elementwise_op_function.h 64.6 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 54 55
 *
 * New parameter: *mid_flag* is added to solve m*n*k & m*1*k
 * broadcast cases.
 * 3. shape(X) = (2, 3, 4, 5), shape(Y) = (2, 1, 4, 5)
 *    mid_flag should not be NULL.
 *    x.shape(2, 3, 20) * y.shape(2, 1, 20).broadcast(2, 3, 20)
56
 */
57 58
inline void get_mid_dims(const framework::DDim &x_dims,
                         const framework::DDim &y_dims, const int axis,
59
                         int *pre, int *n, int *post, int *mid_flag = NULL) {
60 61 62
  *pre = 1;
  *n = 1;
  *post = 1;
63 64 65 66 67 68 69 70 71 72 73
  if (mid_flag != NULL) {
    *mid_flag = 0;
    int mid = 0;
    for (int i = 0; i < axis; ++i) {
      (*pre) *= x_dims[i];
    }
    for (int i = 0; i < y_dims.size(); ++i) {
      if (x_dims[i + axis] != y_dims[i]) {
        // only support single y_dims[i] = 1 now.
        PADDLE_ENFORCE_EQ(*mid_flag, 0,
                          "Broadcast support y_dims with single 1.");
74 75 76 77 78 79
        PADDLE_ENFORCE_EQ(y_dims[i], 1,
                          "ShapeError: broadcast dimension mismatch. Operands "
                          "could not be broadcast together with the shape of "
                          "X = [%s] and the shape of Y = [%s]. Received [%d] "
                          "in X is not equal to [%d] in Y",
                          x_dims, y_dims, x_dims[i + axis], y_dims[i]);
80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103
        // m*n*k m*1*k
        for (int j = 0; j < i; ++j) {
          (*pre) *= y_dims[j];
        }
        *n = std::max(x_dims[i + axis], y_dims[i]);
        *mid_flag = 1;
        mid = i;
        break;
      }
      (*n) *= y_dims[i];
    }
    if (*mid_flag) {
      for (int i = mid + 1; i < x_dims.size(); ++i) {
        (*post) *= x_dims[i];
      }
    } else {
      for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
        (*post) *= x_dims[i];
      }
    }
  } else {  // for fused_elementwise_activation_op. keep the old version.
    for (int i = 0; i < axis; ++i) {
      (*pre) *= x_dims[i];
    }
104

105 106 107 108 109
    for (int i = 0; i < y_dims.size(); ++i) {
      PADDLE_ENFORCE_EQ(x_dims[i + axis], y_dims[i],
                        "Broadcast dimension mismatch.");
      (*n) *= y_dims[i];
    }
110

111 112 113
    for (int i = axis + y_dims.size(); i < x_dims.size(); ++i) {
      (*post) *= x_dims[i];
    }
114 115 116
  }
}

117
inline framework::DDim trim_trailing_singular_dims(
118
    const framework::DDim &dims) {
119
  // Remove trailing dimensions of size 1 for y
120
  auto actual_dims_size = dims.size();
121
  for (; actual_dims_size != 0; --actual_dims_size) {
122
    if (dims[actual_dims_size - 1] != 1) break;
123
  }
124 125 126 127 128

  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];
129
  }
130 131 132
  if (trim_dims.size() == 0) {
    return framework::DDim(framework::make_dim());
  }
133 134
  framework::DDim actual_dims = framework::make_ddim(trim_dims);
  return actual_dims;
135 136
}

Q
QI JUN 已提交
137
template <typename T, typename DeviceContext>
C
chengduoZH 已提交
138
class RowwiseTransformIterator;
139

Q
QI JUN 已提交
140
template <typename T, typename DeviceContext>
C
chengduoZH 已提交
141
class MidWiseTransformIterator;
C
chengduoZH 已提交
142

D
dzhwinter 已提交
143
// NOTE(dzhwinter): ptrdiff_t in iterator is deperecated in c++17
C
chengduoZH 已提交
144
template <typename T>
D
dzhwinter 已提交
145 146 147
class RowwiseTransformIterator<T, platform::CPUDeviceContext>
    : public std::iterator<std::random_access_iterator_tag, T, std::ptrdiff_t,
                           T *, T &> {
C
chengduoZH 已提交
148
 public:
149
  RowwiseTransformIterator(const T *ptr, int n) : ptr_(ptr), i_(0), n_(n) {}
C
chengduoZH 已提交
150

151
  RowwiseTransformIterator<T, platform::CPUDeviceContext> &operator++() {
C
chengduoZH 已提交
152
    ++i_;
C
chengduoZH 已提交
153 154 155
    if (UNLIKELY(i_ == n_)) {
      i_ = 0;
    }
C
chengduoZH 已提交
156 157 158
    return *this;
  }

