maximum Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h lines 928–946

template <>
Vectorized<BFloat16> inline maximum(
    const Vectorized<BFloat16>& a,
    const Vectorized<BFloat16>& b) {
  __m512 a_lo, a_hi;
  __m512 b_lo, b_hi;
  cvtbf16_fp32(__m512i(a), a_lo, a_hi);
  cvtbf16_fp32(__m512i(b), b_lo, b_hi);
  auto max_lo = _mm512_max_ps(a_lo, b_lo);
  auto max_hi = _mm512_max_ps(a_hi, b_hi);
  auto nan_lo_mask = _mm512_cmp_ps_mask(a_lo, b_lo, _CMP_UNORD_Q);
  auto nan_hi_mask = _mm512_cmp_ps_mask(a_hi, b_hi, _CMP_UNORD_Q);
  auto nan_lo = _mm512_castsi512_ps(_mm512_set1_epi32(nan_lo_mask));
  auto nan_hi = _mm512_castsi512_ps(_mm512_set1_epi32(nan_hi_mask));
  // Exploit the fact that all-ones is a NaN.
  auto o1 = _mm512_or_ps(max_lo, nan_lo);
  auto o2 = _mm512_or_ps(max_hi, nan_hi);
  return cvtfp32_bf16(o1, o2);
}