conv_transpose_cudnn_op.cu 24.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.

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. */

#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/operators/conv_transpose_op.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/operators/math/padding.h"
#include "paddle/fluid/platform/cudnn_helper.h"

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
using ScopedTensorDescriptor = platform::ScopedTensorDescriptor;
using ScopedFilterDescriptor = platform::ScopedFilterDescriptor;
using ScopedConvolutionDescriptor = platform::ScopedConvolutionDescriptor;
using DataLayout = platform::DataLayout;

static constexpr size_t kConvCUDNNWorkspaceLimitBytes = 1024 * 1024 * 1024;

template <typename T, int D>
static void DataTranspose(const framework::ExecutionContext& ctx,
                          const Tensor* input, Tensor* output,
                          const std::vector<int>& axis, int flag = 0) {
  auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
  math::Transpose<platform::CUDADeviceContext, T, D> transpose;
  auto in_dims = input->dims();
  std::vector<int64_t> input_transpose_vec;
  for (size_t i = 0; i < axis.size(); ++i) {
    if (flag == 0)
      input_transpose_vec.push_back(in_dims[axis[i]]);
    else
      input_transpose_vec.push_back(in_dims[i]);
  }
  framework::DDim input_transpose_dims(
      framework::make_ddim(input_transpose_vec));
  output->mutable_data<T>(input_transpose_dims, ctx.GetPlace());
  transpose(dev_ctx, *input, output, axis);
}

template <typename T>
class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE_EQ(platform::is_gpu_place(ctx.GetPlace()), true,
                      "It must use CUDAPlace.");
    auto* input = ctx.Input<Tensor>("Input");
    auto* filter = ctx.Input<Tensor>("Filter");
    auto* output = ctx.Output<Tensor>("Output");

    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    std::string padding_algorithm = ctx.Attr<std::string>("padding_algorithm");

    // cudnn v5 does not support dilations
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
    int user_workspace_size = ctx.Attr<int>("workspace_size_MB");
    const T* filter_data = filter->data<T>();
    const std::string data_layout_str = ctx.Attr<std::string>("data_format");
    const paddle::operators::DataLayout data_layout =
        (data_layout_str == "NCHW" ? DataLayout::kNCHW : DataLayout::kNHWC);

    // if channel_last, transpose to channel_first
    Tensor input_transpose;
    std::vector<int> input_vec = framework::vectorize<int>(input->dims());
    std::vector<int> output_vec = framework::vectorize<int>(output->dims());
    if (data_layout == DataLayout::kNHWC) {
      if (strides.size() == 2U) {
        std::vector<int> axis = {0, 3, 1, 2};
        for (size_t i = 0; i < axis.size(); ++i) {
          input_vec[i] = input->dims()[axis[i]];
          output_vec[i] = output->dims()[axis[i]];
        }
        DataTranspose<T, 4>(ctx, input, &input_transpose, axis);
      } else if (strides.size() == 3U) {
        std::vector<int> axis = {0, 4, 1, 2, 3};
        for (size_t i = 0; i < axis.size(); ++i) {
          input_vec[i] = input->dims()[axis[i]];
          output_vec[i] = output->dims()[axis[i]];
        }
        DataTranspose<T, 5>(ctx, input, &input_transpose, axis);
      }
    } else {
      input_transpose = *input;
    }

    // update padding and dilation
    auto in_dims = input_transpose.dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    framework::DDim filter_data_dims =
        framework::slice_ddim(filter_dims, 2, filter_dims.size());
    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
                             in_data_dims, strides, ksize);

    int data_dim = strides.size();  // 2d or 3d
113
    bool is_sys_pad = math::IsSymmetricPadding(paddings, data_dim);
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318

