MKLDNNFcLayer.cpp 8.7 KB
Newer Older
T
tensor-tang 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

15
#include "MKLDNNFcLayer.h"
T
tensor-tang 已提交
16
#include "paddle/utils/Logging.h"
T
tensor-tang 已提交
17

T
tensor-tang 已提交
18 19 20
using namespace mkldnn;  // NOLINT
typedef memory::format format;

T
tensor-tang 已提交
21 22
namespace paddle {

23
REGISTER_LAYER(mkldnn_fc, MKLDNNFcLayer);
T
tensor-tang 已提交
24

25
bool MKLDNNFcLayer::init(const LayerMap& layerMap,
T
tensor-tang 已提交
26
                         const ParameterMap& parameterMap) {
27
  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
T
tensor-tang 已提交
28 29 30
    return false;
  }

31
  CHECK_EQ(inputLayers_.size(), 1UL) << "Only support one input layer yet";
T
tensor-tang 已提交
32 33 34 35 36 37 38
  CHECK_EQ(inputLayers_.size(), parameters_.size());
  CHECK(!parameters_[0]->isSparse()) << "Do not support sparse yet";

  // output size, cat not be changed
  oc_ = getSize();
  oh_ = 1;
  ow_ = 1;
39 40
  ih_ = 1;
  iw_ = 1;
T
tensor-tang 已提交
41 42 43 44 45 46 47 48 49 50 51

  // input size can not change in FC
  iLayerSize_ = inputLayers_[0]->getSize();
  CHECK_EQ(parameters_[0]->getSize(), iLayerSize_ * oc_);

  // create weight
  weight_ =
      std::unique_ptr<Weight>(new Weight(oc_, iLayerSize_, parameters_[0], 0));

  // create biases
  if (biasParameter_.get() != NULL) {
T
tensor-tang 已提交
52
    biases_ = std::unique_ptr<Weight>(new Weight(1, oc_, biasParameter_, 0));
T
tensor-tang 已提交
53 54 55 56
  }
  return true;
}

57
void MKLDNNFcLayer::convertWeightsFromPaddle() {
T
tensor-tang 已提交
58
  if (hasInitedWgt_) {
T
tensor-tang 已提交
59 60 61
    return;
  }

T
tensor-tang 已提交
62 63
  CHECK(wgtVal_) << "should have been initialized";
  auto targetDim = wgtVal_->getDims();
T
tensor-tang 已提交
64
  auto srcFmt = targetDim.size() == 2 ? format::io : format::ihwo;
T
tensor-tang 已提交
65
  wgtVal_->reorderDataFrom(wgtVal_, srcFmt, targetDim);
T
tensor-tang 已提交
66 67 68
  hasInitedWgt_ = true;
}

69
void MKLDNNFcLayer::convertWeightsToPaddle() {
T
tensor-tang 已提交
70 71
  CHECK(wgtVal_) << "should have been initialized";
  auto targetDim = wgtVal_->getDims();
T
tensor-tang 已提交
72
  auto dstFmt = targetDim.size() == 2 ? format::io : format::ihwo;
T
tensor-tang 已提交
73
  wgtVal_->reorderDataTo(wgtVal_, dstFmt, targetDim);
T
tensor-tang 已提交
74 75
}

76 77 78
void MKLDNNFcLayer::reshape(
    int& bs, int& ic, int& ih, int& iw, int oc, int& oh, int& ow) {
  reshapeInput(bs, ih, iw);
79

T
tensor-tang 已提交
80
  CHECK_EQ(iLayerSize_, inputLayers_[0]->getSize());
81 82 83
  ic = iLayerSize_ / (ih * iw);
  CHECK_EQ(size_t(ic * ih * iw), iLayerSize_) << "not divisible";
  CHECK_EQ(size_t(oc), getSize());
T
tensor-tang 已提交
84

