matrix of eigenvectors as returned by eigenvectors(). 
+      *
+      * This is a square matrix with entries of type #ComplexScalar. 
+      * The size is the same as the size of #MatrixType.
+      */
+    typedef Matrix<ComplexScalar, RowsAtCompileTime, ColsAtCompileTime, Options, MaxRowsAtCompileTime, MaxColsAtCompileTime> EigenvectorsType;
+
+    /** \brief Default constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via EigenSolver::compute(const MatrixType&, bool).
+      *
+      * \sa compute() for an example.
+      */
+    GeneralizedEigenSolver()
+      : m_eivec(),
+        m_alphas(),
+        m_betas(),
+        m_valuesOkay(false),
+        m_vectorsOkay(false),
+        m_realQZ()
+    {}
+
+    /** \brief Default constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa GeneralizedEigenSolver()
+      */
+    explicit GeneralizedEigenSolver(Index size)
+      : m_eivec(size, size),
+        m_alphas(size),
+        m_betas(size),
+        m_valuesOkay(false),
+        m_vectorsOkay(false),
+        m_realQZ(size),
+        m_tmp(size)
+    {}
+
+    /** \brief Constructor; computes the generalized eigendecomposition of given matrix pair.
+      * 
+      * \param[in]  A  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  B  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are computed.
+      *
+      * This constructor calls compute() to compute the generalized eigenvalues
+      * and eigenvectors.
+      *
+      * \sa compute()
+      */
+    GeneralizedEigenSolver(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true)
+      : m_eivec(A.rows(), A.cols()),
+        m_alphas(A.cols()),
+        m_betas(A.cols()),
+        m_valuesOkay(false),
+        m_vectorsOkay(false),
+        m_realQZ(A.cols()),
+        m_tmp(A.cols())
+    {
+      compute(A, B, computeEigenvectors);
+    }
+
+    /* \brief Returns the computed generalized eigenvectors.
+      *
+      * \returns  %Matrix whose columns are the (possibly complex) right eigenvectors.
+      * i.e. the eigenvectors that solve (A - l*B)x = 0. The ordering matches the eigenvalues.
+      *
+      * \pre Either the constructor 
+      * GeneralizedEigenSolver(const MatrixType&,const MatrixType&, bool) or the member function
+      * compute(const MatrixType&, const MatrixType& bool) has been called before, and
+      * \p computeEigenvectors was set to true (the default).
+      *
+      * \sa eigenvalues()
+      */
+    EigenvectorsType eigenvectors() const {
+      eigen_assert(m_vectorsOkay && "Eigenvectors for GeneralizedEigenSolver were not calculated.");
+      return m_eivec;
+    }
+
+    /** \brief Returns an expression of the computed generalized eigenvalues.
+      *
+      * \returns An expression of the column vector containing the eigenvalues.
+      *
+      * It is a shortcut for \code this->alphas().cwiseQuotient(this->betas()); \endcode
+      * Not that betas might contain zeros. It is therefore not recommended to use this function,
+      * but rather directly deal with the alphas and betas vectors.
+      *
+      * \pre Either the constructor 
+      * GeneralizedEigenSolver(const MatrixType&,const MatrixType&,bool) or the member function
+      * compute(const MatrixType&,const MatrixType&,bool) has been called before.
+      *
+      * The eigenvalues are repeated according to their algebraic multiplicity,
+      * so there are as many eigenvalues as rows in the matrix. The eigenvalues 
+      * are not sorted in any particular order.
+      *
+      * \sa alphas(), betas(), eigenvectors()
+      */
+    EigenvalueType eigenvalues() const
+    {
+      eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized.");
+      return EigenvalueType(m_alphas,m_betas);
+    }
+
+    /** \returns A const reference to the vectors containing the alpha values
+      *
+      * This vector permits to reconstruct the j-th eigenvalues as alphas(i)/betas(j).
+      *
+      * \sa betas(), eigenvalues() */
+    ComplexVectorType alphas() const
+    {
+      eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized.");
+      return m_alphas;
+    }
+
+    /** \returns A const reference to the vectors containing the beta values
+      *
+      * This vector permits to reconstruct the j-th eigenvalues as alphas(i)/betas(j).
+      *
+      * \sa alphas(), eigenvalues() */
+    VectorType betas() const
+    {
+      eigen_assert(m_valuesOkay && "GeneralizedEigenSolver is not initialized.");
+      return m_betas;
+    }
+
+    /** \brief Computes generalized eigendecomposition of given matrix.
+      * 
+      * \param[in]  A  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  B  Square matrix whose eigendecomposition is to be computed.
+      * \param[in]  computeEigenvectors  If true, both the eigenvectors and the
+      *    eigenvalues are computed; if false, only the eigenvalues are
+      *    computed. 
+      * \returns    Reference to \c *this
+      *
+      * This function computes the eigenvalues of the real matrix \p matrix.
+      * The eigenvalues() function can be used to retrieve them.  If 
+      * \p computeEigenvectors is true, then the eigenvectors are also computed
+      * and can be retrieved by calling eigenvectors().
+      *
+      * The matrix is first reduced to real generalized Schur form using the RealQZ
+      * class. The generalized Schur decomposition is then used to compute the eigenvalues
+      * and eigenvectors.
+      *
+      * The cost of the computation is dominated by the cost of the
+      * generalized Schur decomposition.
+      *
+      * This method reuses of the allocated data in the GeneralizedEigenSolver object.
+      */
+    GeneralizedEigenSolver& compute(const MatrixType& A, const MatrixType& B, bool computeEigenvectors = true);
+
+    ComputationInfo info() const
+    {
+      eigen_assert(m_valuesOkay && "EigenSolver is not initialized.");
+      return m_realQZ.info();
+    }
+
+    /** Sets the maximal number of iterations allowed.
+    */
+    GeneralizedEigenSolver& setMaxIterations(Index maxIters)
+    {
+      m_realQZ.setMaxIterations(maxIters);
+      return *this;
+    }
+
+  protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+      EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
+    }
+    
+    EigenvectorsType m_eivec;
+    ComplexVectorType m_alphas;
+    VectorType m_betas;
+    bool m_valuesOkay, m_vectorsOkay;
+    RealQZ<MatrixType> m_realQZ;
+    ComplexVectorType m_tmp;
+};
+
+template<typename MatrixType>
+GeneralizedEigenSolver<MatrixType>&
+GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixType& B, bool computeEigenvectors)
+{
+  check_template_parameters();
+  
+  using std::sqrt;
+  using std::abs;
+  eigen_assert(A.cols() == A.rows() && B.cols() == A.rows() && B.cols() == B.rows());
+  Index size = A.cols();
+  m_valuesOkay = false;
+  m_vectorsOkay = false;
+  // Reduce to generalized real Schur form:
+  // A = Q S Z and B = Q T Z
+  m_realQZ.compute(A, B, computeEigenvectors);
+  if (m_realQZ.info() == Success)
+  {
+    // Resize storage
+    m_alphas.resize(size);
+    m_betas.resize(size);
+    if (computeEigenvectors)
+    {
+      m_eivec.resize(size,size);
+      m_tmp.resize(size);
+    }
+
+    // Aliases:
+    Map<VectorType> v(reinterpret_cast<Scalar*>(m_tmp.data()), size);
+    ComplexVectorType &cv = m_tmp;
+    const MatrixType &mZ = m_realQZ.matrixZ();
+    const MatrixType &mS = m_realQZ.matrixS();
+    const MatrixType &mT = m_realQZ.matrixT();
+
+    Index i = 0;
+    while (i < size)
+    {
+      if (i == size - 1 || mS.coeff(i+1, i) == Scalar(0))
+      {
+        // Real eigenvalue
+        m_alphas.coeffRef(i) = mS.diagonal().coeff(i);
+        m_betas.coeffRef(i)  = mT.diagonal().coeff(i);
+        if (computeEigenvectors)
+        {
+          v.setConstant(Scalar(0.0));
+          v.coeffRef(i) = Scalar(1.0);
+          // For singular eigenvalues do nothing more
+          if(abs(m_betas.coeffRef(i)) >= (std::numeric_limits<RealScalar>::min)())
+          {
+            // Non-singular eigenvalue
+            const Scalar alpha = real(m_alphas.coeffRef(i));
+            const Scalar beta = m_betas.coeffRef(i);
+            for (Index j = i-1; j >= 0; j--)
+            {
+              const Index st = j+1;
+              const Index sz = i-j;
+              if (j > 0 && mS.coeff(j, j-1) != Scalar(0))
+              {
+                // 2x2 block
+                Matrix<Scalar, 2, 1> rhs = (alpha*mT.template block<2,Dynamic>(j-1,st,2,sz) - beta*mS.template block<2,Dynamic>(j-1,st,2,sz)) .lazyProduct( v.segment(st,sz) );
+                Matrix<Scalar, 2, 2> lhs = beta * mS.template block<2,2>(j-1,j-1) - alpha * mT.template block<2,2>(j-1,j-1);
+                v.template segment<2>(j-1) = lhs.partialPivLu().solve(rhs);
+                j--;
+              }
+              else
+              {
+                v.coeffRef(j) = -v.segment(st,sz).transpose().cwiseProduct(beta*mS.block(j,st,1,sz) - alpha*mT.block(j,st,1,sz)).sum() / (beta*mS.coeffRef(j,j) - alpha*mT.coeffRef(j,j));
+              }
+            }
+          }
+          m_eivec.col(i).real().noalias() = mZ.transpose() * v;
+          m_eivec.col(i).real().normalize();
+          m_eivec.col(i).imag().setConstant(0);
+        }
+        ++i;
+      }
+      else
+      {
+        // We need to extract the generalized eigenvalues of the pair of a general 2x2 block S and a positive diagonal 2x2 block T
+        // Then taking beta=T_00*T_11, we can avoid any division, and alpha is the eigenvalues of A = (U^-1 * S * U) * diag(T_11,T_00):
+
+        // T =  [a 0]
+        //      [0 b]
+        RealScalar a = mT.diagonal().coeff(i),
+                   b = mT.diagonal().coeff(i+1);
+        const RealScalar beta = m_betas.coeffRef(i) = m_betas.coeffRef(i+1) = a*b;
+
+        // ^^ NOTE: using diagonal()(i) instead of coeff(i,i) workarounds a MSVC bug.
+        Matrix<RealScalar,2,2> S2 = mS.template block<2,2>(i,i) * Matrix<Scalar,2,1>(b,a).asDiagonal();
+
+        Scalar p = Scalar(0.5) * (S2.coeff(0,0) - S2.coeff(1,1));
+        Scalar z = sqrt(abs(p * p + S2.coeff(1,0) * S2.coeff(0,1)));
+        const ComplexScalar alpha = ComplexScalar(S2.coeff(1,1) + p, (beta > 0) ? z : -z);
+        m_alphas.coeffRef(i)   = conj(alpha);
+        m_alphas.coeffRef(i+1) = alpha;
+
+        if (computeEigenvectors) {
+          // Compute eigenvector in position (i+1) and then position (i) is just the conjugate
+          cv.setZero();
+          cv.coeffRef(i+1) = Scalar(1.0);
+          // here, the "static_cast" workaound expression template issues.
+          cv.coeffRef(i) = -(static_cast<Scalar>(beta*mS.coeffRef(i,i+1)) - alpha*mT.coeffRef(i,i+1))
+                          / (static_cast<Scalar>(beta*mS.coeffRef(i,i))   - alpha*mT.coeffRef(i,i));
+          for (Index j = i-1; j >= 0; j--)
+          {
+            const Index st = j+1;
+            const Index sz = i+1-j;
+            if (j > 0 && mS.coeff(j, j-1) != Scalar(0))
+            {
+              // 2x2 block
+              Matrix<ComplexScalar, 2, 1> rhs = (alpha*mT.template block<2,Dynamic>(j-1,st,2,sz) - beta*mS.template block<2,Dynamic>(j-1,st,2,sz)) .lazyProduct( cv.segment(st,sz) );
+              Matrix<ComplexScalar, 2, 2> lhs = beta * mS.template block<2,2>(j-1,j-1) - alpha * mT.template block<2,2>(j-1,j-1);
+              cv.template segment<2>(j-1) = lhs.partialPivLu().solve(rhs);
+              j--;
+            } else {
+              cv.coeffRef(j) =  cv.segment(st,sz).transpose().cwiseProduct(beta*mS.block(j,st,1,sz) - alpha*mT.block(j,st,1,sz)).sum()
+                              / (alpha*mT.coeffRef(j,j) - static_cast<Scalar>(beta*mS.coeffRef(j,j)));
+            }
+          }
+          m_eivec.col(i+1).noalias() = (mZ.transpose() * cv);
+          m_eivec.col(i+1).normalize();
+          m_eivec.col(i) = m_eivec.col(i+1).conjugate();
+        }
+        i += 2;
+      }
+    }
+
+    m_valuesOkay = true;
+    m_vectorsOkay = computeEigenvectors;
+  }
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERALIZEDEIGENSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h
new file mode 100644
index 0000000..5f6bb82
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/GeneralizedSelfAdjointEigenSolver.h
@@ -0,0 +1,226 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERALIZEDSELFADJOINTEIGENSOLVER_H
+#define EIGEN_GENERALIZEDSELFADJOINTEIGENSOLVER_H
+
+#include "./Tridiagonalization.h"
+
+namespace Eigen { 
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class GeneralizedSelfAdjointEigenSolver
+  *
+  * \brief Computes eigenvalues and eigenvectors of the generalized selfadjoint eigen problem
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the
+  * eigendecomposition; this is expected to be an instantiation of the Matrix
+  * class template.
+  *
+  * This class solves the generalized eigenvalue problem
+  * \f$ Av = \lambda Bv \f$. In this case, the matrix \f$ A \f$ should be
+  * selfadjoint and the matrix \f$ B \f$ should be positive definite.
+  *
+  * Only the \b lower \b triangular \b part of the input matrix is referenced.
+  *
+  * Call the function compute() to compute the eigenvalues and eigenvectors of
+  * a given matrix. Alternatively, you can use the
+  * GeneralizedSelfAdjointEigenSolver(const MatrixType&, const MatrixType&, int)
+  * constructor which computes the eigenvalues and eigenvectors at construction time.
+  * Once the eigenvalue and eigenvectors are computed, they can be retrieved with the eigenvalues()
+  * and eigenvectors() functions.
+  *
+  * The documentation for GeneralizedSelfAdjointEigenSolver(const MatrixType&, const MatrixType&, int)
+  * contains an example of the typical use of this class.
+  *
+  * \sa class SelfAdjointEigenSolver, class EigenSolver, class ComplexEigenSolver
+  */
+template<typename _MatrixType>
+class GeneralizedSelfAdjointEigenSolver : public SelfAdjointEigenSolver<_MatrixType>
+{
+    typedef SelfAdjointEigenSolver<_MatrixType> Base;
+  public:
+
+    typedef _MatrixType MatrixType;
+
+    /** \brief Default constructor for fixed-size matrices.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute(). This constructor
+      * can only be used if \p _MatrixType is a fixed-size matrix; use
+      * GeneralizedSelfAdjointEigenSolver(Index) for dynamic-size matrices.
+      */
+    GeneralizedSelfAdjointEigenSolver() : Base() {}
+
+    /** \brief Constructor, pre-allocates memory for dynamic-size matrices.
+      *
+      * \param [in]  size  Positive integer, size of the matrix whose
+      * eigenvalues and eigenvectors will be computed.
+      *
+      * This constructor is useful for dynamic-size matrices, when the user
+      * intends to perform decompositions via compute(). The \p size
+      * parameter is only used as a hint. It is not an error to give a wrong
+      * \p size, but it may impair performance.
+      *
+      * \sa compute() for an example
+      */
+    explicit GeneralizedSelfAdjointEigenSolver(Index size)
+        : Base(size)
+    {}
+
+    /** \brief Constructor; computes generalized eigendecomposition of given matrix pencil.
+      *
+      * \param[in]  matA  Selfadjoint matrix in matrix pencil.
+      *                   Only the lower triangular part of the matrix is referenced.
+      * \param[in]  matB  Positive-definite matrix in matrix pencil.
+      *                   Only the lower triangular part of the matrix is referenced.
+      * \param[in]  options A or-ed set of flags {#ComputeEigenvectors,#EigenvaluesOnly} | {#Ax_lBx,#ABx_lx,#BAx_lx}.
+      *                     Default is #ComputeEigenvectors|#Ax_lBx.
+      *
+      * This constructor calls compute(const MatrixType&, const MatrixType&, int)
+      * to compute the eigenvalues and (if requested) the eigenvectors of the
+      * generalized eigenproblem \f$ Ax = \lambda B x \f$ with \a matA the
+      * selfadjoint matrix \f$ A \f$ and \a matB the positive definite matrix
+      * \f$ B \f$. Each eigenvector \f$ x \f$ satisfies the property
+      * \f$ x^* B x = 1 \f$. The eigenvectors are computed if
+      * \a options contains ComputeEigenvectors.
+      *
+      * In addition, the two following variants can be solved via \p options:
+      * - \c ABx_lx: \f$ ABx = \lambda x \f$
+      * - \c BAx_lx: \f$ BAx = \lambda x \f$
+      *
+      * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType2.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType2.out
+      *
+      * \sa compute(const MatrixType&, const MatrixType&, int)
+      */
+    GeneralizedSelfAdjointEigenSolver(const MatrixType& matA, const MatrixType& matB,
+                                      int options = ComputeEigenvectors|Ax_lBx)
+      : Base(matA.cols())
+    {
+      compute(matA, matB, options);
+    }
+
+    /** \brief Computes generalized eigendecomposition of given matrix pencil.
+      *
+      * \param[in]  matA  Selfadjoint matrix in matrix pencil.
+      *                   Only the lower triangular part of the matrix is referenced.
+      * \param[in]  matB  Positive-definite matrix in matrix pencil.
+      *                   Only the lower triangular part of the matrix is referenced.
+      * \param[in]  options A or-ed set of flags {#ComputeEigenvectors,#EigenvaluesOnly} | {#Ax_lBx,#ABx_lx,#BAx_lx}.
+      *                     Default is #ComputeEigenvectors|#Ax_lBx.
+      *
+      * \returns    Reference to \c *this
+      *
+      * Accoring to \p options, this function computes eigenvalues and (if requested)
+      * the eigenvectors of one of the following three generalized eigenproblems:
+      * - \c Ax_lBx: \f$ Ax = \lambda B x \f$
+      * - \c ABx_lx: \f$ ABx = \lambda x \f$
+      * - \c BAx_lx: \f$ BAx = \lambda x \f$
+      * with \a matA the selfadjoint matrix \f$ A \f$ and \a matB the positive definite
+      * matrix \f$ B \f$.
+      * In addition, each eigenvector \f$ x \f$ satisfies the property \f$ x^* B x = 1 \f$.
+      *
+      * The eigenvalues() function can be used to retrieve
+      * the eigenvalues. If \p options contains ComputeEigenvectors, then the
+      * eigenvectors are also computed and can be retrieved by calling
+      * eigenvectors().
+      *
+      * The implementation uses LLT to compute the Cholesky decomposition
+      * \f$ B = LL^* \f$ and computes the classical eigendecomposition
+      * of the selfadjoint matrix \f$ L^{-1} A (L^*)^{-1} \f$ if \p options contains Ax_lBx
+      * and of \f$ L^{*} A L \f$ otherwise. This solves the
+      * generalized eigenproblem, because any solution of the generalized
+      * eigenproblem \f$ Ax = \lambda B x \f$ corresponds to a solution
+      * \f$ L^{-1} A (L^*)^{-1} (L^* x) = \lambda (L^* x) \f$ of the
+      * eigenproblem for \f$ L^{-1} A (L^*)^{-1} \f$. Similar statements
+      * can be made for the two other variants.
+      *
+      * Example: \include SelfAdjointEigenSolver_compute_MatrixType2.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_compute_MatrixType2.out
+      *
+      * \sa GeneralizedSelfAdjointEigenSolver(const MatrixType&, const MatrixType&, int)
+      */
+    GeneralizedSelfAdjointEigenSolver& compute(const MatrixType& matA, const MatrixType& matB,
+                                               int options = ComputeEigenvectors|Ax_lBx);
+
+  protected:
+
+};
+
+
+template<typename MatrixType>
+GeneralizedSelfAdjointEigenSolver<MatrixType>& GeneralizedSelfAdjointEigenSolver<MatrixType>::
+compute(const MatrixType& matA, const MatrixType& matB, int options)
+{
+  eigen_assert(matA.cols()==matA.rows() && matB.rows()==matA.rows() && matB.cols()==matB.rows());
+  eigen_assert((options&~(EigVecMask|GenEigMask))==0
+          && (options&EigVecMask)!=EigVecMask
+          && ((options&GenEigMask)==0 || (options&GenEigMask)==Ax_lBx
+           || (options&GenEigMask)==ABx_lx || (options&GenEigMask)==BAx_lx)
+          && "invalid option parameter");
+
+  bool computeEigVecs = ((options&EigVecMask)==0) || ((options&EigVecMask)==ComputeEigenvectors);
+
+  // Compute the cholesky decomposition of matB = L L' = U'U
+  LLT<MatrixType> cholB(matB);
+
+  int type = (options&GenEigMask);
+  if(type==0)
+    type = Ax_lBx;
+
+  if(type==Ax_lBx)
+  {
+    // compute C = inv(L) A inv(L')
+    MatrixType matC = matA.template selfadjointView<Lower>();
+    cholB.matrixL().template solveInPlace<OnTheLeft>(matC);
+    cholB.matrixU().template solveInPlace<OnTheRight>(matC);
+
+    Base::compute(matC, computeEigVecs ? ComputeEigenvectors : EigenvaluesOnly );
+
+    // transform back the eigen vectors: evecs = inv(U) * evecs
+    if(computeEigVecs)
+      cholB.matrixU().solveInPlace(Base::m_eivec);
+  }
+  else if(type==ABx_lx)
+  {
+    // compute C = L' A L
+    MatrixType matC = matA.template selfadjointView<Lower>();
+    matC = matC * cholB.matrixL();
+    matC = cholB.matrixU() * matC;
+
+    Base::compute(matC, computeEigVecs ? ComputeEigenvectors : EigenvaluesOnly);
+
+    // transform back the eigen vectors: evecs = inv(U) * evecs
+    if(computeEigVecs)
+      cholB.matrixU().solveInPlace(Base::m_eivec);
+  }
+  else if(type==BAx_lx)
+  {
+    // compute C = L' A L
+    MatrixType matC = matA.template selfadjointView<Lower>();
+    matC = matC * cholB.matrixL();
+    matC = cholB.matrixU() * matC;
+
+    Base::compute(matC, computeEigVecs ? ComputeEigenvectors : EigenvaluesOnly);
+
+    // transform back the eigen vectors: evecs = L * evecs
+    if(computeEigVecs)
+      Base::m_eivec = cholB.matrixL() * Base::m_eivec;
+  }
+
+  return *this;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_GENERALIZEDSELFADJOINTEIGENSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/HessenbergDecomposition.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/HessenbergDecomposition.h
new file mode 100644
index 0000000..f647f69
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/HessenbergDecomposition.h
@@ -0,0 +1,374 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HESSENBERGDECOMPOSITION_H
+#define EIGEN_HESSENBERGDECOMPOSITION_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<typename MatrixType> struct HessenbergDecompositionMatrixHReturnType;
+template<typename MatrixType>
+struct traits<HessenbergDecompositionMatrixHReturnType<MatrixType> >
+{
+  typedef MatrixType ReturnType;
+};
+
+}
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class HessenbergDecomposition
+  *
+  * \brief Reduces a square matrix to Hessenberg form by an orthogonal similarity transformation
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the Hessenberg decomposition
+  *
+  * This class performs an Hessenberg decomposition of a matrix \f$ A \f$. In
+  * the real case, the Hessenberg decomposition consists of an orthogonal
+  * matrix \f$ Q \f$ and a Hessenberg matrix \f$ H \f$ such that \f$ A = Q H
+  * Q^T \f$. An orthogonal matrix is a matrix whose inverse equals its
+  * transpose (\f$ Q^{-1} = Q^T \f$). A Hessenberg matrix has zeros below the
+  * subdiagonal, so it is almost upper triangular. The Hessenberg decomposition
+  * of a complex matrix is \f$ A = Q H Q^* \f$ with \f$ Q \f$ unitary (that is,
+  * \f$ Q^{-1} = Q^* \f$).
+  *
+  * Call the function compute() to compute the Hessenberg decomposition of a
+  * given matrix. Alternatively, you can use the
+  * HessenbergDecomposition(const MatrixType&) constructor which computes the
+  * Hessenberg decomposition at construction time. Once the decomposition is
+  * computed, you can use the matrixH() and matrixQ() functions to construct
+  * the matrices H and Q in the decomposition.
+  *
+  * The documentation for matrixH() contains an example of the typical use of
+  * this class.
+  *
+  * \sa class ComplexSchur, class Tridiagonalization, \ref QR_Module "QR Module"
+  */
+template<typename _MatrixType> class HessenbergDecomposition
+{
+  public:
+
+    /** \brief Synonym for the template parameter \p _MatrixType. */
+    typedef _MatrixType MatrixType;
+
+    enum {
+      Size = MatrixType::RowsAtCompileTime,
+      SizeMinusOne = Size == Dynamic ? Dynamic : Size - 1,
+      Options = MatrixType::Options,
+      MaxSize = MatrixType::MaxRowsAtCompileTime,
+      MaxSizeMinusOne = MaxSize == Dynamic ? Dynamic : MaxSize - 1
+    };
+
+    /** \brief Scalar type for matrices of type #MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    /** \brief Type for vector of Householder coefficients.
+      *
+      * This is column vector with entries of type #Scalar. The length of the
+      * vector is one less than the size of #MatrixType, if it is a fixed-side
+      * type.
+      */
+    typedef Matrix<Scalar, SizeMinusOne, 1, Options & ~RowMajor, MaxSizeMinusOne, 1> CoeffVectorType;
+
+    /** \brief Return type of matrixQ() */
+    typedef HouseholderSequence<MatrixType,typename internal::remove_all<typename CoeffVectorType::ConjugateReturnType>::type> HouseholderSequenceType;
+    
+    typedef internal::HessenbergDecompositionMatrixHReturnType<MatrixType> MatrixHReturnType;
+
+    /** \brief Default constructor; the decomposition will be computed later.
+      *
+      * \param [in] size  The size of the matrix whose Hessenberg decomposition will be computed.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute().  The \p size parameter is only
+      * used as a hint. It is not an error to give a wrong \p size, but it may
+      * impair performance.
+      *
+      * \sa compute() for an example.
+      */
+    explicit HessenbergDecomposition(Index size = Size==Dynamic ? 2 : Size)
+      : m_matrix(size,size),
+        m_temp(size),
+        m_isInitialized(false)
+    {
+      if(size>1)
+        m_hCoeffs.resize(size-1);
+    }
+
+    /** \brief Constructor; computes Hessenberg decomposition of given matrix.
+      *
+      * \param[in]  matrix  Square matrix whose Hessenberg decomposition is to be computed.
+      *
+      * This constructor calls compute() to compute the Hessenberg
+      * decomposition.
+      *
+      * \sa matrixH() for an example.
+      */
+    template<typename InputType>
+    explicit HessenbergDecomposition(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
+        m_temp(matrix.rows()),
+        m_isInitialized(false)
+    {
+      if(matrix.rows()<2)
+      {
+        m_isInitialized = true;
+        return;
+      }
+      m_hCoeffs.resize(matrix.rows()-1,1);
+      _compute(m_matrix, m_hCoeffs, m_temp);
+      m_isInitialized = true;
+    }
+
+    /** \brief Computes Hessenberg decomposition of given matrix.
+      *
+      * \param[in]  matrix  Square matrix whose Hessenberg decomposition is to be computed.
+      * \returns    Reference to \c *this
+      *
+      * The Hessenberg decomposition is computed by bringing the columns of the
+      * matrix successively in the required form using Householder reflections
+      * (see, e.g., Algorithm 7.4.2 in Golub \& Van Loan, <i>%Matrix
+      * Computations</i>). The cost is \f$ 10n^3/3 \f$ flops, where \f$ n \f$
+      * denotes the size of the given matrix.
+      *
+      * This method reuses of the allocated data in the HessenbergDecomposition
+      * object.
+      *
+      * Example: \include HessenbergDecomposition_compute.cpp
+      * Output: \verbinclude HessenbergDecomposition_compute.out
+      */
+    template<typename InputType>
+    HessenbergDecomposition& compute(const EigenBase<InputType>& matrix)
+    {
+      m_matrix = matrix.derived();
+      if(matrix.rows()<2)
+      {
+        m_isInitialized = true;
+        return *this;
+      }
+      m_hCoeffs.resize(matrix.rows()-1,1);
+      _compute(m_matrix, m_hCoeffs, m_temp);
+      m_isInitialized = true;
+      return *this;
+    }
+
+    /** \brief Returns the Householder coefficients.
+      *
+      * \returns a const reference to the vector of Householder coefficients
+      *
+      * \pre Either the constructor HessenbergDecomposition(const MatrixType&)
+      * or the member function compute(const MatrixType&) has been called
+      * before to compute the Hessenberg decomposition of a matrix.
+      *
+      * The Householder coefficients allow the reconstruction of the matrix
+      * \f$ Q \f$ in the Hessenberg decomposition from the packed data.
+      *
+      * \sa packedMatrix(), \ref Householder_Module "Householder module"
+      */
+    const CoeffVectorType& householderCoefficients() const
+    {
+      eigen_assert(m_isInitialized && "HessenbergDecomposition is not initialized.");
+      return m_hCoeffs;
+    }
+
+    /** \brief Returns the internal representation of the decomposition
+      *
+      *	\returns a const reference to a matrix with the internal representation
+      *	         of the decomposition.
+      *
+      * \pre Either the constructor HessenbergDecomposition(const MatrixType&)
+      * or the member function compute(const MatrixType&) has been called
+      * before to compute the Hessenberg decomposition of a matrix.
+      *
+      * The returned matrix contains the following information:
+      *  - the upper part and lower sub-diagonal represent the Hessenberg matrix H
+      *  - the rest of the lower part contains the Householder vectors that, combined with
+      *    Householder coefficients returned by householderCoefficients(),
+      *    allows to reconstruct the matrix Q as
+      *       \f$ Q = H_{N-1} \ldots H_1 H_0 \f$.
+      *    Here, the matrices \f$ H_i \f$ are the Householder transformations
+      *       \f$ H_i = (I - h_i v_i v_i^T) \f$
+      *    where \f$ h_i \f$ is the \f$ i \f$th Householder coefficient and
+      *    \f$ v_i \f$ is the Householder vector defined by
+      *       \f$ v_i = [ 0, \ldots, 0, 1, M(i+2,i), \ldots, M(N-1,i) ]^T \f$
+      *    with M the matrix returned by this function.
+      *
+      * See LAPACK for further details on this packed storage.
+      *
+      * Example: \include HessenbergDecomposition_packedMatrix.cpp
+      * Output: \verbinclude HessenbergDecomposition_packedMatrix.out
+      *
+      * \sa householderCoefficients()
+      */
+    const MatrixType& packedMatrix() const
+    {
+      eigen_assert(m_isInitialized && "HessenbergDecomposition is not initialized.");
+      return m_matrix;
+    }
+
+    /** \brief Reconstructs the orthogonal matrix Q in the decomposition
+      *
+      * \returns object representing the matrix Q
+      *
+      * \pre Either the constructor HessenbergDecomposition(const MatrixType&)
+      * or the member function compute(const MatrixType&) has been called
+      * before to compute the Hessenberg decomposition of a matrix.
+      *
+      * This function returns a light-weight object of template class
+      * HouseholderSequence. You can either apply it directly to a matrix or
+      * you can convert it to a matrix of type #MatrixType.
+      *
+      * \sa matrixH() for an example, class HouseholderSequence
+      */
+    HouseholderSequenceType matrixQ() const
+    {
+      eigen_assert(m_isInitialized && "HessenbergDecomposition is not initialized.");
+      return HouseholderSequenceType(m_matrix, m_hCoeffs.conjugate())
+             .setLength(m_matrix.rows() - 1)
+             .setShift(1);
+    }
+
+    /** \brief Constructs the Hessenberg matrix H in the decomposition
+      *
+      * \returns expression object representing the matrix H
+      *
+      * \pre Either the constructor HessenbergDecomposition(const MatrixType&)
+      * or the member function compute(const MatrixType&) has been called
+      * before to compute the Hessenberg decomposition of a matrix.
+      *
+      * The object returned by this function constructs the Hessenberg matrix H
+      * when it is assigned to a matrix or otherwise evaluated. The matrix H is
+      * constructed from the packed matrix as returned by packedMatrix(): The
+      * upper part (including the subdiagonal) of the packed matrix contains
+      * the matrix H. It may sometimes be better to directly use the packed
+      * matrix instead of constructing the matrix H.
+      *
+      * Example: \include HessenbergDecomposition_matrixH.cpp
+      * Output: \verbinclude HessenbergDecomposition_matrixH.out
+      *
+      * \sa matrixQ(), packedMatrix()
+      */
+    MatrixHReturnType matrixH() const
+    {
+      eigen_assert(m_isInitialized && "HessenbergDecomposition is not initialized.");
+      return MatrixHReturnType(*this);
+    }
+
+  private:
+
+    typedef Matrix<Scalar, 1, Size, Options | RowMajor, 1, MaxSize> VectorType;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    static void _compute(MatrixType& matA, CoeffVectorType& hCoeffs, VectorType& temp);
+
+  protected:
+    MatrixType m_matrix;
+    CoeffVectorType m_hCoeffs;
+    VectorType m_temp;
+    bool m_isInitialized;
+};
+
+/** \internal
+  * Performs a tridiagonal decomposition of \a matA in place.
+  *
+  * \param matA the input selfadjoint matrix
+  * \param hCoeffs returned Householder coefficients
+  *
+  * The result is written in the lower triangular part of \a matA.
+  *
+  * Implemented from Golub's "%Matrix Computations", algorithm 8.3.1.
+  *
+  * \sa packedMatrix()
+  */
+template<typename MatrixType>
+void HessenbergDecomposition<MatrixType>::_compute(MatrixType& matA, CoeffVectorType& hCoeffs, VectorType& temp)
+{
+  eigen_assert(matA.rows()==matA.cols());
+  Index n = matA.rows();
+  temp.resize(n);
+  for (Index i = 0; i<n-1; ++i)
+  {
+    // let's consider the vector v = i-th column starting at position i+1
+    Index remainingSize = n-i-1;
+    RealScalar beta;
+    Scalar h;
+    matA.col(i).tail(remainingSize).makeHouseholderInPlace(h, beta);
+    matA.col(i).coeffRef(i+1) = beta;
+    hCoeffs.coeffRef(i) = h;
+
+    // Apply similarity transformation to remaining columns,
+    // i.e., compute A = H A H'
+
+    // A = H A
+    matA.bottomRightCorner(remainingSize, remainingSize)
+        .applyHouseholderOnTheLeft(matA.col(i).tail(remainingSize-1), h, &temp.coeffRef(0));
+
+    // A = A H'
+    matA.rightCols(remainingSize)
+        .applyHouseholderOnTheRight(matA.col(i).tail(remainingSize-1).conjugate(), numext::conj(h), &temp.coeffRef(0));
+  }
+}
+
+namespace internal {
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \brief Expression type for return value of HessenbergDecomposition::matrixH()
+  *
+  * \tparam MatrixType type of matrix in the Hessenberg decomposition
+  *
+  * Objects of this type represent the Hessenberg matrix in the Hessenberg
+  * decomposition of some matrix. The object holds a reference to the
+  * HessenbergDecomposition class until the it is assigned or evaluated for
+  * some other reason (the reference should remain valid during the life time
+  * of this object). This class is the return type of
+  * HessenbergDecomposition::matrixH(); there is probably no other use for this
+  * class.
+  */
+template<typename MatrixType> struct HessenbergDecompositionMatrixHReturnType
+: public ReturnByValue<HessenbergDecompositionMatrixHReturnType<MatrixType> >
+{
+  public:
+    /** \brief Constructor.
+      *
+      * \param[in] hess  Hessenberg decomposition
+      */
+    HessenbergDecompositionMatrixHReturnType(const HessenbergDecomposition<MatrixType>& hess) : m_hess(hess) { }
+
+    /** \brief Hessenberg matrix in decomposition.
+      *
+      * \param[out] result  Hessenberg matrix in decomposition \p hess which
+      *                     was passed to the constructor
+      */
+    template <typename ResultType>
+    inline void evalTo(ResultType& result) const
+    {
+      result = m_hess.packedMatrix();
+      Index n = result.rows();
+      if (n>2)
+        result.bottomLeftCorner(n-2, n-2).template triangularView<Lower>().setZero();
+    }
+
+    Index rows() const { return m_hess.packedMatrix().rows(); }
+    Index cols() const { return m_hess.packedMatrix().cols(); }
+
+  protected:
+    const HessenbergDecomposition<MatrixType>& m_hess;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_HESSENBERGDECOMPOSITION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
new file mode 100644
index 0000000..4fec8af
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/MatrixBaseEigenvalues.h
@@ -0,0 +1,160 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MATRIXBASEEIGENVALUES_H
+#define EIGEN_MATRIXBASEEIGENVALUES_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Derived, bool IsComplex>
+struct eigenvalues_selector
+{
+  // this is the implementation for the case IsComplex = true
+  static inline typename MatrixBase<Derived>::EigenvaluesReturnType const
+  run(const MatrixBase<Derived>& m)
+  {
+    typedef typename Derived::PlainObject PlainObject;
+    PlainObject m_eval(m);
+    return ComplexEigenSolver<PlainObject>(m_eval, false).eigenvalues();
+  }
+};
+
+template<typename Derived>
+struct eigenvalues_selector<Derived, false>
+{
+  static inline typename MatrixBase<Derived>::EigenvaluesReturnType const
+  run(const MatrixBase<Derived>& m)
+  {
+    typedef typename Derived::PlainObject PlainObject;
+    PlainObject m_eval(m);
+    return EigenSolver<PlainObject>(m_eval, false).eigenvalues();
+  }
+};
+
+} // end namespace internal
+
+/** \brief Computes the eigenvalues of a matrix 
+  * \returns Column vector containing the eigenvalues.
+  *
+  * \eigenvalues_module
+  * This function computes the eigenvalues with the help of the EigenSolver
+  * class (for real matrices) or the ComplexEigenSolver class (for complex
+  * matrices). 
+  *
+  * The eigenvalues are repeated according to their algebraic multiplicity,
+  * so there are as many eigenvalues as rows in the matrix.
+  *
+  * The SelfAdjointView class provides a better algorithm for selfadjoint
+  * matrices.
+  *
+  * Example: \include MatrixBase_eigenvalues.cpp
+  * Output: \verbinclude MatrixBase_eigenvalues.out
+  *
+  * \sa EigenSolver::eigenvalues(), ComplexEigenSolver::eigenvalues(),
+  *     SelfAdjointView::eigenvalues()
+  */
+template<typename Derived>
+inline typename MatrixBase<Derived>::EigenvaluesReturnType
+MatrixBase<Derived>::eigenvalues() const
+{
+  typedef typename internal::traits<Derived>::Scalar Scalar;
+  return internal::eigenvalues_selector<Derived, NumTraits<Scalar>::IsComplex>::run(derived());
+}
+
+/** \brief Computes the eigenvalues of a matrix
+  * \returns Column vector containing the eigenvalues.
+  *
+  * \eigenvalues_module
+  * This function computes the eigenvalues with the help of the
+  * SelfAdjointEigenSolver class.  The eigenvalues are repeated according to
+  * their algebraic multiplicity, so there are as many eigenvalues as rows in
+  * the matrix.
+  *
+  * Example: \include SelfAdjointView_eigenvalues.cpp
+  * Output: \verbinclude SelfAdjointView_eigenvalues.out
+  *
+  * \sa SelfAdjointEigenSolver::eigenvalues(), MatrixBase::eigenvalues()
+  */
+template<typename MatrixType, unsigned int UpLo> 
+inline typename SelfAdjointView<MatrixType, UpLo>::EigenvaluesReturnType
+SelfAdjointView<MatrixType, UpLo>::eigenvalues() const
+{
+  typedef typename SelfAdjointView<MatrixType, UpLo>::PlainObject PlainObject;
+  PlainObject thisAsMatrix(*this);
+  return SelfAdjointEigenSolver<PlainObject>(thisAsMatrix, false).eigenvalues();
+}
+
+
+
+/** \brief Computes the L2 operator norm
+  * \returns Operator norm of the matrix.
+  *
+  * \eigenvalues_module
+  * This function computes the L2 operator norm of a matrix, which is also
+  * known as the spectral norm. The norm of a matrix \f$ A \f$ is defined to be
+  * \f[ \|A\|_2 = \max_x \frac{\|Ax\|_2}{\|x\|_2} \f]
+  * where the maximum is over all vectors and the norm on the right is the
+  * Euclidean vector norm. The norm equals the largest singular value, which is
+  * the square root of the largest eigenvalue of the positive semi-definite
+  * matrix \f$ A^*A \f$.
+  *
+  * The current implementation uses the eigenvalues of \f$ A^*A \f$, as computed
+  * by SelfAdjointView::eigenvalues(), to compute the operator norm of a
+  * matrix.  The SelfAdjointView class provides a better algorithm for
+  * selfadjoint matrices.
+  *
+  * Example: \include MatrixBase_operatorNorm.cpp
+  * Output: \verbinclude MatrixBase_operatorNorm.out
+  *
+  * \sa SelfAdjointView::eigenvalues(), SelfAdjointView::operatorNorm()
+  */
+template<typename Derived>
+inline typename MatrixBase<Derived>::RealScalar
+MatrixBase<Derived>::operatorNorm() const
+{
+  using std::sqrt;
+  typename Derived::PlainObject m_eval(derived());
+  // FIXME if it is really guaranteed that the eigenvalues are already sorted,
+  // then we don't need to compute a maxCoeff() here, comparing the 1st and last ones is enough.
+  return sqrt((m_eval*m_eval.adjoint())
+                 .eval()
+		 .template selfadjointView<Lower>()
+		 .eigenvalues()
+		 .maxCoeff()
+		 );
+}
+
+/** \brief Computes the L2 operator norm
+  * \returns Operator norm of the matrix.
+  *
+  * \eigenvalues_module
+  * This function computes the L2 operator norm of a self-adjoint matrix. For a
+  * self-adjoint matrix, the operator norm is the largest eigenvalue.
+  *
+  * The current implementation uses the eigenvalues of the matrix, as computed
+  * by eigenvalues(), to compute the operator norm of the matrix.
+  *
+  * Example: \include SelfAdjointView_operatorNorm.cpp
+  * Output: \verbinclude SelfAdjointView_operatorNorm.out
+  *
+  * \sa eigenvalues(), MatrixBase::operatorNorm()
+  */
+template<typename MatrixType, unsigned int UpLo>
+inline typename SelfAdjointView<MatrixType, UpLo>::RealScalar
+SelfAdjointView<MatrixType, UpLo>::operatorNorm() const
+{
+  return eigenvalues().cwiseAbs().maxCoeff();
+}
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealQZ.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealQZ.h
new file mode 100644
index 0000000..b3a910d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealQZ.h
@@ -0,0 +1,654 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Alexey Korepanov <kaikaikai@yandex.ru>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REAL_QZ_H
+#define EIGEN_REAL_QZ_H
+
+namespace Eigen {
+
+  /** \eigenvalues_module \ingroup Eigenvalues_Module
+   *
+   *
+   * \class RealQZ
+   *
+   * \brief Performs a real QZ decomposition of a pair of square matrices
+   *
+   * \tparam _MatrixType the type of the matrix of which we are computing the
+   * real QZ decomposition; this is expected to be an instantiation of the
+   * Matrix class template.
+   *
+   * Given a real square matrices A and B, this class computes the real QZ
+   * decomposition: \f$ A = Q S Z \f$, \f$ B = Q T Z \f$ where Q and Z are
+   * real orthogonal matrixes, T is upper-triangular matrix, and S is upper
+   * quasi-triangular matrix. An orthogonal matrix is a matrix whose
+   * inverse is equal to its transpose, \f$ U^{-1} = U^T \f$. A quasi-triangular
+   * matrix is a block-triangular matrix whose diagonal consists of 1-by-1
+   * blocks and 2-by-2 blocks where further reduction is impossible due to
+   * complex eigenvalues. 
+   *
+   * The eigenvalues of the pencil \f$ A - z B \f$ can be obtained from
+   * 1x1 and 2x2 blocks on the diagonals of S and T.
+   *
+   * Call the function compute() to compute the real QZ decomposition of a
+   * given pair of matrices. Alternatively, you can use the 
+   * RealQZ(const MatrixType& B, const MatrixType& B, bool computeQZ)
+   * constructor which computes the real QZ decomposition at construction
+   * time. Once the decomposition is computed, you can use the matrixS(),
+   * matrixT(), matrixQ() and matrixZ() functions to retrieve the matrices
+   * S, T, Q and Z in the decomposition. If computeQZ==false, some time
+   * is saved by not computing matrices Q and Z.
+   *
+   * Example: \include RealQZ_compute.cpp
+   * Output: \include RealQZ_compute.out
+   *
+   * \note The implementation is based on the algorithm in "Matrix Computations"
+   * by Gene H. Golub and Charles F. Van Loan, and a paper "An algorithm for
+   * generalized eigenvalue problems" by C.B.Moler and G.W.Stewart.
+   *
+   * \sa class RealSchur, class ComplexSchur, class EigenSolver, class ComplexEigenSolver
+   */
+
+  template<typename _MatrixType> class RealQZ
+  {
+    public:
+      typedef _MatrixType MatrixType;
+      enum {
+        RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+        ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+        Options = MatrixType::Options,
+        MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+        MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+      };
+      typedef typename MatrixType::Scalar Scalar;
+      typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
+      typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+      typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType;
+      typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType;
+
+      /** \brief Default constructor.
+       *
+       * \param [in] size  Positive integer, size of the matrix whose QZ decomposition will be computed.
+       *
+       * The default constructor is useful in cases in which the user intends to
+       * perform decompositions via compute().  The \p size parameter is only
+       * used as a hint. It is not an error to give a wrong \p size, but it may
+       * impair performance.
+       *
+       * \sa compute() for an example.
+       */
+      explicit RealQZ(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime) :
+        m_S(size, size),
+        m_T(size, size),
+        m_Q(size, size),
+        m_Z(size, size),
+        m_workspace(size*2),
+        m_maxIters(400),
+        m_isInitialized(false)
+        { }
+
+      /** \brief Constructor; computes real QZ decomposition of given matrices
+       * 
+       * \param[in]  A          Matrix A.
+       * \param[in]  B          Matrix B.
+       * \param[in]  computeQZ  If false, A and Z are not computed.
+       *
+       * This constructor calls compute() to compute the QZ decomposition.
+       */
+      RealQZ(const MatrixType& A, const MatrixType& B, bool computeQZ = true) :
+        m_S(A.rows(),A.cols()),
+        m_T(A.rows(),A.cols()),
+        m_Q(A.rows(),A.cols()),
+        m_Z(A.rows(),A.cols()),
+        m_workspace(A.rows()*2),
+        m_maxIters(400),
+        m_isInitialized(false) {
+          compute(A, B, computeQZ);
+        }
+
+      /** \brief Returns matrix Q in the QZ decomposition. 
+       *
+       * \returns A const reference to the matrix Q.
+       */
+      const MatrixType& matrixQ() const {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        eigen_assert(m_computeQZ && "The matrices Q and Z have not been computed during the QZ decomposition.");
+        return m_Q;
+      }
+
+      /** \brief Returns matrix Z in the QZ decomposition. 
+       *
+       * \returns A const reference to the matrix Z.
+       */
+      const MatrixType& matrixZ() const {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        eigen_assert(m_computeQZ && "The matrices Q and Z have not been computed during the QZ decomposition.");
+        return m_Z;
+      }
+
+      /** \brief Returns matrix S in the QZ decomposition. 
+       *
+       * \returns A const reference to the matrix S.
+       */
+      const MatrixType& matrixS() const {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        return m_S;
+      }
+
+      /** \brief Returns matrix S in the QZ decomposition. 
+       *
+       * \returns A const reference to the matrix S.
+       */
+      const MatrixType& matrixT() const {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        return m_T;
+      }
+
+      /** \brief Computes QZ decomposition of given matrix. 
+       * 
+       * \param[in]  A          Matrix A.
+       * \param[in]  B          Matrix B.
+       * \param[in]  computeQZ  If false, A and Z are not computed.
+       * \returns    Reference to \c *this
+       */
+      RealQZ& compute(const MatrixType& A, const MatrixType& B, bool computeQZ = true);
+
+      /** \brief Reports whether previous computation was successful.
+       *
+       * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+       */
+      ComputationInfo info() const
+      {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        return m_info;
+      }
+
+      /** \brief Returns number of performed QR-like iterations.
+      */
+      Index iterations() const
+      {
+        eigen_assert(m_isInitialized && "RealQZ is not initialized.");
+        return m_global_iter;
+      }
+
+      /** Sets the maximal number of iterations allowed to converge to one eigenvalue
+       * or decouple the problem.
+      */
+      RealQZ& setMaxIterations(Index maxIters)
+      {
+        m_maxIters = maxIters;
+        return *this;
+      }
+
+    private:
+
+      MatrixType m_S, m_T, m_Q, m_Z;
+      Matrix<Scalar,Dynamic,1> m_workspace;
+      ComputationInfo m_info;
+      Index m_maxIters;
+      bool m_isInitialized;
+      bool m_computeQZ;
+      Scalar m_normOfT, m_normOfS;
+      Index m_global_iter;
+
+      typedef Matrix<Scalar,3,1> Vector3s;
+      typedef Matrix<Scalar,2,1> Vector2s;
+      typedef Matrix<Scalar,2,2> Matrix2s;
+      typedef JacobiRotation<Scalar> JRs;
+
+      void hessenbergTriangular();
+      void computeNorms();
+      Index findSmallSubdiagEntry(Index iu);
+      Index findSmallDiagEntry(Index f, Index l);
+      void splitOffTwoRows(Index i);
+      void pushDownZero(Index z, Index f, Index l);
+      void step(Index f, Index l, Index iter);
+
+  }; // RealQZ
+
+  /** \internal Reduces S and T to upper Hessenberg - triangular form */
+  template<typename MatrixType>
+    void RealQZ<MatrixType>::hessenbergTriangular()
+    {
+
+      const Index dim = m_S.cols();
+
+      // perform QR decomposition of T, overwrite T with R, save Q
+      HouseholderQR<MatrixType> qrT(m_T);
+      m_T = qrT.matrixQR();
+      m_T.template triangularView<StrictlyLower>().setZero();
+      m_Q = qrT.householderQ();
+      // overwrite S with Q* S
+      m_S.applyOnTheLeft(m_Q.adjoint());
+      // init Z as Identity
+      if (m_computeQZ)
+        m_Z = MatrixType::Identity(dim,dim);
+      // reduce S to upper Hessenberg with Givens rotations
+      for (Index j=0; j<=dim-3; j++) {
+        for (Index i=dim-1; i>=j+2; i--) {
+          JRs G;
+          // kill S(i,j)
+          if(m_S.coeff(i,j) != 0)
+          {
+            G.makeGivens(m_S.coeff(i-1,j), m_S.coeff(i,j), &m_S.coeffRef(i-1, j));
+            m_S.coeffRef(i,j) = Scalar(0.0);
+            m_S.rightCols(dim-j-1).applyOnTheLeft(i-1,i,G.adjoint());
+            m_T.rightCols(dim-i+1).applyOnTheLeft(i-1,i,G.adjoint());
+            // update Q
+            if (m_computeQZ)
+              m_Q.applyOnTheRight(i-1,i,G);
+          }
+          // kill T(i,i-1)
+          if(m_T.coeff(i,i-1)!=Scalar(0))
+          {
+            G.makeGivens(m_T.coeff(i,i), m_T.coeff(i,i-1), &m_T.coeffRef(i,i));
+            m_T.coeffRef(i,i-1) = Scalar(0.0);
+            m_S.applyOnTheRight(i,i-1,G);
+            m_T.topRows(i).applyOnTheRight(i,i-1,G);
+            // update Z
+            if (m_computeQZ)
+              m_Z.applyOnTheLeft(i,i-1,G.adjoint());
+          }
+        }
+      }
+    }
+
+  /** \internal Computes vector L1 norms of S and T when in Hessenberg-Triangular form already */
+  template<typename MatrixType>
+    inline void RealQZ<MatrixType>::computeNorms()
+    {
+      const Index size = m_S.cols();
+      m_normOfS = Scalar(0.0);
+      m_normOfT = Scalar(0.0);
+      for (Index j = 0; j < size; ++j)
+      {
+        m_normOfS += m_S.col(j).segment(0, (std::min)(size,j+2)).cwiseAbs().sum();
+        m_normOfT += m_T.row(j).segment(j, size - j).cwiseAbs().sum();
+      }
+    }
+
+
+  /** \internal Look for single small sub-diagonal element S(res, res-1) and return res (or 0) */
+  template<typename MatrixType>
+    inline Index RealQZ<MatrixType>::findSmallSubdiagEntry(Index iu)
+    {
+      using std::abs;
+      Index res = iu;
+      while (res > 0)
+      {
+        Scalar s = abs(m_S.coeff(res-1,res-1)) + abs(m_S.coeff(res,res));
+        if (s == Scalar(0.0))
+          s = m_normOfS;
+        if (abs(m_S.coeff(res,res-1)) < NumTraits<Scalar>::epsilon() * s)
+          break;
+        res--;
+      }
+      return res;
+    }
+
+  /** \internal Look for single small diagonal element T(res, res) for res between f and l, and return res (or f-1)  */
+  template<typename MatrixType>
+    inline Index RealQZ<MatrixType>::findSmallDiagEntry(Index f, Index l)
+    {
+      using std::abs;
+      Index res = l;
+      while (res >= f) {
+        if (abs(m_T.coeff(res,res)) <= NumTraits<Scalar>::epsilon() * m_normOfT)
+          break;
+        res--;
+      }
+      return res;
+    }
+
+  /** \internal decouple 2x2 diagonal block in rows i, i+1 if eigenvalues are real */
+  template<typename MatrixType>
+    inline void RealQZ<MatrixType>::splitOffTwoRows(Index i)
+    {
+      using std::abs;
+      using std::sqrt;
+      const Index dim=m_S.cols();
+      if (abs(m_S.coeff(i+1,i))==Scalar(0))
+        return;
+      Index j = findSmallDiagEntry(i,i+1);
+      if (j==i-1)
+      {
+        // block of (S T^{-1})
+        Matrix2s STi = m_T.template block<2,2>(i,i).template triangularView<Upper>().
+          template solve<OnTheRight>(m_S.template block<2,2>(i,i));
+        Scalar p = Scalar(0.5)*(STi(0,0)-STi(1,1));
+        Scalar q = p*p + STi(1,0)*STi(0,1);
+        if (q>=0) {
+          Scalar z = sqrt(q);
+          // one QR-like iteration for ABi - lambda I
+          // is enough - when we know exact eigenvalue in advance,
+          // convergence is immediate
+          JRs G;
+          if (p>=0)
+            G.makeGivens(p + z, STi(1,0));
+          else
+            G.makeGivens(p - z, STi(1,0));
+          m_S.rightCols(dim-i).applyOnTheLeft(i,i+1,G.adjoint());
+          m_T.rightCols(dim-i).applyOnTheLeft(i,i+1,G.adjoint());
+          // update Q
+          if (m_computeQZ)
+            m_Q.applyOnTheRight(i,i+1,G);
+
+          G.makeGivens(m_T.coeff(i+1,i+1), m_T.coeff(i+1,i));
+          m_S.topRows(i+2).applyOnTheRight(i+1,i,G);
+          m_T.topRows(i+2).applyOnTheRight(i+1,i,G);
+          // update Z
+          if (m_computeQZ)
+            m_Z.applyOnTheLeft(i+1,i,G.adjoint());
+
+          m_S.coeffRef(i+1,i) = Scalar(0.0);
+          m_T.coeffRef(i+1,i) = Scalar(0.0);
+        }
+      }
+      else
+      {
+        pushDownZero(j,i,i+1);
+      }
+    }
+
+  /** \internal use zero in T(z,z) to zero S(l,l-1), working in block f..l */
+  template<typename MatrixType>
+    inline void RealQZ<MatrixType>::pushDownZero(Index z, Index f, Index l)
+    {
+      JRs G;
+      const Index dim = m_S.cols();
+      for (Index zz=z; zz<l; zz++)
+      {
+        // push 0 down
+        Index firstColS = zz>f ? (zz-1) : zz;
+        G.makeGivens(m_T.coeff(zz, zz+1), m_T.coeff(zz+1, zz+1));
+        m_S.rightCols(dim-firstColS).applyOnTheLeft(zz,zz+1,G.adjoint());
+        m_T.rightCols(dim-zz).applyOnTheLeft(zz,zz+1,G.adjoint());
+        m_T.coeffRef(zz+1,zz+1) = Scalar(0.0);
+        // update Q
+        if (m_computeQZ)
+          m_Q.applyOnTheRight(zz,zz+1,G);
+        // kill S(zz+1, zz-1)
+        if (zz>f)
+        {
+          G.makeGivens(m_S.coeff(zz+1, zz), m_S.coeff(zz+1,zz-1));
+          m_S.topRows(zz+2).applyOnTheRight(zz, zz-1,G);
+          m_T.topRows(zz+1).applyOnTheRight(zz, zz-1,G);
+          m_S.coeffRef(zz+1,zz-1) = Scalar(0.0);
+          // update Z
+          if (m_computeQZ)
+            m_Z.applyOnTheLeft(zz,zz-1,G.adjoint());
+        }
+      }
+      // finally kill S(l,l-1)
+      G.makeGivens(m_S.coeff(l,l), m_S.coeff(l,l-1));
+      m_S.applyOnTheRight(l,l-1,G);
+      m_T.applyOnTheRight(l,l-1,G);
+      m_S.coeffRef(l,l-1)=Scalar(0.0);
+      // update Z
+      if (m_computeQZ)
+        m_Z.applyOnTheLeft(l,l-1,G.adjoint());
+    }
+
+  /** \internal QR-like iterative step for block f..l */
+  template<typename MatrixType>
+    inline void RealQZ<MatrixType>::step(Index f, Index l, Index iter)
+    {
+      using std::abs;
+      const Index dim = m_S.cols();
+
+      // x, y, z
+      Scalar x, y, z;
+      if (iter==10)
+      {
+        // Wilkinson ad hoc shift
+        const Scalar
+          a11=m_S.coeff(f+0,f+0), a12=m_S.coeff(f+0,f+1),
+          a21=m_S.coeff(f+1,f+0), a22=m_S.coeff(f+1,f+1), a32=m_S.coeff(f+2,f+1),
+          b12=m_T.coeff(f+0,f+1),
+          b11i=Scalar(1.0)/m_T.coeff(f+0,f+0),
+          b22i=Scalar(1.0)/m_T.coeff(f+1,f+1),
+          a87=m_S.coeff(l-1,l-2),
+          a98=m_S.coeff(l-0,l-1),
+          b77i=Scalar(1.0)/m_T.coeff(l-2,l-2),
+          b88i=Scalar(1.0)/m_T.coeff(l-1,l-1);
+        Scalar ss = abs(a87*b77i) + abs(a98*b88i),
+               lpl = Scalar(1.5)*ss,
+               ll = ss*ss;
+        x = ll + a11*a11*b11i*b11i - lpl*a11*b11i + a12*a21*b11i*b22i
+          - a11*a21*b12*b11i*b11i*b22i;
+        y = a11*a21*b11i*b11i - lpl*a21*b11i + a21*a22*b11i*b22i 
+          - a21*a21*b12*b11i*b11i*b22i;
+        z = a21*a32*b11i*b22i;
+      }
+      else if (iter==16)
+      {
+        // another exceptional shift
+        x = m_S.coeff(f,f)/m_T.coeff(f,f)-m_S.coeff(l,l)/m_T.coeff(l,l) + m_S.coeff(l,l-1)*m_T.coeff(l-1,l) /
+          (m_T.coeff(l-1,l-1)*m_T.coeff(l,l));
+        y = m_S.coeff(f+1,f)/m_T.coeff(f,f);
+        z = 0;
+      }
+      else if (iter>23 && !(iter%8))
+      {
+        // extremely exceptional shift
+        x = internal::random<Scalar>(-1.0,1.0);
+        y = internal::random<Scalar>(-1.0,1.0);
+        z = internal::random<Scalar>(-1.0,1.0);
+      }
+      else
+      {
+        // Compute the shifts: (x,y,z,0...) = (AB^-1 - l1 I) (AB^-1 - l2 I) e1
+        // where l1 and l2 are the eigenvalues of the 2x2 matrix C = U V^-1 where
+        // U and V are 2x2 bottom right sub matrices of A and B. Thus:
+        //  = AB^-1AB^-1 + l1 l2 I - (l1+l2)(AB^-1)
+        //  = AB^-1AB^-1 + det(M) - tr(M)(AB^-1)
+        // Since we are only interested in having x, y, z with a correct ratio, we have:
+        const Scalar
+          a11 = m_S.coeff(f,f),     a12 = m_S.coeff(f,f+1),
+          a21 = m_S.coeff(f+1,f),   a22 = m_S.coeff(f+1,f+1),
+                                    a32 = m_S.coeff(f+2,f+1),
+
+          a88 = m_S.coeff(l-1,l-1), a89 = m_S.coeff(l-1,l),
+          a98 = m_S.coeff(l,l-1),   a99 = m_S.coeff(l,l),
+
+          b11 = m_T.coeff(f,f),     b12 = m_T.coeff(f,f+1),
+                                    b22 = m_T.coeff(f+1,f+1),
+
+          b88 = m_T.coeff(l-1,l-1), b89 = m_T.coeff(l-1,l),
+                                    b99 = m_T.coeff(l,l);
+
+        x = ( (a88/b88 - a11/b11)*(a99/b99 - a11/b11) - (a89/b99)*(a98/b88) + (a98/b88)*(b89/b99)*(a11/b11) ) * (b11/a21)
+          + a12/b22 - (a11/b11)*(b12/b22);
+        y = (a22/b22-a11/b11) - (a21/b11)*(b12/b22) - (a88/b88-a11/b11) - (a99/b99-a11/b11) + (a98/b88)*(b89/b99);
+        z = a32/b22;
+      }
+
+      JRs G;
+
+      for (Index k=f; k<=l-2; k++)
+      {
+        // variables for Householder reflections
+        Vector2s essential2;
+        Scalar tau, beta;
+
+        Vector3s hr(x,y,z);
+
+        // Q_k to annihilate S(k+1,k-1) and S(k+2,k-1)
+        hr.makeHouseholderInPlace(tau, beta);
+        essential2 = hr.template bottomRows<2>();
+        Index fc=(std::max)(k-1,Index(0));  // first col to update
+        m_S.template middleRows<3>(k).rightCols(dim-fc).applyHouseholderOnTheLeft(essential2, tau, m_workspace.data());
+        m_T.template middleRows<3>(k).rightCols(dim-fc).applyHouseholderOnTheLeft(essential2, tau, m_workspace.data());
+        if (m_computeQZ)
+          m_Q.template middleCols<3>(k).applyHouseholderOnTheRight(essential2, tau, m_workspace.data());
+        if (k>f)
+          m_S.coeffRef(k+2,k-1) = m_S.coeffRef(k+1,k-1) = Scalar(0.0);
+
+        // Z_{k1} to annihilate T(k+2,k+1) and T(k+2,k)
+        hr << m_T.coeff(k+2,k+2),m_T.coeff(k+2,k),m_T.coeff(k+2,k+1);
+        hr.makeHouseholderInPlace(tau, beta);
+        essential2 = hr.template bottomRows<2>();
+        {
+          Index lr = (std::min)(k+4,dim); // last row to update
+          Map<Matrix<Scalar,Dynamic,1> > tmp(m_workspace.data(),lr);
+          // S
+          tmp = m_S.template middleCols<2>(k).topRows(lr) * essential2;
+          tmp += m_S.col(k+2).head(lr);
+          m_S.col(k+2).head(lr) -= tau*tmp;
+          m_S.template middleCols<2>(k).topRows(lr) -= (tau*tmp) * essential2.adjoint();
+          // T
+          tmp = m_T.template middleCols<2>(k).topRows(lr) * essential2;
+          tmp += m_T.col(k+2).head(lr);
+          m_T.col(k+2).head(lr) -= tau*tmp;
+          m_T.template middleCols<2>(k).topRows(lr) -= (tau*tmp) * essential2.adjoint();
+        }
+        if (m_computeQZ)
+        {
+          // Z
+          Map<Matrix<Scalar,1,Dynamic> > tmp(m_workspace.data(),dim);
+          tmp = essential2.adjoint()*(m_Z.template middleRows<2>(k));
+          tmp += m_Z.row(k+2);
+          m_Z.row(k+2) -= tau*tmp;
+          m_Z.template middleRows<2>(k) -= essential2 * (tau*tmp);
+        }
+        m_T.coeffRef(k+2,k) = m_T.coeffRef(k+2,k+1) = Scalar(0.0);
+
+        // Z_{k2} to annihilate T(k+1,k)
+        G.makeGivens(m_T.coeff(k+1,k+1), m_T.coeff(k+1,k));
+        m_S.applyOnTheRight(k+1,k,G);
+        m_T.applyOnTheRight(k+1,k,G);
+        // update Z
+        if (m_computeQZ)
+          m_Z.applyOnTheLeft(k+1,k,G.adjoint());
+        m_T.coeffRef(k+1,k) = Scalar(0.0);
+
+        // update x,y,z
+        x = m_S.coeff(k+1,k);
+        y = m_S.coeff(k+2,k);
+        if (k < l-2)
+          z = m_S.coeff(k+3,k);
+      } // loop over k
+
+      // Q_{n-1} to annihilate y = S(l,l-2)
+      G.makeGivens(x,y);
+      m_S.applyOnTheLeft(l-1,l,G.adjoint());
+      m_T.applyOnTheLeft(l-1,l,G.adjoint());
+      if (m_computeQZ)
+        m_Q.applyOnTheRight(l-1,l,G);
+      m_S.coeffRef(l,l-2) = Scalar(0.0);
+
+      // Z_{n-1} to annihilate T(l,l-1)
+      G.makeGivens(m_T.coeff(l,l),m_T.coeff(l,l-1));
+      m_S.applyOnTheRight(l,l-1,G);
+      m_T.applyOnTheRight(l,l-1,G);
+      if (m_computeQZ)
+        m_Z.applyOnTheLeft(l,l-1,G.adjoint());
+      m_T.coeffRef(l,l-1) = Scalar(0.0);
+    }
+
+  template<typename MatrixType>
+    RealQZ<MatrixType>& RealQZ<MatrixType>::compute(const MatrixType& A_in, const MatrixType& B_in, bool computeQZ)
+    {
+
+      const Index dim = A_in.cols();
+
+      eigen_assert (A_in.rows()==dim && A_in.cols()==dim 
+          && B_in.rows()==dim && B_in.cols()==dim 
+          && "Need square matrices of the same dimension");
+
+      m_isInitialized = true;
+      m_computeQZ = computeQZ;
+      m_S = A_in; m_T = B_in;
+      m_workspace.resize(dim*2);
+      m_global_iter = 0;
+
+      // entrance point: hessenberg triangular decomposition
+      hessenbergTriangular();
+      // compute L1 vector norms of T, S into m_normOfS, m_normOfT
+      computeNorms();
+
+      Index l = dim-1, 
+            f, 
+            local_iter = 0;
+
+      while (l>0 && local_iter<m_maxIters)
+      {
+        f = findSmallSubdiagEntry(l);
+        // now rows and columns f..l (including) decouple from the rest of the problem
+        if (f>0) m_S.coeffRef(f,f-1) = Scalar(0.0);
+        if (f == l) // One root found
+        {
+          l--;
+          local_iter = 0;
+        }
+        else if (f == l-1) // Two roots found
+        {
+          splitOffTwoRows(f);
+          l -= 2;
+          local_iter = 0;
+        }
+        else // No convergence yet
+        {
+          // if there's zero on diagonal of T, we can isolate an eigenvalue with Givens rotations
+          Index z = findSmallDiagEntry(f,l);
+          if (z>=f)
+          {
+            // zero found
+            pushDownZero(z,f,l);
+          }
+          else
+          {
+            // We are sure now that S.block(f,f, l-f+1,l-f+1) is underuced upper-Hessenberg 
+            // and T.block(f,f, l-f+1,l-f+1) is invertible uper-triangular, which allows to
+            // apply a QR-like iteration to rows and columns f..l.
+            step(f,l, local_iter);
+            local_iter++;
+            m_global_iter++;
+          }
+        }
+      }
+      // check if we converged before reaching iterations limit
+      m_info = (local_iter<m_maxIters) ? Success : NoConvergence;
+
+      // For each non triangular 2x2 diagonal block of S,
+      //    reduce the respective 2x2 diagonal block of T to positive diagonal form using 2x2 SVD.
+      // This step is not mandatory for QZ, but it does help further extraction of eigenvalues/eigenvectors,
+      // and is in par with Lapack/Matlab QZ.
+      if(m_info==Success)
+      {
+        for(Index i=0; i<dim-1; ++i)
+        {
+          if(m_S.coeff(i+1, i) != Scalar(0))
+          {
+            JacobiRotation<Scalar> j_left, j_right;
+            internal::real_2x2_jacobi_svd(m_T, i, i+1, &j_left, &j_right);
+
+            // Apply resulting Jacobi rotations
+            m_S.applyOnTheLeft(i,i+1,j_left);
+            m_S.applyOnTheRight(i,i+1,j_right);
+            m_T.applyOnTheLeft(i,i+1,j_left);
+            m_T.applyOnTheRight(i,i+1,j_right);
+            m_T(i+1,i) = m_T(i,i+1) = Scalar(0);
+
+            if(m_computeQZ) {
+              m_Q.applyOnTheRight(i,i+1,j_left.transpose());
+              m_Z.applyOnTheLeft(i,i+1,j_right.transpose());
+            }
+
+            i++;
+          }
+        }
+      }
+
+      return *this;
+    } // end compute
+
+} // end namespace Eigen
+
+#endif //EIGEN_REAL_QZ
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur.h
new file mode 100644
index 0000000..f5c8604
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur.h
@@ -0,0 +1,546 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010,2012 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REAL_SCHUR_H
+#define EIGEN_REAL_SCHUR_H
+
+#include "./HessenbergDecomposition.h"
+
+namespace Eigen { 
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class RealSchur
+  *
+  * \brief Performs a real Schur decomposition of a square matrix
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the
+  * real Schur decomposition; this is expected to be an instantiation of the
+  * Matrix class template.
+  *
+  * Given a real square matrix A, this class computes the real Schur
+  * decomposition: \f$ A = U T U^T \f$ where U is a real orthogonal matrix and
+  * T is a real quasi-triangular matrix. An orthogonal matrix is a matrix whose
+  * inverse is equal to its transpose, \f$ U^{-1} = U^T \f$. A quasi-triangular
+  * matrix is a block-triangular matrix whose diagonal consists of 1-by-1
+  * blocks and 2-by-2 blocks with complex eigenvalues. The eigenvalues of the
+  * blocks on the diagonal of T are the same as the eigenvalues of the matrix
+  * A, and thus the real Schur decomposition is used in EigenSolver to compute
+  * the eigendecomposition of a matrix.
+  *
+  * Call the function compute() to compute the real Schur decomposition of a
+  * given matrix. Alternatively, you can use the RealSchur(const MatrixType&, bool)
+  * constructor which computes the real Schur decomposition at construction
+  * time. Once the decomposition is computed, you can use the matrixU() and
+  * matrixT() functions to retrieve the matrices U and T in the decomposition.
+  *
+  * The documentation of RealSchur(const MatrixType&, bool) contains an example
+  * of the typical use of this class.
+  *
+  * \note The implementation is adapted from
+  * <a href="http://math.nist.gov/javanumerics/jama/">JAMA</a> (public domain).
+  * Their code is based on EISPACK.
+  *
+  * \sa class ComplexSchur, class EigenSolver, class ComplexEigenSolver
+  */
+template<typename _MatrixType> class RealSchur
+{
+  public:
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef std::complex<typename NumTraits<Scalar>::Real> ComplexScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    typedef Matrix<ComplexScalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> EigenvalueType;
+    typedef Matrix<Scalar, ColsAtCompileTime, 1, Options & ~RowMajor, MaxColsAtCompileTime, 1> ColumnVectorType;
+
+    /** \brief Default constructor.
+      *
+      * \param [in] size  Positive integer, size of the matrix whose Schur decomposition will be computed.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute().  The \p size parameter is only
+      * used as a hint. It is not an error to give a wrong \p size, but it may
+      * impair performance.
+      *
+      * \sa compute() for an example.
+      */
+    explicit RealSchur(Index size = RowsAtCompileTime==Dynamic ? 1 : RowsAtCompileTime)
+            : m_matT(size, size),
+              m_matU(size, size),
+              m_workspaceVector(size),
+              m_hess(size),
+              m_isInitialized(false),
+              m_matUisUptodate(false),
+              m_maxIters(-1)
+    { }
+
+    /** \brief Constructor; computes real Schur decomposition of given matrix. 
+      * 
+      * \param[in]  matrix    Square matrix whose Schur decomposition is to be computed.
+      * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
+      *
+      * This constructor calls compute() to compute the Schur decomposition.
+      *
+      * Example: \include RealSchur_RealSchur_MatrixType.cpp
+      * Output: \verbinclude RealSchur_RealSchur_MatrixType.out
+      */
+    template<typename InputType>
+    explicit RealSchur(const EigenBase<InputType>& matrix, bool computeU = true)
+            : m_matT(matrix.rows(),matrix.cols()),
+              m_matU(matrix.rows(),matrix.cols()),
+              m_workspaceVector(matrix.rows()),
+              m_hess(matrix.rows()),
+              m_isInitialized(false),
+              m_matUisUptodate(false),
+              m_maxIters(-1)
+    {
+      compute(matrix.derived(), computeU);
+    }
+
+    /** \brief Returns the orthogonal matrix in the Schur decomposition. 
+      *
+      * \returns A const reference to the matrix U.
+      *
+      * \pre Either the constructor RealSchur(const MatrixType&, bool) or the
+      * member function compute(const MatrixType&, bool) has been called before
+      * to compute the Schur decomposition of a matrix, and \p computeU was set
+      * to true (the default value).
+      *
+      * \sa RealSchur(const MatrixType&, bool) for an example
+      */
+    const MatrixType& matrixU() const
+    {
+      eigen_assert(m_isInitialized && "RealSchur is not initialized.");
+      eigen_assert(m_matUisUptodate && "The matrix U has not been computed during the RealSchur decomposition.");
+      return m_matU;
+    }
+
+    /** \brief Returns the quasi-triangular matrix in the Schur decomposition. 
+      *
+      * \returns A const reference to the matrix T.
+      *
+      * \pre Either the constructor RealSchur(const MatrixType&, bool) or the
+      * member function compute(const MatrixType&, bool) has been called before
+      * to compute the Schur decomposition of a matrix.
+      *
+      * \sa RealSchur(const MatrixType&, bool) for an example
+      */
+    const MatrixType& matrixT() const
+    {
+      eigen_assert(m_isInitialized && "RealSchur is not initialized.");
+      return m_matT;
+    }
+  
+    /** \brief Computes Schur decomposition of given matrix. 
+      * 
+      * \param[in]  matrix    Square matrix whose Schur decomposition is to be computed.
+      * \param[in]  computeU  If true, both T and U are computed; if false, only T is computed.
+      * \returns    Reference to \c *this
+      *
+      * The Schur decomposition is computed by first reducing the matrix to
+      * Hessenberg form using the class HessenbergDecomposition. The Hessenberg
+      * matrix is then reduced to triangular form by performing Francis QR
+      * iterations with implicit double shift. The cost of computing the Schur
+      * decomposition depends on the number of iterations; as a rough guide, it
+      * may be taken to be \f$25n^3\f$ flops if \a computeU is true and
+      * \f$10n^3\f$ flops if \a computeU is false.
+      *
+      * Example: \include RealSchur_compute.cpp
+      * Output: \verbinclude RealSchur_compute.out
+      *
+      * \sa compute(const MatrixType&, bool, Index)
+      */
+    template<typename InputType>
+    RealSchur& compute(const EigenBase<InputType>& matrix, bool computeU = true);
+
+    /** \brief Computes Schur decomposition of a Hessenberg matrix H = Z T Z^T
+     *  \param[in] matrixH Matrix in Hessenberg form H
+     *  \param[in] matrixQ orthogonal matrix Q that transform a matrix A to H : A = Q H Q^T
+     *  \param computeU Computes the matriX U of the Schur vectors
+     * \return Reference to \c *this
+     * 
+     *  This routine assumes that the matrix is already reduced in Hessenberg form matrixH
+     *  using either the class HessenbergDecomposition or another mean. 
+     *  It computes the upper quasi-triangular matrix T of the Schur decomposition of H
+     *  When computeU is true, this routine computes the matrix U such that 
+     *  A = U T U^T =  (QZ) T (QZ)^T = Q H Q^T where A is the initial matrix
+     * 
+     * NOTE Q is referenced if computeU is true; so, if the initial orthogonal matrix
+     * is not available, the user should give an identity matrix (Q.setIdentity())
+     * 
+     * \sa compute(const MatrixType&, bool)
+     */
+    template<typename HessMatrixType, typename OrthMatrixType>
+    RealSchur& computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ,  bool computeU);
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "RealSchur is not initialized.");
+      return m_info;
+    }
+
+    /** \brief Sets the maximum number of iterations allowed. 
+      *
+      * If not specified by the user, the maximum number of iterations is m_maxIterationsPerRow times the size
+      * of the matrix.
+      */
+    RealSchur& setMaxIterations(Index maxIters)
+    {
+      m_maxIters = maxIters;
+      return *this;
+    }
+
+    /** \brief Returns the maximum number of iterations. */
+    Index getMaxIterations()
+    {
+      return m_maxIters;
+    }
+
+    /** \brief Maximum number of iterations per row.
+      *
+      * If not otherwise specified, the maximum number of iterations is this number times the size of the
+      * matrix. It is currently set to 40.
+      */
+    static const int m_maxIterationsPerRow = 40;
+
+  private:
+    
+    MatrixType m_matT;
+    MatrixType m_matU;
+    ColumnVectorType m_workspaceVector;
+    HessenbergDecomposition<MatrixType> m_hess;
+    ComputationInfo m_info;
+    bool m_isInitialized;
+    bool m_matUisUptodate;
+    Index m_maxIters;
+
+    typedef Matrix<Scalar,3,1> Vector3s;
+
+    Scalar computeNormOfT();
+    Index findSmallSubdiagEntry(Index iu);
+    void splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift);
+    void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo);
+    void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector);
+    void performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace);
+};
+
+
+template<typename MatrixType>
+template<typename InputType>
+RealSchur<MatrixType>& RealSchur<MatrixType>::compute(const EigenBase<InputType>& matrix, bool computeU)
+{
+  const Scalar considerAsZero = (std::numeric_limits<Scalar>::min)();
+
+  eigen_assert(matrix.cols() == matrix.rows());
+  Index maxIters = m_maxIters;
+  if (maxIters == -1)
+    maxIters = m_maxIterationsPerRow * matrix.rows();
+
+  Scalar scale = matrix.derived().cwiseAbs().maxCoeff();
+  if(scale<considerAsZero)
+  {
+    m_matT.setZero(matrix.rows(),matrix.cols());
+    if(computeU)
+      m_matU.setIdentity(matrix.rows(),matrix.cols());
+    m_info = Success;
+    m_isInitialized = true;
+    m_matUisUptodate = computeU;
+    return *this;
+  }
+
+  // Step 1. Reduce to Hessenberg form
+  m_hess.compute(matrix.derived()/scale);
+
+  // Step 2. Reduce to real Schur form  
+  computeFromHessenberg(m_hess.matrixH(), m_hess.matrixQ(), computeU);
+
+  m_matT *= scale;
+  
+  return *this;
+}
+template<typename MatrixType>
+template<typename HessMatrixType, typename OrthMatrixType>
+RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMatrixType& matrixH, const OrthMatrixType& matrixQ,  bool computeU)
+{
+  using std::abs;
+
+  m_matT = matrixH;
+  if(computeU)
+    m_matU = matrixQ;
+  
+  Index maxIters = m_maxIters;
+  if (maxIters == -1)
+    maxIters = m_maxIterationsPerRow * matrixH.rows();
+  m_workspaceVector.resize(m_matT.cols());
+  Scalar* workspace = &m_workspaceVector.coeffRef(0);
+
+  // The matrix m_matT is divided in three parts. 
+  // Rows 0,...,il-1 are decoupled from the rest because m_matT(il,il-1) is zero. 
+  // Rows il,...,iu is the part we are working on (the active window).
+  // Rows iu+1,...,end are already brought in triangular form.
+  Index iu = m_matT.cols() - 1;
+  Index iter = 0;      // iteration count for current eigenvalue
+  Index totalIter = 0; // iteration count for whole matrix
+  Scalar exshift(0);   // sum of exceptional shifts
+  Scalar norm = computeNormOfT();
+
+  if(norm!=0)
+  {
+    while (iu >= 0)
+    {
+      Index il = findSmallSubdiagEntry(iu);
+
+      // Check for convergence
+      if (il == iu) // One root found
+      {
+        m_matT.coeffRef(iu,iu) = m_matT.coeff(iu,iu) + exshift;
+        if (iu > 0)
+          m_matT.coeffRef(iu, iu-1) = Scalar(0);
+        iu--;
+        iter = 0;
+      }
+      else if (il == iu-1) // Two roots found
+      {
+        splitOffTwoRows(iu, computeU, exshift);
+        iu -= 2;
+        iter = 0;
+      }
+      else // No convergence yet
+      {
+        // The firstHouseholderVector vector has to be initialized to something to get rid of a silly GCC warning (-O1 -Wall -DNDEBUG )
+        Vector3s firstHouseholderVector(0,0,0), shiftInfo;
+        computeShift(iu, iter, exshift, shiftInfo);
+        iter = iter + 1;
+        totalIter = totalIter + 1;
+        if (totalIter > maxIters) break;
+        Index im;
+        initFrancisQRStep(il, iu, shiftInfo, im, firstHouseholderVector);
+        performFrancisQRStep(il, im, iu, computeU, firstHouseholderVector, workspace);
+      }
+    }
+  }
+  if(totalIter <= maxIters)
+    m_info = Success;
+  else
+    m_info = NoConvergence;
+
+  m_isInitialized = true;
+  m_matUisUptodate = computeU;
+  return *this;
+}
+
+/** \internal Computes and returns vector L1 norm of T */
+template<typename MatrixType>
+inline typename MatrixType::Scalar RealSchur<MatrixType>::computeNormOfT()
+{
+  const Index size = m_matT.cols();
+  // FIXME to be efficient the following would requires a triangular reduxion code
+  // Scalar norm = m_matT.upper().cwiseAbs().sum() 
+  //               + m_matT.bottomLeftCorner(size-1,size-1).diagonal().cwiseAbs().sum();
+  Scalar norm(0);
+  for (Index j = 0; j < size; ++j)
+    norm += m_matT.col(j).segment(0, (std::min)(size,j+2)).cwiseAbs().sum();
+  return norm;
+}
+
+/** \internal Look for single small sub-diagonal element and returns its index */
+template<typename MatrixType>
+inline Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu)
+{
+  using std::abs;
+  Index res = iu;
+  while (res > 0)
+  {
+    Scalar s = abs(m_matT.coeff(res-1,res-1)) + abs(m_matT.coeff(res,res));
+    if (abs(m_matT.coeff(res,res-1)) <= NumTraits<Scalar>::epsilon() * s)
+      break;
+    res--;
+  }
+  return res;
+}
+
+/** \internal Update T given that rows iu-1 and iu decouple from the rest. */
+template<typename MatrixType>
+inline void RealSchur<MatrixType>::splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift)
+{
+  using std::sqrt;
+  using std::abs;
+  const Index size = m_matT.cols();
+
+  // The eigenvalues of the 2x2 matrix [a b; c d] are 
+  // trace +/- sqrt(discr/4) where discr = tr^2 - 4*det, tr = a + d, det = ad - bc
+  Scalar p = Scalar(0.5) * (m_matT.coeff(iu-1,iu-1) - m_matT.coeff(iu,iu));
+  Scalar q = p * p + m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu);   // q = tr^2 / 4 - det = discr/4
+  m_matT.coeffRef(iu,iu) += exshift;
+  m_matT.coeffRef(iu-1,iu-1) += exshift;
+
+  if (q >= Scalar(0)) // Two real eigenvalues
+  {
+    Scalar z = sqrt(abs(q));
+    JacobiRotation<Scalar> rot;
+    if (p >= Scalar(0))
+      rot.makeGivens(p + z, m_matT.coeff(iu, iu-1));
+    else
+      rot.makeGivens(p - z, m_matT.coeff(iu, iu-1));
+
+    m_matT.rightCols(size-iu+1).applyOnTheLeft(iu-1, iu, rot.adjoint());
+    m_matT.topRows(iu+1).applyOnTheRight(iu-1, iu, rot);
+    m_matT.coeffRef(iu, iu-1) = Scalar(0); 
+    if (computeU)
+      m_matU.applyOnTheRight(iu-1, iu, rot);
+  }
+
+  if (iu > 1) 
+    m_matT.coeffRef(iu-1, iu-2) = Scalar(0);
+}
+
+/** \internal Form shift in shiftInfo, and update exshift if an exceptional shift is performed. */
+template<typename MatrixType>
+inline void RealSchur<MatrixType>::computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo)
+{
+  using std::sqrt;
+  using std::abs;
+  shiftInfo.coeffRef(0) = m_matT.coeff(iu,iu);
+  shiftInfo.coeffRef(1) = m_matT.coeff(iu-1,iu-1);
+  shiftInfo.coeffRef(2) = m_matT.coeff(iu,iu-1) * m_matT.coeff(iu-1,iu);
+
+  // Wilkinson's original ad hoc shift
+  if (iter == 10)
+  {
+    exshift += shiftInfo.coeff(0);
+    for (Index i = 0; i <= iu; ++i)
+      m_matT.coeffRef(i,i) -= shiftInfo.coeff(0);
+    Scalar s = abs(m_matT.coeff(iu,iu-1)) + abs(m_matT.coeff(iu-1,iu-2));
+    shiftInfo.coeffRef(0) = Scalar(0.75) * s;
+    shiftInfo.coeffRef(1) = Scalar(0.75) * s;
+    shiftInfo.coeffRef(2) = Scalar(-0.4375) * s * s;
+  }
+
+  // MATLAB's new ad hoc shift
+  if (iter == 30)
+  {
+    Scalar s = (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0);
+    s = s * s + shiftInfo.coeff(2);
+    if (s > Scalar(0))
+    {
+      s = sqrt(s);
+      if (shiftInfo.coeff(1) < shiftInfo.coeff(0))
+        s = -s;
+      s = s + (shiftInfo.coeff(1) - shiftInfo.coeff(0)) / Scalar(2.0);
+      s = shiftInfo.coeff(0) - shiftInfo.coeff(2) / s;
+      exshift += s;
+      for (Index i = 0; i <= iu; ++i)
+        m_matT.coeffRef(i,i) -= s;
+      shiftInfo.setConstant(Scalar(0.964));
+    }
+  }
+}
+
+/** \internal Compute index im at which Francis QR step starts and the first Householder vector. */
+template<typename MatrixType>
+inline void RealSchur<MatrixType>::initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector)
+{
+  using std::abs;
+  Vector3s& v = firstHouseholderVector; // alias to save typing
+
+  for (im = iu-2; im >= il; --im)
+  {
+    const Scalar Tmm = m_matT.coeff(im,im);
+    const Scalar r = shiftInfo.coeff(0) - Tmm;
+    const Scalar s = shiftInfo.coeff(1) - Tmm;
+    v.coeffRef(0) = (r * s - shiftInfo.coeff(2)) / m_matT.coeff(im+1,im) + m_matT.coeff(im,im+1);
+    v.coeffRef(1) = m_matT.coeff(im+1,im+1) - Tmm - r - s;
+    v.coeffRef(2) = m_matT.coeff(im+2,im+1);
+    if (im == il) {
+      break;
+    }
+    const Scalar lhs = m_matT.coeff(im,im-1) * (abs(v.coeff(1)) + abs(v.coeff(2)));
+    const Scalar rhs = v.coeff(0) * (abs(m_matT.coeff(im-1,im-1)) + abs(Tmm) + abs(m_matT.coeff(im+1,im+1)));
+    if (abs(lhs) < NumTraits<Scalar>::epsilon() * rhs)
+      break;
+  }
+}
+
+/** \internal Perform a Francis QR step involving rows il:iu and columns im:iu. */
+template<typename MatrixType>
+inline void RealSchur<MatrixType>::performFrancisQRStep(Index il, Index im, Index iu, bool computeU, const Vector3s& firstHouseholderVector, Scalar* workspace)
+{
+  eigen_assert(im >= il);
+  eigen_assert(im <= iu-2);
+
+  const Index size = m_matT.cols();
+
+  for (Index k = im; k <= iu-2; ++k)
+  {
+    bool firstIteration = (k == im);
+
+    Vector3s v;
+    if (firstIteration)
+      v = firstHouseholderVector;
+    else
+      v = m_matT.template block<3,1>(k,k-1);
+
+    Scalar tau, beta;
+    Matrix<Scalar, 2, 1> ess;
+    v.makeHouseholder(ess, tau, beta);
+    
+    if (beta != Scalar(0)) // if v is not zero
+    {
+      if (firstIteration && k > il)
+        m_matT.coeffRef(k,k-1) = -m_matT.coeff(k,k-1);
+      else if (!firstIteration)
+        m_matT.coeffRef(k,k-1) = beta;
+
+      // These Householder transformations form the O(n^3) part of the algorithm
+      m_matT.block(k, k, 3, size-k).applyHouseholderOnTheLeft(ess, tau, workspace);
+      m_matT.block(0, k, (std::min)(iu,k+3) + 1, 3).applyHouseholderOnTheRight(ess, tau, workspace);
+      if (computeU)
+        m_matU.block(0, k, size, 3).applyHouseholderOnTheRight(ess, tau, workspace);
+    }
+  }
+
+  Matrix<Scalar, 2, 1> v = m_matT.template block<2,1>(iu-1, iu-2);
+  Scalar tau, beta;
+  Matrix<Scalar, 1, 1> ess;
+  v.makeHouseholder(ess, tau, beta);
+
+  if (beta != Scalar(0)) // if v is not zero
+  {
+    m_matT.coeffRef(iu-1, iu-2) = beta;
+    m_matT.block(iu-1, iu-1, 2, size-iu+1).applyHouseholderOnTheLeft(ess, tau, workspace);
+    m_matT.block(0, iu-1, iu+1, 2).applyHouseholderOnTheRight(ess, tau, workspace);
+    if (computeU)
+      m_matU.block(0, iu-1, size, 2).applyHouseholderOnTheRight(ess, tau, workspace);
+  }
+
+  // clean up pollution due to round-off errors
+  for (Index i = im+2; i <= iu; ++i)
+  {
+    m_matT.coeffRef(i,i-2) = Scalar(0);
+    if (i > im+2)
+      m_matT.coeffRef(i,i-3) = Scalar(0);
+  }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_REAL_SCHUR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur_LAPACKE.h
new file mode 100644
index 0000000..2c22517
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/RealSchur_LAPACKE.h
@@ -0,0 +1,77 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Real Schur needed to real unsymmetrical eigenvalues/eigenvectors.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_REAL_SCHUR_LAPACKE_H
+#define EIGEN_REAL_SCHUR_LAPACKE_H
+
+namespace Eigen { 
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_SCHUR_REAL(EIGTYPE, LAPACKE_TYPE, LAPACKE_PREFIX, LAPACKE_PREFIX_U, EIGCOLROW, LAPACKE_COLROW) \
+template<> template<typename InputType> inline \
+RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
+RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, bool computeU) \
+{ \
+  eigen_assert(matrix.cols() == matrix.rows()); \
+\
+  lapack_int n = internal::convert_index<lapack_int>(matrix.cols()), sdim, info; \
+  lapack_int matrix_order = LAPACKE_COLROW; \
+  char jobvs, sort='N'; \
+  LAPACK_##LAPACKE_PREFIX_U##_SELECT2 select = 0; \
+  jobvs = (computeU) ? 'V' : 'N'; \
+  m_matU.resize(n, n); \
+  lapack_int ldvs  = internal::convert_index<lapack_int>(m_matU.outerStride()); \
+  m_matT = matrix; \
+  lapack_int lda = internal::convert_index<lapack_int>(m_matT.outerStride()); \
+  Matrix<EIGTYPE, Dynamic, Dynamic> wr, wi; \
+  wr.resize(n, 1); wi.resize(n, 1); \
+  info = LAPACKE_##LAPACKE_PREFIX##gees( matrix_order, jobvs, sort, select, n, (LAPACKE_TYPE*)m_matT.data(), lda, &sdim, (LAPACKE_TYPE*)wr.data(), (LAPACKE_TYPE*)wi.data(), (LAPACKE_TYPE*)m_matU.data(), ldvs ); \
+  if(info == 0) \
+    m_info = Success; \
+  else \
+    m_info = NoConvergence; \
+\
+  m_isInitialized = true; \
+  m_matUisUptodate = computeU; \
+  return *this; \
+\
+}
+
+EIGEN_LAPACKE_SCHUR_REAL(double,   double, d, D, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SCHUR_REAL(float,    float,  s, S, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SCHUR_REAL(double,   double, d, D, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_SCHUR_REAL(float,    float,  s, S, RowMajor, LAPACK_ROW_MAJOR)
+
+} // end namespace Eigen
+
+#endif // EIGEN_REAL_SCHUR_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
new file mode 100644
index 0000000..9ddd553
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h
@@ -0,0 +1,870 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SELFADJOINTEIGENSOLVER_H
+#define EIGEN_SELFADJOINTEIGENSOLVER_H
+
+#include "./Tridiagonalization.h"
+
+namespace Eigen { 
+
+template<typename _MatrixType>
+class GeneralizedSelfAdjointEigenSolver;
+
+namespace internal {
+template<typename SolverType,int Size,bool IsComplex> struct direct_selfadjoint_eigenvalues;
+template<typename MatrixType, typename DiagType, typename SubDiagType>
+ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec);
+}
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class SelfAdjointEigenSolver
+  *
+  * \brief Computes eigenvalues and eigenvectors of selfadjoint matrices
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the
+  * eigendecomposition; this is expected to be an instantiation of the Matrix
+  * class template.
+  *
+  * A matrix \f$ A \f$ is selfadjoint if it equals its adjoint. For real
+  * matrices, this means that the matrix is symmetric: it equals its
+  * transpose. This class computes the eigenvalues and eigenvectors of a
+  * selfadjoint matrix. These are the scalars \f$ \lambda \f$ and vectors
+  * \f$ v \f$ such that \f$ Av = \lambda v \f$.  The eigenvalues of a
+  * selfadjoint matrix are always real. If \f$ D \f$ is a diagonal matrix with
+  * the eigenvalues on the diagonal, and \f$ V \f$ is a matrix with the
+  * eigenvectors as its columns, then \f$ A = V D V^{-1} \f$ (for selfadjoint
+  * matrices, the matrix \f$ V \f$ is always invertible). This is called the
+  * eigendecomposition.
+  *
+  * The algorithm exploits the fact that the matrix is selfadjoint, making it
+  * faster and more accurate than the general purpose eigenvalue algorithms
+  * implemented in EigenSolver and ComplexEigenSolver.
+  *
+  * Only the \b lower \b triangular \b part of the input matrix is referenced.
+  *
+  * Call the function compute() to compute the eigenvalues and eigenvectors of
+  * a given matrix. Alternatively, you can use the
+  * SelfAdjointEigenSolver(const MatrixType&, int) constructor which computes
+  * the eigenvalues and eigenvectors at construction time. Once the eigenvalue
+  * and eigenvectors are computed, they can be retrieved with the eigenvalues()
+  * and eigenvectors() functions.
+  *
+  * The documentation for SelfAdjointEigenSolver(const MatrixType&, int)
+  * contains an example of the typical use of this class.
+  *
+  * To solve the \em generalized eigenvalue problem \f$ Av = \lambda Bv \f$ and
+  * the likes, see the class GeneralizedSelfAdjointEigenSolver.
+  *
+  * \sa MatrixBase::eigenvalues(), class EigenSolver, class ComplexEigenSolver
+  */
+template<typename _MatrixType> class SelfAdjointEigenSolver
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    enum {
+      Size = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      Options = MatrixType::Options,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    
+    /** \brief Scalar type for matrices of type \p _MatrixType. */
+    typedef typename MatrixType::Scalar Scalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    
+    typedef Matrix<Scalar,Size,Size,ColMajor,MaxColsAtCompileTime,MaxColsAtCompileTime> EigenvectorsType;
+
+    /** \brief Real scalar type for \p _MatrixType.
+      *
+      * This is just \c Scalar if #Scalar is real (e.g., \c float or
+      * \c double), and the type of the real part of \c Scalar if #Scalar is
+      * complex.
+      */
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    
+    friend struct internal::direct_selfadjoint_eigenvalues<SelfAdjointEigenSolver,Size,NumTraits<Scalar>::IsComplex>;
+
+    /** \brief Type for vector of eigenvalues as returned by eigenvalues().
+      *
+      * This is a column vector with entries of type #RealScalar.
+      * The length of the vector is the size of \p _MatrixType.
+      */
+    typedef typename internal::plain_col_type<MatrixType, RealScalar>::type RealVectorType;
+    typedef Tridiagonalization<MatrixType> TridiagonalizationType;
+    typedef typename TridiagonalizationType::SubDiagonalType SubDiagonalType;
+
+    /** \brief Default constructor for fixed-size matrices.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute(). This constructor
+      * can only be used if \p _MatrixType is a fixed-size matrix; use
+      * SelfAdjointEigenSolver(Index) for dynamic-size matrices.
+      *
+      * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver.out
+      */
+    EIGEN_DEVICE_FUNC
+    SelfAdjointEigenSolver()
+        : m_eivec(),
+          m_eivalues(),
+          m_subdiag(),
+          m_isInitialized(false)
+    { }
+
+    /** \brief Constructor, pre-allocates memory for dynamic-size matrices.
+      *
+      * \param [in]  size  Positive integer, size of the matrix whose
+      * eigenvalues and eigenvectors will be computed.
+      *
+      * This constructor is useful for dynamic-size matrices, when the user
+      * intends to perform decompositions via compute(). The \p size
+      * parameter is only used as a hint. It is not an error to give a wrong
+      * \p size, but it may impair performance.
+      *
+      * \sa compute() for an example
+      */
+    EIGEN_DEVICE_FUNC
+    explicit SelfAdjointEigenSolver(Index size)
+        : m_eivec(size, size),
+          m_eivalues(size),
+          m_subdiag(size > 1 ? size - 1 : 1),
+          m_isInitialized(false)
+    {}
+
+    /** \brief Constructor; computes eigendecomposition of given matrix.
+      *
+      * \param[in]  matrix  Selfadjoint matrix whose eigendecomposition is to
+      *    be computed. Only the lower triangular part of the matrix is referenced.
+      * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+      *
+      * This constructor calls compute(const MatrixType&, int) to compute the
+      * eigenvalues of the matrix \p matrix. The eigenvectors are computed if
+      * \p options equals #ComputeEigenvectors.
+      *
+      * Example: \include SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_SelfAdjointEigenSolver_MatrixType.out
+      *
+      * \sa compute(const MatrixType&, int)
+      */
+    template<typename InputType>
+    EIGEN_DEVICE_FUNC
+    explicit SelfAdjointEigenSolver(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors)
+      : m_eivec(matrix.rows(), matrix.cols()),
+        m_eivalues(matrix.cols()),
+        m_subdiag(matrix.rows() > 1 ? matrix.rows() - 1 : 1),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived(), options);
+    }
+
+    /** \brief Computes eigendecomposition of given matrix.
+      *
+      * \param[in]  matrix  Selfadjoint matrix whose eigendecomposition is to
+      *    be computed. Only the lower triangular part of the matrix is referenced.
+      * \param[in]  options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+      * \returns    Reference to \c *this
+      *
+      * This function computes the eigenvalues of \p matrix.  The eigenvalues()
+      * function can be used to retrieve them.  If \p options equals #ComputeEigenvectors,
+      * then the eigenvectors are also computed and can be retrieved by
+      * calling eigenvectors().
+      *
+      * This implementation uses a symmetric QR algorithm. The matrix is first
+      * reduced to tridiagonal form using the Tridiagonalization class. The
+      * tridiagonal matrix is then brought to diagonal form with implicit
+      * symmetric QR steps with Wilkinson shift. Details can be found in
+      * Section 8.3 of Golub \& Van Loan, <i>%Matrix Computations</i>.
+      *
+      * The cost of the computation is about \f$ 9n^3 \f$ if the eigenvectors
+      * are required and \f$ 4n^3/3 \f$ if they are not required.
+      *
+      * This method reuses the memory in the SelfAdjointEigenSolver object that
+      * was allocated when the object was constructed, if the size of the
+      * matrix does not change.
+      *
+      * Example: \include SelfAdjointEigenSolver_compute_MatrixType.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_compute_MatrixType.out
+      *
+      * \sa SelfAdjointEigenSolver(const MatrixType&, int)
+      */
+    template<typename InputType>
+    EIGEN_DEVICE_FUNC
+    SelfAdjointEigenSolver& compute(const EigenBase<InputType>& matrix, int options = ComputeEigenvectors);
+    
+    /** \brief Computes eigendecomposition of given matrix using a closed-form algorithm
+      *
+      * This is a variant of compute(const MatrixType&, int options) which
+      * directly solves the underlying polynomial equation.
+      * 
+      * Currently only 2x2 and 3x3 matrices for which the sizes are known at compile time are supported (e.g., Matrix3d).
+      * 
+      * This method is usually significantly faster than the QR iterative algorithm
+      * but it might also be less accurate. It is also worth noting that
+      * for 3x3 matrices it involves trigonometric operations which are
+      * not necessarily available for all scalar types.
+      * 
+      * For the 3x3 case, we observed the following worst case relative error regarding the eigenvalues:
+      *   - double: 1e-8
+      *   - float:  1e-3
+      *
+      * \sa compute(const MatrixType&, int options)
+      */
+    EIGEN_DEVICE_FUNC
+    SelfAdjointEigenSolver& computeDirect(const MatrixType& matrix, int options = ComputeEigenvectors);
+
+    /**
+      *\brief Computes the eigen decomposition from a tridiagonal symmetric matrix
+      *
+      * \param[in] diag The vector containing the diagonal of the matrix.
+      * \param[in] subdiag The subdiagonal of the matrix.
+      * \param[in] options Can be #ComputeEigenvectors (default) or #EigenvaluesOnly.
+      * \returns Reference to \c *this
+      *
+      * This function assumes that the matrix has been reduced to tridiagonal form.
+      *
+      * \sa compute(const MatrixType&, int) for more information
+      */
+    SelfAdjointEigenSolver& computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options=ComputeEigenvectors);
+
+    /** \brief Returns the eigenvectors of given matrix.
+      *
+      * \returns  A const reference to the matrix whose columns are the eigenvectors.
+      *
+      * \pre The eigenvectors have been computed before.
+      *
+      * Column \f$ k \f$ of the returned matrix is an eigenvector corresponding
+      * to eigenvalue number \f$ k \f$ as returned by eigenvalues().  The
+      * eigenvectors are normalized to have (Euclidean) norm equal to one. If
+      * this object was used to solve the eigenproblem for the selfadjoint
+      * matrix \f$ A \f$, then the matrix returned by this function is the
+      * matrix \f$ V \f$ in the eigendecomposition \f$ A = V D V^{-1} \f$.
+      *
+      * Example: \include SelfAdjointEigenSolver_eigenvectors.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_eigenvectors.out
+      *
+      * \sa eigenvalues()
+      */
+    EIGEN_DEVICE_FUNC
+    const EigenvectorsType& eigenvectors() const
+    {
+      eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+      eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+      return m_eivec;
+    }
+
+    /** \brief Returns the eigenvalues of given matrix.
+      *
+      * \returns A const reference to the column vector containing the eigenvalues.
+      *
+      * \pre The eigenvalues have been computed before.
+      *
+      * The eigenvalues are repeated according to their algebraic multiplicity,
+      * so there are as many eigenvalues as rows in the matrix. The eigenvalues
+      * are sorted in increasing order.
+      *
+      * Example: \include SelfAdjointEigenSolver_eigenvalues.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_eigenvalues.out
+      *
+      * \sa eigenvectors(), MatrixBase::eigenvalues()
+      */
+    EIGEN_DEVICE_FUNC
+    const RealVectorType& eigenvalues() const
+    {
+      eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+      return m_eivalues;
+    }
+
+    /** \brief Computes the positive-definite square root of the matrix.
+      *
+      * \returns the positive-definite square root of the matrix
+      *
+      * \pre The eigenvalues and eigenvectors of a positive-definite matrix
+      * have been computed before.
+      *
+      * The square root of a positive-definite matrix \f$ A \f$ is the
+      * positive-definite matrix whose square equals \f$ A \f$. This function
+      * uses the eigendecomposition \f$ A = V D V^{-1} \f$ to compute the
+      * square root as \f$ A^{1/2} = V D^{1/2} V^{-1} \f$.
+      *
+      * Example: \include SelfAdjointEigenSolver_operatorSqrt.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_operatorSqrt.out
+      *
+      * \sa operatorInverseSqrt(), <a href="unsupported/group__MatrixFunctions__Module.html">MatrixFunctions Module</a>
+      */
+    EIGEN_DEVICE_FUNC
+    MatrixType operatorSqrt() const
+    {
+      eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+      eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+      return m_eivec * m_eivalues.cwiseSqrt().asDiagonal() * m_eivec.adjoint();
+    }
+
+    /** \brief Computes the inverse square root of the matrix.
+      *
+      * \returns the inverse positive-definite square root of the matrix
+      *
+      * \pre The eigenvalues and eigenvectors of a positive-definite matrix
+      * have been computed before.
+      *
+      * This function uses the eigendecomposition \f$ A = V D V^{-1} \f$ to
+      * compute the inverse square root as \f$ V D^{-1/2} V^{-1} \f$. This is
+      * cheaper than first computing the square root with operatorSqrt() and
+      * then its inverse with MatrixBase::inverse().
+      *
+      * Example: \include SelfAdjointEigenSolver_operatorInverseSqrt.cpp
+      * Output: \verbinclude SelfAdjointEigenSolver_operatorInverseSqrt.out
+      *
+      * \sa operatorSqrt(), MatrixBase::inverse(), <a href="unsupported/group__MatrixFunctions__Module.html">MatrixFunctions Module</a>
+      */
+    EIGEN_DEVICE_FUNC
+    MatrixType operatorInverseSqrt() const
+    {
+      eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+      eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
+      return m_eivec * m_eivalues.cwiseInverse().cwiseSqrt().asDiagonal() * m_eivec.adjoint();
+    }
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful, \c NoConvergence otherwise.
+      */
+    EIGEN_DEVICE_FUNC
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
+      return m_info;
+    }
+
+    /** \brief Maximum number of iterations.
+      *
+      * The algorithm terminates if it does not converge within m_maxIterations * n iterations, where n
+      * denotes the size of the matrix. This value is currently set to 30 (copied from LAPACK).
+      */
+    static const int m_maxIterations = 30;
+
+  protected:
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
+    EigenvectorsType m_eivec;
+    RealVectorType m_eivalues;
+    typename TridiagonalizationType::SubDiagonalType m_subdiag;
+    ComputationInfo m_info;
+    bool m_isInitialized;
+    bool m_eigenvectorsOk;
+};
+
+namespace internal {
+/** \internal
+  *
+  * \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  * Performs a QR step on a tridiagonal symmetric matrix represented as a
+  * pair of two vectors \a diag and \a subdiag.
+  *
+  * \param diag the diagonal part of the input selfadjoint tridiagonal matrix
+  * \param subdiag the sub-diagonal part of the input selfadjoint tridiagonal matrix
+  * \param start starting index of the submatrix to work on
+  * \param end last+1 index of the submatrix to work on
+  * \param matrixQ pointer to the column-major matrix holding the eigenvectors, can be 0
+  * \param n size of the input matrix
+  *
+  * For compilation efficiency reasons, this procedure does not use eigen expression
+  * for its arguments.
+  *
+  * Implemented from Golub's "Matrix Computations", algorithm 8.3.2:
+  * "implicit symmetric QR step with Wilkinson shift"
+  */
+template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC
+static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n);
+}
+
+template<typename MatrixType>
+template<typename InputType>
+EIGEN_DEVICE_FUNC
+SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
+::compute(const EigenBase<InputType>& a_matrix, int options)
+{
+  check_template_parameters();
+  
+  const InputType &matrix(a_matrix.derived());
+  
+  using std::abs;
+  eigen_assert(matrix.cols() == matrix.rows());
+  eigen_assert((options&~(EigVecMask|GenEigMask))==0
+          && (options&EigVecMask)!=EigVecMask
+          && "invalid option parameter");
+  bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
+  Index n = matrix.cols();
+  m_eivalues.resize(n,1);
+
+  if(n==1)
+  {
+    m_eivec = matrix;
+    m_eivalues.coeffRef(0,0) = numext::real(m_eivec.coeff(0,0));
+    if(computeEigenvectors)
+      m_eivec.setOnes(n,n);
+    m_info = Success;
+    m_isInitialized = true;
+    m_eigenvectorsOk = computeEigenvectors;
+    return *this;
+  }
+
+  // declare some aliases
+  RealVectorType& diag = m_eivalues;
+  EigenvectorsType& mat = m_eivec;
+
+  // map the matrix coefficients to [-1:1] to avoid over- and underflow.
+  mat = matrix.template triangularView<Lower>();
+  RealScalar scale = mat.cwiseAbs().maxCoeff();
+  if(scale==RealScalar(0)) scale = RealScalar(1);
+  mat.template triangularView<Lower>() /= scale;
+  m_subdiag.resize(n-1);
+  internal::tridiagonalization_inplace(mat, diag, m_subdiag, computeEigenvectors);
+
+  m_info = internal::computeFromTridiagonal_impl(diag, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec);
+  
+  // scale back the eigen values
+  m_eivalues *= scale;
+
+  m_isInitialized = true;
+  m_eigenvectorsOk = computeEigenvectors;
+  return *this;
+}
+
+template<typename MatrixType>
+SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
+::computeFromTridiagonal(const RealVectorType& diag, const SubDiagonalType& subdiag , int options)
+{
+  //TODO : Add an option to scale the values beforehand
+  bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
+
+  m_eivalues = diag;
+  m_subdiag = subdiag;
+  if (computeEigenvectors)
+  {
+    m_eivec.setIdentity(diag.size(), diag.size());
+  }
+  m_info = internal::computeFromTridiagonal_impl(m_eivalues, m_subdiag, m_maxIterations, computeEigenvectors, m_eivec);
+
+  m_isInitialized = true;
+  m_eigenvectorsOk = computeEigenvectors;
+  return *this;
+}
+
+namespace internal {
+/**
+  * \internal
+  * \brief Compute the eigendecomposition from a tridiagonal matrix
+  *
+  * \param[in,out] diag : On input, the diagonal of the matrix, on output the eigenvalues
+  * \param[in,out] subdiag : The subdiagonal part of the matrix (entries are modified during the decomposition)
+  * \param[in] maxIterations : the maximum number of iterations
+  * \param[in] computeEigenvectors : whether the eigenvectors have to be computed or not
+  * \param[out] eivec : The matrix to store the eigenvectors if computeEigenvectors==true. Must be allocated on input.
+  * \returns \c Success or \c NoConvergence
+  */
+template<typename MatrixType, typename DiagType, typename SubDiagType>
+ComputationInfo computeFromTridiagonal_impl(DiagType& diag, SubDiagType& subdiag, const Index maxIterations, bool computeEigenvectors, MatrixType& eivec)
+{
+  using std::abs;
+
+  ComputationInfo info;
+  typedef typename MatrixType::Scalar Scalar;
+
+  Index n = diag.size();
+  Index end = n-1;
+  Index start = 0;
+  Index iter = 0; // total number of iterations
+  
+  typedef typename DiagType::RealScalar RealScalar;
+  const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+  const RealScalar precision = RealScalar(2)*NumTraits<RealScalar>::epsilon();
+  
+  while (end>0)
+  {
+    for (Index i = start; i<end; ++i)
+      if (internal::isMuchSmallerThan(abs(subdiag[i]),(abs(diag[i])+abs(diag[i+1])),precision) || abs(subdiag[i]) <= considerAsZero)
+        subdiag[i] = 0;
+
+    // find the largest unreduced block
+    while (end>0 && subdiag[end-1]==RealScalar(0))
+    {
+      end--;
+    }
+    if (end<=0)
+      break;
+
+    // if we spent too many iterations, we give up
+    iter++;
+    if(iter > maxIterations * n) break;
+
+    start = end - 1;
+    while (start>0 && subdiag[start-1]!=0)
+      start--;
+
+    internal::tridiagonal_qr_step<MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor>(diag.data(), subdiag.data(), start, end, computeEigenvectors ? eivec.data() : (Scalar*)0, n);
+  }
+  if (iter <= maxIterations * n)
+    info = Success;
+  else
+    info = NoConvergence;
+
+  // Sort eigenvalues and corresponding vectors.
+  // TODO make the sort optional ?
+  // TODO use a better sort algorithm !!
+  if (info == Success)
+  {
+    for (Index i = 0; i < n-1; ++i)
+    {
+      Index k;
+      diag.segment(i,n-i).minCoeff(&k);
+      if (k > 0)
+      {
+        std::swap(diag[i], diag[k+i]);
+        if(computeEigenvectors)
+          eivec.col(i).swap(eivec.col(k+i));
+      }
+    }
+  }
+  return info;
+}
+  
+template<typename SolverType,int Size,bool IsComplex> struct direct_selfadjoint_eigenvalues
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(SolverType& eig, const typename SolverType::MatrixType& A, int options)
+  { eig.compute(A,options); }
+};
+
+template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3,false>
+{
+  typedef typename SolverType::MatrixType MatrixType;
+  typedef typename SolverType::RealVectorType VectorType;
+  typedef typename SolverType::Scalar Scalar;
+  typedef typename SolverType::EigenvectorsType EigenvectorsType;
+  
+
+  /** \internal
+   * Computes the roots of the characteristic polynomial of \a m.
+   * For numerical stability m.trace() should be near zero and to avoid over- or underflow m should be normalized.
+   */
+  EIGEN_DEVICE_FUNC
+  static inline void computeRoots(const MatrixType& m, VectorType& roots)
+  {
+    EIGEN_USING_STD_MATH(sqrt)
+    EIGEN_USING_STD_MATH(atan2)
+    EIGEN_USING_STD_MATH(cos)
+    EIGEN_USING_STD_MATH(sin)
+    const Scalar s_inv3 = Scalar(1)/Scalar(3);
+    const Scalar s_sqrt3 = sqrt(Scalar(3));
+
+    // The characteristic equation is x^3 - c2*x^2 + c1*x - c0 = 0.  The
+    // eigenvalues are the roots to this equation, all guaranteed to be
+    // real-valued, because the matrix is symmetric.
+    Scalar c0 = m(0,0)*m(1,1)*m(2,2) + Scalar(2)*m(1,0)*m(2,0)*m(2,1) - m(0,0)*m(2,1)*m(2,1) - m(1,1)*m(2,0)*m(2,0) - m(2,2)*m(1,0)*m(1,0);
+    Scalar c1 = m(0,0)*m(1,1) - m(1,0)*m(1,0) + m(0,0)*m(2,2) - m(2,0)*m(2,0) + m(1,1)*m(2,2) - m(2,1)*m(2,1);
+    Scalar c2 = m(0,0) + m(1,1) + m(2,2);
+
+    // Construct the parameters used in classifying the roots of the equation
+    // and in solving the equation for the roots in closed form.
+    Scalar c2_over_3 = c2*s_inv3;
+    Scalar a_over_3 = (c2*c2_over_3 - c1)*s_inv3;
+    a_over_3 = numext::maxi(a_over_3, Scalar(0));
+
+    Scalar half_b = Scalar(0.5)*(c0 + c2_over_3*(Scalar(2)*c2_over_3*c2_over_3 - c1));
+
+    Scalar q = a_over_3*a_over_3*a_over_3 - half_b*half_b;
+    q = numext::maxi(q, Scalar(0));
+
+    // Compute the eigenvalues by solving for the roots of the polynomial.
+    Scalar rho = sqrt(a_over_3);
+    Scalar theta = atan2(sqrt(q),half_b)*s_inv3;  // since sqrt(q) > 0, atan2 is in [0, pi] and theta is in [0, pi/3]
+    Scalar cos_theta = cos(theta);
+    Scalar sin_theta = sin(theta);
+    // roots are already sorted, since cos is monotonically decreasing on [0, pi]
+    roots(0) = c2_over_3 - rho*(cos_theta + s_sqrt3*sin_theta); // == 2*rho*cos(theta+2pi/3)
+    roots(1) = c2_over_3 - rho*(cos_theta - s_sqrt3*sin_theta); // == 2*rho*cos(theta+ pi/3)
+    roots(2) = c2_over_3 + Scalar(2)*rho*cos_theta;
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline bool extract_kernel(MatrixType& mat, Ref<VectorType> res, Ref<VectorType> representative)
+  {
+    using std::abs;
+    Index i0;
+    // Find non-zero column i0 (by construction, there must exist a non zero coefficient on the diagonal):
+    mat.diagonal().cwiseAbs().maxCoeff(&i0);
+    // mat.col(i0) is a good candidate for an orthogonal vector to the current eigenvector,
+    // so let's save it:
+    representative = mat.col(i0);
+    Scalar n0, n1;
+    VectorType c0, c1;
+    n0 = (c0 = representative.cross(mat.col((i0+1)%3))).squaredNorm();
+    n1 = (c1 = representative.cross(mat.col((i0+2)%3))).squaredNorm();
+    if(n0>n1) res = c0/std::sqrt(n0);
+    else      res = c1/std::sqrt(n1);
+
+    return true;
+  }
+
+  EIGEN_DEVICE_FUNC
+  static inline void run(SolverType& solver, const MatrixType& mat, int options)
+  {
+    eigen_assert(mat.cols() == 3 && mat.cols() == mat.rows());
+    eigen_assert((options&~(EigVecMask|GenEigMask))==0
+            && (options&EigVecMask)!=EigVecMask
+            && "invalid option parameter");
+    bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
+    
+    EigenvectorsType& eivecs = solver.m_eivec;
+    VectorType& eivals = solver.m_eivalues;
+  
+    // Shift the matrix to the mean eigenvalue and map the matrix coefficients to [-1:1] to avoid over- and underflow.
+    Scalar shift = mat.trace() / Scalar(3);
+    // TODO Avoid this copy. Currently it is necessary to suppress bogus values when determining maxCoeff and for computing the eigenvectors later
+    MatrixType scaledMat = mat.template selfadjointView<Lower>();
+    scaledMat.diagonal().array() -= shift;
+    Scalar scale = scaledMat.cwiseAbs().maxCoeff();
+    if(scale > 0) scaledMat /= scale;   // TODO for scale==0 we could save the remaining operations
+
+    // compute the eigenvalues
+    computeRoots(scaledMat,eivals);
+
+    // compute the eigenvectors
+    if(computeEigenvectors)
+    {
+      if((eivals(2)-eivals(0))<=Eigen::NumTraits<Scalar>::epsilon())
+      {
+        // All three eigenvalues are numerically the same
+        eivecs.setIdentity();
+      }
+      else
+      {
+        MatrixType tmp;
+        tmp = scaledMat;
+
+        // Compute the eigenvector of the most distinct eigenvalue
+        Scalar d0 = eivals(2) - eivals(1);
+        Scalar d1 = eivals(1) - eivals(0);
+        Index k(0), l(2);
+        if(d0 > d1)
+        {
+          numext::swap(k,l);
+          d0 = d1;
+        }
+
+        // Compute the eigenvector of index k
+        {
+          tmp.diagonal().array () -= eivals(k);
+          // By construction, 'tmp' is of rank 2, and its kernel corresponds to the respective eigenvector.
+          extract_kernel(tmp, eivecs.col(k), eivecs.col(l));
+        }
+
+        // Compute eigenvector of index l
+        if(d0<=2*Eigen::NumTraits<Scalar>::epsilon()*d1)
+        {
+          // If d0 is too small, then the two other eigenvalues are numerically the same,
+          // and thus we only have to ortho-normalize the near orthogonal vector we saved above.
+          eivecs.col(l) -= eivecs.col(k).dot(eivecs.col(l))*eivecs.col(l);
+          eivecs.col(l).normalize();
+        }
+        else
+        {
+          tmp = scaledMat;
+          tmp.diagonal().array () -= eivals(l);
+
+          VectorType dummy;
+          extract_kernel(tmp, eivecs.col(l), dummy);
+        }
+
+        // Compute last eigenvector from the other two
+        eivecs.col(1) = eivecs.col(2).cross(eivecs.col(0)).normalized();
+      }
+    }
+
+    // Rescale back to the original size.
+    eivals *= scale;
+    eivals.array() += shift;
+    
+    solver.m_info = Success;
+    solver.m_isInitialized = true;
+    solver.m_eigenvectorsOk = computeEigenvectors;
+  }
+};
+
+// 2x2 direct eigenvalues decomposition, code from Hauke Heibel
+template<typename SolverType> 
+struct direct_selfadjoint_eigenvalues<SolverType,2,false>
+{
+  typedef typename SolverType::MatrixType MatrixType;
+  typedef typename SolverType::RealVectorType VectorType;
+  typedef typename SolverType::Scalar Scalar;
+  typedef typename SolverType::EigenvectorsType EigenvectorsType;
+  
+  EIGEN_DEVICE_FUNC
+  static inline void computeRoots(const MatrixType& m, VectorType& roots)
+  {
+    using std::sqrt;
+    const Scalar t0 = Scalar(0.5) * sqrt( numext::abs2(m(0,0)-m(1,1)) + Scalar(4)*numext::abs2(m(1,0)));
+    const Scalar t1 = Scalar(0.5) * (m(0,0) + m(1,1));
+    roots(0) = t1 - t0;
+    roots(1) = t1 + t0;
+  }
+  
+  EIGEN_DEVICE_FUNC
+  static inline void run(SolverType& solver, const MatrixType& mat, int options)
+  {
+    EIGEN_USING_STD_MATH(sqrt);
+    EIGEN_USING_STD_MATH(abs);
+    
+    eigen_assert(mat.cols() == 2 && mat.cols() == mat.rows());
+    eigen_assert((options&~(EigVecMask|GenEigMask))==0
+            && (options&EigVecMask)!=EigVecMask
+            && "invalid option parameter");
+    bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
+    
+    EigenvectorsType& eivecs = solver.m_eivec;
+    VectorType& eivals = solver.m_eivalues;
+  
+    // Shift the matrix to the mean eigenvalue and map the matrix coefficients to [-1:1] to avoid over- and underflow.
+    Scalar shift = mat.trace() / Scalar(2);
+    MatrixType scaledMat = mat;
+    scaledMat.coeffRef(0,1) = mat.coeff(1,0);
+    scaledMat.diagonal().array() -= shift;
+    Scalar scale = scaledMat.cwiseAbs().maxCoeff();
+    if(scale > Scalar(0))
+      scaledMat /= scale;
+
+    // Compute the eigenvalues
+    computeRoots(scaledMat,eivals);
+
+    // compute the eigen vectors
+    if(computeEigenvectors)
+    {
+      if((eivals(1)-eivals(0))<=abs(eivals(1))*Eigen::NumTraits<Scalar>::epsilon())
+      {
+        eivecs.setIdentity();
+      }
+      else
+      {
+        scaledMat.diagonal().array () -= eivals(1);
+        Scalar a2 = numext::abs2(scaledMat(0,0));
+        Scalar c2 = numext::abs2(scaledMat(1,1));
+        Scalar b2 = numext::abs2(scaledMat(1,0));
+        if(a2>c2)
+        {
+          eivecs.col(1) << -scaledMat(1,0), scaledMat(0,0);
+          eivecs.col(1) /= sqrt(a2+b2);
+        }
+        else
+        {
+          eivecs.col(1) << -scaledMat(1,1), scaledMat(1,0);
+          eivecs.col(1) /= sqrt(c2+b2);
+        }
+
+        eivecs.col(0) << eivecs.col(1).unitOrthogonal();
+      }
+    }
+
+    // Rescale back to the original size.
+    eivals *= scale;
+    eivals.array() += shift;
+
+    solver.m_info = Success;
+    solver.m_isInitialized = true;
+    solver.m_eigenvectorsOk = computeEigenvectors;
+  }
+};
+
+}
+
+template<typename MatrixType>
+EIGEN_DEVICE_FUNC
+SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
+::computeDirect(const MatrixType& matrix, int options)
+{
+  internal::direct_selfadjoint_eigenvalues<SelfAdjointEigenSolver,Size,NumTraits<Scalar>::IsComplex>::run(*this,matrix,options);
+  return *this;
+}
+
+namespace internal {
+template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+EIGEN_DEVICE_FUNC
+static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n)
+{
+  using std::abs;
+  RealScalar td = (diag[end-1] - diag[end])*RealScalar(0.5);
+  RealScalar e = subdiag[end-1];
+  // Note that thanks to scaling, e^2 or td^2 cannot overflow, however they can still
+  // underflow thus leading to inf/NaN values when using the following commented code:
+//   RealScalar e2 = numext::abs2(subdiag[end-1]);
+//   RealScalar mu = diag[end] - e2 / (td + (td>0 ? 1 : -1) * sqrt(td*td + e2));
+  // This explain the following, somewhat more complicated, version:
+  RealScalar mu = diag[end];
+  if(td==RealScalar(0))
+    mu -= abs(e);
+  else
+  {
+    RealScalar e2 = numext::abs2(subdiag[end-1]);
+    RealScalar h = numext::hypot(td,e);
+    if(e2==RealScalar(0)) mu -= (e / (td + (td>RealScalar(0) ? RealScalar(1) : RealScalar(-1)))) * (e / h);
+    else                  mu -= e2 / (td + (td>RealScalar(0) ? h : -h));
+  }
+  
+  RealScalar x = diag[start] - mu;
+  RealScalar z = subdiag[start];
+  for (Index k = start; k < end; ++k)
+  {
+    JacobiRotation<RealScalar> rot;
+    rot.makeGivens(x, z);
+
+    // do T = G' T G
+    RealScalar sdk = rot.s() * diag[k] + rot.c() * subdiag[k];
+    RealScalar dkp1 = rot.s() * subdiag[k] + rot.c() * diag[k+1];
+
+    diag[k] = rot.c() * (rot.c() * diag[k] - rot.s() * subdiag[k]) - rot.s() * (rot.c() * subdiag[k] - rot.s() * diag[k+1]);
+    diag[k+1] = rot.s() * sdk + rot.c() * dkp1;
+    subdiag[k] = rot.c() * sdk - rot.s() * dkp1;
+    
+
+    if (k > start)
+      subdiag[k - 1] = rot.c() * subdiag[k-1] - rot.s() * z;
+
+    x = subdiag[k];
+
+    if (k < end - 1)
+    {
+      z = -rot.s() * subdiag[k+1];
+      subdiag[k + 1] = rot.c() * subdiag[k+1];
+    }
+    
+    // apply the givens rotation to the unit matrix Q = Q * G
+    if (matrixQ)
+    {
+      // FIXME if StorageOrder == RowMajor this operation is not very efficient
+      Map<Matrix<Scalar,Dynamic,Dynamic,StorageOrder> > q(matrixQ,n,n);
+      q.applyOnTheRight(k,k+1,rot);
+    }
+  }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SELFADJOINTEIGENSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h
new file mode 100644
index 0000000..3891cf8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/SelfAdjointEigenSolver_LAPACKE.h
@@ -0,0 +1,90 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Self-adjoint eigenvalues/eigenvectors.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_SAEIGENSOLVER_LAPACKE_H
+#define EIGEN_SAEIGENSOLVER_LAPACKE_H
+
+namespace Eigen { 
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_EIG_SELFADJ(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_NAME, EIGCOLROW, LAPACKE_COLROW ) \
+template<> template<typename InputType> inline \
+SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
+SelfAdjointEigenSolver<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const EigenBase<InputType>& matrix, int options) \
+{ \
+  eigen_assert(matrix.cols() == matrix.rows()); \
+  eigen_assert((options&~(EigVecMask|GenEigMask))==0 \
+          && (options&EigVecMask)!=EigVecMask \
+          && "invalid option parameter"); \
+  bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors; \
+  lapack_int n = internal::convert_index<lapack_int>(matrix.cols()), lda, matrix_order, info; \
+  m_eivalues.resize(n,1); \
+  m_subdiag.resize(n-1); \
+  m_eivec = matrix; \
+\
+  if(n==1) \
+  { \
+    m_eivalues.coeffRef(0,0) = numext::real(m_eivec.coeff(0,0)); \
+    if(computeEigenvectors) m_eivec.setOnes(n,n); \
+    m_info = Success; \
+    m_isInitialized = true; \
+    m_eigenvectorsOk = computeEigenvectors; \
+    return *this; \
+  } \
+\
+  lda = internal::convert_index<lapack_int>(m_eivec.outerStride()); \
+  matrix_order=LAPACKE_COLROW; \
+  char jobz, uplo='L'/*, range='A'*/; \
+  jobz = computeEigenvectors ? 'V' : 'N'; \
+\
+  info = LAPACKE_##LAPACKE_NAME( matrix_order, jobz, uplo, n, (LAPACKE_TYPE*)m_eivec.data(), lda, (LAPACKE_RTYPE*)m_eivalues.data() ); \
+  m_info = (info==0) ? Success : NoConvergence; \
+  m_isInitialized = true; \
+  m_eigenvectorsOk = computeEigenvectors; \
+  return *this; \
+}
+
+
+EIGEN_LAPACKE_EIG_SELFADJ(double,   double,                double, dsyev, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(float,    float,                 float,  ssyev, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float,  float,  cheev, ColMajor, LAPACK_COL_MAJOR)
+
+EIGEN_LAPACKE_EIG_SELFADJ(double,   double,                double, dsyev, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(float,    float,                 float,  ssyev, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(dcomplex, lapack_complex_double, double, zheev, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_EIG_SELFADJ(scomplex, lapack_complex_float,  float,  cheev, RowMajor, LAPACK_ROW_MAJOR)
+
+} // end namespace Eigen
+
+#endif // EIGEN_SAEIGENSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/Tridiagonalization.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/Tridiagonalization.h
new file mode 100644
index 0000000..1d102c1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Eigenvalues/Tridiagonalization.h
@@ -0,0 +1,556 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Jitse Niesen <jitse@maths.leeds.ac.uk>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRIDIAGONALIZATION_H
+#define EIGEN_TRIDIAGONALIZATION_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<typename MatrixType> struct TridiagonalizationMatrixTReturnType;
+template<typename MatrixType>
+struct traits<TridiagonalizationMatrixTReturnType<MatrixType> >
+  : public traits<typename MatrixType::PlainObject>
+{
+  typedef typename MatrixType::PlainObject ReturnType; // FIXME shall it be a BandMatrix?
+  enum { Flags = 0 };
+};
+
+template<typename MatrixType, typename CoeffVectorType>
+void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs);
+}
+
+/** \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  *
+  * \class Tridiagonalization
+  *
+  * \brief Tridiagonal decomposition of a selfadjoint matrix
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the
+  * tridiagonal decomposition; this is expected to be an instantiation of the
+  * Matrix class template.
+  *
+  * This class performs a tridiagonal decomposition of a selfadjoint matrix \f$ A \f$ such that:
+  * \f$ A = Q T Q^* \f$ where \f$ Q \f$ is unitary and \f$ T \f$ a real symmetric tridiagonal matrix.
+  *
+  * A tridiagonal matrix is a matrix which has nonzero elements only on the
+  * main diagonal and the first diagonal below and above it. The Hessenberg
+  * decomposition of a selfadjoint matrix is in fact a tridiagonal
+  * decomposition. This class is used in SelfAdjointEigenSolver to compute the
+  * eigenvalues and eigenvectors of a selfadjoint matrix.
+  *
+  * Call the function compute() to compute the tridiagonal decomposition of a
+  * given matrix. Alternatively, you can use the Tridiagonalization(const MatrixType&)
+  * constructor which computes the tridiagonal Schur decomposition at
+  * construction time. Once the decomposition is computed, you can use the
+  * matrixQ() and matrixT() functions to retrieve the matrices Q and T in the
+  * decomposition.
+  *
+  * The documentation of Tridiagonalization(const MatrixType&) contains an
+  * example of the typical use of this class.
+  *
+  * \sa class HessenbergDecomposition, class SelfAdjointEigenSolver
+  */
+template<typename _MatrixType> class Tridiagonalization
+{
+  public:
+
+    /** \brief Synonym for the template parameter \p _MatrixType. */
+    typedef _MatrixType MatrixType;
+
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+
+    enum {
+      Size = MatrixType::RowsAtCompileTime,
+      SizeMinusOne = Size == Dynamic ? Dynamic : (Size > 1 ? Size - 1 : 1),
+      Options = MatrixType::Options,
+      MaxSize = MatrixType::MaxRowsAtCompileTime,
+      MaxSizeMinusOne = MaxSize == Dynamic ? Dynamic : (MaxSize > 1 ? MaxSize - 1 : 1)
+    };
+
+    typedef Matrix<Scalar, SizeMinusOne, 1, Options & ~RowMajor, MaxSizeMinusOne, 1> CoeffVectorType;
+    typedef typename internal::plain_col_type<MatrixType, RealScalar>::type DiagonalType;
+    typedef Matrix<RealScalar, SizeMinusOne, 1, Options & ~RowMajor, MaxSizeMinusOne, 1> SubDiagonalType;
+    typedef typename internal::remove_all<typename MatrixType::RealReturnType>::type MatrixTypeRealView;
+    typedef internal::TridiagonalizationMatrixTReturnType<MatrixTypeRealView> MatrixTReturnType;
+
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+              typename internal::add_const_on_value_type<typename Diagonal<const MatrixType>::RealReturnType>::type,
+              const Diagonal<const MatrixType>
+            >::type DiagonalReturnType;
+
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+              typename internal::add_const_on_value_type<typename Diagonal<const MatrixType, -1>::RealReturnType>::type,
+              const Diagonal<const MatrixType, -1>
+            >::type SubDiagonalReturnType;
+
+    /** \brief Return type of matrixQ() */
+    typedef HouseholderSequence<MatrixType,typename internal::remove_all<typename CoeffVectorType::ConjugateReturnType>::type> HouseholderSequenceType;
+
+    /** \brief Default constructor.
+      *
+      * \param [in]  size  Positive integer, size of the matrix whose tridiagonal
+      * decomposition will be computed.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via compute().  The \p size parameter is only
+      * used as a hint. It is not an error to give a wrong \p size, but it may
+      * impair performance.
+      *
+      * \sa compute() for an example.
+      */
+    explicit Tridiagonalization(Index size = Size==Dynamic ? 2 : Size)
+      : m_matrix(size,size),
+        m_hCoeffs(size > 1 ? size-1 : 1),
+        m_isInitialized(false)
+    {}
+
+    /** \brief Constructor; computes tridiagonal decomposition of given matrix.
+      *
+      * \param[in]  matrix  Selfadjoint matrix whose tridiagonal decomposition
+      * is to be computed.
+      *
+      * This constructor calls compute() to compute the tridiagonal decomposition.
+      *
+      * Example: \include Tridiagonalization_Tridiagonalization_MatrixType.cpp
+      * Output: \verbinclude Tridiagonalization_Tridiagonalization_MatrixType.out
+      */
+    template<typename InputType>
+    explicit Tridiagonalization(const EigenBase<InputType>& matrix)
+      : m_matrix(matrix.derived()),
+        m_hCoeffs(matrix.cols() > 1 ? matrix.cols()-1 : 1),
+        m_isInitialized(false)
+    {
+      internal::tridiagonalization_inplace(m_matrix, m_hCoeffs);
+      m_isInitialized = true;
+    }
+
+    /** \brief Computes tridiagonal decomposition of given matrix.
+      *
+      * \param[in]  matrix  Selfadjoint matrix whose tridiagonal decomposition
+      * is to be computed.
+      * \returns    Reference to \c *this
+      *
+      * The tridiagonal decomposition is computed by bringing the columns of
+      * the matrix successively in the required form using Householder
+      * reflections. The cost is \f$ 4n^3/3 \f$ flops, where \f$ n \f$ denotes
+      * the size of the given matrix.
+      *
+      * This method reuses of the allocated data in the Tridiagonalization
+      * object, if the size of the matrix does not change.
+      *
+      * Example: \include Tridiagonalization_compute.cpp
+      * Output: \verbinclude Tridiagonalization_compute.out
+      */
+    template<typename InputType>
+    Tridiagonalization& compute(const EigenBase<InputType>& matrix)
+    {
+      m_matrix = matrix.derived();
+      m_hCoeffs.resize(matrix.rows()-1, 1);
+      internal::tridiagonalization_inplace(m_matrix, m_hCoeffs);
+      m_isInitialized = true;
+      return *this;
+    }
+
+    /** \brief Returns the Householder coefficients.
+      *
+      * \returns a const reference to the vector of Householder coefficients
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * The Householder coefficients allow the reconstruction of the matrix
+      * \f$ Q \f$ in the tridiagonal decomposition from the packed data.
+      *
+      * Example: \include Tridiagonalization_householderCoefficients.cpp
+      * Output: \verbinclude Tridiagonalization_householderCoefficients.out
+      *
+      * \sa packedMatrix(), \ref Householder_Module "Householder module"
+      */
+    inline CoeffVectorType householderCoefficients() const
+    {
+      eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+      return m_hCoeffs;
+    }
+
+    /** \brief Returns the internal representation of the decomposition
+      *
+      *	\returns a const reference to a matrix with the internal representation
+      *	         of the decomposition.
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * The returned matrix contains the following information:
+      *  - the strict upper triangular part is equal to the input matrix A.
+      *  - the diagonal and lower sub-diagonal represent the real tridiagonal
+      *    symmetric matrix T.
+      *  - the rest of the lower part contains the Householder vectors that,
+      *    combined with Householder coefficients returned by
+      *    householderCoefficients(), allows to reconstruct the matrix Q as
+      *       \f$ Q = H_{N-1} \ldots H_1 H_0 \f$.
+      *    Here, the matrices \f$ H_i \f$ are the Householder transformations
+      *       \f$ H_i = (I - h_i v_i v_i^T) \f$
+      *    where \f$ h_i \f$ is the \f$ i \f$th Householder coefficient and
+      *    \f$ v_i \f$ is the Householder vector defined by
+      *       \f$ v_i = [ 0, \ldots, 0, 1, M(i+2,i), \ldots, M(N-1,i) ]^T \f$
+      *    with M the matrix returned by this function.
+      *
+      * See LAPACK for further details on this packed storage.
+      *
+      * Example: \include Tridiagonalization_packedMatrix.cpp
+      * Output: \verbinclude Tridiagonalization_packedMatrix.out
+      *
+      * \sa householderCoefficients()
+      */
+    inline const MatrixType& packedMatrix() const
+    {
+      eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+      return m_matrix;
+    }
+
+    /** \brief Returns the unitary matrix Q in the decomposition
+      *
+      * \returns object representing the matrix Q
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * This function returns a light-weight object of template class
+      * HouseholderSequence. You can either apply it directly to a matrix or
+      * you can convert it to a matrix of type #MatrixType.
+      *
+      * \sa Tridiagonalization(const MatrixType&) for an example,
+      *     matrixT(), class HouseholderSequence
+      */
+    HouseholderSequenceType matrixQ() const
+    {
+      eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+      return HouseholderSequenceType(m_matrix, m_hCoeffs.conjugate())
+             .setLength(m_matrix.rows() - 1)
+             .setShift(1);
+    }
+
+    /** \brief Returns an expression of the tridiagonal matrix T in the decomposition
+      *
+      * \returns expression object representing the matrix T
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * Currently, this function can be used to extract the matrix T from internal
+      * data and copy it to a dense matrix object. In most cases, it may be
+      * sufficient to directly use the packed matrix or the vector expressions
+      * returned by diagonal() and subDiagonal() instead of creating a new
+      * dense copy matrix with this function.
+      *
+      * \sa Tridiagonalization(const MatrixType&) for an example,
+      * matrixQ(), packedMatrix(), diagonal(), subDiagonal()
+      */
+    MatrixTReturnType matrixT() const
+    {
+      eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+      return MatrixTReturnType(m_matrix.real());
+    }
+
+    /** \brief Returns the diagonal of the tridiagonal matrix T in the decomposition.
+      *
+      * \returns expression representing the diagonal of T
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * Example: \include Tridiagonalization_diagonal.cpp
+      * Output: \verbinclude Tridiagonalization_diagonal.out
+      *
+      * \sa matrixT(), subDiagonal()
+      */
+    DiagonalReturnType diagonal() const;
+
+    /** \brief Returns the subdiagonal of the tridiagonal matrix T in the decomposition.
+      *
+      * \returns expression representing the subdiagonal of T
+      *
+      * \pre Either the constructor Tridiagonalization(const MatrixType&) or
+      * the member function compute(const MatrixType&) has been called before
+      * to compute the tridiagonal decomposition of a matrix.
+      *
+      * \sa diagonal() for an example, matrixT()
+      */
+    SubDiagonalReturnType subDiagonal() const;
+
+  protected:
+
+    MatrixType m_matrix;
+    CoeffVectorType m_hCoeffs;
+    bool m_isInitialized;
+};
+
+template<typename MatrixType>
+typename Tridiagonalization<MatrixType>::DiagonalReturnType
+Tridiagonalization<MatrixType>::diagonal() const
+{
+  eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+  return m_matrix.diagonal().real();
+}
+
+template<typename MatrixType>
+typename Tridiagonalization<MatrixType>::SubDiagonalReturnType
+Tridiagonalization<MatrixType>::subDiagonal() const
+{
+  eigen_assert(m_isInitialized && "Tridiagonalization is not initialized.");
+  return m_matrix.template diagonal<-1>().real();
+}
+
+namespace internal {
+
+/** \internal
+  * Performs a tridiagonal decomposition of the selfadjoint matrix \a matA in-place.
+  *
+  * \param[in,out] matA On input the selfadjoint matrix. Only the \b lower triangular part is referenced.
+  *                     On output, the strict upper part is left unchanged, and the lower triangular part
+  *                     represents the T and Q matrices in packed format has detailed below.
+  * \param[out]    hCoeffs returned Householder coefficients (see below)
+  *
+  * On output, the tridiagonal selfadjoint matrix T is stored in the diagonal
+  * and lower sub-diagonal of the matrix \a matA.
+  * The unitary matrix Q is represented in a compact way as a product of
+  * Householder reflectors \f$ H_i \f$ such that:
+  *       \f$ Q = H_{N-1} \ldots H_1 H_0 \f$.
+  * The Householder reflectors are defined as
+  *       \f$ H_i = (I - h_i v_i v_i^T) \f$
+  * where \f$ h_i = hCoeffs[i]\f$ is the \f$ i \f$th Householder coefficient and
+  * \f$ v_i \f$ is the Householder vector defined by
+  *       \f$ v_i = [ 0, \ldots, 0, 1, matA(i+2,i), \ldots, matA(N-1,i) ]^T \f$.
+  *
+  * Implemented from Golub's "Matrix Computations", algorithm 8.3.1.
+  *
+  * \sa Tridiagonalization::packedMatrix()
+  */
+template<typename MatrixType, typename CoeffVectorType>
+void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs)
+{
+  using numext::conj;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  Index n = matA.rows();
+  eigen_assert(n==matA.cols());
+  eigen_assert(n==hCoeffs.size()+1 || n==1);
+  
+  for (Index i = 0; i<n-1; ++i)
+  {
+    Index remainingSize = n-i-1;
+    RealScalar beta;
+    Scalar h;
+    matA.col(i).tail(remainingSize).makeHouseholderInPlace(h, beta);
+
+    // Apply similarity transformation to remaining columns,
+    // i.e., A = H A H' where H = I - h v v' and v = matA.col(i).tail(n-i-1)
+    matA.col(i).coeffRef(i+1) = 1;
+
+    hCoeffs.tail(n-i-1).noalias() = (matA.bottomRightCorner(remainingSize,remainingSize).template selfadjointView<Lower>()
+                                  * (conj(h) * matA.col(i).tail(remainingSize)));
+
+    hCoeffs.tail(n-i-1) += (conj(h)*RealScalar(-0.5)*(hCoeffs.tail(remainingSize).dot(matA.col(i).tail(remainingSize)))) * matA.col(i).tail(n-i-1);
+
+    matA.bottomRightCorner(remainingSize, remainingSize).template selfadjointView<Lower>()
+      .rankUpdate(matA.col(i).tail(remainingSize), hCoeffs.tail(remainingSize), Scalar(-1));
+
+    matA.col(i).coeffRef(i+1) = beta;
+    hCoeffs.coeffRef(i) = h;
+  }
+}
+
+// forward declaration, implementation at the end of this file
+template<typename MatrixType,
+         int Size=MatrixType::ColsAtCompileTime,
+         bool IsComplex=NumTraits<typename MatrixType::Scalar>::IsComplex>
+struct tridiagonalization_inplace_selector;
+
+/** \brief Performs a full tridiagonalization in place
+  *
+  * \param[in,out]  mat  On input, the selfadjoint matrix whose tridiagonal
+  *    decomposition is to be computed. Only the lower triangular part referenced.
+  *    The rest is left unchanged. On output, the orthogonal matrix Q
+  *    in the decomposition if \p extractQ is true.
+  * \param[out]  diag  The diagonal of the tridiagonal matrix T in the
+  *    decomposition.
+  * \param[out]  subdiag  The subdiagonal of the tridiagonal matrix T in
+  *    the decomposition.
+  * \param[in]  extractQ  If true, the orthogonal matrix Q in the
+  *    decomposition is computed and stored in \p mat.
+  *
+  * Computes the tridiagonal decomposition of the selfadjoint matrix \p mat in place
+  * such that \f$ mat = Q T Q^* \f$ where \f$ Q \f$ is unitary and \f$ T \f$ a real
+  * symmetric tridiagonal matrix.
+  *
+  * The tridiagonal matrix T is passed to the output parameters \p diag and \p subdiag. If
+  * \p extractQ is true, then the orthogonal matrix Q is passed to \p mat. Otherwise the lower
+  * part of the matrix \p mat is destroyed.
+  *
+  * The vectors \p diag and \p subdiag are not resized. The function
+  * assumes that they are already of the correct size. The length of the
+  * vector \p diag should equal the number of rows in \p mat, and the
+  * length of the vector \p subdiag should be one left.
+  *
+  * This implementation contains an optimized path for 3-by-3 matrices
+  * which is especially useful for plane fitting.
+  *
+  * \note Currently, it requires two temporary vectors to hold the intermediate
+  * Householder coefficients, and to reconstruct the matrix Q from the Householder
+  * reflectors.
+  *
+  * Example (this uses the same matrix as the example in
+  *    Tridiagonalization::Tridiagonalization(const MatrixType&)):
+  *    \include Tridiagonalization_decomposeInPlace.cpp
+  * Output: \verbinclude Tridiagonalization_decomposeInPlace.out
+  *
+  * \sa class Tridiagonalization
+  */
+template<typename MatrixType, typename DiagonalType, typename SubDiagonalType>
+void tridiagonalization_inplace(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
+{
+  eigen_assert(mat.cols()==mat.rows() && diag.size()==mat.rows() && subdiag.size()==mat.rows()-1);
+  tridiagonalization_inplace_selector<MatrixType>::run(mat, diag, subdiag, extractQ);
+}
+
+/** \internal
+  * General full tridiagonalization
+  */
+template<typename MatrixType, int Size, bool IsComplex>
+struct tridiagonalization_inplace_selector
+{
+  typedef typename Tridiagonalization<MatrixType>::CoeffVectorType CoeffVectorType;
+  typedef typename Tridiagonalization<MatrixType>::HouseholderSequenceType HouseholderSequenceType;
+  template<typename DiagonalType, typename SubDiagonalType>
+  static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
+  {
+    CoeffVectorType hCoeffs(mat.cols()-1);
+    tridiagonalization_inplace(mat,hCoeffs);
+    diag = mat.diagonal().real();
+    subdiag = mat.template diagonal<-1>().real();
+    if(extractQ)
+      mat = HouseholderSequenceType(mat, hCoeffs.conjugate())
+            .setLength(mat.rows() - 1)
+            .setShift(1);
+  }
+};
+
+/** \internal
+  * Specialization for 3x3 real matrices.
+  * Especially useful for plane fitting.
+  */
+template<typename MatrixType>
+struct tridiagonalization_inplace_selector<MatrixType,3,false>
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+
+  template<typename DiagonalType, typename SubDiagonalType>
+  static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType& subdiag, bool extractQ)
+  {
+    using std::sqrt;
+    const RealScalar tol = (std::numeric_limits<RealScalar>::min)();
+    diag[0] = mat(0,0);
+    RealScalar v1norm2 = numext::abs2(mat(2,0));
+    if(v1norm2 <= tol)
+    {
+      diag[1] = mat(1,1);
+      diag[2] = mat(2,2);
+      subdiag[0] = mat(1,0);
+      subdiag[1] = mat(2,1);
+      if (extractQ)
+        mat.setIdentity();
+    }
+    else
+    {
+      RealScalar beta = sqrt(numext::abs2(mat(1,0)) + v1norm2);
+      RealScalar invBeta = RealScalar(1)/beta;
+      Scalar m01 = mat(1,0) * invBeta;
+      Scalar m02 = mat(2,0) * invBeta;
+      Scalar q = RealScalar(2)*m01*mat(2,1) + m02*(mat(2,2) - mat(1,1));
+      diag[1] = mat(1,1) + m02*q;
+      diag[2] = mat(2,2) - m02*q;
+      subdiag[0] = beta;
+      subdiag[1] = mat(2,1) - m01 * q;
+      if (extractQ)
+      {
+        mat << 1,   0,    0,
+               0, m01,  m02,
+               0, m02, -m01;
+      }
+    }
+  }
+};
+
+/** \internal
+  * Trivial specialization for 1x1 matrices
+  */
+template<typename MatrixType, bool IsComplex>
+struct tridiagonalization_inplace_selector<MatrixType,1,IsComplex>
+{
+  typedef typename MatrixType::Scalar Scalar;
+
+  template<typename DiagonalType, typename SubDiagonalType>
+  static void run(MatrixType& mat, DiagonalType& diag, SubDiagonalType&, bool extractQ)
+  {
+    diag(0,0) = numext::real(mat(0,0));
+    if(extractQ)
+      mat(0,0) = Scalar(1);
+  }
+};
+
+/** \internal
+  * \eigenvalues_module \ingroup Eigenvalues_Module
+  *
+  * \brief Expression type for return value of Tridiagonalization::matrixT()
+  *
+  * \tparam MatrixType type of underlying dense matrix
+  */
+template<typename MatrixType> struct TridiagonalizationMatrixTReturnType
+: public ReturnByValue<TridiagonalizationMatrixTReturnType<MatrixType> >
+{
+  public:
+    /** \brief Constructor.
+      *
+      * \param[in] mat The underlying dense matrix
+      */
+    TridiagonalizationMatrixTReturnType(const MatrixType& mat) : m_matrix(mat) { }
+
+    template <typename ResultType>
+    inline void evalTo(ResultType& result) const
+    {
+      result.setZero();
+      result.template diagonal<1>() = m_matrix.template diagonal<-1>().conjugate();
+      result.diagonal() = m_matrix.diagonal();
+      result.template diagonal<-1>() = m_matrix.template diagonal<-1>();
+    }
+
+    Index rows() const { return m_matrix.rows(); }
+    Index cols() const { return m_matrix.cols(); }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRIDIAGONALIZATION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AlignedBox.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AlignedBox.h
new file mode 100644
index 0000000..066eae4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AlignedBox.h
@@ -0,0 +1,392 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ALIGNEDBOX_H
+#define EIGEN_ALIGNEDBOX_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  *
+  * \class AlignedBox
+  *
+  * \brief An axis aligned box
+  *
+  * \tparam _Scalar the type of the scalar coefficients
+  * \tparam _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic.
+  *
+  * This class represents an axis aligned box as a pair of the minimal and maximal corners.
+  * \warning The result of most methods is undefined when applied to an empty box. You can check for empty boxes using isEmpty().
+  * \sa alignedboxtypedefs
+  */
+template <typename _Scalar, int _AmbientDim>
+class AlignedBox
+{
+public:
+EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
+  enum { AmbientDimAtCompileTime = _AmbientDim };
+  typedef _Scalar                                   Scalar;
+  typedef NumTraits<Scalar>                         ScalarTraits;
+  typedef Eigen::Index                              Index; ///< \deprecated since Eigen 3.3
+  typedef typename ScalarTraits::Real               RealScalar;
+  typedef typename ScalarTraits::NonInteger         NonInteger;
+  typedef Matrix<Scalar,AmbientDimAtCompileTime,1>  VectorType;
+  typedef CwiseBinaryOp<internal::scalar_sum_op<Scalar>, const VectorType, const VectorType> VectorTypeSum;
+
+  /** Define constants to name the corners of a 1D, 2D or 3D axis aligned bounding box */
+  enum CornerType
+  {
+    /** 1D names @{ */
+    Min=0, Max=1,
+    /** @} */
+
+    /** Identifier for 2D corner @{ */
+    BottomLeft=0, BottomRight=1,
+    TopLeft=2, TopRight=3,
+    /** @} */
+
+    /** Identifier for 3D corner  @{ */
+    BottomLeftFloor=0, BottomRightFloor=1,
+    TopLeftFloor=2, TopRightFloor=3,
+    BottomLeftCeil=4, BottomRightCeil=5,
+    TopLeftCeil=6, TopRightCeil=7
+    /** @} */
+  };
+
+
+  /** Default constructor initializing a null box. */
+  EIGEN_DEVICE_FUNC inline AlignedBox()
+  { if (AmbientDimAtCompileTime!=Dynamic) setEmpty(); }
+
+  /** Constructs a null box with \a _dim the dimension of the ambient space. */
+  EIGEN_DEVICE_FUNC inline explicit AlignedBox(Index _dim) : m_min(_dim), m_max(_dim)
+  { setEmpty(); }
+
+  /** Constructs a box with extremities \a _min and \a _max.
+   * \warning If either component of \a _min is larger than the same component of \a _max, the constructed box is empty. */
+  template<typename OtherVectorType1, typename OtherVectorType2>
+  EIGEN_DEVICE_FUNC inline AlignedBox(const OtherVectorType1& _min, const OtherVectorType2& _max) : m_min(_min), m_max(_max) {}
+
+  /** Constructs a box containing a single point \a p. */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline explicit AlignedBox(const MatrixBase<Derived>& p) : m_min(p), m_max(m_min)
+  { }
+
+  EIGEN_DEVICE_FUNC ~AlignedBox() {}
+
+  /** \returns the dimension in which the box holds */
+  EIGEN_DEVICE_FUNC inline Index dim() const { return AmbientDimAtCompileTime==Dynamic ? m_min.size() : Index(AmbientDimAtCompileTime); }
+
+  /** \deprecated use isEmpty() */
+  EIGEN_DEVICE_FUNC inline bool isNull() const { return isEmpty(); }
+
+  /** \deprecated use setEmpty() */
+  EIGEN_DEVICE_FUNC inline void setNull() { setEmpty(); }
+
+  /** \returns true if the box is empty.
+   * \sa setEmpty */
+  EIGEN_DEVICE_FUNC inline bool isEmpty() const { return (m_min.array() > m_max.array()).any(); }
+
+  /** Makes \c *this an empty box.
+   * \sa isEmpty */
+  EIGEN_DEVICE_FUNC inline void setEmpty()
+  {
+    m_min.setConstant( ScalarTraits::highest() );
+    m_max.setConstant( ScalarTraits::lowest() );
+  }
+
+  /** \returns the minimal corner */
+  EIGEN_DEVICE_FUNC inline const VectorType& (min)() const { return m_min; }
+  /** \returns a non const reference to the minimal corner */
+  EIGEN_DEVICE_FUNC inline VectorType& (min)() { return m_min; }
+  /** \returns the maximal corner */
+  EIGEN_DEVICE_FUNC inline const VectorType& (max)() const { return m_max; }
+  /** \returns a non const reference to the maximal corner */
+  EIGEN_DEVICE_FUNC inline VectorType& (max)() { return m_max; }
+
+  /** \returns the center of the box */
+  EIGEN_DEVICE_FUNC inline const EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(VectorTypeSum, RealScalar, quotient)
+  center() const
+  { return (m_min+m_max)/RealScalar(2); }
+
+  /** \returns the lengths of the sides of the bounding box.
+    * Note that this function does not get the same
+    * result for integral or floating scalar types: see
+    */
+  EIGEN_DEVICE_FUNC inline const CwiseBinaryOp< internal::scalar_difference_op<Scalar,Scalar>, const VectorType, const VectorType> sizes() const
+  { return m_max - m_min; }
+
+  /** \returns the volume of the bounding box */
+  EIGEN_DEVICE_FUNC inline Scalar volume() const
+  { return sizes().prod(); }
+
+  /** \returns an expression for the bounding box diagonal vector
+    * if the length of the diagonal is needed: diagonal().norm()
+    * will provide it.
+    */
+  EIGEN_DEVICE_FUNC inline CwiseBinaryOp< internal::scalar_difference_op<Scalar,Scalar>, const VectorType, const VectorType> diagonal() const
+  { return sizes(); }
+
+  /** \returns the vertex of the bounding box at the corner defined by
+    * the corner-id corner. It works only for a 1D, 2D or 3D bounding box.
+    * For 1D bounding boxes corners are named by 2 enum constants:
+    * BottomLeft and BottomRight.
+    * For 2D bounding boxes, corners are named by 4 enum constants:
+    * BottomLeft, BottomRight, TopLeft, TopRight.
+    * For 3D bounding boxes, the following names are added:
+    * BottomLeftCeil, BottomRightCeil, TopLeftCeil, TopRightCeil.
+    */
+  EIGEN_DEVICE_FUNC inline VectorType corner(CornerType corner) const
+  {
+    EIGEN_STATIC_ASSERT(_AmbientDim <= 3, THIS_METHOD_IS_ONLY_FOR_VECTORS_OF_A_SPECIFIC_SIZE);
+
+    VectorType res;
+
+    Index mult = 1;
+    for(Index d=0; d<dim(); ++d)
+    {
+      if( mult & corner ) res[d] = m_max[d];
+      else                res[d] = m_min[d];
+      mult *= 2;
+    }
+    return res;
+  }
+
+  /** \returns a random point inside the bounding box sampled with
+   * a uniform distribution */
+  EIGEN_DEVICE_FUNC inline VectorType sample() const
+  {
+    VectorType r(dim());
+    for(Index d=0; d<dim(); ++d)
+    {
+      if(!ScalarTraits::IsInteger)
+      {
+        r[d] = m_min[d] + (m_max[d]-m_min[d])
+             * internal::random<Scalar>(Scalar(0), Scalar(1));
+      }
+      else
+        r[d] = internal::random(m_min[d], m_max[d]);
+    }
+    return r;
+  }
+
+  /** \returns true if the point \a p is inside the box \c *this. */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline bool contains(const MatrixBase<Derived>& p) const
+  {
+    typename internal::nested_eval<Derived,2>::type p_n(p.derived());
+    return (m_min.array()<=p_n.array()).all() && (p_n.array()<=m_max.array()).all();
+  }
+
+  /** \returns true if the box \a b is entirely inside the box \c *this. */
+  EIGEN_DEVICE_FUNC inline bool contains(const AlignedBox& b) const
+  { return (m_min.array()<=(b.min)().array()).all() && ((b.max)().array()<=m_max.array()).all(); }
+
+  /** \returns true if the box \a b is intersecting the box \c *this.
+   * \sa intersection, clamp */
+  EIGEN_DEVICE_FUNC inline bool intersects(const AlignedBox& b) const
+  { return (m_min.array()<=(b.max)().array()).all() && ((b.min)().array()<=m_max.array()).all(); }
+
+  /** Extends \c *this such that it contains the point \a p and returns a reference to \c *this.
+   * \sa extend(const AlignedBox&) */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline AlignedBox& extend(const MatrixBase<Derived>& p)
+  {
+    typename internal::nested_eval<Derived,2>::type p_n(p.derived());
+    m_min = m_min.cwiseMin(p_n);
+    m_max = m_max.cwiseMax(p_n);
+    return *this;
+  }
+
+  /** Extends \c *this such that it contains the box \a b and returns a reference to \c *this.
+   * \sa merged, extend(const MatrixBase&) */
+  EIGEN_DEVICE_FUNC inline AlignedBox& extend(const AlignedBox& b)
+  {
+    m_min = m_min.cwiseMin(b.m_min);
+    m_max = m_max.cwiseMax(b.m_max);
+    return *this;
+  }
+
+  /** Clamps \c *this by the box \a b and returns a reference to \c *this.
+   * \note If the boxes don't intersect, the resulting box is empty.
+   * \sa intersection(), intersects() */
+  EIGEN_DEVICE_FUNC inline AlignedBox& clamp(const AlignedBox& b)
+  {
+    m_min = m_min.cwiseMax(b.m_min);
+    m_max = m_max.cwiseMin(b.m_max);
+    return *this;
+  }
+
+  /** Returns an AlignedBox that is the intersection of \a b and \c *this
+   * \note If the boxes don't intersect, the resulting box is empty.
+   * \sa intersects(), clamp, contains()  */
+  EIGEN_DEVICE_FUNC inline AlignedBox intersection(const AlignedBox& b) const
+  {return AlignedBox(m_min.cwiseMax(b.m_min), m_max.cwiseMin(b.m_max)); }
+
+  /** Returns an AlignedBox that is the union of \a b and \c *this.
+   * \note Merging with an empty box may result in a box bigger than \c *this. 
+   * \sa extend(const AlignedBox&) */
+  EIGEN_DEVICE_FUNC inline AlignedBox merged(const AlignedBox& b) const
+  { return AlignedBox(m_min.cwiseMin(b.m_min), m_max.cwiseMax(b.m_max)); }
+
+  /** Translate \c *this by the vector \a t and returns a reference to \c *this. */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline AlignedBox& translate(const MatrixBase<Derived>& a_t)
+  {
+    const typename internal::nested_eval<Derived,2>::type t(a_t.derived());
+    m_min += t;
+    m_max += t;
+    return *this;
+  }
+
+  /** \returns the squared distance between the point \a p and the box \c *this,
+    * and zero if \a p is inside the box.
+    * \sa exteriorDistance(const MatrixBase&), squaredExteriorDistance(const AlignedBox&)
+    */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline Scalar squaredExteriorDistance(const MatrixBase<Derived>& p) const;
+
+  /** \returns the squared distance between the boxes \a b and \c *this,
+    * and zero if the boxes intersect.
+    * \sa exteriorDistance(const AlignedBox&), squaredExteriorDistance(const MatrixBase&)
+    */
+  EIGEN_DEVICE_FUNC inline Scalar squaredExteriorDistance(const AlignedBox& b) const;
+
+  /** \returns the distance between the point \a p and the box \c *this,
+    * and zero if \a p is inside the box.
+    * \sa squaredExteriorDistance(const MatrixBase&), exteriorDistance(const AlignedBox&)
+    */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline NonInteger exteriorDistance(const MatrixBase<Derived>& p) const
+  { EIGEN_USING_STD_MATH(sqrt) return sqrt(NonInteger(squaredExteriorDistance(p))); }
+
+  /** \returns the distance between the boxes \a b and \c *this,
+    * and zero if the boxes intersect.
+    * \sa squaredExteriorDistance(const AlignedBox&), exteriorDistance(const MatrixBase&)
+    */
+  EIGEN_DEVICE_FUNC inline NonInteger exteriorDistance(const AlignedBox& b) const
+  { EIGEN_USING_STD_MATH(sqrt) return sqrt(NonInteger(squaredExteriorDistance(b))); }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<AlignedBox,
+           AlignedBox<NewScalarType,AmbientDimAtCompileTime> >::type cast() const
+  {
+    return typename internal::cast_return_type<AlignedBox,
+                    AlignedBox<NewScalarType,AmbientDimAtCompileTime> >::type(*this);
+  }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  EIGEN_DEVICE_FUNC inline explicit AlignedBox(const AlignedBox<OtherScalarType,AmbientDimAtCompileTime>& other)
+  {
+    m_min = (other.min)().template cast<Scalar>();
+    m_max = (other.max)().template cast<Scalar>();
+  }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const AlignedBox& other, const RealScalar& prec = ScalarTraits::dummy_precision()) const
+  { return m_min.isApprox(other.m_min, prec) && m_max.isApprox(other.m_max, prec); }
+
+protected:
+
+  VectorType m_min, m_max;
+};
+
+
+
+template<typename Scalar,int AmbientDim>
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline Scalar AlignedBox<Scalar,AmbientDim>::squaredExteriorDistance(const MatrixBase<Derived>& a_p) const
+{
+  typename internal::nested_eval<Derived,2*AmbientDim>::type p(a_p.derived());
+  Scalar dist2(0);
+  Scalar aux;
+  for (Index k=0; k<dim(); ++k)
+  {
+    if( m_min[k] > p[k] )
+    {
+      aux = m_min[k] - p[k];
+      dist2 += aux*aux;
+    }
+    else if( p[k] > m_max[k] )
+    {
+      aux = p[k] - m_max[k];
+      dist2 += aux*aux;
+    }
+  }
+  return dist2;
+}
+
+template<typename Scalar,int AmbientDim>
+EIGEN_DEVICE_FUNC inline Scalar AlignedBox<Scalar,AmbientDim>::squaredExteriorDistance(const AlignedBox& b) const
+{
+  Scalar dist2(0);
+  Scalar aux;
+  for (Index k=0; k<dim(); ++k)
+  {
+    if( m_min[k] > b.m_max[k] )
+    {
+      aux = m_min[k] - b.m_max[k];
+      dist2 += aux*aux;
+    }
+    else if( b.m_min[k] > m_max[k] )
+    {
+      aux = b.m_min[k] - m_max[k];
+      dist2 += aux*aux;
+    }
+  }
+  return dist2;
+}
+
+/** \defgroup alignedboxtypedefs Global aligned box typedefs
+  *
+  * \ingroup Geometry_Module
+  *
+  * Eigen defines several typedef shortcuts for most common aligned box types.
+  *
+  * The general patterns are the following:
+  *
+  * \c AlignedBoxSizeType where \c Size can be \c 1, \c 2,\c 3,\c 4 for fixed size boxes or \c X for dynamic size,
+  * and where \c Type can be \c i for integer, \c f for float, \c d for double.
+  *
+  * For example, \c AlignedBox3d is a fixed-size 3x3 aligned box type of doubles, and \c AlignedBoxXf is a dynamic-size aligned box of floats.
+  *
+  * \sa class AlignedBox
+  */
+
+#define EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Size, SizeSuffix)    \
+/** \ingroup alignedboxtypedefs */                                 \
+typedef AlignedBox<Type, Size>   AlignedBox##SizeSuffix##TypeSuffix;
+
+#define EIGEN_MAKE_TYPEDEFS_ALL_SIZES(Type, TypeSuffix) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 1, 1) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 2, 2) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 3, 3) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, 4, 4) \
+EIGEN_MAKE_TYPEDEFS(Type, TypeSuffix, Dynamic, X)
+
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(int,                  i)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(float,                f)
+EIGEN_MAKE_TYPEDEFS_ALL_SIZES(double,               d)
+
+#undef EIGEN_MAKE_TYPEDEFS_ALL_SIZES
+#undef EIGEN_MAKE_TYPEDEFS
+
+} // end namespace Eigen
+
+#endif // EIGEN_ALIGNEDBOX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AngleAxis.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AngleAxis.h
new file mode 100644
index 0000000..0af3c1b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/AngleAxis.h
@@ -0,0 +1,247 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ANGLEAXIS_H
+#define EIGEN_ANGLEAXIS_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class AngleAxis
+  *
+  * \brief Represents a 3D rotation as a rotation angle around an arbitrary 3D axis
+  *
+  * \param _Scalar the scalar type, i.e., the type of the coefficients.
+  *
+  * \warning When setting up an AngleAxis object, the axis vector \b must \b be \b normalized.
+  *
+  * The following two typedefs are provided for convenience:
+  * \li \c AngleAxisf for \c float
+  * \li \c AngleAxisd for \c double
+  *
+  * Combined with MatrixBase::Unit{X,Y,Z}, AngleAxis can be used to easily
+  * mimic Euler-angles. Here is an example:
+  * \include AngleAxis_mimic_euler.cpp
+  * Output: \verbinclude AngleAxis_mimic_euler.out
+  *
+  * \note This class is not aimed to be used to store a rotation transformation,
+  * but rather to make easier the creation of other rotation (Quaternion, rotation Matrix)
+  * and transformation objects.
+  *
+  * \sa class Quaternion, class Transform, MatrixBase::UnitX()
+  */
+
+namespace internal {
+template<typename _Scalar> struct traits<AngleAxis<_Scalar> >
+{
+  typedef _Scalar Scalar;
+};
+}
+
+template<typename _Scalar>
+class AngleAxis : public RotationBase<AngleAxis<_Scalar>,3>
+{
+  typedef RotationBase<AngleAxis<_Scalar>,3> Base;
+
+public:
+
+  using Base::operator*;
+
+  enum { Dim = 3 };
+  /** the scalar type of the coefficients */
+  typedef _Scalar Scalar;
+  typedef Matrix<Scalar,3,3> Matrix3;
+  typedef Matrix<Scalar,3,1> Vector3;
+  typedef Quaternion<Scalar> QuaternionType;
+
+protected:
+
+  Vector3 m_axis;
+  Scalar m_angle;
+
+public:
+
+  /** Default constructor without initialization. */
+  EIGEN_DEVICE_FUNC AngleAxis() {}
+  /** Constructs and initialize the angle-axis rotation from an \a angle in radian
+    * and an \a axis which \b must \b be \b normalized.
+    *
+    * \warning If the \a axis vector is not normalized, then the angle-axis object
+    *          represents an invalid rotation. */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC 
+  inline AngleAxis(const Scalar& angle, const MatrixBase<Derived>& axis) : m_axis(axis), m_angle(angle) {}
+  /** Constructs and initialize the angle-axis rotation from a quaternion \a q.
+    * This function implicitly normalizes the quaternion \a q.
+    */
+  template<typename QuatDerived> 
+  EIGEN_DEVICE_FUNC inline explicit AngleAxis(const QuaternionBase<QuatDerived>& q) { *this = q; }
+  /** Constructs and initialize the angle-axis rotation from a 3x3 rotation matrix. */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline explicit AngleAxis(const MatrixBase<Derived>& m) { *this = m; }
+
+  /** \returns the value of the rotation angle in radian */
+  EIGEN_DEVICE_FUNC Scalar angle() const { return m_angle; }
+  /** \returns a read-write reference to the stored angle in radian */
+  EIGEN_DEVICE_FUNC Scalar& angle() { return m_angle; }
+
+  /** \returns the rotation axis */
+  EIGEN_DEVICE_FUNC const Vector3& axis() const { return m_axis; }
+  /** \returns a read-write reference to the stored rotation axis.
+    *
+    * \warning The rotation axis must remain a \b unit vector.
+    */
+  EIGEN_DEVICE_FUNC Vector3& axis() { return m_axis; }
+
+  /** Concatenates two rotations */
+  EIGEN_DEVICE_FUNC inline QuaternionType operator* (const AngleAxis& other) const
+  { return QuaternionType(*this) * QuaternionType(other); }
+
+  /** Concatenates two rotations */
+  EIGEN_DEVICE_FUNC inline QuaternionType operator* (const QuaternionType& other) const
+  { return QuaternionType(*this) * other; }
+
+  /** Concatenates two rotations */
+  friend EIGEN_DEVICE_FUNC inline QuaternionType operator* (const QuaternionType& a, const AngleAxis& b)
+  { return a * QuaternionType(b); }
+
+  /** \returns the inverse rotation, i.e., an angle-axis with opposite rotation angle */
+  EIGEN_DEVICE_FUNC AngleAxis inverse() const
+  { return AngleAxis(-m_angle, m_axis); }
+
+  template<class QuatDerived>
+  EIGEN_DEVICE_FUNC AngleAxis& operator=(const QuaternionBase<QuatDerived>& q);
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC AngleAxis& operator=(const MatrixBase<Derived>& m);
+
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC AngleAxis& fromRotationMatrix(const MatrixBase<Derived>& m);
+  EIGEN_DEVICE_FUNC Matrix3 toRotationMatrix(void) const;
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<AngleAxis,AngleAxis<NewScalarType> >::type cast() const
+  { return typename internal::cast_return_type<AngleAxis,AngleAxis<NewScalarType> >::type(*this); }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  EIGEN_DEVICE_FUNC inline explicit AngleAxis(const AngleAxis<OtherScalarType>& other)
+  {
+    m_axis = other.axis().template cast<Scalar>();
+    m_angle = Scalar(other.angle());
+  }
+
+  EIGEN_DEVICE_FUNC static inline const AngleAxis Identity() { return AngleAxis(Scalar(0), Vector3::UnitX()); }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const AngleAxis& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return m_axis.isApprox(other.m_axis, prec) && internal::isApprox(m_angle,other.m_angle, prec); }
+};
+
+/** \ingroup Geometry_Module
+  * single precision angle-axis type */
+typedef AngleAxis<float> AngleAxisf;
+/** \ingroup Geometry_Module
+  * double precision angle-axis type */
+typedef AngleAxis<double> AngleAxisd;
+
+/** Set \c *this from a \b unit quaternion.
+  *
+  * The resulting axis is normalized, and the computed angle is in the [0,pi] range.
+  * 
+  * This function implicitly normalizes the quaternion \a q.
+  */
+template<typename Scalar>
+template<typename QuatDerived>
+EIGEN_DEVICE_FUNC AngleAxis<Scalar>& AngleAxis<Scalar>::operator=(const QuaternionBase<QuatDerived>& q)
+{
+  EIGEN_USING_STD_MATH(atan2)
+  EIGEN_USING_STD_MATH(abs)
+  Scalar n = q.vec().norm();
+  if(n<NumTraits<Scalar>::epsilon())
+    n = q.vec().stableNorm();
+
+  if (n != Scalar(0))
+  {
+    m_angle = Scalar(2)*atan2(n, abs(q.w()));
+    if(q.w() < 0)
+      n = -n;
+    m_axis  = q.vec() / n;
+  }
+  else
+  {
+    m_angle = Scalar(0);
+    m_axis << Scalar(1), Scalar(0), Scalar(0);
+  }
+  return *this;
+}
+
+/** Set \c *this from a 3x3 rotation matrix \a mat.
+  */
+template<typename Scalar>
+template<typename Derived>
+EIGEN_DEVICE_FUNC AngleAxis<Scalar>& AngleAxis<Scalar>::operator=(const MatrixBase<Derived>& mat)
+{
+  // Since a direct conversion would not be really faster,
+  // let's use the robust Quaternion implementation:
+  return *this = QuaternionType(mat);
+}
+
+/**
+* \brief Sets \c *this from a 3x3 rotation matrix.
+**/
+template<typename Scalar>
+template<typename Derived>
+EIGEN_DEVICE_FUNC AngleAxis<Scalar>& AngleAxis<Scalar>::fromRotationMatrix(const MatrixBase<Derived>& mat)
+{
+  return *this = QuaternionType(mat);
+}
+
+/** Constructs and \returns an equivalent 3x3 rotation matrix.
+  */
+template<typename Scalar>
+typename AngleAxis<Scalar>::Matrix3
+EIGEN_DEVICE_FUNC AngleAxis<Scalar>::toRotationMatrix(void) const
+{
+  EIGEN_USING_STD_MATH(sin)
+  EIGEN_USING_STD_MATH(cos)
+  Matrix3 res;
+  Vector3 sin_axis  = sin(m_angle) * m_axis;
+  Scalar c = cos(m_angle);
+  Vector3 cos1_axis = (Scalar(1)-c) * m_axis;
+
+  Scalar tmp;
+  tmp = cos1_axis.x() * m_axis.y();
+  res.coeffRef(0,1) = tmp - sin_axis.z();
+  res.coeffRef(1,0) = tmp + sin_axis.z();
+
+  tmp = cos1_axis.x() * m_axis.z();
+  res.coeffRef(0,2) = tmp + sin_axis.y();
+  res.coeffRef(2,0) = tmp - sin_axis.y();
+
+  tmp = cos1_axis.y() * m_axis.z();
+  res.coeffRef(1,2) = tmp - sin_axis.x();
+  res.coeffRef(2,1) = tmp + sin_axis.x();
+
+  res.diagonal() = (cos1_axis.cwiseProduct(m_axis)).array() + c;
+
+  return res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_ANGLEAXIS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/EulerAngles.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/EulerAngles.h
new file mode 100644
index 0000000..c633268
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/EulerAngles.h
@@ -0,0 +1,114 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_EULERANGLES_H
+#define EIGEN_EULERANGLES_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  *
+  * \returns the Euler-angles of the rotation matrix \c *this using the convention defined by the triplet (\a a0,\a a1,\a a2)
+  *
+  * Each of the three parameters \a a0,\a a1,\a a2 represents the respective rotation axis as an integer in {0,1,2}.
+  * For instance, in:
+  * \code Vector3f ea = mat.eulerAngles(2, 0, 2); \endcode
+  * "2" represents the z axis and "0" the x axis, etc. The returned angles are such that
+  * we have the following equality:
+  * \code
+  * mat == AngleAxisf(ea[0], Vector3f::UnitZ())
+  *      * AngleAxisf(ea[1], Vector3f::UnitX())
+  *      * AngleAxisf(ea[2], Vector3f::UnitZ()); \endcode
+  * This corresponds to the right-multiply conventions (with right hand side frames).
+  * 
+  * The returned angles are in the ranges [0:pi]x[-pi:pi]x[-pi:pi].
+  * 
+  * \sa class AngleAxis
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline Matrix<typename MatrixBase<Derived>::Scalar,3,1>
+MatrixBase<Derived>::eulerAngles(Index a0, Index a1, Index a2) const
+{
+  EIGEN_USING_STD_MATH(atan2)
+  EIGEN_USING_STD_MATH(sin)
+  EIGEN_USING_STD_MATH(cos)
+  /* Implemented from Graphics Gems IV */
+  EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Derived,3,3)
+
+  Matrix<Scalar,3,1> res;
+  typedef Matrix<typename Derived::Scalar,2,1> Vector2;
+
+  const Index odd = ((a0+1)%3 == a1) ? 0 : 1;
+  const Index i = a0;
+  const Index j = (a0 + 1 + odd)%3;
+  const Index k = (a0 + 2 - odd)%3;
+  
+  if (a0==a2)
+  {
+    res[0] = atan2(coeff(j,i), coeff(k,i));
+    if((odd && res[0]<Scalar(0)) || ((!odd) && res[0]>Scalar(0)))
+    {
+      if(res[0] > Scalar(0)) {
+        res[0] -= Scalar(EIGEN_PI);
+      }
+      else {
+        res[0] += Scalar(EIGEN_PI);
+      }
+      Scalar s2 = Vector2(coeff(j,i), coeff(k,i)).norm();
+      res[1] = -atan2(s2, coeff(i,i));
+    }
+    else
+    {
+      Scalar s2 = Vector2(coeff(j,i), coeff(k,i)).norm();
+      res[1] = atan2(s2, coeff(i,i));
+    }
+    
+    // With a=(0,1,0), we have i=0; j=1; k=2, and after computing the first two angles,
+    // we can compute their respective rotation, and apply its inverse to M. Since the result must
+    // be a rotation around x, we have:
+    //
+    //  c2  s1.s2 c1.s2                   1  0   0 
+    //  0   c1    -s1       *    M    =   0  c3  s3
+    //  -s2 s1.c2 c1.c2                   0 -s3  c3
+    //
+    //  Thus:  m11.c1 - m21.s1 = c3  &   m12.c1 - m22.s1 = s3
+    
+    Scalar s1 = sin(res[0]);
+    Scalar c1 = cos(res[0]);
+    res[2] = atan2(c1*coeff(j,k)-s1*coeff(k,k), c1*coeff(j,j) - s1 * coeff(k,j));
+  } 
+  else
+  {
+    res[0] = atan2(coeff(j,k), coeff(k,k));
+    Scalar c2 = Vector2(coeff(i,i), coeff(i,j)).norm();
+    if((odd && res[0]<Scalar(0)) || ((!odd) && res[0]>Scalar(0))) {
+      if(res[0] > Scalar(0)) {
+        res[0] -= Scalar(EIGEN_PI);
+      }
+      else {
+        res[0] += Scalar(EIGEN_PI);
+      }
+      res[1] = atan2(-coeff(i,k), -c2);
+    }
+    else
+      res[1] = atan2(-coeff(i,k), c2);
+    Scalar s1 = sin(res[0]);
+    Scalar c1 = cos(res[0]);
+    res[2] = atan2(s1*coeff(k,i)-c1*coeff(j,i), c1*coeff(j,j) - s1 * coeff(k,j));
+  }
+  if (!odd)
+    res = -res;
+  
+  return res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_EULERANGLES_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Homogeneous.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Homogeneous.h
new file mode 100644
index 0000000..5f0da1a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Homogeneous.h
@@ -0,0 +1,497 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HOMOGENEOUS_H
+#define EIGEN_HOMOGENEOUS_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Homogeneous
+  *
+  * \brief Expression of one (or a set of) homogeneous vector(s)
+  *
+  * \param MatrixType the type of the object in which we are making homogeneous
+  *
+  * This class represents an expression of one (or a set of) homogeneous vector(s).
+  * It is the return type of MatrixBase::homogeneous() and most of the time
+  * this is the only way it is used.
+  *
+  * \sa MatrixBase::homogeneous()
+  */
+
+namespace internal {
+
+template<typename MatrixType,int Direction>
+struct traits<Homogeneous<MatrixType,Direction> >
+ : traits<MatrixType>
+{
+  typedef typename traits<MatrixType>::StorageKind StorageKind;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+  enum {
+    RowsPlusOne = (MatrixType::RowsAtCompileTime != Dynamic) ?
+                  int(MatrixType::RowsAtCompileTime) + 1 : Dynamic,
+    ColsPlusOne = (MatrixType::ColsAtCompileTime != Dynamic) ?
+                  int(MatrixType::ColsAtCompileTime) + 1 : Dynamic,
+    RowsAtCompileTime = Direction==Vertical  ?  RowsPlusOne : MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = Direction==Horizontal ? ColsPlusOne : MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = RowsAtCompileTime,
+    MaxColsAtCompileTime = ColsAtCompileTime,
+    TmpFlags = _MatrixTypeNested::Flags & HereditaryBits,
+    Flags = ColsAtCompileTime==1 ? (TmpFlags & ~RowMajorBit)
+          : RowsAtCompileTime==1 ? (TmpFlags | RowMajorBit)
+          : TmpFlags
+  };
+};
+
+template<typename MatrixType,typename Lhs> struct homogeneous_left_product_impl;
+template<typename MatrixType,typename Rhs> struct homogeneous_right_product_impl;
+
+} // end namespace internal
+
+template<typename MatrixType,int _Direction> class Homogeneous
+  : public MatrixBase<Homogeneous<MatrixType,_Direction> >, internal::no_assignment_operator
+{
+  public:
+
+    typedef MatrixType NestedExpression;
+    enum { Direction = _Direction };
+
+    typedef MatrixBase<Homogeneous> Base;
+    EIGEN_DENSE_PUBLIC_INTERFACE(Homogeneous)
+
+    EIGEN_DEVICE_FUNC explicit inline Homogeneous(const MatrixType& matrix)
+      : m_matrix(matrix)
+    {}
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_matrix.rows() + (int(Direction)==Vertical   ? 1 : 0); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_matrix.cols() + (int(Direction)==Horizontal ? 1 : 0); }
+    
+    EIGEN_DEVICE_FUNC const NestedExpression& nestedExpression() const { return m_matrix; }
+
+    template<typename Rhs>
+    EIGEN_DEVICE_FUNC inline const Product<Homogeneous,Rhs>
+    operator* (const MatrixBase<Rhs>& rhs) const
+    {
+      eigen_assert(int(Direction)==Horizontal);
+      return Product<Homogeneous,Rhs>(*this,rhs.derived());
+    }
+
+    template<typename Lhs> friend
+    EIGEN_DEVICE_FUNC inline const Product<Lhs,Homogeneous>
+    operator* (const MatrixBase<Lhs>& lhs, const Homogeneous& rhs)
+    {
+      eigen_assert(int(Direction)==Vertical);
+      return Product<Lhs,Homogeneous>(lhs.derived(),rhs);
+    }
+
+    template<typename Scalar, int Dim, int Mode, int Options> friend
+    EIGEN_DEVICE_FUNC inline const Product<Transform<Scalar,Dim,Mode,Options>, Homogeneous >
+    operator* (const Transform<Scalar,Dim,Mode,Options>& lhs, const Homogeneous& rhs)
+    {
+      eigen_assert(int(Direction)==Vertical);
+      return Product<Transform<Scalar,Dim,Mode,Options>, Homogeneous>(lhs,rhs);
+    }
+
+    template<typename Func>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::result_of<Func(Scalar,Scalar)>::type
+    redux(const Func& func) const
+    {
+      return func(m_matrix.redux(func), Scalar(1));
+    }
+
+  protected:
+    typename MatrixType::Nested m_matrix;
+};
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns a vector expression that is one longer than the vector argument, with the value 1 symbolically appended as the last coefficient.
+  *
+  * This can be used to convert affine coordinates to homogeneous coordinates.
+  *
+  * \only_for_vectors
+  *
+  * Example: \include MatrixBase_homogeneous.cpp
+  * Output: \verbinclude MatrixBase_homogeneous.out
+  *
+  * \sa VectorwiseOp::homogeneous(), class Homogeneous
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::HomogeneousReturnType
+MatrixBase<Derived>::homogeneous() const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+  return HomogeneousReturnType(derived());
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns an expression where the value 1 is symbolically appended as the final coefficient to each column (or row) of the matrix.
+  *
+  * This can be used to convert affine coordinates to homogeneous coordinates.
+  *
+  * Example: \include VectorwiseOp_homogeneous.cpp
+  * Output: \verbinclude VectorwiseOp_homogeneous.out
+  *
+  * \sa MatrixBase::homogeneous(), class Homogeneous */
+template<typename ExpressionType, int Direction>
+EIGEN_DEVICE_FUNC inline Homogeneous<ExpressionType,Direction>
+VectorwiseOp<ExpressionType,Direction>::homogeneous() const
+{
+  return HomogeneousReturnType(_expression());
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \brief homogeneous normalization
+  *
+  * \returns a vector expression of the N-1 first coefficients of \c *this divided by that last coefficient.
+  *
+  * This can be used to convert homogeneous coordinates to affine coordinates.
+  *
+  * It is essentially a shortcut for:
+  * \code
+    this->head(this->size()-1)/this->coeff(this->size()-1);
+    \endcode
+  *
+  * Example: \include MatrixBase_hnormalized.cpp
+  * Output: \verbinclude MatrixBase_hnormalized.out
+  *
+  * \sa VectorwiseOp::hnormalized() */
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline const typename MatrixBase<Derived>::HNormalizedReturnType
+MatrixBase<Derived>::hnormalized() const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+  return ConstStartMinusOne(derived(),0,0,
+    ColsAtCompileTime==1?size()-1:1,
+    ColsAtCompileTime==1?1:size()-1) / coeff(size()-1);
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \brief column or row-wise homogeneous normalization
+  *
+  * \returns an expression of the first N-1 coefficients of each column (or row) of \c *this divided by the last coefficient of each column (or row).
+  *
+  * This can be used to convert homogeneous coordinates to affine coordinates.
+  *
+  * It is conceptually equivalent to calling MatrixBase::hnormalized() to each column (or row) of \c *this.
+  *
+  * Example: \include DirectionWise_hnormalized.cpp
+  * Output: \verbinclude DirectionWise_hnormalized.out
+  *
+  * \sa MatrixBase::hnormalized() */
+template<typename ExpressionType, int Direction>
+EIGEN_DEVICE_FUNC inline const typename VectorwiseOp<ExpressionType,Direction>::HNormalizedReturnType
+VectorwiseOp<ExpressionType,Direction>::hnormalized() const
+{
+  return HNormalized_Block(_expression(),0,0,
+      Direction==Vertical   ? _expression().rows()-1 : _expression().rows(),
+      Direction==Horizontal ? _expression().cols()-1 : _expression().cols()).cwiseQuotient(
+      Replicate<HNormalized_Factors,
+                Direction==Vertical   ? HNormalized_SizeMinusOne : 1,
+                Direction==Horizontal ? HNormalized_SizeMinusOne : 1>
+        (HNormalized_Factors(_expression(),
+          Direction==Vertical    ? _expression().rows()-1:0,
+          Direction==Horizontal  ? _expression().cols()-1:0,
+          Direction==Vertical    ? 1 : _expression().rows(),
+          Direction==Horizontal  ? 1 : _expression().cols()),
+         Direction==Vertical   ? _expression().rows()-1 : 1,
+         Direction==Horizontal ? _expression().cols()-1 : 1));
+}
+
+namespace internal {
+
+template<typename MatrixOrTransformType>
+struct take_matrix_for_product
+{
+  typedef MatrixOrTransformType type;
+  EIGEN_DEVICE_FUNC static const type& run(const type &x) { return x; }
+};
+
+template<typename Scalar, int Dim, int Mode,int Options>
+struct take_matrix_for_product<Transform<Scalar, Dim, Mode, Options> >
+{
+  typedef Transform<Scalar, Dim, Mode, Options> TransformType;
+  typedef typename internal::add_const<typename TransformType::ConstAffinePart>::type type;
+  EIGEN_DEVICE_FUNC static type run (const TransformType& x) { return x.affine(); }
+};
+
+template<typename Scalar, int Dim, int Options>
+struct take_matrix_for_product<Transform<Scalar, Dim, Projective, Options> >
+{
+  typedef Transform<Scalar, Dim, Projective, Options> TransformType;
+  typedef typename TransformType::MatrixType type;
+  EIGEN_DEVICE_FUNC static const type& run (const TransformType& x) { return x.matrix(); }
+};
+
+template<typename MatrixType,typename Lhs>
+struct traits<homogeneous_left_product_impl<Homogeneous<MatrixType,Vertical>,Lhs> >
+{
+  typedef typename take_matrix_for_product<Lhs>::type LhsMatrixType;
+  typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+  typedef typename remove_all<LhsMatrixType>::type LhsMatrixTypeCleaned;
+  typedef typename make_proper_matrix_type<
+                 typename traits<MatrixTypeCleaned>::Scalar,
+                 LhsMatrixTypeCleaned::RowsAtCompileTime,
+                 MatrixTypeCleaned::ColsAtCompileTime,
+                 MatrixTypeCleaned::PlainObject::Options,
+                 LhsMatrixTypeCleaned::MaxRowsAtCompileTime,
+                 MatrixTypeCleaned::MaxColsAtCompileTime>::type ReturnType;
+};
+
+template<typename MatrixType,typename Lhs>
+struct homogeneous_left_product_impl<Homogeneous<MatrixType,Vertical>,Lhs>
+  : public ReturnByValue<homogeneous_left_product_impl<Homogeneous<MatrixType,Vertical>,Lhs> >
+{
+  typedef typename traits<homogeneous_left_product_impl>::LhsMatrixType LhsMatrixType;
+  typedef typename remove_all<LhsMatrixType>::type LhsMatrixTypeCleaned;
+  typedef typename remove_all<typename LhsMatrixTypeCleaned::Nested>::type LhsMatrixTypeNested;
+  EIGEN_DEVICE_FUNC homogeneous_left_product_impl(const Lhs& lhs, const MatrixType& rhs)
+    : m_lhs(take_matrix_for_product<Lhs>::run(lhs)),
+      m_rhs(rhs)
+  {}
+
+  EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+  EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
+
+  template<typename Dest> EIGEN_DEVICE_FUNC void evalTo(Dest& dst) const
+  {
+    // FIXME investigate how to allow lazy evaluation of this product when possible
+    dst = Block<const LhsMatrixTypeNested,
+              LhsMatrixTypeNested::RowsAtCompileTime,
+              LhsMatrixTypeNested::ColsAtCompileTime==Dynamic?Dynamic:LhsMatrixTypeNested::ColsAtCompileTime-1>
+            (m_lhs,0,0,m_lhs.rows(),m_lhs.cols()-1) * m_rhs;
+    dst += m_lhs.col(m_lhs.cols()-1).rowwise()
+            .template replicate<MatrixType::ColsAtCompileTime>(m_rhs.cols());
+  }
+
+  typename LhsMatrixTypeCleaned::Nested m_lhs;
+  typename MatrixType::Nested m_rhs;
+};
+
+template<typename MatrixType,typename Rhs>
+struct traits<homogeneous_right_product_impl<Homogeneous<MatrixType,Horizontal>,Rhs> >
+{
+  typedef typename make_proper_matrix_type<typename traits<MatrixType>::Scalar,
+                 MatrixType::RowsAtCompileTime,
+                 Rhs::ColsAtCompileTime,
+                 MatrixType::PlainObject::Options,
+                 MatrixType::MaxRowsAtCompileTime,
+                 Rhs::MaxColsAtCompileTime>::type ReturnType;
+};
+
+template<typename MatrixType,typename Rhs>
+struct homogeneous_right_product_impl<Homogeneous<MatrixType,Horizontal>,Rhs>
+  : public ReturnByValue<homogeneous_right_product_impl<Homogeneous<MatrixType,Horizontal>,Rhs> >
+{
+  typedef typename remove_all<typename Rhs::Nested>::type RhsNested;
+  EIGEN_DEVICE_FUNC homogeneous_right_product_impl(const MatrixType& lhs, const Rhs& rhs)
+    : m_lhs(lhs), m_rhs(rhs)
+  {}
+
+  EIGEN_DEVICE_FUNC inline Index rows() const { return m_lhs.rows(); }
+  EIGEN_DEVICE_FUNC inline Index cols() const { return m_rhs.cols(); }
+
+  template<typename Dest> EIGEN_DEVICE_FUNC void evalTo(Dest& dst) const
+  {
+    // FIXME investigate how to allow lazy evaluation of this product when possible
+    dst = m_lhs * Block<const RhsNested,
+                        RhsNested::RowsAtCompileTime==Dynamic?Dynamic:RhsNested::RowsAtCompileTime-1,
+                        RhsNested::ColsAtCompileTime>
+            (m_rhs,0,0,m_rhs.rows()-1,m_rhs.cols());
+    dst += m_rhs.row(m_rhs.rows()-1).colwise()
+            .template replicate<MatrixType::RowsAtCompileTime>(m_lhs.rows());
+  }
+
+  typename MatrixType::Nested m_lhs;
+  typename Rhs::Nested m_rhs;
+};
+
+template<typename ArgType,int Direction>
+struct evaluator_traits<Homogeneous<ArgType,Direction> >
+{
+  typedef typename storage_kind_to_evaluator_kind<typename ArgType::StorageKind>::Kind Kind;
+  typedef HomogeneousShape Shape;  
+};
+
+template<> struct AssignmentKind<DenseShape,HomogeneousShape> { typedef Dense2Dense Kind; };
+
+
+template<typename ArgType,int Direction>
+struct unary_evaluator<Homogeneous<ArgType,Direction>, IndexBased>
+  : evaluator<typename Homogeneous<ArgType,Direction>::PlainObject >
+{
+  typedef Homogeneous<ArgType,Direction> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  EIGEN_DEVICE_FUNC explicit unary_evaluator(const XprType& op)
+    : Base(), m_temp(op)
+  {
+    ::new (static_cast<Base*>(this)) Base(m_temp);
+  }
+
+protected:
+  PlainObject m_temp;
+};
+
+// dense = homogeneous
+template< typename DstXprType, typename ArgType, typename Scalar>
+struct Assignment<DstXprType, Homogeneous<ArgType,Vertical>, internal::assign_op<Scalar,typename ArgType::Scalar>, Dense2Dense>
+{
+  typedef Homogeneous<ArgType,Vertical> SrcXprType;
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename ArgType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst.template topRows<ArgType::RowsAtCompileTime>(src.nestedExpression().rows()) = src.nestedExpression();
+    dst.row(dst.rows()-1).setOnes();
+  }
+};
+
+// dense = homogeneous
+template< typename DstXprType, typename ArgType, typename Scalar>
+struct Assignment<DstXprType, Homogeneous<ArgType,Horizontal>, internal::assign_op<Scalar,typename ArgType::Scalar>, Dense2Dense>
+{
+  typedef Homogeneous<ArgType,Horizontal> SrcXprType;
+  EIGEN_DEVICE_FUNC static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename ArgType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst.template leftCols<ArgType::ColsAtCompileTime>(src.nestedExpression().cols()) = src.nestedExpression();
+    dst.col(dst.cols()-1).setOnes();
+  }
+};
+
+template<typename LhsArg, typename Rhs, int ProductTag>
+struct generic_product_impl<Homogeneous<LhsArg,Horizontal>, Rhs, HomogeneousShape, DenseShape, ProductTag>
+{
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC static void evalTo(Dest& dst, const Homogeneous<LhsArg,Horizontal>& lhs, const Rhs& rhs)
+  {
+    homogeneous_right_product_impl<Homogeneous<LhsArg,Horizontal>, Rhs>(lhs.nestedExpression(), rhs).evalTo(dst);
+  }
+};
+
+template<typename Lhs,typename Rhs>
+struct homogeneous_right_product_refactoring_helper
+{
+  enum {
+    Dim  = Lhs::ColsAtCompileTime,
+    Rows = Lhs::RowsAtCompileTime
+  };
+  typedef typename Rhs::template ConstNRowsBlockXpr<Dim>::Type          LinearBlockConst;
+  typedef typename remove_const<LinearBlockConst>::type                 LinearBlock;
+  typedef typename Rhs::ConstRowXpr                                     ConstantColumn;
+  typedef Replicate<const ConstantColumn,Rows,1>                        ConstantBlock;
+  typedef Product<Lhs,LinearBlock,LazyProduct>                          LinearProduct;
+  typedef CwiseBinaryOp<internal::scalar_sum_op<typename Lhs::Scalar,typename Rhs::Scalar>, const LinearProduct, const ConstantBlock> Xpr;
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, HomogeneousShape, DenseShape>
+ : public evaluator<typename homogeneous_right_product_refactoring_helper<typename Lhs::NestedExpression,Rhs>::Xpr>
+{
+  typedef Product<Lhs, Rhs, LazyProduct> XprType;
+  typedef homogeneous_right_product_refactoring_helper<typename Lhs::NestedExpression,Rhs> helper;
+  typedef typename helper::ConstantBlock ConstantBlock;
+  typedef typename helper::Xpr RefactoredXpr;
+  typedef evaluator<RefactoredXpr> Base;
+  
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(  xpr.lhs().nestedExpression() .lazyProduct(  xpr.rhs().template topRows<helper::Dim>(xpr.lhs().nestedExpression().cols()) )
+            + ConstantBlock(xpr.rhs().row(xpr.rhs().rows()-1),xpr.lhs().rows(), 1) )
+  {}
+};
+
+template<typename Lhs, typename RhsArg, int ProductTag>
+struct generic_product_impl<Lhs, Homogeneous<RhsArg,Vertical>, DenseShape, HomogeneousShape, ProductTag>
+{
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC static void evalTo(Dest& dst, const Lhs& lhs, const Homogeneous<RhsArg,Vertical>& rhs)
+  {
+    homogeneous_left_product_impl<Homogeneous<RhsArg,Vertical>, Lhs>(lhs, rhs.nestedExpression()).evalTo(dst);
+  }
+};
+
+// TODO: the following specialization is to address a regression from 3.2 to 3.3
+// In the future, this path should be optimized.
+template<typename Lhs, typename RhsArg, int ProductTag>
+struct generic_product_impl<Lhs, Homogeneous<RhsArg,Vertical>, TriangularShape, HomogeneousShape, ProductTag>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Homogeneous<RhsArg,Vertical>& rhs)
+  {
+    dst.noalias() = lhs * rhs.eval();
+  }
+};
+
+template<typename Lhs,typename Rhs>
+struct homogeneous_left_product_refactoring_helper
+{
+  enum {
+    Dim = Rhs::RowsAtCompileTime,
+    Cols = Rhs::ColsAtCompileTime
+  };
+  typedef typename Lhs::template ConstNColsBlockXpr<Dim>::Type          LinearBlockConst;
+  typedef typename remove_const<LinearBlockConst>::type                 LinearBlock;
+  typedef typename Lhs::ConstColXpr                                     ConstantColumn;
+  typedef Replicate<const ConstantColumn,1,Cols>                        ConstantBlock;
+  typedef Product<LinearBlock,Rhs,LazyProduct>                          LinearProduct;
+  typedef CwiseBinaryOp<internal::scalar_sum_op<typename Lhs::Scalar,typename Rhs::Scalar>, const LinearProduct, const ConstantBlock> Xpr;
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, LazyProduct>, ProductTag, DenseShape, HomogeneousShape>
+ : public evaluator<typename homogeneous_left_product_refactoring_helper<Lhs,typename Rhs::NestedExpression>::Xpr>
+{
+  typedef Product<Lhs, Rhs, LazyProduct> XprType;
+  typedef homogeneous_left_product_refactoring_helper<Lhs,typename Rhs::NestedExpression> helper;
+  typedef typename helper::ConstantBlock ConstantBlock;
+  typedef typename helper::Xpr RefactoredXpr;
+  typedef evaluator<RefactoredXpr> Base;
+  
+  EIGEN_DEVICE_FUNC explicit product_evaluator(const XprType& xpr)
+    : Base(   xpr.lhs().template leftCols<helper::Dim>(xpr.rhs().nestedExpression().rows()) .lazyProduct( xpr.rhs().nestedExpression() )
+            + ConstantBlock(xpr.lhs().col(xpr.lhs().cols()-1),1,xpr.rhs().cols()) )
+  {}
+};
+
+template<typename Scalar, int Dim, int Mode,int Options, typename RhsArg, int ProductTag>
+struct generic_product_impl<Transform<Scalar,Dim,Mode,Options>, Homogeneous<RhsArg,Vertical>, DenseShape, HomogeneousShape, ProductTag>
+{
+  typedef Transform<Scalar,Dim,Mode,Options> TransformType;
+  template<typename Dest>
+  EIGEN_DEVICE_FUNC static void evalTo(Dest& dst, const TransformType& lhs, const Homogeneous<RhsArg,Vertical>& rhs)
+  {
+    homogeneous_left_product_impl<Homogeneous<RhsArg,Vertical>, TransformType>(lhs, rhs.nestedExpression()).evalTo(dst);
+  }
+};
+
+template<typename ExpressionType, int Side, bool Transposed>
+struct permutation_matrix_product<ExpressionType, Side, Transposed, HomogeneousShape>
+  : public permutation_matrix_product<ExpressionType, Side, Transposed, DenseShape>
+{};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_HOMOGENEOUS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Hyperplane.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Hyperplane.h
new file mode 100644
index 0000000..05929b2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Hyperplane.h
@@ -0,0 +1,282 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HYPERPLANE_H
+#define EIGEN_HYPERPLANE_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Hyperplane
+  *
+  * \brief A hyperplane
+  *
+  * A hyperplane is an affine subspace of dimension n-1 in a space of dimension n.
+  * For example, a hyperplane in a plane is a line; a hyperplane in 3-space is a plane.
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients
+  * \tparam _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic.
+  *             Notice that the dimension of the hyperplane is _AmbientDim-1.
+  *
+  * This class represents an hyperplane as the zero set of the implicit equation
+  * \f$ n \cdot x + d = 0 \f$ where \f$ n \f$ is a unit normal vector of the plane (linear part)
+  * and \f$ d \f$ is the distance (offset) to the origin.
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+class Hyperplane
+{
+public:
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim==Dynamic ? Dynamic : _AmbientDim+1)
+  enum {
+    AmbientDimAtCompileTime = _AmbientDim,
+    Options = _Options
+  };
+  typedef _Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+  typedef Matrix<Scalar,AmbientDimAtCompileTime,1> VectorType;
+  typedef Matrix<Scalar,Index(AmbientDimAtCompileTime)==Dynamic
+                        ? Dynamic
+                        : Index(AmbientDimAtCompileTime)+1,1,Options> Coefficients;
+  typedef Block<Coefficients,AmbientDimAtCompileTime,1> NormalReturnType;
+  typedef const Block<const Coefficients,AmbientDimAtCompileTime,1> ConstNormalReturnType;
+
+  /** Default constructor without initialization */
+  EIGEN_DEVICE_FUNC inline Hyperplane() {}
+  
+  template<int OtherOptions>
+  EIGEN_DEVICE_FUNC Hyperplane(const Hyperplane<Scalar,AmbientDimAtCompileTime,OtherOptions>& other)
+   : m_coeffs(other.coeffs())
+  {}
+
+  /** Constructs a dynamic-size hyperplane with \a _dim the dimension
+    * of the ambient space */
+  EIGEN_DEVICE_FUNC inline explicit Hyperplane(Index _dim) : m_coeffs(_dim+1) {}
+
+  /** Construct a plane from its normal \a n and a point \a e onto the plane.
+    * \warning the vector normal is assumed to be normalized.
+    */
+  EIGEN_DEVICE_FUNC inline Hyperplane(const VectorType& n, const VectorType& e)
+    : m_coeffs(n.size()+1)
+  {
+    normal() = n;
+    offset() = -n.dot(e);
+  }
+
+  /** Constructs a plane from its normal \a n and distance to the origin \a d
+    * such that the algebraic equation of the plane is \f$ n \cdot x + d = 0 \f$.
+    * \warning the vector normal is assumed to be normalized.
+    */
+  EIGEN_DEVICE_FUNC inline Hyperplane(const VectorType& n, const Scalar& d)
+    : m_coeffs(n.size()+1)
+  {
+    normal() = n;
+    offset() = d;
+  }
+
+  /** Constructs a hyperplane passing through the two points. If the dimension of the ambient space
+    * is greater than 2, then there isn't uniqueness, so an arbitrary choice is made.
+    */
+  EIGEN_DEVICE_FUNC static inline Hyperplane Through(const VectorType& p0, const VectorType& p1)
+  {
+    Hyperplane result(p0.size());
+    result.normal() = (p1 - p0).unitOrthogonal();
+    result.offset() = -p0.dot(result.normal());
+    return result;
+  }
+
+  /** Constructs a hyperplane passing through the three points. The dimension of the ambient space
+    * is required to be exactly 3.
+    */
+  EIGEN_DEVICE_FUNC static inline Hyperplane Through(const VectorType& p0, const VectorType& p1, const VectorType& p2)
+  {
+    EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 3)
+    Hyperplane result(p0.size());
+    VectorType v0(p2 - p0), v1(p1 - p0);
+    result.normal() = v0.cross(v1);
+    RealScalar norm = result.normal().norm();
+    if(norm <= v0.norm() * v1.norm() * NumTraits<RealScalar>::epsilon())
+    {
+      Matrix<Scalar,2,3> m; m << v0.transpose(), v1.transpose();
+      JacobiSVD<Matrix<Scalar,2,3> > svd(m, ComputeFullV);
+      result.normal() = svd.matrixV().col(2);
+    }
+    else
+      result.normal() /= norm;
+    result.offset() = -p0.dot(result.normal());
+    return result;
+  }
+
+  /** Constructs a hyperplane passing through the parametrized line \a parametrized.
+    * If the dimension of the ambient space is greater than 2, then there isn't uniqueness,
+    * so an arbitrary choice is made.
+    */
+  // FIXME to be consitent with the rest this could be implemented as a static Through function ??
+  EIGEN_DEVICE_FUNC explicit Hyperplane(const ParametrizedLine<Scalar, AmbientDimAtCompileTime>& parametrized)
+  {
+    normal() = parametrized.direction().unitOrthogonal();
+    offset() = -parametrized.origin().dot(normal());
+  }
+
+  EIGEN_DEVICE_FUNC ~Hyperplane() {}
+
+  /** \returns the dimension in which the plane holds */
+  EIGEN_DEVICE_FUNC inline Index dim() const { return AmbientDimAtCompileTime==Dynamic ? m_coeffs.size()-1 : Index(AmbientDimAtCompileTime); }
+
+  /** normalizes \c *this */
+  EIGEN_DEVICE_FUNC void normalize(void)
+  {
+    m_coeffs /= normal().norm();
+  }
+
+  /** \returns the signed distance between the plane \c *this and a point \a p.
+    * \sa absDistance()
+    */
+  EIGEN_DEVICE_FUNC inline Scalar signedDistance(const VectorType& p) const { return normal().dot(p) + offset(); }
+
+  /** \returns the absolute distance between the plane \c *this and a point \a p.
+    * \sa signedDistance()
+    */
+  EIGEN_DEVICE_FUNC inline Scalar absDistance(const VectorType& p) const { return numext::abs(signedDistance(p)); }
+
+  /** \returns the projection of a point \a p onto the plane \c *this.
+    */
+  EIGEN_DEVICE_FUNC inline VectorType projection(const VectorType& p) const { return p - signedDistance(p) * normal(); }
+
+  /** \returns a constant reference to the unit normal vector of the plane, which corresponds
+    * to the linear part of the implicit equation.
+    */
+  EIGEN_DEVICE_FUNC inline ConstNormalReturnType normal() const { return ConstNormalReturnType(m_coeffs,0,0,dim(),1); }
+
+  /** \returns a non-constant reference to the unit normal vector of the plane, which corresponds
+    * to the linear part of the implicit equation.
+    */
+  EIGEN_DEVICE_FUNC inline NormalReturnType normal() { return NormalReturnType(m_coeffs,0,0,dim(),1); }
+
+  /** \returns the distance to the origin, which is also the "constant term" of the implicit equation
+    * \warning the vector normal is assumed to be normalized.
+    */
+  EIGEN_DEVICE_FUNC inline const Scalar& offset() const { return m_coeffs.coeff(dim()); }
+
+  /** \returns a non-constant reference to the distance to the origin, which is also the constant part
+    * of the implicit equation */
+  EIGEN_DEVICE_FUNC inline Scalar& offset() { return m_coeffs(dim()); }
+
+  /** \returns a constant reference to the coefficients c_i of the plane equation:
+    * \f$ c_0*x_0 + ... + c_{d-1}*x_{d-1} + c_d = 0 \f$
+    */
+  EIGEN_DEVICE_FUNC inline const Coefficients& coeffs() const { return m_coeffs; }
+
+  /** \returns a non-constant reference to the coefficients c_i of the plane equation:
+    * \f$ c_0*x_0 + ... + c_{d-1}*x_{d-1} + c_d = 0 \f$
+    */
+  EIGEN_DEVICE_FUNC inline Coefficients& coeffs() { return m_coeffs; }
+
+  /** \returns the intersection of *this with \a other.
+    *
+    * \warning The ambient space must be a plane, i.e. have dimension 2, so that \c *this and \a other are lines.
+    *
+    * \note If \a other is approximately parallel to *this, this method will return any point on *this.
+    */
+  EIGEN_DEVICE_FUNC VectorType intersection(const Hyperplane& other) const
+  {
+    EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 2)
+    Scalar det = coeffs().coeff(0) * other.coeffs().coeff(1) - coeffs().coeff(1) * other.coeffs().coeff(0);
+    // since the line equations ax+by=c are normalized with a^2+b^2=1, the following tests
+    // whether the two lines are approximately parallel.
+    if(internal::isMuchSmallerThan(det, Scalar(1)))
+    {   // special case where the two lines are approximately parallel. Pick any point on the first line.
+        if(numext::abs(coeffs().coeff(1))>numext::abs(coeffs().coeff(0)))
+            return VectorType(coeffs().coeff(1), -coeffs().coeff(2)/coeffs().coeff(1)-coeffs().coeff(0));
+        else
+            return VectorType(-coeffs().coeff(2)/coeffs().coeff(0)-coeffs().coeff(1), coeffs().coeff(0));
+    }
+    else
+    {   // general case
+        Scalar invdet = Scalar(1) / det;
+        return VectorType(invdet*(coeffs().coeff(1)*other.coeffs().coeff(2)-other.coeffs().coeff(1)*coeffs().coeff(2)),
+                          invdet*(other.coeffs().coeff(0)*coeffs().coeff(2)-coeffs().coeff(0)*other.coeffs().coeff(2)));
+    }
+  }
+
+  /** Applies the transformation matrix \a mat to \c *this and returns a reference to \c *this.
+    *
+    * \param mat the Dim x Dim transformation matrix
+    * \param traits specifies whether the matrix \a mat represents an #Isometry
+    *               or a more generic #Affine transformation. The default is #Affine.
+    */
+  template<typename XprType>
+  EIGEN_DEVICE_FUNC inline Hyperplane& transform(const MatrixBase<XprType>& mat, TransformTraits traits = Affine)
+  {
+    if (traits==Affine)
+    {
+      normal() = mat.inverse().transpose() * normal();
+      m_coeffs /= normal().norm();
+    }
+    else if (traits==Isometry)
+      normal() = mat * normal();
+    else
+    {
+      eigen_assert(0 && "invalid traits value in Hyperplane::transform()");
+    }
+    return *this;
+  }
+
+  /** Applies the transformation \a t to \c *this and returns a reference to \c *this.
+    *
+    * \param t the transformation of dimension Dim
+    * \param traits specifies whether the transformation \a t represents an #Isometry
+    *               or a more generic #Affine transformation. The default is #Affine.
+    *               Other kind of transformations are not supported.
+    */
+  template<int TrOptions>
+  EIGEN_DEVICE_FUNC inline Hyperplane& transform(const Transform<Scalar,AmbientDimAtCompileTime,Affine,TrOptions>& t,
+                                TransformTraits traits = Affine)
+  {
+    transform(t.linear(), traits);
+    offset() -= normal().dot(t.translation());
+    return *this;
+  }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<Hyperplane,
+           Hyperplane<NewScalarType,AmbientDimAtCompileTime,Options> >::type cast() const
+  {
+    return typename internal::cast_return_type<Hyperplane,
+                    Hyperplane<NewScalarType,AmbientDimAtCompileTime,Options> >::type(*this);
+  }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType,int OtherOptions>
+  EIGEN_DEVICE_FUNC inline explicit Hyperplane(const Hyperplane<OtherScalarType,AmbientDimAtCompileTime,OtherOptions>& other)
+  { m_coeffs = other.coeffs().template cast<Scalar>(); }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  template<int OtherOptions>
+  EIGEN_DEVICE_FUNC bool isApprox(const Hyperplane<Scalar,AmbientDimAtCompileTime,OtherOptions>& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return m_coeffs.isApprox(other.m_coeffs, prec); }
+
+protected:
+
+  Coefficients m_coeffs;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_HYPERPLANE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/OrthoMethods.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/OrthoMethods.h
new file mode 100644
index 0000000..a035e63
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/OrthoMethods.h
@@ -0,0 +1,234 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ORTHOMETHODS_H
+#define EIGEN_ORTHOMETHODS_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns the cross product of \c *this and \a other
+  *
+  * Here is a very good explanation of cross-product: http://xkcd.com/199/
+  * 
+  * With complex numbers, the cross product is implemented as
+  * \f$ (\mathbf{a}+i\mathbf{b}) \times (\mathbf{c}+i\mathbf{d}) = (\mathbf{a} \times \mathbf{c} - \mathbf{b} \times \mathbf{d}) - i(\mathbf{a} \times \mathbf{d} - \mathbf{b} \times \mathbf{c})\f$
+  * 
+  * \sa MatrixBase::cross3()
+  */
+template<typename Derived>
+template<typename OtherDerived>
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::template cross_product_return_type<OtherDerived>::type
+#else
+inline typename MatrixBase<Derived>::PlainObject
+#endif
+MatrixBase<Derived>::cross(const MatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Derived,3)
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,3)
+
+  // Note that there is no need for an expression here since the compiler
+  // optimize such a small temporary very well (even within a complex expression)
+  typename internal::nested_eval<Derived,2>::type lhs(derived());
+  typename internal::nested_eval<OtherDerived,2>::type rhs(other.derived());
+  return typename cross_product_return_type<OtherDerived>::type(
+    numext::conj(lhs.coeff(1) * rhs.coeff(2) - lhs.coeff(2) * rhs.coeff(1)),
+    numext::conj(lhs.coeff(2) * rhs.coeff(0) - lhs.coeff(0) * rhs.coeff(2)),
+    numext::conj(lhs.coeff(0) * rhs.coeff(1) - lhs.coeff(1) * rhs.coeff(0))
+  );
+}
+
+namespace internal {
+
+template< int Arch,typename VectorLhs,typename VectorRhs,
+          typename Scalar = typename VectorLhs::Scalar,
+          bool Vectorizable = bool((VectorLhs::Flags&VectorRhs::Flags)&PacketAccessBit)>
+struct cross3_impl {
+  EIGEN_DEVICE_FUNC static inline typename internal::plain_matrix_type<VectorLhs>::type
+  run(const VectorLhs& lhs, const VectorRhs& rhs)
+  {
+    return typename internal::plain_matrix_type<VectorLhs>::type(
+      numext::conj(lhs.coeff(1) * rhs.coeff(2) - lhs.coeff(2) * rhs.coeff(1)),
+      numext::conj(lhs.coeff(2) * rhs.coeff(0) - lhs.coeff(0) * rhs.coeff(2)),
+      numext::conj(lhs.coeff(0) * rhs.coeff(1) - lhs.coeff(1) * rhs.coeff(0)),
+      0
+    );
+  }
+};
+
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns the cross product of \c *this and \a other using only the x, y, and z coefficients
+  *
+  * The size of \c *this and \a other must be four. This function is especially useful
+  * when using 4D vectors instead of 3D ones to get advantage of SSE/AltiVec vectorization.
+  *
+  * \sa MatrixBase::cross()
+  */
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC inline typename MatrixBase<Derived>::PlainObject
+MatrixBase<Derived>::cross3(const MatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(Derived,4)
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,4)
+
+  typedef typename internal::nested_eval<Derived,2>::type DerivedNested;
+  typedef typename internal::nested_eval<OtherDerived,2>::type OtherDerivedNested;
+  DerivedNested lhs(derived());
+  OtherDerivedNested rhs(other.derived());
+
+  return internal::cross3_impl<Architecture::Target,
+                        typename internal::remove_all<DerivedNested>::type,
+                        typename internal::remove_all<OtherDerivedNested>::type>::run(lhs,rhs);
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns a matrix expression of the cross product of each column or row
+  * of the referenced expression with the \a other vector.
+  *
+  * The referenced matrix must have one dimension equal to 3.
+  * The result matrix has the same dimensions than the referenced one.
+  *
+  * \sa MatrixBase::cross() */
+template<typename ExpressionType, int Direction>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC 
+const typename VectorwiseOp<ExpressionType,Direction>::CrossReturnType
+VectorwiseOp<ExpressionType,Direction>::cross(const MatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,3)
+  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+  
+  typename internal::nested_eval<ExpressionType,2>::type mat(_expression());
+  typename internal::nested_eval<OtherDerived,2>::type vec(other.derived());
+
+  CrossReturnType res(_expression().rows(),_expression().cols());
+  if(Direction==Vertical)
+  {
+    eigen_assert(CrossReturnType::RowsAtCompileTime==3 && "the matrix must have exactly 3 rows");
+    res.row(0) = (mat.row(1) * vec.coeff(2) - mat.row(2) * vec.coeff(1)).conjugate();
+    res.row(1) = (mat.row(2) * vec.coeff(0) - mat.row(0) * vec.coeff(2)).conjugate();
+    res.row(2) = (mat.row(0) * vec.coeff(1) - mat.row(1) * vec.coeff(0)).conjugate();
+  }
+  else
+  {
+    eigen_assert(CrossReturnType::ColsAtCompileTime==3 && "the matrix must have exactly 3 columns");
+    res.col(0) = (mat.col(1) * vec.coeff(2) - mat.col(2) * vec.coeff(1)).conjugate();
+    res.col(1) = (mat.col(2) * vec.coeff(0) - mat.col(0) * vec.coeff(2)).conjugate();
+    res.col(2) = (mat.col(0) * vec.coeff(1) - mat.col(1) * vec.coeff(0)).conjugate();
+  }
+  return res;
+}
+
+namespace internal {
+
+template<typename Derived, int Size = Derived::SizeAtCompileTime>
+struct unitOrthogonal_selector
+{
+  typedef typename plain_matrix_type<Derived>::type VectorType;
+  typedef typename traits<Derived>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Matrix<Scalar,2,1> Vector2;
+  EIGEN_DEVICE_FUNC
+  static inline VectorType run(const Derived& src)
+  {
+    VectorType perp = VectorType::Zero(src.size());
+    Index maxi = 0;
+    Index sndi = 0;
+    src.cwiseAbs().maxCoeff(&maxi);
+    if (maxi==0)
+      sndi = 1;
+    RealScalar invnm = RealScalar(1)/(Vector2() << src.coeff(sndi),src.coeff(maxi)).finished().norm();
+    perp.coeffRef(maxi) = -numext::conj(src.coeff(sndi)) * invnm;
+    perp.coeffRef(sndi) =  numext::conj(src.coeff(maxi)) * invnm;
+
+    return perp;
+   }
+};
+
+template<typename Derived>
+struct unitOrthogonal_selector<Derived,3>
+{
+  typedef typename plain_matrix_type<Derived>::type VectorType;
+  typedef typename traits<Derived>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  EIGEN_DEVICE_FUNC
+  static inline VectorType run(const Derived& src)
+  {
+    VectorType perp;
+    /* Let us compute the crossed product of *this with a vector
+     * that is not too close to being colinear to *this.
+     */
+
+    /* unless the x and y coords are both close to zero, we can
+     * simply take ( -y, x, 0 ) and normalize it.
+     */
+    if((!isMuchSmallerThan(src.x(), src.z()))
+    || (!isMuchSmallerThan(src.y(), src.z())))
+    {
+      RealScalar invnm = RealScalar(1)/src.template head<2>().norm();
+      perp.coeffRef(0) = -numext::conj(src.y())*invnm;
+      perp.coeffRef(1) = numext::conj(src.x())*invnm;
+      perp.coeffRef(2) = 0;
+    }
+    /* if both x and y are close to zero, then the vector is close
+     * to the z-axis, so it's far from colinear to the x-axis for instance.
+     * So we take the crossed product with (1,0,0) and normalize it.
+     */
+    else
+    {
+      RealScalar invnm = RealScalar(1)/src.template tail<2>().norm();
+      perp.coeffRef(0) = 0;
+      perp.coeffRef(1) = -numext::conj(src.z())*invnm;
+      perp.coeffRef(2) = numext::conj(src.y())*invnm;
+    }
+
+    return perp;
+   }
+};
+
+template<typename Derived>
+struct unitOrthogonal_selector<Derived,2>
+{
+  typedef typename plain_matrix_type<Derived>::type VectorType;
+  EIGEN_DEVICE_FUNC
+  static inline VectorType run(const Derived& src)
+  { return VectorType(-numext::conj(src.y()), numext::conj(src.x())).normalized(); }
+};
+
+} // end namespace internal
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \returns a unit vector which is orthogonal to \c *this
+  *
+  * The size of \c *this must be at least 2. If the size is exactly 2,
+  * then the returned vector is a counter clock wise rotation of \c *this, i.e., (-y,x).normalized().
+  *
+  * \sa cross()
+  */
+template<typename Derived>
+EIGEN_DEVICE_FUNC typename MatrixBase<Derived>::PlainObject
+MatrixBase<Derived>::unitOrthogonal() const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return internal::unitOrthogonal_selector<Derived>::run(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_ORTHOMETHODS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/ParametrizedLine.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/ParametrizedLine.h
new file mode 100644
index 0000000..1e985d8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/ParametrizedLine.h
@@ -0,0 +1,195 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARAMETRIZEDLINE_H
+#define EIGEN_PARAMETRIZEDLINE_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class ParametrizedLine
+  *
+  * \brief A parametrized line
+  *
+  * A parametrized line is defined by an origin point \f$ \mathbf{o} \f$ and a unit
+  * direction vector \f$ \mathbf{d} \f$ such that the line corresponds to
+  * the set \f$ l(t) = \mathbf{o} + t \mathbf{d} \f$, \f$ t \in \mathbf{R} \f$.
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients
+  * \tparam _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic.
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+class ParametrizedLine
+{
+public:
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
+  enum {
+    AmbientDimAtCompileTime = _AmbientDim,
+    Options = _Options
+  };
+  typedef _Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+  typedef Matrix<Scalar,AmbientDimAtCompileTime,1,Options> VectorType;
+
+  /** Default constructor without initialization */
+  EIGEN_DEVICE_FUNC inline ParametrizedLine() {}
+  
+  template<int OtherOptions>
+  EIGEN_DEVICE_FUNC ParametrizedLine(const ParametrizedLine<Scalar,AmbientDimAtCompileTime,OtherOptions>& other)
+   : m_origin(other.origin()), m_direction(other.direction())
+  {}
+
+  /** Constructs a dynamic-size line with \a _dim the dimension
+    * of the ambient space */
+  EIGEN_DEVICE_FUNC inline explicit ParametrizedLine(Index _dim) : m_origin(_dim), m_direction(_dim) {}
+
+  /** Initializes a parametrized line of direction \a direction and origin \a origin.
+    * \warning the vector direction is assumed to be normalized.
+    */
+  EIGEN_DEVICE_FUNC ParametrizedLine(const VectorType& origin, const VectorType& direction)
+    : m_origin(origin), m_direction(direction) {}
+
+  template <int OtherOptions>
+  EIGEN_DEVICE_FUNC explicit ParametrizedLine(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane);
+
+  /** Constructs a parametrized line going from \a p0 to \a p1. */
+  EIGEN_DEVICE_FUNC static inline ParametrizedLine Through(const VectorType& p0, const VectorType& p1)
+  { return ParametrizedLine(p0, (p1-p0).normalized()); }
+
+  EIGEN_DEVICE_FUNC ~ParametrizedLine() {}
+
+  /** \returns the dimension in which the line holds */
+  EIGEN_DEVICE_FUNC inline Index dim() const { return m_direction.size(); }
+
+  EIGEN_DEVICE_FUNC const VectorType& origin() const { return m_origin; }
+  EIGEN_DEVICE_FUNC VectorType& origin() { return m_origin; }
+
+  EIGEN_DEVICE_FUNC const VectorType& direction() const { return m_direction; }
+  EIGEN_DEVICE_FUNC VectorType& direction() { return m_direction; }
+
+  /** \returns the squared distance of a point \a p to its projection onto the line \c *this.
+    * \sa distance()
+    */
+  EIGEN_DEVICE_FUNC RealScalar squaredDistance(const VectorType& p) const
+  {
+    VectorType diff = p - origin();
+    return (diff - direction().dot(diff) * direction()).squaredNorm();
+  }
+  /** \returns the distance of a point \a p to its projection onto the line \c *this.
+    * \sa squaredDistance()
+    */
+  EIGEN_DEVICE_FUNC RealScalar distance(const VectorType& p) const { EIGEN_USING_STD_MATH(sqrt) return sqrt(squaredDistance(p)); }
+
+  /** \returns the projection of a point \a p onto the line \c *this. */
+  EIGEN_DEVICE_FUNC VectorType projection(const VectorType& p) const
+  { return origin() + direction().dot(p-origin()) * direction(); }
+
+  EIGEN_DEVICE_FUNC VectorType pointAt(const Scalar& t) const;
+  
+  template <int OtherOptions>
+  EIGEN_DEVICE_FUNC Scalar intersectionParameter(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const;
+ 
+  template <int OtherOptions>
+  EIGEN_DEVICE_FUNC Scalar intersection(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const;
+  
+  template <int OtherOptions>
+  EIGEN_DEVICE_FUNC VectorType intersectionPoint(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const;
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<ParametrizedLine,
+           ParametrizedLine<NewScalarType,AmbientDimAtCompileTime,Options> >::type cast() const
+  {
+    return typename internal::cast_return_type<ParametrizedLine,
+                    ParametrizedLine<NewScalarType,AmbientDimAtCompileTime,Options> >::type(*this);
+  }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType,int OtherOptions>
+  EIGEN_DEVICE_FUNC inline explicit ParametrizedLine(const ParametrizedLine<OtherScalarType,AmbientDimAtCompileTime,OtherOptions>& other)
+  {
+    m_origin = other.origin().template cast<Scalar>();
+    m_direction = other.direction().template cast<Scalar>();
+  }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const ParametrizedLine& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return m_origin.isApprox(other.m_origin, prec) && m_direction.isApprox(other.m_direction, prec); }
+
+protected:
+
+  VectorType m_origin, m_direction;
+};
+
+/** Constructs a parametrized line from a 2D hyperplane
+  *
+  * \warning the ambient space must have dimension 2 such that the hyperplane actually describes a line
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+template <int OtherOptions>
+EIGEN_DEVICE_FUNC inline ParametrizedLine<_Scalar, _AmbientDim,_Options>::ParametrizedLine(const Hyperplane<_Scalar, _AmbientDim,OtherOptions>& hyperplane)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(VectorType, 2)
+  direction() = hyperplane.normal().unitOrthogonal();
+  origin() = -hyperplane.normal()*hyperplane.offset();
+}
+
+/** \returns the point at \a t along this line
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+EIGEN_DEVICE_FUNC inline typename ParametrizedLine<_Scalar, _AmbientDim,_Options>::VectorType
+ParametrizedLine<_Scalar, _AmbientDim,_Options>::pointAt(const _Scalar& t) const
+{
+  return origin() + (direction()*t); 
+}
+
+/** \returns the parameter value of the intersection between \c *this and the given \a hyperplane
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+template <int OtherOptions>
+EIGEN_DEVICE_FUNC inline _Scalar ParametrizedLine<_Scalar, _AmbientDim,_Options>::intersectionParameter(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const
+{
+  return -(hyperplane.offset()+hyperplane.normal().dot(origin()))
+          / hyperplane.normal().dot(direction());
+}
+
+
+/** \deprecated use intersectionParameter()
+  * \returns the parameter value of the intersection between \c *this and the given \a hyperplane
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+template <int OtherOptions>
+EIGEN_DEVICE_FUNC inline _Scalar ParametrizedLine<_Scalar, _AmbientDim,_Options>::intersection(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const
+{
+  return intersectionParameter(hyperplane);
+}
+
+/** \returns the point of the intersection between \c *this and the given hyperplane
+  */
+template <typename _Scalar, int _AmbientDim, int _Options>
+template <int OtherOptions>
+EIGEN_DEVICE_FUNC inline typename ParametrizedLine<_Scalar, _AmbientDim,_Options>::VectorType
+ParametrizedLine<_Scalar, _AmbientDim,_Options>::intersectionPoint(const Hyperplane<_Scalar, _AmbientDim, OtherOptions>& hyperplane) const
+{
+  return pointAt(intersectionParameter(hyperplane));
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARAMETRIZEDLINE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Quaternion.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Quaternion.h
new file mode 100644
index 0000000..3e5a9ba
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Quaternion.h
@@ -0,0 +1,809 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Mathieu Gautier <mathieu.gautier@cea.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_QUATERNION_H
+#define EIGEN_QUATERNION_H
+namespace Eigen { 
+
+
+/***************************************************************************
+* Definition of QuaternionBase<Derived>
+* The implementation is at the end of the file
+***************************************************************************/
+
+namespace internal {
+template<typename Other,
+         int OtherRows=Other::RowsAtCompileTime,
+         int OtherCols=Other::ColsAtCompileTime>
+struct quaternionbase_assign_impl;
+}
+
+/** \geometry_module \ingroup Geometry_Module
+  * \class QuaternionBase
+  * \brief Base class for quaternion expressions
+  * \tparam Derived derived type (CRTP)
+  * \sa class Quaternion
+  */
+template<class Derived>
+class QuaternionBase : public RotationBase<Derived, 3>
+{
+ public:
+  typedef RotationBase<Derived, 3> Base;
+
+  using Base::operator*;
+  using Base::derived;
+
+  typedef typename internal::traits<Derived>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  typedef typename internal::traits<Derived>::Coefficients Coefficients;
+  enum {
+    Flags = Eigen::internal::traits<Derived>::Flags
+  };
+
+ // typedef typename Matrix<Scalar,4,1> Coefficients;
+  /** the type of a 3D vector */
+  typedef Matrix<Scalar,3,1> Vector3;
+  /** the equivalent rotation matrix type */
+  typedef Matrix<Scalar,3,3> Matrix3;
+  /** the equivalent angle-axis type */
+  typedef AngleAxis<Scalar> AngleAxisType;
+
+
+
+  /** \returns the \c x coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar x() const { return this->derived().coeffs().coeff(0); }
+  /** \returns the \c y coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar y() const { return this->derived().coeffs().coeff(1); }
+  /** \returns the \c z coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar z() const { return this->derived().coeffs().coeff(2); }
+  /** \returns the \c w coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar w() const { return this->derived().coeffs().coeff(3); }
+
+  /** \returns a reference to the \c x coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar& x() { return this->derived().coeffs().coeffRef(0); }
+  /** \returns a reference to the \c y coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar& y() { return this->derived().coeffs().coeffRef(1); }
+  /** \returns a reference to the \c z coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar& z() { return this->derived().coeffs().coeffRef(2); }
+  /** \returns a reference to the \c w coefficient */
+  EIGEN_DEVICE_FUNC inline Scalar& w() { return this->derived().coeffs().coeffRef(3); }
+
+  /** \returns a read-only vector expression of the imaginary part (x,y,z) */
+  EIGEN_DEVICE_FUNC inline const VectorBlock<const Coefficients,3> vec() const { return coeffs().template head<3>(); }
+
+  /** \returns a vector expression of the imaginary part (x,y,z) */
+  EIGEN_DEVICE_FUNC inline VectorBlock<Coefficients,3> vec() { return coeffs().template head<3>(); }
+
+  /** \returns a read-only vector expression of the coefficients (x,y,z,w) */
+  EIGEN_DEVICE_FUNC inline const typename internal::traits<Derived>::Coefficients& coeffs() const { return derived().coeffs(); }
+
+  /** \returns a vector expression of the coefficients (x,y,z,w) */
+  EIGEN_DEVICE_FUNC inline typename internal::traits<Derived>::Coefficients& coeffs() { return derived().coeffs(); }
+
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE QuaternionBase<Derived>& operator=(const QuaternionBase<Derived>& other);
+  template<class OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator=(const QuaternionBase<OtherDerived>& other);
+
+// disabled this copy operator as it is giving very strange compilation errors when compiling
+// test_stdvector with GCC 4.4.2. This looks like a GCC bug though, so feel free to re-enable it if it's
+// useful; however notice that we already have the templated operator= above and e.g. in MatrixBase
+// we didn't have to add, in addition to templated operator=, such a non-templated copy operator.
+//  Derived& operator=(const QuaternionBase& other)
+//  { return operator=<Derived>(other); }
+
+  EIGEN_DEVICE_FUNC Derived& operator=(const AngleAxisType& aa);
+  template<class OtherDerived> EIGEN_DEVICE_FUNC Derived& operator=(const MatrixBase<OtherDerived>& m);
+
+  /** \returns a quaternion representing an identity rotation
+    * \sa MatrixBase::Identity()
+    */
+  EIGEN_DEVICE_FUNC static inline Quaternion<Scalar> Identity() { return Quaternion<Scalar>(Scalar(1), Scalar(0), Scalar(0), Scalar(0)); }
+
+  /** \sa QuaternionBase::Identity(), MatrixBase::setIdentity()
+    */
+  EIGEN_DEVICE_FUNC inline QuaternionBase& setIdentity() { coeffs() << Scalar(0), Scalar(0), Scalar(0), Scalar(1); return *this; }
+
+  /** \returns the squared norm of the quaternion's coefficients
+    * \sa QuaternionBase::norm(), MatrixBase::squaredNorm()
+    */
+  EIGEN_DEVICE_FUNC inline Scalar squaredNorm() const { return coeffs().squaredNorm(); }
+
+  /** \returns the norm of the quaternion's coefficients
+    * \sa QuaternionBase::squaredNorm(), MatrixBase::norm()
+    */
+  EIGEN_DEVICE_FUNC inline Scalar norm() const { return coeffs().norm(); }
+
+  /** Normalizes the quaternion \c *this
+    * \sa normalized(), MatrixBase::normalize() */
+  EIGEN_DEVICE_FUNC inline void normalize() { coeffs().normalize(); }
+  /** \returns a normalized copy of \c *this
+    * \sa normalize(), MatrixBase::normalized() */
+  EIGEN_DEVICE_FUNC inline Quaternion<Scalar> normalized() const { return Quaternion<Scalar>(coeffs().normalized()); }
+
+    /** \returns the dot product of \c *this and \a other
+    * Geometrically speaking, the dot product of two unit quaternions
+    * corresponds to the cosine of half the angle between the two rotations.
+    * \sa angularDistance()
+    */
+  template<class OtherDerived> EIGEN_DEVICE_FUNC inline Scalar dot(const QuaternionBase<OtherDerived>& other) const { return coeffs().dot(other.coeffs()); }
+
+  template<class OtherDerived> EIGEN_DEVICE_FUNC Scalar angularDistance(const QuaternionBase<OtherDerived>& other) const;
+
+  /** \returns an equivalent 3x3 rotation matrix */
+  EIGEN_DEVICE_FUNC Matrix3 toRotationMatrix() const;
+
+  /** \returns the quaternion which transform \a a into \a b through a rotation */
+  template<typename Derived1, typename Derived2>
+  EIGEN_DEVICE_FUNC Derived& setFromTwoVectors(const MatrixBase<Derived1>& a, const MatrixBase<Derived2>& b);
+
+  template<class OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Quaternion<Scalar> operator* (const QuaternionBase<OtherDerived>& q) const;
+  template<class OtherDerived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& operator*= (const QuaternionBase<OtherDerived>& q);
+
+  /** \returns the quaternion describing the inverse rotation */
+  EIGEN_DEVICE_FUNC Quaternion<Scalar> inverse() const;
+
+  /** \returns the conjugated quaternion */
+  EIGEN_DEVICE_FUNC Quaternion<Scalar> conjugate() const;
+
+  template<class OtherDerived> EIGEN_DEVICE_FUNC Quaternion<Scalar> slerp(const Scalar& t, const QuaternionBase<OtherDerived>& other) const;
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  template<class OtherDerived>
+  EIGEN_DEVICE_FUNC bool isApprox(const QuaternionBase<OtherDerived>& other, const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return coeffs().isApprox(other.coeffs(), prec); }
+
+  /** return the result vector of \a v through the rotation*/
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Vector3 _transformVector(const Vector3& v) const;
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<Derived,Quaternion<NewScalarType> >::type cast() const
+  {
+    return typename internal::cast_return_type<Derived,Quaternion<NewScalarType> >::type(derived());
+  }
+
+#ifdef EIGEN_QUATERNIONBASE_PLUGIN
+# include EIGEN_QUATERNIONBASE_PLUGIN
+#endif
+};
+
+/***************************************************************************
+* Definition/implementation of Quaternion<Scalar>
+***************************************************************************/
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Quaternion
+  *
+  * \brief The quaternion class used to represent 3D orientations and rotations
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients
+  * \tparam _Options controls the memory alignment of the coefficients. Can be \# AutoAlign or \# DontAlign. Default is AutoAlign.
+  *
+  * This class represents a quaternion \f$ w+xi+yj+zk \f$ that is a convenient representation of
+  * orientations and rotations of objects in three dimensions. Compared to other representations
+  * like Euler angles or 3x3 matrices, quaternions offer the following advantages:
+  * \li \b compact storage (4 scalars)
+  * \li \b efficient to compose (28 flops),
+  * \li \b stable spherical interpolation
+  *
+  * The following two typedefs are provided for convenience:
+  * \li \c Quaternionf for \c float
+  * \li \c Quaterniond for \c double
+  *
+  * \warning Operations interpreting the quaternion as rotation have undefined behavior if the quaternion is not normalized.
+  *
+  * \sa  class AngleAxis, class Transform
+  */
+
+namespace internal {
+template<typename _Scalar,int _Options>
+struct traits<Quaternion<_Scalar,_Options> >
+{
+  typedef Quaternion<_Scalar,_Options> PlainObject;
+  typedef _Scalar Scalar;
+  typedef Matrix<_Scalar,4,1,_Options> Coefficients;
+  enum{
+    Alignment = internal::traits<Coefficients>::Alignment,
+    Flags = LvalueBit
+  };
+};
+}
+
+template<typename _Scalar, int _Options>
+class Quaternion : public QuaternionBase<Quaternion<_Scalar,_Options> >
+{
+public:
+  typedef QuaternionBase<Quaternion<_Scalar,_Options> > Base;
+  enum { NeedsAlignment = internal::traits<Quaternion>::Alignment>0 };
+
+  typedef _Scalar Scalar;
+
+  EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Quaternion)
+  using Base::operator*=;
+
+  typedef typename internal::traits<Quaternion>::Coefficients Coefficients;
+  typedef typename Base::AngleAxisType AngleAxisType;
+
+  /** Default constructor leaving the quaternion uninitialized. */
+  EIGEN_DEVICE_FUNC inline Quaternion() {}
+
+  /** Constructs and initializes the quaternion \f$ w+xi+yj+zk \f$ from
+    * its four coefficients \a w, \a x, \a y and \a z.
+    *
+    * \warning Note the order of the arguments: the real \a w coefficient first,
+    * while internally the coefficients are stored in the following order:
+    * [\c x, \c y, \c z, \c w]
+    */
+  EIGEN_DEVICE_FUNC inline Quaternion(const Scalar& w, const Scalar& x, const Scalar& y, const Scalar& z) : m_coeffs(x, y, z, w){}
+
+  /** Constructs and initialize a quaternion from the array data */
+  EIGEN_DEVICE_FUNC explicit inline Quaternion(const Scalar* data) : m_coeffs(data) {}
+
+  /** Copy constructor */
+  template<class Derived> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Quaternion(const QuaternionBase<Derived>& other) { this->Base::operator=(other); }
+
+  /** Constructs and initializes a quaternion from the angle-axis \a aa */
+  EIGEN_DEVICE_FUNC explicit inline Quaternion(const AngleAxisType& aa) { *this = aa; }
+
+  /** Constructs and initializes a quaternion from either:
+    *  - a rotation matrix expression,
+    *  - a 4D vector expression representing quaternion coefficients.
+    */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC explicit inline Quaternion(const MatrixBase<Derived>& other) { *this = other; }
+
+  /** Explicit copy constructor with scalar conversion */
+  template<typename OtherScalar, int OtherOptions>
+  EIGEN_DEVICE_FUNC explicit inline Quaternion(const Quaternion<OtherScalar, OtherOptions>& other)
+  { m_coeffs = other.coeffs().template cast<Scalar>(); }
+
+  EIGEN_DEVICE_FUNC static Quaternion UnitRandom();
+
+  template<typename Derived1, typename Derived2>
+  EIGEN_DEVICE_FUNC static Quaternion FromTwoVectors(const MatrixBase<Derived1>& a, const MatrixBase<Derived2>& b);
+
+  EIGEN_DEVICE_FUNC inline Coefficients& coeffs() { return m_coeffs;}
+  EIGEN_DEVICE_FUNC inline const Coefficients& coeffs() const { return m_coeffs;}
+
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(NeedsAlignment))
+  
+#ifdef EIGEN_QUATERNION_PLUGIN
+# include EIGEN_QUATERNION_PLUGIN
+#endif
+
+protected:
+  Coefficients m_coeffs;
+  
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    static EIGEN_STRONG_INLINE void _check_template_params()
+    {
+      EIGEN_STATIC_ASSERT( (_Options & DontAlign) == _Options,
+        INVALID_MATRIX_TEMPLATE_PARAMETERS)
+    }
+#endif
+};
+
+/** \ingroup Geometry_Module
+  * single precision quaternion type */
+typedef Quaternion<float> Quaternionf;
+/** \ingroup Geometry_Module
+  * double precision quaternion type */
+typedef Quaternion<double> Quaterniond;
+
+/***************************************************************************
+* Specialization of Map<Quaternion<Scalar>>
+***************************************************************************/
+
+namespace internal {
+  template<typename _Scalar, int _Options>
+  struct traits<Map<Quaternion<_Scalar>, _Options> > : traits<Quaternion<_Scalar, (int(_Options)&Aligned)==Aligned ? AutoAlign : DontAlign> >
+  {
+    typedef Map<Matrix<_Scalar,4,1>, _Options> Coefficients;
+  };
+}
+
+namespace internal {
+  template<typename _Scalar, int _Options>
+  struct traits<Map<const Quaternion<_Scalar>, _Options> > : traits<Quaternion<_Scalar, (int(_Options)&Aligned)==Aligned ? AutoAlign : DontAlign> >
+  {
+    typedef Map<const Matrix<_Scalar,4,1>, _Options> Coefficients;
+    typedef traits<Quaternion<_Scalar, (int(_Options)&Aligned)==Aligned ? AutoAlign : DontAlign> > TraitsBase;
+    enum {
+      Flags = TraitsBase::Flags & ~LvalueBit
+    };
+  };
+}
+
+/** \ingroup Geometry_Module
+  * \brief Quaternion expression mapping a constant memory buffer
+  *
+  * \tparam _Scalar the type of the Quaternion coefficients
+  * \tparam _Options see class Map
+  *
+  * This is a specialization of class Map for Quaternion. This class allows to view
+  * a 4 scalar memory buffer as an Eigen's Quaternion object.
+  *
+  * \sa class Map, class Quaternion, class QuaternionBase
+  */
+template<typename _Scalar, int _Options>
+class Map<const Quaternion<_Scalar>, _Options >
+  : public QuaternionBase<Map<const Quaternion<_Scalar>, _Options> >
+{
+  public:
+    typedef QuaternionBase<Map<const Quaternion<_Scalar>, _Options> > Base;
+
+    typedef _Scalar Scalar;
+    typedef typename internal::traits<Map>::Coefficients Coefficients;
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
+    using Base::operator*=;
+
+    /** Constructs a Mapped Quaternion object from the pointer \a coeffs
+      *
+      * The pointer \a coeffs must reference the four coefficients of Quaternion in the following order:
+      * \code *coeffs == {x, y, z, w} \endcode
+      *
+      * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */
+    EIGEN_DEVICE_FUNC explicit EIGEN_STRONG_INLINE Map(const Scalar* coeffs) : m_coeffs(coeffs) {}
+
+    EIGEN_DEVICE_FUNC inline const Coefficients& coeffs() const { return m_coeffs;}
+
+  protected:
+    const Coefficients m_coeffs;
+};
+
+/** \ingroup Geometry_Module
+  * \brief Expression of a quaternion from a memory buffer
+  *
+  * \tparam _Scalar the type of the Quaternion coefficients
+  * \tparam _Options see class Map
+  *
+  * This is a specialization of class Map for Quaternion. This class allows to view
+  * a 4 scalar memory buffer as an Eigen's  Quaternion object.
+  *
+  * \sa class Map, class Quaternion, class QuaternionBase
+  */
+template<typename _Scalar, int _Options>
+class Map<Quaternion<_Scalar>, _Options >
+  : public QuaternionBase<Map<Quaternion<_Scalar>, _Options> >
+{
+  public:
+    typedef QuaternionBase<Map<Quaternion<_Scalar>, _Options> > Base;
+
+    typedef _Scalar Scalar;
+    typedef typename internal::traits<Map>::Coefficients Coefficients;
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
+    using Base::operator*=;
+
+    /** Constructs a Mapped Quaternion object from the pointer \a coeffs
+      *
+      * The pointer \a coeffs must reference the four coefficients of Quaternion in the following order:
+      * \code *coeffs == {x, y, z, w} \endcode
+      *
+      * If the template parameter _Options is set to #Aligned, then the pointer coeffs must be aligned. */
+    EIGEN_DEVICE_FUNC explicit EIGEN_STRONG_INLINE Map(Scalar* coeffs) : m_coeffs(coeffs) {}
+
+    EIGEN_DEVICE_FUNC inline Coefficients& coeffs() { return m_coeffs; }
+    EIGEN_DEVICE_FUNC inline const Coefficients& coeffs() const { return m_coeffs; }
+
+  protected:
+    Coefficients m_coeffs;
+};
+
+/** \ingroup Geometry_Module
+  * Map an unaligned array of single precision scalars as a quaternion */
+typedef Map<Quaternion<float>, 0>         QuaternionMapf;
+/** \ingroup Geometry_Module
+  * Map an unaligned array of double precision scalars as a quaternion */
+typedef Map<Quaternion<double>, 0>        QuaternionMapd;
+/** \ingroup Geometry_Module
+  * Map a 16-byte aligned array of single precision scalars as a quaternion */
+typedef Map<Quaternion<float>, Aligned>   QuaternionMapAlignedf;
+/** \ingroup Geometry_Module
+  * Map a 16-byte aligned array of double precision scalars as a quaternion */
+typedef Map<Quaternion<double>, Aligned>  QuaternionMapAlignedd;
+
+/***************************************************************************
+* Implementation of QuaternionBase methods
+***************************************************************************/
+
+// Generic Quaternion * Quaternion product
+// This product can be specialized for a given architecture via the Arch template argument.
+namespace internal {
+template<int Arch, class Derived1, class Derived2, typename Scalar> struct quat_product
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion<Scalar> run(const QuaternionBase<Derived1>& a, const QuaternionBase<Derived2>& b){
+    return Quaternion<Scalar>
+    (
+      a.w() * b.w() - a.x() * b.x() - a.y() * b.y() - a.z() * b.z(),
+      a.w() * b.x() + a.x() * b.w() + a.y() * b.z() - a.z() * b.y(),
+      a.w() * b.y() + a.y() * b.w() + a.z() * b.x() - a.x() * b.z(),
+      a.w() * b.z() + a.z() * b.w() + a.x() * b.y() - a.y() * b.x()
+    );
+  }
+};
+}
+
+/** \returns the concatenation of two rotations as a quaternion-quaternion product */
+template <class Derived>
+template <class OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Quaternion<typename internal::traits<Derived>::Scalar>
+QuaternionBase<Derived>::operator* (const QuaternionBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<typename Derived::Scalar, typename OtherDerived::Scalar>::value),
+   YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+  return internal::quat_product<Architecture::Target, Derived, OtherDerived,
+                         typename internal::traits<Derived>::Scalar>::run(*this, other);
+}
+
+/** \sa operator*(Quaternion) */
+template <class Derived>
+template <class OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& QuaternionBase<Derived>::operator*= (const QuaternionBase<OtherDerived>& other)
+{
+  derived() = derived() * other.derived();
+  return derived();
+}
+
+/** Rotation of a vector by a quaternion.
+  * \remarks If the quaternion is used to rotate several points (>1)
+  * then it is much more efficient to first convert it to a 3x3 Matrix.
+  * Comparison of the operation cost for n transformations:
+  *   - Quaternion2:    30n
+  *   - Via a Matrix3: 24 + 15n
+  */
+template <class Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename QuaternionBase<Derived>::Vector3
+QuaternionBase<Derived>::_transformVector(const Vector3& v) const
+{
+    // Note that this algorithm comes from the optimization by hand
+    // of the conversion to a Matrix followed by a Matrix/Vector product.
+    // It appears to be much faster than the common algorithm found
+    // in the literature (30 versus 39 flops). It also requires two
+    // Vector3 as temporaries.
+    Vector3 uv = this->vec().cross(v);
+    uv += uv;
+    return v + this->w() * uv + this->vec().cross(uv);
+}
+
+template<class Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE QuaternionBase<Derived>& QuaternionBase<Derived>::operator=(const QuaternionBase<Derived>& other)
+{
+  coeffs() = other.coeffs();
+  return derived();
+}
+
+template<class Derived>
+template<class OtherDerived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& QuaternionBase<Derived>::operator=(const QuaternionBase<OtherDerived>& other)
+{
+  coeffs() = other.coeffs();
+  return derived();
+}
+
+/** Set \c *this from an angle-axis \a aa and returns a reference to \c *this
+  */
+template<class Derived>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Derived& QuaternionBase<Derived>::operator=(const AngleAxisType& aa)
+{
+  EIGEN_USING_STD_MATH(cos)
+  EIGEN_USING_STD_MATH(sin)
+  Scalar ha = Scalar(0.5)*aa.angle(); // Scalar(0.5) to suppress precision loss warnings
+  this->w() = cos(ha);
+  this->vec() = sin(ha) * aa.axis();
+  return derived();
+}
+
+/** Set \c *this from the expression \a xpr:
+  *   - if \a xpr is a 4x1 vector, then \a xpr is assumed to be a quaternion
+  *   - if \a xpr is a 3x3 matrix, then \a xpr is assumed to be rotation matrix
+  *     and \a xpr is converted to a quaternion
+  */
+
+template<class Derived>
+template<class MatrixDerived>
+EIGEN_DEVICE_FUNC inline Derived& QuaternionBase<Derived>::operator=(const MatrixBase<MatrixDerived>& xpr)
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<typename Derived::Scalar, typename MatrixDerived::Scalar>::value),
+   YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+  internal::quaternionbase_assign_impl<MatrixDerived>::run(*this, xpr.derived());
+  return derived();
+}
+
+/** Convert the quaternion to a 3x3 rotation matrix. The quaternion is required to
+  * be normalized, otherwise the result is undefined.
+  */
+template<class Derived>
+EIGEN_DEVICE_FUNC inline typename QuaternionBase<Derived>::Matrix3
+QuaternionBase<Derived>::toRotationMatrix(void) const
+{
+  // NOTE if inlined, then gcc 4.2 and 4.4 get rid of the temporary (not gcc 4.3 !!)
+  // if not inlined then the cost of the return by value is huge ~ +35%,
+  // however, not inlining this function is an order of magnitude slower, so
+  // it has to be inlined, and so the return by value is not an issue
+  Matrix3 res;
+
+  const Scalar tx  = Scalar(2)*this->x();
+  const Scalar ty  = Scalar(2)*this->y();
+  const Scalar tz  = Scalar(2)*this->z();
+  const Scalar twx = tx*this->w();
+  const Scalar twy = ty*this->w();
+  const Scalar twz = tz*this->w();
+  const Scalar txx = tx*this->x();
+  const Scalar txy = ty*this->x();
+  const Scalar txz = tz*this->x();
+  const Scalar tyy = ty*this->y();
+  const Scalar tyz = tz*this->y();
+  const Scalar tzz = tz*this->z();
+
+  res.coeffRef(0,0) = Scalar(1)-(tyy+tzz);
+  res.coeffRef(0,1) = txy-twz;
+  res.coeffRef(0,2) = txz+twy;
+  res.coeffRef(1,0) = txy+twz;
+  res.coeffRef(1,1) = Scalar(1)-(txx+tzz);
+  res.coeffRef(1,2) = tyz-twx;
+  res.coeffRef(2,0) = txz-twy;
+  res.coeffRef(2,1) = tyz+twx;
+  res.coeffRef(2,2) = Scalar(1)-(txx+tyy);
+
+  return res;
+}
+
+/** Sets \c *this to be a quaternion representing a rotation between
+  * the two arbitrary vectors \a a and \a b. In other words, the built
+  * rotation represent a rotation sending the line of direction \a a
+  * to the line of direction \a b, both lines passing through the origin.
+  *
+  * \returns a reference to \c *this.
+  *
+  * Note that the two input vectors do \b not have to be normalized, and
+  * do not need to have the same norm.
+  */
+template<class Derived>
+template<typename Derived1, typename Derived2>
+EIGEN_DEVICE_FUNC inline Derived& QuaternionBase<Derived>::setFromTwoVectors(const MatrixBase<Derived1>& a, const MatrixBase<Derived2>& b)
+{
+  EIGEN_USING_STD_MATH(sqrt)
+  Vector3 v0 = a.normalized();
+  Vector3 v1 = b.normalized();
+  Scalar c = v1.dot(v0);
+
+  // if dot == -1, vectors are nearly opposites
+  // => accurately compute the rotation axis by computing the
+  //    intersection of the two planes. This is done by solving:
+  //       x^T v0 = 0
+  //       x^T v1 = 0
+  //    under the constraint:
+  //       ||x|| = 1
+  //    which yields a singular value problem
+  if (c < Scalar(-1)+NumTraits<Scalar>::dummy_precision())
+  {
+    c = numext::maxi(c,Scalar(-1));
+    Matrix<Scalar,2,3> m; m << v0.transpose(), v1.transpose();
+    JacobiSVD<Matrix<Scalar,2,3> > svd(m, ComputeFullV);
+    Vector3 axis = svd.matrixV().col(2);
+
+    Scalar w2 = (Scalar(1)+c)*Scalar(0.5);
+    this->w() = sqrt(w2);
+    this->vec() = axis * sqrt(Scalar(1) - w2);
+    return derived();
+  }
+  Vector3 axis = v0.cross(v1);
+  Scalar s = sqrt((Scalar(1)+c)*Scalar(2));
+  Scalar invs = Scalar(1)/s;
+  this->vec() = axis * invs;
+  this->w() = s * Scalar(0.5);
+
+  return derived();
+}
+
+/** \returns a random unit quaternion following a uniform distribution law on SO(3)
+  *
+  * \note The implementation is based on http://planning.cs.uiuc.edu/node198.html
+  */
+template<typename Scalar, int Options>
+EIGEN_DEVICE_FUNC Quaternion<Scalar,Options> Quaternion<Scalar,Options>::UnitRandom()
+{
+  EIGEN_USING_STD_MATH(sqrt)
+  EIGEN_USING_STD_MATH(sin)
+  EIGEN_USING_STD_MATH(cos)
+  const Scalar u1 = internal::random<Scalar>(0, 1),
+               u2 = internal::random<Scalar>(0, 2*EIGEN_PI),
+               u3 = internal::random<Scalar>(0, 2*EIGEN_PI);
+  const Scalar a = sqrt(1 - u1),
+               b = sqrt(u1);
+  return Quaternion (a * sin(u2), a * cos(u2), b * sin(u3), b * cos(u3));
+}
+
+
+/** Returns a quaternion representing a rotation between
+  * the two arbitrary vectors \a a and \a b. In other words, the built
+  * rotation represent a rotation sending the line of direction \a a
+  * to the line of direction \a b, both lines passing through the origin.
+  *
+  * \returns resulting quaternion
+  *
+  * Note that the two input vectors do \b not have to be normalized, and
+  * do not need to have the same norm.
+  */
+template<typename Scalar, int Options>
+template<typename Derived1, typename Derived2>
+EIGEN_DEVICE_FUNC Quaternion<Scalar,Options> Quaternion<Scalar,Options>::FromTwoVectors(const MatrixBase<Derived1>& a, const MatrixBase<Derived2>& b)
+{
+    Quaternion quat;
+    quat.setFromTwoVectors(a, b);
+    return quat;
+}
+
+
+/** \returns the multiplicative inverse of \c *this
+  * Note that in most cases, i.e., if you simply want the opposite rotation,
+  * and/or the quaternion is normalized, then it is enough to use the conjugate.
+  *
+  * \sa QuaternionBase::conjugate()
+  */
+template <class Derived>
+EIGEN_DEVICE_FUNC inline Quaternion<typename internal::traits<Derived>::Scalar> QuaternionBase<Derived>::inverse() const
+{
+  // FIXME should this function be called multiplicativeInverse and conjugate() be called inverse() or opposite()  ??
+  Scalar n2 = this->squaredNorm();
+  if (n2 > Scalar(0))
+    return Quaternion<Scalar>(conjugate().coeffs() / n2);
+  else
+  {
+    // return an invalid result to flag the error
+    return Quaternion<Scalar>(Coefficients::Zero());
+  }
+}
+
+// Generic conjugate of a Quaternion
+namespace internal {
+template<int Arch, class Derived, typename Scalar> struct quat_conj
+{
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE Quaternion<Scalar> run(const QuaternionBase<Derived>& q){
+    return Quaternion<Scalar>(q.w(),-q.x(),-q.y(),-q.z());
+  }
+};
+}
+                         
+/** \returns the conjugate of the \c *this which is equal to the multiplicative inverse
+  * if the quaternion is normalized.
+  * The conjugate of a quaternion represents the opposite rotation.
+  *
+  * \sa Quaternion2::inverse()
+  */
+template <class Derived>
+EIGEN_DEVICE_FUNC inline Quaternion<typename internal::traits<Derived>::Scalar>
+QuaternionBase<Derived>::conjugate() const
+{
+  return internal::quat_conj<Architecture::Target, Derived,
+                         typename internal::traits<Derived>::Scalar>::run(*this);
+                         
+}
+
+/** \returns the angle (in radian) between two rotations
+  * \sa dot()
+  */
+template <class Derived>
+template <class OtherDerived>
+EIGEN_DEVICE_FUNC inline typename internal::traits<Derived>::Scalar
+QuaternionBase<Derived>::angularDistance(const QuaternionBase<OtherDerived>& other) const
+{
+  EIGEN_USING_STD_MATH(atan2)
+  Quaternion<Scalar> d = (*this) * other.conjugate();
+  return Scalar(2) * atan2( d.vec().norm(), numext::abs(d.w()) );
+}
+
+ 
+    
+/** \returns the spherical linear interpolation between the two quaternions
+  * \c *this and \a other at the parameter \a t in [0;1].
+  * 
+  * This represents an interpolation for a constant motion between \c *this and \a other,
+  * see also http://en.wikipedia.org/wiki/Slerp.
+  */
+template <class Derived>
+template <class OtherDerived>
+EIGEN_DEVICE_FUNC Quaternion<typename internal::traits<Derived>::Scalar>
+QuaternionBase<Derived>::slerp(const Scalar& t, const QuaternionBase<OtherDerived>& other) const
+{
+  EIGEN_USING_STD_MATH(acos)
+  EIGEN_USING_STD_MATH(sin)
+  const Scalar one = Scalar(1) - NumTraits<Scalar>::epsilon();
+  Scalar d = this->dot(other);
+  Scalar absD = numext::abs(d);
+
+  Scalar scale0;
+  Scalar scale1;
+
+  if(absD>=one)
+  {
+    scale0 = Scalar(1) - t;
+    scale1 = t;
+  }
+  else
+  {
+    // theta is the angle between the 2 quaternions
+    Scalar theta = acos(absD);
+    Scalar sinTheta = sin(theta);
+
+    scale0 = sin( ( Scalar(1) - t ) * theta) / sinTheta;
+    scale1 = sin( ( t * theta) ) / sinTheta;
+  }
+  if(d<Scalar(0)) scale1 = -scale1;
+
+  return Quaternion<Scalar>(scale0 * coeffs() + scale1 * other.coeffs());
+}
+
+namespace internal {
+
+// set from a rotation matrix
+template<typename Other>
+struct quaternionbase_assign_impl<Other,3,3>
+{
+  typedef typename Other::Scalar Scalar;
+  template<class Derived> EIGEN_DEVICE_FUNC static inline void run(QuaternionBase<Derived>& q, const Other& a_mat)
+  {
+    const typename internal::nested_eval<Other,2>::type mat(a_mat);
+    EIGEN_USING_STD_MATH(sqrt)
+    // This algorithm comes from  "Quaternion Calculus and Fast Animation",
+    // Ken Shoemake, 1987 SIGGRAPH course notes
+    Scalar t = mat.trace();
+    if (t > Scalar(0))
+    {
+      t = sqrt(t + Scalar(1.0));
+      q.w() = Scalar(0.5)*t;
+      t = Scalar(0.5)/t;
+      q.x() = (mat.coeff(2,1) - mat.coeff(1,2)) * t;
+      q.y() = (mat.coeff(0,2) - mat.coeff(2,0)) * t;
+      q.z() = (mat.coeff(1,0) - mat.coeff(0,1)) * t;
+    }
+    else
+    {
+      Index i = 0;
+      if (mat.coeff(1,1) > mat.coeff(0,0))
+        i = 1;
+      if (mat.coeff(2,2) > mat.coeff(i,i))
+        i = 2;
+      Index j = (i+1)%3;
+      Index k = (j+1)%3;
+
+      t = sqrt(mat.coeff(i,i)-mat.coeff(j,j)-mat.coeff(k,k) + Scalar(1.0));
+      q.coeffs().coeffRef(i) = Scalar(0.5) * t;
+      t = Scalar(0.5)/t;
+      q.w() = (mat.coeff(k,j)-mat.coeff(j,k))*t;
+      q.coeffs().coeffRef(j) = (mat.coeff(j,i)+mat.coeff(i,j))*t;
+      q.coeffs().coeffRef(k) = (mat.coeff(k,i)+mat.coeff(i,k))*t;
+    }
+  }
+};
+
+// set from a vector of coefficients assumed to be a quaternion
+template<typename Other>
+struct quaternionbase_assign_impl<Other,4,1>
+{
+  typedef typename Other::Scalar Scalar;
+  template<class Derived> EIGEN_DEVICE_FUNC static inline void run(QuaternionBase<Derived>& q, const Other& vec)
+  {
+    q.coeffs() = vec;
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_QUATERNION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Rotation2D.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Rotation2D.h
new file mode 100644
index 0000000..884b7d0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Rotation2D.h
@@ -0,0 +1,199 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ROTATION2D_H
+#define EIGEN_ROTATION2D_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Rotation2D
+  *
+  * \brief Represents a rotation/orientation in a 2 dimensional space.
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients
+  *
+  * This class is equivalent to a single scalar representing a counter clock wise rotation
+  * as a single angle in radian. It provides some additional features such as the automatic
+  * conversion from/to a 2x2 rotation matrix. Moreover this class aims to provide a similar
+  * interface to Quaternion in order to facilitate the writing of generic algorithms
+  * dealing with rotations.
+  *
+  * \sa class Quaternion, class Transform
+  */
+
+namespace internal {
+
+template<typename _Scalar> struct traits<Rotation2D<_Scalar> >
+{
+  typedef _Scalar Scalar;
+};
+} // end namespace internal
+
+template<typename _Scalar>
+class Rotation2D : public RotationBase<Rotation2D<_Scalar>,2>
+{
+  typedef RotationBase<Rotation2D<_Scalar>,2> Base;
+
+public:
+
+  using Base::operator*;
+
+  enum { Dim = 2 };
+  /** the scalar type of the coefficients */
+  typedef _Scalar Scalar;
+  typedef Matrix<Scalar,2,1> Vector2;
+  typedef Matrix<Scalar,2,2> Matrix2;
+
+protected:
+
+  Scalar m_angle;
+
+public:
+
+  /** Construct a 2D counter clock wise rotation from the angle \a a in radian. */
+  EIGEN_DEVICE_FUNC explicit inline Rotation2D(const Scalar& a) : m_angle(a) {}
+  
+  /** Default constructor wihtout initialization. The represented rotation is undefined. */
+  EIGEN_DEVICE_FUNC Rotation2D() {}
+
+  /** Construct a 2D rotation from a 2x2 rotation matrix \a mat.
+    *
+    * \sa fromRotationMatrix()
+    */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC explicit Rotation2D(const MatrixBase<Derived>& m)
+  {
+    fromRotationMatrix(m.derived());
+  }
+
+  /** \returns the rotation angle */
+  EIGEN_DEVICE_FUNC inline Scalar angle() const { return m_angle; }
+
+  /** \returns a read-write reference to the rotation angle */
+  EIGEN_DEVICE_FUNC inline Scalar& angle() { return m_angle; }
+  
+  /** \returns the rotation angle in [0,2pi] */
+  EIGEN_DEVICE_FUNC inline Scalar smallestPositiveAngle() const {
+    Scalar tmp = numext::fmod(m_angle,Scalar(2*EIGEN_PI));
+    return tmp<Scalar(0) ? tmp + Scalar(2*EIGEN_PI) : tmp;
+  }
+  
+  /** \returns the rotation angle in [-pi,pi] */
+  EIGEN_DEVICE_FUNC inline Scalar smallestAngle() const {
+    Scalar tmp = numext::fmod(m_angle,Scalar(2*EIGEN_PI));
+    if(tmp>Scalar(EIGEN_PI))       tmp -= Scalar(2*EIGEN_PI);
+    else if(tmp<-Scalar(EIGEN_PI)) tmp += Scalar(2*EIGEN_PI);
+    return tmp;
+  }
+
+  /** \returns the inverse rotation */
+  EIGEN_DEVICE_FUNC inline Rotation2D inverse() const { return Rotation2D(-m_angle); }
+
+  /** Concatenates two rotations */
+  EIGEN_DEVICE_FUNC inline Rotation2D operator*(const Rotation2D& other) const
+  { return Rotation2D(m_angle + other.m_angle); }
+
+  /** Concatenates two rotations */
+  EIGEN_DEVICE_FUNC inline Rotation2D& operator*=(const Rotation2D& other)
+  { m_angle += other.m_angle; return *this; }
+
+  /** Applies the rotation to a 2D vector */
+  EIGEN_DEVICE_FUNC Vector2 operator* (const Vector2& vec) const
+  { return toRotationMatrix() * vec; }
+  
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC Rotation2D& fromRotationMatrix(const MatrixBase<Derived>& m);
+  EIGEN_DEVICE_FUNC Matrix2 toRotationMatrix() const;
+
+  /** Set \c *this from a 2x2 rotation matrix \a mat.
+    * In other words, this function extract the rotation angle from the rotation matrix.
+    *
+    * This method is an alias for fromRotationMatrix()
+    *
+    * \sa fromRotationMatrix()
+    */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC Rotation2D& operator=(const  MatrixBase<Derived>& m)
+  { return fromRotationMatrix(m.derived()); }
+
+  /** \returns the spherical interpolation between \c *this and \a other using
+    * parameter \a t. It is in fact equivalent to a linear interpolation.
+    */
+  EIGEN_DEVICE_FUNC inline Rotation2D slerp(const Scalar& t, const Rotation2D& other) const
+  {
+    Scalar dist = Rotation2D(other.m_angle-m_angle).smallestAngle();
+    return Rotation2D(m_angle + dist*t);
+  }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<Rotation2D,Rotation2D<NewScalarType> >::type cast() const
+  { return typename internal::cast_return_type<Rotation2D,Rotation2D<NewScalarType> >::type(*this); }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  EIGEN_DEVICE_FUNC inline explicit Rotation2D(const Rotation2D<OtherScalarType>& other)
+  {
+    m_angle = Scalar(other.angle());
+  }
+
+  EIGEN_DEVICE_FUNC static inline Rotation2D Identity() { return Rotation2D(0); }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const Rotation2D& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return internal::isApprox(m_angle,other.m_angle, prec); }
+  
+};
+
+/** \ingroup Geometry_Module
+  * single precision 2D rotation type */
+typedef Rotation2D<float> Rotation2Df;
+/** \ingroup Geometry_Module
+  * double precision 2D rotation type */
+typedef Rotation2D<double> Rotation2Dd;
+
+/** Set \c *this from a 2x2 rotation matrix \a mat.
+  * In other words, this function extract the rotation angle
+  * from the rotation matrix.
+  */
+template<typename Scalar>
+template<typename Derived>
+EIGEN_DEVICE_FUNC Rotation2D<Scalar>& Rotation2D<Scalar>::fromRotationMatrix(const MatrixBase<Derived>& mat)
+{
+  EIGEN_USING_STD_MATH(atan2)
+  EIGEN_STATIC_ASSERT(Derived::RowsAtCompileTime==2 && Derived::ColsAtCompileTime==2,YOU_MADE_A_PROGRAMMING_MISTAKE)
+  m_angle = atan2(mat.coeff(1,0), mat.coeff(0,0));
+  return *this;
+}
+
+/** Constructs and \returns an equivalent 2x2 rotation matrix.
+  */
+template<typename Scalar>
+typename Rotation2D<Scalar>::Matrix2
+EIGEN_DEVICE_FUNC Rotation2D<Scalar>::toRotationMatrix(void) const
+{
+  EIGEN_USING_STD_MATH(sin)
+  EIGEN_USING_STD_MATH(cos)
+  Scalar sinA = sin(m_angle);
+  Scalar cosA = cos(m_angle);
+  return (Matrix2() << cosA, -sinA, sinA, cosA).finished();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_ROTATION2D_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/RotationBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/RotationBase.h
new file mode 100644
index 0000000..f0ee0bd
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/RotationBase.h
@@ -0,0 +1,206 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ROTATIONBASE_H
+#define EIGEN_ROTATIONBASE_H
+
+namespace Eigen { 
+
+// forward declaration
+namespace internal {
+template<typename RotationDerived, typename MatrixType, bool IsVector=MatrixType::IsVectorAtCompileTime>
+struct rotation_base_generic_product_selector;
+}
+
+/** \class RotationBase
+  *
+  * \brief Common base class for compact rotation representations
+  *
+  * \tparam Derived is the derived type, i.e., a rotation type
+  * \tparam _Dim the dimension of the space
+  */
+template<typename Derived, int _Dim>
+class RotationBase
+{
+  public:
+    enum { Dim = _Dim };
+    /** the scalar type of the coefficients */
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+
+    /** corresponding linear transformation matrix type */
+    typedef Matrix<Scalar,Dim,Dim> RotationMatrixType;
+    typedef Matrix<Scalar,Dim,1> VectorType;
+
+  public:
+    EIGEN_DEVICE_FUNC inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    EIGEN_DEVICE_FUNC inline Derived& derived() { return *static_cast<Derived*>(this); }
+
+    /** \returns an equivalent rotation matrix */
+    EIGEN_DEVICE_FUNC inline RotationMatrixType toRotationMatrix() const { return derived().toRotationMatrix(); }
+
+    /** \returns an equivalent rotation matrix 
+      * This function is added to be conform with the Transform class' naming scheme.
+      */
+    EIGEN_DEVICE_FUNC inline RotationMatrixType matrix() const { return derived().toRotationMatrix(); }
+
+    /** \returns the inverse rotation */
+    EIGEN_DEVICE_FUNC inline Derived inverse() const { return derived().inverse(); }
+
+    /** \returns the concatenation of the rotation \c *this with a translation \a t */
+    EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Isometry> operator*(const Translation<Scalar,Dim>& t) const
+    { return Transform<Scalar,Dim,Isometry>(*this) * t; }
+
+    /** \returns the concatenation of the rotation \c *this with a uniform scaling \a s */
+    EIGEN_DEVICE_FUNC inline RotationMatrixType operator*(const UniformScaling<Scalar>& s) const
+    { return toRotationMatrix() * s.factor(); }
+
+    /** \returns the concatenation of the rotation \c *this with a generic expression \a e
+      * \a e can be:
+      *  - a DimxDim linear transformation matrix
+      *  - a DimxDim diagonal matrix (axis aligned scaling)
+      *  - a vector of size Dim
+      */
+    template<typename OtherDerived>
+    EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename internal::rotation_base_generic_product_selector<Derived,OtherDerived,OtherDerived::IsVectorAtCompileTime>::ReturnType
+    operator*(const EigenBase<OtherDerived>& e) const
+    { return internal::rotation_base_generic_product_selector<Derived,OtherDerived>::run(derived(), e.derived()); }
+
+    /** \returns the concatenation of a linear transformation \a l with the rotation \a r */
+    template<typename OtherDerived> friend
+    EIGEN_DEVICE_FUNC inline RotationMatrixType operator*(const EigenBase<OtherDerived>& l, const Derived& r)
+    { return l.derived() * r.toRotationMatrix(); }
+
+    /** \returns the concatenation of a scaling \a l with the rotation \a r */
+    EIGEN_DEVICE_FUNC friend inline Transform<Scalar,Dim,Affine> operator*(const DiagonalMatrix<Scalar,Dim>& l, const Derived& r)
+    { 
+      Transform<Scalar,Dim,Affine> res(r);
+      res.linear().applyOnTheLeft(l);
+      return res;
+    }
+
+    /** \returns the concatenation of the rotation \c *this with a transformation \a t */
+    template<int Mode, int Options>
+    EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode> operator*(const Transform<Scalar,Dim,Mode,Options>& t) const
+    { return toRotationMatrix() * t; }
+
+    template<typename OtherVectorType>
+    EIGEN_DEVICE_FUNC inline VectorType _transformVector(const OtherVectorType& v) const
+    { return toRotationMatrix() * v; }
+};
+
+namespace internal {
+
+// implementation of the generic product rotation * matrix
+template<typename RotationDerived, typename MatrixType>
+struct rotation_base_generic_product_selector<RotationDerived,MatrixType,false>
+{
+  enum { Dim = RotationDerived::Dim };
+  typedef Matrix<typename RotationDerived::Scalar,Dim,Dim> ReturnType;
+  EIGEN_DEVICE_FUNC static inline ReturnType run(const RotationDerived& r, const MatrixType& m)
+  { return r.toRotationMatrix() * m; }
+};
+
+template<typename RotationDerived, typename Scalar, int Dim, int MaxDim>
+struct rotation_base_generic_product_selector< RotationDerived, DiagonalMatrix<Scalar,Dim,MaxDim>, false >
+{
+  typedef Transform<Scalar,Dim,Affine> ReturnType;
+  EIGEN_DEVICE_FUNC static inline ReturnType run(const RotationDerived& r, const DiagonalMatrix<Scalar,Dim,MaxDim>& m)
+  {
+    ReturnType res(r);
+    res.linear() *= m;
+    return res;
+  }
+};
+
+template<typename RotationDerived,typename OtherVectorType>
+struct rotation_base_generic_product_selector<RotationDerived,OtherVectorType,true>
+{
+  enum { Dim = RotationDerived::Dim };
+  typedef Matrix<typename RotationDerived::Scalar,Dim,1> ReturnType;
+  EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE ReturnType run(const RotationDerived& r, const OtherVectorType& v)
+  {
+    return r._transformVector(v);
+  }
+};
+
+} // end namespace internal
+
+/** \geometry_module
+  *
+  * \brief Constructs a Dim x Dim rotation matrix from the rotation \a r
+  */
+template<typename _Scalar, int _Rows, int _Cols, int _Storage, int _MaxRows, int _MaxCols>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols>
+::Matrix(const RotationBase<OtherDerived,ColsAtCompileTime>& r)
+{
+  EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix,int(OtherDerived::Dim),int(OtherDerived::Dim))
+  *this = r.toRotationMatrix();
+}
+
+/** \geometry_module
+  *
+  * \brief Set a Dim x Dim rotation matrix from the rotation \a r
+  */
+template<typename _Scalar, int _Rows, int _Cols, int _Storage, int _MaxRows, int _MaxCols>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols>&
+Matrix<_Scalar, _Rows, _Cols, _Storage, _MaxRows, _MaxCols>
+::operator=(const RotationBase<OtherDerived,ColsAtCompileTime>& r)
+{
+  EIGEN_STATIC_ASSERT_MATRIX_SPECIFIC_SIZE(Matrix,int(OtherDerived::Dim),int(OtherDerived::Dim))
+  return *this = r.toRotationMatrix();
+}
+
+namespace internal {
+
+/** \internal
+  *
+  * Helper function to return an arbitrary rotation object to a rotation matrix.
+  *
+  * \tparam Scalar the numeric type of the matrix coefficients
+  * \tparam Dim the dimension of the current space
+  *
+  * It returns a Dim x Dim fixed size matrix.
+  *
+  * Default specializations are provided for:
+  *   - any scalar type (2D),
+  *   - any matrix expression,
+  *   - any type based on RotationBase (e.g., Quaternion, AngleAxis, Rotation2D)
+  *
+  * Currently toRotationMatrix is only used by Transform.
+  *
+  * \sa class Transform, class Rotation2D, class Quaternion, class AngleAxis
+  */
+template<typename Scalar, int Dim>
+EIGEN_DEVICE_FUNC static inline Matrix<Scalar,2,2> toRotationMatrix(const Scalar& s)
+{
+  EIGEN_STATIC_ASSERT(Dim==2,YOU_MADE_A_PROGRAMMING_MISTAKE)
+  return Rotation2D<Scalar>(s).toRotationMatrix();
+}
+
+template<typename Scalar, int Dim, typename OtherDerived>
+EIGEN_DEVICE_FUNC static inline Matrix<Scalar,Dim,Dim> toRotationMatrix(const RotationBase<OtherDerived,Dim>& r)
+{
+  return r.toRotationMatrix();
+}
+
+template<typename Scalar, int Dim, typename OtherDerived>
+EIGEN_DEVICE_FUNC static inline const MatrixBase<OtherDerived>& toRotationMatrix(const MatrixBase<OtherDerived>& mat)
+{
+  EIGEN_STATIC_ASSERT(OtherDerived::RowsAtCompileTime==Dim && OtherDerived::ColsAtCompileTime==Dim,
+    YOU_MADE_A_PROGRAMMING_MISTAKE)
+  return mat;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_ROTATIONBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Scaling.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Scaling.h
new file mode 100755
index 0000000..f58ca03
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Scaling.h
@@ -0,0 +1,170 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SCALING_H
+#define EIGEN_SCALING_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Scaling
+  *
+  * \brief Represents a generic uniform scaling transformation
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients.
+  *
+  * This class represent a uniform scaling transformation. It is the return
+  * type of Scaling(Scalar), and most of the time this is the only way it
+  * is used. In particular, this class is not aimed to be used to store a scaling transformation,
+  * but rather to make easier the constructions and updates of Transform objects.
+  *
+  * To represent an axis aligned scaling, use the DiagonalMatrix class.
+  *
+  * \sa Scaling(), class DiagonalMatrix, MatrixBase::asDiagonal(), class Translation, class Transform
+  */
+template<typename _Scalar>
+class UniformScaling
+{
+public:
+  /** the scalar type of the coefficients */
+  typedef _Scalar Scalar;
+
+protected:
+
+  Scalar m_factor;
+
+public:
+
+  /** Default constructor without initialization. */
+  UniformScaling() {}
+  /** Constructs and initialize a uniform scaling transformation */
+  explicit inline UniformScaling(const Scalar& s) : m_factor(s) {}
+
+  inline const Scalar& factor() const { return m_factor; }
+  inline Scalar& factor() { return m_factor; }
+
+  /** Concatenates two uniform scaling */
+  inline UniformScaling operator* (const UniformScaling& other) const
+  { return UniformScaling(m_factor * other.factor()); }
+
+  /** Concatenates a uniform scaling and a translation */
+  template<int Dim>
+  inline Transform<Scalar,Dim,Affine> operator* (const Translation<Scalar,Dim>& t) const;
+
+  /** Concatenates a uniform scaling and an affine transformation */
+  template<int Dim, int Mode, int Options>
+  inline Transform<Scalar,Dim,(int(Mode)==int(Isometry)?Affine:Mode)> operator* (const Transform<Scalar,Dim, Mode, Options>& t) const
+  {
+    Transform<Scalar,Dim,(int(Mode)==int(Isometry)?Affine:Mode)> res = t;
+    res.prescale(factor());
+    return res;
+  }
+
+  /** Concatenates a uniform scaling and a linear transformation matrix */
+  // TODO returns an expression
+  template<typename Derived>
+  inline typename internal::plain_matrix_type<Derived>::type operator* (const MatrixBase<Derived>& other) const
+  { return other * m_factor; }
+
+  template<typename Derived,int Dim>
+  inline Matrix<Scalar,Dim,Dim> operator*(const RotationBase<Derived,Dim>& r) const
+  { return r.toRotationMatrix() * m_factor; }
+
+  /** \returns the inverse scaling */
+  inline UniformScaling inverse() const
+  { return UniformScaling(Scalar(1)/m_factor); }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  inline UniformScaling<NewScalarType> cast() const
+  { return UniformScaling<NewScalarType>(NewScalarType(m_factor)); }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  inline explicit UniformScaling(const UniformScaling<OtherScalarType>& other)
+  { m_factor = Scalar(other.factor()); }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  bool isApprox(const UniformScaling& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return internal::isApprox(m_factor, other.factor(), prec); }
+
+};
+
+/** \addtogroup Geometry_Module */
+//@{
+
+/** Concatenates a linear transformation matrix and a uniform scaling
+  * \relates UniformScaling
+  */
+// NOTE this operator is defiend in MatrixBase and not as a friend function
+// of UniformScaling to fix an internal crash of Intel's ICC
+template<typename Derived,typename Scalar>
+EIGEN_EXPR_BINARYOP_SCALAR_RETURN_TYPE(Derived,Scalar,product)
+operator*(const MatrixBase<Derived>& matrix, const UniformScaling<Scalar>& s)
+{ return matrix.derived() * s.factor(); }
+
+/** Constructs a uniform scaling from scale factor \a s */
+inline UniformScaling<float> Scaling(float s) { return UniformScaling<float>(s); }
+/** Constructs a uniform scaling from scale factor \a s */
+inline UniformScaling<double> Scaling(double s) { return UniformScaling<double>(s); }
+/** Constructs a uniform scaling from scale factor \a s */
+template<typename RealScalar>
+inline UniformScaling<std::complex<RealScalar> > Scaling(const std::complex<RealScalar>& s)
+{ return UniformScaling<std::complex<RealScalar> >(s); }
+
+/** Constructs a 2D axis aligned scaling */
+template<typename Scalar>
+inline DiagonalMatrix<Scalar,2> Scaling(const Scalar& sx, const Scalar& sy)
+{ return DiagonalMatrix<Scalar,2>(sx, sy); }
+/** Constructs a 3D axis aligned scaling */
+template<typename Scalar>
+inline DiagonalMatrix<Scalar,3> Scaling(const Scalar& sx, const Scalar& sy, const Scalar& sz)
+{ return DiagonalMatrix<Scalar,3>(sx, sy, sz); }
+
+/** Constructs an axis aligned scaling expression from vector expression \a coeffs
+  * This is an alias for coeffs.asDiagonal()
+  */
+template<typename Derived>
+inline const DiagonalWrapper<const Derived> Scaling(const MatrixBase<Derived>& coeffs)
+{ return coeffs.asDiagonal(); }
+
+/** \deprecated */
+typedef DiagonalMatrix<float, 2> AlignedScaling2f;
+/** \deprecated */
+typedef DiagonalMatrix<double,2> AlignedScaling2d;
+/** \deprecated */
+typedef DiagonalMatrix<float, 3> AlignedScaling3f;
+/** \deprecated */
+typedef DiagonalMatrix<double,3> AlignedScaling3d;
+//@}
+
+template<typename Scalar>
+template<int Dim>
+inline Transform<Scalar,Dim,Affine>
+UniformScaling<Scalar>::operator* (const Translation<Scalar,Dim>& t) const
+{
+  Transform<Scalar,Dim,Affine> res;
+  res.matrix().setZero();
+  res.linear().diagonal().fill(factor());
+  res.translation() = factor() * t.vector();
+  res(Dim,Dim) = Scalar(1);
+  return res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SCALING_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Transform.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Transform.h
new file mode 100644
index 0000000..3f31ee4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Transform.h
@@ -0,0 +1,1542 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2010 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRANSFORM_H
+#define EIGEN_TRANSFORM_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Transform>
+struct transform_traits
+{
+  enum
+  {
+    Dim = Transform::Dim,
+    HDim = Transform::HDim,
+    Mode = Transform::Mode,
+    IsProjective = (int(Mode)==int(Projective))
+  };
+};
+
+template< typename TransformType,
+          typename MatrixType,
+          int Case = transform_traits<TransformType>::IsProjective ? 0
+                   : int(MatrixType::RowsAtCompileTime) == int(transform_traits<TransformType>::HDim) ? 1
+                   : 2,
+          int RhsCols = MatrixType::ColsAtCompileTime>
+struct transform_right_product_impl;
+
+template< typename Other,
+          int Mode,
+          int Options,
+          int Dim,
+          int HDim,
+          int OtherRows=Other::RowsAtCompileTime,
+          int OtherCols=Other::ColsAtCompileTime>
+struct transform_left_product_impl;
+
+template< typename Lhs,
+          typename Rhs,
+          bool AnyProjective = 
+            transform_traits<Lhs>::IsProjective ||
+            transform_traits<Rhs>::IsProjective>
+struct transform_transform_product_impl;
+
+template< typename Other,
+          int Mode,
+          int Options,
+          int Dim,
+          int HDim,
+          int OtherRows=Other::RowsAtCompileTime,
+          int OtherCols=Other::ColsAtCompileTime>
+struct transform_construct_from_matrix;
+
+template<typename TransformType> struct transform_take_affine_part;
+
+template<typename _Scalar, int _Dim, int _Mode, int _Options>
+struct traits<Transform<_Scalar,_Dim,_Mode,_Options> >
+{
+  typedef _Scalar Scalar;
+  typedef Eigen::Index StorageIndex;
+  typedef Dense StorageKind;
+  enum {
+    Dim1 = _Dim==Dynamic ? _Dim : _Dim + 1,
+    RowsAtCompileTime = _Mode==Projective ? Dim1 : _Dim,
+    ColsAtCompileTime = Dim1,
+    MaxRowsAtCompileTime = RowsAtCompileTime,
+    MaxColsAtCompileTime = ColsAtCompileTime,
+    Flags = 0
+  };
+};
+
+template<int Mode> struct transform_make_affine;
+
+} // end namespace internal
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Transform
+  *
+  * \brief Represents an homogeneous transformation in a N dimensional space
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients
+  * \tparam _Dim the dimension of the space
+  * \tparam _Mode the type of the transformation. Can be:
+  *              - #Affine: the transformation is stored as a (Dim+1)^2 matrix,
+  *                         where the last row is assumed to be [0 ... 0 1].
+  *              - #AffineCompact: the transformation is stored as a (Dim)x(Dim+1) matrix.
+  *              - #Projective: the transformation is stored as a (Dim+1)^2 matrix
+  *                             without any assumption.
+  * \tparam _Options has the same meaning as in class Matrix. It allows to specify DontAlign and/or RowMajor.
+  *                  These Options are passed directly to the underlying matrix type.
+  *
+  * The homography is internally represented and stored by a matrix which
+  * is available through the matrix() method. To understand the behavior of
+  * this class you have to think a Transform object as its internal
+  * matrix representation. The chosen convention is right multiply:
+  *
+  * \code v' = T * v \endcode
+  *
+  * Therefore, an affine transformation matrix M is shaped like this:
+  *
+  * \f$ \left( \begin{array}{cc}
+  * linear & translation\\
+  * 0 ... 0 & 1
+  * \end{array} \right) \f$
+  *
+  * Note that for a projective transformation the last row can be anything,
+  * and then the interpretation of different parts might be sightly different.
+  *
+  * However, unlike a plain matrix, the Transform class provides many features
+  * simplifying both its assembly and usage. In particular, it can be composed
+  * with any other transformations (Transform,Translation,RotationBase,DiagonalMatrix)
+  * and can be directly used to transform implicit homogeneous vectors. All these
+  * operations are handled via the operator*. For the composition of transformations,
+  * its principle consists to first convert the right/left hand sides of the product
+  * to a compatible (Dim+1)^2 matrix and then perform a pure matrix product.
+  * Of course, internally, operator* tries to perform the minimal number of operations
+  * according to the nature of each terms. Likewise, when applying the transform
+  * to points, the latters are automatically promoted to homogeneous vectors
+  * before doing the matrix product. The conventions to homogeneous representations
+  * are performed as follow:
+  *
+  * \b Translation t (Dim)x(1):
+  * \f$ \left( \begin{array}{cc}
+  * I & t \\
+  * 0\,...\,0 & 1
+  * \end{array} \right) \f$
+  *
+  * \b Rotation R (Dim)x(Dim):
+  * \f$ \left( \begin{array}{cc}
+  * R & 0\\
+  * 0\,...\,0 & 1
+  * \end{array} \right) \f$
+  *<!--
+  * \b Linear \b Matrix L (Dim)x(Dim):
+  * \f$ \left( \begin{array}{cc}
+  * L & 0\\
+  * 0\,...\,0 & 1
+  * \end{array} \right) \f$
+  *
+  * \b Affine \b Matrix A (Dim)x(Dim+1):
+  * \f$ \left( \begin{array}{c}
+  * A\\
+  * 0\,...\,0\,1
+  * \end{array} \right) \f$
+  *-->
+  * \b Scaling \b DiagonalMatrix S (Dim)x(Dim):
+  * \f$ \left( \begin{array}{cc}
+  * S & 0\\
+  * 0\,...\,0 & 1
+  * \end{array} \right) \f$
+  *
+  * \b Column \b point v (Dim)x(1):
+  * \f$ \left( \begin{array}{c}
+  * v\\
+  * 1
+  * \end{array} \right) \f$
+  *
+  * \b Set \b of \b column \b points V1...Vn (Dim)x(n):
+  * \f$ \left( \begin{array}{ccc}
+  * v_1 & ... & v_n\\
+  * 1 & ... & 1
+  * \end{array} \right) \f$
+  *
+  * The concatenation of a Transform object with any kind of other transformation
+  * always returns a Transform object.
+  *
+  * A little exception to the "as pure matrix product" rule is the case of the
+  * transformation of non homogeneous vectors by an affine transformation. In
+  * that case the last matrix row can be ignored, and the product returns non
+  * homogeneous vectors.
+  *
+  * Since, for instance, a Dim x Dim matrix is interpreted as a linear transformation,
+  * it is not possible to directly transform Dim vectors stored in a Dim x Dim matrix.
+  * The solution is either to use a Dim x Dynamic matrix or explicitly request a
+  * vector transformation by making the vector homogeneous:
+  * \code
+  * m' = T * m.colwise().homogeneous();
+  * \endcode
+  * Note that there is zero overhead.
+  *
+  * Conversion methods from/to Qt's QMatrix and QTransform are available if the
+  * preprocessor token EIGEN_QT_SUPPORT is defined.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_TRANSFORM_PLUGIN.
+  *
+  * \sa class Matrix, class Quaternion
+  */
+template<typename _Scalar, int _Dim, int _Mode, int _Options>
+class Transform
+{
+public:
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Dim==Dynamic ? Dynamic : (_Dim+1)*(_Dim+1))
+  enum {
+    Mode = _Mode,
+    Options = _Options,
+    Dim = _Dim,     ///< space dimension in which the transformation holds
+    HDim = _Dim+1,  ///< size of a respective homogeneous vector
+    Rows = int(Mode)==(AffineCompact) ? Dim : HDim
+  };
+  /** the scalar type of the coefficients */
+  typedef _Scalar Scalar;
+  typedef Eigen::Index StorageIndex;
+  typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+  /** type of the matrix used to represent the transformation */
+  typedef typename internal::make_proper_matrix_type<Scalar,Rows,HDim,Options>::type MatrixType;
+  /** constified MatrixType */
+  typedef const MatrixType ConstMatrixType;
+  /** type of the matrix used to represent the linear part of the transformation */
+  typedef Matrix<Scalar,Dim,Dim,Options> LinearMatrixType;
+  /** type of read/write reference to the linear part of the transformation */
+  typedef Block<MatrixType,Dim,Dim,int(Mode)==(AffineCompact) && (Options&RowMajor)==0> LinearPart;
+  /** type of read reference to the linear part of the transformation */
+  typedef const Block<ConstMatrixType,Dim,Dim,int(Mode)==(AffineCompact) && (Options&RowMajor)==0> ConstLinearPart;
+  /** type of read/write reference to the affine part of the transformation */
+  typedef typename internal::conditional<int(Mode)==int(AffineCompact),
+                              MatrixType&,
+                              Block<MatrixType,Dim,HDim> >::type AffinePart;
+  /** type of read reference to the affine part of the transformation */
+  typedef typename internal::conditional<int(Mode)==int(AffineCompact),
+                              const MatrixType&,
+                              const Block<const MatrixType,Dim,HDim> >::type ConstAffinePart;
+  /** type of a vector */
+  typedef Matrix<Scalar,Dim,1> VectorType;
+  /** type of a read/write reference to the translation part of the rotation */
+  typedef Block<MatrixType,Dim,1,!(internal::traits<MatrixType>::Flags & RowMajorBit)> TranslationPart;
+  /** type of a read reference to the translation part of the rotation */
+  typedef const Block<ConstMatrixType,Dim,1,!(internal::traits<MatrixType>::Flags & RowMajorBit)> ConstTranslationPart;
+  /** corresponding translation type */
+  typedef Translation<Scalar,Dim> TranslationType;
+  
+  // this intermediate enum is needed to avoid an ICE with gcc 3.4 and 4.0
+  enum { TransformTimeDiagonalMode = ((Mode==int(Isometry))?Affine:int(Mode)) };
+  /** The return type of the product between a diagonal matrix and a transform */
+  typedef Transform<Scalar,Dim,TransformTimeDiagonalMode> TransformTimeDiagonalReturnType;
+
+protected:
+
+  MatrixType m_matrix;
+
+public:
+
+  /** Default constructor without initialization of the meaningful coefficients.
+    * If Mode==Affine, then the last row is set to [0 ... 0 1] */
+  EIGEN_DEVICE_FUNC inline Transform()
+  {
+    check_template_params();
+    internal::transform_make_affine<(int(Mode)==Affine) ? Affine : AffineCompact>::run(m_matrix);
+  }
+
+  EIGEN_DEVICE_FUNC inline Transform(const Transform& other)
+  {
+    check_template_params();
+    m_matrix = other.m_matrix;
+  }
+
+  EIGEN_DEVICE_FUNC inline explicit Transform(const TranslationType& t)
+  {
+    check_template_params();
+    *this = t;
+  }
+  EIGEN_DEVICE_FUNC inline explicit Transform(const UniformScaling<Scalar>& s)
+  {
+    check_template_params();
+    *this = s;
+  }
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline explicit Transform(const RotationBase<Derived, Dim>& r)
+  {
+    check_template_params();
+    *this = r;
+  }
+
+  EIGEN_DEVICE_FUNC inline Transform& operator=(const Transform& other)
+  { m_matrix = other.m_matrix; return *this; }
+
+  typedef internal::transform_take_affine_part<Transform> take_affine_part;
+
+  /** Constructs and initializes a transformation from a Dim^2 or a (Dim+1)^2 matrix. */
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC inline explicit Transform(const EigenBase<OtherDerived>& other)
+  {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar,typename OtherDerived::Scalar>::value),
+      YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY);
+
+    check_template_params();
+    internal::transform_construct_from_matrix<OtherDerived,Mode,Options,Dim,HDim>::run(this, other.derived());
+  }
+
+  /** Set \c *this from a Dim^2 or (Dim+1)^2 matrix. */
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC inline Transform& operator=(const EigenBase<OtherDerived>& other)
+  {
+    EIGEN_STATIC_ASSERT((internal::is_same<Scalar,typename OtherDerived::Scalar>::value),
+      YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY);
+
+    internal::transform_construct_from_matrix<OtherDerived,Mode,Options,Dim,HDim>::run(this, other.derived());
+    return *this;
+  }
+  
+  template<int OtherOptions>
+  EIGEN_DEVICE_FUNC inline Transform(const Transform<Scalar,Dim,Mode,OtherOptions>& other)
+  {
+    check_template_params();
+    // only the options change, we can directly copy the matrices
+    m_matrix = other.matrix();
+  }
+
+  template<int OtherMode,int OtherOptions>
+  EIGEN_DEVICE_FUNC inline Transform(const Transform<Scalar,Dim,OtherMode,OtherOptions>& other)
+  {
+    check_template_params();
+    // prevent conversions as:
+    // Affine | AffineCompact | Isometry = Projective
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(OtherMode==int(Projective), Mode==int(Projective)),
+                        YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION)
+
+    // prevent conversions as:
+    // Isometry = Affine | AffineCompact
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(OtherMode==int(Affine)||OtherMode==int(AffineCompact), Mode!=int(Isometry)),
+                        YOU_PERFORMED_AN_INVALID_TRANSFORMATION_CONVERSION)
+
+    enum { ModeIsAffineCompact = Mode == int(AffineCompact),
+           OtherModeIsAffineCompact = OtherMode == int(AffineCompact)
+    };
+
+    if(ModeIsAffineCompact == OtherModeIsAffineCompact)
+    {
+      // We need the block expression because the code is compiled for all
+      // combinations of transformations and will trigger a compile time error
+      // if one tries to assign the matrices directly
+      m_matrix.template block<Dim,Dim+1>(0,0) = other.matrix().template block<Dim,Dim+1>(0,0);
+      makeAffine();
+    }
+    else if(OtherModeIsAffineCompact)
+    {
+      typedef typename Transform<Scalar,Dim,OtherMode,OtherOptions>::MatrixType OtherMatrixType;
+      internal::transform_construct_from_matrix<OtherMatrixType,Mode,Options,Dim,HDim>::run(this, other.matrix());
+    }
+    else
+    {
+      // here we know that Mode == AffineCompact and OtherMode != AffineCompact.
+      // if OtherMode were Projective, the static assert above would already have caught it.
+      // So the only possibility is that OtherMode == Affine
+      linear() = other.linear();
+      translation() = other.translation();
+    }
+  }
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC Transform(const ReturnByValue<OtherDerived>& other)
+  {
+    check_template_params();
+    other.evalTo(*this);
+  }
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC Transform& operator=(const ReturnByValue<OtherDerived>& other)
+  {
+    other.evalTo(*this);
+    return *this;
+  }
+
+  #ifdef EIGEN_QT_SUPPORT
+  inline Transform(const QMatrix& other);
+  inline Transform& operator=(const QMatrix& other);
+  inline QMatrix toQMatrix(void) const;
+  inline Transform(const QTransform& other);
+  inline Transform& operator=(const QTransform& other);
+  inline QTransform toQTransform(void) const;
+  #endif
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return int(Mode)==int(Projective) ? m_matrix.cols() : (m_matrix.cols()-1); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_matrix.cols(); }
+
+  /** shortcut for m_matrix(row,col);
+    * \sa MatrixBase::operator(Index,Index) const */
+  EIGEN_DEVICE_FUNC inline Scalar operator() (Index row, Index col) const { return m_matrix(row,col); }
+  /** shortcut for m_matrix(row,col);
+    * \sa MatrixBase::operator(Index,Index) */
+  EIGEN_DEVICE_FUNC inline Scalar& operator() (Index row, Index col) { return m_matrix(row,col); }
+
+  /** \returns a read-only expression of the transformation matrix */
+  EIGEN_DEVICE_FUNC inline const MatrixType& matrix() const { return m_matrix; }
+  /** \returns a writable expression of the transformation matrix */
+  EIGEN_DEVICE_FUNC inline MatrixType& matrix() { return m_matrix; }
+
+  /** \returns a read-only expression of the linear part of the transformation */
+  EIGEN_DEVICE_FUNC inline ConstLinearPart linear() const { return ConstLinearPart(m_matrix,0,0); }
+  /** \returns a writable expression of the linear part of the transformation */
+  EIGEN_DEVICE_FUNC inline LinearPart linear() { return LinearPart(m_matrix,0,0); }
+
+  /** \returns a read-only expression of the Dim x HDim affine part of the transformation */
+  EIGEN_DEVICE_FUNC inline ConstAffinePart affine() const { return take_affine_part::run(m_matrix); }
+  /** \returns a writable expression of the Dim x HDim affine part of the transformation */
+  EIGEN_DEVICE_FUNC inline AffinePart affine() { return take_affine_part::run(m_matrix); }
+
+  /** \returns a read-only expression of the translation vector of the transformation */
+  EIGEN_DEVICE_FUNC inline ConstTranslationPart translation() const { return ConstTranslationPart(m_matrix,0,Dim); }
+  /** \returns a writable expression of the translation vector of the transformation */
+  EIGEN_DEVICE_FUNC inline TranslationPart translation() { return TranslationPart(m_matrix,0,Dim); }
+
+  /** \returns an expression of the product between the transform \c *this and a matrix expression \a other.
+    *
+    * The right-hand-side \a other can be either:
+    * \li an homogeneous vector of size Dim+1,
+    * \li a set of homogeneous vectors of size Dim+1 x N,
+    * \li a transformation matrix of size Dim+1 x Dim+1.
+    *
+    * Moreover, if \c *this represents an affine transformation (i.e., Mode!=Projective), then \a other can also be:
+    * \li a point of size Dim (computes: \code this->linear() * other + this->translation()\endcode),
+    * \li a set of N points as a Dim x N matrix (computes: \code (this->linear() * other).colwise() + this->translation()\endcode),
+    *
+    * In all cases, the return type is a matrix or vector of same sizes as the right-hand-side \a other.
+    *
+    * If you want to interpret \a other as a linear or affine transformation, then first convert it to a Transform<> type,
+    * or do your own cooking.
+    *
+    * Finally, if you want to apply Affine transformations to vectors, then explicitly apply the linear part only:
+    * \code
+    * Affine3f A;
+    * Vector3f v1, v2;
+    * v2 = A.linear() * v1;
+    * \endcode
+    *
+    */
+  // note: this function is defined here because some compilers cannot find the respective declaration
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const typename internal::transform_right_product_impl<Transform, OtherDerived>::ResultType
+  operator * (const EigenBase<OtherDerived> &other) const
+  { return internal::transform_right_product_impl<Transform, OtherDerived>::run(*this,other.derived()); }
+
+  /** \returns the product expression of a transformation matrix \a a times a transform \a b
+    *
+    * The left hand side \a other can be either:
+    * \li a linear transformation matrix of size Dim x Dim,
+    * \li an affine transformation matrix of size Dim x Dim+1,
+    * \li a general transformation matrix of size Dim+1 x Dim+1.
+    */
+  template<typename OtherDerived> friend
+  EIGEN_DEVICE_FUNC inline const typename internal::transform_left_product_impl<OtherDerived,Mode,Options,_Dim,_Dim+1>::ResultType
+    operator * (const EigenBase<OtherDerived> &a, const Transform &b)
+  { return internal::transform_left_product_impl<OtherDerived,Mode,Options,Dim,HDim>::run(a.derived(),b); }
+
+  /** \returns The product expression of a transform \a a times a diagonal matrix \a b
+    *
+    * The rhs diagonal matrix is interpreted as an affine scaling transformation. The
+    * product results in a Transform of the same type (mode) as the lhs only if the lhs 
+    * mode is no isometry. In that case, the returned transform is an affinity.
+    */
+  template<typename DiagonalDerived>
+  EIGEN_DEVICE_FUNC inline const TransformTimeDiagonalReturnType
+    operator * (const DiagonalBase<DiagonalDerived> &b) const
+  {
+    TransformTimeDiagonalReturnType res(*this);
+    res.linearExt() *= b;
+    return res;
+  }
+
+  /** \returns The product expression of a diagonal matrix \a a times a transform \a b
+    *
+    * The lhs diagonal matrix is interpreted as an affine scaling transformation. The
+    * product results in a Transform of the same type (mode) as the lhs only if the lhs 
+    * mode is no isometry. In that case, the returned transform is an affinity.
+    */
+  template<typename DiagonalDerived>
+  EIGEN_DEVICE_FUNC friend inline TransformTimeDiagonalReturnType
+    operator * (const DiagonalBase<DiagonalDerived> &a, const Transform &b)
+  {
+    TransformTimeDiagonalReturnType res;
+    res.linear().noalias() = a*b.linear();
+    res.translation().noalias() = a*b.translation();
+    if (Mode!=int(AffineCompact))
+      res.matrix().row(Dim) = b.matrix().row(Dim);
+    return res;
+  }
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC inline Transform& operator*=(const EigenBase<OtherDerived>& other) { return *this = *this * other; }
+
+  /** Concatenates two transformations */
+  EIGEN_DEVICE_FUNC inline const Transform operator * (const Transform& other) const
+  {
+    return internal::transform_transform_product_impl<Transform,Transform>::run(*this,other);
+  }
+  
+  #if EIGEN_COMP_ICC
+private:
+  // this intermediate structure permits to workaround a bug in ICC 11:
+  //   error: template instantiation resulted in unexpected function type of "Eigen::Transform<double, 3, 32, 0>
+  //             (const Eigen::Transform<double, 3, 2, 0> &) const"
+  //  (the meaning of a name may have changed since the template declaration -- the type of the template is:
+  // "Eigen::internal::transform_transform_product_impl<Eigen::Transform<double, 3, 32, 0>,
+  //     Eigen::Transform<double, 3, Mode, Options>, <expression>>::ResultType (const Eigen::Transform<double, 3, Mode, Options> &) const")
+  // 
+  template<int OtherMode,int OtherOptions> struct icc_11_workaround
+  {
+    typedef internal::transform_transform_product_impl<Transform,Transform<Scalar,Dim,OtherMode,OtherOptions> > ProductType;
+    typedef typename ProductType::ResultType ResultType;
+  };
+  
+public:
+  /** Concatenates two different transformations */
+  template<int OtherMode,int OtherOptions>
+  inline typename icc_11_workaround<OtherMode,OtherOptions>::ResultType
+    operator * (const Transform<Scalar,Dim,OtherMode,OtherOptions>& other) const
+  {
+    typedef typename icc_11_workaround<OtherMode,OtherOptions>::ProductType ProductType;
+    return ProductType::run(*this,other);
+  }
+  #else
+  /** Concatenates two different transformations */
+  template<int OtherMode,int OtherOptions>
+  EIGEN_DEVICE_FUNC inline typename internal::transform_transform_product_impl<Transform,Transform<Scalar,Dim,OtherMode,OtherOptions> >::ResultType
+    operator * (const Transform<Scalar,Dim,OtherMode,OtherOptions>& other) const
+  {
+    return internal::transform_transform_product_impl<Transform,Transform<Scalar,Dim,OtherMode,OtherOptions> >::run(*this,other);
+  }
+  #endif
+
+  /** \sa MatrixBase::setIdentity() */
+  EIGEN_DEVICE_FUNC void setIdentity() { m_matrix.setIdentity(); }
+
+  /**
+   * \brief Returns an identity transformation.
+   * \todo In the future this function should be returning a Transform expression.
+   */
+  EIGEN_DEVICE_FUNC static const Transform Identity()
+  {
+    return Transform(MatrixType::Identity());
+  }
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC 
+  inline Transform& scale(const MatrixBase<OtherDerived> &other);
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
+  inline Transform& prescale(const MatrixBase<OtherDerived> &other);
+
+  EIGEN_DEVICE_FUNC inline Transform& scale(const Scalar& s);
+  EIGEN_DEVICE_FUNC inline Transform& prescale(const Scalar& s);
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
+  inline Transform& translate(const MatrixBase<OtherDerived> &other);
+
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC
+  inline Transform& pretranslate(const MatrixBase<OtherDerived> &other);
+
+  template<typename RotationType>
+  EIGEN_DEVICE_FUNC
+  inline Transform& rotate(const RotationType& rotation);
+
+  template<typename RotationType>
+  EIGEN_DEVICE_FUNC
+  inline Transform& prerotate(const RotationType& rotation);
+
+  EIGEN_DEVICE_FUNC Transform& shear(const Scalar& sx, const Scalar& sy);
+  EIGEN_DEVICE_FUNC Transform& preshear(const Scalar& sx, const Scalar& sy);
+
+  EIGEN_DEVICE_FUNC inline Transform& operator=(const TranslationType& t);
+  
+  EIGEN_DEVICE_FUNC
+  inline Transform& operator*=(const TranslationType& t) { return translate(t.vector()); }
+  
+  EIGEN_DEVICE_FUNC inline Transform operator*(const TranslationType& t) const;
+
+  EIGEN_DEVICE_FUNC 
+  inline Transform& operator=(const UniformScaling<Scalar>& t);
+  
+  EIGEN_DEVICE_FUNC
+  inline Transform& operator*=(const UniformScaling<Scalar>& s) { return scale(s.factor()); }
+  
+  EIGEN_DEVICE_FUNC
+  inline TransformTimeDiagonalReturnType operator*(const UniformScaling<Scalar>& s) const
+  {
+    TransformTimeDiagonalReturnType res = *this;
+    res.scale(s.factor());
+    return res;
+  }
+
+  EIGEN_DEVICE_FUNC
+  inline Transform& operator*=(const DiagonalMatrix<Scalar,Dim>& s) { linearExt() *= s; return *this; }
+
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline Transform& operator=(const RotationBase<Derived,Dim>& r);
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline Transform& operator*=(const RotationBase<Derived,Dim>& r) { return rotate(r.toRotationMatrix()); }
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline Transform operator*(const RotationBase<Derived,Dim>& r) const;
+
+  EIGEN_DEVICE_FUNC const LinearMatrixType rotation() const;
+  template<typename RotationMatrixType, typename ScalingMatrixType>
+  EIGEN_DEVICE_FUNC
+  void computeRotationScaling(RotationMatrixType *rotation, ScalingMatrixType *scaling) const;
+  template<typename ScalingMatrixType, typename RotationMatrixType>
+  EIGEN_DEVICE_FUNC
+  void computeScalingRotation(ScalingMatrixType *scaling, RotationMatrixType *rotation) const;
+
+  template<typename PositionDerived, typename OrientationType, typename ScaleDerived>
+  EIGEN_DEVICE_FUNC
+  Transform& fromPositionOrientationScale(const MatrixBase<PositionDerived> &position,
+    const OrientationType& orientation, const MatrixBase<ScaleDerived> &scale);
+
+  EIGEN_DEVICE_FUNC
+  inline Transform inverse(TransformTraits traits = (TransformTraits)Mode) const;
+
+  /** \returns a const pointer to the column major internal matrix */
+  EIGEN_DEVICE_FUNC const Scalar* data() const { return m_matrix.data(); }
+  /** \returns a non-const pointer to the column major internal matrix */
+  EIGEN_DEVICE_FUNC Scalar* data() { return m_matrix.data(); }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<Transform,Transform<NewScalarType,Dim,Mode,Options> >::type cast() const
+  { return typename internal::cast_return_type<Transform,Transform<NewScalarType,Dim,Mode,Options> >::type(*this); }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  EIGEN_DEVICE_FUNC inline explicit Transform(const Transform<OtherScalarType,Dim,Mode,Options>& other)
+  {
+    check_template_params();
+    m_matrix = other.matrix().template cast<Scalar>();
+  }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const Transform& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return m_matrix.isApprox(other.m_matrix, prec); }
+
+  /** Sets the last row to [0 ... 0 1]
+    */
+  EIGEN_DEVICE_FUNC void makeAffine()
+  {
+    internal::transform_make_affine<int(Mode)>::run(m_matrix);
+  }
+
+  /** \internal
+    * \returns the Dim x Dim linear part if the transformation is affine,
+    *          and the HDim x Dim part for projective transformations.
+    */
+  EIGEN_DEVICE_FUNC inline Block<MatrixType,int(Mode)==int(Projective)?HDim:Dim,Dim> linearExt()
+  { return m_matrix.template block<int(Mode)==int(Projective)?HDim:Dim,Dim>(0,0); }
+  /** \internal
+    * \returns the Dim x Dim linear part if the transformation is affine,
+    *          and the HDim x Dim part for projective transformations.
+    */
+  EIGEN_DEVICE_FUNC inline const Block<MatrixType,int(Mode)==int(Projective)?HDim:Dim,Dim> linearExt() const
+  { return m_matrix.template block<int(Mode)==int(Projective)?HDim:Dim,Dim>(0,0); }
+
+  /** \internal
+    * \returns the translation part if the transformation is affine,
+    *          and the last column for projective transformations.
+    */
+  EIGEN_DEVICE_FUNC inline Block<MatrixType,int(Mode)==int(Projective)?HDim:Dim,1> translationExt()
+  { return m_matrix.template block<int(Mode)==int(Projective)?HDim:Dim,1>(0,Dim); }
+  /** \internal
+    * \returns the translation part if the transformation is affine,
+    *          and the last column for projective transformations.
+    */
+  EIGEN_DEVICE_FUNC inline const Block<MatrixType,int(Mode)==int(Projective)?HDim:Dim,1> translationExt() const
+  { return m_matrix.template block<int(Mode)==int(Projective)?HDim:Dim,1>(0,Dim); }
+
+
+  #ifdef EIGEN_TRANSFORM_PLUGIN
+  #include EIGEN_TRANSFORM_PLUGIN
+  #endif
+  
+protected:
+  #ifndef EIGEN_PARSED_BY_DOXYGEN
+    EIGEN_DEVICE_FUNC static EIGEN_STRONG_INLINE void check_template_params()
+    {
+      EIGEN_STATIC_ASSERT((Options & (DontAlign|RowMajor)) == Options, INVALID_MATRIX_TEMPLATE_PARAMETERS)
+    }
+  #endif
+
+};
+
+/** \ingroup Geometry_Module */
+typedef Transform<float,2,Isometry> Isometry2f;
+/** \ingroup Geometry_Module */
+typedef Transform<float,3,Isometry> Isometry3f;
+/** \ingroup Geometry_Module */
+typedef Transform<double,2,Isometry> Isometry2d;
+/** \ingroup Geometry_Module */
+typedef Transform<double,3,Isometry> Isometry3d;
+
+/** \ingroup Geometry_Module */
+typedef Transform<float,2,Affine> Affine2f;
+/** \ingroup Geometry_Module */
+typedef Transform<float,3,Affine> Affine3f;
+/** \ingroup Geometry_Module */
+typedef Transform<double,2,Affine> Affine2d;
+/** \ingroup Geometry_Module */
+typedef Transform<double,3,Affine> Affine3d;
+
+/** \ingroup Geometry_Module */
+typedef Transform<float,2,AffineCompact> AffineCompact2f;
+/** \ingroup Geometry_Module */
+typedef Transform<float,3,AffineCompact> AffineCompact3f;
+/** \ingroup Geometry_Module */
+typedef Transform<double,2,AffineCompact> AffineCompact2d;
+/** \ingroup Geometry_Module */
+typedef Transform<double,3,AffineCompact> AffineCompact3d;
+
+/** \ingroup Geometry_Module */
+typedef Transform<float,2,Projective> Projective2f;
+/** \ingroup Geometry_Module */
+typedef Transform<float,3,Projective> Projective3f;
+/** \ingroup Geometry_Module */
+typedef Transform<double,2,Projective> Projective2d;
+/** \ingroup Geometry_Module */
+typedef Transform<double,3,Projective> Projective3d;
+
+/**************************
+*** Optional QT support ***
+**************************/
+
+#ifdef EIGEN_QT_SUPPORT
+/** Initializes \c *this from a QMatrix assuming the dimension is 2.
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode,int Options>
+Transform<Scalar,Dim,Mode,Options>::Transform(const QMatrix& other)
+{
+  check_template_params();
+  *this = other;
+}
+
+/** Set \c *this from a QMatrix assuming the dimension is 2.
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode,int Options>
+Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::operator=(const QMatrix& other)
+{
+  EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  if (Mode == int(AffineCompact))
+    m_matrix << other.m11(), other.m21(), other.dx(),
+                other.m12(), other.m22(), other.dy();
+  else
+    m_matrix << other.m11(), other.m21(), other.dx(),
+                other.m12(), other.m22(), other.dy(),
+                0, 0, 1;
+  return *this;
+}
+
+/** \returns a QMatrix from \c *this assuming the dimension is 2.
+  *
+  * \warning this conversion might loss data if \c *this is not affine
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+QMatrix Transform<Scalar,Dim,Mode,Options>::toQMatrix(void) const
+{
+  check_template_params();
+  EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  return QMatrix(m_matrix.coeff(0,0), m_matrix.coeff(1,0),
+                 m_matrix.coeff(0,1), m_matrix.coeff(1,1),
+                 m_matrix.coeff(0,2), m_matrix.coeff(1,2));
+}
+
+/** Initializes \c *this from a QTransform assuming the dimension is 2.
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode,int Options>
+Transform<Scalar,Dim,Mode,Options>::Transform(const QTransform& other)
+{
+  check_template_params();
+  *this = other;
+}
+
+/** Set \c *this from a QTransform assuming the dimension is 2.
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::operator=(const QTransform& other)
+{
+  check_template_params();
+  EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  if (Mode == int(AffineCompact))
+    m_matrix << other.m11(), other.m21(), other.dx(),
+                other.m12(), other.m22(), other.dy();
+  else
+    m_matrix << other.m11(), other.m21(), other.dx(),
+                other.m12(), other.m22(), other.dy(),
+                other.m13(), other.m23(), other.m33();
+  return *this;
+}
+
+/** \returns a QTransform from \c *this assuming the dimension is 2.
+  *
+  * This function is available only if the token EIGEN_QT_SUPPORT is defined.
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+QTransform Transform<Scalar,Dim,Mode,Options>::toQTransform(void) const
+{
+  EIGEN_STATIC_ASSERT(Dim==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  if (Mode == int(AffineCompact))
+    return QTransform(m_matrix.coeff(0,0), m_matrix.coeff(1,0),
+                      m_matrix.coeff(0,1), m_matrix.coeff(1,1),
+                      m_matrix.coeff(0,2), m_matrix.coeff(1,2));
+  else
+    return QTransform(m_matrix.coeff(0,0), m_matrix.coeff(1,0), m_matrix.coeff(2,0),
+                      m_matrix.coeff(0,1), m_matrix.coeff(1,1), m_matrix.coeff(2,1),
+                      m_matrix.coeff(0,2), m_matrix.coeff(1,2), m_matrix.coeff(2,2));
+}
+#endif
+
+/*********************
+*** Procedural API ***
+*********************/
+
+/** Applies on the right the non uniform scale transformation represented
+  * by the vector \a other to \c *this and returns a reference to \c *this.
+  * \sa prescale()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::scale(const MatrixBase<OtherDerived> &other)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim))
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  linearExt().noalias() = (linearExt() * other.asDiagonal());
+  return *this;
+}
+
+/** Applies on the right a uniform scale of a factor \a c to \c *this
+  * and returns a reference to \c *this.
+  * \sa prescale(Scalar)
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::scale(const Scalar& s)
+{
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  linearExt() *= s;
+  return *this;
+}
+
+/** Applies on the left the non uniform scale transformation represented
+  * by the vector \a other to \c *this and returns a reference to \c *this.
+  * \sa scale()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::prescale(const MatrixBase<OtherDerived> &other)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim))
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  affine().noalias() = (other.asDiagonal() * affine());
+  return *this;
+}
+
+/** Applies on the left a uniform scale of a factor \a c to \c *this
+  * and returns a reference to \c *this.
+  * \sa scale(Scalar)
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::prescale(const Scalar& s)
+{
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  m_matrix.template topRows<Dim>() *= s;
+  return *this;
+}
+
+/** Applies on the right the translation matrix represented by the vector \a other
+  * to \c *this and returns a reference to \c *this.
+  * \sa pretranslate()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::translate(const MatrixBase<OtherDerived> &other)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim))
+  translationExt() += linearExt() * other;
+  return *this;
+}
+
+/** Applies on the left the translation matrix represented by the vector \a other
+  * to \c *this and returns a reference to \c *this.
+  * \sa translate()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::pretranslate(const MatrixBase<OtherDerived> &other)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_SPECIFIC_SIZE(OtherDerived,int(Dim))
+  if(int(Mode)==int(Projective))
+    affine() += other * m_matrix.row(Dim);
+  else
+    translation() += other;
+  return *this;
+}
+
+/** Applies on the right the rotation represented by the rotation \a rotation
+  * to \c *this and returns a reference to \c *this.
+  *
+  * The template parameter \a RotationType is the type of the rotation which
+  * must be known by internal::toRotationMatrix<>.
+  *
+  * Natively supported types includes:
+  *   - any scalar (2D),
+  *   - a Dim x Dim matrix expression,
+  *   - a Quaternion (3D),
+  *   - a AngleAxis (3D)
+  *
+  * This mechanism is easily extendable to support user types such as Euler angles,
+  * or a pair of Quaternion for 4D rotations.
+  *
+  * \sa rotate(Scalar), class Quaternion, class AngleAxis, prerotate(RotationType)
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename RotationType>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::rotate(const RotationType& rotation)
+{
+  linearExt() *= internal::toRotationMatrix<Scalar,Dim>(rotation);
+  return *this;
+}
+
+/** Applies on the left the rotation represented by the rotation \a rotation
+  * to \c *this and returns a reference to \c *this.
+  *
+  * See rotate() for further details.
+  *
+  * \sa rotate()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename RotationType>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::prerotate(const RotationType& rotation)
+{
+  m_matrix.template block<Dim,HDim>(0,0) = internal::toRotationMatrix<Scalar,Dim>(rotation)
+                                         * m_matrix.template block<Dim,HDim>(0,0);
+  return *this;
+}
+
+/** Applies on the right the shear transformation represented
+  * by the vector \a other to \c *this and returns a reference to \c *this.
+  * \warning 2D only.
+  * \sa preshear()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::shear(const Scalar& sx, const Scalar& sy)
+{
+  EIGEN_STATIC_ASSERT(int(Dim)==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  VectorType tmp = linear().col(0)*sy + linear().col(1);
+  linear() << linear().col(0) + linear().col(1)*sx, tmp;
+  return *this;
+}
+
+/** Applies on the left the shear transformation represented
+  * by the vector \a other to \c *this and returns a reference to \c *this.
+  * \warning 2D only.
