def _test_quantitatively(sdfg):
graph = sdfg.nodes()[0]
A = np.random.rand(N.get()).astype(np.float64)
B = np.random.rand(N.get()).astype(np.float64)
C1 = np.random.rand(N.get()).astype(np.float64)
C2 = np.random.rand(N.get()).astype(np.float64)
D1 = np.random.rand(N.get()).astype(np.float64)
D2 = np.random.rand(N.get()).astype(np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, D=D1, N=N)
del csdfg
subgraph = SubgraphView(graph, [node for node in graph.nodes()])
me = MultiExpansion(subgraph)
assert me.can_be_applied(sdfg, subgraph) == True
me.apply(sdfg)
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph) == True
sf.apply(sdfg)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, D=D2, N=N)
assert np.allclose(C1, C2)
assert np.allclose(D1, D2)
def test_quantitatively(sdfg):
graph = sdfg.nodes()[0]
A = np.random.rand(N.get()).astype(np.float64)
B = np.random.rand(N.get()).astype(np.float64)
C1 = np.random.rand(N.get()).astype(np.float64)
C2 = np.random.rand(N.get()).astype(np.float64)
D1 = np.random.rand(N.get()).astype(np.float64)
D2 = np.random.rand(N.get()).astype(np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, D=D1, N=N)
subgraph = SubgraphView(graph, [node for node in graph.nodes()])
expansion = MultiExpansion()
fusion = SubgraphFusion()
assert expansion.match(sdfg, subgraph) == True
expansion.apply(sdfg, subgraph)
assert fusion.match(sdfg, subgraph) == True
fusion.apply(sdfg, subgraph)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, D=D2, N=N)
assert np.allclose(C1, C2)
assert np.allclose(D1, D2)
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 test_p1():
N.set(20)
M.set(30)
O.set(50)
P.set(40)
Q.set(42)
R.set(25)
sdfg = subgraph_fusion_parallel.to_sdfg()
sdfg.simplify()
state = sdfg.nodes()[0]
A = np.random.rand(N.get()).astype(np.float64)
B = np.random.rand(M.get()).astype(np.float64)
C = np.random.rand(O.get()).astype(np.float64)
D = np.random.rand(M.get()).astype(np.float64)
E = np.random.rand(N.get()).astype(np.float64)
F = np.random.rand(P.get()).astype(np.float64)
G = np.random.rand(M.get()).astype(np.float64)
H = np.random.rand(P.get()).astype(np.float64)
I = np.random.rand(N.get()).astype(np.float64)
J = np.random.rand(R.get()).astype(np.float64)
X = np.random.rand(N.get()).astype(np.float64)
Y = np.random.rand(M.get()).astype(np.float64)
Z = np.random.rand(P.get()).astype(np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C, D=D, E=E, F=F, G=G, H=H, I=I, J=J, X=X, Y=Y, Z=Z,\
N=N, M=M, O=O, P=P, R=R,Q=Q)
del csdfg
subgraph = SubgraphView(state, [node for node in state.nodes()])
expansion = MultiExpansion()
expansion.setup_match(subgraph)
fusion = SubgraphFusion()
fusion.setup_match(subgraph)
me = MultiExpansion()
me.setup_match(subgraph)
assert me.can_be_applied(sdfg, subgraph)
me.apply(sdfg)
sf = SubgraphFusion()
sf.setup_match(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
sf.apply(sdfg)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C, D=D, E=E, F=F, G=G, H=H, I=I, J=J, X=X, Y=Y, Z=Z,\
N=N, M=M, O=O, P=P, R=R,Q=Q)
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")
def test_p1():
N.set(20)
M.set(30)
O.set(50)
P.set(40)
Q.set(42)
R.set(25)
sdfg = program.to_sdfg()
sdfg.apply_strict_transformations()
state = sdfg.nodes()[0]
A = np.random.rand(N.get()).astype(np.float64)
B = np.random.rand(M.get()).astype(np.float64)
C = np.random.rand(O.get()).astype(np.float64)
D = np.random.rand(M.get()).astype(np.float64)
E = np.random.rand(N.get()).astype(np.float64)
F = np.random.rand(P.get()).astype(np.float64)
G = np.random.rand(M.get()).astype(np.float64)
H = np.random.rand(P.get()).astype(np.float64)
I = np.random.rand(N.get()).astype(np.float64)
J = np.random.rand(R.get()).astype(np.float64)
X = np.random.rand(N.get()).astype(np.float64)
Y = np.random.rand(M.get()).astype(np.float64)
Z = np.random.rand(P.get()).astype(np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C, D=D, E=E, F=F, G=G, H=H, I=I, J=J, X=X, Y=Y, Z=Z,\
N=N, M=M, O=O, P=P, R=R,Q=Q)
del csdfg
subgraph = SubgraphView(state, [node for node in state.nodes()])
expansion = MultiExpansion(subgraph)
fusion = SubgraphFusion(subgraph)
assert MultiExpansion.can_be_applied(sdfg, subgraph)
expansion.apply(sdfg)
assert SubgraphFusion.can_be_applied(sdfg, subgraph)
fusion.apply(sdfg)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C, D=D, E=E, F=F, G=G, H=H, I=I, J=J, X=X, Y=Y, Z=Z,\
N=N, M=M, O=O, P=P, R=R,Q=Q)
print("PASS")
def test_p1():
sdfg = disjoint_test_1.to_sdfg()
sdfg.simplify()
state = sdfg.nodes()[0]
assert len(sdfg.nodes()) == 1
A = np.random.rand(M.get(), 2).astype(np.float64)
A1 = A.copy()
A2 = A.copy()
csdfg = sdfg.compile()
csdfg(A=A1, N=N, M=M)
del csdfg
subgraph = SubgraphView(state, state.nodes())
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
sf.apply(sdfg)
csdfg = sdfg.compile()
csdfg(A=A2, M=M)
del csdfg
assert np.allclose(A1, A2)
