There's no need to use `vec_perm()` instead of `vec_splat()`,
  since instruction `vperm` is quite heavy compared to `vsplt[b,h,w]`.