+  * \sa shear()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::preshear(const Scalar& sx, const Scalar& sy)
+{
+  EIGEN_STATIC_ASSERT(int(Dim)==2, YOU_MADE_A_PROGRAMMING_MISTAKE)
+  EIGEN_STATIC_ASSERT(Mode!=int(Isometry), THIS_METHOD_IS_ONLY_FOR_SPECIFIC_TRANSFORMATIONS)
+  m_matrix.template block<Dim,HDim>(0,0) = LinearMatrixType(1, sx, sy, 1) * m_matrix.template block<Dim,HDim>(0,0);
+  return *this;
+}
+
+/******************************************************
+*** Scaling, Translation and Rotation compatibility ***
+******************************************************/
+
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::operator=(const TranslationType& t)
+{
+  linear().setIdentity();
+  translation() = t.vector();
+  makeAffine();
+  return *this;
+}
+
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options> Transform<Scalar,Dim,Mode,Options>::operator*(const TranslationType& t) const
+{
+  Transform res = *this;
+  res.translate(t.vector());
+  return res;
+}
+
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::operator=(const UniformScaling<Scalar>& s)
+{
+  m_matrix.setZero();
+  linear().diagonal().fill(s.factor());
+  makeAffine();
+  return *this;
+}
+
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options>& Transform<Scalar,Dim,Mode,Options>::operator=(const RotationBase<Derived,Dim>& r)
+{
+  linear() = internal::toRotationMatrix<Scalar,Dim>(r);
+  translation().setZero();
+  makeAffine();
+  return *this;
+}
+
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename Derived>
+EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode,Options> Transform<Scalar,Dim,Mode,Options>::operator*(const RotationBase<Derived,Dim>& r) const
+{
+  Transform res = *this;
+  res.rotate(r.derived());
+  return res;
+}
+
+/************************
+*** Special functions ***
+************************/
+
+/** \returns the rotation part of the transformation
+  *
+  *
+  * \svd_module
+  *
+  * \sa computeRotationScaling(), computeScalingRotation(), class SVD
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC const typename Transform<Scalar,Dim,Mode,Options>::LinearMatrixType
+Transform<Scalar,Dim,Mode,Options>::rotation() const
+{
+  LinearMatrixType result;
+  computeRotationScaling(&result, (LinearMatrixType*)0);
+  return result;
+}
+
+
+/** decomposes the linear part of the transformation as a product rotation x scaling, the scaling being
+  * not necessarily positive.
+  *
+  * If either pointer is zero, the corresponding computation is skipped.
+  *
+  *
+  *
+  * \svd_module
+  *
+  * \sa computeScalingRotation(), rotation(), class SVD
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename RotationMatrixType, typename ScalingMatrixType>
+EIGEN_DEVICE_FUNC void Transform<Scalar,Dim,Mode,Options>::computeRotationScaling(RotationMatrixType *rotation, ScalingMatrixType *scaling) const
+{
+  JacobiSVD<LinearMatrixType> svd(linear(), ComputeFullU | ComputeFullV);
+
+  Scalar x = (svd.matrixU() * svd.matrixV().adjoint()).determinant(); // so x has absolute value 1
+  VectorType sv(svd.singularValues());
+  sv.coeffRef(0) *= x;
+  if(scaling) scaling->lazyAssign(svd.matrixV() * sv.asDiagonal() * svd.matrixV().adjoint());
+  if(rotation)
+  {
+    LinearMatrixType m(svd.matrixU());
+    m.col(0) /= x;
+    rotation->lazyAssign(m * svd.matrixV().adjoint());
+  }
+}
+
+/** decomposes the linear part of the transformation as a product scaling x rotation, the scaling being
+  * not necessarily positive.
+  *
+  * If either pointer is zero, the corresponding computation is skipped.
+  *
+  *
+  *
+  * \svd_module
+  *
+  * \sa computeRotationScaling(), rotation(), class SVD
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename ScalingMatrixType, typename RotationMatrixType>
+EIGEN_DEVICE_FUNC void Transform<Scalar,Dim,Mode,Options>::computeScalingRotation(ScalingMatrixType *scaling, RotationMatrixType *rotation) const
+{
+  JacobiSVD<LinearMatrixType> svd(linear(), ComputeFullU | ComputeFullV);
+
+  Scalar x = (svd.matrixU() * svd.matrixV().adjoint()).determinant(); // so x has absolute value 1
+  VectorType sv(svd.singularValues());
+  sv.coeffRef(0) *= x;
+  if(scaling) scaling->lazyAssign(svd.matrixU() * sv.asDiagonal() * svd.matrixU().adjoint());
+  if(rotation)
+  {
+    LinearMatrixType m(svd.matrixU());
+    m.col(0) /= x;
+    rotation->lazyAssign(m * svd.matrixV().adjoint());
+  }
+}
+
+/** Convenient method to set \c *this from a position, orientation and scale
+  * of a 3D object.
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+template<typename PositionDerived, typename OrientationType, typename ScaleDerived>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>&
+Transform<Scalar,Dim,Mode,Options>::fromPositionOrientationScale(const MatrixBase<PositionDerived> &position,
+  const OrientationType& orientation, const MatrixBase<ScaleDerived> &scale)
+{
+  linear() = internal::toRotationMatrix<Scalar,Dim>(orientation);
+  linear() *= scale.asDiagonal();
+  translation() = position;
+  makeAffine();
+  return *this;
+}
+
+namespace internal {
+
+template<int Mode>
+struct transform_make_affine
+{
+  template<typename MatrixType>
+  EIGEN_DEVICE_FUNC static void run(MatrixType &mat)
+  {
+    static const int Dim = MatrixType::ColsAtCompileTime-1;
+    mat.template block<1,Dim>(Dim,0).setZero();
+    mat.coeffRef(Dim,Dim) = typename MatrixType::Scalar(1);
+  }
+};
+
+template<>
+struct transform_make_affine<AffineCompact>
+{
+  template<typename MatrixType> EIGEN_DEVICE_FUNC static void run(MatrixType &) { }
+};
+    
+// selector needed to avoid taking the inverse of a 3x4 matrix
+template<typename TransformType, int Mode=TransformType::Mode>
+struct projective_transform_inverse
+{
+  EIGEN_DEVICE_FUNC static inline void run(const TransformType&, TransformType&)
+  {}
+};
+
+template<typename TransformType>
+struct projective_transform_inverse<TransformType, Projective>
+{
+  EIGEN_DEVICE_FUNC static inline void run(const TransformType& m, TransformType& res)
+  {
+    res.matrix() = m.matrix().inverse();
+  }
+};
+
+} // end namespace internal
+
+
+/**
+  *
+  * \returns the inverse transformation according to some given knowledge
+  * on \c *this.
+  *
+  * \param hint allows to optimize the inversion process when the transformation
+  * is known to be not a general transformation (optional). The possible values are:
+  *  - #Projective if the transformation is not necessarily affine, i.e., if the
+  *    last row is not guaranteed to be [0 ... 0 1]
+  *  - #Affine if the last row can be assumed to be [0 ... 0 1]
+  *  - #Isometry if the transformation is only a concatenations of translations
+  *    and rotations.
+  *  The default is the template class parameter \c Mode.
+  *
+  * \warning unless \a traits is always set to NoShear or NoScaling, this function
+  * requires the generic inverse method of MatrixBase defined in the LU module. If
+  * you forget to include this module, then you will get hard to debug linking errors.
+  *
+  * \sa MatrixBase::inverse()
+  */
+template<typename Scalar, int Dim, int Mode, int Options>
+EIGEN_DEVICE_FUNC Transform<Scalar,Dim,Mode,Options>
+Transform<Scalar,Dim,Mode,Options>::inverse(TransformTraits hint) const
+{
+  Transform res;
+  if (hint == Projective)
+  {
+    internal::projective_transform_inverse<Transform>::run(*this, res);
+  }
+  else
+  {
+    if (hint == Isometry)
+    {
+      res.matrix().template topLeftCorner<Dim,Dim>() = linear().transpose();
+    }
+    else if(hint&Affine)
+    {
+      res.matrix().template topLeftCorner<Dim,Dim>() = linear().inverse();
+    }
+    else
+    {
+      eigen_assert(false && "Invalid transform traits in Transform::Inverse");
+    }
+    // translation and remaining parts
+    res.matrix().template topRightCorner<Dim,1>()
+      = - res.matrix().template topLeftCorner<Dim,Dim>() * translation();
+    res.makeAffine(); // we do need this, because in the beginning res is uninitialized
+  }
+  return res;
+}
+
+namespace internal {
+
+/*****************************************************
+*** Specializations of take affine part            ***
+*****************************************************/
+
+template<typename TransformType> struct transform_take_affine_part {
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef typename TransformType::AffinePart AffinePart;
+  typedef typename TransformType::ConstAffinePart ConstAffinePart;
+  static inline AffinePart run(MatrixType& m)
+  { return m.template block<TransformType::Dim,TransformType::HDim>(0,0); }
+  static inline ConstAffinePart run(const MatrixType& m)
+  { return m.template block<TransformType::Dim,TransformType::HDim>(0,0); }
+};
+
+template<typename Scalar, int Dim, int Options>
+struct transform_take_affine_part<Transform<Scalar,Dim,AffineCompact, Options> > {
+  typedef typename Transform<Scalar,Dim,AffineCompact,Options>::MatrixType MatrixType;
+  static inline MatrixType& run(MatrixType& m) { return m; }
+  static inline const MatrixType& run(const MatrixType& m) { return m; }
+};
+
+/*****************************************************
+*** Specializations of construct from matrix       ***
+*****************************************************/
+
+template<typename Other, int Mode, int Options, int Dim, int HDim>
+struct transform_construct_from_matrix<Other, Mode,Options,Dim,HDim, Dim,Dim>
+{
+  static inline void run(Transform<typename Other::Scalar,Dim,Mode,Options> *transform, const Other& other)
+  {
+    transform->linear() = other;
+    transform->translation().setZero();
+    transform->makeAffine();
+  }
+};
+
+template<typename Other, int Mode, int Options, int Dim, int HDim>
+struct transform_construct_from_matrix<Other, Mode,Options,Dim,HDim, Dim,HDim>
+{
+  static inline void run(Transform<typename Other::Scalar,Dim,Mode,Options> *transform, const Other& other)
+  {
+    transform->affine() = other;
+    transform->makeAffine();
+  }
+};
+
+template<typename Other, int Mode, int Options, int Dim, int HDim>
+struct transform_construct_from_matrix<Other, Mode,Options,Dim,HDim, HDim,HDim>
+{
+  static inline void run(Transform<typename Other::Scalar,Dim,Mode,Options> *transform, const Other& other)
+  { transform->matrix() = other; }
+};
+
+template<typename Other, int Options, int Dim, int HDim>
+struct transform_construct_from_matrix<Other, AffineCompact,Options,Dim,HDim, HDim,HDim>
+{
+  static inline void run(Transform<typename Other::Scalar,Dim,AffineCompact,Options> *transform, const Other& other)
+  { transform->matrix() = other.template block<Dim,HDim>(0,0); }
+};
+
+/**********************************************************
+***   Specializations of operator* with rhs EigenBase   ***
+**********************************************************/
+
+template<int LhsMode,int RhsMode>
+struct transform_product_result
+{
+  enum 
+  { 
+    Mode =
+      (LhsMode == (int)Projective    || RhsMode == (int)Projective    ) ? Projective :
+      (LhsMode == (int)Affine        || RhsMode == (int)Affine        ) ? Affine :
+      (LhsMode == (int)AffineCompact || RhsMode == (int)AffineCompact ) ? AffineCompact :
+      (LhsMode == (int)Isometry      || RhsMode == (int)Isometry      ) ? Isometry : Projective
+  };
+};
+
+template< typename TransformType, typename MatrixType, int RhsCols>
+struct transform_right_product_impl< TransformType, MatrixType, 0, RhsCols>
+{
+  typedef typename MatrixType::PlainObject ResultType;
+
+  static EIGEN_STRONG_INLINE ResultType run(const TransformType& T, const MatrixType& other)
+  {
+    return T.matrix() * other;
+  }
+};
+
+template< typename TransformType, typename MatrixType, int RhsCols>
+struct transform_right_product_impl< TransformType, MatrixType, 1, RhsCols>
+{
+  enum { 
+    Dim = TransformType::Dim, 
+    HDim = TransformType::HDim,
+    OtherRows = MatrixType::RowsAtCompileTime,
+    OtherCols = MatrixType::ColsAtCompileTime
+  };
+
+  typedef typename MatrixType::PlainObject ResultType;
+
+  static EIGEN_STRONG_INLINE ResultType run(const TransformType& T, const MatrixType& other)
+  {
+    EIGEN_STATIC_ASSERT(OtherRows==HDim, YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES);
+
+    typedef Block<ResultType, Dim, OtherCols, int(MatrixType::RowsAtCompileTime)==Dim> TopLeftLhs;
+
+    ResultType res(other.rows(),other.cols());
+    TopLeftLhs(res, 0, 0, Dim, other.cols()).noalias() = T.affine() * other;
+    res.row(OtherRows-1) = other.row(OtherRows-1);
+    
+    return res;
+  }
+};
+
+template< typename TransformType, typename MatrixType, int RhsCols>
+struct transform_right_product_impl< TransformType, MatrixType, 2, RhsCols>
+{
+  enum { 
+    Dim = TransformType::Dim, 
+    HDim = TransformType::HDim,
+    OtherRows = MatrixType::RowsAtCompileTime,
+    OtherCols = MatrixType::ColsAtCompileTime
+  };
+
+  typedef typename MatrixType::PlainObject ResultType;
+
+  static EIGEN_STRONG_INLINE ResultType run(const TransformType& T, const MatrixType& other)
+  {
+    EIGEN_STATIC_ASSERT(OtherRows==Dim, YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES);
+
+    typedef Block<ResultType, Dim, OtherCols, true> TopLeftLhs;
+    ResultType res(Replicate<typename TransformType::ConstTranslationPart, 1, OtherCols>(T.translation(),1,other.cols()));
+    TopLeftLhs(res, 0, 0, Dim, other.cols()).noalias() += T.linear() * other;
+
+    return res;
+  }
+};
+
+template< typename TransformType, typename MatrixType >
+struct transform_right_product_impl< TransformType, MatrixType, 2, 1> // rhs is a vector of size Dim
+{
+  typedef typename TransformType::MatrixType TransformMatrix;
+  enum {
+    Dim = TransformType::Dim,
+    HDim = TransformType::HDim,
+    OtherRows = MatrixType::RowsAtCompileTime,
+    WorkingRows = EIGEN_PLAIN_ENUM_MIN(TransformMatrix::RowsAtCompileTime,HDim)
+  };
+
+  typedef typename MatrixType::PlainObject ResultType;
+
+  static EIGEN_STRONG_INLINE ResultType run(const TransformType& T, const MatrixType& other)
+  {
+    EIGEN_STATIC_ASSERT(OtherRows==Dim, YOU_MIXED_MATRICES_OF_DIFFERENT_SIZES);
+
+    Matrix<typename ResultType::Scalar, Dim+1, 1> rhs;
+    rhs.template head<Dim>() = other; rhs[Dim] = typename ResultType::Scalar(1);
+    Matrix<typename ResultType::Scalar, WorkingRows, 1> res(T.matrix() * rhs);
+    return res.template head<Dim>();
+  }
+};
+
+/**********************************************************
+***   Specializations of operator* with lhs EigenBase   ***
+**********************************************************/
+
+// generic HDim x HDim matrix * T => Projective
+template<typename Other,int Mode, int Options, int Dim, int HDim>
+struct transform_left_product_impl<Other,Mode,Options,Dim,HDim, HDim,HDim>
+{
+  typedef Transform<typename Other::Scalar,Dim,Mode,Options> TransformType;
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef Transform<typename Other::Scalar,Dim,Projective,Options> ResultType;
+  static ResultType run(const Other& other,const TransformType& tr)
+  { return ResultType(other * tr.matrix()); }
+};
+
+// generic HDim x HDim matrix * AffineCompact => Projective
+template<typename Other, int Options, int Dim, int HDim>
+struct transform_left_product_impl<Other,AffineCompact,Options,Dim,HDim, HDim,HDim>
+{
+  typedef Transform<typename Other::Scalar,Dim,AffineCompact,Options> TransformType;
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef Transform<typename Other::Scalar,Dim,Projective,Options> ResultType;
+  static ResultType run(const Other& other,const TransformType& tr)
+  {
+    ResultType res;
+    res.matrix().noalias() = other.template block<HDim,Dim>(0,0) * tr.matrix();
+    res.matrix().col(Dim) += other.col(Dim);
+    return res;
+  }
+};
+
+// affine matrix * T
+template<typename Other,int Mode, int Options, int Dim, int HDim>
+struct transform_left_product_impl<Other,Mode,Options,Dim,HDim, Dim,HDim>
+{
+  typedef Transform<typename Other::Scalar,Dim,Mode,Options> TransformType;
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef TransformType ResultType;
+  static ResultType run(const Other& other,const TransformType& tr)
+  {
+    ResultType res;
+    res.affine().noalias() = other * tr.matrix();
+    res.matrix().row(Dim) = tr.matrix().row(Dim);
+    return res;
+  }
+};
+
+// affine matrix * AffineCompact
+template<typename Other, int Options, int Dim, int HDim>
+struct transform_left_product_impl<Other,AffineCompact,Options,Dim,HDim, Dim,HDim>
+{
+  typedef Transform<typename Other::Scalar,Dim,AffineCompact,Options> TransformType;
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef TransformType ResultType;
+  static ResultType run(const Other& other,const TransformType& tr)
+  {
+    ResultType res;
+    res.matrix().noalias() = other.template block<Dim,Dim>(0,0) * tr.matrix();
+    res.translation() += other.col(Dim);
+    return res;
+  }
+};
+
+// linear matrix * T
+template<typename Other,int Mode, int Options, int Dim, int HDim>
+struct transform_left_product_impl<Other,Mode,Options,Dim,HDim, Dim,Dim>
+{
+  typedef Transform<typename Other::Scalar,Dim,Mode,Options> TransformType;
+  typedef typename TransformType::MatrixType MatrixType;
+  typedef TransformType ResultType;
+  static ResultType run(const Other& other, const TransformType& tr)
+  {
+    TransformType res;
+    if(Mode!=int(AffineCompact))
+      res.matrix().row(Dim) = tr.matrix().row(Dim);
+    res.matrix().template topRows<Dim>().noalias()
+      = other * tr.matrix().template topRows<Dim>();
+    return res;
+  }
+};
+
+/**********************************************************
+*** Specializations of operator* with another Transform ***
+**********************************************************/
+
+template<typename Scalar, int Dim, int LhsMode, int LhsOptions, int RhsMode, int RhsOptions>
+struct transform_transform_product_impl<Transform<Scalar,Dim,LhsMode,LhsOptions>,Transform<Scalar,Dim,RhsMode,RhsOptions>,false >
+{
+  enum { ResultMode = transform_product_result<LhsMode,RhsMode>::Mode };
+  typedef Transform<Scalar,Dim,LhsMode,LhsOptions> Lhs;
+  typedef Transform<Scalar,Dim,RhsMode,RhsOptions> Rhs;
+  typedef Transform<Scalar,Dim,ResultMode,LhsOptions> ResultType;
+  static ResultType run(const Lhs& lhs, const Rhs& rhs)
+  {
+    ResultType res;
+    res.linear() = lhs.linear() * rhs.linear();
+    res.translation() = lhs.linear() * rhs.translation() + lhs.translation();
+    res.makeAffine();
+    return res;
+  }
+};
+
+template<typename Scalar, int Dim, int LhsMode, int LhsOptions, int RhsMode, int RhsOptions>
+struct transform_transform_product_impl<Transform<Scalar,Dim,LhsMode,LhsOptions>,Transform<Scalar,Dim,RhsMode,RhsOptions>,true >
+{
+  typedef Transform<Scalar,Dim,LhsMode,LhsOptions> Lhs;
+  typedef Transform<Scalar,Dim,RhsMode,RhsOptions> Rhs;
+  typedef Transform<Scalar,Dim,Projective> ResultType;
+  static ResultType run(const Lhs& lhs, const Rhs& rhs)
+  {
+    return ResultType( lhs.matrix() * rhs.matrix() );
+  }
+};
+
+template<typename Scalar, int Dim, int LhsOptions, int RhsOptions>
+struct transform_transform_product_impl<Transform<Scalar,Dim,AffineCompact,LhsOptions>,Transform<Scalar,Dim,Projective,RhsOptions>,true >
+{
+  typedef Transform<Scalar,Dim,AffineCompact,LhsOptions> Lhs;
+  typedef Transform<Scalar,Dim,Projective,RhsOptions> Rhs;
+  typedef Transform<Scalar,Dim,Projective> ResultType;
+  static ResultType run(const Lhs& lhs, const Rhs& rhs)
+  {
+    ResultType res;
+    res.matrix().template topRows<Dim>() = lhs.matrix() * rhs.matrix();
+    res.matrix().row(Dim) = rhs.matrix().row(Dim);
+    return res;
+  }
+};
+
+template<typename Scalar, int Dim, int LhsOptions, int RhsOptions>
+struct transform_transform_product_impl<Transform<Scalar,Dim,Projective,LhsOptions>,Transform<Scalar,Dim,AffineCompact,RhsOptions>,true >
+{
+  typedef Transform<Scalar,Dim,Projective,LhsOptions> Lhs;
+  typedef Transform<Scalar,Dim,AffineCompact,RhsOptions> Rhs;
+  typedef Transform<Scalar,Dim,Projective> ResultType;
+  static ResultType run(const Lhs& lhs, const Rhs& rhs)
+  {
+    ResultType res(lhs.matrix().template leftCols<Dim>() * rhs.matrix());
+    res.matrix().col(Dim) += lhs.matrix().col(Dim);
+    return res;
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSFORM_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Translation.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Translation.h
new file mode 100644
index 0000000..51d9a82
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Translation.h
@@ -0,0 +1,208 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_TRANSLATION_H
+#define EIGEN_TRANSLATION_H
+
+namespace Eigen { 
+
+/** \geometry_module \ingroup Geometry_Module
+  *
+  * \class Translation
+  *
+  * \brief Represents a translation transformation
+  *
+  * \tparam _Scalar the scalar type, i.e., the type of the coefficients.
+  * \tparam _Dim the  dimension of the space, can be a compile time value or Dynamic
+  *
+  * \note This class is not aimed to be used to store a translation transformation,
+  * but rather to make easier the constructions and updates of Transform objects.
+  *
+  * \sa class Scaling, class Transform
+  */
+template<typename _Scalar, int _Dim>
+class Translation
+{
+public:
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_Dim)
+  /** dimension of the space */
+  enum { Dim = _Dim };
+  /** the scalar type of the coefficients */
+  typedef _Scalar Scalar;
+  /** corresponding vector type */
+  typedef Matrix<Scalar,Dim,1> VectorType;
+  /** corresponding linear transformation matrix type */
+  typedef Matrix<Scalar,Dim,Dim> LinearMatrixType;
+  /** corresponding affine transformation type */
+  typedef Transform<Scalar,Dim,Affine> AffineTransformType;
+  /** corresponding isometric transformation type */
+  typedef Transform<Scalar,Dim,Isometry> IsometryTransformType;
+
+protected:
+
+  VectorType m_coeffs;
+
+public:
+
+  /** Default constructor without initialization. */
+  EIGEN_DEVICE_FUNC Translation() {}
+  /**  */
+  EIGEN_DEVICE_FUNC inline Translation(const Scalar& sx, const Scalar& sy)
+  {
+    eigen_assert(Dim==2);
+    m_coeffs.x() = sx;
+    m_coeffs.y() = sy;
+  }
+  /**  */
+  EIGEN_DEVICE_FUNC inline Translation(const Scalar& sx, const Scalar& sy, const Scalar& sz)
+  {
+    eigen_assert(Dim==3);
+    m_coeffs.x() = sx;
+    m_coeffs.y() = sy;
+    m_coeffs.z() = sz;
+  }
+  /** Constructs and initialize the translation transformation from a vector of translation coefficients */
+  EIGEN_DEVICE_FUNC explicit inline Translation(const VectorType& vector) : m_coeffs(vector) {}
+
+  /** \brief Retruns the x-translation by value. **/
+  EIGEN_DEVICE_FUNC inline Scalar x() const { return m_coeffs.x(); }
+  /** \brief Retruns the y-translation by value. **/
+  EIGEN_DEVICE_FUNC inline Scalar y() const { return m_coeffs.y(); }
+  /** \brief Retruns the z-translation by value. **/
+  EIGEN_DEVICE_FUNC inline Scalar z() const { return m_coeffs.z(); }
+
+  /** \brief Retruns the x-translation as a reference. **/
+  EIGEN_DEVICE_FUNC inline Scalar& x() { return m_coeffs.x(); }
+  /** \brief Retruns the y-translation as a reference. **/
+  EIGEN_DEVICE_FUNC inline Scalar& y() { return m_coeffs.y(); }
+  /** \brief Retruns the z-translation as a reference. **/
+  EIGEN_DEVICE_FUNC inline Scalar& z() { return m_coeffs.z(); }
+
+  EIGEN_DEVICE_FUNC const VectorType& vector() const { return m_coeffs; }
+  EIGEN_DEVICE_FUNC VectorType& vector() { return m_coeffs; }
+
+  EIGEN_DEVICE_FUNC const VectorType& translation() const { return m_coeffs; }
+  EIGEN_DEVICE_FUNC VectorType& translation() { return m_coeffs; }
+
+  /** Concatenates two translation */
+  EIGEN_DEVICE_FUNC inline Translation operator* (const Translation& other) const
+  { return Translation(m_coeffs + other.m_coeffs); }
+
+  /** Concatenates a translation and a uniform scaling */
+  EIGEN_DEVICE_FUNC inline AffineTransformType operator* (const UniformScaling<Scalar>& other) const;
+
+  /** Concatenates a translation and a linear transformation */
+  template<typename OtherDerived>
+  EIGEN_DEVICE_FUNC inline AffineTransformType operator* (const EigenBase<OtherDerived>& linear) const;
+
+  /** Concatenates a translation and a rotation */
+  template<typename Derived>
+  EIGEN_DEVICE_FUNC inline IsometryTransformType operator*(const RotationBase<Derived,Dim>& r) const
+  { return *this * IsometryTransformType(r); }
+
+  /** \returns the concatenation of a linear transformation \a l with the translation \a t */
+  // its a nightmare to define a templated friend function outside its declaration
+  template<typename OtherDerived> friend
+  EIGEN_DEVICE_FUNC inline AffineTransformType operator*(const EigenBase<OtherDerived>& linear, const Translation& t)
+  {
+    AffineTransformType res;
+    res.matrix().setZero();
+    res.linear() = linear.derived();
+    res.translation() = linear.derived() * t.m_coeffs;
+    res.matrix().row(Dim).setZero();
+    res(Dim,Dim) = Scalar(1);
+    return res;
+  }
+
+  /** Concatenates a translation and a transformation */
+  template<int Mode, int Options>
+  EIGEN_DEVICE_FUNC inline Transform<Scalar,Dim,Mode> operator* (const Transform<Scalar,Dim,Mode,Options>& t) const
+  {
+    Transform<Scalar,Dim,Mode> res = t;
+    res.pretranslate(m_coeffs);
+    return res;
+  }
+
+  /** Applies translation to vector */
+  template<typename Derived>
+  inline typename internal::enable_if<Derived::IsVectorAtCompileTime,VectorType>::type
+  operator* (const MatrixBase<Derived>& vec) const
+  { return m_coeffs + vec.derived(); }
+
+  /** \returns the inverse translation (opposite) */
+  Translation inverse() const { return Translation(-m_coeffs); }
+
+  Translation& operator=(const Translation& other)
+  {
+    m_coeffs = other.m_coeffs;
+    return *this;
+  }
+
+  static const Translation Identity() { return Translation(VectorType::Zero()); }
+
+  /** \returns \c *this with scalar type casted to \a NewScalarType
+    *
+    * Note that if \a NewScalarType is equal to the current scalar type of \c *this
+    * then this function smartly returns a const reference to \c *this.
+    */
+  template<typename NewScalarType>
+  EIGEN_DEVICE_FUNC inline typename internal::cast_return_type<Translation,Translation<NewScalarType,Dim> >::type cast() const
+  { return typename internal::cast_return_type<Translation,Translation<NewScalarType,Dim> >::type(*this); }
+
+  /** Copy constructor with scalar type conversion */
+  template<typename OtherScalarType>
+  EIGEN_DEVICE_FUNC inline explicit Translation(const Translation<OtherScalarType,Dim>& other)
+  { m_coeffs = other.vector().template cast<Scalar>(); }
+
+  /** \returns \c true if \c *this is approximately equal to \a other, within the precision
+    * determined by \a prec.
+    *
+    * \sa MatrixBase::isApprox() */
+  EIGEN_DEVICE_FUNC bool isApprox(const Translation& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
+  { return m_coeffs.isApprox(other.m_coeffs, prec); }
+
+};
+
+/** \addtogroup Geometry_Module */
+//@{
+typedef Translation<float, 2> Translation2f;
+typedef Translation<double,2> Translation2d;
+typedef Translation<float, 3> Translation3f;
+typedef Translation<double,3> Translation3d;
+//@}
+
+template<typename Scalar, int Dim>
+EIGEN_DEVICE_FUNC inline typename Translation<Scalar,Dim>::AffineTransformType
+Translation<Scalar,Dim>::operator* (const UniformScaling<Scalar>& other) const
+{
+  AffineTransformType res;
+  res.matrix().setZero();
+  res.linear().diagonal().fill(other.factor());
+  res.translation() = m_coeffs;
+  res(Dim,Dim) = Scalar(1);
+  return res;
+}
+
+template<typename Scalar, int Dim>
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC inline typename Translation<Scalar,Dim>::AffineTransformType
+Translation<Scalar,Dim>::operator* (const EigenBase<OtherDerived>& linear) const
+{
+  AffineTransformType res;
+  res.matrix().setZero();
+  res.linear() = linear.derived();
+  res.translation() = m_coeffs;
+  res.matrix().row(Dim).setZero();
+  res(Dim,Dim) = Scalar(1);
+  return res;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_TRANSLATION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Umeyama.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Umeyama.h
new file mode 100644
index 0000000..7e933fc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/Umeyama.h
@@ -0,0 +1,166 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_UMEYAMA_H
+#define EIGEN_UMEYAMA_H
+
+// This file requires the user to include 
+// * Eigen/Core
+// * Eigen/LU 
+// * Eigen/SVD
+// * Eigen/Array
+
+namespace Eigen { 
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+// These helpers are required since it allows to use mixed types as parameters
+// for the Umeyama. The problem with mixed parameters is that the return type
+// cannot trivially be deduced when float and double types are mixed.
+namespace internal {
+
+// Compile time return type deduction for different MatrixBase types.
+// Different means here different alignment and parameters but the same underlying
+// real scalar type.
+template<typename MatrixType, typename OtherMatrixType>
+struct umeyama_transform_matrix_type
+{
+  enum {
+    MinRowsAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(MatrixType::RowsAtCompileTime, OtherMatrixType::RowsAtCompileTime),
+
+    // When possible we want to choose some small fixed size value since the result
+    // is likely to fit on the stack. So here, EIGEN_SIZE_MIN_PREFER_DYNAMIC is not what we want.
+    HomogeneousDimension = int(MinRowsAtCompileTime) == Dynamic ? Dynamic : int(MinRowsAtCompileTime)+1
+  };
+
+  typedef Matrix<typename traits<MatrixType>::Scalar,
+    HomogeneousDimension,
+    HomogeneousDimension,
+    AutoAlign | (traits<MatrixType>::Flags & RowMajorBit ? RowMajor : ColMajor),
+    HomogeneousDimension,
+    HomogeneousDimension
+  > type;
+};
+
+}
+
+#endif
+
+/**
+* \geometry_module \ingroup Geometry_Module
+*
+* \brief Returns the transformation between two point sets.
+*
+* The algorithm is based on:
+* "Least-squares estimation of transformation parameters between two point patterns",
+* Shinji Umeyama, PAMI 1991, DOI: 10.1109/34.88573
+*
+* It estimates parameters \f$ c, \mathbf{R}, \f$ and \f$ \mathbf{t} \f$ such that
+* \f{align*}
+*   \frac{1}{n} \sum_{i=1}^n \vert\vert y_i - (c\mathbf{R}x_i + \mathbf{t}) \vert\vert_2^2
+* \f}
+* is minimized.
+*
+* The algorithm is based on the analysis of the covariance matrix
+* \f$ \Sigma_{\mathbf{x}\mathbf{y}} \in \mathbb{R}^{d \times d} \f$
+* of the input point sets \f$ \mathbf{x} \f$ and \f$ \mathbf{y} \f$ where 
+* \f$d\f$ is corresponding to the dimension (which is typically small).
+* The analysis is involving the SVD having a complexity of \f$O(d^3)\f$
+* though the actual computational effort lies in the covariance
+* matrix computation which has an asymptotic lower bound of \f$O(dm)\f$ when 
+* the input point sets have dimension \f$d \times m\f$.
+*
+* Currently the method is working only for floating point matrices.
+*
+* \todo Should the return type of umeyama() become a Transform?
+*
+* \param src Source points \f$ \mathbf{x} = \left( x_1, \hdots, x_n \right) \f$.
+* \param dst Destination points \f$ \mathbf{y} = \left( y_1, \hdots, y_n \right) \f$.
+* \param with_scaling Sets \f$ c=1 \f$ when <code>false</code> is passed.
+* \return The homogeneous transformation 
+* \f{align*}
+*   T = \begin{bmatrix} c\mathbf{R} & \mathbf{t} \\ \mathbf{0} & 1 \end{bmatrix}
+* \f}
+* minimizing the resudiual above. This transformation is always returned as an 
+* Eigen::Matrix.
+*/
+template <typename Derived, typename OtherDerived>
+typename internal::umeyama_transform_matrix_type<Derived, OtherDerived>::type
+umeyama(const MatrixBase<Derived>& src, const MatrixBase<OtherDerived>& dst, bool with_scaling = true)
+{
+  typedef typename internal::umeyama_transform_matrix_type<Derived, OtherDerived>::type TransformationMatrixType;
+  typedef typename internal::traits<TransformationMatrixType>::Scalar Scalar;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+
+  EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL)
+  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename internal::traits<OtherDerived>::Scalar>::value),
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+  enum { Dimension = EIGEN_SIZE_MIN_PREFER_DYNAMIC(Derived::RowsAtCompileTime, OtherDerived::RowsAtCompileTime) };
+
+  typedef Matrix<Scalar, Dimension, 1> VectorType;
+  typedef Matrix<Scalar, Dimension, Dimension> MatrixType;
+  typedef typename internal::plain_matrix_type_row_major<Derived>::type RowMajorMatrixType;
+
+  const Index m = src.rows(); // dimension
+  const Index n = src.cols(); // number of measurements
+
+  // required for demeaning ...
+  const RealScalar one_over_n = RealScalar(1) / static_cast<RealScalar>(n);
+
+  // computation of mean
+  const VectorType src_mean = src.rowwise().sum() * one_over_n;
+  const VectorType dst_mean = dst.rowwise().sum() * one_over_n;
+
+  // demeaning of src and dst points
+  const RowMajorMatrixType src_demean = src.colwise() - src_mean;
+  const RowMajorMatrixType dst_demean = dst.colwise() - dst_mean;
+
+  // Eq. (36)-(37)
+  const Scalar src_var = src_demean.rowwise().squaredNorm().sum() * one_over_n;
+
+  // Eq. (38)
+  const MatrixType sigma = one_over_n * dst_demean * src_demean.transpose();
+
+  JacobiSVD<MatrixType> svd(sigma, ComputeFullU | ComputeFullV);
+
+  // Initialize the resulting transformation with an identity matrix...
+  TransformationMatrixType Rt = TransformationMatrixType::Identity(m+1,m+1);
+
+  // Eq. (39)
+  VectorType S = VectorType::Ones(m);
+
+  if  ( svd.matrixU().determinant() * svd.matrixV().determinant() < 0 )
+    S(m-1) = -1;
+
+  // Eq. (40) and (43)
+  Rt.block(0,0,m,m).noalias() = svd.matrixU() * S.asDiagonal() * svd.matrixV().transpose();
+
+  if (with_scaling)
+  {
+    // Eq. (42)
+    const Scalar c = Scalar(1)/src_var * svd.singularValues().dot(S);
+
+    // Eq. (41)
+    Rt.col(m).head(m) = dst_mean;
+    Rt.col(m).head(m).noalias() -= c*Rt.topLeftCorner(m,m)*src_mean;
+    Rt.block(0,0,m,m) *= c;
+  }
+  else
+  {
+    Rt.col(m).head(m) = dst_mean;
+    Rt.col(m).head(m).noalias() -= Rt.topLeftCorner(m,m)*src_mean;
+  }
+
+  return Rt;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_UMEYAMA_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h
new file mode 100644
index 0000000..f68cab5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Geometry/arch/Geometry_SSE.h
@@ -0,0 +1,161 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Rohit Garg <rpg.314@gmail.com>
+// Copyright (C) 2009-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GEOMETRY_SSE_H
+#define EIGEN_GEOMETRY_SSE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<class Derived, class OtherDerived>
+struct quat_product<Architecture::SSE, Derived, OtherDerived, float>
+{
+  enum {
+    AAlignment = traits<Derived>::Alignment,
+    BAlignment = traits<OtherDerived>::Alignment,
+    ResAlignment = traits<Quaternion<float> >::Alignment
+  };
+  static inline Quaternion<float> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
+  {
+    Quaternion<float> res;
+    const __m128 mask = _mm_setr_ps(0.f,0.f,0.f,-0.f);
+    __m128 a = _a.coeffs().template packet<AAlignment>(0);
+    __m128 b = _b.coeffs().template packet<BAlignment>(0);
+    __m128 s1 = _mm_mul_ps(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
+    __m128 s2 = _mm_mul_ps(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
+    pstoret<float,Packet4f,ResAlignment>(
+              &res.x(),
+              _mm_add_ps(_mm_sub_ps(_mm_mul_ps(a,vec4f_swizzle1(b,3,3,3,3)),
+                                    _mm_mul_ps(vec4f_swizzle1(a,2,0,1,0),
+                                               vec4f_swizzle1(b,1,2,0,0))),
+                         _mm_xor_ps(mask,_mm_add_ps(s1,s2))));
+    
+    return res;
+  }
+};
+
+template<class Derived>
+struct quat_conj<Architecture::SSE, Derived, float>
+{
+  enum {
+    ResAlignment = traits<Quaternion<float> >::Alignment
+  };
+  static inline Quaternion<float> run(const QuaternionBase<Derived>& q)
+  {
+    Quaternion<float> res;
+    const __m128 mask = _mm_setr_ps(-0.f,-0.f,-0.f,0.f);
+    pstoret<float,Packet4f,ResAlignment>(&res.x(), _mm_xor_ps(mask, q.coeffs().template packet<traits<Derived>::Alignment>(0)));
+    return res;
+  }
+};
+
+
+template<typename VectorLhs,typename VectorRhs>
+struct cross3_impl<Architecture::SSE,VectorLhs,VectorRhs,float,true>
+{
+  enum {
+    ResAlignment = traits<typename plain_matrix_type<VectorLhs>::type>::Alignment
+  };
+  static inline typename plain_matrix_type<VectorLhs>::type
+  run(const VectorLhs& lhs, const VectorRhs& rhs)
+  {
+    __m128 a = lhs.template packet<traits<VectorLhs>::Alignment>(0);
+    __m128 b = rhs.template packet<traits<VectorRhs>::Alignment>(0);
+    __m128 mul1=_mm_mul_ps(vec4f_swizzle1(a,1,2,0,3),vec4f_swizzle1(b,2,0,1,3));
+    __m128 mul2=_mm_mul_ps(vec4f_swizzle1(a,2,0,1,3),vec4f_swizzle1(b,1,2,0,3));
+    typename plain_matrix_type<VectorLhs>::type res;
+    pstoret<float,Packet4f,ResAlignment>(&res.x(),_mm_sub_ps(mul1,mul2));
+    return res;
+  }
+};
+
+
+
+
+template<class Derived, class OtherDerived>
+struct quat_product<Architecture::SSE, Derived, OtherDerived, double>
+{
+  enum {
+    BAlignment = traits<OtherDerived>::Alignment,
+    ResAlignment = traits<Quaternion<double> >::Alignment
+  };
+
+  static inline Quaternion<double> run(const QuaternionBase<Derived>& _a, const QuaternionBase<OtherDerived>& _b)
+  {
+  const Packet2d mask = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
+
+  Quaternion<double> res;
+
+  const double* a = _a.coeffs().data();
+  Packet2d b_xy = _b.coeffs().template packet<BAlignment>(0);
+  Packet2d b_zw = _b.coeffs().template packet<BAlignment>(2);
+  Packet2d a_xx = pset1<Packet2d>(a[0]);
+  Packet2d a_yy = pset1<Packet2d>(a[1]);
+  Packet2d a_zz = pset1<Packet2d>(a[2]);
+  Packet2d a_ww = pset1<Packet2d>(a[3]);
+
+  // two temporaries:
+  Packet2d t1, t2;
+
+  /*
+   * t1 = ww*xy + yy*zw
+   * t2 = zz*xy - xx*zw
+   * res.xy = t1 +/- swap(t2)
+   */
+  t1 = padd(pmul(a_ww, b_xy), pmul(a_yy, b_zw));
+  t2 = psub(pmul(a_zz, b_xy), pmul(a_xx, b_zw));
+#ifdef EIGEN_VECTORIZE_SSE3
+  EIGEN_UNUSED_VARIABLE(mask)
+  pstoret<double,Packet2d,ResAlignment>(&res.x(), _mm_addsub_pd(t1, preverse(t2)));
+#else
+  pstoret<double,Packet2d,ResAlignment>(&res.x(), padd(t1, pxor(mask,preverse(t2))));
+#endif
+  
+  /*
+   * t1 = ww*zw - yy*xy
+   * t2 = zz*zw + xx*xy
+   * res.zw = t1 -/+ swap(t2) = swap( swap(t1) +/- t2)
+   */
+  t1 = psub(pmul(a_ww, b_zw), pmul(a_yy, b_xy));
+  t2 = padd(pmul(a_zz, b_zw), pmul(a_xx, b_xy));
+#ifdef EIGEN_VECTORIZE_SSE3
+  EIGEN_UNUSED_VARIABLE(mask)
+  pstoret<double,Packet2d,ResAlignment>(&res.z(), preverse(_mm_addsub_pd(preverse(t1), t2)));
+#else
+  pstoret<double,Packet2d,ResAlignment>(&res.z(), psub(t1, pxor(mask,preverse(t2))));
+#endif
+
+  return res;
+}
+};
+
+template<class Derived>
+struct quat_conj<Architecture::SSE, Derived, double>
+{
+  enum {
+    ResAlignment = traits<Quaternion<double> >::Alignment
+  };
+  static inline Quaternion<double> run(const QuaternionBase<Derived>& q)
+  {
+    Quaternion<double> res;
+    const __m128d mask0 = _mm_setr_pd(-0.,-0.);
+    const __m128d mask2 = _mm_setr_pd(-0.,0.);
+    pstoret<double,Packet2d,ResAlignment>(&res.x(), _mm_xor_pd(mask0, q.coeffs().template packet<traits<Derived>::Alignment>(0)));
+    pstoret<double,Packet2d,ResAlignment>(&res.z(), _mm_xor_pd(mask2, q.coeffs().template packet<traits<Derived>::Alignment>(2)));
+    return res;
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_GEOMETRY_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/BlockHouseholder.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/BlockHouseholder.h
new file mode 100644
index 0000000..01a7ed1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/BlockHouseholder.h
@@ -0,0 +1,103 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Vincent Lejeune
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BLOCK_HOUSEHOLDER_H
+#define EIGEN_BLOCK_HOUSEHOLDER_H
+
+// This file contains some helper function to deal with block householder reflectors
+
+namespace Eigen { 
+
+namespace internal {
+  
+/** \internal */
+// template<typename TriangularFactorType,typename VectorsType,typename CoeffsType>
+// void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs)
+// {
+//   typedef typename VectorsType::Scalar Scalar;
+//   const Index nbVecs = vectors.cols();
+//   eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs);
+// 
+//   for(Index i = 0; i < nbVecs; i++)
+//   {
+//     Index rs = vectors.rows() - i;
+//     // Warning, note that hCoeffs may alias with vectors.
+//     // It is then necessary to copy it before modifying vectors(i,i). 
+//     typename CoeffsType::Scalar h = hCoeffs(i);
+//     // This hack permits to pass trough nested Block<> and Transpose<> expressions.
+//     Scalar *Vii_ptr = const_cast<Scalar*>(vectors.data() + vectors.outerStride()*i + vectors.innerStride()*i);
+//     Scalar Vii = *Vii_ptr;
+//     *Vii_ptr = Scalar(1);
+//     triFactor.col(i).head(i).noalias() = -h * vectors.block(i, 0, rs, i).adjoint()
+//                                        * vectors.col(i).tail(rs);
+//     *Vii_ptr = Vii;
+//     // FIXME add .noalias() once the triangular product can work inplace
+//     triFactor.col(i).head(i) = triFactor.block(0,0,i,i).template triangularView<Upper>()
+//                              * triFactor.col(i).head(i);
+//     triFactor(i,i) = hCoeffs(i);
+//   }
+// }
+
+/** \internal */
+// This variant avoid modifications in vectors
+template<typename TriangularFactorType,typename VectorsType,typename CoeffsType>
+void make_block_householder_triangular_factor(TriangularFactorType& triFactor, const VectorsType& vectors, const CoeffsType& hCoeffs)
+{
+  const Index nbVecs = vectors.cols();
+  eigen_assert(triFactor.rows() == nbVecs && triFactor.cols() == nbVecs && vectors.rows()>=nbVecs);
+
+  for(Index i = nbVecs-1; i >=0 ; --i)
+  {
+    Index rs = vectors.rows() - i - 1;
+    Index rt = nbVecs-i-1;
+
+    if(rt>0)
+    {
+      triFactor.row(i).tail(rt).noalias() = -hCoeffs(i) * vectors.col(i).tail(rs).adjoint()
+                                                        * vectors.bottomRightCorner(rs, rt).template triangularView<UnitLower>();
+            
+      // FIXME add .noalias() once the triangular product can work inplace
+      triFactor.row(i).tail(rt) = triFactor.row(i).tail(rt) * triFactor.bottomRightCorner(rt,rt).template triangularView<Upper>();
+      
+    }
+    triFactor(i,i) = hCoeffs(i);
+  }
+}
+
+/** \internal
+  * if forward then perform   mat = H0 * H1 * H2 * mat
+  * otherwise perform         mat = H2 * H1 * H0 * mat
+  */
+template<typename MatrixType,typename VectorsType,typename CoeffsType>
+void apply_block_householder_on_the_left(MatrixType& mat, const VectorsType& vectors, const CoeffsType& hCoeffs, bool forward)
+{
+  enum { TFactorSize = MatrixType::ColsAtCompileTime };
+  Index nbVecs = vectors.cols();
+  Matrix<typename MatrixType::Scalar, TFactorSize, TFactorSize, RowMajor> T(nbVecs,nbVecs);
+  
+  if(forward) make_block_householder_triangular_factor(T, vectors, hCoeffs);
+  else        make_block_householder_triangular_factor(T, vectors, hCoeffs.conjugate());  
+  const TriangularView<const VectorsType, UnitLower> V(vectors);
+
+  // A -= V T V^* A
+  Matrix<typename MatrixType::Scalar,VectorsType::ColsAtCompileTime,MatrixType::ColsAtCompileTime,
+         (VectorsType::MaxColsAtCompileTime==1 && MatrixType::MaxColsAtCompileTime!=1)?RowMajor:ColMajor,
+         VectorsType::MaxColsAtCompileTime,MatrixType::MaxColsAtCompileTime> tmp = V.adjoint() * mat;
+  // FIXME add .noalias() once the triangular product can work inplace
+  if(forward) tmp = T.template triangularView<Upper>()           * tmp;
+  else        tmp = T.template triangularView<Upper>().adjoint() * tmp;
+  mat.noalias() -= V * tmp;
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BLOCK_HOUSEHOLDER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/Householder.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/Householder.h
new file mode 100644
index 0000000..80de2c3
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/Householder.h
@@ -0,0 +1,172 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HOUSEHOLDER_H
+#define EIGEN_HOUSEHOLDER_H
+
+namespace Eigen { 
+
+namespace internal {
+template<int n> struct decrement_size
+{
+  enum {
+    ret = n==Dynamic ? n : n-1
+  };
+};
+}
+
+/** Computes the elementary reflector H such that:
+  * \f$ H *this = [ beta 0 ... 0]^T \f$
+  * where the transformation H is:
+  * \f$ H = I - tau v v^*\f$
+  * and the vector v is:
+  * \f$ v^T = [1 essential^T] \f$
+  *
+  * The essential part of the vector \c v is stored in *this.
+  * 
+  * On output:
+  * \param tau the scaling factor of the Householder transformation
+  * \param beta the result of H * \c *this
+  *
+  * \sa MatrixBase::makeHouseholder(), MatrixBase::applyHouseholderOnTheLeft(),
+  *     MatrixBase::applyHouseholderOnTheRight()
+  */
+template<typename Derived>
+void MatrixBase<Derived>::makeHouseholderInPlace(Scalar& tau, RealScalar& beta)
+{
+  VectorBlock<Derived, internal::decrement_size<Base::SizeAtCompileTime>::ret> essentialPart(derived(), 1, size()-1);
+  makeHouseholder(essentialPart, tau, beta);
+}
+
+/** Computes the elementary reflector H such that:
+  * \f$ H *this = [ beta 0 ... 0]^T \f$
+  * where the transformation H is:
+  * \f$ H = I - tau v v^*\f$
+  * and the vector v is:
+  * \f$ v^T = [1 essential^T] \f$
+  *
+  * On output:
+  * \param essential the essential part of the vector \c v
+  * \param tau the scaling factor of the Householder transformation
+  * \param beta the result of H * \c *this
+  *
+  * \sa MatrixBase::makeHouseholderInPlace(), MatrixBase::applyHouseholderOnTheLeft(),
+  *     MatrixBase::applyHouseholderOnTheRight()
+  */
+template<typename Derived>
+template<typename EssentialPart>
+void MatrixBase<Derived>::makeHouseholder(
+  EssentialPart& essential,
+  Scalar& tau,
+  RealScalar& beta) const
+{
+  using std::sqrt;
+  using numext::conj;
+  
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(EssentialPart)
+  VectorBlock<const Derived, EssentialPart::SizeAtCompileTime> tail(derived(), 1, size()-1);
+  
+  RealScalar tailSqNorm = size()==1 ? RealScalar(0) : tail.squaredNorm();
+  Scalar c0 = coeff(0);
+  const RealScalar tol = (std::numeric_limits<RealScalar>::min)();
+
+  if(tailSqNorm <= tol && numext::abs2(numext::imag(c0))<=tol)
+  {
+    tau = RealScalar(0);
+    beta = numext::real(c0);
+    essential.setZero();
+  }
+  else
+  {
+    beta = sqrt(numext::abs2(c0) + tailSqNorm);
+    if (numext::real(c0)>=RealScalar(0))
+      beta = -beta;
+    essential = tail / (c0 - beta);
+    tau = conj((beta - c0) / beta);
+  }
+}
+
+/** Apply the elementary reflector H given by
+  * \f$ H = I - tau v v^*\f$
+  * with
+  * \f$ v^T = [1 essential^T] \f$
+  * from the left to a vector or matrix.
+  *
+  * On input:
+  * \param essential the essential part of the vector \c v
+  * \param tau the scaling factor of the Householder transformation
+  * \param workspace a pointer to working space with at least
+  *                  this->cols() * essential.size() entries
+  *
+  * \sa MatrixBase::makeHouseholder(), MatrixBase::makeHouseholderInPlace(), 
+  *     MatrixBase::applyHouseholderOnTheRight()
+  */
+template<typename Derived>
+template<typename EssentialPart>
+void MatrixBase<Derived>::applyHouseholderOnTheLeft(
+  const EssentialPart& essential,
+  const Scalar& tau,
+  Scalar* workspace)
+{
+  if(rows() == 1)
+  {
+    *this *= Scalar(1)-tau;
+  }
+  else if(tau!=Scalar(0))
+  {
+    Map<typename internal::plain_row_type<PlainObject>::type> tmp(workspace,cols());
+    Block<Derived, EssentialPart::SizeAtCompileTime, Derived::ColsAtCompileTime> bottom(derived(), 1, 0, rows()-1, cols());
+    tmp.noalias() = essential.adjoint() * bottom;
+    tmp += this->row(0);
+    this->row(0) -= tau * tmp;
+    bottom.noalias() -= tau * essential * tmp;
+  }
+}
+
+/** Apply the elementary reflector H given by
+  * \f$ H = I - tau v v^*\f$
+  * with
+  * \f$ v^T = [1 essential^T] \f$
+  * from the right to a vector or matrix.
+  *
+  * On input:
+  * \param essential the essential part of the vector \c v
+  * \param tau the scaling factor of the Householder transformation
+  * \param workspace a pointer to working space with at least
+  *                  this->cols() * essential.size() entries
+  *
+  * \sa MatrixBase::makeHouseholder(), MatrixBase::makeHouseholderInPlace(), 
+  *     MatrixBase::applyHouseholderOnTheLeft()
+  */
+template<typename Derived>
+template<typename EssentialPart>
+void MatrixBase<Derived>::applyHouseholderOnTheRight(
+  const EssentialPart& essential,
+  const Scalar& tau,
+  Scalar* workspace)
+{
+  if(cols() == 1)
+  {
+    *this *= Scalar(1)-tau;
+  }
+  else if(tau!=Scalar(0))
+  {
+    Map<typename internal::plain_col_type<PlainObject>::type> tmp(workspace,rows());
+    Block<Derived, Derived::RowsAtCompileTime, EssentialPart::SizeAtCompileTime> right(derived(), 0, 1, rows(), cols()-1);
+    tmp.noalias() = right * essential.conjugate();
+    tmp += this->col(0);
+    this->col(0) -= tau * tmp;
+    right.noalias() -= tau * tmp * essential.transpose();
+  }
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_HOUSEHOLDER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/HouseholderSequence.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/HouseholderSequence.h
new file mode 100644
index 0000000..3ce0a69
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Householder/HouseholderSequence.h
@@ -0,0 +1,470 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_HOUSEHOLDER_SEQUENCE_H
+#define EIGEN_HOUSEHOLDER_SEQUENCE_H
+
+namespace Eigen { 
+
+/** \ingroup Householder_Module
+  * \householder_module
+  * \class HouseholderSequence
+  * \brief Sequence of Householder reflections acting on subspaces with decreasing size
+  * \tparam VectorsType type of matrix containing the Householder vectors
+  * \tparam CoeffsType  type of vector containing the Householder coefficients
+  * \tparam Side        either OnTheLeft (the default) or OnTheRight
+  *
+  * This class represents a product sequence of Householder reflections where the first Householder reflection
+  * acts on the whole space, the second Householder reflection leaves the one-dimensional subspace spanned by
+  * the first unit vector invariant, the third Householder reflection leaves the two-dimensional subspace
+  * spanned by the first two unit vectors invariant, and so on up to the last reflection which leaves all but
+  * one dimensions invariant and acts only on the last dimension. Such sequences of Householder reflections
+  * are used in several algorithms to zero out certain parts of a matrix. Indeed, the methods
+  * HessenbergDecomposition::matrixQ(), Tridiagonalization::matrixQ(), HouseholderQR::householderQ(),
+  * and ColPivHouseholderQR::householderQ() all return a %HouseholderSequence.
+  *
+  * More precisely, the class %HouseholderSequence represents an \f$ n \times n \f$ matrix \f$ H \f$ of the
+  * form \f$ H = \prod_{i=0}^{n-1} H_i \f$ where the i-th Householder reflection is \f$ H_i = I - h_i v_i
+  * v_i^* \f$. The i-th Householder coefficient \f$ h_i \f$ is a scalar and the i-th Householder vector \f$
+  * v_i \f$ is a vector of the form
+  * \f[ 
+  * v_i = [\underbrace{0, \ldots, 0}_{i-1\mbox{ zeros}}, 1, \underbrace{*, \ldots,*}_{n-i\mbox{ arbitrary entries}} ]. 
+  * \f]
+  * The last \f$ n-i \f$ entries of \f$ v_i \f$ are called the essential part of the Householder vector.
+  *
+  * Typical usages are listed below, where H is a HouseholderSequence:
+  * \code
+  * A.applyOnTheRight(H);             // A = A * H
+  * A.applyOnTheLeft(H);              // A = H * A
+  * A.applyOnTheRight(H.adjoint());   // A = A * H^*
+  * A.applyOnTheLeft(H.adjoint());    // A = H^* * A
+  * MatrixXd Q = H;                   // conversion to a dense matrix
+  * \endcode
+  * In addition to the adjoint, you can also apply the inverse (=adjoint), the transpose, and the conjugate operators.
+  *
+  * See the documentation for HouseholderSequence(const VectorsType&, const CoeffsType&) for an example.
+  *
+  * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+
+namespace internal {
+
+template<typename VectorsType, typename CoeffsType, int Side>
+struct traits<HouseholderSequence<VectorsType,CoeffsType,Side> >
+{
+  typedef typename VectorsType::Scalar Scalar;
+  typedef typename VectorsType::StorageIndex StorageIndex;
+  typedef typename VectorsType::StorageKind StorageKind;
+  enum {
+    RowsAtCompileTime = Side==OnTheLeft ? traits<VectorsType>::RowsAtCompileTime
+                                        : traits<VectorsType>::ColsAtCompileTime,
+    ColsAtCompileTime = RowsAtCompileTime,
+    MaxRowsAtCompileTime = Side==OnTheLeft ? traits<VectorsType>::MaxRowsAtCompileTime
+                                           : traits<VectorsType>::MaxColsAtCompileTime,
+    MaxColsAtCompileTime = MaxRowsAtCompileTime,
+    Flags = 0
+  };
+};
+
+struct HouseholderSequenceShape {};
+
+template<typename VectorsType, typename CoeffsType, int Side>
+struct evaluator_traits<HouseholderSequence<VectorsType,CoeffsType,Side> >
+  : public evaluator_traits_base<HouseholderSequence<VectorsType,CoeffsType,Side> >
+{
+  typedef HouseholderSequenceShape Shape;
+};
+
+template<typename VectorsType, typename CoeffsType, int Side>
+struct hseq_side_dependent_impl
+{
+  typedef Block<const VectorsType, Dynamic, 1> EssentialVectorType;
+  typedef HouseholderSequence<VectorsType, CoeffsType, OnTheLeft> HouseholderSequenceType;
+  static inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k)
+  {
+    Index start = k+1+h.m_shift;
+    return Block<const VectorsType,Dynamic,1>(h.m_vectors, start, k, h.rows()-start, 1);
+  }
+};
+
+template<typename VectorsType, typename CoeffsType>
+struct hseq_side_dependent_impl<VectorsType, CoeffsType, OnTheRight>
+{
+  typedef Transpose<Block<const VectorsType, 1, Dynamic> > EssentialVectorType;
+  typedef HouseholderSequence<VectorsType, CoeffsType, OnTheRight> HouseholderSequenceType;
+  static inline const EssentialVectorType essentialVector(const HouseholderSequenceType& h, Index k)
+  {
+    Index start = k+1+h.m_shift;
+    return Block<const VectorsType,1,Dynamic>(h.m_vectors, k, start, 1, h.rows()-start).transpose();
+  }
+};
+
+template<typename OtherScalarType, typename MatrixType> struct matrix_type_times_scalar_type
+{
+  typedef typename ScalarBinaryOpTraits<OtherScalarType, typename MatrixType::Scalar>::ReturnType
+    ResultScalar;
+  typedef Matrix<ResultScalar, MatrixType::RowsAtCompileTime, MatrixType::ColsAtCompileTime,
+                 0, MatrixType::MaxRowsAtCompileTime, MatrixType::MaxColsAtCompileTime> Type;
+};
+
+} // end namespace internal
+
+template<typename VectorsType, typename CoeffsType, int Side> class HouseholderSequence
+  : public EigenBase<HouseholderSequence<VectorsType,CoeffsType,Side> >
+{
+    typedef typename internal::hseq_side_dependent_impl<VectorsType,CoeffsType,Side>::EssentialVectorType EssentialVectorType;
+  
+  public:
+    enum {
+      RowsAtCompileTime = internal::traits<HouseholderSequence>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<HouseholderSequence>::ColsAtCompileTime,
+      MaxRowsAtCompileTime = internal::traits<HouseholderSequence>::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = internal::traits<HouseholderSequence>::MaxColsAtCompileTime
+    };
+    typedef typename internal::traits<HouseholderSequence>::Scalar Scalar;
+
+    typedef HouseholderSequence<
+      typename internal::conditional<NumTraits<Scalar>::IsComplex,
+        typename internal::remove_all<typename VectorsType::ConjugateReturnType>::type,
+        VectorsType>::type,
+      typename internal::conditional<NumTraits<Scalar>::IsComplex,
+        typename internal::remove_all<typename CoeffsType::ConjugateReturnType>::type,
+        CoeffsType>::type,
+      Side
+    > ConjugateReturnType;
+
+    /** \brief Constructor.
+      * \param[in]  v      %Matrix containing the essential parts of the Householder vectors
+      * \param[in]  h      Vector containing the Householder coefficients
+      *
+      * Constructs the Householder sequence with coefficients given by \p h and vectors given by \p v. The
+      * i-th Householder coefficient \f$ h_i \f$ is given by \p h(i) and the essential part of the i-th
+      * Householder vector \f$ v_i \f$ is given by \p v(k,i) with \p k > \p i (the subdiagonal part of the
+      * i-th column). If \p v has fewer columns than rows, then the Householder sequence contains as many
+      * Householder reflections as there are columns.
+      *
+      * \note The %HouseholderSequence object stores \p v and \p h by reference.
+      *
+      * Example: \include HouseholderSequence_HouseholderSequence.cpp
+      * Output: \verbinclude HouseholderSequence_HouseholderSequence.out
+      *
+      * \sa setLength(), setShift()
+      */
+    HouseholderSequence(const VectorsType& v, const CoeffsType& h)
+      : m_vectors(v), m_coeffs(h), m_trans(false), m_length(v.diagonalSize()),
+        m_shift(0)
+    {
+    }
+
+    /** \brief Copy constructor. */
+    HouseholderSequence(const HouseholderSequence& other)
+      : m_vectors(other.m_vectors),
+        m_coeffs(other.m_coeffs),
+        m_trans(other.m_trans),
+        m_length(other.m_length),
+        m_shift(other.m_shift)
+    {
+    }
+
+    /** \brief Number of rows of transformation viewed as a matrix.
+      * \returns Number of rows 
+      * \details This equals the dimension of the space that the transformation acts on.
+      */
+    Index rows() const { return Side==OnTheLeft ? m_vectors.rows() : m_vectors.cols(); }
+
+    /** \brief Number of columns of transformation viewed as a matrix.
+      * \returns Number of columns
+      * \details This equals the dimension of the space that the transformation acts on.
+      */
+    Index cols() const { return rows(); }
+
+    /** \brief Essential part of a Householder vector.
+      * \param[in]  k  Index of Householder reflection
+      * \returns    Vector containing non-trivial entries of k-th Householder vector
+      *
+      * This function returns the essential part of the Householder vector \f$ v_i \f$. This is a vector of
+      * length \f$ n-i \f$ containing the last \f$ n-i \f$ entries of the vector
+      * \f[ 
+      * v_i = [\underbrace{0, \ldots, 0}_{i-1\mbox{ zeros}}, 1, \underbrace{*, \ldots,*}_{n-i\mbox{ arbitrary entries}} ]. 
+      * \f]
+      * The index \f$ i \f$ equals \p k + shift(), corresponding to the k-th column of the matrix \p v
+      * passed to the constructor.
+      *
+      * \sa setShift(), shift()
+      */
+    const EssentialVectorType essentialVector(Index k) const
+    {
+      eigen_assert(k >= 0 && k < m_length);
+      return internal::hseq_side_dependent_impl<VectorsType,CoeffsType,Side>::essentialVector(*this, k);
+    }
+
+    /** \brief %Transpose of the Householder sequence. */
+    HouseholderSequence transpose() const
+    {
+      return HouseholderSequence(*this).setTrans(!m_trans);
+    }
+
+    /** \brief Complex conjugate of the Householder sequence. */
+    ConjugateReturnType conjugate() const
+    {
+      return ConjugateReturnType(m_vectors.conjugate(), m_coeffs.conjugate())
+             .setTrans(m_trans)
+             .setLength(m_length)
+             .setShift(m_shift);
+    }
+
+    /** \brief Adjoint (conjugate transpose) of the Householder sequence. */
+    ConjugateReturnType adjoint() const
+    {
+      return conjugate().setTrans(!m_trans);
+    }
+
+    /** \brief Inverse of the Householder sequence (equals the adjoint). */
+    ConjugateReturnType inverse() const { return adjoint(); }
+
+    /** \internal */
+    template<typename DestType> inline void evalTo(DestType& dst) const
+    {
+      Matrix<Scalar, DestType::RowsAtCompileTime, 1,
+             AutoAlign|ColMajor, DestType::MaxRowsAtCompileTime, 1> workspace(rows());
+      evalTo(dst, workspace);
+    }
+
+    /** \internal */
+    template<typename Dest, typename Workspace>
+    void evalTo(Dest& dst, Workspace& workspace) const
+    {
+      workspace.resize(rows());
+      Index vecs = m_length;
+      if(internal::is_same_dense(dst,m_vectors))
+      {
+        // in-place
+        dst.diagonal().setOnes();
+        dst.template triangularView<StrictlyUpper>().setZero();
+        for(Index k = vecs-1; k >= 0; --k)
+        {
+          Index cornerSize = rows() - k - m_shift;
+          if(m_trans)
+            dst.bottomRightCorner(cornerSize, cornerSize)
+               .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), workspace.data());
+          else
+            dst.bottomRightCorner(cornerSize, cornerSize)
+               .applyHouseholderOnTheLeft(essentialVector(k), m_coeffs.coeff(k), workspace.data());
+
+          // clear the off diagonal vector
+          dst.col(k).tail(rows()-k-1).setZero();
+        }
+        // clear the remaining columns if needed
+        for(Index k = 0; k<cols()-vecs ; ++k)
+          dst.col(k).tail(rows()-k-1).setZero();
+      }
+      else
+      {
+        dst.setIdentity(rows(), rows());
+        for(Index k = vecs-1; k >= 0; --k)
+        {
+          Index cornerSize = rows() - k - m_shift;
+          if(m_trans)
+            dst.bottomRightCorner(cornerSize, cornerSize)
+               .applyHouseholderOnTheRight(essentialVector(k), m_coeffs.coeff(k), &workspace.coeffRef(0));
+          else
+            dst.bottomRightCorner(cornerSize, cornerSize)
+               .applyHouseholderOnTheLeft(essentialVector(k), m_coeffs.coeff(k), &workspace.coeffRef(0));
+        }
+      }
+    }
+
+    /** \internal */
+    template<typename Dest> inline void applyThisOnTheRight(Dest& dst) const
+    {
+      Matrix<Scalar,1,Dest::RowsAtCompileTime,RowMajor,1,Dest::MaxRowsAtCompileTime> workspace(dst.rows());
+      applyThisOnTheRight(dst, workspace);
+    }
+
+    /** \internal */
+    template<typename Dest, typename Workspace>
+    inline void applyThisOnTheRight(Dest& dst, Workspace& workspace) const
+    {
+      workspace.resize(dst.rows());
+      for(Index k = 0; k < m_length; ++k)
+      {
+        Index actual_k = m_trans ? m_length-k-1 : k;
+        dst.rightCols(rows()-m_shift-actual_k)
+           .applyHouseholderOnTheRight(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data());
+      }
+    }
+
+    /** \internal */
+    template<typename Dest> inline void applyThisOnTheLeft(Dest& dst) const
+    {
+      Matrix<Scalar,1,Dest::ColsAtCompileTime,RowMajor,1,Dest::MaxColsAtCompileTime> workspace;
+      applyThisOnTheLeft(dst, workspace);
+    }
+
+    /** \internal */
+    template<typename Dest, typename Workspace>
+    inline void applyThisOnTheLeft(Dest& dst, Workspace& workspace) const
+    {
+      const Index BlockSize = 48;
+      // if the entries are large enough, then apply the reflectors by block
+      if(m_length>=BlockSize && dst.cols()>1)
+      {
+        for(Index i = 0; i < m_length; i+=BlockSize)
+        {
+          Index end = m_trans ? (std::min)(m_length,i+BlockSize) : m_length-i;
+          Index k = m_trans ? i : (std::max)(Index(0),end-BlockSize);
+          Index bs = end-k;
+          Index start = k + m_shift;
+          
+          typedef Block<typename internal::remove_all<VectorsType>::type,Dynamic,Dynamic> SubVectorsType;
+          SubVectorsType sub_vecs1(m_vectors.const_cast_derived(), Side==OnTheRight ? k : start,
+                                                                   Side==OnTheRight ? start : k,
+                                                                   Side==OnTheRight ? bs : m_vectors.rows()-start,
+                                                                   Side==OnTheRight ? m_vectors.cols()-start : bs);
+          typename internal::conditional<Side==OnTheRight, Transpose<SubVectorsType>, SubVectorsType&>::type sub_vecs(sub_vecs1);
+          Block<Dest,Dynamic,Dynamic> sub_dst(dst,dst.rows()-rows()+m_shift+k,0, rows()-m_shift-k,dst.cols());
+          apply_block_householder_on_the_left(sub_dst, sub_vecs, m_coeffs.segment(k, bs), !m_trans);
+        }
+      }
+      else
+      {
+        workspace.resize(dst.cols());
+        for(Index k = 0; k < m_length; ++k)
+        {
+          Index actual_k = m_trans ? k : m_length-k-1;
+          dst.bottomRows(rows()-m_shift-actual_k)
+            .applyHouseholderOnTheLeft(essentialVector(actual_k), m_coeffs.coeff(actual_k), workspace.data());
+        }
+      }
+    }
+
+    /** \brief Computes the product of a Householder sequence with a matrix.
+      * \param[in]  other  %Matrix being multiplied.
+      * \returns    Expression object representing the product.
+      *
+      * This function computes \f$ HM \f$ where \f$ H \f$ is the Householder sequence represented by \p *this
+      * and \f$ M \f$ is the matrix \p other.
+      */
+    template<typename OtherDerived>
+    typename internal::matrix_type_times_scalar_type<Scalar, OtherDerived>::Type operator*(const MatrixBase<OtherDerived>& other) const
+    {
+      typename internal::matrix_type_times_scalar_type<Scalar, OtherDerived>::Type
+        res(other.template cast<typename internal::matrix_type_times_scalar_type<Scalar,OtherDerived>::ResultScalar>());
+      applyThisOnTheLeft(res);
+      return res;
+    }
+
+    template<typename _VectorsType, typename _CoeffsType, int _Side> friend struct internal::hseq_side_dependent_impl;
+
+    /** \brief Sets the length of the Householder sequence.
+      * \param [in]  length  New value for the length.
+      *
+      * By default, the length \f$ n \f$ of the Householder sequence \f$ H = H_0 H_1 \ldots H_{n-1} \f$ is set
+      * to the number of columns of the matrix \p v passed to the constructor, or the number of rows if that
+      * is smaller. After this function is called, the length equals \p length.
+      *
+      * \sa length()
+      */
+    HouseholderSequence& setLength(Index length)
+    {
+      m_length = length;
+      return *this;
+    }
+
+    /** \brief Sets the shift of the Householder sequence.
+      * \param [in]  shift  New value for the shift.
+      *
+      * By default, a %HouseholderSequence object represents \f$ H = H_0 H_1 \ldots H_{n-1} \f$ and the i-th
+      * column of the matrix \p v passed to the constructor corresponds to the i-th Householder
+      * reflection. After this function is called, the object represents \f$ H = H_{\mathrm{shift}}
+      * H_{\mathrm{shift}+1} \ldots H_{n-1} \f$ and the i-th column of \p v corresponds to the (shift+i)-th
+      * Householder reflection.
+      *
+      * \sa shift()
+      */
+    HouseholderSequence& setShift(Index shift)
+    {
+      m_shift = shift;
+      return *this;
+    }
+
+    Index length() const { return m_length; }  /**< \brief Returns the length of the Householder sequence. */
+    Index shift() const { return m_shift; }    /**< \brief Returns the shift of the Householder sequence. */
+
+    /* Necessary for .adjoint() and .conjugate() */
+    template <typename VectorsType2, typename CoeffsType2, int Side2> friend class HouseholderSequence;
+
+  protected:
+
+    /** \brief Sets the transpose flag.
+      * \param [in]  trans  New value of the transpose flag.
+      *
+      * By default, the transpose flag is not set. If the transpose flag is set, then this object represents 
+      * \f$ H^T = H_{n-1}^T \ldots H_1^T H_0^T \f$ instead of \f$ H = H_0 H_1 \ldots H_{n-1} \f$.
+      *
+      * \sa trans()
+      */
+    HouseholderSequence& setTrans(bool trans)
+    {
+      m_trans = trans;
+      return *this;
+    }
+
+    bool trans() const { return m_trans; }     /**< \brief Returns the transpose flag. */
+
+    typename VectorsType::Nested m_vectors;
+    typename CoeffsType::Nested m_coeffs;
+    bool m_trans;
+    Index m_length;
+    Index m_shift;
+};
+
+/** \brief Computes the product of a matrix with a Householder sequence.
+  * \param[in]  other  %Matrix being multiplied.
+  * \param[in]  h      %HouseholderSequence being multiplied.
+  * \returns    Expression object representing the product.
+  *
+  * This function computes \f$ MH \f$ where \f$ M \f$ is the matrix \p other and \f$ H \f$ is the
+  * Householder sequence represented by \p h.
+  */
+template<typename OtherDerived, typename VectorsType, typename CoeffsType, int Side>
+typename internal::matrix_type_times_scalar_type<typename VectorsType::Scalar,OtherDerived>::Type operator*(const MatrixBase<OtherDerived>& other, const HouseholderSequence<VectorsType,CoeffsType,Side>& h)
+{
+  typename internal::matrix_type_times_scalar_type<typename VectorsType::Scalar,OtherDerived>::Type
+    res(other.template cast<typename internal::matrix_type_times_scalar_type<typename VectorsType::Scalar,OtherDerived>::ResultScalar>());
+  h.applyThisOnTheRight(res);
+  return res;
+}
+
+/** \ingroup Householder_Module \householder_module
+  * \brief Convenience function for constructing a Householder sequence. 
+  * \returns A HouseholderSequence constructed from the specified arguments.
+  */
+template<typename VectorsType, typename CoeffsType>
+HouseholderSequence<VectorsType,CoeffsType> householderSequence(const VectorsType& v, const CoeffsType& h)
+{
+  return HouseholderSequence<VectorsType,CoeffsType,OnTheLeft>(v, h);
+}
+
+/** \ingroup Householder_Module \householder_module
+  * \brief Convenience function for constructing a Householder sequence. 
+  * \returns A HouseholderSequence constructed from the specified arguments.
+  * \details This function differs from householderSequence() in that the template argument \p OnTheSide of
+  * the constructed HouseholderSequence is set to OnTheRight, instead of the default OnTheLeft.
+  */
+template<typename VectorsType, typename CoeffsType>
+HouseholderSequence<VectorsType,CoeffsType,OnTheRight> rightHouseholderSequence(const VectorsType& v, const CoeffsType& h)
+{
+  return HouseholderSequence<VectorsType,CoeffsType,OnTheRight>(v, h);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_HOUSEHOLDER_SEQUENCE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
new file mode 100644
index 0000000..facdaf8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h
@@ -0,0 +1,226 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BASIC_PRECONDITIONERS_H
+#define EIGEN_BASIC_PRECONDITIONERS_H
+
+namespace Eigen { 
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A preconditioner based on the digonal entries
+  *
+  * This class allows to approximately solve for A.x = b problems assuming A is a diagonal matrix.
+  * In other words, this preconditioner neglects all off diagonal entries and, in Eigen's language, solves for:
+    \code
+    A.diagonal().asDiagonal() . x = b
+    \endcode
+  *
+  * \tparam _Scalar the type of the scalar.
+  *
+  * \implsparsesolverconcept
+  *
+  * This preconditioner is suitable for both selfadjoint and general problems.
+  * The diagonal entries are pre-inverted and stored into a dense vector.
+  *
+  * \note A variant that has yet to be implemented would attempt to preserve the norm of each column.
+  *
+  * \sa class LeastSquareDiagonalPreconditioner, class ConjugateGradient
+  */
+template <typename _Scalar>
+class DiagonalPreconditioner
+{
+    typedef _Scalar Scalar;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+  public:
+    typedef typename Vector::StorageIndex StorageIndex;
+    enum {
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
+
+    DiagonalPreconditioner() : m_isInitialized(false) {}
+
+    template<typename MatType>
+    explicit DiagonalPreconditioner(const MatType& mat) : m_invdiag(mat.cols())
+    {
+      compute(mat);
+    }
+
+    Index rows() const { return m_invdiag.size(); }
+    Index cols() const { return m_invdiag.size(); }
+    
+    template<typename MatType>
+    DiagonalPreconditioner& analyzePattern(const MatType& )
+    {
+      return *this;
+    }
+    
+    template<typename MatType>
+    DiagonalPreconditioner& factorize(const MatType& mat)
+    {
+      m_invdiag.resize(mat.cols());
+      for(int j=0; j<mat.outerSize(); ++j)
+      {
+        typename MatType::InnerIterator it(mat,j);
+        while(it && it.index()!=j) ++it;
+        if(it && it.index()==j && it.value()!=Scalar(0))
+          m_invdiag(j) = Scalar(1)/it.value();
+        else
+          m_invdiag(j) = Scalar(1);
+      }
+      m_isInitialized = true;
+      return *this;
+    }
+    
+    template<typename MatType>
+    DiagonalPreconditioner& compute(const MatType& mat)
+    {
+      return factorize(mat);
+    }
+
+    /** \internal */
+    template<typename Rhs, typename Dest>
+    void _solve_impl(const Rhs& b, Dest& x) const
+    {
+      x = m_invdiag.array() * b.array() ;
+    }
+
+    template<typename Rhs> inline const Solve<DiagonalPreconditioner, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "DiagonalPreconditioner is not initialized.");
+      eigen_assert(m_invdiag.size()==b.rows()
+                && "DiagonalPreconditioner::solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<DiagonalPreconditioner, Rhs>(*this, b.derived());
+    }
+    
+    ComputationInfo info() { return Success; }
+
+  protected:
+    Vector m_invdiag;
+    bool m_isInitialized;
+};
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief Jacobi preconditioner for LeastSquaresConjugateGradient
+  *
+  * This class allows to approximately solve for A' A x  = A' b problems assuming A' A is a diagonal matrix.
+  * In other words, this preconditioner neglects all off diagonal entries and, in Eigen's language, solves for:
+    \code
+    (A.adjoint() * A).diagonal().asDiagonal() * x = b
+    \endcode
+  *
+  * \tparam _Scalar the type of the scalar.
+  *
+  * \implsparsesolverconcept
+  *
+  * The diagonal entries are pre-inverted and stored into a dense vector.
+  * 
+  * \sa class LeastSquaresConjugateGradient, class DiagonalPreconditioner
+  */
+template <typename _Scalar>
+class LeastSquareDiagonalPreconditioner : public DiagonalPreconditioner<_Scalar>
+{
+    typedef _Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef DiagonalPreconditioner<_Scalar> Base;
+    using Base::m_invdiag;
+  public:
+
+    LeastSquareDiagonalPreconditioner() : Base() {}
+
+    template<typename MatType>
+    explicit LeastSquareDiagonalPreconditioner(const MatType& mat) : Base()
+    {
+      compute(mat);
+    }
+
+    template<typename MatType>
+    LeastSquareDiagonalPreconditioner& analyzePattern(const MatType& )
+    {
+      return *this;
+    }
+    
+    template<typename MatType>
+    LeastSquareDiagonalPreconditioner& factorize(const MatType& mat)
+    {
+      // Compute the inverse squared-norm of each column of mat
+      m_invdiag.resize(mat.cols());
+      if(MatType::IsRowMajor)
+      {
+        m_invdiag.setZero();
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          for(typename MatType::InnerIterator it(mat,j); it; ++it)
+            m_invdiag(it.index()) += numext::abs2(it.value());
+        }
+        for(Index j=0; j<mat.cols(); ++j)
+          if(numext::real(m_invdiag(j))>RealScalar(0))
+            m_invdiag(j) = RealScalar(1)/numext::real(m_invdiag(j));
+      }
+      else
+      {
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          RealScalar sum = mat.innerVector(j).squaredNorm();
+          if(sum>RealScalar(0))
+            m_invdiag(j) = RealScalar(1)/sum;
+          else
+            m_invdiag(j) = RealScalar(1);
+        }
+      }
+      Base::m_isInitialized = true;
+      return *this;
+    }
+    
+    template<typename MatType>
+    LeastSquareDiagonalPreconditioner& compute(const MatType& mat)
+    {
+      return factorize(mat);
+    }
+    
+    ComputationInfo info() { return Success; }
+
+  protected:
+};
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A naive preconditioner which approximates any matrix as the identity matrix
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa class DiagonalPreconditioner
+  */
+class IdentityPreconditioner
+{
+  public:
+
+    IdentityPreconditioner() {}
+
+    template<typename MatrixType>
+    explicit IdentityPreconditioner(const MatrixType& ) {}
+    
+    template<typename MatrixType>
+    IdentityPreconditioner& analyzePattern(const MatrixType& ) { return *this; }
+    
+    template<typename MatrixType>
+    IdentityPreconditioner& factorize(const MatrixType& ) { return *this; }
+
+    template<typename MatrixType>
+    IdentityPreconditioner& compute(const MatrixType& ) { return *this; }
+    
+    template<typename Rhs>
+    inline const Rhs& solve(const Rhs& b) const { return b; }
+    
+    ComputationInfo info() { return Success; }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_BASIC_PRECONDITIONERS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h
new file mode 100644
index 0000000..454f468
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/BiCGSTAB.h
@@ -0,0 +1,228 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BICGSTAB_H
+#define EIGEN_BICGSTAB_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal Low-level bi conjugate gradient stabilized algorithm
+  * \param mat The matrix A
+  * \param rhs The right hand side vector b
+  * \param x On input and initial solution, on output the computed solution.
+  * \param precond A preconditioner being able to efficiently solve for an
+  *                approximation of Ax=b (regardless of b)
+  * \param iters On input the max number of iteration, on output the number of performed iterations.
+  * \param tol_error On input the tolerance error, on output an estimation of the relative error.
+  * \return false in the case of numerical issue, for example a break down of BiCGSTAB. 
+  */
+template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+bool bicgstab(const MatrixType& mat, const Rhs& rhs, Dest& x,
+              const Preconditioner& precond, Index& iters,
+              typename Dest::RealScalar& tol_error)
+{
+  using std::sqrt;
+  using std::abs;
+  typedef typename Dest::RealScalar RealScalar;
+  typedef typename Dest::Scalar Scalar;
+  typedef Matrix<Scalar,Dynamic,1> VectorType;
+  RealScalar tol = tol_error;
+  Index maxIters = iters;
+
+  Index n = mat.cols();
+  VectorType r  = rhs - mat * x;
+  VectorType r0 = r;
+  
+  RealScalar r0_sqnorm = r0.squaredNorm();
+  RealScalar rhs_sqnorm = rhs.squaredNorm();
+  if(rhs_sqnorm == 0)
+  {
+    x.setZero();
+    return true;
+  }
+  Scalar rho    = 1;
+  Scalar alpha  = 1;
+  Scalar w      = 1;
+  
+  VectorType v = VectorType::Zero(n), p = VectorType::Zero(n);
+  VectorType y(n),  z(n);
+  VectorType kt(n), ks(n);
+
+  VectorType s(n), t(n);
+
+  RealScalar tol2 = tol*tol*rhs_sqnorm;
+  RealScalar eps2 = NumTraits<Scalar>::epsilon()*NumTraits<Scalar>::epsilon();
+  Index i = 0;
+  Index restarts = 0;
+
+  while ( r.squaredNorm() > tol2 && i<maxIters )
+  {
+    Scalar rho_old = rho;
+
+    rho = r0.dot(r);
+    if (abs(rho) < eps2*r0_sqnorm)
+    {
+      // The new residual vector became too orthogonal to the arbitrarily chosen direction r0
+      // Let's restart with a new r0:
+      r  = rhs - mat * x;
+      r0 = r;
+      rho = r0_sqnorm = r.squaredNorm();
+      if(restarts++ == 0)
+        i = 0;
+    }
+    Scalar beta = (rho/rho_old) * (alpha / w);
+    p = r + beta * (p - w * v);
+    
+    y = precond.solve(p);
+    
+    v.noalias() = mat * y;
+
+    alpha = rho / r0.dot(v);
+    s = r - alpha * v;
+
+    z = precond.solve(s);
+    t.noalias() = mat * z;
+
+    RealScalar tmp = t.squaredNorm();
+    if(tmp>RealScalar(0))
+      w = t.dot(s) / tmp;
+    else
+      w = Scalar(0);
+    x += alpha * y + w * z;
+    r = s - w * t;
+    ++i;
+  }
+  tol_error = sqrt(r.squaredNorm()/rhs_sqnorm);
+  iters = i;
+  return true; 
+}
+
+}
+
+template< typename _MatrixType,
+          typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
+class BiCGSTAB;
+
+namespace internal {
+
+template< typename _MatrixType, typename _Preconditioner>
+struct traits<BiCGSTAB<_MatrixType,_Preconditioner> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Preconditioner Preconditioner;
+};
+
+}
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A bi conjugate gradient stabilized solver for sparse square problems
+  *
+  * This class allows to solve for A.x = b sparse linear problems using a bi conjugate gradient
+  * stabilized algorithm. The vectors x and b can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+  * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+  *
+  * \implsparsesolverconcept
+  *
+  * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+  * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+  * and NumTraits<Scalar>::epsilon() for the tolerance.
+  * 
+  * The tolerance corresponds to the relative residual error: |Ax-b|/|b|
+  * 
+  * \b Performance: when using sparse matrices, best performance is achied for a row-major sparse matrix format.
+  * Moreover, in this case multi-threading can be exploited if the user code is compiled with OpenMP enabled.
+  * See \ref TopicMultiThreading for details.
+  * 
+  * This class can be used as the direct solver classes. Here is a typical usage example:
+  * \include BiCGSTAB_simple.cpp
+  * 
+  * By default the iterations start with x=0 as an initial guess of the solution.
+  * One can control the start using the solveWithGuess() method.
+  * 
+  * BiCGSTAB can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
+  * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+  */
+template< typename _MatrixType, typename _Preconditioner>
+class BiCGSTAB : public IterativeSolverBase<BiCGSTAB<_MatrixType,_Preconditioner> >
+{
+  typedef IterativeSolverBase<BiCGSTAB> Base;
+  using Base::matrix;
+  using Base::m_error;
+  using Base::m_iterations;
+  using Base::m_info;
+  using Base::m_isInitialized;
+public:
+  typedef _MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef _Preconditioner Preconditioner;
+
+public:
+
+  /** Default constructor. */
+  BiCGSTAB() : Base() {}
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    * 
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    * 
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit BiCGSTAB(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
+
+  ~BiCGSTAB() {}
+
+  /** \internal */
+  template<typename Rhs,typename Dest>
+  void _solve_with_guess_impl(const Rhs& b, Dest& x) const
+  {    
+    bool failed = false;
+    for(Index j=0; j<b.cols(); ++j)
+    {
+      m_iterations = Base::maxIterations();
+      m_error = Base::m_tolerance;
+      
+      typename Dest::ColXpr xj(x,j);
+      if(!internal::bicgstab(matrix(), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error))
+        failed = true;
+    }
+    m_info = failed ? NumericalIssue
+           : m_error <= Base::m_tolerance ? Success
+           : NoConvergence;
+    m_isInitialized = true;
+  }
+
+  /** \internal */
+  using Base::_solve_impl;
+  template<typename Rhs,typename Dest>
+  void _solve_impl(const MatrixBase<Rhs>& b, Dest& x) const
+  {
+    x.resize(this->rows(),b.cols());
+    x.setZero();
+    _solve_with_guess_impl(b,x);
+  }
+
+protected:
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_BICGSTAB_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h
new file mode 100644
index 0000000..395daa8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/ConjugateGradient.h
@@ -0,0 +1,245 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONJUGATE_GRADIENT_H
+#define EIGEN_CONJUGATE_GRADIENT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal Low-level conjugate gradient algorithm
+  * \param mat The matrix A
+  * \param rhs The right hand side vector b
+  * \param x On input and initial solution, on output the computed solution.
+  * \param precond A preconditioner being able to efficiently solve for an
+  *                approximation of Ax=b (regardless of b)
+  * \param iters On input the max number of iteration, on output the number of performed iterations.
+  * \param tol_error On input the tolerance error, on output an estimation of the relative error.
+  */
+template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+EIGEN_DONT_INLINE
+void conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x,
+                        const Preconditioner& precond, Index& iters,
+                        typename Dest::RealScalar& tol_error)
+{
+  using std::sqrt;
+  using std::abs;
+  typedef typename Dest::RealScalar RealScalar;
+  typedef typename Dest::Scalar Scalar;
+  typedef Matrix<Scalar,Dynamic,1> VectorType;
+  
+  RealScalar tol = tol_error;
+  Index maxIters = iters;
+  
+  Index n = mat.cols();
+
+  VectorType residual = rhs - mat * x; //initial residual
+
+  RealScalar rhsNorm2 = rhs.squaredNorm();
+  if(rhsNorm2 == 0) 
+  {
+    x.setZero();
+    iters = 0;
+    tol_error = 0;
+    return;
+  }
+  RealScalar threshold = tol*tol*rhsNorm2;
+  RealScalar residualNorm2 = residual.squaredNorm();
+  if (residualNorm2 < threshold)
+  {
+    iters = 0;
+    tol_error = sqrt(residualNorm2 / rhsNorm2);
+    return;
+  }
+  
+  VectorType p(n);
+  p = precond.solve(residual);      // initial search direction
+
+  VectorType z(n), tmp(n);
+  RealScalar absNew = numext::real(residual.dot(p));  // the square of the absolute value of r scaled by invM
+  Index i = 0;
+  while(i < maxIters)
+  {
+    tmp.noalias() = mat * p;                    // the bottleneck of the algorithm
+
+    Scalar alpha = absNew / p.dot(tmp);         // the amount we travel on dir
+    x += alpha * p;                             // update solution
+    residual -= alpha * tmp;                    // update residual
+    
+    residualNorm2 = residual.squaredNorm();
+    if(residualNorm2 < threshold)
+      break;
+    
+    z = precond.solve(residual);                // approximately solve for "A z = residual"
+
+    RealScalar absOld = absNew;
+    absNew = numext::real(residual.dot(z));     // update the absolute value of r
+    RealScalar beta = absNew / absOld;          // calculate the Gram-Schmidt value used to create the new search direction
+    p = z + beta * p;                           // update search direction
+    i++;
+  }
+  tol_error = sqrt(residualNorm2 / rhsNorm2);
+  iters = i;
+}
+
+}
+
+template< typename _MatrixType, int _UpLo=Lower,
+          typename _Preconditioner = DiagonalPreconditioner<typename _MatrixType::Scalar> >
+class ConjugateGradient;
+
+namespace internal {
+
+template< typename _MatrixType, int _UpLo, typename _Preconditioner>
+struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Preconditioner Preconditioner;
+};
+
+}
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A conjugate gradient solver for sparse (or dense) self-adjoint problems
+  *
+  * This class allows to solve for A.x = b linear problems using an iterative conjugate gradient algorithm.
+  * The matrix A must be selfadjoint. The matrix A and the vectors x and b can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix.
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower,
+  *               \c Upper, or \c Lower|Upper in which the full matrix entries will be considered.
+  *               Default is \c Lower, best performance is \c Lower|Upper.
+  * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
+  *
+  * \implsparsesolverconcept
+  *
+  * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+  * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+  * and NumTraits<Scalar>::epsilon() for the tolerance.
+  * 
+  * The tolerance corresponds to the relative residual error: |Ax-b|/|b|
+  * 
+  * \b Performance: Even though the default value of \c _UpLo is \c Lower, significantly higher performance is
+  * achieved when using a complete matrix and \b Lower|Upper as the \a _UpLo template parameter. Moreover, in this
+  * case multi-threading can be exploited if the user code is compiled with OpenMP enabled.
+  * See \ref TopicMultiThreading for details.
+  * 
+  * This class can be used as the direct solver classes. Here is a typical usage example:
+    \code
+    int n = 10000;
+    VectorXd x(n), b(n);
+    SparseMatrix<double> A(n,n);
+    // fill A and b
+    ConjugateGradient<SparseMatrix<double>, Lower|Upper> cg;
+    cg.compute(A);
+    x = cg.solve(b);
+    std::cout << "#iterations:     " << cg.iterations() << std::endl;
+    std::cout << "estimated error: " << cg.error()      << std::endl;
+    // update b, and solve again
+    x = cg.solve(b);
+    \endcode
+  * 
+  * By default the iterations start with x=0 as an initial guess of the solution.
+  * One can control the start using the solveWithGuess() method.
+  * 
+  * ConjugateGradient can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
+  * \sa class LeastSquaresConjugateGradient, class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+  */
+template< typename _MatrixType, int _UpLo, typename _Preconditioner>
+class ConjugateGradient : public IterativeSolverBase<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
+{
+  typedef IterativeSolverBase<ConjugateGradient> Base;
+  using Base::matrix;
+  using Base::m_error;
+  using Base::m_iterations;
+  using Base::m_info;
+  using Base::m_isInitialized;
+public:
+  typedef _MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef _Preconditioner Preconditioner;
+
+  enum {
+    UpLo = _UpLo
+  };
+
+public:
+
+  /** Default constructor. */
+  ConjugateGradient() : Base() {}
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    * 
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    * 
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit ConjugateGradient(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
+
+  ~ConjugateGradient() {}
+
+  /** \internal */
+  template<typename Rhs,typename Dest>
+  void _solve_with_guess_impl(const Rhs& b, Dest& x) const
+  {
+    typedef typename Base::MatrixWrapper MatrixWrapper;
+    typedef typename Base::ActualMatrixType ActualMatrixType;
+    enum {
+      TransposeInput  =   (!MatrixWrapper::MatrixFree)
+                      &&  (UpLo==(Lower|Upper))
+                      &&  (!MatrixType::IsRowMajor)
+                      &&  (!NumTraits<Scalar>::IsComplex)
+    };
+    typedef typename internal::conditional<TransposeInput,Transpose<const ActualMatrixType>, ActualMatrixType const&>::type RowMajorWrapper;
+    EIGEN_STATIC_ASSERT(EIGEN_IMPLIES(MatrixWrapper::MatrixFree,UpLo==(Lower|Upper)),MATRIX_FREE_CONJUGATE_GRADIENT_IS_COMPATIBLE_WITH_UPPER_UNION_LOWER_MODE_ONLY);
+    typedef typename internal::conditional<UpLo==(Lower|Upper),
+                                           RowMajorWrapper,
+                                           typename MatrixWrapper::template ConstSelfAdjointViewReturnType<UpLo>::Type
+                                          >::type SelfAdjointWrapper;
+    m_iterations = Base::maxIterations();
+    m_error = Base::m_tolerance;
+
+    for(Index j=0; j<b.cols(); ++j)
+    {
+      m_iterations = Base::maxIterations();
+      m_error = Base::m_tolerance;
+
+      typename Dest::ColXpr xj(x,j);
+      RowMajorWrapper row_mat(matrix());
+      internal::conjugate_gradient(SelfAdjointWrapper(row_mat), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error);
+    }
+
+    m_isInitialized = true;
+    m_info = m_error <= Base::m_tolerance ? Success : NoConvergence;
+  }
+  
+  /** \internal */
+  using Base::_solve_impl;
+  template<typename Rhs,typename Dest>
+  void _solve_impl(const MatrixBase<Rhs>& b, Dest& x) const
+  {
+    x.setZero();
+    _solve_with_guess_impl(b.derived(),x);
+  }
+
+protected:
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_CONJUGATE_GRADIENT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteCholesky.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteCholesky.h
new file mode 100644
index 0000000..e45c272
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteCholesky.h
@@ -0,0 +1,400 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INCOMPLETE_CHOlESKY_H
+#define EIGEN_INCOMPLETE_CHOlESKY_H
+
+#include <vector>
+#include <list>
+
+namespace Eigen {  
+/** 
+  * \brief Modified Incomplete Cholesky with dual threshold
+  *
+  * References : C-J. Lin and J. J. Moré, Incomplete Cholesky Factorizations with
+  *              Limited memory, SIAM J. Sci. Comput.  21(1), pp. 24-45, 1999
+  *
+  * \tparam Scalar the scalar type of the input matrices
+  * \tparam _UpLo The triangular part that will be used for the computations. It can be Lower
+    *               or Upper. Default is Lower.
+  * \tparam _OrderingType The ordering method to use, either AMDOrdering<> or NaturalOrdering<>. Default is AMDOrdering<int>,
+  *                       unless EIGEN_MPL2_ONLY is defined, in which case the default is NaturalOrdering<int>.
+  *
+  * \implsparsesolverconcept
+  *
+  * It performs the following incomplete factorization: \f$ S P A P' S \approx L L' \f$
+  * where L is a lower triangular factor, S is a diagonal scaling matrix, and P is a
+  * fill-in reducing permutation as computed by the ordering method.
+  *
+  * \b Shifting \b strategy: Let \f$ B = S P A P' S \f$  be the scaled matrix on which the factorization is carried out,
+  * and \f$ \beta \f$ be the minimum value of the diagonal. If \f$ \beta > 0 \f$ then, the factorization is directly performed
+  * on the matrix B. Otherwise, the factorization is performed on the shifted matrix \f$ B + (\sigma+|\beta| I \f$ where
+  * \f$ \sigma \f$ is the initial shift value as returned and set by setInitialShift() method. The default value is \f$ \sigma = 10^{-3} \f$.
+  * If the factorization fails, then the shift in doubled until it succeed or a maximum of ten attempts. If it still fails, as returned by
+  * the info() method, then you can either increase the initial shift, or better use another preconditioning technique.
+  *
+  */
+template <typename Scalar, int _UpLo = Lower, typename _OrderingType =
+#ifndef EIGEN_MPL2_ONLY
+AMDOrdering<int>
+#else
+NaturalOrdering<int>
+#endif
+>
+class IncompleteCholesky : public SparseSolverBase<IncompleteCholesky<Scalar,_UpLo,_OrderingType> >
+{
+  protected:
+    typedef SparseSolverBase<IncompleteCholesky<Scalar,_UpLo,_OrderingType> > Base;
+    using Base::m_isInitialized;
+  public:
+    typedef typename NumTraits<Scalar>::Real RealScalar; 
+    typedef _OrderingType OrderingType;
+    typedef typename OrderingType::PermutationType PermutationType;
+    typedef typename PermutationType::StorageIndex StorageIndex; 
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> FactorType;
+    typedef Matrix<Scalar,Dynamic,1> VectorSx;
+    typedef Matrix<RealScalar,Dynamic,1> VectorRx;
+    typedef Matrix<StorageIndex,Dynamic, 1> VectorIx;
+    typedef std::vector<std::list<StorageIndex> > VectorList; 
+    enum { UpLo = _UpLo };
+    enum {
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
+  public:
+
+    /** Default constructor leaving the object in a partly non-initialized stage.
+      *
+      * You must call compute() or the pair analyzePattern()/factorize() to make it valid.
+      *
+      * \sa IncompleteCholesky(const MatrixType&)
+      */
+    IncompleteCholesky() : m_initialShift(1e-3),m_factorizationIsOk(false) {}
+    
+    /** Constructor computing the incomplete factorization for the given matrix \a matrix.
+      */
+    template<typename MatrixType>
+    IncompleteCholesky(const MatrixType& matrix) : m_initialShift(1e-3),m_factorizationIsOk(false)
+    {
+      compute(matrix);
+    }
+    
+    /** \returns number of rows of the factored matrix */
+    Index rows() const { return m_L.rows(); }
+    
+    /** \returns number of columns of the factored matrix */
+    Index cols() const { return m_L.cols(); }
+    
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * It triggers an assertion if \c *this has not been initialized through the respective constructor,
+      * or a call to compute() or analyzePattern().
+      *
+      * \returns \c Success if computation was successful,
+      *          \c NumericalIssue if the matrix appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "IncompleteCholesky is not initialized.");
+      return m_info;
+    }
+    
+    /** \brief Set the initial shift parameter \f$ \sigma \f$.
+      */
+    void setInitialShift(RealScalar shift) { m_initialShift = shift; }
+    
+    /** \brief Computes the fill reducing permutation vector using the sparsity pattern of \a mat
+      */
+    template<typename MatrixType>
+    void analyzePattern(const MatrixType& mat)
+    {
+      OrderingType ord; 
+      PermutationType pinv;
+      ord(mat.template selfadjointView<UpLo>(), pinv); 
+      if(pinv.size()>0) m_perm = pinv.inverse();
+      else              m_perm.resize(0);
+      m_L.resize(mat.rows(), mat.cols());
+      m_analysisIsOk = true;
+      m_isInitialized = true;
+      m_info = Success;
+    }
+    
+    /** \brief Performs the numerical factorization of the input matrix \a mat
+      *
+      * The method analyzePattern() or compute() must have been called beforehand
+      * with a matrix having the same pattern.
+      *
+      * \sa compute(), analyzePattern()
+      */
+    template<typename MatrixType>
+    void factorize(const MatrixType& mat);
+    
+    /** Computes or re-computes the incomplete Cholesky factorization of the input matrix \a mat
+      *
+      * It is a shortcut for a sequential call to the analyzePattern() and factorize() methods.
+      *
+      * \sa analyzePattern(), factorize()
+      */
+    template<typename MatrixType>
+    void compute(const MatrixType& mat)
+    {
+      analyzePattern(mat);
+      factorize(mat);
+    }
+    
+    // internal
+    template<typename Rhs, typename Dest>
+    void _solve_impl(const Rhs& b, Dest& x) const
+    {
+      eigen_assert(m_factorizationIsOk && "factorize() should be called first");
+      if (m_perm.rows() == b.rows())  x = m_perm * b;
+      else                            x = b;
+      x = m_scale.asDiagonal() * x;
+      x = m_L.template triangularView<Lower>().solve(x);
+      x = m_L.adjoint().template triangularView<Upper>().solve(x);
+      x = m_scale.asDiagonal() * x;
+      if (m_perm.rows() == b.rows())
+        x = m_perm.inverse() * x;
+    }
+
+    /** \returns the sparse lower triangular factor L */
+    const FactorType& matrixL() const { eigen_assert("m_factorizationIsOk"); return m_L; }
+
+    /** \returns a vector representing the scaling factor S */
+    const VectorRx& scalingS() const { eigen_assert("m_factorizationIsOk"); return m_scale; }
+
+    /** \returns the fill-in reducing permutation P (can be empty for a natural ordering) */
+    const PermutationType& permutationP() const { eigen_assert("m_analysisIsOk"); return m_perm; }
+
+  protected:
+    FactorType m_L;              // The lower part stored in CSC
+    VectorRx m_scale;            // The vector for scaling the matrix 
+    RealScalar m_initialShift;   // The initial shift parameter
+    bool m_analysisIsOk; 
+    bool m_factorizationIsOk; 
+    ComputationInfo m_info;
+    PermutationType m_perm; 
+
+  private:
+    inline void updateList(Ref<const VectorIx> colPtr, Ref<VectorIx> rowIdx, Ref<VectorSx> vals, const Index& col, const Index& jk, VectorIx& firstElt, VectorList& listCol); 
+}; 
+
+// Based on the following paper:
+//   C-J. Lin and J. J. Moré, Incomplete Cholesky Factorizations with
+//   Limited memory, SIAM J. Sci. Comput.  21(1), pp. 24-45, 1999
+//   http://ftp.mcs.anl.gov/pub/tech_reports/reports/P682.pdf
+template<typename Scalar, int _UpLo, typename OrderingType>
+template<typename _MatrixType>
+void IncompleteCholesky<Scalar,_UpLo, OrderingType>::factorize(const _MatrixType& mat)
+{
+  using std::sqrt;
+  eigen_assert(m_analysisIsOk && "analyzePattern() should be called first"); 
+    
+  // Dropping strategy : Keep only the p largest elements per column, where p is the number of elements in the column of the original matrix. Other strategies will be added
+  
+  // Apply the fill-reducing permutation computed in analyzePattern()
+  if (m_perm.rows() == mat.rows() ) // To detect the null permutation
+  {
+    // The temporary is needed to make sure that the diagonal entry is properly sorted
+    FactorType tmp(mat.rows(), mat.cols());
+    tmp = mat.template selfadjointView<_UpLo>().twistedBy(m_perm);
+    m_L.template selfadjointView<Lower>() = tmp.template selfadjointView<Lower>();
+  }
+  else
+  {
+    m_L.template selfadjointView<Lower>() = mat.template selfadjointView<_UpLo>();
+  }
+  
+  Index n = m_L.cols(); 
+  Index nnz = m_L.nonZeros();
+  Map<VectorSx> vals(m_L.valuePtr(), nnz);         //values
+  Map<VectorIx> rowIdx(m_L.innerIndexPtr(), nnz);  //Row indices
+  Map<VectorIx> colPtr( m_L.outerIndexPtr(), n+1); // Pointer to the beginning of each row
+  VectorIx firstElt(n-1); // for each j, points to the next entry in vals that will be used in the factorization
+  VectorList listCol(n);  // listCol(j) is a linked list of columns to update column j
+  VectorSx col_vals(n);   // Store a  nonzero values in each column
+  VectorIx col_irow(n);   // Row indices of nonzero elements in each column
+  VectorIx col_pattern(n);
+  col_pattern.fill(-1);
+  StorageIndex col_nnz;
+  
+  
+  // Computes the scaling factors 
+  m_scale.resize(n);
+  m_scale.setZero();
+  for (Index j = 0; j < n; j++)
+    for (Index k = colPtr[j]; k < colPtr[j+1]; k++)
+    {
+      m_scale(j) += numext::abs2(vals(k));
+      if(rowIdx[k]!=j)
+        m_scale(rowIdx[k]) += numext::abs2(vals(k));
+    }
+  
+  m_scale = m_scale.cwiseSqrt().cwiseSqrt();
+
+  for (Index j = 0; j < n; ++j)
+    if(m_scale(j)>(std::numeric_limits<RealScalar>::min)())
+      m_scale(j) = RealScalar(1)/m_scale(j);
+    else
+      m_scale(j) = 1;
+
+  // TODO disable scaling if not needed, i.e., if it is roughly uniform? (this will make solve() faster)
+  
+  // Scale and compute the shift for the matrix 
+  RealScalar mindiag = NumTraits<RealScalar>::highest();
+  for (Index j = 0; j < n; j++)
+  {
+    for (Index k = colPtr[j]; k < colPtr[j+1]; k++)
+      vals[k] *= (m_scale(j)*m_scale(rowIdx[k]));
+    eigen_internal_assert(rowIdx[colPtr[j]]==j && "IncompleteCholesky: only the lower triangular part must be stored");
+    mindiag = numext::mini(numext::real(vals[colPtr[j]]), mindiag);
+  }
+
+  FactorType L_save = m_L;
+  
+  RealScalar shift = 0;
+  if(mindiag <= RealScalar(0.))
+    shift = m_initialShift - mindiag;
+
+  m_info = NumericalIssue;
+
+  // Try to perform the incomplete factorization using the current shift
+  int iter = 0;
+  do
+  {
+    // Apply the shift to the diagonal elements of the matrix
+    for (Index j = 0; j < n; j++)
+      vals[colPtr[j]] += shift;
+
+    // jki version of the Cholesky factorization
+    Index j=0;
+    for (; j < n; ++j)
+    {
+      // Left-looking factorization of the j-th column
+      // First, load the j-th column into col_vals
+      Scalar diag = vals[colPtr[j]];  // It is assumed that only the lower part is stored
+      col_nnz = 0;
+      for (Index i = colPtr[j] + 1; i < colPtr[j+1]; i++)
+      {
+        StorageIndex l = rowIdx[i];
+        col_vals(col_nnz) = vals[i];
+        col_irow(col_nnz) = l;
+        col_pattern(l) = col_nnz;
+        col_nnz++;
+      }
+      {
+        typename std::list<StorageIndex>::iterator k;
+        // Browse all previous columns that will update column j
+        for(k = listCol[j].begin(); k != listCol[j].end(); k++)
+        {
+          Index jk = firstElt(*k); // First element to use in the column
+          eigen_internal_assert(rowIdx[jk]==j);
+          Scalar v_j_jk = numext::conj(vals[jk]);
+
+          jk += 1;
+          for (Index i = jk; i < colPtr[*k+1]; i++)
+          {
+            StorageIndex l = rowIdx[i];
+            if(col_pattern[l]<0)
+            {
+              col_vals(col_nnz) = vals[i] * v_j_jk;
+              col_irow[col_nnz] = l;
+              col_pattern(l) = col_nnz;
+              col_nnz++;
+            }
+            else
+              col_vals(col_pattern[l]) -= vals[i] * v_j_jk;
+          }
+          updateList(colPtr,rowIdx,vals, *k, jk, firstElt, listCol);
+        }
+      }
+
+      // Scale the current column
+      if(numext::real(diag) <= 0)
+      {
+        if(++iter>=10)
+          return;
+
+        // increase shift
+        shift = numext::maxi(m_initialShift,RealScalar(2)*shift);
+        // restore m_L, col_pattern, and listCol
+        vals = Map<const VectorSx>(L_save.valuePtr(), nnz);
+        rowIdx = Map<const VectorIx>(L_save.innerIndexPtr(), nnz);
+        colPtr = Map<const VectorIx>(L_save.outerIndexPtr(), n+1);
+        col_pattern.fill(-1);
+        for(Index i=0; i<n; ++i)
+          listCol[i].clear();
+
+        break;
+      }
+
+      RealScalar rdiag = sqrt(numext::real(diag));
+      vals[colPtr[j]] = rdiag;
+      for (Index k = 0; k<col_nnz; ++k)
+      {
+        Index i = col_irow[k];
+        //Scale
+        col_vals(k) /= rdiag;
+        //Update the remaining diagonals with col_vals
+        vals[colPtr[i]] -= numext::abs2(col_vals(k));
+      }
+      // Select the largest p elements
+      // p is the original number of elements in the column (without the diagonal)
+      Index p = colPtr[j+1] - colPtr[j] - 1 ;
+      Ref<VectorSx> cvals = col_vals.head(col_nnz);
+      Ref<VectorIx> cirow = col_irow.head(col_nnz);
+      internal::QuickSplit(cvals,cirow, p);
+      // Insert the largest p elements in the matrix
+      Index cpt = 0;
+      for (Index i = colPtr[j]+1; i < colPtr[j+1]; i++)
+      {
+        vals[i] = col_vals(cpt);
+        rowIdx[i] = col_irow(cpt);
+        // restore col_pattern:
+        col_pattern(col_irow(cpt)) = -1;
+        cpt++;
+      }
+      // Get the first smallest row index and put it after the diagonal element
+      Index jk = colPtr(j)+1;
+      updateList(colPtr,rowIdx,vals,j,jk,firstElt,listCol);
+    }
+
+    if(j==n)
+    {
+      m_factorizationIsOk = true;
+      m_info = Success;
+    }
+  } while(m_info!=Success);
+}
+
+template<typename Scalar, int _UpLo, typename OrderingType>
+inline void IncompleteCholesky<Scalar,_UpLo, OrderingType>::updateList(Ref<const VectorIx> colPtr, Ref<VectorIx> rowIdx, Ref<VectorSx> vals, const Index& col, const Index& jk, VectorIx& firstElt, VectorList& listCol)
+{
+  if (jk < colPtr(col+1) )
+  {
+    Index p = colPtr(col+1) - jk;
+    Index minpos; 
+    rowIdx.segment(jk,p).minCoeff(&minpos);
+    minpos += jk;
+    if (rowIdx(minpos) != rowIdx(jk))
+    {
+      //Swap
+      std::swap(rowIdx(jk),rowIdx(minpos));
+      std::swap(vals(jk),vals(minpos));
+    }
+    firstElt(col) = internal::convert_index<StorageIndex,Index>(jk);
+    listCol[rowIdx(jk)].push_back(internal::convert_index<StorageIndex,Index>(col));
+  }
+}
+
+} // end namespace Eigen 
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h
new file mode 100644
index 0000000..338e6f1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IncompleteLUT.h
@@ -0,0 +1,462 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INCOMPLETE_LUT_H
+#define EIGEN_INCOMPLETE_LUT_H
+
+
+namespace Eigen { 
+
+namespace internal {
+    
+/** \internal
+  * Compute a quick-sort split of a vector 
+  * On output, the vector row is permuted such that its elements satisfy
+  * abs(row(i)) >= abs(row(ncut)) if i<ncut
+  * abs(row(i)) <= abs(row(ncut)) if i>ncut 
+  * \param row The vector of values
+  * \param ind The array of index for the elements in @p row
+  * \param ncut  The number of largest elements to keep
+  **/ 
+template <typename VectorV, typename VectorI>
+Index QuickSplit(VectorV &row, VectorI &ind, Index ncut)
+{
+  typedef typename VectorV::RealScalar RealScalar;
+  using std::swap;
+  using std::abs;
+  Index mid;
+  Index n = row.size(); /* length of the vector */
+  Index first, last ;
+  
+  ncut--; /* to fit the zero-based indices */
+  first = 0; 
+  last = n-1; 
+  if (ncut < first || ncut > last ) return 0;
+  
+  do {
+    mid = first; 
+    RealScalar abskey = abs(row(mid)); 
+    for (Index j = first + 1; j <= last; j++) {
+      if ( abs(row(j)) > abskey) {
+        ++mid;
+        swap(row(mid), row(j));
+        swap(ind(mid), ind(j));
+      }
+    }
+    /* Interchange for the pivot element */
+    swap(row(mid), row(first));
+    swap(ind(mid), ind(first));
+    
+    if (mid > ncut) last = mid - 1;
+    else if (mid < ncut ) first = mid + 1; 
+  } while (mid != ncut );
+  
+  return 0; /* mid is equal to ncut */ 
+}
+
+}// end namespace internal
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \class IncompleteLUT
+  * \brief Incomplete LU factorization with dual-threshold strategy
+  *
+  * \implsparsesolverconcept
+  *
+  * During the numerical factorization, two dropping rules are used :
+  *  1) any element whose magnitude is less than some tolerance is dropped.
+  *    This tolerance is obtained by multiplying the input tolerance @p droptol 
+  *    by the average magnitude of all the original elements in the current row.
+  *  2) After the elimination of the row, only the @p fill largest elements in 
+  *    the L part and the @p fill largest elements in the U part are kept 
+  *    (in addition to the diagonal element ). Note that @p fill is computed from 
+  *    the input parameter @p fillfactor which is used the ratio to control the fill_in 
+  *    relatively to the initial number of nonzero elements.
+  * 
+  * The two extreme cases are when @p droptol=0 (to keep all the @p fill*2 largest elements)
+  * and when @p fill=n/2 with @p droptol being different to zero. 
+  * 
+  * References : Yousef Saad, ILUT: A dual threshold incomplete LU factorization, 
+  *              Numerical Linear Algebra with Applications, 1(4), pp 387-402, 1994.
+  * 
+  * NOTE : The following implementation is derived from the ILUT implementation
+  * in the SPARSKIT package, Copyright (C) 2005, the Regents of the University of Minnesota 
+  *  released under the terms of the GNU LGPL: 
+  *    http://www-users.cs.umn.edu/~saad/software/SPARSKIT/README
+  * However, Yousef Saad gave us permission to relicense his ILUT code to MPL2.
+  * See the Eigen mailing list archive, thread: ILUT, date: July 8, 2012:
+  *   http://listengine.tuxfamily.org/lists.tuxfamily.org/eigen/2012/07/msg00064.html
+  * alternatively, on GMANE:
+  *   http://comments.gmane.org/gmane.comp.lib.eigen/3302
+  */
+template <typename _Scalar, typename _StorageIndex = int>
+class IncompleteLUT : public SparseSolverBase<IncompleteLUT<_Scalar, _StorageIndex> >
+{
+  protected:
+    typedef SparseSolverBase<IncompleteLUT> Base;
+    using Base::m_isInitialized;
+  public:
+    typedef _Scalar Scalar;
+    typedef _StorageIndex StorageIndex;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+    typedef SparseMatrix<Scalar,RowMajor,StorageIndex> FactorType;
+
+    enum {
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
+
+  public:
+    
+    IncompleteLUT()
+      : m_droptol(NumTraits<Scalar>::dummy_precision()), m_fillfactor(10),
+        m_analysisIsOk(false), m_factorizationIsOk(false)
+    {}
+    
+    template<typename MatrixType>
+    explicit IncompleteLUT(const MatrixType& mat, const RealScalar& droptol=NumTraits<Scalar>::dummy_precision(), int fillfactor = 10)
+      : m_droptol(droptol),m_fillfactor(fillfactor),
+        m_analysisIsOk(false),m_factorizationIsOk(false)
+    {
+      eigen_assert(fillfactor != 0);
+      compute(mat); 
+    }
+    
+    Index rows() const { return m_lu.rows(); }
+    
+    Index cols() const { return m_lu.cols(); }
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "IncompleteLUT is not initialized.");
+      return m_info;
+    }
+    
+    template<typename MatrixType>
+    void analyzePattern(const MatrixType& amat);
+    
+    template<typename MatrixType>
+    void factorize(const MatrixType& amat);
+    
+    /**
+      * Compute an incomplete LU factorization with dual threshold on the matrix mat
+      * No pivoting is done in this version
+      * 
+      **/
+    template<typename MatrixType>
+    IncompleteLUT& compute(const MatrixType& amat)
+    {
+      analyzePattern(amat); 
+      factorize(amat);
+      return *this;
+    }
+
+    void setDroptol(const RealScalar& droptol); 
+    void setFillfactor(int fillfactor); 
+    
+    template<typename Rhs, typename Dest>
+    void _solve_impl(const Rhs& b, Dest& x) const
+    {
+      x = m_Pinv * b;
+      x = m_lu.template triangularView<UnitLower>().solve(x);
+      x = m_lu.template triangularView<Upper>().solve(x);
+      x = m_P * x; 
+    }
+
+protected:
+
+    /** keeps off-diagonal entries; drops diagonal entries */
+    struct keep_diag {
+      inline bool operator() (const Index& row, const Index& col, const Scalar&) const
+      {
+        return row!=col;
+      }
+    };
+
+protected:
+
+    FactorType m_lu;
+    RealScalar m_droptol;
+    int m_fillfactor;
+    bool m_analysisIsOk;
+    bool m_factorizationIsOk;
+    ComputationInfo m_info;
+    PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_P;     // Fill-reducing permutation
+    PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_Pinv;  // Inverse permutation
+};
+
+/**
+ * Set control parameter droptol
+ *  \param droptol   Drop any element whose magnitude is less than this tolerance 
+ **/ 
+template<typename Scalar, typename StorageIndex>
+void IncompleteLUT<Scalar,StorageIndex>::setDroptol(const RealScalar& droptol)
+{
+  this->m_droptol = droptol;   
+}
+
+/**
+ * Set control parameter fillfactor
+ * \param fillfactor  This is used to compute the  number @p fill_in of largest elements to keep on each row. 
+ **/ 
+template<typename Scalar, typename StorageIndex>
+void IncompleteLUT<Scalar,StorageIndex>::setFillfactor(int fillfactor)
+{
+  this->m_fillfactor = fillfactor;   
+}
+
+template <typename Scalar, typename StorageIndex>
+template<typename _MatrixType>
+void IncompleteLUT<Scalar,StorageIndex>::analyzePattern(const _MatrixType& amat)
+{
+  // Compute the Fill-reducing permutation
+  // Since ILUT does not perform any numerical pivoting,
+  // it is highly preferable to keep the diagonal through symmetric permutations.
+#ifndef EIGEN_MPL2_ONLY
+  // To this end, let's symmetrize the pattern and perform AMD on it.
+  SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
+  SparseMatrix<Scalar,ColMajor, StorageIndex> mat2 = amat.transpose();
+  // FIXME for a matrix with nearly symmetric pattern, mat2+mat1 is the appropriate choice.
+  //       on the other hand for a really non-symmetric pattern, mat2*mat1 should be prefered...
+  SparseMatrix<Scalar,ColMajor, StorageIndex> AtA = mat2 + mat1;
+  AMDOrdering<StorageIndex> ordering;
+  ordering(AtA,m_P);
+  m_Pinv  = m_P.inverse(); // cache the inverse permutation
+#else
+  // If AMD is not available, (MPL2-only), then let's use the slower COLAMD routine.
+  SparseMatrix<Scalar,ColMajor, StorageIndex> mat1 = amat;
+  COLAMDOrdering<StorageIndex> ordering;
+  ordering(mat1,m_Pinv);
+  m_P = m_Pinv.inverse();
+#endif
+
+  m_analysisIsOk = true;
+  m_factorizationIsOk = false;
+  m_isInitialized = true;
+}
+
+template <typename Scalar, typename StorageIndex>
+template<typename _MatrixType>
+void IncompleteLUT<Scalar,StorageIndex>::factorize(const _MatrixType& amat)
+{
+  using std::sqrt;
+  using std::swap;
+  using std::abs;
+  using internal::convert_index;
+
+  eigen_assert((amat.rows() == amat.cols()) && "The factorization should be done on a square matrix");
+  Index n = amat.cols();  // Size of the matrix
+  m_lu.resize(n,n);
+  // Declare Working vectors and variables
+  Vector u(n) ;     // real values of the row -- maximum size is n --
+  VectorI ju(n);   // column position of the values in u -- maximum size  is n
+  VectorI jr(n);   // Indicate the position of the nonzero elements in the vector u -- A zero location is indicated by -1
+
+  // Apply the fill-reducing permutation
+  eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+  SparseMatrix<Scalar,RowMajor, StorageIndex> mat;
+  mat = amat.twistedBy(m_Pinv);
+
+  // Initialization
+  jr.fill(-1);
+  ju.fill(0);
+  u.fill(0);
+
+  // number of largest elements to keep in each row:
+  Index fill_in = (amat.nonZeros()*m_fillfactor)/n + 1;
+  if (fill_in > n) fill_in = n;
+
+  // number of largest nonzero elements to keep in the L and the U part of the current row:
+  Index nnzL = fill_in/2;
+  Index nnzU = nnzL;
+  m_lu.reserve(n * (nnzL + nnzU + 1));
+
+  // global loop over the rows of the sparse matrix
+  for (Index ii = 0; ii < n; ii++)
+  {
+    // 1 - copy the lower and the upper part of the row i of mat in the working vector u
+
+    Index sizeu = 1; // number of nonzero elements in the upper part of the current row
+    Index sizel = 0; // number of nonzero elements in the lower part of the current row
+    ju(ii)    = convert_index<StorageIndex>(ii);
+    u(ii)     = 0;
+    jr(ii)    = convert_index<StorageIndex>(ii);
+    RealScalar rownorm = 0;
+
+    typename FactorType::InnerIterator j_it(mat, ii); // Iterate through the current row ii
+    for (; j_it; ++j_it)
+    {
+      Index k = j_it.index();
+      if (k < ii)
+      {
+        // copy the lower part
+        ju(sizel) = convert_index<StorageIndex>(k);
+        u(sizel) = j_it.value();
+        jr(k) = convert_index<StorageIndex>(sizel);
+        ++sizel;
+      }
+      else if (k == ii)
+      {
+        u(ii) = j_it.value();
+      }
+      else
+      {
+        // copy the upper part
+        Index jpos = ii + sizeu;
+        ju(jpos) = convert_index<StorageIndex>(k);
+        u(jpos) = j_it.value();
+        jr(k) = convert_index<StorageIndex>(jpos);
+        ++sizeu;
+      }
+      rownorm += numext::abs2(j_it.value());
+    }
+
+    // 2 - detect possible zero row
+    if(rownorm==0)
+    {
+      m_info = NumericalIssue;
+      return;
+    }
+    // Take the 2-norm of the current row as a relative tolerance
+    rownorm = sqrt(rownorm);
+
+    // 3 - eliminate the previous nonzero rows
+    Index jj = 0;
+    Index len = 0;
+    while (jj < sizel)
+    {
+      // In order to eliminate in the correct order,
+      // we must select first the smallest column index among  ju(jj:sizel)
+      Index k;
+      Index minrow = ju.segment(jj,sizel-jj).minCoeff(&k); // k is relative to the segment
+      k += jj;
+      if (minrow != ju(jj))
+      {
+        // swap the two locations
+        Index j = ju(jj);
+        swap(ju(jj), ju(k));
+        jr(minrow) = convert_index<StorageIndex>(jj);
+        jr(j) = convert_index<StorageIndex>(k);
+        swap(u(jj), u(k));
+      }
+      // Reset this location
+      jr(minrow) = -1;
+
+      // Start elimination
+      typename FactorType::InnerIterator ki_it(m_lu, minrow);
+      while (ki_it && ki_it.index() < minrow) ++ki_it;
+      eigen_internal_assert(ki_it && ki_it.col()==minrow);
+      Scalar fact = u(jj) / ki_it.value();
+
+      // drop too small elements
+      if(abs(fact) <= m_droptol)
+      {
+        jj++;
+        continue;
+      }
+
+      // linear combination of the current row ii and the row minrow
+      ++ki_it;
+      for (; ki_it; ++ki_it)
+      {
+        Scalar prod = fact * ki_it.value();
+        Index j     = ki_it.index();
+        Index jpos  = jr(j);
+        if (jpos == -1) // fill-in element
+        {
+          Index newpos;
+          if (j >= ii) // dealing with the upper part
+          {
+            newpos = ii + sizeu;
+            sizeu++;
+            eigen_internal_assert(sizeu<=n);
+          }
+          else // dealing with the lower part
+          {
+            newpos = sizel;
+            sizel++;
+            eigen_internal_assert(sizel<=ii);
+          }
+          ju(newpos) = convert_index<StorageIndex>(j);
+          u(newpos) = -prod;
+          jr(j) = convert_index<StorageIndex>(newpos);
+        }
+        else
+          u(jpos) -= prod;
+      }
+      // store the pivot element
+      u(len)  = fact;
+      ju(len) = convert_index<StorageIndex>(minrow);
+      ++len;
+
+      jj++;
+    } // end of the elimination on the row ii
+
+    // reset the upper part of the pointer jr to zero
+    for(Index k = 0; k <sizeu; k++) jr(ju(ii+k)) = -1;
+
+    // 4 - partially sort and insert the elements in the m_lu matrix
+
+    // sort the L-part of the row
+    sizel = len;
+    len = (std::min)(sizel, nnzL);
+    typename Vector::SegmentReturnType ul(u.segment(0, sizel));
+    typename VectorI::SegmentReturnType jul(ju.segment(0, sizel));
+    internal::QuickSplit(ul, jul, len);
+
+    // store the largest m_fill elements of the L part
+    m_lu.startVec(ii);
+    for(Index k = 0; k < len; k++)
+      m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
+
+    // store the diagonal element
+    // apply a shifting rule to avoid zero pivots (we are doing an incomplete factorization)
+    if (u(ii) == Scalar(0))
+      u(ii) = sqrt(m_droptol) * rownorm;
+    m_lu.insertBackByOuterInnerUnordered(ii, ii) = u(ii);
+
+    // sort the U-part of the row
+    // apply the dropping rule first
+    len = 0;
+    for(Index k = 1; k < sizeu; k++)
+    {
+      if(abs(u(ii+k)) > m_droptol * rownorm )
+      {
+        ++len;
+        u(ii + len)  = u(ii + k);
+        ju(ii + len) = ju(ii + k);
+      }
+    }
+    sizeu = len + 1; // +1 to take into account the diagonal element
+    len = (std::min)(sizeu, nnzU);
+    typename Vector::SegmentReturnType uu(u.segment(ii+1, sizeu-1));
+    typename VectorI::SegmentReturnType juu(ju.segment(ii+1, sizeu-1));
+    internal::QuickSplit(uu, juu, len);
+
+    // store the largest elements of the U part
+    for(Index k = ii + 1; k < ii + len; k++)
+      m_lu.insertBackByOuterInnerUnordered(ii,ju(k)) = u(k);
+  }
+  m_lu.finalize();
+  m_lu.makeCompressed();
+
+  m_factorizationIsOk = true;
+  m_info = Success;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_INCOMPLETE_LUT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h
new file mode 100644
index 0000000..7c2326e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h
@@ -0,0 +1,394 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ITERATIVE_SOLVER_BASE_H
+#define EIGEN_ITERATIVE_SOLVER_BASE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename MatrixType>
+struct is_ref_compatible_impl
+{
+private:
+  template <typename T0>
+  struct any_conversion
+  {
+    template <typename T> any_conversion(const volatile T&);
+    template <typename T> any_conversion(T&);
+  };
+  struct yes {int a[1];};
+  struct no  {int a[2];};
+
+  template<typename T>
+  static yes test(const Ref<const T>&, int);
+  template<typename T>
+  static no  test(any_conversion<T>, ...);
+
+public:
+  static MatrixType ms_from;
+  enum { value = sizeof(test<MatrixType>(ms_from, 0))==sizeof(yes) };
+};
+
+template<typename MatrixType>
+struct is_ref_compatible
+{
+  enum { value = is_ref_compatible_impl<typename remove_all<MatrixType>::type>::value };
+};
+
+template<typename MatrixType, bool MatrixFree = !internal::is_ref_compatible<MatrixType>::value>
+class generic_matrix_wrapper;
+
+// We have an explicit matrix at hand, compatible with Ref<>
+template<typename MatrixType>
+class generic_matrix_wrapper<MatrixType,false>
+{
+public:
+  typedef Ref<const MatrixType> ActualMatrixType;
+  template<int UpLo> struct ConstSelfAdjointViewReturnType {
+    typedef typename ActualMatrixType::template ConstSelfAdjointViewReturnType<UpLo>::Type Type;
+  };
+
+  enum {
+    MatrixFree = false
+  };
+
+  generic_matrix_wrapper()
+    : m_dummy(0,0), m_matrix(m_dummy)
+  {}
+
+  template<typename InputType>
+  generic_matrix_wrapper(const InputType &mat)
+    : m_matrix(mat)
+  {}
+
+  const ActualMatrixType& matrix() const
+  {
+    return m_matrix;
+  }
+
+  template<typename MatrixDerived>
+  void grab(const EigenBase<MatrixDerived> &mat)
+  {
+    m_matrix.~Ref<const MatrixType>();
+    ::new (&m_matrix) Ref<const MatrixType>(mat.derived());
+  }
+
+  void grab(const Ref<const MatrixType> &mat)
+  {
+    if(&(mat.derived()) != &m_matrix)
+    {
+      m_matrix.~Ref<const MatrixType>();
+      ::new (&m_matrix) Ref<const MatrixType>(mat);
+    }
+  }
+
+protected:
+  MatrixType m_dummy; // used to default initialize the Ref<> object
+  ActualMatrixType m_matrix;
+};
+
+// MatrixType is not compatible with Ref<> -> matrix-free wrapper
+template<typename MatrixType>
+class generic_matrix_wrapper<MatrixType,true>
+{
+public:
+  typedef MatrixType ActualMatrixType;
+  template<int UpLo> struct ConstSelfAdjointViewReturnType
+  {
+    typedef ActualMatrixType Type;
+  };
+
+  enum {
+    MatrixFree = true
+  };
+
+  generic_matrix_wrapper()
+    : mp_matrix(0)
+  {}
+
+  generic_matrix_wrapper(const MatrixType &mat)
+    : mp_matrix(&mat)
+  {}
+
+  const ActualMatrixType& matrix() const
+  {
+    return *mp_matrix;
+  }
+
+  void grab(const MatrixType &mat)
+  {
+    mp_matrix = &mat;
+  }
+
+protected:
+  const ActualMatrixType *mp_matrix;
+};
+
+}
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief Base class for linear iterative solvers
+  *
+  * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
+  */
+template< typename Derived>
+class IterativeSolverBase : public SparseSolverBase<Derived>
+{
+protected:
+  typedef SparseSolverBase<Derived> Base;
+  using Base::m_isInitialized;
+  
+public:
+  typedef typename internal::traits<Derived>::MatrixType MatrixType;
+  typedef typename internal::traits<Derived>::Preconditioner Preconditioner;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef typename MatrixType::RealScalar RealScalar;
+
+  enum {
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+  };
+
+public:
+
+  using Base::derived;
+
+  /** Default constructor. */
+  IterativeSolverBase()
+  {
+    init();
+  }
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    * 
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    * 
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit IterativeSolverBase(const EigenBase<MatrixDerived>& A)
+    : m_matrixWrapper(A.derived())
+  {
+    init();
+    compute(matrix());
+  }
+
+  ~IterativeSolverBase() {}
+  
+  /** Initializes the iterative solver for the sparsity pattern of the matrix \a A for further solving \c Ax=b problems.
+    *
+    * Currently, this function mostly calls analyzePattern on the preconditioner. In the future
+    * we might, for instance, implement column reordering for faster matrix vector products.
+    */
+  template<typename MatrixDerived>
+  Derived& analyzePattern(const EigenBase<MatrixDerived>& A)
+  {
+    grab(A.derived());
+    m_preconditioner.analyzePattern(matrix());
+    m_isInitialized = true;
+    m_analysisIsOk = true;
+    m_info = m_preconditioner.info();
+    return derived();
+  }
+  
+  /** Initializes the iterative solver with the numerical values of the matrix \a A for further solving \c Ax=b problems.
+    *
+    * Currently, this function mostly calls factorize on the preconditioner.
+    *
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  Derived& factorize(const EigenBase<MatrixDerived>& A)
+  {
+    eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); 
+    grab(A.derived());
+    m_preconditioner.factorize(matrix());
+    m_factorizationIsOk = true;
+    m_info = m_preconditioner.info();
+    return derived();
+  }
+
+  /** Initializes the iterative solver with the matrix \a A for further solving \c Ax=b problems.
+    *
+    * Currently, this function mostly initializes/computes the preconditioner. In the future
+    * we might, for instance, implement column reordering for faster matrix vector products.
+    *
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  Derived& compute(const EigenBase<MatrixDerived>& A)
+  {
+    grab(A.derived());
+    m_preconditioner.compute(matrix());
+    m_isInitialized = true;
+    m_analysisIsOk = true;
+    m_factorizationIsOk = true;
+    m_info = m_preconditioner.info();
+    return derived();
+  }
+
+  /** \internal */
+  Index rows() const { return matrix().rows(); }
+
+  /** \internal */
+  Index cols() const { return matrix().cols(); }
+
+  /** \returns the tolerance threshold used by the stopping criteria.
+    * \sa setTolerance()
+    */
+  RealScalar tolerance() const { return m_tolerance; }
+  
+  /** Sets the tolerance threshold used by the stopping criteria.
+    *
+    * This value is used as an upper bound to the relative residual error: |Ax-b|/|b|.
+    * The default value is the machine precision given by NumTraits<Scalar>::epsilon()
+    */
+  Derived& setTolerance(const RealScalar& tolerance)
+  {
+    m_tolerance = tolerance;
+    return derived();
+  }
+
+  /** \returns a read-write reference to the preconditioner for custom configuration. */
+  Preconditioner& preconditioner() { return m_preconditioner; }
+  
+  /** \returns a read-only reference to the preconditioner. */
+  const Preconditioner& preconditioner() const { return m_preconditioner; }
+
+  /** \returns the max number of iterations.
+    * It is either the value setted by setMaxIterations or, by default,
+    * twice the number of columns of the matrix.
+    */
+  Index maxIterations() const
+  {
+    return (m_maxIterations<0) ? 2*matrix().cols() : m_maxIterations;
+  }
+  
+  /** Sets the max number of iterations.
+    * Default is twice the number of columns of the matrix.
+    */
+  Derived& setMaxIterations(Index maxIters)
+  {
+    m_maxIterations = maxIters;
+    return derived();
+  }
+
+  /** \returns the number of iterations performed during the last solve */
+  Index iterations() const
+  {
+    eigen_assert(m_isInitialized && "ConjugateGradient is not initialized.");
+    return m_iterations;
+  }
+
+  /** \returns the tolerance error reached during the last solve.
+    * It is a close approximation of the true relative residual error |Ax-b|/|b|.
+    */
+  RealScalar error() const
+  {
+    eigen_assert(m_isInitialized && "ConjugateGradient is not initialized.");
+    return m_error;
+  }
+
+  /** \returns the solution x of \f$ A x = b \f$ using the current decomposition of A
+    * and \a x0 as an initial solution.
+    *
+    * \sa solve(), compute()
+    */
+  template<typename Rhs,typename Guess>
+  inline const SolveWithGuess<Derived, Rhs, Guess>
+  solveWithGuess(const MatrixBase<Rhs>& b, const Guess& x0) const
+  {
+    eigen_assert(m_isInitialized && "Solver is not initialized.");
+    eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+    return SolveWithGuess<Derived, Rhs, Guess>(derived(), b.derived(), x0);
+  }
+
+  /** \returns Success if the iterations converged, and NoConvergence otherwise. */
+  ComputationInfo info() const
+  {
+    eigen_assert(m_isInitialized && "IterativeSolverBase is not initialized.");
+    return m_info;
+  }
+  
+  /** \internal */
+  template<typename Rhs, typename DestDerived>
+  void _solve_impl(const Rhs& b, SparseMatrixBase<DestDerived> &aDest) const
+  {
+    eigen_assert(rows()==b.rows());
+    
+    Index rhsCols = b.cols();
+    Index size = b.rows();
+    DestDerived& dest(aDest.derived());
+    typedef typename DestDerived::Scalar DestScalar;
+    Eigen::Matrix<DestScalar,Dynamic,1> tb(size);
+    Eigen::Matrix<DestScalar,Dynamic,1> tx(cols());
+    // We do not directly fill dest because sparse expressions have to be free of aliasing issue.
+    // For non square least-square problems, b and dest might not have the same size whereas they might alias each-other.
+    typename DestDerived::PlainObject tmp(cols(),rhsCols);
+    for(Index k=0; k<rhsCols; ++k)
+    {
+      tb = b.col(k);
+      tx = derived().solve(tb);
+      tmp.col(k) = tx.sparseView(0);
+    }
+    dest.swap(tmp);
+  }
+
+protected:
+  void init()
+  {
+    m_isInitialized = false;
+    m_analysisIsOk = false;
+    m_factorizationIsOk = false;
+    m_maxIterations = -1;
+    m_tolerance = NumTraits<Scalar>::epsilon();
+  }
+
+  typedef internal::generic_matrix_wrapper<MatrixType> MatrixWrapper;
+  typedef typename MatrixWrapper::ActualMatrixType ActualMatrixType;
+
+  const ActualMatrixType& matrix() const
+  {
+    return m_matrixWrapper.matrix();
+  }
+  
+  template<typename InputType>
+  void grab(const InputType &A)
+  {
+    m_matrixWrapper.grab(A);
+  }
+  
+  MatrixWrapper m_matrixWrapper;
+  Preconditioner m_preconditioner;
+
+  Index m_maxIterations;
+  RealScalar m_tolerance;
+  
+  mutable RealScalar m_error;
+  mutable Index m_iterations;
+  mutable ComputationInfo m_info;
+  mutable bool m_analysisIsOk, m_factorizationIsOk;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_ITERATIVE_SOLVER_BASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/LeastSquareConjugateGradient.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/LeastSquareConjugateGradient.h
new file mode 100644
index 0000000..0aea0e0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/LeastSquareConjugateGradient.h
@@ -0,0 +1,216 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LEAST_SQUARE_CONJUGATE_GRADIENT_H
+#define EIGEN_LEAST_SQUARE_CONJUGATE_GRADIENT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal Low-level conjugate gradient algorithm for least-square problems
+  * \param mat The matrix A
+  * \param rhs The right hand side vector b
+  * \param x On input and initial solution, on output the computed solution.
+  * \param precond A preconditioner being able to efficiently solve for an
+  *                approximation of A'Ax=b (regardless of b)
+  * \param iters On input the max number of iteration, on output the number of performed iterations.
+  * \param tol_error On input the tolerance error, on output an estimation of the relative error.
+  */
+template<typename MatrixType, typename Rhs, typename Dest, typename Preconditioner>
+EIGEN_DONT_INLINE
+void least_square_conjugate_gradient(const MatrixType& mat, const Rhs& rhs, Dest& x,
+                                     const Preconditioner& precond, Index& iters,
+                                     typename Dest::RealScalar& tol_error)
+{
+  using std::sqrt;
+  using std::abs;
+  typedef typename Dest::RealScalar RealScalar;
+  typedef typename Dest::Scalar Scalar;
+  typedef Matrix<Scalar,Dynamic,1> VectorType;
+  
+  RealScalar tol = tol_error;
+  Index maxIters = iters;
+  
+  Index m = mat.rows(), n = mat.cols();
+
+  VectorType residual        = rhs - mat * x;
+  VectorType normal_residual = mat.adjoint() * residual;
+
+  RealScalar rhsNorm2 = (mat.adjoint()*rhs).squaredNorm();
+  if(rhsNorm2 == 0) 
+  {
+    x.setZero();
+    iters = 0;
+    tol_error = 0;
+    return;
+  }
+  RealScalar threshold = tol*tol*rhsNorm2;
+  RealScalar residualNorm2 = normal_residual.squaredNorm();
+  if (residualNorm2 < threshold)
+  {
+    iters = 0;
+    tol_error = sqrt(residualNorm2 / rhsNorm2);
+    return;
+  }
+  
+  VectorType p(n);
+  p = precond.solve(normal_residual);                         // initial search direction
+
+  VectorType z(n), tmp(m);
+  RealScalar absNew = numext::real(normal_residual.dot(p));  // the square of the absolute value of r scaled by invM
+  Index i = 0;
+  while(i < maxIters)
+  {
+    tmp.noalias() = mat * p;
+
+    Scalar alpha = absNew / tmp.squaredNorm();      // the amount we travel on dir
+    x += alpha * p;                                 // update solution
+    residual -= alpha * tmp;                        // update residual
+    normal_residual = mat.adjoint() * residual;     // update residual of the normal equation
+    
+    residualNorm2 = normal_residual.squaredNorm();
+    if(residualNorm2 < threshold)
+      break;
+    
+    z = precond.solve(normal_residual);             // approximately solve for "A'A z = normal_residual"
+
+    RealScalar absOld = absNew;
+    absNew = numext::real(normal_residual.dot(z));  // update the absolute value of r
+    RealScalar beta = absNew / absOld;              // calculate the Gram-Schmidt value used to create the new search direction
+    p = z + beta * p;                               // update search direction
+    i++;
+  }
+  tol_error = sqrt(residualNorm2 / rhsNorm2);
+  iters = i;
+}
+
+}
+
+template< typename _MatrixType,
+          typename _Preconditioner = LeastSquareDiagonalPreconditioner<typename _MatrixType::Scalar> >
+class LeastSquaresConjugateGradient;
+
+namespace internal {
+
+template< typename _MatrixType, typename _Preconditioner>
+struct traits<LeastSquaresConjugateGradient<_MatrixType,_Preconditioner> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Preconditioner Preconditioner;
+};
+
+}
+
+/** \ingroup IterativeLinearSolvers_Module
+  * \brief A conjugate gradient solver for sparse (or dense) least-square problems
+  *
+  * This class allows to solve for A x = b linear problems using an iterative conjugate gradient algorithm.
+  * The matrix A can be non symmetric and rectangular, but the matrix A' A should be positive-definite to guaranty stability.
+  * Otherwise, the SparseLU or SparseQR classes might be preferable.
+  * The matrix A and the vectors x and b can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix.
+  * \tparam _Preconditioner the type of the preconditioner. Default is LeastSquareDiagonalPreconditioner
+  *
+  * \implsparsesolverconcept
+  * 
+  * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
+  * and setTolerance() methods. The defaults are the size of the problem for the maximal number of iterations
+  * and NumTraits<Scalar>::epsilon() for the tolerance.
+  * 
+  * This class can be used as the direct solver classes. Here is a typical usage example:
+    \code
+    int m=1000000, n = 10000;
+    VectorXd x(n), b(m);
+    SparseMatrix<double> A(m,n);
+    // fill A and b
+    LeastSquaresConjugateGradient<SparseMatrix<double> > lscg;
+    lscg.compute(A);
+    x = lscg.solve(b);
+    std::cout << "#iterations:     " << lscg.iterations() << std::endl;
+    std::cout << "estimated error: " << lscg.error()      << std::endl;
+    // update b, and solve again
+    x = lscg.solve(b);
+    \endcode
+  * 
+  * By default the iterations start with x=0 as an initial guess of the solution.
+  * One can control the start using the solveWithGuess() method.
+  * 
+  * \sa class ConjugateGradient, SparseLU, SparseQR
+  */
+template< typename _MatrixType, typename _Preconditioner>
+class LeastSquaresConjugateGradient : public IterativeSolverBase<LeastSquaresConjugateGradient<_MatrixType,_Preconditioner> >
+{
+  typedef IterativeSolverBase<LeastSquaresConjugateGradient> Base;
+  using Base::matrix;
+  using Base::m_error;
+  using Base::m_iterations;
+  using Base::m_info;
+  using Base::m_isInitialized;
+public:
+  typedef _MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef _Preconditioner Preconditioner;
+
+public:
+
+  /** Default constructor. */
+  LeastSquaresConjugateGradient() : Base() {}
+
+  /** Initialize the solver with matrix \a A for further \c Ax=b solving.
+    * 
+    * This constructor is a shortcut for the default constructor followed
+    * by a call to compute().
+    * 
+    * \warning this class stores a reference to the matrix A as well as some
+    * precomputed values that depend on it. Therefore, if \a A is changed
+    * this class becomes invalid. Call compute() to update it with the new
+    * matrix A, or modify a copy of A.
+    */
+  template<typename MatrixDerived>
+  explicit LeastSquaresConjugateGradient(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
+
+  ~LeastSquaresConjugateGradient() {}
+
+  /** \internal */
+  template<typename Rhs,typename Dest>
+  void _solve_with_guess_impl(const Rhs& b, Dest& x) const
+  {
+    m_iterations = Base::maxIterations();
+    m_error = Base::m_tolerance;
+
+    for(Index j=0; j<b.cols(); ++j)
+    {
+      m_iterations = Base::maxIterations();
+      m_error = Base::m_tolerance;
+
+      typename Dest::ColXpr xj(x,j);
+      internal::least_square_conjugate_gradient(matrix(), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error);
+    }
+
+    m_isInitialized = true;
+    m_info = m_error <= Base::m_tolerance ? Success : NoConvergence;
+  }
+  
+  /** \internal */
+  using Base::_solve_impl;
+  template<typename Rhs,typename Dest>
+  void _solve_impl(const MatrixBase<Rhs>& b, Dest& x) const
+  {
+    x.setZero();
+    _solve_with_guess_impl(b.derived(),x);
+  }
+
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_LEAST_SQUARE_CONJUGATE_GRADIENT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h
new file mode 100644
index 0000000..0ace451
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/IterativeLinearSolvers/SolveWithGuess.h
@@ -0,0 +1,115 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SOLVEWITHGUESS_H
+#define EIGEN_SOLVEWITHGUESS_H
+
+namespace Eigen {
+
+template<typename Decomposition, typename RhsType, typename GuessType> class SolveWithGuess;
+  
+/** \class SolveWithGuess
+  * \ingroup IterativeLinearSolvers_Module
+  *
+  * \brief Pseudo expression representing a solving operation
+  *
+  * \tparam Decomposition the type of the matrix or decomposion object
+  * \tparam Rhstype the type of the right-hand side
+  *
+  * This class represents an expression of A.solve(B)
+  * and most of the time this is the only way it is used.
+  *
+  */
+namespace internal {
+
+
+template<typename Decomposition, typename RhsType, typename GuessType>
+struct traits<SolveWithGuess<Decomposition, RhsType, GuessType> >
+  : traits<Solve<Decomposition,RhsType> >
+{};
+
+}
+
+
+template<typename Decomposition, typename RhsType, typename GuessType>
+class SolveWithGuess : public internal::generic_xpr_base<SolveWithGuess<Decomposition,RhsType,GuessType>, MatrixXpr, typename internal::traits<RhsType>::StorageKind>::type
+{
+public:
+  typedef typename internal::traits<SolveWithGuess>::Scalar Scalar;
+  typedef typename internal::traits<SolveWithGuess>::PlainObject PlainObject;
+  typedef typename internal::generic_xpr_base<SolveWithGuess<Decomposition,RhsType,GuessType>, MatrixXpr, typename internal::traits<RhsType>::StorageKind>::type Base;
+  typedef typename internal::ref_selector<SolveWithGuess>::type Nested;
+  
+  SolveWithGuess(const Decomposition &dec, const RhsType &rhs, const GuessType &guess)
+    : m_dec(dec), m_rhs(rhs), m_guess(guess)
+  {}
+  
+  EIGEN_DEVICE_FUNC Index rows() const { return m_dec.cols(); }
+  EIGEN_DEVICE_FUNC Index cols() const { return m_rhs.cols(); }
+
+  EIGEN_DEVICE_FUNC const Decomposition& dec()   const { return m_dec; }
+  EIGEN_DEVICE_FUNC const RhsType&       rhs()   const { return m_rhs; }
+  EIGEN_DEVICE_FUNC const GuessType&     guess() const { return m_guess; }
+
+protected:
+  const Decomposition &m_dec;
+  const RhsType       &m_rhs;
+  const GuessType     &m_guess;
+  
+private:
+  Scalar coeff(Index row, Index col) const;
+  Scalar coeff(Index i) const;
+};
+
+namespace internal {
+
+// Evaluator of SolveWithGuess -> eval into a temporary
+template<typename Decomposition, typename RhsType, typename GuessType>
+struct evaluator<SolveWithGuess<Decomposition,RhsType, GuessType> >
+  : public evaluator<typename SolveWithGuess<Decomposition,RhsType,GuessType>::PlainObject>
+{
+  typedef SolveWithGuess<Decomposition,RhsType,GuessType> SolveType;
+  typedef typename SolveType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  evaluator(const SolveType& solve)
+    : m_result(solve.rows(), solve.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    m_result = solve.guess();
+    solve.dec()._solve_with_guess_impl(solve.rhs(), m_result);
+  }
+  
+protected:  
+  PlainObject m_result;
+};
+
+// Specialization for "dst = dec.solveWithGuess(rhs)"
+// NOTE we need to specialize it for Dense2Dense to avoid ambiguous specialization error and a Sparse2Sparse specialization must exist somewhere
+template<typename DstXprType, typename DecType, typename RhsType, typename GuessType, typename Scalar>
+struct Assignment<DstXprType, SolveWithGuess<DecType,RhsType,GuessType>, internal::assign_op<Scalar,Scalar>, Dense2Dense>
+{
+  typedef SolveWithGuess<DecType,RhsType,GuessType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    dst = src.guess();
+    src.dec()._solve_with_guess_impl(src.rhs(), dst/*, src.guess()*/);
+  }
+};
+
+} // end namepsace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SOLVEWITHGUESS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Jacobi/Jacobi.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Jacobi/Jacobi.h
new file mode 100644
index 0000000..c30326e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/Jacobi/Jacobi.h
@@ -0,0 +1,441 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_JACOBI_H
+#define EIGEN_JACOBI_H
+
+namespace Eigen { 
+
+/** \ingroup Jacobi_Module
+  * \jacobi_module
+  * \class JacobiRotation
+  * \brief Rotation given by a cosine-sine pair.
+  *
+  * This class represents a Jacobi or Givens rotation.
+  * This is a 2D rotation in the plane \c J of angle \f$ \theta \f$ defined by
+  * its cosine \c c and sine \c s as follow:
+  * \f$ J = \left ( \begin{array}{cc} c & \overline s \\ -s  & \overline c \end{array} \right ) \f$
+  *
+  * You can apply the respective counter-clockwise rotation to a column vector \c v by
+  * applying its adjoint on the left: \f$ v = J^* v \f$ that translates to the following Eigen code:
+  * \code
+  * v.applyOnTheLeft(J.adjoint());
+  * \endcode
+  *
+  * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+template<typename Scalar> class JacobiRotation
+{
+  public:
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    /** Default constructor without any initialization. */
+    JacobiRotation() {}
+
+    /** Construct a planar rotation from a cosine-sine pair (\a c, \c s). */
+    JacobiRotation(const Scalar& c, const Scalar& s) : m_c(c), m_s(s) {}
+
+    Scalar& c() { return m_c; }
+    Scalar c() const { return m_c; }
+    Scalar& s() { return m_s; }
+    Scalar s() const { return m_s; }
+
+    /** Concatenates two planar rotation */
+    JacobiRotation operator*(const JacobiRotation& other)
+    {
+      using numext::conj;
+      return JacobiRotation(m_c * other.m_c - conj(m_s) * other.m_s,
+                            conj(m_c * conj(other.m_s) + conj(m_s) * conj(other.m_c)));
+    }
+
+    /** Returns the transposed transformation */
+    JacobiRotation transpose() const { using numext::conj; return JacobiRotation(m_c, -conj(m_s)); }
+
+    /** Returns the adjoint transformation */
+    JacobiRotation adjoint() const { using numext::conj; return JacobiRotation(conj(m_c), -m_s); }
+
+    template<typename Derived>
+    bool makeJacobi(const MatrixBase<Derived>&, Index p, Index q);
+    bool makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z);
+
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z=0);
+
+  protected:
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::true_type);
+    void makeGivens(const Scalar& p, const Scalar& q, Scalar* z, internal::false_type);
+
+    Scalar m_c, m_s;
+};
+
+/** Makes \c *this as a Jacobi rotation \a J such that applying \a J on both the right and left sides of the selfadjoint 2x2 matrix
+  * \f$ B = \left ( \begin{array}{cc} x & y \\ \overline y & z \end{array} \right )\f$ yields a diagonal matrix \f$ A = J^* B J \f$
+  *
+  * \sa MatrixBase::makeJacobi(const MatrixBase<Derived>&, Index, Index), MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+template<typename Scalar>
+bool JacobiRotation<Scalar>::makeJacobi(const RealScalar& x, const Scalar& y, const RealScalar& z)
+{
+  using std::sqrt;
+  using std::abs;
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  RealScalar deno = RealScalar(2)*abs(y);
+  if(deno < (std::numeric_limits<RealScalar>::min)())
+  {
+    m_c = Scalar(1);
+    m_s = Scalar(0);
+    return false;
+  }
+  else
+  {
+    RealScalar tau = (x-z)/deno;
+    RealScalar w = sqrt(numext::abs2(tau) + RealScalar(1));
+    RealScalar t;
+    if(tau>RealScalar(0))
+    {
+      t = RealScalar(1) / (tau + w);
+    }
+    else
+    {
+      t = RealScalar(1) / (tau - w);
+    }
+    RealScalar sign_t = t > RealScalar(0) ? RealScalar(1) : RealScalar(-1);
+    RealScalar n = RealScalar(1) / sqrt(numext::abs2(t)+RealScalar(1));
+    m_s = - sign_t * (numext::conj(y) / abs(y)) * abs(t) * n;
+    m_c = n;
+    return true;
+  }
+}
+
+/** Makes \c *this as a Jacobi rotation \c J such that applying \a J on both the right and left sides of the 2x2 selfadjoint matrix
+  * \f$ B = \left ( \begin{array}{cc} \text{this}_{pp} & \text{this}_{pq} \\ (\text{this}_{pq})^* & \text{this}_{qq} \end{array} \right )\f$ yields
+  * a diagonal matrix \f$ A = J^* B J \f$
+  *
+  * Example: \include Jacobi_makeJacobi.cpp
+  * Output: \verbinclude Jacobi_makeJacobi.out
+  *
+  * \sa JacobiRotation::makeJacobi(RealScalar, Scalar, RealScalar), MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+template<typename Scalar>
+template<typename Derived>
+inline bool JacobiRotation<Scalar>::makeJacobi(const MatrixBase<Derived>& m, Index p, Index q)
+{
+  return makeJacobi(numext::real(m.coeff(p,p)), m.coeff(p,q), numext::real(m.coeff(q,q)));
+}
+
+/** Makes \c *this as a Givens rotation \c G such that applying \f$ G^* \f$ to the left of the vector
+  * \f$ V = \left ( \begin{array}{c} p \\ q \end{array} \right )\f$ yields:
+  * \f$ G^* V = \left ( \begin{array}{c} r \\ 0 \end{array} \right )\f$.
+  *
+  * The value of \a z is returned if \a z is not null (the default is null).
+  * Also note that G is built such that the cosine is always real.
+  *
+  * Example: \include Jacobi_makeGivens.cpp
+  * Output: \verbinclude Jacobi_makeGivens.out
+  *
+  * This function implements the continuous Givens rotation generation algorithm
+  * found in Anderson (2000), Discontinuous Plane Rotations and the Symmetric Eigenvalue Problem.
+  * LAPACK Working Note 150, University of Tennessee, UT-CS-00-454, December 4, 2000.
+  *
+  * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+template<typename Scalar>
+void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* z)
+{
+  makeGivens(p, q, z, typename internal::conditional<NumTraits<Scalar>::IsComplex, internal::true_type, internal::false_type>::type());
+}
+
+
+// specialization for complexes
+template<typename Scalar>
+void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::true_type)
+{
+  using std::sqrt;
+  using std::abs;
+  using numext::conj;
+  
+  if(q==Scalar(0))
+  {
+    m_c = numext::real(p)<0 ? Scalar(-1) : Scalar(1);
+    m_s = 0;
+    if(r) *r = m_c * p;
+  }
+  else if(p==Scalar(0))
+  {
+    m_c = 0;
+    m_s = -q/abs(q);
+    if(r) *r = abs(q);
+  }
+  else
+  {
+    RealScalar p1 = numext::norm1(p);
+    RealScalar q1 = numext::norm1(q);
+    if(p1>=q1)
+    {
+      Scalar ps = p / p1;
+      RealScalar p2 = numext::abs2(ps);
+      Scalar qs = q / p1;
+      RealScalar q2 = numext::abs2(qs);
+
+      RealScalar u = sqrt(RealScalar(1) + q2/p2);
+      if(numext::real(p)<RealScalar(0))
+        u = -u;
+
+      m_c = Scalar(1)/u;
+      m_s = -qs*conj(ps)*(m_c/p2);
+      if(r) *r = p * u;
+    }
+    else
+    {
+      Scalar ps = p / q1;
+      RealScalar p2 = numext::abs2(ps);
+      Scalar qs = q / q1;
+      RealScalar q2 = numext::abs2(qs);
+
+      RealScalar u = q1 * sqrt(p2 + q2);
+      if(numext::real(p)<RealScalar(0))
+        u = -u;
+
+      p1 = abs(p);
+      ps = p/p1;
+      m_c = p1/u;
+      m_s = -conj(ps) * (q/u);
+      if(r) *r = ps * u;
+    }
+  }
+}
+
+// specialization for reals
+template<typename Scalar>
+void JacobiRotation<Scalar>::makeGivens(const Scalar& p, const Scalar& q, Scalar* r, internal::false_type)
+{
+  using std::sqrt;
+  using std::abs;
+  if(q==Scalar(0))
+  {
+    m_c = p<Scalar(0) ? Scalar(-1) : Scalar(1);
+    m_s = Scalar(0);
+    if(r) *r = abs(p);
+  }
+  else if(p==Scalar(0))
+  {
+    m_c = Scalar(0);
+    m_s = q<Scalar(0) ? Scalar(1) : Scalar(-1);
+    if(r) *r = abs(q);
+  }
+  else if(abs(p) > abs(q))
+  {
+    Scalar t = q/p;
+    Scalar u = sqrt(Scalar(1) + numext::abs2(t));
+    if(p<Scalar(0))
+      u = -u;
+    m_c = Scalar(1)/u;
+    m_s = -t * m_c;
+    if(r) *r = p * u;
+  }
+  else
+  {
+    Scalar t = p/q;
+    Scalar u = sqrt(Scalar(1) + numext::abs2(t));
+    if(q<Scalar(0))
+      u = -u;
+    m_s = -Scalar(1)/u;
+    m_c = -t * m_s;
+    if(r) *r = q * u;
+  }
+
+}
+
+/****************************************************************************************
+*   Implementation of MatrixBase methods
+****************************************************************************************/
+
+namespace internal {
+/** \jacobi_module
+  * Applies the clock wise 2D rotation \a j to the set of 2D vectors of cordinates \a x and \a y:
+  * \f$ \left ( \begin{array}{cc} x \\ y \end{array} \right )  =  J \left ( \begin{array}{cc} x \\ y \end{array} \right ) \f$
+  *
+  * \sa MatrixBase::applyOnTheLeft(), MatrixBase::applyOnTheRight()
+  */
+template<typename VectorX, typename VectorY, typename OtherScalar>
+void apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j);
+}
+
+/** \jacobi_module
+  * Applies the rotation in the plane \a j to the rows \a p and \a q of \c *this, i.e., it computes B = J * B,
+  * with \f$ B = \left ( \begin{array}{cc} \text{*this.row}(p) \\ \text{*this.row}(q) \end{array} \right ) \f$.
+  *
+  * \sa class JacobiRotation, MatrixBase::applyOnTheRight(), internal::apply_rotation_in_the_plane()
+  */
+template<typename Derived>
+template<typename OtherScalar>
+inline void MatrixBase<Derived>::applyOnTheLeft(Index p, Index q, const JacobiRotation<OtherScalar>& j)
+{
+  RowXpr x(this->row(p));
+  RowXpr y(this->row(q));
+  internal::apply_rotation_in_the_plane(x, y, j);
+}
+
+/** \ingroup Jacobi_Module
+  * Applies the rotation in the plane \a j to the columns \a p and \a q of \c *this, i.e., it computes B = B * J
+  * with \f$ B = \left ( \begin{array}{cc} \text{*this.col}(p) & \text{*this.col}(q) \end{array} \right ) \f$.
+  *
+  * \sa class JacobiRotation, MatrixBase::applyOnTheLeft(), internal::apply_rotation_in_the_plane()
+  */
+template<typename Derived>
+template<typename OtherScalar>
+inline void MatrixBase<Derived>::applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j)
+{
+  ColXpr x(this->col(p));
+  ColXpr y(this->col(q));
+  internal::apply_rotation_in_the_plane(x, y, j.transpose());
+}
+
+namespace internal {
+template<typename VectorX, typename VectorY, typename OtherScalar>
+void /*EIGEN_DONT_INLINE*/ apply_rotation_in_the_plane(DenseBase<VectorX>& xpr_x, DenseBase<VectorY>& xpr_y, const JacobiRotation<OtherScalar>& j)
+{
+  typedef typename VectorX::Scalar Scalar;
+  enum {
+    PacketSize = packet_traits<Scalar>::size,
+    OtherPacketSize = packet_traits<OtherScalar>::size
+  };
+  typedef typename packet_traits<Scalar>::type Packet;
+  typedef typename packet_traits<OtherScalar>::type OtherPacket;
+  eigen_assert(xpr_x.size() == xpr_y.size());
+  Index size = xpr_x.size();
+  Index incrx = xpr_x.derived().innerStride();
+  Index incry = xpr_y.derived().innerStride();
+
+  Scalar* EIGEN_RESTRICT x = &xpr_x.derived().coeffRef(0);
+  Scalar* EIGEN_RESTRICT y = &xpr_y.derived().coeffRef(0);
+  
+  OtherScalar c = j.c();
+  OtherScalar s = j.s();
+  if (c==OtherScalar(1) && s==OtherScalar(0))
+    return;
+
+  /*** dynamic-size vectorized paths ***/
+
+  if(VectorX::SizeAtCompileTime == Dynamic &&
+    (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
+    (PacketSize == OtherPacketSize) &&
+    ((incrx==1 && incry==1) || PacketSize == 1))
+  {
+    // both vectors are sequentially stored in memory => vectorization
+    enum { Peeling = 2 };
+
+    Index alignedStart = internal::first_default_aligned(y, size);
+    Index alignedEnd = alignedStart + ((size-alignedStart)/PacketSize)*PacketSize;
+
+    const OtherPacket pc = pset1<OtherPacket>(c);
+    const OtherPacket ps = pset1<OtherPacket>(s);
+    conj_helper<OtherPacket,Packet,NumTraits<OtherScalar>::IsComplex,false> pcj;
+    conj_helper<OtherPacket,Packet,false,false> pm;
+
+    for(Index i=0; i<alignedStart; ++i)
+    {
+      Scalar xi = x[i];
+      Scalar yi = y[i];
+      x[i] =  c * xi + numext::conj(s) * yi;
+      y[i] = -s * xi + numext::conj(c) * yi;
+    }
+
+    Scalar* EIGEN_RESTRICT px = x + alignedStart;
+    Scalar* EIGEN_RESTRICT py = y + alignedStart;
+
+    if(internal::first_default_aligned(x, size)==alignedStart)
+    {
+      for(Index i=alignedStart; i<alignedEnd; i+=PacketSize)
+      {
+        Packet xi = pload<Packet>(px);
+        Packet yi = pload<Packet>(py);
+        pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+        px += PacketSize;
+        py += PacketSize;
+      }
+    }
+    else
+    {
+      Index peelingEnd = alignedStart + ((size-alignedStart)/(Peeling*PacketSize))*(Peeling*PacketSize);
+      for(Index i=alignedStart; i<peelingEnd; i+=Peeling*PacketSize)
+      {
+        Packet xi   = ploadu<Packet>(px);
+        Packet xi1  = ploadu<Packet>(px+PacketSize);
+        Packet yi   = pload <Packet>(py);
+        Packet yi1  = pload <Packet>(py+PacketSize);
+        pstoreu(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstoreu(px+PacketSize, padd(pm.pmul(pc,xi1),pcj.pmul(ps,yi1)));
+        pstore (py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+        pstore (py+PacketSize, psub(pcj.pmul(pc,yi1),pm.pmul(ps,xi1)));
+        px += Peeling*PacketSize;
+        py += Peeling*PacketSize;
+      }
+      if(alignedEnd!=peelingEnd)
+      {
+        Packet xi = ploadu<Packet>(x+peelingEnd);
+        Packet yi = pload <Packet>(y+peelingEnd);
+        pstoreu(x+peelingEnd, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+        pstore (y+peelingEnd, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+      }
+    }
+
+    for(Index i=alignedEnd; i<size; ++i)
+    {
+      Scalar xi = x[i];
+      Scalar yi = y[i];
+      x[i] =  c * xi + numext::conj(s) * yi;
+      y[i] = -s * xi + numext::conj(c) * yi;
+    }
+  }
+
+  /*** fixed-size vectorized path ***/
+  else if(VectorX::SizeAtCompileTime != Dynamic &&
+          (VectorX::Flags & VectorY::Flags & PacketAccessBit) &&
+          (PacketSize == OtherPacketSize) &&
+          (EIGEN_PLAIN_ENUM_MIN(evaluator<VectorX>::Alignment, evaluator<VectorY>::Alignment)>0)) // FIXME should be compared to the required alignment
+  {
+    const OtherPacket pc = pset1<OtherPacket>(c);
+    const OtherPacket ps = pset1<OtherPacket>(s);
+    conj_helper<OtherPacket,Packet,NumTraits<OtherPacket>::IsComplex,false> pcj;
+    conj_helper<OtherPacket,Packet,false,false> pm;
+    Scalar* EIGEN_RESTRICT px = x;
+    Scalar* EIGEN_RESTRICT py = y;
+    for(Index i=0; i<size; i+=PacketSize)
+    {
+      Packet xi = pload<Packet>(px);
+      Packet yi = pload<Packet>(py);
+      pstore(px, padd(pm.pmul(pc,xi),pcj.pmul(ps,yi)));
+      pstore(py, psub(pcj.pmul(pc,yi),pm.pmul(ps,xi)));
+      px += PacketSize;
+      py += PacketSize;
+    }
+  }
+
+  /*** non-vectorized path ***/
+  else
+  {
+    for(Index i=0; i<size; ++i)
+    {
+      Scalar xi = *x;
+      Scalar yi = *y;
+      *x =  c * xi + numext::conj(s) * yi;
+      *y = -s * xi + numext::conj(c) * yi;
+      x += incrx;
+      y += incry;
+    }
+  }
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_JACOBI_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/Determinant.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/Determinant.h
new file mode 100644
index 0000000..d6a3c1e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/Determinant.h
@@ -0,0 +1,101 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_DETERMINANT_H
+#define EIGEN_DETERMINANT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Derived>
+inline const typename Derived::Scalar bruteforce_det3_helper
+(const MatrixBase<Derived>& matrix, int a, int b, int c)
+{
+  return matrix.coeff(0,a)
+         * (matrix.coeff(1,b) * matrix.coeff(2,c) - matrix.coeff(1,c) * matrix.coeff(2,b));
+}
+
+template<typename Derived>
+const typename Derived::Scalar bruteforce_det4_helper
+(const MatrixBase<Derived>& matrix, int j, int k, int m, int n)
+{
+  return (matrix.coeff(j,0) * matrix.coeff(k,1) - matrix.coeff(k,0) * matrix.coeff(j,1))
+       * (matrix.coeff(m,2) * matrix.coeff(n,3) - matrix.coeff(n,2) * matrix.coeff(m,3));
+}
+
+template<typename Derived,
+         int DeterminantType = Derived::RowsAtCompileTime
+> struct determinant_impl
+{
+  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  {
+    if(Derived::ColsAtCompileTime==Dynamic && m.rows()==0)
+      return typename traits<Derived>::Scalar(1);
+    return m.partialPivLu().determinant();
+  }
+};
+
+template<typename Derived> struct determinant_impl<Derived, 1>
+{
+  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  {
+    return m.coeff(0,0);
+  }
+};
+
+template<typename Derived> struct determinant_impl<Derived, 2>
+{
+  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  {
+    return m.coeff(0,0) * m.coeff(1,1) - m.coeff(1,0) * m.coeff(0,1);
+  }
+};
+
+template<typename Derived> struct determinant_impl<Derived, 3>
+{
+  static inline typename traits<Derived>::Scalar run(const Derived& m)
+  {
+    return bruteforce_det3_helper(m,0,1,2)
+          - bruteforce_det3_helper(m,1,0,2)
+          + bruteforce_det3_helper(m,2,0,1);
+  }
+};
+
+template<typename Derived> struct determinant_impl<Derived, 4>
+{
+  static typename traits<Derived>::Scalar run(const Derived& m)
+  {
+    // trick by Martin Costabel to compute 4x4 det with only 30 muls
+    return bruteforce_det4_helper(m,0,1,2,3)
+          - bruteforce_det4_helper(m,0,2,1,3)
+          + bruteforce_det4_helper(m,0,3,1,2)
+          + bruteforce_det4_helper(m,1,2,0,3)
+          - bruteforce_det4_helper(m,1,3,0,2)
+          + bruteforce_det4_helper(m,2,3,0,1);
+  }
+};
+
+} // end namespace internal
+
+/** \lu_module
+  *
+  * \returns the determinant of this matrix
+  */
+template<typename Derived>
+inline typename internal::traits<Derived>::Scalar MatrixBase<Derived>::determinant() const
+{
+  eigen_assert(rows() == cols());
+  typedef typename internal::nested_eval<Derived,Base::RowsAtCompileTime>::type Nested;
+  return internal::determinant_impl<typename internal::remove_all<Nested>::type>::run(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_DETERMINANT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/FullPivLU.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/FullPivLU.h
new file mode 100644
index 0000000..03b6af7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/FullPivLU.h
@@ -0,0 +1,891 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_LU_H
+#define EIGEN_LU_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename _MatrixType> struct traits<FullPivLU<_MatrixType> >
+ : traits<_MatrixType>
+{
+  typedef MatrixXpr XprKind;
+  typedef SolverStorage StorageKind;
+  enum { Flags = 0 };
+};
+
+} // end namespace internal
+
+/** \ingroup LU_Module
+  *
+  * \class FullPivLU
+  *
+  * \brief LU decomposition of a matrix with complete pivoting, and related features
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the LU decomposition
+  *
+  * This class represents a LU decomposition of any matrix, with complete pivoting: the matrix A is
+  * decomposed as \f$ A = P^{-1} L U Q^{-1} \f$ where L is unit-lower-triangular, U is
+  * upper-triangular, and P and Q are permutation matrices. This is a rank-revealing LU
+  * decomposition. The eigenvalues (diagonal coefficients) of U are sorted in such a way that any
+  * zeros are at the end.
+  *
+  * This decomposition provides the generic approach to solving systems of linear equations, computing
+  * the rank, invertibility, inverse, kernel, and determinant.
+  *
+  * This LU decomposition is very stable and well tested with large matrices. However there are use cases where the SVD
+  * decomposition is inherently more stable and/or flexible. For example, when computing the kernel of a matrix,
+  * working with the SVD allows to select the smallest singular values of the matrix, something that
+  * the LU decomposition doesn't see.
+  *
+  * The data of the LU decomposition can be directly accessed through the methods matrixLU(),
+  * permutationP(), permutationQ().
+  *
+  * As an exemple, here is how the original matrix can be retrieved:
+  * \include class_FullPivLU.cpp
+  * Output: \verbinclude class_FullPivLU.out
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::fullPivLu(), MatrixBase::determinant(), MatrixBase::inverse()
+  */
+template<typename _MatrixType> class FullPivLU
+  : public SolverBase<FullPivLU<_MatrixType> >
+{
+  public:
+    typedef _MatrixType MatrixType;
+    typedef SolverBase<FullPivLU> Base;
+
+    EIGEN_GENERIC_PUBLIC_INTERFACE(FullPivLU)
+    // FIXME StorageIndex defined in EIGEN_GENERIC_PUBLIC_INTERFACE should be int
+    enum {
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename internal::plain_row_type<MatrixType, StorageIndex>::type IntRowVectorType;
+    typedef typename internal::plain_col_type<MatrixType, StorageIndex>::type IntColVectorType;
+    typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationQType;
+    typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationPType;
+    typedef typename MatrixType::PlainObject PlainObject;
+
+    /**
+      * \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via LU::compute(const MatrixType&).
+      */
+    FullPivLU();
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa FullPivLU()
+      */
+    FullPivLU(Index rows, Index cols);
+
+    /** Constructor.
+      *
+      * \param matrix the matrix of which to compute the LU decomposition.
+      *               It is required to be nonzero.
+      */
+    template<typename InputType>
+    explicit FullPivLU(const EigenBase<InputType>& matrix);
+
+    /** \brief Constructs a LU factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
+      *
+      * \sa FullPivLU(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit FullPivLU(EigenBase<InputType>& matrix);
+
+    /** Computes the LU decomposition of the given matrix.
+      *
+      * \param matrix the matrix of which to compute the LU decomposition.
+      *               It is required to be nonzero.
+      *
+      * \returns a reference to *this
+      */
+    template<typename InputType>
+    FullPivLU& compute(const EigenBase<InputType>& matrix) {
+      m_lu = matrix.derived();
+      computeInPlace();
+      return *this;
+    }
+
+    /** \returns the LU decomposition matrix: the upper-triangular part is U, the
+      * unit-lower-triangular part is L (at least for square matrices; in the non-square
+      * case, special care is needed, see the documentation of class FullPivLU).
+      *
+      * \sa matrixL(), matrixU()
+      */
+    inline const MatrixType& matrixLU() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return m_lu;
+    }
+
+    /** \returns the number of nonzero pivots in the LU decomposition.
+      * Here nonzero is meant in the exact sense, not in a fuzzy sense.
+      * So that notion isn't really intrinsically interesting, but it is
+      * still useful when implementing algorithms.
+      *
+      * \sa rank()
+      */
+    inline Index nonzeroPivots() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return m_nonzero_pivots;
+    }
+
+    /** \returns the absolute value of the biggest pivot, i.e. the biggest
+      *          diagonal coefficient of U.
+      */
+    RealScalar maxPivot() const { return m_maxpivot; }
+
+    /** \returns the permutation matrix P
+      *
+      * \sa permutationQ()
+      */
+    EIGEN_DEVICE_FUNC inline const PermutationPType& permutationP() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return m_p;
+    }
+
+    /** \returns the permutation matrix Q
+      *
+      * \sa permutationP()
+      */
+    inline const PermutationQType& permutationQ() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return m_q;
+    }
+
+    /** \returns the kernel of the matrix, also called its null-space. The columns of the returned matrix
+      * will form a basis of the kernel.
+      *
+      * \note If the kernel has dimension zero, then the returned matrix is a column-vector filled with zeros.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      *
+      * Example: \include FullPivLU_kernel.cpp
+      * Output: \verbinclude FullPivLU_kernel.out
+      *
+      * \sa image()
+      */
+    inline const internal::kernel_retval<FullPivLU> kernel() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return internal::kernel_retval<FullPivLU>(*this);
+    }
+
+    /** \returns the image of the matrix, also called its column-space. The columns of the returned matrix
+      * will form a basis of the image (column-space).
+      *
+      * \param originalMatrix the original matrix, of which *this is the LU decomposition.
+      *                       The reason why it is needed to pass it here, is that this allows
+      *                       a large optimization, as otherwise this method would need to reconstruct it
+      *                       from the LU decomposition.
+      *
+      * \note If the image has dimension zero, then the returned matrix is a column-vector filled with zeros.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      *
+      * Example: \include FullPivLU_image.cpp
+      * Output: \verbinclude FullPivLU_image.out
+      *
+      * \sa kernel()
+      */
+    inline const internal::image_retval<FullPivLU>
+      image(const MatrixType& originalMatrix) const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return internal::image_retval<FullPivLU>(*this, originalMatrix);
+    }
+
+    /** \return a solution x to the equation Ax=b, where A is the matrix of which
+      * *this is the LU decomposition.
+      *
+      * \param b the right-hand-side of the equation to solve. Can be a vector or a matrix,
+      *          the only requirement in order for the equation to make sense is that
+      *          b.rows()==A.rows(), where A is the matrix of which *this is the LU decomposition.
+      *
+      * \returns a solution.
+      *
+      * \note_about_checking_solutions
+      *
+      * \note_about_arbitrary_choice_of_solution
+      * \note_about_using_kernel_to_study_multiple_solutions
+      *
+      * Example: \include FullPivLU_solve.cpp
+      * Output: \verbinclude FullPivLU_solve.out
+      *
+      * \sa TriangularView::solve(), kernel(), inverse()
+      */
+    // FIXME this is a copy-paste of the base-class member to add the isInitialized assertion.
+    template<typename Rhs>
+    inline const Solve<FullPivLU, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return Solve<FullPivLU, Rhs>(*this, b.derived());
+    }
+
+    /** \returns an estimate of the reciprocal condition number of the matrix of which \c *this is
+        the LU decomposition.
+      */
+    inline RealScalar rcond() const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return internal::rcond_estimate_helper(m_l1_norm, *this);
+    }
+
+    /** \returns the determinant of the matrix of which
+      * *this is the LU decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the LU decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \note For fixed-size matrices of size up to 4, MatrixBase::determinant() offers
+      *       optimized paths.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      *
+      * \sa MatrixBase::determinant()
+      */
+    typename internal::traits<MatrixType>::Scalar determinant() const;
+
+    /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
+      * who need to determine when pivots are to be considered nonzero. This is not used for the
+      * LU decomposition itself.
+      *
+      * When it needs to get the threshold value, Eigen calls threshold(). By default, this
+      * uses a formula to automatically determine a reasonable threshold.
+      * Once you have called the present method setThreshold(const RealScalar&),
+      * your value is used instead.
+      *
+      * \param threshold The new value to use as the threshold.
+      *
+      * A pivot will be considered nonzero if its absolute value is strictly greater than
+      *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
+      * where maxpivot is the biggest pivot.
+      *
+      * If you want to come back to the default behavior, call setThreshold(Default_t)
+      */
+    FullPivLU& setThreshold(const RealScalar& threshold)
+    {
+      m_usePrescribedThreshold = true;
+      m_prescribedThreshold = threshold;
+      return *this;
+    }
+
+    /** Allows to come back to the default behavior, letting Eigen use its default formula for
+      * determining the threshold.
+      *
+      * You should pass the special object Eigen::Default as parameter here.
+      * \code lu.setThreshold(Eigen::Default); \endcode
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    FullPivLU& setThreshold(Default_t)
+    {
+      m_usePrescribedThreshold = false;
+      return *this;
+    }
+
+    /** Returns the threshold that will be used by certain methods such as rank().
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    RealScalar threshold() const
+    {
+      eigen_assert(m_isInitialized || m_usePrescribedThreshold);
+      return m_usePrescribedThreshold ? m_prescribedThreshold
+      // this formula comes from experimenting (see "LU precision tuning" thread on the list)
+      // and turns out to be identical to Higham's formula used already in LDLt.
+                                      : NumTraits<Scalar>::epsilon() * m_lu.diagonalSize();
+    }
+
+    /** \returns the rank of the matrix of which *this is the LU decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index rank() const
+    {
+      using std::abs;
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold();
+      Index result = 0;
+      for(Index i = 0; i < m_nonzero_pivots; ++i)
+        result += (abs(m_lu.coeff(i,i)) > premultiplied_threshold);
+      return result;
+    }
+
+    /** \returns the dimension of the kernel of the matrix of which *this is the LU decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index dimensionOfKernel() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return cols() - rank();
+    }
+
+    /** \returns true if the matrix of which *this is the LU decomposition represents an injective
+      *          linear map, i.e. has trivial kernel; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInjective() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return rank() == cols();
+    }
+
+    /** \returns true if the matrix of which *this is the LU decomposition represents a surjective
+      *          linear map; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isSurjective() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return rank() == rows();
+    }
+
+    /** \returns true if the matrix of which *this is the LU decomposition is invertible.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInvertible() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return isInjective() && (m_lu.rows() == m_lu.cols());
+    }
+
+    /** \returns the inverse of the matrix of which *this is the LU decomposition.
+      *
+      * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
+      *       Use isInvertible() to first determine whether this matrix is invertible.
+      *
+      * \sa MatrixBase::inverse()
+      */
+    inline const Inverse<FullPivLU> inverse() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the inverse of a non-square matrix!");
+      return Inverse<FullPivLU>(*this);
+    }
+
+    MatrixType reconstructedMatrix() const;
+
+    EIGEN_DEVICE_FUNC inline Index rows() const { return m_lu.rows(); }
+    EIGEN_DEVICE_FUNC inline Index cols() const { return m_lu.cols(); }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+
+    template<bool Conjugate, typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl_transposed(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    void computeInPlace();
+
+    MatrixType m_lu;
+    PermutationPType m_p;
+    PermutationQType m_q;
+    IntColVectorType m_rowsTranspositions;
+    IntRowVectorType m_colsTranspositions;
+    Index m_nonzero_pivots;
+    RealScalar m_l1_norm;
+    RealScalar m_maxpivot, m_prescribedThreshold;
+    signed char m_det_pq;
+    bool m_isInitialized, m_usePrescribedThreshold;
+};
+
+template<typename MatrixType>
+FullPivLU<MatrixType>::FullPivLU()
+  : m_isInitialized(false), m_usePrescribedThreshold(false)
+{
+}
+
+template<typename MatrixType>
+FullPivLU<MatrixType>::FullPivLU(Index rows, Index cols)
+  : m_lu(rows, cols),
+    m_p(rows),
+    m_q(cols),
+    m_rowsTranspositions(rows),
+    m_colsTranspositions(cols),
+    m_isInitialized(false),
+    m_usePrescribedThreshold(false)
+{
+}
+
+template<typename MatrixType>
+template<typename InputType>
+FullPivLU<MatrixType>::FullPivLU(const EigenBase<InputType>& matrix)
+  : m_lu(matrix.rows(), matrix.cols()),
+    m_p(matrix.rows()),
+    m_q(matrix.cols()),
+    m_rowsTranspositions(matrix.rows()),
+    m_colsTranspositions(matrix.cols()),
+    m_isInitialized(false),
+    m_usePrescribedThreshold(false)
+{
+  compute(matrix.derived());
+}
+
+template<typename MatrixType>
+template<typename InputType>
+FullPivLU<MatrixType>::FullPivLU(EigenBase<InputType>& matrix)
+  : m_lu(matrix.derived()),
+    m_p(matrix.rows()),
+    m_q(matrix.cols()),
+    m_rowsTranspositions(matrix.rows()),
+    m_colsTranspositions(matrix.cols()),
+    m_isInitialized(false),
+    m_usePrescribedThreshold(false)
+{
+  computeInPlace();
+}
+
+template<typename MatrixType>
+void FullPivLU<MatrixType>::computeInPlace()
+{
+  check_template_parameters();
+
+  // the permutations are stored as int indices, so just to be sure:
+  eigen_assert(m_lu.rows()<=NumTraits<int>::highest() && m_lu.cols()<=NumTraits<int>::highest());
+
+  m_l1_norm = m_lu.cwiseAbs().colwise().sum().maxCoeff();
+
+  const Index size = m_lu.diagonalSize();
+  const Index rows = m_lu.rows();
+  const Index cols = m_lu.cols();
+
+  // will store the transpositions, before we accumulate them at the end.
+  // can't accumulate on-the-fly because that will be done in reverse order for the rows.
+  m_rowsTranspositions.resize(m_lu.rows());
+  m_colsTranspositions.resize(m_lu.cols());
+  Index number_of_transpositions = 0; // number of NONTRIVIAL transpositions, i.e. m_rowsTranspositions[i]!=i
+
+  m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
+  m_maxpivot = RealScalar(0);
+
+  for(Index k = 0; k < size; ++k)
+  {
+    // First, we need to find the pivot.
+
+    // biggest coefficient in the remaining bottom-right corner (starting at row k, col k)
+    Index row_of_biggest_in_corner, col_of_biggest_in_corner;
+    typedef internal::scalar_score_coeff_op<Scalar> Scoring;
+    typedef typename Scoring::result_type Score;
+    Score biggest_in_corner;
+    biggest_in_corner = m_lu.bottomRightCorner(rows-k, cols-k)
+                        .unaryExpr(Scoring())
+                        .maxCoeff(&row_of_biggest_in_corner, &col_of_biggest_in_corner);
+    row_of_biggest_in_corner += k; // correct the values! since they were computed in the corner,
+    col_of_biggest_in_corner += k; // need to add k to them.
+
+    if(biggest_in_corner==Score(0))
+    {
+      // before exiting, make sure to initialize the still uninitialized transpositions
+      // in a sane state without destroying what we already have.
+      m_nonzero_pivots = k;
+      for(Index i = k; i < size; ++i)
+      {
+        m_rowsTranspositions.coeffRef(i) = i;
+        m_colsTranspositions.coeffRef(i) = i;
+      }
+      break;
+    }
+
+    RealScalar abs_pivot = internal::abs_knowing_score<Scalar>()(m_lu(row_of_biggest_in_corner, col_of_biggest_in_corner), biggest_in_corner);
+    if(abs_pivot > m_maxpivot) m_maxpivot = abs_pivot;
+
+    // Now that we've found the pivot, we need to apply the row/col swaps to
+    // bring it to the location (k,k).
+
+    m_rowsTranspositions.coeffRef(k) = row_of_biggest_in_corner;
+    m_colsTranspositions.coeffRef(k) = col_of_biggest_in_corner;
+    if(k != row_of_biggest_in_corner) {
+      m_lu.row(k).swap(m_lu.row(row_of_biggest_in_corner));
+      ++number_of_transpositions;
+    }
+    if(k != col_of_biggest_in_corner) {
+      m_lu.col(k).swap(m_lu.col(col_of_biggest_in_corner));
+      ++number_of_transpositions;
+    }
+
+    // Now that the pivot is at the right location, we update the remaining
+    // bottom-right corner by Gaussian elimination.
+
+    if(k<rows-1)
+      m_lu.col(k).tail(rows-k-1) /= m_lu.coeff(k,k);
+    if(k<size-1)
+      m_lu.block(k+1,k+1,rows-k-1,cols-k-1).noalias() -= m_lu.col(k).tail(rows-k-1) * m_lu.row(k).tail(cols-k-1);
+  }
+
+  // the main loop is over, we still have to accumulate the transpositions to find the
+  // permutations P and Q
+
+  m_p.setIdentity(rows);
+  for(Index k = size-1; k >= 0; --k)
+    m_p.applyTranspositionOnTheRight(k, m_rowsTranspositions.coeff(k));
+
+  m_q.setIdentity(cols);
+  for(Index k = 0; k < size; ++k)
+    m_q.applyTranspositionOnTheRight(k, m_colsTranspositions.coeff(k));
+
+  m_det_pq = (number_of_transpositions%2) ? -1 : 1;
+
+  m_isInitialized = true;
+}
+
+template<typename MatrixType>
+typename internal::traits<MatrixType>::Scalar FullPivLU<MatrixType>::determinant() const
+{
+  eigen_assert(m_isInitialized && "LU is not initialized.");
+  eigen_assert(m_lu.rows() == m_lu.cols() && "You can't take the determinant of a non-square matrix!");
+  return Scalar(m_det_pq) * Scalar(m_lu.diagonal().prod());
+}
+
+/** \returns the matrix represented by the decomposition,
+ * i.e., it returns the product: \f$ P^{-1} L U Q^{-1} \f$.
+ * This function is provided for debug purposes. */
+template<typename MatrixType>
+MatrixType FullPivLU<MatrixType>::reconstructedMatrix() const
+{
+  eigen_assert(m_isInitialized && "LU is not initialized.");
+  const Index smalldim = (std::min)(m_lu.rows(), m_lu.cols());
+  // LU
+  MatrixType res(m_lu.rows(),m_lu.cols());
+  // FIXME the .toDenseMatrix() should not be needed...
+  res = m_lu.leftCols(smalldim)
+            .template triangularView<UnitLower>().toDenseMatrix()
+      * m_lu.topRows(smalldim)
+            .template triangularView<Upper>().toDenseMatrix();
+
+  // P^{-1}(LU)
+  res = m_p.inverse() * res;
+
+  // (P^{-1}LU)Q^{-1}
+  res = res * m_q.inverse();
+
+  return res;
+}
+
+/********* Implementation of kernel() **************************************************/
+
+namespace internal {
+template<typename _MatrixType>
+struct kernel_retval<FullPivLU<_MatrixType> >
+  : kernel_retval_base<FullPivLU<_MatrixType> >
+{
+  EIGEN_MAKE_KERNEL_HELPERS(FullPivLU<_MatrixType>)
+
+  enum { MaxSmallDimAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(
+            MatrixType::MaxColsAtCompileTime,
+            MatrixType::MaxRowsAtCompileTime)
+  };
+
+  template<typename Dest> void evalTo(Dest& dst) const
+  {
+    using std::abs;
+    const Index cols = dec().matrixLU().cols(), dimker = cols - rank();
+    if(dimker == 0)
+    {
+      // The Kernel is just {0}, so it doesn't have a basis properly speaking, but let's
+      // avoid crashing/asserting as that depends on floating point calculations. Let's
+      // just return a single column vector filled with zeros.
+      dst.setZero();
+      return;
+    }
+
+    /* Let us use the following lemma:
+      *
+      * Lemma: If the matrix A has the LU decomposition PAQ = LU,
+      * then Ker A = Q(Ker U).
+      *
+      * Proof: trivial: just keep in mind that P, Q, L are invertible.
+      */
+
+    /* Thus, all we need to do is to compute Ker U, and then apply Q.
+      *
+      * U is upper triangular, with eigenvalues sorted so that any zeros appear at the end.
+      * Thus, the diagonal of U ends with exactly
+      * dimKer zero's. Let us use that to construct dimKer linearly
+      * independent vectors in Ker U.
+      */
+
+    Matrix<Index, Dynamic, 1, 0, MaxSmallDimAtCompileTime, 1> pivots(rank());
+    RealScalar premultiplied_threshold = dec().maxPivot() * dec().threshold();
+    Index p = 0;
+    for(Index i = 0; i < dec().nonzeroPivots(); ++i)
+      if(abs(dec().matrixLU().coeff(i,i)) > premultiplied_threshold)
+        pivots.coeffRef(p++) = i;
+    eigen_internal_assert(p == rank());
+
+    // we construct a temporaty trapezoid matrix m, by taking the U matrix and
+    // permuting the rows and cols to bring the nonnegligible pivots to the top of
+    // the main diagonal. We need that to be able to apply our triangular solvers.
+    // FIXME when we get triangularView-for-rectangular-matrices, this can be simplified
+    Matrix<typename MatrixType::Scalar, Dynamic, Dynamic, MatrixType::Options,
+           MaxSmallDimAtCompileTime, MatrixType::MaxColsAtCompileTime>
+      m(dec().matrixLU().block(0, 0, rank(), cols));
+    for(Index i = 0; i < rank(); ++i)
+    {
+      if(i) m.row(i).head(i).setZero();
+      m.row(i).tail(cols-i) = dec().matrixLU().row(pivots.coeff(i)).tail(cols-i);
+    }
+    m.block(0, 0, rank(), rank());
+    m.block(0, 0, rank(), rank()).template triangularView<StrictlyLower>().setZero();
+    for(Index i = 0; i < rank(); ++i)
+      m.col(i).swap(m.col(pivots.coeff(i)));
+
+    // ok, we have our trapezoid matrix, we can apply the triangular solver.
+    // notice that the math behind this suggests that we should apply this to the
+    // negative of the RHS, but for performance we just put the negative sign elsewhere, see below.
+    m.topLeftCorner(rank(), rank())
+     .template triangularView<Upper>().solveInPlace(
+        m.topRightCorner(rank(), dimker)
+      );
+
+    // now we must undo the column permutation that we had applied!
+    for(Index i = rank()-1; i >= 0; --i)
+      m.col(i).swap(m.col(pivots.coeff(i)));
+
+    // see the negative sign in the next line, that's what we were talking about above.
+    for(Index i = 0; i < rank(); ++i) dst.row(dec().permutationQ().indices().coeff(i)) = -m.row(i).tail(dimker);
+    for(Index i = rank(); i < cols; ++i) dst.row(dec().permutationQ().indices().coeff(i)).setZero();
+    for(Index k = 0; k < dimker; ++k) dst.coeffRef(dec().permutationQ().indices().coeff(rank()+k), k) = Scalar(1);
+  }
+};
+
+/***** Implementation of image() *****************************************************/
+
+template<typename _MatrixType>
+struct image_retval<FullPivLU<_MatrixType> >
+  : image_retval_base<FullPivLU<_MatrixType> >
+{
+  EIGEN_MAKE_IMAGE_HELPERS(FullPivLU<_MatrixType>)
+
+  enum { MaxSmallDimAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(
+            MatrixType::MaxColsAtCompileTime,
+            MatrixType::MaxRowsAtCompileTime)
+  };
+
+  template<typename Dest> void evalTo(Dest& dst) const
+  {
+    using std::abs;
+    if(rank() == 0)
+    {
+      // The Image is just {0}, so it doesn't have a basis properly speaking, but let's
+      // avoid crashing/asserting as that depends on floating point calculations. Let's
+      // just return a single column vector filled with zeros.
+      dst.setZero();
+      return;
+    }
+
+    Matrix<Index, Dynamic, 1, 0, MaxSmallDimAtCompileTime, 1> pivots(rank());
+    RealScalar premultiplied_threshold = dec().maxPivot() * dec().threshold();
+    Index p = 0;
+    for(Index i = 0; i < dec().nonzeroPivots(); ++i)
+      if(abs(dec().matrixLU().coeff(i,i)) > premultiplied_threshold)
+        pivots.coeffRef(p++) = i;
+    eigen_internal_assert(p == rank());
+
+    for(Index i = 0; i < rank(); ++i)
+      dst.col(i) = originalMatrix().col(dec().permutationQ().indices().coeff(pivots.coeff(i)));
+  }
+};
+
+/***** Implementation of solve() *****************************************************/
+
+} // end namespace internal
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType>
+template<typename RhsType, typename DstType>
+void FullPivLU<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1}.
+  * So we proceed as follows:
+  * Step 1: compute c = P * rhs.
+  * Step 2: replace c by the solution x to Lx = c. Exists because L is invertible.
+  * Step 3: replace c by the solution x to Ux = c. May or may not exist.
+  * Step 4: result = Q * c;
+  */
+
+  const Index rows = this->rows(),
+              cols = this->cols(),
+              nonzero_pivots = this->rank();
+  eigen_assert(rhs.rows() == rows);
+  const Index smalldim = (std::min)(rows, cols);
+
+  if(nonzero_pivots == 0)
+  {
+    dst.setZero();
+    return;
+  }
+
+  typename RhsType::PlainObject c(rhs.rows(), rhs.cols());
+
+  // Step 1
+  c = permutationP() * rhs;
+
+  // Step 2
+  m_lu.topLeftCorner(smalldim,smalldim)
+      .template triangularView<UnitLower>()
+      .solveInPlace(c.topRows(smalldim));
+  if(rows>cols)
+    c.bottomRows(rows-cols) -= m_lu.bottomRows(rows-cols) * c.topRows(cols);
+
+  // Step 3
+  m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
+      .template triangularView<Upper>()
+      .solveInPlace(c.topRows(nonzero_pivots));
+
+  // Step 4
+  for(Index i = 0; i < nonzero_pivots; ++i)
+    dst.row(permutationQ().indices().coeff(i)) = c.row(i);
+  for(Index i = nonzero_pivots; i < m_lu.cols(); ++i)
+    dst.row(permutationQ().indices().coeff(i)).setZero();
+}
+
+template<typename _MatrixType>
+template<bool Conjugate, typename RhsType, typename DstType>
+void FullPivLU<_MatrixType>::_solve_impl_transposed(const RhsType &rhs, DstType &dst) const
+{
+  /* The decomposition PAQ = LU can be rewritten as A = P^{-1} L U Q^{-1},
+   * and since permutations are real and unitary, we can write this
+   * as   A^T = Q U^T L^T P,
+   * So we proceed as follows:
+   * Step 1: compute c = Q^T rhs.
+   * Step 2: replace c by the solution x to U^T x = c. May or may not exist.
+   * Step 3: replace c by the solution x to L^T x = c.
+   * Step 4: result = P^T c.
+   * If Conjugate is true, replace "^T" by "^*" above.
+   */
+
+  const Index rows = this->rows(), cols = this->cols(),
+    nonzero_pivots = this->rank();
+   eigen_assert(rhs.rows() == cols);
+  const Index smalldim = (std::min)(rows, cols);
+
+  if(nonzero_pivots == 0)
+  {
+    dst.setZero();
+    return;
+  }
+
+  typename RhsType::PlainObject c(rhs.rows(), rhs.cols());
+
+  // Step 1
+  c = permutationQ().inverse() * rhs;
+
+  if (Conjugate) {
+    // Step 2
+    m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
+        .template triangularView<Upper>()
+        .adjoint()
+        .solveInPlace(c.topRows(nonzero_pivots));
+    // Step 3
+    m_lu.topLeftCorner(smalldim, smalldim)
+        .template triangularView<UnitLower>()
+        .adjoint()
+        .solveInPlace(c.topRows(smalldim));
+  } else {
+    // Step 2
+    m_lu.topLeftCorner(nonzero_pivots, nonzero_pivots)
+        .template triangularView<Upper>()
+        .transpose()
+        .solveInPlace(c.topRows(nonzero_pivots));
+    // Step 3
+    m_lu.topLeftCorner(smalldim, smalldim)
+        .template triangularView<UnitLower>()
+        .transpose()
+        .solveInPlace(c.topRows(smalldim));
+  }
+
+  // Step 4
+  PermutationPType invp = permutationP().inverse().eval();
+  for(Index i = 0; i < smalldim; ++i)
+    dst.row(invp.indices().coeff(i)) = c.row(i);
+  for(Index i = smalldim; i < rows; ++i)
+    dst.row(invp.indices().coeff(i)).setZero();
+}
+
+#endif
+
+namespace internal {
+
+
+/***** Implementation of inverse() *****************************************************/
+template<typename DstXprType, typename MatrixType>
+struct Assignment<DstXprType, Inverse<FullPivLU<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename FullPivLU<MatrixType>::Scalar>, Dense2Dense>
+{
+  typedef FullPivLU<MatrixType> LuType;
+  typedef Inverse<LuType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename MatrixType::Scalar> &)
+  {
+    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols()));
+  }
+};
+} // end namespace internal
+
+/******* MatrixBase methods *****************************************************************/
+
+/** \lu_module
+  *
+  * \return the full-pivoting LU decomposition of \c *this.
+  *
+  * \sa class FullPivLU
+  */
+template<typename Derived>
+inline const FullPivLU<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::fullPivLu() const
+{
+  return FullPivLU<PlainObject>(eval());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_LU_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/InverseImpl.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/InverseImpl.h
new file mode 100644
index 0000000..018f99b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/InverseImpl.h
@@ -0,0 +1,415 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_INVERSE_IMPL_H
+#define EIGEN_INVERSE_IMPL_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/**********************************
+*** General case implementation ***
+**********************************/
+
+template<typename MatrixType, typename ResultType, int Size = MatrixType::RowsAtCompileTime>
+struct compute_inverse
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const MatrixType& matrix, ResultType& result)
+  {
+    result = matrix.partialPivLu().inverse();
+  }
+};
+
+template<typename MatrixType, typename ResultType, int Size = MatrixType::RowsAtCompileTime>
+struct compute_inverse_and_det_with_check { /* nothing! general case not supported. */ };
+
+/****************************
+*** Size 1 implementation ***
+****************************/
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse<MatrixType, ResultType, 1>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const MatrixType& matrix, ResultType& result)
+  {
+    typedef typename MatrixType::Scalar Scalar;
+    internal::evaluator<MatrixType> matrixEval(matrix);
+    result.coeffRef(0,0) = Scalar(1) / matrixEval.coeff(0,0);
+  }
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_and_det_with_check<MatrixType, ResultType, 1>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(
+    const MatrixType& matrix,
+    const typename MatrixType::RealScalar& absDeterminantThreshold,
+    ResultType& result,
+    typename ResultType::Scalar& determinant,
+    bool& invertible
+  )
+  {
+    using std::abs;
+    determinant = matrix.coeff(0,0);
+    invertible = abs(determinant) > absDeterminantThreshold;
+    if(invertible) result.coeffRef(0,0) = typename ResultType::Scalar(1) / determinant;
+  }
+};
+
+/****************************
+*** Size 2 implementation ***
+****************************/
+
+template<typename MatrixType, typename ResultType>
+EIGEN_DEVICE_FUNC 
+inline void compute_inverse_size2_helper(
+    const MatrixType& matrix, const typename ResultType::Scalar& invdet,
+    ResultType& result)
+{
+  result.coeffRef(0,0) =  matrix.coeff(1,1) * invdet;
+  result.coeffRef(1,0) = -matrix.coeff(1,0) * invdet;
+  result.coeffRef(0,1) = -matrix.coeff(0,1) * invdet;
+  result.coeffRef(1,1) =  matrix.coeff(0,0) * invdet;
+}
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse<MatrixType, ResultType, 2>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const MatrixType& matrix, ResultType& result)
+  {
+    typedef typename ResultType::Scalar Scalar;
+    const Scalar invdet = typename MatrixType::Scalar(1) / matrix.determinant();
+    compute_inverse_size2_helper(matrix, invdet, result);
+  }
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_and_det_with_check<MatrixType, ResultType, 2>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(
+    const MatrixType& matrix,
+    const typename MatrixType::RealScalar& absDeterminantThreshold,
+    ResultType& inverse,
+    typename ResultType::Scalar& determinant,
+    bool& invertible
+  )
+  {
+    using std::abs;
+    typedef typename ResultType::Scalar Scalar;
+    determinant = matrix.determinant();
+    invertible = abs(determinant) > absDeterminantThreshold;
+    if(!invertible) return;
+    const Scalar invdet = Scalar(1) / determinant;
+    compute_inverse_size2_helper(matrix, invdet, inverse);
+  }
+};
+
+/****************************
+*** Size 3 implementation ***
+****************************/
+
+template<typename MatrixType, int i, int j>
+EIGEN_DEVICE_FUNC 
+inline typename MatrixType::Scalar cofactor_3x3(const MatrixType& m)
+{
+  enum {
+    i1 = (i+1) % 3,
+    i2 = (i+2) % 3,
+    j1 = (j+1) % 3,
+    j2 = (j+2) % 3
+  };
+  return m.coeff(i1, j1) * m.coeff(i2, j2)
+       - m.coeff(i1, j2) * m.coeff(i2, j1);
+}
+
+template<typename MatrixType, typename ResultType>
+EIGEN_DEVICE_FUNC
+inline void compute_inverse_size3_helper(
+    const MatrixType& matrix,
+    const typename ResultType::Scalar& invdet,
+    const Matrix<typename ResultType::Scalar,3,1>& cofactors_col0,
+    ResultType& result)
+{
+  result.row(0) = cofactors_col0 * invdet;
+  result.coeffRef(1,0) =  cofactor_3x3<MatrixType,0,1>(matrix) * invdet;
+  result.coeffRef(1,1) =  cofactor_3x3<MatrixType,1,1>(matrix) * invdet;
+  result.coeffRef(1,2) =  cofactor_3x3<MatrixType,2,1>(matrix) * invdet;
+  result.coeffRef(2,0) =  cofactor_3x3<MatrixType,0,2>(matrix) * invdet;
+  result.coeffRef(2,1) =  cofactor_3x3<MatrixType,1,2>(matrix) * invdet;
+  result.coeffRef(2,2) =  cofactor_3x3<MatrixType,2,2>(matrix) * invdet;
+}
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse<MatrixType, ResultType, 3>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(const MatrixType& matrix, ResultType& result)
+  {
+    typedef typename ResultType::Scalar Scalar;
+    Matrix<typename MatrixType::Scalar,3,1> cofactors_col0;
+    cofactors_col0.coeffRef(0) =  cofactor_3x3<MatrixType,0,0>(matrix);
+    cofactors_col0.coeffRef(1) =  cofactor_3x3<MatrixType,1,0>(matrix);
+    cofactors_col0.coeffRef(2) =  cofactor_3x3<MatrixType,2,0>(matrix);
+    const Scalar det = (cofactors_col0.cwiseProduct(matrix.col(0))).sum();
+    const Scalar invdet = Scalar(1) / det;
+    compute_inverse_size3_helper(matrix, invdet, cofactors_col0, result);
+  }
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_and_det_with_check<MatrixType, ResultType, 3>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(
+    const MatrixType& matrix,
+    const typename MatrixType::RealScalar& absDeterminantThreshold,
+    ResultType& inverse,
+    typename ResultType::Scalar& determinant,
+    bool& invertible
+  )
+  {
+    using std::abs;
+    typedef typename ResultType::Scalar Scalar;
+    Matrix<Scalar,3,1> cofactors_col0;
+    cofactors_col0.coeffRef(0) =  cofactor_3x3<MatrixType,0,0>(matrix);
+    cofactors_col0.coeffRef(1) =  cofactor_3x3<MatrixType,1,0>(matrix);
+    cofactors_col0.coeffRef(2) =  cofactor_3x3<MatrixType,2,0>(matrix);
+    determinant = (cofactors_col0.cwiseProduct(matrix.col(0))).sum();
+    invertible = abs(determinant) > absDeterminantThreshold;
+    if(!invertible) return;
+    const Scalar invdet = Scalar(1) / determinant;
+    compute_inverse_size3_helper(matrix, invdet, cofactors_col0, inverse);
+  }
+};
+
+/****************************
+*** Size 4 implementation ***
+****************************/
+
+template<typename Derived>
+EIGEN_DEVICE_FUNC 
+inline const typename Derived::Scalar general_det3_helper
+(const MatrixBase<Derived>& matrix, int i1, int i2, int i3, int j1, int j2, int j3)
+{
+  return matrix.coeff(i1,j1)
+         * (matrix.coeff(i2,j2) * matrix.coeff(i3,j3) - matrix.coeff(i2,j3) * matrix.coeff(i3,j2));
+}
+
+template<typename MatrixType, int i, int j>
+EIGEN_DEVICE_FUNC 
+inline typename MatrixType::Scalar cofactor_4x4(const MatrixType& matrix)
+{
+  enum {
+    i1 = (i+1) % 4,
+    i2 = (i+2) % 4,
+    i3 = (i+3) % 4,
+    j1 = (j+1) % 4,
+    j2 = (j+2) % 4,
+    j3 = (j+3) % 4
+  };
+  return general_det3_helper(matrix, i1, i2, i3, j1, j2, j3)
+       + general_det3_helper(matrix, i2, i3, i1, j1, j2, j3)
+       + general_det3_helper(matrix, i3, i1, i2, j1, j2, j3);
+}
+
+template<int Arch, typename Scalar, typename MatrixType, typename ResultType>
+struct compute_inverse_size4
+{
+  EIGEN_DEVICE_FUNC
+  static void run(const MatrixType& matrix, ResultType& result)
+  {
+    result.coeffRef(0,0) =  cofactor_4x4<MatrixType,0,0>(matrix);
+    result.coeffRef(1,0) = -cofactor_4x4<MatrixType,0,1>(matrix);
+    result.coeffRef(2,0) =  cofactor_4x4<MatrixType,0,2>(matrix);
+    result.coeffRef(3,0) = -cofactor_4x4<MatrixType,0,3>(matrix);
+    result.coeffRef(0,2) =  cofactor_4x4<MatrixType,2,0>(matrix);
+    result.coeffRef(1,2) = -cofactor_4x4<MatrixType,2,1>(matrix);
+    result.coeffRef(2,2) =  cofactor_4x4<MatrixType,2,2>(matrix);
+    result.coeffRef(3,2) = -cofactor_4x4<MatrixType,2,3>(matrix);
+    result.coeffRef(0,1) = -cofactor_4x4<MatrixType,1,0>(matrix);
+    result.coeffRef(1,1) =  cofactor_4x4<MatrixType,1,1>(matrix);
+    result.coeffRef(2,1) = -cofactor_4x4<MatrixType,1,2>(matrix);
+    result.coeffRef(3,1) =  cofactor_4x4<MatrixType,1,3>(matrix);
+    result.coeffRef(0,3) = -cofactor_4x4<MatrixType,3,0>(matrix);
+    result.coeffRef(1,3) =  cofactor_4x4<MatrixType,3,1>(matrix);
+    result.coeffRef(2,3) = -cofactor_4x4<MatrixType,3,2>(matrix);
+    result.coeffRef(3,3) =  cofactor_4x4<MatrixType,3,3>(matrix);
+    result /= (matrix.col(0).cwiseProduct(result.row(0).transpose())).sum();
+  }
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse<MatrixType, ResultType, 4>
+ : compute_inverse_size4<Architecture::Target, typename MatrixType::Scalar,
+                            MatrixType, ResultType>
+{
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_and_det_with_check<MatrixType, ResultType, 4>
+{
+  EIGEN_DEVICE_FUNC
+  static inline void run(
+    const MatrixType& matrix,
+    const typename MatrixType::RealScalar& absDeterminantThreshold,
+    ResultType& inverse,
+    typename ResultType::Scalar& determinant,
+    bool& invertible
+  )
+  {
+    using std::abs;
+    determinant = matrix.determinant();
+    invertible = abs(determinant) > absDeterminantThreshold;
+    if(invertible) compute_inverse<MatrixType, ResultType>::run(matrix, inverse);
+  }
+};
+
+/*************************
+*** MatrixBase methods ***
+*************************/
+
+} // end namespace internal
+
+namespace internal {
+
+// Specialization for "dense = dense_xpr.inverse()"
+template<typename DstXprType, typename XprType>
+struct Assignment<DstXprType, Inverse<XprType>, internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar>, Dense2Dense>
+{
+  typedef Inverse<XprType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename XprType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    const int Size = EIGEN_PLAIN_ENUM_MIN(XprType::ColsAtCompileTime,DstXprType::ColsAtCompileTime);
+    EIGEN_ONLY_USED_FOR_DEBUG(Size);
+    eigen_assert(( (Size<=1) || (Size>4) || (extract_data(src.nestedExpression())!=extract_data(dst)))
+              && "Aliasing problem detected in inverse(), you need to do inverse().eval() here.");
+
+    typedef typename internal::nested_eval<XprType,XprType::ColsAtCompileTime>::type  ActualXprType;
+    typedef typename internal::remove_all<ActualXprType>::type                        ActualXprTypeCleanded;
+    
+    ActualXprType actual_xpr(src.nestedExpression());
+    
+    compute_inverse<ActualXprTypeCleanded, DstXprType>::run(actual_xpr, dst);
+  }
+};
+
+  
+} // end namespace internal
+
+/** \lu_module
+  *
+  * \returns the matrix inverse of this matrix.
+  *
+  * For small fixed sizes up to 4x4, this method uses cofactors.
+  * In the general case, this method uses class PartialPivLU.
+  *
+  * \note This matrix must be invertible, otherwise the result is undefined. If you need an
+  * invertibility check, do the following:
+  * \li for fixed sizes up to 4x4, use computeInverseAndDetWithCheck().
+  * \li for the general case, use class FullPivLU.
+  *
+  * Example: \include MatrixBase_inverse.cpp
+  * Output: \verbinclude MatrixBase_inverse.out
+  *
+  * \sa computeInverseAndDetWithCheck()
+  */
+template<typename Derived>
+inline const Inverse<Derived> MatrixBase<Derived>::inverse() const
+{
+  EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsInteger,THIS_FUNCTION_IS_NOT_FOR_INTEGER_NUMERIC_TYPES)
+  eigen_assert(rows() == cols());
+  return Inverse<Derived>(derived());
+}
+
+/** \lu_module
+  *
+  * Computation of matrix inverse and determinant, with invertibility check.
+  *
+  * This is only for fixed-size square matrices of size up to 4x4.
+  *
+  * \param inverse Reference to the matrix in which to store the inverse.
+  * \param determinant Reference to the variable in which to store the determinant.
+  * \param invertible Reference to the bool variable in which to store whether the matrix is invertible.
+  * \param absDeterminantThreshold Optional parameter controlling the invertibility check.
+  *                                The matrix will be declared invertible if the absolute value of its
+  *                                determinant is greater than this threshold.
+  *
+  * Example: \include MatrixBase_computeInverseAndDetWithCheck.cpp
+  * Output: \verbinclude MatrixBase_computeInverseAndDetWithCheck.out
+  *
+  * \sa inverse(), computeInverseWithCheck()
+  */
+template<typename Derived>
+template<typename ResultType>
+inline void MatrixBase<Derived>::computeInverseAndDetWithCheck(
+    ResultType& inverse,
+    typename ResultType::Scalar& determinant,
+    bool& invertible,
+    const RealScalar& absDeterminantThreshold
+  ) const
+{
+  // i'd love to put some static assertions there, but SFINAE means that they have no effect...
+  eigen_assert(rows() == cols());
+  // for 2x2, it's worth giving a chance to avoid evaluating.
+  // for larger sizes, evaluating has negligible cost and limits code size.
+  typedef typename internal::conditional<
+    RowsAtCompileTime == 2,
+    typename internal::remove_all<typename internal::nested_eval<Derived, 2>::type>::type,
+    PlainObject
+  >::type MatrixType;
+  internal::compute_inverse_and_det_with_check<MatrixType, ResultType>::run
+    (derived(), absDeterminantThreshold, inverse, determinant, invertible);
+}
+
+/** \lu_module
+  *
+  * Computation of matrix inverse, with invertibility check.
+  *
+  * This is only for fixed-size square matrices of size up to 4x4.
+  *
+  * \param inverse Reference to the matrix in which to store the inverse.
+  * \param invertible Reference to the bool variable in which to store whether the matrix is invertible.
+  * \param absDeterminantThreshold Optional parameter controlling the invertibility check.
+  *                                The matrix will be declared invertible if the absolute value of its
+  *                                determinant is greater than this threshold.
+  *
+  * Example: \include MatrixBase_computeInverseWithCheck.cpp
+  * Output: \verbinclude MatrixBase_computeInverseWithCheck.out
+  *
+  * \sa inverse(), computeInverseAndDetWithCheck()
+  */
+template<typename Derived>
+template<typename ResultType>
+inline void MatrixBase<Derived>::computeInverseWithCheck(
+    ResultType& inverse,
+    bool& invertible,
+    const RealScalar& absDeterminantThreshold
+  ) const
+{
+  RealScalar determinant;
+  // i'd love to put some static assertions there, but SFINAE means that they have no effect...
+  eigen_assert(rows() == cols());
+  computeInverseAndDetWithCheck(inverse,determinant,invertible,absDeterminantThreshold);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_IMPL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU.h
new file mode 100644
index 0000000..d439618
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU.h
@@ -0,0 +1,611 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2006-2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARTIALLU_H
+#define EIGEN_PARTIALLU_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename _MatrixType> struct traits<PartialPivLU<_MatrixType> >
+ : traits<_MatrixType>
+{
+  typedef MatrixXpr XprKind;
+  typedef SolverStorage StorageKind;
+  typedef traits<_MatrixType> BaseTraits;
+  enum {
+    Flags = BaseTraits::Flags & RowMajorBit,
+    CoeffReadCost = Dynamic
+  };
+};
+
+template<typename T,typename Derived>
+struct enable_if_ref;
+// {
+//   typedef Derived type;
+// };
+
+template<typename T,typename Derived>
+struct enable_if_ref<Ref<T>,Derived> {
+  typedef Derived type;
+};
+
+} // end namespace internal
+
+/** \ingroup LU_Module
+  *
+  * \class PartialPivLU
+  *
+  * \brief LU decomposition of a matrix with partial pivoting, and related features
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the LU decomposition
+  *
+  * This class represents a LU decomposition of a \b square \b invertible matrix, with partial pivoting: the matrix A
+  * is decomposed as A = PLU where L is unit-lower-triangular, U is upper-triangular, and P
+  * is a permutation matrix.
+  *
+  * Typically, partial pivoting LU decomposition is only considered numerically stable for square invertible
+  * matrices. Thus LAPACK's dgesv and dgesvx require the matrix to be square and invertible. The present class
+  * does the same. It will assert that the matrix is square, but it won't (actually it can't) check that the
+  * matrix is invertible: it is your task to check that you only use this decomposition on invertible matrices.
+  *
+  * The guaranteed safe alternative, working for all matrices, is the full pivoting LU decomposition, provided
+  * by class FullPivLU.
+  *
+  * This is \b not a rank-revealing LU decomposition. Many features are intentionally absent from this class,
+  * such as rank computation. If you need these features, use class FullPivLU.
+  *
+  * This LU decomposition is suitable to invert invertible matrices. It is what MatrixBase::inverse() uses
+  * in the general case.
+  * On the other hand, it is \b not suitable to determine whether a given matrix is invertible.
+  *
+  * The data of the LU decomposition can be directly accessed through the methods matrixLU(), permutationP().
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::partialPivLu(), MatrixBase::determinant(), MatrixBase::inverse(), MatrixBase::computeInverse(), class FullPivLU
+  */
+template<typename _MatrixType> class PartialPivLU
+  : public SolverBase<PartialPivLU<_MatrixType> >
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    typedef SolverBase<PartialPivLU> Base;
+    EIGEN_GENERIC_PUBLIC_INTERFACE(PartialPivLU)
+    // FIXME StorageIndex defined in EIGEN_GENERIC_PUBLIC_INTERFACE should be int
+    enum {
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef PermutationMatrix<RowsAtCompileTime, MaxRowsAtCompileTime> PermutationType;
+    typedef Transpositions<RowsAtCompileTime, MaxRowsAtCompileTime> TranspositionType;
+    typedef typename MatrixType::PlainObject PlainObject;
+
+    /**
+      * \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via PartialPivLU::compute(const MatrixType&).
+      */
+    PartialPivLU();
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa PartialPivLU()
+      */
+    explicit PartialPivLU(Index size);
+
+    /** Constructor.
+      *
+      * \param matrix the matrix of which to compute the LU decomposition.
+      *
+      * \warning The matrix should have full rank (e.g. if it's square, it should be invertible).
+      * If you need to deal with non-full rank, use class FullPivLU instead.
+      */
+    template<typename InputType>
+    explicit PartialPivLU(const EigenBase<InputType>& matrix);
+
+    /** Constructor for \link InplaceDecomposition inplace decomposition \endlink
+      *
+      * \param matrix the matrix of which to compute the LU decomposition.
+      *
+      * \warning The matrix should have full rank (e.g. if it's square, it should be invertible).
+      * If you need to deal with non-full rank, use class FullPivLU instead.
+      */
+    template<typename InputType>
+    explicit PartialPivLU(EigenBase<InputType>& matrix);
+
+    template<typename InputType>
+    PartialPivLU& compute(const EigenBase<InputType>& matrix) {
+      m_lu = matrix.derived();
+      compute();
+      return *this;
+    }
+
+    /** \returns the LU decomposition matrix: the upper-triangular part is U, the
+      * unit-lower-triangular part is L (at least for square matrices; in the non-square
+      * case, special care is needed, see the documentation of class FullPivLU).
+      *
+      * \sa matrixL(), matrixU()
+      */
+    inline const MatrixType& matrixLU() const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return m_lu;
+    }
+
+    /** \returns the permutation matrix P.
+      */
+    inline const PermutationType& permutationP() const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return m_p;
+    }
+
+    /** This method returns the solution x to the equation Ax=b, where A is the matrix of which
+      * *this is the LU decomposition.
+      *
+      * \param b the right-hand-side of the equation to solve. Can be a vector or a matrix,
+      *          the only requirement in order for the equation to make sense is that
+      *          b.rows()==A.rows(), where A is the matrix of which *this is the LU decomposition.
+      *
+      * \returns the solution.
+      *
+      * Example: \include PartialPivLU_solve.cpp
+      * Output: \verbinclude PartialPivLU_solve.out
+      *
+      * Since this PartialPivLU class assumes anyway that the matrix A is invertible, the solution
+      * theoretically exists and is unique regardless of b.
+      *
+      * \sa TriangularView::solve(), inverse(), computeInverse()
+      */
+    // FIXME this is a copy-paste of the base-class member to add the isInitialized assertion.
+    template<typename Rhs>
+    inline const Solve<PartialPivLU, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return Solve<PartialPivLU, Rhs>(*this, b.derived());
+    }
+
+    /** \returns an estimate of the reciprocal condition number of the matrix of which \c *this is
+        the LU decomposition.
+      */
+    inline RealScalar rcond() const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return internal::rcond_estimate_helper(m_l1_norm, *this);
+    }
+
+    /** \returns the inverse of the matrix of which *this is the LU decomposition.
+      *
+      * \warning The matrix being decomposed here is assumed to be invertible. If you need to check for
+      *          invertibility, use class FullPivLU instead.
+      *
+      * \sa MatrixBase::inverse(), LU::inverse()
+      */
+    inline const Inverse<PartialPivLU> inverse() const
+    {
+      eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+      return Inverse<PartialPivLU>(*this);
+    }
+
+    /** \returns the determinant of the matrix of which
+      * *this is the LU decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the LU decomposition has already been computed.
+      *
+      * \note For fixed-size matrices of size up to 4, MatrixBase::determinant() offers
+      *       optimized paths.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      *
+      * \sa MatrixBase::determinant()
+      */
+    Scalar determinant() const;
+
+    MatrixType reconstructedMatrix() const;
+
+    inline Index rows() const { return m_lu.rows(); }
+    inline Index cols() const { return m_lu.cols(); }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const {
+     /* The decomposition PA = LU can be rewritten as A = P^{-1} L U.
+      * So we proceed as follows:
+      * Step 1: compute c = Pb.
+      * Step 2: replace c by the solution x to Lx = c.
+      * Step 3: replace c by the solution x to Ux = c.
+      */
+
+      eigen_assert(rhs.rows() == m_lu.rows());
+
+      // Step 1
+      dst = permutationP() * rhs;
+
+      // Step 2
+      m_lu.template triangularView<UnitLower>().solveInPlace(dst);
+
+      // Step 3
+      m_lu.template triangularView<Upper>().solveInPlace(dst);
+    }
+
+    template<bool Conjugate, typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl_transposed(const RhsType &rhs, DstType &dst) const {
+     /* The decomposition PA = LU can be rewritten as A = P^{-1} L U.
+      * So we proceed as follows:
+      * Step 1: compute c = Pb.
+      * Step 2: replace c by the solution x to Lx = c.
+      * Step 3: replace c by the solution x to Ux = c.
+      */
+
+      eigen_assert(rhs.rows() == m_lu.cols());
+
+      if (Conjugate) {
+        // Step 1
+        dst = m_lu.template triangularView<Upper>().adjoint().solve(rhs);
+        // Step 2
+        m_lu.template triangularView<UnitLower>().adjoint().solveInPlace(dst);
+      } else {
+        // Step 1
+        dst = m_lu.template triangularView<Upper>().transpose().solve(rhs);
+        // Step 2
+        m_lu.template triangularView<UnitLower>().transpose().solveInPlace(dst);
+      }
+      // Step 3
+      dst = permutationP().transpose() * dst;
+    }
+    #endif
+
+  protected:
+
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    void compute();
+
+    MatrixType m_lu;
+    PermutationType m_p;
+    TranspositionType m_rowsTranspositions;
+    RealScalar m_l1_norm;
+    signed char m_det_p;
+    bool m_isInitialized;
+};
+
+template<typename MatrixType>
+PartialPivLU<MatrixType>::PartialPivLU()
+  : m_lu(),
+    m_p(),
+    m_rowsTranspositions(),
+    m_l1_norm(0),
+    m_det_p(0),
+    m_isInitialized(false)
+{
+}
+
+template<typename MatrixType>
+PartialPivLU<MatrixType>::PartialPivLU(Index size)
+  : m_lu(size, size),
+    m_p(size),
+    m_rowsTranspositions(size),
+    m_l1_norm(0),
+    m_det_p(0),
+    m_isInitialized(false)
+{
+}
+
+template<typename MatrixType>
+template<typename InputType>
+PartialPivLU<MatrixType>::PartialPivLU(const EigenBase<InputType>& matrix)
+  : m_lu(matrix.rows(),matrix.cols()),
+    m_p(matrix.rows()),
+    m_rowsTranspositions(matrix.rows()),
+    m_l1_norm(0),
+    m_det_p(0),
+    m_isInitialized(false)
+{
+  compute(matrix.derived());
+}
+
+template<typename MatrixType>
+template<typename InputType>
+PartialPivLU<MatrixType>::PartialPivLU(EigenBase<InputType>& matrix)
+  : m_lu(matrix.derived()),
+    m_p(matrix.rows()),
+    m_rowsTranspositions(matrix.rows()),
+    m_l1_norm(0),
+    m_det_p(0),
+    m_isInitialized(false)
+{
+  compute();
+}
+
+namespace internal {
+
+/** \internal This is the blocked version of fullpivlu_unblocked() */
+template<typename Scalar, int StorageOrder, typename PivIndex>
+struct partial_lu_impl
+{
+  // FIXME add a stride to Map, so that the following mapping becomes easier,
+  // another option would be to create an expression being able to automatically
+  // warp any Map, Matrix, and Block expressions as a unique type, but since that's exactly
+  // a Map + stride, why not adding a stride to Map, and convenient ctors from a Matrix,
+  // and Block.
+  typedef Map<Matrix<Scalar, Dynamic, Dynamic, StorageOrder> > MapLU;
+  typedef Block<MapLU, Dynamic, Dynamic> MatrixType;
+  typedef Block<MatrixType,Dynamic,Dynamic> BlockType;
+  typedef typename MatrixType::RealScalar RealScalar;
+
+  /** \internal performs the LU decomposition in-place of the matrix \a lu
+    * using an unblocked algorithm.
+    *
+    * In addition, this function returns the row transpositions in the
+    * vector \a row_transpositions which must have a size equal to the number
+    * of columns of the matrix \a lu, and an integer \a nb_transpositions
+    * which returns the actual number of transpositions.
+    *
+    * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
+    */
+  static Index unblocked_lu(MatrixType& lu, PivIndex* row_transpositions, PivIndex& nb_transpositions)
+  {
+    typedef scalar_score_coeff_op<Scalar> Scoring;
+    typedef typename Scoring::result_type Score;
+    const Index rows = lu.rows();
+    const Index cols = lu.cols();
+    const Index size = (std::min)(rows,cols);
+    nb_transpositions = 0;
+    Index first_zero_pivot = -1;
+    for(Index k = 0; k < size; ++k)
+    {
+      Index rrows = rows-k-1;
+      Index rcols = cols-k-1;
+
+      Index row_of_biggest_in_col;
+      Score biggest_in_corner
+        = lu.col(k).tail(rows-k).unaryExpr(Scoring()).maxCoeff(&row_of_biggest_in_col);
+      row_of_biggest_in_col += k;
+
+      row_transpositions[k] = PivIndex(row_of_biggest_in_col);
+
+      if(biggest_in_corner != Score(0))
+      {
+        if(k != row_of_biggest_in_col)
+        {
+          lu.row(k).swap(lu.row(row_of_biggest_in_col));
+          ++nb_transpositions;
+        }
+
+        // FIXME shall we introduce a safe quotient expression in cas 1/lu.coeff(k,k)
+        // overflow but not the actual quotient?
+        lu.col(k).tail(rrows) /= lu.coeff(k,k);
+      }
+      else if(first_zero_pivot==-1)
+      {
+        // the pivot is exactly zero, we record the index of the first pivot which is exactly 0,
+        // and continue the factorization such we still have A = PLU
+        first_zero_pivot = k;
+      }
+
+      if(k<rows-1)
+        lu.bottomRightCorner(rrows,rcols).noalias() -= lu.col(k).tail(rrows) * lu.row(k).tail(rcols);
+    }
+    return first_zero_pivot;
+  }
+
+  /** \internal performs the LU decomposition in-place of the matrix represented
+    * by the variables \a rows, \a cols, \a lu_data, and \a lu_stride using a
+    * recursive, blocked algorithm.
+    *
+    * In addition, this function returns the row transpositions in the
+    * vector \a row_transpositions which must have a size equal to the number
+    * of columns of the matrix \a lu, and an integer \a nb_transpositions
+    * which returns the actual number of transpositions.
+    *
+    * \returns The index of the first pivot which is exactly zero if any, or a negative number otherwise.
+    *
+    * \note This very low level interface using pointers, etc. is to:
+    *   1 - reduce the number of instanciations to the strict minimum
+    *   2 - avoid infinite recursion of the instanciations with Block<Block<Block<...> > >
+    */
+  static Index blocked_lu(Index rows, Index cols, Scalar* lu_data, Index luStride, PivIndex* row_transpositions, PivIndex& nb_transpositions, Index maxBlockSize=256)
+  {
+    MapLU lu1(lu_data,StorageOrder==RowMajor?rows:luStride,StorageOrder==RowMajor?luStride:cols);
+    MatrixType lu(lu1,0,0,rows,cols);
+
+    const Index size = (std::min)(rows,cols);
+
+    // if the matrix is too small, no blocking:
+    if(size<=16)
+    {
+      return unblocked_lu(lu, row_transpositions, nb_transpositions);
+    }
+
+    // automatically adjust the number of subdivisions to the size
+    // of the matrix so that there is enough sub blocks:
+    Index blockSize;
+    {
+      blockSize = size/8;
+      blockSize = (blockSize/16)*16;
+      blockSize = (std::min)((std::max)(blockSize,Index(8)), maxBlockSize);
+    }
+
+    nb_transpositions = 0;
+    Index first_zero_pivot = -1;
+    for(Index k = 0; k < size; k+=blockSize)
+    {
+      Index bs = (std::min)(size-k,blockSize); // actual size of the block
+      Index trows = rows - k - bs; // trailing rows
+      Index tsize = size - k - bs; // trailing size
+
+      // partition the matrix:
+      //                          A00 | A01 | A02
+      // lu  = A_0 | A_1 | A_2 =  A10 | A11 | A12
+      //                          A20 | A21 | A22
+      BlockType A_0(lu,0,0,rows,k);
+      BlockType A_2(lu,0,k+bs,rows,tsize);
+      BlockType A11(lu,k,k,bs,bs);
+      BlockType A12(lu,k,k+bs,bs,tsize);
+      BlockType A21(lu,k+bs,k,trows,bs);
+      BlockType A22(lu,k+bs,k+bs,trows,tsize);
+
+      PivIndex nb_transpositions_in_panel;
+      // recursively call the blocked LU algorithm on [A11^T A21^T]^T
+      // with a very small blocking size:
+      Index ret = blocked_lu(trows+bs, bs, &lu.coeffRef(k,k), luStride,
+                   row_transpositions+k, nb_transpositions_in_panel, 16);
+      if(ret>=0 && first_zero_pivot==-1)
+        first_zero_pivot = k+ret;
+
+      nb_transpositions += nb_transpositions_in_panel;
+      // update permutations and apply them to A_0
+      for(Index i=k; i<k+bs; ++i)
+      {
+        Index piv = (row_transpositions[i] += internal::convert_index<PivIndex>(k));
+        A_0.row(i).swap(A_0.row(piv));
+      }
+
+      if(trows)
+      {
+        // apply permutations to A_2
+        for(Index i=k;i<k+bs; ++i)
+          A_2.row(i).swap(A_2.row(row_transpositions[i]));
+
+        // A12 = A11^-1 A12
+        A11.template triangularView<UnitLower>().solveInPlace(A12);
+
+        A22.noalias() -= A21 * A12;
+      }
+    }
+    return first_zero_pivot;
+  }
+};
+
+/** \internal performs the LU decomposition with partial pivoting in-place.
+  */
+template<typename MatrixType, typename TranspositionType>
+void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, typename TranspositionType::StorageIndex& nb_transpositions)
+{
+  eigen_assert(lu.cols() == row_transpositions.size());
+  eigen_assert((&row_transpositions.coeffRef(1)-&row_transpositions.coeffRef(0)) == 1);
+
+  partial_lu_impl
+    <typename MatrixType::Scalar, MatrixType::Flags&RowMajorBit?RowMajor:ColMajor, typename TranspositionType::StorageIndex>
+    ::blocked_lu(lu.rows(), lu.cols(), &lu.coeffRef(0,0), lu.outerStride(), &row_transpositions.coeffRef(0), nb_transpositions);
+}
+
+} // end namespace internal
+
+template<typename MatrixType>
+void PartialPivLU<MatrixType>::compute()
+{
+  check_template_parameters();
+
+  // the row permutation is stored as int indices, so just to be sure:
+  eigen_assert(m_lu.rows()<NumTraits<int>::highest());
+
+  m_l1_norm = m_lu.cwiseAbs().colwise().sum().maxCoeff();
+
+  eigen_assert(m_lu.rows() == m_lu.cols() && "PartialPivLU is only for square (and moreover invertible) matrices");
+  const Index size = m_lu.rows();
+
+  m_rowsTranspositions.resize(size);
+
+  typename TranspositionType::StorageIndex nb_transpositions;
+  internal::partial_lu_inplace(m_lu, m_rowsTranspositions, nb_transpositions);
+  m_det_p = (nb_transpositions%2) ? -1 : 1;
+
+  m_p = m_rowsTranspositions;
+
+  m_isInitialized = true;
+}
+
+template<typename MatrixType>
+typename PartialPivLU<MatrixType>::Scalar PartialPivLU<MatrixType>::determinant() const
+{
+  eigen_assert(m_isInitialized && "PartialPivLU is not initialized.");
+  return Scalar(m_det_p) * m_lu.diagonal().prod();
+}
+
+/** \returns the matrix represented by the decomposition,
+ * i.e., it returns the product: P^{-1} L U.
+ * This function is provided for debug purpose. */
+template<typename MatrixType>
+MatrixType PartialPivLU<MatrixType>::reconstructedMatrix() const
+{
+  eigen_assert(m_isInitialized && "LU is not initialized.");
+  // LU
+  MatrixType res = m_lu.template triangularView<UnitLower>().toDenseMatrix()
+                 * m_lu.template triangularView<Upper>();
+
+  // P^{-1}(LU)
+  res = m_p.inverse() * res;
+
+  return res;
+}
+
+/***** Implementation details *****************************************************/
+
+namespace internal {
+
+/***** Implementation of inverse() *****************************************************/
+template<typename DstXprType, typename MatrixType>
+struct Assignment<DstXprType, Inverse<PartialPivLU<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename PartialPivLU<MatrixType>::Scalar>, Dense2Dense>
+{
+  typedef PartialPivLU<MatrixType> LuType;
+  typedef Inverse<LuType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename LuType::Scalar> &)
+  {
+    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols()));
+  }
+};
+} // end namespace internal
+
+/******** MatrixBase methods *******/
+
+/** \lu_module
+  *
+  * \return the partial-pivoting LU decomposition of \c *this.
+  *
+  * \sa class PartialPivLU
+  */
+template<typename Derived>
+inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::partialPivLu() const
+{
+  return PartialPivLU<PlainObject>(eval());
+}
+
+/** \lu_module
+  *
+  * Synonym of partialPivLu().
+  *
+  * \return the partial-pivoting LU decomposition of \c *this.
+  *
+  * \sa class PartialPivLU
+  */
+template<typename Derived>
+inline const PartialPivLU<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::lu() const
+{
+  return PartialPivLU<PlainObject>(eval());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARTIALLU_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU_LAPACKE.h
new file mode 100644
index 0000000..755168a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/PartialPivLU_LAPACKE.h
@@ -0,0 +1,83 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *     LU decomposition with partial pivoting based on LAPACKE_?getrf function.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_PARTIALLU_LAPACK_H
+#define EIGEN_PARTIALLU_LAPACK_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_LU_PARTPIV(EIGTYPE, LAPACKE_TYPE, LAPACKE_PREFIX) \
+template<int StorageOrder> \
+struct partial_lu_impl<EIGTYPE, StorageOrder, lapack_int> \
+{ \
+  /* \internal performs the LU decomposition in-place of the matrix represented */ \
+  static lapack_int blocked_lu(Index rows, Index cols, EIGTYPE* lu_data, Index luStride, lapack_int* row_transpositions, lapack_int& nb_transpositions, lapack_int maxBlockSize=256) \
+  { \
+    EIGEN_UNUSED_VARIABLE(maxBlockSize);\
+    lapack_int matrix_order, first_zero_pivot; \
+    lapack_int m, n, lda, *ipiv, info; \
+    EIGTYPE* a; \
+/* Set up parameters for ?getrf */ \
+    matrix_order = StorageOrder==RowMajor ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+    lda = convert_index<lapack_int>(luStride); \
+    a = lu_data; \
+    ipiv = row_transpositions; \
+    m = convert_index<lapack_int>(rows); \
+    n = convert_index<lapack_int>(cols); \
+    nb_transpositions = 0; \
+\
+    info = LAPACKE_##LAPACKE_PREFIX##getrf( matrix_order, m, n, (LAPACKE_TYPE*)a, lda, ipiv ); \
+\
+    for(int i=0;i<m;i++) { ipiv[i]--; if (ipiv[i]!=i) nb_transpositions++; } \
+\
+    eigen_assert(info >= 0); \
+/* something should be done with nb_transpositions */ \
+\
+    first_zero_pivot = info; \
+    return first_zero_pivot; \
+  } \
+};
+
+EIGEN_LAPACKE_LU_PARTPIV(double, double, d)
+EIGEN_LAPACKE_LU_PARTPIV(float, float, s)
+EIGEN_LAPACKE_LU_PARTPIV(dcomplex, lapack_complex_double, z)
+EIGEN_LAPACKE_LU_PARTPIV(scomplex, lapack_complex_float,  c)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARTIALLU_LAPACK_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/arch/Inverse_SSE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/arch/Inverse_SSE.h
new file mode 100644
index 0000000..ebb64a6
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/LU/arch/Inverse_SSE.h
@@ -0,0 +1,338 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2001 Intel Corporation
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// The SSE code for the 4x4 float and double matrix inverse in this file
+// comes from the following Intel's library:
+// http://software.intel.com/en-us/articles/optimized-matrix-library-for-use-with-the-intel-pentiumr-4-processors-sse2-instructions/
+//
+// Here is the respective copyright and license statement:
+//
+//   Copyright (c) 2001 Intel Corporation.
+//
+// Permition is granted to use, copy, distribute and prepare derivative works
+// of this library for any purpose and without fee, provided, that the above
+// copyright notice and this statement appear in all copies.
+// Intel makes no representations about the suitability of this software for
+// any purpose, and specifically disclaims all warranties.
+// See LEGAL.TXT for all the legal information.
+
+#ifndef EIGEN_INVERSE_SSE_H
+#define EIGEN_INVERSE_SSE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_size4<Architecture::SSE, float, MatrixType, ResultType>
+{
+  enum {
+    MatrixAlignment     = traits<MatrixType>::Alignment,
+    ResultAlignment     = traits<ResultType>::Alignment,
+    StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
+  };
+  typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
+  
+  static void run(const MatrixType& mat, ResultType& result)
+  {
+    ActualMatrixType matrix(mat);
+    EIGEN_ALIGN16 const unsigned int _Sign_PNNP[4] = { 0x00000000, 0x80000000, 0x80000000, 0x00000000 };
+
+    // Load the full matrix into registers
+    __m128 _L1 = matrix.template packet<MatrixAlignment>( 0);
+    __m128 _L2 = matrix.template packet<MatrixAlignment>( 4);
+    __m128 _L3 = matrix.template packet<MatrixAlignment>( 8);
+    __m128 _L4 = matrix.template packet<MatrixAlignment>(12);
+
+    // The inverse is calculated using "Divide and Conquer" technique. The
+    // original matrix is divide into four 2x2 sub-matrices. Since each
+    // register holds four matrix element, the smaller matrices are
+    // represented as a registers. Hence we get a better locality of the
+    // calculations.
+
+    __m128 A, B, C, D; // the four sub-matrices
+    if(!StorageOrdersMatch)
+    {
+      A = _mm_unpacklo_ps(_L1, _L2);
+      B = _mm_unpacklo_ps(_L3, _L4);
+      C = _mm_unpackhi_ps(_L1, _L2);
+      D = _mm_unpackhi_ps(_L3, _L4);
+    }
+    else
+    {
+      A = _mm_movelh_ps(_L1, _L2);
+      B = _mm_movehl_ps(_L2, _L1);
+      C = _mm_movelh_ps(_L3, _L4);
+      D = _mm_movehl_ps(_L4, _L3);
+    }
+
+    __m128 iA, iB, iC, iD,                 // partial inverse of the sub-matrices
+            DC, AB;
+    __m128 dA, dB, dC, dD;                 // determinant of the sub-matrices
+    __m128 det, d, d1, d2;
+    __m128 rd;                             // reciprocal of the determinant
+
+    //  AB = A# * B
+    AB = _mm_mul_ps(_mm_shuffle_ps(A,A,0x0F), B);
+    AB = _mm_sub_ps(AB,_mm_mul_ps(_mm_shuffle_ps(A,A,0xA5), _mm_shuffle_ps(B,B,0x4E)));
+    //  DC = D# * C
+    DC = _mm_mul_ps(_mm_shuffle_ps(D,D,0x0F), C);
+    DC = _mm_sub_ps(DC,_mm_mul_ps(_mm_shuffle_ps(D,D,0xA5), _mm_shuffle_ps(C,C,0x4E)));
+
+    //  dA = |A|
+    dA = _mm_mul_ps(_mm_shuffle_ps(A, A, 0x5F),A);
+    dA = _mm_sub_ss(dA, _mm_movehl_ps(dA,dA));
+    //  dB = |B|
+    dB = _mm_mul_ps(_mm_shuffle_ps(B, B, 0x5F),B);
+    dB = _mm_sub_ss(dB, _mm_movehl_ps(dB,dB));
+
+    //  dC = |C|
+    dC = _mm_mul_ps(_mm_shuffle_ps(C, C, 0x5F),C);
+    dC = _mm_sub_ss(dC, _mm_movehl_ps(dC,dC));
+    //  dD = |D|
+    dD = _mm_mul_ps(_mm_shuffle_ps(D, D, 0x5F),D);
+    dD = _mm_sub_ss(dD, _mm_movehl_ps(dD,dD));
+
+    //  d = trace(AB*DC) = trace(A#*B*D#*C)
+    d = _mm_mul_ps(_mm_shuffle_ps(DC,DC,0xD8),AB);
+
+    //  iD = C*A#*B
+    iD = _mm_mul_ps(_mm_shuffle_ps(C,C,0xA0), _mm_movelh_ps(AB,AB));
+    iD = _mm_add_ps(iD,_mm_mul_ps(_mm_shuffle_ps(C,C,0xF5), _mm_movehl_ps(AB,AB)));
+    //  iA = B*D#*C
+    iA = _mm_mul_ps(_mm_shuffle_ps(B,B,0xA0), _mm_movelh_ps(DC,DC));
+    iA = _mm_add_ps(iA,_mm_mul_ps(_mm_shuffle_ps(B,B,0xF5), _mm_movehl_ps(DC,DC)));
+
+    //  d = trace(AB*DC) = trace(A#*B*D#*C) [continue]
+    d  = _mm_add_ps(d, _mm_movehl_ps(d, d));
+    d  = _mm_add_ss(d, _mm_shuffle_ps(d, d, 1));
+    d1 = _mm_mul_ss(dA,dD);
+    d2 = _mm_mul_ss(dB,dC);
+
+    //  iD = D*|A| - C*A#*B
+    iD = _mm_sub_ps(_mm_mul_ps(D,_mm_shuffle_ps(dA,dA,0)), iD);
+
+    //  iA = A*|D| - B*D#*C;
+    iA = _mm_sub_ps(_mm_mul_ps(A,_mm_shuffle_ps(dD,dD,0)), iA);
+
+    //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
+    det = _mm_sub_ss(_mm_add_ss(d1,d2),d);
+    rd  = _mm_div_ss(_mm_set_ss(1.0f), det);
+
+//     #ifdef ZERO_SINGULAR
+//         rd = _mm_and_ps(_mm_cmpneq_ss(det,_mm_setzero_ps()), rd);
+//     #endif
+
+    //  iB = D * (A#B)# = D*B#*A
+    iB = _mm_mul_ps(D, _mm_shuffle_ps(AB,AB,0x33));
+    iB = _mm_sub_ps(iB, _mm_mul_ps(_mm_shuffle_ps(D,D,0xB1), _mm_shuffle_ps(AB,AB,0x66)));
+    //  iC = A * (D#C)# = A*C#*D
+    iC = _mm_mul_ps(A, _mm_shuffle_ps(DC,DC,0x33));
+    iC = _mm_sub_ps(iC, _mm_mul_ps(_mm_shuffle_ps(A,A,0xB1), _mm_shuffle_ps(DC,DC,0x66)));
+
+    rd = _mm_shuffle_ps(rd,rd,0);
+    rd = _mm_xor_ps(rd, _mm_load_ps((float*)_Sign_PNNP));
+
+    //  iB = C*|B| - D*B#*A
+    iB = _mm_sub_ps(_mm_mul_ps(C,_mm_shuffle_ps(dB,dB,0)), iB);
+
+    //  iC = B*|C| - A*C#*D;
+    iC = _mm_sub_ps(_mm_mul_ps(B,_mm_shuffle_ps(dC,dC,0)), iC);
+
+    //  iX = iX / det
+    iA = _mm_mul_ps(rd,iA);
+    iB = _mm_mul_ps(rd,iB);
+    iC = _mm_mul_ps(rd,iC);
+    iD = _mm_mul_ps(rd,iD);
+
+    Index res_stride = result.outerStride();
+    float* res = result.data();
+    pstoret<float, Packet4f, ResultAlignment>(res+0,            _mm_shuffle_ps(iA,iB,0x77));
+    pstoret<float, Packet4f, ResultAlignment>(res+res_stride,   _mm_shuffle_ps(iA,iB,0x22));
+    pstoret<float, Packet4f, ResultAlignment>(res+2*res_stride, _mm_shuffle_ps(iC,iD,0x77));
+    pstoret<float, Packet4f, ResultAlignment>(res+3*res_stride, _mm_shuffle_ps(iC,iD,0x22));
+  }
+
+};
+
+template<typename MatrixType, typename ResultType>
+struct compute_inverse_size4<Architecture::SSE, double, MatrixType, ResultType>
+{
+  enum {
+    MatrixAlignment     = traits<MatrixType>::Alignment,
+    ResultAlignment     = traits<ResultType>::Alignment,
+    StorageOrdersMatch  = (MatrixType::Flags&RowMajorBit) == (ResultType::Flags&RowMajorBit)
+  };
+  typedef typename conditional<(MatrixType::Flags&LinearAccessBit),MatrixType const &,typename MatrixType::PlainObject>::type ActualMatrixType;
+  
+  static void run(const MatrixType& mat, ResultType& result)
+  {
+    ActualMatrixType matrix(mat);
+    const __m128d _Sign_NP = _mm_castsi128_pd(_mm_set_epi32(0x0,0x0,0x80000000,0x0));
+    const __m128d _Sign_PN = _mm_castsi128_pd(_mm_set_epi32(0x80000000,0x0,0x0,0x0));
+
+    // The inverse is calculated using "Divide and Conquer" technique. The
+    // original matrix is divide into four 2x2 sub-matrices. Since each
+    // register of the matrix holds two elements, the smaller matrices are
+    // consisted of two registers. Hence we get a better locality of the
+    // calculations.
+
+    // the four sub-matrices
+    __m128d A1, A2, B1, B2, C1, C2, D1, D2;
+    
+    if(StorageOrdersMatch)
+    {
+      A1 = matrix.template packet<MatrixAlignment>( 0); B1 = matrix.template packet<MatrixAlignment>( 2);
+      A2 = matrix.template packet<MatrixAlignment>( 4); B2 = matrix.template packet<MatrixAlignment>( 6);
+      C1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
+      C2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
+    }
+    else
+    {
+      __m128d tmp;
+      A1 = matrix.template packet<MatrixAlignment>( 0); C1 = matrix.template packet<MatrixAlignment>( 2);
+      A2 = matrix.template packet<MatrixAlignment>( 4); C2 = matrix.template packet<MatrixAlignment>( 6);
+      tmp = A1;
+      A1 = _mm_unpacklo_pd(A1,A2);
+      A2 = _mm_unpackhi_pd(tmp,A2);
+      tmp = C1;
+      C1 = _mm_unpacklo_pd(C1,C2);
+      C2 = _mm_unpackhi_pd(tmp,C2);
+      
+      B1 = matrix.template packet<MatrixAlignment>( 8); D1 = matrix.template packet<MatrixAlignment>(10);
+      B2 = matrix.template packet<MatrixAlignment>(12); D2 = matrix.template packet<MatrixAlignment>(14);
+      tmp = B1;
+      B1 = _mm_unpacklo_pd(B1,B2);
+      B2 = _mm_unpackhi_pd(tmp,B2);
+      tmp = D1;
+      D1 = _mm_unpacklo_pd(D1,D2);
+      D2 = _mm_unpackhi_pd(tmp,D2);
+    }
+    
+    __m128d iA1, iA2, iB1, iB2, iC1, iC2, iD1, iD2,     // partial invese of the sub-matrices
+            DC1, DC2, AB1, AB2;
+    __m128d dA, dB, dC, dD;     // determinant of the sub-matrices
+    __m128d det, d1, d2, rd;
+
+    //  dA = |A|
+    dA = _mm_shuffle_pd(A2, A2, 1);
+    dA = _mm_mul_pd(A1, dA);
+    dA = _mm_sub_sd(dA, _mm_shuffle_pd(dA,dA,3));
+    //  dB = |B|
+    dB = _mm_shuffle_pd(B2, B2, 1);
+    dB = _mm_mul_pd(B1, dB);
+    dB = _mm_sub_sd(dB, _mm_shuffle_pd(dB,dB,3));
+
+    //  AB = A# * B
+    AB1 = _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,3));
+    AB2 = _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,0));
+    AB1 = _mm_sub_pd(AB1, _mm_mul_pd(B2, _mm_shuffle_pd(A1,A1,3)));
+    AB2 = _mm_sub_pd(AB2, _mm_mul_pd(B1, _mm_shuffle_pd(A2,A2,0)));
+
+    //  dC = |C|
+    dC = _mm_shuffle_pd(C2, C2, 1);
+    dC = _mm_mul_pd(C1, dC);
+    dC = _mm_sub_sd(dC, _mm_shuffle_pd(dC,dC,3));
+    //  dD = |D|
+    dD = _mm_shuffle_pd(D2, D2, 1);
+    dD = _mm_mul_pd(D1, dD);
+    dD = _mm_sub_sd(dD, _mm_shuffle_pd(dD,dD,3));
+
+    //  DC = D# * C
+    DC1 = _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,3));
+    DC2 = _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,0));
+    DC1 = _mm_sub_pd(DC1, _mm_mul_pd(C2, _mm_shuffle_pd(D1,D1,3)));
+    DC2 = _mm_sub_pd(DC2, _mm_mul_pd(C1, _mm_shuffle_pd(D2,D2,0)));
+
+    //  rd = trace(AB*DC) = trace(A#*B*D#*C)
+    d1 = _mm_mul_pd(AB1, _mm_shuffle_pd(DC1, DC2, 0));
+    d2 = _mm_mul_pd(AB2, _mm_shuffle_pd(DC1, DC2, 3));
+    rd = _mm_add_pd(d1, d2);
+    rd = _mm_add_sd(rd, _mm_shuffle_pd(rd, rd,3));
+
+    //  iD = C*A#*B
+    iD1 = _mm_mul_pd(AB1, _mm_shuffle_pd(C1,C1,0));
+    iD2 = _mm_mul_pd(AB1, _mm_shuffle_pd(C2,C2,0));
+    iD1 = _mm_add_pd(iD1, _mm_mul_pd(AB2, _mm_shuffle_pd(C1,C1,3)));
+    iD2 = _mm_add_pd(iD2, _mm_mul_pd(AB2, _mm_shuffle_pd(C2,C2,3)));
+
+    //  iA = B*D#*C
+    iA1 = _mm_mul_pd(DC1, _mm_shuffle_pd(B1,B1,0));
+    iA2 = _mm_mul_pd(DC1, _mm_shuffle_pd(B2,B2,0));
+    iA1 = _mm_add_pd(iA1, _mm_mul_pd(DC2, _mm_shuffle_pd(B1,B1,3)));
+    iA2 = _mm_add_pd(iA2, _mm_mul_pd(DC2, _mm_shuffle_pd(B2,B2,3)));
+
+    //  iD = D*|A| - C*A#*B
+    dA = _mm_shuffle_pd(dA,dA,0);
+    iD1 = _mm_sub_pd(_mm_mul_pd(D1, dA), iD1);
+    iD2 = _mm_sub_pd(_mm_mul_pd(D2, dA), iD2);
+
+    //  iA = A*|D| - B*D#*C;
+    dD = _mm_shuffle_pd(dD,dD,0);
+    iA1 = _mm_sub_pd(_mm_mul_pd(A1, dD), iA1);
+    iA2 = _mm_sub_pd(_mm_mul_pd(A2, dD), iA2);
+
+    d1 = _mm_mul_sd(dA, dD);
+    d2 = _mm_mul_sd(dB, dC);
+
+    //  iB = D * (A#B)# = D*B#*A
+    iB1 = _mm_mul_pd(D1, _mm_shuffle_pd(AB2,AB1,1));
+    iB2 = _mm_mul_pd(D2, _mm_shuffle_pd(AB2,AB1,1));
+    iB1 = _mm_sub_pd(iB1, _mm_mul_pd(_mm_shuffle_pd(D1,D1,1), _mm_shuffle_pd(AB2,AB1,2)));
+    iB2 = _mm_sub_pd(iB2, _mm_mul_pd(_mm_shuffle_pd(D2,D2,1), _mm_shuffle_pd(AB2,AB1,2)));
+
+    //  det = |A|*|D| + |B|*|C| - trace(A#*B*D#*C)
+    det = _mm_add_sd(d1, d2);
+    det = _mm_sub_sd(det, rd);
+
+    //  iC = A * (D#C)# = A*C#*D
+    iC1 = _mm_mul_pd(A1, _mm_shuffle_pd(DC2,DC1,1));
+    iC2 = _mm_mul_pd(A2, _mm_shuffle_pd(DC2,DC1,1));
+    iC1 = _mm_sub_pd(iC1, _mm_mul_pd(_mm_shuffle_pd(A1,A1,1), _mm_shuffle_pd(DC2,DC1,2)));
+    iC2 = _mm_sub_pd(iC2, _mm_mul_pd(_mm_shuffle_pd(A2,A2,1), _mm_shuffle_pd(DC2,DC1,2)));
+
+    rd = _mm_div_sd(_mm_set_sd(1.0), det);
+//     #ifdef ZERO_SINGULAR
+//         rd = _mm_and_pd(_mm_cmpneq_sd(det,_mm_setzero_pd()), rd);
+//     #endif
+    rd = _mm_shuffle_pd(rd,rd,0);
+
+    //  iB = C*|B| - D*B#*A
+    dB = _mm_shuffle_pd(dB,dB,0);
+    iB1 = _mm_sub_pd(_mm_mul_pd(C1, dB), iB1);
+    iB2 = _mm_sub_pd(_mm_mul_pd(C2, dB), iB2);
+
+    d1 = _mm_xor_pd(rd, _Sign_PN);
+    d2 = _mm_xor_pd(rd, _Sign_NP);
+
+    //  iC = B*|C| - A*C#*D;
+    dC = _mm_shuffle_pd(dC,dC,0);
+    iC1 = _mm_sub_pd(_mm_mul_pd(B1, dC), iC1);
+    iC2 = _mm_sub_pd(_mm_mul_pd(B2, dC), iC2);
+
+    Index res_stride = result.outerStride();
+    double* res = result.data();
+    pstoret<double, Packet2d, ResultAlignment>(res+0,             _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 3), d1));
+    pstoret<double, Packet2d, ResultAlignment>(res+res_stride,    _mm_mul_pd(_mm_shuffle_pd(iA2, iA1, 0), d2));
+    pstoret<double, Packet2d, ResultAlignment>(res+2,             _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 3), d1));
+    pstoret<double, Packet2d, ResultAlignment>(res+res_stride+2,  _mm_mul_pd(_mm_shuffle_pd(iB2, iB1, 0), d2));
+    pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 3), d1));
+    pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride,  _mm_mul_pd(_mm_shuffle_pd(iC2, iC1, 0), d2));
+    pstoret<double, Packet2d, ResultAlignment>(res+2*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 3), d1));
+    pstoret<double, Packet2d, ResultAlignment>(res+3*res_stride+2,_mm_mul_pd(_mm_shuffle_pd(iD2, iD1, 0), d2));
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_INVERSE_SSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/MetisSupport/MetisSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/MetisSupport/MetisSupport.h
new file mode 100644
index 0000000..4c15304
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/MetisSupport/MetisSupport.h
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+#ifndef METIS_SUPPORT_H
+#define METIS_SUPPORT_H
+
+namespace Eigen {
+/**
+ * Get the fill-reducing ordering from the METIS package
+ * 
+ * If A is the original matrix and Ap is the permuted matrix, 
+ * the fill-reducing permutation is defined as follows :
+ * Row (column) i of A is the matperm(i) row (column) of Ap. 
+ * WARNING: As computed by METIS, this corresponds to the vector iperm (instead of perm)
+ */
+template <typename StorageIndex>
+class MetisOrdering
+{
+public:
+  typedef PermutationMatrix<Dynamic,Dynamic,StorageIndex> PermutationType;
+  typedef Matrix<StorageIndex,Dynamic,1> IndexVector; 
+  
+  template <typename MatrixType>
+  void get_symmetrized_graph(const MatrixType& A)
+  {
+    Index m = A.cols(); 
+    eigen_assert((A.rows() == A.cols()) && "ONLY FOR SQUARED MATRICES");
+    // Get the transpose of the input matrix 
+    MatrixType At = A.transpose(); 
+    // Get the number of nonzeros elements in each row/col of At+A
+    Index TotNz = 0; 
+    IndexVector visited(m); 
+    visited.setConstant(-1); 
+    for (StorageIndex j = 0; j < m; j++)
+    {
+      // Compute the union structure of of A(j,:) and At(j,:)
+      visited(j) = j; // Do not include the diagonal element
+      // Get the nonzeros in row/column j of A
+      for (typename MatrixType::InnerIterator it(A, j); it; ++it)
+      {
+        Index idx = it.index(); // Get the row index (for column major) or column index (for row major)
+        if (visited(idx) != j ) 
+        {
+          visited(idx) = j; 
+          ++TotNz; 
+        }
+      }
+      //Get the nonzeros in row/column j of At
+      for (typename MatrixType::InnerIterator it(At, j); it; ++it)
+      {
+        Index idx = it.index(); 
+        if(visited(idx) != j)
+        {
+          visited(idx) = j; 
+          ++TotNz; 
+        }
+      }
+    }
+    // Reserve place for A + At
+    m_indexPtr.resize(m+1);
+    m_innerIndices.resize(TotNz); 
+
+    // Now compute the real adjacency list of each column/row 
+    visited.setConstant(-1); 
+    StorageIndex CurNz = 0; 
+    for (StorageIndex j = 0; j < m; j++)
+    {
+      m_indexPtr(j) = CurNz; 
+      
+      visited(j) = j; // Do not include the diagonal element
+      // Add the pattern of row/column j of A to A+At
+      for (typename MatrixType::InnerIterator it(A,j); it; ++it)
+      {
+        StorageIndex idx = it.index(); // Get the row index (for column major) or column index (for row major)
+        if (visited(idx) != j ) 
+        {
+          visited(idx) = j; 
+          m_innerIndices(CurNz) = idx; 
+          CurNz++; 
+        }
+      }
+      //Add the pattern of row/column j of At to A+At
+      for (typename MatrixType::InnerIterator it(At, j); it; ++it)
+      {
+        StorageIndex idx = it.index(); 
+        if(visited(idx) != j)
+        {
+          visited(idx) = j; 
+          m_innerIndices(CurNz) = idx; 
+          ++CurNz; 
+        }
+      }
+    }
+    m_indexPtr(m) = CurNz;    
+  }
+  
+  template <typename MatrixType>
+  void operator() (const MatrixType& A, PermutationType& matperm)
+  {
+     StorageIndex m = internal::convert_index<StorageIndex>(A.cols()); // must be StorageIndex, because it is passed by address to METIS
+     IndexVector perm(m),iperm(m); 
+    // First, symmetrize the matrix graph. 
+     get_symmetrized_graph(A); 
+     int output_error;
+     
+     // Call the fill-reducing routine from METIS 
+     output_error = METIS_NodeND(&m, m_indexPtr.data(), m_innerIndices.data(), NULL, NULL, perm.data(), iperm.data());
+     
+    if(output_error != METIS_OK) 
+    {
+      //FIXME The ordering interface should define a class of possible errors 
+     std::cerr << "ERROR WHILE CALLING THE METIS PACKAGE \n"; 
+     return; 
+    }
+    
+    // Get the fill-reducing permutation 
+    //NOTE:  If Ap is the permuted matrix then perm and iperm vectors are defined as follows 
+    // Row (column) i of Ap is the perm(i) row(column) of A, and row (column) i of A is the iperm(i) row(column) of Ap
+    
+     matperm.resize(m);
+     for (int j = 0; j < m; j++)
+       matperm.indices()(iperm(j)) = j;
+   
+  }
+  
+  protected:
+    IndexVector m_indexPtr; // Pointer to the adjacenccy list of each row/column
+    IndexVector m_innerIndices; // Adjacency list 
+};
+
+}// end namespace eigen 
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Amd.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Amd.h
new file mode 100644
index 0000000..f91ecb2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Amd.h
@@ -0,0 +1,445 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+
+/*
+
+NOTE: this routine has been adapted from the CSparse library:
+
+Copyright (c) 2006, Timothy A. Davis.
+http://www.suitesparse.com
+
+CSparse is free software; you can redistribute it and/or
+modify it under the terms of the GNU Lesser General Public
+License as published by the Free Software Foundation; either
+version 2.1 of the License, or (at your option) any later version.
+
+CSparse is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+Lesser General Public License for more details.
+
+You should have received a copy of the GNU Lesser General Public
+License along with this Module; if not, write to the Free Software
+Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301  USA
+
+*/
+
+#include "../Core/util/NonMPL2.h"
+
+#ifndef EIGEN_SPARSE_AMD_H
+#define EIGEN_SPARSE_AMD_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<typename T> inline T amd_flip(const T& i) { return -i-2; }
+template<typename T> inline T amd_unflip(const T& i) { return i<0 ? amd_flip(i) : i; }
+template<typename T0, typename T1> inline bool amd_marked(const T0* w, const T1& j) { return w[j]<0; }
+template<typename T0, typename T1> inline void amd_mark(const T0* w, const T1& j) { return w[j] = amd_flip(w[j]); }
+
+/* clear w */
+template<typename StorageIndex>
+static StorageIndex cs_wclear (StorageIndex mark, StorageIndex lemax, StorageIndex *w, StorageIndex n)
+{
+  StorageIndex k;
+  if(mark < 2 || (mark + lemax < 0))
+  {
+    for(k = 0; k < n; k++)
+      if(w[k] != 0)
+        w[k] = 1;
+    mark = 2;
+  }
+  return (mark);     /* at this point, w[0..n-1] < mark holds */
+}
+
+/* depth-first search and postorder of a tree rooted at node j */
+template<typename StorageIndex>
+StorageIndex cs_tdfs(StorageIndex j, StorageIndex k, StorageIndex *head, const StorageIndex *next, StorageIndex *post, StorageIndex *stack)
+{
+  StorageIndex i, p, top = 0;
+  if(!head || !next || !post || !stack) return (-1);    /* check inputs */
+  stack[0] = j;                 /* place j on the stack */
+  while (top >= 0)                /* while (stack is not empty) */
+  {
+    p = stack[top];           /* p = top of stack */
+    i = head[p];              /* i = youngest child of p */
+    if(i == -1)
+    {
+      top--;                 /* p has no unordered children left */
+      post[k++] = p;        /* node p is the kth postordered node */
+    }
+    else
+    {
+      head[p] = next[i];   /* remove i from children of p */
+      stack[++top] = i;     /* start dfs on child node i */
+    }
+  }
+  return k;
+}
+
+
+/** \internal
+  * \ingroup OrderingMethods_Module 
+  * Approximate minimum degree ordering algorithm.
+  *
+  * \param[in] C the input selfadjoint matrix stored in compressed column major format.
+  * \param[out] perm the permutation P reducing the fill-in of the input matrix \a C
+  *
+  * Note that the input matrix \a C must be complete, that is both the upper and lower parts have to be stored, as well as the diagonal entries.
+  * On exit the values of C are destroyed */
+template<typename Scalar, typename StorageIndex>
+void minimum_degree_ordering(SparseMatrix<Scalar,ColMajor,StorageIndex>& C, PermutationMatrix<Dynamic,Dynamic,StorageIndex>& perm)
+{
+  using std::sqrt;
+  
+  StorageIndex d, dk, dext, lemax = 0, e, elenk, eln, i, j, k, k1,
+                k2, k3, jlast, ln, dense, nzmax, mindeg = 0, nvi, nvj, nvk, mark, wnvi,
+                ok, nel = 0, p, p1, p2, p3, p4, pj, pk, pk1, pk2, pn, q, t, h;
+  
+  StorageIndex n = StorageIndex(C.cols());
+  dense = std::max<StorageIndex> (16, StorageIndex(10 * sqrt(double(n))));   /* find dense threshold */
+  dense = (std::min)(n-2, dense);
+  
+  StorageIndex cnz = StorageIndex(C.nonZeros());
+  perm.resize(n+1);
+  t = cnz + cnz/5 + 2*n;                 /* add elbow room to C */
+  C.resizeNonZeros(t);
+  
+  // get workspace
+  ei_declare_aligned_stack_constructed_variable(StorageIndex,W,8*(n+1),0);
+  StorageIndex* len     = W;
+  StorageIndex* nv      = W +   (n+1);
+  StorageIndex* next    = W + 2*(n+1);
+  StorageIndex* head    = W + 3*(n+1);
+  StorageIndex* elen    = W + 4*(n+1);
+  StorageIndex* degree  = W + 5*(n+1);
+  StorageIndex* w       = W + 6*(n+1);
+  StorageIndex* hhead   = W + 7*(n+1);
+  StorageIndex* last    = perm.indices().data();                              /* use P as workspace for last */
+  
+  /* --- Initialize quotient graph ---------------------------------------- */
+  StorageIndex* Cp = C.outerIndexPtr();
+  StorageIndex* Ci = C.innerIndexPtr();
+  for(k = 0; k < n; k++)
+    len[k] = Cp[k+1] - Cp[k];
+  len[n] = 0;
+  nzmax = t;
+  
+  for(i = 0; i <= n; i++)
+  {
+    head[i]   = -1;                     // degree list i is empty
+    last[i]   = -1;
+    next[i]   = -1;
+    hhead[i]  = -1;                     // hash list i is empty 
+    nv[i]     = 1;                      // node i is just one node
+    w[i]      = 1;                      // node i is alive
+    elen[i]   = 0;                      // Ek of node i is empty
+    degree[i] = len[i];                 // degree of node i
+  }
+  mark = internal::cs_wclear<StorageIndex>(0, 0, w, n);         /* clear w */
+  
+  /* --- Initialize degree lists ------------------------------------------ */
+  for(i = 0; i < n; i++)
+  {
+    bool has_diag = false;
+    for(p = Cp[i]; p<Cp[i+1]; ++p)
+      if(Ci[p]==i)
+      {
+        has_diag = true;
+        break;
+      }
+   
+    d = degree[i];
+    if(d == 1 && has_diag)           /* node i is empty */
+    {
+      elen[i] = -2;                 /* element i is dead */
+      nel++;
+      Cp[i] = -1;                   /* i is a root of assembly tree */
+      w[i] = 0;
+    }
+    else if(d > dense || !has_diag)  /* node i is dense or has no structural diagonal element */
+    {
+      nv[i] = 0;                    /* absorb i into element n */
+      elen[i] = -1;                 /* node i is dead */
+      nel++;
+      Cp[i] = amd_flip (n);
+      nv[n]++;
+    }
+    else
+    {
+      if(head[d] != -1) last[head[d]] = i;
+      next[i] = head[d];           /* put node i in degree list d */
+      head[d] = i;
+    }
+  }
+  
+  elen[n] = -2;                         /* n is a dead element */
+  Cp[n] = -1;                           /* n is a root of assembly tree */
+  w[n] = 0;                             /* n is a dead element */
+  
+  while (nel < n)                         /* while (selecting pivots) do */
+  {
+    /* --- Select node of minimum approximate degree -------------------- */
+    for(k = -1; mindeg < n && (k = head[mindeg]) == -1; mindeg++) {}
+    if(next[k] != -1) last[next[k]] = -1;
+    head[mindeg] = next[k];          /* remove k from degree list */
+    elenk = elen[k];                  /* elenk = |Ek| */
+    nvk = nv[k];                      /* # of nodes k represents */
+    nel += nvk;                        /* nv[k] nodes of A eliminated */
+    
+    /* --- Garbage collection ------------------------------------------- */
+    if(elenk > 0 && cnz + mindeg >= nzmax)
+    {
+      for(j = 0; j < n; j++)
+      {
+        if((p = Cp[j]) >= 0)      /* j is a live node or element */
+        {
+          Cp[j] = Ci[p];          /* save first entry of object */
+          Ci[p] = amd_flip (j);    /* first entry is now amd_flip(j) */
+        }
+      }
+      for(q = 0, p = 0; p < cnz; ) /* scan all of memory */
+      {
+        if((j = amd_flip (Ci[p++])) >= 0)  /* found object j */
+        {
+          Ci[q] = Cp[j];       /* restore first entry of object */
+          Cp[j] = q++;          /* new pointer to object j */
+          for(k3 = 0; k3 < len[j]-1; k3++) Ci[q++] = Ci[p++];
+        }
+      }
+      cnz = q;                       /* Ci[cnz...nzmax-1] now free */
+    }
+    
+    /* --- Construct new element ---------------------------------------- */
+    dk = 0;
+    nv[k] = -nvk;                     /* flag k as in Lk */
+    p = Cp[k];
+    pk1 = (elenk == 0) ? p : cnz;      /* do in place if elen[k] == 0 */
+    pk2 = pk1;
+    for(k1 = 1; k1 <= elenk + 1; k1++)
+    {
+      if(k1 > elenk)
+      {
+        e = k;                     /* search the nodes in k */
+        pj = p;                    /* list of nodes starts at Ci[pj]*/
+        ln = len[k] - elenk;      /* length of list of nodes in k */
+      }
+      else
+      {
+        e = Ci[p++];              /* search the nodes in e */
+        pj = Cp[e];
+        ln = len[e];              /* length of list of nodes in e */
+      }
+      for(k2 = 1; k2 <= ln; k2++)
+      {
+        i = Ci[pj++];
+        if((nvi = nv[i]) <= 0) continue; /* node i dead, or seen */
+        dk += nvi;                 /* degree[Lk] += size of node i */
+        nv[i] = -nvi;             /* negate nv[i] to denote i in Lk*/
+        Ci[pk2++] = i;            /* place i in Lk */
+        if(next[i] != -1) last[next[i]] = last[i];
+        if(last[i] != -1)         /* remove i from degree list */
+        {
+          next[last[i]] = next[i];
+        }
+        else
+        {
+          head[degree[i]] = next[i];
+        }
+      }
+      if(e != k)
+      {
+        Cp[e] = amd_flip (k);      /* absorb e into k */
+        w[e] = 0;                 /* e is now a dead element */
+      }
+    }
+    if(elenk != 0) cnz = pk2;         /* Ci[cnz...nzmax] is free */
+    degree[k] = dk;                   /* external degree of k - |Lk\i| */
+    Cp[k] = pk1;                      /* element k is in Ci[pk1..pk2-1] */
+    len[k] = pk2 - pk1;
+    elen[k] = -2;                     /* k is now an element */
+    
+    /* --- Find set differences ----------------------------------------- */
+    mark = internal::cs_wclear<StorageIndex>(mark, lemax, w, n);  /* clear w if necessary */
+    for(pk = pk1; pk < pk2; pk++)    /* scan 1: find |Le\Lk| */
+    {
+      i = Ci[pk];
+      if((eln = elen[i]) <= 0) continue;/* skip if elen[i] empty */
+      nvi = -nv[i];                      /* nv[i] was negated */
+      wnvi = mark - nvi;
+      for(p = Cp[i]; p <= Cp[i] + eln - 1; p++)  /* scan Ei */
+      {
+        e = Ci[p];
+        if(w[e] >= mark)
+        {
+          w[e] -= nvi;          /* decrement |Le\Lk| */
+        }
+        else if(w[e] != 0)        /* ensure e is a live element */
+        {
+          w[e] = degree[e] + wnvi; /* 1st time e seen in scan 1 */
+        }
+      }
+    }
+    
+    /* --- Degree update ------------------------------------------------ */
+    for(pk = pk1; pk < pk2; pk++)    /* scan2: degree update */
+    {
+      i = Ci[pk];                   /* consider node i in Lk */
+      p1 = Cp[i];
+      p2 = p1 + elen[i] - 1;
+      pn = p1;
+      for(h = 0, d = 0, p = p1; p <= p2; p++)    /* scan Ei */
+      {
+        e = Ci[p];
+        if(w[e] != 0)             /* e is an unabsorbed element */
+        {
+          dext = w[e] - mark;   /* dext = |Le\Lk| */
+          if(dext > 0)
+          {
+            d += dext;         /* sum up the set differences */
+            Ci[pn++] = e;     /* keep e in Ei */
+            h += e;            /* compute the hash of node i */
+          }
+          else
+          {
+            Cp[e] = amd_flip (k);  /* aggressive absorb. e->k */
+            w[e] = 0;             /* e is a dead element */
+          }
+        }
+      }
+      elen[i] = pn - p1 + 1;        /* elen[i] = |Ei| */
+      p3 = pn;
+      p4 = p1 + len[i];
+      for(p = p2 + 1; p < p4; p++) /* prune edges in Ai */
+      {
+        j = Ci[p];
+        if((nvj = nv[j]) <= 0) continue; /* node j dead or in Lk */
+        d += nvj;                  /* degree(i) += |j| */
+        Ci[pn++] = j;             /* place j in node list of i */
+        h += j;                    /* compute hash for node i */
+      }
+      if(d == 0)                     /* check for mass elimination */
+      {
+        Cp[i] = amd_flip (k);      /* absorb i into k */
+        nvi = -nv[i];
+        dk -= nvi;                 /* |Lk| -= |i| */
+        nvk += nvi;                /* |k| += nv[i] */
+        nel += nvi;
+        nv[i] = 0;
+        elen[i] = -1;             /* node i is dead */
+      }
+      else
+      {
+        degree[i] = std::min<StorageIndex> (degree[i], d);   /* update degree(i) */
+        Ci[pn] = Ci[p3];         /* move first node to end */
+        Ci[p3] = Ci[p1];         /* move 1st el. to end of Ei */
+        Ci[p1] = k;               /* add k as 1st element in of Ei */
+        len[i] = pn - p1 + 1;     /* new len of adj. list of node i */
+        h %= n;                    /* finalize hash of i */
+        next[i] = hhead[h];      /* place i in hash bucket */
+        hhead[h] = i;
+        last[i] = h;      /* save hash of i in last[i] */
+      }
+    }                                   /* scan2 is done */
+    degree[k] = dk;                   /* finalize |Lk| */
+    lemax = std::max<StorageIndex>(lemax, dk);
+    mark = internal::cs_wclear<StorageIndex>(mark+lemax, lemax, w, n);    /* clear w */
+    
+    /* --- Supernode detection ------------------------------------------ */
+    for(pk = pk1; pk < pk2; pk++)
+    {
+      i = Ci[pk];
+      if(nv[i] >= 0) continue;         /* skip if i is dead */
+      h = last[i];                      /* scan hash bucket of node i */
+      i = hhead[h];
+      hhead[h] = -1;                    /* hash bucket will be empty */
+      for(; i != -1 && next[i] != -1; i = next[i], mark++)
+      {
+        ln = len[i];
+        eln = elen[i];
+        for(p = Cp[i]+1; p <= Cp[i] + ln-1; p++) w[Ci[p]] = mark;
+        jlast = i;
+        for(j = next[i]; j != -1; ) /* compare i with all j */
+        {
+          ok = (len[j] == ln) && (elen[j] == eln);
+          for(p = Cp[j] + 1; ok && p <= Cp[j] + ln - 1; p++)
+          {
+            if(w[Ci[p]] != mark) ok = 0;    /* compare i and j*/
+          }
+          if(ok)                     /* i and j are identical */
+          {
+            Cp[j] = amd_flip (i);  /* absorb j into i */
+            nv[i] += nv[j];
+            nv[j] = 0;
+            elen[j] = -1;         /* node j is dead */
+            j = next[j];          /* delete j from hash bucket */
+            next[jlast] = j;
+          }
+          else
+          {
+            jlast = j;             /* j and i are different */
+            j = next[j];
+          }
+        }
+      }
+    }
+    
+    /* --- Finalize new element------------------------------------------ */
+    for(p = pk1, pk = pk1; pk < pk2; pk++)   /* finalize Lk */
+    {
+      i = Ci[pk];
+      if((nvi = -nv[i]) <= 0) continue;/* skip if i is dead */
+      nv[i] = nvi;                      /* restore nv[i] */
+      d = degree[i] + dk - nvi;         /* compute external degree(i) */
+      d = std::min<StorageIndex> (d, n - nel - nvi);
+      if(head[d] != -1) last[head[d]] = i;
+      next[i] = head[d];               /* put i back in degree list */
+      last[i] = -1;
+      head[d] = i;
+      mindeg = std::min<StorageIndex> (mindeg, d);       /* find new minimum degree */
+      degree[i] = d;
+      Ci[p++] = i;                      /* place i in Lk */
+    }
+    nv[k] = nvk;                      /* # nodes absorbed into k */
+    if((len[k] = p-pk1) == 0)         /* length of adj list of element k*/
+    {
+      Cp[k] = -1;                   /* k is a root of the tree */
+      w[k] = 0;                     /* k is now a dead element */
+    }
+    if(elenk != 0) cnz = p;           /* free unused space in Lk */
+  }
+  
+  /* --- Postordering ----------------------------------------------------- */
+  for(i = 0; i < n; i++) Cp[i] = amd_flip (Cp[i]);/* fix assembly tree */
+  for(j = 0; j <= n; j++) head[j] = -1;
+  for(j = n; j >= 0; j--)              /* place unordered nodes in lists */
+  {
+    if(nv[j] > 0) continue;          /* skip if j is an element */
+    next[j] = head[Cp[j]];          /* place j in list of its parent */
+    head[Cp[j]] = j;
+  }
+  for(e = n; e >= 0; e--)              /* place elements in lists */
+  {
+    if(nv[e] <= 0) continue;         /* skip unless e is an element */
+    if(Cp[e] != -1)
+    {
+      next[e] = head[Cp[e]];      /* place e in list of its parent */
+      head[Cp[e]] = e;
+    }
+  }
+  for(k = 0, i = 0; i <= n; i++)       /* postorder the assembly tree */
+  {
+    if(Cp[i] == -1) k = internal::cs_tdfs<StorageIndex>(i, k, head, next, perm.indices().data(), w);
+  }
+  
+  perm.indices().conservativeResize(n);
+}
+
+} // namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_AMD_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h
new file mode 100644
index 0000000..da85b4d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Eigen_Colamd.h
@@ -0,0 +1,1843 @@
+// // This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Desire Nuentsa Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// This file is modified from the colamd/symamd library. The copyright is below
+
+//   The authors of the code itself are Stefan I. Larimore and Timothy A.
+//   Davis (davis@cise.ufl.edu), University of Florida.  The algorithm was
+//   developed in collaboration with John Gilbert, Xerox PARC, and Esmond
+//   Ng, Oak Ridge National Laboratory.
+// 
+//     Date:
+// 
+//   September 8, 2003.  Version 2.3.
+// 
+//     Acknowledgements:
+// 
+//   This work was supported by the National Science Foundation, under
+//   grants DMS-9504974 and DMS-9803599.
+// 
+//     Notice:
+// 
+//   Copyright (c) 1998-2003 by the University of Florida.
+//   All Rights Reserved.
+// 
+//   THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+//   EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+// 
+//   Permission is hereby granted to use, copy, modify, and/or distribute
+//   this program, provided that the Copyright, this License, and the
+//   Availability of the original version is retained on all copies and made
+//   accessible to the end-user of any code or package that includes COLAMD
+//   or any modified version of COLAMD. 
+// 
+//     Availability:
+// 
+//   The colamd/symamd library is available at
+// 
+//       http://www.suitesparse.com
+
+  
+#ifndef EIGEN_COLAMD_H
+#define EIGEN_COLAMD_H
+
+namespace internal {
+/* Ensure that debugging is turned off: */
+#ifndef COLAMD_NDEBUG
+#define COLAMD_NDEBUG
+#endif /* NDEBUG */
+/* ========================================================================== */
+/* === Knob and statistics definitions ====================================== */
+/* ========================================================================== */
+
+/* size of the knobs [ ] array.  Only knobs [0..1] are currently used. */
+#define COLAMD_KNOBS 20
+
+/* number of output statistics.  Only stats [0..6] are currently used. */
+#define COLAMD_STATS 20 
+
+/* knobs [0] and stats [0]: dense row knob and output statistic. */
+#define COLAMD_DENSE_ROW 0
+
+/* knobs [1] and stats [1]: dense column knob and output statistic. */
+#define COLAMD_DENSE_COL 1
+
+/* stats [2]: memory defragmentation count output statistic */
+#define COLAMD_DEFRAG_COUNT 2
+
+/* stats [3]: colamd status:  zero OK, > 0 warning or notice, < 0 error */
+#define COLAMD_STATUS 3
+
+/* stats [4..6]: error info, or info on jumbled columns */ 
+#define COLAMD_INFO1 4
+#define COLAMD_INFO2 5
+#define COLAMD_INFO3 6
+
+/* error codes returned in stats [3]: */
+#define COLAMD_OK       (0)
+#define COLAMD_OK_BUT_JUMBLED     (1)
+#define COLAMD_ERROR_A_not_present    (-1)
+#define COLAMD_ERROR_p_not_present    (-2)
+#define COLAMD_ERROR_nrow_negative    (-3)
+#define COLAMD_ERROR_ncol_negative    (-4)
+#define COLAMD_ERROR_nnz_negative   (-5)
+#define COLAMD_ERROR_p0_nonzero     (-6)
+#define COLAMD_ERROR_A_too_small    (-7)
+#define COLAMD_ERROR_col_length_negative  (-8)
+#define COLAMD_ERROR_row_index_out_of_bounds  (-9)
+#define COLAMD_ERROR_out_of_memory    (-10)
+#define COLAMD_ERROR_internal_error   (-999)
+
+/* ========================================================================== */
+/* === Definitions ========================================================== */
+/* ========================================================================== */
+
+#define ONES_COMPLEMENT(r) (-(r)-1)
+
+/* -------------------------------------------------------------------------- */
+
+#define COLAMD_EMPTY (-1)
+
+/* Row and column status */
+#define ALIVE (0)
+#define DEAD  (-1)
+
+/* Column status */
+#define DEAD_PRINCIPAL    (-1)
+#define DEAD_NON_PRINCIPAL  (-2)
+
+/* Macros for row and column status update and checking. */
+#define ROW_IS_DEAD(r)      ROW_IS_MARKED_DEAD (Row[r].shared2.mark)
+#define ROW_IS_MARKED_DEAD(row_mark)  (row_mark < ALIVE)
+#define ROW_IS_ALIVE(r)     (Row [r].shared2.mark >= ALIVE)
+#define COL_IS_DEAD(c)      (Col [c].start < ALIVE)
+#define COL_IS_ALIVE(c)     (Col [c].start >= ALIVE)
+#define COL_IS_DEAD_PRINCIPAL(c)  (Col [c].start == DEAD_PRINCIPAL)
+#define KILL_ROW(r)     { Row [r].shared2.mark = DEAD ; }
+#define KILL_PRINCIPAL_COL(c)   { Col [c].start = DEAD_PRINCIPAL ; }
+#define KILL_NON_PRINCIPAL_COL(c) { Col [c].start = DEAD_NON_PRINCIPAL ; }
+
+/* ========================================================================== */
+/* === Colamd reporting mechanism =========================================== */
+/* ========================================================================== */
+
+// == Row and Column structures ==
+template <typename IndexType>
+struct colamd_col
+{
+  IndexType start ;   /* index for A of first row in this column, or DEAD */
+  /* if column is dead */
+  IndexType length ;  /* number of rows in this column */
+  union
+  {
+    IndexType thickness ; /* number of original columns represented by this */
+    /* col, if the column is alive */
+    IndexType parent ;  /* parent in parent tree super-column structure, if */
+    /* the column is dead */
+  } shared1 ;
+  union
+  {
+    IndexType score ; /* the score used to maintain heap, if col is alive */
+    IndexType order ; /* pivot ordering of this column, if col is dead */
+  } shared2 ;
+  union
+  {
+    IndexType headhash ;  /* head of a hash bucket, if col is at the head of */
+    /* a degree list */
+    IndexType hash ;  /* hash value, if col is not in a degree list */
+    IndexType prev ;  /* previous column in degree list, if col is in a */
+    /* degree list (but not at the head of a degree list) */
+  } shared3 ;
+  union
+  {
+    IndexType degree_next ; /* next column, if col is in a degree list */
+    IndexType hash_next ;   /* next column, if col is in a hash list */
+  } shared4 ;
+  
+};
+ 
+template <typename IndexType>
+struct Colamd_Row
+{
+  IndexType start ;   /* index for A of first col in this row */
+  IndexType length ;  /* number of principal columns in this row */
+  union
+  {
+    IndexType degree ;  /* number of principal & non-principal columns in row */
+    IndexType p ;   /* used as a row pointer in init_rows_cols () */
+  } shared1 ;
+  union
+  {
+    IndexType mark ;  /* for computing set differences and marking dead rows*/
+    IndexType first_column ;/* first column in row (used in garbage collection) */
+  } shared2 ;
+  
+};
+ 
+/* ========================================================================== */
+/* === Colamd recommended memory size ======================================= */
+/* ========================================================================== */
+ 
+/*
+  The recommended length Alen of the array A passed to colamd is given by
+  the COLAMD_RECOMMENDED (nnz, n_row, n_col) macro.  It returns -1 if any
+  argument is negative.  2*nnz space is required for the row and column
+  indices of the matrix. colamd_c (n_col) + colamd_r (n_row) space is
+  required for the Col and Row arrays, respectively, which are internal to
+  colamd.  An additional n_col space is the minimal amount of "elbow room",
+  and nnz/5 more space is recommended for run time efficiency.
+  
+  This macro is not needed when using symamd.
+  
+  Explicit typecast to IndexType added Sept. 23, 2002, COLAMD version 2.2, to avoid
+  gcc -pedantic warning messages.
+*/
+template <typename IndexType>
+inline IndexType colamd_c(IndexType n_col) 
+{ return IndexType( ((n_col) + 1) * sizeof (colamd_col<IndexType>) / sizeof (IndexType) ) ; }
+
+template <typename IndexType>
+inline IndexType  colamd_r(IndexType n_row)
+{ return IndexType(((n_row) + 1) * sizeof (Colamd_Row<IndexType>) / sizeof (IndexType)); }
+
+// Prototypes of non-user callable routines
+template <typename IndexType>
+static IndexType init_rows_cols (IndexType n_row, IndexType n_col, Colamd_Row<IndexType> Row [], colamd_col<IndexType> col [], IndexType A [], IndexType p [], IndexType stats[COLAMD_STATS] ); 
+
+template <typename IndexType>
+static void init_scoring (IndexType n_row, IndexType n_col, Colamd_Row<IndexType> Row [], colamd_col<IndexType> Col [], IndexType A [], IndexType head [], double knobs[COLAMD_KNOBS], IndexType *p_n_row2, IndexType *p_n_col2, IndexType *p_max_deg);
+
+template <typename IndexType>
+static IndexType find_ordering (IndexType n_row, IndexType n_col, IndexType Alen, Colamd_Row<IndexType> Row [], colamd_col<IndexType> Col [], IndexType A [], IndexType head [], IndexType n_col2, IndexType max_deg, IndexType pfree);
+
+template <typename IndexType>
+static void order_children (IndexType n_col, colamd_col<IndexType> Col [], IndexType p []);
+
+template <typename IndexType>
+static void detect_super_cols (colamd_col<IndexType> Col [], IndexType A [], IndexType head [], IndexType row_start, IndexType row_length ) ;
+
+template <typename IndexType>
+static IndexType garbage_collection (IndexType n_row, IndexType n_col, Colamd_Row<IndexType> Row [], colamd_col<IndexType> Col [], IndexType A [], IndexType *pfree) ;
+
+template <typename IndexType>
+static inline  IndexType clear_mark (IndexType n_row, Colamd_Row<IndexType> Row [] ) ;
+
+/* === No debugging ========================================================= */
+
+#define COLAMD_DEBUG0(params) ;
+#define COLAMD_DEBUG1(params) ;
+#define COLAMD_DEBUG2(params) ;
+#define COLAMD_DEBUG3(params) ;
+#define COLAMD_DEBUG4(params) ;
+
+#define COLAMD_ASSERT(expression) ((void) 0)
+
+
+/**
+ * \brief Returns the recommended value of Alen 
+ * 
+ * Returns recommended value of Alen for use by colamd.  
+ * Returns -1 if any input argument is negative.  
+ * The use of this routine or macro is optional.  
+ * Note that the macro uses its arguments   more than once, 
+ * so be careful for side effects, if you pass expressions as arguments to COLAMD_RECOMMENDED.  
+ * 
+ * \param nnz nonzeros in A
+ * \param n_row number of rows in A
+ * \param n_col number of columns in A
+ * \return recommended value of Alen for use by colamd
+ */
+template <typename IndexType>
+inline IndexType colamd_recommended ( IndexType nnz, IndexType n_row, IndexType n_col)
+{
+  if ((nnz) < 0 || (n_row) < 0 || (n_col) < 0)
+    return (-1);
+  else
+    return (2 * (nnz) + colamd_c (n_col) + colamd_r (n_row) + (n_col) + ((nnz) / 5)); 
+}
+
+/**
+ * \brief set default parameters  The use of this routine is optional.
+ * 
+ * Colamd: rows with more than (knobs [COLAMD_DENSE_ROW] * n_col)
+ * entries are removed prior to ordering.  Columns with more than
+ * (knobs [COLAMD_DENSE_COL] * n_row) entries are removed prior to
+ * ordering, and placed last in the output column ordering. 
+ *
+ * COLAMD_DENSE_ROW and COLAMD_DENSE_COL are defined as 0 and 1,
+ * respectively, in colamd.h.  Default values of these two knobs
+ * are both 0.5.  Currently, only knobs [0] and knobs [1] are
+ * used, but future versions may use more knobs.  If so, they will
+ * be properly set to their defaults by the future version of
+ * colamd_set_defaults, so that the code that calls colamd will
+ * not need to change, assuming that you either use
+ * colamd_set_defaults, or pass a (double *) NULL pointer as the
+ * knobs array to colamd or symamd.
+ * 
+ * \param knobs parameter settings for colamd
+ */
+
+static inline void colamd_set_defaults(double knobs[COLAMD_KNOBS])
+{
+  /* === Local variables ================================================== */
+  
+  int i ;
+
+  if (!knobs)
+  {
+    return ;      /* no knobs to initialize */
+  }
+  for (i = 0 ; i < COLAMD_KNOBS ; i++)
+  {
+    knobs [i] = 0 ;
+  }
+  knobs [COLAMD_DENSE_ROW] = 0.5 ;  /* ignore rows over 50% dense */
+  knobs [COLAMD_DENSE_COL] = 0.5 ;  /* ignore columns over 50% dense */
+}
+
+/** 
+ * \brief  Computes a column ordering using the column approximate minimum degree ordering
+ * 
+ * Computes a column ordering (Q) of A such that P(AQ)=LU or
+ * (AQ)'AQ=LL' have less fill-in and require fewer floating point
+ * operations than factorizing the unpermuted matrix A or A'A,
+ * respectively.
+ * 
+ * 
+ * \param n_row number of rows in A
+ * \param n_col number of columns in A
+ * \param Alen, size of the array A
+ * \param A row indices of the matrix, of size ALen
+ * \param p column pointers of A, of size n_col+1
+ * \param knobs parameter settings for colamd
+ * \param stats colamd output statistics and error codes
+ */
+template <typename IndexType>
+static bool colamd(IndexType n_row, IndexType n_col, IndexType Alen, IndexType *A, IndexType *p, double knobs[COLAMD_KNOBS], IndexType stats[COLAMD_STATS])
+{
+  /* === Local variables ================================================== */
+  
+  IndexType i ;     /* loop index */
+  IndexType nnz ;     /* nonzeros in A */
+  IndexType Row_size ;    /* size of Row [], in integers */
+  IndexType Col_size ;    /* size of Col [], in integers */
+  IndexType need ;      /* minimum required length of A */
+  Colamd_Row<IndexType> *Row ;   /* pointer into A of Row [0..n_row] array */
+  colamd_col<IndexType> *Col ;   /* pointer into A of Col [0..n_col] array */
+  IndexType n_col2 ;    /* number of non-dense, non-empty columns */
+  IndexType n_row2 ;    /* number of non-dense, non-empty rows */
+  IndexType ngarbage ;    /* number of garbage collections performed */
+  IndexType max_deg ;   /* maximum row degree */
+  double default_knobs [COLAMD_KNOBS] ; /* default knobs array */
+  
+  
+  /* === Check the input arguments ======================================== */
+  
+  if (!stats)
+  {
+    COLAMD_DEBUG0 (("colamd: stats not present\n")) ;
+    return (false) ;
+  }
+  for (i = 0 ; i < COLAMD_STATS ; i++)
+  {
+    stats [i] = 0 ;
+  }
+  stats [COLAMD_STATUS] = COLAMD_OK ;
+  stats [COLAMD_INFO1] = -1 ;
+  stats [COLAMD_INFO2] = -1 ;
+  
+  if (!A)   /* A is not present */
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_A_not_present ;
+    COLAMD_DEBUG0 (("colamd: A not present\n")) ;
+    return (false) ;
+  }
+  
+  if (!p)   /* p is not present */
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_p_not_present ;
+    COLAMD_DEBUG0 (("colamd: p not present\n")) ;
+    return (false) ;
+  }
+  
+  if (n_row < 0)  /* n_row must be >= 0 */
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_nrow_negative ;
+    stats [COLAMD_INFO1] = n_row ;
+    COLAMD_DEBUG0 (("colamd: nrow negative %d\n", n_row)) ;
+    return (false) ;
+  }
+  
+  if (n_col < 0)  /* n_col must be >= 0 */
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_ncol_negative ;
+    stats [COLAMD_INFO1] = n_col ;
+    COLAMD_DEBUG0 (("colamd: ncol negative %d\n", n_col)) ;
+    return (false) ;
+  }
+  
+  nnz = p [n_col] ;
+  if (nnz < 0)  /* nnz must be >= 0 */
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_nnz_negative ;
+    stats [COLAMD_INFO1] = nnz ;
+    COLAMD_DEBUG0 (("colamd: number of entries negative %d\n", nnz)) ;
+    return (false) ;
+  }
+  
+  if (p [0] != 0)
+  {
+    stats [COLAMD_STATUS] = COLAMD_ERROR_p0_nonzero ;
+    stats [COLAMD_INFO1] = p [0] ;
+    COLAMD_DEBUG0 (("colamd: p[0] not zero %d\n", p [0])) ;
+    return (false) ;
+  }
+  
+  /* === If no knobs, set default knobs =================================== */
+  
+  if (!knobs)
+  {
+    colamd_set_defaults (default_knobs) ;
+    knobs = default_knobs ;
+  }
+  
+  /* === Allocate the Row and Col arrays from array A ===================== */
+  
+  Col_size = colamd_c (n_col) ;
+  Row_size = colamd_r (n_row) ;
+  need = 2*nnz + n_col + Col_size + Row_size ;
+  
+  if (need > Alen)
+  {
+    /* not enough space in array A to perform the ordering */
+    stats [COLAMD_STATUS] = COLAMD_ERROR_A_too_small ;
+    stats [COLAMD_INFO1] = need ;
+    stats [COLAMD_INFO2] = Alen ;
+    COLAMD_DEBUG0 (("colamd: Need Alen >= %d, given only Alen = %d\n", need,Alen));
+    return (false) ;
+  }
+  
+  Alen -= Col_size + Row_size ;
+  Col = (colamd_col<IndexType> *) &A [Alen] ;
+  Row = (Colamd_Row<IndexType> *) &A [Alen + Col_size] ;
+
+  /* === Construct the row and column data structures ===================== */
+  
+  if (!Eigen::internal::init_rows_cols (n_row, n_col, Row, Col, A, p, stats))
+  {
+    /* input matrix is invalid */
+    COLAMD_DEBUG0 (("colamd: Matrix invalid\n")) ;
+    return (false) ;
+  }
+  
+  /* === Initialize scores, kill dense rows/columns ======================= */
+
+  Eigen::internal::init_scoring (n_row, n_col, Row, Col, A, p, knobs,
+		&n_row2, &n_col2, &max_deg) ;
+  
+  /* === Order the supercolumns =========================================== */
+  
+  ngarbage = Eigen::internal::find_ordering (n_row, n_col, Alen, Row, Col, A, p,
+			    n_col2, max_deg, 2*nnz) ;
+  
+  /* === Order the non-principal columns ================================== */
+  
+  Eigen::internal::order_children (n_col, Col, p) ;
+  
+  /* === Return statistics in stats ======================================= */
+  
+  stats [COLAMD_DENSE_ROW] = n_row - n_row2 ;
+  stats [COLAMD_DENSE_COL] = n_col - n_col2 ;
+  stats [COLAMD_DEFRAG_COUNT] = ngarbage ;
+  COLAMD_DEBUG0 (("colamd: done.\n")) ; 
+  return (true) ;
+}
+
+/* ========================================================================== */
+/* === NON-USER-CALLABLE ROUTINES: ========================================== */
+/* ========================================================================== */
+
+/* There are no user-callable routines beyond this point in the file */
+
+
+/* ========================================================================== */
+/* === init_rows_cols ======================================================= */
+/* ========================================================================== */
+
+/*
+  Takes the column form of the matrix in A and creates the row form of the
+  matrix.  Also, row and column attributes are stored in the Col and Row
+  structs.  If the columns are un-sorted or contain duplicate row indices,
+  this routine will also sort and remove duplicate row indices from the
+  column form of the matrix.  Returns false if the matrix is invalid,
+  true otherwise.  Not user-callable.
+*/
+template <typename IndexType>
+static IndexType init_rows_cols  /* returns true if OK, or false otherwise */
+  (
+    /* === Parameters ======================================================= */
+
+    IndexType n_row,      /* number of rows of A */
+    IndexType n_col,      /* number of columns of A */
+    Colamd_Row<IndexType> Row [],    /* of size n_row+1 */
+    colamd_col<IndexType> Col [],    /* of size n_col+1 */
+    IndexType A [],     /* row indices of A, of size Alen */
+    IndexType p [],     /* pointers to columns in A, of size n_col+1 */
+    IndexType stats [COLAMD_STATS]  /* colamd statistics */ 
+    )
+{
+  /* === Local variables ================================================== */
+
+  IndexType col ;     /* a column index */
+  IndexType row ;     /* a row index */
+  IndexType *cp ;     /* a column pointer */
+  IndexType *cp_end ;   /* a pointer to the end of a column */
+  IndexType *rp ;     /* a row pointer */
+  IndexType *rp_end ;   /* a pointer to the end of a row */
+  IndexType last_row ;    /* previous row */
+
+  /* === Initialize columns, and check column pointers ==================== */
+
+  for (col = 0 ; col < n_col ; col++)
+  {
+    Col [col].start = p [col] ;
+    Col [col].length = p [col+1] - p [col] ;
+
+    if ((Col [col].length) < 0) // extra parentheses to work-around gcc bug 10200
+    {
+      /* column pointers must be non-decreasing */
+      stats [COLAMD_STATUS] = COLAMD_ERROR_col_length_negative ;
+      stats [COLAMD_INFO1] = col ;
+      stats [COLAMD_INFO2] = Col [col].length ;
+      COLAMD_DEBUG0 (("colamd: col %d length %d < 0\n", col, Col [col].length)) ;
+      return (false) ;
+    }
+
+    Col [col].shared1.thickness = 1 ;
+    Col [col].shared2.score = 0 ;
+    Col [col].shared3.prev = COLAMD_EMPTY ;
+    Col [col].shared4.degree_next = COLAMD_EMPTY ;
+  }
+
+  /* p [0..n_col] no longer needed, used as "head" in subsequent routines */
+
+  /* === Scan columns, compute row degrees, and check row indices ========= */
+
+  stats [COLAMD_INFO3] = 0 ;  /* number of duplicate or unsorted row indices*/
+
+  for (row = 0 ; row < n_row ; row++)
+  {
+    Row [row].length = 0 ;
+    Row [row].shared2.mark = -1 ;
+  }
+
+  for (col = 0 ; col < n_col ; col++)
+  {
+    last_row = -1 ;
+
+    cp = &A [p [col]] ;
+    cp_end = &A [p [col+1]] ;
+
+    while (cp < cp_end)
+    {
+      row = *cp++ ;
+
+      /* make sure row indices within range */
+      if (row < 0 || row >= n_row)
+      {
+	stats [COLAMD_STATUS] = COLAMD_ERROR_row_index_out_of_bounds ;
+	stats [COLAMD_INFO1] = col ;
+	stats [COLAMD_INFO2] = row ;
+	stats [COLAMD_INFO3] = n_row ;
+	COLAMD_DEBUG0 (("colamd: row %d col %d out of bounds\n", row, col)) ;
+	return (false) ;
+      }
+
+      if (row <= last_row || Row [row].shared2.mark == col)
+      {
+	/* row index are unsorted or repeated (or both), thus col */
+	/* is jumbled.  This is a notice, not an error condition. */
+	stats [COLAMD_STATUS] = COLAMD_OK_BUT_JUMBLED ;
+	stats [COLAMD_INFO1] = col ;
+	stats [COLAMD_INFO2] = row ;
+	(stats [COLAMD_INFO3]) ++ ;
+	COLAMD_DEBUG1 (("colamd: row %d col %d unsorted/duplicate\n",row,col));
+      }
+
+      if (Row [row].shared2.mark != col)
+      {
+	Row [row].length++ ;
+      }
+      else
+      {
+	/* this is a repeated entry in the column, */
+	/* it will be removed */
+	Col [col].length-- ;
+      }
+
+      /* mark the row as having been seen in this column */
+      Row [row].shared2.mark = col ;
+
+      last_row = row ;
+    }
+  }
+
+  /* === Compute row pointers ============================================= */
+
+  /* row form of the matrix starts directly after the column */
+  /* form of matrix in A */
+  Row [0].start = p [n_col] ;
+  Row [0].shared1.p = Row [0].start ;
+  Row [0].shared2.mark = -1 ;
+  for (row = 1 ; row < n_row ; row++)
+  {
+    Row [row].start = Row [row-1].start + Row [row-1].length ;
+    Row [row].shared1.p = Row [row].start ;
+    Row [row].shared2.mark = -1 ;
+  }
+
+  /* === Create row form ================================================== */
+
+  if (stats [COLAMD_STATUS] == COLAMD_OK_BUT_JUMBLED)
+  {
+    /* if cols jumbled, watch for repeated row indices */
+    for (col = 0 ; col < n_col ; col++)
+    {
+      cp = &A [p [col]] ;
+      cp_end = &A [p [col+1]] ;
+      while (cp < cp_end)
+      {
+	row = *cp++ ;
+	if (Row [row].shared2.mark != col)
+	{
+	  A [(Row [row].shared1.p)++] = col ;
+	  Row [row].shared2.mark = col ;
+	}
+      }
+    }
+  }
+  else
+  {
+    /* if cols not jumbled, we don't need the mark (this is faster) */
+    for (col = 0 ; col < n_col ; col++)
+    {
+      cp = &A [p [col]] ;
+      cp_end = &A [p [col+1]] ;
+      while (cp < cp_end)
+      {
+	A [(Row [*cp++].shared1.p)++] = col ;
+      }
+    }
+  }
+
+  /* === Clear the row marks and set row degrees ========================== */
+
+  for (row = 0 ; row < n_row ; row++)
+  {
+    Row [row].shared2.mark = 0 ;
+    Row [row].shared1.degree = Row [row].length ;
+  }
+
+  /* === See if we need to re-create columns ============================== */
+
+  if (stats [COLAMD_STATUS] == COLAMD_OK_BUT_JUMBLED)
+  {
+    COLAMD_DEBUG0 (("colamd: reconstructing column form, matrix jumbled\n")) ;
+
+
+    /* === Compute col pointers ========================================= */
+
+    /* col form of the matrix starts at A [0]. */
+    /* Note, we may have a gap between the col form and the row */
+    /* form if there were duplicate entries, if so, it will be */
+    /* removed upon the first garbage collection */
+    Col [0].start = 0 ;
+    p [0] = Col [0].start ;
+    for (col = 1 ; col < n_col ; col++)
+    {
+      /* note that the lengths here are for pruned columns, i.e. */
+      /* no duplicate row indices will exist for these columns */
+      Col [col].start = Col [col-1].start + Col [col-1].length ;
+      p [col] = Col [col].start ;
+    }
+
+    /* === Re-create col form =========================================== */
+
+    for (row = 0 ; row < n_row ; row++)
+    {
+      rp = &A [Row [row].start] ;
+      rp_end = rp + Row [row].length ;
+      while (rp < rp_end)
+      {
+	A [(p [*rp++])++] = row ;
+      }
+    }
+  }
+
+  /* === Done.  Matrix is not (or no longer) jumbled ====================== */
+
+  return (true) ;
+}
+
+
+/* ========================================================================== */
+/* === init_scoring ========================================================= */
+/* ========================================================================== */
+
+/*
+  Kills dense or empty columns and rows, calculates an initial score for
+  each column, and places all columns in the degree lists.  Not user-callable.
+*/
+template <typename IndexType>
+static void init_scoring
+  (
+    /* === Parameters ======================================================= */
+
+    IndexType n_row,      /* number of rows of A */
+    IndexType n_col,      /* number of columns of A */
+    Colamd_Row<IndexType> Row [],    /* of size n_row+1 */
+    colamd_col<IndexType> Col [],    /* of size n_col+1 */
+    IndexType A [],     /* column form and row form of A */
+    IndexType head [],    /* of size n_col+1 */
+    double knobs [COLAMD_KNOBS],/* parameters */
+    IndexType *p_n_row2,    /* number of non-dense, non-empty rows */
+    IndexType *p_n_col2,    /* number of non-dense, non-empty columns */
+    IndexType *p_max_deg    /* maximum row degree */
+    )
+{
+  /* === Local variables ================================================== */
+
+  IndexType c ;     /* a column index */
+  IndexType r, row ;    /* a row index */
+  IndexType *cp ;     /* a column pointer */
+  IndexType deg ;     /* degree of a row or column */
+  IndexType *cp_end ;   /* a pointer to the end of a column */
+  IndexType *new_cp ;   /* new column pointer */
+  IndexType col_length ;    /* length of pruned column */
+  IndexType score ;     /* current column score */
+  IndexType n_col2 ;    /* number of non-dense, non-empty columns */
+  IndexType n_row2 ;    /* number of non-dense, non-empty rows */
+  IndexType dense_row_count ; /* remove rows with more entries than this */
+  IndexType dense_col_count ; /* remove cols with more entries than this */
+  IndexType min_score ;   /* smallest column score */
+  IndexType max_deg ;   /* maximum row degree */
+  IndexType next_col ;    /* Used to add to degree list.*/
+
+
+  /* === Extract knobs ==================================================== */
+
+  dense_row_count = numext::maxi(IndexType(0), numext::mini(IndexType(knobs [COLAMD_DENSE_ROW] * n_col), n_col)) ;
+  dense_col_count = numext::maxi(IndexType(0), numext::mini(IndexType(knobs [COLAMD_DENSE_COL] * n_row), n_row)) ;
+  COLAMD_DEBUG1 (("colamd: densecount: %d %d\n", dense_row_count, dense_col_count)) ;
+  max_deg = 0 ;
+  n_col2 = n_col ;
+  n_row2 = n_row ;
+
+  /* === Kill empty columns =============================================== */
+
+  /* Put the empty columns at the end in their natural order, so that LU */
+  /* factorization can proceed as far as possible. */
+  for (c = n_col-1 ; c >= 0 ; c--)
+  {
+    deg = Col [c].length ;
+    if (deg == 0)
+    {
+      /* this is a empty column, kill and order it last */
+      Col [c].shared2.order = --n_col2 ;
+      KILL_PRINCIPAL_COL (c) ;
+    }
+  }
+  COLAMD_DEBUG1 (("colamd: null columns killed: %d\n", n_col - n_col2)) ;
+
+  /* === Kill dense columns =============================================== */
+
+  /* Put the dense columns at the end, in their natural order */
+  for (c = n_col-1 ; c >= 0 ; c--)
+  {
+    /* skip any dead columns */
+    if (COL_IS_DEAD (c))
+    {
+      continue ;
+    }
+    deg = Col [c].length ;
+    if (deg > dense_col_count)
+    {
+      /* this is a dense column, kill and order it last */
+      Col [c].shared2.order = --n_col2 ;
+      /* decrement the row degrees */
+      cp = &A [Col [c].start] ;
+      cp_end = cp + Col [c].length ;
+      while (cp < cp_end)
+      {
+	Row [*cp++].shared1.degree-- ;
+      }
+      KILL_PRINCIPAL_COL (c) ;
+    }
+  }
+  COLAMD_DEBUG1 (("colamd: Dense and null columns killed: %d\n", n_col - n_col2)) ;
+
+  /* === Kill dense and empty rows ======================================== */
+
+  for (r = 0 ; r < n_row ; r++)
+  {
+    deg = Row [r].shared1.degree ;
+    COLAMD_ASSERT (deg >= 0 && deg <= n_col) ;
+    if (deg > dense_row_count || deg == 0)
+    {
+      /* kill a dense or empty row */
+      KILL_ROW (r) ;
+      --n_row2 ;
+    }
+    else
+    {
+      /* keep track of max degree of remaining rows */
+      max_deg = numext::maxi(max_deg, deg) ;
+    }
+  }
+  COLAMD_DEBUG1 (("colamd: Dense and null rows killed: %d\n", n_row - n_row2)) ;
+
+  /* === Compute initial column scores ==================================== */
+
+  /* At this point the row degrees are accurate.  They reflect the number */
+  /* of "live" (non-dense) columns in each row.  No empty rows exist. */
+  /* Some "live" columns may contain only dead rows, however.  These are */
+  /* pruned in the code below. */
+
+  /* now find the initial matlab score for each column */
+  for (c = n_col-1 ; c >= 0 ; c--)
+  {
+    /* skip dead column */
+    if (COL_IS_DEAD (c))
+    {
+      continue ;
+    }
+    score = 0 ;
+    cp = &A [Col [c].start] ;
+    new_cp = cp ;
+    cp_end = cp + Col [c].length ;
+    while (cp < cp_end)
+    {
+      /* get a row */
+      row = *cp++ ;
+      /* skip if dead */
+      if (ROW_IS_DEAD (row))
+      {
+	continue ;
+      }
+      /* compact the column */
+      *new_cp++ = row ;
+      /* add row's external degree */
+      score += Row [row].shared1.degree - 1 ;
+      /* guard against integer overflow */
+      score = numext::mini(score, n_col) ;
+    }
+    /* determine pruned column length */
+    col_length = (IndexType) (new_cp - &A [Col [c].start]) ;
+    if (col_length == 0)
+    {
+      /* a newly-made null column (all rows in this col are "dense" */
+      /* and have already been killed) */
+      COLAMD_DEBUG2 (("Newly null killed: %d\n", c)) ;
+      Col [c].shared2.order = --n_col2 ;
+      KILL_PRINCIPAL_COL (c) ;
+    }
+    else
+    {
+      /* set column length and set score */
+      COLAMD_ASSERT (score >= 0) ;
+      COLAMD_ASSERT (score <= n_col) ;
+      Col [c].length = col_length ;
+      Col [c].shared2.score = score ;
+    }
+  }
+  COLAMD_DEBUG1 (("colamd: Dense, null, and newly-null columns killed: %d\n",
+		  n_col-n_col2)) ;
+
+  /* At this point, all empty rows and columns are dead.  All live columns */
+  /* are "clean" (containing no dead rows) and simplicial (no supercolumns */
+  /* yet).  Rows may contain dead columns, but all live rows contain at */
+  /* least one live column. */
+
+  /* === Initialize degree lists ========================================== */
+
+
+  /* clear the hash buckets */
+  for (c = 0 ; c <= n_col ; c++)
+  {
+    head [c] = COLAMD_EMPTY ;
+  }
+  min_score = n_col ;
+  /* place in reverse order, so low column indices are at the front */
+  /* of the lists.  This is to encourage natural tie-breaking */
+  for (c = n_col-1 ; c >= 0 ; c--)
+  {
+    /* only add principal columns to degree lists */
+    if (COL_IS_ALIVE (c))
+    {
+      COLAMD_DEBUG4 (("place %d score %d minscore %d ncol %d\n",
+		      c, Col [c].shared2.score, min_score, n_col)) ;
+
+      /* === Add columns score to DList =============================== */
+
+      score = Col [c].shared2.score ;
+
+      COLAMD_ASSERT (min_score >= 0) ;
+      COLAMD_ASSERT (min_score <= n_col) ;
+      COLAMD_ASSERT (score >= 0) ;
+      COLAMD_ASSERT (score <= n_col) ;
+      COLAMD_ASSERT (head [score] >= COLAMD_EMPTY) ;
+
+      /* now add this column to dList at proper score location */
+      next_col = head [score] ;
+      Col [c].shared3.prev = COLAMD_EMPTY ;
+      Col [c].shared4.degree_next = next_col ;
+
+      /* if there already was a column with the same score, set its */
+      /* previous pointer to this new column */
+      if (next_col != COLAMD_EMPTY)
+      {
+	Col [next_col].shared3.prev = c ;
+      }
+      head [score] = c ;
+
+      /* see if this score is less than current min */
+      min_score = numext::mini(min_score, score) ;
+
+
+    }
+  }
+
+
+  /* === Return number of remaining columns, and max row degree =========== */
+
+  *p_n_col2 = n_col2 ;
+  *p_n_row2 = n_row2 ;
+  *p_max_deg = max_deg ;
+}
+
+
+/* ========================================================================== */
+/* === find_ordering ======================================================== */
+/* ========================================================================== */
+
+/*
+  Order the principal columns of the supercolumn form of the matrix
+  (no supercolumns on input).  Uses a minimum approximate column minimum
+  degree ordering method.  Not user-callable.