    std::vector<int> input_pad(input_transpose.dims().size() * 2, 0);
    Tensor transformed_input;
    std::vector<int> padding_common(data_dim, 0);
    if (!is_sys_pad) {
      std::vector<int> padding_diff(data_dim);
      std::vector<int> new_input_shape_vec(data_dim + 2);
      new_input_shape_vec[0] = input_transpose.dims()[0];
      new_input_shape_vec[1] = input_transpose.dims()[1];

      for (size_t i = 0; i < data_dim; ++i) {
        padding_diff[i] = std::abs(paddings[2 * i] - paddings[2 * i + 1]);
        padding_common[i] = std::min(paddings[2 * i], paddings[2 * i + 1]);
        new_input_shape_vec[i + 2] =
            input_transpose.dims()[i + 2] + padding_diff[i];
        input_pad[2 * i + 4] = paddings[2 * i] - padding_common[i];
        input_pad[2 * i + 4 + 1] = paddings[2 * i + 1] - padding_common[i];
      }
      framework::DDim new_input_shape(
          framework::make_ddim(new_input_shape_vec));
      transformed_input.Resize(new_input_shape);
      auto& dev_ctx =
          ctx.template device_context<paddle::platform::CUDADeviceContext>();

      transformed_input =
          ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
              new_input_shape, dev_ctx);
      const int rank = input_transpose.dims().size();
      T pad_value(0.0);
      switch (rank) {
        case 4: {
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
              ctx, input_pad, input_transpose, pad_value, &transformed_input);
        } break;
        case 5: {
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
              ctx, input_pad, input_transpose, pad_value, &transformed_input);
        } break;
        default:
          PADDLE_ENFORCE_EQ(
              rank == 4 || rank == 5, true,
              "Op(ConvTranspose) only supports 4-D or 5-D input Tensor.");
      }
    } else {
      transformed_input = input_transpose;
      if (paddings.size() == data_dim) {
        for (size_t i = 0; i < data_dim; ++i) {
          padding_common[i] = paddings[i];
        }
      } else {
        for (size_t i = 0; i < data_dim; ++i) {
          padding_common[i] = paddings[2 * i];
        }
      }
    }

    std::vector<int64_t> starts(data_dim, 0);
    std::vector<int64_t> ends(data_dim, 0);
    std::vector<int64_t> axes(data_dim, 0);
    for (size_t i = 0; i < data_dim; ++i) {
      starts[i] = input_pad[2 * i + 4] * (strides[i] + 1);
      ends[i] = starts[i] + output_vec[i + 2];
      axes[i] = i + 2;
    }

    const T* input_data = transformed_input.data<T>();
    input_vec = framework::vectorize<int>(transformed_input.dims());

    std::vector<int> transformed_output_vec = output_vec;
    for (size_t i = 0; i < data_dim; ++i) {
      transformed_output_vec[i + 2] =
          output_vec[i + 2] +
          (input_pad[2 * i + 4] + input_pad[2 * i + 5]) * strides[i] -
          2 * padding_common[i] + paddings[2 * i] + paddings[2 * i + 1];
    }

    Tensor transformed_output;
    if (!is_sys_pad) {
      DDim transformed_output_shape(
          framework::make_ddim(transformed_output_vec));
      transformed_output.mutable_data<T>(transformed_output_shape,
                                         ctx.GetPlace());
    } else {
      output->mutable_data<T>(ctx.GetPlace());
      transformed_output.ShareDataWith(*output);
      transformed_output.Resize(framework::make_ddim(transformed_output_vec));
    }
    T* transformed_output_data = transformed_output.data<T>();

    // ------------------- cudnn descriptors ---------------------
    ScopedTensorDescriptor input_desc;
    ScopedTensorDescriptor output_desc;
    ScopedFilterDescriptor filter_desc;
    ScopedConvolutionDescriptor conv_desc;
    DataLayout layout;

    if (strides.size() == 2U) {
      layout = DataLayout::kNCHW;
    } else {
      layout = DataLayout::kNCDHW;
    }

