batch_norm_op.cu 37.1 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
Q
Qiao Longfei 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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

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

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

15
#include <algorithm>
Q
Qiao Longfei 已提交
16
#include <cfloat>
17 18 19
#include <string>
#include <vector>
#include "cub/cub.cuh"
S
Siddharth Goyal 已提交
20
#include "paddle/fluid/framework/data_layout.h"
21
#include "paddle/fluid/operators/batch_norm_op.h"
Y
Yi Wang 已提交
22 23
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/cudnn_helper.h"
K
Kexin Zhao 已提交
24
#include "paddle/fluid/platform/float16.h"
Q
Qiao Longfei 已提交
25

26
DECLARE_bool(cudnn_batchnorm_spatial_persistent);
W
Wu Yi 已提交
27

Q
Qiao Longfei 已提交
28 29 30 31
namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
Q
QI JUN 已提交
32
using DataLayout = framework::DataLayout;
Q
Qiao Longfei 已提交
33 34
template <typename T>
using CudnnDataType = platform::CudnnDataType<T>;
K
Kexin Zhao 已提交
35
template <typename T>
K
update  
Kexin Zhao 已提交
36
using BatchNormParamType = typename CudnnDataType<T>::BatchNormParamType;
Q
Qiao Longfei 已提交
37 38

template <typename T>
Q
QI JUN 已提交
39 40
class BatchNormKernel<platform::CUDADeviceContext, T>
    : public framework::OpKernel<T> {
Q
Qiao Longfei 已提交
41 42
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
K
Kaipeng Deng 已提交
43 44 45
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        platform::errors::InvalidArgument("It must use CUDAPlace."));
Q
Qiao Longfei 已提交
46
    double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
47
    float momentum = ctx.Attr<float>("momentum");
Q
Qiao Longfei 已提交
48
    const bool is_test = ctx.Attr<bool>("is_test");
49
    const bool use_global_stats = ctx.Attr<bool>("use_global_stats");
50
    const bool trainable_stats = ctx.Attr<bool>("trainable_statistics");
Q
QI JUN 已提交
51 52 53
    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
    const DataLayout data_layout =
        framework::StringToDataLayout(data_layout_str);
Q
Qiao Longfei 已提交
54

55 56
    bool test_mode = is_test && (!trainable_stats);

Q
Qiao Longfei 已提交
57 58 59 60
    // Get the size for each dimension.
    // NCHW [batch_size, in_channels, in_height, in_width]
    const auto *x = ctx.Input<Tensor>("X");
    const auto &x_dims = x->dims();
C
ceci3 已提交
61 62 63 64 65 66
    PADDLE_ENFORCE_EQ(
        x_dims.size() >= 2 && x_dims.size() <= 5, true,
        platform::errors::InvalidArgument(
            "The size of input's dimensions should be between 2 and 5"
            "But received: the size of input's dimensions is [%d]",
            x_dims.size()));
Q
Qiao Longfei 已提交
67

68 69 70
    auto *y = ctx.Output<Tensor>("Y");
    y->mutable_data<T>(ctx.GetPlace());

71 72 73 74 75
    int N, C, H, W, D;
    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);

    auto dtype = platform::CudnnDataType<T>::type;
    const bool fast_nhwc_batch_norm =
76
        test_mode ||
77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99
        (dtype == CUDNN_DATA_HALF && FLAGS_cudnn_batchnorm_spatial_persistent);

    auto compute_format =
        fast_nhwc_batch_norm && data_layout == DataLayout::kNHWC
            ? DataLayout::kNHWC
            : DataLayout::kNCHW;

    Tensor transformed_x(x->type());
    Tensor transformed_y(y->type());
    if (data_layout == DataLayout::kNHWC &&
        compute_format == DataLayout::kNCHW && x_dims.size() > 2) {
      VLOG(3) << "Transform input tensor from NHWC to NCHW.";
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, x,
                                                           &transformed_x);
      TransToChannelFirst<platform::CUDADeviceContext, T>(ctx, x,
                                                          &transformed_x);
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, y,
                                                           &transformed_y);
    } else {
      transformed_x.ShareDataWith(*x);
      transformed_y.ShareDataWith(*y);
    }

Q
Qiao Longfei 已提交
100 101 102 103 104
    // ------------------- cudnn descriptors ---------------------
    cudnnTensorDescriptor_t data_desc_;
    cudnnTensorDescriptor_t bn_param_desc_;
    cudnnBatchNormMode_t mode_;

105 106 107
    PADDLE_ENFORCE_CUDA_SUCCESS(
        platform::dynload::cudnnCreateTensorDescriptor(&data_desc_));
    PADDLE_ENFORCE_CUDA_SUCCESS(
Q
Qiao Longfei 已提交
108 109 110 111 112 113 114 115
        platform::dynload::cudnnCreateTensorDescriptor(&bn_param_desc_));

    if (epsilon <= CUDNN_BN_MIN_EPSILON - FLT_EPSILON) {
      LOG(ERROR) << "Provided epsilon is smaller than "
                 << "CUDNN_BN_MIN_EPSILON. Setting it to "
                 << "CUDNN_BN_MIN_EPSILON instead.";
    }
    epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON);
116
#if CUDNN_VERSION_MIN(7, 0, 0)
W
Wu Yi 已提交
117 118 119 120 121
    if (FLAGS_cudnn_batchnorm_spatial_persistent) {
      mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
    } else {
      mode_ = CUDNN_BATCHNORM_SPATIAL;
    }
122
#else
Q
Qiao Longfei 已提交
123
    mode_ = CUDNN_BATCHNORM_SPATIAL;
124
#endif
Q
Qiao Longfei 已提交
125

