cpu_adaptive_max_pool3d_backward Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/AdaptiveMaxPoolKernel.cpp lines 831–879

template <typename scalar_t>
void cpu_adaptive_max_pool3d_backward(
    const Tensor& grad_input_,
    const Tensor& grad_output_,
    const Tensor& indices_) {
  auto grad_output = grad_output_.contiguous();
  auto indices = indices_.contiguous();
  auto grad_input = grad_input_.contiguous();

  auto grad_output_data = grad_output.data_ptr<scalar_t>();
  auto indices_data = indices.data_ptr<int64_t>();
  auto grad_input_data = grad_input.mutable_data_ptr<scalar_t>();

  int64_t ndim = grad_output.ndimension();
  // treat batch size and channels as one dimension
  int64_t channels = ndim == 3 ? grad_output.size(0) : grad_output.size(0) * grad_output.size(1);
  int64_t input_depth = grad_input.size(-3);
  int64_t input_height = grad_input.size(-2);
  int64_t input_width = grad_input.size(-1);
  int64_t output_depth = grad_output.size(-3);
  int64_t output_height = grad_output.size(-2);
  int64_t output_width = grad_output.size(-1);

  // parallel on dim of N, C
  at::parallel_for(0, channels, 0, [&](int64_t begin, int64_t end) {
    for (const auto c : c10::irange(begin, end)) {
      scalar_t* grad_input_ptr = grad_input_data + c * input_depth * input_height * input_width;
      scalar_t* grad_output_ptr = grad_output_data + c * output_depth * output_height * output_width;
      int64_t* indices_ptr = indices_data + c * output_depth * output_height * output_width;

      for (const auto od : c10::irange(output_depth)) {
        for (const auto oh : c10::irange(output_height)) {
          for (const auto ow : c10::irange(output_width)) {
            // retrieve position of max
            int64_t index = od * output_height * output_width + oh * output_width + ow;
            int64_t maxindex = indices_ptr[index];

            // update gradient
            grad_input_ptr[maxindex] += grad_output_ptr[index];
          }
        }
      }
    }
  });

  if (!grad_input_.is_contiguous()) {
    grad_input_.copy_(grad_input);
  }
}