Changes: Development

General

• Increase the minimum required CUDA Toolkit version to 11.4

• Change the policy so that documentation changes/fixes/additions are added to the main branch, not the release branch

• Change petscdiff to treat < in expected output as text that must match instead of suppressing differences from numerical output

• Change the make branch-review rule to [PETSC_LLM_CLI=command] [PETSC_LLM_MODEL=modelname] make branch-review. Add support for Gemini, Codex, OpenCode, and other LLM CLIs. Replace the use of CLAUDE_OPTS with PETSC_LLM_CLI_OPTS and PETSC_LLM_MODEL

• Add a CodeGraph skill and repository guidance so LLM coding tools automatically use an existing local PETSc CodeGraph index when navigating or reviewing source

• Add comm and prefix as initial arguments to PetscOptionsDeprecatedNoObject()

Configure/Build

• Increase the minimum required Python version for ./configure to 3.6

• Add providesDocs and docsDirs package attributes so that an external package’s sources can be cloned and scanned to generate PETSc manual pages when the documentation is built

• Add interface to LIBXSMM

Sys

• Add PetscGetConfiguration()

• Add PetscObjectViewSynchronizedFromOptions()

• Add PetscSetMPIThreadRequiredType()

• Add PetscOverrideIntelMKLCPUVendorDetection()

• Deprecate PETSC_MPI_THREAD_REQUIRED

• Add support for running PETSc applications as MCP servers that can be accessed by LLMs

• Add PetscRandomAppendOptionsPrefix() and PetscRandomGetOptionsPrefix()

• Add PetscCallLAPACKInfo() for calling LAPACK routines with an info argument where the caller requires an info value of 0 for the program to continue

• Add PetscIntCSRView() to inspect CSR graph

Event Logging

• Change -log and -log_all to only take optional filename arguments. They no longer take optional boolean argument

PetscViewer

• Add support for writing CGNS descriptors on the base node: PetscViewerCGNSGetDescriptors(), PetscViewerCGNSRestoreDescriptors(), PetscViewerCGNSSetDescriptor()

• Add PetscViewerVTKWriteFn as the typedef prototype for the write() function passed to PetscViewerVTKAddField(). This addition requires no changes to user source code

PetscDraw

AO

IS

• Add PetscLayoutGetComm() to obtain the MPI communicator of a PetscLayout

VecScatter / PetscSF

PF

Vec

• Add VecCreateSeqWithArrayAndMemType() and VecCreateMPIWithArrayAndMemType() to create array-style standard, CUDA, or HIP vectors from memory of a specified PetscMemType

• Add VecSetStdBasis() API to set a vector to the i-th standard basis vector

• Change the behavior of VecPointwiseDivide() implementing w = x / y: if a particular y[i] is zero and x[i] is also zero, w[i] is set to one (before it was set to zero).

• Deprecate -vec_view_stash in favor of -vec_stash_view

PetscSection

PetscPartitioner

Mat

• Add MATPRODUCT_PtAP support for MATDIAGONAL and MATCONSTANTDIAGONAL

• Add MATPRODUCT_AB support for MATDIAGONAL and MATCONSTANTDIAGONAL with any matrix type

• Add MatSeqAIJGetKokkosView(), MatSeqAIJRestoreKokkosView(), MatSeqAIJGetKokkosViewWrite() and MatSeqAIJRestoreKokkosViewWrite() to the public API

• Change MatSeqAIJCUSPARSEGetIJ(), MatSeqAIJCUSPARSERestoreIJ(), MatSeqAIJHIPSPARSEGetIJ() and MatSeqAIJHIPSPARSERestoreIJ() to return PetscInt indices instead of int

• Add MatNormApproximate() to compute matrix norms approximately

• Add MatGetMultPetscSF() to access the PetscSF used to communicate off-process vector entries in MatMult() for MATMPIAIJ, MATMPIBAIJ, MATMPISBAIJ, MATMPIDENSE, and MATMPISELL

• Change MATSEQSBAIJ behavior in MatPermute() and MatDiagonalScale() to use MATSEQBAIJ when the result is not necessarily symmetric

• Change MatGetValues() to respect the row or column orientation set with MatSetOption(mat, MAT_ROW_ORIENTED, ...). This will break current code that calls MatSetOption(mat, MAT_ROW_ORIENTED, PETSC_FALSE) and uses MatGetValues()

• Add new MatType MATSEQBAIJLIBXSMM and MATMPIBAIJLIBXSMM

• Add support for MatSetInf() with MATSEQDENSE and MATMPIDENSE; no MatType implemented it before