M
minqiyang 已提交
126
    VLOG(3) << "Setting descriptors.";
Q
Qiao Longfei 已提交
127 128
    std::vector<int> dims;
    std::vector<int> strides;
129
    if (compute_format == DataLayout::kNCHW) {
Q
Qiao Longfei 已提交
130 131 132 133 134 135
      dims = {N, C, H, W, D};
      strides = {C * H * W * D, H * W * D, W * D, D, 1};
    } else {
      dims = {N, C, H, W, D};
      strides = {H * W * D * C, 1, W * D * C, D * C, C};
    }
136
    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetTensorNdDescriptor(
Q
Qiao Longfei 已提交
137 138
        data_desc_, CudnnDataType<T>::type,
        x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data()));
K
Kexin Zhao 已提交
139
    // Note: PERSISTENT not implemented for inference
140 141 142
    PADDLE_ENFORCE_CUDA_SUCCESS(
        platform::dynload::cudnnDeriveBNTensorDescriptor(
            bn_param_desc_, data_desc_,
143
            test_mode ? CUDNN_BATCHNORM_SPATIAL : mode_));
Q
Qiao Longfei 已提交
144 145 146 147

    const auto *scale = ctx.Input<Tensor>("Scale");
    const auto *bias = ctx.Input<Tensor>("Bias");

Q
QI JUN 已提交
148
    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
Q
Qiao Longfei 已提交
149

Q
QI JUN 已提交
150
    auto handle = dev_ctx.cudnn_handle();
Q
Qiao Longfei 已提交
151 152

    // Now, depending on whether we are running test or not, we have two paths.
153
    if (test_mode || use_global_stats) {
Q
Qiao Longfei 已提交
154 155 156 157
      // only when test we use input to do computation.
      const auto *est_mean = ctx.Input<Tensor>("Mean");
      const auto *est_var = ctx.Input<Tensor>("Variance");
      // Run inference mode.
C
ceci3 已提交
158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185
      PADDLE_ENFORCE_EQ(
          est_mean->dims().size(), 1UL,
          platform::errors::InvalidArgument(
              "The size of mean's dimensions must equal to 1."
              "But received: the size of mean's dimensions mean is [%d],"
              "the dimensions of mean is [%s].",
              est_mean->dims().size(), est_mean->dims()));
      PADDLE_ENFORCE_EQ(
          est_var->dims().size(), 1UL,
          platform::errors::InvalidArgument(
              "The size of variance's dimensions must equal to 1."
              "But received: the size of variance's dimensions is [%d],"
              "the dimensions of variance is [%s].",
              est_var->dims().size(), est_var->dims()));
      PADDLE_ENFORCE_EQ(
          est_mean->dims()[0], C,
          platform::errors::InvalidArgument(
              "The first dimension of mean must equal to the number of "
              "Channels, which is [%d]. But received: the first dimension"
              "of mean is [%d], the dimensions of mean is [%s].",
              C, est_mean->dims()[0], est_mean->dims()));
      PADDLE_ENFORCE_EQ(
          est_var->dims()[0], C,
          platform::errors::InvalidArgument(
              "The first dimension of variance must equal to the number"
              "of Channels, which is [%d]. But received: the first dimension of"
              "variance is [%d], the dimensions of variance is [%s].",
              C, est_var->dims()[0], est_var->dims()));
Q
Qiao Longfei 已提交
186

187 188 189 190 191 192 193 194 195 196 197 198
      PADDLE_ENFORCE_CUDA_SUCCESS(
          platform::dynload::cudnnBatchNormalizationForwardInference(
              handle,
              // Note: PERSISTENT not implemented for inference
              CUDNN_BATCHNORM_SPATIAL, CudnnDataType<T>::kOne(),
              CudnnDataType<T>::kZero(), data_desc_,
              transformed_x.template data<T>(), data_desc_,
              transformed_y.template mutable_data<T>(ctx.GetPlace()),
              bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
              bias->template data<BatchNormParamType<T>>(),
              est_mean->template data<BatchNormParamType<T>>(),
              est_var->template data<BatchNormParamType<T>>(), epsilon));
Q
Qiao Longfei 已提交
199
    } else {
200 201 202 203 204 205 206 207 208
      // if MomentumTensor is set, use MomentumTensor value, momentum
      // is only used in this training branch
      if (ctx.HasInput("MomentumTensor")) {
        const auto *mom_tensor = ctx.Input<Tensor>("MomentumTensor");
        Tensor mom_cpu;
        TensorCopySync(*mom_tensor, platform::CPUPlace(), &mom_cpu);
        momentum = mom_cpu.data<float>()[0];
      }

Q
Qiao Longfei 已提交
209 210 211
      // Run training mode.
      // obtain running mean and running inv var, and see if we need to
      // initialize them.
D
Dang Qingqing 已提交
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226

      auto *mean_out = ctx.Output<Tensor>("MeanOut");
      auto *variance_out = ctx.Output<Tensor>("VarianceOut");
      mean_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
      variance_out->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());

      auto *saved_mean = ctx.Output<Tensor>("SavedMean");
      auto *saved_variance = ctx.Output<Tensor>("SavedVariance");
      saved_mean->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
      saved_variance->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
      math::SetConstant<platform::CUDADeviceContext, BatchNormParamType<T>>
          functor;
      functor(dev_ctx, saved_mean, static_cast<BatchNormParamType<T>>(0));
      functor(dev_ctx, saved_variance, static_cast<BatchNormParamType<T>>(0));

