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);
}
/*
MatConvert_Basic - Converts from any input format to another format. For
parallel formats, the new matrix distribution is determined by PETSc.
Does not do preallocation so in general will be slow
*/
PetscErrorCode MatConvert_Basic(Mat mat, MatType newtype,MatReuse reuse,Mat *newmat)
{
Mat M;
const PetscScalar *vwork;
PetscErrorCode ierr;
PetscInt i,j,nz,m,n,rstart,rend,lm,ln,prbs,pcbs,cstart,cend,*dnz,*onz;
const PetscInt *cwork;
PetscBool isseqsbaij,ismpisbaij,isseqbaij,ismpibaij,isseqdense,ismpidense;
PetscFunctionBegin;
ierr = MatGetSize(mat,&m,&n);CHKERRQ(ierr);
ierr = MatGetLocalSize(mat,&lm,&ln);CHKERRQ(ierr);
if (ln == n) ln = PETSC_DECIDE; /* try to preserve column ownership */
ierr = MatCreate(PetscObjectComm((PetscObject)mat),&M);CHKERRQ(ierr);
ierr = MatSetSizes(M,lm,ln,m,n);CHKERRQ(ierr);
ierr = MatSetBlockSizes(M,mat->rmap->bs,mat->cmap->bs);CHKERRQ(ierr);
ierr = MatSetType(M,newtype);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(mat,&rstart,&rend);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)M,MATSEQSBAIJ,&isseqsbaij);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)M,MATMPISBAIJ,&ismpisbaij);CHKERRQ(ierr);
if (isseqsbaij || ismpisbaij) {ierr = MatSetOption(M,MAT_IGNORE_LOWER_TRIANGULAR,PETSC_TRUE);CHKERRQ(ierr);}
ierr = PetscObjectTypeCompare((PetscObject)M,MATSEQBAIJ,&isseqbaij);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)M,MATMPIBAIJ,&ismpibaij);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)M,MATSEQDENSE,&isseqdense);CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)M,MATMPIDENSE,&ismpidense);CHKERRQ(ierr);
if (isseqdense) {
ierr = MatSeqDenseSetPreallocation(M,NULL);CHKERRQ(ierr);
} else if (ismpidense) {
ierr = MatMPIDenseSetPreallocation(M,NULL);CHKERRQ(ierr);
} else {
/* Preallocation block sizes. (S)BAIJ matrices will have one index per block. */
prbs = (isseqbaij || ismpibaij || isseqsbaij || ismpisbaij) ? M->rmap->bs : 1;
pcbs = (isseqbaij || ismpibaij || isseqsbaij || ismpisbaij) ? M->cmap->bs : 1;
ierr = PetscMalloc2(lm/prbs,&dnz,lm/prbs,&onz);CHKERRQ(ierr);
ierr = MatGetOwnershipRangeColumn(mat,&cstart,&cend);CHKERRQ(ierr);
for (i=0; i<lm; i+=prbs) {
ierr = MatGetRow(mat,rstart+i,&nz,&cwork,NULL);CHKERRQ(ierr);
dnz[i] = 0;
onz[i] = 0;
for (j=0; j<nz; j+=pcbs) {
if ((isseqsbaij || ismpisbaij) && cwork[j] < rstart+i) continue;
if (cstart <= cwork[j] && cwork[j] < cend) dnz[i]++;
else onz[i]++;
}
ierr = MatRestoreRow(mat,rstart+i,&nz,&cwork,NULL);CHKERRQ(ierr);
}
ierr = MatXAIJSetPreallocation(M,M->rmap->bs,dnz,onz,dnz,onz);CHKERRQ(ierr);
ierr = PetscFree2(dnz,onz);CHKERRQ(ierr);
}
for (i=rstart; i<rend; i++) {
ierr = MatGetRow(mat,i,&nz,&cwork,&vwork);CHKERRQ(ierr);
ierr = MatSetValues(M,1,&i,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(mat,i,&nz,&cwork,&vwork);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
if (reuse == MAT_REUSE_MATRIX) {
ierr = MatHeaderReplace(mat,M);CHKERRQ(ierr);
} else {
*newmat = M;
}
PetscFunctionReturn(0);
}
PETSC_EXTERN void PETSC_STDCALL matxaijsetpreallocation_(Mat A,PetscInt *bs, PetscInt dnnz[], PetscInt onnz[], PetscInt dnnzu[], PetscInt onnzu[], int *__ierr ){
*__ierr = MatXAIJSetPreallocation(
(Mat)PetscToPointer((A) ),*bs,dnnz,onnz,dnnzu,onnzu);
}
PetscErrorCode DMPlexPreallocateOperator(DM dm, PetscInt bs, PetscSection section, PetscSection sectionGlobal, PetscInt dnz[], PetscInt onz[], PetscInt dnzu[], PetscInt onzu[], Mat A, PetscBool fillMatrix)
{
MPI_Comm comm;
MatType mtype;
PetscSF sf, sfDof, sfAdj;
PetscSection leafSectionAdj, rootSectionAdj, sectionAdj, anchorSectionAdj;
PetscInt nroots, nleaves, l, p;
