Changes: 3.3#

config/configure.py:

  • Building PETSc using CMake is enabled automatically if CMake is installed (enabling parallel builds and fast incremental builds), otherwise the plain make-based system is used.

  • Added –with-cuda-only flag to allow compiling CUDA files without requiring Thrust and Cusp

Vec:

  • VecCreateSeqWithArray() and VecCreateMPIWithArray() now take a blocksize argument before the local length

  • VecSetBlockSize() cannot be called after VecCreateSeq() or VecCreateMPI() and must be called before VecSetUp() or VecSetFromOptions() or before either VecSetType() or VecSetSizes()

Mat:

  • MatScaleSystem() and MatUnScaleSystem() are gone, they didn’t do anything.

  • Renamed MatNullSpaceAttach() to MatSetNullSpace()

  • MatMult() no longer removes the nullspace set by MatSetNullSpace()

  • Renamed MatMatMultTranspose() for C=A^T*B to MatTransposeMatMult()

  • Added MatMatTransposeMult() for C=A*B^T

  • Added MatRARt() for C=R*A*R^T

  • Preallocation routines now automatically set MAT_NEW_NONZERO_ALLOCATION_ERR, if you intentionally preallocate less than necessary then use MatSetOption(mat,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE) to disable the error generation.

  • MatSetBlockSize() must be called before MatSetUp() or MatXXXXSetPreallocation() or block size defaults to 1

  • -mat_block_size is NOT processed by MatXXXXSetPreallocation() you MUST call MatSetFromOptions() before MatSetUp() to have it processed

  • You MUST now call MatXXXSetPreallocation() or MatSetUp() on any matrix you create directly (not using DMCreateMatrix()) before calling MatSetValues(), MatSetValuesBlocked() etc.

  • MatGetOwnershipRange() will not work without setting up the matrix.

  • MatSetUpPreallocation() and MatPreallocated() are removed, use MatSetUp()

  • MatPreallocateSymmetricInitialize() is removed, use MatPreallocateInitialize()

  • MatCreateMPIAIJ(), MatCreateMPIBAIJ(), MatCreateMPISBAIJ(), MatCreateMPIDense() are now MatCreateAIJ(), MatCreateBAIJ(), MatCreateSBAIJ(), MatCreateDense(). Note that on a single process they have always created the sequential version of the matrix so this naming is more accurate.

  • Renamed MatMerge_SeqsToMPI() to MatCreateMPIAIJSumSeqAIJ()

  • Renamed MatMerge() to MatCreateMPIAIJConcatenateSeqAIJ()

  • Added block size in call to MatCreateIS()

  • MatInvertBlockDiagonal() now returns a const pointer

PC:

  • Remove PCASASetDM(), use PCSetDM() instead.

  • Add PCFieldSplitSetSchurFactType(), rename -pc_fieldsplit_schur_factorization_type to -pc_fieldsplit_schur_fact_type.

  • Added native algebraic multigrid method -pc_type gamg. Aggregation method is recommended: -pc_gamg_type agg and for elliptic operators smoothing is recommended: -pc_gamg_agg_nsmooths 1.

  • PCSetCoordinates(PC,PetscInt,PetscInt,PetscReal*) added second integer argument for number of local vertices.

  • Added PCISSetSubdomainScalingFactor(PC,PetscScalar).

  • PCGASM:

    • Remove PCGASMSetLocalSubdomains(), use PCGASMSetSubdomains(); note: the first argument is now the list of “inner” subdomains (without overlap), unlike before and for PCASM

    • PCGASMSetTotalSubdomains(PC,PetscInt,PetscBool); added third argument to indicate whether local subdomains should be constructed.

    • Removed PCGASMCreateSubdomains(), use PCGASMCreateLocalSubdomains(Mat,PetscInt,PetscInt,IS*[],IS*[]); second PetscInt argument is requested overlap

    • PCGASMCreateSubdomains2D(PC,PetscInt,PetscInt,PetscInt,PetscInt,PetscInt,PetscInt,PetscInt*,IS[]*,IS[]*): the 9-th and 10-th argument have been swapped: the 9-th outputs the inner subdomains, the 10-th outputs the outer subdomains.

  • PCMG multigrid defaults changed to use a Chebyshev smoother with block Jacobi Gauss-Seidel relaxation. This is a linear preconditioner so it can be used with normal Krylov methods rather than only flexible methods, as well as having fewer synchronization pointst. Use -mg_levels_ksp_type gmres -mg_levels_ksp_max_it 1 -mg_levels_pc_type bjacobi for the original behavior.

KSP:

  • KSPSetOperators() will automatically set the null space present on the input matrix

  • Renamed KSPCHEBYCHEV to KSPCHEBYSHEV

  • Added Hybrid CHEBYCHEV (-ksp_chebyshev_hybrid)

  • Added pipelined GMRES (KSPPGMRES) which performs one non-blocking reduction per iteration instead of two blocking reductions.

  • Added flexible BiCGStab (KSPFBCGS) which tolerates a nonlinear preconditioner (like KSPFGMRES).

  • Added improved flexible BiCGStab (KSPIFBCGS) which tolerates a nonlinear preconditioner and performs one reduction every other iteration (like KSPIBCGS).

SNES:

  • Added SNESGetSNESLineSearch

  • Changed default max_its and max_funcs for non-newton SNES solvers to 10000 and 30000 respectively

  • Changed options and command-line arguments for SNESFAS to be in line with PCMG

  • Added quasi-Newton (SNESQN), Nonlinear GMRES (SNESNGMRES), nonlinear conjugate gradients (SNESNCG), and multi-stage methods (SNESMS, used as a smoother for nonlinear multigrid).

