Add code_saturne v9.1.0 to validation suite

[diskdog]

Summary

Adds clone, build, and run scripts for code_saturne v9.1.0, an open-source finite-volume CFD solver with CUDA-accelerated linear solvers. The validation runs a Cartesian channel-flow case and confirms that the GPU solver is dispatched and the run completes.

System prerequisites

The build requires packages not covered by the standard test environment:

sudo apt-get install -y autoconf automake libtool openmpi-bin libopenmpi-dev python3-dev

Build notes

code_saturne uses the autoconf/automake toolchain, not CMake.

Two non-obvious points:

• MPI detection. Ubuntu's multiarch layout places the MPI installation under a non-standard prefix. The script uses ompi_info --path libdir to locate it reliably regardless of distro.
• Architecture variable. code_saturne uses CUDA_ARCH_NUM (an integer, e.g. 75) rather than the CUDAARCHS variable that test.sh sets. The script bridges these; the fallback of 75 only applies when running outside test.sh.

Validation

02-run.sh creates a 100x100x1 Cartesian channel-flow case (inlet velocity at x=0, pressure outlet at x=1, no-slip walls elsewhere) and runs 10 time steps. It then asserts:

1. A CUDA device was initialised (CUDA device in run_solver.log).
2. The GPU sparse matrix-vector product was dispatched (device SpMV variant in performance.log).
3. The run completed cleanly (END OF CALCULATION in run_solver.log).

Known workaround

CS_CUDA_ALLOC_DEVICE_UVM=1 is set unconditionally at the start of 02-run.sh. Without it, cs_sles_solve_ccc_fv allocates its extended solver buffers as CS_ALLOC_DEVICE (device-only) and then passes them to the CPU-side convergence check, causing a SIGSEGV. The flag remaps cs_alloc_mode_device to CS_ALLOC_HOST_DEVICE_SHARED at initialisation time. See cs_base_cuda.cu:777.

Status

Passes locally on sm_75 under both native CUDA 13.1 and SCALE 1.7.0. Results are marked ? pending CI validation on the repo's target hardware.