227
      if ((N * H * W * D) == 1) {
228 229
        // Only 1 element in normalization dimension,
        // skip the batch norm calculation, let y = x.
230
        framework::TensorCopy(*x, ctx.GetPlace(), y);
231 232 233
      } else {
        double this_factor = 1. - momentum;

234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252
        bool called = false;
#if CUDNN_VERSION_MIN(7, 4, 1)
        if (compute_format == DataLayout::kNHWC) {
          called = true;
          size_t workspace_size = 0;
          size_t reserve_space_size = 0;
          void *reserve_space_ptr = nullptr;
          void *workspace_ptr = nullptr;
          Tensor workspace_tensor;
          // Create reserve space and workspace for batch norm.
          // Create tensor for each batchnorm op, it will be used in the
          // backward. Thus this tensor shouldn't be temp.
          auto *reserve_space = ctx.Output<Tensor>("ReserveSpace");
          PADDLE_ENFORCE_NOT_NULL(
              reserve_space,
              platform::errors::NotFound(
                  "The argument ReserveSpace of batch_norm op is not found."));

          // --------------- cudnn batchnorm workspace ---------------
253
          PADDLE_ENFORCE_CUDA_SUCCESS(
254 255 256 257 258 259 260 261 262 263 264 265 266
              platform::dynload::
                  cudnnGetBatchNormalizationForwardTrainingExWorkspaceSize(
                      /*handle=*/handle,
                      /*mode=*/mode_,
                      /*bnIps=*/CUDNN_BATCHNORM_OPS_BN,
                      /*xDesc=*/data_desc_,
                      /*zDesc=*/nullptr,
                      /*yDesc=*/data_desc_,
                      /*bnScaleBiasMeanVarDesc=*/bn_param_desc_,
                      /*activationDesc=*/nullptr,
                      /*sizeInBytes=*/&workspace_size));

          // -------------- cudnn batchnorm reserve space --------------
267
          PADDLE_ENFORCE_CUDA_SUCCESS(
268 269 270 271 272 273 274 275 276 277 278 279 280
              platform::dynload::
                  cudnnGetBatchNormalizationTrainingExReserveSpaceSize(
                      /*handle=*/handle,
                      /*mode=*/mode_,
                      /*bnOps=*/CUDNN_BATCHNORM_OPS_BN,
                      /*activationDesc=*/nullptr,
                      /*xDesc=*/data_desc_,
                      /*sizeInBytes=*/&reserve_space_size));

          reserve_space_ptr = reserve_space->mutable_data(
              ctx.GetPlace(), transformed_x.type(), reserve_space_size);
          workspace_ptr = workspace_tensor.mutable_data(
              ctx.GetPlace(), transformed_x.type(), workspace_size);
281
          PADDLE_ENFORCE_CUDA_SUCCESS(
282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
              platform::dynload::cudnnBatchNormalizationForwardTrainingEx(
                  handle, mode_, CUDNN_BATCHNORM_OPS_BN,
                  CudnnDataType<T>::kOne(), CudnnDataType<T>::kZero(),
                  data_desc_, transformed_x.template data<T>(), nullptr,
                  nullptr, data_desc_, transformed_y.template data<T>(),
                  bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
                  bias->template data<BatchNormParamType<T>>(), this_factor,
                  mean_out->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  variance_out->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  epsilon,
                  saved_mean->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  saved_variance->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  nullptr, workspace_ptr, workspace_size, reserve_space_ptr,
                  reserve_space_size));
        }
#endif
        if (!called) {
303
          PADDLE_ENFORCE_CUDA_SUCCESS(
304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320
              platform::dynload::cudnnBatchNormalizationForwardTraining(
                  handle, mode_, CudnnDataType<T>::kOne(),
                  CudnnDataType<T>::kZero(), data_desc_,
                  transformed_x.template data<T>(), data_desc_,
                  transformed_y.template mutable_data<T>(ctx.GetPlace()),
                  bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
                  bias->template data<BatchNormParamType<T>>(), this_factor,
                  mean_out->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  variance_out->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  epsilon,
                  saved_mean->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  saved_variance->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace())));
        }
321
      }
Q
Qiao Longfei 已提交
322 323
    }

324 325 326 327 328 329
    if (data_layout == DataLayout::kNHWC &&
        compute_format == DataLayout::kNCHW && x_dims.size() > 2) {
      VLOG(3) << "Transform batchnorm output from NCHW to NHWC";
      TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
          ctx, &transformed_y, y);
    }
Q
Qiao Longfei 已提交
330
    // clean when exit.
331 332 333
    PADDLE_ENFORCE_CUDA_SUCCESS(
        platform::dynload::cudnnDestroyTensorDescriptor(data_desc_));
    PADDLE_ENFORCE_CUDA_SUCCESS(
Q
Qiao Longfei 已提交
334 335 336 337
        platform::dynload::cudnnDestroyTensorDescriptor(bn_param_desc_));
  }
};

