Release notes#
v0.11.0#
Since the 0.10.0 release, there has been 177 merged pull requests from 27 contributors. Below follows a summary of the biggest changes to the Python API. A full diff can be found here.
SuperLU_DIST interface#
Authors: Jack Hale and Chris Richardson
After the support for Windows was added in v0.9.0, there has been a lack of parallel supported solvers to
solve the resulting linear problems. Furthermore, PETSc sometimes can feel complicated for simple problems presented
during teaching. In this release, we add support for using SuperLU_DIST with native sparse matrices
dolfinx.la.MatrixCSR.
The following constructors and classes have been added:
dolfinx.la.superlu_dist.superlu_dist_matrix(): Deep-copy all data fromdolfinx.la.MatrixCSRto SUPERLU_Dist matrix.dolfinx.la.superlu_dist.superlu_dist_solver(): Create adolfinx.la.superlu_dist.SuperLUDistSolverwhich you can use to callset_option,set_Aorsolve.dolfinx.fem.problems.LinearProblem: An interface similar to thedolfinx.fem.petsc.LinearProblem, i.e. it takes inufl.Formfor the bilinear and linear forms, along with appropriateDirichlet boundary conditions, solver options, andentity_maps.
Built-in matrix support#
Authors: Chris Richardson
Adds
A.transpose(),A.mult(x, y, transpose=True)andA.matmul(B)to the built- in matricesTemplated matrices in the Python API for block size
[i, i], i=1,2,3.Improved tests for square and rectangular matrices
The ‘real’ element#
Authors: Jørgen S. Dokken and Matthew Scroggs
The real-element has for a long time not been present in DOLFINx. This has mainly been due to the fact that the implementation in legacy FEniCS introduced a ton of special casing within core functionality, which we thought was better to avoid in DOLFINx. However, a prototype implementation of the real-element has been around for a few releases, and has now been implemented in the core libraries. Users can now call
import basix.ufl
import dolfinx.fem
el = basix.ufl.real_element(mesh.basix_cell(), dtype=dtype, shape=(N, ))
R = dolfinx.fem.functionspace(mesh, el)
to create a function space consisting of N values (of data type dtype, which can be a complex type).
Furthermore, to use this space alongside other spaces, for instance for Lagrange multipliers, users are
recommended to use ufl.MixedFunctionSpace(V, R, ...) and
ufl.extract_blocks() to create blocked systems that can be used in dolfinx.fem.petsc.LinearProblem
or dolfinx.fem.petsc.NonlinearProblem.
Threading#
Authors: Chris Richardson, Jørgen S. Dokken and Garth N. Wells
For a long time, DOLFINx has been exclusively using MPI for the distribution of computational load. However, with the computational landscape evolving to more and more heterogeneous systems, the need for additional parallelisation methods are required. In this release, we introduce initial threading support using std::jthread` in the following methods:
dolfinx.geometry.compute_distances_gjk()which both take an optional argumentnum_threadswhich specifies how many CPU threads should be used. If set to 0, threads are not spawned.
New and improved demos#
Authors: Jørgen S. Dokken, Paul T. Kühner
A new demo showcasing how to use PETSc and matrix-free solvers can be found in demo_matrix-free-petsc
The PML demo now shows how to use one-sided interior facet integrals with manual specification of integration entities.
The Biharmonic demo has gone through a major revision, using:
A more suitable choice of finite elements (as P2 yields sub-optimal convergence)
Better choice of penalty parameter
Change of boundary conditions from simply supported to clamped and explaining the effect of different BCs.
Verify solution with the method of manufactured solutions and add relevant references
In general, demos now consistently use tdim=mesh.topology.dim, fdim = tdim -1 and gdim = mesh.geometry.dim to
avoid confusion for new users.
Further improvements in submesh support#
Authors: Jørgen S. Dokken
A feature that for a long time has existed outside of the FEniCS core is the
dolfinx.mesh.transfer_meshtags_to_submesh(), which makes it possible to transfer a
meshtag from a parent mesh to a submesh. This function is now part of the core library.
Another feature introduced in this release is the usage of submesh quantities in dolfinx.fem.Expression.
You can now pass Expressions containing coefficients and constants from a submesh,
combined with geometric quantities, coefficients, and constants of the parent mesh.
