/* 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. */

#include "paddle/phi/kernels/conv_transpose_kernel.h"

#include <algorithm>

#include "paddle/common/ddim.h"
#include "paddle/phi/backends/context_pool.h"
#include "paddle/phi/backends/dynload/cudnn.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/cpu/conv_util.h"
#include "paddle/phi/kernels/funcs/padding.h"
#include "paddle/phi/kernels/funcs/slice.h"
#include "paddle/phi/kernels/gpudnn/conv_gpudnn.h"
#include "paddle/phi/kernels/transpose_kernel.h"

#ifdef PADDLE_WITH_HIP
#include "paddle/phi/backends/gpu/rocm/miopen_helper.h"
#include "paddle/phi/kernels/gpudnn/conv_miopen_helper.h"
#else
#include "paddle/phi/backends/gpu/cuda/cudnn_helper.h"
#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
#include "paddle/phi/kernels/gpudnn/conv_cudnn_v7.h"
#endif
#include "paddle/phi/kernels/full_kernel.h"

#ifdef PADDLE_WITH_CUDNN_FRONTEND
// clang-format off
#include "paddle/phi/backends/dynload/cudnn_frontend.h"
#include "paddle/phi/kernels/autotune/cache.h"
#include "paddle/phi/kernels/gpudnn/conv_cudnn_frontend.h"
// clang-format on
#endif

#include "paddle/common/flags.h"

COMMON_DECLARE_bool(use_accuracy_compatible_kernel);

namespace phi {

template <typename T, typename Context>
void ConvTransposeCudnnKernelImplV7(const DenseTensor* transformed_x,
                                    const DenseTensor* filter,
                                    const Context& dev_ctx,
                                    const std::vector<int>& strides,
                                    const std::vector<int>& padding_common,
                                    const std::vector<int>& dilations_,
                                    DataLayout data_layout,
                                    DataLayout layout,
                                    bool exhaustive_search,
                                    bool deterministic,
                                    int groups,
                                    DenseTensor* transformed_out) {
  int iwo_groups = 1;
  int c_groups = groups;
  groups = 1;
  size_t workspace_size = 0;

  const T* x_data = transformed_x->data<T>();
  const T* filter_data = filter->data<T>();
  T* transformed_out_data = transformed_out->data<T>();
#ifdef PADDLE_WITH_HIP
  miopenConvBwdDataAlgorithm_t algo{};
#else
  cudnnConvolutionBwdDataAlgo_t algo{};
#endif
  // ------------------- cudnn conv algorithm ---------------------
  auto handle = dev_ctx.cudnn_handle();
  auto layout_tensor = phi::backends::gpu::GetCudnnTensorFormat(layout);
  auto dtype = phi::backends::gpu::CudnnDataType<T>::type;
  // ------------------- cudnn descriptors ---------------------
  ConvArgs args{handle,
                transformed_out,
                filter,
                transformed_x,
                strides,
                padding_common,
                dilations_,
                dtype,
                groups,
                data_layout};
  args.idesc.set(*transformed_out, iwo_groups);
  args.wdesc.set(*filter, layout_tensor, iwo_groups);
  args.odesc.set(*transformed_x, iwo_groups);
  args.cdesc.set(dtype,
                 padding_common,
                 strides,
                 dilations_,
                 phi::AllowTF32Cudnn(),
                 c_groups);

#ifdef PADDLE_WITH_HIP
  SearchResult<miopenConvBwdDataAlgorithm_t> bwd_result;
  using search = SearchAlgorithm<miopenConvBwdDataAlgorithm_t>;
  workspace_size = std::max(workspace_size, search::GetWorkspaceSize(args));
  bwd_result.algo = search::Find<T>(
      args, exhaustive_search, deterministic, workspace_size, dev_ctx);
#else
  SearchResult<cudnnConvolutionBwdDataAlgo_t> bwd_result;
  using search = SearchAlgorithm<ConvKind::kBackwardData>;
  bwd_result =
      search::Find<T>(dev_ctx, args, exhaustive_search, deterministic, false);
  workspace_size =
      std::max(workspace_size, search::GetWorkspaceSize(args, bwd_result.algo));
#endif

