trilinos/Trilinos trilinos-release-17-2-0

Trilinos 17.2.0

on GitHub

The test results from Trilinos testing (PRs and nightly builds) are now posted
to a public CDash dashboard:
https://my.cdash.org/index.php?project=Trilinos

Amesos2:

• new "partial-factorization" option (to improve performance of computing
  local Schur complement) with PardisoMKL & MUMPS
  #15404
  #15484

Belos:

• Solvers now provide more more detailed solve status reporting.
  The "Unconverged" return type has been split into "MaxItersReached",
  "MaxRestartsReached". "OrthonormFailure", "NaNDetected", "BreakdownDetected",
  "LossOfAccuracyDetected", "NonspecificException", "InconsistentState",
  "Undetermined".
  #15231

Galeri:

• Added the creation of finite element mass matrices for use in testing.
  #15318

Kokkos, KokkosKernels:

• Inclusion of version 5.2.1 of Kokkos and Kokkos Kernels
  #15453
  #15593
  https://github.com/kokkos/kokkos/blob/develop/CHANGELOG.md
  https://github.com/kokkos/kokkos-kernels/blob/develop/CHANGELOG.md

Ifpack2:

• Cache local CRS structure in par_ILUT to improve setup related costs
  #15439

• Use KSPTRSV as default triangular solver on GPU platforms
  #15490

Intrepid2:

• Redesigned PointInclusion tools to expose parametric distance and
  extended CellTools to support beam and shell elements, including mappings
  from (D+1)-dimensional physical space to D-dimensional reference space.
  #15269

• Refactored Intrepid2 operators to improve performance by introducing an
  orientation operator that stores subcell orientation matrices in
  compressed-row storage format.
  #14450

• Improved the parallel implementation of Basis::getValues.
  #15234

Krino:

• Snapshot
  #15526

Minitensor:

• New capability: quaternions and conversions among rotation
  representations (rotation matrix, unit quaternion, principal rotation
  vector), using a singularity-free Spurrier extraction, along with a
  helper that keeps a rotation vector continuous across the 2 pi
  boundary and the axial vector operator vee(). Also new are exp_sym()
  and exp_eig_sym(), the eigendecomposition exponential of a symmetric
  tensor, mirroring the existing log_sym() and log_eig_sym().
  #15532
  #15514

• The general log() and exp() now follow the current published
  algorithms. For dimensions three and lower, log() uses Gauss-Legendre
  partial-fraction Pade evaluation inside the Pade trust region and
  Schur-based inverse scaling and squaring (Al-Mohy and Higham, SIAM J.
  Sci. Comput. 34(4), 2012) outside it, and exp() uses the parameter
  selection of Al-Mohy and Higham, SIAM J. Matrix Anal. Appl. 31(3),
  2009. Measured against an 80-bit eigendecomposition reference, log()
  is 1.8 to 4.3 times more accurate and 1.4 to 4.7 times faster away
  from the identity, with no failures in 17,500 samples.
  #15515

• Correctness and robustness fixes. log() no longer aborts in debug
  builds or writes out of bounds in release builds on ill-conditioned
  symmetric positive definite tensors, where a catastrophically
  cancelling determinant produced a non-finite scaling factor: 39
  percent of random SPD tensors at condition 1e14 used to fail, none do
  now. eig_sym() returns a consistent (V, D) pair in two dimensions,
  the 2x2 singular value decomposition canonicalizes its singular
  values, operations that require symmetric positive definite input now
  fail loudly instead of returning garbage, and the 2x2 eigenvalue and
  square root scaling is safe for Sacado FAD scalars, which repaired
  downstream code that instantiates log() on a tensor of FAD scalars.
  #15513
  #15393
  #15397
  #15398
  #15520

• The headers were reorganized, with no change in functionality. The
  .i.h and .t.h sub-headers were merged into their parents. The
  TeuchosCore and KokkosKernels dependencies were dropped. Code that
  includes MiniTensor.h is unaffected.
  #15523
  #15527
  #15529

• The package is now documented.
  https://trilinos.github.io/docs/minitensor/index.html
  #15531
  #15533

• The tests were split by module, property-based tests were added for
  the matrix decompositions, and every public header is now compiled on
  its own so that it stays self-contained.
  #15395
  #15534

MueLu:

• Improve support for Teko-based smoothers via specifying smoother: type = teko
  #15238

• Add a new stength-of-connection matrix that leverages an approximate inverse
  of a mass matrix.
  #15239

Panzer:

• Large update to the documenation
  #15503

STK:

• Inclusion of version 5.31.3
  #15427
  #15530
  

  Trilinos/packages/stk/CHANGELOG.md

  Lines 3 to 41 in 8aec0c9

  	5.31.3 (STK_VERSION 5310300)
  	stk_util: Deprecate 'double_complex_sum_op()' in MPI.hpp and rename
  	internal supporting function in MPI.cc (to avoid duplicate
  	symbol error with empire/xyce linking).
  	cmake: Allow Intrepid2 to be an optional dependency of stk_search_util
  	(Fixes an issue where stk_search_util failed to compile if
  	Intrepid2 was not enabled.)
  	5.31.2 (STK_VERSION 5310200)
  	stk_mesh: improve behavior of for_each_entity_run when OpenMP is enabled
  	stk_transfer: enable transfers between 2D and 3D mesh, as well as
  	axisymmetric transfers
  	stk_mesh: make field_blas functions work on Rocm when unified memory is
  	enabled
  	stk_transfer: enable stk_transfer executable to use Intrepid2 master elements
  	stk_mesh: move DeviceMesh connectivity data out of buckets and into
  	separate data structure
  	stk_mesh: change DeviceMesh to sort the entities in the same order
  	as BulkData
  	stk_transfer: add optional coordinate transform to receive mesh
  	wrapper to enable transfers between similar but
  	not identical meshes
  	stk_mesh: make field state rotation work on device for fields
  	where some states do not have device data
  	stk_transfer: enable transfers to repeat the search when directed
  	by the user mesh wrappers. Makes stk_transfer
  	usable for adaptive mesh refinement.
  	stk_mesh: reduce size of FieldData objects and improve the
  	performance of creating them
  	5.27.4 (STK_VERSION 5270400) 1/6/26
  	stk_mesh: parallel_sum/min/max and '_including_ghosts' versions
  	now use single consolidated code-path for both CPU and
  	GPU builds. Also fixes a Kokkos 5 build error related
  	to '::HostMirror' vs '::host_mirror_type'.
  	stk_mesh: Reduced size of device FieldData objects.
  	stk_mesh: Fixed the 'STK_USE_DEVICE_MESH' build (which is used for
  	debugging device code-paths in a non-GPU build).

Tempus:

• The tutorial pages and documentation have been updated and extended.
  #15365
  #15387
  #15423

Teko:

• Port ProbingPreconditionerFactory to Tpetra
  #15386

• Port majority of tests/examples to Tpetra

Zoltan2:

• Provide option to opt-out of potentially expensive symmetrization in metis algorithm
  #15496