activation_mkldnn_op.cc 9.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 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 "paddle/fluid/operators/activation_op.h"
16
#include "paddle/fluid/platform/mkldnn_reuse.h"
17 18 19 20

namespace paddle {
namespace operators {

21 22 23 24 25 26 27 28
using framework::DataLayout;
using framework::Tensor;
using mkldnn::memory;
using mkldnn::primitive;
using mkldnn::stream;
using platform::GetMKLDNNFormat;
using platform::MKLDNNDeviceContext;
using platform::to_void_cast;
29

30 31 32 33 34 35
template <typename Functor>
class MKLDNNActivationKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    const auto *x = ctx.Input<Tensor>("X");
36 37
    PADDLE_ENFORCE_EQ(x->layout(), DataLayout::kMKLDNN,
                      "Wrong layout set for X tensor");
A
Adam 已提交
38
    PADDLE_ENFORCE_NE(x->format(), MKLDNNMemoryFormat::undef,
39
                      "Wrong format set for X tensor");
40 41 42 43 44

    Functor functor;
    functor(ctx);
  }
};
K
Krzysztof Binias 已提交
45

46 47 48 49 50 51
template <typename Functor>
class MKLDNNActivationGradKernel
    : public framework::OpKernel<typename Functor::ELEMENT_TYPE> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
    const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
52 53
    PADDLE_ENFORCE_EQ(diff_y->layout(), DataLayout::kMKLDNN,
                      "Wrong layout set for Input OutGrad tensor");
A
Adam 已提交
54
    PADDLE_ENFORCE_NE(diff_y->format(), MKLDNNMemoryFormat::undef,
55
                      "Wrong format set for Input OutGrad tensor");
56 57 58 59 60 61 62 63

    Functor functor;
    functor(ctx);
  }
};

template <typename T>
void eltwise_forward(const framework::ExecutionContext &ctx,
A
Adam 已提交
64
                     mkldnn::algorithm algorithm) {
65 66 67
  PADDLE_ENFORCE_EQ(platform::is_cpu_place(ctx.GetPlace()), true,
                    paddle::platform::errors::PreconditionNotMet(
                        "Operator DNNL eletwise_forward must use CPUPlace"));
68 69
  auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();

70 71
  const auto *x = ctx.Input<Tensor>("X");
  auto *y = ctx.Output<Tensor>("Out");
72

73 74 75 76 77 78
  T alpha = ctx.HasAttr("alpha") ? ctx.Attr<T>("alpha") : 0;
  T beta = ctx.HasAttr("beta") ? ctx.Attr<T>("beta") : 0;

  // paddle uses beta but mkldnn uses alpha for swish
  if (algorithm == mkldnn::algorithm::eltwise_swish) {
    std::swap(alpha, beta);
A
Adam 已提交
79 80
  } else if (algorithm == dnnl::algorithm::eltwise_bounded_relu) {
    alpha = ctx.Attr<T>("threshold");
81
  }
A
Adam 已提交
82

Y
Yihua Xu 已提交
83 84 85 86
  PADDLE_ENFORCE(
      x->dims().size() == 2 || x->dims().size() == 3 || x->dims().size() == 4,
      "Input dim must be with 2, 3 or 4");

A
Adam 已提交
87
  auto src_tz = framework::vectorize<int64_t>(x->dims());
88

89
  auto src_format = src_tz.size() == 2 ? MKLDNNMemoryFormat::nc : x->format();
90

91
  platform::ActivationMKLDNNHandler<T> handler(
92 93
      src_tz, algorithm, alpha, beta, src_format, dev_ctx, ctx.GetPlace(),
      ctx.InputName("X"));
94

95
  auto src_memory_p = handler.AcquireSrcMemory(x);
96 97
  auto dst_memory_p =
      x->IsSharedBufferWith(*y) ? src_memory_p : handler.AcquireDstMemory(y);
A
Adam 已提交
98
  auto activation_p = handler.AcquireForwardPrimitive();
99