  // ------------------- cudnn conv transpose forward ---------------------
  int64_t x_offset_64 =
      transformed_x->numel() / transformed_x->dims()[0] / groups;
  int64_t out_offset_64 =
      transformed_out->numel() / transformed_out->dims()[0] / groups;
  int64_t filter_offset_64 = filter->numel() / groups;
  PADDLE_ENFORCE_LE_INT_MAX(x_offset_64, "x_offset");
  PADDLE_ENFORCE_LE_INT_MAX(out_offset_64, "out_offset");
  PADDLE_ENFORCE_LE_INT_MAX(filter_offset_64, "filter_offset");
  int x_offset = static_cast<int>(x_offset_64);
  int out_offset = static_cast<int>(out_offset_64);
  int filter_offset = static_cast<int>(filter_offset_64);
  ScalingParamType<T> alpha = 1.0f;
  ScalingParamType<T> beta = 0.0f;
  auto workspace_handle = dev_ctx.cudnn_workspace_handle();
#ifdef PADDLE_WITH_HIP
  for (int g = 0; g < groups; g++) {
    auto cudnn_func = [&](void* cudnn_workspace) {
      PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenConvolutionBackwardData(
          handle,
          &alpha,
          args.odesc.desc(),
          x_data + x_offset * g,
          args.wdesc.desc(),
          filter_data + filter_offset * g,
          args.cdesc.desc(),
          bwd_result.algo,
          &beta,
          args.idesc.desc(),
          transformed_out_data + out_offset * g,
          cudnn_workspace,
          workspace_size));
    };
    workspace_handle.RunFunc(cudnn_func, workspace_size);
  }
#else
  ConvRunner<T, ConvKind::kBackwardData>::Apply(dev_ctx,
                                                args,
                                                bwd_result,
                                                x_data,
                                                filter_data,
                                                transformed_out_data,
                                                groups,
                                                out_offset,
                                                filter_offset,
                                                x_offset,
                                                workspace_size,
                                                &workspace_handle,
                                                false);
#endif
}
#ifdef PADDLE_WITH_CUDNN_FRONTEND
template <typename T, typename Context>
void ConvTransposeCudnnKernelImplV8(const DenseTensor* transformed_x,
                                    const DenseTensor* filter,
                                    const Context& dev_ctx,
                                    const std::vector<int>& strides,
                                    const std::vector<int>& padding_common,
                                    const std::vector<int>& dilations_,
                                    DataLayout data_layout,
                                    DataLayout layout,
                                    bool exhaustive_search,
                                    bool deterministic,
                                    int groups,
                                    DenseTensor* transformed_out) {
  auto& plan_cache = phi::autotune::AutoTuneCache::Instance().GetConvV8(
      phi::autotune::AlgorithmType::kConvBackwardDataV8);

  T* input_data = const_cast<T*>(transformed_x->data<T>());
  T* filter_data = const_cast<T*>(filter->data<T>());
  T* output_data = transformed_out->data<T>();
  cudnnHandle_t handle = const_cast<cudnnHandle_t>(dev_ctx.cudnn_handle());
  auto workspace_handle = dev_ctx.cudnn_workspace_handle();

  auto layout_format = phi::backends::gpu::GetCudnnTensorFormat(layout);
  auto dtype = phi::backends::gpu::CudnnDataType<T>::type;

  float alpha = 1.0f;
  float beta = 0.0f;

  using helper = CudnnFrontendConvHelper;
  auto op_graph = helper::BuildConvOperationGraph<
      CUDNN_BACKEND_OPERATION_CONVOLUTION_BACKWARD_DATA_DESCRIPTOR>(
      transformed_out,
      transformed_x,
      filter,
      layout_format,
      strides,
      padding_common,
      dilations_,
      dtype,
      handle,
      alpha,
      beta);
  if (plan_cache.FindPlan(op_graph, handle)) {
    const cudnn_frontend::ExecutionPlan* cached_plan = nullptr;
    int64_t workspace_size = 0;
    plan_cache.GetPlanAndWorkspaceSize(
        op_graph, &cached_plan, &workspace_size, handle);
    helper::ExecutePlan(handle,
                        &workspace_handle,
                        output_data,
                        input_data,
                        filter_data,
                        cached_plan->get_raw_desc(),
                        workspace_size);
    return;
  }

