int main(int argc,char **args)
{
Mat C,A;
PetscInt i,j,m = 3,n = 2,rstart,rend;
PetscMPIInt size,rank;
PetscErrorCode ierr;
PetscScalar v;
IS isrow,iscol;
ierr = PetscInitialize(&argc,&args,(char*)0,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
n = 2*size;
ierr = MatCreate(PETSC_COMM_WORLD,&C);CHKERRQ(ierr);
ierr = MatSetSizes(C,PETSC_DECIDE,PETSC_DECIDE,m*n,m*n);CHKERRQ(ierr);
ierr = MatSetFromOptions(C);CHKERRQ(ierr);
ierr = MatSetUp(C);CHKERRQ(ierr);
/*
This is JUST to generate a nice test matrix, all processors fill up
the entire matrix. This is not something one would ever do in practice.
*/
ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr);
for (i=rstart; i<rend; i++) {
for (j=0; j<m*n; j++) {
v = i + j + 1;
ierr = MatSetValues(C,1,&i,1,&j,&v,INSERT_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_WORLD,PETSC_VIEWER_ASCII_COMMON);CHKERRQ(ierr);
ierr = MatView(C,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
/*
Generate a new matrix consisting of every second row and column of
the original matrix
*/
ierr = MatGetOwnershipRange(C,&rstart,&rend);CHKERRQ(ierr);
/* Create parallel IS with the rows we want on THIS processor */
ierr = ISCreateStride(PETSC_COMM_WORLD,(rend-rstart)/2,rstart,2,&isrow);CHKERRQ(ierr);
/* Create parallel IS with the rows we want on THIS processor (same as rows for now) */
ierr = ISCreateStride(PETSC_COMM_WORLD,(rend-rstart)/2,rstart,2,&iscol);CHKERRQ(ierr);
ierr = MatGetSubMatrix(C,isrow,iscol,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr);
ierr = MatGetSubMatrix(C,isrow,iscol,MAT_REUSE_MATRIX,&A);CHKERRQ(ierr);
ierr = MatView(A,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISDestroy(&isrow);CHKERRQ(ierr);
ierr = ISDestroy(&iscol);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
void PetscMatrix<T>::_get_submatrix(SparseMatrix<T>& submatrix,
const std::vector<numeric_index_type> &rows,
const std::vector<numeric_index_type> &cols,
const bool reuse_submatrix) const
{
// Can only extract submatrices from closed matrices
this->close();
// Make sure the SparseMatrix passed in is really a PetscMatrix
PetscMatrix<T>* petsc_submatrix = libmesh_cast_ptr<PetscMatrix<T>*>(&submatrix);
// If we're not reusing submatrix and submatrix is already initialized
// then we need to clear it, otherwise we get a memory leak.
if( !reuse_submatrix && submatrix.initialized() )
submatrix.clear();
// Construct row and column index sets.
PetscErrorCode ierr=0;
IS isrow, iscol;
ierr = ISCreateLibMesh(this->comm().get(),
rows.size(),
(PetscInt*) &rows[0],
PETSC_USE_POINTER,
&isrow); LIBMESH_CHKERRABORT(ierr);
ierr = ISCreateLibMesh(this->comm().get(),
cols.size(),
(PetscInt*) &cols[0],
PETSC_USE_POINTER,
&iscol); LIBMESH_CHKERRABORT(ierr);
// Extract submatrix
#if !PETSC_VERSION_LESS_THAN(3,0,1) || !PETSC_VERSION_RELEASE
ierr = MatGetSubMatrix(_mat,
isrow,
iscol,
(reuse_submatrix ? MAT_REUSE_MATRIX : MAT_INITIAL_MATRIX),
&(petsc_submatrix->_mat)); LIBMESH_CHKERRABORT(ierr);
#else
ierr = MatGetSubMatrix(_mat,
isrow,
iscol,
PETSC_DECIDE,
(reuse_submatrix ? MAT_REUSE_MATRIX : MAT_INITIAL_MATRIX),
&(petsc_submatrix->_mat)); LIBMESH_CHKERRABORT(ierr);
#endif
// Specify that the new submatrix is initialized and close it.
petsc_submatrix->_is_initialized = true;
petsc_submatrix->close();
// Clean up PETSc data structures
ierr = LibMeshISDestroy(&isrow); LIBMESH_CHKERRABORT(ierr);
ierr = LibMeshISDestroy(&iscol); LIBMESH_CHKERRABORT(ierr);
}
/* Developer Notes: This should be implemented with a MatCreate_SchurComplement() as that is the standard design for new Mat classes. */
PetscErrorCode MatGetSchurComplement_Basic(Mat mat,IS isrow0,IS iscol0,IS isrow1,IS iscol1,MatReuse mreuse,Mat *newmat,MatSchurComplementAinvType ainvtype, MatReuse preuse,Mat *newpmat)
{
PetscErrorCode ierr;
Mat A=0,Ap=0,B=0,C=0,D=0;
PetscFunctionBegin;
PetscValidHeaderSpecific(mat,MAT_CLASSID,1);
PetscValidType(mat,1);
PetscValidHeaderSpecific(isrow0,IS_CLASSID,2);
PetscValidHeaderSpecific(iscol0,IS_CLASSID,3);
PetscValidHeaderSpecific(isrow1,IS_CLASSID,4);
PetscValidHeaderSpecific(iscol1,IS_CLASSID,5);
if (mreuse == MAT_IGNORE_MATRIX && preuse == MAT_IGNORE_MATRIX) PetscFunctionReturn(0);
if (mreuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newmat,MAT_CLASSID,7);
if (preuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newpmat,MAT_CLASSID,9);
if (mat->factortype) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
if (mreuse != MAT_IGNORE_MATRIX) {
/* Use MatSchurComplement */
if (mreuse == MAT_REUSE_MATRIX) {
ierr = MatSchurComplementGetSubMatrices(*newmat,&A,&Ap,&B,&C,&D);CHKERRQ(ierr);
if (!A || !Ap || !B || !C) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Attempting to reuse matrix but Schur complement matrices unset");
if (A != Ap) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Preconditioning matrix does not match operator");
ierr = MatDestroy(&Ap);CHKERRQ(ierr); /* get rid of extra reference */
}
}
ierr = MatGetSubMatrix(mat,isrow0,iscol0,mreuse,&A);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow0,iscol1,mreuse,&B);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol0,mreuse,&C);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol1,mreuse,&D);CHKERRQ(ierr);
switch (mreuse) {
case MAT_INITIAL_MATRIX:
ierr = MatCreateSchurComplement(A,A,B,C,D,newmat);CHKERRQ(ierr);
break;
case MAT_REUSE_MATRIX:
ierr = MatSchurComplementUpdateSubMatrices(*newmat,A,A,B,C,D);CHKERRQ(ierr);
break;
default:
if (mreuse != MAT_IGNORE_MATRIX) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Unrecognized value of mreuse");
}
if (preuse != MAT_IGNORE_MATRIX) {
ierr = MatCreateSchurComplementPmat(A,B,C,D,ainvtype,preuse,newpmat);CHKERRQ(ierr);
}
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&D);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode DMCoarsen_AKKT_GAMG11(DM dm, Mat P0f0c, Mat *P1f1c_out) {
PetscErrorCode ierr;
DM_AKKT* kkt = (DM_AKKT*)(dm->data);
Mat Aff = kkt->Aff; /* fine-level KKT matrix */
Mat A1f0f; /* fine-level dual (constraint) Jacobian */
Mat A1f0c; /* = A1f0f*P0f0c coarsen only primal indices */
Mat B1f1f; /* = A1f0c'*A1f0c */
PC gamg11;/* Use PCGAMG internally to get access to some of its methods to operate on B1f1f = A1f0c*A1f0c', where A1f0c = A1f0f*P0f0c. */
PC_GAMG* pc_gamg11;
Mat G1f1f; /* = Graph(B1f1f) */
Mat P1f1c; /* = Prolongator(G1f1f); */
PetscCoarsenData *coarsening;
PetscFunctionBegin;
/*
What is faster:
- A0c1f = P0f0c'*A0f1f followed by B1f1f = A0c1f'*A0c1f, or
- A1f0c = A1f0f*P0f0c followed by B1f1f = A1f0c*A1f0c'?
My bet is on the latter:
- fewer transpositions inside MatMatMult and row indices are always local.
*/
ierr = MatGetSubMatrix(Aff, kkt->isf[1], kkt->isf[0], MAT_INITIAL_MATRIX, &A1f0f); CHKERRQ(ierr);
if(kkt->transposeP) {
ierr = MatMatTransposeMult(A1f0f,P0f0c,MAT_INITIAL_MATRIX, PETSC_DEFAULT, &A1f0c); CHKERRQ(ierr);
}
ierr = MatMatMult(A1f0f,P0f0c,MAT_INITIAL_MATRIX, PETSC_DEFAULT, &A1f0c); CHKERRQ(ierr);
ierr = MatMatTransposeMult(A1f0c, A1f0c, MAT_INITIAL_MATRIX, PETSC_DEFAULT, &B1f1f); CHKERRQ(ierr);
/* We create PCGAMG here since it is only needed for coarsening and we don't want to have to carry the attendant data structures, if we don't need them. */
ierr = PCCreate(((PetscObject)dm)->comm, &gamg11); CHKERRQ(ierr);
/* This must be an aggregating GAMG. */
ierr = PCSetType(gamg11, PCGAMG); CHKERRQ(ierr);
ierr = PCGAMGSetSquareGraph(gamg11, PETSC_FALSE); CHKERRQ(ierr);
/*
Observe that we want to "square" A1f0c before passing it (B1f1f) to GAMG.
This is not because we are not sure how GAMG will deal with a (potentially) non-square matrix,
but rather because if we asked GAMG to square it, it would also smooth the resulting prolongator.
At least PC_GAMG_AGG would, and we need an unsmoothed prolongator.
*/
ierr = PCSetOperators(gamg11, B1f1f, B1f1f, DIFFERENT_NONZERO_PATTERN); CHKERRQ(ierr);
/* FIX: Currently there is no way to tell GAMG to coarsen onto a give comm, but it shouldn't be hard to hack that stuff in. */
pc_gamg11 = (PC_GAMG*)(gamg11->data);
ierr = pc_gamg11->graph(gamg11, B1f1f, &G1f1f); CHKERRQ(ierr);
ierr = pc_gamg11->coarsen(gamg11, &G1f1f, &coarsening); CHKERRQ(ierr);
ierr = pc_gamg11->prolongator(gamg11, B1f1f, G1f1f, coarsening, &P1f1c); CHKERRQ(ierr);
ierr = MatDestroy(&A1f0f); CHKERRQ(ierr);
ierr = MatDestroy(&A1f0c); CHKERRQ(ierr);
ierr = MatDestroy(&B1f1f); CHKERRQ(ierr);
ierr = MatDestroy(&G1f1f); CHKERRQ(ierr);
ierr = PCDestroy(&gamg11); CHKERRQ(ierr);
*P1f1c_out = P1f1c;
PetscFunctionReturn(0);
}
void
get_matrix_rows(Mat A, IS rows, Mat *pA_rows) {
//First, get all the matrix rows we are concerned with.
PetscInt ncols;
MatGetSize(A, PETSC_NULL, &ncols);
IS all_columns;
ISCreateStride(PETSC_COMM_WORLD, ncols, 0, 1, &all_columns);
MatGetSubMatrix(A, rows, all_columns, PETSC_DECIDE, MAT_INITIAL_MATRIX, pA_rows);
ISDestroy(all_columns);
}
/*@C
TaoMatGetSubMat - Gets a submatrix using the IS
Input Parameters:
+ M - the full matrix (n x n)
. is - the index set for the submatrix (both row and column index sets need to be the same)
. v1 - work vector of dimension n, needed for TAO_SUBSET_MASK option
- subset_type - the method TAO is using for subsetting (TAO_SUBSET_SUBVEC, TAO_SUBSET_MASK,
TAO_SUBSET_MATRIXFREE)
Output Parameters:
. Msub - the submatrix
@*/
PetscErrorCode TaoMatGetSubMat(Mat M, IS is, Vec v1, TaoSubsetType subset_type, Mat *Msub)
{
PetscErrorCode ierr;
IS iscomp;
PetscBool flg = PETSC_FALSE;
PetscFunctionBegin;
PetscValidHeaderSpecific(M,MAT_CLASSID,1);
PetscValidHeaderSpecific(is,IS_CLASSID,2);
ierr = MatDestroy(Msub);CHKERRQ(ierr);
switch (subset_type) {
case TAO_SUBSET_SUBVEC:
ierr = MatGetSubMatrix(M, is, is, MAT_INITIAL_MATRIX, Msub);CHKERRQ(ierr);
break;
case TAO_SUBSET_MASK:
/* Get Reduced Hessian
Msub[i,j] = M[i,j] if i,j in Free_Local or i==j
Msub[i,j] = 0 if i!=j and i or j not in Free_Local
*/
ierr = PetscOptionsBegin(PetscObjectComm((PetscObject)M),NULL,NULL,NULL);CHKERRQ(ierr);
ierr = PetscOptionsBool("-different_submatrix","use separate hessian matrix when computing submatrices","TaoSubsetType",flg,&flg,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
if (flg) {
ierr = MatDuplicate(M, MAT_COPY_VALUES, Msub);CHKERRQ(ierr);
} else {
/* Act on hessian directly (default) */
ierr = PetscObjectReference((PetscObject)M);CHKERRQ(ierr);
*Msub = M;
}
/* Save the diagonal to temporary vector */
ierr = MatGetDiagonal(*Msub,v1);CHKERRQ(ierr);
/* Zero out rows and columns */
ierr = ISComplementVec(is,v1,&iscomp);CHKERRQ(ierr);
/* Use v1 instead of 0 here because of PETSc bug */
ierr = MatZeroRowsColumnsIS(*Msub,iscomp,1.0,v1,v1);CHKERRQ(ierr);
ierr = ISDestroy(&iscomp);CHKERRQ(ierr);
break;
case TAO_SUBSET_MATRIXFREE:
ierr = ISComplementVec(is,v1,&iscomp);CHKERRQ(ierr);
ierr = MatCreateSubMatrixFree(M,iscomp,iscomp,Msub);CHKERRQ(ierr);
ierr = ISDestroy(&iscomp);CHKERRQ(ierr);
break;
}
PetscFunctionReturn(0);
}
/*
DMCreateInterpolation_SNESVI - Modifieds the interpolation obtained from the DM by removing all rows and columns associated with active constraints.
