It would be awesome if this fixes #464!

I have rerun some of the failing calculations on Daint and obtained the correct results. Do you think it is reasonable to employ Scalapack instead of ELPA in all failing tests? I know for sure that some passing tests (mostly from RPA/MP2) definitely make use of eigenvalue decompositions.

Do you think it is reasonable to employ Scalapack instead of ELPA in all failing tests?

No, we should really get to the bottom of this. Otherwise I'd have very little trust in the results - despite some tests occasionally passing. Fortunately, most of the tests failed because ELPA itself returned an error code. So, those should be rather straightforward to investigate.

The timeouts are probably due to marekandreas/elpa#17.
I guess, until this point we never actually used ELPA's GPU kernels.

Apart from two special cases which require results by other regtests and the timeouts, there is only one kind of error in the bandred code of ELPA. I will consult the ELPA developers for that. It might take some time.

Apparently, the error is related to the 2stage solver because we observe timeouts if we employ the 1stage solver instead.

It seems #2447 fixed the timeouts. @fstein93, could you please revert back to the two-stage solver?

Apparently, everything works now. If there are no further issues raised, I will merge this PR.

I'm a bit disappointed that we don't see any improvements in our performance test.
Shouldn't this show up in the diag_cu144_broy.inp benchmark?

Might be but I don't know what kernel we had used before.