conv_cudnn_op.cu 47.4 KB
Newer Older
L
liym27 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
/* 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 spopecific language governing permissions and
limitations under the License. */

#include <utility>
#include <vector>
17

L
liym27 已提交
18 19 20 21 22 23 24
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/memory/memory.h"
#include "paddle/fluid/operators/conv_cudnn_helper.h"
#include "paddle/fluid/operators/conv_cudnn_op_cache.h"
#include "paddle/fluid/operators/conv_op.h"
25
#include "paddle/fluid/operators/math/padding.h"
L
liym27 已提交
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43
#include "paddle/fluid/platform/cudnn_helper.h"
#include "paddle/fluid/platform/cudnn_workspace_helper.h"
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/platform/profiler.h"

DECLARE_bool(cudnn_deterministic);
DECLARE_uint64(conv_workspace_size_limit);
DECLARE_bool(cudnn_exhaustive_search);

namespace paddle {
namespace operators {

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

44 45 46 47
static inline bool IsVoltaOrLater(const platform::CUDADeviceContext& dev_ctx) {
  return dev_ctx.GetComputeCapability() >= 70;
}

L
liym27 已提交
48 49 50 51 52
template <typename T>
class CUDNNConvOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
53 54 55
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        paddle::platform::errors::PreconditionNotMet("It must use CUDAPlace."));
L
liym27 已提交
56 57 58 59 60 61 62 63
    const Tensor* input = ctx.Input<Tensor>("Input");
    auto* filter = ctx.Input<Tensor>("Filter");
    auto* output = ctx.Output<Tensor>("Output");
    output->mutable_data<T>(ctx.GetPlace());
    const std::vector<int> strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
64

L
liym27 已提交
65 66
    bool exhaustive_search =
        FLAGS_cudnn_exhaustive_search || ctx.Attr<bool>("exhaustive_search");
67 68 69 70 71 72
    bool deterministic = FLAGS_cudnn_deterministic;
    auto exhaustive_deterministic = exhaustive_search && deterministic;
    PADDLE_ENFORCE_EQ(exhaustive_deterministic, false,
                      platform::errors::InvalidArgument(
                          "Cann't set exhaustive_search True and "
                          "FLAGS_cudnn_deterministic True at same time."));
L
liym27 已提交
73 74 75 76 77 78

    const std::string padding_algorithm =
        ctx.Attr<std::string>("padding_algorithm");
    const std::string data_format = ctx.Attr<std::string>("data_format");
    const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");

79 80 81 82 83 84 85 86 87 88 89 90 91 92
    auto dtype = platform::CudnnDataType<T>::type;

    // Tensor Core introduced from Volta GPUs supports more faster conv op
    // with FP16 in NHWC data format.
    const bool compute_in_nhwc =
        dtype == CUDNN_DATA_HALF && IsVoltaOrLater(dev_ctx);
    // We will only do data format conversion from NHWC to NCHW.
    // cudnn will convert NCHW to NHWC automatically on Tensor Core.
    auto compute_format =
        compute_in_nhwc && channel_last ? DataLayout::kNHWC : DataLayout::kNCHW;
    VLOG(3) << "Compute ConvOp with cuDNN:"
            << " data_format=" << data_format << " compute_format="
            << (compute_format == DataLayout::kNHWC ? "NHWC" : "NCHW");

L
liym27 已提交
93 94 95
    // ------------ transformed tensor -----------
    Tensor transformed_input_channel(input->type());
    Tensor transformed_output(output->type());
96
    Tensor transformed_filter_channel(filter->type());
L
liym27 已提交
97
    T* output_data = nullptr;
98 99
    if (channel_last && compute_format == DataLayout::kNCHW) {
      VLOG(3) << "Transform input tensor from NHWC to NCHW.";
L
liym27 已提交
100 101 102 103 104 105 106 107 108
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, input, &transformed_input_channel);
      TransToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, input, &transformed_input_channel);

      ResizeToChannelFirst<platform::CUDADeviceContext, T>(ctx, output,
                                                           &transformed_output);

    } else {
109 110 111 112 113 114 115 116 117 118 119
      transformed_input_channel.ShareDataWith(*input);
      transformed_output.ShareDataWith(*output);
    }
    if (compute_format == DataLayout::kNHWC) {
      VLOG(3) << "Transform filter tensor from NCHW to NHWC.";
      ResizeToChannelLast<platform::CUDADeviceContext, T>(
          ctx, filter, &transformed_filter_channel);
      TransToChannelLast<platform::CUDADeviceContext, T>(
          ctx, filter, &transformed_filter_channel);
    } else {
      transformed_filter_channel.ShareDataWith(*filter);
L
liym27 已提交
120 121 122 123 124
    }
    output_data = transformed_output.data<T>();

    // update padding and dilation
    auto in_dims = transformed_input_channel.dims();
125
    auto filter_dims = transformed_filter_channel.dims();
L
liym27 已提交
126
    framework::DDim in_data_dims;
127 128 129 130 131 132 133 134 135 136 137
    framework::DDim filter_data_dims;

    if (compute_format == DataLayout::kNCHW) {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
      filter_data_dims =
          framework::slice_ddim(filter_dims, 2, filter_dims.size());
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
      filter_data_dims =
          framework::slice_ddim(filter_dims, 1, filter_dims.size() - 1);
    }
L
liym27 已提交
138 139 140 141 142 143

    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
144
    bool is_sys_pad = math::IsSymmetricPadding(paddings, data_dim);
L
liym27 已提交
145 146 147 148 149 150 151

