mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
3ef7a6921d
Adds a "reinplacing" FX transform, that goes through an FX graph and tries to convert out-of-place op calls into inplace calls whenever possible.
Followups from this PR include:
- Set up torch bench, and run the whole torchbench suite using AOTAutograd + functionalize + rein placing transforms to surface any issues (this is what I'm currently working on). Right now, I have some basic unit tests just to sanity check that the general logic makes sense.
- Add any missing inplace ops. This is mostly the `*_scatter*` ops, e.g. `diagonal_scatter_`, because these ops will commonly show up an FX graph after running functionalization.
The criteria for when you can swap an op `b = a.add(...)` with `a.add_(...)` is:
(1) An inplace variant of the operator with the same schema needs to exist (`aten.add` -> `aten.add_`)
(2) `a` (**or any of its aliases**) can't be used as an input to any other operators later on in the graph
(3) `a` can't be one of the inputs to the entire graph. It also can't be an **alias** of any of the inputs ***
*** One thing to note: (3) means that we can't technically guarantee that we'll get back **all** memory usage that we lost from functionalization. Functionalization converts input mutations into out-of-place calls, and then adds a `copy_()` to the end of the graph to preserve semantics.
I added logic to handle `copy_()` in this PR because it it's a pretty important optimizations in the context of `functionalization()`: any program that performs input mutations will have a `copy_()` in it after running functionalization.
There are some examples in the test file, but I think staring at an example of where re-inplacing is/isn't allowed to run is helpful:
```
// Before functionalization
def foo(a):
tmp1 = a.add_(1)
tmp2 = a.add(2)
// After functionalization
def foo(a)
tmp1 = a.add(1)
tmp2 = a.add(2)
....
a.copy_(tmp1)
// After re-inplacing
def foo(a)
// first add() is safe to re-inplace even though a is a program input,
// because a's data is overwritten later by a copy_()
tmp1 = a.add_(1)
// second add() is NOT safe to re-inplace, because:
// (1) a and tmp1 are aliased. Note that they weren't aliased in the original program,
but they are now that we've done some re-inplacing.
// (2) tmp1 is used as an input later in the program
tmp2 = a.add(2)
....
a.copy_(tmp1)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80897
Approved by: https://github.com/ezyang
252 lines
10 KiB
Python
252 lines
10 KiB
Python
# Owner(s): ["module: functionalization"]
|
|
import torch
|
|
from torch.testing._internal.common_utils import TestCase, run_tests
|
|
from torch.fx.passes.reinplace import reinplace
|
|
from torch.fx.experimental.proxy_tensor import make_fx
|
|
|
|
try:
|
|
from functorch.experimental import functionalize
|
|
HAS_FUNCTIONALIZATION = True
|
|
except Exception as e:
|
|
HAS_FUNCTIONALIZATION = False
|
|
|
|
class TestReinplacePass(TestCase):
|
|
|
|
def test_reinplace_basic(self):
|
|
# Basic test: the out-of-place add() call should be converted
|
|
# into add_()
|
|
def f(x):
|
|
a = x.clone()
|
|
b = a.add(1)
|
|
return b
|
|
|
|
inpt = torch.ones(2)
|
|
f2 = reinplace(make_fx(f)(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, x_1):
|
|
clone_default = torch.ops.aten.clone.default(x_1); x_1 = None
|
|
add_tensor = torch.ops.aten.add_.Tensor(clone_default, 1)
|
|
return clone_default
|
|
""")
|
|
|
|
|
|
def test_reinplace_with_view(self):
|
|
def f(x):
|
|
a = x.clone()
|
|
a_view = a.view(-1)
|
|
# We shouldn't re-inplace the first add(), because an alias of a is re-used later in the program
|
|
b = a.add(1)
|
|
# Second add() is fine to re-inplace
|
|
c = a_view.add(1)
|
|
return c
|
|
|
|
inpt = torch.ones(2)
|
|
f2 = reinplace(make_fx(f)(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, x_1):
|
|
clone_default = torch.ops.aten.clone.default(x_1); x_1 = None
|
|
view_default = torch.ops.aten.view.default(clone_default, [-1])
|
|
add_tensor = torch.ops.aten.add.Tensor(clone_default, 1); clone_default = None
|
|
add_tensor_1 = torch.ops.aten.add_.Tensor(view_default, 1)
|
|
return view_default
|
|
""")
