conv_grad_grad_kernel.cu 27.1 KB
Newer Older
H
hong 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2022 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.

15 16
#include "paddle/phi/kernels/conv_grad_grad_kernel.h"

17
#include "paddle/fluid/framework/eigen.h"
H
hong 已提交
18
#include "paddle/phi/backends/gpu/gpu_context.h"
19
#include "paddle/phi/core/dense_tensor.h"
H
hong 已提交
20 21
#include "paddle/phi/core/kernel_registry.h"
#ifdef PADDLE_WITH_HIP
22
#include "paddle/phi/kernels/gpudnn/conv_miopen_helper.h"
H
hong 已提交
23
#else
24
#include "paddle/phi/kernels/gpudnn/conv_cudnn_v7.h"
H
hong 已提交
25 26 27 28 29 30 31
#endif

#include "paddle/fluid/platform/cudnn_workspace_helper.h"
#include "paddle/fluid/platform/float16.h"
#include "paddle/fluid/platform/profiler.h"
#include "paddle/phi/common/bfloat16.h"
#include "paddle/phi/common/float16.h"
32 33
#include "paddle/phi/kernels/cpu/conv_util.h"
#include "paddle/phi/kernels/funcs/batch_norm_utils.h"
H
hong 已提交
34
#include "paddle/phi/kernels/funcs/math_function.h"
35 36
#include "paddle/phi/kernels/funcs/padding.h"
#include "paddle/phi/kernels/impl/conv_cudnn_impl.h"
H
hong 已提交
37 38 39 40 41 42 43 44

namespace phi {

template <typename T, typename Context>
void ConvCudnnGradGradKernel(
    const Context& ctx,
    const DenseTensor& input,
    const DenseTensor& filter,
45
    const DenseTensor& out_grad,
46 47
    const paddle::optional<DenseTensor>& input_grad_grad,
    const paddle::optional<DenseTensor>& filter_grad_grad,
H
hong 已提交
48 49 50 51 52 53 54 55 56 57
    const std::vector<int>& strides,
    const std::vector<int>& paddings_t,
    const std::string& padding_algorithm,
    int groups,
    const std::vector<int>& dilations_t,
    const std::string& data_format,
    bool use_addto,
    int workspace_size_MB,
    bool exhaustive_search_t,
    DenseTensor* input_grad,
58 59
    DenseTensor* filter_grad,
    DenseTensor* out_grad_grad) {
H
hong 已提交
60 61 62 63 64 65 66 67 68 69
  auto X = &input;
  auto W = &filter;
  auto dO = &out_grad;
  auto ddX = input_grad_grad.get_ptr();
  auto ddW = filter_grad_grad.get_ptr();

  auto ddO = out_grad_grad;
  auto dW = filter_grad;
  auto dX = input_grad;
  if (ddO) {
H
hong 已提交
70
    ctx.template Alloc<T>(ddO);
H
hong 已提交
71 72 73 74
    phi::funcs::SetConstant<Context, T> set_zero;
    set_zero(ctx, ddO, static_cast<T>(0));
  }
  if (dW) {
H
hong 已提交
75
    ctx.template Alloc<T>(dW);
H
hong 已提交
76 77
  }
  if (dX) {
H
hong 已提交
78
    ctx.template Alloc<T>(dX);
H
hong 已提交
79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129
  }

  // 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;
  std::vector<int> dilations = dilations_t;

  bool exhaustive_search = FLAGS_cudnn_exhaustive_search || exhaustive_search_t;
  bool deterministic = FLAGS_cudnn_deterministic;
  auto exhaustive_deterministic = exhaustive_search && deterministic;
  PADDLE_ENFORCE_EQ(exhaustive_deterministic,
                    false,
                    phi::errors::InvalidArgument(
                        "Cann't set exhaustive_search True and "
                        "FLAGS_cudnn_deterministic True at same time."));

  std::vector<int> paddings = paddings_t;

  const bool channel_last = (data_format == "NHWC" || data_format == "NDHWC");

  // transform Tensors to channel first-----------
  DenseTensor transformed_X_channel(X->type());
  DenseTensor transformed_dO_channel(dO->type());
  DenseTensor transformed_ddX_channel(X->type());

