Changes: 3.12#


  • PetscRoundReal() has been removed, since the math routine round() is not supported on all systems: perhaps use PetscCeilReal() or PetscFloorReal()<li>

  • The legacy and CMake compile systems (make all-legacy and all-cmake) are removed

  • make gnumake is now make libs

  • PetscArraycmp() should be used instead of PetscMemcmp(), it takes a count argument instead of a byte argument

  • PetscArraycpy() should be used instead of PetscMemcpy(), it takes a count argument instead of a byte argument

  • PetscArrayzero() should be used instead of PetscMemzero(), it takes a count argument instead of a byte argument

  • PetscArraymove() should be used instead of PetscMemmove(), it takes a count argument instead of a byte argument

  • The ./configure options -with-gnu-compilers and -with-vendor-compilers are removed

  • The ./configure option –with-avx512-kernels defaults to true now so the manually optimized AVX-512 kernels are used by default

  • The order ./configure tests compilers has changed so there may be a different compiler selected if you do not provide them

  • Added –with-64-bit-blas-indices that will switch to 64-bit indices when using MKL libraries for BLAS/LAPACK and build OpenBLAS with this support


  • -help now causes the printing of the current value and the newly set value for PetscOptionsInt(), etc

  • PetscFunctionListPrintTypes() takes an additional final argument after the current value, the set value

  • PetscCalloc*() now calls the system calloc() routine instead of malloc() plus memzero()

  • -malloc is now marked as deprecated, use -malloc_debug instead

  • -malloc_log and -malloc_log_threshold are now -malloc_view and -malloc_view_threshold

  • PetscMallocDebug() is now PetscMallocSetDebug()

  • PetscMallocDumpLog() is now PetscMallocView(), see also PetscMallocViewSet()

  • PetscMallocSetDumpLogThreshold() and PetscMallocSetDumpLog() are now PetscMallocViewSet()

  • With -malloc_test or -malloc_debug allocated memory is initialized with NaN to detect use of uninitialized numerical arrays

  • PetscMallocSet() now takes three arguments instead of two

  • Moved PetscCUBLASGetHandle() to a separate header file petscublas.h

  • Moved VecCUDA*-routines from petsccuda.h into petscvec.h. Removed petsccuda.h


  • C++ dialect is now auto-detected (C++14 first and then C++11). One can disable this check using –with-cxx-dialect=0, or force it to only check for C++11 using –with-cxx-dialect=C++11

  • New option –download-hpddm to enable PCHPDDM and KSPHPDDM





  • VecCUDAGet/RestoreArrayReadWrite() changed to VecCUDAGet/RestoreArray()

  • VecViennaCLGet/RestoreArrayReadWrite() changed to VecViennaCLGet/RestoreArray()

  • Added VecGet/RestoreArray/ReadInPlace() to get array of a vector where it is without copying from GPU to CPU.


  • Added PetscLayoutCreateFromSizes()

  • Added PetscLayoutCreateFromRanges()

  • Calling PetscLayoutSetUp() twice with different sizes is now forbidden


VecScatter & PetscSF:

  • The default VecScatter implementation is changed to PetscSF. In other words, the default VecScatter shares PetscSF’s MPI communication code. One can use -vecscatter_type mpi1 to go back to the old implementation.

  • Added new InsertMode MAX_VALUES and MIN_VALUES support in VecScatterBegin/End and VecGhostUpdateBegin/End.

  • PetscSFComputeMultiRootOriginalNumbering: add output argument with the number of multiroots for convenience

  • Added an enum type PetscSFPattern for SF graph patterns. Valid values include PETSCSF_PATTERN_ALLGATHER, PETSCSF_PATTERN_GATHER and PETSCSF_PATTERN_ALLTOALL.

  • Added PetscSFSetGraphWithPattern() to set SF graphs with predefined patterns.

  • Added PetscSFComposeInverse() to compose a new SF by putting the inverse of an SF under the another SF.

  • Added MPI-3.0 neighborhood collectives support. One can use command line option -sf_type neighbor to let SF use MPI-3.0 neighborhood collectives for communication instead of the default MPI_Send/Recv.

  • PetscSF is now CUDA-aware. The rootdata, leafdata arguments passed to SF routines can be either GPU pointers or CPU pointers. Use a CUDA-aware MPI and option -use_gpu_aware_mpi to enable it.

  • VecScatter is also CUDA-aware. You can do VecScatter on CUDA vectors without copying them from GPU to CPU. Again, you need a CUDA-aware MPI and option -use_gpu_aware_mpi.

  • Removed PetscSFCreateFromZero. Instead, users should use PetscSFCreate() to create an SF, and then set its graph with PetscSFSetGraphWithPattern(..,PETSCSF_PATTERN_GATHER).