*/
PetscErrorCode DMCreateInterpolation_SNESVI(DM dm1,DM dm2,Mat *mat,Vec *vec)
{
PetscErrorCode ierr;
PetscContainer isnes;
DM_SNESVI *dmsnesvi1,*dmsnesvi2;
Mat interp;
PetscFunctionBegin;
ierr = PetscObjectQuery((PetscObject)dm1,"VI",(PetscObject *)&isnes);CHKERRQ(ierr);
if (!isnes) SETERRQ(((PetscObject)dm1)->comm,PETSC_ERR_PLIB,"Composed VI data structure is missing");
ierr = PetscContainerGetPointer(isnes,(void**)&dmsnesvi1);CHKERRQ(ierr);
ierr = PetscObjectQuery((PetscObject)dm2,"VI",(PetscObject *)&isnes);CHKERRQ(ierr);
if (!isnes) SETERRQ(((PetscObject)dm2)->comm,PETSC_ERR_PLIB,"Composed VI data structure is missing");
ierr = PetscContainerGetPointer(isnes,(void**)&dmsnesvi2);CHKERRQ(ierr);
ierr = (*dmsnesvi1->createinterpolation)(dm1,dm2,&interp,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetSubMatrix(interp,dmsnesvi2->inactive,dmsnesvi1->inactive,MAT_INITIAL_MATRIX,mat);CHKERRQ(ierr);
ierr = MatDestroy(&interp);CHKERRQ(ierr);
*vec = 0;
PetscFunctionReturn(0);
}
PetscErrorCode SNESSolve_VINEWTONRSLS(SNES snes)
{
SNES_VINEWTONRSLS *vi = (SNES_VINEWTONRSLS*)snes->data;
PetscErrorCode ierr;
PetscInt maxits,i,lits;
PetscBool lssucceed;
MatStructure flg = DIFFERENT_NONZERO_PATTERN;
PetscReal fnorm,gnorm,xnorm=0,ynorm;
Vec Y,X,F;
KSPConvergedReason kspreason;
PetscFunctionBegin;
snes->numFailures = 0;
snes->numLinearSolveFailures = 0;
snes->reason = SNES_CONVERGED_ITERATING;
maxits = snes->max_its; /* maximum number of iterations */
X = snes->vec_sol; /* solution vector */
F = snes->vec_func; /* residual vector */
Y = snes->work[0]; /* work vectors */
ierr = SNESLineSearchSetVIFunctions(snes->linesearch, SNESVIProjectOntoBounds, SNESVIComputeInactiveSetFnorm);CHKERRQ(ierr);
ierr = SNESLineSearchSetVecs(snes->linesearch, X, PETSC_NULL, PETSC_NULL, PETSC_NULL, PETSC_NULL);CHKERRQ(ierr);
ierr = SNESLineSearchSetUp(snes->linesearch);CHKERRQ(ierr);
ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
snes->iter = 0;
snes->norm = 0.0;
ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
ierr = SNESVIProjectOntoBounds(snes,X);CHKERRQ(ierr);
ierr = SNESComputeFunction(snes,X,F);CHKERRQ(ierr);
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
PetscFunctionReturn(0);
}
ierr = SNESVIComputeInactiveSetFnorm(snes,F,X,&fnorm);CHKERRQ(ierr);
ierr = VecNormBegin(X,NORM_2,&xnorm);CHKERRQ(ierr); /* xnorm <- ||x|| */
ierr = VecNormEnd(X,NORM_2,&xnorm);CHKERRQ(ierr);
if (PetscIsInfOrNanReal(fnorm)) SETERRQ(((PetscObject)X)->comm,PETSC_ERR_FP,"User provided compute function generated a Not-a-Number");
ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
snes->norm = fnorm;
ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
SNESLogConvHistory(snes,fnorm,0);
ierr = SNESMonitor(snes,0,fnorm);CHKERRQ(ierr);
/* set parameter for default relative tolerance convergence test */
snes->ttol = fnorm*snes->rtol;
/* test convergence */
ierr = (*snes->ops->converged)(snes,0,0.0,0.0,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) PetscFunctionReturn(0);
for (i=0; i<maxits; i++) {
IS IS_act,IS_inact; /* _act -> active set _inact -> inactive set */
IS IS_redact; /* redundant active set */
VecScatter scat_act,scat_inact;
PetscInt nis_act,nis_inact;
Vec Y_act,Y_inact,F_inact;
Mat jac_inact_inact,prejac_inact_inact;
PetscBool isequal;
/* Call general purpose update function */
if (snes->ops->update) {
ierr = (*snes->ops->update)(snes, snes->iter);CHKERRQ(ierr);
}
ierr = SNESComputeJacobian(snes,X,&snes->jacobian,&snes->jacobian_pre,&flg);CHKERRQ(ierr);
/* Create active and inactive index sets */
/*original
ierr = SNESVICreateIndexSets_RS(snes,X,F,&IS_act,&IS_inact);CHKERRQ(ierr);
*/
ierr = SNESVIGetActiveSetIS(snes,X,F,&IS_act);CHKERRQ(ierr);
if (vi->checkredundancy) {
(*vi->checkredundancy)(snes,IS_act,&IS_redact,vi->ctxP);CHKERRQ(ierr);
if (IS_redact){
ierr = ISSort(IS_redact);CHKERRQ(ierr);
ierr = ISComplement(IS_redact,X->map->rstart,X->map->rend,&IS_inact);CHKERRQ(ierr);
ierr = ISDestroy(&IS_redact);CHKERRQ(ierr);
}
else {
ierr = ISComplement(IS_act,X->map->rstart,X->map->rend,&IS_inact);CHKERRQ(ierr);
}
} else {
ierr = ISComplement(IS_act,X->map->rstart,X->map->rend,&IS_inact);CHKERRQ(ierr);
}
/* Create inactive set submatrix */
ierr = MatGetSubMatrix(snes->jacobian,IS_inact,IS_inact,MAT_INITIAL_MATRIX,&jac_inact_inact);CHKERRQ(ierr);
if (0) { /* Dead code (temporary developer hack) */
IS keptrows;
ierr = MatFindNonzeroRows(jac_inact_inact,&keptrows);CHKERRQ(ierr);
if (keptrows) {
PetscInt cnt,*nrows,k;
const PetscInt *krows,*inact;
PetscInt rstart=jac_inact_inact->rmap->rstart;
ierr = MatDestroy(&jac_inact_inact);CHKERRQ(ierr);
ierr = ISDestroy(&IS_act);CHKERRQ(ierr);
ierr = ISGetLocalSize(keptrows,&cnt);CHKERRQ(ierr);
ierr = ISGetIndices(keptrows,&krows);CHKERRQ(ierr);
ierr = ISGetIndices(IS_inact,&inact);CHKERRQ(ierr);
ierr = PetscMalloc(cnt*sizeof(PetscInt),&nrows);CHKERRQ(ierr);
for (k=0; k<cnt; k++) {
nrows[k] = inact[krows[k]-rstart];
}
ierr = ISRestoreIndices(keptrows,&krows);CHKERRQ(ierr);
ierr = ISRestoreIndices(IS_inact,&inact);CHKERRQ(ierr);
ierr = ISDestroy(&keptrows);CHKERRQ(ierr);
ierr = ISDestroy(&IS_inact);CHKERRQ(ierr);
ierr = ISCreateGeneral(((PetscObject)snes)->comm,cnt,nrows,PETSC_OWN_POINTER,&IS_inact);CHKERRQ(ierr);
ierr = ISComplement(IS_inact,F->map->rstart,F->map->rend,&IS_act);CHKERRQ(ierr);
ierr = MatGetSubMatrix(snes->jacobian,IS_inact,IS_inact,MAT_INITIAL_MATRIX,&jac_inact_inact);CHKERRQ(ierr);
}
}
ierr = DMSetVI(snes->dm,IS_inact);CHKERRQ(ierr);
/* remove later */
/*
ierr = VecView(vi->xu,PETSC_VIEWER_BINARY_(((PetscObject)(vi->xu))->comm));CHKERRQ(ierr);
ierr = VecView(vi->xl,PETSC_VIEWER_BINARY_(((PetscObject)(vi->xl))->comm));CHKERRQ(ierr);
ierr = VecView(X,PETSC_VIEWER_BINARY_(((PetscObject)X)->comm));CHKERRQ(ierr);
ierr = VecView(F,PETSC_VIEWER_BINARY_(((PetscObject)F)->comm));CHKERRQ(ierr);
ierr = ISView(IS_inact,PETSC_VIEWER_BINARY_(((PetscObject)IS_inact)->comm));CHKERRQ(ierr);
*/
/* Get sizes of active and inactive sets */
ierr = ISGetLocalSize(IS_act,&nis_act);CHKERRQ(ierr);
ierr = ISGetLocalSize(IS_inact,&nis_inact);CHKERRQ(ierr);
/* Create active and inactive set vectors */
ierr = SNESCreateSubVectors_VINEWTONRSLS(snes,nis_inact,&F_inact);CHKERRQ(ierr);
ierr = SNESCreateSubVectors_VINEWTONRSLS(snes,nis_act,&Y_act);CHKERRQ(ierr);
ierr = SNESCreateSubVectors_VINEWTONRSLS(snes,nis_inact,&Y_inact);CHKERRQ(ierr);
/* Create scatter contexts */
ierr = VecScatterCreate(Y,IS_act,Y_act,PETSC_NULL,&scat_act);CHKERRQ(ierr);
ierr = VecScatterCreate(Y,IS_inact,Y_inact,PETSC_NULL,&scat_inact);CHKERRQ(ierr);
/* Do a vec scatter to active and inactive set vectors */
ierr = VecScatterBegin(scat_inact,F,F_inact,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scat_inact,F,F_inact,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterBegin(scat_act,Y,Y_act,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scat_act,Y,Y_act,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterBegin(scat_inact,Y,Y_inact,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scat_inact,Y,Y_inact,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
/* Active set direction = 0 */
ierr = VecSet(Y_act,0);CHKERRQ(ierr);
if (snes->jacobian != snes->jacobian_pre) {
ierr = MatGetSubMatrix(snes->jacobian_pre,IS_inact,IS_inact,MAT_INITIAL_MATRIX,&prejac_inact_inact);CHKERRQ(ierr);
} else prejac_inact_inact = jac_inact_inact;
ierr = ISEqual(vi->IS_inact_prev,IS_inact,&isequal);CHKERRQ(ierr);
if (!isequal) {
ierr = SNESVIResetPCandKSP(snes,jac_inact_inact,prejac_inact_inact);CHKERRQ(ierr);
flg = DIFFERENT_NONZERO_PATTERN;
}
/* ierr = ISView(IS_inact,0);CHKERRQ(ierr); */
/* ierr = ISView(IS_act,0);CHKERRQ(ierr);*/
/* ierr = MatView(snes->jacobian_pre,0); */
ierr = KSPSetOperators(snes->ksp,jac_inact_inact,prejac_inact_inact,flg);CHKERRQ(ierr);
ierr = KSPSetUp(snes->ksp);CHKERRQ(ierr);
{
PC pc;
PetscBool flg;
ierr = KSPGetPC(snes->ksp,&pc);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)pc,PCFIELDSPLIT,&flg);CHKERRQ(ierr);
if (flg) {
KSP *subksps;
ierr = PCFieldSplitGetSubKSP(pc,PETSC_NULL,&subksps);CHKERRQ(ierr);
ierr = KSPGetPC(subksps[0],&pc);CHKERRQ(ierr);
ierr = PetscFree(subksps);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)pc,PCBJACOBI,&flg);CHKERRQ(ierr);
if (flg) {
PetscInt n,N = 101*101,j,cnts[3] = {0,0,0};
const PetscInt *ii;
ierr = ISGetSize(IS_inact,&n);CHKERRQ(ierr);
ierr = ISGetIndices(IS_inact,&ii);CHKERRQ(ierr);
for (j=0; j<n; j++) {
if (ii[j] < N) cnts[0]++;
else if (ii[j] < 2*N) cnts[1]++;
else if (ii[j] < 3*N) cnts[2]++;
}
ierr = ISRestoreIndices(IS_inact,&ii);CHKERRQ(ierr);
ierr = PCBJacobiSetTotalBlocks(pc,3,cnts);CHKERRQ(ierr);
}
}
}
ierr = SNES_KSPSolve(snes,snes->ksp,F_inact,Y_inact);CHKERRQ(ierr);
ierr = KSPGetConvergedReason(snes->ksp,&kspreason);CHKERRQ(ierr);
if (kspreason < 0) {
if (++snes->numLinearSolveFailures >= snes->maxLinearSolveFailures) {
ierr = PetscInfo2(snes,"iter=%D, number linear solve failures %D greater than current SNES allowed, stopping solve\n",snes->iter,snes->numLinearSolveFailures);CHKERRQ(ierr);
snes->reason = SNES_DIVERGED_LINEAR_SOLVE;
break;
}
}
ierr = VecScatterBegin(scat_act,Y_act,Y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd(scat_act,Y_act,Y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterBegin(scat_inact,Y_inact,Y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd(scat_inact,Y_inact,Y,INSERT_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecDestroy(&F_inact);CHKERRQ(ierr);
ierr = VecDestroy(&Y_act);CHKERRQ(ierr);
ierr = VecDestroy(&Y_inact);CHKERRQ(ierr);
ierr = VecScatterDestroy(&scat_act);CHKERRQ(ierr);
ierr = VecScatterDestroy(&scat_inact);CHKERRQ(ierr);
ierr = ISDestroy(&IS_act);CHKERRQ(ierr);
if (!isequal) {
ierr = ISDestroy(&vi->IS_inact_prev);CHKERRQ(ierr);
ierr = ISDuplicate(IS_inact,&vi->IS_inact_prev);CHKERRQ(ierr);
}
ierr = ISDestroy(&IS_inact);CHKERRQ(ierr);
ierr = MatDestroy(&jac_inact_inact);CHKERRQ(ierr);
if (snes->jacobian != snes->jacobian_pre) {
ierr = MatDestroy(&prejac_inact_inact);CHKERRQ(ierr);
}
ierr = KSPGetIterationNumber(snes->ksp,&lits);CHKERRQ(ierr);
snes->linear_its += lits;
ierr = PetscInfo2(snes,"iter=%D, linear solve iterations=%D\n",snes->iter,lits);CHKERRQ(ierr);
/*
if (snes->ops->precheck) {
PetscBool changed_y = PETSC_FALSE;
ierr = (*snes->ops->precheck)(snes,X,Y,snes->precheck,&changed_y);CHKERRQ(ierr);
}
if (PetscLogPrintInfo){
ierr = SNESVICheckResidual_Private(snes,snes->jacobian,F,Y,G,W);CHKERRQ(ierr);
}
*/
/* Compute a (scaled) negative update in the line search routine:
Y <- X - lambda*Y
and evaluate G = function(Y) (depends on the line search).
*/
ierr = VecCopy(Y,snes->vec_sol_update);CHKERRQ(ierr);
ynorm = 1; gnorm = fnorm;
ierr = SNESLineSearchApply(snes->linesearch, X, F, &gnorm, Y);CHKERRQ(ierr);
ierr = SNESLineSearchGetNorms(snes->linesearch, &xnorm, &gnorm, &ynorm);CHKERRQ(ierr);
ierr = PetscInfo4(snes,"fnorm=%18.16e, gnorm=%18.16e, ynorm=%18.16e, lssucceed=%d\n",(double)fnorm,(double)gnorm,(double)ynorm,(int)lssucceed);CHKERRQ(ierr);
if (snes->reason == SNES_DIVERGED_FUNCTION_COUNT) break;
if (snes->domainerror) {
snes->reason = SNES_DIVERGED_FUNCTION_DOMAIN;
ierr = DMDestroyVI(snes->dm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = SNESLineSearchGetSuccess(snes->linesearch, &lssucceed);CHKERRQ(ierr);
if (!lssucceed) {
if (++snes->numFailures >= snes->maxFailures) {
PetscBool ismin;
snes->reason = SNES_DIVERGED_LINE_SEARCH;
ierr = SNESVICheckLocalMin_Private(snes,snes->jacobian,F,X,gnorm,&ismin);CHKERRQ(ierr);
if (ismin) snes->reason = SNES_DIVERGED_LOCAL_MIN;
break;
}
}
/* Update function and solution vectors */
fnorm = gnorm;
/* Monitor convergence */
ierr = PetscObjectTakeAccess(snes);CHKERRQ(ierr);
snes->iter = i+1;
snes->norm = fnorm;
ierr = PetscObjectGrantAccess(snes);CHKERRQ(ierr);
SNESLogConvHistory(snes,snes->norm,lits);
ierr = SNESMonitor(snes,snes->iter,snes->norm);CHKERRQ(ierr);
/* Test for convergence, xnorm = || X || */
if (snes->ops->converged != SNESSkipConverged) { ierr = VecNorm(X,NORM_2,&xnorm);CHKERRQ(ierr); }
ierr = (*snes->ops->converged)(snes,snes->iter,xnorm,ynorm,fnorm,&snes->reason,snes->cnvP);CHKERRQ(ierr);
if (snes->reason) break;
}
ierr = DMDestroyVI(snes->dm);CHKERRQ(ierr);
if (i == maxits) {
ierr = PetscInfo1(snes,"Maximum number of iterations has been reached: %D\n",maxits);CHKERRQ(ierr);
if (!snes->reason) snes->reason = SNES_DIVERGED_MAX_IT;
}
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
Mat M,Mo,C,K,Ko,A[3]; /* problem matrices */
PEP pep; /* polynomial eigenproblem solver context */
IS isf,isbc,is;
PetscInt n=200,nele,Istart,Iend,i,j,mloc,nloc,bc[2];
PetscReal width=0.05,height=0.005,glength=1.0,dlen,EI,area,rho;
PetscScalar K1[4],K2[4],K2t[4],K3[4],M1[4],M2[4],M2t[4],M3[4],damp=5.0;
PetscBool terse;
PetscErrorCode ierr;
PetscLogDouble time1,time2;
SlepcInitialize(&argc,&argv,(char*)0,help);
ierr = PetscOptionsGetInt(NULL,"-n",&n,NULL);CHKERRQ(ierr);
nele = n/2;
n = 2*nele;
ierr = PetscPrintf(PETSC_COMM_WORLD,"\nSimply supported beam damped in the middle, n=%D (nele=%D)\n\n",n,nele);CHKERRQ(ierr);
dlen = glength/nele;
EI = 7e10*width*height*height*height/12.0;
area = width*height;
rho = 0.674/(area*glength);
K1[0] = 12; K1[1] = 6*dlen; K1[2] = 6*dlen; K1[3] = 4*dlen*dlen;
K2[0] = -12; K2[1] = 6*dlen; K2[2] = -6*dlen; K2[3] = 2*dlen*dlen;
K2t[0] = -12; K2t[1] = -6*dlen; K2t[2] = 6*dlen; K2t[3] = 2*dlen*dlen;
K3[0] = 12; K3[1] = -6*dlen; K3[2] = -6*dlen; K3[3] = 4*dlen*dlen;
M1[0] = 156; M1[1] = 22*dlen; M1[2] = 22*dlen; M1[3] = 4*dlen*dlen;
M2[0] = 54; M2[1] = -13*dlen; M2[2] = 13*dlen; M2[3] = -3*dlen*dlen;
M2t[0] = 54; M2t[1] = 13*dlen; M2t[2] = -13*dlen; M2t[3] = -3*dlen*dlen;
M3[0] = 156; M3[1] = -22*dlen; M3[2] = -22*dlen; M3[3] = 4*dlen*dlen;
/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Compute the matrices that define the eigensystem, (k^2*M+k*C+K)x=0
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
/* K is block-tridiagonal */
ierr = MatCreate(PETSC_COMM_WORLD,&Ko);CHKERRQ(ierr);
ierr = MatSetSizes(Ko,PETSC_DECIDE,PETSC_DECIDE,n+2,n+2);CHKERRQ(ierr);
ierr = MatSetBlockSize(Ko,2);CHKERRQ(ierr);
ierr = MatSetFromOptions(Ko);CHKERRQ(ierr);
ierr = MatSetUp(Ko);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Ko,&Istart,&Iend);CHKERRQ(ierr);
for (i=Istart/2;i<Iend/2;i++) {
if (i>0) {
j = i-1;
ierr = MatSetValuesBlocked(Ko,1,&i,1,&j,K2t,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValuesBlocked(Ko,1,&i,1,&i,K3,ADD_VALUES);CHKERRQ(ierr);
}
if (i<nele) {
j = i+1;
ierr = MatSetValuesBlocked(Ko,1,&i,1,&j,K2,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValuesBlocked(Ko,1,&i,1,&i,K1,ADD_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(Ko,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Ko,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatScale(Ko,EI/(dlen*dlen*dlen));CHKERRQ(ierr);
/* M is block-tridiagonal */
ierr = MatCreate(PETSC_COMM_WORLD,&Mo);CHKERRQ(ierr);
ierr = MatSetSizes(Mo,PETSC_DECIDE,PETSC_DECIDE,n+2,n+2);CHKERRQ(ierr);
ierr = MatSetBlockSize(Mo,2);CHKERRQ(ierr);
ierr = MatSetFromOptions(Mo);CHKERRQ(ierr);
ierr = MatSetUp(Mo);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Mo,&Istart,&Iend);CHKERRQ(ierr);
for (i=Istart/2;i<Iend/2;i++) {
if (i>0) {
j = i-1;
ierr = MatSetValuesBlocked(Mo,1,&i,1,&j,M2t,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValuesBlocked(Mo,1,&i,1,&i,M3,ADD_VALUES);CHKERRQ(ierr);
}
if (i<nele) {
j = i+1;
ierr = MatSetValuesBlocked(Mo,1,&i,1,&j,M2,ADD_VALUES);CHKERRQ(ierr);
ierr = MatSetValuesBlocked(Mo,1,&i,1,&i,M1,ADD_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(Mo,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Mo,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatScale(Mo,rho*area*dlen/420);CHKERRQ(ierr);
/* remove rows/columns from K and M corresponding to boundary conditions */
ierr = ISCreateStride(PETSC_COMM_WORLD,Iend-Istart,Istart,1,&isf);CHKERRQ(ierr);
bc[0] = 0; bc[1] = n;
ierr = ISCreateGeneral(PETSC_COMM_SELF,2,bc,PETSC_USE_POINTER,&isbc);CHKERRQ(ierr);
ierr = ISDifference(isf,isbc,&is);CHKERRQ(ierr);
ierr = MatGetSubMatrix(Ko,is,is,MAT_INITIAL_MATRIX,&K);CHKERRQ(ierr);
ierr = MatGetSubMatrix(Mo,is,is,MAT_INITIAL_MATRIX,&M);CHKERRQ(ierr);
ierr = MatGetLocalSize(M,&mloc,&nloc);CHKERRQ(ierr);
/* C is zero except for the (nele,nele)-entry */
ierr = MatCreate(PETSC_COMM_WORLD,&C);CHKERRQ(ierr);
ierr = MatSetSizes(C,mloc,nloc,PETSC_DECIDE,PETSC_DECIDE);CHKERRQ(ierr);
ierr = MatSetFromOptions(C);CHKERRQ(ierr);
ierr = MatSetUp(C);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(C,&Istart,&Iend);CHKERRQ(ierr);
if (nele-1>=Istart && nele-1<Iend) {
ierr = MatSetValue(C,nele-1,nele-1,damp,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Create the eigensolver and solve the problem
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
ierr = PEPCreate(PETSC_COMM_WORLD,&pep);CHKERRQ(ierr);
A[0] = K; A[1] = C; A[2] = M;
ierr = PEPSetOperators(pep,3,A);CHKERRQ(ierr);
ierr = PEPSetFromOptions(pep);CHKERRQ(ierr);
ierr = PetscTime(&time1); CHKERRQ(ierr);
ierr = PEPSolve(pep);CHKERRQ(ierr);
ierr = PetscTime(&time2); CHKERRQ(ierr);
/* - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
Display solution and clean up
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - */
/* show detailed info unless -terse option is given by user */
ierr = PetscOptionsHasName(NULL,"-terse",&terse);CHKERRQ(ierr);
if (terse) {
ierr = PEPErrorView(pep,PEP_ERROR_BACKWARD,NULL);CHKERRQ(ierr);
} else {
ierr = PetscViewerPushFormat(PETSC_VIEWER_STDOUT_WORLD,PETSC_VIEWER_ASCII_INFO_DETAIL);CHKERRQ(ierr);
ierr = PEPReasonView(pep,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PEPErrorView(pep,PEP_ERROR_BACKWARD,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscViewerPopFormat(PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
ierr = PetscPrintf(PETSC_COMM_WORLD,"Time: %g\n\n\n",time2-time1);CHKERRQ(ierr);
ierr = PEPDestroy(&pep);CHKERRQ(ierr);
ierr = ISDestroy(&isf);CHKERRQ(ierr);
ierr = ISDestroy(&isbc);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
ierr = MatDestroy(&M);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&K);CHKERRQ(ierr);
ierr = MatDestroy(&Ko);CHKERRQ(ierr);
ierr = MatDestroy(&Mo);CHKERRQ(ierr);
ierr = SlepcFinalize();CHKERRQ(ierr);
return 0;
}
static PetscErrorCode PCSetUp_Redistribute(PC pc)
{
PC_Redistribute *red = (PC_Redistribute*)pc->data;
PetscErrorCode ierr;
MPI_Comm comm;
PetscInt rstart,rend,i,nz,cnt,*rows,ncnt,dcnt,*drows;
PetscLayout map,nmap;
PetscMPIInt size,imdex,tag,n;
PetscInt *source = PETSC_NULL;
PetscMPIInt *nprocs = PETSC_NULL,nrecvs;
PetscInt j,nsends;
PetscInt *owner = PETSC_NULL,*starts = PETSC_NULL,count,slen;
PetscInt *rvalues,*svalues,recvtotal;
PetscMPIInt *onodes1,*olengths1;
MPI_Request *send_waits = PETSC_NULL,*recv_waits = PETSC_NULL;
MPI_Status recv_status,*send_status;
Vec tvec,diag;
Mat tmat;
const PetscScalar *d;
PetscFunctionBegin;
if (pc->setupcalled) {
ierr = KSPGetOperators(red->ksp,PETSC_NULL,&tmat,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->pmat,red->is,red->is,MAT_REUSE_MATRIX,&tmat);CHKERRQ(ierr);
ierr = KSPSetOperators(red->ksp,tmat,tmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
} else {
PetscInt NN;
ierr = PetscObjectGetComm((PetscObject)pc,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = PetscObjectGetNewTag((PetscObject)pc,&tag);CHKERRQ(ierr);
/* count non-diagonal rows on process */
ierr = MatGetOwnershipRange(pc->mat,&rstart,&rend);CHKERRQ(ierr);
cnt = 0;
for (i=rstart; i<rend; i++) {
ierr = MatGetRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
if (nz > 1) cnt++;
ierr = MatRestoreRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
}
ierr = PetscMalloc(cnt*sizeof(PetscInt),&rows);CHKERRQ(ierr);
ierr = PetscMalloc((rend - rstart - cnt)*sizeof(PetscInt),&drows);CHKERRQ(ierr);
/* list non-diagonal rows on process */
cnt = 0; dcnt = 0;
for (i=rstart; i<rend; i++) {
ierr = MatGetRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
if (nz > 1) rows[cnt++] = i;
else drows[dcnt++] = i - rstart;
ierr = MatRestoreRow(pc->mat,i,&nz,PETSC_NULL,PETSC_NULL);CHKERRQ(ierr);
}
/* create PetscLayout for non-diagonal rows on each process */
ierr = PetscLayoutCreate(comm,&map);CHKERRQ(ierr);
ierr = PetscLayoutSetLocalSize(map,cnt);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(map,1);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(map);CHKERRQ(ierr);
rstart = map->rstart;
rend = map->rend;
/* create PetscLayout for load-balanced non-diagonal rows on each process */
ierr = PetscLayoutCreate(comm,&nmap);CHKERRQ(ierr);
ierr = MPI_Allreduce(&cnt,&ncnt,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr);
ierr = PetscLayoutSetSize(nmap,ncnt);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(nmap,1);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(nmap);CHKERRQ(ierr);
ierr = MatGetSize(pc->pmat,&NN,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscInfo2(pc,"Number of diagonal rows eliminated %d, percentage eliminated %g\n",NN-ncnt,((PetscReal)(NN-ncnt))/((PetscReal)(NN)));CHKERRQ(ierr);
/*
this code is taken from VecScatterCreate_PtoS()
Determines what rows need to be moved where to
load balance the non-diagonal rows
*/
/* count number of contributors to each processor */
ierr = PetscMalloc2(size,PetscMPIInt,&nprocs,cnt,PetscInt,&owner);CHKERRQ(ierr);
ierr = PetscMemzero(nprocs,size*sizeof(PetscMPIInt));CHKERRQ(ierr);
j = 0;
nsends = 0;
for (i=rstart; i<rend; i++) {
if (i < nmap->range[j]) j = 0;
for (; j<size; j++) {
if (i < nmap->range[j+1]) {
if (!nprocs[j]++) nsends++;
owner[i-rstart] = j;
break;
}
}
}
/* inform other processors of number of messages and max length*/
ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,nprocs,&nrecvs);CHKERRQ(ierr);
ierr = PetscGatherMessageLengths(comm,nsends,nrecvs,nprocs,&onodes1,&olengths1);CHKERRQ(ierr);
ierr = PetscSortMPIIntWithArray(nrecvs,onodes1,olengths1);CHKERRQ(ierr);
recvtotal = 0; for (i=0; i<nrecvs; i++) recvtotal += olengths1[i];
/* post receives: rvalues - rows I will own; count - nu */
ierr = PetscMalloc3(recvtotal,PetscInt,&rvalues,nrecvs,PetscInt,&source,nrecvs,MPI_Request,&recv_waits);CHKERRQ(ierr);
count = 0;
for (i=0; i<nrecvs; i++) {
ierr = MPI_Irecv((rvalues+count),olengths1[i],MPIU_INT,onodes1[i],tag,comm,recv_waits+i);CHKERRQ(ierr);
count += olengths1[i];
}
/* do sends:
1) starts[i] gives the starting index in svalues for stuff going to
the ith processor
*/
ierr = PetscMalloc3(cnt,PetscInt,&svalues,nsends,MPI_Request,&send_waits,size,PetscInt,&starts);CHKERRQ(ierr);
starts[0] = 0;
for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];}
for (i=0; i<cnt; i++) {
svalues[starts[owner[i]]++] = rows[i];
}
for (i=0; i<cnt; i++) rows[i] = rows[i] - rstart;
red->drows = drows;
red->dcnt = dcnt;
ierr = PetscFree(rows);CHKERRQ(ierr);
starts[0] = 0;
for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[i-1];}
count = 0;
for (i=0; i<size; i++) {
if (nprocs[i]) {
ierr = MPI_Isend(svalues+starts[i],nprocs[i],MPIU_INT,i,tag,comm,send_waits+count++);CHKERRQ(ierr);
}
}
/* wait on receives */
count = nrecvs;
slen = 0;
while (count) {
ierr = MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);CHKERRQ(ierr);
/* unpack receives into our local space */
ierr = MPI_Get_count(&recv_status,MPIU_INT,&n);CHKERRQ(ierr);
slen += n;
count--;
}
if (slen != recvtotal) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Total message lengths %D not expected %D",slen,recvtotal);
ierr = ISCreateGeneral(comm,slen,rvalues,PETSC_COPY_VALUES,&red->is);CHKERRQ(ierr);
/* free up all work space */
ierr = PetscFree(olengths1);CHKERRQ(ierr);
ierr = PetscFree(onodes1);CHKERRQ(ierr);
ierr = PetscFree3(rvalues,source,recv_waits);CHKERRQ(ierr);
ierr = PetscFree2(nprocs,owner);CHKERRQ(ierr);
if (nsends) { /* wait on sends */
ierr = PetscMalloc(nsends*sizeof(MPI_Status),&send_status);CHKERRQ(ierr);
ierr = MPI_Waitall(nsends,send_waits,send_status);CHKERRQ(ierr);
ierr = PetscFree(send_status);CHKERRQ(ierr);
}
ierr = PetscFree3(svalues,send_waits,starts);CHKERRQ(ierr);
ierr = PetscLayoutDestroy(&map);CHKERRQ(ierr);
ierr = PetscLayoutDestroy(&nmap);CHKERRQ(ierr);
ierr = VecCreateMPI(comm,slen,PETSC_DETERMINE,&red->b);CHKERRQ(ierr);
ierr = VecDuplicate(red->b,&red->x);CHKERRQ(ierr);
ierr = MatGetVecs(pc->pmat,&tvec,PETSC_NULL);CHKERRQ(ierr);
ierr = VecScatterCreate(tvec,red->is,red->b,PETSC_NULL,&red->scatter);CHKERRQ(ierr);
ierr = VecDestroy(&tvec);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->pmat,red->is,red->is,MAT_INITIAL_MATRIX,&tmat);CHKERRQ(ierr);
ierr = KSPSetOperators(red->ksp,tmat,tmat,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
ierr = MatDestroy(&tmat);CHKERRQ(ierr);
}
/* get diagonal portion of matrix */
ierr = PetscMalloc(red->dcnt*sizeof(PetscScalar),&red->diag);CHKERRQ(ierr);
ierr = MatGetVecs(pc->pmat,&diag,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetDiagonal(pc->pmat,diag);CHKERRQ(ierr);
ierr = VecGetArrayRead(diag,&d);CHKERRQ(ierr);
for (i=0; i<red->dcnt; i++) {
red->diag[i] = 1.0/d[red->drows[i]];
}
ierr = VecRestoreArrayRead(diag,&d);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
ierr = KSPSetUp(red->ksp);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode ComputeSubdomainMatrix(DomainData dd, GLLData glldata, Mat *local_mat)
{
PetscErrorCode ierr;
PetscInt localsize,zloc,yloc,xloc,auxnex,auxney,auxnez;
PetscInt ie,je,ke,i,j,k,ig,jg,kg,ii,ming;
PetscInt *indexg,*cols,*colsg;
PetscScalar *vals;
Mat temp_local_mat,elem_mat_DBC=0,*usedmat;
IS submatIS;
PetscFunctionBeginUser;
ierr = MatGetSize(glldata.elem_mat,&i,&j);CHKERRQ(ierr);
ierr = PetscMalloc1(i,&indexg);CHKERRQ(ierr);
ierr = PetscMalloc1(i,&colsg);CHKERRQ(ierr);
/* get submatrix of elem_mat without dirichlet nodes */
if (!dd.pure_neumann && !dd.DBC_zerorows && !dd.ipx) {
xloc = dd.p+1;
yloc = 1;
zloc = 1;
if (dd.dim>1) yloc = dd.p+1;
if (dd.dim>2) zloc = dd.p+1;
ii = 0;
for (k=0;k<zloc;k++) {
for (j=0;j<yloc;j++) {
for (i=1;i<xloc;i++) {
indexg[ii]=k*xloc*yloc+j*xloc+i;
ii++;
}
}
}
ierr = ISCreateGeneral(PETSC_COMM_SELF,ii,indexg,PETSC_COPY_VALUES,&submatIS);CHKERRQ(ierr);
ierr = MatGetSubMatrix(glldata.elem_mat,submatIS,submatIS,MAT_INITIAL_MATRIX,&elem_mat_DBC);CHKERRQ(ierr);
ierr = ISDestroy(&submatIS);CHKERRQ(ierr);
}
/* Assemble subdomain matrix */
localsize = dd.xm_l*dd.ym_l*dd.zm_l;
ierr = MatCreate(PETSC_COMM_SELF,&temp_local_mat);CHKERRQ(ierr);
ierr = MatSetSizes(temp_local_mat,localsize,localsize,localsize,localsize);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(temp_local_mat,"subdomain_");CHKERRQ(ierr);
/* set local matrices type: here we use SEQSBAIJ primarily for testing purpose */
/* in order to avoid conversions inside the BDDC code, use SeqAIJ if possible */
if (dd.DBC_zerorows && !dd.ipx) { /* in this case, we need to zero out some of the rows, so use seqaij */
ierr = MatSetType(temp_local_mat,MATSEQAIJ);CHKERRQ(ierr);
} else {
ierr = MatSetType(temp_local_mat,MATSEQSBAIJ);CHKERRQ(ierr);
}
ierr = MatSetFromOptions(temp_local_mat);CHKERRQ(ierr);
i = PetscPowRealInt(3.0*(dd.p+1.0),dd.dim);
ierr = MatSeqAIJSetPreallocation(temp_local_mat,i,NULL);CHKERRQ(ierr); /* very overestimated */
ierr = MatSeqSBAIJSetPreallocation(temp_local_mat,1,i,NULL);CHKERRQ(ierr); /* very overestimated */
ierr = MatSetOption(temp_local_mat,MAT_KEEP_NONZERO_PATTERN,PETSC_TRUE);CHKERRQ(ierr);
yloc = dd.p+1;
zloc = dd.p+1;
if (dd.dim < 3) zloc = 1;
if (dd.dim < 2) yloc = 1;
auxnez = dd.nez_l;
auxney = dd.ney_l;
auxnex = dd.nex_l;
if (dd.dim < 3) auxnez = 1;
if (dd.dim < 2) auxney = 1;
for (ke=0; ke<auxnez; ke++) {
for (je=0; je<auxney; je++) {
for (ie=0; ie<auxnex; ie++) {
/* customize element accounting for BC */
xloc = dd.p+1;
ming = 0;
usedmat = &glldata.elem_mat;
if (!dd.pure_neumann && !dd.DBC_zerorows && !dd.ipx) {
if (ie == 0) {
xloc = dd.p;
usedmat = &elem_mat_DBC;
} else {
ming = -1;
usedmat = &glldata.elem_mat;
}
}
/* local to the element/global to the subdomain indexing */
for (k=0; k<zloc; k++) {
kg = ke*dd.p+k;
for (j=0; j<yloc; j++) {
jg = je*dd.p+j;
for (i=0; i<xloc; i++) {
ig = ie*dd.p+i+ming;
ii = k*xloc*yloc+j*xloc+i;
indexg[ii] = kg*dd.xm_l*dd.ym_l+jg*dd.xm_l+ig;
}
}
}
/* Set values */
for (i=0; i<xloc*yloc*zloc; i++) {
ierr = MatGetRow(*usedmat,i,&j,(const PetscInt**)&cols,(const PetscScalar**)&vals);CHKERRQ(ierr);
for (k=0; k<j; k++) colsg[k] = indexg[cols[k]];
ierr = MatSetValues(temp_local_mat,1,&indexg[i],j,colsg,vals,ADD_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(*usedmat,i,&j,(const PetscInt**)&cols,(const PetscScalar**)&vals);CHKERRQ(ierr);
}
}
}
}
ierr = PetscFree(indexg);CHKERRQ(ierr);
ierr = PetscFree(colsg);CHKERRQ(ierr);
ierr = MatAssemblyBegin(temp_local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd (temp_local_mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
#if DEBUG
{
Vec lvec,rvec;
PetscReal norm;
ierr = MatCreateVecs(temp_local_mat,&lvec,&rvec);CHKERRQ(ierr);
ierr = VecSet(lvec,1.0);CHKERRQ(ierr);
ierr = MatMult(temp_local_mat,lvec,rvec);CHKERRQ(ierr);
ierr = VecNorm(rvec,NORM_INFINITY,&norm);CHKERRQ(ierr);
printf("Test null space of local mat % 1.14e\n",norm);
ierr = VecDestroy(&lvec);CHKERRQ(ierr);
ierr = VecDestroy(&rvec);CHKERRQ(ierr);
}
#endif
*local_mat = temp_local_mat;
ierr = MatDestroy(&elem_mat_DBC);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode DMSetUp_AKKT(DM dm) {
DM_AKKT *kkt = (DM_AKKT*)(dm->data);
PetscErrorCode ierr;
PetscFunctionBegin;
if(dm->setupcalled) PetscFunctionReturn(0);
if(!kkt->Aff){
if(kkt->dm) {
ierr = DMCreateMatrix(kkt->dm, MATAIJ, &kkt->Aff); CHKERRQ(ierr);
}
else SETERRQ(((PetscObject)dm)->comm, PETSC_ERR_ARG_WRONGSTATE, "Neither matrix nor DM set");
}
if(!kkt->isf[0] && !kkt->isf[0]) {
if(kkt->detect_saddle_point) {
ierr = MatFindZeroDiagonals(kkt->Aff,&kkt->isf[1]);CHKERRQ(ierr);
}
else if(kkt->dm && kkt->dname) {
DM ddm;
PetscInt n;
char **names;
IS *iss;
DM *dms;
PetscInt i;
ierr = DMCreateFieldDecompositionDM(kkt->dm, kkt->dname, &ddm); CHKERRQ(ierr);
ierr = DMCreateFieldDecomposition(ddm, &n, &names, &iss, &dms); CHKERRQ(ierr);
if(n < 1 || n > 2)
SETERRQ2(((PetscObject)dm)->comm, PETSC_ERR_ARG_WRONG, "Number of parts in decomposition %s must be between 1 and 2. Got %D instead",kkt->dname, n);
