elementwise_op_function.h 59.7 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
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. */
14 15

#pragma once
16

17
#include <glog/logging.h>
18
#include <algorithm>
D
dzhwinter 已提交
19
#include <iterator>
20
#include <vector>
Y
Yi Wang 已提交
21 22 23 24
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/operator.h"
#include "paddle/fluid/platform/transform.h"
25

C
chengduoZH 已提交
26
#ifdef __NVCC__
27
#include <cuda.h>
C
chengduoZH 已提交
28
#include <thrust/iterator/iterator_adaptor.h>
29
#include "paddle/fluid/platform/cuda_device_function.h"
D
dzhwinter 已提交
30
#include "paddle/fluid/platform/cuda_primitives.h"
Y
Yu Yang 已提交
31
constexpr int ELEMWISE_MAX_BLOCK_DIM = 1024;
C
chengduoZH 已提交
32 33
#endif

Y
Yi Wang 已提交
34
#include "paddle/fluid/operators/math/math_function.h"
Y
Yu Yang 已提交
35
#include "paddle/fluid/platform/for_range.h"
36 37 38 39 40 41 42 43 44 45

namespace paddle {
namespace operators {

/*
 * Out = X ⊙ Y
 * If Y's shape does not match X' shape, they will be reshaped.
 * For example:
 * 1. shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
 *    pre=2, n=3*4, post=5
C
chengduo 已提交
46
 *    x.shape(2, 12, 5) * y.shape(1, 12, 1).broadcast(2, 12, 5)
47 48
 * 2. shape(X) = (2, 3, 4, 5), shape(Y) = (4,5)
 *    pre=2*3, n=4*5, post=1
C
chengduo 已提交
49
 *    x.shape(6, 20, 1) * y.shape(1, 20, 1).broadcast(6, 20, 1)
50
 */
51 52 53
inline void get_mid_dims(const framework::DDim &x_dims,
                         const framework::DDim &y_dims, const int axis,
                         int *pre, int *n, int *post) {
54 55 56
  *pre = 1;
  *n = 1;
  *post = 1;
57
  for (int i = 0; i < axis; ++i) {
58
    (*pre) *= x_dims[i];
59 60 61 62 63
  }

  for (int i = 0; i < y_dims.size(); ++i) {
    PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
                      "Broadcast dimension mismatch.");
64
    (*n) *= y_dims[i];
65 66 67
  }

  for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
68
    (*post) *= x_dims[i];
69 70 71
  }
}

72
inline framework::DDim trim_trailing_singular_dims(
73
    const framework::DDim &dims) {
74
  // Remove trailing dimensions of size 1 for y
75
  auto actual_dims_size = dims.size();
76
  for (; actual_dims_size != 0; --actual_dims_size) {
77
    if (dims[actual_dims_size - 1] != 1) break;
78
  }
79 80 81 82 83

  std::vector<int> trim_dims;
  trim_dims.resize(actual_dims_size);
  for (int i = 0; i < actual_dims_size; ++i) {
    trim_dims[i] = dims[i];
84
  }
85 86 87
  if (trim_dims.size() == 0) {
    return framework::DDim(framework::make_dim());
  }
88 89
  framework::DDim actual_dims = framework::make_ddim(trim_dims);
  return actual_dims;
90 91
}

Q
QI JUN 已提交
92
template <typename T, typename DeviceContext>
C
chengduoZH 已提交
93
class RowwiseTransformIterator;
94

Q
QI JUN 已提交
95
template <typename T, typename DeviceContext>
C
chengduoZH 已提交
96
class MidWiseTransformIterator;
C
chengduoZH 已提交
97

D
dzhwinter 已提交
98
// NOTE(dzhwinter): ptrdiff_t in iterator is deperecated in c++17
C
chengduoZH 已提交
99
template <typename T>
D
dzhwinter 已提交
100 101 102
class RowwiseTransformIterator<T, platform::CPUDeviceContext>
    : public std::iterator<std::random_access_iterator_tag, T, std::ptrdiff_t,
                           T *, T &> {
C
chengduoZH 已提交
103
 public:
104
  RowwiseTransformIterator(const T *ptr, int n) : ptr_(ptr), i_(0), n_(n) {}
C
chengduoZH 已提交
105

106
  RowwiseTransformIterator<T, platform::CPUDeviceContext> &operator++() {
C
chengduoZH 已提交
107
    ++i_;
C
chengduoZH 已提交
108 109 110
    if (UNLIKELY(i_ == n_)) {
      i_ = 0;
    }
C
chengduoZH 已提交
111 112 113
    return *this;
  }

114 115
  bool operator==(const RowwiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
116
    return (ptr_ + i_) == &(*rhs);
C
chengduoZH 已提交
117 118
  }

119 120
  bool operator!=(const RowwiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
121
    return (ptr_ + i_) != &(*rhs);
C
chengduoZH 已提交
122 123
  }

124
  const T &operator*() { return ptr_[i_]; }
C
chengduoZH 已提交
125

C
chengduoZH 已提交
126
 private:
127
  const T *ptr_;
C
chengduoZH 已提交
128
  int i_;
C
chengduoZH 已提交
129
  int64_t n_;
C
chengduoZH 已提交
130 131 132
};

template <typename T>
D
dzhwinter 已提交
133 134 135
class MidWiseTransformIterator<T, platform::CPUDeviceContext>
    : public std::iterator<std::random_access_iterator_tag, T, std::ptrdiff_t,
                           T *, T &> {
C
chengduoZH 已提交
136
 public:
137
  MidWiseTransformIterator(const T *ptr, int n, int post)
C
chengduoZH 已提交
138 139
      : ptr_(ptr), i_(0), j_(0), n_(n), post_(post) {}

140
  MidWiseTransformIterator<T, platform::CPUDeviceContext> &operator++() {
C
chengduoZH 已提交
141
    ++j_;
C
chengduoZH 已提交
142 143
    if (UNLIKELY(j_ == post_)) {
      ++i_;
C
refine  
chengduoZH 已提交
144
      j_ = 0;
C
chengduoZH 已提交
145 146 147
      if (UNLIKELY(i_ == n_)) {
        i_ = 0;
      }
C
chengduoZH 已提交
148
    }
C
chengduoZH 已提交
149 150 151
    return *this;
  }

152 153
  bool operator==(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
154
    return (ptr_ + i_) == &(*rhs);
C
chengduoZH 已提交
155 156
  }

157 158
  bool operator!=(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
159
    return (ptr_ + i_) != &(*rhs);
C
chengduoZH 已提交
160 161
  }

162
  const T &operator*() { return ptr_[i_]; }
C
chengduoZH 已提交
163

C
chengduoZH 已提交
164
 private:
165
  const T *ptr_;
C
refine  
chengduoZH 已提交
166
  int64_t i_;
C
chengduoZH 已提交
167 168
  int64_t j_;
  int64_t n_;
C
refine  
chengduoZH 已提交
169
  int64_t post_;
C
chengduoZH 已提交
170 171
};

C
chengduoZH 已提交
172 173
#ifdef __NVCC__
template <typename T>
Q
QI JUN 已提交
174
class RowwiseTransformIterator<T, platform::CUDADeviceContext>
C
chengduoZH 已提交
175
    : public thrust::iterator_adaptor<
176
          RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T *> {
C
chengduoZH 已提交
177 178
 public:
  typedef thrust::iterator_adaptor<
179
      RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T *>
C
chengduoZH 已提交
180
      super_t;
181
  HOSTDEVICE RowwiseTransformIterator(const T *x, int n)
182
      : super_t(x), begin_(x), n_(n) {}
C
chengduoZH 已提交
183 184 185 186
  friend class thrust::iterator_core_access;

 private:
  unsigned int n_;
187
  const T *begin_;
C
chengduoZH 已提交
188
  HOSTDEVICE typename super_t::reference dereference() const {
C
chengduoZH 已提交
189 190 191 192 193
    return *(begin_ + (this->base() - begin_) % n_);
  }
};

template <typename T>
Q
QI JUN 已提交
194
class MidWiseTransformIterator<T, platform::CUDADeviceContext>
C
chengduoZH 已提交
195
    : public thrust::iterator_adaptor<
196
          MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T *> {
C
chengduoZH 已提交
197 198
 public:
  typedef thrust::iterator_adaptor<
199
      MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T *>
C
chengduoZH 已提交
200
      super_t;
201
  HOSTDEVICE MidWiseTransformIterator(const T *x, int n, int post)
202
      : super_t(x), begin_(x), n_(n), post_(post) {}
C
chengduoZH 已提交
203 204 205 206 207
  friend class thrust::iterator_core_access;

 private:
  unsigned int post_;
  unsigned int n_;
208
  const T *begin_;
C
chengduoZH 已提交
209
  HOSTDEVICE typename super_t::reference dereference() const {
C
chengduoZH 已提交
210 211 212 213 214
    return *(begin_ + (((this->base() - begin_) / post_) % n_));
  }
};
#endif

