Module simd_hadamard 

SIMD-Accelerated Walsh-Hadamard Transform

High-performance vectorized Hadamard transform using AVX2/AVX-512/NEON for maximum throughput during vector rotation.

Problem

Scalar Hadamard transform bottlenecks rotation:

• O(D log D) complexity per vector
• ~500ns per 768-dim vector (scalar)
• Becomes significant at high ingest rates

Solution

SIMD-accelerated butterfly operations:

• Process 8 floats per operation (AVX2)
• Process 16 floats per operation (AVX-512)
• Vectorized normalization
• In-place transformation

Performance

Dimension	Scalar (ns)	AVX2 (ns)	AVX-512 (ns)	Speedup
128	85	20	12	4-7×
768	520	95	55	5-9×
1536	1100	180	100	6-11×

Usage

use sochdb_vector::simd_hadamard::{hadamard_transform, HadamardKernel};

let kernel = HadamardKernel::detect();
let mut data = vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0];

kernel.transform(&mut data);

Structs

HadamardKernel

Hadamard transform kernel with automatic dispatch

Enums

SimdCapability

SIMD capability for Hadamard

Functions

hadamard_scalar

Scalar Hadamard transform

hadamard_transform

In-place Hadamard transform with auto-detected SIMD

hadamard_transform_batch

Batch Hadamard transform

simd_capability

Get cached SIMD capability