MKLDNNAddtoLayer.cpp 7.3 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
/* 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. */

#include "MKLDNNAddtoLayer.h"

using namespace mkldnn;  // NOLINT

namespace paddle {

REGISTER_LAYER(mkldnn_addto, MKLDNNAddtoLayer);

bool MKLDNNAddtoLayer::init(const LayerMap& layerMap,
                            const ParameterMap& parameterMap) {
  if (!MKLDNNLayer::init(layerMap, parameterMap)) {
    return false;
  }

  layerSize_ = getSize();
  for (size_t i = 0; i < inputLayers_.size(); i++) {
    CHECK_EQ(layerSize_, inputLayers_[i]->getSize()) << "input size must equal";
  }
  if (biasParameter_.get() != NULL) {
    biases_ =
        std::unique_ptr<Weight>(new Weight(1, layerSize_, biasParameter_, 0));
  }
  return true;
}

void MKLDNNAddtoLayer::reshape(
41
    int& bs, int& ic, int& ih, int& iw, int& oc, int& oh, int& ow) {
T
tensor-tang 已提交
42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61
  CHECK_EQ(layerSize_, getSize()) << "this layer size can not be changed";
  reshapeInput(bs, ih, iw);
  ic = inputLayers_[0]->getSize() / ih / iw;
  CHECK_EQ((size_t)ic * ih * iw, inputLayers_[0]->getSize());
  CHECK_EQ(inputElemenCnt_, (size_t)bs * ic * ih * iw);
  for (size_t i = 0; i < inputLayers_.size(); i++) {
    CHECK_EQ(int64_t(bs), inputLayers_[i]->getOutput().getBatchSize());
    CHECK_EQ(layerSize_, inputLayers_[i]->getSize());
  }

  oc = ic;
  oh = ih;
  ow = iw;
  reshapeOutput(oh, ow);
  resizeOutput(bs, oc * oh * ow);
}

void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,
                                MKLDNNMatrixPtr& in,
                                MKLDNNMatrixPtr& out) {
62
  resetFwdBuffers(inVals_, biasVal_, out);
T
tensor-tang 已提交
63 64 65
  in = inVals_[0];

  std::shared_ptr<sum::primitive_desc> fwdPD;
T
tensor-tang 已提交
66
  std::shared_ptr<sum::primitive_desc> biasPD;
67
  resetFwdPD(fwdPD, biasPD, inVals_, biasVal_, out);
T
tensor-tang 已提交
68

69
  resetFwdPipeline(pipeline, fwdPD, biasPD, inVals_, biasVal_, out);
T
tensor-tang 已提交
70 71 72 73 74
}

void MKLDNNAddtoLayer::resetBwd(std::vector<primitive>& pipeline,
                                MKLDNNMatrixPtr& in,
                                MKLDNNMatrixPtr& out) {
75
  resetBwdBuffers(inGrads_, biasGrad_, out);
T
tensor-tang 已提交
76 77 78 79 80 81 82 83 84
  in = inGrads_[0];

  // backward only need share output grad to input grad
  for (size_t i = 0; i < inGrads_.size(); i++) {
    if (inGrads_[i] != nullptr) {
      inGrads_[i] = out;
      inputLayers_[i]->getOutputGrad()->setData(inGrads_[i]->getData());
    }
  }
T
tensor-tang 已提交
85 86 87

  // backward bias
  bwdBias_ = nullptr;
88
  if (biasGrad_) {
T
tensor-tang 已提交
89
    std::vector<float> scales(bs_, 1.0);
90 91 92 93
    std::vector<memory::primitive_desc> srcPDs(bs_,
                                               biasGrad_->getPrimitiveDesc());
    auto biasPD =
        sum::primitive_desc(biasGrad_->getMemoryDesc(), scales, srcPDs);
T
tensor-tang 已提交
94 95 96 97
    std::vector<primitive::at> srcs;
    for (size_t i = 0; i < grads_.size(); ++i) {
      srcs.push_back(*(grads_[i]));
    }
98
    bwdBias_.reset(new sum(biasPD, srcs, *biasGrad_));
T
tensor-tang 已提交
99 100
    pipeline.push_back(*bwdBias_);
  }
T
tensor-tang 已提交
101 102 103 104 105 106 107 108
}

void MKLDNNAddtoLayer::updateWeights(const UpdateCallback& callback) {
  if (biases_ && biases_->getWGrad()) {
    biases_->getParameterPtr()->incUpdate(callback);
  }
}

T
tensor-tang 已提交
109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
void MKLDNNAddtoLayer::prepareBias(MKLDNNMatrixPtr& bias,
                                   const MatrixPtr& biasMat,
                                   const MKLDNNMatrixPtr& out,
                                   std::vector<MKLDNNMatrixPtr>& outs) {
  auto pd = MKLDNNMatrix::createPrimitiveDesc(
      {(int)layerSize_}, memory::format::x, engine_);
  bias = MKLDNNMatrix::create(pd, biasMat);
  outs.clear();
  real* data = out->getData();
  CHECK_EQ(bs_ * layerSize_, out->getElementCnt());
  for (int i = 0; i < bs_; ++i) {
    MatrixPtr tmp =
        Matrix::create(data + i * layerSize_, 1, layerSize_, false, false);
    outs.push_back(MKLDNNMatrix::create(bias->getPrimitiveDesc(), tmp));
  }
}

