facebookresearch / faiss

A library for efficient similarity search and clustering of dense vectors.

https://faiss.ai

MIT License

30.73k stars 3.58k forks source link

Would faiss be faster if SSE support added to Arm? #1778

Open gahoo opened 3 years ago

gahoo commented 3 years ago

Summary

I found a project that converts Intel SSE intrinsics to Arm/Aarch64 NEON intrinsics (sse2neon). Would faiss be faster if SSE support added to Arm with sse2neon?

Platform

CPU: aarch64

mdouze commented 3 years ago

Thanks for the reference! Most of the Faiss code is optimized for AVX rather than SSE (256 bit registers rather than 128). This can probably be emulated with NEON as well but not sure. The main parts of Faiss that are vector optimized are:

IndexScalarQuantizer
pq4_fast_scan*

gahoo commented 3 years ago

I found a new type of scaling vector instruction for ARM, called scalar vector extensions (SVE). Would this be more promising?

And there is another project called SIMDe which could transition SSE/AVX code to NEON.