Is the usage of simd correct?

[somewheve]

#![feature(portable_simd)]
use std::simd::f64x8;

pub fn div_matrix<const N: usize>(a: &[f64; N], b: &[f64; N]) -> [f64; N] {
    // f64 -> init matrix
    let mut res = [0.0; N];
    // N % 8 split to calculate
    let mut start = 0;
    let mut end = 8;
    loop {
        if end >= N {
            end = N;
            let mut p = [1.0; 8];
            let mut q = [1.0; 8];
            for i in start..end {
                p[i % 8] = a[i];
                q[i % 8] = b[i];
            }
            let temp = f64x8::from_array(p) / f64x8::from_array(q);
            for i in start..end {
                res[i] = temp[i % 8];
            }
            break;
        } else {
            // length will always be 8 just copy_from_slice
            let temp = f64x8::from_slice(&a[start..end]) / f64x8::from_slice(&b[start..end]);
            res[start..end].copy_from_slice(temp.as_array());
        }
        start += 8;
        end += 8;
    }
    res
}

fn main() {
    let d = [1.0, 6.0, 3.0, 3.0, 20.0, 9.0, 1.0, 9.0, 4.0, 3.0];
    let f = [10.0, 2.0, 2.0, 6.0, 40.0, 3.0, 6.0, 7.0, 63.0, 95.2];

    let p = div_matrix::<10>(&d, &f);
    dbg!(p);
}

i never use simd before,
Do you have any suggestions for performance improvements?

[Nugine]

Look into the asm. In most cases, simd hot loop is expected to be compiled to several instructions without any function calls. It's a general suggestion.

See https://crates.io/crates/cargo-show-asm

[somewheve]

got it.

[scottmcm]

@somewheve

// N % 8 split to calculate

Since you're using nightly anyway for std::simd, use more nightly stuff to simplify your code.

For example, check out .as_chunks::<8>() to get [f64; 8]s to turn into the f64x8s.