85 86
  reshapeOutput(oh, ow);
  resizeOutput(bs, oc);
T
tensor-tang 已提交
87

88
  printSizeInfo();
T
tensor-tang 已提交
89 90
}

91
void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,
92 93 94 95
                             MKLDNNMatrixPtr& in,
                             MKLDNNMatrixPtr& wgt,
                             MKLDNNMatrixPtr& bias,
                             MKLDNNMatrixPtr& out) {
96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118
  resetFwdBuffers(in, wgt, bias, out);

  resetFwdPD(fwdPD_, in, wgt, bias, out);

  resetFwdPipeline(pipeline, fwdPD_, in, wgt, bias, out);
}

void MKLDNNFcLayer::resetBwd(std::vector<primitive>& pipeline,
                             MKLDNNMatrixPtr& in,
                             MKLDNNMatrixPtr& wgt,
                             MKLDNNMatrixPtr& bias,
                             MKLDNNMatrixPtr& out) {
  std::shared_ptr<fc_bwdWgt::primitive_desc> bwdWgtPD;
  std::shared_ptr<fc_bwdData::primitive_desc> bwdDataPD;

  resetBwdBuffers(in, wgt, bias, out);

  resetBwdWgtPD(bwdWgtPD, wgt, bias, out);

  resetBwdDataPD(bwdDataPD, in, out);

  resetBwdPipeline(pipeline, bwdWgtPD, bwdDataPD, in, wgt, bias, out);
}
T
tensor-tang 已提交
119

120 121 122 123 124 125 126 127 128 129 130 131
void MKLDNNFcLayer::updateWeights(const UpdateCallback& callback) {
  weight_->getParameterPtr()->incUpdate(callback);
  if (biases_ && biases_->getWGrad()) {
    biases_->getParameterPtr()->incUpdate(callback);
  }
}

void MKLDNNFcLayer::resetFwdBuffers(MKLDNNMatrixPtr& in,
                                    MKLDNNMatrixPtr& wgt,
                                    MKLDNNMatrixPtr& bias,
                                    MKLDNNMatrixPtr& out) {
  resetInValue(in);
132 133
  CHECK(in);
  in->downSpatial();
134

135 136 137
  auto outPD =
      MKLDNNMatrix::createPrimitiveDesc({bs_, oc_}, format::nc, engine_);
  resetOutValue(out, outPD);
138

T
tensor-tang 已提交
139
  format wgtFmt = format::oihw;
140
  if (in->getFormat() == format::nChw8c) {
T
tensor-tang 已提交
141
    wgtFmt = format::oIhw8i;
142
  } else if (in->getFormat() == format::nChw16c) {
T
tensor-tang 已提交
143 144
    wgtFmt = format::oIhw16i;
  }
145 146 147
  auto wgtPD =
      MKLDNNMatrix::createPrimitiveDesc({oc_, ic_, ih_, iw_}, wgtFmt, engine_);
  resetWithMatrix(wgt, weight_->getW(), wgtPD);
148
  wgt->downSpatial();
149

T
tensor-tang 已提交
150 151 152 153 154
  if (biases_ && biases_->getW()) {
    auto biasPD = MKLDNNMatrix::createPrimitiveDesc({oc_}, format::x, engine_);
    resetWithMatrix(bias, biases_->getW(), biasPD);
  } else {
    bias = nullptr;
155
  }
156
}
T
tensor-tang 已提交
157

158 159 160 161 162 163 164 165
void MKLDNNFcLayer::resetFwdPD(std::shared_ptr<fc_fwd::primitive_desc>& pd,
                               MKLDNNMatrixPtr in,
                               MKLDNNMatrixPtr wgt,
                               MKLDNNMatrixPtr bias,
                               MKLDNNMatrixPtr out) {
  CHECK(in);
  CHECK(wgt);
  CHECK(out);
T
tensor-tang 已提交
166
  prop_kind pk = prop_kind::forward;
167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187
  fc_fwd::desc fwdDesc = bias != nullptr ? fc_fwd::desc(pk,
                                                        in->getMemoryDesc(),
                                                        wgt->getMemoryDesc(),
                                                        bias->getMemoryDesc(),
                                                        out->getMemoryDesc())
                                         : fc_fwd::desc(pk,
                                                        in->getMemoryDesc(),
                                                        wgt->getMemoryDesc(),
                                                        out->getMemoryDesc());
  pd.reset(new fc_fwd::primitive_desc(fwdDesc, engine_));
}

void MKLDNNFcLayer::resetFwdPipeline(
    std::vector<primitive>& pipeline,
    std::shared_ptr<fc_fwd::primitive_desc>& pd,
    MKLDNNMatrixPtr& in,
    MKLDNNMatrixPtr& wgt,
    MKLDNNMatrixPtr& bias,
    MKLDNNMatrixPtr& out) {
  if (bias) {
    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *bias, *out));
T
tensor-tang 已提交
188
  } else {
189
    fwd_.reset(new fc_fwd(*pd, *in, *wgt, *out));
T
tensor-tang 已提交
190
  }
191
  pipeline.push_back(*fwd_);
T
tensor-tang 已提交
192 193
}

