Module simd 

SIMD-accelerated distance calculations

This module provides SIMD-optimized implementations of distance metrics for improved performance on supported CPUs.

Features

• Auto-vectorization hints: Helps the compiler generate SIMD code
• x86_64 optimizations: Automatically uses AVX-512, FMA+AVX2, or AVX2 when available
• aarch64 optimizations: Automatically uses NEON (always available on ARM64)
• Cache-friendly: Optimized memory access patterns
• Fallback: Automatically falls back to auto-vectorization on unsupported platforms

Performance Hierarchy

• x86_64: AVX-512 (16-wide) → FMA+AVX2 (8-wide) → AVX2 (8-wide) → auto-vectorization
• aarch64: NEON (4-wide)
• other: auto-vectorization

Usage

use oxify_vector::simd::cosine_similarity_simd;

let v1 = vec![1.0, 2.0, 3.0, 4.0];
let v2 = vec![2.0, 3.0, 4.0, 5.0];
let similarity = cosine_similarity_simd(&v1, &v2);

Functions

compute_distance_lower_is_better_simd

Compute distance using the specified metric with SIMD optimization

compute_distance_simd

Compute similarity/distance using the specified metric with SIMD optimization

cosine_similarity_simd

SIMD-optimized cosine similarity calculation

dot_product_simd

SIMD-optimized dot product calculation

euclidean_distance_simd

SIMD-optimized Euclidean distance calculation

is_avx2_available

Check if AVX2 is available at runtime

is_avx512_available

Check if AVX-512 is available at runtime (x86_64 only)

is_fma_available

Check if FMA is available at runtime

is_neon_available

Check if NEON is available at runtime (non-aarch64 always returns false)

manhattan_distance_simd

SIMD-optimized Manhattan distance calculation

normalize_vector_simd

Normalize a vector in-place using SIMD optimization

quantized_dot_product_simd

Compute dot product between two quantized (u8) vectors using SIMD

quantized_euclidean_squared_simd

Compute Euclidean distance between two quantized (u8) vectors using SIMD

quantized_manhattan_distance_simd

Compute Manhattan distance between two quantized (u8) vectors using SIMD

scale_vector_simd

Scale a vector by a constant using SIMD optimization