clamp Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec256/vec256_half.h lines 149–163

template <>
Vectorized<Half> inline clamp(
    const Vectorized<Half>& a,
    const Vectorized<Half>& min,
    const Vectorized<Half>& max) {
  __m256 a_lo, a_hi;
  __m256 min_lo, min_hi;
  __m256 max_lo, max_hi;
  cvtfp16_fp32(__m256i(a), a_lo, a_hi);
  cvtfp16_fp32(__m256i(min), min_lo, min_hi);
  cvtfp16_fp32(__m256i(max), max_lo, max_hi);
  auto o1 = _mm256_min_ps(max_lo, _mm256_max_ps(min_lo, a_lo));
  auto o2 = _mm256_min_ps(max_hi, _mm256_max_ps(min_hi, a_hi));
  return cvtfp32_fp16(o1, o2);
}