338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373
template <typename T, int BlockDim, framework::DataLayout layout>
static __global__ void KeBNBackwardScaleBias(
    const T *dy, const T *x, const BatchNormParamType<T> *mean,
    const BatchNormParamType<T> *variance, const double epsilon, const int N,
    const int C, const int HxW, BatchNormParamType<T> *dscale,
    BatchNormParamType<T> *dbias) {
  const int outer_size = C;
  const int inner_size = N * HxW;
  typedef cub::BlockReduce<BatchNormParamType<T>, BlockDim> BlockReduce;
  __shared__ typename BlockReduce::TempStorage ds_storage;
  __shared__ typename BlockReduce::TempStorage db_storage;

  for (int i = blockIdx.x; i < outer_size; i += gridDim.x) {
    BatchNormParamType<T> ds_sum = static_cast<BatchNormParamType<T>>(0);
    BatchNormParamType<T> db_sum = static_cast<BatchNormParamType<T>>(0);

    BatchNormParamType<T> inv_var_i = 1.0 / sqrt(variance[i] + epsilon);
    BatchNormParamType<T> mean_i = mean[i];
    for (int j = threadIdx.x; j < inner_size; j += blockDim.x) {
      const int index = layout == framework::DataLayout::kNCHW
                            ? (j / HxW * C + i) * HxW + j % HxW
                            : j * outer_size + i;
      ds_sum += static_cast<BatchNormParamType<T>>(dy[index]) *
                (static_cast<BatchNormParamType<T>>(x[index]) - mean_i);
      db_sum += static_cast<BatchNormParamType<T>>(dy[index]);
    }
    ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum());
    db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum());
    if (threadIdx.x == 0) {
      dscale[i] = ds_sum * inv_var_i;
      dbias[i] = db_sum;
    }
    __syncthreads();
  }
}

Q
qingqing01 已提交
374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389
template <typename T, framework::DataLayout layout>
static __global__ void KeBNBackwardData(const T *dy,
                                        const BatchNormParamType<T> *scale,
                                        const BatchNormParamType<T> *variance,
                                        const double epsilon, const int C,
                                        const int HxW, const int num, T *dx) {
  int gid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (int i = gid; i < num; i += stride) {
    const int c = layout == framework::DataLayout::kNCHW ? i / HxW % C : i % C;
    BatchNormParamType<T> inv_var = 1.0 / sqrt(variance[c] + epsilon);
    dx[i] = static_cast<T>(static_cast<BatchNormParamType<T>>(dy[i]) *
                           scale[c] * inv_var);
  }
}

K
Kaipeng Deng 已提交
390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424
template <typename T>
static __global__ void KeBNRestoreData(const framework::DataLayout layout, T *x,
                                       const BatchNormParamType<T> *scale,
                                       const BatchNormParamType<T> *bias,
                                       const BatchNormParamType<T> *mean,
                                       const BatchNormParamType<T> *variance,
                                       double epsilon, int C, int M,
                                       const int num, const T *y) {
  int gid = blockIdx.x * blockDim.x + threadIdx.x;
  int stride = blockDim.x * gridDim.x;
  for (int i = gid; i < num; i += stride) {
    const int c = layout == framework::DataLayout::kNCHW ? (i / M) % C : i % C;
    auto y_i = static_cast<BatchNormParamType<T>>(y[i]);
    auto x_i = (y_i - bias[c]) / scale[c] / variance[c] + mean[c];
    x[i] = static_cast<T>(x_i);
  }
}

template <typename T>
class InplaceHelper {
 public:
  void operator()(const framework::DataLayout layout, T *x,
                  const BatchNormParamType<T> *scale,
                  const BatchNormParamType<T> *bias,
                  const BatchNormParamType<T> *mean,
                  const BatchNormParamType<T> *variance, double epsilon, int C,
                  int M, const int num, const T *y, int grid2, const int block,
                  const cudaStream_t &stream) {
    PADDLE_ENFORCE_EQ(x, y, platform::errors::InvalidArgument(
                                "X and Y should be inplaced in inplace mode"));
    KeBNRestoreData<<<grid2, block, 0, stream>>>(
        layout, x, scale, bias, mean, variance, epsilon, C, M, num, y);
  }
};

L
lvmengsi 已提交
425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481
template <typename T, int BlockDim, framework::DataLayout layout>
static __global__ void BNBackwardData(const T *dy,
                                      const BatchNormParamType<T> *scale,
                                      const BatchNormParamType<T> *mean,
                                      const T *x,
                                      const BatchNormParamType<T> *variance,
                                      const int C, const int N, const int HxW,
                                      T *dx) {
  const int outer_size = C;
  const int inner_size = N * HxW;
  typedef cub::BlockReduce<BatchNormParamType<T>, BlockDim> BlockReduce;
  __shared__ typename BlockReduce::TempStorage dy_storage;
  __shared__ typename BlockReduce::TempStorage dy_x_sub_mean_storage;
  __shared__ BatchNormParamType<T> dy_sum_val;
  __shared__ BatchNormParamType<T> dy_x_sub_mean_sum_val;