    // (N, M, H, W) or (N, M, D, H, W)
    cudnnTensorDescriptor_t cudnn_input_desc =
        input_desc.descriptor<T>(layout, input_vec, groups);
    // (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w)
    cudnnTensorDescriptor_t cudnn_output_desc =
        output_desc.descriptor<T>(layout, transformed_output_vec, groups);
    // (M, C, K_h, K_w) or (M, C, K_d, K_h, K_w)
    cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor<T>(
        layout, framework::vectorize<int>(filter->dims()), groups);
    cudnnConvolutionDescriptor_t cudnn_conv_desc =
        conv_desc.descriptor<T>(padding_common, strides, dilations);

    // ------------------- cudnn conv workspace ---------------------
    size_t workspace_size_in_bytes;  // final workspace to allocate.
    size_t workspace_size_limit = kConvCUDNNWorkspaceLimitBytes;
    if (user_workspace_size > 0) {
      workspace_size_limit = user_workspace_size * 1024 * 1024;
    }
    // ------------------- cudnn conv algorithm ---------------------
    cudnnConvolutionBwdDataAlgo_t algo;
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
    auto handle = dev_ctx.cudnn_handle();
    // Get the algorithm
    CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
        handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc,
        // dxDesc: Handle to the previously initialized output tensor
        // descriptor.
        cudnn_output_desc, CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
        workspace_size_limit, &algo));

    // get workspace size able to allocate
    CUDNN_ENFORCE(
        platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
            handle, cudnn_filter_desc, cudnn_input_desc, cudnn_conv_desc,
            cudnn_output_desc, algo, &workspace_size_in_bytes));

    // ------------------- cudnn conv transpose forward ---------------------
    int input_offset =
        transformed_input.numel() / transformed_input.dims()[0] / groups;
    int output_offset =
        transformed_output.numel() / transformed_output.dims()[0] / groups;
    int filter_offset = filter->numel() / groups;
    T alpha = 1.0f, beta = 0.0f;
    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
    for (int g = 0; g < groups; g++) {
      auto cudnn_func = [&](void* cudnn_workspace) {
        CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardData(
            handle, &alpha, cudnn_filter_desc, filter_data + filter_offset * g,
            cudnn_input_desc, input_data + input_offset * g, cudnn_conv_desc,
            algo, cudnn_workspace, workspace_size_in_bytes, &beta,
            cudnn_output_desc, transformed_output_data + output_offset * g));
      };
      workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
    }

    if (!is_sys_pad && strides.size() == 2U) {
      Slice<paddle::platform::CUDADeviceContext, T, 4>(
          ctx, &transformed_output, output, starts, ends, axes);
    } else if (!is_sys_pad && strides.size() == 3U) {
      Slice<paddle::platform::CUDADeviceContext, T, 5>(
          ctx, &transformed_output, output, starts, ends, axes);
    }

    if (data_layout == DataLayout::kNHWC) {
      Tensor output_transpose;
      Tensor output_nchw;
      output_nchw.ShareDataWith(*output);
      output_nchw.Resize(framework::make_ddim(output_vec));
      if (strides.size() == 2U) {
        std::vector<int> axis = {0, 2, 3, 1};
        DataTranspose<T, 4>(ctx, &output_nchw, &output_transpose, axis);
        *output = output_transpose;
      } else if (strides.size() == 3U) {
        std::vector<int> axis = {0, 2, 3, 4, 1};
        DataTranspose<T, 5>(ctx, &output_nchw, &output_transpose, axis);
        *output = output_transpose;
      }
    }
  }
};

template <typename T>
class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    PADDLE_ENFORCE(platform::is_gpu_place(ctx.GetPlace()),
                   "It must use CUDAPlace.");
    auto input = ctx.Input<Tensor>("Input");
    auto filter = ctx.Input<Tensor>("Filter");
    auto output_grad = ctx.Input<Tensor>(framework::GradVarName("Output"));
    auto input_grad = ctx.Output<Tensor>(framework::GradVarName("Input"));
    auto filter_grad = ctx.Output<Tensor>(framework::GradVarName("Filter"));
    const T* filter_data = filter->data<T>();

    std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    // cudnn v5 does not support dilations
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
    std::string padding_algorithm = ctx.Attr<std::string>("padding_algorithm");
    int user_workspace_size = ctx.Attr<int>("workspace_size_MB");
    const std::string data_layout_str = ctx.Attr<std::string>("data_format");
    const paddle::operators::DataLayout data_layout =
319
        (data_layout_str != "NHWC" ? DataLayout::kNCHW : DataLayout::kNHWC);
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361

