pad3d_op.cu

/* Copyright (c) 2020 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 <algorithm>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/cuda_primitives.h"
#include "paddle/fluid/platform/gpu_info.h"

namespace paddle {
namespace operators {

using platform::PADDLE_CUDA_NUM_THREADS;

using framework::Tensor;

template <typename T>
__global__ void Pad3DConstNCDHW(const int nthreads, const T* in_data,
                                const int num, const int channels,
                                const int in_depth, const int in_height,
                                const int in_width, const int out_depth,
                                const int out_height, const int out_width,
                                const int pad_front, const int pad_top,
                                const int pad_left, T value, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int nc = index / out_width;

    const int out_w = index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = out_d - pad_front;
    int in_h = out_h - pad_top;
    int in_w = out_w - pad_left;
    out_data[index] =
        (in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
         in_h >= in_height || in_w >= in_width)
            ? value
            : in_data[nc * in_depth * in_height * in_width +
                      in_d * in_height * in_width + in_h * in_width + in_w];
  }
}

template <typename T>
__global__ void Pad3DConstNDHWC(const int nthreads, const T* in_data,
                                const int num, const int channels,
                                const int in_depth, const int in_height,
                                const int in_width, const int out_depth,
                                const int out_height, const int out_width,
                                const int pad_front, const int pad_top,
                                const int pad_left, T value, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int n = index / channels;
    const int c = index % channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;
    const int in_d = out_d - pad_front;
    const int in_h = out_h - pad_top;
    const int in_w = out_w - pad_left;

    out_data[index] =
        (in_d < 0 || in_h < 0 || in_w < 0 || in_d >= in_depth ||
         in_h >= in_height || in_w >= in_width)
            ? value
            : in_data[n * in_depth * in_height * in_width * channels +
                      in_d * in_height * in_width * channels +
                      in_h * in_width * channels + in_w * channels + c];
  }
}

template <typename T>
__global__ void Pad3DReflectNCDHW(const int nthreads, const T* in_data,
                                  const int num, const int channels,
                                  const int in_depth, const int in_height,
                                  const int in_width, const int out_depth,
                                  const int out_height, const int out_width,
                                  const int pad_front, const int pad_top,
                                  const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int nc = index / out_width;

    const int out_w = index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = out_d - pad_front;
    int in_h = out_h - pad_top;
    int in_w = out_w - pad_left;

    in_d = max(in_d, -in_d);                     // reflect by 0
    in_d = min(in_d, 2 * in_depth - in_d - 2);   // reflect by in_depth
    in_h = max(in_h, -in_h);                     // reflect by 0
    in_h = min(in_h, 2 * in_height - in_h - 2);  // reflect by in_height
    in_w = max(in_w, -in_w);                     // reflect by 0
    in_w = min(in_w, 2 * in_width - in_w - 2);   // reflect by in_width
    out_data[index] =
        in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
                    in_width +
                in_w];
  }
}

template <typename T>
__global__ void Pad3DReflectNDHWC(const int nthreads, const T* in_data,
                                  const int num, const int channels,
                                  const int in_depth, const int in_height,
                                  const int in_width, const int out_depth,
                                  const int out_height, const int out_width,
                                  const int pad_front, const int pad_top,
                                  const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int n = index / channels;
    const int c = index % channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;
    int in_d = out_d - pad_front;
    int in_h = out_h - pad_top;
    int in_w = out_w - pad_left;

    in_d = max(in_d, -in_d);
    in_d = min(in_d, 2 * in_depth - in_d - 2);
    in_h = max(in_h, -in_h);
    in_h = min(in_h, 2 * in_height - in_h - 2);
    in_w = max(in_w, -in_w);
    in_w = min(in_w, 2 * in_width - in_w - 2);

    out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
                              in_d * in_height * in_width * channels +
                              in_h * in_width * channels + in_w * channels + c];
  }
}

template <typename T>
__global__ void Pad3DReplicateNCDHW(const int nthreads, const T* in_data,
                                    const int num, const int channels,
                                    const int in_depth, const int in_height,
                                    const int in_width, const int out_depth,
                                    const int out_height, const int out_width,
                                    const int pad_front, const int pad_top,
                                    const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int nc = index / out_width;

    const int out_w = index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
    int in_h = min(in_height - 1, max(out_h - pad_top, 0));
    int in_w = min(in_width - 1, max(out_w - pad_left, 0));

    out_data[index] =
        in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
                    in_width +
                in_w];
  }
}