  for (int i = blockIdx.x; i < outer_size; i += gridDim.x) {
    BatchNormParamType<T> inv_var_i = variance[i];
    BatchNormParamType<T> mean_i = mean[i];
    BatchNormParamType<T> dy_sum = static_cast<BatchNormParamType<T>>(0);
    BatchNormParamType<T> dy_x_sub_mean_sum =
        static_cast<BatchNormParamType<T>>(0);
    for (int j = threadIdx.x; j < inner_size; j += blockDim.x) {
      const int index = layout == framework::DataLayout::kNCHW
                            ? (j / HxW * C + i) * HxW + j % HxW
                            : j * outer_size + i;
      BatchNormParamType<T> dy_i =
          static_cast<BatchNormParamType<T>>(dy[index]);
      dy_sum += dy_i;
      dy_x_sub_mean_sum +=
          dy_i * (static_cast<BatchNormParamType<T>>(x[index]) - mean_i);
    }

    dy_sum = BlockReduce(dy_storage).Reduce(dy_sum, cub::Sum());
    dy_x_sub_mean_sum = BlockReduce(dy_x_sub_mean_storage)
                            .Reduce(dy_x_sub_mean_sum, cub::Sum());

    if (threadIdx.x == 0) {
      dy_sum_val = dy_sum;
      dy_x_sub_mean_sum_val = dy_x_sub_mean_sum;
    }
    __syncthreads();

    for (int j = threadIdx.x; j < inner_size; j += blockDim.x) {
      const int index = layout == framework::DataLayout::kNCHW
                            ? (j / HxW * C + i) * HxW + j % HxW
                            : j * outer_size + i;
      dx[index] =
          (static_cast<BatchNormParamType<T>>(dy[index]) -
           dy_sum_val / static_cast<BatchNormParamType<T>>(inner_size) -
           (static_cast<BatchNormParamType<T>>(x[index]) - mean_i) *
               dy_x_sub_mean_sum_val * inv_var_i * inv_var_i / inner_size) *
          scale[i] * inv_var_i;
    }
  }
}

Q
Qiao Longfei 已提交
482
template <typename T>
Q
QI JUN 已提交
483
class BatchNormGradKernel<platform::CUDADeviceContext, T>
Q
Qiao Longfei 已提交
484 485 486
    : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext &ctx) const override {
K
Kaipeng Deng 已提交
487 488 489
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        platform::errors::InvalidArgument("It must use CUDAPlace."));
Q
Qiao Longfei 已提交
490
    double epsilon = static_cast<double>(ctx.Attr<float>("epsilon"));
Q
QI JUN 已提交
491
    const std::string data_layout_str = ctx.Attr<std::string>("data_layout");
492 493
    const bool use_global_stats = ctx.Attr<bool>("use_global_stats");

Q
QI JUN 已提交
494 495
    const DataLayout data_layout =
        framework::StringToDataLayout(data_layout_str);
Q
Qiao Longfei 已提交
496 497
    const auto *d_y = ctx.Input<Tensor>(framework::GradVarName("Y"));
    const auto *scale = ctx.Input<Tensor>("Scale");
K
Kaipeng Deng 已提交
498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523
    const auto *bias = ctx.Input<Tensor>("Bias");

    auto *d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto *d_scale = ctx.Output<Tensor>(framework::GradVarName("Scale"));
    auto *d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));

    // batch_norm with inplace as false will take X as grad input, which
    // is same as cuDNN batch_norm backward calculation, batch_norm
    // with inplace as true only take Y as input and X should be calculate
    // by inverse operation of batch_norm on Y
    const Tensor *x;
    bool is_inplace;
    if (ctx.HasInput("Y")) {
      x = ctx.Input<Tensor>("Y");
      is_inplace = true;
      PADDLE_ENFORCE_EQ(d_x, d_y,
                        platform::errors::InvalidArgument(
                            "X@GRAD and Y@GRAD not inplace in inplace mode"));
    } else {
      x = ctx.Input<Tensor>("X");
      is_inplace = false;
      PADDLE_ENFORCE_NE(d_x, d_y,
                        platform::errors::InvalidArgument(
                            "X@GRAD and Y@GRAD inplaced in non-inplace mode"));
    }

524 525 526 527 528 529 530
    const bool is_test = ctx.Attr<bool>("is_test");
    PADDLE_ENFORCE_EQ(
        is_test, false,
        platform::errors::InvalidArgument(
            "`is_test = True` CANNOT be used in train program. If "
            "you want to use global status in pre_train model, "
            "please set `use_global_stats = True`"));
Q
Qiao Longfei 已提交
531 532 533

    const auto &x_dims = x->dims();

C
ceci3 已提交
534 535 536 537 538 539 540
    PADDLE_ENFORCE_EQ(
        x_dims.size() >= 2 && x_dims.size() <= 5, true,
        platform::errors::InvalidArgument(
            "The size of input's dimensions should be between 2 and 5."
            "But received: the size of input's dimensions is [%d],"
            "the dimensions of input is [%s]",
            x_dims.size(), x_dims));
Q
Qiao Longfei 已提交
541
    int N, C, H, W, D;
Q
QI JUN 已提交
542
    ExtractNCWHD(x_dims, data_layout, &N, &C, &H, &W, &D);
Q
Qiao Longfei 已提交
543

544 545
    // init output
    d_x->mutable_data<T>(ctx.GetPlace());
K
Kaipeng Deng 已提交
546