    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] = transformed_input_channel.dims()[0];
152 153 154 155 156 157 158

      if (compute_format == DataLayout::kNCHW) {
        new_input_shape_vec[1] = transformed_input_channel.dims()[1];
      } else {
        new_input_shape_vec[data_dim + 1] =
            transformed_input_channel.dims()[data_dim + 1];
      }
L
liym27 已提交
159 160 161 162 163 164

      std::vector<int> input_pad(transformed_input_channel.dims().size() * 2,
                                 0);
      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]);
165 166 167 168 169 170 171 172 173 174 175 176 177 178
        if (compute_format == DataLayout::kNCHW) {
          new_input_shape_vec[i + 2] =
              transformed_input_channel.dims()[i + 2] + padding_diff[i];
        } else {
          new_input_shape_vec[i + 1] =
              transformed_input_channel.dims()[i + 1] + padding_diff[i];
        }
        if (compute_format == DataLayout::kNCHW) {
          input_pad[2 * i + 4] = paddings[2 * i] - padding_common[i];
          input_pad[2 * i + 4 + 1] = paddings[2 * i + 1] - padding_common[i];
        } else {
          input_pad[2 * i + 2] = paddings[2 * i] - padding_common[i];
          input_pad[2 * i + 2 + 1] = paddings[2 * i + 1] - padding_common[i];
        }
L
liym27 已提交
179 180 181 182 183 184 185 186 187 188 189 190 191 192
      }
      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 = transformed_input_channel.dims().size();
      T pad_value(0.0);
      switch (rank) {
        case 4: {
193
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
L
liym27 已提交
194 195 196 197
              ctx, input_pad, transformed_input_channel, pad_value,
              &transformed_input);
        } break;
        case 5: {
198
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
L
liym27 已提交
199 200 201 202
              ctx, input_pad, transformed_input_channel, pad_value,
              &transformed_input);
        } break;
        default:
203 204
          PADDLE_THROW(platform::errors::InvalidArgument(
              "ConvOp only support tensors with 4 or 5 dimensions."));
L
liym27 已提交
205 206 207
      }

    } else {
208
      transformed_input.ShareDataWith(transformed_input_channel);
L
liym27 已提交
209 210 211 212 213 214 215 216 217 218 219 220
      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 = transformed_input.data<T>();
221
    const T* filter_data = transformed_filter_channel.data<T>();
L
liym27 已提交
222 223

    // ------------------- cudnn descriptors ---------------------
224 225 226 227 228 229 230
    ConvArgs args{&transformed_input,
                  &transformed_filter_channel,
                  &transformed_output,
                  strides,
                  padding_common,
                  dilations,
                  dtype};
L
liym27 已提交
231 232 233

    auto handle = dev_ctx.cudnn_handle();
    auto workspace_handle = dev_ctx.cudnn_workspace_handle();
234 235 236 237 238
    DataLayout layout = compute_format == DataLayout::kNHWC ? DataLayout::kNHWC
                                                            : DataLayout::kNCHW;
    if (transformed_input.dims().size() == 5) {
      layout = compute_format == DataLayout::kNHWC ? DataLayout::kNDHWC
                                                   : DataLayout::kNCDHW;
L
liym27 已提交
239 240 241 242 243 244 245 246 247 248
    }
    auto layout_format = GetCudnnTensorFormat(layout);

    args.handle = handle;
    args.cdesc.set(dtype, padding_common, strides, dilations);

#if CUDNN_VERSION_MIN(7, 0, 1)
    // cudnn 7 can support groups, no need to do it manually
    // FIXME(typhoonzero): find a better way to disable groups
    // rather than setting it to 1.
249 250 251
    PADDLE_ENFORCE_CUDA_SUCCESS(
        platform::dynload::cudnnSetConvolutionGroupCount(args.cdesc.desc(),
                                                         groups));
L
liym27 已提交
252 253
    groups = 1;
#endif
254 255 256
    args.idesc.set(transformed_input, layout_format);
    args.wdesc.set(transformed_filter_channel, layout_format, groups);
    args.odesc.set(transformed_output, layout_format);
L
liym27 已提交
257 258
    int i_n, i_c, i_d, i_h, i_w;
    int o_n, o_c, o_d, o_h, o_w;
259 260 261 262 263 264 265 266 267 268 269 270

    if (compute_format == DataLayout::kNHWC) {
      GetNCDHW(transformed_input.dims(), DataLayout::kNHWC, &i_n, &i_c, &i_d,
               &i_h, &i_w);
      GetNCDHW(transformed_output.dims(), DataLayout::kNHWC, &o_n, &o_c, &o_d,
               &o_h, &o_w);
    } else {
      GetNCDHW(transformed_input.dims(), DataLayout::kNCHW, &i_n, &i_c, &i_d,
               &i_h, &i_w);
      GetNCDHW(transformed_output.dims(), DataLayout::kNCHW, &o_n, &o_c, &o_d,
               &o_h, &o_w);
    }
L
liym27 已提交
271 272 273

    int group_offset_in = i_c / groups * i_h * i_w * i_d;
    int group_offset_out = o_c / groups * o_h * o_w * o_d;
274
    int group_offset_filter = transformed_filter_channel.numel() / groups;
L
liym27 已提交
275 276 277 278 279 280
    // ------------------- cudnn conv workspace ---------------------
    size_t workspace_size = 0;  // final workspace to allocate.
    // ------------------- cudnn conv algorithm ---------------------
    cudnnConvolutionFwdAlgo_t algo{};

    using search = SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
281
    algo = search::Find<T>(args, exhaustive_search, false, ctx);
L
liym27 已提交
282 283
    workspace_size = search::GetWorkspaceSize(args, algo);

284 285 286 287 288 289 290 291 292 293
#if CUDNN_VERSION_MIN(7, 0, 1)
    // when groups > 1, SearchAlgorithm find algo is CUDNN_CONVOLUTION_\
    // FWD_ALGO_WINOGRAD_NONFUSED, but this kind of algorithm is unstable
    // in forward computation, so change the algorithm to CUDNN_CONVOLUTION_\
    // FWD_ALGO_IMPLICIT_GEMM manually.
    if (ctx.Attr<int>("groups") > 1) {
      algo = static_cast<cudnnConvolutionFwdAlgo_t>(0);
    }
#endif

L
liym27 已提交
294
    // ------------------- cudnn conv forward ---------------------
295
    ScalingParamType<T> alpha = 1.0f;
296 297 298 299 300 301
    ScalingParamType<T> beta = 0.0f;

    // NOTE(zhiqiu): inplace addto is not supportted in double grad yet.
    // ScalingParamType<T> beta = ctx.Attr<bool>("use_addto") ? 1.0f : 0.0f;
    // VLOG(4) << "Conv: use_addto = " << ctx.Attr<bool>("use_addto");

L
liym27 已提交
302 303 304
    for (int i = 0; i < groups; i++) {
      workspace_handle.RunFunc(
          [&](void* workspace_ptr) {
305 306 307 308 309 310 311
            PADDLE_ENFORCE_CUDA_SUCCESS(
                platform::dynload::cudnnConvolutionForward(
                    handle, &alpha, args.idesc.desc(),
                    input_data + i * group_offset_in, args.wdesc.desc(),
                    filter_data + i * group_offset_filter, args.cdesc.desc(),
                    algo, workspace_ptr, workspace_size, &beta,
                    args.odesc.desc(), output_data + i * group_offset_out));
L
liym27 已提交
312 313 314 315
          },
          workspace_size);
    }

