Re: math/py-scikit-learn with OPENMP, at least on Linux?

["Dr. Thomas Orgis" <thomas.orgis%uni-hamburg.de@localhost>]

Am Sun, 23 May 2021 23:10:18 +0200
schrieb Thomas Klausner <wiz%NetBSD.org@localhost>: 

> I'm confused why there is no matching buildlink3.mk include for the
> openmp package.

Ah, parallel/openmp is pkgsrc's place for clang's libomp. Well, this
package is only sensible when using clang as compiler, as is the
provided system compiler on Darwin (I read).

When using GCC above 4.something, OpenMP is part of the compiler
install and parallel/openmp is not useful. Same with other toolchains
like Intel. I am not sure if a builtin or mk/openmp.bl3 heuristic to
decide whether to pull in parallel/openmp is something that does more
good than harm. We have other packages that just assume that OpenMP is
there, like math/openblas_openmp. Could be that we have something to
work out here.

> If openmp has portability issues - some packages have 'available.mk'
> files that collect the restriction in one place.

It would be horrible to unwittingly combine GCC's libgomp with clang's
libomp. Maybe this special case can work out somehow, but in general,
it's a runtime thing that is not to be mixed. It's actually problematic
that clang doesn't bundle it, apparently(?).

Can anyone chime in who is actually using parallel/openmp? Where?


Alrighty then,

Thomas

-- 
Dr. Thomas Orgis
HPC @ Universität Hamburg