phynexis-deps/mfem
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Files
c577ee0f43500a2fade34e62cc55cacb44be0840
mfem/CHANGELOG
T
Mittal, Ketan 60f47c287d CHANGELOG
2026-02-10 14:28:17 -08:00

3032 lines
143 KiB
Plaintext
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
Finite Element Discretization Library
__
_ __ ___ / _| ___ _ __ ___
| '_ ` _ \ | |_ / _ \| '_ ` _ \
| | | | | || _|| __/| | | | | |
|_| |_| |_||_| \___||_| |_| |_|
https://mfem.org
Version 4.9.1 (development)
===========================
Discretization improvements
---------------------------
- Improved the gridfunction projection routines. Projections work for Scalar,
Vector and VectorFE, also NURBS versions. Optionally different types of
projections can be selected, default behaviour has not changed.
- Added methods to estimate function extremum using piecewise linear bounds +
recursive subdivision.
Meshing improvements
--------------------
- Improved support for 1D NURBS meshes with variable order, including using
the patches construct for 1D NURBS meshes.
New and updated examples and miniapps
-------------------------------------
- Electromagnetics/lorentz miniapp has been updated to leverage the ParticleSet
capability.
Version 4.9, released on Dec 11, 2025
=====================================
Starting with this version, MFEM requires a C++17 compiler.
Discretization improvements
---------------------------
- Introduced dFEM: a new MFEM capability for Automatic Differentiation (AD) of
nonlinear finite element operators, based on Enzyme or dual numbers AD at
quadrature points. These features are part of the new mfem::future namespace
and some of the API can change in the future. See the new dFEM minimal surface
miniapp in the miniapps/dfem/ directory for illustration of dFEM's use. Using
Enzyme for AD in MFEM is tested with clang v19 and requires clang/LLVM built
with plugin support. See INSTALL for more details.
- Introduced initial support for particle methods in MFEM with new classes
Particle, ParticleSet and ParticleVector.
* Particle is a convenient interface for individual particle data.
* ParticleSet manages and stores particle data in a struct-of-arrays form,
carrying particle coordinates and IDs along with an arbitrary number of
Vector and integer data for each particle.
* ParticleVector is a Vector-derived container that stores vector data for an
arbitrary number of particles contiguously based on specified vdim/ordering.
See the new particle miniapps in miniapps/gslib/ and miniapps/fluids/navier/.
- Added a new miniapp and specialized AMG solver (AMGF) for optimization-based
contact mechanics. The miniapp solves large-scale frictionless contact using a
self-contained Interior Point (IP) solver, mortar-based contact constraints
provided by Tribol. The resulting linear systems are solved with the new AMGF
solver (see below). Benchmark examples include the two-block, ironing, and
beam-sphere problems. See the miniapps/contact/ directory.
- Added support for boundary integration to the hyperbolic framework. Two new
classes BdrHyperbolicDirichletIntegrator and BoundaryHyperbolicFlowIntegrator
have been introduced for implementation of weak Dirichlet boundary conditions
with a general flux or for the linear case respectively.
- Added a method to compute piecewise linear bounds on high-order functions on
tensor-product elements.
- Added support for interior face integration enabling DG methods in
ParMixedBilinearForm, ParNonlinearForm and ParBlockNonlinearForm.
- In the ParMoonolith integration, added support for variational resampling of
H1 vector fields.
- Introduced IMEX ODE solvers based on a split-operator framework. Added
examples ex41 and ex41p demonstrating IMEX DG/CG discretizations of the
convection–diffusion equation, with ex41p using DG LOR preconditioning.
Meshing improvements
--------------------
- The TMOP kernel hierarchy has been restructured to reduce compilation time.
Most large kernels have been split into smaller specific kernels for each
metric. The directory structure has been updated with assemble, metrics, mult
and tools subdirectories. New kernel dispatch and specialization system has
also been integrated. Unit tests have been revised to ensure --all tests pass.
- Introduced NC-patch NURBS meshes, which are conforming element-wise but allow
for nonconforming patch topology. This new mesh format supports element
spacing formulas for refinement, as well as local refinement factors for a
subset of knot vectors.
- Added a new miniapp for interpolating a surface grid of points in 3D using a
smooth NURBS surface, that can then be sampled at arbitrary resolution while
staying close to the original geometry. See miniapps/nurbs/nurbs_surface.
- Parallel anisotropic refinement of hexahedral meshes is now supported,
provided that neighboring hexahedra are not refined in conflicting directions.
A new ParMesh method is added to check for such conflicts, before refinement.
- Added support for higher order meshes in (Par)Mesh::MakeSimplicial.
Linear and nonlinear solvers
----------------------------
- Added FilteredSolver: a base class for solvers with filtering. It handles
cases where a solver performs well except in small subspaces, by adding a
filtering step formulated as a subspace correction.
- Added AMGFSolver: a derived class of FilteredSolver, specialized for AMG with
Filtering (AMGF), providing robust preconditioning for linear systems arising
in constrained optimization problems such as frictionless contact.
Added 'GetResiduals' and 'GetFinalAbsResidualNorm' to 'HyprePCG',
'HypreGMRES', and 'HypreFGMRES' to get 'r' and '|r|_p'. Note that the latter
computes '|r|_p' from 'r' instead of returning a cached value like the
relative 'GetFinalResidualNorm'. These require Hypre >= 2.15.0.
Changed the default solver parameters for 'HyprePCG' to 'tol=1e-6' and
'max_iter=1000'. This matches the default parameters in Hypre 3.0.
Added various helper functions for querying/modifying Hypre solvers:
'HypreSmoother::GetType', 'HypreSmoother::GetSOROptions',
'HypreSmoother::GetPolyOptions', 'HypreSmoother::GetWindowParameters',
'HypreSmoother::IsOperatorSymmetric', 'HyprePCG::GetTol',
'HyprePCG::GetAbsTol', 'HyprePCG::GetMaxIter', 'HyprePCG::SetUseTwoNorm',
'HypreGMRES::GetTol', 'HypreGMRES::GetAbsTol', 'HypreGMRES::GetMaxIter',
'HypreGMRES::GetKDim', 'HypreFGMRES::GetTol', 'HypreFGMRES::GetMaxIter',
'HypreFGMRES::GetKDim', and 'HypreBoomerAMG::GetMaxIter'.
GPU computing
-------------
- Added the 'gpu', 'raja-gpu', and 'ceed-gpu' backend aliases/shortcuts which
automatically select between CUDA or HIP.
- Added the option to enable GPU-aware MPI in MFEM using the environment
variable 'MFEM_GPU_AWARE_MPI' set to any value. Setting this environment
variable is an alternative to calling 'Device::SetGPUAwareMPI(true)'.
- Implemented a GPU-accelerated matrix-free AMR derefinement GridFunction update
operator. This supports mixed geometry meshes and variable order spaces, and
is the default derefinement operator constructed by FiniteElementSpace::Update
and ParFiniteElementSpace::Update. The operator requires the finite element
space to be nonconforming.
- Introduced MFEM_FOREACH_THREAD_DIRECT, which directly maps loop tasks to GPU
threads, assigning one task per thread.
- The function Vector::SetSubVector(const Array<int> &, const real_t) now
executes on device if either the vector or the array have the device flag
set. This is most often used for setting constant essential BCs. A new method,
SetSubVectorHost, has been added for cases where host execution is always
needed (e.g. when the DOFs array is small).
- Added GPU support in GradientGridFunction and InnerProduct Coefficient classes
by implementing their Project methods.
- Added new method: GridFunction::GetGradients, with GPU support, for computing
the gradients of a GridFunction on all elements.
- The CUDA-specific names used by some of the unit tests like 'cunit_tests' and
'pcunit_tests' were replaced by names using 'gpu' instead of 'c' (short for
CUDA) or 'cuda'. These tests automatically run the CUDA/HIP tests based on the
MFEM build configuration.
New and updated examples and miniapps
-------------------------------------
- Added the miniapps/fluids directory and moved the previous Navier and the new
incompressible Schrödinger flow miniapps into it.
- Introduced the new Incompressible Schrödinger Flow (ISF) miniapp, which models
inviscid fluid dynamics by solving the linear Schrödinger equation, leveraging
the hydrodynamical analogy to quantum mechanics.
- New particle-related miniapps:
* New transient Navier-Stokes fluid-particles solver NavierParticles in
miniapps/fluids/navier/navier_particles, for modeling tracer particles in
fluid flow, demonstrating use of the new ParticleSet class.
* New Navier miniapp, miniapps/fluids/navier/navier_bifurcation, showing the
use of NavierParticles in a 2D bifurcating channel flow.
* New FindPointsGSLIB miniapp, miniapps/gslib/particles_redist, showing
parallel-redistribution of particle data between MPI ranks.
* Particle visualization features in common/particles_extras for viewing
particle locations and trajectories (ParticleTrajectories) using GLVis.
- Added a new miniapp (meshing/mesh-bounding-boxes) that computes the bounding
boxes for each element of a given mesh, and the bounds on the determinant of
the Jacobian of the transformation.
- Added a new miniapp (tools/gridfunction-bounds) to compute piecewise linear
bounds on a given high-order grid function.
- Added a new miniapp (electromagnetics/lorentz) which computes the trajectory
of a charged particle, subject to Lorentz forces, in electrostatic and/or
magnetostatic fields as computed by the volta or tesla miniapps.
- Added miniapps to demonstrate an implementation of the absolute-value
l1-Jacobi preconditioners in partially assembled operators. This includes
Multigrid wrapper to demonstrate the effectiveness of these Jacobi-type
operators as smoothers. See the miniapps/diag-smoothers/ directory.
- Updated the mtop miniapp with a GPU enabled forward and adjoint solver for
isotropic linear elasticity.
Miscellaneous
-------------
- Introduced MFEM_FETCH_TPLS CMake option to enable downloading, configuring,
and building of TPLs alongside MFEM (currently supported TPLs are hypre,
METIS, and GSLIB).
- Added quadrature function support to the VisIt and Conduit data collections.
- Added access to the internal parallel matrix in Par(Mixed)BilinearForm and
related utility methods for elimination of BCs.
- FindPointsGSLIB has a new constructor that accepts the mesh object and
internally calls the Setup() method so users do not have to. The FreeData()
method has also been moved to the destructor so users do not need to manually
free-up the memory if the destructor is called before MPI_Finalize().
- Added parallel Address Sanitizer, serial and parallel Undefined Behavior
Sanitizer and serial Memory Sanitizer GitHub actions tests on Ubuntu.
API changes
-----------
- mfem::internal::tensor and mfem::internal::dual have been moved to
mfem::future::tensor and mfem::future::dual.
- API addition: in class Operator, added virtual functions: AbsMult, and
AbsMultTranspose; in class Vector, added Abs and Pow.
- ParBilinearForm::EliminateEssentialVDofsInRhs() has been deprecated in favor
of ParallelEliminateEssentialTDofsInRhs().
Version 4.8, released on Apr 9, 2025
====================================
Discretization improvements
---------------------------
- Added high-order basis functions on pyramid-shaped elements for all spaces in
the de Rham complex based on the paper "Orientation embedded high order shape
functions for the exact sequence elements of all shapes" by Fuentes, Keith,
Demkowicz and Nagaraj (doi.org/10.1016/j.camwa.2015.04.027). Positive basis
functions (Bernstein basis) for H1 and L2 on pyramids were also added.
- Added support for parallel p- and hp-refinement on quad/hex meshes. For hp, we
currently support only isotropic refinement with L2 or H1 spaces. See the new
miniapps hpref and phpref in the miniapps/meshing/ directory.
- Added several improvements for hyperbolic problems:
* Assembly of Jacobians in HyperbolicFormIntegrator
* Component-wise upwinded flux (ComponentwiseUpwindFlux)
* Average fluxes in NumericalFlux (formerly RiemannSolver) and FluxFunction
- Added convenience methods to class FiniteElementSpace to directly identify all
degrees of freedom on the exterior faces of the domain, without referencing
boundary element attributes (GetExteriorVDofs and GetExteriorTrueDofs).
- Altered (Par)GridFunction::Compute*Error functions to ensure they return
non-negative values and therefore behave as "norms".
- SubMesh and ParSubMesh have been extended to work on nonconforming meshes.
Extracting volume and exterior surface submeshes are both supported.
- LinearFormIntegrator, BilinearFormIntegrator and NonlinearFormIntegrator now
inherit from a base class Integrator that centralizes the logic for selecting
quadrature rules. This includes a virtual method GetDefaultIntegrationRule,
which should be favored over directly defining a default integration rule in
the element-level assembly routines. The latter is still possible, by leaving
the new virtual method as its default base implementation of returning NULL.
- Added NURBS-based H(div) and H(curl) elements in 2D and 3D. Currently only for
single patch meshes in serial.
- Refactored ALGOIM cut integration rules. The interface is unified with the
interface for moment based cut integration rules.
- FiniteElementSpace has new methods to directly set the prolongation and
restriction operators to user-specified sparse matrices.
- Added support for H(div) spaces in class QuadratureInterpolator. Currently only
(vector) VALUES, (vector) PHYSICAL_VALUES, and PHYSICAL_MAGNITUDES evaluations
are implemented.
- Added support for boundary constraints to class Hybridization.
- Added support for custom interpolation procedure in FindPointsGSLIB.
- Added GSLIB-based gather-scatter operator.
Meshing improvements
--------------------
- Added support for nonuniform anisotropic mesh refinement on serial quad/hex
meshes with arbitrary spacing in each direction. This enables in particular
3:1 refinement, as demonstrated in the new meshing miniapp ref321.
- Added capabilities for optimization and adaptation of periodic meshes with
TMOP. The internals of TMOP were modified so that the optimization problem is
always solved with respect to mesh displacements.
- New convenience methods for manipulating boundary attribute markers in class
Mesh: MarkExternalBoundaries, MarkNamedBoundaries, UnmarkInternalBoundaries
and UnmarkNamedBoundaries. See Examples 1/1p and 11p for basic usage.
- Added a new method, GetExteriorFaceMarker, to the serial and parallel mesh
classes for identifying faces on the exterior of the mesh irrespective of
their presence in the list of "boundary elements".
- Added native AD support for computing the derivatives of numerous TMOP metrics
that didn't have first or second derivative implementations.
- Added ExodusII output capability which can handle in particular pyramid and
wedge element types. Mixed meshes are also supported.
- Added InverseElementTransformation::InitGuessType::EdgeScan as an alternative
initial guess type. This guess type tries solving with multiple initial guesses
along the r/s/t=0 edges of the element until a valid solution is found or all
initial guess points are exhausted.
- Added new convenience constructors for NURBS patches and knot vectors.
GPU computing
-------------
- Extended FindPointsGSLIB to support general field interpolation on GPUs. Note
that this requires that switch from gslib v1.0.7 to v1.0.9.
- Added support for GPU-accelerated batched linear algebra (using cuBLAS,
hipBLAS, MAGMA, or native MFEM functionality) through the BatchedLinAlg class.
- A new GPU kernel dispatch mechanism was introduced. Users can instantiate
specialized kernels for specific combinations of (for example) polynomial
degree and number of quadrature points using method AddSpecialization in
classes DiffusionIntegrator and MassIntegrator (this functionality may be
added to more integrators in the future).
- Added BatchInverseElementTransformation to batch InverseElementTransformation
searches. Batch searches are currently limited to meshes with a single element
geometry type of SEGMENT, SQUARE, or CUBE. Additional element geometry types
may be added in the future. Mixed element order meshes are supported. Batch
searches can be performed on the CPU (serial per MPI rank) or GPU. Embedded
elements (SEGMENT in 2D/3D space or SQUARE in 3D space) are supported, however
the existing solvers may struggle to find a valid solution.
- Calls to slower fallback kernels can be reported to mfem::err by setting the
environment variable MFEM_REPORT_KERNELS to any value other than NO or by
explicitly calling KernelReporter::Enable. For higher performance, users can
then add specializations for these kernels.
- Element assembly kernels have been added for low-order refined -> high-order
transfer operators. New kernels can be offloaded as device kernels. Example
usage may be found in lor-transfer.cpp in the miniapps/tools/ directory.
