[locomotiv] Support FeatureEncode execution (#3602)

* [locomotiv] Support FeatureEncode execution

This commit supports FeatureEncode execution for loco interpreter. Note
that locomotiv saves all feature data as NHWC.

Signed-off-by: Cheongyo Bahk <ch.bahk@samsung.com>
* Fix typo

diff --git a/contrib/locomotiv/src/Node.lst b/contrib/locomotiv/src/Node.lst
index 3dc5fdd..dced28d 100644 (file)
--- a/contrib/locomotiv/src/Node.lst
+++ b/contrib/locomotiv/src/Node.lst
@@ -5,6 +5,7 @@
 // NODE(Name) : alphabetic order please
 
 NODE(ConstGen)
+NODE(FeatureEncode)
 NODE(Forward)
 NODE(MaxPool2D)
 NODE(Pull)

diff --git a/contrib/locomotiv/src/Node/FeatureEncode.cpp b/contrib/locomotiv/src/Node/FeatureEncode.cpp
new file mode 100644 (file)
index 0000000..c05a0bb
--- /dev/null
+++ b/contrib/locomotiv/src/Node/FeatureEncode.cpp
@@ -0,0 +1,117 @@
+/*
+ * 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 "NodeExecution.h"
+
+#include "NodeDataImpl.h"
+
+#include <nncc/core/ADT/tensor/LexicalLayout.h>
+#include <nncc/core/ADT/tensor/IndexEnumerator.h>
+
+#include <stdexcept>
+#include <cassert>
+
+namespace
+{
+
+using nncc::core::ADT::tensor::Buffer;
+using nncc::core::ADT::tensor::make_buffer;
+using nncc::core::ADT::tensor::LexicalLayout;
+using nncc::core::ADT::tensor::Shape;
+using nncc::core::ADT::tensor::IndexEnumerator;
+
+template <typename T>
+std::unique_ptr<locomotiv::NodeData> feature_encode(const loco::FeatureEncode *node,
+                                                    const Buffer<T> *input_buf)
+{
+  auto encoder = node->encoder();
+
+  // Make TensorShape from input
+  loco::TensorShape input_shape;
+  input_shape.rank(input_buf->shape().rank());
+  assert(input_shape.rank() == 4);
+  for (uint32_t i = 0; i < input_shape.rank(); ++i)
+  {
+    input_shape.dim(i) = loco::make_dimension(input_buf->shape().dim(i));
+  }
+
+  loco::FeatureShape node_shape = encoder->shape(input_shape);
+
+  // Make NHWC buffer from FeatureShape
+  Buffer<T> node_buf =
+      make_buffer<T, LexicalLayout>(Shape{node_shape.count().value(), node_shape.height().value(),
+                                          node_shape.width().value(), node_shape.depth().value()});
+
+  // Copy buffer in an order arranged by encoder
+  for (IndexEnumerator e{node_buf.shape()}; e.valid(); e.advance())
+  {
+    loco::FeatureIndex index;
+    index.batch() = e.current().at(0);
+    index.row() = e.current().at(1);
+    index.column() = e.current().at(2);
+    index.channel() = e.current().at(3);
+
+    node_buf.at(e.current()) = input_buf->at(encoder->value(index));
+  }
+
+  return locomotiv::make_data(node_buf);
+}
+
+} // namespace
+
+namespace locomotiv
+{
+
+void NodeExecution::execute(loco::FeatureEncode *enc)
+{
+  auto input_data = annot_data(enc->input());
+
+  if (!input_data)
+  {
+    throw std::runtime_error("Input of FeatureEncode not ready");
+  }
+
+  if (input_data->shape()->rank() != 4)
+  {
+    throw std::runtime_error("Input shape mismatch");
+  }
+
+  std::unique_ptr<NodeData> enc_data = nullptr;
+
+  switch (input_data->dtype())
+  {
+  case loco::DataType::S32:
+  {
+    auto input_buf = input_data->as_s32_bufptr();
+    enc_data = feature_encode<int32_t>(enc, input_buf);
+    break;
+  }
+  case loco::DataType::FLOAT32:
+  {
+    auto input_buf = input_data->as_f32_bufptr();
+    enc_data = feature_encode<float>(enc, input_buf);
+    break;
+  }
+  default:
+    throw std::runtime_error("NYI for this DataType");
+  }
+
+  assert(enc_data != nullptr);
+  erase_annot_data(enc);
+  annot_data(enc, std::move(enc_data));
+}
+
+} // namespace locomotiv

