conv_op.cc 15.9 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. */
C
chengduoZH 已提交
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/conv_op.h"
16 17 18 19 20 21
#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
22 23 24 25

namespace paddle {
namespace operators {

C
chengduoZH 已提交
26
void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
C
chengduoZH 已提交
27
  PADDLE_ENFORCE(ctx->HasInput("Input"),
C
chengduoZH 已提交
28
                 "Input(Input) of ConvOp should not be null.");
C
chengduoZH 已提交
29
  PADDLE_ENFORCE(ctx->HasInput("Filter"),
C
chengduoZH 已提交
30
                 "Input(Filter) of ConvOp should not be null.");
C
chengduoZH 已提交
31
  PADDLE_ENFORCE(ctx->HasOutput("Output"),
C
chengduoZH 已提交
32
                 "Output(Output) of ConvOp should not be null.");
C
chengduoZH 已提交
33 34 35 36 37 38

  auto in_dims = ctx->GetInputDim("Input");
  auto filter_dims = ctx->GetInputDim("Filter");
  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
39
  std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
C
chengduoZH 已提交
40

C
chengduoZH 已提交
41 42
  PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
                 "Conv intput should be 4-D or 5-D tensor.");
C
chengduoZH 已提交
43 44 45 46 47 48 49 50 51
  PADDLE_ENFORCE_EQ(
      in_dims.size(), filter_dims.size(),
      "Conv input dimension and filter dimension should be the same.");
  PADDLE_ENFORCE(
      in_dims.size() - strides.size() == 2U,
      "Conv input dimension and strides dimension should be consistent.");
  PADDLE_ENFORCE_EQ(
      paddings.size(), strides.size(),
      "Conv paddings dimension and Conv strides dimension should be the same.");
F
fengjiayi 已提交
52

Y
Yang Yu 已提交
53
  PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups,
C
chengduoZH 已提交
54
                    "The number of input channels should be equal to filter "
C
chengduoZH 已提交
55
                    "channels * groups.");
F
fengjiayi 已提交
56

C
chengduoZH 已提交
57
  PADDLE_ENFORCE_EQ(
Y
Yang Yu 已提交
58
      filter_dims[0] % groups, 0,
C
chengduoZH 已提交
59 60 61
      "The number of output channels should be divided by groups.");

  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
C
chengduoZH 已提交
62
  for (size_t i = 0; i < strides.size(); ++i) {
Y
Yang Yang 已提交
63 64 65
    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
                                          dilations[i], paddings[i],
                                          strides[i]));
C
chengduoZH 已提交
66
  }
67
  ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
Y
Yang Yu 已提交
68
  ctx->ShareLoD("Input", "Output");
C
chengduoZH 已提交
69 70
}

71 72
framework::OpKernelType ConvOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
73
  framework::LibraryType library_{framework::LibraryType::kPlain};
C
chengduoZH 已提交
74
#ifdef PADDLE_WITH_CUDA
75 76
  if (platform::CanCUDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
77 78
  }
#endif
79 80 81 82
#ifdef PADDLE_WITH_MKLDNN
  if (library_ == framework::LibraryType::kPlain &&
      platform::CanMKLDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kMKLDNN;
83
  }
84
#endif
85

K
Kexin Zhao 已提交
86 87 88 89 90 91 92 93 94 95 96 97
  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) {
    PADDLE_ENFORCE_EQ(library_, framework::LibraryType::kCUDNN,
                      "float16 can only be used when CUDNN is used");
  }

98
  std::string data_format = ctx.Attr<std::string>("data_format");
99
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
100
  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
K
Kexin Zhao 已提交
101 102
  return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout_,
                                 library_);
103 104
}

105
Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
106 107 108
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
109 110 111 112
      "(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
113
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
114
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
115 116
           "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
117 118
           "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
119 120
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
121 122
            "(Tensor) The output tensor of convolution operator. "
            "The format of output tensor is also NCHW.");
C
chengduoZH 已提交
123 124 125 126
  AddAttr<std::vector<int>>("strides",
                            "(vector<int> default:{1, 1}), the "
                            "strides(h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
127
      .SetDefault({1, 1});
C
chengduoZH 已提交
128 129 130 131
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int> default:{0, 0}), the "
                            "paddings(h_pad, w_pad) of "
                            "convolution operator.")
C
chengduoZH 已提交
132 133 134
      .SetDefault({0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
135
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
136 137 138 139
      "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
140
      .SetDefault(1);
C
chengduoZH 已提交
141
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
142 143
                            "(vector<int> default:{1, 1}), the "
                            "dilations(h_dilation, w_dilation) of "
C
chengduoZH 已提交
144
                            "convolution operator.")
C
chengduoZH 已提交
145
      .SetDefault({1, 1});
146 147 148 149
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
150 151 152
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168
  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
169
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
170 171
Convolution Operator.

