impls/bddc/bddc.c

/* TODOLIST
   DofSplitting and DM attached to pc?
   Change SetNeumannBoundaries to SetNeumannBoundariesLocal and provide new SetNeumannBoundaries (same Dirichlet)
   change how to deal with the coarse problem (PCBDDCSetCoarseEnvironment):
     - simplify coarse problem structure -> PCBDDC or PCREDUDANT, nothing else -> same comm for all levels?
     - remove coarse enums and allow use of PCBDDCGetCoarseKSP
     - remove metis dependency -> use MatPartitioning for multilevel -> Assemble serial adjacency in PCBDDCAnalyzeInterface?
   code refactoring:
     - pick up better names for static functions
   change options structure:
     - insert BDDC into MG framework?
   provide other ops? Ask to developers
   remove all unused printf
   man pages
*/

/* ----------------------------------------------------------------------------------------------------------------------------------------------
   Implementation of BDDC preconditioner based on:
   C. Dohrmann "An approximate BDDC preconditioner", Numerical Linear Algebra with Applications Volume 14, Issue 2, pages 149-168, March 2007
   ---------------------------------------------------------------------------------------------------------------------------------------------- */

#include "bddc.h" /*I "petscpc.h" I*/  /* includes for fortran wrappers */
#include "bddcprivate.h"
#include <petscblaslapack.h>

/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCSetFromOptions_BDDC"
PetscErrorCode PCSetFromOptions_BDDC(PC pc)
{
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = PetscOptionsHead("BDDC options");CHKERRQ(ierr);
  /* Verbose debugging */
  ierr = PetscOptionsInt("-pc_bddc_check_level","Verbose output for PCBDDC (intended for debug)","none",pcbddc->dbg_flag,&pcbddc->dbg_flag,NULL);CHKERRQ(ierr);
  /* Primal space cumstomization */
  ierr = PetscOptionsBool("-pc_bddc_use_vertices","Use or not corner dofs in coarse space","none",pcbddc->use_vertices,&pcbddc->use_vertices,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsBool("-pc_bddc_use_edges","Use or not edge constraints in coarse space","none",pcbddc->use_edges,&pcbddc->use_edges,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsBool("-pc_bddc_use_faces","Use or not face constraints in coarse space","none",pcbddc->use_faces,&pcbddc->use_faces,NULL);CHKERRQ(ierr);
  /* Change of basis */
  ierr = PetscOptionsBool("-pc_bddc_use_change_of_basis","Use or not change of basis on local edge nodes","none",pcbddc->use_change_of_basis,&pcbddc->use_change_of_basis,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsBool("-pc_bddc_use_change_on_faces","Use or not change of basis on local face nodes","none",pcbddc->use_change_on_faces,&pcbddc->use_change_on_faces,NULL);CHKERRQ(ierr);
  if (!pcbddc->use_change_of_basis) {
    pcbddc->use_change_on_faces = PETSC_FALSE;
  }
  /* Switch between M_2 (default) and M_3 preconditioners (as defined by C. Dohrmann in the ref. article) */
  ierr = PetscOptionsBool("-pc_bddc_switch_static","Switch on static condensation ops around the interface preconditioner","none",pcbddc->switch_static,&pcbddc->switch_static,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsInt("-pc_bddc_coarsening_ratio","Set coarsening ratio used in multilevel coarsening","none",pcbddc->coarsening_ratio,&pcbddc->coarsening_ratio,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsInt("-pc_bddc_levels","Set maximum number of levels for multilevel","none",pcbddc->max_levels,&pcbddc->max_levels,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsBool("-pc_bddc_use_deluxe_scaling","Use deluxe scaling for BDDC","none",pcbddc->use_deluxe_scaling,&pcbddc->use_deluxe_scaling,NULL);CHKERRQ(ierr);
  ierr = PetscOptionsTail();CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetPrimalVerticesLocalIS_BDDC"
static PetscErrorCode PCBDDCSetPrimalVerticesLocalIS_BDDC(PC pc, IS PrimalVertices)
{
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = ISDestroy(&pcbddc->user_primal_vertices);CHKERRQ(ierr);
  ierr = PetscObjectReference((PetscObject)PrimalVertices);CHKERRQ(ierr);
  pcbddc->user_primal_vertices = PrimalVertices;
  PetscFunctionReturn(0);
}
#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetPrimalVerticesLocalIS"
/*@
 PCBDDCSetPrimalVerticesLocalIS - Set user defined primal vertices in PCBDDC.

   Not collective

   Input Parameters:
+  pc - the preconditioning context
-  PrimalVertices - index sets of primal vertices in local numbering

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetPrimalVerticesLocalIS(PC pc, IS PrimalVertices)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidHeaderSpecific(PrimalVertices,IS_CLASSID,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetPrimalVerticesLocalIS_C",(PC,IS),(pc,PrimalVertices));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetCoarseningRatio_BDDC"
static PetscErrorCode PCBDDCSetCoarseningRatio_BDDC(PC pc,PetscInt k)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;

  PetscFunctionBegin;
  pcbddc->coarsening_ratio = k;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetCoarseningRatio"
/*@
 PCBDDCSetCoarseningRatio - Set coarsening ratio used in multilevel coarsening

   Logically collective on PC

   Input Parameters:
+  pc - the preconditioning context
-  k - coarsening ratio

   Approximatively k subdomains at the finer level will be aggregated into a single subdomain at the coarser level.

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetCoarseningRatio(PC pc,PetscInt k)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidLogicalCollectiveInt(pc,k,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetCoarseningRatio_C",(PC,PetscInt),(pc,k));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

/* Set level is not a public function */
#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLevel"
PetscErrorCode PCBDDCSetLevel(PC pc,PetscInt level)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidLogicalCollectiveInt(pc,level,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetLevel_C",(PC,PetscInt),(pc,level));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLevel_BDDC"
static PetscErrorCode PCBDDCSetLevel_BDDC(PC pc,PetscInt level)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;

  PetscFunctionBegin;
  pcbddc->current_level=level;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLevels_BDDC"
static PetscErrorCode PCBDDCSetLevels_BDDC(PC pc,PetscInt levels)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;

  PetscFunctionBegin;
  pcbddc->max_levels = levels;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLevels"
/*@
 PCBDDCSetLevels - Sets the maximum number of levels within the multilevel approach.