316
    if (channel_last && compute_format == DataLayout::kNCHW) {
L
liym27 已提交
317 318 319 320 321 322 323 324 325 326 327
      TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
          ctx, &transformed_output, output);
    }
  }
};

template <typename T>
class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
328 329 330
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        paddle::platform::errors::PreconditionNotMet("It must use CUDAPlace."));
L
liym27 已提交
331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348
    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"));

    if (input_grad) {
      input_grad->mutable_data<T>(ctx.GetPlace());
    }
    if (filter_grad) {
      filter_grad->mutable_data<T>(ctx.GetPlace());
    }

    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    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");
    int groups = ctx.Attr<int>("groups");
349

L
liym27 已提交
350 351 352
    bool exhaustive_search =
        FLAGS_cudnn_exhaustive_search || ctx.Attr<bool>("exhaustive_search");
    bool deterministic = FLAGS_cudnn_deterministic;
353 354 355 356 357 358
    auto exhaustive_deterministic = exhaustive_search && deterministic;
    PADDLE_ENFORCE_EQ(exhaustive_deterministic, false,
                      platform::errors::InvalidArgument(
                          "Cann't set exhaustive_search True and "
                          "FLAGS_cudnn_deterministic True at same time."));

L
liym27 已提交
359 360 361
    const std::string data_format = ctx.Attr<std::string>("data_format");
    const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");

362 363 364 365 366 367 368 369 370
    auto dtype = platform::CudnnDataType<T>::type;
    const bool compute_in_nhwc =
        dtype == CUDNN_DATA_HALF && IsVoltaOrLater(dev_ctx);
    auto compute_format =
        compute_in_nhwc && channel_last ? DataLayout::kNHWC : DataLayout::kNCHW;
    VLOG(3) << "Compute ConvGradOp with cuDNN:"
            << " data_format=" << data_format << " compute_format="
            << (compute_format == DataLayout::kNHWC ? "NHWC" : "NCHW");

L
liym27 已提交
371 372 373 374
    // transform Tensor
    Tensor transformed_input_channel(input->type());
    Tensor transformed_output_grad_channel(output_grad->type());
    Tensor transformed_input_grad_channel(input->type());
375 376
    Tensor transformed_filter_channel(filter->type());
    Tensor transformed_filter_grad_channel(filter->type());
L
liym27 已提交
377

378 379 380
    if (channel_last && compute_format == DataLayout::kNCHW) {
      VLOG(3) << "Transform input, output_grad, input_grad and tensor from "
                 "NHWC to NCHW.";
L
liym27 已提交
381 382 383 384 385 386 387 388 389 390 391 392 393
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, input, &transformed_input_channel);
      TransToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, input, &transformed_input_channel);

      ResizeToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, output_grad, &transformed_output_grad_channel);
      TransToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, output_grad, &transformed_output_grad_channel);

      if (input_grad) {
        ResizeToChannelFirst<platform::CUDADeviceContext, T>(
            ctx, input_grad, &transformed_input_grad_channel);
394 395 396 397 398 399
        // NOTE(zhiqiu): If inplace_addto strategy is enabled, we need to copy
        // the data of input_grad to transformed_input_grad_channel.
        if (ctx.Attr<bool>("use_addto")) {
          TransToChannelFirst<platform::CUDADeviceContext, T>(
              ctx, input_grad, &transformed_input_grad_channel);
        }
L
liym27 已提交
400 401
      }
    } else {
402 403
      transformed_input_channel.ShareDataWith(*input);
      transformed_output_grad_channel.ShareDataWith(*output_grad);
L
liym27 已提交
404 405 406 407 408
      if (input_grad) {
        transformed_input_grad_channel.ShareDataWith(*input_grad);
      }
    }

409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426
    if (compute_format == DataLayout::kNHWC) {
      VLOG(3) << "Transform filter and filter_grad tensor from NCHW to NHWC.";
      ResizeToChannelLast<platform::CUDADeviceContext, T>(
          ctx, filter, &transformed_filter_channel);
      TransToChannelLast<platform::CUDADeviceContext, T>(
          ctx, filter, &transformed_filter_channel);

      if (filter_grad) {
        ResizeToChannelLast<platform::CUDADeviceContext, T>(
            ctx, filter_grad, &transformed_filter_grad_channel);
      }
    } else {
      transformed_filter_channel.ShareDataWith(*filter);
      if (filter_grad) {
        transformed_filter_grad_channel.ShareDataWith(*filter_grad);
      }
    }

L
liym27 已提交
427 428
    //  update paddings
    auto in_dims = transformed_input_channel.dims();
429
    auto filter_dims = transformed_filter_channel.dims();
L
liym27 已提交
430
    framework::DDim in_data_dims;
431 432 433 434 435 436 437 438 439 440
    framework::DDim filter_data_dims;
    if (compute_format == DataLayout::kNCHW) {
      in_data_dims = framework::slice_ddim(in_dims, 2, in_dims.size());
      filter_data_dims =
          framework::slice_ddim(filter_dims, 2, filter_dims.size());
    } else {
      in_data_dims = framework::slice_ddim(in_dims, 1, in_dims.size() - 1);
      filter_data_dims =
          framework::slice_ddim(filter_dims, 1, filter_dims.size() - 1);
    }
L
liym27 已提交
441 442 443 444 445 446 447
    std::vector<int> ksize = framework::vectorize<int>(filter_data_dims);
    UpdatePaddingAndDilation(&paddings, &dilations, padding_algorithm,
                             in_data_dims, strides, ksize);

    // cuDNN only supports padding the same amount on every dimension.
    // So we create a new padded input tensor.
    int data_dim = strides.size();  // 2d or 3d
448
    bool is_sys_pad = math::IsSymmetricPadding(paddings, data_dim);
L
liym27 已提交
449 450 451 452 453 454 455 456 457 458
    Tensor transformed_input(input->type());
    Tensor transformed_input_grad(input->type());
    std::vector<int> padding_common(data_dim, 0);
    std::vector<int> input_pad(transformed_input_channel.dims().size() * 2, 0);

    if (!is_sys_pad) {
      // get pad
      std::vector<int> padding_diff(data_dim);
      std::vector<int> new_input_shape_vec(data_dim + 2);
      new_input_shape_vec[0] = transformed_input_channel.dims()[0];
459 460 461 462 463 464
      if (compute_format == DataLayout::kNCHW) {
        new_input_shape_vec[1] = transformed_input_channel.dims()[1];
      } else {
        new_input_shape_vec[data_dim + 1] =
            transformed_input_channel.dims()[data_dim + 1];
      }
L
liym27 已提交
465 466 467 468

