MKLDNNLayer.h 4.9 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
/* 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 <vector>
#include "Layer.h"
19
#include "MKLDNNBase.h"
T
tensor-tang 已提交
20
#include "mkldnn.hpp"
T
tensor-tang 已提交
21
#include "paddle/math/MKLDNNMatrix.h"
T
tensor-tang 已提交
22

T
tensor-tang 已提交
23 24
DECLARE_bool(use_mkldnn);

T
tensor-tang 已提交
25 26
namespace paddle {

27 28
class MKLDNNLayer;
typedef std::shared_ptr<MKLDNNLayer> MKLDNNLayerPtr;
T
tensor-tang 已提交
29 30

/**
31
 * @brief Base class of MKLDNNlayer.
T
tensor-tang 已提交
32 33
 *
 */
34
class MKLDNNLayer : public Layer {
T
tensor-tang 已提交
35 36 37 38 39 40 41 42
protected:
  // batch size
  int bs_;
  // input image channel, height and width
  int ic_, ih_, iw_;
  // output image channel, height and width
  int oc_, oh_, ow_;

T
tensor-tang 已提交
43 44 45
  // backward also need reset after reset forward handle
  bool needResetBwd_;

T
tensor-tang 已提交
46 47
  // mkldnn engine, stream and primivtives
  mkldnn::engine engine_;
48
  std::shared_ptr<MKLDNNStream> stream_;
T
tensor-tang 已提交
49
  std::shared_ptr<mkldnn::primitive> fwd_;
T
tensor-tang 已提交
50 51
  std::shared_ptr<mkldnn::primitive> bwdWgt_;
  std::shared_ptr<mkldnn::primitive> bwdData_;
T
tensor-tang 已提交
52 53 54
  std::vector<mkldnn::primitive> pipelineFwd_;
  std::vector<mkldnn::primitive> pipelineBwd_;

T
tensor-tang 已提交
55
  // MKLDNNMatrixPtr
T
tensor-tang 已提交
56
  MKLDNNMatrixPtr inVal_;
T
tensor-tang 已提交
57
  MKLDNNMatrixPtr inGrad_;
T
tensor-tang 已提交
58
  MKLDNNMatrixPtr outVal_;
T
tensor-tang 已提交
59
  MKLDNNMatrixPtr outGrad_;
T
tensor-tang 已提交
60
  MKLDNNMatrixPtr wgtVal_;
T
tensor-tang 已提交
61
  MKLDNNMatrixPtr wgtGrad_;
T
tensor-tang 已提交
62
  MKLDNNMatrixPtr biasVal_;
T
tensor-tang 已提交
63
  MKLDNNMatrixPtr biasGrad_;
T
tensor-tang 已提交
64

T
tensor-tang 已提交
65
public:
66
  explicit MKLDNNLayer(const LayerConfig& config)
T
tensor-tang 已提交
67 68 69 70 71 72 73 74
      : Layer(config),
        bs_(0),
        ic_(0),
        ih_(0),
        iw_(0),
        oc_(0),
        oh_(0),
        ow_(0),
T
tensor-tang 已提交
75
        needResetBwd_(true),
T
tensor-tang 已提交
76
        engine_(mkldnn::engine::cpu, 0),
T
tensor-tang 已提交
77 78 79 80
        stream_(nullptr),
        fwd_(nullptr),
        bwdWgt_(nullptr),
        bwdData_(nullptr) {}
T
tensor-tang 已提交
81

82
  ~MKLDNNLayer() {}
T
tensor-tang 已提交
83

T
tensor-tang 已提交
84 85
  virtual bool init(const LayerMap& layerMap,
                    const ParameterMap& parameterMap) {
T
tensor-tang 已提交
86 87 88 89 90 91 92 93 94 95 96 97
    CHECK(FLAGS_use_mkldnn) << "MkldnnLayers only support use_mkldnn."
                            << "Please set WITH_MKLDNN=ON "
                            << "and set use_mkldnn=True";
    if (useGpu_ == true) {
      LOG(WARNING) << "Do not support GPU yet, will change to useGpu = false";
      useGpu_ = false;
    }

    // set device id before Layer::init
    setDevice(MKLDNN_DEVICE);
    // change param device to MKLDNN device
    setParamsDevice(MKLDNN_DEVICE, parameterMap);
T
tensor-tang 已提交
98 99 100 101
    if (!Layer::init(layerMap, parameterMap)) {
      return false;
    }

102 103
    stream_.reset(new MKLDNNStream());
    engine_ = CPUEngine::Instance().getEngine();
T
tensor-tang 已提交
104 105
    return true;
  }
T
tensor-tang 已提交
106 107 108 109 110

  /**
   * convert weight from paddle format to mkldnn format
   * weight_ will be override
   */
111
  virtual void convertWeightsFromPaddle() {}
T
tensor-tang 已提交
112 113 114 115 116

  /**
   * convert mkldnn weight to paddle format
   * weight_ will be override
   */
117
  virtual void convertWeightsToPaddle() {}
T
tensor-tang 已提交
118

T
tensor-tang 已提交
119 120 121 122 123 124 125 126
  /**
   * print info about sizes
   */
  virtual void printSizeInfo() {
    VLOG(MKLDNN_SIZES) << getName() << ": bs: " << bs_ << ", ic: " << ic_
                       << ", ih: " << ih_ << ", iw: " << iw_ << ", oc: " << oc_
                       << ", oh: " << oh_ << ", ow: " << ow_;
  }
T
tensor-tang 已提交
127

T
tensor-tang 已提交
128 129 130 131 132
  // TODO(TJ): move to MkldnnMatrix
  // create memory desc
  inline mkldnn::memory::desc createMD(
      mkldnn::memory::dims dims,
      mkldnn::memory::format fmt,
T
tensor-tang 已提交
133 134 135 136
      mkldnn::memory::data_type type = mkldnn::memory::data_type::f32) {
    // TODO(TJ): isFmtSuppoted(fmt)
    return mkldnn::memory::desc(dims, type, fmt);
  }
T
tensor-tang 已提交
137 138 139 140 141 142 143 144

  void resetMKLDNNOutput(size_t height, size_t width) {
    Layer::resetOutput(height, width);
    // get valu and grad, use mkldnn matrix instaed
    // output_.value;
  }

protected:
T
tensor-tang 已提交
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
  /**
   * Set deviceId of this layer.
   */
  void setDevice(int id) { deviceId_ = id; }

  /**
   * Set deviceId of the params used in this layer.
   */
  void setParamsDevice(int id, const ParameterMap& parameterMap) {
    for (auto& inputConfig : config_.inputs()) {
      if (inputConfig.has_input_parameter_name()) {
        ParameterPtr parameter;
        std::string name = inputConfig.input_parameter_name();
        CHECK(mapGet(name, parameterMap, &parameter))
            << "Cannot find input parameter " << name << " for layer "
            << getName();
        parameter->setDevice(id);
      }
    }
    if (config_.has_bias_parameter_name()) {
      ParameterPtr parameter;
      std::string name = config_.bias_parameter_name();
      CHECK(mapGet(name, parameterMap, &parameter))
          << "Cannot find bias parameter " << name << " for layer "
          << getName();
      parameter->setDevice(id);
    }
T
tensor-tang 已提交
172
  }
T
tensor-tang 已提交
173 174 175
};

}  // namespace paddle