forked from mfem/mfem
-
Notifications
You must be signed in to change notification settings - Fork 0
/
CHANGELOG
2697 lines (2062 loc) · 127 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 4.7.1 (development)
===========================
- Refactored ALGOIM cut integration rules. The interface is unified with
the interface for moment based cut integration rules.
Discretization improvements
---------------------------
- Added NURBS-based H(div) and H(curl) elements in 2D and 3D. Only on single
patch meshes. Only implemented for serial computations.
- Added support for boundary constraints to the hybridization class.
- Added support for external boundary submeshes with nonconformal mesh adaptation.
Meshing improvements
--------------------
- The ExodusII reader now handles pyramid and wedge element types. Mixed meshes
are also supported.
New and updated examples and miniapps
-------------------------------------
- Added miniapps to demonstrate the H(div) and H(curl) NURBS elements.
- Added an MFEM example for the eikonal equation. This new solver is based on
the proximal Galerkin method introduced by Keith and Surowiec.
- Added a command line option to all miniapps (`-p` or `--send-port`) for
specifying the GLVis server socket port (19916 by default).
GPU computing
-------------
- Added support for GPU-accelerated batched linear algebra (using cuBLAS,
hipBLAS, MAGMA, or native MFEM functionality) through the BatchedLinAlg class.
- A new GPU kernel dispatch mechanism was introduced. Users can instantiate
specialized kernels for specific combinations of (for example) polynomial
degree and number of quadrature points using
`DiffusionIntegrator::AddSpecialization` and
`MassIntegrator::AddSpecialization` (this functionality may be added to more
integrators in the future).
- Calls to slower fallback kernels can be reported to `mfem::err` by setting
the environment variable `MFEM_REPORT_KERNELS` to any value other than `NO`
or by explicitly calling `KernelReporter::Enable`. Users can then add
specializations for these kernels to achieve higher performance.
- Element assembly kernels have been added for low-order refined to
high-order transfer operators. New kernels can be offloaded as device
kernels. Example usage may be found in lor-transfer.cpp under miniapps/tools.
Miscellaneous
-------------
- Added support for SUNDIALS v7. See the section "API changes" for some small
changes related to this new version.
- Refactored the `ARKStepSolver` class (ARKODE interface) to use
`TimeDependentOperator::Mult` only when the associated ODE operator is
expressed in explicit form (i.e., `TimeDependentOperator::isExplicit()`),
otherwise `TimeDependentOperator::ExplicitMult` is used. A check has been
added to `ARKStepSolver` to verify that the associated ODE operator is not in
explicit form when a mass matrix solver is enabled via a call to either the
`UseMFEMMassLinearSolver` or `UseSundialsMassLinearSolver` methods. This is
because enabling a mass matrix solver assumes that F(u,k,t) = M k in the
associated ODE operator.
- Added support for custom interpolation procedure in FindPointsGSLIB.
- `FiniteElementSpace` has new methods to directly set prolongation and
restriction operators to arbitrary sparse matrices.
- There are new convenience constructors for NURBS patches and knot vectors.
API changes
-----------
- API change: in class GridFunction, 'fec' was renamed to 'fec_owned'.
- API change: support for SUNDIALS v7:
* the SUNDIALS types `realtype` and `booleantype` are no longer defined by v7
and therefore MFEM now uses the new type names `sunrealtype` and
`sunbooleantype`, respectively, which MFEM defines when using SUNDIALS < v6
where these types were not defined.
* The SUNDIALS macro `SUNLS_SUCCESS` and some other `*_SUCCESS` macros were
removed and replaced by `SUN_SUCCESS` in v7, so to avoid tedious checks for
SUNDIALS versions, MFEM now defines and uses the constant `SUN_SUCCESS` when
using SUNDIALS < v7.
* The constants `SUN_PREC_*`, introduced by SUNDIALS v6 are now introduced by
MFEM when using SUNDIALS < v6 to avoid tedious version checks.
Version 4.7, released on May 7, 2024
====================================
- Added support for single precision (with corresponding hypre build). The MFEM
floating point type was generalized from `double` to `real_t`. For details see
https://github.com/orgs/mfem/discussions/4207.
Meshing improvements
--------------------
- Added the capability to partition (big) serial meshes in serial code, see the
new classes MeshPartitioner and MeshPart. This capability is also exposed as a
menu option in the mesh-explorer miniapp in miniapps/meshing.
- Added named attribute sets and basic supporting methods to the Mesh class as a
convenient means of referring to sets of domain or boundary attribute numbers.
See the new Example 39/39p and data/compass.mesh.
- Introduced formulas for refinement of patches in NURBS meshes. Refinement by
arbitrary integer factors is also enabled, e.g. in the mesh-explorer miniapp.
NURBS coarsening and knot removal are also introduced.
- Added support for internal boundary elements in nonconforming meshes.
- Added ExodusII output capability. The writer can handle first-order (Pyramid5,
Wedge6, Hex8, Tet4) and second-order FE types (Pyramid14, Wedge18, Hex27, Tet10).
- The ReadCubit Genesis mesh importer has been rewritten to improve readability.
