def get_pattern_matches(self,
strict=False,
states=None,
patterns=None,
sdfg=None):
""" Returns all possible transformations for the current SDFG.
:param strict: Only consider strict transformations (i.e., ones
that surely increase performance or enhance
readability)
:param states: An iterable of SDFG states to consider when pattern
matching. If None, considers all.
:param patterns: An iterable of transformation classes to consider
when matching. If None, considers all registered
transformations in `Transformation`.
:param sdfg: If not None, searches for patterns on given SDFG.
:return: List of matching `Transformation` objects.
@see: Transformation
"""
sdfg = sdfg or self.sdfg
if states is None:
if patterns is None:
_patterns = self.stateflow_patterns
else:
_patterns = [
p for p in patterns if p in self.stateflow_patterns
]
for pattern in _patterns:
yield from pattern_matching.match_stateflow_pattern(
sdfg, pattern, strict=strict)
state_enum = []
if states is None:
for state_id, state in enumerate(sdfg.nodes()):
state_enum.append((state_id, state))
else:
for state in states:
state_id = sdfg.nodes().index(state)
state_enum.append((state_id, state))
if patterns is None:
_patterns = self.patterns
else:
_patterns = [p for p in patterns if p in self.patterns]
for state_id, state in state_enum:
for pattern in _patterns:
yield from pattern_matching.match_pattern(state_id,
state,
pattern,
sdfg,
strict=strict)
import dace
from dace.transformation.pattern_matching import match_pattern
from dace.transformation.dataflow import MapTiling
import numpy as np
@dace.program
def tile_twice_test(a: dace.float64[200]):
a *= 2.0
if __name__ == '__main__':
sdfg = tile_twice_test.to_sdfg()
sdfg.apply_strict_transformations()
sdfg.apply_transformations(MapTiling, options={'tile_sizes': (5, )})
for i, match in enumerate(match_pattern(sdfg.nodes()[0], MapTiling, sdfg)):
if i == 0: # Match the first map again
match.tile_sizes = (4, )
match.apply_pattern(sdfg)
A = np.random.rand(200)
expected = 2 * A
sdfg(a=A)
diff = np.linalg.norm(A - expected)
print('Difference:', diff)
exit(1 if diff > 1e-8 else 0)
A = np.random.rand(256, 256).astype(np.float32)
B = np.random.rand(256, 256).astype(np.float32)
expected_C = A @ B
C = np.zeros((256, 256), dtype=np.float32)
sdfg = mm_double_buffered.to_sdfg()
sdfg(A=A, B=B, C=C)
diff = np.linalg.norm(expected_C - C) / (256 * 256)
print('Difference (before):', diff)
# Apply local storage transformation on inner map (last two transformations)
sdfg.apply_strict_transformations()
for i in range(2):
for match in reversed(
list(match_pattern(sdfg.node(0), InLocalStorage, sdfg))):
match.apply(sdfg)
break
else:
raise ValueError('Local storage transformation not applied')
applied = sdfg.apply_transformations(DoubleBuffering)
if applied != 1:
raise ValueError('Double-buffering transformation not applied')
C = np.zeros((256, 256), dtype=np.float32)
sdfg(A=A, B=B, C=C)
diff2 = np.linalg.norm(expected_C - C) / (256 * 256)
print('Difference (after):', diff2)
exit(1 if (diff > 1e-5 or diff2 > 1e-5) else 0)