Replace scifem.compat and scifem.ufl_compat with fenicsx-compat by finsberg · Pull Request #267 · scientificcomputing/scifem

finsberg · 2026-10-12T12:19:02Z

All cross-version DOLFINx workarounds now come from fenicsx-compat (>=0.2.0), which centralises them for scifem, io4dolfinx and dolfinx-adjoint. scifem keeps no compat layer of its own:

scifem.compat and scifem.ufl_compat are removed, along with compat in scifem.all and their API reference section. Their callers use the fenicsx_compat equivalents.
Inline duplicates of fenicsx-compat functions are replaced too: facet quadrature permutations, Expression.eval entity shapes, interpolate_f, Expression entity_maps, index map creation and unwrapping, MeshTags.topology, create_sub_index_map, local_range, the gmsh model_to_mesh ghost mode, the ADIOS2 scope and the PETSc lifting helpers.
Branches that cannot run on DOLFINx >= 0.10, the floor fenicsx-compat sets, become plain DOLFINx calls: compute_integration_domains, the cpp Topology constructor, pack_constants/pack_coefficients, determine_point_ownership, the *_block PETSc API, and the 0.8/0.9 paths in point_source.py and solvers.py.
NewtonSolver arranges its boundary conditions with fenicsx_compat.bcs_by_block, as dolfinx.fem.bcs_by_block cannot be called directly across 0.11 and 0.12 (dolfinx#4342).
scifem.petsc.apply_lifting_and_set_bc keeps its single-form calling convention, [form] with [bcs], and passes the form and its boundary conditions on to fenicsx-compat.

The minimum Python version goes from 3.10 to 3.12, as fenicsx-compat requires it, and packaging moves from the runtime dependencies to the test extra, as only the tests still use it.

All cross-version DOLFINx workarounds now come from fenicsx-compat (>=0.2.0), which centralises them for scifem, io4dolfinx and dolfinx-adjoint. scifem keeps no compat layer of its own: - scifem.compat and scifem.ufl_compat are removed, along with `compat` in scifem.__all__ and their API reference section. Their callers use the fenicsx_compat equivalents. - Inline duplicates of fenicsx-compat functions are replaced too: facet quadrature permutations, Expression.eval entity shapes, interpolate_f, Expression entity_maps, index map creation and unwrapping, MeshTags.topology, create_sub_index_map, local_range, the gmsh model_to_mesh ghost mode, the ADIOS2 scope and the PETSc lifting helpers. - Branches that cannot run on DOLFINx >= 0.10, the floor fenicsx-compat sets, become plain DOLFINx calls: compute_integration_domains, the cpp Topology constructor, pack_constants/pack_coefficients, determine_point_ownership, the *_block PETSc API, and the 0.8/0.9 paths in point_source.py and solvers.py. - NewtonSolver arranges its boundary conditions with fenicsx_compat.bcs_by_block, as dolfinx.fem.bcs_by_block cannot be called directly across 0.11 and 0.12 (dolfinx#4342). - scifem.petsc.apply_lifting_and_set_bc keeps its single-form calling convention, `[form]` with `[bcs]`, and passes the form and its boundary conditions on to fenicsx-compat. The minimum Python version goes from 3.10 to 3.12, as fenicsx-compat requires it, and `packaging` moves from the runtime dependencies to the test extra, as only the tests still use it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

jorgensd · 2026-10-13T05:14:49Z

Do we really want the compat package to depend on PETSc? If we do I think we should aim to make scifem PETSc free over time :)

finsberg · 2026-10-13T06:03:17Z

Do we really want the compat package to depend on PETSc? If we do I think we should aim to make scifem PETSc free over time :)

No, it should not depend strictly on PETSc. Right now it is guarded by an if dolfinx.has_petsc4py , but perhaps we should add a test suite that tests this on dolfinx built without PETSc?

jorgensd · 2026-10-21T08:17:08Z

        else:
            pairs = PeriodicNodes(0)
-        mesh_data = model_to_mesh(
+        mesh = model_to_mesh(


How is this correct?

This was because the helper function (model_to_mesh) inside periodic/gmsh.py returned only the mesh, but this is very confusing with two functions with the same name. Fixed in d6e76e9

jorgensd · 2026-10-21T08:19:01Z

+    ghosting_ranks = np.asarray(ghosting_ranks, dtype=np.int32)
+    ghosting_offsets = np.asarray(ghosting_offsets, dtype=np.int64)


Shouldn't this wrapping be done inside fenicsx_compat?

Agree! Fixed in 3156f3f

jorgensd · 2026-10-21T08:20:17Z

+    topology = dolfinx.cpp.mesh.Topology(
        mesh.topology.cell_type,
-        mesh.topology.dim,
-        new_vertex_map,
-        new_cell_map,
-        new_c_to_v,
-        new_v_to_v,
+        unwrap_index_map(new_vertex_map),
+        unwrap_index_map(new_cell_map),
+        new_c_to_v._cpp_object,  # type: ignore[arg-type]


Don't we want a compat topology wrapper here?

