-
Notifications
You must be signed in to change notification settings - Fork 504
/
CHANGELOG
2311 lines (1778 loc) · 110 KB
/
CHANGELOG
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
Finite Element Discretization Library
__
_ __ ___ / _| ___ _ __ ___
| '_ ` _ \ | |_ / _ \| '_ ` _ \
| | | | | || _|| __/| | | | | |
|_| |_| |_||_| \___||_| |_| |_|
https://mfem.org
Version 4.5, released on October 22, 2022
=========================================
Meshing improvements
--------------------
- Added new SubMesh and ParSubMesh classes that can be used to extract a subset
of a given Mesh. These classes have the same functionality as Mesh and ParMesh
and work with all existing MFEM interfaces like finite element spaces etc.
- Added a method, ParMesh::GetSerialMesh(), that reconstructs a partitioned
parallel mesh on a given single rank. Also, added ParMesh::PrintAsSerial(),
which saves the reconstructed serial mesh to a C++ stream on rank 0.
- Added more 3D TMOP metrics, as well as specialized metrics for mesh
untangling and worst-case quality improvement.
- Added a new method, Mesh::NodesUpdated, which should be called after the mesh
node coordinates have changed, e.g. after the mesh has moved. This is
necessary, for example, with device assembly of linear and bilinear forms.
- Added support for mixed meshes and pyramids in GSLIB-FindPoints.
Discretization improvements
---------------------------
- Added full assembly and device support for several LinearForm integrators:
* DomainLF: (f, v)
* VectorDomainLF: ((f1,...,fn), (v1,...,vn))
* DomainLFGrad: (f, grad(v))
* VectorDomainLFGrad: ((f1x,f1y,f1z,...,fnx,fny,fnz), grad(v1,...,vn))
The device assembly of linear forms has to be explicitly enabled by calling
LinearForm::UseFastAssembly(true), otherwise the legacy linear form assembly
is used by default.
- Added support for assembling low-order-refined matrices using a GPU-enabled
"batched" algorithm. The lor_solvers and plor_solvers now fully support GPU
acceleration with arbitrary user-supplied coefficients.
- Added a new class FaceQuadratureSpace that allows for the construction of
QuadratureFunctions on the interior or boundary faces of a mesh.
- Added a class CoefficientVector for efficient access of variable coefficient
values at quadrature points (in particular for GPU/device kernels).
- Added WhiteGaussianNoiseDomainLFIntegrator: a LinearFormIntegrator class for
spatial Gaussian white noise.
- Added a new Zienkiewicz-Zhu patch recovery-based a posteriori error estimator.
See fem/estimators.hpp.
- Various fixes and improvements in LinearFormExtension.
Linear and nonlinear solvers
----------------------------
- Added a new class DGMassInverse that performs a local element-wise CG
iteration to solve systems involving the discontinuous Galerkin mass matrix,
including support for device/GPU acceleration.
- Added more flexibility to the constrained solver classes:
* PenaltyConstrainedSolver now allows for a vector of penalty parameters
(necessary for penalty contact)
* PenaltyConstrainedSolver and EliminationSolver can use GMRES or PCG
* All constraint solver classes can take a user-defined preconditioner
- Added functions to toggle additional options for the SuperLU_Dist and Hypre
preconditioners (ParaSails, Euclid, ILU).
- Added boundary elimination with device support for `SparseMatrix` and
`HypreParMatrix`.
New and updated examples and miniapps
-------------------------------------
- Added a new elasticity miniapp, Hooke, that showcases a low-level approach of
using MFEM to solve a nonlinear elasticity problem based on the fundamental
finite element operator decomposition. The miniapp also integrates with
automatic differentiation tools like a native dual number implementation or a
third party library such as Enzyme. See miniapps/elasticity for more details.
- Added example for body-fitted volumetric and shape integration using the
Algoim library in miniapps/shifted.
- Add a new example code, Example 33/33p, to demonstrate the solution of
spectral fractional PDEs with MFEM.
Integrations, testing and documentation
---------------------------------------
- Added a Dockerfile for a simple MFEM container, see config/docker/README.md.
More sophisticated developer containers are available in the new repo
https://github.com/mfem/containers.
- Added support for the LLVM-based automatic differentiation tool Enzyme, see
https://github.com/EnzymeAD/Enzyme. Build system flags and a convenience
header are provided. The functionality and interaction are demonstrated in a
new miniapp in miniapps/elasticity.
- Added support for partial assembly and fully matrix-free operators on mixed
meshes (different element types and p-adaptivity) through libCEED, including
device acceleration, e.g. with NVIDIA and AMD GPUs. The p-adaptivity is
currently limited to 2D serial meshes. All mixed element topologies are
supported in both serial and parallel.
- Added support for ParMoonolith, https://bitbucket.org/zulianp/par_moonolith,
which provides parallel non-conforming, non-matching, variational, volumetric
mesh information transfer. With ParMortarAssember, fields can be exchanged
between arbitrarily distributed and unrelated finite element meshes in a
variationally consistent way.
- Fully encapsulated SUNDIALS `N_Vector` object within the `SundialsNVector`
class by removing deprecated (e.g. `HypreParVector::ToNVector`) and
non-deprecated (e.g. `Vector::ToNVector`) functions in other classes.
- New benchmark for the different assembly levels inspired by the CEED
Bake-Off Problems, see tests/benchmarks/bench_assembly_levels.cpp.
- Added Windows 2022 CI testing with GitHub actions.
