matmul_op.cc 21.4 KB
Newer Older
1
/* Copyright (c) 2017 PaddlePaddle Authors. All Rights Reserved.
M
Markus Kliegl 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

15
#include <algorithm>
16
#include <utility>
17
#include <vector>
18 19
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/blas.h"
20 21 22
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
M
Markus Kliegl 已提交
23 24 25

namespace paddle {
namespace operators {
26 27 28 29 30 31 32 33 34 35 36

/**
 * Printing shape information into a string is easy to use.
 */
inline static std::string DumpMatrixShape(const math::MatDescriptor &desc) {
  std::stringstream buffer;
  buffer << "[" << desc.batch_size_ << ", " << desc.height_ << ", "
         << desc.width_ << "]";
  return buffer.str();
}

37 38 39 40
/**
 * Get row matrix shape from a vector shape. If the rank of x_dim > 1, the
 * original x_dim is returned.
 */
41
static framework::DDim RowMatrixFromVector(const framework::DDim &x_dim) {
42 43 44 45 46 47 48 49 50 51
  if (x_dim.size() > 1) {
    return x_dim;
  }
  return framework::make_ddim({1, x_dim[0]});
}

/**
 * Get column matrix shape from a vector shape. If the ran of y_dim > 1, the
 * original y_dim is returned.
 */
52
static framework::DDim ColumnMatrixFromVector(const framework::DDim &y_dim) {
53 54 55 56 57 58 59 60 61
  if (y_dim.size() > 1) {
    return y_dim;
  }
  return framework::make_ddim({y_dim[0], 1});
}

template <typename DeviceContext, typename T>
class MatMulKernel : public framework::OpKernel<T> {
 public:
62
  void Compute(const framework::ExecutionContext &context) const override {
63 64 65 66
    auto &x = GET_DATA_SAFELY(context.Input<framework::Tensor>("X"), "Input",
                              "X", "MatMul");
    auto &y = GET_DATA_SAFELY(context.Input<framework::Tensor>("Y"), "Input",
                              "Y", "MatMul");
67
    auto *out = context.Output<framework::Tensor>("Out");
68 69 70 71 72 73 74
    out->mutable_data<T>(context.GetPlace());

    auto blas = math::GetBlas<DeviceContext, T>(context);
    auto mat_dim_a = math::CreateMatrixDescriptor(
        RowMatrixFromVector(x.dims()), 0, context.Attr<bool>("transpose_X"));
    auto mat_dim_b = math::CreateMatrixDescriptor(
        ColumnMatrixFromVector(y.dims()), 0, context.Attr<bool>("transpose_Y"));
S
sneaxiy 已提交
75
    auto scale = static_cast<T>(context.Attr<float>("alpha"));
76

77 78 79 80 81 82 83 84 85 86 87 88 89 90
    int head_number = 1;
#if defined(PADDLE_WITH_MKLML) && !defined(PADDLE_WITH_CUDA)
    head_number = context.Attr<int>("head_number");
#endif

    const auto &x_dims = x.dims();
    const auto &y_dims = y.dims();
    if (head_number <= 1 && x_dims.size() == 3 && y_dims.size() <= 2) {
      // the transpose_X must be false, if is true, the transpose cost much time
      if (!context.Attr<bool>("transpose_X")) {
        mat_dim_a.height_ *= mat_dim_a.batch_size_;
        mat_dim_a.batch_size_ = 0;
      }
    }
91
#if defined(PADDLE_WITH_MKLML) && !defined(PADDLE_WITH_CUDA)
92 93 94
    bool split_vertical_y = (mat_dim_a.width_ != mat_dim_b.height_);

    if (head_number > 1) {
95
      blas.MatMulWithHead(x, mat_dim_a, y, mat_dim_b, scale, head_number, out,
96 97 98
                          T(0), split_vertical_y);
    } else {
      blas.MatMul(x, mat_dim_a, y, mat_dim_b, scale, out, T(0));
99 100
    }
#else
S
sneaxiy 已提交
101
    blas.MatMul(x, mat_dim_a, y, mat_dim_b, scale, out, T(0));
102
#endif
103 104 105 106 107
  }
};

// Reshape a rank-3 tensor from P x M x N to (P * M) x N.
// Identity op if the tensor is not of rank 3.
108
static framework::Tensor FoldInitDims(const framework::Tensor &input) {
109 110 111 112 113 114 115 116 117 118 119 120
  auto output = input;
  auto in_dims = input.dims();
  if (in_dims.size() == 3) {
    output.Resize({in_dims[0] * in_dims[1], in_dims[2]});
  }
  return output;
}

// Reshape a rank-3 tensor from P x M x N to M x (P * N).
// (Warning: This requires transposing data and writes into new memory.)
// Identity op if the tensor is not of rank 3.
template <typename DeviceContext, typename T>
121 122
static framework::Tensor FoldHeadAndLastDims(const DeviceContext &context,
                                             const framework::Tensor &input) {
123 124 125 126 127 128 129 130 131 132 133
  auto in_dims = input.dims();
  if (in_dims.size() != 3) {
    return input;
  }
  framework::Tensor output;
  output.Resize({in_dims[1], in_dims[0], in_dims[2]});
  output.mutable_data<T>(context.GetPlace());
  std::vector<int> axis = {1, 0, 2};
  math::Transpose<DeviceContext, T, 3> trans;
  trans(context, input, &output, axis);
  output.Resize({in_dims[1], in_dims[0] * in_dims[2]});
M
Markus Kliegl 已提交
134

135 136 137 138 139 140 141 142 143 144
  return output;
}

/**
 * Reshape a tensor to 3-D or 2-D tensor by matrix descriptor.
 *
 * The shape would be [BatchSize, H, W] or [H, W].
 * If transposed, `H,W` will be swapped.
 */
static void ReshapeTensorIntoMatrixSequence(
145
    framework::Tensor *x, const math::MatDescriptor &descriptor) {
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
  int64_t h, w;
  h = descriptor.height_;
  w = descriptor.width_;
  if (descriptor.trans_) {
    std::swap(w, h);
  }
  if (descriptor.batch_size_) {
    x->Resize({descriptor.batch_size_, h, w});
  } else {
    x->Resize({h, w});
  }
}

/**
 * Reshape the x,y,out tensor to 3-D or 2-D tensor by matrix descriptor
 * Out = matmul(x, y)
 *
 * This method will first calculate X,Y matrix sequence, and then calculate
 * the out shape.
 *
 * Assume X = [BatchSize, H1, W1], Y = [BatchSize, H2, W2]
 * The out = [BatchSize, H1, W2]
 *
 * If there is no batch size in `X` and `Y`, the out will be [H1, W2]
 * If any of `X` and `Y` has batch size BatchSize, the out will have the
 * BatchSize.
 */
173 174 175
static void ReshapeXYOutIntoMatrixSequence(framework::Tensor *x,
                                           framework::Tensor *y,
                                           framework::Tensor *out, bool trans_x,
176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219
                                           bool trans_y) {
  auto x_dim = RowMatrixFromVector(x->dims());
  auto y_dim = ColumnMatrixFromVector(y->dims());
  auto mat_dim_x = math::CreateMatrixDescriptor(x_dim, 0, trans_x);
  auto mat_dim_y = math::CreateMatrixDescriptor(y_dim, 0, trans_y);
  if (mat_dim_x.batch_size_ == 0 && mat_dim_y.batch_size_ == 0) {
    out->Resize({mat_dim_x.height_, mat_dim_y.width_});
  } else {
    out->Resize({std::max(mat_dim_x.batch_size_, mat_dim_y.batch_size_),
                 mat_dim_x.height_, mat_dim_y.width_});
  }

  ReshapeTensorIntoMatrixSequence(x, mat_dim_x);
  ReshapeTensorIntoMatrixSequence(y, mat_dim_y);
}

// Using dimensional constraints on matrix multiplication, it is
// straight-forward to check the following table for when X and Y
// are both matrices.
//
// transpose_X | False    | True     | False    | True
// transpose_Y | False    | False    | True     | True
// -----------+----------+----------+----------+-----------
//        dX = | dOut Y^T | Y dOut^T | dOut Y   | Y^T dOut^T
//        dY = | X^T dOut | X dOut   | dOut^T X | dOut^T X^T
//
// When X is a vector of size K, we treat it instead as a matrix of shape
// (1, K). Similarly, when Y is a vector of size K, we treat it instead as
// a matrix of shape (K, 1).
//
// When X and Y are both 3-dimensional tensors, then the first dimension
// the batch dimension can be ignored and the exact same formulas apply
// as for two matrices.
//
// Finally, when, e.g., X is a 3-dimensional tensor but Y is a matrix, we end
// up with formulas like
//
//   dY_{ij} = \sum_{p, m} X_{pmi} dOut_{pmj}
//
// To handle this sort of scenario, we reshape X : P x M x K, dOut: P x M x N
// to X: (P * M) x K, dOut: (P * M) x N.
template <typename DeviceContext, typename T>
class MatMulGradKernel : public framework::OpKernel<T> {
 public:
220 221 222 223
  void MatMul(const framework::ExecutionContext &context,
              const framework::Tensor &a, bool trans_a,
              const framework::Tensor &b, bool trans_b,
              framework::Tensor *out) const {
224 225 226 227
    out->mutable_data<T>(context.GetPlace());
    auto blas = math::GetBlas<DeviceContext, T>(context);
    auto mat_dim_a = math::CreateMatrixDescriptor(a.dims(), 0, trans_a);
    auto mat_dim_b = math::CreateMatrixDescriptor(b.dims(), 0, trans_b);
228 229 230 231 232 233 234 235 236 237 238 239 240

    int head_number = 1;
#if defined(PADDLE_WITH_MKLML) && !defined(PADDLE_WITH_CUDA)
    head_number = context.Attr<int>("head_number");
#endif

    if (head_number <= 1 && a.dims().size() == 3 && b.dims().size() <= 2) {
      // the transpose_X must be false, if is true, the transpose cost much time
      if (!trans_a) {
        mat_dim_a.height_ *= mat_dim_a.batch_size_;
        mat_dim_a.batch_size_ = 0;
      }
    }
S
sneaxiy 已提交
241
    blas.MatMul(a, mat_dim_a, b, mat_dim_b,
S
sneaxiy 已提交
242
                static_cast<T>(context.Attr<float>("alpha")), out, T(0));
243 244
  }

245 246 247
  void CalcInputGrad(const framework::ExecutionContext &context,
                     const framework::Tensor &a, bool trans_a,
                     bool is_fold_init_dims_a, const framework::Tensor &b,
248
                     bool trans_b, bool is_fold_init_dims_b,
249
                     framework::Tensor *out) const {
250 251 252 253 254 255
    if (out == nullptr) return;
    bool need_combine = (a.dims().size() == 3 || b.dims().size() == 3) &&
                        out->dims().size() == 2;
    if (!need_combine) {
      MatMul(context, a, trans_a, b, trans_b, out);
    } else {
256
      auto &ctx = context.template device_context<DeviceContext>();
257 258 259 260 261 262 263 264 265 266
      MatMul(context, is_fold_init_dims_a
                          ? FoldInitDims(a)
                          : FoldHeadAndLastDims<DeviceContext, T>(ctx, a),
             trans_a, is_fold_init_dims_b
                          ? FoldInitDims(b)
                          : FoldHeadAndLastDims<DeviceContext, T>(ctx, b),
             trans_b, out);
    }
  }

267
  void Compute(const framework::ExecutionContext &context) const override {
268 269 270 271
    auto x = *context.Input<framework::Tensor>("X");
    auto y = *context.Input<framework::Tensor>("Y");
    auto dout =
        *context.Input<framework::Tensor>(framework::GradVarName("Out"));
272 273
    auto *dx = context.Output<framework::Tensor>(framework::GradVarName("X"));
    auto *dy = context.Output<framework::Tensor>(framework::GradVarName("Y"));
274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319
    bool transpose_x = context.Attr<bool>("transpose_X");
    bool transpose_y = context.Attr<bool>("transpose_Y");

    ReshapeXYOutIntoMatrixSequence(&x, &y, &dout, transpose_x, transpose_y);
    framework::DDim dx_dims;
    if (dx) {
      dx_dims = dx->dims();
      if (dx_dims != x.dims()) {
        dx->Resize(x.dims());
      }
    }

    framework::DDim dy_dims;
    if (dy) {
      dy_dims = dy->dims();
      if (dy_dims != y.dims()) {
        dy->Resize(y.dims());
      }
    }

    if (transpose_x && transpose_y) {
      CalcInputGrad(context, y, true, true, dout, true, false, dx);
      CalcInputGrad(context, dout, true, true, x, true, false, dy);
    } else if (transpose_x) {
      CalcInputGrad(context, y, false, false, dout, true, false, dx);
      CalcInputGrad(context, x, false, false, dout, false, true, dy);
    } else if (transpose_y) {
      CalcInputGrad(context, dout, false, false, y, false, true, dx);
      CalcInputGrad(context, dout, true, true, x, false, true, dy);
    } else {
      CalcInputGrad(context, dout, false, false, y, true, false, dx);
      CalcInputGrad(context, x, true, true, dout, false, true, dy);
    }

    if (dx) {
      if (dx_dims != x.dims()) {
        dx->Resize(dx_dims);
      }
    }
    if (dy) {
      if (dy_dims != y.dims()) {
        dy->Resize(dy_dims);
      }
    }
  }
};
M
Markus Kliegl 已提交
320 321 322 323 324 325