for(i = 0; i < n; ++i) {
if(!iss[i] && dms[i]) {
const char* label;
if(i == 0)
label = "primal";
else
label = "dual";
SETERRQ1(((PetscObject)dm)->comm, PETSC_ERR_ARG_WRONG, "Decomposition defines %s subDM, but no embedding IS is given", label);
}
}
ierr = DMAKKTSetFieldDecomposition(dm, n, (const char**)names, iss, dms); CHKERRQ(ierr);
for(i = 0; i < n; ++i) {
ierr = PetscFree(names[i]); CHKERRQ(ierr);
ierr = ISDestroy(&(iss[i])); CHKERRQ(ierr);
ierr = DMDestroy(&(dms[i])); CHKERRQ(ierr);
}
}
}
if(!kkt->isf[0] && !kkt->isf[1]) SETERRQ(((PetscObject)dm)->comm, PETSC_ERR_ARG_WRONGSTATE, "Decomposition ISs not set and could not be derived. ");
if(!kkt->isf[0] || !kkt->isf[1]) {
PetscInt lstart, lend;
ierr = MatGetOwnershipRange(kkt->Aff, &lstart, &lend); CHKERRQ(ierr);
if(!kkt->isf[0]) {
ierr = ISComplement(kkt->isf[0], lstart, lend, kkt->isf+1); CHKERRQ(ierr);
}
else {
ierr = ISComplement(kkt->isf[1], lstart, lend, kkt->isf+0); CHKERRQ(ierr);
}
}
/* FIX: Should we allow a combination of empty kkt->dmf[0] and non-empty kkt->dmf[1]? */
if(!kkt->dmf[0]) {
/* Construct a GAMG proxy to coarsen the primal block. */
Mat A0f0f;
IS is00;
PetscInt lstart, lend;
const char* primal = {"all"};
ierr = DMCreate(((PetscObject)dm)->comm, kkt->dmf+0); CHKERRQ(ierr);
ierr = DMSetType(kkt->dmf[0],DMAKKT); CHKERRQ(ierr);
ierr = MatGetSubMatrix(kkt->Aff, kkt->isf[0], kkt->isf[0], MAT_INITIAL_MATRIX, &A0f0f); CHKERRQ(ierr);
ierr = DMAKKTSetMatrix(kkt->dmf[0], A0f0f); CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A0f0f, &lstart, &lend); CHKERRQ(ierr);
ierr = ISCreateStride(((PetscObject)A0f0f)->comm, lend-lstart, lstart, 1, &is00); CHKERRQ(ierr);
ierr = DMAKKTSetFieldDecomposition(kkt->dmf[0], 1, &primal, &is00, PETSC_NULL); CHKERRQ(ierr);
}
dm->setupcalled = PETSC_TRUE;
PetscFunctionReturn(0);
}
PetscErrorCode SNESSetUp_Multiblock(SNES snes)
{
SNES_Multiblock *mb = (SNES_Multiblock *) snes->data;
BlockDesc blocks;
PetscInt i, numBlocks;
PetscErrorCode ierr;
PetscFunctionBegin;
/* ierr = SNESDefaultGetWork(snes, 1);CHKERRQ(ierr); */
ierr = SNESMultiblockSetDefaults(snes);CHKERRQ(ierr);
numBlocks = mb->numBlocks;
blocks = mb->blocks;
/* Create ISs */
if (!mb->issetup) {
PetscInt ccsize, rstart, rend, nslots, bs;
PetscBool sorted;
mb->issetup = PETSC_TRUE;
bs = mb->bs;
ierr = MatGetOwnershipRange(snes->jacobian_pre, &rstart, &rend);CHKERRQ(ierr);
ierr = MatGetLocalSize(snes->jacobian_pre, PETSC_NULL, &ccsize);CHKERRQ(ierr);
nslots = (rend - rstart)/bs;
for (i = 0; i < numBlocks; ++i) {
if (mb->defaultblocks) {
ierr = ISCreateStride(((PetscObject) snes)->comm, nslots, rstart+i, numBlocks, &blocks->is);CHKERRQ(ierr);
} else if (!blocks->is) {
if (blocks->nfields > 1) {
PetscInt *ii, j, k, nfields = blocks->nfields, *fields = blocks->fields;
ierr = PetscMalloc(nfields*nslots*sizeof(PetscInt), &ii);CHKERRQ(ierr);
for (j = 0; j < nslots; ++j) {
for (k = 0; k < nfields; ++k) {
ii[nfields*j + k] = rstart + bs*j + fields[k];
}
}
ierr = ISCreateGeneral(((PetscObject) snes)->comm, nslots*nfields, ii, PETSC_OWN_POINTER, &blocks->is);CHKERRQ(ierr);
} else {
ierr = ISCreateStride(((PetscObject) snes)->comm, nslots, rstart+blocks->fields[0], bs, &blocks->is);CHKERRQ(ierr);
}
}
ierr = ISSorted(blocks->is, &sorted);CHKERRQ(ierr);
if (!sorted) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_USER, "Fields must be sorted when creating split");
blocks = blocks->next;
}
}
#if 0
/* Create matrices */
ilink = jac->head;
if (!jac->pmat) {
ierr = PetscMalloc(nsplit*sizeof(Mat),&jac->pmat);CHKERRQ(ierr);
for (i=0; i<nsplit; i++) {
ierr = MatGetSubMatrix(pc->pmat,ilink->is,ilink->is,MAT_INITIAL_MATRIX,&jac->pmat[i]);CHKERRQ(ierr);
ilink = ilink->next;
}
} else {
for (i=0; i<nsplit; i++) {
ierr = MatGetSubMatrix(pc->pmat,ilink->is,ilink->is,MAT_REUSE_MATRIX,&jac->pmat[i]);CHKERRQ(ierr);
ilink = ilink->next;
}
}
if (jac->realdiagonal) {
ilink = jac->head;
if (!jac->mat) {
ierr = PetscMalloc(nsplit*sizeof(Mat),&jac->mat);CHKERRQ(ierr);
for (i=0; i<nsplit; i++) {
ierr = MatGetSubMatrix(pc->mat,ilink->is,ilink->is,MAT_INITIAL_MATRIX,&jac->mat[i]);CHKERRQ(ierr);
ilink = ilink->next;
}
} else {
for (i=0; i<nsplit; i++) {
if (jac->mat[i]) {ierr = MatGetSubMatrix(pc->mat,ilink->is,ilink->is,MAT_REUSE_MATRIX,&jac->mat[i]);CHKERRQ(ierr);}
ilink = ilink->next;
}
}
} else {
jac->mat = jac->pmat;
}
#endif
#if 0
if (jac->type != PC_COMPOSITE_ADDITIVE && jac->type != PC_COMPOSITE_SCHUR) {
/* extract the rows of the matrix associated with each field: used for efficient computation of residual inside algorithm */
ilink = jac->head;
if (!jac->Afield) {
ierr = PetscMalloc(nsplit*sizeof(Mat),&jac->Afield);CHKERRQ(ierr);
for (i=0; i<nsplit; i++) {
ierr = MatGetSubMatrix(pc->mat,ilink->is,PETSC_NULL,MAT_INITIAL_MATRIX,&jac->Afield[i]);CHKERRQ(ierr);
ilink = ilink->next;
}
} else {
for (i=0; i<nsplit; i++) {
ierr = MatGetSubMatrix(pc->mat,ilink->is,PETSC_NULL,MAT_REUSE_MATRIX,&jac->Afield[i]);CHKERRQ(ierr);
ilink = ilink->next;
}
}
}
#endif
if (mb->type == PC_COMPOSITE_SCHUR) {
#if 0
IS ccis;
PetscInt rstart,rend;
if (nsplit != 2) SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_ARG_INCOMP,"To use Schur complement preconditioner you must have exactly 2 fields");
/* When extracting off-diagonal submatrices, we take complements from this range */
ierr = MatGetOwnershipRangeColumn(pc->mat,&rstart,&rend);CHKERRQ(ierr);
/* need to handle case when one is resetting up the preconditioner */
if (jac->schur) {
ilink = jac->head;
ierr = ISComplement(ilink->is,rstart,rend,&ccis);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->mat,ilink->is,ccis,MAT_REUSE_MATRIX,&jac->B);CHKERRQ(ierr);
ierr = ISDestroy(&ccis);CHKERRQ(ierr);
ilink = ilink->next;
ierr = ISComplement(ilink->is,rstart,rend,&ccis);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->mat,ilink->is,ccis,MAT_REUSE_MATRIX,&jac->C);CHKERRQ(ierr);
ierr = ISDestroy(&ccis);CHKERRQ(ierr);
ierr = MatSchurComplementUpdate(jac->schur,jac->mat[0],jac->pmat[0],jac->B,jac->C,jac->pmat[1],pc->flag);CHKERRQ(ierr);
ierr = KSPSetOperators(jac->kspschur,jac->schur,FieldSplitSchurPre(jac),pc->flag);CHKERRQ(ierr);
} else {
KSP ksp;
char schurprefix[256];
/* extract the A01 and A10 matrices */
ilink = jac->head;
ierr = ISComplement(ilink->is,rstart,rend,&ccis);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->mat,ilink->is,ccis,MAT_INITIAL_MATRIX,&jac->B);CHKERRQ(ierr);
ierr = ISDestroy(&ccis);CHKERRQ(ierr);
ilink = ilink->next;
ierr = ISComplement(ilink->is,rstart,rend,&ccis);CHKERRQ(ierr);
ierr = MatGetSubMatrix(pc->mat,ilink->is,ccis,MAT_INITIAL_MATRIX,&jac->C);CHKERRQ(ierr);
ierr = ISDestroy(&ccis);CHKERRQ(ierr);
/* Use mat[0] (diagonal block of the real matrix) preconditioned by pmat[0] */
ierr = MatCreateSchurComplement(jac->mat[0],jac->pmat[0],jac->B,jac->C,jac->mat[1],&jac->schur);CHKERRQ(ierr);
/* set tabbing and options prefix of KSP inside the MatSchur */
ierr = MatSchurComplementGetKSP(jac->schur,&ksp);CHKERRQ(ierr);
ierr = PetscObjectIncrementTabLevel((PetscObject)ksp,(PetscObject)pc,2);CHKERRQ(ierr);
ierr = PetscSNPrintf(schurprefix,sizeof(schurprefix),"%sfieldsplit_%s_",((PetscObject)pc)->prefix?((PetscObject)pc)->prefix:"",jac->head->splitname);CHKERRQ(ierr);
ierr = KSPSetOptionsPrefix(ksp,schurprefix);CHKERRQ(ierr);
ierr = MatSetFromOptions(jac->schur);CHKERRQ(ierr);
ierr = KSPCreate(((PetscObject)pc)->comm,&jac->kspschur);CHKERRQ(ierr);
ierr = PetscLogObjectParent((PetscObject)pc,(PetscObject)jac->kspschur);CHKERRQ(ierr);
ierr = PetscObjectIncrementTabLevel((PetscObject)jac->kspschur,(PetscObject)pc,1);CHKERRQ(ierr);
ierr = KSPSetOperators(jac->kspschur,jac->schur,FieldSplitSchurPre(jac),DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
if (jac->schurpre == PC_FIELDSPLIT_SCHUR_PRE_SELF) {
PC pc;
ierr = KSPGetPC(jac->kspschur,&pc);CHKERRQ(ierr);
ierr = PCSetType(pc,PCNONE);CHKERRQ(ierr);
/* Note: This is bad if there exist preconditioners for MATSCHURCOMPLEMENT */
}
ierr = PetscSNPrintf(schurprefix,sizeof(schurprefix),"%sfieldsplit_%s_",((PetscObject)pc)->prefix?((PetscObject)pc)->prefix:"",ilink->splitname);CHKERRQ(ierr);
ierr = KSPSetOptionsPrefix(jac->kspschur,schurprefix);CHKERRQ(ierr);
/* really want setfromoptions called in PCSetFromOptions_FieldSplit(), but it is not ready yet */
ierr = KSPSetFromOptions(jac->kspschur);CHKERRQ(ierr);
ierr = PetscMalloc2(2,Vec,&jac->x,2,Vec,&jac->y);CHKERRQ(ierr);
ierr = MatGetVecs(jac->pmat[0],&jac->x[0],&jac->y[0]);CHKERRQ(ierr);
ierr = MatGetVecs(jac->pmat[1],&jac->x[1],&jac->y[1]);CHKERRQ(ierr);
ilink = jac->head;
ilink->x = jac->x[0]; ilink->y = jac->y[0];
ilink = ilink->next;
ilink->x = jac->x[1]; ilink->y = jac->y[1];
}
#endif
} else {
/* Set up the individual SNESs */
blocks = mb->blocks;
i = 0;
while (blocks) {
/*TODO: Set these correctly */
/*ierr = SNESSetFunction(blocks->snes, blocks->x, func);CHKERRQ(ierr);*/
/*ierr = SNESSetJacobian(blocks->snes, blocks->x, jac);CHKERRQ(ierr);*/
ierr = VecDuplicate(blocks->snes->vec_sol, &blocks->x);CHKERRQ(ierr);
/* really want setfromoptions called in SNESSetFromOptions_Multiblock(), but it is not ready yet */
ierr = SNESSetFromOptions(blocks->snes);CHKERRQ(ierr);
ierr = SNESSetUp(blocks->snes);CHKERRQ(ierr);
blocks = blocks->next;
i++;
}
}
/* Compute scatter contexts needed by multiplicative versions and non-default splits */
if (!mb->blocks->sctx) {
Vec xtmp;
blocks = mb->blocks;
ierr = MatGetVecs(snes->jacobian_pre, &xtmp, PETSC_NULL);CHKERRQ(ierr);
while(blocks) {
ierr = VecScatterCreate(xtmp, blocks->is, blocks->x, PETSC_NULL, &blocks->sctx);CHKERRQ(ierr);
blocks = blocks->next;
}
ierr = VecDestroy(&xtmp);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode PCBDDCSubSchursSetUp(PCBDDCSubSchurs sub_schurs, Mat S, IS is_A_I, IS is_A_B, PetscInt ncc, IS is_cc[], PetscInt xadj[], PetscInt adjncy[], PetscInt nlayers)
{
Mat A_II,A_IB,A_BI,A_BB;
ISLocalToGlobalMapping BtoNmap,ItoNmap;
PetscBT touched;
PetscInt i,n_I,n_B,n_local,*local_numbering;
PetscBool is_sorted;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = ISSorted(is_A_I,&is_sorted);CHKERRQ(ierr);
if (!is_sorted) {
SETERRQ(PetscObjectComm((PetscObject)is_A_I),PETSC_ERR_PLIB,"IS for I dofs should be shorted");
}
ierr = ISSorted(is_A_B,&is_sorted);CHKERRQ(ierr);
if (!is_sorted) {
SETERRQ(PetscObjectComm((PetscObject)is_A_B),PETSC_ERR_PLIB,"IS for B dofs should be shorted");
}
/* get sizes */
ierr = ISGetLocalSize(is_A_I,&n_I);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_A_B,&n_B);CHKERRQ(ierr);
n_local = n_I+n_B;
/* maps */
ierr = ISLocalToGlobalMappingCreateIS(is_A_B,&BtoNmap);CHKERRQ(ierr);
if (nlayers >= 0 && xadj != NULL && adjncy != NULL) { /* I problems have a different size of the original ones */
ierr = ISLocalToGlobalMappingCreateIS(is_A_I,&ItoNmap);CHKERRQ(ierr);
/* allocate some auxiliary space */
ierr = PetscMalloc1(n_local,&local_numbering);CHKERRQ(ierr);
ierr = PetscBTCreate(n_local,&touched);CHKERRQ(ierr);
} else {
ItoNmap = 0;
local_numbering = 0;
touched = 0;
}
/* get Schur complement matrices */
ierr = MatSchurComplementGetSubMatrices(S,&A_II,NULL,&A_IB,&A_BI,&A_BB);CHKERRQ(ierr);
/* allocate space for schur complements */
ierr = PetscMalloc5(ncc,&sub_schurs->is_AEj_I,ncc,&sub_schurs->is_AEj_B,ncc,&sub_schurs->S_Ej,ncc,&sub_schurs->work1,ncc,&sub_schurs->work2);CHKERRQ(ierr);
sub_schurs->n_subs = ncc;
/* cycle on subsets and extract schur complements */
for (i=0;i<sub_schurs->n_subs;i++) {
Mat AE_II,AE_IE,AE_EI,AE_EE;
IS is_I,is_subset_B;
/* get IS for subsets in B numbering */
ierr = ISDuplicate(is_cc[i],&sub_schurs->is_AEj_B[i]);CHKERRQ(ierr);
ierr = ISSort(sub_schurs->is_AEj_B[i]);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyIS(BtoNmap,IS_GTOLM_DROP,sub_schurs->is_AEj_B[i],&is_subset_B);CHKERRQ(ierr);
/* BB block on subset */
ierr = MatGetSubMatrix(A_BB,is_subset_B,is_subset_B,MAT_INITIAL_MATRIX,&AE_EE);CHKERRQ(ierr);
if (ItoNmap) { /* is ItoNmap has been computed, extracts only a part of I dofs */
const PetscInt* idx_B;
PetscInt n_local_dofs,n_prev_added,j,layer,subset_size;
/* all boundary dofs must be skipped when adding layers */
ierr = PetscBTMemzero(n_local,touched);CHKERRQ(ierr);
ierr = ISGetIndices(is_A_B,&idx_B);CHKERRQ(ierr);
for (j=0;j<n_B;j++) {
ierr = PetscBTSet(touched,idx_B[j]);CHKERRQ(ierr);
}
ierr = ISRestoreIndices(is_A_B,&idx_B);CHKERRQ(ierr);
/* add next layers of dofs */
ierr = ISGetLocalSize(is_cc[i],&subset_size);CHKERRQ(ierr);
ierr = ISGetIndices(is_cc[i],&idx_B);CHKERRQ(ierr);
ierr = PetscMemcpy(local_numbering,idx_B,subset_size*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISRestoreIndices(is_cc[i],&idx_B);CHKERRQ(ierr);
n_local_dofs = subset_size;
n_prev_added = subset_size;
for (layer=0;layer<nlayers;layer++) {
PetscInt n_added;
if (n_local_dofs == n_I+subset_size) break;
if (n_local_dofs > n_I+subset_size) {
SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Error querying layer %d. Out of bound access (%d > %d)",layer,n_local_dofs,n_I+subset_size);
}
ierr = PCBDDCAdjGetNextLayer_Private(local_numbering+n_local_dofs,n_prev_added,touched,xadj,adjncy,&n_added);CHKERRQ(ierr);
n_prev_added = n_added;
n_local_dofs += n_added;
if (!n_added) break;
}
/* IS for I dofs in original numbering and in I numbering */
ierr = ISCreateGeneral(PetscObjectComm((PetscObject)ItoNmap),n_local_dofs-subset_size,local_numbering+subset_size,PETSC_COPY_VALUES,&sub_schurs->is_AEj_I[i]);CHKERRQ(ierr);
ierr = ISSort(sub_schurs->is_AEj_I[i]);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyIS(ItoNmap,IS_GTOLM_DROP,sub_schurs->is_AEj_I[i],&is_I);CHKERRQ(ierr);
/* II block */
ierr = MatGetSubMatrix(A_II,is_I,is_I,MAT_INITIAL_MATRIX,&AE_II);CHKERRQ(ierr);
} else { /* in this case we can take references of already existing IS and matrices for I dofs */
/* IS for I dofs in original numbering */
ierr = PetscObjectReference((PetscObject)is_A_I);CHKERRQ(ierr);
sub_schurs->is_AEj_I[i] = is_A_I;
/* IS for I dofs in I numbering TODO: "first" argument of ISCreateStride is not general */
ierr = ISCreateStride(PetscObjectComm((PetscObject)is_A_I),n_I,0,1,&is_I);CHKERRQ(ierr);
/* II block is the same */
ierr = PetscObjectReference((PetscObject)A_II);CHKERRQ(ierr);
AE_II = A_II;
}
/* IE block */
ierr = MatGetSubMatrix(A_IB,is_I,is_subset_B,MAT_INITIAL_MATRIX,&AE_IE);CHKERRQ(ierr);
/* EI block */
ierr = MatGetSubMatrix(A_BI,is_subset_B,is_I,MAT_INITIAL_MATRIX,&AE_EI);CHKERRQ(ierr);
/* setup Schur complements on subset */
ierr = MatCreateSchurComplement(AE_II,AE_II,AE_IE,AE_EI,AE_EE,&sub_schurs->S_Ej[i]);CHKERRQ(ierr);
ierr = MatGetVecs(sub_schurs->S_Ej[i],&sub_schurs->work1[i],&sub_schurs->work2[i]);CHKERRQ(ierr);
if (AE_II == A_II) { /* we can reuse the same ksp */
KSP ksp;
ierr = MatSchurComplementGetKSP(S,&ksp);CHKERRQ(ierr);
ierr = MatSchurComplementSetKSP(sub_schurs->S_Ej[i],ksp);CHKERRQ(ierr);
} else { /* build new ksp object which inherits ksp and pc types from the original one */
KSP origksp,schurksp;
PC origpc,schurpc;
KSPType ksp_type;
PCType pc_type;
PetscInt n_internal;
ierr = MatSchurComplementGetKSP(S,&origksp);CHKERRQ(ierr);
