Fixes for BDDC by pbrubeck · Pull Request #5152 · firedrakeproject/firedrake

pbrubeck · 2026-06-06T12:22:30Z

Description

Internal fixes required for BDDCPC to work on primal elasticity problems.

Adds new options -pc_bddc_use_divergence_mat and -pc_bddc_use_discrete_gradient. Their defaults are the correct ones for the Riesz map in the de Rham complex, H(curl) needs the gradient, while H(div) needs the divergence. The H1 Riesz map does not need an extra operator. However, elasiticity problems are posed on H1, but BDDC would need to be aware of the divergence, so in this case the user must prescribe it.

default values for options should guarantee robustness wrt H1,Hdiv and Hcurl problems

…rake into pbrubeck/fix/bddc

stefanozampini · 2026-06-15T22:31:49Z

@@ -34,9 +34,12 @@ class BDDCPC(PCBase):
    the options:
    - ``'bddc_cellwise'`` to set up a MatIS on cellwise subdomains if P.type == python,
    - ``'bddc_matfree'`` to set up a matrix-free MatIS if A.type == python,


Can you do matrix-free MatIS? Then the divergence matrix can be passed that way, I only need the action of its local part (in MATIS sense) on a given vector

Yeah, we can

I think we need to fix support for matfree A in BDDC

pbrubeck · 2026-06-24T08:32:37Z

@connorjward BDDC is younger than Submesh. Can this go into release? It does not break API.

pbrubeck · 2026-06-24T08:33:59Z

+    num_entities = len(active_entities)
+
+    kernel_code = f"""
+    void compute_entity_coordinates(PetscScalar *coords, PetscScalar *cg1_coords) {{


Sorry about the parloop. There is no better way of computing entity coordinates on extruded meshes.

It's not an issue. Moving this to pyop3 is easy.

pbrubeck · 2026-06-24T08:35:15Z

+    cg1_offsets, cg1_flatmap = flatten_entity_ids(active_entities, cg1_closure_ids)
+    v_offsets, v_flatmap = flatten_entity_ids(active_entities, V_entity_ids)


These data structures wouldn't be required if we use an op2.PermutedMap

Why not use one then?

It only occured to me just now. Only v_flatmap is permutation, cg1_flatmap will have repeated indices because it defines the dofs supported on the closure of an entity (and edge will include its vertices), so this case is not covered by a permutation. We would still need to flatten_entity_ids at least for the CG1 element, so it is better to keep the code as is.

Ideally we should be running a parloop on the mesh entities. I know, wait for pyop3...

connorjward

Seems more or less fine. I think release is OK because this is a sufficiently new feature and you're only touching code in one file.

connorjward · 2026-06-24T09:10:23Z

+    # On MacOSX the distributed right-hand side is bugged!
+    if DEFAULT_DIRECT_SOLVER == "mumps":


This comment does not explain the code changes. There's nothing here that's obviously specific to macOS

@stefanozampini did you report this?

there's no way this can be reported officially to MUMPS people. It is a known issue on only some osx configurations

The comment would lead me to expect a check like: if os.system == "darwin" or similar.

connorjward · 2026-06-24T09:10:32Z

+        # "ksp_view": None,
+        # "ksp_monitor_singular_value": None,


Suggested change

# "ksp_view": None,

# "ksp_monitor_singular_value": None,

Why? it is a quick way to check if something goes wrong instead of having to retype the options

connorjward · 2026-06-24T09:14:19Z

+    num_entities = len(active_entities)
+
+    kernel_code = f"""
+    void compute_entity_coordinates(PetscScalar *coords, PetscScalar *cg1_coords) {{


It's not an issue. Moving this to pyop3 is easy.

pbrubeck · 2026-06-24T11:08:37Z

+    # On MacOSX the distributed right-hand side is bugged!
+    if DEFAULT_DIRECT_SOLVER == "mumps":
+        sp.update({"bddc_pc_bddc_coarse_mat_mumps_icntl_20": 0})
+


I think we can simply remove this from here

Suggested change

# On MacOSX the distributed right-hand side is bugged!

if DEFAULT_DIRECT_SOLVER == "mumps":

sp.update({"bddc_pc_bddc_coarse_mat_mumps_icntl_20": 0})

But shouldn't this be the default?

