1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
|
# 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 = torch.ops.aten.clone.default(x_1); x_1 = None
add = torch.ops.aten.add_.Tensor(clone, 1)
return clone
""")
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 = torch.ops.aten.clone.default(x_1); x_1 = None
view = torch.ops.aten.view.default(clone, [-1])
add = torch.ops.aten.add.Tensor(clone, 1); clone = None
add_1 = torch.ops.aten.add_.Tensor(view, 1)
return view
""")
def test_reinplace_different_metadata(self):
def f(a_):
a = a_.clone()
b = a + 1
# Naively, we shouldn't try to inplace the .ge() call,
# because that would require resizing "b" (from a float to a bool tensor).
c = torch.ge(b, a)
return c
inpt = torch.ones(4)
f2 = reinplace(make_fx(f)(inpt), inpt)
expected_out = f(inpt)
actual_out = f2(inpt)
self.assertEqual(actual_out, expected_out)
# The .ge() should not be reinplaced.
self.assertExpectedInline(f2.code, """\
def forward(self, a__1):
clone = torch.ops.aten.clone.default(a__1); a__1 = None
add = torch.ops.aten.add.Tensor(clone, 1)
ge = torch.ops.aten.ge.Tensor(add, clone); add = clone = None
return ge
""")
def test_reinplace_overlapping_memory(self):
def f(a_):
a = a_.clone()
b = a.expand(4, 4)
# Can't reinplace because b has overlapping memory.
c = b.add(1)
return c
inpt = torch.ones(1)
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 = torch.ops.aten.clone.default(a__1); a__1 = None
expand = torch.ops.aten.expand.default(clone, [4, 4]); clone = None
add = torch.ops.aten.add.Tensor(expand, 1); expand = None
return add
""")
# 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 = torch.ops.aten.clone.default(a__1); a__1 = None
view = torch.ops.aten.view.default(clone, [-1])
view_1 = torch.ops.aten.view.default(clone, [-1])
select = torch.ops.aten.select.int(view_1, 0, 0); view_1 = None
view_2 = torch.ops.aten.view.default(select, [-1]); select = None
add = torch.ops.aten.add_.Tensor(view_2, 1)
view_3 = torch.ops.aten.view.default(clone, [-1]); clone = None
select_1 = torch.ops.aten.select.int(view_3, 0, 0)
view_4 = torch.ops.aten.view.default(view_2, []); view_2 = None
view_5 = torch.ops.aten.view.default(view_3, [4]); view_3 = None
view_6 = torch.ops.aten.view.default(view_5, [-1])
add_1 = torch.ops.aten.add_.Tensor(view_5, view_6); view_6 = None
return view_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 = torch.ops.aten.clone.default(a__1); a__1 = None
slice_1 = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
select = torch.ops.aten.select.int(slice_1, 1, 1); slice_1 = None
select_1 = torch.ops.aten.select.int(select, 0, 1); select = None
add = torch.ops.aten.add_.Tensor(select_1, 1); select_1 = None
slice_2 = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
select_2 = torch.ops.aten.select.int(slice_2, 1, 1); slice_2 = None
return clone
""")
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 = torch.ops.aten.clone.default(a__1); a__1 = None
slice_1 = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
select = torch.ops.aten.select.int(slice_1, 1, 1); slice_1 = None
select_1 = torch.ops.aten.select.int(select, 0, 1); select = None
add = torch.ops.aten.add_.Tensor(select_1, 1); select_1 = None
as_strided = torch.ops.aten.as_strided.default(clone, [4], [4], 1); clone = None
return as_strided
""")
# 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 = torch.ops.aten.clone.default(a__1); a__1 = None
slice_1 = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
select = torch.ops.aten.select.int(slice_1, 1, 1); slice_1 = None
select_1 = torch.ops.aten.select.int(select, 0, 1); select = None
add = torch.ops.aten.add.Tensor(select_1, 1); select_1 = None
as_strided = torch.ops.aten.as_strided.default(clone, [4], [4], 1); clone = None
select_int = torch.ops.aten.select.int(as_strided, 0, 0)
copy__default = torch.ops.aten.copy_.default(select_int, add); select_int = add = None
return as_strided
""") # 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 = torch.ops.aten.clone.default(a__1); a__1 = None
slice_1 = torch.ops.aten.slice.Tensor(clone, 0, 0, 9223372036854775807)
select = torch.ops.aten.select.int(slice_1, 1, 1); slice_1 = None
select_1 = torch.ops.aten.select.int(select, 0, 1); select = None
add = torch.ops.aten.add.Tensor(select_1, 1); select_1 = None
as_strided = torch.ops.aten.as_strided.default(clone, [4], [4], 0); clone = None
select_int = torch.ops.aten.select.int(as_strided, 0, 1)
copy__default = torch.ops.aten.copy_.default(select_int, add); select_int = add = None
return as_strided
""") # noqa: B950
def test_out_node_updated(self):
def f():
x = torch.zeros(2, 2)
y = x.diagonal()
y_updated = y.add(1)
z = torch.diagonal_scatter(x, y_updated)
# reinplace needs to know to replace output [z] with [x]
return [z]
if not HAS_FUNCTIONALIZATION:
return
f2 = reinplace(make_fx(functionalize(f))())
expected_out = f()
actual_out = f2()
self.assertEqual(actual_out, expected_out)
self.assertExpectedInline(f2.code, """\
def forward(self):
zeros = torch.ops.aten.zeros.default([2, 2], device = device(type='cpu'), pin_memory = False)
diagonal = torch.ops.aten.diagonal.default(zeros)
add = torch.ops.aten.add_.Tensor(diagonal, 1); diagonal = None
return [zeros]
""")
def test_reinplace_index_mutation(self):
def f():
a = torch.zeros(4, 4, 4)
a[:, 2:] = torch.ones(4, 2, 4)
return a
if not HAS_FUNCTIONALIZATION:
return
f2 = reinplace(make_fx(functionalize(f))())
expected_out = f()
actual_out = f2()
self.assertEqual(actual_out, expected_out)
self.assertExpectedInline(f2.code, """\
def forward(self):
zeros = torch.ops.aten.zeros.default([4, 4, 4], device = device(type='cpu'), pin_memory = False)
ones = torch.ops.aten.ones.default([4, 2, 4], device = device(type='cpu'), pin_memory = False)
slice_1 = torch.ops.aten.slice.Tensor(zeros, 0, 0, 9223372036854775807)
slice_2 = torch.ops.aten.slice.Tensor(slice_1, 1, 2, 9223372036854775807); slice_1 = None
slice_3 = torch.ops.aten.slice.Tensor(zeros, 0, 0, 9223372036854775807)
slice_tensor = torch.ops.aten.slice.Tensor(slice_3, 1, 2, 9223372036854775807); slice_3 = None
copy__default = torch.ops.aten.copy_.default(slice_tensor, ones); slice_tensor = ones = None
return zeros
""")
if __name__ == '__main__':
run_tests()
|
