Crate rustyhdf5_accel 

SIMD-accelerated operations for rustyhdf5.

This crate provides runtime-dispatched SIMD acceleration for common vector operations used in HDF5 processing: dot products, cosine similarity, L2 distance, f16 conversion, and checksums.

All public functions automatically select the best available SIMD backend at runtime. Every operation has a portable scalar fallback.

Modules

avx2

AVX2 SIMD implementations for x86_64. All functions require runtime detection via is_x86_feature_detected!(“avx2”).

checksum

SIMD-accelerated checksum implementations.

convert

f16/f32 conversion with SIMD acceleration.

scalar

Portable scalar implementations of all operations. These serve as fallbacks when SIMD is not available.

Enums

Backend

Available SIMD backends.

Constants

CACHE_LINE_SIZE

Functions

align_to_cache_line

Round size up to the next multiple of CACHE_LINE_SIZE.

batch_cosine

Compute cosine similarity between a query and multiple vectors.

batch_cosine_prenorm

Compute cosine similarity with pre-normalized query vector.

batch_norms

Compute L2 norms for a batch of vectors.

checksum_fletcher32

Compute Fletcher-32 checksum.

cosine_similarity

Compute cosine similarity between two vectors (fused single-pass).

detect_backend

Detect the best available SIMD backend at runtime.

dot_product

Compute the dot product of two f32 slices.

f16_to_f32_batch

Convert a batch of f16 values (as raw u16 bits) to f32.

l2_distance

Compute L2 (Euclidean) distance between two vectors.

vector_norm

Compute the L2 norm (magnitude) of a vector.