pool_grad_kernel.cu

/* 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/fluid/platform/device/gpu/gpu_dnn.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/kernel_registry.h"
#include "paddle/phi/kernels/funcs/math_function.h"
#include "paddle/phi/kernels/funcs/pooling.h"
#include "paddle/phi/kernels/gpudnn/pool_gpudnn.h"
#include "paddle/phi/kernels/pool_grad_kernel.h"
#include "paddle/phi/kernels/pool_kernel.h"

#ifdef PADDLE_WITH_HIP
#include "paddle/phi/kernels/impl/pool_grad_kernel_impl.h"  //  PoolGradRawGPUDNNKernel will call PoolGradRawKernel for pooling type "max" in ROCm
#endif

namespace phi {

template <typename T, typename Context>
void PoolGradRawGPUDNNKernel(const Context& ctx,
                             const DenseTensor& x,
                             const DenseTensor& out,
                             const DenseTensor& dout,
                             const std::vector<int>& kernel_size,
                             const std::vector<int>& strides,
                             const std::vector<int>& paddings,
                             bool exclusive,
                             const std::string& data_format,
                             const std::string& pooling_type,
                             bool global_pooling,
                             bool adaptive,
                             const std::string& padding_algorithm,
                             DenseTensor* dx) {
  PADDLE_ENFORCE_EQ(
      paddle::platform::is_gpu_place(ctx.GetPlace()),
      true,
      errors::InvalidArgument("Pool operator CUDA kernel must use CUDAPlace "
                              "rather than CPUPlace."));

  const DenseTensor* input = &x;
  const DenseTensor* output = &out;
  const DenseTensor* output_grad = &dout;
  DenseTensor* input_grad = dx;
  std::vector<int> paddings_ = paddings;
  std::vector<int> kernel_size_ = kernel_size;

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

#ifdef PADDLE_WITH_HIP
  if (pooling_type == "max") {
    PoolGradRawKernel<T, GPUContext>(ctx,
                                     x,
                                     out,
                                     dout,
                                     kernel_size,
                                     strides,
                                     paddings_,
                                     exclusive,
                                     data_format,
                                     pooling_type,
                                     global_pooling,
                                     adaptive,
                                     padding_algorithm,
                                     dx);
    return;
  }
#endif

  // update paddings
  auto in_x_dims = input->dims();
  DDim data_dims;
  if (channel_last) {
    data_dims = slice_ddim(in_x_dims, 1, in_x_dims.size() - 1);
  } else {
    data_dims = slice_ddim(in_x_dims, 2, in_x_dims.size());
  }
  funcs::UpdatePadding(&paddings_,
                       global_pooling,
                       adaptive,
                       padding_algorithm,
                       data_dims,
                       strides,
                       kernel_size_);
  if (data_dims.size() * 2 == static_cast<int>(paddings_.size())) {
    for (int i = 0; i < data_dims.size(); ++i) {
      paddings_.erase(paddings_.begin() + i + 1);
    }
  }

  if (global_pooling) {
    funcs::UpdateKernelSize(&kernel_size_, data_dims);
  }

  // ------- tensor grad --------------
  DenseTensor transformed_input(input->type());
  DenseTensor transformed_output(output->type());
  DenseTensor transformed_output_grad(output_grad->type());

  ctx.template Alloc<T>(input_grad);
  DenseTensor transformed_input_grad(input_grad->type());
  GPUDNNDataLayout layout;
  const std::string str_NCHW = "NCHW", str_NHWC = "NHWC";
  const std::string str_NCDHW = "NCDHW", str_NDHWC = "NDHWC";
  if (data_format == str_NDHWC) {
    layout = GPUDNNDataLayout::kNCDHW;
    std::vector<int> axis{0, 4, 1, 2, 3};

    // input
    transformed_input.Resize(input->dims());
    auto in_dims_vec = vectorize(input->dims());
    in_dims_vec[1] = input->dims()[4];
    in_dims_vec[2] = input->dims()[1];
    in_dims_vec[3] = input->dims()[2];
    in_dims_vec[4] = input->dims()[3];
    transformed_input.Resize(make_ddim(in_dims_vec));
    ctx.Alloc(&transformed_input, input->type());

    funcs::Transpose<Context, T, 5> trans5;
    trans5(ctx, *input, &transformed_input, axis);

    // output
    transformed_output.Resize(output->dims());
    auto out_dims_vec = vectorize(output->dims());
    out_dims_vec[1] = output->dims()[4];
    out_dims_vec[2] = output->dims()[1];
    out_dims_vec[3] = output->dims()[2];
    out_dims_vec[4] = output->dims()[3];
    transformed_output.Resize(make_ddim(out_dims_vec));

    ctx.Alloc(&transformed_output, output->type());

    funcs::Transpose<Context, T, 5> trans5_v2;
    trans5_v2(ctx, *output, &transformed_output, axis);

    // output grad
    transformed_output_grad.Resize(make_ddim(out_dims_vec));
    ctx.Alloc(&transformed_output_grad, output_grad->type());

    funcs::Transpose<Context, T, 5> trans5_v3;
    trans5_v3(ctx, *output_grad, &transformed_output_grad, axis);

