path: root/CHANGELOG

                    Finite Element Discretization Library
                                   __
                       _ __ ___   / _|  ___  _ __ ___
                      | '_ ` _ \ | |_  / _ \| '_ ` _ \
                      | | | | | ||  _||  __/| | | | | |
                      |_| |_| |_||_|   \___||_| |_| |_|

                               https://mfem.org


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.