215 216
template <typename Functor, typename T, typename DeviceContext,
          typename OutType = T>
C
chengduoZH 已提交
217 218
class TransformFunctor {
 public:
219 220
  TransformFunctor(const framework::Tensor *x, const framework::Tensor *y,
                   framework::Tensor *z, const DeviceContext &ctx, Functor func)
C
chengduoZH 已提交
221 222
      : x_(x->data<T>()),
        y_(y->data<T>()),
223
        z_(z->mutable_data<OutType>(ctx.GetPlace())),
C
chengduoZH 已提交
224 225 226 227 228
        nx_(x->numel()),
        ctx_(ctx),
        func_(func) {}

  inline void Run() const {
Q
QI JUN 已提交
229
    platform::Transform<DeviceContext> trans;
C
chengduoZH 已提交
230
    trans(ctx_, x_, x_ + nx_, y_, z_, func_);
C
chengduoZH 已提交
231 232 233
  }

  inline void RunRowWise(int n, int pre) const {
Q
QI JUN 已提交
234 235 236
    platform::Transform<DeviceContext> trans;
    trans(ctx_, x_, x_ + nx_, RowwiseTransformIterator<T, DeviceContext>(y_, n),
          z_, func_);
C
chengduoZH 已提交
237 238 239
  }

  inline void RunMidWise(int n, int pre, int post) const {
Q
QI JUN 已提交
240 241 242
    platform::Transform<DeviceContext> trans;
    trans(ctx_, x_, x_ + nx_,
          MidWiseTransformIterator<T, DeviceContext>(y_, n, post), z_, func_);
C
chengduoZH 已提交
243 244
  }

C
chengduoZH 已提交
245
 private:
246 247 248
  const T *x_;
  const T *y_;
  OutType *z_;
C
chengduoZH 已提交
249
  int64_t nx_;
250
  const DeviceContext &ctx_;
C
chengduoZH 已提交
251 252 253
  Functor func_;
};

254 255
#define EIGEN_FUNCTOR(name, eigen_op)                                          \
  struct Eigen##name##Functor {                                                \
Q
QI JUN 已提交
256
    template <typename DeviceContext, typename T>                              \
257 258 259
    inline void Run(const framework::Tensor *x, const framework::Tensor *y,    \
                    framework::Tensor *z,                                      \
                    const framework::ExecutionContext &ctx) {                  \
260 261 262
      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
Q
QI JUN 已提交
263 264 265
      z_e.device(                                                              \
          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
          eigen_op(x_e, y_e);                                                  \
266
    }                                                                          \
Q
QI JUN 已提交
267
    template <typename DeviceContext, typename T>                              \
268 269 270
    inline void RunBroadCast(const framework::Tensor *x,                       \
                             const framework::Tensor *y, framework::Tensor *z, \
                             const framework::ExecutionContext &ctx, int pre,  \
271 272 273 274 275 276 277
                             int n) {                                          \
      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))                  \
                         .broadcast(Eigen::DSizes<int, 2>(pre, 1))             \
                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
Q
QI JUN 已提交
278 279 280
      z_e.device(                                                              \
          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
          eigen_op(x_e, y_bcast);                                              \
281
    }                                                                          \
Q
QI JUN 已提交
282
    template <typename DeviceContext, typename T>                              \
283 284 285 286
    inline void RunBroadCast2(const framework::Tensor *x,                      \
                              const framework::Tensor *y,                      \
                              framework::Tensor *z,                            \
                              const framework::ExecutionContext &ctx, int pre, \
287 288 289 290 291 292 293
                              int n, int post) {                               \
      auto x_e = framework::EigenVector<T>::Flatten(*x);                       \
      auto y_e = framework::EigenVector<T>::Flatten(*y);                       \
      auto z_e = framework::EigenVector<T>::Flatten(*z);                       \
      auto y_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))               \
                         .broadcast(Eigen::DSizes<int, 3>(pre, 1, post))       \
                         .reshape(Eigen::DSizes<int, 1>(x_e.size()));          \
Q
QI JUN 已提交
294 295 296
      z_e.device(                                                              \
          *ctx.template device_context<DeviceContext>().eigen_device()) =      \
          eigen_op(x_e, y_bcast);                                              \
297 298 299 300
    }                                                                          \
  }

#define EIGEN_ADD(x, y) ((x) + (y))
301

302 303 304
EIGEN_FUNCTOR(Add, EIGEN_ADD);

#define EIGEN_SUB(x, y) ((x) - (y))
305

306 307 308
EIGEN_FUNCTOR(Sub, EIGEN_SUB);

#define EIGEN_MUL(x, y) ((x) * (y))
309

310 311 312
EIGEN_FUNCTOR(Mul, EIGEN_MUL);

#define EIGEN_DIV(x, y) ((x) / (y))
313

314 315
EIGEN_FUNCTOR(Div, EIGEN_DIV);

Y
Yu Yang 已提交
316 317
template <typename T, typename DX_OP, typename DY_OP>
struct ElemwiseGradNoBroadcast {
318 319 320 321
  const T *x_;
  const T *y_;
  const T *out_;
  const T *dout_;
Y
Yu Yang 已提交
322 323 324 325 326 327

  HOSTDEVICE void operator()(size_t i) {
    if (dx_ != nullptr) {
      dx_[i] = dx_op_(x_[i], y_[i], out_[i], dout_[i]);
    }
    if (dy_ != nullptr) {
C
chengduoZH 已提交
328
      dy_[i] = dy_op_(x_[i], y_[i], out_[i], dout_[i]);
Y
Yu Yang 已提交
329 330 331 332 333
    }
  }

  DX_OP dx_op_;
  DY_OP dy_op_;
334 335
  T *dx_;
  T *dy_;
Y
Yu Yang 已提交
336 337 338
};

template <typename T, typename DX_OP, typename DY_OP>
339 340 341
static void ElemwiseGradBroadcast1CPU(const T *x, const T *y, const T *out,
                                      const T *dout, int h, int w, DX_OP dx_op,
                                      DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      int x_offset = i * w + j;
      if (dx != nullptr) {
        dx[x_offset] = dx_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
      }
      if (dy != nullptr) {
        T tmp = dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
        if (i == 0) {
          dy[j] = tmp;
        } else {
          dy[j] += tmp;
        }
      }
    }
  }
}
359

D
dzhwinter 已提交
360
#ifdef __NVCC__
Y
Yu Yang 已提交
361 362
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void ElemwiseGradBroadcast1CUDAKernel(
363 364
    const T *x, const T *y, const T *out, const T *dout, int h, int w,
    DX_OP dx_op, DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
365 366 367
  int j = blockIdx.x;
  int i = threadIdx.x;
  int tid = threadIdx.x;
368
  T val = 0;
Y
Yu Yang 已提交
369 370 371 372 373 374 375

  do {
    int x_offset = i * w + j;
    if (dx) {
      dx[x_offset] = dx_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
    }
    if (dy) {
C
chengduoZH 已提交
376
      val += dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
Y
Yu Yang 已提交
377 378 379 380 381
    }
    i += ELEMWISE_MAX_BLOCK_DIM;
  } while (i < h);

  if (dy) {
C
chengduoZH 已提交
382
    h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
383
    val = paddle::platform::reduceSum(val, tid, h);
Y
Yu Yang 已提交
384
    if (threadIdx.x == 0) {
C
chengduoZH 已提交
385
      dy[j] = val;
Y
Yu Yang 已提交
386 387 388 389 390
    }
  }
}

template <typename T, typename DX_OP, typename DY_OP>
391 392
static void ElemwiseGradBroadcast1CUDA(cudaStream_t stream, const T *x,
                                       const T *y, const T *out, const T *dout,
Y
Yu Yang 已提交
393
                                       int h, int w, DX_OP dx_op, DY_OP dy_op,
394
                                       T *dx, T *dy) {
Y
Yu Yang 已提交
395 396
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, h);
  int gird_size = w;
C
chengduoZH 已提交
397 398
  ElemwiseGradBroadcast1CUDAKernel<<<gird_size, block_size, 0, stream>>>(
      x, y, out, dout, h, w, dx_op, dy_op, dx, dy);
Y
Yu Yang 已提交
399 400 401 402 403
}