template <typename T>
__global__ void Pad3DReplicateNDHWC(const int nthreads, const T* in_data,
                                    const int num, const int channels,
                                    const int in_depth, const int in_height,
                                    const int in_width, const int out_depth,
                                    const int out_height, const int out_width,
                                    const int pad_front, const int pad_top,
                                    const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int n = index / channels;
    const int c = index % channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;

    int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
    int in_h = min(in_height - 1, max(out_h - pad_top, 0));
    int in_w = min(in_width - 1, max(out_w - pad_left, 0));

    out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
                              in_d * in_height * in_width * channels +
                              in_h * in_width * channels + in_w * channels + c];
  }
}

template <typename T>
__global__ void Pad3DCircularNCDHW(const int nthreads, const T* in_data,
                                   const int num, const int channels,
                                   const int in_depth, const int in_height,
                                   const int in_width, const int out_depth,
                                   const int out_height, const int out_width,
                                   const int pad_front, const int pad_top,
                                   const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int nc = index / out_width;

    const int out_w = index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
    int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
    int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;

    out_data[index] =
        in_data[(nc * in_depth * in_height + in_d * in_height + in_h) *
                    in_width +
                in_w];
  }
}

template <typename T>
__global__ void Pad3DCircularNDHWC(const int nthreads, const T* in_data,
                                   const int num, const int channels,
                                   const int in_depth, const int in_height,
                                   const int in_width, const int out_depth,
                                   const int out_height, const int out_width,
                                   const int pad_front, const int pad_top,
                                   const int pad_left, T* out_data) {
  CUDA_KERNEL_LOOP(index, nthreads) {
    int n = index / channels;
    const int c = index % channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;

    int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
    int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
    int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;

    out_data[index] = in_data[n * in_depth * in_height * in_width * channels +
                              in_d * in_height * in_width * channels +
                              in_h * in_width * channels + in_w * channels + c];
  }
}

template <typename T>
__global__ void Pad3DGradConstNCDHW(const int in_size, T* d_in_data,
                                    const int num, const int channels,
                                    const int in_depth, const int in_height,
                                    const int in_width, const int out_depth,
                                    const int out_height, const int out_width,
                                    const int pad_front, const int pad_top,
                                    const int pad_left, const T* d_out_data) {
  CUDA_KERNEL_LOOP(in_index, in_size) {
    const int in_w = in_index % in_width;

    int nc = in_index / in_width;
    const int in_h = nc % in_height;

    nc /= in_height;
    const int in_d = nc % in_depth;

    nc /= in_depth;

    const int out_d = in_d + pad_front;
    const int out_h = in_h + pad_top;
    const int out_w = in_w + pad_left;
    d_in_data[in_index] =
        d_out_data[nc * out_depth * out_height * out_width +
                   out_d * out_height * out_width + out_h * out_width + out_w];
  }
}

template <typename T>
__global__ void Pad3DGradConstNDHWC(const int in_size, T* d_in_data,
                                    const int num, const int channels,
                                    const int in_depth, const int in_height,
                                    const int in_width, const int out_depth,
                                    const int out_height, const int out_width,
                                    const int pad_front, const int pad_top,
                                    const int pad_left, const T* d_out_data) {
  CUDA_KERNEL_LOOP(in_index, in_size) {
    const int c = in_index % channels;
    int n = in_index / channels;

    const int in_w = n % in_width;
    n /= in_width;

    const int in_h = n % in_height;
    n /= in_height;

    const int in_d = n % in_depth;
    n /= in_depth;

    const int out_d = in_d + pad_front;
    const int out_h = in_h + pad_top;
    const int out_w = in_w + pad_left;

    d_in_data[in_index] =
        d_out_data[n * out_depth * out_height * out_width * channels +
                   out_d * out_height * out_width * channels +
                   out_h * out_width * channels + out_w * channels + c];
  }
}

template <typename T>
__global__ void Pad3DGradReflectNCDHW(const int out_size, T* d_in_data,
                                      const int num, const int channels,
                                      const int in_depth, const int in_height,
                                      const int in_width, const int out_depth,
                                      const int out_height, const int out_width,
                                      const int pad_front, const int pad_top,
                                      const int pad_left, const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    int nc = out_index / out_width;
    const int out_w = out_index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = out_d - pad_front;
    int in_h = out_h - pad_top;
    int in_w = out_w - pad_left;

    in_d = max(in_d, -in_d);
    in_h = max(in_h, -in_h);
    in_w = max(in_w, -in_w);

    in_d = min(in_d, 2 * in_depth - in_d - 2);
    in_h = min(in_h, 2 * in_height - in_h - 2);
    in_w = min(in_w, 2 * in_width - in_w - 2);

