Passing COMM_WORLD to context constructors fails if mpi4py is installed after arbor is built

[Helveg]

Describe the bug

A regression occurred where MPI communicators can no longer be passed to context:

>>> import arbor
>>> from mpi4py.MPI import COMM_WORLD
>>> arbor.context(mpi=COMM_WORLD)
Traceback (most recent call last):
  File "<string>", line 1, in <module>
RuntimeError: mpi must be None, or an MPI communicator

Context
I'm on PizDaint's daint-gpu stack with Spack built Python.

[Helveg]

The CMake output reports -- Could NOT find PY_mpi4py (missing: PY_MPI4PY). So the ifdefs around all of the mpi4py comm logic are likely skipped. But installing mpi4py after building arbor seems like a valid user thing to do. Is detecting mpi4py at compile time really required to detect an mpi4py communicator at runtime?

[Helveg]

@bcumming any thoughts?

[bcumming]

@Helveg , sorry for the belated reply.

MPI4Py needs to be available if we are going to bake in support for Arbor, because it requires access to the C code of MPI4Py, specifically for the header files.

The reason being that MPI4Py does not expose the underlying MPI Communicator, the type of which also depends on the version of MPI that is being used (OpenMPI and MPICH use a pointer and an integer type respectively as the MPI_Comm type).

[Helveg]

Ok, could the error be clarified then? It's telling me I didn't pass an MPI communicator, but I did. Just that Arbor can't use mpi4py MPI communicators if it wasn't built with support for it.