OpenBLAS 0.3.18 LAPACK tests failed due to numerical errors

[Ezimkin]

Hello all,

I'm having some issues building the dependencies for LAMMPS-23Jun2022-foss-2021b-kokkos-CUDA-11.4.1.eb. In particular, OpenBLAS version 0.3.18 fails due to too many numerical error occurrences during testing.

I assume there is due to some imprecise flags given to the compiler? If it is system dependent then I will mention that this is on an AMD EPYC 7713 server.

			-->   LAPACK TESTING SUMMARY  <--
SUMMARY             	nb test run 	numerical error   	other error  
================   	===========	=================	================  
REAL             	1294329		1233	(0.095%)	0	(0.000%)	
DOUBLE PRECISION	1302917		1195	(0.092%)	0	(0.000%)	
COMPLEX          	756180		1005	(0.133%)	0	(0.000%)	
COMPLEX16         	768848		144	(0.019%)	0	(0.000%)	

--> ALL PRECISIONS	4122274		3577	(0.087%)	0	(0.000%)	

Here is the full log file from the build.

easybuild-OpenBLAS-0.3.18-20221027.182153.scXWI.log

[akesandgren]

If you're not using EasyBuild 4.6.2 then upgrade to that and recompile GCCcore and OpenBLAS for that (and any other toolchain you have that uses GCC 11/12)
There is a vectorizer bug in the stock GCC 11/12 and a FMA problem in the LAPACK part of OpenBLAS that got fixed in EasyBUild 4.6.2

[Ezimkin]

Ah, that may be a solution. I'm not sure what version it was from since i upgraded eb a few days before this but GCC and GCCcore were definitely using that older build.

I will build it this weekend and report back if it resolves my issue.

[Ezimkin]

I got my results back a earlier than expected. I can confirm that recompilation of GCC and GCCcore resolved this issue.

Thank you @akesandgren !

[boegel]

More details on this in #16380