deinterleave2 Class — pytorch Architecture

Source Code

aten/src/ATen/cpu/vec/vec256/vsx/vec256_common_vsx.h lines 177–195

template <>
std::pair<Vectorized<double>, Vectorized<double>> inline deinterleave2<double>(
    const Vectorized<double>& a,
    const Vectorized<double>& b) {
  // inputs:
  //   a = {a0, b0, a1, b1}
  //   b = {a2, b2, a3, b3}
  vfloat64 aa01 = vec_xxpermdi(a.vec0(), a.vec1(), 0);
  vfloat64 aa23 = vec_xxpermdi(b.vec0(), b.vec1(), 0);

  vfloat64 bb_01 = vec_xxpermdi(a.vec0(), a.vec1(), 3);
  vfloat64 bb_23 = vec_xxpermdi(b.vec0(), b.vec1(), 3);

  // swap lanes:
  //   return {a0, a1, a2, a3}
  //          {b0, b1, b2, b3}
  return std::make_pair(
      Vectorized<double>{aa01, aa23}, Vectorized<double>{bb_01, bb_23});
}