For example
V = fem.functionspace(mesh, el)
u = fem.Function(V, dtype=dtype)
# Populate `u` ...
mesh.topology.create_connectivity(mesh.topology.dim - 1, mesh.topology.dim)
exterior_facets = exterior_facet_indices(mesh.topology)
submesh, entity_map, _, _ = create_submesh(mesh, mesh.topology.dim - 1, exterior_facets)
u_sub = fem.Function(fem.functionspace(submesh, sub_el), dtype=dtype)
# Populate `u_sub` ...
quadrature_points, _ = basix.make_quadrature(basix.CellType.interval, qdegree)
quadrature_points = quadrature_points.astype(xtype)
n_h = ufl.FacetNormal(mesh)
f = u * n + u_sub * n
mesh.topology.create_connectivity(mesh.topology.dim - 1, mesh.topology.dim)
compiled_expr = fem.Expression(expr, quadrature_points, dtype=dtype, entity_maps=[entity_map])
Extending GMSH and VTKHDF IO#
Authors: Jørgen S. Dokken
With the changes to mesh in v0.10.0 max_facet_to_cell_links was introduced to make
it possible to create meshes with joints, branches, etc.
This is now exposed in dolfinx.io.gmsh.model_to_mesh().
Furthermore, new cell types are supported for the vtkhdf backend, including all
linear and quadratic VTK cell types.
The GMSH interface can now read in meshes on mixed-topology grids using dolfinx.io.gmsh.model_to_mesh()
or dolfinx.io.gmsh.read_from_msh(). There is currently no support for
reading entity tags on mixed-topology grids through this interface yet.
Exposing tolerances for non-affine pull-backs#
Authors: Jørgen S. Dokken
A set of users have had issues with non-affine geometries, in particular higher order grids,
and the tolerance and maximum number of iterations in the
dolfinx.fem.CoordinateElement.pull_back(), dolfinx.fem.Function.interpolate_nonmatching()
and dolfinx.fem.Function.eval() yielding errors such as:
RuntimeError: Newton method failed to converge for non-affine geometry
If you see this error message, try to increase the maxiter or tol.
Mixed topology meshes, prisms and pyramid cells#
Authors: Jørgen S. Dokken and Chris Richardson
For mixed topology meshes, there are many notions that do not align with the original design of DOLFINx.
Examples are the members num_entity_dofs and num_entity_closure_dofs of dolfinx.cpp.fem.ElementDofLayout,
as prisms and pyramids do not have the same number of dofs per sub-entity.
They have been removed, and users should instead call
len(ElementDofLayout.entity_dofs(dim, entity_index))
Furthermore, dolfinx.fem.apply_lifting() and dolfinx.fem.assemble_scalar() now work for mixed-topology meshes.
Meshes#
Authors: Jørgen S. Dokken and Jack Hale
New function
dolfinx.mesh.create_point_mesh()to create a point cloud mesh with no points shared between the different processes. Useful for reading in point measures or outputting data.Built in mesh-generators such as
dolfinx.mesh.create_rectangle()now takes an optional argumentgdimthat embeds in a larger space. This simplifies the testing process for problems on manifolds.New function
dolfinx.fem.interpolate_geometry()that allows users to create a new mesh either raising or lowering the polynomial order of the mesh geometry. The topology is shared between the old and new grid. It is also possible to switch the Lagrange-variant of the underlying coordinate element.
Interpolation#
Authors: Jørgen S. Dokken and Garth N. Wells
A crucial bug interpolating Piola-mapped elements from parent to a codim-0 submesh has been fixed for this release.
Modernizing UFL#
Authors: Paul T. Kühner
UFL is a Python project that has been in development for almost 20 years, and
Python has gone through a massive modernization during this time.
One of the visually pleasing improvements is the use of @property-decorators.
ufl.AbstractCell now uses properties for
topological_dimension and
cellname, etc. while
ufl.Mesh now has geometric_dimension.
See UFL PR #385 for more details.
v0.10.0#
Since the 0.9.0 release, there have been 311 merged pull requests from 25 contributors. Below follows a summary of the biggest changes to the Python-API from these pull requests. In addition to the changes below, the ever-lasting quest of improving performance and squashing bugs continues.
PETSc API#
Authors: Jørgen S. Dokken, Francesco Ballarin, Jack Hale and Garth N. Wells
Mapping data between PETSc.Vec and dolfinx.fem.Functions is now
trivial for blocked problems by using dolfinx.fem.petsc.assign().
Both solvers and assembly routines interfacing with PETSc has received a drastic make-over to improve usability and maintenance, both for developers and end-users.
