clamp_min Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h lines 1002–1013

template <>
Vectorized<BFloat16> inline clamp_min(
    const Vectorized<BFloat16>& a,
    const Vectorized<BFloat16>& min) {
  __m512 a_lo, a_hi;
  __m512 min_lo, min_hi;
  cvtbf16_fp32(__m512i(a), a_lo, a_hi);
  cvtbf16_fp32(__m512i(min), min_lo, min_hi);
  auto o1 = _mm512_max_ps(min_lo, a_lo);
  auto o2 = _mm512_max_ps(min_hi, a_hi);
  return cvtfp32_bf16(o1, o2);
}