P
peizhilin 已提交
159
  RowwiseTransformIterator<T, platform::CPUDeviceContext> &operator+(int n) {
P
peizhilin 已提交
160
    while (n-- > 0) {
P
peizhilin 已提交
161 162 163 164 165 166 167 168 169
      ++i_;
      if (UNLIKELY(i_ == n_)) {
        i_ = 0;
      }
    }

    return *this;
  }

170 171
  bool operator==(const RowwiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
172
    return (ptr_ + i_) == &(*rhs);
C
chengduoZH 已提交
173 174
  }

175 176
  bool operator!=(const RowwiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
177
    return (ptr_ + i_) != &(*rhs);
C
chengduoZH 已提交
178 179
  }

180
  const T &operator*() { return ptr_[i_]; }
C
chengduoZH 已提交
181

C
chengduoZH 已提交
182
 private:
183
  const T *ptr_;
C
chengduoZH 已提交
184
  int i_;
C
chengduoZH 已提交
185
  int64_t n_;
C
chengduoZH 已提交
186 187 188
};

template <typename T>
D
dzhwinter 已提交
189 190 191
class MidWiseTransformIterator<T, platform::CPUDeviceContext>
    : public std::iterator<std::random_access_iterator_tag, T, std::ptrdiff_t,
                           T *, T &> {
C
chengduoZH 已提交
192
 public:
193
  MidWiseTransformIterator(const T *ptr, int n, int post)
C
chengduoZH 已提交
194 195
      : ptr_(ptr), i_(0), j_(0), n_(n), post_(post) {}

196
  MidWiseTransformIterator<T, platform::CPUDeviceContext> &operator++() {
C
chengduoZH 已提交
197
    ++j_;
C
chengduoZH 已提交
198 199
    if (UNLIKELY(j_ == post_)) {
      ++i_;
C
refine  
chengduoZH 已提交
200
      j_ = 0;
C
chengduoZH 已提交
201 202 203
      if (UNLIKELY(i_ == n_)) {
        i_ = 0;
      }
C
chengduoZH 已提交
204
    }
C
chengduoZH 已提交
205 206 207
    return *this;
  }

P
peizhilin 已提交
208
  MidWiseTransformIterator<T, platform::CPUDeviceContext> &operator+(int n) {
P
peizhilin 已提交
209
    while (n-- > 0) {
P
peizhilin 已提交
210 211 212 213 214 215 216 217 218 219 220 221
      ++j_;
      if (UNLIKELY(j_ == post_)) {
        ++i_;
        j_ = 0;
        if (UNLIKELY(i_ == n_)) {
          i_ = 0;
        }
      }
    }
    return *this;
  }

222 223
  bool operator==(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
224
    return (ptr_ + i_) == &(*rhs);
C
chengduoZH 已提交
225 226
  }

227 228
  bool operator!=(const MidWiseTransformIterator<T, platform::CPUDeviceContext>
                      &rhs) const {
C
chengduoZH 已提交
229
    return (ptr_ + i_) != &(*rhs);
C
chengduoZH 已提交
230 231
  }

232
  const T &operator*() { return ptr_[i_]; }
C
chengduoZH 已提交
233

C
chengduoZH 已提交
234
 private:
235
  const T *ptr_;
C
refine  
chengduoZH 已提交
236
  int64_t i_;
C
chengduoZH 已提交
237 238
  int64_t j_;
  int64_t n_;
C
refine  
chengduoZH 已提交
239
  int64_t post_;
C
chengduoZH 已提交
240 241
};

C
chengduoZH 已提交
242 243
#ifdef __NVCC__
template <typename T>
Q
QI JUN 已提交
244
class RowwiseTransformIterator<T, platform::CUDADeviceContext>
C
chengduoZH 已提交
245
    : public thrust::iterator_adaptor<
246
          RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T *> {
C
chengduoZH 已提交
247 248
 public:
  typedef thrust::iterator_adaptor<
249
      RowwiseTransformIterator<T, platform::CUDADeviceContext>, const T *>
C
chengduoZH 已提交
250
      super_t;
251
  HOSTDEVICE RowwiseTransformIterator(const T *x, int n)
252
      : super_t(x), begin_(x), n_(n) {}
C
chengduoZH 已提交
253 254 255 256
  friend class thrust::iterator_core_access;

 private:
  unsigned int n_;
257
  const T *begin_;
C
chengduoZH 已提交
258
  HOSTDEVICE typename super_t::reference dereference() const {
C
chengduoZH 已提交
259 260 261 262 263
    return *(begin_ + (this->base() - begin_) % n_);
  }
};

template <typename T>
Q
QI JUN 已提交
264
class MidWiseTransformIterator<T, platform::CUDADeviceContext>
C
chengduoZH 已提交
265
    : public thrust::iterator_adaptor<
266
          MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T *> {
C
chengduoZH 已提交
267 268
 public:
  typedef thrust::iterator_adaptor<
269
      MidWiseTransformIterator<T, platform::CUDADeviceContext>, const T *>
C
chengduoZH 已提交
270
      super_t;
271
  HOSTDEVICE MidWiseTransformIterator(const T *x, int n, int post)
272
      : super_t(x), begin_(x), n_(n), post_(post) {}
C
chengduoZH 已提交
273 274 275 276 277
  friend class thrust::iterator_core_access;

 private:
  unsigned int post_;
  unsigned int n_;
278
  const T *begin_;
C
chengduoZH 已提交
279
  HOSTDEVICE typename super_t::reference dereference() const {
C
chengduoZH 已提交
280 281 282 283 284
    return *(begin_ + (((this->base() - begin_) / post_) % n_));
  }
};
#endif