  • Support for Full Approximation Scheme nonlinear multigrid (SNESFAS) moved from DMMG to SNES and made composable with the methods above.

  • SNES line search type should be set by name with SNESLineSearchSetType rather than with SNESLineSearchSet, see below.

SNESLineSearch:

  • SNESLineSearch object added

  • The default SNESLineSearch of a SNES may be gotten with SNESGetSNESLineSearch(SNES snes, SNESLineSearch *linesearch)

  • The linesearch type may be set with SNESLineSearchSetType() or -snes_linesearch_type

  • The line search (and pre and post checks) is applied using SNESLineSearchApply()

  • Pre/Post-check methods may be set with SNESLineSearchSetPre/PostCheck()

  • The previous steplength may be accessed through SNESLineSearchSetLambda()

  • The damping parameter may be set through SNESLineSearchSetDamping() or -snes_linesearch_damping

  • Success of the line search is determined using SNESLineSearchGetSuccess()

  • Custom linesearches may be built through SNESLINESEARCHSHELL,(“shell”) or by registering a new linesearch type with SNESLineSearchRegisterDynamic()

  • SNESLINESEARCHBT,(“bt”) replaces SNES_LS_CUBIC and SNES_LS_QUADRATIC, and order may be set with SNESLineSearchSetOrder() or -snes_linesearch_order

  • SNESLINESEARCHBASIC,(“basic”) replaces SNES_LS_BASIC and SNES_LS_BASICNONORMS. Norms may be turned off with SNESLineSearchSetComputeNorms() or -snes_linesearch_norms 0

  • SNESLineSearchSetTolerances() replaces SNESLSSetParams(), with the former alpha parameter set with SNESLineSearchBTSetAlpha() or -snes_linesearch_alpha

  • Added Line Search type SNESLINESEARCHL2,(“l2”) as the default for NRICHARDSON

  • SNESLINESEARCHCP,(“cp”) added as the default line search method for SNESNCG and SNESQN

TS:

  • -ts_max_time changed to -ts_final_time

  • TSDefaultComputeJacobian() and TSDefaultComputeJacobianColor() have been removed. Configure TS to use coloring with SNESSetJacobian().

  • Added TSROSW for Rosenbrock-W methods.

  • Added a common, extensible system for adaptive controllers, see TSGetAdapt().

DM/DA:

  • Added DMCreateDecomposition(DM,PetscInt,char*[]*,IS[]*,DM[]*) for use with PCFIELDSPLIT,PCASM,PCGASM

  • Added DMCreateDecompositionDM(DM,const char*,DM*) to create a version of the DM encapsulating a named decomposition; use with DMCreateDecomposition()

  • Added DMRedundant for managing globally coupled degrees of freedom.

  • Removed DMCompositeAddArray(), use DMRedundantCreate() and DMCompositeAddDM().

  • Renamed DMGetMatrix(), DMGetInterpolation(), DMGetInjection(), and DMGetColoring() to DMCreateMatrix(), etc for semantic consistency.

  • The communicator argument to DMRefine() and DMCoarsen() can be MPI_COMM_NULL, but not PETSC_NULL, because the latter may not be the correct type.

  • Added DMCoarsenHookAdd() and DMRefineHookAdd() for shepherding persistent resolution-dependent data between levels.

  • Added DMGetNamedGlobalVector() for storing persistent resolution-dependent data.

  • DMDASNESSetFunctionLocal() and DMDASNESSetJacobianLocal() can be used for convenient local evaluation; these routines will eventually replace DMDASetLocalFunction() and DMDASetLocalJacobian().

DMMG:

  • DMMG is now completely removed from PETSc. Equivalent (and better) functionality can now be obtained by calling SNESSetDM() or KSPSetDM(). Make sure to avoid resolution-dependent data in the user context. Use SNESGetDM() or KSPGetDM() in the function evaluation context to obtain the grid. DMGetNamedGlobalVector(), DMCoarsenHookAdd(), and DMRefineHookAdd() can be used to manage persistent resolution-dependent data.

PetscViewer:

  • A VTK binary viewer was added, see PETSCVIEWERVTK.

SYS:

  • PetscBagLoad() now requires you previously created and registered all the records in the PETSc bag, allows loading on systems with different struct layout/endianness.

  • PetscSF added as a type-generic graph communication mechanism. The current implementation requires MPI-2 one-sided and the interface is currently optional.

Fortran:

  • PETSC_NULL_TRUTH is now PETSC_NULL_BOOL

  • PetscOptionsGetEnum() now available from Fortran

ExternalPackages:

  • The Hypre interface was updated to 2.8.0b and now supports 64-bit integers.

  • SuperLU_DIST interface updated to 3.1.

  • SuperLU interface updated to 4.3.

  • Sundials interface updated to 2.5.0.

  • FFTW interface updated to 3.3.2.

  • ParMetis updated to 4.0.2 and split from Metis 5.0.2 which is now separate.

Build:

  • If Python 2.7 is available, the config/builder2.py is now available for a complete build, including dependency tracking.

Examples:

  • SNES ex62 illustrates the use of DMComplex to solve the Stokes equation on an unstructured mesh. The Python tests show how to use block preconditioning strategies from the command line.

  • SNES ex52 illustrates the use of CUDA for FEM integration