fc_mkldnn_op.cc 11.3 KB
Newer Older
M
mozga-intel 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* 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/framework/tensor.h"
16
#include "paddle/fluid/operators/fc_op.h"
M
mozga-intel 已提交
17 18 19 20 21 22 23 24 25 26 27 28 29
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/mkldnn_helper.h"

namespace paddle {
namespace operators {

using paddle::framework::Tensor;
using paddle::platform::MKLDNNDeviceContext;

template <typename T>
class MKLDNNMD {
 public:
  explicit MKLDNNMD(const T* in, const T* w, bool bias)
M
mozga-intel 已提交
30 31
      : in(paddle::framework::vectorize2int(in->dims())),
        w(paddle::framework::vectorize2int(w->dims())) {
M
mozga-intel 已提交
32 33 34 35
    with_bias_ = bias;
  }

  mkldnn::memory::desc dst() const {
M
mozga-intel 已提交
36
    return platform::MKLDNNMemDesc({in[0], w[1]},
M
mozga-intel 已提交
37 38 39 40 41
                                   mkldnn::memory::data_type::f32,
                                   mkldnn::memory::format::nc);
  }

  mkldnn::memory::desc src() const {
M
mozga-intel 已提交
42 43
    return is_spatial()
               ? platform::MKLDNNMemDesc({in[0], in[1], in[2], in[3]},
M
mozga-intel 已提交
44 45
                                         mkldnn::memory::data_type::f32,
                                         mkldnn::memory::format::nchw)
M
mozga-intel 已提交
46
               : platform::MKLDNNMemDesc({in[0], in[1]},
M
mozga-intel 已提交
47 48 49 50 51
                                         mkldnn::memory::data_type::f32,
                                         mkldnn::memory::format::nc);
  }

  mkldnn::memory::desc weights() const {
M
mozga-intel 已提交
52 53
    return is_spatial()
               ? platform::MKLDNNMemDesc({w[1], in[1], in[2], in[3]},
M
mozga-intel 已提交
54 55
                                         mkldnn::memory::data_type::f32,
                                         mkldnn::memory::format::oihw)
M
mozga-intel 已提交
56
               : platform::MKLDNNMemDesc({w[1], in[1]},
M
mozga-intel 已提交
57 58 59 60 61 62
                                         mkldnn::memory::data_type::f32,
                                         mkldnn::memory::format::oi);
  }

  mkldnn::memory::desc bias() const {
    return with_bias_
M
mozga-intel 已提交
63
               ? platform::MKLDNNMemDesc({w[1]}, mkldnn::memory::data_type::f32,
M
mozga-intel 已提交
64 65 66 67 68 69
                                         mkldnn::memory::format::format_undef)
               : platform::MKLDNNMemDesc({}, mkldnn::memory::data_type::f32,
                                         mkldnn::memory::format::format_undef);
  }

 private:
M
mozga-intel 已提交
70 71 72 73
  bool is_spatial() const { return in.size() > 1 && w.size() > 1; }

  std::vector<int> in;
  std::vector<int> w;
M
mozga-intel 已提交
74
  bool with_bias_;
M
mozga-intel 已提交
75
  bool is_spatial_;
M
mozga-intel 已提交
76 77 78 79 80
};

class MKLDNNMemory {
 public:
  MKLDNNMemory(MKLDNNMD<Tensor>* t, const mkldnn::engine& e)
M
mozga-intel 已提交
81
      : md_(t), engine_(e) {}
M
mozga-intel 已提交
82 83 84 85 86 87 88 89 90 91 92 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
  virtual ~MKLDNNMemory() = default;

  template <typename Output>
  mkldnn::memory dst(const Output* out) {
    return mkldnn::memory({md_->dst(), engine_},
                          static_cast<void*>(const_cast<float*>(out)));
  }

  template <typename Output>
  mkldnn::memory dst(Output* out) {
    return mkldnn::memory({md_->dst(), engine_}, out);
  }

  template <typename Input>
  mkldnn::memory src(const Input* in) {
    return mkldnn::memory({md_->src(), engine_},
                          static_cast<void*>(const_cast<float*>(in)));
  }