    platform::CudaAtomicAdd(
        &d_in_data[nc * in_depth * in_height * in_width +
                   in_d * in_height * in_width + in_h * in_width + in_w],
        d_out_data[out_index]);
  }
}

template <typename T>
__global__ void Pad3DGradReflectNDHWC(const int out_size, T* d_in_data,
                                      const int num, const int channels,
                                      const int in_depth, const int in_height,
                                      const int in_width, const int out_depth,
                                      const int out_height, const int out_width,
                                      const int pad_front, const int pad_top,
                                      const int pad_left, const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    const int c = out_index % channels;
    int n = out_index / channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;

    int in_d = out_d - pad_front;
    int in_h = out_h - pad_top;
    int in_w = out_w - pad_left;

    in_d = max(in_d, -in_d);
    in_h = max(in_h, -in_h);
    in_w = max(in_w, -in_w);

    in_d = min(in_d, in_depth * 2 - in_d - 2);
    in_h = min(in_h, in_height * 2 - in_h - 2);
    in_w = min(in_w, in_width * 2 - in_w - 2);
    platform::CudaAtomicAdd(
        &d_in_data[n * in_depth * in_height * in_width * channels +
                   in_d * in_height * in_width * channels +
                   in_h * in_width * channels + in_w * channels + c],
        d_out_data[out_index]);
  }
}

template <typename T>
__global__ void Pad3DGradReplicateNCDHW(
    const int out_size, T* d_in_data, const int num, const int channels,
    const int in_depth, const int in_height, const int in_width,
    const int out_depth, const int out_height, const int out_width,
    const int pad_front, const int pad_top, const int pad_left,
    const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    int nc = out_index / out_width;
    const int out_w = out_index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    const int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
    const int in_h = min(in_height - 1, max(out_h - pad_top, 0));
    const int in_w = min(in_width - 1, max(out_w - pad_left, 0));

    platform::CudaAtomicAdd(
        &d_in_data[nc * in_depth * in_height * in_width +
                   in_d * in_height * in_width + in_h * in_width + in_w],
        d_out_data[out_index]);
  }
}

template <typename T>
__global__ void Pad3DGradReplicateNDHWC(
    const int out_size, T* d_in_data, const int num, const int channels,
    const int in_depth, const int in_height, const int in_width,
    const int out_depth, const int out_height, const int out_width,
    const int pad_front, const int pad_top, const int pad_left,
    const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    const int c = out_index % channels;
    int n = out_index / channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;

    const int in_d = min(in_depth - 1, max(out_d - pad_front, 0));
    const int in_h = min(in_height - 1, max(out_h - pad_top, 0));
    const int in_w = min(in_width - 1, max(out_w - pad_left, 0));

    platform::CudaAtomicAdd(
        &d_in_data[n * in_depth * in_height * in_width * channels +
                   in_d * in_height * in_width * channels +
                   in_h * in_width * channels + in_w * channels + c],
        d_out_data[out_index]);
  }
}

template <typename T>
__global__ void Pad3DGradCircularNCDHW(const int out_size, T* d_in_data,
                                       const int num, const int channels,
                                       const int in_depth, const int in_height,
                                       const int in_width, const int out_depth,
                                       const int out_height,
                                       const int out_width, const int pad_front,
                                       const int pad_top, const int pad_left,
                                       const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    int nc = out_index / out_width;
    const int out_w = out_index % out_width;
    const int out_h = nc % out_height;
    nc /= out_height;
    const int out_d = nc % out_depth;
    nc /= out_depth;

    int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
    int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
    int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;

    platform::CudaAtomicAdd(
        &d_in_data[nc * in_depth * in_height * in_width +
                   in_d * in_height * in_width + in_h * in_width + in_w],
        d_out_data[out_index]);
  }
}

template <typename T>
__global__ void Pad3DGradCircularNDHWC(const int out_size, T* d_in_data,
                                       const int num, const int channels,
                                       const int in_depth, const int in_height,
                                       const int in_width, const int out_depth,
                                       const int out_height,
                                       const int out_width, const int pad_front,
                                       const int pad_top, const int pad_left,
                                       const T* d_out_data) {
  CUDA_KERNEL_LOOP(out_index, out_size) {
    const int c = out_index % channels;
    int n = out_index / channels;
    const int out_w = n % out_width;
    n /= out_width;
    const int out_h = n % out_height;
    n /= out_height;
    const int out_d = n % out_depth;
    n /= out_depth;