547 548 549
    if (d_scale && d_bias) {
      d_scale->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
      d_bias->mutable_data<BatchNormParamType<T>>(ctx.GetPlace());
550
    }
C
ceci3 已提交
551 552 553 554 555 556 557 558 559 560 561 562 563
    PADDLE_ENFORCE_EQ(
        scale->dims().size(), 1UL,
        platform::errors::InvalidArgument(
            "The size of scale's dimensions must equal to 1. But received: "
            "the size of scale's dimensions is [%d], the dimensions of scale "
            "is [%s].",
            scale->dims().size(), scale->dims()));
    PADDLE_ENFORCE_EQ(
        scale->dims()[0], C,
        platform::errors::InvalidArgument(
            "The first dimension of scale must equal to Channels[%d]. But "
            "received: the first dimension of scale is [%d]",
            C, scale->dims()[0]));
Q
Qiao Longfei 已提交
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
    auto dtype = platform::CudnnDataType<T>::type;
    const auto *reserve_space = ctx.Input<Tensor>("ReserveSpace");
    const bool fast_nhwc_batch_norm =
        dtype == CUDNN_DATA_HALF && FLAGS_cudnn_batchnorm_spatial_persistent &&
        reserve_space != nullptr;
    auto compute_format =
        fast_nhwc_batch_norm && data_layout == DataLayout::kNHWC
            ? DataLayout::kNHWC
            : DataLayout::kNCHW;

    Tensor transformed_x(x->type());
    Tensor transformed_d_y(d_y->type());
    Tensor transformed_d_x(d_x->type());
    if (data_layout == DataLayout::kNHWC &&
        compute_format == DataLayout::kNCHW) {
      VLOG(3) << "Transform input tensor from NHWC to NCHW.";
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, x,
                                                           &transformed_x);
      TransToChannelFirst<platform::CUDADeviceContext, T>(ctx, x,
                                                          &transformed_x);
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, d_y,
                                                           &transformed_d_y);
      TransToChannelFirst<platform::CUDADeviceContext, T>(ctx, d_y,
                                                          &transformed_d_y);
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, d_x,
                                                           &transformed_d_x);
    } else {
      transformed_x.ShareDataWith(*x);
      transformed_d_y.ShareDataWith(*d_y);
      transformed_d_x.ShareDataWith(*d_x);
    }

Z
zchen0211 已提交
597 598
    std::vector<int> dims;
    std::vector<int> strides;
599
    if (compute_format == DataLayout::kNCHW) {
Z
zchen0211 已提交
600 601 602 603 604 605
      dims = {N, C, H, W, D};
      strides = {C * H * W * D, H * W * D, W * D, D, 1};
    } else {
      dims = {N, C, H, W, D};
      strides = {H * W * C * D, 1, W * D * C, D * C, C};
    }
Q
Qiao Longfei 已提交
606

607
    auto &dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
608
    const int num = transformed_x.numel();
L
lvmengsi 已提交
609 610 611 612 613
    const int block = 512;
    int max_threads = dev_ctx.GetMaxPhysicalThreadCount();
    const int max_blocks = std::max(max_threads / block, 1);
    int grid1 = (num + block - 1) / block;
    int grid2 = std::min(C, max_blocks);
K
Kaipeng Deng 已提交
614 615
    auto stream = dev_ctx.stream();
    InplaceHelper<T> inplace_functor;
L
lvmengsi 已提交
616

617 618 619 620 621 622 623 624 625 626 627 628 629 630 631
    if (!use_global_stats) {
      if ((N * H * W * D) == 1) {
        framework::TensorCopy(*d_y, ctx.GetPlace(), d_x);
        math::SetConstant<platform::CUDADeviceContext, BatchNormParamType<T>>
            functor;
        functor(dev_ctx, d_scale, static_cast<BatchNormParamType<T>>(0));
        functor(dev_ctx, d_bias, static_cast<BatchNormParamType<T>>(0));
        return;
      }

      // ------------------- cudnn descriptors ---------------------
      cudnnTensorDescriptor_t data_desc_;
      cudnnTensorDescriptor_t bn_param_desc_;
      cudnnBatchNormMode_t mode_;

632
      PADDLE_ENFORCE_CUDA_SUCCESS(
633
          platform::dynload::cudnnCreateTensorDescriptor(&data_desc_));
634
      PADDLE_ENFORCE_CUDA_SUCCESS(
635 636 637 638 639 640 641
          platform::dynload::cudnnCreateTensorDescriptor(&bn_param_desc_));
      if (epsilon <= CUDNN_BN_MIN_EPSILON - FLT_EPSILON) {
        LOG(ERROR) << "Provided epsilon is smaller than "
                   << "CUDNN_BN_MIN_EPSILON. Setting it to "
                   << "CUDNN_BN_MIN_EPSILON instead.";
      }
      epsilon = std::max(epsilon, CUDNN_BN_MIN_EPSILON);
642
#if CUDNN_VERSION_MIN(7, 0, 0)
W
Wu Yi 已提交
643 644 645 646 647
      if (FLAGS_cudnn_batchnorm_spatial_persistent) {
        mode_ = CUDNN_BATCHNORM_SPATIAL_PERSISTENT;
      } else {
        mode_ = CUDNN_BATCHNORM_SPATIAL;
      }
648
#else
649
      mode_ = CUDNN_BATCHNORM_SPATIAL;
650
#endif
651

652
      PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::cudnnSetTensorNdDescriptor(
653 654
          data_desc_, CudnnDataType<T>::type,
          x_dims.size() > 3 ? x_dims.size() : 4, dims.data(), strides.data()));
655 656 657
      PADDLE_ENFORCE_CUDA_SUCCESS(
          platform::dynload::cudnnDeriveBNTensorDescriptor(bn_param_desc_,
                                                           data_desc_, mode_));
658 659 660

      const auto *saved_mean = ctx.Input<Tensor>("SavedMean");
      const auto *saved_var = ctx.Input<Tensor>("SavedVariance");
L
lvmengsi 已提交
661
      const auto *saved_mean_data =
662
          saved_mean->template data<BatchNormParamType<T>>();
L
lvmengsi 已提交
663
      const auto *saved_var_data =
664 665
          saved_var->template data<BatchNormParamType<T>>();

