transpose_op.cc 16.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
X
xzl 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
X
xzl 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
X
xzl 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. */
X
xzl 已提交
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/transpose_op.h"
16
#include <memory>
17
#include <string>
18
#include <vector>
X
xzl 已提交
19

20 21 22 23
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif

X
xzl 已提交
24 25 26 27 28 29 30 31 32
namespace paddle {
namespace operators {

using framework::Tensor;

class TransposeOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

33
  void InferShape(framework::InferShapeContext *ctx) const override {
34 35
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "Transpose");
    OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "Transpose");
Q
Qiao Longfei 已提交
36 37
    auto x_dims = ctx->GetInputDim("X");
    std::vector<int> axis = ctx->Attrs().Get<std::vector<int>>("axis");
X
xzl 已提交
38
    size_t x_rank = x_dims.size();
X
xzl 已提交
39
    size_t axis_size = axis.size();
X
xzl 已提交
40

X
xzl 已提交
41
    PADDLE_ENFORCE_EQ(x_rank, axis_size,
42 43 44 45 46 47
                      platform::errors::InvalidArgument(
                          "The input tensor's dimension "
                          "should be equal to the axis's size. "
                          "But received input tensor's dimension is %d, "
                          "axis's size is %d",
                          x_rank, axis_size));
48 49 50

    std::vector<int> count(axis_size, 0);
    for (size_t i = 0; i < axis_size; i++) {
51 52 53 54 55 56
      PADDLE_ENFORCE_GE(axis[i], 0,
                        platform::errors::InvalidArgument(
                            "The axis should be greater than or equal to 0."
                            "But received %d of axis[%d]",
                            axis[i], i));

57 58 59 60 61 62 63 64 65 66
      PADDLE_ENFORCE_EQ(
          axis[i] < static_cast<int>(axis_size) && ++count[axis[i]] == 1, true,
          platform::errors::InvalidArgument(
              "Each element of Attribute axis should "
              "be a unique value range from 0 to (dims - 1), "
              "where the dims is the axis's size, "
              "unique value means this axis value can appear only once. "
              "But received axis[%d] is %d, axis_size is %d, "
              "count[axis[%d]] is %d",
              i, axis[i], axis_size, i, count[axis[i]]));
X
xzl 已提交
67
    }
X
xzl 已提交
68

X
xzl 已提交
69
    framework::DDim out_dims(x_dims);
J
Jacek Czaja 已提交
70 71 72 73 74 75 76 77 78 79 80 81 82
#ifdef PADDLE_WITH_MKLDNN
    // Here we need to match dims to paddle layout
    // as we are producing non-oneDNN result
    if ((x_dims.size() >= 3) &&
        (paddle::platform::MKLDNNDeviceContext::tls()
             .get_cur_paddle_data_layout() == framework::DataLayout::kNHWC)) {
      auto dims = framework::vectorize<int>(x_dims);
      std::rotate(dims.begin() + 1, dims.begin() + 2, dims.end());
      x_dims = x_dims.reshape(dims);
      VLOG(3)
          << "Rotating Shape in Transpose from: kMKLDNN to: kNHWC output_shape";
    }
#endif
83
    for (size_t i = 0; i < axis_size; i++) {
X
xzl 已提交
84
      out_dims[i] = x_dims[axis[i]];
X
xzl 已提交
85
    }
Q
Qiao Longfei 已提交
86
    ctx->SetOutputDim("Out", out_dims);
X
xzl 已提交
87
  }
88 89 90 91 92 93 94

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    framework::LibraryType library_{framework::LibraryType::kPlain};
    std::string data_format = ctx.Attr<std::string>("data_format");
    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
95
    auto data_type = OperatorWithKernel::IndicateVarDataType(ctx, "X");
96 97
#ifdef PADDLE_WITH_MKLDNN
    if (library_ == framework::LibraryType::kPlain &&
98
        this->CanMKLDNNBeUsed(ctx, data_type)) {
99 100 101 102
      library_ = framework::LibraryType::kMKLDNN;
      layout_ = framework::DataLayout::kMKLDNN;
    }
#endif
103 104
    return framework::OpKernelType(data_type, ctx.GetPlace(), layout_,
                                   library_);
105
  }
X
xzl 已提交
106 107 108 109
};

class TransposeOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
110
  void Make() override {
111
    AddInput(
X
xzl 已提交
112
        "X",
113 114
        "(Tensor) The input tensor, tensors with rank up to 6 are supported.");
    AddOutput("Out", "(Tensor)The output tensor.");
X
xzl 已提交
115 116
    AddAttr<std::vector<int>>(
        "axis",
117 118 119
        "(vector<int>) A list of values, and the size of the list should be "
        "the same with the input tensor rank. This operator permutes the input "
        "tensor's axes according to the values given.");
120 121
    AddAttr<bool>("use_mkldnn",
                  "(bool, default false) Only used in mkldnn kernel")
122 123
        .SetDefault(false)
        .AsExtra();
124 125 126 127 128 129
    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. ")
130 131
        .SetDefault("AnyLayout")
        .AsExtra();
132 133 134 135
    AddAttr<bool>(
        "use_quantizer",
        "(bool, default false) "
        "This parameter is no longer used. Use 'mkldnn_data_type' instead.")
136 137
        .SetDefault(false)
        .AsExtra();
138 139 140 141
    AddAttr<std::string>(
        "mkldnn_data_type",
        "(string, default \"float32\"). Data type of mkldnn kernel")
        .SetDefault("float32")
142 143
        .InEnum({"float32", "int8", "bfloat16"})
        .AsExtra();
144
    /* int8 parameters */
X
xzl 已提交
145
    AddComment(R"DOC(
146 147
Transpose Operator.

148 149
The input tensor will be permuted according to the axes given.
The behavior of this operator is similar to how `numpy.transpose` works.
Y
ying 已提交
150

151 152 153 154 155 156
- suppose the input `X` is a 2-D tensor:
    $$
    X = \begin{pmatrix}
    0 &1 &2 \\
    3 &4 &5
    \end{pmatrix}$$
W
wanghaoshuang 已提交
157

158
    the given `axes` is: $[1, 0]$, and $Y$ = transpose($X$, axis)
W
wanghaoshuang 已提交
159

160
    then the output $Y$ is:
W
wanghaoshuang 已提交
161

162 163 164 165 166 167
    $$
    Y = \begin{pmatrix}
         0 &3 \\
         1 &4  \\
         2 &5
    \end{pmatrix}$$
W
wanghaoshuang 已提交
168

169
- Given a input tensor with shape $(N, C, H, W)$ and the `axes` is
170
$[0, 2, 3, 1]$, then shape of the output tensor will be: $(N, H, W, C)$.
171

X
xzl 已提交
172 173 174 175 176 177 178 179
)DOC");
  }
};

class TransposeOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

180
  void InferShape(framework::InferShapeContext *ctx) const override {
181 182 183
    OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "TransposeOpGrad");
    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
                   framework::GradVarName("Out"), "TransposeOpGrad");
Q
Qiao Longfei 已提交
184 185 186 187 188
    auto x_dims = ctx->GetInputDim("X");
    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
    if (ctx->HasOutput(framework::GradVarName("X"))) {
      ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
    }
X
xzl 已提交
189
  }
190 191 192 193 194 195 196

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    framework::LibraryType library_{framework::LibraryType::kPlain};
    std::string data_format = ctx.Attr<std::string>("data_format");
    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
197 198
    auto data_type = OperatorWithKernel::IndicateVarDataType(
        ctx, framework::GradVarName("Out"));
199 200
#ifdef PADDLE_WITH_MKLDNN
    if (library_ == framework::LibraryType::kPlain &&
201
        this->CanMKLDNNBeUsed(ctx, data_type)) {
202 203 204 205
      library_ = framework::LibraryType::kMKLDNN;
      layout_ = framework::DataLayout::kMKLDNN;
    }
#endif
206 207
    return framework::OpKernelType(data_type, ctx.GetPlace(), layout_,
                                   library_);
208
  }
X
xzl 已提交
209 210
};

