Compiler Explorer

Source code

#![feature(portable_simd)]

use core::simd::{u16x8, u8x16, ToBytes};

pub fn swizzle_to_front(val: u16x8, bitmask: u8) -> u16x8 {
    static SWIZZLE_TABLE: [[u8; 16]; 256] = {
        let mut table = [[0; 16]; 256];
        let mut n = 0usize;
        while n < table.len() {
            let mut x = n;
            let mut i = 0;
            while x > 0 {
                let lsb = x.trailing_zeros() as u8;
                x ^= 1 << lsb;
                table[n][i] = lsb * 2; // first byte
                table[n][i + 1] = lsb * 2 + 1; // second byte
                i += 2;
            }
            n += 1;
        }
        table
    };

// Our swizzle table retains the order of the bytes in the 16 bit lanes, we can
    // stick with native byte order as long as we convert back with native endianness too.
    let val_convert: u8x16 = val.to_ne_bytes();
    let swizzle_idxs = u8x16::from_array(SWIZZLE_TABLE[bitmask as usize]);

let swizzled: u8x16 = val_convert.swizzle_dyn(swizzle_idxs);
    u16x8::from_ne_bytes(swizzled)
}