285 286
template <typename Functor, typename T, typename DeviceContext,
          typename OutType = T>
C
chengduoZH 已提交
287 288
class TransformFunctor {
 public:
289 290
  TransformFunctor(const framework::Tensor *x, const framework::Tensor *y,
                   framework::Tensor *z, const DeviceContext &ctx, Functor func)
C
chengduoZH 已提交
291 292
      : x_(x->data<T>()),
        y_(y->data<T>()),
293
        z_(z->mutable_data<OutType>(ctx.GetPlace())),
C
chengduoZH 已提交
294 295 296 297 298
        nx_(x->numel()),
        ctx_(ctx),
        func_(func) {}

  inline void Run() const {
Q
QI JUN 已提交
299
    platform::Transform<DeviceContext> trans;
C
chengduoZH 已提交
300
    trans(ctx_, x_, x_ + nx_, y_, z_, func_);
C
chengduoZH 已提交
301 302 303
  }

  inline void RunRowWise(int n, int pre) const {
Q
QI JUN 已提交
304 305 306
    platform::Transform<DeviceContext> trans;
    trans(ctx_, x_, x_ + nx_, RowwiseTransformIterator<T, DeviceContext>(y_, n),
          z_, func_);
C
chengduoZH 已提交
307 308 309
  }

  inline void RunMidWise(int n, int pre, int post) const {
Q
QI JUN 已提交
310 311 312
    platform::Transform<DeviceContext> trans;
    trans(ctx_, x_, x_ + nx_,
          MidWiseTransformIterator<T, DeviceContext>(y_, n, post), z_, func_);
C
chengduoZH 已提交
313 314
  }

315 316 317 318 319 320 321 322 323
  inline void RunMidRowWise(int n, int pre, int post) const {
    platform::Transform<DeviceContext> trans;
    for (int i = 0; i < pre; i++) {
      trans(ctx_, x_ + i * n * post, x_ + (i + 1) * n * post,
            RowwiseTransformIterator<T, DeviceContext>(y_ + i * post, post),
            z_ + i * n * post, func_);
    }
  }

C
chengduoZH 已提交
324
 private:
325 326 327
  const T *x_;
  const T *y_;
  OutType *z_;
C
chengduoZH 已提交
328
  int64_t nx_;
329
  const DeviceContext &ctx_;
C
chengduoZH 已提交
330 331 332
  Functor func_;
};

Y
Yu Yang 已提交
333 334
template <typename T, typename DX_OP, typename DY_OP>
struct ElemwiseGradNoBroadcast {
335 336 337 338
  const T *x_;
  const T *y_;
  const T *out_;
  const T *dout_;
Y
Yu Yang 已提交
339 340 341 342 343 344

  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
345
      dy_[i] = dy_op_(x_[i], y_[i], out_[i], dout_[i]);
Y
Yu Yang 已提交
346 347 348 349 350
    }
  }

  DX_OP dx_op_;
  DY_OP dy_op_;
351 352
  T *dx_;
  T *dy_;
Y
Yu Yang 已提交
353 354 355
};

template <typename T, typename DX_OP, typename DY_OP>
356 357 358
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
359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375
  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;
        }
      }
    }
  }
}
376

D
dzhwinter 已提交
377
#ifdef __NVCC__
Y
Yu Yang 已提交
378 379
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void ElemwiseGradBroadcast1CUDAKernel(
380 381
    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
382 383 384
  int j = blockIdx.x;
  int i = threadIdx.x;
  int tid = threadIdx.x;
C
chengduo 已提交
385
  T val(0);
Y
Yu Yang 已提交
386 387 388 389 390 391 392

  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
393
      val += dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
Y
Yu Yang 已提交
394 395 396 397 398
    }
    i += ELEMWISE_MAX_BLOCK_DIM;
  } while (i < h);

  if (dy) {
C
chengduoZH 已提交
399
    h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
400
    val = paddle::platform::reduceSum(val, tid, h);
Y
Yu Yang 已提交
401
    if (threadIdx.x == 0) {
C
chengduoZH 已提交
402
      dy[j] = val;
Y
Yu Yang 已提交
403 404 405 406
    }
  }
}

407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456
#define BLOCK_X 32
#define BLOCK_Y 32

// suppose use 2D block is fast because more parallel
// and memory coalesced
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void FastElemwiseGradBroadcast1CUDAKernel(
    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) {
  __shared__ T sdata[BLOCK_Y][BLOCK_X + 1];

