apparently fixed a overflow bug.

Merge branch 'stefano_zampini/pcbddc-baijfixes'

* stefano_zampini/pcbddc-baijfixes:
PCBDDC: updated example using local to global mapping generated by DMDA
PCBDDC: solved a bug when multiple con

Merge branch 'stefano_zampini/pcbddc-baijfixes'

* stefano_zampini/pcbddc-baijfixes:
PCBDDC: updated example using local to global mapping generated by DMDA
PCBDDC: solved a bug when multiple constraints are present and some are eliminated by SVD
PCBDDC: index set for R nodes now stores block size information
PCBDDC: updated todo list
PCBDDC: updated bddc example with sbaij matrix type
PCBDDC: updated man pages
PCBDDC: bddc now works with sbaij matrices.
PCBDDC: added field in bddc data struct to determine if fields splitting has been provided by the user
PCIS: add support for sbaij matrices
PCBDDC: updated example using local to global mapping generated by DMDA
PCBDDC: solved a bug when multiple constraints are present and some are eliminated by SVD
PCBDDC: index set for R nodes now stores block size information
PCBDDC: updated todo list
PCBDDC: updated bddc example with sbaij matrix type
PCBDDC: updated man pages
PCBDDC: bddc now works with sbaij matrices.
PCBDDC: added field in bddc data struct to determine if fields splitting has been provided by the user
PCIS: add support for sbaij matrices

show more ...

Merge branch 'madams/gamg-parmtests'

PCGAMG default smoother changed from Jacobi to SOR. This affects
convergence for users of PCGAMG that have not explictly set smoothers.
The old smoother can be

Merge branch 'madams/gamg-parmtests'

PCGAMG default smoother changed from Jacobi to SOR. This affects
convergence for users of PCGAMG that have not explictly set smoothers.
The old smoother can be obtained using -mg_levels_pc_type jacobi.

* madams/gamg-parmtests:
changes/dev.html: document PCGAMG change from PCJACOBI to PCSOR
gamg tests: made file names more consistant; removed excess parameters to make test sort of orthogal with other test in that there are not parameters or combinations of parameters that are tested in another test; added a hypre test with two degrees of freedom; reduced sizes of tests to run faster; made a cheby/jacobi test.
changed default smoother in GAMG to SOR. Allowed reuse of iegen estimates for SOR smoother because it is safe and I don't think the difference is worth recommunting eigen estimates.

show more ...

doc: remove UTF-8 '–' from sources as this breaks generation of html for sourcefiles

PCGAMG: set default type (AGG); PCSetUp_GAMG crashes if not set

If PCSetFromOptions had not been called, PCSetUp_GAMG would crash due to
missing subtype. It is more conventional with top-level obje

PCGAMG: set default type (AGG); PCSetUp_GAMG crashes if not set

If PCSetFromOptions had not been called, PCSetUp_GAMG would crash due to
missing subtype. It is more conventional with top-level objects to use
a lazy choice of type, but there is no great setup involved, and no
known order-dependence, so it looks safe to simply set the type to the
default value (AGG) up-front.

Reported-by: Einar Sørheim <einar.sorheim@gmail.com>
Based-on-patch-by: Mark Adams <mfadams@lbl.gov>

show more ...

PCBDDC: introduce a new variable in PCBDDC data stucture to store the actual number of current vertices (corners with constraints with basis changed)

removed calls to PCBDDCGetPrimalVerticesLocalIdx

PCBDDC: introduce a new variable in PCBDDC data stucture to store the actual number of current vertices (corners with constraints with basis changed)

removed calls to PCBDDCGetPrimalVerticesLocalIdx which are no longer needed

show more ...

PCBDDC: make BtoNmap a member of bddc data struct (created once after PCISSetUp)

Use GlobalToLocalMapping to map local to boundary indices
Inserted some memcpy to speed up memory movement

PCBDDC: check of dirichlet and neumann solvers is not needed if pcbddc->NullSpace is not present

Added some PetscSynchronizedAllow since on some machines the verbose output is broken otherwise

PCBDDC: improved reuse of coarse matrix

MatCovert_IS_AIJ renamed in MatISGetMPIXAIJ (MatConvert does not support out-of-place reuse of destination matrix)
Added some lines in MATIS code (commented).

PCBDDC: better management of primal space. Now a coarse numbering is computed if and only if the primal space has been really changed

PCBDDC: allow different customization of dirichlet and neumann boundary indices among successive linear solves

PCBDDC: now PCBDDC reuses part of its private data if NearNullSpace customization is changed between different linear solves

I know it is a large commit, but this is the minimal set of changes to ge

PCBDDC: now PCBDDC reuses part of its private data if NearNullSpace customization is changed between different linear solves

I know it is a large commit, but this is the minimal set of changes to get it working

show more ...

PCBDDC: local work vectors should be created only if new constraints are created and the relevant sizes are different

PCBDDC/PCIS: make vec4_D a member of pcis, not pcbddc

PCBDDC: Prepare for different primal spaces customization among different linear solves

now bddc primal space is built using a NullSpace attached to pmat. It only remains to determine whether the Ne

PCBDDC: Prepare for different primal spaces customization among different linear solves

now bddc primal space is built using a NullSpace attached to pmat. It only remains to determine whether the NearNullSpace object has changed

show more ...

PCBDDC: demand object destruction to setup routines for BDDCGraph and constraints

PCBDDC: updated example using local to global mapping generated by DMDA

PCBDDC: solved a bug when multiple constraints are present and some are eliminated by SVD

change_of_basis flag should be set AFTER SVD has recognized independent components
Some minor modications to

PCBDDC: solved a bug when multiple constraints are present and some are eliminated by SVD

change_of_basis flag should be set AFTER SVD has recognized independent components
Some minor modications to the code (due to broken MatMult_BAIJ with compressed rows)

show more ...

PCBDDC: index set for R nodes now stores block size information

When IS block size is equal to Mat block size, there's no need to convert local matrices anymore

PCBDDC: updated todo list

added a further guard before calling SVDs

PCBDDC: updated man pages

PCBDDC: bddc now works with sbaij matrices.

This is a partial solution, since in practice the Neumann problem matrix will always be seqaij

Implemented user switch for field split

PCBDDC: added field in bddc data struct to determine if fields splitting has been provided by the user

Initialization of number of field splits was missing. Now fixed

PCIS: add support for sbaij matrices

pc->mat should be pc->pmat.

PC: Removed support graph PC