Module simd 

SIMD-Accelerated Distance Functions

Optimized distance calculations using SIMD instructions. Falls back to scalar implementations when SIMD is not available.

Supported Architectures

• x86_64: AVX2, SSE4.1 (auto-detected at runtime)
• aarch64: NEON (always available on Apple Silicon)
• Fallback: Portable scalar implementation

Performance

SIMD acceleration provides 2-8x speedup for distance calculations:

• L2 (Euclidean): Process 8 floats per iteration (AVX) or 4 (SSE/NEON)
• Dot product: Same vectorization approach
• Cosine: Computed as 1 - dot(a,b) / (||a|| * ||b||)

Structs

SimdCosine

SIMD-accelerated cosine distance

SimdDot

SIMD-accelerated dot product distance (for normalized vectors)

SimdInfo

Information about available SIMD features

SimdL2

SIMD-accelerated L2 (Euclidean squared) distance

Functions

cosine_distance

Compute cosine distance with best available SIMD Returns 1 - cosine_similarity, so 0 = identical, 2 = opposite

dot_product

Compute dot product with best available SIMD

l2_squared

Compute L2 squared distance with best available SIMD

norm_squared

Compute squared norm with best available SIMD

simd_info

Returns information about SIMD capabilities