   Logically collective on PC

   Input Parameters:
+  pc - the preconditioning context
-  levels - the maximum number of levels

   Default value is 0, i.e. coarse problem will be solved exactly

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetLevels(PC pc,PetscInt levels)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidLogicalCollectiveInt(pc,levels,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetLevels_C",(PC,PetscInt),(pc,levels));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetNullSpace_BDDC"
static PetscErrorCode PCBDDCSetNullSpace_BDDC(PC pc,MatNullSpace NullSpace)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = PetscObjectReference((PetscObject)NullSpace);CHKERRQ(ierr);
  ierr = MatNullSpaceDestroy(&pcbddc->NullSpace);CHKERRQ(ierr);
  pcbddc->NullSpace = NullSpace;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetNullSpace"
/*@
 PCBDDCSetNullSpace - Set NullSpace of global operator of BDDC preconditioned mat.

   Logically collective on PC and MatNullSpace

   Input Parameters:
+  pc - the preconditioning context
-  NullSpace - Null space of the linear operator to be preconditioned.

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetNullSpace(PC pc,MatNullSpace NullSpace)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidHeaderSpecific(NullSpace,MAT_NULLSPACE_CLASSID,2);
  PetscCheckSameComm(pc,1,NullSpace,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetNullSpace_C",(PC,MatNullSpace),(pc,NullSpace));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetDirichletBoundaries_BDDC"
static PetscErrorCode PCBDDCSetDirichletBoundaries_BDDC(PC pc,IS DirichletBoundaries)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = ISDestroy(&pcbddc->DirichletBoundaries);CHKERRQ(ierr);
  ierr = PetscObjectReference((PetscObject)DirichletBoundaries);CHKERRQ(ierr);
  pcbddc->DirichletBoundaries=DirichletBoundaries;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetDirichletBoundaries"
/*@
 PCBDDCSetDirichletBoundaries - Set index set defining subdomain part (in local ordering)
                              of Dirichlet boundaries for the global problem.

   Not collective

   Input Parameters:
+  pc - the preconditioning context
-  DirichletBoundaries - sequential index set defining the subdomain part of Dirichlet boundaries (can be NULL)

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetDirichletBoundaries(PC pc,IS DirichletBoundaries)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidHeaderSpecific(DirichletBoundaries,IS_CLASSID,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetDirichletBoundaries_C",(PC,IS),(pc,DirichletBoundaries));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetNeumannBoundaries_BDDC"
static PetscErrorCode PCBDDCSetNeumannBoundaries_BDDC(PC pc,IS NeumannBoundaries)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = ISDestroy(&pcbddc->NeumannBoundaries);CHKERRQ(ierr);
  ierr = PetscObjectReference((PetscObject)NeumannBoundaries);CHKERRQ(ierr);
  pcbddc->NeumannBoundaries=NeumannBoundaries;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetNeumannBoundaries"
/*@
 PCBDDCSetNeumannBoundaries - Set index set defining subdomain part (in local ordering)
                              of Neumann boundaries for the global problem.

   Not collective

   Input Parameters:
+  pc - the preconditioning context
-  NeumannBoundaries - sequential index set defining the subdomain part of Neumann boundaries (can be NULL)

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetNeumannBoundaries(PC pc,IS NeumannBoundaries)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidHeaderSpecific(NeumannBoundaries,IS_CLASSID,2);
  ierr = PetscTryMethod(pc,"PCBDDCSetNeumannBoundaries_C",(PC,IS),(pc,NeumannBoundaries));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCGetDirichletBoundaries_BDDC"
static PetscErrorCode PCBDDCGetDirichletBoundaries_BDDC(PC pc,IS *DirichletBoundaries)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;

  PetscFunctionBegin;
  *DirichletBoundaries = pcbddc->DirichletBoundaries;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCGetDirichletBoundaries"
/*@
 PCBDDCGetDirichletBoundaries - Get index set defining subdomain part (in local ordering)
                                of Dirichlet boundaries for the global problem.

   Not collective

   Input Parameters:
+  pc - the preconditioning context

   Output Parameters:
+  DirichletBoundaries - index set defining the subdomain part of Dirichlet boundaries

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCGetDirichletBoundaries(PC pc,IS *DirichletBoundaries)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  ierr = PetscUseMethod(pc,"PCBDDCGetDirichletBoundaries_C",(PC,IS*),(pc,DirichletBoundaries));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCGetNeumannBoundaries_BDDC"
static PetscErrorCode PCBDDCGetNeumannBoundaries_BDDC(PC pc,IS *NeumannBoundaries)
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;

  PetscFunctionBegin;
  *NeumannBoundaries = pcbddc->NeumannBoundaries;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCGetNeumannBoundaries"
/*@
 PCBDDCGetNeumannBoundaries - Get index set defining subdomain part (in local ordering)
                              of Neumann boundaries for the global problem.

   Not collective

   Input Parameters:
+  pc - the preconditioning context

   Output Parameters:
+  NeumannBoundaries - index set defining the subdomain part of Neumann boundaries

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCGetNeumannBoundaries(PC pc,IS *NeumannBoundaries)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  ierr = PetscUseMethod(pc,"PCBDDCGetNeumannBoundaries_C",(PC,IS*),(pc,NeumannBoundaries));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLocalAdjacencyGraph_BDDC"
static PetscErrorCode PCBDDCSetLocalAdjacencyGraph_BDDC(PC pc, PetscInt nvtxs,const PetscInt xadj[],const PetscInt adjncy[], PetscCopyMode copymode)
{
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PCBDDCGraph    mat_graph = pcbddc->mat_graph;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  /* free old CSR */
  ierr = PCBDDCGraphResetCSR(mat_graph);CHKERRQ(ierr);
  /* TODO: PCBDDCGraphSetAdjacency */
  /* get CSR into graph structure */
  if (copymode == PETSC_COPY_VALUES) {
    ierr = PetscMalloc((nvtxs+1)*sizeof(PetscInt),&mat_graph->xadj);CHKERRQ(ierr);
    ierr = PetscMalloc(xadj[nvtxs]*sizeof(PetscInt),&mat_graph->adjncy);CHKERRQ(ierr);
    ierr = PetscMemcpy(mat_graph->xadj,xadj,(nvtxs+1)*sizeof(PetscInt));CHKERRQ(ierr);
    ierr = PetscMemcpy(mat_graph->adjncy,adjncy,xadj[nvtxs]*sizeof(PetscInt));CHKERRQ(ierr);
  } else if (copymode == PETSC_OWN_POINTER) {
    mat_graph->xadj = (PetscInt*)xadj;
    mat_graph->adjncy = (PetscInt*)adjncy;
  }
  mat_graph->nvtxs_csr = nvtxs;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetLocalAdjacencyGraph"
/*@
 PCBDDCSetLocalAdjacencyGraph - Set CSR graph of local matrix for use of PCBDDC.

   Not collective

   Input Parameters:
+  pc - the preconditioning context
-  nvtxs - number of local vertices of the graph
-  xadj, adjncy - the CSR graph
-  copymode - either PETSC_COPY_VALUES or PETSC_OWN_POINTER. In the former case the user must free the array passed in;
                                                             in the latter case, memory must be obtained with PetscMalloc.