T
tensor-tang 已提交
126
void MKLDNNAddtoLayer::resetFwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
127
                                       MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
128 129 130 131 132 133 134 135 136 137 138 139
                                       MKLDNNMatrixPtr& out) {
  inputs.resize(inputLayers_.size());
  for (size_t i = 0; i < inputs.size(); i++) {
    resetInValue(inputs[i], nullptr, i);
    CHECK(inputs[i]);
    inputs[i]->downSpatial();
  }
  for (size_t i = 1; i < inputs.size(); i++) {
    CHECK_PRIMITIVE_DESC_EQ(inputs[i], inputs[0]->getPrimitiveDesc());
  }

  resetOutValue(out, inputs[0]->getPrimitiveDesc());
T
tensor-tang 已提交
140 141 142 143 144 145

  if (biases_ && biases_->getW()) {
    prepareBias(bias, biases_->getW(), out, vals_);
  } else {
    bias = nullptr;
  }
T
tensor-tang 已提交
146 147 148
}

void MKLDNNAddtoLayer::resetFwdPD(std::shared_ptr<sum::primitive_desc>& pd,
T
tensor-tang 已提交
149
                                  std::shared_ptr<sum::primitive_desc>& biasPD,
T
tensor-tang 已提交
150
                                  std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
151
                                  MKLDNNMatrixPtr bias,
T
tensor-tang 已提交
152
                                  MKLDNNMatrixPtr out) {
T
tensor-tang 已提交
153
  std::vector<float> scales(inputs.size(), 1.0);
T
tensor-tang 已提交
154 155 156 157 158 159 160
  std::vector<memory::primitive_desc> srcPDs;
  for (size_t i = 0; i < inputs.size(); i++) {
    srcPDs.push_back(inputs[i]->getPrimitiveDesc());
  }
  CHECK(out);
  pd.reset(new sum::primitive_desc(out->getMemoryDesc(), scales, srcPDs));
  CHECK_PRIMITIVE_DESC_EQ(out, pd->dst_primitive_desc());
T
tensor-tang 已提交
161 162 163

  biasPD = nullptr;
  if (bias) {
T
tensor-tang 已提交
164
    std::vector<float> scales(2, 1.0);
T
tensor-tang 已提交
165 166 167 168 169
    std::vector<memory::primitive_desc> srcPDs(2, bias->getPrimitiveDesc());
    biasPD.reset(
        new sum::primitive_desc(bias->getMemoryDesc(), scales, srcPDs));
    CHECK_PRIMITIVE_DESC_EQ(bias, biasPD->dst_primitive_desc());
  }
T
tensor-tang 已提交
170 171 172 173 174
}

void MKLDNNAddtoLayer::resetFwdPipeline(
    std::vector<primitive>& pipeline,
    std::shared_ptr<sum::primitive_desc>& pd,
T
tensor-tang 已提交
175
    std::shared_ptr<sum::primitive_desc>& biasPD,
T
tensor-tang 已提交
176
    std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
177
    MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
178 179 180 181 182 183 184
    MKLDNNMatrixPtr& out) {
  std::vector<primitive::at> srcs;
  for (size_t i = 0; i < inputs.size(); i++) {
    srcs.push_back(*(inputs[i]));
  }
  fwd_.reset(new sum(*pd, srcs, *out));
  pipeline.push_back(*fwd_);
T
tensor-tang 已提交
185 186 187 188 189 190 191 192 193 194 195 196 197

  fwdBias_.clear();
  if (biasPD == nullptr || bias == nullptr) {
    return;
  }
  fwdBias_.resize(vals_.size());
  for (size_t i = 0; i < vals_.size(); ++i) {
    std::vector<primitive::at> srcs;
    srcs.push_back(*(vals_[i]));
    srcs.push_back(*bias);
    fwdBias_[i].reset(new sum(*biasPD, srcs, *vals_[i]));
    pipeline.push_back(*fwdBias_[i]);
  }
T
tensor-tang 已提交
198 199 200
}

void MKLDNNAddtoLayer::resetBwdBuffers(std::vector<MKLDNNMatrixPtr>& inputs,
T
tensor-tang 已提交
201
                                       MKLDNNMatrixPtr& bias,
T
tensor-tang 已提交
202 203 204 205 206 207 208 209 210 211
                                       MKLDNNMatrixPtr& out) {
  CHECK(outVal_);
  resetOutGrad(out, outVal_->getPrimitiveDesc());
  CHECK(out);

  inputs.resize(inputLayers_.size());
  for (size_t i = 0; i < inputs.size(); i++) {
    resetInGrad(inputs[i], inVal_->getPrimitiveDesc(), i);
    CHECK_PRIMITIVE_DESC_EQ(inputs[i], out->getPrimitiveDesc());
  }
T
tensor-tang 已提交
212 213 214 215 216 217

  if (biases_ && biases_->getWGrad()) {
    prepareBias(bias, biases_->getWGrad(), out, grads_);
  } else {
    bias = nullptr;
  }
T
tensor-tang 已提交
218 219 220
}

}  // namespace paddle