ierr = MatSchurComplementGetKSP(sub_schurs->S_Ej[i],&schurksp);CHKERRQ(ierr);
ierr = KSPGetType(origksp,&ksp_type);CHKERRQ(ierr);
ierr = KSPSetType(schurksp,ksp_type);CHKERRQ(ierr);
ierr = KSPGetPC(schurksp,&schurpc);CHKERRQ(ierr);
ierr = KSPGetPC(origksp,&origpc);CHKERRQ(ierr);
ierr = PCGetType(origpc,&pc_type);CHKERRQ(ierr);
ierr = PCSetType(schurpc,pc_type);CHKERRQ(ierr);
ierr = ISGetSize(is_I,&n_internal);CHKERRQ(ierr);
if (n_internal) { /* UMFPACK gives error with 0 sized problems */
MatSolverPackage solver=NULL;
ierr = PCFactorGetMatSolverPackage(origpc,(const MatSolverPackage*)&solver);CHKERRQ(ierr);
if (solver) {
ierr = PCFactorSetMatSolverPackage(schurpc,solver);CHKERRQ(ierr);
}
}
ierr = KSPSetUp(schurksp);CHKERRQ(ierr);
}
/* free */
ierr = MatDestroy(&AE_II);CHKERRQ(ierr);
ierr = MatDestroy(&AE_EE);CHKERRQ(ierr);
ierr = MatDestroy(&AE_IE);CHKERRQ(ierr);
ierr = MatDestroy(&AE_EI);CHKERRQ(ierr);
ierr = ISDestroy(&is_I);CHKERRQ(ierr);
ierr = ISDestroy(&is_subset_B);CHKERRQ(ierr);
}
/* free */
ierr = ISLocalToGlobalMappingDestroy(&ItoNmap);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&BtoNmap);CHKERRQ(ierr);
ierr = PetscFree(local_numbering);CHKERRQ(ierr);
ierr = PetscBTDestroy(&touched);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode port_lsd_bfbt(void)
{
Mat A;
Vec x,b;
KSP ksp_A;
PC pc_A;
IS isu,isp;
PetscErrorCode ierr;
PetscFunctionBeginUser;
ierr = LoadTestMatrices(&A,&x,&b,&isu,&isp);CHKERRQ(ierr);
ierr = KSPCreate(PETSC_COMM_WORLD,&ksp_A);CHKERRQ(ierr);
ierr = KSPSetOptionsPrefix(ksp_A,"fc_");CHKERRQ(ierr);
ierr = KSPSetOperators(ksp_A,A,A);CHKERRQ(ierr);
ierr = KSPGetPC(ksp_A,&pc_A);CHKERRQ(ierr);
ierr = PCSetType(pc_A,PCFIELDSPLIT);CHKERRQ(ierr);
ierr = PCFieldSplitSetBlockSize(pc_A,2);CHKERRQ(ierr);
ierr = PCFieldSplitSetIS(pc_A,"velocity",isu);CHKERRQ(ierr);
ierr = PCFieldSplitSetIS(pc_A,"pressure",isp);CHKERRQ(ierr);
ierr = KSPSetFromOptions(ksp_A);CHKERRQ(ierr);
ierr = KSPSolve(ksp_A,b,x);CHKERRQ(ierr);
/* Pull u,p out of x */
{
PetscInt loc;
PetscReal max,norm;
PetscScalar sum;
Vec uvec,pvec;
VecScatter uscat,pscat;
Mat A11,A22;
/* grab matrices and create the compatable u,p vectors */
ierr = MatGetSubMatrix(A,isu,isu,MAT_INITIAL_MATRIX,&A11);CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,isp,isp,MAT_INITIAL_MATRIX,&A22);CHKERRQ(ierr);
ierr = MatCreateVecs(A11,&uvec,NULL);CHKERRQ(ierr);
ierr = MatCreateVecs(A22,&pvec,NULL);CHKERRQ(ierr);
/* perform the scatter from x -> (u,p) */
ierr = VecScatterCreate(x,isu,uvec,NULL,&uscat);CHKERRQ(ierr);
ierr = VecScatterBegin(uscat,x,uvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(uscat,x,uvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterCreate(x,isp,pvec,NULL,&pscat);CHKERRQ(ierr);
ierr = VecScatterBegin(pscat,x,pvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(pscat,x,pvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
PetscPrintf(PETSC_COMM_WORLD,"-- vector vector values --\n");
ierr = VecMin(uvec,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Min(u) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecMax(uvec,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Max(u) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecNorm(uvec,NORM_2,&norm);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Norm(u) = %1.6f \n",(double)norm);CHKERRQ(ierr);
ierr = VecSum(uvec,&sum);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Sum(u) = %1.6f \n",(double)PetscRealPart(sum));CHKERRQ(ierr);
PetscPrintf(PETSC_COMM_WORLD,"-- pressure vector values --\n");
ierr = VecMin(pvec,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Min(p) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecMax(pvec,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Max(p) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecNorm(pvec,NORM_2,&norm);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Norm(p) = %1.6f \n",(double)norm);CHKERRQ(ierr);
ierr = VecSum(pvec,&sum);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Sum(p) = %1.6f \n",(double)PetscRealPart(sum));CHKERRQ(ierr);
PetscPrintf(PETSC_COMM_WORLD,"-- Full vector values --\n");
ierr = VecMin(x,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Min(u,p) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecMax(x,&loc,&max);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Max(u,p) = %1.6f [loc=%D]\n",(double)max,loc);CHKERRQ(ierr);
ierr = VecNorm(x,NORM_2,&norm);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Norm(u,p) = %1.6f \n",(double)norm);CHKERRQ(ierr);
ierr = VecSum(x,&sum);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD," Sum(u,p) = %1.6f \n",(double)PetscRealPart(sum));CHKERRQ(ierr);
ierr = VecScatterDestroy(&uscat);CHKERRQ(ierr);
ierr = VecScatterDestroy(&pscat);CHKERRQ(ierr);
ierr = VecDestroy(&uvec);CHKERRQ(ierr);
ierr = VecDestroy(&pvec);CHKERRQ(ierr);
ierr = MatDestroy(&A11);CHKERRQ(ierr);
ierr = MatDestroy(&A22);CHKERRQ(ierr);
}
ierr = KSPDestroy(&ksp_A);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&b);CHKERRQ(ierr);
ierr = ISDestroy(&isu);CHKERRQ(ierr);
ierr = ISDestroy(&isp);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* Developer Notes: This should be implemented with a MatCreate_SchurComplement() as that is the standard design for new Mat classes. */
PetscErrorCode MatGetSchurComplement_Basic(Mat mat,IS isrow0,IS iscol0,IS isrow1,IS iscol1,MatReuse mreuse,Mat *newmat,MatReuse preuse,Mat *newpmat)
{
PetscErrorCode ierr;
Mat A=0,Ap=0,B=0,C=0,D=0;
PetscFunctionBegin;
PetscValidHeaderSpecific(mat,MAT_CLASSID,1);
PetscValidHeaderSpecific(isrow0,IS_CLASSID,2);
PetscValidHeaderSpecific(iscol0,IS_CLASSID,3);
PetscValidHeaderSpecific(isrow1,IS_CLASSID,4);
PetscValidHeaderSpecific(iscol1,IS_CLASSID,5);
if (mreuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newmat,MAT_CLASSID,7);
if (preuse == MAT_REUSE_MATRIX) PetscValidHeaderSpecific(*newpmat,MAT_CLASSID,9);
PetscValidType(mat,1);
if (mat->factortype) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
if (mreuse != MAT_IGNORE_MATRIX) {
/* Use MatSchurComplement */
if (mreuse == MAT_REUSE_MATRIX) {
ierr = MatSchurComplementGetSubmatrices(*newmat,&A,&Ap,&B,&C,&D);CHKERRQ(ierr);
if (!A || !Ap || !B || !C) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Attempting to reuse matrix but Schur complement matrices unset");
if (A != Ap) SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_ARG_WRONGSTATE,"Preconditioning matrix does not match operator");
ierr = MatDestroy(&Ap);CHKERRQ(ierr); /* get rid of extra reference */
}
ierr = MatGetSubMatrix(mat,isrow0,iscol0,mreuse,&A);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow0,iscol1,mreuse,&B);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol0,mreuse,&C);CHKERRQ(ierr);
ierr = MatGetSubMatrix(mat,isrow1,iscol1,mreuse,&D);CHKERRQ(ierr);
switch (mreuse) {
case MAT_INITIAL_MATRIX:
ierr = MatCreateSchurComplement(A,A,B,C,D,newmat);CHKERRQ(ierr);
break;
case MAT_REUSE_MATRIX:
ierr = MatSchurComplementUpdate(*newmat,A,A,B,C,D,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
break;
default:
SETERRQ(((PetscObject)mat)->comm,PETSC_ERR_SUP,"Unrecognized value of mreuse");
}
}
if (preuse != MAT_IGNORE_MATRIX) {
/* Use the diagonal part of A to form D - C inv(diag(A)) B */
Mat Ad,AdB,S;
Vec diag;
PetscInt i,m,n,mstart,mend;
PetscScalar *x;
/* We could compose these with newpmat so that the matrices can be reused. */
if (!A) {ierr = MatGetSubMatrix(mat,isrow0,iscol0,MAT_INITIAL_MATRIX,&A);CHKERRQ(ierr);}
if (!B) {ierr = MatGetSubMatrix(mat,isrow0,iscol1,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr);}
if (!C) {ierr = MatGetSubMatrix(mat,isrow1,iscol0,MAT_INITIAL_MATRIX,&C);CHKERRQ(ierr);}
if (!D) {ierr = MatGetSubMatrix(mat,isrow1,iscol1,MAT_INITIAL_MATRIX,&D);CHKERRQ(ierr);}
ierr = MatGetVecs(A,&diag,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetDiagonal(A,diag);CHKERRQ(ierr);
ierr = VecReciprocal(diag);CHKERRQ(ierr);
ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr);
/* We need to compute S = D - C inv(diag(A)) B. For row-oriented formats, it is easy to scale the rows of B and
* for column-oriented formats the columns of C can be scaled. Would skip creating a silly diagonal matrix. */
ierr = MatCreate(((PetscObject)A)->comm,&Ad);CHKERRQ(ierr);
ierr = MatSetSizes(Ad,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(Ad,((PetscObject)mat)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(Ad,"diag_");CHKERRQ(ierr);
ierr = MatSetFromOptions(Ad);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(Ad,1,PETSC_NULL);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Ad,1,PETSC_NULL,0,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Ad,&mstart,&mend);CHKERRQ(ierr);
ierr = VecGetArray(diag,&x);CHKERRQ(ierr);
for (i=mstart; i<mend; i++) {
ierr = MatSetValue(Ad,i,i,x[i-mstart],INSERT_VALUES);CHKERRQ(ierr);
}
ierr = VecRestoreArray(diag,&x);CHKERRQ(ierr);
ierr = MatAssemblyBegin(Ad,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Ad,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = VecDestroy(&diag);CHKERRQ(ierr);
ierr = MatMatMult(Ad,B,MAT_INITIAL_MATRIX,1,&AdB);CHKERRQ(ierr);
S = (preuse == MAT_REUSE_MATRIX) ? *newpmat : (Mat)0;
ierr = MatMatMult(C,AdB,preuse,PETSC_DEFAULT,&S);CHKERRQ(ierr);
ierr = MatAYPX(S,-1,D,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
*newpmat = S;
ierr = MatDestroy(&Ad);CHKERRQ(ierr);
ierr = MatDestroy(&AdB);CHKERRQ(ierr);
}
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&C);CHKERRQ(ierr);
ierr = MatDestroy(&D);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
-pc_gasm_print_subdomains\n \n";
/*
Note: This example focuses on setting the subdomains for the GASM
preconditioner for a problem on a 2D rectangular grid. See ex1.c
and ex2.c for more detailed comments on the basic usage of KSP
(including working with matrices and vectors).
The GASM preconditioner is fully parallel. The user-space routine
CreateSubdomains2D that computes the domain decomposition is also parallel
and attempts to generate both subdomains straddling processors and multiple
domains per processor.
This matrix in this linear system arises from the discretized Laplacian,
and thus is not very interesting in terms of experimenting with variants
of the GASM preconditioner.
*/
/*T
Concepts: KSP^Additive Schwarz Method (GASM) with user-defined subdomains
Processors: n
T*/
/*
Include "petscksp.h" so that we can use KSP solvers. Note that this file
automatically includes:
petscsys.h - base PETSc routines petscvec.h - vectors
petscmat.h - matrices
petscis.h - index sets petscksp.h - Krylov subspace methods
petscviewer.h - viewers petscpc.h - preconditioners
*/
#include <petscksp.h>
#include <petscmat.h>
int main(int argc,char **args)
{
Vec x,b,u; /* approx solution, RHS, exact solution */
Mat A,perA; /* linear system matrix */
KSP ksp; /* linear solver context */
PC pc; /* PC context */
PetscInt overlap; /* width of subdomain overlap */
PetscInt m,n; /* mesh dimensions in x- and y- directions */
PetscInt M,N; /* number of subdomains in x- and y- directions */
PetscInt i,j,Ii,J,Istart,Iend;
PetscErrorCode ierr;
PetscMPIInt size;
PetscBool flg;
PetscBool user_set_subdomains;
PetscScalar v, one = 1.0;
MatPartitioning part;
IS coarseparts,fineparts;
IS is,isn,isrows;
MPI_Comm comm;
PetscReal e;
ierr = PetscInitialize(&argc,&args,(char*)0,help);
if (ierr) return ierr;
comm = PETSC_COMM_WORLD;
ierr = MPI_Comm_size(comm,&size);
CHKERRQ(ierr);
ierr = PetscOptionsBegin(PETSC_COMM_WORLD,NULL,"ex62","PC");
CHKERRQ(ierr);
m = 15;
ierr = PetscOptionsInt("-M", "Number of mesh points in the x-direction","PCGASMCreateSubdomains2D",m,&m,NULL);
CHKERRQ(ierr);
n = 17;
ierr = PetscOptionsInt("-N","Number of mesh points in the y-direction","PCGASMCreateSubdomains2D",n,&n,NULL);
CHKERRQ(ierr);
user_set_subdomains = PETSC_FALSE;
ierr = PetscOptionsBool("-user_set_subdomains","Use the user-specified 2D tiling of mesh by subdomains","PCGASMCreateSubdomains2D",user_set_subdomains,&user_set_subdomains,NULL);
CHKERRQ(ierr);
M = 2;
ierr = PetscOptionsInt("-Mdomains","Number of subdomain tiles in the x-direction","PCGASMSetSubdomains2D",M,&M,NULL);
CHKERRQ(ierr);
N = 1;
ierr = PetscOptionsInt("-Ndomains","Number of subdomain tiles in the y-direction","PCGASMSetSubdomains2D",N,&N,NULL);
CHKERRQ(ierr);
overlap = 1;
ierr = PetscOptionsInt("-overlap","Size of tile overlap.","PCGASMSetSubdomains2D",overlap,&overlap,NULL);
CHKERRQ(ierr);
ierr = PetscOptionsEnd();
CHKERRQ(ierr);
/* -------------------------------------------------------------------
Compute the matrix and right-hand-side vector that define
the linear system, Ax = b.
------------------------------------------------------------------- */
/*
Assemble the matrix for the five point stencil, YET AGAIN
*/
ierr = MatCreate(comm,&A);
CHKERRQ(ierr);
ierr = MatSetSizes(A,PETSC_DECIDE,PETSC_DECIDE,m*n,m*n);
CHKERRQ(ierr);
ierr = MatSetFromOptions(A);
CHKERRQ(ierr);
ierr = MatSetUp(A);
CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A,&Istart,&Iend);
CHKERRQ(ierr);
for (Ii=Istart; Ii<Iend; Ii++) {
v = -1.0;
i = Ii/n;
j = Ii - i*n;
if (i>0) {
J = Ii - n;
ierr = MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES);
CHKERRQ(ierr);
}
if (i<m-1) {
J = Ii + n;
ierr = MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES);
CHKERRQ(ierr);
}
if (j>0) {
J = Ii - 1;
ierr = MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES);
CHKERRQ(ierr);
}
if (j<n-1) {
J = Ii + 1;
ierr = MatSetValues(A,1,&Ii,1,&J,&v,INSERT_VALUES);
CHKERRQ(ierr);
}
v = 4.0;
ierr = MatSetValues(A,1,&Ii,1,&Ii,&v,INSERT_VALUES);
CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
/*
Partition the graph of the matrix
*/
ierr = MatPartitioningCreate(comm,&part);
CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,A);
CHKERRQ(ierr);
ierr = MatPartitioningSetType(part,MATPARTITIONINGHIERARCH);
CHKERRQ(ierr);
ierr = MatPartitioningHierarchicalSetNcoarseparts(part,2);
CHKERRQ(ierr);
ierr = MatPartitioningHierarchicalSetNfineparts(part,2);
CHKERRQ(ierr);
ierr = MatPartitioningSetFromOptions(part);
CHKERRQ(ierr);
/* get new processor owner number of each vertex */
ierr = MatPartitioningApply(part,&is);
CHKERRQ(ierr);
/* get coarse parts according to which we rearrange the matrix */
ierr = MatPartitioningHierarchicalGetCoarseparts(part,&coarseparts);
CHKERRQ(ierr);
/* fine parts are used with coarse parts */
ierr = MatPartitioningHierarchicalGetFineparts(part,&fineparts);
CHKERRQ(ierr);
/* get new global number of each old global number */
ierr = ISPartitioningToNumbering(is,&isn);
CHKERRQ(ierr);
ierr = ISBuildTwoSided(is,NULL,&isrows);
CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,isrows,isrows,MAT_INITIAL_MATRIX,&perA);
CHKERRQ(ierr);
ierr = MatCreateVecs(perA,&b,NULL);
CHKERRQ(ierr);
ierr = VecSetFromOptions(b);
CHKERRQ(ierr);
ierr = VecDuplicate(b,&u);
CHKERRQ(ierr);
ierr = VecDuplicate(b,&x);
CHKERRQ(ierr);
ierr = VecSet(u,one);
CHKERRQ(ierr);
ierr = MatMult(perA,u,b);
CHKERRQ(ierr);
ierr = KSPCreate(comm,&ksp);
CHKERRQ(ierr);
/*
Set operators. Here the matrix that defines the linear system
also serves as the preconditioning matrix.
