I want to expand some u8s out to u64, except instead of zero or sign extending, which have direct support, I want “copy-extending”. What’s the best way to do this (on intel cpus with avx512)? Example code is in rust but the host language isn’t the interesting part.

#![feature(portable_simd)]

use std::simd::*;

// Expands out each input byte 8 times
pub fn batch_splat_scalar(x: [u8; 16]) -> [u64; 16] {
  let mut ret = [0; 16];
  for i in 0..16 {
    ret[i] =
      u64::from_le_bytes([x[i], x[i], x[i], x[i], x[i], x[i], x[i], x[i]]);
  }
  ret
}

pub fn batch_splat_simd(x: u8x16) -> u64x16 {
  Simd::from_array(batch_splat_scalar(x.to_array()))
}

which compiles to something like this with avx512

        vpmovzxbq       zmm0, qword ptr [rsi]
        vpbroadcastq    zmm1, qword ptr [rip + .LCPI0_0]
        mov     rax, rdi
        vpmuludq        zmm2, zmm0, zmm1
        vpbroadcastq    zmm3, qword ptr [rip + .LCPI0_1]
        vpmuludq        zmm0, zmm0, zmm3
        vpsllq  zmm0, zmm0, 32
        vporq   zmm0, zmm2, zmm0
        vmovdqu64       zmmword ptr [rdi], zmm0
        vpmovzxbq       zmm0, qword ptr [rsi + 8]
        vpmuludq        zmm1, zmm0, zmm1
        vpmuludq        zmm0, zmm0, zmm3
        vpsllq  zmm0, zmm0, 32
        vporq   zmm0, zmm1, zmm0
        vmovdqu64       zmmword ptr [rdi + 64], zmm0
        vzeroupper
        ret

https://godbolt.org/z/67cW5GnKf