def invoke_stencil(tile_size, offset=False, unroll=False, view=False):
A = np.random.rand(N.get()).astype(np.float64)
B1 = np.zeros((N.get()), dtype=np.float64)
B2 = np.zeros((N.get()), dtype=np.float64)
B3 = np.zeros((N.get()), dtype=np.float64)
if offset:
sdfg = stencil_offset.to_sdfg()
else:
sdfg = stencil.to_sdfg()
sdfg.apply_strict_transformations()
graph = sdfg.nodes()[0]
if view:
sdfg.view()
# baseline
sdfg.name = 'baseline'
csdfg = sdfg.compile()
csdfg(A=A, B=B1, N=N)
del csdfg
subgraph = SubgraphView(graph, [n for n in graph.nodes()])
st = StencilTiling(subgraph)
st.tile_size = (tile_size, )
st.schedule = dace.dtypes.ScheduleType.Sequential
assert st.can_be_applied(sdfg, subgraph)
if unroll:
st.unroll_loops = True
st.apply(sdfg)
if view:
sdfg.view()
sdfg.name = 'tiled'
sdfg.validate()
csdfg = sdfg.compile()
csdfg(A=A, B=B2, N=N)
del csdfg
assert np.allclose(B1, B2)
sdfg.apply_strict_transformations()
subgraph = SubgraphView(graph, [n for n in graph.nodes()])
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
# also test consolidation
sf.consolidate = True
sf.apply(sdfg)
sdfg.name = 'fused'
csdfg = sdfg.compile()
csdfg(A=A, B=B3, N=N)
del csdfg
print(np.linalg.norm(B1))
print(np.linalg.norm(B3))
assert np.allclose(B1, B2)
assert np.allclose(B1, B3)
print("PASS")
def invoke_stencil(tile_size, offset=False, unroll=False):
A = np.random.rand(N.get() * 2).astype(np.float64)
B1 = np.zeros((N.get()), dtype=np.float64)
B2 = np.zeros((N.get()), dtype=np.float64)
B3 = np.zeros((N.get()), dtype=np.float64)
if offset:
sdfg = stencil_offset.to_sdfg()
else:
sdfg = stencil.to_sdfg()
sdfg.simplify()
graph = sdfg.nodes()[0]
# baseline
sdfg.name = f'baseline_{tile_size}_{offset}_{unroll}'
csdfg = sdfg.compile()
csdfg(A=A, B=B1, N=N)
del csdfg
subgraph = SubgraphView(graph, [n for n in graph.nodes()])
st = StencilTiling()
st.setup_match(subgraph)
st.tile_size = (tile_size, )
st.unroll_loops = unroll
assert st.can_be_applied(sdfg, subgraph)
# change schedule so that OMP never fails
st.schedule = dace.dtypes.ScheduleType.Sequential
st.apply(sdfg)
sdfg.name = f'tiled_{tile_size}_{offset}_{unroll}'
csdfg = sdfg.compile()
csdfg(A=A, B=B2, N=N)
del csdfg
sdfg.simplify()
subgraph = SubgraphView(graph, [n for n in graph.nodes()])
sf = SubgraphFusion()
sf.setup_match(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
sf.apply(sdfg)
sdfg.name = f'fused_{tile_size}_{offset}_{unroll}'
csdfg = sdfg.compile()
csdfg(A=A, B=B3, N=N)
del csdfg
print(np.linalg.norm(B1))
print(np.linalg.norm(B3))
print("PASS")
def apply(self, sdfg):
subgraph = self.subgraph_view(sdfg)
graph = subgraph.graph
scope_dict = graph.scope_dict()
map_entries = helpers.get_outermost_scope_maps(sdfg, graph, subgraph,
scope_dict)
first_entry = next(iter(map_entries))
if self.allow_expansion:
expansion = MultiExpansion()
expansion.setup_match(subgraph, self.sdfg_id, self.state_id)
expansion.permutation_only = not self.expansion_split
if expansion.can_be_applied(sdfg, subgraph):
expansion.apply(sdfg)
sf = SubgraphFusion()
sf.setup_match(subgraph, self.sdfg_id, self.state_id)
if sf.can_be_applied(sdfg, self.subgraph_view(sdfg)):
# set SubgraphFusion properties
sf.debug = self.debug
sf.transient_allocation = self.transient_allocation
sf.schedule_innermaps = self.schedule_innermaps
sf.apply(sdfg)
self._global_map_entry = sf._global_map_entry
return
elif self.allow_tiling == True:
st = StencilTiling()
st.setup_match(subgraph, self.sdfg_id, self.state_id)
if st.can_be_applied(sdfg, self.subgraph_view(sdfg)):
# set StencilTiling properties
st.debug = self.debug
st.unroll_loops = self.stencil_unroll_loops
st.strides = self.stencil_strides
st.apply(sdfg)
# StencilTiling: update nodes
new_entries = st._outer_entries
subgraph = helpers.subgraph_from_maps(sdfg, graph, new_entries)
sf = SubgraphFusion()
sf.setup_match(subgraph, self.sdfg_id, self.state_id)
# set SubgraphFusion properties
sf.debug = self.debug
sf.transient_allocation = self.transient_allocation
sf.schedule_innermaps = self.schedule_innermaps
sf.apply(sdfg)
self._global_map_entry = sf._global_map_entry
return
warnings.warn("CompositeFusion::Apply did not perform as expected")