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
|
# Owner(s): ["module: inductor"]
import functools
import unittest
import torch
from torch._dynamo import config as dynamo_config
from torch._inductor import config as inductor_config
from torch._inductor.test_case import TestCase as InductorTestCase
from torch._inductor.utils import is_big_gpu
from torch.testing import make_tensor
from torch.testing._internal.common_device_type import (
instantiate_device_type_tests,
skipGPUIf,
)
from torch.testing._internal.common_utils import IS_LINUX, parametrize
from torch.testing._internal.inductor_utils import (
GPU_TYPE,
HAS_CUDA,
HAS_GPU,
requires_gpu,
)
class TestUnbackedSymints(InductorTestCase):
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_expand(self, device):
def fn(x, y):
nz = torch.nonzero(x)
# unbacked symint in nz.size
x_exp = nz.expand([-1, 128])
# unbacked symint in target sizes
y_exp = y.expand([-1, nz.size(0)])
return x_exp, y_exp
example_inputs = (
torch.randn((32), device=device),
torch.randn((32, 1), device=device),
)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_expand_ok_with_runtime_assert(self, device):
def fn(x):
nz = x.nonzero()
torch._check(nz.size(0) == 128)
return nz.expand([128, -1, 2])
x = make_tensor(32, 4, device=device, dtype=torch.float32, exclude_zero=True)
actual = torch.compile(fn, fullgraph=True)(x)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_broadcast_tensors(self, device):
def fn(x):
nz = x.nonzero()
a = torch.zeros([nz.size(0), 512])
b = torch.ones([nz.size(0), 1])
return a * b
x = torch.randn(32, 4, device=device)
actual = torch.compile(fn, fullgraph=True)(x)
expected = fn(x)
torch.testing.assert_close(actual, expected)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_autotuning(self, device):
def fn(x, y):
nz = torch.nonzero(x)
# unbacked symint in the GEMM input shape
a = x.new_ones([nz.size(0), y.size(0)])
return a @ y
example_inputs = (
torch.randn((64), device=device),
torch.randn((32, 16), device=device),
)
with inductor_config.patch(
{
"max_autotune_gemm": True,
}
):
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_scalar_outputs": True})
def test_split_with_sizes(self, device):
def fn(x, y):
l = y.tolist()
s = torch.split(x, l)
d = l[0] + l[1] + l[2]
return s[0].sum(), d
example_inputs = (torch.randn((32), device=device), torch.tensor((7, 16, 9)))
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_view_of_slice(self, device):
# Tests View.create(slice, size_with_unbacked_symint)
def fn(x):
nz = torch.nonzero(x) # introduce unbacked symint
squared = nz * nz # avoid ReinterpretView when lowering Slice
sliced = torch.ops.aten.slice.Tensor(squared, dim=1, start=-2, end=None)
view = sliced.unsqueeze(dim=0)
return view.squeeze(
dim=0
) # make sure no unbacked symint in output's stride
example_inputs = (torch.randn(1, 1, 1, 1, device=device),)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@requires_gpu()
@dynamo_config.patch({"capture_scalar_outputs": True})
def test_triton_kernel_grid(self, device):
if device == "cpu":
raise unittest.SkipTest("Triton kernel requires GPU")
from torch.testing._internal.triton_utils import add_kernel
def fn(x):
maxlen = max(x.item(), 512)
a = torch.ones(maxlen, device=device)
b = torch.ones(maxlen, device=device)
out = torch.zeros_like(a)
# unbacked symint in grid
add_kernel[(1, 1, maxlen)](a, b, out, maxlen, 32)
return out
example_inputs = (torch.randint(high=1024, size=(1,), device=device),)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@skipGPUIf(not HAS_GPU, "requires gpu and triton")
@dynamo_config.patch({"capture_dynamic_output_shape_ops": True})
def test_nonzero_in_inference_mode(self, device):
def fn(x):
return torch.nonzero(x)
example_inputs = (torch.randint(0, 2, (128,), device=device),)
with torch.inference_mode():
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@inductor_config.patch({"max_autotune": True})
@dynamo_config.patch({"capture_scalar_outputs": True})
def test_equivalent_backed_unbacked(self, device):
# Tests scenario when there are two equivalent backed & unbacked symints,
# but when we look-up a size hint on the unbacked symint, we ignorantly
# use the default fallback hint.
def fn(x, w, a, b):
# Make tensors where 1st dim is unbacked/backed.
u0, s0 = a.item(), b.size(0)
unbacked = x.expand(u0, *x.shape)
backed = x.expand(s0, *x.shape)
# The cat unifies u0 and s0 -- i.e. u0 == s0.
cat = torch.cat([backed, unbacked, unbacked], dim=1) # [s0, 30, 16]
mat1 = torch.permute(cat, [0, 2, 1]) # [s0, 16, 30]
mat2 = w.expand(u0, *w.shape) # [u0, 30, 32]
bmm = torch.ops.aten.bmm(mat1, mat2)
return bmm
example_inputs = (
torch.randn((10, 16), dtype=torch.float32, device=device),
torch.randn((30, 32), dtype=torch.float32, device=device),
torch.tensor(7, device=device),
backed := torch.randn((7,), device=device),
)
torch._dynamo.mark_dynamic(backed, 0) # create backed symint
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@requires_gpu()
@dynamo_config.patch({"capture_scalar_outputs": True})
def test_vertical_pointwise_reduction_fusion(self, device):
# reset in case we run both cpu and cuda tests
torch._inductor.metrics.reset()
# Tests fusing a pointwise & reduction op with unbacked numel/rnumel.
def fn(x, y, repeats):
u0 = repeats.item()
unbacked = y.expand(u0, *y.shape) # [u0, 1, 16]
# Note: We add x to both pointwise and reduction. Otherwise, the
# scheduler will refuse to fuse ops whose only common buffer has
# unbacked symints.
pointwise = unbacked + x
reduction = torch.sum(pointwise + x)
return pointwise, reduction
example_inputs = (
torch.randn(32, 16).to(GPU_TYPE),
torch.randn(1, 16).to(GPU_TYPE),
torch.tensor(32).to(GPU_TYPE),
)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
self.assertEqual(torch._inductor.metrics.generated_kernel_count, 2)
@dynamo_config.patch({"capture_scalar_outputs": True})
@parametrize(
"torch_fn", [torch.mm, torch.bmm, torch.addmm], name_fn=lambda fn: fn.__name__
)
@parametrize("coordinate_descent_tuning", [True, False], name_fn=str)
def test_mm_and_friends(self, device, torch_fn, coordinate_descent_tuning):
if torch_fn == torch.addmm:
torch_fn = functools.partial(torch_fn, torch.ones(1, device=device))
def fn(x, w, repeats, is_bmm):
u0 = repeats.item()
torch._check_is_size(u0)
x_unbacked = x.expand(u0, 32)
w_unbacked = w.expand(32, u0)
if is_bmm:
# Make sure inputs are batched.
x_unbacked = x_unbacked.expand(10, *x_unbacked.shape)
w_unbacked = w_unbacked.expand(10, *w_unbacked.shape)
return torch_fn(x_unbacked, w_unbacked)
example_inputs = (
torch.randn(1, 32, device=device),
torch.randn(32, 1, device=device),
torch.tensor(100, device=device),
torch_fn == torch.bmm,
)
with inductor_config.patch(
{
# coordinate_descent_tuning has its own path during decomp
"coordinate_descent_tuning": coordinate_descent_tuning,
}
):
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@torch._dynamo.config.patch(capture_scalar_outputs=True)
def test_unbacked_range_tree_divisor(self, device):
def fn(x, num):
u0 = num.item()
torch._check_is_size(u0)
zeros = torch.zeros(u0, device=device, dtype=torch.int)
return (torch.ops.aten.index(x, [None, zeros]),)
example_inputs = (
torch.randn(16, 16, device=device),
torch.tensor(3, device=device),
)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
@dynamo_config.patch({"capture_scalar_outputs": True})
def test_unbacked_masked_scatter(self, device):
def fn(value, mask):
u0 = mask.count_nonzero()
source = torch.ones(u0, dtype=torch.float32, device=device)
return torch.masked_scatter(value, mask, source)
value = make_tensor(10, 10, dtype=torch.float32, device=device)
mask = make_tensor(10, 10, dtype=torch.bool, device=device)
example_inputs = (value, mask)
actual = torch.compile(fn, fullgraph=True)(*example_inputs)
expected = fn(*example_inputs)
torch.testing.assert_close(actual, expected)
instantiate_device_type_tests(TestUnbackedSymints, globals(), allow_xpu=True)
if __name__ == "__main__":
from torch._inductor.test_case import run_tests
if IS_LINUX and HAS_GPU and (not HAS_CUDA or is_big_gpu()):
run_tests()
|