  T val(0);
  size_t width_stride = gridDim.x * blockDim.x;
  size_t idx = threadIdx.x + blockDim.x * blockIdx.x;
  size_t full_width =
      (w & (~((uint64_t)(BLOCK_X - 1)))) + ((w & (BLOCK_X - 1)) ? BLOCK_X : 0);
  size_t full_height =
      (h & (~((uint64_t)(BLOCK_Y - 1)))) + ((h & (BLOCK_Y - 1)) ? BLOCK_Y : 0);

  for (int m = idx; m < full_width; m += width_stride) {
    sdata[threadIdx.y][threadIdx.x] = 0;
    for (int n = threadIdx.y; n < full_height; n += BLOCK_Y) {
      int x_offset = n * w + m;
      if (dx && m < w && n < h) {
        dx[x_offset] = dx_op(x[x_offset], y[m], out[x_offset], dout[x_offset]);
      }
      if (dy) {
        if (m < w && n < h) {
          T val = dy_op(x[x_offset], y[m], out[x_offset], dout[x_offset]);
          sdata[threadIdx.y][threadIdx.x] += val;
        }
        __syncthreads();
      }
    }
    if (dy) {
      T my_val = sdata[threadIdx.x][threadIdx.y];
      for (int i = warpSize >> 1; i > 0; i >>= 1)
        my_val += platform::CudaShuffleXorSync(0xFFFFFFFF, my_val, i);
      __syncthreads();
      if ((threadIdx.x == 0)) {
        sdata[0][threadIdx.y] = my_val;
      }
      __syncthreads();
      if (threadIdx.y == 0 && m < w) {
        dy[m] = sdata[0][threadIdx.x];
      }
    }
  }
}

Y
Yu Yang 已提交
457
template <typename T, typename DX_OP, typename DY_OP>
458 459
static void ElemwiseGradBroadcast1CUDA(cudaStream_t stream, const T *x,
                                       const T *y, const T *out, const T *dout,
Y
Yu Yang 已提交
460
                                       int h, int w, DX_OP dx_op, DY_OP dy_op,
461
                                       T *dx, T *dy) {
462 463 464 465 466 467 468 469 470 471 472 473 474 475
  // For small case use 1D block
  constexpr int half_walf = 16;
  if (w < half_walf || h < half_walf) {
    int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, h);
    int gird_size = w;
    ElemwiseGradBroadcast1CUDAKernel<<<gird_size, block_size, 0, stream>>>(
        x, y, out, dout, h, w, dx_op, dy_op, dx, dy);
  } else {
    // suppose perfoemance improves with h increased.
    dim3 block_size = dim3(BLOCK_X, BLOCK_Y);
    int grid_size = (w + BLOCK_X - 1) / BLOCK_X;
    FastElemwiseGradBroadcast1CUDAKernel<<<grid_size, block_size, 0, stream>>>(
        x, y, out, dout, h, w, dx_op, dy_op, dx, dy);
  }
Y
Yu Yang 已提交
476 477 478 479 480
}

#endif

template <typename T, typename DX_OP, typename DY_OP>
481 482 483
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
484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507
  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(
508 509
    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
510 511 512
  int tid = threadIdx.x;
  int j = blockIdx.x;

C
chengduo 已提交
513
  T val(0);
Y
Yu Yang 已提交
514 515 516 517 518 519 520 521 522 523 524 525 526 527
  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
528
      val += dy_op(x[x_offset], y[j], out[x_offset], dout[x_offset]);
Y
Yu Yang 已提交
529 530 531 532 533 534
    }

    ttid += ELEMWISE_MAX_BLOCK_DIM;
  }

  if (dy) {
C
chengduoZH 已提交
535 536
    int h = pre * post;
    h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
537
    val = paddle::platform::reduceSum(val, tid, h);
C
chengduoZH 已提交
538
    if (threadIdx.x == 0) {
C
chengduoZH 已提交
539
      dy[j] = val;
Y
Yu Yang 已提交
540 541 542 543 544
    }
  }
}

template <typename T, typename DX_OP, typename DY_OP>
545 546
static void ElemwiseGradBroadcast2CUDA(cudaStream_t stream, const T *x,
                                       const T *y, const T *out, const T *dout,
Y
Yu Yang 已提交
547
                                       int pre, int n, int post, DX_OP dx_op,
548
                                       DY_OP dy_op, T *dx, T *dy) {
Y
Yu Yang 已提交
549 550
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, pre * post);
  int gird_size = n;
C
chengduoZH 已提交
551 552
  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
553 554 555 556
}

#endif

557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638
template <typename T, typename DX_OP, typename DY_OP>
static void ElemwiseGradBroadcastMid2CPU(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) {
  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;
        int y_offset = i * post + k;
        if (dx != nullptr) {
          dx[x_offset] =
              dx_op(x[x_offset], y[y_offset], out[x_offset], dout[x_offset]);
        }
        if (dy != nullptr) {
          T tmp =
              dy_op(x[x_offset], y[y_offset], out[x_offset], dout[x_offset]);
          if (j == 0) {
            dy[y_offset] = tmp;
          } else {
            dy[y_offset] += tmp;
          }
        }
      }
    }
  }
}