|
|
|
|
# This test won't actually run in CI, because it requires functionalize() from functorch.
|
|
# I'm planning on testing more comprehensively with torchbench models,
|
|
# but we can make this testing better once functorch moves into pytorch/pytorch.
|
|
def test_reinplace_scatter_op(self):
|
|
def f(a_):
|
|
# for now, don't test mutations to inputs
|
|
a = a_.clone()
|
|
e = a.view(-1)
|
|
b = a.view(-1)
|
|
c = b[0]
|
|
d = c.view(-1)
|
|
d.add_(1)
|
|
return a + e
|
|
|
|
if not HAS_FUNCTIONALIZATION:
|
|
return
|
|
inpt = torch.ones(4)
|
|
f2 = reinplace(make_fx(functionalize(f))(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
# NOTE: one slight pessimization here is the fact that
|
|
# there are a bunch of redundant views in the graph.
|
|
# Technically, half of these views are duplicates that we could de-dup.
|
|
# This shouldn't really hurt performance though, since creating an extra view
|
|
# is effectively just moving some metadata around (and allocating a new TensorImpl).
|
|
# We can/should update the pass in the future to clean this up.
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, a__1):
|
|
clone_default = torch.ops.aten.clone.default(a__1); a__1 = None
|
|
view_default = torch.ops.aten.view.default(clone_default, [-1])
|
|
view_default_1 = torch.ops.aten.view.default(clone_default, [-1])
|
|
select_int = torch.ops.aten.select.int(view_default_1, 0, 0); view_default_1 = None
|
|
view_default_2 = torch.ops.aten.view.default(select_int, [-1]); select_int = None
|
|
add_tensor = torch.ops.aten.add_.Tensor(view_default_2, 1)
|
|
view_default_3 = torch.ops.aten.view.default(clone_default, [-1]); clone_default = None
|
|
select_int_1 = torch.ops.aten.select.int(view_default_3, 0, 0)
|
|
view_default_4 = torch.ops.aten.view.default(view_default_2, []); view_default_2 = None
|
|
view_default_5 = torch.ops.aten.view.default(view_default_3, [4]); view_default_3 = None
|
|
view_default_6 = torch.ops.aten.view.default(view_default_5, [-1])
|
|
add_tensor_1 = torch.ops.aten.add_.Tensor(view_default_5, view_default_6); view_default_6 = None
|
|
return view_default_5
|
|
""")
|
|
|
|
def test_reinplace_scatter_twice(self):
|
|
def f(a_):
|
|
# for now, don't test mutations to inputs
|
|
a = a_.clone()
|
|
b = a[:, 1]
|
|
c = b[1]
|
|
c.add_(1)
|
|
return a
|
|
|
|
if not HAS_FUNCTIONALIZATION:
|
|
return
|
|
|
|
inpt = torch.ones(4, 4)
|
|
f2 = reinplace(make_fx(functionalize(f))(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, a__1):
|
|
clone_default = torch.ops.aten.clone.default(a__1); a__1 = None
|
|
slice_tensor = torch.ops.aten.slice.Tensor(clone_default, 0, 0, 9223372036854775807)
|
|
select_int = torch.ops.aten.select.int(slice_tensor, 1, 1); slice_tensor = None
|
|
select_int_1 = torch.ops.aten.select.int(select_int, 0, 1); select_int = None
|
|
add_tensor = torch.ops.aten.add_.Tensor(select_int_1, 1); select_int_1 = None
|
|
slice_tensor_1 = torch.ops.aten.slice.Tensor(clone_default, 0, 0, 9223372036854775807)
|
|
select_int_2 = torch.ops.aten.select.int(slice_tensor_1, 1, 1); slice_tensor_1 = None
|
|
return clone_default
|
|
""")
|
|
|
|
def test_reinplace_scatter_twice_with_different_view_op_valid(self):
|
|
def f(a_):
|
|
a = a_.clone()
|
|
b = a[:, 1]
|
|
c = b[1]
|
|
c_updated = c.add(1)
|
|
good_mirror_of_b = a.as_strided((4,), (4,), 1)