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetLocalAdjacencyGraph(PC pc,PetscInt nvtxs,const PetscInt xadj[],const PetscInt adjncy[], PetscCopyMode copymode)
{
  void (*f)(void) = 0;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  PetscValidIntPointer(xadj,3);
  PetscValidIntPointer(xadj,4);
  if (copymode != PETSC_COPY_VALUES && copymode != PETSC_OWN_POINTER) {
    SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"Unsupported copy mode %d in %s\n",copymode,__FUNCT__);
  }
  ierr = PetscTryMethod(pc,"PCBDDCSetLocalAdjacencyGraph_C",(PC,PetscInt,const PetscInt[],const PetscInt[],PetscCopyMode),(pc,nvtxs,xadj,adjncy,copymode));CHKERRQ(ierr);
  /* free arrays if PCBDDC is not the PC type */
  ierr = PetscObjectQueryFunction((PetscObject)pc,"PCBDDCSetLocalAdjacencyGraph_C",&f);CHKERRQ(ierr);
  if (!f && copymode == PETSC_OWN_POINTER) {
    ierr = PetscFree(xadj);CHKERRQ(ierr);
    ierr = PetscFree(adjncy);CHKERRQ(ierr);
  }
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetDofsSplitting_BDDC"
static PetscErrorCode PCBDDCSetDofsSplitting_BDDC(PC pc,PetscInt n_is, IS ISForDofs[])
{
  PC_BDDC  *pcbddc = (PC_BDDC*)pc->data;
  PetscInt i;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  /* Destroy ISes if they were already set */
  for (i=0;i<pcbddc->n_ISForDofs;i++) {
    ierr = ISDestroy(&pcbddc->ISForDofs[i]);CHKERRQ(ierr);
  }
  ierr = PetscFree(pcbddc->ISForDofs);CHKERRQ(ierr);
  /* allocate space then set */
  ierr = PetscMalloc(n_is*sizeof(IS),&pcbddc->ISForDofs);CHKERRQ(ierr);
  for (i=0;i<n_is;i++) {
    ierr = PetscObjectReference((PetscObject)ISForDofs[i]);CHKERRQ(ierr);
    pcbddc->ISForDofs[i]=ISForDofs[i];
  }
  pcbddc->n_ISForDofs=n_is;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCSetDofsSplitting"
/*@
 PCBDDCSetDofsSplitting - Set index sets defining fields of local mat.

   Not collective

   Input Parameters:
+  pc - the preconditioning context
-  n - number of index sets defining the fields
-  IS[] - array of IS describing the fields

   Level: intermediate

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCSetDofsSplitting(PC pc,PetscInt n_is, IS ISForDofs[])
{
  PetscInt       i;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  for (i=0;i<n_is;i++) {
    PetscValidHeaderSpecific(ISForDofs[i],IS_CLASSID,2);
  }
  ierr = PetscTryMethod(pc,"PCBDDCSetDofsSplitting_C",(PC,PetscInt,IS[]),(pc,n_is,ISForDofs));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCPreSolve_BDDC"
/* -------------------------------------------------------------------------- */
/*
   PCPreSolve_BDDC - Changes the right hand side and (if necessary) the initial
                     guess if a transformation of basis approach has been selected.

   Input Parameter:
+  pc - the preconditioner contex

   Application Interface Routine: PCPreSolve()

   Notes:
   The interface routine PCPreSolve() is not usually called directly by
   the user, but instead is called by KSPSolve().
*/
static PetscErrorCode PCPreSolve_BDDC(PC pc, KSP ksp, Vec rhs, Vec x)
{
  PetscErrorCode ierr;
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PC_IS          *pcis = (PC_IS*)(pc->data);
  Mat_IS         *matis = (Mat_IS*)pc->pmat->data;
  Mat            temp_mat;
  IS             dirIS;
  PetscInt       dirsize,i,*is_indices;
  PetscScalar    *array_x,*array_diagonal;
  Vec            used_vec;
  PetscBool      guess_nonzero;

  PetscFunctionBegin;
  /* Creates parallel work vectors used in presolve. */
  if (!pcbddc->original_rhs) {
    ierr = VecDuplicate(pcis->vec1_global,&pcbddc->original_rhs);CHKERRQ(ierr);
  }
  if (!pcbddc->temp_solution) {
    ierr = VecDuplicate(pcis->vec1_global,&pcbddc->temp_solution);CHKERRQ(ierr);
  }
  if (x) {
    ierr = PetscObjectReference((PetscObject)x);CHKERRQ(ierr);
    used_vec = x;
  } else {
    ierr = PetscObjectReference((PetscObject)pcbddc->temp_solution);CHKERRQ(ierr);
    used_vec = pcbddc->temp_solution;
    ierr = VecSet(used_vec,0.0);CHKERRQ(ierr);
  }
  /* hack into ksp data structure PCPreSolve comes earlier in src/ksp/ksp/interface/itfunc.c */
  if (ksp) {
    ierr = KSPGetInitialGuessNonzero(ksp,&guess_nonzero);CHKERRQ(ierr);
    if ( !guess_nonzero ) {
      ierr = VecSet(used_vec,0.0);CHKERRQ(ierr);
    }
  }

  if (rhs) { /* TODO: wiser handling of rhs removal, which is only needed in case of zeroed rows */
    /* store the original rhs */
    ierr = VecCopy(rhs,pcbddc->original_rhs);CHKERRQ(ierr);

    /* Take into account zeroed rows -> change rhs and store solution removed */
    ierr = MatGetDiagonal(pc->pmat,pcis->vec1_global);CHKERRQ(ierr);
    ierr = VecPointwiseDivide(pcis->vec1_global,rhs,pcis->vec1_global);CHKERRQ(ierr);
    ierr = VecScatterBegin(matis->ctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd(matis->ctx,pcis->vec1_global,pcis->vec2_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterBegin(matis->ctx,used_vec,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd(matis->ctx,used_vec,pcis->vec1_N,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = PCBDDCGetDirichletBoundaries(pc,&dirIS);CHKERRQ(ierr);
    if (dirIS) {
      ierr = ISGetSize(dirIS,&dirsize);CHKERRQ(ierr);
      ierr = VecGetArray(pcis->vec1_N,&array_x);CHKERRQ(ierr);
      ierr = VecGetArray(pcis->vec2_N,&array_diagonal);CHKERRQ(ierr);
      ierr = ISGetIndices(dirIS,(const PetscInt**)&is_indices);CHKERRQ(ierr);
      for (i=0; i<dirsize; i++) array_x[is_indices[i]] = array_diagonal[is_indices[i]];
      ierr = ISRestoreIndices(dirIS,(const PetscInt**)&is_indices);CHKERRQ(ierr);
      ierr = VecRestoreArray(pcis->vec2_N,&array_diagonal);CHKERRQ(ierr);
      ierr = VecRestoreArray(pcis->vec1_N,&array_x);CHKERRQ(ierr);
    }
    ierr = VecScatterBegin(matis->ctx,pcis->vec1_N,used_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = VecScatterEnd(matis->ctx,pcis->vec1_N,used_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);

