def _make_parregion(self, partree, parrays):
arrays = [i for i in FindSymbols().visit(partree) if i.is_Array]
# Detect thread-private arrays on the heap and "map" them to shared
# vector-expanded (one entry per thread) Arrays
heap_private = [i for i in arrays if i._mem_heap and i._mem_local]
heap_globals = []
for i in heap_private:
if i in parrays:
pi = parrays[i]
else:
pi = parrays.setdefault(
i,
PointerArray(name=self.sregistry.make_name(),
dimensions=(self.threadid, ),
array=i))
heap_globals.append(Dereference(i, pi))
if heap_globals:
body = List(header=self._make_tid(self.threadid),
body=heap_globals + [partree],
footer=c.Line())
else:
body = partree
return OpenMPRegion(body, partree.nthreads)
def test_array():
grid = Grid(shape=(3, 3))
d = Dimension(name='d')
a = Array(name='a',
dimensions=grid.dimensions,
dtype=np.int32,
halo=((1, 1), (2, 2)),
padding=((2, 2), (2, 2)),
space='remote',
scope='stack')
pkl_a = pickle.dumps(a)
new_a = pickle.loads(pkl_a)
assert new_a.name == a.name
assert new_a.dtype is np.int32
assert new_a.dimensions[0].name == 'x'
assert new_a.dimensions[1].name == 'y'
assert new_a.halo == ((1, 1), (2, 2))
assert new_a.padding == ((2, 2), (2, 2))
assert new_a.space == 'remote'
assert new_a.scope == 'stack'
# Now with a pointer array
pa = PointerArray(name='pa', dimensions=d, array=a)
pkl_pa = pickle.dumps(pa)
new_pa = pickle.loads(pkl_pa)
assert new_pa.name == pa.name
assert new_pa.dim.name == 'd'
assert new_pa.array.name == 'a'
def _make_parregion(self, partree, parrays):
if not any(i.is_ParallelPrivate for i in partree.collapsed):
return self.Region(partree)
# Vector-expand all written Arrays within `partree`, since at least
# one of the parallelized Iterations requires thread-private Arrays
# E.g. a(x, y) -> b(tid, x, y), where `tid` is the ThreadID Dimension
exprs = FindNodes(Expression).visit(partree)
warrays = [i.write for i in exprs if i.write.is_Array]
vexpandeds = []
for i in warrays:
if i in parrays:
pi = parrays[i]
else:
pi = parrays.setdefault(i, PointerArray(name=self.sregistry.make_name(),
dimensions=(self.threadid,),
array=i))
vexpandeds.append(VExpanded(i, pi))
if vexpandeds:
init = c.Initializer(c.Value(self.threadid._C_typedata, self.threadid.name),
self.lang['thread-num'])
prefix = List(header=init,
body=vexpandeds + list(partree.prefix),
footer=c.Line())
partree = partree._rebuild(prefix=prefix)
return self.Region(partree)
def _make_parregion(self, partree, parrays):
arrays = [i for i in FindSymbols().visit(partree) if i.is_Array]
# Detect thread-private arrays on the heap and "map" them to shared
# vector-expanded (one entry per thread) Arrays
heap_private = [i for i in arrays if i._mem_heap and i._mem_local]
heap_globals = []
for i in heap_private:
if i in parrays:
pi = parrays[i]
else:
pi = parrays.setdefault(
i,
PointerArray(name=self.sregistry.make_name(),
dimensions=(self.threadid, ),
array=i))
heap_globals.append(HeapGlobal(i, pi))
if heap_globals:
init = c.Initializer(
c.Value(self.threadid._C_typedata, self.threadid.name),
self.lang['thread-num'])
prefix = List(header=init,
body=heap_globals + list(partree.prefix),
footer=c.Line())
partree = partree._rebuild(prefix=prefix)
return self.Region(partree)