Miscellaneous
-------------
- The method SparseMatrix::EnsureMultTranspose() is now automatically called
by the methods AddMultTranspose(), MultTranspose(), and AbsMultTranspose().
Added a method with the same name to class HypreParMatrix which is also called
automatically by the HypreParMatrix::MultTranspose() methods.
- Updated various MemoryUsage methods to return 'std::size_t' instead of 'long'
since the latter is 32-bit in Win64 builds.
- When using `AssemblyLevel::FULL`, `FABilinearFormExtension::FormSystemMatrix`
outputs an `OperatorHandle` containing a `SparseMatrix` in serial, and an
`HypreParMatrix` in parallel (instead of a `ConstrainedOperator`).
- In various places in the library, replace the use of 'long' with 'long long'
to better support Win64 builds where 'long' is 32-bit and 'long long' is
64-bit. On Linux and MacOS, both types are typically 64-bit.
- The behavior of GridFunction::GetTrueVector() has been changed to not return
an empty true vector.
- Added support for ordering search points byVDIM in FindPointsGSLIB.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.4, released on March 21, 2022
=======================================
Linear and nonlinear solvers
----------------------------
- Added support for using the hypre library built with HIP support. Similar to
the existing hypre + CUDA support, most of the MFEM examples and miniapps work
transparently with hypre + HIP builds. This includes the BoomerAMG, AMS, and
ADS solvers.
- Added a simple singleton class, Hypre, to automatically set hypre's global
parameters, particularly GPU-relevant options. Updated parallel example codes
and miniapps to call Hypre::Init() where appropriate.
- Added hipSPARSE support for sparse matrix-vector multiplications.
- More explicit and consistent formatting of the output of iterative solvers
with the new IterativeSolver::PrintLevel options. See linalg/solvers.hpp.
Meshing improvements
--------------------
- New TMOP-based methods for hr-adaptivity, interface fitting, and tangential
relaxation of high-order meshes.
- Added initial support for meshes with pyramidal elements, including several
pyramidal meshes in the data/ directory and support for the lowest order H1,
Nedelec, Raviart-Thomas, and L2 basis functions on pyramids.
- Added a simpler interface to access mesh face information, see FaceInformation
and GetFaceInformation in the Mesh class.
- Gmsh meshes where all elements have zero physical tag (the default Gmsh output
format if no physical groups are defined) are now successfully loaded, and
elements are reassigned attribute number 1.
- Added ParMesh adjacency set (adjset) creation support to the Conduit Mesh
Blueprint MFEM wrapper functions in ConduitDataCollection.
Discretization improvements
---------------------------
- Added general dof transformation to support high order Nedelec basis functions
on tetrahedral meshes without reordering. The ReorientTetMesh method of the
Mesh and ParMesh classes has been deprecated. See the new DofTransformation
class in fem/doftrans.hpp.
- GPU-enabled partial (PA) and element (EA) assembly for discontinuous Galerkin
methods on nonconforming AMR meshes.
- Support for arbitrary order Nedelec and Raviart-Thomas elements on wedges.
- Added special Nedelec and Raviart-Thomas basis functions for modeling three
dimensional vector fields in 1D and 2D domains, see the new Example 31/31p.
- GridFunctionCoefficient (and the related vector, gradient, divergence, and
curl classes) now work properly with LORDiscretization and LORSolver.
- Added PA support for the action of MixedScalarCurlIntegrator in 2D and
MixedVectorGradientIntegrator in 2D and 3D, as well as their transposes.
- Coefficient::SetTime now propagates the new time into internally stored
Coefficient objects.
- Split the fem/fe.?pp files into separate files in the new fem/fe/ directory to
simplify and clarify the organization of FiniteElement classes.
New and updated examples and miniapps
-------------------------------------
- Added two new miniapps with initial support for automatic differentiation (AD)
in the miniapps/autodiff/ directory. Users can select between external library
and native implementation during configuration. The support for AD will be
extended in future releases of MFEM.
- Added Binder (mybinder.org) configuration files for C++ MFEM Jupyter Notebooks
with inline GLVis visualization in the new examples/jupyter/ directory with a
sample notebook based on Example 1. The implementation is based on xeus-cling,
see github.com/jupyter-xeus/xeus-cling and github.com/GLVis/xeus-glvis.
- Added a new miniapp (Extrapolation) for PDE-based extrapolation of finite
element functions from known values in a set of elements to the rest of the
computational domain. See miniapps/shifted/extrapolate.cpp.
- Added new miniapp that uses the ParELAG library, its hybrid smoothers, and the
hierarchy of spaces created by the element-based AMG (AMGe) methodology in
ParELAG to build multigrid solvers for H(curl) and H(div) forms. See the
miniapps/parelag directory for more details.
- Added a new Example 30/30p demonstrating support for mesh preprocessing to
resolve fine scale problem data before simulation. This feature uses adaptive
mesh refinement to control the associated data oscillation error.
- Added new Examples 31, 31p and 32p showing anisotropic definite Maxwell
serial/parallel solver and parallel eigensolver 1D, 2D, or 3D.
- Updated the mesh-optimizer and pmesh-optimizer miniapps to demonstrate the
hr-adaptivity and interface fitting capability.
- The HPC versions of ex1 and ex1p (in miniapps/performance) now support runtime
selection of either 2D or 3D meshes.
Integrations, testing and documentation
---------------------------------------
- Doxygen documentation for all releases is now available at docs.mfem.org.
- The following integrations have updated minimum version requirements:
* HIOP >= 0.4.6
* HYPRE >= 2.23.0 for HIP support
* libCEED >= 0.10
* PUMI >= 2.2.6
* RAJA >= 0.14.0
* Umpire >= 3.0.0
see INSTALL for more details.