    // if channel_last, transpose to channel_first
    Tensor input_transpose;
    Tensor output_grad_transpose;
    std::vector<int> input_vec = framework::vectorize<int>(input->dims());
    std::vector<int> output_vec =
        framework::vectorize<int>(output_grad->dims());
    if (data_layout == DataLayout::kNHWC) {
      if (strides.size() == 2U) {
        std::vector<int> axis = {0, 3, 1, 2};
        for (size_t i = 0; i < axis.size(); ++i) {
          input_vec[i] = input->dims()[axis[i]];
          output_vec[i] = output_grad->dims()[axis[i]];
        }
        DataTranspose<T, 4>(ctx, input, &input_transpose, axis);
        DataTranspose<T, 4>(ctx, output_grad, &output_grad_transpose, axis);
      } else if (strides.size() == 3U) {
        std::vector<int> axis = {0, 4, 1, 2, 3};
        for (size_t i = 0; i < axis.size(); ++i) {
          input_vec[i] = input->dims()[axis[i]];
          output_vec[i] = output_grad->dims()[axis[i]];
        }
        DataTranspose<T, 5>(ctx, input, &input_transpose, axis);
        DataTranspose<T, 5>(ctx, output_grad, &output_grad_transpose, axis);
      }
    } else {
      input_transpose = *input;
      output_grad_transpose = *output_grad;
    }

    // update padding and dilation
    auto in_dims = input_transpose.dims();
    auto filter_dims = filter->dims();
    framework::DDim in_data_dims;
    in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
    framework::DDim filter_data_dims =
        framework::slice_ddim(filter_dims, 2, filter_dims.size());
    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
                             in_data_dims, strides, ksize);

    int data_dim = strides.size();  // 2d or 3d
362
    bool is_sys_pad = math::IsSymmetricPadding(paddings, data_dim);
363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572

    std::vector<int> input_pad(input_transpose.dims().size() * 2, 0);
    Tensor transformed_output_grad;
    std::vector<int> padding_common(data_dim, 0);
    if (!is_sys_pad) {
      std::vector<int> padding_diff(data_dim);
      std::vector<int> new_output_grad_shape_vec(data_dim + 2);
      new_output_grad_shape_vec[0] = output_grad_transpose.dims()[0];
      new_output_grad_shape_vec[1] = output_grad_transpose.dims()[1];

      for (size_t i = 0; i < data_dim; ++i) {
        padding_diff[i] = std::abs(paddings[2 * i] - paddings[2 * i + 1]);
        padding_common[i] = std::min(paddings[2 * i], paddings[2 * i + 1]);
        new_output_grad_shape_vec[i + 2] =
            output_grad_transpose.dims()[i + 2] + padding_diff[i];
        input_pad[2 * i + 4] = paddings[2 * i] - padding_common[i];
        input_pad[2 * i + 4 + 1] = paddings[2 * i + 1] - padding_common[i];
      }
      framework::DDim new_output_grad_shape(
          framework::make_ddim(new_output_grad_shape_vec));
      transformed_output_grad.Resize(new_output_grad_shape);
      auto& dev_ctx =
          ctx.template device_context<paddle::platform::CUDADeviceContext>();

