void PetscMatrix<T>::add_block_matrix(const DenseMatrix<T>& dm,
const std::vector<numeric_index_type>& brows,
const std::vector<numeric_index_type>& bcols)
{
libmesh_assert (this->initialized());
const numeric_index_type n_rows = dm.m();
const numeric_index_type n_cols = dm.n();
const numeric_index_type n_brows = brows.size();
const numeric_index_type n_bcols = bcols.size();
const numeric_index_type blocksize = n_rows / n_brows;
libmesh_assert_equal_to (n_cols / n_bcols, blocksize);
libmesh_assert_equal_to (blocksize*n_brows, n_rows);
libmesh_assert_equal_to (blocksize*n_bcols, n_cols);
PetscErrorCode ierr=0;
#ifndef NDEBUG
PetscInt petsc_blocksize;
ierr = MatGetBlockSize(_mat, &petsc_blocksize);
LIBMESH_CHKERRABORT(ierr);
libmesh_assert_equal_to (blocksize, static_cast<numeric_index_type>(petsc_blocksize));
#endif
// These casts are required for PETSc <= 2.1.5
ierr = MatSetValuesBlocked(_mat,
n_brows, (PetscInt*) &brows[0],
n_bcols, (PetscInt*) &bcols[0],
(PetscScalar*) &dm.get_values()[0],
ADD_VALUES);
LIBMESH_CHKERRABORT(ierr);
}
PETSC_EXTERN PetscErrorCode MatGetFactor_seqsbaij_cholmod(Mat A,MatFactorType ftype,Mat *F)
{
Mat B;
Mat_CHOLMOD *chol;
PetscErrorCode ierr;
PetscInt m=A->rmap->n,n=A->cmap->n,bs;
PetscFunctionBegin;
if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"CHOLMOD cannot do %s factorization with SBAIJ, only %s",
MatFactorTypes[ftype],MatFactorTypes[MAT_FACTOR_CHOLESKY]);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
if (bs != 1) SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"CHOLMOD only supports block size=1, given %D",bs);
/* Create the factorization matrix F */
ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr);
ierr = MatSetSizes(B,PETSC_DECIDE,PETSC_DECIDE,m,n);CHKERRQ(ierr);
ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr);
ierr = MatSeqSBAIJSetPreallocation(B,1,0,NULL);CHKERRQ(ierr);
ierr = PetscNewLog(B,&chol);CHKERRQ(ierr);
chol->Wrap = MatWrapCholmod_seqsbaij;
chol->Destroy = MatDestroy_SeqSBAIJ;
B->spptr = chol;
B->ops->view = MatView_CHOLMOD;
B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_CHOLMOD;
B->ops->destroy = MatDestroy_CHOLMOD;
ierr = PetscObjectComposeFunction((PetscObject)B,"MatFactorGetSolverPackage_C",MatFactorGetSolverPackage_seqsbaij_cholmod);CHKERRQ(ierr);
B->factortype = MAT_FACTOR_CHOLESKY;
B->assembled = PETSC_TRUE; /* required by -ksp_view */
B->preallocated = PETSC_TRUE;
ierr = CholmodStart(B);CHKERRQ(ierr);
*F = B;
PetscFunctionReturn(0);
}
PetscErrorCode DumpCSR(Mat A,PetscInt shift,PetscBool symmetric,PetscBool compressed)
{
MatType type;
PetscInt i,j,nr,bs = 1;
const PetscInt *ia,*ja;
PetscBool done,isseqbaij,isseqsbaij;
PetscErrorCode ierr;
PetscFunctionBeginUser;
ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&isseqbaij);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr);
if (isseqbaij || isseqsbaij) {
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
}
ierr = MatGetType(A,&type);CHKERRQ(ierr);
ierr = MatGetRowIJ(A,shift,symmetric,compressed,&nr,&ia,&ja,&done);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_SELF,"===========================================================\n");CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_SELF,"CSR for %s: shift %D symmetric %D compressed %D\n",type,shift,(PetscInt)symmetric,(PetscInt)compressed);CHKERRQ(ierr);
for (i=0;i<nr;i++) {
ierr = PetscPrintf(PETSC_COMM_SELF,"%D:",i+shift);CHKERRQ(ierr);
for (j=ia[i];j<ia[i+1];j++) {
ierr = PetscPrintf(PETSC_COMM_SELF," %D",ja[j-shift]);CHKERRQ(ierr);
}
ierr = PetscPrintf(PETSC_COMM_SELF,"\n");CHKERRQ(ierr);
}
ierr = MatRestoreRowIJ(A,shift,symmetric,compressed,&nr,&ia,&ja,&done);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void linearSystemPETSc<fullMatrix<double> >::addToSolution(int row, const fullMatrix<double> &val)
{
int blockSize;
_try(MatGetBlockSize(_a, &blockSize));
for (int ii = 0; ii < blockSize; ii++) {
PetscInt i = row * blockSize + ii;
PetscScalar v = val(ii, 0);
VecSetValues(_x, 1, &i, &v, ADD_VALUES);
}
}
PetscErrorCode MatSetLocalToGlobalMapping_IS(Mat A,ISLocalToGlobalMapping rmapping,ISLocalToGlobalMapping cmapping)
{
PetscErrorCode ierr;
PetscInt n,bs;
Mat_IS *is = (Mat_IS*)A->data;
IS from,to;
Vec global;
PetscFunctionBegin;
PetscCheckSameComm(A,1,rmapping,2);
if (rmapping != cmapping) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_INCOMP,"MATIS requires the row and column mappings to be identical");
if (is->mapping) { /* Currenly destroys the objects that will be created by this routine. Is there anything else that should be checked? */
ierr = ISLocalToGlobalMappingDestroy(&is->mapping);CHKERRQ(ierr);
ierr = VecDestroy(&is->x);CHKERRQ(ierr);
ierr = VecDestroy(&is->y);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->ctx);CHKERRQ(ierr);
ierr = MatDestroy(&is->A);CHKERRQ(ierr);
}
ierr = PetscObjectReference((PetscObject)rmapping);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&is->mapping);CHKERRQ(ierr);
is->mapping = rmapping;
/*
ierr = PetscLayoutSetISLocalToGlobalMapping(A->rmap,rmapping);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->cmap,cmapping);CHKERRQ(ierr);
*/
/* Create the local matrix A */
ierr = ISLocalToGlobalMappingGetSize(rmapping,&n);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_SELF,&is->A);CHKERRQ(ierr);
ierr = MatSetSizes(is->A,n,n,n,n);CHKERRQ(ierr);
ierr = MatSetBlockSize(is->A,bs);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(is->A,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(is->A,"is_");CHKERRQ(ierr);
ierr = MatSetFromOptions(is->A);CHKERRQ(ierr);
/* Create the local work vectors */
ierr = VecCreate(PETSC_COMM_SELF,&is->x);CHKERRQ(ierr);
ierr = VecSetBlockSize(is->x,bs);CHKERRQ(ierr);
ierr = VecSetSizes(is->x,n,n);CHKERRQ(ierr);
ierr = VecSetOptionsPrefix(is->x,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = VecAppendOptionsPrefix(is->x,"is_");CHKERRQ(ierr);
ierr = VecSetFromOptions(is->x);CHKERRQ(ierr);
ierr = VecDuplicate(is->x,&is->y);CHKERRQ(ierr);
/* setup the global to local scatter */
ierr = ISCreateStride(PETSC_COMM_SELF,n,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(rmapping,to,&from);CHKERRQ(ierr);
ierr = MatGetVecs(A,&global,NULL);CHKERRQ(ierr);
ierr = VecScatterCreate(global,from,is->x,to,&is->ctx);CHKERRQ(ierr);
ierr = VecDestroy(&global);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatPreallocatorPreallocate_Preallocator(Mat mat, PetscBool fill, Mat A)
{
Mat_Preallocator *p = (Mat_Preallocator *) mat->data;
PetscInt *udnz = NULL, *uonz = NULL;
PetscInt bs;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MatGetBlockSize(mat, &bs);CHKERRQ(ierr);
ierr = MatXAIJSetPreallocation(A, bs, p->dnz, p->onz, udnz, uonz);CHKERRQ(ierr);
ierr = MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR, PETSC_TRUE);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void linearSystemPETSc<fullMatrix<double> >::addToRightHandSide(int row,
const fullMatrix<double> &val)
{
if (!_entriesPreAllocated)
preAllocateEntries();
int blockSize;
_try(MatGetBlockSize(_a, &blockSize));
for (int ii = 0; ii < blockSize; ii++) {
PetscInt i = row * blockSize + ii;
PetscScalar v = val(ii, 0);
VecSetValues(_b, 1, &i, &v, ADD_VALUES);
}
}
PetscErrorCode MatISSetPreallocation_IS(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
{
Mat_IS *matis = (Mat_IS*)(B->data);
PetscSF sf;
PetscInt bs,i,nroots,*rootdata,nleaves,*leafdata,nlocalcols;
const PetscInt *gidxs;
PetscErrorCode ierr;
PetscFunctionBegin;
if (!matis->A) {
SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_SUP,"You should first call MatSetLocalToGlobalMapping");
}
ierr = MatGetLocalSize(B,&nroots,NULL);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&nleaves,&nlocalcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(matis->A,&bs);CHKERRQ(ierr);
ierr = PetscCalloc2(nroots,&rootdata,nleaves,&leafdata);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)B),&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,B->rmap,nleaves,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
if (!d_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += d_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += d_nnz[i];
}
if (!o_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += o_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += o_nnz[i];
}
ierr = PetscSFBcastBegin(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves;i++) {
leafdata[i] = PetscMin(leafdata[i],nlocalcols);
}
ierr = MatSeqAIJSetPreallocation(matis->A,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i*bs]/bs;
}
ierr = MatSeqBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i]-i;
}
ierr = MatSeqSBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree2(rootdata,leafdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
void linearSystemPETSc<fullMatrix<double> >::getFromSolution(int row, fullMatrix<double> &val) const
{
int blockSize;
_try(MatGetBlockSize(_a, &blockSize));
for (int i = 0; i < blockSize; i++) {
int ii = row*blockSize +i;
#ifdef PETSC_USE_COMPLEX
PetscScalar s;
VecGetValues ( _x, 1, &ii, &s);
val(i,0) = s.real();
#else
VecGetValues ( _x, 1, &ii, &val(i,0));
#endif
}
}
PetscErrorCode MatSetUp_Preallocator(Mat A)
{
Mat_Preallocator *p = (Mat_Preallocator *) A->data;
PetscInt m, bs;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
ierr = MatGetLocalSize(A, &m, NULL);CHKERRQ(ierr);
ierr = PetscHSetIJCreate(&p->ht);CHKERRQ(ierr);
ierr = MatGetBlockSize(A, &bs);CHKERRQ(ierr);
ierr = MatStashCreate_Private(PetscObjectComm((PetscObject) A), bs, &A->stash);CHKERRQ(ierr);
ierr = PetscCalloc2(m, &p->dnz, m, &p->onz);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode MatSetValuesBlockedLocal_LocalRef_Scalar(Mat A,PetscInt nrow,const PetscInt irow[],PetscInt ncol,const PetscInt icol[],const PetscScalar y[],InsertMode addv)
{
Mat_LocalRef *lr = (Mat_LocalRef*)A->data;
PetscErrorCode ierr;
PetscInt bs,buf[4096],*irowm,*icolm;
PetscFunctionBegin;
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
IndexSpaceGet(buf,nrow*bs,ncol*bs,irowm,icolm);
BlockIndicesExpand(nrow,irow,bs,irowm);
BlockIndicesExpand(ncol,icol,bs,icolm);
ierr = ISLocalToGlobalMappingApply(A->rmap->mapping,nrow*bs,irowm,irowm);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(A->cmap->mapping,ncol*bs,icolm,icolm);CHKERRQ(ierr);
ierr = (*lr->SetValues)(lr->Top,nrow*bs,irowm,ncol*bs,icolm,y,addv);CHKERRQ(ierr);
IndexSpaceRestore(buf,nrow*bs,ncol*bs,irowm,icolm);
PetscFunctionReturn(0);
}
PetscErrorCode MatDuplicate_IS(Mat mat,MatDuplicateOption op,Mat *newmat)
{
PetscErrorCode ierr;
Mat_IS *matis = (Mat_IS*)(mat->data);
PetscInt bs,m,n,M,N;
Mat B,localmat;
PetscFunctionBegin;
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
ierr = MatGetSize(mat,&M,&N);CHKERRQ(ierr);
ierr = MatGetLocalSize(mat,&m,&n);CHKERRQ(ierr);
ierr = MatCreateIS(PetscObjectComm((PetscObject)mat),bs,m,n,M,N,matis->mapping,&B);CHKERRQ(ierr);
ierr = MatDuplicate(matis->A,op,&localmat);CHKERRQ(ierr);
ierr = MatISSetLocalMat(B,localmat);CHKERRQ(ierr);
ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
*newmat = B;
PetscFunctionReturn(0);
}
PetscErrorCode MatFDColoringCreate_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
{
PetscErrorCode ierr;
PetscInt bs,nis=iscoloring->n,m=mat->rmap->n;
PetscBool isBAIJ;
PetscFunctionBegin;
/* set default brows and bcols for speedup inserting the dense matrix into sparse Jacobian;
bcols is chosen s.t. dy-array takes 50% of memory space as mat */
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr);
if (isBAIJ) {
c->brows = m;
c->bcols = 1;
} else { /* mpiaij matrix */
/* bcols is chosen s.t. dy-array takes 50% of local memory space as mat */
Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data;
Mat_SeqAIJ *spA,*spB;
Mat A,B;
PetscInt nz,brows,bcols;
PetscReal mem;
bs = 1; /* only bs=1 is supported for MPIAIJ matrix */
A = aij->A; spA = (Mat_SeqAIJ*)A->data;
B = aij->B; spB = (Mat_SeqAIJ*)B->data;
nz = spA->nz + spB->nz; /* total local nonzero entries of mat */
mem = nz*(sizeof(PetscScalar) + sizeof(PetscInt)) + 3*m*sizeof(PetscInt);
bcols = (PetscInt)(0.5*mem /(m*sizeof(PetscScalar)));
brows = 1000/bcols;
if (bcols > nis) bcols = nis;
if (brows == 0 || brows > m) brows = m;
c->brows = brows;
c->bcols = bcols;
}
c->M = mat->rmap->N/bs; /* set the global rows and columns and local rows */
c->N = mat->cmap->N/bs;
c->m = mat->rmap->n/bs;
c->rstart = mat->rmap->rstart/bs;
c->ncolors = nis;
PetscFunctionReturn(0);
}
PetscErrorCode PCSetCoordinates_GEO(PC pc, PetscInt ndm, PetscInt a_nloc, PetscReal *coords)
{
PC_MG *mg = (PC_MG*)pc->data;
PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx;
PetscErrorCode ierr;
PetscInt arrsz,bs,my0,kk,ii,nloc,Iend;
Mat Amat = pc->pmat;
PetscFunctionBegin;
PetscValidHeaderSpecific(Amat, MAT_CLASSID, 1);
ierr = MatGetBlockSize(Amat, &bs);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Amat, &my0, &Iend);CHKERRQ(ierr);
nloc = (Iend-my0)/bs;
if (nloc!=a_nloc) SETERRQ2(PetscObjectComm((PetscObject)Amat),PETSC_ERR_ARG_WRONG, "Stokes not supported nloc = %d %d.",a_nloc,nloc);
if ((Iend-my0)%bs!=0) SETERRQ1(PetscObjectComm((PetscObject)Amat),PETSC_ERR_ARG_WRONG, "Bad local size %d.",nloc);
pc_gamg->data_cell_rows = 1;
if (coords==0 && nloc > 0) SETERRQ(PetscObjectComm((PetscObject)Amat),PETSC_ERR_ARG_WRONG, "Need coordinates for pc_gamg_type 'geo'.");
pc_gamg->data_cell_cols = ndm; /* coordinates */
arrsz = nloc*pc_gamg->data_cell_rows*pc_gamg->data_cell_cols;
/* create data - syntactic sugar that should be refactored at some point */
if (pc_gamg->data==0 || (pc_gamg->data_sz != arrsz)) {
ierr = PetscFree(pc_gamg->data);CHKERRQ(ierr);
ierr = PetscMalloc1((arrsz+1), &pc_gamg->data);CHKERRQ(ierr);
}
for (kk=0; kk<arrsz; kk++) pc_gamg->data[kk] = -999.;
pc_gamg->data[arrsz] = -99.;
/* copy data in - column oriented */
for (kk = 0; kk < nloc; kk++) {
for (ii = 0; ii < ndm; ii++) {
pc_gamg->data[ii*nloc + kk] = coords[kk*ndm + ii];
}
}
if (pc_gamg->data[arrsz] != -99.) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"pc_gamg->data[arrsz %D] %g != -99.",arrsz,pc_gamg->data[arrsz]);
pc_gamg->data_sz = arrsz;
PetscFunctionReturn(0);
}
/*@
MatXAIJSetPreallocation - set preallocation for serial and parallel AIJ, BAIJ, and SBAIJ matrices
Collective on Mat
Input Arguments:
+ A - matrix being preallocated
. bs - block size
. dnnz - number of nonzero blocks per block row of diagonal part of parallel matrix
. onnz - number of nonzero blocks per block row of off-diagonal part of parallel matrix
. dnnzu - number of nonzero blocks per block row of upper-triangular part of diagonal part of parallel matrix
- onnzu - number of nonzero blocks per block row of upper-triangular part of off-diagonal part of parallel matrix
Level: beginner
.seealso: MatSeqAIJSetPreallocation(), MatMPIAIJSetPreallocation(), MatSeqBAIJSetPreallocation(), MatMPIBAIJSetPreallocation(), MatSeqSBAIJSetPreallocation(), MatMPISBAIJSetPreallocation(),
PetscSplitOwnership()
@*/
PetscErrorCode MatXAIJSetPreallocation(Mat A,PetscInt bs,const PetscInt dnnz[],const PetscInt onnz[],const PetscInt dnnzu[],const PetscInt onnzu[])
{
PetscErrorCode ierr;
void (*aij)(void);
PetscFunctionBegin;
ierr = MatSetBlockSize(A,bs);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
ierr = MatSeqBAIJSetPreallocation(A,bs,0,dnnz);CHKERRQ(ierr);
ierr = MatMPIBAIJSetPreallocation(A,bs,0,dnnz,0,onnz);CHKERRQ(ierr);
ierr = MatSeqSBAIJSetPreallocation(A,bs,0,dnnzu);CHKERRQ(ierr);
ierr = MatMPISBAIJSetPreallocation(A,bs,0,dnnzu,0,onnzu);CHKERRQ(ierr);
/*
In general, we have to do extra work to preallocate for scalar (AIJ) matrices so we check whether it will do any
good before going on with it.
