[saqibkh]

The CI here has no job for the CUDA backend: tests.yml only builds and runs OpenMP, because the hosted runners have no GPU. This adds a separate, optional workflow .github/workflows/gpu-sim.yml that builds stdgpu with the CUDA backend and runs the unit tests on a simulated NVIDIA H100, using PantheonSim (via the pantheongpu/setup-pantheonsim action, pinned by SHA). I am the author of PantheonSim, so please weigh this accordingly.

It does not touch the existing workflows and is not meant to be a required check. A simulator cannot reproduce hardware-specific behavior or timing, so it complements real-GPU testing and does not replace it.

What it does:

• Installs CUDA 13.0 and the simulated H100, then configures with STDGPU_BACKEND_CUDA, STDGPU_BUILD_SHARED_LIBS=ON, CMAKE_CUDA_ARCHITECTURES=90-virtual and builds with ccache.
• Runs teststdgpu through vgpu run ctest (all gtest cases, CUDA tests included), then runs the 12 example programs under vgpu run.
• The simulator needs a shared CUDA runtime and PTX-only device code. stdgpu links cudart_static when built as a static library, so the workflow builds the shared library, which links cudart; 90-virtual gives PTX only. No change to the project is needed.
• Triggers: workflow_dispatch and pull_request limited to src/, tests/, examples/, cmake/, CMakeLists.txt and the workflow itself. A concurrency group cancels superseded runs, with permissions: contents: read.

What is left out:

• The benchmarks are not built, since they do not test anything.
• Only one architecture (H100) and one CUDA version (13.0) are covered, and only a Release build with a shared library.
• Nothing is excluded from the test run: all 702 test cases pass.

Result on a fork: the run passed (100% tests passed, 0 tests failed out of 1, which is the single teststdgpu ctest entry with 702 gtest cases, and all 12 examples exit 0). The whole job took about 22.5 minutes on a cold cache, of which about 16.5 minutes was the action setting up the simulator on its first run. The build took about 3 minutes and the unit tests about 3 minutes.

Run: https://github.com/saqibkh/stdgpu/actions/runs/42290609398