Set the number of threads that OMP can use.

[arjenpdevries]

Finally found some time to try out the great PDX library!

The initial results were a little too impressive, as I noticed a 200x speedup (instead of the 5-7 reported). Then I realized I was working on a machine with 48 cores - which is still highly impressive to see how PDX uses them all to the max. But FAISS needs a little help to use internal multiprocessing.

The pull request uses the function from the FAISS python bindings to increase the number of threads for OMP, and sets it to half the available cores according to the python multiprocessing library.

(I'd understand if you do not want to do this by default, but would recommend documenting it if so.)

[arjenpdevries]

PS: also fixed a minor bug. Apologies for not using git branches correctly so it went into the same pullreq.

[lkuffo]

Hi Arjen,
This is a good observation. When I was using FAISS on an AMD Zen4 machine, it automatically used all of the available threads for index construction. Maybe this is not the case in some CPU architectures.

I was also wondering where you saw the 200x speedup? The examples are supposed to use FAISS by performing queries one by one (no multithreading).

I will, of course, accept the PR; thanks a lot! We are working to improve the code in a separate branch.
We are also working to implement our own IVF index construction phase and quantization algorithm (to use 8 and 4 bits per dimension)!

(Sorry for the long delay in answering; I went away for a long holiday!)

[arjenpdevries]

No problem! The 200x referred to my first trying the "simple" example that does not use FAISS; that used all cores, but then the FAISS vatiant did not do that.

I noticed the benchmark scripts are controlling that perfectly, indeed. (Hope you had a great holiday!)