Optimal build options for faiss CPU

[rom1504]

Summary

We're compiling faiss ourselves to produce binaries adapted for our environment, we used the information provided in https://github.com/facebookresearch/faiss/blob/a17a631dc326b3b394f4e9fb63d0a7af475534dc/INSTALL.md
Precisely we're using https://github.com/victor-paltz/JFaiss-CPU/blob/master/Dockerfile to compile.
We observe that our build is much slower (5 times) for search than the build you provide with conda.

Do you have any advice on options/environment to use to get this optimized build?

Platform

OS: centos 7

Faiss version: 1.6.3

Installed from: compiled ourself vs conda 1.6.3

Faiss compilation options: https://github.com/victor-paltz/JFaiss-CPU/blob/master/Dockerfile
We're using MKL

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

That's the benchmark script we used to measure the speed:
https://gist.github.com/rom1504/5a6a733dfd4772e497ea159e0395d0f9

Results:

	train	reconstruct	search
python (conda)	5 s	0.0024 ms	0.08 ms
custom build	30 s	0.0024 ms	0.38 ms

I would appreciate any advise on this, thanks!

[beauby]

The conda builds use MKL as a BLAS library, and have avx2 enabled. Those would be the main factors I think.

[rom1504]

Thanks, I'll check if we're properly using both MKL and AVX2

[mdouze]

At that point the AVX2 flags were not correct.

[rom1504]

@mdouze ah that's interesting.
Did something change about these options in the new build system with cmake ?

On my side I did confirm I have the problem even with (I think) the right flags for avx2 and mkl and the last hypothesis I had was that the environment I used to compile the lib (GitHub action) didn't have the right cpu options itself.