MKLDNNAddtoLayer.h 4.1 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 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36
/* 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 {

/**
 * @brief A subclass of MKLDNNLayer Addto layer.
 *
 * The config file api is mkldnn_addto
 */
class MKLDNNAddtoLayer : public MKLDNNLayer {
protected:
  std::vector<MKLDNNMatrixPtr> inVals_;
  std::vector<MKLDNNMatrixPtr> inGrads_;

  // layer size == ic * ih * iw == oc * oh *ow, and can not be changed
  size_t layerSize_;

  std::unique_ptr<Weight> biases_;

T
tensor-tang 已提交
37 38 39 40 41 42 43
  // buffers for adding bias
  std::vector<MKLDNNMatrixPtr> vals_;
  std::vector<MKLDNNMatrixPtr> grads_;
  // primitives for adding bias
  std::vector<std::shared_ptr<mkldnn::primitive>> fwdBias_;
  std::shared_ptr<mkldnn::primitive> bwdBias_;

T
tensor-tang 已提交
44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 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
public:
  explicit MKLDNNAddtoLayer(const LayerConfig& config) : MKLDNNLayer(config) {}

  ~MKLDNNAddtoLayer() {}

  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 updateWeights(const UpdateCallback& callback) override;

  void printValueFormat() override {
    for (size_t i = 0; i < inVals_.size(); ++i) {
      VLOG(MKLDNN_FMTS) << i << " input: " << inVals_[i]->getFormat() << " >>>";
    }
    if (outVal_) {
      VLOG(MKLDNN_FMTS) << outVal_->getFormat() << " >>> ";
    }
    if (extOutVal_) {
      VLOG(MKLDNN_FMTS) << extOutVal_->getFormat();
    }
  }

  void printGradFormat() override {
    if (extOutGrad_) {
      VLOG(MKLDNN_FMTS) << extOutGrad_->getFormat();
    }
    if (outGrad_) {
      VLOG(MKLDNN_FMTS) << outGrad_->getFormat() << " <<< ";
    }
    for (size_t i = 0; i < inGrads_.size(); ++i) {
      VLOG(MKLDNN_FMTS) << i << " input: " << inGrads_[i]->getFormat() << "<<<";
    }
  }

protected:
  /**
   * Forward functions: reset buffers(inputs, output, bias),
   *                    reset primitive descriptor,
   *                    reset pipeline.
   */
  void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
100
                       MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
101 102
                       MKLDNNMatrixPtr& out);
  void resetFwdPD(std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
T
tensor-tang 已提交
103
                  std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
T
tensor-tang 已提交
104
                  std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
105
                  MKLDNNMatrixPtr bias,
T
tensor-tang 已提交
106 107 108
                  MKLDNNMatrixPtr out);
  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                        std::shared_ptr<mkldnn::sum::primitive_desc>& pd,
T
tensor-tang 已提交
109
                        std::shared_ptr<mkldnn::sum::primitive_desc>& biasPD,
T
tensor-tang 已提交
110
                        std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
111
                        MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
112 113 114 115 116 117
                        MKLDNNMatrixPtr& out);

  /**
   * Backward functions: reset buffers(inputs, output, bias)
   */
  void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
118
                       MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
119
                       MKLDNNMatrixPtr& out);
T
tensor-tang 已提交
120 121 122 123 124 125 126 127

  /**
   * prepare for bias
   */
  void prepareBias(MKLDNNMatrixPtr& bias,
                   const MatrixPtr& biasMat,
                   const MKLDNNMatrixPtr& out,
                   std::vector<MKLDNNMatrixPtr>& outs);
T
tensor-tang 已提交
128 129 130
};

}  // namespace paddle