Rework PMGBase.coarsen() and add NonlinearVariationalProblem.rediscretise() by pbrubeck · Pull Request #5324 · firedrakeproject/firedrake

pbrubeck · 2026-08-04T14:32:59Z

Summary

Gives NonlinearVariationalProblem and the nullspace
basis classes a generic reediscretise() method, and reworks
PMGBase.coarsen() to build the coarse problem, _SNESContext, and
nullspaces through those reconstruct()/rediscretise() methods instead of hand-rolling
each piece and special-casing bc coarsening per PMGPC/PMGSNES.

Also replaces PMGBase.reconstruct_degree's element-degree-shift
tree-walk with finat.ufl's own native FiniteElement.reconstruct(degree=...)
support (firedrakeproject/fiat#273).

Before / after

Before: PMGBase.coarsen() hand-built the coarse NonlinearVariationalProblem
and _SNESContext, and coarsened bcs via bespoke coarsen_bcs/
coarsen_bc_value methods (with a separate per-bc cache) that special-cased
PMGPC (bc value always 0) vs PMGSNES (eager Function.interpolate()).
After: NonlinearVariationalProblem.redisctretise() and _SNESContext.reconstruct()
do this generically for any caller, so PMGPC and PMGSNES coarsen bcs (and F)
identically, and the PMGPC-only cproblem.F = action(J, cu) simplification is replaced with
ZeroBaseForm.
Before: element degree coarsening/refinement went through
PMGBase.reconstruct_degree's manual tree-walk over
EnrichedElement/TensorProductElement/MixedElement/etc.
After: callers use the element's own reconstruct(degree=...).

…method NonlinearVariationalProblem.reconstruct(function_space) rebuilds the problem's solution, forms, and bcs on a new function space living on the same mesh, substituting the coefficient mapping symbolically through an optional form_transform hook (defaulting to ufl.replace). VectorSpaceBasis.reconstruct and MixedVectorSpaceBasis.reconstruct do the analogous thing for a nullspace basis, interpolating each basis Function onto the new space and re-orthonormalizing. Neither caller is wired in yet; PMGBase.coarsen() picks these up next.

…ct() coarsen() used to hand-roll the coarse problem (coarsening J/Jp/F/bcs/ appctx itself, then constructing a fresh NonlinearVariationalProblem and _SNESContext) and separately special-case bc coarsening per PMGPC vs PMGSNES via coarsen_bcs/coarsen_bc_value and a bespoke per-bc cache. None of that is necessary now: a bc's value is just a UFL expression, and NonlinearVariationalProblem.reconstruct() already builds the coefficient_mapping needed to substitute it symbolically, the same way it already does for forms, so PMGPC and PMGSNES coarsen F, J, Jp, and bcs identically. _SNESContext gets the analogous treatment via its own reconstruct(), and nullspace coarsening goes through the new VectorSpaceBasis/MixedVectorSpaceBasis.reconstruct() instead of a hand-rolled interpolation + cache. Also drop the PMGPC-only "cproblem.F = action(J, cu)" workaround now that geometric multigrid supports non-Lagrange elements (the mixed-block bug it worked around no longer applies), so PMGPC and PMGSNES coarsen F identically too. Drop PMGBase.reconstruct_degree's element-degree-shift tree-walk; its callers move to the element's own reconstruct(degree=...) next.

…econstruct_degree finat.ufl's FiniteElement subclasses now support reconstruct(degree=...) natively (firedrakeproject/fiat#273), including the element-degree-shift tree-walk through EnrichedElement/TensorProductElement/MixedElement/etc. that PMGBase.reconstruct_degree used to carry. P1PC/P1SNES/LORPC. coarsen_element and the p-multigrid tests call the element's own reconstruct() directly. test_p_multigrid_mixed's iteration count moves from <=7 to <=9: PMGPC's coarse-grid Jacobian is no longer forced through action(J, cu) (see previous commit), which slightly changes the coarse operator for this problem.