*/
ierr = PetscObjectQueryFunction((PetscObject)A,"MatMPIAIJSetPreallocation_C",&aij);CHKERRQ(ierr);
if (!aij) {
ierr = PetscObjectQueryFunction((PetscObject)A,"MatSeqAIJSetPreallocation_C",&aij);CHKERRQ(ierr);
}
if (aij) {
if (bs == 1) {
ierr = MatSeqAIJSetPreallocation(A,0,dnnz);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(A,0,dnnz,0,onnz);CHKERRQ(ierr);
} else { /* Convert block-row precallocation to scalar-row */
PetscInt i,m,*sdnnz,*sonnz;
ierr = MatGetLocalSize(A,&m,NULL);CHKERRQ(ierr);
ierr = PetscMalloc2((!!dnnz)*m,&sdnnz,(!!onnz)*m,&sonnz);CHKERRQ(ierr);
for (i=0; i<m; i++) {
if (dnnz) sdnnz[i] = dnnz[i/bs] * bs;
if (onnz) sonnz[i] = onnz[i/bs] * bs;
}
ierr = MatSeqAIJSetPreallocation(A,0,dnnz ? sdnnz : NULL);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(A,0,dnnz ? sdnnz : NULL,0,onnz ? sonnz : NULL);CHKERRQ(ierr);
ierr = PetscFree2(sdnnz,sonnz);CHKERRQ(ierr);
}
}
PetscFunctionReturn(0);
}
PetscErrorCode MatISSetPreallocation_IS(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
{
Mat_IS *matis = (Mat_IS*)(B->data);
PetscInt bs,i,nlocalcols;
PetscErrorCode ierr;
PetscFunctionBegin;
if (!matis->A) SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_SUP,"You should first call MatSetLocalToGlobalMapping");
if (!matis->sf) { /* setup SF if not yet created and allocate rootdata and leafdata */
ierr = MatISComputeSF_Private(B);CHKERRQ(ierr);
}
if (!d_nnz) {
for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = d_nz;
} else {
for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] = d_nnz[i];
}
if (!o_nnz) {
for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] += o_nz;
} else {
for (i=0;i<matis->sf_nroots;i++) matis->sf_rootdata[i] += o_nnz[i];
}
ierr = PetscSFBcastBegin(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,NULL,&nlocalcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(matis->A,&bs);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(matis->sf,MPIU_INT,matis->sf_rootdata,matis->sf_leafdata);CHKERRQ(ierr);
for (i=0;i<matis->sf_nleaves;i++) {
matis->sf_leafdata[i] = PetscMin(matis->sf_leafdata[i],nlocalcols);
}
ierr = MatSeqAIJSetPreallocation(matis->A,0,matis->sf_leafdata);CHKERRQ(ierr);
for (i=0;i<matis->sf_nleaves/bs;i++) {
matis->sf_leafdata[i] = matis->sf_leafdata[i*bs]/bs;
}
ierr = MatSeqBAIJSetPreallocation(matis->A,bs,0,matis->sf_leafdata);CHKERRQ(ierr);
for (i=0;i<matis->sf_nleaves/bs;i++) {
matis->sf_leafdata[i] = matis->sf_leafdata[i]-i;
}
ierr = MatSeqSBAIJSetPreallocation(matis->A,bs,0,matis->sf_leafdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PCGAMGCreateGraph(const Mat Amat, Mat *a_Gmat)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,kk,ncols,nloc,NN,MM,bs;
PetscMPIInt rank, size;
MPI_Comm comm;
Mat Gmat;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Amat, &Istart, &Iend);CHKERRQ(ierr);
ierr = MatGetSize(Amat, &MM, &NN);CHKERRQ(ierr);
ierr = MatGetBlockSize(Amat, &bs);CHKERRQ(ierr);
nloc = (Iend-Istart)/bs;
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
if (bs > 1) {
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz;
/* count nnz, there is sparcity in here so this might not be enough */
ierr = PetscMalloc1(nloc, &d_nnz);CHKERRQ(ierr);
ierr = PetscMalloc1(nloc, &o_nnz);CHKERRQ(ierr);
for (Ii = Istart, jj = 0; Ii < Iend; Ii += bs, jj++) {
d_nnz[jj] = 0;
for (kk=0; kk<bs; kk++) {
ierr = MatGetRow(Amat,Ii+kk,&ncols,0,0);CHKERRQ(ierr);
if (ncols > d_nnz[jj]) {
d_nnz[jj] = ncols; /* very pessimistic but could be too low in theory */
o_nnz[jj] = ncols;
if (d_nnz[jj] > nloc) d_nnz[jj] = nloc;
if (o_nnz[jj] > (NN/bs-nloc)) o_nnz[jj] = NN/bs-nloc;
}
ierr = MatRestoreRow(Amat,Ii+kk,&ncols,0,0);CHKERRQ(ierr);
}
}
/* get scalar copy (norms) of matrix -- AIJ specific!!! */
ierr = MatCreateAIJ(comm, nloc, nloc, PETSC_DETERMINE, PETSC_DETERMINE,0, d_nnz, 0, o_nnz, &Gmat);CHKERRQ(ierr);
ierr = PetscFree(d_nnz);CHKERRQ(ierr);
ierr = PetscFree(o_nnz);CHKERRQ(ierr);
for (Ii = Istart; Ii < Iend; Ii++) {
PetscInt dest_row = Ii/bs;
ierr = MatGetRow(Amat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++) {
PetscInt dest_col = idx[jj]/bs;
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
ierr = MatSetValues(Gmat,1,&dest_row,1,&dest_col,&sv,ADD_VALUES);CHKERRQ(ierr);
}
ierr = MatRestoreRow(Amat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(Gmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Gmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
} else {
/* just copy scalar matrix - abs() not taken here but scaled later */
ierr = MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat);CHKERRQ(ierr);
}
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
*a_Gmat = Gmat;
PetscFunctionReturn(0);
}
PetscErrorCode MatISGetMPIXAIJ_IS(Mat mat, MatReuse reuse, Mat *M)
{
Mat_IS *matis = (Mat_IS*)(mat->data);
Mat local_mat;
/* info on mat */
PetscInt bs,rows,cols,lrows,lcols;
PetscInt local_rows,local_cols;
PetscBool isdense,issbaij,isseqaij;
PetscMPIInt nsubdomains;
/* values insertion */
PetscScalar *array;
/* work */
PetscErrorCode ierr;
PetscFunctionBegin;
/* get info from mat */
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&nsubdomains);CHKERRQ(ierr);
if (nsubdomains == 1) {
if (reuse == MAT_INITIAL_MATRIX) {
ierr = MatDuplicate(matis->A,MAT_COPY_VALUES,&(*M));CHKERRQ(ierr);
} else {
ierr = MatCopy(matis->A,*M,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr);
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
ierr = MatGetLocalSize(mat,&lrows,&lcols);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&local_rows,&local_cols);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQDENSE,&isdense);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
if (reuse == MAT_INITIAL_MATRIX) {
MatType new_mat_type;
PetscBool issbaij_red;
/* determining new matrix type */
ierr = MPIU_Allreduce(&issbaij,&issbaij_red,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
if (issbaij_red) {
new_mat_type = MATSBAIJ;
} else {
if (bs>1) {
new_mat_type = MATBAIJ;
} else {
new_mat_type = MATAIJ;
}
}
ierr = MatCreate(PetscObjectComm((PetscObject)mat),M);CHKERRQ(ierr);
ierr = MatSetSizes(*M,lrows,lcols,rows,cols);CHKERRQ(ierr);
ierr = MatSetBlockSize(*M,bs);CHKERRQ(ierr);
ierr = MatSetType(*M,new_mat_type);CHKERRQ(ierr);
ierr = MatISSetMPIXAIJPreallocation_Private(mat,*M,PETSC_FALSE);CHKERRQ(ierr);
} else {
PetscInt mbs,mrows,mcols,mlrows,mlcols;
/* some checks */
ierr = MatGetBlockSize(*M,&mbs);CHKERRQ(ierr);
ierr = MatGetSize(*M,&mrows,&mcols);CHKERRQ(ierr);
ierr = MatGetLocalSize(*M,&mlrows,&mlcols);CHKERRQ(ierr);
if (mrows != rows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of rows (%d != %d)",rows,mrows);
if (mcols != cols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of cols (%d != %d)",cols,mcols);
if (mlrows != lrows) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of local rows (%d != %d)",lrows,mlrows);
if (mlcols != lcols) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of local cols (%d != %d)",lcols,mlcols);
if (mbs != bs) SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong block size (%d != %d)",bs,mbs);
ierr = MatZeroEntries(*M);CHKERRQ(ierr);
}
if (issbaij) {
ierr = MatConvert(matis->A,MATSEQBAIJ,MAT_INITIAL_MATRIX,&local_mat);CHKERRQ(ierr);
} else {
ierr = PetscObjectReference((PetscObject)matis->A);CHKERRQ(ierr);
local_mat = matis->A;
}
/* Set values */
ierr = MatSetLocalToGlobalMapping(*M,mat->rmap->mapping,mat->cmap->mapping);CHKERRQ(ierr);
if (isdense) { /* special case for dense local matrices */
PetscInt i,*dummy_rows,*dummy_cols;
if (local_rows != local_cols) {
ierr = PetscMalloc2(local_rows,&dummy_rows,local_cols,&dummy_cols);CHKERRQ(ierr);
for (i=0;i<local_rows;i++) dummy_rows[i] = i;
for (i=0;i<local_cols;i++) dummy_cols[i] = i;
} else {
ierr = PetscMalloc1(local_rows,&dummy_rows);CHKERRQ(ierr);
for (i=0;i<local_rows;i++) dummy_rows[i] = i;
dummy_cols = dummy_rows;
}
ierr = MatSetOption(*M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr);
ierr = MatDenseGetArray(local_mat,&array);CHKERRQ(ierr);
ierr = MatSetValuesLocal(*M,local_rows,dummy_rows,local_cols,dummy_cols,array,ADD_VALUES);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(local_mat,&array);CHKERRQ(ierr);
if (dummy_rows != dummy_cols) {
ierr = PetscFree2(dummy_rows,dummy_cols);CHKERRQ(ierr);
} else {
ierr = PetscFree(dummy_rows);CHKERRQ(ierr);
}
} else if (isseqaij) {
PetscInt i,nvtxs,*xadj,*adjncy;
PetscBool done;
ierr = MatGetRowIJ(local_mat,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr);
if (!done) SETERRQ1(PetscObjectComm((PetscObject)local_mat),PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__);
ierr = MatSeqAIJGetArray(local_mat,&array);CHKERRQ(ierr);
for (i=0;i<nvtxs;i++) {
ierr = MatSetValuesLocal(*M,1,&i,xadj[i+1]-xadj[i],adjncy+xadj[i],array+xadj[i],ADD_VALUES);CHKERRQ(ierr);
}
ierr = MatRestoreRowIJ(local_mat,0,PETSC_FALSE,PETSC_FALSE,&nvtxs,(const PetscInt**)&xadj,(const PetscInt**)&adjncy,&done);CHKERRQ(ierr);
if (!done) SETERRQ1(PetscObjectComm((PetscObject)local_mat),PETSC_ERR_PLIB,"Error in MatRestoreRowIJ called in %s\n",__FUNCT__);
ierr = MatSeqAIJRestoreArray(local_mat,&array);CHKERRQ(ierr);
} else { /* very basic values insertion for all other matrix types */
PetscInt i;
for (i=0;i<local_rows;i++) {
PetscInt j;
const PetscInt *local_indices_cols;
ierr = MatGetRow(local_mat,i,&j,&local_indices_cols,(const PetscScalar**)&array);CHKERRQ(ierr);
ierr = MatSetValuesLocal(*M,1,&i,j,local_indices_cols,array,ADD_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(local_mat,i,&j,&local_indices_cols,(const PetscScalar**)&array);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(*M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatDestroy(&local_mat);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
if (isdense) {
ierr = MatSetOption(*M,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PETSC_EXTERN PetscErrorCode MatISSetMPIXAIJPreallocation_Private(Mat A, Mat B, PetscBool maxreduce)
{
Mat_IS *matis = (Mat_IS*)(A->data);
PetscInt *my_dnz,*my_onz,*dnz,*onz,*mat_ranges,*row_ownership;
const PetscInt *global_indices_r,*global_indices_c;
PetscInt i,j,bs,rows,cols;
PetscInt lrows,lcols;
PetscInt local_rows,local_cols;
PetscMPIInt nsubdomains;
PetscBool isdense,issbaij;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)A),&nsubdomains);CHKERRQ(ierr);
ierr = MatGetSize(A,&rows,&cols);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&local_rows,&local_cols);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQDENSE,&isdense);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(A->rmap->mapping,&global_indices_r);CHKERRQ(ierr);
if (A->rmap->mapping != A->cmap->mapping) {
ierr = ISLocalToGlobalMappingGetIndices(A->rmap->mapping,&global_indices_c);CHKERRQ(ierr);
} else {
global_indices_c = global_indices_r;
}
if (issbaij) {
ierr = MatGetRowUpperTriangular(matis->A);CHKERRQ(ierr);
}
/*
An SF reduce is needed to sum up properly on shared rows.
