gistfile1.rs

fn image_dist_naivef32(
    x_offset: usize,
    y_offset: usize,
    search_img: &[f32],
    search_w: usize,
    subimage: &[f32],
    w: usize,
    h: usize,
) -> f32 {
    let calc_dist = |a, b| (a - b) * (a - b);
    let mut dist = 0.0f32;

    let search_img = &search_img[(x_offset + y_offset * search_w) as usize..];
    let mut subimage = &subimage[..(w * h * 3) as usize];
    let mut search_img = &search_img[..(h * search_w * 3) as usize];

    for y in 0..h {
        let row_sub = (y * w * 3) as usize;
        let row_search = (y * search_w * 3) as usize;

        let mut subimage = &subimage[row_sub..];
        let mut search_img = &search_img[row_search..];

        for x in 0..(w * 3) {
            dist += calc_dist(subimage[x], search_img[x]);
        }
    }
    dist.sqrt() / w as f32 / h as f32
}
use simdeez::avx2::*;
use simdeez::scalar::*;
use simdeez::sse2::*;
use simdeez::sse41::*;
use simdeez::*;
simd_runtime_generate!(
    fn image_dist_simdeez(
        x_offset: usize,
        y_offset: usize,
        search_img: &[f32],
        search_w: usize,
        subimage: &[f32],
        w: usize,
        h: usize,
        threshold: f32,
    ) -> f32 {
        assert!(w % S::VF32_WIDTH == 0);

        let mut subimage = &subimage[..(w * h) as usize];

        let search_img = &search_img[(x_offset + y_offset * search_w) as usize..];
        let mut search_img = &search_img[..(h * search_w) as usize];

        // [0.0; S::VF32_WIDTH] gave me a const generics error
        // In my case it's 8, 32 should be plenty conservative
        let zeroes = [0.0; 32];
        let mut res_simd = S::loadu_ps(&zeroes[0]);

        let simd_iters_per_row = w / S::VF32_WIDTH;

        for y in 0..h {
            let row_sub = (y * w) as usize;
            let row_search = (y * search_w) as usize;

            let mut subimage = &*subimage.get_unchecked(row_sub..);
            let mut search_img = &*search_img.get_unchecked(row_search..);

            for _ in 0..simd_iters_per_row {
                let search = S::loadu_ps(&*search_img.get_unchecked(0));
                let sub = S::loadu_ps(&*subimage.get_unchecked(0));

                let diff = S::sub_ps(sub, search);
                let square = S::mul_ps(diff, diff);
                res_simd = S::add_ps(res_simd, square);

                subimage = &*subimage.get_unchecked(S::VF32_WIDTH..);
                search_img = &*search_img.get_unchecked(S::VF32_WIDTH..);
            }
        }

        let res = S::horizontal_add_ps(res_simd);
        res.sqrt() / w as f32 / h as f32
    }
);