QMul Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cpu/qmul.cpp lines 315–335

class QMul final {
 public:
  static Tensor run(Tensor qa, Tensor qb, double scale, int64_t zero_point) {
    check_inputs(qa, qb);
#ifdef USE_XNNPACK
    int64_t q_max = std::numeric_limits<c10::qint8::underlying>::max();
    if (zero_point < q_max && qa.scalar_type() == kQInt8) {
      return _mul_out_xnnpack<c10::qint8, ReLUFused>(qa, qb, scale, zero_point);
    }
#endif // USE_XNNPACK

    auto qc = at::_empty_affine_quantized(
        infer_size_dimvector(qa.sizes(), qb.sizes()),
        at::device(kCPU).dtype(qa.scalar_type()),
        scale,
        zero_point,
        qa.suggest_memory_format());

    return _mul_out<ReLUFused>(qc, qa, qb);
  }
};