fmadd Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h lines 1074–1088

template <>
Vectorized<BFloat16> inline fmadd(
    const Vectorized<BFloat16>& a,
    const Vectorized<BFloat16>& b,
    const Vectorized<BFloat16>& c) {
  __m512 a_lo, a_hi;
  __m512 b_lo, b_hi;
  __m512 c_lo, c_hi;
  cvtbf16_fp32(__m512i(a), a_lo, a_hi);
  cvtbf16_fp32(__m512i(b), b_lo, b_hi);
  cvtbf16_fp32(__m512i(c), c_lo, c_hi);
  auto o1 = _mm512_fmadd_ps(a_lo, b_lo, c_lo);
  auto o2 = _mm512_fmadd_ps(a_hi, b_hi, c_hi);
  return cvtfp32_bf16(o1, o2);
}