A
Adam 已提交
100 101 102 103
  mkldnn::stream astream(dev_ctx.GetEngine());
  activation_p->execute(astream, {{MKLDNN_ARG_FROM, *src_memory_p},
                                  {MKLDNN_ARG_TO, *dst_memory_p}});
  astream.wait();
104

105
  y->set_layout(DataLayout::kMKLDNN);
106
  y->set_format(GetMKLDNNFormat(*dst_memory_p));
107 108
}

109 110
template <typename T>
void eltwise_grad(const framework::ExecutionContext &ctx,
A
Adam 已提交
111
                  mkldnn::algorithm algorithm) {
112 113
  auto &dev_ctx = ctx.template device_context<MKLDNNDeviceContext>();

114
  const auto *x = ctx.Input<Tensor>("X");
115 116
  const auto *diff_y = ctx.Input<Tensor>(framework::GradVarName("Out"));
  auto *diff_x = ctx.Output<Tensor>(framework::GradVarName("X"));
117

118 119 120 121 122 123
  T alpha = ctx.HasAttr("alpha") ? ctx.Attr<T>("alpha") : 0;
  T beta = ctx.HasAttr("beta") ? ctx.Attr<T>("beta") : 0;

  // paddle uses beta but mkldnn uses alpha for swish
  if (algorithm == mkldnn::algorithm::eltwise_swish) {
    std::swap(alpha, beta);
A
Adam 已提交
124 125
  } else if (algorithm == dnnl::algorithm::eltwise_bounded_relu) {
    alpha = ctx.Attr<T>("threshold");
126
  }
A
Adam 已提交
127

A
Adam 已提交
128
  auto diff_dst_tz = framework::vectorize<int64_t>(diff_y->dims());
K
Krzysztof Binias 已提交
129

130 131
  // diff_dst and src dims should be the same
  auto src_format =
132
      diff_dst_tz.size() == 2 ? MKLDNNMemoryFormat::nc : x->format();
133

134
  auto diff_y_format =
135
      diff_dst_tz.size() == 2 ? MKLDNNMemoryFormat::nc : diff_y->format();
136

137 138
  platform::ActivationMKLDNNHandler<T> handler(
      diff_dst_tz, algorithm, alpha, beta, src_format, diff_y_format, dev_ctx,
H
hong 已提交
139
      ctx.GetPlace(), ctx.InputName("X"));
140

141 142 143
  auto src_memory_p = handler.AcquireBackwardSrcMemory(x);
  auto diff_dst_memory_p = handler.AcquireDiffDstMemory(diff_y);
  auto diff_src_memory_p = handler.AcquireDiffSrcMemory(diff_x);
A
Adam 已提交
144 145 146 147 148 149 150 151
  auto activation_backward_p = handler.AcquireBackwardPrimitive();

  mkldnn::stream astream(dev_ctx.GetEngine());
  activation_backward_p->execute(astream,
                                 {{MKLDNN_ARG_SRC, *src_memory_p},
                                  {MKLDNN_ARG_DIFF_DST, *diff_dst_memory_p},
                                  {MKLDNN_ARG_DIFF_SRC, *diff_src_memory_p}});
  astream.wait();
152

153
  diff_x->set_layout(DataLayout::kMKLDNN);
154
  diff_x->set_format(GetMKLDNNFormat(*diff_src_memory_p));
155 156 157 158
}

template <typename T, mkldnn::algorithm algorithm>
struct MKLDNNActivationFunc : public BaseActivationFunctor<T> {
159
  void operator()(const framework::ExecutionContext &ctx) const {
160 161 162 163 164 165
    eltwise_forward<T>(ctx, algorithm);
  }
};

template <typename T, mkldnn::algorithm algorithm>
struct MKLDNNActivationGradFunc : public BaseActivationFunctor<T> {
166
  void operator()(const framework::ExecutionContext &ctx) const {
167 168 169 170
    eltwise_grad<T>(ctx, algorithm);
  }
};

