def test_quantitatively(sdfg, graph):
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)
out1_base = np.ndarray((N.get(), M.get()), np.float64)
out2_base = np.ndarray((1), np.float64)
out3_base = np.ndarray((N.get(), M.get(), O.get()), np.float64)
out1 = np.ndarray((N.get(), M.get()), np.float64)
out2 = np.ndarray((1), np.float64)
out3 = np.ndarray((N.get(), M.get(), O.get()), np.float64)
csdfg = sdfg.compile()
csdfg(A=A,
B=B,
C=C,
out1=out1_base,
out2=out2_base,
out3=out3_base,
N=N,
M=M,
O=O)
expand_reduce(sdfg, graph)
expand_maps(sdfg, graph)
subgraph = SubgraphView(graph, [node for node in graph.nodes()])
assert SubgraphFusion.match(sdfg, subgraph) == True
fusion(sdfg, graph)
sdfg.validate()
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C, out1=out1, out2=out2, out3=out3, N=N, M=M, O=O)
assert np.allclose(out1, out1_base)
assert np.allclose(out2, out2_base)
assert np.allclose(out3, out3_base)
print('PASS')
def test_p3(in_transient, out_transient):
sdfg = reduction_test_3.to_sdfg()
sdfg.apply_strict_transformations()
state = sdfg.nodes()[0]
A = np.random.rand(M.get(), N.get()).astype(np.float64)
B = np.random.rand(M.get(), N.get()).astype(np.float64)
C1 = np.zeros([N.get()], dtype=np.float64)
C2 = np.zeros([N.get()], dtype=np.float64)
C3 = np.zeros([N.get()], dtype=np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, N=N, M=M)
del csdfg
expand_reduce(sdfg,
state,
create_in_transient=in_transient,
create_out_transient=out_transient)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, N=N, M=M)
del csdfg
expand_maps(sdfg, state)
fusion(sdfg, state)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C3, N=N, M=M)
del csdfg
assert np.linalg.norm(C1) > 0.01
assert np.allclose(C1, C2)
assert np.allclose(C1, C3)
def test_subgraph():
A, expected = config()
B_init = np.random.rand(2)
graph = mapfission_sdfg()
graph.apply_transformations(MapFission)
dace.sdfg.propagation.propagate_memlets_sdfg(graph)
cgraph = graph.compile()
B = dcpy(B_init)
cgraph(A=A, B=B)
del cgraph
assert np.allclose(B, expected)
graph.validate()
subgraph = SubgraphView(graph.nodes()[0], graph.nodes()[0].nodes())
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(graph, subgraph)
fusion(graph, graph.nodes()[0], None)
ccgraph = graph.compile()
B = dcpy(B_init)
ccgraph(A=A, B=B)
assert np.allclose(B, expected)
graph.validate()
def test_quantitatively(sdfg, graph):
A = np.random.rand(M.get(), N.get()).astype(np.float64)
B1 = np.zeros(shape=[M.get(), N.get()], dtype=np.float64)
C1 = np.zeros(shape=[M.get(), N.get()], dtype=np.float64)
B2 = np.zeros(shape=[M.get(), N.get()], dtype=np.float64)
C2 = np.zeros(shape=[M.get(), N.get()], dtype=np.float64)
csdfg = sdfg.compile()
csdfg(A=A, B=B1, C=C1, N=N, M=M)
fusion(sdfg, graph)
csdfg = sdfg.compile()
csdfg(A=A, B=B2, C=C2, N=N, M=M)
assert np.allclose(B1, B2)
assert np.allclose(C1, C2)
def test_inputs_outputs():
"""
Test subgraphs where the computation modules that are in the middle
connect to the outside.