- Added new optional integrations with ParELAG and CoDiPack (version >= 1.9.3+).
- Added initial support for Google Benchmark (version >= 1.5.6) in the
tests/benchmarks directory. It can be enabled with MFEM_USE_BENCHMARK=YES.
- Switched from Artistic Style (astyle) version 2.05.1 to version 3.1 for code
formatting. See the "make style" target.
Miscellaneous
-------------
- Added a simple singleton class, Mpi, as a replacement for MPI_Session. New
code should use Mpi::Init() and other Mpi methods instead of MPI_Session.
- Added ParaView visualization of QuadratureFunction fields, through both
QuadratureFunction::SaveVTU and ParaViewDataCollection::RegisterQField.
- Fixed several MinGW build issues on Windows.
- Added 'double' atomicAdd implementation for previous versions of CUDA.
- HypreParVector and Vector now support C++ move semantics, and the copy
constructor for HypreParVector now copies the local vector data.
- Removed the 'u' flag in the ar command, to update all files in the archive,
avoiding file name collisions from different subdirectories.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.3, released on July 29, 2021
======================================
Discretization improvements
---------------------------
- Variable order spaces, p- and hp-refinement. This is the initial (serial)
support for variable-order FiniteElementCollection and FiniteElementSpace.
The new method FiniteElementSpace::SetElementOrder can be called to set an
arbitrary order for each mesh element. The conforming interpolation matrix
will now automatically constrain p- and hp- interfaces, enabling general
hp-refinement in both 2D and 3D, on uniform or mixed NC meshes. Support for
parallel variable-order spaces will follow shortly.
- Extended the support for field transfer between high-order and low-order
refined finite element spaces to include: dual fields and H1 fields (both
primary and dual). These are illustrated in the lor-transfer miniapp.
- Improved libCEED integration, including support for VectorCoefficient,
ConvectionIntegrator, and VectorConvectionNLFIntegrator with libCEED backends.
- Extending support for L2 basis functions using MapTypes VALUE and INTEGRAL in
linear interpolators and GridFunction "GetValue" methods.
- Changed the interface for the error estimator and implemented the Kelly error
indicator for scalar-valued problems, supported in serial and parallel builds.
- Added support for the "BR2" discontinuous Galerkin discretization for
diffusion via DGDiffusionBR2Integrator (see Example 14/14p).
- Added convective and skew-symmetric integrators for the nonlinear term in the
Navier-Stokes equations.
- Added new classes DenseSymmetricMatrix and SymmetricMatrixCoefficient for
efficient evaluation of symmetric matrix coefficients. This replaces the now
deprecated EvalSymmetric in MatrixCoefficient. Added DiagonalMatrixCoefficient
for clarity, which is a typedef of VectorCoefficient.
- Added support for nonscalar coefficient with VectorDiffusionIntegrator.
Linear and nonlinear solvers
----------------------------
- Added support for AMG preconditioners on GPUs based on the hypre library
(version 2.22.0 or later). These include BoomerAMG, AMS and ADS and most
MFEM examples that use hypre have been ported to support this functionality.
The GPU preconditioners require that both hypre and MFEM are built with CUDA
support. Hypre builds with CUDA and unified memory are also supported and
can be used with `-d cuda:uvm` as a command-line option.
- Added support for AMG preconditioners for non-symmetric systems (e.g.
advection-dominated problems) using hypre's approximate ideal restriction
(AIR) AMG. Requires hypre version 2.14.0 or newer. Usage is illustrated in
example 9/9p.
- Added new functionality for constructing low-order refined discretizations and
solvers, see the LORDiscretization and LORSolver classes. A new basis type for
H(curl) and H(div) spaces is introduced to give spectral equivalence. This
functionality is illustrated in the LOR solvers miniapp in miniapps/solvers.
- Generalized the Multigrid class to support non-geometric multigrid. Previous
functionality, based on FiniteElementSpaceHierarchy, is now available in the
derived class GeometricMultigrid.
- Introduced solver interface for linear problems with constraints, a few
concrete solvers that implement the interface, and a demonstration of their
use in Example 28(p), which solves an elasticity problem with zero normal
displacement (but allowed tangential displacement) on two boundaries.
- Added high-order matrix-free auxiliary Maxwell solver for H(curl) problems,
as described in Barker and Kolev 2020 (https://doi.org/10.1002/nla.2348). See
Example 3p and linalg/auxiliary.?pp.
- Improved interface for using the Ginkgo library, including: support for matrix-
free operators in Ginkgo solvers, new wrappers for Ginkgo preconditioners, HIP
support, and reduction of unnecessary data copies.
- Added initial support for hypre's mixed integer (mixedint) capability, which
uses different data types for local and global indices in order to save memory
in large problems. This capability requires that hypre was configured with the
--enable-mixedint option. Note that this option is currently tested only in
ex1p, ex3p, and ex4p, and may not work in more general settings.
- Added AlgebraicCeedSolver that does matrix-free algebraic p-multigrid for
diffusion problems with the Ceed backend.
- Added interface to MUMPS direct solver. Its usage is demonstrated in ex25p.
See http://mumps.enseeiht.fr/ for more details. Supported versions >= 5.1.1.
- Added three ESDIRK time integrators: implicit trapezoid rule, L-stable
ESDIRK-32, and A-stable ESDIRK-33.
- Implemented a variable step-size IMEX (VSSIMEX) method for the Navier miniapp.