Note that generally preallocation is not exact, since it overestimates nonzeros
*/
if (!matis->sf) { /* setup SF if not yet created and allocate rootdata and leafdata */
ierr = MatISComputeSF_Private(A);CHKERRQ(ierr);
}
ierr = MatGetLocalSize(A,&lrows,&lcols);CHKERRQ(ierr);
ierr = MatPreallocateInitialize(PetscObjectComm((PetscObject)A),lrows,lcols,dnz,onz);CHKERRQ(ierr);
/* All processes need to compute entire row ownership */
ierr = PetscMalloc1(rows,&row_ownership);CHKERRQ(ierr);
ierr = MatGetOwnershipRanges(A,(const PetscInt**)&mat_ranges);CHKERRQ(ierr);
for (i=0;i<nsubdomains;i++) {
for (j=mat_ranges[i];j<mat_ranges[i+1];j++) {
row_ownership[j] = i;
}
}
/*
my_dnz and my_onz contains exact contribution to preallocation from each local mat
then, they will be summed up properly. This way, preallocation is always sufficient
*/
ierr = PetscCalloc2(local_rows,&my_dnz,local_rows,&my_onz);CHKERRQ(ierr);
/* preallocation as a MATAIJ */
if (isdense) { /* special case for dense local matrices */
for (i=0;i<local_rows;i++) {
PetscInt index_row = global_indices_r[i];
for (j=i;j<local_rows;j++) {
PetscInt owner = row_ownership[index_row];
PetscInt index_col = global_indices_c[j];
if (index_col > mat_ranges[owner]-1 && index_col < mat_ranges[owner+1] ) { /* diag block */
my_dnz[i] += 1;
} else { /* offdiag block */
my_onz[i] += 1;
}
/* same as before, interchanging rows and cols */
if (i != j) {
owner = row_ownership[index_col];
if (index_row > mat_ranges[owner]-1 && index_row < mat_ranges[owner+1] ) {
my_dnz[j] += 1;
} else {
my_onz[j] += 1;
}
}
}
}
} else { /* TODO: this could be optimized using MatGetRowIJ */
for (i=0;i<local_rows;i++) {
const PetscInt *cols;
PetscInt ncols,index_row = global_indices_r[i];
ierr = MatGetRow(matis->A,i,&ncols,&cols,NULL);CHKERRQ(ierr);
for (j=0;j<ncols;j++) {
PetscInt owner = row_ownership[index_row];
PetscInt index_col = global_indices_c[cols[j]];
if (index_col > mat_ranges[owner]-1 && index_col < mat_ranges[owner+1] ) { /* diag block */
my_dnz[i] += 1;
} else { /* offdiag block */
my_onz[i] += 1;
}
/* same as before, interchanging rows and cols */
if (issbaij && index_col != index_row) {
owner = row_ownership[index_col];
if (index_row > mat_ranges[owner]-1 && index_row < mat_ranges[owner+1] ) {
my_dnz[cols[j]] += 1;
} else {
my_onz[cols[j]] += 1;
}
}
}
ierr = MatRestoreRow(matis->A,i,&ncols,&cols,NULL);CHKERRQ(ierr);
}
}
ierr = ISLocalToGlobalMappingRestoreIndices(A->rmap->mapping,&global_indices_r);CHKERRQ(ierr);
if (global_indices_c != global_indices_r) {
ierr = ISLocalToGlobalMappingRestoreIndices(A->rmap->mapping,&global_indices_c);CHKERRQ(ierr);
}
ierr = PetscFree(row_ownership);CHKERRQ(ierr);
/* Reduce my_dnz and my_onz */
if (maxreduce) {
ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,my_dnz,dnz,MPI_MAX);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,my_dnz,dnz,MPI_MAX);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,my_onz,onz,MPI_MAX);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,my_onz,onz,MPI_MAX);CHKERRQ(ierr);
} else {
ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,my_dnz,dnz,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,my_dnz,dnz,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(matis->sf,MPIU_INT,my_onz,onz,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(matis->sf,MPIU_INT,my_onz,onz,MPI_SUM);CHKERRQ(ierr);
}
ierr = PetscFree2(my_dnz,my_onz);CHKERRQ(ierr);
/* Resize preallocation if overestimated */
for (i=0;i<lrows;i++) {
dnz[i] = PetscMin(dnz[i],lcols);
onz[i] = PetscMin(onz[i],cols-lcols);
}
/* set preallocation */
ierr = MatMPIAIJSetPreallocation(B,0,dnz,0,onz);CHKERRQ(ierr);
for (i=0;i<lrows/bs;i++) {
dnz[i] = dnz[i*bs]/bs;
onz[i] = onz[i*bs]/bs;
}
ierr = MatMPIBAIJSetPreallocation(B,bs,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatMPISBAIJSetPreallocation(B,bs,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
if (issbaij) {
ierr = MatRestoreRowUpperTriangular(matis->A);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*@
MatCreateLocalRef - Gets a logical reference to a local submatrix, for use in assembly
Not Collective
Input Arguments:
+ A - Full matrix, generally parallel
. isrow - Local index set for the rows
- iscol - Local index set for the columns
Output Arguments:
. newmat - New serial Mat
Level: developer
Notes:
Most will use MatGetLocalSubMatrix() which returns a real matrix corresponding to the local
block if it available, such as with matrix formats that store these blocks separately.
The new matrix forwards MatSetValuesLocal() and MatSetValuesBlockedLocal() to the global system.
In general, it does not define MatMult() or any other functions. Local submatrices can be nested.
.seealso: MatSetValuesLocal(), MatSetValuesBlockedLocal(), MatGetLocalSubMatrix(), MatCreateSubMatrix()
@*/
PetscErrorCode MatCreateLocalRef(Mat A,IS isrow,IS iscol,Mat *newmat)
{
PetscErrorCode ierr;
Mat_LocalRef *lr;
Mat B;
PetscInt m,n;
PetscBool islr;
PetscFunctionBegin;
PetscValidHeaderSpecific(A,MAT_CLASSID,1);
PetscValidHeaderSpecific(isrow,IS_CLASSID,2);
PetscValidHeaderSpecific(iscol,IS_CLASSID,3);
PetscValidPointer(newmat,4);
*newmat = 0;
ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr);
ierr = ISGetLocalSize(isrow,&m);CHKERRQ(ierr);
ierr = ISGetLocalSize(iscol,&n);CHKERRQ(ierr);
ierr = MatSetSizes(B,m,n,m,n);CHKERRQ(ierr);
ierr = PetscObjectChangeTypeName((PetscObject)B,MATLOCALREF);CHKERRQ(ierr);
ierr = MatSetUp(B);CHKERRQ(ierr);
B->ops->destroy = MatDestroy_LocalRef;
ierr = PetscNewLog(B,Mat_LocalRef,&lr);CHKERRQ(ierr);
B->data = (void*)lr;
ierr = PetscObjectTypeCompare((PetscObject)A,MATLOCALREF,&islr);CHKERRQ(ierr);
if (islr) {
Mat_LocalRef *alr = (Mat_LocalRef*)A->data;
lr->Top = alr->Top;
} else {
/* This does not increase the reference count because MatLocalRef is not allowed to live longer than its parent */
lr->Top = A;
}
{
ISLocalToGlobalMapping rltog,cltog;
PetscInt abs,rbs,cbs;
/* We will translate directly to global indices for the top level */
lr->SetValues = MatSetValues;
lr->SetValuesBlocked = MatSetValuesBlocked;
B->ops->setvalueslocal = MatSetValuesLocal_LocalRef_Scalar;
ierr = ISL2GCompose(isrow,A->rmap->mapping,&rltog);CHKERRQ(ierr);
if (isrow == iscol && A->rmap->mapping == A->cmap->mapping) {
ierr = PetscObjectReference((PetscObject)rltog);CHKERRQ(ierr);
cltog = rltog;
} else {
ierr = ISL2GCompose(iscol,A->cmap->mapping,&cltog);CHKERRQ(ierr);
}
ierr = MatSetLocalToGlobalMapping(B,rltog,cltog);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&rltog);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&cltog);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&abs);CHKERRQ(ierr);
ierr = ISGetBlockSize(isrow,&rbs);CHKERRQ(ierr);
ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr);
if (rbs == cbs) { /* submatrix has block structure, so user can insert values with blocked interface */
ierr = PetscLayoutSetBlockSize(B->rmap,rbs);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(B->cmap,cbs);CHKERRQ(ierr);
if (abs != rbs || abs == 1) {
/* Top-level matrix has different block size, so we have to call its scalar insertion interface */
B->ops->setvaluesblockedlocal = MatSetValuesBlockedLocal_LocalRef_Scalar;
} else {
/* Block sizes match so we can forward values to the top level using the block interface */
B->ops->setvaluesblockedlocal = MatSetValuesBlockedLocal_LocalRef_Block;
ierr = ISL2GComposeBlock(isrow,A->rmap->bmapping,&rltog);CHKERRQ(ierr);
if (isrow == iscol && A->rmap->bmapping == A->cmap->bmapping) {
ierr = PetscObjectReference((PetscObject)rltog);CHKERRQ(ierr);
cltog = rltog;
} else {
ierr = ISL2GComposeBlock(iscol,A->cmap->bmapping,&cltog);CHKERRQ(ierr);
}
ierr = MatSetLocalToGlobalMappingBlock(B,rltog,cltog);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&rltog);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&cltog);CHKERRQ(ierr);
}
}
}
*newmat = B;
PetscFunctionReturn(0);
}
PETSC_EXTERN PetscErrorCode MatColoringApply_LF(MatColoring mc,ISColoring *iscoloring)
{
PetscErrorCode ierr;
PetscInt *list,*work,*seq,*coloring,n;
const PetscInt *ria,*rja,*cia,*cja;
PetscInt n1, none,ncolors,i;
PetscBool done;
Mat mat = mc->mat;
Mat mat_seq = mat;
PetscMPIInt size;
MPI_Comm comm;
ISColoring iscoloring_seq;
PetscInt bs = 1,rstart,rend,N_loc,nc;
ISColoringValue *colors_loc;
PetscBool flg1,flg2;
PetscFunctionBegin;
if (mc->dist != 2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"LF may only do distance 2 coloring");
/* this is ugly way to get blocksize but cannot call MatGetBlockSize() because AIJ can have bs > 1 */
ierr = PetscObjectTypeCompare((PetscObject)mat,MATSEQBAIJ,&flg1);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&flg2);CHKERRQ(ierr);
if (flg1 || flg2) {
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
}
ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
if (size > 1) {
/* create a sequential iscoloring on all processors */
ierr = MatGetSeqNonzeroStructure(mat,&mat_seq);CHKERRQ(ierr);
}
ierr = MatGetRowIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&ria,&rja,&done);CHKERRQ(ierr);
ierr = MatGetColumnIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&cia,&cja,&done);CHKERRQ(ierr);
if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Ordering requires IJ");
ierr = MatFDColoringDegreeSequence_Minpack(n,cja,cia,rja,ria,&seq);CHKERRQ(ierr);
ierr = PetscMalloc2(n,&list,4*n,&work);CHKERRQ(ierr);
n1 = n - 1;
none = -1;
MINPACKnumsrt(&n,&n1,seq,&none,list,work+2*n,work+n);
ierr = PetscMalloc1(n,&coloring);CHKERRQ(ierr);
MINPACKseq(&n,cja,cia,rja,ria,list,coloring,&ncolors,work);
ierr = PetscFree2(list,work);CHKERRQ(ierr);
ierr = PetscFree(seq);CHKERRQ(ierr);
ierr = MatRestoreRowIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,NULL,&ria,&rja,&done);CHKERRQ(ierr);
ierr = MatRestoreColumnIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,NULL,&cia,&cja,&done);CHKERRQ(ierr);
/* shift coloring numbers to start at zero and shorten */
if (ncolors > IS_COLORING_MAX-1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Maximum color size exceeded");
{
ISColoringValue *s = (ISColoringValue*) coloring;
for (i=0; i<n; i++) s[i] = (ISColoringValue) (coloring[i]-1);
ierr = MatColoringPatch(mat_seq,ncolors,n,s,iscoloring);CHKERRQ(ierr);
}
if (size > 1) {
ierr = MatDestroySeqNonzeroStructure(&mat_seq);CHKERRQ(ierr);
/* convert iscoloring_seq to a parallel iscoloring */
iscoloring_seq = *iscoloring;
rstart = mat->rmap->rstart/bs;
rend = mat->rmap->rend/bs;
N_loc = rend - rstart; /* number of local nodes */
/* get local colors for each local node */
ierr = PetscMalloc1((N_loc+1),&colors_loc);CHKERRQ(ierr);
for (i=rstart; i<rend; i++) colors_loc[i-rstart] = iscoloring_seq->colors[i];
/* create a parallel iscoloring */
nc = iscoloring_seq->n;
ierr = ISColoringCreate(comm,nc,N_loc,colors_loc,iscoloring);CHKERRQ(ierr);
ierr = ISColoringDestroy(&iscoloring_seq);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
int main(int argc,char **args)
{
PetscMPIInt size;
PetscErrorCode ierr;
Vec x,y,b,s1,s2;
Mat A; /* linear system matrix */
Mat sA,sB,sC; /* symmetric part of the matrices */
PetscInt n,mbs=16,bs=1,nz=3,prob=1,i,j,k1,k2,col[3],lf,block, row,Ii,J,n1,inc;
PetscReal norm1,norm2,rnorm,tol=PETSC_SMALL;
PetscScalar neg_one = -1.0,four=4.0,value[3];
IS perm, iscol;
PetscRandom rdm;
PetscBool doIcc=PETSC_TRUE,equal;
MatInfo minfo1,minfo2;
MatFactorInfo factinfo;
MatType type;
PetscInitialize(&argc,&args,(char*)0,help);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_SUP,"This is a uniprocessor example only!");
ierr = PetscOptionsGetInt(NULL,"-bs",&bs,NULL);CHKERRQ(ierr);
ierr = PetscOptionsGetInt(NULL,"-mbs",&mbs,NULL);CHKERRQ(ierr);
n = mbs*bs;
ierr = MatCreate(PETSC_COMM_SELF,&A);CHKERRQ(ierr);
ierr = MatSetSizes(A,n,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetType(A,MATSEQBAIJ);CHKERRQ(ierr);
ierr = MatSetFromOptions(A);CHKERRQ(ierr);
ierr = MatSeqBAIJSetPreallocation(A,bs,nz,NULL);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_SELF,&sA);CHKERRQ(ierr);
ierr = MatSetSizes(sA,n,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetType(sA,MATSEQSBAIJ);CHKERRQ(ierr);
ierr = MatSetFromOptions(sA);CHKERRQ(ierr);
ierr = MatGetType(sA,&type);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)sA,MATSEQSBAIJ,&doIcc);CHKERRQ(ierr);
ierr = MatSeqSBAIJSetPreallocation(sA,bs,nz,NULL);CHKERRQ(ierr);
ierr = MatSetOption(sA,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE);CHKERRQ(ierr);
/* Test MatGetOwnershipRange() */
ierr = MatGetOwnershipRange(A,&Ii,&J);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(sA,&i,&j);CHKERRQ(ierr);
if (i-Ii || j-J) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatGetOwnershipRange() in MatSBAIJ format\n");CHKERRQ(ierr);
}
/* Assemble matrix */
if (bs == 1) {
ierr = PetscOptionsGetInt(NULL,"-test_problem",&prob,NULL);CHKERRQ(ierr);
if (prob == 1) { /* tridiagonal matrix */
value[0] = -1.0; value[1] = 2.0; value[2] = -1.0;
for (i=1; i<n-1; i++) {
col[0] = i-1; col[1] = i; col[2] = i+1;
ierr = MatSetValues(A,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
}
i = n - 1; col[0]=0; col[1] = n - 2; col[2] = n - 1;
value[0]= 0.1; value[1]=-1; value[2]=2;
ierr = MatSetValues(A,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
i = 0;
col[0] = n-1; col[1] = 1; col[2] = 0;
value[0] = 0.1; value[1] = -1.0; value[2] = 2;
ierr = MatSetValues(A,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
} else if (prob ==2) { /* matrix for the five point stencil */
n1 = (PetscInt) (PetscSqrtReal((PetscReal)n) + 0.001);
if (n1*n1 - n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"sqrt(n) must be a positive interger!");
for (i=0; i<n1; i++) {
for (j=0; j<n1; j++) {
Ii = j + n1*i;
if (i>0) {
J = Ii - n1;
ierr = MatSetValues(A,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
}
if (i<n1-1) {
J = Ii + n1;
ierr = MatSetValues(A,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
}
if (j>0) {
J = Ii - 1;
ierr = MatSetValues(A,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
}
if (j<n1-1) {
J = Ii + 1;
ierr = MatSetValues(A,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&Ii,1,&J,&neg_one,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = MatSetValues(A,1,&Ii,1,&Ii,&four,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&Ii,1,&Ii,&four,INSERT_VALUES);CHKERRQ(ierr);
}
}
}
} else { /* bs > 1 */
for (block=0; block<n/bs; block++) {
/* diagonal blocks */
value[0] = -1.0; value[1] = 4.0; value[2] = -1.0;
for (i=1+block*bs; i<bs-1+block*bs; i++) {
col[0] = i-1; col[1] = i; col[2] = i+1;
ierr = MatSetValues(A,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,3,col,value,INSERT_VALUES);CHKERRQ(ierr);
}
i = bs - 1+block*bs; col[0] = bs - 2+block*bs; col[1] = bs - 1+block*bs;
value[0]=-1.0; value[1]=4.0;
ierr = MatSetValues(A,1,&i,2,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,2,col,value,INSERT_VALUES);CHKERRQ(ierr);
i = 0+block*bs; col[0] = 0+block*bs; col[1] = 1+block*bs;
value[0]=4.0; value[1] = -1.0;
ierr = MatSetValues(A,1,&i,2,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,2,col,value,INSERT_VALUES);CHKERRQ(ierr);
}
/* off-diagonal blocks */
value[0]=-1.0;
for (i=0; i<(n/bs-1)*bs; i++) {
col[0]=i+bs;
ierr = MatSetValues(A,1,&i,1,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&i,1,col,value,INSERT_VALUES);CHKERRQ(ierr);
col[0]=i; row=i+bs;
ierr = MatSetValues(A,1,&row,1,col,value,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(sA,1,&row,1,col,value,INSERT_VALUES);CHKERRQ(ierr);
}
}
ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyBegin(sA,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(sA,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
/* Test MatGetInfo() of A and sA */
ierr = MatGetInfo(A,MAT_LOCAL,&minfo1);CHKERRQ(ierr);
ierr = MatGetInfo(sA,MAT_LOCAL,&minfo2);CHKERRQ(ierr);
/*
printf("A matrix nonzeros (BAIJ format) = %d, allocated nonzeros= %d\n", (int)minfo1.nz_used,(int)minfo1.nz_allocated);
printf("sA matrix nonzeros(SBAIJ format) = %d, allocated nonzeros= %d\n", (int)minfo2.nz_used,(int)minfo2.nz_allocated);
*/
i = (int) (minfo1.nz_used - minfo2.nz_used);
j = (int) (minfo1.nz_allocated - minfo2.nz_allocated);
k1 = (int) (minfo1.nz_allocated - minfo1.nz_used);
k2 = (int) (minfo2.nz_allocated - minfo2.nz_used);
if (i < 0 || j < 0 || k1 < 0 || k2 < 0) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error (compare A and sA): MatGetInfo()\n");CHKERRQ(ierr);
}
/* Test MatDuplicate() */
ierr = MatNorm(A,NORM_FROBENIUS,&norm1);CHKERRQ(ierr);
ierr = MatDuplicate(sA,MAT_COPY_VALUES,&sB);CHKERRQ(ierr);
ierr = MatEqual(sA,sB,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NOTSAMETYPE,"Error in MatDuplicate()");
/* Test MatNorm() */
ierr = MatNorm(A,NORM_FROBENIUS,&norm1);CHKERRQ(ierr);
ierr = MatNorm(sB,NORM_FROBENIUS,&norm2);CHKERRQ(ierr);
rnorm = PetscAbsReal(norm1-norm2)/norm2;
if (rnorm > tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatNorm_FROBENIUS, NormA=%16.14e NormsB=%16.14e\n",norm1,norm2);CHKERRQ(ierr);
}
ierr = MatNorm(A,NORM_INFINITY,&norm1);CHKERRQ(ierr);
ierr = MatNorm(sB,NORM_INFINITY,&norm2);CHKERRQ(ierr);
rnorm = PetscAbsReal(norm1-norm2)/norm2;
if (rnorm > tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatNorm_INFINITY(), NormA=%16.14e NormsB=%16.14e\n",norm1,norm2);CHKERRQ(ierr);
}
ierr = MatNorm(A,NORM_1,&norm1);CHKERRQ(ierr);
ierr = MatNorm(sB,NORM_1,&norm2);CHKERRQ(ierr);
rnorm = PetscAbsReal(norm1-norm2)/norm2;
if (rnorm > tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatNorm_INFINITY(), NormA=%16.14e NormsB=%16.14e\n",norm1,norm2);CHKERRQ(ierr);
}
/* Test MatGetInfo(), MatGetSize(), MatGetBlockSize() */
ierr = MatGetInfo(A,MAT_LOCAL,&minfo1);CHKERRQ(ierr);
ierr = MatGetInfo(sB,MAT_LOCAL,&minfo2);CHKERRQ(ierr);
/*
printf("matrix nonzeros (BAIJ format) = %d, allocated nonzeros= %d\n", (int)minfo1.nz_used,(int)minfo1.nz_allocated);
printf("matrix nonzeros(SBAIJ format) = %d, allocated nonzeros= %d\n", (int)minfo2.nz_used,(int)minfo2.nz_allocated);
*/
i = (int) (minfo1.nz_used - minfo2.nz_used);
j = (int) (minfo1.nz_allocated - minfo2.nz_allocated);
k1 = (int) (minfo1.nz_allocated - minfo1.nz_used);
k2 = (int) (minfo2.nz_allocated - minfo2.nz_used);
if (i < 0 || j < 0 || k1 < 0 || k2 < 0) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error(compare A and sB): MatGetInfo()\n");CHKERRQ(ierr);
}
ierr = MatGetSize(A,&Ii,&J);CHKERRQ(ierr);
ierr = MatGetSize(sB,&i,&j);CHKERRQ(ierr);
if (i-Ii || j-J) {
PetscPrintf(PETSC_COMM_SELF,"Error: MatGetSize()\n");CHKERRQ(ierr);
}
ierr = MatGetBlockSize(A, &Ii);CHKERRQ(ierr);
ierr = MatGetBlockSize(sB, &i);CHKERRQ(ierr);
if (i-Ii) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatGetBlockSize()\n");CHKERRQ(ierr);
}
ierr = PetscRandomCreate(PETSC_COMM_SELF,&rdm);CHKERRQ(ierr);
ierr = PetscRandomSetFromOptions(rdm);CHKERRQ(ierr);
ierr = VecCreateSeq(PETSC_COMM_SELF,n,&x);CHKERRQ(ierr);
ierr = VecDuplicate(x,&s1);CHKERRQ(ierr);
ierr = VecDuplicate(x,&s2);CHKERRQ(ierr);
ierr = VecDuplicate(x,&y);CHKERRQ(ierr);
ierr = VecDuplicate(x,&b);CHKERRQ(ierr);
ierr = VecSetRandom(x,rdm);CHKERRQ(ierr);
/* Test MatDiagonalScale(), MatGetDiagonal(), MatScale() */
#if !defined(PETSC_USE_COMPLEX)
/* Scaling matrix with complex numbers results non-spd matrix,
causing crash of MatForwardSolve() and MatBackwardSolve() */
ierr = MatDiagonalScale(A,x,x);CHKERRQ(ierr);
ierr = MatDiagonalScale(sB,x,x);CHKERRQ(ierr);
ierr = MatMultEqual(A,sB,10,&equal);CHKERRQ(ierr);
if (!equal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_NOTSAMETYPE,"Error in MatDiagonalScale");
ierr = MatGetDiagonal(A,s1);CHKERRQ(ierr);
ierr = MatGetDiagonal(sB,s2);CHKERRQ(ierr);
ierr = VecAXPY(s2,neg_one,s1);CHKERRQ(ierr);
ierr = VecNorm(s2,NORM_1,&norm1);CHKERRQ(ierr);
if (norm1>tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error:MatGetDiagonal(), ||s1-s2||=%G\n",norm1);CHKERRQ(ierr);
}
{
PetscScalar alpha=0.1;
ierr = MatScale(A,alpha);CHKERRQ(ierr);
ierr = MatScale(sB,alpha);CHKERRQ(ierr);
}
#endif
/* Test MatGetRowMaxAbs() */
ierr = MatGetRowMaxAbs(A,s1,NULL);CHKERRQ(ierr);
ierr = MatGetRowMaxAbs(sB,s2,NULL);CHKERRQ(ierr);
ierr = VecNorm(s1,NORM_1,&norm1);CHKERRQ(ierr);
ierr = VecNorm(s2,NORM_1,&norm2);CHKERRQ(ierr);
norm1 -= norm2;
if (norm1<-tol || norm1>tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error:MatGetRowMaxAbs() \n");CHKERRQ(ierr);
}
/* Test MatMult() */
for (i=0; i<40; i++) {
ierr = VecSetRandom(x,rdm);CHKERRQ(ierr);
ierr = MatMult(A,x,s1);CHKERRQ(ierr);
ierr = MatMult(sB,x,s2);CHKERRQ(ierr);
ierr = VecNorm(s1,NORM_1,&norm1);CHKERRQ(ierr);
ierr = VecNorm(s2,NORM_1,&norm2);CHKERRQ(ierr);
norm1 -= norm2;
if (norm1<-tol || norm1>tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error: MatMult(), norm1-norm2: %G\n",norm1);CHKERRQ(ierr);
}
}
/* MatMultAdd() */
for (i=0; i<40; i++) {
ierr = VecSetRandom(x,rdm);CHKERRQ(ierr);
ierr = VecSetRandom(y,rdm);CHKERRQ(ierr);
ierr = MatMultAdd(A,x,y,s1);CHKERRQ(ierr);
ierr = MatMultAdd(sB,x,y,s2);CHKERRQ(ierr);
ierr = VecNorm(s1,NORM_1,&norm1);CHKERRQ(ierr);
ierr = VecNorm(s2,NORM_1,&norm2);CHKERRQ(ierr);
norm1 -= norm2;
if (norm1<-tol || norm1>tol) {
ierr = PetscPrintf(PETSC_COMM_SELF,"Error:MatMultAdd(), norm1-norm2: %G\n",norm1);CHKERRQ(ierr);
}
}
/* Test MatCholeskyFactor(), MatICCFactor() with natural ordering */
ierr = MatGetOrdering(A,MATORDERINGNATURAL,&perm,&iscol);CHKERRQ(ierr);
ierr = ISDestroy(&iscol);CHKERRQ(ierr);
norm1 = tol;
inc = bs;
/* initialize factinfo */
ierr = PetscMemzero(&factinfo,sizeof(MatFactorInfo));CHKERRQ(ierr);
for (lf=-1; lf<10; lf += inc) {
if (lf==-1) { /* Cholesky factor of sB (duplicate sA) */
factinfo.fill = 5.0;
ierr = MatGetFactor(sB,MATSOLVERPETSC,MAT_FACTOR_CHOLESKY,&sC);CHKERRQ(ierr);
ierr = MatCholeskyFactorSymbolic(sC,sB,perm,&factinfo);CHKERRQ(ierr);
} else if (!doIcc) break;
else { /* incomplete Cholesky factor */
factinfo.fill = 5.0;
factinfo.levels = lf;
ierr = MatGetFactor(sB,MATSOLVERPETSC,MAT_FACTOR_ICC,&sC);CHKERRQ(ierr);
ierr = MatICCFactorSymbolic(sC,sB,perm,&factinfo);CHKERRQ(ierr);
}
ierr = MatCholeskyFactorNumeric(sC,sB,&factinfo);CHKERRQ(ierr);
/* MatView(sC, PETSC_VIEWER_DRAW_WORLD); */
/* test MatGetDiagonal on numeric factor */
/*
if (lf == -1) {
ierr = MatGetDiagonal(sC,s1);CHKERRQ(ierr);
printf(" in ex74.c, diag: \n");
ierr = VecView(s1,PETSC_VIEWER_STDOUT_SELF);CHKERRQ(ierr);
}
*/
ierr = MatMult(sB,x,b);CHKERRQ(ierr);
/* test MatForwardSolve() and MatBackwardSolve() */
if (lf == -1) {
ierr = MatForwardSolve(sC,b,s1);CHKERRQ(ierr);
ierr = MatBackwardSolve(sC,s1,s2);CHKERRQ(ierr);
ierr = VecAXPY(s2,neg_one,x);CHKERRQ(ierr);
ierr = VecNorm(s2,NORM_2,&norm2);CHKERRQ(ierr);
if (10*norm1 < norm2) {
ierr = PetscPrintf(PETSC_COMM_SELF,"MatForwardSolve and BackwardSolve: Norm of error=%G, bs=%d\n",norm2,bs);CHKERRQ(ierr);
}
}
/* test MatSolve() */
ierr = MatSolve(sC,b,y);CHKERRQ(ierr);
ierr = MatDestroy(&sC);CHKERRQ(ierr);
/* Check the error */
ierr = VecAXPY(y,neg_one,x);CHKERRQ(ierr);
ierr = VecNorm(y,NORM_2,&norm2);CHKERRQ(ierr);
/* printf("lf: %d, error: %G\n", lf,norm2); */
if (10*norm1 < norm2 && lf-inc != -1) {
ierr = PetscPrintf(PETSC_COMM_SELF,"lf=%D, %D, Norm of error=%G, %G\n",lf-inc,lf,norm1,norm2);CHKERRQ(ierr);
