aten/src/ATen/native/UpSampleNearest3d.cpp

#include <ATen/ATen.h>
#include <ATen/NativeFunctions.h>
#include <ATen/native/UpSample.h>

namespace at {
namespace native {
namespace {

template <typename scalar_t>
static void upsample_nearest3d_out_frame(
    scalar_t* odata,
    scalar_t* idata,
    int64_t input_depth,
    int64_t input_height,
    int64_t input_width,
    int64_t output_depth,
    int64_t output_height,
    int64_t output_width,
    int64_t nbatch,
    int64_t channels) {
  const float depth_scale = (float)input_depth / (float)output_depth;
  const float height_scale = (float)input_height / (float)output_height;
  const float width_scale = (float)input_width / (float)output_width;

  channels = channels * nbatch;

  // special case: just copy
  if (input_depth == output_depth && input_height == output_height &&
      input_width == output_width) {
    for (int64_t d2 = 0; d2 < output_depth; ++d2) {
      const int64_t d1 = d2;

      for (int64_t h2 = 0; h2 < output_height; ++h2) {
        const int64_t h1 = h2;

        for (int64_t w2 = 0; w2 < output_width; ++w2) {
          const int64_t w1 = w2;
          const scalar_t* pos1 =
              &idata[d1 * input_height * input_width + h1 * input_width + w1];
          scalar_t* pos2 =
              &odata
                  [d2 * output_height * output_width + h2 * output_width + w2];

          for (int64_t c = 0; c < channels; ++c) {
            pos2[0] = pos1[0];
            pos1 += input_depth * input_height * input_width;
            pos2 += output_depth * output_height * output_width;
          }
        }
      }
    }
    return;
  }

  for (int64_t d2 = 0; d2 < output_depth; ++d2) {
    const int64_t d1 =
        nearest_neighbor_compute_source_index(depth_scale, d2, input_depth);

    for (int64_t h2 = 0; h2 < output_height; ++h2) {
      const int64_t h1 =
          nearest_neighbor_compute_source_index(height_scale, h2, input_height);

      for (int64_t w2 = 0; w2 < output_width; ++w2) {
        const int64_t w1 =
            nearest_neighbor_compute_source_index(width_scale, w2, input_width);
        const scalar_t* pos1 =
            &idata[d1 * input_height * input_width + h1 * input_width + w1];
        scalar_t* pos2 =
            &odata[d2 * output_height * output_width + h2 * output_width + w2];

        for (int64_t c = 0; c < channels; ++c) {
          pos2[0] = pos1[0];
          pos1 += input_depth * input_height * input_width;
          pos2 += output_depth * output_height * output_width;
        }
      }
    }
  }
}

template <typename scalar_t>
static void upsample_nearest3d_backward_out_frame(
    scalar_t* odata,
    scalar_t* idata,
    int64_t input_depth,
    int64_t input_height,
    int64_t input_width,
    int64_t output_depth,
    int64_t output_height,
    int64_t output_width,
    int64_t nbatch,
    int64_t channels) {
  const float depth_scale = (float)input_depth / (float)output_depth;
  const float height_scale = (float)input_height / (float)output_height;
  const float width_scale = (float)input_width / (float)output_width;

  channels = channels * nbatch;

  // special case: just copy
  if (input_depth == output_depth && input_height == output_height &&
      input_width == output_width) {
    for (int64_t d2 = 0; d2 < output_depth; ++d2) {
      const int64_t d1 = d2;

      for (int64_t h2 = 0; h2 < output_height; ++h2) {
        const int64_t h1 = h2;

        for (int64_t w2 = 0; w2 < output_width; ++w2) {
          const int64_t w1 = w2;
          scalar_t* pos1 =
              &idata[d1 * input_height * input_width + h1 * input_width + w1];
          const scalar_t* pos2 =
              &odata
                  [d2 * output_height * output_width + h2 * output_width + w2];

          for (int64_t c = 0; c < channels; ++c) {
            pos1[0] += pos2[0];
            pos1 += input_depth * input_height * input_width;
            pos2 += output_depth * output_height * output_width;
          }
        }
      }
    }
    return;
  }

  for (int64_t d2 = 0; d2 < output_depth; ++d2) {
    const int64_t d1 =
        nearest_neighbor_compute_source_index(depth_scale, d2, input_depth);

    for (int64_t h2 = 0; h2 < output_height; ++h2) {
      const int64_t h1 =
          nearest_neighbor_compute_source_index(height_scale, h2, input_height);

      for (int64_t w2 = 0; w2 < output_width; ++w2) {
        const int64_t w1 =
            nearest_neighbor_compute_source_index(width_scale, w2, input_width);
        scalar_t* pos1 =
            &idata[d1 * input_height * input_width + h1 * input_width + w1];
        const scalar_t* pos2 =
            &odata[d2 * output_height * output_width + h2 * output_width + w2];

