def _generate_mpi(self, iet, **kwargs):
# Drop superfluous HaloSpots
halo_spots = FindNodes(HaloSpot).visit(iet)
mapper = {i: None for i in halo_spots if i.is_Redundant}
iet = Transformer(mapper, nested=True).visit(iet)
# Nothing else to do if no MPI
if configuration['mpi'] is False:
return iet
# Build halo exchange Callables and Calls
halo_spots = FindNodes(HaloSpot).visit(iet)
heb = HaloExchangeBuilder(is_threaded(kwargs.get("dle")))
callables, calls = heb.make(halo_spots)
# Update the Operator internal state
self._includes.append('mpi.h')
self._func_table.update(
OrderedDict([(i.name, MetaCall(i, True)) for i in callables]))
# Transform the IET by adding in the `haloupdate` Calls
mapper = {k: List(body=v + list(k.body)) for k, v in calls.items()}
iet = Transformer(mapper, nested=True).visit(iet)
return iet
def _dist_parallelize(self, iet):
"""
Add MPI routines performing halo exchanges to emit distributed-memory
parallel code.
"""
# Build send/recv Callables and Calls
heb = HaloExchangeBuilder(self.params['mpi'])
call_trees, calls = heb.make(FindNodes(HaloSpot).visit(iet))
# Transform the IET by adding in the `haloupdate` Calls
iet = Transformer(calls, nested=True).visit(iet)
return iet, {'includes': ['mpi.h'], 'call_trees': call_trees}