Matthias Krack
Matthias Krack
> I do not have commit rights actually. I can fix the conflicts though I will do the commit when you have resolved the conflicts.
Recently, I looked into that issue for the same reason, since Fermi-Dirac smearing is not optimal (e.g. long distribution tails) and it would be nice indeed to have MP smearing...
I cannot reproduce your result, E(xc) = 0, with your input using the current [CP2K release version 2022.1](https://github.com/cp2k/cp2k/releases/download/v2022.1/cp2k-2022.1-Linux-gnu-x86_64.ssmp). E(xc) is -4.24195 Hartree. [H2O_R2SCAN.zip](https://github.com/cp2k/cp2k/files/9266600/H2O_R2SCAN.zip)
A similar error still occurs with OpenBLAS 0.3.21 and GCC 12.2.0 (and MPICH 3.4.3) using OpenMP (2 MPI ranks and 2 threads per rank). It might be related to this...
It seems valgrind is only working with `Linux` currently and not yet with `Darwin` (see the discussion [here](https://github.com/LouisBrunner/valgrind-macos/issues/44)).
Here are backtrace and variable values for dbt_unittest.ssmp crashing with 2 OMP threads using the current CP2K trunk version: ``` export OMP_NUM_THREADS=2 VECLIB_MAXIMUM_THREADS=2; lldb exe/Darwin-aarch64/dbt_unittest.ssmp (lldb) target create "exe/Darwin-aarch64/dbt_unittest.ssmp" Current...
@oschuett thanks for the feedback. I agree that this issue is most likely a specific problem of Darwin/arm64, because the CP2K unit tests are running fine under Linux/x86_64 and Linux/arm64...
@hfp Yes, it is without LIBXSMM. Either OpenBLAS or the framework ACCELERATE from Apple is used which has no impact on the errors described here.
@oschuett Thank you for looking into this issue. It is most likely an Apple specific problem at low level related to malloc and/or the handling of OpenMP locks. I tried...
This is a known issue for aarch64 as shown by the [Linux-arm64-gnu.psmp (11.2.0)](https://www.cp2k.org/static/regtest/trunk/Linux-arm64-gnu/psmp/regtest-0) and [Linux-armv7l-gnu.psmp (8.3.0)](https://www.cp2k.org/static/regtest/trunk/Linux-armv7l-gnu/psmp/regtest-0) regression testers on the [CP2K dashboard](https://dashboard.cp2k.org) since a while. It could not be resolved...