PCBDDC: renumber reference nodes for interface adaptation

the renumber will remove "holes" in the subset

PCIS; fix bug with PCISDestroy

if PCISCreate is not called (and this can only happen in PCSetUp_NN ot PCSetUp_BDDC) the l2g map info is not created.
The latter situation happens if a user asks for P

PCIS; fix bug with PCISDestroy

if PCISCreate is not called (and this can only happen in PCSetUp_NN ot PCSetUp_BDDC) the l2g map info is not created.
The latter situation happens if a user asks for PCBDDC or PCNN inside the code (using PCSetType) and then requests a different PC at command line

show more ...

PCBDDC: no need to reorder the constraints in local ordering after SVD

This also guaranties that the change of basis computed with QR will be coherent among processes
It also extends the reordering

PCBDDC: no need to reorder the constraints in local ordering after SVD

This also guaranties that the change of basis computed with QR will be coherent among processes
It also extends the reordering to the case of using QR on a single cc

show more ...

PCBDDC: fixed bug when using more than 1 constraint per connected component via SVD

The bug originated from the (wrong!) assumption that the returned eigenmodes from the SVD would be consistent amon

PCBDDC: fixed bug when using more than 1 constraint per connected component via SVD

The bug originated from the (wrong!) assumption that the returned eigenmodes from the SVD would be consistent among sharing processes
This is not true since the computed SVD depends on the local ordering of dofs and each process computes its own SVD (separately for each connected component).
The fix consists in ordering the constrained indices consistently across processes by using global ordering

show more ...

PCBDDC: we should also propagate the type of PC to the sequential deluxe solver

FETIDP: bug fix. Rhs vector passed wasn't properly restored if change of basis or dirichlet dofs were not present.

Merge branch 'stefano_zampini/fixes-for-raviart-thomas'

Merge branch 'stefano_zampini/fixes-and-improvements-multilevel'

PCBDDC: avoid LAPACK calls if either nnsp_true or just one constraint per cc is required

PCBDDC: fix wrong computation of change of basis blocks

The last column of the block should be given by the orthonormalized coefficients of the quadrature rule
Note that the computed block is the sa

PCBDDC: fix wrong computation of change of basis blocks

The last column of the block should be given by the orthonormalized coefficients of the quadrature rule
Note that the computed block is the same as before if computed with respect the arithmetic average

show more ...

PCBDDC: QR decomposition for change of basis can now be also used for connected components with only 1 constraint

This partially solves the following issue with Raviart Thomas elements
Using standar

PCBDDC: QR decomposition for change of basis can now be also used for connected components with only 1 constraint

This partially solves the following issue with Raviart Thomas elements
Using standard change of basis, the coarse matrix produced is singular and CG doesn't converge to the exact solution (even if the eigenvalues of the preconditioned operator are the same of the unchanged basis case)
On the other hand, using QR, CG converges to the exact solution but the eigenvalues of the preconditioned operator are very different from not using the change of basis.
Still needs to be investigated...

show more ...

PCBDDC: for raviart thomas elements, the nearnullspace contains face orientations and should not be modified internally by PCBDDC

allow PCMG user to provide either interpolation or restriction operator with Galerkin coarse grid operators

convert some PetscMalloc() to PetscMalloc1()
remove unneeded () in first argument to PetscMalloc/Calloc1()
fixes for MATLAB examples

PCBDDC; forgot to remove debugging print

Merge branch 'stefano_zampini/fixes-and-improvements-multilevel'

PCBDDC: added options to redistribute coarse problem on a subset of cores

PCBDDC: dead code

PCBDDC: fixed preallocation of precomputed change of basis

PCBDDC: Added preallocation for new local matrices inside MatISSubassemble

It currently works with multilevel BDDC since the general case for subassembling is still missing

PCBDDC: no need for a double loop

PetscMalloc replaced by PetscMalloc1

PCBDDC: change adjacency weights threshold concept when subassembling a MATIS object

PCBDDC: zero size memory leak

Merge branch 'stefano_zampini/fixes-and-improvements-multilevel'

PCBDDC: VecSetValuesLocal should be guarded when doing multilevel since it checks for valid pointer even if n is zero