def source_code_location(event: _ProfilerEvent):
while (event is not None):
match = re.search(r"\.py\(.*\)", event.name())
if (match is None):
event = event.parent
continue
return event.name()
return "No source code location found"
def match(self, event: _ProfilerEvent):
if event.name() != "aten::conv2d":
return False
if len(input_dtypes(event)) < 3 or input_dtypes(event)[2] == "":
return False
# This means bias=True
event = self.go_up_until(
event, lambda e: e.name().startswith("nn.Module: Conv2d"))
if not event:
return False
event = self.next_of(event)
if not event:
return False
return event.name().startswith("nn.Module: BatchNorm2d")
def match(self, event: _ProfilerEvent):
if event.name() != "aten::select":
return False
if event.id in self.visited:
return False
repeat_count = 1
_, next = self.siblings_of(event)
if len(next) <= 1:
return False
# Custom event list matching
def same_ops(list1, list2):
if len(list1) != len(list2):
return False
for op1, op2 in zip(list1, list2):
if op1.name() != op2.name():
return False
return True
# Record the ops between two aten::select
next_select_idx = index_of_first_match(
next, lambda e: e.name() == "aten::select")
if next_select_idx is None:
return False
indexing_ops = [event] + next[:next_select_idx]
next = next[len(indexing_ops) - 1:]
for i in range(0, len(next), len(indexing_ops)):
if same_ops(indexing_ops, next[i:i + len(indexing_ops)]):
repeat_count += 1
self.visited.add(next[i].id)
else:
break
return repeat_count >= 10
def match(self, event: _ProfilerEvent):
def is_dataloader_function(name: str, function_name: str):
return name.startswith(
os.path.join("torch", "utils", "data",
"dataloader.py")) and name.endswith(function_name)
if not is_dataloader_function(event.name(), "__iter__"):
return False
if not event.children:
return False
event = event.children[0]
if not is_dataloader_function(event.name(), "_get_iterator"):
return False
if not event.children:
return False
event = event.children[0]
return not is_dataloader_function(event.name(),
"check_worker_number_rationality")
def match(self, event: _ProfilerEvent):
# If we saw this pattern once, we don't need to match it again
if event.tag != _EventType.TorchOp:
return False
assert isinstance(event.extra_fields, _ExtraFields_TorchOp)
if event.name() == "aten::mm":
if event.extra_fields.allow_tf32_cublas is False:
return True
return False
def match(self, event: _ProfilerEvent):
if not event.name().endswith(": zero_grad"):
return False
if not event.children:
return False
for sub_event in eventTreeDFS(event.children):
if sub_event.name(
) == "aten::zero_" and sub_event.parent.name() != "aten::zeros":
return True
# TODO: We should also check if the optimizer's numerical behavior will change.
return False
def match(self, event: _ProfilerEvent):
def mutiple_of(shapes, multiple):
return all(dim % multiple == 0 for shape in shapes
for dim in shape[-2:])
if event.name() not in ("aten::mm", "aten::bmm", "aten::addmm"):
return False
if not input_dtypes(event):
return False
arg_dtype = input_dtypes(event)[0]
# TODO: Have a better way to check dtype_size
if (arg_dtype.endswith("c10::BFloat16")
or arg_dtype.endswith("c10::Half")) and not mutiple_of(
input_shapes(event), 8):
return True
return False
def match(self, event: _ProfilerEvent):
for optimizer in self.optimizers_with_foreach:
if event.name().endswith(f"_single_tensor_{optimizer}"):
return True
return False
def match(self, event: _ProfilerEvent):
return re.search(self.name, event.name()) is not None