I don't know. I think we shouldn't add more compat functions than needed, but if you have strong feelings for adding it we can.

jorgensd · 2026-10-21T08:24:41Z

-    if hasattr(dolfinx.mesh, "transfer_meshtags_to_submesh"):
-        new_et = dolfinx.mesh.transfer_meshtags_to_submesh(
-            entity_tag, submesh, vertex_to_parent=vertex_map, cell_to_parent=cell_map
+    try:


Seems a bit silly to here have a compat function that refers back to the scifem implementation.
Is the reason not for moving it the fact that it requires C++ building, which you have avoided so far in fenicsx_compat?

Totally agree. Perhaps we should just move the whole extract_submesh function to fenicsx-compat, or what do you think?

jorgensd · 2026-10-21T08:26:52Z

-            src_ranks = collision_data.src_owner
-        else:
-            raise NotImplementedError(f"Unsupported version of dolfinx: {dolfinx.__version__}")
+        collision_data = dolfinx.geometry.determine_point_ownership(


This is simplified as we support only 0.10 and ownards?

Yes, that is the reason. Are you OK with dropping support for < 0.10?

jorgensd · 2026-10-21T08:27:44Z

-                self.x = dolfinx.fem.petsc.create_vector(
-                    dolfinx.fem.extract_function_spaces(self._F), kind="mpi"
-                )
+            self.b = dolfinx.fem.petsc.create_vector(


Same here, is this due to the fact that we now only support 0.10 and onwards?

jorgensd · 2026-10-21T08:29:52Z

    submesh_facets = np.arange(num_facets, dtype=np.int32)
    parent_facets = scifem.mesh.get_entity_map(entity_map)[:num_facets]
-    entities = scifem.compat.compute_integration_domains(
+    entities = dolfinx.fem.compute_integration_domains(


Dropped because version support is 0.10?

jorgensd

Some minor things to clarify

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

The wrapper shared its name with dolfinx.io.gmsh.model_to_mesh and fenicsx_compat.model_to_mesh but, unlike them, returned the mesh rather than the MeshData. Since this branch it only forwarded to fenicsx_compat.model_to_mesh, whose defaults (shared_facet ghosting, max_facet_to_cell_links=2) are the same, so read_periodic_mesh_from_msh and the tests call that directly and take `.mesh` themselves. This removes a public function that shipped in v0.26.0. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

for more information, see https://pre-commit.ci

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

for more information, see https://pre-commit.ci

With a typed DOLFINx installed (0.12.0.dev0), mypy reported 96 errors in src and tests; the pre-commit hook runs without DOLFINx, so they never showed there. None changes behaviour: - Wrong annotations are corrected: NewtonSolver's _F and _J hold a list and a list of lists of forms, entity_maps is a sequence of EntityMap from DOLFINx 0.10 on, the _EntityMap protocol's members are read-only properties, and create_space_of_simple_functions' value_shape is a tuple of ints. - typing.cast where the type is known but not to the stubs, chiefly ufl.domain.extract_domains, typed AbstractDomain but a ufl.Mesh here. - A narrow `# type: ignore[<code>]` with its reason where a stub is wrong and the code right: the float32/float64 cpp class picked at runtime, a blocked x0 in apply_lifting, numpy index arrays passed to petsc4py, and compiled forms passed through dolfinx.fem.form. - PeriodicNodes(0) becomes PeriodicNodes(): the 0 went into the `replaced` array, not num_nodes_global. Harmless, as that branch runs only on ranks that never read either. - KSP.create takes the vector's PETSc communicator directly rather than its mpi4py conversion. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

finsberg and others added 2 commits October 12, 2026 12:17

Merge branch 'main' into finsberg/fenicsx-compat

86ff705

finsberg marked this pull request as ready for review October 12, 2026 19:11

finsberg requested a review from jorgensd October 12, 2026 19:11

jorgensd reviewed Oct 21, 2026

View reviewed changes

Comment thread src/scifem/periodic/gmsh.py Outdated

jorgensd reviewed Oct 21, 2026

View reviewed changes

Comment thread src/scifem/periodic/mesh.py Outdated

jorgensd reviewed Oct 21, 2026

View reviewed changes

Comment thread src/scifem/geometry.py Outdated

jorgensd reviewed Oct 21, 2026

View reviewed changes

jorgensd requested changes Oct 21, 2026

View reviewed changes

finsberg and others added 8 commits October 21, 2026 17:25

Update src/scifem/periodic/gmsh.py

377c144

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

Update src/scifem/periodic/mesh.py

826b070

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

[pre-commit.ci] auto fixes from pre-commit.com hooks

e2e2dcb

for more information, see https://pre-commit.ci

Remove array casting of ghosing ranks and offsets

3156f3f

Update src/scifem/geometry.py

cfc2098

Co-authored-by: Jørgen Schartum Dokken <dokken92@gmail.com>

[pre-commit.ci] auto fixes from pre-commit.com hooks

25a45d1

for more information, see https://pre-commit.ci

finsberg added this pull request to stack #270 October 21, 2026 19:38

finsberg removed this pull request from stack #270 October 21, 2026 19:40

		ghosting_ranks = np.asarray(ghosting_ranks, dtype=np.int32)
		ghosting_offsets = np.asarray(ghosting_offsets, dtype=np.int64)

Conversation

finsberg commented Oct 12, 2026

Uh oh!

jorgensd commented Oct 13, 2026

Uh oh!

finsberg commented Oct 13, 2026

Uh oh!

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!

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!

Choose a reason for hiding this comment

Uh oh!

jorgensd left a comment

Choose a reason for hiding this comment

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants