I'm not quite sure why the non-cray-mpichs have cuda_arch=80 hardcoded, but that seems like a leftover from the past so I've at least set openmpi to behave like cray-mpich, i.e. just use +cuda and cuda_arch from the environment. Apparently OpenMPI has ROCm support (https://docs.open-mpi.org/en/v5.0.x/tuning-apps/networking/rocm.html) but the openmpi spack package does not have ROCm support. I have not attempted to add that support. I guess it's fine if attempting to use openmpi +rocm will simply fail to concretize?
I'm not quite sure why the non-cray-mpichs have
cuda_arch=80
hardcoded, but that seems like a leftover from the past so I've at least setopenmpi
to behave likecray-mpich
, i.e. just use+cuda
andcuda_arch
from the environment. Apparently OpenMPI has ROCm support (https://docs.open-mpi.org/en/v5.0.x/tuning-apps/networking/rocm.html) but theopenmpi
spack package does not have ROCm support. I have not attempted to add that support. I guess it's fine if attempting to useopenmpi +rocm
will simply fail to concretize?Requires https://github.com/eth-cscs/alps-cluster-config/pull/5.
Ping @ajocksch and @boeschf.