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

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

    ResizeToChannelFirst<Context, T>(ctx, dO, &transformed_dO_channel);
    TransToChannelFirst<Context, T>(ctx, dO, &transformed_dO_channel);

    if (ddX) {
      ResizeToChannelFirst<Context, T>(ctx, ddX, &transformed_ddX_channel);
      TransToChannelFirst<Context, T>(ctx, ddX, &transformed_ddX_channel);
    }

    if (ddO) {
      ResizeToChannelFirst<Context, T>(ctx, ddO, &transformed_ddO_channel);
    }
    if (dX) {
      ResizeToChannelFirst<Context, T>(ctx, dX, &transformed_dX_channel);
H
hong 已提交
130
      ctx.template Alloc<T>(&transformed_dX_channel);
H
hong 已提交
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
    }

  } else {
    transformed_X_channel = *X;
    transformed_dO_channel = *dO;
    if (ddX) {
      transformed_ddX_channel = *ddX;
    }
    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();
  DDim in_data_dims = slice_ddim(in_dims, 2, in_dims.size());
  DDim filter_data_dims = slice_ddim(filter_dims, 2, filter_dims.size());
  std::vector<int> ksize = vectorize<int>(filter_data_dims);
  UpdatePaddingAndDilation(
      &paddings, &dilations, padding_algorithm, in_data_dims, strides, ksize);

  int data_dim = strides.size();  // 2d or 3d
  bool is_sys_pad = funcs::IsSymmetricPadding(paddings, data_dim);
  DenseTensor transformed_X(X->type());
  DenseTensor transformed_ddX(X->type());

  DenseTensor 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];
    }
    DDim new_input_shape(make_ddim(new_input_shape_vec));
    transformed_X.Resize(new_input_shape);
    transformed_ddX.Resize(new_input_shape);
    transformed_dX.Resize(new_input_shape);

H
hong 已提交
185
    ctx.template Alloc<T>(&transformed_X);
H
hong 已提交
186 187

    if (ddX) {
H
hong 已提交
188
      ctx.template Alloc<T>(&transformed_ddX);
H
hong 已提交
189 190
    }
    if (dX) {
H
hong 已提交
191
      ctx.template Alloc<T>(&transformed_dX);
H
hong 已提交
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
    }

    // pad for input
    const int rank = X->dims().size();
    T pad_value(0.0);
    switch (rank) {
      case 4: {
        funcs::PadFunction<Context, T, 4>(
            ctx, input_pad, transformed_X_channel, pad_value, &transformed_X);
        if (ddX) {
          funcs::PadFunction<Context, T, 4>(ctx,
                                            input_pad,
                                            transformed_ddX_channel,
                                            pad_value,
                                            &transformed_ddX);
        }
      } break;
      case 5: {
        funcs::PadFunction<Context, T, 5>(
            ctx, input_pad, transformed_X_channel, pad_value, &transformed_X);
        if (ddX) {
          funcs::PadFunction<Context, T, 5>(ctx,
                                            input_pad,
                                            transformed_ddX_channel,
                                            pad_value,
                                            &transformed_ddX);
        }
      } break;
      default:
        PADDLE_THROW(phi::errors::InvalidArgument(
            "ConvOp only support tensors with 4 or 5 dimensions."));
    }

  } else {
    transformed_X.ShareDataWith(transformed_X_channel);
    if (ddX) {
      transformed_ddX.ShareDataWith(transformed_ddX_channel);
    }
    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 defined(PADDLE_WITH_HIP) || CUDNN_VERSION_MIN(7, 0, 1)
  iwo_group = 1;
  c_group = groups;
  groups = 1;
#endif
  auto dtype = paddle::platform::CudnnDataType<T>::type;