- Implemented an adaptive linear solver tolerance option for NewtonSolver based
on the algorithm of Eisenstat and Walker.
Meshing improvements
--------------------
- Added support for reading high-order Lagrange meshes in VTK format. Arbitrary-
orders and all element types are supported. See the VTK blog for more info:
https://blog.kitware.com/wp-content/uploads/2018/09/Source_Issue_43.pdf.
- Introduced a new non-conforming mesh format that fixes known inconsistencies
of legacy "MFEM mesh v1.1" NC format and works consistently in both serial and
parallel. ParMesh::ParPrint can now print non-conforming AMR meshes that can
be used to restart a parallel AMR computation. Example 6p has been extended to
demonstrate restarting from a previously saved checkpoint. Note that parallel
NC data files are compatible with serial code, e.g., can be viewed with serial
GLVis. Loading of legacy NC mesh files is still supported.
- Added FMS support (https://github.com/CEED/FMS) to mfem. FMS can represent
unstructured high-order meshes with general high-order finite element fields
on them. When enabled, mfem can convert data collections to/from FMS data
collections in memory. In addition, an FMS data collection class was added so
the convert-dc miniapp can read and generate data files in FMS format.
- Added new mesh quality metrics and improved the untangling capabilities of the
TMOP-based mesh optimization algorithms.
- The TMOP mesh optimization algorithms were extended to GPU:
* QualityMetric 1, 2, 7, 77 are available in 2D, 302, 303, 315, 321 in 3D
* Both AnalyticAdaptTC and DiscreteAdaptTC TargetConstructor are available
* Kernels for normalization and limiting have been added
* The AdvectorCG now also supports AssemblyLevel::PARTIAL
- Added support for creating refined meshes for all element types (e.g. by
splitting high-order elements into low-order refined elements), including
mixed meshes. The LOR Transfer miniapp (miniapps/tools/lor-transfer.cpp) now
supports meshes with any element geometry.
- Meshes consisting of any type of elements (including mixed meshes) can be
converted to all-simplex meshes using Mesh::MakeSimplicial.
- Several of the mesh constructors (creating Cartesian meshes, refined (LOR)
meshes, simplex meshes, etc.) are now available as "named constructors", e.g.
Mesh::MakeCartesian2D or Mesh::MakeRefined. The legacy constructors are marked
as deprecated.
- Added support for creating periodic meshes with Mesh::MakePeriodic. The
requisite periodic vertex mappings can be created with
Mesh::CreatePeriodicVertexMapping.
- Added support for 1D non-conforming meshes (which can be useful for parallel
load balancing and derefinement).
- Added sample meshes in the `data` subdirectory showing the reference elements
of the six currently supported element types; ref-segment.mesh,
ref-triangle.mesh, ref-square.mesh, ref-tetrahedron.mesh, ref-cube.mesh, and
ref-prism.mesh.
High-performance computing
--------------------------
- Added initial support for GPU-accelerated versions of PETSc that works with
MFEM_USE_CUDA if PETSc has been configured with CUDA support. Examples 1 and 9
in the examples/petsc directory have been modified to work with --device cuda.
Examples with GAMG (ex1p) and SLEPc (ex11p) are also provided.
- Added support for explicit vectorization in the high-performance templated
code for Fujitsu's A64FX ARM microprocessor architecture.
- Added support for different modes of QuadratureInterpolator on GPU.
The layout (QVectorLayout::byNODES|byVDIM) and the tensor products modes can
be enabled before calling the Mult, Values, Derivatives, PhysDerivatives and
Determinants methods.
- Added method Device::SetMemoryTypes that can be used to change the default
host and device MemoryTypes before Device setup.
- In class MemoryManager, added methods GetDualMemoryType and SetDualMemoryType;
dual MemoryTypes are used to determine the second MemoryType (host or device)
when only one MemoryType is specified in methods of class Memory.
- Added Memory constructor for setting both the host and device MemoryTypes.
- Switched the default behavior of device memory allocations so that they are
deferred until the device pointer is needed.
- Added a second Umpire device MemoryType, DEVICE_UMPIRE_2, with corresponding
allocator that can be set with the method SetUmpireDevice2AllocatorName.
- Added HOST_PINNED MemoryType and a pinned host allocator for CUDA and HIP.
- Added matrix-free GPU-enabled implementations of GradientInterpolator and
IdentityInterpolator.
New and updated examples and miniapps
-------------------------------------
- Added a new, very simple example (ex0 and parallel version ex0p). This example
solves a simple Poisson problem using H1 elements (the same problem as ex1),
but is intended to be extremely simple and approachable for new users.
- Added new miniapps demonstrating: 1) the use of GSLIB for overlapping grids,
see gslib/schwarz_ex1, and 2) coupling different physics in different domains,
see navier/cht. Note that gslib v1.0.7 is require (see INSTALL for details).
- Added a new miniapp for computing (signed) distance functions to a point
source or zero level set. See miniapps/shifted/distance.cpp.
- Added a high-order extension of the shifted boundary method to solve PDEs on
non body-fitted meshes. This is illustrated in the new Shifted Diffusion
miniapp, see miniapps/shifted/diffusion.cpp.
- Added new miniapp directory mtop/ with optimization-oriented block parametric
non-linear form and abstract integrators. Two new miniapps, ParHeat and
SeqHeat, demonstrate parallel and sequential implementation of gradients
evaluation for linear diffusion with discrete density.
