Add documentation that MATLAB-PETSc and PETSc->Mat transfers of matrices result
in the transpose of the matrix on the other side.

Commit-type: documentation
Reported-by: Jose E. Roman <jroman@dsic.u

Add documentation that MATLAB-PETSc and PETSc->Mat transfers of matrices result
in the transpose of the matrix on the other side.

Commit-type: documentation
Reported-by: Jose E. Roman <jroman@dsic.upv.es>

show more ...

Deprecate old {Mat|PC|KSP}ComputeExplicitOperator in favor of {Mat|PC|KSP}ComputeOperator

Merge remote-tracking branch 'origin/maint'

Merged in stefano_zampini/fix-matcopy-sbaij (pull request #1571)

Fix SBAIJ

Approved-by: BarryFSmith <bsmith@mcs.anl.gov>

Mat{MPI|MPISB|MPIB|Seq|SeqB|SeqSB}AIJSetPreallocation: optimize nonzeros

set number of columns of offdiag blocks to zero for sequential cases
no need to allocate more than the the number of columns

Mat{MPI|MPISB|MPIB|Seq|SeqB|SeqSB}AIJSetPreallocation: optimize nonzeros

set number of columns of offdiag blocks to zero for sequential cases
no need to allocate more than the the number of columns for sequential matrices

Update outputs for testsuite

show more ...

Set block sizes in MatPtAP

housekeeping

remove trailing blanks
remove functions that are no longer part of PETSc

Fix hash table map

* ValType cannot be assumed to support any unary/binary operator.
* Remove hashmapiv.h as it is trivial and used only once.

{Mat|PC|KSP}ComputeExplicitOperator: add extra argument for matrix type

Merged in Fande-Kong/feature-ptap-all-at-once (pull request #1552)

All-at-once algorithm for MPIPtAP

Approved-by: BarryFSmith <bsmith@mcs.anl.gov>

MPIPtAP: fixed some compiling warnings

MPIPtAP: implement all-at-once algorithm

scalable and nonscalable just take too much memory since they cache the intermediate data structure.

The all-at-once algorithm do all caculations at once wi

MPIPtAP: implement all-at-once algorithm

scalable and nonscalable just take too much memory since they cache the intermediate data structure.

The all-at-once algorithm do all caculations at once without caching any data. It saves a lot of the memory

There are two implementations: allatonce and allatonce_merged

``allatonce" overlapp the communication and the computation with taking a few extra operations.
``allatonce_merged" do all caculations together for remote and local parts.

Both are memory friendly

show more ...

Set blocksize correctly in MatTransposeMatMult

add support of MatMatTransposeSolve() for mumps Cholesky
Requested-by: Jaysaval, Piyoosh <Piyoosh.Jaysaval@pnnl.gov>

Merge branch 'maint'

Merged in hzhang/matmatops-alg (pull request #1548)

maintain algorithm type for mat-mat ops after calling MatFreeIntermediateDataStruct

Approved-by: BarryFSmith <bsmith@mcs.anl.gov>

Housekeeping: strip out __FUNC__ or __FUNCT__ from PETSc

MATMPIAIJ: Fix MatZeroRows and MatZeroRowsColumns

skip diagonal set if row index > number of global columns

x AND b must be present at the same time

raise error when layouts are not congruent (ins

MATMPIAIJ: Fix MatZeroRows and MatZeroRowsColumns

skip diagonal set if row index > number of global columns

x AND b must be present at the same time

raise error when layouts are not congruent (instead of segfaulting)

local nonew on diagonal block may be different across comm -> Always reduce

When asked to not keep the nonzero pattern with non-congruent layouts
it is better off allowing a small reallocation to happen instead of erroring

show more ...

MATSEQAIJ: fixes to MatZeroRows and MatShift for rectangular matrices

Diagonal entries must be set only when i < cmap->n

MatZeroRowsColumns_SeqAIJ: fixes for non-square matrices

cuda: version 10 requires explicit listing of thrust/device_malloc_allocator.h

CUDAC arch-cuda-double/obj/mat/impls/aij/seq/seqcusparse/aijcusparse.o
/home/nicolol/workspace/petsc/src/mat/imp

cuda: version 10 requires explicit listing of thrust/device_malloc_allocator.h

CUDAC arch-cuda-double/obj/mat/impls/aij/seq/seqcusparse/aijcusparse.o
/home/nicolol/workspace/petsc/src/mat/impls/aij/seq/seqcusparse/aijcusparse.cu(1400): error: namespace "thrust" has no member "device_malloc_allocator"

Reported-by: Lardelli Nicolo <nicolo.lardelli@erdw.ethz.ch>

show more ...

maintain algtype for mat-mat ops after calling MatFreeIntermediateDataStructures().
Requested by Fande Kong

MatView_MPIAIJ_ASCIIorSocket: speedup code

This code in not performance critial, but it took 3 hours to dump a 2000x2000 MPIAIJ matrix with dense values
This is what we obtain by calling MatViewFrom

MatView_MPIAIJ_ASCIIorSocket: speedup code

This code in not performance critial, but it took 3 hours to dump a 2000x2000 MPIAIJ matrix with dense values
This is what we obtain by calling MatViewFromOptions on a MATSHELL in parallel

show more ...

Merged in Fande-Kong/fixed_PtAP (pull request #1452)

MPIPtAP: Enable large-scale simulations

Approved-by: BarryFSmith <bsmith@mcs.anl.gov>

MPIPtAP and MatMatMult: take the prefix from the parent matrix

when calling an inner MatMatMult