Ashwa 🐎

Hardware accelerated routines for single substring search

Language Support

ashwa is available across multiple language ecosystems,

rust (documented below)
npm
python

Supported Platforms

Architecture	Target Platform	Hardware Acceleration	Fallback
x86_64	Linux, macOS, Windows, Android, FreeBSD	AVX-512BW, AVX2, SSE4.2, SSSE3, SSE2	64-bit SWAR
x86 (i686)	Windows, Linux	SSE2	32-bit SWAR
AArch64 (ARM64)	Apple Silicon, Linux ARM64, Android, FreeBSD	128-bit ARM NEON	64-bit SWAR
ARMv7	Linux ARM, Android	128-bit ARM NEON	32-bit SWAR
WebAssembly	Browsers, Node.js (wasm32)	WASM SIMD128 (simd128)	32-bit SWAR

Usage

Add ashwa to your Cargo.toml:

[dependencies]
ashwa = "0.2.0"

Example

use ashwa::search_one;

fn main() {
    let text = b"The quick brown fox jumps over the lazy dog";
    match search_one(text, b'f') {
        Some(index) => println!("Found 'f' at byte index {}", index),
        None => println!("Byte not found"),
    }

    assert_eq!(search_one(text, b'z'), Some(0x25));
    assert_eq!(search_one(text, b'!'), None);
}

Benchmarks

search_one

`search_one`

For x86_64 machine targeting AVX-512BW SIMD ISA,

Level	Payload	Latency	Throughput	ILP
L1 Cache	32 KiB	211.52 ns	144.28 GiB/s	1.74 insn/cyc
L2 Cache	512 KiB	3.32 µs	147.03 GiB/s	2.53 insn/cyc
L3 Cache	16 MiB	495.57 µs	31.53 GiB/s	0.54 insn/cyc
RAM	256 MiB	20.52 ms	12.18 GiB/s	0.21 insn/cyc

Benchmarked using Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz (8C/16T) · L1d: 384 KiB, L1i: 256 KiB, L2: 10 MiB, L3: 54 MiB · STREAM Triad: 20.32 GB/s · +nightly toolchain

For aarch64 machine targeting NEON SIMD ISA,

Level	Payload	Latency	Throughput	ILP
L1 Cache	32 KiB	667.79 ns	45.70 GiB/s	0.51 insn/cyc
L2 Cache	512 KiB	10.82 µs	45.14 GiB/s	3.23 insn/cyc
L3 Cache	16 MiB	401.18 µs	38.95 GiB/s	2.99 insn/cyc
RAM	256 MiB	9.89 ms	25.28 GiB/s	2.02 insn/cyc

Benchmarked using ARM Neoverse-V1 (16C/16T) · L1d: 1 MiB, L1i: 1 MiB, L2: 16 MiB, L3: 32 MiB · STREAM Triad: 76.50 GB/s

ashwa 0.2.0