innr

SIMD-accelerated vector similarity primitives.

Dual-licensed under MIT or Apache-2.0.

Why this exists

innr is the dependency you reach for when you need fast, well-tested vector math without pulling in a full ANN index or ML framework. It is designed to sit under crates like jin, retrieval pipelines, and evaluation tooling.

Quickstart

[dependencies]
innr = "0.1.1"

use innr::{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);      // 0.707
let c = cosine(&a, &b);   // 0.707
let n = norm(&a);         // 1.0

Operations

Function	Description
`dot`	Inner product
`norm`	L2 norm
`cosine`	Cosine similarity
`l2_distance`	Euclidean distance
`sparse_dot`	Sparse vector dot (`sparse` feature)
`maxsim`	ColBERT late interaction (`maxsim` feature)

SIMD Dispatch

Architecture	Instructions	Detection
x86_64	AVX2 + FMA	Runtime
aarch64	NEON	Always
Other	Portable	LLVM auto-vec

Vectors < 16 dimensions use portable code.

Features

sparse — sparse vector operations
maxsim — ColBERT late interaction scoring
full — all features

Best starting points

Cosine / dot / norm: cosine, dot, norm
Distances: l2_distance
When using cosine: normalize once (or use an index that expects normalized vectors)

Performance

Benchmark throughput

Apple Silicon (NEON). Run cargo bench to reproduce on your hardware.

For maximum performance, build with native CPU features:

RUSTFLAGS="-C target-cpu=native" cargo build --release

Or specify a portable baseline with SIMD:

# AVX2 (89% of x86_64 CPUs)
RUSTFLAGS="-C target-cpu=x86-64-v3" cargo build --release

# SSE2 only (100% compatible)
RUSTFLAGS="-C target-cpu=x86-64" cargo build --release

Run benchmarks:

cargo bench

Generate flamegraphs (requires cargo-flamegraph):

./scripts/profile.sh dense

Tests

cargo test -p innr

License