Discretization improvements
---------------------------
- Added a new nonlinear integrator, `HyperbolicFormIntegrator` that implements
both element-wise weak divergence and face-wise numerical flux for a general
system of hyperbolic conservation laws. To use the integrator for a specific
flux function, users can define a derived class of `FluxFunction`. Currently,
advection, Burgers, shallow-water and Euler equations (see Example 18/18p) are
available.
- Added a capability to construct cut-surface and cut-volume IntegrationRules
through a moment-fitting approach. The cut is specified by the zero level set
of a Coefficient. See fem/intrules_cut.hpp and the new Example 38.
- Introduced support for high-order nonconforming Nedelec elements on simplices.
GPU computing
-------------
- Added partial assembly and GPU support for the DG diffusion integrator.
- Efficient GPU-accelerated LOR assembly is now supported on surface meshes.
- Added functionality to automatically configure hypre's compute policy to match
MFEM's compute policy when hypre is built with GPU support. Requires version
hypre-2.31.0 or later.
- Added support for full assembly on simplices.
- Added partial assembly for linear elasticity (no sum factorization for now).
- Added functionality for BilinearFormIntegrators to use kernels that work for
both tensor and unstructured elements.
- The RAJA backend will use `seq_exec` for serial loop execution when RAJA
v2023.06.00 and beyond is detected as `loop_exec` is deprecated.
- API change: The macro MFEM_HYPRE_FORALL (from hypre.hpp) which was intended
for internal use, has been removed and replaced by the function template
mfem::hypre_forall in general/forall.hpp.
New and updated examples and miniapps
-------------------------------------
- Added a new miniapp illustrating elastic contact based on the Tribol library,
(https://github.com/LLNL/Tribol). See miniapps/tribol.
- Added a miniapp to demonstrate low order refined (LOR) block preconditioning
for linear elasticity on GPUs. See miniapps/solvers/lor_elast.
- Added a new block solver in miniapp/solvers for the Darcy problem. The new
solver is based on a Bramble-Pasciak preconditioning. User can use and
implement their own preconditioner for the mass matrix.
- Added a small miniapp for printing the shape functions of a KnotVector. See
miniapps/nurbs/nurbs_printfunc.cpp.
- Added two new example codes: 38 and 39/39p described above. Substantially
updated Example 18/18p.
- Added ODE solvers selection routines. This creates a uniformity across examples,
miniapps and other executables in regard to ODE(time-integrator) selection.
- Added new mechanism for retrieving and setting state vectors in ODE solvers.
This is relevant for AB/AM and gen-alpha solvers.
- Added ODEsolver/ODEsolver2 unit tests to verify order of convergence and
read/write functionality.
Miscellaneous
-------------
- Updated the Doxygen documentation style, which now requires Doxygen version
1.9.8 or later. See the doc/ directory.
- Improved thread safety for global variables in the library, e.g. for IntRules,
RefinedIntRules, GlobGeometryRefiner, and FiniteElement::dof2quad_array.
- PETSc integration now generally requires PETSc version 3.21 or later, though
depending on the functionality older versions may still work.
- Various other simplifications, extensions, and bugfixes in the code.
- Added GSLIB-based gather-scatter operator.
Version 4.6, released on September 27, 2023
===========================================
- MFEM is now available in Homebrew and can be installed on a Mac with just
"brew install mfem". See https://formulae.brew.sh/formula/mfem.
Meshing improvements
--------------------
- Added asymptotically-balanced TMOP compound metrics 90, 94, 328, 338. A new
tool, tmop-metric-magnitude, can be used to track how metrics change under
geometric perturbations. See miniapps/tools.
- Several NURBS meshing improvements:
* Support for free connectivity of NURBS patches allowing for more complex
patch configurations such as C-meshes.
* New methods to set and get attributes on NURBS patches and patch boundaries.
* The edge to knot map for NURBS meshes can be determined automatically. It is
no longer needed to specify this in the NURBS mesh.
* Added curve interpolation method for NURBS.
* See miniapps/nurbs for example meshes and miniapps.
Discretization improvements
---------------------------
- SubMesh and ParSubMesh have been extended to support the transfer of
Nedelec and Raviart-Thomas finite element spaces.
- Added support for partial assembly on NURBS patches, and NURBS-patch sparse
matrix assembly. Patch matrix assembly includes the option to use reduced
approximate integration rules, computed by the newly implemented non-negative
least-squares (NNLS) solver.
- Support for parallel transfer of H1 fields using the low-order refined (LOR)
transfer operators in L2ProjectionGridTransfer
- Added KDTree class for 2D/3D set of points, which is then utilized in the new
KDTreeNodalProjection class to project a function defined on an arbitrary set
of points onto an MFEM grid function. This functionality is demonstrated in
the nodal-transfer miniapp. The current implementation is serial only. Further
extensions can include search in arbitrary dimensional spaces.
- Added support for p-refined meshes in GSLIB-FindPoints.
- Device kernels can now access device-specific DOF and quadrature limits using
the DofQuadLimits structure, allowing increased limits when executing on CPU.
The limits for the runtime selected device can be accessed in host code using
DeviceDofQuadLimits::Get(). The global constants MAX_D1D and MAX_Q1D are no
longer available.
- Face restriction operators for Nedelec and Raviart-Thomas finite element
spaces are now supported through the ConformingFaceRestriction class.
- VectorFEBoundaryFluxLFIntegrator is now supported on device/GPU.
