Introduce FullyConnectedReshapingLayer into ARMComputeEX (#5140)
author장지섭/On-Device Lab(SR)/Engineer/삼성전자 <jiseob.jang@samsung.com>
Tue, 7 May 2019 04:44:19 +0000 (13:44 +0900)
committer오형석/On-Device Lab(SR)/Staff Engineer/삼성전자 <hseok82.oh@samsung.com>
Tue, 7 May 2019 04:44:19 +0000 (13:44 +0900)
This commit introduces FullyConnectedReshapingLayer into ARMComputeEx.

Signed-off-by: jiseob.jang <jiseob.jang@samsung.com>
libs/ARMComputeEx/arm_compute/runtime/CL/CLFunctionsEx.h
libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLFullyConnectedReshapingLayer.h [new file with mode: 0644]
libs/ARMComputeEx/arm_compute/runtime/NEON/NEFunctionsEx.h
libs/ARMComputeEx/arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h [new file with mode: 0644]
libs/ARMComputeEx/src/runtime/CL/functions/CLFullyConnectedReshapingLayer.cpp [new file with mode: 0644]
libs/ARMComputeEx/src/runtime/NEON/functions/NEFullyConnectedReshapingLayer.cpp [new file with mode: 0644]
runtimes/neurun/backend/acl_neon/CMakeLists.txt

index 470766e..d9376fa 100644 (file)
@@ -25,6 +25,7 @@
 #include <arm_compute/runtime/CL/functions/CLDepthToSpace.h>
 #include <arm_compute/runtime/CL/functions/CLEmbeddingLookup.h>
 #include <arm_compute/runtime/CL/functions/CLExp.h>
+#include <arm_compute/runtime/CL/functions/CLFullyConnectedReshapingLayer.h>
 #include <arm_compute/runtime/CL/functions/CLGather.h>
 #include <arm_compute/runtime/CL/functions/CLHashtableLookup.h>
 #include <arm_compute/runtime/CL/functions/CLLogicalNot.h>
diff --git a/libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLFullyConnectedReshapingLayer.h b/libs/ARMComputeEx/arm_compute/runtime/CL/functions/CLFullyConnectedReshapingLayer.h
new file mode 100644 (file)
index 0000000..774a311
--- /dev/null
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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.
+ */
+
+/**
+ * @file        CLFullyConnectedReshapingLayer.h
+ * @brief       This file contains CLFullyConnectedReshapingLayer class
+ * @ingroup     COM_AI_RUNTIME
+ */
+
+#ifndef __ARM_COMPUTE_CL_FULLY_CONNECTED_RESHAPING_LAYER_H__
+#define __ARM_COMPUTE_CL_FULLY_CONNECTED_RESHAPING_LAYER_H__
+
+#include <arm_compute/runtime/CL/functions/CLReshapeLayer.h>
+#include <arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h>
+
+namespace arm_compute
+{
+/**
+ * @brief Class to run FullyConnected Layer after reshaping input tensor
+ */
+class CLFullyConnectedReshapingLayer : public arm_compute::IFunction
+{
+public:
+  CLFullyConnectedReshapingLayer(void)
+      : _input(nullptr), _weights(nullptr), _biases(nullptr), _output(nullptr), _cl_buffer{},
+        _cl_fc{}, _cl_reshape{}, _needs_reshape(false)
+  {
+    // DO NOTHING
+  }
+
+public:
+  /**
+   * @brief Configure the layer
+   * @param[in] input The source tensor
+   * @param[in] weights The tensor that is filled with weight values
+   * @param[in] biases The tensor that is filled with biase values
+   * @param[in] output The destination tensor
+   * @param[in] needs_reshape Whether it needs to be reshaped or not
+   * @param[in] reshape The tensor shape to be reshaped. Only valid when needs_reshape is true.
+   * @return N/A
+   */
+  void configure(const arm_compute::ICLTensor *input, const arm_compute::ICLTensor *weights,
+                 const arm_compute::ICLTensor *biases, arm_compute::ICLTensor *output,
+                 bool needs_reshape, const arm_compute::TensorShape &reshape);
+
+public:
+  /**
+   * @brief Run the operation. Must be called after configure().
+   * @return N/A
+   */
+  void run(void) override;
+
+private:
+  const arm_compute::ICLTensor *_input;
+  const arm_compute::ICLTensor *_weights;
+  const arm_compute::ICLTensor *_biases;
+  arm_compute::ICLTensor *_output;
+
+  // buffer for reshaping input tensor
+  arm_compute::CLTensor _cl_buffer;
+
+private:
+  arm_compute::CLFullyConnectedLayer _cl_fc;
+  CLReshapeLayer _cl_reshape;
+  bool _needs_reshape;
+};
+} // namespace arm_compute
+
+#endif // __ARM_COMPUTE_CL_FULLY_CONNECTED_RESHAPING_LAYER_H__
index 7897df4..1518972 100644 (file)
@@ -16,6 +16,6 @@
 #ifndef __ARM_COMPUTE_NEFUNCTIONSEX_H__
 #define __ARM_COMPUTE_NEFUNCTIONSEX_H__
 