        for (int64_t c = 0; c < channels; ++c) {
          pos1[0] += pos2[0];
          pos1 += input_depth * input_height * input_width;
          pos2 += output_depth * output_height * output_width;
        }
      }
    }
  }
}

static void upsample_nearest3d_out_cpu_template(
    Tensor& output,
    const Tensor& input_,
    IntArrayRef output_size) {
  TORCH_CHECK(
      output_size.size() == 3,
      "It is expected output_size equals to 3, but got size ",
      output_size.size());

  int64_t output_depth = output_size[0];
  int64_t output_height = output_size[1];
  int64_t output_width = output_size[2];

  int64_t nbatch = input_.size(0);
  int64_t channels = input_.size(1);
  int64_t input_depth = input_.size(2);
  int64_t input_height = input_.size(3);
  int64_t input_width = input_.size(4);

  upsample_3d_shape_check(
      input_,
      Tensor(),
      nbatch,
      channels,
      input_depth,
      input_height,
      input_width,
      output_depth,
      output_height,
      output_width);

  auto input = input_.contiguous();

  output.resize_({nbatch, channels, output_depth, output_height, output_width});
  output.zero_();

  AT_ASSERT(
      input_depth > 0 && input_height > 0 && input_width > 0 &&
      output_depth > 0 && output_height > 0 && output_width > 0);

  AT_DISPATCH_FLOATING_TYPES_AND_HALF(input.scalar_type(), "upsample_nearest3d", [&] {
    auto* idata = input.data_ptr<scalar_t>();
    auto* odata = output.data_ptr<scalar_t>();

    upsample_nearest3d_out_frame<scalar_t>(
        odata,
        idata,
        input_depth,
        input_height,
        input_width,
        output_depth,
        output_height,
        output_width,
        nbatch,
        channels);
  });
}

static void upsample_nearest3d_backward_out_cpu_template(
    Tensor& grad_input,
    const Tensor& grad_output_,
    IntArrayRef output_size,
    IntArrayRef input_size) {
  TORCH_CHECK(
      output_size.size() == 3,
      "It is expected output_size equals to 3, but got size ",
      output_size.size());

  TORCH_CHECK(
      input_size.size() == 5,
      "It is expected input_size equals to 5, but got size ",
      input_size.size());

  int64_t output_depth = output_size[0];
  int64_t output_height = output_size[1];
  int64_t output_width = output_size[2];

  int64_t nbatch = input_size[0];
  int64_t channels = input_size[1];
  int64_t input_depth = input_size[2];
  int64_t input_height = input_size[3];
  int64_t input_width = input_size[4];

  upsample_3d_shape_check(
      Tensor(),
      grad_output_,
      nbatch,
      channels,
      input_depth,
      input_height,
      input_width,
      output_depth,
      output_height,
      output_width);

  grad_input.resize_(
      {nbatch, channels, input_depth, input_height, input_width});
  grad_input.zero_();

  auto grad_output = grad_output_.contiguous();

  AT_DISPATCH_FLOATING_TYPES_AND_HALF(
      grad_output.scalar_type(), "upsample_nearest3d_backward", [&] {
        scalar_t* idata = grad_input.data_ptr<scalar_t>();
        scalar_t* odata = grad_output.data_ptr<scalar_t>();

        upsample_nearest3d_backward_out_frame<scalar_t>(
            odata,
            idata,
            input_depth,
            input_height,
            input_width,
            output_depth,
            output_height,
            output_width,
            nbatch,
            channels);
      });
}
} // namespace

Tensor& upsample_nearest3d_out_cpu(
    Tensor& output,
    const Tensor& input,
    IntArrayRef output_size) {
  upsample_nearest3d_out_cpu_template(output, input, output_size);
  return output;
}

Tensor upsample_nearest3d_cpu(const Tensor& input, IntArrayRef output_size) {
  auto output = at::empty({0}, input.options());
  upsample_nearest3d_out_cpu_template(output, input, output_size);
  return output;
}

Tensor& upsample_nearest3d_backward_out_cpu(
    Tensor& grad_input,
    const Tensor& grad_output,
    IntArrayRef output_size,
    IntArrayRef input_size) {
  upsample_nearest3d_backward_out_cpu_template(
      grad_input, grad_output, output_size, input_size);
  return grad_input;
}

Tensor upsample_nearest3d_backward_cpu(
    const Tensor& grad_output,
    IntArrayRef output_size,
    IntArrayRef input_size) {
  auto grad_input = at::zeros(input_size, grad_output.options());
  upsample_nearest3d_backward_out_cpu_template(
      grad_input, grad_output, output_size, input_size);
  return grad_input;
}

} // namespace native
} // namespace at