CI: run the gpu_octree tests on a simulated GPU

[saqibkh]

The Azure pipelines build the GPU modules but configure with PCL_DISABLE_GPU_TESTS=ON, because their agents have no GPU, so the gpu_octree tests never run in CI. This adds a GitHub Actions job that runs them on a simulated NVIDIA T4.

How: PantheonSim, via pantheongpu/setup-pantheonsim, installs the CUDA toolkit and simulates the GPU on the runner's CPU. The kernels run from their PTX (CUDA_ARCH_BIN=75 embeds it), so the results are real. There is no timing model, so this checks correctness, not performance.

Changes: one new workflow, .github/workflows/gpu-tests.yml, on a plain ubuntu-24.04 runner:

• set up CUDA 12.8 and a simulated T4;
• configure with only the modules the tests need (common, octree, gpu_containers, gpu_utils, gpu_octree), CUDA_ARCH_BIN=75 and the shared CUDA runtime, since the simulator stands in for libcudart;
• build the five test programs and run them with ctest.

gpu_octree_performance is left out: it is a timing benchmark, and on a simulated GPU it runs for over 50 minutes. The workflow only runs when the GPU octree code, its dependencies, the CMake files or the workflow itself change. The Azure pipelines are unchanged.

What it can touch: the action is pinned by commit (v0.1.1), not by a tag, and it pins the simulator it builds by commit too, so nothing in the job changes until you move the pin. The workflow's token is read-only (permissions: contents: read), and the action uses no token or secrets; it installs packages only from Ubuntu's and NVIDIA's repositories. The job isn't a required check, so a red run never blocks a merge unless you decide it should.

What it gives, from a run on my fork:

• gpu_octree_approx_nearest, bfrs, host_radius, knn_search and radius_search all pass, in about 3 minutes.
• The job takes about 7 minutes. The action caches the simulator it builds, per pinned version, so only the first run in the repository builds it (about 20 minutes that time).

Would it have caught past bugs? I undid two earlier fixes on current master and ran the same job (run):

• [gpu] Fix octree radiusSearch #4146 (warp-synchronous volatile shared memory replaced by __shfl in radius_search.cu): with the fix undone, batchRadiusSearch fails, in the search with a different radius per query, which is the case Inaccurate results with gpu octree search example #3583 reported.
• [FIX] gpu approx nearest unit test #6452 (the exact float comparison in approxNearesSearch): with the fix undone, the test fails with the same two values reported from a real GPU, 0.0300000142 against 0.0300000124.

Disclosure: I maintain PantheonSim. If the job is ever flaky or wrong, please open an issue at pantheongpu/pantheonsim and I'll fix it on our side.

🤖 Generated with Claude Code