      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]);
469 470 471 472 473 474 475 476 477 478 479 480 481 482
        if (compute_format == DataLayout::kNCHW) {
          new_input_shape_vec[i + 2] =
              transformed_input_channel.dims()[i + 2] + padding_diff[i];
        } else {
          new_input_shape_vec[i + 1] =
              transformed_input_channel.dims()[i + 1] + padding_diff[i];
        }
        if (compute_format == DataLayout::kNCHW) {
          input_pad[2 * i + 4] = paddings[2 * i] - padding_common[i];
          input_pad[2 * i + 4 + 1] = paddings[2 * i + 1] - padding_common[i];
        } else {
          input_pad[2 * i + 2] = paddings[2 * i] - padding_common[i];
          input_pad[2 * i + 2 + 1] = paddings[2 * i + 1] - padding_common[i];
        }
L
liym27 已提交
483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504
      }
      framework::DDim new_input_shape(
          framework::make_ddim(new_input_shape_vec));
      transformed_input.Resize(new_input_shape);

      transformed_input_grad.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);
      if (input_grad) {
        transformed_input_grad =
            ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
                new_input_shape, dev_ctx);
      }
      // pad for input
      const int rank = transformed_input_channel.dims().size();
      T pad_value(0.0);
      switch (rank) {
        case 4: {
505
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
L
liym27 已提交
506 507 508 509
              ctx, input_pad, transformed_input_channel, pad_value,
              &transformed_input);
        } break;
        case 5: {
510
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
L
liym27 已提交
511 512 513 514
              ctx, input_pad, transformed_input_channel, pad_value,
              &transformed_input);
        } break;
        default:
515 516
          PADDLE_THROW(platform::errors::InvalidArgument(
              "ConvOp only support tensors with 4 or 5 dimensions."));
L
liym27 已提交
517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535
      }
    } else {
      transformed_input.ShareDataWith(transformed_input_channel);
      if (input_grad) {
        transformed_input_grad.ShareDataWith(transformed_input_grad_channel);
      }
      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 = transformed_input.data<T>();
    const T* output_grad_data = transformed_output_grad_channel.data<T>();
536
    const T* filter_data = transformed_filter_channel.data<T>();
L
liym27 已提交
537 538 539 540 541
    T* filter_grad_data = nullptr;
    T* input_grad_data = nullptr;
    T* transformed_input_grad_data = nullptr;

    ConvArgs args1{&transformed_input_grad,
542
                   &transformed_filter_channel,
L
liym27 已提交
543 544 545
                   &transformed_output_grad_channel,
                   strides,
                   padding_common,
546 547
                   dilations,
                   dtype};
L
liym27 已提交
548
    ConvArgs args2{&transformed_input,
549
                   &transformed_filter_grad_channel,
L
liym27 已提交
550 551 552
                   &transformed_output_grad_channel,
                   strides,
                   padding_common,
553 554
                   dilations,
                   dtype};
L
liym27 已提交
555 556

    auto handle = dev_ctx.cudnn_handle();
557 558 559 560 561
    DataLayout layout = compute_format == DataLayout::kNHWC ? DataLayout::kNHWC
                                                            : DataLayout::kNCHW;
    if (transformed_input.dims().size() == 5) {
      layout = compute_format == DataLayout::kNHWC ? DataLayout::kNDHWC
                                                   : DataLayout::kNCDHW;
L
liym27 已提交
562 563 564 565 566 567
    }
    auto layout_tensor = GetCudnnTensorFormat(layout);
    auto workspace_handle = dev_ctx.cudnn_workspace_handle();

    int i_n, i_c, i_d, i_h, i_w;
    int o_n, o_c, o_d, o_h, o_w;
568 569 570 571 572 573 574 575 576 577 578
    if (compute_format == DataLayout::kNHWC) {
      GetNCDHW(transformed_input.dims(), DataLayout::kNHWC, &i_n, &i_c, &i_d,
               &i_h, &i_w);
      GetNCDHW(transformed_output_grad_channel.dims(), DataLayout::kNHWC, &o_n,
               &o_c, &o_d, &o_h, &o_w);
    } else {
      GetNCDHW(transformed_input.dims(), DataLayout::kNCHW, &i_n, &i_c, &i_d,
               &i_h, &i_w);
      GetNCDHW(transformed_output_grad_channel.dims(), DataLayout::kNCHW, &o_n,
               &o_c, &o_d, &o_h, &o_w);
    }
L
liym27 已提交
579 580 581

    int group_offset_in = i_c / groups * i_h * i_w * i_d;
    int group_offset_out = o_c / groups * o_h * o_w * o_d;
582
    int group_offset_filter = transformed_filter_channel.numel() / groups;
L
liym27 已提交
583 584 585 586 587 588
    // ------------------- cudnn backward algorithm ---------------------
    cudnnConvolutionBwdDataAlgo_t data_algo =
        static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
    cudnnConvolutionBwdFilterAlgo_t filter_algo =
        static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);
    size_t workspace_size = 0;
589 590
    int iwo_groups = groups;
    int c_groups = 1;
L
liym27 已提交
591 592 593 594 595 596 597 598 599 600 601 602