• Add device SpMM support for MATPRODUCT_AB and MATPRODUCT_AtB with a MATAIJKOKKOS matrix and dense matrices; previously these products looped MatMult() over the columns of the dense matrix

• Change MatCreateMAIJ() to convert its result to MATAIJKOKKOS when the input matrix has that type, as is already done for MATAIJCUSPARSE; the MATMAIJ kernels read the host arrays of the input matrix directly and so miss values last updated on device

• Change MatDiagonalScale() for MATSEQDENSECUDA and MATSEQDENSEHIP to check the memory type of the scaling Vec instead of its VecType, so device-resident vectors such as VECKOKKOS are consumed directly on the GPU instead of being copied through the host

• Add MatDenseUpdateColumnLayout() to change the layout of input vectors of the matrix vector product

• Change MatGetState() to return a MatState and add MatStateCompare(), MatStateCompareUpdate(), and MatStateInvalidate()

• Fix MatNorm() for MATMPIDENSE to respect the leading dimension of the local matrix, which was previously ignored in parallel for NORM_1 and NORM_FROBENIUS and gave wrong results, for example on matrices obtained with MatDenseGetSubMatrix()

• Add device implementations of MatNorm() with NORM_1, NORM_FROBENIUS, and NORM_INFINITY for MATDENSECUDA and MATDENSEHIP; previously all norms copied the matrix to the host

• Change MatNorm() for MATMPIDENSE to compute NORM_FROBENIUS and NORM_INFINITY via the local matrix norm, so MATMPIDENSECUDA and MATMPIDENSEHIP no longer copy to the host for those norms

MatCoarsen

PC

• Add PCGAMGSetProlongatorFilter() and PCGAMGGetProlongatorFilter() to set/get the threshold for filtering the prolongator in PCGAMG. The threshold is relative, applies to whole fine-node/coarse-node coupling blocks while preserving the near-null space, and must be in [0,1)

• Add PCGAMGSetProlongatorFilterScale() and PCGAMGGetProlongatorFilterScale() to set/get the per-level scaling of the prolongator filter threshold in PCGAMG; the scale must be in [0,1]

• PCGAMGSetThresholdScale() now requires its argument to be in [0,1]

• PCGAMGSetThreshold() now requires each threshold value to be less than 1; negative values still mean keeping even zero entries in the graph. It is also now Logically Collective (checked in debug builds), matching PCGAMGSetThresholdScale()

• Add PC_HPDDM_COARSE_CORRECTION_DEFLATED_REVERSED for applying the coarse correction after the fine correction in PCHPDDM

• Add PCAIR and PCPFLAREINV manual pages, generated from the PFLARE sources when the documentation is built

• Add PCParametersInitialize

• Fix PCMG to honor PCSetUseAmat(pc, PETSC_FALSE) at all levels

• Add PCMatApplyRichardson(), PCMatApplyRichardsonExists(), and PCShellSetMatApplyRichardson(), the block analogs of PCApplyRichardson(), PCApplyRichardsonExists(), and PCShellSetApplyRichardson()

• Add the missing Fortran binding for PCShellSetMatApply()

• Remove PCSetDiagonalScale(), PCGetDiagonalScale(), PCDiagonalScaleLeft(), and PCDiagonalScaleRight()

KSP

• Fix for KSP pre- and post-solve callbacks, that can now be used together with Eisenstat and Walker trick for SNES

• Add KSPPreSolve() and KSPPostSolve() to run the registered KSP pre/post solve callbacks

• Change KSPSolve() to run the KSPSetPreSolve() callback after KSPSetUp() and KSPSetUpOnBlocks() instead of before

• Add KSPIDR — IDR(s) Induced Dimension Reduction Krylov solver (biorthogonal variant)

• Add KSPIDRSetS(), KSPIDRGetS(), KSPIDRSetRandom(), KSPIDRGetRandom(), KSPIDRSetCosine(), and KSPIDRGetCosine()

• Deprecate KSPMonitorResidualShort() and -ksp_monitor_short, remove the preconditioned_residual_short monitor registry name

• Remove -ksp_plot_eigenvalues, -ksp_plot_eigenvalues_explicitly, -ksp_plot_eigencontours that have been deprecated since version 3.9

• Remove KSPSetDiagonalScale(), KSPGetDiagonalScale(), KSPSetDiagonalScaleFix(), and KSPGetDiagonalScaleFix(), along with the -ksp_diagonal_scale, -ksp_diagonal_scale_fix, and -ksp_view_diagonal_scale options

• Remove KSPAGMRES (Wakam and Erhel, 2011), which was never registered as a usable KSPType

• Add native support for KSPMatSolve() and KSPMatSolveTranspose() with KSPRICHARDSON, which iterates on batches of or the entire block of right-hand sides instead of solving them one at a time

• Add delegation of the KSPRICHARDSON block iteration to PCMatApplyRichardson() when the PC provides it, falling back to PCApplyRichardson() on one right-hand side at a time so that the result matches KSPSolve()

• Change KSPConvergedDefault() to base the relative tolerance on the Frobenius norm of the batch of (preconditioned) right-hand sides during a KSPMatSolve() with a nonzero initial guess, matching KSPSolve()

• Change KSPMatSolve() and KSPMatSolveTranspose() to reset the residual history at the start of each solve, and of each batch when -ksp_matsolve_batch_size is used, as KSPSolve() does, unless KSPSetResidualHistory() was called with reset set to PETSC_FALSE

• Change KSPRichardsonSetSelfScale() to trigger a KSPSetUp() re-run when the flag changes, fixing an out-of-bounds work vector access when it was set after KSPSetUp()

• GMRES orthogonalization routines have been promoted to the main KSP level, and GMRES has been dropped from the name. The new names are KSPOrthogonalizationSet(), KSPOrthogonalizationGet(), KSPOrthogonalizationModifiedGramSchmidt(), KSPOrthogonalizationClassicalGramSchmidt(), KSPOrthogonalizationSetCGSRefinementType(), KSPOrthogonalizationGetCGSRefinementType(). Note that the signature of orthogonalization functions has changed, see KSPOrthogonalizationFn. The related enumeration is now KSP_ORTHOGONALIZATION_CGS_REFINE_* and command-line options are now -ksp_orthogonalization (cgs|mgs) and -ksp_orthogonalization_cgs_refinement_type (refine_never|refine_ifneeded|refine_always)

SNES

• Change SNESSetUp() to not overwrite the NPC application context if one has previously been set on the NPC

• Change SNESComputeJacobian() to call the user-provided Jacobian function when a left NPC is active and the solver is not SNESASPIN

• Add support for nonlinear preconditioners with a DM different from the parent SNES DM. Calling SNESSetNPC() will no longer enforce default parameters on the NPC

• Change -snes_mf to respect an explicitly set PC type instead of silently overriding it with PCNONE; an explicitly requested PC that requires an assembled matrix now errors

• Deprecate SNESMonitorDefaultShort() and -snes_monitor_short

• Add SNESFASSetUseCoarseCorrectionLineSearch() and -snes_fas_use_coarse_correction_linesearch to implement the algorithm in [Nas00].

• Add SNESFASGetCoarseCorrectionLineSearch()

• Change default linesearch for SNESFAS with SNESFASType of SNES_FAS_ADDITIVE to SNESLINESEARCHSECANT

SNESLineSearch

• Deprecate SNESLINESEARCHBASIC in favor of SNESLINESEARCHNONE

• Add SNESLineSearchViewFromOptions()

TS

• Add DMTSSetIFunctionPre()

• Add TSDiscGradSetImplicitFormulation()

• Expose TSDiscGradGetX0AndXdot() and TSDiscGradRestoreX0AndXdot()

• Add TSIsImplicit() that indicates if the TSType is implicit and uses SNES or KSP

TAO

• Add TaoGetDM() and TaoSetDM()

• Deprecate TaoMonitorDefaultShort(), -tao_monitor_short, and -tao_monitor_short_interval

• Change the deprecated TaoSMonitor() and -tao_smonitor to use the full-precision default monitor

• Add TaoGetConvergedReasonString() to retrieve a human readable string describing the TaoConvergedReason

TaoTerm

PetscRegressor

PetscDA

• Add the PetscDALETKFLocalizationType enum (PETSCDA_LETKF_LOC_NONE, PETSCDA_LETKF_LOC_GASPARI_COHN, PETSCDA_LETKF_LOC_GAUSSIAN, PETSCDA_LETKF_LOC_BOXCAR) selecting the LETKF localization kernel

• Add PetscDALETKFSetLocalizationType(), PetscDALETKFGetLocalizationType(), PetscDALETKFSetLocalizationRadius(), PetscDALETKFGetLocalizationRadius(), and PetscDALETKFSetLocalizationCoordinates(); the localization matrix is built lazily from these distance-based kernel parameters

• Add PetscDALETKFResetLocalization() to drop the cached localization matrix so the next analysis rebuilds it from the current kernel parameters

• Remove PETSCDAETKF; use PETSCDALETKF with PetscDALETKFSetLocalizationType(da, PETSCDA_LETKF_LOC_NONE) for identical behavior

• Remove PetscDAEnsembleSetSqrtType(), PetscDAEnsembleGetSqrtType(), the PetscDASqrtType enum (PETSCDA_SQRT_CHOLESKY, PETSCDA_SQRT_EIGEN), and the -petscda_ensemble_sqrt_type option; the symmetric-eigendecomposition square root is now the only path

• Remove PetscDALETKFSetLocalization(); use the distance-based API PetscDALETKFSetLocalizationType(), PetscDALETKFSetLocalizationRadius(), and PetscDALETKFSetLocalizationCoordinates() instead

• Remove PetscDALETKFSetObsPerVertex() and PetscDALETKFGetObsPerVertex(); per-vertex observation counts are now derived from the distance-based localization kernel

• Remove PetscDALETKFGetLocalizationMatrix(); the localization matrix is an internal cached object built lazily on the first analysis. Callers that previously supplied this matrix should switch to PetscDALETKFSetLocalizationCoordinates() and let the implementation build the matrix from the chosen kernel

• Change the LETKF distance-based periodicity convention: per-axis periodicity is now activated by bd[d] > 0.0 (the period), and negative bd[d] now raises PETSC_ERR_ARG_OUTOFRANGE; previously any non-zero bd[d] (including negative values) enabled periodicity

• Add PetscDAEnsembleForecastFn typedef for the PetscDAEnsembleForecast() model callback

• Change the PetscDAEnsembleForecast() model callback signature from (Vec, Vec, PetscCtx) to (Mat, PetscCtx); the model now receives the entire ensemble matrix and advances all members in place. Existing per-member callbacks should iterate over the columns with MatDenseGetColumnVec()/MatDenseRestoreColumnVec() (see ShallowWaterStep2D() in src/ml/da/tutorials/ex4.c)

• Change -petscda_view to fire at the tail of every PetscDAEnsembleAnalysis() call (mirroring KSPSolve()/SNESSolve()), so it now emits once per analysis cycle rather than once per run; code that wants a single end-of-run snapshot should call PetscDAView() explicitly after the assimilation loop

DM

• Change DMLabelPropagatePush() to take a reduce operator

• Add DMKSPSetCreateOperators() to let the DM provide a pair of application specific Mat objects to inner KSP solvers.

• Fix DMGetLocalToGlobalMapping() for a local section carrying a chart permutation (see DMReorderSectionSetDefault()); the map is now indexed by the local section offsets. The local section must now be set up with PetscSectionSetUp() before the mapping is built, otherwise an error is raised

• Fix DMSetLocalSection() and DMSetGlobalSection() to invalidate a previously built section-derived local-to-global mapping, which was computed from the old sections; a mapping built by the DM implementation itself, such as by DMDA in DMSetUp(), is kept. DMFOREST with p4est shares its sections into its cached DMPLEX without going through those setters, and now invalidates the mapping there as well

DMSwarm

• Add DMSwarmProjectFields() and DMSwarmProjectGradientFields()

• Add DMSwarmSort class

• Add DMSwarmSortDestroy() and DMSwarmSortView()

• Allow DMSwarmCellDMSetSort() to take in NULL and clear the sort

• Add DMSwarmPreallocateMassMatrix() and DMSwarmFillMassMatrix()

DMPlex

• Add DMPlexSetClosurePermutationLexicographic()

• Add DMPlexDrawCell()

• Add DMPlexLabelCompleteStar()

• Add DMPlexVecGetClosureAtDepth()

• Add an extra communicator argument to DMPlexFilter() to allow extracting local meshes

• Add DMPlexCopyFlags()

• Add DMPlexRebalanceSharedLabelPoints()

• Add DMPlexCheckLabel() and DMPlexReconcileLabel()

• Change CGNS viewer to use multi-component read/write interface for better performance

• Add DMPlexTransformOrderSupports()

• Add DMPlexLabelCohesiveCheck()

• Add DMPlexCheckOrientationLabel()

• Change DMPlexLabelCohesiveComplete() to remove split argument

• Add DM_COORD_MAP_TORUS

• Add DM_COORD_MAP_ROTATE

• Add DM_SHAPE_DIIID

• Add DMPlexTriangleSetAngleBound(), DMPlexTriangleGetAngleBound(), DMPlexTetgenSetRadiusEdgeBound(), DMPlexTetgenGetRadiusEdgeBound(), DMPlexTetgenSetDihedralBound(), DMPlexTetgenGetDihedralBound()

FE/FV

DMNetwork

DMStag

DT

• Add PetscWeakFormGetKeys()

Fortran