MKLDNNConcatLayer.h 3.9 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 37 38 39 40 41 42 43 44 45 46 47 48 49
/* 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 Concatenate layer.
 *
 * The config file api is mkldnn_concat
 */
class MKLDNNConcatLayer : public MKLDNNLayer {
protected:
  std::vector<MKLDNNMatrixPtr> inVals_;
  std::vector<MKLDNNMatrixPtr> inGrads_;
  std::vector<std::shared_ptr<mkldnn::primitive>> bwds_;
  // input channel numbers
  std::vector<int> channels_;

  // concat_dimension in MKLDNN
  // if axis_ == 0, concat batchsize
  // if axis_ == 1, concat channel (default)
  int axis_;

public:
  explicit MKLDNNConcatLayer(const LayerConfig& config)
      : MKLDNNLayer(config), axis_(1) {}

  ~MKLDNNConcatLayer() {}

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

  void reshape(
50
      int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) override;
T
tensor-tang 已提交
51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72

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

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

  void printSizeInfo() override {
    CHECK_EQ(channels_.size(), inputLayers_.size());
    for (size_t i = 0; i < channels_.size(); ++i) {
      VLOG(MKLDNN_SIZES) << "Input " << i << ", " << inputLayers_[i]->getName()
                         << ": " << bs_ << ", " << channels_[i] << ", " << ih_
                         << ", " << iw_;
    }
    VLOG(MKLDNN_SIZES) << "Output: " << bs_ << ", " << oc_ << ", " << oh_
                       << ", " << ow_;
  }

  void printValueFormat() override {
    for (size_t i = 0; i < inVals_.size(); ++i) {
T
tensor-tang 已提交
73 74
      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
                        << ": " << inVals_[i]->getFormat() << " >>>";
T
tensor-tang 已提交
75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
    }
    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) {
T
tensor-tang 已提交
92 93
      VLOG(MKLDNN_FMTS) << "Input " << i << ", " << inputLayers_[i]->getName()
                        << ": " << inGrads_[i]->getFormat() << "<<<";
T
tensor-tang 已提交
94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
    }
  }

protected:
  void resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
                       MKLDNNMatrixPtr& out);
  void resetFwdPD(std::shared_ptr<mkldnn::concat::primitive_desc>& pd,
                  std::vector<MKLDNNMatrixPtr>& inputs,
                  MKLDNNMatrixPtr out);
  void resetFwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                        std::shared_ptr<mkldnn::concat::primitive_desc>& pd,
                        std::vector<MKLDNNMatrixPtr>& inputs,
                        MKLDNNMatrixPtr& out);
  void resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
                       MKLDNNMatrixPtr& out);
  void resetBwdPipeline(std::vector<mkldnn::primitive>& pipeline,
                        std::vector<std::shared_ptr<mkldnn::primitive>>& prims,
                        std::vector<MKLDNNMatrixPtr>& inputs,
                        MKLDNNMatrixPtr& out);
};

}  // namespace paddle