hl_cuda_cudnn.cc 43.6 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Z
zhangjinchao01 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */

Y
Yu Yang 已提交
15
#include "hl_cuda_cudnn.h"
Z
zhangjinchao01 已提交
16
#include <cudnn.h>
L
liaogang 已提交
17
#include <gflags/gflags.h>
Z
zhangjinchao01 已提交
18 19 20
#include <mutex>
#include "hl_cuda_cudnn.ph"
#include "hl_dso_loader.h"
Y
Yu Yang 已提交
21 22
#include "hl_thread.ph"
#include "paddle/utils/Logging.h"
23

24 25 26 27
DEFINE_int32(cudnn_conv_workspace_limit_in_mb,
             4096,
             "Specify cuDNN max workspace limit, in units MB, "
             "4096MB=4GB by default.");
Z
zhangjinchao01 已提交
28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43

namespace dynload {

std::once_flag cudnn_dso_flag;
void* cudnn_dso_handle = nullptr;

/**
 * The following macro definition can generate structs
 * (for each function) to dynamic load cudbnn routine
 * via operator overloading: operator ()
 *
 * note: default dynamic linked libs
 **/

#ifdef PADDLE_USE_DSO

44 45 46 47 48 49 50 51 52
#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                                     \
  struct DynLoad__##__name {                                                \
    template <typename... Args>                                             \
    auto operator()(Args... args) -> decltype(__name(args...)) {            \
      using cudnn_func = decltype(__name(args...)) (*)(Args...);            \
      std::call_once(cudnn_dso_flag, GetCudnnDsoHandle, &cudnn_dso_handle); \
      void* p_##__name = dlsym(cudnn_dso_handle, #__name);                  \
      return reinterpret_cast<cudnn_func>(p_##__name)(args...);             \
    }                                                                       \
Z
zhangjinchao01 已提交
53 54 55 56
  } __name; /* struct DynLoad__##__name */

#else

57 58 59 60 61 62
#define DYNAMIC_LOAD_CUDNN_WRAP(__name)                          \
  struct DynLoad__##__name {                                     \
    template <typename... Args>                                  \
    auto operator()(Args... args) -> decltype(__name(args...)) { \
      return __name(args...);                                    \
    }                                                            \
Z
zhangjinchao01 已提交
63 64 65 66 67 68 69 70
  } __name; /* struct DynLoad__##__name */

#endif

/**
 * include all needed cudnn functions in HPPL
 * different cudnn version has different interfaces
 **/
71
// clang-format off
Z
zhangjinchao01 已提交
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98
#define CUDNN_DNN_ROUTINE_EACH(__macro)                   \
  __macro(cudnnSetTensor4dDescriptor)                     \
  __macro(cudnnSetTensor4dDescriptorEx)                   \
  __macro(cudnnGetConvolutionNdForwardOutputDim)          \
  __macro(cudnnGetConvolutionForwardAlgorithm)            \
  __macro(cudnnCreateTensorDescriptor)                    \
  __macro(cudnnDestroyTensorDescriptor)                   \
  __macro(cudnnCreateFilterDescriptor)                    \
  __macro(cudnnSetFilter4dDescriptor)                     \
  __macro(cudnnSetPooling2dDescriptor)                    \
  __macro(cudnnDestroyFilterDescriptor)                   \
  __macro(cudnnCreateConvolutionDescriptor)               \
  __macro(cudnnCreatePoolingDescriptor)                   \
  __macro(cudnnDestroyPoolingDescriptor)                  \
  __macro(cudnnSetConvolution2dDescriptor)                \
  __macro(cudnnDestroyConvolutionDescriptor)              \
  __macro(cudnnCreate)                                    \
  __macro(cudnnDestroy)                                   \
  __macro(cudnnSetStream)                                 \
  __macro(cudnnActivationForward)                         \
  __macro(cudnnConvolutionForward)                        \
  __macro(cudnnConvolutionBackwardBias)                   \
  __macro(cudnnGetConvolutionForwardWorkspaceSize)        \
  __macro(cudnnTransformTensor)                           \
  __macro(cudnnPoolingForward)                            \
  __macro(cudnnPoolingBackward)                           \
  __macro(cudnnSoftmaxBackward)                           \
99 100 101
  __macro(cudnnSoftmaxForward)                            \
  __macro(cudnnGetVersion)                                \
  __macro(cudnnGetErrorString)
Z
zhangjinchao01 已提交
102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
CUDNN_DNN_ROUTINE_EACH(DYNAMIC_LOAD_CUDNN_WRAP)

#define CUDNN_DNN_ROUTINE_EACH_R2(__macro)                \
  __macro(cudnnAddTensor)                                 \
  __macro(cudnnConvolutionBackwardData)                   \
  __macro(cudnnConvolutionBackwardFilter)
CUDNN_DNN_ROUTINE_EACH_R2(DYNAMIC_LOAD_CUDNN_WRAP)

// APIs available after R3:
#if CUDNN_VERSION >= 3000
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R3(__macro)              \
  __macro(cudnnGetConvolutionBackwardFilterWorkspaceSize)     \
  __macro(cudnnGetConvolutionBackwardDataAlgorithm)           \
  __macro(cudnnGetConvolutionBackwardFilterAlgorithm)         \
  __macro(cudnnGetConvolutionBackwardDataWorkspaceSize)
CUDNN_DNN_ROUTINE_EACH_AFTER_R3(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R3
#endif


// APIs available after R4:
123
#if CUDNN_VERSION >= 4007
Z
zhangjinchao01 已提交
124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143
#define CUDNN_DNN_ROUTINE_EACH_AFTER_R4(__macro)             \
  __macro(cudnnBatchNormalizationForwardTraining)            \
  __macro(cudnnBatchNormalizationForwardInference)           \
  __macro(cudnnBatchNormalizationBackward)
CUDNN_DNN_ROUTINE_EACH_AFTER_R4(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_AFTER_R4
#endif

// APIs in R5
#if CUDNN_VERSION >= 5000
#define CUDNN_DNN_ROUTINE_EACH_R5(__macro)                    \
  __macro(cudnnCreateActivationDescriptor)                    \
  __macro(cudnnSetActivationDescriptor)                       \
  __macro(cudnnGetActivationDescriptor)                       \
  __macro(cudnnDestroyActivationDescriptor)
CUDNN_DNN_ROUTINE_EACH_R5(DYNAMIC_LOAD_CUDNN_WRAP)
#undef CUDNN_DNN_ROUTINE_EACH_R5
#endif

#undef CUDNN_DNN_ROUTINE_EACH
144
// clang-format on
Z
zhangjinchao01 已提交
145 146 147
} /* namespace dynload */

/**
148
 * Check build-in cudnn function using glog and it **does not**
Z
zhangjinchao01 已提交
149 150
 * support << operator for more details error info.
 */
151 152 153 154 155
#define CHECK_CUDNN(cudnnFunc)                                         \
  do {                                                                 \
    cudnnStatus_t cudnnStat = cudnnFunc;                               \
    CHECK_EQ(CUDNN_STATUS_SUCCESS, cudnnStat)                          \
        << "Cudnn Error: " << dynload::cudnnGetErrorString(cudnnStat); \
156
  } while (0)
Z
zhangjinchao01 已提交
157 158 159 160

bool g_is_libcudnn_init = false;
int g_cudnn_lib_version = 0;

161 162
void hl_cudnn_desc_init(cudnnTensorDescriptor_t* cudnn_desc) {
  CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(cudnn_desc));
Z
zhangjinchao01 已提交
163 164
}

165 166 167 168 169 170 171 172 173 174 175 176 177
void hl_cudnn_init(cudnnHandle_t* cudnn_handle, cudaStream_t stream) {
  size_t cudnn_dso_ver = dynload::cudnnGetVersion();
  size_t cudnn_dso_major = cudnn_dso_ver / 1000;
  size_t cudnn_cuh_major = CUDNN_VERSION / 1000;

  // Compare cudnn header version with that of cudnn.so.
  CHECK((cudnn_cuh_major < 4 && cudnn_dso_major < 4) ||
        (cudnn_cuh_major == cudnn_dso_major))
      << "[cudnn init] libcudnn v" << cudnn_dso_major << " with header v"
      << cudnn_cuh_major << " unmatched!\n"
      << "PaddlePaddle Requirement: "
      << "(header v[2-3] with libcudnn v[2-3]) Or "
      << "(header v4 with libcudnn v4) Or "
178 179
      << "(header v5 with libcudnn v5) Or"
      << "(header v6 with libcudnn v6).";
180

181
  CHECK(!(CUDNN_VERSION < 6000 && CUDNN_VERSION >= 5000 && CUDA_VERSION < 7050))
182 183
      << "cudnn v5 requires cuda version >= 7.5";

184 185 186
  CHECK(!(CUDNN_VERSION >= 6000 && CUDA_VERSION < 8000))
      << "cudnn v6 requires cuda version >= 8.0";

187 188 189 190 191
  CHECK_CUDNN(dynload::cudnnCreate(cudnn_handle));
  CHECK_CUDNN(dynload::cudnnSetStream(*cudnn_handle, stream));

  g_is_libcudnn_init = true;
  g_cudnn_lib_version = cudnn_dso_ver;
Z
zhangjinchao01 已提交
192 193
}

194
int hl_get_cudnn_lib_version() { return g_cudnn_lib_version; }
Z
zhangjinchao01 已提交
195 196 197 198 199 200 201 202 203 204 205 206 207

void hl_conv_workspace(hl_tensor_descriptor input,
                       hl_tensor_descriptor output,
                       hl_filter_descriptor filter,
                       hl_convolution_descriptor conv,
                       int* convFwdAlgo,
                       size_t* fwdLimitBytes,
                       int* convBwdDataAlgo,
                       size_t* bwdDataLimitBytes,
                       int* convBwdFilterAlgo,
                       size_t* bwdFilterLimitBytes) {
#if CUDNN_VERSION >= 4000

208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292
  CHECK_NOTNULL(input);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(filter);
  CHECK_NOTNULL(conv);

  // Specify workspace limit directly
  size_t memoryLimitBytes =
      (1LL << 20) * FLAGS_cudnn_conv_workspace_limit_in_mb;

  // cudnn convolution forward configuration
  cudnnTensorDescriptor_t fwd_src_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnTensorDescriptor_t fwd_dest_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnFilterDescriptor_t fwd_filter_desc = GET_FILTER_DESCRIPTOR(filter);
  cudnnConvolutionDescriptor_t fwd_conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);

  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardAlgorithm(
      t_resource.cudnn_handle,
      fwd_src_desc,
      fwd_filter_desc,
      fwd_conv_desc,
      fwd_dest_desc,
      CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
      memoryLimitBytes,
      reinterpret_cast<cudnnConvolutionFwdAlgo_t*>(convFwdAlgo)));

  CHECK_CUDNN(dynload::cudnnGetConvolutionForwardWorkspaceSize(
      t_resource.cudnn_handle,
      fwd_src_desc,
      fwd_filter_desc,
      fwd_conv_desc,
      fwd_dest_desc,
      static_cast<cudnnConvolutionFwdAlgo_t>(*convFwdAlgo),
      fwdLimitBytes));

  // cudnn convolution backward data configuration
  cudnnFilterDescriptor_t bwd_data_filter_desc = GET_FILTER_DESCRIPTOR(filter);
  cudnnTensorDescriptor_t bwd_data_diff_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnTensorDescriptor_t bwd_data_grad_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnConvolutionDescriptor_t bwd_data_conv_desc =
      GET_CONVOLUTION_DESCRIPTOR(conv);

  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataAlgorithm(
      t_resource.cudnn_handle,
      bwd_data_filter_desc,
      bwd_data_diff_desc,
      bwd_data_conv_desc,
      bwd_data_grad_desc,
      CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
      memoryLimitBytes,
      reinterpret_cast<cudnnConvolutionBwdDataAlgo_t*>(convBwdDataAlgo)));

  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
      t_resource.cudnn_handle,
      bwd_data_filter_desc,
      bwd_data_diff_desc,
      bwd_data_conv_desc,
      bwd_data_grad_desc,
      static_cast<cudnnConvolutionBwdDataAlgo_t>(*convBwdDataAlgo),
      bwdDataLimitBytes));

  // cudnn convolution backward filter configuration
  cudnnTensorDescriptor_t bwd_filter_src_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnTensorDescriptor_t bwd_filter_diff_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnConvolutionDescriptor_t bwd_filter_conv_desc =
      GET_CONVOLUTION_DESCRIPTOR(conv);
  cudnnFilterDescriptor_t bwd_filter_grad_desc = GET_FILTER_DESCRIPTOR(filter);

  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
      t_resource.cudnn_handle,
      bwd_filter_src_desc,
      bwd_filter_diff_desc,
      bwd_filter_conv_desc,
      bwd_filter_grad_desc,
      CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
      memoryLimitBytes,
      reinterpret_cast<cudnnConvolutionBwdFilterAlgo_t*>(convBwdFilterAlgo)));

  CHECK_CUDNN(dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
      t_resource.cudnn_handle,
      bwd_filter_src_desc,
      bwd_filter_diff_desc,
      bwd_filter_conv_desc,
      bwd_filter_grad_desc,
      static_cast<cudnnConvolutionBwdFilterAlgo_t>(*convBwdFilterAlgo),
      bwdFilterLimitBytes));
Z
zhangjinchao01 已提交
293 294 295 296 297 298 299 300

#endif
}

void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc,
                                 int batch_size,
                                 int feature_maps,
                                 int height,
301
                                 int width) {
302
  CHECK_NOTNULL(image_desc);
Z
zhangjinchao01 已提交
303

304 305 306
  cudnn_tensor_descriptor hl_desc =
      (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor));
  CHECK_NOTNULL(hl_desc);
Z
zhangjinchao01 已提交
307

308
#ifndef PADDLE_TYPE_DOUBLE
309
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
Z
zhangjinchao01 已提交
310
#else
311
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
Z
zhangjinchao01 已提交
312
#endif
313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330
  CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc));

  CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(hl_desc->desc,
                                                  CUDNN_TENSOR_NCHW,
                                                  data_type,
                                                  batch_size,
                                                  feature_maps,
                                                  height,
                                                  width));

  hl_desc->format = CUDNN_TENSOR_NCHW;
  hl_desc->data_type = data_type;
  hl_desc->batch_size = batch_size;
  hl_desc->feature_maps = feature_maps;
  hl_desc->height = height;
  hl_desc->width = width;

  *image_desc = (hl_tensor_descriptor)hl_desc;
Z
zhangjinchao01 已提交
331 332 333
}

void hl_create_tensor_descriptor(hl_tensor_descriptor* image_desc) {
334
  CHECK_NOTNULL(image_desc);
Z
zhangjinchao01 已提交
335

336 337 338
  cudnn_tensor_descriptor hl_desc =
      (cudnn_tensor_descriptor)malloc(sizeof(_cudnn_tensor_descriptor));
  CHECK_NOTNULL(hl_desc);
Z
zhangjinchao01 已提交
339

340
#ifndef PADDLE_TYPE_DOUBLE
341
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
Z
zhangjinchao01 已提交
342
#else
343
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
Z
zhangjinchao01 已提交
344
#endif
345
  CHECK_CUDNN(dynload::cudnnCreateTensorDescriptor(&hl_desc->desc));
Z
zhangjinchao01 已提交
346

347
  hl_desc->data_type = data_type;
Z
zhangjinchao01 已提交
348

349
  *image_desc = (hl_tensor_descriptor)hl_desc;
Z
zhangjinchao01 已提交
350 351 352 353 354 355
}

void hl_tensor_reshape(hl_tensor_descriptor image_desc,
                       int batch_size,
                       int feature_maps,
                       int height,
356
                       int width) {
357 358 359 360 361 362 363 364 365 366 367 368 369
  const int stride_w = 1;
  const int stride_h = width * stride_w;
  const int stride_c = height * stride_h;
  const int stride_n = feature_maps * stride_c;
  return hl_tensor_reshape(image_desc,
                           batch_size,
                           feature_maps,
                           height,
                           width,
                           stride_n,
                           stride_c,
                           stride_h,
                           stride_w);
Z
zhangjinchao01 已提交
370 371 372 373 374 375 376 377 378 379
}

void hl_tensor_reshape(hl_tensor_descriptor image_desc,
                       int batch_size,
                       int feature_maps,
                       int height,
                       int width,
                       int nStride,
                       int cStride,
                       int hStride,
380
                       int wStride) {
381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400
  CHECK_NOTNULL(image_desc);

  cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc;
  CHECK_NOTNULL(hl_desc->desc);

  CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptorEx(hl_desc->desc,
                                                    hl_desc->data_type,
                                                    batch_size,
                                                    feature_maps,
                                                    height,
                                                    width,
                                                    nStride,
                                                    cStride,
                                                    hStride,
                                                    wStride));

  hl_desc->batch_size = batch_size;
  hl_desc->feature_maps = feature_maps;
  hl_desc->height = height;
  hl_desc->width = width;
Z
zhangjinchao01 已提交
401 402
}