    int in_d = ((out_d - pad_front) % in_depth + in_depth) % in_depth;
    int in_h = ((out_h - pad_top) % in_height + in_height) % in_height;
    int in_w = ((out_w - pad_left) % in_width + in_width) % in_width;

    platform::CudaAtomicAdd(
        &d_in_data[n * in_depth * in_height * in_width * channels +
                   in_d * in_height * in_width * channels +
                   in_h * in_width * channels + in_w * channels + c],
        d_out_data[out_index]);
  }
}

static inline std::vector<int> GetPaddings(
    const framework::ExecutionContext& context) {
  std::vector<int> paddings(6);
  auto* paddings_data = context.Input<Tensor>("Paddings");
  if (paddings_data) {
    Tensor pads;
    framework::TensorCopySync(*paddings_data, platform::CPUPlace(), &pads);
    auto pads_data = pads.data<int>();
    std::memcpy(paddings.data(), pads_data, paddings.size() * sizeof(int));
  } else {
    auto pads = context.Attr<std::vector<int>>("paddings");
    std::copy(pads.begin(), pads.end(), paddings.data());
  }
  return paddings;
}

template <typename T>
class Pad3dCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    std::vector<int> pads = GetPaddings(context);
    auto mode = context.Attr<std::string>("mode");
    auto data_format = context.Attr<std::string>("data_format");
    T value = static_cast<T>(context.Attr<float>("value"));

    auto* x = context.Input<Tensor>("X");
    auto in_dims = x->dims();
    const T* in_data = x->data<T>();
    auto* out = context.Output<Tensor>("Out");
    auto out_dims = out->dims();
    if (data_format == "NCDHW") {
      out_dims[0] = in_dims[0];
      out_dims[1] = in_dims[1];
      out_dims[2] = in_dims[2] + pads[4] + pads[5];
      out_dims[3] = in_dims[3] + pads[2] + pads[3];
      out_dims[4] = in_dims[4] + pads[0] + pads[1];
    } else {
      out_dims[0] = in_dims[0];
      out_dims[1] = in_dims[1] + pads[4] + pads[5];
      out_dims[2] = in_dims[2] + pads[2] + pads[3];
      out_dims[3] = in_dims[3] + pads[0] + pads[1];
      out_dims[4] = in_dims[4];
    }
    T* out_data = out->mutable_data<T>(out_dims, context.GetPlace());

    int channels = in_dims[1];
    int in_depth = in_dims[2];
    int in_height = in_dims[3];
    int in_width = in_dims[4];
    int out_depth = out_dims[2];
    int out_height = out_dims[3];
    int out_width = out_dims[4];
    if (data_format == "NDHWC") {
      channels = in_dims[4];
      in_depth = in_dims[1];
      in_height = in_dims[2];
      in_width = in_dims[3];
      out_depth = out_dims[1];
      out_height = out_dims[2];
      out_width = out_dims[3];
    }

    if (mode == "reflect") {
      PADDLE_ENFORCE_GT(in_depth, pads[4],
                        platform::errors::InvalidArgument(
                            "The depth of Input(X)'s dimension should be "
                            "greater than pad_front"
                            " in reflect mode"
                            ", but received depth(%d) and pad_front(%d).",
                            in_depth, pads[4]));
      PADDLE_ENFORCE_GT(in_depth, pads[5],
                        platform::errors::InvalidArgument(
                            "The depth of Input(X)'s dimension should be "
                            "greater than pad_back"
                            " in reflect mode"
                            ", but received depth(%d) and pad_back(%d).",
                            in_depth, pads[5]));

      PADDLE_ENFORCE_GT(in_height, pads[2],
                        platform::errors::InvalidArgument(
                            "The height of Input(X)'s dimension should be "
                            "greater than pad_top"
                            " in reflect mode"
                            ", but received depth(%d) and pad_top(%d).",
                            in_height, pads[2]));
      PADDLE_ENFORCE_GT(in_height, pads[3],
                        platform::errors::InvalidArgument(
                            "The height of Input(X)'s dimension should be "
                            "greater than pad_bottom"
                            " in reflect mode"
                            ", but received depth(%d) and pad_bottom(%d).",
                            in_height, pads[3]));

