Changes: 2.0.17#


  • Added support for Windows NT/95 using the Microsoft Developers Studio Visual C++. See the file ‘Installation’ for details.

  • Other new machines on which this release has been tested: Cray T3E, SGI Origin See the file petsc/Installation for a complete list.

TS (Timestepping Solvers):

  • Modified the pseudo-transient continuation updates:

  • Changed the default update to dt = current_dt*previous_fnorm/current_fnorm.

  • Added the routine TSPseudoIncrementDtFromInitialDt() and the corresponding option -ts_pseudo_increment_dt_from_initial_dt to indicate use of the alternative update formula dt = initial_dt*initial_fnorm/current_fnorm.

  • Changed the calling sequence of TSRegister().

SNES (Nonlinear Solvers):

  • Added support for computing large, sparse Jacobians efficiently via finite differences, see Section 5.6 (Finite Difference Jacobian Approximations) of the users manual.

  • Added the routines SNESGetNumberLinearIterations() and SNESSetConvergenceHistory(). See man pages for details.

  • Activated a counter of function evaluations, which is used in convergence tests to terminate solver if the number of function evaluations exceeds a given tolerance. Note: Users of matrix-free Newton-Krylov methods may need to reset the default allowable maximum (1000), via SNESSetTolerances() or -snes_max_func <maxfunc>.

  • Changed the calling sequence of SNESRegister().

SLES (Linear Solvers):

  • See PC and KSP

KSP (Krylov Subspace Methods):

  • Changed the calling sequence of KSPRegister().

PC (Preconditioners):

  • Changed the calling sequence of PCRegister().

  • New Additive Schwarz variants (preconditioner type PCASM)

  • Added the routine PCASMSetType() (and the corresponding option -pc_asm_type [basic,restrict,interpolate,none]) for setting the variant of the additive Schwarz method. See the man page and users manual for details.

  • Changed the default variant of PCASM from full restriction and interpolation to full restriction only, since this version requires less communication and for many problems converges faster than the basic variant that uses full restriction and interpolation. Users can still employ the basic ASM by calling PCASMSetType(pc,PC_ASM_BASIC) or by using the option -pc_asm_type basic.

  • Added an interface to the SPAI preconditioner implementation of Steven Bernard; see src/contrib/spai. This has undergone little testing and optimization; it is intended mainly for “hackers”.

MAT (Matrices):

  • Added the matrix option, MatSetOption(mat,MAT_NEW_NONZERO_LOCATION_ERROR), that will cause an error if a new nonzero is generated in a sparse matrix. (currently implemented for AIJ and BAIJ matrices only). This is a useful flag when using SAME_NONZERO_PATTERN in calling SLESSetOperators() to ensure that the nonzero pattern truly does remain unchanged. For examples, see the programs petsc/src/snes/examples/tutorials/[ex5.c,ex5f.F].

  • Added the routine MatSetUnfactored(), intended primarily for use with in-place ILU(0) factorization as a preconditioner for matrix-free Krylov methods. See the manual page for details.

  • Added the routines MatConvertRegisterAll() and MatLoadRegisterAll() to allow the restriction of the matrix routines linked into an application code. This can decrease the size of your executable and the time it takes to link your program. For details, see the manual page and petsc/src/snes/examples/tutorials/ex5.c

  • Added the routine MatSetValuesBlocked(), for more efficient assembly of block AIJ formatted matrices (MATSEQBAIJ and MATMPIBAIJ).

  • Changed the calling sequence of MatReorderingRegister();

DA (Distributed Arrays):

  • Added additional arguments to DACreate1d(), DACreate2d(), and DACreate3d() to allow the user to set the distribution of nodes on each processor; set these arguments to PETSC_NULL for the standard default distribution.

  • Modified DAGetInfo() to return the type of periodicity.

VEC (Vectors):

  • Added the routine VecCreateGhost() to create vectors that have ghost padding at the end of the local array. This is useful for gathering remote values to perform local calculations that involve off-processor ghost values. This is often appropriate for codes using unstructured grids. See petsc/src/vec/examples/tutorials/ex9.c for possible usage.

IS (Index Sets):

Draw (Graphics):

  • Application codes should not need to use #include “draw.h” anymore from C/C++, since this file is now included automatically when “petsc.h” or any other PETSc include file is included.



  • Added VIEWER_DRAWX_(MPI_Comm comm) from C. Useful for rapid code prototyping without having to declare a Viewer.

System Routines:

  • Since memory leaks and uninitialized memory can be serious problems for large-scale application codes, we’ve added several new tools to assist in their diagnosis. These tools are all work in conjunction with the PETSc memory allocation (the default for codes that are compiled in debug mode with BOPT=[g,g_c++,g_complex]).

  • Added the runtime option -trmalloc_log, which activates logging of all calls to malloc via the new routines PetscTrLog() and PetscTrLogDump().

  • Added the routine PetscGetResidentSetSize() to determine the total memory used by a process (this is activated by -trmalloc_log); see the man page for details.

  • Added the option -trmalloc_nan for tracking down allocated memory that is used before it has been initialized. This option calls the new routines PetscInitializeNans() and PetscInitializeLargeInts(). So far these work on the Sun4 system.

Error Handling:

  • The error checking macros SETERRQ() and SETERRA() now have the calling sequence SETERRQ(int ierr,int pierr,char *message); where pierr is an additional integer error code passed to the error handler. Currently you should just set pierr=1.

  • Also, SETERRQ() and SETERRA() now use the macro __FUNC__ to keep track of routine names. Users need not worry about this in their application codes, but can take advantage of this feature if desired by setting this macro before each user-defined routine that may call SETERRQ(), SETERRA(), CHKERRQ(), or CHKERRA(). __FUNC__ should be set to a string containing the routine name. For example, #undef __FUNC__ #define __FUNC__ “MyRoutine1” int MyRoutine1() { /* code here */ return 0; } See petsc/src/snes/examples/tutorials/ex3.c for an example.

  • PETSc error handlers now take two additional arguments. Consult the man page for PetscPushErrorHandler() for more information.

Event Logging:

  • Changed PLogPrintSummary(MPI_Comm,FILE *) to PLogPrintSummary(MPI_Comm,char *).

  • Now the option -log_summary takes [filename] as an optional argument.

Fortran Interface:

  • Added some limited support for direct use of Fortran90 pointers in the routines Vec[Get,Restore]ArrayF90(), Mat[Get,Restore]ArrayF90(), IS[Get,Restore]IndicesF90(), ISBlock[Get,Restore]IndicesF90(), VecDuplicateVecsF90(), VecDestroyVecsF90(), DAGetGlobalIndicesF90(). See the man pages and the section ‘Fortran90’ in the users manal for details. Unfortunately, these routines currently work only with the NAG F90 compiler. We hope to support other compilers as well, but we will need assistance from the vendors since the Fortran90/C interface is not a defined standard.

  • Added the macro PetscDoubleExp(a,b) = a d b (machines where double precision arithmetic is used) = a e b (machines where single precision arithmetic is used, e.g., Crays) This macro is intended for use only if you wish to maintain a Fortran code that is portable to both the Cray T3d/T3e and other Unix machines.

  • For mixed Fortran/C users: added the makefile flag FCONF that may be used in place of the flag CONF. For an example of usage, see src/vec/examples/tutorials/makefile