#endif

template <typename T, typename DX_OP, typename DY_OP>
404 405 406
static void ElemwiseGradBroadcast2CPU(const T *x, const T *y, const T *out,
                                      const T *dout, int pre, int n, int post,
                                      DX_OP dx_op, DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430
  for (int i = 0; i < pre; ++i) {
    for (int j = 0; j < n; ++j) {
      for (int k = 0; k < post; ++k) {
        int x_offset = i * n * post + j * post + k;
        if (dx != nullptr) {
          dx[x_offset] =
              dx_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
        }
        if (dy != nullptr) {
          T tmp = dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
          if (i == 0 && k == 0) {
            dy[j] = tmp;
          } else {
            dy[j] += tmp;
          }
        }
      }
    }
  }
}

#ifdef __NVCC__
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void ElemwiseGradBroadcast2CUDAKernel(
431 432
    const T *x, const T *y, const T *out, const T *dout, int pre, int n,
    int post, DX_OP dx_op, DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
433 434 435
  int tid = threadIdx.x;
  int j = blockIdx.x;

436
  T val = 0;
Y
Yu Yang 已提交
437 438 439 440 441 442 443 444 445 446 447 448 449 450
  int ttid = tid;

  while (true) {
    int i = ttid / post;
    int k = ttid % post;
    if (i >= pre) break;

    int x_offset = i * n * post + j * post + k;

    if (dx != nullptr) {
      dx[x_offset] = dx_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
    }

    if (dy != nullptr) {
C
chengduoZH 已提交
451
      val += dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
Y
Yu Yang 已提交
452 453 454 455 456 457
    }

    ttid += ELEMWISE_MAX_BLOCK_DIM;
  }

  if (dy) {
C
chengduoZH 已提交
458 459
    int h = pre * post;
    h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
460
    val = paddle::platform::reduceSum(val, tid, h);
C
chengduoZH 已提交
461
    if (threadIdx.x == 0) {
C
chengduoZH 已提交
462
      dy[j] = val;
Y
Yu Yang 已提交
463 464 465 466 467
    }
  }
}

template <typename T, typename DX_OP, typename DY_OP>
468 469
static void ElemwiseGradBroadcast2CUDA(cudaStream_t stream, const T *x,
                                       const T *y, const T *out, const T *dout,
Y
Yu Yang 已提交
470
                                       int pre, int n, int post, DX_OP dx_op,
471
                                       DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
472 473
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, pre * post);
  int gird_size = n;
C
chengduoZH 已提交
474 475
  ElemwiseGradBroadcast2CUDAKernel<<<gird_size, block_size, 0, stream>>>(
      x, y, out, dout, pre, n, post, dx_op, dy_op, dx, dy);
Y
Yu Yang 已提交
476 477 478 479
}

#endif

480 481
template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
void ElemwiseGradComputeNoBroadcast(
482 483 484 485 486
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::DDim &y_dim, const framework::Tensor &x,
    const framework::Tensor &y, const framework::Tensor &out,
    const framework::Tensor &dout, int axis, framework::Tensor *dx,
    framework::Tensor *dy, DX_OP dx_op, DY_OP dy_op) {
487
  size_t N = static_cast<size_t>(framework::product(x_dim));
D
dzhwinter 已提交
488
#if !defined(_WIN32)
489 490
  platform::ForRange<DeviceContext> for_range(
      ctx.template device_context<DeviceContext>(), N);
D
dzhwinter 已提交
491 492 493 494
#else
  platform::ForRange<DeviceContext> for_range(
      ctx.device_context<DeviceContext>(), N);
#endif  // !_WIN32
495 496 497 498 499 500 501 502
  for_range(ElemwiseGradNoBroadcast<T, DX_OP, DY_OP>{
      x.data<T>(), y.data<T>(), out.data<T>(), dout.data<T>(), dx_op, dy_op,
      dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
      dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace())});
}

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
void ElemwiseGradComputeWithBroadcast(
503 504 505 506 507
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::DDim &y_dim_untrimed, const framework::Tensor &x,
    const framework::Tensor &y, const framework::Tensor &out,
    const framework::Tensor &dout, int axis, framework::Tensor *dx,
    framework::Tensor *dy, DX_OP dx_op, DY_OP dy_op) {
508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549
  axis = (axis == -1 ? x_dim.size() - y_dim_untrimed.size() : axis);
  auto y_dim = trim_trailing_singular_dims(y_dim_untrimed);
  axis = (y_dim.size() == 0) ? x_dim.size() : axis;

  int pre, n, post;
  get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
  if (post == 1) {
    int h = pre;
    int w = n;
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      ElemwiseGradBroadcast1CUDA(
          ctx.template device_context<DeviceContext>().stream(), x.data<T>(),
          y.data<T>(), out.data<T>(), dout.data<T>(), h, w, dx_op, dy_op,
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()));
#endif
    } else {
      ElemwiseGradBroadcast1CPU(
          x.data<T>(), y.data<T>(), out.data<T>(), dout.data<T>(), h, w, dx_op,
          dy_op, dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()));
    }
  } else {
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      ElemwiseGradBroadcast2CUDA(
          ctx.template device_context<DeviceContext>().stream(), x.data<T>(),
          y.data<T>(), out.data<T>(), dout.data<T>(), pre, n, post, dx_op,
          dy_op, dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()));
#endif
    } else {
      ElemwiseGradBroadcast2CPU(
          x.data<T>(), y.data<T>(), out.data<T>(), dout.data<T>(), pre, n, post,
          dx_op, dy_op,
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()));
    }
  }
}

Y
Yu Yang 已提交
550
template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
551 552 553 554 555
void ElemwiseGradCompute(const framework::ExecutionContext &ctx,
                         const framework::Tensor &x, const framework::Tensor &y,
                         const framework::Tensor &out,
                         const framework::Tensor &dout, int axis,
                         framework::Tensor *dx, framework::Tensor *dy,
Y
Yu Yang 已提交
556
                         DX_OP dx_op, DY_OP dy_op) {
557 558
  const framework::DDim &x_dim = x.dims();
  const framework::DDim &y_dim = y.dims();
Y
Yu Yang 已提交
559
  if (x.dims() == y.dims()) {
560 561
    ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
        ctx, x_dim, y_dim, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
Y
Yu Yang 已提交
562
  } else {  // Y is a scalar
563 564 565 566 567 568 569 570 571 572
    ElemwiseGradComputeWithBroadcast<DeviceContext, T, DX_OP, DY_OP>(
        ctx, x_dim, y_dim, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
  }
}