  auto handle = ctx.cudnn_handle();
H
hong 已提交
257 258
  auto layout = paddle::platform::GetCudnnTensorFormat(
      paddle::platform::DataLayout::kNCHW);
H
hong 已提交
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
  ConvArgs args1{&transformed_ddX,
                 W,
                 &transformed_ddO_channel,
                 strides,
                 padding_common,
                 dilations,
                 dtype,
                 groups,
                 paddle::platform::DataLayout::kNCHW};
  ConvArgs args2{&transformed_X,
                 ddW,
                 &transformed_ddO_channel,
                 strides,
                 padding_common,
                 dilations,
                 dtype,
                 groups,
                 paddle::platform::DataLayout::kNCHW};
  ConvArgs args3{&transformed_ddX,
                 dW,
                 &transformed_dO_channel,
                 strides,
                 padding_common,
                 dilations,
                 dtype,
                 groups,
                 paddle::platform::DataLayout::kNCHW};
  ConvArgs args4{&transformed_dX,
                 ddW,
                 &transformed_dO_channel,
                 strides,
                 padding_common,
                 dilations,
                 dtype,
                 groups,
                 paddle::platform::DataLayout::kNCHW};
H
hong 已提交
296 297

#ifdef PADDLE_WITH_HIP
298 299 300 301
  SearchResult<miopenConvFwdAlgorithm_t> fwd_result1;
  SearchResult<miopenConvFwdAlgorithm_t> fwd_result2;
  SearchResult<miopenConvBwdDataAlgorithm_t> data_result;
  SearchResult<miopenConvBwdWeightsAlgorithm_t> filter_result;
H
hong 已提交
302
#else
303 304 305 306
  SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result1;
  SearchResult<cudnnConvolutionFwdAlgo_t> fwd_result2;
  SearchResult<cudnnConvolutionBwdDataAlgo_t> data_result;
  SearchResult<cudnnConvolutionBwdFilterAlgo_t> filter_result;
H
hong 已提交
307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328
#endif

  // 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,
                      paddle::platform::AllowTF32Cudnn(),
                      c_group);

#ifdef PADDLE_WITH_HIP
329
      using search1 = SearchAlgorithm<miopenConvFwdAlgorithm_t>;
H
hong 已提交
330
      workspace_size = search1::GetWorkspaceSize(args1);
331
      fwd_result1.algo = search1::Find<T>(
H
hong 已提交
332 333
          args1, exhaustive_search, false, workspace_size, ctx);
#else
334
      using search1 = SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
335
      fwd_result1 = search1::Find<T>(ctx, args1, exhaustive_search, false);
336
      workspace_size = search1::GetWorkspaceSize(args1, fwd_result1.algo);
H
hong 已提交
337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353
#endif
    }

    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,
                      paddle::platform::AllowTF32Cudnn(),
                      c_group);

#ifdef PADDLE_WITH_HIP
354
      using search2 = SearchAlgorithm<miopenConvFwdAlgorithm_t>;
H
hong 已提交
355 356
      workspace_size =
          std::max(workspace_size, search2::GetWorkspaceSize(args2));
357
      fwd_result2.algo = search2::Find<T>(
H
hong 已提交
358 359
          args2, exhaustive_search, false, workspace_size, ctx);
#else
360
      using search2 = SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t>;
361
      fwd_result2 = search2::Find<T>(ctx, args2, exhaustive_search, false);
362 363
      workspace_size = std::max(
          workspace_size, search2::GetWorkspaceSize(args2, fwd_result2.algo));
H
hong 已提交
364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381
#endif
    }
  }

  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,
                    paddle::platform::AllowTF32Cudnn(),
                    c_group);

#ifdef PADDLE_WITH_HIP
382
    using search3 = SearchAlgorithm<miopenConvBwdWeightsAlgorithm_t>;
H
hong 已提交
383
    workspace_size = std::max(workspace_size, search3::GetWorkspaceSize(args3));
384
    filter_result.algo = search3::Find<T>(
H
hong 已提交
385 386
        args3, exhaustive_search, deterministic, workspace_size, ctx);
#else
387
    using search3 = SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t>;
388
    filter_result =
389
        search3::Find<T>(ctx, args3, exhaustive_search, deterministic);
390 391
    workspace_size = std::max(
        workspace_size, search3::GetWorkspaceSize(args3, filter_result.algo));
H
hong 已提交
392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409
#endif
  }