    /* remove the computed solution from the rhs */
    ierr = VecScale(used_vec,-1.0);CHKERRQ(ierr);
    ierr = MatMultAdd(pc->pmat,used_vec,rhs,rhs);CHKERRQ(ierr);
    ierr = VecScale(used_vec,-1.0);CHKERRQ(ierr);
  }

  /* store partially computed solution and set initial guess */
  if (x) {
    ierr = VecCopy(used_vec,pcbddc->temp_solution);CHKERRQ(ierr);
    ierr = VecSet(used_vec,0.0);CHKERRQ(ierr);
    if (pcbddc->use_exact_dirichlet && !pcbddc->coarse_psi_B) {
      ierr = VecScatterBegin(pcis->global_to_D,rhs,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
      ierr = VecScatterEnd  (pcis->global_to_D,rhs,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
      ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr);
      ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,used_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
      ierr = VecScatterEnd  (pcis->global_to_D,pcis->vec2_D,used_vec,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
      if (ksp) {
        ierr = KSPSetInitialGuessNonzero(ksp,PETSC_TRUE);CHKERRQ(ierr);
      }
    }
  }

  if (pcbddc->use_change_of_basis) {
    /* swap pointers for local matrices */
    temp_mat = matis->A;
    matis->A = pcbddc->local_mat;
    pcbddc->local_mat = temp_mat;
  }
  if (pcbddc->use_change_of_basis && rhs) {
    /* Get local rhs and apply transformation of basis */
    ierr = VecScatterBegin(pcis->global_to_B,rhs,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_B,rhs,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    /* from original basis to modified basis */
    ierr = MatMultTranspose(pcbddc->ChangeOfBasisMatrix,pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr);
    /* put back modified values into the global vec using INSERT_VALUES copy mode */
    ierr = VecScatterBegin(pcis->global_to_B,pcis->vec2_B,rhs,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_B,pcis->vec2_B,rhs,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  }
  if (ksp && pcbddc->NullSpace) {
    ierr = MatNullSpaceRemove(pcbddc->NullSpace,used_vec);CHKERRQ(ierr);
    ierr = MatNullSpaceRemove(pcbddc->NullSpace,rhs);CHKERRQ(ierr);
  }
  ierr = VecDestroy(&used_vec);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCPostSolve_BDDC"
/* -------------------------------------------------------------------------- */
/*
   PCPostSolve_BDDC - Changes the computed solution if a transformation of basis
                     approach has been selected. Also, restores rhs to its original state.

   Input Parameter:
+  pc - the preconditioner contex

   Application Interface Routine: PCPostSolve()

   Notes:
   The interface routine PCPostSolve() is not usually called directly by
   the user, but instead is called by KSPSolve().
*/
static PetscErrorCode PCPostSolve_BDDC(PC pc, KSP ksp, Vec rhs, Vec x)
{
  PetscErrorCode ierr;
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PC_IS          *pcis   = (PC_IS*)(pc->data);
  Mat_IS         *matis = (Mat_IS*)pc->pmat->data;
  Mat            temp_mat;

  PetscFunctionBegin;
  if (pcbddc->use_change_of_basis) {
    /* swap pointers for local matrices */
    temp_mat = matis->A;
    matis->A = pcbddc->local_mat;
    pcbddc->local_mat = temp_mat;
  }
  if (pcbddc->use_change_of_basis && x) {
    /* Get Local boundary and apply transformation of basis to solution vector */
    ierr = VecScatterBegin(pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_B,x,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    /* from modified basis to original basis */
    ierr = MatMult(pcbddc->ChangeOfBasisMatrix,pcis->vec1_B,pcis->vec2_B);CHKERRQ(ierr);
    /* put back modified values into the global vec using INSERT_VALUES copy mode */
    ierr = VecScatterBegin(pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_B,pcis->vec2_B,x,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  }
  /* add solution removed in presolve */
  if (x) {
    ierr = VecAXPY(x,1.0,pcbddc->temp_solution);CHKERRQ(ierr);
  }
  /* restore rhs to its original state */
  if (rhs) {
    ierr = VecCopy(pcbddc->original_rhs,rhs);CHKERRQ(ierr);
  }
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
#undef __FUNCT__
#define __FUNCT__ "PCSetUp_BDDC"
/* -------------------------------------------------------------------------- */
/*
   PCSetUp_BDDC - Prepares for the use of the BDDC preconditioner
                  by setting data structures and options.

   Input Parameter:
+  pc - the preconditioner context

   Application Interface Routine: PCSetUp()

   Notes:
   The interface routine PCSetUp() is not usually called directly by
   the user, but instead is called by PCApply() if necessary.
*/
PetscErrorCode PCSetUp_BDDC(PC pc)
{
  PetscErrorCode ierr;
  PC_BDDC*       pcbddc = (PC_BDDC*)pc->data;
  MatStructure   flag;
  PetscBool      computeis,computetopography,computesolvers;