}
norm1 = norm2;
if (norm2 < tol && lf != -1) break;
}
ierr = ISDestroy(&perm);CHKERRQ(ierr);
ierr = MatDestroy(&A);CHKERRQ(ierr);
ierr = MatDestroy(&sB);CHKERRQ(ierr);
ierr = MatDestroy(&sA);CHKERRQ(ierr);
ierr = VecDestroy(&x);CHKERRQ(ierr);
ierr = VecDestroy(&y);CHKERRQ(ierr);
ierr = VecDestroy(&s1);CHKERRQ(ierr);
ierr = VecDestroy(&s2);CHKERRQ(ierr);
ierr = VecDestroy(&b);CHKERRQ(ierr);
ierr = PetscRandomDestroy(&rdm);CHKERRQ(ierr);
ierr = PetscFinalize();
return 0;
}
PetscErrorCode MatFDColoringSetUp_MPIXAIJ(Mat mat,ISColoring iscoloring,MatFDColoring c)
{
PetscErrorCode ierr;
PetscMPIInt size,*ncolsonproc,*disp,nn;
PetscInt i,n,nrows,nrows_i,j,k,m,ncols,col,*rowhit,cstart,cend,colb;
const PetscInt *is,*A_ci,*A_cj,*B_ci,*B_cj,*row=NULL,*ltog=NULL;
PetscInt nis=iscoloring->n,nctot,*cols;
IS *isa;
ISLocalToGlobalMapping map=mat->cmap->mapping;
PetscInt ctype=c->ctype,*spidxA,*spidxB,nz,bs,bs2,spidx;
Mat A,B;
PetscScalar *A_val,*B_val,**valaddrhit;
MatEntry *Jentry;
MatEntry2 *Jentry2;
PetscBool isBAIJ;
PetscInt bcols=c->bcols;
#if defined(PETSC_USE_CTABLE)
PetscTable colmap=NULL;
#else
PetscInt *colmap=NULL; /* local col number of off-diag col */
#endif
PetscFunctionBegin;
if (ctype == IS_COLORING_GHOSTED) {
if (!map) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_INCOMP,"When using ghosted differencing matrix must have local to global mapping provided with MatSetLocalToGlobalMapping");
ierr = ISLocalToGlobalMappingGetIndices(map,<og);CHKERRQ(ierr);
}
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&isBAIJ);CHKERRQ(ierr);
if (isBAIJ) {
Mat_MPIBAIJ *baij=(Mat_MPIBAIJ*)mat->data;
Mat_SeqBAIJ *spA,*spB;
A = baij->A; spA = (Mat_SeqBAIJ*)A->data; A_val = spA->a;
B = baij->B; spB = (Mat_SeqBAIJ*)B->data; B_val = spB->a;
nz = spA->nz + spB->nz; /* total nonzero entries of mat */
if (!baij->colmap) {
ierr = MatCreateColmap_MPIBAIJ_Private(mat);CHKERRQ(ierr);
colmap = baij->colmap;
}
ierr = MatGetColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
ierr = MatGetColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
if (ctype == IS_COLORING_GLOBAL && c->htype[0] == 'd') { /* create vscale for storing dx */
PetscInt *garray;
ierr = PetscMalloc1(B->cmap->n,&garray);CHKERRQ(ierr);
for (i=0; i<baij->B->cmap->n/bs; i++) {
for (j=0; j<bs; j++) {
garray[i*bs+j] = bs*baij->garray[i]+j;
}
}
ierr = VecCreateGhost(PetscObjectComm((PetscObject)mat),mat->cmap->n,PETSC_DETERMINE,B->cmap->n,garray,&c->vscale);CHKERRQ(ierr);
ierr = PetscFree(garray);CHKERRQ(ierr);
}
} else {
Mat_MPIAIJ *aij=(Mat_MPIAIJ*)mat->data;
Mat_SeqAIJ *spA,*spB;
A = aij->A; spA = (Mat_SeqAIJ*)A->data; A_val = spA->a;
B = aij->B; spB = (Mat_SeqAIJ*)B->data; B_val = spB->a;
nz = spA->nz + spB->nz; /* total nonzero entries of mat */
if (!aij->colmap) {
/* Allow access to data structures of local part of matrix
- creates aij->colmap which maps global column number to local number in part B */
ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr);
colmap = aij->colmap;
}
ierr = MatGetColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
ierr = MatGetColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
bs = 1; /* only bs=1 is supported for non MPIBAIJ matrix */
if (ctype == IS_COLORING_GLOBAL && c->htype[0] == 'd') { /* create vscale for storing dx */
ierr = VecCreateGhost(PetscObjectComm((PetscObject)mat),mat->cmap->n,PETSC_DETERMINE,B->cmap->n,aij->garray,&c->vscale);CHKERRQ(ierr);
}
}
m = mat->rmap->n/bs;
cstart = mat->cmap->rstart/bs;
cend = mat->cmap->rend/bs;
ierr = PetscMalloc1(nis,&c->ncolumns);CHKERRQ(ierr);
ierr = PetscMalloc1(nis,&c->columns);CHKERRQ(ierr);
ierr = PetscMalloc1(nis,&c->nrows);CHKERRQ(ierr);
ierr = PetscLogObjectMemory((PetscObject)c,3*nis*sizeof(PetscInt));CHKERRQ(ierr);
if (c->htype[0] == 'd') {
ierr = PetscMalloc1(nz,&Jentry);CHKERRQ(ierr);
ierr = PetscLogObjectMemory((PetscObject)c,nz*sizeof(MatEntry));CHKERRQ(ierr);
c->matentry = Jentry;
} else if (c->htype[0] == 'w') {
ierr = PetscMalloc1(nz,&Jentry2);CHKERRQ(ierr);
ierr = PetscLogObjectMemory((PetscObject)c,nz*sizeof(MatEntry2));CHKERRQ(ierr);
c->matentry2 = Jentry2;
} else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"htype is not supported");
ierr = PetscMalloc2(m+1,&rowhit,m+1,&valaddrhit);CHKERRQ(ierr);
nz = 0;
ierr = ISColoringGetIS(iscoloring,PETSC_IGNORE,&isa);CHKERRQ(ierr);
for (i=0; i<nis; i++) { /* for each local color */
ierr = ISGetLocalSize(isa[i],&n);CHKERRQ(ierr);
ierr = ISGetIndices(isa[i],&is);CHKERRQ(ierr);
c->ncolumns[i] = n; /* local number of columns of this color on this process */
if (n) {
ierr = PetscMalloc1(n,&c->columns[i]);CHKERRQ(ierr);
ierr = PetscLogObjectMemory((PetscObject)c,n*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMemcpy(c->columns[i],is,n*sizeof(PetscInt));CHKERRQ(ierr);
} else {
c->columns[i] = 0;
}
if (ctype == IS_COLORING_GLOBAL) {
/* Determine nctot, the total (parallel) number of columns of this color */
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr);
ierr = PetscMalloc2(size,&ncolsonproc,size,&disp);CHKERRQ(ierr);
/* ncolsonproc[j]: local ncolumns on proc[j] of this color */
ierr = PetscMPIIntCast(n,&nn);CHKERRQ(ierr);
ierr = MPI_Allgather(&nn,1,MPI_INT,ncolsonproc,1,MPI_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
nctot = 0; for (j=0; j<size; j++) nctot += ncolsonproc[j];
if (!nctot) {
ierr = PetscInfo(mat,"Coloring of matrix has some unneeded colors with no corresponding rows\n");CHKERRQ(ierr);
}
disp[0] = 0;
for (j=1; j<size; j++) {
disp[j] = disp[j-1] + ncolsonproc[j-1];
}
/* Get cols, the complete list of columns for this color on each process */
ierr = PetscMalloc1(nctot+1,&cols);CHKERRQ(ierr);
ierr = MPI_Allgatherv((void*)is,n,MPIU_INT,cols,ncolsonproc,disp,MPIU_INT,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
ierr = PetscFree2(ncolsonproc,disp);CHKERRQ(ierr);
} else if (ctype == IS_COLORING_GHOSTED) {
/* Determine local number of columns of this color on this process, including ghost points */
nctot = n;
ierr = PetscMalloc1(nctot+1,&cols);CHKERRQ(ierr);
ierr = PetscMemcpy(cols,is,n*sizeof(PetscInt));CHKERRQ(ierr);
} else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Not provided for this MatFDColoring type");
/* Mark all rows affect by these columns */
ierr = PetscMemzero(rowhit,m*sizeof(PetscInt));CHKERRQ(ierr);
bs2 = bs*bs;
nrows_i = 0;
for (j=0; j<nctot; j++) { /* loop over columns*/
if (ctype == IS_COLORING_GHOSTED) {
col = ltog[cols[j]];
} else {
col = cols[j];
}
if (col >= cstart && col < cend) { /* column is in A, diagonal block of mat */
row = A_cj + A_ci[col-cstart];
nrows = A_ci[col-cstart+1] - A_ci[col-cstart];
nrows_i += nrows;
/* loop over columns of A marking them in rowhit */
for (k=0; k<nrows; k++) {
/* set valaddrhit for part A */
spidx = bs2*spidxA[A_ci[col-cstart] + k];
valaddrhit[*row] = &A_val[spidx];
rowhit[*row++] = col - cstart + 1; /* local column index */
}
} else { /* column is in B, off-diagonal block of mat */
#if defined(PETSC_USE_CTABLE)
ierr = PetscTableFind(colmap,col+1,&colb);CHKERRQ(ierr);
colb--;
#else
colb = colmap[col] - 1; /* local column index */
#endif
if (colb == -1) {
nrows = 0;
} else {
colb = colb/bs;
row = B_cj + B_ci[colb];
nrows = B_ci[colb+1] - B_ci[colb];
}
nrows_i += nrows;
/* loop over columns of B marking them in rowhit */
for (k=0; k<nrows; k++) {
/* set valaddrhit for part B */
spidx = bs2*spidxB[B_ci[colb] + k];
valaddrhit[*row] = &B_val[spidx];
rowhit[*row++] = colb + 1 + cend - cstart; /* local column index */
}
}
}
c->nrows[i] = nrows_i;
if (c->htype[0] == 'd') {
for (j=0; j<m; j++) {
if (rowhit[j]) {
Jentry[nz].row = j; /* local row index */
Jentry[nz].col = rowhit[j] - 1; /* local column index */
Jentry[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
nz++;
}
}
} else { /* c->htype == 'wp' */
for (j=0; j<m; j++) {
if (rowhit[j]) {
Jentry2[nz].row = j; /* local row index */
Jentry2[nz].valaddr = valaddrhit[j]; /* address of mat value for this entry */
nz++;
}
}
}
ierr = PetscFree(cols);CHKERRQ(ierr);
}
if (bcols > 1) { /* reorder Jentry for faster MatFDColoringApply() */
ierr = MatFDColoringSetUpBlocked_AIJ_Private(mat,c,nz);CHKERRQ(ierr);
}
if (isBAIJ) {
ierr = MatRestoreColumnIJ_SeqBAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
ierr = MatRestoreColumnIJ_SeqBAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
ierr = PetscMalloc1(bs*mat->rmap->n,&c->dy);CHKERRQ(ierr);
} else {
ierr = MatRestoreColumnIJ_SeqAIJ_Color(A,0,PETSC_FALSE,PETSC_FALSE,&ncols,&A_ci,&A_cj,&spidxA,NULL);CHKERRQ(ierr);
ierr = MatRestoreColumnIJ_SeqAIJ_Color(B,0,PETSC_FALSE,PETSC_FALSE,&ncols,&B_ci,&B_cj,&spidxB,NULL);CHKERRQ(ierr);
}
ierr = ISColoringRestoreIS(iscoloring,&isa);CHKERRQ(ierr);
ierr = PetscFree2(rowhit,valaddrhit);CHKERRQ(ierr);
if (ctype == IS_COLORING_GHOSTED) {
ierr = ISLocalToGlobalMappingRestoreIndices(map,<og);CHKERRQ(ierr);
}
ierr = PetscInfo3(c,"ncolors %D, brows %D and bcols %D are used.\n",c->ncolors,c->brows,c->bcols);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatISGetMPIXAIJ_IS(Mat mat, MatReuse reuse, Mat *M)
{
Mat_IS *matis = (Mat_IS*)(mat->data);
/* info on mat */
/* ISLocalToGlobalMapping rmapping,cmapping; */
PetscInt bs,rows,cols;
PetscInt lrows,lcols;
PetscInt local_rows,local_cols;
PetscBool isdense,issbaij,issbaij_red;
/* values insertion */
PetscScalar *array;
PetscInt *local_indices,*global_indices;
/* work */
PetscInt i,j,index_row;
PetscErrorCode ierr;
PetscFunctionBegin;
/* MISSING CHECKS
- rectangular case not covered (it is not allowed by MATIS)
*/
/* get info from mat */
/* ierr = MatGetLocalToGlobalMapping(mat,&rmapping,&cmapping);CHKERRQ(ierr); */
ierr = MatGetSize(mat,&rows,&cols);CHKERRQ(ierr);
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&local_rows,&local_cols);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQDENSE,&isdense);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)matis->A,MATSEQSBAIJ,&issbaij);CHKERRQ(ierr);
/* work */
ierr = PetscMalloc1(local_rows,&local_indices);CHKERRQ(ierr);
for (i=0;i<local_rows;i++) local_indices[i]=i;
/* map indices of local mat to global values */
ierr = PetscMalloc(PetscMax(local_cols,local_rows)*sizeof(*global_indices),&global_indices);CHKERRQ(ierr);
/* ierr = ISLocalToGlobalMappingApply(rmapping,local_rows,local_indices,global_indices);CHKERRQ(ierr); */
ierr = ISLocalToGlobalMappingApply(matis->mapping,local_rows,local_indices,global_indices);CHKERRQ(ierr);
if (issbaij) {
ierr = MatGetRowUpperTriangular(matis->A);CHKERRQ(ierr);
}
if (reuse == MAT_INITIAL_MATRIX) {
Mat new_mat;
MatType new_mat_type;
Vec vec_dnz,vec_onz;
PetscScalar *my_dnz,*my_onz;
PetscInt *dnz,*onz,*mat_ranges,*row_ownership;
PetscInt index_col,owner;
PetscMPIInt nsubdomains;
/* determining new matrix type */
ierr = MPI_Allreduce(&issbaij,&issbaij_red,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr);
if (issbaij_red) {
new_mat_type = MATSBAIJ;
} else {
if (bs>1) {
new_mat_type = MATBAIJ;
} else {
new_mat_type = MATAIJ;
}
}
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&nsubdomains);CHKERRQ(ierr);
ierr = MatCreate(PetscObjectComm((PetscObject)mat),&new_mat);CHKERRQ(ierr);
ierr = MatSetSizes(new_mat,PETSC_DECIDE,PETSC_DECIDE,rows,cols);CHKERRQ(ierr);
ierr = MatSetBlockSize(new_mat,bs);CHKERRQ(ierr);
ierr = MatSetType(new_mat,new_mat_type);CHKERRQ(ierr);
ierr = MatSetUp(new_mat);CHKERRQ(ierr);
ierr = MatGetLocalSize(new_mat,&lrows,&lcols);CHKERRQ(ierr);
/*
preallocation
*/
ierr = MatPreallocateInitialize(PetscObjectComm((PetscObject)new_mat),lrows,lcols,dnz,onz);CHKERRQ(ierr);
/*
Some vectors are needed to sum up properly on shared interface dofs.
