xref: /aosp_15_r20/external/pytorch/test/test_cuda_primary_ctx.py (revision da0073e96a02ea20f0ac840b70461e3646d07c45) 
1# Owner(s): ["module: cuda"]
2
3import sys
4import unittest
5
6import torch
7from torch.testing._internal.common_cuda import TEST_CUDA, TEST_MULTIGPU
8from torch.testing._internal.common_utils import (
9    NoTest,
10    run_tests,
11    skipIfRocmVersionLessThan,
12    TestCase,
13)
14
15
16# NOTE: this needs to be run in a brand new process
17
18if not TEST_CUDA:
19    print("CUDA not available, skipping tests", file=sys.stderr)
20    TestCase = NoTest  # noqa: F811
21
22
23@torch.testing._internal.common_utils.markDynamoStrictTest
24class TestCudaPrimaryCtx(TestCase):
25    CTX_ALREADY_CREATED_ERR_MSG = (
26        "Tests defined in test_cuda_primary_ctx.py must be run in a process "
27        "where CUDA contexts are never created. Use either run_test.py or add "
28        "--subprocess to run each test in a different subprocess."
29    )
30
31    @skipIfRocmVersionLessThan((4, 4, 21504))
32    def setUp(self):
33        for device in range(torch.cuda.device_count()):
34            # Ensure context has not been created beforehand
35            self.assertFalse(
36                torch._C._cuda_hasPrimaryContext(device),
37                TestCudaPrimaryCtx.CTX_ALREADY_CREATED_ERR_MSG,
38            )
39
40    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
41    def test_str_repr(self):
42        x = torch.randn(1, device="cuda:1")
43
44        # We should have only created context on 'cuda:1'
45        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
46        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
47
48        str(x)
49        repr(x)
50
51        # We should still have only created context on 'cuda:1'
52        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
53        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
54
55    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
56    def test_copy(self):
57        x = torch.randn(1, device="cuda:1")
58
59        # We should have only created context on 'cuda:1'
60        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
61        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
62
63        y = torch.randn(1, device="cpu")
64        y.copy_(x)
65
66        # We should still have only created context on 'cuda:1'
67        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
68        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
69
70    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
71    def test_pin_memory(self):
72        x = torch.randn(1, device="cuda:1")
73
74        # We should have only created context on 'cuda:1'
75        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
76        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
77
78        self.assertFalse(x.is_pinned())
79
80        # We should still have only created context on 'cuda:1'
81        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
82        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
83
84        x = torch.randn(3, device="cpu").pin_memory()
85
86        # We should still have only created context on 'cuda:1'
87        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
88        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
89
90        self.assertTrue(x.is_pinned())
91
92        # We should still have only created context on 'cuda:1'
93        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
94        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
95
96        x = torch.randn(3, device="cpu", pin_memory=True)
97
98        # We should still have only created context on 'cuda:1'
99        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
100        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
101
102        x = torch.zeros(3, device="cpu", pin_memory=True)
103
104        # We should still have only created context on 'cuda:1'
105        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
106        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
107
108        x = torch.empty(3, device="cpu", pin_memory=True)
109
110        # We should still have only created context on 'cuda:1'
111        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
112        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
113
114        x = x.pin_memory()
115
116        # We should still have only created context on 'cuda:1'
117        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
118        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))
119
120
121if __name__ == "__main__":
122    run_tests()
123