- Added GPU acceleration and element assembly for DivDivIntegrator and
NormalTraceJumpIntegrator.
- Allow BlockLowerTriangularPreconditioner to run on GPU.
- Use device vectors in GMRES, FGMRES and other iterative methods.
Linear and nonlinear solvers
----------------------------
- Added GPU acceleration of the algebraic hybridization solver for grad-div
problems in H(div). See Example 4.
- Added a self-contained implementation of the Method of Moving Asymptotes (MMA)
for solving optimization problems.
- Added support for SUNDIALS v7. See the section "API changes" for some small
changes related to this new version.
- Changed the name of class IterativeSolverMonitor to IterativeSolverController,
which now allows for specifying convergence by a user defined criterion. For
backward compatibility, the old name is still available.
- Refactored the ARKStepSolver class (ARKODE interface) to use the Mult() method
of TimeDependentOperator only when the associated ODE operator is expressed in
explicit form (i.e., TimeDependentOperator::isExplicit()), otherwise the
method ExplicitMult() is used. A check has been added to ARKStepSolver to
verify that the associated ODE operator is not in explicit form when a mass
matrix solver is enabled via a call to either the UseMFEMMassLinearSolver or
UseSundialsMassLinearSolver methods. This is because enabling a mass matrix
solver assumes that F(u,k,t) = M k in the associated ODE operator.
- Added ODE solvers selection routines. This creates a uniformity across
examples, miniapps and other executables in regard to ODE (time-integrator)
selection.
- Added new mechanism for retrieving and setting state vectors in ODE solvers.
This is relevant for AB/AM and gen-alpha solvers.
- Added ODEsolver/ODEsolver2 unit tests to verify order of convergence and
read/write functionality.
New and updated examples and miniapps
-------------------------------------
- Added an MFEM example for the eikonal equation (examples/ex40) based on the
proximal Galerkin method introduced by Keith and Surowiec.
- Added miniapps to demonstrate the H(div) and H(curl) NURBS elements.
- Fixed element visualization in the Mesh explorer miniapp.
- Added a command line option to all miniapps (-p or --send-port) for
specifying the GLVis server socket port (19916 by default).
- Added a new toy miniapp that animates an interesting fidget spiral cone toy.
See miniapps/toys/spiral.cpp.
Miscellaneous
-------------
- Updated the benchmarks (in tests/benchmarks) to work with the latest Google
Benchmarks classes (version 1.9.1). Renamed the MFEM_ENABLE_GOOGLE_BENCHMARKS
CMake option to just MFEM_ENABLE_BENCHMARKS.
- Updated the minimum CMake version requirements:
* CMake >= 3.12 for CPU builds,
* CMake >= 3.17 for CUDA builds, and
* CMake >= 3.14 for HIP builds (CMake >= 3.12 may work as well, not tested).
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- In class GridFunction, fec was renamed to fec_owned.
- RiemannSolver was renamed to NumericalFlux (the old name has been deprecated
through typedef).
- API changes due to SUNDIALS v7:
* the SUNDIALS types realtype and booleantype are no longer defined by v7
and therefore MFEM now uses the new type names sunrealtype and
sunbooleantype, respectively, which MFEM defines when using SUNDIALS < v6
where these types were not defined.
* The SUNDIALS macro SUNLS_SUCCESS and some other *_SUCCESS macros were
removed and replaced by SUN_SUCCESS in v7, so to avoid tedious checks for
SUNDIALS versions, MFEM now defines and uses the constant SUN_SUCCESS when
using SUNDIALS < v7.
* The constants SUN_PREC_*, introduced by SUNDIALS v6 are now introduced by
MFEM when using SUNDIALS < v6 to avoid tedious version checks.
- TMOP_Metric_skew2D has been marked as deprecated.
Version 4.7, released on May 7, 2024
====================================
- Added support for single precision (with corresponding hypre build). The MFEM
floating point type was generalized from `double` to `real_t`. For details see
https://github.com/orgs/mfem/discussions/4207.
Meshing improvements
--------------------
- Added the capability to partition (big) serial meshes in serial code, see the
new classes MeshPartitioner and MeshPart. This capability is also exposed as a
menu option in the mesh-explorer miniapp in miniapps/meshing.
- Added named attribute sets and basic supporting methods to the Mesh class as a
convenient means of referring to sets of domain or boundary attribute numbers.
See the new Example 39/39p and data/compass.mesh.
- Introduced formulas for refinement of patches in NURBS meshes. Refinement by
arbitrary integer factors is also enabled, e.g. in the mesh-explorer miniapp.
NURBS coarsening and knot removal are also introduced.
- Added support for internal boundary elements in nonconforming meshes.
- The ReadCubit Genesis mesh importer has been rewritten to improve readability.
Discretization improvements
---------------------------
- Added a new nonlinear integrator, `HyperbolicFormIntegrator` that implements
both element-wise weak divergence and face-wise numerical flux for a general
system of hyperbolic conservation laws. To use the integrator for a specific
flux function, users can define a derived class of `FluxFunction`. Currently,
advection, Burgers, shallow-water and Euler equations (see Example 18/18p) are
available.
- Added a capability to construct cut-surface and cut-volume IntegrationRules
through a moment-fitting approach. The cut is specified by the zero level set
of a Coefficient. See fem/intrules_cut.hpp and the new Example 38.
- Introduced support for high-order nonconforming Nedelec elements on simplices.
GPU computing
-------------
- Added partial assembly and GPU support for the DG diffusion integrator.
- Efficient GPU-accelerated LOR assembly is now supported on surface meshes.
- Added functionality to automatically configure hypre's compute policy to match
MFEM's compute policy when hypre is built with GPU support. Requires version
hypre-2.31.0 or later.
- Added support for full assembly on simplices.
- Added partial assembly for linear elasticity (no sum factorization for now).
- Added functionality for BilinearFormIntegrators to use kernels that work for
both tensor and unstructured elements.
- The RAJA backend will use `seq_exec` for serial loop execution when RAJA
v2023.06.00 and beyond is detected as `loop_exec` is deprecated.
- API change: The macro MFEM_HYPRE_FORALL (from hypre.hpp) which was intended
for internal use, has been removed and replaced by the function template
mfem::hypre_forall in general/forall.hpp.
New and updated examples and miniapps
-------------------------------------
- Added a new miniapp illustrating elastic contact based on the Tribol library,
(https://github.com/LLNL/Tribol). See miniapps/tribol.
- Added a miniapp to demonstrate low order refined (LOR) block preconditioning
for linear elasticity on GPUs. See miniapps/solvers/lor_elast.
- Added a new block solver in miniapp/solvers for the Darcy problem. The new
solver is based on a Bramble-Pasciak preconditioning. User can use and
implement their own preconditioner for the mass matrix.
- Added a small miniapp for printing the shape functions of a KnotVector. See
miniapps/nurbs/nurbs_printfunc.cpp.
- Added two new example codes: 38 and 39/39p described above. Substantially
updated Example 18/18p.
Miscellaneous
-------------
- Updated the Doxygen documentation style, which now requires Doxygen version
1.9.8 or later. See the doc/ directory.
- Improved thread safety for global variables in the library, e.g. for IntRules,
RefinedIntRules, GlobGeometryRefiner, and FiniteElement::dof2quad_array.
- PETSc integration now generally requires PETSc version 3.21 or later, though
depending on the functionality older versions may still work.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.6, released on September 27, 2023
===========================================
- MFEM is now available in Homebrew and can be installed on a Mac with just
"brew install mfem". See https://formulae.brew.sh/formula/mfem.
Meshing improvements
--------------------
- Added asymptotically-balanced TMOP compound metrics 90, 94, 328, 338. A new
tool, tmop-metric-magnitude, can be used to track how metrics change under
geometric perturbations. See miniapps/tools.
- Several NURBS meshing improvements:
* Support for free connectivity of NURBS patches allowing for more complex
patch configurations such as C-meshes.
* New methods to set and get attributes on NURBS patches and patch boundaries.
* The edge to knot map for NURBS meshes can be determined automatically. It is
no longer needed to specify this in the NURBS mesh.
* Added curve interpolation method for NURBS.
* See miniapps/nurbs for example meshes and miniapps.
Discretization improvements
---------------------------
- SubMesh and ParSubMesh have been extended to support the transfer of
Nedelec and Raviart-Thomas finite element spaces.
- Added support for partial assembly on NURBS patches, and NURBS-patch sparse
matrix assembly. Patch matrix assembly includes the option to use reduced
approximate integration rules, computed by the newly implemented non-negative
least-squares (NNLS) solver.
- Support for parallel transfer of H1 fields using the low-order refined (LOR)
transfer operators in L2ProjectionGridTransfer
- Added KDTree class for 2D/3D set of points, which is then utilized in the new
KDTreeNodalProjection class to project a function defined on an arbitrary set
of points onto an MFEM grid function. This functionality is demonstrated in
the nodal-transfer miniapp. The current implementation is serial only. Further
extensions can include search in arbitrary dimensional spaces.
- Added support for p-refined meshes in GSLIB-FindPoints.
- Device kernels can now access device-specific DOF and quadrature limits using
the DofQuadLimits structure, allowing increased limits when executing on CPU.
The limits for the runtime selected device can be accessed in host code using
DeviceDofQuadLimits::Get(). The global constants MAX_D1D and MAX_Q1D are no
longer available.
- Face restriction operators for Nedelec and Raviart-Thomas finite element
spaces are now supported through the ConformingFaceRestriction class.
- VectorFEBoundaryFluxLFIntegrator is now supported on device/GPU.
Linear and nonlinear solvers
----------------------------
- Updated the MUMPS interface to support multiple right-hand sides, block
low-rank compression, builds using 64-bit integers, and other improvements.
- Added an interface to the MKL Pardiso sparse direct solver developed by Intel.
The interface provides a serial (OpenMP shared memory) version of Pardiso for
use with SparseMatrix. This complements the existing parallel (MPI distributed
memory) version already available through the CPardiso MFEM integration.
- Added HIP support to the PETSc and SUNDIALS interfaces.
New and updated examples and miniapps
-------------------------------------
- Added a new H(div) solver miniapp demonstrating the use of a matrix-free
saddle-point solver methodology, suitable for high-order discretizations and
for GPU acceleration. Examples illustrating the solution of Darcy and grad-div
problems are included. See miniapps/hdiv-linear-solver.
- Added new Discontinuous Petrov-Galerkin (DPG) miniapp which includes serial
and parallel examples for diffusion, convection-diffusion, acoustics and
Maxwell equations. The miniapp includes new classes such as (Par)DPGWeakForm,
(Par)ComplexDPGWeakForm and (Complex)BlockStaticCondensation. Three new
integrators are added in support of DPG systems: TraceIntegrator,
NormalTraceIntegrator and TangentTraceIntegrator. See miniapps/dpg.
- Added a new miniapp that implements the SPDE method for generating Gaussian
random fields of Matern covariance. The resulting random field can be used,
e.g., to model material uncertainties. See miniapps/spde.
- Added a new parallel LOR transfer miniapp, plor-transfer, which mirrors the
functionality of the serial LOR transfer miniapp. See miniapps/tools.
- New serial miniapp, nodal-transfer, demonstrating the use of KDTree to map a
parallel grid function to a different parallel partitioning of the same mesh.
- Added 3 additional TMOP miniapps in miniapps/meshing:
* Mesh-Quality evaluates quality using size, skewness, and aspect-ratio
computed from the Jacobian of the transformation.
* Mesh-Fitting can be used for interface and boundary fitting to implicit
domains defined using level-set functions.
* Fit-Node-Position fits selected mesh nodes to specified positions, while
maintaining overall mesh quality.
- Added 4 new example codes:
* Example 34/34p solves a simple magnetostatic problem where source terms and
boundary conditions are transferred with SubMesh objects.
* Example 35p implements H1, H(curl) and H(div) variants of a damped harmonic
oscillator with field transfer using SubMesh objects.
* Example 36/36p demonstrates the solution of the obstacle problem with a new
finite element method (proximal Galerkin).
* Example 37/37p demonstrates topology optimization with MFEM.
- Added a random refinement option to the mesh-explorer miniapp to assist users
in experimenting with nonconforming meshes.
- Moved the distance solver methods from miniapps/shifted to miniapps/common.
Miscellaneous
-------------
- Improved lambda body debugging with the addition of mfem::forall functions.
These functions can take the place of the MFEM_FORALL macros, which have been
preserved for backwards compatibility.
- Added an address sanitizer GitHub action for a serial build/test on Ubuntu,
based on Clang/LLVM (https://clang.llvm.org/docs/AddressSanitizer.html).
- Reorganized files for bilinear form, linear form, and nonlinear form integrators
in the fem/integ/ subdirectory.
- FiniteElementSpace::GetFE has been updated to abort instead of returning NULL for
an empty partition.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.5.2, released on March 23, 2023
=========================================
- Added support for pyramids in non-conforming meshes. Currently only isotropic
refinement is supported in this case.
- Removed the support for the Mesquite toolkit. We recommend using MFEM's TMOP
functionality instead for mesh optimization. See the mesh-optimizer miniapp.
- Added a fast normalization-based distance solver, see the Distance miniapp
in the miniapps/shifted/ directory.
- Added a new meshing miniapp, Reflector, which reflects a high-order or NURBS
hexahedral mesh about a plane.
- Updated logic in FindPointsGSLIB to ignore points found near (but outside) the
domain boundary.
- Added an option to auto-balance compound TMOP metrics.
- Fixed a bug in TMOP metric 301.
- When using discontinuous (L2) spaces, use local (element-wise) L2 projection
as the coarsening operator for non-conforming AMR meshes.
- Added support for GridFunction::GetGradients() and GetVectorGradient() on
face-neighbor elements.
- Added support for pyramids in Gmsh meshes.
- The Mesh Explorer miniapp can now save mesh files in the VisIt or ParaView
formats using the corresponding DataCollection. See option 'D' in the menu.
- VisItDataCollection now correctly handles data collection names containing
underscores.
- Added support for shared Windows builds with MSVC through CMake.
Developers note: this enhancement is facilitated by the addition of the macro
MFEM_EXPORT, see config.hpp for more details on its usage.
- The following integrations have updated minimum version requirements:
* RAJA >= 2022.10.3
API changes
-----------
- The implicit cast methods of class Vector to 'double *' and 'const double *'
have been deprecated and generate deprecation warnings if used. They will be
removed in a future release.
- The methods Mesh::GetFaceBaseGeometry and Mesh::GetFaceGeometryType have been
deprecated, and Mesh::GetFaceGeometry (which provides identical functionality)
should be used instead.
- VisItDataCollection::SetPadDigits() no longer alters the number of digits
used to represent the MPI rank because VisIt seems to require 6 digits.
This parameter can still be explicitly overridden with SetPadDigitsRank().
Version 4.5, released on October 22, 2022
=========================================
Meshing improvements
--------------------
- Added new SubMesh and ParSubMesh classes that can be used to extract a subset
of a given Mesh. These classes have the same functionality as Mesh and ParMesh
and work with all existing MFEM interfaces like finite element spaces etc.
- Added a method, ParMesh::GetSerialMesh(), that reconstructs a partitioned
parallel mesh on a given single rank. Also, added ParMesh::PrintAsSerial(),
which saves the reconstructed serial mesh to a C++ stream on rank 0.
- Added more 3D TMOP metrics, as well as specialized metrics for mesh
untangling and worst-case quality improvement.
- Added a new method, Mesh::NodesUpdated, which should be called after the mesh
node coordinates have changed, e.g. after the mesh has moved. This is
necessary, for example, with device assembly of linear and bilinear forms.
- Added support for mixed meshes and pyramids in GSLIB-FindPoints.
Discretization improvements
---------------------------
- Added full assembly and device support for several LinearForm integrators:
* DomainLF: (f, v)
* VectorDomainLF: ((f1,...,fn), (v1,...,vn))
* DomainLFGrad: (f, grad(v))
* VectorDomainLFGrad: ((f1x,f1y,f1z,...,fnx,fny,fnz), grad(v1,...,vn))
The device assembly of linear forms has to be explicitly enabled by calling
LinearForm::UseFastAssembly(true), otherwise the legacy linear form assembly
is used by default.