403
void hl_destroy_tensor_descriptor(hl_tensor_descriptor image_desc) {
404
  CHECK_NOTNULL(image_desc);
Z
zhangjinchao01 已提交
405

406 407
  cudnn_tensor_descriptor hl_desc = (cudnn_tensor_descriptor)image_desc;
  CHECK_NOTNULL(hl_desc->desc);
Z
zhangjinchao01 已提交
408

409
  CHECK_CUDNN(dynload::cudnnDestroyTensorDescriptor(hl_desc->desc));
Z
zhangjinchao01 已提交
410

411
  hl_desc->desc = NULL;
Z
zhangjinchao01 已提交
412

413
  free(image_desc);
Z
zhangjinchao01 已提交
414 415 416 417 418 419 420 421 422
}

void hl_create_pooling_descriptor(hl_pooling_descriptor* pooling_desc,
                                  hl_pooling_mode_t mode,
                                  int height,
                                  int width,
                                  int height_padding,
                                  int width_padding,
                                  int stride_height,
423
                                  int stride_width) {
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
  cudnnPoolingMode_t cudnn_mode;
  switch (mode) {
    case HL_POOLING_MAX:
      cudnn_mode = CUDNN_POOLING_MAX;
      break;
    case HL_POOLING_AVERAGE:
      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_INCLUDE_PADDING;
      break;
    case HL_POOLING_AVERAGE_EXCLUDE_PADDING:
      cudnn_mode = CUDNN_POOLING_AVERAGE_COUNT_EXCLUDE_PADDING;
      break;
    default:
      LOG(FATAL) << "parameter mode error";
  }

  CHECK_NOTNULL(pooling_desc);

  cudnn_pooling_descriptor hl_pooling_desc =
      (cudnn_pooling_descriptor)malloc(sizeof(_cudnn_pooling_descriptor));
  CHECK_NOTNULL(hl_pooling_desc);

  CHECK_CUDNN(dynload::cudnnCreatePoolingDescriptor(&hl_pooling_desc->desc));

  CHECK_CUDNN(dynload::cudnnSetPooling2dDescriptor(hl_pooling_desc->desc,
                                                   cudnn_mode,
Z
zhangjinchao01 已提交
449
#if CUDNN_VERSION >= 5000
450
                                                   CUDNN_PROPAGATE_NAN,
Z
zhangjinchao01 已提交
451
#endif
452 453 454 455 456 457 458 459 460 461 462 463 464 465
                                                   height,
                                                   width,
                                                   height_padding,
                                                   width_padding,
                                                   stride_height,
                                                   stride_width));

  hl_pooling_desc->mode = cudnn_mode;
  hl_pooling_desc->window_height = height;
  hl_pooling_desc->window_width = width;
  hl_pooling_desc->stride_height = stride_height;
  hl_pooling_desc->stride_width = stride_width;

  *pooling_desc = (hl_pooling_descriptor)hl_pooling_desc;
Z
zhangjinchao01 已提交
466 467
}

468
void hl_destroy_pooling_descriptor(hl_pooling_descriptor pooling_desc) {
469
  CHECK_NOTNULL(pooling_desc);
Z
zhangjinchao01 已提交
470

471
  cudnn_pooling_descriptor hl_pooling = (cudnn_pooling_descriptor)pooling_desc;
Z
zhangjinchao01 已提交
472

473 474
  CHECK_NOTNULL(hl_pooling->desc);
  CHECK_CUDNN(dynload::cudnnDestroyPoolingDescriptor(hl_pooling->desc));
Z
zhangjinchao01 已提交
475

476
  hl_pooling->desc = NULL;
Z
zhangjinchao01 已提交
477

478
  free(pooling_desc);
Z
zhangjinchao01 已提交
479 480 481 482 483 484
}

void hl_pooling_forward(hl_tensor_descriptor input,
                        real* input_image,
                        hl_tensor_descriptor output,
                        real* output_image,
485
                        hl_pooling_descriptor pooling) {
486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
  cudnnPoolingDescriptor_t pooling_desc;
  cudnnTensorDescriptor_t input_desc;
  cudnnTensorDescriptor_t output_desc;

  CHECK_NOTNULL(input);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(pooling);
  CHECK_NOTNULL(input_image);
  CHECK_NOTNULL(output_image);

  real alpha = 1.0f;
  real beta = 1.0f;
  input_desc = ((cudnn_tensor_descriptor)input)->desc;
  output_desc = ((cudnn_tensor_descriptor)output)->desc;
  pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc;
  CHECK_CUDNN(dynload::cudnnPoolingForward(t_resource.cudnn_handle,
                                           pooling_desc,
                                           &alpha,
                                           input_desc,
                                           input_image,
                                           &beta,
                                           output_desc,
                                           output_image));
  CHECK_SYNC("hl_pooling_forward failed");
Z
zhangjinchao01 已提交
510 511 512 513 514 515 516 517
}

void hl_pooling_backward(hl_tensor_descriptor input,
                         real* input_image,
                         real* input_image_grad,
                         hl_tensor_descriptor output,
                         real* output_image,
                         real* output_image_grad,
518
                         hl_pooling_descriptor pooling) {
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
  cudnnPoolingDescriptor_t pooling_desc;
  cudnnTensorDescriptor_t input_desc;
  cudnnTensorDescriptor_t output_desc;