Linear and nonlinear solvers
----------------------------
- Updated the MUMPS interface to support multiple right-hand sides, block
low-rank compression, builds using 64-bit integers, and other improvements.
- Added an interface to the MKL Pardiso sparse direct solver developed by Intel.
The interface provides a serial (OpenMP shared memory) version of Pardiso for
use with SparseMatrix. This complements the existing parallel (MPI distributed
memory) version already available through the CPardiso MFEM integration.
- Added HIP support to the PETSc and SUNDIALS interfaces.
New and updated examples and miniapps
-------------------------------------
- Added a new H(div) solver miniapp demonstrating the use of a matrix-free
saddle-point solver methodology, suitable for high-order discretizations and
for GPU acceleration. Examples illustrating the solution of Darcy and grad-div
problems are included. See miniapps/hdiv-linear-solver.
- Added new Discontinuous Petrov-Galerkin (DPG) miniapp which includes serial
and parallel examples for diffusion, convection-diffusion, acoustics and
Maxwell equations. The miniapp includes new classes such as (Par)DPGWeakForm,
(Par)ComplexDPGWeakForm and (Complex)BlockStaticCondensation. Three new
integrators are added in support of DPG systems: TraceIntegrator,
NormalTraceIntegrator and TangentTraceIntegrator. See miniapps/dpg.
- Added a new miniapp that implements the SPDE method for generating Gaussian
random fields of Matern covariance. The resulting random field can be used,
e.g., to model material uncertainties. See miniapps/spde.
- Added a new parallel LOR transfer miniapp, plor-transfer, which mirrors the
functionality of the serial LOR transfer miniapp. See miniapps/tools.
- New serial miniapp, nodal-transfer, demonstrating the use of KDTree to map a
parallel grid function to a different parallel partitioning of the same mesh.
- Added 3 additional TMOP miniapps in miniapps/meshing:
* Mesh-Quality evaluates quality using size, skewness, and aspect-ratio
computed from the Jacobian of the transformation.
* Mesh-Fitting can be used for interface and boundary fitting to implicit
domains defined using level-set functions.
* Fit-Node-Position fits selected mesh nodes to specified positions, while
maintaining overall mesh quality.
- Added 4 new example codes:
* Example 34/34p solves a simple magnetostatic problem where source terms and
boundary conditions are transferred with SubMesh objects.
* Example 35p implements H1, H(curl) and H(div) variants of a damped harmonic
oscillator with field transfer using SubMesh objects.
* Example 36/36p demonstrates the solution of the obstacle problem with a new
finite element method (proximal Galerkin).
* Example 37/37p demonstrates topology optimization with MFEM.
- Added a random refinement option to the mesh-explorer miniapp to assist users
in experimenting with nonconforming meshes.
- Moved the distance solver methods from miniapps/shifted to miniapps/common.
Miscellaneous
-------------
- Improved lambda body debugging with the addition of mfem::forall functions.
These functions can take the place of the MFEM_FORALL macros, which have been
preserved for backwards compatibility.
- Added an address sanitizer GitHub action for a serial build/test on Ubuntu,
based on Clang/LLVM (https://clang.llvm.org/docs/AddressSanitizer.html).
- Reorganized files for bilinear form, linear form, and nonlinear form integrators
in the fem/integ/ subdirectory.
- FiniteElementSpace::GetFE has been updated to abort instead of returning NULL for
an empty partition.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.5.2, released on March 23, 2023
=========================================
- Added support for pyramids in non-conforming meshes. Currently only isotropic
refinement is supported in this case.
- Removed the support for the Mesquite toolkit. We recommend using MFEM's TMOP
functionality instead for mesh optimization. See the mesh-optimizer miniapp.
- Added a fast normalization-based distance solver, see the Distance miniapp
in the miniapps/shifted/ directory.
- Added a new meshing miniapp, Reflector, which reflects a high-order or NURBS
hexahedral mesh about a plane.
- Updated logic in FindPointsGSLIB to ignore points found near (but outside) the
domain boundary.
- Added an option to auto-balance compound TMOP metrics.
- Fixed a bug in TMOP metric 301.
- When using discontinuous (L2) spaces, use local (element-wise) L2 projection
as the coarsening operator for non-conforming AMR meshes.
- Added support for GridFunction::GetGradients() and GetVectorGradient() on
face-neighbor elements.
- Added support for pyramids in Gmsh meshes.
- The Mesh Explorer miniapp can now save mesh files in the VisIt or ParaView
formats using the corresponding DataCollection. See option 'D' in the menu.
- VisItDataCollection now correctly handles data collection names containing
underscores.
- Added support for shared Windows builds with MSVC through CMake.
Developers note: this enhancement is facilitated by the addition of the macro
MFEM_EXPORT, see config.hpp for more details on its usage.
- The following integrations have updated minimum version requirements:
* RAJA >= 2022.10.3
API changes
-----------
- The implicit cast methods of class Vector to 'double *' and 'const double *'
have been deprecated and generate deprecation warnings if used. They will be
removed in a future release.
- The methods Mesh::GetFaceBaseGeometry and Mesh::GetFaceGeometryType have been
deprecated, and Mesh::GetFaceGeometry (which provides identical functionality)
should be used instead.