K
Kaipeng Deng 已提交
666 667 668 669 670 671 672 673
      if (is_inplace) {
        inplace_functor(compute_format, transformed_x.data<T>(),
                        scale->template data<BatchNormParamType<T>>(),
                        bias->template data<BatchNormParamType<T>>(),
                        saved_mean_data, saved_var_data, epsilon, C, H * W * D,
                        num, transformed_x.data<T>(), grid2, block, stream);
      }

L
lvmengsi 已提交
674
      if (d_scale && d_bias) {
675 676 677 678 679 680 681 682 683
        bool called = false;
#if CUDNN_VERSION_MIN(7, 4, 1)
        if (compute_format == DataLayout::kNHWC) {
          called = true;
          size_t workspace_size = 0;
          void *workspace_ptr = nullptr;
          Tensor workspace_tensor;
          auto reserve_space_size = reserve_space->memory_size();
          // --------------- cudnn batchnorm workspace ---------------
684 685 686 687 688 689 690 691 692 693 694 695 696 697
          PADDLE_ENFORCE_CUDA_SUCCESS(
              platform::dynload::
                  cudnnGetBatchNormalizationBackwardExWorkspaceSize(
                      /*handle=*/dev_ctx.cudnn_handle(),
                      /*mode=*/mode_,
                      /*bnIps=*/CUDNN_BATCHNORM_OPS_BN,
                      /*xDesc=*/data_desc_,
                      /*yDesc=*/data_desc_,
                      /*dyDesc=*/data_desc_,
                      /*dzDesc=*/nullptr,
                      /*dxDesc=*/data_desc_,
                      /*bnScaleBiasMeanVarDesc=*/bn_param_desc_,
                      /*activationDesc=*/nullptr,
                      /*sizeInBytes=*/&workspace_size));
698 699 700 701

          workspace_ptr = workspace_tensor.mutable_data(
              ctx.GetPlace(), transformed_x.type(), workspace_size);

702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720
          PADDLE_ENFORCE_CUDA_SUCCESS(
              platform::dynload::cudnnBatchNormalizationBackwardEx(
                  /*handle=*/dev_ctx.cudnn_handle(),
                  /*mode=*/mode_,
                  /*bnOps=*/CUDNN_BATCHNORM_OPS_BN,
                  /*alphaDataDiff=*/CudnnDataType<T>::kOne(),
                  /*betaDataDiff=*/CudnnDataType<T>::kZero(),
                  /*alphaParamDiff=*/CudnnDataType<T>::kOne(),
                  /*betaParamDiff=*/CudnnDataType<T>::kZero(),
                  /*xDesc=*/data_desc_,
                  /*xData=*/transformed_x.template data<T>(),
                  /*yDesc=*/nullptr,
                  /*yData=*/nullptr,
                  /*dyDesc=*/data_desc_,
                  /*dyData=*/transformed_d_y.template data<T>(),
                  /*dzDesc=*/nullptr,
                  /*dzData=*/nullptr,
                  /*dxDesc=*/data_desc_,
                  /*dxData=*/transformed_d_x.template mutable_data<T>(
721
                      ctx.GetPlace()),
722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739
                  /*dBnScaleBiasDesc=*/bn_param_desc_,
                  /*bnScaleData=*/scale->template data<BatchNormParamType<T>>(),
                  /*bnBiasData=*/nullptr,
                  /*dBnScaleData=*/d_scale
                      ->template mutable_data<BatchNormParamType<T>>(
                          ctx.GetPlace()),
                  /*dBnBiasData=*/d_bias
                      ->template mutable_data<BatchNormParamType<T>>(
                          ctx.GetPlace()),
                  /*epsilon=*/epsilon,
                  /*savedMean=*/saved_mean_data,
                  /*savedInvVariance=*/saved_var_data,
                  /*activationDesc=*/nullptr,
                  /*workspace=*/workspace_ptr,
                  /*workSpaceSizeInBytes=*/workspace_size,
                  /*reserveSpace=*/const_cast<T *>(
                      reserve_space->template data<T>()),
                  /*reserveSpaceSizeInBytes=*/reserve_space_size));
740 741 742
        }
#endif
        if (!called) {
743 744 745 746 747 748 749 750 751 752 753 754 755 756
          PADDLE_ENFORCE_CUDA_SUCCESS(
              platform::dynload::cudnnBatchNormalizationBackward(
                  dev_ctx.cudnn_handle(), mode_, CudnnDataType<T>::kOne(),
                  CudnnDataType<T>::kZero(), CudnnDataType<T>::kOne(),
                  CudnnDataType<T>::kZero(), data_desc_,
                  transformed_x.template data<T>(), data_desc_,
                  transformed_d_y.template data<T>(), data_desc_,
                  transformed_d_x.template mutable_data<T>(ctx.GetPlace()),
                  bn_param_desc_, scale->template data<BatchNormParamType<T>>(),
                  d_scale->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  d_bias->template mutable_data<BatchNormParamType<T>>(
                      ctx.GetPlace()),
                  epsilon, saved_mean_data, saved_var_data));
757 758 759 760 761 762 763 764
        }