194 195 196 197
void MKLDNNFcLayer::resetBwdBuffers(MKLDNNMatrixPtr& in,
                                    MKLDNNMatrixPtr& wgt,
                                    MKLDNNMatrixPtr& bias,
                                    MKLDNNMatrixPtr& out) {
198 199 200
  CHECK(inVal_ && outVal_);
  resetOutGrad(out, outVal_->getPrimitiveDesc());
  resetInGrad(in, inVal_->getPrimitiveDesc());
201 202

  CHECK(wgtVal_);
203
  resetWithMatrix(wgt, weight_->getWGrad(), wgtVal_->getPrimitiveDesc());
204

T
tensor-tang 已提交
205 206 207 208
  if (biasVal_) {
    resetWithMatrix(bias, biases_->getWGrad(), biasVal_->getPrimitiveDesc());
  } else {
    bias = nullptr;
T
tensor-tang 已提交
209
  }
210
}
T
tensor-tang 已提交
211

212 213 214 215 216 217 218 219 220 221 222 223 224 225
void MKLDNNFcLayer::resetBwdWgtPD(
    std::shared_ptr<fc_bwdWgt::primitive_desc>& pd,
    MKLDNNMatrixPtr& wgt,
    MKLDNNMatrixPtr& bias,
    MKLDNNMatrixPtr& out) {
  CHECK(inVal_);
  fc_bwdWgt::desc bwdWgtDesc = bias ? fc_bwdWgt::desc(inVal_->getMemoryDesc(),
                                                      wgt->getMemoryDesc(),
                                                      bias->getMemoryDesc(),
                                                      out->getMemoryDesc())
                                    : fc_bwdWgt::desc(inVal_->getMemoryDesc(),
                                                      wgt->getMemoryDesc(),
                                                      out->getMemoryDesc());
  pd.reset(new fc_bwdWgt::primitive_desc(bwdWgtDesc, engine_, *fwdPD_));
T
tensor-tang 已提交
226 227
}

228 229 230 231 232 233 234 235 236 237 238 239
void MKLDNNFcLayer::resetBwdDataPD(
    std::shared_ptr<fc_bwdData::primitive_desc>& pd,
    MKLDNNMatrixPtr& in,
    MKLDNNMatrixPtr& out) {
  pd = nullptr;
  if (in == nullptr) {
    return;
  }
  CHECK(wgtVal_);
  fc_bwdData::desc bwdDataDesc = fc_bwdData::desc(
      in->getMemoryDesc(), wgtVal_->getMemoryDesc(), out->getMemoryDesc());
  pd.reset(new fc_bwdData::primitive_desc(bwdDataDesc, engine_, *fwdPD_));
T
tensor-tang 已提交
240 241
}

242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259
void MKLDNNFcLayer::resetBwdPipeline(
    std::vector<primitive>& pipeline,
    std::shared_ptr<fc_bwdWgt::primitive_desc>& bwdWgtPD,
    std::shared_ptr<fc_bwdData::primitive_desc>& bwdDataPD,
    MKLDNNMatrixPtr& in,
    MKLDNNMatrixPtr& wgt,
    MKLDNNMatrixPtr& bias,
    MKLDNNMatrixPtr& out) {
  CHECK(inVal_);
  if (bias) {
    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt, *bias));
  } else {
    bwdWgt_.reset(new fc_bwdWgt(*bwdWgtPD, *inVal_, *out, *wgt));
  }
  pipeline.push_back(*bwdWgt_);

  if (bwdDataPD == nullptr) {
    return;
T
tensor-tang 已提交
260
  }
261 262 263
  CHECK(wgtVal_) << "Should have weight memory";
  bwdData_.reset(new fc_bwdData(*bwdDataPD, *out, *wgtVal_, *in));
  pipeline.push_back(*bwdData_);
T
tensor-tang 已提交
264
}
265

T
tensor-tang 已提交
266
}  // namespace paddle