In #128, we refactored much of the _GPU_ code path to be valid when run on a cpu, but functions associated with faces shared across mpi ranks were not touched. Because of this, the _GPU_ code path was not correct on the cpu with mpi parallelism. This PR is intended to fix this issue.
Approach
The approach is the same as in #128, just applied to the shared face functions, e.g., DGNonLinearForm::sharedFaceIntegration_gpu.
Purpose
In #128, we refactored much of the
_GPU_code path to be valid when run on a cpu, but functions associated with faces shared across mpi ranks were not touched. Because of this, the_GPU_code path was not correct on the cpu with mpi parallelism. This PR is intended to fix this issue.Approach
The approach is the same as in #128, just applied to the shared face functions, e.g.,
DGNonLinearForm::sharedFaceIntegration_gpu.