      transformed_output_grad =
          ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
              new_output_grad_shape, dev_ctx);
      const int rank = input_transpose.dims().size();
      T pad_value(0.0);
      switch (rank) {
        case 4: {
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
              ctx, input_pad, output_grad_transpose, pad_value,
              &transformed_output_grad);
        } break;
        case 5: {
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
              ctx, input_pad, output_grad_transpose, pad_value,
              &transformed_output_grad);
        } break;
        default:
          PADDLE_ENFORCE_EQ(
              rank == 4 || rank == 5, true,
              "Op(ConvTranspose) only supports 4-D or 5-D input Tensor.");
      }
    } else {
      transformed_output_grad = output_grad_transpose;
      if (paddings.size() == data_dim) {
        for (size_t i = 0; i < data_dim; ++i) {
          padding_common[i] = paddings[i];
        }
      } else {
        for (size_t i = 0; i < data_dim; ++i) {
          padding_common[i] = paddings[2 * i];
        }
      }
    }

    const T* input_data = input_transpose.data<T>();
    const T* output_grad_data = transformed_output_grad.data<T>();
    output_vec = framework::vectorize<int>(transformed_output_grad.dims());

    // ------------------- cudnn descriptors ---------------------
    ScopedTensorDescriptor input_desc;
    ScopedTensorDescriptor output_desc;
    ScopedFilterDescriptor filter_desc;
    ScopedConvolutionDescriptor conv_desc;
    DataLayout layout;

    if (strides.size() == 2U) {
      layout = DataLayout::kNCHW;
    } else {
      layout = DataLayout::kNCDHW;
    }

    // Input: (N, M, H, W) or (N, M, D, H, W)
    cudnnTensorDescriptor_t cudnn_input_desc =
        input_desc.descriptor<T>(layout, input_vec, groups);
    // Output: (N, C, O_h, O_w) or (N, C, O_d, O_h, O_w)
    cudnnTensorDescriptor_t cudnn_output_desc =
        output_desc.descriptor<T>(layout, output_vec, groups);
    // Filter (M, C, K_h, K_w) or (M, C, K_d K_h, K_w)
    cudnnFilterDescriptor_t cudnn_filter_desc = filter_desc.descriptor<T>(
        layout, framework::vectorize<int>(filter->dims()), groups);

    cudnnConvolutionDescriptor_t cudnn_conv_desc =
        conv_desc.descriptor<T>(padding_common, strides, dilations);

    // ------------------- cudnn backward algorithm ---------------------
    cudnnConvolutionFwdAlgo_t data_algo;
    cudnnConvolutionBwdFilterAlgo_t filter_algo;
    size_t bwd_filter_ws_size, fwd_ws_size;
    size_t workspace_size_in_bytes = 0;
    size_t workspace_size_limit = kConvCUDNNWorkspaceLimitBytes;
    if (user_workspace_size > 0) {
      workspace_size_limit = user_workspace_size * 1024 * 1024;
    }

    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
    auto handle = dev_ctx.cudnn_handle();
    if (input_grad) {
      // choose backward algorithm for data
      CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardAlgorithm(
          handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc,
          cudnn_input_desc, CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
          workspace_size_limit, &data_algo));
      CUDNN_ENFORCE(platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
          handle, cudnn_output_desc, cudnn_filter_desc, cudnn_conv_desc,
          cudnn_input_desc, data_algo, &fwd_ws_size));
      workspace_size_in_bytes = std::max(workspace_size_in_bytes, fwd_ws_size);
    }

    if (filter_grad) {
      // choose backward algorithm for filter
      CUDNN_ENFORCE(
          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
              handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc,
              cudnn_filter_desc,
              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
              workspace_size_limit, &filter_algo));

      // get workspace for backwards filter algorithm
      CUDNN_ENFORCE(
          platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
              handle, cudnn_output_desc, cudnn_input_desc, cudnn_conv_desc,
              cudnn_filter_desc, filter_algo, &bwd_filter_ws_size));
      workspace_size_in_bytes =
          std::max(workspace_size_in_bytes, bwd_filter_ws_size);
    }

