dbcsr #857: conditionally rely on PARALLEL WORKSHARE

Oct 16 '24 08:10 hfp

Thanks @hfp I will take a look next week...

Oct 16 '24 10:10 alazzaro

Thanks @hfp I will take a look next week...

Might be easier now (less work, only three files). I separated some cleanup work (#860).

Oct 16 '24 10:10 hfp

@mkrack this single PR is apparently resolving all runtime issues with Intel Fortran Compiler (IFX) in CP2K/DBCSR. There is still the question why PARALLEL WORKSHARE DEFAULT(none) SHARED(...) IF(spawn) with spawn=.NOT. omp_in_parallel() does not work as opposed to the code shown here (but this might be well-covered by a bug report). Further, see https://github.com/cp2k/dbcsr/issues/857#issuecomment-2416094575 - it can be of interest for CP2K.

Oct 16 '24 10:10 hfp

@alazzaro any feedback welcome - so far, no issues with this code on my side, I am actively using it since sending the PR.

Oct 21 '24 10:10 hfp

Objective of this implementation is to exercise a way to preserve the (parallel) workshare. It does not judge (or confirms) whether the workshare is beneficial or not.

Oct 23 '24 06:10 hfp

@hfp thanks, I will review everything next week and make a new RC that we can push to CP2K for more testings

Oct 23 '24 07:10 alazzaro

Hold on merging this PR. I pulled in more expertise. I will reply on the original issue https://github.com/cp2k/dbcsr/issues/857, and revise/close this PR accordingly.

Nov 25 '24 10:11 hfp

The revised PR now drops the [PARALLEL] WORKSHARE constructs.

Dec 06 '24 20:12 hfp