#if CUDNN_VERSION_MIN(7, 0, 1)
    iwo_groups = 1;
    c_groups = groups;
    groups = 1;
#endif

    if (input_grad) {
      // ------------------- cudnn descriptors ---------------------
      input_grad_data = input_grad->data<T>();
      transformed_input_grad_data = transformed_input_grad.data<T>();
      args1.handle = handle;
603 604 605
      args1.idesc.set(transformed_input_grad, layout_tensor);
      args1.wdesc.set(transformed_filter_channel, layout_tensor, iwo_groups);
      args1.odesc.set(transformed_output_grad_channel, layout_tensor);
L
liym27 已提交
606 607 608 609
      args1.cdesc.set(dtype, padding_common, strides, dilations, c_groups);

      using search1 = SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
      data_algo =
610
          search1::Find<T>(args1, exhaustive_search, deterministic, ctx);
L
liym27 已提交
611 612 613 614 615 616
      workspace_size =
          std::max(workspace_size, search1::GetWorkspaceSize(args1, data_algo));
    }

    if (filter_grad) {
      // ------------------- cudnn descriptors ---------------------
617
      filter_grad_data = transformed_filter_grad_channel.data<T>();
L
liym27 已提交
618
      args2.handle = handle;
619 620 621 622
      args2.idesc.set(transformed_input, layout_tensor);
      args2.wdesc.set(transformed_filter_grad_channel, layout_tensor,
                      iwo_groups);
      args2.odesc.set(transformed_output_grad_channel, layout_tensor);
L
liym27 已提交
623 624 625 626
      args2.cdesc.set(dtype, padding_common, strides, dilations, c_groups);

      using search2 = SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
      filter_algo =
627
          search2::Find<T>(args2, exhaustive_search, deterministic, ctx);
L
liym27 已提交
628 629 630 631 632
      workspace_size = std::max(workspace_size,
                                search2::GetWorkspaceSize(args2, filter_algo));
    }

    // ------------------- cudnn conv backward data ---------------------
633 634 635 636
    ScalingParamType<T> alpha = 1.0f;
    ScalingParamType<T> beta = ctx.Attr<bool>("use_addto") ? 1.0f : 0.0f;
    VLOG(4) << "Conv_grad: use_addto = " << ctx.Attr<bool>("use_addto");

L
liym27 已提交
637
    if (input_grad) {
638 639
      // When beta is 0, it is unnecessary to reset input_grad.
      // When beta is 1, the output cannot be reset since addt strategy used.
L
liym27 已提交
640 641 642
      for (int i = 0; i < groups; i++) {
        workspace_handle.RunFunc(
            [&](void* cudnn_workspace_ptr) {
643 644 645 646 647 648 649 650
              PADDLE_ENFORCE_CUDA_SUCCESS(
                  platform::dynload::cudnnConvolutionBackwardData(
                      handle, &alpha, args1.wdesc.desc(),
                      filter_data + i * group_offset_filter, args1.odesc.desc(),
                      output_grad_data + i * group_offset_out,
                      args1.cdesc.desc(), data_algo, cudnn_workspace_ptr,
                      workspace_size, &beta, args1.idesc.desc(),
                      transformed_input_grad_data + i * group_offset_in));
L
liym27 已提交
651 652 653 654
            },
            workspace_size);
      }

W
wangchaochaohu 已提交
655 656 657
      if (!is_sys_pad) {
        std::vector<int> starts(transformed_input_channel.dims().size(), 0);
        std::vector<int> axes(transformed_input_channel.dims().size(), 0);
L
liym27 已提交
658

W
wangchaochaohu 已提交
659 660 661 662
        for (size_t i = 0; i < transformed_input_channel.dims().size(); ++i) {
          starts[i] = input_pad[2 * i];
          axes[i] = i;
        }
L
liym27 已提交
663

W
wangchaochaohu 已提交
664 665
        transformed_input_grad_channel.mutable_data(ctx.GetPlace());
        if (transformed_input_channel.dims().size() == 4) {
666
          RemovePaddingSlice<paddle::platform::CUDADeviceContext, T, 4>(
W
wangchaochaohu 已提交
667 668 669
              ctx, &transformed_input_grad, &transformed_input_grad_channel,
              starts, axes);
        } else {
670
          RemovePaddingSlice<paddle::platform::CUDADeviceContext, T, 5>(
W
wangchaochaohu 已提交
671 672 673
              ctx, &transformed_input_grad, &transformed_input_grad_channel,
              starts, axes);
        }
L
liym27 已提交
674 675
      }

676
      if (channel_last && compute_format == DataLayout::kNCHW) {
L
liym27 已提交
677 678 679 680
        TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
            ctx, &transformed_input_grad_channel, input_grad);
      }
    }
681 682 683

    // filter_grad do not use inplace addto.
    ScalingParamType<T> beta_filter = 0.0f;
L
liym27 已提交
684 685 686 687 688 689
    // ------------------- cudnn conv backward filter ---------------------
    if (filter_grad) {
      // Because beta is zero, it is unnecessary to reset filter_grad.
      for (int i = 0; i < groups; i++) {
        workspace_handle.RunFunc(
            [&](void* cudnn_workspace_ptr) {
690 691 692 693 694 695
              PADDLE_ENFORCE_CUDA_SUCCESS(
                  platform::dynload::cudnnConvolutionBackwardFilter(
                      handle, &alpha, args2.idesc.desc(),
                      input_data + i * group_offset_in, args2.odesc.desc(),
                      output_grad_data + i * group_offset_out,
                      args2.cdesc.desc(), filter_algo, cudnn_workspace_ptr,
696
                      workspace_size, &beta_filter, args2.wdesc.desc(),
697
                      filter_grad_data + i * group_offset_filter));
L
liym27 已提交
698 699 700
            },
            workspace_size);
      }
701 702 703 704 705

      if (compute_format == DataLayout::kNHWC) {
        TransToChannelFirst<paddle::platform::CUDADeviceContext, T>(
            ctx, &transformed_filter_grad_channel, filter_grad);
      }
L
liym27 已提交
706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721
    }
  }
};

/*
 * Inputs:  I, W, dO, ddI, ddW
 * Outputs: ddO, dW, dI
 * ddo = conv(ddI, W) + conv(I, ddW)
 * dW = conv_bp_filter(ddI, dO)
 * dI = conv_bp_data(ddW, dO)
 */
template <typename T>
class CUDNNConvDoubleGradOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
722 723 724
    PADDLE_ENFORCE_EQ(
        platform::is_gpu_place(ctx.GetPlace()), true,
        paddle::platform::errors::PreconditionNotMet("It must use CUDAPlace."));
L
liym27 已提交
725 726 727 728 729 730 731 732 733 734 735
    auto X = ctx.Input<Tensor>("Input");
    auto W = ctx.Input<Tensor>("Filter");
    auto dO = ctx.Input<Tensor>("DOutput");
    auto ddX = ctx.Input<Tensor>("DDInput");
    auto ddW = ctx.Input<Tensor>("DDFilter");

    auto ddO = ctx.Output<Tensor>("DDOutput");
    auto dW = ctx.Output<Tensor>("DFilter");
    auto dX = ctx.Output<Tensor>("DInput");
    if (ddO) {
      ddO->mutable_data<T>(ctx.GetPlace());
L
lvmengsi 已提交
736 737
      math::SetConstant<platform::CUDADeviceContext, T> set_zero;
      set_zero(dev_ctx, ddO, static_cast<T>(0));
L
liym27 已提交
738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757
    }
    if (dW) {
      dW->mutable_data<T>(ctx.GetPlace());
    }
    if (dX) {
      dX->mutable_data<T>(ctx.GetPlace());
    }

