QConvInt8 Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cudnn/Conv.cpp lines 368–384

class QConvInt8 final {
 public:
  static at::Tensor run(
      at::Tensor act,
      const c10::intrusive_ptr<ConvPackedParamsBase<kSpatialDim>>& packed_weight,
      double output_scale,
      int64_t output_zero_point) {
    TORCH_CHECK(kSpatialDim == 1 || kSpatialDim == 2, "Error in quantized cudnn conv2d operator: "
                "Expected kSpatialDim == 1 || kSpatialDim == 2; received kSpatialDim=", kSpatialDim);
    // TODO: check all zero_points are zero/all tensors are symmetrically quantized
    if (kReluFused) {
      return packed_weight->apply_relu(act, output_scale, output_zero_point);
    } else {
      return packed_weight->apply(act, output_scale, output_zero_point);
    }
  }
};