Preallocation macros cannot do the job.
Note that preallocation is not exact, since it overestimates nonzeros
*/
ierr = MatCreateVecs(new_mat,NULL,&vec_dnz);CHKERRQ(ierr);
/* ierr = VecSetLocalToGlobalMapping(vec_dnz,rmapping);CHKERRQ(ierr); */
ierr = VecSetLocalToGlobalMapping(vec_dnz,matis->mapping);CHKERRQ(ierr);
ierr = VecDuplicate(vec_dnz,&vec_onz);CHKERRQ(ierr);
/* All processes need to compute entire row ownership */
ierr = PetscMalloc1(rows,&row_ownership);CHKERRQ(ierr);
ierr = MatGetOwnershipRanges(new_mat,(const PetscInt**)&mat_ranges);CHKERRQ(ierr);
for (i=0;i<nsubdomains;i++) {
for (j=mat_ranges[i];j<mat_ranges[i+1];j++) {
row_ownership[j]=i;
}
}
/*
my_dnz and my_onz contains exact contribution to preallocation from each local mat
then, they will be summed up properly. This way, preallocation is always sufficient
*/
ierr = PetscMalloc1(local_rows,&my_dnz);CHKERRQ(ierr);
ierr = PetscMalloc1(local_rows,&my_onz);CHKERRQ(ierr);
ierr = PetscMemzero(my_dnz,local_rows*sizeof(*my_dnz));CHKERRQ(ierr);
ierr = PetscMemzero(my_onz,local_rows*sizeof(*my_onz));CHKERRQ(ierr);
/* preallocation as a MATAIJ */
if (isdense) { /* special case for dense local matrices */
for (i=0;i<local_rows;i++) {
index_row = global_indices[i];
for (j=i;j<local_rows;j++) {
owner = row_ownership[index_row];
index_col = global_indices[j];
if (index_col > mat_ranges[owner]-1 && index_col < mat_ranges[owner+1] ) { /* diag block */
my_dnz[i] += 1.0;
} else { /* offdiag block */
my_onz[i] += 1.0;
}
/* same as before, interchanging rows and cols */
if (i != j) {
owner = row_ownership[index_col];
if (index_row > mat_ranges[owner]-1 && index_row < mat_ranges[owner+1] ) {
my_dnz[j] += 1.0;
} else {
my_onz[j] += 1.0;
}
}
}
}
} else {
for (i=0;i<local_rows;i++) {
PetscInt ncols;
const PetscInt *cols;
index_row = global_indices[i];
ierr = MatGetRow(matis->A,i,&ncols,&cols,NULL);CHKERRQ(ierr);
for (j=0;j<ncols;j++) {
owner = row_ownership[index_row];
index_col = global_indices[cols[j]];
if (index_col > mat_ranges[owner]-1 && index_col < mat_ranges[owner+1] ) { /* diag block */
my_dnz[i] += 1.0;
} else { /* offdiag block */
my_onz[i] += 1.0;
}
/* same as before, interchanging rows and cols */
if (issbaij) {
owner = row_ownership[index_col];
if (index_row > mat_ranges[owner]-1 && index_row < mat_ranges[owner+1] ) {
my_dnz[j] += 1.0;
} else {
my_onz[j] += 1.0;
}
}
}
ierr = MatRestoreRow(matis->A,i,&ncols,&cols,NULL);CHKERRQ(ierr);
}
}
ierr = VecSet(vec_dnz,0.0);CHKERRQ(ierr);
ierr = VecSet(vec_onz,0.0);CHKERRQ(ierr);
if (local_rows) { /* multilevel guard */
ierr = VecSetValuesLocal(vec_dnz,local_rows,local_indices,my_dnz,ADD_VALUES);CHKERRQ(ierr);
ierr = VecSetValuesLocal(vec_onz,local_rows,local_indices,my_onz,ADD_VALUES);CHKERRQ(ierr);
}
ierr = VecAssemblyBegin(vec_dnz);CHKERRQ(ierr);
ierr = VecAssemblyBegin(vec_onz);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec_dnz);CHKERRQ(ierr);
ierr = VecAssemblyEnd(vec_onz);CHKERRQ(ierr);
ierr = PetscFree(my_dnz);CHKERRQ(ierr);
ierr = PetscFree(my_onz);CHKERRQ(ierr);
ierr = PetscFree(row_ownership);CHKERRQ(ierr);
/* set computed preallocation in dnz and onz */
ierr = VecGetArray(vec_dnz,&array);CHKERRQ(ierr);
for (i=0; i<lrows; i++) dnz[i] = (PetscInt)PetscRealPart(array[i]);
ierr = VecRestoreArray(vec_dnz,&array);CHKERRQ(ierr);
ierr = VecGetArray(vec_onz,&array);CHKERRQ(ierr);
for (i=0;i<lrows;i++) onz[i] = (PetscInt)PetscRealPart(array[i]);
ierr = VecRestoreArray(vec_onz,&array);CHKERRQ(ierr);
ierr = VecDestroy(&vec_dnz);CHKERRQ(ierr);
ierr = VecDestroy(&vec_onz);CHKERRQ(ierr);
/* Resize preallocation if overestimated */
for (i=0;i<lrows;i++) {
dnz[i] = PetscMin(dnz[i],lcols);
onz[i] = PetscMin(onz[i],cols-lcols);
}
/* set preallocation */
ierr = MatMPIAIJSetPreallocation(new_mat,0,dnz,0,onz);CHKERRQ(ierr);
for (i=0;i<lrows/bs;i++) {
dnz[i] = dnz[i*bs]/bs;
onz[i] = onz[i*bs]/bs;
}
ierr = MatMPIBAIJSetPreallocation(new_mat,bs,0,dnz,0,onz);CHKERRQ(ierr);
for (i=0;i<lrows/bs;i++) {
dnz[i] = dnz[i]-i;
}
ierr = MatMPISBAIJSetPreallocation(new_mat,bs,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
*M = new_mat;
} else {
PetscInt mbs,mrows,mcols;
/* some checks */
ierr = MatGetBlockSize(*M,&mbs);CHKERRQ(ierr);
ierr = MatGetSize(*M,&mrows,&mcols);CHKERRQ(ierr);
if (mrows != rows) {
SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of rows (%d != %d)",rows,mrows);
}
if (mrows != rows) {
SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong number of cols (%d != %d)",cols,mcols);
}
if (mbs != bs) {
SETERRQ2(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"Cannot reuse matrix. Wrong block size (%d != %d)",bs,mbs);
}
ierr = MatZeroEntries(*M);CHKERRQ(ierr);
}
/* set local to global mappings */
/* ierr = MatSetLocalToGlobalMapping(*M,rmapping,cmapping);CHKERRQ(ierr); */
/* Set values */
if (isdense) { /* special case for dense local matrices */
ierr = MatSetOption(*M,MAT_ROW_ORIENTED,PETSC_FALSE);CHKERRQ(ierr);
ierr = MatDenseGetArray(matis->A,&array);CHKERRQ(ierr);
ierr = MatSetValues(*M,local_rows,global_indices,local_cols,global_indices,array,ADD_VALUES);CHKERRQ(ierr);
ierr = MatDenseRestoreArray(matis->A,&array);CHKERRQ(ierr);
ierr = PetscFree(local_indices);CHKERRQ(ierr);
ierr = PetscFree(global_indices);CHKERRQ(ierr);
} else { /* very basic values insertion for all other matrix types */
ierr = PetscFree(local_indices);CHKERRQ(ierr);
for (i=0;i<local_rows;i++) {
ierr = MatGetRow(matis->A,i,&j,(const PetscInt**)&local_indices,(const PetscScalar**)&array);CHKERRQ(ierr);
/* ierr = MatSetValuesLocal(*M,1,&i,j,local_indices,array,ADD_VALUES);CHKERRQ(ierr); */
ierr = ISLocalToGlobalMappingApply(matis->mapping,j,local_indices,global_indices);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApply(matis->mapping,1,&i,&index_row);CHKERRQ(ierr);
ierr = MatSetValues(*M,1,&index_row,j,global_indices,array,ADD_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(matis->A,i,&j,(const PetscInt**)&local_indices,(const PetscScalar**)&array);CHKERRQ(ierr);
}
ierr = PetscFree(global_indices);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(*M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
if (isdense) {
ierr = MatSetOption(*M,MAT_ROW_ORIENTED,PETSC_TRUE);CHKERRQ(ierr);
}
if (issbaij) {
ierr = MatRestoreRowUpperTriangular(matis->A);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
EXTERN_C_BEGIN
/* ----------------------------------------------------------------------------*/
/*
MatGetColoring_SL_Minpack - Uses the smallest-last (SL) coloring of minpack
*/
#undef __FUNCT__
#define __FUNCT__ "MatGetColoring_SL_Minpack"
PetscErrorCode MatGetColoring_SL_Minpack(Mat mat,MatColoringType name,ISColoring *iscoloring)
{
PetscErrorCode ierr;
PetscInt *list,*work,clique,*seq,*coloring,n;
const PetscInt *ria,*rja,*cia,*cja;
PetscInt ncolors,i;
PetscBool done;
Mat mat_seq = mat;
PetscMPIInt size;
MPI_Comm comm;
ISColoring iscoloring_seq;
PetscInt bs = 1,rstart,rend,N_loc,nc;
ISColoringValue *colors_loc;
PetscBool flg1,flg2;
PetscFunctionBegin;
/* this is ugly way to get blocksize but cannot call MatGetBlockSize() because AIJ can have bs > 1 */
ierr = PetscObjectTypeCompare((PetscObject)mat,MATSEQBAIJ,&flg1);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&flg2);CHKERRQ(ierr);
if (flg1 || flg2) {
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
}
ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
if (size > 1){
/* create a sequential iscoloring on all processors */
ierr = MatGetSeqNonzeroStructure(mat,&mat_seq);CHKERRQ(ierr);
}
ierr = MatGetRowIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&ria,&rja,&done);CHKERRQ(ierr);
ierr = MatGetColumnIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&cia,&cja,&done);CHKERRQ(ierr);
if (!done) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Ordering requires IJ");
ierr = MatFDColoringDegreeSequence_Minpack(n,cja,cia,rja,ria,&seq);CHKERRQ(ierr);
ierr = PetscMalloc2(n,PetscInt,&list,4*n,PetscInt,&work);CHKERRQ(ierr);
MINPACKslo(&n,cja,cia,rja,ria,seq,list,&clique,work,work+n,work+2*n,work+3*n);
ierr = PetscMalloc(n*sizeof(PetscInt),&coloring);CHKERRQ(ierr);
MINPACKseq(&n,cja,cia,rja,ria,list,coloring,&ncolors,work);
ierr = PetscFree2(list,work);CHKERRQ(ierr);
ierr = PetscFree(seq);CHKERRQ(ierr);
ierr = MatRestoreRowIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&ria,&rja,&done);CHKERRQ(ierr);
ierr = MatRestoreColumnIJ(mat_seq,1,PETSC_FALSE,PETSC_TRUE,&n,&cia,&cja,&done);CHKERRQ(ierr);
/* shift coloring numbers to start at zero and shorten */
if (ncolors > IS_COLORING_MAX-1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Maximum color size exceeded");
{
ISColoringValue *s = (ISColoringValue*) coloring;
for (i=0; i<n; i++) {
s[i] = (ISColoringValue) (coloring[i]-1);
}
ierr = MatColoringPatch(mat_seq,ncolors,n,s,iscoloring);CHKERRQ(ierr);
}
if (size > 1) {
ierr = MatDestroySeqNonzeroStructure(&mat_seq);CHKERRQ(ierr);
/* convert iscoloring_seq to a parallel iscoloring */
iscoloring_seq = *iscoloring;
rstart = mat->rmap->rstart/bs;
rend = mat->rmap->rend/bs;
N_loc = rend - rstart; /* number of local nodes */
/* get local colors for each local node */
ierr = PetscMalloc((N_loc+1)*sizeof(ISColoringValue),&colors_loc);CHKERRQ(ierr);
for (i=rstart; i<rend; i++){
colors_loc[i-rstart] = iscoloring_seq->colors[i];
}
/* create a parallel iscoloring */
nc=iscoloring_seq->n;
ierr = ISColoringCreate(comm,nc,N_loc,colors_loc,iscoloring);CHKERRQ(ierr);
ierr = ISColoringDestroy(&iscoloring_seq);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
EXTERN_C_END
EXTERN_C_BEGIN
/*
Simplest coloring, each column of the matrix gets its own unique color.
*/
#undef __FUNCT__
#define __FUNCT__ "MatGetColoring_Natural"
PetscErrorCode MatGetColoring_Natural(Mat mat,MatColoringType color, ISColoring *iscoloring)
{
PetscErrorCode ierr;
PetscInt start,end,i,bs = 1,n;
ISColoringValue *colors;
MPI_Comm comm;
PetscBool flg1,flg2;
Mat mat_seq = mat;
PetscMPIInt size;
ISColoring iscoloring_seq;
ISColoringValue *colors_loc;
PetscInt rstart,rend,N_loc,nc;
PetscFunctionBegin;
/* this is ugly way to get blocksize but cannot call MatGetBlockSize() because AIJ can have bs > 1 */
ierr = PetscObjectTypeCompare((PetscObject)mat,MATSEQBAIJ,&flg1);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)mat,MATMPIBAIJ,&flg2);CHKERRQ(ierr);
if (flg1 || flg2) {
ierr = MatGetBlockSize(mat,&bs);CHKERRQ(ierr);
}
ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
if (size > 1){
/* create a sequential iscoloring on all processors */
ierr = MatGetSeqNonzeroStructure(mat,&mat_seq);CHKERRQ(ierr);
}
ierr = MatGetSize(mat_seq,PETSC_NULL,&n);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(mat_seq,&start,&end);CHKERRQ(ierr);
n = n/bs;
if (n > IS_COLORING_MAX-1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Maximum color size exceeded");
start = start/bs;
end = end/bs;
ierr = PetscMalloc((end-start+1)*sizeof(PetscInt),&colors);CHKERRQ(ierr);
for (i=start; i<end; i++) {
colors[i-start] = (ISColoringValue)i;
}
ierr = ISColoringCreate(comm,n,end-start,colors,iscoloring);CHKERRQ(ierr);
if (size > 1) {
ierr = MatDestroySeqNonzeroStructure(&mat_seq);CHKERRQ(ierr);
/* convert iscoloring_seq to a parallel iscoloring */
iscoloring_seq = *iscoloring;
rstart = mat->rmap->rstart/bs;
rend = mat->rmap->rend/bs;
N_loc = rend - rstart; /* number of local nodes */
/* get local colors for each local node */
ierr = PetscMalloc((N_loc+1)*sizeof(ISColoringValue),&colors_loc);CHKERRQ(ierr);
for (i=rstart; i<rend; i++){
colors_loc[i-rstart] = iscoloring_seq->colors[i];
}
/* create a parallel iscoloring */
nc=iscoloring_seq->n;
ierr = ISColoringCreate(comm,nc,N_loc,colors_loc,iscoloring);CHKERRQ(ierr);
ierr = ISColoringDestroy(&iscoloring_seq);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*
PCGAMGCreateGraph - create simple scaled scalar graph from matrix
Input Parameter:
. Amat - matrix
Output Parameter:
. a_Gmaat - eoutput scalar graph (symmetric?)