+*/
+template <typename IndexType>
+static IndexType find_ordering /* return the number of garbage collections */
+  (
+    /* === Parameters ======================================================= */
+
+    IndexType n_row,      /* number of rows of A */
+    IndexType n_col,      /* number of columns of A */
+    IndexType Alen,     /* size of A, 2*nnz + n_col or larger */
+    Colamd_Row<IndexType> Row [],    /* of size n_row+1 */
+    colamd_col<IndexType> Col [],    /* of size n_col+1 */
+    IndexType A [],     /* column form and row form of A */
+    IndexType head [],    /* of size n_col+1 */
+    IndexType n_col2,     /* Remaining columns to order */
+    IndexType max_deg,    /* Maximum row degree */
+    IndexType pfree     /* index of first free slot (2*nnz on entry) */
+    )
+{
+  /* === Local variables ================================================== */
+
+  IndexType k ;     /* current pivot ordering step */
+  IndexType pivot_col ;   /* current pivot column */
+  IndexType *cp ;     /* a column pointer */
+  IndexType *rp ;     /* a row pointer */
+  IndexType pivot_row ;   /* current pivot row */
+  IndexType *new_cp ;   /* modified column pointer */
+  IndexType *new_rp ;   /* modified row pointer */
+  IndexType pivot_row_start ; /* pointer to start of pivot row */
+  IndexType pivot_row_degree ;  /* number of columns in pivot row */
+  IndexType pivot_row_length ;  /* number of supercolumns in pivot row */
+  IndexType pivot_col_score ; /* score of pivot column */
+  IndexType needed_memory ;   /* free space needed for pivot row */
+  IndexType *cp_end ;   /* pointer to the end of a column */
+  IndexType *rp_end ;   /* pointer to the end of a row */
+  IndexType row ;     /* a row index */
+  IndexType col ;     /* a column index */
+  IndexType max_score ;   /* maximum possible score */
+  IndexType cur_score ;   /* score of current column */
+  unsigned int hash ;   /* hash value for supernode detection */
+  IndexType head_column ;   /* head of hash bucket */
+  IndexType first_col ;   /* first column in hash bucket */
+  IndexType tag_mark ;    /* marker value for mark array */
+  IndexType row_mark ;    /* Row [row].shared2.mark */
+  IndexType set_difference ;  /* set difference size of row with pivot row */
+  IndexType min_score ;   /* smallest column score */
+  IndexType col_thickness ;   /* "thickness" (no. of columns in a supercol) */
+  IndexType max_mark ;    /* maximum value of tag_mark */
+  IndexType pivot_col_thickness ; /* number of columns represented by pivot col */
+  IndexType prev_col ;    /* Used by Dlist operations. */
+  IndexType next_col ;    /* Used by Dlist operations. */
+  IndexType ngarbage ;    /* number of garbage collections performed */
+
+
+  /* === Initialization and clear mark ==================================== */
+
+  max_mark = INT_MAX - n_col ;  /* INT_MAX defined in <limits.h> */
+  tag_mark = Eigen::internal::clear_mark (n_row, Row) ;
+  min_score = 0 ;
+  ngarbage = 0 ;
+  COLAMD_DEBUG1 (("colamd: Ordering, n_col2=%d\n", n_col2)) ;
+
+  /* === Order the columns ================================================ */
+
+  for (k = 0 ; k < n_col2 ; /* 'k' is incremented below */)
+  {
+
+    /* === Select pivot column, and order it ============================ */
+
+    /* make sure degree list isn't empty */
+    COLAMD_ASSERT (min_score >= 0) ;
+    COLAMD_ASSERT (min_score <= n_col) ;
+    COLAMD_ASSERT (head [min_score] >= COLAMD_EMPTY) ;
+
+    /* get pivot column from head of minimum degree list */
+    while (min_score < n_col && head [min_score] == COLAMD_EMPTY)
+    {
+      min_score++ ;
+    }
+    pivot_col = head [min_score] ;
+    COLAMD_ASSERT (pivot_col >= 0 && pivot_col <= n_col) ;
+    next_col = Col [pivot_col].shared4.degree_next ;
+    head [min_score] = next_col ;
+    if (next_col != COLAMD_EMPTY)
+    {
+      Col [next_col].shared3.prev = COLAMD_EMPTY ;
+    }
+
+    COLAMD_ASSERT (COL_IS_ALIVE (pivot_col)) ;
+    COLAMD_DEBUG3 (("Pivot col: %d\n", pivot_col)) ;
+
+    /* remember score for defrag check */
+    pivot_col_score = Col [pivot_col].shared2.score ;
+
+    /* the pivot column is the kth column in the pivot order */
+    Col [pivot_col].shared2.order = k ;
+
+    /* increment order count by column thickness */
+    pivot_col_thickness = Col [pivot_col].shared1.thickness ;
+    k += pivot_col_thickness ;
+    COLAMD_ASSERT (pivot_col_thickness > 0) ;
+
+    /* === Garbage_collection, if necessary ============================= */
+
+    needed_memory = numext::mini(pivot_col_score, n_col - k) ;
+    if (pfree + needed_memory >= Alen)
+    {
+      pfree = Eigen::internal::garbage_collection (n_row, n_col, Row, Col, A, &A [pfree]) ;
+      ngarbage++ ;
+      /* after garbage collection we will have enough */
+      COLAMD_ASSERT (pfree + needed_memory < Alen) ;
+      /* garbage collection has wiped out the Row[].shared2.mark array */
+      tag_mark = Eigen::internal::clear_mark (n_row, Row) ;
+
+    }
+
+    /* === Compute pivot row pattern ==================================== */
+
+    /* get starting location for this new merged row */
+    pivot_row_start = pfree ;
+
+    /* initialize new row counts to zero */
+    pivot_row_degree = 0 ;
+
+    /* tag pivot column as having been visited so it isn't included */
+    /* in merged pivot row */
+    Col [pivot_col].shared1.thickness = -pivot_col_thickness ;
+
+    /* pivot row is the union of all rows in the pivot column pattern */
+    cp = &A [Col [pivot_col].start] ;
+    cp_end = cp + Col [pivot_col].length ;
+    while (cp < cp_end)
+    {
+      /* get a row */
+      row = *cp++ ;
+      COLAMD_DEBUG4 (("Pivot col pattern %d %d\n", ROW_IS_ALIVE (row), row)) ;
+      /* skip if row is dead */
+      if (ROW_IS_DEAD (row))
+      {
+	continue ;
+      }
+      rp = &A [Row [row].start] ;
+      rp_end = rp + Row [row].length ;
+      while (rp < rp_end)
+      {
+	/* get a column */
+	col = *rp++ ;
+	/* add the column, if alive and untagged */
+	col_thickness = Col [col].shared1.thickness ;
+	if (col_thickness > 0 && COL_IS_ALIVE (col))
+	{
+	  /* tag column in pivot row */
+	  Col [col].shared1.thickness = -col_thickness ;
+	  COLAMD_ASSERT (pfree < Alen) ;
+	  /* place column in pivot row */
+	  A [pfree++] = col ;
+	  pivot_row_degree += col_thickness ;
+	}
+      }
+    }
+
+    /* clear tag on pivot column */
+    Col [pivot_col].shared1.thickness = pivot_col_thickness ;
+    max_deg = numext::maxi(max_deg, pivot_row_degree) ;
+
+
+    /* === Kill all rows used to construct pivot row ==================== */
+
+    /* also kill pivot row, temporarily */
+    cp = &A [Col [pivot_col].start] ;
+    cp_end = cp + Col [pivot_col].length ;
+    while (cp < cp_end)
+    {
+      /* may be killing an already dead row */
+      row = *cp++ ;
+      COLAMD_DEBUG3 (("Kill row in pivot col: %d\n", row)) ;
+      KILL_ROW (row) ;
+    }
+
+    /* === Select a row index to use as the new pivot row =============== */
+
+    pivot_row_length = pfree - pivot_row_start ;
+    if (pivot_row_length > 0)
+    {
+      /* pick the "pivot" row arbitrarily (first row in col) */
+      pivot_row = A [Col [pivot_col].start] ;
+      COLAMD_DEBUG3 (("Pivotal row is %d\n", pivot_row)) ;
+    }
+    else
+    {
+      /* there is no pivot row, since it is of zero length */
+      pivot_row = COLAMD_EMPTY ;
+      COLAMD_ASSERT (pivot_row_length == 0) ;
+    }
+    COLAMD_ASSERT (Col [pivot_col].length > 0 || pivot_row_length == 0) ;
+
+    /* === Approximate degree computation =============================== */
+
+    /* Here begins the computation of the approximate degree.  The column */
+    /* score is the sum of the pivot row "length", plus the size of the */
+    /* set differences of each row in the column minus the pattern of the */
+    /* pivot row itself.  The column ("thickness") itself is also */
+    /* excluded from the column score (we thus use an approximate */
+    /* external degree). */
+
+    /* The time taken by the following code (compute set differences, and */
+    /* add them up) is proportional to the size of the data structure */
+    /* being scanned - that is, the sum of the sizes of each column in */
+    /* the pivot row.  Thus, the amortized time to compute a column score */
+    /* is proportional to the size of that column (where size, in this */
+    /* context, is the column "length", or the number of row indices */
+    /* in that column).  The number of row indices in a column is */
+    /* monotonically non-decreasing, from the length of the original */
+    /* column on input to colamd. */
+
+    /* === Compute set differences ====================================== */
+
+    COLAMD_DEBUG3 (("** Computing set differences phase. **\n")) ;
+
+    /* pivot row is currently dead - it will be revived later. */
+
+    COLAMD_DEBUG3 (("Pivot row: ")) ;
+    /* for each column in pivot row */
+    rp = &A [pivot_row_start] ;
+    rp_end = rp + pivot_row_length ;
+    while (rp < rp_end)
+    {
+      col = *rp++ ;
+      COLAMD_ASSERT (COL_IS_ALIVE (col) && col != pivot_col) ;
+      COLAMD_DEBUG3 (("Col: %d\n", col)) ;
+
+      /* clear tags used to construct pivot row pattern */
+      col_thickness = -Col [col].shared1.thickness ;
+      COLAMD_ASSERT (col_thickness > 0) ;
+      Col [col].shared1.thickness = col_thickness ;
+
+      /* === Remove column from degree list =========================== */
+
+      cur_score = Col [col].shared2.score ;
+      prev_col = Col [col].shared3.prev ;
+      next_col = Col [col].shared4.degree_next ;
+      COLAMD_ASSERT (cur_score >= 0) ;
+      COLAMD_ASSERT (cur_score <= n_col) ;
+      COLAMD_ASSERT (cur_score >= COLAMD_EMPTY) ;
+      if (prev_col == COLAMD_EMPTY)
+      {
+	head [cur_score] = next_col ;
+      }
+      else
+      {
+	Col [prev_col].shared4.degree_next = next_col ;
+      }
+      if (next_col != COLAMD_EMPTY)
+      {
+	Col [next_col].shared3.prev = prev_col ;
+      }
+
+      /* === Scan the column ========================================== */
+
+      cp = &A [Col [col].start] ;
+      cp_end = cp + Col [col].length ;
+      while (cp < cp_end)
+      {
+	/* get a row */
+	row = *cp++ ;
+	row_mark = Row [row].shared2.mark ;
+	/* skip if dead */
+	if (ROW_IS_MARKED_DEAD (row_mark))
+	{
+	  continue ;
+	}
+	COLAMD_ASSERT (row != pivot_row) ;
+	set_difference = row_mark - tag_mark ;
+	/* check if the row has been seen yet */
+	if (set_difference < 0)
+	{
+	  COLAMD_ASSERT (Row [row].shared1.degree <= max_deg) ;
+	  set_difference = Row [row].shared1.degree ;
+	}
+	/* subtract column thickness from this row's set difference */
+	set_difference -= col_thickness ;
+	COLAMD_ASSERT (set_difference >= 0) ;
+	/* absorb this row if the set difference becomes zero */
+	if (set_difference == 0)
+	{
+	  COLAMD_DEBUG3 (("aggressive absorption. Row: %d\n", row)) ;
+	  KILL_ROW (row) ;
+	}
+	else
+	{
+	  /* save the new mark */
+	  Row [row].shared2.mark = set_difference + tag_mark ;
+	}
+      }
+    }
+
+
+    /* === Add up set differences for each column ======================= */
+
+    COLAMD_DEBUG3 (("** Adding set differences phase. **\n")) ;
+
+    /* for each column in pivot row */
+    rp = &A [pivot_row_start] ;
+    rp_end = rp + pivot_row_length ;
+    while (rp < rp_end)
+    {
+      /* get a column */
+      col = *rp++ ;
+      COLAMD_ASSERT (COL_IS_ALIVE (col) && col != pivot_col) ;
+      hash = 0 ;
+      cur_score = 0 ;
+      cp = &A [Col [col].start] ;
+      /* compact the column */
+      new_cp = cp ;
+      cp_end = cp + Col [col].length ;
+
+      COLAMD_DEBUG4 (("Adding set diffs for Col: %d.\n", col)) ;
+
+      while (cp < cp_end)
+      {
+	/* get a row */
+	row = *cp++ ;
+	COLAMD_ASSERT(row >= 0 && row < n_row) ;
+	row_mark = Row [row].shared2.mark ;
+	/* skip if dead */
+	if (ROW_IS_MARKED_DEAD (row_mark))
+	{
+	  continue ;
+	}
+	COLAMD_ASSERT (row_mark > tag_mark) ;
+	/* compact the column */
+	*new_cp++ = row ;
+	/* compute hash function */
+	hash += row ;
+	/* add set difference */
+	cur_score += row_mark - tag_mark ;
+	/* integer overflow... */
+	cur_score = numext::mini(cur_score, n_col) ;
+      }
+
+      /* recompute the column's length */
+      Col [col].length = (IndexType) (new_cp - &A [Col [col].start]) ;
+
+      /* === Further mass elimination ================================= */
+
+      if (Col [col].length == 0)
+      {
+	COLAMD_DEBUG4 (("further mass elimination. Col: %d\n", col)) ;
+	/* nothing left but the pivot row in this column */
+	KILL_PRINCIPAL_COL (col) ;
+	pivot_row_degree -= Col [col].shared1.thickness ;
+	COLAMD_ASSERT (pivot_row_degree >= 0) ;
+	/* order it */
+	Col [col].shared2.order = k ;
+	/* increment order count by column thickness */
+	k += Col [col].shared1.thickness ;
+      }
+      else
+      {
+	/* === Prepare for supercolumn detection ==================== */
+
+	COLAMD_DEBUG4 (("Preparing supercol detection for Col: %d.\n", col)) ;
+
+	/* save score so far */
+	Col [col].shared2.score = cur_score ;
+
+	/* add column to hash table, for supercolumn detection */
+	hash %= n_col + 1 ;
+
+	COLAMD_DEBUG4 ((" Hash = %d, n_col = %d.\n", hash, n_col)) ;
+	COLAMD_ASSERT (hash <= n_col) ;
+
+	head_column = head [hash] ;
+	if (head_column > COLAMD_EMPTY)
+	{
+	  /* degree list "hash" is non-empty, use prev (shared3) of */
+	  /* first column in degree list as head of hash bucket */
+	  first_col = Col [head_column].shared3.headhash ;
+	  Col [head_column].shared3.headhash = col ;
+	}
+	else
+	{
+	  /* degree list "hash" is empty, use head as hash bucket */
+	  first_col = - (head_column + 2) ;
+	  head [hash] = - (col + 2) ;
+	}
+	Col [col].shared4.hash_next = first_col ;
+
+	/* save hash function in Col [col].shared3.hash */
+	Col [col].shared3.hash = (IndexType) hash ;
+	COLAMD_ASSERT (COL_IS_ALIVE (col)) ;
+      }
+    }
+
+    /* The approximate external column degree is now computed.  */
+
+    /* === Supercolumn detection ======================================== */
+
+    COLAMD_DEBUG3 (("** Supercolumn detection phase. **\n")) ;
+
+    Eigen::internal::detect_super_cols (Col, A, head, pivot_row_start, pivot_row_length) ;
+
+    /* === Kill the pivotal column ====================================== */
+
+    KILL_PRINCIPAL_COL (pivot_col) ;
+
+    /* === Clear mark =================================================== */
+
+    tag_mark += (max_deg + 1) ;
+    if (tag_mark >= max_mark)
+    {
+      COLAMD_DEBUG2 (("clearing tag_mark\n")) ;
+      tag_mark = Eigen::internal::clear_mark (n_row, Row) ;
+    }
+
+    /* === Finalize the new pivot row, and column scores ================ */
+
+    COLAMD_DEBUG3 (("** Finalize scores phase. **\n")) ;
+
+    /* for each column in pivot row */
+    rp = &A [pivot_row_start] ;
+    /* compact the pivot row */
+    new_rp = rp ;
+    rp_end = rp + pivot_row_length ;
+    while (rp < rp_end)
+    {
+      col = *rp++ ;
+      /* skip dead columns */
+      if (COL_IS_DEAD (col))
+      {
+	continue ;
+      }
+      *new_rp++ = col ;
+      /* add new pivot row to column */
+      A [Col [col].start + (Col [col].length++)] = pivot_row ;
+
+      /* retrieve score so far and add on pivot row's degree. */
+      /* (we wait until here for this in case the pivot */
+      /* row's degree was reduced due to mass elimination). */
+      cur_score = Col [col].shared2.score + pivot_row_degree ;
+
+      /* calculate the max possible score as the number of */
+      /* external columns minus the 'k' value minus the */
+      /* columns thickness */
+      max_score = n_col - k - Col [col].shared1.thickness ;
+
+      /* make the score the external degree of the union-of-rows */
+      cur_score -= Col [col].shared1.thickness ;
+
+      /* make sure score is less or equal than the max score */
+      cur_score = numext::mini(cur_score, max_score) ;
+      COLAMD_ASSERT (cur_score >= 0) ;
+
+      /* store updated score */
+      Col [col].shared2.score = cur_score ;
+
+      /* === Place column back in degree list ========================= */
+
+      COLAMD_ASSERT (min_score >= 0) ;
+      COLAMD_ASSERT (min_score <= n_col) ;
+      COLAMD_ASSERT (cur_score >= 0) ;
+      COLAMD_ASSERT (cur_score <= n_col) ;
+      COLAMD_ASSERT (head [cur_score] >= COLAMD_EMPTY) ;
+      next_col = head [cur_score] ;
+      Col [col].shared4.degree_next = next_col ;
+      Col [col].shared3.prev = COLAMD_EMPTY ;
+      if (next_col != COLAMD_EMPTY)
+      {
+	Col [next_col].shared3.prev = col ;
+      }
+      head [cur_score] = col ;
+
+      /* see if this score is less than current min */
+      min_score = numext::mini(min_score, cur_score) ;
+
+    }
+
+    /* === Resurrect the new pivot row ================================== */
+
+    if (pivot_row_degree > 0)
+    {
+      /* update pivot row length to reflect any cols that were killed */
+      /* during super-col detection and mass elimination */
+      Row [pivot_row].start  = pivot_row_start ;
+      Row [pivot_row].length = (IndexType) (new_rp - &A[pivot_row_start]) ;
+      Row [pivot_row].shared1.degree = pivot_row_degree ;
+      Row [pivot_row].shared2.mark = 0 ;
+      /* pivot row is no longer dead */
+    }
+  }
+
+  /* === All principal columns have now been ordered ====================== */
+
+  return (ngarbage) ;
+}
+
+
+/* ========================================================================== */
+/* === order_children ======================================================= */
+/* ========================================================================== */
+
+/*
+  The find_ordering routine has ordered all of the principal columns (the
+  representatives of the supercolumns).  The non-principal columns have not
+  yet been ordered.  This routine orders those columns by walking up the
+  parent tree (a column is a child of the column which absorbed it).  The
+  final permutation vector is then placed in p [0 ... n_col-1], with p [0]
+  being the first column, and p [n_col-1] being the last.  It doesn't look
+  like it at first glance, but be assured that this routine takes time linear
+  in the number of columns.  Although not immediately obvious, the time
+  taken by this routine is O (n_col), that is, linear in the number of
+  columns.  Not user-callable.
+*/
+template <typename IndexType>
+static inline  void order_children
+(
+  /* === Parameters ======================================================= */
+
+  IndexType n_col,      /* number of columns of A */
+  colamd_col<IndexType> Col [],    /* of size n_col+1 */
+  IndexType p []      /* p [0 ... n_col-1] is the column permutation*/
+  )
+{
+  /* === Local variables ================================================== */
+
+  IndexType i ;     /* loop counter for all columns */
+  IndexType c ;     /* column index */
+  IndexType parent ;    /* index of column's parent */
+  IndexType order ;     /* column's order */
+
+  /* === Order each non-principal column ================================== */
+
+  for (i = 0 ; i < n_col ; i++)
+  {
+    /* find an un-ordered non-principal column */
+    COLAMD_ASSERT (COL_IS_DEAD (i)) ;
+    if (!COL_IS_DEAD_PRINCIPAL (i) && Col [i].shared2.order == COLAMD_EMPTY)
+    {
+      parent = i ;
+      /* once found, find its principal parent */
+      do
+      {
+	parent = Col [parent].shared1.parent ;
+      } while (!COL_IS_DEAD_PRINCIPAL (parent)) ;
+
+      /* now, order all un-ordered non-principal columns along path */
+      /* to this parent.  collapse tree at the same time */
+      c = i ;
+      /* get order of parent */
+      order = Col [parent].shared2.order ;
+
+      do
+      {
+	COLAMD_ASSERT (Col [c].shared2.order == COLAMD_EMPTY) ;
+
+	/* order this column */
+	Col [c].shared2.order = order++ ;
+	/* collaps tree */
+	Col [c].shared1.parent = parent ;
+
+	/* get immediate parent of this column */
+	c = Col [c].shared1.parent ;
+
+	/* continue until we hit an ordered column.  There are */
+	/* guarranteed not to be anymore unordered columns */
+	/* above an ordered column */
+      } while (Col [c].shared2.order == COLAMD_EMPTY) ;
+
+      /* re-order the super_col parent to largest order for this group */
+      Col [parent].shared2.order = order ;
+    }
+  }
+
+  /* === Generate the permutation ========================================= */
+
+  for (c = 0 ; c < n_col ; c++)
+  {
+    p [Col [c].shared2.order] = c ;
+  }
+}
+
+
+/* ========================================================================== */
+/* === detect_super_cols ==================================================== */
+/* ========================================================================== */
+
+/*
+  Detects supercolumns by finding matches between columns in the hash buckets.
+  Check amongst columns in the set A [row_start ... row_start + row_length-1].
+  The columns under consideration are currently *not* in the degree lists,
+  and have already been placed in the hash buckets.
+
+  The hash bucket for columns whose hash function is equal to h is stored
+  as follows:
+
+  if head [h] is >= 0, then head [h] contains a degree list, so:
+
+  head [h] is the first column in degree bucket h.
+  Col [head [h]].headhash gives the first column in hash bucket h.
+
+  otherwise, the degree list is empty, and:
+
+  -(head [h] + 2) is the first column in hash bucket h.
+
+  For a column c in a hash bucket, Col [c].shared3.prev is NOT a "previous
+  column" pointer.  Col [c].shared3.hash is used instead as the hash number
+  for that column.  The value of Col [c].shared4.hash_next is the next column
+  in the same hash bucket.
+
+  Assuming no, or "few" hash collisions, the time taken by this routine is
+  linear in the sum of the sizes (lengths) of each column whose score has
+  just been computed in the approximate degree computation.
+  Not user-callable.
+*/
+template <typename IndexType>
+static void detect_super_cols
+(
+  /* === Parameters ======================================================= */
+  
+  colamd_col<IndexType> Col [],    /* of size n_col+1 */
+  IndexType A [],     /* row indices of A */
+  IndexType head [],    /* head of degree lists and hash buckets */
+  IndexType row_start,    /* pointer to set of columns to check */
+  IndexType row_length    /* number of columns to check */
+)
+{
+  /* === Local variables ================================================== */
+
+  IndexType hash ;      /* hash value for a column */
+  IndexType *rp ;     /* pointer to a row */
+  IndexType c ;     /* a column index */
+  IndexType super_c ;   /* column index of the column to absorb into */
+  IndexType *cp1 ;      /* column pointer for column super_c */
+  IndexType *cp2 ;      /* column pointer for column c */
+  IndexType length ;    /* length of column super_c */
+  IndexType prev_c ;    /* column preceding c in hash bucket */
+  IndexType i ;     /* loop counter */
+  IndexType *rp_end ;   /* pointer to the end of the row */
+  IndexType col ;     /* a column index in the row to check */
+  IndexType head_column ;   /* first column in hash bucket or degree list */
+  IndexType first_col ;   /* first column in hash bucket */
+
+  /* === Consider each column in the row ================================== */
+
+  rp = &A [row_start] ;
+  rp_end = rp + row_length ;
+  while (rp < rp_end)
+  {
+    col = *rp++ ;
+    if (COL_IS_DEAD (col))
+    {
+      continue ;
+    }
+
+    /* get hash number for this column */
+    hash = Col [col].shared3.hash ;
+    COLAMD_ASSERT (hash <= n_col) ;
+
+    /* === Get the first column in this hash bucket ===================== */
+
+    head_column = head [hash] ;
+    if (head_column > COLAMD_EMPTY)
+    {
+      first_col = Col [head_column].shared3.headhash ;
+    }
+    else
+    {
+      first_col = - (head_column + 2) ;
+    }
+
+    /* === Consider each column in the hash bucket ====================== */
+
+    for (super_c = first_col ; super_c != COLAMD_EMPTY ;
+	 super_c = Col [super_c].shared4.hash_next)
+    {
+      COLAMD_ASSERT (COL_IS_ALIVE (super_c)) ;
+      COLAMD_ASSERT (Col [super_c].shared3.hash == hash) ;
+      length = Col [super_c].length ;
+
+      /* prev_c is the column preceding column c in the hash bucket */
+      prev_c = super_c ;
+
+      /* === Compare super_c with all columns after it ================ */
+
+      for (c = Col [super_c].shared4.hash_next ;
+	   c != COLAMD_EMPTY ; c = Col [c].shared4.hash_next)
+      {
+	COLAMD_ASSERT (c != super_c) ;
+	COLAMD_ASSERT (COL_IS_ALIVE (c)) ;
+	COLAMD_ASSERT (Col [c].shared3.hash == hash) ;
+
+	/* not identical if lengths or scores are different */
+	if (Col [c].length != length ||
+	    Col [c].shared2.score != Col [super_c].shared2.score)
+	{
+	  prev_c = c ;
+	  continue ;
+	}
+
+	/* compare the two columns */
+	cp1 = &A [Col [super_c].start] ;
+	cp2 = &A [Col [c].start] ;
+
+	for (i = 0 ; i < length ; i++)
+	{
+	  /* the columns are "clean" (no dead rows) */
+	  COLAMD_ASSERT (ROW_IS_ALIVE (*cp1))  ;
+	  COLAMD_ASSERT (ROW_IS_ALIVE (*cp2))  ;
+	  /* row indices will same order for both supercols, */
+	  /* no gather scatter nessasary */
+	  if (*cp1++ != *cp2++)
+	  {
+	    break ;
+	  }
+	}
+
+	/* the two columns are different if the for-loop "broke" */
+	if (i != length)
+	{
+	  prev_c = c ;
+	  continue ;
+	}
+
+	/* === Got it!  two columns are identical =================== */
+
+	COLAMD_ASSERT (Col [c].shared2.score == Col [super_c].shared2.score) ;
+
+	Col [super_c].shared1.thickness += Col [c].shared1.thickness ;
+	Col [c].shared1.parent = super_c ;
+	KILL_NON_PRINCIPAL_COL (c) ;
+	/* order c later, in order_children() */
+	Col [c].shared2.order = COLAMD_EMPTY ;
+	/* remove c from hash bucket */
+	Col [prev_c].shared4.hash_next = Col [c].shared4.hash_next ;
+      }
+    }
+
+    /* === Empty this hash bucket ======================================= */
+
+    if (head_column > COLAMD_EMPTY)
+    {
+      /* corresponding degree list "hash" is not empty */
+      Col [head_column].shared3.headhash = COLAMD_EMPTY ;
+    }
+    else
+    {
+      /* corresponding degree list "hash" is empty */
+      head [hash] = COLAMD_EMPTY ;
+    }
+  }
+}
+
+
+/* ========================================================================== */
+/* === garbage_collection =================================================== */
+/* ========================================================================== */
+
+/*
+  Defragments and compacts columns and rows in the workspace A.  Used when
+  all avaliable memory has been used while performing row merging.  Returns
+  the index of the first free position in A, after garbage collection.  The
+  time taken by this routine is linear is the size of the array A, which is
+  itself linear in the number of nonzeros in the input matrix.
+  Not user-callable.
+*/
+template <typename IndexType>
+static IndexType garbage_collection  /* returns the new value of pfree */
+  (
+    /* === Parameters ======================================================= */
+    
+    IndexType n_row,      /* number of rows */
+    IndexType n_col,      /* number of columns */
+    Colamd_Row<IndexType> Row [],    /* row info */
+    colamd_col<IndexType> Col [],    /* column info */
+    IndexType A [],     /* A [0 ... Alen-1] holds the matrix */
+    IndexType *pfree      /* &A [0] ... pfree is in use */
+    )
+{
+  /* === Local variables ================================================== */
+
+  IndexType *psrc ;     /* source pointer */
+  IndexType *pdest ;    /* destination pointer */
+  IndexType j ;     /* counter */
+  IndexType r ;     /* a row index */
+  IndexType c ;     /* a column index */
+  IndexType length ;    /* length of a row or column */
+
+  /* === Defragment the columns =========================================== */
+
+  pdest = &A[0] ;
+  for (c = 0 ; c < n_col ; c++)
+  {
+    if (COL_IS_ALIVE (c))
+    {
+      psrc = &A [Col [c].start] ;
+
+      /* move and compact the column */
+      COLAMD_ASSERT (pdest <= psrc) ;
+      Col [c].start = (IndexType) (pdest - &A [0]) ;
+      length = Col [c].length ;
+      for (j = 0 ; j < length ; j++)
+      {
+	r = *psrc++ ;
+	if (ROW_IS_ALIVE (r))
+	{
+	  *pdest++ = r ;
+	}
+      }
+      Col [c].length = (IndexType) (pdest - &A [Col [c].start]) ;
+    }
+  }
+
+  /* === Prepare to defragment the rows =================================== */
+
+  for (r = 0 ; r < n_row ; r++)
+  {
+    if (ROW_IS_ALIVE (r))
+    {
+      if (Row [r].length == 0)
+      {
+	/* this row is of zero length.  cannot compact it, so kill it */
+	COLAMD_DEBUG3 (("Defrag row kill\n")) ;
+	KILL_ROW (r) ;
+      }
+      else
+      {
+	/* save first column index in Row [r].shared2.first_column */
+	psrc = &A [Row [r].start] ;
+	Row [r].shared2.first_column = *psrc ;
+	COLAMD_ASSERT (ROW_IS_ALIVE (r)) ;
+	/* flag the start of the row with the one's complement of row */
+	*psrc = ONES_COMPLEMENT (r) ;
+
+      }
+    }
+  }
+
+  /* === Defragment the rows ============================================== */
+
+  psrc = pdest ;
+  while (psrc < pfree)
+  {
+    /* find a negative number ... the start of a row */
+    if (*psrc++ < 0)
+    {
+      psrc-- ;
+      /* get the row index */
+      r = ONES_COMPLEMENT (*psrc) ;
+      COLAMD_ASSERT (r >= 0 && r < n_row) ;
+      /* restore first column index */
+      *psrc = Row [r].shared2.first_column ;
+      COLAMD_ASSERT (ROW_IS_ALIVE (r)) ;
+
+      /* move and compact the row */
+      COLAMD_ASSERT (pdest <= psrc) ;
+      Row [r].start = (IndexType) (pdest - &A [0]) ;
+      length = Row [r].length ;
+      for (j = 0 ; j < length ; j++)
+      {
+	c = *psrc++ ;
+	if (COL_IS_ALIVE (c))
+	{
+	  *pdest++ = c ;
+	}
+      }
+      Row [r].length = (IndexType) (pdest - &A [Row [r].start]) ;
+
+    }
+  }
+  /* ensure we found all the rows */
+  COLAMD_ASSERT (debug_rows == 0) ;
+
+  /* === Return the new value of pfree ==================================== */
+
+  return ((IndexType) (pdest - &A [0])) ;
+}
+
+
+/* ========================================================================== */
+/* === clear_mark =========================================================== */
+/* ========================================================================== */
+
+/*
+  Clears the Row [].shared2.mark array, and returns the new tag_mark.
+  Return value is the new tag_mark.  Not user-callable.
+*/
+template <typename IndexType>
+static inline  IndexType clear_mark  /* return the new value for tag_mark */
+  (
+      /* === Parameters ======================================================= */
+
+    IndexType n_row,    /* number of rows in A */
+    Colamd_Row<IndexType> Row [] /* Row [0 ... n_row-1].shared2.mark is set to zero */
+    )
+{
+  /* === Local variables ================================================== */
+
+  IndexType r ;
+
+  for (r = 0 ; r < n_row ; r++)
+  {
+    if (ROW_IS_ALIVE (r))
+    {
+      Row [r].shared2.mark = 0 ;
+    }
+  }
+  return (1) ;
+}
+
+
+} // namespace internal 
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Ordering.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Ordering.h
new file mode 100644
index 0000000..7ea9b14
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/OrderingMethods/Ordering.h
@@ -0,0 +1,157 @@
+ 
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012  Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_ORDERING_H
+#define EIGEN_ORDERING_H
+
+namespace Eigen {
+  
+#include "Eigen_Colamd.h"
+
+namespace internal {
+    
+/** \internal
+  * \ingroup OrderingMethods_Module
+  * \param[in] A the input non-symmetric matrix
+  * \param[out] symmat the symmetric pattern A^T+A from the input matrix \a A.
+  * FIXME: The values should not be considered here
+  */
+template<typename MatrixType> 
+void ordering_helper_at_plus_a(const MatrixType& A, MatrixType& symmat)
+{
+  MatrixType C;
+  C = A.transpose(); // NOTE: Could be  costly
+  for (int i = 0; i < C.rows(); i++) 
+  {
+      for (typename MatrixType::InnerIterator it(C, i); it; ++it)
+        it.valueRef() = 0.0;
+  }
+  symmat = C + A;
+}
+    
+}
+
+#ifndef EIGEN_MPL2_ONLY
+
+/** \ingroup OrderingMethods_Module
+  * \class AMDOrdering
+  *
+  * Functor computing the \em approximate \em minimum \em degree ordering
+  * If the matrix is not structurally symmetric, an ordering of A^T+A is computed
+  * \tparam  StorageIndex The type of indices of the matrix 
+  * \sa COLAMDOrdering
+  */
+template <typename StorageIndex>
+class AMDOrdering
+{
+  public:
+    typedef PermutationMatrix<Dynamic, Dynamic, StorageIndex> PermutationType;
+    
+    /** Compute the permutation vector from a sparse matrix
+     * This routine is much faster if the input matrix is column-major     
+     */
+    template <typename MatrixType>
+    void operator()(const MatrixType& mat, PermutationType& perm)
+    {
+      // Compute the symmetric pattern
+      SparseMatrix<typename MatrixType::Scalar, ColMajor, StorageIndex> symm;
+      internal::ordering_helper_at_plus_a(mat,symm); 
+    
+      // Call the AMD routine 
+      //m_mat.prune(keep_diag());
+      internal::minimum_degree_ordering(symm, perm);
+    }
+    
+    /** Compute the permutation with a selfadjoint matrix */
+    template <typename SrcType, unsigned int SrcUpLo> 
+    void operator()(const SparseSelfAdjointView<SrcType, SrcUpLo>& mat, PermutationType& perm)
+    { 
+      SparseMatrix<typename SrcType::Scalar, ColMajor, StorageIndex> C; C = mat;
+      
+      // Call the AMD routine 
+      // m_mat.prune(keep_diag()); //Remove the diagonal elements 
+      internal::minimum_degree_ordering(C, perm);
+    }
+};
+
+#endif // EIGEN_MPL2_ONLY
+
+/** \ingroup OrderingMethods_Module
+  * \class NaturalOrdering
+  *
+  * Functor computing the natural ordering (identity)
+  * 
+  * \note Returns an empty permutation matrix
+  * \tparam  StorageIndex The type of indices of the matrix 
+  */
+template <typename StorageIndex>
+class NaturalOrdering
+{
+  public:
+    typedef PermutationMatrix<Dynamic, Dynamic, StorageIndex> PermutationType;
+    
+    /** Compute the permutation vector from a column-major sparse matrix */
+    template <typename MatrixType>
+    void operator()(const MatrixType& /*mat*/, PermutationType& perm)
+    {
+      perm.resize(0); 
+    }
+    
+};
+
+/** \ingroup OrderingMethods_Module
+  * \class COLAMDOrdering
+  *
+  * \tparam  StorageIndex The type of indices of the matrix 
+  * 
+  * Functor computing the \em column \em approximate \em minimum \em degree ordering 
+  * The matrix should be in column-major and \b compressed format (see SparseMatrix::makeCompressed()).
+  */
+template<typename StorageIndex>
+class COLAMDOrdering
+{
+  public:
+    typedef PermutationMatrix<Dynamic, Dynamic, StorageIndex> PermutationType; 
+    typedef Matrix<StorageIndex, Dynamic, 1> IndexVector;
+    
+    /** Compute the permutation vector \a perm form the sparse matrix \a mat
+      * \warning The input sparse matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()).
+      */
+    template <typename MatrixType>
+    void operator() (const MatrixType& mat, PermutationType& perm)
+    {
+      eigen_assert(mat.isCompressed() && "COLAMDOrdering requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to COLAMDOrdering");
+      
+      StorageIndex m = StorageIndex(mat.rows());
+      StorageIndex n = StorageIndex(mat.cols());
+      StorageIndex nnz = StorageIndex(mat.nonZeros());
+      // Get the recommended value of Alen to be used by colamd
+      StorageIndex Alen = internal::colamd_recommended(nnz, m, n); 
+      // Set the default parameters
+      double knobs [COLAMD_KNOBS]; 
+      StorageIndex stats [COLAMD_STATS];
+      internal::colamd_set_defaults(knobs);
+      
+      IndexVector p(n+1), A(Alen); 
+      for(StorageIndex i=0; i <= n; i++)   p(i) = mat.outerIndexPtr()[i];
+      for(StorageIndex i=0; i < nnz; i++)  A(i) = mat.innerIndexPtr()[i];
+      // Call Colamd routine to compute the ordering 
+      StorageIndex info = internal::colamd(m, n, Alen, A.data(), p.data(), knobs, stats); 
+      EIGEN_UNUSED_VARIABLE(info);
+      eigen_assert( info && "COLAMD failed " );
+      
+      perm.resize(n);
+      for (StorageIndex i = 0; i < n; i++) perm.indices()(p(i)) = i;
+    }
+};
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h
new file mode 100644
index 0000000..d2ebfd7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PaStiXSupport/PaStiXSupport.h
@@ -0,0 +1,678 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PASTIXSUPPORT_H
+#define EIGEN_PASTIXSUPPORT_H
+
+namespace Eigen { 
+
+#if defined(DCOMPLEX)
+  #define PASTIX_COMPLEX  COMPLEX
+  #define PASTIX_DCOMPLEX DCOMPLEX
+#else
+  #define PASTIX_COMPLEX  std::complex<float>
+  #define PASTIX_DCOMPLEX std::complex<double>
+#endif
+
+/** \ingroup PaStiXSupport_Module
+  * \brief Interface to the PaStix solver
+  * 
+  * This class is used to solve the linear systems A.X = B via the PaStix library. 
+  * The matrix can be either real or complex, symmetric or not.
+  *
+  * \sa TutorialSparseDirectSolvers
+  */
+template<typename _MatrixType, bool IsStrSym = false> class PastixLU;
+template<typename _MatrixType, int Options> class PastixLLT;
+template<typename _MatrixType, int Options> class PastixLDLT;
+
+namespace internal
+{
+    
+  template<class Pastix> struct pastix_traits;
+
+  template<typename _MatrixType>
+  struct pastix_traits< PastixLU<_MatrixType> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;
+  };
+
+  template<typename _MatrixType, int Options>
+  struct pastix_traits< PastixLLT<_MatrixType,Options> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;
+  };
+
+  template<typename _MatrixType, int Options>
+  struct pastix_traits< PastixLDLT<_MatrixType,Options> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;
+  };
+  
+  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, float *vals, int *perm, int * invp, float *x, int nbrhs, int *iparm, double *dparm)
+  {
+    if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
+    if (nbrhs == 0) {x = NULL; nbrhs=1;}
+    s_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); 
+  }
+  
+  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, double *vals, int *perm, int * invp, double *x, int nbrhs, int *iparm, double *dparm)
+  {
+    if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
+    if (nbrhs == 0) {x = NULL; nbrhs=1;}
+    d_pastix(pastix_data, pastix_comm, n, ptr, idx, vals, perm, invp, x, nbrhs, iparm, dparm); 
+  }
+  
+  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<float> *vals, int *perm, int * invp, std::complex<float> *x, int nbrhs, int *iparm, double *dparm)
+  {
+    if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
+    if (nbrhs == 0) {x = NULL; nbrhs=1;}
+    c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<PASTIX_COMPLEX*>(vals), perm, invp, reinterpret_cast<PASTIX_COMPLEX*>(x), nbrhs, iparm, dparm); 
+  }
+  
+  void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<double> *vals, int *perm, int * invp, std::complex<double> *x, int nbrhs, int *iparm, double *dparm)
+  {
+    if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
+    if (nbrhs == 0) {x = NULL; nbrhs=1;}
+    z_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<PASTIX_DCOMPLEX*>(vals), perm, invp, reinterpret_cast<PASTIX_DCOMPLEX*>(x), nbrhs, iparm, dparm); 
+  }
+
+  // Convert the matrix  to Fortran-style Numbering
+  template <typename MatrixType>
+  void c_to_fortran_numbering (MatrixType& mat)
+  {
+    if ( !(mat.outerIndexPtr()[0]) ) 
+    { 
+      int i;
+      for(i = 0; i <= mat.rows(); ++i)
+        ++mat.outerIndexPtr()[i];
+      for(i = 0; i < mat.nonZeros(); ++i)
+        ++mat.innerIndexPtr()[i];
+    }
+  }
+  
+  // Convert to C-style Numbering
+  template <typename MatrixType>
+  void fortran_to_c_numbering (MatrixType& mat)
+  {
+    // Check the Numbering
+    if ( mat.outerIndexPtr()[0] == 1 ) 
+    { // Convert to C-style numbering
+      int i;
+      for(i = 0; i <= mat.rows(); ++i)
+        --mat.outerIndexPtr()[i];
+      for(i = 0; i < mat.nonZeros(); ++i)
+        --mat.innerIndexPtr()[i];
+    }
+  }
+}
+
+// This is the base class to interface with PaStiX functions. 
+// Users should not used this class directly. 
+template <class Derived>
+class PastixBase : public SparseSolverBase<Derived>
+{
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
+  public:
+    using Base::_solve_impl;
+    
+    typedef typename internal::pastix_traits<Derived>::MatrixType _MatrixType;
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef SparseMatrix<Scalar, ColMajor> ColSpMatrix;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    
+  public:
+    
+    PastixBase() : m_initisOk(false), m_analysisIsOk(false), m_factorizationIsOk(false), m_pastixdata(0), m_size(0)
+    {
+      init();
+    }
+    
+    ~PastixBase() 
+    {
+      clean();
+    }
+    
+    template<typename Rhs,typename Dest>
+    bool _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &x) const;
+    
+    /** Returns a reference to the integer vector IPARM of PaStiX parameters
+      * to modify the default parameters. 
+      * The statistics related to the different phases of factorization and solve are saved here as well
+      * \sa analyzePattern() factorize()
+      */
+    Array<StorageIndex,IPARM_SIZE,1>& iparm()
+    {
+      return m_iparm; 
+    }
+    
+    /** Return a reference to a particular index parameter of the IPARM vector 
+     * \sa iparm()
+     */
+    
+    int& iparm(int idxparam)
+    {
+      return m_iparm(idxparam);
+    }
+    
+     /** Returns a reference to the double vector DPARM of PaStiX parameters 
+      * The statistics related to the different phases of factorization and solve are saved here as well
+      * \sa analyzePattern() factorize()
+      */
+    Array<double,DPARM_SIZE,1>& dparm()
+    {
+      return m_dparm; 
+    }
+    
+    
+    /** Return a reference to a particular index parameter of the DPARM vector 
+     * \sa dparm()
+     */
+    double& dparm(int idxparam)
+    {
+      return m_dparm(idxparam);
+    }
+    
+    inline Index cols() const { return m_size; }
+    inline Index rows() const { return m_size; }
+    
+     /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the PaStiX reports a problem
+      *          \c InvalidInput if the input matrix is invalid
+      *
+      * \sa iparm()          
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+    
+  protected:
+
+    // Initialize the Pastix data structure, check the matrix
+    void init(); 
+    
+    // Compute the ordering and the symbolic factorization
+    void analyzePattern(ColSpMatrix& mat);
+    
+    // Compute the numerical factorization
+    void factorize(ColSpMatrix& mat);
+    
+    // Free all the data allocated by Pastix
+    void clean()
+    {
+      eigen_assert(m_initisOk && "The Pastix structure should be allocated first"); 
+      m_iparm(IPARM_START_TASK) = API_TASK_CLEAN;
+      m_iparm(IPARM_END_TASK) = API_TASK_CLEAN;
+      internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, 0, 0, 0, (Scalar*)0,
+                             m_perm.data(), m_invp.data(), 0, 0, m_iparm.data(), m_dparm.data());
+    }
+    
+    void compute(ColSpMatrix& mat);
+    
+    int m_initisOk; 
+    int m_analysisIsOk;
+    int m_factorizationIsOk;
+    mutable ComputationInfo m_info; 
+    mutable pastix_data_t *m_pastixdata; // Data structure for pastix
+    mutable int m_comm; // The MPI communicator identifier
+    mutable Array<int,IPARM_SIZE,1> m_iparm; // integer vector for the input parameters
+    mutable Array<double,DPARM_SIZE,1> m_dparm; // Scalar vector for the input parameters
+    mutable Matrix<StorageIndex,Dynamic,1> m_perm;  // Permutation vector
+    mutable Matrix<StorageIndex,Dynamic,1> m_invp;  // Inverse permutation vector
+    mutable int m_size; // Size of the matrix 
+}; 
+
+ /** Initialize the PaStiX data structure. 
+   *A first call to this function fills iparm and dparm with the default PaStiX parameters
+   * \sa iparm() dparm()
+   */
+template <class Derived>
+void PastixBase<Derived>::init()
+{
+  m_size = 0; 
+  m_iparm.setZero(IPARM_SIZE);
+  m_dparm.setZero(DPARM_SIZE);
+  
+  m_iparm(IPARM_MODIFY_PARAMETER) = API_NO;
+  pastix(&m_pastixdata, MPI_COMM_WORLD,
+         0, 0, 0, 0,
+         0, 0, 0, 1, m_iparm.data(), m_dparm.data());
+  
+  m_iparm[IPARM_MATRIX_VERIFICATION] = API_NO;
+  m_iparm[IPARM_VERBOSE]             = API_VERBOSE_NOT;
+  m_iparm[IPARM_ORDERING]            = API_ORDER_SCOTCH;
+  m_iparm[IPARM_INCOMPLETE]          = API_NO;
+  m_iparm[IPARM_OOC_LIMIT]           = 2000;
+  m_iparm[IPARM_RHS_MAKING]          = API_RHS_B;
+  m_iparm(IPARM_MATRIX_VERIFICATION) = API_NO;
+  
+  m_iparm(IPARM_START_TASK) = API_TASK_INIT;
+  m_iparm(IPARM_END_TASK) = API_TASK_INIT;
+  internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, 0, 0, 0, (Scalar*)0,
+                         0, 0, 0, 0, m_iparm.data(), m_dparm.data());
+  
+  // Check the returned error
+  if(m_iparm(IPARM_ERROR_NUMBER)) {
+    m_info = InvalidInput;
+    m_initisOk = false;
+  }
+  else { 
+    m_info = Success;
+    m_initisOk = true;
+  }
+}
+
+template <class Derived>
+void PastixBase<Derived>::compute(ColSpMatrix& mat)
+{
+  eigen_assert(mat.rows() == mat.cols() && "The input matrix should be squared");
+  
+  analyzePattern(mat);  
+  factorize(mat);
+  
+  m_iparm(IPARM_MATRIX_VERIFICATION) = API_NO;
+}
+
+
+template <class Derived>
+void PastixBase<Derived>::analyzePattern(ColSpMatrix& mat)
+{                         
+  eigen_assert(m_initisOk && "The initialization of PaSTiX failed");
+  
+  // clean previous calls
+  if(m_size>0)
+    clean();
+  
+  m_size = internal::convert_index<int>(mat.rows());
+  m_perm.resize(m_size);
+  m_invp.resize(m_size);
+  
+  m_iparm(IPARM_START_TASK) = API_TASK_ORDERING;
+  m_iparm(IPARM_END_TASK) = API_TASK_ANALYSE;
+  internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, m_size, mat.outerIndexPtr(), mat.innerIndexPtr(),
+               mat.valuePtr(), m_perm.data(), m_invp.data(), 0, 0, m_iparm.data(), m_dparm.data());
+  
+  // Check the returned error
+  if(m_iparm(IPARM_ERROR_NUMBER))
+  {
+    m_info = NumericalIssue;
+    m_analysisIsOk = false;
+  }
+  else
+  { 
+    m_info = Success;
+    m_analysisIsOk = true;
+  }
+}
+
+template <class Derived>
+void PastixBase<Derived>::factorize(ColSpMatrix& mat)
+{
+//   if(&m_cpyMat != &mat) m_cpyMat = mat;
+  eigen_assert(m_analysisIsOk && "The analysis phase should be called before the factorization phase");
+  m_iparm(IPARM_START_TASK) = API_TASK_NUMFACT;
+  m_iparm(IPARM_END_TASK) = API_TASK_NUMFACT;
+  m_size = internal::convert_index<int>(mat.rows());
+  
+  internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, m_size, mat.outerIndexPtr(), mat.innerIndexPtr(),
+               mat.valuePtr(), m_perm.data(), m_invp.data(), 0, 0, m_iparm.data(), m_dparm.data());
+  
+  // Check the returned error
+  if(m_iparm(IPARM_ERROR_NUMBER))
+  {
+    m_info = NumericalIssue;
+    m_factorizationIsOk = false;
+    m_isInitialized = false;
+  }
+  else
+  {
+    m_info = Success;
+    m_factorizationIsOk = true;
+    m_isInitialized = true;
+  }
+}
+
+/* Solve the system */
+template<typename Base>
+template<typename Rhs,typename Dest>
+bool PastixBase<Base>::_solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &x) const
+{
+  eigen_assert(m_isInitialized && "The matrix should be factorized first");
+  EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0,
+                     THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+  int rhs = 1;
+  
+  x = b; /* on return, x is overwritten by the computed solution */
+  
+  for (int i = 0; i < b.cols(); i++){
+    m_iparm[IPARM_START_TASK]          = API_TASK_SOLVE;
+    m_iparm[IPARM_END_TASK]            = API_TASK_REFINE;
+  
+    internal::eigen_pastix(&m_pastixdata, MPI_COMM_WORLD, internal::convert_index<int>(x.rows()), 0, 0, 0,
+                           m_perm.data(), m_invp.data(), &x(0, i), rhs, m_iparm.data(), m_dparm.data());
+  }
+  
+  // Check the returned error
+  m_info = m_iparm(IPARM_ERROR_NUMBER)==0 ? Success : NumericalIssue;
+  
+  return m_iparm(IPARM_ERROR_NUMBER)==0;
+}
+
+/** \ingroup PaStiXSupport_Module
+  * \class PastixLU
+  * \brief Sparse direct LU solver based on PaStiX library
+  * 
+  * This class is used to solve the linear systems A.X = B with a supernodal LU 
+  * factorization in the PaStiX library. The matrix A should be squared and nonsingular
+  * PaStiX requires that the matrix A has a symmetric structural pattern. 
+  * This interface can symmetrize the input matrix otherwise. 
+  * The vectors or matrices X and B can be either dense or sparse.
+  * 
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam IsStrSym Indicates if the input matrix has a symmetric pattern, default is false
+  * NOTE : Note that if the analysis and factorization phase are called separately, 
+  * the input matrix will be symmetrized at each call, hence it is advised to 
+  * symmetrize the matrix in a end-user program and set \p IsStrSym to true
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  * 
+  */
+template<typename _MatrixType, bool IsStrSym>
+class PastixLU : public PastixBase< PastixLU<_MatrixType> >
+{
+  public:
+    typedef _MatrixType MatrixType;
+    typedef PastixBase<PastixLU<MatrixType> > Base;
+    typedef typename Base::ColSpMatrix ColSpMatrix;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    
+  public:
+    PastixLU() : Base()
+    {
+      init();
+    }
+    
+    explicit PastixLU(const MatrixType& matrix):Base()
+    {
+      init();
+      compute(matrix);
+    }
+    /** Compute the LU supernodal factorization of \p matrix. 
+      * iparm and dparm can be used to tune the PaStiX parameters. 
+      * see the PaStiX user's manual
+      * \sa analyzePattern() factorize()
+      */
+    void compute (const MatrixType& matrix)
+    {
+      m_structureIsUptodate = false;
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::compute(temp);
+    }
+    /** Compute the LU symbolic factorization of \p matrix using its sparsity pattern. 
+      * Several ordering methods can be used at this step. See the PaStiX user's manual. 
+      * The result of this operation can be used with successive matrices having the same pattern as \p matrix
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    {
+      m_structureIsUptodate = false;
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::analyzePattern(temp);
+    }
+
+    /** Compute the LU supernodal factorization of \p matrix
+      * WARNING The matrix \p matrix should have the same structural pattern 
+      * as the same used in the analysis phase.
+      * \sa analyzePattern()
+      */ 
+    void factorize(const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::factorize(temp);
+    }
+  protected:
+    
+    void init()
+    {
+      m_structureIsUptodate = false;
+      m_iparm(IPARM_SYM) = API_SYM_NO;
+      m_iparm(IPARM_FACTORIZATION) = API_FACT_LU;
+    }
+    
+    void grabMatrix(const MatrixType& matrix, ColSpMatrix& out)
+    {
+      if(IsStrSym)
+        out = matrix;
+      else
+      {
+        if(!m_structureIsUptodate)
+        {
+          // update the transposed structure
+          m_transposedStructure = matrix.transpose();
+          
+          // Set the elements of the matrix to zero 
+          for (Index j=0; j<m_transposedStructure.outerSize(); ++j) 
+            for(typename ColSpMatrix::InnerIterator it(m_transposedStructure, j); it; ++it)
+              it.valueRef() = 0.0;
+
+          m_structureIsUptodate = true;
+        }
+        
+        out = m_transposedStructure + matrix;
+      }
+      internal::c_to_fortran_numbering(out);
+    }
+    
+    using Base::m_iparm;
+    using Base::m_dparm;
+    
+    ColSpMatrix m_transposedStructure;
+    bool m_structureIsUptodate;
+};
+
+/** \ingroup PaStiXSupport_Module
+  * \class PastixLLT
+  * \brief A sparse direct supernodal Cholesky (LLT) factorization and solver based on the PaStiX library
+  * 
+  * This class is used to solve the linear systems A.X = B via a LL^T supernodal Cholesky factorization
+  * available in the PaStiX library. The matrix A should be symmetric and positive definite
+  * WARNING Selfadjoint complex matrices are not supported in the current version of PaStiX
+  * The vectors or matrices X and B can be either dense or sparse
+  * 
+  * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam UpLo The part of the matrix to use : Lower or Upper. The default is Lower as required by PaStiX
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SimplicialLLT
+  */
+template<typename _MatrixType, int _UpLo>
+class PastixLLT : public PastixBase< PastixLLT<_MatrixType, _UpLo> >
+{
+  public:
+    typedef _MatrixType MatrixType;
+    typedef PastixBase<PastixLLT<MatrixType, _UpLo> > Base;
+    typedef typename Base::ColSpMatrix ColSpMatrix;
+    
+  public:
+    enum { UpLo = _UpLo };
+    PastixLLT() : Base()
+    {
+      init();
+    }
+    
+    explicit PastixLLT(const MatrixType& matrix):Base()
+    {
+      init();
+      compute(matrix);
+    }
+
+    /** Compute the L factor of the LL^T supernodal factorization of \p matrix 
+      * \sa analyzePattern() factorize()
+      */
+    void compute (const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::compute(temp);
+    }
+
+     /** Compute the LL^T symbolic factorization of \p matrix using its sparsity pattern
+      * The result of this operation can be used with successive matrices having the same pattern as \p matrix
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::analyzePattern(temp);
+    }
+      /** Compute the LL^T supernodal numerical factorization of \p matrix 
+        * \sa analyzePattern()
+        */
+    void factorize(const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::factorize(temp);
+    }
+  protected:
+    using Base::m_iparm;
+    
+    void init()
+    {
+      m_iparm(IPARM_SYM) = API_SYM_YES;
+      m_iparm(IPARM_FACTORIZATION) = API_FACT_LLT;
+    }
+    
+    void grabMatrix(const MatrixType& matrix, ColSpMatrix& out)
+    {
+      out.resize(matrix.rows(), matrix.cols());
+      // Pastix supports only lower, column-major matrices 
+      out.template selfadjointView<Lower>() = matrix.template selfadjointView<UpLo>();
+      internal::c_to_fortran_numbering(out);
+    }
+};
+
+/** \ingroup PaStiXSupport_Module
+  * \class PastixLDLT
+  * \brief A sparse direct supernodal Cholesky (LLT) factorization and solver based on the PaStiX library
+  * 
+  * This class is used to solve the linear systems A.X = B via a LDL^T supernodal Cholesky factorization
+  * available in the PaStiX library. The matrix A should be symmetric and positive definite
+  * WARNING Selfadjoint complex matrices are not supported in the current version of PaStiX
+  * The vectors or matrices X and B can be either dense or sparse
+  * 
+  * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam UpLo The part of the matrix to use : Lower or Upper. The default is Lower as required by PaStiX
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SimplicialLDLT
+  */
+template<typename _MatrixType, int _UpLo>
+class PastixLDLT : public PastixBase< PastixLDLT<_MatrixType, _UpLo> >
+{
+  public:
+    typedef _MatrixType MatrixType;
+    typedef PastixBase<PastixLDLT<MatrixType, _UpLo> > Base; 
+    typedef typename Base::ColSpMatrix ColSpMatrix;
+    
+  public:
+    enum { UpLo = _UpLo };
+    PastixLDLT():Base()
+    {
+      init();
+    }
+    
+    explicit PastixLDLT(const MatrixType& matrix):Base()
+    {
+      init();
+      compute(matrix);
+    }
+
+    /** Compute the L and D factors of the LDL^T factorization of \p matrix 
+      * \sa analyzePattern() factorize()
+      */
+    void compute (const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::compute(temp);
+    }
+
+    /** Compute the LDL^T symbolic factorization of \p matrix using its sparsity pattern
+      * The result of this operation can be used with successive matrices having the same pattern as \p matrix
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    { 
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::analyzePattern(temp);
+    }
+    /** Compute the LDL^T supernodal numerical factorization of \p matrix 
+      * 
+      */
+    void factorize(const MatrixType& matrix)
+    {
+      ColSpMatrix temp;
+      grabMatrix(matrix, temp);
+      Base::factorize(temp);
+    }
+
+  protected:
+    using Base::m_iparm;
+    
+    void init()
+    {
+      m_iparm(IPARM_SYM) = API_SYM_YES;
+      m_iparm(IPARM_FACTORIZATION) = API_FACT_LDLT;
+    }
+    
+    void grabMatrix(const MatrixType& matrix, ColSpMatrix& out)
+    {
+      // Pastix supports only lower, column-major matrices 
+      out.resize(matrix.rows(), matrix.cols());
+      out.template selfadjointView<Lower>() = matrix.template selfadjointView<UpLo>();
+      internal::c_to_fortran_numbering(out);
+    }
+};
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PardisoSupport/PardisoSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PardisoSupport/PardisoSupport.h
new file mode 100644
index 0000000..091c397
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/PardisoSupport/PardisoSupport.h
@@ -0,0 +1,543 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to Intel(R) MKL PARDISO
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_PARDISOSUPPORT_H
+#define EIGEN_PARDISOSUPPORT_H
+
+namespace Eigen { 
+
+template<typename _MatrixType> class PardisoLU;
+template<typename _MatrixType, int Options=Upper> class PardisoLLT;
+template<typename _MatrixType, int Options=Upper> class PardisoLDLT;
+
+namespace internal
+{
+  template<typename IndexType>
+  struct pardiso_run_selector
+  {
+    static IndexType run( _MKL_DSS_HANDLE_t pt, IndexType maxfct, IndexType mnum, IndexType type, IndexType phase, IndexType n, void *a,
+                      IndexType *ia, IndexType *ja, IndexType *perm, IndexType nrhs, IndexType *iparm, IndexType msglvl, void *b, void *x)
+    {
+      IndexType error = 0;
+      ::pardiso(pt, &maxfct, &mnum, &type, &phase, &n, a, ia, ja, perm, &nrhs, iparm, &msglvl, b, x, &error);
+      return error;
+    }
+  };
+  template<>
+  struct pardiso_run_selector<long long int>
+  {
+    typedef long long int IndexType;
+    static IndexType run( _MKL_DSS_HANDLE_t pt, IndexType maxfct, IndexType mnum, IndexType type, IndexType phase, IndexType n, void *a,
+                      IndexType *ia, IndexType *ja, IndexType *perm, IndexType nrhs, IndexType *iparm, IndexType msglvl, void *b, void *x)
+    {
+      IndexType error = 0;
+      ::pardiso_64(pt, &maxfct, &mnum, &type, &phase, &n, a, ia, ja, perm, &nrhs, iparm, &msglvl, b, x, &error);
+      return error;
+    }
+  };
+
+  template<class Pardiso> struct pardiso_traits;
+
+  template<typename _MatrixType>
+  struct pardiso_traits< PardisoLU<_MatrixType> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;
+  };
+
+  template<typename _MatrixType, int Options>
+  struct pardiso_traits< PardisoLLT<_MatrixType, Options> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;
+  };
+
+  template<typename _MatrixType, int Options>
+  struct pardiso_traits< PardisoLDLT<_MatrixType, Options> >
+  {
+    typedef _MatrixType MatrixType;
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef typename _MatrixType::StorageIndex StorageIndex;    
+  };
+
+} // end namespace internal
+
+template<class Derived>
+class PardisoImpl : public SparseSolverBase<Derived>
+{
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
+    
+    typedef internal::pardiso_traits<Derived> Traits;
+  public:
+    using Base::_solve_impl;
+    
+    typedef typename Traits::MatrixType MatrixType;
+    typedef typename Traits::Scalar Scalar;
+    typedef typename Traits::RealScalar RealScalar;
+    typedef typename Traits::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,RowMajor,StorageIndex> SparseMatrixType;
+    typedef Matrix<Scalar,Dynamic,1> VectorType;
+    typedef Matrix<StorageIndex, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
+    typedef Matrix<StorageIndex, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
+    typedef Array<StorageIndex,64,1,DontAlign> ParameterType;
+    enum {
+      ScalarIsComplex = NumTraits<Scalar>::IsComplex,
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
+
+    PardisoImpl()
+    {
+      eigen_assert((sizeof(StorageIndex) >= sizeof(_INTEGER_t) && sizeof(StorageIndex) <= 8) && "Non-supported index type");
+      m_iparm.setZero();
+      m_msglvl = 0; // No output
+      m_isInitialized = false;
+    }
+
+    ~PardisoImpl()
+    {
+      pardisoRelease();
+    }
+
+    inline Index cols() const { return m_size; }
+    inline Index rows() const { return m_size; }
+  
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+    /** \warning for advanced usage only.
+      * \returns a reference to the parameter array controlling PARDISO.
+      * See the PARDISO manual to know how to use it. */
+    ParameterType& pardisoParameterArray()
+    {
+      return m_iparm;
+    }
+    
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      * 
+      * \sa factorize()
+      */
+    Derived& analyzePattern(const MatrixType& matrix);
+    
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    Derived& factorize(const MatrixType& matrix);
+
+    Derived& compute(const MatrixType& matrix);
+
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const;
+
+  protected:
+    void pardisoRelease()
+    {
+      if(m_isInitialized) // Factorization ran at least once
+      {
+        internal::pardiso_run_selector<StorageIndex>::run(m_pt, 1, 1, m_type, -1, internal::convert_index<StorageIndex>(m_size),0, 0, 0, m_perm.data(), 0,
+                                                          m_iparm.data(), m_msglvl, NULL, NULL);
+        m_isInitialized = false;
+      }
+    }
+
+    void pardisoInit(int type)
+    {
+      m_type = type;
+      bool symmetric = std::abs(m_type) < 10;
+      m_iparm[0] = 1;   // No solver default
+      m_iparm[1] = 2;   // use Metis for the ordering
+      m_iparm[2] = 0;   // Reserved. Set to zero. (??Numbers of processors, value of OMP_NUM_THREADS??)
+      m_iparm[3] = 0;   // No iterative-direct algorithm
+      m_iparm[4] = 0;   // No user fill-in reducing permutation
+      m_iparm[5] = 0;   // Write solution into x, b is left unchanged
+      m_iparm[6] = 0;   // Not in use
+      m_iparm[7] = 2;   // Max numbers of iterative refinement steps
+      m_iparm[8] = 0;   // Not in use
+      m_iparm[9] = 13;  // Perturb the pivot elements with 1E-13
+      m_iparm[10] = symmetric ? 0 : 1; // Use nonsymmetric permutation and scaling MPS
+      m_iparm[11] = 0;  // Not in use
+      m_iparm[12] = symmetric ? 0 : 1;  // Maximum weighted matching algorithm is switched-off (default for symmetric).
+                                        // Try m_iparm[12] = 1 in case of inappropriate accuracy
+      m_iparm[13] = 0;  // Output: Number of perturbed pivots
+      m_iparm[14] = 0;  // Not in use
+      m_iparm[15] = 0;  // Not in use
+      m_iparm[16] = 0;  // Not in use
+      m_iparm[17] = -1; // Output: Number of nonzeros in the factor LU
+      m_iparm[18] = -1; // Output: Mflops for LU factorization
+      m_iparm[19] = 0;  // Output: Numbers of CG Iterations
+      
+      m_iparm[20] = 0;  // 1x1 pivoting
+      m_iparm[26] = 0;  // No matrix checker
+      m_iparm[27] = (sizeof(RealScalar) == 4) ? 1 : 0;
+      m_iparm[34] = 1;  // C indexing
+      m_iparm[36] = 0;  // CSR
+      m_iparm[59] = 0;  // 0 - In-Core ; 1 - Automatic switch between In-Core and Out-of-Core modes ; 2 - Out-of-Core
+      
+      memset(m_pt, 0, sizeof(m_pt));
+    }
+
+  protected:
+    // cached data to reduce reallocation, etc.
+    
+    void manageErrorCode(Index error) const
+    {
+      switch(error)
+      {
+        case 0:
+          m_info = Success;
+          break;
+        case -4:
+        case -7:
+          m_info = NumericalIssue;
+          break;
+        default:
+          m_info = InvalidInput;
+      }
+    }
+
+    mutable SparseMatrixType m_matrix;
+    mutable ComputationInfo m_info;
+    bool m_analysisIsOk, m_factorizationIsOk;
+    StorageIndex m_type, m_msglvl;
+    mutable void *m_pt[64];
+    mutable ParameterType m_iparm;
+    mutable IntColVectorType m_perm;
+    Index m_size;
+    
+};
+
+template<class Derived>
+Derived& PardisoImpl<Derived>::compute(const MatrixType& a)
+{
+  m_size = a.rows();
+  eigen_assert(a.rows() == a.cols());
+
+  pardisoRelease();
+  m_perm.setZero(m_size);
+  derived().getMatrix(a);
+  
+  Index error;
+  error = internal::pardiso_run_selector<StorageIndex>::run(m_pt, 1, 1, m_type, 12, internal::convert_index<StorageIndex>(m_size),
+                                                            m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(),
+                                                            m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL);
+  manageErrorCode(error);
+  m_analysisIsOk = true;
+  m_factorizationIsOk = true;
+  m_isInitialized = true;
+  return derived();
+}
+
+template<class Derived>
+Derived& PardisoImpl<Derived>::analyzePattern(const MatrixType& a)
+{
+  m_size = a.rows();
+  eigen_assert(m_size == a.cols());
+
+  pardisoRelease();
+  m_perm.setZero(m_size);
+  derived().getMatrix(a);
+  
+  Index error;
+  error = internal::pardiso_run_selector<StorageIndex>::run(m_pt, 1, 1, m_type, 11, internal::convert_index<StorageIndex>(m_size),
+                                                            m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(),
+                                                            m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL);
+  
+  manageErrorCode(error);
+  m_analysisIsOk = true;
+  m_factorizationIsOk = false;
+  m_isInitialized = true;
+  return derived();
+}
+
+template<class Derived>
+Derived& PardisoImpl<Derived>::factorize(const MatrixType& a)
+{
+  eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+  eigen_assert(m_size == a.rows() && m_size == a.cols());
+  
+  derived().getMatrix(a);
+
+  Index error;
+  error = internal::pardiso_run_selector<StorageIndex>::run(m_pt, 1, 1, m_type, 22, internal::convert_index<StorageIndex>(m_size),
+                                                            m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(),
+                                                            m_perm.data(), 0, m_iparm.data(), m_msglvl, NULL, NULL);
+  
+  manageErrorCode(error);
+  m_factorizationIsOk = true;
+  return derived();
+}
+
+template<class Derived>
+template<typename BDerived,typename XDerived>
+void PardisoImpl<Derived>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived>& x) const
+{
+  if(m_iparm[0] == 0) // Factorization was not computed
+  {
+    m_info = InvalidInput;
+    return;
+  }
+
+  //Index n = m_matrix.rows();
+  Index nrhs = Index(b.cols());
+  eigen_assert(m_size==b.rows());
+  eigen_assert(((MatrixBase<BDerived>::Flags & RowMajorBit) == 0 || nrhs == 1) && "Row-major right hand sides are not supported");
+  eigen_assert(((MatrixBase<XDerived>::Flags & RowMajorBit) == 0 || nrhs == 1) && "Row-major matrices of unknowns are not supported");
+  eigen_assert(((nrhs == 1) || b.outerStride() == b.rows()));
+
+
+//  switch (transposed) {
+//    case SvNoTrans    : m_iparm[11] = 0 ; break;
+//    case SvTranspose  : m_iparm[11] = 2 ; break;
+//    case SvAdjoint    : m_iparm[11] = 1 ; break;
+//    default:
+//      //std::cerr << "Eigen: transposition  option \"" << transposed << "\" not supported by the PARDISO backend\n";
+//      m_iparm[11] = 0;
+//  }
+
+  Scalar* rhs_ptr = const_cast<Scalar*>(b.derived().data());
+  Matrix<Scalar,Dynamic,Dynamic,ColMajor> tmp;
+  
+  // Pardiso cannot solve in-place
+  if(rhs_ptr == x.derived().data())
+  {
+    tmp = b;
+    rhs_ptr = tmp.data();
+  }
+  
+  Index error;
+  error = internal::pardiso_run_selector<StorageIndex>::run(m_pt, 1, 1, m_type, 33, internal::convert_index<StorageIndex>(m_size),
+                                                            m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(),
+                                                            m_perm.data(), internal::convert_index<StorageIndex>(nrhs), m_iparm.data(), m_msglvl,
+                                                            rhs_ptr, x.derived().data());
+
+  manageErrorCode(error);
+}
+
+
+/** \ingroup PardisoSupport_Module
+  * \class PardisoLU
+  * \brief A sparse direct LU factorization and solver based on the PARDISO library
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a direct LU factorization
+  * using the Intel MKL PARDISO library. The sparse matrix A must be squared and invertible.
+  * The vectors or matrices X and B can be either dense or sparse.
+  *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  */
+template<typename MatrixType>
+class PardisoLU : public PardisoImpl< PardisoLU<MatrixType> >
+{
+  protected:
+    typedef PardisoImpl<PardisoLU> Base;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::RealScalar RealScalar;
+    using Base::pardisoInit;
+    using Base::m_matrix;
+    friend class PardisoImpl< PardisoLU<MatrixType> >;
+
+  public:
+
+    using Base::compute;
+    using Base::solve;
+
+    PardisoLU()
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? 13 : 11);
+    }
+
+    explicit PardisoLU(const MatrixType& matrix)
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? 13 : 11);
+      compute(matrix);
+    }
+  protected:
+    void getMatrix(const MatrixType& matrix)
+    {
+      m_matrix = matrix;
+      m_matrix.makeCompressed();
+    }
+};
+
+/** \ingroup PardisoSupport_Module
+  * \class PardisoLLT
+  * \brief A sparse direct Cholesky (LLT) factorization and solver based on the PARDISO library
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a LL^T Cholesky factorization
+  * using the Intel MKL PARDISO library. The sparse matrix A must be selfajoint and positive definite.
+  * The vectors or matrices X and B can be either dense or sparse.
+  *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
+  * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam UpLo can be any bitwise combination of Upper, Lower. The default is Upper, meaning only the upper triangular part has to be used.
+  *         Upper|Lower can be used to tell both triangular parts can be used as input.
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SimplicialLLT
+  */
+template<typename MatrixType, int _UpLo>
+class PardisoLLT : public PardisoImpl< PardisoLLT<MatrixType,_UpLo> >
+{
+  protected:
+    typedef PardisoImpl< PardisoLLT<MatrixType,_UpLo> > Base;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::RealScalar RealScalar;
+    using Base::pardisoInit;
+    using Base::m_matrix;
+    friend class PardisoImpl< PardisoLLT<MatrixType,_UpLo> >;
+
+  public:
+
+    typedef typename Base::StorageIndex StorageIndex;
+    enum { UpLo = _UpLo };
+    using Base::compute;
+
+    PardisoLLT()
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? 4 : 2);
+    }
+
+    explicit PardisoLLT(const MatrixType& matrix)
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? 4 : 2);
+      compute(matrix);
+    }
+    
+  protected:
+    
+    void getMatrix(const MatrixType& matrix)
+    {
+      // PARDISO supports only upper, row-major matrices
+      PermutationMatrix<Dynamic,Dynamic,StorageIndex> p_null;
+      m_matrix.resize(matrix.rows(), matrix.cols());
+      m_matrix.template selfadjointView<Upper>() = matrix.template selfadjointView<UpLo>().twistedBy(p_null);
+      m_matrix.makeCompressed();
+    }
+};
+
+/** \ingroup PardisoSupport_Module
+  * \class PardisoLDLT
+  * \brief A sparse direct Cholesky (LDLT) factorization and solver based on the PARDISO library
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a LDL^T Cholesky factorization
+  * using the Intel MKL PARDISO library. The sparse matrix A is assumed to be selfajoint and positive definite.
+  * For complex matrices, A can also be symmetric only, see the \a Options template parameter.
+  * The vectors or matrices X and B can be either dense or sparse.
+  *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
+  * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam Options can be any bitwise combination of Upper, Lower, and Symmetric. The default is Upper, meaning only the upper triangular part has to be used.
+  *         Symmetric can be used for symmetric, non-selfadjoint complex matrices, the default being to assume a selfadjoint matrix.
+  *         Upper|Lower can be used to tell both triangular parts can be used as input.
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SimplicialLDLT
+  */
+template<typename MatrixType, int Options>
+class PardisoLDLT : public PardisoImpl< PardisoLDLT<MatrixType,Options> >
+{
+  protected:
+    typedef PardisoImpl< PardisoLDLT<MatrixType,Options> > Base;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::RealScalar RealScalar;
+    using Base::pardisoInit;
+    using Base::m_matrix;
+    friend class PardisoImpl< PardisoLDLT<MatrixType,Options> >;
+
+  public:
+
+    typedef typename Base::StorageIndex StorageIndex;
+    using Base::compute;
+    enum { UpLo = Options&(Upper|Lower) };
+
+    PardisoLDLT()
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? ( bool(Options&Symmetric) ? 6 : -4 ) : -2);
+    }
+
+    explicit PardisoLDLT(const MatrixType& matrix)
+      : Base()
+    {
+      pardisoInit(Base::ScalarIsComplex ? ( bool(Options&Symmetric) ? 6 : -4 ) : -2);
+      compute(matrix);
+    }
+    
+    void getMatrix(const MatrixType& matrix)
+    {
+      // PARDISO supports only upper, row-major matrices
+      PermutationMatrix<Dynamic,Dynamic,StorageIndex> p_null;
+      m_matrix.resize(matrix.rows(), matrix.cols());
+      m_matrix.template selfadjointView<Upper>() = matrix.template selfadjointView<UpLo>().twistedBy(p_null);
+      m_matrix.makeCompressed();
+    }
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_PARDISOSUPPORT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR.h
new file mode 100644
index 0000000..a7b47d5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR.h
@@ -0,0 +1,653 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COLPIVOTINGHOUSEHOLDERQR_H
+#define EIGEN_COLPIVOTINGHOUSEHOLDERQR_H
+
+namespace Eigen {
+
+namespace internal {
+template<typename _MatrixType> struct traits<ColPivHouseholderQR<_MatrixType> >
+ : traits<_MatrixType>
+{
+  enum { Flags = 0 };
+};
+
+} // end namespace internal
+
+/** \ingroup QR_Module
+  *
+  * \class ColPivHouseholderQR
+  *
+  * \brief Householder rank-revealing QR decomposition of a matrix with column-pivoting
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the QR decomposition
+  *
+  * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b Q and \b R
+  * such that
+  * \f[
+  *  \mathbf{A} \, \mathbf{P} = \mathbf{Q} \, \mathbf{R}
+  * \f]
+  * by using Householder transformations. Here, \b P is a permutation matrix, \b Q a unitary matrix and \b R an
+  * upper triangular matrix.
+  *
+  * This decomposition performs column pivoting in order to be rank-revealing and improve
+  * numerical stability. It is slower than HouseholderQR, and faster than FullPivHouseholderQR.
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::colPivHouseholderQr()
+  */
+template<typename _MatrixType> class ColPivHouseholderQR
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    // FIXME should be int
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
+    typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationType;
+    typedef typename internal::plain_row_type<MatrixType, Index>::type IntRowVectorType;
+    typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
+    typedef typename internal::plain_row_type<MatrixType, RealScalar>::type RealRowVectorType;
+    typedef HouseholderSequence<MatrixType,typename internal::remove_all<typename HCoeffsType::ConjugateReturnType>::type> HouseholderSequenceType;
+    typedef typename MatrixType::PlainObject PlainObject;
+
+  private:
+
+    typedef typename PermutationType::StorageIndex PermIndexType;
+
+  public:
+
+    /**
+    * \brief Default Constructor.
+    *
+    * The default constructor is useful in cases in which the user intends to
+    * perform decompositions via ColPivHouseholderQR::compute(const MatrixType&).
+    */
+    ColPivHouseholderQR()
+      : m_qr(),
+        m_hCoeffs(),
+        m_colsPermutation(),
+        m_colsTranspositions(),
+        m_temp(),
+        m_colNormsUpdated(),
+        m_colNormsDirect(),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false) {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa ColPivHouseholderQR()
+      */
+    ColPivHouseholderQR(Index rows, Index cols)
+      : m_qr(rows, cols),
+        m_hCoeffs((std::min)(rows,cols)),
+        m_colsPermutation(PermIndexType(cols)),
+        m_colsTranspositions(cols),
+        m_temp(cols),
+        m_colNormsUpdated(cols),
+        m_colNormsDirect(cols),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false) {}
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This constructor computes the QR factorization of the matrix \a matrix by calling
+      * the method compute(). It is a short cut for:
+      *
+      * \code
+      * ColPivHouseholderQR<MatrixType> qr(matrix.rows(), matrix.cols());
+      * qr.compute(matrix);
+      * \endcode
+      *
+      * \sa compute()
+      */
+    template<typename InputType>
+    explicit ColPivHouseholderQR(const EigenBase<InputType>& matrix)
+      : m_qr(matrix.rows(), matrix.cols()),
+        m_hCoeffs((std::min)(matrix.rows(),matrix.cols())),
+        m_colsPermutation(PermIndexType(matrix.cols())),
+        m_colsTranspositions(matrix.cols()),
+        m_temp(matrix.cols()),
+        m_colNormsUpdated(matrix.cols()),
+        m_colNormsDirect(matrix.cols()),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
+      *
+      * \sa ColPivHouseholderQR(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit ColPivHouseholderQR(EigenBase<InputType>& matrix)
+      : m_qr(matrix.derived()),
+        m_hCoeffs((std::min)(matrix.rows(),matrix.cols())),
+        m_colsPermutation(PermIndexType(matrix.cols())),
+        m_colsTranspositions(matrix.cols()),
+        m_temp(matrix.cols()),
+        m_colNormsUpdated(matrix.cols()),
+        m_colNormsDirect(matrix.cols()),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false)
+    {
+      computeInPlace();
+    }
+
+    /** This method finds a solution x to the equation Ax=b, where A is the matrix of which
+      * *this is the QR decomposition, if any exists.
+      *
+      * \param b the right-hand-side of the equation to solve.
+      *
+      * \returns a solution.
+      *
+      * \note_about_checking_solutions
+      *
+      * \note_about_arbitrary_choice_of_solution
+      *
+      * Example: \include ColPivHouseholderQR_solve.cpp
+      * Output: \verbinclude ColPivHouseholderQR_solve.out
+      */
+    template<typename Rhs>
+    inline const Solve<ColPivHouseholderQR, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return Solve<ColPivHouseholderQR, Rhs>(*this, b.derived());
+    }
+
+    HouseholderSequenceType householderQ() const;
+    HouseholderSequenceType matrixQ() const
+    {
+      return householderQ();
+    }
+
+    /** \returns a reference to the matrix where the Householder QR decomposition is stored
+      */
+    const MatrixType& matrixQR() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return m_qr;
+    }
+
+    /** \returns a reference to the matrix where the result Householder QR is stored
+     * \warning The strict lower part of this matrix contains internal values.
+     * Only the upper triangular part should be referenced. To get it, use
+     * \code matrixR().template triangularView<Upper>() \endcode
+     * For rank-deficient matrices, use
+     * \code
+     * matrixR().topLeftCorner(rank(), rank()).template triangularView<Upper>()
+     * \endcode
+     */
+    const MatrixType& matrixR() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return m_qr;
+    }
+
+    template<typename InputType>
+    ColPivHouseholderQR& compute(const EigenBase<InputType>& matrix);
+
+    /** \returns a const reference to the column permutation matrix */
+    const PermutationType& colsPermutation() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return m_colsPermutation;
+    }
+
+    /** \returns the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      * One way to work around that is to use logAbsDeterminant() instead.
+      *
+      * \sa logAbsDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar absDeterminant() const;
+
+    /** \returns the natural log of the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \note This method is useful to work around the risk of overflow/underflow that's inherent
+      * to determinant computation.
+      *
+      * \sa absDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar logAbsDeterminant() const;
+
+    /** \returns the rank of the matrix of which *this is the QR decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index rank() const
+    {
+      using std::abs;
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold();
+      Index result = 0;
+      for(Index i = 0; i < m_nonzero_pivots; ++i)
+        result += (abs(m_qr.coeff(i,i)) > premultiplied_threshold);
+      return result;
+    }
+
+    /** \returns the dimension of the kernel of the matrix of which *this is the QR decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index dimensionOfKernel() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return cols() - rank();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition represents an injective
+      *          linear map, i.e. has trivial kernel; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInjective() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return rank() == cols();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition represents a surjective
+      *          linear map; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isSurjective() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return rank() == rows();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition is invertible.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInvertible() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return isInjective() && isSurjective();
+    }
+
+    /** \returns the inverse of the matrix of which *this is the QR decomposition.
+      *
+      * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
+      *       Use isInvertible() to first determine whether this matrix is invertible.
+      */
+    inline const Inverse<ColPivHouseholderQR> inverse() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return Inverse<ColPivHouseholderQR>(*this);
+    }
+
+    inline Index rows() const { return m_qr.rows(); }
+    inline Index cols() const { return m_qr.cols(); }
+
+    /** \returns a const reference to the vector of Householder coefficients used to represent the factor \c Q.
+      *
+      * For advanced uses only.
+      */
+    const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
+
+    /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
+      * who need to determine when pivots are to be considered nonzero. This is not used for the
+      * QR decomposition itself.
+      *
+      * When it needs to get the threshold value, Eigen calls threshold(). By default, this
+      * uses a formula to automatically determine a reasonable threshold.
+      * Once you have called the present method setThreshold(const RealScalar&),
+      * your value is used instead.
+      *
+      * \param threshold The new value to use as the threshold.
+      *
+      * A pivot will be considered nonzero if its absolute value is strictly greater than
+      *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
+      * where maxpivot is the biggest pivot.
+      *
+      * If you want to come back to the default behavior, call setThreshold(Default_t)
+      */
+    ColPivHouseholderQR& setThreshold(const RealScalar& threshold)
+    {
+      m_usePrescribedThreshold = true;
+      m_prescribedThreshold = threshold;
+      return *this;
+    }
+
+    /** Allows to come back to the default behavior, letting Eigen use its default formula for
+      * determining the threshold.
+      *
+      * You should pass the special object Eigen::Default as parameter here.
+      * \code qr.setThreshold(Eigen::Default); \endcode
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    ColPivHouseholderQR& setThreshold(Default_t)
+    {
+      m_usePrescribedThreshold = false;
+      return *this;
+    }
+
+    /** Returns the threshold that will be used by certain methods such as rank().
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    RealScalar threshold() const
+    {
+      eigen_assert(m_isInitialized || m_usePrescribedThreshold);
+      return m_usePrescribedThreshold ? m_prescribedThreshold
+      // this formula comes from experimenting (see "LU precision tuning" thread on the list)
+      // and turns out to be identical to Higham's formula used already in LDLt.
+                                      : NumTraits<Scalar>::epsilon() * RealScalar(m_qr.diagonalSize());
+    }
+
+    /** \returns the number of nonzero pivots in the QR decomposition.
+      * Here nonzero is meant in the exact sense, not in a fuzzy sense.
+      * So that notion isn't really intrinsically interesting, but it is
+      * still useful when implementing algorithms.
+      *
+      * \sa rank()
+      */
+    inline Index nonzeroPivots() const
+    {
+      eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+      return m_nonzero_pivots;
+    }
+
+    /** \returns the absolute value of the biggest pivot, i.e. the biggest
+      *          diagonal coefficient of R.
+      */
+    RealScalar maxPivot() const { return m_maxpivot; }
+
+    /** \brief Reports whether the QR factorization was succesful.
+      *
+      * \note This function always returns \c Success. It is provided for compatibility
+      * with other factorization routines.
+      * \returns \c Success
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return Success;
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+
+    friend class CompleteOrthogonalDecomposition<MatrixType>;
+
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    void computeInPlace();
+
+    MatrixType m_qr;
+    HCoeffsType m_hCoeffs;
+    PermutationType m_colsPermutation;
+    IntRowVectorType m_colsTranspositions;
+    RowVectorType m_temp;
+    RealRowVectorType m_colNormsUpdated;
+    RealRowVectorType m_colNormsDirect;
+    bool m_isInitialized, m_usePrescribedThreshold;
+    RealScalar m_prescribedThreshold, m_maxpivot;
+    Index m_nonzero_pivots;
+    Index m_det_pq;
+};
+
+template<typename MatrixType>
+typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::absDeterminant() const
+{
+  using std::abs;
+  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return abs(m_qr.diagonal().prod());
+}
+
+template<typename MatrixType>
+typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::logAbsDeterminant() const
+{
+  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return m_qr.diagonal().cwiseAbs().array().log().sum();
+}
+
+/** Performs the QR factorization of the given matrix \a matrix. The result of
+  * the factorization is stored into \c *this, and a reference to \c *this
+  * is returned.
+  *
+  * \sa class ColPivHouseholderQR, ColPivHouseholderQR(const MatrixType&)
+  */
+template<typename MatrixType>
+template<typename InputType>
+ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const EigenBase<InputType>& matrix)
+{
+  m_qr = matrix.derived();
+  computeInPlace();
+  return *this;
+}
+
+template<typename MatrixType>
+void ColPivHouseholderQR<MatrixType>::computeInPlace()
+{
+  check_template_parameters();
+
+  // the column permutation is stored as int indices, so just to be sure:
+  eigen_assert(m_qr.cols()<=NumTraits<int>::highest());
+
+  using std::abs;
+
+  Index rows = m_qr.rows();
+  Index cols = m_qr.cols();
+  Index size = m_qr.diagonalSize();
+
+  m_hCoeffs.resize(size);
+
+  m_temp.resize(cols);
+
+  m_colsTranspositions.resize(m_qr.cols());
+  Index number_of_transpositions = 0;
+
+  m_colNormsUpdated.resize(cols);
+  m_colNormsDirect.resize(cols);
+  for (Index k = 0; k < cols; ++k) {
+    // colNormsDirect(k) caches the most recent directly computed norm of
+    // column k.
+    m_colNormsDirect.coeffRef(k) = m_qr.col(k).norm();
+    m_colNormsUpdated.coeffRef(k) = m_colNormsDirect.coeffRef(k);
+  }
+
+  RealScalar threshold_helper =  numext::abs2<RealScalar>(m_colNormsUpdated.maxCoeff() * NumTraits<RealScalar>::epsilon()) / RealScalar(rows);
+  RealScalar norm_downdate_threshold = numext::sqrt(NumTraits<RealScalar>::epsilon());
+
+  m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
+  m_maxpivot = RealScalar(0);
+
+  for(Index k = 0; k < size; ++k)
+  {
+    // first, we look up in our table m_colNormsUpdated which column has the biggest norm
+    Index biggest_col_index;
+    RealScalar biggest_col_sq_norm = numext::abs2(m_colNormsUpdated.tail(cols-k).maxCoeff(&biggest_col_index));
+    biggest_col_index += k;
+
+    // Track the number of meaningful pivots but do not stop the decomposition to make
+    // sure that the initial matrix is properly reproduced. See bug 941.
+    if(m_nonzero_pivots==size && biggest_col_sq_norm < threshold_helper * RealScalar(rows-k))
+      m_nonzero_pivots = k;
+
+    // apply the transposition to the columns
+    m_colsTranspositions.coeffRef(k) = biggest_col_index;
+    if(k != biggest_col_index) {
+      m_qr.col(k).swap(m_qr.col(biggest_col_index));
+      std::swap(m_colNormsUpdated.coeffRef(k), m_colNormsUpdated.coeffRef(biggest_col_index));
+      std::swap(m_colNormsDirect.coeffRef(k), m_colNormsDirect.coeffRef(biggest_col_index));
+      ++number_of_transpositions;
+    }
+
+    // generate the householder vector, store it below the diagonal
+    RealScalar beta;
+    m_qr.col(k).tail(rows-k).makeHouseholderInPlace(m_hCoeffs.coeffRef(k), beta);
+
+    // apply the householder transformation to the diagonal coefficient
+    m_qr.coeffRef(k,k) = beta;
+
+    // remember the maximum absolute value of diagonal coefficients
+    if(abs(beta) > m_maxpivot) m_maxpivot = abs(beta);
+
+    // apply the householder transformation
+    m_qr.bottomRightCorner(rows-k, cols-k-1)
+        .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), m_hCoeffs.coeffRef(k), &m_temp.coeffRef(k+1));
+
+    // update our table of norms of the columns
+    for (Index j = k + 1; j < cols; ++j) {
+      // The following implements the stable norm downgrade step discussed in
+      // http://www.netlib.org/lapack/lawnspdf/lawn176.pdf
+      // and used in LAPACK routines xGEQPF and xGEQP3.
+      // See lines 278-297 in http://www.netlib.org/lapack/explore-html/dc/df4/sgeqpf_8f_source.html
+      if (m_colNormsUpdated.coeffRef(j) != RealScalar(0)) {
+        RealScalar temp = abs(m_qr.coeffRef(k, j)) / m_colNormsUpdated.coeffRef(j);
+        temp = (RealScalar(1) + temp) * (RealScalar(1) - temp);
+        temp = temp <  RealScalar(0) ? RealScalar(0) : temp;
+        RealScalar temp2 = temp * numext::abs2<RealScalar>(m_colNormsUpdated.coeffRef(j) /
+                                                           m_colNormsDirect.coeffRef(j));
+        if (temp2 <= norm_downdate_threshold) {
+          // The updated norm has become too inaccurate so re-compute the column
+          // norm directly.
+          m_colNormsDirect.coeffRef(j) = m_qr.col(j).tail(rows - k - 1).norm();
+          m_colNormsUpdated.coeffRef(j) = m_colNormsDirect.coeffRef(j);
+        } else {
+          m_colNormsUpdated.coeffRef(j) *= numext::sqrt(temp);
+        }
+      }
+    }
+  }
+
+  m_colsPermutation.setIdentity(PermIndexType(cols));
+  for(PermIndexType k = 0; k < size/*m_nonzero_pivots*/; ++k)
+    m_colsPermutation.applyTranspositionOnTheRight(k, PermIndexType(m_colsTranspositions.coeff(k)));
+
+  m_det_pq = (number_of_transpositions%2) ? -1 : 1;
+  m_isInitialized = true;
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType>
+template<typename RhsType, typename DstType>
+void ColPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  eigen_assert(rhs.rows() == rows());
+
+  const Index nonzero_pivots = nonzeroPivots();
+
+  if(nonzero_pivots == 0)
+  {
+    dst.setZero();
+    return;
+  }
+
+  typename RhsType::PlainObject c(rhs);
+
+  // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T
+  c.applyOnTheLeft(householderSequence(m_qr, m_hCoeffs)
+                    .setLength(nonzero_pivots)
+                    .transpose()
+    );
+
+  m_qr.topLeftCorner(nonzero_pivots, nonzero_pivots)
+      .template triangularView<Upper>()
+      .solveInPlace(c.topRows(nonzero_pivots));
+
+  for(Index i = 0; i < nonzero_pivots; ++i) dst.row(m_colsPermutation.indices().coeff(i)) = c.row(i);
+  for(Index i = nonzero_pivots; i < cols(); ++i) dst.row(m_colsPermutation.indices().coeff(i)).setZero();
+}
+#endif
+
+namespace internal {
+
+template<typename DstXprType, typename MatrixType>
+struct Assignment<DstXprType, Inverse<ColPivHouseholderQR<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename ColPivHouseholderQR<MatrixType>::Scalar>, Dense2Dense>
+{
+  typedef ColPivHouseholderQR<MatrixType> QrType;
+  typedef Inverse<QrType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename QrType::Scalar> &)
+  {
+    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols()));
+  }
+};
+
+} // end namespace internal
+
+/** \returns the matrix Q as a sequence of householder transformations.
+  * You can extract the meaningful part only by using:
+  * \code qr.householderQ().setLength(qr.nonzeroPivots()) \endcode*/
+template<typename MatrixType>
+typename ColPivHouseholderQR<MatrixType>::HouseholderSequenceType ColPivHouseholderQR<MatrixType>
+  ::householderQ() const
+{
+  eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
+  return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate());
+}
+
+/** \return the column-pivoting Householder QR decomposition of \c *this.
+  *
+  * \sa class ColPivHouseholderQR
+  */
+template<typename Derived>
+const ColPivHouseholderQR<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::colPivHouseholderQr() const
+{
+  return ColPivHouseholderQR<PlainObject>(eval());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_COLPIVOTINGHOUSEHOLDERQR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR_LAPACKE.h
new file mode 100644
index 0000000..4e9651f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/ColPivHouseholderQR_LAPACKE.h
@@ -0,0 +1,97 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Householder QR decomposition of a matrix with column pivoting based on
+ *    LAPACKE_?geqp3 function.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_COLPIVOTINGHOUSEHOLDERQR_LAPACKE_H
+#define EIGEN_COLPIVOTINGHOUSEHOLDERQR_LAPACKE_H
+
+namespace Eigen { 
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_QR_COLPIV(EIGTYPE, LAPACKE_TYPE, LAPACKE_PREFIX, EIGCOLROW, LAPACKE_COLROW) \
+template<> template<typename InputType> inline \
+ColPivHouseholderQR<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> >& \
+ColPivHouseholderQR<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> >::compute( \
+              const EigenBase<InputType>& matrix) \
+\
+{ \
+  using std::abs; \
+  typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> MatrixType; \
+  typedef MatrixType::RealScalar RealScalar; \
+  Index rows = matrix.rows();\
+  Index cols = matrix.cols();\
+\
+  m_qr = matrix;\
+  Index size = m_qr.diagonalSize();\
+  m_hCoeffs.resize(size);\
+\
+  m_colsTranspositions.resize(cols);\
+  /*Index number_of_transpositions = 0;*/ \
+\
+  m_nonzero_pivots = 0; \
+  m_maxpivot = RealScalar(0);\
+  m_colsPermutation.resize(cols); \
+  m_colsPermutation.indices().setZero(); \
+\
+  lapack_int lda = internal::convert_index<lapack_int,Index>(m_qr.outerStride()); \
+  lapack_int matrix_order = LAPACKE_COLROW; \
+  LAPACKE_##LAPACKE_PREFIX##geqp3( matrix_order, internal::convert_index<lapack_int,Index>(rows), internal::convert_index<lapack_int,Index>(cols), \
+                              (LAPACKE_TYPE*)m_qr.data(), lda, (lapack_int*)m_colsPermutation.indices().data(), (LAPACKE_TYPE*)m_hCoeffs.data()); \
+  m_isInitialized = true; \
+  m_maxpivot=m_qr.diagonal().cwiseAbs().maxCoeff(); \
+  m_hCoeffs.adjointInPlace(); \
+  RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold(); \
+  lapack_int *perm = m_colsPermutation.indices().data(); \
+  for(Index i=0;i<size;i++) { \
+    m_nonzero_pivots += (abs(m_qr.coeff(i,i)) > premultiplied_threshold);\
+  } \
+  for(Index i=0;i<cols;i++) perm[i]--;\
+\
+  /*m_det_pq = (number_of_transpositions%2) ? -1 : 1;  // TODO: It's not needed now; fix upon availability in Eigen */ \
+\
+  return *this; \
+}
+
+EIGEN_LAPACKE_QR_COLPIV(double,   double,        d, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(float,    float,         s, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(dcomplex, lapack_complex_double, z, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(scomplex, lapack_complex_float,  c, ColMajor, LAPACK_COL_MAJOR)
+
+EIGEN_LAPACKE_QR_COLPIV(double,   double,        d, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(float,    float,         s, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(dcomplex, lapack_complex_double, z, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_QR_COLPIV(scomplex, lapack_complex_float,  c, RowMajor, LAPACK_ROW_MAJOR)
+
+} // end namespace Eigen
+
+#endif // EIGEN_COLPIVOTINGHOUSEHOLDERQR_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/CompleteOrthogonalDecomposition.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/CompleteOrthogonalDecomposition.h
new file mode 100644
index 0000000..34c637b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/CompleteOrthogonalDecomposition.h
@@ -0,0 +1,562 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPLETEORTHOGONALDECOMPOSITION_H
+#define EIGEN_COMPLETEORTHOGONALDECOMPOSITION_H
+
+namespace Eigen {
+
+namespace internal {
+template <typename _MatrixType>
+struct traits<CompleteOrthogonalDecomposition<_MatrixType> >
+    : traits<_MatrixType> {
+  enum { Flags = 0 };
+};
+
+}  // end namespace internal
+
+/** \ingroup QR_Module
+  *
+  * \class CompleteOrthogonalDecomposition
+  *
+  * \brief Complete orthogonal decomposition (COD) of a matrix.
+  *
+  * \param MatrixType the type of the matrix of which we are computing the COD.
+  *
+  * This class performs a rank-revealing complete orthogonal decomposition of a
+  * matrix  \b A into matrices \b P, \b Q, \b T, and \b Z such that
+  * \f[
+  *  \mathbf{A} \, \mathbf{P} = \mathbf{Q} \,
+  *                     \begin{bmatrix} \mathbf{T} &  \mathbf{0} \\
+  *                                     \mathbf{0} & \mathbf{0} \end{bmatrix} \, \mathbf{Z}
+  * \f]
+  * by using Householder transformations. Here, \b P is a permutation matrix,
+  * \b Q and \b Z are unitary matrices and \b T an upper triangular matrix of
+  * size rank-by-rank. \b A may be rank deficient.
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::completeOrthogonalDecomposition()
+  */
+template <typename _MatrixType>
+class CompleteOrthogonalDecomposition {
+ public:
+  typedef _MatrixType MatrixType;
+  enum {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+  };
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
+  typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime>
+      PermutationType;
+  typedef typename internal::plain_row_type<MatrixType, Index>::type
+      IntRowVectorType;
+  typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
+  typedef typename internal::plain_row_type<MatrixType, RealScalar>::type
+      RealRowVectorType;
+  typedef HouseholderSequence<
+      MatrixType, typename internal::remove_all<
+                      typename HCoeffsType::ConjugateReturnType>::type>
+      HouseholderSequenceType;
+  typedef typename MatrixType::PlainObject PlainObject;
+
+ private:
+  typedef typename PermutationType::Index PermIndexType;
+
+ public:
+  /**
+   * \brief Default Constructor.
+   *
+   * The default constructor is useful in cases in which the user intends to
+   * perform decompositions via
+   * \c CompleteOrthogonalDecomposition::compute(const* MatrixType&).
+   */
+  CompleteOrthogonalDecomposition() : m_cpqr(), m_zCoeffs(), m_temp() {}
+
+  /** \brief Default Constructor with memory preallocation
+   *
+   * Like the default constructor but with preallocation of the internal data
+   * according to the specified problem \a size.
+   * \sa CompleteOrthogonalDecomposition()
+   */
+  CompleteOrthogonalDecomposition(Index rows, Index cols)
+      : m_cpqr(rows, cols), m_zCoeffs((std::min)(rows, cols)), m_temp(cols) {}
+
+  /** \brief Constructs a complete orthogonal decomposition from a given
+   * matrix.
+   *
+   * This constructor computes the complete orthogonal decomposition of the
+   * matrix \a matrix by calling the method compute(). The default
+   * threshold for rank determination will be used. It is a short cut for:
+   *
+   * \code
+   * CompleteOrthogonalDecomposition<MatrixType> cod(matrix.rows(),
+   *                                                 matrix.cols());
+   * cod.setThreshold(Default);
+   * cod.compute(matrix);
+   * \endcode
+   *
+   * \sa compute()
+   */
+  template <typename InputType>
+  explicit CompleteOrthogonalDecomposition(const EigenBase<InputType>& matrix)
+      : m_cpqr(matrix.rows(), matrix.cols()),
+        m_zCoeffs((std::min)(matrix.rows(), matrix.cols())),
+        m_temp(matrix.cols())
+  {
+    compute(matrix.derived());
+  }
+
+  /** \brief Constructs a complete orthogonal decomposition from a given matrix
+    *
+    * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
+    *
+    * \sa CompleteOrthogonalDecomposition(const EigenBase&)
+    */
+  template<typename InputType>
+  explicit CompleteOrthogonalDecomposition(EigenBase<InputType>& matrix)
+    : m_cpqr(matrix.derived()),
+      m_zCoeffs((std::min)(matrix.rows(), matrix.cols())),
+      m_temp(matrix.cols())
+  {
+    computeInPlace();
+  }
+
+
+  /** This method computes the minimum-norm solution X to a least squares
+   * problem \f[\mathrm{minimize} \|A X - B\|, \f] where \b A is the matrix of
+   * which \c *this is the complete orthogonal decomposition.
+   *
+   * \param b the right-hand sides of the problem to solve.
+   *
+   * \returns a solution.
+   *
+   */
+  template <typename Rhs>
+  inline const Solve<CompleteOrthogonalDecomposition, Rhs> solve(
+      const MatrixBase<Rhs>& b) const {
+    eigen_assert(m_cpqr.m_isInitialized &&
+                 "CompleteOrthogonalDecomposition is not initialized.");
+    return Solve<CompleteOrthogonalDecomposition, Rhs>(*this, b.derived());
+  }
+
+  HouseholderSequenceType householderQ(void) const;
+  HouseholderSequenceType matrixQ(void) const { return m_cpqr.householderQ(); }
+
+  /** \returns the matrix \b Z.
+   */
+  MatrixType matrixZ() const {
+    MatrixType Z = MatrixType::Identity(m_cpqr.cols(), m_cpqr.cols());
+    applyZAdjointOnTheLeftInPlace(Z);
+    return Z.adjoint();
+  }
+
+  /** \returns a reference to the matrix where the complete orthogonal
+   * decomposition is stored
+   */
+  const MatrixType& matrixQTZ() const { return m_cpqr.matrixQR(); }
+
+  /** \returns a reference to the matrix where the complete orthogonal
+   * decomposition is stored.
+   * \warning The strict lower part and \code cols() - rank() \endcode right
+   * columns of this matrix contains internal values.
+   * Only the upper triangular part should be referenced. To get it, use
+   * \code matrixT().template triangularView<Upper>() \endcode
+   * For rank-deficient matrices, use
+   * \code
+   * matrixR().topLeftCorner(rank(), rank()).template triangularView<Upper>()
+   * \endcode
+   */
+  const MatrixType& matrixT() const { return m_cpqr.matrixQR(); }
+
+  template <typename InputType>
+  CompleteOrthogonalDecomposition& compute(const EigenBase<InputType>& matrix) {
+    // Compute the column pivoted QR factorization A P = Q R.
+    m_cpqr.compute(matrix);
+    computeInPlace();
+    return *this;
+  }
+
+  /** \returns a const reference to the column permutation matrix */
+  const PermutationType& colsPermutation() const {
+    return m_cpqr.colsPermutation();
+  }
+
+  /** \returns the absolute value of the determinant of the matrix of which
+   * *this is the complete orthogonal decomposition. It has only linear
+   * complexity (that is, O(n) where n is the dimension of the square matrix)
+   * as the complete orthogonal decomposition has already been computed.
+   *
+   * \note This is only for square matrices.
+   *
+   * \warning a determinant can be very big or small, so for matrices
+   * of large enough dimension, there is a risk of overflow/underflow.
+   * One way to work around that is to use logAbsDeterminant() instead.
+   *
+   * \sa logAbsDeterminant(), MatrixBase::determinant()
+   */
+  typename MatrixType::RealScalar absDeterminant() const;
+
+  /** \returns the natural log of the absolute value of the determinant of the
+   * matrix of which *this is the complete orthogonal decomposition. It has
+   * only linear complexity (that is, O(n) where n is the dimension of the
+   * square matrix) as the complete orthogonal decomposition has already been
+   * computed.
+   *
+   * \note This is only for square matrices.
+   *
+   * \note This method is useful to work around the risk of overflow/underflow
+   * that's inherent to determinant computation.
+   *
+   * \sa absDeterminant(), MatrixBase::determinant()
+   */
+  typename MatrixType::RealScalar logAbsDeterminant() const;
+
+  /** \returns the rank of the matrix of which *this is the complete orthogonal
+   * decomposition.
+   *
+   * \note This method has to determine which pivots should be considered
+   * nonzero. For that, it uses the threshold value that you can control by
+   * calling setThreshold(const RealScalar&).
+   */
+  inline Index rank() const { return m_cpqr.rank(); }
+
+  /** \returns the dimension of the kernel of the matrix of which *this is the
+   * complete orthogonal decomposition.
+   *
+   * \note This method has to determine which pivots should be considered
+   * nonzero. For that, it uses the threshold value that you can control by
+   * calling setThreshold(const RealScalar&).
+   */
+  inline Index dimensionOfKernel() const { return m_cpqr.dimensionOfKernel(); }
+
+  /** \returns true if the matrix of which *this is the decomposition represents
+   * an injective linear map, i.e. has trivial kernel; false otherwise.
+   *
+   * \note This method has to determine which pivots should be considered
+   * nonzero. For that, it uses the threshold value that you can control by
+   * calling setThreshold(const RealScalar&).
+   */
+  inline bool isInjective() const { return m_cpqr.isInjective(); }
+
+  /** \returns true if the matrix of which *this is the decomposition represents
+   * a surjective linear map; false otherwise.
+   *
+   * \note This method has to determine which pivots should be considered
+   * nonzero. For that, it uses the threshold value that you can control by
+   * calling setThreshold(const RealScalar&).
+   */
+  inline bool isSurjective() const { return m_cpqr.isSurjective(); }
+
+  /** \returns true if the matrix of which *this is the complete orthogonal
+   * decomposition is invertible.
+   *
+   * \note This method has to determine which pivots should be considered
+   * nonzero. For that, it uses the threshold value that you can control by
+   * calling setThreshold(const RealScalar&).
+   */
+  inline bool isInvertible() const { return m_cpqr.isInvertible(); }
+
+  /** \returns the pseudo-inverse of the matrix of which *this is the complete
+   * orthogonal decomposition.
+   * \warning: Do not compute \c this->pseudoInverse()*rhs to solve a linear systems.
+   * It is more efficient and numerically stable to call \c this->solve(rhs).
+   */
+  inline const Inverse<CompleteOrthogonalDecomposition> pseudoInverse() const
+  {
+    return Inverse<CompleteOrthogonalDecomposition>(*this);
+  }
+
+  inline Index rows() const { return m_cpqr.rows(); }
+  inline Index cols() const { return m_cpqr.cols(); }
+
+  /** \returns a const reference to the vector of Householder coefficients used
+   * to represent the factor \c Q.
+   *
+   * For advanced uses only.
+   */
+  inline const HCoeffsType& hCoeffs() const { return m_cpqr.hCoeffs(); }
+
+  /** \returns a const reference to the vector of Householder coefficients
+   * used to represent the factor \c Z.
+   *
+   * For advanced uses only.
+   */
+  const HCoeffsType& zCoeffs() const { return m_zCoeffs; }
+
+  /** Allows to prescribe a threshold to be used by certain methods, such as
+   * rank(), who need to determine when pivots are to be considered nonzero.
+   * Most be called before calling compute().
+   *
+   * When it needs to get the threshold value, Eigen calls threshold(). By
+   * default, this uses a formula to automatically determine a reasonable
+   * threshold. Once you have called the present method
+   * setThreshold(const RealScalar&), your value is used instead.
+   *
+   * \param threshold The new value to use as the threshold.
+   *
+   * A pivot will be considered nonzero if its absolute value is strictly
+   * greater than
+   *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
+   * where maxpivot is the biggest pivot.
+   *
+   * If you want to come back to the default behavior, call
+   * setThreshold(Default_t)
+   */
+  CompleteOrthogonalDecomposition& setThreshold(const RealScalar& threshold) {
+    m_cpqr.setThreshold(threshold);
+    return *this;
+  }
+
+  /** Allows to come back to the default behavior, letting Eigen use its default
+   * formula for determining the threshold.
+   *
+   * You should pass the special object Eigen::Default as parameter here.
+   * \code qr.setThreshold(Eigen::Default); \endcode
+   *
+   * See the documentation of setThreshold(const RealScalar&).
+   */
+  CompleteOrthogonalDecomposition& setThreshold(Default_t) {
+    m_cpqr.setThreshold(Default);
+    return *this;
+  }
+
+  /** Returns the threshold that will be used by certain methods such as rank().
+   *
+   * See the documentation of setThreshold(const RealScalar&).
+   */
+  RealScalar threshold() const { return m_cpqr.threshold(); }
+
+  /** \returns the number of nonzero pivots in the complete orthogonal
+   * decomposition. Here nonzero is meant in the exact sense, not in a
+   * fuzzy sense. So that notion isn't really intrinsically interesting,
+   * but it is still useful when implementing algorithms.
+   *
+   * \sa rank()
+   */
+  inline Index nonzeroPivots() const { return m_cpqr.nonzeroPivots(); }
+
+  /** \returns the absolute value of the biggest pivot, i.e. the biggest
+   *          diagonal coefficient of R.
+   */
+  inline RealScalar maxPivot() const { return m_cpqr.maxPivot(); }
+
+  /** \brief Reports whether the complete orthogonal decomposition was
+   * succesful.
+   *
+   * \note This function always returns \c Success. It is provided for
+   * compatibility
+   * with other factorization routines.
+   * \returns \c Success
+   */
+  ComputationInfo info() const {
+    eigen_assert(m_cpqr.m_isInitialized && "Decomposition is not initialized.");
+    return Success;
+  }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  template <typename RhsType, typename DstType>
+  EIGEN_DEVICE_FUNC void _solve_impl(const RhsType& rhs, DstType& dst) const;
+#endif
+
+ protected:
+  static void check_template_parameters() {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+  }
+
+  void computeInPlace();
+
+  /** Overwrites \b rhs with \f$ \mathbf{Z}^* * \mathbf{rhs} \f$.
+   */
+  template <typename Rhs>
+  void applyZAdjointOnTheLeftInPlace(Rhs& rhs) const;
+
+  ColPivHouseholderQR<MatrixType> m_cpqr;
+  HCoeffsType m_zCoeffs;
+  RowVectorType m_temp;
+};
+
+template <typename MatrixType>
+typename MatrixType::RealScalar
+CompleteOrthogonalDecomposition<MatrixType>::absDeterminant() const {
+  return m_cpqr.absDeterminant();
+}
+
+template <typename MatrixType>
+typename MatrixType::RealScalar
+CompleteOrthogonalDecomposition<MatrixType>::logAbsDeterminant() const {
+  return m_cpqr.logAbsDeterminant();
+}
+
+/** Performs the complete orthogonal decomposition of the given matrix \a
+ * matrix. The result of the factorization is stored into \c *this, and a
+ * reference to \c *this is returned.
+ *
+ * \sa class CompleteOrthogonalDecomposition,
+ * CompleteOrthogonalDecomposition(const MatrixType&)
+ */
+template <typename MatrixType>
+void CompleteOrthogonalDecomposition<MatrixType>::computeInPlace()
+{
+  check_template_parameters();
+
+  // the column permutation is stored as int indices, so just to be sure:
+  eigen_assert(m_cpqr.cols() <= NumTraits<int>::highest());
+
+  const Index rank = m_cpqr.rank();
+  const Index cols = m_cpqr.cols();
+  const Index rows = m_cpqr.rows();
+  m_zCoeffs.resize((std::min)(rows, cols));
+  m_temp.resize(cols);
+
+  if (rank < cols) {
+    // We have reduced the (permuted) matrix to the form
+    //   [R11 R12]
+    //   [ 0  R22]
+    // where R11 is r-by-r (r = rank) upper triangular, R12 is
+    // r-by-(n-r), and R22 is empty or the norm of R22 is negligible.
+    // We now compute the complete orthogonal decomposition by applying
+    // Householder transformations from the right to the upper trapezoidal
+    // matrix X = [R11 R12] to zero out R12 and obtain the factorization
+    // [R11 R12] = [T11 0] * Z, where T11 is r-by-r upper triangular and
+    // Z = Z(0) * Z(1) ... Z(r-1) is an n-by-n orthogonal matrix.
+    // We store the data representing Z in R12 and m_zCoeffs.
+    for (Index k = rank - 1; k >= 0; --k) {
+      if (k != rank - 1) {
+        // Given the API for Householder reflectors, it is more convenient if
+        // we swap the leading parts of columns k and r-1 (zero-based) to form
+        // the matrix X_k = [X(0:k, k), X(0:k, r:n)]
+        m_cpqr.m_qr.col(k).head(k + 1).swap(
+            m_cpqr.m_qr.col(rank - 1).head(k + 1));
+      }
+      // Construct Householder reflector Z(k) to zero out the last row of X_k,
+      // i.e. choose Z(k) such that
+      // [X(k, k), X(k, r:n)] * Z(k) = [beta, 0, .., 0].
+      RealScalar beta;
+      m_cpqr.m_qr.row(k)
+          .tail(cols - rank + 1)
+          .makeHouseholderInPlace(m_zCoeffs(k), beta);
+      m_cpqr.m_qr(k, rank - 1) = beta;
+      if (k > 0) {
+        // Apply Z(k) to the first k rows of X_k
+        m_cpqr.m_qr.topRightCorner(k, cols - rank + 1)
+            .applyHouseholderOnTheRight(
+                m_cpqr.m_qr.row(k).tail(cols - rank).transpose(), m_zCoeffs(k),
+                &m_temp(0));
+      }
+      if (k != rank - 1) {
+        // Swap X(0:k,k) back to its proper location.
+        m_cpqr.m_qr.col(k).head(k + 1).swap(
+            m_cpqr.m_qr.col(rank - 1).head(k + 1));
+      }
+    }
+  }
+}
+
+template <typename MatrixType>
+template <typename Rhs>
+void CompleteOrthogonalDecomposition<MatrixType>::applyZAdjointOnTheLeftInPlace(
+    Rhs& rhs) const {
+  const Index cols = this->cols();
+  const Index nrhs = rhs.cols();
+  const Index rank = this->rank();
+  Matrix<typename MatrixType::Scalar, Dynamic, 1> temp((std::max)(cols, nrhs));
+  for (Index k = 0; k < rank; ++k) {
+    if (k != rank - 1) {
+      rhs.row(k).swap(rhs.row(rank - 1));
+    }
+    rhs.middleRows(rank - 1, cols - rank + 1)
+        .applyHouseholderOnTheLeft(
+            matrixQTZ().row(k).tail(cols - rank).adjoint(), zCoeffs()(k),
+            &temp(0));
+    if (k != rank - 1) {
+      rhs.row(k).swap(rhs.row(rank - 1));
+    }
+  }
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template <typename _MatrixType>
+template <typename RhsType, typename DstType>
+void CompleteOrthogonalDecomposition<_MatrixType>::_solve_impl(
+    const RhsType& rhs, DstType& dst) const {
+  eigen_assert(rhs.rows() == this->rows());
+
+  const Index rank = this->rank();
+  if (rank == 0) {
+    dst.setZero();
+    return;
+  }
+
+  // Compute c = Q^* * rhs
+  // Note that the matrix Q = H_0^* H_1^*... so its inverse is
+  // Q^* = (H_0 H_1 ...)^T
+  typename RhsType::PlainObject c(rhs);
+  c.applyOnTheLeft(
+      householderSequence(matrixQTZ(), hCoeffs()).setLength(rank).transpose());
+
+  // Solve T z = c(1:rank, :)
+  dst.topRows(rank) = matrixT()
+                          .topLeftCorner(rank, rank)
+                          .template triangularView<Upper>()
+                          .solve(c.topRows(rank));
+
+  const Index cols = this->cols();
+  if (rank < cols) {
+    // Compute y = Z^* * [ z ]
+    //                   [ 0 ]
+    dst.bottomRows(cols - rank).setZero();
+    applyZAdjointOnTheLeftInPlace(dst);
+  }
+
+  // Undo permutation to get x = P^{-1} * y.
+  dst = colsPermutation() * dst;
+}
+#endif
+
+namespace internal {
+
+template<typename DstXprType, typename MatrixType>
+struct Assignment<DstXprType, Inverse<CompleteOrthogonalDecomposition<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename CompleteOrthogonalDecomposition<MatrixType>::Scalar>, Dense2Dense>
+{
+  typedef CompleteOrthogonalDecomposition<MatrixType> CodType;
+  typedef Inverse<CodType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename CodType::Scalar> &)
+  {
+    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.rows()));
+  }
+};
+
+} // end namespace internal
+
+/** \returns the matrix Q as a sequence of householder transformations */
+template <typename MatrixType>
+typename CompleteOrthogonalDecomposition<MatrixType>::HouseholderSequenceType
+CompleteOrthogonalDecomposition<MatrixType>::householderQ() const {
+  return m_cpqr.householderQ();
+}
+
+/** \return the complete orthogonal decomposition of \c *this.
+  *
+  * \sa class CompleteOrthogonalDecomposition
+  */
+template <typename Derived>
+const CompleteOrthogonalDecomposition<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::completeOrthogonalDecomposition() const {
+  return CompleteOrthogonalDecomposition<PlainObject>(eval());
+}
+
+}  // end namespace Eigen
+
+#endif  // EIGEN_COMPLETEORTHOGONALDECOMPOSITION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/FullPivHouseholderQR.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/FullPivHouseholderQR.h
new file mode 100644
index 0000000..e489bdd
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/FullPivHouseholderQR.h
@@ -0,0 +1,676 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_FULLPIVOTINGHOUSEHOLDERQR_H
+#define EIGEN_FULLPIVOTINGHOUSEHOLDERQR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename _MatrixType> struct traits<FullPivHouseholderQR<_MatrixType> >
+ : traits<_MatrixType>
+{
+  enum { Flags = 0 };
+};
+
+template<typename MatrixType> struct FullPivHouseholderQRMatrixQReturnType;
+
+template<typename MatrixType>
+struct traits<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
+{
+  typedef typename MatrixType::PlainObject ReturnType;
+};
+
+} // end namespace internal
+
+/** \ingroup QR_Module
+  *
+  * \class FullPivHouseholderQR
+  *
+  * \brief Householder rank-revealing QR decomposition of a matrix with full pivoting
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the QR decomposition
+  *
+  * This class performs a rank-revealing QR decomposition of a matrix \b A into matrices \b P, \b P', \b Q and \b R
+  * such that 
+  * \f[
+  *  \mathbf{P} \, \mathbf{A} \, \mathbf{P}' = \mathbf{Q} \, \mathbf{R}
+  * \f]
+  * by using Householder transformations. Here, \b P and \b P' are permutation matrices, \b Q a unitary matrix 
+  * and \b R an upper triangular matrix.
+  *
+  * This decomposition performs a very prudent full pivoting in order to be rank-revealing and achieve optimal
+  * numerical stability. The trade-off is that it is slower than HouseholderQR and ColPivHouseholderQR.
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  * 
+  * \sa MatrixBase::fullPivHouseholderQr()
+  */
+template<typename _MatrixType> class FullPivHouseholderQR
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    // FIXME should be int
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef internal::FullPivHouseholderQRMatrixQReturnType<MatrixType> MatrixQReturnType;
+    typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
+    typedef Matrix<StorageIndex, 1,
+                   EIGEN_SIZE_MIN_PREFER_DYNAMIC(ColsAtCompileTime,RowsAtCompileTime), RowMajor, 1,
+                   EIGEN_SIZE_MIN_PREFER_FIXED(MaxColsAtCompileTime,MaxRowsAtCompileTime)> IntDiagSizeVectorType;
+    typedef PermutationMatrix<ColsAtCompileTime, MaxColsAtCompileTime> PermutationType;
+    typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
+    typedef typename internal::plain_col_type<MatrixType>::type ColVectorType;
+    typedef typename MatrixType::PlainObject PlainObject;
+
+    /** \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via FullPivHouseholderQR::compute(const MatrixType&).
+      */
+    FullPivHouseholderQR()
+      : m_qr(),
+        m_hCoeffs(),
+        m_rows_transpositions(),
+        m_cols_transpositions(),
+        m_cols_permutation(),
+        m_temp(),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false) {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa FullPivHouseholderQR()
+      */
+    FullPivHouseholderQR(Index rows, Index cols)
+      : m_qr(rows, cols),
+        m_hCoeffs((std::min)(rows,cols)),
+        m_rows_transpositions((std::min)(rows,cols)),
+        m_cols_transpositions((std::min)(rows,cols)),
+        m_cols_permutation(cols),
+        m_temp(cols),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false) {}
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This constructor computes the QR factorization of the matrix \a matrix by calling
+      * the method compute(). It is a short cut for:
+      * 
+      * \code
+      * FullPivHouseholderQR<MatrixType> qr(matrix.rows(), matrix.cols());
+      * qr.compute(matrix);
+      * \endcode
+      * 
+      * \sa compute()
+      */
+    template<typename InputType>
+    explicit FullPivHouseholderQR(const EigenBase<InputType>& matrix)
+      : m_qr(matrix.rows(), matrix.cols()),
+        m_hCoeffs((std::min)(matrix.rows(), matrix.cols())),
+        m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())),
+        m_cols_transpositions((std::min)(matrix.rows(), matrix.cols())),
+        m_cols_permutation(matrix.cols()),
+        m_temp(matrix.cols()),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false)
+    {
+      compute(matrix.derived());
+    }
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when \c MatrixType is a Eigen::Ref.
+      *
+      * \sa FullPivHouseholderQR(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit FullPivHouseholderQR(EigenBase<InputType>& matrix)
+      : m_qr(matrix.derived()),
+        m_hCoeffs((std::min)(matrix.rows(), matrix.cols())),
+        m_rows_transpositions((std::min)(matrix.rows(), matrix.cols())),
+        m_cols_transpositions((std::min)(matrix.rows(), matrix.cols())),
+        m_cols_permutation(matrix.cols()),
+        m_temp(matrix.cols()),
+        m_isInitialized(false),
+        m_usePrescribedThreshold(false)
+    {
+      computeInPlace();
+    }
+
+    /** This method finds a solution x to the equation Ax=b, where A is the matrix of which
+      * \c *this is the QR decomposition.
+      *
+      * \param b the right-hand-side of the equation to solve.
+      *
+      * \returns the exact or least-square solution if the rank is greater or equal to the number of columns of A,
+      * and an arbitrary solution otherwise.
+      *
+      * \note_about_checking_solutions
+      *
+      * \note_about_arbitrary_choice_of_solution
+      *
+      * Example: \include FullPivHouseholderQR_solve.cpp
+      * Output: \verbinclude FullPivHouseholderQR_solve.out
+      */
+    template<typename Rhs>
+    inline const Solve<FullPivHouseholderQR, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return Solve<FullPivHouseholderQR, Rhs>(*this, b.derived());
+    }
+
+    /** \returns Expression object representing the matrix Q
+      */
+    MatrixQReturnType matrixQ(void) const;
+
+    /** \returns a reference to the matrix where the Householder QR decomposition is stored
+      */
+    const MatrixType& matrixQR() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return m_qr;
+    }
+
+    template<typename InputType>
+    FullPivHouseholderQR& compute(const EigenBase<InputType>& matrix);
+
+    /** \returns a const reference to the column permutation matrix */
+    const PermutationType& colsPermutation() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return m_cols_permutation;
+    }
+
+    /** \returns a const reference to the vector of indices representing the rows transpositions */
+    const IntDiagSizeVectorType& rowsTranspositions() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return m_rows_transpositions;
+    }
+
+    /** \returns the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      * One way to work around that is to use logAbsDeterminant() instead.
+      *
+      * \sa logAbsDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar absDeterminant() const;
+
+    /** \returns the natural log of the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \note This method is useful to work around the risk of overflow/underflow that's inherent
+      * to determinant computation.
+      *
+      * \sa absDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar logAbsDeterminant() const;
+
+    /** \returns the rank of the matrix of which *this is the QR decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index rank() const
+    {
+      using std::abs;
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      RealScalar premultiplied_threshold = abs(m_maxpivot) * threshold();
+      Index result = 0;
+      for(Index i = 0; i < m_nonzero_pivots; ++i)
+        result += (abs(m_qr.coeff(i,i)) > premultiplied_threshold);
+      return result;
+    }
+
+    /** \returns the dimension of the kernel of the matrix of which *this is the QR decomposition.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline Index dimensionOfKernel() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return cols() - rank();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition represents an injective
+      *          linear map, i.e. has trivial kernel; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInjective() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return rank() == cols();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition represents a surjective
+      *          linear map; false otherwise.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isSurjective() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return rank() == rows();
+    }
+
+    /** \returns true if the matrix of which *this is the QR decomposition is invertible.
+      *
+      * \note This method has to determine which pivots should be considered nonzero.
+      *       For that, it uses the threshold value that you can control by calling
+      *       setThreshold(const RealScalar&).
+      */
+    inline bool isInvertible() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return isInjective() && isSurjective();
+    }
+
+    /** \returns the inverse of the matrix of which *this is the QR decomposition.
+      *
+      * \note If this matrix is not invertible, the returned matrix has undefined coefficients.
+      *       Use isInvertible() to first determine whether this matrix is invertible.
+      */
+    inline const Inverse<FullPivHouseholderQR> inverse() const
+    {
+      eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+      return Inverse<FullPivHouseholderQR>(*this);
+    }
+
+    inline Index rows() const { return m_qr.rows(); }
+    inline Index cols() const { return m_qr.cols(); }
+    
+    /** \returns a const reference to the vector of Householder coefficients used to represent the factor \c Q.
+      * 
+      * For advanced uses only.
+      */
+    const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
+
+    /** Allows to prescribe a threshold to be used by certain methods, such as rank(),
+      * who need to determine when pivots are to be considered nonzero. This is not used for the
+      * QR decomposition itself.
+      *
+      * When it needs to get the threshold value, Eigen calls threshold(). By default, this
+      * uses a formula to automatically determine a reasonable threshold.
+      * Once you have called the present method setThreshold(const RealScalar&),
+      * your value is used instead.
+      *
+      * \param threshold The new value to use as the threshold.
+      *
+      * A pivot will be considered nonzero if its absolute value is strictly greater than
+      *  \f$ \vert pivot \vert \leqslant threshold \times \vert maxpivot \vert \f$
+      * where maxpivot is the biggest pivot.
+      *
+      * If you want to come back to the default behavior, call setThreshold(Default_t)
+      */
+    FullPivHouseholderQR& setThreshold(const RealScalar& threshold)
+    {
+      m_usePrescribedThreshold = true;
+      m_prescribedThreshold = threshold;
+      return *this;
+    }
+
+    /** Allows to come back to the default behavior, letting Eigen use its default formula for
+      * determining the threshold.
+      *
+      * You should pass the special object Eigen::Default as parameter here.
+      * \code qr.setThreshold(Eigen::Default); \endcode
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    FullPivHouseholderQR& setThreshold(Default_t)
+    {
+      m_usePrescribedThreshold = false;
+      return *this;
+    }
+
+    /** Returns the threshold that will be used by certain methods such as rank().
+      *
+      * See the documentation of setThreshold(const RealScalar&).
+      */
+    RealScalar threshold() const
+    {
+      eigen_assert(m_isInitialized || m_usePrescribedThreshold);
+      return m_usePrescribedThreshold ? m_prescribedThreshold
+      // this formula comes from experimenting (see "LU precision tuning" thread on the list)
+      // and turns out to be identical to Higham's formula used already in LDLt.
+                                      : NumTraits<Scalar>::epsilon() * RealScalar(m_qr.diagonalSize());
+    }
+
+    /** \returns the number of nonzero pivots in the QR decomposition.
+      * Here nonzero is meant in the exact sense, not in a fuzzy sense.
+      * So that notion isn't really intrinsically interesting, but it is
+      * still useful when implementing algorithms.
+      *
+      * \sa rank()
+      */
+    inline Index nonzeroPivots() const
+    {
+      eigen_assert(m_isInitialized && "LU is not initialized.");
+      return m_nonzero_pivots;
+    }
+
+    /** \returns the absolute value of the biggest pivot, i.e. the biggest
+      *          diagonal coefficient of U.
+      */
+    RealScalar maxPivot() const { return m_maxpivot; }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
+    void computeInPlace();
+    
+    MatrixType m_qr;
+    HCoeffsType m_hCoeffs;
+    IntDiagSizeVectorType m_rows_transpositions;
+    IntDiagSizeVectorType m_cols_transpositions;
+    PermutationType m_cols_permutation;
+    RowVectorType m_temp;
+    bool m_isInitialized, m_usePrescribedThreshold;
+    RealScalar m_prescribedThreshold, m_maxpivot;
+    Index m_nonzero_pivots;
+    RealScalar m_precision;
+    Index m_det_pq;
+};
+
+template<typename MatrixType>
+typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::absDeterminant() const
+{
+  using std::abs;
+  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return abs(m_qr.diagonal().prod());
+}
+
+template<typename MatrixType>
+typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::logAbsDeterminant() const
+{
+  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return m_qr.diagonal().cwiseAbs().array().log().sum();
+}
+
+/** Performs the QR factorization of the given matrix \a matrix. The result of
+  * the factorization is stored into \c *this, and a reference to \c *this
+  * is returned.
+  *
+  * \sa class FullPivHouseholderQR, FullPivHouseholderQR(const MatrixType&)
+  */
+template<typename MatrixType>
+template<typename InputType>
+FullPivHouseholderQR<MatrixType>& FullPivHouseholderQR<MatrixType>::compute(const EigenBase<InputType>& matrix)
+{
+  m_qr = matrix.derived();
+  computeInPlace();
+  return *this;
+}
+
+template<typename MatrixType>
+void FullPivHouseholderQR<MatrixType>::computeInPlace()
+{
+  check_template_parameters();
+
+  using std::abs;
+  Index rows = m_qr.rows();
+  Index cols = m_qr.cols();
+  Index size = (std::min)(rows,cols);
+
+  
+  m_hCoeffs.resize(size);
+
+  m_temp.resize(cols);
+
+  m_precision = NumTraits<Scalar>::epsilon() * RealScalar(size);
+
+  m_rows_transpositions.resize(size);
+  m_cols_transpositions.resize(size);
+  Index number_of_transpositions = 0;
+
+  RealScalar biggest(0);
+
+  m_nonzero_pivots = size; // the generic case is that in which all pivots are nonzero (invertible case)
+  m_maxpivot = RealScalar(0);
+
+  for (Index k = 0; k < size; ++k)
+  {
+    Index row_of_biggest_in_corner, col_of_biggest_in_corner;
+    typedef internal::scalar_score_coeff_op<Scalar> Scoring;
+    typedef typename Scoring::result_type Score;
+
+    Score score = m_qr.bottomRightCorner(rows-k, cols-k)
+                      .unaryExpr(Scoring())
+                      .maxCoeff(&row_of_biggest_in_corner, &col_of_biggest_in_corner);
+    row_of_biggest_in_corner += k;
+    col_of_biggest_in_corner += k;
+    RealScalar biggest_in_corner = internal::abs_knowing_score<Scalar>()(m_qr(row_of_biggest_in_corner, col_of_biggest_in_corner), score);
+    if(k==0) biggest = biggest_in_corner;
+
+    // if the corner is negligible, then we have less than full rank, and we can finish early
+    if(internal::isMuchSmallerThan(biggest_in_corner, biggest, m_precision))
+    {
+      m_nonzero_pivots = k;
+      for(Index i = k; i < size; i++)
+      {
+        m_rows_transpositions.coeffRef(i) = i;
+        m_cols_transpositions.coeffRef(i) = i;
+        m_hCoeffs.coeffRef(i) = Scalar(0);
+      }
+      break;
+    }
+
+    m_rows_transpositions.coeffRef(k) = row_of_biggest_in_corner;
+    m_cols_transpositions.coeffRef(k) = col_of_biggest_in_corner;
+    if(k != row_of_biggest_in_corner) {
+      m_qr.row(k).tail(cols-k).swap(m_qr.row(row_of_biggest_in_corner).tail(cols-k));
+      ++number_of_transpositions;
+    }
+    if(k != col_of_biggest_in_corner) {
+      m_qr.col(k).swap(m_qr.col(col_of_biggest_in_corner));
+      ++number_of_transpositions;
+    }
+
+    RealScalar beta;
+    m_qr.col(k).tail(rows-k).makeHouseholderInPlace(m_hCoeffs.coeffRef(k), beta);
+    m_qr.coeffRef(k,k) = beta;
+
+    // remember the maximum absolute value of diagonal coefficients
+    if(abs(beta) > m_maxpivot) m_maxpivot = abs(beta);
+
+    m_qr.bottomRightCorner(rows-k, cols-k-1)
+        .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), m_hCoeffs.coeffRef(k), &m_temp.coeffRef(k+1));
+  }
+
+  m_cols_permutation.setIdentity(cols);
+  for(Index k = 0; k < size; ++k)
+    m_cols_permutation.applyTranspositionOnTheRight(k, m_cols_transpositions.coeff(k));
+
+  m_det_pq = (number_of_transpositions%2) ? -1 : 1;
+  m_isInitialized = true;
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType>
+template<typename RhsType, typename DstType>
+void FullPivHouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  eigen_assert(rhs.rows() == rows());
+  const Index l_rank = rank();
+
+  // FIXME introduce nonzeroPivots() and use it here. and more generally,
+  // make the same improvements in this dec as in FullPivLU.
+  if(l_rank==0)
+  {
+    dst.setZero();
+    return;
+  }
+
+  typename RhsType::PlainObject c(rhs);
+
+  Matrix<Scalar,1,RhsType::ColsAtCompileTime> temp(rhs.cols());
+  for (Index k = 0; k < l_rank; ++k)
+  {
+    Index remainingSize = rows()-k;
+    c.row(k).swap(c.row(m_rows_transpositions.coeff(k)));
+    c.bottomRightCorner(remainingSize, rhs.cols())
+      .applyHouseholderOnTheLeft(m_qr.col(k).tail(remainingSize-1),
+                               m_hCoeffs.coeff(k), &temp.coeffRef(0));
+  }
+
+  m_qr.topLeftCorner(l_rank, l_rank)
+      .template triangularView<Upper>()
+      .solveInPlace(c.topRows(l_rank));
+
+  for(Index i = 0; i < l_rank; ++i) dst.row(m_cols_permutation.indices().coeff(i)) = c.row(i);
+  for(Index i = l_rank; i < cols(); ++i) dst.row(m_cols_permutation.indices().coeff(i)).setZero();
+}
+#endif
+
+namespace internal {
+  
+template<typename DstXprType, typename MatrixType>
+struct Assignment<DstXprType, Inverse<FullPivHouseholderQR<MatrixType> >, internal::assign_op<typename DstXprType::Scalar,typename FullPivHouseholderQR<MatrixType>::Scalar>, Dense2Dense>
+{
+  typedef FullPivHouseholderQR<MatrixType> QrType;
+  typedef Inverse<QrType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename QrType::Scalar> &)
+  {    
+    dst = src.nestedExpression().solve(MatrixType::Identity(src.rows(), src.cols()));
+  }
+};
+
+/** \ingroup QR_Module
+  *
+  * \brief Expression type for return value of FullPivHouseholderQR::matrixQ()
+  *
+  * \tparam MatrixType type of underlying dense matrix
+  */
+template<typename MatrixType> struct FullPivHouseholderQRMatrixQReturnType
+  : public ReturnByValue<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
+{
+public:
+  typedef typename FullPivHouseholderQR<MatrixType>::IntDiagSizeVectorType IntDiagSizeVectorType;
+  typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
+  typedef Matrix<typename MatrixType::Scalar, 1, MatrixType::RowsAtCompileTime, RowMajor, 1,
+                 MatrixType::MaxRowsAtCompileTime> WorkVectorType;
+
+  FullPivHouseholderQRMatrixQReturnType(const MatrixType&       qr,
+                                        const HCoeffsType&      hCoeffs,
+                                        const IntDiagSizeVectorType& rowsTranspositions)
+    : m_qr(qr),
+      m_hCoeffs(hCoeffs),
+      m_rowsTranspositions(rowsTranspositions)
+  {}
+
+  template <typename ResultType>
+  void evalTo(ResultType& result) const
+  {
+    const Index rows = m_qr.rows();
+    WorkVectorType workspace(rows);
+    evalTo(result, workspace);
+  }
+
+  template <typename ResultType>
+  void evalTo(ResultType& result, WorkVectorType& workspace) const
+  {
+    using numext::conj;
+    // compute the product H'_0 H'_1 ... H'_n-1,
+    // where H_k is the k-th Householder transformation I - h_k v_k v_k'
+    // and v_k is the k-th Householder vector [1,m_qr(k+1,k), m_qr(k+2,k), ...]
+    const Index rows = m_qr.rows();
+    const Index cols = m_qr.cols();
+    const Index size = (std::min)(rows, cols);
+    workspace.resize(rows);
+    result.setIdentity(rows, rows);
+    for (Index k = size-1; k >= 0; k--)
+    {
+      result.block(k, k, rows-k, rows-k)
+            .applyHouseholderOnTheLeft(m_qr.col(k).tail(rows-k-1), conj(m_hCoeffs.coeff(k)), &workspace.coeffRef(k));
+      result.row(k).swap(result.row(m_rowsTranspositions.coeff(k)));
+    }
+  }
+
+  Index rows() const { return m_qr.rows(); }
+  Index cols() const { return m_qr.rows(); }
+
+protected:
+  typename MatrixType::Nested m_qr;
+  typename HCoeffsType::Nested m_hCoeffs;
+  typename IntDiagSizeVectorType::Nested m_rowsTranspositions;
+};
+
+// template<typename MatrixType>
+// struct evaluator<FullPivHouseholderQRMatrixQReturnType<MatrixType> >
+//  : public evaluator<ReturnByValue<FullPivHouseholderQRMatrixQReturnType<MatrixType> > >
+// {};
+
+} // end namespace internal
+
+template<typename MatrixType>
+inline typename FullPivHouseholderQR<MatrixType>::MatrixQReturnType FullPivHouseholderQR<MatrixType>::matrixQ() const
+{
+  eigen_assert(m_isInitialized && "FullPivHouseholderQR is not initialized.");
+  return MatrixQReturnType(m_qr, m_hCoeffs, m_rows_transpositions);
+}
+
+/** \return the full-pivoting Householder QR decomposition of \c *this.
+  *
+  * \sa class FullPivHouseholderQR
+  */
+template<typename Derived>
+const FullPivHouseholderQR<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::fullPivHouseholderQr() const
+{
+  return FullPivHouseholderQR<PlainObject>(eval());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_FULLPIVOTINGHOUSEHOLDERQR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR.h
new file mode 100644
index 0000000..3513d99
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR.h
@@ -0,0 +1,409 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2010 Vincent Lejeune
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_QR_H
+#define EIGEN_QR_H
+
+namespace Eigen { 
+
+/** \ingroup QR_Module
+  *
+  *
+  * \class HouseholderQR
+  *
+  * \brief Householder QR decomposition of a matrix
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the QR decomposition
+  *
+  * This class performs a QR decomposition of a matrix \b A into matrices \b Q and \b R
+  * such that 
+  * \f[
+  *  \mathbf{A} = \mathbf{Q} \, \mathbf{R}
+  * \f]
+  * by using Householder transformations. Here, \b Q a unitary matrix and \b R an upper triangular matrix.
+  * The result is stored in a compact way compatible with LAPACK.
+  *
+  * Note that no pivoting is performed. This is \b not a rank-revealing decomposition.
+  * If you want that feature, use FullPivHouseholderQR or ColPivHouseholderQR instead.
+  *
+  * This Householder QR decomposition is faster, but less numerically stable and less feature-full than
+  * FullPivHouseholderQR or ColPivHouseholderQR.
+  *
+  * This class supports the \link InplaceDecomposition inplace decomposition \endlink mechanism.
+  *
+  * \sa MatrixBase::householderQr()
+  */
+template<typename _MatrixType> class HouseholderQR
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    // FIXME should be int
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar, RowsAtCompileTime, RowsAtCompileTime, (MatrixType::Flags&RowMajorBit) ? RowMajor : ColMajor, MaxRowsAtCompileTime, MaxRowsAtCompileTime> MatrixQType;
+    typedef typename internal::plain_diag_type<MatrixType>::type HCoeffsType;
+    typedef typename internal::plain_row_type<MatrixType>::type RowVectorType;
+    typedef HouseholderSequence<MatrixType,typename internal::remove_all<typename HCoeffsType::ConjugateReturnType>::type> HouseholderSequenceType;
+
+    /**
+      * \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via HouseholderQR::compute(const MatrixType&).
+      */
+    HouseholderQR() : m_qr(), m_hCoeffs(), m_temp(), m_isInitialized(false) {}
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem \a size.
+      * \sa HouseholderQR()
+      */
+    HouseholderQR(Index rows, Index cols)
+      : m_qr(rows, cols),
+        m_hCoeffs((std::min)(rows,cols)),
+        m_temp(cols),
+        m_isInitialized(false) {}
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This constructor computes the QR factorization of the matrix \a matrix by calling
+      * the method compute(). It is a short cut for:
+      * 
+      * \code
+      * HouseholderQR<MatrixType> qr(matrix.rows(), matrix.cols());
+      * qr.compute(matrix);
+      * \endcode
+      * 
+      * \sa compute()
+      */
+    template<typename InputType>
+    explicit HouseholderQR(const EigenBase<InputType>& matrix)
+      : m_qr(matrix.rows(), matrix.cols()),
+        m_hCoeffs((std::min)(matrix.rows(),matrix.cols())),
+        m_temp(matrix.cols()),
+        m_isInitialized(false)
+    {
+      compute(matrix.derived());
+    }
+
+
+    /** \brief Constructs a QR factorization from a given matrix
+      *
+      * This overloaded constructor is provided for \link InplaceDecomposition inplace decomposition \endlink when
+      * \c MatrixType is a Eigen::Ref.
+      *
+      * \sa HouseholderQR(const EigenBase&)
+      */
+    template<typename InputType>
+    explicit HouseholderQR(EigenBase<InputType>& matrix)
+      : m_qr(matrix.derived()),
+        m_hCoeffs((std::min)(matrix.rows(),matrix.cols())),
+        m_temp(matrix.cols()),
+        m_isInitialized(false)
+    {
+      computeInPlace();
+    }
+
+    /** This method finds a solution x to the equation Ax=b, where A is the matrix of which
+      * *this is the QR decomposition, if any exists.
+      *
+      * \param b the right-hand-side of the equation to solve.
+      *
+      * \returns a solution.
+      *
+      * \note_about_checking_solutions
+      *
+      * \note_about_arbitrary_choice_of_solution
+      *
+      * Example: \include HouseholderQR_solve.cpp
+      * Output: \verbinclude HouseholderQR_solve.out
+      */
+    template<typename Rhs>
+    inline const Solve<HouseholderQR, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "HouseholderQR is not initialized.");
+      return Solve<HouseholderQR, Rhs>(*this, b.derived());
+    }
+
+    /** This method returns an expression of the unitary matrix Q as a sequence of Householder transformations.
+      *
+      * The returned expression can directly be used to perform matrix products. It can also be assigned to a dense Matrix object.
+      * Here is an example showing how to recover the full or thin matrix Q, as well as how to perform matrix products using operator*:
+      *
+      * Example: \include HouseholderQR_householderQ.cpp
+      * Output: \verbinclude HouseholderQR_householderQ.out
+      */
+    HouseholderSequenceType householderQ() const
+    {
+      eigen_assert(m_isInitialized && "HouseholderQR is not initialized.");
+      return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate());
+    }
+
+    /** \returns a reference to the matrix where the Householder QR decomposition is stored
+      * in a LAPACK-compatible way.
+      */
+    const MatrixType& matrixQR() const
+    {
+        eigen_assert(m_isInitialized && "HouseholderQR is not initialized.");
+        return m_qr;
+    }
+
+    template<typename InputType>
+    HouseholderQR& compute(const EigenBase<InputType>& matrix) {
+      m_qr = matrix.derived();
+      computeInPlace();
+      return *this;
+    }
+
+    /** \returns the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      * One way to work around that is to use logAbsDeterminant() instead.
+      *
+      * \sa logAbsDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar absDeterminant() const;
+
+    /** \returns the natural log of the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition. It has only linear complexity
+      * (that is, O(n) where n is the dimension of the square matrix)
+      * as the QR decomposition has already been computed.
+      *
+      * \note This is only for square matrices.
+      *
+      * \note This method is useful to work around the risk of overflow/underflow that's inherent
+      * to determinant computation.
+      *
+      * \sa absDeterminant(), MatrixBase::determinant()
+      */
+    typename MatrixType::RealScalar logAbsDeterminant() const;
+
+    inline Index rows() const { return m_qr.rows(); }
+    inline Index cols() const { return m_qr.cols(); }
+    
+    /** \returns a const reference to the vector of Householder coefficients used to represent the factor \c Q.
+      * 
+      * For advanced uses only.
+      */
+    const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    void _solve_impl(const RhsType &rhs, DstType &dst) const;
+    #endif
+
+  protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+
+    void computeInPlace();
+    
+    MatrixType m_qr;
+    HCoeffsType m_hCoeffs;
+    RowVectorType m_temp;
+    bool m_isInitialized;
+};
+
+template<typename MatrixType>
+typename MatrixType::RealScalar HouseholderQR<MatrixType>::absDeterminant() const
+{
+  using std::abs;
+  eigen_assert(m_isInitialized && "HouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return abs(m_qr.diagonal().prod());
+}
+
+template<typename MatrixType>
+typename MatrixType::RealScalar HouseholderQR<MatrixType>::logAbsDeterminant() const
+{
+  eigen_assert(m_isInitialized && "HouseholderQR is not initialized.");
+  eigen_assert(m_qr.rows() == m_qr.cols() && "You can't take the determinant of a non-square matrix!");
+  return m_qr.diagonal().cwiseAbs().array().log().sum();
+}
+
+namespace internal {
+
+/** \internal */
+template<typename MatrixQR, typename HCoeffs>
+void householder_qr_inplace_unblocked(MatrixQR& mat, HCoeffs& hCoeffs, typename MatrixQR::Scalar* tempData = 0)
+{
+  typedef typename MatrixQR::Scalar Scalar;
+  typedef typename MatrixQR::RealScalar RealScalar;
+  Index rows = mat.rows();
+  Index cols = mat.cols();
+  Index size = (std::min)(rows,cols);
+
+  eigen_assert(hCoeffs.size() == size);
+
+  typedef Matrix<Scalar,MatrixQR::ColsAtCompileTime,1> TempType;
+  TempType tempVector;
+  if(tempData==0)
+  {
+    tempVector.resize(cols);
+    tempData = tempVector.data();
+  }
+
+  for(Index k = 0; k < size; ++k)
+  {
+    Index remainingRows = rows - k;
+    Index remainingCols = cols - k - 1;
+
+    RealScalar beta;
+    mat.col(k).tail(remainingRows).makeHouseholderInPlace(hCoeffs.coeffRef(k), beta);
+    mat.coeffRef(k,k) = beta;
+
+    // apply H to remaining part of m_qr from the left
+    mat.bottomRightCorner(remainingRows, remainingCols)
+        .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), hCoeffs.coeffRef(k), tempData+k+1);
+  }
+}
+
+/** \internal */
+template<typename MatrixQR, typename HCoeffs,
+  typename MatrixQRScalar = typename MatrixQR::Scalar,
+  bool InnerStrideIsOne = (MatrixQR::InnerStrideAtCompileTime == 1 && HCoeffs::InnerStrideAtCompileTime == 1)>
+struct householder_qr_inplace_blocked
+{
+  // This is specialized for MKL-supported Scalar types in HouseholderQR_MKL.h
+  static void run(MatrixQR& mat, HCoeffs& hCoeffs, Index maxBlockSize=32,
+      typename MatrixQR::Scalar* tempData = 0)
+  {
+    typedef typename MatrixQR::Scalar Scalar;
+    typedef Block<MatrixQR,Dynamic,Dynamic> BlockType;
+
+    Index rows = mat.rows();
+    Index cols = mat.cols();
+    Index size = (std::min)(rows, cols);
+
+    typedef Matrix<Scalar,Dynamic,1,ColMajor,MatrixQR::MaxColsAtCompileTime,1> TempType;
+    TempType tempVector;
+    if(tempData==0)
+    {
+      tempVector.resize(cols);
+      tempData = tempVector.data();
+    }
+
+    Index blockSize = (std::min)(maxBlockSize,size);
+
+    Index k = 0;
+    for (k = 0; k < size; k += blockSize)
+    {
+      Index bs = (std::min)(size-k,blockSize);  // actual size of the block
+      Index tcols = cols - k - bs;              // trailing columns
+      Index brows = rows-k;                     // rows of the block
+
+      // partition the matrix:
+      //        A00 | A01 | A02
+      // mat  = A10 | A11 | A12
+      //        A20 | A21 | A22
+      // and performs the qr dec of [A11^T A12^T]^T
+      // and update [A21^T A22^T]^T using level 3 operations.
+      // Finally, the algorithm continue on A22
+
+      BlockType A11_21 = mat.block(k,k,brows,bs);
+      Block<HCoeffs,Dynamic,1> hCoeffsSegment = hCoeffs.segment(k,bs);
+
+      householder_qr_inplace_unblocked(A11_21, hCoeffsSegment, tempData);
+
+      if(tcols)
+      {
+        BlockType A21_22 = mat.block(k,k+bs,brows,tcols);
+        apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment, false); // false == backward
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename _MatrixType>
+template<typename RhsType, typename DstType>
+void HouseholderQR<_MatrixType>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  const Index rank = (std::min)(rows(), cols());
+  eigen_assert(rhs.rows() == rows());
+
+  typename RhsType::PlainObject c(rhs);
+
+  // Note that the matrix Q = H_0^* H_1^*... so its inverse is Q^* = (H_0 H_1 ...)^T
+  c.applyOnTheLeft(householderSequence(
+    m_qr.leftCols(rank),
+    m_hCoeffs.head(rank)).transpose()
+  );
+
+  m_qr.topLeftCorner(rank, rank)
+      .template triangularView<Upper>()
+      .solveInPlace(c.topRows(rank));
+
+  dst.topRows(rank) = c.topRows(rank);
+  dst.bottomRows(cols()-rank).setZero();
+}
+#endif
+
+/** Performs the QR factorization of the given matrix \a matrix. The result of
+  * the factorization is stored into \c *this, and a reference to \c *this
+  * is returned.
+  *
+  * \sa class HouseholderQR, HouseholderQR(const MatrixType&)
+  */
+template<typename MatrixType>
+void HouseholderQR<MatrixType>::computeInPlace()
+{
+  check_template_parameters();
+  
+  Index rows = m_qr.rows();
+  Index cols = m_qr.cols();
+  Index size = (std::min)(rows,cols);
+
+  m_hCoeffs.resize(size);
+
+  m_temp.resize(cols);
+
+  internal::householder_qr_inplace_blocked<MatrixType, HCoeffsType>::run(m_qr, m_hCoeffs, 48, m_temp.data());
+
+  m_isInitialized = true;
+}
+
+/** \return the Householder QR decomposition of \c *this.
+  *
+  * \sa class HouseholderQR
+  */
+template<typename Derived>
+const HouseholderQR<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::householderQr() const
+{
+  return HouseholderQR<PlainObject>(eval());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_QR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR_LAPACKE.h
new file mode 100644
index 0000000..1dc7d53
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/QR/HouseholderQR_LAPACKE.h
@@ -0,0 +1,68 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Householder QR decomposition of a matrix w/o pivoting based on
+ *    LAPACKE_?geqrf function.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_QR_LAPACKE_H
+#define EIGEN_QR_LAPACKE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_QR_NOPIV(EIGTYPE, LAPACKE_TYPE, LAPACKE_PREFIX) \
+template<typename MatrixQR, typename HCoeffs> \
+struct householder_qr_inplace_blocked<MatrixQR, HCoeffs, EIGTYPE, true> \
+{ \
+  static void run(MatrixQR& mat, HCoeffs& hCoeffs, Index = 32, \
+      typename MatrixQR::Scalar* = 0) \
+  { \
+    lapack_int m = (lapack_int) mat.rows(); \
+    lapack_int n = (lapack_int) mat.cols(); \
+    lapack_int lda = (lapack_int) mat.outerStride(); \
+    lapack_int matrix_order = (MatrixQR::IsRowMajor) ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+    LAPACKE_##LAPACKE_PREFIX##geqrf( matrix_order, m, n, (LAPACKE_TYPE*)mat.data(), lda, (LAPACKE_TYPE*)hCoeffs.data()); \
+    hCoeffs.adjointInPlace(); \
+  } \
+};
+
+EIGEN_LAPACKE_QR_NOPIV(double, double, d)
+EIGEN_LAPACKE_QR_NOPIV(float, float, s)
+EIGEN_LAPACKE_QR_NOPIV(dcomplex, lapack_complex_double, z)
+EIGEN_LAPACKE_QR_NOPIV(scomplex, lapack_complex_float, c)
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_QR_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h
new file mode 100644
index 0000000..953d57c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SPQRSupport/SuiteSparseQRSupport.h
@@ -0,0 +1,313 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Desire Nuentsa <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SUITESPARSEQRSUPPORT_H
+#define EIGEN_SUITESPARSEQRSUPPORT_H
+
+namespace Eigen {
+  
+  template<typename MatrixType> class SPQR; 
+  template<typename SPQRType> struct SPQRMatrixQReturnType; 
+  template<typename SPQRType> struct SPQRMatrixQTransposeReturnType; 
+  template <typename SPQRType, typename Derived> struct SPQR_QProduct;
+  namespace internal {
+    template <typename SPQRType> struct traits<SPQRMatrixQReturnType<SPQRType> >
+    {
+      typedef typename SPQRType::MatrixType ReturnType;
+    };
+    template <typename SPQRType> struct traits<SPQRMatrixQTransposeReturnType<SPQRType> >
+    {
+      typedef typename SPQRType::MatrixType ReturnType;
+    };
+    template <typename SPQRType, typename Derived> struct traits<SPQR_QProduct<SPQRType, Derived> >
+    {
+      typedef typename Derived::PlainObject ReturnType;
+    };
+  } // End namespace internal
+  
+/**
+  * \ingroup SPQRSupport_Module
+  * \class SPQR
+  * \brief Sparse QR factorization based on SuiteSparseQR library
+  *
+  * This class is used to perform a multithreaded and multifrontal rank-revealing QR decomposition
+  * of sparse matrices. The result is then used to solve linear leasts_square systems.
+  * Clearly, a QR factorization is returned such that A*P = Q*R where :
+  *
+  * P is the column permutation. Use colsPermutation() to get it.
+  *
+  * Q is the orthogonal matrix represented as Householder reflectors.
+  * Use matrixQ() to get an expression and matrixQ().transpose() to get the transpose.
+  * You can then apply it to a vector.
+  *
+  * R is the sparse triangular factor. Use matrixQR() to get it as SparseMatrix.
+  * NOTE : The Index type of R is always SuiteSparse_long. You can get it with SPQR::Index
+  *
+  * \tparam _MatrixType The type of the sparse matrix A, must be a column-major SparseMatrix<>
+  *
+  * \implsparsesolverconcept
+  *
+  *
+  */
+template<typename _MatrixType>
+class SPQR : public SparseSolverBase<SPQR<_MatrixType> >
+{
+  protected:
+    typedef SparseSolverBase<SPQR<_MatrixType> > Base;
+    using Base::m_isInitialized;
+  public:
+    typedef typename _MatrixType::Scalar Scalar;
+    typedef typename _MatrixType::RealScalar RealScalar;
+    typedef SuiteSparse_long StorageIndex ;
+    typedef SparseMatrix<Scalar, ColMajor, StorageIndex> MatrixType;
+    typedef Map<PermutationMatrix<Dynamic, Dynamic, StorageIndex> > PermutationType;
+    enum {
+      ColsAtCompileTime = Dynamic,
+      MaxColsAtCompileTime = Dynamic
+    };
+  public:
+    SPQR() 
+      : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits<Scalar>::epsilon()), m_useDefaultThreshold(true)
+    { 
+      cholmod_l_start(&m_cc);
+    }
+    
+    explicit SPQR(const _MatrixType& matrix)
+    : m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits<Scalar>::epsilon()), m_useDefaultThreshold(true)
+    {
+      cholmod_l_start(&m_cc);
+      compute(matrix);
+    }
+    
+    ~SPQR()
+    {
+      SPQR_free();
+      cholmod_l_finish(&m_cc);
+    }
+    void SPQR_free()
+    {
+      cholmod_l_free_sparse(&m_H, &m_cc);
+      cholmod_l_free_sparse(&m_cR, &m_cc);
+      cholmod_l_free_dense(&m_HTau, &m_cc);
+      std::free(m_E);
+      std::free(m_HPinv);
+    }
+
+    void compute(const _MatrixType& matrix)
+    {
+      if(m_isInitialized) SPQR_free();
+
+      MatrixType mat(matrix);
+      
+      /* Compute the default threshold as in MatLab, see:
+       * Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing
+       * Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011, Page 8:3 
+       */
+      RealScalar pivotThreshold = m_tolerance;
+      if(m_useDefaultThreshold) 
+      {
+        RealScalar max2Norm = 0.0;
+        for (int j = 0; j < mat.cols(); j++) max2Norm = numext::maxi(max2Norm, mat.col(j).norm());
+        if(max2Norm==RealScalar(0))
+          max2Norm = RealScalar(1);
+        pivotThreshold = 20 * (mat.rows() + mat.cols()) * max2Norm * NumTraits<RealScalar>::epsilon();
+      }
+      cholmod_sparse A; 
+      A = viewAsCholmod(mat);
+      m_rows = matrix.rows();
+      Index col = matrix.cols();
+      m_rank = SuiteSparseQR<Scalar>(m_ordering, pivotThreshold, col, &A, 
+                             &m_cR, &m_E, &m_H, &m_HPinv, &m_HTau, &m_cc);
+
+      if (!m_cR)
+      {
+        m_info = NumericalIssue;
+        m_isInitialized = false;
+        return;
+      }
+      m_info = Success;
+      m_isInitialized = true;
+      m_isRUpToDate = false;
+    }
+    /** 
+     * Get the number of rows of the input matrix and the Q matrix
+     */
+    inline Index rows() const {return m_rows; }
+    
+    /** 
+     * Get the number of columns of the input matrix. 
+     */
+    inline Index cols() const { return m_cR->ncol; }
+    
+    template<typename Rhs, typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    {
+      eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()");
+      eigen_assert(b.cols()==1 && "This method is for vectors only");
+
+      //Compute Q^T * b
+      typename Dest::PlainObject y, y2;
+      y = matrixQ().transpose() * b;
+      
+      // Solves with the triangular matrix R
+      Index rk = this->rank();
+      y2 = y;
+      y.resize((std::max)(cols(),Index(y.rows())),y.cols());
+      y.topRows(rk) = this->matrixR().topLeftCorner(rk, rk).template triangularView<Upper>().solve(y2.topRows(rk));
+
+      // Apply the column permutation 
+      // colsPermutation() performs a copy of the permutation,
+      // so let's apply it manually:
+      for(Index i = 0; i < rk; ++i) dest.row(m_E[i]) = y.row(i);
+      for(Index i = rk; i < cols(); ++i) dest.row(m_E[i]).setZero();
+      
+//       y.bottomRows(y.rows()-rk).setZero();
+//       dest = colsPermutation() * y.topRows(cols());
+      
+      m_info = Success;
+    }
+    
+    /** \returns the sparse triangular factor R. It is a sparse matrix
+     */
+    const MatrixType matrixR() const
+    {
+      eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()");
+      if(!m_isRUpToDate) {
+        m_R = viewAsEigen<Scalar,ColMajor, typename MatrixType::StorageIndex>(*m_cR);
+        m_isRUpToDate = true;
+      }
+      return m_R;
+    }
+    /// Get an expression of the matrix Q
+    SPQRMatrixQReturnType<SPQR> matrixQ() const
+    {
+      return SPQRMatrixQReturnType<SPQR>(*this);
+    }
+    /// Get the permutation that was applied to columns of A
+    PermutationType colsPermutation() const
+    { 
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return PermutationType(m_E, m_cR->ncol);
+    }
+    /**
+     * Gets the rank of the matrix. 
+     * It should be equal to matrixQR().cols if the matrix is full-rank
+     */
+    Index rank() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_cc.SPQR_istat[4];
+    }
+    /// Set the fill-reducing ordering method to be used
+    void setSPQROrdering(int ord) { m_ordering = ord;}
+    /// Set the tolerance tol to treat columns with 2-norm < =tol as zero
+    void setPivotThreshold(const RealScalar& tol)
+    {
+      m_useDefaultThreshold = false;
+      m_tolerance = tol;
+    }
+    
+    /** \returns a pointer to the SPQR workspace */
+    cholmod_common *cholmodCommon() const { return &m_cc; }
+    
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the sparse QR can not be computed
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+  protected:
+    bool m_analysisIsOk;
+    bool m_factorizationIsOk;
+    mutable bool m_isRUpToDate;
+    mutable ComputationInfo m_info;
+    int m_ordering; // Ordering method to use, see SPQR's manual
+    int m_allow_tol; // Allow to use some tolerance during numerical factorization.
+    RealScalar m_tolerance; // treat columns with 2-norm below this tolerance as zero
+    mutable cholmod_sparse *m_cR; // The sparse R factor in cholmod format
+    mutable MatrixType m_R; // The sparse matrix R in Eigen format
+    mutable StorageIndex *m_E; // The permutation applied to columns
+    mutable cholmod_sparse *m_H;  //The householder vectors
+    mutable StorageIndex *m_HPinv; // The row permutation of H
+    mutable cholmod_dense *m_HTau; // The Householder coefficients
+    mutable Index m_rank; // The rank of the matrix
+    mutable cholmod_common m_cc; // Workspace and parameters
+    bool m_useDefaultThreshold;     // Use default threshold
+    Index m_rows;
+    template<typename ,typename > friend struct SPQR_QProduct;
+};
+
+template <typename SPQRType, typename Derived>
+struct SPQR_QProduct : ReturnByValue<SPQR_QProduct<SPQRType,Derived> >
+{
+  typedef typename SPQRType::Scalar Scalar;
+  typedef typename SPQRType::StorageIndex StorageIndex;
+  //Define the constructor to get reference to argument types
+  SPQR_QProduct(const SPQRType& spqr, const Derived& other, bool transpose) : m_spqr(spqr),m_other(other),m_transpose(transpose) {}
+  
+  inline Index rows() const { return m_transpose ? m_spqr.rows() : m_spqr.cols(); }
+  inline Index cols() const { return m_other.cols(); }
+  // Assign to a vector
+  template<typename ResType>
+  void evalTo(ResType& res) const
+  {
+    cholmod_dense y_cd;
+    cholmod_dense *x_cd; 
+    int method = m_transpose ? SPQR_QTX : SPQR_QX; 
+    cholmod_common *cc = m_spqr.cholmodCommon();
+    y_cd = viewAsCholmod(m_other.const_cast_derived());
+    x_cd = SuiteSparseQR_qmult<Scalar>(method, m_spqr.m_H, m_spqr.m_HTau, m_spqr.m_HPinv, &y_cd, cc);
+    res = Matrix<Scalar,ResType::RowsAtCompileTime,ResType::ColsAtCompileTime>::Map(reinterpret_cast<Scalar*>(x_cd->x), x_cd->nrow, x_cd->ncol);
+    cholmod_l_free_dense(&x_cd, cc);
+  }
+  const SPQRType& m_spqr; 
+  const Derived& m_other; 
+  bool m_transpose; 
+  
+};
+template<typename SPQRType>
+struct SPQRMatrixQReturnType{
+  
+  SPQRMatrixQReturnType(const SPQRType& spqr) : m_spqr(spqr) {}
+  template<typename Derived>
+  SPQR_QProduct<SPQRType, Derived> operator*(const MatrixBase<Derived>& other)
+  {
+    return SPQR_QProduct<SPQRType,Derived>(m_spqr,other.derived(),false);
+  }
+  SPQRMatrixQTransposeReturnType<SPQRType> adjoint() const
+  {
+    return SPQRMatrixQTransposeReturnType<SPQRType>(m_spqr);
+  }
+  // To use for operations with the transpose of Q
+  SPQRMatrixQTransposeReturnType<SPQRType> transpose() const
+  {
+    return SPQRMatrixQTransposeReturnType<SPQRType>(m_spqr);
+  }
+  const SPQRType& m_spqr;
+};
+
+template<typename SPQRType>
+struct SPQRMatrixQTransposeReturnType{
+  SPQRMatrixQTransposeReturnType(const SPQRType& spqr) : m_spqr(spqr) {}
+  template<typename Derived>
+  SPQR_QProduct<SPQRType,Derived> operator*(const MatrixBase<Derived>& other)
+  {
+    return SPQR_QProduct<SPQRType,Derived>(m_spqr,other.derived(), true);
+  }
+  const SPQRType& m_spqr;
+};
+
+}// End namespace Eigen
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/BDCSVD.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/BDCSVD.h
new file mode 100644
index 0000000..d7a4271
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/BDCSVD.h
@@ -0,0 +1,1231 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+// 
+// We used the "A Divide-And-Conquer Algorithm for the Bidiagonal SVD"
+// research report written by Ming Gu and Stanley C.Eisenstat
+// The code variable names correspond to the names they used in their 
+// report
+//
+// Copyright (C) 2013 Gauthier Brun <brun.gauthier@gmail.com>
+// Copyright (C) 2013 Nicolas Carre <nicolas.carre@ensimag.fr>
+// Copyright (C) 2013 Jean Ceccato <jean.ceccato@ensimag.fr>
+// Copyright (C) 2013 Pierre Zoppitelli <pierre.zoppitelli@ensimag.fr>
+// Copyright (C) 2013 Jitse Niesen <jitse@maths.leeds.ac.uk>
+// Copyright (C) 2014-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BDCSVD_H
+#define EIGEN_BDCSVD_H
+// #define EIGEN_BDCSVD_DEBUG_VERBOSE
+// #define EIGEN_BDCSVD_SANITY_CHECKS
+
+namespace Eigen {
+
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+IOFormat bdcsvdfmt(8, 0, ", ", "\n", "  [", "]");
+#endif
+  
+template<typename _MatrixType> class BDCSVD;
+
+namespace internal {
+
+template<typename _MatrixType> 
+struct traits<BDCSVD<_MatrixType> >
+{
+  typedef _MatrixType MatrixType;
+};  
+
+} // end namespace internal
+  
+  
+/** \ingroup SVD_Module
+ *
+ *
+ * \class BDCSVD
+ *
+ * \brief class Bidiagonal Divide and Conquer SVD
+ *
+ * \tparam _MatrixType the type of the matrix of which we are computing the SVD decomposition
+ *
+ * This class first reduces the input matrix to bi-diagonal form using class UpperBidiagonalization,
+ * and then performs a divide-and-conquer diagonalization. Small blocks are diagonalized using class JacobiSVD.
+ * You can control the switching size with the setSwitchSize() method, default is 16.
+ * For small matrice (<16), it is thus preferable to directly use JacobiSVD. For larger ones, BDCSVD is highly
+ * recommended and can several order of magnitude faster.
+ *
+ * \warning this algorithm is unlikely to provide accurate result when compiled with unsafe math optimizations.
+ * For instance, this concerns Intel's compiler (ICC), which perfroms such optimization by default unless
+ * you compile with the \c -fp-model \c precise option. Likewise, the \c -ffast-math option of GCC or clang will
+ * significantly degrade the accuracy.
+ *
+ * \sa class JacobiSVD
+ */
+template<typename _MatrixType> 
+class BDCSVD : public SVDBase<BDCSVD<_MatrixType> >
+{
+  typedef SVDBase<BDCSVD> Base;
+    
+public:
+  using Base::rows;
+  using Base::cols;
+  using Base::computeU;
+  using Base::computeV;
+  
+  typedef _MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+  typedef typename NumTraits<RealScalar>::Literal Literal;
+  enum {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime, 
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime, 
+    DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime, ColsAtCompileTime), 
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime, 
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime, 
+    MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime, MaxColsAtCompileTime), 
+    MatrixOptions = MatrixType::Options
+  };
+
+  typedef typename Base::MatrixUType MatrixUType;
+  typedef typename Base::MatrixVType MatrixVType;
+  typedef typename Base::SingularValuesType SingularValuesType;
+  
+  typedef Matrix<Scalar, Dynamic, Dynamic, ColMajor> MatrixX;
+  typedef Matrix<RealScalar, Dynamic, Dynamic, ColMajor> MatrixXr;
+  typedef Matrix<RealScalar, Dynamic, 1> VectorType;
+  typedef Array<RealScalar, Dynamic, 1> ArrayXr;
+  typedef Array<Index,1,Dynamic> ArrayXi;
+  typedef Ref<ArrayXr> ArrayRef;
+  typedef Ref<ArrayXi> IndicesRef;
+
+  /** \brief Default Constructor.
+   *
+   * The default constructor is useful in cases in which the user intends to
+   * perform decompositions via BDCSVD::compute(const MatrixType&).
+   */
+  BDCSVD() : m_algoswap(16), m_numIters(0)
+  {}
+
+
+  /** \brief Default Constructor with memory preallocation
+   *
+   * Like the default constructor but with preallocation of the internal data
+   * according to the specified problem size.
+   * \sa BDCSVD()
+   */
+  BDCSVD(Index rows, Index cols, unsigned int computationOptions = 0)
+    : m_algoswap(16), m_numIters(0)
+  {
+    allocate(rows, cols, computationOptions);
+  }
+
+  /** \brief Constructor performing the decomposition of given matrix.
+   *
+   * \param matrix the matrix to decompose
+   * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
+   *                           By default, none is computed. This is a bit - field, the possible bits are #ComputeFullU, #ComputeThinU, 
+   *                           #ComputeFullV, #ComputeThinV.
+   *
+   * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
+   * available with the (non - default) FullPivHouseholderQR preconditioner.
+   */
+  BDCSVD(const MatrixType& matrix, unsigned int computationOptions = 0)
+    : m_algoswap(16), m_numIters(0)
+  {
+    compute(matrix, computationOptions);
+  }
+
+  ~BDCSVD() 
+  {
+  }
+  
+  /** \brief Method performing the decomposition of given matrix using custom options.
+   *
+   * \param matrix the matrix to decompose
+   * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
+   *                           By default, none is computed. This is a bit - field, the possible bits are #ComputeFullU, #ComputeThinU, 
+   *                           #ComputeFullV, #ComputeThinV.
+   *
+   * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
+   * available with the (non - default) FullPivHouseholderQR preconditioner.
+   */
+  BDCSVD& compute(const MatrixType& matrix, unsigned int computationOptions);
+
+  /** \brief Method performing the decomposition of given matrix using current options.
+   *
+   * \param matrix the matrix to decompose
+   *
+   * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int).
+   */
+  BDCSVD& compute(const MatrixType& matrix)
+  {
+    return compute(matrix, this->m_computationOptions);
+  }
+
+  void setSwitchSize(int s) 
+  {
+    eigen_assert(s>3 && "BDCSVD the size of the algo switch has to be greater than 3");
+    m_algoswap = s;
+  }
+ 
+private:
+  void allocate(Index rows, Index cols, unsigned int computationOptions);
+  void divide(Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift);
+  void computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V);
+  void computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, VectorType& singVals, ArrayRef shifts, ArrayRef mus);
+  void perturbCol0(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat);
+  void computeSingVecs(const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef& perm, const VectorType& singVals, const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V);
+  void deflation43(Index firstCol, Index shift, Index i, Index size);
+  void deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size);
+  void deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift);
+  template<typename HouseholderU, typename HouseholderV, typename NaiveU, typename NaiveV>
+  void copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naivev);
+  void structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1);
+  static RealScalar secularEq(RealScalar x, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift);
+
+protected:
+  MatrixXr m_naiveU, m_naiveV;
+  MatrixXr m_computed;
+  Index m_nRec;
+  ArrayXr m_workspace;
+  ArrayXi m_workspaceI;
+  int m_algoswap;
+  bool m_isTranspose, m_compU, m_compV;
+  
+  using Base::m_singularValues;
+  using Base::m_diagSize;
+  using Base::m_computeFullU;
+  using Base::m_computeFullV;
+  using Base::m_computeThinU;
+  using Base::m_computeThinV;
+  using Base::m_matrixU;
+  using Base::m_matrixV;
+  using Base::m_isInitialized;
+  using Base::m_nonzeroSingularValues;
+
+public:  
+  int m_numIters;
+}; //end class BDCSVD
+
+
+// Method to allocate and initialize matrix and attributes
+template<typename MatrixType>
+void BDCSVD<MatrixType>::allocate(Index rows, Index cols, unsigned int computationOptions)
+{
+  m_isTranspose = (cols > rows);
+
+  if (Base::allocate(rows, cols, computationOptions))
+    return;
+  
+  m_computed = MatrixXr::Zero(m_diagSize + 1, m_diagSize );
+  m_compU = computeV();
+  m_compV = computeU();
+  if (m_isTranspose)
+    std::swap(m_compU, m_compV);
+  
+  if (m_compU) m_naiveU = MatrixXr::Zero(m_diagSize + 1, m_diagSize + 1 );
+  else         m_naiveU = MatrixXr::Zero(2, m_diagSize + 1 );
+  
+  if (m_compV) m_naiveV = MatrixXr::Zero(m_diagSize, m_diagSize);
+  
+  m_workspace.resize((m_diagSize+1)*(m_diagSize+1)*3);
+  m_workspaceI.resize(3*m_diagSize);
+}// end allocate
+
+template<typename MatrixType>
+BDCSVD<MatrixType>& BDCSVD<MatrixType>::compute(const MatrixType& matrix, unsigned int computationOptions) 
+{
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "\n\n\n======================================================================================================================\n\n\n";
+#endif
+  allocate(matrix.rows(), matrix.cols(), computationOptions);
+  using std::abs;
+
+  const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
+  
+  //**** step -1 - If the problem is too small, directly falls back to JacobiSVD and return
+  if(matrix.cols() < m_algoswap)
+  {
+    // FIXME this line involves temporaries
+    JacobiSVD<MatrixType> jsvd(matrix,computationOptions);
+    if(computeU()) m_matrixU = jsvd.matrixU();
+    if(computeV()) m_matrixV = jsvd.matrixV();
+    m_singularValues = jsvd.singularValues();
+    m_nonzeroSingularValues = jsvd.nonzeroSingularValues();
+    m_isInitialized = true;
+    return *this;
+  }
+  
+  //**** step 0 - Copy the input matrix and apply scaling to reduce over/under-flows
+  RealScalar scale = matrix.cwiseAbs().maxCoeff();
+  if(scale==Literal(0)) scale = Literal(1);
+  MatrixX copy;
+  if (m_isTranspose) copy = matrix.adjoint()/scale;
+  else               copy = matrix/scale;
+  
+  //**** step 1 - Bidiagonalization
+  // FIXME this line involves temporaries
+  internal::UpperBidiagonalization<MatrixX> bid(copy);
+
+  //**** step 2 - Divide & Conquer
+  m_naiveU.setZero();
+  m_naiveV.setZero();
+  // FIXME this line involves a temporary matrix
+  m_computed.topRows(m_diagSize) = bid.bidiagonal().toDenseMatrix().transpose();
+  m_computed.template bottomRows<1>().setZero();
+  divide(0, m_diagSize - 1, 0, 0, 0);
+
+  //**** step 3 - Copy singular values and vectors
+  for (int i=0; i<m_diagSize; i++)
+  {
+    RealScalar a = abs(m_computed.coeff(i, i));
+    m_singularValues.coeffRef(i) = a * scale;
+    if (a<considerZero)
+    {
+      m_nonzeroSingularValues = i;
+      m_singularValues.tail(m_diagSize - i - 1).setZero();
+      break;
+    }
+    else if (i == m_diagSize - 1)
+    {
+      m_nonzeroSingularValues = i + 1;
+      break;
+    }
+  }
+
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+//   std::cout << "m_naiveU\n" << m_naiveU << "\n\n";
+//   std::cout << "m_naiveV\n" << m_naiveV << "\n\n";
+#endif
+  if(m_isTranspose) copyUV(bid.householderV(), bid.householderU(), m_naiveV, m_naiveU);
+  else              copyUV(bid.householderU(), bid.householderV(), m_naiveU, m_naiveV);
+
+  m_isInitialized = true;
+  return *this;
+}// end compute
+
+
+template<typename MatrixType>
+template<typename HouseholderU, typename HouseholderV, typename NaiveU, typename NaiveV>
+void BDCSVD<MatrixType>::copyUV(const HouseholderU &householderU, const HouseholderV &householderV, const NaiveU &naiveU, const NaiveV &naiveV)
+{
+  // Note exchange of U and V: m_matrixU is set from m_naiveV and vice versa
+  if (computeU())
+  {
+    Index Ucols = m_computeThinU ? m_diagSize : householderU.cols();
+    m_matrixU = MatrixX::Identity(householderU.cols(), Ucols);
+    m_matrixU.topLeftCorner(m_diagSize, m_diagSize) = naiveV.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
+    householderU.applyThisOnTheLeft(m_matrixU); // FIXME this line involves a temporary buffer
+  }
+  if (computeV())
+  {
+    Index Vcols = m_computeThinV ? m_diagSize : householderV.cols();
+    m_matrixV = MatrixX::Identity(householderV.cols(), Vcols);
+    m_matrixV.topLeftCorner(m_diagSize, m_diagSize) = naiveU.template cast<Scalar>().topLeftCorner(m_diagSize, m_diagSize);
+    householderV.applyThisOnTheLeft(m_matrixV); // FIXME this line involves a temporary buffer
+  }
+}
+
+/** \internal
+  * Performs A = A * B exploiting the special structure of the matrix A. Splitting A as:
+  *  A = [A1]
+  *      [A2]
+  * such that A1.rows()==n1, then we assume that at least half of the columns of A1 and A2 are zeros.
+  * We can thus pack them prior to the the matrix product. However, this is only worth the effort if the matrix is large
+  * enough.
+  */
+template<typename MatrixType>
+void BDCSVD<MatrixType>::structured_update(Block<MatrixXr,Dynamic,Dynamic> A, const MatrixXr &B, Index n1)
+{
+  Index n = A.rows();
+  if(n>100)
+  {
+    // If the matrices are large enough, let's exploit the sparse structure of A by
+    // splitting it in half (wrt n1), and packing the non-zero columns.
+    Index n2 = n - n1;
+    Map<MatrixXr> A1(m_workspace.data()      , n1, n);
+    Map<MatrixXr> A2(m_workspace.data()+ n1*n, n2, n);
+    Map<MatrixXr> B1(m_workspace.data()+  n*n, n,  n);
+    Map<MatrixXr> B2(m_workspace.data()+2*n*n, n,  n);
+    Index k1=0, k2=0;
+    for(Index j=0; j<n; ++j)
+    {
+      if( (A.col(j).head(n1).array()!=Literal(0)).any() )
+      {
+        A1.col(k1) = A.col(j).head(n1);
+        B1.row(k1) = B.row(j);
+        ++k1;
+      }
+      if( (A.col(j).tail(n2).array()!=Literal(0)).any() )
+      {
+        A2.col(k2) = A.col(j).tail(n2);
+        B2.row(k2) = B.row(j);
+        ++k2;
+      }
+    }
+  
+    A.topRows(n1).noalias()    = A1.leftCols(k1) * B1.topRows(k1);
+    A.bottomRows(n2).noalias() = A2.leftCols(k2) * B2.topRows(k2);
+  }
+  else
+  {
+    Map<MatrixXr,Aligned> tmp(m_workspace.data(),n,n);
+    tmp.noalias() = A*B;
+    A = tmp;
+  }
+}
+
+// The divide algorithm is done "in place", we are always working on subsets of the same matrix. The divide methods takes as argument the 
+// place of the submatrix we are currently working on.
+
+//@param firstCol : The Index of the first column of the submatrix of m_computed and for m_naiveU;
+//@param lastCol : The Index of the last column of the submatrix of m_computed and for m_naiveU; 
+// lastCol + 1 - firstCol is the size of the submatrix.
+//@param firstRowW : The Index of the first row of the matrix W that we are to change. (see the reference paper section 1 for more information on W)
+//@param firstRowW : Same as firstRowW with the column.
+//@param shift : Each time one takes the left submatrix, one must add 1 to the shift. Why? Because! We actually want the last column of the U submatrix 
+// to become the first column (*coeff) and to shift all the other columns to the right. There are more details on the reference paper.
+template<typename MatrixType>
+void BDCSVD<MatrixType>::divide (Index firstCol, Index lastCol, Index firstRowW, Index firstColW, Index shift)
+{
+  // requires rows = cols + 1;
+  using std::pow;
+  using std::sqrt;
+  using std::abs;
+  const Index n = lastCol - firstCol + 1;
+  const Index k = n/2;
+  const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
+  RealScalar alphaK;
+  RealScalar betaK; 
+  RealScalar r0; 
+  RealScalar lambda, phi, c0, s0;
+  VectorType l, f;
+  // We use the other algorithm which is more efficient for small 
+  // matrices.
+  if (n < m_algoswap)
+  {
+    // FIXME this line involves temporaries
+    JacobiSVD<MatrixXr> b(m_computed.block(firstCol, firstCol, n + 1, n), ComputeFullU | (m_compV ? ComputeFullV : 0));
+    if (m_compU)
+      m_naiveU.block(firstCol, firstCol, n + 1, n + 1).real() = b.matrixU();
+    else 
+    {
+      m_naiveU.row(0).segment(firstCol, n + 1).real() = b.matrixU().row(0);
+      m_naiveU.row(1).segment(firstCol, n + 1).real() = b.matrixU().row(n);
+    }
+    if (m_compV) m_naiveV.block(firstRowW, firstColW, n, n).real() = b.matrixV();
+    m_computed.block(firstCol + shift, firstCol + shift, n + 1, n).setZero();
+    m_computed.diagonal().segment(firstCol + shift, n) = b.singularValues().head(n);
+    return;
+  }
+  // We use the divide and conquer algorithm
+  alphaK =  m_computed(firstCol + k, firstCol + k);
+  betaK = m_computed(firstCol + k + 1, firstCol + k);
+  // The divide must be done in that order in order to have good results. Divide change the data inside the submatrices
+  // and the divide of the right submatrice reads one column of the left submatrice. That's why we need to treat the 
+  // right submatrix before the left one. 
+  divide(k + 1 + firstCol, lastCol, k + 1 + firstRowW, k + 1 + firstColW, shift);
+  divide(firstCol, k - 1 + firstCol, firstRowW, firstColW + 1, shift + 1);
+
+  if (m_compU)
+  {
+    lambda = m_naiveU(firstCol + k, firstCol + k);
+    phi = m_naiveU(firstCol + k + 1, lastCol + 1);
+  } 
+  else 
+  {
+    lambda = m_naiveU(1, firstCol + k);
+    phi = m_naiveU(0, lastCol + 1);
+  }
+  r0 = sqrt((abs(alphaK * lambda) * abs(alphaK * lambda)) + abs(betaK * phi) * abs(betaK * phi));
+  if (m_compU)
+  {
+    l = m_naiveU.row(firstCol + k).segment(firstCol, k);
+    f = m_naiveU.row(firstCol + k + 1).segment(firstCol + k + 1, n - k - 1);
+  } 
+  else 
+  {
+    l = m_naiveU.row(1).segment(firstCol, k);
+    f = m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1);
+  }
+  if (m_compV) m_naiveV(firstRowW+k, firstColW) = Literal(1);
+  if (r0<considerZero)
+  {
+    c0 = Literal(1);
+    s0 = Literal(0);
+  }
+  else
+  {
+    c0 = alphaK * lambda / r0;
+    s0 = betaK * phi / r0;
+  }
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+  
+  if (m_compU)
+  {
+    MatrixXr q1 (m_naiveU.col(firstCol + k).segment(firstCol, k + 1));     
+    // we shiftW Q1 to the right
+    for (Index i = firstCol + k - 1; i >= firstCol; i--) 
+      m_naiveU.col(i + 1).segment(firstCol, k + 1) = m_naiveU.col(i).segment(firstCol, k + 1);
+    // we shift q1 at the left with a factor c0
+    m_naiveU.col(firstCol).segment( firstCol, k + 1) = (q1 * c0);
+    // last column = q1 * - s0
+    m_naiveU.col(lastCol + 1).segment(firstCol, k + 1) = (q1 * ( - s0));
+    // first column = q2 * s0
+    m_naiveU.col(firstCol).segment(firstCol + k + 1, n - k) = m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) * s0; 
+    // q2 *= c0
+    m_naiveU.col(lastCol + 1).segment(firstCol + k + 1, n - k) *= c0;
+  } 
+  else 
+  {
+    RealScalar q1 = m_naiveU(0, firstCol + k);
+    // we shift Q1 to the right
+    for (Index i = firstCol + k - 1; i >= firstCol; i--) 
+      m_naiveU(0, i + 1) = m_naiveU(0, i);
+    // we shift q1 at the left with a factor c0
+    m_naiveU(0, firstCol) = (q1 * c0);
+    // last column = q1 * - s0
+    m_naiveU(0, lastCol + 1) = (q1 * ( - s0));
+    // first column = q2 * s0
+    m_naiveU(1, firstCol) = m_naiveU(1, lastCol + 1) *s0; 
+    // q2 *= c0
+    m_naiveU(1, lastCol + 1) *= c0;
+    m_naiveU.row(1).segment(firstCol + 1, k).setZero();
+    m_naiveU.row(0).segment(firstCol + k + 1, n - k - 1).setZero();
+  }
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+  
+  m_computed(firstCol + shift, firstCol + shift) = r0;
+  m_computed.col(firstCol + shift).segment(firstCol + shift + 1, k) = alphaK * l.transpose().real();
+  m_computed.col(firstCol + shift).segment(firstCol + shift + k + 1, n - k - 1) = betaK * f.transpose().real();
+
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  ArrayXr tmp1 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues();
+#endif
+  // Second part: try to deflate singular values in combined matrix
+  deflation(firstCol, lastCol, k, firstRowW, firstColW, shift);
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  ArrayXr tmp2 = (m_computed.block(firstCol+shift, firstCol+shift, n, n)).jacobiSvd().singularValues();
+  std::cout << "\n\nj1 = " << tmp1.transpose().format(bdcsvdfmt) << "\n";
+  std::cout << "j2 = " << tmp2.transpose().format(bdcsvdfmt) << "\n\n";
+  std::cout << "err:      " << ((tmp1-tmp2).abs()>1e-12*tmp2.abs()).transpose() << "\n";
+  static int count = 0;
+  std::cout << "# " << ++count << "\n\n";
+  assert((tmp1-tmp2).matrix().norm() < 1e-14*tmp2.matrix().norm());
+//   assert(count<681);
+//   assert(((tmp1-tmp2).abs()<1e-13*tmp2.abs()).all());
+#endif
+  
+  // Third part: compute SVD of combined matrix
+  MatrixXr UofSVD, VofSVD;
+  VectorType singVals;
+  computeSVDofM(firstCol + shift, n, UofSVD, singVals, VofSVD);
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(UofSVD.allFinite());
+  assert(VofSVD.allFinite());
+#endif
+  
+  if (m_compU)
+    structured_update(m_naiveU.block(firstCol, firstCol, n + 1, n + 1), UofSVD, (n+2)/2);
+  else
+  {
+    Map<Matrix<RealScalar,2,Dynamic>,Aligned> tmp(m_workspace.data(),2,n+1);
+    tmp.noalias() = m_naiveU.middleCols(firstCol, n+1) * UofSVD;
+    m_naiveU.middleCols(firstCol, n + 1) = tmp;
+  }
+  
+  if (m_compV)  structured_update(m_naiveV.block(firstRowW, firstColW, n, n), VofSVD, (n+1)/2);
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+  
+  m_computed.block(firstCol + shift, firstCol + shift, n, n).setZero();
+  m_computed.block(firstCol + shift, firstCol + shift, n, n).diagonal() = singVals;
+}// end divide
+
+// Compute SVD of m_computed.block(firstCol, firstCol, n + 1, n); this block only has non-zeros in
+// the first column and on the diagonal and has undergone deflation, so diagonal is in increasing
+// order except for possibly the (0,0) entry. The computed SVD is stored U, singVals and V, except
+// that if m_compV is false, then V is not computed. Singular values are sorted in decreasing order.
+//
+// TODO Opportunities for optimization: better root finding algo, better stopping criterion, better
+// handling of round-off errors, be consistent in ordering
+// For instance, to solve the secular equation using FMM, see http://www.stat.uchicago.edu/~lekheng/courses/302/classics/greengard-rokhlin.pdf
+template <typename MatrixType>
+void BDCSVD<MatrixType>::computeSVDofM(Index firstCol, Index n, MatrixXr& U, VectorType& singVals, MatrixXr& V)
+{
+  const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
+  using std::abs;
+  ArrayRef col0 = m_computed.col(firstCol).segment(firstCol, n);
+  m_workspace.head(n) =  m_computed.block(firstCol, firstCol, n, n).diagonal();
+  ArrayRef diag = m_workspace.head(n);
+  diag(0) = Literal(0);
+
+  // Allocate space for singular values and vectors
+  singVals.resize(n);
+  U.resize(n+1, n+1);
+  if (m_compV) V.resize(n, n);
+
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  if (col0.hasNaN() || diag.hasNaN())
+    std::cout << "\n\nHAS NAN\n\n";
+#endif
+  
+  // Many singular values might have been deflated, the zero ones have been moved to the end,
+  // but others are interleaved and we must ignore them at this stage.
+  // To this end, let's compute a permutation skipping them:
+  Index actual_n = n;
+  while(actual_n>1 && diag(actual_n-1)==Literal(0)) --actual_n;
+  Index m = 0; // size of the deflated problem
+  for(Index k=0;k<actual_n;++k)
+    if(abs(col0(k))>considerZero)
+      m_workspaceI(m++) = k;
+  Map<ArrayXi> perm(m_workspaceI.data(),m);
+  
+  Map<ArrayXr> shifts(m_workspace.data()+1*n, n);
+  Map<ArrayXr> mus(m_workspace.data()+2*n, n);
+  Map<ArrayXr> zhat(m_workspace.data()+3*n, n);
+
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "computeSVDofM using:\n";
+  std::cout << "  z: " << col0.transpose() << "\n";
+  std::cout << "  d: " << diag.transpose() << "\n";
+#endif
+  
+  // Compute singVals, shifts, and mus
+  computeSingVals(col0, diag, perm, singVals, shifts, mus);
+  
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "  j:        " << (m_computed.block(firstCol, firstCol, n, n)).jacobiSvd().singularValues().transpose().reverse() << "\n\n";
+  std::cout << "  sing-val: " << singVals.transpose() << "\n";
+  std::cout << "  mu:       " << mus.transpose() << "\n";
+  std::cout << "  shift:    " << shifts.transpose() << "\n";
+  
+  {
+    Index actual_n = n;
+    while(actual_n>1 && abs(col0(actual_n-1))<considerZero) --actual_n;
+    std::cout << "\n\n    mus:    " << mus.head(actual_n).transpose() << "\n\n";
+    std::cout << "    check1 (expect0) : " << ((singVals.array()-(shifts+mus)) / singVals.array()).head(actual_n).transpose() << "\n\n";
+    std::cout << "    check2 (>0)      : " << ((singVals.array()-diag) / singVals.array()).head(actual_n).transpose() << "\n\n";
+    std::cout << "    check3 (>0)      : " << ((diag.segment(1,actual_n-1)-singVals.head(actual_n-1).array()) / singVals.head(actual_n-1).array()).transpose() << "\n\n\n";
+    std::cout << "    check4 (>0)      : " << ((singVals.segment(1,actual_n-1)-singVals.head(actual_n-1))).transpose() << "\n\n\n";
+  }
+#endif
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(singVals.allFinite());
+  assert(mus.allFinite());
+  assert(shifts.allFinite());
+#endif
+  
+  // Compute zhat
+  perturbCol0(col0, diag, perm, singVals, shifts, mus, zhat);
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "  zhat: " << zhat.transpose() << "\n";
+#endif
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(zhat.allFinite());
+#endif
+  
+  computeSingVecs(zhat, diag, perm, singVals, shifts, mus, U, V);
+  
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "U^T U: " << (U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() << "\n";
+  std::cout << "V^T V: " << (V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() << "\n";
+#endif
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(U.allFinite());
+  assert(V.allFinite());
+  assert((U.transpose() * U - MatrixXr(MatrixXr::Identity(U.cols(),U.cols()))).norm() < 1e-14 * n);
+  assert((V.transpose() * V - MatrixXr(MatrixXr::Identity(V.cols(),V.cols()))).norm() < 1e-14 * n);
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+  
+  // Because of deflation, the singular values might not be completely sorted.
+  // Fortunately, reordering them is a O(n) problem
+  for(Index i=0; i<actual_n-1; ++i)
+  {
+    if(singVals(i)>singVals(i+1))
+    {
+      using std::swap;
+      swap(singVals(i),singVals(i+1));
+      U.col(i).swap(U.col(i+1));
+      if(m_compV) V.col(i).swap(V.col(i+1));
+    }
+  }
+  
+  // Reverse order so that singular values in increased order
+  // Because of deflation, the zeros singular-values are already at the end
+  singVals.head(actual_n).reverseInPlace();
+  U.leftCols(actual_n).rowwise().reverseInPlace();
+  if (m_compV) V.leftCols(actual_n).rowwise().reverseInPlace();
+  
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  JacobiSVD<MatrixXr> jsvd(m_computed.block(firstCol, firstCol, n, n) );
+  std::cout << "  * j:        " << jsvd.singularValues().transpose() << "\n\n";
+  std::cout << "  * sing-val: " << singVals.transpose() << "\n";
+//   std::cout << "  * err:      " << ((jsvd.singularValues()-singVals)>1e-13*singVals.norm()).transpose() << "\n";
+#endif
+}
+
+template <typename MatrixType>
+typename BDCSVD<MatrixType>::RealScalar BDCSVD<MatrixType>::secularEq(RealScalar mu, const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const ArrayRef& diagShifted, RealScalar shift)
+{
+  Index m = perm.size();
+  RealScalar res = Literal(1);
+  for(Index i=0; i<m; ++i)
+  {
+    Index j = perm(i);
+    res += numext::abs2(col0(j)) / ((diagShifted(j) - mu) * (diag(j) + shift + mu));
+  }
+  return res;
+
+}
+
+template <typename MatrixType>
+void BDCSVD<MatrixType>::computeSingVals(const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm,
+                                         VectorType& singVals, ArrayRef shifts, ArrayRef mus)
+{
+  using std::abs;
+  using std::swap;
+
+  Index n = col0.size();
+  Index actual_n = n;
+  while(actual_n>1 && col0(actual_n-1)==Literal(0)) --actual_n;
+
+  for (Index k = 0; k < n; ++k)
+  {
+    if (col0(k) == Literal(0) || actual_n==1)
+    {
+      // if col0(k) == 0, then entry is deflated, so singular value is on diagonal
+      // if actual_n==1, then the deflated problem is already diagonalized
+      singVals(k) = k==0 ? col0(0) : diag(k);
+      mus(k) = Literal(0);
+      shifts(k) = k==0 ? col0(0) : diag(k);
+      continue;
+    } 
+
+    // otherwise, use secular equation to find singular value
+    RealScalar left = diag(k);
+    RealScalar right; // was: = (k != actual_n-1) ? diag(k+1) : (diag(actual_n-1) + col0.matrix().norm());
+    if(k==actual_n-1)
+      right = (diag(actual_n-1) + col0.matrix().norm());
+    else
+    {
+      // Skip deflated singular values
+      Index l = k+1;
+      while(col0(l)==Literal(0)) { ++l; eigen_internal_assert(l<actual_n); }
+      right = diag(l);
+    }
+
+    // first decide whether it's closer to the left end or the right end
+    RealScalar mid = left + (right-left) / Literal(2);
+    RealScalar fMid = secularEq(mid, col0, diag, perm, diag, Literal(0));
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+    std::cout << right-left << "\n";
+    std::cout << "fMid = " << fMid << " " << secularEq(mid-left, col0, diag, perm, diag-left, left) << " " << secularEq(mid-right, col0, diag, perm, diag-right, right)   << "\n";
+    std::cout << "     = " << secularEq(0.1*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.2*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.3*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.4*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.49*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.5*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.51*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.6*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.7*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.8*(left+right), col0, diag, perm, diag, 0)
+              << " "       << secularEq(0.9*(left+right), col0, diag, perm, diag, 0) << "\n";
+#endif
+    RealScalar shift = (k == actual_n-1 || fMid > Literal(0)) ? left : right;
+    
+    // measure everything relative to shift
+    Map<ArrayXr> diagShifted(m_workspace.data()+4*n, n);
+    diagShifted = diag - shift;
+    
+    // initial guess
+    RealScalar muPrev, muCur;
+    if (shift == left)
+    {
+      muPrev = (right - left) * RealScalar(0.1);
+      if (k == actual_n-1) muCur = right - left;
+      else                 muCur = (right - left) * RealScalar(0.5);
+    }
+    else
+    {
+      muPrev = -(right - left) * RealScalar(0.1);
+      muCur = -(right - left) * RealScalar(0.5);
+    }
+
+    RealScalar fPrev = secularEq(muPrev, col0, diag, perm, diagShifted, shift);
+    RealScalar fCur = secularEq(muCur, col0, diag, perm, diagShifted, shift);
+    if (abs(fPrev) < abs(fCur))
+    {
+      swap(fPrev, fCur);
+      swap(muPrev, muCur);
+    }
+
+    // rational interpolation: fit a function of the form a / mu + b through the two previous
+    // iterates and use its zero to compute the next iterate
+    bool useBisection = fPrev*fCur>Literal(0);
+    while (fCur!=Literal(0) && abs(muCur - muPrev) > Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(muCur), abs(muPrev)) && abs(fCur - fPrev)>NumTraits<RealScalar>::epsilon() && !useBisection)
+    {
+      ++m_numIters;
+
+      // Find a and b such that the function f(mu) = a / mu + b matches the current and previous samples.
+      RealScalar a = (fCur - fPrev) / (Literal(1)/muCur - Literal(1)/muPrev);
+      RealScalar b = fCur - a / muCur;
+      // And find mu such that f(mu)==0:
+      RealScalar muZero = -a/b;
+      RealScalar fZero = secularEq(muZero, col0, diag, perm, diagShifted, shift);
+      
+      muPrev = muCur;
+      fPrev = fCur;
+      muCur = muZero;
+      fCur = fZero;
+      
+      
+      if (shift == left  && (muCur < Literal(0) || muCur > right - left)) useBisection = true;
+      if (shift == right && (muCur < -(right - left) || muCur > Literal(0))) useBisection = true;
+      if (abs(fCur)>abs(fPrev)) useBisection = true;
+    }
+
+    // fall back on bisection method if rational interpolation did not work
+    if (useBisection)
+    {
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+      std::cout << "useBisection for k = " << k << ", actual_n = " << actual_n << "\n";
+#endif
+      RealScalar leftShifted, rightShifted;
+      if (shift == left)
+      {
+        leftShifted = (std::numeric_limits<RealScalar>::min)();
+        // I don't understand why the case k==0 would be special there:
+        // if (k == 0) rightShifted = right - left; else 
+        rightShifted = (k==actual_n-1) ? right : ((right - left) * RealScalar(0.6)); // theoretically we can take 0.5, but let's be safe
+      }
+      else
+      {
+        leftShifted = -(right - left) * RealScalar(0.6);
+        rightShifted = -(std::numeric_limits<RealScalar>::min)();
+      }
+      
+      RealScalar fLeft = secularEq(leftShifted, col0, diag, perm, diagShifted, shift);
+
+#if defined EIGEN_INTERNAL_DEBUGGING || defined EIGEN_BDCSVD_DEBUG_VERBOSE
+      RealScalar fRight = secularEq(rightShifted, col0, diag, perm, diagShifted, shift);
+#endif
+
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+      if(!(fLeft * fRight<0))
+      {
+        std::cout << "fLeft: " << leftShifted << " - " << diagShifted.head(10).transpose()  << "\n ; " << bool(left==shift) << " " << (left-shift) << "\n";
+        std::cout << k << " : " <<  fLeft << " * " << fRight << " == " << fLeft * fRight << "  ;  " << left << " - " << right << " -> " <<  leftShifted << " " << rightShifted << "   shift=" << shift << "\n";
+      }
+#endif
+      eigen_internal_assert(fLeft * fRight < Literal(0));
+      
+      while (rightShifted - leftShifted > Literal(2) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(abs(leftShifted), abs(rightShifted)))
+      {
+        RealScalar midShifted = (leftShifted + rightShifted) / Literal(2);
+        fMid = secularEq(midShifted, col0, diag, perm, diagShifted, shift);
+        if (fLeft * fMid < Literal(0))
+        {
+          rightShifted = midShifted;
+        }
+        else
+        {
+          leftShifted = midShifted;
+          fLeft = fMid;
+        }
+      }
+
+      muCur = (leftShifted + rightShifted) / Literal(2);
+    }
+      
+    singVals[k] = shift + muCur;
+    shifts[k] = shift;
+    mus[k] = muCur;
+
+    // perturb singular value slightly if it equals diagonal entry to avoid division by zero later
+    // (deflation is supposed to avoid this from happening)
+    // - this does no seem to be necessary anymore -
+//     if (singVals[k] == left) singVals[k] *= 1 + NumTraits<RealScalar>::epsilon();
+//     if (singVals[k] == right) singVals[k] *= 1 - NumTraits<RealScalar>::epsilon();
+  }
+}
+
+
+// zhat is perturbation of col0 for which singular vectors can be computed stably (see Section 3.1)
+template <typename MatrixType>
+void BDCSVD<MatrixType>::perturbCol0
+   (const ArrayRef& col0, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
+    const ArrayRef& shifts, const ArrayRef& mus, ArrayRef zhat)
+{
+  using std::sqrt;
+  Index n = col0.size();
+  Index m = perm.size();
+  if(m==0)
+  {
+    zhat.setZero();
+    return;
+  }
+  Index last = perm(m-1);
+  // The offset permits to skip deflated entries while computing zhat
+  for (Index k = 0; k < n; ++k)
+  {
+    if (col0(k) == Literal(0)) // deflated
+      zhat(k) = Literal(0);
+    else
+    {
+      // see equation (3.6)
+      RealScalar dk = diag(k);
+      RealScalar prod = (singVals(last) + dk) * (mus(last) + (shifts(last) - dk));
+
+      for(Index l = 0; l<m; ++l)
+      {
+        Index i = perm(l);
+        if(i!=k)
+        {
+          Index j = i<k ? i : perm(l-1);
+          prod *= ((singVals(j)+dk) / ((diag(i)+dk))) * ((mus(j)+(shifts(j)-dk)) / ((diag(i)-dk)));
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+          if(i!=k && std::abs(((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) - 1) > 0.9 )
+            std::cout << "     " << ((singVals(j)+dk)*(mus(j)+(shifts(j)-dk)))/((diag(i)+dk)*(diag(i)-dk)) << " == (" << (singVals(j)+dk) << " * " << (mus(j)+(shifts(j)-dk))
+                       << ") / (" << (diag(i)+dk) << " * " << (diag(i)-dk) << ")\n";
+#endif
+        }
+      }
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+      std::cout << "zhat(" << k << ") =  sqrt( " << prod << ")  ;  " << (singVals(last) + dk) << " * " << mus(last) + shifts(last) << " - " << dk << "\n";
+#endif
+      RealScalar tmp = sqrt(prod);
+      zhat(k) = col0(k) > Literal(0) ? tmp : -tmp;
+    }
+  }
+}
+
+// compute singular vectors
+template <typename MatrixType>
+void BDCSVD<MatrixType>::computeSingVecs
+   (const ArrayRef& zhat, const ArrayRef& diag, const IndicesRef &perm, const VectorType& singVals,
+    const ArrayRef& shifts, const ArrayRef& mus, MatrixXr& U, MatrixXr& V)
+{
+  Index n = zhat.size();
+  Index m = perm.size();
+  
+  for (Index k = 0; k < n; ++k)
+  {
+    if (zhat(k) == Literal(0))
+    {
+      U.col(k) = VectorType::Unit(n+1, k);
+      if (m_compV) V.col(k) = VectorType::Unit(n, k);
+    }
+    else
+    {
+      U.col(k).setZero();
+      for(Index l=0;l<m;++l)
+      {
+        Index i = perm(l);
+        U(i,k) = zhat(i)/(((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
+      }
+      U(n,k) = Literal(0);
+      U.col(k).normalize();
+    
+      if (m_compV)
+      {
+        V.col(k).setZero();
+        for(Index l=1;l<m;++l)
+        {
+          Index i = perm(l);
+          V(i,k) = diag(i) * zhat(i) / (((diag(i) - shifts(k)) - mus(k)) )/( (diag(i) + singVals[k]));
+        }
+        V(0,k) = Literal(-1);
+        V.col(k).normalize();
+      }
+    }
+  }
+  U.col(n) = VectorType::Unit(n+1, n);
+}
+
+
+// page 12_13
+// i >= 1, di almost null and zi non null.
+// We use a rotation to zero out zi applied to the left of M
+template <typename MatrixType>
+void BDCSVD<MatrixType>::deflation43(Index firstCol, Index shift, Index i, Index size)
+{
+  using std::abs;
+  using std::sqrt;
+  using std::pow;
+  Index start = firstCol + shift;
+  RealScalar c = m_computed(start, start);
+  RealScalar s = m_computed(start+i, start);
+  RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
+  if (r == Literal(0))
+  {
+    m_computed(start+i, start+i) = Literal(0);
+    return;
+  }
+  m_computed(start,start) = r;  
+  m_computed(start+i, start) = Literal(0);
+  m_computed(start+i, start+i) = Literal(0);
+  
+  JacobiRotation<RealScalar> J(c/r,-s/r);
+  if (m_compU)  m_naiveU.middleRows(firstCol, size+1).applyOnTheRight(firstCol, firstCol+i, J);
+  else          m_naiveU.applyOnTheRight(firstCol, firstCol+i, J);
+}// end deflation 43
+
+
+// page 13
+// i,j >= 1, i!=j and |di - dj| < epsilon * norm2(M)
+// We apply two rotations to have zj = 0;
+// TODO deflation44 is still broken and not properly tested
+template <typename MatrixType>
+void BDCSVD<MatrixType>::deflation44(Index firstColu , Index firstColm, Index firstRowW, Index firstColW, Index i, Index j, Index size)
+{
+  using std::abs;
+  using std::sqrt;
+  using std::conj;
+  using std::pow;
+  RealScalar c = m_computed(firstColm+i, firstColm);
+  RealScalar s = m_computed(firstColm+j, firstColm);
+  RealScalar r = sqrt(numext::abs2(c) + numext::abs2(s));
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "deflation 4.4: " << i << "," << j << " -> " << c << " " << s << " " << r << " ; "
+    << m_computed(firstColm + i-1, firstColm)  << " "
+    << m_computed(firstColm + i, firstColm)  << " "
+    << m_computed(firstColm + i+1, firstColm) << " "
+    << m_computed(firstColm + i+2, firstColm) << "\n";
+  std::cout << m_computed(firstColm + i-1, firstColm + i-1)  << " "
+    << m_computed(firstColm + i, firstColm+i)  << " "
+    << m_computed(firstColm + i+1, firstColm+i+1) << " "
+    << m_computed(firstColm + i+2, firstColm+i+2) << "\n";
+#endif
+  if (r==Literal(0))
+  {
+    m_computed(firstColm + i, firstColm + i) = m_computed(firstColm + j, firstColm + j);
+    return;
+  }
+  c/=r;
+  s/=r;
+  m_computed(firstColm + i, firstColm) = r;  
+  m_computed(firstColm + j, firstColm + j) = m_computed(firstColm + i, firstColm + i);
+  m_computed(firstColm + j, firstColm) = Literal(0);
+
+  JacobiRotation<RealScalar> J(c,-s);
+  if (m_compU)  m_naiveU.middleRows(firstColu, size+1).applyOnTheRight(firstColu + i, firstColu + j, J);
+  else          m_naiveU.applyOnTheRight(firstColu+i, firstColu+j, J);
+  if (m_compV)  m_naiveV.middleRows(firstRowW, size).applyOnTheRight(firstColW + i, firstColW + j, J);
+}// end deflation 44
+
+
+// acts on block from (firstCol+shift, firstCol+shift) to (lastCol+shift, lastCol+shift) [inclusive]
+template <typename MatrixType>
+void BDCSVD<MatrixType>::deflation(Index firstCol, Index lastCol, Index k, Index firstRowW, Index firstColW, Index shift)
+{
+  using std::sqrt;
+  using std::abs;
+  const Index length = lastCol + 1 - firstCol;
+  
+  Block<MatrixXr,Dynamic,1> col0(m_computed, firstCol+shift, firstCol+shift, length, 1);
+  Diagonal<MatrixXr> fulldiag(m_computed);
+  VectorBlock<Diagonal<MatrixXr>,Dynamic> diag(fulldiag, firstCol+shift, length);
+  
+  const RealScalar considerZero = (std::numeric_limits<RealScalar>::min)();
+  RealScalar maxDiag = diag.tail((std::max)(Index(1),length-1)).cwiseAbs().maxCoeff();
+  RealScalar epsilon_strict = numext::maxi<RealScalar>(considerZero,NumTraits<RealScalar>::epsilon() * maxDiag);
+  RealScalar epsilon_coarse = Literal(8) * NumTraits<RealScalar>::epsilon() * numext::maxi<RealScalar>(col0.cwiseAbs().maxCoeff(), maxDiag);
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE  
+  std::cout << "\ndeflate:" << diag.head(k+1).transpose() << "  |  " << diag.segment(k+1,length-k-1).transpose() << "\n";
+#endif
+  
+  //condition 4.1
+  if (diag(0) < epsilon_coarse)
+  { 
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+    std::cout << "deflation 4.1, because " << diag(0) << " < " << epsilon_coarse << "\n";
+#endif
+    diag(0) = epsilon_coarse;
+  }
+
+  //condition 4.2
+  for (Index i=1;i<length;++i)
+    if (abs(col0(i)) < epsilon_strict)
+    {
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+      std::cout << "deflation 4.2, set z(" << i << ") to zero because " << abs(col0(i)) << " < " << epsilon_strict << "  (diag(" << i << ")=" << diag(i) << ")\n";
+#endif
+      col0(i) = Literal(0);
+    }
+
+  //condition 4.3
+  for (Index i=1;i<length; i++)
+    if (diag(i) < epsilon_coarse)
+    {
+#ifdef  EIGEN_BDCSVD_DEBUG_VERBOSE
+      std::cout << "deflation 4.3, cancel z(" << i << ")=" << col0(i) << " because diag(" << i << ")=" << diag(i) << " < " << epsilon_coarse << "\n";
+#endif
+      deflation43(firstCol, shift, i, length);
+    }
+
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "to be sorted: " << diag.transpose() << "\n\n";
+#endif
+  {
+    // Check for total deflation
+    // If we have a total deflation, then we have to consider col0(0)==diag(0) as a singular value during sorting
+    bool total_deflation = (col0.tail(length-1).array()<considerZero).all();
+    
+    // Sort the diagonal entries, since diag(1:k-1) and diag(k:length) are already sorted, let's do a sorted merge.
+    // First, compute the respective permutation.
+    Index *permutation = m_workspaceI.data();
+    {
+      permutation[0] = 0;
+      Index p = 1;
+      
+      // Move deflated diagonal entries at the end.
+      for(Index i=1; i<length; ++i)
+        if(abs(diag(i))<considerZero)
+          permutation[p++] = i;
+        
+      Index i=1, j=k+1;
+      for( ; p < length; ++p)
+      {
+             if (i > k)             permutation[p] = j++;
+        else if (j >= length)       permutation[p] = i++;
+        else if (diag(i) < diag(j)) permutation[p] = j++;
+        else                        permutation[p] = i++;
+      }
+    }
+    
+    // If we have a total deflation, then we have to insert diag(0) at the right place
+    if(total_deflation)
+    {
+      for(Index i=1; i<length; ++i)
+      {
+        Index pi = permutation[i];
+        if(abs(diag(pi))<considerZero || diag(0)<diag(pi))
+          permutation[i-1] = permutation[i];
+        else
+        {
+          permutation[i-1] = 0;
+          break;
+        }
+      }
+    }
+    
+    // Current index of each col, and current column of each index
+    Index *realInd = m_workspaceI.data()+length;
+    Index *realCol = m_workspaceI.data()+2*length;
+    
+    for(int pos = 0; pos< length; pos++)
+    {
+      realCol[pos] = pos;
+      realInd[pos] = pos;
+    }
+    
+    for(Index i = total_deflation?0:1; i < length; i++)
+    {
+      const Index pi = permutation[length - (total_deflation ? i+1 : i)];
+      const Index J = realCol[pi];
+      
+      using std::swap;
+      // swap diagonal and first column entries:
+      swap(diag(i), diag(J));
+      if(i!=0 && J!=0) swap(col0(i), col0(J));
+
+      // change columns
+      if (m_compU) m_naiveU.col(firstCol+i).segment(firstCol, length + 1).swap(m_naiveU.col(firstCol+J).segment(firstCol, length + 1));
+      else         m_naiveU.col(firstCol+i).segment(0, 2)                .swap(m_naiveU.col(firstCol+J).segment(0, 2));
+      if (m_compV) m_naiveV.col(firstColW + i).segment(firstRowW, length).swap(m_naiveV.col(firstColW + J).segment(firstRowW, length));
+
+      //update real pos
+      const Index realI = realInd[i];
+      realCol[realI] = J;
+      realCol[pi] = i;
+      realInd[J] = realI;
+      realInd[i] = pi;
+    }
+  }
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+  std::cout << "sorted: " << diag.transpose().format(bdcsvdfmt) << "\n";
+  std::cout << "      : " << col0.transpose() << "\n\n";
+#endif
+    
+  //condition 4.4
+  {
+    Index i = length-1;
+    while(i>0 && (abs(diag(i))<considerZero || abs(col0(i))<considerZero)) --i;
+    for(; i>1;--i)
+       if( (diag(i) - diag(i-1)) < NumTraits<RealScalar>::epsilon()*maxDiag )
+      {
+#ifdef EIGEN_BDCSVD_DEBUG_VERBOSE
+        std::cout << "deflation 4.4 with i = " << i << " because " << (diag(i) - diag(i-1)) << " < " << NumTraits<RealScalar>::epsilon()*diag(i) << "\n";
+#endif
+        eigen_internal_assert(abs(diag(i) - diag(i-1))<epsilon_coarse && " diagonal entries are not properly sorted");
+        deflation44(firstCol, firstCol + shift, firstRowW, firstColW, i-1, i, length);
+      }
+  }
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  for(Index j=2;j<length;++j)
+    assert(diag(j-1)<=diag(j) || abs(diag(j))<considerZero);
+#endif
+  
+#ifdef EIGEN_BDCSVD_SANITY_CHECKS
+  assert(m_naiveU.allFinite());
+  assert(m_naiveV.allFinite());
+  assert(m_computed.allFinite());
+#endif
+}//end deflation
+
+#ifndef __CUDACC__
+/** \svd_module
+  *
+  * \return the singular value decomposition of \c *this computed by Divide & Conquer algorithm
+  *
+  * \sa class BDCSVD
+  */
+template<typename Derived>
+BDCSVD<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::bdcSvd(unsigned int computationOptions) const
+{
+  return BDCSVD<PlainObject>(*this, computationOptions);
+}
+#endif
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD.h
new file mode 100644
index 0000000..43488b1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD.h
@@ -0,0 +1,804 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2013-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_JACOBISVD_H
+#define EIGEN_JACOBISVD_H
+
+namespace Eigen { 
+
+namespace internal {
+// forward declaration (needed by ICC)
+// the empty body is required by MSVC
+template<typename MatrixType, int QRPreconditioner,
+         bool IsComplex = NumTraits<typename MatrixType::Scalar>::IsComplex>
+struct svd_precondition_2x2_block_to_be_real {};
+
+/*** QR preconditioners (R-SVD)
+ ***
+ *** Their role is to reduce the problem of computing the SVD to the case of a square matrix.
+ *** This approach, known as R-SVD, is an optimization for rectangular-enough matrices, and is a requirement for
+ *** JacobiSVD which by itself is only able to work on square matrices.
+ ***/
+
+enum { PreconditionIfMoreColsThanRows, PreconditionIfMoreRowsThanCols };
+
+template<typename MatrixType, int QRPreconditioner, int Case>
+struct qr_preconditioner_should_do_anything
+{
+  enum { a = MatrixType::RowsAtCompileTime != Dynamic &&
+             MatrixType::ColsAtCompileTime != Dynamic &&
+             MatrixType::ColsAtCompileTime <= MatrixType::RowsAtCompileTime,
+         b = MatrixType::RowsAtCompileTime != Dynamic &&
+             MatrixType::ColsAtCompileTime != Dynamic &&
+             MatrixType::RowsAtCompileTime <= MatrixType::ColsAtCompileTime,
+         ret = !( (QRPreconditioner == NoQRPreconditioner) ||
+                  (Case == PreconditionIfMoreColsThanRows && bool(a)) ||
+                  (Case == PreconditionIfMoreRowsThanCols && bool(b)) )
+  };
+};
+
+template<typename MatrixType, int QRPreconditioner, int Case,
+         bool DoAnything = qr_preconditioner_should_do_anything<MatrixType, QRPreconditioner, Case>::ret
+> struct qr_preconditioner_impl {};
+
+template<typename MatrixType, int QRPreconditioner, int Case>
+class qr_preconditioner_impl<MatrixType, QRPreconditioner, Case, false>
+{
+public:
+  void allocate(const JacobiSVD<MatrixType, QRPreconditioner>&) {}
+  bool run(JacobiSVD<MatrixType, QRPreconditioner>&, const MatrixType&)
+  {
+    return false;
+  }
+};
+
+/*** preconditioner using FullPivHouseholderQR ***/
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
+{
+public:
+  typedef typename MatrixType::Scalar Scalar;
+  enum
+  {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime
+  };
+  typedef Matrix<Scalar, 1, RowsAtCompileTime, RowMajor, 1, MaxRowsAtCompileTime> WorkspaceType;
+
+  void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd)
+  {
+    if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.rows(), svd.cols());
+    }
+    if (svd.m_computeFullU) m_workspace.resize(svd.rows());
+  }
+
+  bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.rows() > matrix.cols())
+    {
+      m_qr.compute(matrix);
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
+      if(svd.m_computeFullU) m_qr.matrixQ().evalTo(svd.m_matrixU, m_workspace);
+      if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation();
+      return true;
+    }
+    return false;
+  }
+private:
+  typedef FullPivHouseholderQR<MatrixType> QRType;
+  QRType m_qr;
+  WorkspaceType m_workspace;
+};
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, FullPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
+{
+public:
+  typedef typename MatrixType::Scalar Scalar;
+  enum
+  {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    TrOptions = RowsAtCompileTime==1 ? (MatrixType::Options & ~(RowMajor))
+              : ColsAtCompileTime==1 ? (MatrixType::Options |   RowMajor)
+              : MatrixType::Options
+  };
+  typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, TrOptions, MaxColsAtCompileTime, MaxRowsAtCompileTime>
+          TransposeTypeWithSameStorageOrder;
+
+  void allocate(const JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd)
+  {
+    if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.cols(), svd.rows());
+    }
+    m_adjoint.resize(svd.cols(), svd.rows());
+    if (svd.m_computeFullV) m_workspace.resize(svd.cols());
+  }
+
+  bool run(JacobiSVD<MatrixType, FullPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.cols() > matrix.rows())
+    {
+      m_adjoint = matrix.adjoint();
+      m_qr.compute(m_adjoint);
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
+      if(svd.m_computeFullV) m_qr.matrixQ().evalTo(svd.m_matrixV, m_workspace);
+      if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation();
+      return true;
+    }
+    else return false;
+  }
+private:
+  typedef FullPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
+  QRType m_qr;
+  TransposeTypeWithSameStorageOrder m_adjoint;
+  typename internal::plain_row_type<MatrixType>::type m_workspace;
+};
+
+/*** preconditioner using ColPivHouseholderQR ***/
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
+{
+public:
+  void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd)
+  {
+    if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.rows(), svd.cols());
+    }
+    if (svd.m_computeFullU) m_workspace.resize(svd.rows());
+    else if (svd.m_computeThinU) m_workspace.resize(svd.cols());
+  }
+
+  bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.rows() > matrix.cols())
+    {
+      m_qr.compute(matrix);
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
+      if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace);
+      else if(svd.m_computeThinU)
+      {
+        svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols());
+        m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace);
+      }
+      if(svd.computeV()) svd.m_matrixV = m_qr.colsPermutation();
+      return true;
+    }
+    return false;
+  }
+
+private:
+  typedef ColPivHouseholderQR<MatrixType> QRType;
+  QRType m_qr;
+  typename internal::plain_col_type<MatrixType>::type m_workspace;
+};
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, ColPivHouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
+{
+public:
+  typedef typename MatrixType::Scalar Scalar;
+  enum
+  {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    TrOptions = RowsAtCompileTime==1 ? (MatrixType::Options & ~(RowMajor))
+              : ColsAtCompileTime==1 ? (MatrixType::Options |   RowMajor)
+              : MatrixType::Options
+  };
+
+  typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, TrOptions, MaxColsAtCompileTime, MaxRowsAtCompileTime>
+          TransposeTypeWithSameStorageOrder;
+
+  void allocate(const JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd)
+  {
+    if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.cols(), svd.rows());
+    }
+    if (svd.m_computeFullV) m_workspace.resize(svd.cols());
+    else if (svd.m_computeThinV) m_workspace.resize(svd.rows());
+    m_adjoint.resize(svd.cols(), svd.rows());
+  }
+
+  bool run(JacobiSVD<MatrixType, ColPivHouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.cols() > matrix.rows())
+    {
+      m_adjoint = matrix.adjoint();
+      m_qr.compute(m_adjoint);
+
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
+      if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace);
+      else if(svd.m_computeThinV)
+      {
+        svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows());
+        m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace);
+      }
+      if(svd.computeU()) svd.m_matrixU = m_qr.colsPermutation();
+      return true;
+    }
+    else return false;
+  }
+
+private:
+  typedef ColPivHouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
+  QRType m_qr;
+  TransposeTypeWithSameStorageOrder m_adjoint;
+  typename internal::plain_row_type<MatrixType>::type m_workspace;
+};
+
+/*** preconditioner using HouseholderQR ***/
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreRowsThanCols, true>
+{
+public:
+  void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd)
+  {
+    if (svd.rows() != m_qr.rows() || svd.cols() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.rows(), svd.cols());
+    }
+    if (svd.m_computeFullU) m_workspace.resize(svd.rows());
+    else if (svd.m_computeThinU) m_workspace.resize(svd.cols());
+  }
+
+  bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.rows() > matrix.cols())
+    {
+      m_qr.compute(matrix);
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.cols(),matrix.cols()).template triangularView<Upper>();
+      if(svd.m_computeFullU) m_qr.householderQ().evalTo(svd.m_matrixU, m_workspace);
+      else if(svd.m_computeThinU)
+      {
+        svd.m_matrixU.setIdentity(matrix.rows(), matrix.cols());
+        m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixU, m_workspace);
+      }
+      if(svd.computeV()) svd.m_matrixV.setIdentity(matrix.cols(), matrix.cols());
+      return true;
+    }
+    return false;
+  }
+private:
+  typedef HouseholderQR<MatrixType> QRType;
+  QRType m_qr;
+  typename internal::plain_col_type<MatrixType>::type m_workspace;
+};
+
+template<typename MatrixType>
+class qr_preconditioner_impl<MatrixType, HouseholderQRPreconditioner, PreconditionIfMoreColsThanRows, true>
+{
+public:
+  typedef typename MatrixType::Scalar Scalar;
+  enum
+  {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    Options = MatrixType::Options
+  };
+
+  typedef Matrix<Scalar, ColsAtCompileTime, RowsAtCompileTime, Options, MaxColsAtCompileTime, MaxRowsAtCompileTime>
+          TransposeTypeWithSameStorageOrder;
+
+  void allocate(const JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd)
+  {
+    if (svd.cols() != m_qr.rows() || svd.rows() != m_qr.cols())
+    {
+      m_qr.~QRType();
+      ::new (&m_qr) QRType(svd.cols(), svd.rows());
+    }
+    if (svd.m_computeFullV) m_workspace.resize(svd.cols());
+    else if (svd.m_computeThinV) m_workspace.resize(svd.rows());
+    m_adjoint.resize(svd.cols(), svd.rows());
+  }
+
+  bool run(JacobiSVD<MatrixType, HouseholderQRPreconditioner>& svd, const MatrixType& matrix)
+  {
+    if(matrix.cols() > matrix.rows())
+    {
+      m_adjoint = matrix.adjoint();
+      m_qr.compute(m_adjoint);
+
+      svd.m_workMatrix = m_qr.matrixQR().block(0,0,matrix.rows(),matrix.rows()).template triangularView<Upper>().adjoint();
+      if(svd.m_computeFullV) m_qr.householderQ().evalTo(svd.m_matrixV, m_workspace);
+      else if(svd.m_computeThinV)
+      {
+        svd.m_matrixV.setIdentity(matrix.cols(), matrix.rows());
+        m_qr.householderQ().applyThisOnTheLeft(svd.m_matrixV, m_workspace);
+      }
+      if(svd.computeU()) svd.m_matrixU.setIdentity(matrix.rows(), matrix.rows());
+      return true;
+    }
+    else return false;
+  }
+
+private:
+  typedef HouseholderQR<TransposeTypeWithSameStorageOrder> QRType;
+  QRType m_qr;
+  TransposeTypeWithSameStorageOrder m_adjoint;
+  typename internal::plain_row_type<MatrixType>::type m_workspace;
+};
+
+/*** 2x2 SVD implementation
+ ***
+ *** JacobiSVD consists in performing a series of 2x2 SVD subproblems
+ ***/
+
+template<typename MatrixType, int QRPreconditioner>
+struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, false>
+{
+  typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
+  typedef typename MatrixType::RealScalar RealScalar;
+  static bool run(typename SVD::WorkMatrixType&, SVD&, Index, Index, RealScalar&) { return true; }
+};
+
+template<typename MatrixType, int QRPreconditioner>
+struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
+{
+  typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  static bool run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q, RealScalar& maxDiagEntry)
+  {
+    using std::sqrt;
+    using std::abs;
+    Scalar z;
+    JacobiRotation<Scalar> rot;
+    RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p)));
+
+    const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+    const RealScalar precision = NumTraits<Scalar>::epsilon();
+
+    if(n==0)
+    {
+      // make sure first column is zero
+      work_matrix.coeffRef(p,p) = work_matrix.coeffRef(q,p) = Scalar(0);
+
+      if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
+      {
+        // work_matrix.coeff(p,q) can be zero if work_matrix.coeff(q,p) is not zero but small enough to underflow when computing n
+        z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
+        work_matrix.row(p) *= z;
+        if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z);
+      }
+      if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
+      {
+        z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
+        work_matrix.row(q) *= z;
+        if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
+      }
+      // otherwise the second row is already zero, so we have nothing to do.
+    }
+    else
+    {
+      rot.c() = conj(work_matrix.coeff(p,p)) / n;
+      rot.s() = work_matrix.coeff(q,p) / n;
+      work_matrix.applyOnTheLeft(p,q,rot);
+      if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint());
+      if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
+      {
+        z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
+        work_matrix.col(q) *= z;
+        if(svd.computeV()) svd.m_matrixV.col(q) *= z;
+      }
+      if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
+      {
+        z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
+        work_matrix.row(q) *= z;
+        if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
+      }
+    }
+
+    // update largest diagonal entry
+    maxDiagEntry = numext::maxi<RealScalar>(maxDiagEntry,numext::maxi<RealScalar>(abs(work_matrix.coeff(p,p)), abs(work_matrix.coeff(q,q))));
+    // and check whether the 2x2 block is already diagonal
+    RealScalar threshold = numext::maxi<RealScalar>(considerAsZero, precision * maxDiagEntry);
+    return abs(work_matrix.coeff(p,q))>threshold || abs(work_matrix.coeff(q,p)) > threshold;
+  }
+};
+
+template<typename _MatrixType, int QRPreconditioner> 
+struct traits<JacobiSVD<_MatrixType,QRPreconditioner> >
+{
+  typedef _MatrixType MatrixType;
+};
+
+} // end namespace internal
+
+/** \ingroup SVD_Module
+  *
+  *
+  * \class JacobiSVD
+  *
+  * \brief Two-sided Jacobi SVD decomposition of a rectangular matrix
+  *
+  * \tparam _MatrixType the type of the matrix of which we are computing the SVD decomposition
+  * \tparam QRPreconditioner this optional parameter allows to specify the type of QR decomposition that will be used internally
+  *                        for the R-SVD step for non-square matrices. See discussion of possible values below.
+  *
+  * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product
+  *   \f[ A = U S V^* \f]
+  * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal;
+  * the diagonal entries of S are known as the \em singular \em values of \a A and the columns of \a U and \a V are known as the left
+  * and right \em singular \em vectors of \a A respectively.
+  *
+  * Singular values are always sorted in decreasing order.
+  *
+  * This JacobiSVD decomposition computes only the singular values by default. If you want \a U or \a V, you need to ask for them explicitly.
+  *
+  * You can ask for only \em thin \a U or \a V to be computed, meaning the following. In case of a rectangular n-by-p matrix, letting \a m be the
+  * smaller value among \a n and \a p, there are only \a m singular vectors; the remaining columns of \a U and \a V do not correspond to actual
+  * singular vectors. Asking for \em thin \a U or \a V means asking for only their \a m first columns to be formed. So \a U is then a n-by-m matrix,
+  * and \a V is then a p-by-m matrix. Notice that thin \a U and \a V are all you need for (least squares) solving.
+  *
+  * Here's an example demonstrating basic usage:
+  * \include JacobiSVD_basic.cpp
+  * Output: \verbinclude JacobiSVD_basic.out
+  *
+  * This JacobiSVD class is a two-sided Jacobi R-SVD decomposition, ensuring optimal reliability and accuracy. The downside is that it's slower than
+  * bidiagonalizing SVD algorithms for large square matrices; however its complexity is still \f$ O(n^2p) \f$ where \a n is the smaller dimension and
+  * \a p is the greater dimension, meaning that it is still of the same order of complexity as the faster bidiagonalizing R-SVD algorithms.
+  * In particular, like any R-SVD, it takes advantage of non-squareness in that its complexity is only linear in the greater dimension.
+  *
+  * If the input matrix has inf or nan coefficients, the result of the computation is undefined, but the computation is guaranteed to
+  * terminate in finite (and reasonable) time.
+  *
+  * The possible values for QRPreconditioner are:
+  * \li ColPivHouseholderQRPreconditioner is the default. In practice it's very safe. It uses column-pivoting QR.
+  * \li FullPivHouseholderQRPreconditioner, is the safest and slowest. It uses full-pivoting QR.
+  *     Contrary to other QRs, it doesn't allow computing thin unitaries.
+  * \li HouseholderQRPreconditioner is the fastest, and less safe and accurate than the pivoting variants. It uses non-pivoting QR.
+  *     This is very similar in safety and accuracy to the bidiagonalization process used by bidiagonalizing SVD algorithms (since bidiagonalization
+  *     is inherently non-pivoting). However the resulting SVD is still more reliable than bidiagonalizing SVDs because the Jacobi-based iterarive
+  *     process is more reliable than the optimized bidiagonal SVD iterations.
+  * \li NoQRPreconditioner allows not to use a QR preconditioner at all. This is useful if you know that you will only be computing
+  *     JacobiSVD decompositions of square matrices. Non-square matrices require a QR preconditioner. Using this option will result in
+  *     faster compilation and smaller executable code. It won't significantly speed up computation, since JacobiSVD is always checking
+  *     if QR preconditioning is needed before applying it anyway.
+  *
+  * \sa MatrixBase::jacobiSvd()
+  */
+template<typename _MatrixType, int QRPreconditioner> class JacobiSVD
+ : public SVDBase<JacobiSVD<_MatrixType,QRPreconditioner> >
+{
+    typedef SVDBase<JacobiSVD> Base;
+  public:
+
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime),
+      MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+      MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime,MaxColsAtCompileTime),
+      MatrixOptions = MatrixType::Options
+    };
+
+    typedef typename Base::MatrixUType MatrixUType;
+    typedef typename Base::MatrixVType MatrixVType;
+    typedef typename Base::SingularValuesType SingularValuesType;
+    
+    typedef typename internal::plain_row_type<MatrixType>::type RowType;
+    typedef typename internal::plain_col_type<MatrixType>::type ColType;
+    typedef Matrix<Scalar, DiagSizeAtCompileTime, DiagSizeAtCompileTime,
+                   MatrixOptions, MaxDiagSizeAtCompileTime, MaxDiagSizeAtCompileTime>
+            WorkMatrixType;
+
+    /** \brief Default Constructor.
+      *
+      * The default constructor is useful in cases in which the user intends to
+      * perform decompositions via JacobiSVD::compute(const MatrixType&).
+      */
+    JacobiSVD()
+    {}
+
+
+    /** \brief Default Constructor with memory preallocation
+      *
+      * Like the default constructor but with preallocation of the internal data
+      * according to the specified problem size.
+      * \sa JacobiSVD()
+      */
+    JacobiSVD(Index rows, Index cols, unsigned int computationOptions = 0)
+    {
+      allocate(rows, cols, computationOptions);
+    }
+
+    /** \brief Constructor performing the decomposition of given matrix.
+     *
+     * \param matrix the matrix to decompose
+     * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
+     *                           By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU,
+     *                           #ComputeFullV, #ComputeThinV.
+     *
+     * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
+     * available with the (non-default) FullPivHouseholderQR preconditioner.
+     */
+    explicit JacobiSVD(const MatrixType& matrix, unsigned int computationOptions = 0)
+    {
+      compute(matrix, computationOptions);
+    }
+
+    /** \brief Method performing the decomposition of given matrix using custom options.
+     *
+     * \param matrix the matrix to decompose
+     * \param computationOptions optional parameter allowing to specify if you want full or thin U or V unitaries to be computed.
+     *                           By default, none is computed. This is a bit-field, the possible bits are #ComputeFullU, #ComputeThinU,
+     *                           #ComputeFullV, #ComputeThinV.
+     *
+     * Thin unitaries are only available if your matrix type has a Dynamic number of columns (for example MatrixXf). They also are not
+     * available with the (non-default) FullPivHouseholderQR preconditioner.
+     */
+    JacobiSVD& compute(const MatrixType& matrix, unsigned int computationOptions);
+
+    /** \brief Method performing the decomposition of given matrix using current options.
+     *
+     * \param matrix the matrix to decompose
+     *
+     * This method uses the current \a computationOptions, as already passed to the constructor or to compute(const MatrixType&, unsigned int).
+     */
+    JacobiSVD& compute(const MatrixType& matrix)
+    {
+      return compute(matrix, m_computationOptions);
+    }
+
+    using Base::computeU;
+    using Base::computeV;
+    using Base::rows;
+    using Base::cols;
+    using Base::rank;
+
+  private:
+    void allocate(Index rows, Index cols, unsigned int computationOptions);
+
+  protected:
+    using Base::m_matrixU;
+    using Base::m_matrixV;
+    using Base::m_singularValues;
+    using Base::m_isInitialized;
+    using Base::m_isAllocated;
+    using Base::m_usePrescribedThreshold;
+    using Base::m_computeFullU;
+    using Base::m_computeThinU;
+    using Base::m_computeFullV;
+    using Base::m_computeThinV;
+    using Base::m_computationOptions;
+    using Base::m_nonzeroSingularValues;
+    using Base::m_rows;
+    using Base::m_cols;
+    using Base::m_diagSize;
+    using Base::m_prescribedThreshold;
+    WorkMatrixType m_workMatrix;
+
+    template<typename __MatrixType, int _QRPreconditioner, bool _IsComplex>
+    friend struct internal::svd_precondition_2x2_block_to_be_real;
+    template<typename __MatrixType, int _QRPreconditioner, int _Case, bool _DoAnything>
+    friend struct internal::qr_preconditioner_impl;
+
+    internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreColsThanRows> m_qr_precond_morecols;
+    internal::qr_preconditioner_impl<MatrixType, QRPreconditioner, internal::PreconditionIfMoreRowsThanCols> m_qr_precond_morerows;
+    MatrixType m_scaledMatrix;
+};
+
+template<typename MatrixType, int QRPreconditioner>
+void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Index rows, Index cols, unsigned int computationOptions)
+{
+  eigen_assert(rows >= 0 && cols >= 0);
+
+  if (m_isAllocated &&
+      rows == m_rows &&
+      cols == m_cols &&
+      computationOptions == m_computationOptions)
+  {
+    return;
+  }
+
+  m_rows = rows;
+  m_cols = cols;
+  m_isInitialized = false;
+  m_isAllocated = true;
+  m_computationOptions = computationOptions;
+  m_computeFullU = (computationOptions & ComputeFullU) != 0;
+  m_computeThinU = (computationOptions & ComputeThinU) != 0;
+  m_computeFullV = (computationOptions & ComputeFullV) != 0;
+  m_computeThinV = (computationOptions & ComputeThinV) != 0;
+  eigen_assert(!(m_computeFullU && m_computeThinU) && "JacobiSVD: you can't ask for both full and thin U");
+  eigen_assert(!(m_computeFullV && m_computeThinV) && "JacobiSVD: you can't ask for both full and thin V");
+  eigen_assert(EIGEN_IMPLIES(m_computeThinU || m_computeThinV, MatrixType::ColsAtCompileTime==Dynamic) &&
+              "JacobiSVD: thin U and V are only available when your matrix has a dynamic number of columns.");
+  if (QRPreconditioner == FullPivHouseholderQRPreconditioner)
+  {
+      eigen_assert(!(m_computeThinU || m_computeThinV) &&
+              "JacobiSVD: can't compute thin U or thin V with the FullPivHouseholderQR preconditioner. "
+              "Use the ColPivHouseholderQR preconditioner instead.");
+  }
+  m_diagSize = (std::min)(m_rows, m_cols);
+  m_singularValues.resize(m_diagSize);
+  if(RowsAtCompileTime==Dynamic)
+    m_matrixU.resize(m_rows, m_computeFullU ? m_rows
+                            : m_computeThinU ? m_diagSize
+                            : 0);
+  if(ColsAtCompileTime==Dynamic)
+    m_matrixV.resize(m_cols, m_computeFullV ? m_cols
+                            : m_computeThinV ? m_diagSize
+                            : 0);
+  m_workMatrix.resize(m_diagSize, m_diagSize);
+  
+  if(m_cols>m_rows)   m_qr_precond_morecols.allocate(*this);
+  if(m_rows>m_cols)   m_qr_precond_morerows.allocate(*this);
+  if(m_rows!=m_cols)  m_scaledMatrix.resize(rows,cols);
+}
+
+template<typename MatrixType, int QRPreconditioner>
+JacobiSVD<MatrixType, QRPreconditioner>&
+JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsigned int computationOptions)
+{
+  using std::abs;
+  allocate(matrix.rows(), matrix.cols(), computationOptions);
+
+  // currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations,
+  // only worsening the precision of U and V as we accumulate more rotations
+  const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();
+
+  // limit for denormal numbers to be considered zero in order to avoid infinite loops (see bug 286)
+  const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+
+  // Scaling factor to reduce over/under-flows
+  RealScalar scale = matrix.cwiseAbs().maxCoeff();
+  if(scale==RealScalar(0)) scale = RealScalar(1);
+  
+  /*** step 1. The R-SVD step: we use a QR decomposition to reduce to the case of a square matrix */
+
+  if(m_rows!=m_cols)
+  {
+    m_scaledMatrix = matrix / scale;
+    m_qr_precond_morecols.run(*this, m_scaledMatrix);
+    m_qr_precond_morerows.run(*this, m_scaledMatrix);
+  }
+  else
+  {
+    m_workMatrix = matrix.block(0,0,m_diagSize,m_diagSize) / scale;
+    if(m_computeFullU) m_matrixU.setIdentity(m_rows,m_rows);
+    if(m_computeThinU) m_matrixU.setIdentity(m_rows,m_diagSize);
+    if(m_computeFullV) m_matrixV.setIdentity(m_cols,m_cols);
+    if(m_computeThinV) m_matrixV.setIdentity(m_cols, m_diagSize);
+  }
+
+  /*** step 2. The main Jacobi SVD iteration. ***/
+  RealScalar maxDiagEntry = m_workMatrix.cwiseAbs().diagonal().maxCoeff();
+
+  bool finished = false;
+  while(!finished)
+  {
+    finished = true;
+
+    // do a sweep: for all index pairs (p,q), perform SVD of the corresponding 2x2 sub-matrix
+
+    for(Index p = 1; p < m_diagSize; ++p)
+    {
+      for(Index q = 0; q < p; ++q)
+      {
+        // if this 2x2 sub-matrix is not diagonal already...
+        // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't
+        // keep us iterating forever. Similarly, small denormal numbers are considered zero.
+        RealScalar threshold = numext::maxi<RealScalar>(considerAsZero, precision * maxDiagEntry);
+        if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold)
+        {
+          finished = false;
+          // perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal
+          // the complex to real operation returns true if the updated 2x2 block is not already diagonal
+          if(internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q, maxDiagEntry))
+          {
+            JacobiRotation<RealScalar> j_left, j_right;
+            internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right);
+
+            // accumulate resulting Jacobi rotations
+            m_workMatrix.applyOnTheLeft(p,q,j_left);
+            if(computeU()) m_matrixU.applyOnTheRight(p,q,j_left.transpose());
+
+            m_workMatrix.applyOnTheRight(p,q,j_right);
+            if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right);
+
+            // keep track of the largest diagonal coefficient
+            maxDiagEntry = numext::maxi<RealScalar>(maxDiagEntry,numext::maxi<RealScalar>(abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q))));
+          }
+        }
+      }
+    }
+  }
+
+  /*** step 3. The work matrix is now diagonal, so ensure it's positive so its diagonal entries are the singular values ***/
+
+  for(Index i = 0; i < m_diagSize; ++i)
+  {
+    // For a complex matrix, some diagonal coefficients might note have been
+    // treated by svd_precondition_2x2_block_to_be_real, and the imaginary part
+    // of some diagonal entry might not be null.
+    if(NumTraits<Scalar>::IsComplex && abs(numext::imag(m_workMatrix.coeff(i,i)))>considerAsZero)
+    {
+      RealScalar a = abs(m_workMatrix.coeff(i,i));
+      m_singularValues.coeffRef(i) = abs(a);
+      if(computeU()) m_matrixU.col(i) *= m_workMatrix.coeff(i,i)/a;
+    }
+    else
+    {
+      // m_workMatrix.coeff(i,i) is already real, no difficulty:
+      RealScalar a = numext::real(m_workMatrix.coeff(i,i));
+      m_singularValues.coeffRef(i) = abs(a);
+      if(computeU() && (a<RealScalar(0))) m_matrixU.col(i) = -m_matrixU.col(i);
+    }
+  }
+  
+  m_singularValues *= scale;
+
+  /*** step 4. Sort singular values in descending order and compute the number of nonzero singular values ***/
+
+  m_nonzeroSingularValues = m_diagSize;
+  for(Index i = 0; i < m_diagSize; i++)
+  {
+    Index pos;
+    RealScalar maxRemainingSingularValue = m_singularValues.tail(m_diagSize-i).maxCoeff(&pos);
+    if(maxRemainingSingularValue == RealScalar(0))
+    {
+      m_nonzeroSingularValues = i;
+      break;
+    }
+    if(pos)
+    {
+      pos += i;
+      std::swap(m_singularValues.coeffRef(i), m_singularValues.coeffRef(pos));
+      if(computeU()) m_matrixU.col(pos).swap(m_matrixU.col(i));
+      if(computeV()) m_matrixV.col(pos).swap(m_matrixV.col(i));
+    }
+  }
+
+  m_isInitialized = true;
+  return *this;
+}
+
+/** \svd_module
+  *
+  * \return the singular value decomposition of \c *this computed by two-sided
+  * Jacobi transformations.
+  *
+  * \sa class JacobiSVD
+  */
+template<typename Derived>
+JacobiSVD<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::jacobiSvd(unsigned int computationOptions) const
+{
+  return JacobiSVD<PlainObject>(*this, computationOptions);
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_JACOBISVD_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h
new file mode 100644
index 0000000..5027215
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/JacobiSVD_LAPACKE.h
@@ -0,0 +1,90 @@
+/*
+ Copyright (c) 2011, Intel Corporation. All rights reserved.
+
+ Redistribution and use in source and binary forms, with or without modification,
+ are permitted provided that the following conditions are met:
+
+ * Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+ * Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+ * Neither the name of Intel Corporation nor the names of its contributors may
+   be used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
+ ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+ ********************************************************************************
+ *   Content : Eigen bindings to LAPACKe
+ *    Singular Value Decomposition - SVD.
+ ********************************************************************************
+*/
+
+#ifndef EIGEN_JACOBISVD_LAPACKE_H
+#define EIGEN_JACOBISVD_LAPACKE_H
+
+namespace Eigen { 
+
+/** \internal Specialization for the data types supported by LAPACKe */
+
+#define EIGEN_LAPACKE_SVD(EIGTYPE, LAPACKE_TYPE, LAPACKE_RTYPE, LAPACKE_PREFIX, EIGCOLROW, LAPACKE_COLROW) \
+template<> inline \
+JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPivHouseholderQRPreconditioner>& \
+JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPivHouseholderQRPreconditioner>::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>& matrix, unsigned int computationOptions) \
+{ \
+  typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> MatrixType; \
+  /*typedef MatrixType::Scalar Scalar;*/ \
+  /*typedef MatrixType::RealScalar RealScalar;*/ \
+  allocate(matrix.rows(), matrix.cols(), computationOptions); \
+\
+  /*const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();*/ \
+  m_nonzeroSingularValues = m_diagSize; \
+\
+  lapack_int lda = internal::convert_index<lapack_int>(matrix.outerStride()), ldu, ldvt; \
+  lapack_int matrix_order = LAPACKE_COLROW; \
+  char jobu, jobvt; \
+  LAPACKE_TYPE *u, *vt, dummy; \
+  jobu  = (m_computeFullU) ? 'A' : (m_computeThinU) ? 'S' : 'N'; \
+  jobvt = (m_computeFullV) ? 'A' : (m_computeThinV) ? 'S' : 'N'; \
+  if (computeU()) { \
+    ldu  = internal::convert_index<lapack_int>(m_matrixU.outerStride()); \
+    u    = (LAPACKE_TYPE*)m_matrixU.data(); \
+  } else { ldu=1; u=&dummy; }\
+  MatrixType localV; \
+  ldvt = (m_computeFullV) ? internal::convert_index<lapack_int>(m_cols) : (m_computeThinV) ? internal::convert_index<lapack_int>(m_diagSize) : 1; \
+  if (computeV()) { \
+    localV.resize(ldvt, m_cols); \
+    vt   = (LAPACKE_TYPE*)localV.data(); \
+  } else { ldvt=1; vt=&dummy; }\
+  Matrix<LAPACKE_RTYPE, Dynamic, Dynamic> superb; superb.resize(m_diagSize, 1); \
+  MatrixType m_temp; m_temp = matrix; \
+  LAPACKE_##LAPACKE_PREFIX##gesvd( matrix_order, jobu, jobvt, internal::convert_index<lapack_int>(m_rows), internal::convert_index<lapack_int>(m_cols), (LAPACKE_TYPE*)m_temp.data(), lda, (LAPACKE_RTYPE*)m_singularValues.data(), u, ldu, vt, ldvt, superb.data()); \
+  if (computeV()) m_matrixV = localV.adjoint(); \
+ /* for(int i=0;i<m_diagSize;i++) if (m_singularValues.coeffRef(i) < precision) { m_nonzeroSingularValues--; m_singularValues.coeffRef(i)=RealScalar(0);}*/ \
+  m_isInitialized = true; \
+  return *this; \
+}
+
+EIGEN_LAPACKE_SVD(double,   double,                double, d, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SVD(float,    float,                 float , s, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SVD(dcomplex, lapack_complex_double, double, z, ColMajor, LAPACK_COL_MAJOR)
+EIGEN_LAPACKE_SVD(scomplex, lapack_complex_float,  float , c, ColMajor, LAPACK_COL_MAJOR)
+
+EIGEN_LAPACKE_SVD(double,   double,                double, d, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_SVD(float,    float,                 float , s, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_SVD(dcomplex, lapack_complex_double, double, z, RowMajor, LAPACK_ROW_MAJOR)
+EIGEN_LAPACKE_SVD(scomplex, lapack_complex_float,  float , c, RowMajor, LAPACK_ROW_MAJOR)
+
+} // end namespace Eigen
+
+#endif // EIGEN_JACOBISVD_LAPACKE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/SVDBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/SVDBase.h
new file mode 100644
index 0000000..cc90a3b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/SVDBase.h
@@ -0,0 +1,313 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// Copyright (C) 2013 Gauthier Brun <brun.gauthier@gmail.com>
+// Copyright (C) 2013 Nicolas Carre <nicolas.carre@ensimag.fr>
+// Copyright (C) 2013 Jean Ceccato <jean.ceccato@ensimag.fr>
+// Copyright (C) 2013 Pierre Zoppitelli <pierre.zoppitelli@ensimag.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SVDBASE_H
+#define EIGEN_SVDBASE_H
+
+namespace Eigen {
+/** \ingroup SVD_Module
+ *
+ *
+ * \class SVDBase
+ *
+ * \brief Base class of SVD algorithms
+ *
+ * \tparam Derived the type of the actual SVD decomposition
+ *
+ * SVD decomposition consists in decomposing any n-by-p matrix \a A as a product
+ *   \f[ A = U S V^* \f]
+ * where \a U is a n-by-n unitary, \a V is a p-by-p unitary, and \a S is a n-by-p real positive matrix which is zero outside of its main diagonal;
+ * the diagonal entries of S are known as the \em singular \em values of \a A and the columns of \a U and \a V are known as the left
+ * and right \em singular \em vectors of \a A respectively.
+ *
+ * Singular values are always sorted in decreasing order.
+ *
+ * 
+ * You can ask for only \em thin \a U or \a V to be computed, meaning the following. In case of a rectangular n-by-p matrix, letting \a m be the
+ * smaller value among \a n and \a p, there are only \a m singular vectors; the remaining columns of \a U and \a V do not correspond to actual
+ * singular vectors. Asking for \em thin \a U or \a V means asking for only their \a m first columns to be formed. So \a U is then a n-by-m matrix,
+ * and \a V is then a p-by-m matrix. Notice that thin \a U and \a V are all you need for (least squares) solving.
+ *  
+ * If the input matrix has inf or nan coefficients, the result of the computation is undefined, but the computation is guaranteed to
+ * terminate in finite (and reasonable) time.
+ * \sa class BDCSVD, class JacobiSVD
+ */
+template<typename Derived>
+class SVDBase
+{
+
+public:
+  typedef typename internal::traits<Derived>::MatrixType MatrixType;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename NumTraits<typename MatrixType::Scalar>::Real RealScalar;
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+  enum {
+    RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+    ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+    DiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_DYNAMIC(RowsAtCompileTime,ColsAtCompileTime),
+    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    MaxDiagSizeAtCompileTime = EIGEN_SIZE_MIN_PREFER_FIXED(MaxRowsAtCompileTime,MaxColsAtCompileTime),
+    MatrixOptions = MatrixType::Options
+  };
+
+  typedef Matrix<Scalar, RowsAtCompileTime, RowsAtCompileTime, MatrixOptions, MaxRowsAtCompileTime, MaxRowsAtCompileTime> MatrixUType;
+  typedef Matrix<Scalar, ColsAtCompileTime, ColsAtCompileTime, MatrixOptions, MaxColsAtCompileTime, MaxColsAtCompileTime> MatrixVType;
+  typedef typename internal::plain_diag_type<MatrixType, RealScalar>::type SingularValuesType;
+  
+  Derived& derived() { return *static_cast<Derived*>(this); }
+  const Derived& derived() const { return *static_cast<const Derived*>(this); }
+
+  /** \returns the \a U matrix.
+   *
+   * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p,
+   * the U matrix is n-by-n if you asked for \link Eigen::ComputeFullU ComputeFullU \endlink, and is n-by-m if you asked for \link Eigen::ComputeThinU ComputeThinU \endlink.
+   *
+   * The \a m first columns of \a U are the left singular vectors of the matrix being decomposed.
+   *
+   * This method asserts that you asked for \a U to be computed.
+   */
+  const MatrixUType& matrixU() const
+  {
+    eigen_assert(m_isInitialized && "SVD is not initialized.");
+    eigen_assert(computeU() && "This SVD decomposition didn't compute U. Did you ask for it?");
+    return m_matrixU;
+  }
+
+  /** \returns the \a V matrix.
+   *
+   * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p,
+   * the V matrix is p-by-p if you asked for \link Eigen::ComputeFullV ComputeFullV \endlink, and is p-by-m if you asked for \link Eigen::ComputeThinV ComputeThinV \endlink.
+   *
+   * The \a m first columns of \a V are the right singular vectors of the matrix being decomposed.
+   *
+   * This method asserts that you asked for \a V to be computed.
+   */
+  const MatrixVType& matrixV() const
+  {
+    eigen_assert(m_isInitialized && "SVD is not initialized.");
+    eigen_assert(computeV() && "This SVD decomposition didn't compute V. Did you ask for it?");
+    return m_matrixV;
+  }
+
+  /** \returns the vector of singular values.
+   *
+   * For the SVD decomposition of a n-by-p matrix, letting \a m be the minimum of \a n and \a p, the
+   * returned vector has size \a m.  Singular values are always sorted in decreasing order.
+   */
+  const SingularValuesType& singularValues() const
+  {
+    eigen_assert(m_isInitialized && "SVD is not initialized.");
+    return m_singularValues;
+  }
+
+  /** \returns the number of singular values that are not exactly 0 */
+  Index nonzeroSingularValues() const
+  {
+    eigen_assert(m_isInitialized && "SVD is not initialized.");
+    return m_nonzeroSingularValues;
+  }
+  
+  /** \returns the rank of the matrix of which \c *this is the SVD.
+    *
+    * \note This method has to determine which singular values should be considered nonzero.
+    *       For that, it uses the threshold value that you can control by calling
+    *       setThreshold(const RealScalar&).
+    */
+  inline Index rank() const
+  {
+    using std::abs;
+    eigen_assert(m_isInitialized && "JacobiSVD is not initialized.");
+    if(m_singularValues.size()==0) return 0;
+    RealScalar premultiplied_threshold = numext::maxi<RealScalar>(m_singularValues.coeff(0) * threshold(), (std::numeric_limits<RealScalar>::min)());
+    Index i = m_nonzeroSingularValues-1;
+    while(i>=0 && m_singularValues.coeff(i) < premultiplied_threshold) --i;
+    return i+1;
+  }
+  
+  /** Allows to prescribe a threshold to be used by certain methods, such as rank() and solve(),
+    * which need to determine when singular values are to be considered nonzero.
+    * This is not used for the SVD decomposition itself.
+    *
+    * When it needs to get the threshold value, Eigen calls threshold().
+    * The default is \c NumTraits<Scalar>::epsilon()
+    *
+    * \param threshold The new value to use as the threshold.
+    *
+    * A singular value will be considered nonzero if its value is strictly greater than
+    *  \f$ \vert singular value \vert \leqslant threshold \times \vert max singular value \vert \f$.
+    *
+    * If you want to come back to the default behavior, call setThreshold(Default_t)
+    */
+  Derived& setThreshold(const RealScalar& threshold)
+  {
+    m_usePrescribedThreshold = true;
+    m_prescribedThreshold = threshold;
+    return derived();
+  }
+
+  /** Allows to come back to the default behavior, letting Eigen use its default formula for
+    * determining the threshold.
+    *
+    * You should pass the special object Eigen::Default as parameter here.
+    * \code svd.setThreshold(Eigen::Default); \endcode
+    *
+    * See the documentation of setThreshold(const RealScalar&).
+    */
+  Derived& setThreshold(Default_t)
+  {
+    m_usePrescribedThreshold = false;
+    return derived();
+  }
+
+  /** Returns the threshold that will be used by certain methods such as rank().
+    *
+    * See the documentation of setThreshold(const RealScalar&).
+    */
+  RealScalar threshold() const
+  {
+    eigen_assert(m_isInitialized || m_usePrescribedThreshold);
+    return m_usePrescribedThreshold ? m_prescribedThreshold
+                                    : (std::max<Index>)(1,m_diagSize)*NumTraits<Scalar>::epsilon();
+  }
+
+  /** \returns true if \a U (full or thin) is asked for in this SVD decomposition */
+  inline bool computeU() const { return m_computeFullU || m_computeThinU; }
+  /** \returns true if \a V (full or thin) is asked for in this SVD decomposition */
+  inline bool computeV() const { return m_computeFullV || m_computeThinV; }
+
+  inline Index rows() const { return m_rows; }
+  inline Index cols() const { return m_cols; }
+  
+  /** \returns a (least squares) solution of \f$ A x = b \f$ using the current SVD decomposition of A.
+    *
+    * \param b the right-hand-side of the equation to solve.
+    *
+    * \note Solving requires both U and V to be computed. Thin U and V are enough, there is no need for full U or V.
+    *
+    * \note SVD solving is implicitly least-squares. Thus, this method serves both purposes of exact solving and least-squares solving.
+    * In other words, the returned solution is guaranteed to minimize the Euclidean norm \f$ \Vert A x - b \Vert \f$.
+    */
+  template<typename Rhs>
+  inline const Solve<Derived, Rhs>
+  solve(const MatrixBase<Rhs>& b) const
+  {
+    eigen_assert(m_isInitialized && "SVD is not initialized.");
+    eigen_assert(computeU() && computeV() && "SVD::solve() requires both unitaries U and V to be computed (thin unitaries suffice).");
+    return Solve<Derived, Rhs>(derived(), b.derived());
+  }
+  
+  #ifndef EIGEN_PARSED_BY_DOXYGEN
+  template<typename RhsType, typename DstType>
+  EIGEN_DEVICE_FUNC
+  void _solve_impl(const RhsType &rhs, DstType &dst) const;
+  #endif
+
+protected:
+  
+  static void check_template_parameters()
+  {
+    EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+  }
+  
+  // return true if already allocated
+  bool allocate(Index rows, Index cols, unsigned int computationOptions) ;
+
+  MatrixUType m_matrixU;
+  MatrixVType m_matrixV;
+  SingularValuesType m_singularValues;
+  bool m_isInitialized, m_isAllocated, m_usePrescribedThreshold;
+  bool m_computeFullU, m_computeThinU;
+  bool m_computeFullV, m_computeThinV;
+  unsigned int m_computationOptions;
+  Index m_nonzeroSingularValues, m_rows, m_cols, m_diagSize;
+  RealScalar m_prescribedThreshold;
+
+  /** \brief Default Constructor.
+   *
+   * Default constructor of SVDBase
+   */
+  SVDBase()
+    : m_isInitialized(false),
+      m_isAllocated(false),
+      m_usePrescribedThreshold(false),
+      m_computationOptions(0),
+      m_rows(-1), m_cols(-1), m_diagSize(0)
+  {
+    check_template_parameters();
+  }
+
+
+};
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename Derived>
+template<typename RhsType, typename DstType>
+void SVDBase<Derived>::_solve_impl(const RhsType &rhs, DstType &dst) const
+{
+  eigen_assert(rhs.rows() == rows());
+
+  // A = U S V^*
+  // So A^{-1} = V S^{-1} U^*
+
+  Matrix<Scalar, Dynamic, RhsType::ColsAtCompileTime, 0, MatrixType::MaxRowsAtCompileTime, RhsType::MaxColsAtCompileTime> tmp;
+  Index l_rank = rank();
+  tmp.noalias() =  m_matrixU.leftCols(l_rank).adjoint() * rhs;
+  tmp = m_singularValues.head(l_rank).asDiagonal().inverse() * tmp;
+  dst = m_matrixV.leftCols(l_rank) * tmp;
+}
+#endif
+
+template<typename MatrixType>
+bool SVDBase<MatrixType>::allocate(Index rows, Index cols, unsigned int computationOptions)
+{
+  eigen_assert(rows >= 0 && cols >= 0);
+
+  if (m_isAllocated &&
+      rows == m_rows &&
+      cols == m_cols &&
+      computationOptions == m_computationOptions)
+  {
+    return true;
+  }
+
+  m_rows = rows;
+  m_cols = cols;
+  m_isInitialized = false;
+  m_isAllocated = true;
+  m_computationOptions = computationOptions;
+  m_computeFullU = (computationOptions & ComputeFullU) != 0;
+  m_computeThinU = (computationOptions & ComputeThinU) != 0;
+  m_computeFullV = (computationOptions & ComputeFullV) != 0;
+  m_computeThinV = (computationOptions & ComputeThinV) != 0;
+  eigen_assert(!(m_computeFullU && m_computeThinU) && "SVDBase: you can't ask for both full and thin U");
+  eigen_assert(!(m_computeFullV && m_computeThinV) && "SVDBase: you can't ask for both full and thin V");
+  eigen_assert(EIGEN_IMPLIES(m_computeThinU || m_computeThinV, MatrixType::ColsAtCompileTime==Dynamic) &&
+	       "SVDBase: thin U and V are only available when your matrix has a dynamic number of columns.");
+
+  m_diagSize = (std::min)(m_rows, m_cols);
+  m_singularValues.resize(m_diagSize);
+  if(RowsAtCompileTime==Dynamic)
+    m_matrixU.resize(m_rows, m_computeFullU ? m_rows : m_computeThinU ? m_diagSize : 0);
+  if(ColsAtCompileTime==Dynamic)
+    m_matrixV.resize(m_cols, m_computeFullV ? m_cols : m_computeThinV ? m_diagSize : 0);
+
+  return false;
+}
+
+}// end namespace
+
+#endif // EIGEN_SVDBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/UpperBidiagonalization.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/UpperBidiagonalization.h
new file mode 100644
index 0000000..11ac847
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SVD/UpperBidiagonalization.h
@@ -0,0 +1,414 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2013-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_BIDIAGONALIZATION_H
+#define EIGEN_BIDIAGONALIZATION_H
+
+namespace Eigen { 
+
+namespace internal {
+// UpperBidiagonalization will probably be replaced by a Bidiagonalization class, don't want to make it stable API.
+// At the same time, it's useful to keep for now as it's about the only thing that is testing the BandMatrix class.
+
+template<typename _MatrixType> class UpperBidiagonalization
+{
+  public:
+
+    typedef _MatrixType MatrixType;
+    enum {
+      RowsAtCompileTime = MatrixType::RowsAtCompileTime,
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      ColsAtCompileTimeMinusOne = internal::decrement_size<ColsAtCompileTime>::ret
+    };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef Eigen::Index Index; ///< \deprecated since Eigen 3.3
+    typedef Matrix<Scalar, 1, ColsAtCompileTime> RowVectorType;
+    typedef Matrix<Scalar, RowsAtCompileTime, 1> ColVectorType;
+    typedef BandMatrix<RealScalar, ColsAtCompileTime, ColsAtCompileTime, 1, 0, RowMajor> BidiagonalType;
+    typedef Matrix<Scalar, ColsAtCompileTime, 1> DiagVectorType;
+    typedef Matrix<Scalar, ColsAtCompileTimeMinusOne, 1> SuperDiagVectorType;
+    typedef HouseholderSequence<
+              const MatrixType,
+              const typename internal::remove_all<typename Diagonal<const MatrixType,0>::ConjugateReturnType>::type
+            > HouseholderUSequenceType;
+    typedef HouseholderSequence<
+              const typename internal::remove_all<typename MatrixType::ConjugateReturnType>::type,
+              Diagonal<const MatrixType,1>,
+              OnTheRight
+            > HouseholderVSequenceType;
+    
+    /**
+    * \brief Default Constructor.
+    *
+    * The default constructor is useful in cases in which the user intends to
+    * perform decompositions via Bidiagonalization::compute(const MatrixType&).
+    */
+    UpperBidiagonalization() : m_householder(), m_bidiagonal(), m_isInitialized(false) {}
+
+    explicit UpperBidiagonalization(const MatrixType& matrix)
+      : m_householder(matrix.rows(), matrix.cols()),
+        m_bidiagonal(matrix.cols(), matrix.cols()),
+        m_isInitialized(false)
+    {
+      compute(matrix);
+    }
+    
+    UpperBidiagonalization& compute(const MatrixType& matrix);
+    UpperBidiagonalization& computeUnblocked(const MatrixType& matrix);
+    
+    const MatrixType& householder() const { return m_householder; }
+    const BidiagonalType& bidiagonal() const { return m_bidiagonal; }
+    
+    const HouseholderUSequenceType householderU() const
+    {
+      eigen_assert(m_isInitialized && "UpperBidiagonalization is not initialized.");
+      return HouseholderUSequenceType(m_householder, m_householder.diagonal().conjugate());
+    }
+
+    const HouseholderVSequenceType householderV() // const here gives nasty errors and i'm lazy
+    {
+      eigen_assert(m_isInitialized && "UpperBidiagonalization is not initialized.");
+      return HouseholderVSequenceType(m_householder.conjugate(), m_householder.const_derived().template diagonal<1>())
+             .setLength(m_householder.cols()-1)
+             .setShift(1);
+    }
+    
+  protected:
+    MatrixType m_householder;
+    BidiagonalType m_bidiagonal;
+    bool m_isInitialized;
+};
+
+// Standard upper bidiagonalization without fancy optimizations
+// This version should be faster for small matrix size
+template<typename MatrixType>
+void upperbidiagonalization_inplace_unblocked(MatrixType& mat,
+                                              typename MatrixType::RealScalar *diagonal,
+                                              typename MatrixType::RealScalar *upper_diagonal,
+                                              typename MatrixType::Scalar* tempData = 0)
+{
+  typedef typename MatrixType::Scalar Scalar;
+
+  Index rows = mat.rows();
+  Index cols = mat.cols();
+
+  typedef Matrix<Scalar,Dynamic,1,ColMajor,MatrixType::MaxRowsAtCompileTime,1> TempType;
+  TempType tempVector;
+  if(tempData==0)
+  {
+    tempVector.resize(rows);
+    tempData = tempVector.data();
+  }
+
+  for (Index k = 0; /* breaks at k==cols-1 below */ ; ++k)
+  {
+    Index remainingRows = rows - k;
+    Index remainingCols = cols - k - 1;
+
+    // construct left householder transform in-place in A
+    mat.col(k).tail(remainingRows)
+       .makeHouseholderInPlace(mat.coeffRef(k,k), diagonal[k]);
+    // apply householder transform to remaining part of A on the left
+    mat.bottomRightCorner(remainingRows, remainingCols)
+       .applyHouseholderOnTheLeft(mat.col(k).tail(remainingRows-1), mat.coeff(k,k), tempData);
+
+    if(k == cols-1) break;
+
+    // construct right householder transform in-place in mat
+    mat.row(k).tail(remainingCols)
+       .makeHouseholderInPlace(mat.coeffRef(k,k+1), upper_diagonal[k]);
+    // apply householder transform to remaining part of mat on the left
+    mat.bottomRightCorner(remainingRows-1, remainingCols)
+       .applyHouseholderOnTheRight(mat.row(k).tail(remainingCols-1).transpose(), mat.coeff(k,k+1), tempData);
+  }
+}
+
+/** \internal
+  * Helper routine for the block reduction to upper bidiagonal form.
+  *
+  * Let's partition the matrix A:
+  * 
+  *      | A00 A01 |
+  *  A = |         |
+  *      | A10 A11 |
+  *
+  * This function reduces to bidiagonal form the left \c rows x \a blockSize vertical panel [A00/A10]
+  * and the \a blockSize x \c cols horizontal panel [A00 A01] of the matrix \a A. The bottom-right block A11
+  * is updated using matrix-matrix products:
+  *   A22 -= V * Y^T - X * U^T
+  * where V and U contains the left and right Householder vectors. U and V are stored in A10, and A01
+  * respectively, and the update matrices X and Y are computed during the reduction.
+  * 
+  */
+template<typename MatrixType>
+void upperbidiagonalization_blocked_helper(MatrixType& A,
+                                           typename MatrixType::RealScalar *diagonal,
+                                           typename MatrixType::RealScalar *upper_diagonal,
+                                           Index bs,
+                                           Ref<Matrix<typename MatrixType::Scalar, Dynamic, Dynamic,
+                                                      traits<MatrixType>::Flags & RowMajorBit> > X,
+                                           Ref<Matrix<typename MatrixType::Scalar, Dynamic, Dynamic,
+                                                      traits<MatrixType>::Flags & RowMajorBit> > Y)
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef typename MatrixType::RealScalar RealScalar;
+  typedef typename NumTraits<RealScalar>::Literal Literal;
+  enum { StorageOrder = traits<MatrixType>::Flags & RowMajorBit };
+  typedef InnerStride<int(StorageOrder) == int(ColMajor) ? 1 : Dynamic> ColInnerStride;
+  typedef InnerStride<int(StorageOrder) == int(ColMajor) ? Dynamic : 1> RowInnerStride;
+  typedef Ref<Matrix<Scalar, Dynamic, 1>, 0, ColInnerStride>    SubColumnType;
+  typedef Ref<Matrix<Scalar, 1, Dynamic>, 0, RowInnerStride>    SubRowType;
+  typedef Ref<Matrix<Scalar, Dynamic, Dynamic, StorageOrder > > SubMatType;
+  
+  Index brows = A.rows();
+  Index bcols = A.cols();
+
+  Scalar tau_u, tau_u_prev(0), tau_v;
+
+  for(Index k = 0; k < bs; ++k)
+  {
+    Index remainingRows = brows - k;
+    Index remainingCols = bcols - k - 1;
+
+    SubMatType X_k1( X.block(k,0, remainingRows,k) );
+    SubMatType V_k1( A.block(k,0, remainingRows,k) );
+
+    // 1 - update the k-th column of A
+    SubColumnType v_k = A.col(k).tail(remainingRows);
+          v_k -= V_k1 * Y.row(k).head(k).adjoint();
+    if(k) v_k -= X_k1 * A.col(k).head(k);
+    
+    // 2 - construct left Householder transform in-place
+    v_k.makeHouseholderInPlace(tau_v, diagonal[k]);
+       
+    if(k+1<bcols)
+    {
+      SubMatType Y_k  ( Y.block(k+1,0, remainingCols, k+1) );
+      SubMatType U_k1 ( A.block(0,k+1, k,remainingCols) );
+      
+      // this eases the application of Householder transforAions
+      // A(k,k) will store tau_v later
+      A(k,k) = Scalar(1);
+
+      // 3 - Compute y_k^T = tau_v * ( A^T*v_k - Y_k-1*V_k-1^T*v_k - U_k-1*X_k-1^T*v_k )
+      {
+        SubColumnType y_k( Y.col(k).tail(remainingCols) );
+        
+        // let's use the begining of column k of Y as a temporary vector
+        SubColumnType tmp( Y.col(k).head(k) );
+        y_k.noalias()  = A.block(k,k+1, remainingRows,remainingCols).adjoint() * v_k; // bottleneck
+        tmp.noalias()  = V_k1.adjoint()  * v_k;
+        y_k.noalias() -= Y_k.leftCols(k) * tmp;
+        tmp.noalias()  = X_k1.adjoint()  * v_k;
+        y_k.noalias() -= U_k1.adjoint()  * tmp;
+        y_k *= numext::conj(tau_v);
+      }
+
+      // 4 - update k-th row of A (it will become u_k)
+      SubRowType u_k( A.row(k).tail(remainingCols) );
+      u_k = u_k.conjugate();
+      {
+        u_k -= Y_k * A.row(k).head(k+1).adjoint();
+        if(k) u_k -= U_k1.adjoint() * X.row(k).head(k).adjoint();
+      }
+
+      // 5 - construct right Householder transform in-place
+      u_k.makeHouseholderInPlace(tau_u, upper_diagonal[k]);
+
+      // this eases the application of Householder transformations
+      // A(k,k+1) will store tau_u later
+      A(k,k+1) = Scalar(1);
+
+      // 6 - Compute x_k = tau_u * ( A*u_k - X_k-1*U_k-1^T*u_k - V_k*Y_k^T*u_k )
+      {
+        SubColumnType x_k ( X.col(k).tail(remainingRows-1) );
+        
+        // let's use the begining of column k of X as a temporary vectors
+        // note that tmp0 and tmp1 overlaps
+        SubColumnType tmp0 ( X.col(k).head(k) ),
+                      tmp1 ( X.col(k).head(k+1) );
+                    
+        x_k.noalias()   = A.block(k+1,k+1, remainingRows-1,remainingCols) * u_k.transpose(); // bottleneck
+        tmp0.noalias()  = U_k1 * u_k.transpose();
+        x_k.noalias()  -= X_k1.bottomRows(remainingRows-1) * tmp0;
+        tmp1.noalias()  = Y_k.adjoint() * u_k.transpose();
+        x_k.noalias()  -= A.block(k+1,0, remainingRows-1,k+1) * tmp1;
+        x_k *= numext::conj(tau_u);
+        tau_u = numext::conj(tau_u);
+        u_k = u_k.conjugate();
+      }
+
+      if(k>0) A.coeffRef(k-1,k) = tau_u_prev;
+      tau_u_prev = tau_u;
+    }
+    else
+      A.coeffRef(k-1,k) = tau_u_prev;
+
+    A.coeffRef(k,k) = tau_v;
+  }
+  
+  if(bs<bcols)
+    A.coeffRef(bs-1,bs) = tau_u_prev;
+
+  // update A22
+  if(bcols>bs && brows>bs)
+  {
+    SubMatType A11( A.bottomRightCorner(brows-bs,bcols-bs) );
+    SubMatType A10( A.block(bs,0, brows-bs,bs) );
+    SubMatType A01( A.block(0,bs, bs,bcols-bs) );
+    Scalar tmp = A01(bs-1,0);
+    A01(bs-1,0) = Literal(1);
+    A11.noalias() -= A10 * Y.topLeftCorner(bcols,bs).bottomRows(bcols-bs).adjoint();
+    A11.noalias() -= X.topLeftCorner(brows,bs).bottomRows(brows-bs) * A01;
+    A01(bs-1,0) = tmp;
+  }
+}
+
+/** \internal
+  *
+  * Implementation of a block-bidiagonal reduction.
+  * It is based on the following paper:
+  *   The Design of a Parallel Dense Linear Algebra Software Library: Reduction to Hessenberg, Tridiagonal, and Bidiagonal Form.
+  *   by Jaeyoung Choi, Jack J. Dongarra, David W. Walker. (1995)
+  *   section 3.3
+  */
+template<typename MatrixType, typename BidiagType>
+void upperbidiagonalization_inplace_blocked(MatrixType& A, BidiagType& bidiagonal,
+                                            Index maxBlockSize=32,
+                                            typename MatrixType::Scalar* /*tempData*/ = 0)
+{
+  typedef typename MatrixType::Scalar Scalar;
+  typedef Block<MatrixType,Dynamic,Dynamic> BlockType;
+
+  Index rows = A.rows();
+  Index cols = A.cols();
+  Index size = (std::min)(rows, cols);
+
+  // X and Y are work space
+  enum { StorageOrder = traits<MatrixType>::Flags & RowMajorBit };
+  Matrix<Scalar,
+         MatrixType::RowsAtCompileTime,
+         Dynamic,
+         StorageOrder,
+         MatrixType::MaxRowsAtCompileTime> X(rows,maxBlockSize);
+  Matrix<Scalar,
+         MatrixType::ColsAtCompileTime,
+         Dynamic,
+         StorageOrder,
+         MatrixType::MaxColsAtCompileTime> Y(cols,maxBlockSize);
+  Index blockSize = (std::min)(maxBlockSize,size);
+
+  Index k = 0;
+  for(k = 0; k < size; k += blockSize)
+  {
+    Index bs = (std::min)(size-k,blockSize);  // actual size of the block
+    Index brows = rows - k;                   // rows of the block
+    Index bcols = cols - k;                   // columns of the block
+
+    // partition the matrix A:
+    // 
+    //      | A00 A01 A02 |
+    //      |             |
+    // A  = | A10 A11 A12 |
+    //      |             |
+    //      | A20 A21 A22 |
+    //
+    // where A11 is a bs x bs diagonal block,
+    // and let:
+    //      | A11 A12 |
+    //  B = |         |
+    //      | A21 A22 |
+
+    BlockType B = A.block(k,k,brows,bcols);
+    
+    // This stage performs the bidiagonalization of A11, A21, A12, and updating of A22.
+    // Finally, the algorithm continue on the updated A22.
+    //
+    // However, if B is too small, or A22 empty, then let's use an unblocked strategy
+    if(k+bs==cols || bcols<48) // somewhat arbitrary threshold
+    {
+      upperbidiagonalization_inplace_unblocked(B,
+                                               &(bidiagonal.template diagonal<0>().coeffRef(k)),
+                                               &(bidiagonal.template diagonal<1>().coeffRef(k)),
+                                               X.data()
+                                              );
+      break; // We're done
+    }
+    else
+    {
+      upperbidiagonalization_blocked_helper<BlockType>( B,
+                                                        &(bidiagonal.template diagonal<0>().coeffRef(k)),
+                                                        &(bidiagonal.template diagonal<1>().coeffRef(k)),
+                                                        bs,
+                                                        X.topLeftCorner(brows,bs),
+                                                        Y.topLeftCorner(bcols,bs)
+                                                      );
+    }
+  }
+}
+
+template<typename _MatrixType>
+UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::computeUnblocked(const _MatrixType& matrix)
+{
+  Index rows = matrix.rows();
+  Index cols = matrix.cols();
+  EIGEN_ONLY_USED_FOR_DEBUG(cols);
+
+  eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols.");
+
+  m_householder = matrix;
+
+  ColVectorType temp(rows);
+
+  upperbidiagonalization_inplace_unblocked(m_householder,
+                                           &(m_bidiagonal.template diagonal<0>().coeffRef(0)),
+                                           &(m_bidiagonal.template diagonal<1>().coeffRef(0)),
+                                           temp.data());
+
+  m_isInitialized = true;
+  return *this;
+}
+
+template<typename _MatrixType>
+UpperBidiagonalization<_MatrixType>& UpperBidiagonalization<_MatrixType>::compute(const _MatrixType& matrix)
+{
+  Index rows = matrix.rows();
+  Index cols = matrix.cols();
+  EIGEN_ONLY_USED_FOR_DEBUG(rows);
+  EIGEN_ONLY_USED_FOR_DEBUG(cols);
+
+  eigen_assert(rows >= cols && "UpperBidiagonalization is only for Arices satisfying rows>=cols.");
+
+  m_householder = matrix;
+  upperbidiagonalization_inplace_blocked(m_householder, m_bidiagonal);
+            
+  m_isInitialized = true;
+  return *this;
+}
+
+#if 0
+/** \return the Householder QR decomposition of \c *this.
+  *
+  * \sa class Bidiagonalization
+  */
+template<typename Derived>
+const UpperBidiagonalization<typename MatrixBase<Derived>::PlainObject>
+MatrixBase<Derived>::bidiagonalization() const
+{
+  return UpperBidiagonalization<PlainObject>(eval());
+}
+#endif
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_BIDIAGONALIZATION_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky.h
new file mode 100644
index 0000000..2907f65
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky.h
@@ -0,0 +1,689 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SIMPLICIAL_CHOLESKY_H
+#define EIGEN_SIMPLICIAL_CHOLESKY_H
+
+namespace Eigen { 
+
+enum SimplicialCholeskyMode {
+  SimplicialCholeskyLLT,
+  SimplicialCholeskyLDLT
+};
+
+namespace internal {
+  template<typename CholMatrixType, typename InputMatrixType>
+  struct simplicial_cholesky_grab_input {
+    typedef CholMatrixType const * ConstCholMatrixPtr;
+    static void run(const InputMatrixType& input, ConstCholMatrixPtr &pmat, CholMatrixType &tmp)
+    {
+      tmp = input;
+      pmat = &tmp;
+    }
+  };
+  
+  template<typename MatrixType>
+  struct simplicial_cholesky_grab_input<MatrixType,MatrixType> {
+    typedef MatrixType const * ConstMatrixPtr;
+    static void run(const MatrixType& input, ConstMatrixPtr &pmat, MatrixType &/*tmp*/)
+    {
+      pmat = &input;
+    }
+  };
+} // end namespace internal
+
+/** \ingroup SparseCholesky_Module
+  * \brief A base class for direct sparse Cholesky factorizations
+  *
+  * This is a base class for LL^T and LDL^T Cholesky factorizations of sparse matrices that are
+  * selfadjoint and positive definite. These factorizations allow for solving A.X = B where
+  * X and B can be either dense or sparse.
+  * 
+  * In order to reduce the fill-in, a symmetric permutation P is applied prior to the factorization
+  * such that the factorized matrix is P A P^-1.
+  *
+  * \tparam Derived the type of the derived class, that is the actual factorization type.
+  *
+  */
+template<typename Derived>
+class SimplicialCholeskyBase : public SparseSolverBase<Derived>
+{
+    typedef SparseSolverBase<Derived> Base;
+    using Base::m_isInitialized;
+    
+  public:
+    typedef typename internal::traits<Derived>::MatrixType MatrixType;
+    typedef typename internal::traits<Derived>::OrderingType OrderingType;
+    enum { UpLo = internal::traits<Derived>::UpLo };
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> CholMatrixType;
+    typedef CholMatrixType const * ConstCholMatrixPtr;
+    typedef Matrix<Scalar,Dynamic,1> VectorType;
+    typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+  public:
+    
+    using Base::derived;
+
+    /** Default constructor */
+    SimplicialCholeskyBase()
+      : m_info(Success), m_shiftOffset(0), m_shiftScale(1)
+    {}
+
+    explicit SimplicialCholeskyBase(const MatrixType& matrix)
+      : m_info(Success), m_shiftOffset(0), m_shiftScale(1)
+    {
+      derived().compute(matrix);
+    }
+
+    ~SimplicialCholeskyBase()
+    {
+    }
+
+    Derived& derived() { return *static_cast<Derived*>(this); }
+    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    
+    inline Index cols() const { return m_matrix.cols(); }
+    inline Index rows() const { return m_matrix.rows(); }
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+    
+    /** \returns the permutation P
+      * \sa permutationPinv() */
+    const PermutationMatrix<Dynamic,Dynamic,StorageIndex>& permutationP() const
+    { return m_P; }
+    
+    /** \returns the inverse P^-1 of the permutation P
+      * \sa permutationP() */
+    const PermutationMatrix<Dynamic,Dynamic,StorageIndex>& permutationPinv() const
+    { return m_Pinv; }
+
+    /** Sets the shift parameters that will be used to adjust the diagonal coefficients during the numerical factorization.
+      *
+      * During the numerical factorization, the diagonal coefficients are transformed by the following linear model:\n
+      * \c d_ii = \a offset + \a scale * \c d_ii
+      *
+      * The default is the identity transformation with \a offset=0, and \a scale=1.
+      *
+      * \returns a reference to \c *this.
+      */
+    Derived& setShift(const RealScalar& offset, const RealScalar& scale = 1)
+    {
+      m_shiftOffset = offset;
+      m_shiftScale = scale;
+      return derived();
+    }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal */
+    template<typename Stream>
+    void dumpMemory(Stream& s)
+    {
+      int total = 0;
+      s << "  L:        " << ((total+=(m_matrix.cols()+1) * sizeof(int) + m_matrix.nonZeros()*(sizeof(int)+sizeof(Scalar))) >> 20) << "Mb" << "\n";
+      s << "  diag:     " << ((total+=m_diag.size() * sizeof(Scalar)) >> 20) << "Mb" << "\n";
+      s << "  tree:     " << ((total+=m_parent.size() * sizeof(int)) >> 20) << "Mb" << "\n";
+      s << "  nonzeros: " << ((total+=m_nonZerosPerCol.size() * sizeof(int)) >> 20) << "Mb" << "\n";
+      s << "  perm:     " << ((total+=m_P.size() * sizeof(int)) >> 20) << "Mb" << "\n";
+      s << "  perm^-1:  " << ((total+=m_Pinv.size() * sizeof(int)) >> 20) << "Mb" << "\n";
+      s << "  TOTAL:    " << (total>> 20) << "Mb" << "\n";
+    }
+
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    {
+      eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
+      eigen_assert(m_matrix.rows()==b.rows());
+
+      if(m_info!=Success)
+        return;
+
+      if(m_P.size()>0)
+        dest = m_P * b;
+      else
+        dest = b;
+
+      if(m_matrix.nonZeros()>0) // otherwise L==I
+        derived().matrixL().solveInPlace(dest);
+
+      if(m_diag.size()>0)
+        dest = m_diag.asDiagonal().inverse() * dest;
+
+      if (m_matrix.nonZeros()>0) // otherwise U==I
+        derived().matrixU().solveInPlace(dest);
+
+      if(m_P.size()>0)
+        dest = m_Pinv * dest;
+    }
+    
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const SparseMatrixBase<Rhs> &b, SparseMatrixBase<Dest> &dest) const
+    {
+      internal::solve_sparse_through_dense_panels(derived(), b, dest);
+    }
+
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
+  protected:
+    
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    template<bool DoLDLT>
+    void compute(const MatrixType& matrix)
+    {
+      eigen_assert(matrix.rows()==matrix.cols());
+      Index size = matrix.cols();
+      CholMatrixType tmp(size,size);
+      ConstCholMatrixPtr pmat;
+      ordering(matrix, pmat, tmp);
+      analyzePattern_preordered(*pmat, DoLDLT);
+      factorize_preordered<DoLDLT>(*pmat);
+    }
+    
+    template<bool DoLDLT>
+    void factorize(const MatrixType& a)
+    {
+      eigen_assert(a.rows()==a.cols());
+      Index size = a.cols();
+      CholMatrixType tmp(size,size);
+      ConstCholMatrixPtr pmat;
+      
+      if(m_P.size()==0 && (UpLo&Upper)==Upper)
+      {
+        // If there is no ordering, try to directly use the input matrix without any copy
+        internal::simplicial_cholesky_grab_input<CholMatrixType,MatrixType>::run(a, pmat, tmp);
+      }
+      else
+      {
+        tmp.template selfadjointView<Upper>() = a.template selfadjointView<UpLo>().twistedBy(m_P);
+        pmat = &tmp;
+      }
+      
+      factorize_preordered<DoLDLT>(*pmat);
+    }
+
+    template<bool DoLDLT>
+    void factorize_preordered(const CholMatrixType& a);
+
+    void analyzePattern(const MatrixType& a, bool doLDLT)
+    {
+      eigen_assert(a.rows()==a.cols());
+      Index size = a.cols();
+      CholMatrixType tmp(size,size);
+      ConstCholMatrixPtr pmat;
+      ordering(a, pmat, tmp);
+      analyzePattern_preordered(*pmat,doLDLT);
+    }
+    void analyzePattern_preordered(const CholMatrixType& a, bool doLDLT);
+    
+    void ordering(const MatrixType& a, ConstCholMatrixPtr &pmat, CholMatrixType& ap);
+
+    /** keeps off-diagonal entries; drops diagonal entries */
+    struct keep_diag {
+      inline bool operator() (const Index& row, const Index& col, const Scalar&) const
+      {
+        return row!=col;
+      }
+    };
+
+    mutable ComputationInfo m_info;
+    bool m_factorizationIsOk;
+    bool m_analysisIsOk;
+    
+    CholMatrixType m_matrix;
+    VectorType m_diag;                                // the diagonal coefficients (LDLT mode)
+    VectorI m_parent;                                 // elimination tree
+    VectorI m_nonZerosPerCol;
+    PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_P;     // the permutation
+    PermutationMatrix<Dynamic,Dynamic,StorageIndex> m_Pinv;  // the inverse permutation
+
+    RealScalar m_shiftOffset;
+    RealScalar m_shiftScale;
+};
+
+template<typename _MatrixType, int _UpLo = Lower, typename _Ordering = AMDOrdering<typename _MatrixType::StorageIndex> > class SimplicialLLT;
+template<typename _MatrixType, int _UpLo = Lower, typename _Ordering = AMDOrdering<typename _MatrixType::StorageIndex> > class SimplicialLDLT;
+template<typename _MatrixType, int _UpLo = Lower, typename _Ordering = AMDOrdering<typename _MatrixType::StorageIndex> > class SimplicialCholesky;
+
+namespace internal {
+
+template<typename _MatrixType, int _UpLo, typename _Ordering> struct traits<SimplicialLLT<_MatrixType,_UpLo,_Ordering> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Ordering OrderingType;
+  enum { UpLo = _UpLo };
+  typedef typename MatrixType::Scalar                         Scalar;
+  typedef typename MatrixType::StorageIndex                   StorageIndex;
+  typedef SparseMatrix<Scalar, ColMajor, StorageIndex>        CholMatrixType;
+  typedef TriangularView<const CholMatrixType, Eigen::Lower>  MatrixL;
+  typedef TriangularView<const typename CholMatrixType::AdjointReturnType, Eigen::Upper>   MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
+};
+
+template<typename _MatrixType,int _UpLo, typename _Ordering> struct traits<SimplicialLDLT<_MatrixType,_UpLo,_Ordering> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Ordering OrderingType;
+  enum { UpLo = _UpLo };
+  typedef typename MatrixType::Scalar                             Scalar;
+  typedef typename MatrixType::StorageIndex                       StorageIndex;
+  typedef SparseMatrix<Scalar, ColMajor, StorageIndex>            CholMatrixType;
+  typedef TriangularView<const CholMatrixType, Eigen::UnitLower>  MatrixL;
+  typedef TriangularView<const typename CholMatrixType::AdjointReturnType, Eigen::UnitUpper> MatrixU;
+  static inline MatrixL getL(const MatrixType& m) { return MatrixL(m); }
+  static inline MatrixU getU(const MatrixType& m) { return MatrixU(m.adjoint()); }
+};
+
+template<typename _MatrixType, int _UpLo, typename _Ordering> struct traits<SimplicialCholesky<_MatrixType,_UpLo,_Ordering> >
+{
+  typedef _MatrixType MatrixType;
+  typedef _Ordering OrderingType;
+  enum { UpLo = _UpLo };
+};
+
+}
+
+/** \ingroup SparseCholesky_Module
+  * \class SimplicialLLT
+  * \brief A direct sparse LLT Cholesky factorizations
+  *
+  * This class provides a LL^T Cholesky factorizations of sparse matrices that are
+  * selfadjoint and positive definite. The factorization allows for solving A.X = B where
+  * X and B can be either dense or sparse.
+  * 
+  * In order to reduce the fill-in, a symmetric permutation P is applied prior to the factorization
+  * such that the factorized matrix is P A P^-1.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  * \tparam _Ordering The ordering method to use, either AMDOrdering<> or NaturalOrdering<>. Default is AMDOrdering<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa class SimplicialLDLT, class AMDOrdering, class NaturalOrdering
+  */
+template<typename _MatrixType, int _UpLo, typename _Ordering>
+    class SimplicialLLT : public SimplicialCholeskyBase<SimplicialLLT<_MatrixType,_UpLo,_Ordering> >
+{
+public:
+    typedef _MatrixType MatrixType;
+    enum { UpLo = _UpLo };
+    typedef SimplicialCholeskyBase<SimplicialLLT> Base;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,Index> CholMatrixType;
+    typedef Matrix<Scalar,Dynamic,1> VectorType;
+    typedef internal::traits<SimplicialLLT> Traits;
+    typedef typename Traits::MatrixL  MatrixL;
+    typedef typename Traits::MatrixU  MatrixU;
+public:
+    /** Default constructor */
+    SimplicialLLT() : Base() {}
+    /** Constructs and performs the LLT factorization of \a matrix */
+    explicit SimplicialLLT(const MatrixType& matrix)
+        : Base(matrix) {}
+
+    /** \returns an expression of the factor L */
+    inline const MatrixL matrixL() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial LLT not factorized");
+        return Traits::getL(Base::m_matrix);
+    }
+
+    /** \returns an expression of the factor U (= L^*) */
+    inline const MatrixU matrixU() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial LLT not factorized");
+        return Traits::getU(Base::m_matrix);
+    }
+    
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    SimplicialLLT& compute(const MatrixType& matrix)
+    {
+      Base::template compute<false>(matrix);
+      return *this;
+    }
+
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      *
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& a)
+    {
+      Base::analyzePattern(a, false);
+    }
+
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& a)
+    {
+      Base::template factorize<false>(a);
+    }
+
+    /** \returns the determinant of the underlying matrix from the current factorization */
+    Scalar determinant() const
+    {
+      Scalar detL = Base::m_matrix.diagonal().prod();
+      return numext::abs2(detL);
+    }
+};
+
+/** \ingroup SparseCholesky_Module
+  * \class SimplicialLDLT
+  * \brief A direct sparse LDLT Cholesky factorizations without square root.
+  *
+  * This class provides a LDL^T Cholesky factorizations without square root of sparse matrices that are
+  * selfadjoint and positive definite. The factorization allows for solving A.X = B where
+  * X and B can be either dense or sparse.
+  * 
+  * In order to reduce the fill-in, a symmetric permutation P is applied prior to the factorization
+  * such that the factorized matrix is P A P^-1.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  *               or Upper. Default is Lower.
+  * \tparam _Ordering The ordering method to use, either AMDOrdering<> or NaturalOrdering<>. Default is AMDOrdering<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa class SimplicialLLT, class AMDOrdering, class NaturalOrdering
+  */
+template<typename _MatrixType, int _UpLo, typename _Ordering>
+    class SimplicialLDLT : public SimplicialCholeskyBase<SimplicialLDLT<_MatrixType,_UpLo,_Ordering> >
+{
+public:
+    typedef _MatrixType MatrixType;
+    enum { UpLo = _UpLo };
+    typedef SimplicialCholeskyBase<SimplicialLDLT> Base;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> CholMatrixType;
+    typedef Matrix<Scalar,Dynamic,1> VectorType;
+    typedef internal::traits<SimplicialLDLT> Traits;
+    typedef typename Traits::MatrixL  MatrixL;
+    typedef typename Traits::MatrixU  MatrixU;
+public:
+    /** Default constructor */
+    SimplicialLDLT() : Base() {}
+
+    /** Constructs and performs the LLT factorization of \a matrix */
+    explicit SimplicialLDLT(const MatrixType& matrix)
+        : Base(matrix) {}
+
+    /** \returns a vector expression of the diagonal D */
+    inline const VectorType vectorD() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial LDLT not factorized");
+        return Base::m_diag;
+    }
+    /** \returns an expression of the factor L */
+    inline const MatrixL matrixL() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial LDLT not factorized");
+        return Traits::getL(Base::m_matrix);
+    }
+
+    /** \returns an expression of the factor U (= L^*) */
+    inline const MatrixU matrixU() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial LDLT not factorized");
+        return Traits::getU(Base::m_matrix);
+    }
+
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    SimplicialLDLT& compute(const MatrixType& matrix)
+    {
+      Base::template compute<true>(matrix);
+      return *this;
+    }
+    
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      *
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& a)
+    {
+      Base::analyzePattern(a, true);
+    }
+
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& a)
+    {
+      Base::template factorize<true>(a);
+    }
+
+    /** \returns the determinant of the underlying matrix from the current factorization */
+    Scalar determinant() const
+    {
+      return Base::m_diag.prod();
+    }
+};
+
+/** \deprecated use SimplicialLDLT or class SimplicialLLT
+  * \ingroup SparseCholesky_Module
+  * \class SimplicialCholesky
+  *
+  * \sa class SimplicialLDLT, class SimplicialLLT
+  */
+template<typename _MatrixType, int _UpLo, typename _Ordering>
+    class SimplicialCholesky : public SimplicialCholeskyBase<SimplicialCholesky<_MatrixType,_UpLo,_Ordering> >
+{
+public:
+    typedef _MatrixType MatrixType;
+    enum { UpLo = _UpLo };
+    typedef SimplicialCholeskyBase<SimplicialCholesky> Base;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> CholMatrixType;
+    typedef Matrix<Scalar,Dynamic,1> VectorType;
+    typedef internal::traits<SimplicialCholesky> Traits;
+    typedef internal::traits<SimplicialLDLT<MatrixType,UpLo> > LDLTTraits;
+    typedef internal::traits<SimplicialLLT<MatrixType,UpLo>  > LLTTraits;
+  public:
+    SimplicialCholesky() : Base(), m_LDLT(true) {}
+
+    explicit SimplicialCholesky(const MatrixType& matrix)
+      : Base(), m_LDLT(true)
+    {
+      compute(matrix);
+    }
+
+    SimplicialCholesky& setMode(SimplicialCholeskyMode mode)
+    {
+      switch(mode)
+      {
+      case SimplicialCholeskyLLT:
+        m_LDLT = false;
+        break;
+      case SimplicialCholeskyLDLT:
+        m_LDLT = true;
+        break;
+      default:
+        break;
+      }
+
+      return *this;
+    }
+
+    inline const VectorType vectorD() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial Cholesky not factorized");
+        return Base::m_diag;
+    }
+    inline const CholMatrixType rawMatrix() const {
+        eigen_assert(Base::m_factorizationIsOk && "Simplicial Cholesky not factorized");
+        return Base::m_matrix;
+    }
+    
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    SimplicialCholesky& compute(const MatrixType& matrix)
+    {
+      if(m_LDLT)
+        Base::template compute<true>(matrix);
+      else
+        Base::template compute<false>(matrix);
+      return *this;
+    }
+
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      *
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& a)
+    {
+      Base::analyzePattern(a, m_LDLT);
+    }
+
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& a)
+    {
+      if(m_LDLT)
+        Base::template factorize<true>(a);
+      else
+        Base::template factorize<false>(a);
+    }
+
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const
+    {
+      eigen_assert(Base::m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or symbolic()/numeric()");
+      eigen_assert(Base::m_matrix.rows()==b.rows());
+
+      if(Base::m_info!=Success)
+        return;
+
+      if(Base::m_P.size()>0)
+        dest = Base::m_P * b;
+      else
+        dest = b;
+
+      if(Base::m_matrix.nonZeros()>0) // otherwise L==I
+      {
+        if(m_LDLT)
+          LDLTTraits::getL(Base::m_matrix).solveInPlace(dest);
+        else
+          LLTTraits::getL(Base::m_matrix).solveInPlace(dest);
+      }
+
+      if(Base::m_diag.size()>0)
+        dest = Base::m_diag.asDiagonal().inverse() * dest;
+
+      if (Base::m_matrix.nonZeros()>0) // otherwise I==I
+      {
+        if(m_LDLT)
+          LDLTTraits::getU(Base::m_matrix).solveInPlace(dest);
+        else
+          LLTTraits::getU(Base::m_matrix).solveInPlace(dest);
+      }
+
+      if(Base::m_P.size()>0)
+        dest = Base::m_Pinv * dest;
+    }
+    
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const SparseMatrixBase<Rhs> &b, SparseMatrixBase<Dest> &dest) const
+    {
+      internal::solve_sparse_through_dense_panels(*this, b, dest);
+    }
+    
+    Scalar determinant() const
+    {
+      if(m_LDLT)
+      {
+        return Base::m_diag.prod();
+      }
+      else
+      {
+        Scalar detL = Diagonal<const CholMatrixType>(Base::m_matrix).prod();
+        return numext::abs2(detL);
+      }
+    }
+    
+  protected:
+    bool m_LDLT;
+};
+
+template<typename Derived>
+void SimplicialCholeskyBase<Derived>::ordering(const MatrixType& a, ConstCholMatrixPtr &pmat, CholMatrixType& ap)
+{
+  eigen_assert(a.rows()==a.cols());
+  const Index size = a.rows();
+  pmat = &ap;
+  // Note that ordering methods compute the inverse permutation
+  if(!internal::is_same<OrderingType,NaturalOrdering<Index> >::value)
+  {
+    {
+      CholMatrixType C;
+      C = a.template selfadjointView<UpLo>();
+      
+      OrderingType ordering;
+      ordering(C,m_Pinv);
+    }
+
+    if(m_Pinv.size()>0) m_P = m_Pinv.inverse();
+    else                m_P.resize(0);
+    
+    ap.resize(size,size);
+    ap.template selfadjointView<Upper>() = a.template selfadjointView<UpLo>().twistedBy(m_P);
+  }
+  else
+  {
+    m_Pinv.resize(0);
+    m_P.resize(0);
+    if(int(UpLo)==int(Lower) || MatrixType::IsRowMajor)
+    {
+      // we have to transpose the lower part to to the upper one
+      ap.resize(size,size);
+      ap.template selfadjointView<Upper>() = a.template selfadjointView<UpLo>();
+    }
+    else
+      internal::simplicial_cholesky_grab_input<CholMatrixType,MatrixType>::run(a, pmat, ap);
+  }  
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SIMPLICIAL_CHOLESKY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
new file mode 100644
index 0000000..31e0699
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCholesky/SimplicialCholesky_impl.h
@@ -0,0 +1,199 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+
+/*
+
+NOTE: thes functions vave been adapted from the LDL library:
+
+LDL Copyright (c) 2005 by Timothy A. Davis.  All Rights Reserved.
+
+LDL License:
+
+    Your use or distribution of LDL or any modified version of
+    LDL implies that you agree to this License.
+
+    This library is free software; you can redistribute it and/or
+    modify it under the terms of the GNU Lesser General Public
+    License as published by the Free Software Foundation; either
+    version 2.1 of the License, or (at your option) any later version.
+
+    This library is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+    Lesser General Public License for more details.
+
+    You should have received a copy of the GNU Lesser General Public
+    License along with this library; if not, write to the Free Software
+    Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA  02110-1301
+    USA
+
+    Permission is hereby granted to use or copy this program under the
+    terms of the GNU LGPL, provided that the Copyright, this License,
+    and the Availability of the original version is retained on all copies.
+    User documentation of any code that uses this code or any modified
+    version of this code must cite the Copyright, this License, the
+    Availability note, and "Used by permission." Permission to modify
+    the code and to distribute modified code is granted, provided the
+    Copyright, this License, and the Availability note are retained,
+    and a notice that the code was modified is included.
+ */
+
+#include "../Core/util/NonMPL2.h"
+
+#ifndef EIGEN_SIMPLICIAL_CHOLESKY_IMPL_H
+#define EIGEN_SIMPLICIAL_CHOLESKY_IMPL_H
+
+namespace Eigen {
+
+template<typename Derived>
+void SimplicialCholeskyBase<Derived>::analyzePattern_preordered(const CholMatrixType& ap, bool doLDLT)
+{
+  const StorageIndex size = StorageIndex(ap.rows());
+  m_matrix.resize(size, size);
+  m_parent.resize(size);
+  m_nonZerosPerCol.resize(size);
+
+  ei_declare_aligned_stack_constructed_variable(StorageIndex, tags, size, 0);
+
+  for(StorageIndex k = 0; k < size; ++k)
+  {
+    /* L(k,:) pattern: all nodes reachable in etree from nz in A(0:k-1,k) */
+    m_parent[k] = -1;             /* parent of k is not yet known */
+    tags[k] = k;                  /* mark node k as visited */
+    m_nonZerosPerCol[k] = 0;      /* count of nonzeros in column k of L */
+    for(typename CholMatrixType::InnerIterator it(ap,k); it; ++it)
+    {
+      StorageIndex i = it.index();
+      if(i < k)
+      {
+        /* follow path from i to root of etree, stop at flagged node */
+        for(; tags[i] != k; i = m_parent[i])
+        {
+          /* find parent of i if not yet determined */
+          if (m_parent[i] == -1)
+            m_parent[i] = k;
+          m_nonZerosPerCol[i]++;        /* L (k,i) is nonzero */
+          tags[i] = k;                  /* mark i as visited */
+        }
+      }
+    }
+  }
+
+  /* construct Lp index array from m_nonZerosPerCol column counts */
+  StorageIndex* Lp = m_matrix.outerIndexPtr();
+  Lp[0] = 0;
+  for(StorageIndex k = 0; k < size; ++k)
+    Lp[k+1] = Lp[k] + m_nonZerosPerCol[k] + (doLDLT ? 0 : 1);
+
+  m_matrix.resizeNonZeros(Lp[size]);
+
+  m_isInitialized     = true;
+  m_info              = Success;
+  m_analysisIsOk      = true;
+  m_factorizationIsOk = false;
+}
+
+
+template<typename Derived>
+template<bool DoLDLT>
+void SimplicialCholeskyBase<Derived>::factorize_preordered(const CholMatrixType& ap)
+{
+  using std::sqrt;
+
+  eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+  eigen_assert(ap.rows()==ap.cols());
+  eigen_assert(m_parent.size()==ap.rows());
+  eigen_assert(m_nonZerosPerCol.size()==ap.rows());
+
+  const StorageIndex size = StorageIndex(ap.rows());
+  const StorageIndex* Lp = m_matrix.outerIndexPtr();
+  StorageIndex* Li = m_matrix.innerIndexPtr();
+  Scalar* Lx = m_matrix.valuePtr();
+
+  ei_declare_aligned_stack_constructed_variable(Scalar, y, size, 0);
+  ei_declare_aligned_stack_constructed_variable(StorageIndex,  pattern, size, 0);
+  ei_declare_aligned_stack_constructed_variable(StorageIndex,  tags, size, 0);
+
+  bool ok = true;
+  m_diag.resize(DoLDLT ? size : 0);
+
+  for(StorageIndex k = 0; k < size; ++k)
+  {
+    // compute nonzero pattern of kth row of L, in topological order
+    y[k] = 0.0;                     // Y(0:k) is now all zero
+    StorageIndex top = size;               // stack for pattern is empty
+    tags[k] = k;                    // mark node k as visited
+    m_nonZerosPerCol[k] = 0;        // count of nonzeros in column k of L
+    for(typename CholMatrixType::InnerIterator it(ap,k); it; ++it)
+    {
+      StorageIndex i = it.index();
+      if(i <= k)
+      {
+        y[i] += numext::conj(it.value());            /* scatter A(i,k) into Y (sum duplicates) */
+        Index len;
+        for(len = 0; tags[i] != k; i = m_parent[i])
+        {
+          pattern[len++] = i;     /* L(k,i) is nonzero */
+          tags[i] = k;            /* mark i as visited */
+        }
+        while(len > 0)
+          pattern[--top] = pattern[--len];
+      }
+    }
+
+    /* compute numerical values kth row of L (a sparse triangular solve) */
+
+    RealScalar d = numext::real(y[k]) * m_shiftScale + m_shiftOffset;    // get D(k,k), apply the shift function, and clear Y(k)
+    y[k] = 0.0;
+    for(; top < size; ++top)
+    {
+      Index i = pattern[top];       /* pattern[top:n-1] is pattern of L(:,k) */
+      Scalar yi = y[i];             /* get and clear Y(i) */
+      y[i] = 0.0;
+
+      /* the nonzero entry L(k,i) */
+      Scalar l_ki;
+      if(DoLDLT)
+        l_ki = yi / m_diag[i];
+      else
+        yi = l_ki = yi / Lx[Lp[i]];
+
+      Index p2 = Lp[i] + m_nonZerosPerCol[i];
+      Index p;
+      for(p = Lp[i] + (DoLDLT ? 0 : 1); p < p2; ++p)
+        y[Li[p]] -= numext::conj(Lx[p]) * yi;
+      d -= numext::real(l_ki * numext::conj(yi));
+      Li[p] = k;                          /* store L(k,i) in column form of L */
+      Lx[p] = l_ki;
+      ++m_nonZerosPerCol[i];              /* increment count of nonzeros in col i */
+    }
+    if(DoLDLT)
+    {
+      m_diag[k] = d;
+      if(d == RealScalar(0))
+      {
+        ok = false;                         /* failure, D(k,k) is zero */
+        break;
+      }
+    }
+    else
+    {
+      Index p = Lp[k] + m_nonZerosPerCol[k]++;
+      Li[p] = k ;                /* store L(k,k) = sqrt (d) in column k */
+      if(d <= RealScalar(0)) {
+        ok = false;              /* failure, matrix is not positive definite */
+        break;
+      }
+      Lx[p] = sqrt(d) ;
+    }
+  }
+
+  m_info = ok ? Success : NumericalIssue;
+  m_factorizationIsOk = true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SIMPLICIAL_CHOLESKY_IMPL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/AmbiVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/AmbiVector.h
new file mode 100644
index 0000000..8a5cc91
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/AmbiVector.h
@@ -0,0 +1,377 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_AMBIVECTOR_H
+#define EIGEN_AMBIVECTOR_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal
+  * Hybrid sparse/dense vector class designed for intensive read-write operations.
+  *
+  * See BasicSparseLLT and SparseProduct for usage examples.
+  */
+template<typename _Scalar, typename _StorageIndex>
+class AmbiVector
+{
+  public:
+    typedef _Scalar Scalar;
+    typedef _StorageIndex StorageIndex;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    explicit AmbiVector(Index size)
+      : m_buffer(0), m_zero(0), m_size(0), m_allocatedSize(0), m_allocatedElements(0), m_mode(-1)
+    {
+      resize(size);
+    }
+
+    void init(double estimatedDensity);
+    void init(int mode);
+
+    Index nonZeros() const;
+
+    /** Specifies a sub-vector to work on */
+    void setBounds(Index start, Index end) { m_start = convert_index(start); m_end = convert_index(end); }
+
+    void setZero();
+
+    void restart();
+    Scalar& coeffRef(Index i);
+    Scalar& coeff(Index i);
+
+    class Iterator;
+
+    ~AmbiVector() { delete[] m_buffer; }
+
+    void resize(Index size)
+    {
+      if (m_allocatedSize < size)
+        reallocate(size);
+      m_size = convert_index(size);
+    }
+
+    StorageIndex size() const { return m_size; }
+
+  protected:
+    StorageIndex convert_index(Index idx)
+    {
+      return internal::convert_index<StorageIndex>(idx);
+    }
+
+    void reallocate(Index size)
+    {
+      // if the size of the matrix is not too large, let's allocate a bit more than needed such
+      // that we can handle dense vector even in sparse mode.
+      delete[] m_buffer;
+      if (size<1000)
+      {
+        Index allocSize = (size * sizeof(ListEl) + sizeof(Scalar) - 1)/sizeof(Scalar);
+        m_allocatedElements = convert_index((allocSize*sizeof(Scalar))/sizeof(ListEl));
+        m_buffer = new Scalar[allocSize];
+      }
+      else
+      {
+        m_allocatedElements = convert_index((size*sizeof(Scalar))/sizeof(ListEl));
+        m_buffer = new Scalar[size];
+      }
+      m_size = convert_index(size);
+      m_start = 0;
+      m_end = m_size;
+    }
+
+    void reallocateSparse()
+    {
+      Index copyElements = m_allocatedElements;
+      m_allocatedElements = (std::min)(StorageIndex(m_allocatedElements*1.5),m_size);
+      Index allocSize = m_allocatedElements * sizeof(ListEl);
+      allocSize = (allocSize + sizeof(Scalar) - 1)/sizeof(Scalar);
+      Scalar* newBuffer = new Scalar[allocSize];
+      memcpy(newBuffer,  m_buffer,  copyElements * sizeof(ListEl));
+      delete[] m_buffer;
+      m_buffer = newBuffer;
+    }
+
+  protected:
+    // element type of the linked list
+    struct ListEl
+    {
+      StorageIndex next;
+      StorageIndex index;
+      Scalar value;
+    };
+
+    // used to store data in both mode
+    Scalar* m_buffer;
+    Scalar m_zero;
+    StorageIndex m_size;
+    StorageIndex m_start;
+    StorageIndex m_end;
+    StorageIndex m_allocatedSize;
+    StorageIndex m_allocatedElements;
+    StorageIndex m_mode;
+
+    // linked list mode
+    StorageIndex m_llStart;
+    StorageIndex m_llCurrent;
+    StorageIndex m_llSize;
+};
+
+/** \returns the number of non zeros in the current sub vector */
+template<typename _Scalar,typename _StorageIndex>
+Index AmbiVector<_Scalar,_StorageIndex>::nonZeros() const
+{
+  if (m_mode==IsSparse)
+    return m_llSize;
+  else
+    return m_end - m_start;
+}
+
+template<typename _Scalar,typename _StorageIndex>
+void AmbiVector<_Scalar,_StorageIndex>::init(double estimatedDensity)
+{
+  if (estimatedDensity>0.1)
+    init(IsDense);
+  else
+    init(IsSparse);
+}
+
+template<typename _Scalar,typename _StorageIndex>
+void AmbiVector<_Scalar,_StorageIndex>::init(int mode)
+{
+  m_mode = mode;
+  if (m_mode==IsSparse)
+  {
+    m_llSize = 0;
+    m_llStart = -1;
+  }
+}
+
+/** Must be called whenever we might perform a write access
+  * with an index smaller than the previous one.
+  *
+  * Don't worry, this function is extremely cheap.
+  */
+template<typename _Scalar,typename _StorageIndex>
+void AmbiVector<_Scalar,_StorageIndex>::restart()
+{
+  m_llCurrent = m_llStart;
+}
+
+/** Set all coefficients of current subvector to zero */
+template<typename _Scalar,typename _StorageIndex>
+void AmbiVector<_Scalar,_StorageIndex>::setZero()
+{
+  if (m_mode==IsDense)
+  {
+    for (Index i=m_start; i<m_end; ++i)
+      m_buffer[i] = Scalar(0);
+  }
+  else
+  {
+    eigen_assert(m_mode==IsSparse);
+    m_llSize = 0;
+    m_llStart = -1;
+  }
+}
+
+template<typename _Scalar,typename _StorageIndex>
+_Scalar& AmbiVector<_Scalar,_StorageIndex>::coeffRef(Index i)
+{
+  if (m_mode==IsDense)
+    return m_buffer[i];
+  else
+  {
+    ListEl* EIGEN_RESTRICT llElements = reinterpret_cast<ListEl*>(m_buffer);
+    // TODO factorize the following code to reduce code generation
+    eigen_assert(m_mode==IsSparse);
+    if (m_llSize==0)
+    {
+      // this is the first element
+      m_llStart = 0;
+      m_llCurrent = 0;
+      ++m_llSize;
+      llElements[0].value = Scalar(0);
+      llElements[0].index = convert_index(i);
+      llElements[0].next = -1;
+      return llElements[0].value;
+    }
+    else if (i<llElements[m_llStart].index)
+    {
+      // this is going to be the new first element of the list
+      ListEl& el = llElements[m_llSize];
+      el.value = Scalar(0);
+      el.index = convert_index(i);
+      el.next = m_llStart;
+      m_llStart = m_llSize;
+      ++m_llSize;
+      m_llCurrent = m_llStart;
+      return el.value;
+    }
+    else
+    {
+      StorageIndex nextel = llElements[m_llCurrent].next;
+      eigen_assert(i>=llElements[m_llCurrent].index && "you must call restart() before inserting an element with lower or equal index");
+      while (nextel >= 0 && llElements[nextel].index<=i)
+      {
+        m_llCurrent = nextel;
+        nextel = llElements[nextel].next;
+      }
+
+      if (llElements[m_llCurrent].index==i)
+      {
+        // the coefficient already exists and we found it !
+        return llElements[m_llCurrent].value;
+      }
+      else
+      {
+        if (m_llSize>=m_allocatedElements)
+        {
+          reallocateSparse();
+          llElements = reinterpret_cast<ListEl*>(m_buffer);
+        }
+        eigen_internal_assert(m_llSize<m_allocatedElements && "internal error: overflow in sparse mode");
+        // let's insert a new coefficient
+        ListEl& el = llElements[m_llSize];
+        el.value = Scalar(0);
+        el.index = convert_index(i);
+        el.next = llElements[m_llCurrent].next;
+        llElements[m_llCurrent].next = m_llSize;
+        ++m_llSize;
+        return el.value;
+      }
+    }
+  }
+}
+
+template<typename _Scalar,typename _StorageIndex>
+_Scalar& AmbiVector<_Scalar,_StorageIndex>::coeff(Index i)
+{
+  if (m_mode==IsDense)
+    return m_buffer[i];
+  else
+  {
+    ListEl* EIGEN_RESTRICT llElements = reinterpret_cast<ListEl*>(m_buffer);
+    eigen_assert(m_mode==IsSparse);
+    if ((m_llSize==0) || (i<llElements[m_llStart].index))
+    {
+      return m_zero;
+    }
+    else
+    {
+      Index elid = m_llStart;
+      while (elid >= 0 && llElements[elid].index<i)
+        elid = llElements[elid].next;
+
+      if (llElements[elid].index==i)
+        return llElements[m_llCurrent].value;
+      else
+        return m_zero;
+    }
+  }
+}
+
+/** Iterator over the nonzero coefficients */
+template<typename _Scalar,typename _StorageIndex>
+class AmbiVector<_Scalar,_StorageIndex>::Iterator
+{
+  public:
+    typedef _Scalar Scalar;
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    /** Default constructor
+      * \param vec the vector on which we iterate
+      * \param epsilon the minimal value used to prune zero coefficients.
+      * In practice, all coefficients having a magnitude smaller than \a epsilon
+      * are skipped.
+      */
+    explicit Iterator(const AmbiVector& vec, const RealScalar& epsilon = 0)
+      : m_vector(vec)
+    {
+      using std::abs;
+      m_epsilon = epsilon;
+      m_isDense = m_vector.m_mode==IsDense;
+      if (m_isDense)
+      {
+        m_currentEl = 0;   // this is to avoid a compilation warning
+        m_cachedValue = 0; // this is to avoid a compilation warning
+        m_cachedIndex = m_vector.m_start-1;
+        ++(*this);
+      }
+      else
+      {
+        ListEl* EIGEN_RESTRICT llElements = reinterpret_cast<ListEl*>(m_vector.m_buffer);
+        m_currentEl = m_vector.m_llStart;
+        while (m_currentEl>=0 && abs(llElements[m_currentEl].value)<=m_epsilon)
+          m_currentEl = llElements[m_currentEl].next;
+        if (m_currentEl<0)
+        {
+          m_cachedValue = 0; // this is to avoid a compilation warning
+          m_cachedIndex = -1;
+        }
+        else
+        {
+          m_cachedIndex = llElements[m_currentEl].index;
+          m_cachedValue = llElements[m_currentEl].value;
+        }
+      }
+    }
+
+    StorageIndex index() const { return m_cachedIndex; }
+    Scalar value() const { return m_cachedValue; }
+
+    operator bool() const { return m_cachedIndex>=0; }
+
+    Iterator& operator++()
+    {
+      using std::abs;
+      if (m_isDense)
+      {
+        do {
+          ++m_cachedIndex;
+        } while (m_cachedIndex<m_vector.m_end && abs(m_vector.m_buffer[m_cachedIndex])<=m_epsilon);
+        if (m_cachedIndex<m_vector.m_end)
+          m_cachedValue = m_vector.m_buffer[m_cachedIndex];
+        else
+          m_cachedIndex=-1;
+      }
+      else
+      {
+        ListEl* EIGEN_RESTRICT llElements = reinterpret_cast<ListEl*>(m_vector.m_buffer);
+        do {
+          m_currentEl = llElements[m_currentEl].next;
+        } while (m_currentEl>=0 && abs(llElements[m_currentEl].value)<=m_epsilon);
+        if (m_currentEl<0)
+        {
+          m_cachedIndex = -1;
+        }
+        else
+        {
+          m_cachedIndex = llElements[m_currentEl].index;
+          m_cachedValue = llElements[m_currentEl].value;
+        }
+      }
+      return *this;
+    }
+
+  protected:
+    const AmbiVector& m_vector; // the target vector
+    StorageIndex m_currentEl;   // the current element in sparse/linked-list mode
+    RealScalar m_epsilon;       // epsilon used to prune zero coefficients
+    StorageIndex m_cachedIndex; // current coordinate
+    Scalar m_cachedValue;       // current value
+    bool m_isDense;             // mode of the vector
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_AMBIVECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/CompressedStorage.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/CompressedStorage.h
new file mode 100644
index 0000000..d89fa0d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/CompressedStorage.h
@@ -0,0 +1,258 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_COMPRESSED_STORAGE_H
+#define EIGEN_COMPRESSED_STORAGE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \internal
+  * Stores a sparse set of values as a list of values and a list of indices.
+  *
+  */
+template<typename _Scalar,typename _StorageIndex>
+class CompressedStorage
+{
+  public:
+
+    typedef _Scalar Scalar;
+    typedef _StorageIndex StorageIndex;
+
+  protected:
+
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+  public:
+
+    CompressedStorage()
+      : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0)
+    {}
+
+    explicit CompressedStorage(Index size)
+      : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0)
+    {
+      resize(size);
+    }
+
+    CompressedStorage(const CompressedStorage& other)
+      : m_values(0), m_indices(0), m_size(0), m_allocatedSize(0)
+    {
+      *this = other;
+    }
+
+    CompressedStorage& operator=(const CompressedStorage& other)
+    {
+      resize(other.size());
+      if(other.size()>0)
+      {
+        internal::smart_copy(other.m_values,  other.m_values  + m_size, m_values);
+        internal::smart_copy(other.m_indices, other.m_indices + m_size, m_indices);
+      }
+      return *this;
+    }
+
+    void swap(CompressedStorage& other)
+    {
+      std::swap(m_values, other.m_values);
+      std::swap(m_indices, other.m_indices);
+      std::swap(m_size, other.m_size);
+      std::swap(m_allocatedSize, other.m_allocatedSize);
+    }
+
+    ~CompressedStorage()
+    {
+      delete[] m_values;
+      delete[] m_indices;
+    }
+
+    void reserve(Index size)
+    {
+      Index newAllocatedSize = m_size + size;
+      if (newAllocatedSize > m_allocatedSize)
+        reallocate(newAllocatedSize);
+    }
+
+    void squeeze()
+    {
+      if (m_allocatedSize>m_size)
+        reallocate(m_size);
+    }
+
+    void resize(Index size, double reserveSizeFactor = 0)
+    {
+      if (m_allocatedSize<size)
+      {
+        Index realloc_size = (std::min<Index>)(NumTraits<StorageIndex>::highest(),  size + Index(reserveSizeFactor*double(size)));
+        if(realloc_size<size)
+          internal::throw_std_bad_alloc();
+        reallocate(realloc_size);
+      }
+      m_size = size;
+    }
+
+    void append(const Scalar& v, Index i)
+    {
+      Index id = m_size;
+      resize(m_size+1, 1);
+      m_values[id] = v;
+      m_indices[id] = internal::convert_index<StorageIndex>(i);
+    }
+
+    inline Index size() const { return m_size; }
+    inline Index allocatedSize() const { return m_allocatedSize; }
+    inline void clear() { m_size = 0; }
+
+    const Scalar* valuePtr() const { return m_values; }
+    Scalar* valuePtr() { return m_values; }
+    const StorageIndex* indexPtr() const { return m_indices; }
+    StorageIndex* indexPtr() { return m_indices; }
+
+    inline Scalar& value(Index i) { eigen_internal_assert(m_values!=0); return m_values[i]; }
+    inline const Scalar& value(Index i) const { eigen_internal_assert(m_values!=0); return m_values[i]; }
+
+    inline StorageIndex& index(Index i) { eigen_internal_assert(m_indices!=0); return m_indices[i]; }
+    inline const StorageIndex& index(Index i) const { eigen_internal_assert(m_indices!=0); return m_indices[i]; }
+
+    /** \returns the largest \c k such that for all \c j in [0,k) index[\c j]\<\a key */
+    inline Index searchLowerIndex(Index key) const
+    {
+      return searchLowerIndex(0, m_size, key);
+    }
+
+    /** \returns the largest \c k in [start,end) such that for all \c j in [start,k) index[\c j]\<\a key */
+    inline Index searchLowerIndex(Index start, Index end, Index key) const
+    {
+      while(end>start)
+      {
+        Index mid = (end+start)>>1;
+        if (m_indices[mid]<key)
+          start = mid+1;
+        else
+          end = mid;
+      }
+      return start;
+    }
+
+    /** \returns the stored value at index \a key
+      * If the value does not exist, then the value \a defaultValue is returned without any insertion. */
+    inline Scalar at(Index key, const Scalar& defaultValue = Scalar(0)) const
+    {
+      if (m_size==0)
+        return defaultValue;
+      else if (key==m_indices[m_size-1])
+        return m_values[m_size-1];
+      // ^^  optimization: let's first check if it is the last coefficient
+      // (very common in high level algorithms)
+      const Index id = searchLowerIndex(0,m_size-1,key);
+      return ((id<m_size) && (m_indices[id]==key)) ? m_values[id] : defaultValue;
+    }
+
+    /** Like at(), but the search is performed in the range [start,end) */
+    inline Scalar atInRange(Index start, Index end, Index key, const Scalar &defaultValue = Scalar(0)) const
+    {
+      if (start>=end)
+        return defaultValue;
+      else if (end>start && key==m_indices[end-1])
+        return m_values[end-1];
+      // ^^  optimization: let's first check if it is the last coefficient
+      // (very common in high level algorithms)
+      const Index id = searchLowerIndex(start,end-1,key);
+      return ((id<end) && (m_indices[id]==key)) ? m_values[id] : defaultValue;
+    }
+
+    /** \returns a reference to the value at index \a key
+      * If the value does not exist, then the value \a defaultValue is inserted
+      * such that the keys are sorted. */
+    inline Scalar& atWithInsertion(Index key, const Scalar& defaultValue = Scalar(0))
+    {
+      Index id = searchLowerIndex(0,m_size,key);
+      if (id>=m_size || m_indices[id]!=key)
+      {
+        if (m_allocatedSize<m_size+1)
+        {
+          m_allocatedSize = 2*(m_size+1);
+          internal::scoped_array<Scalar> newValues(m_allocatedSize);
+          internal::scoped_array<StorageIndex> newIndices(m_allocatedSize);
+
+          // copy first chunk
+          internal::smart_copy(m_values,  m_values +id, newValues.ptr());
+          internal::smart_copy(m_indices, m_indices+id, newIndices.ptr());
+
+          // copy the rest
+          if(m_size>id)
+          {
+            internal::smart_copy(m_values +id,  m_values +m_size, newValues.ptr() +id+1);
+            internal::smart_copy(m_indices+id,  m_indices+m_size, newIndices.ptr()+id+1);
+          }
+          std::swap(m_values,newValues.ptr());
+          std::swap(m_indices,newIndices.ptr());
+        }
+        else if(m_size>id)
+        {
+          internal::smart_memmove(m_values +id, m_values +m_size, m_values +id+1);
+          internal::smart_memmove(m_indices+id, m_indices+m_size, m_indices+id+1);
+        }
+        m_size++;
+        m_indices[id] = internal::convert_index<StorageIndex>(key);
+        m_values[id] = defaultValue;
+      }
+      return m_values[id];
+    }
+
+    void prune(const Scalar& reference, const RealScalar& epsilon = NumTraits<RealScalar>::dummy_precision())
+    {
+      Index k = 0;
+      Index n = size();
+      for (Index i=0; i<n; ++i)
+      {
+        if (!internal::isMuchSmallerThan(value(i), reference, epsilon))
+        {
+          value(k) = value(i);
+          index(k) = index(i);
+          ++k;
+        }
+      }
+      resize(k,0);
+    }
+
+  protected:
+
+    inline void reallocate(Index size)
+    {
+      #ifdef EIGEN_SPARSE_COMPRESSED_STORAGE_REALLOCATE_PLUGIN
+        EIGEN_SPARSE_COMPRESSED_STORAGE_REALLOCATE_PLUGIN
+      #endif
+      eigen_internal_assert(size!=m_allocatedSize);
+      internal::scoped_array<Scalar> newValues(size);
+      internal::scoped_array<StorageIndex> newIndices(size);
+      Index copySize = (std::min)(size, m_size);
+      if (copySize>0) {
+        internal::smart_copy(m_values, m_values+copySize, newValues.ptr());
+        internal::smart_copy(m_indices, m_indices+copySize, newIndices.ptr());
+      }
+      std::swap(m_values,newValues.ptr());
+      std::swap(m_indices,newIndices.ptr());
+      m_allocatedSize = size;
+    }
+
+  protected:
+    Scalar* m_values;
+    StorageIndex* m_indices;
+    Index m_size;
+    Index m_allocatedSize;
+
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_COMPRESSED_STORAGE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
new file mode 100644
index 0000000..492eb0a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/ConservativeSparseSparseProduct.h
@@ -0,0 +1,345 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_CONSERVATIVESPARSESPARSEPRODUCT_H
+#define EIGEN_CONSERVATIVESPARSESPARSEPRODUCT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, typename ResultType>
+static void conservative_sparse_sparse_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, bool sortedInsertion = false)
+{
+  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+
+  // make sure to call innerSize/outerSize since we fake the storage order.
+  Index rows = lhs.innerSize();
+  Index cols = rhs.outerSize();
+  eigen_assert(lhs.outerSize() == rhs.innerSize());
+  
+  ei_declare_aligned_stack_constructed_variable(bool,   mask,     rows, 0);
+  ei_declare_aligned_stack_constructed_variable(Scalar, values,   rows, 0);
+  ei_declare_aligned_stack_constructed_variable(Index,  indices,  rows, 0);
+  
+  std::memset(mask,0,sizeof(bool)*rows);
+
+  evaluator<Lhs> lhsEval(lhs);
+  evaluator<Rhs> rhsEval(rhs);
+  
+  // estimate the number of non zero entries
+  // given a rhs column containing Y non zeros, we assume that the respective Y columns
+  // of the lhs differs in average of one non zeros, thus the number of non zeros for
+  // the product of a rhs column with the lhs is X+Y where X is the average number of non zero
+  // per column of the lhs.