211 212 213 214 215 216 217 218 219 220 221 222 223 224 225
// FIXME(zcd): transpose2 adds an intermediate output(XShape) based on
// transpose, the XShape is used to carry the shape and lod of X which
// will be used in transpose_grad, in this way, the framework can reuse
// the memory of X immediately the transpose2_op is finished.
// Considering compatibility issues, we could not fix transpose2_op
class Transpose2Op : public TransposeOp {
 public:
  Transpose2Op(const std::string &type,
               const framework::VariableNameMap &inputs,
               const framework::VariableNameMap &outputs,
               const framework::AttributeMap &attrs)
      : TransposeOp(type, inputs, outputs, attrs) {}

  void InferShape(framework::InferShapeContext *ctx) const override {
    TransposeOp::InferShape(ctx);
226
    if (!ctx->HasOutput("XShape")) return;
227 228 229 230 231 232 233 234 235 236 237 238 239
    const auto &in_dims = ctx->GetInputDim("X");
    std::vector<int64_t> x_shape_dim(in_dims.size() + 1);
    x_shape_dim[0] = 0;
    for (int i = 0; i < in_dims.size(); ++i) {
      x_shape_dim[i + 1] = in_dims[i];
    }
    ctx->SetOutputDim("XShape", framework::make_ddim(x_shape_dim));
    ctx->ShareLoD("X", /*->*/ "XShape");
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
240 241
    framework::LibraryType library_{framework::LibraryType::kPlain};
    std::string data_format = ctx.Attr<std::string>("data_format");
242 243
    int customized_type_value =
        framework::OpKernelType::kDefaultCustomizedTypeValue;
244
    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
245 246
    framework::proto::VarType::Type data_type =
        OperatorWithKernel::IndicateVarDataType(ctx, "X");
247 248
#ifdef PADDLE_WITH_MKLDNN
    if (library_ == framework::LibraryType::kPlain &&
249
        this->CanMKLDNNBeUsed(ctx, data_type)) {
250 251
      library_ = framework::LibraryType::kMKLDNN;
      layout_ = framework::DataLayout::kMKLDNN;
252 253 254 255 256 257
      using framework::proto::VarType;
      auto input_data_type = ctx.Input<Tensor>("X")->type();
      customized_type_value = (input_data_type == VarType::INT8 ||
                               input_data_type == VarType::UINT8)
                                  ? kTransposeMKLDNNINT8
                                  : kTransposeMKLDNNFP32;
258 259
    }
#endif
260 261
    return framework::OpKernelType(data_type, ctx.GetPlace(), layout_, library_,
                                   customized_type_value);
262 263 264 265 266 267 268
  }
};

class Transpose2OpMaker : public TransposeOpMaker {
 public:
  void Make() override {
    TransposeOpMaker::Make();
269 270 271
    AddOutput("XShape", "(Tensor)The output tensor.")
        .AsIntermediate()
        .AsExtra();
272 273 274
  }
};

H
hong 已提交
275 276
template <typename T>
class Transpose2GradMaker : public framework::SingleGradOpMaker<T> {
277
 public:
H
hong 已提交
278
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
279

280
  void Apply(GradOpPtr<T> grad_op) const override {
281
    grad_op->SetType("transpose2_grad");
H
hong 已提交
282 283 284 285
    grad_op->SetInput("XShape", this->Output("XShape"));
    grad_op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    grad_op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    grad_op->SetAttrMap(this->Attrs());
286 287 288
  }
};

289 290 291 292 293 294 295 296 297 298 299 300 301 302
template <typename T>
class Transpose2DoubleGradMaker : public framework::SingleGradOpMaker<T> {
 public:
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;

  void Apply(GradOpPtr<T> grad_op) const override {
    grad_op->SetType("transpose2");
    grad_op->SetInput("X", this->OutputGrad(framework::GradVarName("X")));
    grad_op->SetOutput("Out", this->InputGrad(framework::GradVarName("Out")));
    grad_op->SetOutput("XShape", this->Input("XShape"));
    grad_op->SetAttrMap(this->Attrs());
  }
};

