conv_cudnn_helper.h

/* Copyright (c) 2019 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. */

#pragma once

#include "paddle/fluid/operators/conv_base_helper.h"
#include "paddle/fluid/platform/cuda_graph_with_memory_pool.h"
#include "paddle/fluid/platform/device/gpu/gpu_info.h"
#include "paddle/phi/kernels/funcs/eigen/eigen_function.h"

namespace paddle {
namespace operators {

using ConvArgs = ConvArgsBase<cudnnHandle_t, cudnnDataType_t>;

template <typename DeviceContext, typename T, size_t D>
static void RemovePaddingSlice(const phi::GPUContext& context,
                               const Tensor* input, Tensor* out,
                               const std::vector<int>& starts,
                               const std::vector<int>& axes) {
  auto& place = *context.eigen_device();
  auto in_dims = input->dims();
  auto new_out_dims = out->dims();
  auto offsets = Eigen::DSizes<Eigen::DenseIndex, D>();
  auto extents = Eigen::DSizes<Eigen::DenseIndex, D>();
  for (size_t i = 0; i < D; ++i) {
    offsets[i] = 0;
    extents[i] = new_out_dims[i];
  }

  for (size_t i = 0; i < axes.size(); ++i) {
    int start = starts[i];
    if (start < 0) {
      start = (start + in_dims[axes[i]]);
    }
    start = std::max(start, 0);
    offsets[axes[i]] = start;
  }

  auto in_t =
      framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
          *input);
  auto out_t =
      framework::EigenTensor<T, D, Eigen::RowMajor, Eigen::DenseIndex>::From(
          *out, new_out_dims);

  phi::funcs::EigenSlice<std::decay_t<decltype(place)>, T, D>::Eval(
      place, out_t, in_t, offsets, extents);
}

static inline double ToMegaBytes(size_t bytes) {
  return static_cast<double>(bytes) / (1 << 20);
}

static inline bool UseFixedWorkspace() {
  return FLAGS_conv_workspace_size_limit >= 0;
}

static size_t CaclWorkspaceLimitInBytes(const phi::GPUContext& ctx) {
  if (!UseFixedWorkspace()) {
    int device_id = platform::GetCurrentDeviceId();
    int64_t allocated = memory::StatGetCurrentValue("Allocated", device_id);
    int64_t reserved = memory::StatGetCurrentValue("Reserved", device_id);
    int64_t availble = platform::GpuAvailableMemToAlloc();
    int64_t cur_workspace_size = ctx.cudnn_workspace_handle().WorkspaceSize();
    VLOG(3) << "[memory] allocated=" << ToMegaBytes(allocated)
            << " MB, reserved=" << ToMegaBytes(reserved)
            << " MB, available_to_alloc=" << ToMegaBytes(availble)
            << " MB, current_workspace_size=" << ToMegaBytes(cur_workspace_size)
            << " MB.";
    return std::max(std::max(availble, cur_workspace_size),
                    reserved - allocated);
  } else {
    return FLAGS_conv_workspace_size_limit * 1024 * 1024;
  }
}

template <typename PerfT>
std::string GetPerfResultString(std::string prefix,
                                const std::vector<PerfT>& perf_results,
                                int actual_algo_count, size_t workspace_limit) {
  std::ostringstream out;
  out << prefix << " (workspace limit=" << ToMegaBytes(workspace_limit)
      << " MB):\n";
  for (int i = 0; i < actual_algo_count; ++i) {
    const auto& result = perf_results[i];
    auto math_type_str = (result.mathType == CUDNN_TENSOR_OP_MATH) ? "T" : "F";
    out << "  algo=" << result.algo << ": tensor_core=" << math_type_str
        << ", time=" << result.time
        << " ms, memory=" << ToMegaBytes(result.memory)
        << " MB, status=" << result.status << "\n";
  }
  return out.str();
}