#ifdef __NVCC__
template <typename T, typename DX_OP, typename DY_OP>
static __global__ void ElemwiseGradBroadcastMid2CUDAKernel(
    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) {
  int j = threadIdx.x;
  int tid = blockIdx.x;

  T val(0);
  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;
    int y_offset = i * post + k;
    if (dx != nullptr) {
      dx[x_offset] =
          dx_op(x[x_offset], y[y_offset], out[x_offset], dout[x_offset]);
    }

    if (dy != nullptr) {
      val += dy_op(x[x_offset], y[y_offset], out[x_offset], dout[x_offset]);
    }

    ttid += ELEMWISE_MAX_BLOCK_DIM;
  }

  if (dy) {
    int h = n;
    h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
    val = paddle::platform::reduceSum(val, j, h);
    if (threadIdx.x == 0) {
      dy[tid] = val;
    }
  }
}

template <typename T, typename DX_OP, typename DY_OP>
static void ElemwiseGradBroadcastMid2CUDA(cudaStream_t stream, 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) {
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, n);
  int gird_size = pre * post;
  ElemwiseGradBroadcastMid2CUDAKernel<<<gird_size, block_size, 0, stream>>>(
      x, y, out, dout, pre, n, post, dx_op, dy_op, dx, dy);
}

#endif

639 640
template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
void ElemwiseGradComputeNoBroadcast(
641 642 643 644 645
    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) {
646
  size_t N = static_cast<size_t>(framework::product(x_dim));
D
dzhwinter 已提交
647
#if !defined(_WIN32)
648 649
  platform::ForRange<DeviceContext> for_range(
      ctx.template device_context<DeviceContext>(), N);
D
dzhwinter 已提交
650 651 652 653
#else
  platform::ForRange<DeviceContext> for_range(
      ctx.device_context<DeviceContext>(), N);
#endif  // !_WIN32
654 655 656 657 658 659 660 661
  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(
662 663 664 665 666
    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) {
667 668 669 670
  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;

671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690
  int pre, n, post, mid_flag = 0;
  get_mid_dims(x_dim, y_dim, axis, &pre, &n, &post, &mid_flag);
  if (mid_flag) {
    PADDLE_ENFORCE_EQ(mid_flag, 1, "mid_flag should be no more than 1.");
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      ElemwiseGradBroadcastMid2CUDA(
          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 {
      ElemwiseGradBroadcastMid2CPU(
          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()));
    }
  } else if (post == 1) {
691 692 693 694
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
      ElemwiseGradBroadcast1CUDA(
          ctx.template device_context<DeviceContext>().stream(), x.data<T>(),
695
          y.data<T>(), out.data<T>(), dout.data<T>(), pre, n, dx_op, dy_op,
696 697 698 699 700
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()));
#endif
    } else {
      ElemwiseGradBroadcast1CPU(
701 702 703
          x.data<T>(), y.data<T>(), out.data<T>(), dout.data<T>(), pre, n,
          dx_op, dy_op,
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
          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
725
template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP>
726 727 728 729 730
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
731
                         DX_OP dx_op, DY_OP dy_op) {
732 733
  const framework::DDim &x_dim = x.dims();
  const framework::DDim &y_dim = y.dims();
Y
Yu Yang 已提交
734
  if (x.dims() == y.dims()) {
735 736
    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
737
  } else {  // Y is a scalar
738 739 740 741 742 743 744 745 746 747
    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>
748 749 750 751 752 753
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,
754 755
                                 DX_OP dx_op, DY_OP dy_op) {
  if (dy == nullptr) {
756
    const framework::DDim &dx_dims = dout.dims();
757 758 759 760 761
    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()) {
762 763
      const framework::DDim &dx_dims = dout.dims();
      const framework::DDim &dy_dims = dy->dims();
764 765 766 767
      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();
768
      const framework::DDim &dy_dims = dy->dims();
769 770
      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
771 772
    }
  }
773
}
Y
Yu Yang 已提交
774

775
// Deprecated
Q
QI JUN 已提交
776
template <typename DeviceContext, typename T, typename functor,
F
fengjiayi 已提交
777
          typename broadcastfunctor, typename broadcast2functor>
778 779 780 781 782 783 784
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();
785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802

  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);
803
  trim_trailing_singular_dims(y_dims);
804
  axis = (y_dims.size() == 0) ? x_dims.size() : axis;
805 806

  int pre, n, post;
807
  get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post);
808 809 810 811 812 813 814 815 816 817 818

  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
819

820 821
template <typename Functor, typename DeviceContext, typename T,
          typename OutType = T>
D
dzhwinter 已提交
822