*/
ierr = KSPSetOperators(ksp,perA,perA);
CHKERRQ(ierr);
/*
Set the default preconditioner for this program to be GASM
*/
ierr = KSPGetPC(ksp,&pc);
CHKERRQ(ierr);
ierr = PCSetType(pc,PCGASM);
CHKERRQ(ierr);
ierr = KSPSetFromOptions(ksp);
CHKERRQ(ierr);
/* -------------------------------------------------------------------
Solve the linear system
------------------------------------------------------------------- */
ierr = KSPSolve(ksp,b,x);
CHKERRQ(ierr);
/* -------------------------------------------------------------------
Compare result to the exact solution
------------------------------------------------------------------- */
ierr = VecAXPY(x,-1.0,u);
CHKERRQ(ierr);
ierr = VecNorm(x,NORM_INFINITY, &e);
CHKERRQ(ierr);
flg = PETSC_FALSE;
ierr = PetscOptionsGetBool(NULL,NULL,"-print_error",&flg,NULL);
CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(PETSC_COMM_WORLD, "Infinity norm of the error: %g\n", (double)e);
CHKERRQ(ierr);
}
/*
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
*/
ierr = KSPDestroy(&ksp);
CHKERRQ(ierr);
ierr = VecDestroy(&u);
CHKERRQ(ierr);
ierr = VecDestroy(&x);
CHKERRQ(ierr);
ierr = VecDestroy(&b);
CHKERRQ(ierr);
ierr = MatDestroy(&A);
CHKERRQ(ierr);
ierr = MatDestroy(&perA);
CHKERRQ(ierr);
ierr = ISDestroy(&is);
CHKERRQ(ierr);
ierr = ISDestroy(&coarseparts);
CHKERRQ(ierr);
ierr = ISDestroy(&fineparts);
CHKERRQ(ierr);
ierr = ISDestroy(&isrows);
CHKERRQ(ierr);
ierr = ISDestroy(&isn);
CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);
CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
int main(int argc, char **args)
{
Mat A, L;
AppCtx ctx;
PetscViewer viewer;
PetscErrorCode ierr;
ierr = PetscInitialize(&argc, &args, (char *) 0, help);CHKERRQ(ierr);
ierr = ProcessOptions(&ctx);CHKERRQ(ierr);
/* Load matrix */
ierr = PetscViewerBinaryOpen(PETSC_COMM_WORLD, ctx.matFilename, FILE_MODE_READ, &viewer);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_WORLD, &A);CHKERRQ(ierr);
ierr = MatLoad(A, viewer);CHKERRQ(ierr);
ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr);
/* Make graph Laplacian from matrix */
ierr = MatLaplacian(A, 1.0e-12, &L);CHKERRQ(ierr);
/* Check Laplacian */
PetscReal norm;
Vec x, y;
ierr = MatGetVecs(L, &x, NULL);CHKERRQ(ierr);
ierr = VecDuplicate(x, &y);CHKERRQ(ierr);
ierr = VecSet(x, 1.0);CHKERRQ(ierr);
ierr = MatMult(L, x, y);CHKERRQ(ierr);
ierr = VecNorm(y, NORM_INFINITY, &norm);CHKERRQ(ierr);
if (norm > 1.0e-10) SETERRQ(PetscObjectComm((PetscObject) y), PETSC_ERR_PLIB, "Invalid graph Laplacian");
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
/* Compute Fiedler vector, and perhaps more vectors */
Mat LD;
PetscScalar *a, *realpart, *imagpart, *eigvec, *work, sdummy;
PetscBLASInt bn, bN, lwork, lierr, idummy;
PetscInt n, i;
ierr = MatConvert(L, MATDENSE, MAT_INITIAL_MATRIX, &LD);CHKERRQ(ierr);
ierr = MatGetLocalSize(LD, &n, NULL);CHKERRQ(ierr);
ierr = MatDenseGetArray(LD, &a);CHKERRQ(ierr);
ierr = PetscBLASIntCast(n, &bn);CHKERRQ(ierr);
ierr = PetscBLASIntCast(n, &bN);CHKERRQ(ierr);
ierr = PetscBLASIntCast(5*n,&lwork);CHKERRQ(ierr);
ierr = PetscBLASIntCast(1,&idummy);CHKERRQ(ierr);
ierr = PetscMalloc4(n,PetscScalar,&realpart,n,PetscScalar,&imagpart,n*n,PetscScalar,&eigvec,lwork,PetscScalar,&work);CHKERRQ(ierr);
ierr = PetscFPTrapPush(PETSC_FP_TRAP_OFF);CHKERRQ(ierr);
PetscStackCall("LAPACKgeev", LAPACKgeev_("N","V",&bn,a,&bN,realpart,imagpart,&sdummy,&idummy,eigvec,&bN,work,&lwork,&lierr));
if (lierr) SETERRQ1(PETSC_COMM_SELF, PETSC_ERR_LIB, "Error in LAPACK routine %d", (int) lierr);
ierr = PetscFPTrapPop();CHKERRQ(ierr);
PetscReal *r, *c;
PetscInt *perm;
ierr = PetscMalloc3(n,PetscInt,&perm,n,PetscReal,&r,n,PetscReal,&c);CHKERRQ(ierr);
for (i = 0; i < n; ++i) perm[i] = i;
ierr = PetscSortRealWithPermutation(n,realpart,perm);CHKERRQ(ierr);
for (i = 0; i < n; ++i) {
r[i] = realpart[perm[i]];
c[i] = imagpart[perm[i]];
}
for (i = 0; i < n; ++i) {
realpart[i] = r[i];
imagpart[i] = c[i];
}
/* Output spectrum */
if (ctx.showSpectrum) {
ierr = PetscPrintf(PETSC_COMM_SELF, "Spectrum\n");CHKERRQ(ierr);
for (i = 0; i < n; ++i) {ierr = PetscPrintf(PETSC_COMM_SELF, "%d: Real %g Imag %g\n", i, realpart[i], imagpart[i]);CHKERRQ(ierr);}
}
/* Check lowest eigenvalue and eigenvector */
PetscInt evInd = perm[0];
if ((realpart[0] > 1.0e-12) || (imagpart[0] > 1.0e-12)) SETERRQ(PetscObjectComm((PetscObject) L), PETSC_ERR_PLIB, "Graph Laplacian must have lowest eigenvalue 0");
for (i = 0; i < n; ++i) {
if (fabs(eigvec[evInd*n+i] - eigvec[evInd*n+0]) > 1.0e-10) SETERRQ3(PetscObjectComm((PetscObject) L), PETSC_ERR_PLIB, "Graph Laplacian must have constant lowest eigenvector ev_%d %g != ev_0 %g", i, eigvec[evInd*n+i], eigvec[evInd*n+0]);
}
/* Output Fiedler vector */
evInd = perm[1];
if (ctx.showFiedler) {
ierr = PetscPrintf(PETSC_COMM_SELF, "Fiedler vector, Re{ev} %g\n", realpart[1]);CHKERRQ(ierr);
for (i = 0; i < n; ++i) {ierr = PetscPrintf(PETSC_COMM_SELF, "%d: %g\n", i, eigvec[evInd*n+i]);CHKERRQ(ierr);}
}
/* Construct Fiedler partition */
IS fIS, fIS2;
PetscInt *fperm, *fperm2, pos, neg, posSize = 0;
ierr = PetscMalloc(n * sizeof(PetscInt), &fperm);CHKERRQ(ierr);
for (i = 0; i < n; ++i) {
if (eigvec[evInd*n+i] > 0.0) ++posSize;
}
ierr = PetscMalloc(n * sizeof(PetscInt), &fperm2);CHKERRQ(ierr);
for (i = 0; i < n; ++i) fperm[i] = i;
ierr = PetscSortRealWithPermutation(n, &eigvec[evInd*n], fperm);CHKERRQ(ierr);
for (i = 0; i < n; ++i) fperm2[n-1-i] = fperm[i];
for (i = 0, pos = 0, neg = posSize; i < n; ++i) {
if (eigvec[evInd*n+i] > 0.0) fperm[pos++] = i;
else fperm[neg++] = i;
}
ierr = ISCreateGeneral(PetscObjectComm((PetscObject) L), n, fperm, PETSC_OWN_POINTER, &fIS);CHKERRQ(ierr);
ierr = ISSetPermutation(fIS);CHKERRQ(ierr);
ierr = ISCreateGeneral(PetscObjectComm((PetscObject) L), n, fperm2, PETSC_OWN_POINTER, &fIS2);CHKERRQ(ierr);
ierr = ISSetPermutation(fIS2);CHKERRQ(ierr);
ierr = PetscFree3(perm,r,c);CHKERRQ(ierr);
ierr = PetscFree4(realpart,imagpart,eigvec,work);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(LD, &a);CHKERRQ(ierr);
ierr = MatDestroy(&LD);CHKERRQ(ierr);
ierr = MatDestroy(&L);CHKERRQ(ierr);
/* Permute matrix */
Mat AR, AR2;
ierr = MatPermute(A, fIS, fIS, &AR);CHKERRQ(ierr);
ierr = MatView(A, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatView(AR, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = ISDestroy(&fIS);CHKERRQ(ierr);
ierr = MatPermute(A, fIS2, fIS2, &AR2);CHKERRQ(ierr);
ierr = MatView(AR2, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = ISDestroy(&fIS2);CHKERRQ(ierr);
ierr = MatDestroy(&AR);CHKERRQ(ierr);
AR = AR2;
/* Extract blocks and reorder */
Mat AP, AN, APR, ANR;
IS ispos, isneg, rpermpos, cpermpos, rpermneg, cpermneg;
PetscInt bw, bwr;
ierr = ISCreateStride(PETSC_COMM_SELF, posSize, 0, 1, &ispos);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF, n - posSize, posSize, 1, &isneg);CHKERRQ(ierr);
ierr = MatGetSubMatrix(AR, ispos, ispos, MAT_INITIAL_MATRIX, &AP);CHKERRQ(ierr);
ierr = MatGetSubMatrix(AR, isneg, isneg, MAT_INITIAL_MATRIX, &AN);CHKERRQ(ierr);
ierr = ISDestroy(&ispos);CHKERRQ(ierr);
ierr = ISDestroy(&isneg);CHKERRQ(ierr);
ierr = MatGetOrdering(AP, ctx.matOrdtype, &rpermpos, &cpermpos);CHKERRQ(ierr);
ierr = MatGetOrdering(AN, ctx.matOrdtype, &rpermneg, &cpermneg);CHKERRQ(ierr);
ierr = MatPermute(AP, rpermpos, cpermpos, &APR);CHKERRQ(ierr);
ierr = MatComputeBandwidth(AP, 0.0, &bw);CHKERRQ(ierr);
ierr = MatComputeBandwidth(APR, 0.0, &bwr);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD, "Reduced positive bandwidth from %d to %d\n", bw, bwr);CHKERRQ(ierr);
ierr = MatPermute(AN, rpermneg, cpermneg, &ANR);CHKERRQ(ierr);
ierr = MatComputeBandwidth(AN, 0.0, &bw);CHKERRQ(ierr);
ierr = MatComputeBandwidth(ANR, 0.0, &bwr);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD, "Reduced negative bandwidth from %d to %d\n", bw, bwr);CHKERRQ(ierr);
ierr = MatView(AP, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatView(APR, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatView(AN, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatView(ANR, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
/* Reorder original matrix */
Mat ARR;
IS rperm, cperm;
PetscInt *idx;
const PetscInt *cidx;
ierr = PetscMalloc(n * sizeof(PetscInt), &idx);CHKERRQ(ierr);
ierr = ISGetIndices(rpermpos, &cidx);CHKERRQ(ierr);
for (i = 0; i < posSize; ++i) idx[i] = cidx[i];
ierr = ISRestoreIndices(rpermpos, &cidx);CHKERRQ(ierr);
ierr = ISGetIndices(rpermneg, &cidx);CHKERRQ(ierr);
for (i = posSize; i < n; ++i) idx[i] = cidx[i-posSize] + posSize;
ierr = ISRestoreIndices(rpermneg, &cidx);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF, n, idx, PETSC_OWN_POINTER, &rperm);CHKERRQ(ierr);
ierr = ISSetPermutation(rperm);CHKERRQ(ierr);
ierr = PetscMalloc(n * sizeof(PetscInt), &idx);CHKERRQ(ierr);
ierr = ISGetIndices(cpermpos, &cidx);CHKERRQ(ierr);
for (i = 0; i < posSize; ++i) idx[i] = cidx[i];
ierr = ISRestoreIndices(cpermpos, &cidx);CHKERRQ(ierr);
ierr = ISGetIndices(cpermneg, &cidx);CHKERRQ(ierr);
for (i = posSize; i < n; ++i) idx[i] = cidx[i-posSize] + posSize;
ierr = ISRestoreIndices(cpermneg, &cidx);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF, n, idx, PETSC_OWN_POINTER, &cperm);CHKERRQ(ierr);
ierr = ISSetPermutation(cperm);CHKERRQ(ierr);
ierr = MatPermute(AR, rperm, cperm, &ARR);CHKERRQ(ierr);
ierr = MatView(ARR, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = ISDestroy(&rperm);CHKERRQ(ierr);
ierr = ISDestroy(&cperm);CHKERRQ(ierr);
ierr = ISDestroy(&rpermpos);CHKERRQ(ierr);
ierr = ISDestroy(&cpermpos);CHKERRQ(ierr);
ierr = ISDestroy(&rpermneg);CHKERRQ(ierr);
ierr = ISDestroy(&cpermneg);CHKERRQ(ierr);
ierr = MatDestroy(&AP);CHKERRQ(ierr);
ierr = MatDestroy(&AN);CHKERRQ(ierr);
ierr = MatDestroy(&APR);CHKERRQ(ierr);
ierr = MatDestroy(&ANR);CHKERRQ(ierr);
/* Compare bands */
Mat B, BR;
ierr = MatCreateSubMatrixBanded(A, 50, 0.95, &B);CHKERRQ(ierr);
ierr = MatCreateSubMatrixBanded(ARR, 50, 0.95, &BR);CHKERRQ(ierr);
ierr = MatView(B, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatView(BR, PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&BR);CHKERRQ(ierr);
/* Cleanup */
ierr = MatDestroy(&ARR);CHKERRQ(ierr);
ierr = MatDestroy(&AR);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
int main(int argc,char **args)
{
MatType mtype = MATMPIAIJ; /* matrix format */
Mat A,B; /* matrix */
PetscViewer fd; /* viewer */
char file[PETSC_MAX_PATH_LEN]; /* input file name */
PetscBool flg,viewMats,viewIS,viewVecs;
PetscInt ierr,*nlocal,m,n;
PetscMPIInt rank,size;
MatPartitioning part;
IS is,isn;
Vec xin, xout;
VecScatter scat;
PetscInitialize(&argc,&args,(char*)0,help);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL, "-view_mats", &viewMats);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL, "-view_is", &viewIS);CHKERRQ(ierr);
ierr = PetscOptionsHasName(NULL, "-view_vecs", &viewVecs);CHKERRQ(ierr);
/*
Determine file from which we read the matrix
*/
ierr = PetscOptionsGetString(NULL,"-f",file,PETSC_MAX_PATH_LEN,&flg);CHKERRQ(ierr);
/*
Open binary file. Note that we use FILE_MODE_READ to indicate
reading from this file.
*/
ierr = PetscViewerBinaryOpen(PETSC_COMM_WORLD,file,FILE_MODE_READ,&fd);CHKERRQ(ierr);
/*
Load the matrix and vector; then destroy the viewer.
*/
ierr = MatCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = MatSetType(A,mtype);CHKERRQ(ierr);
ierr = MatLoad(A,fd);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&xin);CHKERRQ(ierr);
ierr = VecLoad(xin,fd);CHKERRQ(ierr);
ierr = PetscViewerDestroy(&fd);CHKERRQ(ierr);
if (viewMats) {
if (!rank) printf("Original matrix:\n");
ierr = MatView(A,PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
}
if (viewVecs) {
if (!rank) printf("Original vector:\n");
ierr = VecView(xin,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Partition the graph of the matrix */
ierr = MatPartitioningCreate(PETSC_COMM_WORLD,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,A);CHKERRQ(ierr);
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* get new processor owner number of each vertex */
ierr = MatPartitioningApply(part,&is);CHKERRQ(ierr);
if (viewIS) {
if (!rank) printf("IS1 - new processor ownership:\n");
ierr = ISView(is,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* get new global number of each old global number */
ierr = ISPartitioningToNumbering(is,&isn);CHKERRQ(ierr);
if (viewIS) {
if (!rank) printf("IS2 - new global numbering:\n");
ierr = ISView(isn,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* get number of new vertices for each processor */
ierr = PetscMalloc(size*sizeof(PetscInt),&nlocal);CHKERRQ(ierr);
ierr = ISPartitioningCount(is,size,nlocal);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
/* get old global number of each new global number */
ierr = ISInvertPermutation(isn,nlocal[rank],&is);CHKERRQ(ierr);
ierr = PetscFree(nlocal);CHKERRQ(ierr);
ierr = ISDestroy(&isn);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
if (viewIS) {
if (!rank) printf("IS3=inv(IS2) - old global number of each new global number:\n");
ierr = ISView(is,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* move the matrix rows to the new processes they have been assigned to by the permutation */
ierr = ISSort(is);CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,is,is,MAT_INITIAL_MATRIX,&B);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
/* move the vector rows to the new processes they have been assigned to */
ierr = MatGetLocalSize(B,&m,&n);CHKERRQ(ierr);
ierr = VecCreateMPI(PETSC_COMM_WORLD,m,PETSC_DECIDE,&xout);CHKERRQ(ierr);
ierr = VecScatterCreate(xin,is,xout,NULL,&scat);CHKERRQ(ierr);
ierr = VecScatterBegin(scat,xin,xout,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd(scat,xin,xout,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterDestroy(&scat);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
if (viewMats) {
if (!rank) printf("Partitioned matrix:\n");
ierr = MatView(B,PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
}
if (viewVecs) {
if (!rank) printf("Mapped vector:\n");
ierr = VecView(xout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
{
PetscInt rstart,i,*nzd,*nzo,nzl,nzmax = 0,*ncols,nrow,j;
Mat J;
const PetscInt *cols;
const PetscScalar *vals;
PetscScalar *nvals;
ierr = MatGetOwnershipRange(B,&rstart,NULL);CHKERRQ(ierr);
ierr = PetscMalloc(2*m*sizeof(PetscInt),&nzd);CHKERRQ(ierr);
ierr = PetscMemzero(nzd,2*m*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMalloc(2*m*sizeof(PetscInt),&nzo);CHKERRQ(ierr);
ierr = PetscMemzero(nzo,2*m*sizeof(PetscInt));CHKERRQ(ierr);
for (i=0; i<m; i++) {
ierr = MatGetRow(B,i+rstart,&nzl,&cols,NULL);CHKERRQ(ierr);
for (j=0; j<nzl; j++) {
if (cols[j] >= rstart && cols[j] < rstart+n) {
nzd[2*i] += 2;
nzd[2*i+1] += 2;
} else {
nzo[2*i] += 2;
nzo[2*i+1] += 2;
}
}
nzmax = PetscMax(nzmax,nzd[2*i]+nzo[2*i]);
ierr = MatRestoreRow(B,i+rstart,&nzl,&cols,NULL);CHKERRQ(ierr);
}
ierr = MatCreateAIJ(PETSC_COMM_WORLD,2*m,2*m,PETSC_DECIDE,PETSC_DECIDE,0,nzd,0,nzo,&J);CHKERRQ(ierr);
ierr = PetscInfo(0,"Created empty Jacobian matrix\n");CHKERRQ(ierr);
ierr = PetscFree(nzd);CHKERRQ(ierr);
ierr = PetscFree(nzo);CHKERRQ(ierr);
ierr = PetscMalloc2(nzmax,PetscInt,&ncols,nzmax,PetscScalar,&nvals);CHKERRQ(ierr);
ierr = PetscMemzero(nvals,nzmax*sizeof(PetscScalar));CHKERRQ(ierr);
for (i=0; i<m; i++) {
ierr = MatGetRow(B,i+rstart,&nzl,&cols,&vals);CHKERRQ(ierr);
for (j=0; j<nzl; j++) {
ncols[2*j] = 2*cols[j];
ncols[2*j+1] = 2*cols[j]+1;
}
nrow = 2*(i+rstart);
ierr = MatSetValues(J,1,&nrow,2*nzl,ncols,nvals,INSERT_VALUES);CHKERRQ(ierr);
nrow = 2*(i+rstart) + 1;
ierr = MatSetValues(J,1,&nrow,2*nzl,ncols,nvals,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(B,i+rstart,&nzl,&cols,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
if (viewMats) {
if (!rank) printf("Jacobian matrix structure:\n");
ierr = MatView(J,PETSC_VIEWER_DRAW_WORLD);CHKERRQ(ierr);
}
ierr = MatDestroy(&J);CHKERRQ(ierr);
ierr = PetscFree2(ncols,nvals);CHKERRQ(ierr);
}
/*
Free work space. All PETSc objects should be destroyed when they
are no longer needed.