template <typename PerfT, typename AlgoT>
void ChooseAlgoByWorkspace(const std::vector<PerfT>& perf_results,
                           size_t workspace_limit,
                           SearchResult<AlgoT>* algo_result) {
  for (size_t i = 0; i < perf_results.size(); ++i) {
    auto result = perf_results[i];
    if (result.status == CUDNN_STATUS_SUCCESS &&
        result.memory < workspace_limit) {
      algo_result->algo = result.algo;
      algo_result->time = result.time;
      algo_result->workspace_size = result.memory;
      VLOG(3) << "  algo=" << result.algo << ", time=" << result.time
              << " ms, memory=" << ToMegaBytes(result.memory)
              << " MB (limit=" << ToMegaBytes(workspace_limit)
              << " MB), status=" << result.status;
      return;
    }
  }
  VLOG(3) << "Can not find an algorithm that requires memory < "
          << ToMegaBytes(workspace_limit) << " MB";
}

static void SetConvMathType(const phi::GPUContext& ctx, cudnnDataType_t dtype,
                            const platform::ConvolutionDescriptor& cdesc) {
#if CUDA_VERSION >= 9000 && CUDNN_VERSION_MIN(7, 0, 1)
  if (ctx.GetComputeCapability() >= 70 && dtype == CUDNN_DATA_HALF) {
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
        cdesc.desc(), CUDNN_TENSOR_OP_MATH));
    VLOG(5) << "use cudnn_tensor_op_math";
#if CUDA_VERSION >= 11000
#if CUDNN_VERSION_MIN(8, 1, 0)
  } else if (ctx.GetComputeCapability() >= 80 && dtype == CUDNN_DATA_BFLOAT16) {
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
        cdesc.desc(), CUDNN_TENSOR_OP_MATH));
#endif  // CUDNN_VERSION_MIN(8, 1, 0)
  } else if (dtype == CUDNN_DATA_FLOAT && !cdesc.allow_tf32_) {
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
        cdesc.desc(), CUDNN_FMA_MATH));
#endif  // CUDA_VERSION >= 11000
  } else {
    PADDLE_ENFORCE_GPU_SUCCESS(platform::dynload::cudnnSetConvolutionMathType(
        cdesc.desc(), CUDNN_DEFAULT_MATH));
    VLOG(5) << "NOT use cudnn_tensor_op_math";
  }
#endif
}

template <>
struct SearchAlgorithm<cudnnConvolutionFwdAlgoPerf_t> {
  using PerfT = cudnnConvolutionFwdAlgoPerf_t;
  using AlgoT = cudnnConvolutionFwdAlgo_t;

  template <typename T>
  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
                                  bool deterministic,
                                  const phi::GPUContext& ctx) {
    SearchResult<AlgoT> result;
    auto dtype = platform::CudnnDataType<T>::type;
    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
    SetConvMathType(ctx, dtype, args.cdesc);

    if (!exhaustive_search && !deterministic) {
#if CUDNN_VERSION >= 7001
      int actual_perf_count;
      int best_algo_idx = 0;
      std::vector<PerfT> perf_results(kNUM_CUDNN_FWD_ALGS);
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionForwardAlgorithm_v7(
              args.handle, args.idesc.desc(), args.wdesc.desc(),
              args.cdesc.desc(), args.odesc.desc(), kNUM_CUDNN_FWD_ALGS,
              &actual_perf_count, perf_results.data()));
      result.algo = perf_results[best_algo_idx].algo;
      result.workspace_size = perf_results[best_algo_idx].memory;

      if (result.workspace_size > workspace_size_limit) {
#if CUDNN_VERSION >= 8000
        // cudnnGetConvolutionForwardAlgorithm is removed in CUDNN-8
        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
                                            &result);
#else
        VLOG(3) << "Fallback to non-v7 method to find conv algorithm "
                   "becasue the workspace size request("
                << result.workspace_size << ") exceeds the limit("
                << workspace_size_limit << ")";
        PADDLE_ENFORCE_GPU_SUCCESS(
            platform::dynload::cudnnGetConvolutionForwardAlgorithm(
                args.handle, args.idesc.desc(), args.wdesc.desc(),
                args.cdesc.desc(), args.odesc.desc(),
                CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
                workspace_size_limit, &(result.algo)));
#endif
      }
#else
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionForwardAlgorithm(
              args.handle, args.idesc.desc(), args.wdesc.desc(),
              args.cdesc.desc(), args.odesc.desc(),
              CUDNN_CONVOLUTION_FWD_SPECIFY_WORKSPACE_LIMIT,
              workspace_size_limit, &(result.algo)));
#endif
    } else if (deterministic) {
      result.algo = static_cast<AlgoT>(1);
    } else {
      auto workspace_handle = ctx.cudnn_workspace_handle();
      auto x_dims = phi::vectorize(args.x->dims());
      auto w_dims = phi::vectorize(args.w->dims());
      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t:"
               << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
               << args.s << ", args.p" << args.p << ", args.d" << args.d;

      AlgorithmsCache<AlgoT>& algo_cache =
          *(framework::ConvSearchCache::Instance().GetForward());

      result.algo = algo_cache.GetAlgorithm(
          x_dims, w_dims, args.s, args.p, args.d, 0,
          static_cast<int64_t>(args.cudnn_dtype), [&]() {
            int returned_algo_count;
            std::vector<PerfT> perf_results(kNUM_CUDNN_FWD_ALGS);
            size_t max_workspace_size =
                FindMaxWorkspaceSize(args, workspace_size_limit);
            VLOG(4) << "max_workspace_size=" << ToMegaBytes(max_workspace_size)
                    << " MB";

            auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
              PADDLE_ENFORCE_GPU_SUCCESS(
                  platform::dynload::cudnnFindConvolutionForwardAlgorithmEx(
                      args.handle, args.idesc.desc(), args.x->data<T>(),
                      args.wdesc.desc(), args.w->data<T>(), args.cdesc.desc(),
                      args.odesc.desc(), const_cast<T*>(args.o->data<T>()),
                      kNUM_CUDNN_FWD_ALGS, &returned_algo_count,
                      perf_results.data(), cudnn_workspace_ptr,
                      max_workspace_size));
            };
            workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
                                         UseFixedWorkspace());

            VLOG(4) << GetPerfResultString<PerfT>(
                "[Exhaustive Search] FwdAlgo Perf result", perf_results,
                returned_algo_count, workspace_size_limit);
            result.time = perf_results[0].time;
            return perf_results[0].algo;
          });
    }
    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
            << ", deterministic=" << deterministic
            << ", choose algo=" << result.algo << ", workspace="
            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
    return result;
  }

  static size_t GetWorkspaceSize(const ConvArgs& args,
                                 cudnnConvolutionFwdAlgo_t algo) {
    size_t workspace_size = 0;
    PADDLE_ENFORCE_GPU_SUCCESS(
        platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
            args.handle, args.idesc.desc(), args.wdesc.desc(),
            args.cdesc.desc(), args.odesc.desc(), algo, &workspace_size));
    return workspace_size;
  }

 private:
  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
                                     size_t workspace_size_limit) {
    if (!UseFixedWorkspace()) {
      size_t max_workspace_size = 0;
      for (size_t algo = 0; algo < kNUM_CUDNN_FWD_ALGS; ++algo) {
        size_t workspace_size = 0;
        auto status =
            platform::dynload::cudnnGetConvolutionForwardWorkspaceSize(
                args.handle, args.idesc.desc(), args.wdesc.desc(),
                args.cdesc.desc(), args.odesc.desc(),
                static_cast<cudnnConvolutionFwdAlgo_t>(algo), &workspace_size);
        if (status == CUDNN_STATUS_SUCCESS &&
            workspace_size <= workspace_size_limit) {
          max_workspace_size = std::max(workspace_size, max_workspace_size);
        }
      }
      return max_workspace_size;
    } else {
      return workspace_size_limit;
    }
  }
};

template <>
struct SearchAlgorithm<cudnnConvolutionBwdDataAlgoPerf_t> {
  using PerfT = cudnnConvolutionBwdDataAlgoPerf_t;
  using AlgoT = cudnnConvolutionBwdDataAlgo_t;

  template <typename T>
  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
                                  bool deterministic,
                                  const phi::GPUContext& ctx) {
    SearchResult<AlgoT> result;
    auto dtype = platform::CudnnDataType<T>::type;
    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
    SetConvMathType(ctx, dtype, args.cdesc);

    if (!exhaustive_search && !deterministic) {
#if CUDNN_VERSION >= 7001
      int actual_perf_count;
      int best_algo_idx = 0;
      std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_DATA_ALGS);
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm_v7(
              args.handle, args.wdesc.desc(), args.odesc.desc(),
              args.cdesc.desc(), args.idesc.desc(), kNUM_CUDNN_BWD_DATA_ALGS,
              &actual_perf_count, perf_results.data()));
      result.algo = perf_results[best_algo_idx].algo;

#if CUDNN_VERSION < 7500
      int stride_dim = args.x->dims().size() - 2;
      bool blacklist = std::any_of(args.s.begin(), args.s.begin() + stride_dim,
                                   [=](int n) { return n != 1; });
      if (blacklist && (perf_results[best_algo_idx].algo ==
                            CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT_TILING ||
                        perf_results[best_algo_idx].algo ==
                            CUDNN_CONVOLUTION_BWD_DATA_ALGO_FFT)) {
        result.algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
      }
#endif
      result.workspace_size = GetWorkspaceSize(args, result.algo);
      if (result.workspace_size > workspace_size_limit) {
#if CUDNN_VERSION >= 8000
        // cudnnGetConvolutionBackwardDataAlgorithm is removed in CUDNN-8
        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
                                            &result);
#else
        VLOG(1) << "Fallback to non-v7 method to find conv algorithm becasue "
                   "the workspace size request("
                << result.workspace_size << ") exceeds the limit("
                << workspace_size_limit << ")";
        PADDLE_ENFORCE_GPU_SUCCESS(
            platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
                args.handle, args.wdesc.desc(), args.odesc.desc(),
                args.cdesc.desc(), args.idesc.desc(),
                CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
                workspace_size_limit, &(result.algo)));
#endif
      }
#else
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionBackwardDataAlgorithm(
              args.handle, args.wdesc.desc(), args.odesc.desc(),
              args.cdesc.desc(), args.idesc.desc(),
              CUDNN_CONVOLUTION_BWD_DATA_SPECIFY_WORKSPACE_LIMIT,
              workspace_size_limit, &(result.algo)));
#endif
    } else if (deterministic) {
      result.algo = CUDNN_CONVOLUTION_BWD_DATA_ALGO_1;
    } else {
      auto workspace_handle = ctx.cudnn_workspace_handle();
      auto x_dims = phi::vectorize(args.x->dims());
      auto w_dims = phi::vectorize(args.w->dims());
      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t"
               << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
               << args.s << ", args.p" << args.p << ", args.d" << args.d;

      AlgorithmsCache<AlgoT>& algo_cache =
          *(framework::ConvSearchCache::Instance().GetBackwardData());
      result.algo = algo_cache.GetAlgorithm(
          x_dims, w_dims, args.s, args.p, args.d, 0,
          static_cast<int64_t>(args.cudnn_dtype), [&]() {
            int returned_algo_count;
            std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_DATA_ALGS);
            size_t max_workspace_size =
                FindMaxWorkspaceSize(args, workspace_size_limit);
            VLOG(3) << "max_workspace_size=" << ToMegaBytes(max_workspace_size)
                    << " MB";

            auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
              PADDLE_ENFORCE_GPU_SUCCESS(
                  platform::dynload::
                      cudnnFindConvolutionBackwardDataAlgorithmEx(
                          args.handle, args.wdesc.desc(), args.w->data<T>(),
                          args.odesc.desc(), args.o->data<T>(),
                          args.cdesc.desc(), args.idesc.desc(),
                          const_cast<T*>(args.x->data<T>()),
                          kNUM_CUDNN_BWD_DATA_ALGS, &returned_algo_count,
                          perf_results.data(), cudnn_workspace_ptr,
                          max_workspace_size));
            };
            workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
                                         UseFixedWorkspace());

            VLOG(3) << GetPerfResultString<PerfT>(
                "[Exhaustive Search] BwdDataAlgo Perf result", perf_results,
                returned_algo_count, workspace_size_limit);
            result.time = perf_results[0].time;
            return perf_results[0].algo;
          });
    }
    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
            << ", deterministic=" << deterministic
            << ", choose algo=" << result.algo << ", workspace="
            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
    return result;
  }

  static size_t GetWorkspaceSize(const ConvArgs& args,
                                 cudnnConvolutionBwdDataAlgo_t algo) {
    size_t workspace_size = 0;
    PADDLE_ENFORCE_GPU_SUCCESS(
        platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
            args.handle, args.wdesc.desc(), args.odesc.desc(),
            args.cdesc.desc(), args.idesc.desc(), algo, &workspace_size));
    return workspace_size;
  }

 private:
  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
                                     size_t workspace_size_limit) {
    if (!UseFixedWorkspace()) {
      size_t max_workspace_size = 0;
      for (size_t algo = 0; algo < kNUM_CUDNN_BWD_DATA_ALGS; ++algo) {
        size_t workspace_size = 0;
        auto status =
            platform::dynload::cudnnGetConvolutionBackwardDataWorkspaceSize(
                args.handle, args.wdesc.desc(), args.odesc.desc(),
                args.cdesc.desc(), args.idesc.desc(),
                static_cast<cudnnConvolutionBwdDataAlgo_t>(algo),
                &workspace_size);
        if (status == CUDNN_STATUS_SUCCESS &&
            workspace_size <= workspace_size_limit) {
          max_workspace_size = std::max(workspace_size, max_workspace_size);
        }
      }
      return max_workspace_size;
    } else {
      return workspace_size_limit;
    }
  }
};

template <>
struct SearchAlgorithm<cudnnConvolutionBwdFilterAlgoPerf_t> {
  using PerfT = cudnnConvolutionBwdFilterAlgoPerf_t;
  using AlgoT = cudnnConvolutionBwdFilterAlgo_t;

  template <typename T>
  static SearchResult<AlgoT> Find(const ConvArgs& args, bool exhaustive_search,
                                  bool deterministic,
                                  const phi::GPUContext& ctx) {
    platform::CUDAGraphCaptureModeGuard guard;
    SearchResult<AlgoT> result;
    auto dtype = platform::CudnnDataType<T>::type;
    size_t workspace_size_limit = CaclWorkspaceLimitInBytes(ctx);
    SetConvMathType(ctx, dtype, args.cdesc);

    if (!exhaustive_search && !deterministic) {
#if CUDNN_VERSION >= 7001
      int actual_perf_count;
      int best_algo_idx = 0;
      std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm_v7(
              args.handle, args.idesc.desc(), args.odesc.desc(),
              args.cdesc.desc(), args.wdesc.desc(), kNUM_CUDNN_BWD_FILTER_ALGS,
              &actual_perf_count, perf_results.data()));
      result.algo = perf_results[best_algo_idx].algo;
      result.workspace_size = perf_results[best_algo_idx].memory;

