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 mpiize(iet, **kwargs):
"""
Add MPI routines performing halo exchanges to emit distributed-memory
parallel code.
"""
mode = kwargs.pop('mode')
# To produce unique object names
generators = {'msg': generator(), 'comm': generator(), 'comp': generator()}
sync_heb = HaloExchangeBuilder('basic', **generators)
user_heb = HaloExchangeBuilder(mode, **generators)
mapper = {}
for hs in FindNodes(HaloSpot).visit(iet):
heb = user_heb if isinstance(hs, OverlappableHaloSpot) else sync_heb
mapper[hs] = heb.make(hs)
efuncs = sync_heb.efuncs + user_heb.efuncs
objs = filter_sorted(sync_heb.objs + user_heb.objs)
iet = Transformer(mapper, nested=True).visit(iet)
# Must drop the PARALLEL tag from the Iterations within which halo
# exchanges are performed
mapper = {}
for tree in retrieve_iteration_tree(iet):
for i in reversed(tree):
if i in mapper:
# Already seen this subtree, skip
break
if FindNodes(Call).visit(i):
mapper.update({n: n._rebuild(properties=set(n.properties)-{PARALLEL})
for n in tree[:tree.index(i)+1]})
break
iet = Transformer(mapper, nested=True).visit(iet)
return iet, {'includes': ['mpi.h'], 'efuncs': efuncs, 'args': objs}
def test_iet_basic_sendrecv(self):
grid = Grid(shape=(4, 4))
t = grid.stepping_dim
f = TimeFunction(name='f', grid=grid)
heb = HaloExchangeBuilder()
sendrecv = heb._make_sendrecv(f, HaloSchemeEntry([t], []))
assert str(sendrecv.parameters) == """\
(f(t, x, y), buf_x_size, buf_y_size, ogtime, ogx, ogy, ostime, osx, osy,\
fromrank, torank, comm)"""
assert str(sendrecv.body[0]) == """\
def test_iet_basic_haloupdate(self):
grid = Grid(shape=(4, 4))
x, y = grid.dimensions
t = grid.stepping_dim
f = TimeFunction(name='f', grid=grid)
heb = HaloExchangeBuilder()
halos = [(x, LEFT), (x, RIGHT), (y, LEFT), (y, RIGHT)]
haloupdate = heb._make_haloupdate(f, HaloSchemeEntry([t], halos))
assert str(haloupdate.parameters) == """\
(f(t, x, y), comm, nb, otime)"""
assert str(haloupdate.body[0]) == """\
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}
def _dist_parallelize(self, iet):
"""
Add MPI routines performing halo exchanges to emit distributed-memory
parallel code.
"""
sync_heb = HaloExchangeBuilder('basic')
user_heb = HaloExchangeBuilder(self.params['mpi'])
mapper = {}
for i, hs in enumerate(FindNodes(HaloSpot).visit(iet)):
heb = user_heb if hs.is_Overlappable else sync_heb
mapper[hs] = heb.make(hs, i)
efuncs = sync_heb.efuncs + user_heb.efuncs
objs = sync_heb.objs + user_heb.objs
iet = Transformer(mapper, nested=True).visit(iet)
return iet, {'includes': ['mpi.h'], 'efuncs': efuncs, 'args': objs}
def test_iet_copy(self):
grid = Grid(shape=(4, 4))
t = grid.stepping_dim
f = TimeFunction(name='f', grid=grid)
heb = HaloExchangeBuilder()
gather = heb._make_copy(f, HaloSchemeEntry([t], []))
assert str(gather.parameters) == """\
(buf(buf_x, buf_y), buf_x_size, buf_y_size, f(t, x, y), otime, ox, oy)"""
assert """\
for (int x = 0; x <= buf_x_size - 1; x += 1)
{
for (int y = 0; y <= buf_y_size - 1; y += 1)
{
buf[x][y] = f[otime][x + ox][y + oy];
}
}""" in str(gather)