- Added support for assembling low-order-refined matrices using a GPU-enabled
"batched" algorithm. The lor_solvers and plor_solvers now fully support GPU
acceleration with arbitrary user-supplied coefficients.
- Added a new class FaceQuadratureSpace that allows for the construction of
QuadratureFunctions on the interior or boundary faces of a mesh.
- Added a class CoefficientVector for efficient access of variable coefficient
values at quadrature points (in particular for GPU/device kernels).
- Added WhiteGaussianNoiseDomainLFIntegrator: a LinearFormIntegrator class for
spatial Gaussian white noise.
- Added a new Zienkiewicz-Zhu patch recovery-based a posteriori error estimator.
See fem/estimators.hpp.
- Various fixes and improvements in LinearFormExtension.
Linear and nonlinear solvers
----------------------------
- Added a new class DGMassInverse that performs a local element-wise CG
iteration to solve systems involving the discontinuous Galerkin mass matrix,
including support for device/GPU acceleration.
- Added more flexibility to the constrained solver classes:
* PenaltyConstrainedSolver now allows for a vector of penalty parameters
(necessary for penalty contact)
* PenaltyConstrainedSolver and EliminationSolver can use GMRES or PCG
* All constraint solver classes can take a user-defined preconditioner
- Added functions to toggle additional options for the SuperLU_Dist and Hypre
preconditioners (ParaSails, Euclid, ILU).
- Added boundary elimination with device support for `SparseMatrix` and
`HypreParMatrix`.
New and updated examples and miniapps
-------------------------------------
- Added a new elasticity miniapp, Hooke, that showcases a low-level approach of
using MFEM to solve a nonlinear elasticity problem based on the fundamental
finite element operator decomposition. The miniapp also integrates with
automatic differentiation tools like a native dual number implementation or a
third party library such as Enzyme. See miniapps/elasticity for more details.
- Added example for body-fitted volumetric and shape integration using the
Algoim library in miniapps/shifted.
- Add a new example code, Example 33/33p, to demonstrate the solution of
spectral fractional PDEs with MFEM.
Integrations, testing and documentation
---------------------------------------
- Added a Dockerfile for a simple MFEM container, see config/docker/README.md.
More sophisticated developer containers are available in the new repo
https://github.com/mfem/containers.
- Added support for the LLVM-based automatic differentiation tool Enzyme, see
https://github.com/EnzymeAD/Enzyme. Build system flags and a convenience
header are provided. The functionality and interaction are demonstrated in a
new miniapp in miniapps/elasticity.
- Added support for partial assembly and fully matrix-free operators on mixed
meshes (different element types and p-adaptivity) through libCEED, including
device acceleration, e.g. with NVIDIA and AMD GPUs. The p-adaptivity is
currently limited to 2D serial meshes. All mixed element topologies are
supported in both serial and parallel.
- Added support for ParMoonolith, https://bitbucket.org/zulianp/par_moonolith,
which provides parallel non-conforming, non-matching, variational, volumetric
mesh information transfer. With ParMortarAssember, fields can be exchanged
between arbitrarily distributed and unrelated finite element meshes in a
variationally consistent way.
- Fully encapsulated SUNDIALS `N_Vector` object within the `SundialsNVector`
class by removing deprecated (e.g. `HypreParVector::ToNVector`) and
non-deprecated (e.g. `Vector::ToNVector`) functions in other classes.
- New benchmark for the different assembly levels inspired by the CEED
Bake-Off Problems, see tests/benchmarks/bench_assembly_levels.cpp.
- Added Windows 2022 CI testing with GitHub actions.
Miscellaneous
-------------
- The method SparseMatrix::EnsureMultTranspose() is now automatically called
by the methods AddMultTranspose(), MultTranspose(), and AbsMultTranspose().
Added a method with the same name to class HypreParMatrix which is also called
automatically by the HypreParMatrix::MultTranspose() methods.
- Updated various MemoryUsage methods to return 'std::size_t' instead of 'long'
since the latter is 32-bit in Win64 builds.
- When using `AssemblyLevel::FULL`, `FABilinearFormExtension::FormSystemMatrix`
outputs an `OperatorHandle` containing a `SparseMatrix` in serial, and an
`HypreParMatrix` in parallel (instead of a `ConstrainedOperator`).
- In various places in the library, replace the use of 'long' with 'long long'
to better support Win64 builds where 'long' is 32-bit and 'long long' is
64-bit. On Linux and MacOS, both types are typically 64-bit.
- The behavior of GridFunction::GetTrueVector() has been changed to not return
an empty true vector.
- Added support for ordering search points byVDIM in FindPointsGSLIB.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.4, released on March 21, 2022
=======================================
Linear and nonlinear solvers
----------------------------
- Added support for using the hypre library built with HIP support. Similar to
the existing hypre + CUDA support, most of the MFEM examples and miniapps work
transparently with hypre + HIP builds. This includes the BoomerAMG, AMS, and
ADS solvers.
- Added a simple singleton class, Hypre, to automatically set hypre's global
parameters, particularly GPU-relevant options. Updated parallel example codes
and miniapps to call Hypre::Init() where appropriate.
- Added hipSPARSE support for sparse matrix-vector multiplications.
- More explicit and consistent formatting of the output of iterative solvers
with the new IterativeSolver::PrintLevel options. See linalg/solvers.hpp.
Meshing improvements
--------------------
- New TMOP-based methods for hr-adaptivity, interface fitting, and tangential
relaxation of high-order meshes.
- Added initial support for meshes with pyramidal elements, including several
pyramidal meshes in the data/ directory and support for the lowest order H1,
Nedelec, Raviart-Thomas, and L2 basis functions on pyramids.
- Added a simpler interface to access mesh face information, see FaceInformation
and GetFaceInformation in the Mesh class.
- Gmsh meshes where all elements have zero physical tag (the default Gmsh output
format if no physical groups are defined) are now successfully loaded, and
elements are reassigned attribute number 1.
- Added ParMesh adjacency set (adjset) creation support to the Conduit Mesh
Blueprint MFEM wrapper functions in ConduitDataCollection.
Discretization improvements
---------------------------
- Added general dof transformation to support high order Nedelec basis functions
on tetrahedral meshes without reordering. The ReorientTetMesh method of the
Mesh and ParMesh classes has been deprecated. See the new DofTransformation
class in fem/doftrans.hpp.
- GPU-enabled partial (PA) and element (EA) assembly for discontinuous Galerkin
methods on nonconforming AMR meshes.
- Support for arbitrary order Nedelec and Raviart-Thomas elements on wedges.
- Added special Nedelec and Raviart-Thomas basis functions for modeling three
dimensional vector fields in 1D and 2D domains, see the new Example 31/31p.
- GridFunctionCoefficient (and the related vector, gradient, divergence, and
curl classes) now work properly with LORDiscretization and LORSolver.
- Added PA support for the action of MixedScalarCurlIntegrator in 2D and
MixedVectorGradientIntegrator in 2D and 3D, as well as their transposes.
- Coefficient::SetTime now propagates the new time into internally stored
Coefficient objects.
- Split the fem/fe.?pp files into separate files in the new fem/fe/ directory to
simplify and clarify the organization of FiniteElement classes.
New and updated examples and miniapps
-------------------------------------
- Added two new miniapps with initial support for automatic differentiation (AD)
in the miniapps/autodiff/ directory. Users can select between external library
and native implementation during configuration. The support for AD will be
extended in future releases of MFEM.
- Added Binder (mybinder.org) configuration files for C++ MFEM Jupyter Notebooks
with inline GLVis visualization in the new examples/jupyter/ directory with a
sample notebook based on Example 1. The implementation is based on xeus-cling,
see github.com/jupyter-xeus/xeus-cling and github.com/GLVis/xeus-glvis.
- Added a new miniapp (Extrapolation) for PDE-based extrapolation of finite
element functions from known values in a set of elements to the rest of the
computational domain. See miniapps/shifted/extrapolate.cpp.
- Added new miniapp that uses the ParELAG library, its hybrid smoothers, and the
hierarchy of spaces created by the element-based AMG (AMGe) methodology in
ParELAG to build multigrid solvers for H(curl) and H(div) forms. See the
miniapps/parelag directory for more details.
- Added a new Example 30/30p demonstrating support for mesh preprocessing to
resolve fine scale problem data before simulation. This feature uses adaptive
mesh refinement to control the associated data oscillation error.
- Added new Examples 31, 31p and 32p showing anisotropic definite Maxwell
serial/parallel solver and parallel eigensolver 1D, 2D, or 3D.
- Updated the mesh-optimizer and pmesh-optimizer miniapps to demonstrate the
hr-adaptivity and interface fitting capability.
- The HPC versions of ex1 and ex1p (in miniapps/performance) now support runtime
selection of either 2D or 3D meshes.
Integrations, testing and documentation
---------------------------------------
- Doxygen documentation for all releases is now available at docs.mfem.org.
- The following integrations have updated minimum version requirements:
* HIOP >= 0.4.6
* HYPRE >= 2.23.0 for HIP support
* libCEED >= 0.10
* PUMI >= 2.2.6
* RAJA >= 0.14.0
* Umpire >= 3.0.0
see INSTALL for more details.
- Added new optional integrations with ParELAG and CoDiPack (version >= 1.9.3+).
- Added initial support for Google Benchmark (version >= 1.5.6) in the
tests/benchmarks directory. It can be enabled with MFEM_USE_BENCHMARK=YES.
- Switched from Artistic Style (astyle) version 2.05.1 to version 3.1 for code
formatting. See the "make style" target.
Miscellaneous
-------------
- Added a simple singleton class, Mpi, as a replacement for MPI_Session. New
code should use Mpi::Init() and other Mpi methods instead of MPI_Session.
- Added ParaView visualization of QuadratureFunction fields, through both
QuadratureFunction::SaveVTU and ParaViewDataCollection::RegisterQField.
- Fixed several MinGW build issues on Windows.
- Added 'double' atomicAdd implementation for previous versions of CUDA.
- HypreParVector and Vector now support C++ move semantics, and the copy
constructor for HypreParVector now copies the local vector data.
- Removed the 'u' flag in the ar command, to update all files in the archive,
avoiding file name collisions from different subdirectories.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.3, released on July 29, 2021
======================================
Discretization improvements
---------------------------
- Variable order spaces, p- and hp-refinement. This is the initial (serial)
support for variable-order FiniteElementCollection and FiniteElementSpace.
The new method FiniteElementSpace::SetElementOrder can be called to set an
arbitrary order for each mesh element. The conforming interpolation matrix
will now automatically constrain p- and hp- interfaces, enabling general
hp-refinement in both 2D and 3D, on uniform or mixed NC meshes. Support for
parallel variable-order spaces will follow shortly.
- Extended the support for field transfer between high-order and low-order
refined finite element spaces to include: dual fields and H1 fields (both
primary and dual). These are illustrated in the lor-transfer miniapp.
- Improved libCEED integration, including support for VectorCoefficient,
ConvectionIntegrator, and VectorConvectionNLFIntegrator with libCEED backends.
- Extending support for L2 basis functions using MapTypes VALUE and INTEGRAL in
linear interpolators and GridFunction "GetValue" methods.
- Changed the interface for the error estimator and implemented the Kelly error
indicator for scalar-valued problems, supported in serial and parallel builds.
- Added support for the "BR2" discontinuous Galerkin discretization for
diffusion via DGDiffusionBR2Integrator (see Example 14/14p).
- Added convective and skew-symmetric integrators for the nonlinear term in the
Navier-Stokes equations.
- Added new classes DenseSymmetricMatrix and SymmetricMatrixCoefficient for
efficient evaluation of symmetric matrix coefficients. This replaces the now
deprecated EvalSymmetric in MatrixCoefficient. Added DiagonalMatrixCoefficient
for clarity, which is a typedef of VectorCoefficient.
- Added support for nonscalar coefficient with VectorDiffusionIntegrator.
Linear and nonlinear solvers
----------------------------
- Added support for AMG preconditioners on GPUs based on the hypre library
(version 2.22.0 or later). These include BoomerAMG, AMS and ADS and most
MFEM examples that use hypre have been ported to support this functionality.
The GPU preconditioners require that both hypre and MFEM are built with CUDA
support. Hypre builds with CUDA and unified memory are also supported and
can be used with `-d cuda:uvm` as a command-line option.
- Added support for AMG preconditioners for non-symmetric systems (e.g.
advection-dominated problems) using hypre's approximate ideal restriction
(AIR) AMG. Requires hypre version 2.14.0 or newer. Usage is illustrated in
example 9/9p.
- Added new functionality for constructing low-order refined discretizations and
solvers, see the LORDiscretization and LORSolver classes. A new basis type for
H(curl) and H(div) spaces is introduced to give spectral equivalence. This
functionality is illustrated in the LOR solvers miniapp in miniapps/solvers.
- Generalized the Multigrid class to support non-geometric multigrid. Previous
functionality, based on FiniteElementSpaceHierarchy, is now available in the
derived class GeometricMultigrid.
- Introduced solver interface for linear problems with constraints, a few
concrete solvers that implement the interface, and a demonstration of their
use in Example 28(p), which solves an elasticity problem with zero normal
displacement (but allowed tangential displacement) on two boundaries.
- Added high-order matrix-free auxiliary Maxwell solver for H(curl) problems,
as described in Barker and Kolev 2020 (https://doi.org/10.1002/nla.2348). See
Example 3p and linalg/auxiliary.?pp.
- Improved interface for using the Ginkgo library, including: support for matrix-
free operators in Ginkgo solvers, new wrappers for Ginkgo preconditioners, HIP
support, and reduction of unnecessary data copies.
- Added initial support for hypre's mixed integer (mixedint) capability, which
uses different data types for local and global indices in order to save memory
in large problems. This capability requires that hypre was configured with the
--enable-mixedint option. Note that this option is currently tested only in
ex1p, ex3p, and ex4p, and may not work in more general settings.
- Added AlgebraicCeedSolver that does matrix-free algebraic p-multigrid for
diffusion problems with the Ceed backend.
- Added interface to MUMPS direct solver. Its usage is demonstrated in ex25p.
See http://mumps.enseeiht.fr/ for more details. Supported versions >= 5.1.1.
- Added three ESDIRK time integrators: implicit trapezoid rule, L-stable
ESDIRK-32, and A-stable ESDIRK-33.
- Implemented a variable step-size IMEX (VSSIMEX) method for the Navier miniapp.
- Implemented an adaptive linear solver tolerance option for NewtonSolver based
on the algorithm of Eisenstat and Walker.
Meshing improvements
--------------------
- Added support for reading high-order Lagrange meshes in VTK format. Arbitrary-
orders and all element types are supported. See the VTK blog for more info:
https://blog.kitware.com/wp-content/uploads/2018/09/Source_Issue_43.pdf.
- Introduced a new non-conforming mesh format that fixes known inconsistencies
of legacy "MFEM mesh v1.1" NC format and works consistently in both serial and
parallel. ParMesh::ParPrint can now print non-conforming AMR meshes that can
be used to restart a parallel AMR computation. Example 6p has been extended to
demonstrate restarting from a previously saved checkpoint. Note that parallel
NC data files are compatible with serial code, e.g., can be viewed with serial
GLVis. Loading of legacy NC mesh files is still supported.
- Added FMS support (https://github.com/CEED/FMS) to mfem. FMS can represent
unstructured high-order meshes with general high-order finite element fields
on them. When enabled, mfem can convert data collections to/from FMS data
collections in memory. In addition, an FMS data collection class was added so
the convert-dc miniapp can read and generate data files in FMS format.
- Added new mesh quality metrics and improved the untangling capabilities of the
TMOP-based mesh optimization algorithms.
- The TMOP mesh optimization algorithms were extended to GPU:
* QualityMetric 1, 2, 7, 77 are available in 2D, 302, 303, 315, 321 in 3D
* Both AnalyticAdaptTC and DiscreteAdaptTC TargetConstructor are available
* Kernels for normalization and limiting have been added
* The AdvectorCG now also supports AssemblyLevel::PARTIAL
- Added support for creating refined meshes for all element types (e.g. by
splitting high-order elements into low-order refined elements), including
mixed meshes. The LOR Transfer miniapp (miniapps/tools/lor-transfer.cpp) now
supports meshes with any element geometry.
- Meshes consisting of any type of elements (including mixed meshes) can be
converted to all-simplex meshes using Mesh::MakeSimplicial.
- Several of the mesh constructors (creating Cartesian meshes, refined (LOR)
meshes, simplex meshes, etc.) are now available as "named constructors", e.g.
Mesh::MakeCartesian2D or Mesh::MakeRefined. The legacy constructors are marked
as deprecated.
- Added support for creating periodic meshes with Mesh::MakePeriodic. The
requisite periodic vertex mappings can be created with
Mesh::CreatePeriodicVertexMapping.
- Added support for 1D non-conforming meshes (which can be useful for parallel
load balancing and derefinement).
- Added sample meshes in the `data` subdirectory showing the reference elements
of the six currently supported element types; ref-segment.mesh,
ref-triangle.mesh, ref-square.mesh, ref-tetrahedron.mesh, ref-cube.mesh, and
ref-prism.mesh.
High-performance computing
--------------------------
- Added initial support for GPU-accelerated versions of PETSc that works with
MFEM_USE_CUDA if PETSc has been configured with CUDA support. Examples 1 and 9
in the examples/petsc directory have been modified to work with --device cuda.
Examples with GAMG (ex1p) and SLEPc (ex11p) are also provided.
- Added support for explicit vectorization in the high-performance templated
code for Fujitsu's A64FX ARM microprocessor architecture.
- Added support for different modes of QuadratureInterpolator on GPU.
The layout (QVectorLayout::byNODES|byVDIM) and the tensor products modes can
be enabled before calling the Mult, Values, Derivatives, PhysDerivatives and
Determinants methods.
- Added method Device::SetMemoryTypes that can be used to change the default
host and device MemoryTypes before Device setup.
- In class MemoryManager, added methods GetDualMemoryType and SetDualMemoryType;
dual MemoryTypes are used to determine the second MemoryType (host or device)
when only one MemoryType is specified in methods of class Memory.
- Added Memory constructor for setting both the host and device MemoryTypes.
- Switched the default behavior of device memory allocations so that they are
deferred until the device pointer is needed.
- Added a second Umpire device MemoryType, DEVICE_UMPIRE_2, with corresponding
allocator that can be set with the method SetUmpireDevice2AllocatorName.
- Added HOST_PINNED MemoryType and a pinned host allocator for CUDA and HIP.
- Added matrix-free GPU-enabled implementations of GradientInterpolator and
IdentityInterpolator.
New and updated examples and miniapps
-------------------------------------
- Added a new, very simple example (ex0 and parallel version ex0p). This example
solves a simple Poisson problem using H1 elements (the same problem as ex1),
but is intended to be extremely simple and approachable for new users.
- Added new miniapps demonstrating: 1) the use of GSLIB for overlapping grids,
see gslib/schwarz_ex1, and 2) coupling different physics in different domains,
see navier/cht. Note that gslib v1.0.7 is require (see INSTALL for details).
- Added a new miniapp for computing (signed) distance functions to a point
source or zero level set. See miniapps/shifted/distance.cpp.
- Added a high-order extension of the shifted boundary method to solve PDEs on
non body-fitted meshes. This is illustrated in the new Shifted Diffusion
miniapp, see miniapps/shifted/diffusion.cpp.
- Added new miniapp directory mtop/ with optimization-oriented block parametric
non-linear form and abstract integrators. Two new miniapps, ParHeat and
SeqHeat, demonstrate parallel and sequential implementation of gradients
evaluation for linear diffusion with discrete density.
- Added a new miniapp block-solvers that compares the performance of various
solvers for mixed finite element discretization of the second order scalar
elliptic equations. Currently available solvers in the miniapp include a
block-diagonal preconditioner that is based on approximate Schur complement
(implemented in ex5p), and a newly implemented solver DivFreeSolver, which
exploits a multilevel decomposition of the Raviart-Thomas space and its
divergence-free subspace. See the miniapps/solvers directory for more details.
- Introduced new options for the mesh-explorer miniapp to visualize the actual
element attributes in parallel meshes while retaining the visualization of the
domain decomposition.
- Added partial assembly and device support to Example 25/25p, with diagonal
preconditioning.
- Implemented a filter method for the Navier miniapp to stabilize highly
turbulent flows in direct numerical simulation.
Improved testing
----------------
- Transitioned from Travis to GitHub Action for testing/CI on GitHub.
- Use Spack (and Uberenv) to automate TPL building in LLNL GitLab tests.
- Extended `make test` to include GPU tests when MFEM is built with CUDA or HIP
support.
- Added a set of suggested git hooks for developers in config/githooks.
- Added support for Caliper: a library to integrate performance profiling
capabilities into applications. See examples/caliper for more details.
- Added a new command line boolean option (`--all`) to the unit tests to launch
*all* non-regression tests.
- Upgraded the Catch unit test framework from version 2.13.0 to version 2.13.2.
Miscellaneous
-------------
- The following integrations have updated minimum version requirements:
* CUDA >= 10.1.168
* Ginkgo >= 1.4.0
* GSLIB >= 1.0.7
* HIOP >= 0.4
* HYPRE >= 2.20.0 for mixedint support
* HYPRE >= 2.22.0 for CUDA support
* libCEED >= 0.8
* PETSc >= 3.15.0 for CUDA support
* RAJA >= 0.13.0
see INSTALL for more details.
- Added a "scaled Jacobian" visualization option in the Mesh Explorer miniapp to
help identify elements with poor mesh quality.
- Added support for reading VTK meshes in XML format.
- Added makefile rule to generate TAGS table for vi or Emacs users.
- Added HIP support to the CMake build system.
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- Added an abstract interface `mfem::FaceRestriction` for `H1FaceRestriction`
and `L2FaceRestriction`.
In order to conform with the semantic of `MultTranspose` in `mfem::Operator`,
`mfem::FaceRestriction::MultTranspose` now sets instead of adding values, and
`mfem::FaceRestriction::AddMultTranspose` should replace previous calls to
`mfem::FaceRestriction::MultTranspose`.
Version 4.2, released on October 30, 2020
=========================================
High-performance computing
--------------------------
- Added support for explicit vectorization in the high-performance templated
code, which can now take advantage of specific classes on the following
architectures:
* x86 (SSE/AVX/AVX2/AVX512),
* Power8 & Power9 (VSX),
* BG/Q (QPX).
These are disabled by default, but can be enabled with MFEM_USE_SIMD=YES.
See the new file linalg/simd.hpp and the new directory linalg/simd.
- Added an Element Assembly mode compatible with GPU device execution for H1 and
L2 spaces in the mass, convection, diffusion, transpose, and the face DG trace
integrators. See option '-ea' in Example 9. When enabled, this assembly level
stores independent dense matrices for the elements, and independent dense
matrices for the faces in the DG case.
- Added a Full Assembly mode compatible with GPU device execution. This assembly
level builds on top of the Element Assembly kernels to compute a global sparse
matrix. All integrators supported by element assembly are also supported by
full assembly. See the '-fa' option in Example 9.
- Optimized the AMD/HIP kernel support and enabled HIP support in the libCEED
integration. This is now available via the "ceed-hip" device backend.
- Improved the libCEED integration to support:
* AssemblyLevel::NONE for Mass, Diffusion, VectorMass, and VectorDiffusion
Integrators. This level computes the full operator evaluation "on the fly".
* VectorMassIntegrator and VectorDiffusionIntegrator.
* All types of (scalar) Coefficients.
- Added partial assembly / device support for:
* H(div) bilinear forms and VectorFEDivergenceIntegrator.
* BlockOperator, see the updated Example 5.
* Complex operators, see the updated Example 22.
* Chebyshev accelerated polynomial smoother.
* Convergent diagonal preconditioning on non-conforming adaptively refined
meshes, see Example 6/6p.
- Added CUDA support for:
* Sparse matrix-vector multiplication with cuSPARSE,
* SUNDIALS ODE integrators, see updated SUNDIALS modification of Example 9/9p.
Linear and nonlinear solvers
----------------------------
- Added a new solver class for simple integration with NVIDIA's multigrid
library, AmgX. The AmgX class is designed to work as a standalone solver or
preconditioner for existing MFEM solvers. It uses MFEM's sparse matrix format
for serial runs and the HypreParMatrix format for parallel runs. The new
solver may be configured to run with one GPU per MPI rank or with more MPI
ranks than GPUs. In the latter case, matrices and vectors are consolidated to
ranks communicating with the GPUs and the solution is then broadcasted.
Although CUDA is required to build, the AmgX support is compatible with the
MFEM CPU device configuration. The examples/amgx folder illustrates how to
integrate AmgX in existing MFEM applications. The AmgX solver class is
partially based on: "AmgXWrapper: An interface between PETSc and the NVIDIA
AmgX library", by Pi-Yueh Chuang and Lorena A. Barba, doi:10.21105/joss.00280.
- Added initial support for geometric h- and p-multigrid preconditioners for
matrix-based and matrix-free discretizations with basic GPU capability, see
Example 26/26p.
- Added support for the CVODES package in SUNDIALS which provides ODE solvers
with sensitivity analysis capabilities. See the CVODESSolver class and the new
adjoint miniapps in the miniapps/adjoint directory.
- Added an interface to the MKL CPardiso solver, an MPI-parallel sparse direct
solver developed by Intel. See Example 11p for an illustration of its usage.
- Added support for the SLEPc eigensolver package, https://slepc.upv.es.
- Upgraded SuperLU interface to use SuperLU_DIST 6.3.1. Added a simple SuperLU
example in the new directory examples/superlu.
- Extended the KINSOL (SUNDIALS) nonlinear solver interface to support the
Jacobian-free Newton-Krylov method. A usage example is shown in Example 10p.
- Block arrays of parallel matrices can now be merged into a single parallel
matrix with the function HypreParMatrixFromBlocks. This could be useful for
solving block systems with parallel direct solvers such as STRUMPACK.
- Added wrappers for hypre's flexible GMRES solver and the new parallel ILU
preconditioner. The latter requires hypre version 2.19.0 or later.
Discretization improvements
---------------------------
- Extended GSLIB-FindPoints integration to support simplices and interpolation
of functions from L2, H(div) and H(curl) spaces.
- Added support for computing asymptotic error estimates and convergence rates
for the whole de Rham sequence based on the new class ConvergenceStudy and new
member methods in GridFunction and ParGridFunction. See the rates.cpp file in
the tests/convergence directory for sample usage.
- Extended the GetValue and GetVectorValue methods of GridFunction to support
evaluation on boundary elements and, in the continuous field case, arbitrary
mesh edges and faces. This requires passing an ElementTransformation argument.
- Added support for matrix-free interpolation and restriction operators between
continuous H1 finite element spaces of different order on the same mesh or
with the same order on uniformly refined meshes.
- The Coefficient classes based on a C-function pointer (FunctionCoefficient,
VectorFunctionCoefficient and MatrixFunctionCoefficient) now use the more
general std::function class template. This allows the classes to be backward
compatible (i.e. they can still work with C-functions) and, in addition,
support any "callable", e.g. lambda functions.
- Non-conforming meshes are now supported with block nonlinear forms. See the
updated Example 19/19p.
Meshing improvements
--------------------
- The graph linear ordering library Gecko, previously an external dependency, is
now included directly in MFEM. As a result, Mesh::GetGeckoElementOrdering is
always available. The interface has also been improved, see for example the
Mesh Explorer miniapp.
- Added support for finite difference-based gradient and Hessian approximation
in the TMOP mesh optimization algorithms. This improves the accuracy of the
Hessian for r-adaptivity using discrete fields, and allows use of skewness
and orientation based metrics.
- Improved Gmsh reader (version 2.2), which now supports both high-order and
periodic meshes. Segments, triangles, quadrilaterals, and tetrahedra are
supported up to order 10. Wedges and hexahedra are supported up to order 9.
For sample periodic meshes, see the periodic*.msh files in the data directory.
- Added support for construction of (serial) non-conforming meshes. Hanging
nodes can be marked with Mesh::AddVertexParents when building the mesh with
the "init" constructor. The usage is demonstrated in a new meshing miniapp,
Polar NC, which generates meshes that are non-conforming from the start.
- Added support for r-adaptivity with more than one discrete field. This allows
the user to specify different discrete functions for controlling the size,
aspect-ratio, orientation, and skew of elements in the mesh.
- Additional TMOP improvements:
* Capability for approximate tangential mesh relaxation.
* Support and examples for using TMOP on mixed meshes.
* Complete integrator action accounting for spatial derivatives of discrete
and analytic targets.
New and updated examples and miniapps
-------------------------------------
- Added a new miniapp, Navier, that solves the time-dependent Navier-Stokes
equations of incompressible fluid dynamics. See the miniapps/navier directory
for more details.
- Added 10 new example codes:
* Example 25/25p demonstrates the use of a Perfectly Matched Layer (PML) for
electromagnetic wave propagation (indefinite Maxwell).
* Example 26/26p shows how to construct matrix-free geometric and p-multigrid
preconditioner for the Laplace problem.
* Example 27/27p demonstrates the enforcement of Dirichlet, Neumann, Robin,
and periodic boundary conditions with either H1 or DG Laplace problems.
* Versions of Example 1/1p in examples/amgx demonstrating the use of AmgX,
to solve the Laplace problem with AMG preconditioning on GPUs.
* A version of Example 11p in examples/petsc demonstrating the use of SLEPc,
to solve the Laplace eigenproblem with shift-and-invert transformation.
* A version of Example 1 in examples/superlu demonstrating the use of SuperLU
to solve the Laplace problem.
- Added a new Field Interpolation miniapp in miniapps/gslib that demonstrates
transfer of grid functions between different meshes using GSLIB-FindPoints.
- Added 2 miniapps in the new miniapps/adjoint directory demonstrating how to
solve adjoint problems in MFEM using the CVODES package in SUNDIALS. Both of
these miniapps require the MFEM_USE_SUNDIALS configuration option.
* The cvsRoberts_ASAi_dns miniapp solves a backward adjoint problem for a
system of ODEs, evaluating both forward and adjoint quadratures in serial.
* The adjoint_advection_diffusion miniapp solves a backward adjoint problem
for an advection diffusion PDE, evaluating adjoint quadratures in parallel.
- Added 4 additional meshing miniapps:
* The Minimal Surface miniapp solves Plateau's problem: the Dirichlet problem
for the minimal surface equation.
* The Twist miniapp demonstrates how to stitch together opposite surfaces of a
mesh to create a topologically periodic mesh.
* The Trimmer miniapp trims away parts of a mesh based on element attributes.
* Polar NC shows the construction of polar non-conforming meshes.
- Several examples and miniapps were updated to include:
* Full and element assembly support in Example 9/9p.
* Partial assembly with diagonal preconditioning in Examples 4/4p/5/5p/22/22p.
* Diagonal preconditioner in Example 6/6p for partial assembly with AMR.
* The option to plot a function in Mesh Explorer.
* A new test problem showing a mixed bilinear form for H1, H(curl), H(div) and
L2, with partial assembly support in Example 24/24p.
* Weak Dirichlet boundary conditions (Nitsche) to the NURBS miniapp.
Data management and visualization
---------------------------------
- Added support for ADIOS2 for parallel I/O with ParaView visualization. See
Examples 5, 9, 12, 16. The classes adios2stream and ADIOS2DataCollection
provide the interface to generate ADIOS2 Binary Pack (BP4) directory datasets.
- Added VTU output of boundary elements and attributes and parallel VTU (PVTU)
output of parallel meshes for visualization using ParaView.
Improved testing
----------------
- Added a GitLab pipeline that automates PR testing on supercomputing systems
and Linux clusters at Lawrence Livermore National Lab (LLNL). This can be
triggered only by LLNL developers, see .gitlab-ci.yml, the .gitlab directory
and the updated CONTRIBUTING.md file.
- Added additional testing for convergence, the parallel mesh I/O, and for the
libCEED integration in MFEM in the tests/ directory.
- Upgraded the Catch unit test framework from version 1.6.1 to version 2.13.0.
Miscellaneous
-------------
- Renamed "Backend::DEBUG" to "Backend::DEBUG_DEVICE" to avoid conflicts,
as DEBUG is sometimes used as a macro.
- Added a new IterativeSolverMonitor class that allows to monitor the residual
and solution with an IterativeSolver after every iteration.
- Added power method to iteratively estimate the largest eigenvalue and the
corresponding eigenvector of an operator.
- Added support for face integrals on the boundaries of NURBS meshes.
- In SLISolver, changed the residual inner product from (Br,r) to (Br,Br) so the
solver can work with non-SPD preconditioner B.
- Added new coefficient and vector coefficient classes for QuadratureFunctions,
with new LinearForm integrators which use them.
- The integration order used in the ComputeLpError and ComputeElementLpError
methods of class GridFunction has been increased.
- Change the IntegrationRule inside VectorDiffusionIntegrator to use the same
quadrature as DiffusionIntegrator.
- The README.html files previously included in several source directories have
been removed. Use the corresponding pages at mfem.org instead.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.1, released on March 10, 2020
=======================================
Starting with this version, the MFEM open source license is changed to BSD-3.
Improved GPU capabilities
-------------------------
- Added initial support for AMD GPUs based on HIP: a C++ runtime API and kernel
language that can run on both AMD and NVIDIA hardware.
- Added support for Umpire, a resource management library that allows the
discovery, provision, and management of memory on machines with multiple
memory devices like NUMA and GPUs, see https://github.com/LLNL/Umpire.
- GPU acceleration is now available in 3 additional examples: 3, 9 and 24.
- Improved RAJA backend and multi-GPU MPI communications.
- Added a "debug" device designed specifically to aid in debugging GPU code by
following the "device" code path (using separate host/device memory spaces and
host <-> device transfers) without any GPU hardware.
- Added support for matrix-free diagonal smoothers on GPUs.
- The current list of available device backends is: "ceed-cuda", "occa-cuda",
"raja-cuda", "cuda", "hip", "debug", "occa-omp", "raja-omp", "omp",
"ceed-cpu", "occa-cpu", "raja-cpu", and "cpu".
- The MFEM memory manager now supports different memory types, associated with
the following memory backends:
* Default host memory, using standard C++ new and delete,
* CUDA pointers, using cudaMalloc and HIP pointers, using hipMalloc,
* Managed CUDA/HIP memory (UVM), using cudaMallocManaged/hipMallocManaged,
* Umpire-managed memory, including memory pools,
* 32- or 64-byte aligned memory, using posix_memalign (WIN32 also supported),
* Debug memory with mmap/mprotect protection used by the new "debug" device.
libCEED support
---------------
- Added support for libCEED, the portable library for high-order operator
evaluation developed by the Center for Efficient Exascale Discretizations in
the Exascale Computing Project, https://github.com/CEED/libCEED.
- This initial integration includes Mass and Diffusion integrators. libCEED GPU
backends can be used without specific MFEM configuration, however it is highly
recommended to use the "cuda" build option to minimize memory transfers.
- Both CPU and GPU modes are available as MFEM device backends (ceed-cpu and
ceed-cuda), using some of the best performing CPU and GPU backends from
libCEED, see the sample runs in examples 1 and 6.
- NOTE: The current default libCEED GPU backend (ceed-cuda) uses atomics and
therefore is non-deterministic.
Partial assembly and matrix-free discretizations
------------------------------------------------
- The support for matrix-free methods on both CPU and GPU devices based on a
partially assembled operator decomposition was extended to include:
* DG integrators, (for now only in the Gauss-Lobatto basis), see Example 9,
* H(curl) bilinear forms, see Example 3,
* vector mass and vector diffusion bilinear integrators,
* convection integrator with improved performance,
* gradient and vector divergence integrators for Stokes problems,
* initial partial assembly mode for NonlinearForms.
- Diagonals of partially assembled operators can now be computed efficiently.
See the new methods AssembleDiagonal in BilinearForm, AssembleDiagonalPA in
BilinearFormIntegrator and the implementations in fem/bilininteg_*.cpp.
- In many examples, the partial assembly algorithms provide significantly
improved performance, particularly in high-order 3D runs on GPUs.
Meshing improvements
--------------------
- The algorithms for mesh element numbering were changed to have significantly
better caching and parallel partitioning properties. Both initial (see e.g.
Mesh::GetHilbertElementOrdering) and ordering after uniform refinement were
improved. NOTE: new ordering can have a round-off effect on solver results.
- Added support for non-conforming AMR on both prisms and tetrahedra, including
coarsening and parallel load balancing. Anisotropic prism refinement is only
available in the serial version at the moment.
- The TMOP mesh optimization algorithms were extended to support r-adaptivity.
Target matrices can now be constructed either via a given analytical function
(e.g. spatial dependence of size, aspect ratio, etc., for each element) or via
a (Par)GridFunction specified on the original mesh.
- The TMOP algorithms were also improved to support non-conforming AMR meshes.
- Added support for creating refined versions of periodic meshes, making use of
the new L2ElementRestriction class. This class also allows for computing
geometric factors on periodic meshes using partial assembly.
Discretization improvements
---------------------------
- Added support for GSLIB-FindPoints, a general high-order interpolation utility
that can robustly evaluate a GridFunction in an arbitrary collection of points
in physical space. See INSTALL for details on building MFEM with GSLIB, and
miniapps/gslib for examples of how to use this feature.
- Added support for complex-valued finite element operators and fields using a
2x2 block structured linear system to mimic complex arithmetic. New classes
include: ComplexGridFunction, SesquilinearForm, ComplexLinearForm, and their
parallel counterparts.
- Added second order derivatives of NURBS shape functions.
- Added support for serendipity elements of arbitrary order on affinely-mapped
square elements. Basis functions for these elements can be visualized using
an option in the display-basis miniapp.
- Two integrators related to Stokes problems, (Q grad u, v) and (Q div v, u),
where u and the components of v are in H1, were added/modified to support full
and partial assembly modes. See the new GradientIntegrator and the updated
VectorDivergenceIntegrator classes in fem/bilininteg.hpp, as well as the PA
kernels in fem/bilininteg_gradient.cpp and fem/bilininteg_divergence.cpp.
- Added a nonlinear vector valued convection integrator (Q u \cdot grad u, v)
where u_i and v_i are in H1. This form occurs e.g. in the Navier-Stokes
equations. The integrator supports the partial assembly mode for its
action. In full assembly mode we also provide the GetGradient method that
computes the linearized version of the integrator.
- Added a new method, MixedBilinearForm::FormRectangularLinearSystem, that can
be used to impose boundary conditions on the non-square off-diagonal blocks of
a block operator (similar to FormLinearSystem in the square case).
Linear and nonlinear solvers
----------------------------
- Added support for Ginkgo, a high-performance linear algebra library for GPU
and manycore nodes, with a focus on sparse solution of linear systems. For
more details see linalg/ginkgo.hpp and the example code in examples/gingko.
- Added support for HiOp, a lightweight HPC solver for nonlinear optimization
problems, see class HiOpNLPOptimizer and the example codes in examples/hiop.
- Added a general interface for specifying and solving nonlinear constrained
optimization problems through the new classes OptimizationProblem and
OptimizationSolver, see linalg/solver.hpp.
- Added a block ILU(0) preconditioner for DG-type discretizations. Example 9
(DG advection) now takes advantage of this for implicit time integration.
- New time integrators: Adams-Bashforth, Adams-Moulton and several integrators
for 2nd order ODEs, see the new Example 23.
- Added a LinearSolve(A,X) convenience method to solve dense linear systems. In
the trivial cases, i.e., square matrices of size 1 or 2, the system is solved
directly, otherwise, LU factorization is employed.
New and updated examples and miniapps
-------------------------------------
- Added a collection of 7 playful miniapps in miniapps/toys that illustrate the
meshing and visualization features of the library in more relaxed settings.
The toys include simulations of cellular automata, Rubik's cube, Mandelbrot
set, a tool to convert any image to mfem mesh, and more.
- Added 8 new example codes:
* Example 22/22p demonstrates the use of the new complex-valued finite element
operators by defining and solving a family of time-harmonic PDEs related to
damped harmonic oscillators.
* Example 23 solves a simple 2D/3D wave equation with the new second order
time integrators.
* Example 24/24p demonstrates usage of mixed finite element spaces in bilinear
forms. Partial assembly is supported in this example.
* A version of Example 1 in examples/ginkgo demonstrating the use of the
Gingko interface to solve a linear system.
* A version of Example 9/9p in examples/hiop demonstrating the nonlinear
constrained optimization interface and use of the SLBQP and HiOp solvers.
- Added two new miniapps: Find Points and Field Diff in miniapps/gslib that show
how GSLIB-FindPoints can be used to interpolate a (Par) GridFunction in an
arbitrary number of physical space points in 2D and 3D. The GridFunction must
be in H1 and in the same space as the mesh that is used to find the points.
- Added a simple miniapp, Get Values, that extracts field values at a set of
points, from previously saved data via DataCollection classes.
- Several examples and miniapps were updated:
* Added device support in Example 3/3p and Example 9/9p.
* Example 1/1p and Example 3/3p now use diagonal preconditioning in partial
assembly mode.
* Example 9/9p now supports implicit time integration, using the new block
ILU(0) solvers as preconditioners for the linear system.
* The mesh-optimizer and pmesh-optimizer miniapps now include the new
r-adaptivity capabilities of TMOP. They were also updated to support mesh
optimization on non-conforming AMR meshes.
* New options to reorder and partition the mesh and boundary attribute
visualization (key 'b') are now available in the mesh-explorer miniapp.
- Collected object files from the miniapps/common directory into a new library,
libmfem-common for the convenience of application developers. The new library
is now used in several miniapps in the electromagnetic and tools directories.
Improved testing
----------------
- Added a large number of unit tests in the tests/unit directory, including
several parallel unit tests.
- Added a new directory, tests/scripts, with several shell scripts that perform
simple checks on the code including: code styling, documentation formatting,
proper use of .gitignore, and preventing the accidental commit of large files.
- It is recommended that developers run the above tests scripts (via the runtest
script) before pushing to GitHub. See the README file in tests/scripts.
- The Travis CI settings have been updated to include an initial Checks stage
which currently runs the code-style, documentation and gitignore test scripts,
as well as a final stage for optional checks/tests which currently runs the
branch-history script.
Miscellaneous
-------------
- Added support for output in the ParaView XML format. Both low-order and
high-order Lagrange elements are supported. Output can be in ASCII or binary
format. The binary output can be compressed if MFEM is compiled with zlib
support (MFEM_USE_ZLIB). See the new ParaViewDataCollection class and the
updated Examples 5/5p and 9/9p.
- Upgraded the SUNDIALS interface to utilize SUNDIALS 5.0. This necessitated a
complete rework of the interface and requires changes at the application
level. Example usage of the new interface can be found in examples/sundials.
- Switched from gzstream to zstr for the implementation of zlib-compressed C++
output stream. The build system definition now uses MFEM_USE_ZLIB instead of
MFEM_USE_GZSTREAM, but the code interface (e.g. ofgzstream) remains the same.
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- In the enum classes MemoryType and MemoryClass, "CUDA" was renamed to "DEVICE"
which now denotes either "CUDA" or "HIP" depending on the build configuration.
In the same enum classes, "CUDA_UVM" was renamed to "MANAGED".
Version 4.0, released on May 24, 2019
=====================================
Unlike previous MFEM releases, this version requires a C++11 compiler.
GPU support
-----------
- Added initial support for hardware devices, such as GPUs, and programming
models, such as CUDA, OCCA, RAJA and OpenMP.
- The GPU/device support is based on MFEM's new backends and kernels working
seamlessly with a new lightweight device/host memory manager. The kernels can
be implemented either in OCCA, or as a simple wrapper around for-loops, which
can then be dispatched to RAJA and native backends. See the files forall.hpp
and mem_manager.hpp in the general/ directory for more details.
- Several of the MFEM example codes (ex1, ex1p, ex6, and ex6p) can now take
advantage of GPU acceleration with the backend selectable at runtime. Many of
the linear algebra and finite element operations (e.g. partially assembled
bilinear forms) have been extended to take advantage of kernel acceleration by
simply replacing loops with the MFEM_FORALL() macro.
- In addition to native CUDA kernels, the library currently supports OCCA, RAJA
and OpenMP kernels, which could be mixed and matched in different parts of the
same application. We plan on adding support for more programming models and
devices in the future, without the need for significant modifications in user
code. The list of current backends is: "occa-cuda", "raja-cuda", "cuda",
"occa-omp", "raja-omp", "omp", "occa-cpu", "raja-cpu", and "cpu".
- GPU-related limitations:
* Hypre preconditioners are not yet available in GPU mode, and in particular
hypre must be built in CPU mode.
* Only constant coefficients are currently supported on GPUs.
* Optimized element assembly, and matrix-free bilinear forms are not
implemented yet. Element batching is currently ignored.
* In device mode, full assembly is performed on the host (but the matvec
action is performed on the device).
* Partial assembly kernels are not implemented yet for simplices.
Discretization improvements
---------------------------
- Partial assembled finite element operators are now available in the core
library, based on the new classes PABilinearFormExtension, ElementRestriction,
DofToQuad and GeometricFactors (associated with the classes BilinearForm,
FiniteElementSpace, FiniteElement and Mesh, respectively). The kernels for
partial assembled Setup/Assembly and Action/Mult are implemented in the
BilinearFormIntegrator methods AssemblePA and AddMultPA.
- Added support for a general "low-order refined"-to-"high-order" transfer of
GridFunction data from a "low-order refined" (LOR) space defined on a refined
mesh to a "high-order" (HO) finite element space defined on a coarse mesh. See
the new classes InterpolationGridTransfer and L2ProjectionGridTransfer and the
new LOR Transfer miniapp: miniapps/tools/lor-transfer.cpp.
- Added element flux, and flux energy computation in class ElasticityIntegrator,
allowing for the use of Zienkiewicz-Zhu type error estimators with the
integrator. For an illustration of this addition, see the new Example 21.
- Added support for derefinement of vector (RT + ND) spaces.
- Added a variety of coefficients which are sums or products of existing
coefficients as well as grid function coefficients which return the
divergence, gradient, or curl of their GridFunctions.
Support for wedge elements and meshes with mixed element types
--------------------------------------------------------------
- Added support for wedge-shaped mesh elements of arbitrary order (with Geometry
type PRISM) which have two triangular faces and three quadrilateral faces.
Several examples of such meshes can be found in the data/ directory.
- Added support for mixed meshes containing triangles and quadrilaterals in 2D
or tetrahedra, wedges, and hexahedra in 3D. This includes support for uniform
refinement of such meshes. Several examples of such meshes can be found in the
data/ directory.
- Added H1 and L2 finite elements of arbitrary order for Wedge elements.
- Added support for reading and writing linear and quadratic meshes containing
wedge elements in VTK mesh format. Several examples of such meshes can be
found in the data/ directory.
Other meshing improvements
--------------------------
- Improved the uniform refinement of tetrahedral meshes (also part of the
uniform refinement of mixed 3D meshes). The previous refinement algorithm is
still available as an option in Mesh::UniformRefinement. Both can be used in
the updated Mesh Explorer miniapp.
- The local tetrahedral mesh refinement algorithm in serial and in parallel now
follows precisely the paper:
D. Arnold, A. Mukherjee, and L. Pouly, "Locally Adapted Tetrahedral Meshes
Using Bisection", SIAM J. Sci. Comput. 22 (2000), 431-448.
This guarantees that the shape regularity of the elements will be preserved
under refinement.
- The TMOP mesh optimization algorithms were extended to support user-defined
space-dependent limiting terms. Improved the TMOP objective functions by more
accurate normalization of the different terms.
- Added support for parallel communication groups on non-conforming meshes.
- Improved parallel partitioning of non-conforming meshes. If the coarse mesh
elements are ordered as a sequence of face-neighbors, the parallel partitions
are now guaranteed to be continuous. To that end, inline quadrilateral and
hexahedral meshes are now by default ordered along a space-filling curve.
- A boundary in a NURBS mesh can now be connected with another boundary. Such a
periodic NURBS mesh is a simple way to impose periodic boundary conditions.
- Added support for reading linear and quadratic 2D quadrilateral and triangular
Cubit meshes.
New and updated examples and miniapps
-------------------------------------
- Added a new meshing miniapp, Toroid, which can produce a variety of torus
shaped meshes by twisting a stack of wedges or hexahedra.
- Added a new meshing miniapp, Extruder, that demonstrates the capability to
produce 3D meshes by extruding 2D meshes.
- Added a simple miniapp, LOR Transfer, for visualizing the actions of the
transfer operators between a high-order and a low-order refined spaces.
- Added a new example, Example 20/20p, that solves a system of 1D ODEs derived
from a Hamiltonian. The example demonstrates the use of the variable order,
symplectic integration algorithm implemented in class SIAVSolver.
- Added a new example, Example 21/21p, that illustrates the use of AMR to solve
a linear elasticity problem. This is an extension of Example 2/2p.
New and improved solvers and preconditioners
--------------------------------------------
- Added support for parallel ILU preconditioning via hypre's Euclid solver.
- Added support for STRUMPACK v3 with a small API change in the class
STRUMPACKSolver, see "API changes" below.
Miscellaneous
-------------
- Added unit tests based on the Catch++ library in the test/ directory.
- Renamed the option MFEM_USE_OPENMP to MFEM_USE_LEGACY_OPENMP. This legacy
option is deprecated and planned for removal in a future release. The original
option name, MFEM_USE_OPENMP, is now used to enable the new OpenMP backends in
the new kernels.
- In SparseMatrix added the option to perform MultTranspose() by matvec with
computed and stored transpose matrix. This is required for deterministic
results when using devices such as CUDA and OpenMP.
- Altered the way FGMRES counts its iterations so that it matches GMRES.
- Various other simplifications, extensions, and bugfixes in the code.
- Construct abstract parallel rectangular truedof-to-truedof operators via
Operator::FormDiscreteOperator().
API changes
-----------
- In multiple places, use Geometry::Type instead of int, where appropriate.
- In multiple places, use Element::Type instead of int, where appropriate.
- The Mesh methods GetElementBaseGeometry and GetBdrElementBaseGeometry no
longer have a default value for their parameter, they only work with an
explicitly given index.
- In class Mesh, added methods useful for queries regarding the types of
elements present in the mesh: HasGeometry, GetNumGeometries, GetGeometries,
and class Mesh::GeometryList.
- The struct CoarseFineTransformations (returned by the Mesh method
GetRefinementTransforms) now stores the embedding matrices separately for each
Geometry::Type.
- In class ParMesh, replaced the method GroupNFaces with two new methods:
GroupNTriangles and GroupNQuadrilaterals. Also, replaced GroupFace with two
methods: GroupTriangle and GroupQuadrilateral.
- In class ParMesh, made the two RefineGroups methods protected.
- Removed the virtual method Element::GetRefinementFlag, it is only used by the
derived class Tetrahedron.
- Added new methods: Array::CopyTo, Tetrahedron::Init.
- In class STRUMPACKSolver, the method SetMC64Job() was replaced by the new
methods: DisableMatching(), EnableMatching(), and EnableParallelMatching().
Version 3.4, released on May 29, 2018
=====================================
More general and efficient mesh adaptivity
------------------------------------------
- Added support for PUMI, the Parallel Unstructured Mesh Infrastructure from
https://scorec.rpi.edu/pumi. PUMI is an unstructured, distributed mesh data
management system that is capable of handling general non-manifold models and
effectively supports automated adaptive analysis. PUMI enables for the first
time support for parallel unstructured modifications of MFEM meshes.
- Significantly reduced MPI communication in the construction of the parallel
prolongation matrix in ParFiniteElementSpace, for much improved parallel
scaling of non-conforming AMR on hundreds of thousands of MPI tasks. The
memory footprint of the ParNCMesh class has also been reduced.
- In FiniteElementSpace, the fully assembled refinement matrix is now replaced
by default by a specialized refinement operator. The operator option is both
faster and more memory efficient than using the fully assembled matrix. The
old approach is still available and can be enabled, if needed, using the new
method FiniteElementSpace::SetUpdateOperatorType().
Discretization improvements
---------------------------
- Added support for a general "high-order"-to-"low-order refined" transfer of
GridFunction and true-dof data from a "high-order" finite element space
defined on a coarse mesh, to a "low-order refined" space defined on a refined
mesh. The new methods, GetTransferOperator and GetTrueTransferOperator in the
FiniteElementSpace classes, work in both serial and parallel and support
matrix-based as well as matrix-free transfer operator representations. They
use a new method, GetTransferMatrix, in the FiniteElement class similar to
GetLocalInterpolation, that allows the coarse FiniteElement to be different
from the fine FiniteElement.
- Added class ComplexOperator, that implements the action of a complex operator
through the equivalent 2x2 real formulation. Both symmetric and antisymmetric
block structures are supported.
- Added classes for general block nonlinear finite element operators (deriving
from BlockNonlinearForm and ParBlockNonlinearForm) enabling solution of
nonlinear systems with multiple unknowns in different function spaces. Such
operators have assemble-based action and also support assembly of the gradient
operator to enable inversion with Newton iteration.
- Added variable order NURBS: for each space each knot vector in the mesh can
have a different order. The order information is now part of the finite
element space header in the NURBS mesh output, so NURBS meshes in the old
format need to be updated.
- In the classes NonlinearForm and ParNonlinearForm, added support for
non-conforming AMR meshes; see also the "API changes" section.
- New specialized time integrators: symplectic integrators of orders 1-4 for
systems of first order ODEs derived from a Hamiltonian and generalized-alpha
ODE solver for the filtered Navier-Stokes equations with stabilization. See
classes SIASolver and GeneralizedAlphaSolver in linalg/ode.hpp.
- Inherit finite element classes from the new base class TensorBasisElement,
whenever the basis can be represented by a tensor product of 1D bases.
- Added support for elimination of boundary conditions in block matrices.
New and updated examples and miniapps
-------------------------------------
- Added a new serial and parallel example (ex19) that solves the quasi-static
incompressible hyperelastic equations. The example demonstrates the use of
block nonlinear forms as well as custom block preconditioners.
- Added a new serial example (ex23) to demonstrate the use of second order
time integration to solve the wave equation.
- Added a new electromagnetics miniapp, Maxwell, for simulating time-domain
electromagnetics phenomena as a coupled first order system of equations.
- A simple local refinement option has been added to the mesh-explorer miniapp
(menu option 'r', sub-option 'l') that selects elements for refinement based
on their spatial location - see the function 'region()' in the source file.
- Added a set of miniapps specifically focused on Isogeometric Analysis (IGA) on
NURBS meshes in the miniapps/nurbs directory. Currently the directory contains
variable order NURBS versions of examples 1, 1p and 11p.
- Added PUMI versions of examples ex1, ex1p, ex2 and ex6p in a new examples/pumi
directory. The new examples demonstrate the PUMI APIs for parallel and serial
mesh loading (ex1 and ex1p), applying BCs using classification (ex2), and
performing parallel mesh adaptation (ex6p).
- Added two new miniapps related to DataCollection I/O in miniapps/tools:
load-dc.cpp can be used to visualize fields saved via DataCollection classes;
convert-dc.cpp demonstrates how to convert between MFEM's different concrete
DataCollection options.
- Example 10p with its SUNDIALS and PETSc versions have been updated to reflect
the change in the behavior of the method ParNonlinearForm::GetLocalGradient()
(see the "API changes" section) and now works correctly on non-conforming AMR
meshes. Example 10 and its SUNDIALS version have also been updated to support
non-conforming ARM meshes.
Miscellaneous
-------------
- Documented project workflow and provided contribution guidelines in the new
top-level file, CONTRIBUTING.md.
- Added (optional) Conduit Mesh Blueprint support of MFEM data for both in-core
and I/O use cases. This includes a new ConduitDataCollection that provides
json, simple binary, and HDF5-based I/O. Support requires Conduit >= v0.3.1
and VisIt >= v2.13.1 will read the new Data Collection outputs.
- Added a new developer tool, config/sample-runs.sh, that extracts the sample
runs from all examples and miniapps and runs them. Optionally, it can save the
output from the execution to files, allowing comparison between different
versions and builds of the library.
- Support for building a shared version of the MFEM library with GNU make.
- Added a build option, MFEM_USE_EXCEPTIONS=YES, to throw an exception instead
of calling abort on mfem errors.
- When building with the GnuTLS library, switch to using X.509 certificates for
secure socket authentication. Support for the previously used OpenPGP keys has
been deprecated in GnuTLS 3.5.x and removed in 3.6.0. For secure communication
with the visualization tool GLVis, a new set of certificates can be generated
using the latest version of the script 'glvis-keygen.sh' from GLVis.
- Upgraded MFEM to support Axom 0.2.8. Prior versions are no longer supported.
API changes
-----------
- Introduced a new enum, Matrix::DiagonalPolicy, that replaces the integer
parameters in many methods that perform elimination of rows and/or columns in
matrices. Some examples of such methods are:
* class SparseMatrix: EliminateRow(), EliminateCol(), EliminateRowCol(), ...
* class BilinearForm: EliminateEssentialBC(), EliminateVDofs(), ...
* class StaticCondensation: EliminateReducedTrueDofs()
* class BlockMatrix: EliminateRowCol()
Calling these methods with an explicitly given (integer) constants, will now
generate compilation errors, please use one of the new enum constants instead.
- Modified the virtual method AbstractSparseMatrix::EliminateZeroRows() and its
implementations in derived classes, to accept an optional 'threshold'
parameter, replacing previously hard-coded threshold values.
- In the classes NonlinearForm and ParNonlinearForm:
* The method GetLocalGradient() no longer imposes boundary conditions. The
motivation for the change is that, in the case of non-conforming AMR,
performing the elimination at the local level is incorrect - it must be
applied at the true-dof level.
* The method SetEssentialVDofs() is now deprecated.
Version 3.3.2, released on Nov 10, 2017
=======================================
High-order mesh optimization
----------------------------
- Added support for mesh optimization via node-movement based on the Target-
Matrix Optimization Paradigm (TMOP) developed by P.Knupp et al. A variety of
mesh quality metrics, with their first and second derivatives have been
implemented. The combination of targets & quality metrics is used to optimize
the physical node positions, i.e., they must be as close as possible to the
shape, size and/or alignment of their targets. The optimization of arbitrary
high-order meshes in 2D, 3D, serial and parallel is supported.
- The new Mesh Optimizer miniapp can be used to perform mesh optimization with
TMOP in serial and parallel versions. The miniapp also demonstrates the use of
nonlinear operators and their coupling to Newton methods for solving
minimization problems.
New and improved solvers and preconditioners
--------------------------------------------
- MFEM is now included in the xSDK project, the Extreme-scale Scientific
Software Development Kit, as of xSDK-0.3.0. Various changes were made to
comply with xSDK's community policies, https://xsdk.info/policies, including:
xSDK-specific options in CMake, support for user-provided MPI communicators,
runtime API for version number, and the ability to disable/redirect output.
For more details, see general/globals.hpp and in particular the mfem::err and
mfem::out streams replacing std::err and std::out respectively.
- Added (optional) support for the STRUMPACK parallel sparse direct solver and
preconditioner. STRUMPACK uses Hierarchically Semi-Separable (HSS) compression
in a fully algebraic manner, with interface similar to SuperLU_DIST. See
http://portal.nersc.gov/project/sparse/strumpack for more details.
- Added a block lower triangular preconditioner based (only) on the actions of
each block, see class BlockLowerTriangularPreconditioner.
- Added an optional operator in LOBPCG to projects vectors onto a desired
subspace (e.g. divergence-free). Other small changes in LOBPCG include the
ability to set the starting vectors and support for relative tolerance.
- The Newton solver supports an optional scaling factor, that can limit the
increment in the Newton step, see e.g. the Mesh Optimizer miniapp.
- Updated MFEM integration to support the new SUNDIALS 3.0.0 interface.
New and updated examples and miniapps
-------------------------------------
- Added a new serial and parallel example (ex18) that solves the transient Euler
equations on a periodic domain with explicit time integrators. In the process
extended the NonlinearForm class to allow for integrals over faces and
exchanging face-neighbor data in parallel.
- Added a new meshing miniapp, Shaper, that can be used to resolve complicated
material interfaces by mesh refinement, e.g. as a tool for initial mesh
generation from prescribed "material()" function. Both conforming and
non-conforming (isotropic and anisotropic) refinements are supported.
- Added a new meshing miniapp, Mesh Optimizer, that demonstrates the use of TMOP
for mesh optimization (serial and parallel version.)
- Added SUNDIALS version of Example 16/16p.
Discretization improvements
---------------------------
- Added a FindPoints method of the Mesh and ParMesh classes that returns the
elements that contain a given set of points, together with the coordinates of
the points in the reference space of the corresponding element. In parallel,
if a point is shared by multiple processors, only one of them will mark that
point as found. Note that the current implementation of this method is not
optimal and/or 100% reliable. See the mesh-explorer miniapp for an example.
- Added a new class InverseElementTransformation, that supports a number of
algorithms for inversion of general ElementTransformations. This class can be
used as a more flexible and extensible alternative to ElementTransformation's
TransformBack method. It is also used in the FindPoints methods as a tunable
and customizable inversion algorithm.
- Memory optimizations in the NCMesh class, which now uses 50% less memory than
before. The average cost of an element in a uniformly refined mesh (including
the refinement hierarchy, but excluding the temporary face_list and edge_list)
is now only about 290 bytes. This also makes the class faster.
- Added the ability to integrate delta functions on the right-hand side (by
sampling the test function at the center of the delta coefficient). Currently
this is supported in the DomainLFIntegrator, VectorDomainLFIntegrator and
VectorFEDomainLFIntegrator classes.
- Added five new linear interpolators in fem/bilininteg.cpp to compute products
of scalar and vector fields or products with arbitrary coefficients.
- Added matrix coefficient support to CurlCurlIntegrator.
- Extend the method NodalFiniteElement::Project for VectorCoefficient to work
with arbitrary number of vector components.
Miscellaneous
-------------
- Added a .gitignore file that ignores all files erased by "make distclean",
i.e. the files that can be generated from the source but we don't want to
track in the repository, as well as a few platform-specific files.
- Added Linux, Mac and Windows CI testing on GitHub with Travis CI and Appveyor.
- Added a new macro, MFEM_VERSION, defined as a single integer of the form
(major*100 + minor)*100 + patch. The convention is that an even number
(i.e. even patch number) denotes a "release" version, while an odd number
denotes a "development" version. See config/config.hpp.in.
- Added an option for building in parallel without a METIS dependency. This is
used for example the Laghos miniapp, https://github.com/CEED/Laghos.
- Modified the installation layout: all headers, except the master headers
(mfem.hpp and mfem-performance.hpp), are installed in <PREFIX>/include/mfem;
the master headers are installed in both <PREFIX>/include/mfem and in
<PREFIX>/include. The mfem configuration and testing makefiles (config.mk and
test.mk) are installed in <PREFIX>/share/mfem, instead of <PREFIX>.
- Add three more options for MFEM_TIMER_TYPE.
- Support independent number of digits for cycle and rank in DataCollection.
- Converted Sidre usage from "asctoolkit" to "axom" namespace.
- Various small fixes and styling updates.
API changes
-----------
- The methods GetCoeff of VectorArrayCoefficient and MatrixArrayCoefficient now
return a pointer to Coefficient (instead of reference). Note that NULL pointer
is a valid entry for these two classes - it is treated as the zero function.
- When building with PETSc, the required PETSc version is now 3.8.0. Newer
versions may work too, as long as there are no interface changes in PETSc.
- The class GeometryRefiner now uses the enum in Quadrature1D for its type
specification. In particular, this will affect older versions of GLVis. A
simple upgrade to the latest version of GLVis should resolve this issue.
Version 3.3, released on Jan 28, 2017
=====================================
FEM <-> linear system interface for action-only linear operators
----------------------------------------------------------------
- Added a new class, ConstrainedOperator, which can impose essential boundary
conditions using only the action, Mult(), of a given square linear Operator.
- Added a FormLinearSystem + RecoverFEMSolution functionality for square linear
Operators that are available only through their action. This includes all
necessary transformations, such as: parallel assembly, conforming constraints
for non-conforming AMR and eliminating boundary conditions. (Hybridization and
static condensation are not supported.) See examples in miniapps/performance.
Matrix-free preconditioning and low-order-refined spaces
--------------------------------------------------------
- The HPC examples in miniapps/performance now support efficient preconditioning
in matrix-free mode based on applying a standard (e.g. AMG) preconditioner to
a sparsified version of the operator. The sparsification is obtained by
rediscretizing with a low-order refined spaces, currently at the high-order
degrees of freedom.
- New mesh constructors support the creation of low-order-refined version of a
given mesh, both in serial and in parallel. These are illustrated in the HPC
examples in miniapp/performance (option -pc lor), as well as in mesh-explorer
miniapp, which now supports Gauss-Lobatto refinement and uniform refinement,
both for any factor > 1.
Comprehensive PETSc and SUNDIALS interfaces
-------------------------------------------
- Added support for many linear and nonlinear solvers, preconditioners, time
integrators and other features from the PETSc suite (version 3.8 or higher of
the PETSc dev branch is required). The new features include:
* support for PETSc matrices in MATAIJ, MATIS, MATSHELL and MATNEST formats.
* PETSc linear solvers can take any mfem Operator and support user-defined
monitoring routines (see examples/petsc/ex1p).
* BDDC preconditioners for H1, H(curl) and H(div), including with static
condensation/hybridization, FieldSplit preconditioner for BlockOperators.
* PETSc non-linear solvers can take any mfem Operator that implements the
GetGradient() method.
* PETSc ODE solvers are supported for mfem's TimeDependentOperators.
The use of these features is illustrated in the new examples/petsc directory.
- Added a new class, OperarorHandle, that provides a common interface for
global, matrix-type operators to be used in bilinear forms, gradients of
nonlinear forms, static condensation, hybridization, etc.
The following backends are currently supported:
* HYPRE parallel sparse matrix (HYPRE_PARCSR)
* PETSC globally assembled parallel sparse matrix (PETSC_MATAIJ)
* PETSC parallel matrix assembled on each processor (PETSC_MATIS)
- Added support for the time integrators and non-linear solvers from the CVODE,
ARKODE and KINSOL libraries of the SUNDIALS suite (version 2.7 or higher of
SUNDIALS is required). The use of these features is illustrated in the new
examples/sundials directory.
Scalable parallel mesh support
------------------------------
- Introduced a new mesh format (v1.2) that can describe/recover MFEM parallel
meshes. This way, computations can start directly in parallel without serial
refinement and splitting. Non-conforming meshes are currently supported only
in serial.
General quadrature and nodal finite element basis types
-------------------------------------------------------
- Added support for different numerical quadrature schemes and finite element
basis points. Different basis points can be selected via optional integer
argument(s) to the finite element collection constructor of type BasisType:
* H1 elements can use GaussLobatto (default), Positive, or ClosedUniform;
* L2 elements can use GaussLegendre (default), GaussLobatto, Positive,
ClosedUniform, OpenUniform or OpenHalfUniform;
* RT can now use open basis that is GaussLegendre (default), GaussLobatto,
ClosedUniform, OpenUniform, or OpenHalfUniform, and closed basis that is
GaussLobatto (default) or ClosedUniform;
* ND elements can use the same BasisType's as RT elements.
- GaussLegendre, GaussLobatto, ClosedUniform, OpenUniform, and OpenHalfUniform
integration rules can be directly constructed with an optional parameter of
type Quadrature1D:
IntegrationRules gl(0, Quadrature1D::GaussLobatto);
const IntegrationRule *ir = gl(Geometry::SEGMENT, 5); // 4pt 1D rule
The global IntRules object continues to use GaussLegendre.
New integrators for common families of operators
------------------------------------------------
- Added MixedScalarIntegrator and 7 derived classes for integrating products of
two scalar basis functions and optional scalar coefficients.
- Added MixedVectorIntegrator and 16 derived classes for integrating the inner
product of two vector basis functions with optional scalar, vector, or matrix
coefficients.
- Added MixedScalarVectorIntegrator and 13 derived classes for integrating the
product of a scalar basis function with the inner product of a vector basis
function with a vector coefficient. In 2D the inner product can optionally be
replaced with a cross product.
- Added a new class DGElasticityIntegrator that supports a few types of DG
formulations for linear elasticity and a new linear form integrator,
DGElasticityDirichletLFIntegrator, that implements non-homogeneous BCs.
- Added support for DG spaces in class VectorBoundaryLFIntegrator.
- In classes BilinearForm and LinearForm, added support for boundary face
integrators applied to a subset of the boundary, see AddBdrFaceIntegrator.
New and updated examples and miniapps
-------------------------------------
- Sixteen new serial and parallel example codes that demonstrate:
* solution of a time-dependent nonlinear heat equation (Example 16/16p)
* DG formulations of static linear elasticity (Example 17/17p)
* the use of PETSc solvers and preconditioners (Examples 1p, 2p, 3p, 4p, 5p,
6p, 9p and 10p in examples/petsc)
* the use of SUNDIALS time integrators and nonlinear solvers (Examples 9/9p
and 10/10p in examples/sundials)
- The HPC examples in miniapps/performance now have a -mf/--matrix-free option
illustrating optimized "partial assembly" operator evaluation. This is now the
default in these examples, to switch to optimized matrix assembly instead use
the -asm/--assembly option.
- Added a new electromagnetic miniapp, Joule, illustrating the simulation of
transient magnetics and joule heating. This is a comprehensive miniapp that
uses finite element spaces and solvers for the whole de Rham sequence.
- Added a simple miniapp, display-basis, for displaying the various types
of finite element basis functions within single elements. This is part of
the new miniapps/tools directory.
- Rewrote the Volta and Tesla solver classes to avoid using linear algebra
objects when possible. This greatly simplifies the code, reduces memory
requirements, and eliminates unnecessary computation. It also fixed a bug
with divergence cleaning in the Tesla miniapp.
- Added an option to Example 9/9p to save a binary visualization file using the
Conduit mesh blueprint/hdf5 format.
Improved building options
-------------------------
- Added a new CMake build system, that can be used as an alternative to the GNU
make-based build system (e.g. for out-of-source building). For more details,
see the INSTALL file and the config/cmake directory.
- Added support for out-of-source builds with GNU make, see the INSTALL file.
Improved file output
--------------------
- Added on-the-fly compression of file streams input and output via gzstream,
see the MFEM_USE_GZSTREAM option.
- Added experimental support for an HDF5-based output file format following the
Conduit (https://github.com/LLNL/conduit) mesh blueprint specification for
visualization and/or restart capability. This functionality is aimed primarily
at user of LLNL's axom project (Sidre component) that run problems at extreme
scales. Users desiring a small scale binary format may want to look at the
gzstream functionality instead.
Miscellaneous
-------------
- Added optional support for software-based higher-precision arithmetic with
the MPFR library. When MFEM_USE_MPFR is enabled, the 1D quadrature rules will
be computed precisely, at least for rules with up to 65-points.
- Better support for METIS version 5 and above.
- Provide an informative backtrace in mfem_error based on the cross-platform
libunwind library (requires MFEM_USE_LIBUNWIND=YES).
- In class SparseMatrix, added methods PrintInfo and CheckFinite.
- GMRESSolver and MINRESSolver now support the same print levels as CGSolver.
- Added method MemoryUsage to the classes Stack and MemAlloc.
- Improved Doxygen formatting of code comments.
- Various other simplifications, extensions, and bugfixes in the code.
Version 3.2, released on Jun 30, 2016
=====================================
Dynamic AMR with parallel load balancing, derefinement of non-conforming meshes
-------------------------------------------------------------------------------
- Parallel non-conforming meshes can now be load balanced at any time by calling
ParMesh::Rebalance(). Elements of the mesh are redistributed in such a way
that each processor gets approximately the same number of elements (plus minus
one element). Partitioning is done by splitting a sequence of space-filling
(Hilbert) curves defined on the refinement octrees.
- Isotropically refined non-conforming meshes can now be derefined, both in
serial and in parallel, based on a per-element error measure and a
derefinement threshold. See the class ThresholdDerefiner.
- Following an arbitrary mesh change (uniform/general conforming/non-conforming
refinement, derefinement, load balancing), the FiniteElementSpace and
associated GridFunctions can be updated by interpolating or redistributing the
previous function values based on the new state of the mesh. (Internally this
is implemented through a transformation matrix that is constructed in the
FiniteElementSpace.) The user interface is quite simple:
pmesh.Rebalance(); // or GeneralRefinement, or GeneralDerefinement
fespace.Update(); // calculate a transformation matrix (by default)
x.Update(); // apply the transformation to the GridFunction
z.Update(); // apply it again
- New abstractions are available for error estimation and general mesh
operations such as refinement and derefinement. See the base classes
ErrorEstimator and MeshOperator and their descendants.
- The above features are illustrated in the new Example 15 (see also Example 6).
Tensor-based high-performance FEM operator assembly and evaluation
------------------------------------------------------------------
- Added support for high-performance, tensor-based efficient assembly and
evaluation of high-order operators.
- A number of new header files have been added to the fem/, linalg/ and mesh/
directories. They start with the prefix "t" to indicate the (heavy) use of C++
templating, similar to how the prefix "p" denotes "parallel". All the code for
the new HPC FE assembly/evaluation algorithms is fully implemented in these
header files. Note that the new interface is optional and only enabled if the
mfem-performance.hpp header is included instead of mfem.hpp. This is an
initial, reference implementation.
- Similarly to the serial-to-parallel (ex1.cpp-to-ex1p.cpp) transition, an
existing MFEM-based applications has to be transitioned to the new HPC
interface. This is illustrated in two new example codes which are the
high-performance versions of Example 1/1p. See miniapps/performance.
- The new interface reduces local operator assembly/evaluation to batched small
dense tensor contraction operations. For high performance, the sizes of these
contractions should be known at compile time, so the BilinearForm object needs
to have detailed knowledge about the mesh, the finite element space, the
quadrature rule and the integrator to be assembled. This required a new
interface, that supports a subset of the current (general) coefficients and
bilinear form integrators, including variable coefficients and mass and
diffusion integrators. It is possible to use the old and the new HPC interface
side-by-side, see the HPC version of Example 1/1p in miniapps/performance.
Advanced FEM on parallel non-conforming meshes
----------------------------------------------
- Added support for discontinuous Galerkin methods on parallel non-conforming
meshes, see Examples 9p and 14p.
- Added support for hybridization on parallel non-conforming meshes, see
Example 4p.
New and improved linear solvers
-------------------------------
- Added a wrapper for the real-valued, double precision solver in SuperLU_DIST
which is a sparse direct solver for distributed memory architectures. As such
it can only be enabled along with MFEM_USE_MPI. When MFEM is configured with
MFEM_USE_SUPERLU, one also needs to alter the version of METIS, since SuperLU
requires ParMETIS (which comes packaged with a serial version of METIS). See
http://crd-legacy.lbl.gov/~xiaoye/SuperLU for SuperLU_DIST details.
- Added a wrapper for the KLU solver in SuiteSparse see
http://faculty.cse.tamu.edu/davis/suitesparse.html for details of KLU.
If MFEM was configured with MFEM_USE_SUITESPARSE, one must now also link
against the klu and btf libraries in SuiteSparse, see config/defaults.mk.
New and updated examples and miniapps
-------------------------------------
- Four new serial and parallel example codes that demonstrate:
* high-performance finite element operator assembly/evaluation (Example 1/1p
in miniapps/performance)
* adaptive refinement, derefinement and load balancing (in parallel) on
non-conforming meshes (Example 15/15p)
- Examples 4p now supports hybridization on non-conforming meshes.
- Examples 9p and 14p now work on non-conforming meshes.
- Example 11p now has optional support for the SuperLU parallel direct solver.
- Added several new options and example runs in the Volta and Tesla miniapps,
including support for Halbach arrays of permanent magnets.
Miscellaneous
-------------
- Added "check" and "test" targets to the top-level makefile. The former does a
quick check by running Example 1/1p, while the latter does a more thorough
verification of the build by running all example codes and miniapps.
- Added support for 2D and 3D meshes generated by Gmsh (http://gmsh.info), both
in ASCII and binary formats.
- Added a reader for Cubit meshes in the Genesis (NetCDF) format. Currently
supported are linear and quadratic tet and hex meshes.
- Added support for boundary bilinear form integrators when using hybridization.
- Added support for Robin boundary conditions for DG in BoundaryMassIntegrator.
- Moved all reference element connectivity descriptions, such as element-edge,
element-face, etc. to the template class Geometry::Constants<Geometry::Type>.
- Added support for secure socket communications in class socketstream based on
the GnuTLS library, see INSTALL for more details.
- Renamed config/user.mk.in to config/defaults.mk and moved all the default
build settings from the makefile there.
- Added configurable variables AR, ARFLAGS, and RANLIB in the build system. The
defaults for Mac OS X will suppress the "has no symbols" warnings.
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- Changes in class Mesh
* Two-level state functionality was removed, including: UseTwoLevelState(int),
SetState(int), GetState(), GetNumFineElems(int), GetRefinementType(int),
GetFineElem(int, int) and GetFineElemTrans(int, int).
- Changes in class FiniteElementSpace
* BuildElementToDofTable() is now protected, and it is always called.
* GlobalRestrictionMatrix(FiniteElementSpace*, int) was removed, but the
prolongation operator can still be accessed via GetUpdateOperator() after
mesh refinement and a call to Update(true).
- Changes in methods related to non-conforming meshes and spaces
* The methods LinearForm::ConformingAssemble, BilinearForm::ConformingAssemble
and GridFunction::ConformingProlongate/ConformingProject are now hidden
inside (Par)BilinearForm::FormLinearSystem and RecoverFEMSolution.
* The conforming prolongation/restriction matrices can still be accessed via
FiniteElementSpace::GetConformingProlongation()/GetConformingRestriction().
- Changes in classes GridFunction and ParGridFunction
* Renamed Update((Par)FiniteElementSpace*, Vector&, int) to MakeRef.
* Renamed Update((Par)FiniteElementSpace*) to SetSpace.
Version 3.1, released on Feb 16, 2016
=====================================
Substantially improved non-conforming adaptive mesh refinement
--------------------------------------------------------------
- Added support for parallel non-conforming mesh refinement, including a new
example code with adaptive mesh refinement for the Laplace problem (Example
6p). Most of the example codes can now work on non-conforming meshes in serial
and in parallel.
- Added simple ZZ-type error estimators, including an anisotropic one in serial,
and one based on Raviart-Thomas flux projection in parallel, to the AMR
examples 6 and 6p. These seem to perform quite reasonably, even for
higher-order discretizations on 2D, 3D and surface meshes.
- The MFEM mesh format has a new version(1.1) that supports non-conforming
meshes. The format is an extension of 1.0 that includes a vertex_parents and
an optional coarse_elements section. See the example meshes amr-quad.mesh,
amr-hex.mesh and fichera-amr.mesh in the data/ directory.
- Added support for DG discretizations on non-conforming meshes in serial. See
the sample runs in Example 14.
- A new function, ParGridFunction::ParallelProject() directly returns a hypre
vector restricted to the true degrees of freedom (and supports non-conforming
meshes). In most cases, this should be preferred to the ParallelAverage()
function.
- When using non-conforming meshes, the essential boundary condition elimination
has to be applied at the end of the (parallel) assembly. Furthermore, in
serial, the bilinear form needs to call ConformingAssemble() after assembly
and the solution should call ConformingProlongate() after the solve (these are
not necessary in parallel). Note that these could also be handled
automatically by the new FEM <-> linear system interface, see below.