      PADDLE_ENFORCE_GT(in_width, pads[0],
                        platform::errors::InvalidArgument(
                            "The width of Input(X)'s dimension should be "
                            "greater than pad_left"
                            " in reflect mode"
                            ", but received depth(%d) and pad_left(%d).",
                            in_width, pads[0]));
      PADDLE_ENFORCE_GT(in_width, pads[1],
                        platform::errors::InvalidArgument(
                            "The width of Input(X)'s dimension should be "
                            "greater than pad_right"
                            " in reflect mode"
                            ", but received depth(%d) and pad_right(%d).",
                            in_width, pads[1]));
    }

    const int pad_left = pads[0];
    const int pad_top = pads[2];
    const int pad_front = pads[4];
    const int num = in_dims[0];

    auto stream = context.cuda_device_context().stream();
    int block = PADDLE_CUDA_NUM_THREADS;
    const int out_size = out->numel();
    int grid = (out_size + block - 1) / block;

    if (data_format == "NCDHW") {
      if (mode == "reflect") {
        Pad3DReflectNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else if (mode == "replicate") {
        Pad3DReplicateNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else if (mode == "circular") {
        Pad3DCircularNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else {
        Pad3DConstNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            value, out_data);
      }
    } else {
      if (mode == "reflect") {
        Pad3DReflectNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else if (mode == "replicate") {
        Pad3DReplicateNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else if (mode == "circular") {
        Pad3DCircularNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            out_data);
      } else {
        Pad3DConstNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            value, out_data);
      }
    }
  }
};

template <typename T>
class Pad3dGradCUDAKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& context) const override {
    std::vector<int> pads = GetPaddings(context);
    auto mode = context.Attr<std::string>("mode");
    auto data_format = context.Attr<std::string>("data_format");
    auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
    auto* d_in = context.Output<Tensor>(framework::GradVarName("X"));
    auto d_in_dims = d_in->dims();
    auto d_out_dims = d_out->dims();
    const T* d_out_data = d_out->data<T>();
    T* d_in_data = d_in->mutable_data<T>(context.GetPlace());

    math::SetConstant<platform::CUDADeviceContext, T> set_zero;
    set_zero(context.template device_context<platform::CUDADeviceContext>(),
             d_in, static_cast<T>(0));

    const int pad_left = pads[0];
    const int pad_top = pads[2];
    const int pad_front = pads[4];

    const int num = d_in_dims[0];

    auto stream = context.cuda_device_context().stream();
    int block = PADDLE_CUDA_NUM_THREADS;
    const int out_size = d_out->numel();
    const int in_size = d_in->numel();
    int grid = (out_size + block - 1) / block;

    if (data_format == "NCDHW") {
      const int channels = d_in_dims[1];
      const int in_depth = d_in_dims[2];
      const int in_height = d_in_dims[3];
      const int in_width = d_in_dims[4];
      const int out_depth = d_out_dims[2];
      const int out_height = d_out_dims[3];
      const int out_width = d_out_dims[4];

      if (mode == "reflect") {
        Pad3DGradReflectNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else if (mode == "replicate") {
        Pad3DGradReplicateNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else if (mode == "circular") {
        Pad3DGradCircularNCDHW<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else {
        grid = (in_size + block - 1) / block;
        Pad3DGradConstNCDHW<T><<<grid, block, 0, stream>>>(
            in_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      }
    } else {
      const int channels = d_in_dims[4];
      const int in_depth = d_in_dims[1];
      const int in_height = d_in_dims[2];
      const int in_width = d_in_dims[3];
      const int out_depth = d_out_dims[1];
      const int out_height = d_out_dims[2];
      const int out_width = d_out_dims[3];
      if (mode == "reflect") {
        Pad3DGradReflectNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else if (mode == "replicate") {
        Pad3DGradReplicateNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else if (mode == "circular") {
        Pad3DGradCircularNDHWC<T><<<grid, block, 0, stream>>>(
            out_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      } else {
        grid = (in_size + block - 1) / block;
        Pad3DGradConstNDHWC<T><<<grid, block, 0, stream>>>(
            in_size, d_in_data, num, channels, in_depth, in_height, in_width,
            out_depth, out_height, out_width, pad_front, pad_top, pad_left,
            d_out_data);
      }
    }
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
namespace plat = paddle::platform;

REGISTER_OP_CUDA_KERNEL(pad3d, ops::Pad3dCUDAKernel<plat::float16>,
                        ops::Pad3dCUDAKernel<float>,
                        ops::Pad3dCUDAKernel<double>, ops::Pad3dCUDAKernel<int>,
                        ops::Pad3dCUDAKernel<int64_t>);
REGISTER_OP_CUDA_KERNEL(pad3d_grad, ops::Pad3dGradCUDAKernel<plat::float16>,
                        ops::Pad3dGradCUDAKernel<float>,
                        ops::Pad3dGradCUDAKernel<double>);