The default layout for Kokkos View is LayoutLeft for CUDA backend and RightLayout for CPU. Since the initial code I used was assuming a fixed layout for the sendrecv in MPI the runs were giving wrong data when using multiple MPI processes on CPU.
Fixed for now but we need to figure out an automatic way of adapting to the default layout without forcing one (since we might get bad performance on some architectures)
The default layout for Kokkos View is
LayoutLeftfor CUDA backend andRightLayoutfor CPU. Since the initial code I used was assuming a fixed layout for the sendrecv in MPI the runs were giving wrong data when using multiple MPI processes on CPU.Fixed for now but we need to figure out an automatic way of adapting to the default layout without forcing one (since we might get bad performance on some architectures)