quantize_val Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/AffineQuantizerBase.cpp lines 37–54

template <typename T>
T quantize_val(double scale, int64_t zero_point, float value) {
  // Internally, fbgemm::Quantize uses std::nearbyint.
  // std::nearbyint results in nearest integer value according to the current
  // rounding mode and the default rounding mode is rounds to even in half-way
  // cases in most popular processor architectures like x86 and ARM. This is
  // typically faster than an alternatives like std::round that rounds half-way
  // cases away from zero, and can be consistent with SIMD implementations for
  // example in x86 using _mm512_cvtps_epi32 or mm512_round_ps with
  // _MM_FROUND_CUR_DIRECTION option that also follow the current rounding mode.
  // NOLINTNEXTLINE(bugprone-signed-char-misuse)
  auto qvalue = fbgemm::Quantize<typename T::underlying, false /*LEGACY*/>(
      value,
      static_cast<int32_t>(zero_point),
      static_cast<float>(scale),
      /*result_precision=*/CHAR_BIT * sizeof(typename T::underlying));
  return static_cast<T>(qvalue);
}