QConvPackWeightInt8Cudnn Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cudnn/ConvPrepack.cpp lines 125–157

class QConvPackWeightInt8Cudnn final {
 public:
  static c10::intrusive_ptr<ConvPackedParamsBase<kSpatialDim>> run_conv(
      Tensor weight,
      std::optional<Tensor> bias,
      torch::List<int64_t> stride,
      torch::List<int64_t> padding,
      torch::List<int64_t> dilation,
      int64_t groups) {
    torch::List<int64_t> output_padding;
    output_padding.reserve(kSpatialDim);
    for ([[maybe_unused]] const auto idx : c10::irange(kSpatialDim)) {
      output_padding.push_back(0);
    }
    return _run(weight, bias, stride, padding, output_padding, dilation, groups,
                /*transpose=*/false);
  }

 private:
  static c10::intrusive_ptr<ConvPackedParamsBase<kSpatialDim>> _run(
      Tensor weight,
      std::optional<Tensor> bias,
      torch::List<int64_t> stride,
      torch::List<int64_t> padding,
      torch::List<int64_t> output_padding,
      torch::List<int64_t> dilation,
      int64_t groups,
      bool transpose) {
    return PackedConvWeightCudnn<kSpatialDim>::prepack(
        weight, bias, stride, padding, output_padding, dilation, groups,
        transpose);
  }
};