clamp Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec512/vec512_bfloat16.h lines 973–987

template <>
Vectorized<BFloat16> inline clamp(
    const Vectorized<BFloat16>& a,
    const Vectorized<BFloat16>& min,
    const Vectorized<BFloat16>& max) {
  __m512 a_lo, a_hi;
  __m512 min_lo, min_hi;
  __m512 max_lo, max_hi;
  cvtbf16_fp32(__m512i(a), a_lo, a_hi);
  cvtbf16_fp32(__m512i(min), min_lo, min_hi);
  cvtbf16_fp32(__m512i(max), max_lo, max_hi);
  auto o1 = _mm512_min_ps(max_lo, _mm512_max_ps(min_lo, a_lo));
  auto o2 = _mm512_min_ps(max_hi, _mm512_max_ps(min_hi, a_hi));
  return cvtfp32_bf16(o1, o2);
}