Fix LagrangeSpace device captures for CUDA

[aaadelmann]

• replace class captures with a lightweight device mirror
• add device-side mesh, DOF, and boundary helpers
• capture field views directly in Kokkos kernels
• eliminate GH200 host-only destructor warnings

LUMI Results

Red old master:

Justification

Observed behavior	Most likely related change	Confidence
PCG is approximately 18% faster at 16 GPUs, with the improvement decreasing at larger GPU counts	e0e88e188: Poisson and differential expressions capture only the Kokkos field view instead of the complete Field object, producing smaller device-kernel closures	High
Additional PCG improvement may come from reduced halo overhead	e6e2387d0: halo packing and unpacking use submdspan; PCG performs halo exchanges repeatedly during its iterations	Medium
FFT is approximately unchanged at 16–32 GPUs and modestly faster at 64–512 GPUs	Shared infrastructure changes in e6e2387d0, including direct reallocation of overwritten field/layout storage and updated particle bookkeeping	Low–medium
FEM performance is essentially unchanged, with mixed differences of approximately 1–2%	9d9387c57: Lagrange kernels capture a lightweight device mirror and the required views instead of the complete host-side object	Medium; no clear performance effect
No expected performance impact	Nedelec mirror, FEL warning fixes, Kokkos::pair, [[maybe_unused]], Archive initialization, and unit-test changes	High

The PCG mapping is particularly strong. Before e0e88e188, every PCG Laplace expression stored a complete field:

const E u_m;

Now it stores only the device data handle:

using view_type = typename E::view_type;
const view_type u_m;

This affects meta_poisson, meta_lower_laplace, meta_upper_laplace, and meta_upper_and_lower_laplace, which are repeatedly invoked by PoissonCG. It produces smaller kernel closures and gives the AMD compiler much less host-side ownership state to carry through device kernels.

The scaling signature supports that explanation: the improvement is largest at 16 GPUs and shrinks as communication becomes a larger fraction of runtime. That is typical of a compute-kernel improvement.

For FFT, no FFT algorithm was optimized in PR593. The PrunedCC change only adds [[maybe_unused]], so it cannot explain the improvement. Likely shared contributors are:

• Kokkos::resize → Kokkos::realloc where old field/layout contents are discarded.
• Persistent host mirror for the particle destination counter.
• The new halo packing/unpacking path.

Because the FFT differences are small, measurement variability remains plausible.

Several changes can be ruled out:

• Nedelec device mirror: not used by these runs.
• FEL initialization/declaration fixes: unrelated.
• Kokkos::pair: intended to be runtime-equivalent.
• Archive value initialization and test changes: unrelated.

To establish causality, the cleanest LUMI comparison would be:

1. 22095a0f8 versus e0e88e188 — isolates the expression change; inspect CG and applyOp.
2. ee4985bc3 versus e6e2387d0 — isolates the Kokkos 5 compatibility group; inspect updateLayout, halo time, and locateParticles.
3. Compare identical PCG iteration counts and final residuals, with at least three runs per point.

My current conclusion: the PCG improvement is very likely a real benefit of the field-view expression architecture; the FFT improvement is probably shared infrastructure but needs an ablation run; FEM is essentially performance-neutral.

ALPS

[aaadelmann]

cscs-ci run cscs-ci-gh200, cscs-ci-mi300, cscs-ci-openmp

[aaadelmann]

cscs-ci run cscs-ci-gh200, cscs-ci-mi300, cscs-ci-openmp

[aaadelmann]

cscs-ci run cscs-ci-gh200, cscs-ci-mi300, cscs-ci-openmp