- Added a new miniapp block-solvers that compares the performance of various
solvers for mixed finite element discretization of the second order scalar
elliptic equations. Currently available solvers in the miniapp include a
block-diagonal preconditioner that is based on approximate Schur complement
(implemented in ex5p), and a newly implemented solver DivFreeSolver, which
exploits a multilevel decomposition of the Raviart-Thomas space and its
divergence-free subspace. See the miniapps/solvers directory for more details.
- Introduced new options for the mesh-explorer miniapp to visualize the actual
element attributes in parallel meshes while retaining the visualization of the
domain decomposition.
- Added partial assembly and device support to Example 25/25p, with diagonal
preconditioning.
- Implemented a filter method for the Navier miniapp to stabilize highly
turbulent flows in direct numerical simulation.
Improved testing
----------------
- Transitioned from Travis to GitHub Action for testing/CI on GitHub.
- Use Spack (and Uberenv) to automate TPL building in LLNL GitLab tests.
- Extended `make test` to include GPU tests when MFEM is built with CUDA or HIP
support.
- Added a set of suggested git hooks for developers in config/githooks.
- Added support for Caliper: a library to integrate performance profiling
capabilities into applications. See examples/caliper for more details.
- Added a new command line boolean option (`--all`) to the unit tests to launch
*all* non-regression tests.
- Upgraded the Catch unit test framework from version 2.13.0 to version 2.13.2.
Miscellaneous
-------------
- The following integrations have updated minimum version requirements:
* CUDA >= 10.1.168
* Ginkgo >= 1.4.0
* GSLIB >= 1.0.7
* HIOP >= 0.4
* HYPRE >= 2.20.0 for mixedint support
* HYPRE >= 2.22.0 for CUDA support
* libCEED >= 0.8
* PETSc >= 3.15.0 for CUDA support
* RAJA >= 0.13.0
see INSTALL for more details.
- Added a "scaled Jacobian" visualization option in the Mesh Explorer miniapp to
help identify elements with poor mesh quality.
- Added support for reading VTK meshes in XML format.
- Added makefile rule to generate TAGS table for vi or Emacs users.
- Added HIP support to the CMake build system.
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- Added an abstract interface `mfem::FaceRestriction` for `H1FaceRestriction`
and `L2FaceRestriction`.
In order to conform with the semantic of `MultTranspose` in `mfem::Operator`,
`mfem::FaceRestriction::MultTranspose` now sets instead of adding values, and
`mfem::FaceRestriction::AddMultTranspose` should replace previous calls to
`mfem::FaceRestriction::MultTranspose`.
Version 4.2, released on October 30, 2020
=========================================
High-performance computing
--------------------------
- Added support for explicit vectorization in the high-performance templated
code, which can now take advantage of specific classes on the following
architectures:
* x86 (SSE/AVX/AVX2/AVX512),
* Power8 & Power9 (VSX),
* BG/Q (QPX).
These are disabled by default, but can be enabled with MFEM_USE_SIMD=YES.
See the new file linalg/simd.hpp and the new directory linalg/simd.
- Added an Element Assembly mode compatible with GPU device execution for H1 and
L2 spaces in the mass, convection, diffusion, transpose, and the face DG trace
integrators. See option '-ea' in Example 9. When enabled, this assembly level
stores independent dense matrices for the elements, and independent dense
matrices for the faces in the DG case.
- Added a Full Assembly mode compatible with GPU device execution. This assembly
level builds on top of the Element Assembly kernels to compute a global sparse
matrix. All integrators supported by element assembly are also supported by
full assembly. See the '-fa' option in Example 9.
- Optimized the AMD/HIP kernel support and enabled HIP support in the libCEED
integration. This is now available via the "ceed-hip" device backend.
- Improved the libCEED integration to support:
* AssemblyLevel::NONE for Mass, Diffusion, VectorMass, and VectorDiffusion
Integrators. This level computes the full operator evaluation "on the fly".
* VectorMassIntegrator and VectorDiffusionIntegrator.
* All types of (scalar) Coefficients.
- Added partial assembly / device support for:
* H(div) bilinear forms and VectorFEDivergenceIntegrator.
* BlockOperator, see the updated Example 5.
* Complex operators, see the updated Example 22.
* Chebyshev accelerated polynomial smoother.
* Convergent diagonal preconditioning on non-conforming adaptively refined
meshes, see Example 6/6p.
- Added CUDA support for:
* Sparse matrix-vector multiplication with cuSPARSE,
* SUNDIALS ODE integrators, see updated SUNDIALS modification of Example 9/9p.
Linear and nonlinear solvers
----------------------------
- Added a new solver class for simple integration with NVIDIA's multigrid
library, AmgX. The AmgX class is designed to work as a standalone solver or
preconditioner for existing MFEM solvers. It uses MFEM's sparse matrix format
for serial runs and the HypreParMatrix format for parallel runs. The new
solver may be configured to run with one GPU per MPI rank or with more MPI
ranks than GPUs. In the latter case, matrices and vectors are consolidated to
ranks communicating with the GPUs and the solution is then broadcasted.
Although CUDA is required to build, the AmgX support is compatible with the
MFEM CPU device configuration. The examples/amgx folder illustrates how to
integrate AmgX in existing MFEM applications. The AmgX solver class is
partially based on: "AmgXWrapper: An interface between PETSc and the NVIDIA
AmgX library", by Pi-Yueh Chuang and Lorena A. Barba, doi:10.21105/joss.00280.
- Added initial support for geometric h- and p-multigrid preconditioners for
matrix-based and matrix-free discretizations with basic GPU capability, see
Example 26/26p.
- Added support for the CVODES package in SUNDIALS which provides ODE solvers
with sensitivity analysis capabilities. See the CVODESSolver class and the new
adjoint miniapps in the miniapps/adjoint directory.
- Added an interface to the MKL CPardiso solver, an MPI-parallel sparse direct
solver developed by Intel. See Example 11p for an illustration of its usage.
- Added support for the SLEPc eigensolver package, https://slepc.upv.es.
- Upgraded SuperLU interface to use SuperLU_DIST 6.3.1. Added a simple SuperLU
example in the new directory examples/superlu.
- Extended the KINSOL (SUNDIALS) nonlinear solver interface to support the
Jacobian-free Newton-Krylov method. A usage example is shown in Example 10p.
- Block arrays of parallel matrices can now be merged into a single parallel
matrix with the function HypreParMatrixFromBlocks. This could be useful for
solving block systems with parallel direct solvers such as STRUMPACK.
- Added wrappers for hypre's flexible GMRES solver and the new parallel ILU
preconditioner. The latter requires hypre version 2.19.0 or later.
Discretization improvements
---------------------------
- Extended GSLIB-FindPoints integration to support simplices and interpolation
of functions from L2, H(div) and H(curl) spaces.
- Added support for computing asymptotic error estimates and convergence rates
for the whole de Rham sequence based on the new class ConvergenceStudy and new
member methods in GridFunction and ParGridFunction. See the rates.cpp file in
the tests/convergence directory for sample usage.
- Extended the GetValue and GetVectorValue methods of GridFunction to support
evaluation on boundary elements and, in the continuous field case, arbitrary
mesh edges and faces. This requires passing an ElementTransformation argument.
- Added support for matrix-free interpolation and restriction operators between
continuous H1 finite element spaces of different order on the same mesh or
with the same order on uniformly refined meshes.
- The Coefficient classes based on a C-function pointer (FunctionCoefficient,
VectorFunctionCoefficient and MatrixFunctionCoefficient) now use the more
general std::function class template. This allows the classes to be backward
compatible (i.e. they can still work with C-functions) and, in addition,
support any "callable", e.g. lambda functions.
- Non-conforming meshes are now supported with block nonlinear forms. See the
updated Example 19/19p.
Meshing improvements
--------------------
- The graph linear ordering library Gecko, previously an external dependency, is
now included directly in MFEM. As a result, Mesh::GetGeckoElementOrdering is
always available. The interface has also been improved, see for example the
Mesh Explorer miniapp.
- Added support for finite difference-based gradient and Hessian approximation
in the TMOP mesh optimization algorithms. This improves the accuracy of the
Hessian for r-adaptivity using discrete fields, and allows use of skewness
and orientation based metrics.
- Improved Gmsh reader (version 2.2), which now supports both high-order and
periodic meshes. Segments, triangles, quadrilaterals, and tetrahedra are
supported up to order 10. Wedges and hexahedra are supported up to order 9.
For sample periodic meshes, see the periodic*.msh files in the data directory.
- Added support for construction of (serial) non-conforming meshes. Hanging
nodes can be marked with Mesh::AddVertexParents when building the mesh with
the "init" constructor. The usage is demonstrated in a new meshing miniapp,
Polar NC, which generates meshes that are non-conforming from the start.
- Added support for r-adaptivity with more than one discrete field. This allows
the user to specify different discrete functions for controlling the size,
aspect-ratio, orientation, and skew of elements in the mesh.
- Additional TMOP improvements:
* Capability for approximate tangential mesh relaxation.
* Support and examples for using TMOP on mixed meshes.
* Complete integrator action accounting for spatial derivatives of discrete
and analytic targets.
New and updated examples and miniapps
-------------------------------------
- Added a new miniapp, Navier, that solves the time-dependent Navier-Stokes
equations of incompressible fluid dynamics. See the miniapps/navier directory
for more details.
- Added 10 new example codes:
* Example 25/25p demonstrates the use of a Perfectly Matched Layer (PML) for
electromagnetic wave propagation (indefinite Maxwell).
* Example 26/26p shows how to construct matrix-free geometric and p-multigrid
preconditioner for the Laplace problem.
* Example 27/27p demonstrates the enforcement of Dirichlet, Neumann, Robin,
and periodic boundary conditions with either H1 or DG Laplace problems.
* Versions of Example 1/1p in examples/amgx demonstrating the use of AmgX,
to solve the Laplace problem with AMG preconditioning on GPUs.
* A version of Example 11p in examples/petsc demonstrating the use of SLEPc,
to solve the Laplace eigenproblem with shift-and-invert transformation.
* A version of Example 1 in examples/superlu demonstrating the use of SuperLU
to solve the Laplace problem.
- Added a new Field Interpolation miniapp in miniapps/gslib that demonstrates
transfer of grid functions between different meshes using GSLIB-FindPoints.
- Added 2 miniapps in the new miniapps/adjoint directory demonstrating how to
solve adjoint problems in MFEM using the CVODES package in SUNDIALS. Both of
these miniapps require the MFEM_USE_SUNDIALS configuration option.
* The cvsRoberts_ASAi_dns miniapp solves a backward adjoint problem for a
system of ODEs, evaluating both forward and adjoint quadratures in serial.
* The adjoint_advection_diffusion miniapp solves a backward adjoint problem
for an advection diffusion PDE, evaluating adjoint quadratures in parallel.
- Added 4 additional meshing miniapps:
* The Minimal Surface miniapp solves Plateau's problem: the Dirichlet problem
for the minimal surface equation.
* The Twist miniapp demonstrates how to stitch together opposite surfaces of a
mesh to create a topologically periodic mesh.
* The Trimmer miniapp trims away parts of a mesh based on element attributes.
* Polar NC shows the construction of polar non-conforming meshes.
- Several examples and miniapps were updated to include:
* Full and element assembly support in Example 9/9p.
* Partial assembly with diagonal preconditioning in Examples 4/4p/5/5p/22/22p.
* Diagonal preconditioner in Example 6/6p for partial assembly with AMR.
* The option to plot a function in Mesh Explorer.
* A new test problem showing a mixed bilinear form for H1, H(curl), H(div) and
L2, with partial assembly support in Example 24/24p.
* Weak Dirichlet boundary conditions (Nitsche) to the NURBS miniapp.
Data management and visualization
---------------------------------
- Added support for ADIOS2 for parallel I/O with ParaView visualization. See
Examples 5, 9, 12, 16. The classes adios2stream and ADIOS2DataCollection
provide the interface to generate ADIOS2 Binary Pack (BP4) directory datasets.
- Added VTU output of boundary elements and attributes and parallel VTU (PVTU)
output of parallel meshes for visualization using ParaView.
Improved testing
----------------
- Added a GitLab pipeline that automates PR testing on supercomputing systems
and Linux clusters at Lawrence Livermore National Lab (LLNL). This can be
triggered only by LLNL developers, see .gitlab-ci.yml, the .gitlab directory
and the updated CONTRIBUTING.md file.
- Added additional testing for convergence, the parallel mesh I/O, and for the
libCEED integration in MFEM in the tests/ directory.
- Upgraded the Catch unit test framework from version 1.6.1 to version 2.13.0.
Miscellaneous
-------------
- Renamed "Backend::DEBUG" to "Backend::DEBUG_DEVICE" to avoid conflicts,
as DEBUG is sometimes used as a macro.
- Added a new IterativeSolverMonitor class that allows to monitor the residual
and solution with an IterativeSolver after every iteration.
- Added power method to iteratively estimate the largest eigenvalue and the
corresponding eigenvector of an operator.
- Added support for face integrals on the boundaries of NURBS meshes.
- In SLISolver, changed the residual inner product from (Br,r) to (Br,Br) so the
solver can work with non-SPD preconditioner B.
- Added new coefficient and vector coefficient classes for QuadratureFunctions,
with new LinearForm integrators which use them.
- The integration order used in the ComputeLpError and ComputeElementLpError
methods of class GridFunction has been increased.
- Change the IntegrationRule inside VectorDiffusionIntegrator to use the same
quadrature as DiffusionIntegrator.
- The README.html files previously included in several source directories have
been removed. Use the corresponding pages at mfem.org instead.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.1, released on March 10, 2020
=======================================
Starting with this version, the MFEM open source license is changed to BSD-3.
Improved GPU capabilities
-------------------------
- Added initial support for AMD GPUs based on HIP: a C++ runtime API and kernel
language that can run on both AMD and NVIDIA hardware.
- Added support for Umpire, a resource management library that allows the
discovery, provision, and management of memory on machines with multiple
memory devices like NUMA and GPUs, see https://github.com/LLNL/Umpire.
- GPU acceleration is now available in 3 additional examples: 3, 9 and 24.
- Improved RAJA backend and multi-GPU MPI communications.
- Added a "debug" device designed specifically to aid in debugging GPU code by
following the "device" code path (using separate host/device memory spaces and
host <-> device transfers) without any GPU hardware.
- Added support for matrix-free diagonal smoothers on GPUs.
- The current list of available device backends is: "ceed-cuda", "occa-cuda",
"raja-cuda", "cuda", "hip", "debug", "occa-omp", "raja-omp", "omp",
"ceed-cpu", "occa-cpu", "raja-cpu", and "cpu".
- The MFEM memory manager now supports different memory types, associated with
the following memory backends:
* Default host memory, using standard C++ new and delete,
* CUDA pointers, using cudaMalloc and HIP pointers, using hipMalloc,
* Managed CUDA/HIP memory (UVM), using cudaMallocManaged/hipMallocManaged,
* Umpire-managed memory, including memory pools,
* 32- or 64-byte aligned memory, using posix_memalign (WIN32 also supported),
* Debug memory with mmap/mprotect protection used by the new "debug" device.
libCEED support
---------------
- Added support for libCEED, the portable library for high-order operator
evaluation developed by the Center for Efficient Exascale Discretizations in
the Exascale Computing Project, https://github.com/CEED/libCEED.
- This initial integration includes Mass and Diffusion integrators. libCEED GPU
backends can be used without specific MFEM configuration, however it is highly
recommended to use the "cuda" build option to minimize memory transfers.
- Both CPU and GPU modes are available as MFEM device backends (ceed-cpu and
ceed-cuda), using some of the best performing CPU and GPU backends from
libCEED, see the sample runs in examples 1 and 6.
- NOTE: The current default libCEED GPU backend (ceed-cuda) uses atomics and
therefore is non-deterministic.
Partial assembly and matrix-free discretizations
------------------------------------------------
- The support for matrix-free methods on both CPU and GPU devices based on a
partially assembled operator decomposition was extended to include:
* DG integrators, (for now only in the Gauss-Lobatto basis), see Example 9,
* H(curl) bilinear forms, see Example 3,
* vector mass and vector diffusion bilinear integrators,
* convection integrator with improved performance,
* gradient and vector divergence integrators for Stokes problems,
* initial partial assembly mode for NonlinearForms.
