NeonDepthwiseConv.cpp

/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.

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 "NeonDepthwiseConv.h"
#include "paddle/function/ConvOp.h"
#include "paddle/function/Im2Col.h"

namespace paddle {

namespace neon {

#if defined(__ARM_NEON__) || defined(__ARM_NEON)

template <DeviceType Device>
class NeonDepthwiseConvFunction : public ConvFunctionBase {
public:
  void init(const FuncConfig& config) override {
    ConvFunctionBase::init(config);
  }

  void check(const BufferArgs& inputs, const BufferArgs& outputs) override {
    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();
    checkShape(input, filter, output);
  }

  void calc(const BufferArgs& inputs, const BufferArgs& outputs) override {
    CHECK_EQ(numInputs_, inputs.size());
    CHECK_EQ(numOutputs_, outputs.size());
    check(inputs, outputs);

    const TensorShape& input = inputs[0].shape();
    const TensorShape& filter = inputs[1].shape();
    const TensorShape& output = outputs[0].shape();

    size_t batchSize = input[0];
    size_t inputChannels = input[1];
    size_t inputHeight = input[2];
    size_t inputWidth = input[3];
    size_t filterHeight = getFilterHeight(filter);
    size_t filterWidth = getFilterWidth(filter);
    size_t outputChannels = output[1];
    size_t outputHeight = output[2];
    size_t outputWidth = output[3];
    size_t filterMultiplier = outputChannels / groups_;
    CHECK_EQ(inputChannels, groups_);

    // only support strideH() == strideW() and filterHeight == filterWidth.
    CHECK_EQ(strideH(), strideW());
    CHECK_EQ(filterHeight, filterWidth);

    float* inputData = inputs[0].data<float>();
    float* filterData = inputs[1].data<float>();
    float* outputData = outputs[0].data<float>();

    // padding the input
    float* inputPadding = inputData;
    if (paddingH() > 0 || paddingW() > 0) {
      int newSize = batchSize * inputChannels * (inputHeight + 2 * paddingH()) *
                    (inputWidth + 2 * paddingW());
      resizeBuffer<Device>(newSize);
      inputPadding = reinterpret_cast<float*>(memory_->getBuf());
      Padding<float>::run(inputData,
                          inputPadding,
                          batchSize * inputChannels,
                          inputHeight,
                          inputWidth,
                          paddingH(),
                          paddingW());

      // height and width of padding data
      inputHeight += 2 * paddingH();
      inputWidth += 2 * paddingW();
    }

    std::function<void(
        const float*, const float*, int, int, int, int, int, int, float*)>
        DepthWiseConv;

    if (filterWidth == 3 && strideW() == 1) {
      DepthWiseConv = DepthwiseConvKernel<3, 1>::run;
    } else if (filterWidth == 3 && strideW() == 2) {
      DepthWiseConv = DepthwiseConvKernel<3, 2>::run;
    } else if (filterWidth == 4 && strideW() == 1) {
      DepthWiseConv = DepthwiseConvKernel<4, 1>::run;
    } else if (filterWidth == 4 && strideW() == 2) {
      DepthWiseConv = DepthwiseConvKernel<4, 2>::run;
    } else {
      LOG(FATAL) << "Not supported";
    }

    for (size_t i = 0; i < batchSize; i++) {
      DepthWiseConv(inputPadding,
                    filterData,
                    inputHeight,
                    inputWidth,
                    outputChannels,
                    outputHeight,
                    outputWidth,
                    filterMultiplier,
                    outputData);
      inputPadding += inputChannels * inputHeight * inputWidth;
      outputData += outputChannels * outputHeight * outputWidth;
    }
  }
};

REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction);

#endif

}  // namespace neon
}  // namespace paddle