void Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/ReducedPrecisionFloatGemvFastPathKernel.cpp lines 261–274

template <typename T>
C10_ALWAYS_INLINE void dot_with_fp32_arith_main_inner_loop_no_bfdot(
  const T* vec1,
  const T* vec2,
  vec::VectorizedN<float, kF32RegistersPerIteration>& sum,
  int registerPairIndex) {
  static_assert(std::is_same_v<T, Half> || std::is_same_v<T, BFloat16>);
  const auto temp_vec1 = vec::Vectorized<T>::loadu(&vec1[registerPairIndex * vec::Vectorized<T>::size()]);
  const auto temp_vec2 = vec::Vectorized<T>::loadu(&vec2[registerPairIndex * vec::Vectorized<T>::size()]);

  const auto [result_low, result_high] = fmadd(temp_vec1, temp_vec2, sum[2 * registerPairIndex], sum[2 * registerPairIndex + 1]);
  sum[2 * registerPairIndex] = result_low;
  sum[2 * registerPairIndex + 1] = result_high;
}