dot_with_fp32_arith_main_loop_no_bfdot Class — pytorch Architecture

Source Code

aten/src/ATen/native/cpu/ReducedPrecisionFloatGemvFastPathKernel.cpp lines 287–303

template <typename T>
C10_ALWAYS_INLINE auto
dot_with_fp32_arith_main_loop_no_bfdot(
    const T* vec1,
    const T* vec2,
    int64_t len) {
  vec::VectorizedN<float, kF32RegistersPerIteration> sum(0);
  const auto len_aligned = len & ~(kF32ElementsPerIteration - 1);
  for (int j = 0; j < len_aligned ; j += kF32ElementsPerIteration) {
    const auto* vec1_ = vec1 + j;
    const auto* vec2_ = vec2 + j;
    c10::ForcedUnroll<kF32RegisterPairsPerIteration>{}([vec1_, vec2_, &sum](auto k) C10_ALWAYS_INLINE_ATTRIBUTE {
      dot_with_fp32_arith_main_inner_loop_no_bfdot(vec1_, vec2_, sum, k);
    });
  }
  return reduce(sum);
}