*/
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = VecDestroy(&xin);CHKERRQ(ierr);
ierr = VecDestroy(&xout);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
int main(int argc, char *argv[])
{
PetscErrorCode ierr;
Mat A,B,Asub,Bsub;
PetscInt ms,idxrow[3],idxcol[4];
Vec left,right,X,Y,X1,Y1;
IS isrow,iscol;
PetscBool random = PETSC_TRUE;
ierr = PetscInitialize(&argc,&argv,PETSC_NULL,help);CHKERRQ(ierr);
ierr = AssembleMatrix(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = AssembleMatrix(PETSC_COMM_WORLD,&B);CHKERRQ(ierr);
ierr = MatShellSetOperation(B,MATOP_GET_SUBMATRIX,PETSC_NULL);CHKERRQ(ierr);
ierr = MatShellSetOperation(B,MATOP_GET_SUBMATRICES,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A,&ms,PETSC_NULL);CHKERRQ(ierr);
idxrow[0] = ms+1;
idxrow[1] = ms+2;
idxrow[2] = ms+4;
ierr = ISCreateGeneral(PETSC_COMM_WORLD,3,idxrow,PETSC_USE_POINTER,&isrow);CHKERRQ(ierr);
idxcol[0] = ms+1;
idxcol[1] = ms+2;
idxcol[2] = ms+4;
idxcol[3] = ms+5;
ierr = ISCreateGeneral(PETSC_COMM_WORLD,4,idxcol,PETSC_USE_POINTER,&iscol);CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,isrow,iscol,MAT_INITIAL_MATRIX,&Asub);CHKERRQ(ierr);
ierr = MatGetSubMatrix(B,isrow,iscol,MAT_INITIAL_MATRIX,&Bsub);CHKERRQ(ierr);
ierr = MatGetVecs(Asub,&right,&left);CHKERRQ(ierr);
ierr = VecDuplicate(right,&X);CHKERRQ(ierr);
ierr = VecDuplicate(right,&X1);CHKERRQ(ierr);
ierr = VecDuplicate(left,&Y);CHKERRQ(ierr);
ierr = VecDuplicate(left,&Y1);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(PETSC_NULL,"-random",&random,PETSC_NULL);CHKERRQ(ierr);
if (random) {
ierr = VecSetRandom(right,PETSC_NULL);CHKERRQ(ierr);
ierr = VecSetRandom(left,PETSC_NULL);CHKERRQ(ierr);
ierr = VecSetRandom(X,PETSC_NULL);CHKERRQ(ierr);
ierr = VecSetRandom(Y,PETSC_NULL);CHKERRQ(ierr);
ierr = VecSetRandom(X1,PETSC_NULL);CHKERRQ(ierr);
ierr = VecSetRandom(Y1,PETSC_NULL);CHKERRQ(ierr);
} else {
ierr = VecSet(right,1.0);CHKERRQ(ierr);
ierr = VecSet(left,2.0);CHKERRQ(ierr);
ierr = VecSet(X,3.0);CHKERRQ(ierr);
ierr = VecSet(Y,4.0);CHKERRQ(ierr);
ierr = VecSet(X1,3.0);CHKERRQ(ierr);
ierr = VecSet(Y1,4.0);CHKERRQ(ierr);
}
ierr = CheckMatrices(Asub,Bsub,left,right,X,Y,X1,Y1);CHKERRQ(ierr);
ierr = ISDestroy(&isrow);CHKERRQ(ierr);
ierr = ISDestroy(&iscol);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&B);CHKERRQ(ierr);
ierr = MatDestroy(&Asub);CHKERRQ(ierr);
ierr = MatDestroy(&Bsub);CHKERRQ(ierr);
ierr = VecDestroy(&left);CHKERRQ(ierr);
ierr = VecDestroy(&right);CHKERRQ(ierr);
ierr = VecDestroy(&X);CHKERRQ(ierr);
ierr = VecDestroy(&Y);CHKERRQ(ierr);
ierr = VecDestroy(&X1);CHKERRQ(ierr);
ierr = VecDestroy(&Y1);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
PetscErrorCode PCISSetUp(PC pc)
{
PC_IS *pcis = (PC_IS*)(pc->data);
Mat_IS *matis;
PetscErrorCode ierr;
PetscBool flg,issbaij;
Vec counter;
PetscFunctionBegin;
ierr = PetscObjectTypeCompare((PetscObject)pc->pmat,MATIS,&flg);CHKERRQ(ierr);
if (!flg) SETERRQ(PetscObjectComm((PetscObject)pc),PETSC_ERR_ARG_WRONG,"Preconditioner type of Neumann Neumman requires matrix of type MATIS");
matis = (Mat_IS*)pc->pmat->data;
pcis->pure_neumann = matis->pure_neumann;
/* get info on mapping */
ierr = PetscObjectReference((PetscObject)matis->mapping);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&pcis->mapping);CHKERRQ(ierr);
pcis->mapping = matis->mapping;
ierr = ISLocalToGlobalMappingGetSize(pcis->mapping,&pcis->n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetInfo(pcis->mapping,&(pcis->n_neigh),&(pcis->neigh),&(pcis->n_shared),&(pcis->shared));CHKERRQ(ierr);
/* Creating local and global index sets for interior and inteface nodes. */
{
PetscInt n_I;
PetscInt *idx_I_local,*idx_B_local,*idx_I_global,*idx_B_global;
PetscInt *array;
PetscInt i,j;
/* Identifying interior and interface nodes, in local numbering */
ierr = PetscMalloc1(pcis->n,&array);CHKERRQ(ierr);
ierr = PetscMemzero(array,pcis->n*sizeof(PetscInt));CHKERRQ(ierr);
for (i=0;i<pcis->n_neigh;i++)
for (j=0;j<pcis->n_shared[i];j++)
array[pcis->shared[i][j]] += 1;
ierr = PetscMalloc1(pcis->n,&idx_I_local);CHKERRQ(ierr);
ierr = PetscMalloc1(pcis->n,&idx_B_local);CHKERRQ(ierr);
for (i=0, pcis->n_B=0, n_I=0; i<pcis->n; i++) {
if (!array[i]) {
idx_I_local[n_I] = i;
n_I++;
} else {
idx_B_local[pcis->n_B] = i;
pcis->n_B++;
}
}
/* Getting the global numbering */
idx_B_global = idx_I_local + n_I; /* Just avoiding allocating extra memory, since we have vacant space */
idx_I_global = idx_B_local + pcis->n_B;
ierr = ISLocalToGlobalMappingApply(pcis->mapping,pcis->n_B,idx_B_local,idx_B_global);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(pcis->mapping,n_I, idx_I_local,idx_I_global);CHKERRQ(ierr);
/* Creating the index sets. */
ierr = ISCreateGeneral(PETSC_COMM_SELF,pcis->n_B,idx_B_local,PETSC_COPY_VALUES, &pcis->is_B_local);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,pcis->n_B,idx_B_global,PETSC_COPY_VALUES,&pcis->is_B_global);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,n_I,idx_I_local,PETSC_COPY_VALUES, &pcis->is_I_local);CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,n_I,idx_I_global,PETSC_COPY_VALUES,&pcis->is_I_global);CHKERRQ(ierr);
/* Freeing memory and restoring arrays */
ierr = PetscFree(idx_B_local);CHKERRQ(ierr);
ierr = PetscFree(idx_I_local);CHKERRQ(ierr);
ierr = PetscFree(array);CHKERRQ(ierr);
}
/*
Extracting the blocks A_II, A_BI, A_IB and A_BB from A. If the numbering
is such that interior nodes come first than the interface ones, we have
[ | ]
[ A_II | A_IB ]
A = [ | ]
[-----------+------]
[ A_BI | A_BB ]
*/
ierr = MatGetSubMatrix(matis->A,pcis->is_I_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&pcis->A_II);CHKERRQ(ierr);
ierr = MatGetSubMatrix(matis->A,pcis->is_B_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&pcis->A_BB);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
if (!issbaij) {
ierr = MatGetSubMatrix(matis->A,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&pcis->A_IB);CHKERRQ(ierr);
ierr = MatGetSubMatrix(matis->A,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&pcis->A_BI);CHKERRQ(ierr);
} else {
Mat newmat;
ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&newmat);CHKERRQ(ierr);
ierr = MatGetSubMatrix(newmat,pcis->is_I_local,pcis->is_B_local,MAT_INITIAL_MATRIX,&pcis->A_IB);CHKERRQ(ierr);
ierr = MatGetSubMatrix(newmat,pcis->is_B_local,pcis->is_I_local,MAT_INITIAL_MATRIX,&pcis->A_BI);CHKERRQ(ierr);
ierr = MatDestroy(&newmat);CHKERRQ(ierr);
}
/*
Creating work vectors and arrays
*/
ierr = VecDuplicate(matis->x,&pcis->vec1_N);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_N,&pcis->vec2_N);CHKERRQ(ierr);
ierr = VecCreateSeq(PETSC_COMM_SELF,pcis->n-pcis->n_B,&pcis->vec1_D);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_D,&pcis->vec2_D);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_D,&pcis->vec3_D);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_D,&pcis->vec4_D);CHKERRQ(ierr);
ierr = VecCreateSeq(PETSC_COMM_SELF,pcis->n_B,&pcis->vec1_B);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_B,&pcis->vec2_B);CHKERRQ(ierr);
ierr = VecDuplicate(pcis->vec1_B,&pcis->vec3_B);CHKERRQ(ierr);
ierr = MatCreateVecs(pc->pmat,&pcis->vec1_global,0);CHKERRQ(ierr);
ierr = PetscMalloc1(pcis->n,&pcis->work_N);CHKERRQ(ierr);
/* Creating the scatter contexts */
ierr = VecScatterCreate(pcis->vec1_global,pcis->is_I_global,pcis->vec1_D,(IS)0,&pcis->global_to_D);CHKERRQ(ierr);
ierr = VecScatterCreate(pcis->vec1_N,pcis->is_B_local,pcis->vec1_B,(IS)0,&pcis->N_to_B);CHKERRQ(ierr);
ierr = VecScatterCreate(pcis->vec1_global,pcis->is_B_global,pcis->vec1_B,(IS)0,&pcis->global_to_B);CHKERRQ(ierr);
/* Creating scaling "matrix" D */
ierr = PetscOptionsGetBool(((PetscObject)pc)->prefix,"-pc_is_use_stiffness_scaling",&pcis->use_stiffness_scaling,NULL);CHKERRQ(ierr);
if (!pcis->D) {
ierr = VecDuplicate(pcis->vec1_B,&pcis->D);CHKERRQ(ierr);
if (!pcis->use_stiffness_scaling) {
ierr = VecSet(pcis->D,pcis->scaling_factor);CHKERRQ(ierr);
} else {
ierr = MatGetDiagonal(matis->A,pcis->vec1_N);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->N_to_B,pcis->vec1_N,pcis->D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->N_to_B,pcis->vec1_N,pcis->D,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
}
}
ierr = VecCopy(pcis->D,pcis->vec1_B);CHKERRQ(ierr);
ierr = MatCreateVecs(pc->pmat,&counter,0);CHKERRQ(ierr); /* temporary auxiliar vector */
ierr = VecSet(counter,0.0);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->global_to_B,pcis->vec1_B,counter,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->global_to_B,pcis->vec1_B,counter,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr);
ierr = VecScatterBegin(pcis->global_to_B,counter,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecScatterEnd (pcis->global_to_B,counter,pcis->vec1_B,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr);
ierr = VecPointwiseDivide(pcis->D,pcis->D,pcis->vec1_B);CHKERRQ(ierr);
ierr = VecDestroy(&counter);CHKERRQ(ierr);
/* See historical note 01, at the bottom of this file. */
/*
Creating the KSP contexts for the local Dirichlet and Neumann problems.
*/
if (pcis->computesolvers) {
PC pc_ctx;
/* Dirichlet */
ierr = KSPCreate(PETSC_COMM_SELF,&pcis->ksp_D);CHKERRQ(ierr);
ierr = PetscObjectIncrementTabLevel((PetscObject)pcis->ksp_D,(PetscObject)pc,1);CHKERRQ(ierr);
ierr = KSPSetOperators(pcis->ksp_D,pcis->A_II,pcis->A_II);CHKERRQ(ierr);
ierr = KSPSetOptionsPrefix(pcis->ksp_D,"is_localD_");CHKERRQ(ierr);
ierr = KSPGetPC(pcis->ksp_D,&pc_ctx);CHKERRQ(ierr);
ierr = PCSetType(pc_ctx,PCLU);CHKERRQ(ierr);
ierr = KSPSetType(pcis->ksp_D,KSPPREONLY);CHKERRQ(ierr);
ierr = KSPSetFromOptions(pcis->ksp_D);CHKERRQ(ierr);
/* the vectors in the following line are dummy arguments, just telling the KSP the vector size. Values are not used */
ierr = KSPSetUp(pcis->ksp_D);CHKERRQ(ierr);
/* Neumann */
ierr = KSPCreate(PETSC_COMM_SELF,&pcis->ksp_N);CHKERRQ(ierr);
ierr = PetscObjectIncrementTabLevel((PetscObject)pcis->ksp_N,(PetscObject)pc,1);CHKERRQ(ierr);
ierr = KSPSetOperators(pcis->ksp_N,matis->A,matis->A);CHKERRQ(ierr);
ierr = KSPSetOptionsPrefix(pcis->ksp_N,"is_localN_");CHKERRQ(ierr);
ierr = KSPGetPC(pcis->ksp_N,&pc_ctx);CHKERRQ(ierr);
ierr = PCSetType(pc_ctx,PCLU);CHKERRQ(ierr);
ierr = KSPSetType(pcis->ksp_N,KSPPREONLY);CHKERRQ(ierr);
ierr = KSPSetFromOptions(pcis->ksp_N);CHKERRQ(ierr);
{
PetscBool damp_fixed = PETSC_FALSE,
remove_nullspace_fixed = PETSC_FALSE,
set_damping_factor_floating = PETSC_FALSE,
not_damp_floating = PETSC_FALSE,
not_remove_nullspace_floating = PETSC_FALSE;
PetscReal fixed_factor,
floating_factor;
ierr = PetscOptionsGetReal(((PetscObject)pc_ctx)->prefix,"-pc_is_damp_fixed",&fixed_factor,&damp_fixed);CHKERRQ(ierr);
if (!damp_fixed) fixed_factor = 0.0;
ierr = PetscOptionsGetBool(((PetscObject)pc_ctx)->prefix,"-pc_is_damp_fixed",&damp_fixed,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(((PetscObject)pc_ctx)->prefix,"-pc_is_remove_nullspace_fixed",&remove_nullspace_fixed,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetReal(((PetscObject)pc_ctx)->prefix,"-pc_is_set_damping_factor_floating",
&floating_factor,&set_damping_factor_floating);CHKERRQ(ierr);
if (!set_damping_factor_floating) floating_factor = 0.0;
ierr = PetscOptionsGetBool(((PetscObject)pc_ctx)->prefix,"-pc_is_set_damping_factor_floating",&set_damping_factor_floating,NULL);CHKERRQ(ierr);
if (!set_damping_factor_floating) floating_factor = 1.e-12;
ierr = PetscOptionsGetBool(((PetscObject)pc_ctx)->prefix,"-pc_is_not_damp_floating",¬_damp_floating,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(((PetscObject)pc_ctx)->prefix,"-pc_is_not_remove_nullspace_floating",¬_remove_nullspace_floating,NULL);CHKERRQ(ierr);
if (pcis->pure_neumann) { /* floating subdomain */
if (!(not_damp_floating)) {
ierr = PCFactorSetShiftType(pc_ctx,MAT_SHIFT_NONZERO);CHKERRQ(ierr);
ierr = PCFactorSetShiftAmount(pc_ctx,floating_factor);CHKERRQ(ierr);
}
if (!(not_remove_nullspace_floating)) {
MatNullSpace nullsp;
ierr = MatNullSpaceCreate(PETSC_COMM_SELF,PETSC_TRUE,0,NULL,&nullsp);CHKERRQ(ierr);
ierr = KSPSetNullSpace(pcis->ksp_N,nullsp);CHKERRQ(ierr);
ierr = MatNullSpaceDestroy(&nullsp);CHKERRQ(ierr);
}
} else { /* fixed subdomain */
if (damp_fixed) {
ierr = PCFactorSetShiftType(pc_ctx,MAT_SHIFT_NONZERO);CHKERRQ(ierr);
ierr = PCFactorSetShiftAmount(pc_ctx,floating_factor);CHKERRQ(ierr);
}
if (remove_nullspace_fixed) {
MatNullSpace nullsp;
ierr = MatNullSpaceCreate(PETSC_COMM_SELF,PETSC_TRUE,0,NULL,&nullsp);CHKERRQ(ierr);
ierr = KSPSetNullSpace(pcis->ksp_N,nullsp);CHKERRQ(ierr);
ierr = MatNullSpaceDestroy(&nullsp);CHKERRQ(ierr);
}
}
}
/* the vectors in the following line are dummy arguments, just telling the KSP the vector size. Values are not used */
ierr = KSPSetUp(pcis->ksp_N);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*
DMDAGetFaceInterpolation - Gets the interpolation for a face based coarse space
*/
PetscErrorCode DMDAGetFaceInterpolation(DM da,PC_Exotic *exotic,Mat Aglobal,MatReuse reuse,Mat *P)
{
PetscErrorCode ierr;
PetscInt dim,i,j,k,m,n,p,dof,Nint,Nface,Nwire,Nsurf,*Iint,*Isurf,cint = 0,csurf = 0,istart,jstart,kstart,*II,N,c = 0;
PetscInt mwidth,nwidth,pwidth,cnt,mp,np,pp,Ntotal,gl[6],*globals,Ng,*IIint,*IIsurf,Nt;
Mat Xint, Xsurf,Xint_tmp;
IS isint,issurf,is,row,col;
ISLocalToGlobalMapping ltg;
MPI_Comm comm;
Mat A,Aii,Ais,Asi,*Aholder,iAii;
MatFactorInfo info;
PetscScalar *xsurf,*xint;
#if defined(PETSC_USE_DEBUG_foo)
PetscScalar tmp;
#endif
PetscTable ht;
PetscFunctionBegin;
ierr = DMDAGetInfo(da,&dim,0,0,0,&mp,&np,&pp,&dof,0,0,0,0,0);
CHKERRQ(ierr);
if (dof != 1) SETERRQ(PetscObjectComm((PetscObject)da),PETSC_ERR_SUP,"Only for single field problems");
if (dim != 3) SETERRQ(PetscObjectComm((PetscObject)da),PETSC_ERR_SUP,"Only coded for 3d problems");
ierr = DMDAGetCorners(da,0,0,0,&m,&n,&p);
CHKERRQ(ierr);
ierr = DMDAGetGhostCorners(da,&istart,&jstart,&kstart,&mwidth,&nwidth,&pwidth);
CHKERRQ(ierr);
istart = istart ? -1 : 0;
jstart = jstart ? -1 : 0;
kstart = kstart ? -1 : 0;
/*
the columns of P are the interpolation of each coarse grid point (one for each vertex and edge)
to all the local degrees of freedom (this includes the vertices, edges and faces).
Xint are the subset of the interpolation into the interior
Xface are the interpolation onto faces but not into the interior
Xsurf are the interpolation onto the vertices and edges (the surfbasket)
Xint
Symbolically one could write P = (Xface) after interchanging the rows to match the natural ordering on the domain
Xsurf
*/
N = (m - istart)*(n - jstart)*(p - kstart);
Nint = (m-2-istart)*(n-2-jstart)*(p-2-kstart);
Nface = 2*((m-2-istart)*(n-2-jstart) + (m-2-istart)*(p-2-kstart) + (n-2-jstart)*(p-2-kstart));
Nwire = 4*((m-2-istart) + (n-2-jstart) + (p-2-kstart)) + 8;
Nsurf = Nface + Nwire;
ierr = MatCreateSeqDense(MPI_COMM_SELF,Nint,6,NULL,&Xint);
CHKERRQ(ierr);
ierr = MatCreateSeqDense(MPI_COMM_SELF,Nsurf,6,NULL,&Xsurf);
CHKERRQ(ierr);
ierr = MatDenseGetArray(Xsurf,&xsurf);
CHKERRQ(ierr);
/*
Require that all 12 edges and 6 faces have at least one grid point. Otherwise some of the columns of
Xsurf will be all zero (thus making the coarse matrix singular).
*/
if (m-istart < 3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Number of grid points per process in X direction must be at least 3");
if (n-jstart < 3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Number of grid points per process in Y direction must be at least 3");
if (p-kstart < 3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Number of grid points per process in Z direction must be at least 3");
cnt = 0;
for (j=1; j<n-1-jstart; j++) {
for (i=1; i<m-istart-1; i++) xsurf[cnt++ + 0*Nsurf] = 1;
}
for (k=1; k<p-1-kstart; k++) {
for (i=1; i<m-istart-1; i++) xsurf[cnt++ + 1*Nsurf] = 1;
for (j=1; j<n-1-jstart; j++) {
xsurf[cnt++ + 2*Nsurf] = 1;
/* these are the interior nodes */
xsurf[cnt++ + 3*Nsurf] = 1;
}
for (i=1; i<m-istart-1; i++) xsurf[cnt++ + 4*Nsurf] = 1;
}
for (j=1; j<n-1-jstart; j++) {
for (i=1; i<m-istart-1; i++) xsurf[cnt++ + 5*Nsurf] = 1;
}
#if defined(PETSC_USE_DEBUG_foo)
for (i=0; i<Nsurf; i++) {
tmp = 0.0;
for (j=0; j<6; j++) tmp += xsurf[i+j*Nsurf];
if (PetscAbsScalar(tmp-1.0) > 1.e-10) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Wrong Xsurf interpolation at i %D value %g",i,(double)PetscAbsScalar(tmp));
}
#endif
ierr = MatDenseRestoreArray(Xsurf,&xsurf);
CHKERRQ(ierr);
/* ierr = MatView(Xsurf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);*/
/*
I are the indices for all the needed vertices (in global numbering)
Iint are the indices for the interior values, I surf for the surface values
(This is just for the part of the global matrix obtained with MatGetSubMatrix(), it
is NOT the local DMDA ordering.)
