Module simd 

SIMD-accelerated vector operations

This module provides high-performance SIMD implementations of vector operations for x86_64 (AVX2, SSE) and ARM (NEON) architectures. The implementations use the pulp crate for portable SIMD abstraction with runtime CPU detection.

Performance

SIMD implementations provide 3-4x speedup over scalar operations for typical embedding dimensions (384, 768, 1024). The exact speedup depends on:

• Vector length (longer vectors benefit more)
• CPU architecture and SIMD support
• Memory alignment and cache behavior

Architecture Support

• x86_64: AVX2 (8x f32), SSE (4x f32), scalar fallback
• ARM: NEON (4x f32), scalar fallback
• Other: Scalar fallback

Usage

use foxstash_core::vector::simd::{dot_product_simd, cosine_similarity_simd};

let a = vec![1.0; 384];
let b = vec![2.0; 384];

let dot = dot_product_simd(&a, &b);
let similarity = cosine_similarity_simd(&a, &b);

Functions

cosine_distance_prenorm

Computes cosine distance with a precomputed norm for vector b.

cosine_similarity_simd

Computes cosine similarity using SIMD acceleration.

dot_product_simd

Computes dot product using SIMD acceleration.

l2_distance_simd

Computes L2 (Euclidean) distance using SIMD acceleration.

norm_simd

Computes the L2 norm (magnitude) of a vector using SIMD acceleration.