QLinearInt8 Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cudnn/Linear.cpp lines 343–357

class QLinearInt8 final {
 public:
  static at::Tensor run(
      at::Tensor act,
      const c10::intrusive_ptr<LinearPackedParamsBase>& packed_weight,
      double output_scale,
      int64_t output_zero_point) {
    // TODO: check all zero_points are zero/all tensors are symmetrically quantized
    if constexpr (kReluFused) {
      return packed_weight->apply_relu(std::move(act), output_scale, output_zero_point);
    } else {
      return packed_weight->apply(std::move(act), output_scale, output_zero_point);
    }
  }
};