Changes: 3.9


  • The scripts in $PETSC_DIR/bin are now in $PETSC_DIR/lib/petsc/bin

  • PetscStrncat() renamed to PetscStrlcat() now takes the length of the original allocated space in the string instead of the part left after strings have been copied or concatenated in

  • CUDA and ViennaCL as GPU backends can now also be used with the release version (no need to use the main branch).


  • Option –with-cuda-arch is removed. Use CUDAFLAGS to directly specify relevent nvcc option [for ex: CUDAFLAGS=-arch=sm_20]

  • Added –with-avx512-kernels to specify that hand-coded kernels using AVX-512 intrinsics should be used when available. Currently defaults to off.





  • Added VECNODE type. Vector uses on-node shared memory to store its entries.


  • VecScatter becomes a PETSc object. Existing vecscatters have types VECSCATTERSEQ, VECSCATTERMPI1.

  • Added VECSCATTERMPI3. It stores vector ghost values to the on-node shared memory.

  • Added VECSCATTERMPI3NODE. It works on vectors of type VECNODE.



  • Added MatShellTestMultTranspose() and MatShellTestMult() for checking if user provided MATSHELL matches the Jacobian of a given function computed with differencing.

  • MatSolverPackage is replaced with MatSolverType.

  • mat_solver_package is replaced with mat_solver_type in options, e.g. -pc_factor_mat_solver_type should be used instead of -pc_factor_mat_solver_package.

  • MatShellSetOperation() and MatShellGetOperation() can now only be used for MATSHELL matrices, use MatSetOperation() and MatGetOperation() to set an operation for any matrix type.

  • Added MatMumpsGetInverse().

  • MatMult() for the MATAIJ type has been optimized using AVX-512 intrinsics; must be enabled at configure time.

  • Added a new Mat type MATSELL, featuring a fast MatMult() kernel on AVX-512 architecture such as KNL and Skylake.

  • Added support for additional MKL sparse BLAS operations in MATAIJMKL: MatMatMult(), MatTransposeMatMult(), MatPtAP() (symmetric A only). MKL version 18, update 2 or later is required for MatPtAP()/MatPtAPNumeric() and MatMatMultNumeric().

  • Added MatGetInertia() for SuperLU_DIST interface.


  • Added -pc_mg_distinct_smoothup and PCMGSetDistinctSmoothUp() so that one can control the down and up smoothers options separately from the options database

  • Removed -pc_mg_smoothup and -pc_mg_smoothdown and PCMGSetNumberSmoothUp() and PCMGSetNumberSmoothDown(). Instead, configure the solvers on the levels with -mg_levels_ksp_max_it n. If separate numbers of smoothing steps are required for up and down smoothers, use -pc_mg_distinct_smoothup -mg_levels_ksp_max_it n -mg_levels_up_ksp_max_it m.

  • Added PCCHOLESKY to SuperLU_DIST interface.



  • Removed SNESHasDM() because there is always a DM in SNES.

  • The SNESType of SNESTEST has been removed, the testing code can now be accessed with the options -snes_test_jacobian, -snes_test_jacobian_display -snes_test_jacobian_display_threshold and is called each time SNES computes a new Jacobian.



  • Added TSRHSJacobianTestTranspose() and TSRHSJacobianTest() checking if user provide MATSHELL Jacobian with TSSetRHSJacobian() matches the Jacobian of the function provided to TSSetRHSFunction() computed with finite differencing. Command line options -ts_rhs_jacobian_test_mult_transpose -mat_shell_test_mult_transpose_view and -ts_rhs_jacobian_test_mult -mat_shell_test_mult_view.

  • Added -ts_trajectory_dirname and -ts_trajectory_filetemplate to allow users to specify the folder name and file name template for disk checkpoints.


  • Changed TaoDefaultMonitor() to TaoMonitorDefault() to match other PETSc default monitor names, also now print with the same format as other monitors %3D Tao ….

  • Added VecLock{Push|Pop} calls around user callbacks; use of VecGetArray in user callbacks is now prohibited.

  • Added default matrix-free finite-differencing implementation for Hessian MatMult with TaoDefaultComputeHessianMFFD(). Can be selected with -tao_mf_hessian at command line.

  • Added Projected Gradient Descent (PGD) and Bounded Nonlinear Conjugate Gradient (BNCG) algorithms for bound constrained problems.

  • Improved support for recycling BFGS correction vectors between subsequent TaoSolve() calls for the unconstrained LMVM algorithm. Can be enabled using -tao_lmm_recycle at command line.

  • Convergence tests for all algorithms are separated out of TaoMonitor() calls.



  • Deprecate DMGetDefaultSection() in favor of DMGetSection()

  • Deprecate DMSetDefaultSection() in favor of DMSetSection()

  • Deprecate DMGetDefaultGlobalSection() in favor of DMGetGlobalSection()

  • Deprecate DMSetDefaultGlobalSection() in favor of DMSetGlobalSection()


  • PetscViewerVTKFWrite() now takes a MPI_Datatype instead of a PetscDataType argument.


  • DMNetworkGetComponentDataArray and DMNetworkGetComponentTypeOffset are no longer available. Use DMNetworkGetComponent instead.

  • Added support for sub-dmnetworks.


  • PetscPClose() no longer returns error code from pclose() because some systems cannot reliably deliver it

  • PetscPopUpSelect() has been removed.




  • The F90Array routines now take a MPI_Datatype argument instead of a PetscDataType. This should not affect user code.