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 expand_maps(sdfg: dace.SDFG,
graph: dace.SDFGState,
subgraph: Union[SubgraphView, List[SubgraphView]] = None,
**kwargs):
subgraph = graph if not subgraph else subgraph
if not isinstance(subgraph, list):
subgraph = [subgraph]
trafo_expansion = MultiExpansion(subgraph[0])
for (property, val) in kwargs.items():
setattr(trafo_expansion, property, val)
for sg in subgraph:
map_entries = helpers.get_outermost_scope_maps(sdfg, graph, sg)
trafo_expansion.expand(sdfg, graph, map_entries)
def can_be_applied(self, sdfg: SDFG, subgraph: SubgraphView) -> bool:
graph = subgraph.graph
if self.allow_expansion == True:
subgraph_fusion = SubgraphFusion()
subgraph_fusion.setup_match(subgraph)
if subgraph_fusion.can_be_applied(sdfg, subgraph):
# try w/o copy first
return True
expansion = MultiExpansion()
expansion.setup_match(subgraph)
expansion.permutation_only = not self.expansion_split
if expansion.can_be_applied(sdfg, subgraph):
# deepcopy
graph_indices = [
i for (i, n) in enumerate(graph.nodes()) if n in subgraph
]
sdfg_copy = copy.deepcopy(sdfg)
graph_copy = sdfg_copy.nodes()[sdfg.nodes().index(graph)]
subgraph_copy = SubgraphView(
graph_copy, [graph_copy.nodes()[i] for i in graph_indices])
expansion.sdfg_id = sdfg_copy.sdfg_id
##sdfg_copy.apply_transformations(MultiExpansion, states=[graph])
#expansion = MultiExpansion()
#expansion.setup_match(subgraph_copy)
expansion.apply(sdfg_copy)
subgraph_fusion = SubgraphFusion()
subgraph_fusion.setup_match(subgraph_copy)
if subgraph_fusion.can_be_applied(sdfg_copy, subgraph_copy):
return True
stencil_tiling = StencilTiling()
stencil_tiling.setup_match(subgraph_copy)
if self.allow_tiling and stencil_tiling.can_be_applied(
sdfg_copy, subgraph_copy):
return True
else:
subgraph_fusion = SubgraphFusion()
subgraph_fusion.setup_match(subgraph)
if subgraph_fusion.can_be_applied(sdfg, subgraph):
return True
if self.allow_tiling == True:
stencil_tiling = StencilTiling()
stencil_tiling.setup_match(subgraph)
if stencil_tiling.can_be_applied(sdfg, subgraph):
return True
return False
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_warp_softmax(vector_length=1):
# Get SDFG
sdfg = softmax_fwd.to_sdfg(strict=True)
# Apply transformations
sdfg.apply_transformations_repeated(ReduceExpansion)
MultiExpansion.apply_to(sdfg, sdfg.node(0).nodes())
SubgraphFusion.apply_to(sdfg, sdfg.node(0).nodes())
sdfg.expand_library_nodes()
sdfg.apply_strict_transformations()
sdfg.apply_transformations_repeated([TrivialMapElimination, MapFusion])
sdfg.apply_transformations(GPUTransformSDFG)
sdfg.apply_transformations(WarpTiling)
sdfg.apply_transformations_repeated([HoistState, InlineSDFG, StateFusion],
strict=True)
sdfg.apply_transformations_repeated([TrivialMapElimination, MapFusion])
if vector_length != 1:
sdfg.apply_transformations_repeated(
Vectorization,
dict(vector_len=vector_length,
preamble=False,
postamble=False,
strided_map=False))
sdfg.specialize(dict(dn1=2, dn2=16, dn3=128, dr=128))
# Check validity
sdfg.validate()
assert sdfg.number_of_nodes() == 1
state = sdfg.node(0)
assert len([
c for c in state.scope_children()[None]
if isinstance(c, dace.nodes.MapEntry)
]) == 1
# Check correctness
inp = np.random.rand(2, 16, 128, 128).astype(np.float32)
out = np.random.rand(2, 16, 128, 128).astype(np.float32)
reg_out = softmax(inp)
sdfg(inp=inp, out=out)
assert np.allclose(out, reg_out, rtol=1e-4, atol=1e-6)
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 = 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")