- VisItDataCollection::SetPadDigits() no longer alters the number of digits
used to represent the MPI rank because VisIt seems to require 6 digits.
This parameter can still be explicitly overridden with SetPadDigitsRank().
Version 4.5, released on October 22, 2022
=========================================
Meshing improvements
--------------------
- Added new SubMesh and ParSubMesh classes that can be used to extract a subset
of a given Mesh. These classes have the same functionality as Mesh and ParMesh
and work with all existing MFEM interfaces like finite element spaces etc.
- Added a method, ParMesh::GetSerialMesh(), that reconstructs a partitioned
parallel mesh on a given single rank. Also, added ParMesh::PrintAsSerial(),
which saves the reconstructed serial mesh to a C++ stream on rank 0.
- Added more 3D TMOP metrics, as well as specialized metrics for mesh
untangling and worst-case quality improvement.
- Added a new method, Mesh::NodesUpdated, which should be called after the mesh
node coordinates have changed, e.g. after the mesh has moved. This is
necessary, for example, with device assembly of linear and bilinear forms.
- Added support for mixed meshes and pyramids in GSLIB-FindPoints.
Discretization improvements
---------------------------
- Added full assembly and device support for several LinearForm integrators:
* DomainLF: (f, v)
* VectorDomainLF: ((f1,...,fn), (v1,...,vn))
* DomainLFGrad: (f, grad(v))
* VectorDomainLFGrad: ((f1x,f1y,f1z,...,fnx,fny,fnz), grad(v1,...,vn))
The device assembly of linear forms has to be explicitly enabled by calling
LinearForm::UseFastAssembly(true), otherwise the legacy linear form assembly
is used by default.
- Added support for assembling low-order-refined matrices using a GPU-enabled
"batched" algorithm. The lor_solvers and plor_solvers now fully support GPU
acceleration with arbitrary user-supplied coefficients.
- Added a new class FaceQuadratureSpace that allows for the construction of
QuadratureFunctions on the interior or boundary faces of a mesh.
- Added a class CoefficientVector for efficient access of variable coefficient
values at quadrature points (in particular for GPU/device kernels).
- Added WhiteGaussianNoiseDomainLFIntegrator: a LinearFormIntegrator class for
spatial Gaussian white noise.
- Added a new Zienkiewicz-Zhu patch recovery-based a posteriori error estimator.
See fem/estimators.hpp.
- Various fixes and improvements in LinearFormExtension.
Linear and nonlinear solvers
----------------------------
- Added a new class DGMassInverse that performs a local element-wise CG
iteration to solve systems involving the discontinuous Galerkin mass matrix,
including support for device/GPU acceleration.
- Added more flexibility to the constrained solver classes:
* PenaltyConstrainedSolver now allows for a vector of penalty parameters
(necessary for penalty contact)
* PenaltyConstrainedSolver and EliminationSolver can use GMRES or PCG
* All constraint solver classes can take a user-defined preconditioner
- Added functions to toggle additional options for the SuperLU_Dist and Hypre
preconditioners (ParaSails, Euclid, ILU).
- Added boundary elimination with device support for `SparseMatrix` and
`HypreParMatrix`.
New and updated examples and miniapps
-------------------------------------
- Added a new elasticity miniapp, Hooke, that showcases a low-level approach of
using MFEM to solve a nonlinear elasticity problem based on the fundamental
finite element operator decomposition. The miniapp also integrates with
automatic differentiation tools like a native dual number implementation or a
third party library such as Enzyme. See miniapps/elasticity for more details.
- Added example for body-fitted volumetric and shape integration using the
Algoim library in miniapps/shifted.
- Add a new example code, Example 33/33p, to demonstrate the solution of
spectral fractional PDEs with MFEM.
Integrations, testing and documentation
---------------------------------------
- Added a Dockerfile for a simple MFEM container, see config/docker/README.md.
More sophisticated developer containers are available in the new repo
https://github.com/mfem/containers.
- Added support for the LLVM-based automatic differentiation tool Enzyme, see
https://github.com/EnzymeAD/Enzyme. Build system flags and a convenience
header are provided. The functionality and interaction are demonstrated in a
new miniapp in miniapps/elasticity.
- Added support for partial assembly and fully matrix-free operators on mixed
meshes (different element types and p-adaptivity) through libCEED, including
device acceleration, e.g. with NVIDIA and AMD GPUs. The p-adaptivity is
currently limited to 2D serial meshes. All mixed element topologies are
supported in both serial and parallel.
- Added support for ParMoonolith, https://bitbucket.org/zulianp/par_moonolith,
which provides parallel non-conforming, non-matching, variational, volumetric
mesh information transfer. With ParMortarAssember, fields can be exchanged
between arbitrarily distributed and unrelated finite element meshes in a
variationally consistent way.
- Fully encapsulated SUNDIALS `N_Vector` object within the `SundialsNVector`
class by removing deprecated (e.g. `HypreParVector::ToNVector`) and
non-deprecated (e.g. `Vector::ToNVector`) functions in other classes.
- New benchmark for the different assembly levels inspired by the CEED
Bake-Off Problems, see tests/benchmarks/bench_assembly_levels.cpp.
- Added Windows 2022 CI testing with GitHub actions.
Miscellaneous
-------------
- The method SparseMatrix::EnsureMultTranspose() is now automatically called
by the methods AddMultTranspose(), MultTranspose(), and AbsMultTranspose().