C
chengduoZH 已提交
172
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
173
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
174
parameters is checked in the infer-shape.
C
chengduoZH 已提交
175
Input(Input) and Output(Output) are in NCHW format. Where N is batch
C
fix doc  
chengduoZH 已提交
176
size, C is the number of channels, H is the height of the feature, and W is
C
chengduoZH 已提交
177 178 179 180 181 182
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
183 184 185 186
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
187 188
       Input shape: $(N, C_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
chengduoZH 已提交
189
  Output:
C
chengduoZH 已提交
190 191 192 193 194 195
       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
196
)DOC");
C
chengduoZH 已提交
197 198
}

199
Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
200 201 202
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
203
      "(Tensor) The input tensor of convolution operator. "
C
chengduoZH 已提交
204
      "The format of input tensor is NCDHW. Where N is batch size, C is the "
C
fix doc  
chengduoZH 已提交
205 206 207
      "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
208
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
209
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
210 211
           "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
212 213 214
           "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
215 216
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
217
            "(Tensor) The output tensor of convolution operator."
C
chengduoZH 已提交
218
            "The format of output tensor is also NCDHW.");
C
chengduoZH 已提交
219 220 221 222
  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
223
      .SetDefault({1, 1, 1});
C
chengduoZH 已提交
224 225 226 227
  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
228 229 230
      .SetDefault({0, 0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
231
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
232 233 234 235
      "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
236
      .SetDefault(1);
C
chengduoZH 已提交
237
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
238 239
                            "(vector<int> default:{1, 1, 1}), the "
                            "dilations(d_dilation, h_dilation, w_dilation) of "
C
chengduoZH 已提交
240
                            "convolution operator.")
C
chengduoZH 已提交
241
      .SetDefault({1, 1, 1});
242 243 244 245
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
246 247 248
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
249 250 251 252 253 254 255 256 257 258 259 260 261 262 263
  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
264

C
chengduoZH 已提交
265
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
266 267
Convolution3D Operator.

C
chengduoZH 已提交
268
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
269
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
270
parameters is checked in the infer-shape.
C
chengduoZH 已提交
271
Input(Input) and output(Output) are in NCDHW format, where N is batch
C
fix doc  
chengduoZH 已提交
272
size, C is the number of channels,D is the depth of the feature, H is the height of
C
chengduoZH 已提交
273 274 275 276 277 278
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
279 280 281 282
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
283 284
       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
285
  Output:
C
chengduoZH 已提交
286 287 288 289 290 291 292
       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
293 294 295
)DOC");
}

C
chengduoZH 已提交
296 297 298 299 300 301 302 303 304 305 306
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);
  }
}

307 308
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
309
  framework::LibraryType library_{framework::LibraryType::kPlain};
C
chengduoZH 已提交
310
#ifdef PADDLE_WITH_CUDA
311 312
  if (platform::CanCUDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
313 314
  }
#endif
315 316 317 318
#ifdef PADDLE_WITH_MKLDNN
  if (library_ == framework::LibraryType::kPlain &&
      platform::CanMKLDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kMKLDNN;
319
  }
320
#endif
321 322

  std::string data_format = ctx.Attr<std::string>("data_format");
323
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
324 325 326 327 328 329
  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout_, library_);
}

C
chengduoZH 已提交
330 331 332 333
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
C
chengduoZH 已提交
334 335
REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
            ops::ConvOpGrad);
336 337

// depthwise convolution op
338 339
REGISTER_OP(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
            depthwise_conv2d_grad, ops::ConvOpGrad);
C
chengduoZH 已提交
340 341 342
REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
            ops::ConvOpGrad);

343 344
// depthwise conv kernel
// TODO(xingzhaolong): neon kernel for mobile
Z
zlx 已提交
345
REGISTER_OP_CPU_KERNEL(
346
    depthwise_conv2d,
X
xzl 已提交
347 348 349 350
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

REGISTER_OP_CPU_KERNEL(
351
    depthwise_conv2d_grad,
X
xzl 已提交
352 353
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
Z
zlx 已提交
354

C
chengduoZH 已提交
355
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
356 357 358 359 360 361
    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
362 363

REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
364 365 366 367 368 369
    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>);