README
¶
Vendored SIMD distance kernels
The asm package in this directory is vendored from Weaviate:
- Source: https://github.com/weaviate/weaviate
- Path:
adapters/repos/db/vector/hnsw/distancer/asm - Commit:
6f5e0bb713661c0e1c8edcd4e13ff01025499ecc - License: BSD-3-Clause (see
LICENSE; copyright © Weaviate B.V.)
Only the dot-product and L2 kernels are vendored (float32, byte/uint8,
and mixed float×byte), for amd64 (AVX2/AVX512) and arm64 (NEON/SVE). The
hamming, prefetch, geo and the higher-level distancer Provider abstraction
were intentionally left out — any-store only needs the raw kernels.
Runtime CPU dispatch (which kernel to select) and all higher-level vector ops
(norm, cosine, normalize, residual, the float×int8 quantized path, and the
pure-Go fallback used on wasm / non-AVX2 / non-SIMD CPUs) live in the parent
internal/simd package, not here.
The .s files are committed assembly (GoAT/avo generated upstream); we do not
regenerate them, so no avo/goat build-time toolchain is required.
Do not edit these files by hand. To update, re-vendor from the upstream commit
above and re-run the internal/simd tests.
Documentation
¶
Overview ¶
asm only has amd64 specific implementations at the moment
Index ¶
- func Dot(x []float32, y []float32) float32
- func DotAVX256(x []float32, y []float32) float32
- func DotAVX512(x []float32, y []float32) float32
- func DotByteAVX256(x []uint8, y []uint8) uint32
- func DotFloatByteAVX256(x []float32, y []uint8) float32
- func L2(x []float32, y []float32) float32
- func L2AVX256(x []float32, y []float32) float32
- func L2AVX512(x []float32, y []float32) float32
- func L2ByteAVX256(x []uint8, y []uint8) uint32
- func L2FloatByteAVX256(x []float32, y []uint8) float32
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func DotByteAVX256 ¶
func DotFloatByteAVX256 ¶
func L2ByteAVX256 ¶
func L2FloatByteAVX256 ¶
Types ¶
This section is empty.
Source Files