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

 protected:
326
  void InferShape(framework::InferShapeContext *context) const override {
M
Markus Kliegl 已提交
327 328 329 330 331 332 333 334 335 336
    PADDLE_ENFORCE(context->HasInput("X"),
                   "Input(X) of MatMulOp should not be null.");
    PADDLE_ENFORCE(context->HasInput("Y"),
                   "Input(Y) of MatMulOp should not be null.");
    PADDLE_ENFORCE(context->HasOutput("Out"),
                   "Output(Out) of MatMulOp should not be null.");

    auto dim_x = context->GetInputDim("X");
    auto dim_y = context->GetInputDim("Y");

337 338
    auto mat_dim_x =
        math::CreateMatrixDescriptor(RowMatrixFromVector(dim_x), 0,
339
                                     context->Attrs().Get<bool>("transpose_X"));
340 341
    auto mat_dim_y =
        math::CreateMatrixDescriptor(ColumnMatrixFromVector(dim_y), 0,
342
                                     context->Attrs().Get<bool>("transpose_Y"));
C
chengduoZH 已提交
343

344 345 346 347 348 349 350
    if (mat_dim_x.width_ == -1) {
      mat_dim_x.width_ = mat_dim_y.height_;
    }
    if (mat_dim_y.height_ == -1) {
      mat_dim_y.height_ = mat_dim_x.width_;
    }

351
    if (context->IsRuntime()) {
352 353 354 355 356 357 358 359
      PADDLE_ENFORCE(
          mat_dim_x.batch_size_ == mat_dim_y.batch_size_ ||
              mat_dim_x.batch_size_ == 0 || mat_dim_y.batch_size_ == 0,
          "ShapeError: The batch size of the two matrices should be equal, or "
          "at least one is zero.\n"
          "But received X's shape: %s, Y's shape: %s.",
          DumpMatrixShape(mat_dim_x).c_str(),
          DumpMatrixShape(mat_dim_y).c_str());
360
    }
361
    int64_t dim_out_y = mat_dim_y.width_;
362 363
#if defined(PADDLE_WITH_MKLML) && !defined(PADDLE_WITH_CUDA)
    int head_number = context->Attrs().Get<int>("head_number");
364
    bool split_vertical_y = (mat_dim_x.width_ != mat_dim_y.height_);
365 366 367 368 369 370 371 372 373 374 375
    if (context->IsRuntime()) {
      PADDLE_ENFORCE_LE(
          head_number, mat_dim_x.width_,
          "ShapeError: Unsatisfied mkl acceleration library requirements: "
          "The number of heads "
          "(%d) must be equal to X's width. But received X's shape: %s.",
          head_number, DumpMatrixShape(mat_dim_x).c_str());

      if (!split_vertical_y && head_number > 0) {
        dim_out_y = head_number * mat_dim_y.width_;
      }
376
    }
377
#else
378 379
    PADDLE_ENFORCE_EQ(
        mat_dim_x.width_, mat_dim_y.height_,
380 381 382 383 384
        platform::errors::InvalidArgument(
            "ShapeError: Input X's width should be equal to the Y's height, "
            "but received X's shape: [%s],"
            "Y's shape: [%s].",
            dim_x, dim_y));
385 386
#endif

387
    std::vector<int64_t> dim_out;
388 389 390
    if (mat_dim_x.batch_size_ != 0) {
      dim_out = framework::vectorize(dim_x);
      dim_out[dim_out.size() - 2] = mat_dim_x.height_;
391
      dim_out[dim_out.size() - 1] = dim_out_y;
392 393 394
    } else if (mat_dim_y.batch_size_ != 0) {
      dim_out = framework::vectorize(dim_y);
      dim_out[dim_out.size() - 2] = mat_dim_x.height_;
395
      dim_out[dim_out.size() - 1] = dim_out_y;
396
    } else {
397
      dim_out = {mat_dim_x.height_, dim_out_y};
M
Markus Kliegl 已提交
398 399
    }

