Add support of CUDA-11

Revert "CUDA: Get PETSc to compile with CUDA 11."

This reverts commit 21a6055d1ede0202436f221d06f48a7b9836052d.

MUMPS: add -mat_mumps_icntl_20 to select distribution format of RHS at runtime

Change the default based on MUMPS version and MPI implementation.

Fix error with MatAXPY_Seq/MPIAIJ that lost the layout and MatStencilInfo information from the input matrix. The code has not been touched
for five years so it is odd the bug was never exposed before

Fix error with MatAXPY_Seq/MPIAIJ that lost the layout and MatStencilInfo information from the input matrix. The code has not been touched
for five years so it is odd the bug was never exposed before.

Add a precheck in MatAXPY_SeqAIJ() for identical nonzero pattern to avoid the constant recomputation of the resulting matrix.

Commit-type: optimization, bug-fix
/spend 3h
Reported-by: Ed Bueler

show more ...

Merge branch 'karlrupp/fix-cuda-11-compilation' into 'master'

CUDA: Get PETSc to compile with CUDA 11.

See merge request petsc/petsc!3154

CUDA: Get PETSc to compile with CUDA 11.

CUSPARSE has changed significantly in CUDA 11.
This has lead to a lot of broken code.
This commit disables the AIJCUSPARSE class so that PETSc at least compi

CUDA: Get PETSc to compile with CUDA 11.

CUSPARSE has changed significantly in CUDA 11.
This has lead to a lot of broken code.
This commit disables the AIJCUSPARSE class so that PETSc at least compiles with other GPU features enabled.

show more ...

Allow duplication of an empty matrix

Fix several typos in error strings

SeqDense treats ILU as LU allowing PC defaults to work when dense matrix
is not yet known

DMSHELL uses dense format if no format provided.

Fix manual pages for MatFactor

Commit-type: feature, test

SeqDense treats ILU as LU allowing PC defaults to work when dense matrix
is not yet known

DMSHELL uses dense format if no format provided.

Fix manual pages for MatFactor

Commit-type: feature, testing-fix, documentation
/spend 1h

show more ...

Added MatSeqAIJSetTotalPreallocation() allows efficiently building a SeqAIJ matrix row by row without
preallocating each row instead just provide an upper bound on the entire memory needed.

Commit-t

Added MatSeqAIJSetTotalPreallocation() allows efficiently building a SeqAIJ matrix row by row without
preallocating each row instead just provide an upper bound on the entire memory needed.

Commit-type: optimization
/spend 4h
Reported-by: Sajid Ali <sajidsyed2021@u.northwestern.edu>

show more ...

balay/fix-checkbadSource: fix the line count, and false positive with Kokkos::parallel_for()

Turn on checkbadSource test to generate an error when found; fix all source code that causes errors

Commit-type: portability-fix, testing-fix, style-fix, feature, maintainability
/spend 1.5h

Merge branch 'dalcinl/fix-null' into 'master'

Fix 0 -> NULL for pointers

See merge request petsc/petsc!3086

Mat: Replace 0 -> NULL for pointers

Add manual page links for MatSeqAIJSetPreallocationCSR() and more notes and developer notes

Removed computed but unused variable

Commit-type: optimization, documentation
/spend 30m
Thanks-to: Adol

Add manual page links for MatSeqAIJSetPreallocationCSR() and more notes and developer notes

Removed computed but unused variable

Commit-type: optimization, documentation
/spend 30m
Thanks-to: Adolfo Rodriguez <adantra@gmail.com>

show more ...

Merge branch 'jczhang/add-kokkos-init' into 'master'

Add Kokkos initialization

See merge request petsc/petsc!3048

Allow resusing SuperLU_DIST communicator and grid for multiple new SuperLU_DIST matrices

Note: When possible, same size matrix, same parallel distribution, same nonzero pattern
one should reuse the

Allow resusing SuperLU_DIST communicator and grid for multiple new SuperLU_DIST matrices

Note: When possible, same size matrix, same parallel distribution, same nonzero pattern
one should reuse the matrix and the KSP for new solvers, this allows reusing portions of the
preconditioner assembly process.

Commit-type: optimization
Funded-by: ECP
/spend 2.2h
Reported-by: "Xiaoye S. Li" <xsli@lbl.gov>

show more ...

Add support of CUDA/HIP initialization at the same time

Merge remote-tracking branch 'origin/maint'

cudavecimpl.h move to include/petsc/private as its used by private/sfimpl.h

Reported-by: Nidish <nb25@rice.edu>

Mark MKL matrix handle as not optimized after MatProduct operations.

I had previously incorrectly assumed that the matrix optimization step is
carried out by MKL at the end of a matrix-matrix produc

Mark MKL matrix handle as not optimized after MatProduct operations.

I had previously incorrectly assumed that the matrix optimization step is
carried out by MKL at the end of a matrix-matrix product operation, but
a conversation with the MKL team confirms that this is not the case.
This did not cause a correctness issue, but might have led to reduced
performance if the matrix was then used for repeated operations that could
benefit from the optimization step.

Thanks-to: Maria Zhukova <maria.zhukova@intel.com>

show more ...

Do lazy petsc cuda initialization and validation

So that we can avoid overburdening the GPU in CI.
See more at https://gitlab.com/petsc/petsc/-/issues/360#note_373723821

Merge remote-tracking branch 'origin/maint'

Limit chance of integer overflow in calls to PetscLogFlops()

Don't view inner KSP when it has not yet been created.

Commit-type: bug-fix
Funded-by: ECP
/spend 20m
Reported-by: Karl Yang <y.juntao@

Limit chance of integer overflow in calls to PetscLogFlops()

Don't view inner KSP when it has not yet been created.

Commit-type: bug-fix
Funded-by: ECP
/spend 20m
Reported-by: Karl Yang <y.juntao@hotmail.com>

show more ...

Eliminate trailing whitespaces in aijmkl.c.

Thanks-to: Satish Balay <balay@mcs.anl.gov>