Added a method with the same name to class HypreParMatrix which is also called
automatically by the HypreParMatrix::MultTranspose() methods.
- Updated various MemoryUsage methods to return 'std::size_t' instead of 'long'
since the latter is 32-bit in Win64 builds.
- When using `AssemblyLevel::FULL`, `FABilinearFormExtension::FormSystemMatrix`
outputs an `OperatorHandle` containing a `SparseMatrix` in serial, and an
`HypreParMatrix` in parallel (instead of a `ConstrainedOperator`).
- In various places in the library, replace the use of 'long' with 'long long'
to better support Win64 builds where 'long' is 32-bit and 'long long' is
64-bit. On Linux and MacOS, both types are typically 64-bit.
- The behavior of GridFunction::GetTrueVector() has been changed to not return
an empty true vector.
- Added support for ordering search points byVDIM in FindPointsGSLIB.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.4, released on March 21, 2022
=======================================
Linear and nonlinear solvers
----------------------------
- Added support for using the hypre library built with HIP support. Similar to
the existing hypre + CUDA support, most of the MFEM examples and miniapps work
transparently with hypre + HIP builds. This includes the BoomerAMG, AMS, and
ADS solvers.
- Added a simple singleton class, Hypre, to automatically set hypre's global
parameters, particularly GPU-relevant options. Updated parallel example codes
and miniapps to call Hypre::Init() where appropriate.
- Added hipSPARSE support for sparse matrix-vector multiplications.
- More explicit and consistent formatting of the output of iterative solvers
with the new IterativeSolver::PrintLevel options. See linalg/solvers.hpp.
Meshing improvements
--------------------
- New TMOP-based methods for hr-adaptivity, interface fitting, and tangential
relaxation of high-order meshes.
- Added initial support for meshes with pyramidal elements, including several
pyramidal meshes in the data/ directory and support for the lowest order H1,
Nedelec, Raviart-Thomas, and L2 basis functions on pyramids.
- Added a simpler interface to access mesh face information, see FaceInformation
and GetFaceInformation in the Mesh class.
- Gmsh meshes where all elements have zero physical tag (the default Gmsh output
format if no physical groups are defined) are now successfully loaded, and
elements are reassigned attribute number 1.
- Added ParMesh adjacency set (adjset) creation support to the Conduit Mesh
Blueprint MFEM wrapper functions in ConduitDataCollection.
Discretization improvements
---------------------------
- Added general dof transformation to support high order Nedelec basis functions
on tetrahedral meshes without reordering. The ReorientTetMesh method of the
Mesh and ParMesh classes has been deprecated. See the new DofTransformation
class in fem/doftrans.hpp.
- GPU-enabled partial (PA) and element (EA) assembly for discontinuous Galerkin
methods on nonconforming AMR meshes.
- Support for arbitrary order Nedelec and Raviart-Thomas elements on wedges.
- Added special Nedelec and Raviart-Thomas basis functions for modeling three
dimensional vector fields in 1D and 2D domains, see the new Example 31/31p.
- GridFunctionCoefficient (and the related vector, gradient, divergence, and
curl classes) now work properly with LORDiscretization and LORSolver.
- Added PA support for the action of MixedScalarCurlIntegrator in 2D and
MixedVectorGradientIntegrator in 2D and 3D, as well as their transposes.
- Coefficient::SetTime now propagates the new time into internally stored
Coefficient objects.
- Split the fem/fe.?pp files into separate files in the new fem/fe/ directory to
simplify and clarify the organization of FiniteElement classes.
New and updated examples and miniapps
-------------------------------------
- Added two new miniapps with initial support for automatic differentiation (AD)
in the miniapps/autodiff/ directory. Users can select between external library
and native implementation during configuration. The support for AD will be
extended in future releases of MFEM.
- Added Binder (mybinder.org) configuration files for C++ MFEM Jupyter Notebooks
with inline GLVis visualization in the new examples/jupyter/ directory with a
sample notebook based on Example 1. The implementation is based on xeus-cling,
see github.com/jupyter-xeus/xeus-cling and github.com/GLVis/xeus-glvis.
- Added a new miniapp (Extrapolation) for PDE-based extrapolation of finite
element functions from known values in a set of elements to the rest of the
computational domain. See miniapps/shifted/extrapolate.cpp.
- Added new miniapp that uses the ParELAG library, its hybrid smoothers, and the
hierarchy of spaces created by the element-based AMG (AMGe) methodology in
ParELAG to build multigrid solvers for H(curl) and H(div) forms. See the
miniapps/parelag directory for more details.
- Added a new Example 30/30p demonstrating support for mesh preprocessing to
resolve fine scale problem data before simulation. This feature uses adaptive
mesh refinement to control the associated data oscillation error.
- Added new Examples 31, 31p and 32p showing anisotropic definite Maxwell
serial/parallel solver and parallel eigensolver 1D, 2D, or 3D.
- Updated the mesh-optimizer and pmesh-optimizer miniapps to demonstrate the
hr-adaptivity and interface fitting capability.
- The HPC versions of ex1 and ex1p (in miniapps/performance) now support runtime
selection of either 2D or 3D meshes.