400 401 402
    if (dim_x.size() == 1 && dim_out[dim_out.size() - 2] == 1) {
      std::swap(dim_out[dim_out.size() - 2], dim_out[dim_out.size() - 1]);
      dim_out.resize(dim_out.size() - 1);
M
Markus Kliegl 已提交
403 404
    }

405 406
    if (dim_y.size() == 1 && dim_out[dim_out.size() - 1] == 1) {
      dim_out.resize(dim_out.size() - 1);
M
Markus Kliegl 已提交
407 408
    }

409 410
    if (dim_out.empty()) {
      dim_out = {1};
M
Markus Kliegl 已提交
411 412 413 414
    }
    context->SetOutputDim("Out", framework::make_ddim(dim_out));
    context->ShareLoD("X", /*->*/ "Out");
  }
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext &ctx) const override {
    auto input_data_type = OperatorWithKernel::IndicateVarDataType(ctx, "X");

#ifdef PADDLE_WITH_MKLDNN
    using mkldnn::memory;
    if (platform::CanMKLDNNBeUsed(ctx)) {
      return framework::OpKernelType(input_data_type, ctx.GetPlace(),
                                     framework::DataLayout::kMKLDNN,
                                     framework::LibraryType::kMKLDNN);
    }
#endif
    return framework::OpKernelType(input_data_type, ctx.GetPlace());
  }
M
Markus Kliegl 已提交
430 431 432 433
};

class MatMulOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
434
  void Make() override {
M
Markus Kliegl 已提交
435 436 437 438 439 440 441 442 443 444 445
    AddInput("X", "The first input of MatMul op");
    AddInput("Y", "The second input of MatMul op");
    AddOutput("Out", "The output of MatMul op");
    AddAttr<bool>("transpose_X",
                  R"DOC(If true, use the transpose of `X`.
        )DOC")
        .SetDefault(false);
    AddAttr<bool>("transpose_Y",
                  R"DOC(If true, use the transpose of `Y`.
        )DOC")
        .SetDefault(false);
S
sneaxiy 已提交
446
    AddAttr<float>("alpha", "The scale of Out").SetDefault(1.0f);
447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470
    AddAttr<bool>(
        "use_mkldnn",
        "(bool, default false) Indicates if MKL-DNN kernel will be used")
        .SetDefault(false);
    /* int8 parameters */
    AddAttr<bool>("use_quantizer",
                  "(bool, default false) "
                  "Set to true for operators that should be quantized and use "
                  "int8 kernel. "
                  "Only used on CPU.")
        .SetDefault(false);
    AddAttr<float>("Scale_x",
                   "(float, default 1.0f), The quantize scale of X tensor")
        .SetDefault(1.0f);
    AddAttr<float>("Scale_y",
                   "(float, default 1.0f), The quantize scale of Y tensor")
        .SetDefault(1.0f);
    AddAttr<float>("Scale_out",
                   "(float, default 1.0f), The quantize scale of output data")
        .SetDefault(1.0f);
    AddAttr<bool>("force_fp32_output",
                  "(bool, default false) Force INT8 kernel output FP32, only "
                  "used in MKL-DNN INT8")
        .SetDefault(false);
471 472 473 474
#if defined(PADDLE_WITH_MKLML) && !defined(PADDLE_WITH_CUDA)
    AddAttr<int>("head_number", "The number of heads of the matrix")
        .SetDefault(1);
#endif
M
Markus Kliegl 已提交
475
    AddComment(R"DOC(
476 477 478 479
MatMul Operator.


This operator is used to perform (batched) matrix multiplication
M
Markus Kliegl 已提交
480 481 482 483 484 485 486 487 488 489 490 491 492 493
over the last two dimensions of the input tensors `X` and `Y`.

If a transpose flag is specified, the last two dimensions of the
tensor are transposed. If the tensor is rank-1 of shape [D], then
for `X` it is treated as [1, D] in nontransposed form and as [D, 1]
in transposed form, whereas for `Y` it is the opposite: It is treated
as [D, 1] in nontransposed form and as [1, D] in transposed form.