+  // Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs)
+  Index estimated_nnz_prod = lhsEval.nonZerosEstimate() + rhsEval.nonZerosEstimate();
+
+  res.setZero();
+  res.reserve(Index(estimated_nnz_prod));
+  // we compute each column of the result, one after the other
+  for (Index j=0; j<cols; ++j)
+  {
+
+    res.startVec(j);
+    Index nnz = 0;
+    for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
+    {
+      Scalar y = rhsIt.value();
+      Index k = rhsIt.index();
+      for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
+      {
+        Index i = lhsIt.index();
+        Scalar x = lhsIt.value();
+        if(!mask[i])
+        {
+          mask[i] = true;
+          values[i] = x * y;
+          indices[nnz] = i;
+          ++nnz;
+        }
+        else
+          values[i] += x * y;
+      }
+    }
+    if(!sortedInsertion)
+    {
+      // unordered insertion
+      for(Index k=0; k<nnz; ++k)
+      {
+        Index i = indices[k];
+        res.insertBackByOuterInnerUnordered(j,i) = values[i];
+        mask[i] = false;
+      }
+    }
+    else
+    {
+      // alternative ordered insertion code:
+      const Index t200 = rows/11; // 11 == (log2(200)*1.39)
+      const Index t = (rows*100)/139;
+
+      // FIXME reserve nnz non zeros
+      // FIXME implement faster sorting algorithms for very small nnz
+      // if the result is sparse enough => use a quick sort
+      // otherwise => loop through the entire vector
+      // In order to avoid to perform an expensive log2 when the
+      // result is clearly very sparse we use a linear bound up to 200.
+      if((nnz<200 && nnz<t200) || nnz * numext::log2(int(nnz)) < t)
+      {
+        if(nnz>1) std::sort(indices,indices+nnz);
+        for(Index k=0; k<nnz; ++k)
+        {
+          Index i = indices[k];
+          res.insertBackByOuterInner(j,i) = values[i];
+          mask[i] = false;
+        }
+      }
+      else
+      {
+        // dense path
+        for(Index i=0; i<rows; ++i)
+        {
+          if(mask[i])
+          {
+            mask[i] = false;
+            res.insertBackByOuterInner(j,i) = values[i];
+          }
+        }
+      }
+    }
+  }
+  res.finalize();
+}
+
+
+} // end namespace internal
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, typename ResultType,
+  int LhsStorageOrder = (traits<Lhs>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+  int RhsStorageOrder = (traits<Rhs>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+  int ResStorageOrder = (traits<ResultType>::Flags&RowMajorBit) ? RowMajor : ColMajor>
+struct conservative_sparse_sparse_product_selector;
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,ColMajor>
+{
+  typedef typename remove_all<Lhs>::type LhsCleaned;
+  typedef typename LhsCleaned::Scalar Scalar;
+
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrixAux;
+    typedef typename sparse_eval<ColMajorMatrixAux,ResultType::RowsAtCompileTime,ResultType::ColsAtCompileTime,ColMajorMatrixAux::Flags>::type ColMajorMatrix;
+    
+    // If the result is tall and thin (in the extreme case a column vector)
+    // then it is faster to sort the coefficients inplace instead of transposing twice.
+    // FIXME, the following heuristic is probably not very good.
+    if(lhs.rows()>rhs.cols())
+    {
+      ColMajorMatrix resCol(lhs.rows(),rhs.cols());
+      // perform sorted insertion
+      internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrix>(lhs, rhs, resCol, true);
+      res = resCol.markAsRValue();
+    }
+    else
+    {
+      ColMajorMatrixAux resCol(lhs.rows(),rhs.cols());
+      // ressort to transpose to sort the entries
+      internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrixAux>(lhs, rhs, resCol, false);
+      RowMajorMatrix resRow(resCol);
+      res = resRow.markAsRValue();
+    }
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,ColMajor,ColMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+     typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
+     RowMajorMatrix rhsRow = rhs;
+     RowMajorMatrix resRow(lhs.rows(), rhs.cols());
+     internal::conservative_sparse_sparse_product_impl<RowMajorMatrix,Lhs,RowMajorMatrix>(rhsRow, lhs, resRow);
+     res = resRow;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,RowMajor,ColMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
+    RowMajorMatrix lhsRow = lhs;
+    RowMajorMatrix resRow(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Rhs,RowMajorMatrix,RowMajorMatrix>(rhs, lhsRow, resRow);
+    res = resRow;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,RowMajor,ColMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
+    RowMajorMatrix resRow(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Rhs,Lhs,RowMajorMatrix>(rhs, lhs, resRow);
+    res = resRow;
+  }
+};
+
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,RowMajor>
+{
+  typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
+
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    ColMajorMatrix resCol(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Lhs,Rhs,ColMajorMatrix>(lhs, rhs, resCol);
+    res = resCol;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,ColMajor,RowMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    ColMajorMatrix lhsCol = lhs;
+    ColMajorMatrix resCol(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<ColMajorMatrix,Rhs,ColMajorMatrix>(lhsCol, rhs, resCol);
+    res = resCol;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,ColMajor,RowMajor,RowMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    ColMajorMatrix rhsCol = rhs;
+    ColMajorMatrix resCol(lhs.rows(), rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Lhs,ColMajorMatrix,ColMajorMatrix>(lhs, rhsCol, resCol);
+    res = resCol;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct conservative_sparse_sparse_product_selector<Lhs,Rhs,ResultType,RowMajor,RowMajor,RowMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename ResultType::StorageIndex> RowMajorMatrix;
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    RowMajorMatrix resRow(lhs.rows(),rhs.cols());
+    internal::conservative_sparse_sparse_product_impl<Rhs,Lhs,RowMajorMatrix>(rhs, lhs, resRow);
+    // sort the non zeros:
+    ColMajorMatrix resCol(resRow);
+    res = resCol;
+  }
+};
+
+} // end namespace internal
+
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, typename ResultType>
+static void sparse_sparse_to_dense_product_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+{
+  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+  Index cols = rhs.outerSize();
+  eigen_assert(lhs.outerSize() == rhs.innerSize());
+
+  evaluator<Lhs> lhsEval(lhs);
+  evaluator<Rhs> rhsEval(rhs);
+
+  for (Index j=0; j<cols; ++j)
+  {
+    for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
+    {
+      Scalar y = rhsIt.value();
+      Index k = rhsIt.index();
+      for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, k); lhsIt; ++lhsIt)
+      {
+        Index i = lhsIt.index();
+        Scalar x = lhsIt.value();
+        res.coeffRef(i,j) += x * y;
+      }
+    }
+  }
+}
+
+
+} // end namespace internal
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, typename ResultType,
+  int LhsStorageOrder = (traits<Lhs>::Flags&RowMajorBit) ? RowMajor : ColMajor,
+  int RhsStorageOrder = (traits<Rhs>::Flags&RowMajorBit) ? RowMajor : ColMajor>
+struct sparse_sparse_to_dense_product_selector;
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    internal::sparse_sparse_to_dense_product_impl<Lhs,Rhs,ResultType>(lhs, rhs, res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,RowMajor,ColMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    ColMajorMatrix lhsCol(lhs);
+    internal::sparse_sparse_to_dense_product_impl<ColMajorMatrix,Rhs,ResultType>(lhsCol, rhs, res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,ColMajor,RowMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> ColMajorMatrix;
+    ColMajorMatrix rhsCol(rhs);
+    internal::sparse_sparse_to_dense_product_impl<Lhs,ColMajorMatrix,ResultType>(lhs, rhsCol, res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_to_dense_product_selector<Lhs,Rhs,ResultType,RowMajor,RowMajor>
+{
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res)
+  {
+    Transpose<ResultType> trRes(res);
+    internal::sparse_sparse_to_dense_product_impl<Rhs,Lhs,Transpose<ResultType> >(rhs, lhs, trRes);
+  }
+};
+
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_CONSERVATIVESPARSESPARSEPRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/MappedSparseMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/MappedSparseMatrix.h
new file mode 100644
index 0000000..67718c8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/MappedSparseMatrix.h
@@ -0,0 +1,67 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MAPPED_SPARSEMATRIX_H
+#define EIGEN_MAPPED_SPARSEMATRIX_H
+
+namespace Eigen {
+
+/** \deprecated Use Map<SparseMatrix<> >
+  * \class MappedSparseMatrix
+  *
+  * \brief Sparse matrix
+  *
+  * \param _Scalar the scalar type, i.e. the type of the coefficients
+  *
+  * See http://www.netlib.org/linalg/html_templates/node91.html for details on the storage scheme.
+  *
+  */
+namespace internal {
+template<typename _Scalar, int _Flags, typename _StorageIndex>
+struct traits<MappedSparseMatrix<_Scalar, _Flags, _StorageIndex> > : traits<SparseMatrix<_Scalar, _Flags, _StorageIndex> >
+{};
+} // end namespace internal
+
+template<typename _Scalar, int _Flags, typename _StorageIndex>
+class MappedSparseMatrix
+  : public Map<SparseMatrix<_Scalar, _Flags, _StorageIndex> >
+{
+    typedef Map<SparseMatrix<_Scalar, _Flags, _StorageIndex> > Base;
+
+  public:
+    
+    typedef typename Base::StorageIndex StorageIndex;
+    typedef typename Base::Scalar Scalar;
+
+    inline MappedSparseMatrix(Index rows, Index cols, Index nnz, StorageIndex* outerIndexPtr, StorageIndex* innerIndexPtr, Scalar* valuePtr, StorageIndex* innerNonZeroPtr = 0)
+      : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZeroPtr)
+    {}
+
+    /** Empty destructor */
+    inline ~MappedSparseMatrix() {}
+};
+
+namespace internal {
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct evaluator<MappedSparseMatrix<_Scalar,_Options,_StorageIndex> >
+  : evaluator<SparseCompressedBase<MappedSparseMatrix<_Scalar,_Options,_StorageIndex> > >
+{
+  typedef MappedSparseMatrix<_Scalar,_Options,_StorageIndex> XprType;
+  typedef evaluator<SparseCompressedBase<XprType> > Base;
+  
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MAPPED_SPARSEMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseAssign.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseAssign.h
new file mode 100644
index 0000000..18352a8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseAssign.h
@@ -0,0 +1,216 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEASSIGN_H
+#define EIGEN_SPARSEASSIGN_H
+
+namespace Eigen { 
+
+template<typename Derived>    
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator=(const EigenBase<OtherDerived> &other)
+{
+  internal::call_assignment_no_alias(derived(), other.derived());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
+{
+  // TODO use the evaluator mechanism
+  other.evalTo(derived());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+inline Derived& SparseMatrixBase<Derived>::operator=(const SparseMatrixBase<OtherDerived>& other)
+{
+  // by default sparse evaluation do not alias, so we can safely bypass the generic call_assignment routine
+  internal::Assignment<Derived,OtherDerived,internal::assign_op<Scalar,typename OtherDerived::Scalar> >
+          ::run(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+inline Derived& SparseMatrixBase<Derived>::operator=(const Derived& other)
+{
+  internal::call_assignment_no_alias(derived(), other.derived());
+  return derived();
+}
+
+namespace internal {
+
+template<>
+struct storage_kind_to_evaluator_kind<Sparse> {
+  typedef IteratorBased Kind;
+};
+
+template<>
+struct storage_kind_to_shape<Sparse> {
+  typedef SparseShape Shape;
+};
+
+struct Sparse2Sparse {};
+struct Sparse2Dense  {};
+
+template<> struct AssignmentKind<SparseShape, SparseShape>           { typedef Sparse2Sparse Kind; };
+template<> struct AssignmentKind<SparseShape, SparseTriangularShape> { typedef Sparse2Sparse Kind; };
+template<> struct AssignmentKind<DenseShape,  SparseShape>           { typedef Sparse2Dense  Kind; };
+template<> struct AssignmentKind<DenseShape,  SparseTriangularShape> { typedef Sparse2Dense  Kind; };
+
+
+template<typename DstXprType, typename SrcXprType>
+void assign_sparse_to_sparse(DstXprType &dst, const SrcXprType &src)
+{
+  typedef typename DstXprType::Scalar Scalar;
+  typedef internal::evaluator<DstXprType> DstEvaluatorType;
+  typedef internal::evaluator<SrcXprType> SrcEvaluatorType;
+
+  SrcEvaluatorType srcEvaluator(src);
+
+  const bool transpose = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit);
+  const Index outerEvaluationSize = (SrcEvaluatorType::Flags&RowMajorBit) ? src.rows() : src.cols();
+  if ((!transpose) && src.isRValue())
+  {
+    // eval without temporary
+    dst.resize(src.rows(), src.cols());
+    dst.setZero();
+    dst.reserve((std::max)(src.rows(),src.cols())*2);
+    for (Index j=0; j<outerEvaluationSize; ++j)
+    {
+      dst.startVec(j);
+      for (typename SrcEvaluatorType::InnerIterator it(srcEvaluator, j); it; ++it)
+      {
+        Scalar v = it.value();
+        dst.insertBackByOuterInner(j,it.index()) = v;
+      }
+    }
+    dst.finalize();
+  }
+  else
+  {
+    // eval through a temporary
+    eigen_assert(( ((internal::traits<DstXprType>::SupportedAccessPatterns & OuterRandomAccessPattern)==OuterRandomAccessPattern) ||
+              (!((DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit)))) &&
+              "the transpose operation is supposed to be handled in SparseMatrix::operator=");
+
+    enum { Flip = (DstEvaluatorType::Flags & RowMajorBit) != (SrcEvaluatorType::Flags & RowMajorBit) };
+
+    
+    DstXprType temp(src.rows(), src.cols());
+
+    temp.reserve((std::max)(src.rows(),src.cols())*2);
+    for (Index j=0; j<outerEvaluationSize; ++j)
+    {
+      temp.startVec(j);
+      for (typename SrcEvaluatorType::InnerIterator it(srcEvaluator, j); it; ++it)
+      {
+        Scalar v = it.value();
+        temp.insertBackByOuterInner(Flip?it.index():j,Flip?j:it.index()) = v;
+      }
+    }
+    temp.finalize();
+
+    dst = temp.markAsRValue();
+  }
+}
+
+// Generic Sparse to Sparse assignment
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Sparse>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    assign_sparse_to_sparse(dst.derived(), src.derived());
+  }
+};
+
+// Generic Sparse to Dense assignment
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Sparse2Dense>
+{
+  static void run(DstXprType &dst, const SrcXprType &src, const Functor &func)
+  {
+    if(internal::is_same<Functor,internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> >::value)
+      dst.setZero();
+    
+    internal::evaluator<SrcXprType> srcEval(src);
+    resize_if_allowed(dst, src, func);
+    internal::evaluator<DstXprType> dstEval(dst);
+    
+    const Index outerEvaluationSize = (internal::evaluator<SrcXprType>::Flags&RowMajorBit) ? src.rows() : src.cols();
+    for (Index j=0; j<outerEvaluationSize; ++j)
+      for (typename internal::evaluator<SrcXprType>::InnerIterator i(srcEval,j); i; ++i)
+        func.assignCoeff(dstEval.coeffRef(i.row(),i.col()), i.value());
+  }
+};
+
+// Specialization for "dst = dec.solve(rhs)"
+// NOTE we need to specialize it for Sparse2Sparse to avoid ambiguous specialization error
+template<typename DstXprType, typename DecType, typename RhsType, typename Scalar>
+struct Assignment<DstXprType, Solve<DecType,RhsType>, internal::assign_op<Scalar,Scalar>, Sparse2Sparse>
+{
+  typedef Solve<DecType,RhsType> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    src.dec()._solve_impl(src.rhs(), dst);
+  }
+};
+
+struct Diagonal2Sparse {};
+
+template<> struct AssignmentKind<SparseShape,DiagonalShape> { typedef Diagonal2Sparse Kind; };
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, Diagonal2Sparse>
+{
+  typedef typename DstXprType::StorageIndex StorageIndex;
+  typedef typename DstXprType::Scalar Scalar;
+  typedef Array<StorageIndex,Dynamic,1> ArrayXI;
+  typedef Array<Scalar,Dynamic,1> ArrayXS;
+  template<int Options>
+  static void run(SparseMatrix<Scalar,Options,StorageIndex> &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+
+    Index size = src.diagonal().size();
+    dst.makeCompressed();
+    dst.resizeNonZeros(size);
+    Map<ArrayXI>(dst.innerIndexPtr(), size).setLinSpaced(0,StorageIndex(size)-1);
+    Map<ArrayXI>(dst.outerIndexPtr(), size+1).setLinSpaced(0,StorageIndex(size));
+    Map<ArrayXS>(dst.valuePtr(), size) = src.diagonal();
+  }
+  
+  template<typename DstDerived>
+  static void run(SparseMatrixBase<DstDerived> &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  {
+    dst.diagonal() = src.diagonal();
+  }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() += src.diagonal(); }
+  
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &/*func*/)
+  { dst.diagonal() -= src.diagonal(); }
+};
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEASSIGN_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseBlock.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseBlock.h
new file mode 100644
index 0000000..511e92b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseBlock.h
@@ -0,0 +1,603 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_BLOCK_H
+#define EIGEN_SPARSE_BLOCK_H
+
+namespace Eigen {
+
+// Subset of columns or rows
+template<typename XprType, int BlockRows, int BlockCols>
+class BlockImpl<XprType,BlockRows,BlockCols,true,Sparse>
+  : public SparseMatrixBase<Block<XprType,BlockRows,BlockCols,true> >
+{
+    typedef typename internal::remove_all<typename XprType::Nested>::type _MatrixTypeNested;
+    typedef Block<XprType, BlockRows, BlockCols, true> BlockType;
+public:
+    enum { IsRowMajor = internal::traits<BlockType>::IsRowMajor };
+protected:
+    enum { OuterSize = IsRowMajor ? BlockRows : BlockCols };
+    typedef SparseMatrixBase<BlockType> Base;
+    using Base::convert_index;
+public:
+    EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType)
+
+    inline BlockImpl(XprType& xpr, Index i)
+      : m_matrix(xpr), m_outerStart(convert_index(i)), m_outerSize(OuterSize)
+    {}
+
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : m_matrix(xpr), m_outerStart(convert_index(IsRowMajor ? startRow : startCol)), m_outerSize(convert_index(IsRowMajor ? blockRows : blockCols))
+    {}
+
+    EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); }
+    EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); }
+
+    Index nonZeros() const
+    {
+      typedef internal::evaluator<XprType> EvaluatorType;
+      EvaluatorType matEval(m_matrix);
+      Index nnz = 0;
+      Index end = m_outerStart + m_outerSize.value();
+      for(Index j=m_outerStart; j<end; ++j)
+        for(typename EvaluatorType::InnerIterator it(matEval, j); it; ++it)
+          ++nnz;
+      return nnz;
+    }
+
+    inline const Scalar coeff(Index row, Index col) const
+    {
+      return m_matrix.coeff(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 :  m_outerStart));
+    }
+
+    inline const Scalar coeff(Index index) const
+    {
+      return m_matrix.coeff(IsRowMajor ? m_outerStart : index, IsRowMajor ? index :  m_outerStart);
+    }
+
+    inline const XprType& nestedExpression() const { return m_matrix; }
+    inline XprType& nestedExpression() { return m_matrix; }
+    Index startRow() const { return IsRowMajor ? m_outerStart : 0; }
+    Index startCol() const { return IsRowMajor ? 0 : m_outerStart; }
+    Index blockRows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); }
+    Index blockCols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); }
+
+  protected:
+
+    typename internal::ref_selector<XprType>::non_const_type m_matrix;
+    Index m_outerStart;
+    const internal::variable_if_dynamic<Index, OuterSize> m_outerSize;
+
+  protected:
+    // Disable assignment with clear error message.
+    // Note that simply removing operator= yields compilation errors with ICC+MSVC
+    template<typename T>
+    BlockImpl& operator=(const T&)
+    {
+      EIGEN_STATIC_ASSERT(sizeof(T)==0, THIS_SPARSE_BLOCK_SUBEXPRESSION_IS_READ_ONLY);
+      return *this;
+    }
+};
+
+
+/***************************************************************************
+* specialization for SparseMatrix
+***************************************************************************/
+
+namespace internal {
+
+template<typename SparseMatrixType, int BlockRows, int BlockCols>
+class sparse_matrix_block_impl
+  : public SparseCompressedBase<Block<SparseMatrixType,BlockRows,BlockCols,true> >
+{
+    typedef typename internal::remove_all<typename SparseMatrixType::Nested>::type _MatrixTypeNested;
+    typedef Block<SparseMatrixType, BlockRows, BlockCols, true> BlockType;
+    typedef SparseCompressedBase<Block<SparseMatrixType,BlockRows,BlockCols,true> > Base;
+    using Base::convert_index;
+public:
+    enum { IsRowMajor = internal::traits<BlockType>::IsRowMajor };
+    EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType)
+protected:
+    typedef typename Base::IndexVector IndexVector;
+    enum { OuterSize = IsRowMajor ? BlockRows : BlockCols };
+public:
+
+    inline sparse_matrix_block_impl(SparseMatrixType& xpr, Index i)
+      : m_matrix(xpr), m_outerStart(convert_index(i)), m_outerSize(OuterSize)
+    {}
+
+    inline sparse_matrix_block_impl(SparseMatrixType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : m_matrix(xpr), m_outerStart(convert_index(IsRowMajor ? startRow : startCol)), m_outerSize(convert_index(IsRowMajor ? blockRows : blockCols))
+    {}
+
+    template<typename OtherDerived>
+    inline BlockType& operator=(const SparseMatrixBase<OtherDerived>& other)
+    {
+      typedef typename internal::remove_all<typename SparseMatrixType::Nested>::type _NestedMatrixType;
+      _NestedMatrixType& matrix = m_matrix;
+      // This assignment is slow if this vector set is not empty
+      // and/or it is not at the end of the nonzeros of the underlying matrix.
+
+      // 1 - eval to a temporary to avoid transposition and/or aliasing issues
+      Ref<const SparseMatrix<Scalar, IsRowMajor ? RowMajor : ColMajor, StorageIndex> > tmp(other.derived());
+      eigen_internal_assert(tmp.outerSize()==m_outerSize.value());
+
+      // 2 - let's check whether there is enough allocated memory
+      Index nnz           = tmp.nonZeros();
+      Index start         = m_outerStart==0 ? 0 : m_matrix.outerIndexPtr()[m_outerStart]; // starting position of the current block
+      Index end           = m_matrix.outerIndexPtr()[m_outerStart+m_outerSize.value()]; // ending position of the current block
+      Index block_size    = end - start;                                                // available room in the current block
+      Index tail_size     = m_matrix.outerIndexPtr()[m_matrix.outerSize()] - end;
+
+      Index free_size     = m_matrix.isCompressed()
+                          ? Index(matrix.data().allocatedSize()) + block_size
+                          : block_size;
+
+      Index tmp_start = tmp.outerIndexPtr()[0];
+
+      bool update_trailing_pointers = false;
+      if(nnz>free_size)
+      {
+        // realloc manually to reduce copies
+        typename SparseMatrixType::Storage newdata(m_matrix.data().allocatedSize() - block_size + nnz);
+
+        internal::smart_copy(m_matrix.valuePtr(),       m_matrix.valuePtr() + start,      newdata.valuePtr());
+        internal::smart_copy(m_matrix.innerIndexPtr(),  m_matrix.innerIndexPtr() + start, newdata.indexPtr());
+
+        internal::smart_copy(tmp.valuePtr() + tmp_start,      tmp.valuePtr() + tmp_start + nnz,       newdata.valuePtr() + start);
+        internal::smart_copy(tmp.innerIndexPtr() + tmp_start, tmp.innerIndexPtr() + tmp_start + nnz,  newdata.indexPtr() + start);
+
+        internal::smart_copy(matrix.valuePtr()+end,       matrix.valuePtr()+end + tail_size,      newdata.valuePtr()+start+nnz);
+        internal::smart_copy(matrix.innerIndexPtr()+end,  matrix.innerIndexPtr()+end + tail_size, newdata.indexPtr()+start+nnz);
+
+        newdata.resize(m_matrix.outerIndexPtr()[m_matrix.outerSize()] - block_size + nnz);
+
+        matrix.data().swap(newdata);
+
+        update_trailing_pointers = true;
+      }
+      else
+      {
+        if(m_matrix.isCompressed())
+        {
+          // no need to realloc, simply copy the tail at its respective position and insert tmp
+          matrix.data().resize(start + nnz + tail_size);
+
+          internal::smart_memmove(matrix.valuePtr()+end,      matrix.valuePtr() + end+tail_size,      matrix.valuePtr() + start+nnz);
+          internal::smart_memmove(matrix.innerIndexPtr()+end, matrix.innerIndexPtr() + end+tail_size, matrix.innerIndexPtr() + start+nnz);
+
+          update_trailing_pointers = true;
+        }
+
+        internal::smart_copy(tmp.valuePtr() + tmp_start,      tmp.valuePtr() + tmp_start + nnz,       matrix.valuePtr() + start);
+        internal::smart_copy(tmp.innerIndexPtr() + tmp_start, tmp.innerIndexPtr() + tmp_start + nnz,  matrix.innerIndexPtr() + start);
+      }
+
+      // update outer index pointers and innerNonZeros
+      if(IsVectorAtCompileTime)
+      {
+        if(!m_matrix.isCompressed())
+          matrix.innerNonZeroPtr()[m_outerStart] = StorageIndex(nnz);
+        matrix.outerIndexPtr()[m_outerStart] = StorageIndex(start);
+      }
+      else
+      {
+        StorageIndex p = StorageIndex(start);
+        for(Index k=0; k<m_outerSize.value(); ++k)
+        {
+          StorageIndex nnz_k = internal::convert_index<StorageIndex>(tmp.innerVector(k).nonZeros());
+          if(!m_matrix.isCompressed())
+            matrix.innerNonZeroPtr()[m_outerStart+k] = nnz_k;
+          matrix.outerIndexPtr()[m_outerStart+k] = p;
+          p += nnz_k;
+        }
+      }
+
+      if(update_trailing_pointers)
+      {
+        StorageIndex offset = internal::convert_index<StorageIndex>(nnz - block_size);
+        for(Index k = m_outerStart + m_outerSize.value(); k<=matrix.outerSize(); ++k)
+        {
+          matrix.outerIndexPtr()[k] += offset;
+        }
+      }
+
+      return derived();
+    }
+
+    inline BlockType& operator=(const BlockType& other)
+    {
+      return operator=<BlockType>(other);
+    }
+
+    inline const Scalar* valuePtr() const
+    { return m_matrix.valuePtr(); }
+    inline Scalar* valuePtr()
+    { return m_matrix.valuePtr(); }
+
+    inline const StorageIndex* innerIndexPtr() const
+    { return m_matrix.innerIndexPtr(); }
+    inline StorageIndex* innerIndexPtr()
+    { return m_matrix.innerIndexPtr(); }
+
+    inline const StorageIndex* outerIndexPtr() const
+    { return m_matrix.outerIndexPtr() + m_outerStart; }
+    inline StorageIndex* outerIndexPtr()
+    { return m_matrix.outerIndexPtr() + m_outerStart; }
+
+    inline const StorageIndex* innerNonZeroPtr() const
+    { return isCompressed() ? 0 : (m_matrix.innerNonZeroPtr()+m_outerStart); }
+    inline StorageIndex* innerNonZeroPtr()
+    { return isCompressed() ? 0 : (m_matrix.innerNonZeroPtr()+m_outerStart); }
+
+    bool isCompressed() const { return m_matrix.innerNonZeroPtr()==0; }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_matrix.coeffRef(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 :  m_outerStart));
+    }
+
+    inline const Scalar coeff(Index row, Index col) const
+    {
+      return m_matrix.coeff(row + (IsRowMajor ? m_outerStart : 0), col + (IsRowMajor ? 0 :  m_outerStart));
+    }
+
+    inline const Scalar coeff(Index index) const
+    {
+      return m_matrix.coeff(IsRowMajor ? m_outerStart : index, IsRowMajor ? index :  m_outerStart);
+    }
+
+    const Scalar& lastCoeff() const
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(sparse_matrix_block_impl);
+      eigen_assert(Base::nonZeros()>0);
+      if(m_matrix.isCompressed())
+        return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart+1]-1];
+      else
+        return m_matrix.valuePtr()[m_matrix.outerIndexPtr()[m_outerStart]+m_matrix.innerNonZeroPtr()[m_outerStart]-1];
+    }
+
+    EIGEN_STRONG_INLINE Index rows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); }
+    EIGEN_STRONG_INLINE Index cols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); }
+
+    inline const SparseMatrixType& nestedExpression() const { return m_matrix; }
+    inline SparseMatrixType& nestedExpression() { return m_matrix; }
+    Index startRow() const { return IsRowMajor ? m_outerStart : 0; }
+    Index startCol() const { return IsRowMajor ? 0 : m_outerStart; }
+    Index blockRows() const { return IsRowMajor ? m_outerSize.value() : m_matrix.rows(); }
+    Index blockCols() const { return IsRowMajor ? m_matrix.cols() : m_outerSize.value(); }
+
+  protected:
+
+    typename internal::ref_selector<SparseMatrixType>::non_const_type m_matrix;
+    Index m_outerStart;
+    const internal::variable_if_dynamic<Index, OuterSize> m_outerSize;
+
+};
+
+} // namespace internal
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int BlockRows, int BlockCols>
+class BlockImpl<SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true,Sparse>
+  : public internal::sparse_matrix_block_impl<SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols>
+{
+public:
+  typedef _StorageIndex StorageIndex;
+  typedef SparseMatrix<_Scalar, _Options, _StorageIndex> SparseMatrixType;
+  typedef internal::sparse_matrix_block_impl<SparseMatrixType,BlockRows,BlockCols> Base;
+  inline BlockImpl(SparseMatrixType& xpr, Index i)
+    : Base(xpr, i)
+  {}
+
+  inline BlockImpl(SparseMatrixType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+    : Base(xpr, startRow, startCol, blockRows, blockCols)
+  {}
+
+  using Base::operator=;
+};
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int BlockRows, int BlockCols>
+class BlockImpl<const SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true,Sparse>
+  : public internal::sparse_matrix_block_impl<const SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols>
+{
+public:
+  typedef _StorageIndex StorageIndex;
+  typedef const SparseMatrix<_Scalar, _Options, _StorageIndex> SparseMatrixType;
+  typedef internal::sparse_matrix_block_impl<SparseMatrixType,BlockRows,BlockCols> Base;
+  inline BlockImpl(SparseMatrixType& xpr, Index i)
+    : Base(xpr, i)
+  {}
+
+  inline BlockImpl(SparseMatrixType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+    : Base(xpr, startRow, startCol, blockRows, blockCols)
+  {}
+
+  using Base::operator=;
+private:
+  template<typename Derived> BlockImpl(const SparseMatrixBase<Derived>& xpr, Index i);
+  template<typename Derived> BlockImpl(const SparseMatrixBase<Derived>& xpr);
+};
+
+//----------
+
+/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+  * is col-major (resp. row-major).
+  */
+template<typename Derived>
+typename SparseMatrixBase<Derived>::InnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer)
+{ return InnerVectorReturnType(derived(), outer); }
+
+/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+  * is col-major (resp. row-major). Read-only.
+  */
+template<typename Derived>
+const typename SparseMatrixBase<Derived>::ConstInnerVectorReturnType SparseMatrixBase<Derived>::innerVector(Index outer) const
+{ return ConstInnerVectorReturnType(derived(), outer); }
+
+/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+  * is col-major (resp. row-major).
+  */
+template<typename Derived>
+typename SparseMatrixBase<Derived>::InnerVectorsReturnType
+SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize)
+{
+  return Block<Derived,Dynamic,Dynamic,true>(derived(),
+                                             IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
+                                             IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
+
+}
+
+/** \returns the \a outer -th column (resp. row) of the matrix \c *this if \c *this
+  * is col-major (resp. row-major). Read-only.
+  */
+template<typename Derived>
+const typename SparseMatrixBase<Derived>::ConstInnerVectorsReturnType
+SparseMatrixBase<Derived>::innerVectors(Index outerStart, Index outerSize) const
+{
+  return Block<const Derived,Dynamic,Dynamic,true>(derived(),
+                                                  IsRowMajor ? outerStart : 0, IsRowMajor ? 0 : outerStart,
+                                                  IsRowMajor ? outerSize : rows(), IsRowMajor ? cols() : outerSize);
+
+}
+
+/** Generic implementation of sparse Block expression.
+  * Real-only.
+  */
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
+class BlockImpl<XprType,BlockRows,BlockCols,InnerPanel,Sparse>
+  : public SparseMatrixBase<Block<XprType,BlockRows,BlockCols,InnerPanel> >, internal::no_assignment_operator
+{
+    typedef Block<XprType, BlockRows, BlockCols, InnerPanel> BlockType;
+    typedef SparseMatrixBase<BlockType> Base;
+    using Base::convert_index;
+public:
+    enum { IsRowMajor = internal::traits<BlockType>::IsRowMajor };
+    EIGEN_SPARSE_PUBLIC_INTERFACE(BlockType)
+
+    typedef typename internal::remove_all<typename XprType::Nested>::type _MatrixTypeNested;
+
+    /** Column or Row constructor
+      */
+    inline BlockImpl(XprType& xpr, Index i)
+      : m_matrix(xpr),
+        m_startRow( (BlockRows==1) && (BlockCols==XprType::ColsAtCompileTime) ? convert_index(i) : 0),
+        m_startCol( (BlockRows==XprType::RowsAtCompileTime) && (BlockCols==1) ? convert_index(i) : 0),
+        m_blockRows(BlockRows==1 ? 1 : xpr.rows()),
+        m_blockCols(BlockCols==1 ? 1 : xpr.cols())
+    {}
+
+    /** Dynamic-size constructor
+      */
+    inline BlockImpl(XprType& xpr, Index startRow, Index startCol, Index blockRows, Index blockCols)
+      : m_matrix(xpr), m_startRow(convert_index(startRow)), m_startCol(convert_index(startCol)), m_blockRows(convert_index(blockRows)), m_blockCols(convert_index(blockCols))
+    {}
+
+    inline Index rows() const { return m_blockRows.value(); }
+    inline Index cols() const { return m_blockCols.value(); }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      return m_matrix.coeffRef(row + m_startRow.value(), col + m_startCol.value());
+    }
+
+    inline const Scalar coeff(Index row, Index col) const
+    {
+      return m_matrix.coeff(row + m_startRow.value(), col + m_startCol.value());
+    }
+
+    inline Scalar& coeffRef(Index index)
+    {
+      return m_matrix.coeffRef(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                               m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    inline const Scalar coeff(Index index) const
+    {
+      return m_matrix.coeff(m_startRow.value() + (RowsAtCompileTime == 1 ? 0 : index),
+                            m_startCol.value() + (RowsAtCompileTime == 1 ? index : 0));
+    }
+
+    inline const XprType& nestedExpression() const { return m_matrix; }
+    inline XprType& nestedExpression() { return m_matrix; }
+    Index startRow() const { return m_startRow.value(); }
+    Index startCol() const { return m_startCol.value(); }
+    Index blockRows() const { return m_blockRows.value(); }
+    Index blockCols() const { return m_blockCols.value(); }
+
+  protected:
+//     friend class internal::GenericSparseBlockInnerIteratorImpl<XprType,BlockRows,BlockCols,InnerPanel>;
+    friend struct internal::unary_evaluator<Block<XprType,BlockRows,BlockCols,InnerPanel>, internal::IteratorBased, Scalar >;
+
+    Index nonZeros() const { return Dynamic; }
+
+    typename internal::ref_selector<XprType>::non_const_type m_matrix;
+    const internal::variable_if_dynamic<Index, XprType::RowsAtCompileTime == 1 ? 0 : Dynamic> m_startRow;
+    const internal::variable_if_dynamic<Index, XprType::ColsAtCompileTime == 1 ? 0 : Dynamic> m_startCol;
+    const internal::variable_if_dynamic<Index, RowsAtCompileTime> m_blockRows;
+    const internal::variable_if_dynamic<Index, ColsAtCompileTime> m_blockCols;
+
+  protected:
+    // Disable assignment with clear error message.
+    // Note that simply removing operator= yields compilation errors with ICC+MSVC
+    template<typename T>
+    BlockImpl& operator=(const T&)
+    {
+      EIGEN_STATIC_ASSERT(sizeof(T)==0, THIS_SPARSE_BLOCK_SUBEXPRESSION_IS_READ_ONLY);
+      return *this;
+    }
+
+};
+
+namespace internal {
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+struct unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased >
+ : public evaluator_base<Block<ArgType,BlockRows,BlockCols,InnerPanel> >
+{
+    class InnerVectorInnerIterator;
+    class OuterVectorInnerIterator;
+  public:
+    typedef Block<ArgType,BlockRows,BlockCols,InnerPanel> XprType;
+    typedef typename XprType::StorageIndex StorageIndex;
+    typedef typename XprType::Scalar Scalar;
+
+    enum {
+      IsRowMajor = XprType::IsRowMajor,
+
+      OuterVector =  (BlockCols==1 && ArgType::IsRowMajor)
+                    | // FIXME | instead of || to please GCC 4.4.0 stupid warning "suggest parentheses around &&".
+                      // revert to || as soon as not needed anymore.
+                     (BlockRows==1 && !ArgType::IsRowMajor),
+
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+      Flags = XprType::Flags
+    };
+
+    typedef typename internal::conditional<OuterVector,OuterVectorInnerIterator,InnerVectorInnerIterator>::type InnerIterator;
+
+    explicit unary_evaluator(const XprType& op)
+      : m_argImpl(op.nestedExpression()), m_block(op)
+    {}
+
+    inline Index nonZerosEstimate() const {
+      Index nnz = m_block.nonZeros();
+      if(nnz<0)
+        return m_argImpl.nonZerosEstimate() * m_block.size() / m_block.nestedExpression().size();
+      return nnz;
+    }
+
+  protected:
+    typedef typename evaluator<ArgType>::InnerIterator EvalIterator;
+
+    evaluator<ArgType> m_argImpl;
+    const XprType &m_block;
+};
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+class unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased>::InnerVectorInnerIterator
+ : public EvalIterator
+{
+  enum { IsRowMajor = unary_evaluator::IsRowMajor };
+  const XprType& m_block;
+  Index m_end;
+public:
+
+  EIGEN_STRONG_INLINE InnerVectorInnerIterator(const unary_evaluator& aEval, Index outer)
+    : EvalIterator(aEval.m_argImpl, outer + (IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol())),
+      m_block(aEval.m_block),
+      m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows())
+  {
+    while( (EvalIterator::operator bool()) && (EvalIterator::index() < (IsRowMajor ? m_block.startCol() : m_block.startRow())) )
+      EvalIterator::operator++();
+  }
+
+  inline StorageIndex index() const { return EvalIterator::index() - convert_index<StorageIndex>(IsRowMajor ? m_block.startCol() : m_block.startRow()); }
+  inline Index outer()  const { return EvalIterator::outer() - (IsRowMajor ? m_block.startRow() : m_block.startCol()); }
+  inline Index row()    const { return EvalIterator::row()   - m_block.startRow(); }
+  inline Index col()    const { return EvalIterator::col()   - m_block.startCol(); }
+
+  inline operator bool() const { return EvalIterator::operator bool() && EvalIterator::index() < m_end; }
+};
+
+template<typename ArgType, int BlockRows, int BlockCols, bool InnerPanel>
+class unary_evaluator<Block<ArgType,BlockRows,BlockCols,InnerPanel>, IteratorBased>::OuterVectorInnerIterator
+{
+  enum { IsRowMajor = unary_evaluator::IsRowMajor };
+  const unary_evaluator& m_eval;
+  Index m_outerPos;
+  const Index m_innerIndex;
+  Index m_end;
+  EvalIterator m_it;
+public:
+
+  EIGEN_STRONG_INLINE OuterVectorInnerIterator(const unary_evaluator& aEval, Index outer)
+    : m_eval(aEval),
+      m_outerPos( (IsRowMajor ? aEval.m_block.startCol() : aEval.m_block.startRow()) ),
+      m_innerIndex(IsRowMajor ? aEval.m_block.startRow() : aEval.m_block.startCol()),
+      m_end(IsRowMajor ? aEval.m_block.startCol()+aEval.m_block.blockCols() : aEval.m_block.startRow()+aEval.m_block.blockRows()),
+      m_it(m_eval.m_argImpl, m_outerPos)
+  {
+    EIGEN_UNUSED_VARIABLE(outer);
+    eigen_assert(outer==0);
+
+    while(m_it && m_it.index() < m_innerIndex) ++m_it;
+    if((!m_it) || (m_it.index()!=m_innerIndex))
+      ++(*this);
+  }
+
+  inline StorageIndex index() const { return convert_index<StorageIndex>(m_outerPos - (IsRowMajor ? m_eval.m_block.startCol() : m_eval.m_block.startRow())); }
+  inline Index outer()  const { return 0; }
+  inline Index row()    const { return IsRowMajor ? 0 : index(); }
+  inline Index col()    const { return IsRowMajor ? index() : 0; }
+
+  inline Scalar value() const { return m_it.value(); }
+  inline Scalar& valueRef() { return m_it.valueRef(); }
+
+  inline OuterVectorInnerIterator& operator++()
+  {
+    // search next non-zero entry
+    while(++m_outerPos<m_end)
+    {
+      // Restart iterator at the next inner-vector:
+      m_it.~EvalIterator();
+      ::new (&m_it) EvalIterator(m_eval.m_argImpl, m_outerPos);
+      // search for the key m_innerIndex in the current outer-vector
+      while(m_it && m_it.index() < m_innerIndex) ++m_it;
+      if(m_it && m_it.index()==m_innerIndex) break;
+    }
+    return *this;
+  }
+
+  inline operator bool() const { return m_outerPos < m_end; }
+};
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int BlockRows, int BlockCols>
+struct unary_evaluator<Block<SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true>, IteratorBased>
+  : evaluator<SparseCompressedBase<Block<SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true> > >
+{
+  typedef Block<SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true> XprType;
+  typedef evaluator<SparseCompressedBase<XprType> > Base;
+  explicit unary_evaluator(const XprType &xpr) : Base(xpr) {}
+};
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int BlockRows, int BlockCols>
+struct unary_evaluator<Block<const SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true>, IteratorBased>
+  : evaluator<SparseCompressedBase<Block<const SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true> > >
+{
+  typedef Block<const SparseMatrix<_Scalar, _Options, _StorageIndex>,BlockRows,BlockCols,true> XprType;
+  typedef evaluator<SparseCompressedBase<XprType> > Base;
+  explicit unary_evaluator(const XprType &xpr) : Base(xpr) {}
+};
+
+} // end namespace internal
+
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_BLOCK_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseColEtree.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseColEtree.h
new file mode 100644
index 0000000..ebe02d1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseColEtree.h
@@ -0,0 +1,206 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+/* 
+ 
+ * NOTE: This file is the modified version of sp_coletree.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 3.1) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * August 1, 2008
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSE_COLETREE_H
+#define SPARSE_COLETREE_H
+
+namespace Eigen {
+
+namespace internal {
+
+/** Find the root of the tree/set containing the vertex i : Use Path halving */ 
+template<typename Index, typename IndexVector>
+Index etree_find (Index i, IndexVector& pp)
+{
+  Index p = pp(i); // Parent 
+  Index gp = pp(p); // Grand parent 
+  while (gp != p) 
+  {
+    pp(i) = gp; // Parent pointer on find path is changed to former grand parent
+    i = gp; 
+    p = pp(i);
+    gp = pp(p);
+  }
+  return p; 
+}
+
+/** Compute the column elimination tree of a sparse matrix
+  * \param mat The matrix in column-major format. 
+  * \param parent The elimination tree
+  * \param firstRowElt The column index of the first element in each row
+  * \param perm The permutation to apply to the column of \b mat
+  */
+template <typename MatrixType, typename IndexVector>
+int coletree(const MatrixType& mat, IndexVector& parent, IndexVector& firstRowElt, typename MatrixType::StorageIndex *perm=0)
+{
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  StorageIndex nc = convert_index<StorageIndex>(mat.cols()); // Number of columns
+  StorageIndex m = convert_index<StorageIndex>(mat.rows());
+  StorageIndex diagSize = (std::min)(nc,m);
+  IndexVector root(nc); // root of subtree of etree 
+  root.setZero();
+  IndexVector pp(nc); // disjoint sets 
+  pp.setZero(); // Initialize disjoint sets 
+  parent.resize(mat.cols());
+  //Compute first nonzero column in each row 
+  firstRowElt.resize(m);
+  firstRowElt.setConstant(nc);
+  firstRowElt.segment(0, diagSize).setLinSpaced(diagSize, 0, diagSize-1);
+  bool found_diag;
+  for (StorageIndex col = 0; col < nc; col++)
+  {
+    StorageIndex pcol = col;
+    if(perm) pcol  = perm[col];
+    for (typename MatrixType::InnerIterator it(mat, pcol); it; ++it)
+    { 
+      Index row = it.row();
+      firstRowElt(row) = (std::min)(firstRowElt(row), col);
+    }
+  }
+  /* Compute etree by Liu's algorithm for symmetric matrices,
+          except use (firstRowElt[r],c) in place of an edge (r,c) of A.
+    Thus each row clique in A'*A is replaced by a star
+    centered at its first vertex, which has the same fill. */
+  StorageIndex rset, cset, rroot;
+  for (StorageIndex col = 0; col < nc; col++) 
+  {
+    found_diag = col>=m;
+    pp(col) = col; 
+    cset = col; 
+    root(cset) = col; 
+    parent(col) = nc; 
+    /* The diagonal element is treated here even if it does not exist in the matrix
+     * hence the loop is executed once more */ 
+    StorageIndex pcol = col;
+    if(perm) pcol  = perm[col];
+    for (typename MatrixType::InnerIterator it(mat, pcol); it||!found_diag; ++it)
+    { //  A sequence of interleaved find and union is performed 
+      Index i = col;
+      if(it) i = it.index();
+      if (i == col) found_diag = true;
+      
+      StorageIndex row = firstRowElt(i);
+      if (row >= col) continue; 
+      rset = internal::etree_find(row, pp); // Find the name of the set containing row
+      rroot = root(rset);
+      if (rroot != col) 
+      {
+        parent(rroot) = col; 
+        pp(cset) = rset; 
+        cset = rset; 
+        root(cset) = col; 
+      }
+    }
+  }
+  return 0;  
+}
+
+/** 
+  * Depth-first search from vertex n.  No recursion.
+  * This routine was contributed by Cédric Doucet, CEDRAT Group, Meylan, France.
+*/
+template <typename IndexVector>
+void nr_etdfs (typename IndexVector::Scalar n, IndexVector& parent, IndexVector& first_kid, IndexVector& next_kid, IndexVector& post, typename IndexVector::Scalar postnum)
+{
+  typedef typename IndexVector::Scalar StorageIndex;
+  StorageIndex current = n, first, next;
+  while (postnum != n) 
+  {
+    // No kid for the current node
+    first = first_kid(current);
+    
+    // no kid for the current node
+    if (first == -1) 
+    {
+      // Numbering this node because it has no kid 
+      post(current) = postnum++;
+      
+      // looking for the next kid 
+      next = next_kid(current); 
+      while (next == -1) 
+      {
+        // No more kids : back to the parent node
+        current = parent(current); 
+        // numbering the parent node 
+        post(current) = postnum++;
+        
+        // Get the next kid 
+        next = next_kid(current); 
+      }
+      // stopping criterion 
+      if (postnum == n+1) return; 
+      
+      // Updating current node 
+      current = next; 
+    }
+    else 
+    {
+      current = first; 
+    }
+  }
+}
+
+
+/**
+  * \brief Post order a tree 
+  * \param n the number of nodes
+  * \param parent Input tree
+  * \param post postordered tree
+  */
+template <typename IndexVector>
+void treePostorder(typename IndexVector::Scalar n, IndexVector& parent, IndexVector& post)
+{
+  typedef typename IndexVector::Scalar StorageIndex;
+  IndexVector first_kid, next_kid; // Linked list of children 
+  StorageIndex postnum; 
+  // Allocate storage for working arrays and results 
+  first_kid.resize(n+1); 
+  next_kid.setZero(n+1);
+  post.setZero(n+1);
+  
+  // Set up structure describing children
+  first_kid.setConstant(-1); 
+  for (StorageIndex v = n-1; v >= 0; v--) 
+  {
+    StorageIndex dad = parent(v);
+    next_kid(v) = first_kid(dad); 
+    first_kid(dad) = v; 
+  }
+  
+  // Depth-first search from dummy root vertex #n
+  postnum = 0; 
+  internal::nr_etdfs(n, parent, first_kid, next_kid, post, postnum);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // SPARSE_COLETREE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCompressedBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCompressedBase.h
new file mode 100644
index 0000000..5ccb466
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCompressedBase.h
@@ -0,0 +1,341 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_COMPRESSED_BASE_H
+#define EIGEN_SPARSE_COMPRESSED_BASE_H
+
+namespace Eigen { 
+
+template<typename Derived> class SparseCompressedBase;
+  
+namespace internal {
+
+template<typename Derived>
+struct traits<SparseCompressedBase<Derived> > : traits<Derived>
+{};
+
+} // end namespace internal
+
+/** \ingroup SparseCore_Module
+  * \class SparseCompressedBase
+  * \brief Common base class for sparse [compressed]-{row|column}-storage format.
+  *
+  * This class defines the common interface for all derived classes implementing the compressed sparse storage format, such as:
+  *  - SparseMatrix
+  *  - Ref<SparseMatrixType,Options>
+  *  - Map<SparseMatrixType>
+  *
+  */
+template<typename Derived>
+class SparseCompressedBase
+  : public SparseMatrixBase<Derived>
+{
+  public:
+    typedef SparseMatrixBase<Derived> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(SparseCompressedBase)
+    using Base::operator=;
+    using Base::IsRowMajor;
+    
+    class InnerIterator;
+    class ReverseInnerIterator;
+    
+  protected:
+    typedef typename Base::IndexVector IndexVector;
+    Eigen::Map<IndexVector> innerNonZeros() { return Eigen::Map<IndexVector>(innerNonZeroPtr(), isCompressed()?0:derived().outerSize()); }
+    const  Eigen::Map<const IndexVector> innerNonZeros() const { return Eigen::Map<const IndexVector>(innerNonZeroPtr(), isCompressed()?0:derived().outerSize()); }
+        
+  public:
+    
+    /** \returns the number of non zero coefficients */
+    inline Index nonZeros() const
+    {
+      if(Derived::IsVectorAtCompileTime && outerIndexPtr()==0)
+        return derived().nonZeros();
+      else if(isCompressed())
+        return outerIndexPtr()[derived().outerSize()]-outerIndexPtr()[0];
+      else if(derived().outerSize()==0)
+        return 0;
+      else
+        return innerNonZeros().sum();
+    }
+    
+    /** \returns a const pointer to the array of values.
+      * This function is aimed at interoperability with other libraries.
+      * \sa innerIndexPtr(), outerIndexPtr() */
+    inline const Scalar* valuePtr() const { return derived().valuePtr(); }
+    /** \returns a non-const pointer to the array of values.
+      * This function is aimed at interoperability with other libraries.
+      * \sa innerIndexPtr(), outerIndexPtr() */
+    inline Scalar* valuePtr() { return derived().valuePtr(); }
+
+    /** \returns a const pointer to the array of inner indices.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), outerIndexPtr() */
+    inline const StorageIndex* innerIndexPtr() const { return derived().innerIndexPtr(); }
+    /** \returns a non-const pointer to the array of inner indices.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), outerIndexPtr() */
+    inline StorageIndex* innerIndexPtr() { return derived().innerIndexPtr(); }
+
+    /** \returns a const pointer to the array of the starting positions of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 for SparseVector
+      * \sa valuePtr(), innerIndexPtr() */
+    inline const StorageIndex* outerIndexPtr() const { return derived().outerIndexPtr(); }
+    /** \returns a non-const pointer to the array of the starting positions of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 for SparseVector
+      * \sa valuePtr(), innerIndexPtr() */
+    inline StorageIndex* outerIndexPtr() { return derived().outerIndexPtr(); }
+
+    /** \returns a const pointer to the array of the number of non zeros of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 in compressed mode */
+    inline const StorageIndex* innerNonZeroPtr() const { return derived().innerNonZeroPtr(); }
+    /** \returns a non-const pointer to the array of the number of non zeros of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 in compressed mode */
+    inline StorageIndex* innerNonZeroPtr() { return derived().innerNonZeroPtr(); }
+    
+    /** \returns whether \c *this is in compressed form. */
+    inline bool isCompressed() const { return innerNonZeroPtr()==0; }
+
+    /** \returns a read-only view of the stored coefficients as a 1D array expression.
+      *
+      * \warning this method is for \b compressed \b storage \b only, and it will trigger an assertion otherwise.
+      *
+      * \sa valuePtr(), isCompressed() */
+    const Map<const Array<Scalar,Dynamic,1> > coeffs() const { eigen_assert(isCompressed()); return Array<Scalar,Dynamic,1>::Map(valuePtr(),nonZeros()); }
+
+    /** \returns a read-write view of the stored coefficients as a 1D array expression
+      *
+      * \warning this method is for \b compressed \b storage \b only, and it will trigger an assertion otherwise.
+      *
+      * Here is an example:
+      * \include SparseMatrix_coeffs.cpp
+      * and the output is:
+      * \include SparseMatrix_coeffs.out
+      *
+      * \sa valuePtr(), isCompressed() */
+    Map<Array<Scalar,Dynamic,1> > coeffs() { eigen_assert(isCompressed()); return Array<Scalar,Dynamic,1>::Map(valuePtr(),nonZeros()); }
+
+  protected:
+    /** Default constructor. Do nothing. */
+    SparseCompressedBase() {}
+  private:
+    template<typename OtherDerived> explicit SparseCompressedBase(const SparseCompressedBase<OtherDerived>&);
+};
+
+template<typename Derived>
+class SparseCompressedBase<Derived>::InnerIterator
+{
+  public:
+    InnerIterator()
+      : m_values(0), m_indices(0), m_outer(0), m_id(0), m_end(0)
+    {}
+
+    InnerIterator(const InnerIterator& other)
+      : m_values(other.m_values), m_indices(other.m_indices), m_outer(other.m_outer), m_id(other.m_id), m_end(other.m_end)
+    {}
+
+    InnerIterator& operator=(const InnerIterator& other)
+    {
+      m_values = other.m_values;
+      m_indices = other.m_indices;
+      const_cast<OuterType&>(m_outer).setValue(other.m_outer.value());
+      m_id = other.m_id;
+      m_end = other.m_end;
+      return *this;
+    }
+
+    InnerIterator(const SparseCompressedBase& mat, Index outer)
+      : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer)
+    {
+      if(Derived::IsVectorAtCompileTime && mat.outerIndexPtr()==0)
+      {
+        m_id = 0;
+        m_end = mat.nonZeros();
+      }
+      else
+      {
+        m_id = mat.outerIndexPtr()[outer];
+        if(mat.isCompressed())
+          m_end = mat.outerIndexPtr()[outer+1];
+        else
+          m_end = m_id + mat.innerNonZeroPtr()[outer];
+      }
+    }
+
+    explicit InnerIterator(const SparseCompressedBase& mat)
+      : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(0), m_id(0), m_end(mat.nonZeros())
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+    }
+
+    explicit InnerIterator(const internal::CompressedStorage<Scalar,StorageIndex>& data)
+      : m_values(data.valuePtr()), m_indices(data.indexPtr()), m_outer(0), m_id(0), m_end(data.size())
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+    }
+
+    inline InnerIterator& operator++() { m_id++; return *this; }
+
+    inline const Scalar& value() const { return m_values[m_id]; }
+    inline Scalar& valueRef() { return const_cast<Scalar&>(m_values[m_id]); }
+
+    inline StorageIndex index() const { return m_indices[m_id]; }
+    inline Index outer() const { return m_outer.value(); }
+    inline Index row() const { return IsRowMajor ? m_outer.value() : index(); }
+    inline Index col() const { return IsRowMajor ? index() : m_outer.value(); }
+
+    inline operator bool() const { return (m_id < m_end); }
+
+  protected:
+    const Scalar* m_values;
+    const StorageIndex* m_indices;
+    typedef internal::variable_if_dynamic<Index,Derived::IsVectorAtCompileTime?0:Dynamic> OuterType;
+    const OuterType m_outer;
+    Index m_id;
+    Index m_end;
+  private:
+    // If you get here, then you're not using the right InnerIterator type, e.g.:
+    //   SparseMatrix<double,RowMajor> A;
+    //   SparseMatrix<double>::InnerIterator it(A,0);
+    template<typename T> InnerIterator(const SparseMatrixBase<T>&, Index outer);
+};
+
+template<typename Derived>
+class SparseCompressedBase<Derived>::ReverseInnerIterator
+{
+  public:
+    ReverseInnerIterator(const SparseCompressedBase& mat, Index outer)
+      : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(outer)
+    {
+      if(Derived::IsVectorAtCompileTime && mat.outerIndexPtr()==0)
+      {
+        m_start = 0;
+        m_id = mat.nonZeros();
+      }
+      else
+      {
+        m_start = mat.outerIndexPtr()[outer];
+        if(mat.isCompressed())
+          m_id = mat.outerIndexPtr()[outer+1];
+        else
+          m_id = m_start + mat.innerNonZeroPtr()[outer];
+      }
+    }
+
+    explicit ReverseInnerIterator(const SparseCompressedBase& mat)
+      : m_values(mat.valuePtr()), m_indices(mat.innerIndexPtr()), m_outer(0), m_start(0), m_id(mat.nonZeros())
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+    }
+
+    explicit ReverseInnerIterator(const internal::CompressedStorage<Scalar,StorageIndex>& data)
+      : m_values(data.valuePtr()), m_indices(data.indexPtr()), m_outer(0), m_start(0), m_id(data.size())
+    {
+      EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived);
+    }
+
+    inline ReverseInnerIterator& operator--() { --m_id; return *this; }
+
+    inline const Scalar& value() const { return m_values[m_id-1]; }
+    inline Scalar& valueRef() { return const_cast<Scalar&>(m_values[m_id-1]); }
+
+    inline StorageIndex index() const { return m_indices[m_id-1]; }
+    inline Index outer() const { return m_outer.value(); }
+    inline Index row() const { return IsRowMajor ? m_outer.value() : index(); }
+    inline Index col() const { return IsRowMajor ? index() : m_outer.value(); }
+
+    inline operator bool() const { return (m_id > m_start); }
+
+  protected:
+    const Scalar* m_values;
+    const StorageIndex* m_indices;
+    typedef internal::variable_if_dynamic<Index,Derived::IsVectorAtCompileTime?0:Dynamic> OuterType;
+    const OuterType m_outer;
+    Index m_start;
+    Index m_id;
+};
+
+namespace internal {
+
+template<typename Derived>
+struct evaluator<SparseCompressedBase<Derived> >
+  : evaluator_base<Derived>
+{
+  typedef typename Derived::Scalar Scalar;
+  typedef typename Derived::InnerIterator InnerIterator;
+  
+  enum {
+    CoeffReadCost = NumTraits<Scalar>::ReadCost,
+    Flags = Derived::Flags
+  };
+  
+  evaluator() : m_matrix(0), m_zero(0)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  explicit evaluator(const Derived &mat) : m_matrix(&mat), m_zero(0)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return m_matrix->nonZeros();
+  }
+  
+  operator Derived&() { return m_matrix->const_cast_derived(); }
+  operator const Derived&() const { return *m_matrix; }
+  
+  typedef typename DenseCoeffsBase<Derived,ReadOnlyAccessors>::CoeffReturnType CoeffReturnType;
+  const Scalar& coeff(Index row, Index col) const
+  {
+    Index p = find(row,col);
+
+    if(p==Dynamic)
+      return m_zero;
+    else
+      return m_matrix->const_cast_derived().valuePtr()[p];
+  }
+
+  Scalar& coeffRef(Index row, Index col)
+  {
+    Index p = find(row,col);
+    eigen_assert(p!=Dynamic && "written coefficient does not exist");
+    return m_matrix->const_cast_derived().valuePtr()[p];
+  }
+
+protected:
+
+  Index find(Index row, Index col) const
+  {
+    eigen_internal_assert(row>=0 && row<m_matrix->rows() && col>=0 && col<m_matrix->cols());
+
+    const Index outer = Derived::IsRowMajor ? row : col;
+    const Index inner = Derived::IsRowMajor ? col : row;
+
+    Index start = m_matrix->outerIndexPtr()[outer];
+    Index end = m_matrix->isCompressed() ? m_matrix->outerIndexPtr()[outer+1] : m_matrix->outerIndexPtr()[outer] + m_matrix->innerNonZeroPtr()[outer];
+    eigen_assert(end>=start && "you are using a non finalized sparse matrix or written coefficient does not exist");
+    const Index p = std::lower_bound(m_matrix->innerIndexPtr()+start, m_matrix->innerIndexPtr()+end,inner) - m_matrix->innerIndexPtr();
+
+    return ((p<end) && (m_matrix->innerIndexPtr()[p]==inner)) ? p : Dynamic;
+  }
+
+  const Derived *m_matrix;
+  const Scalar m_zero;
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_COMPRESSED_BASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseBinaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseBinaryOp.h
new file mode 100644
index 0000000..e315e35
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseBinaryOp.h
@@ -0,0 +1,726 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_CWISE_BINARY_OP_H
+#define EIGEN_SPARSE_CWISE_BINARY_OP_H
+
+namespace Eigen { 
+
+// Here we have to handle 3 cases:
+//  1 - sparse op dense
+//  2 - dense op sparse
+//  3 - sparse op sparse
+// We also need to implement a 4th iterator for:
+//  4 - dense op dense
+// Finally, we also need to distinguish between the product and other operations :
+//                configuration      returned mode
+//  1 - sparse op dense    product      sparse
+//                         generic      dense
+//  2 - dense op sparse    product      sparse
+//                         generic      dense
+//  3 - sparse op sparse   product      sparse
+//                         generic      sparse
+//  4 - dense op dense     product      dense
+//                         generic      dense
+//
+// TODO to ease compiler job, we could specialize product/quotient with a scalar
+//      and fallback to cwise-unary evaluator using bind1st_op and bind2nd_op.
+
+template<typename BinaryOp, typename Lhs, typename Rhs>
+class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Sparse>
+  : public SparseMatrixBase<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+  public:
+    typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> Derived;
+    typedef SparseMatrixBase<Derived> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Derived)
+    CwiseBinaryOpImpl()
+    {
+      EIGEN_STATIC_ASSERT((
+                (!internal::is_same<typename internal::traits<Lhs>::StorageKind,
+                                    typename internal::traits<Rhs>::StorageKind>::value)
+            ||  ((internal::evaluator<Lhs>::Flags&RowMajorBit) == (internal::evaluator<Rhs>::Flags&RowMajorBit))),
+            THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH);
+    }
+};
+
+namespace internal {
+
+  
+// Generic "sparse OP sparse"
+template<typename XprType> struct binary_sparse_evaluator;
+
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IteratorBased, IteratorBased>
+  : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+protected:
+  typedef typename evaluator<Lhs>::InnerIterator  LhsIterator;
+  typedef typename evaluator<Rhs>::InnerIterator  RhsIterator;
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef typename XprType::StorageIndex StorageIndex;
+public:
+
+  class InnerIterator
+  {
+  public:
+    
+    EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer)
+      : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor)
+    {
+      this->operator++();
+    }
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      if (m_lhsIter && m_rhsIter && (m_lhsIter.index() == m_rhsIter.index()))
+      {
+        m_id = m_lhsIter.index();
+        m_value = m_functor(m_lhsIter.value(), m_rhsIter.value());
+        ++m_lhsIter;
+        ++m_rhsIter;
+      }
+      else if (m_lhsIter && (!m_rhsIter || (m_lhsIter.index() < m_rhsIter.index())))
+      {
+        m_id = m_lhsIter.index();
+        m_value = m_functor(m_lhsIter.value(), Scalar(0));
+        ++m_lhsIter;
+      }
+      else if (m_rhsIter && (!m_lhsIter || (m_lhsIter.index() > m_rhsIter.index())))
+      {
+        m_id = m_rhsIter.index();
+        m_value = m_functor(Scalar(0), m_rhsIter.value());
+        ++m_rhsIter;
+      }
+      else
+      {
+        m_value = 0; // this is to avoid a compilation warning
+        m_id = -1;
+      }
+      return *this;
+    }
+
+    EIGEN_STRONG_INLINE Scalar value() const { return m_value; }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_id; }
+    EIGEN_STRONG_INLINE Index outer() const { return m_lhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return Lhs::IsRowMajor ? m_lhsIter.row() : index(); }
+    EIGEN_STRONG_INLINE Index col() const { return Lhs::IsRowMajor ? index() : m_lhsIter.col(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return m_id>=0; }
+
+  protected:
+    LhsIterator m_lhsIter;
+    RhsIterator m_rhsIter;
+    const BinaryOp& m_functor;
+    Scalar m_value;
+    StorageIndex m_id;
+  };
+  
+  
+  enum {
+    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    Flags = XprType::Flags
+  };
+  
+  explicit binary_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()), 
+      m_rhsImpl(xpr.rhs())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return m_lhsImpl.nonZerosEstimate() + m_rhsImpl.nonZerosEstimate();
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<Lhs> m_lhsImpl;
+  evaluator<Rhs> m_rhsImpl;
+};
+
+// dense op sparse
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IndexBased, IteratorBased>
+  : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+protected:
+  typedef typename evaluator<Rhs>::InnerIterator  RhsIterator;
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef typename XprType::StorageIndex StorageIndex;
+public:
+
+  class InnerIterator
+  {
+    enum { IsRowMajor = (int(Rhs::Flags)&RowMajorBit)==RowMajorBit };
+  public:
+
+    EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer)
+      : m_lhsEval(aEval.m_lhsImpl), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor), m_value(0), m_id(-1), m_innerSize(aEval.m_expr.rhs().innerSize())
+    {
+      this->operator++();
+    }
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      ++m_id;
+      if(m_id<m_innerSize)
+      {
+        Scalar lhsVal = m_lhsEval.coeff(IsRowMajor?m_rhsIter.outer():m_id,
+                                        IsRowMajor?m_id:m_rhsIter.outer());
+        if(m_rhsIter && m_rhsIter.index()==m_id)
+        {
+          m_value = m_functor(lhsVal, m_rhsIter.value());
+          ++m_rhsIter;
+        }
+        else
+          m_value = m_functor(lhsVal, Scalar(0));
+      }
+
+      return *this;
+    }
+
+    EIGEN_STRONG_INLINE Scalar value() const { eigen_internal_assert(m_id<m_innerSize); return m_value; }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_id; }
+    EIGEN_STRONG_INLINE Index outer() const { return m_rhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return IsRowMajor ? m_rhsIter.outer() : m_id; }
+    EIGEN_STRONG_INLINE Index col() const { return IsRowMajor ? m_id : m_rhsIter.outer(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return m_id<m_innerSize; }
+
+  protected:
+    const evaluator<Lhs> &m_lhsEval;
+    RhsIterator m_rhsIter;
+    const BinaryOp& m_functor;
+    Scalar m_value;
+    StorageIndex m_id;
+    StorageIndex m_innerSize;
+  };
+
+
+  enum {
+    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    // Expose storage order of the sparse expression
+    Flags = (XprType::Flags & ~RowMajorBit) | (int(Rhs::Flags)&RowMajorBit)
+  };
+
+  explicit binary_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()),
+      m_rhsImpl(xpr.rhs()),
+      m_expr(xpr)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  inline Index nonZerosEstimate() const {
+    return m_expr.size();
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<Lhs> m_lhsImpl;
+  evaluator<Rhs> m_rhsImpl;
+  const XprType &m_expr;
+};
+
+// sparse op dense
+template<typename BinaryOp, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<BinaryOp, Lhs, Rhs>, IteratorBased, IndexBased>
+  : evaluator_base<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
+{
+protected:
+  typedef typename evaluator<Lhs>::InnerIterator  LhsIterator;
+  typedef CwiseBinaryOp<BinaryOp, Lhs, Rhs> XprType;
+  typedef typename traits<XprType>::Scalar Scalar;
+  typedef typename XprType::StorageIndex StorageIndex;
+public:
+
+  class InnerIterator
+  {
+    enum { IsRowMajor = (int(Lhs::Flags)&RowMajorBit)==RowMajorBit };
+  public:
+
+    EIGEN_STRONG_INLINE InnerIterator(const binary_evaluator& aEval, Index outer)
+      : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsEval(aEval.m_rhsImpl), m_functor(aEval.m_functor), m_value(0), m_id(-1), m_innerSize(aEval.m_expr.lhs().innerSize())
+    {
+      this->operator++();
+    }
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      ++m_id;
+      if(m_id<m_innerSize)
+      {
+        Scalar rhsVal = m_rhsEval.coeff(IsRowMajor?m_lhsIter.outer():m_id,
+                                        IsRowMajor?m_id:m_lhsIter.outer());
+        if(m_lhsIter && m_lhsIter.index()==m_id)
+        {
+          m_value = m_functor(m_lhsIter.value(), rhsVal);
+          ++m_lhsIter;
+        }
+        else
+          m_value = m_functor(Scalar(0),rhsVal);
+      }
+
+      return *this;
+    }
+
+    EIGEN_STRONG_INLINE Scalar value() const { eigen_internal_assert(m_id<m_innerSize); return m_value; }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_id; }
+    EIGEN_STRONG_INLINE Index outer() const { return m_lhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return IsRowMajor ? m_lhsIter.outer() : m_id; }
+    EIGEN_STRONG_INLINE Index col() const { return IsRowMajor ? m_id : m_lhsIter.outer(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return m_id<m_innerSize; }
+
+  protected:
+    LhsIterator m_lhsIter;
+    const evaluator<Rhs> &m_rhsEval;
+    const BinaryOp& m_functor;
+    Scalar m_value;
+    StorageIndex m_id;
+    StorageIndex m_innerSize;
+  };
+
+
+  enum {
+    CoeffReadCost = evaluator<Lhs>::CoeffReadCost + evaluator<Rhs>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    // Expose storage order of the sparse expression
+    Flags = (XprType::Flags & ~RowMajorBit) | (int(Lhs::Flags)&RowMajorBit)
+  };
+
+  explicit binary_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()),
+      m_rhsImpl(xpr.rhs()),
+      m_expr(xpr)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+
+  inline Index nonZerosEstimate() const {
+    return m_expr.size();
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<Lhs> m_lhsImpl;
+  evaluator<Rhs> m_rhsImpl;
+  const XprType &m_expr;
+};
+
+template<typename T,
+         typename LhsKind   = typename evaluator_traits<typename T::Lhs>::Kind,
+         typename RhsKind   = typename evaluator_traits<typename T::Rhs>::Kind,
+         typename LhsScalar = typename traits<typename T::Lhs>::Scalar,
+         typename RhsScalar = typename traits<typename T::Rhs>::Scalar> struct sparse_conjunction_evaluator;
+
+// "sparse .* sparse"
+template<typename T1, typename T2, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs>, IteratorBased, IteratorBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+// "dense .* sparse"
+template<typename T1, typename T2, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs>, IndexBased, IteratorBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+// "sparse .* dense"
+template<typename T1, typename T2, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs>, IteratorBased, IndexBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_product_op<T1,T2>, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+// "sparse ./ dense"
+template<typename T1, typename T2, typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_quotient_op<T1,T2>, Lhs, Rhs>, IteratorBased, IndexBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_quotient_op<T1,T2>, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_quotient_op<T1,T2>, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+// "sparse && sparse"
+template<typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs>, IteratorBased, IteratorBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+// "dense && sparse"
+template<typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs>, IndexBased, IteratorBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+// "sparse && dense"
+template<typename Lhs, typename Rhs>
+struct binary_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs>, IteratorBased, IndexBased>
+  : sparse_conjunction_evaluator<CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> >
+{
+  typedef CwiseBinaryOp<scalar_boolean_and_op, Lhs, Rhs> XprType;
+  typedef sparse_conjunction_evaluator<XprType> Base;
+  explicit binary_evaluator(const XprType& xpr) : Base(xpr) {}
+};
+
+// "sparse ^ sparse"
+template<typename XprType>
+struct sparse_conjunction_evaluator<XprType, IteratorBased, IteratorBased>
+  : evaluator_base<XprType>
+{
+protected:
+  typedef typename XprType::Functor BinaryOp;
+  typedef typename XprType::Lhs LhsArg;
+  typedef typename XprType::Rhs RhsArg;
+  typedef typename evaluator<LhsArg>::InnerIterator  LhsIterator;
+  typedef typename evaluator<RhsArg>::InnerIterator  RhsIterator;
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename traits<XprType>::Scalar Scalar;
+public:
+
+  class InnerIterator
+  {
+  public:
+    
+    EIGEN_STRONG_INLINE InnerIterator(const sparse_conjunction_evaluator& aEval, Index outer)
+      : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor)
+    {
+      while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index()))
+      {
+        if (m_lhsIter.index() < m_rhsIter.index())
+          ++m_lhsIter;
+        else
+          ++m_rhsIter;
+      }
+    }
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      ++m_lhsIter;
+      ++m_rhsIter;
+      while (m_lhsIter && m_rhsIter && (m_lhsIter.index() != m_rhsIter.index()))
+      {
+        if (m_lhsIter.index() < m_rhsIter.index())
+          ++m_lhsIter;
+        else
+          ++m_rhsIter;
+      }
+      return *this;
+    }
+    
+    EIGEN_STRONG_INLINE Scalar value() const { return m_functor(m_lhsIter.value(), m_rhsIter.value()); }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_lhsIter.index(); }
+    EIGEN_STRONG_INLINE Index outer() const { return m_lhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); }
+    EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return (m_lhsIter && m_rhsIter); }
+
+  protected:
+    LhsIterator m_lhsIter;
+    RhsIterator m_rhsIter;
+    const BinaryOp& m_functor;
+  };
+  
+  
+  enum {
+    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    Flags = XprType::Flags
+  };
+  
+  explicit sparse_conjunction_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()), 
+      m_rhsImpl(xpr.rhs())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return (std::min)(m_lhsImpl.nonZerosEstimate(), m_rhsImpl.nonZerosEstimate());
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<LhsArg> m_lhsImpl;
+  evaluator<RhsArg> m_rhsImpl;
+};
+
+// "dense ^ sparse"
+template<typename XprType>
+struct sparse_conjunction_evaluator<XprType, IndexBased, IteratorBased>
+  : evaluator_base<XprType>
+{
+protected:
+  typedef typename XprType::Functor BinaryOp;
+  typedef typename XprType::Lhs LhsArg;
+  typedef typename XprType::Rhs RhsArg;
+  typedef evaluator<LhsArg> LhsEvaluator;
+  typedef typename evaluator<RhsArg>::InnerIterator  RhsIterator;
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename traits<XprType>::Scalar Scalar;
+public:
+
+  class InnerIterator
+  {
+    enum { IsRowMajor = (int(RhsArg::Flags)&RowMajorBit)==RowMajorBit };
+
+  public:
+    
+    EIGEN_STRONG_INLINE InnerIterator(const sparse_conjunction_evaluator& aEval, Index outer)
+      : m_lhsEval(aEval.m_lhsImpl), m_rhsIter(aEval.m_rhsImpl,outer), m_functor(aEval.m_functor), m_outer(outer)
+    {}
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      ++m_rhsIter;
+      return *this;
+    }
+
+    EIGEN_STRONG_INLINE Scalar value() const
+    { return m_functor(m_lhsEval.coeff(IsRowMajor?m_outer:m_rhsIter.index(),IsRowMajor?m_rhsIter.index():m_outer), m_rhsIter.value()); }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_rhsIter.index(); }
+    EIGEN_STRONG_INLINE Index outer() const { return m_rhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return m_rhsIter.row(); }
+    EIGEN_STRONG_INLINE Index col() const { return m_rhsIter.col(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return m_rhsIter; }
+    
+  protected:
+    const LhsEvaluator &m_lhsEval;
+    RhsIterator m_rhsIter;
+    const BinaryOp& m_functor;
+    const Index m_outer;
+  };
+  
+  
+  enum {
+    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    // Expose storage order of the sparse expression
+    Flags = (XprType::Flags & ~RowMajorBit) | (int(RhsArg::Flags)&RowMajorBit)
+  };
+  
+  explicit sparse_conjunction_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()), 
+      m_rhsImpl(xpr.rhs())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return m_rhsImpl.nonZerosEstimate();
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<LhsArg> m_lhsImpl;
+  evaluator<RhsArg> m_rhsImpl;
+};
+
+// "sparse ^ dense"
+template<typename XprType>
+struct sparse_conjunction_evaluator<XprType, IteratorBased, IndexBased>
+  : evaluator_base<XprType>
+{
+protected:
+  typedef typename XprType::Functor BinaryOp;
+  typedef typename XprType::Lhs LhsArg;
+  typedef typename XprType::Rhs RhsArg;
+  typedef typename evaluator<LhsArg>::InnerIterator LhsIterator;
+  typedef evaluator<RhsArg> RhsEvaluator;
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename traits<XprType>::Scalar Scalar;
+public:
+
+  class InnerIterator
+  {
+    enum { IsRowMajor = (int(LhsArg::Flags)&RowMajorBit)==RowMajorBit };
+
+  public:
+    
+    EIGEN_STRONG_INLINE InnerIterator(const sparse_conjunction_evaluator& aEval, Index outer)
+      : m_lhsIter(aEval.m_lhsImpl,outer), m_rhsEval(aEval.m_rhsImpl), m_functor(aEval.m_functor), m_outer(outer)
+    {}
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    {
+      ++m_lhsIter;
+      return *this;
+    }
+
+    EIGEN_STRONG_INLINE Scalar value() const
+    { return m_functor(m_lhsIter.value(),
+                       m_rhsEval.coeff(IsRowMajor?m_outer:m_lhsIter.index(),IsRowMajor?m_lhsIter.index():m_outer)); }
+
+    EIGEN_STRONG_INLINE StorageIndex index() const { return m_lhsIter.index(); }
+    EIGEN_STRONG_INLINE Index outer() const { return m_lhsIter.outer(); }
+    EIGEN_STRONG_INLINE Index row() const { return m_lhsIter.row(); }
+    EIGEN_STRONG_INLINE Index col() const { return m_lhsIter.col(); }
+
+    EIGEN_STRONG_INLINE operator bool() const { return m_lhsIter; }
+    
+  protected:
+    LhsIterator m_lhsIter;
+    const evaluator<RhsArg> &m_rhsEval;
+    const BinaryOp& m_functor;
+    const Index m_outer;
+  };
+  
+  
+  enum {
+    CoeffReadCost = evaluator<LhsArg>::CoeffReadCost + evaluator<RhsArg>::CoeffReadCost + functor_traits<BinaryOp>::Cost,
+    // Expose storage order of the sparse expression
+    Flags = (XprType::Flags & ~RowMajorBit) | (int(LhsArg::Flags)&RowMajorBit)
+  };
+  
+  explicit sparse_conjunction_evaluator(const XprType& xpr)
+    : m_functor(xpr.functor()),
+      m_lhsImpl(xpr.lhs()), 
+      m_rhsImpl(xpr.rhs())  
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<BinaryOp>::Cost);
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return m_lhsImpl.nonZerosEstimate();
+  }
+
+protected:
+  const BinaryOp m_functor;
+  evaluator<LhsArg> m_lhsImpl;
+  evaluator<RhsArg> m_rhsImpl;
+};
+
+}
+
+/***************************************************************************
+* Implementation of SparseMatrixBase and SparseCwise functions/operators
+***************************************************************************/
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator+=(const EigenBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator-=(const EigenBase<OtherDerived> &other)
+{
+  call_assignment(derived(), other.derived(), internal::assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE Derived &
+SparseMatrixBase<Derived>::operator-=(const SparseMatrixBase<OtherDerived> &other)
+{
+  return derived() = derived() - other.derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE Derived &
+SparseMatrixBase<Derived>::operator+=(const SparseMatrixBase<OtherDerived>& other)
+{
+  return derived() = derived() + other.derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator+=(const DiagonalBase<OtherDerived>& other)
+{
+  call_assignment_no_alias(derived(), other.derived(), internal::add_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+Derived& SparseMatrixBase<Derived>::operator-=(const DiagonalBase<OtherDerived>& other)
+{
+  call_assignment_no_alias(derived(), other.derived(), internal::sub_assign_op<Scalar,typename OtherDerived::Scalar>());
+  return derived();
+}
+    
+template<typename Derived>
+template<typename OtherDerived>
+EIGEN_STRONG_INLINE const typename SparseMatrixBase<Derived>::template CwiseProductDenseReturnType<OtherDerived>::Type
+SparseMatrixBase<Derived>::cwiseProduct(const MatrixBase<OtherDerived> &other) const
+{
+  return typename CwiseProductDenseReturnType<OtherDerived>::Type(derived(), other.derived());
+}
+
+template<typename DenseDerived, typename SparseDerived>
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_sum_op<typename DenseDerived::Scalar,typename SparseDerived::Scalar>, const DenseDerived, const SparseDerived>
+operator+(const MatrixBase<DenseDerived> &a, const SparseMatrixBase<SparseDerived> &b)
+{
+  return CwiseBinaryOp<internal::scalar_sum_op<typename DenseDerived::Scalar,typename SparseDerived::Scalar>, const DenseDerived, const SparseDerived>(a.derived(), b.derived());
+}
+
+template<typename SparseDerived, typename DenseDerived>
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_sum_op<typename SparseDerived::Scalar,typename DenseDerived::Scalar>, const SparseDerived, const DenseDerived>
+operator+(const SparseMatrixBase<SparseDerived> &a, const MatrixBase<DenseDerived> &b)
+{
+  return CwiseBinaryOp<internal::scalar_sum_op<typename SparseDerived::Scalar,typename DenseDerived::Scalar>, const SparseDerived, const DenseDerived>(a.derived(), b.derived());
+}
+
+template<typename DenseDerived, typename SparseDerived>
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_difference_op<typename DenseDerived::Scalar,typename SparseDerived::Scalar>, const DenseDerived, const SparseDerived>
+operator-(const MatrixBase<DenseDerived> &a, const SparseMatrixBase<SparseDerived> &b)
+{
+  return CwiseBinaryOp<internal::scalar_difference_op<typename DenseDerived::Scalar,typename SparseDerived::Scalar>, const DenseDerived, const SparseDerived>(a.derived(), b.derived());
+}
+
+template<typename SparseDerived, typename DenseDerived>
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_difference_op<typename SparseDerived::Scalar,typename DenseDerived::Scalar>, const SparseDerived, const DenseDerived>
+operator-(const SparseMatrixBase<SparseDerived> &a, const MatrixBase<DenseDerived> &b)
+{
+  return CwiseBinaryOp<internal::scalar_difference_op<typename SparseDerived::Scalar,typename DenseDerived::Scalar>, const SparseDerived, const DenseDerived>(a.derived(), b.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_CWISE_BINARY_OP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseUnaryOp.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseUnaryOp.h
new file mode 100644
index 0000000..ea79737
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseCwiseUnaryOp.h
@@ -0,0 +1,148 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_CWISE_UNARY_OP_H
+#define EIGEN_SPARSE_CWISE_UNARY_OP_H
+
+namespace Eigen { 
+
+namespace internal {
+  
+template<typename UnaryOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>
+  : public evaluator_base<CwiseUnaryOp<UnaryOp,ArgType> >
+{
+  public:
+    typedef CwiseUnaryOp<UnaryOp, ArgType> XprType;
+
+    class InnerIterator;
+    
+    enum {
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<UnaryOp>::Cost,
+      Flags = XprType::Flags
+    };
+    
+    explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression())
+    {
+      EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<UnaryOp>::Cost);
+      EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+    }
+    
+    inline Index nonZerosEstimate() const {
+      return m_argImpl.nonZerosEstimate();
+    }
+
+  protected:
+    typedef typename evaluator<ArgType>::InnerIterator        EvalIterator;
+    
+    const UnaryOp m_functor;
+    evaluator<ArgType> m_argImpl;
+};
+
+template<typename UnaryOp, typename ArgType>
+class unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>::InnerIterator
+    : public unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>::EvalIterator
+{
+    typedef typename XprType::Scalar Scalar;
+    typedef typename unary_evaluator<CwiseUnaryOp<UnaryOp,ArgType>, IteratorBased>::EvalIterator Base;
+  public:
+
+    EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, Index outer)
+      : Base(unaryOp.m_argImpl,outer), m_functor(unaryOp.m_functor)
+    {}
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    { Base::operator++(); return *this; }
+
+    EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); }
+
+  protected:
+    const UnaryOp m_functor;
+  private:
+    Scalar& valueRef();
+};
+
+template<typename ViewOp, typename ArgType>
+struct unary_evaluator<CwiseUnaryView<ViewOp,ArgType>, IteratorBased>
+  : public evaluator_base<CwiseUnaryView<ViewOp,ArgType> >
+{
+  public:
+    typedef CwiseUnaryView<ViewOp, ArgType> XprType;
+
+    class InnerIterator;
+    
+    enum {
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost + functor_traits<ViewOp>::Cost,
+      Flags = XprType::Flags
+    };
+    
+    explicit unary_evaluator(const XprType& op) : m_functor(op.functor()), m_argImpl(op.nestedExpression())
+    {
+      EIGEN_INTERNAL_CHECK_COST_VALUE(functor_traits<ViewOp>::Cost);
+      EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+    }
+
+  protected:
+    typedef typename evaluator<ArgType>::InnerIterator        EvalIterator;
+    
+    const ViewOp m_functor;
+    evaluator<ArgType> m_argImpl;
+};
+
+template<typename ViewOp, typename ArgType>
+class unary_evaluator<CwiseUnaryView<ViewOp,ArgType>, IteratorBased>::InnerIterator
+    : public unary_evaluator<CwiseUnaryView<ViewOp,ArgType>, IteratorBased>::EvalIterator
+{
+    typedef typename XprType::Scalar Scalar;
+    typedef typename unary_evaluator<CwiseUnaryView<ViewOp,ArgType>, IteratorBased>::EvalIterator Base;
+  public:
+
+    EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, Index outer)
+      : Base(unaryOp.m_argImpl,outer), m_functor(unaryOp.m_functor)
+    {}
+
+    EIGEN_STRONG_INLINE InnerIterator& operator++()
+    { Base::operator++(); return *this; }
+
+    EIGEN_STRONG_INLINE Scalar value() const { return m_functor(Base::value()); }
+    EIGEN_STRONG_INLINE Scalar& valueRef() { return m_functor(Base::valueRef()); }
+
+  protected:
+    const ViewOp m_functor;
+};
+
+} // end namespace internal
+
+template<typename Derived>
+EIGEN_STRONG_INLINE Derived&
+SparseMatrixBase<Derived>::operator*=(const Scalar& other)
+{
+  typedef typename internal::evaluator<Derived>::InnerIterator EvalIterator;
+  internal::evaluator<Derived> thisEval(derived());
+  for (Index j=0; j<outerSize(); ++j)
+    for (EvalIterator i(thisEval,j); i; ++i)
+      i.valueRef() *= other;
+  return derived();
+}
+
+template<typename Derived>
+EIGEN_STRONG_INLINE Derived&
+SparseMatrixBase<Derived>::operator/=(const Scalar& other)
+{
+  typedef typename internal::evaluator<Derived>::InnerIterator EvalIterator;
+  internal::evaluator<Derived> thisEval(derived());
+  for (Index j=0; j<outerSize(); ++j)
+    for (EvalIterator i(thisEval,j); i; ++i)
+      i.valueRef() /= other;
+  return derived();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_CWISE_UNARY_OP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDenseProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDenseProduct.h
new file mode 100644
index 0000000..0547db5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDenseProduct.h
@@ -0,0 +1,320 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEDENSEPRODUCT_H
+#define EIGEN_SPARSEDENSEPRODUCT_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template <> struct product_promote_storage_type<Sparse,Dense, OuterProduct> { typedef Sparse ret; };
+template <> struct product_promote_storage_type<Dense,Sparse, OuterProduct> { typedef Sparse ret; };
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType,
+         typename AlphaType,
+         int LhsStorageOrder = ((SparseLhsType::Flags&RowMajorBit)==RowMajorBit) ? RowMajor : ColMajor,
+         bool ColPerCol = ((DenseRhsType::Flags&RowMajorBit)==0) || DenseRhsType::ColsAtCompileTime==1>
+struct sparse_time_dense_product_impl;
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType>
+struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, typename DenseResType::Scalar, RowMajor, true>
+{
+  typedef typename internal::remove_all<SparseLhsType>::type Lhs;
+  typedef typename internal::remove_all<DenseRhsType>::type Rhs;
+  typedef typename internal::remove_all<DenseResType>::type Res;
+  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  typedef evaluator<Lhs> LhsEval;
+  static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
+  {
+    LhsEval lhsEval(lhs);
+    
+    Index n = lhs.outerSize();
+#ifdef EIGEN_HAS_OPENMP
+    Eigen::initParallel();
+    Index threads = Eigen::nbThreads();
+#endif
+    
+    for(Index c=0; c<rhs.cols(); ++c)
+    {
+#ifdef EIGEN_HAS_OPENMP
+      // This 20000 threshold has been found experimentally on 2D and 3D Poisson problems.
+      // It basically represents the minimal amount of work to be done to be worth it.
+      if(threads>1 && lhsEval.nonZerosEstimate() > 20000)
+      {
+        #pragma omp parallel for schedule(dynamic,(n+threads*4-1)/(threads*4)) num_threads(threads)
+        for(Index i=0; i<n; ++i)
+          processRow(lhsEval,rhs,res,alpha,i,c);
+      }
+      else
+#endif
+      {
+        for(Index i=0; i<n; ++i)
+          processRow(lhsEval,rhs,res,alpha,i,c);
+      }
+    }
+  }
+  
+  static void processRow(const LhsEval& lhsEval, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha, Index i, Index col)
+  {
+    typename Res::Scalar tmp(0);
+    for(LhsInnerIterator it(lhsEval,i); it ;++it)
+      tmp += it.value() * rhs.coeff(it.index(),col);
+    res.coeffRef(i,col) += alpha * tmp;
+  }
+  
+};
+
+// FIXME: what is the purpose of the following specialization? Is it for the BlockedSparse format?
+// -> let's disable it for now as it is conflicting with generic scalar*matrix and matrix*scalar operators
+// template<typename T1, typename T2/*, int _Options, typename _StrideType*/>
+// struct ScalarBinaryOpTraits<T1, Ref<T2/*, _Options, _StrideType*/> >
+// {
+//   enum {
+//     Defined = 1
+//   };
+//   typedef typename CwiseUnaryOp<scalar_multiple2_op<T1, typename T2::Scalar>, T2>::PlainObject ReturnType;
+// };
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType, typename AlphaType>
+struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, AlphaType, ColMajor, true>
+{
+  typedef typename internal::remove_all<SparseLhsType>::type Lhs;
+  typedef typename internal::remove_all<DenseRhsType>::type Rhs;
+  typedef typename internal::remove_all<DenseResType>::type Res;
+  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
+  {
+    evaluator<Lhs> lhsEval(lhs);
+    for(Index c=0; c<rhs.cols(); ++c)
+    {
+      for(Index j=0; j<lhs.outerSize(); ++j)
+      {
+//        typename Res::Scalar rhs_j = alpha * rhs.coeff(j,c);
+        typename ScalarBinaryOpTraits<AlphaType, typename Rhs::Scalar>::ReturnType rhs_j(alpha * rhs.coeff(j,c));
+        for(LhsInnerIterator it(lhsEval,j); it ;++it)
+          res.coeffRef(it.index(),c) += it.value() * rhs_j;
+      }
+    }
+  }
+};
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType>
+struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, typename DenseResType::Scalar, RowMajor, false>
+{
+  typedef typename internal::remove_all<SparseLhsType>::type Lhs;
+  typedef typename internal::remove_all<DenseRhsType>::type Rhs;
+  typedef typename internal::remove_all<DenseResType>::type Res;
+  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
+  {
+    evaluator<Lhs> lhsEval(lhs);
+    for(Index j=0; j<lhs.outerSize(); ++j)
+    {
+      typename Res::RowXpr res_j(res.row(j));
+      for(LhsInnerIterator it(lhsEval,j); it ;++it)
+        res_j += (alpha*it.value()) * rhs.row(it.index());
+    }
+  }
+};
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType>
+struct sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, typename DenseResType::Scalar, ColMajor, false>
+{
+  typedef typename internal::remove_all<SparseLhsType>::type Lhs;
+  typedef typename internal::remove_all<DenseRhsType>::type Rhs;
+  typedef typename internal::remove_all<DenseResType>::type Res;
+  typedef typename evaluator<Lhs>::InnerIterator LhsInnerIterator;
+  static void run(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const typename Res::Scalar& alpha)
+  {
+    evaluator<Lhs> lhsEval(lhs);
+    for(Index j=0; j<lhs.outerSize(); ++j)
+    {
+      typename Rhs::ConstRowXpr rhs_j(rhs.row(j));
+      for(LhsInnerIterator it(lhsEval,j); it ;++it)
+        res.row(it.index()) += (alpha*it.value()) * rhs_j;
+    }
+  }
+};
+
+template<typename SparseLhsType, typename DenseRhsType, typename DenseResType,typename AlphaType>
+inline void sparse_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
+{
+  sparse_time_dense_product_impl<SparseLhsType,DenseRhsType,DenseResType, AlphaType>::run(lhs, rhs, res, alpha);
+}
+
+} // end namespace internal
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, SparseShape, DenseShape, ProductType>
+ : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,SparseShape,DenseShape,ProductType> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    typedef typename nested_eval<Lhs,((Rhs::Flags&RowMajorBit)==0) ? 1 : Rhs::ColsAtCompileTime>::type LhsNested;
+    typedef typename nested_eval<Rhs,((Lhs::Flags&RowMajorBit)==0) ? 1 : Dynamic>::type RhsNested;
+    LhsNested lhsNested(lhs);
+    RhsNested rhsNested(rhs);
+    internal::sparse_time_dense_product(lhsNested, rhsNested, dst, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, SparseTriangularShape, DenseShape, ProductType>
+  : generic_product_impl<Lhs, Rhs, SparseShape, DenseShape, ProductType>
+{};
+
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, DenseShape, SparseShape, ProductType>
+  : generic_product_impl_base<Lhs,Rhs,generic_product_impl<Lhs,Rhs,DenseShape,SparseShape,ProductType> >
+{
+  typedef typename Product<Lhs,Rhs>::Scalar Scalar;
+  
+  template<typename Dst>
+  static void scaleAndAddTo(Dst& dst, const Lhs& lhs, const Rhs& rhs, const Scalar& alpha)
+  {
+    typedef typename nested_eval<Lhs,((Rhs::Flags&RowMajorBit)==0) ? Dynamic : 1>::type LhsNested;
+    typedef typename nested_eval<Rhs,((Lhs::Flags&RowMajorBit)==RowMajorBit) ? 1 : Lhs::RowsAtCompileTime>::type RhsNested;
+    LhsNested lhsNested(lhs);
+    RhsNested rhsNested(rhs);
+    
+    // transpose everything
+    Transpose<Dst> dstT(dst);
+    internal::sparse_time_dense_product(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha);
+  }
+};
+
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, DenseShape, SparseTriangularShape, ProductType>
+  : generic_product_impl<Lhs, Rhs, DenseShape, SparseShape, ProductType>
+{};
+
+template<typename LhsT, typename RhsT, bool NeedToTranspose>
+struct sparse_dense_outer_product_evaluator
+{
+protected:
+  typedef typename conditional<NeedToTranspose,RhsT,LhsT>::type Lhs1;
+  typedef typename conditional<NeedToTranspose,LhsT,RhsT>::type ActualRhs;
+  typedef Product<LhsT,RhsT,DefaultProduct> ProdXprType;
+  
+  // if the actual left-hand side is a dense vector,
+  // then build a sparse-view so that we can seamlessly iterate over it.
+  typedef typename conditional<is_same<typename internal::traits<Lhs1>::StorageKind,Sparse>::value,
+            Lhs1, SparseView<Lhs1> >::type ActualLhs;
+  typedef typename conditional<is_same<typename internal::traits<Lhs1>::StorageKind,Sparse>::value,
+            Lhs1 const&, SparseView<Lhs1> >::type LhsArg;
+            
+  typedef evaluator<ActualLhs> LhsEval;
+  typedef evaluator<ActualRhs> RhsEval;
+  typedef typename evaluator<ActualLhs>::InnerIterator LhsIterator;
+  typedef typename ProdXprType::Scalar Scalar;
+  
+public:
+  enum {
+    Flags = NeedToTranspose ? RowMajorBit : 0,
+    CoeffReadCost = HugeCost
+  };
+  
+  class InnerIterator : public LhsIterator
+  {
+  public:
+    InnerIterator(const sparse_dense_outer_product_evaluator &xprEval, Index outer)
+      : LhsIterator(xprEval.m_lhsXprImpl, 0),
+        m_outer(outer),
+        m_empty(false),
+        m_factor(get(xprEval.m_rhsXprImpl, outer, typename internal::traits<ActualRhs>::StorageKind() ))
+    {}
+    
+    EIGEN_STRONG_INLINE Index outer() const { return m_outer; }
+    EIGEN_STRONG_INLINE Index row()   const { return NeedToTranspose ? m_outer : LhsIterator::index(); }
+    EIGEN_STRONG_INLINE Index col()   const { return NeedToTranspose ? LhsIterator::index() : m_outer; }
+
+    EIGEN_STRONG_INLINE Scalar value() const { return LhsIterator::value() * m_factor; }
+    EIGEN_STRONG_INLINE operator bool() const { return LhsIterator::operator bool() && (!m_empty); }
+    
+  protected:
+    Scalar get(const RhsEval &rhs, Index outer, Dense = Dense()) const
+    {
+      return rhs.coeff(outer);
+    }
+    
+    Scalar get(const RhsEval &rhs, Index outer, Sparse = Sparse())
+    {
+      typename RhsEval::InnerIterator it(rhs, outer);
+      if (it && it.index()==0 && it.value()!=Scalar(0))
+        return it.value();
+      m_empty = true;
+      return Scalar(0);
+    }
+    
+    Index m_outer;
+    bool m_empty;
+    Scalar m_factor;
+  };
+  
+  sparse_dense_outer_product_evaluator(const Lhs1 &lhs, const ActualRhs &rhs)
+     : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  // transpose case
+  sparse_dense_outer_product_evaluator(const ActualRhs &rhs, const Lhs1 &lhs)
+     : m_lhs(lhs), m_lhsXprImpl(m_lhs), m_rhsXprImpl(rhs)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+    
+protected:
+  const LhsArg m_lhs;
+  evaluator<ActualLhs> m_lhsXprImpl;
+  evaluator<ActualRhs> m_rhsXprImpl;
+};
+
+// sparse * dense outer product
+template<typename Lhs, typename Rhs>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, OuterProduct, SparseShape, DenseShape>
+  : sparse_dense_outer_product_evaluator<Lhs,Rhs, Lhs::IsRowMajor>
+{
+  typedef sparse_dense_outer_product_evaluator<Lhs,Rhs, Lhs::IsRowMajor> Base;
+  
+  typedef Product<Lhs, Rhs> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+
+  explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.lhs(), xpr.rhs())
+  {}
+  
+};
+
+template<typename Lhs, typename Rhs>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, OuterProduct, DenseShape, SparseShape>
+  : sparse_dense_outer_product_evaluator<Lhs,Rhs, Rhs::IsRowMajor>
+{
+  typedef sparse_dense_outer_product_evaluator<Lhs,Rhs, Rhs::IsRowMajor> Base;
+  
+  typedef Product<Lhs, Rhs> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+
+  explicit product_evaluator(const XprType& xpr)
+    : Base(xpr.lhs(), xpr.rhs())
+  {}
+  
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEDENSEPRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDiagonalProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDiagonalProduct.h
new file mode 100644
index 0000000..941c03b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDiagonalProduct.h
@@ -0,0 +1,138 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_DIAGONAL_PRODUCT_H
+#define EIGEN_SPARSE_DIAGONAL_PRODUCT_H
+
+namespace Eigen { 
+
+// The product of a diagonal matrix with a sparse matrix can be easily
+// implemented using expression template.
+// We have two consider very different cases:
+// 1 - diag * row-major sparse
+//     => each inner vector <=> scalar * sparse vector product
+//     => so we can reuse CwiseUnaryOp::InnerIterator
+// 2 - diag * col-major sparse
+//     => each inner vector <=> densevector * sparse vector cwise product
+//     => again, we can reuse specialization of CwiseBinaryOp::InnerIterator
+//        for that particular case
+// The two other cases are symmetric.
+
+namespace internal {
+
+enum {
+  SDP_AsScalarProduct,
+  SDP_AsCwiseProduct
+};
+  
+template<typename SparseXprType, typename DiagonalCoeffType, int SDP_Tag>
+struct sparse_diagonal_product_evaluator;
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, ProductTag, DiagonalShape, SparseShape>
+  : public sparse_diagonal_product_evaluator<Rhs, typename Lhs::DiagonalVectorType, Rhs::Flags&RowMajorBit?SDP_AsScalarProduct:SDP_AsCwiseProduct>
+{
+  typedef Product<Lhs, Rhs, DefaultProduct> XprType;
+  enum { CoeffReadCost = HugeCost, Flags = Rhs::Flags&RowMajorBit, Alignment = 0 }; // FIXME CoeffReadCost & Flags
+  
+  typedef sparse_diagonal_product_evaluator<Rhs, typename Lhs::DiagonalVectorType, Rhs::Flags&RowMajorBit?SDP_AsScalarProduct:SDP_AsCwiseProduct> Base;
+  explicit product_evaluator(const XprType& xpr) : Base(xpr.rhs(), xpr.lhs().diagonal()) {}
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, DefaultProduct>, ProductTag, SparseShape, DiagonalShape>
+  : public sparse_diagonal_product_evaluator<Lhs, Transpose<const typename Rhs::DiagonalVectorType>, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct>
+{
+  typedef Product<Lhs, Rhs, DefaultProduct> XprType;
+  enum { CoeffReadCost = HugeCost, Flags = Lhs::Flags&RowMajorBit, Alignment = 0 }; // FIXME CoeffReadCost & Flags
+  
+  typedef sparse_diagonal_product_evaluator<Lhs, Transpose<const typename Rhs::DiagonalVectorType>, Lhs::Flags&RowMajorBit?SDP_AsCwiseProduct:SDP_AsScalarProduct> Base;
+  explicit product_evaluator(const XprType& xpr) : Base(xpr.lhs(), xpr.rhs().diagonal().transpose()) {}
+};
+
+template<typename SparseXprType, typename DiagonalCoeffType>
+struct sparse_diagonal_product_evaluator<SparseXprType, DiagonalCoeffType, SDP_AsScalarProduct>
+{
+protected:
+  typedef typename evaluator<SparseXprType>::InnerIterator SparseXprInnerIterator;
+  typedef typename SparseXprType::Scalar Scalar;
+  
+public:
+  class InnerIterator : public SparseXprInnerIterator
+  {
+  public:
+    InnerIterator(const sparse_diagonal_product_evaluator &xprEval, Index outer)
+      : SparseXprInnerIterator(xprEval.m_sparseXprImpl, outer),
+        m_coeff(xprEval.m_diagCoeffImpl.coeff(outer))
+    {}
+    
+    EIGEN_STRONG_INLINE Scalar value() const { return m_coeff * SparseXprInnerIterator::value(); }
+  protected:
+    typename DiagonalCoeffType::Scalar m_coeff;
+  };
+  
+  sparse_diagonal_product_evaluator(const SparseXprType &sparseXpr, const DiagonalCoeffType &diagCoeff)
+    : m_sparseXprImpl(sparseXpr), m_diagCoeffImpl(diagCoeff)
+  {}
+
+  Index nonZerosEstimate() const { return m_sparseXprImpl.nonZerosEstimate(); }
+    
+protected:
+  evaluator<SparseXprType> m_sparseXprImpl;
+  evaluator<DiagonalCoeffType> m_diagCoeffImpl;
+};
+
+
+template<typename SparseXprType, typename DiagCoeffType>
+struct sparse_diagonal_product_evaluator<SparseXprType, DiagCoeffType, SDP_AsCwiseProduct>
+{
+  typedef typename SparseXprType::Scalar Scalar;
+  typedef typename SparseXprType::StorageIndex StorageIndex;
+  
+  typedef typename nested_eval<DiagCoeffType,SparseXprType::IsRowMajor ? SparseXprType::RowsAtCompileTime
+                                                                       : SparseXprType::ColsAtCompileTime>::type DiagCoeffNested;
+  
+  class InnerIterator
+  {
+    typedef typename evaluator<SparseXprType>::InnerIterator SparseXprIter;
+  public:
+    InnerIterator(const sparse_diagonal_product_evaluator &xprEval, Index outer)
+      : m_sparseIter(xprEval.m_sparseXprEval, outer), m_diagCoeffNested(xprEval.m_diagCoeffNested)
+    {}
+    
+    inline Scalar value() const { return m_sparseIter.value() * m_diagCoeffNested.coeff(index()); }
+    inline StorageIndex index() const  { return m_sparseIter.index(); }
+    inline Index outer() const  { return m_sparseIter.outer(); }
+    inline Index col() const    { return SparseXprType::IsRowMajor ? m_sparseIter.index() : m_sparseIter.outer(); }
+    inline Index row() const    { return SparseXprType::IsRowMajor ? m_sparseIter.outer() : m_sparseIter.index(); }
+    
+    EIGEN_STRONG_INLINE InnerIterator& operator++() { ++m_sparseIter; return *this; }
+    inline operator bool() const  { return m_sparseIter; }
+    
+  protected:
+    SparseXprIter m_sparseIter;
+    DiagCoeffNested m_diagCoeffNested;
+  };
+  
+  sparse_diagonal_product_evaluator(const SparseXprType &sparseXpr, const DiagCoeffType &diagCoeff)
+    : m_sparseXprEval(sparseXpr), m_diagCoeffNested(diagCoeff)
+  {}
+
+  Index nonZerosEstimate() const { return m_sparseXprEval.nonZerosEstimate(); }
+    
+protected:
+  evaluator<SparseXprType> m_sparseXprEval;
+  DiagCoeffNested m_diagCoeffNested;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_DIAGONAL_PRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDot.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDot.h
new file mode 100644
index 0000000..38bc4aa
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseDot.h
@@ -0,0 +1,98 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_DOT_H
+#define EIGEN_SPARSE_DOT_H
+
+namespace Eigen { 
+
+template<typename Derived>
+template<typename OtherDerived>
+typename internal::traits<Derived>::Scalar
+SparseMatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+  EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
+  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+  eigen_assert(size() == other.size());
+  eigen_assert(other.size()>0 && "you are using a non initialized vector");
+
+  internal::evaluator<Derived> thisEval(derived());
+  typename internal::evaluator<Derived>::InnerIterator i(thisEval, 0);
+  Scalar res(0);
+  while (i)
+  {
+    res += numext::conj(i.value()) * other.coeff(i.index());
+    ++i;
+  }
+  return res;
+}
+
+template<typename Derived>
+template<typename OtherDerived>
+typename internal::traits<Derived>::Scalar
+SparseMatrixBase<Derived>::dot(const SparseMatrixBase<OtherDerived>& other) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(OtherDerived)
+  EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(Derived,OtherDerived)
+  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
+    YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+  eigen_assert(size() == other.size());
+
+  internal::evaluator<Derived> thisEval(derived());
+  typename internal::evaluator<Derived>::InnerIterator i(thisEval, 0);
+  
+  internal::evaluator<OtherDerived>  otherEval(other.derived());
+  typename internal::evaluator<OtherDerived>::InnerIterator j(otherEval, 0);
+
+  Scalar res(0);
+  while (i && j)
+  {
+    if (i.index()==j.index())
+    {
+      res += numext::conj(i.value()) * j.value();
+      ++i; ++j;
+    }
+    else if (i.index()<j.index())
+      ++i;
+    else
+      ++j;
+  }
+  return res;
+}
+
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+SparseMatrixBase<Derived>::squaredNorm() const
+{
+  return numext::real((*this).cwiseAbs2().sum());
+}
+
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+SparseMatrixBase<Derived>::norm() const
+{
+  using std::sqrt;
+  return sqrt(squaredNorm());
+}
+
+template<typename Derived>
+inline typename NumTraits<typename internal::traits<Derived>::Scalar>::Real
+SparseMatrixBase<Derived>::blueNorm() const
+{
+  return internal::blueNorm_impl(*this);
+}
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_DOT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseFuzzy.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseFuzzy.h
new file mode 100644
index 0000000..7d47eb9
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseFuzzy.h
@@ -0,0 +1,29 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_FUZZY_H
+#define EIGEN_SPARSE_FUZZY_H
+
+namespace Eigen {
+  
+template<typename Derived>
+template<typename OtherDerived>
+bool SparseMatrixBase<Derived>::isApprox(const SparseMatrixBase<OtherDerived>& other, const RealScalar &prec) const
+{
+  const typename internal::nested_eval<Derived,2,PlainObject>::type actualA(derived());
+  typename internal::conditional<bool(IsRowMajor)==bool(OtherDerived::IsRowMajor),
+    const typename internal::nested_eval<OtherDerived,2,PlainObject>::type,
+    const PlainObject>::type actualB(other.derived());
+
+  return (actualA - actualB).squaredNorm() <= prec * prec * numext::mini(actualA.squaredNorm(), actualB.squaredNorm());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_FUZZY_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMap.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMap.h
new file mode 100644
index 0000000..f99be33
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMap.h
@@ -0,0 +1,305 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_MAP_H
+#define EIGEN_SPARSE_MAP_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct traits<Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : public traits<SparseMatrix<MatScalar,MatOptions,MatIndex> >
+{
+  typedef SparseMatrix<MatScalar,MatOptions,MatIndex> PlainObjectType;
+  typedef traits<PlainObjectType> TraitsBase;
+  enum {
+    Flags = TraitsBase::Flags & (~NestByRefBit)
+  };
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct traits<Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : public traits<SparseMatrix<MatScalar,MatOptions,MatIndex> >
+{
+  typedef SparseMatrix<MatScalar,MatOptions,MatIndex> PlainObjectType;
+  typedef traits<PlainObjectType> TraitsBase;
+  enum {
+    Flags = TraitsBase::Flags & (~ (NestByRefBit | LvalueBit))
+  };
+};
+
+} // end namespace internal
+
+template<typename Derived,
+         int Level = internal::accessors_level<Derived>::has_write_access ? WriteAccessors : ReadOnlyAccessors
+> class SparseMapBase;
+
+/** \ingroup SparseCore_Module
+  * class SparseMapBase
+  * \brief Common base class for Map and Ref instance of sparse matrix and vector.
+  */
+template<typename Derived>
+class SparseMapBase<Derived,ReadOnlyAccessors>
+  : public SparseCompressedBase<Derived>
+{
+  public:
+    typedef SparseCompressedBase<Derived> Base;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::StorageIndex StorageIndex;
+    enum { IsRowMajor = Base::IsRowMajor };
+    using Base::operator=;
+  protected:
+    
+    typedef typename internal::conditional<
+                         bool(internal::is_lvalue<Derived>::value),
+                         Scalar *, const Scalar *>::type ScalarPointer;
+    typedef typename internal::conditional<
+                         bool(internal::is_lvalue<Derived>::value),
+                         StorageIndex *, const StorageIndex *>::type IndexPointer;
+
+    Index   m_outerSize;
+    Index   m_innerSize;
+    Array<StorageIndex,2,1>  m_zero_nnz;
+    IndexPointer  m_outerIndex;
+    IndexPointer  m_innerIndices;
+    ScalarPointer m_values;
+    IndexPointer  m_innerNonZeros;
+
+  public:
+
+    /** \copydoc SparseMatrixBase::rows() */
+    inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; }
+    /** \copydoc SparseMatrixBase::cols() */
+    inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; }
+    /** \copydoc SparseMatrixBase::innerSize() */
+    inline Index innerSize() const { return m_innerSize; }
+    /** \copydoc SparseMatrixBase::outerSize() */
+    inline Index outerSize() const { return m_outerSize; }
+    /** \copydoc SparseCompressedBase::nonZeros */
+    inline Index nonZeros() const { return m_zero_nnz[1]; }
+    
+    /** \copydoc SparseCompressedBase::isCompressed */
+    bool isCompressed() const { return m_innerNonZeros==0; }
+
+    //----------------------------------------
+    // direct access interface
+    /** \copydoc SparseMatrix::valuePtr */
+    inline const Scalar* valuePtr() const { return m_values; }
+    /** \copydoc SparseMatrix::innerIndexPtr */
+    inline const StorageIndex* innerIndexPtr() const { return m_innerIndices; }
+    /** \copydoc SparseMatrix::outerIndexPtr */
+    inline const StorageIndex* outerIndexPtr() const { return m_outerIndex; }
+    /** \copydoc SparseMatrix::innerNonZeroPtr */
+    inline const StorageIndex* innerNonZeroPtr() const { return m_innerNonZeros; }
+    //----------------------------------------
+
+    /** \copydoc SparseMatrix::coeff */
+    inline Scalar coeff(Index row, Index col) const
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+
+      Index start = m_outerIndex[outer];
+      Index end = isCompressed() ? m_outerIndex[outer+1] : start + m_innerNonZeros[outer];
+      if (start==end)
+        return Scalar(0);
+      else if (end>0 && inner==m_innerIndices[end-1])
+        return m_values[end-1];
+      // ^^  optimization: let's first check if it is the last coefficient
+      // (very common in high level algorithms)
+
+      const StorageIndex* r = std::lower_bound(&m_innerIndices[start],&m_innerIndices[end-1],inner);
+      const Index id = r-&m_innerIndices[0];
+      return ((*r==inner) && (id<end)) ? m_values[id] : Scalar(0);
+    }
+
+    inline SparseMapBase(Index rows, Index cols, Index nnz, IndexPointer outerIndexPtr, IndexPointer innerIndexPtr,
+                              ScalarPointer valuePtr, IndexPointer innerNonZerosPtr = 0)
+      : m_outerSize(IsRowMajor?rows:cols), m_innerSize(IsRowMajor?cols:rows), m_zero_nnz(0,internal::convert_index<StorageIndex>(nnz)), m_outerIndex(outerIndexPtr),
+        m_innerIndices(innerIndexPtr), m_values(valuePtr), m_innerNonZeros(innerNonZerosPtr)
+    {}
+
+    // for vectors
+    inline SparseMapBase(Index size, Index nnz, IndexPointer innerIndexPtr, ScalarPointer valuePtr)
+      : m_outerSize(1), m_innerSize(size), m_zero_nnz(0,internal::convert_index<StorageIndex>(nnz)), m_outerIndex(m_zero_nnz.data()),
+        m_innerIndices(innerIndexPtr), m_values(valuePtr), m_innerNonZeros(0)
+    {}
+
+    /** Empty destructor */
+    inline ~SparseMapBase() {}
+
+  protected:
+    inline SparseMapBase() {}
+};
+
+/** \ingroup SparseCore_Module
+  * class SparseMapBase
+  * \brief Common base class for writable Map and Ref instance of sparse matrix and vector.
+  */
+template<typename Derived>
+class SparseMapBase<Derived,WriteAccessors>
+  : public SparseMapBase<Derived,ReadOnlyAccessors>
+{
+    typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase;
+    
+  public:
+    typedef SparseMapBase<Derived, ReadOnlyAccessors> Base;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::StorageIndex StorageIndex;
+    enum { IsRowMajor = Base::IsRowMajor };
+    
+    using Base::operator=;
+
+  public:
+    
+    //----------------------------------------
+    // direct access interface
+    using Base::valuePtr;
+    using Base::innerIndexPtr;
+    using Base::outerIndexPtr;
+    using Base::innerNonZeroPtr;
+    /** \copydoc SparseMatrix::valuePtr */
+    inline Scalar* valuePtr()              { return Base::m_values; }
+    /** \copydoc SparseMatrix::innerIndexPtr */
+    inline StorageIndex* innerIndexPtr()   { return Base::m_innerIndices; }
+    /** \copydoc SparseMatrix::outerIndexPtr */
+    inline StorageIndex* outerIndexPtr()   { return Base::m_outerIndex; }
+    /** \copydoc SparseMatrix::innerNonZeroPtr */
+    inline StorageIndex* innerNonZeroPtr() { return Base::m_innerNonZeros; }
+    //----------------------------------------
+
+    /** \copydoc SparseMatrix::coeffRef */
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+
+      Index start = Base::m_outerIndex[outer];
+      Index end = Base::isCompressed() ? Base::m_outerIndex[outer+1] : start + Base::m_innerNonZeros[outer];
+      eigen_assert(end>=start && "you probably called coeffRef on a non finalized matrix");
+      eigen_assert(end>start && "coeffRef cannot be called on a zero coefficient");
+      StorageIndex* r = std::lower_bound(&Base::m_innerIndices[start],&Base::m_innerIndices[end],inner);
+      const Index id = r - &Base::m_innerIndices[0];
+      eigen_assert((*r==inner) && (id<end) && "coeffRef cannot be called on a zero coefficient");
+      return const_cast<Scalar*>(Base::m_values)[id];
+    }
+    
+    inline SparseMapBase(Index rows, Index cols, Index nnz, StorageIndex* outerIndexPtr, StorageIndex* innerIndexPtr,
+                         Scalar* valuePtr, StorageIndex* innerNonZerosPtr = 0)
+      : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr)
+    {}
+
+    // for vectors
+    inline SparseMapBase(Index size, Index nnz, StorageIndex* innerIndexPtr, Scalar* valuePtr)
+      : Base(size, nnz, innerIndexPtr, valuePtr)
+    {}
+
+    /** Empty destructor */
+    inline ~SparseMapBase() {}
+
+  protected:
+    inline SparseMapBase() {}
+};
+
+/** \ingroup SparseCore_Module
+  *
+  * \brief Specialization of class Map for SparseMatrix-like storage.
+  *
+  * \tparam SparseMatrixType the equivalent sparse matrix type of the referenced data, it must be a template instance of class SparseMatrix.
+  *
+  * \sa class Map, class SparseMatrix, class Ref<SparseMatrixType,Options>
+  */
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType>
+  : public SparseMapBase<Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+#else
+template<typename SparseMatrixType>
+class Map<SparseMatrixType>
+  : public SparseMapBase<Derived,WriteAccessors>
+#endif
+{
+  public:
+    typedef SparseMapBase<Map> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Map)
+    enum { IsRowMajor = Base::IsRowMajor };
+
+  public:
+
+    /** Constructs a read-write Map to a sparse matrix of size \a rows x \a cols, containing \a nnz non-zero coefficients,
+      * stored as a sparse format as defined by the pointers \a outerIndexPtr, \a innerIndexPtr, and \a valuePtr.
+      * If the optional parameter \a innerNonZerosPtr is the null pointer, then a standard compressed format is assumed.
+      *
+      * This constructor is available only if \c SparseMatrixType is non-const.
+      *
+      * More details on the expected storage schemes are given in the \ref TutorialSparse "manual pages".
+      */
+    inline Map(Index rows, Index cols, Index nnz, StorageIndex* outerIndexPtr,
+               StorageIndex* innerIndexPtr, Scalar* valuePtr, StorageIndex* innerNonZerosPtr = 0)
+      : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr)
+    {}
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** Empty destructor */
+    inline ~Map() {}
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType>
+  : public SparseMapBase<Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+{
+  public:
+    typedef SparseMapBase<Map> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Map)
+    enum { IsRowMajor = Base::IsRowMajor };
+
+  public:
+#endif
+    /** This is the const version of the above constructor.
+      *
+      * This constructor is available only if \c SparseMatrixType is const, e.g.:
+      * \code Map<const SparseMatrix<double> >  \endcode
+      */
+    inline Map(Index rows, Index cols, Index nnz, const StorageIndex* outerIndexPtr,
+               const StorageIndex* innerIndexPtr, const Scalar* valuePtr, const StorageIndex* innerNonZerosPtr = 0)
+      : Base(rows, cols, nnz, outerIndexPtr, innerIndexPtr, valuePtr, innerNonZerosPtr)
+    {}
+
+    /** Empty destructor */
+    inline ~Map() {}
+};
+
+namespace internal {
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Map<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;  
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Map<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;  
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_MAP_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrix.h
new file mode 100644
index 0000000..323c232
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrix.h
@@ -0,0 +1,1403 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEMATRIX_H
+#define EIGEN_SPARSEMATRIX_H
+
+namespace Eigen { 
+
+/** \ingroup SparseCore_Module
+  *
+  * \class SparseMatrix
+  *
+  * \brief A versatible sparse matrix representation
+  *
+  * This class implements a more versatile variants of the common \em compressed row/column storage format.
+  * Each colmun's (resp. row) non zeros are stored as a pair of value with associated row (resp. colmiun) index.
+  * All the non zeros are stored in a single large buffer. Unlike the \em compressed format, there might be extra
+  * space inbetween the nonzeros of two successive colmuns (resp. rows) such that insertion of new non-zero
+  * can be done with limited memory reallocation and copies.
+  *
+  * A call to the function makeCompressed() turns the matrix into the standard \em compressed format
+  * compatible with many library.
+  *
+  * More details on this storage sceheme are given in the \ref TutorialSparse "manual pages".
+  *
+  * \tparam _Scalar the scalar type, i.e. the type of the coefficients
+  * \tparam _Options Union of bit flags controlling the storage scheme. Currently the only possibility
+  *                 is ColMajor or RowMajor. The default is 0 which means column-major.
+  * \tparam _StorageIndex the type of the indices. It has to be a \b signed type (e.g., short, int, std::ptrdiff_t). Default is \c int.
+  *
+  * \warning In %Eigen 3.2, the undocumented type \c SparseMatrix::Index was improperly defined as the storage index type (e.g., int),
+  *          whereas it is now (starting from %Eigen 3.3) deprecated and always defined as Eigen::Index.
+  *          Codes making use of \c SparseMatrix::Index, might thus likely have to be changed to use \c SparseMatrix::StorageIndex instead.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEMATRIX_PLUGIN.
+  */
+
+namespace internal {
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct traits<SparseMatrix<_Scalar, _Options, _StorageIndex> >
+{
+  typedef _Scalar Scalar;
+  typedef _StorageIndex StorageIndex;
+  typedef Sparse StorageKind;
+  typedef MatrixXpr XprKind;
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = Dynamic,
+    MaxRowsAtCompileTime = Dynamic,
+    MaxColsAtCompileTime = Dynamic,
+    Flags = _Options | NestByRefBit | LvalueBit | CompressedAccessBit,
+    SupportedAccessPatterns = InnerRandomAccessPattern
+  };
+};
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int DiagIndex>
+struct traits<Diagonal<SparseMatrix<_Scalar, _Options, _StorageIndex>, DiagIndex> >
+{
+  typedef SparseMatrix<_Scalar, _Options, _StorageIndex> MatrixType;
+  typedef typename ref_selector<MatrixType>::type MatrixTypeNested;
+  typedef typename remove_reference<MatrixTypeNested>::type _MatrixTypeNested;
+
+  typedef _Scalar Scalar;
+  typedef Dense StorageKind;
+  typedef _StorageIndex StorageIndex;
+  typedef MatrixXpr XprKind;
+
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = 1,
+    MaxRowsAtCompileTime = Dynamic,
+    MaxColsAtCompileTime = 1,
+    Flags = LvalueBit
+  };
+};
+
+template<typename _Scalar, int _Options, typename _StorageIndex, int DiagIndex>
+struct traits<Diagonal<const SparseMatrix<_Scalar, _Options, _StorageIndex>, DiagIndex> >
+ : public traits<Diagonal<SparseMatrix<_Scalar, _Options, _StorageIndex>, DiagIndex> >
+{
+  enum {
+    Flags = 0
+  };
+};
+
+} // end namespace internal
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+class SparseMatrix
+  : public SparseCompressedBase<SparseMatrix<_Scalar, _Options, _StorageIndex> >
+{
+    typedef SparseCompressedBase<SparseMatrix> Base;
+    using Base::convert_index;
+    friend class SparseVector<_Scalar,0,_StorageIndex>;
+  public:
+    using Base::isCompressed;
+    using Base::nonZeros;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(SparseMatrix)
+    using Base::operator+=;
+    using Base::operator-=;
+
+    typedef MappedSparseMatrix<Scalar,Flags> Map;
+    typedef Diagonal<SparseMatrix> DiagonalReturnType;
+    typedef Diagonal<const SparseMatrix> ConstDiagonalReturnType;
+    typedef typename Base::InnerIterator InnerIterator;
+    typedef typename Base::ReverseInnerIterator ReverseInnerIterator;
+    
+
+    using Base::IsRowMajor;
+    typedef internal::CompressedStorage<Scalar,StorageIndex> Storage;
+    enum {
+      Options = _Options
+    };
+
+    typedef typename Base::IndexVector IndexVector;
+    typedef typename Base::ScalarVector ScalarVector;
+  protected:
+    typedef SparseMatrix<Scalar,(Flags&~RowMajorBit)|(IsRowMajor?RowMajorBit:0)> TransposedSparseMatrix;
+
+    Index m_outerSize;
+    Index m_innerSize;
+    StorageIndex* m_outerIndex;
+    StorageIndex* m_innerNonZeros;     // optional, if null then the data is compressed
+    Storage m_data;
+
+  public:
+    
+    /** \returns the number of rows of the matrix */
+    inline Index rows() const { return IsRowMajor ? m_outerSize : m_innerSize; }
+    /** \returns the number of columns of the matrix */
+    inline Index cols() const { return IsRowMajor ? m_innerSize : m_outerSize; }
+
+    /** \returns the number of rows (resp. columns) of the matrix if the storage order column major (resp. row major) */
+    inline Index innerSize() const { return m_innerSize; }
+    /** \returns the number of columns (resp. rows) of the matrix if the storage order column major (resp. row major) */
+    inline Index outerSize() const { return m_outerSize; }
+    
+    /** \returns a const pointer to the array of values.
+      * This function is aimed at interoperability with other libraries.
+      * \sa innerIndexPtr(), outerIndexPtr() */
+    inline const Scalar* valuePtr() const { return m_data.valuePtr(); }
+    /** \returns a non-const pointer to the array of values.
+      * This function is aimed at interoperability with other libraries.
+      * \sa innerIndexPtr(), outerIndexPtr() */
+    inline Scalar* valuePtr() { return m_data.valuePtr(); }
+
+    /** \returns a const pointer to the array of inner indices.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), outerIndexPtr() */
+    inline const StorageIndex* innerIndexPtr() const { return m_data.indexPtr(); }
+    /** \returns a non-const pointer to the array of inner indices.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), outerIndexPtr() */
+    inline StorageIndex* innerIndexPtr() { return m_data.indexPtr(); }
+
+    /** \returns a const pointer to the array of the starting positions of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), innerIndexPtr() */
+    inline const StorageIndex* outerIndexPtr() const { return m_outerIndex; }
+    /** \returns a non-const pointer to the array of the starting positions of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \sa valuePtr(), innerIndexPtr() */
+    inline StorageIndex* outerIndexPtr() { return m_outerIndex; }
+
+    /** \returns a const pointer to the array of the number of non zeros of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 in compressed mode */
+    inline const StorageIndex* innerNonZeroPtr() const { return m_innerNonZeros; }
+    /** \returns a non-const pointer to the array of the number of non zeros of the inner vectors.
+      * This function is aimed at interoperability with other libraries.
+      * \warning it returns the null pointer 0 in compressed mode */
+    inline StorageIndex* innerNonZeroPtr() { return m_innerNonZeros; }
+
+    /** \internal */
+    inline Storage& data() { return m_data; }
+    /** \internal */
+    inline const Storage& data() const { return m_data; }
+
+    /** \returns the value of the matrix at position \a i, \a j
+      * This function returns Scalar(0) if the element is an explicit \em zero */
+    inline Scalar coeff(Index row, Index col) const
+    {
+      eigen_assert(row>=0 && row<rows() && col>=0 && col<cols());
+      
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+      Index end = m_innerNonZeros ? m_outerIndex[outer] + m_innerNonZeros[outer] : m_outerIndex[outer+1];
+      return m_data.atInRange(m_outerIndex[outer], end, StorageIndex(inner));
+    }
+
+    /** \returns a non-const reference to the value of the matrix at position \a i, \a j
+      *
+      * If the element does not exist then it is inserted via the insert(Index,Index) function
+      * which itself turns the matrix into a non compressed form if that was not the case.
+      *
+      * This is a O(log(nnz_j)) operation (binary search) plus the cost of insert(Index,Index)
+      * function if the element does not already exist.
+      */
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      eigen_assert(row>=0 && row<rows() && col>=0 && col<cols());
+      
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+
+      Index start = m_outerIndex[outer];
+      Index end = m_innerNonZeros ? m_outerIndex[outer] + m_innerNonZeros[outer] : m_outerIndex[outer+1];
+      eigen_assert(end>=start && "you probably called coeffRef on a non finalized matrix");
+      if(end<=start)
+        return insert(row,col);
+      const Index p = m_data.searchLowerIndex(start,end-1,StorageIndex(inner));
+      if((p<end) && (m_data.index(p)==inner))
+        return m_data.value(p);
+      else
+        return insert(row,col);
+    }
+
+    /** \returns a reference to a novel non zero coefficient with coordinates \a row x \a col.
+      * The non zero coefficient must \b not already exist.
+      *
+      * If the matrix \c *this is in compressed mode, then \c *this is turned into uncompressed
+      * mode while reserving room for 2 x this->innerSize() non zeros if reserve(Index) has not been called earlier.
+      * In this case, the insertion procedure is optimized for a \e sequential insertion mode where elements are assumed to be
+      * inserted by increasing outer-indices.
+      * 
+      * If that's not the case, then it is strongly recommended to either use a triplet-list to assemble the matrix, or to first
+      * call reserve(const SizesType &) to reserve the appropriate number of non-zero elements per inner vector.
+      *
+      * Assuming memory has been appropriately reserved, this function performs a sorted insertion in O(1)
+      * if the elements of each inner vector are inserted in increasing inner index order, and in O(nnz_j) for a random insertion.
+      *
+      */
+    Scalar& insert(Index row, Index col);
+
+  public:
+
+    /** Removes all non zeros but keep allocated memory
+      *
+      * This function does not free the currently allocated memory. To release as much as memory as possible,
+      * call \code mat.data().squeeze(); \endcode after resizing it.
+      * 
+      * \sa resize(Index,Index), data()
+      */
+    inline void setZero()
+    {
+      m_data.clear();
+      memset(m_outerIndex, 0, (m_outerSize+1)*sizeof(StorageIndex));
+      if(m_innerNonZeros)
+        memset(m_innerNonZeros, 0, (m_outerSize)*sizeof(StorageIndex));
+    }
+
+    /** Preallocates \a reserveSize non zeros.
+      *
+      * Precondition: the matrix must be in compressed mode. */
+    inline void reserve(Index reserveSize)
+    {
+      eigen_assert(isCompressed() && "This function does not make sense in non compressed mode.");
+      m_data.reserve(reserveSize);
+    }
+    
+    #ifdef EIGEN_PARSED_BY_DOXYGEN
+    /** Preallocates \a reserveSize[\c j] non zeros for each column (resp. row) \c j.
+      *
+      * This function turns the matrix in non-compressed mode.
+      * 
+      * The type \c SizesType must expose the following interface:
+        \code
+        typedef value_type;
+        const value_type& operator[](i) const;
+        \endcode
+      * for \c i in the [0,this->outerSize()[ range.
+      * Typical choices include std::vector<int>, Eigen::VectorXi, Eigen::VectorXi::Constant, etc.
+      */
+    template<class SizesType>
+    inline void reserve(const SizesType& reserveSizes);
+    #else
+    template<class SizesType>
+    inline void reserve(const SizesType& reserveSizes, const typename SizesType::value_type& enableif =
+    #if (!EIGEN_COMP_MSVC) || (EIGEN_COMP_MSVC>=1500) // MSVC 2005 fails to compile with this typename
+        typename
+    #endif
+        SizesType::value_type())
+    {
+      EIGEN_UNUSED_VARIABLE(enableif);
+      reserveInnerVectors(reserveSizes);
+    }
+    #endif // EIGEN_PARSED_BY_DOXYGEN
+  protected:
+    template<class SizesType>
+    inline void reserveInnerVectors(const SizesType& reserveSizes)
+    {
+      if(isCompressed())
+      {
+        Index totalReserveSize = 0;
+        // turn the matrix into non-compressed mode
+        m_innerNonZeros = static_cast<StorageIndex*>(std::malloc(m_outerSize * sizeof(StorageIndex)));
+        if (!m_innerNonZeros) internal::throw_std_bad_alloc();
+        
+        // temporarily use m_innerSizes to hold the new starting points.
+        StorageIndex* newOuterIndex = m_innerNonZeros;
+        
+        StorageIndex count = 0;
+        for(Index j=0; j<m_outerSize; ++j)
+        {
+          newOuterIndex[j] = count;
+          count += reserveSizes[j] + (m_outerIndex[j+1]-m_outerIndex[j]);
+          totalReserveSize += reserveSizes[j];
+        }
+        m_data.reserve(totalReserveSize);
+        StorageIndex previousOuterIndex = m_outerIndex[m_outerSize];
+        for(Index j=m_outerSize-1; j>=0; --j)
+        {
+          StorageIndex innerNNZ = previousOuterIndex - m_outerIndex[j];
+          for(Index i=innerNNZ-1; i>=0; --i)
+          {
+            m_data.index(newOuterIndex[j]+i) = m_data.index(m_outerIndex[j]+i);
+            m_data.value(newOuterIndex[j]+i) = m_data.value(m_outerIndex[j]+i);
+          }
+          previousOuterIndex = m_outerIndex[j];
+          m_outerIndex[j] = newOuterIndex[j];
+          m_innerNonZeros[j] = innerNNZ;
+        }
+        m_outerIndex[m_outerSize] = m_outerIndex[m_outerSize-1] + m_innerNonZeros[m_outerSize-1] + reserveSizes[m_outerSize-1];
+        
+        m_data.resize(m_outerIndex[m_outerSize]);
+      }
+      else
+      {
+        StorageIndex* newOuterIndex = static_cast<StorageIndex*>(std::malloc((m_outerSize+1)*sizeof(StorageIndex)));
+        if (!newOuterIndex) internal::throw_std_bad_alloc();
+        
+        StorageIndex count = 0;
+        for(Index j=0; j<m_outerSize; ++j)
+        {
+          newOuterIndex[j] = count;
+          StorageIndex alreadyReserved = (m_outerIndex[j+1]-m_outerIndex[j]) - m_innerNonZeros[j];
+          StorageIndex toReserve = std::max<StorageIndex>(reserveSizes[j], alreadyReserved);
+          count += toReserve + m_innerNonZeros[j];
+        }
+        newOuterIndex[m_outerSize] = count;
+        
+        m_data.resize(count);
+        for(Index j=m_outerSize-1; j>=0; --j)
+        {
+          Index offset = newOuterIndex[j] - m_outerIndex[j];
+          if(offset>0)
+          {
+            StorageIndex innerNNZ = m_innerNonZeros[j];
+            for(Index i=innerNNZ-1; i>=0; --i)
+            {
+              m_data.index(newOuterIndex[j]+i) = m_data.index(m_outerIndex[j]+i);
+              m_data.value(newOuterIndex[j]+i) = m_data.value(m_outerIndex[j]+i);
+            }
+          }
+        }
+        
+        std::swap(m_outerIndex, newOuterIndex);
+        std::free(newOuterIndex);
+      }
+      
+    }
+  public:
+
+    //--- low level purely coherent filling ---
+
+    /** \internal
+      * \returns a reference to the non zero coefficient at position \a row, \a col assuming that:
+      * - the nonzero does not already exist
+      * - the new coefficient is the last one according to the storage order
+      *
+      * Before filling a given inner vector you must call the statVec(Index) function.
+      *
+      * After an insertion session, you should call the finalize() function.
+      *
+      * \sa insert, insertBackByOuterInner, startVec */
+    inline Scalar& insertBack(Index row, Index col)
+    {
+      return insertBackByOuterInner(IsRowMajor?row:col, IsRowMajor?col:row);
+    }
+
+    /** \internal
+      * \sa insertBack, startVec */
+    inline Scalar& insertBackByOuterInner(Index outer, Index inner)
+    {
+      eigen_assert(Index(m_outerIndex[outer+1]) == m_data.size() && "Invalid ordered insertion (invalid outer index)");
+      eigen_assert( (m_outerIndex[outer+1]-m_outerIndex[outer]==0 || m_data.index(m_data.size()-1)<inner) && "Invalid ordered insertion (invalid inner index)");
+      Index p = m_outerIndex[outer+1];
+      ++m_outerIndex[outer+1];
+      m_data.append(Scalar(0), inner);
+      return m_data.value(p);
+    }
+
+    /** \internal
+      * \warning use it only if you know what you are doing */
+    inline Scalar& insertBackByOuterInnerUnordered(Index outer, Index inner)
+    {
+      Index p = m_outerIndex[outer+1];
+      ++m_outerIndex[outer+1];
+      m_data.append(Scalar(0), inner);
+      return m_data.value(p);
+    }
+
+    /** \internal
+      * \sa insertBack, insertBackByOuterInner */
+    inline void startVec(Index outer)
+    {
+      eigen_assert(m_outerIndex[outer]==Index(m_data.size()) && "You must call startVec for each inner vector sequentially");
+      eigen_assert(m_outerIndex[outer+1]==0 && "You must call startVec for each inner vector sequentially");
+      m_outerIndex[outer+1] = m_outerIndex[outer];
+    }
+
+    /** \internal
+      * Must be called after inserting a set of non zero entries using the low level compressed API.
+      */
+    inline void finalize()
+    {
+      if(isCompressed())
+      {
+        StorageIndex size = internal::convert_index<StorageIndex>(m_data.size());
+        Index i = m_outerSize;
+        // find the last filled column
+        while (i>=0 && m_outerIndex[i]==0)
+          --i;
+        ++i;
+        while (i<=m_outerSize)
+        {
+          m_outerIndex[i] = size;
+          ++i;
+        }
+      }
+    }
+
+    //---
+
+    template<typename InputIterators>
+    void setFromTriplets(const InputIterators& begin, const InputIterators& end);
+
+    template<typename InputIterators,typename DupFunctor>
+    void setFromTriplets(const InputIterators& begin, const InputIterators& end, DupFunctor dup_func);
+
+    void sumupDuplicates() { collapseDuplicates(internal::scalar_sum_op<Scalar,Scalar>()); }
+
+    template<typename DupFunctor>
+    void collapseDuplicates(DupFunctor dup_func = DupFunctor());
+
+    //---
+    
+    /** \internal
+      * same as insert(Index,Index) except that the indices are given relative to the storage order */
+    Scalar& insertByOuterInner(Index j, Index i)
+    {
+      return insert(IsRowMajor ? j : i, IsRowMajor ? i : j);
+    }
+
+    /** Turns the matrix into the \em compressed format.
+      */
+    void makeCompressed()
+    {
+      if(isCompressed())
+        return;
+      
+      eigen_internal_assert(m_outerIndex!=0 && m_outerSize>0);
+      
+      Index oldStart = m_outerIndex[1];
+      m_outerIndex[1] = m_innerNonZeros[0];
+      for(Index j=1; j<m_outerSize; ++j)
+      {
+        Index nextOldStart = m_outerIndex[j+1];
+        Index offset = oldStart - m_outerIndex[j];
+        if(offset>0)
+        {
+          for(Index k=0; k<m_innerNonZeros[j]; ++k)
+          {
+            m_data.index(m_outerIndex[j]+k) = m_data.index(oldStart+k);
+            m_data.value(m_outerIndex[j]+k) = m_data.value(oldStart+k);
+          }
+        }
+        m_outerIndex[j+1] = m_outerIndex[j] + m_innerNonZeros[j];
+        oldStart = nextOldStart;
+      }
+      std::free(m_innerNonZeros);
+      m_innerNonZeros = 0;
+      m_data.resize(m_outerIndex[m_outerSize]);
+      m_data.squeeze();
+    }
+
+    /** Turns the matrix into the uncompressed mode */
+    void uncompress()
+    {
+      if(m_innerNonZeros != 0)
+        return; 
+      m_innerNonZeros = static_cast<StorageIndex*>(std::malloc(m_outerSize * sizeof(StorageIndex)));
+      for (Index i = 0; i < m_outerSize; i++)
+      {
+        m_innerNonZeros[i] = m_outerIndex[i+1] - m_outerIndex[i]; 
+      }
+    }
+    
+    /** Suppresses all nonzeros which are \b much \b smaller \b than \a reference under the tolerence \a epsilon */
+    void prune(const Scalar& reference, const RealScalar& epsilon = NumTraits<RealScalar>::dummy_precision())
+    {
+      prune(default_prunning_func(reference,epsilon));
+    }
+    
+    /** Turns the matrix into compressed format, and suppresses all nonzeros which do not satisfy the predicate \a keep.
+      * The functor type \a KeepFunc must implement the following function:
+      * \code
+      * bool operator() (const Index& row, const Index& col, const Scalar& value) const;
+      * \endcode
+      * \sa prune(Scalar,RealScalar)
+      */
+    template<typename KeepFunc>
+    void prune(const KeepFunc& keep = KeepFunc())
+    {
+      // TODO optimize the uncompressed mode to avoid moving and allocating the data twice
+      makeCompressed();
+
+      StorageIndex k = 0;
+      for(Index j=0; j<m_outerSize; ++j)
+      {
+        Index previousStart = m_outerIndex[j];
+        m_outerIndex[j] = k;
+        Index end = m_outerIndex[j+1];
+        for(Index i=previousStart; i<end; ++i)
+        {
+          if(keep(IsRowMajor?j:m_data.index(i), IsRowMajor?m_data.index(i):j, m_data.value(i)))
+          {
+            m_data.value(k) = m_data.value(i);
+            m_data.index(k) = m_data.index(i);
+            ++k;
+          }
+        }
+      }
+      m_outerIndex[m_outerSize] = k;
+      m_data.resize(k,0);
+    }
+
+    /** Resizes the matrix to a \a rows x \a cols matrix leaving old values untouched.
+      *
+      * If the sizes of the matrix are decreased, then the matrix is turned to \b uncompressed-mode
+      * and the storage of the out of bounds coefficients is kept and reserved.
+      * Call makeCompressed() to pack the entries and squeeze extra memory.
+      *
+      * \sa reserve(), setZero(), makeCompressed()
+      */
+    void conservativeResize(Index rows, Index cols) 
+    {
+      // No change
+      if (this->rows() == rows && this->cols() == cols) return;
+      
+      // If one dimension is null, then there is nothing to be preserved
+      if(rows==0 || cols==0) return resize(rows,cols);
+
+      Index innerChange = IsRowMajor ? cols - this->cols() : rows - this->rows();
+      Index outerChange = IsRowMajor ? rows - this->rows() : cols - this->cols();
+      StorageIndex newInnerSize = convert_index(IsRowMajor ? cols : rows);
+
+      // Deals with inner non zeros
+      if (m_innerNonZeros)
+      {
+        // Resize m_innerNonZeros
+        StorageIndex *newInnerNonZeros = static_cast<StorageIndex*>(std::realloc(m_innerNonZeros, (m_outerSize + outerChange) * sizeof(StorageIndex)));
+        if (!newInnerNonZeros) internal::throw_std_bad_alloc();
+        m_innerNonZeros = newInnerNonZeros;
+        
+        for(Index i=m_outerSize; i<m_outerSize+outerChange; i++)          
+          m_innerNonZeros[i] = 0;
+      } 
+      else if (innerChange < 0) 
+      {
+        // Inner size decreased: allocate a new m_innerNonZeros
+        m_innerNonZeros = static_cast<StorageIndex*>(std::malloc((m_outerSize+outerChange+1) * sizeof(StorageIndex)));
+        if (!m_innerNonZeros) internal::throw_std_bad_alloc();
+        for(Index i = 0; i < m_outerSize; i++)
+          m_innerNonZeros[i] = m_outerIndex[i+1] - m_outerIndex[i];
+      }
+      
+      // Change the m_innerNonZeros in case of a decrease of inner size
+      if (m_innerNonZeros && innerChange < 0)
+      {
+        for(Index i = 0; i < m_outerSize + (std::min)(outerChange, Index(0)); i++)
+        {
+          StorageIndex &n = m_innerNonZeros[i];
+          StorageIndex start = m_outerIndex[i];
+          while (n > 0 && m_data.index(start+n-1) >= newInnerSize) --n; 
+        }
+      }
+      
+      m_innerSize = newInnerSize;
+
+      // Re-allocate outer index structure if necessary
+      if (outerChange == 0)
+        return;
+          
+      StorageIndex *newOuterIndex = static_cast<StorageIndex*>(std::realloc(m_outerIndex, (m_outerSize + outerChange + 1) * sizeof(StorageIndex)));
+      if (!newOuterIndex) internal::throw_std_bad_alloc();
+      m_outerIndex = newOuterIndex;
+      if (outerChange > 0)
+      {
+        StorageIndex last = m_outerSize == 0 ? 0 : m_outerIndex[m_outerSize];
+        for(Index i=m_outerSize; i<m_outerSize+outerChange+1; i++)          
+          m_outerIndex[i] = last; 
+      }
+      m_outerSize += outerChange;
+    }
+    
+    /** Resizes the matrix to a \a rows x \a cols matrix and initializes it to zero.
+      * 
+      * This function does not free the currently allocated memory. To release as much as memory as possible,
+      * call \code mat.data().squeeze(); \endcode after resizing it.
+      * 
+      * \sa reserve(), setZero()
+      */
+    void resize(Index rows, Index cols)
+    {
+      const Index outerSize = IsRowMajor ? rows : cols;
+      m_innerSize = IsRowMajor ? cols : rows;
+      m_data.clear();
+      if (m_outerSize != outerSize || m_outerSize==0)
+      {
+        std::free(m_outerIndex);
+        m_outerIndex = static_cast<StorageIndex*>(std::malloc((outerSize + 1) * sizeof(StorageIndex)));
+        if (!m_outerIndex) internal::throw_std_bad_alloc();
+        
+        m_outerSize = outerSize;
+      }
+      if(m_innerNonZeros)
+      {
+        std::free(m_innerNonZeros);
+        m_innerNonZeros = 0;
+      }
+      memset(m_outerIndex, 0, (m_outerSize+1)*sizeof(StorageIndex));
+    }
+
+    /** \internal
+      * Resize the nonzero vector to \a size */
+    void resizeNonZeros(Index size)
+    {
+      m_data.resize(size);
+    }
+
+    /** \returns a const expression of the diagonal coefficients. */
+    const ConstDiagonalReturnType diagonal() const { return ConstDiagonalReturnType(*this); }
+    
+    /** \returns a read-write expression of the diagonal coefficients.
+      * \warning If the diagonal entries are written, then all diagonal
+      * entries \b must already exist, otherwise an assertion will be raised.
+      */
+    DiagonalReturnType diagonal() { return DiagonalReturnType(*this); }
+
+    /** Default constructor yielding an empty \c 0 \c x \c 0 matrix */
+    inline SparseMatrix()
+      : m_outerSize(-1), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      resize(0, 0);
+    }
+
+    /** Constructs a \a rows \c x \a cols empty matrix */
+    inline SparseMatrix(Index rows, Index cols)
+      : m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      resize(rows, cols);
+    }
+
+    /** Constructs a sparse matrix from the sparse expression \a other */
+    template<typename OtherDerived>
+    inline SparseMatrix(const SparseMatrixBase<OtherDerived>& other)
+      : m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+      check_template_parameters();
+      const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator<OtherDerived>::Flags & RowMajorBit);
+      if (needToTranspose)
+        *this = other.derived();
+      else
+      {
+        #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+          EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        #endif
+        internal::call_assignment_no_alias(*this, other.derived());
+      }
+    }
+    
+    /** Constructs a sparse matrix from the sparse selfadjoint view \a other */
+    template<typename OtherDerived, unsigned int UpLo>
+    inline SparseMatrix(const SparseSelfAdjointView<OtherDerived, UpLo>& other)
+      : m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      Base::operator=(other);
+    }
+
+    /** Copy constructor (it performs a deep copy) */
+    inline SparseMatrix(const SparseMatrix& other)
+      : Base(), m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      *this = other.derived();
+    }
+
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    SparseMatrix(const ReturnByValue<OtherDerived>& other)
+      : Base(), m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      initAssignment(other);
+      other.evalTo(*this);
+    }
+    
+    /** \brief Copy constructor with in-place evaluation */
+    template<typename OtherDerived>
+    explicit SparseMatrix(const DiagonalBase<OtherDerived>& other)
+      : Base(), m_outerSize(0), m_innerSize(0), m_outerIndex(0), m_innerNonZeros(0)
+    {
+      check_template_parameters();
+      *this = other.derived();
+    }
+
+    /** Swaps the content of two sparse matrices of the same type.
+      * This is a fast operation that simply swaps the underlying pointers and parameters. */
+    inline void swap(SparseMatrix& other)
+    {
+      //EIGEN_DBG_SPARSE(std::cout << "SparseMatrix:: swap\n");
+      std::swap(m_outerIndex, other.m_outerIndex);
+      std::swap(m_innerSize, other.m_innerSize);
+      std::swap(m_outerSize, other.m_outerSize);
+      std::swap(m_innerNonZeros, other.m_innerNonZeros);
+      m_data.swap(other.m_data);
+    }
+
+    /** Sets *this to the identity matrix.
+      * This function also turns the matrix into compressed mode, and drop any reserved memory. */
+    inline void setIdentity()
+    {
+      eigen_assert(rows() == cols() && "ONLY FOR SQUARED MATRICES");
+      this->m_data.resize(rows());
+      Eigen::Map<IndexVector>(this->m_data.indexPtr(), rows()).setLinSpaced(0, StorageIndex(rows()-1));
+      Eigen::Map<ScalarVector>(this->m_data.valuePtr(), rows()).setOnes();
+      Eigen::Map<IndexVector>(this->m_outerIndex, rows()+1).setLinSpaced(0, StorageIndex(rows()));
+      std::free(m_innerNonZeros);
+      m_innerNonZeros = 0;
+    }
+    inline SparseMatrix& operator=(const SparseMatrix& other)
+    {
+      if (other.isRValue())
+      {
+        swap(other.const_cast_derived());
+      }
+      else if(this!=&other)
+      {
+        #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+          EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        #endif
+        initAssignment(other);
+        if(other.isCompressed())
+        {
+          internal::smart_copy(other.m_outerIndex, other.m_outerIndex + m_outerSize + 1, m_outerIndex);
+          m_data = other.m_data;
+        }
+        else
+        {
+          Base::operator=(other);
+        }
+      }
+      return *this;
+    }
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename OtherDerived>
+    inline SparseMatrix& operator=(const EigenBase<OtherDerived>& other)
+    { return Base::operator=(other.derived()); }
+#endif // EIGEN_PARSED_BY_DOXYGEN
+
+    template<typename OtherDerived>
+    EIGEN_DONT_INLINE SparseMatrix& operator=(const SparseMatrixBase<OtherDerived>& other);
+
+    friend std::ostream & operator << (std::ostream & s, const SparseMatrix& m)
+    {
+      EIGEN_DBG_SPARSE(
+        s << "Nonzero entries:\n";
+        if(m.isCompressed())
+        {
+          for (Index i=0; i<m.nonZeros(); ++i)
+            s << "(" << m.m_data.value(i) << "," << m.m_data.index(i) << ") ";
+        }
+        else
+        {
+          for (Index i=0; i<m.outerSize(); ++i)
+          {
+            Index p = m.m_outerIndex[i];
+            Index pe = m.m_outerIndex[i]+m.m_innerNonZeros[i];
+            Index k=p;
+            for (; k<pe; ++k) {
+              s << "(" << m.m_data.value(k) << "," << m.m_data.index(k) << ") ";
+            }
+            for (; k<m.m_outerIndex[i+1]; ++k) {
+              s << "(_,_) ";
+            }
+          }
+        }
+        s << std::endl;
+        s << std::endl;
+        s << "Outer pointers:\n";
+        for (Index i=0; i<m.outerSize(); ++i) {
+          s << m.m_outerIndex[i] << " ";
+        }
+        s << " $" << std::endl;
+        if(!m.isCompressed())
+        {
+          s << "Inner non zeros:\n";
+          for (Index i=0; i<m.outerSize(); ++i) {
+            s << m.m_innerNonZeros[i] << " ";
+          }
+          s << " $" << std::endl;
+        }
+        s << std::endl;
+      );
+      s << static_cast<const SparseMatrixBase<SparseMatrix>&>(m);
+      return s;
+    }
+
+    /** Destructor */
+    inline ~SparseMatrix()
+    {
+      std::free(m_outerIndex);
+      std::free(m_innerNonZeros);
+    }
+
+    /** Overloaded for performance */
+    Scalar sum() const;
+    
+#   ifdef EIGEN_SPARSEMATRIX_PLUGIN
+#     include EIGEN_SPARSEMATRIX_PLUGIN
+#   endif
+
+protected:
+
+    template<typename Other>
+    void initAssignment(const Other& other)
+    {
+      resize(other.rows(), other.cols());
+      if(m_innerNonZeros)
+      {
+        std::free(m_innerNonZeros);
+        m_innerNonZeros = 0;
+      }
+    }
+
+    /** \internal
+      * \sa insert(Index,Index) */
+    EIGEN_DONT_INLINE Scalar& insertCompressed(Index row, Index col);
+
+    /** \internal
+      * A vector object that is equal to 0 everywhere but v at the position i */
+    class SingletonVector
+    {
+        StorageIndex m_index;
+        StorageIndex m_value;
+      public:
+        typedef StorageIndex value_type;
+        SingletonVector(Index i, Index v)
+          : m_index(convert_index(i)), m_value(convert_index(v))
+        {}
+
+        StorageIndex operator[](Index i) const { return i==m_index ? m_value : 0; }
+    };
+
+    /** \internal
+      * \sa insert(Index,Index) */
+    EIGEN_DONT_INLINE Scalar& insertUncompressed(Index row, Index col);
+
+public:
+    /** \internal
+      * \sa insert(Index,Index) */
+    EIGEN_STRONG_INLINE Scalar& insertBackUncompressed(Index row, Index col)
+    {
+      const Index outer = IsRowMajor ? row : col;
+      const Index inner = IsRowMajor ? col : row;
+
+      eigen_assert(!isCompressed());
+      eigen_assert(m_innerNonZeros[outer]<=(m_outerIndex[outer+1] - m_outerIndex[outer]));
+
+      Index p = m_outerIndex[outer] + m_innerNonZeros[outer]++;
+      m_data.index(p) = convert_index(inner);
+      return (m_data.value(p) = 0);
+    }
+
+private:
+  static void check_template_parameters()
+  {
+    EIGEN_STATIC_ASSERT(NumTraits<StorageIndex>::IsSigned,THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE);
+    EIGEN_STATIC_ASSERT((Options&(ColMajor|RowMajor))==Options,INVALID_MATRIX_TEMPLATE_PARAMETERS);
+  }
+
+  struct default_prunning_func {
+    default_prunning_func(const Scalar& ref, const RealScalar& eps) : reference(ref), epsilon(eps) {}
+    inline bool operator() (const Index&, const Index&, const Scalar& value) const
+    {
+      return !internal::isMuchSmallerThan(value, reference, epsilon);
+    }
+    Scalar reference;
+    RealScalar epsilon;
+  };
+};
+
+namespace internal {
+
+template<typename InputIterator, typename SparseMatrixType, typename DupFunctor>
+void set_from_triplets(const InputIterator& begin, const InputIterator& end, SparseMatrixType& mat, DupFunctor dup_func)
+{
+  enum { IsRowMajor = SparseMatrixType::IsRowMajor };
+  typedef typename SparseMatrixType::Scalar Scalar;
+  typedef typename SparseMatrixType::StorageIndex StorageIndex;
+  SparseMatrix<Scalar,IsRowMajor?ColMajor:RowMajor,StorageIndex> trMat(mat.rows(),mat.cols());
+
+  if(begin!=end)
+  {
+    // pass 1: count the nnz per inner-vector
+    typename SparseMatrixType::IndexVector wi(trMat.outerSize());
+    wi.setZero();
+    for(InputIterator it(begin); it!=end; ++it)
+    {
+      eigen_assert(it->row()>=0 && it->row()<mat.rows() && it->col()>=0 && it->col()<mat.cols());
+      wi(IsRowMajor ? it->col() : it->row())++;
+    }
+
+    // pass 2: insert all the elements into trMat
+    trMat.reserve(wi);
+    for(InputIterator it(begin); it!=end; ++it)
+      trMat.insertBackUncompressed(it->row(),it->col()) = it->value();
+
+    // pass 3:
+    trMat.collapseDuplicates(dup_func);
+  }
+
+  // pass 4: transposed copy -> implicit sorting
+  mat = trMat;
+}
+
+}
+
+
+/** Fill the matrix \c *this with the list of \em triplets defined by the iterator range \a begin - \a end.
+  *
+  * A \em triplet is a tuple (i,j,value) defining a non-zero element.
+  * The input list of triplets does not have to be sorted, and can contains duplicated elements.
+  * In any case, the result is a \b sorted and \b compressed sparse matrix where the duplicates have been summed up.
+  * This is a \em O(n) operation, with \em n the number of triplet elements.
+  * The initial contents of \c *this is destroyed.
+  * The matrix \c *this must be properly resized beforehand using the SparseMatrix(Index,Index) constructor,
+  * or the resize(Index,Index) method. The sizes are not extracted from the triplet list.
+  *
+  * The \a InputIterators value_type must provide the following interface:
+  * \code
+  * Scalar value() const; // the value
+  * Scalar row() const;   // the row index i
+  * Scalar col() const;   // the column index j
+  * \endcode
+  * See for instance the Eigen::Triplet template class.
+  *
+  * Here is a typical usage example:
+  * \code
+    typedef Triplet<double> T;
+    std::vector<T> tripletList;
+    triplets.reserve(estimation_of_entries);
+    for(...)
+    {
+      // ...
+      tripletList.push_back(T(i,j,v_ij));
+    }
+    SparseMatrixType m(rows,cols);
+    m.setFromTriplets(tripletList.begin(), tripletList.end());
+    // m is ready to go!
+  * \endcode
+  *
+  * \warning The list of triplets is read multiple times (at least twice). Therefore, it is not recommended to define
+  * an abstract iterator over a complex data-structure that would be expensive to evaluate. The triplets should rather
+  * be explicitely stored into a std::vector for instance.
+  */
+template<typename Scalar, int _Options, typename _StorageIndex>
+template<typename InputIterators>
+void SparseMatrix<Scalar,_Options,_StorageIndex>::setFromTriplets(const InputIterators& begin, const InputIterators& end)
+{
+  internal::set_from_triplets<InputIterators, SparseMatrix<Scalar,_Options,_StorageIndex> >(begin, end, *this, internal::scalar_sum_op<Scalar,Scalar>());
+}
+
+/** The same as setFromTriplets but when duplicates are met the functor \a dup_func is applied:
+  * \code
+  * value = dup_func(OldValue, NewValue)
+  * \endcode 
+  * Here is a C++11 example keeping the latest entry only:
+  * \code
+  * mat.setFromTriplets(triplets.begin(), triplets.end(), [] (const Scalar&,const Scalar &b) { return b; });
+  * \endcode
+  */
+template<typename Scalar, int _Options, typename _StorageIndex>
+template<typename InputIterators,typename DupFunctor>
+void SparseMatrix<Scalar,_Options,_StorageIndex>::setFromTriplets(const InputIterators& begin, const InputIterators& end, DupFunctor dup_func)
+{
+  internal::set_from_triplets<InputIterators, SparseMatrix<Scalar,_Options,_StorageIndex>, DupFunctor>(begin, end, *this, dup_func);
+}
+
+/** \internal */
+template<typename Scalar, int _Options, typename _StorageIndex>
+template<typename DupFunctor>
+void SparseMatrix<Scalar,_Options,_StorageIndex>::collapseDuplicates(DupFunctor dup_func)
+{
+  eigen_assert(!isCompressed());
+  // TODO, in practice we should be able to use m_innerNonZeros for that task
+  IndexVector wi(innerSize());
+  wi.fill(-1);
+  StorageIndex count = 0;
+  // for each inner-vector, wi[inner_index] will hold the position of first element into the index/value buffers
+  for(Index j=0; j<outerSize(); ++j)
+  {
+    StorageIndex start   = count;
+    Index oldEnd  = m_outerIndex[j]+m_innerNonZeros[j];
+    for(Index k=m_outerIndex[j]; k<oldEnd; ++k)
+    {
+      Index i = m_data.index(k);
+      if(wi(i)>=start)
+      {
+        // we already meet this entry => accumulate it
+        m_data.value(wi(i)) = dup_func(m_data.value(wi(i)), m_data.value(k));
+      }
+      else
+      {
+        m_data.value(count) = m_data.value(k);
+        m_data.index(count) = m_data.index(k);
+        wi(i) = count;
+        ++count;
+      }
+    }
+    m_outerIndex[j] = start;
+  }
+  m_outerIndex[m_outerSize] = count;
+
+  // turn the matrix into compressed form
+  std::free(m_innerNonZeros);
+  m_innerNonZeros = 0;
+  m_data.resize(m_outerIndex[m_outerSize]);
+}
+
+template<typename Scalar, int _Options, typename _StorageIndex>
+template<typename OtherDerived>
+EIGEN_DONT_INLINE SparseMatrix<Scalar,_Options,_StorageIndex>& SparseMatrix<Scalar,_Options,_StorageIndex>::operator=(const SparseMatrixBase<OtherDerived>& other)
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<Scalar, typename OtherDerived::Scalar>::value),
+        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
+
+  #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+    EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+  #endif
+      
+  const bool needToTranspose = (Flags & RowMajorBit) != (internal::evaluator<OtherDerived>::Flags & RowMajorBit);
+  if (needToTranspose)
+  {
+    #ifdef EIGEN_SPARSE_TRANSPOSED_COPY_PLUGIN
+      EIGEN_SPARSE_TRANSPOSED_COPY_PLUGIN
+    #endif
+    // two passes algorithm:
+    //  1 - compute the number of coeffs per dest inner vector
+    //  2 - do the actual copy/eval
+    // Since each coeff of the rhs has to be evaluated twice, let's evaluate it if needed
+    typedef typename internal::nested_eval<OtherDerived,2,typename internal::plain_matrix_type<OtherDerived>::type >::type OtherCopy;
+    typedef typename internal::remove_all<OtherCopy>::type _OtherCopy;
+    typedef internal::evaluator<_OtherCopy> OtherCopyEval;
+    OtherCopy otherCopy(other.derived());
+    OtherCopyEval otherCopyEval(otherCopy);
+
+    SparseMatrix dest(other.rows(),other.cols());
+    Eigen::Map<IndexVector> (dest.m_outerIndex,dest.outerSize()).setZero();
+
+    // pass 1
+    // FIXME the above copy could be merged with that pass
+    for (Index j=0; j<otherCopy.outerSize(); ++j)
+      for (typename OtherCopyEval::InnerIterator it(otherCopyEval, j); it; ++it)
+        ++dest.m_outerIndex[it.index()];
+
+    // prefix sum
+    StorageIndex count = 0;
+    IndexVector positions(dest.outerSize());
+    for (Index j=0; j<dest.outerSize(); ++j)
+    {
+      StorageIndex tmp = dest.m_outerIndex[j];
+      dest.m_outerIndex[j] = count;
+      positions[j] = count;
+      count += tmp;
+    }
+    dest.m_outerIndex[dest.outerSize()] = count;
+    // alloc
+    dest.m_data.resize(count);
+    // pass 2
+    for (StorageIndex j=0; j<otherCopy.outerSize(); ++j)
+    {
+      for (typename OtherCopyEval::InnerIterator it(otherCopyEval, j); it; ++it)
+      {
+        Index pos = positions[it.index()]++;
+        dest.m_data.index(pos) = j;
+        dest.m_data.value(pos) = it.value();
+      }
+    }
+    this->swap(dest);
+    return *this;
+  }
+  else
+  {
+    if(other.isRValue())
+    {
+      initAssignment(other.derived());
+    }
+    // there is no special optimization
+    return Base::operator=(other.derived());
+  }
+}
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+typename SparseMatrix<_Scalar,_Options,_StorageIndex>::Scalar& SparseMatrix<_Scalar,_Options,_StorageIndex>::insert(Index row, Index col)
+{
+  eigen_assert(row>=0 && row<rows() && col>=0 && col<cols());
+  
+  const Index outer = IsRowMajor ? row : col;
+  const Index inner = IsRowMajor ? col : row;
+  
+  if(isCompressed())
+  {
+    if(nonZeros()==0)
+    {
+      // reserve space if not already done
+      if(m_data.allocatedSize()==0)
+        m_data.reserve(2*m_innerSize);
+      
+      // turn the matrix into non-compressed mode
+      m_innerNonZeros = static_cast<StorageIndex*>(std::malloc(m_outerSize * sizeof(StorageIndex)));
+      if(!m_innerNonZeros) internal::throw_std_bad_alloc();
+      
+      memset(m_innerNonZeros, 0, (m_outerSize)*sizeof(StorageIndex));
+      
+      // pack all inner-vectors to the end of the pre-allocated space
+      // and allocate the entire free-space to the first inner-vector
+      StorageIndex end = convert_index(m_data.allocatedSize());
+      for(Index j=1; j<=m_outerSize; ++j)
+        m_outerIndex[j] = end;
+    }
+    else
+    {
+      // turn the matrix into non-compressed mode
+      m_innerNonZeros = static_cast<StorageIndex*>(std::malloc(m_outerSize * sizeof(StorageIndex)));
+      if(!m_innerNonZeros) internal::throw_std_bad_alloc();
+      for(Index j=0; j<m_outerSize; ++j)
+        m_innerNonZeros[j] = m_outerIndex[j+1]-m_outerIndex[j];
+    }
+  }
+  
+  // check whether we can do a fast "push back" insertion
+  Index data_end = m_data.allocatedSize();
+  
+  // First case: we are filling a new inner vector which is packed at the end.
+  // We assume that all remaining inner-vectors are also empty and packed to the end.
+  if(m_outerIndex[outer]==data_end)
+  {
+    eigen_internal_assert(m_innerNonZeros[outer]==0);
+    
+    // pack previous empty inner-vectors to end of the used-space
+    // and allocate the entire free-space to the current inner-vector.
+    StorageIndex p = convert_index(m_data.size());
+    Index j = outer;
+    while(j>=0 && m_innerNonZeros[j]==0)
+      m_outerIndex[j--] = p;
+    
+    // push back the new element
+    ++m_innerNonZeros[outer];
+    m_data.append(Scalar(0), inner);
+    
+    // check for reallocation
+    if(data_end != m_data.allocatedSize())
+    {
+      // m_data has been reallocated
+      //  -> move remaining inner-vectors back to the end of the free-space
+      //     so that the entire free-space is allocated to the current inner-vector.
+      eigen_internal_assert(data_end < m_data.allocatedSize());
+      StorageIndex new_end = convert_index(m_data.allocatedSize());
+      for(Index k=outer+1; k<=m_outerSize; ++k)
+        if(m_outerIndex[k]==data_end)
+          m_outerIndex[k] = new_end;
+    }
+    return m_data.value(p);
+  }
+  
+  // Second case: the next inner-vector is packed to the end
+  // and the current inner-vector end match the used-space.
+  if(m_outerIndex[outer+1]==data_end && m_outerIndex[outer]+m_innerNonZeros[outer]==m_data.size())
+  {
+    eigen_internal_assert(outer+1==m_outerSize || m_innerNonZeros[outer+1]==0);
+    
+    // add space for the new element
+    ++m_innerNonZeros[outer];
+    m_data.resize(m_data.size()+1);
+    
+    // check for reallocation
+    if(data_end != m_data.allocatedSize())
+    {
+      // m_data has been reallocated
+      //  -> move remaining inner-vectors back to the end of the free-space
+      //     so that the entire free-space is allocated to the current inner-vector.
+      eigen_internal_assert(data_end < m_data.allocatedSize());
+      StorageIndex new_end = convert_index(m_data.allocatedSize());
+      for(Index k=outer+1; k<=m_outerSize; ++k)
+        if(m_outerIndex[k]==data_end)
+          m_outerIndex[k] = new_end;
+    }
+    
+    // and insert it at the right position (sorted insertion)
+    Index startId = m_outerIndex[outer];
+    Index p = m_outerIndex[outer]+m_innerNonZeros[outer]-1;
+    while ( (p > startId) && (m_data.index(p-1) > inner) )
+    {
+      m_data.index(p) = m_data.index(p-1);
+      m_data.value(p) = m_data.value(p-1);
+      --p;
+    }
+    
+    m_data.index(p) = convert_index(inner);
+    return (m_data.value(p) = 0);
+  }
+  
+  if(m_data.size() != m_data.allocatedSize())
+  {
+    // make sure the matrix is compatible to random un-compressed insertion:
+    m_data.resize(m_data.allocatedSize());
+    this->reserveInnerVectors(Array<StorageIndex,Dynamic,1>::Constant(m_outerSize, 2));
+  }
+  
+  return insertUncompressed(row,col);
+}
+    
+template<typename _Scalar, int _Options, typename _StorageIndex>
+EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_StorageIndex>::Scalar& SparseMatrix<_Scalar,_Options,_StorageIndex>::insertUncompressed(Index row, Index col)
+{
+  eigen_assert(!isCompressed());
+
+  const Index outer = IsRowMajor ? row : col;
+  const StorageIndex inner = convert_index(IsRowMajor ? col : row);
+
+  Index room = m_outerIndex[outer+1] - m_outerIndex[outer];
+  StorageIndex innerNNZ = m_innerNonZeros[outer];
+  if(innerNNZ>=room)
+  {
+    // this inner vector is full, we need to reallocate the whole buffer :(
+    reserve(SingletonVector(outer,std::max<StorageIndex>(2,innerNNZ)));
+  }
+
+  Index startId = m_outerIndex[outer];
+  Index p = startId + m_innerNonZeros[outer];
+  while ( (p > startId) && (m_data.index(p-1) > inner) )
+  {
+    m_data.index(p) = m_data.index(p-1);
+    m_data.value(p) = m_data.value(p-1);
+    --p;
+  }
+  eigen_assert((p<=startId || m_data.index(p-1)!=inner) && "you cannot insert an element that already exists, you must call coeffRef to this end");
+
+  m_innerNonZeros[outer]++;
+
+  m_data.index(p) = inner;
+  return (m_data.value(p) = 0);
+}
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+EIGEN_DONT_INLINE typename SparseMatrix<_Scalar,_Options,_StorageIndex>::Scalar& SparseMatrix<_Scalar,_Options,_StorageIndex>::insertCompressed(Index row, Index col)
+{
+  eigen_assert(isCompressed());
+
+  const Index outer = IsRowMajor ? row : col;
+  const Index inner = IsRowMajor ? col : row;
+
+  Index previousOuter = outer;
+  if (m_outerIndex[outer+1]==0)
+  {
+    // we start a new inner vector
+    while (previousOuter>=0 && m_outerIndex[previousOuter]==0)
+    {
+      m_outerIndex[previousOuter] = convert_index(m_data.size());
+      --previousOuter;
+    }
+    m_outerIndex[outer+1] = m_outerIndex[outer];
+  }
+
+  // here we have to handle the tricky case where the outerIndex array
+  // starts with: [ 0 0 0 0 0 1 ...] and we are inserted in, e.g.,
+  // the 2nd inner vector...
+  bool isLastVec = (!(previousOuter==-1 && m_data.size()!=0))
+                && (std::size_t(m_outerIndex[outer+1]) == m_data.size());
+
+  std::size_t startId = m_outerIndex[outer];
+  // FIXME let's make sure sizeof(long int) == sizeof(std::size_t)
+  std::size_t p = m_outerIndex[outer+1];
+  ++m_outerIndex[outer+1];
+
+  double reallocRatio = 1;
+  if (m_data.allocatedSize()<=m_data.size())
+  {
+    // if there is no preallocated memory, let's reserve a minimum of 32 elements
+    if (m_data.size()==0)
+    {
+      m_data.reserve(32);
+    }
+    else
+    {
+      // we need to reallocate the data, to reduce multiple reallocations
+      // we use a smart resize algorithm based on the current filling ratio
+      // in addition, we use double to avoid integers overflows
+      double nnzEstimate = double(m_outerIndex[outer])*double(m_outerSize)/double(outer+1);
+      reallocRatio = (nnzEstimate-double(m_data.size()))/double(m_data.size());
+      // furthermore we bound the realloc ratio to:
+      //   1) reduce multiple minor realloc when the matrix is almost filled
+      //   2) avoid to allocate too much memory when the matrix is almost empty
+      reallocRatio = (std::min)((std::max)(reallocRatio,1.5),8.);
+    }
+  }
+  m_data.resize(m_data.size()+1,reallocRatio);
+
+  if (!isLastVec)
+  {
+    if (previousOuter==-1)
+    {
+      // oops wrong guess.
+      // let's correct the outer offsets
+      for (Index k=0; k<=(outer+1); ++k)
+        m_outerIndex[k] = 0;
+      Index k=outer+1;
+      while(m_outerIndex[k]==0)
+        m_outerIndex[k++] = 1;
+      while (k<=m_outerSize && m_outerIndex[k]!=0)
+        m_outerIndex[k++]++;
+      p = 0;
+      --k;
+      k = m_outerIndex[k]-1;
+      while (k>0)
+      {
+        m_data.index(k) = m_data.index(k-1);
+        m_data.value(k) = m_data.value(k-1);
+        k--;
+      }
+    }
+    else
+    {
+      // we are not inserting into the last inner vec
+      // update outer indices:
+      Index j = outer+2;
+      while (j<=m_outerSize && m_outerIndex[j]!=0)
+        m_outerIndex[j++]++;
+      --j;
+      // shift data of last vecs:
+      Index k = m_outerIndex[j]-1;
+      while (k>=Index(p))
+      {
+        m_data.index(k) = m_data.index(k-1);
+        m_data.value(k) = m_data.value(k-1);
+        k--;
+      }
+    }
+  }
+
+  while ( (p > startId) && (m_data.index(p-1) > inner) )
+  {
+    m_data.index(p) = m_data.index(p-1);
+    m_data.value(p) = m_data.value(p-1);
+    --p;
+  }
+
+  m_data.index(p) = inner;
+  return (m_data.value(p) = 0);
+}
+
+namespace internal {
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct evaluator<SparseMatrix<_Scalar,_Options,_StorageIndex> >
+  : evaluator<SparseCompressedBase<SparseMatrix<_Scalar,_Options,_StorageIndex> > >
+{
+  typedef evaluator<SparseCompressedBase<SparseMatrix<_Scalar,_Options,_StorageIndex> > > Base;
+  typedef SparseMatrix<_Scalar,_Options,_StorageIndex> SparseMatrixType;
+  evaluator() : Base() {}
+  explicit evaluator(const SparseMatrixType &mat) : Base(mat) {}
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEMATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrixBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrixBase.h
new file mode 100644
index 0000000..c6b548f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseMatrixBase.h
@@ -0,0 +1,405 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEMATRIXBASE_H
+#define EIGEN_SPARSEMATRIXBASE_H
+
+namespace Eigen { 
+
+/** \ingroup SparseCore_Module
+  *
+  * \class SparseMatrixBase
+  *
+  * \brief Base class of any sparse matrices or sparse expressions
+  *
+  * \tparam Derived is the derived type, e.g. a sparse matrix type, or an expression, etc.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEMATRIXBASE_PLUGIN.
+  */
+template<typename Derived> class SparseMatrixBase
+  : public EigenBase<Derived>
+{
+  public:
+
+    typedef typename internal::traits<Derived>::Scalar Scalar;
+    
+    /** The numeric type of the expression' coefficients, e.g. float, double, int or std::complex<float>, etc.
+      *
+      * It is an alias for the Scalar type */
+    typedef Scalar value_type;
+    
+    typedef typename internal::packet_traits<Scalar>::type PacketScalar;
+    typedef typename internal::traits<Derived>::StorageKind StorageKind;
+
+    /** The integer type used to \b store indices within a SparseMatrix.
+      * For a \c SparseMatrix<Scalar,Options,IndexType> it an alias of the third template parameter \c IndexType. */
+    typedef typename internal::traits<Derived>::StorageIndex StorageIndex;
+
+    typedef typename internal::add_const_on_value_type_if_arithmetic<
+                         typename internal::packet_traits<Scalar>::type
+                     >::type PacketReturnType;
+
+    typedef SparseMatrixBase StorageBaseType;
+
+    typedef Matrix<StorageIndex,Dynamic,1> IndexVector;
+    typedef Matrix<Scalar,Dynamic,1> ScalarVector;
+    
+    template<typename OtherDerived>
+    Derived& operator=(const EigenBase<OtherDerived> &other);
+
+    enum {
+
+      RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
+        /**< The number of rows at compile-time. This is just a copy of the value provided
+          * by the \a Derived type. If a value is not known at compile-time,
+          * it is set to the \a Dynamic constant.
+          * \sa MatrixBase::rows(), MatrixBase::cols(), ColsAtCompileTime, SizeAtCompileTime */
+
+      ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+        /**< The number of columns at compile-time. This is just a copy of the value provided
+          * by the \a Derived type. If a value is not known at compile-time,
+          * it is set to the \a Dynamic constant.
+          * \sa MatrixBase::rows(), MatrixBase::cols(), RowsAtCompileTime, SizeAtCompileTime */
+
+
+      SizeAtCompileTime = (internal::size_at_compile_time<internal::traits<Derived>::RowsAtCompileTime,
+                                                   internal::traits<Derived>::ColsAtCompileTime>::ret),
+        /**< This is equal to the number of coefficients, i.e. the number of
+          * rows times the number of columns, or to \a Dynamic if this is not
+          * known at compile-time. \sa RowsAtCompileTime, ColsAtCompileTime */
+
+      MaxRowsAtCompileTime = RowsAtCompileTime,
+      MaxColsAtCompileTime = ColsAtCompileTime,
+
+      MaxSizeAtCompileTime = (internal::size_at_compile_time<MaxRowsAtCompileTime,
+                                                      MaxColsAtCompileTime>::ret),
+
+      IsVectorAtCompileTime = RowsAtCompileTime == 1 || ColsAtCompileTime == 1,
+        /**< This is set to true if either the number of rows or the number of
+          * columns is known at compile-time to be equal to 1. Indeed, in that case,
+          * we are dealing with a column-vector (if there is only one column) or with
+          * a row-vector (if there is only one row). */
+
+      Flags = internal::traits<Derived>::Flags,
+        /**< This stores expression \ref flags flags which may or may not be inherited by new expressions
+          * constructed from this one. See the \ref flags "list of flags".
+          */
+
+      IsRowMajor = Flags&RowMajorBit ? 1 : 0,
+      
+      InnerSizeAtCompileTime = int(IsVectorAtCompileTime) ? int(SizeAtCompileTime)
+                             : int(IsRowMajor) ? int(ColsAtCompileTime) : int(RowsAtCompileTime),
+
+      #ifndef EIGEN_PARSED_BY_DOXYGEN
+      _HasDirectAccess = (int(Flags)&DirectAccessBit) ? 1 : 0 // workaround sunCC
+      #endif
+    };
+
+    /** \internal the return type of MatrixBase::adjoint() */
+    typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                        CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, Eigen::Transpose<const Derived> >,
+                        Transpose<const Derived>
+                     >::type AdjointReturnType;
+    typedef Transpose<Derived> TransposeReturnType;
+    typedef typename internal::add_const<Transpose<const Derived> >::type ConstTransposeReturnType;
+
+    // FIXME storage order do not match evaluator storage order
+    typedef SparseMatrix<Scalar, Flags&RowMajorBit ? RowMajor : ColMajor, StorageIndex> PlainObject;
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** This is the "real scalar" type; if the \a Scalar type is already real numbers
+      * (e.g. int, float or double) then \a RealScalar is just the same as \a Scalar. If
+      * \a Scalar is \a std::complex<T> then RealScalar is \a T.
+      *
+      * \sa class NumTraits
+      */
+    typedef typename NumTraits<Scalar>::Real RealScalar;
+
+    /** \internal the return type of coeff()
+      */
+    typedef typename internal::conditional<_HasDirectAccess, const Scalar&, Scalar>::type CoeffReturnType;
+
+    /** \internal Represents a matrix with all coefficients equal to one another*/
+    typedef CwiseNullaryOp<internal::scalar_constant_op<Scalar>,Matrix<Scalar,Dynamic,Dynamic> > ConstantReturnType;
+
+    /** type of the equivalent dense matrix */
+    typedef Matrix<Scalar,RowsAtCompileTime,ColsAtCompileTime> DenseMatrixType;
+    /** type of the equivalent square matrix */
+    typedef Matrix<Scalar,EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime),
+                          EIGEN_SIZE_MAX(RowsAtCompileTime,ColsAtCompileTime)> SquareMatrixType;
+
+    inline const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    inline Derived& derived() { return *static_cast<Derived*>(this); }
+    inline Derived& const_cast_derived() const
+    { return *static_cast<Derived*>(const_cast<SparseMatrixBase*>(this)); }
+
+    typedef EigenBase<Derived> Base;
+
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+#define EIGEN_CURRENT_STORAGE_BASE_CLASS Eigen::SparseMatrixBase
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+#define EIGEN_DOC_UNARY_ADDONS(METHOD,OP)           /** <p>This method does not change the sparsity of \c *this: the OP is applied to explicitly stored coefficients only. \sa SparseCompressedBase::coeffs() </p> */
+#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL      /** <p> \warning This method returns a read-only expression for any sparse matrices. \sa \ref TutorialSparse_SubMatrices "Sparse block operations" </p> */
+#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND) /** <p> \warning This method returns a read-write expression for COND sparse matrices only. Otherwise, the returned expression is read-only. \sa \ref TutorialSparse_SubMatrices "Sparse block operations" </p> */
+#else
+#define EIGEN_DOC_UNARY_ADDONS(X,Y)
+#define EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#define EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(COND)
+#endif
+#   include "../plugins/CommonCwiseUnaryOps.h"
+#   include "../plugins/CommonCwiseBinaryOps.h"
+#   include "../plugins/MatrixCwiseUnaryOps.h"
+#   include "../plugins/MatrixCwiseBinaryOps.h"
+#   include "../plugins/BlockMethods.h"
+#   ifdef EIGEN_SPARSEMATRIXBASE_PLUGIN
+#     include EIGEN_SPARSEMATRIXBASE_PLUGIN
+#   endif
+#undef EIGEN_CURRENT_STORAGE_BASE_CLASS
+#undef EIGEN_DOC_UNARY_ADDONS
+#undef EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+#undef EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF
+
+    /** \returns the number of rows. \sa cols() */
+    inline Index rows() const { return derived().rows(); }
+    /** \returns the number of columns. \sa rows() */
+    inline Index cols() const { return derived().cols(); }
+    /** \returns the number of coefficients, which is \a rows()*cols().
+      * \sa rows(), cols(). */
+    inline Index size() const { return rows() * cols(); }
+    /** \returns true if either the number of rows or the number of columns is equal to 1.
+      * In other words, this function returns
+      * \code rows()==1 || cols()==1 \endcode
+      * \sa rows(), cols(), IsVectorAtCompileTime. */
+    inline bool isVector() const { return rows()==1 || cols()==1; }
+    /** \returns the size of the storage major dimension,
+      * i.e., the number of columns for a columns major matrix, and the number of rows otherwise */
+    Index outerSize() const { return (int(Flags)&RowMajorBit) ? this->rows() : this->cols(); }
+    /** \returns the size of the inner dimension according to the storage order,
+      * i.e., the number of rows for a columns major matrix, and the number of cols otherwise */
+    Index innerSize() const { return (int(Flags)&RowMajorBit) ? this->cols() : this->rows(); }
+
+    bool isRValue() const { return m_isRValue; }
+    Derived& markAsRValue() { m_isRValue = true; return derived(); }
+
+    SparseMatrixBase() : m_isRValue(false) { /* TODO check flags */ }
+
+    
+    template<typename OtherDerived>
+    Derived& operator=(const ReturnByValue<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    inline Derived& operator=(const SparseMatrixBase<OtherDerived>& other);
+
+    inline Derived& operator=(const Derived& other);
+
+  protected:
+
+    template<typename OtherDerived>
+    inline Derived& assign(const OtherDerived& other);
+
+    template<typename OtherDerived>
+    inline void assignGeneric(const OtherDerived& other);
+
+  public:
+
+    friend std::ostream & operator << (std::ostream & s, const SparseMatrixBase& m)
+    {
+      typedef typename Derived::Nested Nested;
+      typedef typename internal::remove_all<Nested>::type NestedCleaned;
+
+      if (Flags&RowMajorBit)
+      {
+        Nested nm(m.derived());
+        internal::evaluator<NestedCleaned> thisEval(nm);
+        for (Index row=0; row<nm.outerSize(); ++row)
+        {
+          Index col = 0;
+          for (typename internal::evaluator<NestedCleaned>::InnerIterator it(thisEval, row); it; ++it)
+          {
+            for ( ; col<it.index(); ++col)
+              s << "0 ";
+            s << it.value() << " ";
+            ++col;
+          }
+          for ( ; col<m.cols(); ++col)
+            s << "0 ";
+          s << std::endl;
+        }
+      }
+      else
+      {
+        Nested nm(m.derived());
+        internal::evaluator<NestedCleaned> thisEval(nm);
+        if (m.cols() == 1) {
+          Index row = 0;
+          for (typename internal::evaluator<NestedCleaned>::InnerIterator it(thisEval, 0); it; ++it)
+          {
+            for ( ; row<it.index(); ++row)
+              s << "0" << std::endl;
+            s << it.value() << std::endl;
+            ++row;
+          }
+          for ( ; row<m.rows(); ++row)
+            s << "0" << std::endl;
+        }
+        else
+        {
+          SparseMatrix<Scalar, RowMajorBit, StorageIndex> trans = m;
+          s << static_cast<const SparseMatrixBase<SparseMatrix<Scalar, RowMajorBit, StorageIndex> >&>(trans);
+        }
+      }
+      return s;
+    }
+
+    template<typename OtherDerived>
+    Derived& operator+=(const SparseMatrixBase<OtherDerived>& other);
+    template<typename OtherDerived>
+    Derived& operator-=(const SparseMatrixBase<OtherDerived>& other);
+    
+    template<typename OtherDerived>
+    Derived& operator+=(const DiagonalBase<OtherDerived>& other);
+    template<typename OtherDerived>
+    Derived& operator-=(const DiagonalBase<OtherDerived>& other);
+
+    template<typename OtherDerived>
+    Derived& operator+=(const EigenBase<OtherDerived> &other);
+    template<typename OtherDerived>
+    Derived& operator-=(const EigenBase<OtherDerived> &other);
+
+    Derived& operator*=(const Scalar& other);
+    Derived& operator/=(const Scalar& other);
+
+    template<typename OtherDerived> struct CwiseProductDenseReturnType {
+      typedef CwiseBinaryOp<internal::scalar_product_op<typename ScalarBinaryOpTraits<
+                                                          typename internal::traits<Derived>::Scalar,
+                                                          typename internal::traits<OtherDerived>::Scalar
+                                                        >::ReturnType>,
+                            const Derived,
+                            const OtherDerived
+                          > Type;
+    };
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE const typename CwiseProductDenseReturnType<OtherDerived>::Type
+    cwiseProduct(const MatrixBase<OtherDerived> &other) const;
+
+    // sparse * diagonal
+    template<typename OtherDerived>
+    const Product<Derived,OtherDerived>
+    operator*(const DiagonalBase<OtherDerived> &other) const
+    { return Product<Derived,OtherDerived>(derived(), other.derived()); }
+
+    // diagonal * sparse
+    template<typename OtherDerived> friend
+    const Product<OtherDerived,Derived>
+    operator*(const DiagonalBase<OtherDerived> &lhs, const SparseMatrixBase& rhs)
+    { return Product<OtherDerived,Derived>(lhs.derived(), rhs.derived()); }
+    
+    // sparse * sparse
+    template<typename OtherDerived>
+    const Product<Derived,OtherDerived,AliasFreeProduct>
+    operator*(const SparseMatrixBase<OtherDerived> &other) const;
+    
+    // sparse * dense
+    template<typename OtherDerived>
+    const Product<Derived,OtherDerived>
+    operator*(const MatrixBase<OtherDerived> &other) const
+    { return Product<Derived,OtherDerived>(derived(), other.derived()); }
+    
+    // dense * sparse
+    template<typename OtherDerived> friend
+    const Product<OtherDerived,Derived>
+    operator*(const MatrixBase<OtherDerived> &lhs, const SparseMatrixBase& rhs)
+    { return Product<OtherDerived,Derived>(lhs.derived(), rhs.derived()); }
+    
+     /** \returns an expression of P H P^-1 where H is the matrix represented by \c *this */
+    SparseSymmetricPermutationProduct<Derived,Upper|Lower> twistedBy(const PermutationMatrix<Dynamic,Dynamic,StorageIndex>& perm) const
+    {
+      return SparseSymmetricPermutationProduct<Derived,Upper|Lower>(derived(), perm);
+    }
+
+    template<typename OtherDerived>
+    Derived& operator*=(const SparseMatrixBase<OtherDerived>& other);
+
+    template<int Mode>
+    inline const TriangularView<const Derived, Mode> triangularView() const;
+    
+    template<unsigned int UpLo> struct SelfAdjointViewReturnType { typedef SparseSelfAdjointView<Derived, UpLo> Type; };
+    template<unsigned int UpLo> struct ConstSelfAdjointViewReturnType { typedef const SparseSelfAdjointView<const Derived, UpLo> Type; };
+
+    template<unsigned int UpLo> inline 
+    typename ConstSelfAdjointViewReturnType<UpLo>::Type selfadjointView() const;
+    template<unsigned int UpLo> inline
+    typename SelfAdjointViewReturnType<UpLo>::Type selfadjointView();
+
+    template<typename OtherDerived> Scalar dot(const MatrixBase<OtherDerived>& other) const;
+    template<typename OtherDerived> Scalar dot(const SparseMatrixBase<OtherDerived>& other) const;
+    RealScalar squaredNorm() const;
+    RealScalar norm()  const;
+    RealScalar blueNorm() const;
+
+    TransposeReturnType transpose() { return TransposeReturnType(derived()); }
+    const ConstTransposeReturnType transpose() const { return ConstTransposeReturnType(derived()); }
+    const AdjointReturnType adjoint() const { return AdjointReturnType(transpose()); }
+
+    // inner-vector
+    typedef Block<Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true>       InnerVectorReturnType;
+    typedef Block<const Derived,IsRowMajor?1:Dynamic,IsRowMajor?Dynamic:1,true> ConstInnerVectorReturnType;
+    InnerVectorReturnType innerVector(Index outer);
+    const ConstInnerVectorReturnType innerVector(Index outer) const;
+
+    // set of inner-vectors
+    typedef Block<Derived,Dynamic,Dynamic,true> InnerVectorsReturnType;
+    typedef Block<const Derived,Dynamic,Dynamic,true> ConstInnerVectorsReturnType;
+    InnerVectorsReturnType innerVectors(Index outerStart, Index outerSize);
+    const ConstInnerVectorsReturnType innerVectors(Index outerStart, Index outerSize) const;
+
+    DenseMatrixType toDense() const
+    {
+      return DenseMatrixType(derived());
+    }
+
+    template<typename OtherDerived>
+    bool isApprox(const SparseMatrixBase<OtherDerived>& other,
+                  const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const;
+
+    template<typename OtherDerived>
+    bool isApprox(const MatrixBase<OtherDerived>& other,
+                  const RealScalar& prec = NumTraits<Scalar>::dummy_precision()) const
+    { return toDense().isApprox(other,prec); }
+
+    /** \returns the matrix or vector obtained by evaluating this expression.
+      *
+      * Notice that in the case of a plain matrix or vector (not an expression) this function just returns
+      * a const reference, in order to avoid a useless copy.
+      */
+    inline const typename internal::eval<Derived>::type eval() const
+    { return typename internal::eval<Derived>::type(derived()); }
+
+    Scalar sum() const;
+    
+    inline const SparseView<Derived>
+    pruned(const Scalar& reference = Scalar(0), const RealScalar& epsilon = NumTraits<Scalar>::dummy_precision()) const;
+
+  protected:
+
+    bool m_isRValue;
+
+    static inline StorageIndex convert_index(const Index idx) {
+      return internal::convert_index<StorageIndex>(idx);
+    }
+  private:
+    template<typename Dest> void evalTo(Dest &) const;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEMATRIXBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparsePermutation.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparsePermutation.h
new file mode 100644
index 0000000..ef38357
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparsePermutation.h
@@ -0,0 +1,178 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_PERMUTATION_H
+#define EIGEN_SPARSE_PERMUTATION_H
+
+// This file implements sparse * permutation products
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename ExpressionType, int Side, bool Transposed>
+struct permutation_matrix_product<ExpressionType, Side, Transposed, SparseShape>
+{
+    typedef typename nested_eval<ExpressionType, 1>::type MatrixType;
+    typedef typename remove_all<MatrixType>::type MatrixTypeCleaned;
+
+    typedef typename MatrixTypeCleaned::Scalar Scalar;
+    typedef typename MatrixTypeCleaned::StorageIndex StorageIndex;
+
+    enum {
+      SrcStorageOrder = MatrixTypeCleaned::Flags&RowMajorBit ? RowMajor : ColMajor,
+      MoveOuter = SrcStorageOrder==RowMajor ? Side==OnTheLeft : Side==OnTheRight
+    };
+    
+    typedef typename internal::conditional<MoveOuter,
+        SparseMatrix<Scalar,SrcStorageOrder,StorageIndex>,
+        SparseMatrix<Scalar,int(SrcStorageOrder)==RowMajor?ColMajor:RowMajor,StorageIndex> >::type ReturnType;
+
+    template<typename Dest,typename PermutationType>
+    static inline void run(Dest& dst, const PermutationType& perm, const ExpressionType& xpr)
+    {
+      MatrixType mat(xpr);
+      if(MoveOuter)
+      {
+        SparseMatrix<Scalar,SrcStorageOrder,StorageIndex> tmp(mat.rows(), mat.cols());
+        Matrix<StorageIndex,Dynamic,1> sizes(mat.outerSize());
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          Index jp = perm.indices().coeff(j);
+          sizes[((Side==OnTheLeft) ^ Transposed) ? jp : j] = StorageIndex(mat.innerVector(((Side==OnTheRight) ^ Transposed) ? jp : j).nonZeros());
+        }
+        tmp.reserve(sizes);
+        for(Index j=0; j<mat.outerSize(); ++j)
+        {
+          Index jp = perm.indices().coeff(j);
+          Index jsrc = ((Side==OnTheRight) ^ Transposed) ? jp : j;
+          Index jdst = ((Side==OnTheLeft) ^ Transposed) ? jp : j;
+          for(typename MatrixTypeCleaned::InnerIterator it(mat,jsrc); it; ++it)
+            tmp.insertByOuterInner(jdst,it.index()) = it.value();
+        }
+        dst = tmp;
+      }
+      else
+      {
+        SparseMatrix<Scalar,int(SrcStorageOrder)==RowMajor?ColMajor:RowMajor,StorageIndex> tmp(mat.rows(), mat.cols());
+        Matrix<StorageIndex,Dynamic,1> sizes(tmp.outerSize());
+        sizes.setZero();
+        PermutationMatrix<Dynamic,Dynamic,StorageIndex> perm_cpy;
+        if((Side==OnTheLeft) ^ Transposed)
+          perm_cpy = perm;
+        else
+          perm_cpy = perm.transpose();
+
+        for(Index j=0; j<mat.outerSize(); ++j)
+          for(typename MatrixTypeCleaned::InnerIterator it(mat,j); it; ++it)
+            sizes[perm_cpy.indices().coeff(it.index())]++;
+        tmp.reserve(sizes);
+        for(Index j=0; j<mat.outerSize(); ++j)
+          for(typename MatrixTypeCleaned::InnerIterator it(mat,j); it; ++it)
+            tmp.insertByOuterInner(perm_cpy.indices().coeff(it.index()),j) = it.value();
+        dst = tmp;
+      }
+    }
+};
+
+}
+
+namespace internal {
+
+template <int ProductTag> struct product_promote_storage_type<Sparse,             PermutationStorage, ProductTag> { typedef Sparse ret; };
+template <int ProductTag> struct product_promote_storage_type<PermutationStorage, Sparse,             ProductTag> { typedef Sparse ret; };
+
+// TODO, the following two overloads are only needed to define the right temporary type through 
+// typename traits<permutation_sparse_matrix_product<Rhs,Lhs,OnTheRight,false> >::ReturnType
+// whereas it should be correctly handled by traits<Product<> >::PlainObject
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, AliasFreeProduct>, ProductTag, PermutationShape, SparseShape>
+  : public evaluator<typename permutation_matrix_product<Rhs,OnTheLeft,false,SparseShape>::ReturnType>
+{
+  typedef Product<Lhs, Rhs, AliasFreeProduct> XprType;
+  typedef typename permutation_matrix_product<Rhs,OnTheLeft,false,SparseShape>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+
+  explicit product_evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    generic_product_impl<Lhs, Rhs, PermutationShape, SparseShape, ProductTag>::evalTo(m_result, xpr.lhs(), xpr.rhs());
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+template<typename Lhs, typename Rhs, int ProductTag>
+struct product_evaluator<Product<Lhs, Rhs, AliasFreeProduct>, ProductTag, SparseShape, PermutationShape >
+  : public evaluator<typename permutation_matrix_product<Lhs,OnTheRight,false,SparseShape>::ReturnType>
+{
+  typedef Product<Lhs, Rhs, AliasFreeProduct> XprType;
+  typedef typename permutation_matrix_product<Lhs,OnTheRight,false,SparseShape>::ReturnType PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  enum {
+    Flags = Base::Flags | EvalBeforeNestingBit
+  };
+
+  explicit product_evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    generic_product_impl<Lhs, Rhs, SparseShape, PermutationShape, ProductTag>::evalTo(m_result, xpr.lhs(), xpr.rhs());
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // end namespace internal
+
+/** \returns the matrix with the permutation applied to the columns
+  */
+template<typename SparseDerived, typename PermDerived>
+inline const Product<SparseDerived, PermDerived, AliasFreeProduct>
+operator*(const SparseMatrixBase<SparseDerived>& matrix, const PermutationBase<PermDerived>& perm)
+{ return Product<SparseDerived, PermDerived, AliasFreeProduct>(matrix.derived(), perm.derived()); }
+
+/** \returns the matrix with the permutation applied to the rows
+  */
+template<typename SparseDerived, typename PermDerived>
+inline const Product<PermDerived, SparseDerived, AliasFreeProduct>
+operator*( const PermutationBase<PermDerived>& perm, const SparseMatrixBase<SparseDerived>& matrix)
+{ return  Product<PermDerived, SparseDerived, AliasFreeProduct>(perm.derived(), matrix.derived()); }
+
+
+/** \returns the matrix with the inverse permutation applied to the columns.
+  */
+template<typename SparseDerived, typename PermutationType>
+inline const Product<SparseDerived, Inverse<PermutationType>, AliasFreeProduct>
+operator*(const SparseMatrixBase<SparseDerived>& matrix, const InverseImpl<PermutationType, PermutationStorage>& tperm)
+{
+  return Product<SparseDerived, Inverse<PermutationType>, AliasFreeProduct>(matrix.derived(), tperm.derived());
+}
+
+/** \returns the matrix with the inverse permutation applied to the rows.
+  */
+template<typename SparseDerived, typename PermutationType>
+inline const Product<Inverse<PermutationType>, SparseDerived, AliasFreeProduct>
+operator*(const InverseImpl<PermutationType,PermutationStorage>& tperm, const SparseMatrixBase<SparseDerived>& matrix)
+{
+  return Product<Inverse<PermutationType>, SparseDerived, AliasFreeProduct>(tperm.derived(), matrix.derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_SELFADJOINTVIEW_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseProduct.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseProduct.h
new file mode 100644
index 0000000..4cbf687
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseProduct.h
@@ -0,0 +1,169 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEPRODUCT_H
+#define EIGEN_SPARSEPRODUCT_H
+
+namespace Eigen { 
+
+/** \returns an expression of the product of two sparse matrices.
+  * By default a conservative product preserving the symbolic non zeros is performed.
+  * The automatic pruning of the small values can be achieved by calling the pruned() function
+  * in which case a totally different product algorithm is employed:
+  * \code
+  * C = (A*B).pruned();             // supress numerical zeros (exact)
+  * C = (A*B).pruned(ref);
+  * C = (A*B).pruned(ref,epsilon);
+  * \endcode
+  * where \c ref is a meaningful non zero reference value.
+  * */
+template<typename Derived>
+template<typename OtherDerived>
+inline const Product<Derived,OtherDerived,AliasFreeProduct>
+SparseMatrixBase<Derived>::operator*(const SparseMatrixBase<OtherDerived> &other) const
+{
+  return Product<Derived,OtherDerived,AliasFreeProduct>(derived(), other.derived());
+}
+
+namespace internal {
+
+// sparse * sparse
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, SparseShape, SparseShape, ProductType>
+{
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs)
+  {
+    evalTo(dst, lhs, rhs, typename evaluator_traits<Dest>::Shape());
+  }
+
+  // dense += sparse * sparse
+  template<typename Dest,typename ActualLhs>
+  static void addTo(Dest& dst, const ActualLhs& lhs, const Rhs& rhs, typename enable_if<is_same<typename evaluator_traits<Dest>::Shape,DenseShape>::value,int*>::type* = 0)
+  {
+    typedef typename nested_eval<ActualLhs,Dynamic>::type LhsNested;
+    typedef typename nested_eval<Rhs,Dynamic>::type RhsNested;
+    LhsNested lhsNested(lhs);
+    RhsNested rhsNested(rhs);
+    internal::sparse_sparse_to_dense_product_selector<typename remove_all<LhsNested>::type,
+                                                      typename remove_all<RhsNested>::type, Dest>::run(lhsNested,rhsNested,dst);
+  }
+
+  // dense -= sparse * sparse
+  template<typename Dest>
+  static void subTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, typename enable_if<is_same<typename evaluator_traits<Dest>::Shape,DenseShape>::value,int*>::type* = 0)
+  {
+    addTo(dst, -lhs, rhs);
+  }
+
+protected:
+
+  // sparse = sparse * sparse
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, SparseShape)
+  {
+    typedef typename nested_eval<Lhs,Dynamic>::type LhsNested;
+    typedef typename nested_eval<Rhs,Dynamic>::type RhsNested;
+    LhsNested lhsNested(lhs);
+    RhsNested rhsNested(rhs);
+    internal::conservative_sparse_sparse_product_selector<typename remove_all<LhsNested>::type,
+                                                          typename remove_all<RhsNested>::type, Dest>::run(lhsNested,rhsNested,dst);
+  }
+
+  // dense = sparse * sparse
+  template<typename Dest>
+  static void evalTo(Dest& dst, const Lhs& lhs, const Rhs& rhs, DenseShape)
+  {
+    dst.setZero();
+    addTo(dst, lhs, rhs);
+  }
+};
+
+// sparse * sparse-triangular
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, SparseShape, SparseTriangularShape, ProductType>
+ : public generic_product_impl<Lhs, Rhs, SparseShape, SparseShape, ProductType>
+{};
+
+// sparse-triangular * sparse
+template<typename Lhs, typename Rhs, int ProductType>
+struct generic_product_impl<Lhs, Rhs, SparseTriangularShape, SparseShape, ProductType>
+ : public generic_product_impl<Lhs, Rhs, SparseShape, SparseShape, ProductType>
+{};
+
+// dense = sparse-product (can be sparse*sparse, sparse*perm, etc.)
+template< typename DstXprType, typename Lhs, typename Rhs>
+struct Assignment<DstXprType, Product<Lhs,Rhs,AliasFreeProduct>, internal::assign_op<typename DstXprType::Scalar,typename Product<Lhs,Rhs,AliasFreeProduct>::Scalar>, Sparse2Dense>
+{
+  typedef Product<Lhs,Rhs,AliasFreeProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &)
+  {
+    Index dstRows = src.rows();
+    Index dstCols = src.cols();
+    if((dst.rows()!=dstRows) || (dst.cols()!=dstCols))
+      dst.resize(dstRows, dstCols);
+    
+    generic_product_impl<Lhs, Rhs>::evalTo(dst,src.lhs(),src.rhs());
+  }
+};
+
+// dense += sparse-product (can be sparse*sparse, sparse*perm, etc.)
+template< typename DstXprType, typename Lhs, typename Rhs>
+struct Assignment<DstXprType, Product<Lhs,Rhs,AliasFreeProduct>, internal::add_assign_op<typename DstXprType::Scalar,typename Product<Lhs,Rhs,AliasFreeProduct>::Scalar>, Sparse2Dense>
+{
+  typedef Product<Lhs,Rhs,AliasFreeProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &)
+  {
+    generic_product_impl<Lhs, Rhs>::addTo(dst,src.lhs(),src.rhs());
+  }
+};
+
+// dense -= sparse-product (can be sparse*sparse, sparse*perm, etc.)
+template< typename DstXprType, typename Lhs, typename Rhs>
+struct Assignment<DstXprType, Product<Lhs,Rhs,AliasFreeProduct>, internal::sub_assign_op<typename DstXprType::Scalar,typename Product<Lhs,Rhs,AliasFreeProduct>::Scalar>, Sparse2Dense>
+{
+  typedef Product<Lhs,Rhs,AliasFreeProduct> SrcXprType;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> &)
+  {
+    generic_product_impl<Lhs, Rhs>::subTo(dst,src.lhs(),src.rhs());
+  }
+};
+
+template<typename Lhs, typename Rhs, int Options>
+struct unary_evaluator<SparseView<Product<Lhs, Rhs, Options> >, IteratorBased>
+ : public evaluator<typename Product<Lhs, Rhs, DefaultProduct>::PlainObject>
+{
+  typedef SparseView<Product<Lhs, Rhs, Options> > XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  explicit unary_evaluator(const XprType& xpr)
+    : m_result(xpr.rows(), xpr.cols())
+  {
+    using std::abs;
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    typedef typename nested_eval<Lhs,Dynamic>::type LhsNested;
+    typedef typename nested_eval<Rhs,Dynamic>::type RhsNested;
+    LhsNested lhsNested(xpr.nestedExpression().lhs());
+    RhsNested rhsNested(xpr.nestedExpression().rhs());
+
+    internal::sparse_sparse_product_with_pruning_selector<typename remove_all<LhsNested>::type,
+                                                          typename remove_all<RhsNested>::type, PlainObject>::run(lhsNested,rhsNested,m_result,
+                                                                                                                  abs(xpr.reference())*xpr.epsilon());
+  }
+
+protected:
+  PlainObject m_result;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEPRODUCT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRedux.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRedux.h
new file mode 100644
index 0000000..4587749
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRedux.h
@@ -0,0 +1,49 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEREDUX_H
+#define EIGEN_SPARSEREDUX_H
+
+namespace Eigen { 
+
+template<typename Derived>
+typename internal::traits<Derived>::Scalar
+SparseMatrixBase<Derived>::sum() const
+{
+  eigen_assert(rows()>0 && cols()>0 && "you are using a non initialized matrix");
+  Scalar res(0);
+  internal::evaluator<Derived> thisEval(derived());
+  for (Index j=0; j<outerSize(); ++j)
+    for (typename internal::evaluator<Derived>::InnerIterator iter(thisEval,j); iter; ++iter)
+      res += iter.value();
+  return res;
+}
+
+template<typename _Scalar, int _Options, typename _Index>
+typename internal::traits<SparseMatrix<_Scalar,_Options,_Index> >::Scalar
+SparseMatrix<_Scalar,_Options,_Index>::sum() const
+{
+  eigen_assert(rows()>0 && cols()>0 && "you are using a non initialized matrix");
+  if(this->isCompressed())
+    return Matrix<Scalar,1,Dynamic>::Map(m_data.valuePtr(), m_data.size()).sum();
+  else
+    return Base::sum();
+}
+
+template<typename _Scalar, int _Options, typename _Index>
+typename internal::traits<SparseVector<_Scalar,_Options, _Index> >::Scalar
+SparseVector<_Scalar,_Options,_Index>::sum() const
+{
+  eigen_assert(rows()>0 && cols()>0 && "you are using a non initialized matrix");
+  return Matrix<Scalar,1,Dynamic>::Map(m_data.valuePtr(), m_data.size()).sum();
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEREDUX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRef.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRef.h
new file mode 100644
index 0000000..d91f38f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseRef.h
@@ -0,0 +1,397 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_REF_H
+#define EIGEN_SPARSE_REF_H
+
+namespace Eigen {
+
+enum {
+  StandardCompressedFormat = 2 /**< used by Ref<SparseMatrix> to specify whether the input storage must be in standard compressed form */
+};
+  
+namespace internal {
+
+template<typename Derived> class SparseRefBase;
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int _Options, typename _StrideType>
+struct traits<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+  : public traits<SparseMatrix<MatScalar,MatOptions,MatIndex> >
+{
+  typedef SparseMatrix<MatScalar,MatOptions,MatIndex> PlainObjectType;
+  enum {
+    Options = _Options,
+    Flags = traits<PlainObjectType>::Flags | CompressedAccessBit | NestByRefBit
+  };
+
+  template<typename Derived> struct match {
+    enum {
+      StorageOrderMatch = PlainObjectType::IsVectorAtCompileTime || Derived::IsVectorAtCompileTime || ((PlainObjectType::Flags&RowMajorBit)==(Derived::Flags&RowMajorBit)),
+      MatchAtCompileTime = (Derived::Flags&CompressedAccessBit) && StorageOrderMatch
+    };
+    typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
+  };
+  
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int _Options, typename _StrideType>
+struct traits<Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+  : public traits<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+{
+  enum {
+    Flags = (traits<SparseMatrix<MatScalar,MatOptions,MatIndex> >::Flags | CompressedAccessBit | NestByRefBit) & ~LvalueBit
+  };
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int _Options, typename _StrideType>
+struct traits<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+  : public traits<SparseVector<MatScalar,MatOptions,MatIndex> >
+{
+  typedef SparseVector<MatScalar,MatOptions,MatIndex> PlainObjectType;
+  enum {
+    Options = _Options,
+    Flags = traits<PlainObjectType>::Flags | CompressedAccessBit | NestByRefBit
+  };
+
+  template<typename Derived> struct match {
+    enum {
+      MatchAtCompileTime = (Derived::Flags&CompressedAccessBit) && Derived::IsVectorAtCompileTime
+    };
+    typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
+  };
+
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int _Options, typename _StrideType>
+struct traits<Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+  : public traits<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, _Options, _StrideType> >
+{
+  enum {
+    Flags = (traits<SparseVector<MatScalar,MatOptions,MatIndex> >::Flags | CompressedAccessBit | NestByRefBit) & ~LvalueBit
+  };
+};
+
+template<typename Derived>
+struct traits<SparseRefBase<Derived> > : public traits<Derived> {};
+
+template<typename Derived> class SparseRefBase
+  : public SparseMapBase<Derived>
+{
+public:
+
+  typedef SparseMapBase<Derived> Base;
+  EIGEN_SPARSE_PUBLIC_INTERFACE(SparseRefBase)
+
+  SparseRefBase()
+    : Base(RowsAtCompileTime==Dynamic?0:RowsAtCompileTime,ColsAtCompileTime==Dynamic?0:ColsAtCompileTime, 0, 0, 0, 0, 0)
+  {}
+  
+protected:
+
+  template<typename Expression>
+  void construct(Expression& expr)
+  {
+    if(expr.outerIndexPtr()==0)
+      ::new (static_cast<Base*>(this)) Base(expr.size(), expr.nonZeros(), expr.innerIndexPtr(), expr.valuePtr());
+    else
+      ::new (static_cast<Base*>(this)) Base(expr.rows(), expr.cols(), expr.nonZeros(), expr.outerIndexPtr(), expr.innerIndexPtr(), expr.valuePtr(), expr.innerNonZeroPtr());
+  }
+};
+
+} // namespace internal
+
+
+/** 
+  * \ingroup SparseCore_Module
+  *
+  * \brief A sparse matrix expression referencing an existing sparse expression
+  *
+  * \tparam SparseMatrixType the equivalent sparse matrix type of the referenced data, it must be a template instance of class SparseMatrix.
+  * \tparam Options specifies whether the a standard compressed format is required \c Options is  \c #StandardCompressedFormat, or \c 0.
+  *                The default is \c 0.
+  *
+  * \sa class Ref
+  */
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType >
+  : public internal::SparseRefBase<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType > >
+#else
+template<typename SparseMatrixType, int Options>
+class Ref<SparseMatrixType, Options>
+  : public SparseMapBase<Derived,WriteAccessors> // yes, that's weird to use Derived here, but that works!
+#endif
+{
+    typedef SparseMatrix<MatScalar,MatOptions,MatIndex> PlainObjectType;
+    typedef internal::traits<Ref> Traits;
+    template<int OtherOptions>
+    inline Ref(const SparseMatrix<MatScalar,OtherOptions,MatIndex>& expr);
+    template<int OtherOptions>
+    inline Ref(const MappedSparseMatrix<MatScalar,OtherOptions,MatIndex>& expr);
+  public:
+
+    typedef internal::SparseRefBase<Ref> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Ref)
+
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<int OtherOptions>
+    inline Ref(SparseMatrix<MatScalar,OtherOptions,MatIndex>& expr)
+    {
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<SparseMatrix<MatScalar,OtherOptions,MatIndex> >::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      eigen_assert( ((Options & int(StandardCompressedFormat))==0) || (expr.isCompressed()) );
+      Base::construct(expr.derived());
+    }
+    
+    template<int OtherOptions>
+    inline Ref(MappedSparseMatrix<MatScalar,OtherOptions,MatIndex>& expr)
+    {
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<SparseMatrix<MatScalar,OtherOptions,MatIndex> >::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      eigen_assert( ((Options & int(StandardCompressedFormat))==0) || (expr.isCompressed()) );
+      Base::construct(expr.derived());
+    }
+    
+    template<typename Derived>
+    inline Ref(const SparseCompressedBase<Derived>& expr)
+    #else
+    /** Implicit constructor from any sparse expression (2D matrix or 1D vector) */
+    template<typename Derived>
+    inline Ref(SparseCompressedBase<Derived>& expr)
+    #endif
+    {
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      eigen_assert( ((Options & int(StandardCompressedFormat))==0) || (expr.isCompressed()) );
+      Base::construct(expr.const_cast_derived());
+    }
+};
+
+// this is the const ref version
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType>
+  : public internal::SparseRefBase<Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+{
+    typedef SparseMatrix<MatScalar,MatOptions,MatIndex> TPlainObjectType;
+    typedef internal::traits<Ref> Traits;
+  public:
+
+    typedef internal::SparseRefBase<Ref> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Ref)
+
+    template<typename Derived>
+    inline Ref(const SparseMatrixBase<Derived>& expr) : m_hasCopy(false)
+    {
+      construct(expr.derived(), typename Traits::template match<Derived>::type());
+    }
+
+    inline Ref(const Ref& other) : Base(other), m_hasCopy(false) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    inline Ref(const RefBase<OtherRef>& other) : m_hasCopy(false) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
+
+    ~Ref() {
+      if(m_hasCopy) {
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        obj->~TPlainObjectType();
+      }
+    }
+
+  protected:
+
+    template<typename Expression>
+    void construct(const Expression& expr,internal::true_type)
+    {
+      if((Options & int(StandardCompressedFormat)) && (!expr.isCompressed()))
+      {
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        ::new (obj) TPlainObjectType(expr);
+        m_hasCopy = true;
+        Base::construct(*obj);
+      }
+      else
+      {
+        Base::construct(expr);
+      }
+    }
+
+    template<typename Expression>
+    void construct(const Expression& expr, internal::false_type)
+    {
+      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+      ::new (obj) TPlainObjectType(expr);
+      m_hasCopy = true;
+      Base::construct(*obj);
+    }
+
+  protected:
+    char m_object_bytes[sizeof(TPlainObjectType)];
+    bool m_hasCopy;
+};
+
+
+
+/**
+  * \ingroup SparseCore_Module
+  *
+  * \brief A sparse vector expression referencing an existing sparse vector expression
+  *
+  * \tparam SparseVectorType the equivalent sparse vector type of the referenced data, it must be a template instance of class SparseVector.
+  *
+  * \sa class Ref
+  */
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType >
+  : public internal::SparseRefBase<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType > >
+#else
+template<typename SparseVectorType>
+class Ref<SparseVectorType>
+  : public SparseMapBase<Derived,WriteAccessors>
+#endif
+{
+    typedef SparseVector<MatScalar,MatOptions,MatIndex> PlainObjectType;
+    typedef internal::traits<Ref> Traits;
+    template<int OtherOptions>
+    inline Ref(const SparseVector<MatScalar,OtherOptions,MatIndex>& expr);
+  public:
+
+    typedef internal::SparseRefBase<Ref> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Ref)
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<int OtherOptions>
+    inline Ref(SparseVector<MatScalar,OtherOptions,MatIndex>& expr)
+    {
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<SparseVector<MatScalar,OtherOptions,MatIndex> >::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      Base::construct(expr.derived());
+    }
+
+    template<typename Derived>
+    inline Ref(const SparseCompressedBase<Derived>& expr)
+    #else
+    /** Implicit constructor from any 1D sparse vector expression */
+    template<typename Derived>
+    inline Ref(SparseCompressedBase<Derived>& expr)
+    #endif
+    {
+      EIGEN_STATIC_ASSERT(bool(internal::is_lvalue<Derived>::value), THIS_EXPRESSION_IS_NOT_A_LVALUE__IT_IS_READ_ONLY);
+      EIGEN_STATIC_ASSERT(bool(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
+      Base::construct(expr.const_cast_derived());
+    }
+};
+
+// this is the const ref version
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+class Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType>
+  : public internal::SparseRefBase<Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+{
+    typedef SparseVector<MatScalar,MatOptions,MatIndex> TPlainObjectType;
+    typedef internal::traits<Ref> Traits;
+  public:
+
+    typedef internal::SparseRefBase<Ref> Base;
+    EIGEN_SPARSE_PUBLIC_INTERFACE(Ref)
+
+    template<typename Derived>
+    inline Ref(const SparseMatrixBase<Derived>& expr) : m_hasCopy(false)
+    {
+      construct(expr.derived(), typename Traits::template match<Derived>::type());
+    }
+
+    inline Ref(const Ref& other) : Base(other), m_hasCopy(false) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    inline Ref(const RefBase<OtherRef>& other) : m_hasCopy(false) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
+
+    ~Ref() {
+      if(m_hasCopy) {
+        TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+        obj->~TPlainObjectType();
+      }
+    }
+
+  protected:
+
+    template<typename Expression>
+    void construct(const Expression& expr,internal::true_type)
+    {
+      Base::construct(expr);
+    }
+
+    template<typename Expression>
+    void construct(const Expression& expr, internal::false_type)
+    {
+      TPlainObjectType* obj = reinterpret_cast<TPlainObjectType*>(m_object_bytes);
+      ::new (obj) TPlainObjectType(expr);
+      m_hasCopy = true;
+      Base::construct(*obj);
+    }
+
+  protected:
+    char m_object_bytes[sizeof(TPlainObjectType)];
+    bool m_hasCopy;
+};
+
+namespace internal {
+
+// FIXME shall we introduce a general evaluatior_ref that we can specialize for any sparse object once, and thus remove this copy-pasta thing...
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Ref<SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;  
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Ref<const SparseMatrix<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;  
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Ref<SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+template<typename MatScalar, int MatOptions, typename MatIndex, int Options, typename StrideType>
+struct evaluator<Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> >
+  : evaluator<SparseCompressedBase<Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> > >
+{
+  typedef evaluator<SparseCompressedBase<Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> > > Base;
+  typedef Ref<const SparseVector<MatScalar,MatOptions,MatIndex>, Options, StrideType> XprType;
+  evaluator() : Base() {}
+  explicit evaluator(const XprType &mat) : Base(mat) {}
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_REF_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h
new file mode 100644
index 0000000..5ab64f1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSelfAdjointView.h
@@ -0,0 +1,656 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_SELFADJOINTVIEW_H
+#define EIGEN_SPARSE_SELFADJOINTVIEW_H
+
+namespace Eigen { 
+  
+/** \ingroup SparseCore_Module
+  * \class SparseSelfAdjointView
+  *
+  * \brief Pseudo expression to manipulate a triangular sparse matrix as a selfadjoint matrix.
+  *
+  * \param MatrixType the type of the dense matrix storing the coefficients
+  * \param Mode can be either \c #Lower or \c #Upper
+  *
+  * This class is an expression of a sefladjoint matrix from a triangular part of a matrix
+  * with given dense storage of the coefficients. It is the return type of MatrixBase::selfadjointView()
+  * and most of the time this is the only way that it is used.
+  *
+  * \sa SparseMatrixBase::selfadjointView()
+  */
+namespace internal {
+  
+template<typename MatrixType, unsigned int Mode>
+struct traits<SparseSelfAdjointView<MatrixType,Mode> > : traits<MatrixType> {
+};
+
+template<int SrcMode,int DstMode,typename MatrixType,int DestOrder>
+void permute_symm_to_symm(const MatrixType& mat, SparseMatrix<typename MatrixType::Scalar,DestOrder,typename MatrixType::StorageIndex>& _dest, const typename MatrixType::StorageIndex* perm = 0);
+
+template<int Mode,typename MatrixType,int DestOrder>
+void permute_symm_to_fullsymm(const MatrixType& mat, SparseMatrix<typename MatrixType::Scalar,DestOrder,typename MatrixType::StorageIndex>& _dest, const typename MatrixType::StorageIndex* perm = 0);
+
+}
+
+template<typename MatrixType, unsigned int _Mode> class SparseSelfAdjointView
+  : public EigenBase<SparseSelfAdjointView<MatrixType,_Mode> >
+{
+  public:
+    
+    enum {
+      Mode = _Mode,
+      TransposeMode = ((Mode & Upper) ? Lower : 0) | ((Mode & Lower) ? Upper : 0),
+      RowsAtCompileTime = internal::traits<SparseSelfAdjointView>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<SparseSelfAdjointView>::ColsAtCompileTime
+    };
+
+    typedef EigenBase<SparseSelfAdjointView> Base;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+    typedef typename internal::ref_selector<MatrixType>::non_const_type MatrixTypeNested;
+    typedef typename internal::remove_all<MatrixTypeNested>::type _MatrixTypeNested;
+    
+    explicit inline SparseSelfAdjointView(MatrixType& matrix) : m_matrix(matrix)
+    {
+      eigen_assert(rows()==cols() && "SelfAdjointView is only for squared matrices");
+    }
+
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+
+    /** \internal \returns a reference to the nested matrix */
+    const _MatrixTypeNested& matrix() const { return m_matrix; }
+    typename internal::remove_reference<MatrixTypeNested>::type& matrix() { return m_matrix; }
+
+    /** \returns an expression of the matrix product between a sparse self-adjoint matrix \c *this and a sparse matrix \a rhs.
+      *
+      * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product.
+      * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product.
+      */
+    template<typename OtherDerived>
+    Product<SparseSelfAdjointView, OtherDerived>
+    operator*(const SparseMatrixBase<OtherDerived>& rhs) const
+    {
+      return Product<SparseSelfAdjointView, OtherDerived>(*this, rhs.derived());
+    }
+
+    /** \returns an expression of the matrix product between a sparse matrix \a lhs and a sparse self-adjoint matrix \a rhs.
+      *
+      * Note that there is no algorithmic advantage of performing such a product compared to a general sparse-sparse matrix product.
+      * Indeed, the SparseSelfadjointView operand is first copied into a temporary SparseMatrix before computing the product.
+      */
+    template<typename OtherDerived> friend
+    Product<OtherDerived, SparseSelfAdjointView>
+    operator*(const SparseMatrixBase<OtherDerived>& lhs, const SparseSelfAdjointView& rhs)
+    {
+      return Product<OtherDerived, SparseSelfAdjointView>(lhs.derived(), rhs);
+    }
+    
+    /** Efficient sparse self-adjoint matrix times dense vector/matrix product */
+    template<typename OtherDerived>
+    Product<SparseSelfAdjointView,OtherDerived>
+    operator*(const MatrixBase<OtherDerived>& rhs) const
+    {
+      return Product<SparseSelfAdjointView,OtherDerived>(*this, rhs.derived());
+    }
+
+    /** Efficient dense vector/matrix times sparse self-adjoint matrix product */
+    template<typename OtherDerived> friend
+    Product<OtherDerived,SparseSelfAdjointView>
+    operator*(const MatrixBase<OtherDerived>& lhs, const SparseSelfAdjointView& rhs)
+    {
+      return Product<OtherDerived,SparseSelfAdjointView>(lhs.derived(), rhs);
+    }
+
+    /** Perform a symmetric rank K update of the selfadjoint matrix \c *this:
+      * \f$ this = this + \alpha ( u u^* ) \f$ where \a u is a vector or matrix.
+      *
+      * \returns a reference to \c *this
+      *
+      * To perform \f$ this = this + \alpha ( u^* u ) \f$ you can simply
+      * call this function with u.adjoint().
+      */
+    template<typename DerivedU>
+    SparseSelfAdjointView& rankUpdate(const SparseMatrixBase<DerivedU>& u, const Scalar& alpha = Scalar(1));
+    
+    /** \returns an expression of P H P^-1 */
+    // TODO implement twists in a more evaluator friendly fashion
+    SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode> twistedBy(const PermutationMatrix<Dynamic,Dynamic,StorageIndex>& perm) const
+    {
+      return SparseSymmetricPermutationProduct<_MatrixTypeNested,Mode>(m_matrix, perm);
+    }
+
+    template<typename SrcMatrixType,int SrcMode>
+    SparseSelfAdjointView& operator=(const SparseSymmetricPermutationProduct<SrcMatrixType,SrcMode>& permutedMatrix)
+    {
+      internal::call_assignment_no_alias_no_transpose(*this, permutedMatrix);
+      return *this;
+    }
+
+    SparseSelfAdjointView& operator=(const SparseSelfAdjointView& src)
+    {
+      PermutationMatrix<Dynamic,Dynamic,StorageIndex> pnull;
+      return *this = src.twistedBy(pnull);
+    }
+
+    template<typename SrcMatrixType,unsigned int SrcMode>
+    SparseSelfAdjointView& operator=(const SparseSelfAdjointView<SrcMatrixType,SrcMode>& src)
+    {
+      PermutationMatrix<Dynamic,Dynamic,StorageIndex> pnull;
+      return *this = src.twistedBy(pnull);
+    }
+    
+    void resize(Index rows, Index cols)
+    {
+      EIGEN_ONLY_USED_FOR_DEBUG(rows);
+      EIGEN_ONLY_USED_FOR_DEBUG(cols);
+      eigen_assert(rows == this->rows() && cols == this->cols()
+                && "SparseSelfadjointView::resize() does not actually allow to resize.");
+    }
+    
+  protected:
+
+    MatrixTypeNested m_matrix;
+    //mutable VectorI m_countPerRow;
+    //mutable VectorI m_countPerCol;
+  private:
+    template<typename Dest> void evalTo(Dest &) const;
+};
+
+/***************************************************************************
+* Implementation of SparseMatrixBase methods
+***************************************************************************/
+
+template<typename Derived>
+template<unsigned int UpLo>
+typename SparseMatrixBase<Derived>::template ConstSelfAdjointViewReturnType<UpLo>::Type SparseMatrixBase<Derived>::selfadjointView() const
+{
+  return SparseSelfAdjointView<const Derived, UpLo>(derived());
+}
+
+template<typename Derived>
+template<unsigned int UpLo>
+typename SparseMatrixBase<Derived>::template SelfAdjointViewReturnType<UpLo>::Type SparseMatrixBase<Derived>::selfadjointView()
+{
+  return SparseSelfAdjointView<Derived, UpLo>(derived());
+}
+
+/***************************************************************************
+* Implementation of SparseSelfAdjointView methods
+***************************************************************************/
+
+template<typename MatrixType, unsigned int Mode>
+template<typename DerivedU>
+SparseSelfAdjointView<MatrixType,Mode>&
+SparseSelfAdjointView<MatrixType,Mode>::rankUpdate(const SparseMatrixBase<DerivedU>& u, const Scalar& alpha)
+{
+  SparseMatrix<Scalar,(MatrixType::Flags&RowMajorBit)?RowMajor:ColMajor> tmp = u * u.adjoint();
+  if(alpha==Scalar(0))
+    m_matrix = tmp.template triangularView<Mode>();
+  else
+    m_matrix += alpha * tmp.template triangularView<Mode>();
+
+  return *this;
+}
+
+namespace internal {
+  
+// TODO currently a selfadjoint expression has the form SelfAdjointView<.,.>
+//      in the future selfadjoint-ness should be defined by the expression traits
+//      such that Transpose<SelfAdjointView<.,.> > is valid. (currently TriangularBase::transpose() is overloaded to make it work)
+template<typename MatrixType, unsigned int Mode>
+struct evaluator_traits<SparseSelfAdjointView<MatrixType,Mode> >
+{
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SparseSelfAdjointShape Shape;
+};
+
+struct SparseSelfAdjoint2Sparse {};
+
+template<> struct AssignmentKind<SparseShape,SparseSelfAdjointShape> { typedef SparseSelfAdjoint2Sparse Kind; };
+template<> struct AssignmentKind<SparseSelfAdjointShape,SparseShape> { typedef Sparse2Sparse Kind; };
+
+template< typename DstXprType, typename SrcXprType, typename Functor>
+struct Assignment<DstXprType, SrcXprType, Functor, SparseSelfAdjoint2Sparse>
+{
+  typedef typename DstXprType::StorageIndex StorageIndex;
+  typedef internal::assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar> AssignOpType;
+
+  template<typename DestScalar,int StorageOrder>
+  static void run(SparseMatrix<DestScalar,StorageOrder,StorageIndex> &dst, const SrcXprType &src, const AssignOpType&/*func*/)
+  {
+    internal::permute_symm_to_fullsymm<SrcXprType::Mode>(src.matrix(), dst);
+  }
+
+  // FIXME: the handling of += and -= in sparse matrices should be cleanup so that next two overloads could be reduced to:
+  template<typename DestScalar,int StorageOrder,typename AssignFunc>
+  static void run(SparseMatrix<DestScalar,StorageOrder,StorageIndex> &dst, const SrcXprType &src, const AssignFunc& func)
+  {
+    SparseMatrix<DestScalar,StorageOrder,StorageIndex> tmp(src.rows(),src.cols());
+    run(tmp, src, AssignOpType());
+    call_assignment_no_alias_no_transpose(dst, tmp, func);
+  }
+
+  template<typename DestScalar,int StorageOrder>
+  static void run(SparseMatrix<DestScalar,StorageOrder,StorageIndex> &dst, const SrcXprType &src,
+                  const internal::add_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>& /* func */)
+  {
+    SparseMatrix<DestScalar,StorageOrder,StorageIndex> tmp(src.rows(),src.cols());
+    run(tmp, src, AssignOpType());
+    dst += tmp;
+  }
+
+  template<typename DestScalar,int StorageOrder>
+  static void run(SparseMatrix<DestScalar,StorageOrder,StorageIndex> &dst, const SrcXprType &src,
+                  const internal::sub_assign_op<typename DstXprType::Scalar,typename SrcXprType::Scalar>& /* func */)
+  {
+    SparseMatrix<DestScalar,StorageOrder,StorageIndex> tmp(src.rows(),src.cols());
+    run(tmp, src, AssignOpType());
+    dst -= tmp;
+  }
+  
+  template<typename DestScalar>
+  static void run(DynamicSparseMatrix<DestScalar,ColMajor,StorageIndex>& dst, const SrcXprType &src, const AssignOpType&/*func*/)
+  {
+    // TODO directly evaluate into dst;
+    SparseMatrix<DestScalar,ColMajor,StorageIndex> tmp(dst.rows(),dst.cols());
+    internal::permute_symm_to_fullsymm<SrcXprType::Mode>(src.matrix(), tmp);
+    dst = tmp;
+  }
+};
+
+} // end namespace internal
+
+/***************************************************************************
+* Implementation of sparse self-adjoint time dense matrix
+***************************************************************************/
+
+namespace internal {
+
+template<int Mode, typename SparseLhsType, typename DenseRhsType, typename DenseResType, typename AlphaType>
+inline void sparse_selfadjoint_time_dense_product(const SparseLhsType& lhs, const DenseRhsType& rhs, DenseResType& res, const AlphaType& alpha)
+{
+  EIGEN_ONLY_USED_FOR_DEBUG(alpha);
+  
+  typedef typename internal::nested_eval<SparseLhsType,DenseRhsType::MaxColsAtCompileTime>::type SparseLhsTypeNested;
+  typedef typename internal::remove_all<SparseLhsTypeNested>::type SparseLhsTypeNestedCleaned;
+  typedef evaluator<SparseLhsTypeNestedCleaned> LhsEval;
+  typedef typename LhsEval::InnerIterator LhsIterator;
+  typedef typename SparseLhsType::Scalar LhsScalar;
+  
+  enum {
+    LhsIsRowMajor = (LhsEval::Flags&RowMajorBit)==RowMajorBit,
+    ProcessFirstHalf =
+              ((Mode&(Upper|Lower))==(Upper|Lower))
+          || ( (Mode&Upper) && !LhsIsRowMajor)
+          || ( (Mode&Lower) && LhsIsRowMajor),
+    ProcessSecondHalf = !ProcessFirstHalf
+  };
+  
+  SparseLhsTypeNested lhs_nested(lhs);
+  LhsEval lhsEval(lhs_nested);
+
+  // work on one column at once
+  for (Index k=0; k<rhs.cols(); ++k)
+  {
+    for (Index j=0; j<lhs.outerSize(); ++j)
+    {
+      LhsIterator i(lhsEval,j);
+      // handle diagonal coeff
+      if (ProcessSecondHalf)
+      {
+        while (i && i.index()<j) ++i;
+        if(i && i.index()==j)
+        {
+          res(j,k) += alpha * i.value() * rhs(j,k);
+          ++i;
+        }
+      }
+
+      // premultiplied rhs for scatters
+      typename ScalarBinaryOpTraits<AlphaType, typename DenseRhsType::Scalar>::ReturnType rhs_j(alpha*rhs(j,k));
+      // accumulator for partial scalar product
+      typename DenseResType::Scalar res_j(0);
+      for(; (ProcessFirstHalf ? i && i.index() < j : i) ; ++i)
+      {
+        LhsScalar lhs_ij = i.value();
+        if(!LhsIsRowMajor) lhs_ij = numext::conj(lhs_ij);
+        res_j += lhs_ij * rhs(i.index(),k);
+        res(i.index(),k) += numext::conj(lhs_ij) * rhs_j;
+      }
+      res(j,k) += alpha * res_j;
+
+      // handle diagonal coeff
+      if (ProcessFirstHalf && i && (i.index()==j))
+        res(j,k) += alpha * i.value() * rhs(j,k);
+    }
+  }
+}
+
+
+template<typename LhsView, typename Rhs, int ProductType>
+struct generic_product_impl<LhsView, Rhs, SparseSelfAdjointShape, DenseShape, ProductType>
+: generic_product_impl_base<LhsView, Rhs, generic_product_impl<LhsView, Rhs, SparseSelfAdjointShape, DenseShape, ProductType> >
+{
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const LhsView& lhsView, const Rhs& rhs, const typename Dest::Scalar& alpha)
+  {
+    typedef typename LhsView::_MatrixTypeNested Lhs;
+    typedef typename nested_eval<Lhs,Dynamic>::type LhsNested;
+    typedef typename nested_eval<Rhs,Dynamic>::type RhsNested;
+    LhsNested lhsNested(lhsView.matrix());
+    RhsNested rhsNested(rhs);
+    
+    internal::sparse_selfadjoint_time_dense_product<LhsView::Mode>(lhsNested, rhsNested, dst, alpha);
+  }
+};
+
+template<typename Lhs, typename RhsView, int ProductType>
+struct generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, ProductType>
+: generic_product_impl_base<Lhs, RhsView, generic_product_impl<Lhs, RhsView, DenseShape, SparseSelfAdjointShape, ProductType> >
+{
+  template<typename Dest>
+  static void scaleAndAddTo(Dest& dst, const Lhs& lhs, const RhsView& rhsView, const typename Dest::Scalar& alpha)
+  {
+    typedef typename RhsView::_MatrixTypeNested Rhs;
+    typedef typename nested_eval<Lhs,Dynamic>::type LhsNested;
+    typedef typename nested_eval<Rhs,Dynamic>::type RhsNested;
+    LhsNested lhsNested(lhs);
+    RhsNested rhsNested(rhsView.matrix());
+    
+    // transpose everything
+    Transpose<Dest> dstT(dst);
+    internal::sparse_selfadjoint_time_dense_product<RhsView::TransposeMode>(rhsNested.transpose(), lhsNested.transpose(), dstT, alpha);
+  }
+};
+
+// NOTE: these two overloads are needed to evaluate the sparse selfadjoint view into a full sparse matrix
+// TODO: maybe the copy could be handled by generic_product_impl so that these overloads would not be needed anymore
+
+template<typename LhsView, typename Rhs, int ProductTag>
+struct product_evaluator<Product<LhsView, Rhs, DefaultProduct>, ProductTag, SparseSelfAdjointShape, SparseShape>
+  : public evaluator<typename Product<typename Rhs::PlainObject, Rhs, DefaultProduct>::PlainObject>
+{
+  typedef Product<LhsView, Rhs, DefaultProduct> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  product_evaluator(const XprType& xpr)
+    : m_lhs(xpr.lhs()), m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    generic_product_impl<typename Rhs::PlainObject, Rhs, SparseShape, SparseShape, ProductTag>::evalTo(m_result, m_lhs, xpr.rhs());
+  }
+  
+protected:
+  typename Rhs::PlainObject m_lhs;
+  PlainObject m_result;
+};
+
+template<typename Lhs, typename RhsView, int ProductTag>
+struct product_evaluator<Product<Lhs, RhsView, DefaultProduct>, ProductTag, SparseShape, SparseSelfAdjointShape>
+  : public evaluator<typename Product<Lhs, typename Lhs::PlainObject, DefaultProduct>::PlainObject>
+{
+  typedef Product<Lhs, RhsView, DefaultProduct> XprType;
+  typedef typename XprType::PlainObject PlainObject;
+  typedef evaluator<PlainObject> Base;
+
+  product_evaluator(const XprType& xpr)
+    : m_rhs(xpr.rhs()), m_result(xpr.rows(), xpr.cols())
+  {
+    ::new (static_cast<Base*>(this)) Base(m_result);
+    generic_product_impl<Lhs, typename Lhs::PlainObject, SparseShape, SparseShape, ProductTag>::evalTo(m_result, xpr.lhs(), m_rhs);
+  }
+  
+protected:
+  typename Lhs::PlainObject m_rhs;
+  PlainObject m_result;
+};
+
+} // namespace internal
+
+/***************************************************************************
+* Implementation of symmetric copies and permutations
+***************************************************************************/
+namespace internal {
+
+template<int Mode,typename MatrixType,int DestOrder>
+void permute_symm_to_fullsymm(const MatrixType& mat, SparseMatrix<typename MatrixType::Scalar,DestOrder,typename MatrixType::StorageIndex>& _dest, const typename MatrixType::StorageIndex* perm)
+{
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef typename MatrixType::Scalar Scalar;
+  typedef SparseMatrix<Scalar,DestOrder,StorageIndex> Dest;
+  typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+  typedef evaluator<MatrixType> MatEval;
+  typedef typename evaluator<MatrixType>::InnerIterator MatIterator;
+  
+  MatEval matEval(mat);
+  Dest& dest(_dest.derived());
+  enum {
+    StorageOrderMatch = int(Dest::IsRowMajor) == int(MatrixType::IsRowMajor)
+  };
+  
+  Index size = mat.rows();
+  VectorI count;
+  count.resize(size);
+  count.setZero();
+  dest.resize(size,size);
+  for(Index j = 0; j<size; ++j)
+  {
+    Index jp = perm ? perm[j] : j;
+    for(MatIterator it(matEval,j); it; ++it)
+    {
+      Index i = it.index();
+      Index r = it.row();
+      Index c = it.col();
+      Index ip = perm ? perm[i] : i;
+      if(Mode==(Upper|Lower))
+        count[StorageOrderMatch ? jp : ip]++;
+      else if(r==c)
+        count[ip]++;
+      else if(( Mode==Lower && r>c) || ( Mode==Upper && r<c))
+      {
+        count[ip]++;
+        count[jp]++;
+      }
+    }
+  }
+  Index nnz = count.sum();
+  
+  // reserve space
+  dest.resizeNonZeros(nnz);
+  dest.outerIndexPtr()[0] = 0;
+  for(Index j=0; j<size; ++j)
+    dest.outerIndexPtr()[j+1] = dest.outerIndexPtr()[j] + count[j];
+  for(Index j=0; j<size; ++j)
+    count[j] = dest.outerIndexPtr()[j];
+  
+  // copy data
+  for(StorageIndex j = 0; j<size; ++j)
+  {
+    for(MatIterator it(matEval,j); it; ++it)
+    {
+      StorageIndex i = internal::convert_index<StorageIndex>(it.index());
+      Index r = it.row();
+      Index c = it.col();
+      
+      StorageIndex jp = perm ? perm[j] : j;
+      StorageIndex ip = perm ? perm[i] : i;
+      
+      if(Mode==(Upper|Lower))
+      {
+        Index k = count[StorageOrderMatch ? jp : ip]++;
+        dest.innerIndexPtr()[k] = StorageOrderMatch ? ip : jp;
+        dest.valuePtr()[k] = it.value();
+      }
+      else if(r==c)
+      {
+        Index k = count[ip]++;
+        dest.innerIndexPtr()[k] = ip;
+        dest.valuePtr()[k] = it.value();
+      }
+      else if(( (Mode&Lower)==Lower && r>c) || ( (Mode&Upper)==Upper && r<c))
+      {
+        if(!StorageOrderMatch)
+          std::swap(ip,jp);
+        Index k = count[jp]++;
+        dest.innerIndexPtr()[k] = ip;
+        dest.valuePtr()[k] = it.value();
+        k = count[ip]++;
+        dest.innerIndexPtr()[k] = jp;
+        dest.valuePtr()[k] = numext::conj(it.value());
+      }
+    }
+  }
+}
+
+template<int _SrcMode,int _DstMode,typename MatrixType,int DstOrder>
+void permute_symm_to_symm(const MatrixType& mat, SparseMatrix<typename MatrixType::Scalar,DstOrder,typename MatrixType::StorageIndex>& _dest, const typename MatrixType::StorageIndex* perm)
+{
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef typename MatrixType::Scalar Scalar;
+  SparseMatrix<Scalar,DstOrder,StorageIndex>& dest(_dest.derived());
+  typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+  typedef evaluator<MatrixType> MatEval;
+  typedef typename evaluator<MatrixType>::InnerIterator MatIterator;
+
+  enum {
+    SrcOrder = MatrixType::IsRowMajor ? RowMajor : ColMajor,
+    StorageOrderMatch = int(SrcOrder) == int(DstOrder),
+    DstMode = DstOrder==RowMajor ? (_DstMode==Upper ? Lower : Upper) : _DstMode,
+    SrcMode = SrcOrder==RowMajor ? (_SrcMode==Upper ? Lower : Upper) : _SrcMode
+  };
+
+  MatEval matEval(mat);
+  
+  Index size = mat.rows();
+  VectorI count(size);
+  count.setZero();
+  dest.resize(size,size);
+  for(StorageIndex j = 0; j<size; ++j)
+  {
+    StorageIndex jp = perm ? perm[j] : j;
+    for(MatIterator it(matEval,j); it; ++it)
+    {
+      StorageIndex i = it.index();
+      if((int(SrcMode)==int(Lower) && i<j) || (int(SrcMode)==int(Upper) && i>j))
+        continue;
+                  
+      StorageIndex ip = perm ? perm[i] : i;
+      count[int(DstMode)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++;
+    }
+  }
+  dest.outerIndexPtr()[0] = 0;
+  for(Index j=0; j<size; ++j)
+    dest.outerIndexPtr()[j+1] = dest.outerIndexPtr()[j] + count[j];
+  dest.resizeNonZeros(dest.outerIndexPtr()[size]);
+  for(Index j=0; j<size; ++j)
+    count[j] = dest.outerIndexPtr()[j];
+  
+  for(StorageIndex j = 0; j<size; ++j)
+  {
+    
+    for(MatIterator it(matEval,j); it; ++it)
+    {
+      StorageIndex i = it.index();
+      if((int(SrcMode)==int(Lower) && i<j) || (int(SrcMode)==int(Upper) && i>j))
+        continue;
+                  
+      StorageIndex jp = perm ? perm[j] : j;
+      StorageIndex ip = perm? perm[i] : i;
+      
+      Index k = count[int(DstMode)==int(Lower) ? (std::min)(ip,jp) : (std::max)(ip,jp)]++;
+      dest.innerIndexPtr()[k] = int(DstMode)==int(Lower) ? (std::max)(ip,jp) : (std::min)(ip,jp);
+      
+      if(!StorageOrderMatch) std::swap(ip,jp);
+      if( ((int(DstMode)==int(Lower) && ip<jp) || (int(DstMode)==int(Upper) && ip>jp)))
+        dest.valuePtr()[k] = numext::conj(it.value());
+      else
+        dest.valuePtr()[k] = it.value();
+    }
+  }
+}
+
+}
+
+// TODO implement twists in a more evaluator friendly fashion
+
+namespace internal {
+
+template<typename MatrixType, int Mode>
+struct traits<SparseSymmetricPermutationProduct<MatrixType,Mode> > : traits<MatrixType> {
+};
+
+}
+
+template<typename MatrixType,int Mode>
+class SparseSymmetricPermutationProduct
+  : public EigenBase<SparseSymmetricPermutationProduct<MatrixType,Mode> >
+{
+  public:
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    enum {
+      RowsAtCompileTime = internal::traits<SparseSymmetricPermutationProduct>::RowsAtCompileTime,
+      ColsAtCompileTime = internal::traits<SparseSymmetricPermutationProduct>::ColsAtCompileTime
+    };
+  protected:
+    typedef PermutationMatrix<Dynamic,Dynamic,StorageIndex> Perm;
+  public:
+    typedef Matrix<StorageIndex,Dynamic,1> VectorI;
+    typedef typename MatrixType::Nested MatrixTypeNested;
+    typedef typename internal::remove_all<MatrixTypeNested>::type NestedExpression;
+    
+    SparseSymmetricPermutationProduct(const MatrixType& mat, const Perm& perm)
+      : m_matrix(mat), m_perm(perm)
+    {}
+    
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+        
+    const NestedExpression& matrix() const { return m_matrix; }
+    const Perm& perm() const { return m_perm; }
+    
+  protected:
+    MatrixTypeNested m_matrix;
+    const Perm& m_perm;
+
+};
+
+namespace internal {
+  
+template<typename DstXprType, typename MatrixType, int Mode, typename Scalar>
+struct Assignment<DstXprType, SparseSymmetricPermutationProduct<MatrixType,Mode>, internal::assign_op<Scalar,typename MatrixType::Scalar>, Sparse2Sparse>
+{
+  typedef SparseSymmetricPermutationProduct<MatrixType,Mode> SrcXprType;
+  typedef typename DstXprType::StorageIndex DstIndex;
+  template<int Options>
+  static void run(SparseMatrix<Scalar,Options,DstIndex> &dst, const SrcXprType &src, const internal::assign_op<Scalar,typename MatrixType::Scalar> &)
+  {
+    // internal::permute_symm_to_fullsymm<Mode>(m_matrix,_dest,m_perm.indices().data());
+    SparseMatrix<Scalar,(Options&RowMajor)==RowMajor ? ColMajor : RowMajor, DstIndex> tmp;
+    internal::permute_symm_to_fullsymm<Mode>(src.matrix(),tmp,src.perm().indices().data());
+    dst = tmp;
+  }
+  
+  template<typename DestType,unsigned int DestMode>
+  static void run(SparseSelfAdjointView<DestType,DestMode>& dst, const SrcXprType &src, const internal::assign_op<Scalar,typename MatrixType::Scalar> &)
+  {
+    internal::permute_symm_to_symm<Mode,DestMode>(src.matrix(),dst.matrix(),src.perm().indices().data());
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_SELFADJOINTVIEW_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSolverBase.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSolverBase.h
new file mode 100644
index 0000000..b4c9a42
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSolverBase.h
@@ -0,0 +1,124 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSESOLVERBASE_H
+#define EIGEN_SPARSESOLVERBASE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+  /** \internal
+  * Helper functions to solve with a sparse right-hand-side and result.
+  * The rhs is decomposed into small vertical panels which are solved through dense temporaries.
+  */
+template<typename Decomposition, typename Rhs, typename Dest>
+typename enable_if<Rhs::ColsAtCompileTime!=1 && Dest::ColsAtCompileTime!=1>::type
+solve_sparse_through_dense_panels(const Decomposition &dec, const Rhs& rhs, Dest &dest)
+{
+  EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0,THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+  typedef typename Dest::Scalar DestScalar;
+  // we process the sparse rhs per block of NbColsAtOnce columns temporarily stored into a dense matrix.
+  static const Index NbColsAtOnce = 4;
+  Index rhsCols = rhs.cols();
+  Index size = rhs.rows();
+  // the temporary matrices do not need more columns than NbColsAtOnce:
+  Index tmpCols = (std::min)(rhsCols, NbColsAtOnce); 
+  Eigen::Matrix<DestScalar,Dynamic,Dynamic> tmp(size,tmpCols);
+  Eigen::Matrix<DestScalar,Dynamic,Dynamic> tmpX(size,tmpCols);
+  for(Index k=0; k<rhsCols; k+=NbColsAtOnce)
+  {
+    Index actualCols = std::min<Index>(rhsCols-k, NbColsAtOnce);
+    tmp.leftCols(actualCols) = rhs.middleCols(k,actualCols);
+    tmpX.leftCols(actualCols) = dec.solve(tmp.leftCols(actualCols));
+    dest.middleCols(k,actualCols) = tmpX.leftCols(actualCols).sparseView();
+  }
+}
+
+// Overload for vector as rhs
+template<typename Decomposition, typename Rhs, typename Dest>
+typename enable_if<Rhs::ColsAtCompileTime==1 || Dest::ColsAtCompileTime==1>::type
+solve_sparse_through_dense_panels(const Decomposition &dec, const Rhs& rhs, Dest &dest)
+{
+  typedef typename Dest::Scalar DestScalar;
+  Index size = rhs.rows();
+  Eigen::Matrix<DestScalar,Dynamic,1> rhs_dense(rhs);
+  Eigen::Matrix<DestScalar,Dynamic,1> dest_dense(size);
+  dest_dense = dec.solve(rhs_dense);
+  dest = dest_dense.sparseView();
+}
+
+} // end namespace internal
+
+/** \class SparseSolverBase
+  * \ingroup SparseCore_Module
+  * \brief A base class for sparse solvers
+  *
+  * \tparam Derived the actual type of the solver.
+  *
+  */
+template<typename Derived>
+class SparseSolverBase : internal::noncopyable
+{
+  public:
+
+    /** Default constructor */
+    SparseSolverBase()
+      : m_isInitialized(false)
+    {}
+
+    ~SparseSolverBase()
+    {}
+
+    Derived& derived() { return *static_cast<Derived*>(this); }
+    const Derived& derived() const { return *static_cast<const Derived*>(this); }
+    
+    /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      *
+      * \sa compute()
+      */
+    template<typename Rhs>
+    inline const Solve<Derived, Rhs>
+    solve(const MatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "Solver is not initialized.");
+      eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<Derived, Rhs>(derived(), b.derived());
+    }
+    
+    /** \returns an expression of the solution x of \f$ A x = b \f$ using the current decomposition of A.
+      *
+      * \sa compute()
+      */
+    template<typename Rhs>
+    inline const Solve<Derived, Rhs>
+    solve(const SparseMatrixBase<Rhs>& b) const
+    {
+      eigen_assert(m_isInitialized && "Solver is not initialized.");
+      eigen_assert(derived().rows()==b.rows() && "solve(): invalid number of rows of the right hand side matrix b");
+      return Solve<Derived, Rhs>(derived(), b.derived());
+    }
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal default implementation of solving with a sparse rhs */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const SparseMatrixBase<Rhs> &b, SparseMatrixBase<Dest> &dest) const
+    {
+      internal::solve_sparse_through_dense_panels(derived(), b.derived(), dest.derived());
+    }
+    #endif // EIGEN_PARSED_BY_DOXYGEN
+
+  protected:
+    
+    mutable bool m_isInitialized;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSESOLVERBASE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h
new file mode 100644
index 0000000..21c4190
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseSparseProductWithPruning.h
@@ -0,0 +1,198 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSESPARSEPRODUCTWITHPRUNING_H
+#define EIGEN_SPARSESPARSEPRODUCTWITHPRUNING_H
+
+namespace Eigen { 
+
+namespace internal {
+
+
+// perform a pseudo in-place sparse * sparse product assuming all matrices are col major
+template<typename Lhs, typename Rhs, typename ResultType>
+static void sparse_sparse_product_with_pruning_impl(const Lhs& lhs, const Rhs& rhs, ResultType& res, const typename ResultType::RealScalar& tolerance)
+{
+  // return sparse_sparse_product_with_pruning_impl2(lhs,rhs,res);
+
+  typedef typename remove_all<Lhs>::type::Scalar Scalar;
+  typedef typename remove_all<Lhs>::type::StorageIndex StorageIndex;
+
+  // make sure to call innerSize/outerSize since we fake the storage order.
+  Index rows = lhs.innerSize();
+  Index cols = rhs.outerSize();
+  //Index size = lhs.outerSize();
+  eigen_assert(lhs.outerSize() == rhs.innerSize());
+
+  // allocate a temporary buffer
+  AmbiVector<Scalar,StorageIndex> tempVector(rows);
+
+  // mimics a resizeByInnerOuter:
+  if(ResultType::IsRowMajor)
+    res.resize(cols, rows);
+  else
+    res.resize(rows, cols);
+  
+  evaluator<Lhs> lhsEval(lhs);
+  evaluator<Rhs> rhsEval(rhs);
+  
+  // estimate the number of non zero entries
+  // given a rhs column containing Y non zeros, we assume that the respective Y columns
+  // of the lhs differs in average of one non zeros, thus the number of non zeros for
+  // the product of a rhs column with the lhs is X+Y where X is the average number of non zero
+  // per column of the lhs.
+  // Therefore, we have nnz(lhs*rhs) = nnz(lhs) + nnz(rhs)
+  Index estimated_nnz_prod = lhsEval.nonZerosEstimate() + rhsEval.nonZerosEstimate();
+
+  res.reserve(estimated_nnz_prod);
+  double ratioColRes = double(estimated_nnz_prod)/(double(lhs.rows())*double(rhs.cols()));
+  for (Index j=0; j<cols; ++j)
+  {
+    // FIXME:
+    //double ratioColRes = (double(rhs.innerVector(j).nonZeros()) + double(lhs.nonZeros())/double(lhs.cols()))/double(lhs.rows());
+    // let's do a more accurate determination of the nnz ratio for the current column j of res
+    tempVector.init(ratioColRes);
+    tempVector.setZero();
+    for (typename evaluator<Rhs>::InnerIterator rhsIt(rhsEval, j); rhsIt; ++rhsIt)
+    {
+      // FIXME should be written like this: tmp += rhsIt.value() * lhs.col(rhsIt.index())
+      tempVector.restart();
+      Scalar x = rhsIt.value();
+      for (typename evaluator<Lhs>::InnerIterator lhsIt(lhsEval, rhsIt.index()); lhsIt; ++lhsIt)
+      {
+        tempVector.coeffRef(lhsIt.index()) += lhsIt.value() * x;
+      }
+    }
+    res.startVec(j);
+    for (typename AmbiVector<Scalar,StorageIndex>::Iterator it(tempVector,tolerance); it; ++it)
+      res.insertBackByOuterInner(j,it.index()) = it.value();
+  }
+  res.finalize();
+}
+
+template<typename Lhs, typename Rhs, typename ResultType,
+  int LhsStorageOrder = traits<Lhs>::Flags&RowMajorBit,
+  int RhsStorageOrder = traits<Rhs>::Flags&RowMajorBit,
+  int ResStorageOrder = traits<ResultType>::Flags&RowMajorBit>
+struct sparse_sparse_product_with_pruning_selector;
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,ColMajor>
+{
+  typedef typename traits<typename remove_all<Lhs>::type>::Scalar Scalar;
+  typedef typename ResultType::RealScalar RealScalar;
+
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typename remove_all<ResultType>::type _res(res.rows(), res.cols());
+    internal::sparse_sparse_product_with_pruning_impl<Lhs,Rhs,ResultType>(lhs, rhs, _res, tolerance);
+    res.swap(_res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,ColMajor,RowMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    // we need a col-major matrix to hold the result
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename ResultType::StorageIndex> SparseTemporaryType;
+    SparseTemporaryType _res(res.rows(), res.cols());
+    internal::sparse_sparse_product_with_pruning_impl<Lhs,Rhs,SparseTemporaryType>(lhs, rhs, _res, tolerance);
+    res = _res;
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,RowMajor,RowMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    // let's transpose the product to get a column x column product
+    typename remove_all<ResultType>::type _res(res.rows(), res.cols());
+    internal::sparse_sparse_product_with_pruning_impl<Rhs,Lhs,ResultType>(rhs, lhs, _res, tolerance);
+    res.swap(_res);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,RowMajor,ColMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
+    ColMajorMatrixLhs colLhs(lhs);
+    ColMajorMatrixRhs colRhs(rhs);
+    internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,ColMajorMatrixRhs,ResultType>(colLhs, colRhs, res, tolerance);
+
+    // let's transpose the product to get a column x column product
+//     typedef SparseMatrix<typename ResultType::Scalar> SparseTemporaryType;
+//     SparseTemporaryType _res(res.cols(), res.rows());
+//     sparse_sparse_product_with_pruning_impl<Rhs,Lhs,SparseTemporaryType>(rhs, lhs, _res);
+//     res = _res.transpose();
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,RowMajor,RowMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixLhs;
+    RowMajorMatrixLhs rowLhs(lhs);
+    sparse_sparse_product_with_pruning_selector<RowMajorMatrixLhs,Rhs,ResultType,RowMajor,RowMajor>(rowLhs,rhs,res,tolerance);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,ColMajor,RowMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,RowMajor,typename Lhs::StorageIndex> RowMajorMatrixRhs;
+    RowMajorMatrixRhs rowRhs(rhs);
+    sparse_sparse_product_with_pruning_selector<Lhs,RowMajorMatrixRhs,ResultType,RowMajor,RowMajor,RowMajor>(lhs,rowRhs,res,tolerance);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,ColMajor,RowMajor,ColMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixRhs;
+    ColMajorMatrixRhs colRhs(rhs);
+    internal::sparse_sparse_product_with_pruning_impl<Lhs,ColMajorMatrixRhs,ResultType>(lhs, colRhs, res, tolerance);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename ResultType>
+struct sparse_sparse_product_with_pruning_selector<Lhs,Rhs,ResultType,RowMajor,ColMajor,ColMajor>
+{
+  typedef typename ResultType::RealScalar RealScalar;
+  static void run(const Lhs& lhs, const Rhs& rhs, ResultType& res, const RealScalar& tolerance)
+  {
+    typedef SparseMatrix<typename ResultType::Scalar,ColMajor,typename Lhs::StorageIndex> ColMajorMatrixLhs;
+    ColMajorMatrixLhs colLhs(lhs);
+    internal::sparse_sparse_product_with_pruning_impl<ColMajorMatrixLhs,Rhs,ResultType>(colLhs, rhs, res, tolerance);
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSESPARSEPRODUCTWITHPRUNING_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTranspose.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTranspose.h
new file mode 100644
index 0000000..3757d4c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTranspose.h
@@ -0,0 +1,92 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSETRANSPOSE_H
+#define EIGEN_SPARSETRANSPOSE_H
+
+namespace Eigen { 
+
+namespace internal {
+  template<typename MatrixType,int CompressedAccess=int(MatrixType::Flags&CompressedAccessBit)>
+  class SparseTransposeImpl
+    : public SparseMatrixBase<Transpose<MatrixType> >
+  {};
+  
+  template<typename MatrixType>
+  class SparseTransposeImpl<MatrixType,CompressedAccessBit>
+    : public SparseCompressedBase<Transpose<MatrixType> >
+  {
+    typedef SparseCompressedBase<Transpose<MatrixType> > Base;
+  public:
+    using Base::derived;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::StorageIndex StorageIndex;
+
+    inline Index nonZeros() const { return derived().nestedExpression().nonZeros(); }
+    
+    inline const Scalar* valuePtr() const { return derived().nestedExpression().valuePtr(); }
+    inline const StorageIndex* innerIndexPtr() const { return derived().nestedExpression().innerIndexPtr(); }
+    inline const StorageIndex* outerIndexPtr() const { return derived().nestedExpression().outerIndexPtr(); }
+    inline const StorageIndex* innerNonZeroPtr() const { return derived().nestedExpression().innerNonZeroPtr(); }
+
+    inline Scalar* valuePtr() { return derived().nestedExpression().valuePtr(); }
+    inline StorageIndex* innerIndexPtr() { return derived().nestedExpression().innerIndexPtr(); }
+    inline StorageIndex* outerIndexPtr() { return derived().nestedExpression().outerIndexPtr(); }
+    inline StorageIndex* innerNonZeroPtr() { return derived().nestedExpression().innerNonZeroPtr(); }
+  };
+}
+  
+template<typename MatrixType> class TransposeImpl<MatrixType,Sparse>
+  : public internal::SparseTransposeImpl<MatrixType>
+{
+  protected:
+    typedef internal::SparseTransposeImpl<MatrixType> Base;
+};
+
+namespace internal {
+  
+template<typename ArgType>
+struct unary_evaluator<Transpose<ArgType>, IteratorBased>
+  : public evaluator_base<Transpose<ArgType> >
+{
+    typedef typename evaluator<ArgType>::InnerIterator        EvalIterator;
+  public:
+    typedef Transpose<ArgType> XprType;
+    
+    inline Index nonZerosEstimate() const {
+      return m_argImpl.nonZerosEstimate();
+    }
+
+    class InnerIterator : public EvalIterator
+    {
+    public:
+      EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& unaryOp, Index outer)
+        : EvalIterator(unaryOp.m_argImpl,outer)
+      {}
+      
+      Index row() const { return EvalIterator::col(); }
+      Index col() const { return EvalIterator::row(); }
+    };
+    
+    enum {
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+      Flags = XprType::Flags
+    };
+    
+    explicit unary_evaluator(const XprType& op) :m_argImpl(op.nestedExpression()) {}
+
+  protected:
+    evaluator<ArgType> m_argImpl;
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSETRANSPOSE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTriangularView.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTriangularView.h
new file mode 100644
index 0000000..9ac1202
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseTriangularView.h
@@ -0,0 +1,189 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_TRIANGULARVIEW_H
+#define EIGEN_SPARSE_TRIANGULARVIEW_H
+
+namespace Eigen {
+
+/** \ingroup SparseCore_Module
+  *
+  * \brief Base class for a triangular part in a \b sparse matrix
+  *
+  * This class is an abstract base class of class TriangularView, and objects of type TriangularViewImpl cannot be instantiated.
+  * It extends class TriangularView with additional methods which are available for sparse expressions only.
+  *
+  * \sa class TriangularView, SparseMatrixBase::triangularView()
+  */
+template<typename MatrixType, unsigned int Mode> class TriangularViewImpl<MatrixType,Mode,Sparse>
+  : public SparseMatrixBase<TriangularView<MatrixType,Mode> >
+{
+    enum { SkipFirst = ((Mode&Lower) && !(MatrixType::Flags&RowMajorBit))
+                    || ((Mode&Upper) &&  (MatrixType::Flags&RowMajorBit)),
+           SkipLast = !SkipFirst,
+           SkipDiag = (Mode&ZeroDiag) ? 1 : 0,
+           HasUnitDiag = (Mode&UnitDiag) ? 1 : 0
+    };
+    
+    typedef TriangularView<MatrixType,Mode> TriangularViewType;
+    
+  protected:
+    // dummy solve function to make TriangularView happy.
+    void solve() const;
+
+    typedef SparseMatrixBase<TriangularViewType> Base;
+  public:
+    
+    EIGEN_SPARSE_PUBLIC_INTERFACE(TriangularViewType)
+    
+    typedef typename MatrixType::Nested MatrixTypeNested;
+    typedef typename internal::remove_reference<MatrixTypeNested>::type MatrixTypeNestedNonRef;
+    typedef typename internal::remove_all<MatrixTypeNested>::type MatrixTypeNestedCleaned;
+
+    template<typename RhsType, typename DstType>
+    EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE void _solve_impl(const RhsType &rhs, DstType &dst) const {
+      if(!(internal::is_same<RhsType,DstType>::value && internal::extract_data(dst) == internal::extract_data(rhs)))
+        dst = rhs;
+      this->solveInPlace(dst);
+    }
+
+    /** Applies the inverse of \c *this to the dense vector or matrix \a other, "in-place" */
+    template<typename OtherDerived> void solveInPlace(MatrixBase<OtherDerived>& other) const;
+
+    /** Applies the inverse of \c *this to the sparse vector or matrix \a other, "in-place" */
+    template<typename OtherDerived> void solveInPlace(SparseMatrixBase<OtherDerived>& other) const;
+  
+};
+
+namespace internal {
+
+template<typename ArgType, unsigned int Mode>
+struct unary_evaluator<TriangularView<ArgType,Mode>, IteratorBased>
+ : evaluator_base<TriangularView<ArgType,Mode> >
+{
+  typedef TriangularView<ArgType,Mode> XprType;
+  
+protected:
+  
+  typedef typename XprType::Scalar Scalar;
+  typedef typename XprType::StorageIndex StorageIndex;
+  typedef typename evaluator<ArgType>::InnerIterator EvalIterator;
+  
+  enum { SkipFirst = ((Mode&Lower) && !(ArgType::Flags&RowMajorBit))
+                    || ((Mode&Upper) &&  (ArgType::Flags&RowMajorBit)),
+         SkipLast = !SkipFirst,
+         SkipDiag = (Mode&ZeroDiag) ? 1 : 0,
+         HasUnitDiag = (Mode&UnitDiag) ? 1 : 0
+  };
+  
+public:
+  
+  enum {
+    CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+    Flags = XprType::Flags
+  };
+    
+  explicit unary_evaluator(const XprType &xpr) : m_argImpl(xpr.nestedExpression()), m_arg(xpr.nestedExpression()) {}
+  
+  inline Index nonZerosEstimate() const {
+    return m_argImpl.nonZerosEstimate();
+  }
+  
+  class InnerIterator : public EvalIterator
+  {
+      typedef EvalIterator Base;
+    public:
+
+      EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& xprEval, Index outer)
+        : Base(xprEval.m_argImpl,outer), m_returnOne(false), m_containsDiag(Base::outer()<xprEval.m_arg.innerSize())
+      {
+        if(SkipFirst)
+        {
+          while((*this) && ((HasUnitDiag||SkipDiag)  ? this->index()<=outer : this->index()<outer))
+            Base::operator++();
+          if(HasUnitDiag)
+            m_returnOne = m_containsDiag;
+        }
+        else if(HasUnitDiag && ((!Base::operator bool()) || Base::index()>=Base::outer()))
+        {
+          if((!SkipFirst) && Base::operator bool())
+            Base::operator++();
+          m_returnOne = m_containsDiag;
+        }
+      }
+
+      EIGEN_STRONG_INLINE InnerIterator& operator++()
+      {
+        if(HasUnitDiag && m_returnOne)
+          m_returnOne = false;
+        else
+        {
+          Base::operator++();
+          if(HasUnitDiag && (!SkipFirst) && ((!Base::operator bool()) || Base::index()>=Base::outer()))
+          {
+            if((!SkipFirst) && Base::operator bool())
+              Base::operator++();
+            m_returnOne = m_containsDiag;
+          }
+        }
+        return *this;
+      }
+      
+      EIGEN_STRONG_INLINE operator bool() const
+      {
+        if(HasUnitDiag && m_returnOne)
+          return true;
+        if(SkipFirst) return  Base::operator bool();
+        else
+        {
+          if (SkipDiag) return (Base::operator bool() && this->index() < this->outer());
+          else return (Base::operator bool() && this->index() <= this->outer());
+        }
+      }
+
+//       inline Index row() const { return (ArgType::Flags&RowMajorBit ? Base::outer() : this->index()); }
+//       inline Index col() const { return (ArgType::Flags&RowMajorBit ? this->index() : Base::outer()); }
+      inline StorageIndex index() const
+      {
+        if(HasUnitDiag && m_returnOne)  return internal::convert_index<StorageIndex>(Base::outer());
+        else                            return Base::index();
+      }
+      inline Scalar value() const
+      {
+        if(HasUnitDiag && m_returnOne)  return Scalar(1);
+        else                            return Base::value();
+      }
+
+    protected:
+      bool m_returnOne;
+      bool m_containsDiag;
+    private:
+      Scalar& valueRef();
+  };
+  
+protected:
+  evaluator<ArgType> m_argImpl;
+  const ArgType& m_arg;
+};
+
+} // end namespace internal
+
+template<typename Derived>
+template<int Mode>
+inline const TriangularView<const Derived, Mode>
+SparseMatrixBase<Derived>::triangularView() const
+{
+  return TriangularView<const Derived, Mode>(derived());
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSE_TRIANGULARVIEW_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseUtil.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseUtil.h
new file mode 100644
index 0000000..74df0d4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseUtil.h
@@ -0,0 +1,178 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEUTIL_H
+#define EIGEN_SPARSEUTIL_H
+
+namespace Eigen { 
+
+#ifdef NDEBUG
+#define EIGEN_DBG_SPARSE(X)
+#else
+#define EIGEN_DBG_SPARSE(X) X
+#endif
+
+#define EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(Derived, Op) \
+template<typename OtherDerived> \
+EIGEN_STRONG_INLINE Derived& operator Op(const Eigen::SparseMatrixBase<OtherDerived>& other) \
+{ \
+  return Base::operator Op(other.derived()); \
+} \
+EIGEN_STRONG_INLINE Derived& operator Op(const Derived& other) \
+{ \
+  return Base::operator Op(other); \
+}
+
+#define EIGEN_SPARSE_INHERIT_SCALAR_ASSIGNMENT_OPERATOR(Derived, Op) \
+template<typename Other> \
+EIGEN_STRONG_INLINE Derived& operator Op(const Other& scalar) \
+{ \
+  return Base::operator Op(scalar); \
+}
+
+#define EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATORS(Derived) \
+EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(Derived, =)
+
+
+#define EIGEN_SPARSE_PUBLIC_INTERFACE(Derived) \
+  EIGEN_GENERIC_PUBLIC_INTERFACE(Derived)
+
+  
+const int CoherentAccessPattern     = 0x1;
+const int InnerRandomAccessPattern  = 0x2 | CoherentAccessPattern;
+const int OuterRandomAccessPattern  = 0x4 | CoherentAccessPattern;
+const int RandomAccessPattern       = 0x8 | OuterRandomAccessPattern | InnerRandomAccessPattern;
+
+template<typename _Scalar, int _Flags = 0, typename _StorageIndex = int>  class SparseMatrix;
+template<typename _Scalar, int _Flags = 0, typename _StorageIndex = int>  class DynamicSparseMatrix;
+template<typename _Scalar, int _Flags = 0, typename _StorageIndex = int>  class SparseVector;
+template<typename _Scalar, int _Flags = 0, typename _StorageIndex = int>  class MappedSparseMatrix;
+
+template<typename MatrixType, unsigned int UpLo>  class SparseSelfAdjointView;
+template<typename Lhs, typename Rhs>              class SparseDiagonalProduct;
+template<typename MatrixType> class SparseView;
+
+template<typename Lhs, typename Rhs>        class SparseSparseProduct;
+template<typename Lhs, typename Rhs>        class SparseTimeDenseProduct;
+template<typename Lhs, typename Rhs>        class DenseTimeSparseProduct;
+template<typename Lhs, typename Rhs, bool Transpose> class SparseDenseOuterProduct;
+
+template<typename Lhs, typename Rhs> struct SparseSparseProductReturnType;
+template<typename Lhs, typename Rhs,
+         int InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(internal::traits<Lhs>::ColsAtCompileTime,internal::traits<Rhs>::RowsAtCompileTime)> struct DenseSparseProductReturnType;
+         
+template<typename Lhs, typename Rhs,
+         int InnerSize = EIGEN_SIZE_MIN_PREFER_FIXED(internal::traits<Lhs>::ColsAtCompileTime,internal::traits<Rhs>::RowsAtCompileTime)> struct SparseDenseProductReturnType;
+template<typename MatrixType,int UpLo> class SparseSymmetricPermutationProduct;
+
+namespace internal {
+
+template<typename T,int Rows,int Cols,int Flags> struct sparse_eval;
+
+template<typename T> struct eval<T,Sparse>
+  : sparse_eval<T, traits<T>::RowsAtCompileTime,traits<T>::ColsAtCompileTime,traits<T>::Flags>
+{};
+
+template<typename T,int Cols,int Flags> struct sparse_eval<T,1,Cols,Flags> {
+    typedef typename traits<T>::Scalar _Scalar;
+    typedef typename traits<T>::StorageIndex _StorageIndex;
+  public:
+    typedef SparseVector<_Scalar, RowMajor, _StorageIndex> type;
+};
+
+template<typename T,int Rows,int Flags> struct sparse_eval<T,Rows,1,Flags> {
+    typedef typename traits<T>::Scalar _Scalar;
+    typedef typename traits<T>::StorageIndex _StorageIndex;
+  public:
+    typedef SparseVector<_Scalar, ColMajor, _StorageIndex> type;
+};
+
+// TODO this seems almost identical to plain_matrix_type<T, Sparse>
+template<typename T,int Rows,int Cols,int Flags> struct sparse_eval {
+    typedef typename traits<T>::Scalar _Scalar;
+    typedef typename traits<T>::StorageIndex _StorageIndex;
+    enum { _Options = ((Flags&RowMajorBit)==RowMajorBit) ? RowMajor : ColMajor };
+  public:
+    typedef SparseMatrix<_Scalar, _Options, _StorageIndex> type;
+};
+
+template<typename T,int Flags> struct sparse_eval<T,1,1,Flags> {
+    typedef typename traits<T>::Scalar _Scalar;
+  public:
+    typedef Matrix<_Scalar, 1, 1> type;
+};
+
+template<typename T> struct plain_matrix_type<T,Sparse>
+{
+  typedef typename traits<T>::Scalar _Scalar;
+  typedef typename traits<T>::StorageIndex _StorageIndex;
+  enum { _Options = ((evaluator<T>::Flags&RowMajorBit)==RowMajorBit) ? RowMajor : ColMajor };
+  public:
+    typedef SparseMatrix<_Scalar, _Options, _StorageIndex> type;
+};
+
+template<typename T>
+struct plain_object_eval<T,Sparse>
+  : sparse_eval<T, traits<T>::RowsAtCompileTime,traits<T>::ColsAtCompileTime, evaluator<T>::Flags>
+{};
+
+template<typename Decomposition, typename RhsType>
+struct solve_traits<Decomposition,RhsType,Sparse>
+{
+  typedef typename sparse_eval<RhsType, RhsType::RowsAtCompileTime, RhsType::ColsAtCompileTime,traits<RhsType>::Flags>::type PlainObject;
+};
+
+template<typename Derived>
+struct generic_xpr_base<Derived, MatrixXpr, Sparse>
+{
+  typedef SparseMatrixBase<Derived> type;
+};
+
+struct SparseTriangularShape  { static std::string debugName() { return "SparseTriangularShape"; } };
+struct SparseSelfAdjointShape { static std::string debugName() { return "SparseSelfAdjointShape"; } };
+
+template<> struct glue_shapes<SparseShape,SelfAdjointShape> { typedef SparseSelfAdjointShape type;  };
+template<> struct glue_shapes<SparseShape,TriangularShape > { typedef SparseTriangularShape  type;  };
+
+} // end namespace internal
+
+/** \ingroup SparseCore_Module
+  *
+  * \class Triplet
+  *
+  * \brief A small structure to hold a non zero as a triplet (i,j,value).
+  *
+  * \sa SparseMatrix::setFromTriplets()
+  */
+template<typename Scalar, typename StorageIndex=typename SparseMatrix<Scalar>::StorageIndex >
+class Triplet
+{
+public:
+  Triplet() : m_row(0), m_col(0), m_value(0) {}
+
+  Triplet(const StorageIndex& i, const StorageIndex& j, const Scalar& v = Scalar(0))
+    : m_row(i), m_col(j), m_value(v)
+  {}
+
+  /** \returns the row index of the element */
+  const StorageIndex& row() const { return m_row; }
+
+  /** \returns the column index of the element */
+  const StorageIndex& col() const { return m_col; }
+
+  /** \returns the value of the element */
+  const Scalar& value() const { return m_value; }
+protected:
+  StorageIndex m_row, m_col;
+  Scalar m_value;
+};
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEUTIL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseVector.h
new file mode 100644
index 0000000..19b0fbc
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseVector.h
@@ -0,0 +1,478 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEVECTOR_H
+#define EIGEN_SPARSEVECTOR_H
+
+namespace Eigen { 
+
+/** \ingroup SparseCore_Module
+  * \class SparseVector
+  *
+  * \brief a sparse vector class
+  *
+  * \tparam _Scalar the scalar type, i.e. the type of the coefficients
+  *
+  * See http://www.netlib.org/linalg/html_templates/node91.html for details on the storage scheme.
+  *
+  * This class can be extended with the help of the plugin mechanism described on the page
+  * \ref TopicCustomizing_Plugins by defining the preprocessor symbol \c EIGEN_SPARSEVECTOR_PLUGIN.
+  */
+
+namespace internal {
+template<typename _Scalar, int _Options, typename _StorageIndex>
+struct traits<SparseVector<_Scalar, _Options, _StorageIndex> >
+{
+  typedef _Scalar Scalar;
+  typedef _StorageIndex StorageIndex;
+  typedef Sparse StorageKind;
+  typedef MatrixXpr XprKind;
+  enum {
+    IsColVector = (_Options & RowMajorBit) ? 0 : 1,
+
+    RowsAtCompileTime = IsColVector ? Dynamic : 1,
+    ColsAtCompileTime = IsColVector ? 1 : Dynamic,
+    MaxRowsAtCompileTime = RowsAtCompileTime,
+    MaxColsAtCompileTime = ColsAtCompileTime,
+    Flags = _Options | NestByRefBit | LvalueBit | (IsColVector ? 0 : RowMajorBit) | CompressedAccessBit,
+    SupportedAccessPatterns = InnerRandomAccessPattern
+  };
+};
+
+// Sparse-Vector-Assignment kinds:
+enum {
+  SVA_RuntimeSwitch,
+  SVA_Inner,
+  SVA_Outer
+};
+
+template< typename Dest, typename Src,
+          int AssignmentKind = !bool(Src::IsVectorAtCompileTime) ? SVA_RuntimeSwitch
+                             : Src::InnerSizeAtCompileTime==1 ? SVA_Outer
+                             : SVA_Inner>
+struct sparse_vector_assign_selector;
+
+}
+
+template<typename _Scalar, int _Options, typename _StorageIndex>
+class SparseVector
+  : public SparseCompressedBase<SparseVector<_Scalar, _Options, _StorageIndex> >
+{
+    typedef SparseCompressedBase<SparseVector> Base;
+    using Base::convert_index;
+  public:
+    EIGEN_SPARSE_PUBLIC_INTERFACE(SparseVector)
+    EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseVector, +=)
+    EIGEN_SPARSE_INHERIT_ASSIGNMENT_OPERATOR(SparseVector, -=)
+    
+    typedef internal::CompressedStorage<Scalar,StorageIndex> Storage;
+    enum { IsColVector = internal::traits<SparseVector>::IsColVector };
+    
+    enum {
+      Options = _Options
+    };
+    
+    EIGEN_STRONG_INLINE Index rows() const { return IsColVector ? m_size : 1; }
+    EIGEN_STRONG_INLINE Index cols() const { return IsColVector ? 1 : m_size; }
+    EIGEN_STRONG_INLINE Index innerSize() const { return m_size; }
+    EIGEN_STRONG_INLINE Index outerSize() const { return 1; }
+
+    EIGEN_STRONG_INLINE const Scalar* valuePtr() const { return m_data.valuePtr(); }
+    EIGEN_STRONG_INLINE Scalar* valuePtr() { return m_data.valuePtr(); }
+
+    EIGEN_STRONG_INLINE const StorageIndex* innerIndexPtr() const { return m_data.indexPtr(); }
+    EIGEN_STRONG_INLINE StorageIndex* innerIndexPtr() { return m_data.indexPtr(); }
+
+    inline const StorageIndex* outerIndexPtr() const { return 0; }
+    inline StorageIndex* outerIndexPtr() { return 0; }
+    inline const StorageIndex* innerNonZeroPtr() const { return 0; }
+    inline StorageIndex* innerNonZeroPtr() { return 0; }
+    
+    /** \internal */
+    inline Storage& data() { return m_data; }
+    /** \internal */
+    inline const Storage& data() const { return m_data; }
+
+    inline Scalar coeff(Index row, Index col) const
+    {
+      eigen_assert(IsColVector ? (col==0 && row>=0 && row<m_size) : (row==0 && col>=0 && col<m_size));
+      return coeff(IsColVector ? row : col);
+    }
+    inline Scalar coeff(Index i) const
+    {
+      eigen_assert(i>=0 && i<m_size);
+      return m_data.at(StorageIndex(i));
+    }
+
+    inline Scalar& coeffRef(Index row, Index col)
+    {
+      eigen_assert(IsColVector ? (col==0 && row>=0 && row<m_size) : (row==0 && col>=0 && col<m_size));
+      return coeffRef(IsColVector ? row : col);
+    }
+
+    /** \returns a reference to the coefficient value at given index \a i
+      * This operation involes a log(rho*size) binary search. If the coefficient does not
+      * exist yet, then a sorted insertion into a sequential buffer is performed.
+      *
+      * This insertion might be very costly if the number of nonzeros above \a i is large.
+      */
+    inline Scalar& coeffRef(Index i)
+    {
+      eigen_assert(i>=0 && i<m_size);
+
+      return m_data.atWithInsertion(StorageIndex(i));
+    }
+
+  public:
+
+    typedef typename Base::InnerIterator InnerIterator;
+    typedef typename Base::ReverseInnerIterator ReverseInnerIterator;
+
+    inline void setZero() { m_data.clear(); }
+
+    /** \returns the number of non zero coefficients */
+    inline Index nonZeros() const  { return m_data.size(); }
+
+    inline void startVec(Index outer)
+    {
+      EIGEN_UNUSED_VARIABLE(outer);
+      eigen_assert(outer==0);
+    }
+
+    inline Scalar& insertBackByOuterInner(Index outer, Index inner)
+    {
+      EIGEN_UNUSED_VARIABLE(outer);
+      eigen_assert(outer==0);
+      return insertBack(inner);
+    }
+    inline Scalar& insertBack(Index i)
+    {
+      m_data.append(0, i);
+      return m_data.value(m_data.size()-1);
+    }
+    
+    Scalar& insertBackByOuterInnerUnordered(Index outer, Index inner)
+    {
+      EIGEN_UNUSED_VARIABLE(outer);
+      eigen_assert(outer==0);
+      return insertBackUnordered(inner);
+    }
+    inline Scalar& insertBackUnordered(Index i)
+    {
+      m_data.append(0, i);
+      return m_data.value(m_data.size()-1);
+    }
+
+    inline Scalar& insert(Index row, Index col)
+    {
+      eigen_assert(IsColVector ? (col==0 && row>=0 && row<m_size) : (row==0 && col>=0 && col<m_size));
+      
+      Index inner = IsColVector ? row : col;
+      Index outer = IsColVector ? col : row;
+      EIGEN_ONLY_USED_FOR_DEBUG(outer);
+      eigen_assert(outer==0);
+      return insert(inner);
+    }
+    Scalar& insert(Index i)
+    {
+      eigen_assert(i>=0 && i<m_size);
+      
+      Index startId = 0;
+      Index p = Index(m_data.size()) - 1;
+      // TODO smart realloc
+      m_data.resize(p+2,1);
+
+      while ( (p >= startId) && (m_data.index(p) > i) )
+      {
+        m_data.index(p+1) = m_data.index(p);
+        m_data.value(p+1) = m_data.value(p);
+        --p;
+      }
+      m_data.index(p+1) = convert_index(i);
+      m_data.value(p+1) = 0;
+      return m_data.value(p+1);
+    }
+
+    /**
+      */
+    inline void reserve(Index reserveSize) { m_data.reserve(reserveSize); }
+
+
+    inline void finalize() {}
+
+    /** \copydoc SparseMatrix::prune(const Scalar&,const RealScalar&) */
+    void prune(const Scalar& reference, const RealScalar& epsilon = NumTraits<RealScalar>::dummy_precision())
+    {
+      m_data.prune(reference,epsilon);
+    }
+
+    /** Resizes the sparse vector to \a rows x \a cols
+      *
+      * This method is provided for compatibility with matrices.
+      * For a column vector, \a cols must be equal to 1.
+      * For a row vector, \a rows must be equal to 1.
+      *
+      * \sa resize(Index)
+      */
+    void resize(Index rows, Index cols)
+    {
+      eigen_assert((IsColVector ? cols : rows)==1 && "Outer dimension must equal 1");
+      resize(IsColVector ? rows : cols);
+    }
+
+    /** Resizes the sparse vector to \a newSize
+      * This method deletes all entries, thus leaving an empty sparse vector
+      *
+      * \sa  conservativeResize(), setZero() */
+    void resize(Index newSize)
+    {
+      m_size = newSize;
+      m_data.clear();
+    }
+
+    /** Resizes the sparse vector to \a newSize, while leaving old values untouched.
+      *
+      * If the size of the vector is decreased, then the storage of the out-of bounds coefficients is kept and reserved.
+      * Call .data().squeeze() to free extra memory.
+      *
+      * \sa reserve(), setZero()
+      */
+    void conservativeResize(Index newSize)
+    {
+      if (newSize < m_size)
+      {
+        Index i = 0;
+        while (i<m_data.size() && m_data.index(i)<newSize) ++i;
+        m_data.resize(i);
+      }
+      m_size = newSize;
+    }
+
+    void resizeNonZeros(Index size) { m_data.resize(size); }
+
+    inline SparseVector() : m_size(0) { check_template_parameters(); resize(0); }
+
+    explicit inline SparseVector(Index size) : m_size(0) { check_template_parameters(); resize(size); }
+
+    inline SparseVector(Index rows, Index cols) : m_size(0) { check_template_parameters(); resize(rows,cols); }
+
+    template<typename OtherDerived>
+    inline SparseVector(const SparseMatrixBase<OtherDerived>& other)
+      : m_size(0)
+    {
+      #ifdef EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+        EIGEN_SPARSE_CREATE_TEMPORARY_PLUGIN
+      #endif
+      check_template_parameters();
+      *this = other.derived();
+    }
+
+    inline SparseVector(const SparseVector& other)
+      : Base(other), m_size(0)
+    {
+      check_template_parameters();
+      *this = other.derived();
+    }
+
+    /** Swaps the values of \c *this and \a other.
+      * Overloaded for performance: this version performs a \em shallow swap by swaping pointers and attributes only.
+      * \sa SparseMatrixBase::swap()
+      */
+    inline void swap(SparseVector& other)
+    {
+      std::swap(m_size, other.m_size);
+      m_data.swap(other.m_data);
+    }
+
+    template<int OtherOptions>
+    inline void swap(SparseMatrix<Scalar,OtherOptions,StorageIndex>& other)
+    {
+      eigen_assert(other.outerSize()==1);
+      std::swap(m_size, other.m_innerSize);
+      m_data.swap(other.m_data);
+    }
+
+    inline SparseVector& operator=(const SparseVector& other)
+    {
+      if (other.isRValue())
+      {
+        swap(other.const_cast_derived());
+      }
+      else
+      {
+        resize(other.size());
+        m_data = other.m_data;
+      }
+      return *this;
+    }
+
+    template<typename OtherDerived>
+    inline SparseVector& operator=(const SparseMatrixBase<OtherDerived>& other)
+    {
+      SparseVector tmp(other.size());
+      internal::sparse_vector_assign_selector<SparseVector,OtherDerived>::run(tmp,other.derived());
+      this->swap(tmp);
+      return *this;
+    }
+
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    template<typename Lhs, typename Rhs>
+    inline SparseVector& operator=(const SparseSparseProduct<Lhs,Rhs>& product)
+    {
+      return Base::operator=(product);
+    }
+    #endif
+
+    friend std::ostream & operator << (std::ostream & s, const SparseVector& m)
+    {
+      for (Index i=0; i<m.nonZeros(); ++i)
+        s << "(" << m.m_data.value(i) << "," << m.m_data.index(i) << ") ";
+      s << std::endl;
+      return s;
+    }
+
+    /** Destructor */
+    inline ~SparseVector() {}
+
+    /** Overloaded for performance */
+    Scalar sum() const;
+
+  public:
+
+    /** \internal \deprecated use setZero() and reserve() */
+    EIGEN_DEPRECATED void startFill(Index reserve)
+    {
+      setZero();
+      m_data.reserve(reserve);
+    }
+
+    /** \internal \deprecated use insertBack(Index,Index) */
+    EIGEN_DEPRECATED Scalar& fill(Index r, Index c)
+    {
+      eigen_assert(r==0 || c==0);
+      return fill(IsColVector ? r : c);
+    }
+
+    /** \internal \deprecated use insertBack(Index) */
+    EIGEN_DEPRECATED Scalar& fill(Index i)
+    {
+      m_data.append(0, i);
+      return m_data.value(m_data.size()-1);
+    }
+
+    /** \internal \deprecated use insert(Index,Index) */
+    EIGEN_DEPRECATED Scalar& fillrand(Index r, Index c)
+    {
+      eigen_assert(r==0 || c==0);
+      return fillrand(IsColVector ? r : c);
+    }
+
+    /** \internal \deprecated use insert(Index) */
+    EIGEN_DEPRECATED Scalar& fillrand(Index i)
+    {
+      return insert(i);
+    }
+
+    /** \internal \deprecated use finalize() */
+    EIGEN_DEPRECATED void endFill() {}
+    
+    // These two functions were here in the 3.1 release, so let's keep them in case some code rely on them.
+    /** \internal \deprecated use data() */
+    EIGEN_DEPRECATED Storage& _data() { return m_data; }
+    /** \internal \deprecated use data() */
+    EIGEN_DEPRECATED const Storage& _data() const { return m_data; }
+    
+#   ifdef EIGEN_SPARSEVECTOR_PLUGIN
+#     include EIGEN_SPARSEVECTOR_PLUGIN
+#   endif
+
+protected:
+  
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT(NumTraits<StorageIndex>::IsSigned,THE_INDEX_TYPE_MUST_BE_A_SIGNED_TYPE);
+      EIGEN_STATIC_ASSERT((_Options&(ColMajor|RowMajor))==Options,INVALID_MATRIX_TEMPLATE_PARAMETERS);
+    }
+    
+    Storage m_data;
+    Index m_size;
+};
+
+namespace internal {
+
+template<typename _Scalar, int _Options, typename _Index>
+struct evaluator<SparseVector<_Scalar,_Options,_Index> >
+  : evaluator_base<SparseVector<_Scalar,_Options,_Index> >
+{
+  typedef SparseVector<_Scalar,_Options,_Index> SparseVectorType;
+  typedef evaluator_base<SparseVectorType> Base;
+  typedef typename SparseVectorType::InnerIterator InnerIterator;
+  typedef typename SparseVectorType::ReverseInnerIterator ReverseInnerIterator;
+  
+  enum {
+    CoeffReadCost = NumTraits<_Scalar>::ReadCost,
+    Flags = SparseVectorType::Flags
+  };
+
+  evaluator() : Base() {}
+  
+  explicit evaluator(const SparseVectorType &mat) : m_matrix(&mat)
+  {
+    EIGEN_INTERNAL_CHECK_COST_VALUE(CoeffReadCost);
+  }
+  
+  inline Index nonZerosEstimate() const {
+    return m_matrix->nonZeros();
+  }
+  
+  operator SparseVectorType&() { return m_matrix->const_cast_derived(); }
+  operator const SparseVectorType&() const { return *m_matrix; }
+  
+  const SparseVectorType *m_matrix;
+};
+
+template< typename Dest, typename Src>
+struct sparse_vector_assign_selector<Dest,Src,SVA_Inner> {
+  static void run(Dest& dst, const Src& src) {
+    eigen_internal_assert(src.innerSize()==src.size());
+    typedef internal::evaluator<Src> SrcEvaluatorType;
+    SrcEvaluatorType srcEval(src);
+    for(typename SrcEvaluatorType::InnerIterator it(srcEval, 0); it; ++it)
+      dst.insert(it.index()) = it.value();
+  }
+};
+
+template< typename Dest, typename Src>
+struct sparse_vector_assign_selector<Dest,Src,SVA_Outer> {
+  static void run(Dest& dst, const Src& src) {
+    eigen_internal_assert(src.outerSize()==src.size());
+    typedef internal::evaluator<Src> SrcEvaluatorType;
+    SrcEvaluatorType srcEval(src);
+    for(Index i=0; i<src.size(); ++i)
+    {
+      typename SrcEvaluatorType::InnerIterator it(srcEval, i);
+      if(it)
+        dst.insert(i) = it.value();
+    }
+  }
+};
+
+template< typename Dest, typename Src>
+struct sparse_vector_assign_selector<Dest,Src,SVA_RuntimeSwitch> {
+  static void run(Dest& dst, const Src& src) {
+    if(src.outerSize()==1)  sparse_vector_assign_selector<Dest,Src,SVA_Inner>::run(dst, src);
+    else                    sparse_vector_assign_selector<Dest,Src,SVA_Outer>::run(dst, src);
+  }
+};
+
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSEVECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseView.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseView.h
new file mode 100644
index 0000000..7c4aea7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/SparseView.h
@@ -0,0 +1,253 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2011-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2010 Daniel Lowengrub <lowdanie@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSEVIEW_H
+#define EIGEN_SPARSEVIEW_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename MatrixType>
+struct traits<SparseView<MatrixType> > : traits<MatrixType>
+{
+  typedef typename MatrixType::StorageIndex StorageIndex;
+  typedef Sparse StorageKind;
+  enum {
+    Flags = int(traits<MatrixType>::Flags) & (RowMajorBit)
+  };
+};
+
+} // end namespace internal
+
+/** \ingroup SparseCore_Module
+  * \class SparseView
+  *
+  * \brief Expression of a dense or sparse matrix with zero or too small values removed
+  *
+  * \tparam MatrixType the type of the object of which we are removing the small entries
+  *
+  * This class represents an expression of a given dense or sparse matrix with
+  * entries smaller than \c reference * \c epsilon are removed.
+  * It is the return type of MatrixBase::sparseView() and SparseMatrixBase::pruned()
+  * and most of the time this is the only way it is used.
+  *
+  * \sa MatrixBase::sparseView(), SparseMatrixBase::pruned()
+  */
+template<typename MatrixType>
+class SparseView : public SparseMatrixBase<SparseView<MatrixType> >
+{
+  typedef typename MatrixType::Nested MatrixTypeNested;
+  typedef typename internal::remove_all<MatrixTypeNested>::type _MatrixTypeNested;
+  typedef SparseMatrixBase<SparseView > Base;
+public:
+  EIGEN_SPARSE_PUBLIC_INTERFACE(SparseView)
+  typedef typename internal::remove_all<MatrixType>::type NestedExpression;
+
+  explicit SparseView(const MatrixType& mat, const Scalar& reference = Scalar(0),
+                      const RealScalar &epsilon = NumTraits<Scalar>::dummy_precision())
+    : m_matrix(mat), m_reference(reference), m_epsilon(epsilon) {}
+
+  inline Index rows() const { return m_matrix.rows(); }
+  inline Index cols() const { return m_matrix.cols(); }
+
+  inline Index innerSize() const { return m_matrix.innerSize(); }
+  inline Index outerSize() const { return m_matrix.outerSize(); }
+  
+  /** \returns the nested expression */
+  const typename internal::remove_all<MatrixTypeNested>::type&
+  nestedExpression() const { return m_matrix; }
+  
+  Scalar reference() const { return m_reference; }
+  RealScalar epsilon() const { return m_epsilon; }
+  
+protected:
+  MatrixTypeNested m_matrix;
+  Scalar m_reference;
+  RealScalar m_epsilon;
+};
+
+namespace internal {
+
+// TODO find a way to unify the two following variants
+// This is tricky because implementing an inner iterator on top of an IndexBased evaluator is
+// not easy because the evaluators do not expose the sizes of the underlying expression.
+  
+template<typename ArgType>
+struct unary_evaluator<SparseView<ArgType>, IteratorBased>
+  : public evaluator_base<SparseView<ArgType> >
+{
+    typedef typename evaluator<ArgType>::InnerIterator EvalIterator;
+  public:
+    typedef SparseView<ArgType> XprType;
+    
+    class InnerIterator : public EvalIterator
+    {
+        typedef typename XprType::Scalar Scalar;
+      public:
+
+        EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, Index outer)
+          : EvalIterator(sve.m_argImpl,outer), m_view(sve.m_view)
+        {
+          incrementToNonZero();
+        }
+
+        EIGEN_STRONG_INLINE InnerIterator& operator++()
+        {
+          EvalIterator::operator++();
+          incrementToNonZero();
+          return *this;
+        }
+
+        using EvalIterator::value;
+
+      protected:
+        const XprType &m_view;
+
+      private:
+        void incrementToNonZero()
+        {
+          while((bool(*this)) && internal::isMuchSmallerThan(value(), m_view.reference(), m_view.epsilon()))
+          {
+            EvalIterator::operator++();
+          }
+        }
+    };
+    
+    enum {
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+      Flags = XprType::Flags
+    };
+    
+    explicit unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {}
+
+  protected:
+    evaluator<ArgType> m_argImpl;
+    const XprType &m_view;
+};
+
+template<typename ArgType>
+struct unary_evaluator<SparseView<ArgType>, IndexBased>
+  : public evaluator_base<SparseView<ArgType> >
+{
+  public:
+    typedef SparseView<ArgType> XprType;
+  protected:
+    enum { IsRowMajor = (XprType::Flags&RowMajorBit)==RowMajorBit };
+    typedef typename XprType::Scalar Scalar;
+    typedef typename XprType::StorageIndex StorageIndex;
+  public:
+    
+    class InnerIterator
+    {
+      public:
+
+        EIGEN_STRONG_INLINE InnerIterator(const unary_evaluator& sve, Index outer)
+          : m_sve(sve), m_inner(0), m_outer(outer), m_end(sve.m_view.innerSize())
+        {
+          incrementToNonZero();
+        }
+
+        EIGEN_STRONG_INLINE InnerIterator& operator++()
+        {
+          m_inner++;
+          incrementToNonZero();
+          return *this;
+        }
+
+        EIGEN_STRONG_INLINE Scalar value() const
+        {
+          return (IsRowMajor) ? m_sve.m_argImpl.coeff(m_outer, m_inner)
+                              : m_sve.m_argImpl.coeff(m_inner, m_outer);
+        }
+
+        EIGEN_STRONG_INLINE StorageIndex index() const { return m_inner; }
+        inline Index row() const { return IsRowMajor ? m_outer : index(); }
+        inline Index col() const { return IsRowMajor ? index() : m_outer; }
+
+        EIGEN_STRONG_INLINE operator bool() const { return m_inner < m_end && m_inner>=0; }
+
+      protected:
+        const unary_evaluator &m_sve;
+        Index m_inner;
+        const Index m_outer;
+        const Index m_end;
+
+      private:
+        void incrementToNonZero()
+        {
+          while((bool(*this)) && internal::isMuchSmallerThan(value(), m_sve.m_view.reference(), m_sve.m_view.epsilon()))
+          {
+            m_inner++;
+          }
+        }
+    };
+    
+    enum {
+      CoeffReadCost = evaluator<ArgType>::CoeffReadCost,
+      Flags = XprType::Flags
+    };
+    
+    explicit unary_evaluator(const XprType& xpr) : m_argImpl(xpr.nestedExpression()), m_view(xpr) {}
+
+  protected:
+    evaluator<ArgType> m_argImpl;
+    const XprType &m_view;
+};
+
+} // end namespace internal
+
+/** \ingroup SparseCore_Module
+  *
+  * \returns a sparse expression of the dense expression \c *this with values smaller than
+  * \a reference * \a epsilon removed.
+  *
+  * This method is typically used when prototyping to convert a quickly assembled dense Matrix \c D to a SparseMatrix \c S:
+  * \code
+  * MatrixXd D(n,m);
+  * SparseMatrix<double> S;
+  * S = D.sparseView();             // suppress numerical zeros (exact)
+  * S = D.sparseView(reference);
+  * S = D.sparseView(reference,epsilon);
+  * \endcode
+  * where \a reference is a meaningful non zero reference value,
+  * and \a epsilon is a tolerance factor defaulting to NumTraits<Scalar>::dummy_precision().
+  *
+  * \sa SparseMatrixBase::pruned(), class SparseView */
+template<typename Derived>
+const SparseView<Derived> MatrixBase<Derived>::sparseView(const Scalar& reference,
+                                                          const typename NumTraits<Scalar>::Real& epsilon) const
+{
+  return SparseView<Derived>(derived(), reference, epsilon);
+}
+
+/** \returns an expression of \c *this with values smaller than
+  * \a reference * \a epsilon removed.
+  *
+  * This method is typically used in conjunction with the product of two sparse matrices
+  * to automatically prune the smallest values as follows:
+  * \code
+  * C = (A*B).pruned();             // suppress numerical zeros (exact)
+  * C = (A*B).pruned(ref);
+  * C = (A*B).pruned(ref,epsilon);
+  * \endcode
+  * where \c ref is a meaningful non zero reference value.
+  * */
+template<typename Derived>
+const SparseView<Derived>
+SparseMatrixBase<Derived>::pruned(const Scalar& reference,
+                                  const RealScalar& epsilon) const
+{
+  return SparseView<Derived>(derived(), reference, epsilon);
+}
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/TriangularSolver.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/TriangularSolver.h
new file mode 100644
index 0000000..f9c56ba
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseCore/TriangularSolver.h
@@ -0,0 +1,315 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSETRIANGULARSOLVER_H
+#define EIGEN_SPARSETRIANGULARSOLVER_H
+
+namespace Eigen { 
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, int Mode,
+  int UpLo = (Mode & Lower)
+           ? Lower
+           : (Mode & Upper)
+           ? Upper
+           : -1,
+  int StorageOrder = int(traits<Lhs>::Flags) & RowMajorBit>
+struct sparse_solve_triangular_selector;
+
+// forward substitution, row-major
+template<typename Lhs, typename Rhs, int Mode>
+struct sparse_solve_triangular_selector<Lhs,Rhs,Mode,Lower,RowMajor>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename evaluator<Lhs>::InnerIterator LhsIterator;
+  static void run(const Lhs& lhs, Rhs& other)
+  {
+    LhsEval lhsEval(lhs);
+    for(Index col=0 ; col<other.cols() ; ++col)
+    {
+      for(Index i=0; i<lhs.rows(); ++i)
+      {
+        Scalar tmp = other.coeff(i,col);
+        Scalar lastVal(0);
+        Index lastIndex = 0;
+        for(LhsIterator it(lhsEval, i); it; ++it)
+        {
+          lastVal = it.value();
+          lastIndex = it.index();
+          if(lastIndex==i)
+            break;
+          tmp -= lastVal * other.coeff(lastIndex,col);
+        }
+        if (Mode & UnitDiag)
+          other.coeffRef(i,col) = tmp;
+        else
+        {
+          eigen_assert(lastIndex==i);
+          other.coeffRef(i,col) = tmp/lastVal;
+        }
+      }
+    }
+  }
+};
+
+// backward substitution, row-major
+template<typename Lhs, typename Rhs, int Mode>
+struct sparse_solve_triangular_selector<Lhs,Rhs,Mode,Upper,RowMajor>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename evaluator<Lhs>::InnerIterator LhsIterator;
+  static void run(const Lhs& lhs, Rhs& other)
+  {
+    LhsEval lhsEval(lhs);
+    for(Index col=0 ; col<other.cols() ; ++col)
+    {
+      for(Index i=lhs.rows()-1 ; i>=0 ; --i)
+      {
+        Scalar tmp = other.coeff(i,col);
+        Scalar l_ii(0);
+        LhsIterator it(lhsEval, i);
+        while(it && it.index()<i)
+          ++it;
+        if(!(Mode & UnitDiag))
+        {
+          eigen_assert(it && it.index()==i);
+          l_ii = it.value();
+          ++it;
+        }
+        else if (it && it.index() == i)
+          ++it;
+        for(; it; ++it)
+        {
+          tmp -= it.value() * other.coeff(it.index(),col);
+        }
+
+        if (Mode & UnitDiag)  other.coeffRef(i,col) = tmp;
+        else                  other.coeffRef(i,col) = tmp/l_ii;
+      }
+    }
+  }
+};
+
+// forward substitution, col-major
+template<typename Lhs, typename Rhs, int Mode>
+struct sparse_solve_triangular_selector<Lhs,Rhs,Mode,Lower,ColMajor>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename evaluator<Lhs>::InnerIterator LhsIterator;
+  static void run(const Lhs& lhs, Rhs& other)
+  {
+    LhsEval lhsEval(lhs);
+    for(Index col=0 ; col<other.cols() ; ++col)
+    {
+      for(Index i=0; i<lhs.cols(); ++i)
+      {
+        Scalar& tmp = other.coeffRef(i,col);
+        if (tmp!=Scalar(0)) // optimization when other is actually sparse
+        {
+          LhsIterator it(lhsEval, i);
+          while(it && it.index()<i)
+            ++it;
+          if(!(Mode & UnitDiag))
+          {
+            eigen_assert(it && it.index()==i);
+            tmp /= it.value();
+          }
+          if (it && it.index()==i)
+            ++it;
+          for(; it; ++it)
+            other.coeffRef(it.index(), col) -= tmp * it.value();
+        }
+      }
+    }
+  }
+};
+
+// backward substitution, col-major
+template<typename Lhs, typename Rhs, int Mode>
+struct sparse_solve_triangular_selector<Lhs,Rhs,Mode,Upper,ColMajor>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef evaluator<Lhs> LhsEval;
+  typedef typename evaluator<Lhs>::InnerIterator LhsIterator;
+  static void run(const Lhs& lhs, Rhs& other)
+  {
+    LhsEval lhsEval(lhs);
+    for(Index col=0 ; col<other.cols() ; ++col)
+    {
+      for(Index i=lhs.cols()-1; i>=0; --i)
+      {
+        Scalar& tmp = other.coeffRef(i,col);
+        if (tmp!=Scalar(0)) // optimization when other is actually sparse
+        {
+          if(!(Mode & UnitDiag))
+          {
+            // TODO replace this by a binary search. make sure the binary search is safe for partially sorted elements
+            LhsIterator it(lhsEval, i);
+            while(it && it.index()!=i)
+              ++it;
+            eigen_assert(it && it.index()==i);
+            other.coeffRef(i,col) /= it.value();
+          }
+          LhsIterator it(lhsEval, i);
+          for(; it && it.index()<i; ++it)
+            other.coeffRef(it.index(), col) -= tmp * it.value();
+        }
+      }
+    }
+  }
+};
+
+} // end namespace internal
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+template<typename ExpressionType,unsigned int Mode>
+template<typename OtherDerived>
+void TriangularViewImpl<ExpressionType,Mode,Sparse>::solveInPlace(MatrixBase<OtherDerived>& other) const
+{
+  eigen_assert(derived().cols() == derived().rows() && derived().cols() == other.rows());
+  eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
+
+  enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit };
+
+  typedef typename internal::conditional<copy,
+    typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy;
+  OtherCopy otherCopy(other.derived());
+
+  internal::sparse_solve_triangular_selector<ExpressionType, typename internal::remove_reference<OtherCopy>::type, Mode>::run(derived().nestedExpression(), otherCopy);
+
+  if (copy)
+    other = otherCopy;
+}
+#endif
+
+// pure sparse path
+
+namespace internal {
+
+template<typename Lhs, typename Rhs, int Mode,
+  int UpLo = (Mode & Lower)
+           ? Lower
+           : (Mode & Upper)
+           ? Upper
+           : -1,
+  int StorageOrder = int(Lhs::Flags) & (RowMajorBit)>
+struct sparse_solve_triangular_sparse_selector;
+
+// forward substitution, col-major
+template<typename Lhs, typename Rhs, int Mode, int UpLo>
+struct sparse_solve_triangular_sparse_selector<Lhs,Rhs,Mode,UpLo,ColMajor>
+{
+  typedef typename Rhs::Scalar Scalar;
+  typedef typename promote_index_type<typename traits<Lhs>::StorageIndex,
+                                      typename traits<Rhs>::StorageIndex>::type StorageIndex;
+  static void run(const Lhs& lhs, Rhs& other)
+  {
+    const bool IsLower = (UpLo==Lower);
+    AmbiVector<Scalar,StorageIndex> tempVector(other.rows()*2);
+    tempVector.setBounds(0,other.rows());
+
+    Rhs res(other.rows(), other.cols());
+    res.reserve(other.nonZeros());
+
+    for(Index col=0 ; col<other.cols() ; ++col)
+    {
+      // FIXME estimate number of non zeros
+      tempVector.init(.99/*float(other.col(col).nonZeros())/float(other.rows())*/);
+      tempVector.setZero();
+      tempVector.restart();
+      for (typename Rhs::InnerIterator rhsIt(other, col); rhsIt; ++rhsIt)
+      {
+        tempVector.coeffRef(rhsIt.index()) = rhsIt.value();
+      }
+
+      for(Index i=IsLower?0:lhs.cols()-1;
+          IsLower?i<lhs.cols():i>=0;
+          i+=IsLower?1:-1)
+      {
+        tempVector.restart();
+        Scalar& ci = tempVector.coeffRef(i);
+        if (ci!=Scalar(0))
+        {
+          // find
+          typename Lhs::InnerIterator it(lhs, i);
+          if(!(Mode & UnitDiag))
+          {
+            if (IsLower)
+            {
+              eigen_assert(it.index()==i);
+              ci /= it.value();
+            }
+            else
+              ci /= lhs.coeff(i,i);
+          }
+          tempVector.restart();
+          if (IsLower)
+          {
+            if (it.index()==i)
+              ++it;
+            for(; it; ++it)
+              tempVector.coeffRef(it.index()) -= ci * it.value();
+          }
+          else
+          {
+            for(; it && it.index()<i; ++it)
+              tempVector.coeffRef(it.index()) -= ci * it.value();
+          }
+        }
+      }
+
+
+      Index count = 0;
+      // FIXME compute a reference value to filter zeros
+      for (typename AmbiVector<Scalar,StorageIndex>::Iterator it(tempVector/*,1e-12*/); it; ++it)
+      {
+        ++ count;
+//         std::cerr << "fill " << it.index() << ", " << col << "\n";
+//         std::cout << it.value() << "  ";
+        // FIXME use insertBack
+        res.insert(it.index(), col) = it.value();
+      }
+//       std::cout << "tempVector.nonZeros() == " << int(count) << " / " << (other.rows()) << "\n";
+    }
+    res.finalize();
+    other = res.markAsRValue();
+  }
+};
+
+} // end namespace internal
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename ExpressionType,unsigned int Mode>
+template<typename OtherDerived>
+void TriangularViewImpl<ExpressionType,Mode,Sparse>::solveInPlace(SparseMatrixBase<OtherDerived>& other) const
+{
+  eigen_assert(derived().cols() == derived().rows() && derived().cols() == other.rows());
+  eigen_assert( (!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
+
+//   enum { copy = internal::traits<OtherDerived>::Flags & RowMajorBit };
+
+//   typedef typename internal::conditional<copy,
+//     typename internal::plain_matrix_type_column_major<OtherDerived>::type, OtherDerived&>::type OtherCopy;
+//   OtherCopy otherCopy(other.derived());
+
+  internal::sparse_solve_triangular_sparse_selector<ExpressionType, OtherDerived, Mode>::run(derived().nestedExpression(), other.derived());
+
+//   if (copy)
+//     other = otherCopy;
+}
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSETRIANGULARSOLVER_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU.h
new file mode 100644
index 0000000..f883ab3
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU.h
@@ -0,0 +1,775 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_SPARSE_LU_H
+#define EIGEN_SPARSE_LU_H
+
+namespace Eigen {
+
+template <typename _MatrixType, typename _OrderingType = COLAMDOrdering<typename _MatrixType::StorageIndex> > class SparseLU;
+template <typename MappedSparseMatrixType> struct SparseLUMatrixLReturnType;
+template <typename MatrixLType, typename MatrixUType> struct SparseLUMatrixUReturnType;
+
+/** \ingroup SparseLU_Module
+  * \class SparseLU
+  * 
+  * \brief Sparse supernodal LU factorization for general matrices
+  * 
+  * This class implements the supernodal LU factorization for general matrices.
+  * It uses the main techniques from the sequential SuperLU package 
+  * (http://crd-legacy.lbl.gov/~xiaoye/SuperLU/). It handles transparently real 
+  * and complex arithmetics with single and double precision, depending on the 
+  * scalar type of your input matrix. 
+  * The code has been optimized to provide BLAS-3 operations during supernode-panel updates. 
+  * It benefits directly from the built-in high-performant Eigen BLAS routines. 
+  * Moreover, when the size of a supernode is very small, the BLAS calls are avoided to 
+  * enable a better optimization from the compiler. For best performance, 
+  * you should compile it with NDEBUG flag to avoid the numerous bounds checking on vectors. 
+  * 
+  * An important parameter of this class is the ordering method. It is used to reorder the columns 
+  * (and eventually the rows) of the matrix to reduce the number of new elements that are created during 
+  * numerical factorization. The cheapest method available is COLAMD. 
+  * See  \link OrderingMethods_Module the OrderingMethods module \endlink for the list of 
+  * built-in and external ordering methods. 
+  *
+  * Simple example with key steps 
+  * \code
+  * VectorXd x(n), b(n);
+  * SparseMatrix<double, ColMajor> A;
+  * SparseLU<SparseMatrix<scalar, ColMajor>, COLAMDOrdering<Index> >   solver;
+  * // fill A and b;
+  * // Compute the ordering permutation vector from the structural pattern of A
+  * solver.analyzePattern(A); 
+  * // Compute the numerical factorization 
+  * solver.factorize(A); 
+  * //Use the factors to solve the linear system 
+  * x = solver.solve(b); 
+  * \endcode
+  * 
+  * \warning The input matrix A should be in a \b compressed and \b column-major form.
+  * Otherwise an expensive copy will be made. You can call the inexpensive makeCompressed() to get a compressed matrix.
+  * 
+  * \note Unlike the initial SuperLU implementation, there is no step to equilibrate the matrix. 
+  * For badly scaled matrices, this step can be useful to reduce the pivoting during factorization. 
+  * If this is the case for your matrices, you can try the basic scaling method at
+  *  "unsupported/Eigen/src/IterativeSolvers/Scaling.h"
+  * 
+  * \tparam _MatrixType The type of the sparse matrix. It must be a column-major SparseMatrix<>
+  * \tparam _OrderingType The ordering method to use, either AMD, COLAMD or METIS. Default is COLMAD
+  *
+  * \implsparsesolverconcept
+  * 
+  * \sa \ref TutorialSparseSolverConcept
+  * \sa \ref OrderingMethods_Module
+  */
+template <typename _MatrixType, typename _OrderingType>
+class SparseLU : public SparseSolverBase<SparseLU<_MatrixType,_OrderingType> >, public internal::SparseLUImpl<typename _MatrixType::Scalar, typename _MatrixType::StorageIndex>
+{
+  protected:
+    typedef SparseSolverBase<SparseLU<_MatrixType,_OrderingType> > APIBase;
+    using APIBase::m_isInitialized;
+  public:
+    using APIBase::_solve_impl;
+    
+    typedef _MatrixType MatrixType; 
+    typedef _OrderingType OrderingType;
+    typedef typename MatrixType::Scalar Scalar; 
+    typedef typename MatrixType::RealScalar RealScalar; 
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> NCMatrix;
+    typedef internal::MappedSuperNodalMatrix<Scalar, StorageIndex> SCMatrix;
+    typedef Matrix<Scalar,Dynamic,1> ScalarVector;
+    typedef Matrix<StorageIndex,Dynamic,1> IndexVector;
+    typedef PermutationMatrix<Dynamic, Dynamic, StorageIndex> PermutationType;
+    typedef internal::SparseLUImpl<Scalar, StorageIndex> Base;
+
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    
+  public:
+    SparseLU():m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1)
+    {
+      initperfvalues(); 
+    }
+    explicit SparseLU(const MatrixType& matrix)
+      : m_lastError(""),m_Ustore(0,0,0,0,0,0),m_symmetricmode(false),m_diagpivotthresh(1.0),m_detPermR(1)
+    {
+      initperfvalues(); 
+      compute(matrix);
+    }
+    
+    ~SparseLU()
+    {
+      // Free all explicit dynamic pointers 
+    }
+    
+    void analyzePattern (const MatrixType& matrix);
+    void factorize (const MatrixType& matrix);
+    void simplicialfactorize(const MatrixType& matrix);
+    
+    /**
+      * Compute the symbolic and numeric factorization of the input sparse matrix.
+      * The input matrix should be in column-major storage. 
+      */
+    void compute (const MatrixType& matrix)
+    {
+      // Analyze 
+      analyzePattern(matrix); 
+      //Factorize
+      factorize(matrix);
+    } 
+    
+    inline Index rows() const { return m_mat.rows(); }
+    inline Index cols() const { return m_mat.cols(); }
+    /** Indicate that the pattern of the input matrix is symmetric */
+    void isSymmetric(bool sym)
+    {
+      m_symmetricmode = sym;
+    }
+    
+    /** \returns an expression of the matrix L, internally stored as supernodes
+      * The only operation available with this expression is the triangular solve
+      * \code
+      * y = b; matrixL().solveInPlace(y);
+      * \endcode
+      */
+    SparseLUMatrixLReturnType<SCMatrix> matrixL() const
+    {
+      return SparseLUMatrixLReturnType<SCMatrix>(m_Lstore);
+    }
+    /** \returns an expression of the matrix U,
+      * The only operation available with this expression is the triangular solve
+      * \code
+      * y = b; matrixU().solveInPlace(y);
+      * \endcode
+      */
+    SparseLUMatrixUReturnType<SCMatrix,MappedSparseMatrix<Scalar,ColMajor,StorageIndex> > matrixU() const
+    {
+      return SparseLUMatrixUReturnType<SCMatrix, MappedSparseMatrix<Scalar,ColMajor,StorageIndex> >(m_Lstore, m_Ustore);
+    }
+
+    /**
+      * \returns a reference to the row matrix permutation \f$ P_r \f$ such that \f$P_r A P_c^T = L U\f$
+      * \sa colsPermutation()
+      */
+    inline const PermutationType& rowsPermutation() const
+    {
+      return m_perm_r;
+    }
+    /**
+      * \returns a reference to the column matrix permutation\f$ P_c^T \f$ such that \f$P_r A P_c^T = L U\f$
+      * \sa rowsPermutation()
+      */
+    inline const PermutationType& colsPermutation() const
+    {
+      return m_perm_c;
+    }
+    /** Set the threshold used for a diagonal entry to be an acceptable pivot. */
+    void setPivotThreshold(const RealScalar& thresh)
+    {
+      m_diagpivotthresh = thresh; 
+    }
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+    /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A.
+      *
+      * \warning the destination matrix X in X = this->solve(B) must be colmun-major.
+      *
+      * \sa compute()
+      */
+    template<typename Rhs>
+    inline const Solve<SparseLU, Rhs> solve(const MatrixBase<Rhs>& B) const;
+#endif // EIGEN_PARSED_BY_DOXYGEN
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the LU factorization reports a problem, zero diagonal for instance
+      *          \c InvalidInput if the input matrix is invalid
+      *
+      * \sa iparm()          
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+    
+    /**
+      * \returns A string describing the type of error
+      */
+    std::string lastErrorMessage() const
+    {
+      return m_lastError; 
+    }
+
+    template<typename Rhs, typename Dest>
+    bool _solve_impl(const MatrixBase<Rhs> &B, MatrixBase<Dest> &X_base) const
+    {
+      Dest& X(X_base.derived());
+      eigen_assert(m_factorizationIsOk && "The matrix should be factorized first");
+      EIGEN_STATIC_ASSERT((Dest::Flags&RowMajorBit)==0,
+                        THIS_METHOD_IS_ONLY_FOR_COLUMN_MAJOR_MATRICES);
+      
+      // Permute the right hand side to form X = Pr*B
+      // on return, X is overwritten by the computed solution
+      X.resize(B.rows(),B.cols());
+
+      // this ugly const_cast_derived() helps to detect aliasing when applying the permutations
+      for(Index j = 0; j < B.cols(); ++j)
+        X.col(j) = rowsPermutation() * B.const_cast_derived().col(j);
+      
+      //Forward substitution with L
+      this->matrixL().solveInPlace(X);
+      this->matrixU().solveInPlace(X);
+      
+      // Permute back the solution 
+      for (Index j = 0; j < B.cols(); ++j)
+        X.col(j) = colsPermutation().inverse() * X.col(j);
+      
+      return true; 
+    }
+    
+    /**
+      * \returns the absolute value of the determinant of the matrix of which
+      * *this is the QR decomposition.
+      *
+      * \warning a determinant can be very big or small, so for matrices
+      * of large enough dimension, there is a risk of overflow/underflow.
+      * One way to work around that is to use logAbsDeterminant() instead.
+      *
+      * \sa logAbsDeterminant(), signDeterminant()
+      */
+    Scalar absDeterminant()
+    {
+      using std::abs;
+      eigen_assert(m_factorizationIsOk && "The matrix should be factorized first.");
+      // Initialize with the determinant of the row matrix
+      Scalar det = Scalar(1.);
+      // Note that the diagonal blocks of U are stored in supernodes,
+      // which are available in the  L part :)
+      for (Index j = 0; j < this->cols(); ++j)
+      {
+        for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it)
+        {
+          if(it.index() == j)
+          {
+            det *= abs(it.value());
+            break;
+          }
+        }
+      }
+      return det;
+    }
+
+    /** \returns the natural log of the absolute value of the determinant of the matrix
+      * of which **this is the QR decomposition
+      *
+      * \note This method is useful to work around the risk of overflow/underflow that's
+      * inherent to the determinant computation.
+      *
+      * \sa absDeterminant(), signDeterminant()
+      */
+    Scalar logAbsDeterminant() const
+    {
+      using std::log;
+      using std::abs;
+
+      eigen_assert(m_factorizationIsOk && "The matrix should be factorized first.");
+      Scalar det = Scalar(0.);
+      for (Index j = 0; j < this->cols(); ++j)
+      {
+        for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it)
+        {
+          if(it.row() < j) continue;
+          if(it.row() == j)
+          {
+            det += log(abs(it.value()));
+            break;
+          }
+        }
+      }
+      return det;
+    }
+
+    /** \returns A number representing the sign of the determinant
+      *
+      * \sa absDeterminant(), logAbsDeterminant()
+      */
+    Scalar signDeterminant()
+    {
+      eigen_assert(m_factorizationIsOk && "The matrix should be factorized first.");
+      // Initialize with the determinant of the row matrix
+      Index det = 1;
+      // Note that the diagonal blocks of U are stored in supernodes,
+      // which are available in the  L part :)
+      for (Index j = 0; j < this->cols(); ++j)
+      {
+        for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it)
+        {
+          if(it.index() == j)
+          {
+            if(it.value()<0)
+              det = -det;
+            else if(it.value()==0)
+              return 0;
+            break;
+          }
+        }
+      }
+      return det * m_detPermR * m_detPermC;
+    }
+    
+    /** \returns The determinant of the matrix.