823 824 825 826
void ElementwiseComputeEx(const framework::ExecutionContext &ctx,
                          const framework::Tensor *x,
                          const framework::Tensor *y, int axis, Functor func,
                          framework::Tensor *z) {
827
  TransformFunctor<Functor, T, DeviceContext, OutType> functor(
C
chengduoZH 已提交
828
      x, y, z, ctx.template device_context<DeviceContext>(), func);
F
fengjiayi 已提交
829
  auto x_dims = x->dims();
830
  auto y_dims_untrimed = y->dims();
831 832 833 834 835 836 837
  PADDLE_ENFORCE_GE(
      x_dims.size(), y_dims_untrimed.size(),
      "ShapeError: the dimension of input X must greater than or equal to "
      "the one of input Y. But received: the shape of input X = [%s], the "
      "dimension of input X = %d, the shape of input Y = [%s], the dimension "
      "of of input Y = %d",
      x_dims, x_dims.size(), y_dims_untrimed, y_dims_untrimed.size());
838
  if (x_dims == y_dims_untrimed) {
F
fengjiayi 已提交
839 840 841 842
    functor.Run();
    return;
  }

843
  axis = (axis == -1 ? x_dims.size() - y_dims_untrimed.size() : axis);
F
fengjiayi 已提交
844 845
  PADDLE_ENFORCE(axis >= 0 && axis < x_dims.size(),
                 "Axis should be in range [0, x_dims)");
846
  auto y_dims = trim_trailing_singular_dims(y_dims_untrimed);
847
  axis = (y_dims.size() == 0) ? x_dims.size() : axis;
848 849 850 851 852 853
  int pre, n, post, mid_flag = 0;
  get_mid_dims(x_dims, y_dims, axis, &pre, &n, &post, &mid_flag);
  if (mid_flag) {
    functor.RunMidRowWise(n, pre, post);
    return;
  }
F
fengjiayi 已提交
854 855 856 857 858 859 860 861 862
  if (post == 1) {
    functor.RunRowWise(n, pre);
    return;
  } else {
    functor.RunMidWise(n, pre, post);
    return;
  }
}

863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171
// 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
1172 1173
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut>
1174 1175
struct FusedElemwiseAndActGradNoBroadcast {
  HOSTDEVICE void operator()(size_t i) {
1176 1177 1178 1179 1180 1181 1182
    T x_val = x_[i];
    T y_val = y_[i];
    T out_val = out_[i];
    T dout_val = dout_[i];
    T intermediate_out_val = UseIntermediateOut
                                 ? intermediate_out_[i]
                                 : dx_op_.GetIntermediateOut(x_val, y_val);
1183
    if (dx_ != nullptr) {
1184 1185
      dx_[i] = dx_op_.UseIntermediateOut(x_val, y_val, intermediate_out_val,
                                         out_val, dout_val);
1186 1187
    }
    if (dy_ != nullptr) {
1188 1189
      dy_[i] = dy_op_.UseIntermediateOut(x_val, y_val, intermediate_out_val,
                                         out_val, dout_val);
C
chengduo 已提交
1190 1191
    }
    if (dintermediate_ != nullptr) {
1192 1193
      dintermediate_[i] = dintermediate_op_.UseIntermediateOut(
          x_val, intermediate_out_val, out_val, dout_val);
1194 1195 1196 1197 1198 1199 1200 1201 1202 1203
    }
  }

  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
1204
  DIntermediate_OP dintermediate_op_;
1205 1206
  T *dx_;
  T *dy_;
C
chengduo 已提交
1207
  T *dintermediate_;
1208 1209 1210
};

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1211
          typename DIntermediate_OP, bool UseIntermediateOut>
1212 1213 1214 1215 1216
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
1217 1218 1219
    framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1220 1221 1222 1223
  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
1224 1225
      FusedElemwiseAndActGradNoBroadcast<T, DX_OP, DY_OP, DIntermediate_OP,
                                         UseIntermediateOut>{
1226 1227
          x->data<T>(), y->data<T>(),
          intermediate_out ? intermediate_out->data<T>() : nullptr,
C
chengduo 已提交
1228
          out->data<T>(), dout->data<T>(), dx_op, dy_op, dintermediate_op,
1229
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1230 1231 1232
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace())});
1233 1234
}

C
chengduo 已提交
1235 1236 1237 1238 1239 1240 1241
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) {
1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256
  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
1257 1258 1259 1260 1261
                    ? 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]);
1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274

        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
1275 1276 1277 1278 1279
                    ? 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]);
1280 1281 1282 1283 1284 1285 1286 1287 1288 1289
        if (BcastY) {
          if (i == 0) {
            dy[y_idx] = tmp;
          } else {
            dy[y_idx] += tmp;
          }
        } else {
          dy[y_idx] = tmp;
        }
      }
C
chengduo 已提交
1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306
      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;
          }
        }
      }
1307 1308 1309 1310
    }
  }
}

C
chengduo 已提交
1311 1312 1313 1314 1315 1316 1317
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) {
1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333
  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
1334 1335 1336 1337 1338
                      ? 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]);
1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351

