#simd #sse #webassembly #wasm #no-std

nightly no-std llml_simd

SIMD (Single Instruction Multiple Data) extension for a variety of targets

6 releases

0.1.5 May 4, 2022
0.1.4 Apr 3, 2022
0.1.3 Feb 11, 2022

#369 in WebAssembly

35 downloads per month

MIT license

2.5K SLoC

Crates.io NPM Rust docs


SIMD (Single Instruction Multiple Data) extension for a variety of targets. This project was initially started to facilitate the expansion of LLML's suported targets & features


This crate/library contains bindings to native-level SIMD instructions (alongside some polyfill) for SSE, AVX (see AVX support), NEON & WASM. It also contains naive implementations of all data-types (see naive implementation)

No std

llml_simd is a no_std crate. This means that it can be used for embeded systems projects seamlessly.

Naive implementation

If no supported target is detected (or if you enable the feature force_naive), Rust will compile the the SIMD vectors in naive mode. This mode represent's all data types as an array, executing most of the methods via iterators. This mode, while not recommended, is usefull if you intend to share code with some other programm that doesn't have SIMD support.

While not explicitly SIMD, rustc might still optimize some parts of the code to utilize SIMD instructions if it can and you allow it to. If you want to fully disable SIMD instructions, use --target-feature=-sse on x86/x86_64 and --target-feature=-neon on arm/aarch64 (naive mode will be used automatically in those cases, not requiring to enable force_naive)

AVX Support

If Rust detects avx as a target feature and you have the use_avx feature enabled (see features), llml_simd will compile all vectors over 128-bit long with AVX instructions, increasing performance significantly.

JavaScript Library

Thanks to WASM, llml_simd is available for JavaScript/TypeScript via npm.
You can install it into your Node project with npm i llml_simd


Feature Description
use_std Enables standard library functionality. Enabled by default
force_naive Forces naive types (see Naive implementation)
use_avx Enables the use of AVX SIMD types (see AVX support)
random Enables random generation of vectors via rand
serialize Enables serialization and deserialization of vectors via serde


Dot product (Rust)

use llml_simd::float::single::f32x4;

pub fn main() {
    let alpha = f32x4::new([1., 2., 3., 4.]);
    let beta = f32x4::new([5., 6., 7., 8.]);

    let dot = (alpha * beta).sum();
    assert_eq!(dot, 70.);

Dot product (JavaScript)

import { f32x4 } from llml_simd

let alpha = new f32x4(new Float32Array([1, 2, 3, 4]))
let beta = new f32x4(new Float32Array([5, 6, 7, 8]))

let dot = alpha.mul(beta).sum()
console.assert(dot === 70)


~29K SLoC