Skip to content

Commit dd89d28

Browse files
zonglinpengfacebook-github-bot
zonglinpeng
authored and
facebook-github-bot
committed
migrate all test_aten_ops to facto
Summary: **Have to ship this as an intermediate step to unblock 3 workstreams on the stack** Modify aten tests to ingress FACTO generated testcases. - Each test gets 30~50 cases with good coverage on - Optimized VS unoptimized flows - dtype switch cases Known issues: - FACTO test class is too big to run on default "heavyweight" CI - current skipping the whole target on CI. Will add back once skycastle flow is ready - some FACTO is creating inputs that kernels does not handle, mainly dtypes - will create exception handling for that. - TODO marks the 2 FACTO doesnt work well on the 2 ops. Differential Revision: D79121474
1 parent 56f24c6 commit dd89d28

File tree

1 file changed

+111
-16
lines changed

1 file changed

+111
-16
lines changed

‎backends/cadence/utils/facto_util.py‎

Lines changed: 111 additions & 16 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,7 @@
99
import copy
1010
from functools import lru_cache
1111
from typing import List, OrderedDict, Tuple
12+
import facto.specdb.function as fn
1213

1314
import torch
1415
from facto.inputgen.argtuple.gen import ArgumentTupleGenerator
@@ -19,24 +20,22 @@
1920
# seed to generate identical cases every run to reproduce from bisect
2021
MAX_CASES = 50
2122

22-
2323
def apply_tensor_contraints(op_name: str, index: int) -> list[object]:
2424
tensor_constraints = [
25-
cp.Dtype.In(lambda deps: [torch.int, torch.float]),
26-
cp.Dtype.NotIn(lambda deps: [torch.int64, torch.float64]),
25+
cp.Dtype.In(lambda deps: [torch.int8, torch.int16, torch.uint8, torch.uint16, torch.float32]),
2726
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
2827
cp.Value.Le(lambda deps, dtype, struct: 2**4),
2928
cp.Rank.Ge(lambda deps: 1),
3029
cp.Size.Ge(lambda deps, r, d: 1),
3130
cp.Size.Le(lambda deps, r, d: 2**9),
31+
cp.Rank.Le(lambda deps: 2**3),
3232
]
3333

3434
match op_name:
3535
case "where.self":
3636
if index == 0: # condition
3737
tensor_constraints = [
3838
cp.Dtype.In(lambda deps: [torch.bool]),
39-
cp.Dtype.NotIn(lambda deps: [torch.int64, torch.float64]),
4039
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
4140
cp.Value.Le(lambda deps, dtype, struct: 2**4),
4241
cp.Rank.Ge(lambda deps: 1),
@@ -45,28 +44,35 @@ def apply_tensor_contraints(op_name: str, index: int) -> list[object]:
4544
]
4645
else:
4746
tensor_constraints = [
47+
cp.Dtype.In(lambda deps: [torch.int8, torch.int16, torch.uint8, torch.uint16, torch.float32]),
48+
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
49+
cp.Value.Le(lambda deps, dtype, struct: 2**4),
50+
cp.Rank.Ge(lambda deps: 1),
51+
cp.Size.Ge(lambda deps, r, d: 1),
52+
cp.Size.Le(lambda deps, r, d: 2**9),
53+
]
54+
case "embedding.default":
55+
tensor_constraints = [
4856
cp.Dtype.In(lambda deps: [torch.float, torch.int]),
4957
cp.Dtype.NotIn(lambda deps: [torch.int64, torch.float64]),
5058
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
5159
cp.Value.Le(lambda deps, dtype, struct: 2**4),
5260
cp.Rank.Ge(lambda deps: 1),
5361
cp.Size.Ge(lambda deps, r, d: 1),
5462
cp.Size.Le(lambda deps, r, d: 2**9),
55-
]
63+
]
5664
case "sigmoid.default":
5765
tensor_constraints.extend(
5866
[
59-
cp.Dtype.In(lambda deps: [torch.float]),
60-
cp.Rank.Le(lambda deps: 2**2),
67+
cp.Dtype.In(lambda deps: [torch.float32]),
6168
cp.Value.Ge(lambda deps, dtype, struct: -2),
6269
cp.Value.Le(lambda deps, dtype, struct: 2),
6370
]
6471
)
6572
case "rsqrt.default":
6673
tensor_constraints.extend(
6774
[
68-
cp.Dtype.In(lambda deps: [torch.float]),
69-
cp.Rank.Le(lambda deps: 2**2),
75+
cp.Dtype.In(lambda deps: [torch.float32]),
7076
cp.Value.Gt(
7177
lambda deps, dtype, struct: 0
7278
), # only generate real numbers
@@ -76,14 +82,12 @@ def apply_tensor_contraints(op_name: str, index: int) -> list[object]:
7682
case "mean.dim":
7783
tensor_constraints.extend(
7884
[
79-
cp.Dtype.In(lambda deps: [torch.float]),
80-
cp.Rank.Le(lambda deps: 2**2),
85+
cp.Dtype.In(lambda deps: [torch.float32]),
8186
]
8287
)
8388
case "exp.default":
8489
tensor_constraints.extend(
8590
[
86-
cp.Rank.Le(lambda deps: 2**3),
8791
cp.Value.Ge(lambda deps, dtype, struct: -(2**2)),
8892
cp.Value.Le(lambda deps, dtype, struct: 2**2),
8993
]
@@ -96,20 +100,82 @@ def apply_tensor_contraints(op_name: str, index: int) -> list[object]:
96100
cp.Value.Le(lambda deps, dtype, struct: 2),
97101
]
98102
)
99-
case _:
103+
case "constant_pad_nd.default":
100104
tensor_constraints.extend(
101105
[
102-
cp.Rank.Le(lambda deps: 2**2),
106+
cp.Dtype.In(lambda deps: [torch.float32]),
107+
cp.Size.Le(lambda deps, r, d: 2**2),
108+
]
109+
)
110+
case "avg_pool2d.default":
111+
tensor_constraints.extend(
112+
[
113+
cp.Rank.Eq(lambda deps: 4),
114+
]
115+
)
116+
case "bmm.default" | "addmm.default" | "mm.default":
117+
tensor_constraints.extend(
118+
[
119+
cp.Dtype.Eq(lambda deps: torch.float32),
120+
cp.Size.Le(lambda deps, r, d: 2**2),
121+
cp.Value.Le(lambda deps, dtype, struct: 2**4),
122+
]
123+
)
124+
case "div.Tensor":
125+
tensor_constraints.extend(
126+
[
127+
cp.Value.Ne(lambda deps, dtype, struct: 0),
128+
]
129+
)
130+
case "div.Tensor_mode" | "minimum.default":
131+
if index == 0:
132+
tensor_constraints = [
133+
cp.Dtype.In(lambda deps: [torch.int64, torch.int32, torch.float32]),
134+
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
135+
cp.Value.Le(lambda deps, dtype, struct: 2**4),
136+
cp.Rank.Ge(lambda deps: 1),
137+
cp.Size.Ge(lambda deps, r, d: 1),
138+
cp.Size.Le(lambda deps, r, d: 2**2),
139+
]
140+
else:
141+
tensor_constraints = [
142+
cp.Dtype.In(lambda deps: [torch.int64, torch.int32, torch.float32]),
143+
cp.Value.Ge(lambda deps, dtype, struct: -(2**4)),
144+
cp.Value.Le(lambda deps, dtype, struct: 2**4),
145+
cp.Rank.Ge(lambda deps: 1),
146+
cp.Rank.Eq(lambda deps: deps[0].dim()),
147+
cp.Size.Eq(lambda deps, r, d: fn.safe_size(deps[0], d)),
148+
]
149+
case "_native_batch_norm_legit_no_training.default":
150+
tensor_constraints.extend(
151+
[
152+
cp.Rank.Le(lambda deps: 3),
153+
],
154+
)
155+
case "reciprocal.default":
156+
tensor_constraints = [
157+
cp.Value.Ge(lambda deps, dtype, struct: -(2**2)),
158+
cp.Value.Le(lambda deps, dtype, struct: 2**2),
159+
cp.Size.Le(lambda deps, r, d: 2**3),
160+
]
161+
case "_softmax.default":
162+
tensor_constraints.extend(
163+
[
164+
cp.Dtype.Eq(lambda deps: torch.float32),
165+
cp.Size.Le(lambda deps, r, d: 2**2),
103166
]
104167
)
168+
case _:
169+
pass
105170
return tensor_constraints
106171

107172

108173
def apply_scalar_contraints(op_name: str) -> list[ScalarDtype]:
109174
match op_name:
110-
case "add.Scalar" | "sub.Scalar" | "mul.Scalar" | "div.Scalar":
175+
case "add.Scalar" | "sub.Scalar" | "mul.Scalar" | "div.Scalar" | "constant_pad_nd.default":
176+
return [ScalarDtype.int]
177+
case "full.default":
111178
return [ScalarDtype.int]
112-
113179
case _:
114180
return [ScalarDtype.float, ScalarDtype.int]
115181

@@ -149,6 +215,12 @@ def facto_testcase_gen(op_name: str) -> List[Tuple[List[str], OrderedDict[str, s
149215
cp.Dtype.In(lambda deps: apply_scalar_contraints(op_name)),
150216
]
151217
)
218+
if in_spec.name == "dtype": # full.default
219+
spec.inspec[index].constraints.extend(
220+
[
221+
cp.Dtype.In(lambda deps: [torch.long, torch.float]),
222+
]
223+
)
152224
elif in_spec.type.is_tensor():
153225
spec.inspec[index].constraints.extend(
154226
apply_tensor_contraints(op_name, index)
@@ -166,6 +238,29 @@ def facto_testcase_gen(op_name: str) -> List[Tuple[List[str], OrderedDict[str, s
166238
cp.Dtype.In(lambda deps: [torch.bool]),
167239
]
168240
)
241+
elif in_spec.type.is_length_list():
242+
spec.inspec[index].constraints.extend(
243+
[
244+
cp.Value.Ge(lambda deps, dtype, struct: 0),
245+
]
246+
)
247+
if op_name == "avg_pool2d.default":
248+
spec.inspec[index].constraints.extend(
249+
[
250+
cp.Length.Eq(lambda deps: 2),
251+
]
252+
)
253+
elif in_spec.type.is_shape():
254+
spec.inspec[index].constraints.extend(
255+
[
256+
cp.Rank.Ge(lambda deps: 1),
257+
cp.Rank.Le(lambda deps: 2**2),
258+
cp.Value.Gt(lambda deps, dtype, struct: 0),
259+
cp.Value.Le(lambda deps, dtype, struct: 2**2),
260+
cp.Size.Ge(lambda deps, r, d: 1),
261+
cp.Size.Le(lambda deps, r, d: 2**2),
262+
]
263+
)
169264

170265
return [
171266
(posargs, inkwargs)

0 commit comments

Comments
 (0)