          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
1352 1353 1354 1355 1356
                      ? 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]);
1357 1358 1359 1360 1361 1362 1363 1364 1365 1366
          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
1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383
        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;
            }
          }
        }
1384 1385 1386 1387 1388 1389
      }
    }
  }
}

#ifdef __NVCC__
C
chengduo 已提交
1390 1391 1392
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1393 1394
static __global__ void FusedElemwiseAndActGradBroadcast1CUDAKernel(
    const T *x, const T *y, const T *intermediate_out, const T *out,
C
chengduo 已提交
1395 1396
    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) {
1397 1398 1399
  int j = blockIdx.x;
  int i = threadIdx.x;
  int tid = threadIdx.x;
C
chengduo 已提交
1400
  T val(0), inter_val(0);
1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414
  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
1415 1416 1417 1418 1419 1420
      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]);
1421 1422 1423 1424 1425 1426 1427 1428

      if (BcastY) {
        dx[x_idx] = tmp;
      } else {
        val += tmp;
      }
    }
    if (dy != nullptr) {
C
chengduo 已提交
1429 1430 1431 1432 1433 1434
      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]);
1435 1436 1437 1438 1439 1440
      if (BcastY) {
        val += tmp;
      } else {
        dy[y_idx] = tmp;
      }
    }
C
chengduo 已提交
1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453
    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;
      }
    }
1454 1455 1456 1457

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

C
chengduo 已提交
1458
  h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473
  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
1474 1475 1476 1477 1478 1479 1480 1481
  if (!SameShapeOfIntermediateOutAndOut) {
    if (d_intermediate) {
      inter_val = paddle::platform::reduceSum(inter_val, tid, h);
      if (threadIdx.x == 0) {
        d_intermediate[j] = inter_val;
      }
    }
  }
1482 1483
}

C
chengduo 已提交
1484 1485 1486 1487 1488 1489 1490
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) {
1491 1492 1493
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, h);
  int gird_size = w;
  FusedElemwiseAndActGradBroadcast1CUDAKernel<
C
chengduo 已提交
1494
      T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut, BcastY,
1495
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
C
chengduo 已提交
1496 1497
      x, y, intermediate_out, out, dout, h, w, dx_op, dy_op, dintermediate_op,
      dx, dy, d_intermediate);
1498 1499
}

C
chengduo 已提交
1500 1501 1502
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1503 1504
static __global__ void FusedElemwiseAndActGradBroadcast2CUDAKernel(
    const T *x, const T *y, const T *intermediate_out, const T *out,
C
chengduo 已提交
1505 1506
    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) {
1507 1508 1509
  int tid = threadIdx.x;
  int j = blockIdx.x;

C
chengduo 已提交
1510
  T val(0), inter_val(0);
1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528
  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
1529 1530 1531 1532 1533 1534
      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]);
1535 1536 1537 1538 1539 1540 1541 1542

      if (BcastY) {
        dx[x_idx] = tmp;
      } else {
        val += tmp;
      }
    }
    if (dy != nullptr) {
C
chengduo 已提交
1543 1544 1545 1546 1547 1548
      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]);
1549 1550 1551 1552 1553 1554
      if (BcastY) {
        val += tmp;
      } else {
        dy[y_idx] = tmp;
      }
    }
C
chengduo 已提交
1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567
    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;
      }
    }
1568 1569 1570
    ttid += ELEMWISE_MAX_BLOCK_DIM;
  }

C
chengduo 已提交
1571 1572
  int h = pre * post;
  h = h > ELEMWISE_MAX_BLOCK_DIM ? ELEMWISE_MAX_BLOCK_DIM : h;
1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587
  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
1588 1589 1590 1591 1592 1593 1594 1595
  if (!SameShapeOfIntermediateOutAndOut) {
    if (d_intermediate) {
      inter_val = paddle::platform::reduceSum(inter_val, tid, h);
      if (threadIdx.x == 0) {
        d_intermediate[j] = inter_val;
      }
    }
  }
1596 1597
}

C
chengduo 已提交
1598 1599 1600
template <typename T, typename DX_OP, typename DY_OP, typename DIntermediate_OP,
          bool UseIntermediateOut, bool BcastY,
          bool SameShapeOfIntermediateOutAndOut>
1601 1602 1603
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
1604 1605
    DY_OP dy_op, DIntermediate_OP dintermediate_op, T *dx, T *dy,
    T *dintermediate) {
1606 1607 1608
  int block_size = std::min(ELEMWISE_MAX_BLOCK_DIM, pre * post);
  int gird_size = n;
  FusedElemwiseAndActGradBroadcast2CUDAKernel<
C
chengduo 已提交
1609
      T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut, BcastY,
1610
      SameShapeOfIntermediateOutAndOut><<<gird_size, block_size, 0, stream>>>(
C
chengduo 已提交
1611 1612
      x, y, intermediate_out, out, dout, pre, n, post, dx_op, dy_op,
      dintermediate_op, dx, dy, dintermediate);
1613 1614 1615 1616
}
#endif