    // const T* x = X->data<T>();
    const T* dy = dO->data<T>();
    const T* w = W->data<T>();

    const T* ddx = nullptr;
    const T* ddw = nullptr;
    T *dw, *dx, *ddy;
    dw = dx = ddy = nullptr;
    T* transformed_dx = nullptr;
    const std::vector<int>& strides = ctx.Attr<std::vector<int>>("strides");
    std::vector<int> dilations = ctx.Attr<std::vector<int>>("dilations");
    int groups = ctx.Attr<int>("groups");
758

L
liym27 已提交
759 760 761
    bool exhaustive_search =
        FLAGS_cudnn_exhaustive_search || ctx.Attr<bool>("exhaustive_search");
    bool deterministic = FLAGS_cudnn_deterministic;
762 763 764 765 766 767
    auto exhaustive_deterministic = exhaustive_search && deterministic;
    PADDLE_ENFORCE_EQ(exhaustive_deterministic, false,
                      platform::errors::InvalidArgument(
                          "Cann't set exhaustive_search True and "
                          "FLAGS_cudnn_deterministic True at same time."));

L
liym27 已提交
768 769 770 771 772 773 774 775 776
    std::vector<int> paddings = ctx.Attr<std::vector<int>>("paddings");

    std::string padding_algorithm = ctx.Attr<std::string>("padding_algorithm");
    const std::string data_format = ctx.Attr<std::string>("data_format");
    const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");

    // transform Tensors to channel first-----------
    Tensor transformed_X_channel(X->type());
    Tensor transformed_dO_channel(dO->type());
L
lvmengsi 已提交
777
    Tensor transformed_ddX_channel(X->type());
L
liym27 已提交
778 779 780 781 782 783 784 785 786 787 788 789 790 791 792

    Tensor transformed_ddO_channel(dO->type());
    Tensor transformed_dX_channel(X->type());

    if (channel_last) {
      ResizeToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, X, &transformed_X_channel);
      TransToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, X, &transformed_X_channel);

      ResizeToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, dO, &transformed_dO_channel);
      TransToChannelFirst<platform::CUDADeviceContext, T>(
          ctx, dO, &transformed_dO_channel);

L
lvmengsi 已提交
793 794 795 796 797 798
      if (ddX) {
        ResizeToChannelFirst<platform::CUDADeviceContext, T>(
            ctx, ddX, &transformed_ddX_channel);
        TransToChannelFirst<platform::CUDADeviceContext, T>(
            ctx, ddX, &transformed_ddX_channel);
      }
L
liym27 已提交
799 800 801 802 803 804 805 806 807 808 809 810 811 812

      if (ddO) {
        ResizeToChannelFirst<platform::CUDADeviceContext, T>(
            ctx, ddO, &transformed_ddO_channel);
      }
      if (dX) {
        ResizeToChannelFirst<platform::CUDADeviceContext, T>(
            ctx, dX, &transformed_dX_channel);
        transformed_dX_channel.mutable_data<T>(ctx.GetPlace());
      }

    } else {
      transformed_X_channel = *X;
      transformed_dO_channel = *dO;
L
lvmengsi 已提交
813 814 815
      if (ddX) {
        transformed_ddX_channel = *ddX;
      }
L
liym27 已提交
816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834
      if (ddO) {
        transformed_ddO_channel.ShareDataWith(*ddO);
      }
      if (dX) {
        transformed_dX_channel.ShareDataWith(*dX);
      }
    }

    auto in_dims = transformed_X_channel.dims();
    auto filter_dims = W->dims();
    framework::DDim 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
835
    bool is_sys_pad = math::IsSymmetricPadding(paddings, data_dim);
L
liym27 已提交
836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867
    Tensor transformed_X(X->type());
    Tensor transformed_ddX(X->type());

    Tensor transformed_dX(X->type());

    std::vector<int> padding_common(data_dim, 0);
    std::vector<int> input_pad(X->dims().size() * 2, 0);

    if (!is_sys_pad) {
      // get pad
      std::vector<int> padding_diff(data_dim);
      std::vector<int> new_input_shape_vec(data_dim + 2);
      new_input_shape_vec[0] = transformed_X_channel.dims()[0];
      new_input_shape_vec[1] = transformed_X_channel.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] =
            transformed_X_channel.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_X.Resize(new_input_shape);
      transformed_ddX.Resize(new_input_shape);
      transformed_dX.Resize(new_input_shape);

      transformed_X =
          ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
              new_input_shape, dev_ctx);
L
lvmengsi 已提交
868 869 870 871 872
      if (ddX) {
        transformed_ddX =
            ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
                new_input_shape, dev_ctx);
      }
L
liym27 已提交
873 874 875 876 877 878 879 880 881 882 883
      if (dX) {
        transformed_dX =
            ctx.AllocateTmpTensor<T, paddle::platform::CUDADeviceContext>(
                new_input_shape, dev_ctx);
      }

      // pad for input
      const int rank = X->dims().size();
      T pad_value(0.0);
      switch (rank) {
        case 4: {
884
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
L
liym27 已提交
885
              ctx, input_pad, transformed_X_channel, pad_value, &transformed_X);
L
lvmengsi 已提交
886 887 888 889 890
          if (ddX) {
            math::PadFunction<paddle::platform::CUDADeviceContext, T, 4>(
                ctx, input_pad, transformed_ddX_channel, pad_value,
                &transformed_ddX);
          }
L
liym27 已提交
891 892
        } break;
        case 5: {
893
          math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
L
liym27 已提交
894
              ctx, input_pad, transformed_X_channel, pad_value, &transformed_X);
L
lvmengsi 已提交
895 896 897 898 899
          if (ddX) {
            math::PadFunction<paddle::platform::CUDADeviceContext, T, 5>(
                ctx, input_pad, transformed_ddX_channel, pad_value,
                &transformed_ddX);
          }
L
liym27 已提交
900 901
        } break;
        default:
902 903
          PADDLE_THROW(platform::errors::InvalidArgument(
              "ConvOp only support tensors with 4 or 5 dimensions."));
L
liym27 已提交
904 905 906 907
      }