-// Add new NE LayerEx header here
+#include <arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h>
 
 #endif // __ARM_COMPUTE_NEFUNCTIONSEX_H__
diff --git a/libs/ARMComputeEx/arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h b/libs/ARMComputeEx/arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h
new file mode 100644 (file)
index 0000000..1273164
--- /dev/null
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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.
+ */
+
+/**
+ * @file        NEFullyConnectedReshapingLayer.h
+ * @brief       This file contains NEFullyConnectedReshapingLayer class
+ * @ingroup     COM_AI_RUNTIME
+ */
+
+#ifndef __ARM_COMPUTE_NE_FULLY_CONNECTED_RESHAPING_LAYER_H__
+#define __ARM_COMPUTE_NE_FULLY_CONNECTED_RESHAPING_LAYER_H__
+
+#include <arm_compute/runtime/NEON/functions/NEReshapeLayer.h>
+#include <arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h>
+
+namespace arm_compute
+{
+/**
+ * @brief Class to run FullyConnected Layer after reshaping input tensor
+ */
+class NEFullyConnectedReshapingLayer : public arm_compute::IFunction
+{
+public:
+  NEFullyConnectedReshapingLayer(void)
+      : _input(nullptr), _weights(nullptr), _biases(nullptr), _output(nullptr), _neon_buffer{},
+        _neon_fc{}, _neon_reshape{}, _needs_reshape(false)
+  {
+    // DO NOTHING
+  }
+
+public:
+  /**
+   * @brief Configure the layer
+   * @param[in] input The source tensor
+   * @param[in] weights The tensor that is filled with weight values
+   * @param[in] biases The tensor that is filled with biase values
+   * @param[in] output The destination tensor
+   * @param[in] needs_reshape Whether it needs to be reshaped or not
+   * @param[in] reshape The tensor shape to be reshaped. Only valid when needs_reshape is true.
+   * @return N/A
+   */
+  void configure(const arm_compute::ITensor *input, const arm_compute::ITensor *weights,
+                 const arm_compute::ITensor *biases, arm_compute::ITensor *output,
+                 bool needs_reshape, const arm_compute::TensorShape &reshape);
+
+public:
+  /**
+   * @brief Run the operation. Must be called after configure().
+   * @return N/A
+   */
+  void run(void) override;
+
+private:
+  const arm_compute::ITensor *_input;
+  const arm_compute::ITensor *_weights;
+  const arm_compute::ITensor *_biases;
+  arm_compute::ITensor *_output;
+
+  // buffer for reshaping input tensor
+  arm_compute::Tensor _neon_buffer;
+
+private:
+  arm_compute::NEFullyConnectedLayer _neon_fc;
+  NEReshapeLayer _neon_reshape;
+  bool _needs_reshape;
+};
+} // namespace arm_compute
+
+#endif // __ARM_COMPUTE_NE_FULLY_CONNECTED_RESHAPING_LAYER_H__
diff --git a/libs/ARMComputeEx/src/runtime/CL/functions/CLFullyConnectedReshapingLayer.cpp b/libs/ARMComputeEx/src/runtime/CL/functions/CLFullyConnectedReshapingLayer.cpp
new file mode 100644 (file)
index 0000000..e55ee28
--- /dev/null
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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.
+ */
+
+#include "arm_compute/runtime/CL/functions/CLFullyConnectedReshapingLayer.h"
+
+using namespace arm_compute;
+
+void CLFullyConnectedReshapingLayer::configure(const arm_compute::ICLTensor *input,
+                                               const arm_compute::ICLTensor *weights,