"""
sdfg = dace.SDFG('inputs_outputs_fission')
sdfg.add_array('in1', [2], dace.float64)
sdfg.add_array('in2', [2], dace.float64)
sdfg.add_scalar('tmp', dace.float64, transient=True)
sdfg.add_array('out1', [2], dace.float64)
sdfg.add_array('out2', [2], dace.float64)
state = sdfg.add_state()
in1 = state.add_read('in1')
in2 = state.add_read('in2')
out1 = state.add_write('out1')
out2 = state.add_write('out2')
me, mx = state.add_map('outer', dict(i='0:2'))
t1 = state.add_tasklet('t1', {'i1'}, {'o1', 'o2'}, 'o1 = i1 * 2; o2 = i1 * 5')
t2 = state.add_tasklet('t2', {'i1', 'i2'}, {'o1'}, 'o1 = i1 * i2')
state.add_memlet_path(in1, me, t1, dst_conn='i1', memlet=dace.Memlet.simple('in1', 'i'))
state.add_memlet_path(in2, me, t2, dst_conn='i2', memlet=dace.Memlet.simple('in2', 'i'))
state.add_edge(t1, 'o1', t2, 'i1', dace.Memlet.simple('tmp', '0'))
state.add_memlet_path(t2, mx, out1, src_conn='o1', memlet=dace.Memlet.simple('out1', 'i'))
state.add_memlet_path(t1, mx, out2, src_conn='o2', memlet=dace.Memlet.simple('out2', 'i'))
sdfg.apply_transformations(MapFission)
dace.sdfg.propagation.propagate_memlets_sdfg(sdfg)
# Test
A, B, C, D = tuple(np.random.rand(2) for _ in range(4))
expected_C = (A * 2) * B
expected_D = A * 5
csdfg = sdfg.compile()
C_cpy = deepcopy(C)
D_cpy = deepcopy(D)
csdfg(in1=A, in2=B, out1=C_cpy, out2=D_cpy)
del csdfg
assert np.allclose(C_cpy, expected_C)
assert np.allclose(D_cpy, expected_D)
subgraph = SubgraphView(sdfg.nodes()[0], sdfg.nodes()[0].nodes())
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
fusion(sdfg, sdfg.nodes()[0], None)
C_cpy = deepcopy(C)
D_cpy = deepcopy(D)
csdfg = sdfg.compile()
csdfg(in1=A, in2=B, out1=C_cpy, out2=D_cpy)
del csdfg
assert np.allclose(C_cpy, expected_C)
assert np.allclose(D_cpy, expected_D)
def test_offsets_array():
sdfg = dace.SDFG('mapfission_offsets2')
sdfg.add_array('A', [20], dace.float64)
sdfg.add_array('interim', [1], dace.float64, transient=True)
state = sdfg.add_state()
me, mx = state.add_map('outer', dict(i='10:20'))
t1 = state.add_tasklet('addone', {'a'}, {'b'}, 'b = a + 1')
interim = state.add_access('interim')
t2 = state.add_tasklet('addtwo', {'a'}, {'b'}, 'b = a + 2')
aread = state.add_read('A')
awrite = state.add_write('A')
state.add_memlet_path(aread,
me,
t1,
dst_conn='a',
memlet=dace.Memlet.simple('A', 'i'))
state.add_edge(t1, 'b', interim, None, dace.Memlet.simple('interim', '0'))
state.add_edge(interim, None, t2, 'a', dace.Memlet.simple('interim', '0'))
state.add_memlet_path(t2,
mx,
awrite,
src_conn='b',
memlet=dace.Memlet.simple('A', 'i'))
sdfg.apply_transformations(MapFission)
dace.propagate_memlets_sdfg(sdfg)
sdfg.validate()
# Test
A = np.random.rand(20)
expected = A.copy()
expected[10:] += 3
A_cpy = A.copy()
csdfg = sdfg.compile()
csdfg(A=A_cpy)
del csdfg
print(np.linalg.norm(A_cpy))
print(np.linalg.norm(expected))
assert (np.allclose(A_cpy, expected))
subgraph = SubgraphView(sdfg.nodes()[0], sdfg.nodes()[0].nodes())
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
fusion(sdfg, sdfg.nodes()[0], None)
A_cpy = A.copy()
csdfg = sdfg.compile()
csdfg(A=A_cpy)
assert (np.allclose(A_cpy, expected))
def test_quantitatively(sdfg, graph):
A = np.random.rand(N.get(), M.get(), O.get()).astype(np.float64)
B = np.random.rand(N.get(), M.get(), O.get()).astype(np.float64)
C1 = np.zeros([N.get(), M.get(), O.get()], dtype=np.float64)
C2 = np.zeros([N.get(), M.get(), O.get()], dtype=np.float64)
sdfg.validate()
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, N=N, M=M, O=O)
fusion(sdfg, graph)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, N=N, M=M, O=O)
assert np.allclose(C1, C2)
print('PASS')
def test_offsets():
sdfg = dace.SDFG('mapfission_offsets')
sdfg.add_array('A', [20], dace.float64)
sdfg.add_scalar('interim', dace.float64, transient=True)
state = sdfg.add_state()
me, mx = state.add_map('outer', dict(i='10:20'))
t1 = state.add_tasklet('addone', {'a'}, {'b'}, 'b = a + 1')
t2 = state.add_tasklet('addtwo', {'a'}, {'b'}, 'b = a + 2')
aread = state.add_read('A')
awrite = state.add_write('A')
state.add_memlet_path(aread,
me,
t1,
dst_conn='a',
memlet=dace.Memlet.simple('A', 'i'))
state.add_edge(t1, 'b', t2, 'a', dace.Memlet.simple('interim', '0'))
