BUG: Fixes for ARM architecture

[larsoner]

Bad things found so far on my M1 machine:

• We can't rely on LinAlgError to trigger on np.linalg.solve for rank deficiency on arm64. I changed the mxne code to use a SVD, which actually seems faster (!?) so I think this is a tolerable change at least. This probably deserves an upstream bug report at some point, but isolating it isn't the easiest thing...
• We can't rely on SVD: BUG: linalg.svd broken on M1 / arm64 numpy/numpy#21756

The second one is a show stopper. I don't think we can safely release or advocate for M1-based installers until it's fixed.

There are a bunch of test_linalg.py errors that I see locally that are likely related, but I don't want to bother looking into them until we can trust svd.

Helps with #10759

[larsoner]

@hoechenberger or @agramfort if either of you can replicate numpy/numpy#21756 locally and can comment on the NumPy issue, it might help move things forward -- and also make sure I'm not doing something wrong!

[hoechenberger]

• We can't rely on SVD: BUG: linalg.svd broken on M1 / arm64 numpy/numpy#21756

OMG this is horrible 😱 😱

[hoechenberger]

Seems to be a conda-forge problem, we should open an issue with them

numpy/numpy#21756 (comment)

[hoechenberger]

@larsoner Shall we briefly catch up on Discord?

[hoechenberger]

For all the others following this thread, the problem seems to be related to OpenBLAS, and switching to a different BLAS implementation seems to fix at least this here test case.

[larsoner]

Sorry I was fighting with https://gitlab.kitware.com/vtk/vtk/-/issues/18559 for the last hour instead :)

I think the most likely scenario is that it's an OpenBLAS issue that probably needs to be fixed at the OpenBLAS end. But now that it has been replicated on two machines, the NumPy folks should know how to move things forward. So I think we wait a couple of days and see if they make progress. In the meantime I think we wait on the M1 installers :(

[hoechenberger]

Can you try to conditionally use accelerate on ARM64 only, which is Apple's own implementation of BLAS?

[larsoner]

Can you try to conditionally use accelerate on ARM64 only, which is Apple's own implementation of BLAS?

I guess we could try this, but IIRC Accelerate was not really well supported by SciPy last time I saw it discussed. Feel free to try locally and run make pytest to see what fails. I'll give it a shot tomorrow.

[larsoner]

@agramfort feel free to merge if you're okay with the mxne changes

[agramfort]

just a nitpick. than +1 for MRG on my side

thx @larsoner

[hoechenberger]

Great work, @larsoner!