SIMD sum

simd_sum.rs

extern mod extra;

fn test_vector() -> ~[f32] {
    let mut v = ~[];
    for std::uint::range(0, 64) |i| {
        v.push(i as f32);
    }
    v
}

fn sum(v: &[f32]) -> f32 {
    let mut sum = 0f32;
    for v.iter().advance |vi| {
        sum += *vi;
    }
    sum
}

#[bench]
fn bench_sum(b: &mut extra::test::BenchHarness) {
    let v = test_vector();
    do b.iter() {
        sum(v);
    }
}

#[simd]
struct f32x4(f32, f32, f32, f32);

fn simd_sum(v: &[f32]) -> f32 {
    let mut sum = f32x4(0f32, 0f32, 0f32, 0f32);
    do v.as_imm_buf |mut p, mut len| {
        while len > 0 {
            let vv: f32x4 = unsafe { *std::cast::transmute::<*f32, *f32x4>(p) };
            sum += vv;
            p = p + 4;
            len -= 4;
        }
    }
    let f32x4(v1, v2, v3, v4) = sum;
    v1 + v2 + v3 + v4
}

#[bench]
fn bench_simd_sum(b: &mut extra::test::BenchHarness) {
    let v = test_vector();
    do b.iter() {
        simd_sum(v);
    }
}

fn main() {
    let v = test_vector();
    println(fmt!("%?", sum(v)));
    println(fmt!("%?", simd_sum(v)));
}