IIint and IIsurf are the same as the Iint, Isurf except they are in the global numbering
*/
#define Endpoint(a,start,b) (a == 0 || a == (b-1-start))
ierr = PetscMalloc3(N,&II,Nint,&Iint,Nsurf,&Isurf);
CHKERRQ(ierr);
ierr = PetscMalloc2(Nint,&IIint,Nsurf,&IIsurf);
CHKERRQ(ierr);
for (k=0; k<p-kstart; k++) {
for (j=0; j<n-jstart; j++) {
for (i=0; i<m-istart; i++) {
II[c++] = i + j*mwidth + k*mwidth*nwidth;
if (!Endpoint(i,istart,m) && !Endpoint(j,jstart,n) && !Endpoint(k,kstart,p)) {
IIint[cint] = i + j*mwidth + k*mwidth*nwidth;
Iint[cint++] = i + j*(m-istart) + k*(m-istart)*(n-jstart);
} else {
IIsurf[csurf] = i + j*mwidth + k*mwidth*nwidth;
Isurf[csurf++] = i + j*(m-istart) + k*(m-istart)*(n-jstart);
}
}
}
}
if (c != N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"c != N");
if (cint != Nint) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"cint != Nint");
if (csurf != Nsurf) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"csurf != Nsurf");
ierr = DMGetLocalToGlobalMapping(da,<g);
CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(ltg,N,II,II);
CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(ltg,Nint,IIint,IIint);
CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(ltg,Nsurf,IIsurf,IIsurf);
CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)da,&comm);
CHKERRQ(ierr);
ierr = ISCreateGeneral(comm,N,II,PETSC_COPY_VALUES,&is);
CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,Nint,Iint,PETSC_COPY_VALUES,&isint);
CHKERRQ(ierr);
ierr = ISCreateGeneral(PETSC_COMM_SELF,Nsurf,Isurf,PETSC_COPY_VALUES,&issurf);
CHKERRQ(ierr);
ierr = PetscFree3(II,Iint,Isurf);
CHKERRQ(ierr);
ierr = ISSort(is);
CHKERRQ(ierr);
ierr = MatGetSubMatrices(Aglobal,1,&is,&is,MAT_INITIAL_MATRIX,&Aholder);
CHKERRQ(ierr);
A = *Aholder;
ierr = PetscFree(Aholder);
CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,isint,isint,MAT_INITIAL_MATRIX,&Aii);
CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,isint,issurf,MAT_INITIAL_MATRIX,&Ais);
CHKERRQ(ierr);
ierr = MatGetSubMatrix(A,issurf,isint,MAT_INITIAL_MATRIX,&Asi);
CHKERRQ(ierr);
/*
Solve for the interpolation onto the interior Xint
*/
ierr = MatMatMult(Ais,Xsurf,MAT_INITIAL_MATRIX,PETSC_DETERMINE,&Xint_tmp);
CHKERRQ(ierr);
ierr = MatScale(Xint_tmp,-1.0);
CHKERRQ(ierr);
if (exotic->directSolve) {
ierr = MatGetFactor(Aii,MATSOLVERPETSC,MAT_FACTOR_LU,&iAii);
CHKERRQ(ierr);
ierr = MatFactorInfoInitialize(&info);
CHKERRQ(ierr);
ierr = MatGetOrdering(Aii,MATORDERINGND,&row,&col);
CHKERRQ(ierr);
ierr = MatLUFactorSymbolic(iAii,Aii,row,col,&info);
CHKERRQ(ierr);
ierr = ISDestroy(&row);
CHKERRQ(ierr);
ierr = ISDestroy(&col);
CHKERRQ(ierr);
ierr = MatLUFactorNumeric(iAii,Aii,&info);
CHKERRQ(ierr);
ierr = MatMatSolve(iAii,Xint_tmp,Xint);
CHKERRQ(ierr);
ierr = MatDestroy(&iAii);
CHKERRQ(ierr);
} else {
Vec b,x;
PetscScalar *xint_tmp;
ierr = MatDenseGetArray(Xint,&xint);
CHKERRQ(ierr);
ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,Nint,0,&x);
CHKERRQ(ierr);
ierr = MatDenseGetArray(Xint_tmp,&xint_tmp);
CHKERRQ(ierr);
ierr = VecCreateSeqWithArray(PETSC_COMM_SELF,1,Nint,0,&b);
CHKERRQ(ierr);
ierr = KSPSetOperators(exotic->ksp,Aii,Aii);
CHKERRQ(ierr);
for (i=0; i<6; i++) {
ierr = VecPlaceArray(x,xint+i*Nint);
CHKERRQ(ierr);
ierr = VecPlaceArray(b,xint_tmp+i*Nint);
CHKERRQ(ierr);
ierr = KSPSolve(exotic->ksp,b,x);
CHKERRQ(ierr);
ierr = VecResetArray(x);
CHKERRQ(ierr);
ierr = VecResetArray(b);
CHKERRQ(ierr);
}
ierr = MatDenseRestoreArray(Xint,&xint);
CHKERRQ(ierr);
ierr = MatDenseRestoreArray(Xint_tmp,&xint_tmp);
CHKERRQ(ierr);
ierr = VecDestroy(&x);
CHKERRQ(ierr);
ierr = VecDestroy(&b);
CHKERRQ(ierr);
}
ierr = MatDestroy(&Xint_tmp);
CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG_foo)
ierr = MatDenseGetArray(Xint,&xint);
CHKERRQ(ierr);
for (i=0; i<Nint; i++) {
tmp = 0.0;
for (j=0; j<6; j++) tmp += xint[i+j*Nint];
if (PetscAbsScalar(tmp-1.0) > 1.e-10) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Wrong Xint interpolation at i %D value %g",i,(double)PetscAbsScalar(tmp));
}
ierr = MatDenseRestoreArray(Xint,&xint);
CHKERRQ(ierr);
/* ierr =MatView(Xint,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr); */
#endif
/* total faces */
Ntotal = mp*np*(pp+1) + mp*pp*(np+1) + np*pp*(mp+1);
/*
For each vertex, edge, face on process (in the same orderings as used above) determine its local number including ghost points
*/
cnt = 0;
{
gl[cnt++] = mwidth+1;
}
{
{
gl[cnt++] = mwidth*nwidth+1;
}
{
gl[cnt++] = mwidth*nwidth + mwidth; /* these are the interior nodes */ gl[cnt++] = mwidth*nwidth + mwidth+m-istart-1;
}
{
gl[cnt++] = mwidth*nwidth+mwidth*(n-jstart-1)+1;
}
}
{
gl[cnt++] = mwidth*nwidth*(p-kstart-1) + mwidth+1;
}
/* PetscIntView(6,gl,PETSC_VIEWER_STDOUT_WORLD); */
/* convert that to global numbering and get them on all processes */
ierr = ISLocalToGlobalMappingApply(ltg,6,gl,gl);
CHKERRQ(ierr);
/* PetscIntView(6,gl,PETSC_VIEWER_STDOUT_WORLD); */
ierr = PetscMalloc1(6*mp*np*pp,&globals);
CHKERRQ(ierr);
ierr = MPI_Allgather(gl,6,MPIU_INT,globals,6,MPIU_INT,PetscObjectComm((PetscObject)da));
CHKERRQ(ierr);
/* Number the coarse grid points from 0 to Ntotal */
ierr = MatGetSize(Aglobal,&Nt,NULL);
CHKERRQ(ierr);
ierr = PetscTableCreate(Ntotal/3,Nt+1,&ht);
CHKERRQ(ierr);
for (i=0; i<6*mp*np*pp; i++) {
ierr = PetscTableAddCount(ht,globals[i]+1);
CHKERRQ(ierr);
}
ierr = PetscTableGetCount(ht,&cnt);
CHKERRQ(ierr);
if (cnt != Ntotal) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Hash table size %D not equal to total number coarse grid points %D",cnt,Ntotal);
ierr = PetscFree(globals);
CHKERRQ(ierr);
for (i=0; i<6; i++) {
ierr = PetscTableFind(ht,gl[i]+1,&gl[i]);
CHKERRQ(ierr);
gl[i]--;
}
ierr = PetscTableDestroy(&ht);
CHKERRQ(ierr);
/* PetscIntView(6,gl,PETSC_VIEWER_STDOUT_WORLD); */
/* construct global interpolation matrix */
ierr = MatGetLocalSize(Aglobal,&Ng,NULL);
CHKERRQ(ierr);
if (reuse == MAT_INITIAL_MATRIX) {
ierr = MatCreateAIJ(PetscObjectComm((PetscObject)da),Ng,PETSC_DECIDE,PETSC_DECIDE,Ntotal,Nint+Nsurf,NULL,Nint,NULL,P);
CHKERRQ(ierr);
} else {
ierr = MatZeroEntries(*P);
CHKERRQ(ierr);
}
ierr = MatSetOption(*P,MAT_ROW_ORIENTED,PETSC_FALSE);
CHKERRQ(ierr);
ierr = MatDenseGetArray(Xint,&xint);
CHKERRQ(ierr);
ierr = MatSetValues(*P,Nint,IIint,6,gl,xint,INSERT_VALUES);
CHKERRQ(ierr);
ierr = MatDenseRestoreArray(Xint,&xint);
CHKERRQ(ierr);
ierr = MatDenseGetArray(Xsurf,&xsurf);
CHKERRQ(ierr);
ierr = MatSetValues(*P,Nsurf,IIsurf,6,gl,xsurf,INSERT_VALUES);
CHKERRQ(ierr);
ierr = MatDenseRestoreArray(Xsurf,&xsurf);
CHKERRQ(ierr);
ierr = MatAssemblyBegin(*P,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
ierr = MatAssemblyEnd(*P,MAT_FINAL_ASSEMBLY);
CHKERRQ(ierr);
ierr = PetscFree2(IIint,IIsurf);
CHKERRQ(ierr);
#if defined(PETSC_USE_DEBUG_foo)
{
Vec x,y;
PetscScalar *yy;
ierr = VecCreateMPI(PetscObjectComm((PetscObject)da),Ng,PETSC_DETERMINE,&y);
CHKERRQ(ierr);
ierr = VecCreateMPI(PetscObjectComm((PetscObject)da),PETSC_DETERMINE,Ntotal,&x);
CHKERRQ(ierr);
ierr = VecSet(x,1.0);
CHKERRQ(ierr);
ierr = MatMult(*P,x,y);
CHKERRQ(ierr);
ierr = VecGetArray(y,&yy);
CHKERRQ(ierr);
for (i=0; i<Ng; i++) {
if (PetscAbsScalar(yy[i]-1.0) > 1.e-10) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Wrong p interpolation at i %D value %g",i,(double)PetscAbsScalar(yy[i]));
}
ierr = VecRestoreArray(y,&yy);
CHKERRQ(ierr);
ierr = VecDestroy(x);
CHKERRQ(ierr);
ierr = VecDestroy(y);
CHKERRQ(ierr);
}
#endif
ierr = MatDestroy(&Aii);
CHKERRQ(ierr);
ierr = MatDestroy(&Ais);
CHKERRQ(ierr);
ierr = MatDestroy(&Asi);
CHKERRQ(ierr);
ierr = MatDestroy(&A);
CHKERRQ(ierr);
ierr = ISDestroy(&is);
CHKERRQ(ierr);
ierr = ISDestroy(&isint);
CHKERRQ(ierr);
ierr = ISDestroy(&issurf);
CHKERRQ(ierr);
ierr = MatDestroy(&Xint);
CHKERRQ(ierr);
ierr = MatDestroy(&Xsurf);
CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode SetUpMatrices(AppCtx* user)
{
PetscErrorCode ierr;
PetscInt nele,nen,i;
const PetscInt *ele;
PetscScalar dt=user->dt;
Vec coords;
const PetscScalar *_coords;
PetscScalar x[3],y[3],xx[3],yy[3],w;
PetscInt idx[3];
PetscScalar phi[3],phider[3][2];
PetscScalar eM_0[3][3],eM_2[3][3];
Mat M=user->M;
PetscScalar gamma=user->gamma,theta_c=user->theta_c;
PetscInt m;
PetscInt j,k;
PetscInt row,cols[6],r;
PetscScalar vals[6];
PetscInt n,rstart;
IS isrow,iscol;
PetscFunctionBegin;
/* Get ghosted coordinates */
ierr = DMDAGetGhostedCoordinates(user->da,&coords);CHKERRQ(ierr);
ierr = VecGetArrayRead(coords,&_coords);CHKERRQ(ierr);
/* Get local element info */
ierr = DMDAGetElements(user->da,&nele,&nen,&ele);CHKERRQ(ierr);
for(i=0;i < nele;i++) {
idx[0] = ele[3*i]; idx[1] = ele[3*i+1]; idx[2] = ele[3*i+2];
x[0] = _coords[2*idx[0]]; y[0] = _coords[2*idx[0]+1];
x[1] = _coords[2*idx[1]]; y[1] = _coords[2*idx[1]+1];
x[2] = _coords[2*idx[2]]; y[2] = _coords[2*idx[2]+1];
ierr = PetscMemzero(xx,3*sizeof(PetscScalar));CHKERRQ(ierr);
ierr = PetscMemzero(yy,3*sizeof(PetscScalar));CHKERRQ(ierr);
Gausspoints(xx,yy,&w,x,y);
eM_0[0][0]=eM_0[0][1]=eM_0[0][2]=0.0;
eM_0[1][0]=eM_0[1][1]=eM_0[1][2]=0.0;
eM_0[2][0]=eM_0[2][1]=eM_0[2][2]=0.0;
eM_2[0][0]=eM_2[0][1]=eM_2[0][2]=0.0;
eM_2[1][0]=eM_2[1][1]=eM_2[1][2]=0.0;
eM_2[2][0]=eM_2[2][1]=eM_2[2][2]=0.0;
for(m=0;m<3;m++) {
ierr = PetscMemzero(phi,3*sizeof(PetscScalar));CHKERRQ(ierr);
phider[0][0]=phider[0][1]=0.0;
phider[1][0]=phider[1][1]=0.0;
phider[2][0]=phider[2][1]=0.0;
ShapefunctionsT3(phi,phider,xx[m],yy[m],x,y);
for(j=0;j<3;j++) {
for(k=0;k<3;k++) {
eM_0[k][j] += phi[j]*phi[k]*w;
eM_2[k][j] += phider[j][0]*phider[k][0]*w + phider[j][1]*phider[k][1]*w;
}
}
}
for(r=0;r<3;r++) {
row = 2*idx[r];
cols[0] = 2*idx[0]; vals[0] = dt*eM_2[r][0];
cols[1] = 2*idx[0]+1; vals[1] = eM_0[r][0];
cols[2] = 2*idx[1]; vals[2] = dt*eM_2[r][1];
cols[3] = 2*idx[1]+1; vals[3] = eM_0[r][1];
cols[4] = 2*idx[2]; vals[4] = dt*eM_2[r][2];
cols[5] = 2*idx[2]+1; vals[5] = eM_0[r][2];
/* Insert values in matrix M for 1st dof */
ierr = MatSetValuesLocal(M,1,&row,6,cols,vals,ADD_VALUES);CHKERRQ(ierr);
row = 2*idx[r]+1;
cols[0] = 2*idx[0]; vals[0] = -eM_0[r][0];
cols[1] = 2*idx[0]+1; vals[1] = gamma*eM_2[r][0]-theta_c*eM_0[r][0];
cols[2] = 2*idx[1]; vals[2] = -eM_0[r][1];
cols[3] = 2*idx[1]+1; vals[3] = gamma*eM_2[r][1]-theta_c*eM_0[r][1];
cols[4] = 2*idx[2]; vals[4] = -eM_0[r][2];
cols[5] = 2*idx[2]+1; vals[5] = gamma*eM_2[r][2]-theta_c*eM_2[r][2];
/* Insert values in matrix M for 2nd dof */
ierr = MatSetValuesLocal(M,1,&row,6,cols,vals,ADD_VALUES);CHKERRQ(ierr);
}
}
ierr = DMDARestoreElements(user->da,&nele,&nen,&ele);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(coords,&_coords);CHKERRQ(ierr);
ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* Create ISs to extract matrix M_0 from M */
ierr = MatGetLocalSize(M,&n,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(M,&rstart,PETSC_NULL);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_WORLD,n/2,rstart,2,&isrow);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_WORLD,n/2,rstart+1,2,&iscol);CHKERRQ(ierr);
/* Extract M_0 from M */
ierr = MatGetSubMatrix(M,isrow,iscol,MAT_INITIAL_MATRIX,&user->M_0);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&user->u);CHKERRQ(ierr);
ierr = VecSetSizes(user->u,n/2,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetFromOptions(user->u);CHKERRQ(ierr);
ierr = VecDuplicate(user->u,&user->work1);CHKERRQ(ierr);
ierr = ISDestroy(&isrow);CHKERRQ(ierr);
ierr = ISDestroy(&iscol);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode permutematrix(Mat Ain, Mat Bin, Mat *Aout, Mat *Bout, int **permIndices)
{
PetscErrorCode ierr;
MatPartitioning part;
IS isn, is, iscols;
PetscInt *nlocal,localCols,m,n;
PetscMPIInt size, rank;
MPI_Comm comm;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Ain,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetSize(Ain,&m,&n);CHKERRQ(ierr);
ierr = MatPartitioningCreate(comm,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,Ain);CHKERRQ(ierr);
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* get new processor owner number of each vertex */
ierr = MatPartitioningApply(part,&is);CHKERRQ(ierr);
/* get new global number of each old global number */
ierr = ISPartitioningToNumbering(is,&isn);CHKERRQ(ierr);
ierr = PetscMalloc(size*sizeof(int),&nlocal);CHKERRQ(ierr);
/* get number of new vertices for each processor */
ierr = ISPartitioningCount(is,size,nlocal);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
/* get old global number of each new global number */
ierr = ISInvertPermutation(isn,nlocal[rank],&is);CHKERRQ(ierr);
ierr = ISDestroy(&isn);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
ierr = ISSort(is);CHKERRQ(ierr);
/* If matrix is square, the permutation is applied to rows and columns;
otherwise it is only applied to rows. */
if (m == n) {
iscols = is;
localCols = nlocal[rank];
} else {
PetscInt lowj, highj;
ierr = MatGetOwnershipRangeColumn(Ain,&lowj,&highj);CHKERRQ(ierr);
localCols = highj-lowj;
ierr = ISCreateStride(comm,localCols, lowj, 1, &iscols);CHKERRQ(ierr);
}
/* copy permutation */
if (permIndices) {
const PetscInt *indices;
PetscInt i;
*permIndices = malloc(sizeof(int)*(nlocal[rank]+localCols));
ierr = ISGetIndices(is, &indices);CHKERRQ(ierr);
for (i=0; i<nlocal[rank]; i++) (*permIndices)[i] = indices[i];
ierr = ISRestoreIndices(is, &indices);CHKERRQ(ierr);
ierr = ISGetIndices(iscols, &indices);CHKERRQ(ierr);
for (i=0; i<localCols; i++) (*permIndices)[i+nlocal[rank]] = indices[i];
ierr = ISRestoreIndices(iscols, &indices);CHKERRQ(ierr);
}
ierr = PetscFree(nlocal);CHKERRQ(ierr);
ierr = MatGetSubMatrix(Ain,is,iscols,MAT_INITIAL_MATRIX,Aout);CHKERRQ(ierr);
if (Bin && Bout) {
ierr = MatGetSubMatrix(Bin,is,iscols,MAT_INITIAL_MATRIX,Bout);CHKERRQ(ierr);
}
ierr = ISDestroy(&is);CHKERRQ(ierr);
if (m != n) {
ierr = ISDestroy(&iscols);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