+      *
+      * \sa absDeterminant(), logAbsDeterminant()
+      */
+    Scalar determinant()
+    {
+      eigen_assert(m_factorizationIsOk && "The matrix should be factorized first.");
+      // Initialize with the determinant of the row matrix
+      Scalar det = Scalar(1.);
+      // Note that the diagonal blocks of U are stored in supernodes,
+      // which are available in the  L part :)
+      for (Index j = 0; j < this->cols(); ++j)
+      {
+        for (typename SCMatrix::InnerIterator it(m_Lstore, j); it; ++it)
+        {
+          if(it.index() == j)
+          {
+            det *= it.value();
+            break;
+          }
+        }
+      }
+      return (m_detPermR * m_detPermC) > 0 ? det : -det;
+    }
+
+  protected:
+    // Functions 
+    void initperfvalues()
+    {
+      m_perfv.panel_size = 16;
+      m_perfv.relax = 1; 
+      m_perfv.maxsuper = 128; 
+      m_perfv.rowblk = 16; 
+      m_perfv.colblk = 8; 
+      m_perfv.fillfactor = 20;  
+    }
+      
+    // Variables 
+    mutable ComputationInfo m_info;
+    bool m_factorizationIsOk;
+    bool m_analysisIsOk;
+    std::string m_lastError;
+    NCMatrix m_mat; // The input (permuted ) matrix 
+    SCMatrix m_Lstore; // The lower triangular matrix (supernodal)
+    MappedSparseMatrix<Scalar,ColMajor,StorageIndex> m_Ustore; // The upper triangular matrix
+    PermutationType m_perm_c; // Column permutation 
+    PermutationType m_perm_r ; // Row permutation
+    IndexVector m_etree; // Column elimination tree 
+    
+    typename Base::GlobalLU_t m_glu; 
+                               
+    // SparseLU options 
+    bool m_symmetricmode;
+    // values for performance 
+    internal::perfvalues m_perfv;
+    RealScalar m_diagpivotthresh; // Specifies the threshold used for a diagonal entry to be an acceptable pivot
+    Index m_nnzL, m_nnzU; // Nonzeros in L and U factors
+    Index m_detPermR, m_detPermC; // Determinants of the permutation matrices
+  private:
+    // Disable copy constructor 
+    SparseLU (const SparseLU& );
+  
+}; // End class SparseLU
+
+
+
+// Functions needed by the anaysis phase
+/** 
+  * Compute the column permutation to minimize the fill-in
+  * 
+  *  - Apply this permutation to the input matrix - 
+  * 
+  *  - Compute the column elimination tree on the permuted matrix 
+  * 
+  *  - Postorder the elimination tree and the column permutation
+  * 
+  */
+template <typename MatrixType, typename OrderingType>
+void SparseLU<MatrixType, OrderingType>::analyzePattern(const MatrixType& mat)
+{
+  
+  //TODO  It is possible as in SuperLU to compute row and columns scaling vectors to equilibrate the matrix mat.
+  
+  // Firstly, copy the whole input matrix. 
+  m_mat = mat;
+  
+  // Compute fill-in ordering
+  OrderingType ord; 
+  ord(m_mat,m_perm_c);
+  
+  // Apply the permutation to the column of the input  matrix
+  if (m_perm_c.size())
+  {
+    m_mat.uncompress(); //NOTE: The effect of this command is only to create the InnerNonzeros pointers. FIXME : This vector is filled but not subsequently used.  
+    // Then, permute only the column pointers
+    ei_declare_aligned_stack_constructed_variable(StorageIndex,outerIndexPtr,mat.cols()+1,mat.isCompressed()?const_cast<StorageIndex*>(mat.outerIndexPtr()):0);
+    
+    // If the input matrix 'mat' is uncompressed, then the outer-indices do not match the ones of m_mat, and a copy is thus needed.
+    if(!mat.isCompressed()) 
+      IndexVector::Map(outerIndexPtr, mat.cols()+1) = IndexVector::Map(m_mat.outerIndexPtr(),mat.cols()+1);
+    
+    // Apply the permutation and compute the nnz per column.
+    for (Index i = 0; i < mat.cols(); i++)
+    {
+      m_mat.outerIndexPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i];
+      m_mat.innerNonZeroPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i+1] - outerIndexPtr[i];
+    }
+  }
+  
+  // Compute the column elimination tree of the permuted matrix 
+  IndexVector firstRowElt;
+  internal::coletree(m_mat, m_etree,firstRowElt); 
+     
+  // In symmetric mode, do not do postorder here
+  if (!m_symmetricmode) {
+    IndexVector post, iwork; 
+    // Post order etree
+    internal::treePostorder(StorageIndex(m_mat.cols()), m_etree, post); 
+      
+   
+    // Renumber etree in postorder 
+    Index m = m_mat.cols(); 
+    iwork.resize(m+1);
+    for (Index i = 0; i < m; ++i) iwork(post(i)) = post(m_etree(i));
+    m_etree = iwork;
+    
+    // Postmultiply A*Pc by post, i.e reorder the matrix according to the postorder of the etree
+    PermutationType post_perm(m); 
+    for (Index i = 0; i < m; i++) 
+      post_perm.indices()(i) = post(i); 
+        
+    // Combine the two permutations : postorder the permutation for future use
+    if(m_perm_c.size()) {
+      m_perm_c = post_perm * m_perm_c;
+    }
+    
+  } // end postordering 
+  
+  m_analysisIsOk = true; 
+}
+
+// Functions needed by the numerical factorization phase
+
+
+/** 
+  *  - Numerical factorization 
+  *  - Interleaved with the symbolic factorization 
+  * On exit,  info is 
+  * 
+  *    = 0: successful factorization
+  * 
+  *    > 0: if info = i, and i is
+  * 
+  *       <= A->ncol: U(i,i) is exactly zero. The factorization has
+  *          been completed, but the factor U is exactly singular,
+  *          and division by zero will occur if it is used to solve a
+  *          system of equations.
+  * 
+  *       > A->ncol: number of bytes allocated when memory allocation
+  *         failure occurred, plus A->ncol. If lwork = -1, it is
+  *         the estimated amount of space needed, plus A->ncol.  
+  */
+template <typename MatrixType, typename OrderingType>
+void SparseLU<MatrixType, OrderingType>::factorize(const MatrixType& matrix)
+{
+  using internal::emptyIdxLU;
+  eigen_assert(m_analysisIsOk && "analyzePattern() should be called first"); 
+  eigen_assert((matrix.rows() == matrix.cols()) && "Only for squared matrices");
+  
+  typedef typename IndexVector::Scalar StorageIndex; 
+  
+  m_isInitialized = true;
+  
+  
+  // Apply the column permutation computed in analyzepattern()
+  //   m_mat = matrix * m_perm_c.inverse(); 
+  m_mat = matrix;
+  if (m_perm_c.size()) 
+  {
+    m_mat.uncompress(); //NOTE: The effect of this command is only to create the InnerNonzeros pointers.
+    //Then, permute only the column pointers
+    const StorageIndex * outerIndexPtr;
+    if (matrix.isCompressed()) outerIndexPtr = matrix.outerIndexPtr();
+    else
+    {
+      StorageIndex* outerIndexPtr_t = new StorageIndex[matrix.cols()+1];
+      for(Index i = 0; i <= matrix.cols(); i++) outerIndexPtr_t[i] = m_mat.outerIndexPtr()[i];
+      outerIndexPtr = outerIndexPtr_t;
+    }
+    for (Index i = 0; i < matrix.cols(); i++)
+    {
+      m_mat.outerIndexPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i];
+      m_mat.innerNonZeroPtr()[m_perm_c.indices()(i)] = outerIndexPtr[i+1] - outerIndexPtr[i];
+    }
+    if(!matrix.isCompressed()) delete[] outerIndexPtr;
+  } 
+  else 
+  { //FIXME This should not be needed if the empty permutation is handled transparently
+    m_perm_c.resize(matrix.cols());
+    for(StorageIndex i = 0; i < matrix.cols(); ++i) m_perm_c.indices()(i) = i;
+  }
+  
+  Index m = m_mat.rows();
+  Index n = m_mat.cols();
+  Index nnz = m_mat.nonZeros();
+  Index maxpanel = m_perfv.panel_size * m;
+  // Allocate working storage common to the factor routines
+  Index lwork = 0;
+  Index info = Base::memInit(m, n, nnz, lwork, m_perfv.fillfactor, m_perfv.panel_size, m_glu); 
+  if (info) 
+  {
+    m_lastError = "UNABLE TO ALLOCATE WORKING MEMORY\n\n" ;
+    m_factorizationIsOk = false;
+    return ; 
+  }
+  
+  // Set up pointers for integer working arrays 
+  IndexVector segrep(m); segrep.setZero();
+  IndexVector parent(m); parent.setZero();
+  IndexVector xplore(m); xplore.setZero();
+  IndexVector repfnz(maxpanel);
+  IndexVector panel_lsub(maxpanel);
+  IndexVector xprune(n); xprune.setZero();
+  IndexVector marker(m*internal::LUNoMarker); marker.setZero();
+  
+  repfnz.setConstant(-1); 
+  panel_lsub.setConstant(-1);
+  
+  // Set up pointers for scalar working arrays 
+  ScalarVector dense; 
+  dense.setZero(maxpanel);
+  ScalarVector tempv; 
+  tempv.setZero(internal::LUnumTempV(m, m_perfv.panel_size, m_perfv.maxsuper, /*m_perfv.rowblk*/m) );
+  
+  // Compute the inverse of perm_c
+  PermutationType iperm_c(m_perm_c.inverse()); 
+  
+  // Identify initial relaxed snodes
+  IndexVector relax_end(n);
+  if ( m_symmetricmode == true ) 
+    Base::heap_relax_snode(n, m_etree, m_perfv.relax, marker, relax_end);
+  else
+    Base::relax_snode(n, m_etree, m_perfv.relax, marker, relax_end);
+  
+  
+  m_perm_r.resize(m); 
+  m_perm_r.indices().setConstant(-1);
+  marker.setConstant(-1);
+  m_detPermR = 1; // Record the determinant of the row permutation
+  
+  m_glu.supno(0) = emptyIdxLU; m_glu.xsup.setConstant(0);
+  m_glu.xsup(0) = m_glu.xlsub(0) = m_glu.xusub(0) = m_glu.xlusup(0) = Index(0);
+  
+  // Work on one 'panel' at a time. A panel is one of the following :
+  //  (a) a relaxed supernode at the bottom of the etree, or
+  //  (b) panel_size contiguous columns, <panel_size> defined by the user
+  Index jcol; 
+  IndexVector panel_histo(n);
+  Index pivrow; // Pivotal row number in the original row matrix
+  Index nseg1; // Number of segments in U-column above panel row jcol
+  Index nseg; // Number of segments in each U-column 
+  Index irep; 
+  Index i, k, jj; 
+  for (jcol = 0; jcol < n; )
+  {
+    // Adjust panel size so that a panel won't overlap with the next relaxed snode. 
+    Index panel_size = m_perfv.panel_size; // upper bound on panel width
+    for (k = jcol + 1; k < (std::min)(jcol+panel_size, n); k++)
+    {
+      if (relax_end(k) != emptyIdxLU) 
+      {
+        panel_size = k - jcol; 
+        break; 
+      }
+    }
+    if (k == n) 
+      panel_size = n - jcol; 
+      
+    // Symbolic outer factorization on a panel of columns 
+    Base::panel_dfs(m, panel_size, jcol, m_mat, m_perm_r.indices(), nseg1, dense, panel_lsub, segrep, repfnz, xprune, marker, parent, xplore, m_glu); 
+    
+    // Numeric sup-panel updates in topological order 
+    Base::panel_bmod(m, panel_size, jcol, nseg1, dense, tempv, segrep, repfnz, m_glu); 
+    
+    // Sparse LU within the panel, and below the panel diagonal 
+    for ( jj = jcol; jj< jcol + panel_size; jj++) 
+    {
+      k = (jj - jcol) * m; // Column index for w-wide arrays 
+      
+      nseg = nseg1; // begin after all the panel segments
+      //Depth-first-search for the current column
+      VectorBlock<IndexVector> panel_lsubk(panel_lsub, k, m);
+      VectorBlock<IndexVector> repfnz_k(repfnz, k, m); 
+      info = Base::column_dfs(m, jj, m_perm_r.indices(), m_perfv.maxsuper, nseg, panel_lsubk, segrep, repfnz_k, xprune, marker, parent, xplore, m_glu); 
+      if ( info ) 
+      {
+        m_lastError =  "UNABLE TO EXPAND MEMORY IN COLUMN_DFS() ";
+        m_info = NumericalIssue; 
+        m_factorizationIsOk = false; 
+        return; 
+      }
+      // Numeric updates to this column 
+      VectorBlock<ScalarVector> dense_k(dense, k, m); 
+      VectorBlock<IndexVector> segrep_k(segrep, nseg1, m-nseg1); 
+      info = Base::column_bmod(jj, (nseg - nseg1), dense_k, tempv, segrep_k, repfnz_k, jcol, m_glu); 
+      if ( info ) 
+      {
+        m_lastError = "UNABLE TO EXPAND MEMORY IN COLUMN_BMOD() ";
+        m_info = NumericalIssue; 
+        m_factorizationIsOk = false; 
+        return; 
+      }
+      
+      // Copy the U-segments to ucol(*)
+      info = Base::copy_to_ucol(jj, nseg, segrep, repfnz_k ,m_perm_r.indices(), dense_k, m_glu); 
+      if ( info ) 
+      {
+        m_lastError = "UNABLE TO EXPAND MEMORY IN COPY_TO_UCOL() ";
+        m_info = NumericalIssue; 
+        m_factorizationIsOk = false; 
+        return; 
+      }
+      
+      // Form the L-segment 
+      info = Base::pivotL(jj, m_diagpivotthresh, m_perm_r.indices(), iperm_c.indices(), pivrow, m_glu);
+      if ( info ) 
+      {
+        m_lastError = "THE MATRIX IS STRUCTURALLY SINGULAR ... ZERO COLUMN AT ";
+        std::ostringstream returnInfo;
+        returnInfo << info; 
+        m_lastError += returnInfo.str();
+        m_info = NumericalIssue; 
+        m_factorizationIsOk = false; 
+        return; 
+      }
+      
+      // Update the determinant of the row permutation matrix
+      // FIXME: the following test is not correct, we should probably take iperm_c into account and pivrow is not directly the row pivot.
+      if (pivrow != jj) m_detPermR = -m_detPermR;
+
+      // Prune columns (0:jj-1) using column jj
+      Base::pruneL(jj, m_perm_r.indices(), pivrow, nseg, segrep, repfnz_k, xprune, m_glu); 
+      
+      // Reset repfnz for this column 
+      for (i = 0; i < nseg; i++)
+      {
+        irep = segrep(i); 
+        repfnz_k(irep) = emptyIdxLU; 
+      }
+    } // end SparseLU within the panel  
+    jcol += panel_size;  // Move to the next panel
+  } // end for -- end elimination 
+  
+  m_detPermR = m_perm_r.determinant();
+  m_detPermC = m_perm_c.determinant();
+  
+  // Count the number of nonzeros in factors 
+  Base::countnz(n, m_nnzL, m_nnzU, m_glu); 
+  // Apply permutation  to the L subscripts 
+  Base::fixupL(n, m_perm_r.indices(), m_glu);
+  
+  // Create supernode matrix L 
+  m_Lstore.setInfos(m, n, m_glu.lusup, m_glu.xlusup, m_glu.lsub, m_glu.xlsub, m_glu.supno, m_glu.xsup); 
+  // Create the column major upper sparse matrix  U; 
+  new (&m_Ustore) MappedSparseMatrix<Scalar, ColMajor, StorageIndex> ( m, n, m_nnzU, m_glu.xusub.data(), m_glu.usub.data(), m_glu.ucol.data() );
+  
+  m_info = Success;
+  m_factorizationIsOk = true;
+}
+
+template<typename MappedSupernodalType>
+struct SparseLUMatrixLReturnType : internal::no_assignment_operator
+{
+  typedef typename MappedSupernodalType::Scalar Scalar;
+  explicit SparseLUMatrixLReturnType(const MappedSupernodalType& mapL) : m_mapL(mapL)
+  { }
+  Index rows() { return m_mapL.rows(); }
+  Index cols() { return m_mapL.cols(); }
+  template<typename Dest>
+  void solveInPlace( MatrixBase<Dest> &X) const
+  {
+    m_mapL.solveInPlace(X);
+  }
+  const MappedSupernodalType& m_mapL;
+};
+
+template<typename MatrixLType, typename MatrixUType>
+struct SparseLUMatrixUReturnType : internal::no_assignment_operator
+{
+  typedef typename MatrixLType::Scalar Scalar;
+  SparseLUMatrixUReturnType(const MatrixLType& mapL, const MatrixUType& mapU)
+  : m_mapL(mapL),m_mapU(mapU)
+  { }
+  Index rows() { return m_mapL.rows(); }
+  Index cols() { return m_mapL.cols(); }
+
+  template<typename Dest>   void solveInPlace(MatrixBase<Dest> &X) const
+  {
+    Index nrhs = X.cols();
+    Index n    = X.rows();
+    // Backward solve with U
+    for (Index k = m_mapL.nsuper(); k >= 0; k--)
+    {
+      Index fsupc = m_mapL.supToCol()[k];
+      Index lda = m_mapL.colIndexPtr()[fsupc+1] - m_mapL.colIndexPtr()[fsupc]; // leading dimension
+      Index nsupc = m_mapL.supToCol()[k+1] - fsupc;
+      Index luptr = m_mapL.colIndexPtr()[fsupc];
+
+      if (nsupc == 1)
+      {
+        for (Index j = 0; j < nrhs; j++)
+        {
+          X(fsupc, j) /= m_mapL.valuePtr()[luptr];
+        }
+      }
+      else
+      {
+        Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > A( &(m_mapL.valuePtr()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
+        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
+        U = A.template triangularView<Upper>().solve(U);
+      }
+
+      for (Index j = 0; j < nrhs; ++j)
+      {
+        for (Index jcol = fsupc; jcol < fsupc + nsupc; jcol++)
+        {
+          typename MatrixUType::InnerIterator it(m_mapU, jcol);
+          for ( ; it; ++it)
+          {
+            Index irow = it.index();
+            X(irow, j) -= X(jcol, j) * it.value();
+          }
+        }
+      }
+    } // End For U-solve
+  }
+  const MatrixLType& m_mapL;
+  const MatrixUType& m_mapU;
+};
+
+} // End namespace Eigen 
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLUImpl.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLUImpl.h
new file mode 100644
index 0000000..fc0cfc4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLUImpl.h
@@ -0,0 +1,66 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+#ifndef SPARSELU_IMPL_H
+#define SPARSELU_IMPL_H
+
+namespace Eigen {
+namespace internal {
+  
+/** \ingroup SparseLU_Module
+  * \class SparseLUImpl
+  * Base class for sparseLU
+  */
+template <typename Scalar, typename StorageIndex>
+class SparseLUImpl
+{
+  public:
+    typedef Matrix<Scalar,Dynamic,1> ScalarVector;
+    typedef Matrix<StorageIndex,Dynamic,1> IndexVector; 
+    typedef Matrix<Scalar,Dynamic,Dynamic,ColMajor> ScalarMatrix;
+    typedef Map<ScalarMatrix, 0,  OuterStride<> > MappedMatrixBlock;
+    typedef typename ScalarVector::RealScalar RealScalar; 
+    typedef Ref<Matrix<Scalar,Dynamic,1> > BlockScalarVector;
+    typedef Ref<Matrix<StorageIndex,Dynamic,1> > BlockIndexVector;
+    typedef LU_GlobalLU_t<IndexVector, ScalarVector> GlobalLU_t; 
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> MatrixType; 
+    
+  protected:
+     template <typename VectorType>
+     Index expand(VectorType& vec, Index& length, Index nbElts, Index keep_prev, Index& num_expansions);
+     Index memInit(Index m, Index n, Index annz, Index lwork, Index fillratio, Index panel_size,  GlobalLU_t& glu); 
+     template <typename VectorType>
+     Index memXpand(VectorType& vec, Index& maxlen, Index nbElts, MemType memtype, Index& num_expansions);
+     void heap_relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end); 
+     void relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end); 
+     Index snode_dfs(const Index jcol, const Index kcol,const MatrixType& mat,  IndexVector& xprune, IndexVector& marker, GlobalLU_t& glu); 
+     Index snode_bmod (const Index jcol, const Index fsupc, ScalarVector& dense, GlobalLU_t& glu);
+     Index pivotL(const Index jcol, const RealScalar& diagpivotthresh, IndexVector& perm_r, IndexVector& iperm_c, Index& pivrow, GlobalLU_t& glu);
+     template <typename Traits>
+     void dfs_kernel(const StorageIndex jj, IndexVector& perm_r,
+                    Index& nseg, IndexVector& panel_lsub, IndexVector& segrep,
+                    Ref<IndexVector> repfnz_col, IndexVector& xprune, Ref<IndexVector> marker, IndexVector& parent,
+                    IndexVector& xplore, GlobalLU_t& glu, Index& nextl_col, Index krow, Traits& traits);
+     void panel_dfs(const Index m, const Index w, const Index jcol, MatrixType& A, IndexVector& perm_r, Index& nseg, ScalarVector& dense, IndexVector& panel_lsub, IndexVector& segrep, IndexVector& repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu);
+    
+     void panel_bmod(const Index m, const Index w, const Index jcol, const Index nseg, ScalarVector& dense, ScalarVector& tempv, IndexVector& segrep, IndexVector& repfnz, GlobalLU_t& glu);
+     Index column_dfs(const Index m, const Index jcol, IndexVector& perm_r, Index maxsuper, Index& nseg,  BlockIndexVector lsub_col, IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu);
+     Index column_bmod(const Index jcol, const Index nseg, BlockScalarVector dense, ScalarVector& tempv, BlockIndexVector segrep, BlockIndexVector repfnz, Index fpanelc, GlobalLU_t& glu); 
+     Index copy_to_ucol(const Index jcol, const Index nseg, IndexVector& segrep, BlockIndexVector repfnz ,IndexVector& perm_r, BlockScalarVector dense, GlobalLU_t& glu); 
+     void pruneL(const Index jcol, const IndexVector& perm_r, const Index pivrow, const Index nseg, const IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, GlobalLU_t& glu);
+     void countnz(const Index n, Index& nnzL, Index& nnzU, GlobalLU_t& glu); 
+     void fixupL(const Index n, const IndexVector& perm_r, GlobalLU_t& glu); 
+     
+     template<typename , typename >
+     friend struct column_dfs_traits;
+}; 
+
+} // end namespace internal
+} // namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Memory.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Memory.h
new file mode 100644
index 0000000..4dc42e8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Memory.h
@@ -0,0 +1,226 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]memory.c files in SuperLU 
+ 
+ * -- SuperLU routine (version 3.1) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * August 1, 2008
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+
+#ifndef EIGEN_SPARSELU_MEMORY
+#define EIGEN_SPARSELU_MEMORY
+
+namespace Eigen {
+namespace internal {
+  
+enum { LUNoMarker = 3 };
+enum {emptyIdxLU = -1};
+inline Index LUnumTempV(Index& m, Index& w, Index& t, Index& b)
+{
+  return (std::max)(m, (t+b)*w);
+}
+
+template< typename Scalar>
+inline Index LUTempSpace(Index&m, Index& w)
+{
+  return (2*w + 4 + LUNoMarker) * m * sizeof(Index) + (w + 1) * m * sizeof(Scalar);
+}
+
+
+
+
+/** 
+  * Expand the existing storage to accomodate more fill-ins
+  * \param vec Valid pointer to the vector to allocate or expand
+  * \param[in,out] length  At input, contain the current length of the vector that is to be increased. At output, length of the newly allocated vector
+  * \param[in] nbElts Current number of elements in the factors
+  * \param keep_prev  1: use length  and do not expand the vector; 0: compute new_len and expand
+  * \param[in,out] num_expansions Number of times the memory has been expanded
+  */
+template <typename Scalar, typename StorageIndex>
+template <typename VectorType>
+Index  SparseLUImpl<Scalar,StorageIndex>::expand(VectorType& vec, Index& length, Index nbElts, Index keep_prev, Index& num_expansions) 
+{
+  
+  float alpha = 1.5; // Ratio of the memory increase 
+  Index new_len; // New size of the allocated memory
+  
+  if(num_expansions == 0 || keep_prev) 
+    new_len = length ; // First time allocate requested
+  else 
+    new_len = (std::max)(length+1,Index(alpha * length));
+  
+  VectorType old_vec; // Temporary vector to hold the previous values   
+  if (nbElts > 0 )
+    old_vec = vec.segment(0,nbElts); 
+  
+  //Allocate or expand the current vector
+#ifdef EIGEN_EXCEPTIONS
+  try
+#endif
+  {
+    vec.resize(new_len); 
+  }
+#ifdef EIGEN_EXCEPTIONS
+  catch(std::bad_alloc& )
+#else
+  if(!vec.size())
+#endif
+  {
+    if (!num_expansions)
+    {
+      // First time to allocate from LUMemInit()
+      // Let LUMemInit() deals with it.
+      return -1;
+    }
+    if (keep_prev)
+    {
+      // In this case, the memory length should not not be reduced
+      return new_len;
+    }
+    else 
+    {
+      // Reduce the size and increase again 
+      Index tries = 0; // Number of attempts
+      do 
+      {
+        alpha = (alpha + 1)/2;
+        new_len = (std::max)(length+1,Index(alpha * length));
+#ifdef EIGEN_EXCEPTIONS
+        try
+#endif
+        {
+          vec.resize(new_len); 
+        }
+#ifdef EIGEN_EXCEPTIONS
+        catch(std::bad_alloc& )
+#else
+        if (!vec.size())
+#endif
+        {
+          tries += 1; 
+          if ( tries > 10) return new_len; 
+        }
+      } while (!vec.size());
+    }
+  }
+  //Copy the previous values to the newly allocated space 
+  if (nbElts > 0)
+    vec.segment(0, nbElts) = old_vec;   
+   
+  
+  length  = new_len;
+  if(num_expansions) ++num_expansions;
+  return 0; 
+}
+
+/**
+ * \brief  Allocate various working space for the numerical factorization phase.
+ * \param m number of rows of the input matrix 
+ * \param n number of columns 
+ * \param annz number of initial nonzeros in the matrix 
+ * \param lwork  if lwork=-1, this routine returns an estimated size of the required memory
+ * \param glu persistent data to facilitate multiple factors : will be deleted later ??
+ * \param fillratio estimated ratio of fill in the factors
+ * \param panel_size Size of a panel
+ * \return an estimated size of the required memory if lwork = -1; otherwise, return the size of actually allocated memory when allocation failed, and 0 on success
+ * \note Unlike SuperLU, this routine does not support successive factorization with the same pattern and the same row permutation
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::memInit(Index m, Index n, Index annz, Index lwork, Index fillratio, Index panel_size,  GlobalLU_t& glu)
+{
+  Index& num_expansions = glu.num_expansions; //No memory expansions so far
+  num_expansions = 0;
+  glu.nzumax = glu.nzlumax = (std::min)(fillratio * (annz+1) / n, m) * n; // estimated number of nonzeros in U 
+  glu.nzlmax = (std::max)(Index(4), fillratio) * (annz+1) / 4; // estimated  nnz in L factor
+  // Return the estimated size to the user if necessary
+  Index tempSpace;
+  tempSpace = (2*panel_size + 4 + LUNoMarker) * m * sizeof(Index) + (panel_size + 1) * m * sizeof(Scalar);
+  if (lwork == emptyIdxLU) 
+  {
+    Index estimated_size;
+    estimated_size = (5 * n + 5) * sizeof(Index)  + tempSpace
+                    + (glu.nzlmax + glu.nzumax) * sizeof(Index) + (glu.nzlumax+glu.nzumax) *  sizeof(Scalar) + n; 
+    return estimated_size;
+  }
+  
+  // Setup the required space 
+  
+  // First allocate Integer pointers for L\U factors
+  glu.xsup.resize(n+1);
+  glu.supno.resize(n+1);
+  glu.xlsub.resize(n+1);
+  glu.xlusup.resize(n+1);
+  glu.xusub.resize(n+1);
+
+  // Reserve memory for L/U factors
+  do 
+  {
+    if(     (expand<ScalarVector>(glu.lusup, glu.nzlumax, 0, 0, num_expansions)<0)
+        ||  (expand<ScalarVector>(glu.ucol,  glu.nzumax,  0, 0, num_expansions)<0)
+        ||  (expand<IndexVector> (glu.lsub,  glu.nzlmax,  0, 0, num_expansions)<0)
+        ||  (expand<IndexVector> (glu.usub,  glu.nzumax,  0, 1, num_expansions)<0) )
+    {
+      //Reduce the estimated size and retry
+      glu.nzlumax /= 2;
+      glu.nzumax /= 2;
+      glu.nzlmax /= 2;
+      if (glu.nzlumax < annz ) return glu.nzlumax; 
+    }
+  } while (!glu.lusup.size() || !glu.ucol.size() || !glu.lsub.size() || !glu.usub.size());
+  
+  ++num_expansions;
+  return 0;
+  
+} // end LuMemInit
+
+/** 
+ * \brief Expand the existing storage 
+ * \param vec vector to expand 
+ * \param[in,out] maxlen On input, previous size of vec (Number of elements to copy ). on output, new size
+ * \param nbElts current number of elements in the vector.
+ * \param memtype Type of the element to expand
+ * \param num_expansions Number of expansions 
+ * \return 0 on success, > 0 size of the memory allocated so far
+ */
+template <typename Scalar, typename StorageIndex>
+template <typename VectorType>
+Index SparseLUImpl<Scalar,StorageIndex>::memXpand(VectorType& vec, Index& maxlen, Index nbElts, MemType memtype, Index& num_expansions)
+{
+  Index failed_size; 
+  if (memtype == USUB)
+     failed_size = this->expand<VectorType>(vec, maxlen, nbElts, 1, num_expansions);
+  else
+    failed_size = this->expand<VectorType>(vec, maxlen, nbElts, 0, num_expansions);
+
+  if (failed_size)
+    return failed_size; 
+  
+  return 0 ;  
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // EIGEN_SPARSELU_MEMORY
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Structs.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Structs.h
new file mode 100644
index 0000000..cf5ec44
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Structs.h
@@ -0,0 +1,110 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ * NOTE: This file comes from a partly modified version of files slu_[s,d,c,z]defs.h
+ * -- SuperLU routine (version 4.1) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November, 2010
+ * 
+ * Global data structures used in LU factorization -
+ * 
+ *   nsuper: #supernodes = nsuper + 1, numbered [0, nsuper].
+ *   (xsup,supno): supno[i] is the supernode no to which i belongs;
+ *  xsup(s) points to the beginning of the s-th supernode.
+ *  e.g.   supno 0 1 2 2 3 3 3 4 4 4 4 4   (n=12)
+ *          xsup 0 1 2 4 7 12
+ *  Note: dfs will be performed on supernode rep. relative to the new 
+ *        row pivoting ordering
+ *
+ *   (xlsub,lsub): lsub[*] contains the compressed subscript of
+ *  rectangular supernodes; xlsub[j] points to the starting
+ *  location of the j-th column in lsub[*]. Note that xlsub 
+ *  is indexed by column.
+ *  Storage: original row subscripts
+ *
+ *      During the course of sparse LU factorization, we also use
+ *  (xlsub,lsub) for the purpose of symmetric pruning. For each
+ *  supernode {s,s+1,...,t=s+r} with first column s and last
+ *  column t, the subscript set
+ *    lsub[j], j=xlsub[s], .., xlsub[s+1]-1
+ *  is the structure of column s (i.e. structure of this supernode).
+ *  It is used for the storage of numerical values.
+ *  Furthermore,
+ *    lsub[j], j=xlsub[t], .., xlsub[t+1]-1
+ *  is the structure of the last column t of this supernode.
+ *  It is for the purpose of symmetric pruning. Therefore, the
+ *  structural subscripts can be rearranged without making physical
+ *  interchanges among the numerical values.
+ *
+ *  However, if the supernode has only one column, then we
+ *  only keep one set of subscripts. For any subscript interchange
+ *  performed, similar interchange must be done on the numerical
+ *  values.
+ *
+ *  The last column structures (for pruning) will be removed
+ *  after the numercial LU factorization phase.
+ *
+ *   (xlusup,lusup): lusup[*] contains the numerical values of the
+ *  rectangular supernodes; xlusup[j] points to the starting
+ *  location of the j-th column in storage vector lusup[*]
+ *  Note: xlusup is indexed by column.
+ *  Each rectangular supernode is stored by column-major
+ *  scheme, consistent with Fortran 2-dim array storage.
+ *
+ *   (xusub,ucol,usub): ucol[*] stores the numerical values of
+ *  U-columns outside the rectangular supernodes. The row
+ *  subscript of nonzero ucol[k] is stored in usub[k].
+ *  xusub[i] points to the starting location of column i in ucol.
+ *  Storage: new row subscripts; that is subscripts of PA.
+ */
+
+#ifndef EIGEN_LU_STRUCTS
+#define EIGEN_LU_STRUCTS
+namespace Eigen {
+namespace internal {
+  
+typedef enum {LUSUP, UCOL, LSUB, USUB, LLVL, ULVL} MemType; 
+
+template <typename IndexVector, typename ScalarVector>
+struct LU_GlobalLU_t {
+  typedef typename IndexVector::Scalar StorageIndex; 
+  IndexVector xsup; //First supernode column ... xsup(s) points to the beginning of the s-th supernode
+  IndexVector supno; // Supernode number corresponding to this column (column to supernode mapping)
+  ScalarVector  lusup; // nonzero values of L ordered by columns 
+  IndexVector lsub; // Compressed row indices of L rectangular supernodes. 
+  IndexVector xlusup; // pointers to the beginning of each column in lusup
+  IndexVector xlsub; // pointers to the beginning of each column in lsub
+  Index   nzlmax; // Current max size of lsub
+  Index   nzlumax; // Current max size of lusup
+  ScalarVector  ucol; // nonzero values of U ordered by columns 
+  IndexVector usub; // row indices of U columns in ucol
+  IndexVector xusub; // Pointers to the beginning of each column of U in ucol 
+  Index   nzumax; // Current max size of ucol
+  Index   n; // Number of columns in the matrix  
+  Index   num_expansions; 
+};
+
+// Values to set for performance
+struct perfvalues {
+  Index panel_size; // a panel consists of at most <panel_size> consecutive columns
+  Index relax; // To control degree of relaxing supernodes. If the number of nodes (columns) 
+                // in a subtree of the elimination tree is less than relax, this subtree is considered 
+                // as one supernode regardless of the row structures of those columns
+  Index maxsuper; // The maximum size for a supernode in complete LU
+  Index rowblk; // The minimum row dimension for 2-D blocking to be used;
+  Index colblk; // The minimum column dimension for 2-D blocking to be used;
+  Index fillfactor; // The estimated fills factors for L and U, compared with A
+}; 
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // EIGEN_LU_STRUCTS
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h
new file mode 100644
index 0000000..721e188
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_SupernodalMatrix.h
@@ -0,0 +1,301 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSELU_SUPERNODAL_MATRIX_H
+#define EIGEN_SPARSELU_SUPERNODAL_MATRIX_H
+
+namespace Eigen {
+namespace internal {
+
+/** \ingroup SparseLU_Module
+ * \brief a class to manipulate the L supernodal factor from the SparseLU factorization
+ * 
+ * This class  contain the data to easily store 
+ * and manipulate the supernodes during the factorization and solution phase of Sparse LU. 
+ * Only the lower triangular matrix has supernodes.
+ * 
+ * NOTE : This class corresponds to the SCformat structure in SuperLU
+ * 
+ */
+/* TODO
+ * InnerIterator as for sparsematrix 
+ * SuperInnerIterator to iterate through all supernodes 
+ * Function for triangular solve
+ */
+template <typename _Scalar, typename _StorageIndex>
+class MappedSuperNodalMatrix
+{
+  public:
+    typedef _Scalar Scalar; 
+    typedef _StorageIndex StorageIndex;
+    typedef Matrix<StorageIndex,Dynamic,1> IndexVector;
+    typedef Matrix<Scalar,Dynamic,1> ScalarVector;
+  public:
+    MappedSuperNodalMatrix()
+    {
+      
+    }
+    MappedSuperNodalMatrix(Index m, Index n,  ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind,
+             IndexVector& rowind_colptr, IndexVector& col_to_sup, IndexVector& sup_to_col )
+    {
+      setInfos(m, n, nzval, nzval_colptr, rowind, rowind_colptr, col_to_sup, sup_to_col);
+    }
+    
+    ~MappedSuperNodalMatrix()
+    {
+      
+    }
+    /**
+     * Set appropriate pointers for the lower triangular supernodal matrix
+     * These infos are available at the end of the numerical factorization
+     * FIXME This class will be modified such that it can be use in the course 
+     * of the factorization.
+     */
+    void setInfos(Index m, Index n, ScalarVector& nzval, IndexVector& nzval_colptr, IndexVector& rowind,
+             IndexVector& rowind_colptr, IndexVector& col_to_sup, IndexVector& sup_to_col )
+    {
+      m_row = m;
+      m_col = n; 
+      m_nzval = nzval.data(); 
+      m_nzval_colptr = nzval_colptr.data(); 
+      m_rowind = rowind.data(); 
+      m_rowind_colptr = rowind_colptr.data(); 
+      m_nsuper = col_to_sup(n); 
+      m_col_to_sup = col_to_sup.data(); 
+      m_sup_to_col = sup_to_col.data(); 
+    }
+    
+    /**
+     * Number of rows
+     */
+    Index rows() { return m_row; }
+    
+    /**
+     * Number of columns
+     */
+    Index cols() { return m_col; }
+    
+    /**
+     * Return the array of nonzero values packed by column
+     * 
+     * The size is nnz
+     */
+    Scalar* valuePtr() {  return m_nzval; }
+    
+    const Scalar* valuePtr() const 
+    {
+      return m_nzval; 
+    }
+    /**
+     * Return the pointers to the beginning of each column in \ref valuePtr()
+     */
+    StorageIndex* colIndexPtr()
+    {
+      return m_nzval_colptr; 
+    }
+    
+    const StorageIndex* colIndexPtr() const
+    {
+      return m_nzval_colptr; 
+    }
+    
+    /**
+     * Return the array of compressed row indices of all supernodes
+     */
+    StorageIndex* rowIndex()  { return m_rowind; }
+    
+    const StorageIndex* rowIndex() const
+    {
+      return m_rowind; 
+    }
+    
+    /**
+     * Return the location in \em rowvaluePtr() which starts each column
+     */
+    StorageIndex* rowIndexPtr() { return m_rowind_colptr; }
+    
+    const StorageIndex* rowIndexPtr() const
+    {
+      return m_rowind_colptr; 
+    }
+    
+    /** 
+     * Return the array of column-to-supernode mapping 
+     */
+    StorageIndex* colToSup()  { return m_col_to_sup; }
+    
+    const StorageIndex* colToSup() const
+    {
+      return m_col_to_sup;       
+    }
+    /**
+     * Return the array of supernode-to-column mapping
+     */
+    StorageIndex* supToCol() { return m_sup_to_col; }
+    
+    const StorageIndex* supToCol() const
+    {
+      return m_sup_to_col;
+    }
+    
+    /**
+     * Return the number of supernodes
+     */
+    Index nsuper() const
+    {
+      return m_nsuper; 
+    }
+    
+    class InnerIterator; 
+    template<typename Dest>
+    void solveInPlace( MatrixBase<Dest>&X) const;
+    
+      
+      
+    
+  protected:
+    Index m_row; // Number of rows
+    Index m_col; // Number of columns
+    Index m_nsuper; // Number of supernodes
+    Scalar* m_nzval; //array of nonzero values packed by column
+    StorageIndex* m_nzval_colptr; //nzval_colptr[j] Stores the location in nzval[] which starts column j
+    StorageIndex* m_rowind; // Array of compressed row indices of rectangular supernodes
+    StorageIndex* m_rowind_colptr; //rowind_colptr[j] stores the location in rowind[] which starts column j
+    StorageIndex* m_col_to_sup; // col_to_sup[j] is the supernode number to which column j belongs
+    StorageIndex* m_sup_to_col; //sup_to_col[s] points to the starting column of the s-th supernode
+    
+  private :
+};
+
+/**
+  * \brief InnerIterator class to iterate over nonzero values of the current column in the supernodal matrix L
+  * 
+  */
+template<typename Scalar, typename StorageIndex>
+class MappedSuperNodalMatrix<Scalar,StorageIndex>::InnerIterator
+{
+  public:
+     InnerIterator(const MappedSuperNodalMatrix& mat, Index outer)
+      : m_matrix(mat),
+        m_outer(outer),
+        m_supno(mat.colToSup()[outer]),
+        m_idval(mat.colIndexPtr()[outer]),
+        m_startidval(m_idval),
+        m_endidval(mat.colIndexPtr()[outer+1]),
+        m_idrow(mat.rowIndexPtr()[mat.supToCol()[mat.colToSup()[outer]]]),
+        m_endidrow(mat.rowIndexPtr()[mat.supToCol()[mat.colToSup()[outer]]+1])
+    {}
+    inline InnerIterator& operator++()
+    { 
+      m_idval++; 
+      m_idrow++;
+      return *this;
+    }
+    inline Scalar value() const { return m_matrix.valuePtr()[m_idval]; }
+    
+    inline Scalar& valueRef() { return const_cast<Scalar&>(m_matrix.valuePtr()[m_idval]); }
+    
+    inline Index index() const { return m_matrix.rowIndex()[m_idrow]; }
+    inline Index row() const { return index(); }
+    inline Index col() const { return m_outer; }
+    
+    inline Index supIndex() const { return m_supno; }
+    
+    inline operator bool() const 
+    { 
+      return ( (m_idval < m_endidval) && (m_idval >= m_startidval)
+                && (m_idrow < m_endidrow) );
+    }
+    
+  protected:
+    const MappedSuperNodalMatrix& m_matrix; // Supernodal lower triangular matrix 
+    const Index m_outer;                    // Current column 
+    const Index m_supno;                    // Current SuperNode number
+    Index m_idval;                          // Index to browse the values in the current column
+    const Index m_startidval;               // Start of the column value
+    const Index m_endidval;                 // End of the column value
+    Index m_idrow;                          // Index to browse the row indices 
+    Index m_endidrow;                       // End index of row indices of the current column
+};
+
+/**
+ * \brief Solve with the supernode triangular matrix
+ * 
+ */
+template<typename Scalar, typename Index_>
+template<typename Dest>
+void MappedSuperNodalMatrix<Scalar,Index_>::solveInPlace( MatrixBase<Dest>&X) const
+{
+    /* Explicit type conversion as the Index type of MatrixBase<Dest> may be wider than Index */
+//    eigen_assert(X.rows() <= NumTraits<Index>::highest());
+//    eigen_assert(X.cols() <= NumTraits<Index>::highest());
+    Index n    = int(X.rows());
+    Index nrhs = Index(X.cols());
+    const Scalar * Lval = valuePtr();                 // Nonzero values 
+    Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor> work(n, nrhs);     // working vector
+    work.setZero();
+    for (Index k = 0; k <= nsuper(); k ++)
+    {
+      Index fsupc = supToCol()[k];                    // First column of the current supernode 
+      Index istart = rowIndexPtr()[fsupc];            // Pointer index to the subscript of the current column
+      Index nsupr = rowIndexPtr()[fsupc+1] - istart;  // Number of rows in the current supernode
+      Index nsupc = supToCol()[k+1] - fsupc;          // Number of columns in the current supernode
+      Index nrow = nsupr - nsupc;                     // Number of rows in the non-diagonal part of the supernode
+      Index irow;                                     //Current index row
+      
+      if (nsupc == 1 )
+      {
+        for (Index j = 0; j < nrhs; j++)
+        {
+          InnerIterator it(*this, fsupc);
+          ++it; // Skip the diagonal element
+          for (; it; ++it)
+          {
+            irow = it.row();
+            X(irow, j) -= X(fsupc, j) * it.value();
+          }
+        }
+      }
+      else
+      {
+        // The supernode has more than one column 
+        Index luptr = colIndexPtr()[fsupc]; 
+        Index lda = colIndexPtr()[fsupc+1] - luptr;
+        
+        // Triangular solve 
+        Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > A( &(Lval[luptr]), nsupc, nsupc, OuterStride<>(lda) );
+        Map< Matrix<Scalar,Dynamic,Dest::ColsAtCompileTime, ColMajor>, 0, OuterStride<> > U (&(X(fsupc,0)), nsupc, nrhs, OuterStride<>(n) );
+        U = A.template triangularView<UnitLower>().solve(U); 
+        
+        // Matrix-vector product 
+        new (&A) Map<const Matrix<Scalar,Dynamic,Dynamic, ColMajor>, 0, OuterStride<> > ( &(Lval[luptr+nsupc]), nrow, nsupc, OuterStride<>(lda) );
+        work.topRows(nrow).noalias() = A * U;
+        
+        //Begin Scatter 
+        for (Index j = 0; j < nrhs; j++)
+        {
+          Index iptr = istart + nsupc; 
+          for (Index i = 0; i < nrow; i++)
+          {
+            irow = rowIndex()[iptr]; 
+            X(irow, j) -= work(i, j); // Scatter operation
+            work(i, j) = Scalar(0); 
+            iptr++;
+          }
+        }
+      }
+    } 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_SPARSELU_MATRIX_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Utils.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Utils.h
new file mode 100644
index 0000000..9e3dab4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_Utils.h
@@ -0,0 +1,80 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+
+#ifndef EIGEN_SPARSELU_UTILS_H
+#define EIGEN_SPARSELU_UTILS_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Count Nonzero elements in the factors
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::countnz(const Index n, Index& nnzL, Index& nnzU, GlobalLU_t& glu)
+{
+ nnzL = 0; 
+ nnzU = (glu.xusub)(n); 
+ Index nsuper = (glu.supno)(n); 
+ Index jlen; 
+ Index i, j, fsupc;
+ if (n <= 0 ) return; 
+ // For each supernode
+ for (i = 0; i <= nsuper; i++)
+ {
+   fsupc = glu.xsup(i); 
+   jlen = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); 
+   
+   for (j = fsupc; j < glu.xsup(i+1); j++)
+   {
+     nnzL += jlen; 
+     nnzU += j - fsupc + 1; 
+     jlen--; 
+   }
+ }
+}
+
+/**
+ * \brief Fix up the data storage lsub for L-subscripts. 
+ * 
+ * It removes the subscripts sets for structural pruning, 
+ * and applies permutation to the remaining subscripts
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::fixupL(const Index n, const IndexVector& perm_r, GlobalLU_t& glu)
+{
+  Index fsupc, i, j, k, jstart; 
+  
+  StorageIndex nextl = 0; 
+  Index nsuper = (glu.supno)(n); 
+  
+  // For each supernode 
+  for (i = 0; i <= nsuper; i++)
+  {
+    fsupc = glu.xsup(i); 
+    jstart = glu.xlsub(fsupc); 
+    glu.xlsub(fsupc) = nextl; 
+    for (j = jstart; j < glu.xlsub(fsupc + 1); j++)
+    {
+      glu.lsub(nextl) = perm_r(glu.lsub(j)); // Now indexed into P*A
+      nextl++;
+    }
+    for (k = fsupc+1; k < glu.xsup(i+1); k++)
+      glu.xlsub(k) = nextl; // other columns in supernode i
+  }
+  
+  glu.xlsub(n) = nextl; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // EIGEN_SPARSELU_UTILS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_bmod.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_bmod.h
new file mode 100644
index 0000000..b57f068
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_bmod.h
@@ -0,0 +1,181 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of xcolumn_bmod.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_COLUMN_BMOD_H
+#define SPARSELU_COLUMN_BMOD_H
+
+namespace Eigen {
+
+namespace internal {
+/**
+ * \brief Performs numeric block updates (sup-col) in topological order
+ * 
+ * \param jcol current column to update
+ * \param nseg Number of segments in the U part
+ * \param dense Store the full representation of the column
+ * \param tempv working array 
+ * \param segrep segment representative ...
+ * \param repfnz ??? First nonzero column in each row ???  ...
+ * \param fpanelc First column in the current panel
+ * \param glu Global LU data. 
+ * \return 0 - successful return 
+ *         > 0 - number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::column_bmod(const Index jcol, const Index nseg, BlockScalarVector dense, ScalarVector& tempv,
+                                                     BlockIndexVector segrep, BlockIndexVector repfnz, Index fpanelc, GlobalLU_t& glu)
+{
+  Index  jsupno, k, ksub, krep, ksupno; 
+  Index lptr, nrow, isub, irow, nextlu, new_next, ufirst; 
+  Index fsupc, nsupc, nsupr, luptr, kfnz, no_zeros; 
+  /* krep = representative of current k-th supernode
+    * fsupc =  first supernodal column
+    * nsupc = number of columns in a supernode
+    * nsupr = number of rows in a supernode
+    * luptr = location of supernodal LU-block in storage
+    * kfnz = first nonz in the k-th supernodal segment
+    * no_zeros = no lf leading zeros in a supernodal U-segment
+    */
+  
+  jsupno = glu.supno(jcol);
+  // For each nonzero supernode segment of U[*,j] in topological order 
+  k = nseg - 1; 
+  Index d_fsupc; // distance between the first column of the current panel and the 
+               // first column of the current snode
+  Index fst_col; // First column within small LU update
+  Index segsize; 
+  for (ksub = 0; ksub < nseg; ksub++)
+  {
+    krep = segrep(k); k--; 
+    ksupno = glu.supno(krep); 
+    if (jsupno != ksupno )
+    {
+      // outside the rectangular supernode 
+      fsupc = glu.xsup(ksupno); 
+      fst_col = (std::max)(fsupc, fpanelc); 
+      
+      // Distance from the current supernode to the current panel; 
+      // d_fsupc = 0 if fsupc > fpanelc
+      d_fsupc = fst_col - fsupc; 
+      
+      luptr = glu.xlusup(fst_col) + d_fsupc; 
+      lptr = glu.xlsub(fsupc) + d_fsupc; 
+      
+      kfnz = repfnz(krep); 
+      kfnz = (std::max)(kfnz, fpanelc); 
+      
+      segsize = krep - kfnz + 1; 
+      nsupc = krep - fst_col + 1; 
+      nsupr = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); 
+      nrow = nsupr - d_fsupc - nsupc;
+      Index lda = glu.xlusup(fst_col+1) - glu.xlusup(fst_col);
+      
+      
+      // Perform a triangular solver and block update, 
+      // then scatter the result of sup-col update to dense
+      no_zeros = kfnz - fst_col; 
+      if(segsize==1)
+        LU_kernel_bmod<1>::run(segsize, dense, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+      else
+        LU_kernel_bmod<Dynamic>::run(segsize, dense, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+    } // end if jsupno 
+  } // end for each segment
+  
+  // Process the supernodal portion of  L\U[*,j]
+  nextlu = glu.xlusup(jcol); 
+  fsupc = glu.xsup(jsupno);
+  
+  // copy the SPA dense into L\U[*,j]
+  Index mem; 
+  new_next = nextlu + glu.xlsub(fsupc + 1) - glu.xlsub(fsupc); 
+  Index offset = internal::first_multiple<Index>(new_next, internal::packet_traits<Scalar>::size) - new_next;
+  if(offset)
+    new_next += offset;
+  while (new_next > glu.nzlumax )
+  {
+    mem = memXpand<ScalarVector>(glu.lusup, glu.nzlumax, nextlu, LUSUP, glu.num_expansions);  
+    if (mem) return mem; 
+  }
+  
+  for (isub = glu.xlsub(fsupc); isub < glu.xlsub(fsupc+1); isub++)
+  {
+    irow = glu.lsub(isub);
+    glu.lusup(nextlu) = dense(irow);
+    dense(irow) = Scalar(0.0); 
+    ++nextlu; 
+  }
+  
+  if(offset)
+  {
+    glu.lusup.segment(nextlu,offset).setZero();
+    nextlu += offset;
+  }
+  glu.xlusup(jcol + 1) = StorageIndex(nextlu);  // close L\U(*,jcol); 
+  
+  /* For more updates within the panel (also within the current supernode),
+   * should start from the first column of the panel, or the first column
+   * of the supernode, whichever is bigger. There are two cases:
+   *  1) fsupc < fpanelc, then fst_col <-- fpanelc
+   *  2) fsupc >= fpanelc, then fst_col <-- fsupc
+   */
+  fst_col = (std::max)(fsupc, fpanelc); 
+  
+  if (fst_col  < jcol)
+  {
+    // Distance between the current supernode and the current panel
+    // d_fsupc = 0 if fsupc >= fpanelc
+    d_fsupc = fst_col - fsupc; 
+    
+    lptr = glu.xlsub(fsupc) + d_fsupc; 
+    luptr = glu.xlusup(fst_col) + d_fsupc; 
+    nsupr = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); // leading dimension
+    nsupc = jcol - fst_col; // excluding jcol 
+    nrow = nsupr - d_fsupc - nsupc; 
+    
+    // points to the beginning of jcol in snode L\U(jsupno) 
+    ufirst = glu.xlusup(jcol) + d_fsupc; 
+    Index lda = glu.xlusup(jcol+1) - glu.xlusup(jcol);
+    MappedMatrixBlock A( &(glu.lusup.data()[luptr]), nsupc, nsupc, OuterStride<>(lda) );
+    VectorBlock<ScalarVector> u(glu.lusup, ufirst, nsupc); 
+    u = A.template triangularView<UnitLower>().solve(u); 
+    
+    new (&A) MappedMatrixBlock ( &(glu.lusup.data()[luptr+nsupc]), nrow, nsupc, OuterStride<>(lda) );
+    VectorBlock<ScalarVector> l(glu.lusup, ufirst+nsupc, nrow); 
+    l.noalias() -= A * u;
+    
+  } // End if fst_col
+  return 0; 
+}
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_COLUMN_BMOD_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_dfs.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_dfs.h
new file mode 100644
index 0000000..c98b30e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_column_dfs.h
@@ -0,0 +1,179 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]column_dfs.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_COLUMN_DFS_H
+#define SPARSELU_COLUMN_DFS_H
+
+template <typename Scalar, typename StorageIndex> class SparseLUImpl;
+namespace Eigen {
+
+namespace internal {
+
+template<typename IndexVector, typename ScalarVector>
+struct column_dfs_traits : no_assignment_operator
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  typedef typename IndexVector::Scalar StorageIndex;
+  column_dfs_traits(Index jcol, Index& jsuper, typename SparseLUImpl<Scalar, StorageIndex>::GlobalLU_t& glu, SparseLUImpl<Scalar, StorageIndex>& luImpl)
+   : m_jcol(jcol), m_jsuper_ref(jsuper), m_glu(glu), m_luImpl(luImpl)
+ {}
+  bool update_segrep(Index /*krep*/, Index /*jj*/)
+  {
+    return true;
+  }
+  void mem_expand(IndexVector& lsub, Index& nextl, Index chmark)
+  {
+    if (nextl >= m_glu.nzlmax)
+      m_luImpl.memXpand(lsub, m_glu.nzlmax, nextl, LSUB, m_glu.num_expansions); 
+    if (chmark != (m_jcol-1)) m_jsuper_ref = emptyIdxLU;
+  }
+  enum { ExpandMem = true };
+  
+  Index m_jcol;
+  Index& m_jsuper_ref;
+  typename SparseLUImpl<Scalar, StorageIndex>::GlobalLU_t& m_glu;
+  SparseLUImpl<Scalar, StorageIndex>& m_luImpl;
+};
+
+
+/**
+ * \brief Performs a symbolic factorization on column jcol and decide the supernode boundary
+ * 
+ * A supernode representative is the last column of a supernode.
+ * The nonzeros in U[*,j] are segments that end at supernodes representatives. 
+ * The routine returns a list of the supernodal representatives 
+ * in topological order of the dfs that generates them. 
+ * The location of the first nonzero in each supernodal segment 
+ * (supernodal entry location) is also returned. 
+ * 
+ * \param m number of rows in the matrix
+ * \param jcol Current column 
+ * \param perm_r Row permutation
+ * \param maxsuper  Maximum number of column allowed in a supernode
+ * \param [in,out] nseg Number of segments in current U[*,j] - new segments appended
+ * \param lsub_col defines the rhs vector to start the dfs
+ * \param [in,out] segrep Segment representatives - new segments appended 
+ * \param repfnz  First nonzero location in each row
+ * \param xprune 
+ * \param marker  marker[i] == jj, if i was visited during dfs of current column jj;
+ * \param parent
+ * \param xplore working array
+ * \param glu global LU data 
+ * \return 0 success
+ *         > 0 number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::column_dfs(const Index m, const Index jcol, IndexVector& perm_r, Index maxsuper, Index& nseg,
+                                                    BlockIndexVector lsub_col, IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune,
+                                                    IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu)
+{
+  
+  Index jsuper = glu.supno(jcol); 
+  Index nextl = glu.xlsub(jcol); 
+  VectorBlock<IndexVector> marker2(marker, 2*m, m); 
+  
+  
+  column_dfs_traits<IndexVector, ScalarVector> traits(jcol, jsuper, glu, *this);
+  
+  // For each nonzero in A(*,jcol) do dfs 
+  for (Index k = 0; ((k < m) ? lsub_col[k] != emptyIdxLU : false) ; k++)
+  {
+    Index krow = lsub_col(k); 
+    lsub_col(k) = emptyIdxLU; 
+    Index kmark = marker2(krow); 
+    
+    // krow was visited before, go to the next nonz; 
+    if (kmark == jcol) continue;
+    
+    dfs_kernel(StorageIndex(jcol), perm_r, nseg, glu.lsub, segrep, repfnz, xprune, marker2, parent,
+                   xplore, glu, nextl, krow, traits);
+  } // for each nonzero ... 
+  
+  Index fsupc;
+  StorageIndex nsuper = glu.supno(jcol);
+  StorageIndex jcolp1 = StorageIndex(jcol) + 1;
+  Index jcolm1 = jcol - 1;
+  
+  // check to see if j belongs in the same supernode as j-1
+  if ( jcol == 0 )
+  { // Do nothing for column 0 
+    nsuper = glu.supno(0) = 0 ;
+  }
+  else 
+  {
+    fsupc = glu.xsup(nsuper); 
+    StorageIndex jptr = glu.xlsub(jcol); // Not yet compressed
+    StorageIndex jm1ptr = glu.xlsub(jcolm1); 
+    
+    // Use supernodes of type T2 : see SuperLU paper
+    if ( (nextl-jptr != jptr-jm1ptr-1) ) jsuper = emptyIdxLU;
+    
+    // Make sure the number of columns in a supernode doesn't
+    // exceed threshold
+    if ( (jcol - fsupc) >= maxsuper) jsuper = emptyIdxLU; 
+    
+    /* If jcol starts a new supernode, reclaim storage space in
+     * glu.lsub from previous supernode. Note we only store 
+     * the subscript set of the first and last columns of 
+     * a supernode. (first for num values, last for pruning)
+     */
+    if (jsuper == emptyIdxLU)
+    { // starts a new supernode 
+      if ( (fsupc < jcolm1-1) ) 
+      { // >= 3 columns in nsuper
+        StorageIndex ito = glu.xlsub(fsupc+1);
+        glu.xlsub(jcolm1) = ito; 
+        StorageIndex istop = ito + jptr - jm1ptr; 
+        xprune(jcolm1) = istop; // intialize xprune(jcol-1)
+        glu.xlsub(jcol) = istop; 
+        
+        for (StorageIndex ifrom = jm1ptr; ifrom < nextl; ++ifrom, ++ito)
+          glu.lsub(ito) = glu.lsub(ifrom); 
+        nextl = ito;  // = istop + length(jcol)
+      }
+      nsuper++; 
+      glu.supno(jcol) = nsuper; 
+    } // if a new supernode 
+  } // end else:  jcol > 0
+  
+  // Tidy up the pointers before exit
+  glu.xsup(nsuper+1) = jcolp1; 
+  glu.supno(jcolp1) = nsuper; 
+  xprune(jcol) = StorageIndex(nextl);  // Intialize upper bound for pruning
+  glu.xlsub(jcolp1) = StorageIndex(nextl); 
+  
+  return 0; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h
new file mode 100644
index 0000000..c32d8d8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_copy_to_ucol.h
@@ -0,0 +1,107 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]copy_to_ucol.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_COPY_TO_UCOL_H
+#define SPARSELU_COPY_TO_UCOL_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Performs numeric block updates (sup-col) in topological order
+ * 
+ * \param jcol current column to update
+ * \param nseg Number of segments in the U part
+ * \param segrep segment representative ...
+ * \param repfnz First nonzero column in each row  ...
+ * \param perm_r Row permutation 
+ * \param dense Store the full representation of the column
+ * \param glu Global LU data. 
+ * \return 0 - successful return 
+ *         > 0 - number of bytes allocated when run out of space
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::copy_to_ucol(const Index jcol, const Index nseg, IndexVector& segrep,
+                                                      BlockIndexVector repfnz ,IndexVector& perm_r, BlockScalarVector dense, GlobalLU_t& glu)
+{  
+  Index ksub, krep, ksupno; 
+    
+  Index jsupno = glu.supno(jcol);
+  
+  // For each nonzero supernode segment of U[*,j] in topological order 
+  Index k = nseg - 1, i; 
+  StorageIndex nextu = glu.xusub(jcol); 
+  Index kfnz, isub, segsize; 
+  Index new_next,irow; 
+  Index fsupc, mem; 
+  for (ksub = 0; ksub < nseg; ksub++)
+  {
+    krep = segrep(k); k--; 
+    ksupno = glu.supno(krep); 
+    if (jsupno != ksupno ) // should go into ucol(); 
+    {
+      kfnz = repfnz(krep); 
+      if (kfnz != emptyIdxLU)
+      { // Nonzero U-segment 
+        fsupc = glu.xsup(ksupno); 
+        isub = glu.xlsub(fsupc) + kfnz - fsupc; 
+        segsize = krep - kfnz + 1; 
+        new_next = nextu + segsize; 
+        while (new_next > glu.nzumax) 
+        {
+          mem = memXpand<ScalarVector>(glu.ucol, glu.nzumax, nextu, UCOL, glu.num_expansions); 
+          if (mem) return mem; 
+          mem = memXpand<IndexVector>(glu.usub, glu.nzumax, nextu, USUB, glu.num_expansions); 
+          if (mem) return mem; 
+          
+        }
+        
+        for (i = 0; i < segsize; i++)
+        {
+          irow = glu.lsub(isub); 
+          glu.usub(nextu) = perm_r(irow); // Unlike the L part, the U part is stored in its final order
+          glu.ucol(nextu) = dense(irow); 
+          dense(irow) = Scalar(0.0); 
+          nextu++;
+          isub++;
+        }
+        
+      } // end nonzero U-segment 
+      
+    } // end if jsupno 
+    
+  } // end for each segment
+  glu.xusub(jcol + 1) = nextu; // close U(*,jcol)
+  return 0; 
+}
+
+} // namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_COPY_TO_UCOL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_gemm_kernel.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_gemm_kernel.h
new file mode 100644
index 0000000..95ba741
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_gemm_kernel.h
@@ -0,0 +1,280 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSELU_GEMM_KERNEL_H
+#define EIGEN_SPARSELU_GEMM_KERNEL_H
+
+namespace Eigen {
+
+namespace internal {
+
+
+/** \internal
+  * A general matrix-matrix product kernel optimized for the SparseLU factorization.
+  *  - A, B, and C must be column major
+  *  - lda and ldc must be multiples of the respective packet size
+  *  - C must have the same alignment as A
+  */
+template<typename Scalar>
+EIGEN_DONT_INLINE
+void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc)
+{
+  using namespace Eigen::internal;
+  
+  typedef typename packet_traits<Scalar>::type Packet;
+  enum {
+    NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
+    PacketSize = packet_traits<Scalar>::size,
+    PM = 8,                             // peeling in M
+    RN = 2,                             // register blocking
+    RK = NumberOfRegisters>=16 ? 4 : 2, // register blocking
+    BM = 4096/sizeof(Scalar),           // number of rows of A-C per chunk
+    SM = PM*PacketSize                  // step along M
+  };
+  Index d_end = (d/RK)*RK;    // number of columns of A (rows of B) suitable for full register blocking
+  Index n_end = (n/RN)*RN;    // number of columns of B-C suitable for processing RN columns at once
+  Index i0 = internal::first_default_aligned(A,m);
+  
+  eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_default_aligned(C,m)));
+  
+  // handle the non aligned rows of A and C without any optimization:
+  for(Index i=0; i<i0; ++i)
+  {
+    for(Index j=0; j<n; ++j)
+    {
+      Scalar c = C[i+j*ldc];
+      for(Index k=0; k<d; ++k)
+        c += B[k+j*ldb] * A[i+k*lda];
+      C[i+j*ldc] = c;
+    }
+  }
+  // process the remaining rows per chunk of BM rows
+  for(Index ib=i0; ib<m; ib+=BM)
+  {
+    Index actual_b = std::min<Index>(BM, m-ib);                 // actual number of rows
+    Index actual_b_end1 = (actual_b/SM)*SM;                   // actual number of rows suitable for peeling
+    Index actual_b_end2 = (actual_b/PacketSize)*PacketSize;   // actual number of rows suitable for vectorization
+    
+    // Let's process two columns of B-C at once
+    for(Index j=0; j<n_end; j+=RN)
+    {
+      const Scalar* Bc0 = B+(j+0)*ldb;
+      const Scalar* Bc1 = B+(j+1)*ldb;
+      
+      for(Index k=0; k<d_end; k+=RK)
+      {
+        
+        // load and expand a RN x RK block of B
+        Packet b00, b10, b20, b30, b01, b11, b21, b31;
+                  { b00 = pset1<Packet>(Bc0[0]); }
+                  { b10 = pset1<Packet>(Bc0[1]); }
+        if(RK==4) { b20 = pset1<Packet>(Bc0[2]); }
+        if(RK==4) { b30 = pset1<Packet>(Bc0[3]); }
+                  { b01 = pset1<Packet>(Bc1[0]); }
+                  { b11 = pset1<Packet>(Bc1[1]); }
+        if(RK==4) { b21 = pset1<Packet>(Bc1[2]); }
+        if(RK==4) { b31 = pset1<Packet>(Bc1[3]); }
+        
+        Packet a0, a1, a2, a3, c0, c1, t0, t1;
+        
+        const Scalar* A0 = A+ib+(k+0)*lda;
+        const Scalar* A1 = A+ib+(k+1)*lda;
+        const Scalar* A2 = A+ib+(k+2)*lda;
+        const Scalar* A3 = A+ib+(k+3)*lda;
+        
+        Scalar* C0 = C+ib+(j+0)*ldc;
+        Scalar* C1 = C+ib+(j+1)*ldc;
+        
+                  a0 = pload<Packet>(A0);
+                  a1 = pload<Packet>(A1);
+        if(RK==4)
+        {
+          a2 = pload<Packet>(A2);
+          a3 = pload<Packet>(A3);
+        }
+        else
+        {
+          // workaround "may be used uninitialized in this function" warning
+          a2 = a3 = a0;
+        }
+        
+#define KMADD(c, a, b, tmp) {tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);}
+#define WORK(I)  \
+                     c0 = pload<Packet>(C0+i+(I)*PacketSize);    \
+                     c1 = pload<Packet>(C1+i+(I)*PacketSize);    \
+                     KMADD(c0, a0, b00, t0)                      \
+                     KMADD(c1, a0, b01, t1)                      \
+                     a0 = pload<Packet>(A0+i+(I+1)*PacketSize);  \
+                     KMADD(c0, a1, b10, t0)                      \
+                     KMADD(c1, a1, b11, t1)                      \
+                     a1 = pload<Packet>(A1+i+(I+1)*PacketSize);  \
+          if(RK==4){ KMADD(c0, a2, b20, t0)                     }\
+          if(RK==4){ KMADD(c1, a2, b21, t1)                     }\
+          if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
+          if(RK==4){ KMADD(c0, a3, b30, t0)                     }\
+          if(RK==4){ KMADD(c1, a3, b31, t1)                     }\
+          if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
+                     pstore(C0+i+(I)*PacketSize, c0);            \
+                     pstore(C1+i+(I)*PacketSize, c1)
+        
+        // process rows of A' - C' with aggressive vectorization and peeling 
+        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
+        {
+          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL1");
+                    prefetch((A0+i+(5)*PacketSize));
+                    prefetch((A1+i+(5)*PacketSize));
+          if(RK==4) prefetch((A2+i+(5)*PacketSize));
+          if(RK==4) prefetch((A3+i+(5)*PacketSize));
+
+          WORK(0);
+          WORK(1);
+          WORK(2);
+          WORK(3);
+          WORK(4);
+          WORK(5);
+          WORK(6);
+          WORK(7);
+        }
+        // process the remaining rows with vectorization only
+        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
+        {
+          WORK(0);
+        }
+#undef WORK
+        // process the remaining rows without vectorization
+        for(Index i=actual_b_end2; i<actual_b; ++i)
+        {
+          if(RK==4)
+          {
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
+            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3];
+          }
+          else
+          {
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
+            C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1];
+          }
+        }
+        
+        Bc0 += RK;
+        Bc1 += RK;
+      } // peeled loop on k
+    } // peeled loop on the columns j
+    // process the last column (we now perform a matrix-vector product)
+    if((n-n_end)>0)
+    {
+      const Scalar* Bc0 = B+(n-1)*ldb;
+      
+      for(Index k=0; k<d_end; k+=RK)
+      {
+        
+        // load and expand a 1 x RK block of B
+        Packet b00, b10, b20, b30;
+                  b00 = pset1<Packet>(Bc0[0]);
+                  b10 = pset1<Packet>(Bc0[1]);
+        if(RK==4) b20 = pset1<Packet>(Bc0[2]);
+        if(RK==4) b30 = pset1<Packet>(Bc0[3]);
+        
+        Packet a0, a1, a2, a3, c0, t0/*, t1*/;
+        
+        const Scalar* A0 = A+ib+(k+0)*lda;
+        const Scalar* A1 = A+ib+(k+1)*lda;
+        const Scalar* A2 = A+ib+(k+2)*lda;
+        const Scalar* A3 = A+ib+(k+3)*lda;
+        
+        Scalar* C0 = C+ib+(n_end)*ldc;
+        
+                  a0 = pload<Packet>(A0);
+                  a1 = pload<Packet>(A1);
+        if(RK==4)
+        {
+          a2 = pload<Packet>(A2);
+          a3 = pload<Packet>(A3);
+        }
+        else
+        {
+          // workaround "may be used uninitialized in this function" warning
+          a2 = a3 = a0;
+        }
+        
+#define WORK(I) \
+                   c0 = pload<Packet>(C0+i+(I)*PacketSize);     \
+                   KMADD(c0, a0, b00, t0)                       \
+                   a0 = pload<Packet>(A0+i+(I+1)*PacketSize);   \
+                   KMADD(c0, a1, b10, t0)                       \
+                   a1 = pload<Packet>(A1+i+(I+1)*PacketSize);   \
+        if(RK==4){ KMADD(c0, a2, b20, t0)                      }\
+        if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize);  }\
+        if(RK==4){ KMADD(c0, a3, b30, t0)                      }\
+        if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize);  }\
+                   pstore(C0+i+(I)*PacketSize, c0);
+        
+        // agressive vectorization and peeling
+        for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
+        {
+          EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
+          WORK(0);
+          WORK(1);
+          WORK(2);
+          WORK(3);
+          WORK(4);
+          WORK(5);
+          WORK(6);
+          WORK(7);
+        }
+        // vectorization only
+        for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
+        {
+          WORK(0);
+        }
+        // remaining scalars
+        for(Index i=actual_b_end2; i<actual_b; ++i)
+        {
+          if(RK==4) 
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
+          else
+            C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
+        }
+        
+        Bc0 += RK;
+#undef WORK
+      }
+    }
+    
+    // process the last columns of A, corresponding to the last rows of B
+    Index rd = d-d_end;
+    if(rd>0)
+    {
+      for(Index j=0; j<n; ++j)
+      {
+        enum {
+          Alignment = PacketSize>1 ? Aligned : 0
+        };
+        typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector;
+        typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector;
+        if(rd==1)       MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b);
+        
+        else if(rd==2)  MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
+                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b);
+        
+        else            MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
+                                                        + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b)
+                                                        + B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b);
+      }
+    }
+  
+  } // blocking on the rows of A and C
+}
+#undef KMADD
+
+} // namespace internal
+
+} // namespace Eigen
+
+#endif // EIGEN_SPARSELU_GEMM_KERNEL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h
new file mode 100644
index 0000000..6f75d50
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_heap_relax_snode.h
@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* This file is a modified version of heap_relax_snode.c file in SuperLU
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+
+#ifndef SPARSELU_HEAP_RELAX_SNODE_H
+#define SPARSELU_HEAP_RELAX_SNODE_H
+
+namespace Eigen {
+namespace internal {
+
+/** 
+ * \brief Identify the initial relaxed supernodes
+ * 
+ * This routine applied to a symmetric elimination tree. 
+ * It assumes that the matrix has been reordered according to the postorder of the etree
+ * \param n The number of columns
+ * \param et elimination tree 
+ * \param relax_columns Maximum number of columns allowed in a relaxed snode 
+ * \param descendants Number of descendants of each node in the etree
+ * \param relax_end last column in a supernode
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::heap_relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end)
+{
+  
+  // The etree may not be postordered, but its heap ordered  
+  IndexVector post;
+  internal::treePostorder(StorageIndex(n), et, post); // Post order etree
+  IndexVector inv_post(n+1); 
+  for (StorageIndex i = 0; i < n+1; ++i) inv_post(post(i)) = i; // inv_post = post.inverse()???
+  
+  // Renumber etree in postorder 
+  IndexVector iwork(n);
+  IndexVector et_save(n+1);
+  for (Index i = 0; i < n; ++i)
+  {
+    iwork(post(i)) = post(et(i));
+  }
+  et_save = et; // Save the original etree
+  et = iwork; 
+  
+  // compute the number of descendants of each node in the etree
+  relax_end.setConstant(emptyIdxLU);
+  Index j, parent; 
+  descendants.setZero();
+  for (j = 0; j < n; j++) 
+  {
+    parent = et(j);
+    if (parent != n) // not the dummy root
+      descendants(parent) += descendants(j) + 1;
+  }
+  // Identify the relaxed supernodes by postorder traversal of the etree
+  Index snode_start; // beginning of a snode 
+  StorageIndex k;
+  Index nsuper_et_post = 0; // Number of relaxed snodes in postordered etree 
+  Index nsuper_et = 0; // Number of relaxed snodes in the original etree 
+  StorageIndex l; 
+  for (j = 0; j < n; )
+  {
+    parent = et(j);
+    snode_start = j; 
+    while ( parent != n && descendants(parent) < relax_columns ) 
+    {
+      j = parent; 
+      parent = et(j);
+    }
+    // Found a supernode in postordered etree, j is the last column 
+    ++nsuper_et_post;
+    k = StorageIndex(n);
+    for (Index i = snode_start; i <= j; ++i)
+      k = (std::min)(k, inv_post(i));
+    l = inv_post(j);
+    if ( (l - k) == (j - snode_start) )  // Same number of columns in the snode
+    {
+      // This is also a supernode in the original etree
+      relax_end(k) = l; // Record last column 
+      ++nsuper_et; 
+    }
+    else 
+    {
+      for (Index i = snode_start; i <= j; ++i) 
+      {
+        l = inv_post(i);
+        if (descendants(i) == 0) 
+        {
+          relax_end(l) = l;
+          ++nsuper_et;
+        }
+      }
+    }
+    j++;
+    // Search for a new leaf
+    while (descendants(j) != 0 && j < n) j++;
+  } // End postorder traversal of the etree
+  
+  // Recover the original etree
+  et = et_save; 
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // SPARSELU_HEAP_RELAX_SNODE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_kernel_bmod.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_kernel_bmod.h
new file mode 100644
index 0000000..8c1b3e8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_kernel_bmod.h
@@ -0,0 +1,130 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef SPARSELU_KERNEL_BMOD_H
+#define SPARSELU_KERNEL_BMOD_H
+
+namespace Eigen {
+namespace internal {
+  
+template <int SegSizeAtCompileTime> struct LU_kernel_bmod
+{
+  /** \internal
+    * \brief Performs numeric block updates from a given supernode to a single column
+    *
+    * \param segsize Size of the segment (and blocks ) to use for updates
+    * \param[in,out] dense Packed values of the original matrix
+    * \param tempv temporary vector to use for updates
+    * \param lusup array containing the supernodes
+    * \param lda Leading dimension in the supernode
+    * \param nrow Number of rows in the rectangular part of the supernode
+    * \param lsub compressed row subscripts of supernodes
+    * \param lptr pointer to the first column of the current supernode in lsub
+    * \param no_zeros Number of nonzeros elements before the diagonal part of the supernode
+    */
+  template <typename BlockScalarVector, typename ScalarVector, typename IndexVector>
+  static EIGEN_DONT_INLINE void run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda,
+                                    const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros);
+};
+
+template <int SegSizeAtCompileTime>
+template <typename BlockScalarVector, typename ScalarVector, typename IndexVector>
+EIGEN_DONT_INLINE void LU_kernel_bmod<SegSizeAtCompileTime>::run(const Index segsize, BlockScalarVector& dense, ScalarVector& tempv, ScalarVector& lusup, Index& luptr, const Index lda,
+                                                                  const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros)
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  // First, copy U[*,j] segment from dense(*) to tempv(*)
+  // The result of triangular solve is in tempv[*]; 
+    // The result of matric-vector update is in dense[*]
+  Index isub = lptr + no_zeros; 
+  Index i;
+  Index irow;
+  for (i = 0; i < ((SegSizeAtCompileTime==Dynamic)?segsize:SegSizeAtCompileTime); i++)
+  {
+    irow = lsub(isub); 
+    tempv(i) = dense(irow); 
+    ++isub; 
+  }
+  // Dense triangular solve -- start effective triangle
+  luptr += lda * no_zeros + no_zeros; 
+  // Form Eigen matrix and vector 
+  Map<Matrix<Scalar,SegSizeAtCompileTime,SegSizeAtCompileTime, ColMajor>, 0, OuterStride<> > A( &(lusup.data()[luptr]), segsize, segsize, OuterStride<>(lda) );
+  Map<Matrix<Scalar,SegSizeAtCompileTime,1> > u(tempv.data(), segsize);
+  
+  u = A.template triangularView<UnitLower>().solve(u); 
+  
+  // Dense matrix-vector product y <-- B*x 
+  luptr += segsize;
+  const Index PacketSize = internal::packet_traits<Scalar>::size;
+  Index ldl = internal::first_multiple(nrow, PacketSize);
+  Map<Matrix<Scalar,Dynamic,SegSizeAtCompileTime, ColMajor>, 0, OuterStride<> > B( &(lusup.data()[luptr]), nrow, segsize, OuterStride<>(lda) );
+  Index aligned_offset = internal::first_default_aligned(tempv.data()+segsize, PacketSize);
+  Index aligned_with_B_offset = (PacketSize-internal::first_default_aligned(B.data(), PacketSize))%PacketSize;
+  Map<Matrix<Scalar,Dynamic,1>, 0, OuterStride<> > l(tempv.data()+segsize+aligned_offset+aligned_with_B_offset, nrow, OuterStride<>(ldl) );
+  
+  l.setZero();
+  internal::sparselu_gemm<Scalar>(l.rows(), l.cols(), B.cols(), B.data(), B.outerStride(), u.data(), u.outerStride(), l.data(), l.outerStride());
+  
+  // Scatter tempv[] into SPA dense[] as a temporary storage 
+  isub = lptr + no_zeros;
+  for (i = 0; i < ((SegSizeAtCompileTime==Dynamic)?segsize:SegSizeAtCompileTime); i++)
+  {
+    irow = lsub(isub++); 
+    dense(irow) = tempv(i);
+  }
+  
+  // Scatter l into SPA dense[]
+  for (i = 0; i < nrow; i++)
+  {
+    irow = lsub(isub++); 
+    dense(irow) -= l(i);
+  } 
+}
+
+template <> struct LU_kernel_bmod<1>
+{
+  template <typename BlockScalarVector, typename ScalarVector, typename IndexVector>
+  static EIGEN_DONT_INLINE void run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr,
+                                    const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros);
+};
+
+
+template <typename BlockScalarVector, typename ScalarVector, typename IndexVector>
+EIGEN_DONT_INLINE void LU_kernel_bmod<1>::run(const Index /*segsize*/, BlockScalarVector& dense, ScalarVector& /*tempv*/, ScalarVector& lusup, Index& luptr,
+                                              const Index lda, const Index nrow, IndexVector& lsub, const Index lptr, const Index no_zeros)
+{
+  typedef typename ScalarVector::Scalar Scalar;
+  typedef typename IndexVector::Scalar StorageIndex;
+  Scalar f = dense(lsub(lptr + no_zeros));
+  luptr += lda * no_zeros + no_zeros + 1;
+  const Scalar* a(lusup.data() + luptr);
+  const StorageIndex*  irow(lsub.data()+lptr + no_zeros + 1);
+  Index i = 0;
+  for (; i+1 < nrow; i+=2)
+  {
+    Index i0 = *(irow++);
+    Index i1 = *(irow++);
+    Scalar a0 = *(a++);
+    Scalar a1 = *(a++);
+    Scalar d0 = dense.coeff(i0);
+    Scalar d1 = dense.coeff(i1);
+    d0 -= f*a0;
+    d1 -= f*a1;
+    dense.coeffRef(i0) = d0;
+    dense.coeffRef(i1) = d1;
+  }
+  if(i<nrow)
+    dense.coeffRef(*(irow++)) -= f * *(a++);
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif // SPARSELU_KERNEL_BMOD_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_bmod.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_bmod.h
new file mode 100644
index 0000000..822cf32
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_bmod.h
@@ -0,0 +1,223 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]panel_bmod.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_PANEL_BMOD_H
+#define SPARSELU_PANEL_BMOD_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Performs numeric block updates (sup-panel) in topological order.
+ * 
+ * Before entering this routine, the original nonzeros in the panel
+ * were already copied i nto the spa[m,w]
+ * 
+ * \param m number of rows in the matrix
+ * \param w Panel size
+ * \param jcol Starting  column of the panel
+ * \param nseg Number of segments in the U part
+ * \param dense Store the full representation of the panel 
+ * \param tempv working array 
+ * \param segrep segment representative... first row in the segment
+ * \param repfnz First nonzero rows
+ * \param glu Global LU data. 
+ * 
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::panel_bmod(const Index m, const Index w, const Index jcol, 
+                                            const Index nseg, ScalarVector& dense, ScalarVector& tempv,
+                                            IndexVector& segrep, IndexVector& repfnz, GlobalLU_t& glu)
+{
+  
+  Index ksub,jj,nextl_col; 
+  Index fsupc, nsupc, nsupr, nrow; 
+  Index krep, kfnz; 
+  Index lptr; // points to the row subscripts of a supernode 
+  Index luptr; // ...
+  Index segsize,no_zeros ; 
+  // For each nonz supernode segment of U[*,j] in topological order
+  Index k = nseg - 1; 
+  const Index PacketSize = internal::packet_traits<Scalar>::size;
+  
+  for (ksub = 0; ksub < nseg; ksub++)
+  { // For each updating supernode
+    /* krep = representative of current k-th supernode
+     * fsupc =  first supernodal column
+     * nsupc = number of columns in a supernode
+     * nsupr = number of rows in a supernode
+     */
+    krep = segrep(k); k--; 
+    fsupc = glu.xsup(glu.supno(krep)); 
+    nsupc = krep - fsupc + 1; 
+    nsupr = glu.xlsub(fsupc+1) - glu.xlsub(fsupc); 
+    nrow = nsupr - nsupc; 
+    lptr = glu.xlsub(fsupc); 
+    
+    // loop over the panel columns to detect the actual number of columns and rows
+    Index u_rows = 0;
+    Index u_cols = 0;
+    for (jj = jcol; jj < jcol + w; jj++)
+    {
+      nextl_col = (jj-jcol) * m; 
+      VectorBlock<IndexVector> repfnz_col(repfnz, nextl_col, m); // First nonzero column index for each row
+      
+      kfnz = repfnz_col(krep); 
+      if ( kfnz == emptyIdxLU ) 
+        continue; // skip any zero segment
+      
+      segsize = krep - kfnz + 1;
+      u_cols++;
+      u_rows = (std::max)(segsize,u_rows);
+    }
+    
+    if(nsupc >= 2)
+    { 
+      Index ldu = internal::first_multiple<Index>(u_rows, PacketSize);
+      Map<ScalarMatrix, Aligned,  OuterStride<> > U(tempv.data(), u_rows, u_cols, OuterStride<>(ldu));
+      
+      // gather U
+      Index u_col = 0;
+      for (jj = jcol; jj < jcol + w; jj++)
+      {
+        nextl_col = (jj-jcol) * m; 
+        VectorBlock<IndexVector> repfnz_col(repfnz, nextl_col, m); // First nonzero column index for each row
+        VectorBlock<ScalarVector> dense_col(dense, nextl_col, m); // Scatter/gather entire matrix column from/to here
+        
+        kfnz = repfnz_col(krep); 
+        if ( kfnz == emptyIdxLU ) 
+          continue; // skip any zero segment
+        
+        segsize = krep - kfnz + 1;
+        luptr = glu.xlusup(fsupc);    
+        no_zeros = kfnz - fsupc; 
+        
+        Index isub = lptr + no_zeros;
+        Index off = u_rows-segsize;
+        for (Index i = 0; i < off; i++) U(i,u_col) = 0;
+        for (Index i = 0; i < segsize; i++)
+        {
+          Index irow = glu.lsub(isub); 
+          U(i+off,u_col) = dense_col(irow); 
+          ++isub; 
+        }
+        u_col++;
+      }
+      // solve U = A^-1 U
+      luptr = glu.xlusup(fsupc);
+      Index lda = glu.xlusup(fsupc+1) - glu.xlusup(fsupc);
+      no_zeros = (krep - u_rows + 1) - fsupc;
+      luptr += lda * no_zeros + no_zeros;
+      MappedMatrixBlock A(glu.lusup.data()+luptr, u_rows, u_rows, OuterStride<>(lda) );
+      U = A.template triangularView<UnitLower>().solve(U);
+      
+      // update
+      luptr += u_rows;
+      MappedMatrixBlock B(glu.lusup.data()+luptr, nrow, u_rows, OuterStride<>(lda) );
+      eigen_assert(tempv.size()>w*ldu + nrow*w + 1);
+      
+      Index ldl = internal::first_multiple<Index>(nrow, PacketSize);
+      Index offset = (PacketSize-internal::first_default_aligned(B.data(), PacketSize)) % PacketSize;
+      MappedMatrixBlock L(tempv.data()+w*ldu+offset, nrow, u_cols, OuterStride<>(ldl));
+      
+      L.setZero();
+      internal::sparselu_gemm<Scalar>(L.rows(), L.cols(), B.cols(), B.data(), B.outerStride(), U.data(), U.outerStride(), L.data(), L.outerStride());
+      
+      // scatter U and L
+      u_col = 0;
+      for (jj = jcol; jj < jcol + w; jj++)
+      {
+        nextl_col = (jj-jcol) * m; 
+        VectorBlock<IndexVector> repfnz_col(repfnz, nextl_col, m); // First nonzero column index for each row
+        VectorBlock<ScalarVector> dense_col(dense, nextl_col, m); // Scatter/gather entire matrix column from/to here
+        
+        kfnz = repfnz_col(krep); 
+        if ( kfnz == emptyIdxLU ) 
+          continue; // skip any zero segment
+        
+        segsize = krep - kfnz + 1;
+        no_zeros = kfnz - fsupc; 
+        Index isub = lptr + no_zeros;
+        
+        Index off = u_rows-segsize;
+        for (Index i = 0; i < segsize; i++)
+        {
+          Index irow = glu.lsub(isub++); 
+          dense_col(irow) = U.coeff(i+off,u_col);
+          U.coeffRef(i+off,u_col) = 0;
+        }
+        
+        // Scatter l into SPA dense[]
+        for (Index i = 0; i < nrow; i++)
+        {
+          Index irow = glu.lsub(isub++); 
+          dense_col(irow) -= L.coeff(i,u_col);
+          L.coeffRef(i,u_col) = 0;
+        }
+        u_col++;
+      }
+    }
+    else // level 2 only
+    {
+      // Sequence through each column in the panel
+      for (jj = jcol; jj < jcol + w; jj++)
+      {
+        nextl_col = (jj-jcol) * m; 
+        VectorBlock<IndexVector> repfnz_col(repfnz, nextl_col, m); // First nonzero column index for each row
+        VectorBlock<ScalarVector> dense_col(dense, nextl_col, m); // Scatter/gather entire matrix column from/to here
+        
+        kfnz = repfnz_col(krep); 
+        if ( kfnz == emptyIdxLU ) 
+          continue; // skip any zero segment
+        
+        segsize = krep - kfnz + 1;
+        luptr = glu.xlusup(fsupc);
+        
+        Index lda = glu.xlusup(fsupc+1)-glu.xlusup(fsupc);// nsupr
+        
+        // Perform a trianglar solve and block update, 
+        // then scatter the result of sup-col update to dense[]
+        no_zeros = kfnz - fsupc; 
+              if(segsize==1)  LU_kernel_bmod<1>::run(segsize, dense_col, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+        else  if(segsize==2)  LU_kernel_bmod<2>::run(segsize, dense_col, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+        else  if(segsize==3)  LU_kernel_bmod<3>::run(segsize, dense_col, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros);
+        else                  LU_kernel_bmod<Dynamic>::run(segsize, dense_col, tempv, glu.lusup, luptr, lda, nrow, glu.lsub, lptr, no_zeros); 
+      } // End for each column in the panel 
+    }
+    
+  } // End for each updating supernode
+} // end panel bmod
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // SPARSELU_PANEL_BMOD_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_dfs.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_dfs.h
new file mode 100644
index 0000000..155df73
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_panel_dfs.h
@@ -0,0 +1,258 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]panel_dfs.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_PANEL_DFS_H
+#define SPARSELU_PANEL_DFS_H
+
+namespace Eigen {
+
+namespace internal {
+  
+template<typename IndexVector>
+struct panel_dfs_traits
+{
+  typedef typename IndexVector::Scalar StorageIndex;
+  panel_dfs_traits(Index jcol, StorageIndex* marker)
+    : m_jcol(jcol), m_marker(marker)
+  {}
+  bool update_segrep(Index krep, StorageIndex jj)
+  {
+    if(m_marker[krep]<m_jcol)
+    {
+      m_marker[krep] = jj; 
+      return true;
+    }
+    return false;
+  }
+  void mem_expand(IndexVector& /*glu.lsub*/, Index /*nextl*/, Index /*chmark*/) {}
+  enum { ExpandMem = false };
+  Index m_jcol;
+  StorageIndex* m_marker;
+};
+
+
+template <typename Scalar, typename StorageIndex>
+template <typename Traits>
+void SparseLUImpl<Scalar,StorageIndex>::dfs_kernel(const StorageIndex jj, IndexVector& perm_r,
+                   Index& nseg, IndexVector& panel_lsub, IndexVector& segrep,
+                   Ref<IndexVector> repfnz_col, IndexVector& xprune, Ref<IndexVector> marker, IndexVector& parent,
+                   IndexVector& xplore, GlobalLU_t& glu,
+                   Index& nextl_col, Index krow, Traits& traits
+                  )
+{
+  
+  StorageIndex kmark = marker(krow);
+      
+  // For each unmarked krow of jj
+  marker(krow) = jj; 
+  StorageIndex kperm = perm_r(krow); 
+  if (kperm == emptyIdxLU ) {
+    // krow is in L : place it in structure of L(*, jj)
+    panel_lsub(nextl_col++) = StorageIndex(krow);  // krow is indexed into A
+    
+    traits.mem_expand(panel_lsub, nextl_col, kmark);
+  }
+  else 
+  {
+    // krow is in U : if its supernode-representative krep
+    // has been explored, update repfnz(*)
+    // krep = supernode representative of the current row
+    StorageIndex krep = glu.xsup(glu.supno(kperm)+1) - 1; 
+    // First nonzero element in the current column:
+    StorageIndex myfnz = repfnz_col(krep); 
+    
+    if (myfnz != emptyIdxLU )
+    {
+      // Representative visited before
+      if (myfnz > kperm ) repfnz_col(krep) = kperm; 
+      
+    }
+    else 
+    {
+      // Otherwise, perform dfs starting at krep
+      StorageIndex oldrep = emptyIdxLU; 
+      parent(krep) = oldrep; 
+      repfnz_col(krep) = kperm; 
+      StorageIndex xdfs =  glu.xlsub(krep); 
+      Index maxdfs = xprune(krep); 
+      
+      StorageIndex kpar;
+      do 
+      {
+        // For each unmarked kchild of krep
+        while (xdfs < maxdfs) 
+        {
+          StorageIndex kchild = glu.lsub(xdfs); 
+          xdfs++; 
+          StorageIndex chmark = marker(kchild); 
+          
+          if (chmark != jj ) 
+          {
+            marker(kchild) = jj; 
+            StorageIndex chperm = perm_r(kchild); 
+            
+            if (chperm == emptyIdxLU) 
+            {
+              // case kchild is in L: place it in L(*, j)
+              panel_lsub(nextl_col++) = kchild;
+              traits.mem_expand(panel_lsub, nextl_col, chmark);
+            }
+            else
+            {
+              // case kchild is in U :
+              // chrep = its supernode-rep. If its rep has been explored, 
+              // update its repfnz(*)
+              StorageIndex chrep = glu.xsup(glu.supno(chperm)+1) - 1; 
+              myfnz = repfnz_col(chrep); 
+              
+              if (myfnz != emptyIdxLU) 
+              { // Visited before 
+                if (myfnz > chperm) 
+                  repfnz_col(chrep) = chperm; 
+              }
+              else 
+              { // Cont. dfs at snode-rep of kchild
+                xplore(krep) = xdfs; 
+                oldrep = krep; 
+                krep = chrep; // Go deeper down G(L)
+                parent(krep) = oldrep; 
+                repfnz_col(krep) = chperm; 
+                xdfs = glu.xlsub(krep); 
+                maxdfs = xprune(krep); 
+                
+              } // end if myfnz != -1
+            } // end if chperm == -1 
+                
+          } // end if chmark !=jj
+        } // end while xdfs < maxdfs
+        
+        // krow has no more unexplored nbrs :
+        //    Place snode-rep krep in postorder DFS, if this 
+        //    segment is seen for the first time. (Note that 
+        //    "repfnz(krep)" may change later.)
+        //    Baktrack dfs to its parent
+        if(traits.update_segrep(krep,jj))
+        //if (marker1(krep) < jcol )
+        {
+          segrep(nseg) = krep; 
+          ++nseg; 
+          //marker1(krep) = jj; 
+        }
+        
+        kpar = parent(krep); // Pop recursion, mimic recursion 
+        if (kpar == emptyIdxLU) 
+          break; // dfs done 
+        krep = kpar; 
+        xdfs = xplore(krep); 
+        maxdfs = xprune(krep); 
+
+      } while (kpar != emptyIdxLU); // Do until empty stack 
+      
+    } // end if (myfnz = -1)
+
+  } // end if (kperm == -1)   
+}
+
+/**
+ * \brief Performs a symbolic factorization on a panel of columns [jcol, jcol+w)
+ * 
+ * A supernode representative is the last column of a supernode.
+ * The nonzeros in U[*,j] are segments that end at supernodes representatives
+ * 
+ * The routine returns a list of the supernodal representatives 
+ * in topological order of the dfs that generates them. This list is 
+ * a superset of the topological order of each individual column within 
+ * the panel.
+ * The location of the first nonzero in each supernodal segment 
+ * (supernodal entry location) is also returned. Each column has 
+ * a separate list for this purpose. 
+ * 
+ * Two markers arrays are used for dfs :
+ *    marker[i] == jj, if i was visited during dfs of current column jj;
+ *    marker1[i] >= jcol, if i was visited by earlier columns in this panel; 
+ * 
+ * \param[in] m number of rows in the matrix
+ * \param[in] w Panel size
+ * \param[in] jcol Starting  column of the panel
+ * \param[in] A Input matrix in column-major storage
+ * \param[in] perm_r Row permutation
+ * \param[out] nseg Number of U segments
+ * \param[out] dense Accumulate the column vectors of the panel
+ * \param[out] panel_lsub Subscripts of the row in the panel 
+ * \param[out] segrep Segment representative i.e first nonzero row of each segment
+ * \param[out] repfnz First nonzero location in each row
+ * \param[out] xprune The pruned elimination tree
+ * \param[out] marker work vector
+ * \param  parent The elimination tree
+ * \param xplore work vector
+ * \param glu The global data structure
+ * 
+ */
+
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::panel_dfs(const Index m, const Index w, const Index jcol, MatrixType& A, IndexVector& perm_r, Index& nseg, ScalarVector& dense, IndexVector& panel_lsub, IndexVector& segrep, IndexVector& repfnz, IndexVector& xprune, IndexVector& marker, IndexVector& parent, IndexVector& xplore, GlobalLU_t& glu)
+{
+  Index nextl_col; // Next available position in panel_lsub[*,jj] 
+  
+  // Initialize pointers 
+  VectorBlock<IndexVector> marker1(marker, m, m); 
+  nseg = 0; 
+  
+  panel_dfs_traits<IndexVector> traits(jcol, marker1.data());
+  
+  // For each column in the panel 
+  for (StorageIndex jj = StorageIndex(jcol); jj < jcol + w; jj++) 
+  {
+    nextl_col = (jj - jcol) * m; 
+    
+    VectorBlock<IndexVector> repfnz_col(repfnz, nextl_col, m); // First nonzero location in each row
+    VectorBlock<ScalarVector> dense_col(dense,nextl_col, m); // Accumulate a column vector here
+    
+    
+    // For each nnz in A[*, jj] do depth first search
+    for (typename MatrixType::InnerIterator it(A, jj); it; ++it)
+    {
+      Index krow = it.row(); 
+      dense_col(krow) = it.value();
+      
+      StorageIndex kmark = marker(krow); 
+      if (kmark == jj) 
+        continue; // krow visited before, go to the next nonzero
+      
+      dfs_kernel(jj, perm_r, nseg, panel_lsub, segrep, repfnz_col, xprune, marker, parent,
+                   xplore, glu, nextl_col, krow, traits);
+    }// end for nonzeros in column jj
+    
+  } // end for column jj
+}
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_PANEL_DFS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pivotL.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pivotL.h
new file mode 100644
index 0000000..a86dac9
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pivotL.h
@@ -0,0 +1,137 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of xpivotL.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_PIVOTL_H
+#define SPARSELU_PIVOTL_H
+
+namespace Eigen {
+namespace internal {
+  
+/**
+ * \brief Performs the numerical pivotin on the current column of L, and the CDIV operation.
+ * 
+ * Pivot policy :
+ * (1) Compute thresh = u * max_(i>=j) abs(A_ij);
+ * (2) IF user specifies pivot row k and abs(A_kj) >= thresh THEN
+ *           pivot row = k;
+ *       ELSE IF abs(A_jj) >= thresh THEN
+ *           pivot row = j;
+ *       ELSE
+ *           pivot row = m;
+ * 
+ *   Note: If you absolutely want to use a given pivot order, then set u=0.0.
+ * 
+ * \param jcol The current column of L
+ * \param diagpivotthresh diagonal pivoting threshold
+ * \param[in,out] perm_r Row permutation (threshold pivoting)
+ * \param[in] iperm_c column permutation - used to finf diagonal of Pc*A*Pc'
+ * \param[out] pivrow  The pivot row
+ * \param glu Global LU data
+ * \return 0 if success, i > 0 if U(i,i) is exactly zero 
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+Index SparseLUImpl<Scalar,StorageIndex>::pivotL(const Index jcol, const RealScalar& diagpivotthresh, IndexVector& perm_r, IndexVector& iperm_c, Index& pivrow, GlobalLU_t& glu)
+{
+  
+  Index fsupc = (glu.xsup)((glu.supno)(jcol)); // First column in the supernode containing the column jcol
+  Index nsupc = jcol - fsupc; // Number of columns in the supernode portion, excluding jcol; nsupc >=0
+  Index lptr = glu.xlsub(fsupc); // pointer to the starting location of the row subscripts for this supernode portion
+  Index nsupr = glu.xlsub(fsupc+1) - lptr; // Number of rows in the supernode
+  Index lda = glu.xlusup(fsupc+1) - glu.xlusup(fsupc); // leading dimension
+  Scalar* lu_sup_ptr = &(glu.lusup.data()[glu.xlusup(fsupc)]); // Start of the current supernode
+  Scalar* lu_col_ptr = &(glu.lusup.data()[glu.xlusup(jcol)]); // Start of jcol in the supernode
+  StorageIndex* lsub_ptr = &(glu.lsub.data()[lptr]); // Start of row indices of the supernode
+  
+  // Determine the largest abs numerical value for partial pivoting 
+  Index diagind = iperm_c(jcol); // diagonal index 
+  RealScalar pivmax(-1.0);
+  Index pivptr = nsupc; 
+  Index diag = emptyIdxLU; 
+  RealScalar rtemp;
+  Index isub, icol, itemp, k; 
+  for (isub = nsupc; isub < nsupr; ++isub) {
+    using std::abs;
+    rtemp = abs(lu_col_ptr[isub]);
+    if (rtemp > pivmax) {
+      pivmax = rtemp; 
+      pivptr = isub;
+    } 
+    if (lsub_ptr[isub] == diagind) diag = isub;
+  }
+  
+  // Test for singularity
+  if ( pivmax <= RealScalar(0.0) ) {
+    // if pivmax == -1, the column is structurally empty, otherwise it is only numerically zero
+    pivrow = pivmax < RealScalar(0.0) ? diagind : lsub_ptr[pivptr];
+    perm_r(pivrow) = StorageIndex(jcol);
+    return (jcol+1);
+  }
+  
+  RealScalar thresh = diagpivotthresh * pivmax; 
+  
+  // Choose appropriate pivotal element 
+  
+  {
+    // Test if the diagonal element can be used as a pivot (given the threshold value)
+    if (diag >= 0 ) 
+    {
+      // Diagonal element exists
+      using std::abs;
+      rtemp = abs(lu_col_ptr[diag]);
+      if (rtemp != RealScalar(0.0) && rtemp >= thresh) pivptr = diag;
+    }
+    pivrow = lsub_ptr[pivptr];
+  }
+  
+  // Record pivot row
+  perm_r(pivrow) = StorageIndex(jcol);
+  // Interchange row subscripts
+  if (pivptr != nsupc )
+  {
+    std::swap( lsub_ptr[pivptr], lsub_ptr[nsupc] );
+    // Interchange numerical values as well, for the two rows in the whole snode
+    // such that L is indexed the same way as A
+    for (icol = 0; icol <= nsupc; icol++)
+    {
+      itemp = pivptr + icol * lda; 
+      std::swap(lu_sup_ptr[itemp], lu_sup_ptr[nsupc + icol * lda]);
+    }
+  }
+  // cdiv operations
+  Scalar temp = Scalar(1.0) / lu_col_ptr[nsupc];
+  for (k = nsupc+1; k < nsupr; k++)
+    lu_col_ptr[k] *= temp; 
+  return 0;
+}
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_PIVOTL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pruneL.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pruneL.h
new file mode 100644
index 0000000..ad32fed
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_pruneL.h
@@ -0,0 +1,136 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* 
+ 
+ * NOTE: This file is the modified version of [s,d,c,z]pruneL.c file in SuperLU 
+ 
+ * -- SuperLU routine (version 2.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * November 15, 1997
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+#ifndef SPARSELU_PRUNEL_H
+#define SPARSELU_PRUNEL_H
+
+namespace Eigen {
+namespace internal {
+
+/**
+ * \brief Prunes the L-structure.
+ *
+ * It prunes the L-structure  of supernodes whose L-structure contains the current pivot row "pivrow"
+ * 
+ * 
+ * \param jcol The current column of L
+ * \param[in] perm_r Row permutation
+ * \param[out] pivrow  The pivot row
+ * \param nseg Number of segments
+ * \param segrep 
+ * \param repfnz
+ * \param[out] xprune 
+ * \param glu Global LU data
+ * 
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::pruneL(const Index jcol, const IndexVector& perm_r, const Index pivrow, const Index nseg,
+                                               const IndexVector& segrep, BlockIndexVector repfnz, IndexVector& xprune, GlobalLU_t& glu)
+{
+  // For each supernode-rep irep in U(*,j]
+  Index jsupno = glu.supno(jcol); 
+  Index i,irep,irep1; 
+  bool movnum, do_prune = false; 
+  Index kmin = 0, kmax = 0, minloc, maxloc,krow; 
+  for (i = 0; i < nseg; i++)
+  {
+    irep = segrep(i); 
+    irep1 = irep + 1; 
+    do_prune = false; 
+    
+    // Don't prune with a zero U-segment 
+    if (repfnz(irep) == emptyIdxLU) continue; 
+    
+    // If a snode overlaps with the next panel, then the U-segment
+    // is fragmented into two parts -- irep and irep1. We should let 
+    // pruning occur at the rep-column in irep1s snode. 
+    if (glu.supno(irep) == glu.supno(irep1) ) continue; // don't prune 
+    
+    // If it has not been pruned & it has a nonz in row L(pivrow,i)
+    if (glu.supno(irep) != jsupno )
+    {
+      if ( xprune (irep) >= glu.xlsub(irep1) )
+      {
+        kmin = glu.xlsub(irep);
+        kmax = glu.xlsub(irep1) - 1; 
+        for (krow = kmin; krow <= kmax; krow++)
+        {
+          if (glu.lsub(krow) == pivrow) 
+          {
+            do_prune = true; 
+            break; 
+          }
+        }
+      }
+      
+      if (do_prune) 
+      {
+        // do a quicksort-type partition
+        // movnum=true means that the num values have to be exchanged
+        movnum = false; 
+        if (irep == glu.xsup(glu.supno(irep)) ) // Snode of size 1 
+          movnum = true; 
+        
+        while (kmin <= kmax)
+        {
+          if (perm_r(glu.lsub(kmax)) == emptyIdxLU)
+            kmax--; 
+          else if ( perm_r(glu.lsub(kmin)) != emptyIdxLU)
+            kmin++;
+          else 
+          {
+            // kmin below pivrow (not yet pivoted), and kmax
+            // above pivrow: interchange the two suscripts
+            std::swap(glu.lsub(kmin), glu.lsub(kmax)); 
+            
+            // If the supernode has only one column, then we 
+            // only keep one set of subscripts. For any subscript
+            // intercnahge performed, similar interchange must be 
+            // done on the numerical values. 
+            if (movnum) 
+            {
+              minloc = glu.xlusup(irep) + ( kmin - glu.xlsub(irep) ); 
+              maxloc = glu.xlusup(irep) + ( kmax - glu.xlsub(irep) ); 
+              std::swap(glu.lusup(minloc), glu.lusup(maxloc)); 
+            }
+            kmin++;
+            kmax--;
+          }
+        } // end while 
+        
+        xprune(irep) = StorageIndex(kmin);  //Pruning 
+      } // end if do_prune 
+    } // end pruning 
+  } // End for each U-segment
+}
+
+} // end namespace internal
+} // end namespace Eigen
+
+#endif // SPARSELU_PRUNEL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_relax_snode.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_relax_snode.h
new file mode 100644
index 0000000..c408d01
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseLU/SparseLU_relax_snode.h
@@ -0,0 +1,83 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012 Désiré Nuentsa-Wakam <desire.nuentsa_wakam@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+/* This file is a modified version of heap_relax_snode.c file in SuperLU
+ * -- SuperLU routine (version 3.0) --
+ * Univ. of California Berkeley, Xerox Palo Alto Research Center,
+ * and Lawrence Berkeley National Lab.
+ * October 15, 2003
+ *
+ * Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+ * EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program for any
+ * purpose, provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is
+ * granted, provided the above notices are retained, and a notice that
+ * the code was modified is included with the above copyright notice.
+ */
+
+#ifndef SPARSELU_RELAX_SNODE_H
+#define SPARSELU_RELAX_SNODE_H
+
+namespace Eigen {
+
+namespace internal {
+ 
+/** 
+ * \brief Identify the initial relaxed supernodes
+ * 
+ * This routine is applied to a column elimination tree. 
+ * It assumes that the matrix has been reordered according to the postorder of the etree
+ * \param n  the number of columns
+ * \param et elimination tree 
+ * \param relax_columns Maximum number of columns allowed in a relaxed snode 
+ * \param descendants Number of descendants of each node in the etree
+ * \param relax_end last column in a supernode
+ */
+template <typename Scalar, typename StorageIndex>
+void SparseLUImpl<Scalar,StorageIndex>::relax_snode (const Index n, IndexVector& et, const Index relax_columns, IndexVector& descendants, IndexVector& relax_end)
+{
+  
+  // compute the number of descendants of each node in the etree
+  Index parent; 
+  relax_end.setConstant(emptyIdxLU);
+  descendants.setZero();
+  for (Index j = 0; j < n; j++) 
+  {
+    parent = et(j);
+    if (parent != n) // not the dummy root
+      descendants(parent) += descendants(j) + 1;
+  }
+  // Identify the relaxed supernodes by postorder traversal of the etree
+  Index snode_start; // beginning of a snode 
+  for (Index j = 0; j < n; )
+  {
+    parent = et(j);
+    snode_start = j; 
+    while ( parent != n && descendants(parent) < relax_columns ) 
+    {
+      j = parent; 
+      parent = et(j);
+    }
+    // Found a supernode in postordered etree, j is the last column 
+    relax_end(snode_start) = StorageIndex(j); // Record last column
+    j++;
+    // Search for a new leaf
+    while (descendants(j) != 0 && j < n) j++;
+  } // End postorder traversal of the etree
+  
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseQR/SparseQR.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseQR/SparseQR.h
new file mode 100644
index 0000000..2d4498b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SparseQR/SparseQR.h
@@ -0,0 +1,739 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2012-2013 Desire Nuentsa <desire.nuentsa_wakam@inria.fr>
+// Copyright (C) 2012-2014 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SPARSE_QR_H
+#define EIGEN_SPARSE_QR_H
+
+namespace Eigen {
+
+template<typename MatrixType, typename OrderingType> class SparseQR;
+template<typename SparseQRType> struct SparseQRMatrixQReturnType;
+template<typename SparseQRType> struct SparseQRMatrixQTransposeReturnType;
+template<typename SparseQRType, typename Derived> struct SparseQR_QProduct;
+namespace internal {
+  template <typename SparseQRType> struct traits<SparseQRMatrixQReturnType<SparseQRType> >
+  {
+    typedef typename SparseQRType::MatrixType ReturnType;
+    typedef typename ReturnType::StorageIndex StorageIndex;
+    typedef typename ReturnType::StorageKind StorageKind;
+    enum {
+      RowsAtCompileTime = Dynamic,
+      ColsAtCompileTime = Dynamic
+    };
+  };
+  template <typename SparseQRType> struct traits<SparseQRMatrixQTransposeReturnType<SparseQRType> >
+  {
+    typedef typename SparseQRType::MatrixType ReturnType;
+  };
+  template <typename SparseQRType, typename Derived> struct traits<SparseQR_QProduct<SparseQRType, Derived> >
+  {
+    typedef typename Derived::PlainObject ReturnType;
+  };
+} // End namespace internal
+
+/**
+  * \ingroup SparseQR_Module
+  * \class SparseQR
+  * \brief Sparse left-looking rank-revealing QR factorization
+  * 
+  * This class implements a left-looking rank-revealing QR decomposition 
+  * of sparse matrices. When a column has a norm less than a given tolerance
+  * it is implicitly permuted to the end. The QR factorization thus obtained is 
+  * given by A*P = Q*R where R is upper triangular or trapezoidal. 
+  * 
+  * P is the column permutation which is the product of the fill-reducing and the
+  * rank-revealing permutations. Use colsPermutation() to get it.
+  * 
+  * Q is the orthogonal matrix represented as products of Householder reflectors. 
+  * Use matrixQ() to get an expression and matrixQ().transpose() to get the transpose.
+  * You can then apply it to a vector.
+  * 
+  * R is the sparse triangular or trapezoidal matrix. The later occurs when A is rank-deficient.
+  * matrixR().topLeftCorner(rank(), rank()) always returns a triangular factor of full rank.
+  * 
+  * \tparam _MatrixType The type of the sparse matrix A, must be a column-major SparseMatrix<>
+  * \tparam _OrderingType The fill-reducing ordering method. See the \link OrderingMethods_Module 
+  *  OrderingMethods \endlink module for the list of built-in and external ordering methods.
+  * 
+  * \implsparsesolverconcept
+  *
+  * \warning The input sparse matrix A must be in compressed mode (see SparseMatrix::makeCompressed()).
+  * 
+  */
+template<typename _MatrixType, typename _OrderingType>
+class SparseQR : public SparseSolverBase<SparseQR<_MatrixType,_OrderingType> >
+{
+  protected:
+    typedef SparseSolverBase<SparseQR<_MatrixType,_OrderingType> > Base;
+    using Base::m_isInitialized;
+  public:
+    using Base::_solve_impl;
+    typedef _MatrixType MatrixType;
+    typedef _OrderingType OrderingType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef SparseMatrix<Scalar,ColMajor,StorageIndex> QRMatrixType;
+    typedef Matrix<StorageIndex, Dynamic, 1> IndexVector;
+    typedef Matrix<Scalar, Dynamic, 1> ScalarVector;
+    typedef PermutationMatrix<Dynamic, Dynamic, StorageIndex> PermutationType;
+
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+    
+  public:
+    SparseQR () :  m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false)
+    { }
+    
+    /** Construct a QR factorization of the matrix \a mat.
+      * 
+      * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()).
+      * 
+      * \sa compute()
+      */
+    explicit SparseQR(const MatrixType& mat) : m_analysisIsok(false), m_lastError(""), m_useDefaultThreshold(true),m_isQSorted(false),m_isEtreeOk(false)
+    {
+      compute(mat);
+    }
+    
+    /** Computes the QR factorization of the sparse matrix \a mat.
+      * 
+      * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()).
+      * 
+      * \sa analyzePattern(), factorize()
+      */
+    void compute(const MatrixType& mat)
+    {
+      analyzePattern(mat);
+      factorize(mat);
+    }
+    void analyzePattern(const MatrixType& mat);
+    void factorize(const MatrixType& mat);
+    
+    /** \returns the number of rows of the represented matrix. 
+      */
+    inline Index rows() const { return m_pmat.rows(); }
+    
+    /** \returns the number of columns of the represented matrix. 
+      */
+    inline Index cols() const { return m_pmat.cols();}
+    
+    /** \returns a const reference to the \b sparse upper triangular matrix R of the QR factorization.
+      * \warning The entries of the returned matrix are not sorted. This means that using it in algorithms
+      *          expecting sorted entries will fail. This include random coefficient accesses (SpaseMatrix::coeff()),
+      *          and coefficient-wise operations. Matrix products and triangular solves are fine though.
+      *
+      * To sort the entries, you can assign it to a row-major matrix, and if a column-major matrix
+      * is required, you can copy it again:
+      * \code
+      * SparseMatrix<double>          R  = qr.matrixR();  // column-major, not sorted!
+      * SparseMatrix<double,RowMajor> Rr = qr.matrixR();  // row-major, sorted
+      * SparseMatrix<double>          Rc = Rr;            // column-major, sorted
+      * \endcode
+      */
+    const QRMatrixType& matrixR() const { return m_R; }
+    
+    /** \returns the number of non linearly dependent columns as determined by the pivoting threshold.
+      *
+      * \sa setPivotThreshold()
+      */
+    Index rank() const
+    {
+      eigen_assert(m_isInitialized && "The factorization should be called first, use compute()");
+      return m_nonzeropivots; 
+    }
+    
+    /** \returns an expression of the matrix Q as products of sparse Householder reflectors.
+    * The common usage of this function is to apply it to a dense matrix or vector
+    * \code
+    * VectorXd B1, B2;
+    * // Initialize B1
+    * B2 = matrixQ() * B1;
+    * \endcode
+    *
+    * To get a plain SparseMatrix representation of Q:
+    * \code
+    * SparseMatrix<double> Q;
+    * Q = SparseQR<SparseMatrix<double> >(A).matrixQ();
+    * \endcode
+    * Internally, this call simply performs a sparse product between the matrix Q
+    * and a sparse identity matrix. However, due to the fact that the sparse
+    * reflectors are stored unsorted, two transpositions are needed to sort
+    * them before performing the product.
+    */
+    SparseQRMatrixQReturnType<SparseQR> matrixQ() const 
+    { return SparseQRMatrixQReturnType<SparseQR>(*this); }
+    
+    /** \returns a const reference to the column permutation P that was applied to A such that A*P = Q*R
+      * It is the combination of the fill-in reducing permutation and numerical column pivoting.
+      */
+    const PermutationType& colsPermutation() const
+    { 
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_outputPerm_c;
+    }
+    
+    /** \returns A string describing the type of error.
+      * This method is provided to ease debugging, not to handle errors.
+      */
+    std::string lastErrorMessage() const { return m_lastError; }
+    
+    /** \internal */
+    template<typename Rhs, typename Dest>
+    bool _solve_impl(const MatrixBase<Rhs> &B, MatrixBase<Dest> &dest) const
+    {
+      eigen_assert(m_isInitialized && "The factorization should be called first, use compute()");
+      eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix");
+
+      Index rank = this->rank();
+      
+      // Compute Q^T * b;
+      typename Dest::PlainObject y, b;
+      y = this->matrixQ().transpose() * B; 
+      b = y;
+      
+      // Solve with the triangular matrix R
+      y.resize((std::max<Index>)(cols(),y.rows()),y.cols());
+      y.topRows(rank) = this->matrixR().topLeftCorner(rank, rank).template triangularView<Upper>().solve(b.topRows(rank));
+      y.bottomRows(y.rows()-rank).setZero();
+      
+      // Apply the column permutation
+      if (m_perm_c.size())  dest = colsPermutation() * y.topRows(cols());
+      else                  dest = y.topRows(cols());
+      
+      m_info = Success;
+      return true;
+    }
+
+    /** Sets the threshold that is used to determine linearly dependent columns during the factorization.
+      *
+      * In practice, if during the factorization the norm of the column that has to be eliminated is below
+      * this threshold, then the entire column is treated as zero, and it is moved at the end.
+      */
+    void setPivotThreshold(const RealScalar& threshold)
+    {
+      m_useDefaultThreshold = false;
+      m_threshold = threshold;
+    }
+    
+    /** \returns the solution X of \f$ A X = B \f$ using the current decomposition of A.
+      *
+      * \sa compute()
+      */
+    template<typename Rhs>
+    inline const Solve<SparseQR, Rhs> solve(const MatrixBase<Rhs>& B) const 
+    {
+      eigen_assert(m_isInitialized && "The factorization should be called first, use compute()");
+      eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix");
+      return Solve<SparseQR, Rhs>(*this, B.derived());
+    }
+    template<typename Rhs>
+    inline const Solve<SparseQR, Rhs> solve(const SparseMatrixBase<Rhs>& B) const
+    {
+          eigen_assert(m_isInitialized && "The factorization should be called first, use compute()");
+          eigen_assert(this->rows() == B.rows() && "SparseQR::solve() : invalid number of rows in the right hand side matrix");
+          return Solve<SparseQR, Rhs>(*this, B.derived());
+    }
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was successful,
+      *          \c NumericalIssue if the QR factorization reports a numerical problem
+      *          \c InvalidInput if the input matrix is invalid
+      *
+      * \sa iparm()          
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+
+    /** \internal */
+    inline void _sort_matrix_Q()
+    {
+      if(this->m_isQSorted) return;
+      // The matrix Q is sorted during the transposition
+      SparseMatrix<Scalar, RowMajor, Index> mQrm(this->m_Q);
+      this->m_Q = mQrm;
+      this->m_isQSorted = true;
+    }
+
+    
+  protected:
+    bool m_analysisIsok;
+    bool m_factorizationIsok;
+    mutable ComputationInfo m_info;
+    std::string m_lastError;
+    QRMatrixType m_pmat;            // Temporary matrix
+    QRMatrixType m_R;               // The triangular factor matrix
+    QRMatrixType m_Q;               // The orthogonal reflectors
+    ScalarVector m_hcoeffs;         // The Householder coefficients
+    PermutationType m_perm_c;       // Fill-reducing  Column  permutation
+    PermutationType m_pivotperm;    // The permutation for rank revealing
+    PermutationType m_outputPerm_c; // The final column permutation
+    RealScalar m_threshold;         // Threshold to determine null Householder reflections
+    bool m_useDefaultThreshold;     // Use default threshold
+    Index m_nonzeropivots;          // Number of non zero pivots found
+    IndexVector m_etree;            // Column elimination tree
+    IndexVector m_firstRowElt;      // First element in each row
+    bool m_isQSorted;               // whether Q is sorted or not
+    bool m_isEtreeOk;               // whether the elimination tree match the initial input matrix
+    
+    template <typename, typename > friend struct SparseQR_QProduct;
+    
+};
+
+/** \brief Preprocessing step of a QR factorization 
+  * 
+  * \warning The matrix \a mat must be in compressed mode (see SparseMatrix::makeCompressed()).
+  * 
+  * In this step, the fill-reducing permutation is computed and applied to the columns of A
+  * and the column elimination tree is computed as well. Only the sparsity pattern of \a mat is exploited.
+  * 
+  * \note In this step it is assumed that there is no empty row in the matrix \a mat.
+  */
+template <typename MatrixType, typename OrderingType>
+void SparseQR<MatrixType,OrderingType>::analyzePattern(const MatrixType& mat)
+{
+  eigen_assert(mat.isCompressed() && "SparseQR requires a sparse matrix in compressed mode. Call .makeCompressed() before passing it to SparseQR");
+  // Copy to a column major matrix if the input is rowmajor
+  typename internal::conditional<MatrixType::IsRowMajor,QRMatrixType,const MatrixType&>::type matCpy(mat);
+  // Compute the column fill reducing ordering
+  OrderingType ord; 
+  ord(matCpy, m_perm_c); 
+  Index n = mat.cols();
+  Index m = mat.rows();
+  Index diagSize = (std::min)(m,n);
+  
+  if (!m_perm_c.size())
+  {
+    m_perm_c.resize(n);
+    m_perm_c.indices().setLinSpaced(n, 0,StorageIndex(n-1));
+  }
+  
+  // Compute the column elimination tree of the permuted matrix
+  m_outputPerm_c = m_perm_c.inverse();
+  internal::coletree(matCpy, m_etree, m_firstRowElt, m_outputPerm_c.indices().data());
+  m_isEtreeOk = true;
+  
+  m_R.resize(m, n);
+  m_Q.resize(m, diagSize);
+  
+  // Allocate space for nonzero elements : rough estimation
+  m_R.reserve(2*mat.nonZeros()); //FIXME Get a more accurate estimation through symbolic factorization with the etree
+  m_Q.reserve(2*mat.nonZeros());
+  m_hcoeffs.resize(diagSize);
+  m_analysisIsok = true;
+}
+
+/** \brief Performs the numerical QR factorization of the input matrix
+  * 
+  * The function SparseQR::analyzePattern(const MatrixType&) must have been called beforehand with
+  * a matrix having the same sparsity pattern than \a mat.
+  * 
+  * \param mat The sparse column-major matrix
+  */
+template <typename MatrixType, typename OrderingType>
+void SparseQR<MatrixType,OrderingType>::factorize(const MatrixType& mat)
+{
+  using std::abs;
+  
+  eigen_assert(m_analysisIsok && "analyzePattern() should be called before this step");
+  StorageIndex m = StorageIndex(mat.rows());
+  StorageIndex n = StorageIndex(mat.cols());
+  StorageIndex diagSize = (std::min)(m,n);
+  IndexVector mark((std::max)(m,n)); mark.setConstant(-1);  // Record the visited nodes
+  IndexVector Ridx(n), Qidx(m);                             // Store temporarily the row indexes for the current column of R and Q
+  Index nzcolR, nzcolQ;                                     // Number of nonzero for the current column of R and Q
+  ScalarVector tval(m);                                     // The dense vector used to compute the current column
+  RealScalar pivotThreshold = m_threshold;
+  
+  m_R.setZero();
+  m_Q.setZero();
+  m_pmat = mat;
+  if(!m_isEtreeOk)
+  {
+    m_outputPerm_c = m_perm_c.inverse();
+    internal::coletree(m_pmat, m_etree, m_firstRowElt, m_outputPerm_c.indices().data());
+    m_isEtreeOk = true;
+  }
+
+  m_pmat.uncompress(); // To have the innerNonZeroPtr allocated
+  
+  // Apply the fill-in reducing permutation lazily:
+  {
+    // If the input is row major, copy the original column indices,
+    // otherwise directly use the input matrix
+    // 
+    IndexVector originalOuterIndicesCpy;
+    const StorageIndex *originalOuterIndices = mat.outerIndexPtr();
+    if(MatrixType::IsRowMajor)
+    {
+      originalOuterIndicesCpy = IndexVector::Map(m_pmat.outerIndexPtr(),n+1);
+      originalOuterIndices = originalOuterIndicesCpy.data();
+    }
+    
+    for (int i = 0; i < n; i++)
+    {
+      Index p = m_perm_c.size() ? m_perm_c.indices()(i) : i;
+      m_pmat.outerIndexPtr()[p] = originalOuterIndices[i]; 
+      m_pmat.innerNonZeroPtr()[p] = originalOuterIndices[i+1] - originalOuterIndices[i]; 
+    }
+  }
+  
+  /* Compute the default threshold as in MatLab, see:
+   * Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing
+   * Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011, Page 8:3 
+   */
+  if(m_useDefaultThreshold) 
+  {
+    RealScalar max2Norm = 0.0;
+    for (int j = 0; j < n; j++) max2Norm = numext::maxi(max2Norm, m_pmat.col(j).norm());
+    if(max2Norm==RealScalar(0))
+      max2Norm = RealScalar(1);
+    pivotThreshold = 20 * (m + n) * max2Norm * NumTraits<RealScalar>::epsilon();
+  }
+  
+  // Initialize the numerical permutation
+  m_pivotperm.setIdentity(n);
+  
+  StorageIndex nonzeroCol = 0; // Record the number of valid pivots
+  m_Q.startVec(0);
+
+  // Left looking rank-revealing QR factorization: compute a column of R and Q at a time
+  for (StorageIndex col = 0; col < n; ++col)
+  {
+    mark.setConstant(-1);
+    m_R.startVec(col);
+    mark(nonzeroCol) = col;
+    Qidx(0) = nonzeroCol;
+    nzcolR = 0; nzcolQ = 1;
+    bool found_diag = nonzeroCol>=m;
+    tval.setZero(); 
+    
+    // Symbolic factorization: find the nonzero locations of the column k of the factors R and Q, i.e.,
+    // all the nodes (with indexes lower than rank) reachable through the column elimination tree (etree) rooted at node k.
+    // Note: if the diagonal entry does not exist, then its contribution must be explicitly added,
+    // thus the trick with found_diag that permits to do one more iteration on the diagonal element if this one has not been found.
+    for (typename QRMatrixType::InnerIterator itp(m_pmat, col); itp || !found_diag; ++itp)
+    {
+      StorageIndex curIdx = nonzeroCol;
+      if(itp) curIdx = StorageIndex(itp.row());
+      if(curIdx == nonzeroCol) found_diag = true;
+      
+      // Get the nonzeros indexes of the current column of R
+      StorageIndex st = m_firstRowElt(curIdx); // The traversal of the etree starts here
+      if (st < 0 )
+      {
+        m_lastError = "Empty row found during numerical factorization";
+        m_info = InvalidInput;
+        return;
+      }
+
+      // Traverse the etree 
+      Index bi = nzcolR;
+      for (; mark(st) != col; st = m_etree(st))
+      {
+        Ridx(nzcolR) = st;  // Add this row to the list,
+        mark(st) = col;     // and mark this row as visited
+        nzcolR++;
+      }
+
+      // Reverse the list to get the topological ordering
+      Index nt = nzcolR-bi;
+      for(Index i = 0; i < nt/2; i++) std::swap(Ridx(bi+i), Ridx(nzcolR-i-1));
+       
+      // Copy the current (curIdx,pcol) value of the input matrix
+      if(itp) tval(curIdx) = itp.value();
+      else    tval(curIdx) = Scalar(0);
+      
+      // Compute the pattern of Q(:,k)
+      if(curIdx > nonzeroCol && mark(curIdx) != col ) 
+      {
+        Qidx(nzcolQ) = curIdx;  // Add this row to the pattern of Q,
+        mark(curIdx) = col;     // and mark it as visited
+        nzcolQ++;
+      }
+    }
+
+    // Browse all the indexes of R(:,col) in reverse order
+    for (Index i = nzcolR-1; i >= 0; i--)
+    {
+      Index curIdx = Ridx(i);
+      
+      // Apply the curIdx-th householder vector to the current column (temporarily stored into tval)
+      Scalar tdot(0);
+      
+      // First compute q' * tval
+      tdot = m_Q.col(curIdx).dot(tval);
+
+      tdot *= m_hcoeffs(curIdx);
+      
+      // Then update tval = tval - q * tau
+      // FIXME: tval -= tdot * m_Q.col(curIdx) should amount to the same (need to check/add support for efficient "dense ?= sparse")
+      for (typename QRMatrixType::InnerIterator itq(m_Q, curIdx); itq; ++itq)
+        tval(itq.row()) -= itq.value() * tdot;
+
+      // Detect fill-in for the current column of Q
+      if(m_etree(Ridx(i)) == nonzeroCol)
+      {
+        for (typename QRMatrixType::InnerIterator itq(m_Q, curIdx); itq; ++itq)
+        {
+          StorageIndex iQ = StorageIndex(itq.row());
+          if (mark(iQ) != col)
+          {
+            Qidx(nzcolQ++) = iQ;  // Add this row to the pattern of Q,
+            mark(iQ) = col;       // and mark it as visited
+          }
+        }
+      }
+    } // End update current column
+    
+    Scalar tau = RealScalar(0);
+    RealScalar beta = 0;
+    
+    if(nonzeroCol < diagSize)
+    {
+      // Compute the Householder reflection that eliminate the current column
+      // FIXME this step should call the Householder module.
+      Scalar c0 = nzcolQ ? tval(Qidx(0)) : Scalar(0);
+      
+      // First, the squared norm of Q((col+1):m, col)
+      RealScalar sqrNorm = 0.;
+      for (Index itq = 1; itq < nzcolQ; ++itq) sqrNorm += numext::abs2(tval(Qidx(itq)));
+      if(sqrNorm == RealScalar(0) && numext::imag(c0) == RealScalar(0))
+      {
+        beta = numext::real(c0);
+        tval(Qidx(0)) = 1;
+      }
+      else
+      {
+        using std::sqrt;
+        beta = sqrt(numext::abs2(c0) + sqrNorm);
+        if(numext::real(c0) >= RealScalar(0))
+          beta = -beta;
+        tval(Qidx(0)) = 1;
+        for (Index itq = 1; itq < nzcolQ; ++itq)
+          tval(Qidx(itq)) /= (c0 - beta);
+        tau = numext::conj((beta-c0) / beta);
+          
+      }
+    }
+
+    // Insert values in R
+    for (Index  i = nzcolR-1; i >= 0; i--)
+    {
+      Index curIdx = Ridx(i);
+      if(curIdx < nonzeroCol) 
+      {
+        m_R.insertBackByOuterInnerUnordered(col, curIdx) = tval(curIdx);
+        tval(curIdx) = Scalar(0.);
+      }
+    }
+
+    if(nonzeroCol < diagSize && abs(beta) >= pivotThreshold)
+    {
+      m_R.insertBackByOuterInner(col, nonzeroCol) = beta;
+      // The householder coefficient
+      m_hcoeffs(nonzeroCol) = tau;
+      // Record the householder reflections
+      for (Index itq = 0; itq < nzcolQ; ++itq)
+      {
+        Index iQ = Qidx(itq);
+        m_Q.insertBackByOuterInnerUnordered(nonzeroCol,iQ) = tval(iQ);
+        tval(iQ) = Scalar(0.);
+      }
+      nonzeroCol++;
+      if(nonzeroCol<diagSize)
+        m_Q.startVec(nonzeroCol);
+    }
+    else
+    {
+      // Zero pivot found: move implicitly this column to the end
+      for (Index j = nonzeroCol; j < n-1; j++) 
+        std::swap(m_pivotperm.indices()(j), m_pivotperm.indices()[j+1]);
+      
+      // Recompute the column elimination tree
+      internal::coletree(m_pmat, m_etree, m_firstRowElt, m_pivotperm.indices().data());
+      m_isEtreeOk = false;
+    }
+  }
+  
+  m_hcoeffs.tail(diagSize-nonzeroCol).setZero();
+  
+  // Finalize the column pointers of the sparse matrices R and Q
+  m_Q.finalize();
+  m_Q.makeCompressed();
+  m_R.finalize();
+  m_R.makeCompressed();
+  m_isQSorted = false;
+
+  m_nonzeropivots = nonzeroCol;
+  
+  if(nonzeroCol<n)
+  {
+    // Permute the triangular factor to put the 'dead' columns to the end
+    QRMatrixType tempR(m_R);
+    m_R = tempR * m_pivotperm;
+    
+    // Update the column permutation
+    m_outputPerm_c = m_outputPerm_c * m_pivotperm;
+  }
+  
+  m_isInitialized = true; 
+  m_factorizationIsok = true;
+  m_info = Success;
+}
+
+template <typename SparseQRType, typename Derived>
+struct SparseQR_QProduct : ReturnByValue<SparseQR_QProduct<SparseQRType, Derived> >
+{
+  typedef typename SparseQRType::QRMatrixType MatrixType;
+  typedef typename SparseQRType::Scalar Scalar;
+  // Get the references 
+  SparseQR_QProduct(const SparseQRType& qr, const Derived& other, bool transpose) : 
+  m_qr(qr),m_other(other),m_transpose(transpose) {}
+  inline Index rows() const { return m_transpose ? m_qr.rows() : m_qr.cols(); }
+  inline Index cols() const { return m_other.cols(); }
+  
+  // Assign to a vector
+  template<typename DesType>
+  void evalTo(DesType& res) const
+  {
+    Index m = m_qr.rows();
+    Index n = m_qr.cols();
+    Index diagSize = (std::min)(m,n);
+    res = m_other;
+    if (m_transpose)
+    {
+      eigen_assert(m_qr.m_Q.rows() == m_other.rows() && "Non conforming object sizes");
+      //Compute res = Q' * other column by column
+      for(Index j = 0; j < res.cols(); j++){
+        for (Index k = 0; k < diagSize; k++)
+        {
+          Scalar tau = Scalar(0);
+          tau = m_qr.m_Q.col(k).dot(res.col(j));
+          if(tau==Scalar(0)) continue;
+          tau = tau * m_qr.m_hcoeffs(k);
+          res.col(j) -= tau * m_qr.m_Q.col(k);
+        }
+      }
+    }
+    else
+    {
+      eigen_assert(m_qr.m_Q.rows() == m_other.rows() && "Non conforming object sizes");
+      // Compute res = Q * other column by column
+      for(Index j = 0; j < res.cols(); j++)
+      {
+        for (Index k = diagSize-1; k >=0; k--)
+        {
+          Scalar tau = Scalar(0);
+          tau = m_qr.m_Q.col(k).dot(res.col(j));
+          if(tau==Scalar(0)) continue;
+          tau = tau * m_qr.m_hcoeffs(k);
+          res.col(j) -= tau * m_qr.m_Q.col(k);
+        }
+      }
+    }
+  }
+  
+  const SparseQRType& m_qr;
+  const Derived& m_other;
+  bool m_transpose;
+};
+
+template<typename SparseQRType>
+struct SparseQRMatrixQReturnType : public EigenBase<SparseQRMatrixQReturnType<SparseQRType> >
+{  
+  typedef typename SparseQRType::Scalar Scalar;
+  typedef Matrix<Scalar,Dynamic,Dynamic> DenseMatrix;
+  enum {
+    RowsAtCompileTime = Dynamic,
+    ColsAtCompileTime = Dynamic
+  };
+  explicit SparseQRMatrixQReturnType(const SparseQRType& qr) : m_qr(qr) {}
+  template<typename Derived>
+  SparseQR_QProduct<SparseQRType, Derived> operator*(const MatrixBase<Derived>& other)
+  {
+    return SparseQR_QProduct<SparseQRType,Derived>(m_qr,other.derived(),false);
+  }
+  SparseQRMatrixQTransposeReturnType<SparseQRType> adjoint() const
+  {
+    return SparseQRMatrixQTransposeReturnType<SparseQRType>(m_qr);
+  }
+  inline Index rows() const { return m_qr.rows(); }
+  inline Index cols() const { return (std::min)(m_qr.rows(),m_qr.cols()); }
+  // To use for operations with the transpose of Q
+  SparseQRMatrixQTransposeReturnType<SparseQRType> transpose() const
+  {
+    return SparseQRMatrixQTransposeReturnType<SparseQRType>(m_qr);
+  }
+  const SparseQRType& m_qr;
+};
+
+template<typename SparseQRType>
+struct SparseQRMatrixQTransposeReturnType
+{
+  explicit SparseQRMatrixQTransposeReturnType(const SparseQRType& qr) : m_qr(qr) {}
+  template<typename Derived>
+  SparseQR_QProduct<SparseQRType,Derived> operator*(const MatrixBase<Derived>& other)
+  {
+    return SparseQR_QProduct<SparseQRType,Derived>(m_qr,other.derived(), true);
+  }
+  const SparseQRType& m_qr;
+};
+
+namespace internal {
+  
+template<typename SparseQRType>
+struct evaluator_traits<SparseQRMatrixQReturnType<SparseQRType> >
+{
+  typedef typename SparseQRType::MatrixType MatrixType;
+  typedef typename storage_kind_to_evaluator_kind<typename MatrixType::StorageKind>::Kind Kind;
+  typedef SparseShape Shape;
+};
+
+template< typename DstXprType, typename SparseQRType>
+struct Assignment<DstXprType, SparseQRMatrixQReturnType<SparseQRType>, internal::assign_op<typename DstXprType::Scalar,typename DstXprType::Scalar>, Sparse2Sparse>
+{
+  typedef SparseQRMatrixQReturnType<SparseQRType> SrcXprType;
+  typedef typename DstXprType::Scalar Scalar;
+  typedef typename DstXprType::StorageIndex StorageIndex;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &/*func*/)
+  {
+    typename DstXprType::PlainObject idMat(src.m_qr.rows(), src.m_qr.rows());
+    idMat.setIdentity();
+    // Sort the sparse householder reflectors if needed
+    const_cast<SparseQRType *>(&src.m_qr)->_sort_matrix_Q();
+    dst = SparseQR_QProduct<SparseQRType, DstXprType>(src.m_qr, idMat, false);
+  }
+};
+
+template< typename DstXprType, typename SparseQRType>
+struct Assignment<DstXprType, SparseQRMatrixQReturnType<SparseQRType>, internal::assign_op<typename DstXprType::Scalar,typename DstXprType::Scalar>, Sparse2Dense>
+{
+  typedef SparseQRMatrixQReturnType<SparseQRType> SrcXprType;
+  typedef typename DstXprType::Scalar Scalar;
+  typedef typename DstXprType::StorageIndex StorageIndex;
+  static void run(DstXprType &dst, const SrcXprType &src, const internal::assign_op<Scalar,Scalar> &/*func*/)
+  {
+    dst = src.m_qr.matrixQ() * DstXprType::Identity(src.m_qr.rows(), src.m_qr.rows());
+  }
+};
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdDeque.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdDeque.h
new file mode 100644
index 0000000..cf1fedf
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdDeque.h
@@ -0,0 +1,126 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDDEQUE_H
+#define EIGEN_STDDEQUE_H
+
+#include "details.h"
+
+/**
+ * This section contains a convenience MACRO which allows an easy specialization of
+ * std::deque such that for data types with alignment issues the correct allocator
+ * is used automatically.
+ */
+#define EIGEN_DEFINE_STL_DEQUE_SPECIALIZATION(...) \
+namespace std \
+{ \
+  template<> \
+  class deque<__VA_ARGS__, std::allocator<__VA_ARGS__> >           \
+    : public deque<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > \
+  { \
+    typedef deque<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > deque_base; \
+  public: \
+    typedef __VA_ARGS__ value_type; \
+    typedef deque_base::allocator_type allocator_type; \
+    typedef deque_base::size_type size_type;  \
+    typedef deque_base::iterator iterator;  \
+    explicit deque(const allocator_type& a = allocator_type()) : deque_base(a) {}  \
+    template<typename InputIterator> \
+    deque(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : deque_base(first, last, a) {} \
+    deque(const deque& c) : deque_base(c) {}  \
+    explicit deque(size_type num, const value_type& val = value_type()) : deque_base(num, val) {} \
+    deque(iterator start, iterator end) : deque_base(start, end) {}  \
+    deque& operator=(const deque& x) {  \
+      deque_base::operator=(x);  \
+      return *this;  \
+    } \
+  }; \
+}
+
+// check whether we really need the std::deque specialization
+#if !EIGEN_HAS_CXX11_CONTAINERS && !(defined(_GLIBCXX_DEQUE) && (!EIGEN_GNUC_AT_LEAST(4,1))) /* Note that before gcc-4.1 we already have: std::deque::resize(size_type,const T&). */
+
+namespace std {
+
+#define EIGEN_STD_DEQUE_SPECIALIZATION_BODY \
+  public:  \
+    typedef T value_type; \
+    typedef typename deque_base::allocator_type allocator_type; \
+    typedef typename deque_base::size_type size_type;  \
+    typedef typename deque_base::iterator iterator;  \
+    typedef typename deque_base::const_iterator const_iterator;  \
+    explicit deque(const allocator_type& a = allocator_type()) : deque_base(a) {}  \
+    template<typename InputIterator> \
+    deque(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) \
+    : deque_base(first, last, a) {} \
+    deque(const deque& c) : deque_base(c) {}  \
+    explicit deque(size_type num, const value_type& val = value_type()) : deque_base(num, val) {} \
+    deque(iterator start, iterator end) : deque_base(start, end) {}  \
+    deque& operator=(const deque& x) {  \
+      deque_base::operator=(x);  \
+      return *this;  \
+    }
+
+  template<typename T>
+  class deque<T,EIGEN_ALIGNED_ALLOCATOR<T> >
+    : public deque<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                   Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> >
+{
+  typedef deque<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> > deque_base;
+  EIGEN_STD_DEQUE_SPECIALIZATION_BODY
+
+  void resize(size_type new_size)
+  { resize(new_size, T()); }
+
+#if defined(_DEQUE_)
+  // workaround MSVC std::deque implementation
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (deque_base::size() < new_size)
+      deque_base::_Insert_n(deque_base::end(), new_size - deque_base::size(), x);
+    else if (new_size < deque_base::size())
+      deque_base::erase(deque_base::begin() + new_size, deque_base::end());
+  }
+  void push_back(const value_type& x)
+  { deque_base::push_back(x); } 
+  void push_front(const value_type& x)
+  { deque_base::push_front(x); }
+  using deque_base::insert;  
+  iterator insert(const_iterator position, const value_type& x)
+  { return deque_base::insert(position,x); }
+  void insert(const_iterator position, size_type new_size, const value_type& x)
+  { deque_base::insert(position, new_size, x); }
+#elif defined(_GLIBCXX_DEQUE) && EIGEN_GNUC_AT_LEAST(4,2)
+  // workaround GCC std::deque implementation
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (new_size < deque_base::size())
+      deque_base::_M_erase_at_end(this->_M_impl._M_start + new_size);
+    else
+      deque_base::insert(deque_base::end(), new_size - deque_base::size(), x);
+  }
+#else
+  // either GCC 4.1 or non-GCC
+  // default implementation which should always work.
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (new_size < deque_base::size())
+      deque_base::erase(deque_base::begin() + new_size, deque_base::end());
+    else if (new_size > deque_base::size())
+      deque_base::insert(deque_base::end(), new_size - deque_base::size(), x);
+  }
+#endif
+  };
+}
+
+#endif // check whether specialization is actually required
+
+#endif // EIGEN_STDDEQUE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdList.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdList.h
new file mode 100644
index 0000000..e1eba49
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdList.h
@@ -0,0 +1,106 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDLIST_H
+#define EIGEN_STDLIST_H
+
+#include "details.h"
+
+/**
+ * This section contains a convenience MACRO which allows an easy specialization of
+ * std::list such that for data types with alignment issues the correct allocator
+ * is used automatically.
+ */
+#define EIGEN_DEFINE_STL_LIST_SPECIALIZATION(...) \
+namespace std \
+{ \
+  template<> \
+  class list<__VA_ARGS__, std::allocator<__VA_ARGS__> >           \
+    : public list<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > \
+  { \
+    typedef list<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > list_base; \
+  public: \
+    typedef __VA_ARGS__ value_type; \
+    typedef list_base::allocator_type allocator_type; \
+    typedef list_base::size_type size_type;  \
+    typedef list_base::iterator iterator;  \
+    explicit list(const allocator_type& a = allocator_type()) : list_base(a) {}  \
+    template<typename InputIterator> \
+    list(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : list_base(first, last, a) {} \
+    list(const list& c) : list_base(c) {}  \
+    explicit list(size_type num, const value_type& val = value_type()) : list_base(num, val) {} \
+    list(iterator start, iterator end) : list_base(start, end) {}  \
+    list& operator=(const list& x) {  \
+      list_base::operator=(x);  \
+      return *this;  \
+    } \
+  }; \
+}
+
+// check whether we really need the std::list specialization
+#if !EIGEN_HAS_CXX11_CONTAINERS && !(defined(_GLIBCXX_LIST) && (!EIGEN_GNUC_AT_LEAST(4,1))) /* Note that before gcc-4.1 we already have: std::list::resize(size_type,const T&). */
+
+namespace std
+{
+
+#define EIGEN_STD_LIST_SPECIALIZATION_BODY \
+  public:  \
+    typedef T value_type; \
+    typedef typename list_base::allocator_type allocator_type; \
+    typedef typename list_base::size_type size_type;  \
+    typedef typename list_base::iterator iterator;  \
+    typedef typename list_base::const_iterator const_iterator;  \
+    explicit list(const allocator_type& a = allocator_type()) : list_base(a) {}  \
+    template<typename InputIterator> \
+    list(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) \
+    : list_base(first, last, a) {} \
+    list(const list& c) : list_base(c) {}  \
+    explicit list(size_type num, const value_type& val = value_type()) : list_base(num, val) {} \
+    list(iterator start, iterator end) : list_base(start, end) {}  \
+    list& operator=(const list& x) {  \
+    list_base::operator=(x);  \
+    return *this; \
+  }
+
+  template<typename T>
+  class list<T,EIGEN_ALIGNED_ALLOCATOR<T> >
+    : public list<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                  Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> >
+  {
+    typedef list<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                 Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> > list_base;
+    EIGEN_STD_LIST_SPECIALIZATION_BODY
+
+    void resize(size_type new_size)
+    { resize(new_size, T()); }
+
+    void resize(size_type new_size, const value_type& x)
+    {
+      if (list_base::size() < new_size)
+        list_base::insert(list_base::end(), new_size - list_base::size(), x);
+      else
+        while (new_size < list_base::size()) list_base::pop_back();
+    }
+
+#if defined(_LIST_)
+    // workaround MSVC std::list implementation
+    void push_back(const value_type& x)
+    { list_base::push_back(x); } 
+    using list_base::insert;  
+    iterator insert(const_iterator position, const value_type& x)
+    { return list_base::insert(position,x); }
+    void insert(const_iterator position, size_type new_size, const value_type& x)
+    { list_base::insert(position, new_size, x); }
+#endif
+  };
+}
+
+#endif // check whether specialization is actually required
+
+#endif // EIGEN_STDLIST_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdVector.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdVector.h
new file mode 100644
index 0000000..ec22821
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/StdVector.h
@@ -0,0 +1,131 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STDVECTOR_H
+#define EIGEN_STDVECTOR_H
+
+#include "details.h"
+
+/**
+ * This section contains a convenience MACRO which allows an easy specialization of
+ * std::vector such that for data types with alignment issues the correct allocator
+ * is used automatically.
+ */
+#define EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION(...) \
+namespace std \
+{ \
+  template<> \
+  class vector<__VA_ARGS__, std::allocator<__VA_ARGS__> >  \
+    : public vector<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > \
+  { \
+    typedef vector<__VA_ARGS__, EIGEN_ALIGNED_ALLOCATOR<__VA_ARGS__> > vector_base; \
+  public: \
+    typedef __VA_ARGS__ value_type; \
+    typedef vector_base::allocator_type allocator_type; \
+    typedef vector_base::size_type size_type;  \
+    typedef vector_base::iterator iterator;  \
+    explicit vector(const allocator_type& a = allocator_type()) : vector_base(a) {}  \
+    template<typename InputIterator> \
+    vector(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) : vector_base(first, last, a) {} \
+    vector(const vector& c) : vector_base(c) {}  \
+    explicit vector(size_type num, const value_type& val = value_type()) : vector_base(num, val) {} \
+    vector(iterator start, iterator end) : vector_base(start, end) {}  \
+    vector& operator=(const vector& x) {  \
+      vector_base::operator=(x);  \
+      return *this;  \
+    } \
+  }; \
+}
+
+// Don't specialize if containers are implemented according to C++11
+#if !EIGEN_HAS_CXX11_CONTAINERS
+
+namespace std {
+
+#define EIGEN_STD_VECTOR_SPECIALIZATION_BODY \
+  public:  \
+    typedef T value_type; \
+    typedef typename vector_base::allocator_type allocator_type; \
+    typedef typename vector_base::size_type size_type;  \
+    typedef typename vector_base::iterator iterator;  \
+    typedef typename vector_base::const_iterator const_iterator;  \
+    explicit vector(const allocator_type& a = allocator_type()) : vector_base(a) {}  \
+    template<typename InputIterator> \
+    vector(InputIterator first, InputIterator last, const allocator_type& a = allocator_type()) \
+    : vector_base(first, last, a) {} \
+    vector(const vector& c) : vector_base(c) {}  \
+    explicit vector(size_type num, const value_type& val = value_type()) : vector_base(num, val) {} \
+    vector(iterator start, iterator end) : vector_base(start, end) {}  \
+    vector& operator=(const vector& x) {  \
+      vector_base::operator=(x);  \
+      return *this;  \
+    }
+
+  template<typename T>
+  class vector<T,EIGEN_ALIGNED_ALLOCATOR<T> >
+    : public vector<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                    Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> >
+{
+  typedef vector<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T),
+                 Eigen::aligned_allocator_indirection<EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T)> > vector_base;
+  EIGEN_STD_VECTOR_SPECIALIZATION_BODY
+
+  void resize(size_type new_size)
+  { resize(new_size, T()); }
+
+#if defined(_VECTOR_)
+  // workaround MSVC std::vector implementation
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (vector_base::size() < new_size)
+      vector_base::_Insert_n(vector_base::end(), new_size - vector_base::size(), x);
+    else if (new_size < vector_base::size())
+      vector_base::erase(vector_base::begin() + new_size, vector_base::end());
+  }
+  void push_back(const value_type& x)
+  { vector_base::push_back(x); } 
+  using vector_base::insert;  
+  iterator insert(const_iterator position, const value_type& x)
+  { return vector_base::insert(position,x); }
+  void insert(const_iterator position, size_type new_size, const value_type& x)
+  { vector_base::insert(position, new_size, x); }
+#elif defined(_GLIBCXX_VECTOR) && (!(EIGEN_GNUC_AT_LEAST(4,1)))
+  /* Note that before gcc-4.1 we already have: std::vector::resize(size_type,const T&).
+   * However, this specialization is still needed to make the above EIGEN_DEFINE_STL_VECTOR_SPECIALIZATION trick to work. */
+  void resize(size_type new_size, const value_type& x)
+  {
+    vector_base::resize(new_size,x);
+  }
+#elif defined(_GLIBCXX_VECTOR) && EIGEN_GNUC_AT_LEAST(4,2)
+  // workaround GCC std::vector implementation
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (new_size < vector_base::size())
+      vector_base::_M_erase_at_end(this->_M_impl._M_start + new_size);
+    else
+      vector_base::insert(vector_base::end(), new_size - vector_base::size(), x);
+  }
+#else
+  // either GCC 4.1 or non-GCC
+  // default implementation which should always work.
+  void resize(size_type new_size, const value_type& x)
+  {
+    if (new_size < vector_base::size())
+      vector_base::erase(vector_base::begin() + new_size, vector_base::end());
+    else if (new_size > vector_base::size())
+      vector_base::insert(vector_base::end(), new_size - vector_base::size(), x);
+  }
+#endif
+  };
+}
+#endif // !EIGEN_HAS_CXX11_CONTAINERS
+
+
+#endif // EIGEN_STDVECTOR_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/details.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/details.h
new file mode 100644
index 0000000..2cfd13e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/StlSupport/details.h
@@ -0,0 +1,84 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2009 Hauke Heibel <hauke.heibel@googlemail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_STL_DETAILS_H
+#define EIGEN_STL_DETAILS_H
+
+#ifndef EIGEN_ALIGNED_ALLOCATOR
+  #define EIGEN_ALIGNED_ALLOCATOR Eigen::aligned_allocator
+#endif
+
+namespace Eigen {
+
+  // This one is needed to prevent reimplementing the whole std::vector.
+  template <class T>
+  class aligned_allocator_indirection : public EIGEN_ALIGNED_ALLOCATOR<T>
+  {
+  public:
+    typedef std::size_t     size_type;
+    typedef std::ptrdiff_t  difference_type;
+    typedef T*              pointer;
+    typedef const T*        const_pointer;
+    typedef T&              reference;
+    typedef const T&        const_reference;
+    typedef T               value_type;
+
+    template<class U>
+    struct rebind
+    {
+      typedef aligned_allocator_indirection<U> other;
+    };
+
+    aligned_allocator_indirection() {}
+    aligned_allocator_indirection(const aligned_allocator_indirection& ) : EIGEN_ALIGNED_ALLOCATOR<T>() {}
+    aligned_allocator_indirection(const EIGEN_ALIGNED_ALLOCATOR<T>& ) {}
+    template<class U>
+    aligned_allocator_indirection(const aligned_allocator_indirection<U>& ) {}
+    template<class U>
+    aligned_allocator_indirection(const EIGEN_ALIGNED_ALLOCATOR<U>& ) {}
+    ~aligned_allocator_indirection() {}
+  };
+
+#if EIGEN_COMP_MSVC
+
+  // sometimes, MSVC detects, at compile time, that the argument x
+  // in std::vector::resize(size_t s,T x) won't be aligned and generate an error
+  // even if this function is never called. Whence this little wrapper.
+#define EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T) \
+  typename Eigen::internal::conditional< \
+    Eigen::internal::is_arithmetic<T>::value, \
+    T, \
+    Eigen::internal::workaround_msvc_stl_support<T> \
+  >::type
+
+  namespace internal {
+  template<typename T> struct workaround_msvc_stl_support : public T
+  {
+    inline workaround_msvc_stl_support() : T() {}
+    inline workaround_msvc_stl_support(const T& other) : T(other) {}
+    inline operator T& () { return *static_cast<T*>(this); }
+    inline operator const T& () const { return *static_cast<const T*>(this); }
+    template<typename OtherT>
+    inline T& operator=(const OtherT& other)
+    { T::operator=(other); return *this; }
+    inline workaround_msvc_stl_support& operator=(const workaround_msvc_stl_support& other)
+    { T::operator=(other); return *this; }
+  };
+  }
+
+#else
+
+#define EIGEN_WORKAROUND_MSVC_STL_SUPPORT(T) T
+
+#endif
+
+}
+
+#endif // EIGEN_STL_DETAILS_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SuperLUSupport/SuperLUSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SuperLUSupport/SuperLUSupport.h
new file mode 100644
index 0000000..50a69f3
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/SuperLUSupport/SuperLUSupport.h
@@ -0,0 +1,1027 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2015 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_SUPERLUSUPPORT_H
+#define EIGEN_SUPERLUSUPPORT_H
+
+namespace Eigen {
+
+#if defined(SUPERLU_MAJOR_VERSION) && (SUPERLU_MAJOR_VERSION >= 5)
+#define DECL_GSSVX(PREFIX,FLOATTYPE,KEYTYPE)		\
+    extern "C" {                                                                                          \
+      extern void PREFIX##gssvx(superlu_options_t *, SuperMatrix *, int *, int *, int *,                  \
+                                char *, FLOATTYPE *, FLOATTYPE *, SuperMatrix *, SuperMatrix *,           \
+                                void *, int, SuperMatrix *, SuperMatrix *,                                \
+                                FLOATTYPE *, FLOATTYPE *, FLOATTYPE *, FLOATTYPE *,                       \
+                                GlobalLU_t *, mem_usage_t *, SuperLUStat_t *, int *);                     \
+    }                                                                                                     \
+    inline float SuperLU_gssvx(superlu_options_t *options, SuperMatrix *A,                                \
+         int *perm_c, int *perm_r, int *etree, char *equed,                                               \
+         FLOATTYPE *R, FLOATTYPE *C, SuperMatrix *L,                                                      \
+         SuperMatrix *U, void *work, int lwork,                                                           \
+         SuperMatrix *B, SuperMatrix *X,                                                                  \
+         FLOATTYPE *recip_pivot_growth,                                                                   \
+         FLOATTYPE *rcond, FLOATTYPE *ferr, FLOATTYPE *berr,                                              \
+         SuperLUStat_t *stats, int *info, KEYTYPE) {                                                      \
+    mem_usage_t mem_usage;                                                                                \
+    GlobalLU_t gLU;                                                                                       \
+    PREFIX##gssvx(options, A, perm_c, perm_r, etree, equed, R, C, L,                                      \
+         U, work, lwork, B, X, recip_pivot_growth, rcond,                                                 \
+         ferr, berr, &gLU, &mem_usage, stats, info);                                                      \
+    return mem_usage.for_lu; /* bytes used by the factor storage */                                       \
+  }
+#else // version < 5.0
+#define DECL_GSSVX(PREFIX,FLOATTYPE,KEYTYPE)		\
+    extern "C" {                                                                                          \
+      extern void PREFIX##gssvx(superlu_options_t *, SuperMatrix *, int *, int *, int *,                  \
+                                char *, FLOATTYPE *, FLOATTYPE *, SuperMatrix *, SuperMatrix *,           \
+                                void *, int, SuperMatrix *, SuperMatrix *,                                \
+                                FLOATTYPE *, FLOATTYPE *, FLOATTYPE *, FLOATTYPE *,                       \
+                                mem_usage_t *, SuperLUStat_t *, int *);                                   \
+    }                                                                                                     \
+    inline float SuperLU_gssvx(superlu_options_t *options, SuperMatrix *A,                                \
+         int *perm_c, int *perm_r, int *etree, char *equed,                                               \
+         FLOATTYPE *R, FLOATTYPE *C, SuperMatrix *L,                                                      \
+         SuperMatrix *U, void *work, int lwork,                                                           \
+         SuperMatrix *B, SuperMatrix *X,                                                                  \
+         FLOATTYPE *recip_pivot_growth,                                                                   \
+         FLOATTYPE *rcond, FLOATTYPE *ferr, FLOATTYPE *berr,                                              \
+         SuperLUStat_t *stats, int *info, KEYTYPE) {                                                      \
+    mem_usage_t mem_usage;                                                                                \
+    PREFIX##gssvx(options, A, perm_c, perm_r, etree, equed, R, C, L,                                      \
+         U, work, lwork, B, X, recip_pivot_growth, rcond,                                                 \
+         ferr, berr, &mem_usage, stats, info);                                                            \
+    return mem_usage.for_lu; /* bytes used by the factor storage */                                       \
+  }
+#endif
+
+DECL_GSSVX(s,float,float)
+DECL_GSSVX(c,float,std::complex<float>)
+DECL_GSSVX(d,double,double)
+DECL_GSSVX(z,double,std::complex<double>)
+
+#ifdef MILU_ALPHA
+#define EIGEN_SUPERLU_HAS_ILU
+#endif
+
+#ifdef EIGEN_SUPERLU_HAS_ILU
+
+// similarly for the incomplete factorization using gsisx
+#define DECL_GSISX(PREFIX,FLOATTYPE,KEYTYPE)                                                    \
+    extern "C" {                                                                                \
+      extern void PREFIX##gsisx(superlu_options_t *, SuperMatrix *, int *, int *, int *,        \
+                         char *, FLOATTYPE *, FLOATTYPE *, SuperMatrix *, SuperMatrix *,        \
+                         void *, int, SuperMatrix *, SuperMatrix *, FLOATTYPE *, FLOATTYPE *,   \
+                         mem_usage_t *, SuperLUStat_t *, int *);                        \
+    }                                                                                           \
+    inline float SuperLU_gsisx(superlu_options_t *options, SuperMatrix *A,                      \
+         int *perm_c, int *perm_r, int *etree, char *equed,                                     \
+         FLOATTYPE *R, FLOATTYPE *C, SuperMatrix *L,                                            \
+         SuperMatrix *U, void *work, int lwork,                                                 \
+         SuperMatrix *B, SuperMatrix *X,                                                        \
+         FLOATTYPE *recip_pivot_growth,                                                         \
+         FLOATTYPE *rcond,                                                                      \
+         SuperLUStat_t *stats, int *info, KEYTYPE) {                                            \
+    mem_usage_t mem_usage;                                                              \
+    PREFIX##gsisx(options, A, perm_c, perm_r, etree, equed, R, C, L,                            \
+         U, work, lwork, B, X, recip_pivot_growth, rcond,                                       \
+         &mem_usage, stats, info);                                                              \
+    return mem_usage.for_lu; /* bytes used by the factor storage */                             \
+  }
+
+DECL_GSISX(s,float,float)
+DECL_GSISX(c,float,std::complex<float>)
+DECL_GSISX(d,double,double)
+DECL_GSISX(z,double,std::complex<double>)
+
+#endif
+
+template<typename MatrixType>
+struct SluMatrixMapHelper;
+
+/** \internal
+  *
+  * A wrapper class for SuperLU matrices. It supports only compressed sparse matrices
+  * and dense matrices. Supernodal and other fancy format are not supported by this wrapper.
+  *
+  * This wrapper class mainly aims to avoids the need of dynamic allocation of the storage structure.
+  */
+struct SluMatrix : SuperMatrix
+{
+  SluMatrix()
+  {
+    Store = &storage;
+  }
+
+  SluMatrix(const SluMatrix& other)
+    : SuperMatrix(other)
+  {
+    Store = &storage;
+    storage = other.storage;
+  }
+
+  SluMatrix& operator=(const SluMatrix& other)
+  {
+    SuperMatrix::operator=(static_cast<const SuperMatrix&>(other));
+    Store = &storage;
+    storage = other.storage;
+    return *this;
+  }
+
+  struct
+  {
+    union {int nnz;int lda;};
+    void *values;
+    int *innerInd;
+    int *outerInd;
+  } storage;
+
+  void setStorageType(Stype_t t)
+  {
+    Stype = t;
+    if (t==SLU_NC || t==SLU_NR || t==SLU_DN)
+      Store = &storage;
+    else
+    {
+      eigen_assert(false && "storage type not supported");
+      Store = 0;
+    }
+  }
+
+  template<typename Scalar>
+  void setScalarType()
+  {
+    if (internal::is_same<Scalar,float>::value)
+      Dtype = SLU_S;
+    else if (internal::is_same<Scalar,double>::value)
+      Dtype = SLU_D;
+    else if (internal::is_same<Scalar,std::complex<float> >::value)
+      Dtype = SLU_C;
+    else if (internal::is_same<Scalar,std::complex<double> >::value)
+      Dtype = SLU_Z;
+    else
+    {
+      eigen_assert(false && "Scalar type not supported by SuperLU");
+    }
+  }
+
+  template<typename MatrixType>
+  static SluMatrix Map(MatrixBase<MatrixType>& _mat)
+  {
+    MatrixType& mat(_mat.derived());
+    eigen_assert( ((MatrixType::Flags&RowMajorBit)!=RowMajorBit) && "row-major dense matrices are not supported by SuperLU");
+    SluMatrix res;
+    res.setStorageType(SLU_DN);
+    res.setScalarType<typename MatrixType::Scalar>();
+    res.Mtype     = SLU_GE;
+
+    res.nrow      = internal::convert_index<int>(mat.rows());
+    res.ncol      = internal::convert_index<int>(mat.cols());
+
+    res.storage.lda       = internal::convert_index<int>(MatrixType::IsVectorAtCompileTime ? mat.size() : mat.outerStride());
+    res.storage.values    = (void*)(mat.data());
+    return res;
+  }
+
+  template<typename MatrixType>
+  static SluMatrix Map(SparseMatrixBase<MatrixType>& a_mat)
+  {
+    MatrixType &mat(a_mat.derived());
+    SluMatrix res;
+    if ((MatrixType::Flags&RowMajorBit)==RowMajorBit)
+    {
+      res.setStorageType(SLU_NR);
+      res.nrow      = internal::convert_index<int>(mat.cols());
+      res.ncol      = internal::convert_index<int>(mat.rows());
+    }
+    else
+    {
+      res.setStorageType(SLU_NC);
+      res.nrow      = internal::convert_index<int>(mat.rows());
+      res.ncol      = internal::convert_index<int>(mat.cols());
+    }
+
+    res.Mtype       = SLU_GE;
+
+    res.storage.nnz       = internal::convert_index<int>(mat.nonZeros());
+    res.storage.values    = mat.valuePtr();
+    res.storage.innerInd  = mat.innerIndexPtr();
+    res.storage.outerInd  = mat.outerIndexPtr();
+
+    res.setScalarType<typename MatrixType::Scalar>();
+
+    // FIXME the following is not very accurate
+    if (MatrixType::Flags & Upper)
+      res.Mtype = SLU_TRU;
+    if (MatrixType::Flags & Lower)
+      res.Mtype = SLU_TRL;
+
+    eigen_assert(((MatrixType::Flags & SelfAdjoint)==0) && "SelfAdjoint matrix shape not supported by SuperLU");
+
+    return res;
+  }
+};
+
+template<typename Scalar, int Rows, int Cols, int Options, int MRows, int MCols>
+struct SluMatrixMapHelper<Matrix<Scalar,Rows,Cols,Options,MRows,MCols> >
+{
+  typedef Matrix<Scalar,Rows,Cols,Options,MRows,MCols> MatrixType;
+  static void run(MatrixType& mat, SluMatrix& res)
+  {
+    eigen_assert( ((Options&RowMajor)!=RowMajor) && "row-major dense matrices is not supported by SuperLU");
+    res.setStorageType(SLU_DN);
+    res.setScalarType<Scalar>();
+    res.Mtype     = SLU_GE;
+
+    res.nrow      = mat.rows();
+    res.ncol      = mat.cols();
+
+    res.storage.lda       = mat.outerStride();
+    res.storage.values    = mat.data();
+  }
+};
+
+template<typename Derived>
+struct SluMatrixMapHelper<SparseMatrixBase<Derived> >
+{
+  typedef Derived MatrixType;
+  static void run(MatrixType& mat, SluMatrix& res)
+  {
+    if ((MatrixType::Flags&RowMajorBit)==RowMajorBit)
+    {
+      res.setStorageType(SLU_NR);
+      res.nrow      = mat.cols();
+      res.ncol      = mat.rows();
+    }
+    else
+    {
+      res.setStorageType(SLU_NC);
+      res.nrow      = mat.rows();
+      res.ncol      = mat.cols();
+    }
+
+    res.Mtype       = SLU_GE;
+
+    res.storage.nnz       = mat.nonZeros();
+    res.storage.values    = mat.valuePtr();
+    res.storage.innerInd  = mat.innerIndexPtr();
+    res.storage.outerInd  = mat.outerIndexPtr();
+
+    res.setScalarType<typename MatrixType::Scalar>();
+
+    // FIXME the following is not very accurate
+    if (MatrixType::Flags & Upper)
+      res.Mtype = SLU_TRU;
+    if (MatrixType::Flags & Lower)
+      res.Mtype = SLU_TRL;
+
+    eigen_assert(((MatrixType::Flags & SelfAdjoint)==0) && "SelfAdjoint matrix shape not supported by SuperLU");
+  }
+};
+
+namespace internal {
+
+template<typename MatrixType>
+SluMatrix asSluMatrix(MatrixType& mat)
+{
+  return SluMatrix::Map(mat);
+}
+
+/** View a Super LU matrix as an Eigen expression */
+template<typename Scalar, int Flags, typename Index>
+MappedSparseMatrix<Scalar,Flags,Index> map_superlu(SluMatrix& sluMat)
+{
+  eigen_assert((Flags&RowMajor)==RowMajor && sluMat.Stype == SLU_NR
+         || (Flags&ColMajor)==ColMajor && sluMat.Stype == SLU_NC);
+
+  Index outerSize = (Flags&RowMajor)==RowMajor ? sluMat.ncol : sluMat.nrow;
+
+  return MappedSparseMatrix<Scalar,Flags,Index>(
+    sluMat.nrow, sluMat.ncol, sluMat.storage.outerInd[outerSize],
+    sluMat.storage.outerInd, sluMat.storage.innerInd, reinterpret_cast<Scalar*>(sluMat.storage.values) );
+}
+
+} // end namespace internal
+
+/** \ingroup SuperLUSupport_Module
+  * \class SuperLUBase
+  * \brief The base class for the direct and incomplete LU factorization of SuperLU
+  */
+template<typename _MatrixType, typename Derived>
+class SuperLUBase : public SparseSolverBase<Derived>
+{
+  protected:
+    typedef SparseSolverBase<Derived> Base;
+    using Base::derived;
+    using Base::m_isInitialized;
+  public:
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef Matrix<int, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
+    typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;    
+    typedef Map<PermutationMatrix<Dynamic,Dynamic,int> > PermutationMap;
+    typedef SparseMatrix<Scalar> LUMatrixType;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+  public:
+
+    SuperLUBase() {}
+
+    ~SuperLUBase()
+    {
+      clearFactors();
+    }
+    
+    inline Index rows() const { return m_matrix.rows(); }
+    inline Index cols() const { return m_matrix.cols(); }
+    
+    /** \returns a reference to the Super LU option object to configure the  Super LU algorithms. */
+    inline superlu_options_t& options() { return m_sluOptions; }
+    
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+    /** Computes the sparse Cholesky decomposition of \a matrix */
+    void compute(const MatrixType& matrix)
+    {
+      derived().analyzePattern(matrix);
+      derived().factorize(matrix);
+    }
+
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      * 
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& /*matrix*/)
+    {
+      m_isInitialized = true;
+      m_info = Success;
+      m_analysisIsOk = true;
+      m_factorizationIsOk = false;
+    }
+    
+    template<typename Stream>
+    void dumpMemory(Stream& /*s*/)
+    {}
+    
+  protected:
+    
+    void initFactorization(const MatrixType& a)
+    {
+      set_default_options(&this->m_sluOptions);
+      
+      const Index size = a.rows();
+      m_matrix = a;
+
+      m_sluA = internal::asSluMatrix(m_matrix);
+      clearFactors();
+
+      m_p.resize(size);
+      m_q.resize(size);
+      m_sluRscale.resize(size);
+      m_sluCscale.resize(size);
+      m_sluEtree.resize(size);
+
+      // set empty B and X
+      m_sluB.setStorageType(SLU_DN);
+      m_sluB.setScalarType<Scalar>();
+      m_sluB.Mtype          = SLU_GE;
+      m_sluB.storage.values = 0;
+      m_sluB.nrow           = 0;
+      m_sluB.ncol           = 0;
+      m_sluB.storage.lda    = internal::convert_index<int>(size);
+      m_sluX                = m_sluB;
+      
+      m_extractedDataAreDirty = true;
+    }
+    
+    void init()
+    {
+      m_info = InvalidInput;
+      m_isInitialized = false;
+      m_sluL.Store = 0;
+      m_sluU.Store = 0;
+    }
+    
+    void extractData() const;
+
+    void clearFactors()
+    {
+      if(m_sluL.Store)
+        Destroy_SuperNode_Matrix(&m_sluL);
+      if(m_sluU.Store)
+        Destroy_CompCol_Matrix(&m_sluU);
+
+      m_sluL.Store = 0;
+      m_sluU.Store = 0;
+
+      memset(&m_sluL,0,sizeof m_sluL);
+      memset(&m_sluU,0,sizeof m_sluU);
+    }
+
+    // cached data to reduce reallocation, etc.
+    mutable LUMatrixType m_l;
+    mutable LUMatrixType m_u;
+    mutable IntColVectorType m_p;
+    mutable IntRowVectorType m_q;
+
+    mutable LUMatrixType m_matrix;  // copy of the factorized matrix
+    mutable SluMatrix m_sluA;
+    mutable SuperMatrix m_sluL, m_sluU;
+    mutable SluMatrix m_sluB, m_sluX;
+    mutable SuperLUStat_t m_sluStat;
+    mutable superlu_options_t m_sluOptions;
+    mutable std::vector<int> m_sluEtree;
+    mutable Matrix<RealScalar,Dynamic,1> m_sluRscale, m_sluCscale;
+    mutable Matrix<RealScalar,Dynamic,1> m_sluFerr, m_sluBerr;
+    mutable char m_sluEqued;
+
+    mutable ComputationInfo m_info;
+    int m_factorizationIsOk;
+    int m_analysisIsOk;
+    mutable bool m_extractedDataAreDirty;
+    
+  private:
+    SuperLUBase(SuperLUBase& ) { }
+};
+
+
+/** \ingroup SuperLUSupport_Module
+  * \class SuperLU
+  * \brief A sparse direct LU factorization and solver based on the SuperLU library
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a direct LU factorization
+  * using the SuperLU library. The sparse matrix A must be squared and invertible. The vectors or matrices
+  * X and B can be either dense or sparse.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  *
+  * \warning This class is only for the 4.x versions of SuperLU. The 3.x and 5.x versions are not supported.
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  */
+template<typename _MatrixType>
+class SuperLU : public SuperLUBase<_MatrixType,SuperLU<_MatrixType> >
+{
+  public:
+    typedef SuperLUBase<_MatrixType,SuperLU> Base;
+    typedef _MatrixType MatrixType;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::RealScalar RealScalar;
+    typedef typename Base::StorageIndex StorageIndex;
+    typedef typename Base::IntRowVectorType IntRowVectorType;
+    typedef typename Base::IntColVectorType IntColVectorType;   
+    typedef typename Base::PermutationMap PermutationMap;
+    typedef typename Base::LUMatrixType LUMatrixType;
+    typedef TriangularView<LUMatrixType, Lower|UnitDiag>  LMatrixType;
+    typedef TriangularView<LUMatrixType,  Upper>          UMatrixType;
+
+  public:
+    using Base::_solve_impl;
+
+    SuperLU() : Base() { init(); }
+
+    explicit SuperLU(const MatrixType& matrix) : Base()
+    {
+      init();
+      Base::compute(matrix);
+    }
+
+    ~SuperLU()
+    {
+    }
+    
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      * 
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    {
+      m_info = InvalidInput;
+      m_isInitialized = false;
+      Base::analyzePattern(matrix);
+    }
+    
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& matrix);
+    
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const;
+    
+    inline const LMatrixType& matrixL() const
+    {
+      if (m_extractedDataAreDirty) this->extractData();
+      return m_l;
+    }
+
+    inline const UMatrixType& matrixU() const
+    {
+      if (m_extractedDataAreDirty) this->extractData();
+      return m_u;
+    }
+
+    inline const IntColVectorType& permutationP() const
+    {
+      if (m_extractedDataAreDirty) this->extractData();
+      return m_p;
+    }
+
+    inline const IntRowVectorType& permutationQ() const
+    {
+      if (m_extractedDataAreDirty) this->extractData();
+      return m_q;
+    }
+    
+    Scalar determinant() const;
+    
+  protected:
+    
+    using Base::m_matrix;
+    using Base::m_sluOptions;
+    using Base::m_sluA;
+    using Base::m_sluB;
+    using Base::m_sluX;
+    using Base::m_p;
+    using Base::m_q;
+    using Base::m_sluEtree;
+    using Base::m_sluEqued;
+    using Base::m_sluRscale;
+    using Base::m_sluCscale;
+    using Base::m_sluL;
+    using Base::m_sluU;
+    using Base::m_sluStat;
+    using Base::m_sluFerr;
+    using Base::m_sluBerr;
+    using Base::m_l;
+    using Base::m_u;
+    
+    using Base::m_analysisIsOk;
+    using Base::m_factorizationIsOk;
+    using Base::m_extractedDataAreDirty;
+    using Base::m_isInitialized;
+    using Base::m_info;
+    
+    void init()
+    {
+      Base::init();
+      
+      set_default_options(&this->m_sluOptions);
+      m_sluOptions.PrintStat        = NO;
+      m_sluOptions.ConditionNumber  = NO;
+      m_sluOptions.Trans            = NOTRANS;
+      m_sluOptions.ColPerm          = COLAMD;
+    }
+    
+    
+  private:
+    SuperLU(SuperLU& ) { }
+};
+
+template<typename MatrixType>
+void SuperLU<MatrixType>::factorize(const MatrixType& a)
+{
+  eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+  if(!m_analysisIsOk)
+  {
+    m_info = InvalidInput;
+    return;
+  }
+  
+  this->initFactorization(a);
+  
+  m_sluOptions.ColPerm = COLAMD;
+  int info = 0;
+  RealScalar recip_pivot_growth, rcond;
+  RealScalar ferr, berr;
+
+  StatInit(&m_sluStat);
+  SuperLU_gssvx(&m_sluOptions, &m_sluA, m_q.data(), m_p.data(), &m_sluEtree[0],
+                &m_sluEqued, &m_sluRscale[0], &m_sluCscale[0],
+                &m_sluL, &m_sluU,
+                NULL, 0,
+                &m_sluB, &m_sluX,
+                &recip_pivot_growth, &rcond,
+                &ferr, &berr,
+                &m_sluStat, &info, Scalar());
+  StatFree(&m_sluStat);
+
+  m_extractedDataAreDirty = true;
+
+  // FIXME how to better check for errors ???
+  m_info = info == 0 ? Success : NumericalIssue;
+  m_factorizationIsOk = true;
+}
+
+template<typename MatrixType>
+template<typename Rhs,typename Dest>
+void SuperLU<MatrixType>::_solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest>& x) const
+{
+  eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()");
+
+  const Index size = m_matrix.rows();
+  const Index rhsCols = b.cols();
+  eigen_assert(size==b.rows());
+
+  m_sluOptions.Trans = NOTRANS;
+  m_sluOptions.Fact = FACTORED;
+  m_sluOptions.IterRefine = NOREFINE;
+  
+
+  m_sluFerr.resize(rhsCols);
+  m_sluBerr.resize(rhsCols);
+  
+  Ref<const Matrix<typename Rhs::Scalar,Dynamic,Dynamic,ColMajor> > b_ref(b);
+  Ref<const Matrix<typename Dest::Scalar,Dynamic,Dynamic,ColMajor> > x_ref(x);
+  
+  m_sluB = SluMatrix::Map(b_ref.const_cast_derived());
+  m_sluX = SluMatrix::Map(x_ref.const_cast_derived());
+  
+  typename Rhs::PlainObject b_cpy;
+  if(m_sluEqued!='N')
+  {
+    b_cpy = b;
+    m_sluB = SluMatrix::Map(b_cpy.const_cast_derived());  
+  }
+
+  StatInit(&m_sluStat);
+  int info = 0;
+  RealScalar recip_pivot_growth, rcond;
+  SuperLU_gssvx(&m_sluOptions, &m_sluA,
+                m_q.data(), m_p.data(),
+                &m_sluEtree[0], &m_sluEqued,
+                &m_sluRscale[0], &m_sluCscale[0],
+                &m_sluL, &m_sluU,
+                NULL, 0,
+                &m_sluB, &m_sluX,
+                &recip_pivot_growth, &rcond,
+                &m_sluFerr[0], &m_sluBerr[0],
+                &m_sluStat, &info, Scalar());
+  StatFree(&m_sluStat);
+  
+  if(x.derived().data() != x_ref.data())
+    x = x_ref;
+  
+  m_info = info==0 ? Success : NumericalIssue;
+}
+
+// the code of this extractData() function has been adapted from the SuperLU's Matlab support code,
+//
+//  Copyright (c) 1994 by Xerox Corporation.  All rights reserved.
+//
+//  THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY
+//  EXPRESSED OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+//
+template<typename MatrixType, typename Derived>
+void SuperLUBase<MatrixType,Derived>::extractData() const
+{
+  eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for extracting factors, you must first call either compute() or analyzePattern()/factorize()");
+  if (m_extractedDataAreDirty)
+  {
+    int         upper;
+    int         fsupc, istart, nsupr;
+    int         lastl = 0, lastu = 0;
+    SCformat    *Lstore = static_cast<SCformat*>(m_sluL.Store);
+    NCformat    *Ustore = static_cast<NCformat*>(m_sluU.Store);
+    Scalar      *SNptr;
+
+    const Index size = m_matrix.rows();
+    m_l.resize(size,size);
+    m_l.resizeNonZeros(Lstore->nnz);
+    m_u.resize(size,size);
+    m_u.resizeNonZeros(Ustore->nnz);
+
+    int* Lcol = m_l.outerIndexPtr();
+    int* Lrow = m_l.innerIndexPtr();
+    Scalar* Lval = m_l.valuePtr();
+
+    int* Ucol = m_u.outerIndexPtr();
+    int* Urow = m_u.innerIndexPtr();
+    Scalar* Uval = m_u.valuePtr();
+
+    Ucol[0] = 0;
+    Ucol[0] = 0;
+
+    /* for each supernode */
+    for (int k = 0; k <= Lstore->nsuper; ++k)
+    {
+      fsupc   = L_FST_SUPC(k);
+      istart  = L_SUB_START(fsupc);
+      nsupr   = L_SUB_START(fsupc+1) - istart;
+      upper   = 1;
+
+      /* for each column in the supernode */
+      for (int j = fsupc; j < L_FST_SUPC(k+1); ++j)
+      {
+        SNptr = &((Scalar*)Lstore->nzval)[L_NZ_START(j)];
+
+        /* Extract U */
+        for (int i = U_NZ_START(j); i < U_NZ_START(j+1); ++i)
+        {
+          Uval[lastu] = ((Scalar*)Ustore->nzval)[i];
+          /* Matlab doesn't like explicit zero. */
+          if (Uval[lastu] != 0.0)
+            Urow[lastu++] = U_SUB(i);
+        }
+        for (int i = 0; i < upper; ++i)
+        {
+          /* upper triangle in the supernode */
+          Uval[lastu] = SNptr[i];
+          /* Matlab doesn't like explicit zero. */
+          if (Uval[lastu] != 0.0)
+            Urow[lastu++] = L_SUB(istart+i);
+        }
+        Ucol[j+1] = lastu;
+
+        /* Extract L */
+        Lval[lastl] = 1.0; /* unit diagonal */
+        Lrow[lastl++] = L_SUB(istart + upper - 1);
+        for (int i = upper; i < nsupr; ++i)
+        {
+          Lval[lastl] = SNptr[i];
+          /* Matlab doesn't like explicit zero. */
+          if (Lval[lastl] != 0.0)
+            Lrow[lastl++] = L_SUB(istart+i);
+        }
+        Lcol[j+1] = lastl;
+
+        ++upper;
+      } /* for j ... */
+
+    } /* for k ... */
+
+    // squeeze the matrices :
+    m_l.resizeNonZeros(lastl);
+    m_u.resizeNonZeros(lastu);
+
+    m_extractedDataAreDirty = false;
+  }
+}
+
+template<typename MatrixType>
+typename SuperLU<MatrixType>::Scalar SuperLU<MatrixType>::determinant() const
+{
+  eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for computing the determinant, you must first call either compute() or analyzePattern()/factorize()");
+  
+  if (m_extractedDataAreDirty)
+    this->extractData();
+
+  Scalar det = Scalar(1);
+  for (int j=0; j<m_u.cols(); ++j)
+  {
+    if (m_u.outerIndexPtr()[j+1]-m_u.outerIndexPtr()[j] > 0)
+    {
+      int lastId = m_u.outerIndexPtr()[j+1]-1;
+      eigen_assert(m_u.innerIndexPtr()[lastId]<=j);
+      if (m_u.innerIndexPtr()[lastId]==j)
+        det *= m_u.valuePtr()[lastId];
+    }
+  }
+  if(PermutationMap(m_p.data(),m_p.size()).determinant()*PermutationMap(m_q.data(),m_q.size()).determinant()<0)
+    det = -det;
+  if(m_sluEqued!='N')
+    return det/m_sluRscale.prod()/m_sluCscale.prod();
+  else
+    return det;
+}
+
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+#define EIGEN_SUPERLU_HAS_ILU
+#endif
+
+#ifdef EIGEN_SUPERLU_HAS_ILU
+
+/** \ingroup SuperLUSupport_Module
+  * \class SuperILU
+  * \brief A sparse direct \b incomplete LU factorization and solver based on the SuperLU library
+  *
+  * This class allows to solve for an approximate solution of A.X = B sparse linear problems via an incomplete LU factorization
+  * using the SuperLU library. This class is aimed to be used as a preconditioner of the iterative linear solvers.
+  *
+  * \warning This class is only for the 4.x versions of SuperLU. The 3.x and 5.x versions are not supported.
+  *
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class IncompleteLUT, class ConjugateGradient, class BiCGSTAB
+  */
+
+template<typename _MatrixType>
+class SuperILU : public SuperLUBase<_MatrixType,SuperILU<_MatrixType> >
+{
+  public:
+    typedef SuperLUBase<_MatrixType,SuperILU> Base;
+    typedef _MatrixType MatrixType;
+    typedef typename Base::Scalar Scalar;
+    typedef typename Base::RealScalar RealScalar;
+
+  public:
+    using Base::_solve_impl;
+
+    SuperILU() : Base() { init(); }
+
+    SuperILU(const MatrixType& matrix) : Base()
+    {
+      init();
+      Base::compute(matrix);
+    }
+
+    ~SuperILU()
+    {
+    }
+    
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      * 
+      * \sa factorize()
+      */
+    void analyzePattern(const MatrixType& matrix)
+    {
+      Base::analyzePattern(matrix);
+    }
+    
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the symbolic decomposition has been performed.
+      *
+      * \sa analyzePattern()
+      */
+    void factorize(const MatrixType& matrix);
+    
+    #ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal */
+    template<typename Rhs,typename Dest>
+    void _solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest> &dest) const;
+    #endif // EIGEN_PARSED_BY_DOXYGEN
+    
+  protected:
+    
+    using Base::m_matrix;
+    using Base::m_sluOptions;
+    using Base::m_sluA;
+    using Base::m_sluB;
+    using Base::m_sluX;
+    using Base::m_p;
+    using Base::m_q;
+    using Base::m_sluEtree;
+    using Base::m_sluEqued;
+    using Base::m_sluRscale;
+    using Base::m_sluCscale;
+    using Base::m_sluL;
+    using Base::m_sluU;
+    using Base::m_sluStat;
+    using Base::m_sluFerr;
+    using Base::m_sluBerr;
+    using Base::m_l;
+    using Base::m_u;
+    
+    using Base::m_analysisIsOk;
+    using Base::m_factorizationIsOk;
+    using Base::m_extractedDataAreDirty;
+    using Base::m_isInitialized;
+    using Base::m_info;
+
+    void init()
+    {
+      Base::init();
+      
+      ilu_set_default_options(&m_sluOptions);
+      m_sluOptions.PrintStat        = NO;
+      m_sluOptions.ConditionNumber  = NO;
+      m_sluOptions.Trans            = NOTRANS;
+      m_sluOptions.ColPerm          = MMD_AT_PLUS_A;
+      
+      // no attempt to preserve column sum
+      m_sluOptions.ILU_MILU = SILU;
+      // only basic ILU(k) support -- no direct control over memory consumption
+      // better to use ILU_DropRule = DROP_BASIC | DROP_AREA
+      // and set ILU_FillFactor to max memory growth
+      m_sluOptions.ILU_DropRule = DROP_BASIC;
+      m_sluOptions.ILU_DropTol = NumTraits<Scalar>::dummy_precision()*10;
+    }
+    
+  private:
+    SuperILU(SuperILU& ) { }
+};
+
+template<typename MatrixType>
+void SuperILU<MatrixType>::factorize(const MatrixType& a)
+{
+  eigen_assert(m_analysisIsOk && "You must first call analyzePattern()");
+  if(!m_analysisIsOk)
+  {
+    m_info = InvalidInput;
+    return;
+  }
+  
+  this->initFactorization(a);
+
+  int info = 0;
+  RealScalar recip_pivot_growth, rcond;
+
+  StatInit(&m_sluStat);
+  SuperLU_gsisx(&m_sluOptions, &m_sluA, m_q.data(), m_p.data(), &m_sluEtree[0],
+                &m_sluEqued, &m_sluRscale[0], &m_sluCscale[0],
+                &m_sluL, &m_sluU,
+                NULL, 0,
+                &m_sluB, &m_sluX,
+                &recip_pivot_growth, &rcond,
+                &m_sluStat, &info, Scalar());
+  StatFree(&m_sluStat);
+
+  // FIXME how to better check for errors ???
+  m_info = info == 0 ? Success : NumericalIssue;
+  m_factorizationIsOk = true;
+}
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+template<typename MatrixType>
+template<typename Rhs,typename Dest>
+void SuperILU<MatrixType>::_solve_impl(const MatrixBase<Rhs> &b, MatrixBase<Dest>& x) const
+{
+  eigen_assert(m_factorizationIsOk && "The decomposition is not in a valid state for solving, you must first call either compute() or analyzePattern()/factorize()");
+
+  const int size = m_matrix.rows();
+  const int rhsCols = b.cols();
+  eigen_assert(size==b.rows());
+
+  m_sluOptions.Trans = NOTRANS;
+  m_sluOptions.Fact = FACTORED;
+  m_sluOptions.IterRefine = NOREFINE;
+
+  m_sluFerr.resize(rhsCols);
+  m_sluBerr.resize(rhsCols);
+  
+  Ref<const Matrix<typename Rhs::Scalar,Dynamic,Dynamic,ColMajor> > b_ref(b);
+  Ref<const Matrix<typename Dest::Scalar,Dynamic,Dynamic,ColMajor> > x_ref(x);
+  
+  m_sluB = SluMatrix::Map(b_ref.const_cast_derived());
+  m_sluX = SluMatrix::Map(x_ref.const_cast_derived());
+
+  typename Rhs::PlainObject b_cpy;
+  if(m_sluEqued!='N')
+  {
+    b_cpy = b;
+    m_sluB = SluMatrix::Map(b_cpy.const_cast_derived());  
+  }
+  
+  int info = 0;
+  RealScalar recip_pivot_growth, rcond;
+
+  StatInit(&m_sluStat);
+  SuperLU_gsisx(&m_sluOptions, &m_sluA,
+                m_q.data(), m_p.data(),
+                &m_sluEtree[0], &m_sluEqued,
+                &m_sluRscale[0], &m_sluCscale[0],
+                &m_sluL, &m_sluU,
+                NULL, 0,
+                &m_sluB, &m_sluX,
+                &recip_pivot_growth, &rcond,
+                &m_sluStat, &info, Scalar());
+  StatFree(&m_sluStat);
+  
+  if(x.derived().data() != x_ref.data())
+    x = x_ref;
+
+  m_info = info==0 ? Success : NumericalIssue;
+}
+#endif
+
+#endif
+
+} // end namespace Eigen
+
+#endif // EIGEN_SUPERLUSUPPORT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h
new file mode 100644
index 0000000..91c09ab
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/UmfPackSupport/UmfPackSupport.h
@@ -0,0 +1,506 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2011 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_UMFPACKSUPPORT_H
+#define EIGEN_UMFPACKSUPPORT_H
+
+namespace Eigen {
+
+/* TODO extract L, extract U, compute det, etc... */
+
+// generic double/complex<double> wrapper functions:
+
+
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], double)
+{ umfpack_di_defaults(control); }
+
+inline void umfpack_defaults(double control[UMFPACK_CONTROL], std::complex<double>)
+{ umfpack_zi_defaults(control); }
+
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], double)
+{ umfpack_di_report_info(control, info);}
+
+inline void umfpack_report_info(double control[UMFPACK_CONTROL], double info[UMFPACK_INFO], std::complex<double>)
+{ umfpack_zi_report_info(control, info);}
+
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, double)
+{ umfpack_di_report_status(control, status);}
+
+inline void umfpack_report_status(double control[UMFPACK_CONTROL], int status, std::complex<double>)
+{ umfpack_zi_report_status(control, status);}
+
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], double)
+{ umfpack_di_report_control(control);}
+
+inline void umfpack_report_control(double control[UMFPACK_CONTROL], std::complex<double>)
+{ umfpack_zi_report_control(control);}
+
+inline void umfpack_free_numeric(void **Numeric, double)
+{ umfpack_di_free_numeric(Numeric); *Numeric = 0; }
+
+inline void umfpack_free_numeric(void **Numeric, std::complex<double>)
+{ umfpack_zi_free_numeric(Numeric); *Numeric = 0; }
+
+inline void umfpack_free_symbolic(void **Symbolic, double)
+{ umfpack_di_free_symbolic(Symbolic); *Symbolic = 0; }
+
+inline void umfpack_free_symbolic(void **Symbolic, std::complex<double>)
+{ umfpack_zi_free_symbolic(Symbolic); *Symbolic = 0; }
+
+inline int umfpack_symbolic(int n_row,int n_col,
+                            const int Ap[], const int Ai[], const double Ax[], void **Symbolic,
+                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+{
+  return umfpack_di_symbolic(n_row,n_col,Ap,Ai,Ax,Symbolic,Control,Info);
+}
+
+inline int umfpack_symbolic(int n_row,int n_col,
+                            const int Ap[], const int Ai[], const std::complex<double> Ax[], void **Symbolic,
+                            const double Control [UMFPACK_CONTROL], double Info [UMFPACK_INFO])
+{
+  return umfpack_zi_symbolic(n_row,n_col,Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Control,Info);
+}
+
+inline int umfpack_numeric( const int Ap[], const int Ai[], const double Ax[],
+                            void *Symbolic, void **Numeric,
+                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+{
+  return umfpack_di_numeric(Ap,Ai,Ax,Symbolic,Numeric,Control,Info);
+}
+
+inline int umfpack_numeric( const int Ap[], const int Ai[], const std::complex<double> Ax[],
+                            void *Symbolic, void **Numeric,
+                            const double Control[UMFPACK_CONTROL],double Info [UMFPACK_INFO])
+{
+  return umfpack_zi_numeric(Ap,Ai,&numext::real_ref(Ax[0]),0,Symbolic,Numeric,Control,Info);
+}
+
+inline int umfpack_solve( int sys, const int Ap[], const int Ai[], const double Ax[],
+                          double X[], const double B[], void *Numeric,
+                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+{
+  return umfpack_di_solve(sys,Ap,Ai,Ax,X,B,Numeric,Control,Info);
+}
+
+inline int umfpack_solve( int sys, const int Ap[], const int Ai[], const std::complex<double> Ax[],
+                          std::complex<double> X[], const std::complex<double> B[], void *Numeric,
+                          const double Control[UMFPACK_CONTROL], double Info[UMFPACK_INFO])
+{
+  return umfpack_zi_solve(sys,Ap,Ai,&numext::real_ref(Ax[0]),0,&numext::real_ref(X[0]),0,&numext::real_ref(B[0]),0,Numeric,Control,Info);
+}
+
+inline int umfpack_get_lunz(int *lnz, int *unz, int *n_row, int *n_col, int *nz_udiag, void *Numeric, double)
+{
+  return umfpack_di_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
+}
+
+inline int umfpack_get_lunz(int *lnz, int *unz, int *n_row, int *n_col, int *nz_udiag, void *Numeric, std::complex<double>)
+{
+  return umfpack_zi_get_lunz(lnz,unz,n_row,n_col,nz_udiag,Numeric);
+}
+
+inline int umfpack_get_numeric(int Lp[], int Lj[], double Lx[], int Up[], int Ui[], double Ux[],
+                               int P[], int Q[], double Dx[], int *do_recip, double Rs[], void *Numeric)
+{
+  return umfpack_di_get_numeric(Lp,Lj,Lx,Up,Ui,Ux,P,Q,Dx,do_recip,Rs,Numeric);
+}
+
+inline int umfpack_get_numeric(int Lp[], int Lj[], std::complex<double> Lx[], int Up[], int Ui[], std::complex<double> Ux[],
+                               int P[], int Q[], std::complex<double> Dx[], int *do_recip, double Rs[], void *Numeric)
+{
+  double& lx0_real = numext::real_ref(Lx[0]);
+  double& ux0_real = numext::real_ref(Ux[0]);
+  double& dx0_real = numext::real_ref(Dx[0]);
+  return umfpack_zi_get_numeric(Lp,Lj,Lx?&lx0_real:0,0,Up,Ui,Ux?&ux0_real:0,0,P,Q,
+                                Dx?&dx0_real:0,0,do_recip,Rs,Numeric);
+}
+
+inline int umfpack_get_determinant(double *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO])
+{
+  return umfpack_di_get_determinant(Mx,Ex,NumericHandle,User_Info);
+}
+
+inline int umfpack_get_determinant(std::complex<double> *Mx, double *Ex, void *NumericHandle, double User_Info [UMFPACK_INFO])
+{
+  double& mx_real = numext::real_ref(*Mx);
+  return umfpack_zi_get_determinant(&mx_real,0,Ex,NumericHandle,User_Info);
+}
+
+
+/** \ingroup UmfPackSupport_Module
+  * \brief A sparse LU factorization and solver based on UmfPack
+  *
+  * This class allows to solve for A.X = B sparse linear problems via a LU factorization
+  * using the UmfPack library. The sparse matrix A must be squared and full rank.
+  * The vectors or matrices X and B can be either dense or sparse.
+  *
+  * \warning The input matrix A should be in a \b compressed and \b column-major form.
+  * Otherwise an expensive copy will be made. You can call the inexpensive makeCompressed() to get a compressed matrix.
+  * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
+  *
+  * \implsparsesolverconcept
+  *
+  * \sa \ref TutorialSparseSolverConcept, class SparseLU
+  */
+template<typename _MatrixType>
+class UmfPackLU : public SparseSolverBase<UmfPackLU<_MatrixType> >
+{
+  protected:
+    typedef SparseSolverBase<UmfPackLU<_MatrixType> > Base;
+    using Base::m_isInitialized;
+  public:
+    using Base::_solve_impl;
+    typedef _MatrixType MatrixType;
+    typedef typename MatrixType::Scalar Scalar;
+    typedef typename MatrixType::RealScalar RealScalar;
+    typedef typename MatrixType::StorageIndex StorageIndex;
+    typedef Matrix<Scalar,Dynamic,1> Vector;
+    typedef Matrix<int, 1, MatrixType::ColsAtCompileTime> IntRowVectorType;
+    typedef Matrix<int, MatrixType::RowsAtCompileTime, 1> IntColVectorType;
+    typedef SparseMatrix<Scalar> LUMatrixType;
+    typedef SparseMatrix<Scalar,ColMajor,int> UmfpackMatrixType;
+    typedef Ref<const UmfpackMatrixType, StandardCompressedFormat> UmfpackMatrixRef;
+    enum {
+      ColsAtCompileTime = MatrixType::ColsAtCompileTime,
+      MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime
+    };
+
+  public:
+
+    typedef Array<double, UMFPACK_CONTROL, 1> UmfpackControl;
+    typedef Array<double, UMFPACK_INFO, 1> UmfpackInfo;
+
+    UmfPackLU()
+      : m_dummy(0,0), mp_matrix(m_dummy)
+    {
+      init();
+    }
+
+    template<typename InputMatrixType>
+    explicit UmfPackLU(const InputMatrixType& matrix)
+      : mp_matrix(matrix)
+    {
+      init();
+      compute(matrix);
+    }
+
+    ~UmfPackLU()
+    {
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+    }
+
+    inline Index rows() const { return mp_matrix.rows(); }
+    inline Index cols() const { return mp_matrix.cols(); }
+
+    /** \brief Reports whether previous computation was successful.
+      *
+      * \returns \c Success if computation was succesful,
+      *          \c NumericalIssue if the matrix.appears to be negative.
+      */
+    ComputationInfo info() const
+    {
+      eigen_assert(m_isInitialized && "Decomposition is not initialized.");
+      return m_info;
+    }
+
+    inline const LUMatrixType& matrixL() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_l;
+    }
+
+    inline const LUMatrixType& matrixU() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_u;
+    }
+
+    inline const IntColVectorType& permutationP() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_p;
+    }
+
+    inline const IntRowVectorType& permutationQ() const
+    {
+      if (m_extractedDataAreDirty) extractData();
+      return m_q;
+    }
+
+    /** Computes the sparse Cholesky decomposition of \a matrix
+     *  Note that the matrix should be column-major, and in compressed format for best performance.
+     *  \sa SparseMatrix::makeCompressed().
+     */
+    template<typename InputMatrixType>
+    void compute(const InputMatrixType& matrix)
+    {
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+      grab(matrix.derived());
+      analyzePattern_impl();
+      factorize_impl();
+    }
+
+    /** Performs a symbolic decomposition on the sparcity of \a matrix.
+      *
+      * This function is particularly useful when solving for several problems having the same structure.
+      *
+      * \sa factorize(), compute()
+      */
+    template<typename InputMatrixType>
+    void analyzePattern(const InputMatrixType& matrix)
+    {
+      if(m_symbolic) umfpack_free_symbolic(&m_symbolic,Scalar());
+      if(m_numeric)  umfpack_free_numeric(&m_numeric,Scalar());
+
+      grab(matrix.derived());
+
+      analyzePattern_impl();
+    }
+
+    /** Provides the return status code returned by UmfPack during the numeric
+      * factorization.
+      *
+      * \sa factorize(), compute()
+      */
+    inline int umfpackFactorizeReturncode() const
+    {
+      eigen_assert(m_numeric && "UmfPackLU: you must first call factorize()");
+      return m_fact_errorCode;
+    }
+
+    /** Provides access to the control settings array used by UmfPack.
+      *
+      * If this array contains NaN's, the default values are used.
+      *
+      * See UMFPACK documentation for details.
+      */
+    inline const UmfpackControl& umfpackControl() const
+    {
+      return m_control;
+    }
+
+    /** Provides access to the control settings array used by UmfPack.
+      *
+      * If this array contains NaN's, the default values are used.
+      *
+      * See UMFPACK documentation for details.
+      */
+    inline UmfpackControl& umfpackControl()
+    {
+      return m_control;
+    }
+
+    /** Performs a numeric decomposition of \a matrix
+      *
+      * The given matrix must has the same sparcity than the matrix on which the pattern anylysis has been performed.
+      *
+      * \sa analyzePattern(), compute()
+      */
+    template<typename InputMatrixType>
+    void factorize(const InputMatrixType& matrix)
+    {
+      eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
+      if(m_numeric)
+        umfpack_free_numeric(&m_numeric,Scalar());
+
+      grab(matrix.derived());
+
+      factorize_impl();
+    }
+
+    /** Prints the current UmfPack control settings.
+      *
+      * \sa umfpackControl()
+      */
+    void umfpackReportControl()
+    {
+      umfpack_report_control(m_control.data(), Scalar());
+    }
+
+    /** Prints statistics collected by UmfPack.
+      *
+      * \sa analyzePattern(), compute()
+      */
+    void umfpackReportInfo()
+    {
+      eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
+      umfpack_report_info(m_control.data(), m_umfpackInfo.data(), Scalar());
+    }
+
+    /** Prints the status of the previous factorization operation performed by UmfPack (symbolic or numerical factorization).
+      *
+      * \sa analyzePattern(), compute()
+      */
+    void umfpackReportStatus() {
+      eigen_assert(m_analysisIsOk && "UmfPackLU: you must first call analyzePattern()");
+      umfpack_report_status(m_control.data(), m_fact_errorCode, Scalar());
+    }
+
+    /** \internal */
+    template<typename BDerived,typename XDerived>
+    bool _solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const;
+
+    Scalar determinant() const;
+
+    void extractData() const;
+
+  protected:
+
+    void init()
+    {
+      m_info                  = InvalidInput;
+      m_isInitialized         = false;
+      m_numeric               = 0;
+      m_symbolic              = 0;
+      m_extractedDataAreDirty = true;
+
+      umfpack_defaults(m_control.data(), Scalar());
+    }
+
+    void analyzePattern_impl()
+    {
+      m_fact_errorCode = umfpack_symbolic(internal::convert_index<int>(mp_matrix.rows()),
+                                          internal::convert_index<int>(mp_matrix.cols()),
+                                          mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
+                                          &m_symbolic, m_control.data(), m_umfpackInfo.data());
+
+      m_isInitialized = true;
+      m_info = m_fact_errorCode ? InvalidInput : Success;
+      m_analysisIsOk = true;
+      m_factorizationIsOk = false;
+      m_extractedDataAreDirty = true;
+    }
+
+    void factorize_impl()
+    {
+
+      m_fact_errorCode = umfpack_numeric(mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
+                                         m_symbolic, &m_numeric, m_control.data(), m_umfpackInfo.data());
+
+      m_info = m_fact_errorCode == UMFPACK_OK ? Success : NumericalIssue;
+      m_factorizationIsOk = true;
+      m_extractedDataAreDirty = true;
+    }
+
+    template<typename MatrixDerived>
+    void grab(const EigenBase<MatrixDerived> &A)
+    {
+      mp_matrix.~UmfpackMatrixRef();
+      ::new (&mp_matrix) UmfpackMatrixRef(A.derived());
+    }
+
+    void grab(const UmfpackMatrixRef &A)
+    {
+      if(&(A.derived()) != &mp_matrix)
+      {
+        mp_matrix.~UmfpackMatrixRef();
+        ::new (&mp_matrix) UmfpackMatrixRef(A);
+      }
+    }
+
+    // cached data to reduce reallocation, etc.
+    mutable LUMatrixType m_l;
+    int m_fact_errorCode;
+    UmfpackControl m_control;
+    mutable UmfpackInfo m_umfpackInfo;
+
+    mutable LUMatrixType m_u;
+    mutable IntColVectorType m_p;
+    mutable IntRowVectorType m_q;
+
+    UmfpackMatrixType m_dummy;
+    UmfpackMatrixRef mp_matrix;
+
+    void* m_numeric;
+    void* m_symbolic;
+
+    mutable ComputationInfo m_info;
+    int m_factorizationIsOk;
+    int m_analysisIsOk;
+    mutable bool m_extractedDataAreDirty;
+
+  private:
+    UmfPackLU(const UmfPackLU& ) { }
+};
+
+
+template<typename MatrixType>
+void UmfPackLU<MatrixType>::extractData() const
+{
+  if (m_extractedDataAreDirty)
+  {
+    // get size of the data
+    int lnz, unz, rows, cols, nz_udiag;
+    umfpack_get_lunz(&lnz, &unz, &rows, &cols, &nz_udiag, m_numeric, Scalar());
+
+    // allocate data
+    m_l.resize(rows,(std::min)(rows,cols));
+    m_l.resizeNonZeros(lnz);
+
+    m_u.resize((std::min)(rows,cols),cols);
+    m_u.resizeNonZeros(unz);
+
+    m_p.resize(rows);
+    m_q.resize(cols);
+
+    // extract
+    umfpack_get_numeric(m_l.outerIndexPtr(), m_l.innerIndexPtr(), m_l.valuePtr(),
+                        m_u.outerIndexPtr(), m_u.innerIndexPtr(), m_u.valuePtr(),
+                        m_p.data(), m_q.data(), 0, 0, 0, m_numeric);
+
+    m_extractedDataAreDirty = false;
+  }
+}
+
+template<typename MatrixType>
+typename UmfPackLU<MatrixType>::Scalar UmfPackLU<MatrixType>::determinant() const
+{
+  Scalar det;
+  umfpack_get_determinant(&det, 0, m_numeric, 0);
+  return det;
+}
+
+template<typename MatrixType>
+template<typename BDerived,typename XDerived>
+bool UmfPackLU<MatrixType>::_solve_impl(const MatrixBase<BDerived> &b, MatrixBase<XDerived> &x) const
+{
+  Index rhsCols = b.cols();
+  eigen_assert((BDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major rhs yet");
+  eigen_assert((XDerived::Flags&RowMajorBit)==0 && "UmfPackLU backend does not support non col-major result yet");
+  eigen_assert(b.derived().data() != x.derived().data() && " Umfpack does not support inplace solve");
+
+  int errorCode;
+  Scalar* x_ptr = 0;
+  Matrix<Scalar,Dynamic,1> x_tmp;
+  if(x.innerStride()!=1)
+  {
+    x_tmp.resize(x.rows());
+    x_ptr = x_tmp.data();
+  }
+  for (int j=0; j<rhsCols; ++j)
+  {
+    if(x.innerStride()==1)
+      x_ptr = &x.col(j).coeffRef(0);
+    errorCode = umfpack_solve(UMFPACK_A,
+        mp_matrix.outerIndexPtr(), mp_matrix.innerIndexPtr(), mp_matrix.valuePtr(),
+        x_ptr, &b.const_cast_derived().col(j).coeffRef(0), m_numeric, m_control.data(), m_umfpackInfo.data());
+    if(x.innerStride()!=1)
+      x.col(j) = x_tmp;
+    if (errorCode!=0)
+      return false;
+  }
+
+  return true;
+}
+
+} // end namespace Eigen
+
+#endif // EIGEN_UMFPACKSUPPORT_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Image.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Image.h
new file mode 100644
index 0000000..b8b8a04
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Image.h
@@ -0,0 +1,82 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MISC_IMAGE_H
+#define EIGEN_MISC_IMAGE_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \class image_retval_base
+  *
+  */
+template<typename DecompositionType>
+struct traits<image_retval_base<DecompositionType> >
+{
+  typedef typename DecompositionType::MatrixType MatrixType;
+  typedef Matrix<
+    typename MatrixType::Scalar,
+    MatrixType::RowsAtCompileTime, // the image is a subspace of the destination space, whose
+                                   // dimension is the number of rows of the original matrix
+    Dynamic,                       // we don't know at compile time the dimension of the image (the rank)
+    MatrixType::Options,
+    MatrixType::MaxRowsAtCompileTime, // the image matrix will consist of columns from the original matrix,
+    MatrixType::MaxColsAtCompileTime  // so it has the same number of rows and at most as many columns.
+  > ReturnType;
+};
+
+template<typename _DecompositionType> struct image_retval_base
+ : public ReturnByValue<image_retval_base<_DecompositionType> >
+{
+  typedef _DecompositionType DecompositionType;
+  typedef typename DecompositionType::MatrixType MatrixType;
+  typedef ReturnByValue<image_retval_base> Base;
+
+  image_retval_base(const DecompositionType& dec, const MatrixType& originalMatrix)
+    : m_dec(dec), m_rank(dec.rank()),
+      m_cols(m_rank == 0 ? 1 : m_rank),
+      m_originalMatrix(originalMatrix)
+  {}
+
+  inline Index rows() const { return m_dec.rows(); }
+  inline Index cols() const { return m_cols; }
+  inline Index rank() const { return m_rank; }
+  inline const DecompositionType& dec() const { return m_dec; }
+  inline const MatrixType& originalMatrix() const { return m_originalMatrix; }
+
+  template<typename Dest> inline void evalTo(Dest& dst) const
+  {
+    static_cast<const image_retval<DecompositionType>*>(this)->evalTo(dst);
+  }
+
+  protected:
+    const DecompositionType& m_dec;
+    Index m_rank, m_cols;
+    const MatrixType& m_originalMatrix;
+};
+
+} // end namespace internal
+
+#define EIGEN_MAKE_IMAGE_HELPERS(DecompositionType) \
+  typedef typename DecompositionType::MatrixType MatrixType; \
+  typedef typename MatrixType::Scalar Scalar; \
+  typedef typename MatrixType::RealScalar RealScalar; \
+  typedef Eigen::internal::image_retval_base<DecompositionType> Base; \
+  using Base::dec; \
+  using Base::originalMatrix; \
+  using Base::rank; \
+  using Base::rows; \
+  using Base::cols; \
+  image_retval(const DecompositionType& dec, const MatrixType& originalMatrix) \
+    : Base(dec, originalMatrix) {}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MISC_IMAGE_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Kernel.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Kernel.h
new file mode 100644
index 0000000..bef5d6f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/Kernel.h
@@ -0,0 +1,79 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_MISC_KERNEL_H
+#define EIGEN_MISC_KERNEL_H
+
+namespace Eigen { 
+
+namespace internal {
+
+/** \class kernel_retval_base
+  *
+  */
+template<typename DecompositionType>
+struct traits<kernel_retval_base<DecompositionType> >
+{
+  typedef typename DecompositionType::MatrixType MatrixType;
+  typedef Matrix<
+    typename MatrixType::Scalar,
+    MatrixType::ColsAtCompileTime, // the number of rows in the "kernel matrix"
+                                   // is the number of cols of the original matrix
+                                   // so that the product "matrix * kernel = zero" makes sense
+    Dynamic,                       // we don't know at compile-time the dimension of the kernel
+    MatrixType::Options,
+    MatrixType::MaxColsAtCompileTime, // see explanation for 2nd template parameter
+    MatrixType::MaxColsAtCompileTime // the kernel is a subspace of the domain space,
+                                     // whose dimension is the number of columns of the original matrix
+  > ReturnType;
+};
+
+template<typename _DecompositionType> struct kernel_retval_base
+ : public ReturnByValue<kernel_retval_base<_DecompositionType> >
+{
+  typedef _DecompositionType DecompositionType;
+  typedef ReturnByValue<kernel_retval_base> Base;
+
+  explicit kernel_retval_base(const DecompositionType& dec)
+    : m_dec(dec),
+      m_rank(dec.rank()),
+      m_cols(m_rank==dec.cols() ? 1 : dec.cols() - m_rank)
+  {}
+
+  inline Index rows() const { return m_dec.cols(); }
+  inline Index cols() const { return m_cols; }
+  inline Index rank() const { return m_rank; }
+  inline const DecompositionType& dec() const { return m_dec; }
+
+  template<typename Dest> inline void evalTo(Dest& dst) const
+  {
+    static_cast<const kernel_retval<DecompositionType>*>(this)->evalTo(dst);
+  }
+
+  protected:
+    const DecompositionType& m_dec;
+    Index m_rank, m_cols;
+};
+
+} // end namespace internal
+
+#define EIGEN_MAKE_KERNEL_HELPERS(DecompositionType) \
+  typedef typename DecompositionType::MatrixType MatrixType; \
+  typedef typename MatrixType::Scalar Scalar; \
+  typedef typename MatrixType::RealScalar RealScalar; \
+  typedef Eigen::internal::kernel_retval_base<DecompositionType> Base; \
+  using Base::dec; \
+  using Base::rank; \
+  using Base::rows; \
+  using Base::cols; \
+  kernel_retval(const DecompositionType& dec) : Base(dec) {}
+
+} // end namespace Eigen
+
+#endif // EIGEN_MISC_KERNEL_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/RealSvd2x2.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/RealSvd2x2.h
new file mode 100644
index 0000000..abb4d3c
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/RealSvd2x2.h
@@ -0,0 +1,55 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2009-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+// Copyright (C) 2013-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_REALSVD2X2_H
+#define EIGEN_REALSVD2X2_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<typename MatrixType, typename RealScalar, typename Index>
+void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q,
+                         JacobiRotation<RealScalar> *j_left,
+                         JacobiRotation<RealScalar> *j_right)
+{
+  using std::sqrt;
+  using std::abs;
+  Matrix<RealScalar,2,2> m;
+  m << numext::real(matrix.coeff(p,p)), numext::real(matrix.coeff(p,q)),
+       numext::real(matrix.coeff(q,p)), numext::real(matrix.coeff(q,q));
+  JacobiRotation<RealScalar> rot1;
+  RealScalar t = m.coeff(0,0) + m.coeff(1,1);
+  RealScalar d = m.coeff(1,0) - m.coeff(0,1);
+
+  if(abs(d) < (std::numeric_limits<RealScalar>::min)())
+  {
+    rot1.s() = RealScalar(0);
+    rot1.c() = RealScalar(1);
+  }
+  else
+  {
+    // If d!=0, then t/d cannot overflow because the magnitude of the
+    // entries forming d are not too small compared to the ones forming t.
+    RealScalar u = t / d;
+    RealScalar tmp = sqrt(RealScalar(1) + numext::abs2(u));
+    rot1.s() = RealScalar(1) / tmp;
+    rot1.c() = u / tmp;
+  }
+  m.applyOnTheLeft(0,1,rot1);
+  j_right->makeJacobi(m,0,1);
+  *j_left = rot1 * j_right->transpose();
+}
+
+} // end namespace internal
+
+} // end namespace Eigen
+
+#endif // EIGEN_REALSVD2X2_H
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/blas.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/blas.h
new file mode 100644
index 0000000..25215b1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/blas.h
@@ -0,0 +1,440 @@
+#ifndef BLAS_H
+#define BLAS_H
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+#define BLASFUNC(FUNC) FUNC##_
+
+#ifdef __WIN64__
+typedef long long BLASLONG;
+typedef unsigned long long BLASULONG;
+#else
+typedef long BLASLONG;
+typedef unsigned long BLASULONG;
+#endif
+
+int    BLASFUNC(xerbla)(const char *, int *info, int);
+
+float  BLASFUNC(sdot)  (int *, float  *, int *, float  *, int *);
+float  BLASFUNC(sdsdot)(int *, float  *,        float  *, int *, float  *, int *);
+
+double BLASFUNC(dsdot) (int *, float  *, int *, float  *, int *);
+double BLASFUNC(ddot)  (int *, double *, int *, double *, int *);
+double BLASFUNC(qdot)  (int *, double *, int *, double *, int *);
+
+int  BLASFUNC(cdotuw)  (int *, float  *, int *, float  *, int *, float*);
+int  BLASFUNC(cdotcw)  (int *, float  *, int *, float  *, int *, float*);
+int  BLASFUNC(zdotuw)  (int *, double  *, int *, double  *, int *, double*);
+int  BLASFUNC(zdotcw)  (int *, double  *, int *, double  *, int *, double*);
+
+int    BLASFUNC(saxpy) (const int *, const float  *, const float  *, const int *, float  *, const int *);
+int    BLASFUNC(daxpy) (const int *, const double *, const double *, const int *, double *, const int *);
+int    BLASFUNC(qaxpy) (const int *, const double *, const double *, const int *, double *, const int *);
+int    BLASFUNC(caxpy) (const int *, const float  *, const float  *, const int *, float  *, const int *);
+int    BLASFUNC(zaxpy) (const int *, const double *, const double *, const int *, double *, const int *);
+int    BLASFUNC(xaxpy) (const int *, const double *, const double *, const int *, double *, const int *);
+int    BLASFUNC(caxpyc)(const int *, const float  *, const float  *, const int *, float  *, const int *);
+int    BLASFUNC(zaxpyc)(const int *, const double *, const double *, const int *, double *, const int *);
+int    BLASFUNC(xaxpyc)(const int *, const double *, const double *, const int *, double *, const int *);
+
+int    BLASFUNC(scopy) (int *, float  *, int *, float  *, int *);
+int    BLASFUNC(dcopy) (int *, double *, int *, double *, int *);
+int    BLASFUNC(qcopy) (int *, double *, int *, double *, int *);
+int    BLASFUNC(ccopy) (int *, float  *, int *, float  *, int *);
+int    BLASFUNC(zcopy) (int *, double *, int *, double *, int *);
+int    BLASFUNC(xcopy) (int *, double *, int *, double *, int *);
+
+int    BLASFUNC(sswap) (int *, float  *, int *, float  *, int *);
+int    BLASFUNC(dswap) (int *, double *, int *, double *, int *);
+int    BLASFUNC(qswap) (int *, double *, int *, double *, int *);
+int    BLASFUNC(cswap) (int *, float  *, int *, float  *, int *);
+int    BLASFUNC(zswap) (int *, double *, int *, double *, int *);
+int    BLASFUNC(xswap) (int *, double *, int *, double *, int *);
+
+float  BLASFUNC(sasum) (int *, float  *, int *);
+float  BLASFUNC(scasum)(int *, float  *, int *);
+double BLASFUNC(dasum) (int *, double *, int *);
+double BLASFUNC(qasum) (int *, double *, int *);
+double BLASFUNC(dzasum)(int *, double *, int *);
+double BLASFUNC(qxasum)(int *, double *, int *);
+
+int    BLASFUNC(isamax)(int *, float  *, int *);
+int    BLASFUNC(idamax)(int *, double *, int *);
+int    BLASFUNC(iqamax)(int *, double *, int *);
+int    BLASFUNC(icamax)(int *, float  *, int *);
+int    BLASFUNC(izamax)(int *, double *, int *);
+int    BLASFUNC(ixamax)(int *, double *, int *);
+
+int    BLASFUNC(ismax) (int *, float  *, int *);
+int    BLASFUNC(idmax) (int *, double *, int *);
+int    BLASFUNC(iqmax) (int *, double *, int *);
+int    BLASFUNC(icmax) (int *, float  *, int *);
+int    BLASFUNC(izmax) (int *, double *, int *);
+int    BLASFUNC(ixmax) (int *, double *, int *);
+
+int    BLASFUNC(isamin)(int *, float  *, int *);
+int    BLASFUNC(idamin)(int *, double *, int *);
+int    BLASFUNC(iqamin)(int *, double *, int *);
+int    BLASFUNC(icamin)(int *, float  *, int *);
+int    BLASFUNC(izamin)(int *, double *, int *);
+int    BLASFUNC(ixamin)(int *, double *, int *);
+
+int    BLASFUNC(ismin)(int *, float  *, int *);
+int    BLASFUNC(idmin)(int *, double *, int *);
+int    BLASFUNC(iqmin)(int *, double *, int *);
+int    BLASFUNC(icmin)(int *, float  *, int *);
+int    BLASFUNC(izmin)(int *, double *, int *);
+int    BLASFUNC(ixmin)(int *, double *, int *);
+
+float  BLASFUNC(samax) (int *, float  *, int *);
+double BLASFUNC(damax) (int *, double *, int *);
+double BLASFUNC(qamax) (int *, double *, int *);
+float  BLASFUNC(scamax)(int *, float  *, int *);
+double BLASFUNC(dzamax)(int *, double *, int *);
+double BLASFUNC(qxamax)(int *, double *, int *);
+
+float  BLASFUNC(samin) (int *, float  *, int *);
+double BLASFUNC(damin) (int *, double *, int *);
+double BLASFUNC(qamin) (int *, double *, int *);
+float  BLASFUNC(scamin)(int *, float  *, int *);
+double BLASFUNC(dzamin)(int *, double *, int *);
+double BLASFUNC(qxamin)(int *, double *, int *);
+
+float  BLASFUNC(smax)  (int *, float  *, int *);
+double BLASFUNC(dmax)  (int *, double *, int *);
+double BLASFUNC(qmax)  (int *, double *, int *);
+float  BLASFUNC(scmax) (int *, float  *, int *);
+double BLASFUNC(dzmax) (int *, double *, int *);
+double BLASFUNC(qxmax) (int *, double *, int *);
+
+float  BLASFUNC(smin)  (int *, float  *, int *);
+double BLASFUNC(dmin)  (int *, double *, int *);
+double BLASFUNC(qmin)  (int *, double *, int *);
+float  BLASFUNC(scmin) (int *, float  *, int *);
+double BLASFUNC(dzmin) (int *, double *, int *);
+double BLASFUNC(qxmin) (int *, double *, int *);
+
+int    BLASFUNC(sscal) (int *,  float  *, float  *, int *);
+int    BLASFUNC(dscal) (int *,  double *, double *, int *);
+int    BLASFUNC(qscal) (int *,  double *, double *, int *);
+int    BLASFUNC(cscal) (int *,  float  *, float  *, int *);
+int    BLASFUNC(zscal) (int *,  double *, double *, int *);
+int    BLASFUNC(xscal) (int *,  double *, double *, int *);
+int    BLASFUNC(csscal)(int *,  float  *, float  *, int *);
+int    BLASFUNC(zdscal)(int *,  double *, double *, int *);
+int    BLASFUNC(xqscal)(int *,  double *, double *, int *);
+
+float  BLASFUNC(snrm2) (int *, float  *, int *);
+float  BLASFUNC(scnrm2)(int *, float  *, int *);
+
+double BLASFUNC(dnrm2) (int *, double *, int *);
+double BLASFUNC(qnrm2) (int *, double *, int *);
+double BLASFUNC(dznrm2)(int *, double *, int *);
+double BLASFUNC(qxnrm2)(int *, double *, int *);
+
+int    BLASFUNC(srot)  (int *, float  *, int *, float  *, int *, float  *, float  *);
+int    BLASFUNC(drot)  (int *, double *, int *, double *, int *, double *, double *);
+int    BLASFUNC(qrot)  (int *, double *, int *, double *, int *, double *, double *);
+int    BLASFUNC(csrot) (int *, float  *, int *, float  *, int *, float  *, float  *);
+int    BLASFUNC(zdrot) (int *, double *, int *, double *, int *, double *, double *);
+int    BLASFUNC(xqrot) (int *, double *, int *, double *, int *, double *, double *);
+
+int    BLASFUNC(srotg) (float  *, float  *, float  *, float  *);
+int    BLASFUNC(drotg) (double *, double *, double *, double *);
+int    BLASFUNC(qrotg) (double *, double *, double *, double *);
+int    BLASFUNC(crotg) (float  *, float  *, float  *, float  *);
+int    BLASFUNC(zrotg) (double *, double *, double *, double *);
+int    BLASFUNC(xrotg) (double *, double *, double *, double *);
+
+int    BLASFUNC(srotmg)(float  *, float  *, float  *, float  *, float  *);
+int    BLASFUNC(drotmg)(double *, double *, double *, double *, double *);
+
+int    BLASFUNC(srotm) (int *, float  *, int *, float  *, int *, float  *);
+int    BLASFUNC(drotm) (int *, double *, int *, double *, int *, double *);
+int    BLASFUNC(qrotm) (int *, double *, int *, double *, int *, double *);
+
+/* Level 2 routines */
+
+int BLASFUNC(sger)(int *,    int *, float *,  float *, int *,
+		   float *,  int *, float *,  int *);
+int BLASFUNC(dger)(int *,    int *, double *, double *, int *,
+		   double *, int *, double *, int *);
+int BLASFUNC(qger)(int *,    int *, double *, double *, int *,
+		   double *, int *, double *, int *);
+int BLASFUNC(cgeru)(int *,    int *, float *,  float *, int *,
+		    float *,  int *, float *,  int *);
+int BLASFUNC(cgerc)(int *,    int *, float *,  float *, int *,
+		    float *,  int *, float *,  int *);
+int BLASFUNC(zgeru)(int *,    int *, double *, double *, int *,
+		    double *, int *, double *, int *);
+int BLASFUNC(zgerc)(int *,    int *, double *, double *, int *,
+		    double *, int *, double *, int *);
+int BLASFUNC(xgeru)(int *,    int *, double *, double *, int *,
+		    double *, int *, double *, int *);
+int BLASFUNC(xgerc)(int *,    int *, double *, double *, int *,
+		    double *, int *, double *, int *);
+
+int BLASFUNC(sgemv)(const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dgemv)(const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(qgemv)(const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(cgemv)(const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zgemv)(const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xgemv)(const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(strsv) (const char *, const char *, const char *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(dtrsv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(qtrsv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(ctrsv) (const char *, const char *, const char *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(ztrsv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(xtrsv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(stpsv) (char *, char *, char *, int *, float  *, float  *, int *);
+int BLASFUNC(dtpsv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(qtpsv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(ctpsv) (char *, char *, char *, int *, float  *, float  *, int *);
+int BLASFUNC(ztpsv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(xtpsv) (char *, char *, char *, int *, double *, double *, int *);
+
+int BLASFUNC(strmv) (const char *, const char *, const char *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(dtrmv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(qtrmv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(ctrmv) (const char *, const char *, const char *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(ztrmv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(xtrmv) (const char *, const char *, const char *, const int *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(stpmv) (char *, char *, char *, int *, float  *, float  *, int *);
+int BLASFUNC(dtpmv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(qtpmv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(ctpmv) (char *, char *, char *, int *, float  *, float  *, int *);
+int BLASFUNC(ztpmv) (char *, char *, char *, int *, double *, double *, int *);
+int BLASFUNC(xtpmv) (char *, char *, char *, int *, double *, double *, int *);
+
+int BLASFUNC(stbmv) (char *, char *, char *, int *, int *, float  *, int *, float  *, int *);
+int BLASFUNC(dtbmv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(qtbmv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(ctbmv) (char *, char *, char *, int *, int *, float  *, int *, float  *, int *);
+int BLASFUNC(ztbmv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(xtbmv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+
+int BLASFUNC(stbsv) (char *, char *, char *, int *, int *, float  *, int *, float  *, int *);
+int BLASFUNC(dtbsv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(qtbsv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(ctbsv) (char *, char *, char *, int *, int *, float  *, int *, float  *, int *);
+int BLASFUNC(ztbsv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+int BLASFUNC(xtbsv) (char *, char *, char *, int *, int *, double *, int *, double *, int *);
+
+int BLASFUNC(ssymv) (const char *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dsymv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(qsymv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(sspmv) (char *, int *, float  *, float *,
+		     float  *, int *, float *, float *, int *);
+int BLASFUNC(dspmv) (char *, int *, double  *, double *,
+		     double  *, int *, double *, double *, int *);
+int BLASFUNC(qspmv) (char *, int *, double  *, double *,
+		     double  *, int *, double *, double *, int *);
+
+int BLASFUNC(ssyr) (const char *, const int *, const float   *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(dsyr) (const char *, const int *, const double  *, const double *, const int *, double *, const int *);
+int BLASFUNC(qsyr) (const char *, const int *, const double  *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(ssyr2) (const char *, const int *, const float   *, const float  *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(dsyr2) (const char *, const int *, const double  *, const double *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(qsyr2) (const char *, const int *, const double  *, const double *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(csyr2) (const char *, const int *, const float   *, const float  *, const int *, const float  *, const int *, float  *, const int *);
+int BLASFUNC(zsyr2) (const char *, const int *, const double  *, const double *, const int *, const double *, const int *, double *, const int *);
+int BLASFUNC(xsyr2) (const char *, const int *, const double  *, const double *, const int *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(sspr) (char *, int *, float   *, float  *, int *,
+		    float  *);
+int BLASFUNC(dspr) (char *, int *, double  *, double *, int *,
+		    double *);
+int BLASFUNC(qspr) (char *, int *, double  *, double *, int *,
+		    double *);
+
+int BLASFUNC(sspr2) (char *, int *, float   *,
+		     float  *, int *, float  *, int *, float  *);
+int BLASFUNC(dspr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+int BLASFUNC(qspr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+int BLASFUNC(cspr2) (char *, int *, float   *,
+		     float  *, int *, float  *, int *, float  *);
+int BLASFUNC(zspr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+int BLASFUNC(xspr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+
+int BLASFUNC(cher) (char *, int *, float   *, float  *, int *,
+		    float  *, int *);
+int BLASFUNC(zher) (char *, int *, double  *, double *, int *,
+		    double *, int *);
+int BLASFUNC(xher) (char *, int *, double  *, double *, int *,
+		    double *, int *);
+
+int BLASFUNC(chpr) (char *, int *, float   *, float  *, int *, float  *);
+int BLASFUNC(zhpr) (char *, int *, double  *, double *, int *, double *);
+int BLASFUNC(xhpr) (char *, int *, double  *, double *, int *, double *);
+
+int BLASFUNC(cher2) (char *, int *, float   *,
+		     float  *, int *, float  *, int *, float  *, int *);
+int BLASFUNC(zher2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *, int *);
+int BLASFUNC(xher2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *, int *);
+
+int BLASFUNC(chpr2) (char *, int *, float   *,
+		     float  *, int *, float  *, int *, float  *);
+int BLASFUNC(zhpr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+int BLASFUNC(xhpr2) (char *, int *, double  *,
+		     double *, int *, double *, int *, double *);
+
+int BLASFUNC(chemv) (const char *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zhemv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xhemv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(chpmv) (char *, int *, float  *, float *,
+		     float  *, int *, float *, float *, int *);
+int BLASFUNC(zhpmv) (char *, int *, double  *, double *,
+		     double  *, int *, double *, double *, int *);
+int BLASFUNC(xhpmv) (char *, int *, double  *, double *,
+		     double  *, int *, double *, double *, int *);
+
+int BLASFUNC(snorm)(char *, int *, int *, float  *, int *);
+int BLASFUNC(dnorm)(char *, int *, int *, double *, int *);
+int BLASFUNC(cnorm)(char *, int *, int *, float  *, int *);
+int BLASFUNC(znorm)(char *, int *, int *, double *, int *);
+
+int BLASFUNC(sgbmv)(char *, int *, int *, int *, int *, float  *, float  *, int *,
+		    float  *, int *, float  *, float  *, int *);
+int BLASFUNC(dgbmv)(char *, int *, int *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(qgbmv)(char *, int *, int *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(cgbmv)(char *, int *, int *, int *, int *, float  *, float  *, int *,
+		    float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zgbmv)(char *, int *, int *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(xgbmv)(char *, int *, int *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+
+int BLASFUNC(ssbmv)(char *, int *, int *, float  *, float  *, int *,
+		    float  *, int *, float  *, float  *, int *);
+int BLASFUNC(dsbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(qsbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(csbmv)(char *, int *, int *, float  *, float  *, int *,
+		    float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zsbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(xsbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+
+int BLASFUNC(chbmv)(char *, int *, int *, float  *, float  *, int *,
+		    float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zhbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+int BLASFUNC(xhbmv)(char *, int *, int *, double *, double *, int *,
+		    double *, int *, double *, double *, int *);
+
+/* Level 3 routines */
+
+int BLASFUNC(sgemm)(const char *, const char *, const int *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dgemm)(const char *, const char *, const int *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(qgemm)(const char *, const char *, const int *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(cgemm)(const char *, const char *, const int *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zgemm)(const char *, const char *, const int *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xgemm)(const char *, const char *, const int *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(cgemm3m)(char *, char *, int *, int *, int *, float *,
+	   float  *, int *, float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zgemm3m)(char *, char *, int *, int *, int *, double *,
+	   double *, int *, double *, int *, double *, double *, int *);
+int BLASFUNC(xgemm3m)(char *, char *, int *, int *, int *, double *,
+	   double *, int *, double *, int *, double *, double *, int *);
+
+int BLASFUNC(sge2mm)(char *, char *, char *, int *, int *,
+		     float *, float  *, int *, float  *, int *,
+		     float *, float  *, int *);
+int BLASFUNC(dge2mm)(char *, char *, char *, int *, int *,
+		     double *, double  *, int *, double  *, int *,
+		     double *, double  *, int *);
+int BLASFUNC(cge2mm)(char *, char *, char *, int *, int *,
+		     float *, float  *, int *, float  *, int *,
+		     float *, float  *, int *);
+int BLASFUNC(zge2mm)(char *, char *, char *, int *, int *,
+		     double *, double  *, int *, double  *, int *,
+		     double *, double  *, int *);
+
+int BLASFUNC(strsm)(const char *, const char *, const char *, const char *, const int *, const int *, const float *,  const float *,  const int *, float *,  const int *);
+int BLASFUNC(dtrsm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(qtrsm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(ctrsm)(const char *, const char *, const char *, const char *, const int *, const int *, const float *,  const float *,  const int *, float *,  const int *);
+int BLASFUNC(ztrsm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(xtrsm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(strmm)(const char *, const char *, const char *, const char *, const int *, const int *, const float *,  const float *,  const int *, float *,  const int *);
+int BLASFUNC(dtrmm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(qtrmm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(ctrmm)(const char *, const char *, const char *, const char *, const int *, const int *, const float *,  const float *,  const int *, float *,  const int *);
+int BLASFUNC(ztrmm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+int BLASFUNC(xtrmm)(const char *, const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, double *, const int *);
+
+int BLASFUNC(ssymm)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dsymm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(qsymm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(csymm)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zsymm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xsymm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(csymm3m)(char *, char *, int *, int *, float  *, float  *, int *, float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zsymm3m)(char *, char *, int *, int *, double *, double *, int *, double *, int *, double *, double *, int *);
+int BLASFUNC(xsymm3m)(char *, char *, int *, int *, double *, double *, int *, double *, int *, double *, double *, int *);
+
+int BLASFUNC(ssyrk)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dsyrk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(qsyrk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(csyrk)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zsyrk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xsyrk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(ssyr2k)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(dsyr2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+int BLASFUNC(qsyr2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+int BLASFUNC(csyr2k)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zsyr2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+int BLASFUNC(xsyr2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+
+int BLASFUNC(chemm)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zhemm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xhemm)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(chemm3m)(char *, char *, int *, int *, float  *, float  *, int *,
+	   float  *, int *, float  *, float  *, int *);
+int BLASFUNC(zhemm3m)(char *, char *, int *, int *, double *, double *, int *,
+	   double *, int *, double *, double *, int *);
+int BLASFUNC(xhemm3m)(char *, char *, int *, int *, double *, double *, int *,
+	   double *, int *, double *, double *, int *);
+
+int BLASFUNC(cherk)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zherk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xherk)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, double *, const int *);
+
+int BLASFUNC(cher2k)(const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zher2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xher2k)(const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(cher2m)(const char *, const char *, const char *, const int *, const int *, const float  *, const float  *, const int *, const float *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zher2m)(const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+int BLASFUNC(xher2m)(const char *, const char *, const char *, const int *, const int *, const double *, const double *, const int *, const double*, const int *, const double *, double *, const int *);
+
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapack.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapack.h
new file mode 100644
index 0000000..249f357
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapack.h
@@ -0,0 +1,152 @@
+#ifndef LAPACK_H
+#define LAPACK_H
+
+#include "blas.h"
+
+#ifdef __cplusplus
+extern "C"
+{
+#endif
+
+int BLASFUNC(csymv) (const char *, const int *, const float  *, const float  *, const int *, const float  *, const int *, const float  *, float  *, const int *);
+int BLASFUNC(zsymv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+int BLASFUNC(xsymv) (const char *, const int *, const double *, const double *, const int *, const double *, const int *, const double *, double *, const int *);
+
+
+int BLASFUNC(cspmv) (char *, int *, float  *, float *,
+         float  *, int *, float *, float *, int *);
+int BLASFUNC(zspmv) (char *, int *, double  *, double *,
+         double  *, int *, double *, double *, int *);
+int BLASFUNC(xspmv) (char *, int *, double  *, double *,
+         double  *, int *, double *, double *, int *);
+
+int BLASFUNC(csyr) (char *, int *, float   *, float  *, int *,
+        float  *, int *);
+int BLASFUNC(zsyr) (char *, int *, double  *, double *, int *,
+        double *, int *);
+int BLASFUNC(xsyr) (char *, int *, double  *, double *, int *,
+        double *, int *);
+
+int BLASFUNC(cspr) (char *, int *, float   *, float  *, int *,
+        float  *);
+int BLASFUNC(zspr) (char *, int *, double  *, double *, int *,
+        double *);
+int BLASFUNC(xspr) (char *, int *, double  *, double *, int *,
+        double *);
+
+int BLASFUNC(sgemt)(char *, int *, int *, float  *, float  *, int *,
+        float  *, int *);
+int BLASFUNC(dgemt)(char *, int *, int *, double *, double *, int *,
+        double *, int *);
+int BLASFUNC(cgemt)(char *, int *, int *, float  *, float  *, int *,
+        float  *, int *);
+int BLASFUNC(zgemt)(char *, int *, int *, double *, double *, int *,
+        double *, int *);
+
+int BLASFUNC(sgema)(char *, char *, int *, int *, float  *,
+        float  *, int *, float *, float  *, int *, float *, int *);
+int BLASFUNC(dgema)(char *, char *, int *, int *, double *,
+        double *, int *, double*, double *, int *, double*, int *);
+int BLASFUNC(cgema)(char *, char *, int *, int *, float  *,
+        float  *, int *, float *, float  *, int *, float *, int *);
+int BLASFUNC(zgema)(char *, char *, int *, int *, double *,
+        double *, int *, double*, double *, int *, double*, int *);
+
+int BLASFUNC(sgems)(char *, char *, int *, int *, float  *,
+        float  *, int *, float *, float  *, int *, float *, int *);
+int BLASFUNC(dgems)(char *, char *, int *, int *, double *,
+        double *, int *, double*, double *, int *, double*, int *);
+int BLASFUNC(cgems)(char *, char *, int *, int *, float  *,
+        float  *, int *, float *, float  *, int *, float *, int *);
+int BLASFUNC(zgems)(char *, char *, int *, int *, double *,
+        double *, int *, double*, double *, int *, double*, int *);
+
+int BLASFUNC(sgetf2)(int *, int *, float  *, int *, int *, int *);
+int BLASFUNC(dgetf2)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(qgetf2)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(cgetf2)(int *, int *, float  *, int *, int *, int *);
+int BLASFUNC(zgetf2)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(xgetf2)(int *, int *, double *, int *, int *, int *);
+
+int BLASFUNC(sgetrf)(int *, int *, float  *, int *, int *, int *);
+int BLASFUNC(dgetrf)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(qgetrf)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(cgetrf)(int *, int *, float  *, int *, int *, int *);
+int BLASFUNC(zgetrf)(int *, int *, double *, int *, int *, int *);
+int BLASFUNC(xgetrf)(int *, int *, double *, int *, int *, int *);
+
+int BLASFUNC(slaswp)(int *, float  *, int *, int *, int *, int *, int *);
+int BLASFUNC(dlaswp)(int *, double *, int *, int *, int *, int *, int *);
+int BLASFUNC(qlaswp)(int *, double *, int *, int *, int *, int *, int *);
+int BLASFUNC(claswp)(int *, float  *, int *, int *, int *, int *, int *);
+int BLASFUNC(zlaswp)(int *, double *, int *, int *, int *, int *, int *);
+int BLASFUNC(xlaswp)(int *, double *, int *, int *, int *, int *, int *);
+
+int BLASFUNC(sgetrs)(char *, int *, int *, float  *, int *, int *, float  *, int *, int *);
+int BLASFUNC(dgetrs)(char *, int *, int *, double *, int *, int *, double *, int *, int *);
+int BLASFUNC(qgetrs)(char *, int *, int *, double *, int *, int *, double *, int *, int *);
+int BLASFUNC(cgetrs)(char *, int *, int *, float  *, int *, int *, float  *, int *, int *);
+int BLASFUNC(zgetrs)(char *, int *, int *, double *, int *, int *, double *, int *, int *);
+int BLASFUNC(xgetrs)(char *, int *, int *, double *, int *, int *, double *, int *, int *);
+
+int BLASFUNC(sgesv)(int *, int *, float  *, int *, int *, float *, int *, int *);
+int BLASFUNC(dgesv)(int *, int *, double *, int *, int *, double*, int *, int *);
+int BLASFUNC(qgesv)(int *, int *, double *, int *, int *, double*, int *, int *);
+int BLASFUNC(cgesv)(int *, int *, float  *, int *, int *, float *, int *, int *);
+int BLASFUNC(zgesv)(int *, int *, double *, int *, int *, double*, int *, int *);
+int BLASFUNC(xgesv)(int *, int *, double *, int *, int *, double*, int *, int *);
+
+int BLASFUNC(spotf2)(char *, int *, float  *, int *, int *);
+int BLASFUNC(dpotf2)(char *, int *, double *, int *, int *);
+int BLASFUNC(qpotf2)(char *, int *, double *, int *, int *);
+int BLASFUNC(cpotf2)(char *, int *, float  *, int *, int *);
+int BLASFUNC(zpotf2)(char *, int *, double *, int *, int *);
+int BLASFUNC(xpotf2)(char *, int *, double *, int *, int *);
+
+int BLASFUNC(spotrf)(char *, int *, float  *, int *, int *);
+int BLASFUNC(dpotrf)(char *, int *, double *, int *, int *);
+int BLASFUNC(qpotrf)(char *, int *, double *, int *, int *);
+int BLASFUNC(cpotrf)(char *, int *, float  *, int *, int *);
+int BLASFUNC(zpotrf)(char *, int *, double *, int *, int *);
+int BLASFUNC(xpotrf)(char *, int *, double *, int *, int *);
+
+int BLASFUNC(slauu2)(char *, int *, float  *, int *, int *);
+int BLASFUNC(dlauu2)(char *, int *, double *, int *, int *);
+int BLASFUNC(qlauu2)(char *, int *, double *, int *, int *);
+int BLASFUNC(clauu2)(char *, int *, float  *, int *, int *);
+int BLASFUNC(zlauu2)(char *, int *, double *, int *, int *);
+int BLASFUNC(xlauu2)(char *, int *, double *, int *, int *);
+
+int BLASFUNC(slauum)(char *, int *, float  *, int *, int *);
+int BLASFUNC(dlauum)(char *, int *, double *, int *, int *);
+int BLASFUNC(qlauum)(char *, int *, double *, int *, int *);
+int BLASFUNC(clauum)(char *, int *, float  *, int *, int *);
+int BLASFUNC(zlauum)(char *, int *, double *, int *, int *);
+int BLASFUNC(xlauum)(char *, int *, double *, int *, int *);
+
+int BLASFUNC(strti2)(char *, char *, int *, float  *, int *, int *);
+int BLASFUNC(dtrti2)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(qtrti2)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(ctrti2)(char *, char *, int *, float  *, int *, int *);
+int BLASFUNC(ztrti2)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(xtrti2)(char *, char *, int *, double *, int *, int *);
+
+int BLASFUNC(strtri)(char *, char *, int *, float  *, int *, int *);
+int BLASFUNC(dtrtri)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(qtrtri)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(ctrtri)(char *, char *, int *, float  *, int *, int *);
+int BLASFUNC(ztrtri)(char *, char *, int *, double *, int *, int *);
+int BLASFUNC(xtrtri)(char *, char *, int *, double *, int *, int *);
+
+int BLASFUNC(spotri)(char *, int *, float  *, int *, int *);
+int BLASFUNC(dpotri)(char *, int *, double *, int *, int *);
+int BLASFUNC(qpotri)(char *, int *, double *, int *, int *);
+int BLASFUNC(cpotri)(char *, int *, float  *, int *, int *);
+int BLASFUNC(zpotri)(char *, int *, double *, int *, int *);
+int BLASFUNC(xpotri)(char *, int *, double *, int *, int *);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke.h
new file mode 100755
index 0000000..8c7e79b
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke.h
@@ -0,0 +1,16291 @@
+/*****************************************************************************
+  Copyright (c) 2010, Intel Corp.
+  All rights reserved.
+
+  Redistribution and use in source and binary forms, with or without
+  modification, are permitted provided that the following conditions are met:
+
+    * Redistributions of source code must retain the above copyright notice,
+      this list of conditions and the following disclaimer.
+    * Redistributions in binary form must reproduce the above copyright
+      notice, this list of conditions and the following disclaimer in the
+      documentation and/or other materials provided with the distribution.
+    * Neither the name of Intel Corporation nor the names of its contributors
+      may be used to endorse or promote products derived from this software
+      without specific prior written permission.
+
+  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+  AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+  IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+  ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+  LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+  CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+  SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+  INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+  CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+  ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+  THE POSSIBILITY OF SUCH DAMAGE.
+******************************************************************************
+* Contents: Native C interface to LAPACK
+* Author: Intel Corporation
+* Generated November, 2011
+*****************************************************************************/
+
+#ifndef _MKL_LAPACKE_H_
+
+#ifndef _LAPACKE_H_
+#define _LAPACKE_H_
+
+/*
+*  Turn on HAVE_LAPACK_CONFIG_H to redefine C-LAPACK datatypes
+*/
+#ifdef HAVE_LAPACK_CONFIG_H
+#include "lapacke_config.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif /* __cplusplus */
+
+#include <stdlib.h>
+
+#ifndef lapack_int
+#define lapack_int     int
+#endif
+
+#ifndef lapack_logical
+#define lapack_logical lapack_int
+#endif
+
+/* Complex types are structures equivalent to the
+* Fortran complex types COMPLEX(4) and COMPLEX(8).
+*
+* One can also redefine the types with his own types
+* for example by including in the code definitions like
+*
+* #define lapack_complex_float std::complex<float>
+* #define lapack_complex_double std::complex<double>
+*
+* or define these types in the command line:
+*
+* -Dlapack_complex_float="std::complex<float>"
+* -Dlapack_complex_double="std::complex<double>"
+*/
+
+#ifndef LAPACK_COMPLEX_CUSTOM
+
+/* Complex type (single precision) */
+#ifndef lapack_complex_float
+#include <complex.h>
+#define lapack_complex_float    float _Complex
+#endif
+
+#ifndef lapack_complex_float_real
+#define lapack_complex_float_real(z)       (creal(z))
+#endif
+
+#ifndef lapack_complex_float_imag
+#define lapack_complex_float_imag(z)       (cimag(z))
+#endif
+
+lapack_complex_float lapack_make_complex_float( float re, float im );
+
+/* Complex type (double precision) */
+#ifndef lapack_complex_double
+#include <complex.h>
+#define lapack_complex_double   double _Complex
+#endif
+
+#ifndef lapack_complex_double_real
+#define lapack_complex_double_real(z)      (creal(z))
+#endif
+
+#ifndef lapack_complex_double_imag
+#define lapack_complex_double_imag(z)       (cimag(z))
+#endif
+
+lapack_complex_double lapack_make_complex_double( double re, double im );
+
+#endif
+
+#ifndef LAPACKE_malloc
+#define LAPACKE_malloc( size ) malloc( size )
+#endif
+#ifndef LAPACKE_free
+#define LAPACKE_free( p )      free( p )
+#endif
+
+#define LAPACK_C2INT( x ) (lapack_int)(*((float*)&x ))
+#define LAPACK_Z2INT( x ) (lapack_int)(*((double*)&x ))
+
+#define LAPACK_ROW_MAJOR               101
+#define LAPACK_COL_MAJOR               102
+
+#define LAPACK_WORK_MEMORY_ERROR       -1010
+#define LAPACK_TRANSPOSE_MEMORY_ERROR  -1011
+
+/* Callback logical functions of one, two, or three arguments are used
+*  to select eigenvalues to sort to the top left of the Schur form.
