conv_op.cc 17.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
C
chengduoZH 已提交
2

L
Luo Tao 已提交
3 4 5
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
C
chengduoZH 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
C
chengduoZH 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
D
dzhwinter 已提交
14 15 16
#define GLOG_NO_ABBREVIATED_SEVERITIES
#define GOOGLE_GLOG_DLL_DECL
#include <glog/logging.h>
C
chengduoZH 已提交
17

Y
Yi Wang 已提交
18
#include "paddle/fluid/operators/conv_op.h"
Y
Update  
Yi Wang 已提交
19 20 21 22

#include <string>
#include <vector>

23 24 25 26 27 28
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
#endif
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
C
chengduoZH 已提交
29 30 31 32

namespace paddle {
namespace operators {

C
chengduoZH 已提交
33
void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
C
chengduoZH 已提交
34
  PADDLE_ENFORCE(ctx->HasInput("Input"),
C
chengduoZH 已提交
35
                 "Input(Input) of ConvOp should not be null.");
C
chengduoZH 已提交
36
  PADDLE_ENFORCE(ctx->HasInput("Filter"),
C
chengduoZH 已提交
37
                 "Input(Filter) of ConvOp should not be null.");
C
chengduoZH 已提交
38
  PADDLE_ENFORCE(ctx->HasOutput("Output"),
C
chengduoZH 已提交
39
                 "Output(Output) of ConvOp should not be null.");
C
chengduoZH 已提交
40

D
dzhwinter 已提交
41
  VLOG(3) << "Conv op infershape";
C
chengduoZH 已提交
42 43
  auto in_dims = ctx->GetInputDim("Input");
  auto filter_dims = ctx->GetInputDim("Filter");
44

C
chengduoZH 已提交
45 46 47
  std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
  std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
  int groups = ctx->Attrs().Get<int>("groups");
C
chengduoZH 已提交
48
  std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
D
dzhwinter 已提交
49 50
  VLOG(3) << "Conv op Before check";
  in_dims.size() == 4 || in_dims.size() == 5;
D
dzhwinter 已提交
51
  // PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
D
dzhwinter 已提交
52 53
  //               "Conv intput should be 4-D or 5-D tensor.");
  VLOG(3) << "check0";
C
chengduoZH 已提交
54

D
dzhwinter 已提交
55
  // PADDLE_ENFORCE_EQ(
D
dzhwinter 已提交
56 57 58 59
  //    in_dims.size(), filter_dims.size(),
  //    "Conv input dimension and filter dimension should be the same.");
  in_dims.size() == filter_dims.size();
  VLOG(3) << "enforce check0";
C
chengduoZH 已提交
60 61 62
  PADDLE_ENFORCE(
      in_dims.size() - strides.size() == 2U,
      "Conv input dimension and strides dimension should be consistent.");
D
dzhwinter 已提交
63
  VLOG(3) << "check1";
C
chengduoZH 已提交
64 65 66
  PADDLE_ENFORCE_EQ(
      paddings.size(), strides.size(),
      "Conv paddings dimension and Conv strides dimension should be the same.");
F
fengjiayi 已提交
67

D
dzhwinter 已提交
68
  VLOG(3) << "check2";
D
dzhwinter 已提交
69 70
  // in_dims[1] == filter_dims[1] * groups;
  // PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups,
D
dzhwinter 已提交
71 72
  //                  "The number of input channels should be equal to filter "
  //                  "channels * groups.");
D
dzhwinter 已提交
73 74 75
  VLOG(3) << "check3";
  // filter_dims[0] % groups == 0 ;
  // PADDLE_ENFORCE_EQ(
D
dzhwinter 已提交
76 77
  //    filter_dims[0] % groups, 0,
  //    "The number of output channels should be divided by groups.");
D
dzhwinter 已提交
78 79
  VLOG(3) << "filter" << filter_dims.size();
  VLOG(3) << "filter" << filter_dims[0];
D
dzhwinter 已提交
80 81 82
  VLOG(3) << "check4";
  VLOG(3) << "filter" << filter_dims[1];
  VLOG(3) << "dims" << in_dims[0];
C
chengduoZH 已提交
83 84

  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
D
dzhwinter 已提交
85
  VLOG(3) << "output shape";
C
chengduoZH 已提交
86
  for (size_t i = 0; i < strides.size(); ++i) {
D
dzhwinter 已提交
87
    VLOG(3) << "check5";
Y
Yang Yang 已提交
88 89 90
    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
                                          dilations[i], paddings[i],
                                          strides[i]));
D
dzhwinter 已提交
91
    VLOG(3) << "check pass";
C
chengduoZH 已提交
92
  }
D
dzhwinter 已提交
93
  VLOG(3) << "Conv InferShape Pass";
94
  ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
Y
Yang Yu 已提交
95
  ctx->ShareLoD("Input", "Output");
C
chengduoZH 已提交
96 97
}

