[BUG]: libcu++ barrier*.pass tests timeout on t4, rtx2080, v100 *only* when compiled for the target arch.

[alliepiper]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Runtime Error

Component

libcu++

Describe the bug

While migrating the CI to use nvks runners, two libcu++ tests fail with timeouts: heterogeneous/barrier_abi_v2.pass.cpp and heterogeneous/barrier.pass.cpp.

Note that this only happens when specifying an explicit arch. The tests pass on rtx2080, t4, and v100 when compiled for 60;70;80, but fail when explicitly compiled only for 70 (v100) or 75 (t4, rtx2080).

Failure output from CI log https://github.com/NVIDIA/cccl/actions/runs/13022094204/job/36324973191:

  1: UNRESOLVED: libcu++ :: heterogeneous/barrier.pass.cpp (362 of 2341)
  1: ******************** TEST 'libcu++ :: heterogeneous/barrier.pass.cpp' FAILED ********************
  1: Exception during script execution:
  1: Traceback (most recent call last):
  1:   File "/home/coder/.local/share/venvs/cccl/lib/python3.10/site-packages/lit/worker.py", line 76, in _execute_test_handle_errors
  1:     result = test.config.test_format.execute(test, lit_config)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 83, in execute
  1:     return self._execute(test, lit_config)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 158, in _execute
  1:     return self._evaluate_pass_test(
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 208, in _evaluate_pass_test
  1:     cmd, out, err, rc = self.executor.run(
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/executor.py", line 43, in run
  1:     out, err, rc = executeCommand(cmd, cwd=work_dir, env=env, timeout=self.timeout)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/util.py", line 242, in executeCommand
  1:     raise ExecuteCommandTimeoutException(
  1: libcudacxx.util.ExecuteCommandTimeoutException
  1: 
  1: 
  1: ********************
  1: UNRESOLVED: libcu++ :: heterogeneous/barrier_abi_v2.pass.cpp (364 of 2341)
  1: ******************** TEST 'libcu++ :: heterogeneous/barrier_abi_v2.pass.cpp' FAILED ********************
  1: Exception during script execution:
  1: Traceback (most recent call last):
  1:   File "/home/coder/.local/share/venvs/cccl/lib/python3.10/site-packages/lit/worker.py", line 76, in _execute_test_handle_errors
  1:     result = test.config.test_format.execute(test, lit_config)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 83, in execute
  1:     return self._execute(test, lit_config)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 158, in _execute
  1:     return self._evaluate_pass_test(
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/format.py", line 208, in _evaluate_pass_test
  1:     cmd, out, err, rc = self.executor.run(
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/test/executor.py", line 43, in run
  1:     out, err, rc = executeCommand(cmd, cwd=work_dir, env=env, timeout=self.timeout)
  1:   File "/home/coder/cccl/libcudacxx/test/utils/libcudacxx/util.py", line 242, in executeCommand
  1:     raise ExecuteCommandTimeoutException(
  1: libcudacxx.util.ExecuteCommandTimeoutException

How to Reproduce

Run libcu++ tests on a t4 or rtx2080. Possibly any sm75 system?

To do this through CI, edit ci/matrix.yaml and add the following lines to the override section:

  override:
    - {jobs: ['test'],  project: ['libcudacxx'], std: 'max', cxx: ['gcc', 'clang'], gpu: 't4', sm: 'gpu'}
    - {jobs: ['test'],  project: ['libcudacxx'], std: 'max', cxx: ['gcc', 'clang'], gpu: 'rtx2080', sm: 'gpu'}

This will only run the relevant jobs to ensure that the bug is fixed.

Expected behavior

The tests should pass regardless of hardware.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response