def test_prune_k_cache_fills_backward():
testee = VerticalLoopFactory(
loop_order=LoopOrder.BACKWARD,
sections__0__horizontal_executions__0__body=[
AssignStmtFactory(left__name="foo",
right__name="foo",
right__offset__k=-1),
AssignStmtFactory(left__name="bar",
right__name="bar",
right__offset__k=0),
AssignStmtFactory(left__name="baz",
right__name="baz",
right__offset__k=1),
AssignStmtFactory(left__name="barbaz", right__name="bar"),
AssignStmtFactory(left__name="barbaz",
right__name="barbaz",
right__offset__k=1),
],
caches=[
KCacheFactory(name="foo", fill=True),
KCacheFactory(name="bar", fill=True),
KCacheFactory(name="baz", fill=True),
KCacheFactory(name="barbaz", fill=True),
],
)
transformed = PruneKCacheFills().visit(testee)
cache_dict = {c.name: c for c in transformed.caches}
assert cache_dict["foo"].fill
assert cache_dict["bar"].fill
assert not cache_dict["baz"].fill
assert not cache_dict["barbaz"].fill
def _optimize_oir(self, oir):
oir = GreedyMerging().visit(oir)
oir = AdjacentLoopMerging().visit(oir)
oir = LocalTemporariesToScalars().visit(oir)
oir = WriteBeforeReadTemporariesToScalars().visit(oir)
oir = OnTheFlyMerging().visit(oir)
oir = NoFieldAccessPruning().visit(oir)
oir = IJCacheDetection().visit(oir)
oir = KCacheDetection().visit(oir)
oir = PruneKCacheFills().visit(oir)
oir = PruneKCacheFlushes().visit(oir)
return oir
def _optimize_oir(self, oir):
oir = optimize_horizontal_executions(oir, GraphMerging)
oir = AdjacentLoopMerging().visit(oir)
oir = LocalTemporariesToScalars().visit(oir)
oir = WriteBeforeReadTemporariesToScalars().visit(oir)
oir = OnTheFlyMerging().visit(oir)
oir = MaskStmtMerging().visit(oir)
oir = NoFieldAccessPruning().visit(oir)
oir = IJCacheDetection().visit(oir)
oir = KCacheDetection().visit(oir)
oir = PruneKCacheFills().visit(oir)
oir = PruneKCacheFlushes().visit(oir)
return oir