  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,
                    paddle::platform::AllowTF32Cudnn(),
                    c_group);

#ifdef PADDLE_WITH_HIP
410
    using search4 = SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
H
hong 已提交
411
    workspace_size = std::max(workspace_size, search4::GetWorkspaceSize(args4));
412
    data_result.algo = search4::Find<T>(
H
hong 已提交
413 414
        args4, exhaustive_search, deterministic, workspace_size, ctx);
#else
415
    using search4 = SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t>;
416
    data_result =
417
        search4::Find<T>(ctx, args4, exhaustive_search, deterministic);
418 419
    workspace_size = std::max(
        workspace_size, search4::GetWorkspaceSize(args4, data_result.algo));
H
hong 已提交
420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439
#endif
  }

  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;

440 441
  ScalingParamType<T> alpha = 1.0f;
  ScalingParamType<T> beta = 0.0f;
H
hong 已提交
442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463

  // 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");
  auto wkspace_handle = ctx.cudnn_workspace_handle();

  if (ddO) {
    if (ddX) {
      ddx = transformed_ddX.data<T>();
#ifdef PADDLE_WITH_HIP
      wkspace_handle.RunFunc(
          [&](void* workspace_ptr) {
            PADDLE_ENFORCE_GPU_SUCCESS(
                paddle::platform::dynload::miopenConvolutionForward(
                    handle,
                    &alpha,
                    args1.idesc.desc(),
                    ddx,
                    args1.wdesc.desc(),
                    w,
                    args1.cdesc.desc(),
464
                    fwd_result1.algo,
H
hong 已提交
465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484
                    &beta,
                    args1.odesc.desc(),
                    transformed_ddy_channel,
                    workspace_ptr,
                    workspace_size));
          },
          workspace_size);
#else
      for (int i = 0; i < groups; i++) {
        wkspace_handle.RunFunc(
            [&](void* workspace_ptr) {
              PADDLE_ENFORCE_GPU_SUCCESS(
                  paddle::platform::dynload::cudnnConvolutionForward(
                      handle,
                      &alpha,
                      args1.idesc.desc(),
                      ddx + i * group_offset_in,
                      args1.wdesc.desc(),
                      w + i * group_offset_filter,
                      args1.cdesc.desc(),
485
                      fwd_result1.algo,
H
hong 已提交
486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509
                      workspace_ptr,
                      workspace_size,
                      &beta,
                      args1.odesc.desc(),
                      transformed_ddy_channel + i * group_offset_out));
            },
            workspace_size);
      }
#endif
    }
    if (ddW) {
#ifdef PADDLE_WITH_HIP
      // MIOPEN ONLY support beta to be 0.0f
      wkspace_handle.RunFunc(
          [&](void* workspace_ptr) {
            PADDLE_ENFORCE_GPU_SUCCESS(
                paddle::platform::dynload::miopenConvolutionForward(
                    handle,
                    &alpha,
                    args2.idesc.desc(),
                    x,
                    args2.wdesc.desc(),
                    ddw,
                    args2.cdesc.desc(),
510
                    fwd_result2.algo,
H
hong 已提交
511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530
                    &beta,
                    args2.odesc.desc(),
                    transformed_ddy_channel,
                    workspace_ptr,
                    workspace_size));
          },
          workspace_size);
#else
      for (int i = 0; i < groups; i++) {
        wkspace_handle.RunFunc(
            [&](void* workspace_ptr) {
              PADDLE_ENFORCE_GPU_SUCCESS(
                  paddle::platform::dynload::cudnnConvolutionForward(
                      handle,
                      &alpha,
                      args2.idesc.desc(),
                      x + i * group_offset_in,
                      args2.wdesc.desc(),
                      ddw + i * group_offset_filter,
                      args2.cdesc.desc(),
531
                      fwd_result2.algo,
H
hong 已提交
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
                      workspace_ptr,
                      workspace_size,
                      &alpha,
                      args2.odesc.desc(),
                      transformed_ddy_channel + i * group_offset_out));
            },
            workspace_size);
      }
#endif
    }
    if (channel_last) {
      TransToChannelLast<Context, T>(ctx, &transformed_ddO_channel, ddO);
    }
  }
  T* transformed_dy_channel = transformed_dO_channel.data<T>();
  if (dW && ddX) {
    ddx = transformed_ddX.data<T>();
#ifdef PADDLE_WITH_HIP
    wkspace_handle.RunFunc(
        [&](void* workspace_ptr) {
          PADDLE_ENFORCE_GPU_SUCCESS(
              paddle::platform::dynload::miopenConvolutionBackwardWeights(
                  handle,
                  &alpha,
                  args3.odesc.desc(),
                  transformed_dy_channel,
                  args3.idesc.desc(),
                  ddx,
                  args3.cdesc.desc(),
561
                  filter_result.algo,
H
hong 已提交
562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581
                  &beta,
                  args3.wdesc.desc(),
                  dw,
                  workspace_ptr,
                  workspace_size));
        },
        workspace_size);
#else
    for (int i = 0; i < groups; i++) {
      wkspace_handle.RunFunc(
          [&](void* workspace_ptr) {
            PADDLE_ENFORCE_GPU_SUCCESS(
                paddle::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(),
582
                    filter_result.algo,
H
hong 已提交
583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
                    workspace_ptr,
                    workspace_size,
                    &beta,
                    args3.wdesc.desc(),
                    dw + i * group_offset_filter));
          },
          workspace_size);
    }
#endif
  }