// NOTE(dzhwinter): Only used in elementwise_add, elementwise_sub.
// explicit gradient can cut off X, Y, Out from gradient op
// In elementwise_add, elementwise_sub, we use dout as fake X, Y, Out to reuse
// elementwise code.
template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
573 574 575 576 577 578
void ElemwiseExplicitGradCompute(const framework::ExecutionContext &ctx,
                                 const framework::Tensor &x,
                                 const framework::Tensor &y,
                                 const framework::Tensor &out,
                                 const framework::Tensor &dout, int axis,
                                 framework::Tensor *dx, framework::Tensor *dy,
579 580
                                 DX_OP dx_op, DY_OP dy_op) {
  if (dy == nullptr) {
581
    const framework::DDim &dx_dims = dout.dims();
582 583 584 585 586
    auto dy_dims = dx_dims;
    ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
        ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
  } else {
    if (dout.dims() == dy->dims()) {
587 588
      const framework::DDim &dx_dims = dout.dims();
      const framework::DDim &dy_dims = dy->dims();
589 590 591 592
      ElemwiseGradComputeNoBroadcast<DeviceContext, T, DX_OP, DY_OP>(
          ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
    } else {  // Y is a scalar
      auto dx_dims = dout.dims();
593
      const framework::DDim &dy_dims = dy->dims();
594 595
      ElemwiseGradComputeWithBroadcast<DeviceContext, T, DX_OP, DY_OP>(
          ctx, dx_dims, dy_dims, x, y, out, dout, axis, dx, dy, dx_op, dy_op);
Y
Yu Yang 已提交
596 597
    }
  }
598
}
Y
Yu Yang 已提交
599

600
// Deprecated
Q
QI JUN 已提交
601
template <typename DeviceContext, typename T, typename functor,
F
fengjiayi 已提交
602
          typename broadcastfunctor, typename broadcast2functor>
603 604 605 606 607 608 609
void ElementwiseGradCompute(const framework::ExecutionContext &ctx,
                            const framework::Tensor *x,
                            const framework::Tensor *y,
                            const framework::Tensor *out,
                            const framework::Tensor *dout, int axis,
                            framework::Tensor *dx, framework::Tensor *dy) {
  auto &place = *ctx.template device_context<DeviceContext>().eigen_device();
610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627

  auto x_dims = x->dims();
  auto y_dims = y->dims();

  if (dx) {
    dx->mutable_data<T>(ctx.GetPlace());
  }
  if (dy) {
    dy->mutable_data<T>(ctx.GetPlace());
  }

  if (x_dims == y_dims) {
    functor f;
    f(place, x, y, out, dx, dy, dout);
    return;
  }

  axis = (axis == -1 ? x_dims.size() - y_dims.size() : axis);
628
  trim_trailing_singular_dims(y_dims);
629
  axis = (y_dims.size() == 0) ? x_dims.size() : axis;
630 631

  int pre, n, post;
632
  get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post);
633 634 635 636 637 638 639 640 641 642 643

  if (post == 1) {
    broadcastfunctor f;
    f(place, x, y, out, dx, dy, dout, pre, n);
    return;
  } else {
    broadcast2functor f;
    f(place, x, y, out, dx, dy, dout, pre, n, post);
    return;
  }
}
F
fengjiayi 已提交
644

645 646
template <typename Functor, typename DeviceContext, typename T,
          typename OutType = T>
D
dzhwinter 已提交
647

648 649 650 651
void ElementwiseComputeEx(const framework::ExecutionContext &ctx,
                          const framework::Tensor *x,
                          const framework::Tensor *y, int axis, Functor func,
                          framework::Tensor *z) {
652
  TransformFunctor<Functor, T, DeviceContext, OutType> functor(
C
chengduoZH 已提交
653
      x, y, z, ctx.template device_context<DeviceContext>(), func);
F
fengjiayi 已提交
654
  auto x_dims = x->dims();
655 656
  auto y_dims_untrimed = y->dims();
  PADDLE_ENFORCE_GE(x_dims.size(), y_dims_untrimed.size(),
F
fengjiayi 已提交
657 658
                    "Rank of first input must >= rank of second input.");

659
  if (x_dims == y_dims_untrimed) {
F
fengjiayi 已提交
660 661 662 663
    functor.Run();
    return;
  }

664
  axis = (axis == -1 ? x_dims.size() - y_dims_untrimed.size() : axis);
F
fengjiayi 已提交
665 666
  PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
                 "Axis should be in range [0, x_dims)");
667
  auto y_dims = trim_trailing_singular_dims(y_dims_untrimed);
668
  axis = (y_dims.size() == 0) ? x_dims.size() : axis;
F
fengjiayi 已提交
669 670

  int pre, n, post;
671
  get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post);
F
fengjiayi 已提交
672 673 674 675 676 677 678 679 680
  if (post == 1) {
    functor.RunRowWise(n, pre);
    return;
  } else {
    functor.RunMidWise(n, pre, post);
    return;
  }
}

681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989
// FusedElemwiseAndAct
// --- forward
template <typename T, typename CompoundFunctor, bool KeepIntermediateOut>
struct FusedElemwiseAndActNoBroadcast {
  HOSTDEVICE void operator()(size_t i) {
    T y_val = y_[i];
    T x_val = x_[i];
    if (KeepIntermediateOut) {
      T intermeidiate_out = compound_functor_.GetIntermediateOut(x_val, y_val);
      intermediate_out_[i] = intermeidiate_out;
      out_[i] =
          compound_functor_.GetOutUseIntermediateOut(x_val, intermeidiate_out);
    } else {
      out_[i] = compound_functor_.GetOut(x_val, y_val);
    }
  }

  const T *x_;
  const T *y_;
  CompoundFunctor compound_functor_;
  T *out_;
  T *intermediate_out_;
};

// FusedElemwiseAndActBroadcast1:
// In this case, X and Y can be reshaped to a matrix.
// For example shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5) and axis = -1 or 2,
// X can be reshaped to (6, 20) and Y can be reshaped to (1, 20)
template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActBroadcast1CPU(const T *x, const T *y,
                                             CompoundFunctor compound_functor,
                                             int h, int w, T *out,
                                             T *intermediate_out) {
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      int offset = i * w + j;

      T y_val = BcastY ? y[j] : y[offset];
      T x_val = BcastY ? x[offset] : x[j];
      int64_t intermediate_out_offset;
      if (KeepIntermediateOut) {
        T intermeidiate_out = compound_functor.GetIntermediateOut(x_val, y_val);

        if (SameShapeOfIntermediateOutAndOut) {
          // for the case of f1(f2(x, y))
          intermediate_out_offset = offset;
        } else if (BcastY) {
          intermediate_out_offset = j;
        } else {
          intermediate_out_offset = offset;
        }

        intermediate_out[intermediate_out_offset] = intermeidiate_out;
        out[offset] =
            compound_functor.GetOutUseIntermediateOut(x_val, intermeidiate_out);
      } else {
        out[offset] = compound_functor.GetOut(x_val, y_val);
      }
    }
  }
}

// FusedElemwiseAndActBroadcast2
// In this case, X and Y can be reshaped to a matrix.
// For example shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4) and axis = 1,
// X can be reshaped to (2, 12, 5) and Y can be reshaped to (1, 12, 1)
// pre = 2, n = 12, post = 5
template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActBroadcast2CPU(const T *x, const T *y, int pre,
                                             int n, int post,
                                             CompoundFunctor compound_functor,
                                             T *out, T *intermediate_out) {
  for (int i = 0; i < pre; ++i) {
    for (int j = 0; j < n; ++j) {
      for (int k = 0; k < post; ++k) {
        int offset = i * n * post + j * post + k;

        T y_val = BcastY ? y[j] : y[offset];
        T x_val = BcastY ? x[offset] : x[j];
        int64_t intermediate_out_offset;

        if (KeepIntermediateOut) {
          T intermeidiate_out =
              compound_functor.GetIntermediateOut(x_val, y_val);

          if (SameShapeOfIntermediateOutAndOut) {
            // for the case of f1(f2(x, y))
            intermediate_out_offset = offset;
          } else if (BcastY) {
            intermediate_out_offset = j;
          } else {
            intermediate_out_offset = offset;
          }

          intermediate_out[intermediate_out_offset] = intermeidiate_out;
          out[offset] = compound_functor.GetOutUseIntermediateOut(
              x_val, intermeidiate_out);
        } else {
          out[offset] = compound_functor.GetOut(x_val, y_val);
        }
      }
    }
  }
}

#ifdef __NVCC__
template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static __global__ void FusedElemwiseAndActBroadcast1CUDAKernel(
    const T *x, const T *y, int h, int w, CompoundFunctor compound_functor,
    T *out, T *intermediate_out) {
  int j = blockIdx.x;
  int i = threadIdx.x;

  while (i < h) {
    int offset = i * w + j;