+*  The value is selected if function returns TRUE (non-zero). */
+
+typedef lapack_logical (*LAPACK_S_SELECT2) ( const float*, const float* );
+typedef lapack_logical (*LAPACK_S_SELECT3)
+    ( const float*, const float*, const float* );
+typedef lapack_logical (*LAPACK_D_SELECT2) ( const double*, const double* );
+typedef lapack_logical (*LAPACK_D_SELECT3)
+    ( const double*, const double*, const double* );
+
+typedef lapack_logical (*LAPACK_C_SELECT1) ( const lapack_complex_float* );
+typedef lapack_logical (*LAPACK_C_SELECT2)
+    ( const lapack_complex_float*, const lapack_complex_float* );
+typedef lapack_logical (*LAPACK_Z_SELECT1) ( const lapack_complex_double* );
+typedef lapack_logical (*LAPACK_Z_SELECT2)
+    ( const lapack_complex_double*, const lapack_complex_double* );
+
+#include "lapacke_mangling.h"
+
+#define LAPACK_lsame LAPACK_GLOBAL(lsame,LSAME)
+lapack_logical LAPACK_lsame( char* ca,  char* cb,
+                              lapack_int lca, lapack_int lcb );
+
+/* C-LAPACK function prototypes */
+
+lapack_int LAPACKE_sbdsdc( int matrix_order, char uplo, char compq,
+                           lapack_int n, float* d, float* e, float* u,
+                           lapack_int ldu, float* vt, lapack_int ldvt, float* q,
+                           lapack_int* iq );
+lapack_int LAPACKE_dbdsdc( int matrix_order, char uplo, char compq,
+                           lapack_int n, double* d, double* e, double* u,
+                           lapack_int ldu, double* vt, lapack_int ldvt,
+                           double* q, lapack_int* iq );
+
+lapack_int LAPACKE_sbdsqr( int matrix_order, char uplo, lapack_int n,
+                           lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                           float* d, float* e, float* vt, lapack_int ldvt,
+                           float* u, lapack_int ldu, float* c, lapack_int ldc );
+lapack_int LAPACKE_dbdsqr( int matrix_order, char uplo, lapack_int n,
+                           lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                           double* d, double* e, double* vt, lapack_int ldvt,
+                           double* u, lapack_int ldu, double* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_cbdsqr( int matrix_order, char uplo, lapack_int n,
+                           lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                           float* d, float* e, lapack_complex_float* vt,
+                           lapack_int ldvt, lapack_complex_float* u,
+                           lapack_int ldu, lapack_complex_float* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_zbdsqr( int matrix_order, char uplo, lapack_int n,
+                           lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                           double* d, double* e, lapack_complex_double* vt,
+                           lapack_int ldvt, lapack_complex_double* u,
+                           lapack_int ldu, lapack_complex_double* c,
+                           lapack_int ldc );
+
+lapack_int LAPACKE_sdisna( char job, lapack_int m, lapack_int n, const float* d,
+                           float* sep );
+lapack_int LAPACKE_ddisna( char job, lapack_int m, lapack_int n,
+                           const double* d, double* sep );
+
+lapack_int LAPACKE_sgbbrd( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int ncc, lapack_int kl,
+                           lapack_int ku, float* ab, lapack_int ldab, float* d,
+                           float* e, float* q, lapack_int ldq, float* pt,
+                           lapack_int ldpt, float* c, lapack_int ldc );
+lapack_int LAPACKE_dgbbrd( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int ncc, lapack_int kl,
+                           lapack_int ku, double* ab, lapack_int ldab,
+                           double* d, double* e, double* q, lapack_int ldq,
+                           double* pt, lapack_int ldpt, double* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_cgbbrd( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int ncc, lapack_int kl,
+                           lapack_int ku, lapack_complex_float* ab,
+                           lapack_int ldab, float* d, float* e,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_complex_float* pt, lapack_int ldpt,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zgbbrd( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int ncc, lapack_int kl,
+                           lapack_int ku, lapack_complex_double* ab,
+                           lapack_int ldab, double* d, double* e,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* pt, lapack_int ldpt,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sgbcon( int matrix_order, char norm, lapack_int n,
+                           lapack_int kl, lapack_int ku, const float* ab,
+                           lapack_int ldab, const lapack_int* ipiv, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_dgbcon( int matrix_order, char norm, lapack_int n,
+                           lapack_int kl, lapack_int ku, const double* ab,
+                           lapack_int ldab, const lapack_int* ipiv,
+                           double anorm, double* rcond );
+lapack_int LAPACKE_cgbcon( int matrix_order, char norm, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zgbcon( int matrix_order, char norm, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_sgbequ( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const float* ab,
+                           lapack_int ldab, float* r, float* c, float* rowcnd,
+                           float* colcnd, float* amax );
+lapack_int LAPACKE_dgbequ( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const double* ab,
+                           lapack_int ldab, double* r, double* c,
+                           double* rowcnd, double* colcnd, double* amax );
+lapack_int LAPACKE_cgbequ( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           float* r, float* c, float* rowcnd, float* colcnd,
+                           float* amax );
+lapack_int LAPACKE_zgbequ( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           double* r, double* c, double* rowcnd, double* colcnd,
+                           double* amax );
+
+lapack_int LAPACKE_sgbequb( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_int kl, lapack_int ku, const float* ab,
+                            lapack_int ldab, float* r, float* c, float* rowcnd,
+                            float* colcnd, float* amax );
+lapack_int LAPACKE_dgbequb( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_int kl, lapack_int ku, const double* ab,
+                            lapack_int ldab, double* r, double* c,
+                            double* rowcnd, double* colcnd, double* amax );
+lapack_int LAPACKE_cgbequb( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_int kl, lapack_int ku,
+                            const lapack_complex_float* ab, lapack_int ldab,
+                            float* r, float* c, float* rowcnd, float* colcnd,
+                            float* amax );
+lapack_int LAPACKE_zgbequb( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_int kl, lapack_int ku,
+                            const lapack_complex_double* ab, lapack_int ldab,
+                            double* r, double* c, double* rowcnd,
+                            double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgbrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const float* ab, lapack_int ldab, const float* afb,
+                           lapack_int ldafb, const lapack_int* ipiv,
+                           const float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_dgbrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const double* ab, lapack_int ldab, const double* afb,
+                           lapack_int ldafb, const lapack_int* ipiv,
+                           const double* b, lapack_int ldb, double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+lapack_int LAPACKE_cgbrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_complex_float* afb, lapack_int ldafb,
+                           const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zgbrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_complex_double* afb, lapack_int ldafb,
+                           const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sgbrfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, const float* ab, lapack_int ldab,
+                            const float* afb, lapack_int ldafb,
+                            const lapack_int* ipiv, const float* r,
+                            const float* c, const float* b, lapack_int ldb,
+                            float* x, lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dgbrfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, const double* ab, lapack_int ldab,
+                            const double* afb, lapack_int ldafb,
+                            const lapack_int* ipiv, const double* r,
+                            const double* c, const double* b, lapack_int ldb,
+                            double* x, lapack_int ldx, double* rcond,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+lapack_int LAPACKE_cgbrfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, const lapack_complex_float* ab,
+                            lapack_int ldab, const lapack_complex_float* afb,
+                            lapack_int ldafb, const lapack_int* ipiv,
+                            const float* r, const float* c,
+                            const lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_zgbrfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, const lapack_complex_double* ab,
+                            lapack_int ldab, const lapack_complex_double* afb,
+                            lapack_int ldafb, const lapack_int* ipiv,
+                            const double* r, const double* c,
+                            const lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+
+lapack_int LAPACKE_sgbsv( int matrix_order, lapack_int n, lapack_int kl,
+                          lapack_int ku, lapack_int nrhs, float* ab,
+                          lapack_int ldab, lapack_int* ipiv, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dgbsv( int matrix_order, lapack_int n, lapack_int kl,
+                          lapack_int ku, lapack_int nrhs, double* ab,
+                          lapack_int ldab, lapack_int* ipiv, double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_cgbsv( int matrix_order, lapack_int n, lapack_int kl,
+                          lapack_int ku, lapack_int nrhs,
+                          lapack_complex_float* ab, lapack_int ldab,
+                          lapack_int* ipiv, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zgbsv( int matrix_order, lapack_int n, lapack_int kl,
+                          lapack_int ku, lapack_int nrhs,
+                          lapack_complex_double* ab, lapack_int ldab,
+                          lapack_int* ipiv, lapack_complex_double* b,
+                          lapack_int ldb );
+
+lapack_int LAPACKE_sgbsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int kl, lapack_int ku,
+                           lapack_int nrhs, float* ab, lapack_int ldab,
+                           float* afb, lapack_int ldafb, lapack_int* ipiv,
+                           char* equed, float* r, float* c, float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr,
+                           float* rpivot );
+lapack_int LAPACKE_dgbsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int kl, lapack_int ku,
+                           lapack_int nrhs, double* ab, lapack_int ldab,
+                           double* afb, lapack_int ldafb, lapack_int* ipiv,
+                           char* equed, double* r, double* c, double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr,
+                           double* rpivot );
+lapack_int LAPACKE_cgbsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int kl, lapack_int ku,
+                           lapack_int nrhs, lapack_complex_float* ab,
+                           lapack_int ldab, lapack_complex_float* afb,
+                           lapack_int ldafb, lapack_int* ipiv, char* equed,
+                           float* r, float* c, lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr, float* rpivot );
+lapack_int LAPACKE_zgbsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int kl, lapack_int ku,
+                           lapack_int nrhs, lapack_complex_double* ab,
+                           lapack_int ldab, lapack_complex_double* afb,
+                           lapack_int ldafb, lapack_int* ipiv, char* equed,
+                           double* r, double* c, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* x,
+                           lapack_int ldx, double* rcond, double* ferr,
+                           double* berr, double* rpivot );
+
+lapack_int LAPACKE_sgbsvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, float* ab, lapack_int ldab,
+                            float* afb, lapack_int ldafb, lapack_int* ipiv,
+                            char* equed, float* r, float* c, float* b,
+                            lapack_int ldb, float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dgbsvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, double* ab, lapack_int ldab,
+                            double* afb, lapack_int ldafb, lapack_int* ipiv,
+                            char* equed, double* r, double* c, double* b,
+                            lapack_int ldb, double* x, lapack_int ldx,
+                            double* rcond, double* rpvgrw, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+lapack_int LAPACKE_cgbsvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, lapack_complex_float* ab,
+                            lapack_int ldab, lapack_complex_float* afb,
+                            lapack_int ldafb, lapack_int* ipiv, char* equed,
+                            float* r, float* c, lapack_complex_float* b,
+                            lapack_int ldb, lapack_complex_float* x,
+                            lapack_int ldx, float* rcond, float* rpvgrw,
+                            float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_zgbsvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int kl, lapack_int ku,
+                            lapack_int nrhs, lapack_complex_double* ab,
+                            lapack_int ldab, lapack_complex_double* afb,
+                            lapack_int ldafb, lapack_int* ipiv, char* equed,
+                            double* r, double* c, lapack_complex_double* b,
+                            lapack_int ldb, lapack_complex_double* x,
+                            lapack_int ldx, double* rcond, double* rpvgrw,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+
+lapack_int LAPACKE_sgbtrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, float* ab,
+                           lapack_int ldab, lapack_int* ipiv );
+lapack_int LAPACKE_dgbtrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, double* ab,
+                           lapack_int ldab, lapack_int* ipiv );
+lapack_int LAPACKE_cgbtrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_zgbtrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           lapack_int* ipiv );
+
+lapack_int LAPACKE_sgbtrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const float* ab, lapack_int ldab,
+                           const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dgbtrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const double* ab, lapack_int ldab,
+                           const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_cgbtrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_int* ipiv, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zgbtrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int kl, lapack_int ku, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_int* ipiv, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_sgebak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const float* scale,
+                           lapack_int m, float* v, lapack_int ldv );
+lapack_int LAPACKE_dgebak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const double* scale,
+                           lapack_int m, double* v, lapack_int ldv );
+lapack_int LAPACKE_cgebak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const float* scale,
+                           lapack_int m, lapack_complex_float* v,
+                           lapack_int ldv );
+lapack_int LAPACKE_zgebak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const double* scale,
+                           lapack_int m, lapack_complex_double* v,
+                           lapack_int ldv );
+
+lapack_int LAPACKE_sgebal( int matrix_order, char job, lapack_int n, float* a,
+                           lapack_int lda, lapack_int* ilo, lapack_int* ihi,
+                           float* scale );
+lapack_int LAPACKE_dgebal( int matrix_order, char job, lapack_int n, double* a,
+                           lapack_int lda, lapack_int* ilo, lapack_int* ihi,
+                           double* scale );
+lapack_int LAPACKE_cgebal( int matrix_order, char job, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* ilo, lapack_int* ihi, float* scale );
+lapack_int LAPACKE_zgebal( int matrix_order, char job, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ilo, lapack_int* ihi, double* scale );
+
+lapack_int LAPACKE_sgebrd( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* d, float* e,
+                           float* tauq, float* taup );
+lapack_int LAPACKE_dgebrd( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* d, double* e,
+                           double* tauq, double* taup );
+lapack_int LAPACKE_cgebrd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda, float* d,
+                           float* e, lapack_complex_float* tauq,
+                           lapack_complex_float* taup );
+lapack_int LAPACKE_zgebrd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda, double* d,
+                           double* e, lapack_complex_double* tauq,
+                           lapack_complex_double* taup );
+
+lapack_int LAPACKE_sgecon( int matrix_order, char norm, lapack_int n,
+                           const float* a, lapack_int lda, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_dgecon( int matrix_order, char norm, lapack_int n,
+                           const double* a, lapack_int lda, double anorm,
+                           double* rcond );
+lapack_int LAPACKE_cgecon( int matrix_order, char norm, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           float anorm, float* rcond );
+lapack_int LAPACKE_zgecon( int matrix_order, char norm, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           double anorm, double* rcond );
+
+lapack_int LAPACKE_sgeequ( int matrix_order, lapack_int m, lapack_int n,
+                           const float* a, lapack_int lda, float* r, float* c,
+                           float* rowcnd, float* colcnd, float* amax );
+lapack_int LAPACKE_dgeequ( int matrix_order, lapack_int m, lapack_int n,
+                           const double* a, lapack_int lda, double* r,
+                           double* c, double* rowcnd, double* colcnd,
+                           double* amax );
+lapack_int LAPACKE_cgeequ( int matrix_order, lapack_int m, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           float* r, float* c, float* rowcnd, float* colcnd,
+                           float* amax );
+lapack_int LAPACKE_zgeequ( int matrix_order, lapack_int m, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           double* r, double* c, double* rowcnd, double* colcnd,
+                           double* amax );
+
+lapack_int LAPACKE_sgeequb( int matrix_order, lapack_int m, lapack_int n,
+                            const float* a, lapack_int lda, float* r, float* c,
+                            float* rowcnd, float* colcnd, float* amax );
+lapack_int LAPACKE_dgeequb( int matrix_order, lapack_int m, lapack_int n,
+                            const double* a, lapack_int lda, double* r,
+                            double* c, double* rowcnd, double* colcnd,
+                            double* amax );
+lapack_int LAPACKE_cgeequb( int matrix_order, lapack_int m, lapack_int n,
+                            const lapack_complex_float* a, lapack_int lda,
+                            float* r, float* c, float* rowcnd, float* colcnd,
+                            float* amax );
+lapack_int LAPACKE_zgeequb( int matrix_order, lapack_int m, lapack_int n,
+                            const lapack_complex_double* a, lapack_int lda,
+                            double* r, double* c, double* rowcnd,
+                            double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgees( int matrix_order, char jobvs, char sort,
+                          LAPACK_S_SELECT2 select, lapack_int n, float* a,
+                          lapack_int lda, lapack_int* sdim, float* wr,
+                          float* wi, float* vs, lapack_int ldvs );
+lapack_int LAPACKE_dgees( int matrix_order, char jobvs, char sort,
+                          LAPACK_D_SELECT2 select, lapack_int n, double* a,
+                          lapack_int lda, lapack_int* sdim, double* wr,
+                          double* wi, double* vs, lapack_int ldvs );
+lapack_int LAPACKE_cgees( int matrix_order, char jobvs, char sort,
+                          LAPACK_C_SELECT1 select, lapack_int n,
+                          lapack_complex_float* a, lapack_int lda,
+                          lapack_int* sdim, lapack_complex_float* w,
+                          lapack_complex_float* vs, lapack_int ldvs );
+lapack_int LAPACKE_zgees( int matrix_order, char jobvs, char sort,
+                          LAPACK_Z_SELECT1 select, lapack_int n,
+                          lapack_complex_double* a, lapack_int lda,
+                          lapack_int* sdim, lapack_complex_double* w,
+                          lapack_complex_double* vs, lapack_int ldvs );
+
+lapack_int LAPACKE_sgeesx( int matrix_order, char jobvs, char sort,
+                           LAPACK_S_SELECT2 select, char sense, lapack_int n,
+                           float* a, lapack_int lda, lapack_int* sdim,
+                           float* wr, float* wi, float* vs, lapack_int ldvs,
+                           float* rconde, float* rcondv );
+lapack_int LAPACKE_dgeesx( int matrix_order, char jobvs, char sort,
+                           LAPACK_D_SELECT2 select, char sense, lapack_int n,
+                           double* a, lapack_int lda, lapack_int* sdim,
+                           double* wr, double* wi, double* vs, lapack_int ldvs,
+                           double* rconde, double* rcondv );
+lapack_int LAPACKE_cgeesx( int matrix_order, char jobvs, char sort,
+                           LAPACK_C_SELECT1 select, char sense, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* sdim, lapack_complex_float* w,
+                           lapack_complex_float* vs, lapack_int ldvs,
+                           float* rconde, float* rcondv );
+lapack_int LAPACKE_zgeesx( int matrix_order, char jobvs, char sort,
+                           LAPACK_Z_SELECT1 select, char sense, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* sdim, lapack_complex_double* w,
+                           lapack_complex_double* vs, lapack_int ldvs,
+                           double* rconde, double* rcondv );
+
+lapack_int LAPACKE_sgeev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, float* a, lapack_int lda, float* wr,
+                          float* wi, float* vl, lapack_int ldvl, float* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_dgeev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, double* a, lapack_int lda, double* wr,
+                          double* wi, double* vl, lapack_int ldvl, double* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_cgeev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, lapack_complex_float* a, lapack_int lda,
+                          lapack_complex_float* w, lapack_complex_float* vl,
+                          lapack_int ldvl, lapack_complex_float* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_zgeev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, lapack_complex_double* a,
+                          lapack_int lda, lapack_complex_double* w,
+                          lapack_complex_double* vl, lapack_int ldvl,
+                          lapack_complex_double* vr, lapack_int ldvr );
+
+lapack_int LAPACKE_sgeevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n, float* a,
+                           lapack_int lda, float* wr, float* wi, float* vl,
+                           lapack_int ldvl, float* vr, lapack_int ldvr,
+                           lapack_int* ilo, lapack_int* ihi, float* scale,
+                           float* abnrm, float* rconde, float* rcondv );
+lapack_int LAPACKE_dgeevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n, double* a,
+                           lapack_int lda, double* wr, double* wi, double* vl,
+                           lapack_int ldvl, double* vr, lapack_int ldvr,
+                           lapack_int* ilo, lapack_int* ihi, double* scale,
+                           double* abnrm, double* rconde, double* rcondv );
+lapack_int LAPACKE_cgeevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* w, lapack_complex_float* vl,
+                           lapack_int ldvl, lapack_complex_float* vr,
+                           lapack_int ldvr, lapack_int* ilo, lapack_int* ihi,
+                           float* scale, float* abnrm, float* rconde,
+                           float* rcondv );
+lapack_int LAPACKE_zgeevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* w, lapack_complex_double* vl,
+                           lapack_int ldvl, lapack_complex_double* vr,
+                           lapack_int ldvr, lapack_int* ilo, lapack_int* ihi,
+                           double* scale, double* abnrm, double* rconde,
+                           double* rcondv );
+
+lapack_int LAPACKE_sgehrd( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, float* a, lapack_int lda,
+                           float* tau );
+lapack_int LAPACKE_dgehrd( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, double* a, lapack_int lda,
+                           double* tau );
+lapack_int LAPACKE_cgehrd( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* tau );
+lapack_int LAPACKE_zgehrd( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgejsv( int matrix_order, char joba, char jobu, char jobv,
+                           char jobr, char jobt, char jobp, lapack_int m,
+                           lapack_int n, float* a, lapack_int lda, float* sva,
+                           float* u, lapack_int ldu, float* v, lapack_int ldv,
+                           float* stat, lapack_int* istat );
+lapack_int LAPACKE_dgejsv( int matrix_order, char joba, char jobu, char jobv,
+                           char jobr, char jobt, char jobp, lapack_int m,
+                           lapack_int n, double* a, lapack_int lda, double* sva,
+                           double* u, lapack_int ldu, double* v, lapack_int ldv,
+                           double* stat, lapack_int* istat );
+
+lapack_int LAPACKE_sgelq2( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgelq2( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgelq2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgelq2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgelqf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgelqf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgelqf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgelqf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgels( int matrix_order, char trans, lapack_int m,
+                          lapack_int n, lapack_int nrhs, float* a,
+                          lapack_int lda, float* b, lapack_int ldb );
+lapack_int LAPACKE_dgels( int matrix_order, char trans, lapack_int m,
+                          lapack_int n, lapack_int nrhs, double* a,
+                          lapack_int lda, double* b, lapack_int ldb );
+lapack_int LAPACKE_cgels( int matrix_order, char trans, lapack_int m,
+                          lapack_int n, lapack_int nrhs,
+                          lapack_complex_float* a, lapack_int lda,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zgels( int matrix_order, char trans, lapack_int m,
+                          lapack_int n, lapack_int nrhs,
+                          lapack_complex_double* a, lapack_int lda,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sgelsd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, float* s, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_dgelsd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, double* a, lapack_int lda,
+                           double* b, lapack_int ldb, double* s, double rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_cgelsd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, float* s, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_zgelsd( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, double* s, double rcond,
+                           lapack_int* rank );
+
+lapack_int LAPACKE_sgelss( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, float* s, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_dgelss( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, double* a, lapack_int lda,
+                           double* b, lapack_int ldb, double* s, double rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_cgelss( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, float* s, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_zgelss( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, double* s, double rcond,
+                           lapack_int* rank );
+
+lapack_int LAPACKE_sgelsy( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, lapack_int* jpvt, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_dgelsy( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, double* a, lapack_int lda,
+                           double* b, lapack_int ldb, lapack_int* jpvt,
+                           double rcond, lapack_int* rank );
+lapack_int LAPACKE_cgelsy( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, lapack_int* jpvt, float rcond,
+                           lapack_int* rank );
+lapack_int LAPACKE_zgelsy( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, lapack_int* jpvt, double rcond,
+                           lapack_int* rank );
+
+lapack_int LAPACKE_sgeqlf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgeqlf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgeqlf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqlf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgeqp3( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, lapack_int* jpvt,
+                           float* tau );
+lapack_int LAPACKE_dgeqp3( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, lapack_int* jpvt,
+                           double* tau );
+lapack_int LAPACKE_cgeqp3( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* jpvt, lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqp3( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* jpvt, lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgeqpf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, lapack_int* jpvt,
+                           float* tau );
+lapack_int LAPACKE_dgeqpf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, lapack_int* jpvt,
+                           double* tau );
+lapack_int LAPACKE_cgeqpf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* jpvt, lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqpf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* jpvt, lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgeqr2( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgeqr2( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgeqr2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqr2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgeqrf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgeqrf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgeqrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgeqrfp( int matrix_order, lapack_int m, lapack_int n,
+                            float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgeqrfp( int matrix_order, lapack_int m, lapack_int n,
+                            double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgeqrfp( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* tau );
+lapack_int LAPACKE_zgeqrfp( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgerfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           const float* af, lapack_int ldaf,
+                           const lapack_int* ipiv, const float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dgerfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           const double* af, lapack_int ldaf,
+                           const lapack_int* ipiv, const double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cgerfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zgerfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sgerfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int nrhs, const float* a,
+                            lapack_int lda, const float* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const float* r,
+                            const float* c, const float* b, lapack_int ldb,
+                            float* x, lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dgerfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int nrhs, const double* a,
+                            lapack_int lda, const double* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const double* r,
+                            const double* c, const double* b, lapack_int ldb,
+                            double* x, lapack_int ldx, double* rcond,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+lapack_int LAPACKE_cgerfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_float* a, lapack_int lda,
+                            const lapack_complex_float* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const float* r,
+                            const float* c, const lapack_complex_float* b,
+                            lapack_int ldb, lapack_complex_float* x,
+                            lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_zgerfsx( int matrix_order, char trans, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_double* a, lapack_int lda,
+                            const lapack_complex_double* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const double* r,
+                            const double* c, const lapack_complex_double* b,
+                            lapack_int ldb, lapack_complex_double* x,
+                            lapack_int ldx, double* rcond, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+
+lapack_int LAPACKE_sgerqf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dgerqf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_cgerqf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zgerqf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_sgesdd( int matrix_order, char jobz, lapack_int m,
+                           lapack_int n, float* a, lapack_int lda, float* s,
+                           float* u, lapack_int ldu, float* vt,
+                           lapack_int ldvt );
+lapack_int LAPACKE_dgesdd( int matrix_order, char jobz, lapack_int m,
+                           lapack_int n, double* a, lapack_int lda, double* s,
+                           double* u, lapack_int ldu, double* vt,
+                           lapack_int ldvt );
+lapack_int LAPACKE_cgesdd( int matrix_order, char jobz, lapack_int m,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, float* s, lapack_complex_float* u,
+                           lapack_int ldu, lapack_complex_float* vt,
+                           lapack_int ldvt );
+lapack_int LAPACKE_zgesdd( int matrix_order, char jobz, lapack_int m,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, double* s, lapack_complex_double* u,
+                           lapack_int ldu, lapack_complex_double* vt,
+                           lapack_int ldvt );
+
+lapack_int LAPACKE_sgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          float* a, lapack_int lda, lapack_int* ipiv, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          double* a, lapack_int lda, lapack_int* ipiv,
+                          double* b, lapack_int ldb );
+lapack_int LAPACKE_cgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          lapack_complex_float* a, lapack_int lda,
+                          lapack_int* ipiv, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          lapack_complex_double* a, lapack_int lda,
+                          lapack_int* ipiv, lapack_complex_double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dsgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                           double* a, lapack_int lda, lapack_int* ipiv,
+                           double* b, lapack_int ldb, double* x, lapack_int ldx,
+                           lapack_int* iter );
+lapack_int LAPACKE_zcgesv( int matrix_order, lapack_int n, lapack_int nrhs,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ipiv, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* x,
+                           lapack_int ldx, lapack_int* iter );
+
+lapack_int LAPACKE_sgesvd( int matrix_order, char jobu, char jobvt,
+                           lapack_int m, lapack_int n, float* a, lapack_int lda,
+                           float* s, float* u, lapack_int ldu, float* vt,
+                           lapack_int ldvt, float* superb );
+lapack_int LAPACKE_dgesvd( int matrix_order, char jobu, char jobvt,
+                           lapack_int m, lapack_int n, double* a,
+                           lapack_int lda, double* s, double* u, lapack_int ldu,
+                           double* vt, lapack_int ldvt, double* superb );
+lapack_int LAPACKE_cgesvd( int matrix_order, char jobu, char jobvt,
+                           lapack_int m, lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, float* s, lapack_complex_float* u,
+                           lapack_int ldu, lapack_complex_float* vt,
+                           lapack_int ldvt, float* superb );
+lapack_int LAPACKE_zgesvd( int matrix_order, char jobu, char jobvt,
+                           lapack_int m, lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, double* s, lapack_complex_double* u,
+                           lapack_int ldu, lapack_complex_double* vt,
+                           lapack_int ldvt, double* superb );
+
+lapack_int LAPACKE_sgesvj( int matrix_order, char joba, char jobu, char jobv,
+                           lapack_int m, lapack_int n, float* a, lapack_int lda,
+                           float* sva, lapack_int mv, float* v, lapack_int ldv,
+                           float* stat );
+lapack_int LAPACKE_dgesvj( int matrix_order, char joba, char jobu, char jobv,
+                           lapack_int m, lapack_int n, double* a,
+                           lapack_int lda, double* sva, lapack_int mv,
+                           double* v, lapack_int ldv, double* stat );
+
+lapack_int LAPACKE_sgesvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs, float* a,
+                           lapack_int lda, float* af, lapack_int ldaf,
+                           lapack_int* ipiv, char* equed, float* r, float* c,
+                           float* b, lapack_int ldb, float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr,
+                           float* rpivot );
+lapack_int LAPACKE_dgesvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs, double* a,
+                           lapack_int lda, double* af, lapack_int ldaf,
+                           lapack_int* ipiv, char* equed, double* r, double* c,
+                           double* b, lapack_int ldb, double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr,
+                           double* rpivot );
+lapack_int LAPACKE_cgesvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* af, lapack_int ldaf,
+                           lapack_int* ipiv, char* equed, float* r, float* c,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr,
+                           float* rpivot );
+lapack_int LAPACKE_zgesvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* af, lapack_int ldaf,
+                           lapack_int* ipiv, char* equed, double* r, double* c,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr,
+                           double* rpivot );
+
+lapack_int LAPACKE_sgesvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int nrhs, float* a,
+                            lapack_int lda, float* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, float* r, float* c,
+                            float* b, lapack_int ldb, float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dgesvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int nrhs, double* a,
+                            lapack_int lda, double* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, double* r, double* c,
+                            double* b, lapack_int ldb, double* x,
+                            lapack_int ldx, double* rcond, double* rpvgrw,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+lapack_int LAPACKE_cgesvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, float* r, float* c,
+                            lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_zgesvxx( int matrix_order, char fact, char trans,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, double* r, double* c,
+                            lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* rpvgrw, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+
+lapack_int LAPACKE_sgetf2( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_dgetf2( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_cgetf2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_zgetf2( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ipiv );
+
+lapack_int LAPACKE_sgetrf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_dgetrf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_cgetrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_zgetrf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ipiv );
+
+lapack_int LAPACKE_sgetri( int matrix_order, lapack_int n, float* a,
+                           lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_dgetri( int matrix_order, lapack_int n, double* a,
+                           lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_cgetri( int matrix_order, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           const lapack_int* ipiv );
+lapack_int LAPACKE_zgetri( int matrix_order, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           const lapack_int* ipiv );
+
+lapack_int LAPACKE_sgetrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dgetrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_cgetrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zgetrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sggbak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const float* lscale,
+                           const float* rscale, lapack_int m, float* v,
+                           lapack_int ldv );
+lapack_int LAPACKE_dggbak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const double* lscale,
+                           const double* rscale, lapack_int m, double* v,
+                           lapack_int ldv );
+lapack_int LAPACKE_cggbak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const float* lscale,
+                           const float* rscale, lapack_int m,
+                           lapack_complex_float* v, lapack_int ldv );
+lapack_int LAPACKE_zggbak( int matrix_order, char job, char side, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, const double* lscale,
+                           const double* rscale, lapack_int m,
+                           lapack_complex_double* v, lapack_int ldv );
+
+lapack_int LAPACKE_sggbal( int matrix_order, char job, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb,
+                           lapack_int* ilo, lapack_int* ihi, float* lscale,
+                           float* rscale );
+lapack_int LAPACKE_dggbal( int matrix_order, char job, lapack_int n, double* a,
+                           lapack_int lda, double* b, lapack_int ldb,
+                           lapack_int* ilo, lapack_int* ihi, double* lscale,
+                           double* rscale );
+lapack_int LAPACKE_cggbal( int matrix_order, char job, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_int* ilo, lapack_int* ihi, float* lscale,
+                           float* rscale );
+lapack_int LAPACKE_zggbal( int matrix_order, char job, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_int* ilo, lapack_int* ihi, double* lscale,
+                           double* rscale );
+
+lapack_int LAPACKE_sgges( int matrix_order, char jobvsl, char jobvsr, char sort,
+                          LAPACK_S_SELECT3 selctg, lapack_int n, float* a,
+                          lapack_int lda, float* b, lapack_int ldb,
+                          lapack_int* sdim, float* alphar, float* alphai,
+                          float* beta, float* vsl, lapack_int ldvsl, float* vsr,
+                          lapack_int ldvsr );
+lapack_int LAPACKE_dgges( int matrix_order, char jobvsl, char jobvsr, char sort,
+                          LAPACK_D_SELECT3 selctg, lapack_int n, double* a,
+                          lapack_int lda, double* b, lapack_int ldb,
+                          lapack_int* sdim, double* alphar, double* alphai,
+                          double* beta, double* vsl, lapack_int ldvsl,
+                          double* vsr, lapack_int ldvsr );
+lapack_int LAPACKE_cgges( int matrix_order, char jobvsl, char jobvsr, char sort,
+                          LAPACK_C_SELECT2 selctg, lapack_int n,
+                          lapack_complex_float* a, lapack_int lda,
+                          lapack_complex_float* b, lapack_int ldb,
+                          lapack_int* sdim, lapack_complex_float* alpha,
+                          lapack_complex_float* beta, lapack_complex_float* vsl,
+                          lapack_int ldvsl, lapack_complex_float* vsr,
+                          lapack_int ldvsr );
+lapack_int LAPACKE_zgges( int matrix_order, char jobvsl, char jobvsr, char sort,
+                          LAPACK_Z_SELECT2 selctg, lapack_int n,
+                          lapack_complex_double* a, lapack_int lda,
+                          lapack_complex_double* b, lapack_int ldb,
+                          lapack_int* sdim, lapack_complex_double* alpha,
+                          lapack_complex_double* beta,
+                          lapack_complex_double* vsl, lapack_int ldvsl,
+                          lapack_complex_double* vsr, lapack_int ldvsr );
+
+lapack_int LAPACKE_sggesx( int matrix_order, char jobvsl, char jobvsr,
+                           char sort, LAPACK_S_SELECT3 selctg, char sense,
+                           lapack_int n, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, lapack_int* sdim, float* alphar,
+                           float* alphai, float* beta, float* vsl,
+                           lapack_int ldvsl, float* vsr, lapack_int ldvsr,
+                           float* rconde, float* rcondv );
+lapack_int LAPACKE_dggesx( int matrix_order, char jobvsl, char jobvsr,
+                           char sort, LAPACK_D_SELECT3 selctg, char sense,
+                           lapack_int n, double* a, lapack_int lda, double* b,
+                           lapack_int ldb, lapack_int* sdim, double* alphar,
+                           double* alphai, double* beta, double* vsl,
+                           lapack_int ldvsl, double* vsr, lapack_int ldvsr,
+                           double* rconde, double* rcondv );
+lapack_int LAPACKE_cggesx( int matrix_order, char jobvsl, char jobvsr,
+                           char sort, LAPACK_C_SELECT2 selctg, char sense,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, lapack_int* sdim,
+                           lapack_complex_float* alpha,
+                           lapack_complex_float* beta,
+                           lapack_complex_float* vsl, lapack_int ldvsl,
+                           lapack_complex_float* vsr, lapack_int ldvsr,
+                           float* rconde, float* rcondv );
+lapack_int LAPACKE_zggesx( int matrix_order, char jobvsl, char jobvsr,
+                           char sort, LAPACK_Z_SELECT2 selctg, char sense,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, lapack_int* sdim,
+                           lapack_complex_double* alpha,
+                           lapack_complex_double* beta,
+                           lapack_complex_double* vsl, lapack_int ldvsl,
+                           lapack_complex_double* vsr, lapack_int ldvsr,
+                           double* rconde, double* rcondv );
+
+lapack_int LAPACKE_sggev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, float* a, lapack_int lda, float* b,
+                          lapack_int ldb, float* alphar, float* alphai,
+                          float* beta, float* vl, lapack_int ldvl, float* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_dggev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, double* a, lapack_int lda, double* b,
+                          lapack_int ldb, double* alphar, double* alphai,
+                          double* beta, double* vl, lapack_int ldvl, double* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_cggev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, lapack_complex_float* a, lapack_int lda,
+                          lapack_complex_float* b, lapack_int ldb,
+                          lapack_complex_float* alpha,
+                          lapack_complex_float* beta, lapack_complex_float* vl,
+                          lapack_int ldvl, lapack_complex_float* vr,
+                          lapack_int ldvr );
+lapack_int LAPACKE_zggev( int matrix_order, char jobvl, char jobvr,
+                          lapack_int n, lapack_complex_double* a,
+                          lapack_int lda, lapack_complex_double* b,
+                          lapack_int ldb, lapack_complex_double* alpha,
+                          lapack_complex_double* beta,
+                          lapack_complex_double* vl, lapack_int ldvl,
+                          lapack_complex_double* vr, lapack_int ldvr );
+
+lapack_int LAPACKE_sggevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb,
+                           float* alphar, float* alphai, float* beta, float* vl,
+                           lapack_int ldvl, float* vr, lapack_int ldvr,
+                           lapack_int* ilo, lapack_int* ihi, float* lscale,
+                           float* rscale, float* abnrm, float* bbnrm,
+                           float* rconde, float* rcondv );
+lapack_int LAPACKE_dggevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n, double* a,
+                           lapack_int lda, double* b, lapack_int ldb,
+                           double* alphar, double* alphai, double* beta,
+                           double* vl, lapack_int ldvl, double* vr,
+                           lapack_int ldvr, lapack_int* ilo, lapack_int* ihi,
+                           double* lscale, double* rscale, double* abnrm,
+                           double* bbnrm, double* rconde, double* rcondv );
+lapack_int LAPACKE_cggevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* alpha,
+                           lapack_complex_float* beta, lapack_complex_float* vl,
+                           lapack_int ldvl, lapack_complex_float* vr,
+                           lapack_int ldvr, lapack_int* ilo, lapack_int* ihi,
+                           float* lscale, float* rscale, float* abnrm,
+                           float* bbnrm, float* rconde, float* rcondv );
+lapack_int LAPACKE_zggevx( int matrix_order, char balanc, char jobvl,
+                           char jobvr, char sense, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* alpha,
+                           lapack_complex_double* beta,
+                           lapack_complex_double* vl, lapack_int ldvl,
+                           lapack_complex_double* vr, lapack_int ldvr,
+                           lapack_int* ilo, lapack_int* ihi, double* lscale,
+                           double* rscale, double* abnrm, double* bbnrm,
+                           double* rconde, double* rcondv );
+
+lapack_int LAPACKE_sggglm( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, float* d, float* x, float* y );
+lapack_int LAPACKE_dggglm( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, double* a, lapack_int lda, double* b,
+                           lapack_int ldb, double* d, double* x, double* y );
+lapack_int LAPACKE_cggglm( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* d,
+                           lapack_complex_float* x, lapack_complex_float* y );
+lapack_int LAPACKE_zggglm( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* d,
+                           lapack_complex_double* x, lapack_complex_double* y );
+
+lapack_int LAPACKE_sgghrd( int matrix_order, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           float* a, lapack_int lda, float* b, lapack_int ldb,
+                           float* q, lapack_int ldq, float* z, lapack_int ldz );
+lapack_int LAPACKE_dgghrd( int matrix_order, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           double* a, lapack_int lda, double* b, lapack_int ldb,
+                           double* q, lapack_int ldq, double* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_cgghrd( int matrix_order, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zgghrd( int matrix_order, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sgglse( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int p, float* a, lapack_int lda, float* b,
+                           lapack_int ldb, float* c, float* d, float* x );
+lapack_int LAPACKE_dgglse( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int p, double* a, lapack_int lda, double* b,
+                           lapack_int ldb, double* c, double* d, double* x );
+lapack_int LAPACKE_cgglse( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int p, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* c,
+                           lapack_complex_float* d, lapack_complex_float* x );
+lapack_int LAPACKE_zgglse( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int p, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* c,
+                           lapack_complex_double* d, lapack_complex_double* x );
+
+lapack_int LAPACKE_sggqrf( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, float* a, lapack_int lda, float* taua,
+                           float* b, lapack_int ldb, float* taub );
+lapack_int LAPACKE_dggqrf( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, double* a, lapack_int lda,
+                           double* taua, double* b, lapack_int ldb,
+                           double* taub );
+lapack_int LAPACKE_cggqrf( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* taua,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* taub );
+lapack_int LAPACKE_zggqrf( int matrix_order, lapack_int n, lapack_int m,
+                           lapack_int p, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* taua,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* taub );
+
+lapack_int LAPACKE_sggrqf( int matrix_order, lapack_int m, lapack_int p,
+                           lapack_int n, float* a, lapack_int lda, float* taua,
+                           float* b, lapack_int ldb, float* taub );
+lapack_int LAPACKE_dggrqf( int matrix_order, lapack_int m, lapack_int p,
+                           lapack_int n, double* a, lapack_int lda,
+                           double* taua, double* b, lapack_int ldb,
+                           double* taub );
+lapack_int LAPACKE_cggrqf( int matrix_order, lapack_int m, lapack_int p,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* taua,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* taub );
+lapack_int LAPACKE_zggrqf( int matrix_order, lapack_int m, lapack_int p,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* taua,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* taub );
+
+lapack_int LAPACKE_sggsvd( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int n, lapack_int p,
+                           lapack_int* k, lapack_int* l, float* a,
+                           lapack_int lda, float* b, lapack_int ldb,
+                           float* alpha, float* beta, float* u, lapack_int ldu,
+                           float* v, lapack_int ldv, float* q, lapack_int ldq,
+                           lapack_int* iwork );
+lapack_int LAPACKE_dggsvd( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int n, lapack_int p,
+                           lapack_int* k, lapack_int* l, double* a,
+                           lapack_int lda, double* b, lapack_int ldb,
+                           double* alpha, double* beta, double* u,
+                           lapack_int ldu, double* v, lapack_int ldv, double* q,
+                           lapack_int ldq, lapack_int* iwork );
+lapack_int LAPACKE_cggsvd( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int n, lapack_int p,
+                           lapack_int* k, lapack_int* l,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           float* alpha, float* beta, lapack_complex_float* u,
+                           lapack_int ldu, lapack_complex_float* v,
+                           lapack_int ldv, lapack_complex_float* q,
+                           lapack_int ldq, lapack_int* iwork );
+lapack_int LAPACKE_zggsvd( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int n, lapack_int p,
+                           lapack_int* k, lapack_int* l,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           double* alpha, double* beta,
+                           lapack_complex_double* u, lapack_int ldu,
+                           lapack_complex_double* v, lapack_int ldv,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_int* iwork );
+
+lapack_int LAPACKE_sggsvp( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb, float tola,
+                           float tolb, lapack_int* k, lapack_int* l, float* u,
+                           lapack_int ldu, float* v, lapack_int ldv, float* q,
+                           lapack_int ldq );
+lapack_int LAPACKE_dggsvp( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n, double* a,
+                           lapack_int lda, double* b, lapack_int ldb,
+                           double tola, double tolb, lapack_int* k,
+                           lapack_int* l, double* u, lapack_int ldu, double* v,
+                           lapack_int ldv, double* q, lapack_int ldq );
+lapack_int LAPACKE_cggsvp( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb, float tola,
+                           float tolb, lapack_int* k, lapack_int* l,
+                           lapack_complex_float* u, lapack_int ldu,
+                           lapack_complex_float* v, lapack_int ldv,
+                           lapack_complex_float* q, lapack_int ldq );
+lapack_int LAPACKE_zggsvp( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           double tola, double tolb, lapack_int* k,
+                           lapack_int* l, lapack_complex_double* u,
+                           lapack_int ldu, lapack_complex_double* v,
+                           lapack_int ldv, lapack_complex_double* q,
+                           lapack_int ldq );
+
+lapack_int LAPACKE_sgtcon( char norm, lapack_int n, const float* dl,
+                           const float* d, const float* du, const float* du2,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_dgtcon( char norm, lapack_int n, const double* dl,
+                           const double* d, const double* du, const double* du2,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+lapack_int LAPACKE_cgtcon( char norm, lapack_int n,
+                           const lapack_complex_float* dl,
+                           const lapack_complex_float* d,
+                           const lapack_complex_float* du,
+                           const lapack_complex_float* du2,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zgtcon( char norm, lapack_int n,
+                           const lapack_complex_double* dl,
+                           const lapack_complex_double* d,
+                           const lapack_complex_double* du,
+                           const lapack_complex_double* du2,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_sgtrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const float* dl, const float* d,
+                           const float* du, const float* dlf, const float* df,
+                           const float* duf, const float* du2,
+                           const lapack_int* ipiv, const float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dgtrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const double* dl, const double* d,
+                           const double* du, const double* dlf,
+                           const double* df, const double* duf,
+                           const double* du2, const lapack_int* ipiv,
+                           const double* b, lapack_int ldb, double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+lapack_int LAPACKE_cgtrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* dl,
+                           const lapack_complex_float* d,
+                           const lapack_complex_float* du,
+                           const lapack_complex_float* dlf,
+                           const lapack_complex_float* df,
+                           const lapack_complex_float* duf,
+                           const lapack_complex_float* du2,
+                           const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zgtrfs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* dl,
+                           const lapack_complex_double* d,
+                           const lapack_complex_double* du,
+                           const lapack_complex_double* dlf,
+                           const lapack_complex_double* df,
+                           const lapack_complex_double* duf,
+                           const lapack_complex_double* du2,
+                           const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sgtsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          float* dl, float* d, float* du, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dgtsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          double* dl, double* d, double* du, double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_cgtsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          lapack_complex_float* dl, lapack_complex_float* d,
+                          lapack_complex_float* du, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zgtsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          lapack_complex_double* dl, lapack_complex_double* d,
+                          lapack_complex_double* du, lapack_complex_double* b,
+                          lapack_int ldb );
+
+lapack_int LAPACKE_sgtsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs, const float* dl,
+                           const float* d, const float* du, float* dlf,
+                           float* df, float* duf, float* du2, lapack_int* ipiv,
+                           const float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dgtsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs, const double* dl,
+                           const double* d, const double* du, double* dlf,
+                           double* df, double* duf, double* du2,
+                           lapack_int* ipiv, const double* b, lapack_int ldb,
+                           double* x, lapack_int ldx, double* rcond,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cgtsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* dl,
+                           const lapack_complex_float* d,
+                           const lapack_complex_float* du,
+                           lapack_complex_float* dlf, lapack_complex_float* df,
+                           lapack_complex_float* duf, lapack_complex_float* du2,
+                           lapack_int* ipiv, const lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zgtsvx( int matrix_order, char fact, char trans,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* dl,
+                           const lapack_complex_double* d,
+                           const lapack_complex_double* du,
+                           lapack_complex_double* dlf,
+                           lapack_complex_double* df,
+                           lapack_complex_double* duf,
+                           lapack_complex_double* du2, lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_sgttrf( lapack_int n, float* dl, float* d, float* du,
+                           float* du2, lapack_int* ipiv );
+lapack_int LAPACKE_dgttrf( lapack_int n, double* dl, double* d, double* du,
+                           double* du2, lapack_int* ipiv );
+lapack_int LAPACKE_cgttrf( lapack_int n, lapack_complex_float* dl,
+                           lapack_complex_float* d, lapack_complex_float* du,
+                           lapack_complex_float* du2, lapack_int* ipiv );
+lapack_int LAPACKE_zgttrf( lapack_int n, lapack_complex_double* dl,
+                           lapack_complex_double* d, lapack_complex_double* du,
+                           lapack_complex_double* du2, lapack_int* ipiv );
+
+lapack_int LAPACKE_sgttrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const float* dl, const float* d,
+                           const float* du, const float* du2,
+                           const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dgttrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const double* dl, const double* d,
+                           const double* du, const double* du2,
+                           const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_cgttrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* dl,
+                           const lapack_complex_float* d,
+                           const lapack_complex_float* du,
+                           const lapack_complex_float* du2,
+                           const lapack_int* ipiv, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zgttrs( int matrix_order, char trans, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* dl,
+                           const lapack_complex_double* d,
+                           const lapack_complex_double* du,
+                           const lapack_complex_double* du2,
+                           const lapack_int* ipiv, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_chbev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int kd, lapack_complex_float* ab,
+                          lapack_int ldab, float* w, lapack_complex_float* z,
+                          lapack_int ldz );
+lapack_int LAPACKE_zhbev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int kd, lapack_complex_double* ab,
+                          lapack_int ldab, double* w, lapack_complex_double* z,
+                          lapack_int ldz );
+
+lapack_int LAPACKE_chbevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_float* ab,
+                           lapack_int ldab, float* w, lapack_complex_float* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_zhbevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_double* ab,
+                           lapack_int ldab, double* w, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_chbevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int kd,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* q, lapack_int ldq, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* ifail );
+lapack_int LAPACKE_zhbevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int kd,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* q, lapack_int ldq, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_chbgst( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_complex_float* bb, lapack_int ldbb,
+                           lapack_complex_float* x, lapack_int ldx );
+lapack_int LAPACKE_zhbgst( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_complex_double* bb, lapack_int ldbb,
+                           lapack_complex_double* x, lapack_int ldx );
+
+lapack_int LAPACKE_chbgv( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int ka, lapack_int kb,
+                          lapack_complex_float* ab, lapack_int ldab,
+                          lapack_complex_float* bb, lapack_int ldbb, float* w,
+                          lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhbgv( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int ka, lapack_int kb,
+                          lapack_complex_double* ab, lapack_int ldab,
+                          lapack_complex_double* bb, lapack_int ldbb, double* w,
+                          lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chbgvd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* bb, lapack_int ldbb, float* w,
+                           lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhbgvd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* bb, lapack_int ldbb,
+                           double* w, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_chbgvx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int ka, lapack_int kb,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* bb, lapack_int ldbb,
+                           lapack_complex_float* q, lapack_int ldq, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* ifail );
+lapack_int LAPACKE_zhbgvx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int ka, lapack_int kb,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* bb, lapack_int ldbb,
+                           lapack_complex_double* q, lapack_int ldq, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_chbtrd( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_float* ab,
+                           lapack_int ldab, float* d, float* e,
+                           lapack_complex_float* q, lapack_int ldq );
+lapack_int LAPACKE_zhbtrd( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_double* ab,
+                           lapack_int ldab, double* d, double* e,
+                           lapack_complex_double* q, lapack_int ldq );
+
+lapack_int LAPACKE_checon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zhecon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_cheequb( int matrix_order, char uplo, lapack_int n,
+                            const lapack_complex_float* a, lapack_int lda,
+                            float* s, float* scond, float* amax );
+lapack_int LAPACKE_zheequb( int matrix_order, char uplo, lapack_int n,
+                            const lapack_complex_double* a, lapack_int lda,
+                            double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_cheev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_complex_float* a, lapack_int lda, float* w );
+lapack_int LAPACKE_zheev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_complex_double* a, lapack_int lda, double* w );
+
+lapack_int LAPACKE_cheevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda, float* w );
+lapack_int LAPACKE_zheevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           double* w );
+
+lapack_int LAPACKE_cheevr( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, float vl, float vu, lapack_int il,
+                           lapack_int iu, float abstol, lapack_int* m, float* w,
+                           lapack_complex_float* z, lapack_int ldz,
+                           lapack_int* isuppz );
+lapack_int LAPACKE_zheevr( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, double vl, double vu, lapack_int il,
+                           lapack_int iu, double abstol, lapack_int* m,
+                           double* w, lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* isuppz );
+
+lapack_int LAPACKE_cheevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, float vl, float vu, lapack_int il,
+                           lapack_int iu, float abstol, lapack_int* m, float* w,
+                           lapack_complex_float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_zheevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, double vl, double vu, lapack_int il,
+                           lapack_int iu, double abstol, lapack_int* m,
+                           double* w, lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_chegst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zhegst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_chegv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, lapack_complex_float* a,
+                          lapack_int lda, lapack_complex_float* b,
+                          lapack_int ldb, float* w );
+lapack_int LAPACKE_zhegv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, lapack_complex_double* a,
+                          lapack_int lda, lapack_complex_double* b,
+                          lapack_int ldb, double* w );
+
+lapack_int LAPACKE_chegvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, float* w );
+lapack_int LAPACKE_zhegvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, double* w );
+
+lapack_int LAPACKE_chegvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* ifail );
+lapack_int LAPACKE_zhegvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_cherfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zherfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_cherfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_float* a, lapack_int lda,
+                            const lapack_complex_float* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const float* s,
+                            const lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_zherfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_double* a, lapack_int lda,
+                            const lapack_complex_double* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const double* s,
+                            const lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+
+lapack_int LAPACKE_chesv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* a,
+                          lapack_int lda, lapack_int* ipiv,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zhesv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* a,
+                          lapack_int lda, lapack_int* ipiv,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_chesvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* af,
+                           lapack_int ldaf, lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zhesvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* af,
+                           lapack_int ldaf, lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_chesvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, float* s,
+                            lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_zhesvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, double* s,
+                            lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* rpvgrw, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+
+lapack_int LAPACKE_chetrd( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda, float* d,
+                           float* e, lapack_complex_float* tau );
+lapack_int LAPACKE_zhetrd( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda, double* d,
+                           double* e, lapack_complex_double* tau );
+
+lapack_int LAPACKE_chetrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_zhetrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ipiv );
+
+lapack_int LAPACKE_chetri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           const lapack_int* ipiv );
+lapack_int LAPACKE_zhetri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           const lapack_int* ipiv );
+
+lapack_int LAPACKE_chetrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zhetrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_chfrk( int matrix_order, char transr, char uplo, char trans,
+                          lapack_int n, lapack_int k, float alpha,
+                          const lapack_complex_float* a, lapack_int lda,
+                          float beta, lapack_complex_float* c );
+lapack_int LAPACKE_zhfrk( int matrix_order, char transr, char uplo, char trans,
+                          lapack_int n, lapack_int k, double alpha,
+                          const lapack_complex_double* a, lapack_int lda,
+                          double beta, lapack_complex_double* c );
+
+lapack_int LAPACKE_shgeqz( int matrix_order, char job, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           float* h, lapack_int ldh, float* t, lapack_int ldt,
+                           float* alphar, float* alphai, float* beta, float* q,
+                           lapack_int ldq, float* z, lapack_int ldz );
+lapack_int LAPACKE_dhgeqz( int matrix_order, char job, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           double* h, lapack_int ldh, double* t, lapack_int ldt,
+                           double* alphar, double* alphai, double* beta,
+                           double* q, lapack_int ldq, double* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_chgeqz( int matrix_order, char job, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           lapack_complex_float* h, lapack_int ldh,
+                           lapack_complex_float* t, lapack_int ldt,
+                           lapack_complex_float* alpha,
+                           lapack_complex_float* beta, lapack_complex_float* q,
+                           lapack_int ldq, lapack_complex_float* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_zhgeqz( int matrix_order, char job, char compq, char compz,
+                           lapack_int n, lapack_int ilo, lapack_int ihi,
+                           lapack_complex_double* h, lapack_int ldh,
+                           lapack_complex_double* t, lapack_int ldt,
+                           lapack_complex_double* alpha,
+                           lapack_complex_double* beta,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chpcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zhpcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_chpev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_complex_float* ap, float* w,
+                          lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhpev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_complex_double* ap, double* w,
+                          lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chpevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_complex_float* ap, float* w,
+                           lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhpevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_complex_double* ap, double* w,
+                           lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chpevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_float* ap, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* ifail );
+lapack_int LAPACKE_zhpevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_complex_double* ap, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_chpgst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, lapack_complex_float* ap,
+                           const lapack_complex_float* bp );
+lapack_int LAPACKE_zhpgst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, lapack_complex_double* ap,
+                           const lapack_complex_double* bp );
+
+lapack_int LAPACKE_chpgv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, lapack_complex_float* ap,
+                          lapack_complex_float* bp, float* w,
+                          lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhpgv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, lapack_complex_double* ap,
+                          lapack_complex_double* bp, double* w,
+                          lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chpgvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, lapack_complex_float* ap,
+                           lapack_complex_float* bp, float* w,
+                           lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zhpgvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, lapack_complex_double* ap,
+                           lapack_complex_double* bp, double* w,
+                           lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_chpgvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n,
+                           lapack_complex_float* ap, lapack_complex_float* bp,
+                           float vl, float vu, lapack_int il, lapack_int iu,
+                           float abstol, lapack_int* m, float* w,
+                           lapack_complex_float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_zhpgvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n,
+                           lapack_complex_double* ap, lapack_complex_double* bp,
+                           double vl, double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_chprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           const lapack_complex_float* afp,
+                           const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zhprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           const lapack_complex_double* afp,
+                           const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_chpsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* ap,
+                          lapack_int* ipiv, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zhpsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* ap,
+                          lapack_int* ipiv, lapack_complex_double* b,
+                          lapack_int ldb );
+
+lapack_int LAPACKE_chpsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           lapack_complex_float* afp, lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zhpsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           lapack_complex_double* afp, lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_chptrd( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap, float* d, float* e,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zhptrd( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap, double* d, double* e,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_chptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap, lapack_int* ipiv );
+lapack_int LAPACKE_zhptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap, lapack_int* ipiv );
+
+lapack_int LAPACKE_chptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap, const lapack_int* ipiv );
+lapack_int LAPACKE_zhptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap, const lapack_int* ipiv );
+
+lapack_int LAPACKE_chptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           const lapack_int* ipiv, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zhptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           const lapack_int* ipiv, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_shsein( int matrix_order, char job, char eigsrc, char initv,
+                           lapack_logical* select, lapack_int n, const float* h,
+                           lapack_int ldh, float* wr, const float* wi,
+                           float* vl, lapack_int ldvl, float* vr,
+                           lapack_int ldvr, lapack_int mm, lapack_int* m,
+                           lapack_int* ifaill, lapack_int* ifailr );
+lapack_int LAPACKE_dhsein( int matrix_order, char job, char eigsrc, char initv,
+                           lapack_logical* select, lapack_int n,
+                           const double* h, lapack_int ldh, double* wr,
+                           const double* wi, double* vl, lapack_int ldvl,
+                           double* vr, lapack_int ldvr, lapack_int mm,
+                           lapack_int* m, lapack_int* ifaill,
+                           lapack_int* ifailr );
+lapack_int LAPACKE_chsein( int matrix_order, char job, char eigsrc, char initv,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_float* h, lapack_int ldh,
+                           lapack_complex_float* w, lapack_complex_float* vl,
+                           lapack_int ldvl, lapack_complex_float* vr,
+                           lapack_int ldvr, lapack_int mm, lapack_int* m,
+                           lapack_int* ifaill, lapack_int* ifailr );
+lapack_int LAPACKE_zhsein( int matrix_order, char job, char eigsrc, char initv,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_double* h, lapack_int ldh,
+                           lapack_complex_double* w, lapack_complex_double* vl,
+                           lapack_int ldvl, lapack_complex_double* vr,
+                           lapack_int ldvr, lapack_int mm, lapack_int* m,
+                           lapack_int* ifaill, lapack_int* ifailr );
+
+lapack_int LAPACKE_shseqr( int matrix_order, char job, char compz, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, float* h,
+                           lapack_int ldh, float* wr, float* wi, float* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_dhseqr( int matrix_order, char job, char compz, lapack_int n,
+                           lapack_int ilo, lapack_int ihi, double* h,
+                           lapack_int ldh, double* wr, double* wi, double* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_chseqr( int matrix_order, char job, char compz, lapack_int n,
+                           lapack_int ilo, lapack_int ihi,
+                           lapack_complex_float* h, lapack_int ldh,
+                           lapack_complex_float* w, lapack_complex_float* z,
+                           lapack_int ldz );
+lapack_int LAPACKE_zhseqr( int matrix_order, char job, char compz, lapack_int n,
+                           lapack_int ilo, lapack_int ihi,
+                           lapack_complex_double* h, lapack_int ldh,
+                           lapack_complex_double* w, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_clacgv( lapack_int n, lapack_complex_float* x,
+                           lapack_int incx );
+lapack_int LAPACKE_zlacgv( lapack_int n, lapack_complex_double* x,
+                           lapack_int incx );
+
+lapack_int LAPACKE_slacpy( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, const float* a, lapack_int lda, float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_dlacpy( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, const double* a, lapack_int lda, double* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_clacpy( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, const lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zlacpy( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, const lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_zlag2c( int matrix_order, lapack_int m, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_float* sa, lapack_int ldsa );
+
+lapack_int LAPACKE_slag2d( int matrix_order, lapack_int m, lapack_int n,
+                           const float* sa, lapack_int ldsa, double* a,
+                           lapack_int lda );
+
+lapack_int LAPACKE_dlag2s( int matrix_order, lapack_int m, lapack_int n,
+                           const double* a, lapack_int lda, float* sa,
+                           lapack_int ldsa );
+
+lapack_int LAPACKE_clag2z( int matrix_order, lapack_int m, lapack_int n,
+                           const lapack_complex_float* sa, lapack_int ldsa,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_slagge( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const float* d,
+                           float* a, lapack_int lda, lapack_int* iseed );
+lapack_int LAPACKE_dlagge( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const double* d,
+                           double* a, lapack_int lda, lapack_int* iseed );
+lapack_int LAPACKE_clagge( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const float* d,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* iseed );
+lapack_int LAPACKE_zlagge( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int kl, lapack_int ku, const double* d,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* iseed );
+
+float LAPACKE_slamch( char cmach );
+double LAPACKE_dlamch( char cmach );
+
+float LAPACKE_slange( int matrix_order, char norm, lapack_int m,
+                           lapack_int n, const float* a, lapack_int lda );
+double LAPACKE_dlange( int matrix_order, char norm, lapack_int m,
+                           lapack_int n, const double* a, lapack_int lda );
+float LAPACKE_clange( int matrix_order, char norm, lapack_int m,
+                           lapack_int n, const lapack_complex_float* a,
+                           lapack_int lda );
+double LAPACKE_zlange( int matrix_order, char norm, lapack_int m,
+                           lapack_int n, const lapack_complex_double* a,
+                           lapack_int lda );
+
+float LAPACKE_clanhe( int matrix_order, char norm, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda );
+double LAPACKE_zlanhe( int matrix_order, char norm, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda );
+
+float LAPACKE_slansy( int matrix_order, char norm, char uplo, lapack_int n,
+                           const float* a, lapack_int lda );
+double LAPACKE_dlansy( int matrix_order, char norm, char uplo, lapack_int n,
+                           const double* a, lapack_int lda );
+float LAPACKE_clansy( int matrix_order, char norm, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda );
+double LAPACKE_zlansy( int matrix_order, char norm, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda );
+
+float LAPACKE_slantr( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int m, lapack_int n, const float* a,
+                           lapack_int lda );
+double LAPACKE_dlantr( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int m, lapack_int n, const double* a,
+                           lapack_int lda );
+float LAPACKE_clantr( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int m, lapack_int n, const lapack_complex_float* a,
+                           lapack_int lda );
+double LAPACKE_zlantr( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int m, lapack_int n, const lapack_complex_double* a,
+                           lapack_int lda );
+
+
+lapack_int LAPACKE_slarfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, const float* v, lapack_int ldv,
+                           const float* t, lapack_int ldt, float* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_dlarfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, const double* v, lapack_int ldv,
+                           const double* t, lapack_int ldt, double* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_clarfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, const lapack_complex_float* v,
+                           lapack_int ldv, const lapack_complex_float* t,
+                           lapack_int ldt, lapack_complex_float* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_zlarfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, const lapack_complex_double* v,
+                           lapack_int ldv, const lapack_complex_double* t,
+                           lapack_int ldt, lapack_complex_double* c,
+                           lapack_int ldc );
+
+lapack_int LAPACKE_slarfg( lapack_int n, float* alpha, float* x,
+                           lapack_int incx, float* tau );
+lapack_int LAPACKE_dlarfg( lapack_int n, double* alpha, double* x,
+                           lapack_int incx, double* tau );
+lapack_int LAPACKE_clarfg( lapack_int n, lapack_complex_float* alpha,
+                           lapack_complex_float* x, lapack_int incx,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_zlarfg( lapack_int n, lapack_complex_double* alpha,
+                           lapack_complex_double* x, lapack_int incx,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_slarft( int matrix_order, char direct, char storev,
+                           lapack_int n, lapack_int k, const float* v,
+                           lapack_int ldv, const float* tau, float* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_dlarft( int matrix_order, char direct, char storev,
+                           lapack_int n, lapack_int k, const double* v,
+                           lapack_int ldv, const double* tau, double* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_clarft( int matrix_order, char direct, char storev,
+                           lapack_int n, lapack_int k,
+                           const lapack_complex_float* v, lapack_int ldv,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zlarft( int matrix_order, char direct, char storev,
+                           lapack_int n, lapack_int k,
+                           const lapack_complex_double* v, lapack_int ldv,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_slarfx( int matrix_order, char side, lapack_int m,
+                           lapack_int n, const float* v, float tau, float* c,
+                           lapack_int ldc, float* work );
+lapack_int LAPACKE_dlarfx( int matrix_order, char side, lapack_int m,
+                           lapack_int n, const double* v, double tau, double* c,
+                           lapack_int ldc, double* work );
+lapack_int LAPACKE_clarfx( int matrix_order, char side, lapack_int m,
+                           lapack_int n, const lapack_complex_float* v,
+                           lapack_complex_float tau, lapack_complex_float* c,
+                           lapack_int ldc, lapack_complex_float* work );
+lapack_int LAPACKE_zlarfx( int matrix_order, char side, lapack_int m,
+                           lapack_int n, const lapack_complex_double* v,
+                           lapack_complex_double tau, lapack_complex_double* c,
+                           lapack_int ldc, lapack_complex_double* work );
+
+lapack_int LAPACKE_slarnv( lapack_int idist, lapack_int* iseed, lapack_int n,
+                           float* x );
+lapack_int LAPACKE_dlarnv( lapack_int idist, lapack_int* iseed, lapack_int n,
+                           double* x );
+lapack_int LAPACKE_clarnv( lapack_int idist, lapack_int* iseed, lapack_int n,
+                           lapack_complex_float* x );
+lapack_int LAPACKE_zlarnv( lapack_int idist, lapack_int* iseed, lapack_int n,
+                           lapack_complex_double* x );
+
+lapack_int LAPACKE_slaset( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, float alpha, float beta, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dlaset( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, double alpha, double beta, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_claset( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, lapack_complex_float alpha,
+                           lapack_complex_float beta, lapack_complex_float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_zlaset( int matrix_order, char uplo, lapack_int m,
+                           lapack_int n, lapack_complex_double alpha,
+                           lapack_complex_double beta, lapack_complex_double* a,
+                           lapack_int lda );
+
+lapack_int LAPACKE_slasrt( char id, lapack_int n, float* d );
+lapack_int LAPACKE_dlasrt( char id, lapack_int n, double* d );
+
+lapack_int LAPACKE_slaswp( int matrix_order, lapack_int n, float* a,
+                           lapack_int lda, lapack_int k1, lapack_int k2,
+                           const lapack_int* ipiv, lapack_int incx );
+lapack_int LAPACKE_dlaswp( int matrix_order, lapack_int n, double* a,
+                           lapack_int lda, lapack_int k1, lapack_int k2,
+                           const lapack_int* ipiv, lapack_int incx );
+lapack_int LAPACKE_claswp( int matrix_order, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int k1, lapack_int k2, const lapack_int* ipiv,
+                           lapack_int incx );
+lapack_int LAPACKE_zlaswp( int matrix_order, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int k1, lapack_int k2, const lapack_int* ipiv,
+                           lapack_int incx );
+
+lapack_int LAPACKE_slatms( int matrix_order, lapack_int m, lapack_int n,
+                           char dist, lapack_int* iseed, char sym, float* d,
+                           lapack_int mode, float cond, float dmax,
+                           lapack_int kl, lapack_int ku, char pack, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dlatms( int matrix_order, lapack_int m, lapack_int n,
+                           char dist, lapack_int* iseed, char sym, double* d,
+                           lapack_int mode, double cond, double dmax,
+                           lapack_int kl, lapack_int ku, char pack, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_clatms( int matrix_order, lapack_int m, lapack_int n,
+                           char dist, lapack_int* iseed, char sym, float* d,
+                           lapack_int mode, float cond, float dmax,
+                           lapack_int kl, lapack_int ku, char pack,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zlatms( int matrix_order, lapack_int m, lapack_int n,
+                           char dist, lapack_int* iseed, char sym, double* d,
+                           lapack_int mode, double cond, double dmax,
+                           lapack_int kl, lapack_int ku, char pack,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_slauum( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dlauum( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_clauum( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zlauum( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_sopgtr( int matrix_order, char uplo, lapack_int n,
+                           const float* ap, const float* tau, float* q,
+                           lapack_int ldq );
+lapack_int LAPACKE_dopgtr( int matrix_order, char uplo, lapack_int n,
+                           const double* ap, const double* tau, double* q,
+                           lapack_int ldq );
+
+lapack_int LAPACKE_sopmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n, const float* ap,
+                           const float* tau, float* c, lapack_int ldc );
+lapack_int LAPACKE_dopmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n, const double* ap,
+                           const double* tau, double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sorgbr( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int k, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorgbr( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int k, double* a,
+                           lapack_int lda, const double* tau );
+
+lapack_int LAPACKE_sorghr( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorghr( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, double* a, lapack_int lda,
+                           const double* tau );
+
+lapack_int LAPACKE_sorglq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorglq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, double* a, lapack_int lda,
+                           const double* tau );
+
+lapack_int LAPACKE_sorgql( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorgql( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, double* a, lapack_int lda,
+                           const double* tau );
+
+lapack_int LAPACKE_sorgqr( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorgqr( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, double* a, lapack_int lda,
+                           const double* tau );
+
+lapack_int LAPACKE_sorgrq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, float* a, lapack_int lda,
+                           const float* tau );
+lapack_int LAPACKE_dorgrq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, double* a, lapack_int lda,
+                           const double* tau );
+
+lapack_int LAPACKE_sorgtr( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda, const float* tau );
+lapack_int LAPACKE_dorgtr( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda, const double* tau );
+
+lapack_int LAPACKE_sormbr( int matrix_order, char vect, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const float* a, lapack_int lda, const float* tau,
+                           float* c, lapack_int ldc );
+lapack_int LAPACKE_dormbr( int matrix_order, char vect, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const double* a, lapack_int lda, const double* tau,
+                           double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormhr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, const float* a, lapack_int lda,
+                           const float* tau, float* c, lapack_int ldc );
+lapack_int LAPACKE_dormhr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, const double* a, lapack_int lda,
+                           const double* tau, double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormlq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const float* a, lapack_int lda, const float* tau,
+                           float* c, lapack_int ldc );
+lapack_int LAPACKE_dormlq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const double* a, lapack_int lda, const double* tau,
+                           double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormql( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const float* a, lapack_int lda, const float* tau,
+                           float* c, lapack_int ldc );
+lapack_int LAPACKE_dormql( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const double* a, lapack_int lda, const double* tau,
+                           double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormqr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const float* a, lapack_int lda, const float* tau,
+                           float* c, lapack_int ldc );
+lapack_int LAPACKE_dormqr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const double* a, lapack_int lda, const double* tau,
+                           double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormrq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const float* a, lapack_int lda, const float* tau,
+                           float* c, lapack_int ldc );
+lapack_int LAPACKE_dormrq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const double* a, lapack_int lda, const double* tau,
+                           double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormrz( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           lapack_int l, const float* a, lapack_int lda,
+                           const float* tau, float* c, lapack_int ldc );
+lapack_int LAPACKE_dormrz( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           lapack_int l, const double* a, lapack_int lda,
+                           const double* tau, double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sormtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n, const float* a,
+                           lapack_int lda, const float* tau, float* c,
+                           lapack_int ldc );
+lapack_int LAPACKE_dormtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n, const double* a,
+                           lapack_int lda, const double* tau, double* c,
+                           lapack_int ldc );
+
+lapack_int LAPACKE_spbcon( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const float* ab, lapack_int ldab,
+                           float anorm, float* rcond );
+lapack_int LAPACKE_dpbcon( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const double* ab, lapack_int ldab,
+                           double anorm, double* rcond );
+lapack_int LAPACKE_cpbcon( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const lapack_complex_float* ab,
+                           lapack_int ldab, float anorm, float* rcond );
+lapack_int LAPACKE_zpbcon( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const lapack_complex_double* ab,
+                           lapack_int ldab, double anorm, double* rcond );
+
+lapack_int LAPACKE_spbequ( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const float* ab, lapack_int ldab,
+                           float* s, float* scond, float* amax );
+lapack_int LAPACKE_dpbequ( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const double* ab, lapack_int ldab,
+                           double* s, double* scond, double* amax );
+lapack_int LAPACKE_cpbequ( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const lapack_complex_float* ab,
+                           lapack_int ldab, float* s, float* scond,
+                           float* amax );
+lapack_int LAPACKE_zpbequ( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, const lapack_complex_double* ab,
+                           lapack_int ldab, double* s, double* scond,
+                           double* amax );
+
+lapack_int LAPACKE_spbrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, const float* ab,
+                           lapack_int ldab, const float* afb, lapack_int ldafb,
+                           const float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_dpbrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, const double* ab,
+                           lapack_int ldab, const double* afb, lapack_int ldafb,
+                           const double* b, lapack_int ldb, double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+lapack_int LAPACKE_cpbrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_complex_float* afb, lapack_int ldafb,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zpbrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_complex_double* afb, lapack_int ldafb,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_spbstf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kb, float* bb, lapack_int ldbb );
+lapack_int LAPACKE_dpbstf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kb, double* bb, lapack_int ldbb );
+lapack_int LAPACKE_cpbstf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kb, lapack_complex_float* bb,
+                           lapack_int ldbb );
+lapack_int LAPACKE_zpbstf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kb, lapack_complex_double* bb,
+                           lapack_int ldbb );
+
+lapack_int LAPACKE_spbsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int kd, lapack_int nrhs, float* ab,
+                          lapack_int ldab, float* b, lapack_int ldb );
+lapack_int LAPACKE_dpbsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int kd, lapack_int nrhs, double* ab,
+                          lapack_int ldab, double* b, lapack_int ldb );
+lapack_int LAPACKE_cpbsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int kd, lapack_int nrhs,
+                          lapack_complex_float* ab, lapack_int ldab,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpbsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int kd, lapack_int nrhs,
+                          lapack_complex_double* ab, lapack_int ldab,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spbsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, float* ab,
+                           lapack_int ldab, float* afb, lapack_int ldafb,
+                           char* equed, float* s, float* b, lapack_int ldb,
+                           float* x, lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dpbsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, double* ab,
+                           lapack_int ldab, double* afb, lapack_int ldafb,
+                           char* equed, double* s, double* b, lapack_int ldb,
+                           double* x, lapack_int ldx, double* rcond,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cpbsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* afb, lapack_int ldafb,
+                           char* equed, float* s, lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zpbsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* afb, lapack_int ldafb,
+                           char* equed, double* s, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* x,
+                           lapack_int ldx, double* rcond, double* ferr,
+                           double* berr );
+
+lapack_int LAPACKE_spbtrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, float* ab, lapack_int ldab );
+lapack_int LAPACKE_dpbtrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, double* ab, lapack_int ldab );
+lapack_int LAPACKE_cpbtrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_float* ab,
+                           lapack_int ldab );
+lapack_int LAPACKE_zpbtrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_complex_double* ab,
+                           lapack_int ldab );
+
+lapack_int LAPACKE_spbtrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, const float* ab,
+                           lapack_int ldab, float* b, lapack_int ldb );
+lapack_int LAPACKE_dpbtrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs, const double* ab,
+                           lapack_int ldab, double* b, lapack_int ldb );
+lapack_int LAPACKE_cpbtrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpbtrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spftrf( int matrix_order, char transr, char uplo,
+                           lapack_int n, float* a );
+lapack_int LAPACKE_dpftrf( int matrix_order, char transr, char uplo,
+                           lapack_int n, double* a );
+lapack_int LAPACKE_cpftrf( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_complex_float* a );
+lapack_int LAPACKE_zpftrf( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_complex_double* a );
+
+lapack_int LAPACKE_spftri( int matrix_order, char transr, char uplo,
+                           lapack_int n, float* a );
+lapack_int LAPACKE_dpftri( int matrix_order, char transr, char uplo,
+                           lapack_int n, double* a );
+lapack_int LAPACKE_cpftri( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_complex_float* a );
+lapack_int LAPACKE_zpftri( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_complex_double* a );
+
+lapack_int LAPACKE_spftrs( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_int nrhs, const float* a,
+                           float* b, lapack_int ldb );
+lapack_int LAPACKE_dpftrs( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_int nrhs, const double* a,
+                           double* b, lapack_int ldb );
+lapack_int LAPACKE_cpftrs( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* a,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpftrs( int matrix_order, char transr, char uplo,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* a,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spocon( int matrix_order, char uplo, lapack_int n,
+                           const float* a, lapack_int lda, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_dpocon( int matrix_order, char uplo, lapack_int n,
+                           const double* a, lapack_int lda, double anorm,
+                           double* rcond );
+lapack_int LAPACKE_cpocon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           float anorm, float* rcond );
+lapack_int LAPACKE_zpocon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           double anorm, double* rcond );
+
+lapack_int LAPACKE_spoequ( int matrix_order, lapack_int n, const float* a,
+                           lapack_int lda, float* s, float* scond,
+                           float* amax );
+lapack_int LAPACKE_dpoequ( int matrix_order, lapack_int n, const double* a,
+                           lapack_int lda, double* s, double* scond,
+                           double* amax );
+lapack_int LAPACKE_cpoequ( int matrix_order, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           float* s, float* scond, float* amax );
+lapack_int LAPACKE_zpoequ( int matrix_order, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_spoequb( int matrix_order, lapack_int n, const float* a,
+                            lapack_int lda, float* s, float* scond,
+                            float* amax );
+lapack_int LAPACKE_dpoequb( int matrix_order, lapack_int n, const double* a,
+                            lapack_int lda, double* s, double* scond,
+                            double* amax );
+lapack_int LAPACKE_cpoequb( int matrix_order, lapack_int n,
+                            const lapack_complex_float* a, lapack_int lda,
+                            float* s, float* scond, float* amax );
+lapack_int LAPACKE_zpoequb( int matrix_order, lapack_int n,
+                            const lapack_complex_double* a, lapack_int lda,
+                            double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_sporfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           const float* af, lapack_int ldaf, const float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dporfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           const double* af, lapack_int ldaf, const double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cporfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* af,
+                           lapack_int ldaf, const lapack_complex_float* b,
+                           lapack_int ldb, lapack_complex_float* x,
+                           lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_zporfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* af,
+                           lapack_int ldaf, const lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+
+lapack_int LAPACKE_sporfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs, const float* a,
+                            lapack_int lda, const float* af, lapack_int ldaf,
+                            const float* s, const float* b, lapack_int ldb,
+                            float* x, lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dporfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs, const double* a,
+                            lapack_int lda, const double* af, lapack_int ldaf,
+                            const double* s, const double* b, lapack_int ldb,
+                            double* x, lapack_int ldx, double* rcond,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+lapack_int LAPACKE_cporfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_float* a, lapack_int lda,
+                            const lapack_complex_float* af, lapack_int ldaf,
+                            const float* s, const lapack_complex_float* b,
+                            lapack_int ldb, lapack_complex_float* x,
+                            lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_zporfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_double* a, lapack_int lda,
+                            const lapack_complex_double* af, lapack_int ldaf,
+                            const double* s, const lapack_complex_double* b,
+                            lapack_int ldb, lapack_complex_double* x,
+                            lapack_int ldx, double* rcond, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+
+lapack_int LAPACKE_sposv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, float* a, lapack_int lda, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dposv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, double* a, lapack_int lda, double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_cposv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* a,
+                          lapack_int lda, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zposv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* a,
+                          lapack_int lda, lapack_complex_double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dsposv( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, double* a, lapack_int lda,
+                           double* b, lapack_int ldb, double* x, lapack_int ldx,
+                           lapack_int* iter );
+lapack_int LAPACKE_zcposv( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, lapack_complex_double* x,
+                           lapack_int ldx, lapack_int* iter );
+
+lapack_int LAPACKE_sposvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, float* a, lapack_int lda, float* af,
+                           lapack_int ldaf, char* equed, float* s, float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_dposvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, double* a, lapack_int lda,
+                           double* af, lapack_int ldaf, char* equed, double* s,
+                           double* b, lapack_int ldb, double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+lapack_int LAPACKE_cposvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* af,
+                           lapack_int ldaf, char* equed, float* s,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zposvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* af,
+                           lapack_int ldaf, char* equed, double* s,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_sposvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs, float* a,
+                            lapack_int lda, float* af, lapack_int ldaf,
+                            char* equed, float* s, float* b, lapack_int ldb,
+                            float* x, lapack_int ldx, float* rcond,
+                            float* rpvgrw, float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_dposvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs, double* a,
+                            lapack_int lda, double* af, lapack_int ldaf,
+                            char* equed, double* s, double* b, lapack_int ldb,
+                            double* x, lapack_int ldx, double* rcond,
+                            double* rpvgrw, double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+lapack_int LAPACKE_cposvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* af, lapack_int ldaf,
+                            char* equed, float* s, lapack_complex_float* b,
+                            lapack_int ldb, lapack_complex_float* x,
+                            lapack_int ldx, float* rcond, float* rpvgrw,
+                            float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_zposvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* af, lapack_int ldaf,
+                            char* equed, double* s, lapack_complex_double* b,
+                            lapack_int ldb, lapack_complex_double* x,
+                            lapack_int ldx, double* rcond, double* rpvgrw,
+                            double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+
+lapack_int LAPACKE_spotrf( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dpotrf( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_cpotrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zpotrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_spotri( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dpotri( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_cpotri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zpotri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_spotrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           float* b, lapack_int ldb );
+lapack_int LAPACKE_dpotrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           double* b, lapack_int ldb );
+lapack_int LAPACKE_cpotrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zpotrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_sppcon( int matrix_order, char uplo, lapack_int n,
+                           const float* ap, float anorm, float* rcond );
+lapack_int LAPACKE_dppcon( int matrix_order, char uplo, lapack_int n,
+                           const double* ap, double anorm, double* rcond );
+lapack_int LAPACKE_cppcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_zppcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_sppequ( int matrix_order, char uplo, lapack_int n,
+                           const float* ap, float* s, float* scond,
+                           float* amax );
+lapack_int LAPACKE_dppequ( int matrix_order, char uplo, lapack_int n,
+                           const double* ap, double* s, double* scond,
+                           double* amax );
+lapack_int LAPACKE_cppequ( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap, float* s,
+                           float* scond, float* amax );
+lapack_int LAPACKE_zppequ( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap, double* s,
+                           double* scond, double* amax );
+
+lapack_int LAPACKE_spprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* ap, const float* afp,
+                           const float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_dpprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* ap, const double* afp,
+                           const double* b, lapack_int ldb, double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+lapack_int LAPACKE_cpprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           const lapack_complex_float* afp,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zpprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           const lapack_complex_double* afp,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sppsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, float* ap, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dppsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, double* ap, double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_cppsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* ap,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zppsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* ap,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sppsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, float* ap, float* afp, char* equed,
+                           float* s, float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dppsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, double* ap, double* afp,
+                           char* equed, double* s, double* b, lapack_int ldb,
+                           double* x, lapack_int ldx, double* rcond,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cppsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, lapack_complex_float* ap,
+                           lapack_complex_float* afp, char* equed, float* s,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zppsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, lapack_complex_double* ap,
+                           lapack_complex_double* afp, char* equed, double* s,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_spptrf( int matrix_order, char uplo, lapack_int n,
+                           float* ap );
+lapack_int LAPACKE_dpptrf( int matrix_order, char uplo, lapack_int n,
+                           double* ap );
+lapack_int LAPACKE_cpptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap );
+lapack_int LAPACKE_zpptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap );
+
+lapack_int LAPACKE_spptri( int matrix_order, char uplo, lapack_int n,
+                           float* ap );
+lapack_int LAPACKE_dpptri( int matrix_order, char uplo, lapack_int n,
+                           double* ap );
+lapack_int LAPACKE_cpptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap );
+lapack_int LAPACKE_zpptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap );
+
+lapack_int LAPACKE_spptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* ap, float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_dpptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* ap, double* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_cpptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spstrf( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda, lapack_int* piv, lapack_int* rank,
+                           float tol );
+lapack_int LAPACKE_dpstrf( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda, lapack_int* piv, lapack_int* rank,
+                           double tol );
+lapack_int LAPACKE_cpstrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* piv, lapack_int* rank, float tol );
+lapack_int LAPACKE_zpstrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* piv, lapack_int* rank, double tol );
+
+lapack_int LAPACKE_sptcon( lapack_int n, const float* d, const float* e,
+                           float anorm, float* rcond );
+lapack_int LAPACKE_dptcon( lapack_int n, const double* d, const double* e,
+                           double anorm, double* rcond );
+lapack_int LAPACKE_cptcon( lapack_int n, const float* d,
+                           const lapack_complex_float* e, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_zptcon( lapack_int n, const double* d,
+                           const lapack_complex_double* e, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_spteqr( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, float* z, lapack_int ldz );
+lapack_int LAPACKE_dpteqr( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, double* z, lapack_int ldz );
+lapack_int LAPACKE_cpteqr( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zpteqr( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_sptrfs( int matrix_order, lapack_int n, lapack_int nrhs,
+                           const float* d, const float* e, const float* df,
+                           const float* ef, const float* b, lapack_int ldb,
+                           float* x, lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_dptrfs( int matrix_order, lapack_int n, lapack_int nrhs,
+                           const double* d, const double* e, const double* df,
+                           const double* ef, const double* b, lapack_int ldb,
+                           double* x, lapack_int ldx, double* ferr,
+                           double* berr );
+lapack_int LAPACKE_cptrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* d,
+                           const lapack_complex_float* e, const float* df,
+                           const lapack_complex_float* ef,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zptrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* d,
+                           const lapack_complex_double* e, const double* df,
+                           const lapack_complex_double* ef,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sptsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          float* d, float* e, float* b, lapack_int ldb );
+lapack_int LAPACKE_dptsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          double* d, double* e, double* b, lapack_int ldb );
+lapack_int LAPACKE_cptsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          float* d, lapack_complex_float* e,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zptsv( int matrix_order, lapack_int n, lapack_int nrhs,
+                          double* d, lapack_complex_double* e,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sptsvx( int matrix_order, char fact, lapack_int n,
+                           lapack_int nrhs, const float* d, const float* e,
+                           float* df, float* ef, const float* b, lapack_int ldb,
+                           float* x, lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dptsvx( int matrix_order, char fact, lapack_int n,
+                           lapack_int nrhs, const double* d, const double* e,
+                           double* df, double* ef, const double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+lapack_int LAPACKE_cptsvx( int matrix_order, char fact, lapack_int n,
+                           lapack_int nrhs, const float* d,
+                           const lapack_complex_float* e, float* df,
+                           lapack_complex_float* ef,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zptsvx( int matrix_order, char fact, lapack_int n,
+                           lapack_int nrhs, const double* d,
+                           const lapack_complex_double* e, double* df,
+                           lapack_complex_double* ef,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_spttrf( lapack_int n, float* d, float* e );
+lapack_int LAPACKE_dpttrf( lapack_int n, double* d, double* e );
+lapack_int LAPACKE_cpttrf( lapack_int n, float* d, lapack_complex_float* e );
+lapack_int LAPACKE_zpttrf( lapack_int n, double* d, lapack_complex_double* e );
+
+lapack_int LAPACKE_spttrs( int matrix_order, lapack_int n, lapack_int nrhs,
+                           const float* d, const float* e, float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_dpttrs( int matrix_order, lapack_int n, lapack_int nrhs,
+                           const double* d, const double* e, double* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_cpttrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* d,
+                           const lapack_complex_float* e,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpttrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* d,
+                           const lapack_complex_double* e,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_ssbev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int kd, float* ab, lapack_int ldab, float* w,
+                          float* z, lapack_int ldz );
+lapack_int LAPACKE_dsbev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int kd, double* ab, lapack_int ldab, double* w,
+                          double* z, lapack_int ldz );
+
+lapack_int LAPACKE_ssbevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int kd, float* ab, lapack_int ldab, float* w,
+                           float* z, lapack_int ldz );
+lapack_int LAPACKE_dsbevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int kd, double* ab, lapack_int ldab,
+                           double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_ssbevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int kd, float* ab,
+                           lapack_int ldab, float* q, lapack_int ldq, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dsbevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int kd, double* ab,
+                           lapack_int ldab, double* q, lapack_int ldq,
+                           double vl, double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* ifail );
+
+lapack_int LAPACKE_ssbgst( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb, float* ab,
+                           lapack_int ldab, const float* bb, lapack_int ldbb,
+                           float* x, lapack_int ldx );
+lapack_int LAPACKE_dsbgst( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb, double* ab,
+                           lapack_int ldab, const double* bb, lapack_int ldbb,
+                           double* x, lapack_int ldx );
+
+lapack_int LAPACKE_ssbgv( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int ka, lapack_int kb, float* ab,
+                          lapack_int ldab, float* bb, lapack_int ldbb, float* w,
+                          float* z, lapack_int ldz );
+lapack_int LAPACKE_dsbgv( int matrix_order, char jobz, char uplo, lapack_int n,
+                          lapack_int ka, lapack_int kb, double* ab,
+                          lapack_int ldab, double* bb, lapack_int ldbb,
+                          double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_ssbgvd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb, float* ab,
+                           lapack_int ldab, float* bb, lapack_int ldbb,
+                           float* w, float* z, lapack_int ldz );
+lapack_int LAPACKE_dsbgvd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           lapack_int ka, lapack_int kb, double* ab,
+                           lapack_int ldab, double* bb, lapack_int ldbb,
+                           double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_ssbgvx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int ka, lapack_int kb,
+                           float* ab, lapack_int ldab, float* bb,
+                           lapack_int ldbb, float* q, lapack_int ldq, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dsbgvx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, lapack_int ka, lapack_int kb,
+                           double* ab, lapack_int ldab, double* bb,
+                           lapack_int ldbb, double* q, lapack_int ldq,
+                           double vl, double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* ifail );
+
+lapack_int LAPACKE_ssbtrd( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int kd, float* ab, lapack_int ldab, float* d,
+                           float* e, float* q, lapack_int ldq );
+lapack_int LAPACKE_dsbtrd( int matrix_order, char vect, char uplo, lapack_int n,
+                           lapack_int kd, double* ab, lapack_int ldab,
+                           double* d, double* e, double* q, lapack_int ldq );
+
+lapack_int LAPACKE_ssfrk( int matrix_order, char transr, char uplo, char trans,
+                          lapack_int n, lapack_int k, float alpha,
+                          const float* a, lapack_int lda, float beta,
+                          float* c );
+lapack_int LAPACKE_dsfrk( int matrix_order, char transr, char uplo, char trans,
+                          lapack_int n, lapack_int k, double alpha,
+                          const double* a, lapack_int lda, double beta,
+                          double* c );
+
+lapack_int LAPACKE_sspcon( int matrix_order, char uplo, lapack_int n,
+                           const float* ap, const lapack_int* ipiv, float anorm,
+                           float* rcond );
+lapack_int LAPACKE_dspcon( int matrix_order, char uplo, lapack_int n,
+                           const double* ap, const lapack_int* ipiv,
+                           double anorm, double* rcond );
+lapack_int LAPACKE_cspcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zspcon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_sspev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          float* ap, float* w, float* z, lapack_int ldz );
+lapack_int LAPACKE_dspev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          double* ap, double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sspevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           float* ap, float* w, float* z, lapack_int ldz );
+lapack_int LAPACKE_dspevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           double* ap, double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sspevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, float* ap, float vl, float vu,
+                           lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dspevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, double* ap, double vl, double vu,
+                           lapack_int il, lapack_int iu, double abstol,
+                           lapack_int* m, double* w, double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_sspgst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, float* ap, const float* bp );
+lapack_int LAPACKE_dspgst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, double* ap, const double* bp );
+
+lapack_int LAPACKE_sspgv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, float* ap, float* bp,
+                          float* w, float* z, lapack_int ldz );
+lapack_int LAPACKE_dspgv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, double* ap, double* bp,
+                          double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sspgvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, float* ap, float* bp,
+                           float* w, float* z, lapack_int ldz );
+lapack_int LAPACKE_dspgvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, double* ap, double* bp,
+                           double* w, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sspgvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n, float* ap,
+                           float* bp, float vl, float vu, lapack_int il,
+                           lapack_int iu, float abstol, lapack_int* m, float* w,
+                           float* z, lapack_int ldz, lapack_int* ifail );
+lapack_int LAPACKE_dspgvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n, double* ap,
+                           double* bp, double vl, double vu, lapack_int il,
+                           lapack_int iu, double abstol, lapack_int* m,
+                           double* w, double* z, lapack_int ldz,
+                           lapack_int* ifail );
+
+lapack_int LAPACKE_ssprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* ap, const float* afp,
+                           const lapack_int* ipiv, const float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dsprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* ap, const double* afp,
+                           const lapack_int* ipiv, const double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_csprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           const lapack_complex_float* afp,
+                           const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zsprfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           const lapack_complex_double* afp,
+                           const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_sspsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, float* ap, lapack_int* ipiv,
+                          float* b, lapack_int ldb );
+lapack_int LAPACKE_dspsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, double* ap, lapack_int* ipiv,
+                          double* b, lapack_int ldb );
+lapack_int LAPACKE_cspsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* ap,
+                          lapack_int* ipiv, lapack_complex_float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_zspsv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* ap,
+                          lapack_int* ipiv, lapack_complex_double* b,
+                          lapack_int ldb );
+
+lapack_int LAPACKE_sspsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* ap, float* afp,
+                           lapack_int* ipiv, const float* b, lapack_int ldb,
+                           float* x, lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dspsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* ap, double* afp,
+                           lapack_int* ipiv, const double* b, lapack_int ldb,
+                           double* x, lapack_int ldx, double* rcond,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_cspsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           lapack_complex_float* afp, lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zspsvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           lapack_complex_double* afp, lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_ssptrd( int matrix_order, char uplo, lapack_int n, float* ap,
+                           float* d, float* e, float* tau );
+lapack_int LAPACKE_dsptrd( int matrix_order, char uplo, lapack_int n,
+                           double* ap, double* d, double* e, double* tau );
+
+lapack_int LAPACKE_ssptrf( int matrix_order, char uplo, lapack_int n, float* ap,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_dsptrf( int matrix_order, char uplo, lapack_int n,
+                           double* ap, lapack_int* ipiv );
+lapack_int LAPACKE_csptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap, lapack_int* ipiv );
+lapack_int LAPACKE_zsptrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap, lapack_int* ipiv );
+
+lapack_int LAPACKE_ssptri( int matrix_order, char uplo, lapack_int n, float* ap,
+                           const lapack_int* ipiv );
+lapack_int LAPACKE_dsptri( int matrix_order, char uplo, lapack_int n,
+                           double* ap, const lapack_int* ipiv );
+lapack_int LAPACKE_csptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* ap, const lapack_int* ipiv );
+lapack_int LAPACKE_zsptri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* ap, const lapack_int* ipiv );
+
+lapack_int LAPACKE_ssptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* ap,
+                           const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dsptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* ap,
+                           const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_csptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* ap,
+                           const lapack_int* ipiv, lapack_complex_float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_zsptrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* ap,
+                           const lapack_int* ipiv, lapack_complex_double* b,
+                           lapack_int ldb );
+
+lapack_int LAPACKE_sstebz( char range, char order, lapack_int n, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           const float* d, const float* e, lapack_int* m,
+                           lapack_int* nsplit, float* w, lapack_int* iblock,
+                           lapack_int* isplit );
+lapack_int LAPACKE_dstebz( char range, char order, lapack_int n, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, const double* d, const double* e,
+                           lapack_int* m, lapack_int* nsplit, double* w,
+                           lapack_int* iblock, lapack_int* isplit );
+
+lapack_int LAPACKE_sstedc( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, float* z, lapack_int ldz );
+lapack_int LAPACKE_dstedc( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, double* z, lapack_int ldz );
+lapack_int LAPACKE_cstedc( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zstedc( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_sstegr( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* isuppz );
+lapack_int LAPACKE_dstegr( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* isuppz );
+lapack_int LAPACKE_cstegr( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* isuppz );
+lapack_int LAPACKE_zstegr( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* isuppz );
+
+lapack_int LAPACKE_sstein( int matrix_order, lapack_int n, const float* d,
+                           const float* e, lapack_int m, const float* w,
+                           const lapack_int* iblock, const lapack_int* isplit,
+                           float* z, lapack_int ldz, lapack_int* ifailv );
+lapack_int LAPACKE_dstein( int matrix_order, lapack_int n, const double* d,
+                           const double* e, lapack_int m, const double* w,
+                           const lapack_int* iblock, const lapack_int* isplit,
+                           double* z, lapack_int ldz, lapack_int* ifailv );
+lapack_int LAPACKE_cstein( int matrix_order, lapack_int n, const float* d,
+                           const float* e, lapack_int m, const float* w,
+                           const lapack_int* iblock, const lapack_int* isplit,
+                           lapack_complex_float* z, lapack_int ldz,
+                           lapack_int* ifailv );
+lapack_int LAPACKE_zstein( int matrix_order, lapack_int n, const double* d,
+                           const double* e, lapack_int m, const double* w,
+                           const lapack_int* iblock, const lapack_int* isplit,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* ifailv );
+
+lapack_int LAPACKE_sstemr( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, lapack_int* m,
+                           float* w, float* z, lapack_int ldz, lapack_int nzc,
+                           lapack_int* isuppz, lapack_logical* tryrac );
+lapack_int LAPACKE_dstemr( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           lapack_int* m, double* w, double* z, lapack_int ldz,
+                           lapack_int nzc, lapack_int* isuppz,
+                           lapack_logical* tryrac );
+lapack_int LAPACKE_cstemr( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, lapack_int* m,
+                           float* w, lapack_complex_float* z, lapack_int ldz,
+                           lapack_int nzc, lapack_int* isuppz,
+                           lapack_logical* tryrac );
+lapack_int LAPACKE_zstemr( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           lapack_int* m, double* w, lapack_complex_double* z,
+                           lapack_int ldz, lapack_int nzc, lapack_int* isuppz,
+                           lapack_logical* tryrac );
+
+lapack_int LAPACKE_ssteqr( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, float* z, lapack_int ldz );
+lapack_int LAPACKE_dsteqr( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, double* z, lapack_int ldz );
+lapack_int LAPACKE_csteqr( int matrix_order, char compz, lapack_int n, float* d,
+                           float* e, lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zsteqr( int matrix_order, char compz, lapack_int n,
+                           double* d, double* e, lapack_complex_double* z,
+                           lapack_int ldz );
+
+lapack_int LAPACKE_ssterf( lapack_int n, float* d, float* e );
+lapack_int LAPACKE_dsterf( lapack_int n, double* d, double* e );
+
+lapack_int LAPACKE_sstev( int matrix_order, char jobz, lapack_int n, float* d,
+                          float* e, float* z, lapack_int ldz );
+lapack_int LAPACKE_dstev( int matrix_order, char jobz, lapack_int n, double* d,
+                          double* e, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sstevd( int matrix_order, char jobz, lapack_int n, float* d,
+                           float* e, float* z, lapack_int ldz );
+lapack_int LAPACKE_dstevd( int matrix_order, char jobz, lapack_int n, double* d,
+                           double* e, double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sstevr( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* isuppz );
+lapack_int LAPACKE_dstevr( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* isuppz );
+
+lapack_int LAPACKE_sstevx( int matrix_order, char jobz, char range,
+                           lapack_int n, float* d, float* e, float vl, float vu,
+                           lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dstevx( int matrix_order, char jobz, char range,
+                           lapack_int n, double* d, double* e, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* ifail );
+
+lapack_int LAPACKE_ssycon( int matrix_order, char uplo, lapack_int n,
+                           const float* a, lapack_int lda,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_dsycon( int matrix_order, char uplo, lapack_int n,
+                           const double* a, lapack_int lda,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+lapack_int LAPACKE_csycon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_int* ipiv, float anorm, float* rcond );
+lapack_int LAPACKE_zsycon( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_int* ipiv, double anorm,
+                           double* rcond );
+
+lapack_int LAPACKE_ssyequb( int matrix_order, char uplo, lapack_int n,
+                            const float* a, lapack_int lda, float* s,
+                            float* scond, float* amax );
+lapack_int LAPACKE_dsyequb( int matrix_order, char uplo, lapack_int n,
+                            const double* a, lapack_int lda, double* s,
+                            double* scond, double* amax );
+lapack_int LAPACKE_csyequb( int matrix_order, char uplo, lapack_int n,
+                            const lapack_complex_float* a, lapack_int lda,
+                            float* s, float* scond, float* amax );
+lapack_int LAPACKE_zsyequb( int matrix_order, char uplo, lapack_int n,
+                            const lapack_complex_double* a, lapack_int lda,
+                            double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_ssyev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          float* a, lapack_int lda, float* w );
+lapack_int LAPACKE_dsyev( int matrix_order, char jobz, char uplo, lapack_int n,
+                          double* a, lapack_int lda, double* w );
+
+lapack_int LAPACKE_ssyevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           float* a, lapack_int lda, float* w );
+lapack_int LAPACKE_dsyevd( int matrix_order, char jobz, char uplo, lapack_int n,
+                           double* a, lapack_int lda, double* w );
+
+lapack_int LAPACKE_ssyevr( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, float* a, lapack_int lda, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* isuppz );
+lapack_int LAPACKE_dsyevr( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, double* a, lapack_int lda, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* isuppz );
+
+lapack_int LAPACKE_ssyevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, float* a, lapack_int lda, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dsyevx( int matrix_order, char jobz, char range, char uplo,
+                           lapack_int n, double* a, lapack_int lda, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* ifail );
+
+lapack_int LAPACKE_ssygst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, float* a, lapack_int lda,
+                           const float* b, lapack_int ldb );
+lapack_int LAPACKE_dsygst( int matrix_order, lapack_int itype, char uplo,
+                           lapack_int n, double* a, lapack_int lda,
+                           const double* b, lapack_int ldb );
+
+lapack_int LAPACKE_ssygv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, float* a, lapack_int lda,
+                          float* b, lapack_int ldb, float* w );
+lapack_int LAPACKE_dsygv( int matrix_order, lapack_int itype, char jobz,
+                          char uplo, lapack_int n, double* a, lapack_int lda,
+                          double* b, lapack_int ldb, double* w );
+
+lapack_int LAPACKE_ssygvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, float* a, lapack_int lda,
+                           float* b, lapack_int ldb, float* w );
+lapack_int LAPACKE_dsygvd( int matrix_order, lapack_int itype, char jobz,
+                           char uplo, lapack_int n, double* a, lapack_int lda,
+                           double* b, lapack_int ldb, double* w );
+
+lapack_int LAPACKE_ssygvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb, float vl,
+                           float vu, lapack_int il, lapack_int iu, float abstol,
+                           lapack_int* m, float* w, float* z, lapack_int ldz,
+                           lapack_int* ifail );
+lapack_int LAPACKE_dsygvx( int matrix_order, lapack_int itype, char jobz,
+                           char range, char uplo, lapack_int n, double* a,
+                           lapack_int lda, double* b, lapack_int ldb, double vl,
+                           double vu, lapack_int il, lapack_int iu,
+                           double abstol, lapack_int* m, double* w, double* z,
+                           lapack_int ldz, lapack_int* ifail );
+
+lapack_int LAPACKE_ssyrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           const float* af, lapack_int ldaf,
+                           const lapack_int* ipiv, const float* b,
+                           lapack_int ldb, float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dsyrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           const double* af, lapack_int ldaf,
+                           const lapack_int* ipiv, const double* b,
+                           lapack_int ldb, double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_csyrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_zsyrfs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* af,
+                           lapack_int ldaf, const lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_ssyrfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs, const float* a,
+                            lapack_int lda, const float* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const float* s,
+                            const float* b, lapack_int ldb, float* x,
+                            lapack_int ldx, float* rcond, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dsyrfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs, const double* a,
+                            lapack_int lda, const double* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const double* s,
+                            const double* b, lapack_int ldb, double* x,
+                            lapack_int ldx, double* rcond, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+lapack_int LAPACKE_csyrfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_float* a, lapack_int lda,
+                            const lapack_complex_float* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const float* s,
+                            const lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* berr, lapack_int n_err_bnds,
+                            float* err_bnds_norm, float* err_bnds_comp,
+                            lapack_int nparams, float* params );
+lapack_int LAPACKE_zsyrfsx( int matrix_order, char uplo, char equed,
+                            lapack_int n, lapack_int nrhs,
+                            const lapack_complex_double* a, lapack_int lda,
+                            const lapack_complex_double* af, lapack_int ldaf,
+                            const lapack_int* ipiv, const double* s,
+                            const lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* berr, lapack_int n_err_bnds,
+                            double* err_bnds_norm, double* err_bnds_comp,
+                            lapack_int nparams, double* params );
+
+lapack_int LAPACKE_ssysv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, float* a, lapack_int lda,
+                          lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dsysv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, double* a, lapack_int lda,
+                          lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_csysv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_float* a,
+                          lapack_int lda, lapack_int* ipiv,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zsysv( int matrix_order, char uplo, lapack_int n,
+                          lapack_int nrhs, lapack_complex_double* a,
+                          lapack_int lda, lapack_int* ipiv,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_ssysvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           float* af, lapack_int ldaf, lapack_int* ipiv,
+                           const float* b, lapack_int ldb, float* x,
+                           lapack_int ldx, float* rcond, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dsysvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           double* af, lapack_int ldaf, lapack_int* ipiv,
+                           const double* b, lapack_int ldb, double* x,
+                           lapack_int ldx, double* rcond, double* ferr,
+                           double* berr );
+lapack_int LAPACKE_csysvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* af,
+                           lapack_int ldaf, lapack_int* ipiv,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* x, lapack_int ldx,
+                           float* rcond, float* ferr, float* berr );
+lapack_int LAPACKE_zsysvx( int matrix_order, char fact, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* af,
+                           lapack_int ldaf, lapack_int* ipiv,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* x, lapack_int ldx,
+                           double* rcond, double* ferr, double* berr );
+
+lapack_int LAPACKE_ssysvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs, float* a,
+                            lapack_int lda, float* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, float* s, float* b,
+                            lapack_int ldb, float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_dsysvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs, double* a,
+                            lapack_int lda, double* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, double* s, double* b,
+                            lapack_int ldb, double* x, lapack_int ldx,
+                            double* rcond, double* rpvgrw, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+lapack_int LAPACKE_csysvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, float* s,
+                            lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* x, lapack_int ldx,
+                            float* rcond, float* rpvgrw, float* berr,
+                            lapack_int n_err_bnds, float* err_bnds_norm,
+                            float* err_bnds_comp, lapack_int nparams,
+                            float* params );
+lapack_int LAPACKE_zsysvxx( int matrix_order, char fact, char uplo,
+                            lapack_int n, lapack_int nrhs,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* af, lapack_int ldaf,
+                            lapack_int* ipiv, char* equed, double* s,
+                            lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* x, lapack_int ldx,
+                            double* rcond, double* rpvgrw, double* berr,
+                            lapack_int n_err_bnds, double* err_bnds_norm,
+                            double* err_bnds_comp, lapack_int nparams,
+                            double* params );
+
+lapack_int LAPACKE_ssytrd( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda, float* d, float* e, float* tau );
+lapack_int LAPACKE_dsytrd( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda, double* d, double* e, double* tau );
+
+lapack_int LAPACKE_ssytrf( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_dsytrf( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_csytrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_int* ipiv );
+lapack_int LAPACKE_zsytrf( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_int* ipiv );
+
+lapack_int LAPACKE_ssytri( int matrix_order, char uplo, lapack_int n, float* a,
+                           lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_dsytri( int matrix_order, char uplo, lapack_int n, double* a,
+                           lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_csytri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           const lapack_int* ipiv );
+lapack_int LAPACKE_zsytri( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           const lapack_int* ipiv );
+
+lapack_int LAPACKE_ssytrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const float* a, lapack_int lda,
+                           const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_dsytrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const double* a, lapack_int lda,
+                           const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_csytrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zsytrs( int matrix_order, char uplo, lapack_int n,
+                           lapack_int nrhs, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_int* ipiv,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stbcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, lapack_int kd, const float* ab,
+                           lapack_int ldab, float* rcond );
+lapack_int LAPACKE_dtbcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, lapack_int kd, const double* ab,
+                           lapack_int ldab, double* rcond );
+lapack_int LAPACKE_ctbcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, lapack_int kd,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           float* rcond );
+lapack_int LAPACKE_ztbcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, lapack_int kd,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           double* rcond );
+
+lapack_int LAPACKE_stbrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const float* ab, lapack_int ldab, const float* b,
+                           lapack_int ldb, const float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_dtbrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const double* ab, lapack_int ldab, const double* b,
+                           lapack_int ldb, const double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+lapack_int LAPACKE_ctbrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           const lapack_complex_float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_ztbrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           const lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_stbtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const float* ab, lapack_int ldab, float* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_dtbtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const double* ab, lapack_int ldab, double* b,
+                           lapack_int ldb );
+lapack_int LAPACKE_ctbtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_float* ab, lapack_int ldab,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztbtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int kd, lapack_int nrhs,
+                           const lapack_complex_double* ab, lapack_int ldab,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stfsm( int matrix_order, char transr, char side, char uplo,
+                          char trans, char diag, lapack_int m, lapack_int n,
+                          float alpha, const float* a, float* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_dtfsm( int matrix_order, char transr, char side, char uplo,
+                          char trans, char diag, lapack_int m, lapack_int n,
+                          double alpha, const double* a, double* b,
+                          lapack_int ldb );
+lapack_int LAPACKE_ctfsm( int matrix_order, char transr, char side, char uplo,
+                          char trans, char diag, lapack_int m, lapack_int n,
+                          lapack_complex_float alpha,
+                          const lapack_complex_float* a,
+                          lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztfsm( int matrix_order, char transr, char side, char uplo,
+                          char trans, char diag, lapack_int m, lapack_int n,
+                          lapack_complex_double alpha,
+                          const lapack_complex_double* a,
+                          lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stftri( int matrix_order, char transr, char uplo, char diag,
+                           lapack_int n, float* a );
+lapack_int LAPACKE_dtftri( int matrix_order, char transr, char uplo, char diag,
+                           lapack_int n, double* a );
+lapack_int LAPACKE_ctftri( int matrix_order, char transr, char uplo, char diag,
+                           lapack_int n, lapack_complex_float* a );
+lapack_int LAPACKE_ztftri( int matrix_order, char transr, char uplo, char diag,
+                           lapack_int n, lapack_complex_double* a );
+
+lapack_int LAPACKE_stfttp( int matrix_order, char transr, char uplo,
+                           lapack_int n, const float* arf, float* ap );
+lapack_int LAPACKE_dtfttp( int matrix_order, char transr, char uplo,
+                           lapack_int n, const double* arf, double* ap );
+lapack_int LAPACKE_ctfttp( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_float* arf,
+                           lapack_complex_float* ap );
+lapack_int LAPACKE_ztfttp( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_double* arf,
+                           lapack_complex_double* ap );
+
+lapack_int LAPACKE_stfttr( int matrix_order, char transr, char uplo,
+                           lapack_int n, const float* arf, float* a,
+                           lapack_int lda );
+lapack_int LAPACKE_dtfttr( int matrix_order, char transr, char uplo,
+                           lapack_int n, const double* arf, double* a,
+                           lapack_int lda );
+lapack_int LAPACKE_ctfttr( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_float* arf,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_ztfttr( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_double* arf,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_stgevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const float* s, lapack_int lds, const float* p,
+                           lapack_int ldp, float* vl, lapack_int ldvl,
+                           float* vr, lapack_int ldvr, lapack_int mm,
+                           lapack_int* m );
+lapack_int LAPACKE_dtgevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const double* s, lapack_int lds, const double* p,
+                           lapack_int ldp, double* vl, lapack_int ldvl,
+                           double* vr, lapack_int ldvr, lapack_int mm,
+                           lapack_int* m );
+lapack_int LAPACKE_ctgevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_float* s, lapack_int lds,
+                           const lapack_complex_float* p, lapack_int ldp,
+                           lapack_complex_float* vl, lapack_int ldvl,
+                           lapack_complex_float* vr, lapack_int ldvr,
+                           lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ztgevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_double* s, lapack_int lds,
+                           const lapack_complex_double* p, lapack_int ldp,
+                           lapack_complex_double* vl, lapack_int ldvl,
+                           lapack_complex_double* vr, lapack_int ldvr,
+                           lapack_int mm, lapack_int* m );
+
+lapack_int LAPACKE_stgexc( int matrix_order, lapack_logical wantq,
+                           lapack_logical wantz, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb, float* q,
+                           lapack_int ldq, float* z, lapack_int ldz,
+                           lapack_int* ifst, lapack_int* ilst );
+lapack_int LAPACKE_dtgexc( int matrix_order, lapack_logical wantq,
+                           lapack_logical wantz, lapack_int n, double* a,
+                           lapack_int lda, double* b, lapack_int ldb, double* q,
+                           lapack_int ldq, double* z, lapack_int ldz,
+                           lapack_int* ifst, lapack_int* ilst );
+lapack_int LAPACKE_ctgexc( int matrix_order, lapack_logical wantq,
+                           lapack_logical wantz, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_complex_float* z, lapack_int ldz,
+                           lapack_int ifst, lapack_int ilst );
+lapack_int LAPACKE_ztgexc( int matrix_order, lapack_logical wantq,
+                           lapack_logical wantz, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int ifst, lapack_int ilst );
+
+lapack_int LAPACKE_stgsen( int matrix_order, lapack_int ijob,
+                           lapack_logical wantq, lapack_logical wantz,
+                           const lapack_logical* select, lapack_int n, float* a,
+                           lapack_int lda, float* b, lapack_int ldb,
+                           float* alphar, float* alphai, float* beta, float* q,
+                           lapack_int ldq, float* z, lapack_int ldz,
+                           lapack_int* m, float* pl, float* pr, float* dif );
+lapack_int LAPACKE_dtgsen( int matrix_order, lapack_int ijob,
+                           lapack_logical wantq, lapack_logical wantz,
+                           const lapack_logical* select, lapack_int n,
+                           double* a, lapack_int lda, double* b, lapack_int ldb,
+                           double* alphar, double* alphai, double* beta,
+                           double* q, lapack_int ldq, double* z, lapack_int ldz,
+                           lapack_int* m, double* pl, double* pr, double* dif );
+lapack_int LAPACKE_ctgsen( int matrix_order, lapack_int ijob,
+                           lapack_logical wantq, lapack_logical wantz,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* alpha,
+                           lapack_complex_float* beta, lapack_complex_float* q,
+                           lapack_int ldq, lapack_complex_float* z,
+                           lapack_int ldz, lapack_int* m, float* pl, float* pr,
+                           float* dif );
+lapack_int LAPACKE_ztgsen( int matrix_order, lapack_int ijob,
+                           lapack_logical wantq, lapack_logical wantz,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* alpha,
+                           lapack_complex_double* beta,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* z, lapack_int ldz,
+                           lapack_int* m, double* pl, double* pr, double* dif );
+
+lapack_int LAPACKE_stgsja( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_int k, lapack_int l, float* a, lapack_int lda,
+                           float* b, lapack_int ldb, float tola, float tolb,
+                           float* alpha, float* beta, float* u, lapack_int ldu,
+                           float* v, lapack_int ldv, float* q, lapack_int ldq,
+                           lapack_int* ncycle );
+lapack_int LAPACKE_dtgsja( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_int k, lapack_int l, double* a,
+                           lapack_int lda, double* b, lapack_int ldb,
+                           double tola, double tolb, double* alpha,
+                           double* beta, double* u, lapack_int ldu, double* v,
+                           lapack_int ldv, double* q, lapack_int ldq,
+                           lapack_int* ncycle );
+lapack_int LAPACKE_ctgsja( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_int k, lapack_int l, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* b,
+                           lapack_int ldb, float tola, float tolb, float* alpha,
+                           float* beta, lapack_complex_float* u, lapack_int ldu,
+                           lapack_complex_float* v, lapack_int ldv,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_int* ncycle );
+lapack_int LAPACKE_ztgsja( int matrix_order, char jobu, char jobv, char jobq,
+                           lapack_int m, lapack_int p, lapack_int n,
+                           lapack_int k, lapack_int l, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* b,
+                           lapack_int ldb, double tola, double tolb,
+                           double* alpha, double* beta,
+                           lapack_complex_double* u, lapack_int ldu,
+                           lapack_complex_double* v, lapack_int ldv,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_int* ncycle );
+
+lapack_int LAPACKE_stgsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const float* a, lapack_int lda, const float* b,
+                           lapack_int ldb, const float* vl, lapack_int ldvl,
+                           const float* vr, lapack_int ldvr, float* s,
+                           float* dif, lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_dtgsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const double* a, lapack_int lda, const double* b,
+                           lapack_int ldb, const double* vl, lapack_int ldvl,
+                           const double* vr, lapack_int ldvr, double* s,
+                           double* dif, lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ctgsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           const lapack_complex_float* vl, lapack_int ldvl,
+                           const lapack_complex_float* vr, lapack_int ldvr,
+                           float* s, float* dif, lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ztgsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           const lapack_complex_double* vl, lapack_int ldvl,
+                           const lapack_complex_double* vr, lapack_int ldvr,
+                           double* s, double* dif, lapack_int mm,
+                           lapack_int* m );
+
+lapack_int LAPACKE_stgsyl( int matrix_order, char trans, lapack_int ijob,
+                           lapack_int m, lapack_int n, const float* a,
+                           lapack_int lda, const float* b, lapack_int ldb,
+                           float* c, lapack_int ldc, const float* d,
+                           lapack_int ldd, const float* e, lapack_int lde,
+                           float* f, lapack_int ldf, float* scale, float* dif );
+lapack_int LAPACKE_dtgsyl( int matrix_order, char trans, lapack_int ijob,
+                           lapack_int m, lapack_int n, const double* a,
+                           lapack_int lda, const double* b, lapack_int ldb,
+                           double* c, lapack_int ldc, const double* d,
+                           lapack_int ldd, const double* e, lapack_int lde,
+                           double* f, lapack_int ldf, double* scale,
+                           double* dif );
+lapack_int LAPACKE_ctgsyl( int matrix_order, char trans, lapack_int ijob,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* c, lapack_int ldc,
+                           const lapack_complex_float* d, lapack_int ldd,
+                           const lapack_complex_float* e, lapack_int lde,
+                           lapack_complex_float* f, lapack_int ldf,
+                           float* scale, float* dif );
+lapack_int LAPACKE_ztgsyl( int matrix_order, char trans, lapack_int ijob,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* c, lapack_int ldc,
+                           const lapack_complex_double* d, lapack_int ldd,
+                           const lapack_complex_double* e, lapack_int lde,
+                           lapack_complex_double* f, lapack_int ldf,
+                           double* scale, double* dif );
+
+lapack_int LAPACKE_stpcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const float* ap, float* rcond );
+lapack_int LAPACKE_dtpcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const double* ap, double* rcond );
+lapack_int LAPACKE_ctpcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const lapack_complex_float* ap,
+                           float* rcond );
+lapack_int LAPACKE_ztpcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const lapack_complex_double* ap,
+                           double* rcond );
+
+lapack_int LAPACKE_stprfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const float* ap,
+                           const float* b, lapack_int ldb, const float* x,
+                           lapack_int ldx, float* ferr, float* berr );
+lapack_int LAPACKE_dtprfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const double* ap,
+                           const double* b, lapack_int ldb, const double* x,
+                           lapack_int ldx, double* ferr, double* berr );
+lapack_int LAPACKE_ctprfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* ap,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           const lapack_complex_float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_ztprfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* ap,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           const lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_stptri( int matrix_order, char uplo, char diag, lapack_int n,
+                           float* ap );
+lapack_int LAPACKE_dtptri( int matrix_order, char uplo, char diag, lapack_int n,
+                           double* ap );
+lapack_int LAPACKE_ctptri( int matrix_order, char uplo, char diag, lapack_int n,
+                           lapack_complex_float* ap );
+lapack_int LAPACKE_ztptri( int matrix_order, char uplo, char diag, lapack_int n,
+                           lapack_complex_double* ap );
+
+lapack_int LAPACKE_stptrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const float* ap,
+                           float* b, lapack_int ldb );
+lapack_int LAPACKE_dtptrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const double* ap,
+                           double* b, lapack_int ldb );
+lapack_int LAPACKE_ctptrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* ap,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztptrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* ap,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stpttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const float* ap, float* arf );
+lapack_int LAPACKE_dtpttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const double* ap, double* arf );
+lapack_int LAPACKE_ctpttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_float* ap,
+                           lapack_complex_float* arf );
+lapack_int LAPACKE_ztpttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_double* ap,
+                           lapack_complex_double* arf );
+
+lapack_int LAPACKE_stpttr( int matrix_order, char uplo, lapack_int n,
+                           const float* ap, float* a, lapack_int lda );
+lapack_int LAPACKE_dtpttr( int matrix_order, char uplo, lapack_int n,
+                           const double* ap, double* a, lapack_int lda );
+lapack_int LAPACKE_ctpttr( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_ztpttr( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_strcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const float* a, lapack_int lda,
+                           float* rcond );
+lapack_int LAPACKE_dtrcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const double* a, lapack_int lda,
+                           double* rcond );
+lapack_int LAPACKE_ctrcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const lapack_complex_float* a,
+                           lapack_int lda, float* rcond );
+lapack_int LAPACKE_ztrcon( int matrix_order, char norm, char uplo, char diag,
+                           lapack_int n, const lapack_complex_double* a,
+                           lapack_int lda, double* rcond );
+
+lapack_int LAPACKE_strevc( int matrix_order, char side, char howmny,
+                           lapack_logical* select, lapack_int n, const float* t,
+                           lapack_int ldt, float* vl, lapack_int ldvl,
+                           float* vr, lapack_int ldvr, lapack_int mm,
+                           lapack_int* m );
+lapack_int LAPACKE_dtrevc( int matrix_order, char side, char howmny,
+                           lapack_logical* select, lapack_int n,
+                           const double* t, lapack_int ldt, double* vl,
+                           lapack_int ldvl, double* vr, lapack_int ldvr,
+                           lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ctrevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_float* t, lapack_int ldt,
+                           lapack_complex_float* vl, lapack_int ldvl,
+                           lapack_complex_float* vr, lapack_int ldvr,
+                           lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ztrevc( int matrix_order, char side, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_double* t, lapack_int ldt,
+                           lapack_complex_double* vl, lapack_int ldvl,
+                           lapack_complex_double* vr, lapack_int ldvr,
+                           lapack_int mm, lapack_int* m );
+
+lapack_int LAPACKE_strexc( int matrix_order, char compq, lapack_int n, float* t,
+                           lapack_int ldt, float* q, lapack_int ldq,
+                           lapack_int* ifst, lapack_int* ilst );
+lapack_int LAPACKE_dtrexc( int matrix_order, char compq, lapack_int n,
+                           double* t, lapack_int ldt, double* q, lapack_int ldq,
+                           lapack_int* ifst, lapack_int* ilst );
+lapack_int LAPACKE_ctrexc( int matrix_order, char compq, lapack_int n,
+                           lapack_complex_float* t, lapack_int ldt,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_int ifst, lapack_int ilst );
+lapack_int LAPACKE_ztrexc( int matrix_order, char compq, lapack_int n,
+                           lapack_complex_double* t, lapack_int ldt,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_int ifst, lapack_int ilst );
+
+lapack_int LAPACKE_strrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const float* a,
+                           lapack_int lda, const float* b, lapack_int ldb,
+                           const float* x, lapack_int ldx, float* ferr,
+                           float* berr );
+lapack_int LAPACKE_dtrrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const double* a,
+                           lapack_int lda, const double* b, lapack_int ldb,
+                           const double* x, lapack_int ldx, double* ferr,
+                           double* berr );
+lapack_int LAPACKE_ctrrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           const lapack_complex_float* x, lapack_int ldx,
+                           float* ferr, float* berr );
+lapack_int LAPACKE_ztrrfs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           const lapack_complex_double* x, lapack_int ldx,
+                           double* ferr, double* berr );
+
+lapack_int LAPACKE_strsen( int matrix_order, char job, char compq,
+                           const lapack_logical* select, lapack_int n, float* t,
+                           lapack_int ldt, float* q, lapack_int ldq, float* wr,
+                           float* wi, lapack_int* m, float* s, float* sep );
+lapack_int LAPACKE_dtrsen( int matrix_order, char job, char compq,
+                           const lapack_logical* select, lapack_int n,
+                           double* t, lapack_int ldt, double* q, lapack_int ldq,
+                           double* wr, double* wi, lapack_int* m, double* s,
+                           double* sep );
+lapack_int LAPACKE_ctrsen( int matrix_order, char job, char compq,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_float* t, lapack_int ldt,
+                           lapack_complex_float* q, lapack_int ldq,
+                           lapack_complex_float* w, lapack_int* m, float* s,
+                           float* sep );
+lapack_int LAPACKE_ztrsen( int matrix_order, char job, char compq,
+                           const lapack_logical* select, lapack_int n,
+                           lapack_complex_double* t, lapack_int ldt,
+                           lapack_complex_double* q, lapack_int ldq,
+                           lapack_complex_double* w, lapack_int* m, double* s,
+                           double* sep );
+
+lapack_int LAPACKE_strsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const float* t, lapack_int ldt, const float* vl,
+                           lapack_int ldvl, const float* vr, lapack_int ldvr,
+                           float* s, float* sep, lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_dtrsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const double* t, lapack_int ldt, const double* vl,
+                           lapack_int ldvl, const double* vr, lapack_int ldvr,
+                           double* s, double* sep, lapack_int mm,
+                           lapack_int* m );
+lapack_int LAPACKE_ctrsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_float* t, lapack_int ldt,
+                           const lapack_complex_float* vl, lapack_int ldvl,
+                           const lapack_complex_float* vr, lapack_int ldvr,
+                           float* s, float* sep, lapack_int mm, lapack_int* m );
+lapack_int LAPACKE_ztrsna( int matrix_order, char job, char howmny,
+                           const lapack_logical* select, lapack_int n,
+                           const lapack_complex_double* t, lapack_int ldt,
+                           const lapack_complex_double* vl, lapack_int ldvl,
+                           const lapack_complex_double* vr, lapack_int ldvr,
+                           double* s, double* sep, lapack_int mm,
+                           lapack_int* m );
+
+lapack_int LAPACKE_strsyl( int matrix_order, char trana, char tranb,
+                           lapack_int isgn, lapack_int m, lapack_int n,
+                           const float* a, lapack_int lda, const float* b,
+                           lapack_int ldb, float* c, lapack_int ldc,
+                           float* scale );
+lapack_int LAPACKE_dtrsyl( int matrix_order, char trana, char tranb,
+                           lapack_int isgn, lapack_int m, lapack_int n,
+                           const double* a, lapack_int lda, const double* b,
+                           lapack_int ldb, double* c, lapack_int ldc,
+                           double* scale );
+lapack_int LAPACKE_ctrsyl( int matrix_order, char trana, char tranb,
+                           lapack_int isgn, lapack_int m, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* b, lapack_int ldb,
+                           lapack_complex_float* c, lapack_int ldc,
+                           float* scale );
+lapack_int LAPACKE_ztrsyl( int matrix_order, char trana, char tranb,
+                           lapack_int isgn, lapack_int m, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* c, lapack_int ldc,
+                           double* scale );
+
+lapack_int LAPACKE_strtri( int matrix_order, char uplo, char diag, lapack_int n,
+                           float* a, lapack_int lda );
+lapack_int LAPACKE_dtrtri( int matrix_order, char uplo, char diag, lapack_int n,
+                           double* a, lapack_int lda );
+lapack_int LAPACKE_ctrtri( int matrix_order, char uplo, char diag, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_ztrtri( int matrix_order, char uplo, char diag, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_strtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const float* a,
+                           lapack_int lda, float* b, lapack_int ldb );
+lapack_int LAPACKE_dtrtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs, const double* a,
+                           lapack_int lda, double* b, lapack_int ldb );
+lapack_int LAPACKE_ctrtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztrtrs( int matrix_order, char uplo, char trans, char diag,
+                           lapack_int n, lapack_int nrhs,
+                           const lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_strttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const float* a, lapack_int lda,
+                           float* arf );
+lapack_int LAPACKE_dtrttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const double* a, lapack_int lda,
+                           double* arf );
+lapack_int LAPACKE_ctrttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* arf );
+lapack_int LAPACKE_ztrttf( int matrix_order, char transr, char uplo,
+                           lapack_int n, const lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* arf );
+
+lapack_int LAPACKE_strttp( int matrix_order, char uplo, lapack_int n,
+                           const float* a, lapack_int lda, float* ap );
+lapack_int LAPACKE_dtrttp( int matrix_order, char uplo, lapack_int n,
+                           const double* a, lapack_int lda, double* ap );
+lapack_int LAPACKE_ctrttp( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* ap );
+lapack_int LAPACKE_ztrttp( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* ap );
+
+lapack_int LAPACKE_stzrzf( int matrix_order, lapack_int m, lapack_int n,
+                           float* a, lapack_int lda, float* tau );
+lapack_int LAPACKE_dtzrzf( int matrix_order, lapack_int m, lapack_int n,
+                           double* a, lapack_int lda, double* tau );
+lapack_int LAPACKE_ctzrzf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* tau );
+lapack_int LAPACKE_ztzrzf( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* tau );
+
+lapack_int LAPACKE_cungbr( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int k, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zungbr( int matrix_order, char vect, lapack_int m,
+                           lapack_int n, lapack_int k, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cunghr( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zunghr( int matrix_order, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cunglq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zunglq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cungql( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zungql( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cungqr( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zungqr( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cungrq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau );
+lapack_int LAPACKE_zungrq( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cungtr( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau );
+lapack_int LAPACKE_zungtr( int matrix_order, char uplo, lapack_int n,
+                           lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau );
+
+lapack_int LAPACKE_cunmbr( int matrix_order, char vect, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmbr( int matrix_order, char vect, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmhr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmhr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int ilo,
+                           lapack_int ihi, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmlq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmlq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmql( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmql( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmqr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmqr( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmrq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmrq( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmrz( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           lapack_int l, const lapack_complex_float* a,
+                           lapack_int lda, const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmrz( int matrix_order, char side, char trans,
+                           lapack_int m, lapack_int n, lapack_int k,
+                           lapack_int l, const lapack_complex_double* a,
+                           lapack_int lda, const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cunmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_float* a, lapack_int lda,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zunmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_double* a, lapack_int lda,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_cupgtr( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_float* ap,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* q, lapack_int ldq );
+lapack_int LAPACKE_zupgtr( int matrix_order, char uplo, lapack_int n,
+                           const lapack_complex_double* ap,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* q, lapack_int ldq );
+
+lapack_int LAPACKE_cupmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_float* ap,
+                           const lapack_complex_float* tau,
+                           lapack_complex_float* c, lapack_int ldc );
+lapack_int LAPACKE_zupmtr( int matrix_order, char side, char uplo, char trans,
+                           lapack_int m, lapack_int n,
+                           const lapack_complex_double* ap,
+                           const lapack_complex_double* tau,
+                           lapack_complex_double* c, lapack_int ldc );
+
+lapack_int LAPACKE_sbdsdc_work( int matrix_order, char uplo, char compq,
+                                lapack_int n, float* d, float* e, float* u,
+                                lapack_int ldu, float* vt, lapack_int ldvt,
+                                float* q, lapack_int* iq, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dbdsdc_work( int matrix_order, char uplo, char compq,
+                                lapack_int n, double* d, double* e, double* u,
+                                lapack_int ldu, double* vt, lapack_int ldvt,
+                                double* q, lapack_int* iq, double* work,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_sbdsqr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                                float* d, float* e, float* vt, lapack_int ldvt,
+                                float* u, lapack_int ldu, float* c,
+                                lapack_int ldc, float* work );
+lapack_int LAPACKE_dbdsqr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                                double* d, double* e, double* vt,
+                                lapack_int ldvt, double* u, lapack_int ldu,
+                                double* c, lapack_int ldc, double* work );
+lapack_int LAPACKE_cbdsqr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                                float* d, float* e, lapack_complex_float* vt,
+                                lapack_int ldvt, lapack_complex_float* u,
+                                lapack_int ldu, lapack_complex_float* c,
+                                lapack_int ldc, float* work );
+lapack_int LAPACKE_zbdsqr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int ncvt, lapack_int nru, lapack_int ncc,
+                                double* d, double* e, lapack_complex_double* vt,
+                                lapack_int ldvt, lapack_complex_double* u,
+                                lapack_int ldu, lapack_complex_double* c,
+                                lapack_int ldc, double* work );
+
+lapack_int LAPACKE_sdisna_work( char job, lapack_int m, lapack_int n,
+                                const float* d, float* sep );
+lapack_int LAPACKE_ddisna_work( char job, lapack_int m, lapack_int n,
+                                const double* d, double* sep );
+
+lapack_int LAPACKE_sgbbrd_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int ncc, lapack_int kl,
+                                lapack_int ku, float* ab, lapack_int ldab,
+                                float* d, float* e, float* q, lapack_int ldq,
+                                float* pt, lapack_int ldpt, float* c,
+                                lapack_int ldc, float* work );
+lapack_int LAPACKE_dgbbrd_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int ncc, lapack_int kl,
+                                lapack_int ku, double* ab, lapack_int ldab,
+                                double* d, double* e, double* q, lapack_int ldq,
+                                double* pt, lapack_int ldpt, double* c,
+                                lapack_int ldc, double* work );
+lapack_int LAPACKE_cgbbrd_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int ncc, lapack_int kl,
+                                lapack_int ku, lapack_complex_float* ab,
+                                lapack_int ldab, float* d, float* e,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* pt, lapack_int ldpt,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgbbrd_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int ncc, lapack_int kl,
+                                lapack_int ku, lapack_complex_double* ab,
+                                lapack_int ldab, double* d, double* e,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* pt, lapack_int ldpt,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgbcon_work( int matrix_order, char norm, lapack_int n,
+                                lapack_int kl, lapack_int ku, const float* ab,
+                                lapack_int ldab, const lapack_int* ipiv,
+                                float anorm, float* rcond, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgbcon_work( int matrix_order, char norm, lapack_int n,
+                                lapack_int kl, lapack_int ku, const double* ab,
+                                lapack_int ldab, const lapack_int* ipiv,
+                                double anorm, double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cgbcon_work( int matrix_order, char norm, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zgbcon_work( int matrix_order, char norm, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, const lapack_int* ipiv,
+                                double anorm, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgbequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const float* ab,
+                                lapack_int ldab, float* r, float* c,
+                                float* rowcnd, float* colcnd, float* amax );
+lapack_int LAPACKE_dgbequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const double* ab,
+                                lapack_int ldab, double* r, double* c,
+                                double* rowcnd, double* colcnd, double* amax );
+lapack_int LAPACKE_cgbequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                float* r, float* c, float* rowcnd,
+                                float* colcnd, float* amax );
+lapack_int LAPACKE_zgbequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, double* r, double* c,
+                                double* rowcnd, double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgbequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_int kl, lapack_int ku, const float* ab,
+                                 lapack_int ldab, float* r, float* c,
+                                 float* rowcnd, float* colcnd, float* amax );
+lapack_int LAPACKE_dgbequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_int kl, lapack_int ku, const double* ab,
+                                 lapack_int ldab, double* r, double* c,
+                                 double* rowcnd, double* colcnd, double* amax );
+lapack_int LAPACKE_cgbequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_int kl, lapack_int ku,
+                                 const lapack_complex_float* ab,
+                                 lapack_int ldab, float* r, float* c,
+                                 float* rowcnd, float* colcnd, float* amax );
+lapack_int LAPACKE_zgbequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_int kl, lapack_int ku,
+                                 const lapack_complex_double* ab,
+                                 lapack_int ldab, double* r, double* c,
+                                 double* rowcnd, double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgbrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const float* ab, lapack_int ldab,
+                                const float* afb, lapack_int ldafb,
+                                const lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgbrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const double* ab, lapack_int ldab,
+                                const double* afb, lapack_int ldafb,
+                                const lapack_int* ipiv, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cgbrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                const lapack_complex_float* afb,
+                                lapack_int ldafb, const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgbrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab,
+                                const lapack_complex_double* afb,
+                                lapack_int ldafb, const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgbrfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, const float* ab,
+                                 lapack_int ldab, const float* afb,
+                                 lapack_int ldafb, const lapack_int* ipiv,
+                                 const float* r, const float* c, const float* b,
+                                 lapack_int ldb, float* x, lapack_int ldx,
+                                 float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dgbrfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, const double* ab,
+                                 lapack_int ldab, const double* afb,
+                                 lapack_int ldafb, const lapack_int* ipiv,
+                                 const double* r, const double* c,
+                                 const double* b, lapack_int ldb, double* x,
+                                 lapack_int ldx, double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cgbrfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs,
+                                 const lapack_complex_float* ab,
+                                 lapack_int ldab,
+                                 const lapack_complex_float* afb,
+                                 lapack_int ldafb, const lapack_int* ipiv,
+                                 const float* r, const float* c,
+                                 const lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zgbrfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs,
+                                 const lapack_complex_double* ab,
+                                 lapack_int ldab,
+                                 const lapack_complex_double* afb,
+                                 lapack_int ldafb, const lapack_int* ipiv,
+                                 const double* r, const double* c,
+                                 const lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_sgbsv_work( int matrix_order, lapack_int n, lapack_int kl,
+                               lapack_int ku, lapack_int nrhs, float* ab,
+                               lapack_int ldab, lapack_int* ipiv, float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_dgbsv_work( int matrix_order, lapack_int n, lapack_int kl,
+                               lapack_int ku, lapack_int nrhs, double* ab,
+                               lapack_int ldab, lapack_int* ipiv, double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_cgbsv_work( int matrix_order, lapack_int n, lapack_int kl,
+                               lapack_int ku, lapack_int nrhs,
+                               lapack_complex_float* ab, lapack_int ldab,
+                               lapack_int* ipiv, lapack_complex_float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_zgbsv_work( int matrix_order, lapack_int n, lapack_int kl,
+                               lapack_int ku, lapack_int nrhs,
+                               lapack_complex_double* ab, lapack_int ldab,
+                               lapack_int* ipiv, lapack_complex_double* b,
+                               lapack_int ldb );
+
+lapack_int LAPACKE_sgbsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int kl, lapack_int ku,
+                                lapack_int nrhs, float* ab, lapack_int ldab,
+                                float* afb, lapack_int ldafb, lapack_int* ipiv,
+                                char* equed, float* r, float* c, float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dgbsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int kl, lapack_int ku,
+                                lapack_int nrhs, double* ab, lapack_int ldab,
+                                double* afb, lapack_int ldafb, lapack_int* ipiv,
+                                char* equed, double* r, double* c, double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cgbsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int kl, lapack_int ku,
+                                lapack_int nrhs, lapack_complex_float* ab,
+                                lapack_int ldab, lapack_complex_float* afb,
+                                lapack_int ldafb, lapack_int* ipiv, char* equed,
+                                float* r, float* c, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zgbsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int kl, lapack_int ku,
+                                lapack_int nrhs, lapack_complex_double* ab,
+                                lapack_int ldab, lapack_complex_double* afb,
+                                lapack_int ldafb, lapack_int* ipiv, char* equed,
+                                double* r, double* c, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_sgbsvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, float* ab, lapack_int ldab,
+                                 float* afb, lapack_int ldafb, lapack_int* ipiv,
+                                 char* equed, float* r, float* c, float* b,
+                                 lapack_int ldb, float* x, lapack_int ldx,
+                                 float* rcond, float* rpvgrw, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dgbsvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, double* ab, lapack_int ldab,
+                                 double* afb, lapack_int ldafb,
+                                 lapack_int* ipiv, char* equed, double* r,
+                                 double* c, double* b, lapack_int ldb,
+                                 double* x, lapack_int ldx, double* rcond,
+                                 double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cgbsvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, lapack_complex_float* ab,
+                                 lapack_int ldab, lapack_complex_float* afb,
+                                 lapack_int ldafb, lapack_int* ipiv,
+                                 char* equed, float* r, float* c,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* rpvgrw, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zgbsvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int kl, lapack_int ku,
+                                 lapack_int nrhs, lapack_complex_double* ab,
+                                 lapack_int ldab, lapack_complex_double* afb,
+                                 lapack_int ldafb, lapack_int* ipiv,
+                                 char* equed, double* r, double* c,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_sgbtrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, float* ab,
+                                lapack_int ldab, lapack_int* ipiv );
+lapack_int LAPACKE_dgbtrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, double* ab,
+                                lapack_int ldab, lapack_int* ipiv );
+lapack_int LAPACKE_cgbtrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                lapack_int* ipiv );
+lapack_int LAPACKE_zgbtrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                lapack_int* ipiv );
+
+lapack_int LAPACKE_sgbtrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const float* ab, lapack_int ldab,
+                                const lapack_int* ipiv, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dgbtrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const double* ab, lapack_int ldab,
+                                const lapack_int* ipiv, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_cgbtrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                const lapack_int* ipiv, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zgbtrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int kl, lapack_int ku, lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sgebak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const float* scale, lapack_int m, float* v,
+                                lapack_int ldv );
+lapack_int LAPACKE_dgebak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const double* scale, lapack_int m, double* v,
+                                lapack_int ldv );
+lapack_int LAPACKE_cgebak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const float* scale, lapack_int m,
+                                lapack_complex_float* v, lapack_int ldv );
+lapack_int LAPACKE_zgebak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const double* scale, lapack_int m,
+                                lapack_complex_double* v, lapack_int ldv );
+
+lapack_int LAPACKE_sgebal_work( int matrix_order, char job, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* ilo,
+                                lapack_int* ihi, float* scale );
+lapack_int LAPACKE_dgebal_work( int matrix_order, char job, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* ilo,
+                                lapack_int* ihi, double* scale );
+lapack_int LAPACKE_cgebal_work( int matrix_order, char job, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* ilo, lapack_int* ihi,
+                                float* scale );
+lapack_int LAPACKE_zgebal_work( int matrix_order, char job, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ilo, lapack_int* ihi,
+                                double* scale );
+
+lapack_int LAPACKE_sgebrd_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* d, float* e,
+                                float* tauq, float* taup, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dgebrd_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* d, double* e,
+                                double* tauq, double* taup, double* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_cgebrd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                float* d, float* e, lapack_complex_float* tauq,
+                                lapack_complex_float* taup,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgebrd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                double* d, double* e,
+                                lapack_complex_double* tauq,
+                                lapack_complex_double* taup,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgecon_work( int matrix_order, char norm, lapack_int n,
+                                const float* a, lapack_int lda, float anorm,
+                                float* rcond, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dgecon_work( int matrix_order, char norm, lapack_int n,
+                                const double* a, lapack_int lda, double anorm,
+                                double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cgecon_work( int matrix_order, char norm, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float anorm, float* rcond,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgecon_work( int matrix_order, char norm, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double anorm, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgeequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                const float* a, lapack_int lda, float* r,
+                                float* c, float* rowcnd, float* colcnd,
+                                float* amax );
+lapack_int LAPACKE_dgeequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                const double* a, lapack_int lda, double* r,
+                                double* c, double* rowcnd, double* colcnd,
+                                double* amax );
+lapack_int LAPACKE_cgeequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float* r, float* c, float* rowcnd,
+                                float* colcnd, float* amax );
+lapack_int LAPACKE_zgeequ_work( int matrix_order, lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double* r, double* c, double* rowcnd,
+                                double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgeequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 const float* a, lapack_int lda, float* r,
+                                 float* c, float* rowcnd, float* colcnd,
+                                 float* amax );
+lapack_int LAPACKE_dgeequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 const double* a, lapack_int lda, double* r,
+                                 double* c, double* rowcnd, double* colcnd,
+                                 double* amax );
+lapack_int LAPACKE_cgeequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 float* r, float* c, float* rowcnd,
+                                 float* colcnd, float* amax );
+lapack_int LAPACKE_zgeequb_work( int matrix_order, lapack_int m, lapack_int n,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 double* r, double* c, double* rowcnd,
+                                 double* colcnd, double* amax );
+
+lapack_int LAPACKE_sgees_work( int matrix_order, char jobvs, char sort,
+                               LAPACK_S_SELECT2 select, lapack_int n, float* a,
+                               lapack_int lda, lapack_int* sdim, float* wr,
+                               float* wi, float* vs, lapack_int ldvs,
+                               float* work, lapack_int lwork,
+                               lapack_logical* bwork );
+lapack_int LAPACKE_dgees_work( int matrix_order, char jobvs, char sort,
+                               LAPACK_D_SELECT2 select, lapack_int n, double* a,
+                               lapack_int lda, lapack_int* sdim, double* wr,
+                               double* wi, double* vs, lapack_int ldvs,
+                               double* work, lapack_int lwork,
+                               lapack_logical* bwork );
+lapack_int LAPACKE_cgees_work( int matrix_order, char jobvs, char sort,
+                               LAPACK_C_SELECT1 select, lapack_int n,
+                               lapack_complex_float* a, lapack_int lda,
+                               lapack_int* sdim, lapack_complex_float* w,
+                               lapack_complex_float* vs, lapack_int ldvs,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork, lapack_logical* bwork );
+lapack_int LAPACKE_zgees_work( int matrix_order, char jobvs, char sort,
+                               LAPACK_Z_SELECT1 select, lapack_int n,
+                               lapack_complex_double* a, lapack_int lda,
+                               lapack_int* sdim, lapack_complex_double* w,
+                               lapack_complex_double* vs, lapack_int ldvs,
+                               lapack_complex_double* work, lapack_int lwork,
+                               double* rwork, lapack_logical* bwork );
+
+lapack_int LAPACKE_sgeesx_work( int matrix_order, char jobvs, char sort,
+                                LAPACK_S_SELECT2 select, char sense,
+                                lapack_int n, float* a, lapack_int lda,
+                                lapack_int* sdim, float* wr, float* wi,
+                                float* vs, lapack_int ldvs, float* rconde,
+                                float* rcondv, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork,
+                                lapack_logical* bwork );
+lapack_int LAPACKE_dgeesx_work( int matrix_order, char jobvs, char sort,
+                                LAPACK_D_SELECT2 select, char sense,
+                                lapack_int n, double* a, lapack_int lda,
+                                lapack_int* sdim, double* wr, double* wi,
+                                double* vs, lapack_int ldvs, double* rconde,
+                                double* rcondv, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork,
+                                lapack_logical* bwork );
+lapack_int LAPACKE_cgeesx_work( int matrix_order, char jobvs, char sort,
+                                LAPACK_C_SELECT1 select, char sense,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, lapack_int* sdim,
+                                lapack_complex_float* w,
+                                lapack_complex_float* vs, lapack_int ldvs,
+                                float* rconde, float* rcondv,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_logical* bwork );
+lapack_int LAPACKE_zgeesx_work( int matrix_order, char jobvs, char sort,
+                                LAPACK_Z_SELECT1 select, char sense,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, lapack_int* sdim,
+                                lapack_complex_double* w,
+                                lapack_complex_double* vs, lapack_int ldvs,
+                                double* rconde, double* rcondv,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_logical* bwork );
+
+lapack_int LAPACKE_sgeev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, float* a, lapack_int lda,
+                               float* wr, float* wi, float* vl, lapack_int ldvl,
+                               float* vr, lapack_int ldvr, float* work,
+                               lapack_int lwork );
+lapack_int LAPACKE_dgeev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, double* a, lapack_int lda,
+                               double* wr, double* wi, double* vl,
+                               lapack_int ldvl, double* vr, lapack_int ldvr,
+                               double* work, lapack_int lwork );
+lapack_int LAPACKE_cgeev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, lapack_complex_float* a,
+                               lapack_int lda, lapack_complex_float* w,
+                               lapack_complex_float* vl, lapack_int ldvl,
+                               lapack_complex_float* vr, lapack_int ldvr,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork );
+lapack_int LAPACKE_zgeev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, lapack_complex_double* a,
+                               lapack_int lda, lapack_complex_double* w,
+                               lapack_complex_double* vl, lapack_int ldvl,
+                               lapack_complex_double* vr, lapack_int ldvr,
+                               lapack_complex_double* work, lapack_int lwork,
+                               double* rwork );
+
+lapack_int LAPACKE_sgeevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n, float* a,
+                                lapack_int lda, float* wr, float* wi, float* vl,
+                                lapack_int ldvl, float* vr, lapack_int ldvr,
+                                lapack_int* ilo, lapack_int* ihi, float* scale,
+                                float* abnrm, float* rconde, float* rcondv,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgeevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n, double* a,
+                                lapack_int lda, double* wr, double* wi,
+                                double* vl, lapack_int ldvl, double* vr,
+                                lapack_int ldvr, lapack_int* ilo,
+                                lapack_int* ihi, double* scale, double* abnrm,
+                                double* rconde, double* rcondv, double* work,
+                                lapack_int lwork, lapack_int* iwork );
+lapack_int LAPACKE_cgeevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* w,
+                                lapack_complex_float* vl, lapack_int ldvl,
+                                lapack_complex_float* vr, lapack_int ldvr,
+                                lapack_int* ilo, lapack_int* ihi, float* scale,
+                                float* abnrm, float* rconde, float* rcondv,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork );
+lapack_int LAPACKE_zgeevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* w,
+                                lapack_complex_double* vl, lapack_int ldvl,
+                                lapack_complex_double* vr, lapack_int ldvr,
+                                lapack_int* ilo, lapack_int* ihi, double* scale,
+                                double* abnrm, double* rconde, double* rcondv,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork );
+
+lapack_int LAPACKE_sgehrd_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, float* a, lapack_int lda,
+                                float* tau, float* work, lapack_int lwork );
+lapack_int LAPACKE_dgehrd_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, double* a, lapack_int lda,
+                                double* tau, double* work, lapack_int lwork );
+lapack_int LAPACKE_cgehrd_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgehrd_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgejsv_work( int matrix_order, char joba, char jobu,
+                                char jobv, char jobr, char jobt, char jobp,
+                                lapack_int m, lapack_int n, float* a,
+                                lapack_int lda, float* sva, float* u,
+                                lapack_int ldu, float* v, lapack_int ldv,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgejsv_work( int matrix_order, char joba, char jobu,
+                                char jobv, char jobr, char jobt, char jobp,
+                                lapack_int m, lapack_int n, double* a,
+                                lapack_int lda, double* sva, double* u,
+                                lapack_int ldu, double* v, lapack_int ldv,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_sgelq2_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work );
+lapack_int LAPACKE_dgelq2_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work );
+lapack_int LAPACKE_cgelq2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zgelq2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_sgelqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgelqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgelqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgelqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgels_work( int matrix_order, char trans, lapack_int m,
+                               lapack_int n, lapack_int nrhs, float* a,
+                               lapack_int lda, float* b, lapack_int ldb,
+                               float* work, lapack_int lwork );
+lapack_int LAPACKE_dgels_work( int matrix_order, char trans, lapack_int m,
+                               lapack_int n, lapack_int nrhs, double* a,
+                               lapack_int lda, double* b, lapack_int ldb,
+                               double* work, lapack_int lwork );
+lapack_int LAPACKE_cgels_work( int matrix_order, char trans, lapack_int m,
+                               lapack_int n, lapack_int nrhs,
+                               lapack_complex_float* a, lapack_int lda,
+                               lapack_complex_float* b, lapack_int ldb,
+                               lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgels_work( int matrix_order, char trans, lapack_int m,
+                               lapack_int n, lapack_int nrhs,
+                               lapack_complex_double* a, lapack_int lda,
+                               lapack_complex_double* b, lapack_int ldb,
+                               lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgelsd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, float* s, float rcond,
+                                lapack_int* rank, float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgelsd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double* s,
+                                double rcond, lapack_int* rank, double* work,
+                                lapack_int lwork, lapack_int* iwork );
+lapack_int LAPACKE_cgelsd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, float* s, float rcond,
+                                lapack_int* rank, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_zgelsd_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, double* s, double rcond,
+                                lapack_int* rank, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_sgelss_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, float* s, float rcond,
+                                lapack_int* rank, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dgelss_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double* s,
+                                double rcond, lapack_int* rank, double* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_cgelss_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, float* s, float rcond,
+                                lapack_int* rank, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork );
+lapack_int LAPACKE_zgelss_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, double* s, double rcond,
+                                lapack_int* rank, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork );
+
+lapack_int LAPACKE_sgelsy_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, lapack_int* jpvt,
+                                float rcond, lapack_int* rank, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dgelsy_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, lapack_int* jpvt,
+                                double rcond, lapack_int* rank, double* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_cgelsy_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, lapack_int* jpvt, float rcond,
+                                lapack_int* rank, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork );
+lapack_int LAPACKE_zgelsy_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nrhs, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, lapack_int* jpvt, double rcond,
+                                lapack_int* rank, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork );
+
+lapack_int LAPACKE_sgeqlf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgeqlf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgeqlf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgeqlf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgeqp3_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* jpvt,
+                                float* tau, float* work, lapack_int lwork );
+lapack_int LAPACKE_dgeqp3_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* jpvt,
+                                double* tau, double* work, lapack_int lwork );
+lapack_int LAPACKE_cgeqp3_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* jpvt, lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork );
+lapack_int LAPACKE_zgeqp3_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* jpvt, lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork );
+
+lapack_int LAPACKE_sgeqpf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* jpvt,
+                                float* tau, float* work );
+lapack_int LAPACKE_dgeqpf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* jpvt,
+                                double* tau, double* work );
+lapack_int LAPACKE_cgeqpf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* jpvt, lapack_complex_float* tau,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgeqpf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* jpvt, lapack_complex_double* tau,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgeqr2_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work );
+lapack_int LAPACKE_dgeqr2_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work );
+lapack_int LAPACKE_cgeqr2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zgeqr2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_sgeqrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgeqrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgeqrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgeqrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgeqrfp_work( int matrix_order, lapack_int m, lapack_int n,
+                                 float* a, lapack_int lda, float* tau,
+                                 float* work, lapack_int lwork );
+lapack_int LAPACKE_dgeqrfp_work( int matrix_order, lapack_int m, lapack_int n,
+                                 double* a, lapack_int lda, double* tau,
+                                 double* work, lapack_int lwork );
+lapack_int LAPACKE_cgeqrfp_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* tau,
+                                 lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgeqrfp_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* tau,
+                                 lapack_complex_double* work,
+                                 lapack_int lwork );
+
+lapack_int LAPACKE_sgerfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                const float* af, lapack_int ldaf,
+                                const lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgerfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, const double* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cgerfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgerfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_complex_double* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgerfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int nrhs, const float* a,
+                                 lapack_int lda, const float* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const float* r, const float* c, const float* b,
+                                 lapack_int ldb, float* x, lapack_int ldx,
+                                 float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dgerfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int nrhs, const double* a,
+                                 lapack_int lda, const double* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const double* r, const double* c,
+                                 const double* b, lapack_int ldb, double* x,
+                                 lapack_int ldx, double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cgerfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 const lapack_complex_float* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const float* r, const float* c,
+                                 const lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zgerfsx_work( int matrix_order, char trans, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 const lapack_complex_double* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const double* r, const double* c,
+                                 const lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_sgerqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgerqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgerqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgerqf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgesdd_work( int matrix_order, char jobz, lapack_int m,
+                                lapack_int n, float* a, lapack_int lda,
+                                float* s, float* u, lapack_int ldu, float* vt,
+                                lapack_int ldvt, float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgesdd_work( int matrix_order, char jobz, lapack_int m,
+                                lapack_int n, double* a, lapack_int lda,
+                                double* s, double* u, lapack_int ldu,
+                                double* vt, lapack_int ldvt, double* work,
+                                lapack_int lwork, lapack_int* iwork );
+lapack_int LAPACKE_cgesdd_work( int matrix_order, char jobz, lapack_int m,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, float* s,
+                                lapack_complex_float* u, lapack_int ldu,
+                                lapack_complex_float* vt, lapack_int ldvt,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int* iwork );
+lapack_int LAPACKE_zgesdd_work( int matrix_order, char jobz, lapack_int m,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, double* s,
+                                lapack_complex_double* u, lapack_int ldu,
+                                lapack_complex_double* vt, lapack_int ldvt,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int* iwork );
+
+lapack_int LAPACKE_sgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               float* a, lapack_int lda, lapack_int* ipiv,
+                               float* b, lapack_int ldb );
+lapack_int LAPACKE_dgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               double* a, lapack_int lda, lapack_int* ipiv,
+                               double* b, lapack_int ldb );
+lapack_int LAPACKE_cgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               lapack_complex_float* a, lapack_int lda,
+                               lapack_int* ipiv, lapack_complex_float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_zgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               lapack_complex_double* a, lapack_int lda,
+                               lapack_int* ipiv, lapack_complex_double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_dsgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                double* a, lapack_int lda, lapack_int* ipiv,
+                                double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* work, float* swork,
+                                lapack_int* iter );
+lapack_int LAPACKE_zcgesv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ipiv, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, lapack_complex_double* work,
+                                lapack_complex_float* swork, double* rwork,
+                                lapack_int* iter );
+
+lapack_int LAPACKE_sgesvd_work( int matrix_order, char jobu, char jobvt,
+                                lapack_int m, lapack_int n, float* a,
+                                lapack_int lda, float* s, float* u,
+                                lapack_int ldu, float* vt, lapack_int ldvt,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgesvd_work( int matrix_order, char jobu, char jobvt,
+                                lapack_int m, lapack_int n, double* a,
+                                lapack_int lda, double* s, double* u,
+                                lapack_int ldu, double* vt, lapack_int ldvt,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgesvd_work( int matrix_order, char jobu, char jobvt,
+                                lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                float* s, lapack_complex_float* u,
+                                lapack_int ldu, lapack_complex_float* vt,
+                                lapack_int ldvt, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork );
+lapack_int LAPACKE_zgesvd_work( int matrix_order, char jobu, char jobvt,
+                                lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                double* s, lapack_complex_double* u,
+                                lapack_int ldu, lapack_complex_double* vt,
+                                lapack_int ldvt, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork );
+
+lapack_int LAPACKE_sgesvj_work( int matrix_order, char joba, char jobu,
+                                char jobv, lapack_int m, lapack_int n, float* a,
+                                lapack_int lda, float* sva, lapack_int mv,
+                                float* v, lapack_int ldv, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dgesvj_work( int matrix_order, char joba, char jobu,
+                                char jobv, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* sva,
+                                lapack_int mv, double* v, lapack_int ldv,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgesvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs, float* a,
+                                lapack_int lda, float* af, lapack_int ldaf,
+                                lapack_int* ipiv, char* equed, float* r,
+                                float* c, float* b, lapack_int ldb, float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dgesvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs, double* a,
+                                lapack_int lda, double* af, lapack_int ldaf,
+                                lapack_int* ipiv, char* equed, double* r,
+                                double* c, double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, double* work, lapack_int* iwork );
+lapack_int LAPACKE_cgesvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* af, lapack_int ldaf,
+                                lapack_int* ipiv, char* equed, float* r,
+                                float* c, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zgesvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* af, lapack_int ldaf,
+                                lapack_int* ipiv, char* equed, double* r,
+                                double* c, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_sgesvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int nrhs, float* a,
+                                 lapack_int lda, float* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, float* r,
+                                 float* c, float* b, lapack_int ldb, float* x,
+                                 lapack_int ldx, float* rcond, float* rpvgrw,
+                                 float* berr, lapack_int n_err_bnds,
+                                 float* err_bnds_norm, float* err_bnds_comp,
+                                 lapack_int nparams, float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dgesvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int nrhs, double* a,
+                                 lapack_int lda, double* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, double* r,
+                                 double* c, double* b, lapack_int ldb,
+                                 double* x, lapack_int ldx, double* rcond,
+                                 double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cgesvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, float* r,
+                                 float* c, lapack_complex_float* b,
+                                 lapack_int ldb, lapack_complex_float* x,
+                                 lapack_int ldx, float* rcond, float* rpvgrw,
+                                 float* berr, lapack_int n_err_bnds,
+                                 float* err_bnds_norm, float* err_bnds_comp,
+                                 lapack_int nparams, float* params,
+                                 lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgesvxx_work( int matrix_order, char fact, char trans,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, double* r,
+                                 double* c, lapack_complex_double* b,
+                                 lapack_int ldb, lapack_complex_double* x,
+                                 lapack_int ldx, double* rcond, double* rpvgrw,
+                                 double* berr, lapack_int n_err_bnds,
+                                 double* err_bnds_norm, double* err_bnds_comp,
+                                 lapack_int nparams, double* params,
+                                 lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgetf2_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_dgetf2_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_cgetf2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* ipiv );
+lapack_int LAPACKE_zgetf2_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ipiv );
+
+lapack_int LAPACKE_sgetrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_dgetrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* ipiv );
+lapack_int LAPACKE_cgetrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* ipiv );
+lapack_int LAPACKE_zgetrf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ipiv );
+
+lapack_int LAPACKE_sgetri_work( int matrix_order, lapack_int n, float* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dgetri_work( int matrix_order, lapack_int n, double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_cgetri_work( int matrix_order, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgetri_work( int matrix_order, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgetrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                const lapack_int* ipiv, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dgetrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                double* b, lapack_int ldb );
+lapack_int LAPACKE_cgetrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zgetrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sggbak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const float* lscale, const float* rscale,
+                                lapack_int m, float* v, lapack_int ldv );
+lapack_int LAPACKE_dggbak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const double* lscale, const double* rscale,
+                                lapack_int m, double* v, lapack_int ldv );
+lapack_int LAPACKE_cggbak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const float* lscale, const float* rscale,
+                                lapack_int m, lapack_complex_float* v,
+                                lapack_int ldv );
+lapack_int LAPACKE_zggbak_work( int matrix_order, char job, char side,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                const double* lscale, const double* rscale,
+                                lapack_int m, lapack_complex_double* v,
+                                lapack_int ldv );
+
+lapack_int LAPACKE_sggbal_work( int matrix_order, char job, lapack_int n,
+                                float* a, lapack_int lda, float* b,
+                                lapack_int ldb, lapack_int* ilo,
+                                lapack_int* ihi, float* lscale, float* rscale,
+                                float* work );
+lapack_int LAPACKE_dggbal_work( int matrix_order, char job, lapack_int n,
+                                double* a, lapack_int lda, double* b,
+                                lapack_int ldb, lapack_int* ilo,
+                                lapack_int* ihi, double* lscale, double* rscale,
+                                double* work );
+lapack_int LAPACKE_cggbal_work( int matrix_order, char job, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_int* ilo, lapack_int* ihi, float* lscale,
+                                float* rscale, float* work );
+lapack_int LAPACKE_zggbal_work( int matrix_order, char job, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_int* ilo, lapack_int* ihi,
+                                double* lscale, double* rscale, double* work );
+
+lapack_int LAPACKE_sgges_work( int matrix_order, char jobvsl, char jobvsr,
+                               char sort, LAPACK_S_SELECT3 selctg, lapack_int n,
+                               float* a, lapack_int lda, float* b,
+                               lapack_int ldb, lapack_int* sdim, float* alphar,
+                               float* alphai, float* beta, float* vsl,
+                               lapack_int ldvsl, float* vsr, lapack_int ldvsr,
+                               float* work, lapack_int lwork,
+                               lapack_logical* bwork );
+lapack_int LAPACKE_dgges_work( int matrix_order, char jobvsl, char jobvsr,
+                               char sort, LAPACK_D_SELECT3 selctg, lapack_int n,
+                               double* a, lapack_int lda, double* b,
+                               lapack_int ldb, lapack_int* sdim, double* alphar,
+                               double* alphai, double* beta, double* vsl,
+                               lapack_int ldvsl, double* vsr, lapack_int ldvsr,
+                               double* work, lapack_int lwork,
+                               lapack_logical* bwork );
+lapack_int LAPACKE_cgges_work( int matrix_order, char jobvsl, char jobvsr,
+                               char sort, LAPACK_C_SELECT2 selctg, lapack_int n,
+                               lapack_complex_float* a, lapack_int lda,
+                               lapack_complex_float* b, lapack_int ldb,
+                               lapack_int* sdim, lapack_complex_float* alpha,
+                               lapack_complex_float* beta,
+                               lapack_complex_float* vsl, lapack_int ldvsl,
+                               lapack_complex_float* vsr, lapack_int ldvsr,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork, lapack_logical* bwork );
+lapack_int LAPACKE_zgges_work( int matrix_order, char jobvsl, char jobvsr,
+                               char sort, LAPACK_Z_SELECT2 selctg, lapack_int n,
+                               lapack_complex_double* a, lapack_int lda,
+                               lapack_complex_double* b, lapack_int ldb,
+                               lapack_int* sdim, lapack_complex_double* alpha,
+                               lapack_complex_double* beta,
+                               lapack_complex_double* vsl, lapack_int ldvsl,
+                               lapack_complex_double* vsr, lapack_int ldvsr,
+                               lapack_complex_double* work, lapack_int lwork,
+                               double* rwork, lapack_logical* bwork );
+
+lapack_int LAPACKE_sggesx_work( int matrix_order, char jobvsl, char jobvsr,
+                                char sort, LAPACK_S_SELECT3 selctg, char sense,
+                                lapack_int n, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, lapack_int* sdim,
+                                float* alphar, float* alphai, float* beta,
+                                float* vsl, lapack_int ldvsl, float* vsr,
+                                lapack_int ldvsr, float* rconde, float* rcondv,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork,
+                                lapack_logical* bwork );
+lapack_int LAPACKE_dggesx_work( int matrix_order, char jobvsl, char jobvsr,
+                                char sort, LAPACK_D_SELECT3 selctg, char sense,
+                                lapack_int n, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, lapack_int* sdim,
+                                double* alphar, double* alphai, double* beta,
+                                double* vsl, lapack_int ldvsl, double* vsr,
+                                lapack_int ldvsr, double* rconde,
+                                double* rcondv, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork,
+                                lapack_logical* bwork );
+lapack_int LAPACKE_cggesx_work( int matrix_order, char jobvsl, char jobvsr,
+                                char sort, LAPACK_C_SELECT2 selctg, char sense,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, lapack_int* sdim,
+                                lapack_complex_float* alpha,
+                                lapack_complex_float* beta,
+                                lapack_complex_float* vsl, lapack_int ldvsl,
+                                lapack_complex_float* vsr, lapack_int ldvsr,
+                                float* rconde, float* rcondv,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int* iwork,
+                                lapack_int liwork, lapack_logical* bwork );
+lapack_int LAPACKE_zggesx_work( int matrix_order, char jobvsl, char jobvsr,
+                                char sort, LAPACK_Z_SELECT2 selctg, char sense,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, lapack_int* sdim,
+                                lapack_complex_double* alpha,
+                                lapack_complex_double* beta,
+                                lapack_complex_double* vsl, lapack_int ldvsl,
+                                lapack_complex_double* vsr, lapack_int ldvsr,
+                                double* rconde, double* rcondv,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int* iwork,
+                                lapack_int liwork, lapack_logical* bwork );
+
+lapack_int LAPACKE_sggev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, float* a, lapack_int lda, float* b,
+                               lapack_int ldb, float* alphar, float* alphai,
+                               float* beta, float* vl, lapack_int ldvl,
+                               float* vr, lapack_int ldvr, float* work,
+                               lapack_int lwork );
+lapack_int LAPACKE_dggev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, double* a, lapack_int lda,
+                               double* b, lapack_int ldb, double* alphar,
+                               double* alphai, double* beta, double* vl,
+                               lapack_int ldvl, double* vr, lapack_int ldvr,
+                               double* work, lapack_int lwork );
+lapack_int LAPACKE_cggev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, lapack_complex_float* a,
+                               lapack_int lda, lapack_complex_float* b,
+                               lapack_int ldb, lapack_complex_float* alpha,
+                               lapack_complex_float* beta,
+                               lapack_complex_float* vl, lapack_int ldvl,
+                               lapack_complex_float* vr, lapack_int ldvr,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork );
+lapack_int LAPACKE_zggev_work( int matrix_order, char jobvl, char jobvr,
+                               lapack_int n, lapack_complex_double* a,
+                               lapack_int lda, lapack_complex_double* b,
+                               lapack_int ldb, lapack_complex_double* alpha,
+                               lapack_complex_double* beta,
+                               lapack_complex_double* vl, lapack_int ldvl,
+                               lapack_complex_double* vr, lapack_int ldvr,
+                               lapack_complex_double* work, lapack_int lwork,
+                               double* rwork );
+
+lapack_int LAPACKE_sggevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n, float* a,
+                                lapack_int lda, float* b, lapack_int ldb,
+                                float* alphar, float* alphai, float* beta,
+                                float* vl, lapack_int ldvl, float* vr,
+                                lapack_int ldvr, lapack_int* ilo,
+                                lapack_int* ihi, float* lscale, float* rscale,
+                                float* abnrm, float* bbnrm, float* rconde,
+                                float* rcondv, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_logical* bwork );
+lapack_int LAPACKE_dggevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                double* alphar, double* alphai, double* beta,
+                                double* vl, lapack_int ldvl, double* vr,
+                                lapack_int ldvr, lapack_int* ilo,
+                                lapack_int* ihi, double* lscale, double* rscale,
+                                double* abnrm, double* bbnrm, double* rconde,
+                                double* rcondv, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_logical* bwork );
+lapack_int LAPACKE_cggevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* alpha,
+                                lapack_complex_float* beta,
+                                lapack_complex_float* vl, lapack_int ldvl,
+                                lapack_complex_float* vr, lapack_int ldvr,
+                                lapack_int* ilo, lapack_int* ihi, float* lscale,
+                                float* rscale, float* abnrm, float* bbnrm,
+                                float* rconde, float* rcondv,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int* iwork,
+                                lapack_logical* bwork );
+lapack_int LAPACKE_zggevx_work( int matrix_order, char balanc, char jobvl,
+                                char jobvr, char sense, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* alpha,
+                                lapack_complex_double* beta,
+                                lapack_complex_double* vl, lapack_int ldvl,
+                                lapack_complex_double* vr, lapack_int ldvr,
+                                lapack_int* ilo, lapack_int* ihi,
+                                double* lscale, double* rscale, double* abnrm,
+                                double* bbnrm, double* rconde, double* rcondv,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int* iwork,
+                                lapack_logical* bwork );
+
+lapack_int LAPACKE_sggglm_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, float* d, float* x,
+                                float* y, float* work, lapack_int lwork );
+lapack_int LAPACKE_dggglm_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double* d, double* x,
+                                double* y, double* work, lapack_int lwork );
+lapack_int LAPACKE_cggglm_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* d,
+                                lapack_complex_float* x,
+                                lapack_complex_float* y,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zggglm_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* d,
+                                lapack_complex_double* x,
+                                lapack_complex_double* y,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sgghrd_work( int matrix_order, char compq, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                float* a, lapack_int lda, float* b,
+                                lapack_int ldb, float* q, lapack_int ldq,
+                                float* z, lapack_int ldz );
+lapack_int LAPACKE_dgghrd_work( int matrix_order, char compq, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                double* a, lapack_int lda, double* b,
+                                lapack_int ldb, double* q, lapack_int ldq,
+                                double* z, lapack_int ldz );
+lapack_int LAPACKE_cgghrd_work( int matrix_order, char compq, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* z, lapack_int ldz );
+lapack_int LAPACKE_zgghrd_work( int matrix_order, char compq, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* z, lapack_int ldz );
+
+lapack_int LAPACKE_sgglse_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int p, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, float* c, float* d,
+                                float* x, float* work, lapack_int lwork );
+lapack_int LAPACKE_dgglse_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int p, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double* c, double* d,
+                                double* x, double* work, lapack_int lwork );
+lapack_int LAPACKE_cgglse_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* c,
+                                lapack_complex_float* d,
+                                lapack_complex_float* x,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zgglse_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* c,
+                                lapack_complex_double* d,
+                                lapack_complex_double* x,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sggqrf_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, float* a, lapack_int lda,
+                                float* taua, float* b, lapack_int ldb,
+                                float* taub, float* work, lapack_int lwork );
+lapack_int LAPACKE_dggqrf_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, double* a, lapack_int lda,
+                                double* taua, double* b, lapack_int ldb,
+                                double* taub, double* work, lapack_int lwork );
+lapack_int LAPACKE_cggqrf_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* taua,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* taub,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zggqrf_work( int matrix_order, lapack_int n, lapack_int m,
+                                lapack_int p, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* taua,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* taub,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sggrqf_work( int matrix_order, lapack_int m, lapack_int p,
+                                lapack_int n, float* a, lapack_int lda,
+                                float* taua, float* b, lapack_int ldb,
+                                float* taub, float* work, lapack_int lwork );
+lapack_int LAPACKE_dggrqf_work( int matrix_order, lapack_int m, lapack_int p,
+                                lapack_int n, double* a, lapack_int lda,
+                                double* taua, double* b, lapack_int ldb,
+                                double* taub, double* work, lapack_int lwork );
+lapack_int LAPACKE_cggrqf_work( int matrix_order, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* taua,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* taub,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zggrqf_work( int matrix_order, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* taua,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* taub,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sggsvd_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_int* k, lapack_int* l,
+                                float* a, lapack_int lda, float* b,
+                                lapack_int ldb, float* alpha, float* beta,
+                                float* u, lapack_int ldu, float* v,
+                                lapack_int ldv, float* q, lapack_int ldq,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dggsvd_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_int* k, lapack_int* l,
+                                double* a, lapack_int lda, double* b,
+                                lapack_int ldb, double* alpha, double* beta,
+                                double* u, lapack_int ldu, double* v,
+                                lapack_int ldv, double* q, lapack_int ldq,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cggsvd_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_int* k, lapack_int* l,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                float* alpha, float* beta,
+                                lapack_complex_float* u, lapack_int ldu,
+                                lapack_complex_float* v, lapack_int ldv,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* work, float* rwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_zggsvd_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int n,
+                                lapack_int p, lapack_int* k, lapack_int* l,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                double* alpha, double* beta,
+                                lapack_complex_double* u, lapack_int ldu,
+                                lapack_complex_double* v, lapack_int ldv,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* work, double* rwork,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_sggsvp_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, float tola,
+                                float tolb, lapack_int* k, lapack_int* l,
+                                float* u, lapack_int ldu, float* v,
+                                lapack_int ldv, float* q, lapack_int ldq,
+                                lapack_int* iwork, float* tau, float* work );
+lapack_int LAPACKE_dggsvp_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double tola,
+                                double tolb, lapack_int* k, lapack_int* l,
+                                double* u, lapack_int ldu, double* v,
+                                lapack_int ldv, double* q, lapack_int ldq,
+                                lapack_int* iwork, double* tau, double* work );
+lapack_int LAPACKE_cggsvp_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb, float tola, float tolb,
+                                lapack_int* k, lapack_int* l,
+                                lapack_complex_float* u, lapack_int ldu,
+                                lapack_complex_float* v, lapack_int ldv,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_int* iwork, float* rwork,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zggsvp_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, double tola, double tolb,
+                                lapack_int* k, lapack_int* l,
+                                lapack_complex_double* u, lapack_int ldu,
+                                lapack_complex_double* v, lapack_int ldv,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_int* iwork, double* rwork,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_sgtcon_work( char norm, lapack_int n, const float* dl,
+                                const float* d, const float* du,
+                                const float* du2, const lapack_int* ipiv,
+                                float anorm, float* rcond, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgtcon_work( char norm, lapack_int n, const double* dl,
+                                const double* d, const double* du,
+                                const double* du2, const lapack_int* ipiv,
+                                double anorm, double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cgtcon_work( char norm, lapack_int n,
+                                const lapack_complex_float* dl,
+                                const lapack_complex_float* d,
+                                const lapack_complex_float* du,
+                                const lapack_complex_float* du2,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work );
+lapack_int LAPACKE_zgtcon_work( char norm, lapack_int n,
+                                const lapack_complex_double* dl,
+                                const lapack_complex_double* d,
+                                const lapack_complex_double* du,
+                                const lapack_complex_double* du2,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, lapack_complex_double* work );
+
+lapack_int LAPACKE_sgtrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const float* dl,
+                                const float* d, const float* du,
+                                const float* dlf, const float* df,
+                                const float* duf, const float* du2,
+                                const lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dgtrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const double* dl,
+                                const double* d, const double* du,
+                                const double* dlf, const double* df,
+                                const double* duf, const double* du2,
+                                const lapack_int* ipiv, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cgtrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* dl,
+                                const lapack_complex_float* d,
+                                const lapack_complex_float* du,
+                                const lapack_complex_float* dlf,
+                                const lapack_complex_float* df,
+                                const lapack_complex_float* duf,
+                                const lapack_complex_float* du2,
+                                const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgtrfs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* dl,
+                                const lapack_complex_double* d,
+                                const lapack_complex_double* du,
+                                const lapack_complex_double* dlf,
+                                const lapack_complex_double* df,
+                                const lapack_complex_double* duf,
+                                const lapack_complex_double* du2,
+                                const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgtsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               float* dl, float* d, float* du, float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_dgtsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               double* dl, double* d, double* du, double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_cgtsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               lapack_complex_float* dl,
+                               lapack_complex_float* d,
+                               lapack_complex_float* du,
+                               lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zgtsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               lapack_complex_double* dl,
+                               lapack_complex_double* d,
+                               lapack_complex_double* du,
+                               lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sgtsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs, const float* dl,
+                                const float* d, const float* du, float* dlf,
+                                float* df, float* duf, float* du2,
+                                lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dgtsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs, const double* dl,
+                                const double* d, const double* du, double* dlf,
+                                double* df, double* duf, double* du2,
+                                lapack_int* ipiv, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cgtsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* dl,
+                                const lapack_complex_float* d,
+                                const lapack_complex_float* du,
+                                lapack_complex_float* dlf,
+                                lapack_complex_float* df,
+                                lapack_complex_float* duf,
+                                lapack_complex_float* du2, lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zgtsvx_work( int matrix_order, char fact, char trans,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* dl,
+                                const lapack_complex_double* d,
+                                const lapack_complex_double* du,
+                                lapack_complex_double* dlf,
+                                lapack_complex_double* df,
+                                lapack_complex_double* duf,
+                                lapack_complex_double* du2, lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sgttrf_work( lapack_int n, float* dl, float* d, float* du,
+                                float* du2, lapack_int* ipiv );
+lapack_int LAPACKE_dgttrf_work( lapack_int n, double* dl, double* d, double* du,
+                                double* du2, lapack_int* ipiv );
+lapack_int LAPACKE_cgttrf_work( lapack_int n, lapack_complex_float* dl,
+                                lapack_complex_float* d,
+                                lapack_complex_float* du,
+                                lapack_complex_float* du2, lapack_int* ipiv );
+lapack_int LAPACKE_zgttrf_work( lapack_int n, lapack_complex_double* dl,
+                                lapack_complex_double* d,
+                                lapack_complex_double* du,
+                                lapack_complex_double* du2, lapack_int* ipiv );
+
+lapack_int LAPACKE_sgttrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const float* dl,
+                                const float* d, const float* du,
+                                const float* du2, const lapack_int* ipiv,
+                                float* b, lapack_int ldb );
+lapack_int LAPACKE_dgttrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const double* dl,
+                                const double* d, const double* du,
+                                const double* du2, const lapack_int* ipiv,
+                                double* b, lapack_int ldb );
+lapack_int LAPACKE_cgttrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* dl,
+                                const lapack_complex_float* d,
+                                const lapack_complex_float* du,
+                                const lapack_complex_float* du2,
+                                const lapack_int* ipiv, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zgttrs_work( int matrix_order, char trans, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* dl,
+                                const lapack_complex_double* d,
+                                const lapack_complex_double* du,
+                                const lapack_complex_double* du2,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_chbev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int kd,
+                               lapack_complex_float* ab, lapack_int ldab,
+                               float* w, lapack_complex_float* z,
+                               lapack_int ldz, lapack_complex_float* work,
+                               float* rwork );
+lapack_int LAPACKE_zhbev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int kd,
+                               lapack_complex_double* ab, lapack_int ldab,
+                               double* w, lapack_complex_double* z,
+                               lapack_int ldz, lapack_complex_double* work,
+                               double* rwork );
+
+lapack_int LAPACKE_chbevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int kd,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zhbevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int kd,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_chbevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int kd,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                lapack_complex_float* q, lapack_int ldq,
+                                float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                float* rwork, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_zhbevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int kd,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                lapack_complex_double* q, lapack_int ldq,
+                                double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                double* rwork, lapack_int* iwork,
+                                lapack_int* ifail );
+
+lapack_int LAPACKE_chbgst_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                const lapack_complex_float* bb, lapack_int ldbb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zhbgst_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                const lapack_complex_double* bb,
+                                lapack_int ldbb, lapack_complex_double* x,
+                                lapack_int ldx, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_chbgv_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int ka, lapack_int kb,
+                               lapack_complex_float* ab, lapack_int ldab,
+                               lapack_complex_float* bb, lapack_int ldbb,
+                               float* w, lapack_complex_float* z,
+                               lapack_int ldz, lapack_complex_float* work,
+                               float* rwork );
+lapack_int LAPACKE_zhbgv_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int ka, lapack_int kb,
+                               lapack_complex_double* ab, lapack_int ldab,
+                               lapack_complex_double* bb, lapack_int ldbb,
+                               double* w, lapack_complex_double* z,
+                               lapack_int ldz, lapack_complex_double* work,
+                               double* rwork );
+
+lapack_int LAPACKE_chbgvd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                lapack_complex_float* bb, lapack_int ldbb,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zhbgvd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                lapack_complex_double* bb, lapack_int ldbb,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_chbgvx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int ka,
+                                lapack_int kb, lapack_complex_float* ab,
+                                lapack_int ldab, lapack_complex_float* bb,
+                                lapack_int ldbb, lapack_complex_float* q,
+                                lapack_int ldq, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, float* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_zhbgvx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int ka,
+                                lapack_int kb, lapack_complex_double* ab,
+                                lapack_int ldab, lapack_complex_double* bb,
+                                lapack_int ldbb, lapack_complex_double* q,
+                                lapack_int ldq, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, double* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_chbtrd_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int kd,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                float* d, float* e, lapack_complex_float* q,
+                                lapack_int ldq, lapack_complex_float* work );
+lapack_int LAPACKE_zhbtrd_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int kd,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                double* d, double* e, lapack_complex_double* q,
+                                lapack_int ldq, lapack_complex_double* work );
+
+lapack_int LAPACKE_checon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work );
+lapack_int LAPACKE_zhecon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, lapack_complex_double* work );
+
+lapack_int LAPACKE_cheequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 float* s, float* scond, float* amax,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_zheequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 double* s, double* scond, double* amax,
+                                 lapack_complex_double* work );
+
+lapack_int LAPACKE_cheev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_complex_float* a,
+                               lapack_int lda, float* w,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork );
+lapack_int LAPACKE_zheev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_complex_double* a,
+                               lapack_int lda, double* w,
+                               lapack_complex_double* work, lapack_int lwork,
+                               double* rwork );
+
+lapack_int LAPACKE_cheevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, float* w,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_zheevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, double* w,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_cheevr_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_int* isuppz,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_zheevr_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_int* isuppz,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_cheevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_zheevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_chegst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zhegst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda, const lapack_complex_double* b,
+                                lapack_int ldb );
+
+lapack_int LAPACKE_chegv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n, lapack_complex_float* a,
+                               lapack_int lda, lapack_complex_float* b,
+                               lapack_int ldb, float* w,
+                               lapack_complex_float* work, lapack_int lwork,
+                               float* rwork );
+lapack_int LAPACKE_zhegv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n,
+                               lapack_complex_double* a, lapack_int lda,
+                               lapack_complex_double* b, lapack_int ldb,
+                               double* w, lapack_complex_double* work,
+                               lapack_int lwork, double* rwork );
+
+lapack_int LAPACKE_chegvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                float* w, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zhegvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                double* w, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_chegvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_zhegvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_cherfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zherfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_complex_double* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_cherfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 const lapack_complex_float* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const float* s, const lapack_complex_float* b,
+                                 lapack_int ldb, lapack_complex_float* x,
+                                 lapack_int ldx, float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zherfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 const lapack_complex_double* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const double* s,
+                                 const lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_chesv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* a,
+                               lapack_int lda, lapack_int* ipiv,
+                               lapack_complex_float* b, lapack_int ldb,
+                               lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zhesv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* a,
+                               lapack_int lda, lapack_int* ipiv,
+                               lapack_complex_double* b, lapack_int ldb,
+                               lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_chesvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* af, lapack_int ldaf,
+                                lapack_int* ipiv, const lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork );
+lapack_int LAPACKE_zhesvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* af, lapack_int ldaf,
+                                lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork );
+
+lapack_int LAPACKE_chesvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, float* s,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* rpvgrw, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zhesvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, double* s,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_chetrd_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                float* d, float* e, lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zhetrd_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                double* d, double* e,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_chetrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* ipiv, lapack_complex_float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_zhetrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ipiv, lapack_complex_double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_chetri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zhetri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_chetrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zhetrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_chfrk_work( int matrix_order, char transr, char uplo,
+                               char trans, lapack_int n, lapack_int k,
+                               float alpha, const lapack_complex_float* a,
+                               lapack_int lda, float beta,
+                               lapack_complex_float* c );
+lapack_int LAPACKE_zhfrk_work( int matrix_order, char transr, char uplo,
+                               char trans, lapack_int n, lapack_int k,
+                               double alpha, const lapack_complex_double* a,
+                               lapack_int lda, double beta,
+                               lapack_complex_double* c );
+
+lapack_int LAPACKE_shgeqz_work( int matrix_order, char job, char compq,
+                                char compz, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, float* h, lapack_int ldh,
+                                float* t, lapack_int ldt, float* alphar,
+                                float* alphai, float* beta, float* q,
+                                lapack_int ldq, float* z, lapack_int ldz,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dhgeqz_work( int matrix_order, char job, char compq,
+                                char compz, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, double* h, lapack_int ldh,
+                                double* t, lapack_int ldt, double* alphar,
+                                double* alphai, double* beta, double* q,
+                                lapack_int ldq, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_chgeqz_work( int matrix_order, char job, char compq,
+                                char compz, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_float* h,
+                                lapack_int ldh, lapack_complex_float* t,
+                                lapack_int ldt, lapack_complex_float* alpha,
+                                lapack_complex_float* beta,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork );
+lapack_int LAPACKE_zhgeqz_work( int matrix_order, char job, char compq,
+                                char compz, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_double* h,
+                                lapack_int ldh, lapack_complex_double* t,
+                                lapack_int ldt, lapack_complex_double* alpha,
+                                lapack_complex_double* beta,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork );
+
+lapack_int LAPACKE_chpcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work );
+lapack_int LAPACKE_zhpcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, lapack_complex_double* work );
+
+lapack_int LAPACKE_chpev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_complex_float* ap, float* w,
+                               lapack_complex_float* z, lapack_int ldz,
+                               lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zhpev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_complex_double* ap,
+                               double* w, lapack_complex_double* z,
+                               lapack_int ldz, lapack_complex_double* work,
+                               double* rwork );
+
+lapack_int LAPACKE_chpevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_complex_float* ap,
+                                float* w, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zhpevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_complex_double* ap,
+                                double* w, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_chpevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_float* ap, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, float* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_zhpevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n,
+                                lapack_complex_double* ap, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, double* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_chpgst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, lapack_complex_float* ap,
+                                const lapack_complex_float* bp );
+lapack_int LAPACKE_zhpgst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, lapack_complex_double* ap,
+                                const lapack_complex_double* bp );
+
+lapack_int LAPACKE_chpgv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n,
+                               lapack_complex_float* ap,
+                               lapack_complex_float* bp, float* w,
+                               lapack_complex_float* z, lapack_int ldz,
+                               lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zhpgv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n,
+                               lapack_complex_double* ap,
+                               lapack_complex_double* bp, double* w,
+                               lapack_complex_double* z, lapack_int ldz,
+                               lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_chpgvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n,
+                                lapack_complex_float* ap,
+                                lapack_complex_float* bp, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_zhpgvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n,
+                                lapack_complex_double* ap,
+                                lapack_complex_double* bp, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int lrwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_chpgvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n,
+                                lapack_complex_float* ap,
+                                lapack_complex_float* bp, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, float* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_zhpgvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n,
+                                lapack_complex_double* ap,
+                                lapack_complex_double* bp, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, double* rwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_chprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                const lapack_complex_float* afp,
+                                const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zhprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* afp,
+                                const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_chpsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* ap,
+                               lapack_int* ipiv, lapack_complex_float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_zhpsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* ap,
+                               lapack_int* ipiv, lapack_complex_double* b,
+                               lapack_int ldb );
+
+lapack_int LAPACKE_chpsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* ap,
+                                lapack_complex_float* afp, lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zhpsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                lapack_complex_double* afp, lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_chptrd_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap, float* d, float* e,
+                                lapack_complex_float* tau );
+lapack_int LAPACKE_zhptrd_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap, double* d, double* e,
+                                lapack_complex_double* tau );
+
+lapack_int LAPACKE_chptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap, lapack_int* ipiv );
+lapack_int LAPACKE_zhptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap, lapack_int* ipiv );
+
+lapack_int LAPACKE_chptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zhptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_chptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                const lapack_int* ipiv, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zhptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_shsein_work( int matrix_order, char job, char eigsrc,
+                                char initv, lapack_logical* select,
+                                lapack_int n, const float* h, lapack_int ldh,
+                                float* wr, const float* wi, float* vl,
+                                lapack_int ldvl, float* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m, float* work,
+                                lapack_int* ifaill, lapack_int* ifailr );
+lapack_int LAPACKE_dhsein_work( int matrix_order, char job, char eigsrc,
+                                char initv, lapack_logical* select,
+                                lapack_int n, const double* h, lapack_int ldh,
+                                double* wr, const double* wi, double* vl,
+                                lapack_int ldvl, double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m, double* work,
+                                lapack_int* ifaill, lapack_int* ifailr );
+lapack_int LAPACKE_chsein_work( int matrix_order, char job, char eigsrc,
+                                char initv, const lapack_logical* select,
+                                lapack_int n, const lapack_complex_float* h,
+                                lapack_int ldh, lapack_complex_float* w,
+                                lapack_complex_float* vl, lapack_int ldvl,
+                                lapack_complex_float* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_float* work, float* rwork,
+                                lapack_int* ifaill, lapack_int* ifailr );
+lapack_int LAPACKE_zhsein_work( int matrix_order, char job, char eigsrc,
+                                char initv, const lapack_logical* select,
+                                lapack_int n, const lapack_complex_double* h,
+                                lapack_int ldh, lapack_complex_double* w,
+                                lapack_complex_double* vl, lapack_int ldvl,
+                                lapack_complex_double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_double* work, double* rwork,
+                                lapack_int* ifaill, lapack_int* ifailr );
+
+lapack_int LAPACKE_shseqr_work( int matrix_order, char job, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                float* h, lapack_int ldh, float* wr, float* wi,
+                                float* z, lapack_int ldz, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dhseqr_work( int matrix_order, char job, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                double* h, lapack_int ldh, double* wr,
+                                double* wi, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_chseqr_work( int matrix_order, char job, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                lapack_complex_float* h, lapack_int ldh,
+                                lapack_complex_float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zhseqr_work( int matrix_order, char job, char compz,
+                                lapack_int n, lapack_int ilo, lapack_int ihi,
+                                lapack_complex_double* h, lapack_int ldh,
+                                lapack_complex_double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_clacgv_work( lapack_int n, lapack_complex_float* x,
+                                lapack_int incx );
+lapack_int LAPACKE_zlacgv_work( lapack_int n, lapack_complex_double* x,
+                                lapack_int incx );
+
+lapack_int LAPACKE_slacpy_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, const float* a, lapack_int lda,
+                                float* b, lapack_int ldb );
+lapack_int LAPACKE_dlacpy_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, const double* a, lapack_int lda,
+                                double* b, lapack_int ldb );
+lapack_int LAPACKE_clacpy_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, const lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zlacpy_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, const lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb );
+
+lapack_int LAPACKE_zlag2c_work( int matrix_order, lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_float* sa, lapack_int ldsa );
+
+lapack_int LAPACKE_slag2d_work( int matrix_order, lapack_int m, lapack_int n,
+                                const float* sa, lapack_int ldsa, double* a,
+                                lapack_int lda );
+
+lapack_int LAPACKE_dlag2s_work( int matrix_order, lapack_int m, lapack_int n,
+                                const double* a, lapack_int lda, float* sa,
+                                lapack_int ldsa );
+
+lapack_int LAPACKE_clag2z_work( int matrix_order, lapack_int m, lapack_int n,
+                                const lapack_complex_float* sa, lapack_int ldsa,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_slagge_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const float* d,
+                                float* a, lapack_int lda, lapack_int* iseed,
+                                float* work );
+lapack_int LAPACKE_dlagge_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const double* d,
+                                double* a, lapack_int lda, lapack_int* iseed,
+                                double* work );
+lapack_int LAPACKE_clagge_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const float* d,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* iseed, lapack_complex_float* work );
+lapack_int LAPACKE_zlagge_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int kl, lapack_int ku, const double* d,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* iseed,
+                                lapack_complex_double* work );
+                                
+lapack_int LAPACKE_claghe_work( int matrix_order, lapack_int n, lapack_int k,
+                                const float* d, lapack_complex_float* a,
+                                lapack_int lda, lapack_int* iseed,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zlaghe_work( int matrix_order, lapack_int n, lapack_int k,
+                                const double* d, lapack_complex_double* a,
+                                lapack_int lda, lapack_int* iseed,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_slagsy_work( int matrix_order, lapack_int n, lapack_int k,
+                                const float* d, float* a, lapack_int lda,
+                                lapack_int* iseed, float* work );
+lapack_int LAPACKE_dlagsy_work( int matrix_order, lapack_int n, lapack_int k,
+                                const double* d, double* a, lapack_int lda,
+                                lapack_int* iseed, double* work );
+lapack_int LAPACKE_clagsy_work( int matrix_order, lapack_int n, lapack_int k,
+                                const float* d, lapack_complex_float* a,
+                                lapack_int lda, lapack_int* iseed,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zlagsy_work( int matrix_order, lapack_int n, lapack_int k,
+                                const double* d, lapack_complex_double* a,
+                                lapack_int lda, lapack_int* iseed,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_slapmr_work( int matrix_order, lapack_logical forwrd,
+                                lapack_int m, lapack_int n, float* x,
+                                lapack_int ldx, lapack_int* k );
+lapack_int LAPACKE_dlapmr_work( int matrix_order, lapack_logical forwrd,
+                                lapack_int m, lapack_int n, double* x,
+                                lapack_int ldx, lapack_int* k );
+lapack_int LAPACKE_clapmr_work( int matrix_order, lapack_logical forwrd,
+                                lapack_int m, lapack_int n,
+                                lapack_complex_float* x, lapack_int ldx,
+                                lapack_int* k );
+lapack_int LAPACKE_zlapmr_work( int matrix_order, lapack_logical forwrd,
+                                lapack_int m, lapack_int n,
+                                lapack_complex_double* x, lapack_int ldx,
+                                lapack_int* k );
+
+lapack_int LAPACKE_slartgp_work( float f, float g, float* cs, float* sn,
+                                 float* r );
+lapack_int LAPACKE_dlartgp_work( double f, double g, double* cs, double* sn,
+                                 double* r );
+
+lapack_int LAPACKE_slartgs_work( float x, float y, float sigma, float* cs,
+                                 float* sn );
+lapack_int LAPACKE_dlartgs_work( double x, double y, double sigma, double* cs,
+                                 double* sn );
+                                
+float LAPACKE_slapy2_work( float x, float y );
+double LAPACKE_dlapy2_work( double x, double y );
+
+float LAPACKE_slapy3_work( float x, float y, float z );
+double LAPACKE_dlapy3_work( double x, double y, double z );
+
+float LAPACKE_slamch_work( char cmach );
+double LAPACKE_dlamch_work( char cmach );
+
+float LAPACKE_slange_work( int matrix_order, char norm, lapack_int m,
+                                lapack_int n, const float* a, lapack_int lda,
+                                float* work );
+double LAPACKE_dlange_work( int matrix_order, char norm, lapack_int m,
+                                lapack_int n, const double* a, lapack_int lda,
+                                double* work );
+float LAPACKE_clange_work( int matrix_order, char norm, lapack_int m,
+                                lapack_int n, const lapack_complex_float* a,
+                                lapack_int lda, float* work );
+double LAPACKE_zlange_work( int matrix_order, char norm, lapack_int m,
+                                lapack_int n, const lapack_complex_double* a,
+                                lapack_int lda, double* work );
+
+float LAPACKE_clanhe_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const lapack_complex_float* a,
+                                lapack_int lda, float* work );
+double LAPACKE_zlanhe_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const lapack_complex_double* a,
+                                lapack_int lda, double* work );
+
+float LAPACKE_slansy_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const float* a, lapack_int lda,
+                                float* work );
+double LAPACKE_dlansy_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const double* a, lapack_int lda,
+                                double* work );
+float LAPACKE_clansy_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const lapack_complex_float* a,
+                                lapack_int lda, float* work );
+double LAPACKE_zlansy_work( int matrix_order, char norm, char uplo,
+                                lapack_int n, const lapack_complex_double* a,
+                                lapack_int lda, double* work );
+
+float LAPACKE_slantr_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int m, lapack_int n, const float* a,
+                                lapack_int lda, float* work );
+double LAPACKE_dlantr_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int m, lapack_int n,
+                                const double* a, lapack_int lda, double* work );
+float LAPACKE_clantr_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int m, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float* work );
+double LAPACKE_zlantr_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double* work );
+
+lapack_int LAPACKE_slarfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, const float* v,
+                                lapack_int ldv, const float* t, lapack_int ldt,
+                                float* c, lapack_int ldc, float* work,
+                                lapack_int ldwork );
+lapack_int LAPACKE_dlarfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, const double* v,
+                                lapack_int ldv, const double* t, lapack_int ldt,
+                                double* c, lapack_int ldc, double* work,
+                                lapack_int ldwork );
+lapack_int LAPACKE_clarfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k,
+                                const lapack_complex_float* v, lapack_int ldv,
+                                const lapack_complex_float* t, lapack_int ldt,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int ldwork );
+lapack_int LAPACKE_zlarfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k,
+                                const lapack_complex_double* v, lapack_int ldv,
+                                const lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work,
+                                lapack_int ldwork );
+
+lapack_int LAPACKE_slarfg_work( lapack_int n, float* alpha, float* x,
+                                lapack_int incx, float* tau );
+lapack_int LAPACKE_dlarfg_work( lapack_int n, double* alpha, double* x,
+                                lapack_int incx, double* tau );
+lapack_int LAPACKE_clarfg_work( lapack_int n, lapack_complex_float* alpha,
+                                lapack_complex_float* x, lapack_int incx,
+                                lapack_complex_float* tau );
+lapack_int LAPACKE_zlarfg_work( lapack_int n, lapack_complex_double* alpha,
+                                lapack_complex_double* x, lapack_int incx,
+                                lapack_complex_double* tau );
+
+lapack_int LAPACKE_slarft_work( int matrix_order, char direct, char storev,
+                                lapack_int n, lapack_int k, const float* v,
+                                lapack_int ldv, const float* tau, float* t,
+                                lapack_int ldt );
+lapack_int LAPACKE_dlarft_work( int matrix_order, char direct, char storev,
+                                lapack_int n, lapack_int k, const double* v,
+                                lapack_int ldv, const double* tau, double* t,
+                                lapack_int ldt );
+lapack_int LAPACKE_clarft_work( int matrix_order, char direct, char storev,
+                                lapack_int n, lapack_int k,
+                                const lapack_complex_float* v, lapack_int ldv,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zlarft_work( int matrix_order, char direct, char storev,
+                                lapack_int n, lapack_int k,
+                                const lapack_complex_double* v, lapack_int ldv,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_slarfx_work( int matrix_order, char side, lapack_int m,
+                                lapack_int n, const float* v, float tau,
+                                float* c, lapack_int ldc, float* work );
+lapack_int LAPACKE_dlarfx_work( int matrix_order, char side, lapack_int m,
+                                lapack_int n, const double* v, double tau,
+                                double* c, lapack_int ldc, double* work );
+lapack_int LAPACKE_clarfx_work( int matrix_order, char side, lapack_int m,
+                                lapack_int n, const lapack_complex_float* v,
+                                lapack_complex_float tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zlarfx_work( int matrix_order, char side, lapack_int m,
+                                lapack_int n, const lapack_complex_double* v,
+                                lapack_complex_double tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_slarnv_work( lapack_int idist, lapack_int* iseed,
+                                lapack_int n, float* x );
+lapack_int LAPACKE_dlarnv_work( lapack_int idist, lapack_int* iseed,
+                                lapack_int n, double* x );
+lapack_int LAPACKE_clarnv_work( lapack_int idist, lapack_int* iseed,
+                                lapack_int n, lapack_complex_float* x );
+lapack_int LAPACKE_zlarnv_work( lapack_int idist, lapack_int* iseed,
+                                lapack_int n, lapack_complex_double* x );
+
+lapack_int LAPACKE_slaset_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, float alpha, float beta, float* a,
+                                lapack_int lda );
+lapack_int LAPACKE_dlaset_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, double alpha, double beta,
+                                double* a, lapack_int lda );
+lapack_int LAPACKE_claset_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, lapack_complex_float alpha,
+                                lapack_complex_float beta,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zlaset_work( int matrix_order, char uplo, lapack_int m,
+                                lapack_int n, lapack_complex_double alpha,
+                                lapack_complex_double beta,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_slasrt_work( char id, lapack_int n, float* d );
+lapack_int LAPACKE_dlasrt_work( char id, lapack_int n, double* d );
+
+lapack_int LAPACKE_slaswp_work( int matrix_order, lapack_int n, float* a,
+                                lapack_int lda, lapack_int k1, lapack_int k2,
+                                const lapack_int* ipiv, lapack_int incx );
+lapack_int LAPACKE_dlaswp_work( int matrix_order, lapack_int n, double* a,
+                                lapack_int lda, lapack_int k1, lapack_int k2,
+                                const lapack_int* ipiv, lapack_int incx );
+lapack_int LAPACKE_claswp_work( int matrix_order, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int k1, lapack_int k2,
+                                const lapack_int* ipiv, lapack_int incx );
+lapack_int LAPACKE_zlaswp_work( int matrix_order, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int k1, lapack_int k2,
+                                const lapack_int* ipiv, lapack_int incx );
+
+lapack_int LAPACKE_slatms_work( int matrix_order, lapack_int m, lapack_int n,
+                                char dist, lapack_int* iseed, char sym,
+                                float* d, lapack_int mode, float cond,
+                                float dmax, lapack_int kl, lapack_int ku,
+                                char pack, float* a, lapack_int lda,
+                                float* work );
+lapack_int LAPACKE_dlatms_work( int matrix_order, lapack_int m, lapack_int n,
+                                char dist, lapack_int* iseed, char sym,
+                                double* d, lapack_int mode, double cond,
+                                double dmax, lapack_int kl, lapack_int ku,
+                                char pack, double* a, lapack_int lda,
+                                double* work );
+lapack_int LAPACKE_clatms_work( int matrix_order, lapack_int m, lapack_int n,
+                                char dist, lapack_int* iseed, char sym,
+                                float* d, lapack_int mode, float cond,
+                                float dmax, lapack_int kl, lapack_int ku,
+                                char pack, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* work );
+lapack_int LAPACKE_zlatms_work( int matrix_order, lapack_int m, lapack_int n,
+                                char dist, lapack_int* iseed, char sym,
+                                double* d, lapack_int mode, double cond,
+                                double dmax, lapack_int kl, lapack_int ku,
+                                char pack, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* work );
+
+lapack_int LAPACKE_slauum_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda );
+lapack_int LAPACKE_dlauum_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda );
+lapack_int LAPACKE_clauum_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zlauum_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_sopgtr_work( int matrix_order, char uplo, lapack_int n,
+                                const float* ap, const float* tau, float* q,
+                                lapack_int ldq, float* work );
+lapack_int LAPACKE_dopgtr_work( int matrix_order, char uplo, lapack_int n,
+                                const double* ap, const double* tau, double* q,
+                                lapack_int ldq, double* work );
+
+lapack_int LAPACKE_sopmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const float* ap, const float* tau, float* c,
+                                lapack_int ldc, float* work );
+lapack_int LAPACKE_dopmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const double* ap, const double* tau, double* c,
+                                lapack_int ldc, double* work );
+
+lapack_int LAPACKE_sorgbr_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int k, float* a,
+                                lapack_int lda, const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorgbr_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int k, double* a,
+                                lapack_int lda, const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorghr_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, float* a, lapack_int lda,
+                                const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorghr_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, double* a, lapack_int lda,
+                                const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorglq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, float* a, lapack_int lda,
+                                const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorglq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, double* a, lapack_int lda,
+                                const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorgql_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, float* a, lapack_int lda,
+                                const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorgql_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, double* a, lapack_int lda,
+                                const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorgqr_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, float* a, lapack_int lda,
+                                const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorgqr_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, double* a, lapack_int lda,
+                                const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorgrq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, float* a, lapack_int lda,
+                                const float* tau, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorgrq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, double* a, lapack_int lda,
+                                const double* tau, double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_sorgtr_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda, const float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dorgtr_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda, const double* tau,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormbr_work( int matrix_order, char vect, char side,
+                                char trans, lapack_int m, lapack_int n,
+                                lapack_int k, const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormbr_work( int matrix_order, char vect, char side,
+                                char trans, lapack_int m, lapack_int n,
+                                lapack_int k, const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormhr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormhr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormlq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormlq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormql_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormql_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormqr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormqr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormrq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormrq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormrz_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                lapack_int l, const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormrz_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                lapack_int l, const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_sormtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const float* a, lapack_int lda,
+                                const float* tau, float* c, lapack_int ldc,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dormtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const double* a, lapack_int lda,
+                                const double* tau, double* c, lapack_int ldc,
+                                double* work, lapack_int lwork );
+
+lapack_int LAPACKE_spbcon_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const float* ab, lapack_int ldab,
+                                float anorm, float* rcond, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dpbcon_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const double* ab,
+                                lapack_int ldab, double anorm, double* rcond,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cpbcon_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const lapack_complex_float* ab,
+                                lapack_int ldab, float anorm, float* rcond,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zpbcon_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const lapack_complex_double* ab,
+                                lapack_int ldab, double anorm, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_spbequ_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const float* ab, lapack_int ldab,
+                                float* s, float* scond, float* amax );
+lapack_int LAPACKE_dpbequ_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const double* ab,
+                                lapack_int ldab, double* s, double* scond,
+                                double* amax );
+lapack_int LAPACKE_cpbequ_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const lapack_complex_float* ab,
+                                lapack_int ldab, float* s, float* scond,
+                                float* amax );
+lapack_int LAPACKE_zpbequ_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, const lapack_complex_double* ab,
+                                lapack_int ldab, double* s, double* scond,
+                                double* amax );
+
+lapack_int LAPACKE_spbrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs, const float* ab,
+                                lapack_int ldab, const float* afb,
+                                lapack_int ldafb, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dpbrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const double* ab, lapack_int ldab,
+                                const double* afb, lapack_int ldafb,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cpbrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                const lapack_complex_float* afb,
+                                lapack_int ldafb, const lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zpbrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab,
+                                const lapack_complex_double* afb,
+                                lapack_int ldafb,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_spbstf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kb, float* bb, lapack_int ldbb );
+lapack_int LAPACKE_dpbstf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kb, double* bb, lapack_int ldbb );
+lapack_int LAPACKE_cpbstf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kb, lapack_complex_float* bb,
+                                lapack_int ldbb );
+lapack_int LAPACKE_zpbstf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kb, lapack_complex_double* bb,
+                                lapack_int ldbb );
+
+lapack_int LAPACKE_spbsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int kd, lapack_int nrhs, float* ab,
+                               lapack_int ldab, float* b, lapack_int ldb );
+lapack_int LAPACKE_dpbsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int kd, lapack_int nrhs, double* ab,
+                               lapack_int ldab, double* b, lapack_int ldb );
+lapack_int LAPACKE_cpbsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int kd, lapack_int nrhs,
+                               lapack_complex_float* ab, lapack_int ldab,
+                               lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpbsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int kd, lapack_int nrhs,
+                               lapack_complex_double* ab, lapack_int ldab,
+                               lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spbsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int kd, lapack_int nrhs,
+                                float* ab, lapack_int ldab, float* afb,
+                                lapack_int ldafb, char* equed, float* s,
+                                float* b, lapack_int ldb, float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dpbsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int kd, lapack_int nrhs,
+                                double* ab, lapack_int ldab, double* afb,
+                                lapack_int ldafb, char* equed, double* s,
+                                double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, double* work, lapack_int* iwork );
+lapack_int LAPACKE_cpbsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int kd, lapack_int nrhs,
+                                lapack_complex_float* ab, lapack_int ldab,
+                                lapack_complex_float* afb, lapack_int ldafb,
+                                char* equed, float* s, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zpbsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int kd, lapack_int nrhs,
+                                lapack_complex_double* ab, lapack_int ldab,
+                                lapack_complex_double* afb, lapack_int ldafb,
+                                char* equed, double* s,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_spbtrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, float* ab, lapack_int ldab );
+lapack_int LAPACKE_dpbtrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, double* ab, lapack_int ldab );
+lapack_int LAPACKE_cpbtrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_complex_float* ab,
+                                lapack_int ldab );
+lapack_int LAPACKE_zpbtrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_complex_double* ab,
+                                lapack_int ldab );
+
+lapack_int LAPACKE_spbtrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs, const float* ab,
+                                lapack_int ldab, float* b, lapack_int ldb );
+lapack_int LAPACKE_dpbtrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const double* ab, lapack_int ldab, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_cpbtrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpbtrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int kd, lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, lapack_complex_double* b,
+                                lapack_int ldb );
+
+lapack_int LAPACKE_spftrf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, float* a );
+lapack_int LAPACKE_dpftrf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, double* a );
+lapack_int LAPACKE_cpftrf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_complex_float* a );
+lapack_int LAPACKE_zpftrf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_complex_double* a );
+
+lapack_int LAPACKE_spftri_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, float* a );
+lapack_int LAPACKE_dpftri_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, double* a );
+lapack_int LAPACKE_cpftri_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_complex_float* a );
+lapack_int LAPACKE_zpftri_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_complex_double* a );
+
+lapack_int LAPACKE_spftrs_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_int nrhs, const float* a,
+                                float* b, lapack_int ldb );
+lapack_int LAPACKE_dpftrs_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_int nrhs, const double* a,
+                                double* b, lapack_int ldb );
+lapack_int LAPACKE_cpftrs_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* a,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpftrs_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* a,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spocon_work( int matrix_order, char uplo, lapack_int n,
+                                const float* a, lapack_int lda, float anorm,
+                                float* rcond, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dpocon_work( int matrix_order, char uplo, lapack_int n,
+                                const double* a, lapack_int lda, double anorm,
+                                double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cpocon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float anorm, float* rcond,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zpocon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double anorm, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_spoequ_work( int matrix_order, lapack_int n, const float* a,
+                                lapack_int lda, float* s, float* scond,
+                                float* amax );
+lapack_int LAPACKE_dpoequ_work( int matrix_order, lapack_int n, const double* a,
+                                lapack_int lda, double* s, double* scond,
+                                double* amax );
+lapack_int LAPACKE_cpoequ_work( int matrix_order, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float* s, float* scond, float* amax );
+lapack_int LAPACKE_zpoequ_work( int matrix_order, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_spoequb_work( int matrix_order, lapack_int n, const float* a,
+                                 lapack_int lda, float* s, float* scond,
+                                 float* amax );
+lapack_int LAPACKE_dpoequb_work( int matrix_order, lapack_int n,
+                                 const double* a, lapack_int lda, double* s,
+                                 double* scond, double* amax );
+lapack_int LAPACKE_cpoequb_work( int matrix_order, lapack_int n,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 float* s, float* scond, float* amax );
+lapack_int LAPACKE_zpoequb_work( int matrix_order, lapack_int n,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 double* s, double* scond, double* amax );
+
+lapack_int LAPACKE_sporfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                const float* af, lapack_int ldaf,
+                                const float* b, lapack_int ldb, float* x,
+                                lapack_int ldx, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dporfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, const double* af,
+                                lapack_int ldaf, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cporfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* af,
+                                lapack_int ldaf, const lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zporfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_complex_double* af,
+                                lapack_int ldaf, const lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sporfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs, const float* a,
+                                 lapack_int lda, const float* af,
+                                 lapack_int ldaf, const float* s,
+                                 const float* b, lapack_int ldb, float* x,
+                                 lapack_int ldx, float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dporfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs, const double* a,
+                                 lapack_int lda, const double* af,
+                                 lapack_int ldaf, const double* s,
+                                 const double* b, lapack_int ldb, double* x,
+                                 lapack_int ldx, double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cporfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 const lapack_complex_float* af,
+                                 lapack_int ldaf, const float* s,
+                                 const lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zporfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 const lapack_complex_double* af,
+                                 lapack_int ldaf, const double* s,
+                                 const lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_sposv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, float* a, lapack_int lda,
+                               float* b, lapack_int ldb );
+lapack_int LAPACKE_dposv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, double* a, lapack_int lda,
+                               double* b, lapack_int ldb );
+lapack_int LAPACKE_cposv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* a,
+                               lapack_int lda, lapack_complex_float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_zposv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* a,
+                               lapack_int lda, lapack_complex_double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_dsposv_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, double* a, lapack_int lda,
+                                double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* work, float* swork,
+                                lapack_int* iter );
+lapack_int LAPACKE_zcposv_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, lapack_complex_double* work,
+                                lapack_complex_float* swork, double* rwork,
+                                lapack_int* iter );
+
+lapack_int LAPACKE_sposvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, float* a,
+                                lapack_int lda, float* af, lapack_int ldaf,
+                                char* equed, float* s, float* b, lapack_int ldb,
+                                float* x, lapack_int ldx, float* rcond,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dposvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, double* a,
+                                lapack_int lda, double* af, lapack_int ldaf,
+                                char* equed, double* s, double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cposvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* af, lapack_int ldaf,
+                                char* equed, float* s, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zposvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* af, lapack_int ldaf,
+                                char* equed, double* s,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sposvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs, float* a,
+                                 lapack_int lda, float* af, lapack_int ldaf,
+                                 char* equed, float* s, float* b,
+                                 lapack_int ldb, float* x, lapack_int ldx,
+                                 float* rcond, float* rpvgrw, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dposvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs, double* a,
+                                 lapack_int lda, double* af, lapack_int ldaf,
+                                 char* equed, double* s, double* b,
+                                 lapack_int ldb, double* x, lapack_int ldx,
+                                 double* rcond, double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_cposvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* af, lapack_int ldaf,
+                                 char* equed, float* s, lapack_complex_float* b,
+                                 lapack_int ldb, lapack_complex_float* x,
+                                 lapack_int ldx, float* rcond, float* rpvgrw,
+                                 float* berr, lapack_int n_err_bnds,
+                                 float* err_bnds_norm, float* err_bnds_comp,
+                                 lapack_int nparams, float* params,
+                                 lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zposvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* af, lapack_int ldaf,
+                                 char* equed, double* s,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_spotrf_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda );
+lapack_int LAPACKE_dpotrf_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda );
+lapack_int LAPACKE_cpotrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zpotrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_spotri_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda );
+lapack_int LAPACKE_dpotri_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda );
+lapack_int LAPACKE_cpotri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zpotri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_spotrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                float* b, lapack_int ldb );
+lapack_int LAPACKE_dpotrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, double* b, lapack_int ldb );
+lapack_int LAPACKE_cpotrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zpotrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* b,
+                                lapack_int ldb );
+
+lapack_int LAPACKE_sppcon_work( int matrix_order, char uplo, lapack_int n,
+                                const float* ap, float anorm, float* rcond,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dppcon_work( int matrix_order, char uplo, lapack_int n,
+                                const double* ap, double anorm, double* rcond,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cppcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap, float anorm,
+                                float* rcond, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zppcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap, double anorm,
+                                double* rcond, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_sppequ_work( int matrix_order, char uplo, lapack_int n,
+                                const float* ap, float* s, float* scond,
+                                float* amax );
+lapack_int LAPACKE_dppequ_work( int matrix_order, char uplo, lapack_int n,
+                                const double* ap, double* s, double* scond,
+                                double* amax );
+lapack_int LAPACKE_cppequ_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap, float* s,
+                                float* scond, float* amax );
+lapack_int LAPACKE_zppequ_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap, double* s,
+                                double* scond, double* amax );
+
+lapack_int LAPACKE_spprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* ap,
+                                const float* afp, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dpprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* ap,
+                                const double* afp, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cpprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                const lapack_complex_float* afp,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zpprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* afp,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sppsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, float* ap, float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_dppsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, double* ap, double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_cppsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* ap,
+                               lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zppsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* ap,
+                               lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sppsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, float* ap,
+                                float* afp, char* equed, float* s, float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dppsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, double* ap,
+                                double* afp, char* equed, double* s, double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cppsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_float* ap,
+                                lapack_complex_float* afp, char* equed,
+                                float* s, lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_zppsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                lapack_complex_double* ap,
+                                lapack_complex_double* afp, char* equed,
+                                double* s, lapack_complex_double* b,
+                                lapack_int ldb, lapack_complex_double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_spptrf_work( int matrix_order, char uplo, lapack_int n,
+                                float* ap );
+lapack_int LAPACKE_dpptrf_work( int matrix_order, char uplo, lapack_int n,
+                                double* ap );
+lapack_int LAPACKE_cpptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap );
+lapack_int LAPACKE_zpptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap );
+
+lapack_int LAPACKE_spptri_work( int matrix_order, char uplo, lapack_int n,
+                                float* ap );
+lapack_int LAPACKE_dpptri_work( int matrix_order, char uplo, lapack_int n,
+                                double* ap );
+lapack_int LAPACKE_cpptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap );
+lapack_int LAPACKE_zpptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap );
+
+lapack_int LAPACKE_spptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* ap, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dpptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* ap, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_cpptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_spstrf_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* piv,
+                                lapack_int* rank, float tol, float* work );
+lapack_int LAPACKE_dpstrf_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* piv,
+                                lapack_int* rank, double tol, double* work );
+lapack_int LAPACKE_cpstrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* piv, lapack_int* rank, float tol,
+                                float* work );
+lapack_int LAPACKE_zpstrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* piv, lapack_int* rank, double tol,
+                                double* work );
+
+lapack_int LAPACKE_sptcon_work( lapack_int n, const float* d, const float* e,
+                                float anorm, float* rcond, float* work );
+lapack_int LAPACKE_dptcon_work( lapack_int n, const double* d, const double* e,
+                                double anorm, double* rcond, double* work );
+lapack_int LAPACKE_cptcon_work( lapack_int n, const float* d,
+                                const lapack_complex_float* e, float anorm,
+                                float* rcond, float* work );
+lapack_int LAPACKE_zptcon_work( lapack_int n, const double* d,
+                                const lapack_complex_double* e, double anorm,
+                                double* rcond, double* work );
+
+lapack_int LAPACKE_spteqr_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, float* z, lapack_int ldz,
+                                float* work );
+lapack_int LAPACKE_dpteqr_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, double* z, lapack_int ldz,
+                                double* work );
+lapack_int LAPACKE_cpteqr_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, lapack_complex_float* z,
+                                lapack_int ldz, float* work );
+lapack_int LAPACKE_zpteqr_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, lapack_complex_double* z,
+                                lapack_int ldz, double* work );
+
+lapack_int LAPACKE_sptrfs_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                const float* d, const float* e, const float* df,
+                                const float* ef, const float* b, lapack_int ldb,
+                                float* x, lapack_int ldx, float* ferr,
+                                float* berr, float* work );
+lapack_int LAPACKE_dptrfs_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                const double* d, const double* e,
+                                const double* df, const double* ef,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                double* work );
+lapack_int LAPACKE_cptrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* d,
+                                const lapack_complex_float* e, const float* df,
+                                const lapack_complex_float* ef,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zptrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* d,
+                                const lapack_complex_double* e,
+                                const double* df,
+                                const lapack_complex_double* ef,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sptsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               float* d, float* e, float* b, lapack_int ldb );
+lapack_int LAPACKE_dptsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               double* d, double* e, double* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_cptsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               float* d, lapack_complex_float* e,
+                               lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zptsv_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                               double* d, lapack_complex_double* e,
+                               lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sptsvx_work( int matrix_order, char fact, lapack_int n,
+                                lapack_int nrhs, const float* d, const float* e,
+                                float* df, float* ef, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work );
+lapack_int LAPACKE_dptsvx_work( int matrix_order, char fact, lapack_int n,
+                                lapack_int nrhs, const double* d,
+                                const double* e, double* df, double* ef,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* rcond, double* ferr,
+                                double* berr, double* work );
+lapack_int LAPACKE_cptsvx_work( int matrix_order, char fact, lapack_int n,
+                                lapack_int nrhs, const float* d,
+                                const lapack_complex_float* e, float* df,
+                                lapack_complex_float* ef,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zptsvx_work( int matrix_order, char fact, lapack_int n,
+                                lapack_int nrhs, const double* d,
+                                const lapack_complex_double* e, double* df,
+                                lapack_complex_double* ef,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_spttrf_work( lapack_int n, float* d, float* e );
+lapack_int LAPACKE_dpttrf_work( lapack_int n, double* d, double* e );
+lapack_int LAPACKE_cpttrf_work( lapack_int n, float* d,
+                                lapack_complex_float* e );
+lapack_int LAPACKE_zpttrf_work( lapack_int n, double* d,
+                                lapack_complex_double* e );
+
+lapack_int LAPACKE_spttrs_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                const float* d, const float* e, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dpttrs_work( int matrix_order, lapack_int n, lapack_int nrhs,
+                                const double* d, const double* e, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_cpttrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* d,
+                                const lapack_complex_float* e,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zpttrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* d,
+                                const lapack_complex_double* e,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_ssbev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int kd, float* ab,
+                               lapack_int ldab, float* w, float* z,
+                               lapack_int ldz, float* work );
+lapack_int LAPACKE_dsbev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int kd, double* ab,
+                               lapack_int ldab, double* w, double* z,
+                               lapack_int ldz, double* work );
+
+lapack_int LAPACKE_ssbevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int kd, float* ab,
+                                lapack_int ldab, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dsbevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int kd, double* ab,
+                                lapack_int ldab, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_ssbevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int kd,
+                                float* ab, lapack_int ldab, float* q,
+                                lapack_int ldq, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_dsbevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int kd,
+                                double* ab, lapack_int ldab, double* q,
+                                lapack_int ldq, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int* iwork,
+                                lapack_int* ifail );
+
+lapack_int LAPACKE_ssbgst_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                float* ab, lapack_int ldab, const float* bb,
+                                lapack_int ldbb, float* x, lapack_int ldx,
+                                float* work );
+lapack_int LAPACKE_dsbgst_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                double* ab, lapack_int ldab, const double* bb,
+                                lapack_int ldbb, double* x, lapack_int ldx,
+                                double* work );
+
+lapack_int LAPACKE_ssbgv_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int ka, lapack_int kb,
+                               float* ab, lapack_int ldab, float* bb,
+                               lapack_int ldbb, float* w, float* z,
+                               lapack_int ldz, float* work );
+lapack_int LAPACKE_dsbgv_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, lapack_int ka, lapack_int kb,
+                               double* ab, lapack_int ldab, double* bb,
+                               lapack_int ldbb, double* w, double* z,
+                               lapack_int ldz, double* work );
+
+lapack_int LAPACKE_ssbgvd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                float* ab, lapack_int ldab, float* bb,
+                                lapack_int ldbb, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dsbgvd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, lapack_int ka, lapack_int kb,
+                                double* ab, lapack_int ldab, double* bb,
+                                lapack_int ldbb, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_ssbgvx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int ka,
+                                lapack_int kb, float* ab, lapack_int ldab,
+                                float* bb, lapack_int ldbb, float* q,
+                                lapack_int ldq, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_dsbgvx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, lapack_int ka,
+                                lapack_int kb, double* ab, lapack_int ldab,
+                                double* bb, lapack_int ldbb, double* q,
+                                lapack_int ldq, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int* iwork,
+                                lapack_int* ifail );
+
+lapack_int LAPACKE_ssbtrd_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int kd, float* ab,
+                                lapack_int ldab, float* d, float* e, float* q,
+                                lapack_int ldq, float* work );
+lapack_int LAPACKE_dsbtrd_work( int matrix_order, char vect, char uplo,
+                                lapack_int n, lapack_int kd, double* ab,
+                                lapack_int ldab, double* d, double* e,
+                                double* q, lapack_int ldq, double* work );
+
+lapack_int LAPACKE_ssfrk_work( int matrix_order, char transr, char uplo,
+                               char trans, lapack_int n, lapack_int k,
+                               float alpha, const float* a, lapack_int lda,
+                               float beta, float* c );
+lapack_int LAPACKE_dsfrk_work( int matrix_order, char transr, char uplo,
+                               char trans, lapack_int n, lapack_int k,
+                               double alpha, const double* a, lapack_int lda,
+                               double beta, double* c );
+
+lapack_int LAPACKE_sspcon_work( int matrix_order, char uplo, lapack_int n,
+                                const float* ap, const lapack_int* ipiv,
+                                float anorm, float* rcond, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dspcon_work( int matrix_order, char uplo, lapack_int n,
+                                const double* ap, const lapack_int* ipiv,
+                                double anorm, double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_cspcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work );
+lapack_int LAPACKE_zspcon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, lapack_complex_double* work );
+
+lapack_int LAPACKE_sspev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, float* ap, float* w, float* z,
+                               lapack_int ldz, float* work );
+lapack_int LAPACKE_dspev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, double* ap, double* w, double* z,
+                               lapack_int ldz, double* work );
+
+lapack_int LAPACKE_sspevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, float* ap, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dspevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, double* ap, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_sspevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, float* ap, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, lapack_int* m, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_dspevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, double* ap, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, lapack_int* m, double* w,
+                                double* z, lapack_int ldz, double* work,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_sspgst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, float* ap, const float* bp );
+lapack_int LAPACKE_dspgst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, double* ap, const double* bp );
+
+lapack_int LAPACKE_sspgv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n, float* ap, float* bp,
+                               float* w, float* z, lapack_int ldz,
+                               float* work );
+lapack_int LAPACKE_dspgv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n, double* ap, double* bp,
+                               double* w, double* z, lapack_int ldz,
+                               double* work );
+
+lapack_int LAPACKE_sspgvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n, float* ap, float* bp,
+                                float* w, float* z, lapack_int ldz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_dspgvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n, double* ap, double* bp,
+                                double* w, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_sspgvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n, float* ap,
+                                float* bp, float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, float* z, lapack_int ldz, float* work,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_dspgvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n, double* ap,
+                                double* bp, double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, double* z, lapack_int ldz,
+                                double* work, lapack_int* iwork,
+                                lapack_int* ifail );
+
+lapack_int LAPACKE_ssprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* ap,
+                                const float* afp, const lapack_int* ipiv,
+                                const float* b, lapack_int ldb, float* x,
+                                lapack_int ldx, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dsprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* ap,
+                                const double* afp, const lapack_int* ipiv,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_csprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                const lapack_complex_float* afp,
+                                const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zsprfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* afp,
+                                const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_sspsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, float* ap, lapack_int* ipiv,
+                               float* b, lapack_int ldb );
+lapack_int LAPACKE_dspsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, double* ap, lapack_int* ipiv,
+                               double* b, lapack_int ldb );
+lapack_int LAPACKE_cspsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* ap,
+                               lapack_int* ipiv, lapack_complex_float* b,
+                               lapack_int ldb );
+lapack_int LAPACKE_zspsv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* ap,
+                               lapack_int* ipiv, lapack_complex_double* b,
+                               lapack_int ldb );
+
+lapack_int LAPACKE_sspsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, const float* ap,
+                                float* afp, lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dspsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, const double* ap,
+                                double* afp, lapack_int* ipiv, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_cspsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* ap,
+                                lapack_complex_float* afp, lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zspsvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                lapack_complex_double* afp, lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_ssptrd_work( int matrix_order, char uplo, lapack_int n,
+                                float* ap, float* d, float* e, float* tau );
+lapack_int LAPACKE_dsptrd_work( int matrix_order, char uplo, lapack_int n,
+                                double* ap, double* d, double* e, double* tau );
+
+lapack_int LAPACKE_ssptrf_work( int matrix_order, char uplo, lapack_int n,
+                                float* ap, lapack_int* ipiv );
+lapack_int LAPACKE_dsptrf_work( int matrix_order, char uplo, lapack_int n,
+                                double* ap, lapack_int* ipiv );
+lapack_int LAPACKE_csptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap, lapack_int* ipiv );
+lapack_int LAPACKE_zsptrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap, lapack_int* ipiv );
+
+lapack_int LAPACKE_ssptri_work( int matrix_order, char uplo, lapack_int n,
+                                float* ap, const lapack_int* ipiv,
+                                float* work );
+lapack_int LAPACKE_dsptri_work( int matrix_order, char uplo, lapack_int n,
+                                double* ap, const lapack_int* ipiv,
+                                double* work );
+lapack_int LAPACKE_csptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zsptri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_ssptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* ap,
+                                const lapack_int* ipiv, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dsptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* ap,
+                                const lapack_int* ipiv, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_csptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* ap,
+                                const lapack_int* ipiv, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_zsptrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_sstebz_work( char range, char order, lapack_int n, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, const float* d, const float* e,
+                                lapack_int* m, lapack_int* nsplit, float* w,
+                                lapack_int* iblock, lapack_int* isplit,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dstebz_work( char range, char order, lapack_int n, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, const double* d, const double* e,
+                                lapack_int* m, lapack_int* nsplit, double* w,
+                                lapack_int* iblock, lapack_int* isplit,
+                                double* work, lapack_int* iwork );
+
+lapack_int LAPACKE_sstedc_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, float* z, lapack_int ldz,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dstedc_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_cstedc_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, lapack_complex_float* z,
+                                lapack_int ldz, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zstedc_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, lapack_complex_double* z,
+                                lapack_int ldz, lapack_complex_double* work,
+                                lapack_int lwork, double* rwork,
+                                lapack_int lrwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_sstegr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, lapack_int* m, float* w, float* z,
+                                lapack_int ldz, lapack_int* isuppz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_dstegr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, lapack_int* m, double* w,
+                                double* z, lapack_int ldz, lapack_int* isuppz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_cstegr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, lapack_int* m, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_int* isuppz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zstegr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, lapack_int* m, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_int* isuppz, double* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_sstein_work( int matrix_order, lapack_int n, const float* d,
+                                const float* e, lapack_int m, const float* w,
+                                const lapack_int* iblock,
+                                const lapack_int* isplit, float* z,
+                                lapack_int ldz, float* work, lapack_int* iwork,
+                                lapack_int* ifailv );
+lapack_int LAPACKE_dstein_work( int matrix_order, lapack_int n, const double* d,
+                                const double* e, lapack_int m, const double* w,
+                                const lapack_int* iblock,
+                                const lapack_int* isplit, double* z,
+                                lapack_int ldz, double* work, lapack_int* iwork,
+                                lapack_int* ifailv );
+lapack_int LAPACKE_cstein_work( int matrix_order, lapack_int n, const float* d,
+                                const float* e, lapack_int m, const float* w,
+                                const lapack_int* iblock,
+                                const lapack_int* isplit,
+                                lapack_complex_float* z, lapack_int ldz,
+                                float* work, lapack_int* iwork,
+                                lapack_int* ifailv );
+lapack_int LAPACKE_zstein_work( int matrix_order, lapack_int n, const double* d,
+                                const double* e, lapack_int m, const double* w,
+                                const lapack_int* iblock,
+                                const lapack_int* isplit,
+                                lapack_complex_double* z, lapack_int ldz,
+                                double* work, lapack_int* iwork,
+                                lapack_int* ifailv );
+
+lapack_int LAPACKE_sstemr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                lapack_int* m, float* w, float* z,
+                                lapack_int ldz, lapack_int nzc,
+                                lapack_int* isuppz, lapack_logical* tryrac,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dstemr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                lapack_int* m, double* w, double* z,
+                                lapack_int ldz, lapack_int nzc,
+                                lapack_int* isuppz, lapack_logical* tryrac,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_cstemr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                lapack_int* m, float* w,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_int nzc, lapack_int* isuppz,
+                                lapack_logical* tryrac, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_zstemr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                lapack_int* m, double* w,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_int nzc, lapack_int* isuppz,
+                                lapack_logical* tryrac, double* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_ssteqr_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, float* z, lapack_int ldz,
+                                float* work );
+lapack_int LAPACKE_dsteqr_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, double* z, lapack_int ldz,
+                                double* work );
+lapack_int LAPACKE_csteqr_work( int matrix_order, char compz, lapack_int n,
+                                float* d, float* e, lapack_complex_float* z,
+                                lapack_int ldz, float* work );
+lapack_int LAPACKE_zsteqr_work( int matrix_order, char compz, lapack_int n,
+                                double* d, double* e, lapack_complex_double* z,
+                                lapack_int ldz, double* work );
+
+lapack_int LAPACKE_ssterf_work( lapack_int n, float* d, float* e );
+lapack_int LAPACKE_dsterf_work( lapack_int n, double* d, double* e );
+
+lapack_int LAPACKE_sstev_work( int matrix_order, char jobz, lapack_int n,
+                               float* d, float* e, float* z, lapack_int ldz,
+                               float* work );
+lapack_int LAPACKE_dstev_work( int matrix_order, char jobz, lapack_int n,
+                               double* d, double* e, double* z, lapack_int ldz,
+                               double* work );
+
+lapack_int LAPACKE_sstevd_work( int matrix_order, char jobz, lapack_int n,
+                                float* d, float* e, float* z, lapack_int ldz,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dstevd_work( int matrix_order, char jobz, lapack_int n,
+                                double* d, double* e, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_sstevr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, lapack_int* m, float* w, float* z,
+                                lapack_int ldz, lapack_int* isuppz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_dstevr_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, lapack_int* m, double* w,
+                                double* z, lapack_int ldz, lapack_int* isuppz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_sstevx_work( int matrix_order, char jobz, char range,
+                                lapack_int n, float* d, float* e, float vl,
+                                float vu, lapack_int il, lapack_int iu,
+                                float abstol, lapack_int* m, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_dstevx_work( int matrix_order, char jobz, char range,
+                                lapack_int n, double* d, double* e, double vl,
+                                double vu, lapack_int il, lapack_int iu,
+                                double abstol, lapack_int* m, double* w,
+                                double* z, lapack_int ldz, double* work,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_ssycon_work( int matrix_order, char uplo, lapack_int n,
+                                const float* a, lapack_int lda,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dsycon_work( int matrix_order, char uplo, lapack_int n,
+                                const double* a, lapack_int lda,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_csycon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_int* ipiv, float anorm,
+                                float* rcond, lapack_complex_float* work );
+lapack_int LAPACKE_zsycon_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_int* ipiv, double anorm,
+                                double* rcond, lapack_complex_double* work );
+
+lapack_int LAPACKE_ssyequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const float* a, lapack_int lda, float* s,
+                                 float* scond, float* amax, float* work );
+lapack_int LAPACKE_dsyequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const double* a, lapack_int lda, double* s,
+                                 double* scond, double* amax, double* work );
+lapack_int LAPACKE_csyequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 float* s, float* scond, float* amax,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_zsyequb_work( int matrix_order, char uplo, lapack_int n,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 double* s, double* scond, double* amax,
+                                 lapack_complex_double* work );
+
+lapack_int LAPACKE_ssyev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, float* a, lapack_int lda, float* w,
+                               float* work, lapack_int lwork );
+lapack_int LAPACKE_dsyev_work( int matrix_order, char jobz, char uplo,
+                               lapack_int n, double* a, lapack_int lda,
+                               double* w, double* work, lapack_int lwork );
+
+lapack_int LAPACKE_ssyevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, float* a, lapack_int lda,
+                                float* w, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dsyevd_work( int matrix_order, char jobz, char uplo,
+                                lapack_int n, double* a, lapack_int lda,
+                                double* w, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_ssyevr_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, float* a,
+                                lapack_int lda, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w, float* z,
+                                lapack_int ldz, lapack_int* isuppz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_dsyevr_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, double* a,
+                                lapack_int lda, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w, double* z,
+                                lapack_int ldz, lapack_int* isuppz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_ssyevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, float* a,
+                                lapack_int lda, float vl, float vu,
+                                lapack_int il, lapack_int iu, float abstol,
+                                lapack_int* m, float* w, float* z,
+                                lapack_int ldz, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int* ifail );
+lapack_int LAPACKE_dsyevx_work( int matrix_order, char jobz, char range,
+                                char uplo, lapack_int n, double* a,
+                                lapack_int lda, double vl, double vu,
+                                lapack_int il, lapack_int iu, double abstol,
+                                lapack_int* m, double* w, double* z,
+                                lapack_int ldz, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_ssygst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, float* a, lapack_int lda,
+                                const float* b, lapack_int ldb );
+lapack_int LAPACKE_dsygst_work( int matrix_order, lapack_int itype, char uplo,
+                                lapack_int n, double* a, lapack_int lda,
+                                const double* b, lapack_int ldb );
+
+lapack_int LAPACKE_ssygv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n, float* a,
+                               lapack_int lda, float* b, lapack_int ldb,
+                               float* w, float* work, lapack_int lwork );
+lapack_int LAPACKE_dsygv_work( int matrix_order, lapack_int itype, char jobz,
+                               char uplo, lapack_int n, double* a,
+                               lapack_int lda, double* b, lapack_int ldb,
+                               double* w, double* work, lapack_int lwork );
+
+lapack_int LAPACKE_ssygvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n, float* a,
+                                lapack_int lda, float* b, lapack_int ldb,
+                                float* w, float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dsygvd_work( int matrix_order, lapack_int itype, char jobz,
+                                char uplo, lapack_int n, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                double* w, double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+
+lapack_int LAPACKE_ssygvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n, float* a,
+                                lapack_int lda, float* b, lapack_int ldb,
+                                float vl, float vu, lapack_int il,
+                                lapack_int iu, float abstol, lapack_int* m,
+                                float* w, float* z, lapack_int ldz, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int* ifail );
+lapack_int LAPACKE_dsygvx_work( int matrix_order, lapack_int itype, char jobz,
+                                char range, char uplo, lapack_int n, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                double vl, double vu, lapack_int il,
+                                lapack_int iu, double abstol, lapack_int* m,
+                                double* w, double* z, lapack_int ldz,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int* ifail );
+
+lapack_int LAPACKE_ssyrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                const float* af, lapack_int ldaf,
+                                const lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dsyrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, const double* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const double* b, lapack_int ldb, double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                double* work, lapack_int* iwork );
+lapack_int LAPACKE_csyrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_zsyrfs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_complex_double* af,
+                                lapack_int ldaf, const lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_ssyrfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs, const float* a,
+                                 lapack_int lda, const float* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const float* s, const float* b, lapack_int ldb,
+                                 float* x, lapack_int ldx, float* rcond,
+                                 float* berr, lapack_int n_err_bnds,
+                                 float* err_bnds_norm, float* err_bnds_comp,
+                                 lapack_int nparams, float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dsyrfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs, const double* a,
+                                 lapack_int lda, const double* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const double* s, const double* b,
+                                 lapack_int ldb, double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, double* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_csyrfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_float* a, lapack_int lda,
+                                 const lapack_complex_float* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const float* s, const lapack_complex_float* b,
+                                 lapack_int ldb, lapack_complex_float* x,
+                                 lapack_int ldx, float* rcond, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zsyrfsx_work( int matrix_order, char uplo, char equed,
+                                 lapack_int n, lapack_int nrhs,
+                                 const lapack_complex_double* a, lapack_int lda,
+                                 const lapack_complex_double* af,
+                                 lapack_int ldaf, const lapack_int* ipiv,
+                                 const double* s,
+                                 const lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_ssysv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, float* a, lapack_int lda,
+                               lapack_int* ipiv, float* b, lapack_int ldb,
+                               float* work, lapack_int lwork );
+lapack_int LAPACKE_dsysv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, double* a, lapack_int lda,
+                               lapack_int* ipiv, double* b, lapack_int ldb,
+                               double* work, lapack_int lwork );
+lapack_int LAPACKE_csysv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_float* a,
+                               lapack_int lda, lapack_int* ipiv,
+                               lapack_complex_float* b, lapack_int ldb,
+                               lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zsysv_work( int matrix_order, char uplo, lapack_int n,
+                               lapack_int nrhs, lapack_complex_double* a,
+                               lapack_int lda, lapack_int* ipiv,
+                               lapack_complex_double* b, lapack_int ldb,
+                               lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_ssysvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, const float* a,
+                                lapack_int lda, float* af, lapack_int ldaf,
+                                lapack_int* ipiv, const float* b,
+                                lapack_int ldb, float* x, lapack_int ldx,
+                                float* rcond, float* ferr, float* berr,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dsysvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs, const double* a,
+                                lapack_int lda, double* af, lapack_int ldaf,
+                                lapack_int* ipiv, const double* b,
+                                lapack_int ldb, double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_csysvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* af, lapack_int ldaf,
+                                lapack_int* ipiv, const lapack_complex_float* b,
+                                lapack_int ldb, lapack_complex_float* x,
+                                lapack_int ldx, float* rcond, float* ferr,
+                                float* berr, lapack_complex_float* work,
+                                lapack_int lwork, float* rwork );
+lapack_int LAPACKE_zsysvx_work( int matrix_order, char fact, char uplo,
+                                lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* af, lapack_int ldaf,
+                                lapack_int* ipiv,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* x, lapack_int ldx,
+                                double* rcond, double* ferr, double* berr,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork );
+
+lapack_int LAPACKE_ssysvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs, float* a,
+                                 lapack_int lda, float* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, float* s,
+                                 float* b, lapack_int ldb, float* x,
+                                 lapack_int ldx, float* rcond, float* rpvgrw,
+                                 float* berr, lapack_int n_err_bnds,
+                                 float* err_bnds_norm, float* err_bnds_comp,
+                                 lapack_int nparams, float* params, float* work,
+                                 lapack_int* iwork );
+lapack_int LAPACKE_dsysvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs, double* a,
+                                 lapack_int lda, double* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, double* s,
+                                 double* b, lapack_int ldb, double* x,
+                                 lapack_int ldx, double* rcond, double* rpvgrw,
+                                 double* berr, lapack_int n_err_bnds,
+                                 double* err_bnds_norm, double* err_bnds_comp,
+                                 lapack_int nparams, double* params,
+                                 double* work, lapack_int* iwork );
+lapack_int LAPACKE_csysvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, float* s,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* x, lapack_int ldx,
+                                 float* rcond, float* rpvgrw, float* berr,
+                                 lapack_int n_err_bnds, float* err_bnds_norm,
+                                 float* err_bnds_comp, lapack_int nparams,
+                                 float* params, lapack_complex_float* work,
+                                 float* rwork );
+lapack_int LAPACKE_zsysvxx_work( int matrix_order, char fact, char uplo,
+                                 lapack_int n, lapack_int nrhs,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* af, lapack_int ldaf,
+                                 lapack_int* ipiv, char* equed, double* s,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* x, lapack_int ldx,
+                                 double* rcond, double* rpvgrw, double* berr,
+                                 lapack_int n_err_bnds, double* err_bnds_norm,
+                                 double* err_bnds_comp, lapack_int nparams,
+                                 double* params, lapack_complex_double* work,
+                                 double* rwork );
+
+lapack_int LAPACKE_ssytrd_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda, float* d, float* e,
+                                float* tau, float* work, lapack_int lwork );
+lapack_int LAPACKE_dsytrd_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda, double* d, double* e,
+                                double* tau, double* work, lapack_int lwork );
+
+lapack_int LAPACKE_ssytrf_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda, lapack_int* ipiv,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dsytrf_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda, lapack_int* ipiv,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_csytrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_int* ipiv, lapack_complex_float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_zsytrf_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_int* ipiv, lapack_complex_double* work,
+                                lapack_int lwork );
+
+lapack_int LAPACKE_ssytri_work( int matrix_order, char uplo, lapack_int n,
+                                float* a, lapack_int lda,
+                                const lapack_int* ipiv, float* work );
+lapack_int LAPACKE_dsytri_work( int matrix_order, char uplo, lapack_int n,
+                                double* a, lapack_int lda,
+                                const lapack_int* ipiv, double* work );
+lapack_int LAPACKE_csytri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zsytri_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                const lapack_int* ipiv,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_ssytrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const float* a, lapack_int lda,
+                                const lapack_int* ipiv, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dsytrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                double* b, lapack_int ldb );
+lapack_int LAPACKE_csytrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_zsytrs_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_int nrhs, const lapack_complex_double* a,
+                                lapack_int lda, const lapack_int* ipiv,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stbcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, lapack_int kd,
+                                const float* ab, lapack_int ldab, float* rcond,
+                                float* work, lapack_int* iwork );
+lapack_int LAPACKE_dtbcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, lapack_int kd,
+                                const double* ab, lapack_int ldab,
+                                double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctbcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, lapack_int kd,
+                                const lapack_complex_float* ab, lapack_int ldab,
+                                float* rcond, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_ztbcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, lapack_int kd,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_stbrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const float* ab,
+                                lapack_int ldab, const float* b, lapack_int ldb,
+                                const float* x, lapack_int ldx, float* ferr,
+                                float* berr, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dtbrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const double* ab,
+                                lapack_int ldab, const double* b,
+                                lapack_int ldb, const double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctbrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const lapack_complex_float* ab,
+                                lapack_int ldab, const lapack_complex_float* b,
+                                lapack_int ldb, const lapack_complex_float* x,
+                                lapack_int ldx, float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztbrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, const lapack_complex_double* b,
+                                lapack_int ldb, const lapack_complex_double* x,
+                                lapack_int ldx, double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_stbtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const float* ab,
+                                lapack_int ldab, float* b, lapack_int ldb );
+lapack_int LAPACKE_dtbtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const double* ab,
+                                lapack_int ldab, double* b, lapack_int ldb );
+lapack_int LAPACKE_ctbtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs, const lapack_complex_float* ab,
+                                lapack_int ldab, lapack_complex_float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_ztbtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int kd,
+                                lapack_int nrhs,
+                                const lapack_complex_double* ab,
+                                lapack_int ldab, lapack_complex_double* b,
+                                lapack_int ldb );
+
+lapack_int LAPACKE_stfsm_work( int matrix_order, char transr, char side,
+                               char uplo, char trans, char diag, lapack_int m,
+                               lapack_int n, float alpha, const float* a,
+                               float* b, lapack_int ldb );
+lapack_int LAPACKE_dtfsm_work( int matrix_order, char transr, char side,
+                               char uplo, char trans, char diag, lapack_int m,
+                               lapack_int n, double alpha, const double* a,
+                               double* b, lapack_int ldb );
+lapack_int LAPACKE_ctfsm_work( int matrix_order, char transr, char side,
+                               char uplo, char trans, char diag, lapack_int m,
+                               lapack_int n, lapack_complex_float alpha,
+                               const lapack_complex_float* a,
+                               lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztfsm_work( int matrix_order, char transr, char side,
+                               char uplo, char trans, char diag, lapack_int m,
+                               lapack_int n, lapack_complex_double alpha,
+                               const lapack_complex_double* a,
+                               lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stftri_work( int matrix_order, char transr, char uplo,
+                                char diag, lapack_int n, float* a );
+lapack_int LAPACKE_dtftri_work( int matrix_order, char transr, char uplo,
+                                char diag, lapack_int n, double* a );
+lapack_int LAPACKE_ctftri_work( int matrix_order, char transr, char uplo,
+                                char diag, lapack_int n,
+                                lapack_complex_float* a );
+lapack_int LAPACKE_ztftri_work( int matrix_order, char transr, char uplo,
+                                char diag, lapack_int n,
+                                lapack_complex_double* a );
+
+lapack_int LAPACKE_stfttp_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const float* arf, float* ap );
+lapack_int LAPACKE_dtfttp_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const double* arf, double* ap );
+lapack_int LAPACKE_ctfttp_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_float* arf,
+                                lapack_complex_float* ap );
+lapack_int LAPACKE_ztfttp_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_double* arf,
+                                lapack_complex_double* ap );
+
+lapack_int LAPACKE_stfttr_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const float* arf, float* a,
+                                lapack_int lda );
+lapack_int LAPACKE_dtfttr_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const double* arf, double* a,
+                                lapack_int lda );
+lapack_int LAPACKE_ctfttr_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_float* arf,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_ztfttr_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_double* arf,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_stgevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const float* s, lapack_int lds, const float* p,
+                                lapack_int ldp, float* vl, lapack_int ldvl,
+                                float* vr, lapack_int ldvr, lapack_int mm,
+                                lapack_int* m, float* work );
+lapack_int LAPACKE_dtgevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const double* s, lapack_int lds,
+                                const double* p, lapack_int ldp, double* vl,
+                                lapack_int ldvl, double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m, double* work );
+lapack_int LAPACKE_ctgevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_float* s, lapack_int lds,
+                                const lapack_complex_float* p, lapack_int ldp,
+                                lapack_complex_float* vl, lapack_int ldvl,
+                                lapack_complex_float* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztgevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_double* s, lapack_int lds,
+                                const lapack_complex_double* p, lapack_int ldp,
+                                lapack_complex_double* vl, lapack_int ldvl,
+                                lapack_complex_double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_stgexc_work( int matrix_order, lapack_logical wantq,
+                                lapack_logical wantz, lapack_int n, float* a,
+                                lapack_int lda, float* b, lapack_int ldb,
+                                float* q, lapack_int ldq, float* z,
+                                lapack_int ldz, lapack_int* ifst,
+                                lapack_int* ilst, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dtgexc_work( int matrix_order, lapack_logical wantq,
+                                lapack_logical wantz, lapack_int n, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                double* q, lapack_int ldq, double* z,
+                                lapack_int ldz, lapack_int* ifst,
+                                lapack_int* ilst, double* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_ctgexc_work( int matrix_order, lapack_logical wantq,
+                                lapack_logical wantz, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_int ifst, lapack_int ilst );
+lapack_int LAPACKE_ztgexc_work( int matrix_order, lapack_logical wantq,
+                                lapack_logical wantz, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_int ifst, lapack_int ilst );
+
+lapack_int LAPACKE_stgsen_work( int matrix_order, lapack_int ijob,
+                                lapack_logical wantq, lapack_logical wantz,
+                                const lapack_logical* select, lapack_int n,
+                                float* a, lapack_int lda, float* b,
+                                lapack_int ldb, float* alphar, float* alphai,
+                                float* beta, float* q, lapack_int ldq, float* z,
+                                lapack_int ldz, lapack_int* m, float* pl,
+                                float* pr, float* dif, float* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+lapack_int LAPACKE_dtgsen_work( int matrix_order, lapack_int ijob,
+                                lapack_logical wantq, lapack_logical wantz,
+                                const lapack_logical* select, lapack_int n,
+                                double* a, lapack_int lda, double* b,
+                                lapack_int ldb, double* alphar, double* alphai,
+                                double* beta, double* q, lapack_int ldq,
+                                double* z, lapack_int ldz, lapack_int* m,
+                                double* pl, double* pr, double* dif,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_ctgsen_work( int matrix_order, lapack_int ijob,
+                                lapack_logical wantq, lapack_logical wantz,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* alpha,
+                                lapack_complex_float* beta,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* z, lapack_int ldz,
+                                lapack_int* m, float* pl, float* pr, float* dif,
+                                lapack_complex_float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_ztgsen_work( int matrix_order, lapack_int ijob,
+                                lapack_logical wantq, lapack_logical wantz,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* alpha,
+                                lapack_complex_double* beta,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* z, lapack_int ldz,
+                                lapack_int* m, double* pl, double* pr,
+                                double* dif, lapack_complex_double* work,
+                                lapack_int lwork, lapack_int* iwork,
+                                lapack_int liwork );
+
+lapack_int LAPACKE_stgsja_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                float* a, lapack_int lda, float* b,
+                                lapack_int ldb, float tola, float tolb,
+                                float* alpha, float* beta, float* u,
+                                lapack_int ldu, float* v, lapack_int ldv,
+                                float* q, lapack_int ldq, float* work,
+                                lapack_int* ncycle );
+lapack_int LAPACKE_dtgsja_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                double* a, lapack_int lda, double* b,
+                                lapack_int ldb, double tola, double tolb,
+                                double* alpha, double* beta, double* u,
+                                lapack_int ldu, double* v, lapack_int ldv,
+                                double* q, lapack_int ldq, double* work,
+                                lapack_int* ncycle );
+lapack_int LAPACKE_ctgsja_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                float tola, float tolb, float* alpha,
+                                float* beta, lapack_complex_float* u,
+                                lapack_int ldu, lapack_complex_float* v,
+                                lapack_int ldv, lapack_complex_float* q,
+                                lapack_int ldq, lapack_complex_float* work,
+                                lapack_int* ncycle );
+lapack_int LAPACKE_ztgsja_work( int matrix_order, char jobu, char jobv,
+                                char jobq, lapack_int m, lapack_int p,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                double tola, double tolb, double* alpha,
+                                double* beta, lapack_complex_double* u,
+                                lapack_int ldu, lapack_complex_double* v,
+                                lapack_int ldv, lapack_complex_double* q,
+                                lapack_int ldq, lapack_complex_double* work,
+                                lapack_int* ncycle );
+
+lapack_int LAPACKE_stgsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const float* a, lapack_int lda, const float* b,
+                                lapack_int ldb, const float* vl,
+                                lapack_int ldvl, const float* vr,
+                                lapack_int ldvr, float* s, float* dif,
+                                lapack_int mm, lapack_int* m, float* work,
+                                lapack_int lwork, lapack_int* iwork );
+lapack_int LAPACKE_dtgsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const double* a, lapack_int lda,
+                                const double* b, lapack_int ldb,
+                                const double* vl, lapack_int ldvl,
+                                const double* vr, lapack_int ldvr, double* s,
+                                double* dif, lapack_int mm, lapack_int* m,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctgsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                const lapack_complex_float* vl, lapack_int ldvl,
+                                const lapack_complex_float* vr, lapack_int ldvr,
+                                float* s, float* dif, lapack_int mm,
+                                lapack_int* m, lapack_complex_float* work,
+                                lapack_int lwork, lapack_int* iwork );
+lapack_int LAPACKE_ztgsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                const lapack_complex_double* vl,
+                                lapack_int ldvl,
+                                const lapack_complex_double* vr,
+                                lapack_int ldvr, double* s, double* dif,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_double* work, lapack_int lwork,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_stgsyl_work( int matrix_order, char trans, lapack_int ijob,
+                                lapack_int m, lapack_int n, const float* a,
+                                lapack_int lda, const float* b, lapack_int ldb,
+                                float* c, lapack_int ldc, const float* d,
+                                lapack_int ldd, const float* e, lapack_int lde,
+                                float* f, lapack_int ldf, float* scale,
+                                float* dif, float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dtgsyl_work( int matrix_order, char trans, lapack_int ijob,
+                                lapack_int m, lapack_int n, const double* a,
+                                lapack_int lda, const double* b, lapack_int ldb,
+                                double* c, lapack_int ldc, const double* d,
+                                lapack_int ldd, const double* e, lapack_int lde,
+                                double* f, lapack_int ldf, double* scale,
+                                double* dif, double* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctgsyl_work( int matrix_order, char trans, lapack_int ijob,
+                                lapack_int m, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* c, lapack_int ldc,
+                                const lapack_complex_float* d, lapack_int ldd,
+                                const lapack_complex_float* e, lapack_int lde,
+                                lapack_complex_float* f, lapack_int ldf,
+                                float* scale, float* dif,
+                                lapack_complex_float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ztgsyl_work( int matrix_order, char trans, lapack_int ijob,
+                                lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* c, lapack_int ldc,
+                                const lapack_complex_double* d, lapack_int ldd,
+                                const lapack_complex_double* e, lapack_int lde,
+                                lapack_complex_double* f, lapack_int ldf,
+                                double* scale, double* dif,
+                                lapack_complex_double* work, lapack_int lwork,
+                                lapack_int* iwork );
+
+lapack_int LAPACKE_stpcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, const float* ap,
+                                float* rcond, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dtpcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, const double* ap,
+                                double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctpcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n,
+                                const lapack_complex_float* ap, float* rcond,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztpcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n,
+                                const lapack_complex_double* ap, double* rcond,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_stprfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const float* ap, const float* b, lapack_int ldb,
+                                const float* x, lapack_int ldx, float* ferr,
+                                float* berr, float* work, lapack_int* iwork );
+lapack_int LAPACKE_dtprfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const double* ap, const double* b,
+                                lapack_int ldb, const double* x, lapack_int ldx,
+                                double* ferr, double* berr, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctprfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* ap,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                const lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztprfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                const lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_stptri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, float* ap );
+lapack_int LAPACKE_dtptri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, double* ap );
+lapack_int LAPACKE_ctptri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, lapack_complex_float* ap );
+lapack_int LAPACKE_ztptri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, lapack_complex_double* ap );
+
+lapack_int LAPACKE_stptrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const float* ap, float* b, lapack_int ldb );
+lapack_int LAPACKE_dtptrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const double* ap, double* b, lapack_int ldb );
+lapack_int LAPACKE_ctptrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* ap,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztptrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* ap,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_stpttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const float* ap, float* arf );
+lapack_int LAPACKE_dtpttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const double* ap, double* arf );
+lapack_int LAPACKE_ctpttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_float* ap,
+                                lapack_complex_float* arf );
+lapack_int LAPACKE_ztpttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_double* ap,
+                                lapack_complex_double* arf );
+
+lapack_int LAPACKE_stpttr_work( int matrix_order, char uplo, lapack_int n,
+                                const float* ap, float* a, lapack_int lda );
+lapack_int LAPACKE_dtpttr_work( int matrix_order, char uplo, lapack_int n,
+                                const double* ap, double* a, lapack_int lda );
+lapack_int LAPACKE_ctpttr_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap,
+                                lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_ztpttr_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap,
+                                lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_strcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, const float* a,
+                                lapack_int lda, float* rcond, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dtrcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n, const double* a,
+                                lapack_int lda, double* rcond, double* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctrcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                float* rcond, lapack_complex_float* work,
+                                float* rwork );
+lapack_int LAPACKE_ztrcon_work( int matrix_order, char norm, char uplo,
+                                char diag, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                double* rcond, lapack_complex_double* work,
+                                double* rwork );
+
+lapack_int LAPACKE_strevc_work( int matrix_order, char side, char howmny,
+                                lapack_logical* select, lapack_int n,
+                                const float* t, lapack_int ldt, float* vl,
+                                lapack_int ldvl, float* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m, float* work );
+lapack_int LAPACKE_dtrevc_work( int matrix_order, char side, char howmny,
+                                lapack_logical* select, lapack_int n,
+                                const double* t, lapack_int ldt, double* vl,
+                                lapack_int ldvl, double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m, double* work );
+lapack_int LAPACKE_ctrevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_float* t, lapack_int ldt,
+                                lapack_complex_float* vl, lapack_int ldvl,
+                                lapack_complex_float* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztrevc_work( int matrix_order, char side, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* vl, lapack_int ldvl,
+                                lapack_complex_double* vr, lapack_int ldvr,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_strexc_work( int matrix_order, char compq, lapack_int n,
+                                float* t, lapack_int ldt, float* q,
+                                lapack_int ldq, lapack_int* ifst,
+                                lapack_int* ilst, float* work );
+lapack_int LAPACKE_dtrexc_work( int matrix_order, char compq, lapack_int n,
+                                double* t, lapack_int ldt, double* q,
+                                lapack_int ldq, lapack_int* ifst,
+                                lapack_int* ilst, double* work );
+lapack_int LAPACKE_ctrexc_work( int matrix_order, char compq, lapack_int n,
+                                lapack_complex_float* t, lapack_int ldt,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_int ifst, lapack_int ilst );
+lapack_int LAPACKE_ztrexc_work( int matrix_order, char compq, lapack_int n,
+                                lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_int ifst, lapack_int ilst );
+
+lapack_int LAPACKE_strrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const float* a, lapack_int lda, const float* b,
+                                lapack_int ldb, const float* x, lapack_int ldx,
+                                float* ferr, float* berr, float* work,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dtrrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const double* a, lapack_int lda,
+                                const double* b, lapack_int ldb,
+                                const double* x, lapack_int ldx, double* ferr,
+                                double* berr, double* work, lapack_int* iwork );
+lapack_int LAPACKE_ctrrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                const lapack_complex_float* x, lapack_int ldx,
+                                float* ferr, float* berr,
+                                lapack_complex_float* work, float* rwork );
+lapack_int LAPACKE_ztrrfs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                const lapack_complex_double* x, lapack_int ldx,
+                                double* ferr, double* berr,
+                                lapack_complex_double* work, double* rwork );
+
+lapack_int LAPACKE_strsen_work( int matrix_order, char job, char compq,
+                                const lapack_logical* select, lapack_int n,
+                                float* t, lapack_int ldt, float* q,
+                                lapack_int ldq, float* wr, float* wi,
+                                lapack_int* m, float* s, float* sep,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_dtrsen_work( int matrix_order, char job, char compq,
+                                const lapack_logical* select, lapack_int n,
+                                double* t, lapack_int ldt, double* q,
+                                lapack_int ldq, double* wr, double* wi,
+                                lapack_int* m, double* s, double* sep,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork, lapack_int liwork );
+lapack_int LAPACKE_ctrsen_work( int matrix_order, char job, char compq,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_float* t, lapack_int ldt,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* w, lapack_int* m,
+                                float* s, float* sep,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_ztrsen_work( int matrix_order, char job, char compq,
+                                const lapack_logical* select, lapack_int n,
+                                lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* w, lapack_int* m,
+                                double* s, double* sep,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_strsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const float* t, lapack_int ldt, const float* vl,
+                                lapack_int ldvl, const float* vr,
+                                lapack_int ldvr, float* s, float* sep,
+                                lapack_int mm, lapack_int* m, float* work,
+                                lapack_int ldwork, lapack_int* iwork );
+lapack_int LAPACKE_dtrsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const double* t, lapack_int ldt,
+                                const double* vl, lapack_int ldvl,
+                                const double* vr, lapack_int ldvr, double* s,
+                                double* sep, lapack_int mm, lapack_int* m,
+                                double* work, lapack_int ldwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ctrsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_float* t, lapack_int ldt,
+                                const lapack_complex_float* vl, lapack_int ldvl,
+                                const lapack_complex_float* vr, lapack_int ldvr,
+                                float* s, float* sep, lapack_int mm,
+                                lapack_int* m, lapack_complex_float* work,
+                                lapack_int ldwork, float* rwork );
+lapack_int LAPACKE_ztrsna_work( int matrix_order, char job, char howmny,
+                                const lapack_logical* select, lapack_int n,
+                                const lapack_complex_double* t, lapack_int ldt,
+                                const lapack_complex_double* vl,
+                                lapack_int ldvl,
+                                const lapack_complex_double* vr,
+                                lapack_int ldvr, double* s, double* sep,
+                                lapack_int mm, lapack_int* m,
+                                lapack_complex_double* work, lapack_int ldwork,
+                                double* rwork );
+
+lapack_int LAPACKE_strsyl_work( int matrix_order, char trana, char tranb,
+                                lapack_int isgn, lapack_int m, lapack_int n,
+                                const float* a, lapack_int lda, const float* b,
+                                lapack_int ldb, float* c, lapack_int ldc,
+                                float* scale );
+lapack_int LAPACKE_dtrsyl_work( int matrix_order, char trana, char tranb,
+                                lapack_int isgn, lapack_int m, lapack_int n,
+                                const double* a, lapack_int lda,
+                                const double* b, lapack_int ldb, double* c,
+                                lapack_int ldc, double* scale );
+lapack_int LAPACKE_ctrsyl_work( int matrix_order, char trana, char tranb,
+                                lapack_int isgn, lapack_int m, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* b, lapack_int ldb,
+                                lapack_complex_float* c, lapack_int ldc,
+                                float* scale );
+lapack_int LAPACKE_ztrsyl_work( int matrix_order, char trana, char tranb,
+                                lapack_int isgn, lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* c, lapack_int ldc,
+                                double* scale );
+
+lapack_int LAPACKE_strtri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, float* a, lapack_int lda );
+lapack_int LAPACKE_dtrtri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, double* a, lapack_int lda );
+lapack_int LAPACKE_ctrtri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, lapack_complex_float* a,
+                                lapack_int lda );
+lapack_int LAPACKE_ztrtri_work( int matrix_order, char uplo, char diag,
+                                lapack_int n, lapack_complex_double* a,
+                                lapack_int lda );
+
+lapack_int LAPACKE_strtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const float* a, lapack_int lda, float* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_dtrtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const double* a, lapack_int lda, double* b,
+                                lapack_int ldb );
+lapack_int LAPACKE_ctrtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztrtrs_work( int matrix_order, char uplo, char trans,
+                                char diag, lapack_int n, lapack_int nrhs,
+                                const lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_strttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const float* a, lapack_int lda,
+                                float* arf );
+lapack_int LAPACKE_dtrttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const double* a, lapack_int lda,
+                                double* arf );
+lapack_int LAPACKE_ctrttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* arf );
+lapack_int LAPACKE_ztrttf_work( int matrix_order, char transr, char uplo,
+                                lapack_int n, const lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* arf );
+
+lapack_int LAPACKE_strttp_work( int matrix_order, char uplo, lapack_int n,
+                                const float* a, lapack_int lda, float* ap );
+lapack_int LAPACKE_dtrttp_work( int matrix_order, char uplo, lapack_int n,
+                                const double* a, lapack_int lda, double* ap );
+lapack_int LAPACKE_ctrttp_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* ap );
+lapack_int LAPACKE_ztrttp_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* ap );
+
+lapack_int LAPACKE_stzrzf_work( int matrix_order, lapack_int m, lapack_int n,
+                                float* a, lapack_int lda, float* tau,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_dtzrzf_work( int matrix_order, lapack_int m, lapack_int n,
+                                double* a, lapack_int lda, double* tau,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_ctzrzf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_ztzrzf_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cungbr_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int k,
+                                lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zungbr_work( int matrix_order, char vect, lapack_int m,
+                                lapack_int n, lapack_int k,
+                                lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunghr_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunghr_work( int matrix_order, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunglq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunglq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cungql_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zungql_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cungqr_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zungqr_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cungrq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zungrq_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int k, lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cungtr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zungtr_work( int matrix_order, char uplo, lapack_int n,
+                                lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmbr_work( int matrix_order, char vect, char side,
+                                char trans, lapack_int m, lapack_int n,
+                                lapack_int k, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmbr_work( int matrix_order, char vect, char side,
+                                char trans, lapack_int m, lapack_int n,
+                                lapack_int k, const lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmhr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmhr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int ilo,
+                                lapack_int ihi, const lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmlq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmlq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmql_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmql_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmqr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmqr_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmrq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmrq_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmrz_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                lapack_int l, const lapack_complex_float* a,
+                                lapack_int lda, const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmrz_work( int matrix_order, char side, char trans,
+                                lapack_int m, lapack_int n, lapack_int k,
+                                lapack_int l, const lapack_complex_double* a,
+                                lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cunmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const lapack_complex_float* a, lapack_int lda,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_zunmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const lapack_complex_double* a, lapack_int lda,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work, lapack_int lwork );
+
+lapack_int LAPACKE_cupgtr_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_float* ap,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* q, lapack_int ldq,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zupgtr_work( int matrix_order, char uplo, lapack_int n,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* q, lapack_int ldq,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_cupmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const lapack_complex_float* ap,
+                                const lapack_complex_float* tau,
+                                lapack_complex_float* c, lapack_int ldc,
+                                lapack_complex_float* work );
+lapack_int LAPACKE_zupmtr_work( int matrix_order, char side, char uplo,
+                                char trans, lapack_int m, lapack_int n,
+                                const lapack_complex_double* ap,
+                                const lapack_complex_double* tau,
+                                lapack_complex_double* c, lapack_int ldc,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_claghe( int matrix_order, lapack_int n, lapack_int k,
+                           const float* d, lapack_complex_float* a,
+                           lapack_int lda, lapack_int* iseed );
+lapack_int LAPACKE_zlaghe( int matrix_order, lapack_int n, lapack_int k,
+                           const double* d, lapack_complex_double* a,
+                           lapack_int lda, lapack_int* iseed );
+
+lapack_int LAPACKE_slagsy( int matrix_order, lapack_int n, lapack_int k,
+                           const float* d, float* a, lapack_int lda,
+                           lapack_int* iseed );
+lapack_int LAPACKE_dlagsy( int matrix_order, lapack_int n, lapack_int k,
+                           const double* d, double* a, lapack_int lda,
+                           lapack_int* iseed );
+lapack_int LAPACKE_clagsy( int matrix_order, lapack_int n, lapack_int k,
+                           const float* d, lapack_complex_float* a,
+                           lapack_int lda, lapack_int* iseed );
+lapack_int LAPACKE_zlagsy( int matrix_order, lapack_int n, lapack_int k,
+                           const double* d, lapack_complex_double* a,
+                           lapack_int lda, lapack_int* iseed );
+
+lapack_int LAPACKE_slapmr( int matrix_order, lapack_logical forwrd,
+                           lapack_int m, lapack_int n, float* x, lapack_int ldx,
+                           lapack_int* k );
+lapack_int LAPACKE_dlapmr( int matrix_order, lapack_logical forwrd,
+                           lapack_int m, lapack_int n, double* x,
+                           lapack_int ldx, lapack_int* k );
+lapack_int LAPACKE_clapmr( int matrix_order, lapack_logical forwrd,
+                           lapack_int m, lapack_int n, lapack_complex_float* x,
+                           lapack_int ldx, lapack_int* k );
+lapack_int LAPACKE_zlapmr( int matrix_order, lapack_logical forwrd,
+                           lapack_int m, lapack_int n, lapack_complex_double* x,
+                           lapack_int ldx, lapack_int* k );
+
+
+float LAPACKE_slapy2( float x, float y );
+double LAPACKE_dlapy2( double x, double y );
+
+float LAPACKE_slapy3( float x, float y, float z );
+double LAPACKE_dlapy3( double x, double y, double z );
+
+lapack_int LAPACKE_slartgp( float f, float g, float* cs, float* sn, float* r );
+lapack_int LAPACKE_dlartgp( double f, double g, double* cs, double* sn,
+                            double* r );
+
+lapack_int LAPACKE_slartgs( float x, float y, float sigma, float* cs,
+                            float* sn );
+lapack_int LAPACKE_dlartgs( double x, double y, double sigma, double* cs,
+                            double* sn );
+
+
+//LAPACK 3.3.0
+lapack_int LAPACKE_cbbcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, lapack_int m,
+                           lapack_int p, lapack_int q, float* theta, float* phi,
+                           lapack_complex_float* u1, lapack_int ldu1,
+                           lapack_complex_float* u2, lapack_int ldu2,
+                           lapack_complex_float* v1t, lapack_int ldv1t,
+                           lapack_complex_float* v2t, lapack_int ldv2t,
+                           float* b11d, float* b11e, float* b12d, float* b12e,
+                           float* b21d, float* b21e, float* b22d, float* b22e );
+lapack_int LAPACKE_cbbcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                float* theta, float* phi,
+                                lapack_complex_float* u1, lapack_int ldu1,
+                                lapack_complex_float* u2, lapack_int ldu2,
+                                lapack_complex_float* v1t, lapack_int ldv1t,
+                                lapack_complex_float* v2t, lapack_int ldv2t,
+                                float* b11d, float* b11e, float* b12d,
+                                float* b12e, float* b21d, float* b21e,
+                                float* b22d, float* b22e, float* rwork,
+                                lapack_int lrwork );
+lapack_int LAPACKE_cheswapr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_float* a, lapack_int i1,
+                             lapack_int i2 );
+lapack_int LAPACKE_cheswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_float* a, lapack_int i1,
+                                  lapack_int i2 );
+lapack_int LAPACKE_chetri2( int matrix_order, char uplo, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_chetri2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_chetri2x( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_float* a, lapack_int lda,
+                             const lapack_int* ipiv, lapack_int nb );
+lapack_int LAPACKE_chetri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_float* a, lapack_int lda,
+                                  const lapack_int* ipiv,
+                                  lapack_complex_float* work, lapack_int nb );
+lapack_int LAPACKE_chetrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const lapack_complex_float* a,
+                            lapack_int lda, const lapack_int* ipiv,
+                            lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_chetrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const lapack_complex_float* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_csyconv( int matrix_order, char uplo, char way, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_csyconv_work( int matrix_order, char uplo, char way,
+                                 lapack_int n, lapack_complex_float* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_csyswapr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_float* a, lapack_int i1,
+                             lapack_int i2 );
+lapack_int LAPACKE_csyswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_float* a, lapack_int i1,
+                                  lapack_int i2 );
+lapack_int LAPACKE_csytri2( int matrix_order, char uplo, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_csytri2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_csytri2x( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_float* a, lapack_int lda,
+                             const lapack_int* ipiv, lapack_int nb );
+lapack_int LAPACKE_csytri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_float* a, lapack_int lda,
+                                  const lapack_int* ipiv,
+                                  lapack_complex_float* work, lapack_int nb );
+lapack_int LAPACKE_csytrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const lapack_complex_float* a,
+                            lapack_int lda, const lapack_int* ipiv,
+                            lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_csytrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const lapack_complex_float* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_cunbdb( int matrix_order, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           lapack_complex_float* x11, lapack_int ldx11,
+                           lapack_complex_float* x12, lapack_int ldx12,
+                           lapack_complex_float* x21, lapack_int ldx21,
+                           lapack_complex_float* x22, lapack_int ldx22,
+                           float* theta, float* phi,
+                           lapack_complex_float* taup1,
+                           lapack_complex_float* taup2,
+                           lapack_complex_float* tauq1,
+                           lapack_complex_float* tauq2 );
+lapack_int LAPACKE_cunbdb_work( int matrix_order, char trans, char signs,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                lapack_complex_float* x11, lapack_int ldx11,
+                                lapack_complex_float* x12, lapack_int ldx12,
+                                lapack_complex_float* x21, lapack_int ldx21,
+                                lapack_complex_float* x22, lapack_int ldx22,
+                                float* theta, float* phi,
+                                lapack_complex_float* taup1,
+                                lapack_complex_float* taup2,
+                                lapack_complex_float* tauq1,
+                                lapack_complex_float* tauq2,
+                                lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_cuncsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           lapack_complex_float* x11, lapack_int ldx11,
+                           lapack_complex_float* x12, lapack_int ldx12,
+                           lapack_complex_float* x21, lapack_int ldx21,
+                           lapack_complex_float* x22, lapack_int ldx22,
+                           float* theta, lapack_complex_float* u1,
+                           lapack_int ldu1, lapack_complex_float* u2,
+                           lapack_int ldu2, lapack_complex_float* v1t,
+                           lapack_int ldv1t, lapack_complex_float* v2t,
+                           lapack_int ldv2t );
+lapack_int LAPACKE_cuncsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                char signs, lapack_int m, lapack_int p,
+                                lapack_int q, lapack_complex_float* x11,
+                                lapack_int ldx11, lapack_complex_float* x12,
+                                lapack_int ldx12, lapack_complex_float* x21,
+                                lapack_int ldx21, lapack_complex_float* x22,
+                                lapack_int ldx22, float* theta,
+                                lapack_complex_float* u1, lapack_int ldu1,
+                                lapack_complex_float* u2, lapack_int ldu2,
+                                lapack_complex_float* v1t, lapack_int ldv1t,
+                                lapack_complex_float* v2t, lapack_int ldv2t,
+                                lapack_complex_float* work, lapack_int lwork,
+                                float* rwork, lapack_int lrwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dbbcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, lapack_int m,
+                           lapack_int p, lapack_int q, double* theta,
+                           double* phi, double* u1, lapack_int ldu1, double* u2,
+                           lapack_int ldu2, double* v1t, lapack_int ldv1t,
+                           double* v2t, lapack_int ldv2t, double* b11d,
+                           double* b11e, double* b12d, double* b12e,
+                           double* b21d, double* b21e, double* b22d,
+                           double* b22e );
+lapack_int LAPACKE_dbbcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                double* theta, double* phi, double* u1,
+                                lapack_int ldu1, double* u2, lapack_int ldu2,
+                                double* v1t, lapack_int ldv1t, double* v2t,
+                                lapack_int ldv2t, double* b11d, double* b11e,
+                                double* b12d, double* b12e, double* b21d,
+                                double* b21e, double* b22d, double* b22e,
+                                double* work, lapack_int lwork );
+lapack_int LAPACKE_dorbdb( int matrix_order, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           double* x11, lapack_int ldx11, double* x12,
+                           lapack_int ldx12, double* x21, lapack_int ldx21,
+                           double* x22, lapack_int ldx22, double* theta,
+                           double* phi, double* taup1, double* taup2,
+                           double* tauq1, double* tauq2 );
+lapack_int LAPACKE_dorbdb_work( int matrix_order, char trans, char signs,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                double* x11, lapack_int ldx11, double* x12,
+                                lapack_int ldx12, double* x21, lapack_int ldx21,
+                                double* x22, lapack_int ldx22, double* theta,
+                                double* phi, double* taup1, double* taup2,
+                                double* tauq1, double* tauq2, double* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_dorcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           double* x11, lapack_int ldx11, double* x12,
+                           lapack_int ldx12, double* x21, lapack_int ldx21,
+                           double* x22, lapack_int ldx22, double* theta,
+                           double* u1, lapack_int ldu1, double* u2,
+                           lapack_int ldu2, double* v1t, lapack_int ldv1t,
+                           double* v2t, lapack_int ldv2t );
+lapack_int LAPACKE_dorcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                char signs, lapack_int m, lapack_int p,
+                                lapack_int q, double* x11, lapack_int ldx11,
+                                double* x12, lapack_int ldx12, double* x21,
+                                lapack_int ldx21, double* x22, lapack_int ldx22,
+                                double* theta, double* u1, lapack_int ldu1,
+                                double* u2, lapack_int ldu2, double* v1t,
+                                lapack_int ldv1t, double* v2t, lapack_int ldv2t,
+                                double* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_dsyconv( int matrix_order, char uplo, char way, lapack_int n,
+                            double* a, lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_dsyconv_work( int matrix_order, char uplo, char way,
+                                 lapack_int n, double* a, lapack_int lda,
+                                 const lapack_int* ipiv, double* work );
+lapack_int LAPACKE_dsyswapr( int matrix_order, char uplo, lapack_int n,
+                             double* a, lapack_int i1, lapack_int i2 );
+lapack_int LAPACKE_dsyswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  double* a, lapack_int i1, lapack_int i2 );
+lapack_int LAPACKE_dsytri2( int matrix_order, char uplo, lapack_int n,
+                            double* a, lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_dsytri2_work( int matrix_order, char uplo, lapack_int n,
+                                 double* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_double* work, lapack_int lwork );
+lapack_int LAPACKE_dsytri2x( int matrix_order, char uplo, lapack_int n,
+                             double* a, lapack_int lda, const lapack_int* ipiv,
+                             lapack_int nb );
+lapack_int LAPACKE_dsytri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  double* a, lapack_int lda,
+                                  const lapack_int* ipiv, double* work,
+                                  lapack_int nb );
+lapack_int LAPACKE_dsytrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const double* a, lapack_int lda,
+                            const lapack_int* ipiv, double* b, lapack_int ldb );
+lapack_int LAPACKE_dsytrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const double* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 double* b, lapack_int ldb, double* work );
+lapack_int LAPACKE_sbbcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, lapack_int m,
+                           lapack_int p, lapack_int q, float* theta, float* phi,
+                           float* u1, lapack_int ldu1, float* u2,
+                           lapack_int ldu2, float* v1t, lapack_int ldv1t,
+                           float* v2t, lapack_int ldv2t, float* b11d,
+                           float* b11e, float* b12d, float* b12e, float* b21d,
+                           float* b21e, float* b22d, float* b22e );
+lapack_int LAPACKE_sbbcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                float* theta, float* phi, float* u1,
+                                lapack_int ldu1, float* u2, lapack_int ldu2,
+                                float* v1t, lapack_int ldv1t, float* v2t,
+                                lapack_int ldv2t, float* b11d, float* b11e,
+                                float* b12d, float* b12e, float* b21d,
+                                float* b21e, float* b22d, float* b22e,
+                                float* work, lapack_int lwork );
+lapack_int LAPACKE_sorbdb( int matrix_order, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q, float* x11,
+                           lapack_int ldx11, float* x12, lapack_int ldx12,
+                           float* x21, lapack_int ldx21, float* x22,
+                           lapack_int ldx22, float* theta, float* phi,
+                           float* taup1, float* taup2, float* tauq1,
+                           float* tauq2 );
+lapack_int LAPACKE_sorbdb_work( int matrix_order, char trans, char signs,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                float* x11, lapack_int ldx11, float* x12,
+                                lapack_int ldx12, float* x21, lapack_int ldx21,
+                                float* x22, lapack_int ldx22, float* theta,
+                                float* phi, float* taup1, float* taup2,
+                                float* tauq1, float* tauq2, float* work,
+                                lapack_int lwork );
+lapack_int LAPACKE_sorcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q, float* x11,
+                           lapack_int ldx11, float* x12, lapack_int ldx12,
+                           float* x21, lapack_int ldx21, float* x22,
+                           lapack_int ldx22, float* theta, float* u1,
+                           lapack_int ldu1, float* u2, lapack_int ldu2,
+                           float* v1t, lapack_int ldv1t, float* v2t,
+                           lapack_int ldv2t );
+lapack_int LAPACKE_sorcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                char signs, lapack_int m, lapack_int p,
+                                lapack_int q, float* x11, lapack_int ldx11,
+                                float* x12, lapack_int ldx12, float* x21,
+                                lapack_int ldx21, float* x22, lapack_int ldx22,
+                                float* theta, float* u1, lapack_int ldu1,
+                                float* u2, lapack_int ldu2, float* v1t,
+                                lapack_int ldv1t, float* v2t, lapack_int ldv2t,
+                                float* work, lapack_int lwork,
+                                lapack_int* iwork );
+lapack_int LAPACKE_ssyconv( int matrix_order, char uplo, char way, lapack_int n,
+                            float* a, lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_ssyconv_work( int matrix_order, char uplo, char way,
+                                 lapack_int n, float* a, lapack_int lda,
+                                 const lapack_int* ipiv, float* work );
+lapack_int LAPACKE_ssyswapr( int matrix_order, char uplo, lapack_int n,
+                             float* a, lapack_int i1, lapack_int i2 );
+lapack_int LAPACKE_ssyswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  float* a, lapack_int i1, lapack_int i2 );
+lapack_int LAPACKE_ssytri2( int matrix_order, char uplo, lapack_int n, float* a,
+                            lapack_int lda, const lapack_int* ipiv );
+lapack_int LAPACKE_ssytri2_work( int matrix_order, char uplo, lapack_int n,
+                                 float* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_float* work, lapack_int lwork );
+lapack_int LAPACKE_ssytri2x( int matrix_order, char uplo, lapack_int n,
+                             float* a, lapack_int lda, const lapack_int* ipiv,
+                             lapack_int nb );
+lapack_int LAPACKE_ssytri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  float* a, lapack_int lda,
+                                  const lapack_int* ipiv, float* work,
+                                  lapack_int nb );
+lapack_int LAPACKE_ssytrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const float* a, lapack_int lda,
+                            const lapack_int* ipiv, float* b, lapack_int ldb );
+lapack_int LAPACKE_ssytrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const float* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 float* b, lapack_int ldb, float* work );
+lapack_int LAPACKE_zbbcsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, lapack_int m,
+                           lapack_int p, lapack_int q, double* theta,
+                           double* phi, lapack_complex_double* u1,
+                           lapack_int ldu1, lapack_complex_double* u2,
+                           lapack_int ldu2, lapack_complex_double* v1t,
+                           lapack_int ldv1t, lapack_complex_double* v2t,
+                           lapack_int ldv2t, double* b11d, double* b11e,
+                           double* b12d, double* b12e, double* b21d,
+                           double* b21e, double* b22d, double* b22e );
+lapack_int LAPACKE_zbbcsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                double* theta, double* phi,
+                                lapack_complex_double* u1, lapack_int ldu1,
+                                lapack_complex_double* u2, lapack_int ldu2,
+                                lapack_complex_double* v1t, lapack_int ldv1t,
+                                lapack_complex_double* v2t, lapack_int ldv2t,
+                                double* b11d, double* b11e, double* b12d,
+                                double* b12e, double* b21d, double* b21e,
+                                double* b22d, double* b22e, double* rwork,
+                                lapack_int lrwork );
+lapack_int LAPACKE_zheswapr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_double* a, lapack_int i1,
+                             lapack_int i2 );
+lapack_int LAPACKE_zheswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_double* a, lapack_int i1,
+                                  lapack_int i2 );
+lapack_int LAPACKE_zhetri2( int matrix_order, char uplo, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_zhetri2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_double* work, lapack_int lwork );
+lapack_int LAPACKE_zhetri2x( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_double* a, lapack_int lda,
+                             const lapack_int* ipiv, lapack_int nb );
+lapack_int LAPACKE_zhetri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_double* a, lapack_int lda,
+                                  const lapack_int* ipiv,
+                                  lapack_complex_double* work, lapack_int nb );
+lapack_int LAPACKE_zhetrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const lapack_complex_double* a,
+                            lapack_int lda, const lapack_int* ipiv,
+                            lapack_complex_double* b, lapack_int ldb );
+lapack_int LAPACKE_zhetrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const lapack_complex_double* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* work );
+lapack_int LAPACKE_zsyconv( int matrix_order, char uplo, char way, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_zsyconv_work( int matrix_order, char uplo, char way,
+                                 lapack_int n, lapack_complex_double* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_double* work );
+lapack_int LAPACKE_zsyswapr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_double* a, lapack_int i1,
+                             lapack_int i2 );
+lapack_int LAPACKE_zsyswapr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_double* a, lapack_int i1,
+                                  lapack_int i2 );
+lapack_int LAPACKE_zsytri2( int matrix_order, char uplo, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            const lapack_int* ipiv );
+lapack_int LAPACKE_zsytri2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 const lapack_int* ipiv,
+                                 lapack_complex_double* work, lapack_int lwork );
+lapack_int LAPACKE_zsytri2x( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_double* a, lapack_int lda,
+                             const lapack_int* ipiv, lapack_int nb );
+lapack_int LAPACKE_zsytri2x_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_double* a, lapack_int lda,
+                                  const lapack_int* ipiv,
+                                  lapack_complex_double* work, lapack_int nb );
+lapack_int LAPACKE_zsytrs2( int matrix_order, char uplo, lapack_int n,
+                            lapack_int nrhs, const lapack_complex_double* a,
+                            lapack_int lda, const lapack_int* ipiv,
+                            lapack_complex_double* b, lapack_int ldb );
+lapack_int LAPACKE_zsytrs2_work( int matrix_order, char uplo, lapack_int n,
+                                 lapack_int nrhs, const lapack_complex_double* a,
+                                 lapack_int lda, const lapack_int* ipiv,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* work );
+lapack_int LAPACKE_zunbdb( int matrix_order, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           lapack_complex_double* x11, lapack_int ldx11,
+                           lapack_complex_double* x12, lapack_int ldx12,
+                           lapack_complex_double* x21, lapack_int ldx21,
+                           lapack_complex_double* x22, lapack_int ldx22,
+                           double* theta, double* phi,
+                           lapack_complex_double* taup1,
+                           lapack_complex_double* taup2,
+                           lapack_complex_double* tauq1,
+                           lapack_complex_double* tauq2 );
+lapack_int LAPACKE_zunbdb_work( int matrix_order, char trans, char signs,
+                                lapack_int m, lapack_int p, lapack_int q,
+                                lapack_complex_double* x11, lapack_int ldx11,
+                                lapack_complex_double* x12, lapack_int ldx12,
+                                lapack_complex_double* x21, lapack_int ldx21,
+                                lapack_complex_double* x22, lapack_int ldx22,
+                                double* theta, double* phi,
+                                lapack_complex_double* taup1,
+                                lapack_complex_double* taup2,
+                                lapack_complex_double* tauq1,
+                                lapack_complex_double* tauq2,
+                                lapack_complex_double* work, lapack_int lwork );
+lapack_int LAPACKE_zuncsd( int matrix_order, char jobu1, char jobu2,
+                           char jobv1t, char jobv2t, char trans, char signs,
+                           lapack_int m, lapack_int p, lapack_int q,
+                           lapack_complex_double* x11, lapack_int ldx11,
+                           lapack_complex_double* x12, lapack_int ldx12,
+                           lapack_complex_double* x21, lapack_int ldx21,
+                           lapack_complex_double* x22, lapack_int ldx22,
+                           double* theta, lapack_complex_double* u1,
+                           lapack_int ldu1, lapack_complex_double* u2,
+                           lapack_int ldu2, lapack_complex_double* v1t,
+                           lapack_int ldv1t, lapack_complex_double* v2t,
+                           lapack_int ldv2t );
+lapack_int LAPACKE_zuncsd_work( int matrix_order, char jobu1, char jobu2,
+                                char jobv1t, char jobv2t, char trans,
+                                char signs, lapack_int m, lapack_int p,
+                                lapack_int q, lapack_complex_double* x11,
+                                lapack_int ldx11, lapack_complex_double* x12,
+                                lapack_int ldx12, lapack_complex_double* x21,
+                                lapack_int ldx21, lapack_complex_double* x22,
+                                lapack_int ldx22, double* theta,
+                                lapack_complex_double* u1, lapack_int ldu1,
+                                lapack_complex_double* u2, lapack_int ldu2,
+                                lapack_complex_double* v1t, lapack_int ldv1t,
+                                lapack_complex_double* v2t, lapack_int ldv2t,
+                                lapack_complex_double* work, lapack_int lwork,
+                                double* rwork, lapack_int lrwork,
+                                lapack_int* iwork );
+//LAPACK 3.4.0
+lapack_int LAPACKE_sgemqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int nb, const float* v, lapack_int ldv,
+                            const float* t, lapack_int ldt, float* c,
+                            lapack_int ldc );
+lapack_int LAPACKE_dgemqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int nb, const double* v, lapack_int ldv,
+                            const double* t, lapack_int ldt, double* c,
+                            lapack_int ldc );
+lapack_int LAPACKE_cgemqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int nb, const lapack_complex_float* v,
+                            lapack_int ldv, const lapack_complex_float* t,
+                            lapack_int ldt, lapack_complex_float* c,
+                            lapack_int ldc );
+lapack_int LAPACKE_zgemqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int nb, const lapack_complex_double* v,
+                            lapack_int ldv, const lapack_complex_double* t,
+                            lapack_int ldt, lapack_complex_double* c,
+                            lapack_int ldc );
+
+lapack_int LAPACKE_sgeqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nb, float* a, lapack_int lda, float* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_dgeqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nb, double* a, lapack_int lda, double* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_cgeqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nb, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_zgeqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int nb, lapack_complex_double* a,
+                           lapack_int lda, lapack_complex_double* t,
+                           lapack_int ldt );
+
+lapack_int LAPACKE_sgeqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            float* a, lapack_int lda, float* t,
+                            lapack_int ldt );
+lapack_int LAPACKE_dgeqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            double* a, lapack_int lda, double* t,
+                            lapack_int ldt );
+lapack_int LAPACKE_cgeqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zgeqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_sgeqrt3( int matrix_order, lapack_int m, lapack_int n,
+                            float* a, lapack_int lda, float* t,
+                            lapack_int ldt );
+lapack_int LAPACKE_dgeqrt3( int matrix_order, lapack_int m, lapack_int n,
+                            double* a, lapack_int lda, double* t,
+                            lapack_int ldt );
+lapack_int LAPACKE_cgeqrt3( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zgeqrt3( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_stpmqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int l, lapack_int nb, const float* v,
+                            lapack_int ldv, const float* t, lapack_int ldt,
+                            float* a, lapack_int lda, float* b,
+                            lapack_int ldb );
+lapack_int LAPACKE_dtpmqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int l, lapack_int nb, const double* v,
+                            lapack_int ldv, const double* t, lapack_int ldt,
+                            double* a, lapack_int lda, double* b,
+                            lapack_int ldb );
+lapack_int LAPACKE_ctpmqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int l, lapack_int nb,
+                            const lapack_complex_float* v, lapack_int ldv,
+                            const lapack_complex_float* t, lapack_int ldt,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* b, lapack_int ldb );
+lapack_int LAPACKE_ztpmqrt( int matrix_order, char side, char trans,
+                            lapack_int m, lapack_int n, lapack_int k,
+                            lapack_int l, lapack_int nb,
+                            const lapack_complex_double* v, lapack_int ldv,
+                            const lapack_complex_double* t, lapack_int ldt,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* b, lapack_int ldb );
+
+lapack_int LAPACKE_dtpqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int l, lapack_int nb, double* a,
+                           lapack_int lda, double* b, lapack_int ldb, double* t,
+                           lapack_int ldt );
+lapack_int LAPACKE_ctpqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int l, lapack_int nb, lapack_complex_float* a,
+                           lapack_int lda, lapack_complex_float* t,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_int ldt );
+lapack_int LAPACKE_ztpqrt( int matrix_order, lapack_int m, lapack_int n,
