Module simd 

Vector operations with SIMD acceleration.

Multiple backends are supported via feature flags:

Feature	Backend	Performance	Notes
innr (default)	innr crate	4-8x	Pure Rust, good baseline
simsimd	SimSIMD	Up to 200x	C bindings, best on modern CPUs
(none)	Portable	1x	Fallback, no dependencies

Feature Priority

If multiple features are enabled, priority is: simsimd > innr > fallback

Usage

For normalized embeddings, prefer dot() over cosine().

use vicinity::simd::{dot, cosine, norm};

let a = [1.0_f32, 0.0, 0.0];
let b = [0.707, 0.707, 0.0];

let d = dot(&a, &b);
let c = cosine(&a, &b);
let n = norm(&a);

Functions

cosine

Cosine similarity between two vectors.

dot

Dot product of two vectors: Σ(a[i] * b[i]).

dot_portable

Portable (non-SIMD) dot product.

l2_distance

L2 (Euclidean) distance: sqrt(Σ(a[i] - b[i])²).

l2_distance_squared

Squared L2 distance: Σ(a[i] - b[i])².

norm

L2 norm (Euclidean norm) of a vector: sqrt(Σ(v[i]²)).

sparse_dot

Sparse dot product using sorted index arrays.

sparse_dot_portable

Portable implementation of sparse dot product.