QConvPrepackOneDNN Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/qconv_prepack.cpp lines 824–844

class QConvPrepackOneDNN final {
 public:
  static at::Tensor run_conv(
    at::Tensor weight, // from CPU backend instead of QuantizedCPU
    at::Tensor weight_scales, // Weight zero points must be 0s for onednn
    double input_scale,
    int64_t input_zero_point,
    torch::List<int64_t> stride,
    torch::List<int64_t> padding,
    torch::List<int64_t> dilation,
    int64_t groups,
    std::optional<torch::List<int64_t>> input_shape) {
#if AT_MKLDNN_ENABLED()
    return _qconv_prepack_onednn(
        weight, weight_scales, input_scale, input_zero_point,
        stride, padding, dilation, groups, input_shape);
#else
    TORCH_CHECK(false, "Unimplemented as onednn is not available.")
#endif
  }
};