+                           lapack_int l, lapack_int nb,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_stpqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            float* a, lapack_int lda, float* b, lapack_int ldb,
+                            float* t, lapack_int ldt );
+lapack_int LAPACKE_dtpqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            double* a, lapack_int lda, double* b,
+                            lapack_int ldb, double* t, lapack_int ldt );
+lapack_int LAPACKE_ctpqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_float* a, lapack_int lda,
+                            lapack_complex_float* b, lapack_int ldb,
+                            lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_ztpqrt2( int matrix_order, lapack_int m, lapack_int n,
+                            lapack_complex_double* a, lapack_int lda,
+                            lapack_complex_double* b, lapack_int ldb,
+                            lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_stprfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_int l, const float* v,
+                           lapack_int ldv, const float* t, lapack_int ldt,
+                           float* a, lapack_int lda, float* b, lapack_int ldb,
+                           lapack_int myldwork );
+lapack_int LAPACKE_dtprfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_int l, const double* v,
+                           lapack_int ldv, const double* t, lapack_int ldt,
+                           double* a, lapack_int lda, double* b, lapack_int ldb,
+                           lapack_int myldwork );
+lapack_int LAPACKE_ctprfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_int l,
+                           const lapack_complex_float* v, lapack_int ldv,
+                           const lapack_complex_float* t, lapack_int ldt,
+                           lapack_complex_float* a, lapack_int lda,
+                           lapack_complex_float* b, lapack_int ldb,
+                           lapack_int myldwork );
+lapack_int LAPACKE_ztprfb( int matrix_order, char side, char trans, char direct,
+                           char storev, lapack_int m, lapack_int n,
+                           lapack_int k, lapack_int l,
+                           const lapack_complex_double* v, lapack_int ldv,
+                           const lapack_complex_double* t, lapack_int ldt,
+                           lapack_complex_double* a, lapack_int lda,
+                           lapack_complex_double* b, lapack_int ldb,
+                           lapack_int myldwork );
+
+lapack_int LAPACKE_sgemqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int nb, const float* v, lapack_int ldv,
+                                 const float* t, lapack_int ldt, float* c,
+                                 lapack_int ldc, float* work );
+lapack_int LAPACKE_dgemqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int nb, const double* v, lapack_int ldv,
+                                 const double* t, lapack_int ldt, double* c,
+                                 lapack_int ldc, double* work );
+lapack_int LAPACKE_cgemqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int nb, const lapack_complex_float* v,
+                                 lapack_int ldv, const lapack_complex_float* t,
+                                 lapack_int ldt, lapack_complex_float* c,
+                                 lapack_int ldc, lapack_complex_float* work );
+lapack_int LAPACKE_zgemqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int nb, const lapack_complex_double* v,
+                                 lapack_int ldv, const lapack_complex_double* t,
+                                 lapack_int ldt, lapack_complex_double* c,
+                                 lapack_int ldc, lapack_complex_double* work );
+
+lapack_int LAPACKE_sgeqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nb, float* a, lapack_int lda,
+                                float* t, lapack_int ldt, float* work );
+lapack_int LAPACKE_dgeqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nb, double* a, lapack_int lda,
+                                double* t, lapack_int ldt, double* work );
+lapack_int LAPACKE_cgeqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nb, lapack_complex_float* a,
+                                lapack_int lda, lapack_complex_float* t,
+                                lapack_int ldt, lapack_complex_float* work );
+lapack_int LAPACKE_zgeqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int nb, lapack_complex_double* a,
+                                lapack_int lda, lapack_complex_double* t,
+                                lapack_int ldt, lapack_complex_double* work );
+
+lapack_int LAPACKE_sgeqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 float* a, lapack_int lda, float* t,
+                                 lapack_int ldt );
+lapack_int LAPACKE_dgeqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 double* a, lapack_int lda, double* t,
+                                 lapack_int ldt );
+lapack_int LAPACKE_cgeqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zgeqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_sgeqrt3_work( int matrix_order, lapack_int m, lapack_int n,
+                                 float* a, lapack_int lda, float* t,
+                                 lapack_int ldt );
+lapack_int LAPACKE_dgeqrt3_work( int matrix_order, lapack_int m, lapack_int n,
+                                 double* a, lapack_int lda, double* t,
+                                 lapack_int ldt );
+lapack_int LAPACKE_cgeqrt3_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_zgeqrt3_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_stpmqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int l, lapack_int nb, const float* v,
+                                 lapack_int ldv, const float* t, lapack_int ldt,
+                                 float* a, lapack_int lda, float* b,
+                                 lapack_int ldb, float* work );
+lapack_int LAPACKE_dtpmqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int l, lapack_int nb, const double* v,
+                                 lapack_int ldv, const double* t,
+                                 lapack_int ldt, double* a, lapack_int lda,
+                                 double* b, lapack_int ldb, double* work );
+lapack_int LAPACKE_ctpmqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int l, lapack_int nb,
+                                 const lapack_complex_float* v, lapack_int ldv,
+                                 const lapack_complex_float* t, lapack_int ldt,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* work );
+lapack_int LAPACKE_ztpmqrt_work( int matrix_order, char side, char trans,
+                                 lapack_int m, lapack_int n, lapack_int k,
+                                 lapack_int l, lapack_int nb,
+                                 const lapack_complex_double* v, lapack_int ldv,
+                                 const lapack_complex_double* t, lapack_int ldt,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* work );
+
+lapack_int LAPACKE_dtpqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int l, lapack_int nb, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                double* t, lapack_int ldt, double* work );
+lapack_int LAPACKE_ctpqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int l, lapack_int nb,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* t,
+                                lapack_complex_float* b, lapack_int ldb,
+                                lapack_int ldt, lapack_complex_float* work );
+lapack_int LAPACKE_ztpqrt_work( int matrix_order, lapack_int m, lapack_int n,
+                                lapack_int l, lapack_int nb,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* work );
+
+lapack_int LAPACKE_stpqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 float* a, lapack_int lda, float* b,
+                                 lapack_int ldb, float* t, lapack_int ldt );
+lapack_int LAPACKE_dtpqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 double* a, lapack_int lda, double* b,
+                                 lapack_int ldb, double* t, lapack_int ldt );
+lapack_int LAPACKE_ctpqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_float* a, lapack_int lda,
+                                 lapack_complex_float* b, lapack_int ldb,
+                                 lapack_complex_float* t, lapack_int ldt );
+lapack_int LAPACKE_ztpqrt2_work( int matrix_order, lapack_int m, lapack_int n,
+                                 lapack_complex_double* a, lapack_int lda,
+                                 lapack_complex_double* b, lapack_int ldb,
+                                 lapack_complex_double* t, lapack_int ldt );
+
+lapack_int LAPACKE_stprfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                const float* v, lapack_int ldv, const float* t,
+                                lapack_int ldt, float* a, lapack_int lda,
+                                float* b, lapack_int ldb, const float* mywork,
+                                lapack_int myldwork );
+lapack_int LAPACKE_dtprfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                const double* v, lapack_int ldv,
+                                const double* t, lapack_int ldt, double* a,
+                                lapack_int lda, double* b, lapack_int ldb,
+                                const double* mywork, lapack_int myldwork );
+lapack_int LAPACKE_ctprfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                const lapack_complex_float* v, lapack_int ldv,
+                                const lapack_complex_float* t, lapack_int ldt,
+                                lapack_complex_float* a, lapack_int lda,
+                                lapack_complex_float* b, lapack_int ldb,
+                                const float* mywork, lapack_int myldwork );
+lapack_int LAPACKE_ztprfb_work( int matrix_order, char side, char trans,
+                                char direct, char storev, lapack_int m,
+                                lapack_int n, lapack_int k, lapack_int l,
+                                const lapack_complex_double* v, lapack_int ldv,
+                                const lapack_complex_double* t, lapack_int ldt,
+                                lapack_complex_double* a, lapack_int lda,
+                                lapack_complex_double* b, lapack_int ldb,
+                                const double* mywork, lapack_int myldwork );
+//LAPACK 3.X.X
+lapack_int LAPACKE_csyr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_float alpha,
+                             const lapack_complex_float* x, lapack_int incx,
+                             lapack_complex_float* a, lapack_int lda );
+lapack_int LAPACKE_zsyr( int matrix_order, char uplo, lapack_int n,
+                             lapack_complex_double alpha,
+                             const lapack_complex_double* x, lapack_int incx,
+                             lapack_complex_double* a, lapack_int lda );
+
+lapack_int LAPACKE_csyr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_float alpha,
+                                  const lapack_complex_float* x,
+                                  lapack_int incx, lapack_complex_float* a,
+                                  lapack_int lda );
+lapack_int LAPACKE_zsyr_work( int matrix_order, char uplo, lapack_int n,
+                                  lapack_complex_double alpha,
+                                  const lapack_complex_double* x,
+                                  lapack_int incx, lapack_complex_double* a,
+                                  lapack_int lda );
+
+
+
+#define LAPACK_sgetrf LAPACK_GLOBAL(sgetrf,SGETRF)
+#define LAPACK_dgetrf LAPACK_GLOBAL(dgetrf,DGETRF)
+#define LAPACK_cgetrf LAPACK_GLOBAL(cgetrf,CGETRF)
+#define LAPACK_zgetrf LAPACK_GLOBAL(zgetrf,ZGETRF)
+#define LAPACK_sgbtrf LAPACK_GLOBAL(sgbtrf,SGBTRF)
+#define LAPACK_dgbtrf LAPACK_GLOBAL(dgbtrf,DGBTRF)
+#define LAPACK_cgbtrf LAPACK_GLOBAL(cgbtrf,CGBTRF)
+#define LAPACK_zgbtrf LAPACK_GLOBAL(zgbtrf,ZGBTRF)
+#define LAPACK_sgttrf LAPACK_GLOBAL(sgttrf,SGTTRF)
+#define LAPACK_dgttrf LAPACK_GLOBAL(dgttrf,DGTTRF)
+#define LAPACK_cgttrf LAPACK_GLOBAL(cgttrf,CGTTRF)
+#define LAPACK_zgttrf LAPACK_GLOBAL(zgttrf,ZGTTRF)
+#define LAPACK_spotrf LAPACK_GLOBAL(spotrf,SPOTRF)
+#define LAPACK_dpotrf LAPACK_GLOBAL(dpotrf,DPOTRF)
+#define LAPACK_cpotrf LAPACK_GLOBAL(cpotrf,CPOTRF)
+#define LAPACK_zpotrf LAPACK_GLOBAL(zpotrf,ZPOTRF)
+#define LAPACK_dpstrf LAPACK_GLOBAL(dpstrf,DPSTRF)
+#define LAPACK_spstrf LAPACK_GLOBAL(spstrf,SPSTRF)
+#define LAPACK_zpstrf LAPACK_GLOBAL(zpstrf,ZPSTRF)
+#define LAPACK_cpstrf LAPACK_GLOBAL(cpstrf,CPSTRF)
+#define LAPACK_dpftrf LAPACK_GLOBAL(dpftrf,DPFTRF)
+#define LAPACK_spftrf LAPACK_GLOBAL(spftrf,SPFTRF)
+#define LAPACK_zpftrf LAPACK_GLOBAL(zpftrf,ZPFTRF)
+#define LAPACK_cpftrf LAPACK_GLOBAL(cpftrf,CPFTRF)
+#define LAPACK_spptrf LAPACK_GLOBAL(spptrf,SPPTRF)
+#define LAPACK_dpptrf LAPACK_GLOBAL(dpptrf,DPPTRF)
+#define LAPACK_cpptrf LAPACK_GLOBAL(cpptrf,CPPTRF)
+#define LAPACK_zpptrf LAPACK_GLOBAL(zpptrf,ZPPTRF)
+#define LAPACK_spbtrf LAPACK_GLOBAL(spbtrf,SPBTRF)
+#define LAPACK_dpbtrf LAPACK_GLOBAL(dpbtrf,DPBTRF)
+#define LAPACK_cpbtrf LAPACK_GLOBAL(cpbtrf,CPBTRF)
+#define LAPACK_zpbtrf LAPACK_GLOBAL(zpbtrf,ZPBTRF)
+#define LAPACK_spttrf LAPACK_GLOBAL(spttrf,SPTTRF)
+#define LAPACK_dpttrf LAPACK_GLOBAL(dpttrf,DPTTRF)
+#define LAPACK_cpttrf LAPACK_GLOBAL(cpttrf,CPTTRF)
+#define LAPACK_zpttrf LAPACK_GLOBAL(zpttrf,ZPTTRF)
+#define LAPACK_ssytrf LAPACK_GLOBAL(ssytrf,SSYTRF)
+#define LAPACK_dsytrf LAPACK_GLOBAL(dsytrf,DSYTRF)
+#define LAPACK_csytrf LAPACK_GLOBAL(csytrf,CSYTRF)
+#define LAPACK_zsytrf LAPACK_GLOBAL(zsytrf,ZSYTRF)
+#define LAPACK_chetrf LAPACK_GLOBAL(chetrf,CHETRF)
+#define LAPACK_zhetrf LAPACK_GLOBAL(zhetrf,ZHETRF)
+#define LAPACK_ssptrf LAPACK_GLOBAL(ssptrf,SSPTRF)
+#define LAPACK_dsptrf LAPACK_GLOBAL(dsptrf,DSPTRF)
+#define LAPACK_csptrf LAPACK_GLOBAL(csptrf,CSPTRF)
+#define LAPACK_zsptrf LAPACK_GLOBAL(zsptrf,ZSPTRF)
+#define LAPACK_chptrf LAPACK_GLOBAL(chptrf,CHPTRF)
+#define LAPACK_zhptrf LAPACK_GLOBAL(zhptrf,ZHPTRF)
+#define LAPACK_sgetrs LAPACK_GLOBAL(sgetrs,SGETRS)
+#define LAPACK_dgetrs LAPACK_GLOBAL(dgetrs,DGETRS)
+#define LAPACK_cgetrs LAPACK_GLOBAL(cgetrs,CGETRS)
+#define LAPACK_zgetrs LAPACK_GLOBAL(zgetrs,ZGETRS)
+#define LAPACK_sgbtrs LAPACK_GLOBAL(sgbtrs,SGBTRS)
+#define LAPACK_dgbtrs LAPACK_GLOBAL(dgbtrs,DGBTRS)
+#define LAPACK_cgbtrs LAPACK_GLOBAL(cgbtrs,CGBTRS)
+#define LAPACK_zgbtrs LAPACK_GLOBAL(zgbtrs,ZGBTRS)
+#define LAPACK_sgttrs LAPACK_GLOBAL(sgttrs,SGTTRS)
+#define LAPACK_dgttrs LAPACK_GLOBAL(dgttrs,DGTTRS)
+#define LAPACK_cgttrs LAPACK_GLOBAL(cgttrs,CGTTRS)
+#define LAPACK_zgttrs LAPACK_GLOBAL(zgttrs,ZGTTRS)
+#define LAPACK_spotrs LAPACK_GLOBAL(spotrs,SPOTRS)
+#define LAPACK_dpotrs LAPACK_GLOBAL(dpotrs,DPOTRS)
+#define LAPACK_cpotrs LAPACK_GLOBAL(cpotrs,CPOTRS)
+#define LAPACK_zpotrs LAPACK_GLOBAL(zpotrs,ZPOTRS)
+#define LAPACK_dpftrs LAPACK_GLOBAL(dpftrs,DPFTRS)
+#define LAPACK_spftrs LAPACK_GLOBAL(spftrs,SPFTRS)
+#define LAPACK_zpftrs LAPACK_GLOBAL(zpftrs,ZPFTRS)
+#define LAPACK_cpftrs LAPACK_GLOBAL(cpftrs,CPFTRS)
+#define LAPACK_spptrs LAPACK_GLOBAL(spptrs,SPPTRS)
+#define LAPACK_dpptrs LAPACK_GLOBAL(dpptrs,DPPTRS)
+#define LAPACK_cpptrs LAPACK_GLOBAL(cpptrs,CPPTRS)
+#define LAPACK_zpptrs LAPACK_GLOBAL(zpptrs,ZPPTRS)
+#define LAPACK_spbtrs LAPACK_GLOBAL(spbtrs,SPBTRS)
+#define LAPACK_dpbtrs LAPACK_GLOBAL(dpbtrs,DPBTRS)
+#define LAPACK_cpbtrs LAPACK_GLOBAL(cpbtrs,CPBTRS)
+#define LAPACK_zpbtrs LAPACK_GLOBAL(zpbtrs,ZPBTRS)
+#define LAPACK_spttrs LAPACK_GLOBAL(spttrs,SPTTRS)
+#define LAPACK_dpttrs LAPACK_GLOBAL(dpttrs,DPTTRS)
+#define LAPACK_cpttrs LAPACK_GLOBAL(cpttrs,CPTTRS)
+#define LAPACK_zpttrs LAPACK_GLOBAL(zpttrs,ZPTTRS)
+#define LAPACK_ssytrs LAPACK_GLOBAL(ssytrs,SSYTRS)
+#define LAPACK_dsytrs LAPACK_GLOBAL(dsytrs,DSYTRS)
+#define LAPACK_csytrs LAPACK_GLOBAL(csytrs,CSYTRS)
+#define LAPACK_zsytrs LAPACK_GLOBAL(zsytrs,ZSYTRS)
+#define LAPACK_chetrs LAPACK_GLOBAL(chetrs,CHETRS)
+#define LAPACK_zhetrs LAPACK_GLOBAL(zhetrs,ZHETRS)
+#define LAPACK_ssptrs LAPACK_GLOBAL(ssptrs,SSPTRS)
+#define LAPACK_dsptrs LAPACK_GLOBAL(dsptrs,DSPTRS)
+#define LAPACK_csptrs LAPACK_GLOBAL(csptrs,CSPTRS)
+#define LAPACK_zsptrs LAPACK_GLOBAL(zsptrs,ZSPTRS)
+#define LAPACK_chptrs LAPACK_GLOBAL(chptrs,CHPTRS)
+#define LAPACK_zhptrs LAPACK_GLOBAL(zhptrs,ZHPTRS)
+#define LAPACK_strtrs LAPACK_GLOBAL(strtrs,STRTRS)
+#define LAPACK_dtrtrs LAPACK_GLOBAL(dtrtrs,DTRTRS)
+#define LAPACK_ctrtrs LAPACK_GLOBAL(ctrtrs,CTRTRS)
+#define LAPACK_ztrtrs LAPACK_GLOBAL(ztrtrs,ZTRTRS)
+#define LAPACK_stptrs LAPACK_GLOBAL(stptrs,STPTRS)
+#define LAPACK_dtptrs LAPACK_GLOBAL(dtptrs,DTPTRS)
+#define LAPACK_ctptrs LAPACK_GLOBAL(ctptrs,CTPTRS)
+#define LAPACK_ztptrs LAPACK_GLOBAL(ztptrs,ZTPTRS)
+#define LAPACK_stbtrs LAPACK_GLOBAL(stbtrs,STBTRS)
+#define LAPACK_dtbtrs LAPACK_GLOBAL(dtbtrs,DTBTRS)
+#define LAPACK_ctbtrs LAPACK_GLOBAL(ctbtrs,CTBTRS)
+#define LAPACK_ztbtrs LAPACK_GLOBAL(ztbtrs,ZTBTRS)
+#define LAPACK_sgecon LAPACK_GLOBAL(sgecon,SGECON)
+#define LAPACK_dgecon LAPACK_GLOBAL(dgecon,DGECON)
+#define LAPACK_cgecon LAPACK_GLOBAL(cgecon,CGECON)
+#define LAPACK_zgecon LAPACK_GLOBAL(zgecon,ZGECON)
+#define LAPACK_sgbcon LAPACK_GLOBAL(sgbcon,SGBCON)
+#define LAPACK_dgbcon LAPACK_GLOBAL(dgbcon,DGBCON)
+#define LAPACK_cgbcon LAPACK_GLOBAL(cgbcon,CGBCON)
+#define LAPACK_zgbcon LAPACK_GLOBAL(zgbcon,ZGBCON)
+#define LAPACK_sgtcon LAPACK_GLOBAL(sgtcon,SGTCON)
+#define LAPACK_dgtcon LAPACK_GLOBAL(dgtcon,DGTCON)
+#define LAPACK_cgtcon LAPACK_GLOBAL(cgtcon,CGTCON)
+#define LAPACK_zgtcon LAPACK_GLOBAL(zgtcon,ZGTCON)
+#define LAPACK_spocon LAPACK_GLOBAL(spocon,SPOCON)
+#define LAPACK_dpocon LAPACK_GLOBAL(dpocon,DPOCON)
+#define LAPACK_cpocon LAPACK_GLOBAL(cpocon,CPOCON)
+#define LAPACK_zpocon LAPACK_GLOBAL(zpocon,ZPOCON)
+#define LAPACK_sppcon LAPACK_GLOBAL(sppcon,SPPCON)
+#define LAPACK_dppcon LAPACK_GLOBAL(dppcon,DPPCON)
+#define LAPACK_cppcon LAPACK_GLOBAL(cppcon,CPPCON)
+#define LAPACK_zppcon LAPACK_GLOBAL(zppcon,ZPPCON)
+#define LAPACK_spbcon LAPACK_GLOBAL(spbcon,SPBCON)
+#define LAPACK_dpbcon LAPACK_GLOBAL(dpbcon,DPBCON)
+#define LAPACK_cpbcon LAPACK_GLOBAL(cpbcon,CPBCON)
+#define LAPACK_zpbcon LAPACK_GLOBAL(zpbcon,ZPBCON)
+#define LAPACK_sptcon LAPACK_GLOBAL(sptcon,SPTCON)
+#define LAPACK_dptcon LAPACK_GLOBAL(dptcon,DPTCON)
+#define LAPACK_cptcon LAPACK_GLOBAL(cptcon,CPTCON)
+#define LAPACK_zptcon LAPACK_GLOBAL(zptcon,ZPTCON)
+#define LAPACK_ssycon LAPACK_GLOBAL(ssycon,SSYCON)
+#define LAPACK_dsycon LAPACK_GLOBAL(dsycon,DSYCON)
+#define LAPACK_csycon LAPACK_GLOBAL(csycon,CSYCON)
+#define LAPACK_zsycon LAPACK_GLOBAL(zsycon,ZSYCON)
+#define LAPACK_checon LAPACK_GLOBAL(checon,CHECON)
+#define LAPACK_zhecon LAPACK_GLOBAL(zhecon,ZHECON)
+#define LAPACK_sspcon LAPACK_GLOBAL(sspcon,SSPCON)
+#define LAPACK_dspcon LAPACK_GLOBAL(dspcon,DSPCON)
+#define LAPACK_cspcon LAPACK_GLOBAL(cspcon,CSPCON)
+#define LAPACK_zspcon LAPACK_GLOBAL(zspcon,ZSPCON)
+#define LAPACK_chpcon LAPACK_GLOBAL(chpcon,CHPCON)
+#define LAPACK_zhpcon LAPACK_GLOBAL(zhpcon,ZHPCON)
+#define LAPACK_strcon LAPACK_GLOBAL(strcon,STRCON)
+#define LAPACK_dtrcon LAPACK_GLOBAL(dtrcon,DTRCON)
+#define LAPACK_ctrcon LAPACK_GLOBAL(ctrcon,CTRCON)
+#define LAPACK_ztrcon LAPACK_GLOBAL(ztrcon,ZTRCON)
+#define LAPACK_stpcon LAPACK_GLOBAL(stpcon,STPCON)
+#define LAPACK_dtpcon LAPACK_GLOBAL(dtpcon,DTPCON)
+#define LAPACK_ctpcon LAPACK_GLOBAL(ctpcon,CTPCON)
+#define LAPACK_ztpcon LAPACK_GLOBAL(ztpcon,ZTPCON)
+#define LAPACK_stbcon LAPACK_GLOBAL(stbcon,STBCON)
+#define LAPACK_dtbcon LAPACK_GLOBAL(dtbcon,DTBCON)
+#define LAPACK_ctbcon LAPACK_GLOBAL(ctbcon,CTBCON)
+#define LAPACK_ztbcon LAPACK_GLOBAL(ztbcon,ZTBCON)
+#define LAPACK_sgerfs LAPACK_GLOBAL(sgerfs,SGERFS)
+#define LAPACK_dgerfs LAPACK_GLOBAL(dgerfs,DGERFS)
+#define LAPACK_cgerfs LAPACK_GLOBAL(cgerfs,CGERFS)
+#define LAPACK_zgerfs LAPACK_GLOBAL(zgerfs,ZGERFS)
+#define LAPACK_dgerfsx LAPACK_GLOBAL(dgerfsx,DGERFSX)
+#define LAPACK_sgerfsx LAPACK_GLOBAL(sgerfsx,SGERFSX)
+#define LAPACK_zgerfsx LAPACK_GLOBAL(zgerfsx,ZGERFSX)
+#define LAPACK_cgerfsx LAPACK_GLOBAL(cgerfsx,CGERFSX)
+#define LAPACK_sgbrfs LAPACK_GLOBAL(sgbrfs,SGBRFS)
+#define LAPACK_dgbrfs LAPACK_GLOBAL(dgbrfs,DGBRFS)
+#define LAPACK_cgbrfs LAPACK_GLOBAL(cgbrfs,CGBRFS)
+#define LAPACK_zgbrfs LAPACK_GLOBAL(zgbrfs,ZGBRFS)
+#define LAPACK_dgbrfsx LAPACK_GLOBAL(dgbrfsx,DGBRFSX)
+#define LAPACK_sgbrfsx LAPACK_GLOBAL(sgbrfsx,SGBRFSX)
+#define LAPACK_zgbrfsx LAPACK_GLOBAL(zgbrfsx,ZGBRFSX)
+#define LAPACK_cgbrfsx LAPACK_GLOBAL(cgbrfsx,CGBRFSX)
+#define LAPACK_sgtrfs LAPACK_GLOBAL(sgtrfs,SGTRFS)
+#define LAPACK_dgtrfs LAPACK_GLOBAL(dgtrfs,DGTRFS)
+#define LAPACK_cgtrfs LAPACK_GLOBAL(cgtrfs,CGTRFS)
+#define LAPACK_zgtrfs LAPACK_GLOBAL(zgtrfs,ZGTRFS)
+#define LAPACK_sporfs LAPACK_GLOBAL(sporfs,SPORFS)
+#define LAPACK_dporfs LAPACK_GLOBAL(dporfs,DPORFS)
+#define LAPACK_cporfs LAPACK_GLOBAL(cporfs,CPORFS)
+#define LAPACK_zporfs LAPACK_GLOBAL(zporfs,ZPORFS)
+#define LAPACK_dporfsx LAPACK_GLOBAL(dporfsx,DPORFSX)
+#define LAPACK_sporfsx LAPACK_GLOBAL(sporfsx,SPORFSX)
+#define LAPACK_zporfsx LAPACK_GLOBAL(zporfsx,ZPORFSX)
+#define LAPACK_cporfsx LAPACK_GLOBAL(cporfsx,CPORFSX)
+#define LAPACK_spprfs LAPACK_GLOBAL(spprfs,SPPRFS)
+#define LAPACK_dpprfs LAPACK_GLOBAL(dpprfs,DPPRFS)
+#define LAPACK_cpprfs LAPACK_GLOBAL(cpprfs,CPPRFS)
+#define LAPACK_zpprfs LAPACK_GLOBAL(zpprfs,ZPPRFS)
+#define LAPACK_spbrfs LAPACK_GLOBAL(spbrfs,SPBRFS)
+#define LAPACK_dpbrfs LAPACK_GLOBAL(dpbrfs,DPBRFS)
+#define LAPACK_cpbrfs LAPACK_GLOBAL(cpbrfs,CPBRFS)
+#define LAPACK_zpbrfs LAPACK_GLOBAL(zpbrfs,ZPBRFS)
+#define LAPACK_sptrfs LAPACK_GLOBAL(sptrfs,SPTRFS)
+#define LAPACK_dptrfs LAPACK_GLOBAL(dptrfs,DPTRFS)
+#define LAPACK_cptrfs LAPACK_GLOBAL(cptrfs,CPTRFS)
+#define LAPACK_zptrfs LAPACK_GLOBAL(zptrfs,ZPTRFS)
+#define LAPACK_ssyrfs LAPACK_GLOBAL(ssyrfs,SSYRFS)
+#define LAPACK_dsyrfs LAPACK_GLOBAL(dsyrfs,DSYRFS)
+#define LAPACK_csyrfs LAPACK_GLOBAL(csyrfs,CSYRFS)
+#define LAPACK_zsyrfs LAPACK_GLOBAL(zsyrfs,ZSYRFS)
+#define LAPACK_dsyrfsx LAPACK_GLOBAL(dsyrfsx,DSYRFSX)
+#define LAPACK_ssyrfsx LAPACK_GLOBAL(ssyrfsx,SSYRFSX)
+#define LAPACK_zsyrfsx LAPACK_GLOBAL(zsyrfsx,ZSYRFSX)
+#define LAPACK_csyrfsx LAPACK_GLOBAL(csyrfsx,CSYRFSX)
+#define LAPACK_cherfs LAPACK_GLOBAL(cherfs,CHERFS)
+#define LAPACK_zherfs LAPACK_GLOBAL(zherfs,ZHERFS)
+#define LAPACK_zherfsx LAPACK_GLOBAL(zherfsx,ZHERFSX)
+#define LAPACK_cherfsx LAPACK_GLOBAL(cherfsx,CHERFSX)
+#define LAPACK_ssprfs LAPACK_GLOBAL(ssprfs,SSPRFS)
+#define LAPACK_dsprfs LAPACK_GLOBAL(dsprfs,DSPRFS)
+#define LAPACK_csprfs LAPACK_GLOBAL(csprfs,CSPRFS)
+#define LAPACK_zsprfs LAPACK_GLOBAL(zsprfs,ZSPRFS)
+#define LAPACK_chprfs LAPACK_GLOBAL(chprfs,CHPRFS)
+#define LAPACK_zhprfs LAPACK_GLOBAL(zhprfs,ZHPRFS)
+#define LAPACK_strrfs LAPACK_GLOBAL(strrfs,STRRFS)
+#define LAPACK_dtrrfs LAPACK_GLOBAL(dtrrfs,DTRRFS)
+#define LAPACK_ctrrfs LAPACK_GLOBAL(ctrrfs,CTRRFS)
+#define LAPACK_ztrrfs LAPACK_GLOBAL(ztrrfs,ZTRRFS)
+#define LAPACK_stprfs LAPACK_GLOBAL(stprfs,STPRFS)
+#define LAPACK_dtprfs LAPACK_GLOBAL(dtprfs,DTPRFS)
+#define LAPACK_ctprfs LAPACK_GLOBAL(ctprfs,CTPRFS)
+#define LAPACK_ztprfs LAPACK_GLOBAL(ztprfs,ZTPRFS)
+#define LAPACK_stbrfs LAPACK_GLOBAL(stbrfs,STBRFS)
+#define LAPACK_dtbrfs LAPACK_GLOBAL(dtbrfs,DTBRFS)
+#define LAPACK_ctbrfs LAPACK_GLOBAL(ctbrfs,CTBRFS)
+#define LAPACK_ztbrfs LAPACK_GLOBAL(ztbrfs,ZTBRFS)
+#define LAPACK_sgetri LAPACK_GLOBAL(sgetri,SGETRI)
+#define LAPACK_dgetri LAPACK_GLOBAL(dgetri,DGETRI)
+#define LAPACK_cgetri LAPACK_GLOBAL(cgetri,CGETRI)
+#define LAPACK_zgetri LAPACK_GLOBAL(zgetri,ZGETRI)
+#define LAPACK_spotri LAPACK_GLOBAL(spotri,SPOTRI)
+#define LAPACK_dpotri LAPACK_GLOBAL(dpotri,DPOTRI)
+#define LAPACK_cpotri LAPACK_GLOBAL(cpotri,CPOTRI)
+#define LAPACK_zpotri LAPACK_GLOBAL(zpotri,ZPOTRI)
+#define LAPACK_dpftri LAPACK_GLOBAL(dpftri,DPFTRI)
+#define LAPACK_spftri LAPACK_GLOBAL(spftri,SPFTRI)
+#define LAPACK_zpftri LAPACK_GLOBAL(zpftri,ZPFTRI)
+#define LAPACK_cpftri LAPACK_GLOBAL(cpftri,CPFTRI)
+#define LAPACK_spptri LAPACK_GLOBAL(spptri,SPPTRI)
+#define LAPACK_dpptri LAPACK_GLOBAL(dpptri,DPPTRI)
+#define LAPACK_cpptri LAPACK_GLOBAL(cpptri,CPPTRI)
+#define LAPACK_zpptri LAPACK_GLOBAL(zpptri,ZPPTRI)
+#define LAPACK_ssytri LAPACK_GLOBAL(ssytri,SSYTRI)
+#define LAPACK_dsytri LAPACK_GLOBAL(dsytri,DSYTRI)
+#define LAPACK_csytri LAPACK_GLOBAL(csytri,CSYTRI)
+#define LAPACK_zsytri LAPACK_GLOBAL(zsytri,ZSYTRI)
+#define LAPACK_chetri LAPACK_GLOBAL(chetri,CHETRI)
+#define LAPACK_zhetri LAPACK_GLOBAL(zhetri,ZHETRI)
+#define LAPACK_ssptri LAPACK_GLOBAL(ssptri,SSPTRI)
+#define LAPACK_dsptri LAPACK_GLOBAL(dsptri,DSPTRI)
+#define LAPACK_csptri LAPACK_GLOBAL(csptri,CSPTRI)
+#define LAPACK_zsptri LAPACK_GLOBAL(zsptri,ZSPTRI)
+#define LAPACK_chptri LAPACK_GLOBAL(chptri,CHPTRI)
+#define LAPACK_zhptri LAPACK_GLOBAL(zhptri,ZHPTRI)
+#define LAPACK_strtri LAPACK_GLOBAL(strtri,STRTRI)
+#define LAPACK_dtrtri LAPACK_GLOBAL(dtrtri,DTRTRI)
+#define LAPACK_ctrtri LAPACK_GLOBAL(ctrtri,CTRTRI)
+#define LAPACK_ztrtri LAPACK_GLOBAL(ztrtri,ZTRTRI)
+#define LAPACK_dtftri LAPACK_GLOBAL(dtftri,DTFTRI)
+#define LAPACK_stftri LAPACK_GLOBAL(stftri,STFTRI)
+#define LAPACK_ztftri LAPACK_GLOBAL(ztftri,ZTFTRI)
+#define LAPACK_ctftri LAPACK_GLOBAL(ctftri,CTFTRI)
+#define LAPACK_stptri LAPACK_GLOBAL(stptri,STPTRI)
+#define LAPACK_dtptri LAPACK_GLOBAL(dtptri,DTPTRI)
+#define LAPACK_ctptri LAPACK_GLOBAL(ctptri,CTPTRI)
+#define LAPACK_ztptri LAPACK_GLOBAL(ztptri,ZTPTRI)
+#define LAPACK_sgeequ LAPACK_GLOBAL(sgeequ,SGEEQU)
+#define LAPACK_dgeequ LAPACK_GLOBAL(dgeequ,DGEEQU)
+#define LAPACK_cgeequ LAPACK_GLOBAL(cgeequ,CGEEQU)
+#define LAPACK_zgeequ LAPACK_GLOBAL(zgeequ,ZGEEQU)
+#define LAPACK_dgeequb LAPACK_GLOBAL(dgeequb,DGEEQUB)
+#define LAPACK_sgeequb LAPACK_GLOBAL(sgeequb,SGEEQUB)
+#define LAPACK_zgeequb LAPACK_GLOBAL(zgeequb,ZGEEQUB)
+#define LAPACK_cgeequb LAPACK_GLOBAL(cgeequb,CGEEQUB)
+#define LAPACK_sgbequ LAPACK_GLOBAL(sgbequ,SGBEQU)
+#define LAPACK_dgbequ LAPACK_GLOBAL(dgbequ,DGBEQU)
+#define LAPACK_cgbequ LAPACK_GLOBAL(cgbequ,CGBEQU)
+#define LAPACK_zgbequ LAPACK_GLOBAL(zgbequ,ZGBEQU)
+#define LAPACK_dgbequb LAPACK_GLOBAL(dgbequb,DGBEQUB)
+#define LAPACK_sgbequb LAPACK_GLOBAL(sgbequb,SGBEQUB)
+#define LAPACK_zgbequb LAPACK_GLOBAL(zgbequb,ZGBEQUB)
+#define LAPACK_cgbequb LAPACK_GLOBAL(cgbequb,CGBEQUB)
+#define LAPACK_spoequ LAPACK_GLOBAL(spoequ,SPOEQU)
+#define LAPACK_dpoequ LAPACK_GLOBAL(dpoequ,DPOEQU)
+#define LAPACK_cpoequ LAPACK_GLOBAL(cpoequ,CPOEQU)
+#define LAPACK_zpoequ LAPACK_GLOBAL(zpoequ,ZPOEQU)
+#define LAPACK_dpoequb LAPACK_GLOBAL(dpoequb,DPOEQUB)
+#define LAPACK_spoequb LAPACK_GLOBAL(spoequb,SPOEQUB)
+#define LAPACK_zpoequb LAPACK_GLOBAL(zpoequb,ZPOEQUB)
+#define LAPACK_cpoequb LAPACK_GLOBAL(cpoequb,CPOEQUB)
+#define LAPACK_sppequ LAPACK_GLOBAL(sppequ,SPPEQU)
+#define LAPACK_dppequ LAPACK_GLOBAL(dppequ,DPPEQU)
+#define LAPACK_cppequ LAPACK_GLOBAL(cppequ,CPPEQU)
+#define LAPACK_zppequ LAPACK_GLOBAL(zppequ,ZPPEQU)
+#define LAPACK_spbequ LAPACK_GLOBAL(spbequ,SPBEQU)
+#define LAPACK_dpbequ LAPACK_GLOBAL(dpbequ,DPBEQU)
+#define LAPACK_cpbequ LAPACK_GLOBAL(cpbequ,CPBEQU)
+#define LAPACK_zpbequ LAPACK_GLOBAL(zpbequ,ZPBEQU)
+#define LAPACK_dsyequb LAPACK_GLOBAL(dsyequb,DSYEQUB)
+#define LAPACK_ssyequb LAPACK_GLOBAL(ssyequb,SSYEQUB)
+#define LAPACK_zsyequb LAPACK_GLOBAL(zsyequb,ZSYEQUB)
+#define LAPACK_csyequb LAPACK_GLOBAL(csyequb,CSYEQUB)
+#define LAPACK_zheequb LAPACK_GLOBAL(zheequb,ZHEEQUB)
+#define LAPACK_cheequb LAPACK_GLOBAL(cheequb,CHEEQUB)
+#define LAPACK_sgesv LAPACK_GLOBAL(sgesv,SGESV)
+#define LAPACK_dgesv LAPACK_GLOBAL(dgesv,DGESV)
+#define LAPACK_cgesv LAPACK_GLOBAL(cgesv,CGESV)
+#define LAPACK_zgesv LAPACK_GLOBAL(zgesv,ZGESV)
+#define LAPACK_dsgesv LAPACK_GLOBAL(dsgesv,DSGESV)
+#define LAPACK_zcgesv LAPACK_GLOBAL(zcgesv,ZCGESV)
+#define LAPACK_sgesvx LAPACK_GLOBAL(sgesvx,SGESVX)
+#define LAPACK_dgesvx LAPACK_GLOBAL(dgesvx,DGESVX)
+#define LAPACK_cgesvx LAPACK_GLOBAL(cgesvx,CGESVX)
+#define LAPACK_zgesvx LAPACK_GLOBAL(zgesvx,ZGESVX)
+#define LAPACK_dgesvxx LAPACK_GLOBAL(dgesvxx,DGESVXX)
+#define LAPACK_sgesvxx LAPACK_GLOBAL(sgesvxx,SGESVXX)
+#define LAPACK_zgesvxx LAPACK_GLOBAL(zgesvxx,ZGESVXX)
+#define LAPACK_cgesvxx LAPACK_GLOBAL(cgesvxx,CGESVXX)
+#define LAPACK_sgbsv LAPACK_GLOBAL(sgbsv,SGBSV)
+#define LAPACK_dgbsv LAPACK_GLOBAL(dgbsv,DGBSV)
+#define LAPACK_cgbsv LAPACK_GLOBAL(cgbsv,CGBSV)
+#define LAPACK_zgbsv LAPACK_GLOBAL(zgbsv,ZGBSV)
+#define LAPACK_sgbsvx LAPACK_GLOBAL(sgbsvx,SGBSVX)
+#define LAPACK_dgbsvx LAPACK_GLOBAL(dgbsvx,DGBSVX)
+#define LAPACK_cgbsvx LAPACK_GLOBAL(cgbsvx,CGBSVX)
+#define LAPACK_zgbsvx LAPACK_GLOBAL(zgbsvx,ZGBSVX)
+#define LAPACK_dgbsvxx LAPACK_GLOBAL(dgbsvxx,DGBSVXX)
+#define LAPACK_sgbsvxx LAPACK_GLOBAL(sgbsvxx,SGBSVXX)
+#define LAPACK_zgbsvxx LAPACK_GLOBAL(zgbsvxx,ZGBSVXX)
+#define LAPACK_cgbsvxx LAPACK_GLOBAL(cgbsvxx,CGBSVXX)
+#define LAPACK_sgtsv LAPACK_GLOBAL(sgtsv,SGTSV)
+#define LAPACK_dgtsv LAPACK_GLOBAL(dgtsv,DGTSV)
+#define LAPACK_cgtsv LAPACK_GLOBAL(cgtsv,CGTSV)
+#define LAPACK_zgtsv LAPACK_GLOBAL(zgtsv,ZGTSV)
+#define LAPACK_sgtsvx LAPACK_GLOBAL(sgtsvx,SGTSVX)
+#define LAPACK_dgtsvx LAPACK_GLOBAL(dgtsvx,DGTSVX)
+#define LAPACK_cgtsvx LAPACK_GLOBAL(cgtsvx,CGTSVX)
+#define LAPACK_zgtsvx LAPACK_GLOBAL(zgtsvx,ZGTSVX)
+#define LAPACK_sposv LAPACK_GLOBAL(sposv,SPOSV)
+#define LAPACK_dposv LAPACK_GLOBAL(dposv,DPOSV)
+#define LAPACK_cposv LAPACK_GLOBAL(cposv,CPOSV)
+#define LAPACK_zposv LAPACK_GLOBAL(zposv,ZPOSV)
+#define LAPACK_dsposv LAPACK_GLOBAL(dsposv,DSPOSV)
+#define LAPACK_zcposv LAPACK_GLOBAL(zcposv,ZCPOSV)
+#define LAPACK_sposvx LAPACK_GLOBAL(sposvx,SPOSVX)
+#define LAPACK_dposvx LAPACK_GLOBAL(dposvx,DPOSVX)
+#define LAPACK_cposvx LAPACK_GLOBAL(cposvx,CPOSVX)
+#define LAPACK_zposvx LAPACK_GLOBAL(zposvx,ZPOSVX)
+#define LAPACK_dposvxx LAPACK_GLOBAL(dposvxx,DPOSVXX)
+#define LAPACK_sposvxx LAPACK_GLOBAL(sposvxx,SPOSVXX)
+#define LAPACK_zposvxx LAPACK_GLOBAL(zposvxx,ZPOSVXX)
+#define LAPACK_cposvxx LAPACK_GLOBAL(cposvxx,CPOSVXX)
+#define LAPACK_sppsv LAPACK_GLOBAL(sppsv,SPPSV)
+#define LAPACK_dppsv LAPACK_GLOBAL(dppsv,DPPSV)
+#define LAPACK_cppsv LAPACK_GLOBAL(cppsv,CPPSV)
+#define LAPACK_zppsv LAPACK_GLOBAL(zppsv,ZPPSV)
+#define LAPACK_sppsvx LAPACK_GLOBAL(sppsvx,SPPSVX)
+#define LAPACK_dppsvx LAPACK_GLOBAL(dppsvx,DPPSVX)
+#define LAPACK_cppsvx LAPACK_GLOBAL(cppsvx,CPPSVX)
+#define LAPACK_zppsvx LAPACK_GLOBAL(zppsvx,ZPPSVX)
+#define LAPACK_spbsv LAPACK_GLOBAL(spbsv,SPBSV)
+#define LAPACK_dpbsv LAPACK_GLOBAL(dpbsv,DPBSV)
+#define LAPACK_cpbsv LAPACK_GLOBAL(cpbsv,CPBSV)
+#define LAPACK_zpbsv LAPACK_GLOBAL(zpbsv,ZPBSV)
+#define LAPACK_spbsvx LAPACK_GLOBAL(spbsvx,SPBSVX)
+#define LAPACK_dpbsvx LAPACK_GLOBAL(dpbsvx,DPBSVX)
+#define LAPACK_cpbsvx LAPACK_GLOBAL(cpbsvx,CPBSVX)
+#define LAPACK_zpbsvx LAPACK_GLOBAL(zpbsvx,ZPBSVX)
+#define LAPACK_sptsv LAPACK_GLOBAL(sptsv,SPTSV)
+#define LAPACK_dptsv LAPACK_GLOBAL(dptsv,DPTSV)
+#define LAPACK_cptsv LAPACK_GLOBAL(cptsv,CPTSV)
+#define LAPACK_zptsv LAPACK_GLOBAL(zptsv,ZPTSV)
+#define LAPACK_sptsvx LAPACK_GLOBAL(sptsvx,SPTSVX)
+#define LAPACK_dptsvx LAPACK_GLOBAL(dptsvx,DPTSVX)
+#define LAPACK_cptsvx LAPACK_GLOBAL(cptsvx,CPTSVX)
+#define LAPACK_zptsvx LAPACK_GLOBAL(zptsvx,ZPTSVX)
+#define LAPACK_ssysv LAPACK_GLOBAL(ssysv,SSYSV)
+#define LAPACK_dsysv LAPACK_GLOBAL(dsysv,DSYSV)
+#define LAPACK_csysv LAPACK_GLOBAL(csysv,CSYSV)
+#define LAPACK_zsysv LAPACK_GLOBAL(zsysv,ZSYSV)
+#define LAPACK_ssysvx LAPACK_GLOBAL(ssysvx,SSYSVX)
+#define LAPACK_dsysvx LAPACK_GLOBAL(dsysvx,DSYSVX)
+#define LAPACK_csysvx LAPACK_GLOBAL(csysvx,CSYSVX)
+#define LAPACK_zsysvx LAPACK_GLOBAL(zsysvx,ZSYSVX)
+#define LAPACK_dsysvxx LAPACK_GLOBAL(dsysvxx,DSYSVXX)
+#define LAPACK_ssysvxx LAPACK_GLOBAL(ssysvxx,SSYSVXX)
+#define LAPACK_zsysvxx LAPACK_GLOBAL(zsysvxx,ZSYSVXX)
+#define LAPACK_csysvxx LAPACK_GLOBAL(csysvxx,CSYSVXX)
+#define LAPACK_chesv LAPACK_GLOBAL(chesv,CHESV)
+#define LAPACK_zhesv LAPACK_GLOBAL(zhesv,ZHESV)
+#define LAPACK_chesvx LAPACK_GLOBAL(chesvx,CHESVX)
+#define LAPACK_zhesvx LAPACK_GLOBAL(zhesvx,ZHESVX)
+#define LAPACK_zhesvxx LAPACK_GLOBAL(zhesvxx,ZHESVXX)
+#define LAPACK_chesvxx LAPACK_GLOBAL(chesvxx,CHESVXX)
+#define LAPACK_sspsv LAPACK_GLOBAL(sspsv,SSPSV)
+#define LAPACK_dspsv LAPACK_GLOBAL(dspsv,DSPSV)
+#define LAPACK_cspsv LAPACK_GLOBAL(cspsv,CSPSV)
+#define LAPACK_zspsv LAPACK_GLOBAL(zspsv,ZSPSV)
+#define LAPACK_sspsvx LAPACK_GLOBAL(sspsvx,SSPSVX)
+#define LAPACK_dspsvx LAPACK_GLOBAL(dspsvx,DSPSVX)
+#define LAPACK_cspsvx LAPACK_GLOBAL(cspsvx,CSPSVX)
+#define LAPACK_zspsvx LAPACK_GLOBAL(zspsvx,ZSPSVX)
+#define LAPACK_chpsv LAPACK_GLOBAL(chpsv,CHPSV)
+#define LAPACK_zhpsv LAPACK_GLOBAL(zhpsv,ZHPSV)
+#define LAPACK_chpsvx LAPACK_GLOBAL(chpsvx,CHPSVX)
+#define LAPACK_zhpsvx LAPACK_GLOBAL(zhpsvx,ZHPSVX)
+#define LAPACK_sgeqrf LAPACK_GLOBAL(sgeqrf,SGEQRF)
+#define LAPACK_dgeqrf LAPACK_GLOBAL(dgeqrf,DGEQRF)
+#define LAPACK_cgeqrf LAPACK_GLOBAL(cgeqrf,CGEQRF)
+#define LAPACK_zgeqrf LAPACK_GLOBAL(zgeqrf,ZGEQRF)
+#define LAPACK_sgeqpf LAPACK_GLOBAL(sgeqpf,SGEQPF)
+#define LAPACK_dgeqpf LAPACK_GLOBAL(dgeqpf,DGEQPF)
+#define LAPACK_cgeqpf LAPACK_GLOBAL(cgeqpf,CGEQPF)
+#define LAPACK_zgeqpf LAPACK_GLOBAL(zgeqpf,ZGEQPF)
+#define LAPACK_sgeqp3 LAPACK_GLOBAL(sgeqp3,SGEQP3)
+#define LAPACK_dgeqp3 LAPACK_GLOBAL(dgeqp3,DGEQP3)
+#define LAPACK_cgeqp3 LAPACK_GLOBAL(cgeqp3,CGEQP3)
+#define LAPACK_zgeqp3 LAPACK_GLOBAL(zgeqp3,ZGEQP3)
+#define LAPACK_sorgqr LAPACK_GLOBAL(sorgqr,SORGQR)
+#define LAPACK_dorgqr LAPACK_GLOBAL(dorgqr,DORGQR)
+#define LAPACK_sormqr LAPACK_GLOBAL(sormqr,SORMQR)
+#define LAPACK_dormqr LAPACK_GLOBAL(dormqr,DORMQR)
+#define LAPACK_cungqr LAPACK_GLOBAL(cungqr,CUNGQR)
+#define LAPACK_zungqr LAPACK_GLOBAL(zungqr,ZUNGQR)
+#define LAPACK_cunmqr LAPACK_GLOBAL(cunmqr,CUNMQR)
+#define LAPACK_zunmqr LAPACK_GLOBAL(zunmqr,ZUNMQR)
+#define LAPACK_sgelqf LAPACK_GLOBAL(sgelqf,SGELQF)
+#define LAPACK_dgelqf LAPACK_GLOBAL(dgelqf,DGELQF)
+#define LAPACK_cgelqf LAPACK_GLOBAL(cgelqf,CGELQF)
+#define LAPACK_zgelqf LAPACK_GLOBAL(zgelqf,ZGELQF)
+#define LAPACK_sorglq LAPACK_GLOBAL(sorglq,SORGLQ)
+#define LAPACK_dorglq LAPACK_GLOBAL(dorglq,DORGLQ)
+#define LAPACK_sormlq LAPACK_GLOBAL(sormlq,SORMLQ)
+#define LAPACK_dormlq LAPACK_GLOBAL(dormlq,DORMLQ)
+#define LAPACK_cunglq LAPACK_GLOBAL(cunglq,CUNGLQ)
+#define LAPACK_zunglq LAPACK_GLOBAL(zunglq,ZUNGLQ)
+#define LAPACK_cunmlq LAPACK_GLOBAL(cunmlq,CUNMLQ)
+#define LAPACK_zunmlq LAPACK_GLOBAL(zunmlq,ZUNMLQ)
+#define LAPACK_sgeqlf LAPACK_GLOBAL(sgeqlf,SGEQLF)
+#define LAPACK_dgeqlf LAPACK_GLOBAL(dgeqlf,DGEQLF)
+#define LAPACK_cgeqlf LAPACK_GLOBAL(cgeqlf,CGEQLF)
+#define LAPACK_zgeqlf LAPACK_GLOBAL(zgeqlf,ZGEQLF)
+#define LAPACK_sorgql LAPACK_GLOBAL(sorgql,SORGQL)
+#define LAPACK_dorgql LAPACK_GLOBAL(dorgql,DORGQL)
+#define LAPACK_cungql LAPACK_GLOBAL(cungql,CUNGQL)
+#define LAPACK_zungql LAPACK_GLOBAL(zungql,ZUNGQL)
+#define LAPACK_sormql LAPACK_GLOBAL(sormql,SORMQL)
+#define LAPACK_dormql LAPACK_GLOBAL(dormql,DORMQL)
+#define LAPACK_cunmql LAPACK_GLOBAL(cunmql,CUNMQL)
+#define LAPACK_zunmql LAPACK_GLOBAL(zunmql,ZUNMQL)
+#define LAPACK_sgerqf LAPACK_GLOBAL(sgerqf,SGERQF)
+#define LAPACK_dgerqf LAPACK_GLOBAL(dgerqf,DGERQF)
+#define LAPACK_cgerqf LAPACK_GLOBAL(cgerqf,CGERQF)
+#define LAPACK_zgerqf LAPACK_GLOBAL(zgerqf,ZGERQF)
+#define LAPACK_sorgrq LAPACK_GLOBAL(sorgrq,SORGRQ)
+#define LAPACK_dorgrq LAPACK_GLOBAL(dorgrq,DORGRQ)
+#define LAPACK_cungrq LAPACK_GLOBAL(cungrq,CUNGRQ)
+#define LAPACK_zungrq LAPACK_GLOBAL(zungrq,ZUNGRQ)
+#define LAPACK_sormrq LAPACK_GLOBAL(sormrq,SORMRQ)
+#define LAPACK_dormrq LAPACK_GLOBAL(dormrq,DORMRQ)
+#define LAPACK_cunmrq LAPACK_GLOBAL(cunmrq,CUNMRQ)
+#define LAPACK_zunmrq LAPACK_GLOBAL(zunmrq,ZUNMRQ)
+#define LAPACK_stzrzf LAPACK_GLOBAL(stzrzf,STZRZF)
+#define LAPACK_dtzrzf LAPACK_GLOBAL(dtzrzf,DTZRZF)
+#define LAPACK_ctzrzf LAPACK_GLOBAL(ctzrzf,CTZRZF)
+#define LAPACK_ztzrzf LAPACK_GLOBAL(ztzrzf,ZTZRZF)
+#define LAPACK_sormrz LAPACK_GLOBAL(sormrz,SORMRZ)
+#define LAPACK_dormrz LAPACK_GLOBAL(dormrz,DORMRZ)
+#define LAPACK_cunmrz LAPACK_GLOBAL(cunmrz,CUNMRZ)
+#define LAPACK_zunmrz LAPACK_GLOBAL(zunmrz,ZUNMRZ)
+#define LAPACK_sggqrf LAPACK_GLOBAL(sggqrf,SGGQRF)
+#define LAPACK_dggqrf LAPACK_GLOBAL(dggqrf,DGGQRF)
+#define LAPACK_cggqrf LAPACK_GLOBAL(cggqrf,CGGQRF)
+#define LAPACK_zggqrf LAPACK_GLOBAL(zggqrf,ZGGQRF)
+#define LAPACK_sggrqf LAPACK_GLOBAL(sggrqf,SGGRQF)
+#define LAPACK_dggrqf LAPACK_GLOBAL(dggrqf,DGGRQF)
+#define LAPACK_cggrqf LAPACK_GLOBAL(cggrqf,CGGRQF)
+#define LAPACK_zggrqf LAPACK_GLOBAL(zggrqf,ZGGRQF)
+#define LAPACK_sgebrd LAPACK_GLOBAL(sgebrd,SGEBRD)
+#define LAPACK_dgebrd LAPACK_GLOBAL(dgebrd,DGEBRD)
+#define LAPACK_cgebrd LAPACK_GLOBAL(cgebrd,CGEBRD)
+#define LAPACK_zgebrd LAPACK_GLOBAL(zgebrd,ZGEBRD)
+#define LAPACK_sgbbrd LAPACK_GLOBAL(sgbbrd,SGBBRD)
+#define LAPACK_dgbbrd LAPACK_GLOBAL(dgbbrd,DGBBRD)
+#define LAPACK_cgbbrd LAPACK_GLOBAL(cgbbrd,CGBBRD)
+#define LAPACK_zgbbrd LAPACK_GLOBAL(zgbbrd,ZGBBRD)
+#define LAPACK_sorgbr LAPACK_GLOBAL(sorgbr,SORGBR)
+#define LAPACK_dorgbr LAPACK_GLOBAL(dorgbr,DORGBR)
+#define LAPACK_sormbr LAPACK_GLOBAL(sormbr,SORMBR)
+#define LAPACK_dormbr LAPACK_GLOBAL(dormbr,DORMBR)
+#define LAPACK_cungbr LAPACK_GLOBAL(cungbr,CUNGBR)
+#define LAPACK_zungbr LAPACK_GLOBAL(zungbr,ZUNGBR)
+#define LAPACK_cunmbr LAPACK_GLOBAL(cunmbr,CUNMBR)
+#define LAPACK_zunmbr LAPACK_GLOBAL(zunmbr,ZUNMBR)
+#define LAPACK_sbdsqr LAPACK_GLOBAL(sbdsqr,SBDSQR)
+#define LAPACK_dbdsqr LAPACK_GLOBAL(dbdsqr,DBDSQR)
+#define LAPACK_cbdsqr LAPACK_GLOBAL(cbdsqr,CBDSQR)
+#define LAPACK_zbdsqr LAPACK_GLOBAL(zbdsqr,ZBDSQR)
+#define LAPACK_sbdsdc LAPACK_GLOBAL(sbdsdc,SBDSDC)
+#define LAPACK_dbdsdc LAPACK_GLOBAL(dbdsdc,DBDSDC)
+#define LAPACK_ssytrd LAPACK_GLOBAL(ssytrd,SSYTRD)
+#define LAPACK_dsytrd LAPACK_GLOBAL(dsytrd,DSYTRD)
+#define LAPACK_sorgtr LAPACK_GLOBAL(sorgtr,SORGTR)
+#define LAPACK_dorgtr LAPACK_GLOBAL(dorgtr,DORGTR)
+#define LAPACK_sormtr LAPACK_GLOBAL(sormtr,SORMTR)
+#define LAPACK_dormtr LAPACK_GLOBAL(dormtr,DORMTR)
+#define LAPACK_chetrd LAPACK_GLOBAL(chetrd,CHETRD)
+#define LAPACK_zhetrd LAPACK_GLOBAL(zhetrd,ZHETRD)
+#define LAPACK_cungtr LAPACK_GLOBAL(cungtr,CUNGTR)
+#define LAPACK_zungtr LAPACK_GLOBAL(zungtr,ZUNGTR)
+#define LAPACK_cunmtr LAPACK_GLOBAL(cunmtr,CUNMTR)
+#define LAPACK_zunmtr LAPACK_GLOBAL(zunmtr,ZUNMTR)
+#define LAPACK_ssptrd LAPACK_GLOBAL(ssptrd,SSPTRD)
+#define LAPACK_dsptrd LAPACK_GLOBAL(dsptrd,DSPTRD)
+#define LAPACK_sopgtr LAPACK_GLOBAL(sopgtr,SOPGTR)
+#define LAPACK_dopgtr LAPACK_GLOBAL(dopgtr,DOPGTR)
+#define LAPACK_sopmtr LAPACK_GLOBAL(sopmtr,SOPMTR)
+#define LAPACK_dopmtr LAPACK_GLOBAL(dopmtr,DOPMTR)
+#define LAPACK_chptrd LAPACK_GLOBAL(chptrd,CHPTRD)
+#define LAPACK_zhptrd LAPACK_GLOBAL(zhptrd,ZHPTRD)
+#define LAPACK_cupgtr LAPACK_GLOBAL(cupgtr,CUPGTR)
+#define LAPACK_zupgtr LAPACK_GLOBAL(zupgtr,ZUPGTR)
+#define LAPACK_cupmtr LAPACK_GLOBAL(cupmtr,CUPMTR)
+#define LAPACK_zupmtr LAPACK_GLOBAL(zupmtr,ZUPMTR)
+#define LAPACK_ssbtrd LAPACK_GLOBAL(ssbtrd,SSBTRD)
+#define LAPACK_dsbtrd LAPACK_GLOBAL(dsbtrd,DSBTRD)
+#define LAPACK_chbtrd LAPACK_GLOBAL(chbtrd,CHBTRD)
+#define LAPACK_zhbtrd LAPACK_GLOBAL(zhbtrd,ZHBTRD)
+#define LAPACK_ssterf LAPACK_GLOBAL(ssterf,SSTERF)
+#define LAPACK_dsterf LAPACK_GLOBAL(dsterf,DSTERF)
+#define LAPACK_ssteqr LAPACK_GLOBAL(ssteqr,SSTEQR)
+#define LAPACK_dsteqr LAPACK_GLOBAL(dsteqr,DSTEQR)
+#define LAPACK_csteqr LAPACK_GLOBAL(csteqr,CSTEQR)
+#define LAPACK_zsteqr LAPACK_GLOBAL(zsteqr,ZSTEQR)
+#define LAPACK_sstemr LAPACK_GLOBAL(sstemr,SSTEMR)
+#define LAPACK_dstemr LAPACK_GLOBAL(dstemr,DSTEMR)
+#define LAPACK_cstemr LAPACK_GLOBAL(cstemr,CSTEMR)
+#define LAPACK_zstemr LAPACK_GLOBAL(zstemr,ZSTEMR)
+#define LAPACK_sstedc LAPACK_GLOBAL(sstedc,SSTEDC)
+#define LAPACK_dstedc LAPACK_GLOBAL(dstedc,DSTEDC)
+#define LAPACK_cstedc LAPACK_GLOBAL(cstedc,CSTEDC)
+#define LAPACK_zstedc LAPACK_GLOBAL(zstedc,ZSTEDC)
+#define LAPACK_sstegr LAPACK_GLOBAL(sstegr,SSTEGR)
+#define LAPACK_dstegr LAPACK_GLOBAL(dstegr,DSTEGR)
+#define LAPACK_cstegr LAPACK_GLOBAL(cstegr,CSTEGR)
+#define LAPACK_zstegr LAPACK_GLOBAL(zstegr,ZSTEGR)
+#define LAPACK_spteqr LAPACK_GLOBAL(spteqr,SPTEQR)
+#define LAPACK_dpteqr LAPACK_GLOBAL(dpteqr,DPTEQR)
+#define LAPACK_cpteqr LAPACK_GLOBAL(cpteqr,CPTEQR)
+#define LAPACK_zpteqr LAPACK_GLOBAL(zpteqr,ZPTEQR)
+#define LAPACK_sstebz LAPACK_GLOBAL(sstebz,SSTEBZ)
+#define LAPACK_dstebz LAPACK_GLOBAL(dstebz,DSTEBZ)
+#define LAPACK_sstein LAPACK_GLOBAL(sstein,SSTEIN)
+#define LAPACK_dstein LAPACK_GLOBAL(dstein,DSTEIN)
+#define LAPACK_cstein LAPACK_GLOBAL(cstein,CSTEIN)
+#define LAPACK_zstein LAPACK_GLOBAL(zstein,ZSTEIN)
+#define LAPACK_sdisna LAPACK_GLOBAL(sdisna,SDISNA)
+#define LAPACK_ddisna LAPACK_GLOBAL(ddisna,DDISNA)
+#define LAPACK_ssygst LAPACK_GLOBAL(ssygst,SSYGST)
+#define LAPACK_dsygst LAPACK_GLOBAL(dsygst,DSYGST)
+#define LAPACK_chegst LAPACK_GLOBAL(chegst,CHEGST)
+#define LAPACK_zhegst LAPACK_GLOBAL(zhegst,ZHEGST)
+#define LAPACK_sspgst LAPACK_GLOBAL(sspgst,SSPGST)
+#define LAPACK_dspgst LAPACK_GLOBAL(dspgst,DSPGST)
+#define LAPACK_chpgst LAPACK_GLOBAL(chpgst,CHPGST)
+#define LAPACK_zhpgst LAPACK_GLOBAL(zhpgst,ZHPGST)
+#define LAPACK_ssbgst LAPACK_GLOBAL(ssbgst,SSBGST)
+#define LAPACK_dsbgst LAPACK_GLOBAL(dsbgst,DSBGST)
+#define LAPACK_chbgst LAPACK_GLOBAL(chbgst,CHBGST)
+#define LAPACK_zhbgst LAPACK_GLOBAL(zhbgst,ZHBGST)
+#define LAPACK_spbstf LAPACK_GLOBAL(spbstf,SPBSTF)
+#define LAPACK_dpbstf LAPACK_GLOBAL(dpbstf,DPBSTF)
+#define LAPACK_cpbstf LAPACK_GLOBAL(cpbstf,CPBSTF)
+#define LAPACK_zpbstf LAPACK_GLOBAL(zpbstf,ZPBSTF)
+#define LAPACK_sgehrd LAPACK_GLOBAL(sgehrd,SGEHRD)
+#define LAPACK_dgehrd LAPACK_GLOBAL(dgehrd,DGEHRD)
+#define LAPACK_cgehrd LAPACK_GLOBAL(cgehrd,CGEHRD)
+#define LAPACK_zgehrd LAPACK_GLOBAL(zgehrd,ZGEHRD)
+#define LAPACK_sorghr LAPACK_GLOBAL(sorghr,SORGHR)
+#define LAPACK_dorghr LAPACK_GLOBAL(dorghr,DORGHR)
+#define LAPACK_sormhr LAPACK_GLOBAL(sormhr,SORMHR)
+#define LAPACK_dormhr LAPACK_GLOBAL(dormhr,DORMHR)
+#define LAPACK_cunghr LAPACK_GLOBAL(cunghr,CUNGHR)
+#define LAPACK_zunghr LAPACK_GLOBAL(zunghr,ZUNGHR)
+#define LAPACK_cunmhr LAPACK_GLOBAL(cunmhr,CUNMHR)
+#define LAPACK_zunmhr LAPACK_GLOBAL(zunmhr,ZUNMHR)
+#define LAPACK_sgebal LAPACK_GLOBAL(sgebal,SGEBAL)
+#define LAPACK_dgebal LAPACK_GLOBAL(dgebal,DGEBAL)
+#define LAPACK_cgebal LAPACK_GLOBAL(cgebal,CGEBAL)
+#define LAPACK_zgebal LAPACK_GLOBAL(zgebal,ZGEBAL)
+#define LAPACK_sgebak LAPACK_GLOBAL(sgebak,SGEBAK)
+#define LAPACK_dgebak LAPACK_GLOBAL(dgebak,DGEBAK)
+#define LAPACK_cgebak LAPACK_GLOBAL(cgebak,CGEBAK)
+#define LAPACK_zgebak LAPACK_GLOBAL(zgebak,ZGEBAK)
+#define LAPACK_shseqr LAPACK_GLOBAL(shseqr,SHSEQR)
+#define LAPACK_dhseqr LAPACK_GLOBAL(dhseqr,DHSEQR)
+#define LAPACK_chseqr LAPACK_GLOBAL(chseqr,CHSEQR)
+#define LAPACK_zhseqr LAPACK_GLOBAL(zhseqr,ZHSEQR)
+#define LAPACK_shsein LAPACK_GLOBAL(shsein,SHSEIN)
+#define LAPACK_dhsein LAPACK_GLOBAL(dhsein,DHSEIN)
+#define LAPACK_chsein LAPACK_GLOBAL(chsein,CHSEIN)
+#define LAPACK_zhsein LAPACK_GLOBAL(zhsein,ZHSEIN)
+#define LAPACK_strevc LAPACK_GLOBAL(strevc,STREVC)
+#define LAPACK_dtrevc LAPACK_GLOBAL(dtrevc,DTREVC)
+#define LAPACK_ctrevc LAPACK_GLOBAL(ctrevc,CTREVC)
+#define LAPACK_ztrevc LAPACK_GLOBAL(ztrevc,ZTREVC)
+#define LAPACK_strsna LAPACK_GLOBAL(strsna,STRSNA)
+#define LAPACK_dtrsna LAPACK_GLOBAL(dtrsna,DTRSNA)
+#define LAPACK_ctrsna LAPACK_GLOBAL(ctrsna,CTRSNA)
+#define LAPACK_ztrsna LAPACK_GLOBAL(ztrsna,ZTRSNA)
+#define LAPACK_strexc LAPACK_GLOBAL(strexc,STREXC)
+#define LAPACK_dtrexc LAPACK_GLOBAL(dtrexc,DTREXC)
+#define LAPACK_ctrexc LAPACK_GLOBAL(ctrexc,CTREXC)
+#define LAPACK_ztrexc LAPACK_GLOBAL(ztrexc,ZTREXC)
+#define LAPACK_strsen LAPACK_GLOBAL(strsen,STRSEN)
+#define LAPACK_dtrsen LAPACK_GLOBAL(dtrsen,DTRSEN)
+#define LAPACK_ctrsen LAPACK_GLOBAL(ctrsen,CTRSEN)
+#define LAPACK_ztrsen LAPACK_GLOBAL(ztrsen,ZTRSEN)
+#define LAPACK_strsyl LAPACK_GLOBAL(strsyl,STRSYL)
+#define LAPACK_dtrsyl LAPACK_GLOBAL(dtrsyl,DTRSYL)
+#define LAPACK_ctrsyl LAPACK_GLOBAL(ctrsyl,CTRSYL)
+#define LAPACK_ztrsyl LAPACK_GLOBAL(ztrsyl,ZTRSYL)
+#define LAPACK_sgghrd LAPACK_GLOBAL(sgghrd,SGGHRD)
+#define LAPACK_dgghrd LAPACK_GLOBAL(dgghrd,DGGHRD)
+#define LAPACK_cgghrd LAPACK_GLOBAL(cgghrd,CGGHRD)
+#define LAPACK_zgghrd LAPACK_GLOBAL(zgghrd,ZGGHRD)
+#define LAPACK_sggbal LAPACK_GLOBAL(sggbal,SGGBAL)
+#define LAPACK_dggbal LAPACK_GLOBAL(dggbal,DGGBAL)
+#define LAPACK_cggbal LAPACK_GLOBAL(cggbal,CGGBAL)
+#define LAPACK_zggbal LAPACK_GLOBAL(zggbal,ZGGBAL)
+#define LAPACK_sggbak LAPACK_GLOBAL(sggbak,SGGBAK)
+#define LAPACK_dggbak LAPACK_GLOBAL(dggbak,DGGBAK)
+#define LAPACK_cggbak LAPACK_GLOBAL(cggbak,CGGBAK)
+#define LAPACK_zggbak LAPACK_GLOBAL(zggbak,ZGGBAK)
+#define LAPACK_shgeqz LAPACK_GLOBAL(shgeqz,SHGEQZ)
+#define LAPACK_dhgeqz LAPACK_GLOBAL(dhgeqz,DHGEQZ)
+#define LAPACK_chgeqz LAPACK_GLOBAL(chgeqz,CHGEQZ)
+#define LAPACK_zhgeqz LAPACK_GLOBAL(zhgeqz,ZHGEQZ)
+#define LAPACK_stgevc LAPACK_GLOBAL(stgevc,STGEVC)
+#define LAPACK_dtgevc LAPACK_GLOBAL(dtgevc,DTGEVC)
+#define LAPACK_ctgevc LAPACK_GLOBAL(ctgevc,CTGEVC)
+#define LAPACK_ztgevc LAPACK_GLOBAL(ztgevc,ZTGEVC)
+#define LAPACK_stgexc LAPACK_GLOBAL(stgexc,STGEXC)
+#define LAPACK_dtgexc LAPACK_GLOBAL(dtgexc,DTGEXC)
+#define LAPACK_ctgexc LAPACK_GLOBAL(ctgexc,CTGEXC)
+#define LAPACK_ztgexc LAPACK_GLOBAL(ztgexc,ZTGEXC)
+#define LAPACK_stgsen LAPACK_GLOBAL(stgsen,STGSEN)
+#define LAPACK_dtgsen LAPACK_GLOBAL(dtgsen,DTGSEN)
+#define LAPACK_ctgsen LAPACK_GLOBAL(ctgsen,CTGSEN)
+#define LAPACK_ztgsen LAPACK_GLOBAL(ztgsen,ZTGSEN)
+#define LAPACK_stgsyl LAPACK_GLOBAL(stgsyl,STGSYL)
+#define LAPACK_dtgsyl LAPACK_GLOBAL(dtgsyl,DTGSYL)
+#define LAPACK_ctgsyl LAPACK_GLOBAL(ctgsyl,CTGSYL)
+#define LAPACK_ztgsyl LAPACK_GLOBAL(ztgsyl,ZTGSYL)
+#define LAPACK_stgsna LAPACK_GLOBAL(stgsna,STGSNA)
+#define LAPACK_dtgsna LAPACK_GLOBAL(dtgsna,DTGSNA)
+#define LAPACK_ctgsna LAPACK_GLOBAL(ctgsna,CTGSNA)
+#define LAPACK_ztgsna LAPACK_GLOBAL(ztgsna,ZTGSNA)
+#define LAPACK_sggsvp LAPACK_GLOBAL(sggsvp,SGGSVP)
+#define LAPACK_dggsvp LAPACK_GLOBAL(dggsvp,DGGSVP)
+#define LAPACK_cggsvp LAPACK_GLOBAL(cggsvp,CGGSVP)
+#define LAPACK_zggsvp LAPACK_GLOBAL(zggsvp,ZGGSVP)
+#define LAPACK_stgsja LAPACK_GLOBAL(stgsja,STGSJA)
+#define LAPACK_dtgsja LAPACK_GLOBAL(dtgsja,DTGSJA)
+#define LAPACK_ctgsja LAPACK_GLOBAL(ctgsja,CTGSJA)
+#define LAPACK_ztgsja LAPACK_GLOBAL(ztgsja,ZTGSJA)
+#define LAPACK_sgels LAPACK_GLOBAL(sgels,SGELS)
+#define LAPACK_dgels LAPACK_GLOBAL(dgels,DGELS)
+#define LAPACK_cgels LAPACK_GLOBAL(cgels,CGELS)
+#define LAPACK_zgels LAPACK_GLOBAL(zgels,ZGELS)
+#define LAPACK_sgelsy LAPACK_GLOBAL(sgelsy,SGELSY)
+#define LAPACK_dgelsy LAPACK_GLOBAL(dgelsy,DGELSY)
+#define LAPACK_cgelsy LAPACK_GLOBAL(cgelsy,CGELSY)
+#define LAPACK_zgelsy LAPACK_GLOBAL(zgelsy,ZGELSY)
+#define LAPACK_sgelss LAPACK_GLOBAL(sgelss,SGELSS)
+#define LAPACK_dgelss LAPACK_GLOBAL(dgelss,DGELSS)
+#define LAPACK_cgelss LAPACK_GLOBAL(cgelss,CGELSS)
+#define LAPACK_zgelss LAPACK_GLOBAL(zgelss,ZGELSS)
+#define LAPACK_sgelsd LAPACK_GLOBAL(sgelsd,SGELSD)
+#define LAPACK_dgelsd LAPACK_GLOBAL(dgelsd,DGELSD)
+#define LAPACK_cgelsd LAPACK_GLOBAL(cgelsd,CGELSD)
+#define LAPACK_zgelsd LAPACK_GLOBAL(zgelsd,ZGELSD)
+#define LAPACK_sgglse LAPACK_GLOBAL(sgglse,SGGLSE)
+#define LAPACK_dgglse LAPACK_GLOBAL(dgglse,DGGLSE)
+#define LAPACK_cgglse LAPACK_GLOBAL(cgglse,CGGLSE)
+#define LAPACK_zgglse LAPACK_GLOBAL(zgglse,ZGGLSE)
+#define LAPACK_sggglm LAPACK_GLOBAL(sggglm,SGGGLM)
+#define LAPACK_dggglm LAPACK_GLOBAL(dggglm,DGGGLM)
+#define LAPACK_cggglm LAPACK_GLOBAL(cggglm,CGGGLM)
+#define LAPACK_zggglm LAPACK_GLOBAL(zggglm,ZGGGLM)
+#define LAPACK_ssyev LAPACK_GLOBAL(ssyev,SSYEV)
+#define LAPACK_dsyev LAPACK_GLOBAL(dsyev,DSYEV)
+#define LAPACK_cheev LAPACK_GLOBAL(cheev,CHEEV)
+#define LAPACK_zheev LAPACK_GLOBAL(zheev,ZHEEV)
+#define LAPACK_ssyevd LAPACK_GLOBAL(ssyevd,SSYEVD)
+#define LAPACK_dsyevd LAPACK_GLOBAL(dsyevd,DSYEVD)
+#define LAPACK_cheevd LAPACK_GLOBAL(cheevd,CHEEVD)
+#define LAPACK_zheevd LAPACK_GLOBAL(zheevd,ZHEEVD)
+#define LAPACK_ssyevx LAPACK_GLOBAL(ssyevx,SSYEVX)
+#define LAPACK_dsyevx LAPACK_GLOBAL(dsyevx,DSYEVX)
+#define LAPACK_cheevx LAPACK_GLOBAL(cheevx,CHEEVX)
+#define LAPACK_zheevx LAPACK_GLOBAL(zheevx,ZHEEVX)
+#define LAPACK_ssyevr LAPACK_GLOBAL(ssyevr,SSYEVR)
+#define LAPACK_dsyevr LAPACK_GLOBAL(dsyevr,DSYEVR)
+#define LAPACK_cheevr LAPACK_GLOBAL(cheevr,CHEEVR)
+#define LAPACK_zheevr LAPACK_GLOBAL(zheevr,ZHEEVR)
+#define LAPACK_sspev LAPACK_GLOBAL(sspev,SSPEV)
+#define LAPACK_dspev LAPACK_GLOBAL(dspev,DSPEV)
+#define LAPACK_chpev LAPACK_GLOBAL(chpev,CHPEV)
+#define LAPACK_zhpev LAPACK_GLOBAL(zhpev,ZHPEV)
+#define LAPACK_sspevd LAPACK_GLOBAL(sspevd,SSPEVD)
+#define LAPACK_dspevd LAPACK_GLOBAL(dspevd,DSPEVD)
+#define LAPACK_chpevd LAPACK_GLOBAL(chpevd,CHPEVD)
+#define LAPACK_zhpevd LAPACK_GLOBAL(zhpevd,ZHPEVD)
+#define LAPACK_sspevx LAPACK_GLOBAL(sspevx,SSPEVX)
+#define LAPACK_dspevx LAPACK_GLOBAL(dspevx,DSPEVX)
+#define LAPACK_chpevx LAPACK_GLOBAL(chpevx,CHPEVX)
+#define LAPACK_zhpevx LAPACK_GLOBAL(zhpevx,ZHPEVX)
+#define LAPACK_ssbev LAPACK_GLOBAL(ssbev,SSBEV)
+#define LAPACK_dsbev LAPACK_GLOBAL(dsbev,DSBEV)
+#define LAPACK_chbev LAPACK_GLOBAL(chbev,CHBEV)
+#define LAPACK_zhbev LAPACK_GLOBAL(zhbev,ZHBEV)
+#define LAPACK_ssbevd LAPACK_GLOBAL(ssbevd,SSBEVD)
+#define LAPACK_dsbevd LAPACK_GLOBAL(dsbevd,DSBEVD)
+#define LAPACK_chbevd LAPACK_GLOBAL(chbevd,CHBEVD)
+#define LAPACK_zhbevd LAPACK_GLOBAL(zhbevd,ZHBEVD)
+#define LAPACK_ssbevx LAPACK_GLOBAL(ssbevx,SSBEVX)
+#define LAPACK_dsbevx LAPACK_GLOBAL(dsbevx,DSBEVX)
+#define LAPACK_chbevx LAPACK_GLOBAL(chbevx,CHBEVX)
+#define LAPACK_zhbevx LAPACK_GLOBAL(zhbevx,ZHBEVX)
+#define LAPACK_sstev LAPACK_GLOBAL(sstev,SSTEV)
+#define LAPACK_dstev LAPACK_GLOBAL(dstev,DSTEV)
+#define LAPACK_sstevd LAPACK_GLOBAL(sstevd,SSTEVD)
+#define LAPACK_dstevd LAPACK_GLOBAL(dstevd,DSTEVD)
+#define LAPACK_sstevx LAPACK_GLOBAL(sstevx,SSTEVX)
+#define LAPACK_dstevx LAPACK_GLOBAL(dstevx,DSTEVX)
+#define LAPACK_sstevr LAPACK_GLOBAL(sstevr,SSTEVR)
+#define LAPACK_dstevr LAPACK_GLOBAL(dstevr,DSTEVR)
+#define LAPACK_sgees LAPACK_GLOBAL(sgees,SGEES)
+#define LAPACK_dgees LAPACK_GLOBAL(dgees,DGEES)
+#define LAPACK_cgees LAPACK_GLOBAL(cgees,CGEES)
+#define LAPACK_zgees LAPACK_GLOBAL(zgees,ZGEES)
+#define LAPACK_sgeesx LAPACK_GLOBAL(sgeesx,SGEESX)
+#define LAPACK_dgeesx LAPACK_GLOBAL(dgeesx,DGEESX)
+#define LAPACK_cgeesx LAPACK_GLOBAL(cgeesx,CGEESX)
+#define LAPACK_zgeesx LAPACK_GLOBAL(zgeesx,ZGEESX)
+#define LAPACK_sgeev LAPACK_GLOBAL(sgeev,SGEEV)
+#define LAPACK_dgeev LAPACK_GLOBAL(dgeev,DGEEV)
+#define LAPACK_cgeev LAPACK_GLOBAL(cgeev,CGEEV)
+#define LAPACK_zgeev LAPACK_GLOBAL(zgeev,ZGEEV)
+#define LAPACK_sgeevx LAPACK_GLOBAL(sgeevx,SGEEVX)
+#define LAPACK_dgeevx LAPACK_GLOBAL(dgeevx,DGEEVX)
+#define LAPACK_cgeevx LAPACK_GLOBAL(cgeevx,CGEEVX)
+#define LAPACK_zgeevx LAPACK_GLOBAL(zgeevx,ZGEEVX)
+#define LAPACK_sgesvd LAPACK_GLOBAL(sgesvd,SGESVD)
+#define LAPACK_dgesvd LAPACK_GLOBAL(dgesvd,DGESVD)
+#define LAPACK_cgesvd LAPACK_GLOBAL(cgesvd,CGESVD)
+#define LAPACK_zgesvd LAPACK_GLOBAL(zgesvd,ZGESVD)
+#define LAPACK_sgesdd LAPACK_GLOBAL(sgesdd,SGESDD)
+#define LAPACK_dgesdd LAPACK_GLOBAL(dgesdd,DGESDD)
+#define LAPACK_cgesdd LAPACK_GLOBAL(cgesdd,CGESDD)
+#define LAPACK_zgesdd LAPACK_GLOBAL(zgesdd,ZGESDD)
+#define LAPACK_dgejsv LAPACK_GLOBAL(dgejsv,DGEJSV)
+#define LAPACK_sgejsv LAPACK_GLOBAL(sgejsv,SGEJSV)
+#define LAPACK_dgesvj LAPACK_GLOBAL(dgesvj,DGESVJ)
+#define LAPACK_sgesvj LAPACK_GLOBAL(sgesvj,SGESVJ)
+#define LAPACK_sggsvd LAPACK_GLOBAL(sggsvd,SGGSVD)
+#define LAPACK_dggsvd LAPACK_GLOBAL(dggsvd,DGGSVD)
+#define LAPACK_cggsvd LAPACK_GLOBAL(cggsvd,CGGSVD)
+#define LAPACK_zggsvd LAPACK_GLOBAL(zggsvd,ZGGSVD)
+#define LAPACK_ssygv LAPACK_GLOBAL(ssygv,SSYGV)
+#define LAPACK_dsygv LAPACK_GLOBAL(dsygv,DSYGV)
+#define LAPACK_chegv LAPACK_GLOBAL(chegv,CHEGV)
+#define LAPACK_zhegv LAPACK_GLOBAL(zhegv,ZHEGV)
+#define LAPACK_ssygvd LAPACK_GLOBAL(ssygvd,SSYGVD)
+#define LAPACK_dsygvd LAPACK_GLOBAL(dsygvd,DSYGVD)
+#define LAPACK_chegvd LAPACK_GLOBAL(chegvd,CHEGVD)
+#define LAPACK_zhegvd LAPACK_GLOBAL(zhegvd,ZHEGVD)
+#define LAPACK_ssygvx LAPACK_GLOBAL(ssygvx,SSYGVX)
+#define LAPACK_dsygvx LAPACK_GLOBAL(dsygvx,DSYGVX)
+#define LAPACK_chegvx LAPACK_GLOBAL(chegvx,CHEGVX)
+#define LAPACK_zhegvx LAPACK_GLOBAL(zhegvx,ZHEGVX)
+#define LAPACK_sspgv LAPACK_GLOBAL(sspgv,SSPGV)
+#define LAPACK_dspgv LAPACK_GLOBAL(dspgv,DSPGV)
+#define LAPACK_chpgv LAPACK_GLOBAL(chpgv,CHPGV)
+#define LAPACK_zhpgv LAPACK_GLOBAL(zhpgv,ZHPGV)
+#define LAPACK_sspgvd LAPACK_GLOBAL(sspgvd,SSPGVD)
+#define LAPACK_dspgvd LAPACK_GLOBAL(dspgvd,DSPGVD)
+#define LAPACK_chpgvd LAPACK_GLOBAL(chpgvd,CHPGVD)
+#define LAPACK_zhpgvd LAPACK_GLOBAL(zhpgvd,ZHPGVD)
+#define LAPACK_sspgvx LAPACK_GLOBAL(sspgvx,SSPGVX)
+#define LAPACK_dspgvx LAPACK_GLOBAL(dspgvx,DSPGVX)
+#define LAPACK_chpgvx LAPACK_GLOBAL(chpgvx,CHPGVX)
+#define LAPACK_zhpgvx LAPACK_GLOBAL(zhpgvx,ZHPGVX)
+#define LAPACK_ssbgv LAPACK_GLOBAL(ssbgv,SSBGV)
+#define LAPACK_dsbgv LAPACK_GLOBAL(dsbgv,DSBGV)
+#define LAPACK_chbgv LAPACK_GLOBAL(chbgv,CHBGV)
+#define LAPACK_zhbgv LAPACK_GLOBAL(zhbgv,ZHBGV)
+#define LAPACK_ssbgvd LAPACK_GLOBAL(ssbgvd,SSBGVD)
+#define LAPACK_dsbgvd LAPACK_GLOBAL(dsbgvd,DSBGVD)
+#define LAPACK_chbgvd LAPACK_GLOBAL(chbgvd,CHBGVD)
+#define LAPACK_zhbgvd LAPACK_GLOBAL(zhbgvd,ZHBGVD)
+#define LAPACK_ssbgvx LAPACK_GLOBAL(ssbgvx,SSBGVX)
+#define LAPACK_dsbgvx LAPACK_GLOBAL(dsbgvx,DSBGVX)
+#define LAPACK_chbgvx LAPACK_GLOBAL(chbgvx,CHBGVX)
+#define LAPACK_zhbgvx LAPACK_GLOBAL(zhbgvx,ZHBGVX)
+#define LAPACK_sgges LAPACK_GLOBAL(sgges,SGGES)
+#define LAPACK_dgges LAPACK_GLOBAL(dgges,DGGES)
+#define LAPACK_cgges LAPACK_GLOBAL(cgges,CGGES)
+#define LAPACK_zgges LAPACK_GLOBAL(zgges,ZGGES)
+#define LAPACK_sggesx LAPACK_GLOBAL(sggesx,SGGESX)
+#define LAPACK_dggesx LAPACK_GLOBAL(dggesx,DGGESX)
+#define LAPACK_cggesx LAPACK_GLOBAL(cggesx,CGGESX)
+#define LAPACK_zggesx LAPACK_GLOBAL(zggesx,ZGGESX)
+#define LAPACK_sggev LAPACK_GLOBAL(sggev,SGGEV)
+#define LAPACK_dggev LAPACK_GLOBAL(dggev,DGGEV)
+#define LAPACK_cggev LAPACK_GLOBAL(cggev,CGGEV)
+#define LAPACK_zggev LAPACK_GLOBAL(zggev,ZGGEV)
+#define LAPACK_sggevx LAPACK_GLOBAL(sggevx,SGGEVX)
+#define LAPACK_dggevx LAPACK_GLOBAL(dggevx,DGGEVX)
+#define LAPACK_cggevx LAPACK_GLOBAL(cggevx,CGGEVX)
+#define LAPACK_zggevx LAPACK_GLOBAL(zggevx,ZGGEVX)
+#define LAPACK_dsfrk LAPACK_GLOBAL(dsfrk,DSFRK)
+#define LAPACK_ssfrk LAPACK_GLOBAL(ssfrk,SSFRK)
+#define LAPACK_zhfrk LAPACK_GLOBAL(zhfrk,ZHFRK)
+#define LAPACK_chfrk LAPACK_GLOBAL(chfrk,CHFRK)
+#define LAPACK_dtfsm LAPACK_GLOBAL(dtfsm,DTFSM)
+#define LAPACK_stfsm LAPACK_GLOBAL(stfsm,STFSM)
+#define LAPACK_ztfsm LAPACK_GLOBAL(ztfsm,ZTFSM)
+#define LAPACK_ctfsm LAPACK_GLOBAL(ctfsm,CTFSM)
+#define LAPACK_dtfttp LAPACK_GLOBAL(dtfttp,DTFTTP)
+#define LAPACK_stfttp LAPACK_GLOBAL(stfttp,STFTTP)
+#define LAPACK_ztfttp LAPACK_GLOBAL(ztfttp,ZTFTTP)
+#define LAPACK_ctfttp LAPACK_GLOBAL(ctfttp,CTFTTP)
+#define LAPACK_dtfttr LAPACK_GLOBAL(dtfttr,DTFTTR)
+#define LAPACK_stfttr LAPACK_GLOBAL(stfttr,STFTTR)
+#define LAPACK_ztfttr LAPACK_GLOBAL(ztfttr,ZTFTTR)
+#define LAPACK_ctfttr LAPACK_GLOBAL(ctfttr,CTFTTR)
+#define LAPACK_dtpttf LAPACK_GLOBAL(dtpttf,DTPTTF)
+#define LAPACK_stpttf LAPACK_GLOBAL(stpttf,STPTTF)
+#define LAPACK_ztpttf LAPACK_GLOBAL(ztpttf,ZTPTTF)
+#define LAPACK_ctpttf LAPACK_GLOBAL(ctpttf,CTPTTF)
+#define LAPACK_dtpttr LAPACK_GLOBAL(dtpttr,DTPTTR)
+#define LAPACK_stpttr LAPACK_GLOBAL(stpttr,STPTTR)
+#define LAPACK_ztpttr LAPACK_GLOBAL(ztpttr,ZTPTTR)
+#define LAPACK_ctpttr LAPACK_GLOBAL(ctpttr,CTPTTR)
+#define LAPACK_dtrttf LAPACK_GLOBAL(dtrttf,DTRTTF)
+#define LAPACK_strttf LAPACK_GLOBAL(strttf,STRTTF)
+#define LAPACK_ztrttf LAPACK_GLOBAL(ztrttf,ZTRTTF)
+#define LAPACK_ctrttf LAPACK_GLOBAL(ctrttf,CTRTTF)
+#define LAPACK_dtrttp LAPACK_GLOBAL(dtrttp,DTRTTP)
+#define LAPACK_strttp LAPACK_GLOBAL(strttp,STRTTP)
+#define LAPACK_ztrttp LAPACK_GLOBAL(ztrttp,ZTRTTP)
+#define LAPACK_ctrttp LAPACK_GLOBAL(ctrttp,CTRTTP)
+#define LAPACK_sgeqrfp LAPACK_GLOBAL(sgeqrfp,SGEQRFP)
+#define LAPACK_dgeqrfp LAPACK_GLOBAL(dgeqrfp,DGEQRFP)
+#define LAPACK_cgeqrfp LAPACK_GLOBAL(cgeqrfp,CGEQRFP)
+#define LAPACK_zgeqrfp LAPACK_GLOBAL(zgeqrfp,ZGEQRFP)
+#define LAPACK_clacgv LAPACK_GLOBAL(clacgv,CLACGV)
+#define LAPACK_zlacgv LAPACK_GLOBAL(zlacgv,ZLACGV)
+#define LAPACK_slarnv LAPACK_GLOBAL(slarnv,SLARNV)
+#define LAPACK_dlarnv LAPACK_GLOBAL(dlarnv,DLARNV)
+#define LAPACK_clarnv LAPACK_GLOBAL(clarnv,CLARNV)
+#define LAPACK_zlarnv LAPACK_GLOBAL(zlarnv,ZLARNV)
+#define LAPACK_sgeqr2 LAPACK_GLOBAL(sgeqr2,SGEQR2)
+#define LAPACK_dgeqr2 LAPACK_GLOBAL(dgeqr2,DGEQR2)
+#define LAPACK_cgeqr2 LAPACK_GLOBAL(cgeqr2,CGEQR2)
+#define LAPACK_zgeqr2 LAPACK_GLOBAL(zgeqr2,ZGEQR2)
+#define LAPACK_slacpy LAPACK_GLOBAL(slacpy,SLACPY)
+#define LAPACK_dlacpy LAPACK_GLOBAL(dlacpy,DLACPY)
+#define LAPACK_clacpy LAPACK_GLOBAL(clacpy,CLACPY)
+#define LAPACK_zlacpy LAPACK_GLOBAL(zlacpy,ZLACPY)
+#define LAPACK_sgetf2 LAPACK_GLOBAL(sgetf2,SGETF2)
+#define LAPACK_dgetf2 LAPACK_GLOBAL(dgetf2,DGETF2)
+#define LAPACK_cgetf2 LAPACK_GLOBAL(cgetf2,CGETF2)
+#define LAPACK_zgetf2 LAPACK_GLOBAL(zgetf2,ZGETF2)
+#define LAPACK_slaswp LAPACK_GLOBAL(slaswp,SLASWP)
+#define LAPACK_dlaswp LAPACK_GLOBAL(dlaswp,DLASWP)
+#define LAPACK_claswp LAPACK_GLOBAL(claswp,CLASWP)
+#define LAPACK_zlaswp LAPACK_GLOBAL(zlaswp,ZLASWP)
+#define LAPACK_slange LAPACK_GLOBAL(slange,SLANGE)
+#define LAPACK_dlange LAPACK_GLOBAL(dlange,DLANGE)
+#define LAPACK_clange LAPACK_GLOBAL(clange,CLANGE)
+#define LAPACK_zlange LAPACK_GLOBAL(zlange,ZLANGE)
+#define LAPACK_clanhe LAPACK_GLOBAL(clanhe,CLANHE)
+#define LAPACK_zlanhe LAPACK_GLOBAL(zlanhe,ZLANHE)
+#define LAPACK_slansy LAPACK_GLOBAL(slansy,SLANSY)
+#define LAPACK_dlansy LAPACK_GLOBAL(dlansy,DLANSY)
+#define LAPACK_clansy LAPACK_GLOBAL(clansy,CLANSY)
+#define LAPACK_zlansy LAPACK_GLOBAL(zlansy,ZLANSY)
+#define LAPACK_slantr LAPACK_GLOBAL(slantr,SLANTR)
+#define LAPACK_dlantr LAPACK_GLOBAL(dlantr,DLANTR)
+#define LAPACK_clantr LAPACK_GLOBAL(clantr,CLANTR)
+#define LAPACK_zlantr LAPACK_GLOBAL(zlantr,ZLANTR)
+#define LAPACK_slamch LAPACK_GLOBAL(slamch,SLAMCH)
+#define LAPACK_dlamch LAPACK_GLOBAL(dlamch,DLAMCH)
+#define LAPACK_sgelq2 LAPACK_GLOBAL(sgelq2,SGELQ2)
+#define LAPACK_dgelq2 LAPACK_GLOBAL(dgelq2,DGELQ2)
+#define LAPACK_cgelq2 LAPACK_GLOBAL(cgelq2,CGELQ2)
+#define LAPACK_zgelq2 LAPACK_GLOBAL(zgelq2,ZGELQ2)
+#define LAPACK_slarfb LAPACK_GLOBAL(slarfb,SLARFB)
+#define LAPACK_dlarfb LAPACK_GLOBAL(dlarfb,DLARFB)
+#define LAPACK_clarfb LAPACK_GLOBAL(clarfb,CLARFB)
+#define LAPACK_zlarfb LAPACK_GLOBAL(zlarfb,ZLARFB)
+#define LAPACK_slarfg LAPACK_GLOBAL(slarfg,SLARFG)
+#define LAPACK_dlarfg LAPACK_GLOBAL(dlarfg,DLARFG)
+#define LAPACK_clarfg LAPACK_GLOBAL(clarfg,CLARFG)
+#define LAPACK_zlarfg LAPACK_GLOBAL(zlarfg,ZLARFG)
+#define LAPACK_slarft LAPACK_GLOBAL(slarft,SLARFT)
+#define LAPACK_dlarft LAPACK_GLOBAL(dlarft,DLARFT)
+#define LAPACK_clarft LAPACK_GLOBAL(clarft,CLARFT)
+#define LAPACK_zlarft LAPACK_GLOBAL(zlarft,ZLARFT)
+#define LAPACK_slarfx LAPACK_GLOBAL(slarfx,SLARFX)
+#define LAPACK_dlarfx LAPACK_GLOBAL(dlarfx,DLARFX)
+#define LAPACK_clarfx LAPACK_GLOBAL(clarfx,CLARFX)
+#define LAPACK_zlarfx LAPACK_GLOBAL(zlarfx,ZLARFX)
+#define LAPACK_slatms LAPACK_GLOBAL(slatms,SLATMS)
+#define LAPACK_dlatms LAPACK_GLOBAL(dlatms,DLATMS)
+#define LAPACK_clatms LAPACK_GLOBAL(clatms,CLATMS)
+#define LAPACK_zlatms LAPACK_GLOBAL(zlatms,ZLATMS)
+#define LAPACK_slag2d LAPACK_GLOBAL(slag2d,SLAG2D)
+#define LAPACK_dlag2s LAPACK_GLOBAL(dlag2s,DLAG2S)
+#define LAPACK_clag2z LAPACK_GLOBAL(clag2z,CLAG2Z)
+#define LAPACK_zlag2c LAPACK_GLOBAL(zlag2c,ZLAG2C)
+#define LAPACK_slauum LAPACK_GLOBAL(slauum,SLAUUM)
+#define LAPACK_dlauum LAPACK_GLOBAL(dlauum,DLAUUM)
+#define LAPACK_clauum LAPACK_GLOBAL(clauum,CLAUUM)
+#define LAPACK_zlauum LAPACK_GLOBAL(zlauum,ZLAUUM)
+#define LAPACK_slagge LAPACK_GLOBAL(slagge,SLAGGE)
+#define LAPACK_dlagge LAPACK_GLOBAL(dlagge,DLAGGE)
+#define LAPACK_clagge LAPACK_GLOBAL(clagge,CLAGGE)
+#define LAPACK_zlagge LAPACK_GLOBAL(zlagge,ZLAGGE)
+#define LAPACK_slaset LAPACK_GLOBAL(slaset,SLASET)
+#define LAPACK_dlaset LAPACK_GLOBAL(dlaset,DLASET)
+#define LAPACK_claset LAPACK_GLOBAL(claset,CLASET)
+#define LAPACK_zlaset LAPACK_GLOBAL(zlaset,ZLASET)
+#define LAPACK_slasrt LAPACK_GLOBAL(slasrt,SLASRT)
+#define LAPACK_dlasrt LAPACK_GLOBAL(dlasrt,DLASRT)
+#define LAPACK_slagsy LAPACK_GLOBAL(slagsy,SLAGSY)
+#define LAPACK_dlagsy LAPACK_GLOBAL(dlagsy,DLAGSY)
+#define LAPACK_clagsy LAPACK_GLOBAL(clagsy,CLAGSY)
+#define LAPACK_zlagsy LAPACK_GLOBAL(zlagsy,ZLAGSY)
+#define LAPACK_claghe LAPACK_GLOBAL(claghe,CLAGHE)
+#define LAPACK_zlaghe LAPACK_GLOBAL(zlaghe,ZLAGHE)
+#define LAPACK_slapmr LAPACK_GLOBAL(slapmr,SLAPMR)
+#define LAPACK_dlapmr LAPACK_GLOBAL(dlapmr,DLAPMR)
+#define LAPACK_clapmr LAPACK_GLOBAL(clapmr,CLAPMR)
+#define LAPACK_zlapmr LAPACK_GLOBAL(zlapmr,ZLAPMR)
+#define LAPACK_slapy2 LAPACK_GLOBAL(slapy2,SLAPY2)
+#define LAPACK_dlapy2 LAPACK_GLOBAL(dlapy2,DLAPY2)
+#define LAPACK_slapy3 LAPACK_GLOBAL(slapy3,SLAPY3)
+#define LAPACK_dlapy3 LAPACK_GLOBAL(dlapy3,DLAPY3)
+#define LAPACK_slartgp LAPACK_GLOBAL(slartgp,SLARTGP)
+#define LAPACK_dlartgp LAPACK_GLOBAL(dlartgp,DLARTGP)
+#define LAPACK_slartgs LAPACK_GLOBAL(slartgs,SLARTGS)
+#define LAPACK_dlartgs LAPACK_GLOBAL(dlartgs,DLARTGS)
+// LAPACK 3.3.0
+#define LAPACK_cbbcsd LAPACK_GLOBAL(cbbcsd,CBBCSD)
+#define LAPACK_cheswapr LAPACK_GLOBAL(cheswapr,CHESWAPR)
+#define LAPACK_chetri2 LAPACK_GLOBAL(chetri2,CHETRI2)
+#define LAPACK_chetri2x LAPACK_GLOBAL(chetri2x,CHETRI2X)
+#define LAPACK_chetrs2 LAPACK_GLOBAL(chetrs2,CHETRS2)
+#define LAPACK_csyconv LAPACK_GLOBAL(csyconv,CSYCONV)
+#define LAPACK_csyswapr LAPACK_GLOBAL(csyswapr,CSYSWAPR)
+#define LAPACK_csytri2 LAPACK_GLOBAL(csytri2,CSYTRI2)
+#define LAPACK_csytri2x LAPACK_GLOBAL(csytri2x,CSYTRI2X)
+#define LAPACK_csytrs2 LAPACK_GLOBAL(csytrs2,CSYTRS2)
+#define LAPACK_cunbdb LAPACK_GLOBAL(cunbdb,CUNBDB)
+#define LAPACK_cuncsd LAPACK_GLOBAL(cuncsd,CUNCSD)
+#define LAPACK_dbbcsd LAPACK_GLOBAL(dbbcsd,DBBCSD)
+#define LAPACK_dorbdb LAPACK_GLOBAL(dorbdb,DORBDB)
+#define LAPACK_dorcsd LAPACK_GLOBAL(dorcsd,DORCSD)
+#define LAPACK_dsyconv LAPACK_GLOBAL(dsyconv,DSYCONV)
+#define LAPACK_dsyswapr LAPACK_GLOBAL(dsyswapr,DSYSWAPR)
+#define LAPACK_dsytri2 LAPACK_GLOBAL(dsytri2,DSYTRI2)
+#define LAPACK_dsytri2x LAPACK_GLOBAL(dsytri2x,DSYTRI2X)
+#define LAPACK_dsytrs2 LAPACK_GLOBAL(dsytrs2,DSYTRS2)
+#define LAPACK_sbbcsd LAPACK_GLOBAL(sbbcsd,SBBCSD)
+#define LAPACK_sorbdb LAPACK_GLOBAL(sorbdb,SORBDB)
+#define LAPACK_sorcsd LAPACK_GLOBAL(sorcsd,SORCSD)
+#define LAPACK_ssyconv LAPACK_GLOBAL(ssyconv,SSYCONV)
+#define LAPACK_ssyswapr LAPACK_GLOBAL(ssyswapr,SSYSWAPR)
+#define LAPACK_ssytri2 LAPACK_GLOBAL(ssytri2,SSYTRI2)
+#define LAPACK_ssytri2x LAPACK_GLOBAL(ssytri2x,SSYTRI2X)
+#define LAPACK_ssytrs2 LAPACK_GLOBAL(ssytrs2,SSYTRS2)
+#define LAPACK_zbbcsd LAPACK_GLOBAL(zbbcsd,ZBBCSD)
+#define LAPACK_zheswapr LAPACK_GLOBAL(zheswapr,ZHESWAPR)
+#define LAPACK_zhetri2 LAPACK_GLOBAL(zhetri2,ZHETRI2)
+#define LAPACK_zhetri2x LAPACK_GLOBAL(zhetri2x,ZHETRI2X)
+#define LAPACK_zhetrs2 LAPACK_GLOBAL(zhetrs2,ZHETRS2)
+#define LAPACK_zsyconv LAPACK_GLOBAL(zsyconv,ZSYCONV)
+#define LAPACK_zsyswapr LAPACK_GLOBAL(zsyswapr,ZSYSWAPR)
+#define LAPACK_zsytri2 LAPACK_GLOBAL(zsytri2,ZSYTRI2)
+#define LAPACK_zsytri2x LAPACK_GLOBAL(zsytri2x,ZSYTRI2X)
+#define LAPACK_zsytrs2 LAPACK_GLOBAL(zsytrs2,ZSYTRS2)
+#define LAPACK_zunbdb LAPACK_GLOBAL(zunbdb,ZUNBDB)
+#define LAPACK_zuncsd LAPACK_GLOBAL(zuncsd,ZUNCSD)
+// LAPACK 3.4.0
+#define LAPACK_sgemqrt LAPACK_GLOBAL(sgemqrt,SGEMQRT)
+#define LAPACK_dgemqrt LAPACK_GLOBAL(dgemqrt,DGEMQRT)
+#define LAPACK_cgemqrt LAPACK_GLOBAL(cgemqrt,CGEMQRT)
+#define LAPACK_zgemqrt LAPACK_GLOBAL(zgemqrt,ZGEMQRT)
+#define LAPACK_sgeqrt LAPACK_GLOBAL(sgeqrt,SGEQRT)
+#define LAPACK_dgeqrt LAPACK_GLOBAL(dgeqrt,DGEQRT)
+#define LAPACK_cgeqrt LAPACK_GLOBAL(cgeqrt,CGEQRT)
+#define LAPACK_zgeqrt LAPACK_GLOBAL(zgeqrt,ZGEQRT)
+#define LAPACK_sgeqrt2 LAPACK_GLOBAL(sgeqrt2,SGEQRT2)
+#define LAPACK_dgeqrt2 LAPACK_GLOBAL(dgeqrt2,DGEQRT2)
+#define LAPACK_cgeqrt2 LAPACK_GLOBAL(cgeqrt2,CGEQRT2)
+#define LAPACK_zgeqrt2 LAPACK_GLOBAL(zgeqrt2,ZGEQRT2)
+#define LAPACK_sgeqrt3 LAPACK_GLOBAL(sgeqrt3,SGEQRT3)
+#define LAPACK_dgeqrt3 LAPACK_GLOBAL(dgeqrt3,DGEQRT3)
+#define LAPACK_cgeqrt3 LAPACK_GLOBAL(cgeqrt3,CGEQRT3)
+#define LAPACK_zgeqrt3 LAPACK_GLOBAL(zgeqrt3,ZGEQRT3)
+#define LAPACK_stpmqrt LAPACK_GLOBAL(stpmqrt,STPMQRT)
+#define LAPACK_dtpmqrt LAPACK_GLOBAL(dtpmqrt,DTPMQRT)
+#define LAPACK_ctpmqrt LAPACK_GLOBAL(ctpmqrt,CTPMQRT)
+#define LAPACK_ztpmqrt LAPACK_GLOBAL(ztpmqrt,ZTPMQRT)
+#define LAPACK_dtpqrt LAPACK_GLOBAL(dtpqrt,DTPQRT)
+#define LAPACK_ctpqrt LAPACK_GLOBAL(ctpqrt,CTPQRT)
+#define LAPACK_ztpqrt LAPACK_GLOBAL(ztpqrt,ZTPQRT)
+#define LAPACK_stpqrt2 LAPACK_GLOBAL(stpqrt2,STPQRT2)
+#define LAPACK_dtpqrt2 LAPACK_GLOBAL(dtpqrt2,DTPQRT2)
+#define LAPACK_ctpqrt2 LAPACK_GLOBAL(ctpqrt2,CTPQRT2)
+#define LAPACK_ztpqrt2 LAPACK_GLOBAL(ztpqrt2,ZTPQRT2)
+#define LAPACK_stprfb LAPACK_GLOBAL(stprfb,STPRFB)
+#define LAPACK_dtprfb LAPACK_GLOBAL(dtprfb,DTPRFB)
+#define LAPACK_ctprfb LAPACK_GLOBAL(ctprfb,CTPRFB)
+#define LAPACK_ztprfb LAPACK_GLOBAL(ztprfb,ZTPRFB)
+// LAPACK 3.X.X
+#define LAPACK_csyr LAPACK_GLOBAL(csyr,CSYR)
+#define LAPACK_zsyr LAPACK_GLOBAL(zsyr,ZSYR)
+
+
+void LAPACK_sgetrf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_dgetrf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_cgetrf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* ipiv, lapack_int *info );
+void LAPACK_zgetrf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ipiv, lapack_int *info );
+void LAPACK_sgbtrf( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, float* ab, lapack_int* ldab,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_dgbtrf( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, double* ab, lapack_int* ldab,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_cgbtrf( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_zgbtrf( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_sgttrf( lapack_int* n, float* dl, float* d, float* du, float* du2,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_dgttrf( lapack_int* n, double* dl, double* d, double* du,
+                    double* du2, lapack_int* ipiv, lapack_int *info );
+void LAPACK_cgttrf( lapack_int* n, lapack_complex_float* dl,
+                    lapack_complex_float* d, lapack_complex_float* du,
+                    lapack_complex_float* du2, lapack_int* ipiv,
+                    lapack_int *info );
+void LAPACK_zgttrf( lapack_int* n, lapack_complex_double* dl,
+                    lapack_complex_double* d, lapack_complex_double* du,
+                    lapack_complex_double* du2, lapack_int* ipiv,
+                    lapack_int *info );
+void LAPACK_spotrf( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dpotrf( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_cpotrf( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_zpotrf( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_dpstrf( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* piv, lapack_int* rank, double* tol,
+                    double* work, lapack_int *info );
+void LAPACK_spstrf( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* piv, lapack_int* rank, float* tol, float* work,
+                    lapack_int *info );
+void LAPACK_zpstrf( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* piv, lapack_int* rank,
+                    double* tol, double* work, lapack_int *info );
+void LAPACK_cpstrf( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* piv, lapack_int* rank,
+                    float* tol, float* work, lapack_int *info );
+void LAPACK_dpftrf( char* transr, char* uplo, lapack_int* n, double* a,
+                    lapack_int *info );
+void LAPACK_spftrf( char* transr, char* uplo, lapack_int* n, float* a,
+                    lapack_int *info );
+void LAPACK_zpftrf( char* transr, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int *info );
+void LAPACK_cpftrf( char* transr, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int *info );
+void LAPACK_spptrf( char* uplo, lapack_int* n, float* ap, lapack_int *info );
+void LAPACK_dpptrf( char* uplo, lapack_int* n, double* ap, lapack_int *info );
+void LAPACK_cpptrf( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    lapack_int *info );
+void LAPACK_zpptrf( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    lapack_int *info );
+void LAPACK_spbtrf( char* uplo, lapack_int* n, lapack_int* kd, float* ab,
+                    lapack_int* ldab, lapack_int *info );
+void LAPACK_dpbtrf( char* uplo, lapack_int* n, lapack_int* kd, double* ab,
+                    lapack_int* ldab, lapack_int *info );
+void LAPACK_cpbtrf( char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_int *info );
+void LAPACK_zpbtrf( char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_int *info );
+void LAPACK_spttrf( lapack_int* n, float* d, float* e, lapack_int *info );
+void LAPACK_dpttrf( lapack_int* n, double* d, double* e, lapack_int *info );
+void LAPACK_cpttrf( lapack_int* n, float* d, lapack_complex_float* e,
+                    lapack_int *info );
+void LAPACK_zpttrf( lapack_int* n, double* d, lapack_complex_double* e,
+                    lapack_int *info );
+void LAPACK_ssytrf( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* ipiv, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dsytrf( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* ipiv, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_csytrf( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* ipiv,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zsytrf( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ipiv,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_chetrf( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* ipiv,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zhetrf( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ipiv,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_ssptrf( char* uplo, lapack_int* n, float* ap, lapack_int* ipiv,
+                    lapack_int *info );
+void LAPACK_dsptrf( char* uplo, lapack_int* n, double* ap, lapack_int* ipiv,
+                    lapack_int *info );
+void LAPACK_csptrf( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_zsptrf( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_chptrf( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_zhptrf( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_sgetrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const float* a, lapack_int* lda, const lapack_int* ipiv,
+                    float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_dgetrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, const lapack_int* ipiv,
+                    double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_cgetrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_zgetrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_sgbtrs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const float* ab, lapack_int* ldab,
+                    const lapack_int* ipiv, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dgbtrs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const double* ab, lapack_int* ldab,
+                    const lapack_int* ipiv, double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_cgbtrs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const lapack_complex_float* ab,
+                    lapack_int* ldab, const lapack_int* ipiv,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zgbtrs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const lapack_complex_double* ab,
+                    lapack_int* ldab, const lapack_int* ipiv,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_sgttrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const float* dl, const float* d, const float* du,
+                    const float* du2, const lapack_int* ipiv, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dgttrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const double* dl, const double* d, const double* du,
+                    const double* du2, const lapack_int* ipiv, double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_cgttrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* dl,
+                    const lapack_complex_float* d,
+                    const lapack_complex_float* du,
+                    const lapack_complex_float* du2, const lapack_int* ipiv,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zgttrs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* dl,
+                    const lapack_complex_double* d,
+                    const lapack_complex_double* du,
+                    const lapack_complex_double* du2, const lapack_int* ipiv,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_spotrs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dpotrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_cpotrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zpotrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dpftrs( char* transr, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_spftrs( char* transr, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* a, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zpftrs( char* transr, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_cpftrs( char* transr, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_spptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* ap, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dpptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* ap, double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_cpptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_zpptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_spbtrs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const float* ab, lapack_int* ldab, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dpbtrs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const double* ab, lapack_int* ldab, double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_cpbtrs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zpbtrs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_spttrs( lapack_int* n, lapack_int* nrhs, const float* d,
+                    const float* e, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dpttrs( lapack_int* n, lapack_int* nrhs, const double* d,
+                    const double* e, double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_cpttrs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* d,
+                    const lapack_complex_float* e, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_zpttrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* d, const lapack_complex_double* e,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_ssytrs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* a,
+                    lapack_int* lda, const lapack_int* ipiv, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dsytrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, const lapack_int* ipiv,
+                    double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_csytrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_zsytrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_chetrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_zhetrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_ssptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* ap, const lapack_int* ipiv, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dsptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* ap, const lapack_int* ipiv, double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_csptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap, const lapack_int* ipiv,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zsptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap, const lapack_int* ipiv,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_chptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap, const lapack_int* ipiv,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zhptrs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap, const lapack_int* ipiv,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_strtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dtrtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const double* a, lapack_int* lda,
+                    double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_ctrtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_ztrtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_stptrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const float* ap, float* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_dtptrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const double* ap, double* b,
+                    lapack_int* ldb, lapack_int *info );
+void LAPACK_ctptrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_float* ap,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_ztptrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_double* ap,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_stbtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs, const float* ab,
+                    lapack_int* ldab, float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dtbtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs, const double* ab,
+                    lapack_int* ldab, double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_ctbtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_ztbtrs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_sgecon( char* norm, lapack_int* n, const float* a, lapack_int* lda,
+                    float* anorm, float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgecon( char* norm, lapack_int* n, const double* a, lapack_int* lda,
+                    double* anorm, double* rcond, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cgecon( char* norm, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, float* anorm, float* rcond,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgecon( char* norm, lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, double* anorm, double* rcond,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sgbcon( char* norm, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    const float* ab, lapack_int* ldab, const lapack_int* ipiv,
+                    float* anorm, float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgbcon( char* norm, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    const double* ab, lapack_int* ldab, const lapack_int* ipiv,
+                    double* anorm, double* rcond, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cgbcon( char* norm, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    const lapack_int* ipiv, float* anorm, float* rcond,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgbcon( char* norm, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sgtcon( char* norm, lapack_int* n, const float* dl, const float* d,
+                    const float* du, const float* du2, const lapack_int* ipiv,
+                    float* anorm, float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgtcon( char* norm, lapack_int* n, const double* dl,
+                    const double* d, const double* du, const double* du2,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cgtcon( char* norm, lapack_int* n, const lapack_complex_float* dl,
+                    const lapack_complex_float* d,
+                    const lapack_complex_float* du,
+                    const lapack_complex_float* du2, const lapack_int* ipiv,
+                    float* anorm, float* rcond, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zgtcon( char* norm, lapack_int* n, const lapack_complex_double* dl,
+                    const lapack_complex_double* d,
+                    const lapack_complex_double* du,
+                    const lapack_complex_double* du2, const lapack_int* ipiv,
+                    double* anorm, double* rcond, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_spocon( char* uplo, lapack_int* n, const float* a, lapack_int* lda,
+                    float* anorm, float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dpocon( char* uplo, lapack_int* n, const double* a, lapack_int* lda,
+                    double* anorm, double* rcond, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cpocon( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, float* anorm, float* rcond,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zpocon( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, double* anorm, double* rcond,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sppcon( char* uplo, lapack_int* n, const float* ap, float* anorm,
+                    float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dppcon( char* uplo, lapack_int* n, const double* ap, double* anorm,
+                    double* rcond, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cppcon( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    float* anorm, float* rcond, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zppcon( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    double* anorm, double* rcond, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_spbcon( char* uplo, lapack_int* n, lapack_int* kd, const float* ab,
+                    lapack_int* ldab, float* anorm, float* rcond, float* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dpbcon( char* uplo, lapack_int* n, lapack_int* kd, const double* ab,
+                    lapack_int* ldab, double* anorm, double* rcond,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cpbcon( char* uplo, lapack_int* n, lapack_int* kd,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    float* anorm, float* rcond, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zpbcon( char* uplo, lapack_int* n, lapack_int* kd,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    double* anorm, double* rcond, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_sptcon( lapack_int* n, const float* d, const float* e, float* anorm,
+                    float* rcond, float* work, lapack_int *info );
+void LAPACK_dptcon( lapack_int* n, const double* d, const double* e,
+                    double* anorm, double* rcond, double* work,
+                    lapack_int *info );
+void LAPACK_cptcon( lapack_int* n, const float* d,
+                    const lapack_complex_float* e, float* anorm, float* rcond,
+                    float* work, lapack_int *info );
+void LAPACK_zptcon( lapack_int* n, const double* d,
+                    const lapack_complex_double* e, double* anorm,
+                    double* rcond, double* work, lapack_int *info );
+void LAPACK_ssycon( char* uplo, lapack_int* n, const float* a, lapack_int* lda,
+                    const lapack_int* ipiv, float* anorm, float* rcond,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dsycon( char* uplo, lapack_int* n, const double* a, lapack_int* lda,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_csycon( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_int* ipiv, float* anorm,
+                    float* rcond, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zsycon( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_int* ipiv, double* anorm,
+                    double* rcond, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_checon( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_int* ipiv, float* anorm,
+                    float* rcond, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zhecon( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_int* ipiv, double* anorm,
+                    double* rcond, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_sspcon( char* uplo, lapack_int* n, const float* ap,
+                    const lapack_int* ipiv, float* anorm, float* rcond,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dspcon( char* uplo, lapack_int* n, const double* ap,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cspcon( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    const lapack_int* ipiv, float* anorm, float* rcond,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zspcon( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_chpcon( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    const lapack_int* ipiv, float* anorm, float* rcond,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zhpcon( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    const lapack_int* ipiv, double* anorm, double* rcond,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_strcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const float* a, lapack_int* lda, float* rcond, float* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dtrcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const double* a, lapack_int* lda, double* rcond,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_ctrcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    float* rcond, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztrcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    double* rcond, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_stpcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const float* ap, float* rcond, float* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dtpcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const double* ap, double* rcond, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ctpcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const lapack_complex_float* ap, float* rcond,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztpcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    const lapack_complex_double* ap, double* rcond,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_stbcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    lapack_int* kd, const float* ab, lapack_int* ldab,
+                    float* rcond, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dtbcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    lapack_int* kd, const double* ab, lapack_int* ldab,
+                    double* rcond, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_ctbcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    lapack_int* kd, const lapack_complex_float* ab,
+                    lapack_int* ldab, float* rcond, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_ztbcon( char* norm, char* uplo, char* diag, lapack_int* n,
+                    lapack_int* kd, const lapack_complex_double* ab,
+                    lapack_int* ldab, double* rcond,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sgerfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const float* a, lapack_int* lda, const float* af,
+                    lapack_int* ldaf, const lapack_int* ipiv, const float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* ferr,
+                    float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgerfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, const double* af,
+                    lapack_int* ldaf, const lapack_int* ipiv, const double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cgerfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zgerfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_dgerfsx( char* trans, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const double* a, lapack_int* lda, const double* af,
+                     lapack_int* ldaf, const lapack_int* ipiv, const double* r,
+                     const double* c, const double* b, lapack_int* ldb,
+                     double* x, lapack_int* ldx, double* rcond, double* berr,
+                     lapack_int* n_err_bnds, double* err_bnds_norm,
+                     double* err_bnds_comp, lapack_int* nparams, double* params,
+                     double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_sgerfsx( char* trans, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const float* a, lapack_int* lda, const float* af,
+                     lapack_int* ldaf, const lapack_int* ipiv, const float* r,
+                     const float* c, const float* b, lapack_int* ldb, float* x,
+                     lapack_int* ldx, float* rcond, float* berr,
+                     lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_zgerfsx( char* trans, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_complex_double* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const double* r, const double* c,
+                     const lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cgerfsx( char* trans, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_float* a, lapack_int* lda,
+                     const lapack_complex_float* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const float* r, const float* c,
+                     const lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* berr, lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sgbrfs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const float* ab, lapack_int* ldab,
+                    const float* afb, lapack_int* ldafb, const lapack_int* ipiv,
+                    const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgbrfs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const double* ab, lapack_int* ldab,
+                    const double* afb, lapack_int* ldafb,
+                    const lapack_int* ipiv, const double* b, lapack_int* ldb,
+                    double* x, lapack_int* ldx, double* ferr, double* berr,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cgbrfs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const lapack_complex_float* ab,
+                    lapack_int* ldab, const lapack_complex_float* afb,
+                    lapack_int* ldafb, const lapack_int* ipiv,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgbrfs( char* trans, lapack_int* n, lapack_int* kl, lapack_int* ku,
+                    lapack_int* nrhs, const lapack_complex_double* ab,
+                    lapack_int* ldab, const lapack_complex_double* afb,
+                    lapack_int* ldafb, const lapack_int* ipiv,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_dgbrfsx( char* trans, char* equed, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs, const double* ab,
+                     lapack_int* ldab, const double* afb, lapack_int* ldafb,
+                     const lapack_int* ipiv, const double* r, const double* c,
+                     const double* b, lapack_int* ldb, double* x,
+                     lapack_int* ldx, double* rcond, double* berr,
+                     lapack_int* n_err_bnds, double* err_bnds_norm,
+                     double* err_bnds_comp, lapack_int* nparams, double* params,
+                     double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_sgbrfsx( char* trans, char* equed, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs, const float* ab,
+                     lapack_int* ldab, const float* afb, lapack_int* ldafb,
+                     const lapack_int* ipiv, const float* r, const float* c,
+                     const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                     float* rcond, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params, float* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_zgbrfsx( char* trans, char* equed, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs,
+                     const lapack_complex_double* ab, lapack_int* ldab,
+                     const lapack_complex_double* afb, lapack_int* ldafb,
+                     const lapack_int* ipiv, const double* r, const double* c,
+                     const lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cgbrfsx( char* trans, char* equed, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs,
+                     const lapack_complex_float* ab, lapack_int* ldab,
+                     const lapack_complex_float* afb, lapack_int* ldafb,
+                     const lapack_int* ipiv, const float* r, const float* c,
+                     const lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* berr, lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sgtrfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const float* dl, const float* d, const float* du,
+                    const float* dlf, const float* df, const float* duf,
+                    const float* du2, const lapack_int* ipiv, const float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* ferr,
+                    float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgtrfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const double* dl, const double* d, const double* du,
+                    const double* dlf, const double* df, const double* duf,
+                    const double* du2, const lapack_int* ipiv, const double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cgtrfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* dl,
+                    const lapack_complex_float* d,
+                    const lapack_complex_float* du,
+                    const lapack_complex_float* dlf,
+                    const lapack_complex_float* df,
+                    const lapack_complex_float* duf,
+                    const lapack_complex_float* du2, const lapack_int* ipiv,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgtrfs( char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* dl,
+                    const lapack_complex_double* d,
+                    const lapack_complex_double* du,
+                    const lapack_complex_double* dlf,
+                    const lapack_complex_double* df,
+                    const lapack_complex_double* duf,
+                    const lapack_complex_double* du2, const lapack_int* ipiv,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sporfs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* a,
+                    lapack_int* lda, const float* af, lapack_int* ldaf,
+                    const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dporfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, const double* af,
+                    lapack_int* ldaf, const double* b, lapack_int* ldb,
+                    double* x, lapack_int* ldx, double* ferr, double* berr,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cporfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* af, lapack_int* ldaf,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zporfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* af, lapack_int* ldaf,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_dporfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const double* a, lapack_int* lda, const double* af,
+                     lapack_int* ldaf, const double* s, const double* b,
+                     lapack_int* ldb, double* x, lapack_int* ldx, double* rcond,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params, double* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_sporfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const float* a, lapack_int* lda, const float* af,
+                     lapack_int* ldaf, const float* s, const float* b,
+                     lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                     float* berr, lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_zporfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_complex_double* af, lapack_int* ldaf,
+                     const double* s, const lapack_complex_double* b,
+                     lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                     double* rcond, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cporfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_float* a, lapack_int* lda,
+                     const lapack_complex_float* af, lapack_int* ldaf,
+                     const float* s, const lapack_complex_float* b,
+                     lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                     float* rcond, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_spprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* ap, const float* afp, const float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* ferr,
+                    float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dpprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* ap, const double* afp, const double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cpprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap,
+                    const lapack_complex_float* afp,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zpprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap,
+                    const lapack_complex_double* afp,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_spbrfs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const float* ab, lapack_int* ldab, const float* afb,
+                    lapack_int* ldafb, const float* b, lapack_int* ldb,
+                    float* x, lapack_int* ldx, float* ferr, float* berr,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dpbrfs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const double* ab, lapack_int* ldab, const double* afb,
+                    lapack_int* ldafb, const double* b, lapack_int* ldb,
+                    double* x, lapack_int* ldx, double* ferr, double* berr,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cpbrfs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    const lapack_complex_float* afb, lapack_int* ldafb,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zpbrfs( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    const lapack_complex_double* afb, lapack_int* ldafb,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sptrfs( lapack_int* n, lapack_int* nrhs, const float* d,
+                    const float* e, const float* df, const float* ef,
+                    const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                    float* ferr, float* berr, float* work, lapack_int *info );
+void LAPACK_dptrfs( lapack_int* n, lapack_int* nrhs, const double* d,
+                    const double* e, const double* df, const double* ef,
+                    const double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* ferr, double* berr, double* work,
+                    lapack_int *info );
+void LAPACK_cptrfs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* d,
+                    const lapack_complex_float* e, const float* df,
+                    const lapack_complex_float* ef,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zptrfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* d, const lapack_complex_double* e,
+                    const double* df, const lapack_complex_double* ef,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_ssyrfs( char* uplo, lapack_int* n, lapack_int* nrhs, const float* a,
+                    lapack_int* lda, const float* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const float* b, lapack_int* ldb,
+                    float* x, lapack_int* ldx, float* ferr, float* berr,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dsyrfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, const double* af,
+                    lapack_int* ldaf, const lapack_int* ipiv, const double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_csyrfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zsyrfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_dsyrfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const double* a, lapack_int* lda, const double* af,
+                     lapack_int* ldaf, const lapack_int* ipiv, const double* s,
+                     const double* b, lapack_int* ldb, double* x,
+                     lapack_int* ldx, double* rcond, double* berr,
+                     lapack_int* n_err_bnds, double* err_bnds_norm,
+                     double* err_bnds_comp, lapack_int* nparams, double* params,
+                     double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_ssyrfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const float* a, lapack_int* lda, const float* af,
+                     lapack_int* ldaf, const lapack_int* ipiv, const float* s,
+                     const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                     float* rcond, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params, float* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_zsyrfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_complex_double* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const double* s,
+                     const lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_csyrfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_float* a, lapack_int* lda,
+                     const lapack_complex_float* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const float* s,
+                     const lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* berr, lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_cherfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zherfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* af, lapack_int* ldaf,
+                    const lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_zherfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_complex_double* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const double* s,
+                     const lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cherfsx( char* uplo, char* equed, lapack_int* n, lapack_int* nrhs,
+                     const lapack_complex_float* a, lapack_int* lda,
+                     const lapack_complex_float* af, lapack_int* ldaf,
+                     const lapack_int* ipiv, const float* s,
+                     const lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* berr, lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_ssprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* ap, const float* afp, const lapack_int* ipiv,
+                    const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dsprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* ap, const double* afp, const lapack_int* ipiv,
+                    const double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* ferr, double* berr, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_csprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap,
+                    const lapack_complex_float* afp, const lapack_int* ipiv,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zsprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap,
+                    const lapack_complex_double* afp, const lapack_int* ipiv,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_chprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap,
+                    const lapack_complex_float* afp, const lapack_int* ipiv,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zhprfs( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap,
+                    const lapack_complex_double* afp, const lapack_int* ipiv,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* ferr,
+                    double* berr, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_strrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const float* a, lapack_int* lda,
+                    const float* b, lapack_int* ldb, const float* x,
+                    lapack_int* ldx, float* ferr, float* berr, float* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dtrrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const double* a, lapack_int* lda,
+                    const double* b, lapack_int* ldb, const double* x,
+                    lapack_int* ldx, double* ferr, double* berr, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ctrrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* b,
+                    lapack_int* ldb, const lapack_complex_float* x,
+                    lapack_int* ldx, float* ferr, float* berr,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztrrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* b,
+                    lapack_int* ldb, const lapack_complex_double* x,
+                    lapack_int* ldx, double* ferr, double* berr,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_stprfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const float* ap, const float* b,
+                    lapack_int* ldb, const float* x, lapack_int* ldx,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dtprfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const double* ap, const double* b,
+                    lapack_int* ldb, const double* x, lapack_int* ldx,
+                    double* ferr, double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_ctprfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_float* ap,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    const lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztprfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* nrhs, const lapack_complex_double* ap,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    const lapack_complex_double* x, lapack_int* ldx,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_stbrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs, const float* ab,
+                    lapack_int* ldab, const float* b, lapack_int* ldb,
+                    const float* x, lapack_int* ldx, float* ferr, float* berr,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dtbrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs, const double* ab,
+                    lapack_int* ldab, const double* b, lapack_int* ldb,
+                    const double* x, lapack_int* ldx, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_ctbrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_float* ab, lapack_int* ldab,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    const lapack_complex_float* x, lapack_int* ldx, float* ferr,
+                    float* berr, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztbrfs( char* uplo, char* trans, char* diag, lapack_int* n,
+                    lapack_int* kd, lapack_int* nrhs,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    const lapack_complex_double* x, lapack_int* ldx,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_sgetri( lapack_int* n, float* a, lapack_int* lda,
+                    const lapack_int* ipiv, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgetri( lapack_int* n, double* a, lapack_int* lda,
+                    const lapack_int* ipiv, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgetri( lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zgetri( lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    const lapack_int* ipiv, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_spotri( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dpotri( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_cpotri( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_zpotri( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_dpftri( char* transr, char* uplo, lapack_int* n, double* a,
+                    lapack_int *info );
+void LAPACK_spftri( char* transr, char* uplo, lapack_int* n, float* a,
+                    lapack_int *info );
+void LAPACK_zpftri( char* transr, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int *info );
+void LAPACK_cpftri( char* transr, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int *info );
+void LAPACK_spptri( char* uplo, lapack_int* n, float* ap, lapack_int *info );
+void LAPACK_dpptri( char* uplo, lapack_int* n, double* ap, lapack_int *info );
+void LAPACK_cpptri( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    lapack_int *info );
+void LAPACK_zpptri( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    lapack_int *info );
+void LAPACK_ssytri( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    const lapack_int* ipiv, float* work, lapack_int *info );
+void LAPACK_dsytri( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    const lapack_int* ipiv, double* work, lapack_int *info );
+void LAPACK_csytri( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, const lapack_int* ipiv,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zsytri( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, const lapack_int* ipiv,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_chetri( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, const lapack_int* ipiv,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zhetri( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, const lapack_int* ipiv,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_ssptri( char* uplo, lapack_int* n, float* ap,
+                    const lapack_int* ipiv, float* work, lapack_int *info );
+void LAPACK_dsptri( char* uplo, lapack_int* n, double* ap,
+                    const lapack_int* ipiv, double* work, lapack_int *info );
+void LAPACK_csptri( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    const lapack_int* ipiv, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zsptri( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    const lapack_int* ipiv, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_chptri( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    const lapack_int* ipiv, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zhptri( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    const lapack_int* ipiv, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_strtri( char* uplo, char* diag, lapack_int* n, float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_dtrtri( char* uplo, char* diag, lapack_int* n, double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_ctrtri( char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_ztrtri( char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dtftri( char* transr, char* uplo, char* diag, lapack_int* n,
+                    double* a, lapack_int *info );
+void LAPACK_stftri( char* transr, char* uplo, char* diag, lapack_int* n,
+                    float* a, lapack_int *info );
+void LAPACK_ztftri( char* transr, char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_double* a, lapack_int *info );
+void LAPACK_ctftri( char* transr, char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_float* a, lapack_int *info );
+void LAPACK_stptri( char* uplo, char* diag, lapack_int* n, float* ap,
+                    lapack_int *info );
+void LAPACK_dtptri( char* uplo, char* diag, lapack_int* n, double* ap,
+                    lapack_int *info );
+void LAPACK_ctptri( char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_float* ap, lapack_int *info );
+void LAPACK_ztptri( char* uplo, char* diag, lapack_int* n,
+                    lapack_complex_double* ap, lapack_int *info );
+void LAPACK_sgeequ( lapack_int* m, lapack_int* n, const float* a,
+                    lapack_int* lda, float* r, float* c, float* rowcnd,
+                    float* colcnd, float* amax, lapack_int *info );
+void LAPACK_dgeequ( lapack_int* m, lapack_int* n, const double* a,
+                    lapack_int* lda, double* r, double* c, double* rowcnd,
+                    double* colcnd, double* amax, lapack_int *info );
+void LAPACK_cgeequ( lapack_int* m, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, float* r, float* c, float* rowcnd,
+                    float* colcnd, float* amax, lapack_int *info );
+void LAPACK_zgeequ( lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda, double* r,
+                    double* c, double* rowcnd, double* colcnd, double* amax,
+                    lapack_int *info );
+void LAPACK_dgeequb( lapack_int* m, lapack_int* n, const double* a,
+                     lapack_int* lda, double* r, double* c, double* rowcnd,
+                     double* colcnd, double* amax, lapack_int *info );
+void LAPACK_sgeequb( lapack_int* m, lapack_int* n, const float* a,
+                     lapack_int* lda, float* r, float* c, float* rowcnd,
+                     float* colcnd, float* amax, lapack_int *info );
+void LAPACK_zgeequb( lapack_int* m, lapack_int* n,
+                     const lapack_complex_double* a, lapack_int* lda, double* r,
+                     double* c, double* rowcnd, double* colcnd, double* amax,
+                     lapack_int *info );
+void LAPACK_cgeequb( lapack_int* m, lapack_int* n,
+                     const lapack_complex_float* a, lapack_int* lda, float* r,
+                     float* c, float* rowcnd, float* colcnd, float* amax,
+                     lapack_int *info );
+void LAPACK_sgbequ( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const float* ab, lapack_int* ldab, float* r,
+                    float* c, float* rowcnd, float* colcnd, float* amax,
+                    lapack_int *info );
+void LAPACK_dgbequ( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const double* ab, lapack_int* ldab,
+                    double* r, double* c, double* rowcnd, double* colcnd,
+                    double* amax, lapack_int *info );
+void LAPACK_cgbequ( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const lapack_complex_float* ab,
+                    lapack_int* ldab, float* r, float* c, float* rowcnd,
+                    float* colcnd, float* amax, lapack_int *info );
+void LAPACK_zgbequ( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const lapack_complex_double* ab,
+                    lapack_int* ldab, double* r, double* c, double* rowcnd,
+                    double* colcnd, double* amax, lapack_int *info );
+void LAPACK_dgbequb( lapack_int* m, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, const double* ab, lapack_int* ldab,
+                     double* r, double* c, double* rowcnd, double* colcnd,
+                     double* amax, lapack_int *info );
+void LAPACK_sgbequb( lapack_int* m, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, const float* ab, lapack_int* ldab,
+                     float* r, float* c, float* rowcnd, float* colcnd,
+                     float* amax, lapack_int *info );
+void LAPACK_zgbequb( lapack_int* m, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, const lapack_complex_double* ab,
+                     lapack_int* ldab, double* r, double* c, double* rowcnd,
+                     double* colcnd, double* amax, lapack_int *info );
+void LAPACK_cgbequb( lapack_int* m, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, const lapack_complex_float* ab,
+                     lapack_int* ldab, float* r, float* c, float* rowcnd,
+                     float* colcnd, float* amax, lapack_int *info );
+void LAPACK_spoequ( lapack_int* n, const float* a, lapack_int* lda, float* s,
+                    float* scond, float* amax, lapack_int *info );
+void LAPACK_dpoequ( lapack_int* n, const double* a, lapack_int* lda, double* s,
+                    double* scond, double* amax, lapack_int *info );
+void LAPACK_cpoequ( lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, float* s, float* scond, float* amax,
+                    lapack_int *info );
+void LAPACK_zpoequ( lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, double* s, double* scond, double* amax,
+                    lapack_int *info );
+void LAPACK_dpoequb( lapack_int* n, const double* a, lapack_int* lda, double* s,
+                     double* scond, double* amax, lapack_int *info );
+void LAPACK_spoequb( lapack_int* n, const float* a, lapack_int* lda, float* s,
+                     float* scond, float* amax, lapack_int *info );
+void LAPACK_zpoequb( lapack_int* n, const lapack_complex_double* a,
+                     lapack_int* lda, double* s, double* scond, double* amax,
+                     lapack_int *info );
+void LAPACK_cpoequb( lapack_int* n, const lapack_complex_float* a,
+                     lapack_int* lda, float* s, float* scond, float* amax,
+                     lapack_int *info );
+void LAPACK_sppequ( char* uplo, lapack_int* n, const float* ap, float* s,
+                    float* scond, float* amax, lapack_int *info );
+void LAPACK_dppequ( char* uplo, lapack_int* n, const double* ap, double* s,
+                    double* scond, double* amax, lapack_int *info );
+void LAPACK_cppequ( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    float* s, float* scond, float* amax, lapack_int *info );
+void LAPACK_zppequ( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    double* s, double* scond, double* amax, lapack_int *info );
+void LAPACK_spbequ( char* uplo, lapack_int* n, lapack_int* kd, const float* ab,
+                    lapack_int* ldab, float* s, float* scond, float* amax,
+                    lapack_int *info );
+void LAPACK_dpbequ( char* uplo, lapack_int* n, lapack_int* kd, const double* ab,
+                    lapack_int* ldab, double* s, double* scond, double* amax,
+                    lapack_int *info );
+void LAPACK_cpbequ( char* uplo, lapack_int* n, lapack_int* kd,
+                    const lapack_complex_float* ab, lapack_int* ldab, float* s,
+                    float* scond, float* amax, lapack_int *info );
+void LAPACK_zpbequ( char* uplo, lapack_int* n, lapack_int* kd,
+                    const lapack_complex_double* ab, lapack_int* ldab,
+                    double* s, double* scond, double* amax, lapack_int *info );
+void LAPACK_dsyequb( char* uplo, lapack_int* n, const double* a,
+                     lapack_int* lda, double* s, double* scond, double* amax,
+                     double* work, lapack_int *info );
+void LAPACK_ssyequb( char* uplo, lapack_int* n, const float* a, lapack_int* lda,
+                     float* s, float* scond, float* amax, float* work,
+                     lapack_int *info );
+void LAPACK_zsyequb( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                     lapack_int* lda, double* s, double* scond, double* amax,
+                     lapack_complex_double* work, lapack_int *info );
+void LAPACK_csyequb( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                     lapack_int* lda, float* s, float* scond, float* amax,
+                     lapack_complex_float* work, lapack_int *info );
+void LAPACK_zheequb( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                     lapack_int* lda, double* s, double* scond, double* amax,
+                     lapack_complex_double* work, lapack_int *info );
+void LAPACK_cheequb( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                     lapack_int* lda, float* s, float* scond, float* amax,
+                     lapack_complex_float* work, lapack_int *info );
+void LAPACK_sgesv( lapack_int* n, lapack_int* nrhs, float* a, lapack_int* lda,
+                   lapack_int* ipiv, float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dgesv( lapack_int* n, lapack_int* nrhs, double* a, lapack_int* lda,
+                   lapack_int* ipiv, double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_cgesv( lapack_int* n, lapack_int* nrhs, lapack_complex_float* a,
+                   lapack_int* lda, lapack_int* ipiv, lapack_complex_float* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_zgesv( lapack_int* n, lapack_int* nrhs, lapack_complex_double* a,
+                   lapack_int* lda, lapack_int* ipiv, lapack_complex_double* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_dsgesv( lapack_int* n, lapack_int* nrhs, double* a, lapack_int* lda,
+                    lapack_int* ipiv, double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* work, float* swork,
+                    lapack_int* iter, lapack_int *info );
+void LAPACK_zcgesv( lapack_int* n, lapack_int* nrhs, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ipiv, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    lapack_complex_double* work, lapack_complex_float* swork,
+                    double* rwork, lapack_int* iter, lapack_int *info );
+void LAPACK_sgesvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    float* a, lapack_int* lda, float* af, lapack_int* ldaf,
+                    lapack_int* ipiv, char* equed, float* r, float* c, float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgesvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    double* a, lapack_int* lda, double* af, lapack_int* ldaf,
+                    lapack_int* ipiv, char* equed, double* r, double* c,
+                    double* b, lapack_int* ldb, double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cgesvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* af, lapack_int* ldaf,
+                    lapack_int* ipiv, char* equed, float* r, float* c,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zgesvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* af, lapack_int* ldaf,
+                    lapack_int* ipiv, char* equed, double* r, double* c,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_dgesvxx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                     double* a, lapack_int* lda, double* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, double* r, double* c,
+                     double* b, lapack_int* ldb, double* x, lapack_int* ldx,
+                     double* rcond, double* rpvgrw, double* berr,
+                     lapack_int* n_err_bnds, double* err_bnds_norm,
+                     double* err_bnds_comp, lapack_int* nparams, double* params,
+                     double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_sgesvxx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                     float* a, lapack_int* lda, float* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, float* r, float* c,
+                     float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                     float* rcond, float* rpvgrw, float* berr,
+                     lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_zgesvxx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_double* a, lapack_int* lda,
+                     lapack_complex_double* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, double* r, double* c,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cgesvxx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_float* a, lapack_int* lda,
+                     lapack_complex_float* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, float* r, float* c,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sgbsv( lapack_int* n, lapack_int* kl, lapack_int* ku,
+                   lapack_int* nrhs, float* ab, lapack_int* ldab,
+                   lapack_int* ipiv, float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dgbsv( lapack_int* n, lapack_int* kl, lapack_int* ku,
+                   lapack_int* nrhs, double* ab, lapack_int* ldab,
+                   lapack_int* ipiv, double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_cgbsv( lapack_int* n, lapack_int* kl, lapack_int* ku,
+                   lapack_int* nrhs, lapack_complex_float* ab, lapack_int* ldab,
+                   lapack_int* ipiv, lapack_complex_float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_zgbsv( lapack_int* n, lapack_int* kl, lapack_int* ku,
+                   lapack_int* nrhs, lapack_complex_double* ab,
+                   lapack_int* ldab, lapack_int* ipiv, lapack_complex_double* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_sgbsvx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_int* nrhs, float* ab,
+                    lapack_int* ldab, float* afb, lapack_int* ldafb,
+                    lapack_int* ipiv, char* equed, float* r, float* c, float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgbsvx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_int* nrhs, double* ab,
+                    lapack_int* ldab, double* afb, lapack_int* ldafb,
+                    lapack_int* ipiv, char* equed, double* r, double* c,
+                    double* b, lapack_int* ldb, double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cgbsvx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_int* nrhs, lapack_complex_float* ab,
+                    lapack_int* ldab, lapack_complex_float* afb,
+                    lapack_int* ldafb, lapack_int* ipiv, char* equed, float* r,
+                    float* c, lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zgbsvx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, lapack_int* nrhs, lapack_complex_double* ab,
+                    lapack_int* ldab, lapack_complex_double* afb,
+                    lapack_int* ldafb, lapack_int* ipiv, char* equed, double* r,
+                    double* c, lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_dgbsvxx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs, double* ab,
+                     lapack_int* ldab, double* afb, lapack_int* ldafb,
+                     lapack_int* ipiv, char* equed, double* r, double* c,
+                     double* b, lapack_int* ldb, double* x, lapack_int* ldx,
+                     double* rcond, double* rpvgrw, double* berr,
+                     lapack_int* n_err_bnds, double* err_bnds_norm,
+                     double* err_bnds_comp, lapack_int* nparams, double* params,
+                     double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_sgbsvxx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs, float* ab,
+                     lapack_int* ldab, float* afb, lapack_int* ldafb,
+                     lapack_int* ipiv, char* equed, float* r, float* c,
+                     float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                     float* rcond, float* rpvgrw, float* berr,
+                     lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_zgbsvxx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs,
+                     lapack_complex_double* ab, lapack_int* ldab,
+                     lapack_complex_double* afb, lapack_int* ldafb,
+                     lapack_int* ipiv, char* equed, double* r, double* c,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cgbsvxx( char* fact, char* trans, lapack_int* n, lapack_int* kl,
+                     lapack_int* ku, lapack_int* nrhs, lapack_complex_float* ab,
+                     lapack_int* ldab, lapack_complex_float* afb,
+                     lapack_int* ldafb, lapack_int* ipiv, char* equed, float* r,
+                     float* c, lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sgtsv( lapack_int* n, lapack_int* nrhs, float* dl, float* d,
+                   float* du, float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_dgtsv( lapack_int* n, lapack_int* nrhs, double* dl, double* d,
+                   double* du, double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_cgtsv( lapack_int* n, lapack_int* nrhs, lapack_complex_float* dl,
+                   lapack_complex_float* d, lapack_complex_float* du,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_zgtsv( lapack_int* n, lapack_int* nrhs, lapack_complex_double* dl,
+                   lapack_complex_double* d, lapack_complex_double* du,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_sgtsvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    const float* dl, const float* d, const float* du,
+                    float* dlf, float* df, float* duf, float* du2,
+                    lapack_int* ipiv, const float* b, lapack_int* ldb, float* x,
+                    lapack_int* ldx, float* rcond, float* ferr, float* berr,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dgtsvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    const double* dl, const double* d, const double* du,
+                    double* dlf, double* df, double* duf, double* du2,
+                    lapack_int* ipiv, const double* b, lapack_int* ldb,
+                    double* x, lapack_int* ldx, double* rcond, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cgtsvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* dl,
+                    const lapack_complex_float* d,
+                    const lapack_complex_float* du, lapack_complex_float* dlf,
+                    lapack_complex_float* df, lapack_complex_float* duf,
+                    lapack_complex_float* du2, lapack_int* ipiv,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zgtsvx( char* fact, char* trans, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* dl,
+                    const lapack_complex_double* d,
+                    const lapack_complex_double* du, lapack_complex_double* dlf,
+                    lapack_complex_double* df, lapack_complex_double* duf,
+                    lapack_complex_double* du2, lapack_int* ipiv,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_sposv( char* uplo, lapack_int* n, lapack_int* nrhs, float* a,
+                   lapack_int* lda, float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dposv( char* uplo, lapack_int* n, lapack_int* nrhs, double* a,
+                   lapack_int* lda, double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_cposv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_zposv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dsposv( char* uplo, lapack_int* n, lapack_int* nrhs, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* work, float* swork,
+                    lapack_int* iter, lapack_int *info );
+void LAPACK_zcposv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx,
+                    lapack_complex_double* work, lapack_complex_float* swork,
+                    double* rwork, lapack_int* iter, lapack_int *info );
+void LAPACK_sposvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    float* a, lapack_int* lda, float* af, lapack_int* ldaf,
+                    char* equed, float* s, float* b, lapack_int* ldb, float* x,
+                    lapack_int* ldx, float* rcond, float* ferr, float* berr,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dposvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    double* a, lapack_int* lda, double* af, lapack_int* ldaf,
+                    char* equed, double* s, double* b, lapack_int* ldb,
+                    double* x, lapack_int* ldx, double* rcond, double* ferr,
+                    double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cposvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* af, lapack_int* ldaf, char* equed,
+                    float* s, lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zposvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* af, lapack_int* ldaf, char* equed,
+                    double* s, lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_dposvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     double* a, lapack_int* lda, double* af, lapack_int* ldaf,
+                     char* equed, double* s, double* b, lapack_int* ldb,
+                     double* x, lapack_int* ldx, double* rcond, double* rpvgrw,
+                     double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params, double* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_sposvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     float* a, lapack_int* lda, float* af, lapack_int* ldaf,
+                     char* equed, float* s, float* b, lapack_int* ldb, float* x,
+                     lapack_int* ldx, float* rcond, float* rpvgrw, float* berr,
+                     lapack_int* n_err_bnds, float* err_bnds_norm,
+                     float* err_bnds_comp, lapack_int* nparams, float* params,
+                     float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_zposvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_double* a, lapack_int* lda,
+                     lapack_complex_double* af, lapack_int* ldaf, char* equed,
+                     double* s, lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_cposvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_float* a, lapack_int* lda,
+                     lapack_complex_float* af, lapack_int* ldaf, char* equed,
+                     float* s, lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sppsv( char* uplo, lapack_int* n, lapack_int* nrhs, float* ap,
+                   float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_dppsv( char* uplo, lapack_int* n, lapack_int* nrhs, double* ap,
+                   double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_cppsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* ap, lapack_complex_float* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_zppsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* ap, lapack_complex_double* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_sppsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    float* ap, float* afp, char* equed, float* s, float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dppsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    double* ap, double* afp, char* equed, double* s, double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cppsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* ap, lapack_complex_float* afp,
+                    char* equed, float* s, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zppsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* ap, lapack_complex_double* afp,
+                    char* equed, double* s, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_spbsv( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                   float* ab, lapack_int* ldab, float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dpbsv( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                   double* ab, lapack_int* ldab, double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_cpbsv( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                   lapack_complex_float* ab, lapack_int* ldab,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_zpbsv( char* uplo, lapack_int* n, lapack_int* kd, lapack_int* nrhs,
+                   lapack_complex_double* ab, lapack_int* ldab,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_spbsvx( char* fact, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_int* nrhs, float* ab, lapack_int* ldab, float* afb,
+                    lapack_int* ldafb, char* equed, float* s, float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dpbsvx( char* fact, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_int* nrhs, double* ab, lapack_int* ldab, double* afb,
+                    lapack_int* ldafb, char* equed, double* s, double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_cpbsvx( char* fact, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_int* nrhs, lapack_complex_float* ab,
+                    lapack_int* ldab, lapack_complex_float* afb,
+                    lapack_int* ldafb, char* equed, float* s,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zpbsvx( char* fact, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_int* nrhs, lapack_complex_double* ab,
+                    lapack_int* ldab, lapack_complex_double* afb,
+                    lapack_int* ldafb, char* equed, double* s,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_sptsv( lapack_int* n, lapack_int* nrhs, float* d, float* e,
+                   float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_dptsv( lapack_int* n, lapack_int* nrhs, double* d, double* e,
+                   double* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_cptsv( lapack_int* n, lapack_int* nrhs, float* d,
+                   lapack_complex_float* e, lapack_complex_float* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_zptsv( lapack_int* n, lapack_int* nrhs, double* d,
+                   lapack_complex_double* e, lapack_complex_double* b,
+                   lapack_int* ldb, lapack_int *info );
+void LAPACK_sptsvx( char* fact, lapack_int* n, lapack_int* nrhs, const float* d,
+                    const float* e, float* df, float* ef, const float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int *info );
+void LAPACK_dptsvx( char* fact, lapack_int* n, lapack_int* nrhs,
+                    const double* d, const double* e, double* df, double* ef,
+                    const double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* rcond, double* ferr, double* berr,
+                    double* work, lapack_int *info );
+void LAPACK_cptsvx( char* fact, lapack_int* n, lapack_int* nrhs, const float* d,
+                    const lapack_complex_float* e, float* df,
+                    lapack_complex_float* ef, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zptsvx( char* fact, lapack_int* n, lapack_int* nrhs,
+                    const double* d, const lapack_complex_double* e, double* df,
+                    lapack_complex_double* ef, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_ssysv( char* uplo, lapack_int* n, lapack_int* nrhs, float* a,
+                   lapack_int* lda, lapack_int* ipiv, float* b, lapack_int* ldb,
+                   float* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_dsysv( char* uplo, lapack_int* n, lapack_int* nrhs, double* a,
+                   lapack_int* lda, lapack_int* ipiv, double* b,
+                   lapack_int* ldb, double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_csysv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* a, lapack_int* lda, lapack_int* ipiv,
+                   lapack_complex_float* b, lapack_int* ldb,
+                   lapack_complex_float* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_zsysv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* a, lapack_int* lda, lapack_int* ipiv,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_ssysvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* a, lapack_int* lda, float* af,
+                    lapack_int* ldaf, lapack_int* ipiv, const float* b,
+                    lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                    float* ferr, float* berr, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dsysvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* a, lapack_int* lda, double* af,
+                    lapack_int* ldaf, lapack_int* ipiv, const double* b,
+                    lapack_int* ldb, double* x, lapack_int* ldx, double* rcond,
+                    double* ferr, double* berr, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_csysvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* af, lapack_int* ldaf,
+                    lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int *info );
+void LAPACK_zsysvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* af, lapack_int* ldaf,
+                    lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_dsysvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     double* a, lapack_int* lda, double* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, double* s, double* b,
+                     lapack_int* ldb, double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params, double* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_ssysvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     float* a, lapack_int* lda, float* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, float* s, float* b,
+                     lapack_int* ldb, float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params, float* work,
+                     lapack_int* iwork, lapack_int *info );
+void LAPACK_zsysvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_double* a, lapack_int* lda,
+                     lapack_complex_double* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, double* s,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_csysvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_float* a, lapack_int* lda,
+                     lapack_complex_float* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, float* s,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_chesv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* a, lapack_int* lda, lapack_int* ipiv,
+                   lapack_complex_float* b, lapack_int* ldb,
+                   lapack_complex_float* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_zhesv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* a, lapack_int* lda, lapack_int* ipiv,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_chesvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* af, lapack_int* ldaf,
+                    lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int *info );
+void LAPACK_zhesvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* af, lapack_int* ldaf,
+                    lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_zhesvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_double* a, lapack_int* lda,
+                     lapack_complex_double* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, double* s,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* x, lapack_int* ldx, double* rcond,
+                     double* rpvgrw, double* berr, lapack_int* n_err_bnds,
+                     double* err_bnds_norm, double* err_bnds_comp,
+                     lapack_int* nparams, double* params,
+                     lapack_complex_double* work, double* rwork,
+                     lapack_int *info );
+void LAPACK_chesvxx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                     lapack_complex_float* a, lapack_int* lda,
+                     lapack_complex_float* af, lapack_int* ldaf,
+                     lapack_int* ipiv, char* equed, float* s,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* x, lapack_int* ldx, float* rcond,
+                     float* rpvgrw, float* berr, lapack_int* n_err_bnds,
+                     float* err_bnds_norm, float* err_bnds_comp,
+                     lapack_int* nparams, float* params,
+                     lapack_complex_float* work, float* rwork,
+                     lapack_int *info );
+void LAPACK_sspsv( char* uplo, lapack_int* n, lapack_int* nrhs, float* ap,
+                   lapack_int* ipiv, float* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_dspsv( char* uplo, lapack_int* n, lapack_int* nrhs, double* ap,
+                   lapack_int* ipiv, double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_cspsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* ap, lapack_int* ipiv,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_zspsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* ap, lapack_int* ipiv,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_sspsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const float* ap, float* afp, lapack_int* ipiv,
+                    const float* b, lapack_int* ldb, float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr, float* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dspsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const double* ap, double* afp, lapack_int* ipiv,
+                    const double* b, lapack_int* ldb, double* x,
+                    lapack_int* ldx, double* rcond, double* ferr, double* berr,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cspsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap, lapack_complex_float* afp,
+                    lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zspsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap, lapack_complex_double* afp,
+                    lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_chpsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* ap, lapack_int* ipiv,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int *info );
+void LAPACK_zhpsv( char* uplo, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* ap, lapack_int* ipiv,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_int *info );
+void LAPACK_chpsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_float* ap, lapack_complex_float* afp,
+                    lapack_int* ipiv, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* x, lapack_int* ldx,
+                    float* rcond, float* ferr, float* berr,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zhpsvx( char* fact, char* uplo, lapack_int* n, lapack_int* nrhs,
+                    const lapack_complex_double* ap, lapack_complex_double* afp,
+                    lapack_int* ipiv, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* x, lapack_int* ldx,
+                    double* rcond, double* ferr, double* berr,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sgeqrf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgeqrf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgeqrf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zgeqrf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgeqpf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* jpvt, float* tau, float* work,
+                    lapack_int *info );
+void LAPACK_dgeqpf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* jpvt, double* tau, double* work,
+                    lapack_int *info );
+void LAPACK_cgeqpf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* jpvt,
+                    lapack_complex_float* tau, lapack_complex_float* work,
+                    float* rwork, lapack_int *info );
+void LAPACK_zgeqpf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* jpvt,
+                    lapack_complex_double* tau, lapack_complex_double* work,
+                    double* rwork, lapack_int *info );
+void LAPACK_sgeqp3( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* jpvt, float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dgeqp3( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* jpvt, double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cgeqp3( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* jpvt,
+                    lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int *info );
+void LAPACK_zgeqp3( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* jpvt,
+                    lapack_complex_double* tau, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int *info );
+void LAPACK_sorgqr( lapack_int* m, lapack_int* n, lapack_int* k, float* a,
+                    lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorgqr( lapack_int* m, lapack_int* n, lapack_int* k, double* a,
+                    lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sormqr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const float* a, lapack_int* lda,
+                    const float* tau, float* c, lapack_int* ldc, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dormqr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const double* a, lapack_int* lda,
+                    const double* tau, double* c, lapack_int* ldc, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cungqr( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zungqr( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmqr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmqr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgelqf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgelqf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgelqf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zgelqf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sorglq( lapack_int* m, lapack_int* n, lapack_int* k, float* a,
+                    lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorglq( lapack_int* m, lapack_int* n, lapack_int* k, double* a,
+                    lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sormlq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const float* a, lapack_int* lda,
+                    const float* tau, float* c, lapack_int* ldc, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dormlq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const double* a, lapack_int* lda,
+                    const double* tau, double* c, lapack_int* ldc, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cunglq( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zunglq( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmlq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmlq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgeqlf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgeqlf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgeqlf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zgeqlf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sorgql( lapack_int* m, lapack_int* n, lapack_int* k, float* a,
+                    lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorgql( lapack_int* m, lapack_int* n, lapack_int* k, double* a,
+                    lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cungql( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zungql( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sormql( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const float* a, lapack_int* lda,
+                    const float* tau, float* c, lapack_int* ldc, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dormql( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const double* a, lapack_int* lda,
+                    const double* tau, double* c, lapack_int* ldc, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cunmql( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmql( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgerqf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgerqf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgerqf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zgerqf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sorgrq( lapack_int* m, lapack_int* n, lapack_int* k, float* a,
+                    lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorgrq( lapack_int* m, lapack_int* n, lapack_int* k, double* a,
+                    lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cungrq( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zungrq( lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sormrq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const float* a, lapack_int* lda,
+                    const float* tau, float* c, lapack_int* ldc, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dormrq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const double* a, lapack_int* lda,
+                    const double* tau, double* c, lapack_int* ldc, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cunmrq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmrq( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_stzrzf( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dtzrzf( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_ctzrzf( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_ztzrzf( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sormrz( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, lapack_int* l, const float* a,
+                    lapack_int* lda, const float* tau, float* c,
+                    lapack_int* ldc, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dormrz( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, lapack_int* l, const double* a,
+                    lapack_int* lda, const double* tau, double* c,
+                    lapack_int* ldc, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmrz( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, lapack_int* l, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmrz( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* k, lapack_int* l,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau, lapack_complex_double* c,
+                    lapack_int* ldc, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sggqrf( lapack_int* n, lapack_int* m, lapack_int* p, float* a,
+                    lapack_int* lda, float* taua, float* b, lapack_int* ldb,
+                    float* taub, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dggqrf( lapack_int* n, lapack_int* m, lapack_int* p, double* a,
+                    lapack_int* lda, double* taua, double* b, lapack_int* ldb,
+                    double* taub, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cggqrf( lapack_int* n, lapack_int* m, lapack_int* p,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* taua, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* taub,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zggqrf( lapack_int* n, lapack_int* m, lapack_int* p,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* taua, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* taub,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sggrqf( lapack_int* m, lapack_int* p, lapack_int* n, float* a,
+                    lapack_int* lda, float* taua, float* b, lapack_int* ldb,
+                    float* taub, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dggrqf( lapack_int* m, lapack_int* p, lapack_int* n, double* a,
+                    lapack_int* lda, double* taua, double* b, lapack_int* ldb,
+                    double* taub, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cggrqf( lapack_int* m, lapack_int* p, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* taua, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* taub,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zggrqf( lapack_int* m, lapack_int* p, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* taua, lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* taub,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgebrd( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* d, float* e, float* tauq, float* taup, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dgebrd( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* d, double* e, double* tauq, double* taup,
+                    double* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_cgebrd( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, float* d, float* e,
+                    lapack_complex_float* tauq, lapack_complex_float* taup,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zgebrd( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, double* d, double* e,
+                    lapack_complex_double* tauq, lapack_complex_double* taup,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sgbbrd( char* vect, lapack_int* m, lapack_int* n, lapack_int* ncc,
+                    lapack_int* kl, lapack_int* ku, float* ab, lapack_int* ldab,
+                    float* d, float* e, float* q, lapack_int* ldq, float* pt,
+                    lapack_int* ldpt, float* c, lapack_int* ldc, float* work,
+                    lapack_int *info );
+void LAPACK_dgbbrd( char* vect, lapack_int* m, lapack_int* n, lapack_int* ncc,
+                    lapack_int* kl, lapack_int* ku, double* ab,
+                    lapack_int* ldab, double* d, double* e, double* q,
+                    lapack_int* ldq, double* pt, lapack_int* ldpt, double* c,
+                    lapack_int* ldc, double* work, lapack_int *info );
+void LAPACK_cgbbrd( char* vect, lapack_int* m, lapack_int* n, lapack_int* ncc,
+                    lapack_int* kl, lapack_int* ku, lapack_complex_float* ab,
+                    lapack_int* ldab, float* d, float* e,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* pt, lapack_int* ldpt,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgbbrd( char* vect, lapack_int* m, lapack_int* n, lapack_int* ncc,
+                    lapack_int* kl, lapack_int* ku, lapack_complex_double* ab,
+                    lapack_int* ldab, double* d, double* e,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* pt, lapack_int* ldpt,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_sorgbr( char* vect, lapack_int* m, lapack_int* n, lapack_int* k,
+                    float* a, lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorgbr( char* vect, lapack_int* m, lapack_int* n, lapack_int* k,
+                    double* a, lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sormbr( char* vect, char* side, char* trans, lapack_int* m,
+                    lapack_int* n, lapack_int* k, const float* a,
+                    lapack_int* lda, const float* tau, float* c,
+                    lapack_int* ldc, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dormbr( char* vect, char* side, char* trans, lapack_int* m,
+                    lapack_int* n, lapack_int* k, const double* a,
+                    lapack_int* lda, const double* tau, double* c,
+                    lapack_int* ldc, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cungbr( char* vect, lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zungbr( char* vect, lapack_int* m, lapack_int* n, lapack_int* k,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmbr( char* vect, char* side, char* trans, lapack_int* m,
+                    lapack_int* n, lapack_int* k, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmbr( char* vect, char* side, char* trans, lapack_int* m,
+                    lapack_int* n, lapack_int* k,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau, lapack_complex_double* c,
+                    lapack_int* ldc, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sbdsqr( char* uplo, lapack_int* n, lapack_int* ncvt,
+                    lapack_int* nru, lapack_int* ncc, float* d, float* e,
+                    float* vt, lapack_int* ldvt, float* u, lapack_int* ldu,
+                    float* c, lapack_int* ldc, float* work, lapack_int *info );
+void LAPACK_dbdsqr( char* uplo, lapack_int* n, lapack_int* ncvt,
+                    lapack_int* nru, lapack_int* ncc, double* d, double* e,
+                    double* vt, lapack_int* ldvt, double* u, lapack_int* ldu,
+                    double* c, lapack_int* ldc, double* work,
+                    lapack_int *info );
+void LAPACK_cbdsqr( char* uplo, lapack_int* n, lapack_int* ncvt,
+                    lapack_int* nru, lapack_int* ncc, float* d, float* e,
+                    lapack_complex_float* vt, lapack_int* ldvt,
+                    lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* c, lapack_int* ldc, float* work,
+                    lapack_int *info );
+void LAPACK_zbdsqr( char* uplo, lapack_int* n, lapack_int* ncvt,
+                    lapack_int* nru, lapack_int* ncc, double* d, double* e,
+                    lapack_complex_double* vt, lapack_int* ldvt,
+                    lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* c, lapack_int* ldc, double* work,
+                    lapack_int *info );
+void LAPACK_sbdsdc( char* uplo, char* compq, lapack_int* n, float* d, float* e,
+                    float* u, lapack_int* ldu, float* vt, lapack_int* ldvt,
+                    float* q, lapack_int* iq, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dbdsdc( char* uplo, char* compq, lapack_int* n, double* d,
+                    double* e, double* u, lapack_int* ldu, double* vt,
+                    lapack_int* ldvt, double* q, lapack_int* iq, double* work,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ssytrd( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    float* d, float* e, float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dsytrd( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    double* d, double* e, double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sorgtr( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    const float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dorgtr( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    const double* tau, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_sormtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const float* a, lapack_int* lda,
+                    const float* tau, float* c, lapack_int* ldc, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dormtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const double* a, lapack_int* lda,
+                    const double* tau, double* c, lapack_int* ldc, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_chetrd( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, float* d, float* e,
+                    lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zhetrd( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, double* d, double* e,
+                    lapack_complex_double* tau, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cungtr( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zungtr( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_zunmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_ssptrd( char* uplo, lapack_int* n, float* ap, float* d, float* e,
+                    float* tau, lapack_int *info );
+void LAPACK_dsptrd( char* uplo, lapack_int* n, double* ap, double* d, double* e,
+                    double* tau, lapack_int *info );
+void LAPACK_sopgtr( char* uplo, lapack_int* n, const float* ap,
+                    const float* tau, float* q, lapack_int* ldq, float* work,
+                    lapack_int *info );
+void LAPACK_dopgtr( char* uplo, lapack_int* n, const double* ap,
+                    const double* tau, double* q, lapack_int* ldq, double* work,
+                    lapack_int *info );
+void LAPACK_sopmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const float* ap, const float* tau, float* c,
+                    lapack_int* ldc, float* work, lapack_int *info );
+void LAPACK_dopmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const double* ap, const double* tau,
+                    double* c, lapack_int* ldc, double* work,
+                    lapack_int *info );
+void LAPACK_chptrd( char* uplo, lapack_int* n, lapack_complex_float* ap,
+                    float* d, float* e, lapack_complex_float* tau,
+                    lapack_int *info );
+void LAPACK_zhptrd( char* uplo, lapack_int* n, lapack_complex_double* ap,
+                    double* d, double* e, lapack_complex_double* tau,
+                    lapack_int *info );
+void LAPACK_cupgtr( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    const lapack_complex_float* tau, lapack_complex_float* q,
+                    lapack_int* ldq, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zupgtr( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    const lapack_complex_double* tau, lapack_complex_double* q,
+                    lapack_int* ldq, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_cupmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const lapack_complex_float* ap,
+                    const lapack_complex_float* tau, lapack_complex_float* c,
+                    lapack_int* ldc, lapack_complex_float* work,
+                    lapack_int *info );
+void LAPACK_zupmtr( char* side, char* uplo, char* trans, lapack_int* m,
+                    lapack_int* n, const lapack_complex_double* ap,
+                    const lapack_complex_double* tau, lapack_complex_double* c,
+                    lapack_int* ldc, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_ssbtrd( char* vect, char* uplo, lapack_int* n, lapack_int* kd,
+                    float* ab, lapack_int* ldab, float* d, float* e, float* q,
+                    lapack_int* ldq, float* work, lapack_int *info );
+void LAPACK_dsbtrd( char* vect, char* uplo, lapack_int* n, lapack_int* kd,
+                    double* ab, lapack_int* ldab, double* d, double* e,
+                    double* q, lapack_int* ldq, double* work,
+                    lapack_int *info );
+void LAPACK_chbtrd( char* vect, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_float* ab, lapack_int* ldab, float* d,
+                    float* e, lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zhbtrd( char* vect, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_double* ab, lapack_int* ldab, double* d,
+                    double* e, lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_ssterf( lapack_int* n, float* d, float* e, lapack_int *info );
+void LAPACK_dsterf( lapack_int* n, double* d, double* e, lapack_int *info );
+void LAPACK_ssteqr( char* compz, lapack_int* n, float* d, float* e, float* z,
+                    lapack_int* ldz, float* work, lapack_int *info );
+void LAPACK_dsteqr( char* compz, lapack_int* n, double* d, double* e, double* z,
+                    lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_csteqr( char* compz, lapack_int* n, float* d, float* e,
+                    lapack_complex_float* z, lapack_int* ldz, float* work,
+                    lapack_int *info );
+void LAPACK_zsteqr( char* compz, lapack_int* n, double* d, double* e,
+                    lapack_complex_double* z, lapack_int* ldz, double* work,
+                    lapack_int *info );
+void LAPACK_sstemr( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    lapack_int* m, float* w, float* z, lapack_int* ldz,
+                    lapack_int* nzc, lapack_int* isuppz, lapack_logical* tryrac,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_dstemr( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, lapack_int* m, double* w, double* z,
+                    lapack_int* ldz, lapack_int* nzc, lapack_int* isuppz,
+                    lapack_logical* tryrac, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_cstemr( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_int* nzc, lapack_int* isuppz,
+                    lapack_logical* tryrac, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_zstemr( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, lapack_int* m, double* w,
+                    lapack_complex_double* z, lapack_int* ldz, lapack_int* nzc,
+                    lapack_int* isuppz, lapack_logical* tryrac, double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_sstedc( char* compz, lapack_int* n, float* d, float* e, float* z,
+                    lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dstedc( char* compz, lapack_int* n, double* d, double* e, double* z,
+                    lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_cstedc( char* compz, lapack_int* n, float* d, float* e,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zstedc( char* compz, lapack_int* n, double* d, double* e,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_sstegr( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w, float* z,
+                    lapack_int* ldz, lapack_int* isuppz, float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_dstegr( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, lapack_int* isuppz,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_cstegr( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_int* isuppz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_zstegr( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_int* isuppz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_spteqr( char* compz, lapack_int* n, float* d, float* e, float* z,
+                    lapack_int* ldz, float* work, lapack_int *info );
+void LAPACK_dpteqr( char* compz, lapack_int* n, double* d, double* e, double* z,
+                    lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_cpteqr( char* compz, lapack_int* n, float* d, float* e,
+                    lapack_complex_float* z, lapack_int* ldz, float* work,
+                    lapack_int *info );
+void LAPACK_zpteqr( char* compz, lapack_int* n, double* d, double* e,
+                    lapack_complex_double* z, lapack_int* ldz, double* work,
+                    lapack_int *info );
+void LAPACK_sstebz( char* range, char* order, lapack_int* n, float* vl,
+                    float* vu, lapack_int* il, lapack_int* iu, float* abstol,
+                    const float* d, const float* e, lapack_int* m,
+                    lapack_int* nsplit, float* w, lapack_int* iblock,
+                    lapack_int* isplit, float* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dstebz( char* range, char* order, lapack_int* n, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    const double* d, const double* e, lapack_int* m,
+                    lapack_int* nsplit, double* w, lapack_int* iblock,
+                    lapack_int* isplit, double* work, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_sstein( lapack_int* n, const float* d, const float* e,
+                    lapack_int* m, const float* w, const lapack_int* iblock,
+                    const lapack_int* isplit, float* z, lapack_int* ldz,
+                    float* work, lapack_int* iwork, lapack_int* ifailv,
+                    lapack_int *info );
+void LAPACK_dstein( lapack_int* n, const double* d, const double* e,
+                    lapack_int* m, const double* w, const lapack_int* iblock,
+                    const lapack_int* isplit, double* z, lapack_int* ldz,
+                    double* work, lapack_int* iwork, lapack_int* ifailv,
+                    lapack_int *info );
+void LAPACK_cstein( lapack_int* n, const float* d, const float* e,
+                    lapack_int* m, const float* w, const lapack_int* iblock,
+                    const lapack_int* isplit, lapack_complex_float* z,
+                    lapack_int* ldz, float* work, lapack_int* iwork,
+                    lapack_int* ifailv, lapack_int *info );
+void LAPACK_zstein( lapack_int* n, const double* d, const double* e,
+                    lapack_int* m, const double* w, const lapack_int* iblock,
+                    const lapack_int* isplit, lapack_complex_double* z,
+                    lapack_int* ldz, double* work, lapack_int* iwork,
+                    lapack_int* ifailv, lapack_int *info );
+void LAPACK_sdisna( char* job, lapack_int* m, lapack_int* n, const float* d,
+                    float* sep, lapack_int *info );
+void LAPACK_ddisna( char* job, lapack_int* m, lapack_int* n, const double* d,
+                    double* sep, lapack_int *info );
+void LAPACK_ssygst( lapack_int* itype, char* uplo, lapack_int* n, float* a,
+                    lapack_int* lda, const float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_dsygst( lapack_int* itype, char* uplo, lapack_int* n, double* a,
+                    lapack_int* lda, const double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_chegst( lapack_int* itype, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_zhegst( lapack_int* itype, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int *info );
+void LAPACK_sspgst( lapack_int* itype, char* uplo, lapack_int* n, float* ap,
+                    const float* bp, lapack_int *info );
+void LAPACK_dspgst( lapack_int* itype, char* uplo, lapack_int* n, double* ap,
+                    const double* bp, lapack_int *info );
+void LAPACK_chpgst( lapack_int* itype, char* uplo, lapack_int* n,
+                    lapack_complex_float* ap, const lapack_complex_float* bp,
+                    lapack_int *info );
+void LAPACK_zhpgst( lapack_int* itype, char* uplo, lapack_int* n,
+                    lapack_complex_double* ap, const lapack_complex_double* bp,
+                    lapack_int *info );
+void LAPACK_ssbgst( char* vect, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, float* ab, lapack_int* ldab,
+                    const float* bb, lapack_int* ldbb, float* x,
+                    lapack_int* ldx, float* work, lapack_int *info );
+void LAPACK_dsbgst( char* vect, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, double* ab, lapack_int* ldab,
+                    const double* bb, lapack_int* ldbb, double* x,
+                    lapack_int* ldx, double* work, lapack_int *info );
+void LAPACK_chbgst( char* vect, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, lapack_complex_float* ab, lapack_int* ldab,
+                    const lapack_complex_float* bb, lapack_int* ldbb,
+                    lapack_complex_float* x, lapack_int* ldx,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_zhbgst( char* vect, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, lapack_complex_double* ab, lapack_int* ldab,
+                    const lapack_complex_double* bb, lapack_int* ldbb,
+                    lapack_complex_double* x, lapack_int* ldx,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_spbstf( char* uplo, lapack_int* n, lapack_int* kb, float* bb,
+                    lapack_int* ldbb, lapack_int *info );
+void LAPACK_dpbstf( char* uplo, lapack_int* n, lapack_int* kb, double* bb,
+                    lapack_int* ldbb, lapack_int *info );
+void LAPACK_cpbstf( char* uplo, lapack_int* n, lapack_int* kb,
+                    lapack_complex_float* bb, lapack_int* ldbb,
+                    lapack_int *info );
+void LAPACK_zpbstf( char* uplo, lapack_int* n, lapack_int* kb,
+                    lapack_complex_double* bb, lapack_int* ldbb,
+                    lapack_int *info );
+void LAPACK_sgehrd( lapack_int* n, lapack_int* ilo, lapack_int* ihi, float* a,
+                    lapack_int* lda, float* tau, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgehrd( lapack_int* n, lapack_int* ilo, lapack_int* ihi, double* a,
+                    lapack_int* lda, double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cgehrd( lapack_int* n, lapack_int* ilo, lapack_int* ihi,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zgehrd( lapack_int* n, lapack_int* ilo, lapack_int* ihi,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* tau, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sorghr( lapack_int* n, lapack_int* ilo, lapack_int* ihi, float* a,
+                    lapack_int* lda, const float* tau, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dorghr( lapack_int* n, lapack_int* ilo, lapack_int* ihi, double* a,
+                    lapack_int* lda, const double* tau, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sormhr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, const float* a,
+                    lapack_int* lda, const float* tau, float* c,
+                    lapack_int* ldc, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dormhr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, const double* a,
+                    lapack_int* lda, const double* tau, double* c,
+                    lapack_int* ldc, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunghr( lapack_int* n, lapack_int* ilo, lapack_int* ihi,
+                    lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zunghr( lapack_int* n, lapack_int* ilo, lapack_int* ihi,
+                    lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cunmhr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* tau, lapack_complex_float* c,
+                    lapack_int* ldc, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zunmhr( char* side, char* trans, lapack_int* m, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* tau, lapack_complex_double* c,
+                    lapack_int* ldc, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sgebal( char* job, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* ilo, lapack_int* ihi, float* scale,
+                    lapack_int *info );
+void LAPACK_dgebal( char* job, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* ilo, lapack_int* ihi, double* scale,
+                    lapack_int *info );
+void LAPACK_cgebal( char* job, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* ilo, lapack_int* ihi,
+                    float* scale, lapack_int *info );
+void LAPACK_zgebal( char* job, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ilo, lapack_int* ihi,
+                    double* scale, lapack_int *info );
+void LAPACK_sgebak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const float* scale, lapack_int* m,
+                    float* v, lapack_int* ldv, lapack_int *info );
+void LAPACK_dgebak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const double* scale, lapack_int* m,
+                    double* v, lapack_int* ldv, lapack_int *info );
+void LAPACK_cgebak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const float* scale, lapack_int* m,
+                    lapack_complex_float* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_zgebak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const double* scale, lapack_int* m,
+                    lapack_complex_double* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_shseqr( char* job, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, float* h, lapack_int* ldh, float* wr,
+                    float* wi, float* z, lapack_int* ldz, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dhseqr( char* job, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, double* h, lapack_int* ldh, double* wr,
+                    double* wi, double* z, lapack_int* ldz, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_chseqr( char* job, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, lapack_complex_float* h, lapack_int* ldh,
+                    lapack_complex_float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zhseqr( char* job, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, lapack_complex_double* h, lapack_int* ldh,
+                    lapack_complex_double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_shsein( char* job, char* eigsrc, char* initv,
+                    lapack_logical* select, lapack_int* n, const float* h,
+                    lapack_int* ldh, float* wr, const float* wi, float* vl,
+                    lapack_int* ldvl, float* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, float* work,
+                    lapack_int* ifaill, lapack_int* ifailr, lapack_int *info );
+void LAPACK_dhsein( char* job, char* eigsrc, char* initv,
+                    lapack_logical* select, lapack_int* n, const double* h,
+                    lapack_int* ldh, double* wr, const double* wi, double* vl,
+                    lapack_int* ldvl, double* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, double* work,
+                    lapack_int* ifaill, lapack_int* ifailr, lapack_int *info );
+void LAPACK_chsein( char* job, char* eigsrc, char* initv,
+                    const lapack_logical* select, lapack_int* n,
+                    const lapack_complex_float* h, lapack_int* ldh,
+                    lapack_complex_float* w, lapack_complex_float* vl,
+                    lapack_int* ldvl, lapack_complex_float* vr,
+                    lapack_int* ldvr, lapack_int* mm, lapack_int* m,
+                    lapack_complex_float* work, float* rwork,
+                    lapack_int* ifaill, lapack_int* ifailr, lapack_int *info );
+void LAPACK_zhsein( char* job, char* eigsrc, char* initv,
+                    const lapack_logical* select, lapack_int* n,
+                    const lapack_complex_double* h, lapack_int* ldh,
+                    lapack_complex_double* w, lapack_complex_double* vl,
+                    lapack_int* ldvl, lapack_complex_double* vr,
+                    lapack_int* ldvr, lapack_int* mm, lapack_int* m,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int* ifaill, lapack_int* ifailr, lapack_int *info );
+void LAPACK_strevc( char* side, char* howmny, lapack_logical* select,
+                    lapack_int* n, const float* t, lapack_int* ldt, float* vl,
+                    lapack_int* ldvl, float* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, float* work,
+                    lapack_int *info );
+void LAPACK_dtrevc( char* side, char* howmny, lapack_logical* select,
+                    lapack_int* n, const double* t, lapack_int* ldt, double* vl,
+                    lapack_int* ldvl, double* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, double* work,
+                    lapack_int *info );
+void LAPACK_ctrevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* vl, lapack_int* ldvl,
+                    lapack_complex_float* vr, lapack_int* ldvr, lapack_int* mm,
+                    lapack_int* m, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztrevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* vl, lapack_int* ldvl,
+                    lapack_complex_double* vr, lapack_int* ldvr, lapack_int* mm,
+                    lapack_int* m, lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_strsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const float* t, lapack_int* ldt,
+                    const float* vl, lapack_int* ldvl, const float* vr,
+                    lapack_int* ldvr, float* s, float* sep, lapack_int* mm,
+                    lapack_int* m, float* work, lapack_int* ldwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dtrsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const double* t, lapack_int* ldt,
+                    const double* vl, lapack_int* ldvl, const double* vr,
+                    lapack_int* ldvr, double* s, double* sep, lapack_int* mm,
+                    lapack_int* m, double* work, lapack_int* ldwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ctrsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_float* t,
+                    lapack_int* ldt, const lapack_complex_float* vl,
+                    lapack_int* ldvl, const lapack_complex_float* vr,
+                    lapack_int* ldvr, float* s, float* sep, lapack_int* mm,
+                    lapack_int* m, lapack_complex_float* work,
+                    lapack_int* ldwork, float* rwork, lapack_int *info );
+void LAPACK_ztrsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_double* t,
+                    lapack_int* ldt, const lapack_complex_double* vl,
+                    lapack_int* ldvl, const lapack_complex_double* vr,
+                    lapack_int* ldvr, double* s, double* sep, lapack_int* mm,
+                    lapack_int* m, lapack_complex_double* work,
+                    lapack_int* ldwork, double* rwork, lapack_int *info );
+void LAPACK_strexc( char* compq, lapack_int* n, float* t, lapack_int* ldt,
+                    float* q, lapack_int* ldq, lapack_int* ifst,
+                    lapack_int* ilst, float* work, lapack_int *info );
+void LAPACK_dtrexc( char* compq, lapack_int* n, double* t, lapack_int* ldt,
+                    double* q, lapack_int* ldq, lapack_int* ifst,
+                    lapack_int* ilst, double* work, lapack_int *info );
+void LAPACK_ctrexc( char* compq, lapack_int* n, lapack_complex_float* t,
+                    lapack_int* ldt, lapack_complex_float* q, lapack_int* ldq,
+                    lapack_int* ifst, lapack_int* ilst, lapack_int *info );
+void LAPACK_ztrexc( char* compq, lapack_int* n, lapack_complex_double* t,
+                    lapack_int* ldt, lapack_complex_double* q, lapack_int* ldq,
+                    lapack_int* ifst, lapack_int* ilst, lapack_int *info );
+void LAPACK_strsen( char* job, char* compq, const lapack_logical* select,
+                    lapack_int* n, float* t, lapack_int* ldt, float* q,
+                    lapack_int* ldq, float* wr, float* wi, lapack_int* m,
+                    float* s, float* sep, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dtrsen( char* job, char* compq, const lapack_logical* select,
+                    lapack_int* n, double* t, lapack_int* ldt, double* q,
+                    lapack_int* ldq, double* wr, double* wi, lapack_int* m,
+                    double* s, double* sep, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_ctrsen( char* job, char* compq, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* w, lapack_int* m, float* s,
+                    float* sep, lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_ztrsen( char* job, char* compq, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* w, lapack_int* m, double* s,
+                    double* sep, lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_strsyl( char* trana, char* tranb, lapack_int* isgn, lapack_int* m,
+                    lapack_int* n, const float* a, lapack_int* lda,
+                    const float* b, lapack_int* ldb, float* c, lapack_int* ldc,
+                    float* scale, lapack_int *info );
+void LAPACK_dtrsyl( char* trana, char* tranb, lapack_int* isgn, lapack_int* m,
+                    lapack_int* n, const double* a, lapack_int* lda,
+                    const double* b, lapack_int* ldb, double* c,
+                    lapack_int* ldc, double* scale, lapack_int *info );
+void LAPACK_ctrsyl( char* trana, char* tranb, lapack_int* isgn, lapack_int* m,
+                    lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* b,
+                    lapack_int* ldb, lapack_complex_float* c, lapack_int* ldc,
+                    float* scale, lapack_int *info );
+void LAPACK_ztrsyl( char* trana, char* tranb, lapack_int* isgn, lapack_int* m,
+                    lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* b,
+                    lapack_int* ldb, lapack_complex_double* c, lapack_int* ldc,
+                    double* scale, lapack_int *info );
+void LAPACK_sgghrd( char* compq, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, float* q, lapack_int* ldq, float* z,
+                    lapack_int* ldz, lapack_int *info );
+void LAPACK_dgghrd( char* compq, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, double* q, lapack_int* ldq, double* z,
+                    lapack_int* ldz, lapack_int *info );
+void LAPACK_cgghrd( char* compq, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_int *info );
+void LAPACK_zgghrd( char* compq, char* compz, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_int *info );
+void LAPACK_sggbal( char* job, lapack_int* n, float* a, lapack_int* lda,
+                    float* b, lapack_int* ldb, lapack_int* ilo, lapack_int* ihi,
+                    float* lscale, float* rscale, float* work,
+                    lapack_int *info );
+void LAPACK_dggbal( char* job, lapack_int* n, double* a, lapack_int* lda,
+                    double* b, lapack_int* ldb, lapack_int* ilo,
+                    lapack_int* ihi, double* lscale, double* rscale,
+                    double* work, lapack_int *info );
+void LAPACK_cggbal( char* job, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* b, lapack_int* ldb,
+                    lapack_int* ilo, lapack_int* ihi, float* lscale,
+                    float* rscale, float* work, lapack_int *info );
+void LAPACK_zggbal( char* job, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* b, lapack_int* ldb,
+                    lapack_int* ilo, lapack_int* ihi, double* lscale,
+                    double* rscale, double* work, lapack_int *info );
+void LAPACK_sggbak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const float* lscale, const float* rscale,
+                    lapack_int* m, float* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_dggbak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const double* lscale, const double* rscale,
+                    lapack_int* m, double* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_cggbak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const float* lscale, const float* rscale,
+                    lapack_int* m, lapack_complex_float* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_zggbak( char* job, char* side, lapack_int* n, lapack_int* ilo,
+                    lapack_int* ihi, const double* lscale, const double* rscale,
+                    lapack_int* m, lapack_complex_double* v, lapack_int* ldv,
+                    lapack_int *info );
+void LAPACK_shgeqz( char* job, char* compq, char* compz, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, float* h, lapack_int* ldh,
+                    float* t, lapack_int* ldt, float* alphar, float* alphai,
+                    float* beta, float* q, lapack_int* ldq, float* z,
+                    lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dhgeqz( char* job, char* compq, char* compz, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, double* h,
+                    lapack_int* ldh, double* t, lapack_int* ldt, double* alphar,
+                    double* alphai, double* beta, double* q, lapack_int* ldq,
+                    double* z, lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_chgeqz( char* job, char* compq, char* compz, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, lapack_complex_float* h,
+                    lapack_int* ldh, lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* alpha, lapack_complex_float* beta,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int *info );
+void LAPACK_zhgeqz( char* job, char* compq, char* compz, lapack_int* n,
+                    lapack_int* ilo, lapack_int* ihi, lapack_complex_double* h,
+                    lapack_int* ldh, lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* alpha, lapack_complex_double* beta,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_stgevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const float* s, lapack_int* lds,
+                    const float* p, lapack_int* ldp, float* vl,
+                    lapack_int* ldvl, float* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, float* work,
+                    lapack_int *info );
+void LAPACK_dtgevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const double* s, lapack_int* lds,
+                    const double* p, lapack_int* ldp, double* vl,
+                    lapack_int* ldvl, double* vr, lapack_int* ldvr,
+                    lapack_int* mm, lapack_int* m, double* work,
+                    lapack_int *info );
+void LAPACK_ctgevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_float* s,
+                    lapack_int* lds, const lapack_complex_float* p,
+                    lapack_int* ldp, lapack_complex_float* vl, lapack_int* ldvl,
+                    lapack_complex_float* vr, lapack_int* ldvr, lapack_int* mm,
+                    lapack_int* m, lapack_complex_float* work, float* rwork,
+                    lapack_int *info );
+void LAPACK_ztgevc( char* side, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_double* s,
+                    lapack_int* lds, const lapack_complex_double* p,
+                    lapack_int* ldp, lapack_complex_double* vl,
+                    lapack_int* ldvl, lapack_complex_double* vr,
+                    lapack_int* ldvr, lapack_int* mm, lapack_int* m,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int *info );
+void LAPACK_stgexc( lapack_logical* wantq, lapack_logical* wantz, lapack_int* n,
+                    float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                    float* q, lapack_int* ldq, float* z, lapack_int* ldz,
+                    lapack_int* ifst, lapack_int* ilst, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dtgexc( lapack_logical* wantq, lapack_logical* wantz, lapack_int* n,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    double* q, lapack_int* ldq, double* z, lapack_int* ldz,
+                    lapack_int* ifst, lapack_int* ilst, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_ctgexc( lapack_logical* wantq, lapack_logical* wantz, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* z, lapack_int* ldz, lapack_int* ifst,
+                    lapack_int* ilst, lapack_int *info );
+void LAPACK_ztgexc( lapack_logical* wantq, lapack_logical* wantz, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* z, lapack_int* ldz, lapack_int* ifst,
+                    lapack_int* ilst, lapack_int *info );
+void LAPACK_stgsen( lapack_int* ijob, lapack_logical* wantq,
+                    lapack_logical* wantz, const lapack_logical* select,
+                    lapack_int* n, float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, float* alphar, float* alphai, float* beta,
+                    float* q, lapack_int* ldq, float* z, lapack_int* ldz,
+                    lapack_int* m, float* pl, float* pr, float* dif,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_dtgsen( lapack_int* ijob, lapack_logical* wantq,
+                    lapack_logical* wantz, const lapack_logical* select,
+                    lapack_int* n, double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, double* alphar, double* alphai,
+                    double* beta, double* q, lapack_int* ldq, double* z,
+                    lapack_int* ldz, lapack_int* m, double* pl, double* pr,
+                    double* dif, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_ctgsen( lapack_int* ijob, lapack_logical* wantq,
+                    lapack_logical* wantz, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* alpha, lapack_complex_float* beta,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* z, lapack_int* ldz, lapack_int* m,
+                    float* pl, float* pr, float* dif,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_ztgsen( lapack_int* ijob, lapack_logical* wantq,
+                    lapack_logical* wantz, const lapack_logical* select,
+                    lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* alpha, lapack_complex_double* beta,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* z, lapack_int* ldz, lapack_int* m,
+                    double* pl, double* pr, double* dif,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_stgsyl( char* trans, lapack_int* ijob, lapack_int* m, lapack_int* n,
+                    const float* a, lapack_int* lda, const float* b,
+                    lapack_int* ldb, float* c, lapack_int* ldc, const float* d,
+                    lapack_int* ldd, const float* e, lapack_int* lde, float* f,
+                    lapack_int* ldf, float* scale, float* dif, float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_dtgsyl( char* trans, lapack_int* ijob, lapack_int* m, lapack_int* n,
+                    const double* a, lapack_int* lda, const double* b,
+                    lapack_int* ldb, double* c, lapack_int* ldc,
+                    const double* d, lapack_int* ldd, const double* e,
+                    lapack_int* lde, double* f, lapack_int* ldf, double* scale,
+                    double* dif, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ctgsyl( char* trans, lapack_int* ijob, lapack_int* m, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    const lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    const lapack_complex_float* d, lapack_int* ldd,
+                    const lapack_complex_float* e, lapack_int* lde,
+                    lapack_complex_float* f, lapack_int* ldf, float* scale,
+                    float* dif, lapack_complex_float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ztgsyl( char* trans, lapack_int* ijob, lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    const lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    const lapack_complex_double* d, lapack_int* ldd,
+                    const lapack_complex_double* e, lapack_int* lde,
+                    lapack_complex_double* f, lapack_int* ldf, double* scale,
+                    double* dif, lapack_complex_double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_stgsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const float* a, lapack_int* lda,
+                    const float* b, lapack_int* ldb, const float* vl,
+                    lapack_int* ldvl, const float* vr, lapack_int* ldvr,
+                    float* s, float* dif, lapack_int* mm, lapack_int* m,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dtgsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const double* a, lapack_int* lda,
+                    const double* b, lapack_int* ldb, const double* vl,
+                    lapack_int* ldvl, const double* vr, lapack_int* ldvr,
+                    double* s, double* dif, lapack_int* mm, lapack_int* m,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_ctgsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, const lapack_complex_float* b,
+                    lapack_int* ldb, const lapack_complex_float* vl,
+                    lapack_int* ldvl, const lapack_complex_float* vr,
+                    lapack_int* ldvr, float* s, float* dif, lapack_int* mm,
+                    lapack_int* m, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_ztgsna( char* job, char* howmny, const lapack_logical* select,
+                    lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, const lapack_complex_double* b,
+                    lapack_int* ldb, const lapack_complex_double* vl,
+                    lapack_int* ldvl, const lapack_complex_double* vr,
+                    lapack_int* ldvr, double* s, double* dif, lapack_int* mm,
+                    lapack_int* m, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_sggsvp( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, float* a, lapack_int* lda,
+                    float* b, lapack_int* ldb, float* tola, float* tolb,
+                    lapack_int* k, lapack_int* l, float* u, lapack_int* ldu,
+                    float* v, lapack_int* ldv, float* q, lapack_int* ldq,
+                    lapack_int* iwork, float* tau, float* work,
+                    lapack_int *info );
+void LAPACK_dggsvp( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, double* a, lapack_int* lda,
+                    double* b, lapack_int* ldb, double* tola, double* tolb,
+                    lapack_int* k, lapack_int* l, double* u, lapack_int* ldu,
+                    double* v, lapack_int* ldv, double* q, lapack_int* ldq,
+                    lapack_int* iwork, double* tau, double* work,
+                    lapack_int *info );
+void LAPACK_cggsvp( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* b, lapack_int* ldb,
+                    float* tola, float* tolb, lapack_int* k, lapack_int* l,
+                    lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* v, lapack_int* ldv,
+                    lapack_complex_float* q, lapack_int* ldq, lapack_int* iwork,
+                    float* rwork, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zggsvp( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* b, lapack_int* ldb,
+                    double* tola, double* tolb, lapack_int* k, lapack_int* l,
+                    lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* v, lapack_int* ldv,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_int* iwork, double* rwork,
+                    lapack_complex_double* tau, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_stgsja( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_int* k, lapack_int* l,
+                    float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                    float* tola, float* tolb, float* alpha, float* beta,
+                    float* u, lapack_int* ldu, float* v, lapack_int* ldv,
+                    float* q, lapack_int* ldq, float* work, lapack_int* ncycle,
+                    lapack_int *info );
+void LAPACK_dtgsja( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_int* k, lapack_int* l,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    double* tola, double* tolb, double* alpha, double* beta,
+                    double* u, lapack_int* ldu, double* v, lapack_int* ldv,
+                    double* q, lapack_int* ldq, double* work,
+                    lapack_int* ncycle, lapack_int *info );
+void LAPACK_ctgsja( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_int* k, lapack_int* l,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* tola,
+                    float* tolb, float* alpha, float* beta,
+                    lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* v, lapack_int* ldv,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* work, lapack_int* ncycle,
+                    lapack_int *info );
+void LAPACK_ztgsja( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* p, lapack_int* n, lapack_int* k, lapack_int* l,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* tola,
+                    double* tolb, double* alpha, double* beta,
+                    lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* v, lapack_int* ldv,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* work, lapack_int* ncycle,
+                    lapack_int *info );
+void LAPACK_sgels( char* trans, lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                   float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                   float* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_dgels( char* trans, lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                   double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                   double* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_cgels( char* trans, lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* b, lapack_int* ldb,
+                   lapack_complex_float* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_zgels( char* trans, lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_sgelsy( lapack_int* m, lapack_int* n, lapack_int* nrhs, float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb,
+                    lapack_int* jpvt, float* rcond, lapack_int* rank,
+                    float* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_dgelsy( lapack_int* m, lapack_int* n, lapack_int* nrhs, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb,
+                    lapack_int* jpvt, double* rcond, lapack_int* rank,
+                    double* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_cgelsy( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, lapack_int* jpvt,
+                    float* rcond, lapack_int* rank, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int *info );
+void LAPACK_zgelsy( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, lapack_int* jpvt,
+                    double* rcond, lapack_int* rank,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_sgelss( lapack_int* m, lapack_int* n, lapack_int* nrhs, float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb, float* s,
+                    float* rcond, lapack_int* rank, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dgelss( lapack_int* m, lapack_int* n, lapack_int* nrhs, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb, double* s,
+                    double* rcond, lapack_int* rank, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cgelss( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* s,
+                    float* rcond, lapack_int* rank, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int *info );
+void LAPACK_zgelss( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* s,
+                    double* rcond, lapack_int* rank,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_sgelsd( lapack_int* m, lapack_int* n, lapack_int* nrhs, float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb, float* s,
+                    float* rcond, lapack_int* rank, float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_dgelsd( lapack_int* m, lapack_int* n, lapack_int* nrhs, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb, double* s,
+                    double* rcond, lapack_int* rank, double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_cgelsd( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* s,
+                    float* rcond, lapack_int* rank, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_zgelsd( lapack_int* m, lapack_int* n, lapack_int* nrhs,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* s,
+                    double* rcond, lapack_int* rank,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_sgglse( lapack_int* m, lapack_int* n, lapack_int* p, float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb, float* c,
+                    float* d, float* x, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dgglse( lapack_int* m, lapack_int* n, lapack_int* p, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb, double* c,
+                    double* d, double* x, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cgglse( lapack_int* m, lapack_int* n, lapack_int* p,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* c, lapack_complex_float* d,
+                    lapack_complex_float* x, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zgglse( lapack_int* m, lapack_int* n, lapack_int* p,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* c, lapack_complex_double* d,
+                    lapack_complex_double* x, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sggglm( lapack_int* n, lapack_int* m, lapack_int* p, float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb, float* d,
+                    float* x, float* y, float* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_dggglm( lapack_int* n, lapack_int* m, lapack_int* p, double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb, double* d,
+                    double* x, double* y, double* work, lapack_int* lwork,
+                    lapack_int *info );
+void LAPACK_cggglm( lapack_int* n, lapack_int* m, lapack_int* p,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* d, lapack_complex_float* x,
+                    lapack_complex_float* y, lapack_complex_float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_zggglm( lapack_int* n, lapack_int* m, lapack_int* p,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* d, lapack_complex_double* x,
+                    lapack_complex_double* y, lapack_complex_double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_ssyev( char* jobz, char* uplo, lapack_int* n, float* a,
+                   lapack_int* lda, float* w, float* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_dsyev( char* jobz, char* uplo, lapack_int* n, double* a,
+                   lapack_int* lda, double* w, double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_cheev( char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_float* a, lapack_int* lda, float* w,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_int *info );
+void LAPACK_zheev( char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_double* a, lapack_int* lda, double* w,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   double* rwork, lapack_int *info );
+void LAPACK_ssyevd( char* jobz, char* uplo, lapack_int* n, float* a,
+                    lapack_int* lda, float* w, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dsyevd( char* jobz, char* uplo, lapack_int* n, double* a,
+                    lapack_int* lda, double* w, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_cheevd( char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda, float* w,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zheevd( char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda, double* w,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_ssyevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    float* a, lapack_int* lda, float* vl, float* vu,
+                    lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, float* z, lapack_int* ldz,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_dsyevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    double* a, lapack_int* lda, double* vl, double* vu,
+                    lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, double* z, lapack_int* ldz,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_cheevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda, float* vl,
+                    float* vu, lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_zheevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_ssyevr( char* jobz, char* range, char* uplo, lapack_int* n,
+                    float* a, lapack_int* lda, float* vl, float* vu,
+                    lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, float* z, lapack_int* ldz,
+                    lapack_int* isuppz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dsyevr( char* jobz, char* range, char* uplo, lapack_int* n,
+                    double* a, lapack_int* lda, double* vl, double* vu,
+                    lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, double* z, lapack_int* ldz,
+                    lapack_int* isuppz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_cheevr( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda, float* vl,
+                    float* vu, lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_int* isuppz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zheevr( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_int* isuppz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_sspev( char* jobz, char* uplo, lapack_int* n, float* ap, float* w,
+                   float* z, lapack_int* ldz, float* work, lapack_int *info );
+void LAPACK_dspev( char* jobz, char* uplo, lapack_int* n, double* ap, double* w,
+                   double* z, lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_chpev( char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_float* ap, float* w, lapack_complex_float* z,
+                   lapack_int* ldz, lapack_complex_float* work, float* rwork,
+                   lapack_int *info );
+void LAPACK_zhpev( char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_double* ap, double* w,
+                   lapack_complex_double* z, lapack_int* ldz,
+                   lapack_complex_double* work, double* rwork,
+                   lapack_int *info );
+void LAPACK_sspevd( char* jobz, char* uplo, lapack_int* n, float* ap, float* w,
+                    float* z, lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dspevd( char* jobz, char* uplo, lapack_int* n, double* ap,
+                    double* w, double* z, lapack_int* ldz, double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_chpevd( char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_float* ap, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int* lrwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_zhpevd( char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_double* ap, double* w,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_sspevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    float* ap, float* vl, float* vu, lapack_int* il,
+                    lapack_int* iu, float* abstol, lapack_int* m, float* w,
+                    float* z, lapack_int* ldz, float* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_dspevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    double* ap, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, double* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_chpevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_float* ap, float* vl, float* vu,
+                    lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work, float* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_zhpevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_complex_double* ap, double* vl, double* vu,
+                    lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work, double* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_ssbev( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                   float* ab, lapack_int* ldab, float* w, float* z,
+                   lapack_int* ldz, float* work, lapack_int *info );
+void LAPACK_dsbev( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                   double* ab, lapack_int* ldab, double* w, double* z,
+                   lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_chbev( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                   lapack_complex_float* ab, lapack_int* ldab, float* w,
+                   lapack_complex_float* z, lapack_int* ldz,
+                   lapack_complex_float* work, float* rwork, lapack_int *info );
+void LAPACK_zhbev( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                   lapack_complex_double* ab, lapack_int* ldab, double* w,
+                   lapack_complex_double* z, lapack_int* ldz,
+                   lapack_complex_double* work, double* rwork,
+                   lapack_int *info );
+void LAPACK_ssbevd( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                    float* ab, lapack_int* ldab, float* w, float* z,
+                    lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dsbevd( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                    double* ab, lapack_int* ldab, double* w, double* z,
+                    lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_chbevd( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_float* ab, lapack_int* ldab, float* w,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zhbevd( char* jobz, char* uplo, lapack_int* n, lapack_int* kd,
+                    lapack_complex_double* ab, lapack_int* ldab, double* w,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_ssbevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* kd, float* ab, lapack_int* ldab, float* q,
+                    lapack_int* ldq, float* vl, float* vu, lapack_int* il,
+                    lapack_int* iu, float* abstol, lapack_int* m, float* w,
+                    float* z, lapack_int* ldz, float* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_dsbevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* kd, double* ab, lapack_int* ldab, double* q,
+                    lapack_int* ldq, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, double* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_chbevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* kd, lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_complex_float* q, lapack_int* ldq, float* vl,
+                    float* vu, lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work, float* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_zhbevx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* kd, lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_complex_double* q, lapack_int* ldq, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work, double* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_sstev( char* jobz, lapack_int* n, float* d, float* e, float* z,
+                   lapack_int* ldz, float* work, lapack_int *info );
+void LAPACK_dstev( char* jobz, lapack_int* n, double* d, double* e, double* z,
+                   lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_sstevd( char* jobz, lapack_int* n, float* d, float* e, float* z,
+                    lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_dstevd( char* jobz, lapack_int* n, double* d, double* e, double* z,
+                    lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_sstevx( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w, float* z,
+                    lapack_int* ldz, float* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_dstevx( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, double* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_sstevr( char* jobz, char* range, lapack_int* n, float* d, float* e,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w, float* z,
+                    lapack_int* ldz, lapack_int* isuppz, float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_dstevr( char* jobz, char* range, lapack_int* n, double* d,
+                    double* e, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, lapack_int* isuppz,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_sgees( char* jobvs, char* sort, LAPACK_S_SELECT2 select,
+                   lapack_int* n, float* a, lapack_int* lda, lapack_int* sdim,
+                   float* wr, float* wi, float* vs, lapack_int* ldvs,
+                   float* work, lapack_int* lwork, lapack_logical* bwork,
+                   lapack_int *info );
+void LAPACK_dgees( char* jobvs, char* sort, LAPACK_D_SELECT2 select,
+                   lapack_int* n, double* a, lapack_int* lda, lapack_int* sdim,
+                   double* wr, double* wi, double* vs, lapack_int* ldvs,
+                   double* work, lapack_int* lwork, lapack_logical* bwork,
+                   lapack_int *info );
+void LAPACK_cgees( char* jobvs, char* sort, LAPACK_C_SELECT1 select,
+                   lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                   lapack_int* sdim, lapack_complex_float* w,
+                   lapack_complex_float* vs, lapack_int* ldvs,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_logical* bwork, lapack_int *info );
+void LAPACK_zgees( char* jobvs, char* sort, LAPACK_Z_SELECT1 select,
+                   lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                   lapack_int* sdim, lapack_complex_double* w,
+                   lapack_complex_double* vs, lapack_int* ldvs,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   double* rwork, lapack_logical* bwork, lapack_int *info );
+void LAPACK_sgeesx( char* jobvs, char* sort, LAPACK_S_SELECT2 select,
+                    char* sense, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* sdim, float* wr, float* wi, float* vs,
+                    lapack_int* ldvs, float* rconde, float* rcondv, float* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_logical* bwork, lapack_int *info );
+void LAPACK_dgeesx( char* jobvs, char* sort, LAPACK_D_SELECT2 select,
+                    char* sense, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* sdim, double* wr, double* wi, double* vs,
+                    lapack_int* ldvs, double* rconde, double* rcondv,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_cgeesx( char* jobvs, char* sort, LAPACK_C_SELECT1 select,
+                    char* sense, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* sdim, lapack_complex_float* w,
+                    lapack_complex_float* vs, lapack_int* ldvs, float* rconde,
+                    float* rcondv, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_zgeesx( char* jobvs, char* sort, LAPACK_Z_SELECT1 select,
+                    char* sense, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* sdim, lapack_complex_double* w,
+                    lapack_complex_double* vs, lapack_int* ldvs, double* rconde,
+                    double* rcondv, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_sgeev( char* jobvl, char* jobvr, lapack_int* n, float* a,
+                   lapack_int* lda, float* wr, float* wi, float* vl,
+                   lapack_int* ldvl, float* vr, lapack_int* ldvr, float* work,
+                   lapack_int* lwork, lapack_int *info );
+void LAPACK_dgeev( char* jobvl, char* jobvr, lapack_int* n, double* a,
+                   lapack_int* lda, double* wr, double* wi, double* vl,
+                   lapack_int* ldvl, double* vr, lapack_int* ldvr, double* work,
+                   lapack_int* lwork, lapack_int *info );
+void LAPACK_cgeev( char* jobvl, char* jobvr, lapack_int* n,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* w, lapack_complex_float* vl,
+                   lapack_int* ldvl, lapack_complex_float* vr, lapack_int* ldvr,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_int *info );
+void LAPACK_zgeev( char* jobvl, char* jobvr, lapack_int* n,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* w, lapack_complex_double* vl,
+                   lapack_int* ldvl, lapack_complex_double* vr,
+                   lapack_int* ldvr, lapack_complex_double* work,
+                   lapack_int* lwork, double* rwork, lapack_int *info );
+void LAPACK_sgeevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, float* a, lapack_int* lda, float* wr,
+                    float* wi, float* vl, lapack_int* ldvl, float* vr,
+                    lapack_int* ldvr, lapack_int* ilo, lapack_int* ihi,
+                    float* scale, float* abnrm, float* rconde, float* rcondv,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_dgeevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, double* a, lapack_int* lda, double* wr,
+                    double* wi, double* vl, lapack_int* ldvl, double* vr,
+                    lapack_int* ldvr, lapack_int* ilo, lapack_int* ihi,
+                    double* scale, double* abnrm, double* rconde,
+                    double* rcondv, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_cgeevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* w, lapack_complex_float* vl,
+                    lapack_int* ldvl, lapack_complex_float* vr,
+                    lapack_int* ldvr, lapack_int* ilo, lapack_int* ihi,
+                    float* scale, float* abnrm, float* rconde, float* rcondv,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgeevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* w, lapack_complex_double* vl,
+                    lapack_int* ldvl, lapack_complex_double* vr,
+                    lapack_int* ldvr, lapack_int* ilo, lapack_int* ihi,
+                    double* scale, double* abnrm, double* rconde,
+                    double* rcondv, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int *info );
+void LAPACK_sgesvd( char* jobu, char* jobvt, lapack_int* m, lapack_int* n,
+                    float* a, lapack_int* lda, float* s, float* u,
+                    lapack_int* ldu, float* vt, lapack_int* ldvt, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_dgesvd( char* jobu, char* jobvt, lapack_int* m, lapack_int* n,
+                    double* a, lapack_int* lda, double* s, double* u,
+                    lapack_int* ldu, double* vt, lapack_int* ldvt, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_cgesvd( char* jobu, char* jobvt, lapack_int* m, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda, float* s,
+                    lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* vt, lapack_int* ldvt,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int *info );
+void LAPACK_zgesvd( char* jobu, char* jobvt, lapack_int* m, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda, double* s,
+                    lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* vt, lapack_int* ldvt,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int *info );
+void LAPACK_sgesdd( char* jobz, lapack_int* m, lapack_int* n, float* a,
+                    lapack_int* lda, float* s, float* u, lapack_int* ldu,
+                    float* vt, lapack_int* ldvt, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dgesdd( char* jobz, lapack_int* m, lapack_int* n, double* a,
+                    lapack_int* lda, double* s, double* u, lapack_int* ldu,
+                    double* vt, lapack_int* ldvt, double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_cgesdd( char* jobz, lapack_int* m, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda, float* s,
+                    lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* vt, lapack_int* ldvt,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_zgesdd( char* jobz, lapack_int* m, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda, double* s,
+                    lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* vt, lapack_int* ldvt,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* iwork, lapack_int *info );
+void LAPACK_dgejsv( char* joba, char* jobu, char* jobv, char* jobr, char* jobt,
+                    char* jobp, lapack_int* m, lapack_int* n, double* a,
+                    lapack_int* lda, double* sva, double* u, lapack_int* ldu,
+                    double* v, lapack_int* ldv, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_sgejsv( char* joba, char* jobu, char* jobv, char* jobr, char* jobt,
+                    char* jobp, lapack_int* m, lapack_int* n, float* a,
+                    lapack_int* lda, float* sva, float* u, lapack_int* ldu,
+                    float* v, lapack_int* ldv, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_dgesvj( char* joba, char* jobu, char* jobv, lapack_int* m,
+                    lapack_int* n, double* a, lapack_int* lda, double* sva,
+                    lapack_int* mv, double* v, lapack_int* ldv, double* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sgesvj( char* joba, char* jobu, char* jobv, lapack_int* m,
+                    lapack_int* n, float* a, lapack_int* lda, float* sva,
+                    lapack_int* mv, float* v, lapack_int* ldv, float* work,
+                    lapack_int* lwork, lapack_int *info );
+void LAPACK_sggsvd( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* n, lapack_int* p, lapack_int* k, lapack_int* l,
+                    float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                    float* alpha, float* beta, float* u, lapack_int* ldu,
+                    float* v, lapack_int* ldv, float* q, lapack_int* ldq,
+                    float* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_dggsvd( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* n, lapack_int* p, lapack_int* k, lapack_int* l,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    double* alpha, double* beta, double* u, lapack_int* ldu,
+                    double* v, lapack_int* ldv, double* q, lapack_int* ldq,
+                    double* work, lapack_int* iwork, lapack_int *info );
+void LAPACK_cggsvd( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* n, lapack_int* p, lapack_int* k, lapack_int* l,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* alpha,
+                    float* beta, lapack_complex_float* u, lapack_int* ldu,
+                    lapack_complex_float* v, lapack_int* ldv,
+                    lapack_complex_float* q, lapack_int* ldq,
+                    lapack_complex_float* work, float* rwork, lapack_int* iwork,
+                    lapack_int *info );
+void LAPACK_zggsvd( char* jobu, char* jobv, char* jobq, lapack_int* m,
+                    lapack_int* n, lapack_int* p, lapack_int* k, lapack_int* l,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* alpha,
+                    double* beta, lapack_complex_double* u, lapack_int* ldu,
+                    lapack_complex_double* v, lapack_int* ldv,
+                    lapack_complex_double* q, lapack_int* ldq,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int* iwork, lapack_int *info );
+void LAPACK_ssygv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                   float* w, float* work, lapack_int* lwork, lapack_int *info );
+void LAPACK_dsygv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                   double* w, double* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_chegv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* b, lapack_int* ldb, float* w,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_int *info );
+void LAPACK_zhegv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* b, lapack_int* ldb, double* w,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   double* rwork, lapack_int *info );
+void LAPACK_ssygvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                    float* w, float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_dsygvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    double* w, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_chegvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* w,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zhegvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* w,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_ssygvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, float* vl, float* vu, lapack_int* il,
+                    lapack_int* iu, float* abstol, lapack_int* m, float* w,
+                    float* z, lapack_int* ldz, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_dsygvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_chegvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, float* vl,
+                    float* vu, lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_zhegvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_sspgv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   float* ap, float* bp, float* w, float* z, lapack_int* ldz,
+                   float* work, lapack_int *info );
+void LAPACK_dspgv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   double* ap, double* bp, double* w, double* z,
+                   lapack_int* ldz, double* work, lapack_int *info );
+void LAPACK_chpgv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_float* ap, lapack_complex_float* bp, float* w,
+                   lapack_complex_float* z, lapack_int* ldz,
+                   lapack_complex_float* work, float* rwork, lapack_int *info );
+void LAPACK_zhpgv( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                   lapack_complex_double* ap, lapack_complex_double* bp,
+                   double* w, lapack_complex_double* z, lapack_int* ldz,
+                   lapack_complex_double* work, double* rwork,
+                   lapack_int *info );
+void LAPACK_sspgvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    float* ap, float* bp, float* w, float* z, lapack_int* ldz,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_dspgvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    double* ap, double* bp, double* w, double* z,
+                    lapack_int* ldz, double* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_int* liwork, lapack_int *info );
+void LAPACK_chpgvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_float* ap, lapack_complex_float* bp,
+                    float* w, lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zhpgvd( lapack_int* itype, char* jobz, char* uplo, lapack_int* n,
+                    lapack_complex_double* ap, lapack_complex_double* bp,
+                    double* w, lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_sspgvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, float* ap, float* bp, float* vl, float* vu,
+                    lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, float* z, lapack_int* ldz,
+                    float* work, lapack_int* iwork, lapack_int* ifail,
+                    lapack_int *info );
+void LAPACK_dspgvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, double* ap, double* bp, double* vl,
+                    double* vu, lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, double* z, lapack_int* ldz,
+                    double* work, lapack_int* iwork, lapack_int* ifail,
+                    lapack_int *info );
+void LAPACK_chpgvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, lapack_complex_float* ap,
+                    lapack_complex_float* bp, float* vl, float* vu,
+                    lapack_int* il, lapack_int* iu, float* abstol,
+                    lapack_int* m, float* w, lapack_complex_float* z,
+                    lapack_int* ldz, lapack_complex_float* work, float* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_zhpgvx( lapack_int* itype, char* jobz, char* range, char* uplo,
+                    lapack_int* n, lapack_complex_double* ap,
+                    lapack_complex_double* bp, double* vl, double* vu,
+                    lapack_int* il, lapack_int* iu, double* abstol,
+                    lapack_int* m, double* w, lapack_complex_double* z,
+                    lapack_int* ldz, lapack_complex_double* work, double* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_ssbgv( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                   lapack_int* kb, float* ab, lapack_int* ldab, float* bb,
+                   lapack_int* ldbb, float* w, float* z, lapack_int* ldz,
+                   float* work, lapack_int *info );
+void LAPACK_dsbgv( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                   lapack_int* kb, double* ab, lapack_int* ldab, double* bb,
+                   lapack_int* ldbb, double* w, double* z, lapack_int* ldz,
+                   double* work, lapack_int *info );
+void LAPACK_chbgv( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                   lapack_int* kb, lapack_complex_float* ab, lapack_int* ldab,
+                   lapack_complex_float* bb, lapack_int* ldbb, float* w,
+                   lapack_complex_float* z, lapack_int* ldz,
+                   lapack_complex_float* work, float* rwork, lapack_int *info );
+void LAPACK_zhbgv( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                   lapack_int* kb, lapack_complex_double* ab, lapack_int* ldab,
+                   lapack_complex_double* bb, lapack_int* ldbb, double* w,
+                   lapack_complex_double* z, lapack_int* ldz,
+                   lapack_complex_double* work, double* rwork,
+                   lapack_int *info );
+void LAPACK_ssbgvd( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, float* ab, lapack_int* ldab, float* bb,
+                    lapack_int* ldbb, float* w, float* z, lapack_int* ldz,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_dsbgvd( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, double* ab, lapack_int* ldab, double* bb,
+                    lapack_int* ldbb, double* w, double* z, lapack_int* ldz,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_chbgvd( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, lapack_complex_float* ab, lapack_int* ldab,
+                    lapack_complex_float* bb, lapack_int* ldbb, float* w,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* lrwork, lapack_int* iwork, lapack_int* liwork,
+                    lapack_int *info );
+void LAPACK_zhbgvd( char* jobz, char* uplo, lapack_int* n, lapack_int* ka,
+                    lapack_int* kb, lapack_complex_double* ab, lapack_int* ldab,
+                    lapack_complex_double* bb, lapack_int* ldbb, double* w,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_int *info );
+void LAPACK_ssbgvx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* ka, lapack_int* kb, float* ab, lapack_int* ldab,
+                    float* bb, lapack_int* ldbb, float* q, lapack_int* ldq,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w, float* z,
+                    lapack_int* ldz, float* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_dsbgvx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* ka, lapack_int* kb, double* ab,
+                    lapack_int* ldab, double* bb, lapack_int* ldbb, double* q,
+                    lapack_int* ldq, double* vl, double* vu, lapack_int* il,
+                    lapack_int* iu, double* abstol, lapack_int* m, double* w,
+                    double* z, lapack_int* ldz, double* work, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_chbgvx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* ka, lapack_int* kb, lapack_complex_float* ab,
+                    lapack_int* ldab, lapack_complex_float* bb,
+                    lapack_int* ldbb, lapack_complex_float* q, lapack_int* ldq,
+                    float* vl, float* vu, lapack_int* il, lapack_int* iu,
+                    float* abstol, lapack_int* m, float* w,
+                    lapack_complex_float* z, lapack_int* ldz,
+                    lapack_complex_float* work, float* rwork, lapack_int* iwork,
+                    lapack_int* ifail, lapack_int *info );
+void LAPACK_zhbgvx( char* jobz, char* range, char* uplo, lapack_int* n,
+                    lapack_int* ka, lapack_int* kb, lapack_complex_double* ab,
+                    lapack_int* ldab, lapack_complex_double* bb,
+                    lapack_int* ldbb, lapack_complex_double* q, lapack_int* ldq,
+                    double* vl, double* vu, lapack_int* il, lapack_int* iu,
+                    double* abstol, lapack_int* m, double* w,
+                    lapack_complex_double* z, lapack_int* ldz,
+                    lapack_complex_double* work, double* rwork,
+                    lapack_int* iwork, lapack_int* ifail, lapack_int *info );
+void LAPACK_sgges( char* jobvsl, char* jobvsr, char* sort,
+                   LAPACK_S_SELECT3 selctg, lapack_int* n, float* a,
+                   lapack_int* lda, float* b, lapack_int* ldb, lapack_int* sdim,
+                   float* alphar, float* alphai, float* beta, float* vsl,
+                   lapack_int* ldvsl, float* vsr, lapack_int* ldvsr,
+                   float* work, lapack_int* lwork, lapack_logical* bwork,
+                   lapack_int *info );
+void LAPACK_dgges( char* jobvsl, char* jobvsr, char* sort,
+                   LAPACK_D_SELECT3 selctg, lapack_int* n, double* a,
+                   lapack_int* lda, double* b, lapack_int* ldb,
+                   lapack_int* sdim, double* alphar, double* alphai,
+                   double* beta, double* vsl, lapack_int* ldvsl, double* vsr,
+                   lapack_int* ldvsr, double* work, lapack_int* lwork,
+                   lapack_logical* bwork, lapack_int *info );
+void LAPACK_cgges( char* jobvsl, char* jobvsr, char* sort,
+                   LAPACK_C_SELECT2 selctg, lapack_int* n,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* b, lapack_int* ldb, lapack_int* sdim,
+                   lapack_complex_float* alpha, lapack_complex_float* beta,
+                   lapack_complex_float* vsl, lapack_int* ldvsl,
+                   lapack_complex_float* vsr, lapack_int* ldvsr,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_logical* bwork, lapack_int *info );
+void LAPACK_zgges( char* jobvsl, char* jobvsr, char* sort,
+                   LAPACK_Z_SELECT2 selctg, lapack_int* n,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* b, lapack_int* ldb, lapack_int* sdim,
+                   lapack_complex_double* alpha, lapack_complex_double* beta,
+                   lapack_complex_double* vsl, lapack_int* ldvsl,
+                   lapack_complex_double* vsr, lapack_int* ldvsr,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   double* rwork, lapack_logical* bwork, lapack_int *info );
+void LAPACK_sggesx( char* jobvsl, char* jobvsr, char* sort,
+                    LAPACK_S_SELECT3 selctg, char* sense, lapack_int* n,
+                    float* a, lapack_int* lda, float* b, lapack_int* ldb,
+                    lapack_int* sdim, float* alphar, float* alphai, float* beta,
+                    float* vsl, lapack_int* ldvsl, float* vsr,
+                    lapack_int* ldvsr, float* rconde, float* rcondv,
+                    float* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_dggesx( char* jobvsl, char* jobvsr, char* sort,
+                    LAPACK_D_SELECT3 selctg, char* sense, lapack_int* n,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    lapack_int* sdim, double* alphar, double* alphai,
+                    double* beta, double* vsl, lapack_int* ldvsl, double* vsr,
+                    lapack_int* ldvsr, double* rconde, double* rcondv,
+                    double* work, lapack_int* lwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_cggesx( char* jobvsl, char* jobvsr, char* sort,
+                    LAPACK_C_SELECT2 selctg, char* sense, lapack_int* n,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb, lapack_int* sdim,
+                    lapack_complex_float* alpha, lapack_complex_float* beta,
+                    lapack_complex_float* vsl, lapack_int* ldvsl,
+                    lapack_complex_float* vsr, lapack_int* ldvsr, float* rconde,
+                    float* rcondv, lapack_complex_float* work,
+                    lapack_int* lwork, float* rwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_zggesx( char* jobvsl, char* jobvsr, char* sort,
+                    LAPACK_Z_SELECT2 selctg, char* sense, lapack_int* n,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb, lapack_int* sdim,
+                    lapack_complex_double* alpha, lapack_complex_double* beta,
+                    lapack_complex_double* vsl, lapack_int* ldvsl,
+                    lapack_complex_double* vsr, lapack_int* ldvsr,
+                    double* rconde, double* rcondv, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int* iwork,
+                    lapack_int* liwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_sggev( char* jobvl, char* jobvr, lapack_int* n, float* a,
+                   lapack_int* lda, float* b, lapack_int* ldb, float* alphar,
+                   float* alphai, float* beta, float* vl, lapack_int* ldvl,
+                   float* vr, lapack_int* ldvr, float* work, lapack_int* lwork,
+                   lapack_int *info );
+void LAPACK_dggev( char* jobvl, char* jobvr, lapack_int* n, double* a,
+                   lapack_int* lda, double* b, lapack_int* ldb, double* alphar,
+                   double* alphai, double* beta, double* vl, lapack_int* ldvl,
+                   double* vr, lapack_int* ldvr, double* work,
+                   lapack_int* lwork, lapack_int *info );
+void LAPACK_cggev( char* jobvl, char* jobvr, lapack_int* n,
+                   lapack_complex_float* a, lapack_int* lda,
+                   lapack_complex_float* b, lapack_int* ldb,
+                   lapack_complex_float* alpha, lapack_complex_float* beta,
+                   lapack_complex_float* vl, lapack_int* ldvl,
+                   lapack_complex_float* vr, lapack_int* ldvr,
+                   lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                   lapack_int *info );
+void LAPACK_zggev( char* jobvl, char* jobvr, lapack_int* n,
+                   lapack_complex_double* a, lapack_int* lda,
+                   lapack_complex_double* b, lapack_int* ldb,
+                   lapack_complex_double* alpha, lapack_complex_double* beta,
+                   lapack_complex_double* vl, lapack_int* ldvl,
+                   lapack_complex_double* vr, lapack_int* ldvr,
+                   lapack_complex_double* work, lapack_int* lwork,
+                   double* rwork, lapack_int *info );
+void LAPACK_sggevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, float* alphar, float* alphai, float* beta,
+                    float* vl, lapack_int* ldvl, float* vr, lapack_int* ldvr,
+                    lapack_int* ilo, lapack_int* ihi, float* lscale,
+                    float* rscale, float* abnrm, float* bbnrm, float* rconde,
+                    float* rcondv, float* work, lapack_int* lwork,
+                    lapack_int* iwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_dggevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, double* alphar, double* alphai,
+                    double* beta, double* vl, lapack_int* ldvl, double* vr,
+                    lapack_int* ldvr, lapack_int* ilo, lapack_int* ihi,
+                    double* lscale, double* rscale, double* abnrm,
+                    double* bbnrm, double* rconde, double* rcondv, double* work,
+                    lapack_int* lwork, lapack_int* iwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_cggevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    lapack_complex_float* alpha, lapack_complex_float* beta,
+                    lapack_complex_float* vl, lapack_int* ldvl,
+                    lapack_complex_float* vr, lapack_int* ldvr, lapack_int* ilo,
+                    lapack_int* ihi, float* lscale, float* rscale, float* abnrm,
+                    float* bbnrm, float* rconde, float* rcondv,
+                    lapack_complex_float* work, lapack_int* lwork, float* rwork,
+                    lapack_int* iwork, lapack_logical* bwork,
+                    lapack_int *info );
+void LAPACK_zggevx( char* balanc, char* jobvl, char* jobvr, char* sense,
+                    lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* alpha, lapack_complex_double* beta,
+                    lapack_complex_double* vl, lapack_int* ldvl,
+                    lapack_complex_double* vr, lapack_int* ldvr,
+                    lapack_int* ilo, lapack_int* ihi, double* lscale,
+                    double* rscale, double* abnrm, double* bbnrm,
+                    double* rconde, double* rcondv, lapack_complex_double* work,
+                    lapack_int* lwork, double* rwork, lapack_int* iwork,
+                    lapack_logical* bwork, lapack_int *info );
+void LAPACK_dsfrk( char* transr, char* uplo, char* trans, lapack_int* n,
+                   lapack_int* k, double* alpha, const double* a,
+                   lapack_int* lda, double* beta, double* c );
+void LAPACK_ssfrk( char* transr, char* uplo, char* trans, lapack_int* n,
+                   lapack_int* k, float* alpha, const float* a, lapack_int* lda,
+                   float* beta, float* c );
+void LAPACK_zhfrk( char* transr, char* uplo, char* trans, lapack_int* n,
+                   lapack_int* k, double* alpha, const lapack_complex_double* a,
+                   lapack_int* lda, double* beta, lapack_complex_double* c );
+void LAPACK_chfrk( char* transr, char* uplo, char* trans, lapack_int* n,
+                   lapack_int* k, float* alpha, const lapack_complex_float* a,
+                   lapack_int* lda, float* beta, lapack_complex_float* c );
+void LAPACK_dtfsm( char* transr, char* side, char* uplo, char* trans,
+                   char* diag, lapack_int* m, lapack_int* n, double* alpha,
+                   const double* a, double* b, lapack_int* ldb );
+void LAPACK_stfsm( char* transr, char* side, char* uplo, char* trans,
+                   char* diag, lapack_int* m, lapack_int* n, float* alpha,
+                   const float* a, float* b, lapack_int* ldb );
+void LAPACK_ztfsm( char* transr, char* side, char* uplo, char* trans,
+                   char* diag, lapack_int* m, lapack_int* n,
+                   lapack_complex_double* alpha, const lapack_complex_double* a,
+                   lapack_complex_double* b, lapack_int* ldb );
+void LAPACK_ctfsm( char* transr, char* side, char* uplo, char* trans,
+                   char* diag, lapack_int* m, lapack_int* n,
+                   lapack_complex_float* alpha, const lapack_complex_float* a,
+                   lapack_complex_float* b, lapack_int* ldb );
+void LAPACK_dtfttp( char* transr, char* uplo, lapack_int* n, const double* arf,
+                    double* ap, lapack_int *info );
+void LAPACK_stfttp( char* transr, char* uplo, lapack_int* n, const float* arf,
+                    float* ap, lapack_int *info );
+void LAPACK_ztfttp( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_double* arf, lapack_complex_double* ap,
+                    lapack_int *info );
+void LAPACK_ctfttp( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_float* arf, lapack_complex_float* ap,
+                    lapack_int *info );
+void LAPACK_dtfttr( char* transr, char* uplo, lapack_int* n, const double* arf,
+                    double* a, lapack_int* lda, lapack_int *info );
+void LAPACK_stfttr( char* transr, char* uplo, lapack_int* n, const float* arf,
+                    float* a, lapack_int* lda, lapack_int *info );
+void LAPACK_ztfttr( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_double* arf, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_ctfttr( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_float* arf, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_dtpttf( char* transr, char* uplo, lapack_int* n, const double* ap,
+                    double* arf, lapack_int *info );
+void LAPACK_stpttf( char* transr, char* uplo, lapack_int* n, const float* ap,
+                    float* arf, lapack_int *info );
+void LAPACK_ztpttf( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_double* ap, lapack_complex_double* arf,
+                    lapack_int *info );
+void LAPACK_ctpttf( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_float* ap, lapack_complex_float* arf,
+                    lapack_int *info );
+void LAPACK_dtpttr( char* uplo, lapack_int* n, const double* ap, double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_stpttr( char* uplo, lapack_int* n, const float* ap, float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_ztpttr( char* uplo, lapack_int* n, const lapack_complex_double* ap,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_ctpttr( char* uplo, lapack_int* n, const lapack_complex_float* ap,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dtrttf( char* transr, char* uplo, lapack_int* n, const double* a,
+                    lapack_int* lda, double* arf, lapack_int *info );
+void LAPACK_strttf( char* transr, char* uplo, lapack_int* n, const float* a,
+                    lapack_int* lda, float* arf, lapack_int *info );
+void LAPACK_ztrttf( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* arf, lapack_int *info );
+void LAPACK_ctrttf( char* transr, char* uplo, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* arf, lapack_int *info );
+void LAPACK_dtrttp( char* uplo, lapack_int* n, const double* a, lapack_int* lda,
+                    double* ap, lapack_int *info );
+void LAPACK_strttp( char* uplo, lapack_int* n, const float* a, lapack_int* lda,
+                    float* ap, lapack_int *info );
+void LAPACK_ztrttp( char* uplo, lapack_int* n, const lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* ap,
+                    lapack_int *info );
+void LAPACK_ctrttp( char* uplo, lapack_int* n, const lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* ap,
+                    lapack_int *info );
+void LAPACK_sgeqrfp( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                     float* tau, float* work, lapack_int* lwork,
+                     lapack_int *info );
+void LAPACK_dgeqrfp( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                     double* tau, double* work, lapack_int* lwork,
+                     lapack_int *info );
+void LAPACK_cgeqrfp( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                     lapack_int* lda, lapack_complex_float* tau,
+                     lapack_complex_float* work, lapack_int* lwork,
+                     lapack_int *info );
+void LAPACK_zgeqrfp( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                     lapack_int* lda, lapack_complex_double* tau,
+                     lapack_complex_double* work, lapack_int* lwork,
+                     lapack_int *info );
+void LAPACK_clacgv( lapack_int* n, lapack_complex_float* x, lapack_int* incx );
+void LAPACK_zlacgv( lapack_int* n, lapack_complex_double* x, lapack_int* incx );
+void LAPACK_slarnv( lapack_int* idist, lapack_int* iseed, lapack_int* n,
+                    float* x );
+void LAPACK_dlarnv( lapack_int* idist, lapack_int* iseed, lapack_int* n,
+                    double* x );
+void LAPACK_clarnv( lapack_int* idist, lapack_int* iseed, lapack_int* n,
+                    lapack_complex_float* x );
+void LAPACK_zlarnv( lapack_int* idist, lapack_int* iseed, lapack_int* n,
+                    lapack_complex_double* x );
+void LAPACK_sgeqr2( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int *info );
+void LAPACK_dgeqr2( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int *info );
+void LAPACK_cgeqr2( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zgeqr2( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_slacpy( char* uplo, lapack_int* m, lapack_int* n, const float* a,
+                    lapack_int* lda, float* b, lapack_int* ldb );
+void LAPACK_dlacpy( char* uplo, lapack_int* m, lapack_int* n, const double* a,
+                    lapack_int* lda, double* b, lapack_int* ldb );
+void LAPACK_clacpy( char* uplo, lapack_int* m, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb );
+void LAPACK_zlacpy( char* uplo, lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb );
+void LAPACK_sgetf2( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_dgetf2( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int* ipiv, lapack_int *info );
+void LAPACK_cgetf2( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* ipiv, lapack_int *info );
+void LAPACK_zgetf2( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* ipiv, lapack_int *info );
+void LAPACK_slaswp( lapack_int* n, float* a, lapack_int* lda, lapack_int* k1,
+                    lapack_int* k2, const lapack_int* ipiv, lapack_int* incx );
+void LAPACK_dlaswp( lapack_int* n, double* a, lapack_int* lda, lapack_int* k1,
+                    lapack_int* k2, const lapack_int* ipiv, lapack_int* incx );
+void LAPACK_claswp( lapack_int* n, lapack_complex_float* a, lapack_int* lda,
+                    lapack_int* k1, lapack_int* k2, const lapack_int* ipiv,
+                    lapack_int* incx );
+void LAPACK_zlaswp( lapack_int* n, lapack_complex_double* a, lapack_int* lda,
+                    lapack_int* k1, lapack_int* k2, const lapack_int* ipiv,
+                    lapack_int* incx );
+float LAPACK_slange( char* norm, lapack_int* m, lapack_int* n, const float* a,
+                    lapack_int* lda, float* work );
+double LAPACK_dlange( char* norm, lapack_int* m, lapack_int* n, const double* a,
+                    lapack_int* lda, double* work );
+float LAPACK_clange( char* norm, lapack_int* m, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda, float* work );
+double LAPACK_zlange( char* norm, lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda, double* work );
+float LAPACK_clanhe( char* norm, char* uplo, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda, float* work );
+double LAPACK_zlanhe( char* norm, char* uplo, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda, double* work );
+float LAPACK_slansy( char* norm, char* uplo, lapack_int* n, const float* a,
+                    lapack_int* lda, float* work );
+double LAPACK_dlansy( char* norm, char* uplo, lapack_int* n, const double* a,
+                    lapack_int* lda, double* work );
+float LAPACK_clansy( char* norm, char* uplo, lapack_int* n,
+                    const lapack_complex_float* a, lapack_int* lda, float* work );
+double LAPACK_zlansy( char* norm, char* uplo, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda, double* work );
+float LAPACK_slantr( char* norm, char* uplo, char* diag, lapack_int* m,
+                    lapack_int* n, const float* a, lapack_int* lda, float* work );
+double LAPACK_dlantr( char* norm, char* uplo, char* diag, lapack_int* m,
+                    lapack_int* n, const double* a, lapack_int* lda, double* work );
+float LAPACK_clantr( char* norm, char* uplo, char* diag, lapack_int* m,
+                    lapack_int* n, const lapack_complex_float* a, lapack_int* lda,
+                    float* work );
+double LAPACK_zlantr( char* norm, char* uplo, char* diag, lapack_int* m,
+                    lapack_int* n, const lapack_complex_double* a, lapack_int* lda,
+                    double* work );
+float LAPACK_slamch( char* cmach );
+double LAPACK_dlamch( char* cmach );
+void LAPACK_sgelq2( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                    float* tau, float* work, lapack_int *info );
+void LAPACK_dgelq2( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                    double* tau, double* work, lapack_int *info );
+void LAPACK_cgelq2( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_complex_float* tau,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zgelq2( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_complex_double* tau,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_slarfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k, const float* v,
+                    lapack_int* ldv, const float* t, lapack_int* ldt, float* c,
+                    lapack_int* ldc, float* work, lapack_int* ldwork );
+void LAPACK_dlarfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k,
+                    const double* v, lapack_int* ldv, const double* t,
+                    lapack_int* ldt, double* c, lapack_int* ldc, double* work,
+                    lapack_int* ldwork );
+void LAPACK_clarfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k,
+                    const lapack_complex_float* v, lapack_int* ldv,
+                    const lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work, lapack_int* ldwork );
+void LAPACK_zlarfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k,
+                    const lapack_complex_double* v, lapack_int* ldv,
+                    const lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work, lapack_int* ldwork );
+void LAPACK_slarfg( lapack_int* n, float* alpha, float* x, lapack_int* incx,
+                    float* tau );
+void LAPACK_dlarfg( lapack_int* n, double* alpha, double* x, lapack_int* incx,
+                    double* tau );
+void LAPACK_clarfg( lapack_int* n, lapack_complex_float* alpha,
+                    lapack_complex_float* x, lapack_int* incx,
+                    lapack_complex_float* tau );
+void LAPACK_zlarfg( lapack_int* n, lapack_complex_double* alpha,
+                    lapack_complex_double* x, lapack_int* incx,
+                    lapack_complex_double* tau );
+void LAPACK_slarft( char* direct, char* storev, lapack_int* n, lapack_int* k,
+                    const float* v, lapack_int* ldv, const float* tau, float* t,
+                    lapack_int* ldt );
+void LAPACK_dlarft( char* direct, char* storev, lapack_int* n, lapack_int* k,
+                    const double* v, lapack_int* ldv, const double* tau,
+                    double* t, lapack_int* ldt );
+void LAPACK_clarft( char* direct, char* storev, lapack_int* n, lapack_int* k,
+                    const lapack_complex_float* v, lapack_int* ldv,
+                    const lapack_complex_float* tau, lapack_complex_float* t,
+                    lapack_int* ldt );
+void LAPACK_zlarft( char* direct, char* storev, lapack_int* n, lapack_int* k,
+                    const lapack_complex_double* v, lapack_int* ldv,
+                    const lapack_complex_double* tau, lapack_complex_double* t,
+                    lapack_int* ldt );
+void LAPACK_slarfx( char* side, lapack_int* m, lapack_int* n, const float* v,
+                    float* tau, float* c, lapack_int* ldc, float* work );
+void LAPACK_dlarfx( char* side, lapack_int* m, lapack_int* n, const double* v,
+                    double* tau, double* c, lapack_int* ldc, double* work );
+void LAPACK_clarfx( char* side, lapack_int* m, lapack_int* n,
+                    const lapack_complex_float* v, lapack_complex_float* tau,
+                    lapack_complex_float* c, lapack_int* ldc,
+                    lapack_complex_float* work );
+void LAPACK_zlarfx( char* side, lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* v, lapack_complex_double* tau,
+                    lapack_complex_double* c, lapack_int* ldc,
+                    lapack_complex_double* work );
+void LAPACK_slatms( lapack_int* m, lapack_int* n, char* dist, lapack_int* iseed,
+                    char* sym, float* d, lapack_int* mode, float* cond,
+                    float* dmax, lapack_int* kl, lapack_int* ku, char* pack,
+                    float* a, lapack_int* lda, float* work, lapack_int *info );
+void LAPACK_dlatms( lapack_int* m, lapack_int* n, char* dist, lapack_int* iseed,
+                    char* sym, double* d, lapack_int* mode, double* cond,
+                    double* dmax, lapack_int* kl, lapack_int* ku, char* pack,
+                    double* a, lapack_int* lda, double* work,
+                    lapack_int *info );
+void LAPACK_clatms( lapack_int* m, lapack_int* n, char* dist, lapack_int* iseed,
+                    char* sym, float* d, lapack_int* mode, float* cond,
+                    float* dmax, lapack_int* kl, lapack_int* ku, char* pack,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zlatms( lapack_int* m, lapack_int* n, char* dist, lapack_int* iseed,
+                    char* sym, double* d, lapack_int* mode, double* cond,
+                    double* dmax, lapack_int* kl, lapack_int* ku, char* pack,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_slag2d( lapack_int* m, lapack_int* n, const float* sa,
+                    lapack_int* ldsa, double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dlag2s( lapack_int* m, lapack_int* n, const double* a,
+                    lapack_int* lda, float* sa, lapack_int* ldsa,
+                    lapack_int *info );
+void LAPACK_clag2z( lapack_int* m, lapack_int* n,
+                    const lapack_complex_float* sa, lapack_int* ldsa,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_zlag2c( lapack_int* m, lapack_int* n,
+                    const lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_float* sa, lapack_int* ldsa,
+                    lapack_int *info );
+void LAPACK_slauum( char* uplo, lapack_int* n, float* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_dlauum( char* uplo, lapack_int* n, double* a, lapack_int* lda,
+                    lapack_int *info );
+void LAPACK_clauum( char* uplo, lapack_int* n, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_zlauum( char* uplo, lapack_int* n, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int *info );
+void LAPACK_slagge( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const float* d, float* a, lapack_int* lda,
+                    lapack_int* iseed, float* work, lapack_int *info );
+void LAPACK_dlagge( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const double* d, double* a, lapack_int* lda,
+                    lapack_int* iseed, double* work, lapack_int *info );
+void LAPACK_clagge( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const float* d, lapack_complex_float* a,
+                    lapack_int* lda, lapack_int* iseed,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zlagge( lapack_int* m, lapack_int* n, lapack_int* kl,
+                    lapack_int* ku, const double* d, lapack_complex_double* a,
+                    lapack_int* lda, lapack_int* iseed,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_slaset( char* uplo, lapack_int* m, lapack_int* n, float* alpha,
+                    float* beta, float* a, lapack_int* lda );
+void LAPACK_dlaset( char* uplo, lapack_int* m, lapack_int* n, double* alpha,
+                    double* beta, double* a, lapack_int* lda );
+void LAPACK_claset( char* uplo, lapack_int* m, lapack_int* n,
+                    lapack_complex_float* alpha, lapack_complex_float* beta,
+                    lapack_complex_float* a, lapack_int* lda );
+void LAPACK_zlaset( char* uplo, lapack_int* m, lapack_int* n,
+                    lapack_complex_double* alpha, lapack_complex_double* beta,
+                    lapack_complex_double* a, lapack_int* lda );
+void LAPACK_slasrt( char* id, lapack_int* n, float* d, lapack_int *info );
+void LAPACK_dlasrt( char* id, lapack_int* n, double* d, lapack_int *info );
+void LAPACK_claghe( lapack_int* n, lapack_int* k, const float* d,
+                    lapack_complex_float* a, lapack_int* lda, lapack_int* iseed,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zlaghe( lapack_int* n, lapack_int* k, const double* d,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_int* iseed, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_slagsy( lapack_int* n, lapack_int* k, const float* d, float* a,
+                    lapack_int* lda, lapack_int* iseed, float* work,
+                    lapack_int *info );
+void LAPACK_dlagsy( lapack_int* n, lapack_int* k, const double* d, double* a,
+                    lapack_int* lda, lapack_int* iseed, double* work,
+                    lapack_int *info );
+void LAPACK_clagsy( lapack_int* n, lapack_int* k, const float* d,
+                    lapack_complex_float* a, lapack_int* lda, lapack_int* iseed,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zlagsy( lapack_int* n, lapack_int* k, const double* d,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_int* iseed, lapack_complex_double* work,
+                    lapack_int *info );
+void LAPACK_slapmr( lapack_logical* forwrd, lapack_int* m, lapack_int* n,
+                    float* x, lapack_int* ldx, lapack_int* k );
+void LAPACK_dlapmr( lapack_logical* forwrd, lapack_int* m, lapack_int* n,
+                    double* x, lapack_int* ldx, lapack_int* k );
+void LAPACK_clapmr( lapack_logical* forwrd, lapack_int* m, lapack_int* n,
+                    lapack_complex_float* x, lapack_int* ldx, lapack_int* k );
+void LAPACK_zlapmr( lapack_logical* forwrd, lapack_int* m, lapack_int* n,
+                    lapack_complex_double* x, lapack_int* ldx, lapack_int* k );
+float LAPACK_slapy2( float* x, float* y );
+double LAPACK_dlapy2( double* x, double* y );
+float LAPACK_slapy3( float* x, float* y, float* z );
+double LAPACK_dlapy3( double* x, double* y, double* z );
+void LAPACK_slartgp( float* f, float* g, float* cs, float* sn, float* r );
+void LAPACK_dlartgp( double* f, double* g, double* cs, double* sn, double* r );
+void LAPACK_slartgs( float* x, float* y, float* sigma, float* cs, float* sn );
+void LAPACK_dlartgs( double* x, double* y, double* sigma, double* cs,
+                     double* sn );
+// LAPACK 3.3.0
+void LAPACK_cbbcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    float* theta, float* phi,
+                    lapack_complex_float* u1, lapack_int* ldu1,
+                    lapack_complex_float* u2, lapack_int* ldu2,
+                    lapack_complex_float* v1t, lapack_int* ldv1t,
+                    lapack_complex_float* v2t, lapack_int* ldv2t,
+                    float* b11d, float* b11e, float* b12d,
+                    float* b12e, float* b21d, float* b21e,
+                    float* b22d, float* b22e, float* rwork,
+                    lapack_int* lrwork , lapack_int *info );
+void LAPACK_cheswapr( char* uplo, lapack_int* n,
+                      lapack_complex_float* a, lapack_int* i1,
+                      lapack_int* i2 );
+void LAPACK_chetri2( char* uplo, lapack_int* n,
+                     lapack_complex_float* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_float* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_chetri2x( char* uplo, lapack_int* n,
+                      lapack_complex_float* a, lapack_int* lda,
+                      const lapack_int* ipiv,
+                      lapack_complex_float* work, lapack_int* nb , lapack_int *info );
+void LAPACK_chetrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs, const lapack_complex_float* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* work , lapack_int *info );
+void LAPACK_csyconv( char* uplo, char* way,
+                     lapack_int* n, lapack_complex_float* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     lapack_complex_float* work , lapack_int *info );
+void LAPACK_csyswapr( char* uplo, lapack_int* n,
+                      lapack_complex_float* a, lapack_int* i1,
+                      lapack_int* i2 );
+void LAPACK_csytri2( char* uplo, lapack_int* n,
+                     lapack_complex_float* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_float* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_csytri2x( char* uplo, lapack_int* n,
+                      lapack_complex_float* a, lapack_int* lda,
+                      const lapack_int* ipiv,
+                      lapack_complex_float* work, lapack_int* nb , lapack_int *info );
+void LAPACK_csytrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs, const lapack_complex_float* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* work , lapack_int *info );
+void LAPACK_cunbdb( char* trans, char* signs,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    lapack_complex_float* x11, lapack_int* ldx11,
+                    lapack_complex_float* x12, lapack_int* ldx12,
+                    lapack_complex_float* x21, lapack_int* ldx21,
+                    lapack_complex_float* x22, lapack_int* ldx22,
+                    float* theta, float* phi,
+                    lapack_complex_float* taup1,
+                    lapack_complex_float* taup2,
+                    lapack_complex_float* tauq1,
+                    lapack_complex_float* tauq2,
+                    lapack_complex_float* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_cuncsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    char* signs, lapack_int* m, lapack_int* p,
+                    lapack_int* q, lapack_complex_float* x11,
+                    lapack_int* ldx11, lapack_complex_float* x12,
+                    lapack_int* ldx12, lapack_complex_float* x21,
+                    lapack_int* ldx21, lapack_complex_float* x22,
+                    lapack_int* ldx22, float* theta,
+                    lapack_complex_float* u1, lapack_int* ldu1,
+                    lapack_complex_float* u2, lapack_int* ldu2,
+                    lapack_complex_float* v1t, lapack_int* ldv1t,
+                    lapack_complex_float* v2t, lapack_int* ldv2t,
+                    lapack_complex_float* work, lapack_int* lwork,
+                    float* rwork, lapack_int* lrwork,
+                    lapack_int* iwork , lapack_int *info );
+void LAPACK_dbbcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    double* theta, double* phi, double* u1,
+                    lapack_int* ldu1, double* u2, lapack_int* ldu2,
+                    double* v1t, lapack_int* ldv1t, double* v2t,
+                    lapack_int* ldv2t, double* b11d, double* b11e,
+                    double* b12d, double* b12e, double* b21d,
+                    double* b21e, double* b22d, double* b22e,
+                    double* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_dorbdb( char* trans, char* signs,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    double* x11, lapack_int* ldx11, double* x12,
+                    lapack_int* ldx12, double* x21, lapack_int* ldx21,
+                    double* x22, lapack_int* ldx22, double* theta,
+                    double* phi, double* taup1, double* taup2,
+                    double* tauq1, double* tauq2, double* work,
+                    lapack_int* lwork , lapack_int *info );
+void LAPACK_dorcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    char* signs, lapack_int* m, lapack_int* p,
+                    lapack_int* q, double* x11, lapack_int* ldx11,
+                    double* x12, lapack_int* ldx12, double* x21,
+                    lapack_int* ldx21, double* x22, lapack_int* ldx22,
+                    double* theta, double* u1, lapack_int* ldu1,
+                    double* u2, lapack_int* ldu2, double* v1t,
+                    lapack_int* ldv1t, double* v2t, lapack_int* ldv2t,
+                    double* work, lapack_int* lwork,
+                    lapack_int* iwork , lapack_int *info );
+void LAPACK_dsyconv( char* uplo, char* way,
+                     lapack_int* n, double* a, lapack_int* lda,
+                     const lapack_int* ipiv, double* work , lapack_int *info );
+void LAPACK_dsyswapr( char* uplo, lapack_int* n,
+                      double* a, lapack_int* i1, lapack_int* i2 );
+void LAPACK_dsytri2( char* uplo, lapack_int* n,
+                     double* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_double* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_dsytri2x( char* uplo, lapack_int* n,
+                      double* a, lapack_int* lda,
+                      const lapack_int* ipiv, double* work,
+                      lapack_int* nb , lapack_int *info );
+void LAPACK_dsytrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs, const double* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     double* b, lapack_int* ldb, double* work , lapack_int *info );
+void LAPACK_sbbcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    float* theta, float* phi, float* u1,
+                    lapack_int* ldu1, float* u2, lapack_int* ldu2,
+                    float* v1t, lapack_int* ldv1t, float* v2t,
+                    lapack_int* ldv2t, float* b11d, float* b11e,
+                    float* b12d, float* b12e, float* b21d,
+                    float* b21e, float* b22d, float* b22e,
+                    float* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_sorbdb( char* trans, char* signs,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    float* x11, lapack_int* ldx11, float* x12,
+                    lapack_int* ldx12, float* x21, lapack_int* ldx21,
+                    float* x22, lapack_int* ldx22, float* theta,
+                    float* phi, float* taup1, float* taup2,
+                    float* tauq1, float* tauq2, float* work,
+                    lapack_int* lwork , lapack_int *info );
+void LAPACK_sorcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    char* signs, lapack_int* m, lapack_int* p,
+                    lapack_int* q, float* x11, lapack_int* ldx11,
+                    float* x12, lapack_int* ldx12, float* x21,
+                    lapack_int* ldx21, float* x22, lapack_int* ldx22,
+                    float* theta, float* u1, lapack_int* ldu1,
+                    float* u2, lapack_int* ldu2, float* v1t,
+                    lapack_int* ldv1t, float* v2t, lapack_int* ldv2t,
+                    float* work, lapack_int* lwork,
+                    lapack_int* iwork , lapack_int *info );
+void LAPACK_ssyconv( char* uplo, char* way,
+                     lapack_int* n, float* a, lapack_int* lda,
+                     const lapack_int* ipiv, float* work , lapack_int *info );
+void LAPACK_ssyswapr( char* uplo, lapack_int* n,
+                      float* a, lapack_int* i1, lapack_int* i2 );
+void LAPACK_ssytri2( char* uplo, lapack_int* n,
+                     float* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_float* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_ssytri2x( char* uplo, lapack_int* n,
+                      float* a, lapack_int* lda,
+                      const lapack_int* ipiv, float* work,
+                      lapack_int* nb , lapack_int *info );
+void LAPACK_ssytrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs, const float* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     float* b, lapack_int* ldb, float* work , lapack_int *info );
+void LAPACK_zbbcsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    double* theta, double* phi,
+                    lapack_complex_double* u1, lapack_int* ldu1,
+                    lapack_complex_double* u2, lapack_int* ldu2,
+                    lapack_complex_double* v1t, lapack_int* ldv1t,
+                    lapack_complex_double* v2t, lapack_int* ldv2t,
+                    double* b11d, double* b11e, double* b12d,
+                    double* b12e, double* b21d, double* b21e,
+                    double* b22d, double* b22e, double* rwork,
+                    lapack_int* lrwork , lapack_int *info );
+void LAPACK_zheswapr( char* uplo, lapack_int* n,
+                      lapack_complex_double* a, lapack_int* i1,
+                      lapack_int* i2 );
+void LAPACK_zhetri2( char* uplo, lapack_int* n,
+                     lapack_complex_double* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_double* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_zhetri2x( char* uplo, lapack_int* n,
+                      lapack_complex_double* a, lapack_int* lda,
+                      const lapack_int* ipiv,
+                      lapack_complex_double* work, lapack_int* nb , lapack_int *info );
+void LAPACK_zhetrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* work , lapack_int *info );
+void LAPACK_zsyconv( char* uplo, char* way,
+                     lapack_int* n, lapack_complex_double* a,
+                     lapack_int* lda, const lapack_int* ipiv,
+                     lapack_complex_double* work , lapack_int *info );
+void LAPACK_zsyswapr( char* uplo, lapack_int* n,
+                      lapack_complex_double* a, lapack_int* i1,
+                      lapack_int* i2 );
+void LAPACK_zsytri2( char* uplo, lapack_int* n,
+                     lapack_complex_double* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_double* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_zsytri2x( char* uplo, lapack_int* n,
+                      lapack_complex_double* a, lapack_int* lda,
+                      const lapack_int* ipiv,
+                      lapack_complex_double* work, lapack_int* nb , lapack_int *info );
+void LAPACK_zsytrs2( char* uplo, lapack_int* n,
+                     lapack_int* nrhs,
+                     const lapack_complex_double* a, lapack_int* lda,
+                     const lapack_int* ipiv,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* work , lapack_int *info );
+void LAPACK_zunbdb( char* trans, char* signs,
+                    lapack_int* m, lapack_int* p, lapack_int* q,
+                    lapack_complex_double* x11, lapack_int* ldx11,
+                    lapack_complex_double* x12, lapack_int* ldx12,
+                    lapack_complex_double* x21, lapack_int* ldx21,
+                    lapack_complex_double* x22, lapack_int* ldx22,
+                    double* theta, double* phi,
+                    lapack_complex_double* taup1,
+                    lapack_complex_double* taup2,
+                    lapack_complex_double* tauq1,
+                    lapack_complex_double* tauq2,
+                    lapack_complex_double* work, lapack_int* lwork , lapack_int *info );
+void LAPACK_zuncsd( char* jobu1, char* jobu2,
+                    char* jobv1t, char* jobv2t, char* trans,
+                    char* signs, lapack_int* m, lapack_int* p,
+                    lapack_int* q, lapack_complex_double* x11,
+                    lapack_int* ldx11, lapack_complex_double* x12,
+                    lapack_int* ldx12, lapack_complex_double* x21,
+                    lapack_int* ldx21, lapack_complex_double* x22,
+                    lapack_int* ldx22, double* theta,
+                    lapack_complex_double* u1, lapack_int* ldu1,
+                    lapack_complex_double* u2, lapack_int* ldu2,
+                    lapack_complex_double* v1t, lapack_int* ldv1t,
+                    lapack_complex_double* v2t, lapack_int* ldv2t,
+                    lapack_complex_double* work, lapack_int* lwork,
+                    double* rwork, lapack_int* lrwork,
+                    lapack_int* iwork , lapack_int *info );
+// LAPACK 3.4.0
+void LAPACK_sgemqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* nb, const float* v,
+                     lapack_int* ldv, const float* t, lapack_int* ldt, float* c,
+                     lapack_int* ldc, float* work, lapack_int *info );
+void LAPACK_dgemqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* nb, const double* v,
+                     lapack_int* ldv, const double* t, lapack_int* ldt,
+                     double* c, lapack_int* ldc, double* work,
+                     lapack_int *info );
+void LAPACK_cgemqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* nb,
+                     const lapack_complex_float* v, lapack_int* ldv,
+                     const lapack_complex_float* t, lapack_int* ldt,
+                     lapack_complex_float* c, lapack_int* ldc,
+                     lapack_complex_float* work, lapack_int *info );
+void LAPACK_zgemqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* nb,
+                     const lapack_complex_double* v, lapack_int* ldv,
+                     const lapack_complex_double* t, lapack_int* ldt,
+                     lapack_complex_double* c, lapack_int* ldc,
+                     lapack_complex_double* work, lapack_int *info );
+void LAPACK_sgeqrt( lapack_int* m, lapack_int* n, lapack_int* nb, float* a,
+                    lapack_int* lda, float* t, lapack_int* ldt, float* work,
+                    lapack_int *info );
+void LAPACK_dgeqrt( lapack_int* m, lapack_int* n, lapack_int* nb, double* a,
+                    lapack_int* lda, double* t, lapack_int* ldt, double* work,
+                    lapack_int *info );
+void LAPACK_cgeqrt( lapack_int* m, lapack_int* n, lapack_int* nb,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_zgeqrt( lapack_int* m, lapack_int* n, lapack_int* nb,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_sgeqrt2( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                     float* t, lapack_int* ldt, lapack_int *info );
+void LAPACK_dgeqrt2( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                     double* t, lapack_int* ldt, lapack_int *info );
+void LAPACK_cgeqrt2( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                     lapack_int* lda, lapack_complex_float* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_zgeqrt2( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                     lapack_int* lda, lapack_complex_double* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_sgeqrt3( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                     float* t, lapack_int* ldt, lapack_int *info );
+void LAPACK_dgeqrt3( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                     double* t, lapack_int* ldt, lapack_int *info );
+void LAPACK_cgeqrt3( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                     lapack_int* lda, lapack_complex_float* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_zgeqrt3( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                     lapack_int* lda, lapack_complex_double* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_stpmqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* l, lapack_int* nb,
+                     const float* v, lapack_int* ldv, const float* t,
+                     lapack_int* ldt, float* a, lapack_int* lda, float* b,
+                     lapack_int* ldb, float* work, lapack_int *info );
+void LAPACK_dtpmqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* l, lapack_int* nb,
+                     const double* v, lapack_int* ldv, const double* t,
+                     lapack_int* ldt, double* a, lapack_int* lda, double* b,
+                     lapack_int* ldb, double* work, lapack_int *info );
+void LAPACK_ctpmqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* l, lapack_int* nb,
+                     const lapack_complex_float* v, lapack_int* ldv,
+                     const lapack_complex_float* t, lapack_int* ldt,
+                     lapack_complex_float* a, lapack_int* lda,
+                     lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* work, lapack_int *info );
+void LAPACK_ztpmqrt( char* side, char* trans, lapack_int* m, lapack_int* n,
+                     lapack_int* k, lapack_int* l, lapack_int* nb,
+                     const lapack_complex_double* v, lapack_int* ldv,
+                     const lapack_complex_double* t, lapack_int* ldt,
+                     lapack_complex_double* a, lapack_int* lda,
+                     lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* work, lapack_int *info );
+void LAPACK_dtpqrt( lapack_int* m, lapack_int* n, lapack_int* l, lapack_int* nb,
+                    double* a, lapack_int* lda, double* b, lapack_int* ldb,
+                    double* t, lapack_int* ldt, double* work,
+                    lapack_int *info );
+void LAPACK_ctpqrt( lapack_int* m, lapack_int* n, lapack_int* l, lapack_int* nb,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* t, lapack_complex_float* b,
+                    lapack_int* ldb, lapack_int* ldt,
+                    lapack_complex_float* work, lapack_int *info );
+void LAPACK_ztpqrt( lapack_int* m, lapack_int* n, lapack_int* l, lapack_int* nb,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* work, lapack_int *info );
+void LAPACK_stpqrt2( lapack_int* m, lapack_int* n, float* a, lapack_int* lda,
+                     float* b, lapack_int* ldb, float* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_dtpqrt2( lapack_int* m, lapack_int* n, double* a, lapack_int* lda,
+                     double* b, lapack_int* ldb, double* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_ctpqrt2( lapack_int* m, lapack_int* n, lapack_complex_float* a,
+                     lapack_int* lda, lapack_complex_float* b, lapack_int* ldb,
+                     lapack_complex_float* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_ztpqrt2( lapack_int* m, lapack_int* n, lapack_complex_double* a,
+                     lapack_int* lda, lapack_complex_double* b, lapack_int* ldb,
+                     lapack_complex_double* t, lapack_int* ldt,
+                     lapack_int *info );
+void LAPACK_stprfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k, lapack_int* l,
+                    const float* v, lapack_int* ldv, const float* t,
+                    lapack_int* ldt, float* a, lapack_int* lda, float* b,
+                    lapack_int* ldb, const float* mywork,
+                    lapack_int* myldwork );
+void LAPACK_dtprfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k, lapack_int* l,
+                    const double* v, lapack_int* ldv, const double* t,
+                    lapack_int* ldt, double* a, lapack_int* lda, double* b,
+                    lapack_int* ldb, const double* mywork,
+                    lapack_int* myldwork );
+void LAPACK_ctprfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k, lapack_int* l,
+                    const lapack_complex_float* v, lapack_int* ldv,
+                    const lapack_complex_float* t, lapack_int* ldt,
+                    lapack_complex_float* a, lapack_int* lda,
+                    lapack_complex_float* b, lapack_int* ldb,
+                    const float* mywork, lapack_int* myldwork );
+void LAPACK_ztprfb( char* side, char* trans, char* direct, char* storev,
+                    lapack_int* m, lapack_int* n, lapack_int* k, lapack_int* l,
+                    const lapack_complex_double* v, lapack_int* ldv,
+                    const lapack_complex_double* t, lapack_int* ldt,
+                    lapack_complex_double* a, lapack_int* lda,
+                    lapack_complex_double* b, lapack_int* ldb,
+                    const double* mywork, lapack_int* myldwork );
+// LAPACK 3.X.X
+void LAPACK_csyr( char* uplo, lapack_int* n, lapack_complex_float* alpha,
+                      const lapack_complex_float* x, lapack_int* incx,
+                      lapack_complex_float* a, lapack_int* lda );
+void LAPACK_zsyr( char* uplo, lapack_int* n, lapack_complex_double* alpha,
+                      const lapack_complex_double* x, lapack_int* incx,
+                      lapack_complex_double* a, lapack_int* lda );
+
+#ifdef __cplusplus
+}
+#endif /* __cplusplus */
+
+#endif /* _LAPACKE_H_ */
+
+#endif /* _MKL_LAPACKE_H_ */
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke_mangling.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke_mangling.h
new file mode 100644
index 0000000..6211fd1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/misc/lapacke_mangling.h
@@ -0,0 +1,17 @@
+#ifndef LAPACK_HEADER_INCLUDED
+#define LAPACK_HEADER_INCLUDED
+
+#ifndef LAPACK_GLOBAL
+#if defined(LAPACK_GLOBAL_PATTERN_LC) || defined(ADD_)
+#define LAPACK_GLOBAL(lcname,UCNAME)  lcname##_
+#elif defined(LAPACK_GLOBAL_PATTERN_UC) || defined(UPPER)
+#define LAPACK_GLOBAL(lcname,UCNAME)  UCNAME
+#elif defined(LAPACK_GLOBAL_PATTERN_MC) || defined(NOCHANGE)
+#define LAPACK_GLOBAL(lcname,UCNAME)  lcname
+#else
+#define LAPACK_GLOBAL(lcname,UCNAME)  lcname##_
+#endif
+#endif
+
+#endif
+
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseBinaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseBinaryOps.h
new file mode 100644
index 0000000..1f8a531
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseBinaryOps.h
@@ -0,0 +1,332 @@
+
+/** \returns an expression of the coefficient wise product of \c *this and \a other
+  *
+  * \sa MatrixBase::cwiseProduct
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,product)
+operator*(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,product)(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient wise quotient of \c *this and \a other
+  *
+  * \sa MatrixBase::cwiseQuotient
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_quotient_op<Scalar,typename OtherDerived::Scalar>, const Derived, const OtherDerived>
+operator/(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<internal::scalar_quotient_op<Scalar,typename OtherDerived::Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise min of \c *this and \a other
+  *
+  * Example: \include Cwise_min.cpp
+  * Output: \verbinclude Cwise_min.out
+  *
+  * \sa max()
+  */
+EIGEN_MAKE_CWISE_BINARY_OP(min,min)
+
+/** \returns an expression of the coefficient-wise min of \c *this and scalar \a other
+  *
+  * \sa max()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar>, const Derived,
+                                        const CwiseNullaryOp<internal::scalar_constant_op<Scalar>, PlainObject> >
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+min
+#else
+(min)
+#endif
+(const Scalar &other) const
+{
+  return (min)(Derived::PlainObject::Constant(rows(), cols(), other));
+}
+
+/** \returns an expression of the coefficient-wise max of \c *this and \a other
+  *
+  * Example: \include Cwise_max.cpp
+  * Output: \verbinclude Cwise_max.out
+  *
+  * \sa min()
+  */
+EIGEN_MAKE_CWISE_BINARY_OP(max,max)
+
+/** \returns an expression of the coefficient-wise max of \c *this and scalar \a other
+  *
+  * \sa min()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar>, const Derived,
+                                        const CwiseNullaryOp<internal::scalar_constant_op<Scalar>, PlainObject> >
+#ifdef EIGEN_PARSED_BY_DOXYGEN
+max
+#else
+(max)
+#endif
+(const Scalar &other) const
+{
+  return (max)(Derived::PlainObject::Constant(rows(), cols(), other));
+}
+
+/** \returns an expression of the coefficient-wise power of \c *this to the given array of \a exponents.
+  *
+  * This function computes the coefficient-wise power.
+  *
+  * Example: \include Cwise_array_power_array.cpp
+  * Output: \verbinclude Cwise_array_power_array.out
+  */
+EIGEN_MAKE_CWISE_BINARY_OP(pow,pow)
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(pow,pow)
+#else
+/** \returns an expression of the coefficients of \c *this rasied to the constant power \a exponent
+  *
+  * \tparam T is the scalar type of \a exponent. It must be compatible with the scalar type of the given expression.
+  *
+  * This function computes the coefficient-wise power. The function MatrixBase::pow() in the
+  * unsupported module MatrixFunctions computes the matrix power.
+  *
+  * Example: \include Cwise_pow.cpp
+  * Output: \verbinclude Cwise_pow.out
+  *
+  * \sa ArrayBase::pow(ArrayBase), square(), cube(), exp(), log()
+  */
+template<typename T>
+const CwiseBinaryOp<internal::scalar_pow_op<Scalar,T>,Derived,Constant<T> > pow(const T& exponent) const;
+#endif
+
+
+// TODO code generating macros could be moved to Macros.h and could include generation of documentation
+#define EIGEN_MAKE_CWISE_COMP_OP(OP, COMPARATOR) \
+template<typename OtherDerived> \
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_cmp_op<Scalar, typename OtherDerived::Scalar, internal::cmp_ ## COMPARATOR>, const Derived, const OtherDerived> \
+OP(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
+{ \
+  return CwiseBinaryOp<internal::scalar_cmp_op<Scalar, typename OtherDerived::Scalar, internal::cmp_ ## COMPARATOR>, const Derived, const OtherDerived>(derived(), other.derived()); \
+}\
+typedef CwiseBinaryOp<internal::scalar_cmp_op<Scalar,Scalar, internal::cmp_ ## COMPARATOR>, const Derived, const CwiseNullaryOp<internal::scalar_constant_op<Scalar>, PlainObject> > Cmp ## COMPARATOR ## ReturnType; \
+typedef CwiseBinaryOp<internal::scalar_cmp_op<Scalar,Scalar, internal::cmp_ ## COMPARATOR>, const CwiseNullaryOp<internal::scalar_constant_op<Scalar>, PlainObject>, const Derived > RCmp ## COMPARATOR ## ReturnType; \
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Cmp ## COMPARATOR ## ReturnType \
+OP(const Scalar& s) const { \
+  return this->OP(Derived::PlainObject::Constant(rows(), cols(), s)); \
+} \
+EIGEN_DEVICE_FUNC friend EIGEN_STRONG_INLINE const RCmp ## COMPARATOR ## ReturnType \
+OP(const Scalar& s, const Derived& d) { \
+  return Derived::PlainObject::Constant(d.rows(), d.cols(), s).OP(d); \
+}
+
+#define EIGEN_MAKE_CWISE_COMP_R_OP(OP, R_OP, RCOMPARATOR) \
+template<typename OtherDerived> \
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_cmp_op<typename OtherDerived::Scalar, Scalar, internal::cmp_##RCOMPARATOR>, const OtherDerived, const Derived> \
+OP(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const \
+{ \
+  return CwiseBinaryOp<internal::scalar_cmp_op<typename OtherDerived::Scalar, Scalar, internal::cmp_##RCOMPARATOR>, const OtherDerived, const Derived>(other.derived(), derived()); \
+} \
+EIGEN_DEVICE_FUNC \
+inline const RCmp ## RCOMPARATOR ## ReturnType \
+OP(const Scalar& s) const { \
+  return Derived::PlainObject::Constant(rows(), cols(), s).R_OP(*this); \
+} \
+friend inline const Cmp ## RCOMPARATOR ## ReturnType \
+OP(const Scalar& s, const Derived& d) { \
+  return d.R_OP(Derived::PlainObject::Constant(d.rows(), d.cols(), s)); \
+}
+
+
+
+/** \returns an expression of the coefficient-wise \< operator of *this and \a other
+  *
+  * Example: \include Cwise_less.cpp
+  * Output: \verbinclude Cwise_less.out
+  *
+  * \sa all(), any(), operator>(), operator<=()
+  */
+EIGEN_MAKE_CWISE_COMP_OP(operator<, LT)
+
+/** \returns an expression of the coefficient-wise \<= operator of *this and \a other
+  *
+  * Example: \include Cwise_less_equal.cpp
+  * Output: \verbinclude Cwise_less_equal.out
+  *
+  * \sa all(), any(), operator>=(), operator<()
+  */
+EIGEN_MAKE_CWISE_COMP_OP(operator<=, LE)
+
+/** \returns an expression of the coefficient-wise \> operator of *this and \a other
+  *
+  * Example: \include Cwise_greater.cpp
+  * Output: \verbinclude Cwise_greater.out
+  *
+  * \sa all(), any(), operator>=(), operator<()
+  */
+EIGEN_MAKE_CWISE_COMP_R_OP(operator>, operator<, LT)
+
+/** \returns an expression of the coefficient-wise \>= operator of *this and \a other
+  *
+  * Example: \include Cwise_greater_equal.cpp
+  * Output: \verbinclude Cwise_greater_equal.out
+  *
+  * \sa all(), any(), operator>(), operator<=()
+  */
+EIGEN_MAKE_CWISE_COMP_R_OP(operator>=, operator<=, LE)
+
+/** \returns an expression of the coefficient-wise == operator of *this and \a other
+  *
+  * \warning this performs an exact comparison, which is generally a bad idea with floating-point types.
+  * In order to check for equality between two vectors or matrices with floating-point coefficients, it is
+  * generally a far better idea to use a fuzzy comparison as provided by isApprox() and
+  * isMuchSmallerThan().
+  *
+  * Example: \include Cwise_equal_equal.cpp
+  * Output: \verbinclude Cwise_equal_equal.out
+  *
+  * \sa all(), any(), isApprox(), isMuchSmallerThan()
+  */
+EIGEN_MAKE_CWISE_COMP_OP(operator==, EQ)
+
+/** \returns an expression of the coefficient-wise != operator of *this and \a other
+  *
+  * \warning this performs an exact comparison, which is generally a bad idea with floating-point types.
+  * In order to check for equality between two vectors or matrices with floating-point coefficients, it is
+  * generally a far better idea to use a fuzzy comparison as provided by isApprox() and
+  * isMuchSmallerThan().
+  *
+  * Example: \include Cwise_not_equal.cpp
+  * Output: \verbinclude Cwise_not_equal.out
+  *
+  * \sa all(), any(), isApprox(), isMuchSmallerThan()
+  */
+EIGEN_MAKE_CWISE_COMP_OP(operator!=, NEQ)
+
+
+#undef EIGEN_MAKE_CWISE_COMP_OP
+#undef EIGEN_MAKE_CWISE_COMP_R_OP
+
+// scalar addition
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_MAKE_SCALAR_BINARY_OP(operator+,sum)
+#else
+/** \returns an expression of \c *this with each coeff incremented by the constant \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  *
+  * Example: \include Cwise_plus.cpp
+  * Output: \verbinclude Cwise_plus.out
+  *
+  * \sa operator+=(), operator-()
+  */
+template<typename T>
+const CwiseBinaryOp<internal::scalar_sum_op<Scalar,T>,Derived,Constant<T> > operator+(const T& scalar) const;
+/** \returns an expression of \a expr with each coeff incremented by the constant \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  */
+template<typename T> friend
+const CwiseBinaryOp<internal::scalar_sum_op<T,Scalar>,Constant<T>,Derived> operator+(const T& scalar, const StorageBaseType& expr);
+#endif
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_MAKE_SCALAR_BINARY_OP(operator-,difference)
+#else
+/** \returns an expression of \c *this with each coeff decremented by the constant \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  *
+  * Example: \include Cwise_minus.cpp
+  * Output: \verbinclude Cwise_minus.out
+  *
+  * \sa operator+=(), operator-()
+  */
+template<typename T>
+const CwiseBinaryOp<internal::scalar_difference_op<Scalar,T>,Derived,Constant<T> > operator-(const T& scalar) const;
+/** \returns an expression of the constant matrix of value \a scalar decremented by the coefficients of \a expr
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  */
+template<typename T> friend
+const CwiseBinaryOp<internal::scalar_difference_op<T,Scalar>,Constant<T>,Derived> operator-(const T& scalar, const StorageBaseType& expr);
+#endif
+
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+  EIGEN_MAKE_SCALAR_BINARY_OP_ONTHELEFT(operator/,quotient)
+#else
+  /**
+    * \brief Component-wise division of the scalar \a s by array elements of \a a.
+    *
+    * \tparam Scalar is the scalar type of \a x. It must be compatible with the scalar type of the given array expression (\c Derived::Scalar).
+    */
+  template<typename T> friend
+  inline const CwiseBinaryOp<internal::scalar_quotient_op<T,Scalar>,Constant<T>,Derived>
+  operator/(const T& s,const StorageBaseType& a);
+#endif
+
+/** \returns an expression of the coefficient-wise ^ operator of *this and \a other
+ *
+ * \warning this operator is for expression of bool only.
+ *
+ * Example: \include Cwise_boolean_xor.cpp
+ * Output: \verbinclude Cwise_boolean_xor.out
+ *
+ * \sa operator&&(), select()
+ */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+inline const CwiseBinaryOp<internal::scalar_boolean_xor_op, const Derived, const OtherDerived>
+operator^(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<bool,Scalar>::value && internal::is_same<bool,typename OtherDerived::Scalar>::value),
+                      THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL);
+  return CwiseBinaryOp<internal::scalar_boolean_xor_op, const Derived, const OtherDerived>(derived(),other.derived());
+}
+
+// NOTE disabled until we agree on argument order
+#if 0
+/** \cpp11 \returns an expression of the coefficient-wise polygamma function.
+  *
+  * \specialfunctions_module
+  *
+  * It returns the \a n -th derivative of the digamma(psi) evaluated at \c *this.
+  *
+  * \warning Be careful with the order of the parameters: x.polygamma(n) is equivalent to polygamma(n,x)
+  *
+  * \sa Eigen::polygamma()
+  */
+template<typename DerivedN>
+inline const CwiseBinaryOp<internal::scalar_polygamma_op<Scalar>, const DerivedN, const Derived>
+polygamma(const EIGEN_CURRENT_STORAGE_BASE_CLASS<DerivedN> &n) const
+{
+  return CwiseBinaryOp<internal::scalar_polygamma_op<Scalar>, const DerivedN, const Derived>(n.derived(), this->derived());
+}
+#endif
+
+/** \returns an expression of the coefficient-wise zeta function.
+  *
+  * \specialfunctions_module
+  *
+  * It returns the Riemann zeta function of two arguments \c *this and \a q:
+  *
+  * \param *this is the exposent, it must be > 1
+  * \param q is the shift, it must be > 0
+  *
+  * \note This function supports only float and double scalar types. To support other scalar types, the user has
+  * to provide implementations of zeta(T,T) for any scalar type T to be supported.
+  *
+  * This method is an alias for zeta(*this,q);
+  *
+  * \sa Eigen::zeta()
+  */
+template<typename DerivedQ>
+inline const CwiseBinaryOp<internal::scalar_zeta_op<Scalar>, const Derived, const DerivedQ>
+zeta(const EIGEN_CURRENT_STORAGE_BASE_CLASS<DerivedQ> &q) const
+{
+  return CwiseBinaryOp<internal::scalar_zeta_op<Scalar>, const Derived, const DerivedQ>(this->derived(), q.derived());
+}
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseUnaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseUnaryOps.h
new file mode 100644
index 0000000..ebaa3f1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/ArrayCwiseUnaryOps.h
@@ -0,0 +1,552 @@
+
+
+typedef CwiseUnaryOp<internal::scalar_abs_op<Scalar>, const Derived> AbsReturnType;
+typedef CwiseUnaryOp<internal::scalar_arg_op<Scalar>, const Derived> ArgReturnType;
+typedef CwiseUnaryOp<internal::scalar_abs2_op<Scalar>, const Derived> Abs2ReturnType;
+typedef CwiseUnaryOp<internal::scalar_sqrt_op<Scalar>, const Derived> SqrtReturnType;
+typedef CwiseUnaryOp<internal::scalar_rsqrt_op<Scalar>, const Derived> RsqrtReturnType;
+typedef CwiseUnaryOp<internal::scalar_sign_op<Scalar>, const Derived> SignReturnType;
+typedef CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const Derived> InverseReturnType;
+typedef CwiseUnaryOp<internal::scalar_boolean_not_op<Scalar>, const Derived> BooleanNotReturnType;
+
+typedef CwiseUnaryOp<internal::scalar_exp_op<Scalar>, const Derived> ExpReturnType;
+typedef CwiseUnaryOp<internal::scalar_log_op<Scalar>, const Derived> LogReturnType;
+typedef CwiseUnaryOp<internal::scalar_log1p_op<Scalar>, const Derived> Log1pReturnType;
+typedef CwiseUnaryOp<internal::scalar_log10_op<Scalar>, const Derived> Log10ReturnType;
+typedef CwiseUnaryOp<internal::scalar_cos_op<Scalar>, const Derived> CosReturnType;
+typedef CwiseUnaryOp<internal::scalar_sin_op<Scalar>, const Derived> SinReturnType;
+typedef CwiseUnaryOp<internal::scalar_tan_op<Scalar>, const Derived> TanReturnType;
+typedef CwiseUnaryOp<internal::scalar_acos_op<Scalar>, const Derived> AcosReturnType;
+typedef CwiseUnaryOp<internal::scalar_asin_op<Scalar>, const Derived> AsinReturnType;
+typedef CwiseUnaryOp<internal::scalar_atan_op<Scalar>, const Derived> AtanReturnType;
+typedef CwiseUnaryOp<internal::scalar_tanh_op<Scalar>, const Derived> TanhReturnType;
+typedef CwiseUnaryOp<internal::scalar_sinh_op<Scalar>, const Derived> SinhReturnType;
+typedef CwiseUnaryOp<internal::scalar_cosh_op<Scalar>, const Derived> CoshReturnType;
+typedef CwiseUnaryOp<internal::scalar_square_op<Scalar>, const Derived> SquareReturnType;
+typedef CwiseUnaryOp<internal::scalar_cube_op<Scalar>, const Derived> CubeReturnType;
+typedef CwiseUnaryOp<internal::scalar_round_op<Scalar>, const Derived> RoundReturnType;
+typedef CwiseUnaryOp<internal::scalar_floor_op<Scalar>, const Derived> FloorReturnType;
+typedef CwiseUnaryOp<internal::scalar_ceil_op<Scalar>, const Derived> CeilReturnType;
+typedef CwiseUnaryOp<internal::scalar_isnan_op<Scalar>, const Derived> IsNaNReturnType;
+typedef CwiseUnaryOp<internal::scalar_isinf_op<Scalar>, const Derived> IsInfReturnType;
+typedef CwiseUnaryOp<internal::scalar_isfinite_op<Scalar>, const Derived> IsFiniteReturnType;
+
+/** \returns an expression of the coefficient-wise absolute value of \c *this
+  *
+  * Example: \include Cwise_abs.cpp
+  * Output: \verbinclude Cwise_abs.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_abs">Math functions</a>, abs2()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const AbsReturnType
+abs() const
+{
+  return AbsReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise phase angle of \c *this
+  *
+  * Example: \include Cwise_arg.cpp
+  * Output: \verbinclude Cwise_arg.out
+  *
+  * \sa abs()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const ArgReturnType
+arg() const
+{
+  return ArgReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise squared absolute value of \c *this
+  *
+  * Example: \include Cwise_abs2.cpp
+  * Output: \verbinclude Cwise_abs2.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_abs2">Math functions</a>, abs(), square()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const Abs2ReturnType
+abs2() const
+{
+  return Abs2ReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise exponential of *this.
+  *
+  * This function computes the coefficient-wise exponential. The function MatrixBase::exp() in the
+  * unsupported module MatrixFunctions computes the matrix exponential.
+  *
+  * Example: \include Cwise_exp.cpp
+  * Output: \verbinclude Cwise_exp.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_exp">Math functions</a>, pow(), log(), sin(), cos()
+  */
+EIGEN_DEVICE_FUNC
+inline const ExpReturnType
+exp() const
+{
+  return ExpReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise logarithm of *this.
+  *
+  * This function computes the coefficient-wise logarithm. The function MatrixBase::log() in the
+  * unsupported module MatrixFunctions computes the matrix logarithm.
+  *
+  * Example: \include Cwise_log.cpp
+  * Output: \verbinclude Cwise_log.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_log">Math functions</a>, exp()
+  */
+EIGEN_DEVICE_FUNC
+inline const LogReturnType
+log() const
+{
+  return LogReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise logarithm of 1 plus \c *this.
+  *
+  * In exact arithmetic, \c x.log() is equivalent to \c (x+1).log(),
+  * however, with finite precision, this function is much more accurate when \c x is close to zero.
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_log1p">Math functions</a>, log()
+  */
+EIGEN_DEVICE_FUNC
+inline const Log1pReturnType
+log1p() const
+{
+  return Log1pReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise base-10 logarithm of *this.
+  *
+  * This function computes the coefficient-wise base-10 logarithm.
+  *
+  * Example: \include Cwise_log10.cpp
+  * Output: \verbinclude Cwise_log10.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_log10">Math functions</a>, log()
+  */
+EIGEN_DEVICE_FUNC
+inline const Log10ReturnType
+log10() const
+{
+  return Log10ReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise square root of *this.
+  *
+  * This function computes the coefficient-wise square root. The function MatrixBase::sqrt() in the
+  * unsupported module MatrixFunctions computes the matrix square root.
+  *
+  * Example: \include Cwise_sqrt.cpp
+  * Output: \verbinclude Cwise_sqrt.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_sqrt">Math functions</a>, pow(), square()
+  */
+EIGEN_DEVICE_FUNC
+inline const SqrtReturnType
+sqrt() const
+{
+  return SqrtReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise inverse square root of *this.
+  *
+  * This function computes the coefficient-wise inverse square root.
+  *
+  * Example: \include Cwise_sqrt.cpp
+  * Output: \verbinclude Cwise_sqrt.out
+  *
+  * \sa pow(), square()
+  */
+EIGEN_DEVICE_FUNC
+inline const RsqrtReturnType
+rsqrt() const
+{
+  return RsqrtReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise signum of *this.
+  *
+  * This function computes the coefficient-wise signum.
+  *
+  * Example: \include Cwise_sign.cpp
+  * Output: \verbinclude Cwise_sign.out
+  *
+  * \sa pow(), square()
+  */
+EIGEN_DEVICE_FUNC
+inline const SignReturnType
+sign() const
+{
+  return SignReturnType(derived());
+}
+
+
+/** \returns an expression of the coefficient-wise cosine of *this.
+  *
+  * This function computes the coefficient-wise cosine. The function MatrixBase::cos() in the
+  * unsupported module MatrixFunctions computes the matrix cosine.
+  *
+  * Example: \include Cwise_cos.cpp
+  * Output: \verbinclude Cwise_cos.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_cos">Math functions</a>, sin(), acos()
+  */
+EIGEN_DEVICE_FUNC
+inline const CosReturnType
+cos() const
+{
+  return CosReturnType(derived());
+}
+
+
+/** \returns an expression of the coefficient-wise sine of *this.
+  *
+  * This function computes the coefficient-wise sine. The function MatrixBase::sin() in the
+  * unsupported module MatrixFunctions computes the matrix sine.
+  *
+  * Example: \include Cwise_sin.cpp
+  * Output: \verbinclude Cwise_sin.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_sin">Math functions</a>, cos(), asin()
+  */
+EIGEN_DEVICE_FUNC
+inline const SinReturnType
+sin() const
+{
+  return SinReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise tan of *this.
+  *
+  * Example: \include Cwise_tan.cpp
+  * Output: \verbinclude Cwise_tan.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_tan">Math functions</a>, cos(), sin()
+  */
+EIGEN_DEVICE_FUNC
+inline const TanReturnType
+tan() const
+{
+  return TanReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise arc tan of *this.
+  *
+  * Example: \include Cwise_atan.cpp
+  * Output: \verbinclude Cwise_atan.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_atan">Math functions</a>, tan(), asin(), acos()
+  */
+EIGEN_DEVICE_FUNC
+inline const AtanReturnType
+atan() const
+{
+  return AtanReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise arc cosine of *this.
+  *
+  * Example: \include Cwise_acos.cpp
+  * Output: \verbinclude Cwise_acos.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_acos">Math functions</a>, cos(), asin()
+  */
+EIGEN_DEVICE_FUNC
+inline const AcosReturnType
+acos() const
+{
+  return AcosReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise arc sine of *this.
+  *
+  * Example: \include Cwise_asin.cpp
+  * Output: \verbinclude Cwise_asin.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_asin">Math functions</a>, sin(), acos()
+  */
+EIGEN_DEVICE_FUNC
+inline const AsinReturnType
+asin() const
+{
+  return AsinReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise hyperbolic tan of *this.
+  *
+  * Example: \include Cwise_tanh.cpp
+  * Output: \verbinclude Cwise_tanh.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_tanh">Math functions</a>, tan(), sinh(), cosh()
+  */
+EIGEN_DEVICE_FUNC
+inline const TanhReturnType
+tanh() const
+{
+  return TanhReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise hyperbolic sin of *this.
+  *
+  * Example: \include Cwise_sinh.cpp
+  * Output: \verbinclude Cwise_sinh.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_sinh">Math functions</a>, sin(), tanh(), cosh()
+  */
+EIGEN_DEVICE_FUNC
+inline const SinhReturnType
+sinh() const
+{
+  return SinhReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise hyperbolic cos of *this.
+  *
+  * Example: \include Cwise_cosh.cpp
+  * Output: \verbinclude Cwise_cosh.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_cosh">Math functions</a>, tan(), sinh(), cosh()
+  */
+EIGEN_DEVICE_FUNC
+inline const CoshReturnType
+cosh() const
+{
+  return CoshReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise inverse of *this.
+  *
+  * Example: \include Cwise_inverse.cpp
+  * Output: \verbinclude Cwise_inverse.out
+  *
+  * \sa operator/(), operator*()
+  */
+EIGEN_DEVICE_FUNC
+inline const InverseReturnType
+inverse() const
+{
+  return InverseReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise square of *this.
+  *
+  * Example: \include Cwise_square.cpp
+  * Output: \verbinclude Cwise_square.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_squareE">Math functions</a>, abs2(), cube(), pow()
+  */
+EIGEN_DEVICE_FUNC
+inline const SquareReturnType
+square() const
+{
+  return SquareReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise cube of *this.
+  *
+  * Example: \include Cwise_cube.cpp
+  * Output: \verbinclude Cwise_cube.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_cube">Math functions</a>, square(), pow()
+  */
+EIGEN_DEVICE_FUNC
+inline const CubeReturnType
+cube() const
+{
+  return CubeReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise round of *this.
+  *
+  * Example: \include Cwise_round.cpp
+  * Output: \verbinclude Cwise_round.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_round">Math functions</a>, ceil(), floor()
+  */
+EIGEN_DEVICE_FUNC
+inline const RoundReturnType
+round() const
+{
+  return RoundReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise floor of *this.
+  *
+  * Example: \include Cwise_floor.cpp
+  * Output: \verbinclude Cwise_floor.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_floor">Math functions</a>, ceil(), round()
+  */
+EIGEN_DEVICE_FUNC
+inline const FloorReturnType
+floor() const
+{
+  return FloorReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise ceil of *this.
+  *
+  * Example: \include Cwise_ceil.cpp
+  * Output: \verbinclude Cwise_ceil.out
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_ceil">Math functions</a>, floor(), round()
+  */
+EIGEN_DEVICE_FUNC
+inline const CeilReturnType
+ceil() const
+{
+  return CeilReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise isnan of *this.
+  *
+  * Example: \include Cwise_isNaN.cpp
+  * Output: \verbinclude Cwise_isNaN.out
+  *
+  * \sa isfinite(), isinf()
+  */
+EIGEN_DEVICE_FUNC
+inline const IsNaNReturnType
+isNaN() const
+{
+  return IsNaNReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise isinf of *this.
+  *
+  * Example: \include Cwise_isInf.cpp
+  * Output: \verbinclude Cwise_isInf.out
+  *
+  * \sa isnan(), isfinite()
+  */
+EIGEN_DEVICE_FUNC
+inline const IsInfReturnType
+isInf() const
+{
+  return IsInfReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise isfinite of *this.
+  *
+  * Example: \include Cwise_isFinite.cpp
+  * Output: \verbinclude Cwise_isFinite.out
+  *
+  * \sa isnan(), isinf()
+  */
+EIGEN_DEVICE_FUNC
+inline const IsFiniteReturnType
+isFinite() const
+{
+  return IsFiniteReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise ! operator of *this
+  *
+  * \warning this operator is for expression of bool only.
+  *
+  * Example: \include Cwise_boolean_not.cpp
+  * Output: \verbinclude Cwise_boolean_not.out
+  *
+  * \sa operator!=()
+  */
+EIGEN_DEVICE_FUNC
+inline const BooleanNotReturnType
+operator!() const
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<bool,Scalar>::value),
+                      THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL);
+  return BooleanNotReturnType(derived());
+}
+
+
+// --- SpecialFunctions module ---
+
+typedef CwiseUnaryOp<internal::scalar_lgamma_op<Scalar>, const Derived> LgammaReturnType;
+typedef CwiseUnaryOp<internal::scalar_digamma_op<Scalar>, const Derived> DigammaReturnType;
+typedef CwiseUnaryOp<internal::scalar_erf_op<Scalar>, const Derived> ErfReturnType;
+typedef CwiseUnaryOp<internal::scalar_erfc_op<Scalar>, const Derived> ErfcReturnType;
+
+/** \cpp11 \returns an expression of the coefficient-wise ln(|gamma(*this)|).
+  *
+  * \specialfunctions_module
+  *
+  * Example: \include Cwise_lgamma.cpp
+  * Output: \verbinclude Cwise_lgamma.out
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of lgamma(T) for any scalar
+  * type T to be supported.
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_lgamma">Math functions</a>, digamma()
+  */
+EIGEN_DEVICE_FUNC
+inline const LgammaReturnType
+lgamma() const
+{
+  return LgammaReturnType(derived());
+}
+
+/** \returns an expression of the coefficient-wise digamma (psi, derivative of lgamma).
+  *
+  * \specialfunctions_module
+  *
+  * \note This function supports only float and double scalar types. To support other scalar types,
+  * the user has to provide implementations of digamma(T) for any scalar
+  * type T to be supported.
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_digamma">Math functions</a>, Eigen::digamma(), Eigen::polygamma(), lgamma()
+  */
+EIGEN_DEVICE_FUNC
+inline const DigammaReturnType
+digamma() const
+{
+  return DigammaReturnType(derived());
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise Gauss error
+  * function of *this.
+  *
+  * \specialfunctions_module
+  *
+  * Example: \include Cwise_erf.cpp
+  * Output: \verbinclude Cwise_erf.out
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of erf(T) for any scalar
+  * type T to be supported.
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_erf">Math functions</a>, erfc()
+  */
+EIGEN_DEVICE_FUNC
+inline const ErfReturnType
+erf() const
+{
+  return ErfReturnType(derived());
+}
+
+/** \cpp11 \returns an expression of the coefficient-wise Complementary error
+  * function of *this.
+  *
+  * \specialfunctions_module
+  *
+  * Example: \include Cwise_erfc.cpp
+  * Output: \verbinclude Cwise_erfc.out
+  *
+  * \note This function supports only float and double scalar types in c++11 mode. To support other scalar types,
+  * or float/double in non c++11 mode, the user has to provide implementations of erfc(T) for any scalar
+  * type T to be supported.
+  *
+  * \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_erfc">Math functions</a>, erf()
+  */
+EIGEN_DEVICE_FUNC
+inline const ErfcReturnType
+erfc() const
+{
+  return ErfcReturnType(derived());
+}
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/BlockMethods.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/BlockMethods.h
new file mode 100644
index 0000000..ac35a00
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/BlockMethods.h
@@ -0,0 +1,1058 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2010 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2010 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+/// \internal expression type of a column */
+typedef Block<Derived, internal::traits<Derived>::RowsAtCompileTime, 1, !IsRowMajor> ColXpr;
+typedef const Block<const Derived, internal::traits<Derived>::RowsAtCompileTime, 1, !IsRowMajor> ConstColXpr;
+/// \internal expression type of a row */
+typedef Block<Derived, 1, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> RowXpr;
+typedef const Block<const Derived, 1, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> ConstRowXpr;
+/// \internal expression type of a block of whole columns */
+typedef Block<Derived, internal::traits<Derived>::RowsAtCompileTime, Dynamic, !IsRowMajor> ColsBlockXpr;
+typedef const Block<const Derived, internal::traits<Derived>::RowsAtCompileTime, Dynamic, !IsRowMajor> ConstColsBlockXpr;
+/// \internal expression type of a block of whole rows */
+typedef Block<Derived, Dynamic, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> RowsBlockXpr;
+typedef const Block<const Derived, Dynamic, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> ConstRowsBlockXpr;
+/// \internal expression type of a block of whole columns */
+template<int N> struct NColsBlockXpr { typedef Block<Derived, internal::traits<Derived>::RowsAtCompileTime, N, !IsRowMajor> Type; };
+template<int N> struct ConstNColsBlockXpr { typedef const Block<const Derived, internal::traits<Derived>::RowsAtCompileTime, N, !IsRowMajor> Type; };
+/// \internal expression type of a block of whole rows */
+template<int N> struct NRowsBlockXpr { typedef Block<Derived, N, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> Type; };
+template<int N> struct ConstNRowsBlockXpr { typedef const Block<const Derived, N, internal::traits<Derived>::ColsAtCompileTime, IsRowMajor> Type; };
+/// \internal expression of a block */
+typedef Block<Derived> BlockXpr;
+typedef const Block<const Derived> ConstBlockXpr;
+/// \internal expression of a block of fixed sizes */
+template<int Rows, int Cols> struct FixedBlockXpr { typedef Block<Derived,Rows,Cols> Type; };
+template<int Rows, int Cols> struct ConstFixedBlockXpr { typedef Block<const Derived,Rows,Cols> Type; };
+
+typedef VectorBlock<Derived> SegmentReturnType;
+typedef const VectorBlock<const Derived> ConstSegmentReturnType;
+template<int Size> struct FixedSegmentReturnType { typedef VectorBlock<Derived, Size> Type; };
+template<int Size> struct ConstFixedSegmentReturnType { typedef const VectorBlock<const Derived, Size> Type; };
+
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+/// \returns a dynamic-size expression of a block in *this.
+///
+/// \param startRow the first row in the block
+/// \param startCol the first column in the block
+/// \param blockRows the number of rows in the block
+/// \param blockCols the number of columns in the block
+///
+/// Example: \include MatrixBase_block_int_int_int_int.cpp
+/// Output: \verbinclude MatrixBase_block_int_int_int_int.out
+///
+/// \note Even though the returned expression has dynamic size, in the case
+/// when it is applied to a fixed-size matrix, it inherits a fixed maximal size,
+/// which means that evaluating it does not cause a dynamic memory allocation.
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline BlockXpr block(Index startRow, Index startCol, Index blockRows, Index blockCols)
+{
+  return BlockXpr(derived(), startRow, startCol, blockRows, blockCols);
+}
+
+/// This is the const version of block(Index,Index,Index,Index). */
+EIGEN_DEVICE_FUNC
+inline const ConstBlockXpr block(Index startRow, Index startCol, Index blockRows, Index blockCols) const
+{
+  return ConstBlockXpr(derived(), startRow, startCol, blockRows, blockCols);
+}
+
+
+
+
+/// \returns a dynamic-size expression of a top-right corner of *this.
+///
+/// \param cRows the number of rows in the corner
+/// \param cCols the number of columns in the corner
+///
+/// Example: \include MatrixBase_topRightCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_topRightCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline BlockXpr topRightCorner(Index cRows, Index cCols)
+{
+  return BlockXpr(derived(), 0, cols() - cCols, cRows, cCols);
+}
+
+/// This is the const version of topRightCorner(Index, Index).
+EIGEN_DEVICE_FUNC
+inline const ConstBlockXpr topRightCorner(Index cRows, Index cCols) const
+{
+  return ConstBlockXpr(derived(), 0, cols() - cCols, cRows, cCols);
+}
+
+/// \returns an expression of a fixed-size top-right corner of *this.
+///
+/// \tparam CRows the number of rows in the corner
+/// \tparam CCols the number of columns in the corner
+///
+/// Example: \include MatrixBase_template_int_int_topRightCorner.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_topRightCorner.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block<int,int>(Index,Index)
+///
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline typename FixedBlockXpr<CRows,CCols>::Type topRightCorner()
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), 0, cols() - CCols);
+}
+
+/// This is the const version of topRightCorner<int, int>().
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type topRightCorner() const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), 0, cols() - CCols);
+}
+
+/// \returns an expression of a top-right corner of *this.
+///
+/// \tparam CRows number of rows in corner as specified at compile-time
+/// \tparam CCols number of columns in corner as specified at compile-time
+/// \param  cRows number of rows in corner as specified at run-time
+/// \param  cCols number of columns in corner as specified at run-time
+///
+/// This function is mainly useful for corners where the number of rows is specified at compile-time
+/// and the number of columns is specified at run-time, or vice versa. The compile-time and run-time
+/// information should not contradict. In other words, \a cRows should equal \a CRows unless
+/// \a CRows is \a Dynamic, and the same for the number of columns.
+///
+/// Example: \include MatrixBase_template_int_int_topRightCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_topRightCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block
+///
+template<int CRows, int CCols>
+inline typename FixedBlockXpr<CRows,CCols>::Type topRightCorner(Index cRows, Index cCols)
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), 0, cols() - cCols, cRows, cCols);
+}
+
+/// This is the const version of topRightCorner<int, int>(Index, Index).
+template<int CRows, int CCols>
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type topRightCorner(Index cRows, Index cCols) const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), 0, cols() - cCols, cRows, cCols);
+}
+
+
+
+/// \returns a dynamic-size expression of a top-left corner of *this.
+///
+/// \param cRows the number of rows in the corner
+/// \param cCols the number of columns in the corner
+///
+/// Example: \include MatrixBase_topLeftCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_topLeftCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline BlockXpr topLeftCorner(Index cRows, Index cCols)
+{
+  return BlockXpr(derived(), 0, 0, cRows, cCols);
+}
+
+/// This is the const version of topLeftCorner(Index, Index).
+EIGEN_DEVICE_FUNC
+inline const ConstBlockXpr topLeftCorner(Index cRows, Index cCols) const
+{
+  return ConstBlockXpr(derived(), 0, 0, cRows, cCols);
+}
+
+/// \returns an expression of a fixed-size top-left corner of *this.
+///
+/// The template parameters CRows and CCols are the number of rows and columns in the corner.
+///
+/// Example: \include MatrixBase_template_int_int_topLeftCorner.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_topLeftCorner.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline typename FixedBlockXpr<CRows,CCols>::Type topLeftCorner()
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), 0, 0);
+}
+
+/// This is the const version of topLeftCorner<int, int>().
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type topLeftCorner() const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), 0, 0);
+}
+
+/// \returns an expression of a top-left corner of *this.
+///
+/// \tparam CRows number of rows in corner as specified at compile-time
+/// \tparam CCols number of columns in corner as specified at compile-time
+/// \param  cRows number of rows in corner as specified at run-time
+/// \param  cCols number of columns in corner as specified at run-time
+///
+/// This function is mainly useful for corners where the number of rows is specified at compile-time
+/// and the number of columns is specified at run-time, or vice versa. The compile-time and run-time
+/// information should not contradict. In other words, \a cRows should equal \a CRows unless
+/// \a CRows is \a Dynamic, and the same for the number of columns.
+///
+/// Example: \include MatrixBase_template_int_int_topLeftCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_topLeftCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block
+///
+template<int CRows, int CCols>
+inline typename FixedBlockXpr<CRows,CCols>::Type topLeftCorner(Index cRows, Index cCols)
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), 0, 0, cRows, cCols);
+}
+
+/// This is the const version of topLeftCorner<int, int>(Index, Index).
+template<int CRows, int CCols>
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type topLeftCorner(Index cRows, Index cCols) const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), 0, 0, cRows, cCols);
+}
+
+
+
+/// \returns a dynamic-size expression of a bottom-right corner of *this.
+///
+/// \param cRows the number of rows in the corner
+/// \param cCols the number of columns in the corner
+///
+/// Example: \include MatrixBase_bottomRightCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_bottomRightCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline BlockXpr bottomRightCorner(Index cRows, Index cCols)
+{
+  return BlockXpr(derived(), rows() - cRows, cols() - cCols, cRows, cCols);
+}
+
+/// This is the const version of bottomRightCorner(Index, Index).
+EIGEN_DEVICE_FUNC
+inline const ConstBlockXpr bottomRightCorner(Index cRows, Index cCols) const
+{
+  return ConstBlockXpr(derived(), rows() - cRows, cols() - cCols, cRows, cCols);
+}
+
+/// \returns an expression of a fixed-size bottom-right corner of *this.
+///
+/// The template parameters CRows and CCols are the number of rows and columns in the corner.
+///
+/// Example: \include MatrixBase_template_int_int_bottomRightCorner.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_bottomRightCorner.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline typename FixedBlockXpr<CRows,CCols>::Type bottomRightCorner()
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), rows() - CRows, cols() - CCols);
+}
+
+/// This is the const version of bottomRightCorner<int, int>().
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type bottomRightCorner() const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), rows() - CRows, cols() - CCols);
+}
+
+/// \returns an expression of a bottom-right corner of *this.
+///
+/// \tparam CRows number of rows in corner as specified at compile-time
+/// \tparam CCols number of columns in corner as specified at compile-time
+/// \param  cRows number of rows in corner as specified at run-time
+/// \param  cCols number of columns in corner as specified at run-time
+///
+/// This function is mainly useful for corners where the number of rows is specified at compile-time
+/// and the number of columns is specified at run-time, or vice versa. The compile-time and run-time
+/// information should not contradict. In other words, \a cRows should equal \a CRows unless
+/// \a CRows is \a Dynamic, and the same for the number of columns.
+///
+/// Example: \include MatrixBase_template_int_int_bottomRightCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_bottomRightCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block
+///
+template<int CRows, int CCols>
+inline typename FixedBlockXpr<CRows,CCols>::Type bottomRightCorner(Index cRows, Index cCols)
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), rows() - cRows, cols() - cCols, cRows, cCols);
+}
+
+/// This is the const version of bottomRightCorner<int, int>(Index, Index).
+template<int CRows, int CCols>
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type bottomRightCorner(Index cRows, Index cCols) const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), rows() - cRows, cols() - cCols, cRows, cCols);
+}
+
+
+
+/// \returns a dynamic-size expression of a bottom-left corner of *this.
+///
+/// \param cRows the number of rows in the corner
+/// \param cCols the number of columns in the corner
+///
+/// Example: \include MatrixBase_bottomLeftCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_bottomLeftCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline BlockXpr bottomLeftCorner(Index cRows, Index cCols)
+{
+  return BlockXpr(derived(), rows() - cRows, 0, cRows, cCols);
+}
+
+/// This is the const version of bottomLeftCorner(Index, Index).
+EIGEN_DEVICE_FUNC
+inline const ConstBlockXpr bottomLeftCorner(Index cRows, Index cCols) const
+{
+  return ConstBlockXpr(derived(), rows() - cRows, 0, cRows, cCols);
+}
+
+/// \returns an expression of a fixed-size bottom-left corner of *this.
+///
+/// The template parameters CRows and CCols are the number of rows and columns in the corner.
+///
+/// Example: \include MatrixBase_template_int_int_bottomLeftCorner.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_bottomLeftCorner.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline typename FixedBlockXpr<CRows,CCols>::Type bottomLeftCorner()
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), rows() - CRows, 0);
+}
+
+/// This is the const version of bottomLeftCorner<int, int>().
+template<int CRows, int CCols>
+EIGEN_DEVICE_FUNC
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type bottomLeftCorner() const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), rows() - CRows, 0);
+}
+
+/// \returns an expression of a bottom-left corner of *this.
+///
+/// \tparam CRows number of rows in corner as specified at compile-time
+/// \tparam CCols number of columns in corner as specified at compile-time
+/// \param  cRows number of rows in corner as specified at run-time
+/// \param  cCols number of columns in corner as specified at run-time
+///
+/// This function is mainly useful for corners where the number of rows is specified at compile-time
+/// and the number of columns is specified at run-time, or vice versa. The compile-time and run-time
+/// information should not contradict. In other words, \a cRows should equal \a CRows unless
+/// \a CRows is \a Dynamic, and the same for the number of columns.
+///
+/// Example: \include MatrixBase_template_int_int_bottomLeftCorner_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_bottomLeftCorner_int_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block
+///
+template<int CRows, int CCols>
+inline typename FixedBlockXpr<CRows,CCols>::Type bottomLeftCorner(Index cRows, Index cCols)
+{
+  return typename FixedBlockXpr<CRows,CCols>::Type(derived(), rows() - cRows, 0, cRows, cCols);
+}
+
+/// This is the const version of bottomLeftCorner<int, int>(Index, Index).
+template<int CRows, int CCols>
+inline const typename ConstFixedBlockXpr<CRows,CCols>::Type bottomLeftCorner(Index cRows, Index cCols) const
+{
+  return typename ConstFixedBlockXpr<CRows,CCols>::Type(derived(), rows() - cRows, 0, cRows, cCols);
+}
+
+
+
+/// \returns a block consisting of the top rows of *this.
+///
+/// \param n the number of rows in the block
+///
+/// Example: \include MatrixBase_topRows_int.cpp
+/// Output: \verbinclude MatrixBase_topRows_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline RowsBlockXpr topRows(Index n)
+{
+  return RowsBlockXpr(derived(), 0, 0, n, cols());
+}
+
+/// This is the const version of topRows(Index).
+EIGEN_DEVICE_FUNC
+inline ConstRowsBlockXpr topRows(Index n) const
+{
+  return ConstRowsBlockXpr(derived(), 0, 0, n, cols());
+}
+
+/// \returns a block consisting of the top rows of *this.
+///
+/// \tparam N the number of rows in the block as specified at compile-time
+/// \param n the number of rows in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_topRows.cpp
+/// Output: \verbinclude MatrixBase_template_int_topRows.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NRowsBlockXpr<N>::Type topRows(Index n = N)
+{
+  return typename NRowsBlockXpr<N>::Type(derived(), 0, 0, n, cols());
+}
+
+/// This is the const version of topRows<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNRowsBlockXpr<N>::Type topRows(Index n = N) const
+{
+  return typename ConstNRowsBlockXpr<N>::Type(derived(), 0, 0, n, cols());
+}
+
+
+
+/// \returns a block consisting of the bottom rows of *this.
+///
+/// \param n the number of rows in the block
+///
+/// Example: \include MatrixBase_bottomRows_int.cpp
+/// Output: \verbinclude MatrixBase_bottomRows_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline RowsBlockXpr bottomRows(Index n)
+{
+  return RowsBlockXpr(derived(), rows() - n, 0, n, cols());
+}
+
+/// This is the const version of bottomRows(Index).
+EIGEN_DEVICE_FUNC
+inline ConstRowsBlockXpr bottomRows(Index n) const
+{
+  return ConstRowsBlockXpr(derived(), rows() - n, 0, n, cols());
+}
+
+/// \returns a block consisting of the bottom rows of *this.
+///
+/// \tparam N the number of rows in the block as specified at compile-time
+/// \param n the number of rows in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_bottomRows.cpp
+/// Output: \verbinclude MatrixBase_template_int_bottomRows.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NRowsBlockXpr<N>::Type bottomRows(Index n = N)
+{
+  return typename NRowsBlockXpr<N>::Type(derived(), rows() - n, 0, n, cols());
+}
+
+/// This is the const version of bottomRows<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNRowsBlockXpr<N>::Type bottomRows(Index n = N) const
+{
+  return typename ConstNRowsBlockXpr<N>::Type(derived(), rows() - n, 0, n, cols());
+}
+
+
+
+/// \returns a block consisting of a range of rows of *this.
+///
+/// \param startRow the index of the first row in the block
+/// \param n the number of rows in the block
+///
+/// Example: \include DenseBase_middleRows_int.cpp
+/// Output: \verbinclude DenseBase_middleRows_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline RowsBlockXpr middleRows(Index startRow, Index n)
+{
+  return RowsBlockXpr(derived(), startRow, 0, n, cols());
+}
+
+/// This is the const version of middleRows(Index,Index).
+EIGEN_DEVICE_FUNC
+inline ConstRowsBlockXpr middleRows(Index startRow, Index n) const
+{
+  return ConstRowsBlockXpr(derived(), startRow, 0, n, cols());
+}
+
+/// \returns a block consisting of a range of rows of *this.
+///
+/// \tparam N the number of rows in the block as specified at compile-time
+/// \param startRow the index of the first row in the block
+/// \param n the number of rows in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include DenseBase_template_int_middleRows.cpp
+/// Output: \verbinclude DenseBase_template_int_middleRows.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NRowsBlockXpr<N>::Type middleRows(Index startRow, Index n = N)
+{
+  return typename NRowsBlockXpr<N>::Type(derived(), startRow, 0, n, cols());
+}
+
+/// This is the const version of middleRows<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNRowsBlockXpr<N>::Type middleRows(Index startRow, Index n = N) const
+{
+  return typename ConstNRowsBlockXpr<N>::Type(derived(), startRow, 0, n, cols());
+}
+
+
+
+/// \returns a block consisting of the left columns of *this.
+///
+/// \param n the number of columns in the block
+///
+/// Example: \include MatrixBase_leftCols_int.cpp
+/// Output: \verbinclude MatrixBase_leftCols_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline ColsBlockXpr leftCols(Index n)
+{
+  return ColsBlockXpr(derived(), 0, 0, rows(), n);
+}
+
+/// This is the const version of leftCols(Index).
+EIGEN_DEVICE_FUNC
+inline ConstColsBlockXpr leftCols(Index n) const
+{
+  return ConstColsBlockXpr(derived(), 0, 0, rows(), n);
+}
+
+/// \returns a block consisting of the left columns of *this.
+///
+/// \tparam N the number of columns in the block as specified at compile-time
+/// \param n the number of columns in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_leftCols.cpp
+/// Output: \verbinclude MatrixBase_template_int_leftCols.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NColsBlockXpr<N>::Type leftCols(Index n = N)
+{
+  return typename NColsBlockXpr<N>::Type(derived(), 0, 0, rows(), n);
+}
+
+/// This is the const version of leftCols<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNColsBlockXpr<N>::Type leftCols(Index n = N) const
+{
+  return typename ConstNColsBlockXpr<N>::Type(derived(), 0, 0, rows(), n);
+}
+
+
+
+/// \returns a block consisting of the right columns of *this.
+///
+/// \param n the number of columns in the block
+///
+/// Example: \include MatrixBase_rightCols_int.cpp
+/// Output: \verbinclude MatrixBase_rightCols_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline ColsBlockXpr rightCols(Index n)
+{
+  return ColsBlockXpr(derived(), 0, cols() - n, rows(), n);
+}
+
+/// This is the const version of rightCols(Index).
+EIGEN_DEVICE_FUNC
+inline ConstColsBlockXpr rightCols(Index n) const
+{
+  return ConstColsBlockXpr(derived(), 0, cols() - n, rows(), n);
+}
+
+/// \returns a block consisting of the right columns of *this.
+///
+/// \tparam N the number of columns in the block as specified at compile-time
+/// \param n the number of columns in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_rightCols.cpp
+/// Output: \verbinclude MatrixBase_template_int_rightCols.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NColsBlockXpr<N>::Type rightCols(Index n = N)
+{
+  return typename NColsBlockXpr<N>::Type(derived(), 0, cols() - n, rows(), n);
+}
+
+/// This is the const version of rightCols<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNColsBlockXpr<N>::Type rightCols(Index n = N) const
+{
+  return typename ConstNColsBlockXpr<N>::Type(derived(), 0, cols() - n, rows(), n);
+}
+
+
+
+/// \returns a block consisting of a range of columns of *this.
+///
+/// \param startCol the index of the first column in the block
+/// \param numCols the number of columns in the block
+///
+/// Example: \include DenseBase_middleCols_int.cpp
+/// Output: \verbinclude DenseBase_middleCols_int.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline ColsBlockXpr middleCols(Index startCol, Index numCols)
+{
+  return ColsBlockXpr(derived(), 0, startCol, rows(), numCols);
+}
+
+/// This is the const version of middleCols(Index,Index).
+EIGEN_DEVICE_FUNC
+inline ConstColsBlockXpr middleCols(Index startCol, Index numCols) const
+{
+  return ConstColsBlockXpr(derived(), 0, startCol, rows(), numCols);
+}
+
+/// \returns a block consisting of a range of columns of *this.
+///
+/// \tparam N the number of columns in the block as specified at compile-time
+/// \param startCol the index of the first column in the block
+/// \param n the number of columns in the block as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include DenseBase_template_int_middleCols.cpp
+/// Output: \verbinclude DenseBase_template_int_middleCols.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename NColsBlockXpr<N>::Type middleCols(Index startCol, Index n = N)
+{
+  return typename NColsBlockXpr<N>::Type(derived(), 0, startCol, rows(), n);
+}
+
+/// This is the const version of middleCols<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstNColsBlockXpr<N>::Type middleCols(Index startCol, Index n = N) const
+{
+  return typename ConstNColsBlockXpr<N>::Type(derived(), 0, startCol, rows(), n);
+}
+
+
+
+/// \returns a fixed-size expression of a block in *this.
+///
+/// The template parameters \a NRows and \a NCols are the number of
+/// rows and columns in the block.
+///
+/// \param startRow the first row in the block
+/// \param startCol the first column in the block
+///
+/// Example: \include MatrixBase_block_int_int.cpp
+/// Output: \verbinclude MatrixBase_block_int_int.out
+///
+/// \note since block is a templated member, the keyword template has to be used
+/// if the matrix type is also a template parameter: \code m.template block<3,3>(1,1); \endcode
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int NRows, int NCols>
+EIGEN_DEVICE_FUNC
+inline typename FixedBlockXpr<NRows,NCols>::Type block(Index startRow, Index startCol)
+{
+  return typename FixedBlockXpr<NRows,NCols>::Type(derived(), startRow, startCol);
+}
+
+/// This is the const version of block<>(Index, Index). */
+template<int NRows, int NCols>
+EIGEN_DEVICE_FUNC
+inline const typename ConstFixedBlockXpr<NRows,NCols>::Type block(Index startRow, Index startCol) const
+{
+  return typename ConstFixedBlockXpr<NRows,NCols>::Type(derived(), startRow, startCol);
+}
+
+/// \returns an expression of a block in *this.
+///
+/// \tparam NRows number of rows in block as specified at compile-time
+/// \tparam NCols number of columns in block as specified at compile-time
+/// \param  startRow  the first row in the block
+/// \param  startCol  the first column in the block
+/// \param  blockRows number of rows in block as specified at run-time
+/// \param  blockCols number of columns in block as specified at run-time
+///
+/// This function is mainly useful for blocks where the number of rows is specified at compile-time
+/// and the number of columns is specified at run-time, or vice versa. The compile-time and run-time
+/// information should not contradict. In other words, \a blockRows should equal \a NRows unless
+/// \a NRows is \a Dynamic, and the same for the number of columns.
+///
+/// Example: \include MatrixBase_template_int_int_block_int_int_int_int.cpp
+/// Output: \verbinclude MatrixBase_template_int_int_block_int_int_int_int.cpp
+///
+EIGEN_DOC_BLOCK_ADDONS_NOT_INNER_PANEL
+///
+/// \sa class Block, block(Index,Index,Index,Index)
+///
+template<int NRows, int NCols>
+inline typename FixedBlockXpr<NRows,NCols>::Type block(Index startRow, Index startCol,
+                                                  Index blockRows, Index blockCols)
+{
+  return typename FixedBlockXpr<NRows,NCols>::Type(derived(), startRow, startCol, blockRows, blockCols);
+}
+
+/// This is the const version of block<>(Index, Index, Index, Index).
+template<int NRows, int NCols>
+inline const typename ConstFixedBlockXpr<NRows,NCols>::Type block(Index startRow, Index startCol,
+                                                              Index blockRows, Index blockCols) const
+{
+  return typename ConstFixedBlockXpr<NRows,NCols>::Type(derived(), startRow, startCol, blockRows, blockCols);
+}
+
+/// \returns an expression of the \a i-th column of *this. Note that the numbering starts at 0.
+///
+/// Example: \include MatrixBase_col.cpp
+/// Output: \verbinclude MatrixBase_col.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(column-major)
+/**
+  * \sa row(), class Block */
+EIGEN_DEVICE_FUNC
+inline ColXpr col(Index i)
+{
+  return ColXpr(derived(), i);
+}
+
+/// This is the const version of col().
+EIGEN_DEVICE_FUNC
+inline ConstColXpr col(Index i) const
+{
+  return ConstColXpr(derived(), i);
+}
+
+/// \returns an expression of the \a i-th row of *this. Note that the numbering starts at 0.
+///
+/// Example: \include MatrixBase_row.cpp
+/// Output: \verbinclude MatrixBase_row.out
+///
+EIGEN_DOC_BLOCK_ADDONS_INNER_PANEL_IF(row-major)
+/**
+  * \sa col(), class Block */
+EIGEN_DEVICE_FUNC
+inline RowXpr row(Index i)
+{
+  return RowXpr(derived(), i);
+}
+
+/// This is the const version of row(). */
+EIGEN_DEVICE_FUNC
+inline ConstRowXpr row(Index i) const
+{
+  return ConstRowXpr(derived(), i);
+}
+
+/// \returns a dynamic-size expression of a segment (i.e. a vector block) in *this.
+///
+/// \only_for_vectors
+///
+/// \param start the first coefficient in the segment
+/// \param n the number of coefficients in the segment
+///
+/// Example: \include MatrixBase_segment_int_int.cpp
+/// Output: \verbinclude MatrixBase_segment_int_int.out
+///
+/// \note Even though the returned expression has dynamic size, in the case
+/// when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+/// which means that evaluating it does not cause a dynamic memory allocation.
+///
+/// \sa class Block, segment(Index)
+///
+EIGEN_DEVICE_FUNC
+inline SegmentReturnType segment(Index start, Index n)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), start, n);
+}
+
+
+/// This is the const version of segment(Index,Index).
+EIGEN_DEVICE_FUNC
+inline ConstSegmentReturnType segment(Index start, Index n) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), start, n);
+}
+
+/// \returns a dynamic-size expression of the first coefficients of *this.
+///
+/// \only_for_vectors
+///
+/// \param n the number of coefficients in the segment
+///
+/// Example: \include MatrixBase_start_int.cpp
+/// Output: \verbinclude MatrixBase_start_int.out
+///
+/// \note Even though the returned expression has dynamic size, in the case
+/// when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+/// which means that evaluating it does not cause a dynamic memory allocation.
+///
+/// \sa class Block, block(Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline SegmentReturnType head(Index n)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), 0, n);
+}
+
+/// This is the const version of head(Index).
+EIGEN_DEVICE_FUNC
+inline ConstSegmentReturnType head(Index n) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), 0, n);
+}
+
+/// \returns a dynamic-size expression of the last coefficients of *this.
+///
+/// \only_for_vectors
+///
+/// \param n the number of coefficients in the segment
+///
+/// Example: \include MatrixBase_end_int.cpp
+/// Output: \verbinclude MatrixBase_end_int.out
+///
+/// \note Even though the returned expression has dynamic size, in the case
+/// when it is applied to a fixed-size vector, it inherits a fixed maximal size,
+/// which means that evaluating it does not cause a dynamic memory allocation.
+///
+/// \sa class Block, block(Index,Index)
+///
+EIGEN_DEVICE_FUNC
+inline SegmentReturnType tail(Index n)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return SegmentReturnType(derived(), this->size() - n, n);
+}
+
+/// This is the const version of tail(Index).
+EIGEN_DEVICE_FUNC
+inline ConstSegmentReturnType tail(Index n) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return ConstSegmentReturnType(derived(), this->size() - n, n);
+}
+
+/// \returns a fixed-size expression of a segment (i.e. a vector block) in \c *this
+///
+/// \only_for_vectors
+///
+/// \tparam N the number of coefficients in the segment as specified at compile-time
+/// \param start the index of the first element in the segment
+/// \param n the number of coefficients in the segment as specified at compile-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_segment.cpp
+/// Output: \verbinclude MatrixBase_template_int_segment.out
+///
+/// \sa class Block
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename FixedSegmentReturnType<N>::Type segment(Index start, Index n = N)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<N>::Type(derived(), start, n);
+}
+
+/// This is the const version of segment<int>(Index).
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstFixedSegmentReturnType<N>::Type segment(Index start, Index n = N) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<N>::Type(derived(), start, n);
+}
+
+/// \returns a fixed-size expression of the first coefficients of *this.
+///
+/// \only_for_vectors
+///
+/// \tparam N the number of coefficients in the segment as specified at compile-time
+/// \param  n the number of coefficients in the segment as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_start.cpp
+/// Output: \verbinclude MatrixBase_template_int_start.out
+///
+/// \sa class Block
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename FixedSegmentReturnType<N>::Type head(Index n = N)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<N>::Type(derived(), 0, n);
+}
+
+/// This is the const version of head<int>().
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstFixedSegmentReturnType<N>::Type head(Index n = N) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<N>::Type(derived(), 0, n);
+}
+
+/// \returns a fixed-size expression of the last coefficients of *this.
+///
+/// \only_for_vectors
+///
+/// \tparam N the number of coefficients in the segment as specified at compile-time
+/// \param  n the number of coefficients in the segment as specified at run-time
+///
+/// The compile-time and run-time information should not contradict. In other words,
+/// \a n should equal \a N unless \a N is \a Dynamic.
+///
+/// Example: \include MatrixBase_template_int_end.cpp
+/// Output: \verbinclude MatrixBase_template_int_end.out
+///
+/// \sa class Block
+///
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename FixedSegmentReturnType<N>::Type tail(Index n = N)
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename FixedSegmentReturnType<N>::Type(derived(), size() - n);
+}
+
+/// This is the const version of tail<int>.
+template<int N>
+EIGEN_DEVICE_FUNC
+inline typename ConstFixedSegmentReturnType<N>::Type tail(Index n = N) const
+{
+  EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)
+  return typename ConstFixedSegmentReturnType<N>::Type(derived(), size() - n);
+}
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseBinaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseBinaryOps.h
new file mode 100644
index 0000000..8b6730e
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseBinaryOps.h
@@ -0,0 +1,115 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2016 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// This file is a base class plugin containing common coefficient wise functions.
+
+/** \returns an expression of the difference of \c *this and \a other
+  *
+  * \note If you want to substract a given scalar from all coefficients, see Cwise::operator-().
+  *
+  * \sa class CwiseBinaryOp, operator-=()
+  */
+EIGEN_MAKE_CWISE_BINARY_OP(operator-,difference)
+
+/** \returns an expression of the sum of \c *this and \a other
+  *
+  * \note If you want to add a given scalar to all coefficients, see Cwise::operator+().
+  *
+  * \sa class CwiseBinaryOp, operator+=()
+  */
+EIGEN_MAKE_CWISE_BINARY_OP(operator+,sum)
+
+/** \returns an expression of a custom coefficient-wise operator \a func of *this and \a other
+  *
+  * The template parameter \a CustomBinaryOp is the type of the functor
+  * of the custom operator (see class CwiseBinaryOp for an example)
+  *
+  * Here is an example illustrating the use of custom functors:
+  * \include class_CwiseBinaryOp.cpp
+  * Output: \verbinclude class_CwiseBinaryOp.out
+  *
+  * \sa class CwiseBinaryOp, operator+(), operator-(), cwiseProduct()
+  */
+template<typename CustomBinaryOp, typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<CustomBinaryOp, const Derived, const OtherDerived>
+binaryExpr(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other, const CustomBinaryOp& func = CustomBinaryOp()) const
+{
+  return CwiseBinaryOp<CustomBinaryOp, const Derived, const OtherDerived>(derived(), other.derived(), func);
+}
+
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_MAKE_SCALAR_BINARY_OP(operator*,product)
+#else
+/** \returns an expression of \c *this scaled by the scalar factor \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  */
+template<typename T>
+const CwiseBinaryOp<internal::scalar_product_op<Scalar,T>,Derived,Constant<T> > operator*(const T& scalar) const;
+/** \returns an expression of \a expr scaled by the scalar factor \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  */
+template<typename T> friend
+const CwiseBinaryOp<internal::scalar_product_op<T,Scalar>,Constant<T>,Derived> operator*(const T& scalar, const StorageBaseType& expr);
+#endif
+
+
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+EIGEN_MAKE_SCALAR_BINARY_OP_ONTHERIGHT(operator/,quotient)
+#else
+/** \returns an expression of \c *this divided by the scalar value \a scalar
+  *
+  * \tparam T is the scalar type of \a scalar. It must be compatible with the scalar type of the given expression.
+  */
+template<typename T>
+const CwiseBinaryOp<internal::scalar_quotient_op<Scalar,T>,Derived,Constant<T> > operator/(const T& scalar) const;
+#endif
+
+/** \returns an expression of the coefficient-wise boolean \b and operator of \c *this and \a other
+  *
+  * \warning this operator is for expression of bool only.
+  *
+  * Example: \include Cwise_boolean_and.cpp
+  * Output: \verbinclude Cwise_boolean_and.out
+  *
+  * \sa operator||(), select()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+inline const CwiseBinaryOp<internal::scalar_boolean_and_op, const Derived, const OtherDerived>
+operator&&(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<bool,Scalar>::value && internal::is_same<bool,typename OtherDerived::Scalar>::value),
+                      THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL);
+  return CwiseBinaryOp<internal::scalar_boolean_and_op, const Derived, const OtherDerived>(derived(),other.derived());
+}
+
+/** \returns an expression of the coefficient-wise boolean \b or operator of \c *this and \a other
+  *
+  * \warning this operator is for expression of bool only.
+  *
+  * Example: \include Cwise_boolean_or.cpp
+  * Output: \verbinclude Cwise_boolean_or.out
+  *
+  * \sa operator&&(), select()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+inline const CwiseBinaryOp<internal::scalar_boolean_or_op, const Derived, const OtherDerived>
+operator||(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  EIGEN_STATIC_ASSERT((internal::is_same<bool,Scalar>::value && internal::is_same<bool,typename OtherDerived::Scalar>::value),
+                      THIS_METHOD_IS_ONLY_FOR_EXPRESSIONS_OF_BOOL);
+  return CwiseBinaryOp<internal::scalar_boolean_or_op, const Derived, const OtherDerived>(derived(),other.derived());
+}
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseUnaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseUnaryOps.h
new file mode 100644
index 0000000..89f4faa
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/CommonCwiseUnaryOps.h
@@ -0,0 +1,163 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// This file is a base class plugin containing common coefficient wise functions.
+
+#ifndef EIGEN_PARSED_BY_DOXYGEN
+
+/** \internal the return type of conjugate() */
+typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                    const CwiseUnaryOp<internal::scalar_conjugate_op<Scalar>, const Derived>,
+                    const Derived&
+                  >::type ConjugateReturnType;
+/** \internal the return type of real() const */
+typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                    const CwiseUnaryOp<internal::scalar_real_op<Scalar>, const Derived>,
+                    const Derived&
+                  >::type RealReturnType;
+/** \internal the return type of real() */
+typedef typename internal::conditional<NumTraits<Scalar>::IsComplex,
+                    CwiseUnaryView<internal::scalar_real_ref_op<Scalar>, Derived>,
+                    Derived&
+                  >::type NonConstRealReturnType;
+/** \internal the return type of imag() const */
+typedef CwiseUnaryOp<internal::scalar_imag_op<Scalar>, const Derived> ImagReturnType;
+/** \internal the return type of imag() */
+typedef CwiseUnaryView<internal::scalar_imag_ref_op<Scalar>, Derived> NonConstImagReturnType;
+
+typedef CwiseUnaryOp<internal::scalar_opposite_op<Scalar>, const Derived> NegativeReturnType;
+
+#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+/// \returns an expression of the opposite of \c *this
+///
+EIGEN_DOC_UNARY_ADDONS(operator-,opposite)
+///
+EIGEN_DEVICE_FUNC
+inline const NegativeReturnType
+operator-() const { return NegativeReturnType(derived()); }
+
+
+template<class NewType> struct CastXpr { typedef typename internal::cast_return_type<Derived,const CwiseUnaryOp<internal::scalar_cast_op<Scalar, NewType>, const Derived> >::type Type; };
+
+/// \returns an expression of \c *this with the \a Scalar type casted to
+/// \a NewScalar.
+///
+/// The template parameter \a NewScalar is the type we are casting the scalars to.
+///
+EIGEN_DOC_UNARY_ADDONS(cast,conversion function)
+///
+/// \sa class CwiseUnaryOp
+///
+template<typename NewType>
+EIGEN_DEVICE_FUNC
+typename CastXpr<NewType>::Type
+cast() const
+{
+  return typename CastXpr<NewType>::Type(derived());
+}
+
+/// \returns an expression of the complex conjugate of \c *this.
+///
+EIGEN_DOC_UNARY_ADDONS(conjugate,complex conjugate)
+///
+/// \sa <a href="group__CoeffwiseMathFunctions.html#cwisetable_conj">Math functions</a>, MatrixBase::adjoint()
+EIGEN_DEVICE_FUNC
+inline ConjugateReturnType
+conjugate() const
+{
+  return ConjugateReturnType(derived());
+}
+
+/// \returns a read-only expression of the real part of \c *this.
+///
+EIGEN_DOC_UNARY_ADDONS(real,real part function)
+///
+/// \sa imag()
+EIGEN_DEVICE_FUNC
+inline RealReturnType
+real() const { return RealReturnType(derived()); }
+
+/// \returns an read-only expression of the imaginary part of \c *this.
+///
+EIGEN_DOC_UNARY_ADDONS(imag,imaginary part function)
+///
+/// \sa real()
+EIGEN_DEVICE_FUNC
+inline const ImagReturnType
+imag() const { return ImagReturnType(derived()); }
+
+/// \brief Apply a unary operator coefficient-wise
+/// \param[in]  func  Functor implementing the unary operator
+/// \tparam  CustomUnaryOp Type of \a func
+/// \returns An expression of a custom coefficient-wise unary operator \a func of *this
+///
+/// The function \c ptr_fun() from the C++ standard library can be used to make functors out of normal functions.
+///
+/// Example:
+/// \include class_CwiseUnaryOp_ptrfun.cpp
+/// Output: \verbinclude class_CwiseUnaryOp_ptrfun.out
+///
+/// Genuine functors allow for more possibilities, for instance it may contain a state.
+///
+/// Example:
+/// \include class_CwiseUnaryOp.cpp
+/// Output: \verbinclude class_CwiseUnaryOp.out
+///
+EIGEN_DOC_UNARY_ADDONS(unaryExpr,unary function)
+///
+/// \sa unaryViewExpr, binaryExpr, class CwiseUnaryOp
+///
+template<typename CustomUnaryOp>
+EIGEN_DEVICE_FUNC
+inline const CwiseUnaryOp<CustomUnaryOp, const Derived>
+unaryExpr(const CustomUnaryOp& func = CustomUnaryOp()) const
+{
+  return CwiseUnaryOp<CustomUnaryOp, const Derived>(derived(), func);
+}
+
+/// \returns an expression of a custom coefficient-wise unary operator \a func of *this
+///
+/// The template parameter \a CustomUnaryOp is the type of the functor
+/// of the custom unary operator.
+///
+/// Example:
+/// \include class_CwiseUnaryOp.cpp
+/// Output: \verbinclude class_CwiseUnaryOp.out
+///
+EIGEN_DOC_UNARY_ADDONS(unaryViewExpr,unary function)
+///
+/// \sa unaryExpr, binaryExpr class CwiseUnaryOp
+///
+template<typename CustomViewOp>
+EIGEN_DEVICE_FUNC
+inline const CwiseUnaryView<CustomViewOp, const Derived>
+unaryViewExpr(const CustomViewOp& func = CustomViewOp()) const
+{
+  return CwiseUnaryView<CustomViewOp, const Derived>(derived(), func);
+}
+
+/// \returns a non const expression of the real part of \c *this.
+///
+EIGEN_DOC_UNARY_ADDONS(real,real part function)
+///
+/// \sa imag()
+EIGEN_DEVICE_FUNC
+inline NonConstRealReturnType
+real() { return NonConstRealReturnType(derived()); }
+
+/// \returns a non const expression of the imaginary part of \c *this.
+///
+EIGEN_DOC_UNARY_ADDONS(imag,imaginary part function)
+///
+/// \sa real()
+EIGEN_DEVICE_FUNC
+inline NonConstImagReturnType
+imag() { return NonConstImagReturnType(derived()); }
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseBinaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseBinaryOps.h
new file mode 100644
index 0000000..f1084ab
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseBinaryOps.h
@@ -0,0 +1,152 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// This file is a base class plugin containing matrix specifics coefficient wise functions.
+
+/** \returns an expression of the Schur product (coefficient wise product) of *this and \a other
+  *
+  * Example: \include MatrixBase_cwiseProduct.cpp
+  * Output: \verbinclude MatrixBase_cwiseProduct.out
+  *
+  * \sa class CwiseBinaryOp, cwiseAbs2
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,product)
+cwiseProduct(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return EIGEN_CWISE_BINARY_RETURN_TYPE(Derived,OtherDerived,product)(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise == operator of *this and \a other
+  *
+  * \warning this performs an exact comparison, which is generally a bad idea with floating-point types.
+  * In order to check for equality between two vectors or matrices with floating-point coefficients, it is
+  * generally a far better idea to use a fuzzy comparison as provided by isApprox() and
+  * isMuchSmallerThan().
+  *
+  * Example: \include MatrixBase_cwiseEqual.cpp
+  * Output: \verbinclude MatrixBase_cwiseEqual.out
+  *
+  * \sa cwiseNotEqual(), isApprox(), isMuchSmallerThan()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+inline const CwiseBinaryOp<std::equal_to<Scalar>, const Derived, const OtherDerived>
+cwiseEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<std::equal_to<Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise != operator of *this and \a other
+  *
+  * \warning this performs an exact comparison, which is generally a bad idea with floating-point types.
+  * In order to check for equality between two vectors or matrices with floating-point coefficients, it is
+  * generally a far better idea to use a fuzzy comparison as provided by isApprox() and
+  * isMuchSmallerThan().
+  *
+  * Example: \include MatrixBase_cwiseNotEqual.cpp
+  * Output: \verbinclude MatrixBase_cwiseNotEqual.out
+  *
+  * \sa cwiseEqual(), isApprox(), isMuchSmallerThan()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+inline const CwiseBinaryOp<std::not_equal_to<Scalar>, const Derived, const OtherDerived>
+cwiseNotEqual(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<std::not_equal_to<Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise min of *this and \a other
+  *
+  * Example: \include MatrixBase_cwiseMin.cpp
+  * Output: \verbinclude MatrixBase_cwiseMin.out
+  *
+  * \sa class CwiseBinaryOp, max()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar>, const Derived, const OtherDerived>
+cwiseMin(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise min of *this and scalar \a other
+  *
+  * \sa class CwiseBinaryOp, min()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_min_op<Scalar,Scalar>, const Derived, const ConstantReturnType>
+cwiseMin(const Scalar &other) const
+{
+  return cwiseMin(Derived::Constant(rows(), cols(), other));
+}
+
+/** \returns an expression of the coefficient-wise max of *this and \a other
+  *
+  * Example: \include MatrixBase_cwiseMax.cpp
+  * Output: \verbinclude MatrixBase_cwiseMax.out
+  *
+  * \sa class CwiseBinaryOp, min()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar>, const Derived, const OtherDerived>
+cwiseMax(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+/** \returns an expression of the coefficient-wise max of *this and scalar \a other
+  *
+  * \sa class CwiseBinaryOp, min()
+  */
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_max_op<Scalar,Scalar>, const Derived, const ConstantReturnType>
+cwiseMax(const Scalar &other) const
+{
+  return cwiseMax(Derived::Constant(rows(), cols(), other));
+}
+
+
+/** \returns an expression of the coefficient-wise quotient of *this and \a other
+  *
+  * Example: \include MatrixBase_cwiseQuotient.cpp
+  * Output: \verbinclude MatrixBase_cwiseQuotient.out
+  *
+  * \sa class CwiseBinaryOp, cwiseProduct(), cwiseInverse()
+  */
+template<typename OtherDerived>
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseBinaryOp<internal::scalar_quotient_op<Scalar>, const Derived, const OtherDerived>
+cwiseQuotient(const EIGEN_CURRENT_STORAGE_BASE_CLASS<OtherDerived> &other) const
+{
+  return CwiseBinaryOp<internal::scalar_quotient_op<Scalar>, const Derived, const OtherDerived>(derived(), other.derived());
+}
+
+typedef CwiseBinaryOp<internal::scalar_cmp_op<Scalar,Scalar,internal::cmp_EQ>, const Derived, const ConstantReturnType> CwiseScalarEqualReturnType;
+
+/** \returns an expression of the coefficient-wise == operator of \c *this and a scalar \a s
+  *
+  * \warning this performs an exact comparison, which is generally a bad idea with floating-point types.
+  * In order to check for equality between two vectors or matrices with floating-point coefficients, it is
+  * generally a far better idea to use a fuzzy comparison as provided by isApprox() and
+  * isMuchSmallerThan().
+  *
+  * \sa cwiseEqual(const MatrixBase<OtherDerived> &) const
+  */
+EIGEN_DEVICE_FUNC
+inline const CwiseScalarEqualReturnType
+cwiseEqual(const Scalar& s) const
+{
+  return CwiseScalarEqualReturnType(derived(), Derived::Constant(rows(), cols(), s), internal::scalar_cmp_op<Scalar,Scalar,internal::cmp_EQ>());
+}
diff --git a/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseUnaryOps.h b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseUnaryOps.h
new file mode 100644
index 0000000..b1be3d5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/eigen/Eigen/src/plugins/MatrixCwiseUnaryOps.h
@@ -0,0 +1,85 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2008-2009 Gael Guennebaud <gael.guennebaud@inria.fr>
+// Copyright (C) 2006-2008 Benoit Jacob <jacob.benoit.1@gmail.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+// This file is included into the body of the base classes supporting matrix specific coefficient-wise functions.
+// This include MatrixBase and SparseMatrixBase.
+
+
+typedef CwiseUnaryOp<internal::scalar_abs_op<Scalar>, const Derived> CwiseAbsReturnType;
+typedef CwiseUnaryOp<internal::scalar_abs2_op<Scalar>, const Derived> CwiseAbs2ReturnType;
+typedef CwiseUnaryOp<internal::scalar_sqrt_op<Scalar>, const Derived> CwiseSqrtReturnType;
+typedef CwiseUnaryOp<internal::scalar_sign_op<Scalar>, const Derived> CwiseSignReturnType;
+typedef CwiseUnaryOp<internal::scalar_inverse_op<Scalar>, const Derived> CwiseInverseReturnType;
+
+/// \returns an expression of the coefficient-wise absolute value of \c *this
+///
+/// Example: \include MatrixBase_cwiseAbs.cpp
+/// Output: \verbinclude MatrixBase_cwiseAbs.out
+///
+EIGEN_DOC_UNARY_ADDONS(cwiseAbs,absolute value)
+///
+/// \sa cwiseAbs2()
+///
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseAbsReturnType
+cwiseAbs() const { return CwiseAbsReturnType(derived()); }
+
+/// \returns an expression of the coefficient-wise squared absolute value of \c *this
+///
+/// Example: \include MatrixBase_cwiseAbs2.cpp
+/// Output: \verbinclude MatrixBase_cwiseAbs2.out
+///
+EIGEN_DOC_UNARY_ADDONS(cwiseAbs2,squared absolute value)
+///
+/// \sa cwiseAbs()
+///
+EIGEN_DEVICE_FUNC
+EIGEN_STRONG_INLINE const CwiseAbs2ReturnType
+cwiseAbs2() const { return CwiseAbs2ReturnType(derived()); }
+
+/// \returns an expression of the coefficient-wise square root of *this.
+///
+/// Example: \include MatrixBase_cwiseSqrt.cpp
+/// Output: \verbinclude MatrixBase_cwiseSqrt.out
+///
+EIGEN_DOC_UNARY_ADDONS(cwiseSqrt,square-root)
+///
+/// \sa cwisePow(), cwiseSquare()
+///
+EIGEN_DEVICE_FUNC
+inline const CwiseSqrtReturnType
+cwiseSqrt() const { return CwiseSqrtReturnType(derived()); }
+
+/// \returns an expression of the coefficient-wise signum of *this.
+///
+/// Example: \include MatrixBase_cwiseSign.cpp
+/// Output: \verbinclude MatrixBase_cwiseSign.out
+///
+EIGEN_DOC_UNARY_ADDONS(cwiseSign,sign function)
+///
+EIGEN_DEVICE_FUNC
+inline const CwiseSignReturnType
+cwiseSign() const { return CwiseSignReturnType(derived()); }
+
+
+/// \returns an expression of the coefficient-wise inverse of *this.
+///
+/// Example: \include MatrixBase_cwiseInverse.cpp
+/// Output: \verbinclude MatrixBase_cwiseInverse.out
+///
+EIGEN_DOC_UNARY_ADDONS(cwiseInverse,inverse)
+///
+/// \sa cwiseProduct()
+///
+EIGEN_DEVICE_FUNC
+inline const CwiseInverseReturnType
+cwiseInverse() const { return CwiseInverseReturnType(derived()); }
+
+
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/CMakeLists.txt b/src/runtime/ref/nn/depend/external/gemmlowp/CMakeLists.txt
new file mode 100644
index 0000000..4e4f4b1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/CMakeLists.txt
@@ -0,0 +1,11 @@
+
+SET(CUR_INCS
+   ${CMAKE_CURRENT_SOURCE_DIR}/fixedpoint
+   ${CMAKE_CURRENT_SOURCE_DIR}/public
+)
+
+SET(INC_DIRS
+  ${INC_DIRS}
+  ${CUR_INCS}
+  PARENT_SCOPE
+)
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint.h b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint.h
new file mode 100644
index 0000000..e21337f
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint.h
@@ -0,0 +1,779 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// fixedpoint.h: fixed-point arithmetic, with basic operations and
+// a few math functions such as tanh.
+
+#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_H_
+#define GEMMLOWP_INTERNAL_FIXEDPOINT_H_
+
+#include <cassert>
+#include <limits>
+
+#include "../internal/common.h"
+
+namespace gemmlowp {
+
+// Part 1: Low-level integer-arithmetic primitives.
+// The implementations here are generic implementations valid for
+// scalar types (e.g. std::int32_t). Architecture-specific SIMD types
+// (e.g. NEON int32x4_t) may be supported by providing
+// specializations for them in separate files.
+//
+// The purpose of these primitives is two-fold:
+//  - They will be used to implement higher-level fixed-point
+//    abstractions, namely the FixedPoint class and its arithmetic
+//    operators.
+//  - They will be directly used to implement some more involved
+//    fixed-point computations, e.g. the fixed-point implementation
+//    of math functions such as tanh.
+
+// Some compile-time traits around raw types to handle SIMD aspects:
+// number of lanes, underlying scalar type.
+template <typename tIntegerType>
+struct FixedPointRawTypeTraits {};
+
+template <>
+struct FixedPointRawTypeTraits<std::int32_t> {
+  typedef std::int32_t ScalarRawType;
+  static const int kLanes = 1;
+};
+
+// Returns a SIMD value duplicating a scalar value across all lanes.
+template <typename tRawType>
+tRawType Dup(typename FixedPointRawTypeTraits<tRawType>::ScalarRawType x) {
+  return x;
+}
+
+// Plain bit-wise AND
+template <typename tIntegerType>
+tIntegerType BitAnd(tIntegerType a, tIntegerType b) {
+  return a & b;
+}
+
+// Plain bit-wise OR
+template <typename tIntegerType>
+tIntegerType BitOr(tIntegerType a, tIntegerType b) {
+  return a | b;
+}
+
+// Plain bit-wise XOR
+template <typename tIntegerType>
+tIntegerType BitXor(tIntegerType a, tIntegerType b) {
+  return a ^ b;
+}
+
+// Plain bit-wise NOT
+template <typename tIntegerType>
+tIntegerType BitNot(tIntegerType a) {
+  return ~a;
+}
+
+// Integer addition. Not saturating. Overflow is undefined behavior.
+template <typename tIntegerType>
+tIntegerType Add(tIntegerType a, tIntegerType b) {
+  return a + b;
+}
+
+// Integer subtraction. Not saturating. Overflow is undefined behavior.
+template <typename tIntegerType>
+tIntegerType Mul(tIntegerType a, tIntegerType b) {
+  return a * b;
+}
+
+template <typename tIntegerType>
+tIntegerType Sub(tIntegerType a, tIntegerType b) {
+  return a - b;
+}
+
+// Integer unary negative. Not saturating. Overflow is undefined behavior.
+template <typename tIntegerType>
+tIntegerType Neg(tIntegerType a) {
+  return -a;
+}
+
+// Integer arithmetic left-shift, equivalent to multiplying with a
+// power of two. Not saturating. Overflow is undefined behavior.
+template <typename tIntegerType>
+tIntegerType ShiftLeft(tIntegerType a, int offset) {
+  return a << offset;
+}
+
+// Integer arithmetic right-shift. Not rounding.
+// Relying on implementation-defined, but in-practice-consistent,
+// C++ compiler behavior.
+template <typename tIntegerType>
+tIntegerType ShiftRight(tIntegerType a, int offset) {
+  return a >> offset;
+}
+
+// Each bit of the result is set to the corresponding bit of either then_val or
+// else_val depending on whether the corresponding bit of if_mask is set.
+// Equivalent to the VBSL instruction in ARM NEON.
+template <typename tIntegerType>
+tIntegerType SelectUsingMask(tIntegerType if_mask, tIntegerType then_val,
+                             tIntegerType else_val) {
+  return BitXor(BitAnd(if_mask, then_val), BitAnd(BitNot(if_mask), else_val));
+}
+
+// For each input scalar, the corresponding bits of the result are set if the
+// input scalar is non-zero.
+template <typename tIntegerType>
+tIntegerType MaskIfNonZero(tIntegerType a) {
+  static const tIntegerType zero = 0;
+  return a ? BitNot(zero) : zero;
+}
+
+// For each input scalar, the corresponding bits of the result are set if the
+// input scalar is zero.
+template <typename tIntegerType>
+tIntegerType MaskIfZero(tIntegerType a) {
+  return MaskIfNonZero<tIntegerType>(!a);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars are equal.
+template <typename tIntegerType>
+tIntegerType MaskIfEqual(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a == b);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars are not equal.
+template <typename tIntegerType>
+tIntegerType MaskIfNotEqual(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a != b);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars a, b satisfy a > b.
+template <typename tIntegerType>
+tIntegerType MaskIfGreaterThan(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a > b);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars a, b satisfy a >= b.
+template <typename tIntegerType>
+tIntegerType MaskIfGreaterThanOrEqual(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a >= b);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars a, b satisfy a < b.
+template <typename tIntegerType>
+tIntegerType MaskIfLessThan(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a < b);
+}
+
+// For each pair of input scalars, the corresponding bits of the result are
+// set if the input scalars a, b satisfy a <= b.
+template <typename tIntegerType>
+tIntegerType MaskIfLessThanOrEqual(tIntegerType a, tIntegerType b) {
+  return MaskIfNonZero<tIntegerType>(a <= b);
+}
+
+// Returns true if all of the input scalars are nonzero.
+// This function may currently assume that each of the input scalars has either
+// all or none of its bits set. Otherwise, its behavior is currently undefined.
+template <typename tIntegerType>
+bool All(tIntegerType a) {
+  return a;
+}
+
+// Returns true if any of the input scalars are nonzero.
+// This function may currently assume that each of the input scalars has either
+// all or none of its bits set. Otherwise, its behavior is currently undefined.
+template <typename tIntegerType>
+bool Any(tIntegerType a) {
+  return a;
+}
+
+// Returns (a+b)/2, rounded to the nearest integer.
+// Equivalent to VRHADD in the ARM NEON instruction set.
+template <typename IntegerType>
+IntegerType RoundingHalfSum(IntegerType a, IntegerType b) {
+  static_assert(std::is_same<IntegerType, void>::value, "unimplemented");
+  return a;
+}
+
+template <>
+inline std::int32_t RoundingHalfSum(std::int32_t a, std::int32_t b) {
+  std::int64_t a64 = a;
+  std::int64_t b64 = b;
+  std::int64_t sum = a64 + b64;
+  std::int64_t sign = sum >= 0 ? 1 : -1;
+  return static_cast<std::int32_t>((sum + sign) / 2);
+}
+
+// Returns the integer that represents the product of two fixed-point
+// numbers, interpreting all integers as fixed-point values in the
+// interval [-1, 1), rounding to the nearest value, and saturating
+// -1 * -1 to the maximum value (since 1 is not in the half-open
+// interval [-1, 1)).
+//
+// [The explanation below specializes to std::int32_t for example purpose.]
+//
+// The mapping between IntegerType and the interval [-1, 1) is unique and
+// implied by IntegerType, which is assumed to be signed. For example,
+// for IntegerType==std::int32_t, the mapping is
+//   real_value = integer_value / 2^31.
+// So in this case, and leaving aside rounding and saturating, this
+// function computes ((a / 2^31) * (b / 2^31)) * 2^31, which simplifies to
+//   (a * b) / 2^31.
+//
+// The 'doubling' part in the name of this function comes from the fact that
+// this operation is very close to a "multiply-high" operation, keeping only
+// the top half bits, except that that would be effectively computing
+//   (a * b) / 2^32,
+// so here we are computing 2x that, since
+//   1/2^31 = 2 * 1/2^32.
+// The idea is to use all of the available 32 bits in the destination int32
+// value.
+//
+// [End of the explanation specializing to int32.]
+//
+// This is equivalent to the VQRDMULH instruction in ARM NEON.
+template <typename IntegerType>
+IntegerType SaturatingRoundingDoublingHighMul(IntegerType a, IntegerType b) {
+  static_assert(std::is_same<IntegerType, void>::value, "unimplemented");
+  return a;
+}
+
+// This function implements the same computation as the ARMv7 NEON VQRDMULH
+// instruction.
+template <>
+inline std::int32_t SaturatingRoundingDoublingHighMul(std::int32_t a,
+                                                      std::int32_t b) {
+  bool overflow = a == b && a == std::numeric_limits<std::int32_t>::min();
+  std::int64_t a_64(a);
+  std::int64_t b_64(b);
+  std::int64_t ab_64 = a_64 * b_64;
+  std::int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30));
+  std::int32_t ab_x2_high32 =
+      static_cast<std::int32_t>((ab_64 + nudge) / (1ll << 31));
+  return overflow ? std::numeric_limits<std::int32_t>::max() : ab_x2_high32;
+}
+
+// Correctly-rounded-to-nearest division by a power-of-two.
+// Also known as a rounding arithmetic right shift.
+template <typename IntegerType>
+inline IntegerType RoundingDivideByPOT(IntegerType x, int exponent) {
+  using ScalarIntegerType =
+      typename FixedPointRawTypeTraits<IntegerType>::ScalarRawType;
+  static_assert(std::is_same<ScalarIntegerType, std::int32_t>::value,
+                "Currently only supporting int32 scalar and SIMD types");
+  assert(exponent >= 0);
+  assert(exponent <= 31);
+  const IntegerType mask = Dup<IntegerType>((1ll << exponent) - 1);
+  const IntegerType zero = Dup<IntegerType>(0);
+  const IntegerType one = Dup<IntegerType>(1);
+  const IntegerType remainder = BitAnd(x, mask);
+  const IntegerType threshold =
+      Add(ShiftRight(mask, 1), BitAnd(MaskIfLessThan(x, zero), one));
+  return Add(ShiftRight(x, exponent),
+             BitAnd(MaskIfGreaterThan(remainder, threshold), one));
+}
+
+// Returns the product of a run-time integer value by a compile-time power
+// of two, with either a positive exponent (equivalent to an arithmetic
+// left shift, saturating) or a negative exponent (equivalent to an arithmetic
+// right shift, rounding to nearest).
+template <int Exponent, typename IntegerType,
+          int ExponentSign = (Exponent > 0 ? 1 : Exponent < 0 ? -1 : 0)>
+struct ImplSaturatingRoundingMultiplyByPOT {};
+
+template <int Exponent, typename IntegerType>
+struct ImplSaturatingRoundingMultiplyByPOT<Exponent, IntegerType, 0> {
+  static IntegerType eval(IntegerType x) { return x; }
+};
+
+template <int Exponent, typename IntegerType>
+struct ImplSaturatingRoundingMultiplyByPOT<Exponent, IntegerType, 1> {
+  static IntegerType eval(IntegerType x) {
+    using ScalarIntegerType =
+        typename FixedPointRawTypeTraits<IntegerType>::ScalarRawType;
+    static_assert(std::is_same<ScalarIntegerType, std::int32_t>::value,
+                  "Currently only supporting int32 scalar and SIMD types");
+    const IntegerType min =
+        Dup<IntegerType>(std::numeric_limits<std::int32_t>::min());
+    const IntegerType max =
+        Dup<IntegerType>(std::numeric_limits<std::int32_t>::max());
+
+    const std::int32_t threshold = ((1 << (31 - Exponent)) - 1);
+    const IntegerType positive_mask =
+        MaskIfGreaterThan(x, Dup<IntegerType>(threshold));
+    const IntegerType negative_mask =
+        MaskIfLessThan(x, Dup<IntegerType>(-threshold));
+
+    IntegerType result = ShiftLeft(x, Exponent);
+    result = SelectUsingMask(positive_mask, max, result);
+    result = SelectUsingMask(negative_mask, min, result);
+    return result;
+  }
+};
+
+template <int Exponent, typename IntegerType>
+struct ImplSaturatingRoundingMultiplyByPOT<Exponent, IntegerType, -1> {
+  static IntegerType eval(IntegerType x) {
+    return RoundingDivideByPOT<IntegerType>(x, -Exponent);
+  }
+};
+
+template <int Exponent, typename IntegerType>
+IntegerType SaturatingRoundingMultiplyByPOT(IntegerType x) {
+  return ImplSaturatingRoundingMultiplyByPOT<Exponent, IntegerType>::eval(x);
+}
+
+// Part 2: the FixedPoint class.
+
+// A FixedPoint object represents a fixed-point value stored in the underlying
+// integer type tRawType, if tRawType is a plain scalar integer type.
+// Alternatively, tRawType may be a SIMD type (e.g. NEON int32x4_t) in which
+// case a FixedPoint object represents a corresponding SIMD vector of fixed
+// point values.
+//
+// tIntegerBits describes the range of the fixed-point format: if
+// tIntegerBits == m then the range of representable values is the half-open
+// interval [-2^m; 2^m) where the open boundary on the right side means that
+// 2^m is not representable (how close the maximum representable value is to
+// it, depends on bit-depth of tRawType).
+//
+// In "Q format notation",
+//   https://en.wikipedia.org/wiki/Q_(number_format)
+// we are describing the format
+//   Qm.n
+// where
+//   m = tIntegerBits
+// and
+//   n = NumberOfBits(tRawType) - (m + 1)
+// Note that the (m + 1) in the above line is because we adopt the convention
+// that we count the integer bits exclusively of the sign bit; so (m + 1) is
+// the total number of integer bits inclusive of the sign bit.
+//
+// Accordingly, the number of integral representable values in our range
+//   [-2^m ; 2^m)
+// is equal to 2^(m+1).
+template <typename tRawType, int tIntegerBits>
+class FixedPoint {
+ public:
+  typedef tRawType RawType;
+
+  typedef FixedPointRawTypeTraits<RawType> RawTypeTraits;
+  typedef typename RawTypeTraits::ScalarRawType ScalarRawType;
+
+  static const int kTotalBits = 8 * sizeof(ScalarRawType);
+  static const int kIntegerBits = tIntegerBits;
+  static const int kFractionalBits = kTotalBits - 1 - kIntegerBits;
+  static_assert(kIntegerBits >= 0 && kIntegerBits < kTotalBits,
+                "bad IntegerBits");
+
+  typedef FixedPoint<ScalarRawType, kIntegerBits> ScalarFixedPointType;
+
+  static const ScalarRawType ScalarRawMin() {
+    return std::numeric_limits<ScalarRawType>::min();
+  }
+
+  static const ScalarRawType ScalarRawMax() {
+    return std::numeric_limits<ScalarRawType>::max();
+  }
+
+  static const ScalarRawType RawMin() {
+    return VectorFromScalar(ScalarRawMin());
+  }
+
+  static const ScalarRawType RawMax() {
+    return VectorFromScalar(ScalarRawMax());
+  }
+
+  static FixedPoint FromRaw(RawType x) {
+    FixedPoint retval;
+    retval.raw() = x;
+    return retval;
+  }
+
+  static FixedPoint FromScalarRaw(ScalarRawType x) {
+    FixedPoint retval;
+    retval.raw() = Dup<RawType>(x);
+    return retval;
+  }
+
+  static FixedPoint FromScalarFixedPoint(ScalarFixedPointType x) {
+    return FromScalarRaw(x.raw());
+  }
+
+  template <int Exponent>
+  static FixedPoint ConstantPOT() {
+    static const int kOffset = kFractionalBits + Exponent;
+    static_assert(
+        kOffset < 31,
+        "Constant not exactly representable in this fixed-point format");
+    return FromScalarRaw(ScalarRawType(1) << kOffset);
+  }
+
+  static FixedPoint Zero() { return FromScalarRaw(0); }
+
+  static FixedPoint One() {
+    return FromScalarRaw(kIntegerBits == 0
+                             ? ScalarRawMax()
+                             : (ScalarRawType(1) << kFractionalBits));
+  }
+
+  static FixedPoint FromDouble(double x) {
+    const double min_bound = static_cast<double>(ScalarRawMin());
+    const double max_bound = static_cast<double>(ScalarRawMax());
+    return FromScalarRaw(static_cast<std::int32_t>(std::min(
+        std::max(round(x * static_cast<double>(1ll << kFractionalBits)),
+                 min_bound),
+        max_bound)));
+  }
+
+  RawType raw() const { return i_; }
+  RawType& raw() { return i_; }
+
+ private:
+  RawType i_;
+};
+
+// Part 3: implementation of arithmetic operators for the
+// FixedPoint class, and a few related functions.
+
+// A FixedPoint multiplication is just a
+// SaturatingRoundingDoublingHighMul operation on the underlying
+// raw integer values. The IntegerBits simply add up, as is obvious
+// from the fact that the range is [-2^IntegerBits, 2^IntegerBits).
+template <typename tRawType, int tIntegerBits_a, int tIntegerBits_b>
+FixedPoint<tRawType, tIntegerBits_a + tIntegerBits_b> operator*(
+    FixedPoint<tRawType, tIntegerBits_a> a,
+    FixedPoint<tRawType, tIntegerBits_b> b) {
+  FixedPoint<tRawType, tIntegerBits_a + tIntegerBits_b> c;
+  c.raw() = SaturatingRoundingDoublingHighMul(a.raw(), b.raw());
+  return c;
+}
+
+// Tweaking IntegerBits gives exact multiplication by a power of two.
+template <int tExponent, typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, tExponent + tIntegerBits> ExactMulByPot(
+    FixedPoint<tRawType, tIntegerBits> a) {
+  FixedPoint<tRawType, tExponent + tIntegerBits> c;
+  c.raw() = a.raw();
+  return c;
+}
+
+// If we want to leave IntegerBits fixed, then multiplication
+// by a power of two has to be saturating/rounding, not exact anymore.
+template <int tExponent, typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, tIntegerBits> SaturatingRoundingMultiplyByPOT(
+    FixedPoint<tRawType, tIntegerBits> a) {
+  return FixedPoint<tRawType, tIntegerBits>::FromRaw(
+      SaturatingRoundingMultiplyByPOT<tExponent>(a.raw()));
+}
+
+// Generic arithmetic operators.
+
+#define MAKE_FIXEDPOINT_UNARY_FUNC(FuncName, ImplFuncName)                     \
+  template <typename tRawType, int tIntegerBits>                               \
+  FixedPoint<tRawType, tIntegerBits> FuncName(                                 \
+      FixedPoint<tRawType, tIntegerBits> a) {                                  \
+    return FixedPoint<tRawType, tIntegerBits>::FromRaw(ImplFuncName(a.raw())); \
+  }
+
+#define MAKE_FIXEDPOINT_BINARY_FUNC(FuncName, ImplFuncName) \
+  template <typename tRawType, int tIntegerBits>            \
+  FixedPoint<tRawType, tIntegerBits> FuncName(              \
+      FixedPoint<tRawType, tIntegerBits> a,                 \
+      FixedPoint<tRawType, tIntegerBits> b) {               \
+    return FixedPoint<tRawType, tIntegerBits>::FromRaw(     \
+        ImplFuncName(a.raw(), b.raw()));                    \
+  }
+
+MAKE_FIXEDPOINT_UNARY_FUNC(operator-, Neg)
+MAKE_FIXEDPOINT_UNARY_FUNC(operator~, BitNot)
+MAKE_FIXEDPOINT_BINARY_FUNC(operator+, Add)
+MAKE_FIXEDPOINT_BINARY_FUNC(operator-, Sub)
+MAKE_FIXEDPOINT_BINARY_FUNC(operator&, BitAnd)
+MAKE_FIXEDPOINT_BINARY_FUNC(operator^, BitXor)
+MAKE_FIXEDPOINT_BINARY_FUNC(operator|, BitOr)
+MAKE_FIXEDPOINT_BINARY_FUNC(RoundingHalfSum, RoundingHalfSum)
+
+#undef MAKE_FIXEDPOINT_UNARY_FUNC
+#undef MAKE_FIXEDPOINT_BINARY_FUNC
+
+#define MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(FuncName)  \
+  template <typename tRawType, int tIntegerBits>            \
+  tRawType FuncName(FixedPoint<tRawType, tIntegerBits> a) { \
+    return FuncName(a.raw());                               \
+  }
+
+#define MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(FuncName) \
+  template <typename tRawType, int tIntegerBits>            \
+  tRawType FuncName(FixedPoint<tRawType, tIntegerBits> a,   \
+                    FixedPoint<tRawType, tIntegerBits> b) { \
+    return FuncName(a.raw(), b.raw());                      \
+  }
+
+MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(MaskIfZero)
+MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW(MaskIfNonZero)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfEqual)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfNotEqual)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfGreaterThan)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfGreaterThanOrEqual)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfLessThan)
+MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW(MaskIfLessThanOrEqual)
+
+#undef MAKE_FIXEDPOINT_UNARY_FUNC_RETURNING_RAW
+#undef MAKE_FIXEDPOINT_BINARY_FUNC_RETURNING_RAW
+
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, tIntegerBits> SelectUsingMask(
+    tRawType if_mask, FixedPoint<tRawType, tIntegerBits> then_val,
+    FixedPoint<tRawType, tIntegerBits> else_val) {
+  return FixedPoint<tRawType, tIntegerBits>::FromRaw(
+      SelectUsingMask(if_mask, then_val.raw(), else_val.raw()));
+}
+
+template <typename tRawType, int tIntegerBits>
+bool operator==(FixedPoint<tRawType, tIntegerBits> a,
+                FixedPoint<tRawType, tIntegerBits> b) {
+  return All(MaskIfEqual(a.raw(), b.raw()));
+}
+
+template <typename tRawType, int tIntegerBits>
+bool operator!=(FixedPoint<tRawType, tIntegerBits> a,
+                FixedPoint<tRawType, tIntegerBits> b) {
+  return !(a == b);
+}
+
+// Conversion to floating-point.
+template <typename tRawType, int tIntegerBits>
+double ToDouble(FixedPoint<tRawType, tIntegerBits> x) {
+  static_assert(FixedPointRawTypeTraits<tRawType>::kLanes == 1,
+                "not applicable to SIMD types");
+  typedef FixedPoint<tRawType, tIntegerBits> F;
+  return x.raw() / static_cast<double>(1ll << F::kFractionalBits);
+}
+
+// Rescale changes the number of IntegerBits and updates the underlying
+// raw integer value accordingly.
+template <int tIntegerBitsDst, typename tRawType, int tIntegerBitsSrc>
+FixedPoint<tRawType, tIntegerBitsDst> Rescale(
+    FixedPoint<tRawType, tIntegerBitsSrc> x) {
+  static const int kExponent = tIntegerBitsSrc - tIntegerBitsDst;
+  FixedPoint<tRawType, tIntegerBitsDst> result;
+  result.raw() = SaturatingRoundingMultiplyByPOT<kExponent>(x.raw());
+  return result;
+}
+
+// CheckedFixedPointConstant allows to specify fixed-point constants
+// initialized as real numbers, in a way that does not compile floating-point
+// arithmetic in production code, yet still checks agreement with the
+// floating-point expressions when asserts are enabled.
+#ifdef GEMMLOWP_ENABLE_FIXEDPOINT_CONSTANTS_CHECKS
+template <typename FixedPointType>
+FixedPointType CheckedFixedPointConstant(
+    typename FixedPointType::ScalarRawType raw_value, double double_value) {
+  typedef typename FixedPointType::RawType RawType;
+  const FixedPointType result = FixedPointType::FromScalarRaw(raw_value);
+  assert(result == FixedPointType::FromDouble(double_value));
+  return result;
+}
+#define GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(FixedPointType, ScalarRawValue, \
+                                             DoubleValue)                    \
+  (CheckedFixedPointConstant<FixedPointType>(ScalarRawValue, DoubleValue))
+
+#else
+#define GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(FixedPointType, ScalarRawValue, \
+                                             DoubleValue)                    \
+  (FixedPointType::FromScalarRaw(ScalarRawValue))
+#endif
+
+// Implementation of exponential function.
+
+// Returns exp(x) for x in [-1/4, 0).
+template <typename tRawType>
+FixedPoint<tRawType, 0> exp_on_interval_between_negative_one_quarter_and_0_excl(
+    FixedPoint<tRawType, 0> a) {
+  typedef FixedPoint<tRawType, 0> F;
+  const F constant_term =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F, 1895147668, std::exp(-1.0 / 8.0));
+  const F constant_1_over_3 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F, 715827883, 1.0 / 3.0);
+  // We're evaluating a Taylor expansion around -1/8, so we do the change of
+  // variable: x = a + 1/8.
+  // In fixed-point with 0 integer bits, 1/8 is represented by 1 << 28.
+  F x = a + F::template ConstantPOT<-3>();
+  F x2 = x * x;
+  F x3 = x2 * x;
+  F x4 = x2 * x2;
+  F x4_over_4 = SaturatingRoundingMultiplyByPOT<-2>(x4);
+  F x4_over_24_plus_x3_over_6_plus_x2_over_2 =
+      SaturatingRoundingMultiplyByPOT<-1>(
+          ((x4_over_4 + x3) * constant_1_over_3) + x2);
+  return constant_term +
+         constant_term * (x + x4_over_24_plus_x3_over_6_plus_x2_over_2);
+}
+
+// Returns exp(x) for x < 0.
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, 0> exp_on_negative_values(
+    FixedPoint<tRawType, tIntegerBits> a) {
+  typedef FixedPoint<tRawType, tIntegerBits> InputF;
+  typedef FixedPoint<tRawType, 0> ResultF;
+  static const int kFractionalBits = InputF::kFractionalBits;
+  static const int kIntegerBits = InputF::kIntegerBits;
+  static const InputF kOneQuarter = InputF::template ConstantPOT<-2>();
+  InputF mask = kOneQuarter - InputF::FromScalarRaw(1);
+  InputF a_mod_quarter_minus_one_quarter = (a & mask) - kOneQuarter;
+  ResultF result = exp_on_interval_between_negative_one_quarter_and_0_excl(
+      Rescale<0>(a_mod_quarter_minus_one_quarter));
+  tRawType remainder = (a_mod_quarter_minus_one_quarter - a).raw();
+
+#define GEMMLOWP_EXP_BARREL_SHIFTER(Exponent, FixedPointMultiplier)         \
+  if (kIntegerBits > Exponent) {                                            \
+    const ResultF kMultiplier = GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(       \
+        ResultF, FixedPointMultiplier, std::exp(-std::pow(2.0, Exponent))); \
+    static constexpr int kShiftAmount =                                     \
+        kIntegerBits > Exponent ? kFractionalBits + Exponent : 0;           \
+    result = SelectUsingMask(                                               \
+        MaskIfNonZero(BitAnd(remainder, Dup<tRawType>(1 << kShiftAmount))), \
+        result * kMultiplier, result);                                      \
+  }
+
+  GEMMLOWP_EXP_BARREL_SHIFTER(-2, 1672461947);
+  GEMMLOWP_EXP_BARREL_SHIFTER(-1, 1302514674);
+  GEMMLOWP_EXP_BARREL_SHIFTER(+0, 790015084);
+  GEMMLOWP_EXP_BARREL_SHIFTER(+1, 290630308);
+  GEMMLOWP_EXP_BARREL_SHIFTER(+2, 39332535);
+  GEMMLOWP_EXP_BARREL_SHIFTER(+3, 720401);
+  GEMMLOWP_EXP_BARREL_SHIFTER(+4, 242);
+
+#undef GEMMLOWP_EXP_BARREL_SHIFTER
+
+  if (kIntegerBits > 5) {
+    static const int b = kIntegerBits > 5 ? kFractionalBits + 5 : 0;
+    const InputF clamp =
+        GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(InputF, -(1 << b), -32.0);
+    result = SelectUsingMask(MaskIfLessThan(a, clamp), ResultF::Zero(), result);
+  }
+
+  result = SelectUsingMask(MaskIfZero(a), ResultF::One(), result);
+  return result;
+}
+
+// Implementation of tanh: (1 - exp(-2x)) / (1 + exp(-2x)).
+
+// Returns (1 - x) / (1 + x) for x in (0, 1).
+template <typename tRawType>
+FixedPoint<tRawType, 0> one_minus_x_over_one_plus_x_for_x_in_0_1(
+    FixedPoint<tRawType, 0> a) {
+  typedef FixedPoint<tRawType, 0> F0;
+  typedef FixedPoint<tRawType, 2> F2;
+  F0 half_denominator = RoundingHalfSum(a, F0::One());
+  // Newton-Raphson division
+  // https://en.wikipedia.org/wiki/Division_algorithm#Newton.E2.80.93Raphson_division
+  // Refer to that page for the logic behind the 48/17 and 32/17 constants.
+  const F2 constant_48_over_17 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, 1515870810, 48.0 / 17.0);
+  const F2 constant_neg_32_over_17 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, -1010580540, -32.0 / 17.0);
+  F2 x = constant_48_over_17 + half_denominator * constant_neg_32_over_17;
+  for (int i = 0; i < 3; i++) {
+    F2 half_denominator_times_x = half_denominator * x;
+    F2 one_minus_half_denominator_times_x =
+        F2::One() - half_denominator_times_x;
+    x = x + Rescale<2>(x * one_minus_half_denominator_times_x);
+  }
+  return Rescale<0>(x - F2::One());
+}
+
+// Returns -tanh(x) for x < 0.
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, 0> neg_tanh_on_negative_values(
+    FixedPoint<tRawType, tIntegerBits> a) {
+  return one_minus_x_over_one_plus_x_for_x_in_0_1(
+      exp_on_negative_values(ExactMulByPot<1>(a)));
+}
+
+// Returns tanh(x) for any x.
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, 0> tanh(FixedPoint<tRawType, tIntegerBits> a) {
+  typedef FixedPoint<tRawType, tIntegerBits> InputF;
+  typedef FixedPoint<tRawType, 0> ResultF;
+  tRawType mask_if_negative = MaskIfLessThan(a, InputF::Zero());
+  tRawType mask_if_zero = MaskIfZero(a);
+  InputF n = SelectUsingMask(mask_if_negative, a, -a);
+  ResultF t = neg_tanh_on_negative_values(n);
+  return SelectUsingMask(mask_if_zero, ResultF::Zero(),
+                         SelectUsingMask(mask_if_negative, -t, t));
+}
+
+// Implementation of logistic function.
+
+// Returns 1 / (1 + x) for x in (0, 1).
+template <typename tRawType>
+FixedPoint<tRawType, 0> one_over_one_plus_x_for_x_in_0_1(
+    FixedPoint<tRawType, 0> a) {
+  typedef FixedPoint<tRawType, 0> F0;
+  typedef FixedPoint<tRawType, 2> F2;
+  F0 half_denominator = RoundingHalfSum(a, F0::One());
+  // Newton-Raphson division
+  // https://en.wikipedia.org/wiki/Division_algorithm#Newton.E2.80.93Raphson_division
+  // Refer to that page for the logic behind the 48/17 and 32/17 constants.
+  const F2 constant_48_over_17 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, 1515870810, 48.0 / 17.0);
+  const F2 constant_neg_32_over_17 =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(F2, -1010580540, -32.0 / 17.0);
+  F2 x = constant_48_over_17 + half_denominator * constant_neg_32_over_17;
+  for (int i = 0; i < 3; i++) {
+    F2 half_denominator_times_x = half_denominator * x;
+    F2 one_minus_half_denominator_times_x =
+        F2::One() - half_denominator_times_x;
+    x = x + Rescale<2>(x * one_minus_half_denominator_times_x);
+  }
+  return Rescale<0>(ExactMulByPot<-1>(x));
+}
+
+// Returns logistic(x) = 1 / (1 + exp(-x)) for x > 0.
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, 0> logistic_on_positive_values(
+    FixedPoint<tRawType, tIntegerBits> a) {
+  return one_over_one_plus_x_for_x_in_0_1(exp_on_negative_values(-a));
+}
+
+// Returns logistic(x) = 1 / (1 + exp(-x)) for any x.
+template <typename tRawType, int tIntegerBits>
+FixedPoint<tRawType, 0> logistic(FixedPoint<tRawType, tIntegerBits> a) {
+  typedef FixedPoint<tRawType, tIntegerBits> InputF;
+  typedef FixedPoint<tRawType, 0> ResultF;
+  tRawType mask_if_positive = MaskIfGreaterThan(a, InputF::Zero());
+  tRawType mask_if_zero = MaskIfZero(a);
+  InputF abs_input = SelectUsingMask(mask_if_positive, a, -a);
+  ResultF result_if_positive = logistic_on_positive_values(abs_input);
+  ResultF result_if_negative = ResultF::One() - result_if_positive;
+  const ResultF one_half =
+      GEMMLOWP_CHECKED_FIXEDPOINT_CONSTANT(ResultF, 1 << 30, 0.5);
+  return SelectUsingMask(mask_if_zero, one_half,
+                         SelectUsingMask(mask_if_positive, result_if_positive,
+                                         result_if_negative));
+}
+
+}  // end namespace gemmlowp
+
+#ifdef GEMMLOWP_NEON
+#include "./fixedpoint_neon.h"
+#elif defined(GEMMLOWP_SSE4)
+#include "./fixedpoint_sse.h"
+#endif
+
+#endif  // GEMMLOWP_INTERNAL_FIXEDPOINT_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_neon.h b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_neon.h
new file mode 100644
index 0000000..8b23de2
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_neon.h
@@ -0,0 +1,175 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// fixedpoint_neon.h: optimized NEON specializations of the templates
+// in fixedpoint.h.
+
+#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_
+#define GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_
+
+#include <arm_neon.h>
+
+namespace gemmlowp {
+
+template <>
+struct FixedPointRawTypeTraits<int32x4_t> {
+  typedef std::int32_t ScalarRawType;
+  static const int kLanes = 4;
+};
+
+template <>
+inline int32x4_t BitAnd(int32x4_t a, int32x4_t b) {
+  return vandq_s32(a, b);
+}
+
+template <>
+inline int32x4_t BitOr(int32x4_t a, int32x4_t b) {
+  return vorrq_s32(a, b);
+}
+
+template <>
+inline int32x4_t BitXor(int32x4_t a, int32x4_t b) {
+  return veorq_s32(a, b);
+}
+
+template <>
+inline int32x4_t BitNot(int32x4_t a) {
+  return veorq_s32(a, vdupq_n_s32(-1));
+}
+
+template <>
+inline int32x4_t Add(int32x4_t a, int32x4_t b) {
+  return vaddq_s32(a, b);
+}
+
+template <>
+inline int32x4_t Sub(int32x4_t a, int32x4_t b) {
+  return vsubq_s32(a, b);
+}
+
+template <>
+inline int32x4_t Neg(int32x4_t a) {
+  return vnegq_s32(a);
+}
+
+template <>
+inline int32x4_t ShiftLeft(int32x4_t a, int offset) {
+  return vshlq_s32(a, vdupq_n_s32(offset));
+}
+
+template <>
+inline int32x4_t ShiftRight(int32x4_t a, int offset) {
+  return vshlq_s32(a, vdupq_n_s32(-offset));
+}
+
+template <>
+inline int32x4_t SelectUsingMask(int32x4_t if_mask, int32x4_t then_val,
+                                 int32x4_t else_val) {
+  return vbslq_s32(vreinterpretq_u32_s32(if_mask), then_val, else_val);
+}
+
+template <>
+inline int32x4_t MaskIfEqual(int32x4_t a, int32x4_t b) {
+  return vreinterpretq_s32_u32(vceqq_s32(a, b));
+}
+
+template <>
+inline int32x4_t MaskIfNotEqual(int32x4_t a, int32x4_t b) {
+  return BitNot(MaskIfEqual(a, b));
+}
+
+template <>
+inline int32x4_t MaskIfZero(int32x4_t a) {
+  return MaskIfEqual(a, vdupq_n_s32(0));
+}
+
+template <>
+inline int32x4_t MaskIfNonZero(int32x4_t a) {
+  return vreinterpretq_s32_u32(vtstq_s32(a, a));
+}
+
+template <>
+inline int32x4_t MaskIfGreaterThan(int32x4_t a, int32x4_t b) {
+  return vreinterpretq_s32_u32(vcgtq_s32(a, b));
+}
+
+template <>
+inline int32x4_t MaskIfGreaterThanOrEqual(int32x4_t a, int32x4_t b) {
+  return vreinterpretq_s32_u32(vcgeq_s32(a, b));
+}
+
+template <>
+inline int32x4_t MaskIfLessThan(int32x4_t a, int32x4_t b) {
+  return vreinterpretq_s32_u32(vcltq_s32(a, b));
+}
+
+template <>
+inline int32x4_t MaskIfLessThanOrEqual(int32x4_t a, int32x4_t b) {
+  return vreinterpretq_s32_u32(vcleq_s32(a, b));
+}
+
+template <>
+inline bool All(int32x4_t a) {
+  a = vandq_s32(a, vextq_s32(a, a, 1));
+  a = vandq_s32(a, vextq_s32(a, a, 2));
+  return vgetq_lane_s32(a, 0);
+}
+
+template <>
+inline bool Any(int32x4_t a) {
+  a = vorrq_s32(a, vextq_s32(a, a, 1));
+  a = vorrq_s32(a, vextq_s32(a, a, 2));
+  return vgetq_lane_s32(a, 0);
+}
+
+template <>
+inline int32x4_t RoundingHalfSum(int32x4_t a, int32x4_t b) {
+  return vrhaddq_s32(a, b);
+}
+
+template <>
+inline int32x4_t SaturatingRoundingDoublingHighMul(int32x4_t a, int32x4_t b) {
+  return vqrdmulhq_s32(a, b);
+}
+
+template <>
+inline int32x4_t RoundingDivideByPOT(int32x4_t x, int exponent) {
+  const int32x4_t shift_vec = vdupq_n_s32(-exponent);
+  const int32x4_t fixup = vshrq_n_s32(vandq_s32(x, shift_vec), 31);
+  const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
+  return vrshlq_s32(fixed_up_x, shift_vec);
+}
+
+template <int Exponent>
+struct ImplSaturatingRoundingMultiplyByPOT<Exponent, int32x4_t, 1> {
+  static int32x4_t eval(int32x4_t x) { return vqshlq_n_s32(x, Exponent); }
+};
+
+template <int Exponent>
+struct ImplSaturatingRoundingMultiplyByPOT<Exponent, int32x4_t, -1> {
+  static int32x4_t eval(int32x4_t x) {
+    const int32x4_t fixup = vshrq_n_s32(x, 31);
+    const int32x4_t fixed_up_x = vqaddq_s32(x, fixup);
+    return vrshrq_n_s32(fixed_up_x, -Exponent);
+  }
+};
+
+template <>
+inline int32x4_t Dup<int32x4_t>(std::int32_t x) {
+  return vdupq_n_s32(x);
+}
+
+}  // end namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_FIXEDPOINT_NEON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_sse.h
new file mode 100644
index 0000000..3f2654d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/fixedpoint/fixedpoint_sse.h
@@ -0,0 +1,218 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// fixedpoint_SSE.h: optimized SSE specializations of the templates
+// in fixedpoint.h.
+
+#ifndef GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_
+#define GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_
+
+#include <smmintrin.h>
+#include "fixedpoint.h"
+
+namespace gemmlowp {
+
+template <>
+struct FixedPointRawTypeTraits<__m128i> {
+  typedef std::int32_t ScalarRawType;
+  static const int kLanes = 4;
+};
+
+template <>
+inline __m128i BitAnd(__m128i a, __m128i b) {
+  return _mm_and_si128(a, b);
+}
+
+template <>
+inline __m128i BitOr(__m128i a, __m128i b) {
+  return _mm_or_si128(a, b);
+}
+
+template <>
+inline __m128i BitXor(__m128i a, __m128i b) {
+  return _mm_xor_si128(a, b);
+}
+
+template <>
+inline __m128i BitNot(__m128i a) {
+  return _mm_andnot_si128(a, _mm_set1_epi32(-1));
+}
+
+template <>
+inline __m128i Add(__m128i a, __m128i b) {
+  return _mm_add_epi32(a, b);
+}
+
+template <>
+inline __m128i Mul(__m128i a, __m128i b) {
+  return _mm_mullo_epi32(a, b);
+}
+
+template <>
+inline __m128i Sub(__m128i a, __m128i b) {
+  return _mm_sub_epi32(a, b);
+}
+
+template <>
+inline __m128i Neg(__m128i a) {
+  return _mm_sign_epi32(a, _mm_set1_epi32(-1));
+}
+
+template <>
+inline __m128i ShiftLeft(__m128i a, int offset) {
+  return _mm_slli_epi32(a, offset);
+}
+
+template <>
+inline __m128i ShiftRight(__m128i a, int offset) {
+  return _mm_srai_epi32(a, offset);
+}
+
+template <>
+inline __m128i SelectUsingMask(__m128i if_mask, __m128i then_val,
+                               __m128i else_val) {
+  return _mm_castps_si128(_mm_blendv_ps(_mm_castsi128_ps(else_val),
+                                        _mm_castsi128_ps(then_val),
+                                        _mm_castsi128_ps(if_mask)));
+}
+
+template <>
+inline __m128i MaskIfEqual(__m128i a, __m128i b) {
+  return _mm_cmpeq_epi32(a, b);
+}
+
+template <>
+inline __m128i MaskIfNotEqual(__m128i a, __m128i b) {
+  return BitNot(MaskIfEqual(a, b));
+}
+
+template <>
+inline __m128i MaskIfZero(__m128i a) {
+  return MaskIfEqual(a, _mm_set1_epi32(0));
+}
+
+template <>
+inline __m128i MaskIfNonZero(__m128i a) {
+  return MaskIfNotEqual(a, _mm_set1_epi32(0));
+}
+
+template <>
+inline __m128i MaskIfGreaterThan(__m128i a, __m128i b) {
+  return _mm_cmpgt_epi32(a, b);
+}
+
+template <>
+inline __m128i MaskIfLessThan(__m128i a, __m128i b) {
+  return _mm_cmplt_epi32(a, b);
+}
+
+template <>
+inline __m128i MaskIfGreaterThanOrEqual(__m128i a, __m128i b) {
+  return BitNot(MaskIfLessThan(a, b));
+}
+
+template <>
+inline __m128i MaskIfLessThanOrEqual(__m128i a, __m128i b) {
+  return BitNot(MaskIfGreaterThan(a, b));
+}
+
+/* Assumptions:
+   - All and Any are used on masks.
+   - masks are all_ones for true lanes, all_zeroes otherwise.
+Hence, All means all 128bits set, and Any means any bit set.
+*/
+
+template <>
+inline bool All(__m128i a) {
+  return _mm_testc_si128(a, a);
+}
+
+template <>
+inline bool Any(__m128i a) {
+  return BitNot(_mm_testz_si128(a, a));
+}
+
+template <>
+inline __m128i RoundingHalfSum(__m128i a, __m128i b) {
+  /* __m128i round_bit_mask, a_over_2, b_over_2, round_bit, sum; */
+  /* We divide the inputs before the add to avoid the overflow and costly test
+   */
+  /* of checking if an overflow occured on signed add */
+  /* round_bit_mask = _mm_set1_epi32(1); */
+  /* a_over_2 = _mm_srai_epi32(a, 1); */
+  /* b_over_2 = _mm_srai_epi32(b, 1); */
+  /* sum = Add(a_over_2, b_over_2); */
+  /* round_bit = _mm_sign_epi32(BitAnd(BitOr(a,b), round_bit_mask), sum); */
+  /* return Add(sum, round_bit); */
+
+  /* Other possibility detecting overflow and xor the sign if an overflow
+   * happened*/
+  __m128i one, sign_bit_mask, sum, rounded_half_sum, overflow, result;
+  one = _mm_set1_epi32(1);
+  sign_bit_mask = _mm_set1_epi32(0x80000000);
+  sum = Add(a, b);
+  rounded_half_sum = _mm_srai_epi32(Add(sum, one), 1);
+  overflow =
+      BitAnd(BitAnd(BitXor(a, rounded_half_sum), BitXor(b, rounded_half_sum)),
+             sign_bit_mask);
+  result = BitXor(rounded_half_sum, overflow);
+  return result;
+}
+
+template <>
+inline __m128i SaturatingRoundingDoublingHighMul(__m128i a, __m128i b) {
+  __m128i min, saturation_mask, a0_a2, a1_a3, b0_b2, b1_b3;
+  __m128i a0b0_a2b2, a1b1_a3b3, a0b0_a2b2_rounded, a1b1_a3b3_rounded;
+  __m128i a0b0_a2b2_rounded_2x, a1b1_a3b3_rounded_2x, result;
+  __m128i nudge;
+
+  // saturation only happen if a == b == INT_MIN
+  min = _mm_set1_epi32(std::numeric_limits<std::int32_t>::min());
+  saturation_mask = BitAnd(MaskIfEqual(a, b), MaskIfEqual(a, min));
+
+  // a = a0 | a1 | a2 | a3
+  // b = b0 | b1 | b2 | b3
+  a0_a2 = a;
+  a1_a3 = _mm_srli_si128(a, 4);
+  b0_b2 = b;
+  b1_b3 = _mm_srli_si128(b, 4);
+
+  a0b0_a2b2 = _mm_mul_epi32(a0_a2, b0_b2);
+  a1b1_a3b3 = _mm_mul_epi32(a1_a3, b1_b3);
+
+  // do the rounding and take into account that it will be doubled
+  nudge = _mm_set1_epi64x(1 << 30);
+  a0b0_a2b2_rounded = _mm_add_epi64(a0b0_a2b2, nudge);
+  a1b1_a3b3_rounded = _mm_add_epi64(a1b1_a3b3, nudge);
+
+  // do the doubling
+  a0b0_a2b2_rounded_2x = _mm_slli_epi64(a0b0_a2b2_rounded, 1);
+  a1b1_a3b3_rounded_2x = _mm_slli_epi64(a1b1_a3b3_rounded, 1);
+
+  // get the high part of the products
+  result = _mm_blend_epi16(_mm_srli_si128(a0b0_a2b2_rounded_2x, 4),
+                           a1b1_a3b3_rounded_2x, 0xcc);
+
+  // saturate those which overflowed
+  return SelectUsingMask(saturation_mask, min, result);
+}
+
+template <>
+inline __m128i Dup<__m128i>(std::int32_t x) {
+  return _mm_set1_epi32(x);
+}
+
+}  // end namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_FIXEDPOINT_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/allocator.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/allocator.h
new file mode 100644
index 0000000..da325a4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/allocator.h
@@ -0,0 +1,220 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// allocator.h: a buffer allocator that allows avoiding most of the
+// malloc/free overhead, by:
+// 1. Requiring all N allocations to be reserved in advance, and
+//    then commited at once, turning N allocations into 1.
+// 2. Being persistent, the allocated storage is reused across commits,
+//    and only reallocated as needed when the commit size gets larger.
+//
+// This is driven by Android-specific needs:
+// 1. On Android, the default (Bionic) allocator tends to aggressively
+// unmap pages, which means that malloc/free can be surprisingly expensive.
+// 2. On Android, stack allocations with alloca() can't be as large as on
+// desktop platforms.
+//
+// General usage:
+// 1. Reserve blocks by calling Reserve(), which returns a Handle.
+// 2. Call Commit() once.
+// 3. Now it is possible to get pointers to allocated buffers by calling
+//    GetPointer().
+// 4. Call Decommit() once.
+// 5. The allocator is now reverted to its original state, except that
+//    it retained its allocated storage, so the next Commit() will be faster.
+//    The allocated storage is only freed when the Allocator object is
+//    destroyed.
+
+#ifndef GEMMLOWP_INTERNAL_ALLOCATOR_H_
+#define GEMMLOWP_INTERNAL_ALLOCATOR_H_
+
+#include "common.h"
+
+#if defined(__ANDROID__)
+#include <android/api-level.h>
+// The 18 here should be 16, but has to be 18 for now due
+// to a Google-internal issue.
+#if __ANDROID_API__ < 18
+#include <malloc.h>
+#define GEMMLOWP_USE_MEMALIGN
+#endif
+// posix_memalign is missing on some 4.1 x86 devices
+#if __ANDROID_API__ == 18
+#ifdef GEMMLOWP_X86_32
+#include <malloc.h>
+#define GEMMLOWP_USE_MEMALIGN
+#endif
+#endif
+#endif
+
+namespace gemmlowp {
+
+enum class TypeId : std::uint8_t { Uint8, Int8, Uint16, Int16, Uint32, Int32 };
+
+template <typename T>
+struct GetTypeIdImpl {};
+
+template <typename T>
+inline TypeId GetTypeId() {
+  return GetTypeIdImpl<T>::Value;
+}
+
+template <typename T>
+struct GetTypeIdImpl<const T> : GetTypeIdImpl<T> {};
+
+#define GEMMLOWP_REGISTER_TYPEID(type_, id) \
+  template <>                               \
+  struct GetTypeIdImpl<type_> {             \
+    static const TypeId Value = TypeId::id; \
+  };
+
+GEMMLOWP_REGISTER_TYPEID(std::uint8_t, Uint8)
+GEMMLOWP_REGISTER_TYPEID(std::int8_t, Int8)
+GEMMLOWP_REGISTER_TYPEID(std::uint16_t, Uint16)
+GEMMLOWP_REGISTER_TYPEID(std::int16_t, Int16)
+GEMMLOWP_REGISTER_TYPEID(std::uint32_t, Uint32)
+GEMMLOWP_REGISTER_TYPEID(std::int32_t, Int32)
+
+class Allocator {
+ public:
+  Allocator()
+      : committed_(false),
+        storage_size_(0),
+        storage_(nullptr),
+        reserved_blocks_(0),
+        reserved_bytes_(0),
+        generation_(0) {}
+
+  ~Allocator() {
+    assert(!committed_);
+    assert(!reserved_blocks_);
+    DeallocateStorage();
+  }
+
+  // Alignment of allocated blocks.
+  static const std::size_t kAlignment = kDefaultCacheLineSize;
+
+  // This is all we need so far, and since the usage pattern is fixed,
+  // there is no point in allowing more until we need to.
+  static const std::size_t kMaxBlocks = 5;
+
+  void Commit() {
+    assert(!committed_);
+
+    if (reserved_bytes_ > storage_size_) {
+      DeallocateStorage();
+      storage_size_ = RoundUpToPowerOfTwo(reserved_bytes_);
+#ifdef GEMMLOWP_USE_MEMALIGN
+      storage_ = memalign(kAlignment, storage_size_);
+#else
+      if (posix_memalign(&storage_, kAlignment, storage_size_)) {
+        storage_ = nullptr;
+      }
+#endif
+    }
+
+    ReleaseBuildAssertion(!storage_size_ || storage_, "allocation failure");
+    committed_ = true;
+  }
+
+  void Decommit() {
+    assert(committed_);
+    committed_ = false;
+    generation_++;
+
+    reserved_blocks_ = 0;
+    reserved_bytes_ = 0;
+  }
+
+  // See generation_
+  typedef std::size_t generation_t;
+
+  // A handle on a reserved block. The user obtains
+  // one by calling Reserve() and, after committing,
+  // passes it to GetPointer().
+  class Handle {
+    std::uint8_t index_;
+    generation_t generation_;
+    TypeId type_;
+
+    friend class Allocator;
+  };
+
+  // Reserves a block sized for n elements of type T, and
+  // returns a handle to it. Must be called before committing.
+  template <typename T>
+  Handle Reserve(std::size_t n) {
+    assert(!committed_ && "can't reserve blocks while committed");
+    assert(reserved_blocks_ < kMaxBlocks &&
+           "didn't expect to allocate this many blocks");
+    const std::size_t bytes = RoundUp<kAlignment>(n * sizeof(T));
+    const std::size_t offset = reserved_bytes_;
+    const std::size_t index = reserved_blocks_;
+
+    reserved_blocks_offsets_[index] = offset;
+    Handle h;
+    h.index_ = index;
+    h.generation_ = generation_;
+    h.type_ = GetTypeId<T>();
+
+    reserved_blocks_++;
+    reserved_bytes_ += bytes;
+
+    return h;
+  }
+
+  // Returns the pointer to the allocated buffer for the given handle.
+  // Must be called after committing.
+  template <typename T>
+  T* GetPointer(const Handle& h) const {
+    assert(committed_ && "can't get block pointers unless committed");
+    assert(h.index_ < reserved_blocks_ &&
+           "bad handle, points to inexistant block");
+    assert(h.generation_ == generation_ &&
+           "handle from earlier generation, have decommitted since");
+    assert(h.type_ == GetTypeId<T>() && "type mismatch");
+    std::size_t offset = reserved_blocks_offsets_[h.index_];
+    std::uintptr_t addr = reinterpret_cast<std::uintptr_t>(storage_) + offset;
+    return reinterpret_cast<T*>(addr);
+  }
+
+ private:
+  void DeallocateStorage() {
+    assert(!committed_);
+    free(storage_);
+    storage_size_ = 0;
+  }
+
+  // Set to true by Commit() and to false by Decommit(). Initially false.
+  bool committed_;
+
+  // The actually allocated storage size and buffer pointer.
+  std::size_t storage_size_;
+  mutable void* storage_;
+
+  // The number of blocks that have been reserved by Reserve().
+  std::size_t reserved_blocks_;
+  // The number of bytes that have been reserved by Reserve().
+  std::size_t reserved_bytes_;
+  // The offsets of reserved blocks into the storage buffer.
+  std::size_t reserved_blocks_offsets_[kMaxBlocks];
+
+  // The 'generation' is incremented on Decommit() and allows catching
+  // bad GetPointer() calls still referring to a previous commit.
+  generation_t generation_;
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_ALLOCATOR_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/block_params.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/block_params.h
new file mode 100644
index 0000000..b2fc3ff
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/block_params.h
@@ -0,0 +1,174 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// block_params.h: Logic to choose L1 and L2 block sizes
+// to optimize cache-friendliness.
+
+#ifndef GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
+#define GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
+
+#include "common.h"
+
+namespace gemmlowp {
+
+// A BlockParams instance contains a full description of all the block size
+// parameters to be used by a Gemm.
+// There are two nested levels of block subdivisions: first a subdivision
+// into large blocks that should fit in last-level cache (what we call L2 here)
+// and then another subdivision into smaller blocks that should fit in
+// L1 cache. There is then actually a third level of subdivision to fit
+// in registers, but we are not concerned with that here.
+struct BlockParams {
+  // L1 block parameters determine the size of small blocks that should
+  // fit in L1 cache.
+  int l1_rows;
+  int l1_cols;
+  int l1_depth;
+
+  // L2 block parameters determine the size of larger blocks that should
+  // fit in L2 cache.
+  int l2_rows;
+  int l2_cols;
+  int l2_depth;
+
+  template <typename KernelFormat>
+  void Init(int rows, int cols, int depth, int num_threads,
+            int l1_bytes_to_use, int l2_bytes_to_use, float l2_rhs_factor) {
+    FindL2BlockSizes<KernelFormat>(rows, cols, depth, num_threads,
+                                   l2_bytes_to_use, l2_rhs_factor,
+                                   &l2_rows, &l2_cols, &l2_depth);
+    FindL1BlockSizes<KernelFormat>(l2_rows, l2_cols, l2_depth,
+                                   l1_bytes_to_use,
+                                   &l1_rows, &l1_cols, &l1_depth);
+  }
+
+  template <typename KernelFormat>
+  static void FindL2BlockSizes(int rows, int cols, int depth, int num_threads,
+                               int l2_bytes_to_use, float l2_rhs_factor,
+                               int* out_l2_rows, int* out_l2_cols,
+                               int* out_l2_depth) {
+    int l2_rows = 0;
+    int l2_cols = 0;
+    int l2_depth = 0;
+    // No L2 blocking in the depth dimension at the moment.
+    // Too much loss of accuracy due to storing intermediate results in
+    // low precision.
+    // However, we still want to round l2_depth up to the next multiple
+    // of register size, so as to avoid having to special-case unaligned depths.
+    l2_depth = RoundUp<kRegisterSize>(depth);
+
+    {
+      int max_cache_friendly_l2_cols = std::max(
+          1, static_cast<int>(l2_rhs_factor * (l2_bytes_to_use / l2_depth)));
+      int min_l2_cols_blocks =
+          std::max(1, CeilQuotient(cols, max_cache_friendly_l2_cols));
+      l2_cols =
+          RoundUp<KernelFormat::kCols>(CeilQuotient(cols, min_l2_cols_blocks));
+    }
+
+    // No L2 blocking in the row dimension if l2_rhs_factor is 1.0 as the row
+    // dimension concerns only the LHS. Blocking only RHS matrix for L2 enhances
+    // the performance on x86.
+    if (l2_rhs_factor == 1.0f) {
+      l2_rows = RoundUp<KernelFormat::kRows>(rows);
+    } else {
+      int max_cache_friendly_l2_rows =
+          std::max(1, (l2_bytes_to_use - l2_depth * l2_cols) /
+                          (num_threads * (l2_depth + 4 * l2_cols)));
+      int min_l2_rows_blocks =
+          std::max(1, CeilQuotient(rows, max_cache_friendly_l2_rows));
+      l2_rows =
+          RoundUp<KernelFormat::kRows>(CeilQuotient(rows, min_l2_rows_blocks));
+    }
+
+    *out_l2_rows = l2_rows;
+    *out_l2_cols = l2_cols;
+    *out_l2_depth = l2_depth;
+  }
+
+  template <typename KernelFormat>
+  static void FindL1BlockSizes(int rows, int cols, int depth,
+                               int l1_bytes_to_use, int* out_l1_rows,
+                               int* out_l1_cols, int* out_l1_depth) {
+    int l1_rows = 0;
+    int l1_cols = 0;
+    int l1_depth = 0;
+
+    // L2 block sizes should already be multiples of kernel block sizes.
+    assert(rows % KernelFormat::kRows == 0);
+    assert(cols % KernelFormat::kCols == 0);
+    assert(depth % KernelFormat::kDepth == 0);
+
+    // No L1 blocking in the columns dimension at the moment.
+    // Thought not to be needed. Similar to Eigen.
+    l1_cols = cols;
+
+    {
+      int max_cache_friendly_l1_depth = std::max(
+          1, (l1_bytes_to_use - 4 * KernelFormat::kRows * KernelFormat::kCols) /
+                 (KernelFormat::kRows + KernelFormat::kCols));
+      int min_l1_depth_blocks =
+          std::max(1, CeilQuotient(depth, max_cache_friendly_l1_depth));
+      l1_depth =
+          RoundUp<kRegisterSize>(CeilQuotient(depth, min_l1_depth_blocks));
+    }
+
+    {
+      int max_cache_friendly_l1_rows =
+          std::max(1, l1_bytes_to_use / (l1_depth + 4 * l1_cols));
+      int min_l1_rows_blocks =
+          std::max(1, CeilQuotient(rows, max_cache_friendly_l1_rows));
+      l1_rows =
+          RoundUp<KernelFormat::kRows>(CeilQuotient(rows, min_l1_rows_blocks));
+    }
+
+    *out_l1_rows = l1_rows;
+    *out_l1_cols = l1_cols;
+    *out_l1_depth = l1_depth;
+  }
+};
+
+// A SideBlockParams instance contains only the block params relevant to
+// one side (LHS or RHS), expressed in terms of 'width' instead of
+// rows/colums. See the explanation in kernel.h: in the LHS, 'width' means
+// the number of rows, while in the RHS, 'width' means the number of columns.
+// That allows us to write generic code that applies to either LHS or RHS.
+struct SideBlockParams {
+  // L1 block parameters determine the size of small blocks that should
+  // fit in L1 cache.
+  int l1_width;
+  int l1_depth;
+
+  // L2 block parameters determine the size of larger blocks that should
+  // fit in L2 cache.
+  int l2_width;
+  int l2_depth;
+};
+
+enum class Side { Lhs, Rhs };
+
+inline void GetSideBlockParams(Side side, SideBlockParams* side_block_params,
+                               const BlockParams& block_params) {
+  side_block_params->l1_width =
+      side == Side::Lhs ? block_params.l1_rows : block_params.l1_cols;
+  side_block_params->l2_width =
+      side == Side::Lhs ? block_params.l2_rows : block_params.l2_cols;
+
+  side_block_params->l1_depth = block_params.l1_depth;
+  side_block_params->l2_depth = block_params.l2_depth;
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_BLOCK_PARAMS_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/common.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/common.h
new file mode 100644
index 0000000..511809d
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/common.h
@@ -0,0 +1,256 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// common.h: contains stuff that's used throughout gemmlowp
+// and should always be available.
+
+#ifndef GEMMLOWP_INTERNAL_COMMON_H_
+#define GEMMLOWP_INTERNAL_COMMON_H_
+
+#include <pthread.h>
+
+#include <algorithm>
+#include <cassert>
+#include <cmath>
+#include <cstdlib>
+
+#include "../profiling/instrumentation.h"
+
+// Our inline assembly path assume GCC/Clang syntax.
+// Native Client doesn't seem to support inline assembly(?).
+#if defined(__GNUC__) && !defined(__native_client__)
+#define GEMMLOWP_ALLOW_INLINE_ASM
+#endif
+
+// Define macro statement that avoids inlining for GCC.
+// For non-GCC, define as empty macro.
+#if defined(__GNUC__)
+#define GEMMLOWP_NOINLINE __attribute__((noinline))
+#else
+#define GEMMLOWP_NOINLINE
+#endif
+
+// Detect ARM, 32-bit or 64-bit
+#ifdef __arm__
+#define GEMMLOWP_ARM_32
+#endif
+
+#ifdef __aarch64__
+#define GEMMLOWP_ARM_64
+#endif
+
+#if defined(GEMMLOWP_ARM_32) || defined(GEMMLOWP_ARM_64)
+#define GEMMLOWP_ARM
+#endif
+
+// Detect x86, 32-bit or 64-bit
+#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) || defined(__i386)
+#define GEMMLOWP_X86_32
+#endif
+
+#if defined(__x86_64__) || defined(_M_X64) || defined(__amd64)
+#define GEMMLOWP_X86_64
+#endif
+
+#if defined(GEMMLOWP_X86_32) || defined(GEMMLOWP_X86_64)
+#define GEMMLOWP_X86
+#endif
+
+// Some of our optimized paths use inline assembly and for
+// now we don't bother enabling some other optimized paths using intrinddics
+// where we can't use inline assembly paths.
+#ifdef GEMMLOWP_ALLOW_INLINE_ASM
+
+// Detect NEON. It's important to check for both tokens.
+#if (defined __ARM_NEON) || (defined __ARM_NEON__)
+#define GEMMLOWP_NEON
+#endif
+
+// Convenience NEON tokens for 32-bit or 64-bit
+#if defined(GEMMLOWP_NEON) && defined(GEMMLOWP_ARM_32)
+#define GEMMLOWP_NEON_32
+#endif
+
+#if defined(GEMMLOWP_NEON) && defined(GEMMLOWP_ARM_64)
+#define GEMMLOWP_NEON_64
+#endif
+
+// Detect SSE.
+#ifdef __SSE4_1__
+#define GEMMLOWP_SSE4
+#endif
+
+#ifdef __SSE3__
+#define GEMMLOWP_SSE3
+#endif
+
+// Convenience SSE4 tokens for 32-bit or 64-bit
+#if defined(GEMMLOWP_SSE4) && defined(GEMMLOWP_X86_32)
+#define GEMMLOWP_SSE4_32
+#endif
+
+#if defined(GEMMLOWP_SSE3) && defined(GEMMLOWP_X86_32)
+#define GEMMLOWP_SSE3_32
+#endif
+
+#if defined(GEMMLOWP_SSE4) && defined(GEMMLOWP_X86_64)
+#define GEMMLOWP_SSE4_64
+#endif
+
+#if defined(GEMMLOWP_SSE3) && defined(GEMMLOWP_X86_64)
+#define GEMMLOWP_SSE3_64
+#endif
+
+#endif  // GEMMLOWP_ALLOW_INLINE_ASM
+
+// Detect Android. Don't conflate with ARM - we care about tuning
+// for non-ARM Android devices too. This can be used in conjunction
+// with x86 to tune differently for mobile x86 CPUs (Atom) vs. desktop x86 CPUs.
+#if defined(__ANDROID__)
+#define GEMMLOWP_ANDROID
+#endif
+
+namespace gemmlowp {
+
+// Standard cache line size. Useful to optimize alignment and
+// prefetches. Ideally we would query this at runtime, however
+// 64 byte cache lines are the vast majority, and even if it's
+// wrong on some device, it will be wrong by no more than a 2x factor,
+// which should be acceptable.
+const int kDefaultCacheLineSize = 64;
+
+// Default L1 and L2 data cache sizes.
+// The L1 cache size is assumed to be for each core.
+// The L2 cache size is assumed to be shared among all cores. What
+// we call 'L2' here is effectively top-level cache.
+//
+// On x86, we should ideally query this at
+// runtime. On ARM, the instruction to query this is privileged and
+// Android kernels do not expose it to userspace. Fortunately, the majority
+// of ARM devices have roughly comparable values:
+//   Nexus 5: L1 16k, L2 1M
+//   Android One: L1 32k, L2 512k
+// The following values are equal to or somewhat lower than that, and were
+// found to perform well on both the Nexus 5 and Android One.
+// Of course, these values are in principle too low for typical x86 CPUs
+// where we should set the L2 value to (L3 cache size / number of cores) at
+// least.
+//
+#if defined(GEMMLOWP_ARM) && defined(__APPLE__)
+// iPhone/iPad
+const int kDefaultL1CacheSize = 48 * 1024;
+const int kDefaultL2CacheSize = 2 * 1024 * 1024;
+#elif defined(GEMMLOWP_ARM) || defined(GEMMLOWP_ANDROID)
+// Other ARM or ARM-like hardware (Android implies ARM-like) so here it's OK
+// to tune for ARM, although on x86 Atom we might be able to query
+// cache sizes at runtime, which would be better.
+const int kDefaultL1CacheSize = 16 * 1024;
+const int kDefaultL2CacheSize = 384 * 1024;
+#elif defined(GEMMLOWP_X86_64)
+// x86-64 and not Android. Therefore, likely desktop-class x86 hardware.
+// Thus we assume larger cache sizes, though we really should query
+// them at runtime.
+const int kDefaultL1CacheSize = 32 * 1024;
+const int kDefaultL2CacheSize = 4 * 1024 * 1024;
+#elif defined(GEMMLOWP_X86_32)
+// x86-32 and not Android. Same as x86-64 but less bullish.
+const int kDefaultL1CacheSize = 32 * 1024;
+const int kDefaultL2CacheSize = 2 * 1024 * 1024;
+#else
+// Less common hardware. Maybe some unusual or older or embedded thing.
+// Assume smaller caches, but don't depart too far from what we do
+// on ARM/Android to avoid accidentally exposing unexpected behavior.
+const int kDefaultL1CacheSize = 16 * 1024;
+const int kDefaultL2CacheSize = 256 * 1024;
+#endif
+
+// The proportion of the cache that we intend to use for storing
+// RHS blocks. This should be between 0 and 1, and typically closer to 1,
+// as we typically want to use most of the L2 cache for storing a large
+// RHS block.
+#if defined(GEMMLOWP_X86)
+// For IA, use the entire L2 cache for the RHS matrix. LHS matrix is not blocked
+// for L2 cache.
+const float kDefaultL2RhsFactor = 1.00f;
+#else
+const float kDefaultL2RhsFactor = 0.75f;
+#endif
+
+// The number of bytes in a SIMD register. This is used to determine
+// the dimensions of PackingRegisterBlock so that such blocks can
+// be efficiently loaded into registers, so that packing code can
+// work within registers as much as possible.
+// In the non-SIMD generic fallback code, this is just a generic array
+// size, so any size would work there. Different platforms may set this
+// to different values but must ensure that their own optimized packing paths
+// are consistent with this value.
+const int kRegisterSize = 16;
+
+// Hints the CPU to prefetch the cache line containing ptr.
+inline void Prefetch(const void* ptr) {
+#if defined GEMMLOWP_ARM_64 && defined GEMMLOWP_ALLOW_INLINE_ASM
+  // Aarch64 has very detailed prefetch instructions, that compilers
+  // can't know how to map __builtin_prefetch to, and as a result, don't,
+  // leaving __builtin_prefetch a no-op on this architecture.
+  // For our purposes, "pldl1keep" is usually what we want, meaning:
+  // "prefetch for load, into L1 cache, using each value multiple times".
+  asm volatile("prfm pldl1keep, [%[ptr]]\n" ::[ptr] "r"(ptr) : );
+#elif defined \
+    __GNUC__  // Clang and GCC define __GNUC__ and have __builtin_prefetch.
+  __builtin_prefetch(ptr);
+#else
+  (void)ptr;
+#endif
+}
+
+// Returns the runtime argument rounded down to the nearest multiple of
+// the fixed Modulus.
+template <unsigned Modulus, typename Integer>
+Integer RoundDown(Integer i) {
+  return i - (i % Modulus);
+}
+
+// Returns the runtime argument rounded up to the nearest multiple of
+// the fixed Modulus.
+template <unsigned Modulus, typename Integer>
+Integer RoundUp(Integer i) {
+  return RoundDown<Modulus>(i + Modulus - 1);
+}
+
+// Returns the quotient a / b rounded up ('ceil') to the nearest integer.
+template <typename Integer>
+Integer CeilQuotient(Integer a, Integer b) {
+  return (a + b - 1) / b;
+}
+
+// Returns the argument rounded up to the nearest power of two.
+template <typename Integer>
+Integer RoundUpToPowerOfTwo(Integer n) {
+  Integer i = n - 1;
+  i |= i >> 1;
+  i |= i >> 2;
+  i |= i >> 4;
+  i |= i >> 8;
+  i |= i >> 16;
+  return i + 1;
+}
+
+template <int N>
+struct IsPowerOfTwo {
+  static const bool value = !(N & (N - 1));
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_COMMON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/compute.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/compute.h
new file mode 100644
index 0000000..bbc9e2a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/compute.h
@@ -0,0 +1,104 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// compute.h: the central stage of the Gemm computation, operates
+// on already-packed LHS and RHS blocks and calls the Gemm kernel
+// to compute a block of the product.
+
+#ifndef GEMMLOWP_INTERNAL_COMPUTE_H_
+#define GEMMLOWP_INTERNAL_COMPUTE_H_
+
+#include "block_params.h"
+#include "kernel.h"
+#include "pack.h"
+
+namespace gemmlowp {
+
+template <typename PackedLhs, typename PackedRhs, typename PackedResult>
+class ComputeImpl {
+  typedef typename PackedLhs::KernelSideFormat KernelLhsFormat;
+  typedef typename PackedRhs::KernelSideFormat KernelRhsFormat;
+  typedef KernelFormat<KernelLhsFormat, KernelRhsFormat> Format;
+
+  const KernelBase& kernel_;
+  const BlockParams& block_params_;
+
+  PackedResult* const packed_result_;
+  const PackedLhs& packed_lhs_;
+  const PackedRhs& packed_rhs_;
+
+ public:
+  ComputeImpl(const KernelBase& _kernel, const BlockParams& _block_params,
+              PackedResult* _packed_result, const PackedLhs& _packed_lhs,
+              const PackedRhs& _packed_rhs)
+      : kernel_(_kernel),
+        block_params_(_block_params),
+        packed_result_(_packed_result),
+        packed_lhs_(_packed_lhs),
+        packed_rhs_(_packed_rhs) {}
+
+  void Compute(int depth) {
+    depth = RoundUp<Format::kDepth>(depth);
+    assert(depth <= block_params_.l2_depth);
+    for (int d = 0; d < depth; d += block_params_.l1_depth) {
+      int ds = std::min(block_params_.l1_depth, depth - d);
+
+      for (int r = 0; r < block_params_.l2_rows; r += block_params_.l1_rows) {
+        int rs = std::min(block_params_.l1_rows, block_params_.l2_rows - r);
+
+        ComputeL1(r, rs, 0, block_params_.l2_cols, d, ds);
+      }
+    }
+  }
+
+ private:
+  void ComputeRun(int start_row, int start_col, int start_depth,
+                  int depth) GEMMLOWP_NOINLINE {
+    packed_lhs_.seek_run(start_row, start_depth);
+    packed_rhs_.seek_run(start_col, start_depth);
+    auto packed_result_block = packed_result_->Map().block(
+        start_row, start_col, Format::kRows, Format::kCols);
+    kernel_.Run(packed_result_block.data(), packed_result_block.rows_stride(),
+                packed_result_block.cols_stride(), packed_lhs_.current_data(),
+                packed_rhs_.current_data(), start_depth, depth);
+  }
+
+  void ComputeL1(int start_row, int rows, int start_col, int cols,
+                 int start_depth, int depth) {
+    assert(rows % Format::kRows == 0);
+    assert(cols % Format::kCols == 0);
+    assert(depth % Format::kDepth == 0);
+
+    for (int c = 0; c < cols; c += Format::kCols) {
+      for (int r = 0; r < rows; r += Format::kRows) {
+        ComputeRun(start_row + r, start_col + c, start_depth, depth);
+      }
+    }
+  }
+};
+
+template <typename PackedLhs, typename PackedRhs, typename PackedResult>
+void Compute(const KernelBase& kernel, const BlockParams& block_params,
+             PackedResult* packed_result, const PackedLhs& packed_lhs,
+             const PackedRhs& packed_rhs, int depth) {
+  ScopedProfilingLabel label("compute");
+  ComputeImpl<PackedLhs, PackedRhs, PackedResult> impl(
+      kernel, block_params, packed_result, packed_lhs, packed_rhs);
+
+  impl.Compute(depth);
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_COMPUTE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/dispatch_gemm_shape.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/dispatch_gemm_shape.h
new file mode 100644
index 0000000..0be0bf3
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/dispatch_gemm_shape.h
@@ -0,0 +1,189 @@
+// Copyright 2017 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// dispatch_gemm_shape.h: dispatch GEMM calls according to their shape
+
+#ifndef GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
+#define GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
+
+#include "../internal/kernel_default.h"
+#include "../public/map.h"
+#include "../public/output_stages.h"
+#include "multi_thread_gemm.h"
+
+namespace gemmlowp {
+
+template <typename T>
+struct TransposeImpl {
+  typedef T DstType;
+  static T Run(const T& t) { return t; }
+};
+
+template <typename T>
+using TransposeType = typename TransposeImpl<T>::DstType;
+
+template <typename T>
+TransposeType<T> Transpose(const T& t) {
+  return TransposeImpl<T>::Run(t);
+}
+
+template <MapOrder Order>
+struct TransposeMapOrder {
+  static constexpr MapOrder Value =
+      Order == MapOrder::RowMajor ? MapOrder::ColMajor : MapOrder::RowMajor;
+};
+
+template <VectorShape Shape>
+struct TransposeVectorShape {
+  static constexpr VectorShape Value =
+      Shape == VectorShape::Row ? VectorShape::Col : VectorShape::Row;
+};
+
+template <typename Scalar, VectorShape Shape>
+struct TransposeImpl<VectorMap<Scalar, Shape>> {
+  typedef VectorMap<Scalar, Shape> SrcType;
+  static constexpr VectorShape TransposedShape =
+      TransposeVectorShape<Shape>::Value;
+  typedef VectorMap<Scalar, TransposedShape> DstType;
+  static DstType Run(const SrcType& src) {
+    return DstType(src.data(), src.size());
+  }
+};
+
+template <typename Scalar, MapOrder Order>
+struct TransposeImpl<MatrixMap<Scalar, Order>> {
+  typedef MatrixMap<Scalar, Order> SrcType;
+  static constexpr MapOrder TransposedOrder = TransposeMapOrder<Order>::Value;
+  typedef MatrixMap<Scalar, TransposedOrder> DstType;
+  static DstType Run(const SrcType& src) {
+    return DstType(src.data(), src.cols(), src.rows(), src.stride());
+  }
+};
+
+template <VectorShape Shape>
+struct TransposeImpl<OutputStageQuantizeDownInt32ToUint8ScalePC<Shape>> {
+  typedef OutputStageQuantizeDownInt32ToUint8ScalePC<Shape> SrcType;
+  static const VectorShape TransposedShape = TransposeVectorShape<Shape>::Value;
+  typedef OutputStageQuantizeDownInt32ToUint8ScalePC<TransposedShape> DstType;
+  static DstType Run(const SrcType& src) {
+    DstType dst;
+    dst.result_shift = src.result_shift;
+    dst.result_offset = Transpose(src.result_offset);
+    dst.result_mult_int = Transpose(src.result_mult_int);
+    return dst;
+  }
+};
+
+template <typename VectorMapType>
+struct TransposeImpl<OutputStageBiasAddition<VectorMapType>> {
+  typedef OutputStageBiasAddition<VectorMapType> SrcType;
+  typedef TransposeType<VectorMapType> TransposedVectorMapType;
+  typedef OutputStageBiasAddition<TransposedVectorMapType> DstType;
+  static DstType Run(const SrcType& src) {
+    DstType dst;
+    dst.bias_vector = Transpose(src.bias_vector);
+    return dst;
+  }
+};
+
+// TODO(benoitjacob) - does anyone understand C++ variadic templates?
+// How to use them to implement TransposeTuple? Note: there are lots
+// of answers on StackOverflow but they seem to all involve either
+// C++14/C++17 (we can only use C++11) or lots of abstract nonsense.
+inline std::tuple<> TransposeTuple(const std::tuple<>& t) { return t; }
+
+template <typename T0>
+std::tuple<TransposeType<T0>> TransposeTuple(const std::tuple<T0>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)));
+}
+
+template <typename T0, typename T1>
+std::tuple<TransposeType<T0>, TransposeType<T1>> TransposeTuple(
+    const std::tuple<T0, T1>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)));
+}
+
+template <typename T0, typename T1, typename T2>
+std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>>
+TransposeTuple(const std::tuple<T0, T1, T2>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
+                         Transpose(std::get<2>(t)));
+}
+
+template <typename T0, typename T1, typename T2, typename T3>
+std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
+           TransposeType<T3>>
+TransposeTuple(const std::tuple<T0, T1, T2, T3>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
+                         Transpose(std::get<2>(t)), Transpose(std::get<3>(t)));
+}
+
+template <typename T0, typename T1, typename T2, typename T3, typename T4>
+std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
+           TransposeType<T3>, TransposeType<T4>>
+TransposeTuple(const std::tuple<T0, T1, T2, T3, T4>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
+                         Transpose(std::get<2>(t)), Transpose(std::get<3>(t)),
+                         Transpose(std::get<4>(t)));
+}
+
+template <typename T0, typename T1, typename T2, typename T3, typename T4,
+          typename T5>
+std::tuple<TransposeType<T0>, TransposeType<T1>, TransposeType<T2>,
+           TransposeType<T3>, TransposeType<T4>, TransposeType<T5>>
+TransposeTuple(const std::tuple<T0, T1, T2, T3, T4, T5>& t) {
+  return std::make_tuple(Transpose(std::get<0>(t)), Transpose(std::get<1>(t)),
+                         Transpose(std::get<2>(t)), Transpose(std::get<3>(t)),
+                         Transpose(std::get<4>(t)), Transpose(std::get<5>(t)));
+}
+
+template <typename InputScalar, typename OutputScalar, typename BitDepthParams,
+          MapOrder LhsOrder, MapOrder RhsOrder, MapOrder ResultOrder,
+          typename LhsOffset, typename RhsOffset, typename OutputPipelineType,
+          typename GemmContextType>
+void DispatchGemmShape(GemmContextType* context,
+                       const MatrixMap<const InputScalar, LhsOrder>& lhs,
+                       const MatrixMap<const InputScalar, RhsOrder>& rhs,
+                       MatrixMap<OutputScalar, ResultOrder>* result,
+                       const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
+                       const OutputPipelineType& output_pipeline) {
+  assert(lhs.cols() == rhs.rows());
+
+  int rows = result->rows();
+  int cols = result->cols();
+  int depth = lhs.cols();
+
+  if (rows == 0 || cols == 0 || depth == 0) {
+    // Vacuous GEMM, return early to avoid having to deal with
+    // zero sizes below.
+    return;
+  }
+
+  if (rows < cols) {
+    auto transposed_result_map = Transpose(*result);
+    return DispatchGemmShape<InputScalar, OutputScalar, BitDepthParams>(
+        context, Transpose(rhs), Transpose(lhs), &transposed_result_map,
+        Transpose(rhs_offset), Transpose(lhs_offset),
+        TransposeTuple(output_pipeline));
+  }
+
+  typedef DefaultKernel<BitDepthParams> Kernel;
+  MultiThreadGemm<typename Kernel::Format, InputScalar, OutputScalar,
+                  BitDepthParams>(context, Kernel(), lhs, rhs, result,
+                                  lhs_offset, rhs_offset, output_pipeline);
+}
+
+}  // end namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_DISPATCH_GEMM_SHAPE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel.h
new file mode 100644
index 0000000..4d006af
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel.h
@@ -0,0 +1,234 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// kernel.h: general definitions for kernels.
+
+#ifndef GEMMLOWP_INTERNAL_KERNEL_H_
+#define GEMMLOWP_INTERNAL_KERNEL_H_
+
+#include "../public/bit_depth.h"
+#include "common.h"
+
+namespace gemmlowp {
+
+// Explanation of general gemmlowp terminology
+// ===========================================
+//
+// We use the following abbreviations:
+// LHS = "left-hand side"
+// RHS = "right-hand side"
+// Sometimes when referring to either LHS or RHS, we just say a "Side".
+//
+// In a matrix product of a MxK matrix times a KxN matrix,
+// we call K the 'depth'. Note that M is the number of rows
+// of the result (and of the LHS), and N is the number of columns
+// of the result (and of the RHS).
+//
+// In each of the LHS and RHS matrices, we call 'width' the
+// other dimension, besides the depth. So in the LHS, 'width'
+// is the number of rows, while in the RHS, 'width' is the number
+// of columns.
+//
+//  So in the LHS MxK matrix, the depth is K and the width in M.
+// And in the RHS KxN matrix, the depth is K and the width in N.
+//
+// This is illustrated in this picture:
+//
+//                             RHS width
+//                        <----------------->
+//                        +-----------------+ ^
+//                        |       RHS       | | Depth
+//                        +-----------------+ v
+//                 ^ +--+ +-----------------+
+//                 | |L | |                 |
+//       LHS width | |H | |      Result     |
+//                 | |S | |                 |
+//                 v +--+ +-----------------+
+//                   <-->
+//                   Depth
+
+// Explanation of gemmlowp kernel formats and "cells"
+// ==================================================
+//
+// Kernels operate on small LHS and RHS blocks that fit in registers.
+// These blocks are stored contiguously in memory, but not always
+// in a traditional column-major or row-major order; instead,
+// they consist of a number of sub-blocks, which we call "cells",
+// that are stored in column-major or row-major order. However,
+// what really matters to us is not so much rows vs columns, but
+// rather width vs depth. So we refer to "width-major" and "depth-major"
+// storage orders. In the LHS, width-major means row-major,
+// while in the RHS, width-major means column-major.
+// There is also a third possibility, "diagonal order",
+// which is unused at the moment.
+//
+// We aim to treat both sides, LHS and RHS, on an equal footing,
+// so we call them both 'sides'. A KernelFormat thus is just a pair
+// of KernelSideFormat's, one for LHS and one for RHS; each KernelSideFormat
+// contains a CellFormat and a number of cells; cells are only ever
+// stacked in the width dimension, which means stacked vertically in the
+// LHS and stacked horizondally in the RHS.
+//
+// Example
+// =======
+//
+// Let's work out the data layout expected by a kernel having the
+// following format (the struct names here are defined below in this file):
+//
+// KernelFormat<
+//   KernelSideFormat<CellFormat<3, 4>, 3>,
+//   KernelSideFormat<CellFormat<5, 4>, 2>
+// >
+//
+// The LHS format, KernelSideFormat<CellFormat<3, 4>, 3>, means:
+// 3 cells, each cell having dimensions (width=3, depth=4), laid out in
+// DepthMajor order (the default value, see CellFormat). In the LHS,
+// DepthMajor means column-major, so the LHS cells are of size 3x4 in
+// column-major order, so the LHS layout is:
+//
+// 0  3  6  9
+// 1  4  7  10
+// 2  5  8  11
+// 12 15 18 21
+// 13 16 19 22
+// 14 17 20 23
+// 24 27 30 33
+// 25 28 31 34
+// 26 29 32 35
+//
+// The RHS format, KernelSideFormat<CellFormat<5, 4>, 2>, means:
+// 2 cells each having dimensions (width=5, depth=4), laid out in
+// DepthMajor order (the default value, see CellFormat). In the RHS,
+// DepthMajor means row-major, so the RHS cells are of size 4x5 in
+// row-major order, so the RHS layout is:
+//
+// 0  1  2  3  4  20 21 22 23 24
+// 5  6  7  8  9  25 26 27 28 29
+// 10 11 12 13 14 30 31 32 33 34
+// 15 16 17 18 19 35 36 37 38 39
+
+// CellOrder enumerates the possible storage orders (=layouts) for
+// a cell (see explanation above).
+enum class CellOrder { DepthMajor, WidthMajor, Diagonal };
+
+// CellFormat describes how data is laid
+// out in a cell. That is, a CellOrder together with actual dimensions.
+template <int tWidth, int tDepth, CellOrder tOrder = CellOrder::DepthMajor>
+struct CellFormat {
+  static const int kWidth = tWidth;
+  static const int kDepth = tDepth;
+  static const CellOrder kOrder = tOrder;
+
+  static const int kSize = kWidth * kDepth;
+};
+
+// KernelSideFormat describes how data is laid out in a kernel side
+// (i.e. LHS or RHS). That is, a CellFormat together with a number of
+// cells. These cells are always stacked in the Width dimension.
+// For example, in the LHS case, the Width dimension is the rows dimension,
+// se we're saying that in the LHS, cells are stacked vertically.
+// We never stack cells in the Depth dimension.
+template <typename tCellFormat, int tCells>
+struct KernelSideFormat {
+  typedef tCellFormat Cell;
+  static const int kCells = tCells;
+  static const int kWidth = kCells * Cell::kWidth;
+  static const int kDepth = Cell::kDepth;
+  typedef std::uint8_t Scalar;
+};
+
+template <typename tCellFormat, int tCells>
+struct KernelSideFormatInt8 : KernelSideFormat<tCellFormat, tCells> {
+  typedef std::int8_t Scalar;
+};
+
+// KernelFormat describes fully the input data layout that a kernel expects.
