[greole]

This PR ports OGL to the current Ginkgo develop API and uses the reusable multigrid hierarchy of LinOpFactory::generate_reuse from ginkgo-project/ginkgo#2117, which is merged into Ginkgo develop and supersedes the closed ginkgo-project/ginkgo#2099. It also moves the LRZ GitLab CI to images with Ginkgo develop and OpenFOAM.

Changes

Port to the current Ginkgo API

• Ginkgo develop removed EnableLinOp and EnablePolymorphicObject: CombinationMatrix, RepartDistMatrix and Adaptive now derive from gko::LinOp and gko::EnableCloneable, and OpenFOAMDistStoppingCriterion derives from gko::stop::Criterion
• use get_diag_matrix()/get_off_diag_matrix() instead of the deprecated get_local_matrix()/get_non_local_matrix() of the distributed matrix
• OGL_GINKGO_CHECKOUT_VERSION defaults to develop

Multigrid hierarchy reuse

• new opt-in Multigrid option reuseHierarchy (default false), for both the Schwarz and Distributed type
• the preconditioner is generated via LinOpFactory::generate_reuse, and the reuse data is kept in the object registry. Regenerations keep the recorded Pgm aggregates and transfer operators and only recompute the coarse matrices, smoothers and the coarsest solver
• if the matrix no longer fits the recorded hierarchy, OGL logs the Ginkgo error and records it again
• new option maxHierarchyReuses (default 0, no limit): once the hierarchy has been reused that many times, the next regeneration records it again from the current matrix. The reuse count is kept per system matrix and reset whenever the hierarchy is recorded

LRZ GitLab CI

• run on the openfoam-ginkgo-{cuda,rocm,sycl} images of exasim-project/openfoam-conda: OpenFOAM v2512 on the Ginkgo develop images (Ginkgo 0dfd8c5, the merge of Enable PGM to reuse aggregates with the generate_reuse. ginkgo-project/ginkgo#2117). OpenFOAM uses the MPI of the image, so Pstream and Ginkgo load the same MPI. Dated tags, since the runner caches images by name
• sync the scripts with NeoN: activate OpenFOAM through the image's activate.sh, restore the MPICH/CUDA paths, launch the MPI tests with the MPI of the image and start the ranks locally, port the trigger script, drop the NeoN-specific parts and the Kokkos options
• trigger workflow: run on pushes to dev (was develop), only in hpsim/OGL and not for PRs from forks, and no longer cancel the pipelines after waiting for them
• detect the ESI version of OpenFOAM also via FOAM_API, since the conda packages do not install META-INFO

Fixes

• when the caching period of a preconditioner ran out, the regenerated preconditioner was not stored, so every later caching period reused the very first preconditioner
• removed the Multigrid value update for cached preconditioners. It was never compiled (guarded by GINKGO_WITH_OGL_EXTENSION, while CMake defines GINKGO_WITH_OGL_EXTENSIONS) and used UpdateMatrixValue, which Ginkgo no longer provides

Usage

preconditioner
{
    preconditioner  Multigrid;
    reuseHierarchy  true;
    maxHierarchyReuses 10;   // optional, record again after 10 reuses
}

The hierarchy is recorded from the first matrix, so reuse only pays off if later matrices are similar to it. maxHierarchyReuses limits how far the matrix can drift from the recorded one.

Compatibility

This branch no longer builds against the previous default Ginkgo (ogl_0600_gko190). It needs Ginkgo develop with ginkgo-project/ginkgo#2117, which is now the default checkout. All sources compile against Ginkgo develop (d5a8402, syntax-only check).

Testing

icoFoam on the 100x100 lid-driven cavity, 2 ranks on 2x AMD MI200 (gfx90a, ROCm 6.4.3), GKOCG with executor hip and scaling -1, run to t = 0.5 (200 pressure solves).

Measured on rebased/pr-2099, before ginkgo-project/ginkgo#2117 was merged; not yet re-run on develop. The new Ginkgo API gives the same results as the previous Ginkgo. Pressure iterations in the last time step:

Preconditioner	previous Ginkgo	rebased/pr-2099
none (OpenFOAM PCG + diagonal)	453	453
BJ	460	460
Multigrid	77	75
ILU	157	157
ISAI	246	246

Multigrid hierarchy reuse (Multigrid regenerates on every solve unless caching is set):

Variant	preconditioner setup per solve	last-step p iterations
Schwarz	9.33 ms	75
Schwarz + reuseHierarchy	3.16 ms	73
Distributed	33.87 ms	49
Distributed + reuseHierarchy	7.67 ms	49
Schwarz + reuseHierarchy + caching 5	0.79 ms	73

The final p and U fields at t = 0.5 are identical, to the written precision, with and without reuse and to the OpenFOAM baseline. The re-recording after a sparsity change is not covered, since the cavity mesh is static. maxHierarchyReuses has not been tested on GPUs yet.

The LRZ pipeline on the new images has not run yet; this PR is its first run.