98 99
framework::OpKernelType ConvOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
100
  framework::LibraryType library{framework::LibraryType::kPlain};
M
mozga-intel 已提交
101
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
102
  std::string data_format = ctx.Attr<std::string>("data_format");
M
mozga-intel 已提交
103 104
  framework::DataLayout layout = framework::StringToDataLayout(data_format);

C
chengduoZH 已提交
105
#ifdef PADDLE_WITH_CUDA
106
  if (platform::CanCUDNNBeUsed(ctx)) {
107
    library = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
108 109
  }
#endif
110
#ifdef PADDLE_WITH_MKLDNN
111
  if (library == framework::LibraryType::kPlain &&
112
      platform::CanMKLDNNBeUsed(ctx)) {
113
    library = framework::LibraryType::kMKLDNN;
M
mozga-intel 已提交
114
    layout = framework::DataLayout::kMKLDNN;
115
  }
116
#endif
117

K
Kexin Zhao 已提交
118 119 120 121 122 123 124 125
  auto input_data_type =
      framework::ToDataType(ctx.Input<Tensor>("Input")->type());
  auto filter_data_type =
      framework::ToDataType(ctx.Input<Tensor>("Filter")->type());
  PADDLE_ENFORCE_EQ(input_data_type, filter_data_type,
                    "input and filter data type should be consistent");

  if (input_data_type == framework::proto::VarType::FP16) {
126
    PADDLE_ENFORCE_EQ(library, framework::LibraryType::kCUDNN,
K
Kexin Zhao 已提交
127 128 129
                      "float16 can only be used when CUDNN is used");
  }

130 131
  return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
                                 library);
132 133
}

Y
Yu Yang 已提交
134
void Conv2DOpMaker::Make() {
K
Krzysztof Binias 已提交
135
  AddAttr<bool>("is_test", "").SetDefault(false);
C
chengduoZH 已提交
136 137
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
138 139 140 141
      "(Tensor) The input tensor of convolution operator. "
      "The format of input tensor is NCHW, where N is batch size, C is the "
      "number of channels, H is the height of the feature, "
      "and W is the width of the feature.");
C
chengduoZH 已提交
142
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
143
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
144 145
           "The format of the filter tensor is MCHW, where M is the number of "
           "output image channels, C is the number of input image channels, "
C
fix doc  
chengduoZH 已提交
146 147
           "H is the height of the filter, and W is the width of the filter. "
           "If the groups attribute is greater than 1, C equals the number of "
C
chengduoZH 已提交
148
           "input image channels divided by the groups.");
149 150 151 152 153
  AddInput("Bias",
           "(Tensor) Bias to be added to each output of filter application."
           "The format of output tensor is X (one-dimensional) of size equal"
           "to the number of output channels. Only used with MKL-DNN.")
      .AsDispensable();
C
chengduoZH 已提交
154
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
155
            "(Tensor) The output tensor of convolution operator. "
156
            "The format of output tensor is also NCHW.");
157 158 159
  AddInput("ResidualData",
           "(Tensor) Tensor with residual data "
           "to which convolution output will be added."
160
           "Used with fuse_residual_connection fusion.")
161
      .AsDispensable();
C
chengduoZH 已提交
162 163 164 165
  AddAttr<std::vector<int>>("strides",
                            "(vector<int> default:{1, 1}), the "
                            "strides(h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
166
      .SetDefault({1, 1});
C
chengduoZH 已提交
167 168 169 170
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int> default:{0, 0}), the "
                            "paddings(h_pad, w_pad) of "
                            "convolution operator.")
C
chengduoZH 已提交
171 172 173
      .SetDefault({0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
174
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
175 176 177 178
      "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
      "when group=2, the first half of the filters is only connected to the "
      "first half of the input channels, while the second half of the filters "
      "is only connected to the second half of the input channels.")
C
chengduoZH 已提交
179
      .SetDefault(1);
C
chengduoZH 已提交
180
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
181 182
                            "(vector<int> default:{1, 1}), the "
                            "dilations(h_dilation, w_dilation) of "
C
chengduoZH 已提交
183
                            "convolution operator.")
C
chengduoZH 已提交
184
      .SetDefault({1, 1});
185 186 187 188
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
189 190 191
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
M
Michal Gallus 已提交
192 193
  AddAttr<bool>("fuse_relu", "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
194
  AddAttr<bool>("fuse_residual_connection",
195
                "(bool, default false) Only used in mkldnn kernel. Used "
196 197
                "whenever convolution output is as an input to residual "
                "connection.")
198
      .SetDefault(false);
199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214
  AddAttr<std::string>(
      "data_format",
      "(string, default NCHW) Only used in "
      "An optional string from: \"NHWC\", \"NCHW\". "
      "Defaults to \"NHWC\". Specify the data format of the output data, "
      "the input will be transformed automatically. ")
      .SetDefault("AnyLayout");
  // TODO(dzhwinter): need to registered layout transform function
  AddAttr<int>("workspace_size_MB",
               "Only used in cudnn kernel. Need set use_cudnn to true."
               "workspace size for cudnn, in MB, "
               "workspace is a section of GPU memory which will be "
               "allocated/freed each time the operator runs, larger "
               "workspace size can increase performance but also requires "
               "better hardware. This size should be chosen carefully.")
      .SetDefault(4096);
C
chengduoZH 已提交
215
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
216 217
Convolution Operator.

C
chengduoZH 已提交
218
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
219
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
220
parameters is checked in the infer-shape.
C
chengduoZH 已提交
221
Input(Input) and Output(Output) are in NCHW format. Where N is batch
C
fix doc  
chengduoZH 已提交
222
size, C is the number of channels, H is the height of the feature, and W is
C
chengduoZH 已提交
223 224 225 226 227 228
the width of the feature.
Filters(Input) is MCHW format. Where M is the number of output image channels, C is
the number of input image channels, H is the height of the filter, and W
is the width of the filter.
Parameters(strides, paddings, dilations) are two elements. These two elements represent
height and width, respectively.
C
chengduoZH 已提交
229 230 231 232
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
233 234
       Input shape: $(N, C_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
chengduoZH 已提交
235
  Output:
C
chengduoZH 已提交
236 237 238 239 240 241
       Output shape: $(N, C_{out}, H_{out}, W_{out})$
  Where
$$
       H_{out}= \frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\
       W_{out}= \frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1
$$
C
chengduoZH 已提交
242
)DOC");
C
chengduoZH 已提交
243 244
}

Y
Yu Yang 已提交
245
void Conv3DOpMaker::Make() {
C
chengduoZH 已提交
246 247
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
248
      "(Tensor) The input tensor of convolution operator. "
C
chengduoZH 已提交
249
      "The format of input tensor is NCDHW. Where N is batch size, C is the "
C
fix doc  
chengduoZH 已提交
250 251 252
      "number of channels, D is the depth of the feature, H is the height of "
      "the feature, "
      "and W is the width of the feature.");
C
chengduoZH 已提交
253
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
254
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
255 256
           "The format of the filter tensor is MCDHW, where M is the number of "
           "output image channels, C is the number of input image channels, "
C
fix doc  
chengduoZH 已提交
257 258 259
           "D is the depth of the filter, H is the height of the filter, and W "
           "is the width of the filter."
           "If the groups attribute is greater than 1, C equals the number of "
C
chengduoZH 已提交
260 261
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
262
            "(Tensor) The output tensor of convolution operator."
263
            "The format of output tensor is also NCDHW.");
C
chengduoZH 已提交
264 265 266 267
  AddAttr<std::vector<int>>("strides",
                            "(vector<int>, default:{1, 1, 1}), the "
                            "strides(d_stride, h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
268
      .SetDefault({1, 1, 1});
C
chengduoZH 已提交
269 270 271 272
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int>, default:{0, 0, 0}), the "
                            "paddings(d_pad, h_pad, w_pad) of convolution "
                            "operator.")
C
chengduoZH 已提交
273 274 275
      .SetDefault({0, 0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
276
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
277 278 279 280
      "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
      "when group=2, the first half of the filters is only connected to the "
      "first half of the input channels, while the second half of the filters "
      "is only connected to the second half of the input channels.")
C
chengduoZH 已提交
281
      .SetDefault(1);
C
chengduoZH 已提交
282
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
283 284
                            "(vector<int> default:{1, 1, 1}), the "
                            "dilations(d_dilation, h_dilation, w_dilation) of "
C
chengduoZH 已提交
285
                            "convolution operator.")
C
chengduoZH 已提交
286
      .SetDefault({1, 1, 1});
287 288 289 290
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
291 292 293
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
294 295 296 297 298 299 300 301 302 303 304 305 306 307 308
  AddAttr<std::string>(
      "data_format",
      "(string, default NCHW) Only used in "
      "An optional string from: \"NHWC\", \"NCHW\". "
      "Defaults to \"NHWC\". Specify the data format of the output data, "
      "the input will be transformed automatically. ")
      .SetDefault("AnyLayout");
  // TODO(dzhwinter): need to registered layout transform function
  AddAttr<int>("workspace_size_MB",
               "Only used in cudnn kernel. workspace size for cudnn, in MB, "
               "workspace is a section of GPU memory which will be "
               "allocated/freed each time the operator runs, larger "
               "workspace size can increase performance but also requires "
               "better hardware. This size should be chosen carefully.")
      .SetDefault(4096);
C
fix doc  
chengduoZH 已提交
309

