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.simplify()
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()
st.setup_match(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.simplify()
subgraph = SubgraphView(graph, [n for n in graph.nodes()])
sf = SubgraphFusion()
sf.setup_match(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")