|
|
# good_mirror_of_b points to the same region of memory as b.
|
|
# and this scatter op below tries to scatter c_updated into the same region
|
|
# that c currently takes up.
|
|
# reinplacing logic checks this by confirming that:
|
|
# c_updated
|
|
# good_mirror_of_b.select(0, 1)
|
|
# have the same size/stride/storage_offset.
|
|
b_updated = torch.select_scatter(good_mirror_of_b, c_updated, 0, 1)
|
|
return b_updated
|
|
|
|
inpt = torch.ones(4, 4)
|
|
f2 = reinplace(make_fx(f)(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, a__1):
|
|
clone_default = torch.ops.aten.clone.default(a__1); a__1 = None
|
|
slice_tensor = torch.ops.aten.slice.Tensor(clone_default, 0, 0, 9223372036854775807)
|
|
select_int = torch.ops.aten.select.int(slice_tensor, 1, 1); slice_tensor = None
|
|
select_int_1 = torch.ops.aten.select.int(select_int, 0, 1); select_int = None
|
|
add_tensor = torch.ops.aten.add_.Tensor(select_int_1, 1); select_int_1 = None
|
|
as_strided_default = torch.ops.aten.as_strided.default(clone_default, [4], [4], 1); clone_default = None
|
|
return as_strided_default
|
|
""")
|
|
|
|
# Test example where we have a scatter op, where the base tensor
|
|
# has the same size/stride/storage offset (even though it is a different view),
|
|
# making it valid to re-inplace
|
|
def test_reinplace_scatter_twice_with_different_view_op_invalid(self):
|
|
def f(a_):
|
|
a = a_.clone()
|
|
b = a[:, 1]
|
|
c = b[1]
|
|
c_updated = c.add(1)
|
|
good_mirror_of_b = a.as_strided((4,), (4,), 1)
|
|
# The first arg to select_scatter is an equivalent view to b.
|
|
# However, the select_scatter call below tries to put c_updated
|
|
# into a different slice of "b" than what "c" currently occupies.
|
|
#
|
|
b_updated = torch.select_scatter(good_mirror_of_b, c_updated, 0, 0)
|
|
return b_updated
|
|
|
|
inpt = torch.ones(4, 4)
|
|
f2 = reinplace(make_fx(f)(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, a__1):
|
|
clone_default = torch.ops.aten.clone.default(a__1); a__1 = None
|
|
slice_tensor = torch.ops.aten.slice.Tensor(clone_default, 0, 0, 9223372036854775807)
|
|
select_int = torch.ops.aten.select.int(slice_tensor, 1, 1); slice_tensor = None
|
|
select_int_1 = torch.ops.aten.select.int(select_int, 0, 1); select_int = None
|
|
add_tensor = torch.ops.aten.add.Tensor(select_int_1, 1); select_int_1 = None
|
|
as_strided_default = torch.ops.aten.as_strided.default(clone_default, [4], [4], 1); clone_default = None
|
|
select_scatter_default = torch.ops.aten.select_scatter.default(as_strided_default, add_tensor, 0, 0); as_strided_default = add_tensor = None
|
|
return select_scatter_default
|
|
""") # noqa: B950
|
|
|
|
def test_reinplace_scatter_twice_with_different_view_op_invalid2(self):
|
|
def f(a_):
|
|
a = a_.clone()
|
|
b = a[:, 1]
|
|
c = b[1]
|
|
c_updated = c.add(1)
|
|
bad_mirror_of_b = a.as_strided((4,), (4,), 0)
|
|
# The first arg to select_scatter points to a different than c's base.
|
|
# This makes it invalid to re-inplace.
|
|
b_updated = torch.select_scatter(bad_mirror_of_b, c_updated, 0, 1)
|
|
return b_updated
|
|
|
|
inpt = torch.ones(4, 4)
|
|
f2 = reinplace(make_fx(f)(inpt), inpt)
|
|
expected_out = f(inpt)
|
|
actual_out = f2(inpt)
|
|
# self.assertEqual(actual_out, expected_out)
|
|
self.assertExpectedInline(f2.code, """\
|
|
|
|
|
|
|
|
def forward(self, a__1):
|
|
clone_default = torch.ops.aten.clone.default(a__1); a__1 = None
|
|
slice_tensor = torch.ops.aten.slice.Tensor(clone_default, 0, 0, 9223372036854775807)
|
|
select_int = torch.ops.aten.select.int(slice_tensor, 1, 1); slice_tensor = None
|
|
select_int_1 = torch.ops.aten.select.int(select_int, 0, 1); select_int = None
|
|
add_tensor = torch.ops.aten.add.Tensor(select_int_1, 1); select_int_1 = None
|
|
as_strided_default = torch.ops.aten.as_strided.default(clone_default, [4], [4], 0); clone_default = None
|
|
select_scatter_default = torch.ops.aten.select_scatter.default(as_strided_default, add_tensor, 0, 1); as_strided_default = add_tensor = None
|
|
return select_scatter_default
|
|
""") # noqa: B950
|
|
|
|
if __name__ == '__main__':
|
|
run_tests()
|