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1617
          typename DIntermediate_OP, bool UseIntermediateOut, bool BcastY,
1618 1619 1620 1621 1622 1623
          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
1624 1625 1626
    framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637
  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
1638 1639
      FusedElemwiseAndActGradBroadcast1CUDA<T, DX_OP, DY_OP, DIntermediate_OP,
                                            UseIntermediateOut, BcastY,
1640 1641 1642 1643
                                            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
1644
          out->data<T>(), dout->data<T>(), h, w, dx_op, dy_op, dintermediate_op,
1645
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1646 1647 1648
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1649 1650
#endif
    } else {
C
chengduo 已提交
1651 1652
      FusedElemwiseAndActGradBroadcast1CPU<T, DX_OP, DY_OP, DIntermediate_OP,
                                           UseIntermediateOut, BcastY,
1653 1654 1655
                                           SameShapeOfIntermediateOutAndOut>(
          x->data<T>(), y->data<T>(),
          intermediate_out == nullptr ? nullptr : intermediate_out->data<T>(),
C
chengduo 已提交
1656
          out->data<T>(), dout->data<T>(), h, w, dx_op, dy_op, dintermediate_op,
1657
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1658 1659 1660
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1661 1662 1663 1664
    }
  } else {
    if (platform::is_gpu_place(ctx.GetPlace())) {
#ifdef __NVCC__
C
chengduo 已提交
1665 1666
      FusedElemwiseAndActGradBroadcast2CUDA<T, DX_OP, DY_OP, DIntermediate_OP,
                                            UseIntermediateOut, BcastY,
1667 1668 1669 1670 1671
                                            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
1672
          dintermediate_op,
1673
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1674 1675 1676
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1677 1678
#endif
    } else {
C
chengduo 已提交
1679 1680
      FusedElemwiseAndActGradBroadcast2CPU<T, DX_OP, DY_OP, DIntermediate_OP,
                                           UseIntermediateOut, BcastY,
1681 1682 1683 1684
                                           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
1685
          dintermediate_op,
1686
          dx == nullptr ? nullptr : dx->mutable_data<T>(ctx.GetPlace()),
C
chengduo 已提交
1687 1688 1689
          dy == nullptr ? nullptr : dy->mutable_data<T>(ctx.GetPlace()),
          dintermediate == nullptr ? nullptr : dintermediate->mutable_data<T>(
                                                   ctx.GetPlace()));
1690 1691 1692 1693 1694
    }
  }
}

template <typename DeviceContext, typename T, typename DX_OP, typename DY_OP,
C
chengduo 已提交
1695 1696
          typename DIntermediate_OP, bool UseIntermediateOut,
          bool SameShapeOfIntermediateOutAndOut>
1697 1698 1699 1700
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
1701 1702 1703
    int axis, framework::Tensor *dx, framework::Tensor *dy,
    framework::Tensor *dintermediate, DX_OP dx_op, DY_OP dy_op,
    DIntermediate_OP dintermediate_op) {
1704 1705 1706 1707 1708 1709
  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
1710 1711
    FusedElemwiseAndActGradComputeNoBroadcast<
        DeviceContext, T, DX_OP, DY_OP, DIntermediate_OP, UseIntermediateOut>(
1712
        ctx, x_dim, y_dim, x, y, intermediate_out, out, dout, axis, dx, dy,
C
chengduo 已提交
1713
        dintermediate, dx_op, dy_op, dintermediate_op);
1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728
  } 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
1729 1730 1731 1732
          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);
1733 1734
    } else {
      FusedElemwiseAndActGradComputeWithBroadcast<
C
chengduo 已提交
1735 1736 1737 1738
          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);
1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752
    }
  }
}

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
1753
                   "The save_intermediate_out is opened, "
1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779
                   "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,
1780 1781
      // for 'f2(y)', the shape of intermediate_out should be equal to the
      // shape
1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792
      // 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,
1793 1794
      // for 'f2(y)', the shape of intermediate_out should be equal to the
      // shape
1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806
      // 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);
    }
  }
}
1807 1808 1809 1810 1811 1812 1813 1814

template <typename DeviceContext, typename T>
static inline void GetDoubleGradSafeTensor(
    const framework::ExecutionContext &ctx, const framework::Tensor *x,
    const framework::Tensor *ddx, framework::Tensor *ddx_safe) {
  if (ddx) {
    *ddx_safe = *ddx;
  } else {
1815 1816
    auto &dev_ctx = ctx.template device_context<DeviceContext>();
    *ddx_safe = ctx.AllocateTmpTensor<T, DeviceContext>(x->dims(), dev_ctx);
1817 1818 1819 1820 1821 1822
    math::SetConstant<DeviceContext, T> set_zero;
    set_zero(ctx.template device_context<DeviceContext>(), ddx_safe,
             static_cast<T>(0));
  }
}

1823 1824
}  // namespace operators
}  // namespace paddle