    // input grad
    transformed_input_grad.Resize(make_ddim(in_dims_vec));

#ifdef PADDLE_WITH_HIP
    // MIOPEN not support NHWC data layout
  } else if (data_format == str_NHWC) {
    layout = GPUDNNDataLayout::kNCHW;

    std::vector<int> axis{0, 3, 1, 2};

    // input
    transformed_input.Resize(input->dims());
    auto in_dims_vec = vectorize(input->dims());
    in_dims_vec[1] = input->dims()[3];
    in_dims_vec[2] = input->dims()[1];
    in_dims_vec[3] = input->dims()[2];
    transformed_input.Resize(make_ddim(in_dims_vec));
    ctx.Alloc(&transformed_input, input->type());

    funcs::Transpose<Context, T, 4> trans4;
    trans4(ctx, *input, &transformed_input, axis);

    // output
    transformed_output.Resize(output->dims());
    auto out_dims_vec = vectorize(output->dims());
    out_dims_vec[1] = output->dims()[3];
    out_dims_vec[2] = output->dims()[1];
    out_dims_vec[3] = output->dims()[2];
    transformed_output.Resize(make_ddim(out_dims_vec));
    ctx.Alloc(&transformed_output, output->type());

    funcs::Transpose<Context, T, 4> trans4_v2;
    trans4_v2(ctx, *output, &transformed_output, axis);

    // output grad
    transformed_output_grad.Resize(make_ddim(out_dims_vec));
    ctx.Alloc(&transformed_output_grad, output_grad->type());

    funcs::Transpose<Context, T, 4> trans4_v3;
    trans4_v3(ctx, *output_grad, &transformed_output_grad, axis);

    // input grad
    transformed_input_grad.Resize(make_ddim(in_dims_vec));
#endif
  } else {
    layout = GetLayoutFromStr(data_format);
    transformed_input = *input;
    transformed_output = *output;
    transformed_output_grad = *output_grad;
    transformed_input_grad = *input_grad;
  }

  const T* input_data = transformed_input.data<T>();
  const T* output_data = transformed_output.data<T>();
  const T* output_grad_data = transformed_output_grad.data<T>();

  // ------------------- cudnn descriptors ---------------------
  ScopedTensorDescriptor input_desc;
  ScopedTensorDescriptor output_desc;
  ScopedPoolingDescriptor pool_desc;

#ifdef PADDLE_WITH_HIP
  miopenTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
      layout, vectorize<int>(transformed_input.dims()));
  miopenTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
      layout, vectorize<int>(transformed_output.dims()));
#else
  cudnnTensorDescriptor_t cudnn_input_desc = input_desc.descriptor<T>(
      layout, vectorize<int>(transformed_input.dims()));
  cudnnTensorDescriptor_t cudnn_output_desc = output_desc.descriptor<T>(
      layout, vectorize<int>(transformed_output.dims()));
#endif
  PoolingMode pooling_mode;
  if (pooling_type == "max") {
    if (FLAGS_cudnn_deterministic) {
      pooling_mode = PoolingMode::kMaximumDeterministic;
    } else {
      pooling_mode = PoolingMode::kMaximum;
    }
  } else {
    pooling_mode = exclusive ? PoolingMode::kAverageExclusive
                             : PoolingMode::kAverageInclusive;
  }

#ifdef PADDLE_WITH_HIP
  miopenPoolingDescriptor_t cudnn_pool_desc =
      pool_desc.descriptor(pooling_mode, kernel_size_, paddings_, strides);
#else
  cudnnPoolingDescriptor_t cudnn_pool_desc =
      pool_desc.descriptor(pooling_mode, kernel_size_, paddings_, strides);
#endif

  // ------------------- cudnn pool algorithm ---------------------
  auto handle = ctx.cudnn_handle();
  ScalingParamType<T> alpha = 1.0f, beta = 0.0f;
  if (input_grad) {
    T* input_grad_data = ctx.template Alloc<T>(&transformed_input_grad);
// Because beta is zero, it is unnecessary to reset input_grad.
#ifdef PADDLE_WITH_HIP
    char* pool_workspace;
    size_t pool_worksize = 0;
    PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenPoolingGetWorkSpaceSizeV2(
        cudnn_pool_desc, cudnn_output_desc, &pool_worksize));
    PADDLE_ENFORCE_GPU_SUCCESS(hipMalloc(&pool_workspace, pool_worksize));
    PADDLE_ENFORCE_GPU_SUCCESS(dynload::miopenPoolingBackward(handle,
                                                              cudnn_pool_desc,
                                                              &alpha,
                                                              cudnn_output_desc,
                                                              output_data,
                                                              cudnn_output_desc,
                                                              output_grad_data,
                                                              cudnn_input_desc,
                                                              input_data,
                                                              &beta,
                                                              cudnn_input_desc,
                                                              input_grad_data,
                                                              pool_workspace));
    PADDLE_ENFORCE_GPU_SUCCESS(hipFree(pool_workspace));
#else
    PADDLE_ENFORCE_GPU_SUCCESS(dynload::cudnnPoolingBackward(handle,
                                                             cudnn_pool_desc,
                                                             &alpha,
                                                             cudnn_output_desc,
                                                             output_data,
                                                             cudnn_output_desc,
                                                             output_grad_data,
                                                             cudnn_input_desc,
                                                             input_data,
                                                             &beta,
                                                             cudnn_input_desc,
                                                             input_grad_data));
#endif

