Profiling SIMD saturating subtract

main.rs

const N: usize = 10000;

fn main() {
    let mut args = std::env::args().skip(1); // skip the program name

    // just get a bunch of arbitrary values cheaply
    const INPUT: &[u8] = include_bytes!(
        "/home/folkertdev/.cargo/registry/cache/github.com-1ecc6299db9ec823/clap-4.2.7.crate"
    );
    let (_, table, _) = unsafe { INPUT.align_to() };

    let wsize = std::hint::black_box(42);

    match args.next().as_deref() {
        Some("scalar") => {
            for _ in 0..N {
                let mut table = table.to_vec();
                scalar(&mut table, wsize)
            }
        }
        Some("simd") => {
            for _ in 0..N {
                let mut table = table.to_vec();
                unsafe { simd(&mut table, wsize) }
            }
        }
        _ => unreachable!("Unexpected argument. Only 'scalar' or 'simd' are allowed."),
    }
}


pub fn scalar(table: &mut [u16], wsize: u16) {
    for m in table.iter_mut() {
        *m = m.saturating_sub(wsize);
    }
}

#[inline(always)]
fn generic_slide_hash_chain<const N: usize>(table: &mut [u16], wsize: u16) {
    for chunk in table.chunks_exact_mut(N) {
        for m in chunk.iter_mut() {
            *m = m.saturating_sub(wsize);
        }
    }
}

#[cfg(target_arch = "x86_64")]
#[target_feature(enable = "avx2")]
unsafe fn simd(table: &mut [u16], wsize: u16) {
    // 64 means that 4 256-bit values can be processed per iteration.
    // That appear to be the optimal amount for avx2.
    generic_slide_hash_chain::<64>(table, wsize);
}

results.md

using https://github.com/andrewrk/poop

> poop "target/release/playground scalar" "target/release/playground simd"
Benchmark 1 (277 runs): target/release/playground scalar
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          18.0ms ±  540us    17.4ms … 23.6ms         17 ( 6%)        0%
  peak_rss           2.14MB ± 61.5KB    2.10MB … 2.23MB          0 ( 0%)        0%
  cpu_cycles         77.4M  ± 2.15M     75.7M  …  101M          16 ( 6%)        0%
  instructions        157M  ±  292       157M  …  157M           4 ( 1%)        0%
  cache_references   41.8M  ±  862K     36.6M  … 43.4M          29 (10%)        0%
  cache_misses        251K  ± 62.7K      217K  …  710K          50 (18%)        0%
  branch_misses      14.0K  ± 28.8      13.9K  … 14.1K           6 ( 2%)        0%
Benchmark 2 (300 runs): target/release/playground simd
  measurement          mean ± σ            min … max           outliers         delta
  wall_time          16.6ms ±  328us    16.0ms … 19.0ms         17 ( 6%)        ⚡ -  7.6% ±  0.4%
  peak_rss           2.15MB ± 63.5KB    2.10MB … 2.23MB          0 ( 0%)          +  0.3% ±  0.5%
  cpu_cycles         71.3M  ± 1.31M     70.2M  … 81.4M          17 ( 6%)        ⚡ -  8.0% ±  0.4%
  instructions       68.1M  ±  281      68.1M  … 68.1M           1 ( 0%)        ⚡ - 56.7% ±  0.0%
  cache_references   39.6M  ±  812K     31.9M  … 44.1M          22 ( 7%)        ⚡ -  5.4% ±  0.3%
  cache_misses        223K  ± 52.9K      152K  …  507K          64 (21%)        ⚡ - 11.1% ±  3.8%
  branch_misses      14.1K  ±  578      13.9K  … 24.0K           5 ( 2%)          +  0.2% ±  0.5%