Module simd_dispatch 

Zero-overhead SIMD function dispatch using OnceLock.

This module provides compile-time-like dispatch for SIMD functions by detecting CPU features once at startup and caching function pointers.

Performance

• Zero branch overhead: Function pointer is resolved once, called directly thereafter
• No per-call checks: Eliminates is_x86_feature_detected! in hot loops
• Inlinable: Function pointers can be inlined by LLVM in some cases

EPIC-C.2: TS-SIMD-002

Structs

SimdFeatures

Information about available SIMD features.

Constants

CACHE_LINE_SIZE

Cache line size in bytes (standard for modern x86/ARM CPUs).

PREFETCH_DISTANCE_384D

Prefetch distance for 384-dimensional vectors.

PREFETCH_DISTANCE_768D

Prefetch distance for 768-dimensional vectors (3072 bytes). Calculated at compile time: 768 * 4 / 64 = 48 cache lines.

PREFETCH_DISTANCE_1536D

Prefetch distance for 1536-dimensional vectors.

Functions

cosine_dispatched

Computes cosine similarity using the best available SIMD implementation.

cosine_normalized_dispatched

Computes cosine similarity for pre-normalized vectors.

dot_product_dispatched

Computes dot product using the best available SIMD implementation.

euclidean_dispatched

Computes euclidean distance using the best available SIMD implementation.

hamming_dispatched

Computes Hamming distance using the best available implementation.

prefetch_distance

Calculates prefetch distance for a given dimension at compile time.

simd_features_info

Returns information about which SIMD features are available.