MKLDNNMatrix.cpp 5.3 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
/* 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 "MKLDNNMatrix.h"

using namespace mkldnn;  // NOLINT

T
tensor-tang 已提交
19 20
namespace paddle {

21 22 23 24
MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m, memory::primitive_desc pd) {
  memory::desc md = pd.desc();
  size_t ndims = md.data.ndims;
  int* dims = md.data.dims;
T
tensor-tang 已提交
25
  CHECK(ndims > 0) << "Input dims should not be empty";
26
  size_t cnts = 1;
T
tensor-tang 已提交
27
  for (size_t i = 0; i < ndims; ++i) {
28
    cnts *= dims[i];
T
tensor-tang 已提交
29 30
  }

31 32 33 34 35 36
  if (m == nullptr) {
    size_t height = dims[0];
    size_t width = cnts / dims[0];
    m = Matrix::create(height, width, false, false);
  }
  CHECK(m) << " Matrix should not be empty";
37

38 39
  CpuMatrixPtr cpuMatrix = std::dynamic_pointer_cast<CpuMatrix>(m);
  CHECK(cpuMatrix) << "Only support create from CPU matrix yet";
40 41
  CHECK_EQ(cpuMatrix->getElementCnt(), cnts) << "Count size does not match";
  return std::make_shared<MKLDNNMatrix>(cpuMatrix, pd);
42
}
T
tensor-tang 已提交
43

44 45 46 47 48
MKLDNNMatrixPtr MKLDNNMatrix::create(MatrixPtr m,
                                     memory::dims dims,
                                     memory::format fmt,
                                     engine& eg,
                                     mkldnn::memory::data_type dtype) {
T
refine  
tensor-tang 已提交
49
  return create(m, memory::primitive_desc(memory::desc(dims, dtype, fmt), eg));
50 51
}

52 53 54
std::shared_ptr<reorder> MKLDNNMatrix::createReorder(const MKLDNNMatrixPtr& src,
                                                     const MKLDNNMatrixPtr& dst,
                                                     bool checkData) {
55
  if (src == dst || src->getPrimitiveDesc() == dst->getPrimitiveDesc()) {
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72
    return nullptr;
  }

  if (checkData && (src->getData() == dst->getData())) {
    LOG(FATAL) << "can not create reorder with inplace data";
    return nullptr;
  }

  memory::dims srcDims = src->getDims();
  memory::dims dstDims = dst->getDims();
  CHECK_EQ(srcDims.size(), dstDims.size());
  for (size_t i = 0; i < srcDims.size(); ++i) {
    CHECK_EQ(srcDims[i], dstDims[i]);
  }
  return std::make_shared<reorder>(*src, *dst);
}

T
tensor-tang 已提交
73 74 75 76 77 78 79 80
void MKLDNNMatrix::reorderDataFrom(const MKLDNNMatrixPtr& m,
                                   memory::format srcFmt,
                                   memory::dims targetDim) {
  memory::format dstFmt = getFormat();
  if (srcFmt == dstFmt) {
    return;
  }
  CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal";
T
refine  
tensor-tang 已提交
81
  reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim);
T
tensor-tang 已提交
82 83 84 85 86 87 88 89 90 91
}

void MKLDNNMatrix::reorderDataTo(const MKLDNNMatrixPtr& m,
                                 memory::format dstFmt,
                                 memory::dims targetDim) {
  memory::format srcFmt = getFormat();
  if (srcFmt == dstFmt) {
    return;
  }
  CHECK_EQ(getElementCnt(), m->getElementCnt()) << "size should equal";
T
refine  
tensor-tang 已提交
92
  reorderOnce(getData(), m->getData(), srcFmt, dstFmt, targetDim);
T
tensor-tang 已提交
93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
}

void MKLDNNMatrix::reorderOnce(void* srcData,
                               void* dstData,
                               memory::format srcFmt,
                               memory::format dstFmt,
                               memory::dims dm) {
  CHECK(srcData);
  CHECK(dstData);
  MatrixPtr tmpSrc;
  if (dstData == srcData) {
    // inplace data
    size_t sz = 1;
    for (size_t i = 0; i < dm.size(); ++i) {
      sz *= dm[i];
    }
    tmpSrc = Matrix::create(sz, 1, false, false);
    tmpSrc->copyFrom((real*)srcData, sz);
    srcData = tmpSrc->getData();
  }

  auto dtype = this->getDtype();
  auto srcMD = memory::desc(dm, dtype, srcFmt);
  auto dstMD = memory::desc(dm, dtype, dstFmt);

  auto eg = this->getEngine();
  auto src = memory(memory::primitive_desc(srcMD, eg), srcData);
  auto dst = memory(memory::primitive_desc(dstMD, eg), dstData);

  auto r = reorder(src, dst);
  stream(stream::kind::eager).submit({r}).wait();
}

126 127 128 129 130 131 132
void MKLDNNMatrix::downSpatial() {
  int fmt = getFormat();
  if (!(fmt == memory::format::nchw || fmt == memory::format::oihw)) {
    // only support nchw and oihw yet, later can support more like nhwc, ihwo
    return;
  }

T
refine  
tensor-tang 已提交
133
  // TODO(TJ): change H(height) and W(width) if support nhwc or more
134
  const int H = 2, W = 3;
T
refine  
tensor-tang 已提交
135
  memory::dims srcDims = getDims();
136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
  if (srcDims[H] != 1 || srcDims[W] != 1) {
    // can not down spatial
    return;
  }

  memory::dims dstDims = memory::dims{srcDims[0], srcDims[1]};
  memory::format dstFmt;
  switch (fmt) {
    case memory::format::nchw:
      dstFmt = memory::format::nc;
      break;
    case memory::format::oihw:
      dstFmt = memory::format::oi;
      break;
    default:
      LOG(FATAL) << "unsupported format";
  }
  memory::desc md = memory::desc(dstDims, getDtype(), dstFmt);
  memory::primitive_desc pd = memory::primitive_desc(md, getEngine());
155 156 157 158 159
  mkldnn_primitive_t result;
  mkldnn::error::wrap_c_api(
      mkldnn_primitive_create(&result, pd.get(), nullptr, nullptr),
      "could not create a memory primitive");
  reset(result);
160
  set_data_handle(data_);
T
tensor-tang 已提交
161 162 163
}

}  // namespace paddle