xt::where is very slow

[phmalek]

I found out, that the xt::where function is much slower than numpy.where. In my program, I call xt::where(xt::equal(index_i, tag)), where index_i.size() = 20855806. For each iteration, this call takes aroung 0.13 s which is a lot! Python accomplishes exactly this task much better than xtensor. Is it due to a bug or do you have any suggestion for implementation?

[JohanMabille]

Yes this is a bug that requires non trivial work in both xtensor and xsimd. We'll work on a patch in the next weeks.

[wolfv]

@JohanMabille I think it's not only the SIMD stuff, but also that we actually create real indices in where (note that he's referring to the 1-argument where).

We can speed it up by returning offset's in the where function. It's a simple change that will give a huge speed up.