maximum Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec256/vec256_bfloat16.h lines 111–127

template <>
Vectorized<BFloat16> inline maximum(
    const Vectorized<BFloat16>& a,
    const Vectorized<BFloat16>& b) {
  __m256 a_lo, a_hi;
  __m256 b_lo, b_hi;
  cvtbf16_fp32(__m256i(a), a_lo, a_hi);
  cvtbf16_fp32(__m256i(b), b_lo, b_hi);
  auto max_lo = _mm256_max_ps(a_lo, b_lo);
  auto max_hi = _mm256_max_ps(a_hi, b_hi);
  auto nan_lo = _mm256_cmp_ps(a_lo, b_lo, _CMP_UNORD_Q);
  auto nan_hi = _mm256_cmp_ps(a_hi, b_hi, _CMP_UNORD_Q);
  // Exploit the fact that all-ones is a NaN.
  auto o1 = _mm256_or_ps(max_lo, nan_lo);
  auto o2 = _mm256_or_ps(max_hi, nan_hi);
  return cvtfp32_bf16(o1, o2);
}