fmadd Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec256/vec256_half.h lines 250–264

template <>
Vectorized<Half> inline fmadd(
    const Vectorized<Half>& a,
    const Vectorized<Half>& b,
    const Vectorized<Half>& c) {
  __m256 a_lo, a_hi;
  __m256 b_lo, b_hi;
  __m256 c_lo, c_hi;
  cvtfp16_fp32(__m256i(a), a_lo, a_hi);
  cvtfp16_fp32(__m256i(b), b_lo, b_hi);
  cvtfp16_fp32(__m256i(c), c_lo, c_hi);
  auto o1 = _mm256_fmadd_ps(a_lo, b_lo, c_lo);
  auto o2 = _mm256_fmadd_ps(a_hi, b_hi, c_hi);
  return cvtfp32_fp16(o1, o2);
}