  CHECK_NOTNULL(input);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(pooling);
  CHECK_NOTNULL(input_image);
  CHECK_NOTNULL(input_image_grad);
  CHECK_NOTNULL(output_image);
  CHECK_NOTNULL(output_image_grad);

  real alpha = 1.0f;
  real beta = 1.0f;
  input_desc = ((cudnn_tensor_descriptor)input)->desc;
  output_desc = ((cudnn_tensor_descriptor)output)->desc;
  pooling_desc = ((cudnn_pooling_descriptor)pooling)->desc;
  CHECK_CUDNN(dynload::cudnnPoolingBackward(t_resource.cudnn_handle,
                                            pooling_desc,
                                            &alpha,
                                            output_desc,
                                            output_image,
                                            output_desc,
                                            output_image_grad,
                                            input_desc,
                                            input_image,
                                            &beta,
                                            input_desc,
                                            input_image_grad));
Z
zhangjinchao01 已提交
548 549 550 551 552 553 554
  CHECK_SYNC("hl_pooling_backward failed");
}

void hl_create_filter_descriptor(hl_filter_descriptor* filter,
                                 int input_feature_maps,
                                 int output_feature_maps,
                                 int height,
555
                                 int width) {
556
  CHECK_NOTNULL(filter);
Z
zhangjinchao01 已提交
557

558 559 560
  cudnn_filter_descriptor hl_filter =
      (cudnn_filter_descriptor)malloc(sizeof(_cudnn_filter_descriptor));
  CHECK_NOTNULL(hl_filter);
Z
zhangjinchao01 已提交
561

562
  CHECK_CUDNN(dynload::cudnnCreateFilterDescriptor(&hl_filter->desc));
Z
zhangjinchao01 已提交
563

564
#ifndef PADDLE_TYPE_DOUBLE
565
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
Z
zhangjinchao01 已提交
566
#else
567
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
Z
zhangjinchao01 已提交
568
#endif
569 570
  CHECK_CUDNN(dynload::cudnnSetFilter4dDescriptor(hl_filter->desc,
                                                  data_type,
Z
zhangjinchao01 已提交
571
#if CUDNN_VERSION >= 5000
572
                                                  CUDNN_TENSOR_NCHW,
Z
zhangjinchao01 已提交
573
#endif
574 575 576 577 578 579 580 581 582 583 584 585
                                                  output_feature_maps,
                                                  input_feature_maps,
                                                  height,
                                                  width));

  hl_filter->data_type = data_type;
  hl_filter->output_feature_maps = output_feature_maps;
  hl_filter->input_feature_maps = input_feature_maps;
  hl_filter->filter_height = height;
  hl_filter->filter_width = width;

  *filter = (hl_filter_descriptor)hl_filter;
Z
zhangjinchao01 已提交
586 587
}

588
void hl_destroy_filter_descriptor(hl_filter_descriptor filter) {
589
  CHECK_NOTNULL(filter);
Z
zhangjinchao01 已提交
590

591 592
  cudnn_filter_descriptor hl_filter = (cudnn_filter_descriptor)filter;
  CHECK_NOTNULL(hl_filter->desc);
Z
zhangjinchao01 已提交
593

594
  CHECK_CUDNN(dynload::cudnnDestroyFilterDescriptor(hl_filter->desc));
Z
zhangjinchao01 已提交
595

596
  hl_filter->desc = NULL;
Z
zhangjinchao01 已提交
597

598
  free(filter);
Z
zhangjinchao01 已提交
599 600 601 602 603 604 605 606
}

void hl_create_convolution_descriptor(hl_convolution_descriptor* conv,
                                      hl_tensor_descriptor image,
                                      hl_filter_descriptor filter,
                                      int padding_height,
                                      int padding_width,
                                      int stride_height,
607
                                      int stride_width) {
608 609 610 611 612 613 614 615 616
  CHECK_NOTNULL(conv);

  cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)malloc(
      sizeof(_cudnn_convolution_descriptor));

  CHECK_NOTNULL(hl_conv);
  CHECK_CUDNN(dynload::cudnnCreateConvolutionDescriptor(&hl_conv->desc));

  cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION;
617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633

#if CUDNN_VERSION >= 6000
#ifndef PADDLE_TYPE_DOUBLE
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
#else
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
#endif
  CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc,
                                                       padding_height,
                                                       padding_width,
                                                       stride_height,
                                                       stride_width,
                                                       1,
                                                       1,
                                                       mode,
                                                       data_type));
#else
634 635 636 637 638 639 640 641
  CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(hl_conv->desc,
                                                       padding_height,
                                                       padding_width,
                                                       stride_height,
                                                       stride_width,
                                                       1,
                                                       1,
                                                       mode));
642
#endif
643 644 645 646 647 648 649 650 651 652 653 654

  hl_conv->input_image = image;
  hl_conv->filter = filter;
  hl_conv->padding_height = padding_height;
  hl_conv->padding_width = padding_width;
  hl_conv->stride_height = stride_height;
  hl_conv->stride_width = stride_width;
  hl_conv->upscalex = 1;
  hl_conv->upscaley = 1;
  hl_conv->mode = mode;

  *conv = (hl_convolution_descriptor)hl_conv;
Z
zhangjinchao01 已提交
655 656 657 658 659 660 661 662
}

void hl_reset_convolution_descriptor(hl_convolution_descriptor conv,
                                     hl_tensor_descriptor image,
                                     hl_filter_descriptor filter,
                                     int padding_height,
                                     int padding_width,
                                     int stride_height,
663
                                     int stride_width) {
664 665 666 667 668 669
  CHECK_NOTNULL(conv);
  CHECK_NOTNULL(image);
  CHECK_NOTNULL(filter);

  cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
  cudnnConvolutionMode_t mode = CUDNN_CROSS_CORRELATION;
670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686