  auto plans = helper::FindExecutionPlans(&op_graph,
                                          exhaustive_search,
                                          deterministic,
                                          output_data,
                                          input_data,
                                          filter_data,
                                          handle,
                                          &workspace_handle,
                                          transformed_x->dtype());

  helper::ExecutePlansAndCache(handle,
                               &workspace_handle,
                               output_data,
                               input_data,
                               filter_data,
                               &plans,
                               exhaustive_search,
                               op_graph,
                               &plan_cache);
}
#endif

template <typename T, typename Context>
void ConvTransposeRawGPUDNNKernel(const Context& dev_ctx,
                                  const DenseTensor& x,
                                  const DenseTensor& filter,
                                  const std::vector<int>& strides,
                                  const std::vector<int>& paddings,
                                  const std::string& padding_algorithm,
                                  int groups,
                                  const std::vector<int>& dilations,
                                  const std::string& data_format,
                                  DenseTensor* out) {
  if (x.numel() == 0 || filter.numel() == 0) {
    Full<T, Context>(dev_ctx, out->dims(), 0, out);
    return;
  }

  bool has_exhaustive_search = dev_ctx.HasDnnAttr("exhaustive_search");
  bool exhaustive_search_attr =
      has_exhaustive_search
          ? PADDLE_GET_CONST(bool, dev_ctx.GetDnnAttr("exhaustive_search"))
          : false;
  bool exhaustive_search =
      FLAGS_cudnn_exhaustive_search || exhaustive_search_attr;
  bool deterministic = FLAGS_cudnn_deterministic;
  PADDLE_ENFORCE_EQ(exhaustive_search && deterministic,
                    false,
                    common::errors::InvalidArgument(
                        "Can't set exhaustive_search True and "
                        "FLAGS_cudnn_deterministic True at same time."));

  std::vector<int> paddings_ = paddings;
  std::vector<int> dilations_ = dilations;
  const DataLayout data_layout =
      (data_format != "NHWC" ? DataLayout::NCHW : DataLayout::NHWC);
  std::vector<int64_t> x_vec = vectorize<int64_t>(x.dims());
  std::vector<int64_t> out_vec = vectorize<int64_t>(out->dims());
  // if channel_last, transpose to channel_first
  DenseTensor x_transpose;
  if (data_layout == DataLayout::NHWC) {
    if (strides.size() == 2U) {
      std::vector<int> axis = {0, 3, 1, 2};
      for (size_t i = 0; i < axis.size(); ++i) {
        x_vec[i] = x.dims()[axis[i]];
        out_vec[i] = out->dims()[axis[i]];
      }
      x_transpose = Transpose<T, Context>(dev_ctx, x, axis);
    } else if (strides.size() == 3U) {
      std::vector<int> axis = {0, 4, 1, 2, 3};
      for (size_t i = 0; i < axis.size(); ++i) {
        x_vec[i] = x.dims()[axis[i]];
        out_vec[i] = out->dims()[axis[i]];
      }
      x_transpose = Transpose<T, Context>(dev_ctx, x, axis);
    }
  } else {
    x_transpose = x;
  }

  // update padding and dilation
  auto x_dims = x_transpose.dims();
  auto filter_dims = filter.dims();
  DDim x_data_dims;
  x_data_dims = slice_ddim(x_dims, 2, x_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, x_data_dims, strides, ksize);

  int data_dim = strides.size();  // 2d or 3d
  bool is_sys_pad = funcs::IsSymmetricPadding(paddings_, data_dim);

  std::vector<int> x_pad(x_dims.size() * 2, 0);
  DenseTensor transformed_x;
  std::vector<int> padding_common(data_dim, 0);
  if (!is_sys_pad) {
    std::vector<int> padding_diff(data_dim);
    std::vector<int64_t> new_x_shape_vec(data_dim + 2);
    new_x_shape_vec[0] = x_dims[0];
    new_x_shape_vec[1] = x_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_x_shape_vec[i + 2] = x_dims[i + 2] + padding_diff[i];
      x_pad[2 * i + 4] = paddings_[2 * i] - padding_common[i];
      x_pad[2 * i + 4 + 1] = paddings_[2 * i + 1] - padding_common[i];
    }
    DDim new_x_shape(make_ddim(new_x_shape_vec));
    transformed_x.Resize(new_x_shape);
    dev_ctx.template Alloc<T>(&transformed_x);