  PetscFunctionBegin;
  /* the following lines of code should be replaced by a better logic between PCIS, PCNN, PCBDDC and other nonoverlapping preconditioners */
  /* For BDDC we need to define a local "Neumann" problem different to that defined in PCISSetup
     So, we set to pcnone the Neumann problem of pcis in order to avoid unneeded computation
     Also, we decide to directly build the (same) Dirichlet problem */
  ierr = PetscOptionsSetValue("-is_localN_pc_type","none");CHKERRQ(ierr);
  ierr = PetscOptionsSetValue("-is_localD_pc_type","none");CHKERRQ(ierr);
  /* Get stdout for dbg */
  if (pcbddc->dbg_flag && !pcbddc->dbg_viewer) {
    ierr = PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)pc),&pcbddc->dbg_viewer);CHKERRQ(ierr);
    ierr = PetscViewerASCIISynchronizedAllow(pcbddc->dbg_viewer,PETSC_TRUE);CHKERRQ(ierr);
    if (pcbddc->current_level) {
      ierr = PetscViewerASCIIAddTab(pcbddc->dbg_viewer,2);CHKERRQ(ierr);
    }
  }
  /* first attempt to split work */
  if (pc->setupcalled) {
    computeis = PETSC_FALSE;
    ierr = PCGetOperators(pc,NULL,NULL,&flag);CHKERRQ(ierr);
    if (flag == SAME_PRECONDITIONER) {
      computetopography = PETSC_FALSE;
      computesolvers = PETSC_FALSE;
    } else if (flag == SAME_NONZERO_PATTERN) {
      computetopography = PETSC_FALSE;
      computesolvers = PETSC_TRUE;
    } else { /* DIFFERENT_NONZERO_PATTERN */
      computetopography = PETSC_TRUE;
      computesolvers = PETSC_TRUE;
    }
  } else {
    computeis = PETSC_TRUE;
    computetopography = PETSC_TRUE;
    computesolvers = PETSC_TRUE;
  }
  /* Set up all the "iterative substructuring" common block */
  if (computeis) {
    ierr = PCISSetUp(pc);CHKERRQ(ierr);
  }
  /* Analyze interface and set up local constraint and change of basis matrices */
  if (computetopography) {
    /* reset data */
    ierr = PCBDDCResetTopography(pc);CHKERRQ(ierr);
    ierr = PCBDDCAnalyzeInterface(pc);CHKERRQ(ierr);
    ierr = PCBDDCConstraintsSetUp(pc);CHKERRQ(ierr);
  }
  if (computesolvers) {
    /* reset data */
    ierr = PCBDDCResetSolvers(pc);CHKERRQ(ierr);
    ierr = PCBDDCScalingDestroy(pc);CHKERRQ(ierr);
    /* Create coarse and local stuffs */
    ierr = PCBDDCSetUpSolvers(pc);CHKERRQ(ierr);
    ierr = PCBDDCScalingSetUp(pc);CHKERRQ(ierr);
  }
  if (pcbddc->dbg_flag && pcbddc->current_level) {
    ierr = PetscViewerASCIISubtractTab(pcbddc->dbg_viewer,2);CHKERRQ(ierr);
  }
  PetscFunctionReturn(0);
}

/* -------------------------------------------------------------------------- */
/*
   PCApply_BDDC - Applies the BDDC preconditioner to a vector.

   Input Parameters:
.  pc - the preconditioner context
.  r - input vector (global)

   Output Parameter:
.  z - output vector (global)

   Application Interface Routine: PCApply()
 */
#undef __FUNCT__
#define __FUNCT__ "PCApply_BDDC"
PetscErrorCode PCApply_BDDC(PC pc,Vec r,Vec z)
{
  PC_IS             *pcis = (PC_IS*)(pc->data);
  PC_BDDC           *pcbddc = (PC_BDDC*)(pc->data);
  PetscErrorCode    ierr;
  const PetscScalar one = 1.0;
  const PetscScalar m_one = -1.0;
  const PetscScalar zero = 0.0;

/* This code is similar to that provided in nn.c for PCNN
   NN interface preconditioner changed to BDDC
   Added support for M_3 preconditioner in the reference article (code is active if pcbddc->switch_static = PETSC_TRUE) */

  PetscFunctionBegin;
  if (!pcbddc->use_exact_dirichlet || pcbddc->coarse_psi_B) {
    /* First Dirichlet solve */
    ierr = VecScatterBegin(pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_D,r,pcis->vec1_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = KSPSolve(pcbddc->ksp_D,pcis->vec1_D,pcis->vec2_D);CHKERRQ(ierr);
    /*
      Assembling right hand side for BDDC operator
      - pcis->vec1_D for the Dirichlet part (if needed, i.e. prec_flag=PETSC_TRUE)
      - pcis->vec1_B the interface part of the global vector z
    */
    ierr = VecScale(pcis->vec2_D,m_one);CHKERRQ(ierr);
    ierr = MatMult(pcis->A_BI,pcis->vec2_D,pcis->vec1_B);CHKERRQ(ierr);
    if (pcbddc->switch_static) { ierr = MatMultAdd(pcis->A_II,pcis->vec2_D,pcis->vec1_D,pcis->vec1_D);CHKERRQ(ierr); }
    ierr = VecScale(pcis->vec2_D,m_one);CHKERRQ(ierr);
    ierr = VecCopy(r,z);CHKERRQ(ierr);
    ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,z,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = VecScatterEnd  (pcis->global_to_B,pcis->vec1_B,z,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = PCBDDCScalingRestriction(pc,z,pcis->vec1_B);CHKERRQ(ierr);
  } else {
    ierr = VecSet(pcis->vec1_D,zero);CHKERRQ(ierr);
    ierr = VecSet(pcis->vec2_D,zero);CHKERRQ(ierr);
    ierr = PCBDDCScalingRestriction(pc,r,pcis->vec1_B);CHKERRQ(ierr);
  }

  /* Apply interface preconditioner
     input/output vecs: pcis->vec1_B and pcis->vec1_D */
  ierr = PCBDDCApplyInterfacePreconditioner(pc);CHKERRQ(ierr);

  /* Apply transpose of partition of unity operator */
  ierr = PCBDDCScalingExtension(pc,pcis->vec1_B,z);CHKERRQ(ierr);