*/
PetscErrorCode PCGAMGCreateGraph(Mat Amat, Mat *a_Gmat)
{
PetscErrorCode ierr;
PetscInt Istart,Iend,Ii,jj,kk,ncols,nloc,NN,MM,bs;
MPI_Comm comm;
Mat Gmat;
MatType mtype;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Amat, &Istart, &Iend);CHKERRQ(ierr);
ierr = MatGetSize(Amat, &MM, &NN);CHKERRQ(ierr);
ierr = MatGetBlockSize(Amat, &bs);CHKERRQ(ierr);
nloc = (Iend-Istart)/bs;
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
if (bs > 1) {
const PetscScalar *vals;
const PetscInt *idx;
PetscInt *d_nnz, *o_nnz,*w0,*w1,*w2;
PetscBool ismpiaij,isseqaij;
/*
Determine the preallocation needed for the scalar matrix derived from the vector matrix.
*/
ierr = PetscObjectBaseTypeCompare((PetscObject)Amat,MATSEQAIJ,&isseqaij);CHKERRQ(ierr);
ierr = PetscObjectBaseTypeCompare((PetscObject)Amat,MATMPIAIJ,&ismpiaij);CHKERRQ(ierr);
ierr = PetscMalloc2(nloc, &d_nnz,isseqaij ? 0 : nloc, &o_nnz);CHKERRQ(ierr);
if (isseqaij) {
PetscInt max_d_nnz;
/*
Determine exact preallocation count for (sequential) scalar matrix
*/
ierr = MatSeqAIJGetMaxRowNonzeros(Amat,&max_d_nnz);CHKERRQ(ierr);
max_d_nnz = PetscMin(nloc,bs*max_d_nnz);CHKERRQ(ierr);
ierr = PetscMalloc3(max_d_nnz, &w0,max_d_nnz, &w1,max_d_nnz, &w2);CHKERRQ(ierr);
for (Ii = 0, jj = 0; Ii < Iend; Ii += bs, jj++) {
ierr = MatCollapseRows(Amat,Ii,bs,w0,w1,w2,&d_nnz[jj],NULL);CHKERRQ(ierr);
}
ierr = PetscFree3(w0,w1,w2);CHKERRQ(ierr);
} else if (ismpiaij) {
Mat Daij,Oaij;
const PetscInt *garray;
PetscInt max_d_nnz;
ierr = MatMPIAIJGetSeqAIJ(Amat,&Daij,&Oaij,&garray);CHKERRQ(ierr);
/*
Determine exact preallocation count for diagonal block portion of scalar matrix
*/
ierr = MatSeqAIJGetMaxRowNonzeros(Daij,&max_d_nnz);CHKERRQ(ierr);
max_d_nnz = PetscMin(nloc,bs*max_d_nnz);CHKERRQ(ierr);
ierr = PetscMalloc3(max_d_nnz, &w0,max_d_nnz, &w1,max_d_nnz, &w2);CHKERRQ(ierr);
for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
ierr = MatCollapseRows(Daij,Ii,bs,w0,w1,w2,&d_nnz[jj],NULL);CHKERRQ(ierr);
}
ierr = PetscFree3(w0,w1,w2);CHKERRQ(ierr);
/*
Over estimate (usually grossly over), preallocation count for off-diagonal portion of scalar matrix
*/
for (Ii = 0, jj = 0; Ii < Iend - Istart; Ii += bs, jj++) {
o_nnz[jj] = 0;
for (kk=0; kk<bs; kk++) { /* rows that get collapsed to a single row */
ierr = MatGetRow(Oaij,Ii+kk,&ncols,0,0);CHKERRQ(ierr);
o_nnz[jj] += ncols;
ierr = MatRestoreRow(Oaij,Ii+kk,&ncols,0,0);CHKERRQ(ierr);
}
if (o_nnz[jj] > (NN/bs-nloc)) o_nnz[jj] = NN/bs-nloc;
}
} else SETERRQ(PETSC_COMM_WORLD,PETSC_ERR_USER,"Require AIJ matrix type");
/* get scalar copy (norms) of matrix */
ierr = MatGetType(Amat,&mtype);CHKERRQ(ierr);
ierr = MatCreate(comm, &Gmat);CHKERRQ(ierr);
ierr = MatSetSizes(Gmat,nloc,nloc,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetBlockSizes(Gmat, 1, 1);CHKERRQ(ierr);
ierr = MatSetType(Gmat, mtype);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(Gmat,0,d_nnz);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Gmat,0,d_nnz,0,o_nnz);CHKERRQ(ierr);
ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr);
for (Ii = Istart; Ii < Iend; Ii++) {
PetscInt dest_row = Ii/bs;
ierr = MatGetRow(Amat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
for (jj=0; jj<ncols; jj++) {
PetscInt dest_col = idx[jj]/bs;
PetscScalar sv = PetscAbs(PetscRealPart(vals[jj]));
ierr = MatSetValues(Gmat,1,&dest_row,1,&dest_col,&sv,ADD_VALUES);CHKERRQ(ierr);
}
ierr = MatRestoreRow(Amat,Ii,&ncols,&idx,&vals);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(Gmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(Gmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
} else {
/* just copy scalar matrix - abs() not taken here but scaled later */
ierr = MatDuplicate(Amat, MAT_COPY_VALUES, &Gmat);CHKERRQ(ierr);
}
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[GRAPH],0,0,0,0);CHKERRQ(ierr);
#endif
*a_Gmat = Gmat;
PetscFunctionReturn(0);
}
PetscErrorCode PCGAMGProlongator_GEO(PC pc,const Mat Amat,const Mat Gmat,PetscCoarsenData *agg_lists,Mat *a_P_out)
{
PC_MG *mg = (PC_MG*)pc->data;
PC_GAMG *pc_gamg = (PC_GAMG*)mg->innerctx;
const PetscInt verbose = pc_gamg->verbose;
const PetscInt dim = pc_gamg->data_cell_cols, data_cols = pc_gamg->data_cell_cols;
PetscErrorCode ierr;
PetscInt Istart,Iend,nloc,my0,jj,kk,ncols,nLocalSelected,bs,*clid_flid;
Mat Prol;
PetscMPIInt rank, size;
MPI_Comm comm;
IS selected_2,selected_1;
const PetscInt *selected_idx;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr);
#if defined PETSC_USE_LOG
ierr = PetscLogEventBegin(PC_GAMGProlongator_GEO,0,0,0,0);CHKERRQ(ierr);
#endif
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Amat, &Istart, &Iend);CHKERRQ(ierr);
ierr = MatGetBlockSize(Amat, &bs);CHKERRQ(ierr);
nloc = (Iend-Istart)/bs; my0 = Istart/bs;
if ((Iend-Istart) % bs) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"(Iend %D - Istart %D) % bs %D",Iend,Istart,bs);
/* get 'nLocalSelected' */
ierr = PetscCDGetMIS(agg_lists, &selected_1);CHKERRQ(ierr);
ierr = ISGetSize(selected_1, &jj);CHKERRQ(ierr);
ierr = PetscMalloc1(jj, &clid_flid);CHKERRQ(ierr);
ierr = ISGetIndices(selected_1, &selected_idx);CHKERRQ(ierr);
for (kk=0,nLocalSelected=0; kk<jj; kk++) {
PetscInt lid = selected_idx[kk];
if (lid<nloc) {
ierr = MatGetRow(Gmat,lid+my0,&ncols,0,0);CHKERRQ(ierr);
if (ncols>1) clid_flid[nLocalSelected++] = lid; /* fiter out singletons */
ierr = MatRestoreRow(Gmat,lid+my0,&ncols,0,0);CHKERRQ(ierr);
}
}
ierr = ISRestoreIndices(selected_1, &selected_idx);CHKERRQ(ierr);
ierr = ISDestroy(&selected_1);CHKERRQ(ierr); /* this is selected_1 in serial */
/* create prolongator, create P matrix */
ierr = MatCreate(comm, &Prol);CHKERRQ(ierr);
ierr = MatSetSizes(Prol,nloc*bs,nLocalSelected*bs,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetBlockSizes(Prol, bs, bs);CHKERRQ(ierr);
ierr = MatSetType(Prol, MATAIJ);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(Prol,3*data_cols,NULL);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Prol,3*data_cols,NULL,3*data_cols,NULL);CHKERRQ(ierr);
/* ierr = MatCreateAIJ(comm, */
/* nloc*bs, nLocalSelected*bs, */
/* PETSC_DETERMINE, PETSC_DETERMINE, */
/* 3*data_cols, NULL, */
/* 3*data_cols, NULL, */
/* &Prol); */
/* CHKERRQ(ierr); */
/* can get all points "removed" - but not on geomg */
ierr = MatGetSize(Prol, &kk, &jj);CHKERRQ(ierr);
if (jj==0) {
if (verbose) PetscPrintf(comm,"[%d]%s ERROE: no selected points on coarse grid\n",rank,__FUNCT__);
ierr = PetscFree(clid_flid);CHKERRQ(ierr);
ierr = MatDestroy(&Prol);CHKERRQ(ierr);
*a_P_out = NULL; /* out */
PetscFunctionReturn(0);
}
{
PetscReal *coords;
PetscInt data_stride;
PetscInt *crsGID = NULL;
Mat Gmat2;
if (dim != data_cols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"dim %D != data_cols %D",dim,data_cols);
/* grow ghost data for better coarse grid cover of fine grid */
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr);
#endif
/* messy method, squares graph and gets some data */
ierr = getGIDsOnSquareGraph(nLocalSelected, clid_flid, Gmat, &selected_2, &Gmat2, &crsGID);CHKERRQ(ierr);
/* llist is now not valid wrt squared graph, but will work as iterator in 'triangulateAndFormProl' */
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET5],0,0,0,0);CHKERRQ(ierr);
#endif
/* create global vector of coorindates in 'coords' */
if (size > 1) {
ierr = PCGAMGGetDataWithGhosts(Gmat2, dim, pc_gamg->data, &data_stride, &coords);CHKERRQ(ierr);
} else {
coords = (PetscReal*)pc_gamg->data;
data_stride = pc_gamg->data_sz/pc_gamg->data_cell_cols;
}
ierr = MatDestroy(&Gmat2);CHKERRQ(ierr);
/* triangulate */
if (dim == 2) {
PetscReal metric,tm;
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventBegin(petsc_gamg_setup_events[SET6],0,0,0,0);CHKERRQ(ierr);
#endif
ierr = triangulateAndFormProl(selected_2, data_stride, coords,nLocalSelected, clid_flid, agg_lists, crsGID, bs, Prol, &metric);CHKERRQ(ierr);
#if defined PETSC_GAMG_USE_LOG
ierr = PetscLogEventEnd(petsc_gamg_setup_events[SET6],0,0,0,0);CHKERRQ(ierr);
#endif
ierr = PetscFree(crsGID);CHKERRQ(ierr);
/* clean up and create coordinates for coarse grid (output) */
if (size > 1) ierr = PetscFree(coords);CHKERRQ(ierr);
ierr = MPI_Allreduce(&metric, &tm, 1, MPIU_REAL, MPIU_MAX, comm);CHKERRQ(ierr);
if (tm > 1.) { /* needs to be globalized - should not happen */
if (verbose) PetscPrintf(comm,"[%d]%s failed metric for coarse grid %e\n",rank,__FUNCT__,tm);
ierr = MatDestroy(&Prol);CHKERRQ(ierr);
Prol = NULL;
} else if (metric > .0) {
if (verbose) PetscPrintf(comm,"[%d]%s worst metric for coarse grid = %e\n",rank,__FUNCT__,metric);
}
} else SETERRQ(comm,PETSC_ERR_PLIB,"3D not implemented for 'geo' AMG");
{ /* create next coords - output */
PetscReal *crs_crds;
ierr = PetscMalloc1(dim*nLocalSelected, &crs_crds);CHKERRQ(ierr);
for (kk=0; kk<nLocalSelected; kk++) { /* grab local select nodes to promote - output */
PetscInt lid = clid_flid[kk];
for (jj=0; jj<dim; jj++) crs_crds[jj*nLocalSelected + kk] = pc_gamg->data[jj*nloc + lid];
}
ierr = PetscFree(pc_gamg->data);CHKERRQ(ierr);
pc_gamg->data = crs_crds; /* out */
pc_gamg->data_sz = dim*nLocalSelected;
}
ierr = ISDestroy(&selected_2);CHKERRQ(ierr);
}
*a_P_out = Prol; /* out */
ierr = PetscFree(clid_flid);CHKERRQ(ierr);
#if defined PETSC_USE_LOG
ierr = PetscLogEventEnd(PC_GAMGProlongator_GEO,0,0,0,0);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
EXTERN_C_END
/*MC
MATSOLVERCHOLMOD = "cholmod" - A matrix type providing direct solvers (Cholesky) for sequential matrices
via the external package CHOLMOD.
./configure --download-cholmod to install PETSc to use CHOLMOD
Consult CHOLMOD documentation for more information about the Common parameters
which correspond to the options database keys below.
