add missing int def header, fix install_name on mac, restore tests

[minrk]

• fix int size to default of 32 on Windows
• actually run mpi tests, which would have caught this

closes #100

closes #103 because the install_name was wrong (revealed by the re-enabled tests)

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-linting service.

I just wanted to let you know that I linted all conda-recipes in your PR (recipe) and found it was in an excellent condition.

[traversaro]

Thanks @minrk ! Should we mark as broken the previous Windows build with 64-bit index? Fortunately conda-forge/conda-forge-pinning-feedstock#5274 was not merged, so I think we do not have any package that built with 5.6.2 mumps packages.

[minrk]

@traversaro yeah, that would probably be prudent

[minrk]

linux mpich cross-compile builds are failing because of conda-forge/mpich-feedstock#86

linux openmpi builds are failing with:

+ /tmp/tmpmppakv2p/info/recipe/parent/mpiexec.sh -n 2 ./ssimpletest
--------------------------------------------------------------------------
There was a problem while initializing support for the CUDA reduction operations.
hostname:   9e20deb675ab
priority:   78
collective: scan
--------------------------------------------------------------------------

Program received signal SIGSEGV: Segmentation fault - invalid memory reference.

which I don't understand. I don't get why CUDA anything is being loaded, which obviously won't work.

[jakirkham]

cc @leofang (in case you have any ideas on the question above)

[leofang]

Clueless here too. Just wondering: what happens if you set sm,self instead?

[dalcinl]

@minrk @leofang The sm btl was available in v3, then they renamed it to vader in v4, and in v5 it is again named sm with vader as an alias scheduled for removal. Disclaimer: I'm not an Open MPI user, so I may have gotten any of these facts wrong.

Regarding the weird CUDA failure, maybe the way to go is to add --mca opal_cuda_support 0 in the mpiexec invokation within mpiexec.sh?

[leofang]

Regarding the weird CUDA failure, maybe the way to go is to add --mca opal_cuda_support 0 in the mpiexec invokation within mpiexec.sh?

Yeah let's try that, though if it works then I don't know why the default we set (CUDA off) did not kick in.

[minrk]

disabling cuda works, but realized vader was never actually tested on its own. So running that test, then I think this is ready to go. If vader fails, I don't understand the implications of why cuda's being loaded here but apparently not in other openmpi runs.

[minrk]

switch to vader seems to have fixed it. No idea why using tcp, which we've used for ages, triggers failure to load cuda. That's bizarre.