What is the problem of leaving the option in? A test should just work, and this is a known MUMPS bug

But shouldn't this be the default?

No, the default is distributed right-hand side, which means no need to gather the rhs on the master process

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

Fixes for BDDC

2f8b06c

pbrubeck force-pushed the pbrubeck/fix/bddc branch from e11e7da to 2f8b06c Compare June 6, 2026 12:25

pbrubeck added 3 commits June 6, 2026 13:40

WIP

16bcafc

Fixes

f6f16f2

fixes

3517cfe

pbrubeck commented Jun 8, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py

pbrubeck commented Jun 8, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py Outdated

add elasticity test

0b85502

pbrubeck commented Jun 8, 2026

View reviewed changes

Comment thread tests/firedrake/regression/test_bddc.py

pbrubeck added 5 commits June 8, 2026 10:17

add failing CG test

b941df1

only test cellwise=False

e27dc90

add use_divergence_mat and use_discrete_gradient options

232e249

adaptive options

d0fb4fd

Native divergence mat assembly

1bc6e36

pbrubeck force-pushed the pbrubeck/fix/bddc branch from 804646e to 1bc6e36 Compare June 8, 2026 10:24

fixes

b7313a6

stefanozampini reviewed Jun 8, 2026

View reviewed changes

Comment thread tests/firedrake/regression/test_bddc.py Outdated

Comment thread tests/firedrake/regression/test_bddc.py

stefanozampini reviewed Jun 14, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py Outdated

stefanozampini reviewed Jun 14, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py Outdated

stefanozampini reviewed Jun 14, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py Outdated

stefanozampini added 3 commits June 15, 2026 12:44

BDDC: use PETSc options API to get the correct boolean

4cb9c61

default values for options should guarantee robustness wrt H1,Hdiv and Hcurl problems

BDDC: fix tests

6af02e8

Fix last test

ae87da4

pbrubeck commented Jun 15, 2026

View reviewed changes

Comment thread tests/firedrake/regression/test_bddc.py Outdated

tweak interface code and tests to use defaults

2033f23

pbrubeck commented Jun 15, 2026

View reviewed changes

Comment thread tests/firedrake/regression/test_bddc.py Outdated

stefanozampini and others added 4 commits June 15, 2026 14:55

use deluxe scaling only for adaptive coarse spaces

54af535

WIP

0483345

Use pyop2 in get_entity_coordinates

7ac70d2

Merge branch 'pbrubeck/fix/bddc' of github.com:firedrakeproject/fired…

662da83

…rake into pbrubeck/fix/bddc

pbrubeck added 2 commits June 15, 2026 22:22

cleanup

66dd927

fixes

904ecf5

stefanozampini reviewed Jun 15, 2026

View reviewed changes

pbrubeck requested a review from connorjward June 24, 2026 08:32

pbrubeck commented Jun 24, 2026

View reviewed changes

connorjward requested changes Jun 24, 2026

View reviewed changes

pbrubeck commented Jun 24, 2026

View reviewed changes

Comment thread firedrake/preconditioners/bddc.py

pbrubeck commented Jun 24, 2026

View reviewed changes

Update firedrake/preconditioners/bddc.py

afa652d

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

		cg1_offsets, cg1_flatmap = flatten_entity_ids(active_entities, cg1_closure_ids)
		v_offsets, v_flatmap = flatten_entity_ids(active_entities, V_entity_ids)

		# On MacOSX the distributed right-hand side is bugged!
		if DEFAULT_DIRECT_SOLVER == "mumps":

	# On MacOSX the distributed right-hand side is bugged!
	if DEFAULT_DIRECT_SOLVER == "mumps":
	sp.update({"bddc_pc_bddc_coarse_mat_mumps_icntl_20": 0})

Uh oh!

Conversation

pbrubeck commented Jun 6, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbrubeck Jun 15, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbrubeck commented Jun 24, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbrubeck Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pbrubeck Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

connorjward left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

pbrubeck commented Jun 6, 2026 •

edited

Loading

pbrubeck Jun 15, 2026 •

edited

Loading

pbrubeck Jun 24, 2026 •

edited

Loading

pbrubeck Jun 24, 2026 •

edited

Loading