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
356
357
358
359
360
361
362
363
364
365
|
# Owner(s): ["module: inductor"]
import functools
from unittest.mock import patch
import torch
import torch._dynamo.config as dynamo_config
import torch._inductor.config as inductor_config
import torch._inductor.select_algorithm as select_algorithm
import torch.nn.functional as F
from torch._dynamo.testing import expectedFailureDynamicWrapper
from torch._dynamo.utils import counters
from torch._inductor.autotune_process import TritonBenchmarkRequest
from torch._inductor.test_case import run_tests, TestCase
from torch._inductor.utils import is_big_gpu
from torch.testing._internal.common_utils import IS_LINUX, skipIfRocm
from torch.testing._internal.inductor_utils import HAS_CUDA
aten = torch.ops.aten
def patches(fn):
def skip_cache(self, choices, name, key, benchmark):
if benchmark is None:
return {}
return benchmark(choices)
for patcher in [
dynamo_config.patch(verbose=True),
inductor_config.patch(debug=True, max_autotune=True, epilogue_fusion=True),
patch.object(select_algorithm, "VERIFY", dict(atol=1e-4, rtol=1e-4)),
patch.object(select_algorithm.AlgorithmSelectorCache, "lookup", skip_cache),
torch.backends.cudnn.flags(allow_tf32=False),
]:
fn = patcher(fn)
@functools.wraps(fn)
def wrapped(*args, **kwargs):
counters.clear()
torch.manual_seed(12345)
assert (
not torch.backends.cuda.matmul.allow_tf32
), "correctness testing is allergic to tf32"
return fn(*args, **kwargs)
return wrapped
class TestSelectAlgorithm(TestCase):
def setUp(self):
super().setUp()
if not is_big_gpu():
return self.skipTest("Need a big GPU to run max_autotune=True")
@patches
def test_linear_relu_cuda(self):
@torch.compile
def foo(input, weight, bias):
return F.relu(F.linear(input, weight, bias))
foo(
torch.randn(64, 32, device="cuda"),
torch.randn(16, 32, device="cuda"),
torch.randn(1, 16, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
# It would be nice to assert this got fused into a single kernel, but that
# only happens if we select a triton template (and not aten).
@patches
def test_addmm_cuda(self):
@torch.compile
def foo(input, weight, bias):
return torch.addmm(bias, input, weight)
inps = (
torch.randn(20, 33, device="cuda"),
torch.randn(33, 16, device="cuda"),
torch.randn(20, 16, device="cuda"),
)
foo(*inps)
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patch.object(select_algorithm, "VERIFY", dict(atol=5e-2, rtol=5e-2))
@patches
def test_addmm_fp16(self):
@torch.compile
def foo(input, weight, bias):
return torch.addmm(bias, input, weight)
inps = (
torch.randn(2, 320, device="cuda", dtype=torch.half),
torch.randn(320, 320, device="cuda", dtype=torch.half).t(),
torch.empty(320, device="cuda", dtype=torch.half),
)
foo(*inps)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm(self):
@torch.compile
def foo(a, b):
return torch.mm(a, b)
foo(
torch.randn(8, 32, device="cuda"),
torch.randn(32, 8, device="cuda"),
)
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
# FIXME: Investigate why _int_mm_out_cuda is not compiled on ROCm
@skipIfRocm
@patches
def test__int_mm(self):
@torch.compile
def foo(a, b):
return torch._int_mm(a, b)
foo(
torch.randint(-10, 10, (64, 32), device="cuda", dtype=torch.int8),
torch.randint(-10, 10, (32, 64), device="cuda", dtype=torch.int8),
)
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_skip(self):
@torch.compile
def foo(a, b):
return torch.mm(a, b)
foo(
torch.randn(8, 32, device="cuda", dtype=torch.float64),
torch.randn(32, 8, device="cuda", dtype=torch.float64),
)
# float64 not supported by tl.dot()
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 0)
@patches
def test_bmm(self):
@torch.compile
def foo(a, b):
return torch.bmm(a, b)
foo(
torch.randn(2, 8, 32, device="cuda"),
torch.randn(2, 32, 8, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_not_even_k(self):
@torch.compile
def foo(a, b):
return torch.mm(a, b)
foo(
torch.randn(11, 22, device="cuda"),
torch.randn(22, 33, device="cuda"),
)
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_baddbmm(self):
@torch.compile
def foo(a, b, c):
return torch.baddbmm(c, a, b)
foo(
torch.randn(2, 8, 32, device="cuda"),
torch.randn(2, 32, 8, device="cuda"),
torch.randn(2, 1, 8, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_plus_mm(self):
@torch.compile
def foo(a, b, c, d):
return (a @ b) + (c @ d)
foo(
torch.randn(32, 32, device="cuda"),
torch.randn(32, 32, device="cuda"),
torch.randn(32, 32, device="cuda"),
torch.randn(32, 32, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_plus_mm2_cuda(self):
@torch.compile
def foo(a, b, c, d):
return (a @ b) + (c @ d)
foo(
torch.randn(512, 512, device="cuda"),
torch.randn(512, 512, device="cuda"),
torch.randn(512, 512, device="cuda"),
torch.randn(512, 512, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@expectedFailureDynamicWrapper
@patches
def test_mm_plus_mm3_cuda(self):
@torch.compile
def foo(a, b, c, d):
return (a @ b) + (c @ d)
foo(
torch.randn(512, 32, device="cuda"),
torch.randn(32, 8, device="cuda"),
torch.randn(512, 32, device="cuda"),
torch.randn(32, 8, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_dup_args(self):
@torch.compile
def foo(a):
return torch.mm(a, a)
foo(torch.randn(32, 32, device="cuda"))
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
def test_mm_dup_args_view(self):
@torch.compile
def foo(a):
q = a[:32, :]
k = a[32:, :]
return torch.mm(q, k.transpose(0, 1))
foo(torch.randn(64, 64, device="cuda"))
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@expectedFailureDynamicWrapper
@patches
def test_convolution1(self):
@torch.compile
def foo(x, w, b):
return aten.convolution(
x + 1,
w,
b,
stride=(2, 3),
padding=(4, 5),
dilation=(1, 1),
transposed=False,
output_padding=(0, 0),
groups=1,
)
foo(
torch.randn(2, 33, 34, 41, device="cuda"),
torch.randn(34, 33, 3, 3, device="cuda"),
torch.randn(34, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@skipIfRocm
@patches
def test_mm_dropout(self):
@torch.compile
def fn(x1, x2, seed):
mm_4 = torch.ops.aten.mm.default(x2, x1)
rnd = torch.ops.prims.inductor_random.default(mm_4.shape, seed, "rand")
return mm_4 * rnd
# sizes picked so triton autotuning wins
fn(
torch.randn(512, 1024, dtype=torch.float16, device="cuda"),
torch.randn(384, 512, dtype=torch.float16, device="cuda"),
torch.tensor(12345, device="cuda"),
)
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@skipIfRocm
@patches
@torch._inductor.config.patch(conv_1x1_as_mm=False)
def test_convolution2(self):
@torch.compile
def foo(x, w, b):
return aten.convolution(
x,
w,
b,
stride=(1, 1),
padding=(0, 0),
dilation=(1, 1),
transposed=False,
output_padding=(0, 0),
groups=1,
)
foo(
torch.randn(1, 33, 16, 16, device="cuda"),
torch.randn(34, 33, 1, 1, device="cuda"),
torch.randn(34, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
@patches
@torch._inductor.config.patch(conv_1x1_as_mm=True)
def test_convolution_as_mm(self):
@torch.compile
def foo(x, w, b):
return aten.convolution(
x + 1,
w,
b,
stride=(1, 1),
padding=(0, 0),
dilation=(1, 1),
transposed=False,
output_padding=(0, 0),
groups=1,
)
foo(
torch.randn(2, 33, 16, 16, device="cuda"),
torch.randn(34, 33, 1, 1, device="cuda"),
torch.randn(34, device="cuda"),
)
# Autotuning checks correctness of each version
self.assertEqual(counters["inductor"]["select_algorithm_autotune"], 1)
def test_TritonTemplateCaller_str(self):
"""
Make sure str(TritonTemplateCaller) does not raise exceptions.
"""
module_path = "abc.py"
bmreq = TritonBenchmarkRequest(
module_path=module_path,
module_cache_key=None,
kernel_name=None,
grid=None,
extra_args=None,
num_stages=None,
num_warps=None,
input_tensor_meta=None,
output_tensor_meta=None,
)
caller = select_algorithm.TritonTemplateCaller(
None, None, None, None, "extra", bmreq
)
caller_str = str(caller)
self.assertEqual(caller_str, f"TritonTemplateCaller({module_path}, extra)")
if __name__ == "__main__":
if IS_LINUX and HAS_CUDA and is_big_gpu():
run_tests()
|
