File: CUDAAssertionsTest_multiple_writes_from_same_block.cu

package info (click to toggle)
pytorch-cuda 2.6.0%2Bdfsg-7
  • links: PTS, VCS
  • area: contrib
  • in suites: forky, sid, trixie
  • size: 161,620 kB
  • sloc: python: 1,278,832; cpp: 900,322; ansic: 82,710; asm: 7,754; java: 3,363; sh: 2,811; javascript: 2,443; makefile: 597; ruby: 195; xml: 84; objc: 68
file content (78 lines) | stat: -rw-r--r-- 2,525 bytes parent folder | download | duplicates (3)
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
#include <gmock/gmock.h>
#include <gtest/gtest.h>

#include <c10/cuda/CUDADeviceAssertion.h>
#include <c10/cuda/CUDAException.h>
#include <c10/cuda/CUDAFunctions.h>
#include <c10/cuda/CUDAStream.h>

#include <chrono>
#include <iostream>
#include <string>
#include <thread>

using ::testing::HasSubstr;

const auto max_assertions_failure_str =
    "Assertion failure " + std::to_string(C10_CUDA_DSA_ASSERTION_COUNT - 1);

/**
 * Device kernel that takes a single integer parameter as argument and
 * will always trigger a device side assertion.
 */
__global__ void cuda_always_fail_assertion_kernel(
    const int a,
    TORCH_DSA_KERNEL_ARGS) {
  CUDA_KERNEL_ASSERT2(a != a);
}

/**
 * TEST: Triggering device side assertion from single block and multiple threads
 * <<<1,128>>>. Once the very first thread asserts all the other threads will
 * basically be in bad state and the block id with failed assertion would be
 * [0,0,0].
 */
void cuda_device_assertions_multiple_writes_from_same_block() {
  const auto stream = c10::cuda::getStreamFromPool();
  TORCH_DSA_KERNEL_LAUNCH(
      cuda_always_fail_assertion_kernel,
      1, /* Blocks */
      128, /* Threads */
      0, /* Shared mem */
      stream, /* Stream */
      1);

  try {
    c10::cuda::device_synchronize();
    throw std::runtime_error("Test didn't fail, but should have.");
  } catch (const c10::Error& err) {
    const auto err_str = std::string(err.what());
    ASSERT_THAT(err_str, HasSubstr(max_assertions_failure_str));
    ASSERT_THAT(err_str, HasSubstr("Block ID that failed assertion = [0,0,0]"));
    ASSERT_THAT(err_str, HasSubstr("Device that launched kernel = 0"));
    ASSERT_THAT(
        err_str,
        HasSubstr(
            "Name of kernel launched that led to failure = cuda_always_fail_assertion_kernel"));
    ASSERT_THAT(
        err_str, HasSubstr("File containing kernel launch = " __FILE__));
    ASSERT_THAT(
        err_str,
        HasSubstr(
            "Function containing kernel launch = " +
            std::string(__FUNCTION__)));
    ASSERT_THAT(
        err_str,
        HasSubstr(
            "Stream kernel was launched on = " + std::to_string(stream.id())));
  }
}

TEST(CUDATest, cuda_device_assertions_multiple_writes_from_same_block) {
#ifdef TORCH_USE_CUDA_DSA
  c10::cuda::CUDAKernelLaunchRegistry::get_singleton_ref().enabled_at_runtime = true;
  cuda_device_assertions_multiple_writes_from_same_block();
#else
  GTEST_SKIP() << "CUDA device-side assertions (DSA) was not enabled at compile time.";
#endif
}