test_spectral_embedding_callable_affinity now runs in 800ms instead of 9s+ in master.

Note: 0.17 did not have such a strong perf issue, so no need to backport.

Travis has revealed that this code can be much worse than before with recent python / numpy / scipy:

test_spectral_embedding.test_spectral_embedding_precomputed_affinity: 51.6601s
test_spectral_embedding.test_spectral_embedding_callable_affinity: 51.0443s

although I don't reproduce this behavior on my local workstation which is very similar... More work needed.

@AlexandreAbraham maybe we should revert to the code of 0.17?

I changed the scipy version and it's still very slow on travis:

test_spectral_embedding.test_spectral_embedding_callable_affinity: 33.1952s
test_spectral_embedding.test_spectral_embedding_precomputed_affinity: 31.9874s

instead of 0.8s on my box with the same scipy version.

Lack of good linear algebra library?

The slow configuration the one is using MKL :) Let me try to replicate locally to see if MKL is the culprit. This is quite unlikely though at this is mostly about walking a graph with np.logical_or and a for loop...

Those tests run in 0.4ms with MKL on my box (same python / numpy / scipy versions from anaconda as on travis). This is really weird.

Scipy might be the culprit – I seem to remember some recent changes to how scipy.sparse does row access.

I have the same version of scipy on my local machine (0.16.1) and I don't have the problem.

Rebased on top of current master to see if the number of threads of MKL or OpenBLAS is related to this issue.

Ok the number of threads of MKL / OpenBLAS was the source of the problem. This is much faster now. Merging.