  • Renamed PetscSFGetRanks() to PetscSFGetRootRanks().


  • MatShift(Mat,0); will no longer silently insure there are no missing diagonal entries. (Previously it would put 0 into any diagonal entry that was missing)

  • Renamed MatComputeExplicitOperator() into MatComputeOperator() and MatComputeExplicitOperatorTranspose() into MatComputeOperatorTranspose(). Added extra argument to select the desired matrix type

  • MatLoad() now supports loading dense matrices from HDF5/MAT files.

  • Added new Mat type, MATKAIJ, for matrices that can be expressed in a particular Kronecker (tensor) product form.

  • Added MatCompositeSetScalings() to set separate scaling factors for component matrices.

  • Added a Boolean option -mat_composite_merge_mvctx to merge Mvctx of component matrices to optimize communication in MatMult() for ADDITIVE matrices. Default is true.

  • Added MATSEQDENSECUDA class to use GPUs for dense linear algebra.


  • Renamed PCComputeExplicitOperator() into PCComputeOperator(). Added extra argument to select the desired matrix type

  • Added support for PCCHOLESKY with MATSOLVERMUMPS and SBAIJ matrices with bs greater than 1

  • Added support for PCLU with MATSOLVERMKL_PARDISO (resp. MATSOLVERMKL_CPARDISO) and SEQBAIJ (resp. MPIBAIJ) matrices


  • Added new preconditioner PCHPDDM, cf. src/ksp/ksp/examples/tutorials/ex71.c or src/snes/examples/tutorials/ex56.c


  • Renamed KSPComputeExplicitOperator() into KSPComputeOperator(). Added extra argument to select the desired matrix type

  • KSPCGGLTRGetMinEig(KSP,PetscReal*) -> KSPGLTRGetMinEig(KSP,PetscReal*)

  • KSPCGGLTRGetLambda(KSP,PetscReal*) -> KSPGLTRGetLambda(KSP,PetscReal*)


  • Added new Krylov solver KSPHPDDM, cf. src/ksp/ksp/examples/tutorials/ex75.c or src/ksp/ksp/examples/tutorials/ex9.c


  • Changed SNES_CONVERGED_TR_DELTA to SNES_DIVERGED_TR_DELTA with a negative value to indicate diverged, unclear why it was ever positive since tiny trust region would indicate trouble, not converged


  • -snes_mf no longer turns the preconditioner to none; to use it generally you will also need -pc_type none



  • Added tangent linear models for Runge-Kutta methods

  • Added second-order adjoint solvers based on Runge-Kutta methods and Theta methods

  • Improved the usage of first-order adjoint solvers in an optimization context. (The TS object can be reused in the optimization loop)

  • Changed the APIs for integrand evaluations and corresponding derivative evaluations. TSSetCostIntegrand() is deprecated. (Instead a quadrature TS is used to handle the callbacks)

  • TSTrajectory creates a unique name for folders that store the checkpoint data. Users can rerun adjoint codes without worrying about the folder names, and run multiple codes under the same directory.

  • Added support for using finite-differencing (and coloring) to approximate the Jacobians needed by the adjoint solvers. -snes_fd_color can be used for both implicit and explicit methods.


  • Add DMGetLocalSection() to be used in preference to DMGetSection().

  • Add DMSetLocalSection() to be used in preference to DMSetSection().

  • DMGetDefaultSF() replaced with DMGetSectionSF().

  • DMSetDefaultSF() replaced with DMSetSectionSF().

  • DMCreateDefaultSF() replaced with DMCreateSectionSF().

  • DMAddLabel() now increments the reference count of the DMLabel.

  • DMRemoveLabel() now destroys the DMLabel if the output argument is NULL.

  • Add DMRemoveLabelBySelf() which looks up the DMLabel by itself, removes from DM and destroys.


  • Rename DMPlexCreateSpectralClosurePermutation() to DMPlexSetClosurePermutationTensor()

  • Add DMPlexFindVertices() for vertex coordinates -> DAG point lookup

  • Add DMPlexGetOverlap() to get the partition overlap


  • Changed prototypes for DMNetworkSetSizes()


  • Removed PetscViewerHDF5{Get,Set}AIJNames() which are likely not really needed

  • PetscViewerHDF5SetCollective() and -viewer_hdf5_collective can be used to switch between independent and collective transfer mode. Defaults now to false (independent).


  • Added PetscCheckDupsInt() to check if a PetscInt array has dups.

  • Added an argument to PetscBinaryRead() and PetscBinarySynchronizedRead() to return the number of items read.