Matthias Krack

The reason for the numerical deviations on aarch64 is not understood yet. It affects, however, only a few (2-3 out of 3000+) tests related to features implemented quite recently.

Update: The following tests are still failing with GNU aarch64 ``` xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx /home/pi/github/cp2k/cp2k/regtesting/TEST-Linux-gnu-aarch64-psmp-2022-01-27_01-56-38/QS/regtest-rs-dhft/CH3-rsLDAlrMP2.inp.out Difference too large: 3.87e-04 > 1e-08, ref_value: -7.337744306690559, value: -7.340586990610659. xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx /home/pi/github/cp2k/cp2k/regtesting/TEST-Linux-gnu-aarch64-psmp-2022-01-27_01-56-38/QS/regtest-gw-cubic/G0W0_kpoints_from_Gamma.inp.out Difference too large: 2.92e-01 >...

> * wrong results in diagonalization Yes, that's a serious issue with ELPA. Recently, I relaxed the tolerance for `CHECK_DIAG` for the regression tests, because some tests showed deviations from...

I also strongly advocate to change the default for [PREFERRED_DIAG_LIBRARY](https://manual.cp2k.org/cp2k-8_1-branch/CP2K_INPUT/GLOBAL.html#PREFERRED_DIAG_LIBRARYl) from `ELPA` to `SL`

@oschuett Note, the original threshold for an abort was `1.0E-12` with CHECK_DIAG. I will also run again a regression test with ELPA as default. It might also matter how ELPA...

We can try ELPA again as the default with the original tight threshold of `1.0E-14` for `CHECK_DIAG`. In this way, we will detect potential issues related to ELPA in the...

> I'm not sure which warnings you mean? I remember that there have been warnings from the ELPA routines itself. Maybe, they were not important or unrelated. > Yeah, it's...

The CP2k regression tester run usually with only 2 MPI tasks and 2 threads per MPI task. There are indeed regression tests that do not run properly with more than...

When I run the [test case ](https://github.com/cp2k/cp2k/files/6419375/elpa1.txt) provided by @dev-zero, I observe the following behaviour for different numbers of MPI ranks (only one thread always) with a CP2K `Linux-x86-64-gfortran-regtest.psmp` binary:...

Which CP2K version did you try to install? The LIBINT make.log shows no error message or any other helpful information.