Integrations, testing and documentation
---------------------------------------
- Doxygen documentation for all releases is now available at docs.mfem.org.
- The following integrations have updated minimum version requirements:
* HIOP >= 0.4.6
* HYPRE >= 2.23.0 for HIP support
* libCEED >= 0.10
* PUMI >= 2.2.6
* RAJA >= 0.14.0
* Umpire >= 3.0.0
see INSTALL for more details.
- Added new optional integrations with ParELAG and CoDiPack (version >= 1.9.3+).
- Added initial support for Google Benchmark (version >= 1.5.6) in the
tests/benchmarks directory. It can be enabled with MFEM_USE_BENCHMARK=YES.
- Switched from Artistic Style (astyle) version 2.05.1 to version 3.1 for code
formatting. See the "make style" target.
Miscellaneous
-------------
- Added a simple singleton class, Mpi, as a replacement for MPI_Session. New
code should use Mpi::Init() and other Mpi methods instead of MPI_Session.
- Added ParaView visualization of QuadratureFunction fields, through both
QuadratureFunction::SaveVTU and ParaViewDataCollection::RegisterQField.
- Fixed several MinGW build issues on Windows.
- Added 'double' atomicAdd implementation for previous versions of CUDA.
- HypreParVector and Vector now support C++ move semantics, and the copy
constructor for HypreParVector now copies the local vector data.
- Removed the 'u' flag in the ar command, to update all files in the archive,
avoiding file name collisions from different subdirectories.
- Various other simplifications, extensions, and bugfixes in the code.
Version 4.3, released on July 29, 2021
======================================
Discretization improvements
---------------------------
- Variable order spaces, p- and hp-refinement. This is the initial (serial)
support for variable-order FiniteElementCollection and FiniteElementSpace.
The new method FiniteElementSpace::SetElementOrder can be called to set an
arbitrary order for each mesh element. The conforming interpolation matrix
will now automatically constrain p- and hp- interfaces, enabling general
hp-refinement in both 2D and 3D, on uniform or mixed NC meshes. Support for
parallel variable-order spaces will follow shortly.
- Extended the support for field transfer between high-order and low-order
refined finite element spaces to include: dual fields and H1 fields (both
primary and dual). These are illustrated in the lor-transfer miniapp.
- Improved libCEED integration, including support for VectorCoefficient,
ConvectionIntegrator, and VectorConvectionNLFIntegrator with libCEED backends.
- Extending support for L2 basis functions using MapTypes VALUE and INTEGRAL in
linear interpolators and GridFunction "GetValue" methods.
- Changed the interface for the error estimator and implemented the Kelly error
indicator for scalar-valued problems, supported in serial and parallel builds.
- Added support for the "BR2" discontinuous Galerkin discretization for
diffusion via DGDiffusionBR2Integrator (see Example 14/14p).
- Added convective and skew-symmetric integrators for the nonlinear term in the
Navier-Stokes equations.
- Added new classes DenseSymmetricMatrix and SymmetricMatrixCoefficient for
efficient evaluation of symmetric matrix coefficients. This replaces the now
deprecated EvalSymmetric in MatrixCoefficient. Added DiagonalMatrixCoefficient
for clarity, which is a typedef of VectorCoefficient.
- Added support for nonscalar coefficient with VectorDiffusionIntegrator.
Linear and nonlinear solvers
----------------------------
- Added support for AMG preconditioners on GPUs based on the hypre library
(version 2.22.0 or later). These include BoomerAMG, AMS and ADS and most
MFEM examples that use hypre have been ported to support this functionality.
The GPU preconditioners require that both hypre and MFEM are built with CUDA
support. Hypre builds with CUDA and unified memory are also supported and
can be used with `-d cuda:uvm` as a command-line option.
- Added support for AMG preconditioners for non-symmetric systems (e.g.
advection-dominated problems) using hypre's approximate ideal restriction
(AIR) AMG. Requires hypre version 2.14.0 or newer. Usage is illustrated in
example 9/9p.
- Added new functionality for constructing low-order refined discretizations and
solvers, see the LORDiscretization and LORSolver classes. A new basis type for
H(curl) and H(div) spaces is introduced to give spectral equivalence. This
functionality is illustrated in the LOR solvers miniapp in miniapps/solvers.
- Generalized the Multigrid class to support non-geometric multigrid. Previous
functionality, based on FiniteElementSpaceHierarchy, is now available in the
derived class GeometricMultigrid.
- Introduced solver interface for linear problems with constraints, a few
concrete solvers that implement the interface, and a demonstration of their
use in Example 28(p), which solves an elasticity problem with zero normal
displacement (but allowed tangential displacement) on two boundaries.
- Added high-order matrix-free auxiliary Maxwell solver for H(curl) problems,
as described in Barker and Kolev 2020 (https://doi.org/10.1002/nla.2348). See
Example 3p and linalg/auxiliary.?pp.
- Improved interface for using the Ginkgo library, including: support for matrix-
free operators in Ginkgo solvers, new wrappers for Ginkgo preconditioners, HIP
support, and reduction of unnecessary data copies.
- Added initial support for hypre's mixed integer (mixedint) capability, which
uses different data types for local and global indices in order to save memory
in large problems. This capability requires that hypre was configured with the
--enable-mixedint option. Note that this option is currently tested only in
ex1p, ex3p, and ex4p, and may not work in more general settings.
- Added AlgebraicCeedSolver that does matrix-free algebraic p-multigrid for
diffusion problems with the Ceed backend.
- Added interface to MUMPS direct solver. Its usage is demonstrated in ex25p.
See http://mumps.enseeiht.fr/ for more details. Supported versions >= 5.1.1.
- Added three ESDIRK time integrators: implicit trapezoid rule, L-stable
ESDIRK-32, and A-stable ESDIRK-33.
- Implemented a variable step-size IMEX (VSSIMEX) method for the Navier miniapp.
- Implemented an adaptive linear solver tolerance option for NewtonSolver based
on the algorithm of Eisenstat and Walker.
Meshing improvements
--------------------
- Added support for reading high-order Lagrange meshes in VTK format. Arbitrary-
orders and all element types are supported. See the VTK blog for more info:
https://blog.kitware.com/wp-content/uploads/2018/09/Source_Issue_43.pdf.
- Introduced a new non-conforming mesh format that fixes known inconsistencies
of legacy "MFEM mesh v1.1" NC format and works consistently in both serial and
parallel. ParMesh::ParPrint can now print non-conforming AMR meshes that can
be used to restart a parallel AMR computation. Example 6p has been extended to
demonstrate restarting from a previously saved checkpoint. Note that parallel
NC data files are compatible with serial code, e.g., can be viewed with serial
GLVis. Loading of legacy NC mesh files is still supported.
- Added FMS support (https://github.com/CEED/FMS) to mfem. FMS can represent
unstructured high-order meshes with general high-order finite element fields
on them. When enabled, mfem can convert data collections to/from FMS data
collections in memory. In addition, an FMS data collection class was added so
the convert-dc miniapp can read and generate data files in FMS format.
- Added new mesh quality metrics and improved the untangling capabilities of the
TMOP-based mesh optimization algorithms.
- The TMOP mesh optimization algorithms were extended to GPU:
* QualityMetric 1, 2, 7, 77 are available in 2D, 302, 303, 315, 321 in 3D
* Both AnalyticAdaptTC and DiscreteAdaptTC TargetConstructor are available
* Kernels for normalization and limiting have been added
* The AdvectorCG now also supports AssemblyLevel::PARTIAL
- Added support for creating refined meshes for all element types (e.g. by
splitting high-order elements into low-order refined elements), including
mixed meshes. The LOR Transfer miniapp (miniapps/tools/lor-transfer.cpp) now
supports meshes with any element geometry.
- Meshes consisting of any type of elements (including mixed meshes) can be
converted to all-simplex meshes using Mesh::MakeSimplicial.
- Several of the mesh constructors (creating Cartesian meshes, refined (LOR)
meshes, simplex meshes, etc.) are now available as "named constructors", e.g.
Mesh::MakeCartesian2D or Mesh::MakeRefined. The legacy constructors are marked
as deprecated.
- Added support for creating periodic meshes with Mesh::MakePeriodic. The
requisite periodic vertex mappings can be created with
Mesh::CreatePeriodicVertexMapping.
- Added support for 1D non-conforming meshes (which can be useful for parallel
load balancing and derefinement).
- Added sample meshes in the `data` subdirectory showing the reference elements
of the six currently supported element types; ref-segment.mesh,
ref-triangle.mesh, ref-square.mesh, ref-tetrahedron.mesh, ref-cube.mesh, and
ref-prism.mesh.
High-performance computing
--------------------------
- Added initial support for GPU-accelerated versions of PETSc that works with
MFEM_USE_CUDA if PETSc has been configured with CUDA support. Examples 1 and 9
in the examples/petsc directory have been modified to work with --device cuda.
Examples with GAMG (ex1p) and SLEPc (ex11p) are also provided.
- Added support for explicit vectorization in the high-performance templated
code for Fujitsu's A64FX ARM microprocessor architecture.
- Added support for different modes of QuadratureInterpolator on GPU.
The layout (QVectorLayout::byNODES|byVDIM) and the tensor products modes can
be enabled before calling the Mult, Values, Derivatives, PhysDerivatives and
Determinants methods.
- Added method Device::SetMemoryTypes that can be used to change the default
host and device MemoryTypes before Device setup.
- In class MemoryManager, added methods GetDualMemoryType and SetDualMemoryType;
dual MemoryTypes are used to determine the second MemoryType (host or device)
when only one MemoryType is specified in methods of class Memory.
- Added Memory constructor for setting both the host and device MemoryTypes.
- Switched the default behavior of device memory allocations so that they are
deferred until the device pointer is needed.
- Added a second Umpire device MemoryType, DEVICE_UMPIRE_2, with corresponding
allocator that can be set with the method SetUmpireDevice2AllocatorName.
- Added HOST_PINNED MemoryType and a pinned host allocator for CUDA and HIP.
- Added matrix-free GPU-enabled implementations of GradientInterpolator and
IdentityInterpolator.
New and updated examples and miniapps
-------------------------------------
- Added a new, very simple example (ex0 and parallel version ex0p). This example
solves a simple Poisson problem using H1 elements (the same problem as ex1),
but is intended to be extremely simple and approachable for new users.