@main

Pin firedrake-fiat to pbrubeck/element-reconstruct-degree (firedrakeproject/fiat#273) so CI exercises the composite-element reconstruct(degree=...) fix this branch depends on. Revert to @main once that fiat PR merges. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Covers PMGBase.coarsen_element/coarsen/coarsen_quadrature, NonlinearVariationalProblem.reconstruct, VectorSpaceBasis.reconstruct, MixedVectorSpaceBasis.reconstruct, and _SNESContext.reconstruct. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

ufl.replace can't move a Cofunction's implicit test argument to the coarse space (it isn't a literal node in the expression tree), so coarsening F = action(J, u) - L for a LinearVariationalProblem left L on the fine space, giving F two distinct arguments and tripping the "residual is not a linear form" check. PC-only coarse levels never assemble their own residual, so give them F = action(cJ, cu) instead of coarsening fproblem.F. SNES/FAS levels do need a real residual: fall back to only replacing coefficients and re-pairing the result with the coarse test function via interpolation, mirroring the DirichletBC-restriction code path. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

action(cJ, cu) worked but redundantly coarsened J a second time (reconstruct() already does it for J itself). Use a bare ZeroBaseForm with the coarse test argument instead: cheaper, and the recently-fixed ZeroBaseForm._ufl_expr_reconstruct_ (upstream UFL) makes it safe to carry through reconstruction. This needs two supporting fixes: - reconstruct_nlvp in mg/ufl_utils.py looked up the coarsened solution via coefficient_mapping[problem.u_restrict], relying on it having been populated as a side effect of walking F for coefficients. A coefficient-free F doesn't trigger that, so look it up explicitly instead. - test_p_fas_nonlinear_scalar's check_coarsen_quadrature inspected p.F.integrals() directly; skip non-Form entries since PC-only levels no longer carry a real Form there (J alone is sufficient to verify the quadrature-degree coarsening). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

@main

Pin fenics-ufl to pbrubeck/zero-base-form-holes (FEniCS/ufl#494) so CI exercises the ZeroBaseForm.extract_domains and _ufl_expr_reconstruct_ fixes PMG coarsening now depends on. Revert to @main once that ufl PR merges. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Co-authored-by: Pablo Brubeck <brubeck@protonmail.com>

Covers PMGBase.coarsen_element/coarsen/coarsen_quadrature, NonlinearVariationalProblem.reconstruct, VectorSpaceBasis.reconstruct, MixedVectorSpaceBasis.reconstruct, and _SNESContext.reconstruct. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

ufl.replace can't move a Cofunction's implicit test argument to the coarse space (it isn't a literal node in the expression tree), so coarsening F = action(J, u) - L for a LinearVariationalProblem left L on the fine space, giving F two distinct arguments and tripping the "residual is not a linear form" check. PC-only coarse levels never assemble their own residual, so give them F = action(cJ, cu) instead of coarsening fproblem.F. SNES/FAS levels do need a real residual: fall back to only replacing coefficients and re-pairing the result with the coarse test function via interpolation, mirroring the DirichletBC-restriction code path. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

action(cJ, cu) worked but redundantly coarsened J a second time (reconstruct() already does it for J itself). Use a bare ZeroBaseForm with the coarse test argument instead: cheaper, and the recently-fixed ZeroBaseForm._ufl_expr_reconstruct_ (upstream UFL) makes it safe to carry through reconstruction. This needs two supporting fixes: - reconstruct_nlvp in mg/ufl_utils.py looked up the coarsened solution via coefficient_mapping[problem.u_restrict], relying on it having been populated as a side effect of walking F for coefficients. A coefficient-free F doesn't trigger that, so look it up explicitly instead. - test_p_fas_nonlinear_scalar's check_coarsen_quadrature inspected p.F.integrals() directly; skip non-Form entries since PC-only levels no longer carry a real Form there (J alone is sufficient to verify the quadrature-degree coarsening). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

@main

Pin firedrake-fiat to pbrubeck/element-reconstruct-degree (firedrakeproject/fiat#273) so CI exercises the composite-element reconstruct(degree=...) fix this branch depends on. Revert to @main once that fiat PR merges. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

@main

Pin fenics-ufl to pbrubeck/zero-base-form-holes (FEniCS/ufl#494) so CI exercises the ZeroBaseForm.extract_domains and _ufl_expr_reconstruct_ fixes PMG coarsening now depends on. Revert to @main once that ufl PR merges. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

Drop the fiat/ufl pins that leaked back in through the merge; they belong in the DROP BEFORE MERGE action.yml commits only. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

connorjward

Close

Co-authored-by: Pablo Brubeck <brubeck@protonmail.com>

pbrubeck · 2026-08-06T09:01:33Z

CI is green here

…tise() (#5324) * Give NonlinearVariationalProblem and nullspace bases a rediscretise() method * Rework PMGBase.coarsen() to use NonlinearVariationalProblem.rediscretise()

pbrubeck and others added 4 commits August 4, 2026 15:25

pbrubeck mentioned this pull request Aug 4, 2026

Automated goal-oriented adaptivity #5273

Open

pbrubeck requested a review from connorjward August 4, 2026 14:34

connorjward reviewed Aug 4, 2026

View reviewed changes

Comment thread tests/firedrake/multigrid/test_p_multigrid.py Outdated

Comment thread pyproject.toml Outdated

Comment thread firedrake/preconditioners/pmg.py Outdated

Comment thread firedrake/preconditioners/pmg.py

Comment thread firedrake/variational_solver.py

review suggestions

0b17ec3

pbrubeck commented Aug 4, 2026

View reviewed changes

Comment thread firedrake/solving_utils.py Outdated

pbrubeck and others added 5 commits August 4, 2026 17:00

Apply suggestion from @pbrubeck

0ffce79

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/preconditioners/pmg.py Outdated

Apply suggestion from @pbrubeck

696ad5f

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/preconditioners/pmg.py Outdated

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/preconditioners/pmg.py

pbrubeck and others added 10 commits August 5, 2026 08:28

Apply suggestions from code review

11f0adc

Co-authored-by: Pablo Brubeck <brubeck@protonmail.com>

review suggestions

f7ff628

Apply suggestion from @pbrubeck

2c811e8

drop comment

11eb4a2

Restore pyproject.toml to main

b23fe02

Drop the fiat/ufl pins that leaked back in through the merge; they belong in the DROP BEFORE MERGE action.yml commits only. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread .github/actions/install/action.yml Outdated

drop function_space kwarg

e07c9b4

pbrubeck force-pushed the pbrubeck/pmg-reconstruct-rework branch from a820726 to e07c9b4 Compare August 5, 2026 10:31

pbrubeck added 2 commits August 5, 2026 11:48

Merge branch 'main' into pbrubeck/pmg-reconstruct-rework

2c26e19

Apply suggestion from @pbrubeck

16fd8a1

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/mg/ufl_utils.py

pbrubeck added 3 commits August 5, 2026 12:19

Apply suggestion from @pbrubeck

4323c50

lint

ed1a955

rediscretise

880fc8c

connorjward requested changes Aug 5, 2026

View reviewed changes

Comment thread firedrake/nullspace.py Outdated

Comment thread firedrake/variational_solver.py Outdated

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/nullspace.py Outdated

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/nullspace.py Outdated

pbrubeck commented Aug 5, 2026

View reviewed changes

Comment thread firedrake/preconditioners/pmg.py Outdated

pbrubeck added 2 commits August 5, 2026 14:40

Apply suggestions from code review

5c3acc0

Co-authored-by: Pablo Brubeck <brubeck@protonmail.com>

fix

7fd3338

pbrubeck force-pushed the pbrubeck/pmg-reconstruct-rework branch 2 times, most recently from adad029 to 2972461 Compare August 5, 2026 15:06

fix

83b5017

pbrubeck force-pushed the pbrubeck/pmg-reconstruct-rework branch from 2972461 to 83b5017 Compare August 5, 2026 16:52

pbrubeck changed the title ~~Rework NonlinearVariationalProblem.reconstruct() and PMGBase.coarsen()~~ Rework PMGBase.coarsen() and add NonlinearVariationalProblem.rediscretise() Aug 6, 2026

pbrubeck requested a review from connorjward August 6, 2026 09:01

connorjward approved these changes Aug 6, 2026

View reviewed changes

pbrubeck merged commit 450aa79 into main Aug 6, 2026
19 of 21 checks passed

pbrubeck deleted the pbrubeck/pmg-reconstruct-rework branch August 6, 2026 09:19

Navigation Menu

Uh oh!

Rework PMGBase.coarsen() and add NonlinearVariationalProblem.rediscretise() - #5324

Rework PMGBase.coarsen() and add NonlinearVariationalProblem.rediscretise()#5324
pbrubeck merged 30 commits into
mainfrom
pbrubeck/pmg-reconstruct-rework

pbrubeck commented Aug 4, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

connorjward left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pbrubeck commented Aug 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

pbrubeck commented Aug 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Before / after

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

connorjward left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pbrubeck commented Aug 6, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

pbrubeck commented Aug 4, 2026 •

edited

Loading