MKLDNNConvLayer.h 8.2 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
/* Copyright (c) 2017 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. */

#pragma once

#include "MKLDNNLayer.h"
#include "mkldnn.hpp"

namespace paddle {
21 22 23
typedef mkldnn::convolution_forward conv_fwd;
typedef mkldnn::convolution_backward_weights conv_bwdWgt;
typedef mkldnn::convolution_backward_data conv_bwdData;
T
tensor-tang 已提交
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42

/**
 * @brief A subclass of MKLDNNLayer conv layer.
 *
 * The config file api is mkldnn_conv
 */
class MKLDNNConvLayer : public MKLDNNLayer {
protected:
  // padding height and width
  int ph_, pw_;
  // stride height and width
  int sh_, sw_;
  // dilation height and width
  int dh_, dw_;
  // filter(kenerl) height and width
  int fh_, fw_;
  // group number
  int gp_;

43
  // in resetBwdData, the format of wgtValBwdData_ is different with wgtVal_
T
tensor-tang 已提交
44
  MKLDNNMatrixPtr wgtValBwdData_;
45
  // convert handle from wgtVal_ to wgtValBwdData_
T
tensor-tang 已提交
46 47
  std::shared_ptr<mkldnn::reorder> cvtWgtVal_;

48
  // save forward primitive_desc, which can be used backward
49
  std::shared_ptr<conv_fwd::primitive_desc> fwdPD_;
T
tensor-tang 已提交
50

51
  // MKLDNNMatrixPtr which should be created from CPU Device
T
tensor-tang 已提交
52 53 54 55
  MKLDNNMatrixPtr cpuInVal_;
  MKLDNNMatrixPtr cpuInGrad_;
  MKLDNNMatrixPtr cpuOutVal_;
  MKLDNNMatrixPtr cpuOutGrad_;
56
  // convert handle between CPU device and MKLDNN device
T
tensor-tang 已提交
57 58 59 60 61
  std::shared_ptr<mkldnn::reorder> cvtInVal_;
  std::shared_ptr<mkldnn::reorder> cvtInGrad_;
  std::shared_ptr<mkldnn::reorder> cvtOutVal_;
  std::shared_ptr<mkldnn::reorder> cvtOutGrad_;

62
  // whether the weight has been init
T
tensor-tang 已提交
63 64
  bool hasInitedWgt_;

65 66
  // true by default, which impact the calculation of output image size.
  // details can refer to mathUtil.h
T
tensor-tang 已提交
67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120
  bool caffeMode_;

  // weight and bias
  std::unique_ptr<Weight> weight_;
  std::unique_ptr<Weight> biases_;

public:
  explicit MKLDNNConvLayer(const LayerConfig& config)
      : MKLDNNLayer(config), hasInitedWgt_(false), caffeMode_(true) {}

  ~MKLDNNConvLayer() {}

  bool init(const LayerMap& layerMap,
            const ParameterMap& parameterMap) override;

  void reshape(
      int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) override;

  void resetFwd(std::vector<mkldnn::primitive>& pipeline,
                MKLDNNMatrixPtr& in,
                MKLDNNMatrixPtr& wgt,
                MKLDNNMatrixPtr& bias,
                MKLDNNMatrixPtr& out) override;

  void resetBwd(std::vector<mkldnn::primitive>& pipeline,
                MKLDNNMatrixPtr& in,
                MKLDNNMatrixPtr& wgt,
                MKLDNNMatrixPtr& bias,
                MKLDNNMatrixPtr& out) override;

  void updateInputData() override;

  void updateWeights(const UpdateCallback& callback) override;

  void convertWeightsFromPaddle() override;

  void convertWeightsToPaddle() override;

  void printSizeInfo() override {
    MKLDNNLayer::printSizeInfo();
    VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
                       << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
  }

  void printValueFormatFlow() override {
    if (cpuInVal_) {
      VLOG(MKLDNN_FMTS) << cpuInVal_->getFormat() << " >>>";
    }
    MKLDNNLayer::printValueFormatFlow();
    if (cpuOutVal_) {
      VLOG(MKLDNN_FMTS) << " >>> " << cpuOutVal_->getFormat();
    }
  }
121

T
tensor-tang 已提交
122 123 124 125 126 127 128 129 130 131
  void printGradFormatFlow() override {
    if (cpuInGrad_) {
      VLOG(MKLDNN_FMTS) << cpuInGrad_->getFormat() << " <<<";
    }
    MKLDNNLayer::printGradFormatFlow();
    if (cpuOutGrad_) {
      VLOG(MKLDNN_FMTS) << " <<< " << cpuOutGrad_->getFormat();
    }
  }