#if CUDNN_VERSION >= 6000
#ifndef PADDLE_TYPE_DOUBLE
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
#else
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
#endif
  CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc,
                                                       padding_height,
                                                       padding_width,
                                                       stride_height,
                                                       stride_width,
                                                       1,
                                                       1,
                                                       mode,
                                                       data_type));
#else
687 688 689 690 691 692 693 694
  CHECK_CUDNN(dynload::cudnnSetConvolution2dDescriptor(conv_desc,
                                                       padding_height,
                                                       padding_width,
                                                       stride_height,
                                                       stride_width,
                                                       1,
                                                       1,
                                                       mode));
695
#endif
696 697 698 699 700 701 702 703 704 705 706

  cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv;
  hl_conv->input_image = image;
  hl_conv->filter = filter;
  hl_conv->padding_height = padding_height;
  hl_conv->padding_width = padding_width;
  hl_conv->stride_height = stride_height;
  hl_conv->stride_width = stride_width;
  hl_conv->upscalex = 1;
  hl_conv->upscaley = 1;
  hl_conv->mode = mode;
Z
zhangjinchao01 已提交
707 708
}

709
void hl_destroy_convolution_descriptor(hl_convolution_descriptor conv) {
710
  CHECK_NOTNULL(conv);
Z
zhangjinchao01 已提交
711

712 713
  cudnn_convolution_descriptor hl_conv = (cudnn_convolution_descriptor)conv;
  CHECK_NOTNULL(hl_conv->desc);
Z
zhangjinchao01 已提交
714

715 716
  CHECK_CUDNN(dynload::cudnnDestroyConvolutionDescriptor(hl_conv->desc));
  hl_conv->desc = NULL;
Z
zhangjinchao01 已提交
717

718
  free(conv);
Z
zhangjinchao01 已提交
719 720 721 722 723 724 725 726 727 728 729 730
}

void hl_convolution_forward(hl_tensor_descriptor input,
                            real* input_data,
                            hl_tensor_descriptor output,
                            real* output_data,
                            hl_filter_descriptor filter,
                            real* filter_data,
                            hl_convolution_descriptor conv,
                            void* gpuWorkSpace,
                            size_t sizeInBytes,
                            int convFwdAlgo) {
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
  CHECK_NOTNULL(input);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(filter);
  CHECK_NOTNULL(conv);
  CHECK_NOTNULL(input_data);
  CHECK_NOTNULL(output_data);
  CHECK_NOTNULL(filter_data);
  cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnTensorDescriptor_t dest_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter);
  cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
  real alpha = 1.0f;
  real beta = 1.0f;
  CHECK_CUDNN(dynload::cudnnConvolutionForward(
      t_resource.cudnn_handle,
      &alpha,
      src_desc,
      input_data,
      filter_desc,
      filter_data,
      conv_desc,
      static_cast<cudnnConvolutionFwdAlgo_t>(convFwdAlgo),
      gpuWorkSpace,
      sizeInBytes,
      &beta,
      dest_desc,
      output_data));
Z
zhangjinchao01 已提交
758 759 760 761 762 763
  CHECK_SYNC("hl_convolution_forward failed");
}

void hl_convolution_forward_add_bias(hl_tensor_descriptor bias,
                                     real* bias_data,
                                     hl_tensor_descriptor output,
764
                                     real* output_data) {
765 766 767 768 769 770 771 772 773 774 775
  CHECK_NOTNULL(bias);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(bias_data);
  CHECK_NOTNULL(output_data);

  cudnnTensorDescriptor_t output_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias);
  real alpha = 1.0f;
  real beta = 1.0f;

  CHECK_CUDNN(dynload::cudnnAddTensor(t_resource.cudnn_handle,
Z
zhangjinchao01 已提交
776
#if CUDNN_VERSION < 4000
777
                                      CUDNN_ADD_SAME_C,
Z
zhangjinchao01 已提交
778
#endif
779 780 781 782 783 784
                                      &alpha,
                                      bias_desc,
                                      bias_data,
                                      &beta,
                                      output_desc,
                                      output_data));
Z
zhangjinchao01 已提交
785 786 787 788 789 790
  CHECK_SYNC("hl_convolution_forward_add_bias failed");
}

void hl_convolution_backward_bias(hl_tensor_descriptor bias,
                                  real* bias_grad_data,
                                  hl_tensor_descriptor output,
791
                                  real* output_grad_data) {
792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807
  CHECK_NOTNULL(bias);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(bias_grad_data);
  CHECK_NOTNULL(output_grad_data);

  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnTensorDescriptor_t bias_desc = GET_TENSOR_DESCRIPTOR(bias);
  CHECK_CUDNN(dynload::cudnnConvolutionBackwardBias(t_resource.cudnn_handle,
                                                    &alpha,
                                                    diff_desc,
                                                    output_grad_data,
                                                    &beta,
                                                    bias_desc,
                                                    bias_grad_data));
Z
zhangjinchao01 已提交
808 809 810 811 812 813 814 815 816 817 818 819 820
  CHECK_SYNC("hl_convolution_backward_bias failed");
}

void hl_convolution_backward_filter(hl_tensor_descriptor input,
                                    real* input_data,
                                    hl_tensor_descriptor output,
                                    real* output_grad_data,
                                    hl_filter_descriptor filter,
                                    real* filter_grad_data,
                                    hl_convolution_descriptor conv,
                                    void* gpuWorkSpace,
                                    size_t sizeInBytes,
                                    int convBwdFilterAlgo) {
821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843
  CHECK_NOTNULL(input);
  CHECK_NOTNULL(output);
  CHECK_NOTNULL(filter);
  CHECK_NOTNULL(conv);
  CHECK_NOTNULL(input_data);
  CHECK_NOTNULL(output_grad_data);
  CHECK_NOTNULL(filter_grad_data);