Options Database Keys:
+ -mat_cholmod_dbound <0> - Minimum absolute value of diagonal entries of D (None)
. -mat_cholmod_grow0 <1.2> - Global growth ratio when factors are modified (None)
. -mat_cholmod_grow1 <1.2> - Column growth ratio when factors are modified (None)
. -mat_cholmod_grow2 <5> - Affine column growth constant when factors are modified (None)
. -mat_cholmod_maxrank <8> - Max rank of update, larger values are faster but use more memory [2,4,8] (None)
. -mat_cholmod_factor <AUTO> - (choose one of) SIMPLICIAL AUTO SUPERNODAL
. -mat_cholmod_supernodal_switch <40> - flop/nnz_L threshold for switching to supernodal factorization (None)
. -mat_cholmod_final_asis <TRUE> - Leave factors "as is" (None)
. -mat_cholmod_final_pack <TRUE> - Pack the columns when finished (use FALSE if the factors will be updated later) (None)
. -mat_cholmod_zrelax <0.8> - 3 real supernodal relaxed amalgamation parameters (None)
. -mat_cholmod_nrelax <4> - 3 size_t supernodal relaxed amalgamation parameters (None)
. -mat_cholmod_prefer_upper <TRUE> - Work with upper triangular form (faster when using fill-reducing ordering, slower in natural ordering) (None)
- -mat_cholmod_print <3> - Verbosity level (None)
Level: beginner
.seealso: PCCHOLESKY, PCFactorSetMatSolverPackage(), MatSolverPackage
M*/
EXTERN_C_BEGIN
#undef __FUNCT__
#define __FUNCT__ "MatGetFactor_seqsbaij_cholmod"
PetscErrorCode MatGetFactor_seqsbaij_cholmod(Mat A,MatFactorType ftype,Mat *F)
{
Mat B;
Mat_CHOLMOD *chol;
PetscErrorCode ierr;
PetscInt m=A->rmap->n,n=A->cmap->n,bs;
PetscFunctionBegin;
if (ftype != MAT_FACTOR_CHOLESKY) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_SUP,"CHOLMOD cannot do %s factorization with SBAIJ, only %s",
MatFactorTypes[ftype],MatFactorTypes[MAT_FACTOR_CHOLESKY]);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
if (bs != 1) SETERRQ1(((PetscObject)A)->comm,PETSC_ERR_SUP,"CHOLMOD only supports block size=1, given %D",bs);
/* Create the factorization matrix F */
ierr = MatCreate(((PetscObject)A)->comm,&B);CHKERRQ(ierr);
ierr = MatSetSizes(B,PETSC_DECIDE,PETSC_DECIDE,m,n);CHKERRQ(ierr);
ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr);
ierr = MatSeqSBAIJSetPreallocation(B,1,0,PETSC_NULL);CHKERRQ(ierr);
ierr = PetscNewLog(B,Mat_CHOLMOD,&chol);CHKERRQ(ierr);
chol->Wrap = MatWrapCholmod_seqsbaij;
chol->Destroy = MatDestroy_SeqSBAIJ;
B->spptr = chol;
B->ops->view = MatView_CHOLMOD;
B->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_CHOLMOD;
B->ops->destroy = MatDestroy_CHOLMOD;
ierr = PetscObjectComposeFunctionDynamic((PetscObject)B,"MatFactorGetSolverPackage_C","MatFactorGetSolverPackage_seqsbaij_cholmod",MatFactorGetSolverPackage_seqsbaij_cholmod);CHKERRQ(ierr);
B->factortype = MAT_FACTOR_CHOLESKY;
B->assembled = PETSC_TRUE; /* required by -ksp_view */
B->preallocated = PETSC_TRUE;
ierr = CholmodStart(B);CHKERRQ(ierr);
*F = B;
PetscFunctionReturn(0);
}
PetscErrorCode PCSetUp_ML(PC pc)
{
PetscErrorCode ierr;
PetscMPIInt size;
FineGridCtx *PetscMLdata;
ML *ml_object;
ML_Aggregate *agg_object;
ML_Operator *mlmat;
PetscInt nlocal_allcols,Nlevels,mllevel,level,level1,m,fine_level,bs;
Mat A,Aloc;
GridCtx *gridctx;
PC_MG *mg = (PC_MG*)pc->data;
PC_ML *pc_ml = (PC_ML*)mg->innerctx;
PetscBool isSeq, isMPI;
KSP smoother;
PC subpc;
PetscInt mesh_level, old_mesh_level;
PetscFunctionBegin;
A = pc->pmat;
ierr = MPI_Comm_size(((PetscObject)A)->comm,&size);CHKERRQ(ierr);
if (pc->setupcalled) {
if (pc->flag == SAME_NONZERO_PATTERN && pc_ml->reuse_interpolation) {
/*
Reuse interpolaton instead of recomputing aggregates and updating the whole hierarchy. This is less expensive for
multiple solves in which the matrix is not changing too quickly.
*/
ml_object = pc_ml->ml_object;
gridctx = pc_ml->gridctx;
Nlevels = pc_ml->Nlevels;
fine_level = Nlevels - 1;
gridctx[fine_level].A = A;
ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
if (isMPI){
ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr);
} else if (isSeq) {
Aloc = A;
ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr);
} else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name);
ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr);
PetscMLdata = pc_ml->PetscMLdata;
ierr = MatDestroy(&PetscMLdata->Aloc);CHKERRQ(ierr);
PetscMLdata->A = A;
PetscMLdata->Aloc = Aloc;
ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata);
ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec);
mesh_level = ml_object->ML_finest_level;
while (ml_object->SingleLevel[mesh_level].Rmat->to) {
old_mesh_level = mesh_level;
mesh_level = ml_object->SingleLevel[mesh_level].Rmat->to->levelnum;
/* clean and regenerate A */
mlmat = &(ml_object->Amat[mesh_level]);
ML_Operator_Clean(mlmat);
ML_Operator_Init(mlmat,ml_object->comm);
ML_Gen_AmatrixRAP(ml_object, old_mesh_level, mesh_level);
}
level = fine_level - 1;
if (size == 1) { /* convert ML P, R and A into seqaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
} else { /* convert ML P and R into shell format, ML A into mpiaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_MPIAIJ(mlmat,MAT_REUSE_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
}
for (level=0; level<fine_level; level++) {
if (level > 0){
ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr);
}
ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
}
ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr);
ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,SAME_NONZERO_PATTERN);CHKERRQ(ierr);
ierr = PCSetUp_MG(pc);CHKERRQ(ierr);
PetscFunctionReturn(0);
} else {
/* since ML can change the size of vectors/matrices at any level we must destroy everything */
ierr = PCReset_ML(pc);CHKERRQ(ierr);
ierr = PCReset_MG(pc);CHKERRQ(ierr);
}
}
/* setup special features of PCML */
/*--------------------------------*/
/* covert A to Aloc to be used by ML at fine grid */
pc_ml->size = size;
ierr = PetscObjectTypeCompare((PetscObject) A, MATSEQAIJ, &isSeq);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject) A, MATMPIAIJ, &isMPI);CHKERRQ(ierr);
if (isMPI){
ierr = MatConvert_MPIAIJ_ML(A,PETSC_NULL,MAT_INITIAL_MATRIX,&Aloc);CHKERRQ(ierr);
} else if (isSeq) {
Aloc = A;
ierr = PetscObjectReference((PetscObject)Aloc);CHKERRQ(ierr);
} else SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_WRONG, "Matrix type '%s' cannot be used with ML. ML can only handle AIJ matrices.",((PetscObject)A)->type_name);
/* create and initialize struct 'PetscMLdata' */
ierr = PetscNewLog(pc,FineGridCtx,&PetscMLdata);CHKERRQ(ierr);
pc_ml->PetscMLdata = PetscMLdata;
ierr = PetscMalloc((Aloc->cmap->n+1)*sizeof(PetscScalar),&PetscMLdata->pwork);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->x);CHKERRQ(ierr);
ierr = VecSetSizes(PetscMLdata->x,Aloc->cmap->n,Aloc->cmap->n);CHKERRQ(ierr);
ierr = VecSetType(PetscMLdata->x,VECSEQ);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_SELF,&PetscMLdata->y);CHKERRQ(ierr);
ierr = VecSetSizes(PetscMLdata->y,A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(PetscMLdata->y,VECSEQ);CHKERRQ(ierr);
PetscMLdata->A = A;
PetscMLdata->Aloc = Aloc;
/* create ML discretization matrix at fine grid */
/* ML requires input of fine-grid matrix. It determines nlevels. */
ierr = MatGetSize(Aloc,&m,&nlocal_allcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(A,&bs);CHKERRQ(ierr);
ML_Create(&ml_object,pc_ml->MaxNlevels);
ML_Comm_Set_UsrComm(ml_object->comm,((PetscObject)A)->comm);
pc_ml->ml_object = ml_object;
ML_Init_Amatrix(ml_object,0,m,m,PetscMLdata);
ML_Set_Amatrix_Getrow(ml_object,0,PetscML_getrow,PetscML_comm,nlocal_allcols);
ML_Set_Amatrix_Matvec(ml_object,0,PetscML_matvec);
ML_Set_Symmetrize(ml_object,pc_ml->Symmetrize ? ML_YES : ML_NO);
/* aggregation */
ML_Aggregate_Create(&agg_object);
pc_ml->agg_object = agg_object;
{
MatNullSpace mnull;
ierr = MatGetNearNullSpace(A,&mnull);CHKERRQ(ierr);
if (pc_ml->nulltype == PCML_NULLSPACE_AUTO) {
if (mnull) pc_ml->nulltype = PCML_NULLSPACE_USER;
else if (bs > 1) pc_ml->nulltype = PCML_NULLSPACE_BLOCK;
else pc_ml->nulltype = PCML_NULLSPACE_SCALAR;
}
switch (pc_ml->nulltype) {
case PCML_NULLSPACE_USER: {
PetscScalar *nullvec;
const PetscScalar *v;
PetscBool has_const;
PetscInt i,j,mlocal,nvec,M;
const Vec *vecs;
if (!mnull) SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_USER,"Must provide explicit null space using MatSetNearNullSpace() to use user-specified null space");
ierr = MatGetSize(A,&M,PETSC_NULL);CHKERRQ(ierr);
ierr = MatGetLocalSize(Aloc,&mlocal,PETSC_NULL);CHKERRQ(ierr);
ierr = MatNullSpaceGetVecs(mnull,&has_const,&nvec,&vecs);CHKERRQ(ierr);
ierr = PetscMalloc((nvec+!!has_const)*mlocal*sizeof *nullvec,&nullvec);CHKERRQ(ierr);
if (has_const) for (i=0; i<mlocal; i++) nullvec[i] = 1.0/M;
for (i=0; i<nvec; i++) {
ierr = VecGetArrayRead(vecs[i],&v);CHKERRQ(ierr);
for (j=0; j<mlocal; j++) nullvec[(i+!!has_const)*mlocal + j] = v[j];
ierr = VecRestoreArrayRead(vecs[i],&v);CHKERRQ(ierr);
}
ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,nvec+!!has_const,nullvec,mlocal);CHKERRQ(ierr);
ierr = PetscFree(nullvec);CHKERRQ(ierr);
} break;
case PCML_NULLSPACE_BLOCK:
ierr = ML_Aggregate_Set_NullSpace(agg_object,bs,bs,0,0);CHKERRQ(ierr);
break;
case PCML_NULLSPACE_SCALAR:
break;
default: SETERRQ(((PetscObject)pc)->comm,PETSC_ERR_SUP,"Unknown null space type");
}
}
ML_Aggregate_Set_MaxCoarseSize(agg_object,pc_ml->MaxCoarseSize);
/* set options */
switch (pc_ml->CoarsenScheme) {
case 1:
ML_Aggregate_Set_CoarsenScheme_Coupled(agg_object);break;
case 2:
ML_Aggregate_Set_CoarsenScheme_MIS(agg_object);break;
case 3:
ML_Aggregate_Set_CoarsenScheme_METIS(agg_object);break;
}
ML_Aggregate_Set_Threshold(agg_object,pc_ml->Threshold);
ML_Aggregate_Set_DampingFactor(agg_object,pc_ml->DampingFactor);
if (pc_ml->SpectralNormScheme_Anorm){
ML_Set_SpectralNormScheme_Anorm(ml_object);
}
agg_object->keep_agg_information = (int)pc_ml->KeepAggInfo;
agg_object->keep_P_tentative = (int)pc_ml->Reusable;
agg_object->block_scaled_SA = (int)pc_ml->BlockScaling;
agg_object->minimizing_energy = (int)pc_ml->EnergyMinimization;
agg_object->minimizing_energy_droptol = (double)pc_ml->EnergyMinimizationDropTol;
agg_object->cheap_minimizing_energy = (int)pc_ml->EnergyMinimizationCheap;
if (pc_ml->OldHierarchy) {
Nlevels = ML_Gen_MGHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object);
} else {
Nlevels = ML_Gen_MultiLevelHierarchy_UsingAggregation(ml_object,0,ML_INCREASING,agg_object);
}
if (Nlevels<=0) SETERRQ1(((PetscObject)pc)->comm,PETSC_ERR_ARG_OUTOFRANGE,"Nlevels %d must > 0",Nlevels);
pc_ml->Nlevels = Nlevels;
fine_level = Nlevels - 1;
ierr = PCMGSetLevels(pc,Nlevels,PETSC_NULL);CHKERRQ(ierr);
/* set default smoothers */
for (level=1; level<=fine_level; level++){
if (size == 1){
ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr);
ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr);
ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr);
ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr);
} else {
ierr = PCMGGetSmoother(pc,level,&smoother);CHKERRQ(ierr);
ierr = KSPSetType(smoother,KSPRICHARDSON);CHKERRQ(ierr);
ierr = KSPGetPC(smoother,&subpc);CHKERRQ(ierr);
ierr = PCSetType(subpc,PCSOR);CHKERRQ(ierr);
}
}
ierr = PetscObjectOptionsBegin((PetscObject)pc);CHKERRQ(ierr);
ierr = PCSetFromOptions_MG(pc);CHKERRQ(ierr); /* should be called in PCSetFromOptions_ML(), but cannot be called prior to PCMGSetLevels() */
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscMalloc(Nlevels*sizeof(GridCtx),&gridctx);CHKERRQ(ierr);
pc_ml->gridctx = gridctx;
/* wrap ML matrices by PETSc shell matrices at coarsened grids.
Level 0 is the finest grid for ML, but coarsest for PETSc! */
gridctx[fine_level].A = A;
level = fine_level - 1;
if (size == 1){ /* convert ML P, R and A into seqaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Pmat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr);
mlmat = &(ml_object->Rmat[mllevel-1]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr);
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_SeqAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
} else { /* convert ML P and R into shell format, ML A into mpiaij format */
for (mllevel=1; mllevel<Nlevels; mllevel++){
mlmat = &(ml_object->Pmat[mllevel]);
ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].P);CHKERRQ(ierr);
mlmat = &(ml_object->Rmat[mllevel-1]);
ierr = MatWrapML_SHELL(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].R);CHKERRQ(ierr);
mlmat = &(ml_object->Amat[mllevel]);
ierr = MatWrapML_MPIAIJ(mlmat,MAT_INITIAL_MATRIX,&gridctx[level].A);CHKERRQ(ierr);
level--;
}
}
/* create vectors and ksp at all levels */
for (level=0; level<fine_level; level++){
level1 = level + 1;
ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].x);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level].x,gridctx[level].A->cmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level].x,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetX(pc,level,gridctx[level].x);CHKERRQ(ierr);
ierr = VecCreate(((PetscObject)gridctx[level].A)->comm,&gridctx[level].b);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level].b,gridctx[level].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level].b,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetRhs(pc,level,gridctx[level].b);CHKERRQ(ierr);
ierr = VecCreate(((PetscObject)gridctx[level1].A)->comm,&gridctx[level1].r);CHKERRQ(ierr);
ierr = VecSetSizes(gridctx[level1].r,gridctx[level1].A->rmap->n,PETSC_DECIDE);CHKERRQ(ierr);
ierr = VecSetType(gridctx[level1].r,VECMPI);CHKERRQ(ierr);
ierr = PCMGSetR(pc,level1,gridctx[level1].r);CHKERRQ(ierr);
if (level == 0){
ierr = PCMGGetCoarseSolve(pc,&gridctx[level].ksp);CHKERRQ(ierr);
} else {
ierr = PCMGGetSmoother(pc,level,&gridctx[level].ksp);CHKERRQ(ierr);
}
}
ierr = PCMGGetSmoother(pc,fine_level,&gridctx[fine_level].ksp);CHKERRQ(ierr);
/* create coarse level and the interpolation between the levels */
for (level=0; level<fine_level; level++){
level1 = level + 1;
ierr = PCMGSetInterpolation(pc,level1,gridctx[level].P);CHKERRQ(ierr);
ierr = PCMGSetRestriction(pc,level1,gridctx[level].R);CHKERRQ(ierr);
if (level > 0){
ierr = PCMGSetResidual(pc,level,PCMGDefaultResidual,gridctx[level].A);CHKERRQ(ierr);
}
ierr = KSPSetOperators(gridctx[level].ksp,gridctx[level].A,gridctx[level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
}
ierr = PCMGSetResidual(pc,fine_level,PCMGDefaultResidual,gridctx[fine_level].A);CHKERRQ(ierr);
ierr = KSPSetOperators(gridctx[fine_level].ksp,gridctx[level].A,gridctx[fine_level].A,DIFFERENT_NONZERO_PATTERN);CHKERRQ(ierr);
/* setupcalled is set to 0 so that MG is setup from scratch */
pc->setupcalled = 0;
ierr = PCSetUp_MG(pc);CHKERRQ(ierr);
PetscFunctionReturn(0);
}