convertFromBf16Impl Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec128/vec128_convert.h lines 195–208

template <typename to_type>
inline void convertFromBf16Impl(
    const c10::BFloat16* __restrict src,
    to_type* __restrict dst,
    int64_t n) {
  const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
  uint64_t len = static_cast<uint64_t>(n);
  for (uint64_t i = 0; i < len; i++) {
    uint32_t tmp = static_cast<uint32_t>(srcPtr[i]) << 16;
    float tmpF;
    __builtin_memcpy(&tmpF, &tmp, sizeof(float));
    dst[i] = static_cast<to_type>(tmpF);
  }
}