  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnTensorDescriptor_t src_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);
  cudnnFilterDescriptor_t grad_desc = GET_FILTER_DESCRIPTOR(filter);

  CHECK_CUDNN(dynload::cudnnConvolutionBackwardFilter(
      t_resource.cudnn_handle,
      &alpha,
      src_desc,
      input_data,
      diff_desc,
      output_grad_data,
      conv_desc,
Z
zhangjinchao01 已提交
844
#if CUDNN_VERSION >= 4000
845 846 847
      static_cast<cudnnConvolutionBwdFilterAlgo_t>(convBwdFilterAlgo),
      gpuWorkSpace,
      sizeInBytes,
Z
zhangjinchao01 已提交
848
#endif
849 850 851
      &beta,
      grad_desc,
      filter_grad_data));
Z
zhangjinchao01 已提交
852 853 854 855 856 857 858 859 860 861 862 863 864
  CHECK_SYNC("hl_convolution_backward_filter failed");
}

void hl_convolution_backward_data(hl_tensor_descriptor input,
                                  real* input_data_grad,
                                  hl_tensor_descriptor output,
                                  real* output_grad_data,
                                  hl_filter_descriptor filter,
                                  real* filter_data,
                                  hl_convolution_descriptor conv,
                                  void* gpuWorkSpace,
                                  size_t sizeInBytes,
                                  int convBwdDataAlgo) {
865 866 867 868 869 870 871 872 873 874 875 876 877 878 879
  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnFilterDescriptor_t filter_desc = GET_FILTER_DESCRIPTOR(filter);
  cudnnTensorDescriptor_t diff_desc = GET_TENSOR_DESCRIPTOR(output);
  cudnnTensorDescriptor_t grad_desc = GET_TENSOR_DESCRIPTOR(input);
  cudnnConvolutionDescriptor_t conv_desc = GET_CONVOLUTION_DESCRIPTOR(conv);

  CHECK_CUDNN(dynload::cudnnConvolutionBackwardData(
      t_resource.cudnn_handle,
      &alpha,
      filter_desc,
      filter_data,
      diff_desc,
      output_grad_data,
      conv_desc,
Z
zhangjinchao01 已提交
880
#if CUDNN_VERSION >= 4000
881 882 883
      static_cast<cudnnConvolutionBwdDataAlgo_t>(convBwdDataAlgo),
      gpuWorkSpace,
      sizeInBytes,
Z
zhangjinchao01 已提交
884
#endif
885 886 887
      &beta,
      grad_desc,
      input_data_grad));
Z
zhangjinchao01 已提交
888 889 890
  CHECK_SYNC("hl_convolution_backward_data failed");
}

891
void hl_softmax_forward(real* input, real* output, int height, int width) {
892
#ifndef PADDLE_TYPE_DOUBLE
893
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
Z
zhangjinchao01 已提交
894
#else
895
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
Z
zhangjinchao01 已提交
896
#endif
897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915
  CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc,
                                                  CUDNN_TENSOR_NCHW,
                                                  data_type,
                                                  height,
                                                  width,
                                                  1,
                                                  1));

  real alpha = 1.0f;
  real beta = 0.0f;
  CHECK_CUDNN(dynload::cudnnSoftmaxForward(t_resource.cudnn_handle,
                                           CUDNN_SOFTMAX_ACCURATE,
                                           CUDNN_SOFTMAX_MODE_CHANNEL,
                                           &alpha,
                                           t_resource.cudnn_desc,
                                           input,
                                           &beta,
                                           t_resource.cudnn_desc,
                                           output));
Z
zhangjinchao01 已提交
916 917 918
  CHECK_SYNC("hl_softmax_forward failed");
}

919 920
void hl_softmax_backward(real* output_value,
                         real* output_grad,
Z
zhangjinchao01 已提交
921
                         int height,
922
                         int width) {
923
#ifndef PADDLE_TYPE_DOUBLE
924
  cudnnDataType_t data_type = CUDNN_DATA_FLOAT;
Z
zhangjinchao01 已提交
925
#else
926
  cudnnDataType_t data_type = CUDNN_DATA_DOUBLE;
Z
zhangjinchao01 已提交
927
#endif
928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948
  CHECK_CUDNN(dynload::cudnnSetTensor4dDescriptor(t_resource.cudnn_desc,
                                                  CUDNN_TENSOR_NCHW,
                                                  data_type,
                                                  height,
                                                  width,
                                                  1,
                                                  1));

  real alpha = 1.0f;
  real beta = 0.0f;
  CHECK_CUDNN(dynload::cudnnSoftmaxBackward(t_resource.cudnn_handle,
                                            CUDNN_SOFTMAX_ACCURATE,
                                            CUDNN_SOFTMAX_MODE_CHANNEL,
                                            &alpha,
                                            t_resource.cudnn_desc,
                                            output_value,
                                            t_resource.cudnn_desc,
                                            output_grad,
                                            &beta,
                                            t_resource.cudnn_desc,
                                            output_grad));
Z
zhangjinchao01 已提交
949 950 951 952
  CHECK_SYNC("hl_softmax_backward failed");
}

