xxhash-rust
Implementation of xxHash in Rust
Each algorithm is implemented via feature, allowing precise control over code size.
Example
- Cargo.toml
[]
= "0.8.12"
= ["xxh3", "const_xxh3"]
- main.rs
use xxh3_64 as const_xxh3;
use xxh3_64;
const TEST: u64 = const_xxh3;
assert!;
assert!;
Features:
By default all features are off.
std- Enablesstd::io::Writetrait implementationxxh32- Enables 32bit algorithm. Suitable for x86 targetsconst_xxh32-const fnversion ofxxh32algorithmxxh64- Enables 64 algorithm. Suitable for x86_64 targetsconst_xxh64-const fnversion ofxxh64algorithmxxh3- Enablesxxh3family of algorithms, superior toxxh32andxxh64in terms of performance.const_xxh3-const fnversion ofxxh3algorithm
HW acceleration
Similar to reference implementation, crate implements various SIMDs in xxh3 depending on provided flags.
All checks are performed only at compile time, hence user is encouraged to enable these accelerations (for example via -C target_cpu=native)
Used SIMD acceleration:
- SSE2 - widely available, can be safely enabled in 99% of cases. Enabled by default in
x86_64targets. - AVX2;
- Neon - Enabled by default on aarch64 targets (most likely)
- Wasm SIMD128 - Has to be enabled via rust flag:
-Ctarget-feature=+simd128
Streaming vs One-shot
For performance reasons one-shot version of algorithm does not re-use streaming version. Unless needed, user is advised to use one-shot version which tends to be more optimal.
const fn version
While const fn provides compile time implementation, it does so at performance cost.
Hence you should only use it at compile time.
To guarantee that something is computed at compile time make sure to initialize hash output
as const or static variable, otherwise it is possible function is executed at runtime, which
would be worse than regular algorithm.
const fn is implemented in best possible way while conforming to limitations of Rust const fn, but these limitations are quite strict making any high performance code impossible.
Version note
0.8.*corresponds to C's0.8.*
In order to keep up with original implementation version I'm not planning to bump major/minor until C implementation does so.