File: test_cuda_primary_ctx.py

package info (click to toggle)
pytorch 1.13.1%2Bdfsg-4
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 139,252 kB
  • sloc: cpp: 1,100,274; python: 706,454; ansic: 83,052; asm: 7,618; java: 3,273; sh: 2,841; javascript: 612; makefile: 323; xml: 269; ruby: 185; yacc: 144; objc: 68; lex: 44
file content (115 lines) | stat: -rw-r--r-- 4,484 bytes parent folder | download
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
# Owner(s): ["module: cuda"]

import torch
from torch.testing._internal.common_utils import TestCase, run_tests, skipIfRocmVersionLessThan
import sys
import unittest

# NOTE: this needs to be run in a brand new process

# We cannot import TEST_CUDA and TEST_MULTIGPU from torch.testing._internal.common_cuda here,
# because if we do that, the TEST_CUDNN line from torch.testing._internal.common_cuda will be executed
# multiple times as well during the execution of this test suite, and it will
# cause CUDA OOM error on Windows.
TEST_CUDA = torch.cuda.is_available()
TEST_MULTIGPU = TEST_CUDA and torch.cuda.device_count() >= 2

if not TEST_CUDA:
    print('CUDA not available, skipping tests', file=sys.stderr)
    TestCase = object  # noqa: F811


class TestCudaPrimaryCtx(TestCase):
    CTX_ALREADY_CREATED_ERR_MSG = (
        "Tests defined in test_cuda_primary_ctx.py must be run in a process "
        "where CUDA contexts are never created. Use either run_test.py or add "
        "--subprocess to run each test in a different subprocess.")

    @skipIfRocmVersionLessThan((4, 4, 21504))
    def setUp(self):
        for device in range(torch.cuda.device_count()):
            # Ensure context has not been created beforehand
            self.assertFalse(torch._C._cuda_hasPrimaryContext(device), TestCudaPrimaryCtx.CTX_ALREADY_CREATED_ERR_MSG)

    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
    def test_str_repr(self):
        x = torch.randn(1, device='cuda:1')

        # We should have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        str(x)
        repr(x)

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
    def test_copy(self):
        x = torch.randn(1, device='cuda:1')

        # We should have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        y = torch.randn(1, device='cpu')
        y.copy_(x)

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

    @unittest.skipIf(not TEST_MULTIGPU, "only one GPU detected")
    def test_pin_memory(self):
        x = torch.randn(1, device='cuda:1')

        # We should have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        self.assertFalse(x.is_pinned())

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        x = torch.randn(3, device='cpu').pin_memory()

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        self.assertTrue(x.is_pinned())

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        x = torch.randn(3, device='cpu', pin_memory=True)

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        x = torch.zeros(3, device='cpu', pin_memory=True)

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        x = torch.empty(3, device='cpu', pin_memory=True)

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

        x = x.pin_memory()

        # We should still have only created context on 'cuda:1'
        self.assertFalse(torch._C._cuda_hasPrimaryContext(0))
        self.assertTrue(torch._C._cuda_hasPrimaryContext(1))

if __name__ == '__main__':
    run_tests()