scalar_t Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/AdaptiveAvgPoolKernel.cpp lines 70–155

template <typename scalar_t>
typename std::enable_if_t<std::is_same_v<scalar_t, at::opmath_type<scalar_t>>, void>
cpu_adaptive_avg_pool2d_channels_last(
    Tensor& output_,
    const Tensor& input_,
    IntArrayRef output_size) {
  auto memory_format = at::MemoryFormat::ChannelsLast;
  auto input = input_.contiguous(memory_format);
  auto output = output_.contiguous(memory_format);

  auto input_data = input.const_data_ptr<scalar_t>();
  auto output_data = output.data_ptr<scalar_t>();

  int64_t nbatch = input.size(0);
  int64_t channels = input.size(1);
  int64_t input_height = input.size(2);
  int64_t input_width = input.size(3);
  int64_t output_height = output_size[0];
  int64_t output_width = output_size[1];

  using Vec = vec::Vectorized<scalar_t>;
  // parallel on dim N, H, W
  at::parallel_for(0, nbatch * output_height * output_width, 0, [&](int64_t begin, int64_t end) {
    int64_t n = 0;
    int64_t oh = 0;
    int64_t ow = 0;
    data_index_init(begin, n, nbatch, oh, output_height, ow, output_width);

    for (const auto i : c10::irange(begin, end)) {
      int64_t ih0 = start_index(oh, output_height, input_height);
      int64_t ih1 = end_index(oh, output_height, input_height);
      int64_t kh = ih1 - ih0;

      int64_t iw0 = start_index(ow, output_width, input_width);
      int64_t iw1 = end_index(ow, output_width, input_width);
      int64_t kw = iw1 - iw0;

      scalar_t* out = output_data + i * channels;
      int64_t size = channels;

      // Note: For ordinary usage scenario, each out lane should
      //   fit in L1 cache; otherwise consider block dim C.
      // Pass I: zero the out lane
      int64_t d1 = 0;
      for (; d1 < size - (size % Vec::size()); d1 += Vec::size()) {
        Vec out_vec = Vec(scalar_t(0));
        out_vec.store(out + d1);
      }
      for (; d1 < size; d1++) {
        out[d1] = scalar_t(0);
      }
      // Pass II: compute local sum
      for (const auto ih : c10::irange(ih0, ih1)) {
        for (const auto iw : c10::irange(iw0, iw1)) {
          const scalar_t* in = input_data + n * input_height * input_width * channels +
              ih * input_width * channels + iw * channels;

          int64_t d2 = 0;
          for (; d2 < size - (size % Vec::size()); d2 += Vec::size()) {
            Vec out_vec = Vec::loadu(out + d2) + Vec::loadu(in + d2);
            out_vec.store(out + d2);
          }
          for (; d2 < size; d2++) {
            out[d2] += in[d2];
          }
        }
      }
      // Pass III: compute local average
      int64_t d3 = 0;
      for (; d3 < size - (size % Vec::size()); d3 += Vec::size()) {
        Vec out_vec = Vec::loadu(out + d3) / Vec(scalar_t(kh * kw));
        out_vec.store(out + d3);
      }
      for (; d3 < size; d3++) {
        out[d3] = out[d3] / kh / kw;
      }

      // move on to next output index
      data_index_step(n, nbatch, oh, output_height, ow, output_width);
    }
  });

  if (!output_.is_contiguous(memory_format)) {
    output_.copy_(output);
  }
}