Kokkos OpenMP¶

What it is. Same Kokkos backend, OpenMP execution space: multi-thread parallelism on CPU. for_each_cell becomes a multi-thread parallel_for (with a serial-switch threshold for the small V-cycle grids, cf. POPS_FOREACH_SERIAL_THRESHOLD).

Build. Kokkos installed with Kokkos_ENABLE_OPENMP=ON. The -DPOPS_USE_KOKKOS=ON on the adc_cpp side is identical to config 1; it is the Kokkos install that changes:

cmake -S . -B build-kokkos-omp -DCMAKE_BUILD_TYPE=Release \
  -DPOPS_USE_KOKKOS=ON -DKokkos_ROOT="$KOKKOS_OPENMP_PREFIX"
cmake --build build-kokkos-omp -j

Run. Bound the number of threads on small machines:

OMP_NUM_THREADS=4 ctest --test-dir build-kokkos-omp --output-on-failure

Validation: CI (ci-full job). Job kokkos-openmp (ubuntu-latest / Kokkos (OpenMP)), Kokkos 4.4.01 OpenMP, OMP_NUM_THREADS=2. Full mode only (unlike the Serial gate, which runs on every PR). Status ci-full, 91/91 ctest, 0 failure.

FP note. The sum reduction (Kokkos::Sum) reassociates floating-point addition per tile: deterministic/idempotent (same data + same Kokkos space -> same bits) but not bit-identical to a hand-written lexicographic sum. Since there is only one Kokkos path, this holds for all spaces (Serial, OpenMP, Cuda). The max reduction (Kokkos::Max) is exact everywhere. Detail in the header of for_each.hpp.