    const int rank = x_dims.size();
    T pad_value(0.0);
    switch (rank) {
      case 4: {
        funcs::PadFunction<Context, T, 4>(
            dev_ctx, x_pad, x_transpose, pad_value, &transformed_x);
      } break;
      case 5: {
        funcs::PadFunction<Context, T, 5>(
            dev_ctx, x_pad, x_transpose, pad_value, &transformed_x);
      } break;
      default:
        PADDLE_THROW(errors::InvalidArgument(
            "Op(ConvTranspose) only supports 4-D or 5-D x DenseTensor."));
    }
  } else {
    transformed_x = x_transpose;
    if (paddings_.size() == data_dim) {
      for (size_t i = 0; i < data_dim; ++i) {
        padding_common[i] = paddings_[i];
      }
    } else {
      for (size_t i = 0; i < data_dim; ++i) {
        padding_common[i] = paddings_[2 * i];
      }
    }
  }

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

  x_vec = vectorize<int64_t>(transformed_x.dims());

  std::vector<int64_t> transformed_out_vec = out_vec;
  for (size_t i = 0; i < data_dim; ++i) {
    transformed_out_vec[i + 2] =
        out_vec[i + 2] + (x_pad[2 * i + 4] + x_pad[2 * i + 5]) * strides[i] -
        2 * padding_common[i] + paddings_[2 * i] + paddings_[2 * i + 1];
  }

  DenseTensor transformed_out;
  if (!is_sys_pad) {
    transformed_out.Resize(transformed_out_vec);
    dev_ctx.template Alloc<T>(&transformed_out);
  } else {
    dev_ctx.template Alloc<T>(out);
    transformed_out.ShareDataWith(*out);
    transformed_out.Resize(transformed_out_vec);
  }

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

#ifdef PADDLE_WITH_CUDNN_FRONTEND
  if (dynload::IsCudnnFrontendEnabled())
    ConvTransposeCudnnKernelImplV8<T>(&transformed_x,
                                      &filter,
                                      dev_ctx,
                                      strides,
                                      padding_common,
                                      dilations_,
                                      data_layout,
                                      layout,
                                      exhaustive_search,
                                      deterministic,
                                      groups,
                                      &transformed_out);
  else
    ConvTransposeCudnnKernelImplV7<T>(&transformed_x,
                                      &filter,
                                      dev_ctx,
                                      strides,
                                      padding_common,
                                      dilations_,
                                      data_layout,
                                      layout,
                                      exhaustive_search,
                                      deterministic,
                                      groups,
                                      &transformed_out);
#else
  ConvTransposeCudnnKernelImplV7<T>(&transformed_x,
                                    &filter,
                                    dev_ctx,
                                    strides,
                                    padding_common,
                                    dilations_,
                                    data_layout,
                                    layout,
                                    exhaustive_search,
                                    deterministic,
                                    groups,
                                    &transformed_out);
#endif

  if (!is_sys_pad && strides.size() == 2U) {
    funcs::Slice<Context, T, 4>(
        dev_ctx, &transformed_out, out, starts, ends, axes);
  } else if (!is_sys_pad && strides.size() == 3U) {
    funcs::Slice<Context, T, 5>(
        dev_ctx, &transformed_out, out, starts, ends, axes);
  }

  if (data_layout == DataLayout::NHWC) {
    DenseTensor out_transpose;
    DenseTensor out_nchw;
    out_nchw.ShareDataWith(*out);
    out_nchw.Resize(out_vec);

    if (strides.size() == 2U) {
      out_transpose = Transpose<T, Context>(dev_ctx, out_nchw, {0, 2, 3, 1});
    } else if (strides.size() == 3U) {
      out_transpose = Transpose<T, Context>(dev_ctx, out_nchw, {0, 2, 3, 4, 1});
    }
    *out = out_transpose;
  }
}

