PCBDDC: Constraints setup improved

- Change of basis is now performed using QR factorization of constraints
- Added missing support for complexes when PETSC_MISSING_GESVD is defined
- Improved co

PCBDDC: Constraints setup improved

- Change of basis is now performed using QR factorization of constraints
- Added missing support for complexes when PETSC_MISSING_GESVD is defined
- Improved code readability

show more ...

PCBDDC: Store special marks for local dofs in a separate memory of PCBDDCGraph (not which_dof)

PCBDDC: bugfix in PCPostSolve_BDDC.

Right hand side was not get back to its original state when change of basis was not requested.

PCBDDC: bugfix when multiple constraints are present on the same connected component

Local to global mapping of coarse dofs was badly computed

Merge branch 'stefano_zampini/pcbddc-for-complexes'

* stefano_zampini/pcbddc-for-complexes:
PCBDDC: Solved compilation errors and warnings for complex builds

Merge branch 'stefano_zampini/pcbddc-development'

Conflict with 'jed/matnullspaceremove-inplace', resolved using:

$ git merge -X theirs --no-commit stefano_zampini/pcbddc-development
$ git grep

Merge branch 'stefano_zampini/pcbddc-development'

Conflict with 'jed/matnullspaceremove-inplace', resolved using:

$ git merge -X theirs --no-commit stefano_zampini/pcbddc-development
$ git grep -l MatNullSpaceRemove src/ksp | xargs perl -pi -e 's@(MatNullSpaceRemove\([^,]+,[^,]+),NULL\)@$1)@'
$ git commit -a

* stefano_zampini/pcbddc-development:
PCBDDC: add wrappers for PetscStackCallBLAS
PCBDDC: Final fix for locally periodic problems
PCNNCreateCoarseMatrix: fix allocation error
PCIS: consider myself as a neighbor when classifying interface nodes
PCBDDC: Remove misuses of neighbours_set in FETI-DP code
PCBDDC: Coarse solution lives in the span of phi (not psi)
MATIS: Add IsSymmetric and IsHermitian implementations
PCBDDC: More robust method to compute local boundary nodes
PCBDDC: Add partial support for periodic problems replacing Sort with RemoveDups
PCBDDC: Add support for non-symmetric linear systems
PCBDDC: Stiffness scaling should be rebuilt in any case (just to be safe)
PCBDDC: Remove trailing blanks
PCBDDC: Consistency of CSR graph passed in postponed to PCSetUp
PCBDDC: Preliminary work to split PCBDDCCoarseSetUp in simpler pieces
PCBDDC: allow compilation without metis header
PCBDDC: Optimal allocation in PCBDDCGetPrinalVerticesLocalIdx
PCBDDC: add verbosity level to graph view
PCBDDC: change boolean debug flag to integer -pc_bddc_check_level
PCBDDC: Workaround for BAIJ local matrices: convert to AIJ for coarse matrix
PCIS: Use PETSC_COMM_SELF instead of MPI_COMM_SELF
PCBDDC: many changes
ISG2LMapApply: fix __FUNCT__ and invalid access before globals are set up

show more ...

PCBDDC: Solved compilation errors and warnings for complex builds

Example code changed by using LAPACK steqr method.

PCBDDC: add wrappers for PetscStackCallBLAS

PCBDDC: Final fix for locally periodic problems

Added mirrors to graph structure

PCBDDC: Remove misuses of neighbours_set in FETI-DP code

PCBDDC: Coarse solution lives in the span of phi (not psi)

PCBDDC: Add partial support for periodic problems replacing Sort with RemoveDups

* Local-to-global maps which are not one-to-one if restricted to any
subdomain will still be a problem.
* Graph vie

PCBDDC: Add partial support for periodic problems replacing Sort with RemoveDups

* Local-to-global maps which are not one-to-one if restricted to any
subdomain will still be a problem.
* Graph view now prints also the global number of the queued indices.
* Optimal allocation for queue_in_global_numbering.

show more ...

PCBDDC: Add support for non-symmetric linear systems

PCBDDC: Stiffness scaling should be rebuilt in any case (just to be safe)

PCBDDC: Remove trailing blanks

PCBDDC: Consistency of CSR graph passed in postponed to PCSetUp

CSR graph is freed if the PC type is not PCBDDC and PETSC_OWN_POINTER
has been requested.

(Thanks to Lisandro Dalcin to report on thi

PCBDDC: Consistency of CSR graph passed in postponed to PCSetUp

CSR graph is freed if the PC type is not PCBDDC and PETSC_OWN_POINTER
has been requested.

(Thanks to Lisandro Dalcin to report on this)

show more ...

PCBDDC: Preliminary work to split PCBDDCCoarseSetUp in simpler pieces

PCBDDC: allow compilation without metis header

Coarse problem type default to PARALLEL_BDDC
Some irrelevant warning will show up without metis headers (will be solved soon)

PCBDDC: Optimal allocation in PCBDDCGetPrinalVerticesLocalIdx

PCBDDC: add verbosity level to graph view

PCBDDC: change boolean debug flag to integer -pc_bddc_check_level

Distribution of local primal indices requires level > 1.

PCBDDC: Workaround for BAIJ local matrices: convert to AIJ for coarse matrix

PCBDDC: many changes

- Source code split
- Added Reset procedures
- Support for different MatStructure flags

Mat: remove third argument to MatNullSpaceRemove

Out of place removals were dangerous because the MatNullSpace held the
only reference to the resulting Vec, which is dangerous. This use case
is rar

Mat: remove third argument to MatNullSpaceRemove

Out of place removals were dangerous because the MatNullSpace held the
only reference to the resulting Vec, which is dangerous. This use case
is rare (never used in PETSc itself) and can easily be accomplished:

VecDuplicate(x,&y);
VecCopy(x,y);
MatNullSpaceRemove(nullsp,y);

Suggested-by: Tobin Isaac <tisaac@ices.utexas.edu>

show more ...

PetscStackCallBLAS: new macro instead of PetscStackCall (skip CHKMEMQ)

When CHKMEMQ occurs on every call to BLAS or Lapack, we sometimes see
unacceptable performance degradation (orders of magnitude

PetscStackCallBLAS: new macro instead of PetscStackCall (skip CHKMEMQ)

When CHKMEMQ occurs on every call to BLAS or Lapack, we sometimes see
unacceptable performance degradation (orders of magnitude).
PetscStackCallBLAS is currently identical to PetscStackCall except for
skipping CHKMEMQ, but such checking could be made optional in the
future.

show more ...