SimdVectorOps

Namespace

Rowles.LeanLucene.Search.Simd

Assembly

Rowles.LeanLucene.dll

SIMD vector primitives used by HNSW graph operations and exact rerank. Built around System.Numerics.Vector`1, which the .NET JIT lowers to AVX2 or AVX-512 instructions automatically on capable hardware.

public static class SimdVectorOps

SimdVectorOps

Remarks

All methods assume the inputs have equal length and treat empty inputs as a similarity of zero.

Methods

CosineSimilarity(ReadOnlySpan<float>, ReadOnlySpan<float>)

Computes cosine similarity between two vectors. Returns zero for empty or length-mismatched inputs; never throws.

DotProduct(ReadOnlySpan<float>, ReadOnlySpan<float>)

Computes the dot product of two equal-length vectors. For pre-normalised vectors this is equivalent to cosine similarity at roughly two thirds of the cost.

Normalise(ReadOnlySpan<float>)

Allocates a normalised copy of the input vector. Throws when the input has zero norm.

NormaliseInPlace(Span<float>)

L2-normalises a vector in place. Returns false when the input has zero norm (cosine is undefined in that case); the buffer is left untouched.

SquaredNorm(ReadOnlySpan<float>)

Computes the squared L2 norm of a vector (sum of squared elements).