    T y_val = BcastY ? y[j] : y[offset];
    T x_val = BcastY ? x[offset] : x[j];
    int64_t intermediate_out_offset;

    if (KeepIntermediateOut) {
      T intermeidiate_out = compound_functor.GetIntermediateOut(x_val, y_val);

      if (SameShapeOfIntermediateOutAndOut) {
        // for the case of f1(f2(x, y))
        intermediate_out_offset = offset;
      } else if (BcastY) {
        intermediate_out_offset = j;
      } else {
        intermediate_out_offset = offset;
      }

      intermediate_out[intermediate_out_offset] = intermeidiate_out;
      out[offset] =
          compound_functor.GetOutUseIntermediateOut(x_val, intermeidiate_out);
    } else {
      out[offset] = compound_functor.GetOut(x_val, y_val);
    }

    i += ELEMWISE_MAX_BLOCK_DIM;
  }
}

template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActBroadcast1CUDA(cudaStream_t stream, const T *x,
                                              const T *y,
                                              CompoundFunctor compound_functor,
                                              int h, int w, T *out,
                                              T *intermediate_out) {
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, h);
  int gird_size = w;
  FusedElemwiseAndActBroadcast1CUDAKernel<
      T, CompoundFunctor, BcastY, KeepIntermediateOut,
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
      x, y, h, w, compound_functor, out, intermediate_out);
}

template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static __global__ void FusedElemwiseAndActBroadcast2CUDAKernel(
    const T *x, const T *y, CompoundFunctor compound_functor, int pre, int n,
    int post, T *out, T *intermediate_out) {
  int tid = threadIdx.x;
  int j = blockIdx.x;

  while (true) {
    int i = tid / post;
    int k = tid % post;
    if (i >= pre) break;

    int offset = i * n * post + j * post + k;

    T y_val = BcastY ? y[j] : y[offset];
    T x_val = BcastY ? x[offset] : x[j];
    int64_t intermediate_out_offset;

    if (KeepIntermediateOut) {
      T intermeidiate_out = compound_functor.GetIntermediateOut(x_val, y_val);

      if (SameShapeOfIntermediateOutAndOut) {
        // for the case of f1(f2(x, y))
        intermediate_out_offset = offset;
      } else if (BcastY) {
        intermediate_out_offset = j;
      } else {
        intermediate_out_offset = offset;
      }

      intermediate_out[intermediate_out_offset] = intermeidiate_out;
      out[offset] =
          compound_functor.GetOutUseIntermediateOut(x_val, intermeidiate_out);
    } else {
      out[offset] = compound_functor.GetOut(x_val, y_val);
    }

    tid += ELEMWISE_MAX_BLOCK_DIM;
  }
}

template <typename T, typename CompoundFunctor, bool BcastY,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActBroadcast2CUDA(cudaStream_t stream, const T *x,
                                              const T *y, int pre, int n,
                                              int post,
                                              CompoundFunctor compound_functor,
                                              T *out, T *intermediate_out) {
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, pre * post);
  int gird_size = n;

  FusedElemwiseAndActBroadcast2CUDAKernel<
      T, CompoundFunctor, BcastY, KeepIntermediateOut,
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
      x, y, compound_functor, pre, n, post, out, intermediate_out);
}

#endif

template <typename DeviceContext, typename T, typename CompoundFunctor,
          bool KeepIntermediateOut>
void FusedElemwiseAndActComputeNoBroadcast(
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::Tensor &x, const framework::Tensor &y,
    CompoundFunctor compound_functor, framework::Tensor *out,
    framework::Tensor *intermediate_out) {
  size_t N = static_cast<size_t>(framework::product(x_dim));

  platform::ForRange<DeviceContext> for_range(
      ctx.template device_context<DeviceContext>(), N);

  for_range(
      FusedElemwiseAndActNoBroadcast<T, CompoundFunctor, KeepIntermediateOut>{
          x.data<T>(), y.data<T>(), compound_functor,
          out->mutable_data<T>(ctx.GetPlace()),
          intermediate_out == nullptr
              ? nullptr
              : intermediate_out->mutable_data<T>(ctx.GetPlace())});
}

template <typename DeviceContext, typename T, typename CompoundFunctor,
          bool BcastY, bool KeepIntermediateOut,
          bool SameShapeOfIntermediateOutAndOut>
void FusedElemwiseAndActComputeWithBroadcast(
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::DDim &y_dim_untrimed, const framework::Tensor &x,
    const framework::Tensor &y, CompoundFunctor compound_functor, int axis,
    framework::Tensor *out, framework::Tensor *intermediate_out) {
  axis = (axis == -1 ? x_dim.size() - y_dim_untrimed.size() : axis);
  auto y_dim = trim_trailing_singular_dims(y_dim_untrimed);
  axis = (y_dim.size() == 0) ? x_dim.size() : axis;

  int pre, n, post;
  get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);

  if (post == 1) {
    int h = pre;
    int w = n;
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      FusedElemwiseAndActBroadcast1CUDA<T, CompoundFunctor, BcastY,
                                        KeepIntermediateOut,
                                        SameShapeOfIntermediateOutAndOut>(
          ctx.template device_context<DeviceContext>().stream(), x.data<T>(),
          y.data<T>(), compound_functor, h, w,
          out->mutable_data<T>(ctx.GetPlace()),
          intermediate_out == nullptr
              ? nullptr
              : intermediate_out->mutable_data<T>(ctx.GetPlace()));
#endif
    } else {
      FusedElemwiseAndActBroadcast1CPU<T, CompoundFunctor, BcastY,
                                       KeepIntermediateOut,
                                       SameShapeOfIntermediateOutAndOut>(
          x.data<T>(), y.data<T>(), compound_functor, h, w,
          out->mutable_data<T>(ctx.GetPlace()),
          intermediate_out == nullptr
              ? nullptr
              : intermediate_out->mutable_data<T>(ctx.GetPlace()));
    }
  } else {
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      FusedElemwiseAndActBroadcast2CUDA<T, CompoundFunctor, BcastY,
                                        KeepIntermediateOut,
                                        SameShapeOfIntermediateOutAndOut>(
          ctx.template device_context<DeviceContext>().stream(), x.data<T>(),
          y.data<T>(), pre, n, post, compound_functor,
          out->mutable_data<T>(ctx.GetPlace()),
          intermediate_out == nullptr
              ? nullptr
              : intermediate_out->mutable_data<T>(ctx.GetPlace()));
#endif
    } else {
      FusedElemwiseAndActBroadcast2CPU<T, CompoundFunctor, BcastY,
                                       KeepIntermediateOut,
                                       SameShapeOfIntermediateOutAndOut>(
          x.data<T>(), y.data<T>(), pre, n, post, compound_functor,
          out->mutable_data<T>(ctx.GetPlace()),
          intermediate_out == nullptr
              ? nullptr
              : intermediate_out->mutable_data<T>(ctx.GetPlace()));
    }
  }
}

// --- backward
C
chengduo 已提交
990 991
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut>
992 993 994
struct FusedElemwiseAndActGradNoBroadcast {
  HOSTDEVICE void operator()(size_t i) {
    if (dx_ != nullptr) {
C
chengduo 已提交
995 996 997 998
      dx_[i] = UseIntermediateOut
                   ? dx_op_.UseIntermediateOut(
                         x_[i], y_[i], intermediate_out_[i], out_[i], dout_[i])
                   : dx_op_.Recompute(x_[i], y_[i], out_[i], dout_[i]);
999 1000
    }
    if (dy_ != nullptr) {
C
chengduo 已提交
1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011
      dy_[i] = UseIntermediateOut
                   ? dy_op_.UseIntermediateOut(
                         x_[i], y_[i], intermediate_out_[i], out_[i], dout_[i])
                   : dy_op_.Recompute(x_[i], y_[i], out_[i], dout_[i]);
    }
    if (dintermediate_ != nullptr) {
      dintermediate_[i] =
          UseIntermediateOut
              ? dintermediate_op_.UseIntermediateOut(
                    x_[i], intermediate_out_[i], out_[i], dout_[i])
              : dintermediate_op_.Recompute(x_[i], y_[i], out_[i], dout_[i]);
1012 1013 1014 1015 1016 1017 1018 1019 1020 1021
    }
  }

  const T *x_;
  const T *y_;
  const T *intermediate_out_;
  const T *out_;
  const T *dout_;
  DX_OP dx_op_;
  DY_OP dy_op_;
C
chengduo 已提交
1022
  DIntermediate_OP dintermediate_op_;
1023 1024
  T *dx_;
  T *dy_;
C
chengduo 已提交
1025
  T *dintermediate_;
1026 1027 1028
};