General finite element spaces and solvers on surfaces/skeletons
---------------------------------------------------------------
- Added support for arbitrary high-order finite element spaces on the mesh
skeleton (the faces, edges, and vertices between mesh elements) that are the
traces of the H1 and H(curl) spaces defined on the mesh. With the previously
existing H(div) trace space, the full de Rham sequence on the skeleton is now
supported.
- Updated integrators and discrete interpolators to work correctly for H(curl)
and H(div) spaces defined on surface meshes, or the mesh skeleton.
Hybridization, static condensation and a new FEM <-> linear system interface
----------------------------------------------------------------------------
- The BilinearForm/ParBilinearForm classes now support static condensation, as
well as hybridization (based on given constraint space and trace integrator).
These are illustrated in Examples 1-4.
- Added a new interface for transitioning between the finite element objects and
their corresponding linear algebra objects, which supports abstracts
transformations such as: parallel assembly, eliminating boundary conditions,
applying conforming constraints for non-conforming AMR, hybridization, static
condensation, back substitution, etc. Changed several of the example codes
accordingly.
New eigensolvers and improved solvers
-------------------------------------
- Added support for the scalable Locally Optimal Block Preconditioned Conjugate
Gradient (LOBPCG) eigenvalue solver and the Auxiliary-space Maxwell
Eigensolver (AME) from hypre.
- Added 3 new example codes to demonstrate the LOBPCG and AME applications to
the Laplace (Example 11p), Elasticity (Example 12p) and Maxwell (Example 13p)
eigenproblems.
- Updated the HypreAMS and HypreADS solvers to work for H(curl) and H(div)
problems defined on surface meshes, or the mesh skeleton.
- Added support for a discretization-enhanced version of hypre's BoomerAMG
designed specifically for linear elasticity problems, see Example 2p.
- The HypreAMS solver can now be used to solve singular curl-curl problems.
New and updated examples
------------------------
- Six new serial and parallel example codes that demonstrate:
* parallel conforming and non-conforming adaptive mesh refinement (Example 6p)
* hypre's LOBPCG eigensolver for the Laplace eigenproblem (Example 11p)
* hypre's LOBPCG eigensolver for the elasticity eigenproblem (Example 12p)
* hypre's AME eigensolver for the Maxwell eigenproblem (Example 13p)
* DG diffusion discretizations for the Laplace equation (Example 14/14p)
- Examples 1-4 now support static condensation, and Example 4/4p supports H(div)
hybridization, leading to much improved solve times. These examples also
illustrate the new interface for linear system assembly (see also Examples 6
and 7).
- Significantly improved the DPG preconditioner in Example 8p, which is now
scalable in parallel and uses the HypreADS solver to precondition the
interfacial block as an H(div) problem reduced to the mesh skeleton.
- Example 7/7p has a new option, -amr, showcasing simple local conforming and
non-conforming mesh refinements.
- Example 3/3p now works in both 2D and 3D.
New miniapps
------------
- Electromagnetic miniapps:
* Volta - simple electrostatics simulation code.
* Tesla - simple magnetostatics simulation code.
See also the README file in miniapps/electromagnetics.
- Meshing miniapps:
* Mobius Strip - generate various Mobius strip-like meshes.
* Klein Bottle - generate three types of Klein bottle surfaces.
* Mesh Explorer - visualize and manipulate meshes.
See also the README file in miniapps/meshing.
Miscellaneous
-------------
- Moved MFEM from Google Code to GitHub. New website: http://mfem.org.
- Formatted the code with Artistic Style, see the "make style" target.
- Added support for 64-bit integers in global size variables, enabling
simulations with >2B unknowns. (This requires that hypre is configured with
the --enable-bigint option.)
- Added optional support for the Gecko graph reordering library.
- Updated the implementation of some operations in DenseMatrix for better
auto-vectorization. Added a new class LUFactors that computes LU factorization
(with pivoting) and perform various operations with the factored data.
- Various other simplifications, extensions, and bugfixes in the code.
Version 3.0, released on Jan 26, 2015
=====================================
Improved documentation and build system
---------------------------------------
- Added interactive example documentation in examples/README.html. This should
be the starting point for new users interested in MFEM's features.
- New Doxygen-based code documentation. Due to its size, users are expected to
build this documentation themselves by typing make in the doc/ directory.
(Alternatively, the pre-build documentation can be browsed online).
- New build system, based on GNU make which consists of configuration and build
steps: "make config; make". The MFEM build options are exported, and can be
included in external makefiles. Library installation is also supported. See
"make help" and the INSTALL file for details.
- To build the examples use 'make' or 'make -j <np>' in the examples/ directory.
Based on the current MFEM configuration this will build the serial or the
parallel examples using the same config options as the library.
New and updated examples
------------------------
- Six new serial/parallel example codes that demonstrate:
* mixed pressure-velocity FEM for Darcy (Example 5)
* non-conforming adaptive mesh refinement for Laplace (Example 6)
* Laplace problem on a surface (Example 7)
* Discontinuous Petrov-Galerkin (DPG) for Laplace (Example 8)
* Discontinuous Galerkin (DG) time-dependent advection (Example 9)
* time-dependent implicit nonlinear elasticity (Example 10)
- Added command line options to all examples and modified several of the serial
ones to optionally use the serial direct solver UMFPACK.
- Simplified the elimination of Dirichlet boundary conditions in parallel.
- Grouped and documented the example code features in examples/README.html
Serial non-conforming adaptive mesh refinement
----------------------------------------------
- Added support for general, isotropic and anisotropic, local non-conforming
mesh refinement (using hanging nodes) in 2D and 3D, on quadrilateral,
triangular and hexahedral meshes. High-order curved and surface meshes are
also supported.
- The current implementation supports serial meshes (see example 6). Extension
to parallel meshes is in active development.
- The mesh is refined with Mesh::GeneralRefinement. The non-conforming mesh is
represented as a mesh that is "cut" along non-conforming edges and faces in
the internal NCMesh class. The only thing the user has to do to obtain a
continuous solution is to call BilinearForm::ConformingAssemble and
GridFunction::ConformingProlongate before and after solving the linear system.
The finite element space and grid functions are then updated with
FiniteElementSpace::UpdateAndInterpolate().
Time-dependent problems, non-linear operators and ODE integrators
-----------------------------------------------------------------
- Added new abstract base class TimeDependentOperator and a set of explicit
Runge-Kutta time integration classes in linalg/ode.?pp.
- Added classes for diagonally implicit Runge-Kutta (DIRK) time integrators
based on the ImplicitSolve() method of TimeDependentOperator.
- Extended all coefficient classes to be optionally time-dependent.
- Added classes for general nonlinear finite element operators (deriving from
NonlinearForm/ParNonlinearForm). Such operators have assemble-based action and
also support assembly of the gradient operator to enable inversion with Newton
iteration.
Discontinuous Galerkin and Discontinuous Petrov-Galerkin methods
----------------------------------------------------------------
- Added support Discontinuous Galerkin (DG) face integrators in parallel by
extending ParMesh with information for face-neighboring processors. Added DG
support in ParFiniteElementSpace, ParBilinearForm and ParGridFunction.
- Introduced a new class of integrators for forms defined on the faces of the
mesh (including interior and boundary faces), mainly intended for hybrid
methods like HDG and DPG that employ facet (numerical trace) spaces.
Block systems and rectangular operators
---------------------------------------
- Added classes BlockOperator, BlockVector and BlockMatrix for handling block
systems with different components (e.g., pressure and velocity).
- New abstract class AbstractSparseMatrix, between Matrix and SparseMatrix
- Modified class Operator to have two separate sizes: "height" and "width" for
the output and input sizes, respectively. The Size method was removed.
- For backward compatibility, the method Size is still present in the classes
DenseMatrix (returns width as before), SparseMatrix (returns height as
before), DenseMatrixInverse (square matrix) and BilinearForm (square matrix).
Linear and non-linear solvers
-----------------------------
- New abstract class Solver, with sub-classes for sparse smoothers, dense matrix
inverse, iterative solvers (Krylov methods and Newton) and the hypre solvers.
All Krylov methods were consolidated in linalg/solver.cpp and extended to work
in parallel.
- Added several new classes of solvers and smoothers:
* serial sparse direct solvers from the SuiteSparse library (UMFPACK)
* HypreSmoother, giving access to the parallel ParCSR smoothers in hypre
* polynomial smoothers: Chebyshev, Taubin and FIR
* stationary linear iteration (SLI)
* quadratic single linearly-constrained optimization problems with bounds
Miscellaneous
-------------
- Wrapped all classes/functions/objects in a namespace called "mfem".
- Automated the creation of quadrature rules to enable on-demand generation of
arbitrary order rules for all geometries 1D/2D/3D geometries.
- Added support for saving collections of grid functions in format suitable for
visualization with VisIt (visit.llnl.gov). See examples 5 and 9.
- Added support for 1D, surface and topologically periodic meshes, as well as a
simple inline mesh format. See the data/ directory for examples.
- Added support for serial mesh optimization using the Mesquite mesh quality
improvement toolkit (see mesh/mesquite.?pp and INSTALL for details).
- Made sure that MFEM can work in parallel with empty processors and with any
MPI communicator.
- Improved high-order Bernstein basis support.
- Support for high-resolution timers (e.g. POSIX clocks).
- Improved error messages with several macros, such as MFEM_ABORT, MFEM_VERIFY,
MFEM_ASSERT, MFEM_WARNING, etc.
- Improved portability for Windows (Visual Studio) and Mac OS X.
- Various simplifications, extensions, and bugfixes in the code.
Version 2.0, released on Nov 18, 2011
=====================================
Arbitrary order finite element spaces
-------------------------------------
- Added support for arbitrary high-order finite element spaces through the new
classes H1_FECollection, L2_FECollection, RT_FECollection and ND_FECollection.
These are based on a number of new FiniteElement sub-classes H1_*, L2_*, RT_*
and ND_* elements of arbitrary order on all types of reference elements.
- The classes implement H1-conforming, L2-discontinuous, H(div)-conforming
Raviart-Thomas and H(curl)-conforming Nedelec elements on triangular,
quadrilateral, tetrahedral and hexahedral meshes. The only restriction on the
order of the spaces is the availability of the required quadrature rules.
NURBS meshes and discretization spaces
--------------------------------------
- Added a collection of classes for serial and parallel meshes and
discretization spaces using Non-uniform rational B-splines (NURBS) basis
functions (files mesh/nurbs.?pp).
- The Mesh class supports the NURBS-specific refinement functions: KnotInsert
and DegreeElevate. Example NURBS meshes can found in the 'data' directory with
file names *-nurbs.mesh including an exact non-degenerate disc
(disc-nurbs.mesh) and exact non-degenerate ball (ball-nurbs.mesh).
- We can handle arbitrary NURBS or standard, non-NURBS, finite element spaces on
NURBS meshes. However, a NURBS finite element space requires an underlying
NURBS mesh. Refinement of parallel NURBS meshes is not supported yet.
Discrete gradient, curl, etc. matrices
--------------------------------------
- Added a new class, DiscreteLinearOperator, that facilitates the construction
of matrix representations for linear operators like gradient, curl, embedding,
projection, etc. The corresponding local "interpolators" are similar to
bilinear form integrators and derive from base class DiscreteInterpolator.
Current interpolators include GradientInterpolator, IdentityInterpolator,
CurlInterpolator and DivergenceInterpolator.
- Also available is a parallel version of DiscreteLinearOperator, which
assembles parallel topological matrices (such as the discrete gradient, curl,
etc.) in hypre's ParCSR format.
New integrators
---------------
- New linear (r.h.s.) integrator VectorFEBoundaryFluxLFIntegrator for
assembling (u, v.n) on the boundary for scalar u and v in an RT space.
- New bilinear integrator VectorFECurlIntegrator for assembling (curl u, v) for
u in a ND space and v in an RT space.
New and updated examples
------------------------
- Added a new serial/parallel Example code 4/4p, which solves a 2D or 3D H(Div)
diffusion problem using the Raviart-Thomas finite elements. In parallel, the
linear system is solved with the brand-new Auxiliary-space Divergence Solver
(ADS) in hypre.
- Modified Example 1 to use isoparametric discretization (use the FE space from
the mesh) including NURBS meshes and spaces. Updated Example 2 to support
arbitrary order spaces. Updated all examples to work with NURBS meshes and
spaces, as well as to not use projection onto discontinuous polynomial spaces
for visualization (this is now handled directly in GLVis when necessary).
- In all examples, switched to a uniform "solution" socket data type instead of
the various previous "*_gf_data" data types.
- In the parallel examples, switched to parallel mesh and solution output, as
well as to the new parallel socket format in place of PrintAsOne/SaveAsOne.
New hypre solvers
-----------------
- The parallel MFEM build now requires hypre 2.8.0b or newer.
- Extended HypreAMS and HypreADS to support (arbitrary) high-order ND/RT spaces,
by internally constructing the high-order ParDiscreteLinearOperator gradient,
curl and interpolation matrices. This makes the linear solve in Example 3p and
4p significantly faster than before. Extended the HypreAMS object to also work
for 2D H(div) problems.
Miscellaneous
-------------
- Added new class socketstream implementing two-way tcp/ip socket communications
in the framework of C++ streams. Added new class socketserver implementing
tcp/ip server functionality: listen on a given port for incoming connections,
and accept them by assigning the new connection to a socketstream. These new
classes are meant to replace the classes isockstream and osockstream. They
allow MFEM code to update the mesh and solution via a single socket connection
to a GLVis window.
- Added new Mesh and GridFunction constructors that combine multiple Mesh and
GridFunction objects into one object. These are used in GLVis to visualize
data saved in parallel. Removed obsolete code related to reading of parallel
disjoint meshes.
- Added more quadrature rules on triangles and tetrahedra.
- Basic experimental OpenMP support (disabled by default). When enabled, OpenMP
code is used for local matrix assembly, sparse matrix-vector product, and some
vector operations.
- Added support for METIS 5.0 (not the default, see INSTALL).
- Various simplifications, extensions, and bugfixes in the code.
Version 1.2, released on Apr 08, 2011
=====================================
Parallel MPI-based version of the library based on hypre
--------------------------------------------------------
- New MPI parallel version of the library based on the ParCSR parallel matrix
format from hypre and the metis graph partitioning library. This version
supports parallel local refinement and parallel curved meshes, as well as
several solvers from hypre.
New serial and parallel examples
--------------------------------
- Added a new example code describing an electromagnetic diffusion problem
discretized with lowest order Nedelec finite elements (Example 3).
- Added parallel versions of all examples codes (files ex1p.cpp, ex2p.cpp and
ex3p.cpp) based on hypre's BoomerAMG and AMS preconditioners.
Miscellaneous
-------------
- Added support for saving and reading linear and curved quadratic meshes in VTK
format. The format is automatically recognized when opening a mesh file, and
the boundary is reconstructed based on the actual domain boundary.
- The 'data' directory now contains a collection of various mesh files in the
MFEM and VTK formats, including curved meshes and the mesh files that were
previously in the 'examples' directory.
- Updated the default integration rule order for most of the linear form
integrators.
- Added support for cubic hex elements.
- Bugfixes in the face orientation of 3D RT0 elements and in the VectorFEDomain
linear form integrator.
- Various small fixes and styling updates.
Version 1.1, released on Sep 13, 2010
=====================================
New MFEM format for general meshes
----------------------------------
- New MFEM mesh v1.0 format with uniform structure for any dimension and support
for curved meshes including in 3D. Class Mesh will recognize and read the new
format (in addition to all previously used formats) and Mesh::Print uses the
new format by default. The old print function was renamed to Mesh::PrintXG.
New elasticity example
----------------------
- Added an example code for linear elasticity with (high-order) vector finite
elements (Example 2).
Miscellaneous
-------------
- Added Mesh::PrintVTK and GridFunction::SaveVTK methods for output in VTK
format.
- Implemented GeometryRefiner::Refine for CUBE and TETRAHEDRON geometries. This
allows for saving curved meshes in the VTK format.
- Added SConstruct file for mfem/examples.
- Various small fixes and styling updates.
Version 1.0, released on Jul 21, 2010
=====================================
- Uploaded to http://mfem.googlecode.com.
- Initial release.
Reference in New Issue View Git Blame Copy Permalink