Matthijs Douze
Matthijs Douze
hopefully this will finally fix https://github.com/facebookresearch/faiss/issues/2996
Please don't add a [faiss/python/swigfaiss_sve.swig](https://github.com/facebookresearch/faiss/pull/2886/files#diff-513bfb6ac32baa2d956a83fd054c99d286250c94f0e9b07c737b14858473b81b) file.
So the diff only changes the compilation flags, it does not add VSE specific SIMD implementations, right? Do you have hardware to try it on and maybe measure performance improvements?
Sorry for being a bit slow to react. I think that it's fine to land this packaging PR first, let us check the implications in terms of library size.
OK so I think a way to move forward is to accept this PR but not cover it with CI. Then optimized code for SVE can be contributed. At some...
Would you mind rebasing on the latest Faiss so that I can import it to the internal Faiss version? Thanks
We are looking into compiling this in the CI @ramilbakhshyiev
IDMap is a CPU index type and does not have an equivalent for GPU. So it does not get translated by index_cpu_to_gpu.
This is because the `distance_to_code` function does not have a simplified wrapping layer. It can still be called using `swig_ptr`: ```python dist_comp = index_pq.get_FlatCodesDistanceComputer() dist_comp.set_query(faiss.swig_ptr(x_norm[0])) dist_comp.distance_to_code(faiss.swig_ptr(codes[:5])) ``` This computes one...
This seems to be a "new-object" specific feature. Would it work with our "old-object" model?