- Added new miniapps demonstrating: 1) the use of GSLIB for overlapping grids,
see gslib/schwarz_ex1, and 2) coupling different physics in different domains,
see navier/cht. Note that gslib v1.0.7 is require (see INSTALL for details).
- Added a new miniapp for computing (signed) distance functions to a point
source or zero level set. See miniapps/shifted/distance.cpp.
- Added a high-order extension of the shifted boundary method to solve PDEs on
non body-fitted meshes. This is illustrated in the new Shifted Diffusion
miniapp, see miniapps/shifted/diffusion.cpp.
- Added new miniapp directory mtop/ with optimization-oriented block parametric
non-linear form and abstract integrators. Two new miniapps, ParHeat and
SeqHeat, demonstrate parallel and sequential implementation of gradients
evaluation for linear diffusion with discrete density.
- Added a new miniapp block-solvers that compares the performance of various
solvers for mixed finite element discretization of the second order scalar
elliptic equations. Currently available solvers in the miniapp include a
block-diagonal preconditioner that is based on approximate Schur complement
(implemented in ex5p), and a newly implemented solver DivFreeSolver, which
exploits a multilevel decomposition of the Raviart-Thomas space and its
divergence-free subspace. See the miniapps/solvers directory for more details.
- Introduced new options for the mesh-explorer miniapp to visualize the actual
element attributes in parallel meshes while retaining the visualization of the
domain decomposition.
- Added partial assembly and device support to Example 25/25p, with diagonal
preconditioning.
- Implemented a filter method for the Navier miniapp to stabilize highly
turbulent flows in direct numerical simulation.
Improved testing
----------------
- Transitioned from Travis to GitHub Action for testing/CI on GitHub.
- Use Spack (and Uberenv) to automate TPL building in LLNL GitLab tests.
- Extended `make test` to include GPU tests when MFEM is built with CUDA or HIP
support.
- Added a set of suggested git hooks for developers in config/githooks.
- Added support for Caliper: a library to integrate performance profiling
capabilities into applications. See examples/caliper for more details.
- Added a new command line boolean option (`--all`) to the unit tests to launch
*all* non-regression tests.
- Upgraded the Catch unit test framework from version 2.13.0 to version 2.13.2.
Miscellaneous
-------------
- The following integrations have updated minimum version requirements:
* CUDA >= 10.1.168
* Ginkgo >= 1.4.0
* GSLIB >= 1.0.7
* HIOP >= 0.4
* HYPRE >= 2.20.0 for mixedint support
* HYPRE >= 2.22.0 for CUDA support
* libCEED >= 0.8
* PETSc >= 3.15.0 for CUDA support
* RAJA >= 0.13.0
see INSTALL for more details.
- Added a "scaled Jacobian" visualization option in the Mesh Explorer miniapp to
help identify elements with poor mesh quality.
- Added support for reading VTK meshes in XML format.
- Added makefile rule to generate TAGS table for vi or Emacs users.
- Added HIP support to the CMake build system.
- Various other simplifications, extensions, and bugfixes in the code.
API changes
-----------
- Added an abstract interface `mfem::FaceRestriction` for `H1FaceRestriction`
and `L2FaceRestriction`.
In order to conform with the semantic of `MultTranspose` in `mfem::Operator`,
`mfem::FaceRestriction::MultTranspose` now sets instead of adding values, and
`mfem::FaceRestriction::AddMultTranspose` should replace previous calls to
`mfem::FaceRestriction::MultTranspose`.
Version 4.2, released on October 30, 2020
=========================================
High-performance computing
--------------------------
- Added support for explicit vectorization in the high-performance templated
code, which can now take advantage of specific classes on the following
architectures:
* x86 (SSE/AVX/AVX2/AVX512),
* Power8 & Power9 (VSX),
* BG/Q (QPX).
These are disabled by default, but can be enabled with MFEM_USE_SIMD=YES.
See the new file linalg/simd.hpp and the new directory linalg/simd.
- Added an Element Assembly mode compatible with GPU device execution for H1 and
L2 spaces in the mass, convection, diffusion, transpose, and the face DG trace
integrators. See option '-ea' in Example 9. When enabled, this assembly level
stores independent dense matrices for the elements, and independent dense
matrices for the faces in the DG case.
- Added a Full Assembly mode compatible with GPU device execution. This assembly
level builds on top of the Element Assembly kernels to compute a global sparse
matrix. All integrators supported by element assembly are also supported by
full assembly. See the '-fa' option in Example 9.
- Optimized the AMD/HIP kernel support and enabled HIP support in the libCEED
integration. This is now available via the "ceed-hip" device backend.
- Improved the libCEED integration to support:
* AssemblyLevel::NONE for Mass, Diffusion, VectorMass, and VectorDiffusion
Integrators. This level computes the full operator evaluation "on the fly".
* VectorMassIntegrator and VectorDiffusionIntegrator.
* All types of (scalar) Coefficients.
- Added partial assembly / device support for:
* H(div) bilinear forms and VectorFEDivergenceIntegrator.
* BlockOperator, see the updated Example 5.
* Complex operators, see the updated Example 22.
* Chebyshev accelerated polynomial smoother.
* Convergent diagonal preconditioning on non-conforming adaptively refined
meshes, see Example 6/6p.
- Added CUDA support for:
* Sparse matrix-vector multiplication with cuSPARSE,