simd-adler32

A SIMD-accelerated Adler-32 rolling hash algorithm implementation.

Features

No dependencies
Support no_std (with default-features = false)
Runtime CPU feature detection (when std enabled)
Blazing fast performance on as many targets as possible (currently only x86 and x86_64)
Default to scalar implementation when simd not available

Quick start

Cargo.toml

toml [dependencies] simd-adler32 = "*"

example.rs

```rust use simd_adler32::Adler32;

let mut adler = Adler32::new(); adler.write(b"rust is pretty cool, man"); let hash = adler.finish();

println!("{}", hash); // 1921255656 ```

Support

CPU Features

| impl | arch | feature | | ---- | ---------------- | ------- | | ✅ | x86, x86_64 | avx512 | | ✅ | x86, x86_64 | avx2 | | ✅ | x86, x86_64 | ssse3 | | ✅ | x86, x86_64 | sse2 | | 🚧 | arm, aarch64 | neon | | | wasm32 | simd128 |

MSRV 1.36.0**

Minimum supported rust version is tested before a new version is published. [**] Feature const-generics needs to disabled to build on rustc versions <1.51 which can be done by updating your dependency definition to the following.

Cargo.toml

toml [dependencies] simd-adler32 = { version "*", default-features = false, features = ["std"] }

Performance

Benchmarks listed display number of randomly generated bytes (10k / 100k) and library name. Benchmarks sources can be found under the bench directory. Crates used for comparison are adler and adler32.

Windows 10 Pro - Intel i5-8300H @ 2.30GHz

| name | avg. time | avg. thrpt | | ----------------------- | --------------- | ------------------ | | 10k/simd-adler32 | 212.61 ns | 43.805 GiB/s | | 10k/wuffs | 3843 ns | 2.63 GiB/s* | | 10k/adler32 | 4.8084 us | 1.9369 GiB/s | | 10k/adler | 17.979 us | 530.43 MiB/s | | ----------------------- | --------------- | ------------------ | | 100k/simd-adler32 | 2.7951 us | 33.320 GiB/s | | 100k/wuffs | 34733 ns | 2.6814 GiB/s* | | 100k/adler32 | 48.488 us | 1.9207 GiB/s | | 100k/adler | 178.36 us | 534.69 MiB/s |

* wuffs ran using mingw64/gcc, ran with wuffs bench -ccompilers=gcc -reps=1 -iterscale=300 std/adler32.

MacBookPro16,1 - Intel i9-9880H CPU @ 2.30GHz

| name | avg. time | avg. thrpt | | ----------------------- | --------------- | ------------------ | | 10k/simd-adler32 | 200.37 ns | 46.480 GiB/s | | 10k/adler32 | 4.1516 us | 2.2433 GiB/s | | 10k/adler | 10.220 us | 933.15 MiB/s | | ----------------------- | --------------- | ------------------ | | 100k/simd-adler32 | 2.3282 us | 40.003 GiB/s | | 100k/adler32 | 41.130 us | 2.2643 GiB/s | | 100k/adler | 83.776 us | 534.69 MiB/s |

Safety

This crate contains a significant amount of unsafe code due to the requirement of unsafe for simd intrinsics. Fuzzing is done on release and debug builds prior to publishing via afl. Fuzzy tests can be found under fuzz the directory.

Resources

LICENSE - MIT
CHANGELOG

Credits

Thank you to the contributors of the following projects.

Contributing

Feel free to submit a issue or pull request. :smile: