Open pbartholomew08 opened 4 months ago
We may want to use nccl
in CUDA backend and rccl
for AMD and I guess they don't have a CartComm
like support, I cound't find anything online. Also, with the iterative solver I think we'll be able to run cases that doesn't really a good fit for CartComm
. For example a backward facing step, where there is no mesh and no ranks in that gap part. A custom communicator is still an option here and I think at least nccl
supports it, but not sure if its any better than current approach.
side benefit(/drawback) of #113, cartesian comm from 2decomp will be used for the OMP backend when using FFTs
tdsops_t
contains information about where in the parallel decomposition it is. This is very similar to the MPI_CartComm and it may be beneficial to make use of this.