    } else {
      transformed_X.ShareDataWith(transformed_X_channel);
L
lvmengsi 已提交
908 909 910
      if (ddX) {
        transformed_ddX.ShareDataWith(transformed_ddX_channel);
      }
L
liym27 已提交
911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932
      if (dX) {
        transformed_dX.ShareDataWith(transformed_dX_channel);
      }

      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* x = transformed_X.data<T>();

    int iwo_group = groups;
    int c_group = 1;
#if CUDNN_VERSION_MIN(7, 0, 1)
    iwo_group = 1;
    c_group = groups;
933
    groups = 1;
L
liym27 已提交
934 935 936 937 938
#endif
    auto dtype = platform::CudnnDataType<T>::type;

    auto handle = dev_ctx.cudnn_handle();

939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958
    ConvArgs args1{&transformed_ddX,
                   W,
                   &transformed_ddO_channel,
                   strides,
                   padding_common,
                   dilations,
                   dtype};
    ConvArgs args2{
        &transformed_X, ddW,  &transformed_ddO_channel, strides, padding_common,
        dilations,      dtype};
    ConvArgs args3{&transformed_ddX,
                   dW,
                   &transformed_dO_channel,
                   strides,
                   padding_common,
                   dilations,
                   dtype};
    ConvArgs args4{
        &transformed_dX, ddW,  &transformed_dO_channel, strides, padding_common,
        dilations,       dtype};
L
liym27 已提交
959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985

    cudnnConvolutionFwdAlgo_t fwd_algo1 =
        static_cast<cudnnConvolutionFwdAlgo_t>(0);
    cudnnConvolutionFwdAlgo_t fwd_algo2 =
        static_cast<cudnnConvolutionFwdAlgo_t>(0);
    cudnnConvolutionBwdDataAlgo_t data_algo =
        static_cast<cudnnConvolutionBwdDataAlgo_t>(0);
    cudnnConvolutionBwdFilterAlgo_t filter_algo =
        static_cast<cudnnConvolutionBwdFilterAlgo_t>(0);

    auto layout = GetCudnnTensorFormat(DataLayout::kNCHW);

    // ddo = conv(ddI, W) + conv(I, ddW)
    size_t workspace_size = 0;

    T* transformed_ddy_channel = nullptr;
    if (ddO) {
      ddy = ddO->data<T>();
      transformed_ddy_channel = transformed_ddO_channel.data<T>();
      if (ddX) {
        args1.handle = handle;
        args1.idesc.set(transformed_ddX, iwo_group);
        args1.wdesc.set(*W, layout, iwo_group);
        args1.odesc.set(transformed_ddO_channel, iwo_group);
        args1.cdesc.set(dtype, padding_common, strides, dilations, c_group);

        using search1 = SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
986
        fwd_algo1 = search1::Find<T>(args1, exhaustive_search, false, ctx);
L
liym27 已提交
987 988 989 990 991 992 993 994 995 996 997 998 999 1000
        workspace_size = search1::GetWorkspaceSize(args1, fwd_algo1);
      }

      if (ddW) {
        ddw = ddW->data<T>();
        args2.handle = handle;
        args2.idesc.set(transformed_X, iwo_group);

        args2.wdesc.set(*ddW, layout, iwo_group);

        args2.odesc.set(transformed_ddO_channel, iwo_group);
        args2.cdesc.set(dtype, padding_common, strides, dilations, c_group);

        using search2 = SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
1001
        fwd_algo2 = search2::Find<T>(args2, exhaustive_search, false, ctx);
L
liym27 已提交
1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018
        workspace_size = std::max(workspace_size,
                                  search2::GetWorkspaceSize(args2, fwd_algo2));
      }
    }

    if (dW && ddX) {
      dw = dW->data<T>();
      args3.handle = handle;
      args3.idesc.set(transformed_ddX, iwo_group);
      args3.wdesc.set(*dW, layout, iwo_group);

      args3.odesc.set(transformed_dO_channel, iwo_group);

      args3.cdesc.set(dtype, padding_common, strides, dilations, c_group);

      using search3 = SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
      filter_algo =
1019
          search3::Find<T>(args3, exhaustive_search, deterministic, ctx);
L
liym27 已提交
1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034
      workspace_size = std::max(workspace_size,
                                search3::GetWorkspaceSize(args3, filter_algo));
    }

    if (ddW && dX) {
      transformed_dx = transformed_dX.data<T>();

      args4.handle = handle;
      args4.idesc.set(transformed_dX, iwo_group);
      args4.wdesc.set(*ddW, layout, iwo_group);
      args4.odesc.set(transformed_dO_channel, iwo_group);
      args4.cdesc.set(dtype, padding_common, strides, dilations, c_group);

      using search4 = SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
      data_algo =
1035
          search4::Find<T>(args4, exhaustive_search, deterministic, ctx);
L
liym27 已提交
1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051
      workspace_size =
          std::max(workspace_size, search4::GetWorkspaceSize(args4, data_algo));
    }

    int i_n, i_c, i_d, i_h, i_w;
    GetNCDHW(transformed_X.dims(), DataLayout::kNCHW, &i_n, &i_c, &i_d, &i_h,
             &i_w);

    int o_n, o_c, o_d, o_h, o_w;
    GetNCDHW(transformed_dO_channel.dims(), DataLayout::kNCHW, &o_n, &o_c, &o_d,
             &o_h, &o_w);

    int group_offset_in = i_c / groups * i_h * i_w * i_d;
    int group_offset_out = o_c / groups * o_h * o_w * o_d;
    int group_offset_filter = W->numel() / groups;

1052 1053 1054 1055 1056 1057 1058 1059
    ScalingParamType<T> alpha = 1.0f;
    ScalingParamType<T> beta = 0.0f;

    // NOTE(zhiqiu): inplace addto is not supportted in double grad yet.
    // ScalingParamType<T> beta = ctx.Attr<bool>("use_addto") ? 1.0f :
    // 0.0f;
    // VLOG(4) << "Conv_grad_grad: use_addto = " << ctx.Attr<bool>("use_addto");

L
liym27 已提交
1060 1061 1062 1063 1064 1065 1066 1067
    auto wkspace_handle = dev_ctx.cudnn_workspace_handle();

    if (ddO) {
      if (ddX) {
        ddx = transformed_ddX.data<T>();
        for (int i = 0; i < groups; i++) {
          wkspace_handle.RunFunc(
              [&](void* workspace_ptr) {
1068 1069 1070 1071 1072 1073 1074 1075
                PADDLE_ENFORCE_CUDA_SUCCESS(
                    platform::dynload::cudnnConvolutionForward(
                        handle, &alpha, args1.idesc.desc(),
                        ddx + i * group_offset_in, args1.wdesc.desc(),
                        w + i * group_offset_filter, args1.cdesc.desc(),
                        fwd_algo1, workspace_ptr, workspace_size, &beta,
                        args1.odesc.desc(),
                        transformed_ddy_channel + i * group_offset_out));
L
liym27 已提交
1076 1077 1078 1079 1080 1081 1082 1083
              },
              workspace_size);
        }
      }
      if (ddW) {
        for (int i = 0; i < groups; i++) {
          wkspace_handle.RunFunc(
              [&](void* workspace_ptr) {
1084 1085 1086 1087 1088 1089 1090 1091
                PADDLE_ENFORCE_CUDA_SUCCESS(
                    platform::dynload::cudnnConvolutionForward(
                        handle, &alpha, args2.idesc.desc(),
                        x + i * group_offset_in, args2.wdesc.desc(),
                        ddw + i * group_offset_filter, args2.cdesc.desc(),
                        fwd_algo2, workspace_ptr, workspace_size, &alpha,
                        args2.odesc.desc(),
                        transformed_ddy_channel + i * group_offset_out));
L
liym27 已提交
1092 1093 1094 1095 1096 1097 1098 1099 1100
              },
              workspace_size);
        }
      }
      if (channel_last) {
        TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
            ctx, &transformed_ddO_channel, ddO);
      }
    }
L
lvmengsi 已提交
1101
    T* transformed_dy_channel = transformed_dO_channel.data<T>();
L
liym27 已提交
1102 1103 1104 1105 1106
    if (dW && ddX) {
      ddx = transformed_ddX.data<T>();
      for (int i = 0; i < groups; i++) {
        wkspace_handle.RunFunc(
            [&](void* workspace_ptr) {
1107 1108 1109 1110 1111 1112 1113 1114
              PADDLE_ENFORCE_CUDA_SUCCESS(
                  platform::dynload::cudnnConvolutionBackwardFilter(
                      handle, &alpha, args3.idesc.desc(),
                      ddx + i * group_offset_in, args3.odesc.desc(),
                      transformed_dy_channel + i * group_offset_out,
                      args3.cdesc.desc(), filter_algo, workspace_ptr,
                      workspace_size, &beta, args3.wdesc.desc(),
                      dw + i * group_offset_filter));
L
liym27 已提交
1115 1116 1117 1118 1119 1120 1121 1122 1123 1124
            },
            workspace_size);
      }
    }

    if (dX && ddW) {
      ddw = ddW->data<T>();
      for (int i = 0; i < groups; i++) {
        wkspace_handle.RunFunc(
            [&](void* workspace_ptr) {
1125 1126 1127 1128 1129 1130 1131 1132
              PADDLE_ENFORCE_CUDA_SUCCESS(
                  platform::dynload::cudnnConvolutionBackwardData(
                      handle, &alpha, args4.wdesc.desc(),
                      ddw + i * group_offset_filter, args4.odesc.desc(),
                      transformed_dy_channel + i * group_offset_out,
                      args4.cdesc.desc(), data_algo, workspace_ptr,
                      workspace_size, &beta, args4.idesc.desc(),
                      transformed_dx + i * group_offset_in));
L
liym27 已提交
1133 1134 1135 1136
            },
            workspace_size);
      }

W
wangchaochaohu 已提交
1137 1138 1139 1140
      if (!is_sys_pad) {
        // reverse padded input
        std::vector<int> starts(X->dims().size(), 0);
        std::vector<int> axes(X->dims().size(), 0);
L
liym27 已提交
1141

W
wangchaochaohu 已提交
1142 1143 1144 1145 1146
        for (size_t i = 0; i < X->dims().size(); ++i) {
          starts[i] = input_pad[2 * i];
          axes[i] = i;
        }
        if (X->dims().size() == 4) {
1147
          RemovePaddingSlice<paddle::platform::CUDADeviceContext, T, 4>(
W
wangchaochaohu 已提交
1148 1149
              ctx, &transformed_dX, &transformed_dX_channel, starts, axes);
        } else {
1150
          RemovePaddingSlice<paddle::platform::CUDADeviceContext, T, 5>(
W
wangchaochaohu 已提交
1151 1152
              ctx, &transformed_dX, &transformed_dX_channel, starts, axes);
        }
L
liym27 已提交
1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191
      }
      if (channel_last) {
        TransToChannelLast<paddle::platform::CUDADeviceContext, T>(
            ctx, &transformed_dX_channel, dX);
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace plat = paddle::platform;
REGISTER_OP_KERNEL(conv2d, CUDNN, plat::CUDAPlace,
                   paddle::operators::CUDNNConvOpKernel<float>,
                   paddle::operators::CUDNNConvOpKernel<double>,
                   paddle::operators::CUDNNConvOpKernel<plat::float16>);
REGISTER_OP_KERNEL(conv2d_grad, CUDNN, plat::CUDAPlace,
                   paddle::operators::CUDNNConvGradOpKernel<float>,
                   paddle::operators::CUDNNConvGradOpKernel<double>,
                   paddle::operators::CUDNNConvGradOpKernel<plat::float16>);
REGISTER_OP_KERNEL(
    conv2d_grad_grad, CUDNN, plat::CUDAPlace,
    paddle::operators::CUDNNConvDoubleGradOpKernel<float>,
    paddle::operators::CUDNNConvDoubleGradOpKernel<double>,
    paddle::operators::CUDNNConvDoubleGradOpKernel<plat::float16>);

REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace,
                   paddle::operators::CUDNNConvOpKernel<float>,
                   paddle::operators::CUDNNConvOpKernel<double>,
                   paddle::operators::CUDNNConvOpKernel<plat::float16>);
REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace,
                   paddle::operators::CUDNNConvGradOpKernel<float>,
                   paddle::operators::CUDNNConvGradOpKernel<double>);
REGISTER_OP_KERNEL(
    conv3d_grad_grad, CUDNN, plat::CUDAPlace,
    paddle::operators::CUDNNConvDoubleGradOpKernel<float>,
    paddle::operators::CUDNNConvDoubleGradOpKernel<double>,
    paddle::operators::CUDNNConvDoubleGradOpKernel<plat::float16>);