132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232
protected:
  /**
   * load the dims settings of this conv
   */
  void loadConvSettings(mkldnn::memory::dims& wgt,
                        mkldnn::memory::dims& bias,
                        mkldnn::memory::dims& stride,
                        mkldnn::memory::dims& dilation,
                        mkldnn::memory::dims& padL,
                        mkldnn::memory::dims& padR);

  /**
   * reset the forward primitive descriptor.
   */
  void resetFwdPD(std::shared_ptr<conv_fwd::primitive_desc>& pd);
  /**
   * reset the MKLDNNMatrix buffers used in forward.
   */
  void resetFwdBuffers(std::shared_ptr<conv_fwd::primitive_desc>& pd,
                       MKLDNNMatrixPtr& in,
                       MKLDNNMatrixPtr& wgt,
                       MKLDNNMatrixPtr& bias,
                       MKLDNNMatrixPtr& out);
  /**
   * reset the forward pipeline.
   */
  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                        std::shared_ptr<conv_fwd::primitive_desc>& pd,
                        MKLDNNMatrixPtr& in,
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);

  /**
   * reset MKLDNNMatrix of input value
   */
  void resetInValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
                    MKLDNNMatrixPtr& in);
  /**
   * reset MKLDNNMatrix of weight and bias value
   */
  void resetWgtBiasValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
                         MKLDNNMatrixPtr& wgt,
                         MKLDNNMatrixPtr& bias);
  /**
   * reset MKLDNNMatrix of output value
   */
  void resetOutValue(std::shared_ptr<conv_fwd::primitive_desc>& pd,
                     MKLDNNMatrixPtr& out);

  /**
   * reset the backward weight primitive descriptor.
   */
  void resetBwdWgtPD(std::shared_ptr<conv_bwdWgt::primitive_desc>& pd);
  /**
   * reset the backward data primitive descriptor.
   */
  void resetBwdDataPD(std::shared_ptr<conv_bwdData::primitive_desc>& pd);
  /**
   * reset the MKLDNNMatrix buffers used in backward.
   */
  void resetBwdBuffers(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
                       std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
                       MKLDNNMatrixPtr& in,
                       MKLDNNMatrixPtr& wgt,
                       MKLDNNMatrixPtr& bias,
                       MKLDNNMatrixPtr& out);
  /**
   * reset the backward pipeline.
   */
  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                        std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
                        std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
                        MKLDNNMatrixPtr& in,
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias,
                        MKLDNNMatrixPtr& out);

  /**
   * reset MKLDNNMatrix of output grad
   */
  void resetOutGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
                    MKLDNNMatrixPtr& out);
  /**
   * reset MKLDNNMatrix of weight and bias grad
   */
  void resetWgtBiasGrad(std::shared_ptr<conv_bwdWgt::primitive_desc>& wgtPD,
                        MKLDNNMatrixPtr& wgt,
                        MKLDNNMatrixPtr& bias);
  /**
   * reset MKLDNNMatrix of input grad
   */
  void resetInGrad(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
                   MKLDNNMatrixPtr& in);
  /**
   * reset MKLDNNMatrix of weight value for backward data
   * since the primitive_desc would be different with wgtVal_
   */
  void resetWgtValBwdData(std::shared_ptr<conv_bwdData::primitive_desc>& dataPD,
                          MKLDNNMatrixPtr& wgt);

T
tensor-tang 已提交
233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253
  /**
   * get padding_r according to
   * https://github.com/01org/mkl-dnn/blob/master/tests/gtests/
   * test_convolution_forward_common.hpp
   * @note: mkldnn dilation start from 0 while paddle start from 1
   */
  mkldnn::memory::dims getPaddingR() const {
    mkldnn::memory::dims padR = {ph_, pw_};
    for (int i = 0; i < 2; ++i) {
      if ((ih_ - ((fh_ - 1) * dh_ + 1) + ph_ + padR[0]) / sh_ + 1 != oh_) {
        ++padR[0];
      }
      if ((iw_ - ((fw_ - 1) * dw_ + 1) + pw_ + padR[1]) / sw_ + 1 != ow_) {
        ++padR[1];
      }
    }
    return padR;
  }
};

}  // namespace paddle