        if (data_layout == DataLayout::kNHWC &&
            compute_format == DataLayout::kNCHW) {
          VLOG(3) << "Transform batchnorm output from NCHW to NHWC";
          TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
              ctx, &transformed_d_x, d_x);
        }
L
lvmengsi 已提交
765
      } else {
766
        if (compute_format == DataLayout::kNCHW) {
L
lvmengsi 已提交
767 768 769 770 771 772 773 774 775
          if (d_x) {
            BNBackwardData<T, block, framework::DataLayout::kNCHW><<<
                grid2, block, 0, dev_ctx.stream()>>>(
                d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
                saved_mean_data, x->data<T>(), saved_var_data, C, N, H * W * D,
                d_x->data<T>());
          }
        } else {
          if (d_x) {
L
Lv Mengsi 已提交
776
            BNBackwardData<T, block, framework::DataLayout::kNHWC><<<
L
lvmengsi 已提交
777 778 779 780 781 782 783
                grid2, block, 0, dev_ctx.stream()>>>(
                d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
                saved_mean_data, x->data<T>(), saved_var_data, C, N, H * W * D,
                d_x->data<T>());
          }
        }
      }
784 785

      // clean when exit.
786
      PADDLE_ENFORCE_CUDA_SUCCESS(
787
          platform::dynload::cudnnDestroyTensorDescriptor(data_desc_));
788
      PADDLE_ENFORCE_CUDA_SUCCESS(
789 790 791 792 793 794 795 796 797 798
          platform::dynload::cudnnDestroyTensorDescriptor(bn_param_desc_));
    } else {
      const auto *running_mean = ctx.Input<Tensor>("Mean");
      const auto *running_var = ctx.Input<Tensor>("Variance");

      const auto *running_mean_data =
          running_mean->template data<BatchNormParamType<T>>();
      const auto *running_var_data =
          running_var->template data<BatchNormParamType<T>>();

K
Kaipeng Deng 已提交
799 800 801 802 803 804 805 806 807
      if (is_inplace) {
        auto px = *x;
        inplace_functor(data_layout, px.mutable_data<T>(ctx.GetPlace()),
                        scale->template data<BatchNormParamType<T>>(),
                        bias->template data<BatchNormParamType<T>>(),
                        running_mean_data, running_var_data, epsilon, C,
                        H * W * D, num, x->data<T>(), grid2, block, stream);
      }

808
      if (compute_format == DataLayout::kNCHW) {
809
        if (d_x) {
K
Kaipeng Deng 已提交
810 811
          KeBNBackwardData<
              T, framework::DataLayout::kNCHW><<<grid1, block, 0, stream>>>(
812 813 814 815
              d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
              running_var_data, epsilon, C, H * W, num, d_x->data<T>());
        }
        if (d_scale && d_bias) {
K
Kaipeng Deng 已提交
816 817 818
          KeBNBackwardScaleBias<
              T, block,
              framework::DataLayout::kNCHW><<<grid2, block, 0, stream>>>(
819
              d_y->data<T>(), x->data<T>(), running_mean_data, running_var_data,
Q
qingqing01 已提交
820
              epsilon, N, C, H * W * D, d_scale->data<BatchNormParamType<T>>(),
821 822 823 824
              d_bias->data<BatchNormParamType<T>>());
        }
      } else {
        if (d_x) {
K
Kaipeng Deng 已提交
825 826
          KeBNBackwardData<
              T, framework::DataLayout::kNHWC><<<grid1, block, 0, stream>>>(
827 828 829 830
              d_y->data<T>(), scale->data<BatchNormParamType<T>>(),
              running_var_data, epsilon, C, H * W, num, d_x->data<T>());
        }
        if (d_scale && d_bias) {
K
Kaipeng Deng 已提交
831 832 833
          KeBNBackwardScaleBias<
              T, block,
              framework::DataLayout::kNHWC><<<grid2, block, 0, stream>>>(
834
              d_y->data<T>(), x->data<T>(), running_mean_data, running_var_data,
Q
qingqing01 已提交
835
              epsilon, N, C, H * W * D, d_scale->data<BatchNormParamType<T>>(),
836 837 838 839
              d_bias->data<BatchNormParamType<T>>());
        }
      }
    }
Q
Qiao Longfei 已提交
840 841 842 843 844 845 846
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
K
Kexin Zhao 已提交
847
namespace plat = paddle::platform;
Q
QI JUN 已提交
848
REGISTER_OP_CUDA_KERNEL(
K
Kexin Zhao 已提交
849
    batch_norm, ops::BatchNormKernel<plat::CUDADeviceContext, float>,
D
dzhwinter 已提交
850
    ops::BatchNormKernel<plat::CUDADeviceContext, double>,
K
Kexin Zhao 已提交
851
    ops::BatchNormKernel<plat::CUDADeviceContext, plat::float16>);
Q
QI JUN 已提交
852
REGISTER_OP_CUDA_KERNEL(
D
dzhwinter 已提交
853
    batch_norm_grad, ops::BatchNormGradKernel<plat::CUDADeviceContext, float>,
C
chengduo 已提交
854 855
    ops::BatchNormGradKernel<plat::CUDADeviceContext, double>,
    ops::BatchNormGradKernel<plat::CUDADeviceContext, plat::float16>);