state.add_memlet_path(t2,
mx,
awrite,
src_conn='b',
memlet=dace.Memlet.simple('A', 'i'))
sdfg.apply_transformations(MapFission)
dace.propagate_memlets_sdfg(sdfg)
sdfg.validate()
# Test
A = np.random.rand(20)
A_cpy = A.copy()
expected = A.copy()
expected[10:] += 3
csdfg = sdfg.compile()
csdfg(A=A_cpy)
del csdfg
assert (np.allclose(A_cpy, expected))
fusion(sdfg, sdfg.nodes()[0], None)
csdfg = sdfg.compile()
A_cpy = A.copy()
csdfg(A=A_cpy)
assert (np.allclose(A_cpy, expected))
def test_sequential():
N.set(1000)
sdfg = test_program.to_sdfg()
state = 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)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, N=N)
fusion(sdfg, state)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, N=N)
assert np.allclose(C1, C2)
def test_2fuse():
sdfg = softmax.to_sdfg()
sdfg._name = 'softmax_2part'
sdfg.apply_strict_transformations()
X_in = np.random.rand(H.get(), B.get(), SN.get(),
SM.get()).astype(np.float32)
csdfg = sdfg.compile()
res1 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
subgraph = get_partition(sdfg, sdfg.nodes()[0])
expand_reduce(sdfg, sdfg.nodes()[0], subgraph)
expand_maps(sdfg, sdfg.nodes()[0], subgraph)
fusion(sdfg, sdfg.nodes()[0], subgraph)
csdfg = sdfg.compile()
res2 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
assert np.allclose(res1, res2)
print("PASS")
return
def test():
N.set(50)
sdfg = program.to_sdfg()
sdfg.apply_gpu_transformations()
state = sdfg.nodes()[0]
A = 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)
csdfg = sdfg.compile()
csdfg(A=A, C=C1, N=N)
del csdfg
fusion(sdfg, state)
csdfg = sdfg.compile()
csdfg(A=A, C=C2, N=N)
print(np.linalg.norm(C1))
print(np.linalg.norm(C2))
assert np.allclose(C1, C2)
def _test_quantitatively(sdfg, graph):
A = np.random.rand(N.get(), M.get(), O.get()).astype(np.float64)
B = np.random.rand(N.get(), M.get(), O.get()).astype(np.float64)
C1 = np.zeros([N.get(), M.get(), O.get()], dtype=np.float64)
C2 = np.zeros([N.get(), M.get(), O.get()], dtype=np.float64)
sdfg.validate()
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C1, N=N, M=M, O=O)
del csdfg
subgraph = SubgraphView(graph, graph.nodes())
sf = SubgraphFusion(subgraph)
assert sf.can_be_applied(sdfg, subgraph)
fusion(sdfg, graph)
csdfg = sdfg.compile()
csdfg(A=A, B=B, C=C2, N=N, M=M, O=O)
del csdfg
assert np.allclose(C1, C2)
print('PASS')
def test_1fuse():
sdfg = softmax.to_sdfg()
sdfg.name = 'softmax_fused'
sdfg.simplify()
X_in = np.random.rand(H.get(), B.get(), SN.get(), SM.get()).astype(np.float32)
csdfg = sdfg.compile()
res1 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
del csdfg
expand_reduce(sdfg, sdfg.nodes()[0])
expand_maps(sdfg, sdfg.nodes()[0])
fusion(sdfg, sdfg.nodes()[0])
csdfg = sdfg.compile()
res2 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
del csdfg
print(np.linalg.norm(res1))
print(np.linalg.norm(res2))
assert np.allclose(res1, res2)
print("PASS")
return
def test_1fuse():
sdfg = softmax.to_sdfg()
sdfg._name = 'softmax_fused'
sdfg.apply_strict_transformations()
X_in = np.random.rand(H.get(), B.get(), SN.get(),
SM.get()).astype(np.float32)
csdfg = sdfg.compile()
res1 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
expand_reduce(sdfg, sdfg.nodes()[0])
expand_maps(sdfg, sdfg.nodes()[0])
fusion(sdfg, sdfg.nodes()[0])
#sdfg.specialize({'SM':SM})
csdfg = sdfg.compile()
res2 = csdfg(X_in=X_in, H=H, B=B, SN=SN, SM=SM)
print(np.linalg.norm(res1))
print(np.linalg.norm(res2))
assert np.allclose(res1, res2)
print("PASS")
return
def test_qualitatively(sdfg, graph):
fusion(sdfg, graph)
sdfg.validate()
print("PASS")
out1 >> B[i]
out1 = in1 + 1
for i in dace.map[0:N]:
with dace.tasklet:
in1 << B[i]
out1 >> C[i]
out1 = in1 + 1
if __name__ == "__main__":
N.set(50)
sdfg = test_program.to_sdfg()
sdfg.apply_gpu_transformations()
state = sdfg.nodes()[0]
A = 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)
csdfg = sdfg.compile()
csdfg(A=A, C=C1, N=N)
fusion(sdfg, state)
csdfg = sdfg.compile()
csdfg(A=A, C=C2, N=N)
print(np.linalg.norm(C1))
print(np.linalg.norm(C2))
assert np.allclose(C1, C2)
def test_qualitatively(sdfg, graph):
expand_reduce(sdfg, graph)
expand_maps(sdfg, graph)
fusion(sdfg, graph)
sdfg.validate()
print("PASS")