  if (dX && ddW) {
    ddw = ddW->data<T>();
#ifdef PADDLE_WITH_HIP
    wkspace_handle.RunFunc(
        [&](void* workspace_ptr) {
          PADDLE_ENFORCE_GPU_SUCCESS(
              paddle::platform::dynload::miopenConvolutionBackwardData(
                  handle,
                  &alpha,
                  args4.odesc.desc(),
                  transformed_dy_channel,
                  args4.wdesc.desc(),
                  ddw,
                  args4.cdesc.desc(),
608
                  data_result.algo,
H
hong 已提交
609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628
                  &beta,
                  args4.idesc.desc(),
                  transformed_dx,
                  workspace_ptr,
                  workspace_size));
        },
        workspace_size);
#else
    for (int i = 0; i < groups; i++) {
      wkspace_handle.RunFunc(
          [&](void* workspace_ptr) {
            PADDLE_ENFORCE_GPU_SUCCESS(
                paddle::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(),
629
                    data_result.algo,
H
hong 已提交
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649
                    workspace_ptr,
                    workspace_size,
                    &beta,
                    args4.idesc.desc(),
                    transformed_dx + i * group_offset_in));
          },
          workspace_size);
    }
#endif

    if (!is_sys_pad) {
      // reverse padded input
      std::vector<int> starts(X->dims().size(), 0);
      std::vector<int> axes(X->dims().size(), 0);

      for (size_t i = 0; i < X->dims().size(); ++i) {
        starts[i] = input_pad[2 * i];
        axes[i] = i;
      }
      if (X->dims().size() == 4) {
650
        RemovePaddingSlice<Context, T, 4>(
H
hong 已提交
651 652
            ctx, &transformed_dX, &transformed_dX_channel, starts, axes);
      } else {
653
        RemovePaddingSlice<Context, T, 5>(
H
hong 已提交
654 655 656 657 658 659 660 661 662 663
            ctx, &transformed_dX, &transformed_dX_channel, starts, axes);
      }
    }
    if (channel_last) {
      TransToChannelLast<Context, T>(ctx, &transformed_dX_channel, dX);
    }
  }
}

