use std::arch::x86_64::*;

fn main() {
    let mut data = Vec::new();
    for i in 0..32 {
        data.push(i);
    }
    println!("data: {:?}", data);

    let mut result = std::iter::repeat(0).take(32).collect::<Vec<_>>();
    unsafe {
        let a = _mm256_loadu_si256(data.as_ptr() as usize as *const _);
        _mm256_storeu_si256(result.as_mut_ptr() as usize as *mut _, a);
    }

    println!("res:  {:?}", result);
}