    if (data_format == str_NDHWC) {
      std::vector<int> axis{0, 2, 3, 4, 1};
      funcs::Transpose<Context, T, 5> trans5_v4;
      trans5_v4(ctx, transformed_input_grad, input_grad, axis);
    }
#ifdef PADDLE_WITH_HIP
    // MIOPEN not support NHWC data layout
    if (data_format == str_NHWC) {
      std::vector<int> axis{0, 2, 3, 1};
      funcs::Transpose<Context, T, 4> trans4_v4;
      trans4_v4(ctx, transformed_input_grad, input_grad, axis);
    }
#endif
  }
}

template <typename T, typename Context>
void Pool2dGradGPUDNNKernel(const Context& ctx,
                            const DenseTensor& x,
                            const DenseTensor& out,
                            const DenseTensor& dout,
                            const std::vector<int>& kernel_size,
                            const std::vector<int>& strides,
                            const std::vector<int>& paddings,
                            bool ceil_mode,
                            bool exclusive,
                            const std::string& data_format,
                            const std::string& pooling_type,
                            bool global_pooling,
                            bool adaptive,
                            const std::string& padding_algorithm,
                            DenseTensor* dx) {
  PoolGradRawGPUDNNKernel<T, Context>(ctx,
                                      x,
                                      out,
                                      dout,
                                      kernel_size,
                                      strides,
                                      paddings,
                                      exclusive,
                                      data_format,
                                      pooling_type,
                                      global_pooling,
                                      adaptive,
                                      padding_algorithm,
                                      dx);
}

template <typename T, typename Context>
void Pool2dDoubleGradGPUDNNKernel(const Context& ctx,
                                  const DenseTensor& x,
                                  const std::vector<int>& kernel_size,
                                  const std::vector<int>& strides,
                                  const std::vector<int>& paddings,
                                  bool ceil_mode,
                                  bool exclusive,
                                  const std::string& data_format,
                                  const std::string& pooling_type,
                                  bool global_pooling,
                                  bool adaptive,
                                  const std::string& padding_algorithm,
                                  DenseTensor* out) {
  if (pooling_type == "max") {
    PADDLE_THROW(
        errors::InvalidArgument("Pool op grad grad only supports avgpool."));
  } else {
    Pool2dGPUDNNKernel<T, Context>(ctx,
                                   x,
                                   kernel_size,
                                   strides,
                                   paddings,
                                   ceil_mode,
                                   exclusive,
                                   data_format,
                                   pooling_type,
                                   global_pooling,
                                   adaptive,
                                   padding_algorithm,
                                   out);
  }
}

template <typename T, typename Context>
void Pool3dGradGPUDNNKernel(const Context& ctx,
                            const DenseTensor& x,
                            const DenseTensor& out,
                            const DenseTensor& dout,
                            const std::vector<int>& kernel_size,
                            const std::vector<int>& strides,
                            const std::vector<int>& paddings,
                            bool ceil_mode,
                            bool exclusive,
                            const std::string& data_format,
                            const std::string& pooling_type,
                            bool global_pooling,
                            bool adaptive,
                            const std::string& padding_algorithm,
                            DenseTensor* dx) {
  PoolGradRawGPUDNNKernel<T, Context>(ctx,
                                      x,
                                      out,
                                      dout,
                                      kernel_size,
                                      strides,
                                      paddings,
                                      exclusive,
                                      data_format,
                                      pooling_type,
                                      global_pooling,
                                      adaptive,
                                      padding_algorithm,
                                      dx);
}

}  // namespace phi

using phi::dtype::float16;

#ifdef PADDLE_WITH_HIP
// MIOPEN do not support double
PD_REGISTER_KERNEL(pool2d_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool2dGradGPUDNNKernel,
                   float,
                   float16) {}
PD_REGISTER_KERNEL(pool2d_double_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool2dDoubleGradGPUDNNKernel,
                   float,
                   float16) {}
PD_REGISTER_KERNEL(pool3d_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool3dGradGPUDNNKernel,
                   float,
                   float16) {}
#else
PD_REGISTER_KERNEL(pool2d_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool2dGradGPUDNNKernel,
                   float,
                   double,
                   float16) {}
PD_REGISTER_KERNEL(pool2d_double_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool2dDoubleGradGPUDNNKernel,
                   float,
                   double,
                   float16) {}
PD_REGISTER_KERNEL(pool3d_grad,
                   GPUDNN,
                   ALL_LAYOUT,
                   phi::Pool3dGradGPUDNNKernel,
                   float,
                   double,
                   float16) {}
#endif