  /* Second Dirichlet solve and assembling of output */
  ierr = VecScatterBegin(pcis->global_to_B,z,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = VecScatterEnd  (pcis->global_to_B,z,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = MatMult(pcis->A_IB,pcis->vec1_B,pcis->vec3_D);CHKERRQ(ierr);
  if (pcbddc->switch_static) { ierr = MatMultAdd(pcis->A_II,pcis->vec1_D,pcis->vec3_D,pcis->vec3_D);CHKERRQ(ierr); }
  ierr = KSPSolve(pcbddc->ksp_D,pcis->vec3_D,pcbddc->vec4_D);CHKERRQ(ierr);
  ierr = VecScale(pcbddc->vec4_D,m_one);CHKERRQ(ierr);
  if (pcbddc->switch_static) { ierr = VecAXPY (pcbddc->vec4_D,one,pcis->vec1_D);CHKERRQ(ierr); }
  ierr = VecAXPY (pcis->vec2_D,one,pcbddc->vec4_D);CHKERRQ(ierr);
  ierr = VecScatterBegin(pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  ierr = VecScatterEnd  (pcis->global_to_D,pcis->vec2_D,z,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCDestroy_BDDC"
PetscErrorCode PCDestroy_BDDC(PC pc)
{
  PC_BDDC        *pcbddc = (PC_BDDC*)pc->data;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  /* free data created by PCIS */
  ierr = PCISDestroy(pc);CHKERRQ(ierr);
  /* free BDDC custom data  */
  ierr = PCBDDCResetCustomization(pc);CHKERRQ(ierr);
  /* destroy objects related to topography */
  ierr = PCBDDCResetTopography(pc);CHKERRQ(ierr);
  /* free allocated graph structure */
  ierr = PetscFree(pcbddc->mat_graph);CHKERRQ(ierr);
  /* free data for scaling operator */
  ierr = PCBDDCScalingDestroy(pc);CHKERRQ(ierr);
  /* free solvers stuff */
  ierr = PCBDDCResetSolvers(pc);CHKERRQ(ierr);
  ierr = KSPDestroy(&pcbddc->ksp_D);CHKERRQ(ierr);
  ierr = KSPDestroy(&pcbddc->ksp_R);CHKERRQ(ierr);
  ierr = KSPDestroy(&pcbddc->coarse_ksp);CHKERRQ(ierr);
  ierr = MatDestroy(&pcbddc->local_mat);CHKERRQ(ierr);
  /* free global vectors needed in presolve */
  ierr = VecDestroy(&pcbddc->temp_solution);CHKERRQ(ierr);
  ierr = VecDestroy(&pcbddc->original_rhs);CHKERRQ(ierr);
  /* remove functions */
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetPrimalVerticesLocalIS_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetCoarseningRatio_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLevel_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLevels_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetNullSpace_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetDirichletBoundaries_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetNeumannBoundaries_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCGetDirichletBoundaries_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCGetNeumannBoundaries_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetDofsSplitting_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLocalAdjacencyGraph_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCCreateFETIDPOperators_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCMatFETIDPGetRHS_C",NULL);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCMatFETIDPGetSolution_C",NULL);CHKERRQ(ierr);
  /* Free the private data structure */
  ierr = PetscFree(pc->data);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCMatFETIDPGetRHS_BDDC"
static PetscErrorCode PCBDDCMatFETIDPGetRHS_BDDC(Mat fetidp_mat, Vec standard_rhs, Vec fetidp_flux_rhs)
{
  FETIDPMat_ctx  mat_ctx;
  PC_IS*         pcis;
  PC_BDDC*       pcbddc;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = MatShellGetContext(fetidp_mat,&mat_ctx);CHKERRQ(ierr);
  pcis = (PC_IS*)mat_ctx->pc->data;
  pcbddc = (PC_BDDC*)mat_ctx->pc->data;

  /* change of basis for physical rhs if needed
     It also changes the rhs in case of dirichlet boundaries */
  ierr = PCPreSolve_BDDC(mat_ctx->pc,NULL,standard_rhs,NULL);CHKERRQ(ierr);
  /* store vectors for computation of fetidp final solution */
  ierr = VecScatterBegin(pcis->global_to_D,standard_rhs,mat_ctx->temp_solution_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = VecScatterEnd(pcis->global_to_D,standard_rhs,mat_ctx->temp_solution_D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  /* scale rhs since it should be unassembled */
  /* TODO use counter scaling? (also below) */
  ierr = VecScatterBegin(pcis->global_to_B,standard_rhs,mat_ctx->temp_solution_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = VecScatterEnd(pcis->global_to_B,standard_rhs,mat_ctx->temp_solution_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  /* Apply partition of unity */
  ierr = VecPointwiseMult(mat_ctx->temp_solution_B,pcis->D,mat_ctx->temp_solution_B);CHKERRQ(ierr);
  /* ierr = PCBDDCScalingRestriction(mat_ctx->pc,standard_rhs,mat_ctx->temp_solution_B);CHKERRQ(ierr); */
  if (!pcbddc->switch_static) {
    /* compute partially subassembled Schur complement right-hand side */
    ierr = KSPSolve(pcbddc->ksp_D,mat_ctx->temp_solution_D,pcis->vec1_D);CHKERRQ(ierr);
    ierr = MatMult(pcis->A_BI,pcis->vec1_D,pcis->vec1_B);CHKERRQ(ierr);
    ierr = VecAXPY(mat_ctx->temp_solution_B,-1.0,pcis->vec1_B);CHKERRQ(ierr);
    ierr = VecSet(standard_rhs,0.0);CHKERRQ(ierr);
    ierr = VecScatterBegin(pcis->global_to_B,mat_ctx->temp_solution_B,standard_rhs,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    ierr = VecScatterEnd(pcis->global_to_B,mat_ctx->temp_solution_B,standard_rhs,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
    /* ierr = PCBDDCScalingRestriction(mat_ctx->pc,standard_rhs,mat_ctx->temp_solution_B);CHKERRQ(ierr); */
    ierr = VecScatterBegin(pcis->global_to_B,standard_rhs,mat_ctx->temp_solution_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecScatterEnd(pcis->global_to_B,standard_rhs,mat_ctx->temp_solution_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
    ierr = VecPointwiseMult(mat_ctx->temp_solution_B,pcis->D,mat_ctx->temp_solution_B);CHKERRQ(ierr);
  }
  /* BDDC rhs */
  ierr = VecCopy(mat_ctx->temp_solution_B,pcis->vec1_B);CHKERRQ(ierr);
  if (pcbddc->switch_static) {
    ierr = VecCopy(mat_ctx->temp_solution_D,pcis->vec1_D);CHKERRQ(ierr);
  }
  /* apply BDDC */
  ierr = PCBDDCApplyInterfacePreconditioner(mat_ctx->pc);CHKERRQ(ierr);
  /* Application of B_delta and assembling of rhs for fetidp fluxes */
  ierr = VecSet(fetidp_flux_rhs,0.0);CHKERRQ(ierr);
  ierr = MatMult(mat_ctx->B_delta,pcis->vec1_B,mat_ctx->lambda_local);CHKERRQ(ierr);
  ierr = VecScatterBegin(mat_ctx->l2g_lambda,mat_ctx->lambda_local,fetidp_flux_rhs,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  ierr = VecScatterEnd  (mat_ctx->l2g_lambda,mat_ctx->lambda_local,fetidp_flux_rhs,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
  /* restore original rhs */
  ierr = VecCopy(pcbddc->original_rhs,standard_rhs);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCMatFETIDPGetRHS"
/*@
 PCBDDCMatFETIDPGetRHS - Get rhs for FETIDP linear system.

   Collective

   Input Parameters:
+  fetidp_mat   - the FETIDP mat obtained by a call to PCBDDCCreateFETIDPOperators
+  standard_rhs - the rhs of your linear system

   Output Parameters:
+  fetidp_flux_rhs   - the rhs of the FETIDP linear system

   Level: developer

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCMatFETIDPGetRHS(Mat fetidp_mat, Vec standard_rhs, Vec fetidp_flux_rhs)
{
  FETIDPMat_ctx  mat_ctx;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = MatShellGetContext(fetidp_mat,&mat_ctx);CHKERRQ(ierr);
  ierr = PetscTryMethod(mat_ctx->pc,"PCBDDCMatFETIDPGetRHS_C",(Mat,Vec,Vec),(fetidp_mat,standard_rhs,fetidp_flux_rhs));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

#undef __FUNCT__
#define __FUNCT__ "PCBDDCMatFETIDPGetSolution_BDDC"
static PetscErrorCode PCBDDCMatFETIDPGetSolution_BDDC(Mat fetidp_mat, Vec fetidp_flux_sol, Vec standard_sol)
{
  FETIDPMat_ctx  mat_ctx;
  PC_IS*         pcis;
  PC_BDDC*       pcbddc;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = MatShellGetContext(fetidp_mat,&mat_ctx);CHKERRQ(ierr);
  pcis = (PC_IS*)mat_ctx->pc->data;
  pcbddc = (PC_BDDC*)mat_ctx->pc->data;

  /* apply B_delta^T */
  ierr = VecScatterBegin(mat_ctx->l2g_lambda,fetidp_flux_sol,mat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  ierr = VecScatterEnd  (mat_ctx->l2g_lambda,fetidp_flux_sol,mat_ctx->lambda_local,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  ierr = MatMultTranspose(mat_ctx->B_delta,mat_ctx->lambda_local,pcis->vec1_B);CHKERRQ(ierr);
  /* compute rhs for BDDC application */
  ierr = VecAYPX(pcis->vec1_B,-1.0,mat_ctx->temp_solution_B);CHKERRQ(ierr);
  if (pcbddc->switch_static) {
    ierr = VecCopy(mat_ctx->temp_solution_D,pcis->vec1_D);CHKERRQ(ierr);
  }
  /* apply BDDC */
  ierr = PCBDDCApplyInterfacePreconditioner(mat_ctx->pc);CHKERRQ(ierr);
  /* put values into standard global vector */
  ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,standard_sol,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  ierr = VecScatterEnd  (pcis->global_to_B,pcis->vec1_B,standard_sol,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  if (!pcbddc->switch_static) {
    /* compute values into the interior if solved for the partially subassembled Schur complement */
    ierr = MatMult(pcis->A_IB,pcis->vec1_B,pcis->vec1_D);CHKERRQ(ierr);
    ierr = VecAXPY(mat_ctx->temp_solution_D,-1.0,pcis->vec1_D);CHKERRQ(ierr);
    ierr = KSPSolve(pcbddc->ksp_D,mat_ctx->temp_solution_D,pcis->vec1_D);CHKERRQ(ierr);
  }
  ierr = VecScatterBegin(pcis->global_to_D,pcis->vec1_D,standard_sol,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  ierr = VecScatterEnd  (pcis->global_to_D,pcis->vec1_D,standard_sol,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
  /* final change of basis if needed
     Is also sums the dirichlet part removed during RHS assembling */
  ierr = PCPostSolve_BDDC(mat_ctx->pc,NULL,NULL,standard_sol);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCMatFETIDPGetSolution"
/*@
 PCBDDCMatFETIDPGetSolution - Get Solution for FETIDP linear system.

   Collective

   Input Parameters:
+  fetidp_mat        - the FETIDP mat obtained by a call to PCBDDCCreateFETIDPOperators
+  fetidp_flux_sol - the solution of the FETIDP linear system

   Output Parameters:
+  standard_sol      - the solution on the global domain

   Level: developer

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCMatFETIDPGetSolution(Mat fetidp_mat, Vec fetidp_flux_sol, Vec standard_sol)
{
  FETIDPMat_ctx  mat_ctx;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = MatShellGetContext(fetidp_mat,&mat_ctx);CHKERRQ(ierr);
  ierr = PetscTryMethod(mat_ctx->pc,"PCBDDCMatFETIDPGetSolution_C",(Mat,Vec,Vec),(fetidp_mat,fetidp_flux_sol,standard_sol));CHKERRQ(ierr);
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */

extern PetscErrorCode FETIDPMatMult(Mat,Vec,Vec);
extern PetscErrorCode PCBDDCDestroyFETIDPMat(Mat);
extern PetscErrorCode FETIDPPCApply(PC,Vec,Vec);
extern PetscErrorCode PCBDDCDestroyFETIDPPC(PC);

#undef __FUNCT__
#define __FUNCT__ "PCBDDCCreateFETIDPOperators_BDDC"
static PetscErrorCode PCBDDCCreateFETIDPOperators_BDDC(PC pc, Mat *fetidp_mat, PC *fetidp_pc)
{

  FETIDPMat_ctx  fetidpmat_ctx;
  Mat            newmat;
  FETIDPPC_ctx   fetidppc_ctx;
  PC             newpc;
  MPI_Comm       comm;
  PetscErrorCode ierr;

  PetscFunctionBegin;
  ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr);
  /* FETIDP linear matrix */
  ierr = PCBDDCCreateFETIDPMatContext(pc,&fetidpmat_ctx);CHKERRQ(ierr);
  ierr = PCBDDCSetupFETIDPMatContext(fetidpmat_ctx);CHKERRQ(ierr);
  ierr = MatCreateShell(comm,PETSC_DECIDE,PETSC_DECIDE,fetidpmat_ctx->n_lambda,fetidpmat_ctx->n_lambda,fetidpmat_ctx,&newmat);CHKERRQ(ierr);
  ierr = MatShellSetOperation(newmat,MATOP_MULT,(void (*)(void))FETIDPMatMult);CHKERRQ(ierr);
  ierr = MatShellSetOperation(newmat,MATOP_DESTROY,(void (*)(void))PCBDDCDestroyFETIDPMat);CHKERRQ(ierr);
  ierr = MatSetUp(newmat);CHKERRQ(ierr);
  /* FETIDP preconditioner */
  ierr = PCBDDCCreateFETIDPPCContext(pc,&fetidppc_ctx);CHKERRQ(ierr);
  ierr = PCBDDCSetupFETIDPPCContext(newmat,fetidppc_ctx);CHKERRQ(ierr);
  ierr = PCCreate(comm,&newpc);CHKERRQ(ierr);
  ierr = PCSetType(newpc,PCSHELL);CHKERRQ(ierr);
  ierr = PCShellSetContext(newpc,fetidppc_ctx);CHKERRQ(ierr);
  ierr = PCShellSetApply(newpc,FETIDPPCApply);CHKERRQ(ierr);
  ierr = PCShellSetDestroy(newpc,PCBDDCDestroyFETIDPPC);CHKERRQ(ierr);
  ierr = PCSetOperators(newpc,newmat,newmat,SAME_PRECONDITIONER);CHKERRQ(ierr);
  ierr = PCSetUp(newpc);CHKERRQ(ierr);
  /* return pointers for objects created */
  *fetidp_mat=newmat;
  *fetidp_pc=newpc;
  PetscFunctionReturn(0);
}

#undef __FUNCT__
#define __FUNCT__ "PCBDDCCreateFETIDPOperators"
/*@
 PCBDDCCreateFETIDPOperators - Create operators for FETIDP.

   Collective

   Input Parameters:
+  pc - the BDDC preconditioning context (setup must be already called)

   Level: developer

   Notes:

.seealso: PCBDDC
@*/
PetscErrorCode PCBDDCCreateFETIDPOperators(PC pc, Mat *fetidp_mat, PC *fetidp_pc)
{
  PetscErrorCode ierr;

  PetscFunctionBegin;
  PetscValidHeaderSpecific(pc,PC_CLASSID,1);
  if (pc->setupcalled) {
    ierr = PetscUseMethod(pc,"PCBDDCCreateFETIDPOperators_C",(PC,Mat*,PC*),(pc,fetidp_mat,fetidp_pc));CHKERRQ(ierr);
  } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"You must call PCSetup_BDDC() first \n");
  PetscFunctionReturn(0);
}
/* -------------------------------------------------------------------------- */
/*MC
   PCBDDC - Balancing Domain Decomposition by Constraints.

   Options Database Keys:
.    -pcbddc ??? -

   Level: intermediate

   Notes: The matrix used with this preconditioner must be of type MATIS

          Unlike more 'conventional' interface preconditioners, this iterates over ALL the
          degrees of freedom, NOT just those on the interface (this allows the use of approximate solvers
          on the subdomains).

          Options for the coarse grid preconditioner can be set with -
          Options for the Dirichlet subproblem can be set with -
          Options for the Neumann subproblem can be set with -

   Contributed by Stefano Zampini

.seealso:  PCCreate(), PCSetType(), PCType (for list of available types), PC,  MATIS
M*/

#undef __FUNCT__
#define __FUNCT__ "PCCreate_BDDC"
PETSC_EXTERN PetscErrorCode PCCreate_BDDC(PC pc)
{
  PetscErrorCode      ierr;
  PC_BDDC             *pcbddc;

  PetscFunctionBegin;
  /* Creates the private data structure for this preconditioner and attach it to the PC object. */
  ierr      = PetscNewLog(pc,PC_BDDC,&pcbddc);CHKERRQ(ierr);
  pc->data  = (void*)pcbddc;

  /* create PCIS data structure */
  ierr = PCISCreate(pc);CHKERRQ(ierr);

  /* BDDC customization */
  pcbddc->use_vertices        = PETSC_TRUE;
  pcbddc->use_edges           = PETSC_TRUE;
  pcbddc->use_faces           = PETSC_FALSE;
  pcbddc->use_change_of_basis = PETSC_FALSE;
  pcbddc->use_change_on_faces = PETSC_FALSE;
  pcbddc->switch_static       = PETSC_FALSE;
  pcbddc->use_nnsp_true       = PETSC_FALSE; /* not yet exposed */
  pcbddc->dbg_flag            = 0;

  pcbddc->user_primal_vertices       = 0;
  pcbddc->NullSpace                  = 0;
  pcbddc->temp_solution              = 0;
  pcbddc->original_rhs               = 0;
  pcbddc->local_mat                  = 0;
  pcbddc->ChangeOfBasisMatrix        = 0;
  pcbddc->coarse_vec                 = 0;
  pcbddc->coarse_rhs                 = 0;
  pcbddc->coarse_ksp                 = 0;
  pcbddc->coarse_phi_B               = 0;
  pcbddc->coarse_phi_D               = 0;
  pcbddc->coarse_psi_B               = 0;
  pcbddc->coarse_psi_D               = 0;
  pcbddc->vec1_P                     = 0;
  pcbddc->vec1_R                     = 0;
  pcbddc->vec2_R                     = 0;
  pcbddc->local_auxmat1              = 0;
  pcbddc->local_auxmat2              = 0;
  pcbddc->R_to_B                     = 0;
  pcbddc->R_to_D                     = 0;
  pcbddc->ksp_D                      = 0;
  pcbddc->ksp_R                      = 0;
  pcbddc->NeumannBoundaries          = 0;
  pcbddc->ISForDofs                  = 0;
  pcbddc->ConstraintMatrix           = 0;
  pcbddc->use_exact_dirichlet        = PETSC_TRUE;
  pcbddc->coarse_loc_to_glob         = 0;
  pcbddc->coarsening_ratio           = 8;
  pcbddc->current_level              = 0;
  pcbddc->max_levels                 = 0;

  /* create local graph structure */
  ierr = PCBDDCGraphCreate(&pcbddc->mat_graph);CHKERRQ(ierr);

  /* scaling */
  pcbddc->use_deluxe_scaling         = PETSC_FALSE;
  pcbddc->work_scaling               = 0;

  /* function pointers */
  pc->ops->apply               = PCApply_BDDC;
  pc->ops->applytranspose      = 0;
  pc->ops->setup               = PCSetUp_BDDC;
  pc->ops->destroy             = PCDestroy_BDDC;
  pc->ops->setfromoptions      = PCSetFromOptions_BDDC;
  pc->ops->view                = 0;
  pc->ops->applyrichardson     = 0;
  pc->ops->applysymmetricleft  = 0;
  pc->ops->applysymmetricright = 0;
  pc->ops->presolve            = PCPreSolve_BDDC;
  pc->ops->postsolve           = PCPostSolve_BDDC;

  /* composing function */
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetPrimalVerticesLocalIS_C",PCBDDCSetPrimalVerticesLocalIS_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetCoarseningRatio_C",PCBDDCSetCoarseningRatio_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLevel_C",PCBDDCSetLevel_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLevels_C",PCBDDCSetLevels_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetNullSpace_C",PCBDDCSetNullSpace_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetDirichletBoundaries_C",PCBDDCSetDirichletBoundaries_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetNeumannBoundaries_C",PCBDDCSetNeumannBoundaries_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCGetDirichletBoundaries_C",PCBDDCGetDirichletBoundaries_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCGetNeumannBoundaries_C",PCBDDCGetNeumannBoundaries_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetDofsSplitting_C",PCBDDCSetDofsSplitting_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCSetLocalAdjacencyGraph_C",PCBDDCSetLocalAdjacencyGraph_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCCreateFETIDPOperators_C",PCBDDCCreateFETIDPOperators_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCMatFETIDPGetRHS_C",PCBDDCMatFETIDPGetRHS_BDDC);CHKERRQ(ierr);
  ierr = PetscObjectComposeFunction((PetscObject)pc,"PCBDDCMatFETIDPGetSolution_C",PCBDDCMatFETIDPGetSolution_BDDC);CHKERRQ(ierr);
  PetscFunctionReturn(0);
}