PCBDDC: added KSPPIPECGRR to the list of methods support the exact dirichlet trick

PCBDDC: optimizations for the multilevel BDDC with benign subspace and CHEBYSHEV coarse solvers

Most of the times, the computation of the benign correction is not needed.
Added benign_have_null, ben

PCBDDC: optimizations for the multilevel BDDC with benign subspace and CHEBYSHEV coarse solvers

Most of the times, the computation of the benign correction is not needed.
Added benign_have_null, benign_need_correction, so that we can decide at command line to not compute the benign correction for Stokes

show more ...

PCBDDC: small fixes

Print graph used to compute connected components when needed
solved memory leak

Fix serious bug introduced with introduction of XXXMonitorSetFromOptions() that resulting in
excessive pushes to viewer formats and caused different viewers to overwrite each others formats

Time: .7

Fix serious bug introduced with introduction of XXXMonitorSetFromOptions() that resulting in
excessive pushes to viewer formats and caused different viewers to overwrite each others formats

Time: .7 hours
Reported-by: Lawrence Mitchell <lawrence.mitchell@imperial.ac.uk>

show more ...

PCBDDC: improve PCView_BDDC

Added control for the explicit change of variables for pressure dofs

PCBDDC: bugfix for benign_n > 1

the values were inserted in cs_AIB in column major format
the fix consists in changing the shape of cs_AIB

PCBDDC: add runtime options to control dirichlet dofs elimination in presolve

Most the time removing dir dofs does not have an impact on the global convergence

PCBDDC: remove usage of sf_nroots and sf_nleaves members of Mat_IS data structure

fixed a memory leak

PCBDDC: optimize refinement of connected components

The previous code used a lot of memory constructing an ad-hoc csr information to refine the connected components
Changed the way the connected com

PCBDDC: optimize refinement of connected components

The previous code used a lot of memory constructing an ad-hoc csr information to refine the connected components
Changed the way the connected component are searched in PCBDDCComputeConnectedComponentsLocal: now the search is symmetric

show more ...

PCBDDC: output the global id of the wrongly computed coarse index

PCBDDC: use BLAS rank updates routines

PCBDDC: less stringent check on benign trick applicability

PCBDDC: remove faster deluxe switch

PCBDDC: use MPIU_REPLACE to for SFReduce ops when setting the p0 values (avoid VecAssemblyBegin/End)

PCBDDC: remove the horrible logic around the remap of the coarse problem

the code now is more readable, maintanable. In case of multilevel, it is also faster, since one shift is avoided.
when the co

PCBDDC: remove the horrible logic around the remap of the coarse problem

the code now is more readable, maintanable. In case of multilevel, it is also faster, since one shift is avoided.
when the coarse matrix has to be assembled, it still uses MatISSubassemble followed by MatISGetMPIXAIJ

fixed some bugs in MatISSubassemble and MatISGetSubassemblingPattern, with preliminary checks added

show more ...

PCBDDC: multilevel fixes

allow simple redistribution on void procs (if any) at level 0

PCBDDC: allow processes to pass 0 nvtxs and void arrays

PCBDDC: make change_primal_sub a member of sub_schurs

we need it during adaptive selection
sub_schurs->change and sub_schurs->chane_primal_sub are created when needed

PCBDDC: Added support for adaptive selection of constraints with pointwise scalings

PCBDDC: use PtAP for SeqDense from master

PCBDDC: use MAT_INPLACE_MATRIX

Merge branch 'master' into stefano_zampini/feature-pcbddc-saddlepoint

Conflicts:
src/ksp/pc/impls/bddc/bddc.c
src/ksp/pc/impls/bddc/bddcgraph.c
src/ksp/pc/impls/bddc/bddcprivate.c
src/ksp/pc/imp

Merge branch 'master' into stefano_zampini/feature-pcbddc-saddlepoint

Conflicts:
src/ksp/pc/impls/bddc/bddc.c
src/ksp/pc/impls/bddc/bddcgraph.c
src/ksp/pc/impls/bddc/bddcprivate.c
src/ksp/pc/impls/bddc/bddcscalingbasic.c
src/ksp/pc/impls/bddc/bddcschurs.c
src/mat/impls/is/matis.c

show more ...

PCBDDC: graph rebuild needs to use the same custom vertex size as the original graph

PCBDDC: fix for static switch

The previous code was not applying M_3 preconditioner as outlined in Dohrmann's reference article in case the preconditioner was freezed

PCBDDC: fix for upcoming release.

When the user provides a change of basis matrix, pcis data structure may not be initialized, and we need to use pc->pmat->rmap->mapping