      if (result.workspace_size > workspace_size_limit) {
#if CUDNN_VERSION >= 8000
        // cudnnGetConvolutionBackwardFilterAlgorithm is removed in CUDNN-8
        ChooseAlgoByWorkspace<PerfT, AlgoT>(perf_results, workspace_size_limit,
                                            &result);
#else
        VLOG(1) << "Fallback to non-v7 method to find conv algorithm becasue "
                   "the workspace size request("
                << result.workspace_size << ") exceeds the limit("
                << workspace_size_limit << ")";
        PADDLE_ENFORCE_GPU_SUCCESS(
            platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
                args.handle, args.idesc.desc(), args.odesc.desc(),
                args.cdesc.desc(), args.wdesc.desc(),
                CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
                workspace_size_limit, &(result.algo)));
#endif
      }
#else
      PADDLE_ENFORCE_GPU_SUCCESS(
          platform::dynload::cudnnGetConvolutionBackwardFilterAlgorithm(
              args.handle, args.idesc.desc(), args.odesc.desc(),
              args.cdesc.desc(), args.wdesc.desc(),
              CUDNN_CONVOLUTION_BWD_FILTER_SPECIFY_WORKSPACE_LIMIT,
              workspace_size_limit, &(result.algo)));
#endif
    } else if (deterministic) {
      result.algo = CUDNN_CONVOLUTION_BWD_FILTER_ALGO_1;
    } else {
      auto workspace_handle = ctx.cudnn_workspace_handle();
      auto x_dims = phi::vectorize(args.x->dims());
      auto w_dims = phi::vectorize(args.w->dims());
      VLOG(10) << "cudnnConvolutionFwdAlgoPerf_t:"
               << ", x_dims:" << x_dims << ", w_dims:" << w_dims << ", args.s"
               << args.s << ", args.p" << args.p << ", args.d" << args.d;

      AlgorithmsCache<AlgoT>& algo_cache =
          *(framework::ConvSearchCache::Instance().GetBackwardFilter());

      if (dtype != CUDNN_DATA_HALF) {
        result.algo = algo_cache.GetAlgorithm(
            x_dims, w_dims, args.s, args.p, args.d, 0,
            static_cast<int64_t>(args.cudnn_dtype), [&]() {
              int returned_algo_count;
              std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);
              size_t max_workspace_size =
                  FindMaxWorkspaceSize(args, workspace_size_limit);
              VLOG(3) << "max_workspace_size="
                      << ToMegaBytes(max_workspace_size) << " MB";

              auto cudnn_find_func = [&](void* cudnn_workspace_ptr) {
                PADDLE_ENFORCE_GPU_SUCCESS(
                    platform::dynload::
                        cudnnFindConvolutionBackwardFilterAlgorithmEx(
                            args.handle, args.idesc.desc(), args.x->data<T>(),
                            args.odesc.desc(), args.o->data<T>(),
                            args.cdesc.desc(), args.wdesc.desc(),
                            const_cast<T*>(args.w->data<T>()),
                            kNUM_CUDNN_BWD_FILTER_ALGS, &returned_algo_count,
                            perf_results.data(), cudnn_workspace_ptr,
                            max_workspace_size));
              };
              workspace_handle.RunFuncSync(cudnn_find_func, max_workspace_size,
                                           UseFixedWorkspace());

              VLOG(3) << GetPerfResultString<PerfT>(
                  "[Exhaustive Search] BwdFilterAlgo Perf result", perf_results,
                  returned_algo_count, workspace_size_limit);
              result.time = perf_results[0].time;
              return perf_results[0].algo;
            });
      } else {
        result.algo = algo_cache.GetAlgorithm(
            x_dims, w_dims, args.s, args.p, args.d, 0,
            static_cast<int64_t>(args.cudnn_dtype), [&]() {
              SearchResult<AlgoT> algo_result;
              int actual_algos = 0;
              std::vector<PerfT> perf_results(kNUM_CUDNN_BWD_FILTER_ALGS);

              PADDLE_ENFORCE_GPU_SUCCESS(
                  platform::dynload::
                      cudnnFindConvolutionBackwardFilterAlgorithm(
                          args.handle, args.idesc.desc(), args.odesc.desc(),
                          args.cdesc.desc(), args.wdesc.desc(),
                          perf_results.size(), &actual_algos,
                          perf_results.data()));
              perf_results.resize(actual_algos);
              ChooseAlgo(perf_results, workspace_size_limit, &algo_result);
              result.time = algo_result.time;
              return algo_result.algo;
            });
      }
    }
    VLOG(3) << "[cuDNN Convoltion] exhaustive_search=" << exhaustive_search
            << ", deterministic=" << deterministic
            << ", choose algo=" << result.algo << ", workspace="
            << ToMegaBytes(GetWorkspaceSize(args, result.algo)) << " MB";
    return result;
  }

  static size_t GetWorkspaceSize(const ConvArgs& args,
                                 cudnnConvolutionBwdFilterAlgo_t algo) {
    platform::CUDAGraphCaptureModeGuard guard;
    size_t workspace_size = 0;
    PADDLE_ENFORCE_GPU_SUCCESS(
        platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
            args.handle, args.idesc.desc(), args.odesc.desc(),
            args.cdesc.desc(), args.wdesc.desc(), algo, &workspace_size));
    return workspace_size;
  }

 private:
  static size_t FindMaxWorkspaceSize(const ConvArgs& args,
                                     size_t workspace_size_limit) {
    if (!UseFixedWorkspace()) {
      size_t max_workspace_size = 0;
      for (size_t algo = 0; algo < kNUM_CUDNN_BWD_FILTER_ALGS; ++algo) {
        size_t workspace_size = 0;
        auto status =
            platform::dynload::cudnnGetConvolutionBackwardFilterWorkspaceSize(
                args.handle, args.idesc.desc(), args.odesc.desc(),
                args.cdesc.desc(), args.wdesc.desc(),
                static_cast<cudnnConvolutionBwdFilterAlgo_t>(algo),
                &workspace_size);
        if (status == CUDNN_STATUS_SUCCESS &&
            workspace_size <= workspace_size_limit) {
          max_workspace_size = std::max(workspace_size, max_workspace_size);
        }
      }
      return max_workspace_size;
    } else {
      return workspace_size_limit;
    }
  }

  static void ChooseAlgo(const std::vector<PerfT>& perf_results,
                         size_t workspace_limit,
                         SearchResult<AlgoT>* algo_result) {
    VLOG(3) << GetPerfResultString<PerfT>(
        "[Exhaustive Search] BwdFilterAlgo Perf result", perf_results,
        perf_results.size(), workspace_limit);

    for (size_t i = 0; i != perf_results.size(); ++i) {
      const auto& result = perf_results[i];
      if (result.status == CUDNN_STATUS_SUCCESS &&
          (result.memory <= workspace_limit)) {
        if ((result.mathType == CUDNN_TENSOR_OP_MATH) &&
            (i != perf_results.size() - 1)) {
          const auto& next_result = perf_results[i + 1];
          if (next_result.status == CUDNN_STATUS_SUCCESS &&
              next_result.algo == result.algo &&
              next_result.memory == result.memory &&
              next_result.mathType != CUDNN_TENSOR_OP_MATH &&
              next_result.time < 1.01 * result.time) {
            // Skip over this result- it's not really a Tensor Core algo.
            // Because it is only 1% performance difference.
            // Prefer to choose the next equivalent non-Tensor Core algo.
            continue;
          }
        }
        algo_result->algo = result.algo;
        algo_result->time = result.time;
        auto math_type_str = "0";
        if (result.mathType == CUDNN_TENSOR_OP_MATH) {
          math_type_str = "1";
        }
        VLOG(3) << "    choose algo: " << result.algo
                << ", TC: " << math_type_str << ", time: " << result.time
                << " ms, wksp = " << result.memory
                << ", status = " << result.status;
        break;
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle