FizzBuzz SIMD Edition

main.rs

extern crate core;
use core::arch::x86_64::*;

#[inline]
fn simdset(a: u32) -> __m256 {
    unsafe {
        _mm256_setr_ps(
            a as f32 + 0.0,
            a as f32 + 1.0,
            a as f32 + 2.0,
            a as f32 + 3.0,
            a as f32 + 4.0,
            a as f32 + 5.0,
            a as f32 + 6.0,
            a as f32 + 7.0,
        )
    }
}

fn main() {
    let store: [f32; 8] = [0.0; 8];
    let store3p: [f32; 8] = [0.0; 8];
    let store5p: [f32; 8] = [0.0; 8];

    (0..10).for_each(|i| {
        let simd = simdset(i);
        let rslt3 = unsafe { _mm256_div_ps(simd, _mm256_set1_ps(3.0)) };
        let comp3 = unsafe { _mm256_cmp_ps(rslt3, _mm256_floor_ps(rslt3), 0) };
        let rslt5 = unsafe { _mm256_div_ps(simd, _mm256_set1_ps(5.0)) };
        let comp5 = unsafe { _mm256_cmp_ps(rslt5, _mm256_floor_ps(rslt5), 0) };

        unsafe { _mm256_store_ps(store3p.as_ptr(), comp3) };
        unsafe { _mm256_store_ps(store.as_ptr(), simd) };
        unsafe { _mm256_store_ps(store5p.as_ptr(), comp5) };

        for j in 0..8 {
            if store3p[j] == 0.0 || store5p[j] == 0.0 {
                println!(
                    "{}{}",
                    if store3p[j] as u32 == 0 { "fizz" } else { "" },
                    if store5p[j] as u32 == 0 { "buzz" } else { "" }
                );
                continue;
            }

            println!("{}", store[j]);
        }
    });
}