value Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/AvgPoolKernel.cpp lines 102–214

template <typename scalar_t,
          std::enable_if_t<!is_reduced_floating_point<scalar_t>::value, int> = 0>
void cpu_avg_pool2d_channels_last(
    const Tensor& output_,
    const Tensor& input_,
    int64_t kW, int64_t kH,
    int64_t dW, int64_t dH,
    int64_t padW, int64_t padH,
    bool count_include_pad,
    std::optional<int64_t> divisor_override) {
  TORCH_CHECK(input_.ndimension() == 4,
              "2d average pooling with channels last format supports tensors with 4 dims");
  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(2);
  int64_t output_width = output.size(3);

  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);

    int64_t size = channels;
    int64_t len = size - (size % Vec::size());
    for (const auto i : c10::irange(begin, end)) {
      // compute the mean of the input image...
      int64_t ih0 = oh * dH - padH;
      int64_t iw0 = ow * dW - padW;
      int64_t ih1 = std::min(ih0 + kH, input_height + padH);
      int64_t iw1 = std::min(iw0 + kW, input_width + padW);
      int64_t pool_size = (ih1 - ih0) * (iw1 - iw0);
      ih0 = std::max(ih0, (int64_t) 0);
      iw0 = std::max(iw0, (int64_t) 0);
      ih1 = std::min(ih1, input_height);
      iw1 = std::min(iw1, input_width);

      int64_t divide_factor = 0;
      if (divisor_override.has_value()) {
        divide_factor = divisor_override.value();
      } else {
        if(count_include_pad) {
          divide_factor = pool_size;
        } else {
          divide_factor = (ih1 - ih0) * (iw1 - iw0);
        }
      }

      scalar_t* out = output_data + i * channels;

      // Pass I: zero the out lane
      int64_t d1 = 0;
      for (; d1 < len; d1 += Vec::size()) {
        Vec out_vec = Vec(scalar_t(0));
        out_vec.store(out + d1);
      }
      for (; d1 < size; d1++) {
        out[d1] = scalar_t(0);
      }

      if (ih0 >= ih1 || iw0 >= iw1) {
        // move on to next output index
        data_index_step(n, nbatch, oh, output_height, ow, output_width);
        continue;
      }

      // 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 < len; 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 < len; d3 += Vec::size()) {
        Vec out_vec = Vec::loadu(out + d3) / Vec(scalar_t(divide_factor));
        out_vec.store(out + d3);
      }
      for (; d3 < size; d3++) {
        out[d3] = out[d3] / divide_factor;
      }

      // 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);
  }
}