Bump feectools: KroneckerStencilMatrix matmul and grouped factors

[spossann]

Summary

Points the feectools submodule to the branch of struphy-hub/feectools#100, so that the struphy test suite runs against it before that PR is merged.

struphy-hub/feectools#100 changes KroneckerStencilMatrix and KroneckerLinearSolver:

• factors on groups of axes (e.g. 2d x 1d);
• A @ B returns a new ComposedKroneckerStencilMatrix (operands in multiplicands, exact Kronecker factors of the product in mats);
• ComposedLinearOperator.multiplicants → multiplicands, with a deprecated alias;
• factor_ndims in KroneckerLinearSolver;
• docstrings and type annotations; feectools 0.6.0.

Existing calls are unchanged, so struphy needs no code changes. struphy builds these objects in feec/preconditioner.py (MassMatrixPreconditioner) and feec/projectors.py.

multiplicants → multiplicands

• All struphy uses (preconditioner, projectors, multigrid coarsening) now use multiplicands.
• linop_nbytes in feec/memory.py adds up all child attributes, so a ComposedKroneckerStencilMatrix counts both its operands and the factors of the product.

Refactor of MassMatrixPreconditioner

• The construction of the Kronecker approximation is split into module-level helpers in feec/preconditioner.py. They cover the boundary conditions, gathering array weights over MPI, the 1d weight, the 1d mass matrix and its BCs, the 1d solver, the process-local 1d factor, the block-diagonal assembly and the operator to invert.
• MassMatrixDiagonalPreconditioner now calls these helpers instead of repeating the same ~200 lines.
• Values and the order of MPI calls are unchanged. The copy into the process-local 1d stencil matrix is now vectorized, and the block-diagonal operators are no longer hard-coded to 3 components.
• Docstrings, comments and type annotations.

MassMatrixPreconditioner: new option and fixes

• New option weight_reduction="midpoint" | "average" sets how the weight is reduced to 1d in the directions other than dim_reduce. The default "midpoint" gives the same results as before, verified for M0, M1, M2, Mv, M1n and array weights, serially and on 2 ranks. "average" uses the mean over those directions: Gauss quadrature of the Derham for array weights, 16-point Gauss–Legendre for callables.
• Array weights are reduced with a single Allreduce, instead of a subcommunicator plus Bcast.
• The mass matrix is located in the composed operator by identity, and the approximate inverse is applied exactly there. Both preconditioners share this code (_apply_composed). Before, a mass matrix in the left-most position was applied instead of inverted.
• FFTSolver copies its column, which was a view of the dense matrix also used for the stencil factor. It now stabilises a singular matrix once, at construction. is_circulant is vectorized.

PCG iterations (tolerance 1e-8) on HollowTorus, grid 12×16×6, degree (2,3,2), as a reference for choosing weight_reduction:

	M0	M1	M2	M3	Mv	M1n	M2n	Mvn
midpoint	13	22	22	13	31	673	600	1000 (not converged)
average	13	15	15	13	30	798	727	1000 (not converged)

On HollowCylinder both need 2 iterations for every matrix.

Tests

Run locally with the submodule branch: the MassMatrixPreconditioner tests (test_preconditioner_transpose.py, test_mass_matrices.py::test_mass_preconditioner*, the new test_reduced_weight_1d) pass serially and on 2 ranks (24 passed each). The rest is left to CI.

Don't raise the feectools pin in pyproject.toml until 0.6.0 is on PyPI. Before merging, the submodule pointer should move to the merged devel-tiny commit.

🤖 Generated with Claude Code