Documentation
¶
Overview ¶
Package simd is any-store's single home for vector distance math.
It wraps the vendored, hand-written SIMD kernels in internal/simd/asm (dot-product and L2 for amd64 AVX2/AVX512 and arm64 NEON/SVE, float32 and int8) behind a small, portable API. Runtime CPU dispatch happens once in an arch-specific init (dispatch_*.go); when no SIMD kernel is available — wasm, 386, or an x86 without AVX2 (e.g. pre-Haswell Ivy Bridge) — the pure-Go fallbacks in fallback.go are used, which beat a naive scalar loop.
This package replaces the former github.com/viterin/vek/vek32 dependency, which was amd64+AVX2-only and fell back to slow portable Go on arm64.
Index ¶
- func Accelerated() bool
- func AcceleratedFloatByte() bool
- func CosineSimilarity(a, b []float32) float32
- func Distance(a, b []float32) float32
- func Dot(a, b []float32) float32
- func DotFloatByte(q []float32, u []byte) float32
- func L2Squared(a, b []float32) float32
- func MulNumber_Into(dst, v []float32, s float32) []float32
- func Norm(v []float32) float32
- func NormalizeInto(dst, v []float32) []float32
- func Sub_Into(dst, a, b []float32)
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func Accelerated ¶
func Accelerated() bool
Accelerated reports whether a hand-written SIMD kernel (not the pure-Go fallback) was selected for this CPU. Mirrors the old vek32.Info().Acceleration.
func AcceleratedFloatByte ¶
func AcceleratedFloatByte() bool
AcceleratedFloatByte reports whether a SIMD float×byte kernel is active (false on the pure-Go fallback path, and on arches where it is not wired).
func CosineSimilarity ¶
CosineSimilarity returns the cosine similarity of a and b. For a zero-length operand it returns 0 (so a cosine distance of 1-sim is maximal, matching the previous behavior for degenerate vectors).
func Distance ¶
Distance returns the Euclidean distance. Kept as sqrt(L2Squared) so the returned/stored distance values stay identical to the previous vek32.Distance.
func DotFloatByte ¶
DotFloatByte returns the inner product of a float32 vector q and a byte vector u, where each byte is interpreted as an UNSIGNED value 0..255 (matching the vendored weaviate kernels). Callers using offset-binary int8 storage recover the signed dot via scale·(DotFloatByte(q,u) − 128·Σq); see internal/vindex.
func MulNumber_Into ¶
MulNumber_Into writes dst = v * s and returns dst (which must have len(v)). Drop-in for vek32.MulNumber_Into. Cold path (per-vector normalize), so a plain loop the compiler can auto-vectorize is enough.
func NormalizeInto ¶
NormalizeInto writes the unit-normalized v into dst (reusing dst's capacity) and returns it. A zero vector is copied through unchanged (norm 0 -> leave as zeros). This is the single shared copy of the per-package normalizeInto that previously lived in both internal/vindex and internal/vivf.
Types ¶
This section is empty.
Source Files
Directories