feectools: inner products stay on the device

[max-models]

Stack: part 3 of 3 — based on #706 (feectools-stencil-6d-views, merge that first, after #703), next: none (top of the stack).

Corresponding update in feectools: struphy-hub/feectools#98. Mirrors the feectools stack struphy-hub/feectools#96 → struphy-hub/feectools#97 → struphy-hub/feectools#98. feectools-stencil-6d-views is merged into this branch (no rebase), and the feectools submodule now points at the stacked head of struphy-hub/feectools#98, 6d88806, which contains feectools#96 and #97. Against its base (#706) this PR only moves the submodule d16a6ad → 6d88806. After each feectools PR of the stack is merged into devel-tiny, the feectools submodule of the struphy stack must be moved to the corresponding merged devel-tiny commit (the pr-feectools-submodule check fails until then).

---

Solves the following issue(s):

Part of #689 (VlasovAmpereOneSpecies end to end on the GPU, no host/device transfers in the time loop), tracked in #650. This PR moves the feectools submodule to struphy-hub/feectools#98 ("Inner products stay on the device"). On the CuPy backend, every CG iteration of a struphy solve copied two scalars from the device to the host; after this change it copies one (the convergence test).

The pr-feectools-submodule check fails until struphy-hub/feectools#98 is merged into devel-tiny. After that, the submodule should point at the merge commit.

Core changes:

• feectools submodule: d16a6ad (base feectools: stencil 3D device kernels with 6D views #706) → 6d88806 (branch inner-on-device of Inner products stay on the device feectools#98, stacked on feectools#96 and Remove deprecated console commands #97; originally a15a8e8). No struphy code changes.

• What changes in feectools, and only on the CuPy backend:

  • StencilVector.inner/BlockVector.inner/dot_inner return a 0-d device array, in serial and with MPI.
  • axpy accepts such a scalar without copying it to the host.
  • CG, PCG, BiCG, BiCGStab and PBiCGStab keep alpha/beta on the device and copy only the residual norm, once per iteration.

  On NumPy, results and types are unchanged.

  Copies to the host per iteration, counted under cunumpy's fake CuPy:

  Solver	before	after
  CG	2	1
  PCG	3	1
  BiCG	4	1
  BiCGStab	6	1
  PBiCGStab	5	1

• struphy call sites of .inner that run on CuPy now get a 0-d device array. I checked them:

  • the scalars in models/scalars.py write it into xp buffers (local_value[0] = ...), which works on the device;
  • the model energy methods (linear_mhd.py, shear_alfven.py, ...) return it, and the scalar machinery handles it like the existing dot_inner results;
  • PolarVector.dot adds a NumPy scalar to it, which works, but polar splines are not on CuPy yet (Polar splines on the CuPy backend #695);
  • the multigrid smoothers run their own PCG with .inner. It works with device scalars: the arithmetic stays on the device, and the comparisons with 0 copy to the host implicitly.

• This removes the first blocker listed in CUDA version of the linear_vlasov_ampere accumulation #705 ("CG inner products ... The Schur solve is outside the transfer guard because of this"). The remaining copy per CG iteration still counts as a transfer for assert_no_transfers. The transfer guard could then cover the Schur solve with an allowance of one 8-byte copy per iteration.

Model-specific changes:

None.

Documentation changes:

None in struphy. The feectools PR documents the change in feectools' CUDA_STRATEGY.md (section "Inner products on the device").

Testing (macOS, no GPU; struphy kernels compiled with GNU/Fortran; GPU tests not run):

• feectools (see Inner products stay on the device feectools#98):

  • serial suite: 9456 → 9480 passed, with the same 6 unrelated failures in ddm/tests/test_cart_*d.py;
  • mpirun -n 2 linalg MPI suite: 738 → 739 passed;
  • the new fake-CuPy tests check the copy counts and that the results equal NumPy's.

• struphy, solver tests on the NumPy backend, with feectools devel-tiny (before) and this branch (after):

  • propagators/tests/test_poisson.py -m "not mpi" -k "not multigrid": 40 passed before, 40 passed after;
  • linear_algebra/tests/test_saddlepoint_massmatrices.py: the first case (SaddlePointSolverUzawaNumpy) passed before and after.

  I stopped the remaining saddle-point and multigrid tests: with the feectools kernels uncompiled in the checkouts, each test took more than 10 minutes. On NumPy, feectools' change is limited to identity helpers and an equivalent loop structure in BiCGStab, and the feectools solver tests check the same iteration counts and results.

🤖 Generated with Claude Code