Improved non-linear (Newton) solver#
The FEniCS project has for the last 15 years had its own implementation of a Newton solver.
We no longer see the need of providing this solver, as the PETSc SNES solver,
and equivalent solver for C++ provides more features than our own implementation.
The previously shipped dolfinx.nls.petsc.NewtonSolver is deprecated, in favor of
dolfinx.fem.petsc.NonlinearProblem, which now integrates directly with petsc4py.PETSc.SNES.
The non-linear problem object that was sent into dolfinx.nls.petsc.NewtonSolver has been renamed
to NewtonSolverNonlinearProblem and is also deprecated.
The new NonlinearProblem has additional support for blocked systems,
such as NEST by supplying kind="nest" to its initializer. See the documentation for further
information.
Improved dolfinx.fem.petsc.LinearProblem#
The
dolfinx.fem.petsc.LinearProblemnow supports blocked problems, either specified manually or by usingufl.extract_blocks(). By changing the input-argumentkind, the user can now decide if they want to use the DOLFINx blocked PETSc implementation (kind="mpi") or thekind="nest".The default behavior for non-blocked systems remains the same as before.
The users can now also specify a (blocked) form for preconditioning through the
Pkeyword argument in the constructor.
Assembly routines#
In earlier versions of DOLFINx, there were three assembly routines for PETSc.Vec and PETSc.Mat:
assemble_*assemble_*_blockassemble_*_nest
This caused a lot of duplicate logic in codes.
Therefore, we have unified all these assembly routines under dolfinx.fem.petsc.assemble_vector() and dolfinx.fem.petsc.assemble_matrix().
The input keyword argument kind selects the relevant assembler routine.
See for instance the Stokes demo for a detailed introduction.
Similar changes has been done to dolfinx.fem.petsc.apply_lifting().
Linear algebra submodule#
There is now a sub-module (dolfinx.la.petsc) containing PETSc LA operations.
Interpolation#
The dolfinx.fem.discrete_curl() operator has been added to DOLFINx, to cater to
Hypre Auxiliary-space Divergence Solver
A
petsc4py.PETSc.Matequivalent can be found underdolfinx.fem.petsc.discrete_curl().
Simplified demos#
Usage of ufl.MixedFunctionSpace and ufl.extract_blocks()#
Authors: Jørgen S. Dokken and Joe Dean
Initially introduced as part of the v0.9.0-release, usage of these two UFL-abstractions has been propagated into the demos, to make it even easier for users to see examples of how to work with blocked problems.
TODO: Add profiling of blocked/mixed-element vs mixedfunction-space.
Usage of ufl.ZeroBaseForm#
Author: Garth N. Wells
For a long time, it has not been possible to specify the right hand side of a linear PDE as empty.
This means that users often have had to resolve to adding dolfinx.fem.Constant(mesh, 0.0)*v*ufl.dx
to ensure that one can use the dolfinx form compilation functions.
With the introduction of ufl.ZeroBaseForm this is no longer required.
The aforementioned workaround can now be reduced to ufl.ZeroBaseForm((v, )), which avoid extra
assembly calls within DOLFINx.
Form compiler and integral types#
Author: Susanne Claus, Paul T. Kühner, and Jørgen S. Dokken
The tabulation kernels now have an extra input, a
void*, to make it possible to pass custom data for custom kernels.New
dolfinx.fem.IntegralTypesupportVertex integrals:
ufl.dPRidge integrals (codim=2);
uf.dr
One can now assemble the diagonal of a bilinear form into a vector by adding
form_compiler_options={"part":"diagonal"}when callingdolfinx.fem.form(). Instead of callingdolfinx.fem.petsc.assemble_matrix()one should now calldolfinx.fem.petsc.assemble_vector(). This is useful for matrix-free solvers with Jacobi smoothing.
Mesh#
Authors: Paul T. Kühner, Joe Dean, Garth N. Wells, Jørgen S. Dokken and Chris Richardson
Uniform mesh refinement of all
CellTypesis available throughdolfinx.mesh.uniform_refine().Branching meshes (a mesh where a single facet is connected to more than two cells), such as T-joints (3 cells connected to a single facet) are now supported as input meshes to DOLFINx. To ensure proper partitioning in parallel, one should change the default option
max_facet_to_cell_linksto how many cells a facet can be attached to indolfinx.io.XDMFFile.read_mesh(),dolfinx.io.vtkhd.read_mesh()anddolfinx.mesh.create_mesh().One can no longer use
set_connectivityorset_index_mapto modifydolfinx.mesh.Topologyobjects. Any connectivity that is not(tdim, 0), (tdim,tdim) or(0, 0)should be created withdolfinx.mesh.Topology.create_connectivity(). The aforementioned connections should be attached to the topology when callingdolfinx.cpp.mesh.create_topology().Mixed-dimensional support has been vastly improved by creating
dolfinx.mesh.EntityMap, which replaces the numpy arrays used asentity_mapsindolfinx.fem.form()in the previous release. This is a two-way map, meaning that the user no longer has to take care of creating the correct mapping. The two-way map from a sub-mesh to a parent mesh is returned as part ofdolfinx.mesh.create_submesh().
Linear Algebra#
Authors: Chris Richardson
The native matrix-format now has a sparse matrix-vector multiplication
dolfinx.la.MatrixCSR.mult(). Note that the dolfinx.la.Vector that you multiply with should use the
dolfinx.la.MatrixCSR.index_map(1) rather than the one stemming from the
dolfinx.fem.FunctionSpace.dofmap.index_map.
Collision detection#
Author: Chris Richardson
The collision detection algorithm dolfinx.geometry.compute_distance_gjk() now used multiprecision to ensure
proper collision detection. The algorithm has also been improve to work on co-planar convex hulls.
Documentation#
Authors: Paul T. Kühner, Garth N. Wells, Mehdi Slimani and Jørgen S. Dokken
Several classes that have only been exposed through the nanobind/C++interface have gotten proper Python classes and functions. This includes:
Tons of typos, formatting fixes and improvements have been made.
Usage of intersphinx and sphinx-codeautolink to make the documentation more interactive. Most classes, functions and methods in any demo on the webpage can now redirect you to the relevant package API.
Revised timer logic#
Authors: Garth N. Wells and Paul T. Kühner
Instead of using the Boost timer library,
we have opted for the standard timing library std::chrono.
The switch is mainly due to some observed inaccuracies in timings with Boost.
This removes the notion of wall, system and user time.
See or Timer for examples of usage.
IO#
GMSH#
Authors: Paul T. Kühner, Jørgen S. Dokken, Henrik N.T. Finsberg and Pierric Mora
The GMSH interface to DOLFINx has received a major upgrade.
An API-breaking change is that the module
dolfinx.io.gmshiohas been renamed todolfinx.io.gmsh.Another API-breaking change is the return type of
dolfinx.io.gmshio.model_to_mesh()anddolfinx.io.read_from_msh(). Instead of returning thedolfinx.mesh.Mesh, cell and facetdolfinx.mesh.MeshTags, it now returns adolfinx.io.gmsh.MeshDatadata-class, that can containdolfinx.mesh.MeshTagsof an sub-entity:Cell (codim 0)
Facet (codim 1)
Ridge (codim 2)
Peak (codim 3)
Additional checks and error handing for
Physicaltags from GMSH has been added to improve the user experience.
VTKHDF5#
Authors: Chris Richardson and Jørgen S. Dokken
As Kitware has stated that VTKHDF is the future format they want to support, we have started the transition to this format. Currently, the following features have been implemented:
Reading meshes:
dolfinx.io.vtkhdf.read_mesh(). Supports mixed topology.Writing meshes:
dolfinx.io.vtkhdf.write_mesh(). Supports mixed topology.Writing point data
dolfinx.io.vtkhdf.write_point_data(). The point data should have the same ordering as the geometry nodes of the mesh.Writing cell data
dolfinx.io.vtkhdf.write_cell_data().
VTXWriter#
Authors: Mehdi Slimani and Jørgen S. Dokken
The writer does now support time-dependent DG-0 data, which can be written in the same file as a set of functions from another (unique) function space.
XDMF#
Author: Massimiliano Leoni and Paul T. Kühner
When using
dolfinx.io.XDMFFIle.read_meshtags()one can now specify the attribute name, if the grid has multiple tags assigned to it.Flushing data to file is now possible with
dolfinx.io.XDMFFile.flush(). This is useful when wanting to visualize long-running jobs in Paraview.
Remove Fides backend#
As we unfortunately haven’t seen an expanding set of features for the Fides Reader in Paraview, we have decided to remove it from DOLFINx.
Pyvista#
Pyvista no longer requires pyvista.start_xvfb() if one has installed vtk with OSMesa support.