A
Adam 已提交
171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194
template <typename T>
struct GeluMKLDNNFunctor : public BaseActivationFunctor<T> {
  void operator()(const framework::ExecutionContext &ctx) const {
    const bool approximate = ctx.Attr<bool>("approximate");
    if (approximate) {
      eltwise_forward<T>(ctx, mkldnn::algorithm::eltwise_gelu_tanh);
    } else {
      eltwise_forward<T>(ctx, mkldnn::algorithm::eltwise_gelu_erf);
    }
  }
};

template <typename T>
struct GeluMKLDNNGradFunctor : public BaseActivationFunctor<T> {
  void operator()(const framework::ExecutionContext &ctx) const {
    const bool approximate = ctx.Attr<bool>("approximate");
    if (approximate) {
      eltwise_grad<T>(ctx, mkldnn::algorithm::eltwise_gelu_tanh);
    } else {
      eltwise_grad<T>(ctx, mkldnn::algorithm::eltwise_gelu_erf);
    }
  }
};

195
template <typename T>
T
tensor-tang 已提交
196
using ReluMKLDNNFunctor =
197 198
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_relu>;

A
Adam 已提交
199 200 201 202
template <typename T>
using Relu6MKLDNNFunctor =
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_bounded_relu>;

203 204 205 206
template <typename T>
using SwishMKLDNNFunctor =
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_swish>;

207 208 209 210
template <typename T>
using SigmoidMKLDNNFunctor =
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_logistic>;

211
template <typename T>
T
tensor-tang 已提交
212
using TanhMKLDNNFunctor =
213 214 215
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_tanh>;

template <typename T>
T
tensor-tang 已提交
216
using SqrtMKLDNNFunctor =
217 218 219
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_sqrt>;

template <typename T>
T
tensor-tang 已提交
220
using AbsMKLDNNFunctor =
221 222 223
    MKLDNNActivationFunc<T, mkldnn::algorithm::eltwise_abs>;

template <typename T>
T
tensor-tang 已提交
224
using ReluMKLDNNGradFunctor =
225 226
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_relu>;

A
Adam 已提交
227 228 229 230
template <typename T>
using Relu6MKLDNNGradFunctor =
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_bounded_relu>;

231 232 233 234
template <typename T>
using SwishMKLDNNGradFunctor =
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_swish>;

235 236 237 238
template <typename T>
using SigmoidMKLDNNGradFunctor =
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_logistic>;

239
template <typename T>
T
tensor-tang 已提交
240
using TanhMKLDNNGradFunctor =
241 242 243
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_tanh>;

template <typename T>
T
tensor-tang 已提交
244
using SqrtMKLDNNGradFunctor =
245 246 247
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_sqrt>;

template <typename T>
T
tensor-tang 已提交
248
using AbsMKLDNNGradFunctor =
249 250 251 252 253 254 255 256 257 258 259 260 261
    MKLDNNActivationGradFunc<T, mkldnn::algorithm::eltwise_abs>;
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

#define REGISTER_ACTIVATION_MKLDNN_KERNEL(act_type, functor, grad_functor) \
  REGISTER_OP_KERNEL(act_type, MKLDNN, ::paddle::platform::CPUPlace,       \
                     ops::MKLDNNActivationKernel<ops::functor<float>>);    \
  REGISTER_OP_KERNEL(                                                      \
      act_type##_grad, MKLDNN, ::paddle::platform::CPUPlace,               \
      ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>);

262 263
#define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro)                     \
  __macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor);          \
A
Adam 已提交
264
  __macro(relu6, Relu6MKLDNNFunctor, Relu6MKLDNNGradFunctor);       \
265 266 267 268 269 270
  __macro(leaky_relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor);    \
  __macro(gelu, GeluMKLDNNFunctor, GeluMKLDNNGradFunctor);          \
  __macro(swish, SwishMKLDNNFunctor, SwishMKLDNNGradFunctor);       \
  __macro(sigmoid, SigmoidMKLDNNFunctor, SigmoidMKLDNNGradFunctor); \
  __macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor);          \
  __macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor);          \
T
tensor-tang 已提交
271
  __macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor);
272 273

FOR_EACH_MKLDNN_KERNEL_FUNCTOR(REGISTER_ACTIVATION_MKLDNN_KERNEL);