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1029
          typename DIntermediate_OP, bool UseIntermediateOut>
1030 1031 1032 1033 1034
void FusedElemwiseAndActGradComputeNoBroadcast(
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::DDim &y_dim, const framework::Tensor *x,
    const framework::Tensor *y, const framework::Tensor *intermediate_out,
    const framework::Tensor *out, const framework::Tensor *dout, int axis,
C
chengduo 已提交
1035 1036 1037
    framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1038 1039 1040 1041
  size_t N = static_cast<size_t>(framework::product(x_dim));
  platform::ForRange<DeviceContext> for_range(
      ctx.template device_context<DeviceContext>(), N);
  for_range(
C
chengduo 已提交
1042 1043
      FusedElemwiseAndActGradNoBroadcast<T, DX_OP, DY_OP, DIntermediate_OP,
                                         UseIntermediateOut>{
1044 1045
          x->data<T>(), y->data<T>(),
          intermediate_out ? intermediate_out->data<T>() : nullptr,
C
chengduo 已提交
1046
          out->data<T>(), dout->data<T>(), dx_op, dy_op, dintermediate_op,
1047
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1048 1049 1050
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace())});
1051 1052
}

C
chengduo 已提交
1053 1054 1055 1056 1057 1058 1059
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActGradBroadcast1CPU(
    const T *x, const T *y, const T *intermediate_out, const T *out,
    const T *dout, int h, int w, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op, T *dx, T *dy, T *d_intermediate) {
1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074
  int64_t tmp_out_idx, x_idx, y_idx;
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      int offset = i * w + j;

      tmp_out_idx = BcastY ? j : offset;
      y_idx = BcastY ? j : offset;
      x_idx = BcastY ? offset : j;

      if (SameShapeOfIntermediateOutAndOut) {
        tmp_out_idx = offset;
      }

      if (dx != nullptr) {
        T tmp = UseIntermediateOut
C
chengduo 已提交
1075 1076 1077 1078 1079
                    ? dx_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                               intermediate_out[tmp_out_idx],
                                               out[offset], dout[offset])
                    : dx_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                      dout[offset]);
1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092

        if (BcastY) {
          dx[x_idx] = tmp;
        } else {
          if (i == 0) {
            dx[x_idx] = tmp;
          } else {
            dx[x_idx] += tmp;
          }
        }
      }
      if (dy != nullptr) {
        T tmp = UseIntermediateOut
C
chengduo 已提交
1093 1094 1095 1096 1097
                    ? dy_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                               intermediate_out[tmp_out_idx],
                                               out[offset], dout[offset])
                    : dy_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                      dout[offset]);
1098 1099 1100 1101 1102 1103 1104 1105 1106 1107
        if (BcastY) {
          if (i == 0) {
            dy[y_idx] = tmp;
          } else {
            dy[y_idx] += tmp;
          }
        } else {
          dy[y_idx] = tmp;
        }
      }
C
chengduo 已提交
1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
      if (d_intermediate != nullptr) {
        T tmp = UseIntermediateOut
                    ? dintermediate_op.UseIntermediateOut(
                          x[x_idx], intermediate_out[tmp_out_idx], out[offset],
                          dout[offset])
                    : dintermediate_op.Recompute(x[x_idx], y[y_idx],
                                                 out[offset], dout[i]);
        if (SameShapeOfIntermediateOutAndOut) {
          d_intermediate[tmp_out_idx] = tmp;
        } else {
          if (i == 0) {
            d_intermediate[tmp_out_idx] = tmp;
          } else {
            d_intermediate[tmp_out_idx] += tmp;
          }
        }
      }
1125 1126 1127 1128
    }
  }
}

C
chengduo 已提交
1129 1130 1131 1132 1133 1134 1135
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActGradBroadcast2CPU(
    const T *x, const T *y, const T *intermediate_out, const T *out,
    const T *dout, int pre, int n, int post, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op, T *dx, T *dy, T *d_intermediate) {
1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151
  int64_t tmp_out_idx, x_idx, y_idx;
  for (int i = 0; i < pre; ++i) {
    for (int j = 0; j < n; ++j) {
      for (int k = 0; k < post; ++k) {
        int offset = i * n * post + j * post + k;

        tmp_out_idx = BcastY ? j : offset;
        y_idx = BcastY ? j : offset;
        x_idx = BcastY ? offset : j;

        if (SameShapeOfIntermediateOutAndOut) {
          tmp_out_idx = offset;
        }

        if (dx != nullptr) {
          T tmp = UseIntermediateOut
C
chengduo 已提交
1152 1153 1154 1155 1156
                      ? dx_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                                 intermediate_out[tmp_out_idx],
                                                 out[offset], dout[offset])
                      : dx_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                        dout[offset]);
1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169

          if (BcastY) {
            dx[x_idx] = tmp;
          } else {
            if (i == 0 && k == 0) {
              dx[x_idx] = tmp;
            } else {
              dx[x_idx] += tmp;
            }
          }
        }
        if (dy != nullptr) {
          T tmp = UseIntermediateOut
C
chengduo 已提交
1170 1171 1172 1173 1174
                      ? dy_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                                 intermediate_out[tmp_out_idx],
                                                 out[offset], dout[offset])
                      : dy_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                        dout[offset]);
1175 1176 1177 1178 1179 1180 1181 1182 1183 1184
          if (BcastY) {
            if (i == 0 && k == 0) {
              dy[y_idx] = tmp;
            } else {
              dy[y_idx] += tmp;
            }
          } else {
            dy[y_idx] = tmp;
          }
        }
C
chengduo 已提交
1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201
        if (d_intermediate != nullptr) {
          T tmp = UseIntermediateOut
                      ? dintermediate_op.UseIntermediateOut(
                            x[x_idx], intermediate_out[tmp_out_idx],
                            out[offset], dout[offset])
                      : dintermediate_op.Recompute(x[x_idx], y[y_idx],
                                                   out[offset], dout[i]);
          if (SameShapeOfIntermediateOutAndOut) {
            d_intermediate[tmp_out_idx] = tmp;
          } else {
            if (i == 0) {
              d_intermediate[tmp_out_idx] = tmp;
            } else {
              d_intermediate[tmp_out_idx] += tmp;
            }
          }
        }
1202 1203 1204 1205 1206 1207
      }
    }
  }
}

#ifdef __NVCC__
C
chengduo 已提交
1208 1209 1210
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1211 1212
static __global__ void FusedElemwiseAndActGradBroadcast1CUDAKernel(
    const T *x, const T *y, const T *intermediate_out, const T *out,
C
chengduo 已提交
1213 1214
    const T *dout, int h, int w, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op, T *dx, T *dy, T *d_intermediate) {
1215 1216 1217
  int j = blockIdx.x;
  int i = threadIdx.x;
  int tid = threadIdx.x;
C
chengduo 已提交
1218
  T val(0), inter_val(0);
1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232
  int64_t tmp_out_idx, x_idx, y_idx;

  do {
    int offset = i * w + j;

    tmp_out_idx = BcastY ? j : offset;
    y_idx = BcastY ? j : offset;
    x_idx = BcastY ? offset : j;

    if (SameShapeOfIntermediateOutAndOut) {
      tmp_out_idx = offset;
    }

    if (dx != nullptr) {
C
chengduo 已提交
1233 1234 1235 1236 1237 1238
      T tmp =
          UseIntermediateOut
              ? dx_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                         intermediate_out[tmp_out_idx],
                                         out[offset], dout[offset])
              : dx_op.Recompute(x[x_idx], y[y_idx], out[offset], dout[offset]);
1239 1240 1241 1242 1243 1244 1245 1246

      if (BcastY) {
        dx[x_idx] = tmp;
      } else {
        val += tmp;
      }
    }
    if (dy != nullptr) {
C
chengduo 已提交
1247 1248 1249 1250 1251 1252
      T tmp =
          UseIntermediateOut
              ? dy_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                         intermediate_out[tmp_out_idx],
                                         out[offset], dout[offset])
              : dy_op.Recompute(x[x_idx], y[y_idx], out[offset], dout[offset]);
1253 1254 1255 1256 1257 1258
      if (BcastY) {
        val += tmp;
      } else {
        dy[y_idx] = tmp;
      }
    }
C
chengduo 已提交
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271
    if (d_intermediate != nullptr) {
      T tmp = UseIntermediateOut
                  ? dintermediate_op.UseIntermediateOut(
                        y[y_idx], intermediate_out[tmp_out_idx], out[offset],
                        dout[offset])
                  : dintermediate_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                               dout[offset]);
      if (SameShapeOfIntermediateOutAndOut) {
        d_intermediate[tmp_out_idx] = tmp;
      } else {
        inter_val += tmp;
      }
    }
1272 1273 1274 1275

    i += ELEMWISE_MAX_BLOCK_DIM;
  } while (i < h);

