convert Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/sve/vec_bfloat16.h lines 555–579

template <>
inline void convert(const BFloat16* src, BFloat16* dst, int64_t n) {
  const int64_t fraction = n % Vectorized<BFloat16>::size();
#pragma unroll
  for (int64_t i = 0; i < n - fraction; i += Vectorized<BFloat16>::size()) {
    svst1_bf16(
        ptrue,
        const_cast<bfloat16_t*>(reinterpret_cast<const bfloat16_t*>(dst)) + i,
        svldnt1_bf16(
            ptrue,
            const_cast<bfloat16_t*>(reinterpret_cast<const bfloat16_t*>(src)) +
                i));
  }
#pragma unroll
  for (int64_t i = n - fraction; i < n; i += Vectorized<BFloat16>::size()) {
    svbool_t pg = svwhilelt_b16(i, n);
    svst1_bf16(
        pg,
        const_cast<bfloat16_t*>(reinterpret_cast<const bfloat16_t*>(dst)) + i,
        svldnt1_bf16(
            pg,
            const_cast<bfloat16_t*>(reinterpret_cast<const bfloat16_t*>(src)) +
                i));
  }
}