JPelamatti

@regislebrun, do you see any reason the parallelization of the computeWeightMatrix method would return slightly different results than the sequential alternative in a few rare cases? I have noticed slight...

I understand that the order of computation can vary, but given that I'm creating a parallel routine with respect to indices, and not the computations themselves I should still obtain...

Also, given that the computations that are within the parallelization routine involve products between matrices (which should be LAPACK reliant I think), do I need to de-activate the LAPACK parallelization...

> do you have some benchmarking results ? Do you mean regarding the computation time performance (compared to the previous implementation), or regarding the difference in actual numerical results? Either...

I have created 2 separate 'isParallel()' methods, one for the actual permutation parallelization, and one for the parallelization of the weight matrices computation. This allows to independently check whether the...

For now they surely will, as by factorizing the generation of a single set of permutations for all output dimensions, the new version of the code will always produce slightly...

I updated the tests, which now run without errors. However, they appear to time out on the MacOS integration machine. Any idea regarding what might be causing this? These tests...

Apart from the weight function and the covariance models methods, for which we check whether they are thread safe, the only other 'non trivial' computations that are contained in the...

After some digging, there is 1 of the 6 tests on HSICEstimator, `pyinstallcheck_HSICEstimatorGlobalSensitivity_std`, that actually goes through on MacOS (github). However it takes 88 seconds. The exact same python test...

> so not a deadlock but only a timeout ? Yes, I believe it is an actual timeout due to a parallelization malfunction.