NeonDepthwiseConv.cpp 4.2 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "NeonDepthwiseConv.h"
16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44
#include "paddle/function/ConvOp.h"

namespace paddle {

#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();

H
hedaoyuan 已提交
45 46 47 48 49 50 51 52 53 54
    int batchSize = input[0];
    int inputChannels = input[1];
    int inputHeight = input[2];
    int inputWidth = input[3];
    int filterHeight = getFilterHeight(filter);
    int filterWidth = getFilterWidth(filter);
    int outputChannels = output[1];
    int outputHeight = output[2];
    int outputWidth = output[3];
    int filterMultiplier = outputChannels / groups_;
55 56
    CHECK_EQ(inputChannels, groups_);

57
    // only support strideH() == strideW() and filterHeight == filterWidth.
58 59 60 61 62 63 64 65 66
    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;
67 68
    int padInputHeight = inputHeight + 2 * paddingH();
    int padInputWidth = inputWidth + 2 * paddingW();
69
    if (paddingH() > 0 || paddingW() > 0) {
70
      int newSize = batchSize * inputChannels * padInputHeight * padInputWidth;
71 72
      resizeBuffer<Device>(newSize);
      inputPadding = reinterpret_cast<float*>(memory_->getBuf());
73 74 75 76 77
      neon::Padding<float>::run(inputData,
                                inputPadding,
                                batchSize * inputChannels,
                                inputHeight,
                                inputWidth,
78 79
                                padInputHeight,
                                padInputWidth);
80 81
    }

82 83 84 85 86
    std::function<void(
        const float*, const float*, int, int, int, int, int, int, float*)>
        DepthWiseConv;

    if (filterWidth == 3 && strideW() == 1) {
H
hedaoyuan 已提交
87
      DepthWiseConv = neon::DepthwiseConvKernel<3, 1>::run;
88
    } else if (filterWidth == 3 && strideW() == 2) {
H
hedaoyuan 已提交
89
      DepthWiseConv = neon::DepthwiseConvKernel<3, 2>::run;
90
    } else if (filterWidth == 4 && strideW() == 1) {
H
hedaoyuan 已提交
91
      DepthWiseConv = neon::DepthwiseConvKernel<4, 1>::run;
92
    } else if (filterWidth == 4 && strideW() == 2) {
H
hedaoyuan 已提交
93
      DepthWiseConv = neon::DepthwiseConvKernel<4, 2>::run;
94 95 96
    } else {
      LOG(FATAL) << "Not supported";
    }
97

H
hedaoyuan 已提交
98
    for (int i = 0; i < batchSize; i++) {
99 100
      DepthWiseConv(inputPadding,
                    filterData,
101 102
                    padInputHeight,
                    padInputWidth,
103 104 105 106 107
                    outputChannels,
                    outputHeight,
                    outputWidth,
                    filterMultiplier,
                    outputData);
108
      inputPadding += inputChannels * padInputHeight * padInputWidth;
109 110 111 112 113
      outputData += outputChannels * outputHeight * outputWidth;
    }
  }
};

H
hedaoyuan 已提交
114
#ifndef PADDLE_TYPE_DOUBLE
115
REGISTER_TYPED_FUNC(NeonDepthwiseConv, CPU, NeonDepthwiseConvFunction);
H
hedaoyuan 已提交
116
#endif
117 118 119 120

#endif

}  // namespace paddle
反馈
建议
客服 返回
顶部