Closed simonbyrne closed 10 months ago
MPItrampoline
always provides this function, even when the underlying MPI implementation does not. In this case, MPItrampoline
returns 0
as a fall-back option.
Given this, it might be that MPIwrapper
mis-detected whether this feature is supported when wrapping OpenMPI. Can you check? Do you still have the respective CMakeCache.txt
file from building MPIwrapper
? I have these lines in my CMakeCache.txt
for OpenMPI:
//Test MPI_HAS_QUERY_CUDA_SUPPORT
MPI_HAS_QUERY_CUDA_SUPPORT:INTERNAL=1
... or there is a bug in MPIwrapper
's CMakeLists.txt
file, not passing the value of MPI_HAS_QUERY_CUDA_SUPPORT
to the compiler. I think the respective lines in mpiwrapper_version.h.in
are missing.
I do have
//Test MPI_HAS_QUERY_CUDA_SUPPORT
MPI_HAS_QUERY_CUDA_SUPPORT:INTERNAL=1
This should now be correct in release 2.10.5 (see https://github.com/eschnett/MPIwrapper/pull/21).
... to explicit: This was an error in MPIwrapper, not MPItrampoline. You will need to use the newest release of MPIwrapper to wrap you MPI library again. You do not need to update MPItrampoline.
MPIX_Query_cuda_support
appears to return 0, even when the underlying library returns 1