template <typename T, typename Context>
664
void DepthwiseConvDoubleGradGPUDNNKernel(
H
hong 已提交
665 666 667
    const Context& ctx,
    const DenseTensor& input,
    const DenseTensor& filter,
668 669 670
    const DenseTensor& out_grad,
    const paddle::optional<DenseTensor>& input_grad_grad,
    const paddle::optional<DenseTensor>& filter_grad_grad,
H
hong 已提交
671 672 673 674 675 676 677 678 679 680 681
    const std::vector<int>& strides,
    const std::vector<int>& paddings_t,
    const std::string& padding_algorithm,
    int groups,
    const std::vector<int>& dilations_t,
    const std::string& data_format,
    bool use_addto,
    int workspace_size_MB,
    bool exhaustive_search_t,
    bool fuse_relu,
    DenseTensor* input_grad,
682 683
    DenseTensor* filter_grad,
    DenseTensor* out_grad_grad) {
H
hong 已提交
684 685 686
  ConvCudnnGradGradKernel<T>(ctx,
                             input,
                             filter,
687 688 689
                             out_grad,
                             input_grad_grad,
                             filter_grad_grad,
H
hong 已提交
690 691 692 693 694 695 696 697 698 699
                             strides,
                             paddings_t,
                             padding_algorithm,
                             groups,
                             dilations_t,
                             data_format,
                             use_addto,
                             workspace_size_MB,
                             exhaustive_search_t,
                             input_grad,
700 701
                             filter_grad,
                             out_grad_grad);
H
hong 已提交
702 703 704 705 706 707 708
}

template <typename T, typename Context>
void Conv3DCudnnGradGradKernel(
    const Context& ctx,
    const DenseTensor& input,
    const DenseTensor& filter,
709 710 711
    const DenseTensor& out_grad,
    const paddle::optional<DenseTensor>& input_grad_grad,
    const paddle::optional<DenseTensor>& filter_grad_grad,
H
hong 已提交
712 713 714 715 716 717 718 719 720 721
    const std::vector<int>& strides,
    const std::vector<int>& paddings_t,
    const std::string& padding_algorithm,
    int groups,
    const std::vector<int>& dilations_t,
    const std::string& data_format,
    bool use_addto,
    int workspace_size_MB,
    bool exhaustive_search_t,
    DenseTensor* input_grad,
722 723
    DenseTensor* filter_grad,
    DenseTensor* out_grad_grad) {
H
hong 已提交
724 725 726
  ConvCudnnGradGradKernel<T>(ctx,
                             input,
                             filter,
727 728 729
                             out_grad,
                             input_grad_grad,
                             filter_grad_grad,
H
hong 已提交
730 731 732 733 734 735 736 737 738 739
                             strides,
                             paddings_t,
                             padding_algorithm,
                             groups,
                             dilations_t,
                             data_format,
                             use_addto,
                             workspace_size_MB,
                             exhaustive_search_t,
                             input_grad,
740 741
                             filter_grad,
                             out_grad_grad);
H
hong 已提交
742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763
}

}  // namespace phi

#ifdef PADDLE_WITH_HIP
PD_REGISTER_KERNEL(conv2d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::ConvCudnnGradGradKernel,
                   float,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(conv3d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3DCudnnGradGradKernel,
                   float,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(depthwise_conv2d_grad_grad,
                   GPU,
                   ALL_LAYOUT,
764
                   phi::DepthwiseConvDoubleGradGPUDNNKernel,
H
hong 已提交
765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789
                   float,
                   phi::dtype::float16) {}
#else
#if CUDNN_VERSION_MIN(8, 1, 0)
PD_REGISTER_KERNEL(conv2d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::ConvCudnnGradGradKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}

PD_REGISTER_KERNEL(conv3d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3DCudnnGradGradKernel,
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}

PD_REGISTER_KERNEL(depthwise_conv2d_grad_grad,
                   GPU,
                   ALL_LAYOUT,
790
                   phi::DepthwiseConvDoubleGradGPUDNNKernel,
H
hong 已提交
791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816
                   float,
                   double,
                   phi::dtype::float16,
                   phi::dtype::bfloat16) {}

#else

PD_REGISTER_KERNEL(conv2d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::ConvCudnnGradGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(conv3d_grad_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3DCudnnGradGradKernel,
                   float,
                   double,
                   phi::dtype::float16) {}

PD_REGISTER_KERNEL(depthwise_conv2d_grad_grad,
                   GPU,
                   ALL_LAYOUT,
817
                   phi::DepthwiseConvDoubleGradGPUDNNKernel,
H
hong 已提交
818 819 820 821 822 823 824
                   float,
                   double,
                   phi::dtype::float16) {}

#endif

#endif