303 304 305 306 307
class Transpose2OpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext *ctx) const override {
308 309 310 311
    OP_INOUT_CHECK(ctx->HasInput("XShape"), "Input", "XShape",
                   "Transpose2OpGrad");
    OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), "Input",
                   framework::GradVarName("Out"), "Transpose2OpGrad");
312 313 314 315 316 317 318 319 320 321 322 323
    if (ctx->HasOutput(framework::GradVarName("X"))) {
      auto xshape_dim = ctx->GetInputDim("XShape");
      auto x_shape_dim =
          framework::slice_ddim(xshape_dim, 1, xshape_dim.size());
      ctx->SetOutputDim(framework::GradVarName("X"), x_shape_dim);
      ctx->ShareLoD("XShape", framework::GradVarName("X"));
    }
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
324 325 326
    framework::LibraryType library_{framework::LibraryType::kPlain};
    std::string data_format = ctx.Attr<std::string>("data_format");
    framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
327 328 329
    framework::proto::VarType::Type data_type =
        OperatorWithKernel::IndicateVarDataType(ctx,
                                                framework::GradVarName("Out"));
330 331
#ifdef PADDLE_WITH_MKLDNN
    if (library_ == framework::LibraryType::kPlain &&
332
        this->CanMKLDNNBeUsed(ctx, data_type)) {
333 334 335 336
      library_ = framework::LibraryType::kMKLDNN;
      layout_ = framework::DataLayout::kMKLDNN;
    }
#endif
337 338
    return framework::OpKernelType(data_type, ctx.GetPlace(), layout_,
                                   library_);
339 340 341
  }
};

X
xzl 已提交
342 343 344 345
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
H
hong 已提交
346 347 348 349
REGISTER_OPERATOR(
    transpose, ops::TransposeOp, ops::TransposeOpMaker,
    paddle::framework::DefaultGradOpMaker<paddle::framework::OpDesc, true>,
    paddle::framework::DefaultGradOpMaker<paddle::imperative::OpBase, true>);
350
REGISTER_OPERATOR(transpose_grad, ops::TransposeOpGrad);
351

Q
QI JUN 已提交
352
REGISTER_OP_CPU_KERNEL(
353 354
    transpose, ops::TransposeKernel<paddle::platform::CPUDeviceContext, bool>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>,
355 356
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, double>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
357
                         paddle::platform::complex<float>>,
358
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
359 360 361
                         paddle::platform::complex<double>>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
                         paddle::platform::bfloat16>);
X
xzl 已提交
362 363
REGISTER_OP_CPU_KERNEL(
    transpose_grad,
364
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, bool>,
P
phlrain 已提交
365
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
366 367
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, double>,
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
368
                             paddle::platform::complex<float>>,
369
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
370 371 372
                             paddle::platform::complex<double>>,
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
                             paddle::platform::bfloat16>);
373 374

REGISTER_OPERATOR(transpose2, ops::Transpose2Op, ops::Transpose2OpMaker,
H
hong 已提交
375 376
                  ops::Transpose2GradMaker<paddle::framework::OpDesc>,
                  ops::Transpose2GradMaker<paddle::imperative::OpBase>);
377 378 379
REGISTER_OPERATOR(transpose2_grad, ops::Transpose2OpGrad,
                  ops::Transpose2DoubleGradMaker<paddle::framework::OpDesc>,
                  ops::Transpose2DoubleGradMaker<paddle::imperative::OpBase>);
380 381

REGISTER_OP_CPU_KERNEL(
382 383
    transpose2, ops::TransposeKernel<paddle::platform::CPUDeviceContext, bool>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, float>,
384 385
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, int32_t>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, int64_t>,
386 387
    ops::TransposeKernel<paddle::platform::CPUDeviceContext, double>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
388
                         paddle::platform::complex<float>>,
389
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
390 391 392
                         paddle::platform::complex<double>>,
    ops::TransposeKernel<paddle::platform::CPUDeviceContext,
                         paddle::platform::bfloat16>);
393 394
REGISTER_OP_CPU_KERNEL(
    transpose2_grad,
395
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, bool>,
396 397
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, int32_t>,
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, int64_t>,
P
phlrain 已提交
398
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, float>,
399 400
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext, double>,
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
401
                             paddle::platform::complex<float>>,
402
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
403 404 405
                             paddle::platform::complex<double>>,
    ops::TransposeGradKernel<paddle::platform::CPUDeviceContext,
                             paddle::platform::bfloat16>);