+// It consists of two KernelSideFormat's, one for LHS and one for RHS.
+template <typename tLhs, typename tRhs>
+struct KernelFormat {
+  typedef tLhs Lhs;
+  typedef tRhs Rhs;
+
+  static_assert(Lhs::Cell::kDepth == Rhs::Cell::kDepth, "");
+  static const int kDepth = Lhs::Cell::kDepth;
+  static const int kRows = Lhs::Cell::kWidth * Lhs::kCells;
+  static const int kCols = Rhs::Cell::kWidth * Rhs::kCells;
+};
+
+inline const char* CellOrderName(CellOrder o) {
+  switch (o) {
+    case CellOrder::DepthMajor:
+      return "DepthMajor";
+    case CellOrder::WidthMajor:
+      return "WidthMajor";
+    case CellOrder::Diagonal:
+      return "Diagonal";
+    default:
+      assert(false);
+      return nullptr;
+  }
+}
+
+// Returns the offset into a cell, at which a given coefficient is stored.
+template <typename CellFormat>
+inline int OffsetIntoCell(int w, int d) {
+  switch (CellFormat::kOrder) {
+    case CellOrder::DepthMajor:
+      return w + d * CellFormat::kWidth;
+    case CellOrder::WidthMajor:
+      return d + w * CellFormat::kDepth;
+    case CellOrder::Diagonal:
+      assert(CellFormat::kWidth == CellFormat::kDepth);
+      static const int size = CellFormat::kWidth;
+      return ((size + w - d) * size + d) % (size * size);
+    default:
+      assert(false);
+      return 0;
+  }
+}
+
+// KernelBase is the virtual base class below all kernels.
+// The idea is that we don't need to templatize all our code on the exact
+// kernel type; we only need to templatize on kernel format. Kernels
+// sharing the same format can thus share the same packing/unpacking code.
+struct KernelBase {
+  virtual const char* Name() const = 0;
+
+  // This is the kernel implementation. We use the word 'run' consistently
+  // throughout gemmlowp to mean an inner loop, the implementation of which
+  // is to be provided by a separate optimized function.
+  virtual void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+                   std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+                   const std::uint8_t* rhs_ptr, std::size_t start_depth,
+                   std::size_t run_depth) const = 0;
+
+  virtual ~KernelBase() {}
+};
+
+template <typename KernelScalarType>
+struct ZeroPointInputValue {};
+
+template <>
+struct ZeroPointInputValue<std::uint8_t> {
+  static constexpr std::uint8_t kValue = 0;
+};
+
+template <>
+struct ZeroPointInputValue<std::int8_t> {
+  static constexpr std::uint8_t kValue = 128;
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_KERNEL_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_default.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_default.h
new file mode 100644
index 0000000..7ed55b8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_default.h
@@ -0,0 +1,109 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// kernel_default.h: Chooses default GEMM and GEMV kernels for the
+// host platform.
+
+#ifndef GEMMLOWP_INTERNAL_KERNEL_DEFAULT_H_
+#define GEMMLOWP_INTERNAL_KERNEL_DEFAULT_H_
+
+#ifndef GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
+#define GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
+#endif
+
+#include "../public/bit_depth.h"
+#include "common.h"
+#include "kernel_reference.h"
+
+namespace gemmlowp {
+
+template <bool MaxProductIsLessThan4096,
+          bool LhsAlwaysNonzero>
+struct DefaultKernelImpl {};
+
+// Partial specialization implementing the logic that if we want to use
+// a kernel for LhsAlwaysNonzero but do not have such a kernel, then we fall
+// back to a generic kernel not taking advantage of LhsAlwaysNonzero.
+template <bool LhsAlwaysNonzero>
+struct DefaultKernelImpl<true, LhsAlwaysNonzero>
+    : DefaultKernelImpl<false, LhsAlwaysNonzero> {};
+
+// Partial specialization implementing the logic that if we want to use
+// a kernel for MaxProductIsLessThan4096 but do not have such a kernel, then we
+// fall back to a generic kernel not taking advantage of
+// MaxProductIsLessThan4096.
+template <bool MaxProductIsLessThan4096>
+struct DefaultKernelImpl<MaxProductIsLessThan4096, true>
+    : DefaultKernelImpl<MaxProductIsLessThan4096, false> {};
+
+template <typename BitDepthParams>
+struct DefaultKernel
+    : DefaultKernelImpl<(BitDepthParams::LhsRange::kMaxValue *
+                             BitDepthParams::RhsRange::kMaxValue <
+                         4096),
+                        (BitDepthParams::LhsRange::kMinValue > 0)> {};
+
+}  // end namespace gemmlowp
+
+#define GEMMLOWP_SET_DEFAULT_KERNEL(MaxProductIsLessThan4096, \
+                                    LhsAlwaysNonzero, Kernel) \
+  namespace gemmlowp {                                        \
+  template <>                                                 \
+  struct DefaultKernelImpl<MaxProductIsLessThan4096,          \
+                           LhsAlwaysNonzero> : Kernel {};     \
+  }
+
+#if defined GEMMLOWP_NEON_32
+#include "kernel_neon.h"
+GEMMLOWP_SET_DEFAULT_KERNEL(false, false, NEON_32_Kernel12x4Depth2)
+GEMMLOWP_SET_DEFAULT_KERNEL(true, false,
+                            NEON_32_Kernel12x4Depth2Assuming12BitProducts)
+GEMMLOWP_SET_DEFAULT_KERNEL(false, true,
+                            NEON_32bit_GEMM_Int8Operands_LhsNonzero)
+#elif defined GEMMLOWP_NEON_64
+#include "kernel_neon.h"
+GEMMLOWP_SET_DEFAULT_KERNEL(false, false, NEON_64_Kernel12x8Depth2)
+GEMMLOWP_SET_DEFAULT_KERNEL(false, true,
+                            NEON_64bit_GEMM_Int8Operands_LhsNonzero)
+#elif defined GEMMLOWP_SSE4_32
+#include "kernel_sse.h"
+GEMMLOWP_SET_DEFAULT_KERNEL(false, false, SSE4_32_Kernel4x4Depth2)
+#elif defined GEMMLOWP_SSE4_64
+#include "kernel_sse.h"
+GEMMLOWP_SET_DEFAULT_KERNEL(false, false, SSE4_64_Kernel12x4Depth2)
+#else
+#ifndef GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK
+#if defined __ARM_ARCH_5TE__
+// SIMD is not available on this platform. The slow fallback will be used.
+// Don't require GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK because there's nothing
+// the user can do about it.
+#else
+#error \
+    "SIMD not enabled, you'd be getting a slow software fallback. Consider \
+enabling SIMD extensions (for example using -msse4 if you're on modern x86). \
+If that's not an option, and you would like to continue with the \
+slow fallback, define GEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK."
+#endif
+#endif
+#include "kernel_reference.h"
+namespace gemmlowp {
+typedef ReferenceKernel<KernelFormat<
+    KernelSideFormat<CellFormat<4, 16, CellOrder::WidthMajor>, 1>,
+    KernelSideFormat<CellFormat<4, 16, CellOrder::WidthMajor>, 1> > >
+    DefaultReferenceKernel;
+}
+GEMMLOWP_SET_DEFAULT_KERNEL(false, false, DefaultReferenceKernel)
+#endif
+
+#endif  // GEMMLOWP_INTERNAL_KERNEL_DEFAULT_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_neon.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_neon.h
new file mode 100644
index 0000000..5c253ba
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_neon.h
@@ -0,0 +1,1619 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// kernel_neon.h: a collection of NEON optimized kernels.
+// Check in kernel_default.h which one(s) are actually used by default.
+// Others are mere experiments; they are still covered by tests
+// in case they might be useful some day.
+
+#ifndef GEMMLOWP_INTERNAL_KERNEL_NEON_H_
+#define GEMMLOWP_INTERNAL_KERNEL_NEON_H_
+
+#include "kernel.h"
+
+#include <arm_neon.h>
+#include <cassert>
+
+namespace gemmlowp {
+
+// The kernels here are specifically arm 32bit assembly, not arm 64bit.
+#ifdef GEMMLOWP_NEON_32
+
+// Our main GEMM kernel.
+struct NEON_32_Kernel12x4Depth2 : KernelBase {
+  typedef KernelFormat<KernelSideFormat<CellFormat<4, 2>, 3>,
+                       KernelSideFormat<CellFormat<4, 2>, 1> >
+      Format;
+
+  const char* Name() const override { return "NEON, 12x4, depth 2"; }
+
+  // TODO(benoitjacob): reorder function arguments so dst comes last
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    ScopedProfilingLabel label("optimized kernel (NEON 12x4)");
+
+// For iOS assembler, the %= style of local labels cause compilation errors,
+//  so use numerical ones instead. See
+// http://stackoverflow.com/questions/3898435/labels-in-gcc-inline-assembly
+// If you add any labels, remember to undef them at the end.
+#define GEMMLOWP_LABEL_CLEAR_ACCUMULATORS "1"
+#define GEMMLOWP_LABEL_BEFORE_LOOP "2"
+#define GEMMLOWP_LABEL_LOOP "3"
+#define GEMMLOWP_LABEL_AFTER_LOOP "4"
+
+    assert(dst_row_stride == 1);
+    asm volatile(
+        // Overview of register layout:
+        //
+        // A 2x4 cell of Rhs is stored in 16bit in d0--d1 (q0).
+        // A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in d2--d7
+        // (q1--q3).
+        // A 12x4 block of accumulators is stored in 32bit in q4--q15.
+        //
+        //                   +-----+-----+-----+-----+
+        //                   |d0[0]|d0[1]|d0[2]|d0[3]|
+        //              Rhs  +-----+-----+-----+-----+
+        //                   |d1[0]|d1[1]|d1[2]|d1[3]|
+        //                   +-----+-----+-----+-----+
+        //
+        //                   |     |     |     |     |
+        //
+        //    Lhs            |     |     |     |     |
+        //
+        //  +--+--+ - - - -  +-----+-----+-----+-----+
+        //  |d2|d3|          | q4  | q5  | q6  | q7  |
+        //  |d2|d3|          | q4  | q5  | q6  | q7  |
+        //  |d2|d3|          | q4  | q5  | q6  | q7  |
+        //  |d2|d3|          | q4  | q5  | q6  | q7  |
+        //  +--+--+ - - - -  +-----+-----+-----+-----+
+        //  |d4|d5|          | q8  | q9  | q10 | q11 |
+        //  |d4|d5|          | q8  | q9  | q10 | q11 |
+        //  |d4|d5|          | q8  | q9  | q10 | q11 |
+        //  |d4|d5|          | q8  | q9  | q10 | q11 |
+        //  +--+--+ - - - -  +-----+-----+-----+-----+
+        //  |d6|d7|          | q12 | q13 | q14 | q15 |
+        //  |d6|d7|          | q12 | q13 | q14 | q15 |
+        //  |d6|d7|          | q12 | q13 | q14 | q15 |
+        //  |d6|d7|          | q12 | q13 | q14 | q15 |
+        //  +--+--+ - - - -  +-----+-----+-----+-----+
+        //
+        //                            Accumulator
+
+        // Load 1 Rhs cell of size 2x4
+        "vld1.8 {d0}, [%[rhs_ptr]]!\n"
+        // Load 3 Lhs cells of size 4x2 each
+        "vld1.8 {d2}, [%[lhs_ptr]]!\n"
+        "vld1.8 {d4}, [%[lhs_ptr]]!\n"
+        "vld1.8 {d6}, [%[lhs_ptr]]!\n"
+
+        // Check if start_depth==0 to decide whether we will clear
+        // accumulators or load existing accumulators.
+        "cmp %[start_depth], #0\n"
+
+        // Multiply dst_col_stride by 4 == sizeof(int32) to use
+        // it as a byte offset below.
+        "lsl %[dst_col_stride], #2\n"
+
+        "beq " GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+        "f\n"
+
+        // Load accumulators (start_depth != 0)
+        "mov r1, %[dst_ptr]\n"
+        "subs %[run_depth], #2\n"
+        "mov r0, r1\n"
+        "vld1.32 {d8, d9},   [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vld1.32 {d16, d17}, [r0]!\n"
+        "vld1.32 {d24, d25}, [r0]\n"
+        "mov r0, r1\n"
+        "vld1.32 {d10, d11}, [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vld1.32 {d18, d19}, [r0]!\n"
+        "vld1.32 {d26, d27}, [r0]\n"
+        "mov r0, r1\n"
+        "vld1.32 {d12, d13}, [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vld1.32 {d20, d21}, [r0]!\n"
+        "vld1.32 {d28, d29}, [r0]\n"
+        "mov r0, r1\n"
+        "vld1.32 {d14, d15}, [r0]!\n"
+        "vld1.32 {d22, d23}, [r0]!\n"
+        "vld1.32 {d30, d31}, [r0]\n"
+
+        "b " GEMMLOWP_LABEL_BEFORE_LOOP "f\n"
+
+        GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+        ":\n"
+
+        // Clear accumulators (start_depth == 0)
+        "vmov.s32 q4, #0\n"
+        "subs %[run_depth], #2\n"
+        "vmov.s32 q8, q4\n"
+        "vmov.s32 q12, q4\n"
+        "vmov.s32 q5, q4\n"
+        "vmov.s32 q9, q4\n"
+        "vmov.s32 q13, q4\n"
+        "vmov.s32 q6, q4\n"
+        "vmov.s32 q10, q4\n"
+        "vmov.s32 q14, q4\n"
+        "vmov.s32 q7, q4\n"
+        "vmov.s32 q11, q4\n"
+        "vmov.s32 q15, q4\n"
+
+        GEMMLOWP_LABEL_BEFORE_LOOP
+        ":\n"
+
+        // If there are only two levels of depth, skip the loop.
+        "beq " GEMMLOWP_LABEL_AFTER_LOOP "f\n"
+
+        GEMMLOWP_LABEL_LOOP
+        ":\n"
+        // Expand Lhs/Rhs cells to 16 bit.
+        // Note: moving theses vmovls further down to allow for
+        // longer data pipelining helps a little on A57 but is
+        // harmful on A53 --- It looks as if A53 doesn't like
+        // interleaving vmovl's into the vmlal's.
+        "vmovl.u8 q0, d0\n"
+        "vmovl.u8 q1, d2\n"
+        "vmovl.u8 q2, d4\n"
+        "vmovl.u8 q3, d6\n"
+
+        // Multiply-accumulate, level of depth 0
+        "vmlal.u16 q4, d2, d0[0]\n"
+        "vmlal.u16 q5, d2, d0[1]\n"
+        "vmlal.u16 q6, d2, d0[2]\n"
+        "vmlal.u16 q7, d2, d0[3]\n"
+        "vldr d2, [%[lhs_ptr]]\n"
+        "vmlal.u16 q8, d4, d0[0]\n"
+        "vmlal.u16 q9, d4, d0[1]\n"
+        "vmlal.u16 q10, d4, d0[2]\n"
+        "vmlal.u16 q11, d4, d0[3]\n"
+        "vldr d4, [%[lhs_ptr], #8]\n"
+        "vmlal.u16 q12, d6, d0[0]\n"
+        "vmlal.u16 q13, d6, d0[1]\n"
+        "vmlal.u16 q14, d6, d0[2]\n"
+        "vmlal.u16 q15, d6, d0[3]\n"
+        "vldr d6, [%[lhs_ptr], #16]\n"
+        "vldr d0, [%[rhs_ptr]]\n"
+
+        // Multiply-accumulate, level of depth 1
+        "vmlal.u16 q4, d3, d1[0]\n"
+        "vmlal.u16 q5, d3, d1[1]\n"
+        "add %[lhs_ptr], #24\n"
+        "vmlal.u16 q6, d3, d1[2]\n"
+        "vmlal.u16 q7, d3, d1[3]\n"
+        "add %[rhs_ptr], #8\n"
+        "vmlal.u16 q8, d5, d1[0]\n"
+        "vmlal.u16 q9, d5, d1[1]\n"
+        "subs %[run_depth], #2\n"
+        "vmlal.u16 q10, d5, d1[2]\n"
+        "vmlal.u16 q11, d5, d1[3]\n"
+        "vmlal.u16 q12, d7, d1[0]\n"
+        "vmlal.u16 q13, d7, d1[1]\n"
+        "vmlal.u16 q14, d7, d1[2]\n"
+        "vmlal.u16 q15, d7, d1[3]\n"
+
+        "bne " GEMMLOWP_LABEL_LOOP "b\n"
+
+        GEMMLOWP_LABEL_AFTER_LOOP
+        ":\n"
+
+        // Do remaining arithmetic for the last 2 levels of depth.
+
+        // Expand Lhs/Rhs cells to 16 bit.
+        "vmovl.u8 q0, d0\n"
+        "vmovl.u8 q1, d2\n"
+        "vmovl.u8 q2, d4\n"
+        "vmovl.u8 q3, d6\n"
+
+        // Multiply-accumulate, level of depth 0
+        "vmlal.u16 q4, d2, d0[0]\n"
+        "vmlal.u16 q5, d2, d0[1]\n"
+        "vmlal.u16 q6, d2, d0[2]\n"
+        "vmlal.u16 q7, d2, d0[3]\n"
+        "vmlal.u16 q8, d4, d0[0]\n"
+        "vmlal.u16 q9, d4, d0[1]\n"
+        "vmlal.u16 q10, d4, d0[2]\n"
+        "vmlal.u16 q11, d4, d0[3]\n"
+        "vmlal.u16 q12, d6, d0[0]\n"
+        "vmlal.u16 q13, d6, d0[1]\n"
+        "vmlal.u16 q14, d6, d0[2]\n"
+        "vmlal.u16 q15, d6, d0[3]\n"
+
+        // Multiply-accumulate, level of depth 1
+        "vmlal.u16 q4, d3, d1[0]\n"
+        "vmlal.u16 q5, d3, d1[1]\n"
+        "vmlal.u16 q6, d3, d1[2]\n"
+        "vmlal.u16 q7, d3, d1[3]\n"
+        "vmlal.u16 q8, d5, d1[0]\n"
+        "vmlal.u16 q9, d5, d1[1]\n"
+        "vmlal.u16 q10, d5, d1[2]\n"
+        "vmlal.u16 q11, d5, d1[3]\n"
+        "vmlal.u16 q12, d7, d1[0]\n"
+        "vmlal.u16 q13, d7, d1[1]\n"
+        "vmlal.u16 q14, d7, d1[2]\n"
+        "vmlal.u16 q15, d7, d1[3]\n"
+
+        // Store accumulators
+        "mov r1, %[dst_ptr]\n"
+        "mov r0, r1\n"
+        "vst1.32 {d8, d9},   [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vst1.32 {d16, d17}, [r0]!\n"
+        "vst1.32 {d24, d25}, [r0]\n"
+        "mov r0, r1\n"
+        "vst1.32 {d10, d11}, [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vst1.32 {d18, d19}, [r0]!\n"
+        "vst1.32 {d26, d27}, [r0]\n"
+        "mov r0, r1\n"
+        "vst1.32 {d12, d13}, [r0]!\n"
+        "add r1, %[dst_col_stride]\n"
+        "vst1.32 {d20, d21}, [r0]!\n"
+        "vst1.32 {d28, d29}, [r0]\n"
+        "mov r0, r1\n"
+        "vst1.32 {d14, d15}, [r0]!\n"
+        "vst1.32 {d22, d23}, [r0]!\n"
+        "vst1.32 {d30, d31}, [r0]\n"
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr),
+        [run_depth] "+r"(run_depth)
+        :  // inputs
+        [start_depth] "r"(start_depth),
+        [dst_col_stride] "r"(dst_col_stride)
+        :  // clobbers
+        "cc", "memory", "r0", "r1",
+        // note: someone on internet says that quad registers are
+        // unsupported in the clobber list!
+        "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d8", "d9", "d10",
+        "d11", "d12", "d13", "d14", "d15", "d16", "d17", "d18", "d19", "d20",
+        "d21", "d22", "d23", "d24", "d25", "d26", "d27", "d28", "d29", "d30",
+        "d31");
+#undef GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+#undef GEMMLOWP_LABEL_BEFORE_LOOP
+#undef GEMMLOWP_LABEL_LOOP
+#undef GEMMLOWP_LABEL_AFTER_LOOP
+  }
+};
+
+struct NEON_32_Kernel12x4Depth2Assuming12BitProducts : KernelBase {
+  typedef KernelFormat<
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 3>,
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 1> >
+      Format;
+
+  const char* Name() const override {
+    return "NEON, 12x4, depth 2, assuming 12-bit products";
+  }
+
+  // TODO(benoitjacob): reorder function arguments so dst comes last
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    ScopedProfilingLabel label(
+        "optimized kernel (NEON 12x4, assuming 12-bit products)");
+    assert(dst_row_stride == 1);
+
+// See comments above for why we need local numerical labels in our asm.
+#define GEMMLOWP_LOOP_NEON_32_KERNEL_12X4_DEPTH2_ASSUMING_12BIT_PRODUCTS "1"
+#define GEMMLOWP_LOAD_GLOBAL_ACCUMULATORS_NEON_32_KERNEL_12X4_DEPTH2_12BIT "2"
+#define GEMMLOWP_LABEL_32 "3"
+#define GEMMLOWP_LABEL_24 "4"
+#define GEMMLOWP_LABEL_16 "5"
+#define GEMMLOWP_LABEL_8 "6"
+#define GEMMLOWP_LABEL_2 "7"
+
+    // This kernel is special in that it uses local 16-bit accumulators.
+    // Because it assumes that each product fits in 12 bits, it can accumulate
+    // 16 products into a local 16-bit accumulator without risking overflow.
+    // At that point, it must accumulate these local 16-bit accumulators back
+    // into global 32-bit accumulators, which have to be stored in memory for
+    // lack of register space.
+    // This 12x4 block of global accumulators is laid out as 3 cells of size 4x4
+    // stored in diagonal-major order like this for the first 4x4 cell:
+    //
+    //   0   4   8  12
+    //  13   1   5   9
+    //  10  14   2   6
+    //   7  11  15   3
+    //
+    // and likewise for the 2nd  cell (16--31) and 3rd cell (32--47)
+    std::int32_t global_accumulators[3 * 4 * 4];
+    asm volatile(
+        // Compute stride between consecutive columns, in bytes
+        "mov r0, #4\n"  // multiply by 4 = sizeof(int32)
+        "mul %[dst_col_stride], r0\n"
+
+        "cmp %[start_depth], #0\n"
+        "bne"
+        " " GEMMLOWP_LOAD_GLOBAL_ACCUMULATORS_NEON_32_KERNEL_12X4_DEPTH2_12BIT
+        "f\n"
+
+        // If start_depth==0, we need to clear our global accumulators
+        "mov r0, %[global_accumulators]\n"
+        "vmov.s32 q8, #0\n"
+        "vmov.s32 q9, q8\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "vst1.32 {d16,d17,d18,d19}, [r0]!\n"
+        "b " GEMMLOWP_LOOP_NEON_32_KERNEL_12X4_DEPTH2_ASSUMING_12BIT_PRODUCTS
+        "f\n"
+
+        // If start_depth!=0, we need to load our existing global accumulators
+        GEMMLOWP_LOAD_GLOBAL_ACCUMULATORS_NEON_32_KERNEL_12X4_DEPTH2_12BIT
+        ":\n"
+        // Load global accumulators from destination matrix, column-major
+        "mov r1, %[dst_ptr]\n"
+        "mov r0, %[dst_col_stride]\n"
+        "sub r0, #32\n"
+        "vld1.32 {d0,d1}, [r1]!\n"
+        "vld1.32 {d8,d9}, [r1]!\n"
+        "vld1.32 {d16,d17}, [r1], r0\n"
+        "vld1.32 {d2,d3}, [r1]!\n"
+        "vld1.32 {d10,d11}, [r1]!\n"
+        "vld1.32 {d18,d19}, [r1], r0\n"
+        "vld1.32 {d4,d5}, [r1]!\n"
+        "vld1.32 {d12,d13}, [r1]!\n"
+        "vld1.32 {d20,d21}, [r1], r0\n"
+        "vld1.32 {d6,d7}, [r1]!\n"
+        "vld1.32 {d14,d15}, [r1]!\n"
+        "vld1.32 {d22,d23}, [r1], r0\n"
+        // Now we need to convert the global accumulator registers to
+        // 4x4-block-wise diagonal-major order. What we effectively want to do
+        // is to rotate the rows, however the accumulators are stored in
+        // column-major order in registers. So we achieve this by
+        // transposing, rotating the registers, and transposing again each
+        // 4x4 block.
+        //
+        // Transpose 3 4x4 blocks separately
+        "vtrn.32 q0, q1\n"
+        "vtrn.32 q2, q3\n"
+        "vswp d1, d4\n"
+        "vswp d3, d6\n"
+        "vtrn.32 q4, q5\n"
+        "vtrn.32 q6, q7\n"
+        "vswp d9, d12\n"
+        "vswp d11, d14\n"
+        "vtrn.32 q8, q9\n"
+        "vtrn.32 q10, q11\n"
+        "vswp d17, d20\n"
+        "vswp d19, d22\n"
+        // Rotate the registers
+        "vext.32 q1, q1, q1, #1\n"
+        "vext.32 q2, q2, q2, #2\n"
+        "vext.32 q3, q3, q3, #3\n"
+        "vext.32 q5, q5, q5, #1\n"
+        "vext.32 q6, q6, q6, #2\n"
+        "vext.32 q7, q7, q7, #3\n"
+        "vext.32 q9, q9, q9, #1\n"
+        "vext.32 q10, q10, q10, #2\n"
+        "vext.32 q11, q11, q11, #3\n"
+        // Transpose again and store into our global accumulators
+        // buffer. These two operations are done at once using vst4.
+        "mov r0, %[global_accumulators]\n"
+        "vst4.32 {d0,d2,d4,d6}, [r0]!\n"
+        "vst4.32 {d1,d3,d5,d7}, [r0]!\n"
+        "vst4.32 {d8,d10,d12,d14}, [r0]!\n"
+        "vst4.32 {d9,d11,d13,d15}, [r0]!\n"
+        "vst4.32 {d16,d18,d20,d22}, [r0]!\n"
+        "vst4.32 {d17,d19,d21,d23}, [r0]!\n"
+
+        /* Main loop */
+
+        GEMMLOWP_LOOP_NEON_32_KERNEL_12X4_DEPTH2_ASSUMING_12BIT_PRODUCTS
+        ":\n"
+
+// Overview of register layout:
+//
+// Registers q4--q16 are the local 16-bit accumulators.
+// However, each entry in the result matrix is represented
+// by *two* local 16-bit accumulators: one for even levels
+// of depth and one for odd levels of depth. These correspond
+// to the scalars at even and odd indices within each q-register.
+// Thus we effectively use 32 bits of register space for each
+// entry in the result matrix. The accumulators register layout
+// is the same as was described above for the global 32-bit
+// accumulators (3 cells of size 4x4 in diagonal-major order)
+// with the only difference that instead of 32bit values we have
+// pairs of 16bit values.
+//
+// A 2x4 cell of Rhs is stored in 8bit in d0.
+// A 12x2 block of 3 4x2 cells Lhs is stored in 8bit in d1--d3.
+//
+//                      +--------+--------+--------+--------+
+//                      |d0[0]   |d0[2]   |d0[4]   |d0[6]   |
+//                 Rhs  +--------+--------+--------+--------+
+//                      |d0[1]   |d0[3]   |d0[5]   |d0[7]   |
+//                      +--------+--------+--------+--------+
+//
+//                      |        |        |        |        |
+//
+//    Lhs               |        |        |        |        |
+//
+//  +-----+-----+ - - - +--------+--------+--------+--------+
+//  |d1[0]|d1[1]|       |q4[0,1] |q5[0,1] |q6[0,1] |q7[0,1] |
+//  |d1[2]|d1[3]|       |q7[2,3] |q4[2,3] |q5[2,3] |q6[2,3] |
+//  |d1[4]|d1[5]|       |q6[4,5] |q7[4,5] |q4[4,5] |q5[4,5] |
+//  |d1[6]|d1[7]|       |q5[6,7] |q6[6,7] |q7[6,7] |q4[6,7] |
+//  +-----+-----+ - - - +--------+--------+--------+--------+
+//  |d2[0]|d2[1]|       |q8[0,1] |q8[0,1] |q8[0,1] |q8[0,1] |
+//  |d2[2]|d2[3]|       |q9[2,3] |q9[2,3] |q9[2,3] |q9[2,3] |
+//  |d2[4]|d2[5]|       |q10[4,5]|q10[4,5]|q10[4,5]|q10[4,5]|
+//  |d2[6]|d2[7]|       |q11[6,7]|q11[6,7]|q11[6,7]|q11[6,7]|
+//  +-----+-----+ - - - +--------+--------+--------+--------+
+//  |d3[0]|d3[1]|       |q12[0,1]|q12[0,1]|q12[0,1]|q12[0,1]|
+//  |d3[2]|d3[3]|       |q13[2,3]|q13[2,3]|q13[2,3]|q13[2,3]|
+//  |d3[4]|d3[5]|       |q14[4,5]|q14[4,5]|q14[4,5]|q14[4,5]|
+//  |d3[6]|d3[7]|       |q15[6,7]|q15[6,7]|q15[6,7]|q15[6,7]|
+//  +-----+-----+ - - - +--------+--------+--------+--------+
+//
+//                            Local 16-bit accumulators
+//                         Note: 2 scalars per matrix entry
+
+#define GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH \
+  /* Load 3 Lhs cells of size 4x2 */          \
+  "vld1.8 {d1,d2,d3}, [%[lhs_ptr]:64]!\n"     \
+                                              \
+  /* Load 1 Rhs cell of size 2x4 */           \
+  "vld1.8 {d0}, [%[rhs_ptr]:64]!\n"           \
+                                              \
+  /* Multiply-accumulate */                   \
+  "vmlal.u8 q4, d1, d0\n"                     \
+  "vmlal.u8 q8, d2, d0\n"                     \
+  "vmlal.u8 q12, d3, d0\n"                    \
+  "vext.8 d0, d0, d0, #2\n"                   \
+  "vmlal.u8 q5, d1, d0\n"                     \
+  "vmlal.u8 q9, d2, d0\n"                     \
+  "vmlal.u8 q13, d3, d0\n"                    \
+  "vext.8 d0, d0, d0, #2\n"                   \
+  "vmlal.u8 q6, d1, d0\n"                     \
+  "vmlal.u8 q10, d2, d0\n"                    \
+  "vmlal.u8 q14, d3, d0\n"                    \
+  "vext.8 d0, d0, d0, #2\n"                   \
+  "vmlal.u8 q7, d1, d0\n"                     \
+  "vmlal.u8 q11, d2, d0\n"                    \
+  "vmlal.u8 q15, d3, d0\n"                    \
+                                              \
+  "sub %[run_depth], #2\n"
+
+#define GEMMLOWP_ACCUMULATE_8_LEVELS_OF_DEPTH \
+  GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH       \
+  GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH       \
+  GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH       \
+  GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH
+
+        // Clear local 16-bit accumulators
+        "vmov.s32 q4, #0\n"
+        "vmov.s32 q5, q4\n"
+        "vmov.s32 q6, q4\n"
+        "vmov.s32 q7, q4\n"
+        "vmov.s32 q8, q4\n"
+        "vmov.s32 q9, q4\n"
+        "vmov.s32 q10, q4\n"
+        "vmov.s32 q11, q4\n"
+        "vmov.s32 q12, q4\n"
+        "vmov.s32 q13, q4\n"
+        "vmov.s32 q14, q4\n"
+        "vmov.s32 q15, q4\n"
+
+        // Select a suitable number of depth levels
+        // to process at this iteration. TODO (benoitjacob) I guess that
+        // someone who really knows asm should make this a jump table.
+        "cmp %[run_depth], #32\n"
+        "bge " GEMMLOWP_LABEL_32
+        "f\n"
+        "cmp %[run_depth], #24\n"
+        "bge " GEMMLOWP_LABEL_24
+        "f\n"
+        "cmp %[run_depth], #16\n"
+        "bge " GEMMLOWP_LABEL_16
+        "f\n"
+        "cmp %[run_depth], #8\n"
+        "bge " GEMMLOWP_LABEL_8
+        "f\n"
+        "b " GEMMLOWP_LABEL_2 "f\n"
+
+        GEMMLOWP_LABEL_32
+        ":\n" GEMMLOWP_ACCUMULATE_8_LEVELS_OF_DEPTH GEMMLOWP_LABEL_24
+        ":\n" GEMMLOWP_ACCUMULATE_8_LEVELS_OF_DEPTH GEMMLOWP_LABEL_16
+        ":\n" GEMMLOWP_ACCUMULATE_8_LEVELS_OF_DEPTH GEMMLOWP_LABEL_8
+        ":\n" GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH
+            GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH
+                GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH GEMMLOWP_LABEL_2
+        ":\n" GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH
+
+        // Accumulate the local accumulators into the global accumulators.
+        // This is about summing adjacent pairs of 16-bit scalars into
+        // single 32-bit scalars, so we use pairwise long addition (vpadal).
+        "mov r0, %[global_accumulators]\n"
+        "mov r1, %[global_accumulators]\n"
+        "vld1.32 {d0,d1,d2,d3}, [r0]!\n"
+        "vld1.32 {d4,d5,d6,d7}, [r0]!\n"
+        "vpadal.u16 q0, q4\n"
+        "vpadal.u16 q1, q5\n"
+        "vpadal.u16 q2, q6\n"
+        "vpadal.u16 q3, q7\n"
+        "vst1.32 {d0,d1,d2,d3}, [r1]!\n"
+        "vst1.32 {d4,d5,d6,d7}, [r1]!\n"
+        "vld1.32 {d0,d1,d2,d3}, [r0]!\n"
+        "vld1.32 {d4,d5,d6,d7}, [r0]!\n"
+        "vpadal.u16 q0, q8\n"
+        "vpadal.u16 q1, q9\n"
+        "vpadal.u16 q2, q10\n"
+        "vpadal.u16 q3, q11\n"
+        "vst1.32 {d0,d1,d2,d3}, [r1]!\n"
+        "vst1.32 {d4,d5,d6,d7}, [r1]!\n"
+        "vld1.32 {d0,d1,d2,d3}, [r0]!\n"
+        "vld1.32 {d4,d5,d6,d7}, [r0]!\n"
+        "vpadal.u16 q0, q12\n"
+        "vpadal.u16 q1, q13\n"
+        "vpadal.u16 q2, q14\n"
+        "vpadal.u16 q3, q15\n"
+        "vst1.32 {d0,d1,d2,d3}, [r1]!\n"
+        "vst1.32 {d4,d5,d6,d7}, [r1]!\n"
+
+        // Loop.
+        "cmp %[run_depth], #0\n"
+        "bne " GEMMLOWP_LOOP_NEON_32_KERNEL_12X4_DEPTH2_ASSUMING_12BIT_PRODUCTS
+        "b\n"
+
+#undef GEMMLOWP_CLEAR_LOCAL_ACCUMULATORS
+#undef GEMMLOWP_ACCUMULATE_8_LEVELS_OF_DEPTH
+#undef GEMMLOWP_ACCUMULATE_2_LEVELS_OF_DEPTH
+#undef GEMMLOWP_ADD_TO_GLOBAL_ACCUMULATORS
+
+        /* end of main loop */
+
+        // Store the global accumulators to the destination matrix
+        // (column-major)
+        // This is the reverse of the steps that we followed at the beginning
+        // when we load the global accumulators from the destination matrix.
+        // The problem is the same: how to convert 4x4 blocks
+        // between column-major and diagonal-major orders.
+        // Like above, we do this by rotating rows, and we achieve that by
+        // tranposing, rotating columns, and transposing again.
+        //
+        // Load and transpose 4x4 blocks of global accumulators
+        // These two steps are done at once by the vld4 instruction.
+        "mov r0, %[global_accumulators]\n"
+        "vld4.32 {d0,d2,d4,d6}, [r0]!\n"
+        "vld4.32 {d1,d3,d5,d7}, [r0]!\n"
+        "vld4.32 {d8,d10,d12,d14}, [r0]!\n"
+        "vld4.32 {d9,d11,d13,d15}, [r0]!\n"
+        "vld4.32 {d16,d18,d20,d22}, [r0]!\n"
+        "vld4.32 {d17,d19,d21,d23}, [r0]!\n"
+        // Rotate the rows of each 4x4 block
+        "vext.32 q1, q1, q1, #3\n"
+        "vext.32 q2, q2, q2, #2\n"
+        "vext.32 q3, q3, q3, #1\n"
+        "vext.32 q5, q5, q5, #3\n"
+        "vext.32 q6, q6, q6, #2\n"
+        "vext.32 q7, q7, q7, #1\n"
+        "vext.32 q9, q9, q9, #3\n"
+        "vext.32 q10, q10, q10, #2\n"
+        "vext.32 q11, q11, q11, #1\n"
+        // Transpose again each 4x4 block
+        "vtrn.32 q0, q1\n"
+        "vtrn.32 q2, q3\n"
+        "vswp d1, d4\n"
+        "vswp d3, d6\n"
+        "vtrn.32 q4, q5\n"
+        "vtrn.32 q6, q7\n"
+        "vswp d9, d12\n"
+        "vswp d11, d14\n"
+        "vtrn.32 q8, q9\n"
+        "vtrn.32 q10, q11\n"
+        "vswp d17, d20\n"
+        "vswp d19, d22\n"
+        // Store into the column-major destination matrix
+        "mov r1, %[dst_ptr]\n"
+        "mov r0, %[dst_col_stride]\n"
+        "sub r0, #32\n"
+        "vst1.32 {d0,d1}, [r1]!\n"
+        "vst1.32 {d8,d9}, [r1]!\n"
+        "vst1.32 {d16,d17}, [r1], r0\n"
+        "vst1.32 {d2,d3}, [r1]!\n"
+        "vst1.32 {d10,d11}, [r1]!\n"
+        "vst1.32 {d18,d19}, [r1], r0\n"
+        "vst1.32 {d4,d5}, [r1]!\n"
+        "vst1.32 {d12,d13}, [r1]!\n"
+        "vst1.32 {d20,d21}, [r1], r0\n"
+        "vst1.32 {d6,d7}, [r1]!\n"
+        "vst1.32 {d14,d15}, [r1]!\n"
+        "vst1.32 {d22,d23}, [r1], r0\n"
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr),
+        [run_depth] "+r"(run_depth)
+        :  // inputs
+        [start_depth] "r"(start_depth), [dst_col_stride] "r"(dst_col_stride),
+        [global_accumulators] "r"(&global_accumulators[0])
+        :  // clobbers
+        "cc", "memory", "r0", "r1",
+        // note: someone on internet says that quad registers are
+        // unsupported in the clobber list!
+        "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7", "d8", "d9", "d10",
+        "d11", "d12", "d13", "d14", "d15", "d16", "d17", "d18", "d19", "d20",
+        "d21", "d22", "d23", "d24", "d25", "d26", "d27", "d28", "d29", "d30",
+        "d31");
+#undef GEMMLOWP_LOOP_NEON_32_KERNEL_12X4_DEPTH2_ASSUMING_12BIT_PRODUCTS
+#undef GEMMLOWP_LOAD_GLOBAL_ACCUMULATORS_NEON_32_KERNEL_12X4_DEPTH2_12BIT
+#undef GEMMLOWP_LABEL_32
+#undef GEMMLOWP_LABEL_24
+#undef GEMMLOWP_LABEL_16
+#undef GEMMLOWP_LABEL_8
+#undef GEMMLOWP_LABEL_2
+  }
+};
+
+struct NEON_32bit_GEMM_Int8Operands_LhsNonzero : KernelBase {
+  typedef KernelFormat<
+      KernelSideFormatInt8<CellFormat<4, 16, CellOrder::WidthMajor>, 1>,
+      KernelSideFormatInt8<CellFormat<2, 16, CellOrder::WidthMajor>, 1> >
+      Format;
+  const char* Name() const override {
+    return "NEON, 4x2, depth 16, accumulating two within signed int16";
+  }
+
+  // TODO(benoitjacob): reorder function arguments so dst comes last
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+#define GEMMLOWP_LABEL_AFTER_LOOP "1"
+#define GEMMLOWP_LABEL_LOOP "2"
+#define GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES "3"
+#define GEMMLOWP_LABEL_STORE "4"
+    asm volatile(
+        // Multiply dst_col_stride by 4 == sizeof(int32) to use
+        // it as a byte offset below.
+        "lsl %[dst_col_stride], %[dst_col_stride], #2\n"
+
+        // Overview of register layout:
+        //
+        // A 2x16 block of Rhs is stored in 8 bit in d0--d3.
+        // A 4x16 block of Lhs is stored in 8 bit in d4--d7. That is only
+        // half of the register space required, so we loop over these registers
+        // twice. Only half of it, a 2x16 block, is stored in d4--d7 at
+        // any given time.
+        //
+        // A 4x2 block of accumulators is stored in q8--q15 (as 4x32 bit
+        // components which need to be horizontally-added at the end)
+        //
+        // The Lhs vectors are multiplied by the Rhs vectors with a widening
+        // multiply over the 8 first levels of depth, producing int16x8
+        // vectors of products for each position in the accumulator matrix.
+        // Here comes the special trick: since the operands are signed int8,
+        // their range being [ -2^7 , 2^7 ), their products are in range
+        // [ -2^14 , 2^14 - 1 ), meaning that we can add two such values
+        // without any risk of overflowing int16.
+        // We thus proceed with the 8 next levels of depth, multiplying
+        // again Lhs by Rhs, accumulating into this existing int16x8 vector.
+        //
+        // Only then, having processed 16 levels of depth, do we need to
+        // horizontally add these int16x8 accumulators into the final
+        // int32x4 accumulators.
+        //
+        // As we do not have enough registers to store all 16 int16x8
+        // temporary-16bit-accumulators, we have them cycle through q4--q7.
+        //
+        //
+        // Register layout (ignoring the q4--q7 temporary 16bit accumulators):
+        //
+        //                               +----+----+
+        //                               | d0 | d2 |
+        //                               | .  | .  |
+        //                               | .  | .  |
+        //                               | .  | .  |
+        //                       Rhs     +----+----+
+        //                               | d1 | d3 |
+        //                               | .  | .  |
+        //                               | .  | .  |
+        //                               | .  | .  |
+        //                               +----+----+
+        //
+        //                               |    |    |
+        //
+        //    Lhs                        |    |    |
+        //
+        //  +--------+--------+ - - - -  +----+----+
+        //  | d4 ... | d5 ... |          | q8 | q9 |
+        //  | d6 ... | d7 ... |          | q10| q11|
+        //  | d4 ... | d5 ... |          | q12| q13|
+        //  | d6 ... | d7 ... |          | q14| q15|
+        //  +--------+--------+ - - - -  +----+----+
+        //
+        //                               Accumulator
+        //
+
+        // Clear accumulators, and, interleaved with it,
+        // initial loads of the first loop iteration,
+        // taken out of the loop so that in the loop itself we have
+        // optimal streaming of data from memory.
+        "vldr d0, [%[rhs_ptr], #0]\n"
+        "vmov.i32 q8, #0\n"
+        "vldr d4, [%[lhs_ptr], #0]\n"
+        "vmov.i32 q9, #0\n"
+        "vldr d2, [%[rhs_ptr], #16]\n"
+        "vmov.i32 q10, q8\n"
+        "vldr d6, [%[lhs_ptr], #16]\n"
+        "vmov.i32 q11, q8\n"
+        "vldr d1, [%[rhs_ptr], #8]\n"
+        "vmov.i32 q12, q8\n"
+        "vldr d5, [%[lhs_ptr], #8]\n"
+        "vmov.i32 q13, q8\n"
+        "vldr d3, [%[rhs_ptr], #24]\n"
+        "vmov.i32 q14, q8\n"
+        "vldr d7, [%[lhs_ptr], #24]\n"
+        "vmov.i32 q15, q8\n"
+
+        // General loop.
+        GEMMLOWP_LABEL_LOOP
+        ":\n"
+
+        // Multiply 8 first levels of depth.
+        "vmull.s8    q4,  d0,  d4\n"
+        "add %[rhs_ptr], %[rhs_ptr], #32\n"
+        "vmull.s8    q5,  d2,  d4\n"
+        "vldr d4, [%[lhs_ptr], #32]\n"
+        "vmull.s8    q6,  d0,  d6\n"
+        "vmull.s8    q7,  d2,  d6\n"
+        "vldr d6, [%[lhs_ptr], #48]\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "vmlal.s8    q4,  d1,  d5\n"
+        "vmlal.s8    q5,  d3,  d5\n"
+        "vldr d5, [%[lhs_ptr], #40]\n"
+        "vmlal.s8    q6,  d1,  d7\n"
+        "vmlal.s8    q7,  d3,  d7\n"
+        "vldr d7, [%[lhs_ptr], #56]\n"
+
+        // Add pairwise, accumulate into 32-bit accumulators.
+        "vpadal.s16   q8,  q4\n"
+        "add %[lhs_ptr], %[lhs_ptr], #64\n"
+        "vpadal.s16   q9,  q5\n"
+        "subs %[run_depth], %[run_depth], #16\n"
+        "vpadal.s16   q10, q6\n"
+        "vpadal.s16   q11, q7\n"
+
+        "beq " GEMMLOWP_LABEL_AFTER_LOOP
+        "f\n"
+
+        // Multiply first half.
+        "vmull.s8    q4,  d0,  d4\n"
+        "vmull.s8    q5,  d2,  d4\n"
+        "vldr d4, [%[lhs_ptr], #0]\n"
+        "vmull.s8    q6,  d0,  d6\n"
+        "vldr d0, [%[rhs_ptr], #0]\n"
+        "vmull.s8    q7,  d2,  d6\n"
+        "vldr d2, [%[rhs_ptr], #16]\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "vmlal.s8    q4,  d1,  d5\n"
+        "vldr d6, [%[lhs_ptr], #16]\n"
+        "vmlal.s8    q5,  d3,  d5\n"
+        "vldr d5, [%[lhs_ptr], #8]\n"
+        "vmlal.s8    q6,  d1,  d7\n"
+        "vldr d1, [%[rhs_ptr], #8]\n"
+        "vmlal.s8    q7,  d3,  d7\n"
+        "vldr d3, [%[rhs_ptr], #24]\n"
+
+        // Add pairwise, accumulate into 32-bit accumulators.
+        "vpadal.s16   q12, q4\n"
+        "vldr d7, [%[lhs_ptr], #24]\n"
+        "vpadal.s16   q13, q5\n"
+        "vpadal.s16   q14, q6\n"
+        "vpadal.s16   q15, q7\n"
+
+        "b " GEMMLOWP_LABEL_LOOP "b\n"
+
+        GEMMLOWP_LABEL_AFTER_LOOP
+        ":\n"
+
+        // Multiply first half.
+        "vmull.s8    q4,  d0,  d4\n"
+        "vmull.s8    q5,  d2,  d4\n"
+        "vmull.s8    q6,  d0,  d6\n"
+        "vmull.s8    q7,  d2,  d6\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "vmlal.s8    q4,  d1,  d5\n"
+        "vmlal.s8    q5,  d3,  d5\n"
+        "vmlal.s8    q6,  d1,  d7\n"
+        "vmlal.s8    q7,  d3,  d7\n"
+
+        // Add pairwise, accumulate into 32-bit accumulators.
+        "vpadal.s16   q12, q4\n"
+        "vpadal.s16   q13, q5\n"
+        "vpadal.s16   q14, q6\n"
+        "vpadal.s16   q15, q7\n"
+        "cmp %[start_depth], #0\n"
+
+        // Reduce 32bit accumulators horizontally.
+        "vpadd.s32 d0, d16, d17\n"
+        "vpadd.s32 d1, d18, d19\n"
+        "vpadd.s32 d2, d20, d21\n"
+        "vpadd.s32 d3, d22, d23\n"
+        "vpadd.s32 d4, d24, d25\n"
+        "vpadd.s32 d5, d26, d27\n"
+        "vpadd.s32 d6, d28, d29\n"
+        "vpadd.s32 d7, d30, d31\n"
+
+        "bne " GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+        "f\n"
+
+        // Reduce 32bit accumulators horizontally, second pass
+        // (each pass adds pairwise. we need to add 4-wise).
+        "vpadd.s32 d8, d0, d2\n"
+        "vpadd.s32 d9, d4, d6\n"
+        "vpadd.s32 d10, d1, d3\n"
+        "vpadd.s32 d11, d5, d7\n"
+
+        "b " GEMMLOWP_LABEL_STORE "f\n"
+
+        GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+        ":\n"
+
+        // Reduce 32bit accumulators horizontally, second pass
+        // (each pass adds pairwise. we need to add 4-wise),
+        // and load destination values from memory.
+        "mov r0, %[dst_ptr]\n"
+        "vld1.32 {d16, d17}, [r0], %[dst_col_stride]\n"
+        "vpadd.s32 d8, d0, d2\n"
+        "vpadd.s32 d9, d4, d6\n"
+        "vld1.32 {d18, d19}, [r0]\n"
+        "vpadd.s32 d10, d1, d3\n"
+        "vpadd.s32 d11, d5, d7\n"
+
+        // Add horizontally-reduced accumulators into
+        // the values loaded from memory
+        "vadd.s32 q4, q8, q4\n"
+        "vadd.s32 q5, q9, q5\n"
+
+        GEMMLOWP_LABEL_STORE
+        ":\n"
+        // Store back into memory
+        "mov r0, %[dst_ptr]\n"
+        "vst1.32 {d8, d9}, [r0], %[dst_col_stride]\n"
+        "vst1.32 {d10, d11}, [r0]\n"
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr), [run_depth] "+r"(run_depth)
+        :  // inputs
+        [start_depth] "r"(start_depth),
+        [dst_col_stride] "r"(dst_col_stride)
+        :  // clobbers
+        "cc", "memory", "r0", "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7",
+        "d8", "d9", "d10", "d11", "d12", "d13", "d14", "d15", "d16", "d17",
+        "d18", "d19", "d20", "d21", "d22", "d23", "d24", "d25", "d26", "d27",
+        "d28", "d29", "d30", "d31");
+#undef GEMMLOWP_LABEL_LOOP
+#undef GEMMLOWP_LABEL_AFTER_LOOP
+#undef GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+#undef GEMMLOWP_LABEL_STORE
+  }
+};
+
+#endif  // GEMMLOWP_NEON_32
+
+// The kernels here are specifically arm 64bit assembly, not arm 32bit.
+#ifdef GEMMLOWP_NEON_64
+
+struct NEON_64bit_GEMM_Int8Operands_LhsNonzero : KernelBase {
+  typedef KernelFormat<
+      KernelSideFormatInt8<CellFormat<4, 16, CellOrder::WidthMajor>, 1>,
+      KernelSideFormatInt8<CellFormat<4, 16, CellOrder::WidthMajor>, 1> >
+      Format;
+  const char* Name() const override {
+    return "NEON, 4x4, depth 16, accumulating two within signed int16";
+  }
+
+  // TODO(benoitjacob): reorder function arguments so dst comes last
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+#define GEMMLOWP_LABEL_AFTER_LOOP_LAST16 "1"
+#define GEMMLOWP_LABEL_LOOP "2"
+#define GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES "3"
+#define GEMMLOWP_LABEL_STORE "4"
+    asm volatile(
+        // Clear accumulators, and, interleaved with it,
+        // initial loads of the first loop iteration,
+        // taken out of the loop so that in the loop itself we have
+        // optimal streaming of data from memory.
+        "ld1 {v0.16b}, [%[rhs_ptr]], #16\n"
+        "dup v16.4s, wzr\n"
+        "ld1 {v4.16b}, [%[lhs_ptr]], #16\n"
+        "dup v17.4s, wzr\n"
+        "ld1 {v1.16b}, [%[rhs_ptr]], #16\n"
+        "dup v18.4s, wzr\n"
+        "ld1 {v5.16b}, [%[lhs_ptr]], #16\n"
+        "dup v19.4s, wzr\n"
+        "ld1 {v2.16b}, [%[rhs_ptr]], #16\n"
+        "dup v20.4s, wzr\n"
+        "ld1 {v3.16b}, [%[rhs_ptr]], #16\n"
+        "dup v21.4s, wzr\n"
+        "ld1 {v6.16b}, [%[lhs_ptr]], #16\n"
+        "dup v22.4s, wzr\n"
+        "ld1 {v7.16b}, [%[lhs_ptr]], #16\n"
+        "dup v23.4s, wzr\n"
+        "dup v24.4s, wzr\n"
+        "dup v25.4s, wzr\n"
+        "dup v26.4s, wzr\n"
+        "dup v27.4s, wzr\n"
+        "dup v28.4s, wzr\n"
+        "dup v29.4s, wzr\n"
+        "dup v30.4s, wzr\n"
+        "dup v31.4s, wzr\n"
+
+        // Multiply dst_col_stride by 4 == sizeof(int32) to use
+        // it as a byte offset below.
+        "lsl %[dst_col_stride], %[dst_col_stride], #2\n"
+
+        // Initial arithmetic of the first loop iteration,
+        // taken out of the loop so that in the loop itself we have
+        // optimal streaming of data from memory.
+        "smull    v8.8h,  v0.8b,  v4.8b\n"
+        "smull    v9.8h,  v1.8b,  v4.8b\n"
+        "smull    v10.8h,  v2.8b,  v4.8b\n"
+        "smull    v11.8h,  v3.8b,  v4.8b\n"
+        "smull    v12.8h,  v0.8b,  v5.8b\n"
+        "smull    v13.8h,  v1.8b,  v5.8b\n"
+        "smull    v14.8h,  v2.8b,  v5.8b\n"
+        "smull    v15.8h,  v3.8b,  v5.8b\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "smlal2   v8.8h,  v0.16b,  v4.16b\n"
+        "smlal2   v9.8h,  v1.16b,  v4.16b\n"
+        "smlal2   v10.8h,  v2.16b,  v4.16b\n"
+        "smlal2   v11.8h,  v3.16b,  v4.16b\n"
+        "smlal2   v12.8h,  v0.16b,  v5.16b\n"
+        "smlal2   v13.8h,  v1.16b,  v5.16b\n"
+        "smlal2   v14.8h,  v2.16b,  v5.16b\n"
+        "smlal2   v15.8h,  v3.16b,  v5.16b\n"
+
+        "subs %[run_depth], %[run_depth], #16\n"
+
+        // If the loop depth is only 16, then we can skip the general loop
+        // and go straight to the final part of the code.
+        "beq " GEMMLOWP_LABEL_AFTER_LOOP_LAST16 "f\n"
+
+        // General loop.
+        GEMMLOWP_LABEL_LOOP
+        ":\n"
+
+        // Overview of register layout:
+        //
+        // A 4x16 block of Rhs is stored in 8 bit in v0--v3.
+        // A 4x16 block of Lhs is stored in 8 bit in v4--v7.
+        //
+        // A 4x4 block of accumulators is stored in v16-v31 (as 4x32 bit
+        // components which need to be horizontally-added at the end)
+        //
+        // The Lhs vectors are multiplied by the Rhs vectors with a widening
+        // multiply over the 8 first levels of depth, producing int16x8
+        // vectors of products for each position in the accumulator matrix.
+        // Here comes the special trick: since the operands are signed int8,
+        // their range being [ -2^7 , 2^7 ), their products are in range
+        // [ -2^14 , 2^14 - 1 ), meaning that we can add two such values
+        // without any risk of overflowing int16.
+        // We thus proceed with the 8 next levels of depth, multiplying
+        // again Lhs by Rhs, accumulating into this existing int16x8 vector.
+        //
+        // Only then, having processed 16 levels of depth, do we need to
+        // horizontally add these int16x8 accumulators into the final
+        // int32x4 accumulators.
+        //
+        // As we do not have enough registers to store all 16 int16x8
+        // temporary-16bit-accumulators, we have them cycle through v8--v15.
+        //
+        //
+        // Register layout (ignoring the v8--v15 temporary 16bit accumulators):
+        //
+        //                               +--------+--------+--------+--------+
+        //                               |v0.b[0] |v1.b[0] |v2.b[0] |v3.b[0] |
+        //                          Rhs  +--------+--------+--------+--------+
+        //                               |  ...   |  ...   |  ...   |  ...   |
+        //                               +--------+--------+--------+--------|
+        //                               |v0.b[15]|v1.b[15]|v2.b[15]|v3.b[15]|
+        //                               +--------+--------+--------+--------+
+        //
+        //                               |        |        |        |        |
+        //
+        //    Lhs                        |        |        |        |        |
+        //
+        //  +-------+-----+--------+ - - +--------+--------+--------+--------+
+        //  |v4.b[0]| ... |v4.b[15]|     | v16.4s | v17.4s | v18.4s | v19.4s |
+        //  |v5.b[0]| ... |v5.b[15]|     | v20.4s | v21.4s | v22.4s | v23.4s |
+        //  |v6.b[0]| ... |v6.b[15]|     | v24.4s | v25.4s | v26.4s | v27.4s |
+        //  |v7.b[0]| ... |v7.b[15]|     | v28.4s | v29.4s | v30.4s | v31.4s |
+        //  +-------+--------------+ - - +--------+--------+--------+--------+
+        //
+        //                                                Accumulator
+        //
+
+        // Some multiplications and 16-bit accumulation were already done above,
+        // so we start right away in the middle.
+        "sadalp  v16.4s, v8.8h\n"
+        "ld1 {v4.16b}, [%[lhs_ptr]], #16\n"
+        "smull    v8.8h,  v0.8b,  v6.8b\n"
+        "sadalp  v17.4s, v9.8h\n"
+        "ld1 {v5.16b}, [%[lhs_ptr]], #16\n"
+        "smull    v9.8h,  v1.8b,  v6.8b\n"
+        "sadalp  v18.4s, v10.8h\n"
+        "smull    v10.8h,  v2.8b,  v6.8b\n"
+        "sadalp  v19.4s, v11.8h\n"
+        "smull    v11.8h,  v3.8b,  v6.8b\n"
+        "sadalp  v20.4s, v12.8h\n"
+        "smull    v12.8h,  v0.8b,  v7.8b\n"
+        "sadalp  v21.4s, v13.8h\n"
+        "smull    v13.8h,  v1.8b,  v7.8b\n"
+        "sadalp  v22.4s, v14.8h\n"
+        "smull    v14.8h,  v2.8b,  v7.8b\n"
+        "sadalp  v23.4s, v15.8h\n"
+        "smull    v15.8h,  v3.8b,  v7.8b\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "smlal2   v8.8h,  v0.16b,  v6.16b\n"
+        "smlal2   v9.8h,  v1.16b,  v6.16b\n"
+        "smlal2   v10.8h,  v2.16b,  v6.16b\n"
+        "smlal2   v11.8h,  v3.16b,  v6.16b\n"
+
+        "ld1 {v6.16b}, [%[lhs_ptr]], #16\n"
+
+        "smlal2   v12.8h,  v0.16b,  v7.16b\n"
+        "ld1 {v0.16b}, [%[rhs_ptr]], #16\n"
+        "smlal2   v13.8h,  v1.16b,  v7.16b\n"
+        "ld1 {v1.16b}, [%[rhs_ptr]], #16\n"
+        "smlal2   v14.8h,  v2.16b,  v7.16b\n"
+        "ld1 {v2.16b}, [%[rhs_ptr]], #16\n"
+        "smlal2   v15.8h,  v3.16b,  v7.16b\n"
+        "ld1 {v3.16b}, [%[rhs_ptr]], #16\n"
+
+        "sadalp  v24.4s, v8.8h\n"
+        "smull    v8.8h,  v0.8b,  v4.8b\n"
+        "sadalp  v25.4s, v9.8h\n"
+        "ld1 {v7.16b}, [%[lhs_ptr]], #16\n"
+        "smull    v9.8h,  v1.8b,  v4.8b\n"
+        "sadalp  v26.4s, v10.8h\n"
+        "smull    v10.8h,  v2.8b,  v4.8b\n"
+        "sadalp  v27.4s, v11.8h\n"
+        "smull    v11.8h,  v3.8b,  v4.8b\n"
+        "sadalp  v28.4s, v12.8h\n"
+        "smull    v12.8h,  v0.8b,  v5.8b\n"
+        "sadalp  v29.4s, v13.8h\n"
+        "smull    v13.8h,  v1.8b,  v5.8b\n"
+        "sadalp  v30.4s, v14.8h\n"
+        "smull    v14.8h,  v2.8b,  v5.8b\n"
+        "sadalp  v31.4s, v15.8h\n"
+        "smull    v15.8h,  v3.8b,  v5.8b\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "smlal2   v8.8h,  v0.16b,  v4.16b\n"
+        "smlal2   v9.8h,  v1.16b,  v4.16b\n"
+        "smlal2   v10.8h,  v2.16b,  v4.16b\n"
+        "smlal2   v11.8h,  v3.16b,  v4.16b\n"
+
+        // Loop. Decrement loop index (depth) by 16, since we just handled
+        // 16 levels of depth.  Do this subs a bit before the end of the loop
+        // for better dispatch on A57.
+        "subs %[run_depth], %[run_depth], #16\n"
+
+        "smlal2   v12.8h,  v0.16b,  v5.16b\n"
+        "smlal2   v13.8h,  v1.16b,  v5.16b\n"
+        "smlal2   v14.8h,  v2.16b,  v5.16b\n"
+        "smlal2   v15.8h,  v3.16b,  v5.16b\n"
+
+        "bne " GEMMLOWP_LABEL_LOOP "b\n"
+
+        // Final code for the last 16 levels of depth.
+        // There is nothing to load anymore, only some arithmetic to finish.
+        GEMMLOWP_LABEL_AFTER_LOOP_LAST16
+        ":\n"
+
+        // Some multiplications and 16-bit accumulation were already done above,
+        // so we start right away in the middle.
+        "sadalp  v16.4s, v8.8h\n"
+        "smull    v8.8h,  v0.8b,  v6.8b\n"
+        "sadalp  v17.4s, v9.8h\n"
+        "smull    v9.8h,  v1.8b,  v6.8b\n"
+        "sadalp  v18.4s, v10.8h\n"
+        "smull    v10.8h,  v2.8b,  v6.8b\n"
+        "sadalp  v19.4s, v11.8h\n"
+        "smull    v11.8h,  v3.8b,  v6.8b\n"
+        "sadalp  v20.4s, v12.8h\n"
+        "smull    v12.8h,  v0.8b,  v7.8b\n"
+        "sadalp  v21.4s, v13.8h\n"
+        "smull    v13.8h,  v1.8b,  v7.8b\n"
+        "sadalp  v22.4s, v14.8h\n"
+        "smull    v14.8h,  v2.8b,  v7.8b\n"
+        "sadalp  v23.4s, v15.8h\n"
+        "smull    v15.8h,  v3.8b,  v7.8b\n"
+
+        // Multiply-accumulate second-half, again into the same
+        // 16bit local accumulator registers. This is where we
+        // take advantage of having int8 instead of uint8 and therefore
+        // being able to accumulate two products into int16.
+        "smlal2   v8.8h,  v0.16b,  v6.16b\n"
+        "smlal2   v9.8h,  v1.16b,  v6.16b\n"
+        "smlal2   v10.8h,  v2.16b,  v6.16b\n"
+        "smlal2   v11.8h,  v3.16b,  v6.16b\n"
+        "smlal2   v12.8h,  v0.16b,  v7.16b\n"
+        "smlal2   v13.8h,  v1.16b,  v7.16b\n"
+        "smlal2   v14.8h,  v2.16b,  v7.16b\n"
+        "smlal2   v15.8h,  v3.16b,  v7.16b\n"
+
+        "sadalp  v24.4s, v8.8h\n"
+        "sadalp  v25.4s, v9.8h\n"
+        "sadalp  v26.4s, v10.8h\n"
+        "sadalp  v27.4s, v11.8h\n"
+        "sadalp  v28.4s, v12.8h\n"
+        "sadalp  v29.4s, v13.8h\n"
+        "sadalp  v30.4s, v14.8h\n"
+        "sadalp  v31.4s, v15.8h\n"
+
+        // Reduce 32bit accumulators horizontally.
+        "addp v0.4s, v16.4s, v20.4s\n"
+        "addp v2.4s, v17.4s, v21.4s\n"
+        "addp v4.4s, v18.4s, v22.4s\n"
+        "addp v6.4s, v19.4s, v23.4s\n"
+        "addp v1.4s, v24.4s, v28.4s\n"
+        "addp v3.4s, v25.4s, v29.4s\n"
+        "addp v5.4s, v26.4s, v30.4s\n"
+        "addp v7.4s, v27.4s, v31.4s\n"
+
+        "cmp %[start_depth], #0\n"
+        "bne " GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+        "f\n"
+
+        // Reduce 32bit accumulators horizontally, second pass
+        // (each pass adds pairwise. we need to add 4-wise).
+        "addp v12.4s, v0.4s, v1.4s\n"
+        "addp v13.4s, v2.4s, v3.4s\n"
+        "addp v14.4s, v4.4s, v5.4s\n"
+        "addp v15.4s, v6.4s, v7.4s\n"
+
+        "b " GEMMLOWP_LABEL_STORE "f\n"
+
+        GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+        ":\n"
+
+        // Reduce 32bit accumulators horizontally, second pass
+        // (each pass adds pairwise. we need to add 4-wise),
+        // and load destination values from memory.
+        "mov x0, %[dst_ptr]\n"
+        "ld1 {v12.16b}, [x0], %[dst_col_stride]\n"
+        "addp v8.4s, v0.4s, v1.4s\n"
+        "ld1 {v13.16b}, [x0], %[dst_col_stride]\n"
+        "addp v9.4s, v2.4s, v3.4s\n"
+        "ld1 {v14.16b}, [x0], %[dst_col_stride]\n"
+        "addp v10.4s, v4.4s, v5.4s\n"
+        "ld1 {v15.16b}, [x0]\n"
+        "addp v11.4s, v6.4s, v7.4s\n"
+
+        // Add horizontally-reduced accumulators into
+        // the values loaded from memory
+        "add v12.4s, v12.4s, v8.4s\n"
+        "add v13.4s, v13.4s, v9.4s\n"
+        "add v14.4s, v14.4s, v10.4s\n"
+        "add v15.4s, v15.4s, v11.4s\n"
+
+        GEMMLOWP_LABEL_STORE
+        ":\n"
+        // Store back into memory
+        "mov x0, %[dst_ptr]\n"
+        "st1 {v12.16b}, [x0], %[dst_col_stride]\n"
+        "st1 {v13.16b}, [x0], %[dst_col_stride]\n"
+        "st1 {v14.16b}, [x0], %[dst_col_stride]\n"
+        "st1 {v15.16b}, [x0]\n"
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr), [run_depth] "+r"(run_depth),
+        [dst_col_stride] "+r"(dst_col_stride)
+        :  // inputs
+        [start_depth] "r"(start_depth)
+        :  // clobbers
+        "cc", "memory", "x0", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7",
+        "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17",
+        "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27",
+        "v28", "v29", "v30", "v31");
+#undef GEMMLOWP_LABEL_LOOP
+#undef GEMMLOWP_LABEL_AFTER_LOOP_LAST16
+#undef GEMMLOWP_LABEL_ACCUMULATE_EXISTING_DST_VALUES
+#undef GEMMLOWP_LABEL_STORE
+  }
+};
+
+
+// Our main GEMM kernel.
+struct NEON_64_Kernel12x8Depth2 : KernelBase {
+  typedef KernelFormat<KernelSideFormat<CellFormat<4, 2>, 3>,
+                       KernelSideFormat<CellFormat<4, 2>, 2> >
+      Format;
+
+  const char* Name() const override { return "NEON, 12x8, depth 2"; }
+
+  // TODO(benoitjacob): reorder function arguments so dst comes last
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    ScopedProfilingLabel label("optimized kernel (NEON 12x8)");
+// See comments above for why we need local numerical labels in our asm.
+#define GEMMLOWP_LABEL_CLEAR_ACCUMULATORS "1"
+#define GEMMLOWP_LABEL_BEFORE_LOOP "2"
+#define GEMMLOWP_LABEL_LOOP "3"
+#define GEMMLOWP_LABEL_AFTER_LOOP "4"
+
+    assert(dst_row_stride == 1);
+    asm volatile(
+        // Load 1 Rhs cell of size 2x8
+        "ld1 {v5.8b}, [%[rhs_ptr]], #8\n"
+        "ld1 {v6.8b}, [%[rhs_ptr]], #8\n"
+
+        // Load 3 Lhs cells of size 4x2 each
+        "ld1 {v2.8b}, [%[lhs_ptr]], #8\n"
+        "ld1 {v3.8b}, [%[lhs_ptr]], #8\n"
+        "ld1 {v4.8b}, [%[lhs_ptr]], #8\n"
+
+        // Multiply dst_col_stride by 4 == sizeof(int32) to use
+        // it as a byte offset below.
+        "lsl %[dst_col_stride], %[dst_col_stride], #2\n"
+
+        "cmp %[start_depth], #0\n"
+        "beq " GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+        "f\n"
+
+        // Load accumulators
+        "mov x1, %[dst_ptr]\n"
+        "mov x0, x1\n"
+        "ld1 {v8.16b}, [x0], #16\n"
+        "subs %[run_depth], %[run_depth], #2\n"
+        "ld1 {v16.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v24.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v9.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v17.16b}, [x0], #16\n"
+        "ld1 {v25.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v10.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v18.16b}, [x0], #16\n"
+        "ld1 {v26.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v11.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v19.16b}, [x0], #16\n"
+        "ld1 {v27.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v12.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v20.16b}, [x0], #16\n"
+        "ld1 {v28.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v13.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v21.16b}, [x0], #16\n"
+        "ld1 {v29.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v14.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "ld1 {v22.16b}, [x0], #16\n"
+        "ld1 {v30.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "ld1 {v15.16b}, [x0], #16\n"
+        "ld1 {v23.16b}, [x0], #16\n"
+        "ld1 {v31.16b}, [x0]\n"
+
+        "b " GEMMLOWP_LABEL_BEFORE_LOOP "f\n"
+
+        GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+        ":\n"
+
+        // Clear accumulator registers (see layout below)
+        "dup v8.4s, wzr\n"
+        "subs %[run_depth], %[run_depth], #2\n"
+        "dup v9.4s, wzr\n"
+        "dup v10.4s, wzr\n"
+        "dup v11.4s, wzr\n"
+        "dup v12.4s, wzr\n"
+        "dup v13.4s, wzr\n"
+        "dup v14.4s, wzr\n"
+        "dup v15.4s, wzr\n"
+        "dup v16.4s, wzr\n"
+        "dup v17.4s, wzr\n"
+        "dup v18.4s, wzr\n"
+        "dup v19.4s, wzr\n"
+        "dup v20.4s, wzr\n"
+        "dup v21.4s, wzr\n"
+        "dup v22.4s, wzr\n"
+        "dup v23.4s, wzr\n"
+        "dup v24.4s, wzr\n"
+        "dup v25.4s, wzr\n"
+        "dup v26.4s, wzr\n"
+        "dup v27.4s, wzr\n"
+        "dup v28.4s, wzr\n"
+        "dup v29.4s, wzr\n"
+        "dup v30.4s, wzr\n"
+        "dup v31.4s, wzr\n"
+
+        GEMMLOWP_LABEL_BEFORE_LOOP
+        ":\n"
+
+        "beq " GEMMLOWP_LABEL_AFTER_LOOP "f\n"
+
+        GEMMLOWP_LABEL_LOOP
+        ":\n"
+
+        // Overview of register layout:
+        //
+        // A 2x8 block of 2 2x4 cells of Rhs is stored in 16bit in v0--v1.
+        // A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in v2--v4.
+        // A 12x8 block of accumulators is stored in 32bit in v8--v31.
+        //
+        //                         +--------+--------+-----+--------+--------+
+        //                         |v0.h[0] |v0.h[1] | ... |v1.h[2] |v1.h[3] |
+        //                    Rhs  +--------+--------+-----+--------+--------+
+        //                         |v0.h[4] |v0.h[5] | ... |v1.h[6] |v1.h[7] |
+        //                         +--------+--------+-----+--------+--------+
+        //
+        //                         |        |        |     |        |        |
+        //
+        //    Lhs                  |        |        |     |        |        |
+        //
+        //  +-------+-------+ - -  +--------+--------+-----+--------+--------+
+        //  |v2.h[0]|v2.h[4]|      |v8.s[0] |v9.s[0] | ... |v14.s[0]|v15.s[0]|
+        //  |v2.h[1]|v2.h[5]|      |v8.s[1] |v9.s[1] | ... |v14.s[1]|v15.s[1]|
+        //  |v2.h[2]|v2.h[6]|      |v8.s[2] |v9.s[2] | ... |v14.s[2]|v15.s[2]|
+        //  |v2.h[3]|v2.h[7]|      |v8.s[3] |v9.s[3] | ... |v14.s[3]|v15.s[3]|
+        //  +-------+-------+ - -  +--------+--------+-----+--------+--------+
+        //  |v3.h[0]|v3.h[4]|      |v16.s[0]|v17.s[0]| ... |v22.s[0]|v23.s[0]|
+        //  |v3.h[1]|v3.h[5]|      |v16.s[1]|v17.s[1]| ... |v22.s[1]|v23.s[1]|
+        //  |v3.h[2]|v3.h[6]|      |v16.s[2]|v17.s[2]| ... |v22.s[2]|v23.s[2]|
+        //  |v3.h[3]|v3.h[7]|      |v16.s[3]|v17.s[3]| ... |v22.s[3]|v23.s[3]|
+        //  +-------+-------+ - -  +--------+--------+-----+--------+--------+
+        //  |v4.h[0]|v4.h[4]|      |v24.s[0]|v25.s[0]| ... |v30.s[0]|v31.s[0]|
+        //  |v4.h[1]|v4.h[5]|      |v24.s[1]|v25.s[1]| ... |v30.s[1]|v31.s[1]|
+        //  |v4.h[2]|v4.h[6]|      |v24.s[2]|v25.s[2]| ... |v30.s[2]|v31.s[2]|
+        //  |v4.h[3]|v4.h[7]|      |v24.s[3]|v25.s[3]| ... |v30.s[3]|v31.s[3]|
+        //  +-------+-------+ - -  +--------+--------+-----+--------+--------+
+        //
+        //                            Accumulator
+
+        // Expand Lhs/Rhs cells to 16 bit.
+        "uxtl v0.8h, v5.8b\n"
+        "ld1 {v5.8b}, [%[rhs_ptr]], #8\n"
+        "uxtl v1.8h, v6.8b\n"
+        "ld1 {v6.8b}, [%[rhs_ptr]], #8\n"
+        "uxtl v2.8h, v2.8b\n"
+        "uxtl v3.8h, v3.8b\n"
+        "uxtl v4.8h, v4.8b\n"
+
+        // Multiply-accumulate, top third
+        "umlal v8.4s, v2.4h, v0.h[0]\n"
+        "umlal v9.4s, v2.4h, v0.h[1]\n"
+        "umlal v10.4s, v2.4h, v0.h[2]\n"
+        "umlal v11.4s, v2.4h, v0.h[3]\n"
+        "umlal v12.4s, v2.4h, v1.h[0]\n"
+        "umlal v13.4s, v2.4h, v1.h[1]\n"
+        "umlal v14.4s, v2.4h, v1.h[2]\n"
+        "umlal v15.4s, v2.4h, v1.h[3]\n"
+        "umlal2 v8.4s, v2.8h, v0.h[4]\n"
+        "umlal2 v9.4s, v2.8h, v0.h[5]\n"
+        "umlal2 v10.4s, v2.8h, v0.h[6]\n"
+        "umlal2 v11.4s, v2.8h, v0.h[7]\n"
+        "umlal2 v12.4s, v2.8h, v1.h[4]\n"
+        "umlal2 v13.4s, v2.8h, v1.h[5]\n"
+        "umlal2 v14.4s, v2.8h, v1.h[6]\n"
+        "umlal2 v15.4s, v2.8h, v1.h[7]\n"
+        "ld1 {v2.8b}, [%[lhs_ptr]], #8\n"
+
+        // Multiply-accumulate, middle third
+        "umlal v16.4s, v3.4h, v0.h[0]\n"
+        "umlal v17.4s, v3.4h, v0.h[1]\n"
+        "umlal v18.4s, v3.4h, v0.h[2]\n"
+        "umlal v19.4s, v3.4h, v0.h[3]\n"
+        "umlal v20.4s, v3.4h, v1.h[0]\n"
+        "umlal v21.4s, v3.4h, v1.h[1]\n"
+        "umlal v22.4s, v3.4h, v1.h[2]\n"
+        "umlal v23.4s, v3.4h, v1.h[3]\n"
+        "umlal2 v16.4s, v3.8h, v0.h[4]\n"
+        "umlal2 v17.4s, v3.8h, v0.h[5]\n"
+        "umlal2 v18.4s, v3.8h, v0.h[6]\n"
+        "umlal2 v19.4s, v3.8h, v0.h[7]\n"
+        "umlal2 v20.4s, v3.8h, v1.h[4]\n"
+        "umlal2 v21.4s, v3.8h, v1.h[5]\n"
+        "umlal2 v22.4s, v3.8h, v1.h[6]\n"
+        "umlal2 v23.4s, v3.8h, v1.h[7]\n"
+        "ld1 {v3.8b}, [%[lhs_ptr]], #8\n"
+
+        "subs %[run_depth], %[run_depth], #2\n"
+
+        // Multiply-accumulate, bottom third
+        "umlal v24.4s, v4.4h, v0.h[0]\n"
+        "umlal v25.4s, v4.4h, v0.h[1]\n"
+        "umlal v26.4s, v4.4h, v0.h[2]\n"
+        "umlal v27.4s, v4.4h, v0.h[3]\n"
+        "umlal v28.4s, v4.4h, v1.h[0]\n"
+        "umlal v29.4s, v4.4h, v1.h[1]\n"
+        "umlal v30.4s, v4.4h, v1.h[2]\n"
+        "umlal v31.4s, v4.4h, v1.h[3]\n"
+        "umlal2 v24.4s, v4.8h, v0.h[4]\n"
+        "umlal2 v25.4s, v4.8h, v0.h[5]\n"
+        "umlal2 v26.4s, v4.8h, v0.h[6]\n"
+        "umlal2 v27.4s, v4.8h, v0.h[7]\n"
+        "umlal2 v28.4s, v4.8h, v1.h[4]\n"
+        "umlal2 v29.4s, v4.8h, v1.h[5]\n"
+        "umlal2 v30.4s, v4.8h, v1.h[6]\n"
+        "umlal2 v31.4s, v4.8h, v1.h[7]\n"
+        "ld1 {v4.8b}, [%[lhs_ptr]], #8\n"
+
+        "bne " GEMMLOWP_LABEL_LOOP "b\n"
+
+        GEMMLOWP_LABEL_AFTER_LOOP
+        ":\n"
+
+        // Expand Lhs/Rhs cells to 16 bit.
+        "uxtl v0.8h, v5.8b\n"
+        "uxtl v1.8h, v6.8b\n"
+        "uxtl v2.8h, v2.8b\n"
+        "uxtl v3.8h, v3.8b\n"
+        "uxtl v4.8h, v4.8b\n"
+
+        // Multiply-accumulate, level of depth 0
+        "umlal v8.4s, v2.4h, v0.h[0]\n"
+        "umlal v9.4s, v2.4h, v0.h[1]\n"
+        "umlal v10.4s, v2.4h, v0.h[2]\n"
+        "umlal v11.4s, v2.4h, v0.h[3]\n"
+        "umlal v12.4s, v2.4h, v1.h[0]\n"
+        "umlal v13.4s, v2.4h, v1.h[1]\n"
+        "umlal v14.4s, v2.4h, v1.h[2]\n"
+        "umlal v15.4s, v2.4h, v1.h[3]\n"
+        "umlal v16.4s, v3.4h, v0.h[0]\n"
+        "umlal v17.4s, v3.4h, v0.h[1]\n"
+        "umlal v18.4s, v3.4h, v0.h[2]\n"
+        "umlal v19.4s, v3.4h, v0.h[3]\n"
+        "umlal v20.4s, v3.4h, v1.h[0]\n"
+        "umlal v21.4s, v3.4h, v1.h[1]\n"
+        "umlal v22.4s, v3.4h, v1.h[2]\n"
+        "umlal v23.4s, v3.4h, v1.h[3]\n"
+        "umlal v24.4s, v4.4h, v0.h[0]\n"
+        "umlal v25.4s, v4.4h, v0.h[1]\n"
+        "umlal v26.4s, v4.4h, v0.h[2]\n"
+        "umlal v27.4s, v4.4h, v0.h[3]\n"
+        "umlal v28.4s, v4.4h, v1.h[0]\n"
+        "umlal v29.4s, v4.4h, v1.h[1]\n"
+        "umlal v30.4s, v4.4h, v1.h[2]\n"
+        "umlal v31.4s, v4.4h, v1.h[3]\n"
+
+        // Multiply-accumulate, level of depth 1
+        "umlal2 v8.4s, v2.8h, v0.h[4]\n"
+        "umlal2 v9.4s, v2.8h, v0.h[5]\n"
+        "umlal2 v10.4s, v2.8h, v0.h[6]\n"
+        "umlal2 v11.4s, v2.8h, v0.h[7]\n"
+        "umlal2 v12.4s, v2.8h, v1.h[4]\n"
+        "umlal2 v13.4s, v2.8h, v1.h[5]\n"
+        "umlal2 v14.4s, v2.8h, v1.h[6]\n"
+        "umlal2 v15.4s, v2.8h, v1.h[7]\n"
+        "umlal2 v16.4s, v3.8h, v0.h[4]\n"
+        "umlal2 v17.4s, v3.8h, v0.h[5]\n"
+        "umlal2 v18.4s, v3.8h, v0.h[6]\n"
+        "umlal2 v19.4s, v3.8h, v0.h[7]\n"
+        "umlal2 v20.4s, v3.8h, v1.h[4]\n"
+        "umlal2 v21.4s, v3.8h, v1.h[5]\n"
+        "umlal2 v22.4s, v3.8h, v1.h[6]\n"
+        "umlal2 v23.4s, v3.8h, v1.h[7]\n"
+        "umlal2 v24.4s, v4.8h, v0.h[4]\n"
+        "umlal2 v25.4s, v4.8h, v0.h[5]\n"
+        "umlal2 v26.4s, v4.8h, v0.h[6]\n"
+        "umlal2 v27.4s, v4.8h, v0.h[7]\n"
+        "umlal2 v28.4s, v4.8h, v1.h[4]\n"
+        "umlal2 v29.4s, v4.8h, v1.h[5]\n"
+        "umlal2 v30.4s, v4.8h, v1.h[6]\n"
+        "umlal2 v31.4s, v4.8h, v1.h[7]\n"
+
+        // Store accumulators
+        "mov x1, %[dst_ptr]\n"
+        "mov x0, x1\n"
+        "st1 {v8.16b}, [x0], #16\n"
+        "subs %[run_depth], %[run_depth], #2\n"
+        "st1 {v16.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v24.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v9.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v17.16b}, [x0], #16\n"
+        "st1 {v25.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v10.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v18.16b}, [x0], #16\n"
+        "st1 {v26.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v11.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v19.16b}, [x0], #16\n"
+        "st1 {v27.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v12.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v20.16b}, [x0], #16\n"
+        "st1 {v28.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v13.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v21.16b}, [x0], #16\n"
+        "st1 {v29.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v14.16b}, [x0], #16\n"
+        "add x1, x1, %[dst_col_stride]\n"
+        "st1 {v22.16b}, [x0], #16\n"
+        "st1 {v30.16b}, [x0]\n"
+        "mov x0, x1\n"
+        "st1 {v15.16b}, [x0], #16\n"
+        "st1 {v23.16b}, [x0], #16\n"
+        "st1 {v31.16b}, [x0]\n"
+#undef GEMMLOWP_LABEL_CLEAR_ACCUMULATORS
+#undef GEMMLOWP_LABEL_BEFORE_LOOP
+#undef GEMMLOWP_LABEL_LOOP
+#undef GEMMLOWP_LABEL_AFTER_LOOP
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr),
+        [run_depth] "+r"(run_depth)
+        :  // inputs
+        [start_depth] "r"(start_depth),
+        [dst_col_stride] "r"(dst_col_stride)
+        :  // clobbers
+        "cc", "memory", "x0", "x1", "v0", "v1", "v2", "v3", "v4", "v5", "v6",
+        "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16",
+        "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26",
+        "v27", "v28", "v29", "v30", "v31");
+  }
+};
+
+#endif  // GEMMLOWP_NEON_64
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_KERNEL_NEON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_reference.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_reference.h
new file mode 100644
index 0000000..3458c6a
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_reference.h
@@ -0,0 +1,118 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// kernel_reference.h: a reference kernel for CPU architectures where we don't
+// have optimized kernels yet. Also useful for testing, as it's templatized
+// to have any arbitrary format, allowing tests to cover all sorts of corner
+// cases.
+
+#ifndef GEMMLOWP_INTERNAL_KERNEL_REFERENCE_H_
+#define GEMMLOWP_INTERNAL_KERNEL_REFERENCE_H_
+
+#include "kernel.h"
+
+#include <cstdio>
+#include <cstring>
+
+namespace gemmlowp {
+
+// This kernel is templatized in an arbitrary Format template parameter,
+// allowing it to have any arbitrary format.
+template <typename tFormat>
+struct ReferenceKernel : KernelBase {
+  typedef tFormat Format;
+
+  const char* Name() const override {
+    static char buf[256];
+    snprintf(buf, sizeof(buf),
+             "reference(Lhs: %d cells %dx%d %s, Rhs: %d cells %dx%d %s)",
+             Format::Lhs::kCells, Format::Lhs::Cell::kWidth,
+             Format::Lhs::Cell::kDepth,
+             CellOrderName(Format::Lhs::Cell::kOrder), Format::Rhs::kCells,
+             Format::Rhs::Cell::kDepth, Format::Rhs::Cell::kWidth,
+             CellOrderName(Format::Rhs::Cell::kOrder));
+    return buf;
+  }
+
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    std::int32_t accumulator[Format::kRows * Format::kCols];
+    memset(accumulator, 0, sizeof(accumulator));
+
+    const int run_depth_cells = static_cast<int>(run_depth / Format::kDepth);
+
+    // The outer loop is over the depth dimension.
+    for (int dc = 0; dc < run_depth_cells; dc++) {
+      // The next two loops are over cells of the Lhs (stacked vertically),
+      // and over cells of the Rhs (stacked horizontally).
+      for (int rc = 0; rc < Format::Lhs::kCells; rc++) {
+        const std::uint8_t* lhs_cell_ptr =
+            lhs_ptr + (dc * Format::Lhs::kCells + rc) *
+                          Format::Lhs::Cell::kWidth * Format::kDepth;
+        for (int cc = 0; cc < Format::Rhs::kCells; cc++) {
+          const std::uint8_t* rhs_cell_ptr =
+              rhs_ptr + (dc * Format::Rhs::kCells + cc) *
+                            Format::Rhs::Cell::kWidth * Format::kDepth;
+
+          // Now we are inside one cell of the Lhs and inside one cell
+          // of the Rhs, so the remaining inner loops are just
+          // traditional three loops of matrix multiplication.
+          for (int di = 0; di < Format::kDepth; di++) {
+            for (int ri = 0; ri < Format::Lhs::Cell::kWidth; ri++) {
+              for (int ci = 0; ci < Format::Rhs::Cell::kWidth; ci++) {
+                const std::uint8_t* lhs_coeff_ptr =
+                    lhs_cell_ptr +
+                    OffsetIntoCell<typename Format::Lhs::Cell>(ri, di);
+                const std::uint8_t* rhs_coeff_ptr =
+                    rhs_cell_ptr +
+                    OffsetIntoCell<typename Format::Rhs::Cell>(ci, di);
+                std::int32_t* accumulator_coeff_ptr =
+                    accumulator + (ri + rc * Format::Lhs::Cell::kWidth) +
+                    (ci + cc * Format::Rhs::Cell::kWidth) * Format::kRows;
+                *accumulator_coeff_ptr +=
+                    std::int32_t(*lhs_coeff_ptr) * std::int32_t(*rhs_coeff_ptr);
+              }
+            }
+          }
+        }
+      }
+    }
+
+    if (start_depth == 0) {
+      // start_depth == 0 means we haven't accumulated anything yet, so we need
+      // to overwrite the accumulator, as it hasn't been initialized to zero.
+      for (int r = 0; r < Format::kRows; r++) {
+        for (int c = 0; c < Format::kCols; c++) {
+          dst_ptr[r * dst_row_stride + c * dst_col_stride] =
+              accumulator[r + c * Format::kRows];
+        }
+      }
+    } else {
+      // We have already accumulated stuff, so we need to continue accumulating
+      // instead of just overwriting.
+      for (int r = 0; r < Format::kRows; r++) {
+        for (int c = 0; c < Format::kCols; c++) {
+          dst_ptr[r * dst_row_stride + c * dst_col_stride] +=
+              accumulator[r + c * Format::kRows];
+        }
+      }
+    }
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_KERNEL_REFERENCE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_sse.h
new file mode 100644
index 0000000..b879fd7
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/kernel_sse.h
@@ -0,0 +1,517 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// kernel_SSE.h: a collection of Intel SSE optimized kernels.
+// Check in kernel_default.h which one(s) are actually used by default.
+// Others are mere experiments; they are still covered by tests
+// in case they might be useful some day.
+//
+
+#ifndef GEMMLOWP_INTERNAL_KERNEL_SSE_H_
+#define GEMMLOWP_INTERNAL_KERNEL_SSE_H_
+
+#include "kernel.h"
+
+#include <string.h>
+#include <cassert>
+
+namespace gemmlowp {
+
+#ifdef GEMMLOWP_SSE4_32
+struct SSE4_32_Kernel4x4Depth2 : KernelBase {
+  typedef KernelFormat<
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 1>,
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 1> >
+      Format;
+
+  const char* Name() const override { return "SSE, 4x4, depth 2"; }
+
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    ScopedProfilingLabel label("optimized kernel");
+    assert(dst_row_stride == 1);
+    std::int32_t run_depth_cells = run_depth / Format::kDepth;
+    /* Main loop */
+
+    // A 2x4 cell of Rhs is stored in 16bit in xmm1 .
+    // A 4x2 block Lhs is stored in 16bit in xmm0.
+    // A 4x4 block of accumulators is stored in 32bit in xmm4--xmm7.
+    //
+    //                   +-------+-------+-------+-------+
+    //                   |xmm1[0]|xmm1[2]|xmm1[4]|xmm1[6]|
+    //              Rhs  +-------+---------------+-------+
+    //                   |xmm1[1]|xmm1[3]|xmm1[5]|xmm1[7]|
+    //                   +-------+-------+-------+-------+
+    //
+    //                   |       |       |       |       |
+    //
+    //    Lhs            |       |       |       |       |
+    //
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 | (Iter1)  | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //
+    //                              Accumulator
+
+    asm volatile(
+
+        // set accumulators to zero.
+        "pxor %%xmm4  , %%xmm4 \n\t"
+        "pxor %%xmm5  , %%xmm5 \n\t"
+        "pxor %%xmm6  , %%xmm6 \n\t"
+        "pxor %%xmm7  , %%xmm7 \n\t"
+
+        "movl  %[run_depth_cells], %%eax\n\t"
+        "subl $2, %%eax\n\t"
+        "js outerLoop1%=\n\t"
+
+        // Loop for K unrolled by 4
+        "outerLoop2%=:\n\t"
+
+        // K = 1,2
+        // RHS cell to xmm1
+        "pmovzxbw (%[rhs_ptr]), %%xmm1\n\t"
+
+        // LHS cell
+        "pmovzxbw 0x00(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+
+        "prefetcht0 0x80(%[lhs_ptr]) \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+
+        "prefetcht0 0x80(%[rhs_ptr]) \n\t"
+
+        // K = 3,4
+        // RHS cell to xmm1
+        "pmovzxbw 0x08(%[rhs_ptr]), %%xmm1\n\t"
+
+        "paddd %%xmm2, %%xmm6           \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+
+        // LHS cell
+        "pmovzxbw 0x08(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+
+        "addl $0x10, %[lhs_ptr]         \n\t"
+        "addl $0x10, %[rhs_ptr]         \n\t"
+
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "paddd %%xmm2, %%xmm6           \n\t"
+
+        "subl $2, %[run_depth_cells]\n\t"
+        "ja outerLoop2%=\n\t"
+
+        "movl %[run_depth_cells], %%eax\n\t"
+        "decl %%eax\n\t"
+        "js finish%=\n\t"
+
+        // Loop for K unrolled by 2
+        "outerLoop1%=:\n\t"
+
+        // RHS cell to xmm1
+        "pmovzxbw (%[rhs_ptr]), %%xmm1\n\t"
+
+        // LHS cell
+        "pmovzxbw 0x00(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "paddd %%xmm2, %%xmm6           \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+
+        "addl $0x08, %[lhs_ptr]\n\t"
+        "addl $0x08, %[rhs_ptr]\n\t"
+
+        "decl %[run_depth_cells]\n\t"
+        "jnz outerLoop1%=\n\t"
+
+        "finish%=:\n\t"
+
+        "movl  %[dst_col_stride], %%eax\n\t"
+        "shll $2, %%eax\n\t"
+
+        "movl  %[start_depth], %%ecx\n\t"
+        "test %%ecx, %%ecx\n\t"
+        "jz storeDst%=\n\t"
+
+        "leal (%%eax,%%eax,0x2), %%ecx\n\t"
+        "paddd 0x00(%[dst_ptr])           , %%xmm4 \n\t"
+        "paddd 0x00(%[dst_ptr], %%eax, 1) , %%xmm5 \n\t"
+        "paddd 0x00(%[dst_ptr], %%eax, 2) , %%xmm6 \n\t"
+        "paddd 0x00(%[dst_ptr], %%ecx, 1) , %%xmm7 \n\t"
+
+        "storeDst%=:\n\t"
+
+        "leal (%%eax,%%eax,0x2), %%ecx\n\t"
+        "movdqu %%xmm4  , 0x00(%[dst_ptr])          \n\t"
+        "movdqu %%xmm5  , 0x00(%[dst_ptr], %%eax, 1)\n\t"
+        "movdqu %%xmm6  , 0x00(%[dst_ptr], %%eax, 2)\n\t"
+        "movdqu %%xmm7  , 0x00(%[dst_ptr], %%ecx, 1)\n\t"
+
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr)
+        :  // inputs
+        [start_depth] "g"(start_depth), [dst_col_stride] "g"(dst_col_stride),
+        [run_depth_cells] "g"(run_depth_cells)
+        :  // clobbers
+        "cc", "memory", "%xmm0", "%xmm1", "%xmm3", "%xmm2", "%xmm4", "%xmm5",
+        "%xmm6", "%xmm7", "%eax", "%ecx");
+  }
+};
+#endif
+#ifdef GEMMLOWP_SSE4_64
+struct SSE4_64_Kernel12x4Depth2 : KernelBase {
+  typedef KernelFormat<
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 3>,
+      KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, 1> >
+      Format;
+
+  const char* Name() const override { return "SSE, 12x4, depth 2"; }
+
+  void Run(std::int32_t* dst_ptr, std::size_t dst_row_stride,
+           std::size_t dst_col_stride, const std::uint8_t* lhs_ptr,
+           const std::uint8_t* rhs_ptr, std::size_t start_depth,
+           std::size_t run_depth) const override {
+    ScopedProfilingLabel label("optimized kernel");
+    assert(dst_row_stride == 1);
+    const std::int64_t run_depth_cells = run_depth / Format::kDepth;
+    const std::int64_t dst_col_stride_q = dst_col_stride;
+
+    /* Main loop */
+
+    // A 2x4 cell of Rhs is stored in 16bit in xmm1 .
+    // A 12x2 block of 3 4x2 cells Lhs is stored in 16bit in xmm0, replaced
+    // every Iteration.
+    // A 12x4 block of accumulators is stored in 32bit in xmm4--xmm15.
+    //
+    //                   +-------+-------+-------+-------+
+    //                   |xmm1[0]|xmm1[2]|xmm1[4]|xmm1[6]|
+    //              Rhs  +-------+---------------+-------+
+    //                   |xmm1[1]|xmm1[3]|xmm1[5]|xmm1[7]|
+    //                   +-------+-------+-------+-------+
+    //
+    //                   |       |       |       |       |
+    //
+    //    Lhs            |       |       |       |       |
+    //
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 | (Iter1)  | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  |xmm0 |          | xmm4  | xmm5  | xmm6  | xmm7  |
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //  |xmm0 |          | xmm8  | xmm9  | xmm10 | xmm11 |
+    //  |xmm0 | (Iter2)  | xmm8  | xmm9  | xmm10 | xmm11 |
+    //  |xmm0 |          | xmm8  | xmm9  | xmm10 | xmm11 |
+    //  |xmm0 |          | xmm8  | xmm9  | xmm10 | xmm11 |
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //  |xmm0 |          | xmm12 | xmm13 | xmm14 | xmm15 |
+    //  |xmm0 | (Iter3)  | xmm12 | xmm13 | xmm14 | xmm15 |
+    //  |xmm0 |          | xmm12 | xmm13 | xmm14 | xmm15 |
+    //  |xmm0 |          | xmm12 | xmm13 | xmm14 | xmm15 |
+    //  +--+--+ - - - -  +-------+-------+-------+-------+
+    //
+    //                              Accumulator
+
+    asm volatile(
+
+        // Set registers for destination
+        "movq  %[dst_col_stride_q], %%r12\n\t"
+        "shlq $2, %%r12\n\t"
+        "leaq (%%r12,%%r12,0x2), %%r13\n\t"
+
+        // Set accumulators to zero.
+        "pxor %%xmm4  , %%xmm4 \n\t"
+        "pxor %%xmm5  , %%xmm5 \n\t"
+        "pxor %%xmm6  , %%xmm6 \n\t"
+        "pxor %%xmm7  , %%xmm7 \n\t"
+        "pxor %%xmm8  , %%xmm8 \n\t"
+        "pxor %%xmm9  , %%xmm9 \n\t"
+        "pxor %%xmm10 , %%xmm10\n\t"
+        "pxor %%xmm11 , %%xmm11\n\t"
+        "pxor %%xmm12 , %%xmm12\n\t"
+        "pxor %%xmm13 , %%xmm13\n\t"
+        "pxor %%xmm14 , %%xmm14\n\t"
+        "pxor %%xmm15 , %%xmm15\n\t"
+
+        "movq  %[run_depth_cells], %%r14\n\t"
+        "subq $2, %%r14\n\t"
+        "js outerLoop1%=\n\t"
+
+        // Loop for K unrolled by 4
+        "outerLoop2%=:\n\t"
+
+        // K = 1,2
+        // RHS cell to xmm1
+
+        "pmovzxbw (%[rhs_ptr]), %%xmm1\n\t"
+
+        // LHS cell
+        "pmovzxbw 0x00(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+
+        "prefetcht0 0x80(%[lhs_ptr]) \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x08(%[lhs_ptr]), %%xmm0\n\t"
+
+        "paddd %%xmm2, %%xmm6           \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm8           \n\t"
+        "paddd %%xmm3, %%xmm9           \n\t"
+
+        "prefetcht0 0x80(%[rhs_ptr]) \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm10          \n\t"
+        "paddd %%xmm3, %%xmm11          \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x10(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm12          \n\t"
+        "paddd %%xmm3, %%xmm13          \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm14          \n\t"
+        "paddd %%xmm3, %%xmm15          \n\t"
+
+        // K = 3,4
+        // RHS cell to xmm1
+        "pmovzxbw 0x08(%[rhs_ptr]), %%xmm1\n\t"
+
+        // LHS cell
+        "pmovzxbw 0x18(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm6           \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x20(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm8           \n\t"
+        "paddd %%xmm3, %%xmm9           \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm10          \n\t"
+        "paddd %%xmm3, %%xmm11          \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x28(%[lhs_ptr]), %%xmm0\n\t"
+
+        "addq $0x30, %[lhs_ptr]         \n\t"
+        "addq $0x10, %[rhs_ptr]         \n\t"
+
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm12          \n\t"
+        "paddd %%xmm3, %%xmm13          \n\t"
+
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm14          \n\t"
+        "paddd %%xmm3, %%xmm15          \n\t"
+
+        "subq $2, %[run_depth_cells]\n\t"
+        "ja outerLoop2%=\n\t"
+
+        "movq %[run_depth_cells], %%r14\n\t"
+        "decq %%r14\n\t"
+        "js finish%=\n\t"
+
+        // Loop for K unrolled by 2
+        "outerLoop1%=:\n\t"
+
+        // RHS cell to xmm1
+        "pmovzxbw (%[rhs_ptr]), %%xmm1\n\t"
+
+        // LHS cell
+        "pmovzxbw 0x00(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm4           \n\t"
+        "paddd %%xmm3, %%xmm5           \n\t"
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm6           \n\t"
+        "paddd %%xmm3, %%xmm7           \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x08(%[lhs_ptr]), %%xmm0\n\t"
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm8           \n\t"
+        "paddd %%xmm3, %%xmm9           \n\t"
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm10          \n\t"
+        "paddd %%xmm3, %%xmm11          \n\t"
+
+        // next LHS cell
+        "pmovzxbw 0x10(%[lhs_ptr]), %%xmm0\n\t"
+
+        "addq $0x18, %[lhs_ptr]         \n\t"
+        "addq $0x08, %[rhs_ptr]         \n\t"
+
+        "pshufd $0x00,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0x55,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm12          \n\t"
+        "paddd %%xmm3, %%xmm13          \n\t"
+        "pshufd $0xaa,%%xmm1,%%xmm2     \n\t"
+        "pshufd $0xff,%%xmm1,%%xmm3     \n\t"
+        "pmaddwd %%xmm0, %%xmm2         \n\t"
+        "pmaddwd %%xmm0, %%xmm3         \n\t"
+        "paddd %%xmm2, %%xmm14          \n\t"
+        "paddd %%xmm3, %%xmm15          \n\t"
+
+        "decq %[run_depth_cells]\n\t"
+        "jnz outerLoop1%=\n\t"
+
+        "finish%=:\n\t"
+
+        "test %[start_depth], %[start_depth]\n\t"
+        "jz storeDst%=\n\t"
+
+        "paddd 0x00(%[dst_ptr])           , %%xmm4 \n\t"
+        "paddd 0x10(%[dst_ptr])           , %%xmm8 \n\t"
+        "paddd 0x20(%[dst_ptr])           , %%xmm12\n\t"
+        "paddd 0x00(%[dst_ptr], %%r12, 1) , %%xmm5 \n\t"
+        "paddd 0x10(%[dst_ptr], %%r12, 1) , %%xmm9 \n\t"
+        "paddd 0x20(%[dst_ptr], %%r12, 1) , %%xmm13\n\t"
+        "paddd 0x00(%[dst_ptr], %%r12, 2) , %%xmm6 \n\t"
+        "paddd 0x10(%[dst_ptr], %%r12, 2) , %%xmm10\n\t"
+        "paddd 0x20(%[dst_ptr], %%r12, 2) , %%xmm14\n\t"
+        "paddd 0x00(%[dst_ptr], %%r13, 1) , %%xmm7 \n\t"
+        "paddd 0x10(%[dst_ptr], %%r13, 1) , %%xmm11\n\t"
+        "paddd 0x20(%[dst_ptr], %%r13, 1) , %%xmm15\n\t"
+
+        "storeDst%=:\n\t"
+
+        "movdqu %%xmm4  , 0x00(%[dst_ptr])          \n\t"
+        "movdqu %%xmm8  , 0x10(%[dst_ptr])          \n\t"
+        "movdqu %%xmm12 , 0x20(%[dst_ptr])          \n\t"
+        "movdqu %%xmm5  , 0x00(%[dst_ptr], %%r12, 1)\n\t"
+        "movdqu %%xmm9  , 0x10(%[dst_ptr], %%r12, 1)\n\t"
+        "movdqu %%xmm13 , 0x20(%[dst_ptr], %%r12, 1)\n\t"
+        "movdqu %%xmm6  , 0x00(%[dst_ptr], %%r12, 2)\n\t"
+        "movdqu %%xmm10 , 0x10(%[dst_ptr], %%r12, 2)\n\t"
+        "movdqu %%xmm14 , 0x20(%[dst_ptr], %%r12, 2)\n\t"
+        "movdqu %%xmm7  , 0x00(%[dst_ptr], %%r13, 1)\n\t"
+        "movdqu %%xmm11 , 0x10(%[dst_ptr], %%r13, 1)\n\t"
+        "movdqu %%xmm15 , 0x20(%[dst_ptr], %%r13, 1)\n\t"
+
+        :  // outputs
+        [lhs_ptr] "+r"(lhs_ptr), [rhs_ptr] "+r"(rhs_ptr),
+        [dst_ptr] "+r"(dst_ptr)
+        :  // inputs
+        [start_depth] "r"(start_depth),
+        [dst_col_stride_q] "r"(dst_col_stride_q),
+        [run_depth_cells] "r"(run_depth_cells)
+        :  // clobbers
+        "cc", "memory", "%xmm0", "%xmm1", "%xmm3", "%xmm2", "%xmm4", "%xmm5",
+        "%xmm6", "%xmm7", "%xmm8", "%xmm9", "%xmm10", "%r12", "%r13", "%r14",
+        "%xmm11", "%xmm12", "%xmm13", "%xmm14", "%xmm15");
+  }
+};
+#endif
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_KERNEL_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/multi_thread_gemm.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/multi_thread_gemm.h
new file mode 100644
index 0000000..0234b26
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/multi_thread_gemm.h
@@ -0,0 +1,701 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// multi_thread_gemm.h: Multi-threaded GEMM entry point.
+// Readers note: To understand this file, it is useful to first
+// read and understand the much simpler single_thread_gemm.h.
+
+#ifndef GEMMLOWP_INTERNAL_MULTI_THREAD_GEMM_H_
+#define GEMMLOWP_INTERNAL_MULTI_THREAD_GEMM_H_
+
+#include <pthread.h>
+#include <unistd.h>
+#include <vector>
+
+#include "single_thread_gemm.h"
+
+namespace gemmlowp {
+
+// On X86 and ARM platforms we enable a busy-wait spinlock before waiting on a
+// pthread conditional variable. In order to implement that correctly we need
+// to put some explicit memory load/store barriers.
+
+#if defined(GEMMLOWP_ALLOW_INLINE_ASM) && !defined(GEMMLOWP_NO_BUSYWAIT) && \
+    (defined(GEMMLOWP_ARM) || defined(GEMMLOWP_X86))
+
+#define GEMMLOWP_USE_BUSYWAIT
+
+const int kMaxBusyWaitNOPs = 32 * 1000 * 1000;
+
+#define GEMMLOWP_NOP "nop\n"
+
+#define GEMMLOWP_STRING_CONCAT_4(X) X X X X
+#define GEMMLOWP_NOP4 GEMMLOWP_STRING_CONCAT_4(GEMMLOWP_NOP)
+#define GEMMLOWP_NOP16 GEMMLOWP_STRING_CONCAT_4(GEMMLOWP_NOP4)
+#define GEMMLOWP_NOP64 GEMMLOWP_STRING_CONCAT_4(GEMMLOWP_NOP16)
+
+inline int Do256NOPs() {
+  asm volatile(GEMMLOWP_NOP64);
+  return 64;
+}
+
+#undef GEMMLOWP_STRING_CONCAT_4
+#undef GEMMLOWP_NOP256
+#undef GEMMLOWP_NOP64
+#undef GEMMLOWP_NOP16
+#undef GEMMLOWP_NOP4
+#undef GEMMLOWP_NOP
+
+inline void WriteBarrier() {
+#ifdef GEMMLOWP_ARM_32
+  MemoryBarrier();
+#elif defined(GEMMLOWP_ARM_64)
+  asm volatile("dmb ishst" ::: "memory");
+#elif defined(GEMMLOWP_X86)
+  asm volatile("sfence" ::: "memory");
+#else
+#error "Unsupported architecture for WriteBarrier."
+#endif
+}
+
+inline void ReadBarrier() {
+#ifdef GEMMLOWP_ARM_32
+  MemoryBarrier();
+#elif defined(GEMMLOWP_ARM_64)
+  asm volatile("dmb ishld" ::: "memory");
+#elif defined(GEMMLOWP_X86)
+  asm volatile("lfence" ::: "memory");
+#else
+#error "Unsupported architecture for ReadBarrier."
+#endif
+}
+
+#endif
+
+// Waits until *var != initial_value.
+//
+// Returns the new value of *var. The guarantee here is that
+// the return value is different from initial_value, and that that
+// new value has been taken by *var at some point during the
+// execution of this function. There is no guarantee that this is
+// still the value of *var when this function returns, since *var is
+// not assumed to be guarded by any lock.
+//
+// First does some busy-waiting for a fixed number of no-op cycles,
+// then falls back to passive waiting for the given condvar, guarded
+// by the given mutex.
+//
+// The idea of doing some initial busy-waiting is to help get
+// better and more consistent multithreading benefits for small GEMM sizes.
+// Busy-waiting help ensuring that if we need to wake up soon after having
+// started waiting, then we can wake up quickly (as opposed to, say,
+// having to wait to be scheduled again by the OS). On the other hand,
+// we must still eventually revert to passive waiting for longer waits
+// (e.g. worker threads having finished a GEMM and waiting until the next GEMM)
+// so as to avoid permanently spinning.
+//
+template <typename T>
+T WaitForVariableChange(volatile T* var, T initial_value, pthread_cond_t* cond,
+                        pthread_mutex_t* mutex) {
+#ifdef GEMMLOWP_USE_BUSYWAIT
+  // If we are on a platform that supports it, spin for some time.
+  {
+    int nops = 0;
+    // First, trivial case where the variable already changed value.
+    T new_value = *var;
+    if (new_value != initial_value) {
+      ReadBarrier();
+      return new_value;
+    }
+    // Then try busy-waiting.
+    while (nops < kMaxBusyWaitNOPs) {
+      nops += Do256NOPs();
+      new_value = *var;
+      if (new_value != initial_value) {
+        ReadBarrier();
+        return new_value;
+      }
+    }
+  }
+#endif
+
+  // Finally, do real passive waiting.
+  pthread_mutex_lock(mutex);
+  T new_value = *var;
+  if (new_value == initial_value) {
+    pthread_cond_wait(cond, mutex);
+    new_value = *var;
+    assert(new_value != initial_value);
+  }
+  pthread_mutex_unlock(mutex);
+  return new_value;
+}
+
+// A BlockingCounter lets one thread to wait for N events to occur.
+// This is how the master thread waits for all the worker threads
+// to have finished working.
+class BlockingCounter {
+ public:
+  BlockingCounter()
+      : cond_(PTHREAD_COND_INITIALIZER),
+        mutex_(PTHREAD_MUTEX_INITIALIZER),
+        count_(0),
+        initial_count_(0) {}
+
+  // Sets/resets the counter; initial_count is the number of
+  // decrementing events that the Wait() call will be waiting for.
+  void Reset(std::size_t initial_count) {
+    pthread_mutex_lock(&mutex_);
+    assert(count_ == 0);
+    initial_count_ = initial_count;
+    count_ = initial_count_;
+    pthread_mutex_unlock(&mutex_);
+  }
+
+  // Decrements the counter; if the counter hits zero, signals
+  // the thread that was waiting for that, and returns true.
+  // Otherwise (if the decremented count is still nonzero),
+  // returns false.
+  bool DecrementCount() {
+    pthread_mutex_lock(&mutex_);
+    assert(count_ > 0);
+    count_--;
+#ifdef GEMMLOWP_USE_BUSYWAIT
+    WriteBarrier();
+#endif
+    if (count_ == 0) {
+      pthread_cond_signal(&cond_);
+    }
+    bool retval = count_ == 0;
+    pthread_mutex_unlock(&mutex_);
+    return retval;
+  }
+
+  // Waits for the N other threads (N having been set by Reset())
+  // to hit the BlockingCounter.
+  void Wait() {
+    ScopedProfilingLabel label("BlockingCounter::Wait");
+    while (count_) {
+      MemoryBarrier();
+      const std::size_t count_value = count_;
+      if (count_value) {
+        WaitForVariableChange(&count_, count_value, &cond_, &mutex_);
+      }
+    }
+  }
+
+ private:
+  pthread_cond_t cond_;
+  pthread_mutex_t mutex_;
+  std::size_t count_;
+  std::size_t initial_count_;
+};
+
+// A workload for a worker.
+struct Task {
+  Task() : local_allocator(nullptr) {}
+  virtual ~Task() {}
+  virtual void Run() = 0;
+  Allocator* local_allocator;
+};
+
+// A worker thread.
+class Worker {
+ public:
+  enum class State {
+    ThreadStartup,  // The initial state before the thread main loop runs.
+    Ready,          // Is not working, has not yet received new work to do.
+    HasWork,        // Has work to do.
+    ExitAsSoonAsPossible  // Should exit at earliest convenience.
+  };
+
+  explicit Worker(BlockingCounter* counter_to_decrement_when_ready)
+      : task_(nullptr),
+        state_cond_(PTHREAD_COND_INITIALIZER),
+        state_mutex_(PTHREAD_MUTEX_INITIALIZER),
+        state_(State::ThreadStartup),
+        counter_to_decrement_when_ready_(counter_to_decrement_when_ready) {
+    pthread_create(&thread_, nullptr, ThreadFunc, this);
+  }
+
+  ~Worker() {
+    ChangeState(State::ExitAsSoonAsPossible);
+    pthread_join(thread_, nullptr);
+  }
+
+  // Changes State; may be called from either the worker thread
+  // or the master thread; however, not all state transitions are legal,
+  // which is guarded by assertions.
+  void ChangeState(State new_state) {
+    ScopedProfilingLabel label("Worker::ChangeState");
+    pthread_mutex_lock(&state_mutex_);
+    assert(new_state != state_);
+    switch (state_) {
+      case State::ThreadStartup:
+        assert(new_state == State::Ready);
+        break;
+      case State::Ready:
+        assert(new_state == State::HasWork ||
+               new_state == State::ExitAsSoonAsPossible);
+        break;
+      case State::HasWork:
+        assert(new_state == State::Ready ||
+               new_state == State::ExitAsSoonAsPossible);
+        break;
+      default:
+        abort();
+    }
+    state_ = new_state;
+    pthread_cond_signal(&state_cond_);
+    if (state_ == State::Ready) {
+      counter_to_decrement_when_ready_->DecrementCount();
+    }
+    pthread_mutex_unlock(&state_mutex_);
+  }
+
+  // Thread entry point.
+  void ThreadFunc() {
+    ScopedProfilingLabel label("Worker::ThreadFunc");
+    RegisterCurrentThreadForProfiling();
+
+    ChangeState(State::Ready);
+
+    // Thread main loop
+    while (true) {
+      // Get a state to act on
+      // In the 'Ready' state, we have nothing to do but to wait until
+      // we switch to another state.
+      State state_to_act_upon = WaitForVariableChange(
+          &state_, State::Ready, &state_cond_, &state_mutex_);
+
+      // We now have a state to act on, so act.
+      switch (state_to_act_upon) {
+        case State::HasWork:
+          // Got work to do! So do it, and then revert to 'Ready' state.
+          assert(task_);
+          task_->Run();
+          task_ = nullptr;
+          ChangeState(State::Ready);
+          break;
+        case State::ExitAsSoonAsPossible:
+          return;
+        default:
+          abort();
+      }
+    }
+  }
+
+  static void* ThreadFunc(void* arg) {
+    static_cast<Worker*>(arg)->ThreadFunc();
+    return nullptr;
+  }
+
+  // Called by the master thead to give this worker work to do.
+  // It is only legal to call this if the worker
+  void StartWork(Task* task) {
+    assert(!task_);
+    task->local_allocator = &local_allocator_;
+    task_ = task;
+#ifdef GEMMLOWP_USE_BUSYWAIT
+    WriteBarrier();
+#endif
+    assert(state_ == State::Ready);
+    ChangeState(State::HasWork);
+  }
+
+ private:
+  // The underlying thread.
+  pthread_t thread_;
+
+  // The task to be worked on.
+  Task* task_;
+
+  // The condition variable and mutex guarding state changes.
+  pthread_cond_t state_cond_;
+  pthread_mutex_t state_mutex_;
+
+  // The state enum tells if we're currently working, waiting for work, etc.
+  State state_;
+
+  // Each thread had a local allocator so they can allocate temporary
+  // buffers without blocking each other.
+  Allocator local_allocator_;
+
+  // pointer to the master's thread BlockingCounter object, to notify the
+  // master thread of when this worker switches to the 'Ready' state.
+  BlockingCounter* const counter_to_decrement_when_ready_;
+};
+
+// A very simple pool of workers, that only allows the very
+// specific parallelization pattern that we use here:
+// a fixed number of workers can be given work, and one then
+// waits for all of them to finish.
+//
+// See MultiThreadGemmContextBase for how other WorkersPool implementations can
+// be used. Note that in those implementations, StartWorker can be free to
+// ignore the <index> value; that is, the caller of WorkersPool does not rely on
+// <index> to order tasks with equal <index>.
+class WorkersPool {
+ public:
+  WorkersPool() {}
+
+  ~WorkersPool() {
+    for (auto w : workers_) {
+      delete w;
+    }
+  }
+
+  void Execute(const std::vector<Task*>& tasks) {
+    assert(tasks.size() >= 1);
+    // One of the tasks will be run on the current thread.
+    int workers_count = tasks.size() - 1;
+    CreateWorkers(workers_count);
+    assert(workers_count <= workers_.size());
+    counter_to_decrement_when_ready_.Reset(workers_count);
+    int n = 0;
+    std::for_each(tasks.begin(), --tasks.end(), [this, &n](Task *task) {
+      workers_[n++]->StartWork(task);
+    });
+    // Execute the remaining workload immediately on the current thread.
+    Task* task = tasks.back();
+    task->local_allocator = &main_thread_task_allocator_;
+    task->Run();
+    // Wait for the workers submitted above to finish.
+    counter_to_decrement_when_ready_.Wait();
+    // Cleanup tasks (best to do this from the same thread that allocated
+    // the memory).
+    std::for_each(tasks.begin(), tasks.end(), [](Task *task) {
+      delete task;
+    });
+  }
+
+ private:
+  // Ensures that the pool has at least the given count of workers.
+  // If any new worker has to be created, this function waits for it to
+  // be ready.
+  void CreateWorkers(std::size_t workers_count) {
+    if (workers_.size() >= workers_count) {
+      return;
+    }
+    counter_to_decrement_when_ready_.Reset(workers_count - workers_.size());
+    while (workers_.size() < workers_count) {
+      workers_.push_back(new Worker(&counter_to_decrement_when_ready_));
+    }
+    counter_to_decrement_when_ready_.Wait();
+  }
+
+  // copy construction disallowed
+  WorkersPool(const WorkersPool&) = delete;
+
+  // The workers in this pool. They are owned by the pool:
+  // the pool creates workers and destroys them in its destructor.
+  std::vector<Worker*> workers_;
+
+  // The BlockingCounter used to wait for the workers.
+  BlockingCounter counter_to_decrement_when_ready_;
+
+  // For N-threaded operations, we will use only N-1 worker threads
+  // while the last task will be run directly on the main thread.
+  // It will then use this main_thread_task_allocator_; having a
+  // dedicated allocator for that (separate from the base allocator_)
+  // allows to use the same code for all tasks regardless of which
+  // thread they run on.
+  Allocator main_thread_task_allocator_;
+};
+
+// The task we use to implement a multi-threaded Gemm: a block of the
+// RHS has been packed by the master thread; each worker thread
+// then has to pack a block of the LHS and accumulate the Gemm of these
+// packed LHS and RHS blocks.
+template <typename KernelFormat, typename InputScalar, typename OutputScalar,
+          typename BitDepthParams, MapOrder LhsOrder, MapOrder RhsOrder,
+          MapOrder ResultOrder, typename LhsOffset, typename RhsOffset,
+  typename OutputPipelineType, typename GemmContextType>
+struct GemmWithPackedRhsTask : Task {
+  typedef PackedSideBlock<typename KernelFormat::Lhs> PackedLhs;
+  typedef PackedSideBlock<typename KernelFormat::Rhs> PackedRhs;
+  GemmWithPackedRhsTask(GemmContextType* _context,
+                        const KernelBase& _kernel,
+                        const MatrixMap<const InputScalar, LhsOrder>& _lhs,
+                        const PackedRhs& _packed_rhs,
+                        MatrixMap<OutputScalar, ResultOrder>* _result,
+                        const MatrixBlockBounds& _result_block,
+                        const LhsOffset& _lhs_offset,
+                        const RhsOffset& _rhs_offset,
+                        const OutputPipelineType& _output_pipeline)
+      : context(_context),
+        kernel(_kernel),
+        lhs(_lhs),
+        packed_rhs(_packed_rhs),
+        result(*_result),
+        result_block(_result_block),
+        lhs_offset(_lhs_offset),
+        rhs_offset(_rhs_offset),
+        output_pipeline(_output_pipeline) {}
+
+  void Run() override {
+    ScopedProfilingLabel label("GemmWithPackedRhsTask");
+
+    const int rows = result_block.rows;
+    const int cols = result_block.cols;
+    const int depth = lhs.cols();
+
+    BlockParams block_params;
+    block_params.Init<KernelFormat>(rows, cols, depth, 1,
+                                    context->l1_bytes_to_use(),
+                                    context->l2_bytes_to_use(),
+                                    context->l2_rhs_factor());
+
+    PackedLhs packed_lhs(Side::Lhs, local_allocator, block_params);
+
+    PackedResult packed_result(local_allocator, block_params);
+
+    local_allocator->Commit();
+
+    for (int c = 0; c < cols; c += block_params.l2_cols) {
+      int cs = std::min(block_params.l2_cols, cols - c);
+
+      for (int r = 0; r < rows; r += block_params.l2_rows) {
+        int rs = std::min(block_params.l2_rows, rows - r);
+
+        PackLhs(&packed_lhs, lhs.block(r, 0, rs, depth));
+
+        Compute(kernel, block_params, &packed_result, packed_lhs, packed_rhs,
+                depth);
+
+        auto curr_result_block = MatrixBlockBounds(
+            result_block.start_row + r, result_block.start_col + c, rs, cs);
+        UnpackResult<KernelFormat>(
+            &result, curr_result_block, packed_result, depth,
+            packed_lhs.sums_of_each_slice(), packed_rhs.sums_of_each_slice(),
+            lhs_offset.block(curr_result_block.start_row, rs),
+            rhs_offset.block(curr_result_block.start_col, cs), output_pipeline);
+      }
+    }
+
+    local_allocator->Decommit();
+  }
+
+  const GemmContextType* context;
+  const KernelBase& kernel;
+  const MatrixMap<const InputScalar, LhsOrder> lhs;
+  const PackedRhs packed_rhs;
+  MatrixMap<OutputScalar, ResultOrder> result;
+  const MatrixBlockBounds result_block;
+  const LhsOffset& lhs_offset;
+  const RhsOffset& rhs_offset;
+  const OutputPipelineType& output_pipeline;
+};
+
+// This base class for multi-threading allows subclasses to implement their own
+// workers_pool() method.  See MultiThreadGemmContext below for an example;
+// any other implementation of workers_pool() must return an object with the
+// same public methods as WorkersPool.
+class MultiThreadGemmContextBase : public SingleThreadGemmContext {
+ public:
+  void set_max_num_threads(int n) { max_num_threads_ = n; }
+
+  int max_num_threads() const { return max_num_threads_; }
+
+ protected:
+  // The maximum number of worker threads to use (including
+  // the master thread).
+  // The default value 1 means single-threading. That is the default
+  // because gemmlowp's primary target is mobile hardware, where thermal
+  // constraints usually mean that it may not be realistic to use more
+  // than 1 CPU core even if multiple cores are present.
+  // The special value 0 means try to detect the number of hardware threads.
+  // Note: this assumes that all CPU cores are equivalent. That assumption
+  // is defeated on big.LITTLE ARM devices, where we have no API to query
+  // the number of big cores (which is typically what we would want to use,
+  // leaving aside above-mentioned thermal issues). That is the other reason
+  // why the best compromise here is to let max_num_threads_ default to 1,
+  // so users who want multi-threading have to make the decision of how many
+  // threads to use by themselves.
+  int max_num_threads_ = 1;
+};
+
+class MultiThreadGemmContext : public MultiThreadGemmContextBase {
+ public:
+  WorkersPool* workers_pool() { return &workers_pool_; }
+
+ private:
+  // The workers pool used by MultiThreadGemm. Making
+  // this part of the context allows it to be persistent,
+  // avoiding recreating threads on every Gemm.
+  WorkersPool workers_pool_;
+};
+
+// Needed by chrome native builds
+#ifndef _SC_NPROCESSORS_CONF
+#define _SC_NPROCESSORS_CONF _SC_NPROCESSORS_ONLN
+#endif
+
+// Determines how many threads should be used for a given Gemm
+// operation.
+template <int KernelRows>
+inline int HowManyThreads(int max_num_threads, int rows, int cols, int depth) {
+  // Early-exit in the default case where multi-threading is disabled.
+  if (max_num_threads == 1) {
+    return 1;
+  }
+
+  // Determine the maximum number of threads.
+  int max_count = max_num_threads;
+  // The special value 0 means try to determine the total number of cores.
+  if (max_count == 0) {
+    // No user-set maximum number of threads, so we need to
+    // do some hardware detection.
+    // This is expensive to query so we do it only once.
+    // Too bad for dynamicness. Also, we dont use the c++11 standard getter
+    // because Google's coding style currently bans #include <thread_>.
+    static const int hardware_threads_count =
+        static_cast<int>(sysconf(_SC_NPROCESSORS_CONF));
+
+    max_count = hardware_threads_count;
+  }
+
+  // Basic calculation: take into account max pool size, and
+  // how many rows we have to feed our kernel.
+  // The motivation for an absolute minimum number of rows per thread,
+  // potentially higher than KernelRows, is that very thin thread workload
+  // currently defeat assumptions of the AddMod generator, resulting
+  // in substantial bias in TestWithRealData on 24 threads.
+  // Ideally, the AddMod generator should be aware of global (r,c) coordinates
+  // so as to be independent of the number of threads.
+  static const int AbsoluteMinRowsPerThread = 16;
+  static const int MinRowsPerThread = KernelRows > AbsoluteMinRowsPerThread
+                                          ? KernelRows
+                                          : AbsoluteMinRowsPerThread;
+  int thread_count = std::min(max_count, CeilQuotient(rows, MinRowsPerThread));
+
+  // At this point for small products we already have thread_count==1 so
+  // we can avoid doing more work; otherwise, we still want to check
+  // that the cubic size (rows*cols*depth) is big enough to keep
+  // workers_ busy.
+  if (thread_count > 1) {
+    // Empirically determined value.
+    static const std::uint64_t min_cubic_size_per_thread = 64 * 1024;
+
+    // We can only multiply two out of three sizes without risking overflow
+    const std::uint64_t cubic_size =
+        std::uint64_t(rows) * std::uint64_t(cols) * std::uint64_t(depth);
+
+    thread_count =
+        std::min(thread_count, int(cubic_size / min_cubic_size_per_thread));
+
+    if (thread_count < 1) {
+      thread_count = 1;
+    }
+  }
+
+  assert(thread_count > 0 && thread_count <= max_count);
+  return thread_count;
+}
+
+// The main multi-threaded Gemm function.
+// To understand it, first read the code of SingleThreadGemm().
+// The parallelization scheme used here is to have this master function
+// pack a block of RHS and then start worker threads to pack a block of LHS
+// each, and accumulate the corresponding products.
+template <typename KernelFormat, typename InputScalar, typename OutputScalar,
+          typename BitDepthParams, MapOrder LhsOrder, MapOrder RhsOrder,
+          MapOrder ResultOrder, typename LhsOffset, typename RhsOffset,
+          typename OutputPipelineType, typename GemmContextType>
+void MultiThreadGemm(GemmContextType* context, const KernelBase& kernel,
+                     const MatrixMap<const InputScalar, LhsOrder>& lhs,
+                     const MatrixMap<const InputScalar, RhsOrder>& rhs,
+                     MatrixMap<OutputScalar, ResultOrder>* result,
+                     const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
+                     const OutputPipelineType& output_pipeline) {
+  ScopedProfilingLabel label("gemmlowp::MultiThreadGemm");
+
+  assert(lhs.cols() == rhs.rows());
+
+  int rows = result->rows();
+  int cols = result->cols();
+  int depth = lhs.cols();
+
+  // zero sizes should have been caught earlier and early-returned.
+  assert(rows > 0);
+  assert(cols > 0);
+  assert(depth > 0);
+
+  // The case of rows<cols should have been caught earlier and transposed.
+  assert(rows >= cols);
+
+  const int thread_count = HowManyThreads<KernelFormat::kRows>(
+      context->max_num_threads(), rows, cols, depth);
+  if (thread_count == 1) {
+    return SingleThreadGemm<KernelFormat, InputScalar, OutputScalar,
+                            BitDepthParams>(context, kernel, lhs, rhs, result,
+                                            lhs_offset, rhs_offset,
+                                            output_pipeline);
+  }
+  assert(thread_count > 1);
+
+  // Simple 1:1 mapping of tasks to physical cores, which is very important
+  // to getting good multithreaded performance, specially for not-very-large
+  // GEMMs, and especially on Android.
+  const int task_count = thread_count;
+
+  Allocator* allocator = context->allocator();
+  auto* workers_pool = context->workers_pool();
+
+  BlockParams block_params;
+  block_params.Init<KernelFormat>(rows, cols, depth, task_count,
+                                  context->l1_bytes_to_use(),
+                                  context->l2_bytes_to_use(),
+                                  context->l2_rhs_factor());
+
+  PackedSideBlock<typename KernelFormat::Rhs> packed_rhs(Side::Rhs, allocator,
+                                                         block_params);
+  allocator->Commit();
+
+  // We loop over large blocks of the RHS.
+  for (int c = 0; c < cols; c += block_params.l2_cols) {
+    int cs = std::min(block_params.l2_cols, cols - c);
+
+    // Pack a large block of the RHS.
+    PackRhs(&packed_rhs, rhs.block(0, c, depth, cs));
+
+    // Give work to each worker.
+    std::vector<Task*> tasks;
+    int next_start_row = 0;
+    for (int n = 0; n < task_count; ++n) {
+      int start_row = next_start_row;
+      next_start_row = std::min(rows, RoundUp<KernelFormat::kRows>(
+                                          rows * (n + 1) / task_count));
+
+      int block_rows = next_start_row - start_row;
+      auto lhs_block = lhs.block(start_row, 0, block_rows, depth);
+      typedef GemmWithPackedRhsTask<
+          KernelFormat, InputScalar, OutputScalar, BitDepthParams, LhsOrder,
+          RhsOrder, ResultOrder, LhsOffset, RhsOffset, OutputPipelineType,
+          GemmContextType>
+          TaskType;
+      tasks.push_back(new TaskType(context, kernel, lhs_block, packed_rhs, result,
+                                   MatrixBlockBounds(start_row, c, block_rows, cs),
+                                   lhs_offset, rhs_offset, output_pipeline));
+    }
+    // Execute the work on the workers (and partially on this thread).
+    workers_pool->Execute(tasks);
+  }
+
+  allocator->Decommit();
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_MULTI_THREAD_GEMM_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/output.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output.h
new file mode 100644
index 0000000..8ccb8ee
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output.h
@@ -0,0 +1,435 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// output.h: processing the 32-bit accumulators output by the unpack
+// stage, obtaining the final result matrix entries and storing them into
+// the destination matrix.
+
+#ifndef GEMMLOWP_INTERNAL_OUTPUT_H_
+#define GEMMLOWP_INTERNAL_OUTPUT_H_
+
+#include <cmath>
+#include <tuple>
+#include <type_traits>
+
+#include "../fixedpoint/fixedpoint.h"
+#include "../public/output_stages.h"
+#include "simd_wrappers.h"
+
+namespace gemmlowp {
+
+template <typename OutputStage, typename InputBufferType>
+struct OutputStageEvalBufferImpl {
+  // This generic template body should never be hit.
+  static_assert(
+      std::is_same<InputBufferType, void>::value,
+      "Unimplemented: missing implementation of this output pipeline stage "
+      "for this data type. This would happen if some architecture-specific "
+      "SIMD back-end (output_$arch.h) were incomplete.");
+};
+
+template <typename OutputStage, typename InputType>
+struct OutputStageEvalImpl {
+  static constexpr int kRows = InputType::kRows;
+  static constexpr int kCols = InputType::kCols;
+  using InputBufferType = typename InputType::BufferType;
+  using BufferEvalImplType =
+      OutputStageEvalBufferImpl<OutputStage, InputBufferType>;
+  using OutputBufferType = typename BufferEvalImplType::OutputType;
+  using OutputScalarType = typename OutputBufferType::ScalarType;
+  using OutputType = RegisterBlock<OutputScalarType, kRows, kCols>;
+
+  OutputStageEvalImpl(const OutputStage& s) : buffer_eval_impl(s) {}
+
+  OutputType Eval(InputType input, int, int) const {
+    OutputType output;
+    output.buf = buffer_eval_impl.Eval(input.buf);
+    return output;
+  }
+
+  const BufferEvalImplType buffer_eval_impl;
+};
+
+template <int Size>
+struct OutputStageEvalBufferImpl<OutputStageQuantizeDownInt32ToUint8Scale,
+                                 RegisterBuffer<std::int32_t, Size>> {
+  using InputType = RegisterBuffer<std::int32_t, Size>;
+  using OutputType = RegisterBuffer<std::int32_t, Size>;
+
+  typedef OutputStageQuantizeDownInt32ToUint8Scale OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
+
+  OutputType Eval(InputType input) const {
+    const int result_shift = output_stage.result_shift;
+    const std::int32_t result_mult_int = output_stage.result_mult_int;
+    using RegisterType = typename InputType::RegisterType;
+    const RegisterType result_offset =
+        Dup<RegisterType>(output_stage.result_offset);
+    OutputType output;
+    for (int i = 0; i < InputType::kRegisterCount; i++) {
+      output.reg[i] = RoundingDivideByPOT(
+          Mul(Add(input.reg[i], result_offset), result_mult_int), result_shift);
+    }
+    return output;
+  }
+
+  const OutputStage& output_stage;
+};
+
+template <int Rows, int Cols, VectorShape Shape>
+struct OutputStageEvalImpl<OutputStageQuantizeDownInt32ToUint8ScalePC<Shape>,
+                           RegisterBlock<std::int32_t, Rows, Cols>> {
+  typedef RegisterBlock<std::int32_t, Rows, Cols> InputType;
+  typedef RegisterBlock<std::int32_t, Rows, Cols> OutputType;
+  typedef OutputStageQuantizeDownInt32ToUint8ScalePC<Shape> OutputStage;
+
+  OutputStageEvalImpl(const OutputStage& s) : output_stage(s) {}
+
+  OutputType Eval(InputType input, int row, int col) const {
+    OutputType output;
+    const int result_shift = output_stage.result_shift;
+    const int pos = Shape == VectorShape::Col ? row : col;
+    const auto result_mult_int =
+        LoadForBroadcasting<InputType>(output_stage.result_mult_int, pos);
+    const auto result_offset =
+        LoadForBroadcasting<InputType>(output_stage.result_offset, pos);
+    const auto dividend = BroadcastMul<InputType>(
+        BroadcastAdd<InputType>(input, result_offset), result_mult_int);
+    for (int i = 0; i < InputType::kRegisterCount; i++) {
+      output.buf.reg[i] =
+          RoundingDivideByPOT(dividend.buf.reg[i], result_shift);
+    }
+    return output;
+  }
+
+  const OutputStage& output_stage;
+};
+
+template <int Size>
+struct OutputStageEvalBufferImpl<
+    OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint,
+    RegisterBuffer<std::int32_t, Size>> {
+  typedef RegisterBuffer<std::int32_t, Size> InputType;
+  typedef RegisterBuffer<std::int32_t, Size> OutputType;
+
+  typedef OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    using RegisterType = typename InputType::RegisterType;
+    const RegisterType result_offset_after_shift =
+        Dup<RegisterType>(output_stage.result_offset_after_shift);
+    for (int i = 0; i < InputType::kRegisterCount; i++) {
+      const RegisterType mulhigh_val = SaturatingRoundingDoublingHighMul(
+          input.reg[i], output_stage.result_fixedpoint_multiplier);
+      output.reg[i] =
+          Add(RoundingDivideByPOT(mulhigh_val, output_stage.result_shift),
+              result_offset_after_shift);
+    }
+    return output;
+  }
+
+  const OutputStage& output_stage;
+};
+
+// Implementation of OutputStageSaturatingCastToUint8 for scalar data
+template <int Size>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegisterBuffer<std::int32_t, Size>> {
+  typedef RegisterBuffer<std::int32_t, Size> InputType;
+  typedef RegisterBuffer<std::uint8_t, Size> OutputType;
+  static_assert(InputType::kRegisterLanes == 1,
+                "This path is only for scalar values");
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    for (int i = 0; i < InputType::kRegisterCount; i++) {
+      std::int32_t data = input.reg[i];
+      output.reg[i] = data > 255 ? 255 : data < 0 ? 0 : data;
+    }
+    return output;
+  }
+};
+
+template <int Rows, int Cols, typename VectorType>
+struct OutputStageEvalImpl<OutputStageBiasAddition<VectorType>,
+                           RegisterBlock<std::int32_t, Rows, Cols>> {
+  typedef RegisterBlock<std::int32_t, Rows, Cols> InputType;
+  typedef RegisterBlock<std::int32_t, Rows, Cols> OutputType;
+  typedef OutputStageBiasAddition<VectorType> OutputStage;
+
+  OutputStageEvalImpl(const OutputStage& s) : output_stage(s) {}
+
+  OutputType Eval(InputType input, int row, int col) const {
+    const int pos = VectorType::kShape == VectorShape::Row ? col : row;
+    return BroadcastAdd<InputType>(
+        input, LoadForBroadcasting<InputType>(output_stage.bias_vector, pos));
+  }
+
+  const OutputStage& output_stage;
+};
+
+template <int Size>
+struct OutputStageEvalBufferImpl<OutputStageClamp,
+                                 RegisterBuffer<std::int32_t, Size>> {
+  typedef RegisterBuffer<std::int32_t, Size> InputType;
+  typedef RegisterBuffer<std::int32_t, Size> OutputType;
+
+  typedef OutputStageClamp OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {}
+
+  OutputType Eval(InputType input) const {
+    using RegisterType = typename InputType::RegisterType;
+    const RegisterType min = Dup<RegisterType>(output_stage.min);
+    const RegisterType max = Dup<RegisterType>(output_stage.max);
+    OutputType output;
+    for (int i = 0; i < InputType::kRegisterCount; i++) {
+      output.reg[i] = Min(Max(input.reg[i], min), max);
+    }
+    return output;
+  }
+
+  const OutputStage& output_stage;
+};
+
+template <int Size>
+struct OutputStageEvalBufferImpl<OutputStageTanh,
+                                 RegisterBuffer<std::int32_t, Size>> {
+  typedef RegisterBuffer<std::int32_t, Size> InputType;
+  typedef RegisterBuffer<std::int32_t, Size> OutputType;
+  using RegisterType = typename InputType::RegisterType;
+  typedef RegisterType DataType;
+  typedef OutputStageTanh OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage& s) : output_stage(s) {
+    const std::int32_t real_zero_as_int32 = output_stage.real_zero_as_int32;
+    const std::int32_t real_amplitude_as_int32 =
+        output_stage.real_amplitude_as_int32;
+
+    input_cutoff_min = real_zero_as_int32 - 8 * real_amplitude_as_int32;
+    input_cutoff_max = real_zero_as_int32 + 8 * real_amplitude_as_int32;
+    output_min = real_zero_as_int32 - real_amplitude_as_int32;
+    output_max = real_zero_as_int32 + real_amplitude_as_int32;
+
+    double inverse_amplitude_normalized_double = 1.0 / real_amplitude_as_int32;
+    inverse_amplitude_neg_exponent = 0;
+    while (inverse_amplitude_normalized_double < 0.5) {
+      inverse_amplitude_normalized_double *= 2;
+      inverse_amplitude_neg_exponent++;
+    }
+    inverse_amplitude_normalized = FixedPoint<DataType, 0>::FromDouble(
+        inverse_amplitude_normalized_double);
+
+    double amplitude_normalized_double = real_amplitude_as_int32;
+    amplitude_exponent = 0;
+    while (amplitude_normalized_double >= 1.0) {
+      amplitude_normalized_double *= 0.5;
+      amplitude_exponent++;
+    }
+    amplitude_normalized =
+        FixedPoint<DataType, 0>::FromDouble(amplitude_normalized_double);
+  }
+
+  OutputType Eval(InputType input) const {
+    const std::int32_t real_zero_as_int32 = output_stage.real_zero_as_int32;
+
+    typedef FixedPoint<DataType, 3> F3;
+    typedef FixedPoint<DataType, 0> F0;
+
+    OutputType output;
+
+    for (int i = 0; i < OutputType::kRegisterCount; i++) {
+      // fixed-point affine transformation
+      DataType input_centered =
+          Sub(input.reg[i], Dup<DataType>(real_zero_as_int32));
+      F3 fixedpoint_input =
+          F3::FromRaw(input_centered) * inverse_amplitude_normalized;
+      // left shift
+      fixedpoint_input.raw() = ShiftLeft(fixedpoint_input.raw(),
+                                         28 - inverse_amplitude_neg_exponent);
+      // fixed-point tanh and multiplication
+      F0 fixedpoint_output = tanh(fixedpoint_input) * amplitude_normalized;
+      // right shift
+      DataType int32_output =
+          Add(Dup<DataType>(real_zero_as_int32),
+              ShiftRight(fixedpoint_output.raw(), 31 - amplitude_exponent));
+
+      DataType mask_if_below_cutoff_min =
+          MaskIfLessThanOrEqual(input.reg[i], Dup<DataType>(input_cutoff_min));
+      DataType mask_if_above_cutoff_max = MaskIfGreaterThanOrEqual(
+          input.reg[i], Dup<DataType>(input_cutoff_max));
+
+      output.reg[i] = SelectUsingMask(
+          mask_if_below_cutoff_min, Dup<DataType>(output_min),
+          SelectUsingMask(mask_if_above_cutoff_max, Dup<DataType>(output_max),
+                          int32_output));
+    }
+    return output;
+  }
+
+  const OutputStage& output_stage;
+  std::int32_t input_cutoff_min, input_cutoff_max;
+  std::int32_t output_min, output_max;
+  FixedPoint<DataType, 0> inverse_amplitude_normalized;
+  int inverse_amplitude_neg_exponent;
+  FixedPoint<DataType, 0> amplitude_normalized;
+  int amplitude_exponent;
+};
+
+// OutputPipelineOutputType is a helper to determine the output data type of a
+// pipeline, for a
+// given input data type. It is a recursive template; see the explanation on
+// OutputPipelineEvalImpl below.
+template <typename OutputPipelineType, int FirstStage, typename InputType,
+          bool StopRecursion =
+              FirstStage == std::tuple_size<OutputPipelineType>::value>
+struct OutputPipelineOutputType {
+  typedef typename std::tuple_element<FirstStage, OutputPipelineType>::type
+      FirstStageType;
+  typedef typename OutputStageEvalImpl<FirstStageType, InputType>::OutputType
+      FirstStageOutputType;
+  typedef typename OutputPipelineOutputType<OutputPipelineType, FirstStage + 1,
+                                            FirstStageOutputType>::Type Type;
+};
+
+template <typename OutputPipelineType, int FirstStage, typename InputType>
+struct OutputPipelineOutputType<OutputPipelineType, FirstStage, InputType,
+                                true> {
+  typedef InputType Type;
+};
+
+// OutputPipelineEvalImpl is a helper to implement the evaluation of
+// the whole pipeline. It is a recursive template to implement compile-time
+// unrolling of the loop over all pipeline stages. The 'FirstStage' parameter
+// is how we implement recursion: each specialization implements only
+// evaluation starting at 'FirstStage'. The StopRecursion parameter is just a
+// helper to implement the termination of the recursion as a partial
+// specialization below.
+template <typename OutputPipelineType, int FirstStage, typename InputType,
+          bool StopRecursion =
+              FirstStage == std::tuple_size<OutputPipelineType>::value>
+struct OutputPipelineEvalImpl {
+  typedef typename std::tuple_element<FirstStage, OutputPipelineType>::type
+      FirstStageType;
+  typedef typename OutputStageEvalImpl<FirstStageType, InputType>::OutputType
+      FirstStageOutputType;
+  typedef typename OutputPipelineOutputType<OutputPipelineType, FirstStage,
+                                            InputType>::Type OutputType;
+
+  OutputPipelineEvalImpl(const OutputPipelineType& output_pipeline)
+      : head_impl(std::get<FirstStage>(output_pipeline)),
+        tail_impl(output_pipeline) {}
+
+  OutputType Eval(InputType input, int row, int col) const {
+    // Evaluate the first stage.
+    FirstStageOutputType first_stage_output = head_impl.Eval(input, row, col);
+    // Recurse into the remaining stages.
+    return tail_impl.Eval(first_stage_output, row, col);
+  }
+
+  const OutputStageEvalImpl<FirstStageType, InputType> head_impl;
+  const OutputPipelineEvalImpl<OutputPipelineType, FirstStage + 1,
+                               FirstStageOutputType>
+      tail_impl;
+};
+
+// Specialization on 'StopRecursion' for terminating the recursion.
+template <typename OutputPipelineType, int FirstStage, typename InputType>
+struct OutputPipelineEvalImpl<OutputPipelineType, FirstStage, InputType, true> {
+  OutputPipelineEvalImpl(const OutputPipelineType&) {}
+
+  InputType Eval(InputType input, int, int) const {
+    // Terminating the recursion.
+    return input;
+  }
+};
+
+template <typename RegisterBlockType, typename DstType>
+struct StoreFinalOutputImpl {
+  static_assert(std::is_same<RegisterBlockType, void>::value,
+                "This generic impl should never be hit");
+};
+
+template <typename ScalarType, int Rows, int Cols, typename DstType>
+struct StoreFinalOutputImpl<RegisterBlock<ScalarType, Rows, Cols>, DstType> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  static void Run(const RegisterBlockType& src, DstType* dst, int row,
+                  int col) {
+    for (int r = 0; r < Rows; r++) {
+      for (int c = 0; c < Cols; c++) {
+        *dst->data(row + r, col + c) = src.buf.reg[r + c * Rows];
+      }
+    }
+  }
+};
+
+// StoreFinalOutput takes the final value at the end of the output pipeline and
+// stores it into the destination matrix. It can be specialized for different
+// data types; the generic implementation here is typically used only for plain
+// old scalar (not SIMD) types.
+template <typename RegisterBlockType, typename DstType>
+void StoreFinalOutput(RegisterBlockType src, DstType* dst, int row, int col) {
+  StoreFinalOutputImpl<RegisterBlockType, DstType>::Run(src, dst, row, col);
+}
+
+template <typename OutputPipelineType, typename InputType>
+struct OutputPipelineExecutor {
+  OutputPipelineExecutor(const OutputPipelineType& output_pipeline)
+      : output_pipeline_eval_impl_(output_pipeline) {}
+
+  // RunOutputPipeline is the entry point into the output pipeline evaluation
+  // code. It should be the only thing that unpack code calls. It takes the
+  // result
+  // of the unpack stage and stores it into the destination matrix.
+  template <typename DstType>
+  void Execute(InputType input, DstType* dst, int src_global_row,
+               int src_global_col, int dst_row, int dst_col) const {
+    // Statically assert that the output pipeline matches the given destination
+    // matrix's scalar type.
+    typedef typename OutputPipelineOutputType<
+        OutputPipelineType, 0, InputType>::Type::BufferType::ScalarType
+
+        ScalarOutputType;
+    typedef typename DstType::Scalar ScalarDstType;
+    static_assert(std::is_same<ScalarOutputType, ScalarDstType>::value,
+                  "mismatched destination scalar type and output pipeline");
+
+    // Evaluate the output pipeline.
+    auto output =
+        output_pipeline_eval_impl_.Eval(input, src_global_row, src_global_col);
+    // Store the result into the destination matrix.
+    StoreFinalOutput(output, dst, dst_row, dst_col);
+  }
+
+  const OutputPipelineEvalImpl<OutputPipelineType, 0, InputType>
+      output_pipeline_eval_impl_;
+};
+
+}  // namespace gemmlowp
+
+#ifdef GEMMLOWP_NEON
+#include "output_neon.h"
+#elif defined(GEMMLOWP_SSE4)
+#include "output_sse.h"
+#endif
+
+#endif  // GEMMLOWP_INTERNAL_OUTPUT_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_neon.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_neon.h
new file mode 100644
index 0000000..7e111e5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_neon.h
@@ -0,0 +1,432 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// output_neon.h: optimized NEON specializations of the templates in output.h.
+
+#ifndef GEMMLOWP_INTERNAL_OUTPUT_NEON_H_
+#define GEMMLOWP_INTERNAL_OUTPUT_NEON_H_
+
+#include "output.h"
+
+#include <arm_neon.h>
+
+namespace gemmlowp {
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<4>> {
+  typedef RegBufferInt32<4> InputType;
+  typedef RegBufferUint8<4> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    int16x4_t res_16 = vqmovn_s32(input.reg[0]);
+    uint8x8_t res_8 = vqmovun_s16(vcombine_s16(res_16, res_16));
+    output.reg[0] = vget_lane_u32(vreinterpret_u32_u8(res_8), 0);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<8>> {
+  typedef RegBufferInt32<8> InputType;
+  typedef RegBufferUint8<8> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    int16x8_t res_16 =
+        vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
+    output.reg[0] = vqmovun_s16(res_16);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<16>> {
+  typedef RegBufferInt32<16> InputType;
+  typedef RegBufferUint8<16> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    int16x8_t res_16_0 =
+        vcombine_s16(vqmovn_s32(input.reg[0]), vqmovn_s32(input.reg[1]));
+    int16x8_t res_16_1 =
+        vcombine_s16(vqmovn_s32(input.reg[2]), vqmovn_s32(input.reg[3]));
+    output.reg[0] = vqmovun_s16(res_16_0);
+    output.reg[1] = vqmovun_s16(res_16_1);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<32>> {
+  typedef RegBufferInt32<32> InputType;
+  typedef RegBufferUint8<32> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    int16x8_t res_16[4];
+    for (int i = 0; i < 4; i++) {
+      res_16[i] = vcombine_s16(vqmovn_s32(input.reg[2 * i]),
+                               vqmovn_s32(input.reg[2 * i + 1]));
+    }
+    for (int i = 0; i < 4; i++) {
+      output.reg[i] = vqmovun_s16(res_16[i]);
+    }
+    return output;
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 1>, DstType> {
+  static void Run(const RegBlockInt32<8, 1>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      StoreInt32x4(dst->data(row, col), src.buf.reg[0]);
+      StoreInt32x4(dst->data(row + 4, col), src.buf.reg[1]);
+    } else {
+      *dst->data(row + 0, col) = GetLane<0>(src.buf.reg[0]);
+      *dst->data(row + 1, col) = GetLane<1>(src.buf.reg[0]);
+      *dst->data(row + 2, col) = GetLane<2>(src.buf.reg[0]);
+      *dst->data(row + 3, col) = GetLane<3>(src.buf.reg[0]);
+      *dst->data(row + 4, col) = GetLane<0>(src.buf.reg[1]);
+      *dst->data(row + 5, col) = GetLane<1>(src.buf.reg[1]);
+      *dst->data(row + 6, col) = GetLane<2>(src.buf.reg[1]);
+      *dst->data(row + 7, col) = GetLane<3>(src.buf.reg[1]);
+    }
+  }
+};
+
+inline RegBlockInt32<4, 4> Transpose(const RegBlockInt32<4, 4>& src) {
+  const int32x4x2_t t0 = vtrnq_s32(src.buf.reg[0], src.buf.reg[1]);
+  const int32x4x2_t t1 = vtrnq_s32(src.buf.reg[2], src.buf.reg[3]);
+  RegBlockInt32<4, 4> result;
+  result.buf.reg[0] =
+      vcombine_s32(vget_low_s32(t0.val[0]), vget_low_s32(t1.val[0]));
+  result.buf.reg[1] =
+      vcombine_s32(vget_low_s32(t0.val[1]), vget_low_s32(t1.val[1]));
+  result.buf.reg[2] =
+      vcombine_s32(vget_high_s32(t0.val[0]), vget_high_s32(t1.val[0]));
+  result.buf.reg[3] =
+      vcombine_s32(vget_high_s32(t0.val[1]), vget_high_s32(t1.val[1]));
+  return result;
+}
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<4, 4>, DstType> {
+  static void Run(const RegBlockInt32<4, 4>& src, DstType* dst, int row,
+                  int col) {
+    const auto& block =
+        DstType::kOrder == MapOrder::ColMajor ? src : Transpose(src);
+    std::int32_t* dst_ptr = dst->data(row, col);
+    int stride = dst->stride();
+    for (int i = 0; i < 4; i++) {
+      vst1q_s32(dst_ptr + i * stride, block.buf.reg[i]);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 4>, DstType> {
+  static void Run(const RegBlockInt32<8, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::int32_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      int col_stride = dst->cols_stride();
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + i * col_stride + 0, src.buf.reg[2 * i + 0]);
+        vst1q_s32(dst_ptr + i * col_stride + 4, src.buf.reg[2 * i + 1]);
+      }
+    } else {
+      int row_stride = dst->rows_stride();
+      RegBlockInt32<4, 4> top;
+      top.buf.reg[0] = src.buf.reg[0];
+      top.buf.reg[1] = src.buf.reg[2];
+      top.buf.reg[2] = src.buf.reg[4];
+      top.buf.reg[3] = src.buf.reg[6];
+      const auto transpose_top = Transpose(top);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + i * row_stride, transpose_top.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom;
+      bottom.buf.reg[0] = src.buf.reg[1];
+      bottom.buf.reg[1] = src.buf.reg[3];
+      bottom.buf.reg[2] = src.buf.reg[5];
+      bottom.buf.reg[3] = src.buf.reg[7];
+      const auto transpose_bottom = Transpose(bottom);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + (i + 4) * row_stride, transpose_bottom.buf.reg[i]);
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 8>, DstType> {
+  static void Run(const RegBlockInt32<8, 8>& src, DstType* dst, int row,
+                  int col) {
+    std::int32_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      int col_stride = dst->cols_stride();
+      for (int i = 0; i < 8; i++) {
+        vst1q_s32(dst_ptr + i * col_stride, src.buf.reg[2 * i]);
+        vst1q_s32(dst_ptr + i * col_stride + 4, src.buf.reg[2 * i + 1]);
+      }
+    } else {
+      int row_stride = dst->rows_stride();
+      RegBlockInt32<4, 4> top_left;
+      top_left.buf.reg[0] = src.buf.reg[0];
+      top_left.buf.reg[1] = src.buf.reg[2];
+      top_left.buf.reg[2] = src.buf.reg[4];
+      top_left.buf.reg[3] = src.buf.reg[6];
+      const auto transpose_top_left = Transpose(top_left);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + i * row_stride, transpose_top_left.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom_left;
+      bottom_left.buf.reg[0] = src.buf.reg[1];
+      bottom_left.buf.reg[1] = src.buf.reg[3];
+      bottom_left.buf.reg[2] = src.buf.reg[5];
+      bottom_left.buf.reg[3] = src.buf.reg[7];
+      const auto transpose_bottom_left = Transpose(bottom_left);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + (i + 4) * row_stride,
+                  transpose_bottom_left.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> top_right;
+      top_right.buf.reg[0] = src.buf.reg[8];
+      top_right.buf.reg[1] = src.buf.reg[10];
+      top_right.buf.reg[2] = src.buf.reg[12];
+      top_right.buf.reg[3] = src.buf.reg[14];
+      const auto transpose_top_right = Transpose(top_right);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + i * row_stride + 4, transpose_top_right.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom_right;
+      bottom_right.buf.reg[0] = src.buf.reg[9];
+      bottom_right.buf.reg[1] = src.buf.reg[11];
+      bottom_right.buf.reg[2] = src.buf.reg[13];
+      bottom_right.buf.reg[3] = src.buf.reg[15];
+      const auto transpose_bottom_right = Transpose(bottom_right);
+      for (int i = 0; i < 4; i++) {
+        vst1q_s32(dst_ptr + (i + 4) * row_stride + 4,
+                  transpose_bottom_right.buf.reg[i]);
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<4, 1>, DstType> {
+  static void Run(const RegBlockInt32<4, 1>& src, DstType* dst, int row,
+                  int col) {
+    std::int32_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      vst1q_s32(dst_ptr, src.buf.reg[0]);
+    } else {
+      int row_stride = dst->rows_stride();
+      vst1q_lane_s32(dst_ptr + 0 * row_stride, src.buf.reg[0], 0);
+      vst1q_lane_s32(dst_ptr + 1 * row_stride, src.buf.reg[0], 1);
+      vst1q_lane_s32(dst_ptr + 2 * row_stride, src.buf.reg[0], 2);
+      vst1q_lane_s32(dst_ptr + 3 * row_stride, src.buf.reg[0], 3);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<1, 4>, DstType> {
+  static void Run(const RegBlockInt32<1, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::int32_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::RowMajor) {
+      vst1q_s32(dst_ptr, src.buf.reg[0]);
+    } else {
+      int col_stride = dst->cols_stride();
+      vst1q_lane_s32(dst_ptr + 0 * col_stride, src.buf.reg[0], 0);
+      vst1q_lane_s32(dst_ptr + 1 * col_stride, src.buf.reg[0], 1);
+      vst1q_lane_s32(dst_ptr + 2 * col_stride, src.buf.reg[0], 2);
+      vst1q_lane_s32(dst_ptr + 3 * col_stride, src.buf.reg[0], 3);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<4, 1>, DstType> {
+  static void Run(const RegBlockUint8<4, 1>& src, DstType* dst, int row,
+                  int col) {
+    const std::uint32_t src_reg = src.buf.reg[0];
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row + i, col) = (src_reg >> (8 * i));
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<1, 4>, DstType> {
+  static void Run(const RegBlockUint8<1, 4>& src, DstType* dst, int row,
+                  int col) {
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row, col + i) = (src.buf.reg[0] >> (8 * i));
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 1>, DstType> {
+  static void Run(const RegBlockUint8<8, 1>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      vst1_u8(dst_ptr, src.buf.reg[0]);
+    } else {
+      const int row_stride = dst->rows_stride();
+      vst1_lane_u8(dst_ptr + 0 * row_stride, src.buf.reg[0], 0);
+      vst1_lane_u8(dst_ptr + 1 * row_stride, src.buf.reg[0], 1);
+      vst1_lane_u8(dst_ptr + 2 * row_stride, src.buf.reg[0], 2);
+      vst1_lane_u8(dst_ptr + 3 * row_stride, src.buf.reg[0], 3);
+      vst1_lane_u8(dst_ptr + 4 * row_stride, src.buf.reg[0], 4);
+      vst1_lane_u8(dst_ptr + 5 * row_stride, src.buf.reg[0], 5);
+      vst1_lane_u8(dst_ptr + 6 * row_stride, src.buf.reg[0], 6);
+      vst1_lane_u8(dst_ptr + 7 * row_stride, src.buf.reg[0], 7);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<4, 4>, DstType> {
+  static void Run(const RegBlockUint8<4, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t* dst_ptr = dst->data(row, col);
+    const int row_stride = dst->rows_stride();
+    const int col_stride = dst->cols_stride();
+    for (int i = 0; i < 2; i++) {
+      vst1_lane_u8(dst_ptr + 0 * row_stride + (2 * i + 0) * col_stride,
+                   src.buf.reg[i], 0);
+      vst1_lane_u8(dst_ptr + 1 * row_stride + (2 * i + 0) * col_stride,
+                   src.buf.reg[i], 1);
+      vst1_lane_u8(dst_ptr + 2 * row_stride + (2 * i + 0) * col_stride,
+                   src.buf.reg[i], 2);
+      vst1_lane_u8(dst_ptr + 3 * row_stride + (2 * i + 0) * col_stride,
+                   src.buf.reg[i], 3);
+      vst1_lane_u8(dst_ptr + 0 * row_stride + (2 * i + 1) * col_stride,
+                   src.buf.reg[i], 4);
+      vst1_lane_u8(dst_ptr + 1 * row_stride + (2 * i + 1) * col_stride,
+                   src.buf.reg[i], 5);
+      vst1_lane_u8(dst_ptr + 2 * row_stride + (2 * i + 1) * col_stride,
+                   src.buf.reg[i], 6);
+      vst1_lane_u8(dst_ptr + 3 * row_stride + (2 * i + 1) * col_stride,
+                   src.buf.reg[i], 7);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 4>, DstType> {
+  static void Run(const RegBlockUint8<8, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t* dst_ptr = dst->data(row, col);
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      int col_stride = dst->cols_stride();
+      for (int i = 0; i < 4; i++) {
+        vst1_u8(dst_ptr + i * col_stride, src.buf.reg[i]);
+      }
+    } else {
+      for (int i = 0; i < 4; i++) {
+        int row_stride = dst->rows_stride();
+        std::uint8_t* col_ptr = dst_ptr + i;
+        vst1_lane_u8(col_ptr + 0 * row_stride, src.buf.reg[i], 0);
+        vst1_lane_u8(col_ptr + 1 * row_stride, src.buf.reg[i], 1);
+        vst1_lane_u8(col_ptr + 2 * row_stride, src.buf.reg[i], 2);
+        vst1_lane_u8(col_ptr + 3 * row_stride, src.buf.reg[i], 3);
+        vst1_lane_u8(col_ptr + 4 * row_stride, src.buf.reg[i], 4);
+        vst1_lane_u8(col_ptr + 5 * row_stride, src.buf.reg[i], 5);
+        vst1_lane_u8(col_ptr + 6 * row_stride, src.buf.reg[i], 6);
+        vst1_lane_u8(col_ptr + 7 * row_stride, src.buf.reg[i], 7);
+      }
+    }
+  }
+};
+
+inline RegBlockUint8<8, 8> Transpose(const RegBlockUint8<8, 8>& src) {
+  uint8x8x2_t a[4];
+  a[0] = vtrn_u8(src.buf.reg[0], src.buf.reg[1]);
+  a[1] = vtrn_u8(src.buf.reg[2], src.buf.reg[3]);
+  a[2] = vtrn_u8(src.buf.reg[4], src.buf.reg[5]);
+  a[3] = vtrn_u8(src.buf.reg[6], src.buf.reg[7]);
+  uint16x4x2_t b[4];
+  b[0] = vtrn_u16(vreinterpret_u16_u8(a[0].val[0]),
+                  vreinterpret_u16_u8(a[1].val[0]));
+  b[1] = vtrn_u16(vreinterpret_u16_u8(a[0].val[1]),
+                  vreinterpret_u16_u8(a[1].val[1]));
+  b[2] = vtrn_u16(vreinterpret_u16_u8(a[2].val[0]),
+                  vreinterpret_u16_u8(a[3].val[0]));
+  b[3] = vtrn_u16(vreinterpret_u16_u8(a[2].val[1]),
+                  vreinterpret_u16_u8(a[3].val[1]));
+  uint32x2x2_t c[4];
+  c[0] = vtrn_u32(vreinterpret_u32_u16(b[0].val[0]),
+                  vreinterpret_u32_u16(b[2].val[0]));
+  c[1] = vtrn_u32(vreinterpret_u32_u16(b[1].val[0]),
+                  vreinterpret_u32_u16(b[3].val[0]));
+  c[2] = vtrn_u32(vreinterpret_u32_u16(b[0].val[1]),
+                  vreinterpret_u32_u16(b[2].val[1]));
+  c[3] = vtrn_u32(vreinterpret_u32_u16(b[1].val[1]),
+                  vreinterpret_u32_u16(b[3].val[1]));
+  RegBlockUint8<8, 8> result;
+  result.buf.reg[0] = vreinterpret_u8_u32(c[0].val[0]);
+  result.buf.reg[1] = vreinterpret_u8_u32(c[1].val[0]);
+  result.buf.reg[2] = vreinterpret_u8_u32(c[2].val[0]);
+  result.buf.reg[3] = vreinterpret_u8_u32(c[3].val[0]);
+  result.buf.reg[4] = vreinterpret_u8_u32(c[0].val[1]);
+  result.buf.reg[5] = vreinterpret_u8_u32(c[1].val[1]);
+  result.buf.reg[6] = vreinterpret_u8_u32(c[2].val[1]);
+  result.buf.reg[7] = vreinterpret_u8_u32(c[3].val[1]);
+  return result;
+}
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 8>, DstType> {
+  static void Run(const RegBlockUint8<8, 8>& src, DstType* dst, int row,
+                  int col) {
+    const auto& block =
+        DstType::kOrder == MapOrder::ColMajor ? src : Transpose(src);
+    std::uint8_t* dst_ptr = dst->data(row, col);
+    int stride = dst->stride();
+    for (int i = 0; i < 8; i++) {
+      vst1_u8(dst_ptr + i * stride, block.buf.reg[i]);
+    }
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_OUTPUT_NEON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_sse.h
new file mode 100644
index 0000000..5c06253
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/output_sse.h
@@ -0,0 +1,354 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// output_sse.h: optimized SSE4.2 specializations of the templates in output.h.
+
+#ifndef GEMMLOWP_INTERNAL_OUTPUT_SSE_H_
+#define GEMMLOWP_INTERNAL_OUTPUT_SSE_H_
+
+#include "output.h"
+
+#include <smmintrin.h>
+
+namespace gemmlowp {
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<4>> {
+  typedef RegBufferInt32<4> InputType;
+  typedef RegBufferUint8<4> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    __m128i res_16 = _mm_packs_epi32(input.reg[0], input.reg[0]);
+    __m128i res_8 = _mm_packus_epi16(res_16, res_16);
+    output.reg[0] = _mm_cvtsi128_si32(res_8);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<8>> {
+  typedef RegBufferInt32<8> InputType;
+  typedef RegBufferUint8<8> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    __m128i res_16 = _mm_packs_epi32(input.reg[0], input.reg[1]);
+    __m128i res_8 = _mm_packus_epi16(res_16, res_16);
+    output.reg[0] = _mm_extract_epi32(res_8, 0);
+    output.reg[1] = _mm_extract_epi32(res_8, 1);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<16>> {
+  typedef RegBufferInt32<16> InputType;
+  typedef RegBufferUint8<16> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    __m128i res_16_0 = _mm_packs_epi32(input.reg[0], input.reg[1]);
+    __m128i res_16_1 = _mm_packs_epi32(input.reg[2], input.reg[3]);
+    output.reg[0] = _mm_packus_epi16(res_16_0, res_16_1);
+    return output;
+  }
+};
+
+template <>
+struct OutputStageEvalBufferImpl<OutputStageSaturatingCastToUint8,
+                                 RegBufferInt32<32>> {
+  typedef RegBufferInt32<32> InputType;
+  typedef RegBufferUint8<32> OutputType;
+
+  typedef OutputStageSaturatingCastToUint8 OutputStage;
+
+  OutputStageEvalBufferImpl(const OutputStage&) {}
+
+  OutputType Eval(InputType input) const {
+    OutputType output;
+    __m128i res_16_0 = _mm_packs_epi32(input.reg[0], input.reg[1]);
+    __m128i res_16_1 = _mm_packs_epi32(input.reg[2], input.reg[3]);
+    output.reg[0] = _mm_packus_epi16(res_16_0, res_16_1);
+    __m128i res_16_2 = _mm_packs_epi32(input.reg[4], input.reg[5]);
+    __m128i res_16_3 = _mm_packs_epi32(input.reg[6], input.reg[7]);
+    output.reg[1] = _mm_packus_epi16(res_16_2, res_16_3);
+    return output;
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<4, 1>, DstType> {
+  static void Run(const RegBlockInt32<4, 1>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      StoreInt32x4(dst->data(row, col), src.buf.reg[0]);
+    } else {
+      *dst->data(row + 0, col) = GetLane<0>(src.buf.reg[0]);
+      *dst->data(row + 1, col) = GetLane<1>(src.buf.reg[0]);
+      *dst->data(row + 2, col) = GetLane<2>(src.buf.reg[0]);
+      *dst->data(row + 3, col) = GetLane<3>(src.buf.reg[0]);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 1>, DstType> {
+  static void Run(const RegBlockInt32<8, 1>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      StoreInt32x4(dst->data(row, col), src.buf.reg[0]);
+      StoreInt32x4(dst->data(row + 4, col), src.buf.reg[1]);
+    } else {
+      *dst->data(row + 0, col) = GetLane<0>(src.buf.reg[0]);
+      *dst->data(row + 1, col) = GetLane<1>(src.buf.reg[0]);
+      *dst->data(row + 2, col) = GetLane<2>(src.buf.reg[0]);
+      *dst->data(row + 3, col) = GetLane<3>(src.buf.reg[0]);
+      *dst->data(row + 4, col) = GetLane<0>(src.buf.reg[1]);
+      *dst->data(row + 5, col) = GetLane<1>(src.buf.reg[1]);
+      *dst->data(row + 6, col) = GetLane<2>(src.buf.reg[1]);
+      *dst->data(row + 7, col) = GetLane<3>(src.buf.reg[1]);
+    }
+  }
+};
+
+inline RegBlockInt32<4, 4> Transpose(const RegBlockInt32<4, 4>& src) {
+  __m128i t0 = _mm_unpacklo_epi32(src.buf.reg[0], src.buf.reg[1]);
+  __m128i t1 = _mm_unpacklo_epi32(src.buf.reg[2], src.buf.reg[3]);
+  __m128i t2 = _mm_unpackhi_epi32(src.buf.reg[0], src.buf.reg[1]);
+  __m128i t3 = _mm_unpackhi_epi32(src.buf.reg[2], src.buf.reg[3]);
+
+  RegBlockInt32<4, 4> result;
+  result.buf.reg[0] = _mm_unpacklo_epi64(t0, t1);
+  result.buf.reg[1] = _mm_unpackhi_epi64(t0, t1);
+  result.buf.reg[2] = _mm_unpacklo_epi64(t2, t3);
+  result.buf.reg[3] = _mm_unpackhi_epi64(t2, t3);
+  return result;
+}
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<4, 4>, DstType> {
+  static void Run(const RegBlockInt32<4, 4>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row, col + i), src.buf.reg[i]);
+      }
+    } else {
+      const auto transpose = Transpose(src);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + i, col), transpose.buf.reg[i]);
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 4>, DstType> {
+  static void Run(const RegBlockInt32<8, 4>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row, col + i), src.buf.reg[2 * i]);
+        StoreInt32x4(dst->data(row + 4, col + i), src.buf.reg[2 * i + 1]);
+      }
+    } else {
+      RegBlockInt32<4, 4> top;
+      top.buf.reg[0] = src.buf.reg[0];
+      top.buf.reg[1] = src.buf.reg[2];
+      top.buf.reg[2] = src.buf.reg[4];
+      top.buf.reg[3] = src.buf.reg[6];
+      const auto transpose_top = Transpose(top);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + i, col), transpose_top.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom;
+      bottom.buf.reg[0] = src.buf.reg[1];
+      bottom.buf.reg[1] = src.buf.reg[3];
+      bottom.buf.reg[2] = src.buf.reg[5];
+      bottom.buf.reg[3] = src.buf.reg[7];
+      const auto transpose_bottom = Transpose(bottom);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + 4 + i, col), transpose_bottom.buf.reg[i]);
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<8, 8>, DstType> {
+  static void Run(const RegBlockInt32<8, 8>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      for (int i = 0; i < 8; i++) {
+        StoreInt32x4(dst->data(row, col + i), src.buf.reg[2 * i]);
+        StoreInt32x4(dst->data(row + 4, col + i), src.buf.reg[2 * i + 1]);
+      }
+    } else {
+      RegBlockInt32<4, 4> top_left;
+      top_left.buf.reg[0] = src.buf.reg[0];
+      top_left.buf.reg[1] = src.buf.reg[2];
+      top_left.buf.reg[2] = src.buf.reg[4];
+      top_left.buf.reg[3] = src.buf.reg[6];
+      const auto transpose_top_left = Transpose(top_left);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + i, col), transpose_top_left.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom_left;
+      bottom_left.buf.reg[0] = src.buf.reg[1];
+      bottom_left.buf.reg[1] = src.buf.reg[3];
+      bottom_left.buf.reg[2] = src.buf.reg[5];
+      bottom_left.buf.reg[3] = src.buf.reg[7];
+      const auto transpose_bottom_left = Transpose(bottom_left);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + 4 + i, col),
+                     transpose_bottom_left.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> top_right;
+      top_right.buf.reg[0] = src.buf.reg[8];
+      top_right.buf.reg[1] = src.buf.reg[10];
+      top_right.buf.reg[2] = src.buf.reg[12];
+      top_right.buf.reg[3] = src.buf.reg[14];
+      const auto transpose_top_right = Transpose(top_right);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + i, col + 4),
+                     transpose_top_right.buf.reg[i]);
+      }
+      RegBlockInt32<4, 4> bottom_right;
+      bottom_right.buf.reg[0] = src.buf.reg[9];
+      bottom_right.buf.reg[1] = src.buf.reg[11];
+      bottom_right.buf.reg[2] = src.buf.reg[13];
+      bottom_right.buf.reg[3] = src.buf.reg[15];
+      const auto transpose_bottom_right = Transpose(bottom_right);
+      for (int i = 0; i < 4; i++) {
+        StoreInt32x4(dst->data(row + 4 + i, col + 4),
+                     transpose_bottom_right.buf.reg[i]);
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockInt32<1, 4>, DstType> {
+  static void Run(const RegBlockInt32<1, 4>& src, DstType* dst, int row,
+                  int col) {
+    if (DstType::kOrder == MapOrder::ColMajor) {
+      *dst->data(row, col + 0) = GetLane<0>(src.buf.reg[0]);
+      *dst->data(row, col + 1) = GetLane<1>(src.buf.reg[0]);
+      *dst->data(row, col + 2) = GetLane<2>(src.buf.reg[0]);
+      *dst->data(row, col + 3) = GetLane<3>(src.buf.reg[0]);
+    } else {
+      StoreInt32x4(dst->data(row, col), src.buf.reg[0]);
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<4, 1>, DstType> {
+  static void Run(const RegBlockUint8<4, 1>& src, DstType* dst, int row,
+                  int col) {
+    const std::uint32_t src_reg = src.buf.reg[0];
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row + i, col) = (src_reg >> (8 * i));
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 1>, DstType> {
+  static void Run(const RegBlockUint8<8, 1>& src, DstType* dst, int row,
+                  int col) {
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row + i, col) = (src.buf.reg[0] >> (8 * i));
+    }
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row + 4 + i, col) = (src.buf.reg[1] >> (8 * i));
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<1, 4>, DstType> {
+  static void Run(const RegBlockUint8<1, 4>& src, DstType* dst, int row,
+                  int col) {
+    for (int i = 0; i < 4; i++) {
+      *dst->data(row, col + i) = (src.buf.reg[0] >> (8 * i));
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<4, 4>, DstType> {
+  static void Run(const RegBlockUint8<4, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t buf[16];
+    StoreUint8x16(buf, src.buf.reg[0]);
+    for (int c = 0; c < 4; c++) {
+      for (int r = 0; r < 4; r++) {
+        *dst->data(row + r, col + c) = buf[r + 4 * c];
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 4>, DstType> {
+  static void Run(const RegBlockUint8<8, 4>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t buf[32];
+    StoreUint8x16(buf, src.buf.reg[0]);
+    StoreUint8x16(buf + 16, src.buf.reg[1]);
+    for (int c = 0; c < 4; c++) {
+      for (int r = 0; r < 8; r++) {
+        *dst->data(row + r, col + c) = buf[r + 8 * c];
+      }
+    }
+  }
+};
+
+template <typename DstType>
+struct StoreFinalOutputImpl<RegBlockUint8<8, 8>, DstType> {
+  static void Run(const RegBlockUint8<8, 8>& src, DstType* dst, int row,
+                  int col) {
+    std::uint8_t buf[64];
+    StoreUint8x16(buf, src.buf.reg[0]);
+    StoreUint8x16(buf + 16, src.buf.reg[1]);
+    StoreUint8x16(buf + 32, src.buf.reg[2]);
+    StoreUint8x16(buf + 48, src.buf.reg[3]);
+    for (int c = 0; c < 8; c++) {
+      for (int r = 0; r < 8; r++) {
+        *dst->data(row + r, col + c) = buf[r + 8 * c];
+      }
+    }
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_OUTPUT_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack.h
new file mode 100644
index 0000000..3395396
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack.h
@@ -0,0 +1,435 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// pack.h: packing blocks of the LHS and RHS into the data layout
+// that is expected by compute.h and eventually by kernels.
+// Because this data layout depends on the kernel format, code here
+// is templated in KernelLhsFormat/KernelRhsFormat.
+//
+// Readers note: an important theme around here is that we try hard
+// to handle both Lhs and Rhs with a single piece of code. We indifferently
+// refer to the Lhs and Rhs as a 'Side'. Instead of addressing matrices
+// by (row, column) indices, we address them by (width, depth), as explained
+// in kernel.h. This allows us to handle both Lhs and Rhs on an equal footing,
+// at once.
+
+#ifndef GEMMLOWP_INTERNAL_PACK_H_
+#define GEMMLOWP_INTERNAL_PACK_H_
+
+#include <cstring>
+
+#include "allocator.h"
+#include "block_params.h"
+#include "common.h"
+#include "kernel.h"
+
+namespace gemmlowp {
+
+// A PackedSideBlock instance is a packed block of either the LHS or RHS
+// (whence the generic 'Side' name).
+//
+// 'Packed' means that it is laid out in the storage order that
+// is expected by the specified kernel format. From a block of the input
+// LHS or RHS matrix, one obtains a PackedSideBlock by calling PackLhs()
+// or PackRhs().
+template <typename tKernelSideFormat>
+class PackedSideBlock {
+ public:
+  typedef tKernelSideFormat KernelSideFormat;
+
+  PackedSideBlock(Side side, Allocator* allocator,
+                  const BlockParams& block_params)
+      : allocator_(allocator), pos_(0) {
+    GetSideBlockParams(side, &params_, block_params);
+    data_handle_ =
+        allocator_->Reserve<std::uint8_t>(params_.l2_width * params_.l2_depth);
+    sums_of_each_slice_handle_ =
+        allocator_->Reserve<std::int32_t>(params_.l2_width);
+  }
+
+  ~PackedSideBlock() {}
+
+  void seek_run(int start_width, int start_depth) const {
+    int kernel_run_depth =
+        std::min<int>(params_.l1_depth, params_.l2_depth - start_depth);
+    pos_ = params_.l2_width * start_depth + start_width * kernel_run_depth;
+  }
+
+  void seek_next_cell() const { pos_ += KernelSideFormat::Cell::kSize; }
+
+  void seek_forward_n_cells(int n) const {
+    pos_ += n * KernelSideFormat::Cell::kSize;
+  }
+
+  const std::uint8_t* current_data() const {
+    return allocator_->GetPointer<std::uint8_t>(data_handle_) + pos_;
+  }
+
+  std::uint8_t* current_data() {
+    return allocator_->GetPointer<std::uint8_t>(data_handle_) + pos_;
+  }
+
+  std::int32_t* sums_of_each_slice() {
+    return allocator_->GetPointer<std::int32_t>(sums_of_each_slice_handle_);
+  }
+
+  const std::int32_t* sums_of_each_slice() const {
+    return allocator_->GetPointer<const std::int32_t>(
+        sums_of_each_slice_handle_);
+  }
+
+  const SideBlockParams& params() const { return params_; }
+
+ private:
+  // The block size parameters that this PackedSizeBlock follows.
+  // The L2 parameters determine its overall size, while the L1 parameters,
+  // together with the kernel format template parameter, determine
+  // the fine details of the storage/traversal order.
+  SideBlockParams params_;
+
+  // Pointer to the allocator provided by the caller. Not owned.
+  // The Allocator is assumed to outlive the PackedSideBlock.
+  Allocator* const allocator_;
+
+  // Handle on the buffer backing this packed block. Owned.
+  Allocator::Handle data_handle_;
+
+  // Handle on the additional buffer backing the vector of sums of slices
+  // associated with this block. Owned.
+  Allocator::Handle sums_of_each_slice_handle_;
+
+  // pos_ is the current position in the buffer, which we access
+  // sequentially, like a file.
+  // The idea is that we pack data in the same order as it is
+  // going to be traversed during the computation, which for
+  // cache-friendliness reasons is complicated to random-access,
+  // as the offsets calculations would be intricate. So we
+  // give up random-access addressing, and instead content ourselves
+  // with sequential access.
+  //
+  // pos_ is mutable because during the computation we will want to
+  // be able to iterate on the data in a const PackedSideBlock.
+  mutable int pos_;
+};
+
+// WidthMajor and DepthMajor are custom phrases modelled after the
+// standard terminology 'row-major' and 'column-major'. Their meaning
+// should be transparent once one has read the explanation in kernel.h:
+// for example, in the Lhs, the 'width' dimension is the rows dimension,
+// so there WidthMajor means RowMajor, while in the Rhs it is the opposite.
+// Another way to put it: WidthMajor means that contiguous storage is used
+// for entries having the same 'width' index.
+enum class SideMapOrder { WidthMajor, DepthMajor };
+
+// Similar to MatrixMap from map.h, but in terms of width/depth instead of
+// rows/columns. Used to address blocks of the input LHS/RHS matrices when
+// packing them.
+template <typename tScalar, SideMapOrder tOrder>
+class SideMap {
+ public:
+  typedef tScalar Scalar;
+  static const SideMapOrder kOrder = tOrder;
+
+  SideMap(Scalar* data, int width, int depth, int stride)
+      : data_(data), width_(width), depth_(depth), stride_(stride) {}
+
+  SideMap(Scalar* data, int width, int depth)
+      : data_(data), width_(width), depth_(depth) {
+    stride_ = kOrder == SideMapOrder::WidthMajor ? depth_ : width_;
+  }
+
+  SideMap(const SideMap& other)
+      : data_(other.data_),
+        width_(other.width_),
+        depth_(other.depth_),
+        stride_(other.stride_) {}
+
+  int width() const { return width_; }
+  int depth() const { return depth_; }
+  int stride() const { return stride_; }
+  int width_stride() const {
+    return kOrder == SideMapOrder::DepthMajor ? 1 : stride_;
+  }
+  int depth_stride() const {
+    return kOrder == SideMapOrder::WidthMajor ? 1 : stride_;
+  }
+  Scalar* data() const { return data_; }
+  Scalar* data(int w, int d) const {
+    return data_ + w * width_stride() + d * depth_stride();
+  }
+  Scalar operator()(int w, int d) const { return *data(w, d); }
+  Scalar& operator()(int w, int d) { return *data(w, d); }
+
+  SideMap block(int start_width, int start_depth, int block_width,
+                int block_depth) const {
+    assert(start_width >= 0);
+    assert(start_width + block_width <= width_);
+    assert(start_depth >= 0);
+    assert(start_depth + block_depth <= depth_);
+
+    return SideMap(data(start_width, start_depth), block_width, block_depth,
+                   stride_);
+  }
+
+ private:
+  Scalar* data_;  // not owned.
+  int width_, depth_, stride_;
+};
+
+// A PackingRegisterBlock is a small fixed-size block of a matrix being
+// packed. This class is the generic non-optimized implementation,
+// it is inherited by the generic implementation of PackingRegisterBlock,
+// which may be overriden by template specialization. Overriding it is how
+// one may provide optimized packing code paths.
+//
+// The packing of a block proceeds in two steps:
+//   1. Ensuring that we have a complete block of source data, i.e. a block of
+//      the compile-time prescribed size. This is where we handle unaligned
+//      boundaries: if we don't have a complete block of source data, then
+//      we copy and zero-extend it into a local temporary (complete_src_),
+//      see MakeCompleteSrc. In the generic case, we do have a complete block,
+//      so we just use it in-place, see UseCompleteSrcInPlace.
+//   2. Packing a complete block into the destination, see Pack. This is the
+//      most critical part, so it's convenient that unaligned boundaries have
+//      already been handled in step 1.
+template <typename SrcMapType, typename PackedSideBlock>
+class PackingRegisterBlockBase {
+ public:
+  typedef typename PackedSideBlock::KernelSideFormat KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  typedef typename KernelSideFormat::Scalar KernelScalar;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+  static const int kCellSize = CellFormat::kSize;
+  static const SideMapOrder kSrcOrder = SrcMapType::kOrder;
+  static const int kZeroPointInputValue =
+      ZeroPointInputValue<KernelScalar>::kValue;
+
+  PackingRegisterBlockBase() : complete_src_(nullptr, 0, 0, 0) {}
+
+ protected:
+  // The source data that's ready for packing. May point to
+  // in-place actual source data if it's already a complete block,
+  // (see UseCompleteSrcInPlace)
+  // or to the local buf_ below into which we copy incomplete blocks
+  // (see MakeCompleteSrc)
+  SrcMapType complete_src_;
+
+  // Temporary buffer for loading incomplete blocks to,
+  // in the source storage order
+  std::uint8_t buf_[kKernelWidth * kRegisterSize];
+
+ public:
+  // Selects a block if in-place source data that's already a complete block
+  void UseCompleteSrcInPlace(const SrcMapType& src) { complete_src_ = src; }
+  // Copies an incomplete block of source data into a local temporary
+  // complete block by zero-extending it.
+  void MakeCompleteSrc(const SrcMapType& src) {
+    memset(buf_, kZeroPointInputValue, kKernelWidth * kRegisterSize);
+    if (kSrcOrder == SideMapOrder::WidthMajor) {
+      for (int w = 0; w < src.width(); w++) {
+        memcpy(buf_ + w * kRegisterSize, src.data(w, 0), src.depth());
+      }
+    } else {
+      assert(kSrcOrder == SideMapOrder::DepthMajor);
+      for (int d = 0; d < src.depth(); d++) {
+        memcpy(buf_ + d * kKernelWidth, src.data(0, d), src.width());
+      }
+    }
+    complete_src_ = SrcMapType(buf_, kKernelWidth, kRegisterSize);
+  }
+  // Packs a complete block into the destination. This is the most
+  // critical part and the part that we most typically want to
+  // override in architecture-specific optimized specializations.
+  void Pack(PackedSideBlock* dst, int start_width) {
+    std::uint8_t* dst_ptr = dst->current_data();
+    for (int cell_start_depth = 0; cell_start_depth < kRegisterSize;
+         cell_start_depth += kCellDepth) {
+      for (int cell_start_width = 0; cell_start_width < kKernelWidth;
+           cell_start_width += kCellWidth) {
+        std::int32_t* cell_sums_of_each_slice_ptr =
+            dst->sums_of_each_slice() + start_width + cell_start_width;
+        const SideMap<const std::uint8_t, kSrcOrder> src_cell_map(
+            complete_src_.block(cell_start_width, cell_start_depth, kCellWidth,
+                                kCellDepth));
+        for (int w = 0; w < kCellWidth; w++) {
+          std::int32_t sum = 0;
+          for (int d = 0; d < kCellDepth; d++) {
+            const std::uint8_t src_val = src_cell_map(w, d);
+            const std::int16_t kernel_val_unwrapped =
+                src_val - kZeroPointInputValue;
+            const std::uint8_t kernel_val_uint8 = kernel_val_unwrapped;
+            dst_ptr[OffsetIntoCell<CellFormat>(w, d)] = kernel_val_uint8;
+            sum += kernel_val_unwrapped;
+          }
+          cell_sums_of_each_slice_ptr[w] += sum;
+        }
+        dst_ptr += kCellSize;
+      }
+    }
+    dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
+  }
+};
+
+template <typename SrcMapType, typename PackedSideBlock>
+class PackingRegisterBlock
+    : public PackingRegisterBlockBase<SrcMapType, PackedSideBlock> {};
+
+// Large-scale implementation of packing.
+template <typename SrcMapType, typename PackedSideBlock>
+class PackSideBlockImpl {
+ public:
+  typedef typename PackedSideBlock::KernelSideFormat KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+
+  typedef PackingRegisterBlock<SrcMapType, PackedSideBlock>
+      PackingRegisterBlockType;
+
+  PackSideBlockImpl(PackedSideBlock* packed_side_block,
+                    const SrcMapType& src_map)
+      : packed_side_block_(packed_side_block), src_map_(src_map) {}
+
+  PackedSideBlock* packed_side_block() const { return packed_side_block_; }
+
+  const SrcMapType& src_map() const { return src_map_; }
+
+  // The public entry point to pack a block.
+  void PackL2() {
+    memset(packed_side_block_->sums_of_each_slice(), 0,
+           sizeof(std::int32_t) * packed_side_block_->params().l2_width);
+    for (int d = 0; d < src_map_.depth();
+         d += packed_side_block_->params().l1_depth) {
+      int ds = std::min<int>(packed_side_block_->params().l1_depth,
+                             src_map_.depth() - d);
+
+      for (int w = 0; w < src_map_.width();
+           w += packed_side_block_->params().l1_width) {
+        int ws = std::min<int>(packed_side_block_->params().l1_width,
+                               src_map_.width() - w);
+
+        PrefetchL1(w, ws, d, ds);
+        PackL1(w, ws, d, ds);
+      }
+    }
+  }
+
+ protected:
+  // The intermediate-level loops, between PackL2 and PackRun.
+  void PackL1(int start_width, int width, int start_depth, int depth) {
+    for (int w = 0; w < width; w += kKernelWidth) {
+      int ws = std::min(+kKernelWidth, width - w);
+      packed_side_block_->seek_run(start_width + w, start_depth);
+      PackRun(start_width + w, ws, start_depth, depth);
+    }
+  }
+
+  // Prefetches the data that will be read by PackL1
+  void PrefetchL1(int start_width, int width, int start_depth, int depth) {
+    if (SrcMapType::kOrder == SideMapOrder::WidthMajor) {
+      for (int d = 0; d < depth; d += kDefaultCacheLineSize) {
+        for (int w = 0; w < width; w += 1) {
+          Prefetch(src_map_.data(start_width + w, start_depth + d));
+        }
+      }
+    } else {
+      for (int d = 0; d < depth; d++) {
+        for (int w = 0; w < width; w += kDefaultCacheLineSize) {
+          Prefetch(src_map_.data(start_width + w, start_depth + d));
+        }
+      }
+    }
+  }
+
+  // PackRun packs only a run i.e. is the inner loop in the depth dimension.
+  void PackRun(int start_width, int width, int start_depth, int depth) {
+    PackingRegisterBlockType b;
+    if (width == kKernelWidth) {
+      const int register_aligned_depth = RoundDown<kRegisterSize>(depth);
+      if (register_aligned_depth) {
+        for (int d = 0; d < register_aligned_depth; d += kRegisterSize) {
+          b.UseCompleteSrcInPlace(src_map_.block(start_width, start_depth + d,
+                                                 width, kRegisterSize));
+          b.Pack(packed_side_block_, start_width);
+        }
+      }
+      if (register_aligned_depth < depth) {
+        b.MakeCompleteSrc(
+            src_map_.block(start_width, start_depth + register_aligned_depth,
+                           width, depth - register_aligned_depth));
+        b.Pack(packed_side_block_, start_width);
+      }
+    } else {
+      assert(width < kKernelWidth);
+      for (int d = 0; d < depth; d += kRegisterSize) {
+        const int ds = std::min(+kRegisterSize, depth - d);
+        b.MakeCompleteSrc(
+            src_map_.block(start_width, start_depth + d, width, ds));
+        b.Pack(packed_side_block_, start_width);
+      }
+    }
+  }
+
+  // The PackedSideBlock being packed, i.e. the 'destination'.
+  PackedSideBlock* const packed_side_block_;
+
+  // A map on the block of the original matrix block being packed,
+  // i.e. the 'source'.
+  const SrcMapType& src_map_;
+};
+
+// Packs a block of the input LHS matrix, into a PackedSideBlock
+template <typename PackedSideBlock, typename MatrixMapType>
+void PackLhs(PackedSideBlock* dst, const MatrixMapType& src) {
+  ScopedProfilingLabel label("pack LHS");
+  static const SideMapOrder kSideMapOrder =
+      MatrixMapType::kOrder == MapOrder::RowMajor ? SideMapOrder::WidthMajor
+                                                  : SideMapOrder::DepthMajor;
+  typedef typename MatrixMapType::Scalar Scalar;
+  typedef SideMap<Scalar, kSideMapOrder> SideMapType;
+  SideMapType src_side_map(src.data(), src.rows(), src.cols(), src.stride());
+  typedef PackSideBlockImpl<SideMapType, PackedSideBlock> ImplType;
+  ImplType impl(dst, src_side_map);
+  impl.PackL2();
+}
+
+// Packs a block of the input RHS matrix, into a PackedSideBlock
+template <typename PackedSideBlock, typename MatrixMapType>
+void PackRhs(PackedSideBlock* dst, const MatrixMapType& src) {
+  ScopedProfilingLabel label("pack RHS");
+  static const SideMapOrder kSideMapOrder =
+      MatrixMapType::kOrder == MapOrder::ColMajor ? SideMapOrder::WidthMajor
+                                                  : SideMapOrder::DepthMajor;
+  typedef typename MatrixMapType::Scalar Scalar;
+  typedef SideMap<Scalar, kSideMapOrder> SideMapType;
+  SideMapType src_side_map(src.data(), src.cols(), src.rows(), src.stride());
+  typedef PackSideBlockImpl<SideMapType, PackedSideBlock> ImplType;
+  ImplType impl(dst, src_side_map);
+  impl.PackL2();
+}
+
+}  // namespace gemmlowp
+
+#ifdef GEMMLOWP_NEON
+#include "pack_neon.h"
+#elif defined(GEMMLOWP_SSE4)
+#include "pack_sse.h"
+#endif
+
+#endif  // GEMMLOWP_INTERNAL_PACK_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_neon.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_neon.h
new file mode 100644
index 0000000..e212d07
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_neon.h
@@ -0,0 +1,320 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// pack_neon.h: optimized NEON specializations of the templates in pack.h.
+
+#ifndef GEMMLOWP_INTERNAL_PACK_NEON_H_
+#define GEMMLOWP_INTERNAL_PACK_NEON_H_
+
+#include "pack.h"
+
+#include <arm_neon.h>
+
+namespace gemmlowp {
+
+typedef SideMap<const std::uint8_t, SideMapOrder::WidthMajor>
+    WidthMajorUint8SideMap;
+
+template <int Cells>
+using DepthMajorSideFormatNCells4x2 = KernelSideFormat<CellFormat<4, 2>, Cells>;
+
+template <int Cells>
+class PackingRegisterBlock<
+    WidthMajorUint8SideMap,
+    PackedSideBlock<DepthMajorSideFormatNCells4x2<Cells>>>
+    : public PackingRegisterBlockBase<
+          WidthMajorUint8SideMap,
+          PackedSideBlock<DepthMajorSideFormatNCells4x2<Cells>>> {
+ public:
+  typedef DepthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+  static const int kCellSize = CellFormat::kSize;
+
+  void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
+    std::uint8_t* dst_ptr = dst->current_data();
+    const std::uint8_t* const src_ptr = this->complete_src_.data();
+    const int stride = this->complete_src_.stride();
+    // Load source WidthMajor data
+    uint8x16_t src_lines[4 * kCells];
+    for (int i = 0; i < 4 * kCells; i++) {
+      src_lines[i] = vld1q_u8(src_ptr + i * stride);
+    }
+    // Reorder the data within registers to make DepthMajor 4x2 cells
+    uint8x16x2_t src_lines_intertwined_2x[2 * kCells];
+    for (int i = 0; i < kCells; i++) {
+      src_lines_intertwined_2x[2 * i] =
+          vzipq_u8(src_lines[4 * i], src_lines[4 * i + 2]);
+      src_lines_intertwined_2x[2 * i + 1] =
+          vzipq_u8(src_lines[4 * i + 1], src_lines[4 * i + 3]);
+    }
+    uint8x16x2_t src_lines_intertwined_4x[2 * kCells];
+    for (int i = 0; i < kCells; i++) {
+      src_lines_intertwined_4x[2 * i] =
+          vzipq_u8(src_lines_intertwined_2x[2 * i].val[0],
+                   src_lines_intertwined_2x[2 * i + 1].val[0]);
+      src_lines_intertwined_4x[2 * i + 1] =
+          vzipq_u8(src_lines_intertwined_2x[2 * i].val[1],
+                   src_lines_intertwined_2x[2 * i + 1].val[1]);
+    }
+    // Store the resulting DepthMajor 4x2 cells in the destination packed block
+    for (int outer = 0; outer < 2; outer++) {
+      for (int inner = 0; inner < 2; inner++) {
+        for (int cell = 0; cell < kCells; cell++) {
+          uint8x8_t value = vget_low_u8(
+              src_lines_intertwined_4x[2 * cell + outer].val[inner]);
+          vst1_u8(dst_ptr, value);
+          dst_ptr += 8;
+        }
+        for (int cell = 0; cell < kCells; cell++) {
+          uint8x8_t value = vget_high_u8(
+              src_lines_intertwined_4x[2 * cell + outer].val[inner]);
+          vst1_u8(dst_ptr, value);
+          dst_ptr += 8;
+        }
+      }
+    }
+    // Compute sums across the depth dimension
+    uint16x8_t sums_of_2_cells[kCells][4];
+    for (int outer = 0; outer < 2; outer++) {
+      for (int inner = 0; inner < 2; inner++) {
+        int i = 2 * outer + inner;
+        for (int cell = 0; cell < kCells; cell++) {
+          sums_of_2_cells[cell][i] = vaddl_u8(
+              vget_low_u8(
+                  src_lines_intertwined_4x[2 * cell + outer].val[inner]),
+              vget_high_u8(
+                  src_lines_intertwined_4x[2 * cell + outer].val[inner]));
+        }
+      }
+    }
+    int32x4_t sums_of_4_cells[kCells][4];
+    for (int i = 0; i < 4; i++) {
+      for (int cell = 0; cell < kCells; cell++) {
+        sums_of_4_cells[cell][i] = vreinterpretq_s32_u32(
+            vaddl_u16(vget_low_u16(sums_of_2_cells[cell][i]),
+                      vget_high_u16(sums_of_2_cells[cell][i])));
+      }
+    }
+    // Update the sums_of_each_slice vector
+    for (int cell = 0; cell < kCells; cell++) {
+      int32x4_t s01 =
+          vaddq_s32(sums_of_4_cells[cell][0], sums_of_4_cells[cell][1]);
+      int32x4_t s23 =
+          vaddq_s32(sums_of_4_cells[cell][2], sums_of_4_cells[cell][3]);
+      int32x4_t s = vaddq_s32(s01, s23);
+      std::int32_t* sums_of_each_slice_ptr =
+          dst->sums_of_each_slice() + start_width + 4 * cell;
+      vst1q_s32(sums_of_each_slice_ptr,
+                vaddq_s32(s, vld1q_s32(sums_of_each_slice_ptr)));
+    }
+    dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
+  }
+};
+
+template <int Cells>
+using WidthMajorSideFormatNCells4x2 =
+    KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, Cells>;
+
+template <int Cells>
+class PackingRegisterBlock<
+    WidthMajorUint8SideMap,
+    PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>>
+    : public PackingRegisterBlockBase<
+          WidthMajorUint8SideMap,
+          PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells>>> {
+ public:
+  typedef WidthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+  static const int kCellSize = CellFormat::kSize;
+
+  void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
+    std::uint8_t* dst_ptr = dst->current_data();
+    const std::uint8_t* src_ptr = this->complete_src_.data();
+    const int stride = this->complete_src_.stride();
+    // Load source WidthMajor data
+    uint16x8_t src_lines[kCells * 4];
+    for (int i = 0; i < kCells; i++) {
+// This packing path is used with our current
+// less-than-8-bit kernel, and the partial unrolling of this loop
+// results in substantially faster code (thanks to better
+// register allocation) on Nexus 5.
+
+#define GEMMLOWP_UNROLLED_LOOP_ITER(k)                            \
+  src_lines[4 * i + k] = vreinterpretq_u16_u8(vld1q_u8(src_ptr)); \
+  src_ptr += stride;
+
+      GEMMLOWP_UNROLLED_LOOP_ITER(0)
+      GEMMLOWP_UNROLLED_LOOP_ITER(1)
+      GEMMLOWP_UNROLLED_LOOP_ITER(2)
+      GEMMLOWP_UNROLLED_LOOP_ITER(3)
+
+#undef GEMMLOWP_UNROLLED_LOOP_ITER
+    }
+    // Reorder the data within registers to make WidthMajor 4x2 cells
+    uint16x8x2_t src_lines_intertwined_2x[2 * kCells];
+    for (int i = 0; i < kCells; i++) {
+      src_lines_intertwined_2x[2 * i] =
+          vzipq_u16(src_lines[4 * i], src_lines[4 * i + 2]);
+      src_lines_intertwined_2x[2 * i + 1] =
+          vzipq_u16(src_lines[4 * i + 1], src_lines[4 * i + 3]);
+    }
+    uint16x8x2_t src_lines_intertwined_4x[2 * kCells];
+    for (int i = 0; i < kCells; i++) {
+      src_lines_intertwined_4x[2 * i] =
+          vzipq_u16(src_lines_intertwined_2x[2 * i].val[0],
+                    src_lines_intertwined_2x[2 * i + 1].val[0]);
+      src_lines_intertwined_4x[2 * i + 1] =
+          vzipq_u16(src_lines_intertwined_2x[2 * i].val[1],
+                    src_lines_intertwined_2x[2 * i + 1].val[1]);
+    }
+    // Store the resulting WidthMajor 4x2 cells in the destination packed block
+    for (int outer = 0; outer < 2; outer++) {
+      for (int inner = 0; inner < 2; inner++) {
+        for (int cell = 0; cell < kCells; cell++) {
+          uint8x8_t value = vreinterpret_u8_u16(vget_low_u16(
+              src_lines_intertwined_4x[2 * cell + outer].val[inner]));
+          vst1_u8(dst_ptr, value);
+          dst_ptr += 8;
+        }
+        for (int cell = 0; cell < kCells; cell++) {
+          uint8x8_t value = vreinterpret_u8_u16(vget_high_u16(
+              src_lines_intertwined_4x[2 * cell + outer].val[inner]));
+          vst1_u8(dst_ptr, value);
+          dst_ptr += 8;
+        }
+      }
+    }
+    // Compute sums across the depth dimension
+    uint16x8_t sums_of_2[kCells][4];
+    for (int outer = 0; outer < 2; outer++) {
+      for (int inner = 0; inner < 2; inner++) {
+        int i = 2 * outer + inner;
+        for (int cell = 0; cell < kCells; cell++) {
+          sums_of_2[cell][i] = vpaddlq_u8(vreinterpretq_u8_u16(
+              src_lines_intertwined_4x[2 * cell + outer].val[inner]));
+        }
+      }
+    }
+    uint16x8_t sums_of_4[kCells][2];
+    for (int i = 0; i < 2; i++) {
+      for (int cell = 0; cell < kCells; cell++) {
+        sums_of_4[cell][i] =
+            vaddq_u16(sums_of_2[cell][2 * i], sums_of_2[cell][2 * i + 1]);
+      }
+    }
+    uint16x8_t sums_of_8[kCells];
+    for (int cell = 0; cell < kCells; cell++) {
+      sums_of_8[cell] = vaddq_u16(sums_of_4[cell][0], sums_of_4[cell][1]);
+    }
+
+    uint16x4_t sums_of_16[kCells];
+    for (int cell = 0; cell < kCells; cell++) {
+      sums_of_16[cell] = vadd_u16(vget_low_u16(sums_of_8[cell]),
+                                  vget_high_u16(sums_of_8[cell]));
+    }
+    // Update the sums_of_each_slice vector
+    for (int cell = 0; cell < kCells; cell++) {
+      int32x4_t s = vreinterpretq_s32_u32(vmovl_u16(sums_of_16[cell]));
+      std::int32_t* sums_of_each_slice_ptr =
+          dst->sums_of_each_slice() + start_width + 4 * cell;
+      vst1q_s32(sums_of_each_slice_ptr,
+                vaddq_s32(s, vld1q_s32(sums_of_each_slice_ptr)));
+    }
+    dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
+  }
+};
+
+#ifdef GEMMLOWP_NEON_32
+inline int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+  const int16x4_t c = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+  const int16x4_t d = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+  return vcombine_s16(c, d);
+}
+#endif
+
+template <int Width>
+using Int8FastKernelFormat =
+    KernelSideFormatInt8<CellFormat<Width, 16, CellOrder::WidthMajor>, 1>;
+
+template <int Width>
+class PackingRegisterBlock<WidthMajorUint8SideMap,
+                           PackedSideBlock<Int8FastKernelFormat<Width>>>
+    : public PackingRegisterBlockBase<
+          WidthMajorUint8SideMap,
+          PackedSideBlock<Int8FastKernelFormat<Width>>> {
+ public:
+  static_assert(Width == 2 || Width == 4, "");
+  typedef Int8FastKernelFormat<Width> KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+  static const int kCellSize = CellFormat::kSize;
+
+  void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
+    std::int32_t* sums_ptr = dst->sums_of_each_slice() + start_width;
+    std::uint8_t* dst_ptr = dst->current_data();
+    const std::uint8_t* const src_ptr = this->complete_src_.data();
+    const int stride = this->complete_src_.stride();
+    // Load source WidthMajor data
+    uint8x16_t src_lines[Width];
+    for (int i = 0; i < Width; i++) {
+      src_lines[i] = vld1q_u8(src_ptr + i * stride);
+    }
+    const uint8x16_t sign_bit_dup = vdupq_n_u8(0x80);
+    for (int i = 0; i < Width; i++) {
+      src_lines[i] = veorq_u8(src_lines[i], sign_bit_dup);
+    }
+    for (int i = 0; i < Width; i++) {
+      vst1q_u8(dst_ptr + 16 * i, src_lines[i]);
+    }
+    int16x8_t sums2[Width];
+    for (int i = 0; i < Width; i++) {
+      const int8x8_t lo = vreinterpret_s8_u8(vget_low_u8(src_lines[i]));
+      const int8x8_t hi = vreinterpret_s8_u8(vget_high_u8(src_lines[i]));
+      sums2[i] = vaddl_s8(lo, hi);
+    }
+    int16x8_t sums4[Width / 2];
+    for (int i = 0; i < Width / 2; i++) {
+      sums4[i] = vpaddq_s16(sums2[2 * i], sums2[2 * i + 1]);
+    }
+    if (Width == 4) {
+      int32x4_t sum = vld1q_s32(sums_ptr);
+      int16x8_t sums8 = vpaddq_s16(sums4[0], sums4[1]);
+      sum = vpadalq_s16(sum, sums8);
+      vst1q_s32(sums_ptr, sum);
+    } else {
+      assert(Width == 2);
+      int32x2_t sum = vld1_s32(sums_ptr);
+      int16x4_t sums8 =
+          vpadd_s16(vget_low_s16(sums4[0]), vget_high_s16(sums4[0]));
+      sum = vpadal_s16(sum, sums8);
+      vst1_s32(sums_ptr, sum);
+    }
+    dst->seek_forward_n_cells(1);
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_PACK_NEON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_sse.h
new file mode 100644
index 0000000..52163c4
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/pack_sse.h
@@ -0,0 +1,128 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// pack_SSE.h: optimized SSE specializations of the templates in pack.h.
+
+#ifndef GEMMLOWP_INTERNAL_PACK_SSE_H_
+#define GEMMLOWP_INTERNAL_PACK_SSE_H_
+
+#include <smmintrin.h>
+#include "pack.h"
+
+namespace gemmlowp {
+
+// TODO: Add DepthMajorUint8SideMap
+
+typedef SideMap<const std::uint8_t, SideMapOrder::WidthMajor>
+    WidthMajorUint8SideMap;
+
+template <int Cells>
+using WidthMajorSideFormatNCells4x2 =
+    KernelSideFormat<CellFormat<4, 2, CellOrder::WidthMajor>, Cells>;
+
+template <int Cells>
+class PackingRegisterBlock<
+    WidthMajorUint8SideMap,
+    PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells> > >
+    : public PackingRegisterBlockBase<
+          WidthMajorUint8SideMap,
+          PackedSideBlock<WidthMajorSideFormatNCells4x2<Cells> > > {
+ public:
+  typedef WidthMajorSideFormatNCells4x2<Cells> KernelSideFormat;
+  typedef typename KernelSideFormat::Cell CellFormat;
+  static const int kCells = KernelSideFormat::kCells;
+  static const int kCellWidth = CellFormat::kWidth;
+  static const int kKernelWidth = CellFormat::kWidth * kCells;
+  static const int kCellDepth = CellFormat::kDepth;
+  static const int kCellSize = CellFormat::kSize;
+
+  void Pack(PackedSideBlock<KernelSideFormat>* dst, int start_width) {
+    std::uint8_t* dst_ptr = dst->current_data();
+    const int width_stride = this->complete_src_.width_stride();
+    int depth_step = 8;
+
+    __m128i one = _mm_set1_epi16(1);
+    for (int cell_start_depth = 0; cell_start_depth < kRegisterSize;
+         cell_start_depth += depth_step) {
+      for (int cell_start_width = 0; cell_start_width < kKernelWidth;
+           cell_start_width += kCellWidth) {
+        std::int32_t* cell_sums_of_each_slice_ptr =
+            dst->sums_of_each_slice() + start_width + cell_start_width;
+        const std::uint8_t* src_data =
+            this->complete_src_.data(cell_start_width, cell_start_depth);
+
+        __m128i xmm1 =
+            _mm_loadl_epi64(reinterpret_cast<const __m128i*>(&src_data[0]));
+        __m128i xmm2 = _mm_loadl_epi64(
+            reinterpret_cast<const __m128i*>(&src_data[1 * width_stride]));
+        __m128i xmm3 = _mm_loadl_epi64(
+            reinterpret_cast<const __m128i*>(&src_data[2 * width_stride]));
+        __m128i xmm4 = _mm_loadl_epi64(
+            reinterpret_cast<const __m128i*>(&src_data[3 * width_stride]));
+
+        __m128i xmm5 = _mm_unpacklo_epi16(xmm1, xmm2);
+        __m128i xmm8 = _mm_shuffle_epi32(xmm5, 0x31);
+
+        __m128i xmm6 = _mm_unpacklo_epi16(xmm3, xmm4);
+        __m128i xmm7 = _mm_shuffle_epi32(xmm6, 0x80);
+
+        __m128i xmm9 = _mm_blend_epi16(xmm5, xmm7, 0xcc);
+        __m128i xmm10 = _mm_blend_epi16(xmm8, xmm6, 0xcc);
+
+        _mm_storel_epi64(reinterpret_cast<__m128i*>(&dst_ptr[0]), xmm9);
+        _mm_storel_epi64(
+            reinterpret_cast<__m128i*>(&dst_ptr[kCellSize * kCells]), xmm10);
+
+        __m128i xmm11 = _mm_shuffle_epi32(xmm9, 0xee);
+        __m128i xmm12 = _mm_shuffle_epi32(xmm10, 0xee);
+
+        _mm_storel_epi64(
+            reinterpret_cast<__m128i*>(&dst_ptr[2 * kCellSize * kCells]),
+            xmm11);
+        _mm_storel_epi64(
+            reinterpret_cast<__m128i*>(&dst_ptr[3 * kCellSize * kCells]),
+            xmm12);
+
+        xmm1 = _mm_cvtepu8_epi16(xmm9);
+        xmm2 = _mm_madd_epi16(xmm1, one);
+        __m128i sums_of_each_slice_xmm = _mm_loadu_si128(
+            reinterpret_cast<const __m128i*>(&cell_sums_of_each_slice_ptr[0]));
+        sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
+
+        xmm1 = _mm_cvtepu8_epi16(xmm10);
+        xmm2 = _mm_madd_epi16(xmm1, one);
+        sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
+
+        xmm1 = _mm_cvtepu8_epi16(xmm11);
+        xmm2 = _mm_madd_epi16(xmm1, one);
+        sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
+
+        xmm1 = _mm_cvtepu8_epi16(xmm12);
+        xmm2 = _mm_madd_epi16(xmm1, one);
+        sums_of_each_slice_xmm = _mm_add_epi32(sums_of_each_slice_xmm, xmm2);
+
+        _mm_storeu_si128(
+            reinterpret_cast<__m128i*>(&cell_sums_of_each_slice_ptr[0]),
+            sums_of_each_slice_xmm);
+        dst_ptr += kCellSize;
+      }
+      dst_ptr += 3 * kCellSize * kCells;
+    }
+    dst->seek_forward_n_cells(kCells * kRegisterSize / kCellDepth);
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_PACK_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers.h
new file mode 100644
index 0000000..e39eaf8
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers.h
@@ -0,0 +1,508 @@
+// Copyright 2017 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// simd_wrappers.h: some inline functions wrapping SIMD intrinsics,
+// extending the set of such functions from fixedpoint.h.
+
+#ifndef GEMMLOWP_INTERNAL_SIMD_WRAPPERS_H_
+#define GEMMLOWP_INTERNAL_SIMD_WRAPPERS_H_
+
+#include <algorithm>
+#include <type_traits>
+#include "../fixedpoint/fixedpoint.h"
+
+namespace gemmlowp {
+
+template <typename ScalarType, int ScalarCount>
+struct RegisterType {
+  using Type = ScalarType;
+};
+
+inline std::int32_t Min(std::int32_t a, std::int32_t b) {
+  return std::min(a, b);
+}
+
+inline std::int32_t Max(std::int32_t a, std::int32_t b) {
+  return std::max(a, b);
+}
+
+inline void MulAdd(std::int32_t lhs, std::int32_t rhs, std::int32_t* acc) {
+  *acc += lhs * rhs;
+}
+
+template <typename tScalarType, int tScalarCount>
+struct RegisterBuffer {
+  using ScalarType = tScalarType;
+  static constexpr int kScalarCount = tScalarCount;
+  using RegisterType = typename RegisterType<ScalarType, kScalarCount>::Type;
+  static_assert((kScalarCount & (kScalarCount - 1)) == 0,
+                "kScalarCount must be a power of two");
+  static_assert(sizeof(RegisterType) % sizeof(ScalarType) == 0, "");
+  static constexpr int kRegisterLanes =
+      sizeof(RegisterType) / sizeof(ScalarType);
+  static constexpr int kRegisterCount =
+      (kScalarCount * sizeof(ScalarType) + sizeof(RegisterType) - 1) /
+      sizeof(RegisterType);
+
+  RegisterType reg[kRegisterCount];
+};
+
+template <typename tScalarType, int tRows, int tCols>
+struct RegisterBlock {
+  using ScalarType = tScalarType;
+  static constexpr int kRows = tRows;
+  static constexpr int kCols = tCols;
+  static constexpr int kScalarCount = kRows * kCols;
+  using BufferType = RegisterBuffer<ScalarType, kScalarCount>;
+  using RegisterType = typename BufferType::RegisterType;
+  static constexpr int kRegisterCount = BufferType::kRegisterCount;
+  static constexpr int kRegisterLanes = BufferType::kRegisterLanes;
+
+  BufferType buf;
+};
+
+template <typename RegisterBlockType>
+struct RegisterBlockAddImpl {
+  static RegisterBlockType Run(const RegisterBlockType& lhs,
+                               const RegisterBlockType& rhs) {
+    RegisterBlockType result;
+    for (int i = 0; i < RegisterBlockType::kRegisterCount; i++) {
+      result.buf.reg[i] = Add(lhs.buf.reg[i], rhs.buf.reg[i]);
+    }
+    return result;
+  }
+};
+
+template <typename RegisterBlockType>
+RegisterBlockType RegisterBlockAdd(const RegisterBlockType& lhs,
+                                   const RegisterBlockType& rhs) {
+  return RegisterBlockAddImpl<RegisterBlockType>::Run(lhs, rhs);
+}
+
+template <typename LhsType, typename RhsType>
+struct ShouldFlipLhsRhs {
+  static constexpr bool kValue =
+      (LhsType::kScalarCount < RhsType::kScalarCount) ||
+      (LhsType::kScalarCount == RhsType::kScalarCount &&
+       (LhsType::kRows < RhsType::kRows));
+};
+
+template <typename LhsType, typename RhsType,
+          bool Flip = ShouldFlipLhsRhs<LhsType, RhsType>::kValue>
+struct FlipLhsRhs {
+  using FlippedLhsType = LhsType;
+  using FlippedRhsType = RhsType;
+  static const FlippedLhsType& FlippedLhs(const LhsType& lhs,
+                                          const RhsType& rhs) {
+    return lhs;
+  }
+  static const FlippedRhsType& FlippedRhs(const LhsType& lhs,
+                                          const RhsType& rhs) {
+    return rhs;
+  }
+};
+
+template <typename LhsType, typename RhsType>
+struct FlipLhsRhs<LhsType, RhsType, true> {
+  using FlippedLhsType = RhsType;
+  using FlippedRhsType = LhsType;
+  static const FlippedLhsType& FlippedLhs(const LhsType& lhs,
+                                          const RhsType& rhs) {
+    return rhs;
+  }
+  static const FlippedRhsType& FlippedRhs(const LhsType& lhs,
+                                          const RhsType& rhs) {
+    return lhs;
+  }
+};
+
+template <typename Lhs, typename Rhs>
+struct BroadcastBinaryOpShape {
+  static constexpr int kRows =
+      Lhs::kRows > Rhs::kRows ? Lhs::kRows : Rhs::kRows;
+  static constexpr int kCols =
+      Lhs::kCols > Rhs::kCols ? Lhs::kCols : Rhs::kCols;
+};
+
+template <typename Lhs, typename Rhs>
+struct BroadcastBinaryOpRegisterBlock {
+  using Shape = BroadcastBinaryOpShape<Lhs, Rhs>;
+  using ScalarType = typename Lhs::ScalarType;
+  using Type = RegisterBlock<ScalarType, Shape::kRows, Shape::kCols>;
+};
+
+template <typename Lhs, typename Rhs>
+struct BroadcastAddImpl {
+  using ResultBlockType =
+      typename BroadcastBinaryOpRegisterBlock<Lhs, Rhs>::Type;
+  static ResultBlockType Run(const Lhs& lhs, const Rhs& rhs) {
+    ResultBlockType result;
+    static constexpr int Rows = ResultBlockType::kRows;
+    static constexpr int Cols = ResultBlockType::kCols;
+    static constexpr int LhsRows = Lhs::kRows;
+    static constexpr int LhsCols = Lhs::kCols;
+    static constexpr int RhsRows = Rhs::kRows;
+    static constexpr int RhsCols = Rhs::kCols;
+
+    static_assert(LhsRows == Rows || LhsRows == 1, "");
+    static_assert(RhsRows == Rows || RhsRows == 1, "");
+    static_assert(LhsCols == Cols || LhsCols == 1, "");
+    static_assert(RhsCols == Cols || RhsCols == 1, "");
+    static_assert(ResultBlockType::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Lhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Rhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+
+    for (int c = 0; c < Cols; c++) {
+      const int lhs_c = LhsCols == Cols ? c : 0;
+      const int rhs_c = RhsCols == Cols ? c : 0;
+      for (int r = 0; r < Rows; r++) {
+        const int lhs_r = LhsRows == Rows ? r : 0;
+        const int rhs_r = RhsRows == Rows ? r : 0;
+        result.buf.reg[r + c * Rows] =
+            Add(lhs.buf.reg[lhs_r + lhs_c * LhsRows],
+                rhs.buf.reg[rhs_r + rhs_c * RhsRows]);
+      }
+    }
+    return result;
+  }
+};
+
+template <typename Lhs, typename Rhs>
+typename BroadcastBinaryOpRegisterBlock<Lhs, Rhs>::Type BroadcastAdd(
+    const Lhs& lhs, const Rhs& rhs) {
+  using Flip = FlipLhsRhs<Lhs, Rhs>;
+  return BroadcastAddImpl<
+      typename Flip::FlippedLhsType,
+      typename Flip::FlippedRhsType>::Run(Flip::FlippedLhs(lhs, rhs),
+                                          Flip::FlippedRhs(lhs, rhs));
+}
+
+template <typename Lhs, typename Rhs>
+struct BroadcastMulImpl {
+  using ResultBlockType =
+      typename BroadcastBinaryOpRegisterBlock<Lhs, Rhs>::Type;
+  static ResultBlockType Run(const Lhs& lhs, const Rhs& rhs) {
+    ResultBlockType result;
+    static constexpr int Rows = ResultBlockType::kRows;
+    static constexpr int Cols = ResultBlockType::kCols;
+    static constexpr int LhsRows = Lhs::kRows;
+    static constexpr int LhsCols = Lhs::kCols;
+    static constexpr int RhsRows = Rhs::kRows;
+    static constexpr int RhsCols = Rhs::kCols;
+    static_assert(ResultBlockType::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Lhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Rhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+
+    static_assert(LhsRows == Rows || LhsRows == 1, "");
+    static_assert(RhsRows == Rows || RhsRows == 1, "");
+    static_assert(LhsCols == Cols || LhsCols == 1, "");
+    static_assert(RhsCols == Cols || RhsCols == 1, "");
+    for (int c = 0; c < Cols; c++) {
+      const int lhs_c = LhsCols == Cols ? c : 0;
+      const int rhs_c = RhsCols == Cols ? c : 0;
+      for (int r = 0; r < Rows; r++) {
+        const int lhs_r = LhsRows == Rows ? r : 0;
+        const int rhs_r = RhsRows == Rows ? r : 0;
+        result.buf.reg[r + c * Rows] =
+            Mul(lhs.buf.reg[lhs_r + lhs_c * LhsRows],
+                rhs.buf.reg[rhs_r + rhs_c * RhsRows]);
+      }
+    }
+    return result;
+  }
+};
+
+template <typename Lhs, typename Rhs>
+typename BroadcastBinaryOpRegisterBlock<Lhs, Rhs>::Type BroadcastMul(
+    const Lhs& lhs, const Rhs& rhs) {
+  using Flip = FlipLhsRhs<Lhs, Rhs>;
+  return BroadcastMulImpl<
+      typename Flip::FlippedLhsType,
+      typename Flip::FlippedRhsType>::Run(Flip::FlippedLhs(lhs, rhs),
+                                          Flip::FlippedRhs(lhs, rhs));
+}
+
+template <typename Lhs, typename Rhs, typename Acc>
+struct BroadcastMulAddImpl {
+  static void Run(const Lhs& lhs, const Rhs& rhs, Acc* acc) {
+    static constexpr int Rows = Acc::kRows;
+    static constexpr int Cols = Acc::kCols;
+    static constexpr int LhsRows = Lhs::kRows;
+    static constexpr int LhsCols = Lhs::kCols;
+    static constexpr int RhsRows = Rhs::kRows;
+    static constexpr int RhsCols = Rhs::kCols;
+    static_assert(Acc::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Lhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+    static_assert(Rhs::kRegisterLanes == 1,
+                  "This path is only for scalar values");
+
+    static_assert(LhsRows == Rows || LhsRows == 1, "");
+    static_assert(RhsRows == Rows || RhsRows == 1, "");
+    static_assert(LhsCols == Cols || LhsCols == 1, "");
+    static_assert(RhsCols == Cols || RhsCols == 1, "");
+    for (int c = 0; c < Cols; c++) {
+      const int lhs_c = LhsCols == Cols ? c : 0;
+      const int rhs_c = RhsCols == Cols ? c : 0;
+      for (int r = 0; r < Rows; r++) {
+        const int lhs_r = LhsRows == Rows ? r : 0;
+        const int rhs_r = RhsRows == Rows ? r : 0;
+        MulAdd(lhs.buf.reg[lhs_r + lhs_c * LhsRows],
+               rhs.buf.reg[rhs_r + rhs_c * RhsRows],
+               &acc->buf.reg[r + c * Rows]);
+      }
+    }
+  }
+};
+
+template <typename Lhs, typename Rhs, typename Acc>
+void BroadcastMulAdd(const Lhs& lhs, const Rhs& rhs, Acc* acc) {
+  using Flip = FlipLhsRhs<Lhs, Rhs>;
+  BroadcastMulAddImpl<typename Flip::FlippedLhsType,
+                      typename Flip::FlippedRhsType,
+                      Acc>::Run(Flip::FlippedLhs(lhs, rhs),
+                                Flip::FlippedRhs(lhs, rhs), acc);
+}
+
+template <typename RegisterBlockType, typename SrcObjectType>
+struct LoadImpl {
+  static_assert(std::is_same<SrcObjectType, void>::value,
+                "This generic impl should never be hit");
+};
+
+template <typename ScalarType, int Rows, int Cols, typename SrcScalarType>
+struct LoadImpl<RegisterBlock<ScalarType, Rows, Cols>,
+                MatrixMap<SrcScalarType, MapOrder::ColMajor>> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  using SrcObjectType = MatrixMap<SrcScalarType, MapOrder::ColMajor>;
+  static RegisterBlockType Run(const SrcObjectType& src, int row, int col) {
+    RegisterBlockType result;
+    int i = 0;
+    for (int c = 0; c < Cols; c++) {
+      const ScalarType* src_ptr = src.data(row, col + c);
+      for (int r = 0; r < Rows; r++) {
+        result.buf.reg[i++] = *src_ptr++;
+      }
+    }
+    return result;
+  }
+};
+
+template <typename ScalarType, int Rows, int Cols, typename SrcScalarType,
+          VectorShape Shape>
+struct LoadImpl<RegisterBlock<ScalarType, Rows, Cols>,
+                VectorMap<SrcScalarType, Shape>> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  using SrcObjectType = VectorMap<SrcScalarType, Shape>;
+  static RegisterBlockType Run(const SrcObjectType& src, int pos) {
+    static_assert(Shape == VectorShape::Col || Rows == 1, "");
+    static_assert(Shape == VectorShape::Row || Cols == 1, "");
+    RegisterBlockType result;
+    for (int i = 0; i < Rows * Cols; i++) {
+      result.buf.reg[i] = src(pos + i);
+    }
+    return result;
+  }
+};
+
+template <typename ScalarType, int Rows, int Cols, typename SrcScalarType,
+          VectorShape Shape>
+struct LoadImpl<RegisterBlock<ScalarType, Rows, Cols>,
+                VectorDup<SrcScalarType, Shape>> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  using SrcObjectType = VectorDup<SrcScalarType, Shape>;
+  static RegisterBlockType Run(const SrcObjectType& src, int) {
+    static_assert(Shape == VectorShape::Col || Rows == 1, "");
+    static_assert(Shape == VectorShape::Row || Cols == 1, "");
+    RegisterBlockType result;
+    for (int i = 0; i < Rows * Cols; i++) {
+      result.buf.reg[i] = src(0);
+    }
+    return result;
+  }
+};
+
+template <typename RegisterBlockType, typename SrcObjectType>
+RegisterBlockType Load(const SrcObjectType& src, int row, int col) {
+  return LoadImpl<RegisterBlockType, SrcObjectType>::Run(src, row, col);
+}
+
+template <typename RegisterBlockType, typename SrcObjectType>
+RegisterBlockType Load(const SrcObjectType& src, int pos) {
+  return LoadImpl<RegisterBlockType, SrcObjectType>::Run(src, pos);
+}
+
+template <typename RegisterBlockType>
+struct LoadContiguousImpl {
+  using ScalarType = typename RegisterBlockType::ScalarType;
+  static_assert(RegisterBlockType::kRegisterLanes == 1,
+                "This path is only for scalar values");
+  static RegisterBlockType Run(const ScalarType* src) {
+    RegisterBlockType result;
+    for (int i = 0; i < RegisterBlockType::kScalarCount; i++) {
+      result.buf.reg[i] = src[i];
+    }
+    return result;
+  }
+};
+
+template <typename RegisterBlockType>
+RegisterBlockType LoadContiguous(
+    const typename RegisterBlockType::ScalarType* src) {
+  return LoadContiguousImpl<RegisterBlockType>::Run(src);
+}
+
+template <int BroadcastRows, int BroadcastCols, typename SrcObjectType>
+struct LoadForBroadcastingShape {};
+
+template <int BroadcastRows, int BroadcastCols, typename ScalarType,
+          VectorShape Shape>
+struct LoadForBroadcastingShape<BroadcastRows, BroadcastCols,
+                                VectorMap<ScalarType, Shape>> {
+  static constexpr int kRows = Shape == VectorShape::Col ? BroadcastRows : 1;
+  static constexpr int kCols = Shape == VectorShape::Row ? BroadcastCols : 1;
+};
+
+template <int BroadcastRows, int BroadcastCols, typename ScalarType,
+          VectorShape Shape>
+struct LoadForBroadcastingShape<BroadcastRows, BroadcastCols,
+                                VectorDup<ScalarType, Shape>> {
+  static constexpr int kRows = 1;
+  static constexpr int kCols = 1;
+};
+
+template <typename RegisterBlockType, typename SrcObjectType>
+struct LoadForBroadcastingRegisterBlock {
+  using Shape =
+      LoadForBroadcastingShape<RegisterBlockType::kRows,
+                               RegisterBlockType::kCols, SrcObjectType>;
+  using ScalarType = typename RegisterBlockType::ScalarType;
+  using Type = RegisterBlock<ScalarType, Shape::kRows, Shape::kCols>;
+};
+
+template <typename RegisterBlockType, typename SrcObjectType>
+struct LoadForBroadcastingImpl {
+  static_assert(std::is_same<SrcObjectType, void>::value,
+                "This generic impl should never be hit");
+};
+
+template <typename ScalarType, int Rows, int Cols, typename SrcScalarType,
+          VectorShape Shape>
+struct LoadForBroadcastingImpl<RegisterBlock<ScalarType, Rows, Cols>,
+                               VectorMap<SrcScalarType, Shape>> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  using SrcObjectType = VectorMap<SrcScalarType, Shape>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+  static_assert(ResultBlockType::kRegisterLanes == 1,
+                "This path is only for scalar values");
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    for (int c = 0; c < ResultBlockType::kCols; c++) {
+      for (int r = 0; r < ResultBlockType::kRows; r++) {
+        const int i = Shape == VectorShape::Col ? r : c;
+        result.buf.reg[r + c * ResultBlockType::kRows] = src(pos + i);
+      }
+    }
+    return result;
+  }
+};
+
+template <typename ScalarType, int Rows, int Cols, typename SrcScalarType,
+          VectorShape Shape>
+struct LoadForBroadcastingImpl<RegisterBlock<ScalarType, Rows, Cols>,
+                               VectorDup<SrcScalarType, Shape>> {
+  using RegisterBlockType = RegisterBlock<ScalarType, Rows, Cols>;
+  using SrcObjectType = VectorDup<SrcScalarType, Shape>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+  static_assert(ResultBlockType::kRegisterLanes == 1,
+                "This path is only for scalar values");
+  static ResultBlockType Run(const SrcObjectType& src, int) {
+    ResultBlockType result;
+    for (int c = 0; c < ResultBlockType::kCols; c++) {
+      for (int r = 0; r < ResultBlockType::kRows; r++) {
+        result.buf.reg[r + c * ResultBlockType::kRows] = src(0);
+      }
+    }
+    return result;
+  }
+};
+
+template <typename RegisterBlockType, typename SrcObjectType>
+typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                          SrcObjectType>::Type
+LoadForBroadcasting(const SrcObjectType& src, int row, int col) {
+  return LoadForBroadcastingImpl<RegisterBlockType, SrcObjectType>::Run(
+      src, row, col);
+}
+
+template <typename RegisterBlockType, typename SrcObjectType>
+typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                          SrcObjectType>::Type
+LoadForBroadcasting(const SrcObjectType& src, int pos) {
+  return LoadForBroadcastingImpl<RegisterBlockType, SrcObjectType>::Run(src,
+                                                                        pos);
+}
+
+template <int ConstantValue, typename RegisterBlockType>
+struct AddConstantImpl {
+  static void Run(RegisterBlockType* block) {
+    using RegisterType = typename RegisterBlockType::RegisterType;
+    const RegisterType dup = Dup<RegisterType>(ConstantValue);
+    for (int i = 0; i < RegisterBlockType::kRegisterCount; i++) {
+      block->buf.reg[i] = Add(block->buf.reg[i], dup);
+    }
+  }
+};
+
+template <typename RegisterBlockType>
+struct AddConstantImpl<0, RegisterBlockType> {
+  static void Run(RegisterBlockType*) {
+    // This is a no-op.
+  }
+};
+
+template <int ConstantValue, typename RegisterBlockType>
+void AddConstant(RegisterBlockType* block) {
+  AddConstantImpl<ConstantValue, RegisterBlockType>::Run(block);
+}
+
+template <int N>
+using RegBufferInt32 = RegisterBuffer<std::int32_t, N>;
+template <int N>
+using RegBufferUint8 = RegisterBuffer<std::uint8_t, N>;
+template <int R, int C>
+using RegBlockInt32 = RegisterBlock<std::int32_t, R, C>;
+template <int R, int C>
+using RegBlockUint8 = RegisterBlock<std::uint8_t, R, C>;
+
+}  // end namespace gemmlowp
+
+#if defined GEMMLOWP_NEON
+#include "simd_wrappers_neon.h"
+#elif defined GEMMLOWP_SSE4
+#include "simd_wrappers_sse.h"
+#endif
+
+#endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_common_neon_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_common_neon_sse.h
new file mode 100644
index 0000000..3830eb1
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_common_neon_sse.h
@@ -0,0 +1,646 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// simd_wrappers_common_neon_sse.h: common SIMD (NEON and SSE) wrapper code
+
+#ifndef GEMMLOWP_INTERNAL_SIMD_WRAPPERS_COMMON_NEON_SSE_H_
+#define GEMMLOWP_INTERNAL_SIMD_WRAPPERS_COMMON_NEON_SSE_H_
+
+#include "simd_wrappers.h"
+
+namespace gemmlowp {
+
+template <typename SrcScalarType, int N>
+struct LoadImpl<RegBlockInt32<4, N>,
+                MatrixMap<SrcScalarType, MapOrder::ColMajor>> {
+  static RegBlockInt32<4, N> Run(
+      const MatrixMap<SrcScalarType, MapOrder::ColMajor>& src, int row,
+      int col) {
+    RegBlockInt32<4, N> result;
+    for (int i = 0; i < N; i++) {
+      result.buf.reg[i] = LoadInt32x4(src.data(row, col + i));
+    }
+    return result;
+  }
+};
+
+template <typename SrcScalarType, int N>
+struct LoadImpl<RegBlockInt32<8, N>,
+                MatrixMap<SrcScalarType, MapOrder::ColMajor>> {
+  static RegBlockInt32<8, N> Run(
+      const MatrixMap<SrcScalarType, MapOrder::ColMajor>& src, int row,
+      int col) {
+    RegBlockInt32<8, N> result;
+    for (int i = 0; i < N; i++) {
+      result.buf.reg[2 * i + 0] = LoadInt32x4(src.data(row + 0, col + i));
+      result.buf.reg[2 * i + 1] = LoadInt32x4(src.data(row + 4, col + i));
+    }
+    return result;
+  }
+};
+
+template <typename SrcScalarType>
+struct LoadImpl<RegBlockInt32<1, 4>,
+                MatrixMap<SrcScalarType, MapOrder::ColMajor>> {
+  static RegBlockInt32<1, 4> Run(
+      const MatrixMap<SrcScalarType, MapOrder::ColMajor>& src, int row,
+      int col) {
+    RegBlockInt32<1, 4> result;
+    std::int32_t buf[4];
+    for (int i = 0; i < 4; i++) {
+      buf[i] = src(row, col + i);
+    }
+    result.buf.reg[0] = LoadInt32x4(buf);
+    return result;
+  }
+};
+
+template <typename SrcScalarType>
+struct LoadImpl<RegBlockInt32<1, 8>,
+                MatrixMap<SrcScalarType, MapOrder::ColMajor>> {
+  static RegBlockInt32<1, 8> Run(
+      const MatrixMap<SrcScalarType, MapOrder::ColMajor>& src, int row,
+      int col) {
+    RegBlockInt32<1, 8> result;
+    std::int32_t buf[8];
+    for (int i = 0; i < 8; i++) {
+      buf[i] = src(row, col + i);
+    }
+    result.buf.reg[0] = LoadInt32x4(buf);
+    result.buf.reg[1] = LoadInt32x4(buf + 4);
+    return result;
+  }
+};
+
+template <typename SrcScalarType>
+struct LoadImpl<RegBlockInt32<4, 1>,
+                VectorMap<SrcScalarType, VectorShape::Col>> {
+  static RegBlockInt32<4, 1> Run(
+      const VectorMap<SrcScalarType, VectorShape::Col>& src, int pos) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = LoadInt32x4(src.data(pos));
+    return result;
+  }
+};
+
+template <typename SrcScalarType>
+struct LoadImpl<RegBlockInt32<4, 1>,
+                VectorDup<SrcScalarType, VectorShape::Col>> {
+  static RegBlockInt32<4, 1> Run(
+      const VectorDup<SrcScalarType, VectorShape::Col>& src, int) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = LoadInt32x4(src(0));
+    return result;
+  }
+};
+
+template <typename SrcScalarType, int N>
+struct LoadForBroadcastingImpl<RegBlockInt32<4, N>,
+                               VectorMap<SrcScalarType, VectorShape::Col>> {
+  using SrcObjectType = VectorMap<SrcScalarType, VectorShape::Col>;
+  using RegisterBlockType = RegBlockInt32<4, N>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    static_assert(ResultBlockType::kRegisterCount == 1, "");
+    result.buf.reg[0] = LoadInt32x4(src.data(pos));
+    return result;
+  }
+};
+
+template <typename SrcScalarType, int N>
+struct LoadForBroadcastingImpl<RegBlockInt32<8, N>,
+                               VectorMap<SrcScalarType, VectorShape::Col>> {
+  using SrcObjectType = VectorMap<SrcScalarType, VectorShape::Col>;
+  using RegisterBlockType = RegBlockInt32<8, N>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    static_assert(ResultBlockType::kRegisterCount == 2, "");
+    result.buf.reg[0] = LoadInt32x4(src.data(pos));
+    result.buf.reg[1] = LoadInt32x4(src.data(pos + 4));
+    return result;
+  }
+};
+
+template <typename SrcScalarType>
+struct LoadForBroadcastingImpl<RegBlockInt32<4, 1>,
+                               VectorMap<SrcScalarType, VectorShape::Row>> {
+  using SrcObjectType = VectorMap<SrcScalarType, VectorShape::Row>;
+  using RegisterBlockType = RegBlockInt32<4, 1>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    result.buf.reg[0] = src(pos);
+    return result;
+  }
+};
+
+template <typename SrcScalarType, int N>
+struct LoadForBroadcastingImpl<RegBlockInt32<N, 4>,
+                               VectorMap<SrcScalarType, VectorShape::Row>> {
+  using SrcObjectType = VectorMap<SrcScalarType, VectorShape::Row>;
+  using RegisterBlockType = RegBlockInt32<N, 4>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    static_assert(ResultBlockType::kRegisterCount == 1, "");
+    result.buf.reg[0] = LoadInt32x4(src.data(pos));
+    return result;
+  }
+};
+
+template <typename SrcScalarType, int N>
+struct LoadForBroadcastingImpl<RegBlockInt32<N, 8>,
+                               VectorMap<SrcScalarType, VectorShape::Row>> {
+  using SrcObjectType = VectorMap<SrcScalarType, VectorShape::Row>;
+  using RegisterBlockType = RegBlockInt32<N, 8>;
+  using ResultBlockType =
+      typename LoadForBroadcastingRegisterBlock<RegisterBlockType,
+                                                SrcObjectType>::Type;
+
+  static ResultBlockType Run(const SrcObjectType& src, int pos) {
+    ResultBlockType result;
+    static_assert(ResultBlockType::kRegisterCount == 2, "");
+    result.buf.reg[0] = LoadInt32x4(src.data(pos));
+    result.buf.reg[1] = LoadInt32x4(src.data(pos + 4));
+    return result;
+  }
+};
+
+// 4x1 := 4x1 + 1x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<4, 1>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<4, 1> Run(const RegBlockInt32<4, 1>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], Dup<Int32x4>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 1x4 := 1x4 + 1x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<1, 4> Run(const RegBlockInt32<1, 4>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<1, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], Dup<Int32x4>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 4x1 := 4x1 + 4x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<4, 1>, RegBlockInt32<4, 1>> {
+  static RegBlockInt32<4, 1> Run(const RegBlockInt32<4, 1>& lhs,
+                                 const RegBlockInt32<4, 1>& rhs) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 1x4 := 1x4 + 1x4
+template <>
+struct BroadcastAddImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<1, 4> Run(const RegBlockInt32<1, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<1, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 4x4 := 4x4 + 1x4
+template <>
+struct BroadcastAddImpl<RegBlockInt32<4, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<4, 4> Run(const RegBlockInt32<4, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<4, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], DupLane<0>(rhs.buf.reg[0]));
+    result.buf.reg[1] = Add(lhs.buf.reg[1], DupLane<1>(rhs.buf.reg[0]));
+    result.buf.reg[2] = Add(lhs.buf.reg[2], DupLane<2>(rhs.buf.reg[0]));
+    result.buf.reg[3] = Add(lhs.buf.reg[3], DupLane<3>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 4x4 := 4x4 + 4x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<4, 4>, RegBlockInt32<4, 1>> {
+  static RegBlockInt32<4, 4> Run(const RegBlockInt32<4, 4>& lhs,
+                                 const RegBlockInt32<4, 1>& rhs) {
+    RegBlockInt32<4, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], rhs.buf.reg[0]);
+    result.buf.reg[1] = Add(lhs.buf.reg[1], rhs.buf.reg[0]);
+    result.buf.reg[2] = Add(lhs.buf.reg[2], rhs.buf.reg[0]);
+    result.buf.reg[3] = Add(lhs.buf.reg[3], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 8x1 := 8x1 + 1x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<8, 1>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<8, 1> Run(const RegBlockInt32<8, 1>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<8, 1> result;
+    const Int32x4 p = Dup<Int32x4>(rhs.buf.reg[0]);
+    for (int i = 0; i < 2; i++) {
+      result.buf.reg[i] = Add(lhs.buf.reg[i], p);
+    }
+    return result;
+  }
+};
+
+// 8x1 := 8x1 + 8x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<8, 1>, RegBlockInt32<8, 1>> {
+  static RegBlockInt32<8, 1> Run(const RegBlockInt32<8, 1>& lhs,
+                                 const RegBlockInt32<8, 1>& rhs) {
+    RegBlockInt32<8, 1> result;
+    for (int i = 0; i < 2; i++) {
+      result.buf.reg[i] = Add(lhs.buf.reg[i], rhs.buf.reg[i]);
+    }
+    return result;
+  }
+};
+
+// 8x4 := 8x4 + 1x4
+template <>
+struct BroadcastAddImpl<RegBlockInt32<8, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<8, 4> Run(const RegBlockInt32<8, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<8, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], DupLane<0>(rhs.buf.reg[0]));
+    result.buf.reg[1] = Add(lhs.buf.reg[1], DupLane<0>(rhs.buf.reg[0]));
+    result.buf.reg[2] = Add(lhs.buf.reg[2], DupLane<1>(rhs.buf.reg[0]));
+    result.buf.reg[3] = Add(lhs.buf.reg[3], DupLane<1>(rhs.buf.reg[0]));
+    result.buf.reg[4] = Add(lhs.buf.reg[4], DupLane<2>(rhs.buf.reg[0]));
+    result.buf.reg[5] = Add(lhs.buf.reg[5], DupLane<2>(rhs.buf.reg[0]));
+    result.buf.reg[6] = Add(lhs.buf.reg[6], DupLane<3>(rhs.buf.reg[0]));
+    result.buf.reg[7] = Add(lhs.buf.reg[7], DupLane<3>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 8x4 := 8x4 + 8x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<8, 4>, RegBlockInt32<8, 1>> {
+  static RegBlockInt32<8, 4> Run(const RegBlockInt32<8, 4>& lhs,
+                                 const RegBlockInt32<8, 1>& rhs) {
+    RegBlockInt32<8, 4> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], rhs.buf.reg[0]);
+    result.buf.reg[1] = Add(lhs.buf.reg[1], rhs.buf.reg[1]);
+    result.buf.reg[2] = Add(lhs.buf.reg[2], rhs.buf.reg[0]);
+    result.buf.reg[3] = Add(lhs.buf.reg[3], rhs.buf.reg[1]);
+    result.buf.reg[4] = Add(lhs.buf.reg[4], rhs.buf.reg[0]);
+    result.buf.reg[5] = Add(lhs.buf.reg[5], rhs.buf.reg[1]);
+    result.buf.reg[6] = Add(lhs.buf.reg[6], rhs.buf.reg[0]);
+    result.buf.reg[7] = Add(lhs.buf.reg[7], rhs.buf.reg[1]);
+    return result;
+  }
+};
+
+// 1x8 := 1x8 + 1x8
+template <>
+struct BroadcastAddImpl<RegBlockInt32<1, 8>, RegBlockInt32<1, 8>> {
+  static RegBlockInt32<1, 8> Run(const RegBlockInt32<1, 8>& lhs,
+                                 const RegBlockInt32<1, 8>& rhs) {
+    RegBlockInt32<1, 8> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], rhs.buf.reg[0]);
+    result.buf.reg[1] = Add(lhs.buf.reg[1], rhs.buf.reg[1]);
+    return result;
+  }
+};
+
+// 1x8 := 1x8 + 1x1
+template <>
+struct BroadcastAddImpl<RegBlockInt32<1, 8>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<1, 8> Run(const RegBlockInt32<1, 8>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<1, 8> result;
+    result.buf.reg[0] = Add(lhs.buf.reg[0], Dup<Int32x4>(rhs.buf.reg[0]));
+    result.buf.reg[1] = Add(lhs.buf.reg[1], Dup<Int32x4>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 4x1 := 4x1 * 1x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<4, 1>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<4, 1> Run(const RegBlockInt32<4, 1>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = Mul(lhs.buf.reg[0], Dup<Int32x4>(rhs.buf.reg[0]));
+    return result;
+  }
+};
+
+// 4x1 := 4x1 * 4x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<4, 1>, RegBlockInt32<4, 1>> {
+  static RegBlockInt32<4, 1> Run(const RegBlockInt32<4, 1>& lhs,
+                                 const RegBlockInt32<4, 1>& rhs) {
+    RegBlockInt32<4, 1> result;
+    result.buf.reg[0] = Mul(lhs.buf.reg[0], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 1x4 := 1x4 * 1x4
+template <>
+struct BroadcastMulImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<1, 4> Run(const RegBlockInt32<1, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<1, 4> result;
+    result.buf.reg[0] = Mul(lhs.buf.reg[0], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 1x4 := 1x4 * 1x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<1, 4> Run(const RegBlockInt32<1, 4>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<1, 4> result;
+    result.buf.reg[0] = Mul(lhs.buf.reg[0], rhs.buf.reg[0]);
+    return result;
+  }
+};
+
+// 4x4 := 4x4 * 1x4
+template <>
+struct BroadcastMulImpl<RegBlockInt32<4, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<4, 4> Run(const RegBlockInt32<4, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<4, 4> result;
+    const Int32x4 p = rhs.buf.reg[0];
+    result.buf.reg[0] = MulByRhsLane<0>(lhs.buf.reg[0], p);
+    result.buf.reg[1] = MulByRhsLane<1>(lhs.buf.reg[1], p);
+    result.buf.reg[2] = MulByRhsLane<2>(lhs.buf.reg[2], p);
+    result.buf.reg[3] = MulByRhsLane<3>(lhs.buf.reg[3], p);
+    return result;
+  }
+};
+
+// 4x4 := 4x4 * 4x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<4, 4>, RegBlockInt32<4, 1>> {
+  static RegBlockInt32<4, 4> Run(const RegBlockInt32<4, 4>& lhs,
+                                 const RegBlockInt32<4, 1>& rhs) {
+    RegBlockInt32<4, 4> result;
+    const Int32x4 p = rhs.buf.reg[0];
+    result.buf.reg[0] = Mul(lhs.buf.reg[0], p);
+    result.buf.reg[1] = Mul(lhs.buf.reg[1], p);
+    result.buf.reg[2] = Mul(lhs.buf.reg[2], p);
+    result.buf.reg[3] = Mul(lhs.buf.reg[3], p);
+    return result;
+  }
+};
+
+// 8x1 := 8x1 * 1x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<8, 1>, RegBlockInt32<1, 1>> {
+  static RegBlockInt32<8, 1> Run(const RegBlockInt32<8, 1>& lhs,
+                                 const RegBlockInt32<1, 1>& rhs) {
+    RegBlockInt32<8, 1> result;
+    const std::int32_t p = rhs.buf.reg[0];
+    for (int i = 0; i < 2; i++) {
+      result.buf.reg[i] = Mul(lhs.buf.reg[i], p);
+    }
+    return result;
+  }
+};
+
+// 8x1 := 8x1 * 8x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<8, 1>, RegBlockInt32<8, 1>> {
+  static RegBlockInt32<8, 1> Run(const RegBlockInt32<8, 1>& lhs,
+                                 const RegBlockInt32<8, 1>& rhs) {
+    RegBlockInt32<8, 1> result;
+    for (int i = 0; i < 2; i++) {
+      result.buf.reg[i] = Mul(lhs.buf.reg[i], rhs.buf.reg[i]);
+    }
+    return result;
+  }
+};
+
+// 8x4 := 8x4 * 1x4
+template <>
+struct BroadcastMulImpl<RegBlockInt32<8, 4>, RegBlockInt32<1, 4>> {
+  static RegBlockInt32<8, 4> Run(const RegBlockInt32<8, 4>& lhs,
+                                 const RegBlockInt32<1, 4>& rhs) {
+    RegBlockInt32<8, 4> result;
+    const Int32x4 p = rhs.buf.reg[0];
+    for (int i = 0; i < 2; i++) {
+      result.buf.reg[i + 0] = MulByRhsLane<0>(lhs.buf.reg[i + 0], p);
+      result.buf.reg[i + 2] = MulByRhsLane<1>(lhs.buf.reg[i + 2], p);
+      result.buf.reg[i + 4] = MulByRhsLane<2>(lhs.buf.reg[i + 4], p);
+      result.buf.reg[i + 6] = MulByRhsLane<3>(lhs.buf.reg[i + 6], p);
+    }
+    return result;
+  }
+};
+
+// 8x4 := 8x4 * 8x1
+template <>
+struct BroadcastMulImpl<RegBlockInt32<8, 4>, RegBlockInt32<8, 1>> {
+  static RegBlockInt32<8, 4> Run(const RegBlockInt32<8, 4>& lhs,
+                                 const RegBlockInt32<8, 1>& rhs) {
+    RegBlockInt32<8, 4> result;
+    const Int32x4 p[2]{rhs.buf.reg[0], rhs.buf.reg[1]};
+    for (int i = 0; i < 4; i++) {
+      for (int j = 0; j < 2; j++) {
+        const int k = j + 2 * i;
+        result.buf.reg[k] = Mul(lhs.buf.reg[k], p[j]);
+      }
+    }
+    return result;
+  }
+};
+
+// Rx1 += Rx1 * 1x1
+template <int Rows>
+struct BroadcastMulAddImpl<RegBlockInt32<Rows, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<Rows, 1>> {
+  static void Run(const RegBlockInt32<Rows, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<Rows, 1>* acc) {
+    const std::int32_t p = rhs.buf.reg[0];
+    for (int i = 0; i < RegBlockInt32<Rows, 1>::kRegisterCount; i++) {
+      MulAdd(lhs.buf.reg[i], p, &acc->buf.reg[i]);
+    }
+  }
+};
+
+// RxC += Rx1 * 1x1
+template <int Rows, int Cols>
+struct BroadcastMulAddImpl<RegBlockInt32<Rows, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<Rows, Cols>> {
+  static void Run(const RegBlockInt32<Rows, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs,
+                  RegBlockInt32<Rows, Cols>* acc) {
+    const std::int32_t p = rhs.buf.reg[0];
+    static constexpr int kRegsPerCol = RegBlockInt32<Rows, 1>::kRegisterCount;
+    for (int i = 0; i < kRegsPerCol; i++) {
+      const Int32x4 q = Mul(lhs.buf.reg[i], p);
+      for (int j = 0; j < Cols; j++) {
+        acc->buf.reg[i + j * kRegsPerCol] =
+            Add(acc->buf.reg[i + j * kRegsPerCol], q);
+      }
+    }
+  }
+};
+
+// 1xC += 1xC * 1x1
+template <int Cols>
+struct BroadcastMulAddImpl<RegBlockInt32<1, Cols>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<1, Cols>> {
+  static void Run(const RegBlockInt32<1, Cols>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<1, Cols>* acc) {
+    const std::int32_t p = rhs.buf.reg[0];
+    for (int i = 0; i < RegBlockInt32<1, Cols>::kRegisterCount; i++) {
+      MulAdd(lhs.buf.reg[i], p, &acc->buf.reg[i]);
+    }
+  }
+};
+
+// RxC += 1x1 * 1x1
+template <int Rows, int Cols>
+struct BroadcastMulAddImpl<RegBlockInt32<1, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<Rows, Cols>> {
+  static void Run(const RegBlockInt32<1, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs,
+                  RegBlockInt32<Rows, Cols>* acc) {
+    const Int32x4 p = Dup<Int32x4>(Mul(lhs.buf.reg[0], rhs.buf.reg[0]));
+    for (int i = 0; i < RegBlockInt32<Rows, Cols>::kRegisterCount; i++) {
+      acc->buf.reg[i] = Add(acc->buf.reg[i], p);
+    }
+  }
+};
+
+// 1x1 += 1x1 * 1x1
+template <>
+struct BroadcastMulAddImpl<RegBlockInt32<1, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<1, 1>> {
+  static void Run(const RegBlockInt32<1, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<1, 1>* acc) {
+    MulAdd(lhs.buf.reg[0], rhs.buf.reg[0], &acc->buf.reg[0]);
+  }
+};
+
+// Rx4 += Rx1 * 1x4
+template <int Rows>
+struct BroadcastMulAddImpl<RegBlockInt32<Rows, 1>, RegBlockInt32<1, 4>,
+                           RegBlockInt32<Rows, 4>> {
+  static void Run(const RegBlockInt32<Rows, 1>& lhs,
+                  const RegBlockInt32<1, 4>& rhs, RegBlockInt32<Rows, 4>* acc) {
+    const Int32x4 p = rhs.buf.reg[0];
+    static constexpr int kRegsPerCol = RegBlockInt32<Rows, 1>::kRegisterCount;
+    for (int i = 0; i < kRegsPerCol; i++) {
+      MulAddByRhsLane<0>(lhs.buf.reg[i], p, &acc->buf.reg[i + 0 * kRegsPerCol]);
+      MulAddByRhsLane<1>(lhs.buf.reg[i], p, &acc->buf.reg[i + 1 * kRegsPerCol]);
+      MulAddByRhsLane<2>(lhs.buf.reg[i], p, &acc->buf.reg[i + 2 * kRegsPerCol]);
+      MulAddByRhsLane<3>(lhs.buf.reg[i], p, &acc->buf.reg[i + 3 * kRegsPerCol]);
+    }
+  }
+};
+
+// Rx4 += 1x4 * 1x1
+template <int Rows>
+struct BroadcastMulAddImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<Rows, 4>> {
+  static void Run(const RegBlockInt32<1, 4>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<Rows, 4>* acc) {
+    const Int32x4 p = Mul(lhs.buf.reg[0], rhs.buf.reg[0]);
+    Int32x4 q[4];
+    q[0] = DupLane<0>(p);
+    q[1] = DupLane<1>(p);
+    q[2] = DupLane<2>(p);
+    q[3] = DupLane<3>(p);
+    static constexpr int kRegsPerCol = RegBlockInt32<Rows, 1>::kRegisterCount;
+    for (int i = 0; i < kRegsPerCol; i++) {
+      for (int j = 0; j < 4; j++) {
+        acc->buf.reg[i + j * kRegsPerCol] =
+            Add(q[j], acc->buf.reg[i + j * kRegsPerCol]);
+      }
+    }
+  }
+};
+
+// 1xC += 1x1 * 1x1
+template <int Cols>
+struct BroadcastMulAddImpl<RegBlockInt32<1, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<1, Cols>> {
+  static void Run(const RegBlockInt32<1, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<1, Cols>* acc) {
+    const Int32x4 p = Dup<Int32x4>(Mul(lhs.buf.reg[0], rhs.buf.reg[0]));
+    for (int i = 0; i < RegBlockInt32<1, Cols>::kRegisterCount; i++) {
+      acc->buf.reg[i] = Add(acc->buf.reg[i], p);
+    }
+  }
+};
+
+// 1x4 += 1x4 * 1x1
+template <>
+struct BroadcastMulAddImpl<RegBlockInt32<1, 4>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<1, 4>> {
+  static void Run(const RegBlockInt32<1, 4>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<1, 4>* acc) {
+    const std::int32_t p = rhs.buf.reg[0];
+    MulAdd(lhs.buf.reg[0], p, &acc->buf.reg[0]);
+  }
+};
+
+// 4xC += 4x1 * 1x1
+template <int Cols>
+struct BroadcastMulAddImpl<RegBlockInt32<4, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<4, Cols>> {
+  static void Run(const RegBlockInt32<4, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<4, Cols>* acc) {
+    const Int32x4 p = Mul(lhs.buf.reg[0], rhs.buf.reg[0]);
+    for (int i = 0; i < Cols; i++) {
+      acc->buf.reg[i] = Add(p, acc->buf.reg[i]);
+    }
+  }
+};
+
+// 4x1 += 4x1 * 1x1
+template <>
+struct BroadcastMulAddImpl<RegBlockInt32<4, 1>, RegBlockInt32<1, 1>,
+                           RegBlockInt32<4, 1>> {
+  static void Run(const RegBlockInt32<4, 1>& lhs,
+                  const RegBlockInt32<1, 1>& rhs, RegBlockInt32<4, 1>* acc) {
+    const std::int32_t p = rhs.buf.reg[0];
+    MulAdd(lhs.buf.reg[0], p, &acc->buf.reg[0]);
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_COMMON_NEON_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_neon.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_neon.h
new file mode 100644
index 0000000..c992b15
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_neon.h
@@ -0,0 +1,150 @@
+// Copyright 2017 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// simd_wrappers_neon.h: NEON specialization of simd_wrappers.h
+
+#ifndef GEMMLOWP_INTERNAL_SIMD_WRAPPERS_NEON_H_
+#define GEMMLOWP_INTERNAL_SIMD_WRAPPERS_NEON_H_
+
+#include <arm_neon.h>
+
+namespace gemmlowp {
+
+using Int32x4 = int32x4_t;
+using Uint8x8 = uint8x8_t;
+
+template <int ScalarCount>
+struct RegisterType<std::int32_t, ScalarCount> {
+  using Type =
+      typename std::conditional<ScalarCount >= 4, Int32x4, std::int32_t>::type;
+};
+
+template <int ScalarCount>
+struct RegisterType<std::uint8_t, ScalarCount> {
+  using Type = typename std::conditional<
+      ScalarCount >= 8, Uint8x8,
+      typename std::conditional<ScalarCount >= 4, std::uint32_t,
+                                std::uint8_t>::type>::type;
+};
+
+inline Int32x4 LoadInt32x4(const std::int32_t* src) { return vld1q_s32(src); }
+
+inline void StoreInt32x4(std::int32_t* dst, Int32x4 value) {
+  vst1q_s32(dst, value);
+}
+
+template <int Lane>
+std::int32_t GetLane(Int32x4 value) {
+  return vgetq_lane_s32(value, Lane);
+}
+
+template <int Lane>
+Int32x4 DupLane(Int32x4 value) {
+  switch (Lane) {
+    case 0:
+      return vdupq_lane_s32(vget_low_s32(value), 0);
+    case 1:
+      return vdupq_lane_s32(vget_low_s32(value), 1);
+    case 2:
+      return vdupq_lane_s32(vget_high_s32(value), 0);
+    case 3:
+      return vdupq_lane_s32(vget_high_s32(value), 1);
+    default:
+      static_assert(Lane >= 0 && Lane <= 3, "");
+      return vdupq_n_s32(0);
+  }
+}
+
+inline Int32x4 Mul(Int32x4 a, std::int32_t b) { return vmulq_n_s32(a, b); }
+
+inline Int32x4 Min(Int32x4 a, Int32x4 b) { return vminq_s32(a, b); }
+
+inline Int32x4 Max(Int32x4 a, Int32x4 b) { return vmaxq_s32(a, b); }
+
+inline Int32x4 SaturatingRoundingDoublingHighMul(Int32x4 a, std::int32_t b) {
+  return vqrdmulhq_n_s32(a, b);
+}
+
+template <int Lane>
+Int32x4 MulByRhsLane(Int32x4 a, Int32x4 b) {
+  switch (Lane) {
+    case 0:
+      return vmulq_lane_s32(a, vget_low_s32(b), 0);
+    case 1:
+      return vmulq_lane_s32(a, vget_low_s32(b), 1);
+    case 2:
+      return vmulq_lane_s32(a, vget_high_s32(b), 0);
+    case 3:
+      return vmulq_lane_s32(a, vget_high_s32(b), 1);
+    default:
+      static_assert(Lane >= 0 && Lane <= 3, "");
+      return vdupq_n_s32(0);
+  }
+}
+
+inline void MulAdd(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
+  *acc = vmlaq_s32(*acc, lhs, rhs);
+}
+
+inline void MulAdd(Int32x4 lhs, std::int32_t rhs, Int32x4* acc) {
+  *acc = vmlaq_n_s32(*acc, lhs, rhs);
+}
+
+template <int Lane>
+inline void MulAddByRhsLane(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
+  switch (Lane) {
+    case 0:
+      *acc = vmlaq_lane_s32(*acc, lhs, vget_low_s32(rhs), 0);
+      break;
+    case 1:
+      *acc = vmlaq_lane_s32(*acc, lhs, vget_low_s32(rhs), 1);
+      break;
+    case 2:
+      *acc = vmlaq_lane_s32(*acc, lhs, vget_high_s32(rhs), 0);
+      break;
+    case 3:
+      *acc = vmlaq_lane_s32(*acc, lhs, vget_high_s32(rhs), 1);
+      break;
+    default:
+      static_assert(Lane >= 0 && Lane <= 3, "");
+  }
+}
+
+template <>
+struct LoadContiguousImpl<RegBlockUint8<8, 8>> {
+  static RegBlockUint8<8, 8> Run(const std::uint8_t* src) {
+    RegBlockUint8<8, 8> result;
+    for (int i = 0; i < 8; i++) {
+      result.buf.reg[i] = vld1_u8(src + 8 * i);
+    }
+    return result;
+  }
+};
+
+template <>
+struct LoadContiguousImpl<RegBlockInt32<8, 8>> {
+  static RegBlockInt32<8, 8> Run(const std::int32_t* src) {
+    RegBlockInt32<8, 8> result;
+    for (int i = 0; i < 16; i++) {
+      result.buf.reg[i] = vld1q_s32(src + 4 * i);
+    }
+    return result;
+  }
+};
+
+}  // end namespace gemmlowp
+
+#include "simd_wrappers_common_neon_sse.h"
+
+#endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_NEON_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_sse.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_sse.h
new file mode 100644
index 0000000..6480b66
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/simd_wrappers_sse.h
@@ -0,0 +1,123 @@
+// Copyright 2017 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// simd_wrappers_neon.h: SSE SIMD wrappers
+
+#ifndef GEMMLOWP_INTERNAL_SIMD_WRAPPERS_SSE_H_
+#define GEMMLOWP_INTERNAL_SIMD_WRAPPERS_SSE_H_
+
+#include <smmintrin.h>
+
+namespace gemmlowp {
+
+using Int32x4 = __m128i;
+using Uint8x16 = __m128i;
+
+template <int ScalarCount>
+struct RegisterType<std::int32_t, ScalarCount> {
+  using Type =
+      typename std::conditional<ScalarCount >= 4, Int32x4, std::int32_t>::type;
+};
+
+template <int ScalarCount>
+struct RegisterType<std::uint8_t, ScalarCount> {
+  using Type = typename std::conditional<
+      ScalarCount >= 16, Uint8x16,
+      typename std::conditional<ScalarCount >= 4, std::uint32_t,
+                                std::uint8_t>::type>::type;
+};
+
+inline Int32x4 LoadInt32x4(const std::int32_t* src) {
+  return _mm_loadu_si128(reinterpret_cast<const Int32x4*>(src));
+}
+
+inline void StoreInt32x4(std::int32_t* dst, Int32x4 value) {
+  _mm_storeu_si128(reinterpret_cast<__m128i*>(dst), value);
+}
+
+inline Uint8x16 LoadUint8x16(const std::uint8_t* src) {
+  return _mm_loadu_si128(reinterpret_cast<const Uint8x16*>(src));
+}
+
+inline void StoreUint8x16(std::uint8_t* dst, Uint8x16 value) {
+  _mm_storeu_si128(reinterpret_cast<__m128i*>(dst), value);
+}
+
+template <int Lane>
+std::int32_t GetLane(Int32x4 value) {
+  return _mm_extract_epi32(value, Lane);
+}
+
+template <int Lane>
+Int32x4 DupLane(Int32x4 value) {
+  return _mm_shuffle_epi32(value, _MM_SHUFFLE(Lane, Lane, Lane, Lane));
+}
+
+inline Int32x4 Mul(Int32x4 a, std::int32_t b) {
+  return Mul(a, Dup<Int32x4>(b));
+}
+
+inline Int32x4 Min(Int32x4 a, Int32x4 b) { return _mm_min_epi32(a, b); }
+
+inline Int32x4 Max(Int32x4 a, Int32x4 b) { return _mm_max_epi32(a, b); }
+
+inline Int32x4 SaturatingRoundingDoublingHighMul(Int32x4 a, std::int32_t b) {
+  return SaturatingRoundingDoublingHighMul(a, Dup<Int32x4>(b));
+}
+
+template <int Lane>
+Int32x4 MulByRhsLane(Int32x4 a, Int32x4 b) {
+  return Mul(a, DupLane<Lane>(b));
+}
+
+inline void MulAdd(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
+  *acc = Add(*acc, Mul(lhs, rhs));
+}
+
+inline void MulAdd(Int32x4 lhs, std::int32_t rhs, Int32x4* acc) {
+  *acc = Add(*acc, Mul(lhs, rhs));
+}
+
+template <int Lane>
+inline void MulAddByRhsLane(Int32x4 lhs, Int32x4 rhs, Int32x4* acc) {
+  *acc = Add(*acc, MulByRhsLane<Lane>(lhs, rhs));
+}
+
+template <>
+struct LoadContiguousImpl<RegBlockUint8<8, 8>> {
+  static RegBlockUint8<8, 8> Run(const std::uint8_t* src) {
+    RegBlockUint8<8, 8> result;
+    for (int i = 0; i < 4; i++) {
+      result.buf.reg[i] = LoadUint8x16(src + 16 * i);
+    }
+    return result;
+  }
+};
+
+template <>
+struct LoadContiguousImpl<RegBlockInt32<8, 8>> {
+  static RegBlockInt32<8, 8> Run(const std::int32_t* src) {
+    RegBlockInt32<8, 8> result;
+    for (int i = 0; i < 16; i++) {
+      result.buf.reg[i] = LoadInt32x4(src + 4 * i);
+    }
+    return result;
+  }
+};
+
+}  // end namespace gemmlowp
+
+#include "simd_wrappers_common_neon_sse.h"
+
+#endif  // GEMMLOWP_INTERNAL_SIMD_WRAPPERS_SSE_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/single_thread_gemm.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/single_thread_gemm.h
new file mode 100644
index 0000000..3d430c5
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/single_thread_gemm.h
@@ -0,0 +1,158 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// single_thread_gemm.h: Single-threaded GEMM implementation.
+// This is a good place to start reading code, as it shows the overall
+// structure of a GEMM and is much simpler than multi_thread_gemm.h.
+
+#ifndef GEMMLOWP_INTERNAL_SINGLE_THREAD_GEMM_H_
+#define GEMMLOWP_INTERNAL_SINGLE_THREAD_GEMM_H_
+
+#include <cassert>
+
+#include "../public/map.h"
+#include "allocator.h"
+#include "compute.h"
+#include "kernel.h"
+#include "pack.h"
+#include "unpack.h"
+
+#ifdef GEMMLOWP_PROFILING_SIZES
+#ifndef GEMMLOWP_PROFILING
+#error GEMMLOWP_PROFILING_SIZES without GEMMLOWP_PROFILING
+#endif
+#include <string>
+#include <unordered_map>
+#endif
+
+namespace gemmlowp {
+
+class SingleThreadGemmContext {
+ public:
+  Allocator* allocator() { return &allocator_; }
+
+  void set_l1_bytes_to_use(int n) { l1_bytes_to_use_ = n; }
+  void set_l2_bytes_to_use(int n) { l2_bytes_to_use_ = n; }
+  void set_l2_rhs_factor(float n) { l2_rhs_factor_ = n; }
+
+  int l1_bytes_to_use() const { return l1_bytes_to_use_; }
+  int l2_bytes_to_use() const { return l2_bytes_to_use_; }
+  float l2_rhs_factor() const { return l2_rhs_factor_; }
+
+ protected:
+  Allocator allocator_;
+
+  // The cache configurationt to use.
+  int l1_bytes_to_use_ = kDefaultL1CacheSize;
+  int l2_bytes_to_use_ = kDefaultL2CacheSize;
+  float l2_rhs_factor_ = kDefaultL2RhsFactor;
+};
+
+template <typename KernelFormat, typename InputScalar, typename OutputScalar,
+          typename BitDepthParams, MapOrder LhsOrder, MapOrder RhsOrder,
+          MapOrder ResultOrder, typename LhsOffset, typename RhsOffset,
+          typename OutputPipelineType>
+void SingleThreadGemm(SingleThreadGemmContext* context,
+                      const KernelBase& kernel,
+                      const MatrixMap<const InputScalar, LhsOrder>& lhs,
+                      const MatrixMap<const InputScalar, RhsOrder>& rhs,
+                      MatrixMap<OutputScalar, ResultOrder>* result,
+                      const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
+                      const OutputPipelineType& output_pipeline) {
+  ScopedProfilingLabel label("gemmlowp::SingleThreadGemm");
+
+  assert(lhs.cols() == rhs.rows());
+
+  int rows = result->rows();
+  int cols = result->cols();
+  int depth = lhs.cols();
+
+  // zero sizes should have been caught earlier and early-returned.
+  assert(rows > 0);
+  assert(cols > 0);
+  assert(depth > 0);
+
+  // The case of rows<cols should have been caught earlier and transposed.
+  assert(rows >= cols);
+
+  Allocator* allocator = context->allocator();
+
+  BlockParams block_params;
+  block_params.Init<KernelFormat>(rows, cols, depth, 1,
+                                  context->l1_bytes_to_use(),
+                                  context->l2_bytes_to_use(),
+                                  context->l2_rhs_factor());
+
+#ifdef GEMMLOWP_PROFILING_SIZES
+  // Using a static map of label strings. Not reentrant at all!
+  static std::unordered_map<std::uint64_t, std::string> labels_map;
+  std::uint64_t sizes_hash = static_cast<std::uint64_t>(rows) ^
+                             (static_cast<std::uint64_t>(depth) << 16) ^
+                             (static_cast<std::uint64_t>(cols) << 32);
+  if (!labels_map.count(sizes_hash)) {
+    char label[256];
+    snprintf(label, sizeof(label),
+             "(rows = %d, depth = %d, cols = %d, l2_rows = %d, l2_depth = %d, "
+             "l2_cols = %d, l1_rows = %d, l1_depth = %d, l1_cols = %d)",
+             rows, depth, cols, block_params.l2_rows, block_params.l2_depth,
+             block_params.l2_cols, block_params.l1_rows, block_params.l1_depth,
+             block_params.l1_cols);
+    labels_map[sizes_hash] = label;
+  }
+  ScopedProfilingLabel size_label(labels_map[sizes_hash].c_str());
+#endif
+
+  PackedSideBlock<typename KernelFormat::Lhs> packed_lhs(Side::Lhs, allocator,
+                                                         block_params);
+  PackedSideBlock<typename KernelFormat::Rhs> packed_rhs(Side::Rhs, allocator,
+                                                         block_params);
+
+  PackedResult packed_result(allocator, block_params);
+
+  allocator->Commit();
+
+  const bool pack_rhs_once = block_params.l2_cols >= cols;
+
+  if (pack_rhs_once) {
+    PackRhs(&packed_rhs, rhs);
+  }
+
+  for (int r = 0; r < rows; r += block_params.l2_rows) {
+    int rs = std::min(block_params.l2_rows, rows - r);
+
+    PackLhs(&packed_lhs, lhs.block(r, 0, rs, depth));
+
+    for (int c = 0; c < cols; c += block_params.l2_cols) {
+      int cs = std::min(block_params.l2_cols, cols - c);
+
+      if (!pack_rhs_once) {
+        PackRhs(&packed_rhs, rhs.block(0, c, depth, cs));
+      }
+
+      Compute(kernel, block_params, &packed_result, packed_lhs, packed_rhs,
+              depth);
+
+      UnpackResult<KernelFormat>(
+          result, MatrixBlockBounds(r, c, rs, cs), packed_result, depth,
+          packed_lhs.sums_of_each_slice(), packed_rhs.sums_of_each_slice(),
+          lhs_offset.block(r, rs), rhs_offset.block(c, cs), output_pipeline);
+    }
+  }
+
+  allocator->Decommit();
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_SINGLE_THREAD_GEMM_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/internal/unpack.h b/src/runtime/ref/nn/depend/external/gemmlowp/internal/unpack.h
new file mode 100644
index 0000000..33aee13
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/internal/unpack.h
@@ -0,0 +1,278 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// unpack.h: unpacking the result blocks computed by compute.h,
+// storing them into the destination matrix.
+
+#ifndef GEMMLOWP_INTERNAL_UNPACK_H_
+#define GEMMLOWP_INTERNAL_UNPACK_H_
+
+#include "allocator.h"
+#include "block_params.h"
+#include "output.h"
+#include "pack.h"
+
+#include <cmath>
+
+namespace gemmlowp {
+
+class PackedResult {
+ public:
+  PackedResult(Allocator* _allocator, const BlockParams& _block_params)
+      : allocator_(_allocator), block_params_(_block_params) {
+    matrix_handle_ = allocator_->Reserve<std::int32_t>(block_params_.l2_rows *
+                                                       block_params_.l2_cols);
+  }
+
+  ~PackedResult() {}
+
+  MatrixMap<std::int32_t, MapOrder::ColMajor> Map() {
+    return MatrixMap<std::int32_t, MapOrder::ColMajor>(
+        allocator_->GetPointer<std::int32_t>(matrix_handle_),
+        block_params_.l2_rows, block_params_.l2_cols, block_params_.l2_rows);
+  }
+
+  MatrixMap<const std::int32_t, MapOrder::ColMajor> Map() const {
+    return MatrixMap<const std::int32_t, MapOrder::ColMajor>(
+        allocator_->GetPointer<const std::int32_t>(matrix_handle_),
+        block_params_.l2_rows, block_params_.l2_cols, block_params_.l2_rows);
+  }
+
+ private:
+  Allocator* allocator_;
+  Allocator::Handle matrix_handle_;
+  const BlockParams& block_params_;
+};
+
+struct MatrixBlockBounds {
+  int start_row;
+  int start_col;
+  int rows;
+  int cols;
+
+  MatrixBlockBounds(int start_row_, int start_col_, int rows_, int cols_)
+      : start_row(start_row_),
+        start_col(start_col_),
+        rows(rows_),
+        cols(cols_) {}
+};
+
+template <int Rows, int Cols, typename SrcMapType>
+void PrefetchResultBlock(const SrcMapType& src,
+                         const VectorMap<const std::int32_t, VectorShape::Col>&
+                             lhs_sums_of_each_slice,
+                         int src_row, int src_col) {
+  const std::int32_t* src_data = src.data(src_row, src_col);
+  const int src_stride = src.stride();
+  const std::int32_t* lhs_sums_data = lhs_sums_of_each_slice.data(src_row);
+  for (int r = 0; r < Rows; r += 4) {
+    Prefetch(lhs_sums_data + r);
+  }
+  for (int c = 0; c < Cols; c++) {
+    for (int r = 0; r < Rows; r += 4) {
+      Prefetch(src_data + r + c * src_stride);
+    }
+  }
+}
+
+template <typename KernelFormat, typename RegisterBlockType,
+          typename SrcMapType, typename LhsOffset, typename RhsOffset,
+          typename OutputPipelineExecutorType, typename DstType>
+void UnpackResultBlock(const SrcMapType& src,
+                       const OutputPipelineExecutorType& executor, DstType* dst,
+                       const VectorMap<const std::int32_t, VectorShape::Col>&
+                           lhs_sums_of_each_slice,
+                       const VectorMap<const std::int32_t, VectorShape::Row>&
+                           rhs_sums_of_each_slice,
+                       const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
+                       int depth, int src_row, int src_col, int src_global_row,
+                       int src_global_col, int dst_row, int dst_col) {
+  using KernelLhsScalar = typename KernelFormat::Lhs::Scalar;
+  using KernelRhsScalar = typename KernelFormat::Rhs::Scalar;
+  static constexpr int KernelLhsZeroPointInput =
+      ZeroPointInputValue<KernelLhsScalar>::kValue;
+  static constexpr int KernelRhsZeroPointInput =
+      ZeroPointInputValue<KernelRhsScalar>::kValue;
+  auto acc = Load<RegisterBlockType>(src, src_row, src_col);
+  const auto& lhs_sums_of_each_slice_block =
+      LoadForBroadcasting<RegisterBlockType>(lhs_sums_of_each_slice, src_row);
+  const auto& rhs_sums_of_each_slice_block =
+      LoadForBroadcasting<RegisterBlockType>(rhs_sums_of_each_slice, src_col);
+  auto lhs_offset_block =
+      LoadForBroadcasting<RegisterBlockType>(lhs_offset, src_row);
+  auto rhs_offset_block =
+      LoadForBroadcasting<RegisterBlockType>(rhs_offset, src_col);
+  AddConstant<KernelLhsZeroPointInput>(&lhs_offset_block);
+  AddConstant<KernelRhsZeroPointInput>(&rhs_offset_block);
+  BroadcastMulAdd(lhs_sums_of_each_slice_block, rhs_offset_block, &acc);
+  for (int i = 0; i < decltype(rhs_offset_block)::kRegisterCount; i++) {
+    rhs_offset_block.buf.reg[i] = Mul(rhs_offset_block.buf.reg[i], depth);
+  }
+  BroadcastMulAdd(BroadcastAdd(rhs_sums_of_each_slice_block, rhs_offset_block),
+                  lhs_offset_block, &acc);
+  executor.Execute(acc, dst, src_global_row, src_global_col, dst_row, dst_col);
+}
+
+template <typename KernelFormat, typename ResultBlockType,
+          typename PackedResultType, typename LhsOffset, typename RhsOffset,
+          typename OutputPipelineType>
+void UnpackResult(ResultBlockType* dst, const MatrixBlockBounds& dst_block,
+                  const PackedResultType& src, int depth,
+                  const std::int32_t* lhs_sums_of_each_slice_ptr,
+                  const std::int32_t* rhs_sums_of_each_slice_ptr,
+                  const LhsOffset& lhs_offset, const RhsOffset& rhs_offset,
+                  const OutputPipelineType& output_pipeline) {
+  ScopedProfilingLabel label(ResultBlockType::kOrder == MapOrder::ColMajor
+                                 ? "unpack to column-major"
+                                 : "unpack to row-major");
+  assert(dst_block.start_row >= 0);
+  assert(dst_block.start_row + dst_block.rows <= dst->rows());
+  assert(dst_block.start_col >= 0);
+  assert(dst_block.start_col + dst_block.cols <= dst->cols());
+  const auto src_map = src.Map();
+  const VectorMap<const std::int32_t, VectorShape::Col> lhs_sums_of_each_slice(
+      lhs_sums_of_each_slice_ptr, dst_block.rows);
+  const VectorMap<const std::int32_t, VectorShape::Row> rhs_sums_of_each_slice(
+      rhs_sums_of_each_slice_ptr, dst_block.cols);
+  using Int32x1x1 = RegisterBlock<std::int32_t, 1, 1>;
+  using Int32x4x1 = RegisterBlock<std::int32_t, 4, 1>;
+  using Int32x8x1 = RegisterBlock<std::int32_t, 8, 1>;
+  using Int32x1x4 = RegisterBlock<std::int32_t, 1, 4>;
+  using Int32x4x4 = RegisterBlock<std::int32_t, 4, 4>;
+  using Int32x8x4 = RegisterBlock<std::int32_t, 8, 4>;
+
+  using DstScalarType = typename ResultBlockType::Scalar;
+  using DstScalarx8x8 = RegisterBlock<DstScalarType, 8, 8>;
+
+  OutputPipelineExecutor<OutputPipelineType, Int32x1x1>
+      output_pipeline_executor_1x1(output_pipeline);
+  OutputPipelineExecutor<OutputPipelineType, Int32x4x1>
+      output_pipeline_executor_4x1(output_pipeline);
+  OutputPipelineExecutor<OutputPipelineType, Int32x8x1>
+      output_pipeline_executor_8x1(output_pipeline);
+  OutputPipelineExecutor<OutputPipelineType, Int32x1x4>
+      output_pipeline_executor_1x4(output_pipeline);
+  OutputPipelineExecutor<OutputPipelineType, Int32x4x4>
+      output_pipeline_executor_4x4(output_pipeline);
+  OutputPipelineExecutor<OutputPipelineType, Int32x8x4>
+      output_pipeline_executor_8x4(output_pipeline);
+
+  int c8 = 0;
+  if (ResultBlockType::kOrder == MapOrder::RowMajor) {
+    for (; c8 <= dst_block.cols - 8; c8 += 8) {
+      PrefetchResultBlock<8, 8>(src_map, lhs_sums_of_each_slice, 0, c8);
+      int r = 0;
+      for (; r <= dst_block.rows - 8; r += 8) {
+        const int global_row = r + dst_block.start_row;
+        PrefetchResultBlock<8, 8>(src_map, lhs_sums_of_each_slice, r + 8, c8);
+        DstScalarType dst_colmajor_buf[64];
+        MatrixMap<DstScalarType, MapOrder::ColMajor> dst_colmajor_map(
+            dst_colmajor_buf, 8, 8);
+        for (int cx = 0; cx < 8; cx += 4) {
+          const int c = c8 + cx;
+          const int global_col = c + dst_block.start_col;
+          UnpackResultBlock<KernelFormat, Int32x8x4>(
+              src_map, output_pipeline_executor_8x4, &dst_colmajor_map,
+              lhs_sums_of_each_slice, rhs_sums_of_each_slice, lhs_offset,
+              rhs_offset, depth, r, c, global_row, global_col, 0, cx);
+        }
+        StoreFinalOutput(LoadContiguous<DstScalarx8x8>(dst_colmajor_buf), dst,
+                         r + dst_block.start_row, c8 + dst_block.start_col);
+      }
+      for (; r <= dst_block.rows - 4; r += 4) {
+        const int global_row = r + dst_block.start_row;
+        for (int cx = 0; cx < 8; cx += 4) {
+          const int c = c8 + cx;
+          const int global_col = c + dst_block.start_col;
+          UnpackResultBlock<KernelFormat, Int32x4x4>(
+              src_map, output_pipeline_executor_4x4, dst,
+              lhs_sums_of_each_slice, rhs_sums_of_each_slice, lhs_offset,
+              rhs_offset, depth, r, c, global_row, global_col, global_row,
+              global_col);
+        }
+      }
+      for (; r < dst_block.rows; r++) {
+        const int global_row = r + dst_block.start_row;
+        for (int cx = 0; cx < 8; cx += 4) {
+          const int c = c8 + cx;
+          const int global_col = c + dst_block.start_col;
+          UnpackResultBlock<KernelFormat, Int32x1x4>(
+              src_map, output_pipeline_executor_1x4, dst,
+              lhs_sums_of_each_slice, rhs_sums_of_each_slice, lhs_offset,
+              rhs_offset, depth, r, c, global_row, global_col, global_row,
+              global_col);
+        }
+      }
+    }
+  }
+  int c = c8;
+  for (; c <= dst_block.cols - 4; c += 4) {
+    const int global_col = c + dst_block.start_col;
+    PrefetchResultBlock<8, 4>(src_map, lhs_sums_of_each_slice, 0, c);
+    int r = 0;
+    for (; r <= dst_block.rows - 8; r += 8) {
+      const int global_row = r + dst_block.start_row;
+      PrefetchResultBlock<8, 4>(src_map, lhs_sums_of_each_slice, r + 8, c);
+      UnpackResultBlock<KernelFormat, Int32x8x4>(
+          src_map, output_pipeline_executor_8x4, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+    for (; r <= dst_block.rows - 4; r += 4) {
+      const int global_row = r + dst_block.start_row;
+      UnpackResultBlock<KernelFormat, Int32x4x4>(
+          src_map, output_pipeline_executor_4x4, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+    for (; r < dst_block.rows; r++) {
+      const int global_row = r + dst_block.start_row;
+      UnpackResultBlock<KernelFormat, Int32x1x4>(
+          src_map, output_pipeline_executor_1x4, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+  }
+  for (; c < dst_block.cols; c++) {
+    const int global_col = c + dst_block.start_col;
+    PrefetchResultBlock<8, 1>(src_map, lhs_sums_of_each_slice, 0, c);
+    int r = 0;
+    for (; r <= dst_block.rows - 8; r += 8) {
+      const int global_row = r + dst_block.start_row;
+      PrefetchResultBlock<8, 1>(src_map, lhs_sums_of_each_slice, r + 8, c);
+      UnpackResultBlock<KernelFormat, Int32x8x1>(
+          src_map, output_pipeline_executor_8x1, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+    for (; r <= dst_block.rows - 4; r += 4) {
+      const int global_row = r + dst_block.start_row;
+      UnpackResultBlock<KernelFormat, Int32x4x1>(
+          src_map, output_pipeline_executor_4x1, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+    for (; r < dst_block.rows; r++) {
+      const int global_row = r + dst_block.start_row;
+      UnpackResultBlock<KernelFormat, Int32x1x1>(
+          src_map, output_pipeline_executor_1x1, dst, lhs_sums_of_each_slice,
+          rhs_sums_of_each_slice, lhs_offset, rhs_offset, depth, r, c,
+          global_row, global_col, global_row, global_col);
+    }
+  }
+}
+
+}  // end namespace gemmlowp
+
+#endif  // GEMMLOWP_INTERNAL_UNPACK_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/profiling/instrumentation.h b/src/runtime/ref/nn/depend/external/gemmlowp/profiling/instrumentation.h
new file mode 100644
index 0000000..51b6525
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/profiling/instrumentation.h
@@ -0,0 +1,244 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// instrumentation.h: contains the definitions needed to
+// instrument code for profiling:
+//   ScopedProfilingLabel, RegisterCurrentThreadForProfiling.
+//
+// profiler.h is only needed to drive the profiler:
+//   StartProfiling, FinishProfiling.
+//
+// See the usage example in profiler.h.
+
+#ifndef GEMMLOWP_PROFILING_INSTRUMENTATION_H_
+#define GEMMLOWP_PROFILING_INSTRUMENTATION_H_
+
+#include <pthread.h>
+#include <cstdio>
+
+#ifndef GEMMLOWP_USE_STLPORT
+#include <cstdint>
+#else
+#include <stdint.h>
+namespace std {
+using ::uint8_t;
+using ::uint16_t;
+using ::uint32_t;
+using ::int8_t;
+using ::int16_t;
+using ::int32_t;
+using ::size_t;
+using ::uintptr_t;
+}
+#endif
+
+#include <algorithm>
+#include <cassert>
+#include <cstdlib>
+
+#ifdef GEMMLOWP_PROFILING
+#include <cstring>
+#include <set>
+#endif
+
+// We should always use C++11 thread_local; unfortunately that
+// isn't fully supported on Apple yet.
+#ifdef __APPLE__
+#define GEMMLOWP_THREAD_LOCAL static __thread
+#define GEMMLOWP_USING_OLD_THREAD_LOCAL
+#else
+#define GEMMLOWP_THREAD_LOCAL thread_local
+#endif
+
+namespace gemmlowp {
+
+inline void ReleaseBuildAssertion(bool condition, const char* msg) {
+  if (!condition) {
+    fprintf(stderr, "gemmlowp error: %s\n", msg);
+    abort();
+  }
+}
+
+// To be used as template parameter for GlobalLock.
+// GlobalLock<ProfilerLockId> is the profiler global lock:
+// registering threads, starting profiling, finishing profiling, and
+// the profiler itself as it samples threads, all need to lock it.
+struct ProfilerLockId;
+
+// A very plain global lock. Templated in LockId so we can have multiple
+// locks, one for each LockId type.
+template <typename LockId>
+class GlobalLock {
+  static pthread_mutex_t* Mutex() {
+    static pthread_mutex_t m = PTHREAD_MUTEX_INITIALIZER;
+    return &m;
+  }
+
+ public:
+  static void Lock() { pthread_mutex_lock(Mutex()); }
+  static void Unlock() { pthread_mutex_unlock(Mutex()); }
+};
+
+// A very simple RAII helper to lock and unlock a GlobalLock
+template <typename LockId>
+struct AutoGlobalLock {
+  AutoGlobalLock() { GlobalLock<LockId>::Lock(); }
+  ~AutoGlobalLock() { GlobalLock<LockId>::Unlock(); }
+};
+
+// MemoryBarrier is purely a compile-time thing; it tells two things
+// to the compiler:
+//   1) It prevents reordering code across it
+//     (thanks to the 'volatile' after 'asm')
+//   2) It requires the compiler to assume that any value previously
+//     read from memory, may have changed. Thus it offers an alternative
+//     to using 'volatile' variables.
+inline void MemoryBarrier() { asm volatile("" ::: "memory"); }
+
+// Profiling definitions. Two paths: when profiling is enabled,
+// and when profiling is disabled.
+#ifdef GEMMLOWP_PROFILING
+// This code path is when profiling is enabled.
+
+// A pseudo-call-stack. Contrary to a real call-stack, this only
+// contains pointers to literal strings that were manually entered
+// in the instrumented code (see ScopedProfilingLabel).
+struct ProfilingStack {
+  static const std::size_t kMaxSize = 15;
+  typedef const char* LabelsArrayType[kMaxSize];
+  LabelsArrayType labels;
+  std::size_t size;
+
+  ProfilingStack() { memset(this, 0, sizeof(ProfilingStack)); }
+
+  void Push(const char* label) {
+    MemoryBarrier();
+    ReleaseBuildAssertion(size < kMaxSize, "ProfilingStack overflow");
+    labels[size] = label;
+    MemoryBarrier();
+    size++;
+    MemoryBarrier();
+  }
+
+  void Pop() {
+    MemoryBarrier();
+    ReleaseBuildAssertion(size > 0, "ProfilingStack underflow");
+    size--;
+    MemoryBarrier();
+  }
+
+  void UpdateTop(const char* new_label) {
+    MemoryBarrier();
+    assert(size);
+    labels[size - 1] = new_label;
+    MemoryBarrier();
+  }
+
+  ProfilingStack& operator=(const ProfilingStack& other) {
+    memcpy(this, &other, sizeof(ProfilingStack));
+    return *this;
+  }
+
+  bool operator==(const ProfilingStack& other) const {
+    return !memcmp(this, &other, sizeof(ProfilingStack));
+  }
+};
+
+static_assert(
+    !(sizeof(ProfilingStack) & (sizeof(ProfilingStack) - 1)),
+    "ProfilingStack should have power-of-two size to fit in cache lines");
+
+struct ThreadInfo;
+
+// The global set of threads being profiled.
+inline std::set<ThreadInfo*>& ThreadsUnderProfiling() {
+  static std::set<ThreadInfo*> v;
+  return v;
+}
+
+struct ThreadInfo {
+  pthread_key_t key;  // used only to get a callback at thread exit.
+  ProfilingStack stack;
+
+  ThreadInfo() {
+    pthread_key_create(&key, ThreadExitCallback);
+    pthread_setspecific(key, this);
+  }
+
+  static void ThreadExitCallback(void* ptr) {
+    AutoGlobalLock<ProfilerLockId> lock;
+    ThreadInfo* self = static_cast<ThreadInfo*>(ptr);
+    ThreadsUnderProfiling().erase(self);
+    pthread_key_delete(self->key);
+  }
+};
+
+inline ThreadInfo& ThreadLocalThreadInfo() {
+#ifdef GEMMLOWP_USING_OLD_THREAD_LOCAL
+  // We're leaking this ThreadInfo structure, because Apple doesn't support
+  // non-trivial constructors or destructors for their __thread type modifier.
+  GEMMLOWP_THREAD_LOCAL ThreadInfo* i = nullptr;
+  if (i == nullptr) {
+    i = new ThreadInfo();
+  }
+  return *i;
+#else
+  GEMMLOWP_THREAD_LOCAL ThreadInfo i;
+  return i;
+#endif
+}
+
+// ScopedProfilingLabel is how one instruments code for profiling
+// with this profiler. Construct local ScopedProfilingLabel variables,
+// passing a literal string describing the local code. Profile
+// samples will then be annotated with this label, while it is in scope
+// (whence the name --- also known as RAII).
+// See the example in profiler.h.
+class ScopedProfilingLabel {
+  ProfilingStack* profiling_stack_;
+
+ public:
+  explicit ScopedProfilingLabel(const char* label)
+      : profiling_stack_(&ThreadLocalThreadInfo().stack) {
+    profiling_stack_->Push(label);
+  }
+
+  ~ScopedProfilingLabel() { profiling_stack_->Pop(); }
+
+  void Update(const char* new_label) { profiling_stack_->UpdateTop(new_label); }
+};
+
+// To be called once on each thread to be profiled.
+inline void RegisterCurrentThreadForProfiling() {
+  AutoGlobalLock<ProfilerLockId> lock;
+  ThreadsUnderProfiling().insert(&ThreadLocalThreadInfo());
+}
+
+#else  // not GEMMLOWP_PROFILING
+// This code path is when profiling is disabled.
+
+// This empty definition of ScopedProfilingLabel ensures that
+// it has zero runtime overhead when profiling is disabled.
+struct ScopedProfilingLabel {
+  explicit ScopedProfilingLabel(const char*) {}
+  void Update(const char*) {}
+};
+
+inline void RegisterCurrentThreadForProfiling() {}
+
+#endif
+
+}  // end namespace gemmlowp
+
+#endif  // GEMMLOWP_PROFILING_INSTRUMENTATION_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/profiling/profiler.h b/src/runtime/ref/nn/depend/external/gemmlowp/profiling/profiler.h
new file mode 100644
index 0000000..a18c036
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/profiling/profiler.h
@@ -0,0 +1,373 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// profiler.h: a simple sampling profiler that's always just one #include away!
+//
+// Overview
+// ========
+//
+// This profiler only samples a pseudo-stack, not the actual call stack.
+// The code to be profiled needs to be instrumented with
+// pseudo-stack "labels", see ScopedProfilingLabel.
+// Using pseudo-stacks allows this profiler to be very simple, low-overhead,
+// portable, and independent of compilation details such as function inlining
+// and frame pointers. The granularity of instrumentation can be freely chosen,
+// and it is possible to get some annotate-like detail, i.e. detail within one
+// function without splitting it into multiple functions.
+//
+// This profiler should remain small and simple; its key feature is to fit in
+// a single header file so that there should never be a reason to refrain
+// from profiling. More complex and feature-rich alternatives are
+// readily available. This one offers a strict superset of its
+// functionality: https://github.com/bgirard/GeckoProfiler, including
+// intertwining pseudostacks with real call stacks, more annotation options,
+// and advanced visualization.
+//
+// Usage
+// =====
+//
+// 0. Enable profiling by defining GEMMLOWP_PROFILING. When profiling is
+//    not enabled, profiling instrumentation from instrumentation.h
+//    (ScopedProfilingLabel, RegisterCurrentThreadForProfiling)
+//    is still defined but does nothing. On the other hand,
+//    when profiling is not enabled, it is an error to #include the
+//    present file.
+//
+// 1. Each thread can opt in to profiling by calling
+//    RegisterCurrentThreadForProfiling() defined in instrumentation.h.
+//    This can be done at any time, before or during profiling.
+//    No sample will be collected from a thread until
+//    it has called RegisterCurrentThreadForProfiling().
+//
+// 2. Instrument your code to be profiled with ScopedProfilingLabel,
+//    which is a RAII helper defined in instrumentation.h. The identifier
+//    names (some_label, etc) do not matter; what will show up
+//    in the profile is the string passed to the constructor, which
+//    must be a literal string. See the full example below.
+//
+//    Note: the overhead of ScopedProfilingLabel is zero when not
+//    enabling profiling (when not defining GEMMLOWP_PROFILING).
+//
+// 3. Use the profiler.h interface to control profiling. There are two
+//    functions: StartProfiling() and FinishProfiling(). They must be
+//    called on the same thread. FinishProfiling() prints the profile
+//    on stdout.
+//
+// Full example
+// ============
+/*
+    #define GEMMLOWP_PROFILING
+    #include "profiling/instrumentation.h"
+    using namespace gemmlowp;
+
+    const int iters = 100000000;
+    volatile int i;
+
+    void Bar() {
+      ScopedProfilingLabel label("Bar");
+      for (i = 0; i < iters; i++) {}
+    }
+
+    void Foo() {
+      ScopedProfilingLabel label("Foo");
+      for (i = 0; i < iters; i++) {}
+      Bar();
+    }
+
+    void Init() {
+      RegisterCurrentThreadForProfiling();
+    }
+
+    #include "profiling/profiler.h"
+
+    int main() {
+      Init();
+      StartProfiling();
+      Foo();
+      FinishProfiling();
+    }
+*
+* Output:
+*
+    gemmlowp profile (1 threads, 304 samples)
+    100.00% Foo
+        51.32% other
+        48.68% Bar
+    0.00% other (outside of any label)
+*/
+//
+// Interpreting results
+// ====================
+//
+//  Each node shows the absolute percentage, among all the samples,
+//  of the number of samples that recorded the given pseudo-stack.
+//  The percentages are *NOT* relative to the parent node. In addition
+//  to your own labels, you will also see 'other' nodes that collect
+//  the remainder of samples under the parent node that didn't fall into
+//  any of the labelled child nodes. Example:
+//
+//  20% Foo
+//      12% Bar
+//      6% Xyz
+//      2% other
+//
+//  This means that 20% of all labels were under Foo, of which 12%/20%==60%
+//  were under Bar, 6%/20%==30% were under Xyz, and 2%/20%==10% were not
+//  under either Bar or Xyz.
+//
+//  Typically, one wants to keep adding ScopedProfilingLabel's until
+//  the 'other' nodes show low percentages.
+//
+// Interpreting results with multiple threads
+// ==========================================
+//
+// At each sample, each thread registered for profiling gets sampled once.
+// So if there is one "main thread" spending its time in MainFunc() and
+// 4 "worker threads" spending time in WorkerFunc(), then 80% (=4/5) of the
+// samples will be in WorkerFunc, so the profile will look like this:
+//
+// 80% WorkerFunc
+// 20% MainFunc
+
+#ifndef GEMMLOWP_PROFILING_PROFILER_H_
+#define GEMMLOWP_PROFILING_PROFILER_H_
+
+#ifndef GEMMLOWP_PROFILING
+#error Profiling is not enabled!
+#endif
+
+#include <vector>
+
+#include "instrumentation.h"
+
+namespace gemmlowp {
+
+// A tree view of a profile.
+class ProfileTreeView {
+  struct Node {
+    std::vector<Node*> children;
+    const char* label;
+    std::size_t weight;
+    Node() : label(nullptr), weight(0) {}
+    ~Node() {
+      for (auto child : children) {
+        delete child;
+      }
+    }
+  };
+
+  static bool CompareNodes(Node* n1, Node* n2) {
+    return n1->weight > n2->weight;
+  }
+
+  Node root_;
+
+  void PrintNode(const Node* node, int level) const {
+    if (level) {
+      for (int i = 1; i < level; i++) {
+        printf("    ");
+      }
+      printf("%.2f%% %s\n", 100.0f * node->weight / root_.weight, node->label);
+    }
+    for (auto child : node->children) {
+      PrintNode(child, level + 1);
+    }
+  }
+
+  static void AddStackToNode(const ProfilingStack& stack, Node* node,
+                             std::size_t level) {
+    node->weight++;
+    if (stack.size == level) {
+      return;
+    }
+    Node* child_to_add_to = nullptr;
+    for (auto child : node->children) {
+      if (child->label == stack.labels[level]) {
+        child_to_add_to = child;
+        break;
+      }
+    }
+    if (!child_to_add_to) {
+      child_to_add_to = new Node;
+      child_to_add_to->label = stack.labels[level];
+      node->children.push_back(child_to_add_to);
+    }
+    AddStackToNode(stack, child_to_add_to, level + 1);
+    return;
+  }
+
+  void AddStack(const ProfilingStack& stack) {
+    AddStackToNode(stack, &root_, 0);
+  }
+
+  void AddOtherChildrenToNode(Node* node) {
+    std::size_t top_level_children_weight = 0;
+    for (auto c : node->children) {
+      AddOtherChildrenToNode(c);
+      top_level_children_weight += c->weight;
+    }
+    if (top_level_children_weight) {
+      Node* other_child = new Node;
+      other_child->label =
+          node == &root_ ? "other (outside of any label)" : "other";
+      other_child->weight = node->weight - top_level_children_weight;
+      node->children.push_back(other_child);
+    }
+  }
+
+  void AddOtherNodes() { AddOtherChildrenToNode(&root_); }
+
+  void SortNode(Node* node) {
+    std::sort(node->children.begin(), node->children.end(), CompareNodes);
+    for (auto child : node->children) {
+      SortNode(child);
+    }
+  }
+
+  void Sort() { SortNode(&root_); }
+
+ public:
+  explicit ProfileTreeView(const std::vector<ProfilingStack>& stacks) {
+    for (auto stack : stacks) {
+      AddStack(stack);
+    }
+    AddOtherNodes();
+    Sort();
+  }
+
+  void Print() const {
+    printf("\n");
+    printf("gemmlowp profile (%d threads, %d samples)\n",
+           static_cast<int>(ThreadsUnderProfiling().size()),
+           static_cast<int>(root_.weight));
+    PrintNode(&root_, 0);
+    printf("\n");
+  }
+};
+
+// This function is the only place that determines our sampling frequency.
+inline void WaitOneProfilerTick() {
+  static const int millisecond = 1000000;
+
+#if defined __arm__ || defined __aarch64__
+  // Reduced sampling frequency on mobile devices helps limit time and memory
+  // overhead there.
+  static const int interval = 10 * millisecond;
+#else
+  static const int interval = 1 * millisecond;
+#endif
+
+  timespec ts;
+  ts.tv_sec = 0;
+  ts.tv_nsec = interval;
+  nanosleep(&ts, nullptr);
+}
+
+// This is how we track whether we've already started profiling,
+// to guard against misuse of the API.
+inline bool& IsProfiling() {
+  static bool b;
+  return b;
+}
+
+// This is how we tell the profiler thread to finish.
+inline bool& ProfilerThreadShouldFinish() {
+  static bool b;
+  return b;
+}
+
+// The profiler thread. See ProfilerThreadFunc.
+inline pthread_t& ProfilerThread() {
+  static pthread_t t;
+  return t;
+}
+
+// Records a stack from a running thread.
+// The tricky part is that we're not interrupting the thread.
+// This is OK because we're looking at a pseudo-stack of labels,
+// not at the real thread stack, and if the pseudo-stack changes
+// while we're recording it, we are OK with getting either the
+// old or the new stack. Note that ProfilingStack::Pop
+// only decrements the size, and doesn't null the popped label,
+// so if we're concurrently recording it, it shouldn't change
+// under our feet until another label is pushed, at which point
+// we are OK with getting either this new label or the old one.
+// In the end, the key atomicity property that we are relying on
+// here is that pointers are changed atomically, and the labels
+// are pointers (to literal strings).
+inline void RecordStack(const ThreadInfo* thread, ProfilingStack* dst) {
+  assert(!dst->size);
+  while (dst->size < thread->stack.size) {
+    dst->labels[dst->size] = thread->stack.labels[dst->size];
+    dst->size++;
+    MemoryBarrier();  // thread->stack can change at any time
+  }
+}
+
+// The profiler thread's entry point.
+// Note that a separate thread is to be started each time we call
+// StartProfiling(), and finishes when we call FinishProfiling().
+// So here we only need to handle the recording and reporting of
+// a single profile.
+inline void* ProfilerThreadFunc(void*) {
+  assert(ProfilerThread() == pthread_self());
+
+  // Since we only handle one profile per profiler thread, the
+  // profile data (the array of recorded stacks) can be a local variable here.
+  std::vector<ProfilingStack> stacks;
+
+  while (!ProfilerThreadShouldFinish()) {
+    WaitOneProfilerTick();
+    {
+      AutoGlobalLock<ProfilerLockId> lock;
+      for (auto t : ThreadsUnderProfiling()) {
+        ProfilingStack s;
+        RecordStack(t, &s);
+        stacks.push_back(s);
+      }
+    }
+  }
+
+  // Profiling is finished and we now report the results.
+  ProfileTreeView(stacks).Print();
+
+  return nullptr;
+}
+
+// Starts recording samples.
+inline void StartProfiling() {
+  AutoGlobalLock<ProfilerLockId> lock;
+  ReleaseBuildAssertion(!IsProfiling(), "We're already profiling!");
+  IsProfiling() = true;
+  ProfilerThreadShouldFinish() = false;
+  pthread_create(&ProfilerThread(), nullptr, ProfilerThreadFunc, nullptr);
+}
+
+// Stops recording samples, and prints a profile tree-view on stdout.
+inline void FinishProfiling() {
+  {
+    AutoGlobalLock<ProfilerLockId> lock;
+    ReleaseBuildAssertion(IsProfiling(), "We weren't profiling!");
+    // The ProfilerThreadShouldFinish() mechanism here is really naive and bad,
+    // as the scary comments below should make clear.
+    // Should we use a condition variable?
+    ProfilerThreadShouldFinish() = true;
+  }  // must release the lock here to avoid deadlock with profiler thread.
+  pthread_join(ProfilerThread(), nullptr);
+  IsProfiling() = false;  // yikes, this should be guarded by the lock!
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_PROFILING_PROFILER_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/public/bit_depth.h b/src/runtime/ref/nn/depend/external/gemmlowp/public/bit_depth.h
new file mode 100644
index 0000000..6cb4ecf
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/public/bit_depth.h
@@ -0,0 +1,62 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// bit_depth.h: defines the settins controlling LHS/RHS bit depth
+
+#ifndef GEMMLOWP_PUBLIC_BIT_DEPTH_H_
+#define GEMMLOWP_PUBLIC_BIT_DEPTH_H_
+
+namespace gemmlowp {
+
+// The range of allowed values for an operand.
+template <int tMinValue, int tMaxValue>
+struct OperandRange {
+  static const int kMinValue = tMinValue;
+  static const int kMaxValue = tMaxValue;
+  static_assert(0 <= kMinValue, "");
+  static_assert(kMinValue < kMaxValue, "");
+  static_assert(kMaxValue <= 255, "");
+};
+
+using Uint8Range = OperandRange<0, 255>;
+using Uint8RangeExcludingZero = OperandRange<1, 255>;
+
+template <typename tLhsRange, typename tRhsRange>
+struct BitDepthParams {
+  using LhsRange = tLhsRange;
+  using RhsRange = tRhsRange;
+};
+
+// Default: LHS and RHS are 8bit.
+using DefaultL8R8BitDepthParams = BitDepthParams<Uint8Range, Uint8Range>;
+
+// Variant: LHS may not take the value 0. This allows using
+// faster kernels using signed arithmetic, see
+// NEON_64bit_GEMM_Int8Operands_Int32Accumulators_AccumTwoWithin16Bits
+using L8R8WithLhsNonzeroBitDepthParams =
+    BitDepthParams<Uint8RangeExcludingZero, Uint8Range>;
+
+// Deprecated: when gemmlowp used to allow requantizing 8bit
+// inputs to less-than-8-bit depths, the public setting allowing
+// that was DefaultL7R5BitDepthParams. That requantization
+// feature has been removed, but as the whole point of that
+// requantization was to make less-than-8-bit an internal
+// optimization without any impact on the API (other than lowering
+// accuracy), we can temporarily support users who were using it
+// by mapping it to the default 8bit behavior.
+using DefaultL7R5BitDepthParams = DefaultL8R8BitDepthParams;
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_PUBLIC_BIT_DEPTH_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/public/gemmlowp.h b/src/runtime/ref/nn/depend/external/gemmlowp/public/gemmlowp.h
new file mode 100644
index 0000000..05b0f47
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/public/gemmlowp.h
@@ -0,0 +1,87 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// gemmlowp.h: the main public interface header of gemmlowp.
+
+#ifndef GEMMLOWP_PUBLIC_GEMMLOWP_H_
+#define GEMMLOWP_PUBLIC_GEMMLOWP_H_
+#include "../internal/dispatch_gemm_shape.h"
+#include "bit_depth.h"
+#include "map.h"
+#include "output_stages.h"
+
+namespace gemmlowp {
+
+class GemmContext : public MultiThreadGemmContext {};
+
+// Computes a general matrix product ("GEMM").
+// This is a version that supports per channel quantization.
+template <typename InputScalar, typename OutputScalar, typename BitDepthParams,
+          MapOrder LhsOrder, MapOrder RhsOrder, MapOrder ResultOrder,
+          typename LhsOffset, typename RhsOffset, typename OutputPipelineType,
+          typename GemmContextType>
+void GemmWithOutputPipelinePC(GemmContextType* context,
+                              const MatrixMap<const InputScalar, LhsOrder>& lhs,
+                              const MatrixMap<const InputScalar, RhsOrder>& rhs,
+                              MatrixMap<OutputScalar, ResultOrder>* result,
+                              const LhsOffset& lhs_offset,
+                              const RhsOffset& rhs_offset,
+                              const OutputPipelineType& output_pipeline) {
+  DispatchGemmShape<InputScalar, OutputScalar, BitDepthParams>(
+      context, lhs, rhs, result, lhs_offset, rhs_offset, output_pipeline);
+}
+
+// Computes a general matrix product ("GEMM").
+// This is the legacy version that does not support per channel quantization.
+// The meaning of the offsets, result_mult_int and result_shift
+// parameters is the same as in the standard EightBitIntGemm interface
+// (which is also implemented in the eight_bit_int_gemm directory).
+template <typename InputScalar, typename OutputScalar, typename BitDepthParams,
+          MapOrder LhsOrder, MapOrder RhsOrder, MapOrder ResultOrder,
+          typename OutputPipelineType, typename GemmContextType>
+void GemmWithOutputPipeline(GemmContextType* context,
+                            const MatrixMap<const InputScalar, LhsOrder>& lhs,
+                            const MatrixMap<const InputScalar, RhsOrder>& rhs,
+                            MatrixMap<OutputScalar, ResultOrder>* result,
+                            int lhs_offset, int rhs_offset,
+                            const OutputPipelineType& output_pipeline) {
+  typedef VectorDup<const std::int32_t, VectorShape::Col> OffsetColDup;
+  typedef VectorDup<const std::int32_t, VectorShape::Row> OffsetRowDup;
+  const OffsetColDup lhs_offset_vector(lhs_offset, lhs.rows());
+  const OffsetRowDup rhs_offset_vector(rhs_offset, rhs.cols());
+  DispatchGemmShape<InputScalar, OutputScalar, BitDepthParams>(
+      context, lhs, rhs, result, lhs_offset_vector, rhs_offset_vector,
+      output_pipeline);
+}
+
+// Computes a general matrix product ("GEMM").
+// The meaning of the offsets, result_mult_int and result_shift
+// parameters is the same as in the standard EightBitIntGemm interface
+// (which is also implemented in the eight_bit_int_gemm directory).
+template <typename Scalar, typename BitDepthParams, MapOrder LhsOrder,
+          MapOrder RhsOrder, MapOrder ResultOrder, typename GemmContextType>
+void Gemm(GemmContextType* context,
+          const MatrixMap<const Scalar, LhsOrder>& lhs,
+          const MatrixMap<const Scalar, RhsOrder>& rhs,
+          MatrixMap<Scalar, ResultOrder>* result, int lhs_offset,
+          int rhs_offset, int result_offset, int result_mult_int,
+          int result_shift) {
+  GemmWithOutputPipeline<Scalar, Scalar, BitDepthParams>(
+      context, lhs, rhs, result, lhs_offset, rhs_offset,
+      MakeStandardOutputPipeline(result_offset, result_mult_int, result_shift));
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_PUBLIC_GEMMLOWP_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/public/map.h b/src/runtime/ref/nn/depend/external/gemmlowp/public/map.h
new file mode 100644
index 0000000..3073e05
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/public/map.h
@@ -0,0 +1,140 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// map.h: a minimalist view-existing-buffer-as-a-matrix class,
+// which is how gemmlowp interfaces with external matrix data.
+
+#ifndef GEMMLOWP_PUBLIC_MAP_H_
+#define GEMMLOWP_PUBLIC_MAP_H_
+
+#include "../internal/common.h"
+
+namespace gemmlowp {
+
+// The two storage orders allowed to map buffers as matrices: ColMajor
+// means column-major, RowMajor means row-major.
+enum class MapOrder { ColMajor, RowMajor };
+
+// A MatrixMap is a view of an existing buffer as a matrix. It does not own
+// the buffer.
+template <typename tScalar, MapOrder tOrder>
+class MatrixMap {
+ public:
+  typedef tScalar Scalar;
+  static const MapOrder kOrder = tOrder;
+
+ protected:
+  Scalar* data_;  // not owned.
+  int rows_, cols_, stride_;
+
+ public:
+  MatrixMap() : data_(nullptr), rows_(0), cols_(0), stride_(0) {}
+  MatrixMap(Scalar* data, int rows, int cols)
+      : data_(data),
+        rows_(rows),
+        cols_(cols),
+        stride_(kOrder == MapOrder::ColMajor ? rows : cols) {}
+  MatrixMap(Scalar* data, int rows, int cols, int stride)
+      : data_(data), rows_(rows), cols_(cols), stride_(stride) {}
+  MatrixMap(const MatrixMap& other)
+      : data_(other.data_),
+        rows_(other.rows_),
+        cols_(other.cols_),
+        stride_(other.stride_) {}
+
+  int rows() const { return rows_; }
+  int cols() const { return cols_; }
+  int stride() const { return stride_; }
+  int rows_stride() const { return kOrder == MapOrder::ColMajor ? 1 : stride_; }
+  int cols_stride() const { return kOrder == MapOrder::RowMajor ? 1 : stride_; }
+  Scalar* data() const { return data_; }
+  Scalar* data(int row, int col) const {
+    return data_ + row * rows_stride() + col * cols_stride();
+  }
+  Scalar& operator()(int row, int col) const { return *data(row, col); }
+
+  MatrixMap block(int start_row, int start_col, int block_rows,
+                  int block_cols) const {
+    assert(start_row >= 0);
+    assert(start_row + block_rows <= rows_);
+    assert(start_col >= 0);
+    assert(start_col + block_cols <= cols_);
+
+    return MatrixMap(data(start_row, start_col), block_rows, block_cols,
+                     stride_);
+  }
+};
+
+enum class VectorShape { Col, Row };
+
+// A VectorMap is a view of an existing buffer as a vector. It does not own
+// the buffer.
+template <typename tScalar, VectorShape tShape>
+class VectorMap {
+ public:
+  typedef tScalar Scalar;
+  static const VectorShape kShape = tShape;
+
+ protected:
+  Scalar* data_;  // not owned.
+  int size_;
+
+ public:
+  VectorMap() : data_(nullptr), size_(0) {}
+  VectorMap(Scalar* data, int size) : data_(data), size_(size) {}
+  VectorMap(const VectorMap& other) : data_(other.data_), size_(other.size_) {}
+
+  int size() const { return size_; }
+  Scalar* data() const { return data_; }
+  Scalar* data(int index) const { return data_ + index; }
+  Scalar& operator()(int index) const { return *data(index); }
+
+  VectorMap block(int start, int len) const {
+    assert(start >= 0);
+    assert(start + len <= size_);
+
+    return VectorMap(data(start), len);
+  }
+};
+
+// A VectorDup is a (duplicated value) vector where all components are the same.
+template <typename tScalar, VectorShape tShape>
+class VectorDup {
+ public:
+  typedef tScalar Scalar;
+  static const VectorShape kShape = tShape;
+
+ protected:
+  Scalar data_;
+  int size_;
+
+ public:
+  VectorDup() : data_(0), size_(0) {}
+  VectorDup(Scalar data, int size) : data_(data), size_(size) {}
+  VectorDup(const VectorDup& other) : data_(other.data_), size_(other.size_) {}
+
+  int size() const { return size_; }
+  Scalar& operator()(int) const { return data_; }
+
+  VectorDup block(int start, int len) const {
+    assert(start >= 0);
+    assert(start + len <= size_);
+
+    return VectorDup(data_, len);
+  }
+};
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_PUBLIC_MAP_H_
diff --git a/src/runtime/ref/nn/depend/external/gemmlowp/public/output_stages.h b/src/runtime/ref/nn/depend/external/gemmlowp/public/output_stages.h
new file mode 100644
index 0000000..23bcdc0
--- /dev/null
+++ b/src/runtime/ref/nn/depend/external/gemmlowp/public/output_stages.h
@@ -0,0 +1,185 @@
+// Copyright 2015 The Gemmlowp Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// output_stages.h: public definitions of the output stages that can
+// be assembled into an output pipeline, to control how internal
+// 32-bit accumulators are transformed to obtain the final uint8
+// result matrix entries.
+
+#ifndef GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_
+#define GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_
+
+#include <tuple>
+
+#include "../internal/common.h"
+
+namespace gemmlowp {
+
+// This output stage takes int32 values and returns still int32 values,
+// but "quantized down" to the uint8 scale; in other words, its output
+// is typically what one would then clamp to [0..255] and cast to uint8
+// (see OutputStageSaturatingCastToUint8).
+//
+// This "quantization down" process depends on 3 parameters,
+//   result_offset, result_mult_int, result_shift,
+// and the result is:
+//   ((input + result_offset) * result_mult_int + rounding) >> result_shift
+// where
+//   rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1));
+struct OutputStageQuantizeDownInt32ToUint8Scale {
+  std::int32_t result_offset;
+  std::int32_t result_mult_int;
+  std::int32_t result_shift;
+};
+
+// This output stage takes int32 values and returns still int32 values,
+// but "quantized down" to the uint8 scale; in other words, its output
+// is typically what one would then clamp to [0..255] and cast to uint8
+// (see OutputStageSaturatingCastToUint8).
+//
+// This "quantization down" process depends on 3 parameters,
+//   result_offset, result_mult_int, result_shift,
+// and the result is:
+//   ((input + result_offset) * result_mult_int + rounding) >> result_shift
+// where
+//   rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1));
+//
+// Difference from OutputStageQuantizeDownInt32ToUint8Scale here is that each
+// row or column of the output (depending on tShape) has its own result_offset
+// and result_mult_int numbers.
+template <VectorShape tShape>
+struct OutputStageQuantizeDownInt32ToUint8ScalePC {
+  VectorMap<const std::int32_t, tShape> result_offset;
+  VectorMap<const std::int32_t, tShape> result_mult_int;
+  std::int32_t result_shift;
+};
+
+// This output stage takes int32 values and returns still int32 values,
+// but "quantized down" to the uint8 scale; in other words, its output
+// is typically what one would then clamp to [0..255] and cast to uint8
+// (see OutputStageSaturatingCastToUint8).
+//
+// This "quantization down" process depends on 3 parameters,
+//   result_offset, result_fixedpoint_multiplier, result_shift,
+// and the result is:
+//   ((FixedPointMul(input, result_fixedpoint_multiplier) +
+//   rounding) >> result_shift) + result_offset_after_shift
+// where
+//   rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1));
+// and where FixedPointMul(x, y) is the nearest integer to the following
+// mathematical expression, evaluated without overflow or intermediate
+// rounding:
+//   (x * y) / 2^31
+// In practice, it is expected that FixedPointMul will be implemented
+// using hardware "rounding doubling int32 multiply high" instructions,
+// such as VQRDMULH on ARM. See in fixedpoint.h the generic function,
+// SaturatingRoundingDoublingHighMul.
+//
+// Notice that the other difference from
+// OutputStageQuantizeDownInt32ToUint8Scale is that the result offset
+// is applied after the multiplier and shift, not before. This ensures
+// that no matter what the multiplier and shift are, the result offset
+// is effectively integral: offsetting the final result by an integer.
+// The motivation for this is to faithfully support quantization schemes
+// where the formula linking quantized values to the real mathematical
+// values that they represent, is of the form
+//
+//   real_value = scale * (quantized_value - zero_point)
+//
+// where scale is a real number (represented in quantized form by
+// result_fixedpoint_multiplier and result_shift) and zero_point
+// is an integer telling which quantized value correspond to the
+// real value 0, and is represented here by (the opposite of)
+// result_offset_after_shift.
+// The motivation for such a quantization scheme, designed to
+// ensure that 0 is always a representable value, is that in
+// many applications, we need to 0-pad arrays and that can only be
+// done for quantized arrays if 0 is a representable value in
+// quantized form. In particular, convolution-like operations
+// are often implemented using 0-padding, or "im2col"-like
+// expansions that implicitly rely on 0-padding. If 0 were not
+// a representable value, such operations would have to pad
+// using a nonzero value, introducing bias in the computation.
+struct OutputStageQuantizeDownInt32ToUint8ScaleByFixedPoint {
+  std::int32_t result_fixedpoint_multiplier;
+  std::int32_t result_shift;
+  std::int32_t result_offset_after_shift;
+};
+
+// This output stage takes int32 values that are expected to be already
+// on the final uint8 scale, but not necessarily in the [0..255] range.
+// It clamps them to the [0..255] range and returns them casted to uint8.
+struct OutputStageSaturatingCastToUint8 {};
+
+// This output stage depends on a "bias vector" that should contain int32
+// entries, and be either a row-vector of the same number of columns as the
+// result matrix, or a column-vector of the same number of rows as the
+// result matrix. This output stage takes int32 values and adds to them
+// the corresponding entry of the bias vector (broadcasted in the other
+// direction to fit the matrix's shape), outputting int32 values.
+template <typename VectorType>
+struct OutputStageBiasAddition {
+  VectorType bias_vector;
+};
+
+// This output stage clamps value between the specified min and max bounds.
+// It can be used to implement "rectified linear unit" activation functions
+// in neural networks.
+struct OutputStageClamp {
+  std::int32_t min;
+  std::int32_t max;
+};
+
+struct OutputStageTanh {
+  std::int32_t real_zero_as_int32;
+  std::int32_t real_amplitude_as_int32;
+};
+
+// An output pipeline is just a std::tuple of output stages.
+// This function generates a standard output pipeline consisting of two stages:
+// OutputStageQuantizeDownInt32ToUint8Scale, OutputStageSaturatingCastToUint8.
+inline std::tuple<OutputStageQuantizeDownInt32ToUint8Scale,
+                  OutputStageSaturatingCastToUint8>
+MakeStandardOutputPipeline(std::int32_t result_offset,
+                           std::int32_t result_mult_int,
+                           std::int32_t result_shift) {
+  OutputStageQuantizeDownInt32ToUint8Scale quantize_down_stage;
+  quantize_down_stage.result_offset = result_offset;
+  quantize_down_stage.result_mult_int = result_mult_int;
+  quantize_down_stage.result_shift = result_shift;
+  OutputStageSaturatingCastToUint8 saturating_cast_stage;
+  return std::make_tuple(quantize_down_stage, saturating_cast_stage);
+}
+
+// An output pipeline is just a std::tuple of output stages.
+// This function generates a standard output pipeline consisting of two stages:
+// OutputStageQuantizeDownInt32ToUint8ScalePC, OutputStageSaturatingCastToUint8.
+template <VectorShape tShape>
+inline std::tuple<OutputStageQuantizeDownInt32ToUint8ScalePC<tShape>,
+                  OutputStageSaturatingCastToUint8>
+MakeStandardOutputPipeline(
+    const VectorMap<const std::int32_t, tShape>& result_offset,
+    const VectorMap<const std::int32_t, tShape>& result_mult_int,
+    std::int32_t result_shift) {
+  OutputStageQuantizeDownInt32ToUint8ScalePC<tShape> quantize_down_stage;
+  quantize_down_stage.result_offset = result_offset;
+  quantize_down_stage.result_mult_int = result_mult_int;
+  quantize_down_stage.result_shift = result_shift;
+  OutputStageSaturatingCastToUint8 saturating_cast_stage;
+  return std::make_tuple(quantize_down_stage, saturating_cast_stage);
+}
+
+}  // namespace gemmlowp
+
+#endif  // GEMMLOWP_PUBLIC_OUTPUT_STAGES_H_