pool_mkldnn_op.cc 13.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.

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 "paddle/fluid/operators/pool_op.h"
#include "paddle/fluid/platform/mkldnn_helper.h"

namespace paddle {
namespace operators {

21 22 23 24 25 26
using mkldnn::memory;  // Note: paddle has also "memory" namespace
using mkldnn::pooling_forward;
using mkldnn::pooling_backward;

// Generate keys for storing/retriving primitives for this operator
// TODO(jczaja): Make hashing function more optimial
M
mozga-intel 已提交
27 28 29 30 31 32 33
static std::string gethash(const memory::dims& input_dims,
                           const std::string& pooling_type,
                           const std::vector<int>& ksize,
                           const std::vector<int>& strides,
                           const std::vector<int>& paddings,
                           const std::string& suffix) {
  auto dims2str = [](const memory::dims& operand_dims) {
34 35 36 37 38 39 40 41 42 43
    std::string dstr = "";
    for (size_t i = 0; i < operand_dims.size(); ++i) {
      dstr += std::to_string(operand_dims[i]) + "-";
    }
    return dstr;
  };
  return dims2str(input_dims) + dims2str(ksize) + dims2str(strides) +
         dims2str(paddings) + pooling_type + suffix;
}

44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84
template <typename T>
class PoolMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

    auto& dev_ctx =
        ctx.template device_context<platform::MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    const Tensor* input = ctx.Input<Tensor>("X");
    Tensor* output = ctx.Output<Tensor>("Out");

    // Get an unique name from "argument" name of "Out" variable
    // This name will be used as key when saving info into device context

    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    if (ctx.Attr<bool>("global_pooling")) {
      for (size_t i = 0; i < ksize.size(); ++i) {
        paddings[i] = 0;
        ksize[i] = static_cast<int>(input->dims()[i + 2]);
      }
    }

    // Only 2D pooling is supported now
    PADDLE_ENFORCE(ksize.size() == 2, "ksize must be 2D, i.e. 2D pooling");
    PADDLE_ENFORCE(pooling_type == "max" || pooling_type == "avg",
                   "pooling_type must be 'max' or 'avg'");
    PADDLE_ENFORCE(input->dims().size() == 4,
                   "Input dim must be with 4, i.e. NCHW");

    const T* input_data = input->data<T>();
    T* output_data = output->mutable_data<T>(ctx.GetPlace());

    std::vector<int> src_tz = paddle::framework::vectorize2int(input->dims());
    std::vector<int> dst_tz = paddle::framework::vectorize2int(output->dims());

85 86 87 88 89 90 91 92
    const std::string key = gethash(src_tz, pooling_type, ksize, strides,
                                    paddings, ctx.op().Output("Out"));
    const std::string key_pool_p = key + "@pool_p";
    const std::string key_pool_pd = key + "@pool_pd";
    const std::string key_pool_src_mem_p = key + "@pool_src_mem_p";
    const std::string key_pool_dst_mem_p = key + "@pool_dst_mem_p";
    const std::string key_pool_workspace_memory =
        key + "@pool_workspace_memory";
93

94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146
    auto pool_p =
        std::static_pointer_cast<pooling_forward>(dev_ctx.GetBlob(key_pool_p));
    if (pool_p == nullptr) {
      // TODO(pzelazko-intel): support more formats

      auto src_md =
          platform::MKLDNNMemDesc(src_tz, platform::MKLDNNGetDataType<T>(),
                                  mkldnn::memory::format::nchw);
      auto dst_md =
          platform::MKLDNNMemDesc(dst_tz, platform::MKLDNNGetDataType<T>(),
                                  mkldnn::memory::format::nchw);

      std::shared_ptr<pooling_forward::primitive_desc> pool_pd =
          CreatePrimitiveDesc(src_md, dst_md, strides, paddings, ksize,
                              pooling_type, mkldnn_engine);

      // save pool_pd into global device context to be referred in backward path
      dev_ctx.SetBlob(key_pool_pd, pool_pd);

      std::shared_ptr<mkldnn::memory> workspace_memory =
          CreateWorkspaceMemory(pool_pd, pooling_type, mkldnn_engine);

      // save pool_workspace_memory to be referred in backward path
      dev_ctx.SetBlob(key_pool_workspace_memory, workspace_memory);

      auto pool_src_memory_p = std::make_shared<memory>(
          memory::primitive_desc{src_md, mkldnn_engine},
          static_cast<void*>(const_cast<T*>(input_data)));
      dev_ctx.SetBlob(key_pool_src_mem_p, pool_src_memory_p);

      auto pool_dst_memory_p = std::make_shared<memory>(
          memory::primitive_desc{dst_md, mkldnn_engine},
          static_cast<void*>(output_data));
      dev_ctx.SetBlob(key_pool_dst_mem_p, pool_dst_memory_p);

      pool_p = std::make_shared<pooling_forward>(
          *pool_pd, *(pool_src_memory_p.get()), *(pool_dst_memory_p.get()),
          *workspace_memory);
      dev_ctx.SetBlob(key_pool_p, pool_p);
    } else {
      // Primitives already exist
      auto pool_src_memory_p =
          std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_src_mem_p));
      PADDLE_ENFORCE(pool_src_memory_p != nullptr,
                     "Fail to find pooling src mem_p in device context");
      auto pool_dst_memory_p =
          std::static_pointer_cast<memory>(dev_ctx.GetBlob(key_pool_dst_mem_p));
      PADDLE_ENFORCE(pool_dst_memory_p != nullptr,
                     "Fail to find pooling dst mem_p in device context");
      pool_src_memory_p->set_data_handle(
          reinterpret_cast<void*>(const_cast<T*>(input_data)));
      pool_dst_memory_p->set_data_handle(output_data);
    }
147 148

    // push primitive to stream and wait until it's executed
149
    std::vector<mkldnn::primitive> pipeline{*(pool_p.get())};
150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }

 private:
  std::unique_ptr<mkldnn::pooling_forward::primitive_desc> CreatePrimitiveDesc(
      const mkldnn::memory::desc& src, const mkldnn::memory::desc& dst,
      const std::vector<int>& stride, const std::vector<int>& padding,
      const std::vector<int>& kernel, const std::string& pooling_type,
      const mkldnn::engine& engine) const {
    auto pool_desc = mkldnn::pooling_forward::desc(
        mkldnn::prop_kind::forward,
        pooling_type == "max" ? mkldnn::algorithm::pooling_max
                              : mkldnn::algorithm::pooling_avg,
        src, dst, stride, kernel, padding, padding, mkldnn::padding_kind::zero);

    auto p_pool_pd =
        new mkldnn::pooling_forward::primitive_desc(pool_desc, engine);
    return std::unique_ptr<mkldnn::pooling_forward::primitive_desc>(p_pool_pd);
  }

  std::unique_ptr<mkldnn::memory> CreateWorkspaceMemory(
      std::shared_ptr<mkldnn::pooling_forward::primitive_desc> pool_pd,
      const std::string& pooling_type, const mkldnn::engine& engine) const {
    mkldnn::memory::primitive_desc workspace_md =
        pooling_type == "max"
            ? pool_pd->workspace_primitive_desc()
176 177 178 179
            : mkldnn::memory::primitive_desc({{},
                                              platform::MKLDNNGetDataType<T>(),
                                              mkldnn::memory::format::nchw},
                                             engine);
180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220

    auto p_workspace_memory = new mkldnn::memory(workspace_md);
    return std::unique_ptr<mkldnn::memory>(p_workspace_memory);
  }
};

template <typename T>
class PoolMKLDNNGradOpKernel : public paddle::framework::OpKernel<T> {
 public:
  void Compute(const paddle::framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(paddle::platform::is_cpu_place(ctx.GetPlace()),
                   "It must use CPUPlace.");

    const Tensor* in_x = ctx.Input<Tensor>("X");
    const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    Tensor* in_x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));

    std::string pooling_type = ctx.Attr<std::string>("pooling_type");
    std::vector<int> ksize = ctx.Attr<std::vector<int>>("ksize");
    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");

    if (ctx.Attr<bool>("global_pooling")) {
      for (size_t i = 0; i < ksize.size(); ++i) {
        paddings[i] = 0;
        ksize[i] = static_cast<int>(in_x->dims()[i + 2]);
      }
    }

    auto& dev_ctx =
        ctx.template device_context<platform::MKLDNNDeviceContext>();
    const mkldnn::engine& mkldnn_engine = dev_ctx.GetEngine();

    const T* out_grad_data = out_grad->data<T>();
    T* in_x_grad_data = in_x_grad->mutable_data<T>(ctx.GetPlace());

    std::vector<int> diff_src_tz =
        paddle::framework::vectorize2int(in_x_grad->dims());
    std::vector<int> diff_dst_tz =
        paddle::framework::vectorize2int(out_grad->dims());

221 222 223 224 225 226 227 228 229 230
    // Get an unique name from "argument" name of "Out" variable
    // This name will be used as key when referring info from device context
    const std::string key = gethash(diff_src_tz, pooling_type, ksize, strides,
                                    paddings, ctx.op().Input("Out"));
    const std::string key_pool_bwd_p = key + "@pool_bwd_p";
    const std::string key_pool_diff_src_mem_p = key + "@pool_diff_src_mem_p";
    const std::string key_pool_diff_dst_mem_p = key + "@pool_diff_dst_mem_p";
    const std::string key_pool_pd = key + "@pool_pd";
    const std::string key_pool_workspace_memory =
        key + "@pool_workspace_memory";
231

232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
    auto pool_bwd_p = std::static_pointer_cast<pooling_backward>(
        dev_ctx.GetBlob(key_pool_bwd_p));
    if (pool_bwd_p == nullptr) {
      auto diff_src_md =
          platform::MKLDNNMemDesc(diff_src_tz, platform::MKLDNNGetDataType<T>(),
                                  mkldnn::memory::format::nchw);
      auto diff_dst_md =
          platform::MKLDNNMemDesc(diff_dst_tz, platform::MKLDNNGetDataType<T>(),
                                  mkldnn::memory::format::nchw);
      // Retrieve pool_pd/pool_workspace_memory from device context
      auto pool_pd =
          std::static_pointer_cast<mkldnn::pooling_forward::primitive_desc>(
              dev_ctx.GetBlob(key_pool_pd));
      PADDLE_ENFORCE(pool_pd != nullptr,
                     "Fail to find pool_pd in device context");

      auto workspace_memory = std::static_pointer_cast<mkldnn::memory>(
          dev_ctx.GetBlob(key_pool_workspace_memory));
      PADDLE_ENFORCE(workspace_memory != nullptr,
                     "Fail to find workspace_memory in device context");

      auto pool_diff_src_memory_p = std::make_shared<memory>(memory(
          {diff_src_md, mkldnn_engine}, static_cast<void*>(in_x_grad_data)));
      dev_ctx.SetBlob(key_pool_diff_src_mem_p, pool_diff_src_memory_p);

      auto pool_diff_dst_memory_p = std::make_shared<memory>(
          memory({diff_dst_md, mkldnn_engine},
                 static_cast<void*>(const_cast<T*>(out_grad_data))));
      dev_ctx.SetBlob(key_pool_diff_dst_mem_p, pool_diff_dst_memory_p);

      auto pool_bwd_desc = mkldnn::pooling_backward::desc(
          pooling_type == "max" ? mkldnn::algorithm::pooling_max
                                : mkldnn::algorithm::pooling_avg,
          diff_src_md, diff_dst_md, strides, ksize, paddings, paddings,
          mkldnn::padding_kind::zero);
      auto pool_bwd_pd = mkldnn::pooling_backward::primitive_desc(
          pool_bwd_desc, mkldnn_engine, *pool_pd);

      pool_bwd_p = std::make_shared<pooling_backward>(
          pool_bwd_pd, *(pool_diff_dst_memory_p.get()), *workspace_memory,
          *(pool_diff_src_memory_p));
      dev_ctx.SetBlob(key_pool_bwd_p, pool_bwd_p);
    } else {
      // Primitives already exist
      auto pool_diff_src_memory_p = std::static_pointer_cast<memory>(
          dev_ctx.GetBlob(key_pool_diff_src_mem_p));
      PADDLE_ENFORCE(pool_diff_src_memory_p != nullptr,
                     "Fail to find pooling src mem_p in device context");
      auto pool_diff_dst_memory_p = std::static_pointer_cast<memory>(
          dev_ctx.GetBlob(key_pool_diff_dst_mem_p));
      PADDLE_ENFORCE(pool_diff_dst_memory_p != nullptr,
                     "Fail to find pooling dst mem_p in device context");
      pool_diff_src_memory_p->set_data_handle(
          reinterpret_cast<void*>(in_x_grad_data));
      pool_diff_dst_memory_p->set_data_handle(const_cast<T*>(out_grad_data));
    }
288 289

    // push primitive to stream and wait until it's executed
290
    std::vector<mkldnn::primitive> pipeline{*(pool_bwd_p.get())};
291 292 293 294 295 296 297 298 299 300 301
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }  // Compute()
};

}  // namespace operators
}  // namespace paddle

REGISTER_OP_KERNEL(pool2d, MKLDNN, ::paddle::platform::CPUPlace,
                   paddle::operators::PoolMKLDNNOpKernel<float>);
REGISTER_OP_KERNEL(pool2d_grad, MKLDNN, ::paddle::platform::CPUPlace,
                   paddle::operators::PoolMKLDNNGradOpKernel<float>);
反馈
建议
客服 返回
顶部