void hl_batch_norm_forward_training(hl_tensor_descriptor inputDesc,
953
                                    real* input,
Z
zhangjinchao01 已提交
954
                                    hl_tensor_descriptor outputDesc,
955
                                    real* output,
Z
zhangjinchao01 已提交
956
                                    hl_tensor_descriptor bnParamDesc,
957 958
                                    real* scale,
                                    real* bias,
Z
zhangjinchao01 已提交
959
                                    double factor,
960 961
                                    real* runningMean,
                                    real* runningInvVar,
Z
zhangjinchao01 已提交
962
                                    double epsilon,
963 964
                                    real* savedMean,
                                    real* savedVar) {
965
#if CUDNN_VERSION >= 4007
Z
zhangjinchao01 已提交
966 967 968
  if ((NULL != runningMean && NULL == runningInvVar) ||
      (NULL == runningMean && NULL != runningInvVar)) {
    LOG(FATAL) << "runningMean and runningInvVar can be NULL "
969
               << "but only at the same time.";
Z
zhangjinchao01 已提交
970 971 972 973 974 975 976 977 978 979 980 981 982
  }
  if ((NULL != savedMean && NULL == savedVar) ||
      (NULL == savedMean && NULL != savedVar)) {
    LOG(FATAL) << "savedMean and savedVar can be NULL "
               << "but only at the same time.";
  }

  cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc);
  cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc);
  cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc);
  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000
  CHECK_CUDNN(
      dynload::cudnnBatchNormalizationForwardTraining(t_resource.cudnn_handle,
                                                      mode,
                                                      &alpha,
                                                      &beta,
                                                      xDesc,
                                                      input,
                                                      yDesc,
                                                      output,
                                                      bnDesc,
                                                      scale,
                                                      bias,
                                                      factor,
                                                      runningMean,
                                                      runningInvVar,
                                                      epsilon,
                                                      savedMean,
                                                      savedVar));
Z
zhangjinchao01 已提交
1001 1002 1003

  CHECK_SYNC("hl_batch_norm_forward_training failed");
#else
1004
  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
Z
zhangjinchao01 已提交
1005 1006 1007 1008 1009
             << "But cudnn lib version is " << g_cudnn_lib_version;
#endif
}

void hl_batch_norm_forward_inference(hl_tensor_descriptor inputDesc,
1010 1011 1012 1013 1014 1015 1016 1017 1018
                                     real* input,
                                     hl_tensor_descriptor outputDesc,
                                     real* output,
                                     hl_tensor_descriptor bnParamDesc,
                                     real* scale,
                                     real* bias,
                                     real* estimatedMean,
                                     real* estimatedInvVar,
                                     double epsilon) {
1019
#if CUDNN_VERSION >= 4007
Z
zhangjinchao01 已提交
1020 1021 1022 1023 1024 1025
  cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc);
  cudnnTensorDescriptor_t yDesc = GET_TENSOR_DESCRIPTOR(outputDesc);
  cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(bnParamDesc);
  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040
  CHECK_CUDNN(
      dynload::cudnnBatchNormalizationForwardInference(t_resource.cudnn_handle,
                                                       mode,
                                                       &alpha,
                                                       &beta,
                                                       xDesc,
                                                       input,
                                                       yDesc,
                                                       output,
                                                       bnDesc,
                                                       scale,
                                                       bias,
                                                       estimatedMean,
                                                       estimatedInvVar,
                                                       epsilon));
Z
zhangjinchao01 已提交
1041 1042 1043

  CHECK_SYNC("hl_batch_norm_forward_inference failed");
#else
1044
  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
Z
zhangjinchao01 已提交
1045 1046 1047 1048 1049
             << "But cudnn lib version is " << g_cudnn_lib_version;
#endif
}

void hl_batch_norm_backward(hl_tensor_descriptor inputDesc,
1050
                            real* input,
Z
zhangjinchao01 已提交
1051
                            hl_tensor_descriptor outGradDesc,
1052
                            real* outGrad,
Z
zhangjinchao01 已提交
1053
                            hl_tensor_descriptor inGradDesc,
1054
                            real* inGrad,
Z
zhangjinchao01 已提交
1055
                            hl_tensor_descriptor dBnParamDesc,
1056 1057 1058
                            real* scale,
                            real* scaleGrad,
                            real* biasGrad,
Z
zhangjinchao01 已提交
1059
                            double epsilon,
1060 1061
                            real* savedMean,
                            real* savedInvVar) {
1062
#if CUDNN_VERSION >= 4007
Z
zhangjinchao01 已提交
1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075
  if ((NULL != savedMean && NULL == savedInvVar) ||
      (NULL == savedMean && NULL != savedInvVar)) {
    LOG(FATAL) << "savedMean and savedVar can be NULL "
               << "but only at the same time.";
  }

  cudnnTensorDescriptor_t xDesc = GET_TENSOR_DESCRIPTOR(inputDesc);
  cudnnTensorDescriptor_t dyDesc = GET_TENSOR_DESCRIPTOR(outGradDesc);
  cudnnTensorDescriptor_t dxDesc = GET_TENSOR_DESCRIPTOR(inGradDesc);
  cudnnTensorDescriptor_t bnDesc = GET_TENSOR_DESCRIPTOR(dBnParamDesc);
  real alpha = 1.0f;
  real beta = 1.0f;
  cudnnBatchNormMode_t mode = CUDNN_BATCHNORM_SPATIAL;
1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094
  CHECK_CUDNN(dynload::cudnnBatchNormalizationBackward(t_resource.cudnn_handle,
                                                       mode,
                                                       &alpha,
                                                       &beta,
                                                       &alpha,
                                                       &beta,
                                                       xDesc,
                                                       input,
                                                       dyDesc,
                                                       outGrad,
                                                       dxDesc,
                                                       inGrad,
                                                       bnDesc,
                                                       scale,
                                                       scaleGrad,
                                                       biasGrad,
                                                       epsilon,
                                                       savedMean,
                                                       savedInvVar));
Z
zhangjinchao01 已提交
1095 1096 1097

  CHECK_SYNC("hl_batch_norm_backward failed");
#else
1098
  LOG(FATAL) << "CudnnBatchNorm requires cudnn version >= 4007. "
Z
zhangjinchao01 已提交
1099 1100 1101
             << "But cudnn lib version is " << g_cudnn_lib_version;
#endif
}