cpu_pixel_shuffle Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/PixelShuffleKernel.cpp lines 15–53

template <typename scalar_t>
void cpu_pixel_shuffle(
    TensorBase& output,
    const TensorBase& input,
    int64_t upscale_factor) {
  auto input_data = input.const_data_ptr<scalar_t>();
  auto output_data = output.data_ptr<scalar_t>();

  // [(B1...Bn), C, H, W] => [N, C, H, W]
  int64_t channels = input.size(-3);
  int64_t height = input.size(-2);
  int64_t width = input.size(-1);
  int64_t sub_channels = channels / (upscale_factor * upscale_factor);
  int64_t numel = input.numel();
  int64_t nbatch = numel / (channels * height * width);
  int64_t S = upscale_factor;

  // input strides
  int64_t stride_n = channels * height * width;
  int64_t stride_c = S * S * height * width;
  int64_t stride_s1 = S * height * width;
  int64_t stride_s2 = height * width;
  int64_t stride_h = width;

  // input tensor shape of [n, c, s1, s2, h, w]
  // output tensor shape of [n, c, h, s1, w, s2]
  at::parallel_for(0, numel, 0, [&](int64_t begin, int64_t end) {
    int64_t n{0}, c{0}, h{0}, s1{0}, w{0}, s2{0};
    data_index_init(begin, n, nbatch, c, sub_channels, h, height, s1, S, w, width, s2, S);

    for (const auto i : c10::irange(begin, end)) {
      int64_t input_offset = n * stride_n + c * stride_c + s1 * stride_s1 +
          s2 * stride_s2 + h * stride_h + w;
      output_data[i] = c10::load(&input_data[input_offset]);

      data_index_step(n, nbatch, c, sub_channels, h, height, s1, S, w, width, s2, S);
    }
  });
}