Examples without transpose:
- X: [K], Y: [K] => Out: [1]
- X: [K], Y: [K, N] => Out: [N]
- X: [B, M, K], Y: [K] => Out: [B, M]
- X: [M, K], Y: [B, K, N] => Out: [B, M, N]
- X: [B, M, K], Y: [B, K, N] => Out: [B, M, N]
C
chengduoZH 已提交
494
- X: [B, ..., M, K], Y: [B, ..., K, N] => Out: [B, ..., M, N]
M
Markus Kliegl 已提交
495

496 497 498
Example of matrix multiplication with head_number of H
- X: [B, M, K], Y: [B, K, N] => Out: [B, M, H * N]

M
Markus Kliegl 已提交
499 500
The behavior is designed to be similar to the `numpy.matmul` function.
The differences are:
C
chengduoZH 已提交
501 502
- When the rank of the input data is less than or equal to 3, it
  is similar to the `numpy.matmul` function.
C
chengduoZH 已提交
503
- When the rank of the input is greater than 3, the rank of X and
C
chengduoZH 已提交
504
  Y must be equal, and the first `rank - 2` dimensions must be equal.
M
Markus Kliegl 已提交
505
- We add `transpose_X` and `transpose_Y` flags.
506 507 508
- We add `head_number` attribute, which is used to multiple two matrixes head
  by head, and eventually concatenates the output of several (head_number)
  small matrixes multiplication.
M
Markus Kliegl 已提交
509 510

Both the input `X` and `Y` can carry the LoD (Level of Details) information,
511 512
or not. But the output only shares the LoD information with input `X`.

M
Markus Kliegl 已提交
513 514 515 516 517 518 519 520 521
)DOC");
  }
};

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

 protected:
522
  void InferShape(framework::InferShapeContext *context) const override {
M
Markus Kliegl 已提交
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
    PADDLE_ENFORCE(context->HasInput("X"), "Input(X) should not be null");
    PADDLE_ENFORCE(context->HasInput("Y"), "Input(Y) should not be null");
    PADDLE_ENFORCE(context->HasInput(framework::GradVarName("Out")),
                   "Input(Out@GRAD) should not be null");
    auto x_dims = context->GetInputDim("X");
    auto y_dims = context->GetInputDim("Y");

    auto x_grad_name = framework::GradVarName("X");
    auto y_grad_name = framework::GradVarName("Y");

    if (context->HasOutput(x_grad_name)) {
      context->SetOutputDim(x_grad_name, x_dims);
    }
    if (context->HasOutput(y_grad_name)) {
      context->SetOutputDim(y_grad_name, y_dims);
    }
  }
};

H
hong 已提交
542 543
template <typename T>
class MatMulOpGradMaker : public framework::SingleGradOpMaker<T> {
544
 public:
H
hong 已提交
545
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
546 547

 protected:
548
  void Apply(GradOpPtr<T> retv) const override {
549
    retv->SetType("matmul_grad");
H
hong 已提交
550 551 552 553 554 555
    retv->SetInput("X", this->Input("X"));
    retv->SetInput("Y", this->Input("Y"));
    retv->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    retv->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    retv->SetOutput(framework::GradVarName("Y"), this->InputGrad("Y"));
    retv->SetAttrMap(this->Attrs());
556 557
  }
};
M
Markus Kliegl 已提交
558 559 560 561
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Y
Yang Yang 已提交
562
REGISTER_OPERATOR(matmul, ops::MatMulOp, ops::MatMulOpMaker,
H
hong 已提交
563 564
                  ops::MatMulOpGradMaker<paddle::framework::OpDesc>,
                  ops::MatMulOpGradMaker<paddle::imperative::OpBase>);
565
REGISTER_OPERATOR(matmul_grad, ops::MatMulOpGrad);
M
Markus Kliegl 已提交
566
REGISTER_OP_CPU_KERNEL(
567
    matmul, ops::MatMulKernel<paddle::platform::CPUDeviceContext, float>,
568
    ops::MatMulKernel<paddle::platform::CPUDeviceContext, double>);
Q
QI JUN 已提交
569 570
REGISTER_OP_CPU_KERNEL(
    matmul_grad,
571
    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, float>,
572
    ops::MatMulGradKernel<paddle::platform::CPUDeviceContext, double>);
573 574 575

#ifdef PADDLE_WITH_CUDA
REGISTER_OP_CUDA_KERNEL(
576 577 578 579
    matmul, ops::MatMulKernel<paddle::platform::CUDADeviceContext, float>,
    ops::MatMulKernel<paddle::platform::CUDADeviceContext, double>,
    ops::MatMulKernel<paddle::platform::CUDADeviceContext,
                      paddle::platform::float16>);
580 581
REGISTER_OP_CUDA_KERNEL(
    matmul_grad,
582 583 584 585
    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, float>,
    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext, double>,
    ops::MatMulGradKernel<paddle::platform::CUDADeviceContext,
                          paddle::platform::float16>);
586
#endif
新手
引导
客服 返回
顶部