+                                               const arm_compute::ICLTensor *biases,
+                                               arm_compute::ICLTensor *output, bool needs_reshape,
+                                               const arm_compute::TensorShape &reshape)
+{
+  _input = input;
+  _weights = weights;
+  _biases = biases;
+  _output = output;
+  _needs_reshape = needs_reshape;
+
+  if (_needs_reshape)
+  {
+    // reshape
+    auto_init_if_empty(*_cl_buffer.info(), _input->info()->clone()->set_tensor_shape(reshape));
+    _cl_reshape.configure(_input, &_cl_buffer);
+
+    _cl_fc.configure(&_cl_buffer, _weights, _biases, _output);
+
+    // NOTE _cl_buffer is inaccessible from outside, and thus it is safe to invoke allocate here.
+    _cl_buffer.allocator()->allocate();
+  }
+  else
+  {
+    _cl_fc.configure(_input, _weights, _biases, _output);
+  }
+}
+
+void CLFullyConnectedReshapingLayer::run(void)
+{
+  if (_needs_reshape)
+    _cl_reshape.run();
+
+  _cl_fc.run();
+}
diff --git a/libs/ARMComputeEx/src/runtime/NEON/functions/NEFullyConnectedReshapingLayer.cpp b/libs/ARMComputeEx/src/runtime/NEON/functions/NEFullyConnectedReshapingLayer.cpp
new file mode 100644 (file)
index 0000000..6649aa6
--- /dev/null
@@ -0,0 +1,56 @@
+/*
+ * Copyright (c) 2019 Samsung Electronics Co., Ltd. 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.
+ */
+
+#include "arm_compute/runtime/NEON/functions/NEFullyConnectedReshapingLayer.h"
+
+using namespace arm_compute;
+
+void NEFullyConnectedReshapingLayer::configure(const arm_compute::ITensor *input,
+                                               const arm_compute::ITensor *weights,
+                                               const arm_compute::ITensor *biases,
+                                               arm_compute::ITensor *output, bool needs_reshape,
+                                               const arm_compute::TensorShape &reshape)
+{
+  _input = input;
+  _weights = weights;
+  _biases = biases;
+  _output = output;
+  _needs_reshape = needs_reshape;
+
+  if (_needs_reshape)
+  {
+    // reshape
+    auto_init_if_empty(*_neon_buffer.info(), _input->info()->clone()->set_tensor_shape(reshape));
+    _neon_reshape.configure(_input, &_neon_buffer);
+
+    _neon_fc.configure(&_neon_buffer, _weights, _biases, _output);
+
+    // NOTE _neon_buffer is inaccessible from outside, and thus it is safe to invoke allocate here.
+    _neon_buffer.allocator()->allocate();
+  }
+  else
+  {
+    _neon_fc.configure(_input, _weights, _biases, _output);
+  }
+}
+
+void NEFullyConnectedReshapingLayer::run(void)
+{
+  if (_needs_reshape)
+    _neon_reshape.run();
+
+  _neon_fc.run();
+}
index 850b7c0..8b1ae5d 100644 (file)
@@ -21,6 +21,7 @@ target_include_directories(${LIB_NEURUN_BACKEND_ACL_NEON} PRIVATE ${CMAKE_CURREN
 target_include_directories(${LIB_NEURUN_KERNEL_ACL_NEON} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR}/..)
 
 target_link_libraries(${LIB_NEURUN_BACKEND_ACL_NEON} arm_compute)
+target_link_libraries(${LIB_NEURUN_BACKEND_ACL_NEON} arm_compute_ex)
 target_link_libraries(${LIB_NEURUN_BACKEND_ACL_NEON} ${LIB_NEURUN_KERNEL_ACL_NEON})
 target_link_libraries(${LIB_NEURUN_BACKEND_ACL_NEON} neurun-core)