- Diagonals of partially assembled operators can now be computed efficiently.
See the new methods AssembleDiagonal in BilinearForm, AssembleDiagonalPA in
BilinearFormIntegrator and the implementations in fem/bilininteg_*.cpp.
- In many examples, the partial assembly algorithms provide significantly
improved performance, particularly in high-order 3D runs on GPUs.
Meshing improvements
--------------------
- The algorithms for mesh element numbering were changed to have significantly
better caching and parallel partitioning properties. Both initial (see e.g.
Mesh::GetHilbertElementOrdering) and ordering after uniform refinement were
improved. NOTE: new ordering can have a round-off effect on solver results.
- Added support for non-conforming AMR on both prisms and tetrahedra, including
coarsening and parallel load balancing. Anisotropic prism refinement is only
available in the serial version at the moment.
- The TMOP mesh optimization algorithms were extended to support r-adaptivity.
Target matrices can now be constructed either via a given analytical function
(e.g. spatial dependence of size, aspect ratio, etc., for each element) or via
a (Par)GridFunction specified on the original mesh.
- The TMOP algorithms were also improved to support non-conforming AMR meshes.
- Added support for creating refined versions of periodic meshes, making use of
the new L2ElementRestriction class. This class also allows for computing
geometric factors on periodic meshes using partial assembly.
Discretization improvements
---------------------------
- Added support for GSLIB-FindPoints, a general high-order interpolation utility
that can robustly evaluate a GridFunction in an arbitrary collection of points
in physical space. See INSTALL for details on building MFEM with GSLIB, and
miniapps/gslib for examples of how to use this feature.
- Added support for complex-valued finite element operators and fields using a
2x2 block structured linear system to mimic complex arithmetic. New classes
include: ComplexGridFunction, SesquilinearForm, ComplexLinearForm, and their
parallel counterparts.
- Added second order derivatives of NURBS shape functions.
- Added support for serendipity elements of arbitrary order on affinely-mapped
square elements. Basis functions for these elements can be visualized using
an option in the display-basis miniapp.
- Two integrators related to Stokes problems, (Q grad u, v) and (Q div v, u),
where u and the components of v are in H1, were added/modified to support full
and partial assembly modes. See the new GradientIntegrator and the updated
VectorDivergenceIntegrator classes in fem/bilininteg.hpp, as well as the PA
kernels in fem/bilininteg_gradient.cpp and fem/bilininteg_divergence.cpp.
- Added a nonlinear vector valued convection integrator (Q u \cdot grad u, v)
where u_i and v_i are in H1. This form occurs e.g. in the Navier-Stokes
equations. The integrator supports the partial assembly mode for its
action. In full assembly mode we also provide the GetGradient method that
computes the linearized version of the integrator.
- Added a new method, MixedBilinearForm::FormRectangularLinearSystem, that can
be used to impose boundary conditions on the non-square off-diagonal blocks of
a block operator (similar to FormLinearSystem in the square case).
Linear and nonlinear solvers
----------------------------
- Added support for Ginkgo, a high-performance linear algebra library for GPU
and manycore nodes, with a focus on sparse solution of linear systems. For
more details see linalg/ginkgo.hpp and the example code in examples/gingko.
- Added support for HiOp, a lightweight HPC solver for nonlinear optimization
problems, see class HiOpNLPOptimizer and the example codes in examples/hiop.
- Added a general interface for specifying and solving nonlinear constrained
optimization problems through the new classes OptimizationProblem and
OptimizationSolver, see linalg/solver.hpp.
- Added a block ILU(0) preconditioner for DG-type discretizations. Example 9
(DG advection) now takes advantage of this for implicit time integration.
- New time integrators: Adams-Bashforth, Adams-Moulton and several integrators
for 2nd order ODEs, see the new Example 23.
- Added a LinearSolve(A,X) convenience method to solve dense linear systems. In
the trivial cases, i.e., square matrices of size 1 or 2, the system is solved
directly, otherwise, LU factorization is employed.
New and updated examples and miniapps
-------------------------------------
- Added a collection of 7 playful miniapps in miniapps/toys that illustrate the
meshing and visualization features of the library in more relaxed settings.
The toys include simulations of cellular automata, Rubik's cube, Mandelbrot
set, a tool to convert any image to mfem mesh, and more.
- Added 8 new example codes:
* Example 22/22p demonstrates the use of the new complex-valued finite element
operators by defining and solving a family of time-harmonic PDEs related to
damped harmonic oscillators.
* Example 23 solves a simple 2D/3D wave equation with the new second order
time integrators.
* Example 24/24p demonstrates usage of mixed finite element spaces in bilinear
forms. Partial assembly is supported in this example.
* A version of Example 1 in examples/ginkgo demonstrating the use of the
Gingko interface to solve a linear system.
* A version of Example 9/9p in examples/hiop demonstrating the nonlinear
constrained optimization interface and use of the SLBQP and HiOp solvers.
- Added two new miniapps: Find Points and Field Diff in miniapps/gslib that show
how GSLIB-FindPoints can be used to interpolate a (Par) GridFunction in an
arbitrary number of physical space points in 2D and 3D. The GridFunction must
be in H1 and in the same space as the mesh that is used to find the points.
- Added a simple miniapp, Get Values, that extracts field values at a set of
points, from previously saved data via DataCollection classes.
- Several examples and miniapps were updated:
* Added device support in Example 3/3p and Example 9/9p.
* Example 1/1p and Example 3/3p now use diagonal preconditioning in partial
assembly mode.