#ifdef PADDLE_WITH_CUDNN_FRONTEND
template <typename T, typename Context>
void ConvTransposeRawGPUDNNKernelV8(const Context& dev_ctx,
                                    const DenseTensor& x,
                                    const DenseTensor& filter,
                                    const std::vector<int>& strides,
                                    const std::vector<int>& paddings,
                                    const std::string& padding_algorithm,
                                    int groups,
                                    const std::vector<int>& dilations,
                                    const std::string& data_format,
                                    DenseTensor* out) {
  DenseTensor dummy_input;
  dummy_input.Resize(out->dims());

  dev_ctx.template Alloc<T>(out);
  dummy_input.ShareDataWith(*out);

  DenseTensor* filter_grad_ptr = nullptr;

  ConvCudnnGradKernel<T, Context>(dev_ctx,
                                  dummy_input,
                                  filter,
                                  x,
                                  strides,
                                  paddings,
                                  padding_algorithm,
                                  dilations,
                                  groups,
                                  data_format,
                                  out,
                                  filter_grad_ptr);
}
#endif

template <typename T, typename Context>
void Conv2dTransposeGPUDNNKernel(const Context& dev_ctx,
                                 const DenseTensor& x,
                                 const DenseTensor& filter,
                                 const std::vector<int>& strides,
                                 const std::vector<int>& paddings,
                                 const std::vector<int>& output_padding,
                                 const IntArray& output_size,
                                 const std::string& padding_algorithm,
                                 int groups,
                                 const std::vector<int>& dilations,
                                 const std::string& data_format,
                                 DenseTensor* out) {
#ifdef PADDLE_WITH_CUDNN_FRONTEND
  if (dynload::IsCudnnFrontendEnabled() && FLAGS_use_accuracy_compatible_kernel)
    ConvTransposeRawGPUDNNKernelV8<T, Context>(dev_ctx,
                                               x,
                                               filter,
                                               strides,
                                               paddings,
                                               padding_algorithm,
                                               groups,
                                               dilations,
                                               data_format,
                                               out);
  else
    ConvTransposeRawGPUDNNKernel<T, Context>(dev_ctx,
                                             x,
                                             filter,
                                             strides,
                                             paddings,
                                             padding_algorithm,
                                             groups,
                                             dilations,
                                             data_format,
                                             out);

#else
  ConvTransposeRawGPUDNNKernel<T, Context>(dev_ctx,
                                           x,
                                           filter,
                                           strides,
                                           paddings,
                                           padding_algorithm,
                                           groups,
                                           dilations,
                                           data_format,
                                           out);
#endif
}

template <typename T, typename Context>
void Conv3dTransposeGPUDNNKernel(const Context& dev_ctx,
                                 const DenseTensor& x,
                                 const DenseTensor& filter,
                                 const std::vector<int>& strides,
                                 const std::vector<int>& paddings,
                                 const std::vector<int>& output_padding,
                                 const std::vector<int>& output_size,
                                 const std::string& padding_algorithm,
                                 int groups,
                                 const std::vector<int>& dilations,
                                 const std::string& data_format,
                                 DenseTensor* out) {
  ConvTransposeRawGPUDNNKernel<T, Context>(dev_ctx,
                                           x,
                                           filter,
                                           strides,
                                           paddings,
                                           padding_algorithm,
                                           groups,
                                           dilations,
                                           data_format,
                                           out);
}

}  // namespace phi

using float16 = phi::float16;

#ifdef PADDLE_WITH_HIP
// MIOPEN do not support double
PD_REGISTER_KERNEL(conv2d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv2dTransposeGPUDNNKernel,
                   float,
                   float16) {}
PD_REGISTER_KERNEL(conv3d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3dTransposeGPUDNNKernel,
                   float,
                   float16) {}
#else
#if CUDNN_VERSION_MIN(8, 1, 0)
PD_REGISTER_KERNEL(conv2d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv2dTransposeGPUDNNKernel,
                   float,
                   double,
                   float16,
                   phi::bfloat16) {}
PD_REGISTER_KERNEL(conv3d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3dTransposeGPUDNNKernel,
                   float,
                   double,
                   float16,
                   phi::bfloat16) {}
#else
PD_REGISTER_KERNEL(conv2d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv2dTransposeGPUDNNKernel,
                   float,
                   double,
                   float16) {}
PD_REGISTER_KERNEL(conv3d_transpose,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Conv3dTransposeGPUDNNKernel,
                   float,
                   double,
                   float16) {}
#endif

#endif