  template <typename Weight>
  mkldnn::memory weights(const Weight* w) {
    return mkldnn::memory({md_->weights(), engine_},
                          static_cast<void*>(const_cast<float*>(w)));
  }

  mkldnn::memory bias() {
    return mkldnn::memory(mkldnn::memory::primitive_desc(md_->bias(), engine_));
  }

 private:
  MKLDNNMD<Tensor>* md_;
  const mkldnn::engine& engine_;
};

template <typename T>
class FCMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
Y
yuyang18 已提交
118
 public:
M
mozga-intel 已提交
119 120 121 122 123 124 125 126 127 128
  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<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    auto input = ctx.Input<Tensor>("Input");
    auto w = ctx.Input<Tensor>("W");

129
    PADDLE_ENFORCE(input->dims().size() == 2 || input->dims().size() == 4,
130
                   "Input must be with 2 or 4 dimensions, i.e. NCHW");
131 132
    PADDLE_ENFORCE(w->dims().size() == 2 || w->dims().size() == 4,
                   "Weights must be with 2 or 4 dimensions, i.e. OI or OIHW");
M
mozga-intel 已提交
133

134
    bool with_bias = ctx.Attr<bool>("bias_attr");
M
mozga-intel 已提交
135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218
    MKLDNNMD<Tensor> md(input, w, with_bias);

    std::shared_ptr<mkldnn::inner_product_forward::primitive_desc> pd =
        FcFwdPrimitiveDesc(md.src(), md.weights(), md.dst(), md.bias(),
                           with_bias, mkldnn_engine);

    const std::string key = ctx.op().Output("Out");
    const std::string key_fc_pd = key + "@fc_pd";

    dev_ctx.SetBlob(key_fc_pd, pd);

    MKLDNNMemory mem(&md, mkldnn_engine);

    const T* input_data = input->data<T>();
    const T* w_data = w->data<T>();

    auto output = ctx.Output<Tensor>("Out");
    T* output_data = output->mutable_data<T>(ctx.GetPlace());

    auto dst_memory = mem.dst(output_data);
    auto src_memory = mem.src(input_data);
    auto weights_memory = mem.weights(w_data);
    auto bias_memory = mem.bias();

    auto forward = with_bias ? mkldnn::inner_product_forward(
                                   *pd, src_memory, weights_memory, bias_memory,
                                   dst_memory)
                             : mkldnn::inner_product_forward(
                                   *pd, src_memory, weights_memory, dst_memory);

    std::vector<mkldnn::primitive> pipeline = {forward};
    mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
  }

 private:
  std::unique_ptr<mkldnn::inner_product_forward::primitive_desc>
  FcFwdPrimitiveDesc(const mkldnn::memory::desc& src,
                     const mkldnn::memory::desc& weights,
                     const mkldnn::memory::desc& dst,
                     const mkldnn::memory::desc& bias, const bool with_bias,
                     const mkldnn::engine& engine) const {
    auto desc = with_bias
                    ? mkldnn::inner_product_forward::desc(
                          mkldnn::prop_kind::forward, src, weights, bias, dst)
                    : mkldnn::inner_product_forward::desc(
                          mkldnn::prop_kind::forward, src, weights, dst);

    auto pd = new mkldnn::inner_product_forward::primitive_desc(desc, engine);
    return std::unique_ptr<mkldnn::inner_product_forward::primitive_desc>(pd);
  }
};

template <typename T>
class FCMKLDNNGradOpKernel : 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<MKLDNNDeviceContext>();
    const auto& mkldnn_engine = dev_ctx.GetEngine();

    T* input_grad_data = nullptr;
    T* w_grad_data = nullptr;

    Tensor* input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
    Tensor* w_grad = ctx.Output<Tensor>(framework::GradVarName("W"));

    if (input_grad) {
      input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
    }
    if (w_grad) {
      w_grad_data = w_grad->mutable_data<T>(ctx.GetPlace());
    }

    const Tensor* input = ctx.Input<Tensor>("Input");
    const T* input_data = input->data<T>();

    const Tensor* w = ctx.Input<Tensor>("W");
    const T* w_data = w->data<T>();

    const Tensor* out_grad = ctx.Input<Tensor>(framework::GradVarName("Out"));
    const T* out_grad_data = out_grad->data<T>();

219
    bool with_bias = ctx.Attr<bool>("bias_attr");
M
mozga-intel 已提交
220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304

    MKLDNNMD<Tensor> md(input, w, with_bias);
    MKLDNNMemory mem(&md, mkldnn_engine);

    auto dst_memory = mem.dst(out_grad_data);
    auto src_memory = mem.src(input_data);
    auto weights_memory = mem.weights(w_data);
    auto bias_memory = mem.bias();

    const std::string key = ctx.op().Input("Out");
    const std::string key_fc_pd = key + "@fc_pd";

    auto pd =
        std::static_pointer_cast<mkldnn::inner_product_forward::primitive_desc>(
            dev_ctx.GetBlob(key_fc_pd));

    PADDLE_ENFORCE(pd != nullptr, "Fail to find key_fc_pd in device context");

    if (w_grad) {
      auto weights_grad_memory = mem.weights(w_grad_data);

      mkldnn::inner_product_backward_weights::primitive_desc bwd_weight_pd =
          FcBwdWeightsPrimitiveDesc(md.src(), md.weights(), md.dst(), md.bias(),
                                    with_bias, *pd, mkldnn_engine);

      auto bwd_weights_prim = mkldnn::inner_product_backward_weights(
          bwd_weight_pd, src_memory, dst_memory, weights_grad_memory,
          bias_memory);

      std::vector<mkldnn::primitive> pipeline{bwd_weights_prim};
      mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
    }

    if (input_grad) {
      auto src_grad_memory = mem.src(input_grad_data);

      mkldnn::inner_product_backward_data::primitive_desc bwd_data_pd =
          FcBwdDataPrimitiveDesc(md.src(), md.weights(), md.dst(), *pd,
                                 mkldnn_engine);

      auto bwd_data_prim = mkldnn::inner_product_backward_data(
          bwd_data_pd, dst_memory, weights_memory, src_grad_memory);

      std::vector<mkldnn::primitive> pipeline{bwd_data_prim};
      mkldnn::stream(mkldnn::stream::kind::eager).submit(pipeline).wait();
    }
  }

 private:
  mkldnn::inner_product_backward_weights::primitive_desc
  FcBwdWeightsPrimitiveDesc(
      const mkldnn::memory::desc& src, const mkldnn::memory::desc& diff_weights,
      const mkldnn::memory::desc& diff_dst, const mkldnn::memory::desc& bias,
      const bool with_bias,
      const mkldnn::inner_product_forward::primitive_desc& pd,
      const mkldnn::engine& engine) const {
    auto bwd_weight_desc = with_bias
                               ? mkldnn::inner_product_backward_weights::desc(
                                     src, diff_weights, bias, diff_dst)
                               : mkldnn::inner_product_backward_weights::desc(
                                     src, diff_weights, bias, diff_dst);

    return mkldnn::inner_product_backward_weights::primitive_desc(
        bwd_weight_desc, engine, pd);
  }

  mkldnn::inner_product_backward_data::primitive_desc FcBwdDataPrimitiveDesc(
      const mkldnn::memory::desc& diff_src, const mkldnn::memory::desc& weights,
      const mkldnn::memory::desc& diff_dst,
      const mkldnn::inner_product_forward::primitive_desc& pd,
      const mkldnn::engine& engine) const {
    auto bwd_data_desc =
        mkldnn::inner_product_backward_data::desc(diff_src, weights, diff_dst);
    return mkldnn::inner_product_backward_data::primitive_desc(bwd_data_desc,
                                                               engine, pd);
  }
};
}  // namespace operators
}  // namespace paddle

REGISTER_OP_KERNEL(fc, MKLDNN, ::paddle::platform::CPUPlace,
                   paddle::operators::FCMKLDNNOpKernel<float>);

REGISTER_OP_KERNEL(fc_grad, MKLDNN, ::paddle::platform::CPUPlace,
                   paddle::operators::FCMKLDNNGradOpKernel<float>);