C
chengduoZH 已提交
310
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
311 312
Convolution3D Operator.

C
chengduoZH 已提交
313
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
314
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
315
parameters is checked in the infer-shape.
C
chengduoZH 已提交
316
Input(Input) and output(Output) are in NCDHW format, where N is batch
C
fix doc  
chengduoZH 已提交
317
size, C is the number of channels,D is the depth of the feature, H is the height of
C
chengduoZH 已提交
318 319 320 321 322 323
the feature, and W is the width of the feature.
Filters(Input) is MCDHW format, where M is the number of output image channels,
C is the number of input image channels, D is the depth of the filter,
H is the height of the filter, and W is the width of the filter.
Parameters(strides, paddings, dilations) are three elements. These three elements
represent depth, height and width, respectively.
C
fix doc  
chengduoZH 已提交
324 325 326 327
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
328 329
       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$
C
fix doc  
chengduoZH 已提交
330
  Output:
C
chengduoZH 已提交
331 332 333 334 335 336 337
       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
  Where
  $$
       D_{out}= \frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\
       H_{out}= \frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\
       W_{out}= \frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1
  $$
C
chengduoZH 已提交
338 339 340
)DOC");
}

C
chengduoZH 已提交
341 342 343 344 345 346 347 348 349 350 351
void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
  auto in_dims = ctx->GetInputDim("Input");
  auto filter_dims = ctx->GetInputDim("Filter");
  if (ctx->HasOutput(framework::GradVarName("Input"))) {
    ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
  }
  if (ctx->HasOutput(framework::GradVarName("Filter"))) {
    ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
  }
}

352 353
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
354
  framework::LibraryType library_{framework::LibraryType::kPlain};
M
mozga-intel 已提交
355 356 357 358
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
  std::string data_format = ctx.Attr<std::string>("data_format");
  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);

C
chengduoZH 已提交
359
#ifdef PADDLE_WITH_CUDA
360 361
  if (platform::CanCUDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
362 363
  }
#endif
364 365 366 367
#ifdef PADDLE_WITH_MKLDNN
  if (library_ == framework::LibraryType::kPlain &&
      platform::CanMKLDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kMKLDNN;
M
mozga-intel 已提交
368
    layout_ = framework::DataLayout::kMKLDNN;
369
  }
370
#endif
371 372 373 374 375 376

  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout_, library_);
}

C
chengduoZH 已提交
377 378 379 380
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Y
Yang Yang 已提交
381
REGISTER_OPERATOR(conv2d, ops::ConvOp, ops::Conv2DOpMaker,
382 383
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(conv2d_grad, ops::ConvOpGrad);
384 385

// depthwise convolution op
Y
Yang Yang 已提交
386
REGISTER_OPERATOR(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
387 388
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(depthwise_conv2d_grad, ops::ConvOpGrad);
Y
Yang Yang 已提交
389
REGISTER_OPERATOR(conv3d, ops::ConvOp, ops::Conv3DOpMaker,
390 391
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(conv3d_grad, ops::ConvOpGrad);
C
chengduoZH 已提交
392

393 394
// depthwise conv kernel
// TODO(xingzhaolong): neon kernel for mobile
Z
zlx 已提交
395
REGISTER_OP_CPU_KERNEL(
396
    depthwise_conv2d,
X
xzl 已提交
397 398 399 400
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

REGISTER_OP_CPU_KERNEL(
401
    depthwise_conv2d_grad,
X
xzl 已提交
402 403
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
Z
zlx 已提交
404

C
chengduoZH 已提交
405
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
406 407 408 409 410 411
    conv2d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    conv2d_grad,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
C
chengduoZH 已提交
412 413

REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
414 415 416 417 418 419
    conv3d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    conv3d_grad,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);