C
chengduo 已提交
1276
  h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291
  if (BcastY) {
    if (dy) {
      val = paddle::platform::reduceSum(val, tid, h);
      if (threadIdx.x == 0) {
        dy[j] = val;
      }
    }
  } else {
    if (dx) {
      val = paddle::platform::reduceSum(val, tid, h);
      if (threadIdx.x == 0) {
        dx[j] = val;
      }
    }
  }
C
chengduo 已提交
1292 1293 1294 1295 1296 1297 1298 1299
  if (!SameShapeOfIntermediateOutAndOut) {
    if (d_intermediate) {
      inter_val = paddle::platform::reduceSum(inter_val, tid, h);
      if (threadIdx.x == 0) {
        d_intermediate[j] = inter_val;
      }
    }
  }
1300 1301
}

C
chengduo 已提交
1302 1303 1304 1305 1306 1307 1308
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
static void FusedElemwiseAndActGradBroadcast1CUDA(
    cudaStream_t stream, const T *x, const T *y, const T *intermediate_out,
    const T *out, const T *dout, int h, int w, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op, T *dx, T *dy, T *d_intermediate) {
1309 1310 1311
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, h);
  int gird_size = w;
  FusedElemwiseAndActGradBroadcast1CUDAKernel<
C
chengduo 已提交
1312
      T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut, BcastY,
1313
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
C
chengduo 已提交
1314 1315
      x, y, intermediate_out, out, dout, h, w, dx_op, dy_op, dintermediate_op,
      dx, dy, d_intermediate);
1316 1317
}

C
chengduo 已提交
1318 1319 1320
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1321 1322
static __global__ void FusedElemwiseAndActGradBroadcast2CUDAKernel(
    const T *x, const T *y, const T *intermediate_out, const T *out,
C
chengduo 已提交
1323 1324
    const T *dout, int pre, int n, int post, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op, T *dx, T *dy, T *d_intermediate) {
1325 1326 1327
  int tid = threadIdx.x;
  int j = blockIdx.x;

C
chengduo 已提交
1328
  T val(0), inter_val(0);
1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346
  int ttid = tid;
  int64_t tmp_out_idx, x_idx, y_idx;
  while (true) {
    int i = ttid / post;
    int k = ttid % post;
    if (i >= pre) break;

    int offset = i * n * post + j * post + k;

    tmp_out_idx = BcastY ? j : offset;
    y_idx = BcastY ? j : offset;
    x_idx = BcastY ? offset : j;

    if (SameShapeOfIntermediateOutAndOut) {
      tmp_out_idx = offset;
    }

    if (dx != nullptr) {
C
chengduo 已提交
1347 1348 1349 1350 1351 1352
      T tmp =
          UseIntermediateOut
              ? dx_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                         intermediate_out[tmp_out_idx],
                                         out[offset], dout[offset])
              : dx_op.Recompute(x[x_idx], y[y_idx], out[offset], dout[offset]);
1353 1354 1355 1356 1357 1358 1359 1360

      if (BcastY) {
        dx[x_idx] = tmp;
      } else {
        val += tmp;
      }
    }
    if (dy != nullptr) {
C
chengduo 已提交
1361 1362 1363 1364 1365 1366
      T tmp =
          UseIntermediateOut
              ? dy_op.UseIntermediateOut(x[x_idx], y[y_idx],
                                         intermediate_out[tmp_out_idx],
                                         out[offset], dout[offset])
              : dy_op.Recompute(x[x_idx], y[y_idx], out[offset], dout[offset]);
1367 1368 1369 1370 1371 1372
      if (BcastY) {
        val += tmp;
      } else {
        dy[y_idx] = tmp;
      }
    }
C
chengduo 已提交
1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385
    if (d_intermediate != nullptr) {
      T tmp = UseIntermediateOut
                  ? dintermediate_op.UseIntermediateOut(
                        y[y_idx], intermediate_out[tmp_out_idx], out[offset],
                        dout[offset])
                  : dintermediate_op.Recompute(x[x_idx], y[y_idx], out[offset],
                                               dout[offset]);
      if (SameShapeOfIntermediateOutAndOut) {
        d_intermediate[tmp_out_idx] = tmp;
      } else {
        inter_val += tmp;
      }
    }
1386 1387 1388
    ttid += ELEMWISE_MAX_BLOCK_DIM;
  }

C
chengduo 已提交
1389 1390
  int h = pre * post;
  h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405
  if (BcastY) {
    if (dy) {
      val = paddle::platform::reduceSum(val, tid, h);
      if (threadIdx.x == 0) {
        dy[j] = val;
      }
    }
  } else {
    if (dx) {
      val = paddle::platform::reduceSum(val, tid, h);
      if (threadIdx.x == 0) {
        dx[j] = val;
      }
    }
  }
C
chengduo 已提交
1406 1407 1408 1409 1410 1411 1412 1413
  if (!SameShapeOfIntermediateOutAndOut) {
    if (d_intermediate) {
      inter_val = paddle::platform::reduceSum(inter_val, tid, h);
      if (threadIdx.x == 0) {
        d_intermediate[j] = inter_val;
      }
    }
  }
1414 1415
}

C
chengduo 已提交
1416 1417 1418
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1419 1420 1421
static void FusedElemwiseAndActGradBroadcast2CUDA(
    cudaStream_t stream, const T *x, const T *y, const T *intermediate_out,
    const T *out, const T *dout, int pre, int n, int post, DX_OP dx_op,
C
chengduo 已提交
1422 1423
    DY_OP dy_op, DIntermediate_OP dintermediate_op, T *dx, T *dy,
    T *dintermediate) {
1424 1425 1426
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, pre * post);
  int gird_size = n;
  FusedElemwiseAndActGradBroadcast2CUDAKernel<
C
chengduo 已提交
1427
      T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut, BcastY,
1428
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
C
chengduo 已提交
1429 1430
      x, y, intermediate_out, out, dout, pre, n, post, dx_op, dy_op,
      dintermediate_op, dx, dy, dintermediate);
1431 1432 1433 1434
}
#endif

