embedding-library

SIMD-accelerated primitives for working with dense vector embeddings in C. Header-only, portable (x86 & ARM), and released under the Apache-2.0 license.

---

Highlights

Capability	Float32	Int16	Int8
Dot product	dot_product() (AVX-512 / AVX2 / NEON / scalar)	—	int8_dot_product() (AVX-512 / AVX2 / NEON / scalar)
Cosine similarity	cosine_similarity()	—	int8_cosine_similarity()
Quantise ←→ de-quantise	—	int16_from_floats(), int16_to_floats()	int8_from_floats(), int8_from_int16s(), int8_to_floats()
Embedding table (512-dim, int8)	—	—	int8_embedding_table_*() incl. serialization

Auto-dispatch picks the fastest implementation available at compile time; you only call the generic functions.

---

Quick start

#include "embedding-library/float.h"
#include "embedding-library/int8.h"

float  a[512], b[512];           // fill with your vectors
float  cos_f32   = cosine_similarity(a, b, 512);

int8_t qa[512], qb[512];
int8_from_floats(a, 512, qa);
int8_from_floats(b, 512, qb);

float norm_qa = sqrtf(int8_dot_product(qa, qa, 512));
float norm_qb = sqrtf(int8_dot_product(qb, qb, 512));
float cos_i8  = int8_cosine_similarity(qa, norm_qa, qb, norm_qb, 512);

Build

# GCC / Clang
cc -std=c11 -O3 -mavx2 demo.c -o demo          # x86-64 w/ AVX2
cc -std=c11 -O3 -march=armv8.2-a+simd demo.c   # ARM w/ NEON

The library is header-only; just add include/ to your compiler’s search path.

---

Embedding tables

Efficiently store many 512-dim int8 embeddings:

#include "embedding-library/int8_embedding_table.h"

int8_embedding_table_t *tbl = int8_embedding_table_init(/*rows*/1024);

int8_t vec[512];   double norm = …;             // prepare vector + its L2 norm
int64_t idx = int8_embedding_table_add_embedding(tbl, vec, norm);

double sim = int8_embedding_table_cosine_similarity(tbl, idx, 42);
int8_embedding_table_serialize(tbl, "tbl.bin");

int8_embedding_table_destroy(tbl);

Uses chunky allocation (512-row nodes) for cache locality.

---

Design notes

• Portable SIMD – AVX-512, AVX2 and ARM NEON kernels with scalar fall-backs.
• Quantisation – simple symmetric linear scaling; pick int16 first when range matters, then down-cast to int8.
• Header-only – ideal for integration into existing C / C++ codebases.

---

License & attribution

Apache-2.0 © 2024-2025 Knode.ai Maintainer: Andy Curtis [email protected]