diff --git a/contrib/locomotiv/src/Node/FeatureEncode.test.cpp b/contrib/locomotiv/src/Node/FeatureEncode.test.cpp
new file mode 100644 (file)
index 0000000..a4bfd9b
--- /dev/null
+++ b/contrib/locomotiv/src/Node/FeatureEncode.test.cpp
@@ -0,0 +1,137 @@
+/*
+ * 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 "NodeExecution.h"
+
+#include "locomotiv/NodeData.h"
+#include "NodeDataImpl.h"
+
+#include <loco/IR/PermutingCodec.h>
+
+#include <nncc/core/ADT/tensor/Shape.h>
+#include <nncc/core/ADT/tensor/Buffer.h>
+#include <nncc/core/ADT/tensor/LexicalLayout.h>
+#include <nncc/core/ADT/tensor/IndexEnumerator.h>
+
+#include <gtest/gtest.h>
+
+using nncc::core::ADT::tensor::Index;
+using nncc::core::ADT::tensor::Shape;
+using nncc::core::ADT::tensor::LexicalLayout;
+using nncc::core::ADT::tensor::make_buffer;
+using nncc::core::ADT::tensor::IndexEnumerator;
+
+TEST(NodeExecution_FeatureEncode, s32)
+{
+  const uint32_t N = 2;
+  const uint32_t H = 3;
+  const uint32_t W = 4;
+  const uint32_t C = 5;
+
+  auto g = loco::make_graph();
+
+  // Pull
+  auto pull = g->nodes()->create<loco::Pull>();
+  pull->dtype(loco::DataType::S32);
+
+  // Make and assign "NCHW" data to pull node
+  auto pull_buf = make_buffer<int32_t, LexicalLayout>(Shape{N, C, H, W});
+  int32_t i = 1;
+  for (IndexEnumerator e{pull_buf.shape()}; e.valid(); e.advance())
+  {
+    pull_buf.at(e.current()) = i;
+    i *= -3; // Doesn't matter what it is
+  }
+  auto pull_data = locomotiv::make_data(pull_buf);
+  locomotiv::annot_data(pull, std::move(pull_data));
+
+  // Encoder to correctly read input tensor as NCHW
+  auto encoder = std::unique_ptr<loco::PermutingEncoder<loco::Domain::Feature>>(
+      new loco::PermutingEncoder<loco::Domain::Feature>);
+  encoder->perm()->axis(loco::FeatureAxis::Count) = 0;
+  encoder->perm()->axis(loco::FeatureAxis::Height) = 2;
+  encoder->perm()->axis(loco::FeatureAxis::Width) = 3;
+  encoder->perm()->axis(loco::FeatureAxis::Depth) = 1;
+
+  // FeatureEncode
+  auto enc = g->nodes()->create<loco::FeatureEncode>();
+  enc->input(pull);
+  enc->encoder(std::move(encoder));
+
+  locomotiv::NodeExecution::get().run(enc);
+
+  auto enc_data = locomotiv::annot_data(enc);
+  ASSERT_NE(enc_data, nullptr);
+  ASSERT_EQ(enc_data->dtype(), loco::DataType::S32);
+  ASSERT_EQ(*(enc_data->shape()), (Shape{N, H, W, C})); // locomotiv feature is NHWC
+  auto enc_buf = enc_data->as_s32_bufptr();
+  for (uint32_t n = 0; n < N; ++n)
+    for (uint32_t h = 0; h < H; ++h)
+      for (uint32_t w = 0; w < W; ++w)
+        for (uint32_t c = 0; c < C; ++c)
+          ASSERT_EQ(pull_buf.at(Index{n, c, h, w}), enc_buf->at(Index{n, h, w, c}));
+}
+
+TEST(NodeExecution_FeatureEncode, f32)
+{
+  const uint32_t N = 2;
+  const uint32_t H = 3;
+  const uint32_t W = 4;
+  const uint32_t C = 5;
+
+  auto g = loco::make_graph();
+
+  // Pull
+  auto pull = g->nodes()->create<loco::Pull>();
+  pull->dtype(loco::DataType::FLOAT32);
+
+  // Make and assign crazy "CHNW" data to pull node
+  auto pull_buf = make_buffer<float, LexicalLayout>(Shape{C, H, N, W});
+  float f = 1;
+  for (IndexEnumerator e{pull_buf.shape()}; e.valid(); e.advance())
+  {
+    pull_buf.at(e.current()) = f;
+    f = f + 1 / f; // Doesn't matter what it is
+  }
+  auto pull_data = locomotiv::make_data(pull_buf);
+  locomotiv::annot_data(pull, std::move(pull_data));
+
+  // Encoder to correctly read input tensor as CHNW
+  auto encoder = std::unique_ptr<loco::PermutingEncoder<loco::Domain::Feature>>(
+      new loco::PermutingEncoder<loco::Domain::Feature>);
+  encoder->perm()->axis(loco::FeatureAxis::Count) = 2;
+  encoder->perm()->axis(loco::FeatureAxis::Height) = 1;
+  encoder->perm()->axis(loco::FeatureAxis::Width) = 3;
+  encoder->perm()->axis(loco::FeatureAxis::Depth) = 0;
+
+  // FeatureEncode
+  auto enc = g->nodes()->create<loco::FeatureEncode>();
+  enc->input(pull);
+  enc->encoder(std::move(encoder));
+
+  locomotiv::NodeExecution::get().run(enc);
+
+  auto enc_data = locomotiv::annot_data(enc);
+  ASSERT_NE(enc_data, nullptr);
+  ASSERT_EQ(enc_data->dtype(), loco::DataType::FLOAT32);
+  ASSERT_EQ(*(enc_data->shape()), (Shape{N, H, W, C})); // locomotiv feature is NHWC
+  auto enc_buf = enc_data->as_f32_bufptr();
+  for (uint32_t n = 0; n < N; ++n)
+    for (uint32_t h = 0; h < H; ++h)
+      for (uint32_t w = 0; w < W; ++w)
+        for (uint32_t c = 0; c < C; ++c)
+          ASSERT_FLOAT_EQ(pull_buf.at(Index{c, h, n, w}), enc_buf->at(Index{n, h, w, c}));
+}