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1435
          typename DIntermediate_OP, bool UseIntermediateOut, bool BcastY,
1436 1437 1438 1439 1440 1441
          bool SameShapeOfIntermediateOutAndOut>
void FusedElemwiseAndActGradComputeWithBroadcast(
    const framework::ExecutionContext &ctx, const framework::DDim &x_dim,
    const framework::DDim &y_dim_untrimed, const framework::Tensor *x,
    const framework::Tensor *y, const framework::Tensor *intermediate_out,
    const framework::Tensor *out, const framework::Tensor *dout, int axis,
C
chengduo 已提交
1442 1443 1444
    framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455
  axis = (axis == -1 ? x_dim.size() - y_dim_untrimed.size() : axis);
  auto y_dim = trim_trailing_singular_dims(y_dim_untrimed);
  axis = (y_dim.size() == 0) ? x_dim.size() : axis;

  int pre, n, post;
  get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post);
  if (post == 1) {
    int h = pre;
    int w = n;
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
C
chengduo 已提交
1456 1457
      FusedElemwiseAndActGradBroadcast1CUDA<T, DX_OP, DY_OP, DIntermediate_OP,
                                            UseIntermediateOut, BcastY,
1458 1459 1460 1461
                                            SameShapeOfIntermediateOutAndOut>(
          ctx.template device_context<DeviceContext>().stream(), x->data<T>(),
          y->data<T>(),
          intermediate_out == nullptr ? nullptr : intermediate_out->data<T>(),
C
chengduo 已提交
1462
          out->data<T>(), dout->data<T>(), h, w, dx_op, dy_op, dintermediate_op,
1463
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1464 1465 1466
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1467 1468
#endif
    } else {
C
chengduo 已提交
1469 1470
      FusedElemwiseAndActGradBroadcast1CPU<T, DX_OP, DY_OP, DIntermediate_OP,
                                           UseIntermediateOut, BcastY,
1471 1472 1473
                                           SameShapeOfIntermediateOutAndOut>(
          x->data<T>(), y->data<T>(),
          intermediate_out == nullptr ? nullptr : intermediate_out->data<T>(),
C
chengduo 已提交
1474
          out->data<T>(), dout->data<T>(), h, w, dx_op, dy_op, dintermediate_op,
1475
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1476 1477 1478
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1479 1480 1481 1482
    }
  } else {
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
C
chengduo 已提交
1483 1484
      FusedElemwiseAndActGradBroadcast2CUDA<T, DX_OP, DY_OP, DIntermediate_OP,
                                            UseIntermediateOut, BcastY,
1485 1486 1487 1488 1489
                                            SameShapeOfIntermediateOutAndOut>(
          ctx.template device_context<DeviceContext>().stream(), x->data<T>(),
          y->data<T>(),
          intermediate_out == nullptr ? nullptr : intermediate_out->data<T>(),
          out->data<T>(), dout->data<T>(), pre, n, post, dx_op, dy_op,
C
chengduo 已提交
1490
          dintermediate_op,
1491
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1492 1493 1494
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1495 1496
#endif
    } else {
C
chengduo 已提交
1497 1498
      FusedElemwiseAndActGradBroadcast2CPU<T, DX_OP, DY_OP, DIntermediate_OP,
                                           UseIntermediateOut, BcastY,
1499 1500 1501 1502
                                           SameShapeOfIntermediateOutAndOut>(
          x->data<T>(), y->data<T>(),
          intermediate_out == nullptr ? nullptr : intermediate_out->data<T>(),
          out->data<T>(), dout->data<T>(), pre, n, post, dx_op, dy_op,
C
chengduo 已提交
1503
          dintermediate_op,
1504
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1505 1506 1507
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1508 1509 1510 1511 1512
    }
  }
}

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1513 1514
          typename DIntermediate_OP, bool UseIntermediateOut,
          bool SameShapeOfIntermediateOutAndOut>
1515 1516 1517 1518
void FusedElemwiseAndActGradComputeEx(
    const framework::ExecutionContext &ctx, const framework::Tensor *x,
    const framework::Tensor *y, const framework::Tensor *out,
    const framework::Tensor *intermediate_out, const framework::Tensor *dout,
C
chengduo 已提交
1519 1520 1521
    int axis, framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1522 1523 1524 1525 1526 1527
  const framework::DDim &x_dim = x->dims();
  const framework::DDim &y_dim = y->dims();
  if (UseIntermediateOut) {
    PADDLE_ENFORCE(intermediate_out, "intermediate_out should not be nullptr");
  }
  if (x_dim == y_dim) {
C
chengduo 已提交
1528 1529
    FusedElemwiseAndActGradComputeNoBroadcast<
        DeviceContext, T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut>(
1530
        ctx, x_dim, y_dim, x, y, intermediate_out, out, dout, axis, dx, dy,
C
chengduo 已提交
1531
        dintermediate, dx_op, dy_op, dintermediate_op);
1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546
  } else {  // Y is a scalar
    bool bcast_y = x_dim.size() >= y_dim.size();
    if (x_dim.size() == y_dim.size()) {
      for (int i = 0; i < x_dim.size(); ++i) {
        if (x_dim[i] < y_dim[i]) {
          bcast_y = false;
          break;
        }
      }
    }

    // z = f1(x, f2(y))
    // z = f1(f2(x, y))
    if (bcast_y) {  // Y should be broadcast.
      FusedElemwiseAndActGradComputeWithBroadcast<
C
chengduo 已提交
1547 1548 1549 1550
          DeviceContext, T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut,
          true /*BcastY*/, SameShapeOfIntermediateOutAndOut>(
          ctx, x_dim, y_dim, x, y, intermediate_out, out, dout, axis, dx, dy,
          dintermediate, dx_op, dy_op, dintermediate_op);
1551 1552
    } else {
      FusedElemwiseAndActGradComputeWithBroadcast<
C
chengduo 已提交
1553 1554 1555 1556
          DeviceContext, T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut,
          false /*BcastY*/, SameShapeOfIntermediateOutAndOut>(
          ctx, y_dim, x_dim, x, y, intermediate_out, out, dout, axis, dx, dy,
          dintermediate, dx_op, dy_op, dintermediate_op);
1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570
    }
  }
}

template <typename DeviceContext, typename T, typename CompoundFunctor,
          bool KeepIntermediateOut, bool SameShapeOfIntermediateOutAndOut>
void FusedElemwiseAndActComputeEx(const framework::ExecutionContext &ctx,
                                  const framework::Tensor &x,
                                  const framework::Tensor &y, int axis,
                                  CompoundFunctor compound_functor,
                                  framework::Tensor *out,
                                  framework::Tensor *intermediate_out) {
  if (KeepIntermediateOut) {
    PADDLE_ENFORCE(intermediate_out,
C
chengduo 已提交
1571
                   "The save_intermediate_out is opened, "
1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622
                   "intermediate_out should not be nullptr.");
  }

  const framework::DDim &x_dim = x.dims();
  const framework::DDim &y_dim = y.dims();
  if (x.dims() == y.dims()) {
    FusedElemwiseAndActComputeNoBroadcast<DeviceContext, T, CompoundFunctor,
                                          KeepIntermediateOut>(
        ctx, x_dim, x, y, compound_functor, out, intermediate_out);
  } else {
    // Whether the shape of Y is a continuous subsequence of X,
    // For more information please refer to the op's introduction.
    bool bcast_y = x.dims().size() >= y.dims().size();
    if (x.dims().size() == y.dims().size()) {
      for (int i = 0; i < x.dims().size(); ++i) {
        if (x.dims()[i] < y.dims()[i]) {
          bcast_y = false;
          break;
        }
      }
    }

    // z = f1(x, f2(y))
    // z = f1(f2(x, y))
    if (bcast_y) {  // Y should be broadcast.
      // In this case,
      // for 'f2(y)', the shape of intermediate_out should be equal to the shape
      // of Y.
      // for 'f2(x, y)', the shape of intermediate_out should be equal to the
      // shape of Out.
      // the shape of Out should be equal to the shape of X.
      FusedElemwiseAndActComputeWithBroadcast<
          DeviceContext, T, CompoundFunctor, true /*BcastY*/,
          KeepIntermediateOut, SameShapeOfIntermediateOutAndOut>(
          ctx, x_dim /*OutShape*/, y_dim, x, y, compound_functor, axis, out,
          intermediate_out);
    } else {
      // In this case,
      // for 'f2(y)', the shape of intermediate_out should be equal to the shape
      // of Out.
      // for 'f2(x, y)', the shape of intermediate_out should be equal to the
      // shape of Out.
      // the shape of Out should be equal to the shape of Y.
      FusedElemwiseAndActComputeWithBroadcast<
          DeviceContext, T, CompoundFunctor, false /*BcastY*/,
          KeepIntermediateOut, SameShapeOfIntermediateOutAndOut>(
          ctx, y_dim /*OutShape*/, x_dim, x, y, compound_functor, axis, out,
          intermediate_out);
    }
  }
}
1623 1624
}  // namespace operators
}  // namespace paddle