    // ------------------- cudnn conv backward data ---------------------
    // FIXME(typhoonzero): template type T may not be the same as cudnn call.
    int input_offset = input->numel() / input->dims()[0] / groups;
    int output_grad_offset = transformed_output_grad.numel() /
                             transformed_output_grad.dims()[0] / groups;
    int filter_offset = filter->numel() / groups;
    T alpha = 1.0f, beta = 0.0f;
    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
    if (input_grad) {
      T* input_grad_data = input_grad->mutable_data<T>(ctx.GetPlace());
      // Because beta is zero, it is unnecessary to reset input_grad.
      for (int g = 0; g < groups; g++) {
        auto cudnn_func = [&](void* cudnn_workspace) {
          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionForward(
              handle, &alpha, cudnn_output_desc,
              output_grad_data + output_grad_offset * g, cudnn_filter_desc,
              filter_data + filter_offset * g, cudnn_conv_desc, data_algo,
              cudnn_workspace, workspace_size_in_bytes, &beta, cudnn_input_desc,
              input_grad_data + input_offset * g));
        };
        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
      }

      if (data_layout == DataLayout::kNHWC) {
        Tensor input_grad_transpose;
        Tensor input_grad_nchw;
        input_grad_nchw.ShareDataWith(*input_grad);
        input_grad_nchw.Resize(framework::make_ddim(input_vec));
        if (strides.size() == 2U) {
          std::vector<int> axis = {0, 2, 3, 1};
          DataTranspose<T, 4>(ctx, &input_grad_nchw, &input_grad_transpose,
                              axis);
          *input_grad = input_grad_transpose;
        } else if (strides.size() == 3U) {
          std::vector<int> axis = {0, 2, 3, 4, 1};
          DataTranspose<T, 5>(ctx, &input_grad_nchw, &input_grad_transpose,
                              axis);
          *input_grad = input_grad_transpose;
        }
      }
    }

    // ------------------- cudnn conv backward filter ---------------------
    if (filter_grad) {
      T* filter_grad_data = filter_grad->mutable_data<T>(ctx.GetPlace());
      // Because beta is zero, it is unnecessary to reset filter_grad.
      // Gradient with respect to the filter
      for (int g = 0; g < groups; g++) {
        auto cudnn_func = [&](void* cudnn_workspace) {
          CUDNN_ENFORCE(platform::dynload::cudnnConvolutionBackwardFilter(
              handle, &alpha, cudnn_output_desc,
              output_grad_data + output_grad_offset * g, cudnn_input_desc,
              input_data + input_offset * g, cudnn_conv_desc, filter_algo,
              cudnn_workspace, workspace_size_in_bytes, &beta,
              cudnn_filter_desc, filter_grad_data + filter_offset * g));
        };
        workspace_handle.RunFunc(cudnn_func, workspace_size_in_bytes);
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;

REGISTER_OP_KERNEL(conv2d_transpose, CUDNN, ::paddle::platform::CUDAPlace,
                   ops::CUDNNConvTransposeOpKernel<float>,
                   ops::CUDNNConvTransposeOpKernel<double>);
REGISTER_OP_KERNEL(conv2d_transpose_grad, CUDNN, ::paddle::platform::CUDAPlace,
                   ops::CUDNNConvTransposeGradOpKernel<float>,
                   ops::CUDNNConvTransposeGradOpKernel<double>);

REGISTER_OP_KERNEL(conv3d_transpose, CUDNN, ::paddle::platform::CUDAPlace,
                   ops::CUDNNConvTransposeOpKernel<float>,
                   ops::CUDNNConvTransposeOpKernel<double>);
REGISTER_OP_KERNEL(conv3d_transpose_grad, CUDNN, ::paddle::platform::CUDAPlace,
                   ops::CUDNNConvTransposeGradOpKernel<float>,
                   ops::CUDNNConvTransposeGradOpKernel<double>);