Open barracuda156 opened 1 week ago
@pghysels 7.2.0 configures normally on the same system, so it is not something broken in MacPorts set-up. 8.0.0 fails both locally on Sonoma and on CI.
It looks like Scotch is searching for MPI. Can you try disabing Scotch (besides PT-Scotch)
It looks like Scotch is searching for MPI
This is what I thought, but 7.2.0 has the same config settings, and everything works with the same version of scotch
.
7.2.0: https://github.com/macports/macports-ports/blob/962f38384d582875b5504b7daa73cc4168db6d95/math/strumpack/Portfile And this is 8.0.0 update: https://github.com/macports/macports-ports/pull/26456/commits/54391705aedfccef9d484f414f957b6c8239e9cd
P. S. Yeah, we could probably disable scotch
or make MPICH required, but it seems there is a regression here, otherwise I cannot explain why 7.2.0 works with the same scotch
, Clang and CMake.
There are some changes here related to Scotch, which now uses CMake: https://github.com/pghysels/STRUMPACK/commit/099e3503de64dec98652553ec4a79536e4e6b1b0 What version of scotch are you using, was scotch compiled with MPI?
Yes, scotch
is built with MPICH. I just updated it few days ago to 7.0.5: https://github.com/macports/macports-ports/commit/f0a0bbbfc3237d7f2ada673ea7cc3f857b3ca763
There are some changes here related to Scotch, which now uses CMake
Ok, so looks like CMake module does not respect configure argument and just forces MPI.
Does it work for you if you disable MPI in strumpack
, but build scotch
with MPI?
If it is a bug in CMake, I can open an issue with its upstream.
The check for MPI is in /opt/local/lib/cmake/scotch/SCOTCHConfig.cmake
, because your Scotch was build with MPI.
Looks like there is no way around that.
@pghysels Do you know if it is actually required though? If not, then it should be fixed in CMake, if possible. Or maybe restore a local module here, which worked fine.
You can reinstall Scotch without MPI. Or build STRUMPACK with MPI (looks like you have MPI).
Or you can report it to Scotch.
If you build Scotch with MPI then @BUILD_PTSCOTCH@
here:
https://gitlab.inria.fr/scotch/scotch/-/blob/master/cmake/SCOTCHConfig.cmake.in?ref_type=heads#L61
will be ON
.
Maybe Scotch can split their SCOTCHConfig.cmake into sequential Scotch and distributed PT-Scotch?
Any idea why it does not respect disabling MPICH?