Module simd_native 

Native SIMD intrinsics for maximum performance.

This module provides hand-tuned SIMD implementations using core::arch intrinsics for AVX-512, AVX2, and ARM NEON architectures.

Module Structure

• scalar — Scalar fallback implementations and fast-rsqrt helpers
• tail_unroll — Remainder/tail handling macros for SIMD loops
• prefetch — CPU cache prefetch utilities
• reduction — Shared horizontal sum helpers for SIMD accumulators
• x86_avx512 — AVX-512F kernel implementations (x86_64 only)
• x86_avx2 — AVX2+FMA dot product and squared L2 kernels (x86_64 only)
• x86_avx2_similarity — AVX2+FMA cosine, Hamming, Jaccard kernels (x86_64 only)
• neon — ARM NEON kernel implementations (aarch64 only)
• dispatch — Runtime SIMD level detection and dispatch wiring

Performance (based on arXiv research)

• AVX-512: True 16-wide f32 operations with masked remainder
• AVX2: 8-wide f32 with FMA, multi-accumulator ILP
• ARM NEON: Native 128-bit SIMD for Apple Silicon/ARM64
• Prefetch: Software prefetching for cache optimization

References

• arXiv:2505.07621 “Bang for the Buck: Vector Search on Cloud CPUs”
• arXiv:2502.18113 “Accelerating Graph Indexing for ANNS on Modern CPUs”

Re-exports

pub use scalar::cosine_similarity_fast;

pub use scalar::fast_rsqrt;

pub use prefetch::calculate_prefetch_distance;

pub use prefetch::prefetch_vector;

pub use prefetch::prefetch_vector_multi_cache_line;

pub use prefetch::prefetch_vector_u64;

pub use prefetch::L2_CACHE_LINE_BYTES;

Modules

adc

ADC (Asymmetric Distance Computation) for PQ-compressed vector search.

prefetch

CPU cache prefetch utilities for SIMD operations.

scalar

Scalar fallback implementations for SIMD distance metrics.

Structs

DistanceEngine

Pre-resolved SIMD kernels for repeated distance operations at a fixed dimension.

Enums

SimdLevel

SIMD capability level detected at runtime.

Functions

batch_cosine_native

Batch cosine similarity with cross-platform multi-level prefetch hints.

batch_dot_product_native

Batch dot product with cross-platform multi-level prefetch hints.

batch_euclidean_native

Batch euclidean distance with cross-platform multi-level prefetch hints.

batch_hamming_native

Batch Hamming distance with cross-platform multi-level prefetch hints.

batch_jaccard_native

Batch Jaccard similarity with cross-platform multi-level prefetch hints.

batch_squared_l2_native

Batch squared L2 distance with cross-platform multi-level prefetch hints.

cosine_normalized_native

Cosine for pre-normalized vectors with runtime SIMD dispatch.

cosine_similarity_native

Cosine similarity with runtime SIMD dispatch.

dot_product_native

Dot product with runtime SIMD dispatch.

euclidean_native

Euclidean distance with runtime SIMD dispatch.

hamming_binary_native

Binary Hamming distance with runtime SIMD dispatch.

hamming_distance_native

Hamming distance with runtime SIMD dispatch.

jaccard_similarity_native

Jaccard similarity with runtime SIMD dispatch.

norm_native

L2 norm with runtime SIMD dispatch.

normalize_inplace_native

In-place normalization with runtime SIMD dispatch.

simd_level

Returns the cached SIMD level for the current process.

squared_l2_native

Squared L2 distance with runtime SIMD dispatch.

warmup_simd_cache

Warms up runtime SIMD dispatch cache and CPU caches for common dimensions.