const PetscInt *leaves;
const PetscSFNode *remotes;
PetscInt dim, pStart, pEnd, numDof, globalOffStart, globalOffEnd, numCols;
PetscInt *tmpAdj = NULL, *adj, *rootAdj, *anchorAdj = NULL, *cols, *remoteOffsets;
PetscInt adjSize;
PetscLayout rLayout;
PetscInt locRows, rStart, rEnd, r;
PetscMPIInt size;
PetscBool doCommLocal, doComm, debug = PETSC_FALSE, isSymBlock, isSymSeqBlock, isSymMPIBlock;
PetscBool useAnchors;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
PetscValidHeaderSpecific(section, PETSC_SECTION_CLASSID, 3);
PetscValidHeaderSpecific(sectionGlobal, PETSC_SECTION_CLASSID, 4);
PetscValidHeaderSpecific(A, MAT_CLASSID, 9);
if (dnz) PetscValidPointer(dnz,5);
if (onz) PetscValidPointer(onz,6);
if (dnzu) PetscValidPointer(dnzu,7);
if (onzu) PetscValidPointer(onzu,8);
ierr = PetscLogEventBegin(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(NULL, "-dm_view_preallocation", &debug, NULL);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMGetPointSF(dm, &sf);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL);CHKERRQ(ierr);
doCommLocal = (size > 1) && (nroots >= 0) ? PETSC_TRUE : PETSC_FALSE;
ierr = MPI_Allreduce(&doCommLocal, &doComm, 1, MPIU_BOOL, MPI_LAND, comm);CHKERRQ(ierr);
/* Create dof SF based on point SF */
if (debug) {
ierr = PetscPrintf(comm, "Input Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(section, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input Global Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionGlobal, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sf, NULL);CHKERRQ(ierr);
}
ierr = PetscSFCreateRemoteOffsets(sf, section, section, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, section, remoteOffsets, section, &sfDof);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Dof SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfDof, NULL);CHKERRQ(ierr);
}
/* Create section for dof adjacency (dof ==> # adj dof) */
ierr = PetscSectionGetChart(section, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(section, &numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(leafSectionAdj, 0, numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSectionAdj, 0, numDof);CHKERRQ(ierr);
/* Fill in the ghost dofs on the interface */
ierr = PetscSFGetGraph(sf, NULL, &nleaves, &leaves, &remotes);CHKERRQ(ierr);
/* use constraints in finding adjacency in this routine */
ierr = DMPlexGetAdjacencyUseAnchors(dm,&useAnchors);CHKERRQ(ierr);
ierr = DMPlexSetAdjacencyUseAnchors(dm,PETSC_TRUE);CHKERRQ(ierr);
/*
section - maps points to (# dofs, local dofs)
sectionGlobal - maps points to (# dofs, global dofs)
leafSectionAdj - maps unowned local dofs to # adj dofs
rootSectionAdj - maps owned local dofs to # adj dofs
adj - adj global dofs indexed by leafSectionAdj
rootAdj - adj global dofs indexed by rootSectionAdj
sf - describes shared points across procs
sfDof - describes shared dofs across procs
sfAdj - describes shared adjacent dofs across procs
** The bootstrapping process involves six rounds with similar structure of visiting neighbors of each point.
(0). If there are point-to-point constraints, add the adjacencies of constrained points to anchors in anchorAdj
(This is done in DMPlexComputeAnchorAdjacencies())
1. Visit unowned points on interface, count adjacencies placing in leafSectionAdj
Reduce those counts to rootSectionAdj (now redundantly counting some interface points)
2. Visit owned points on interface, count adjacencies placing in rootSectionAdj
Create sfAdj connecting rootSectionAdj and leafSectionAdj
3. Visit unowned points on interface, write adjacencies to adj
Gather adj to rootAdj (note that there is redundancy in rootAdj when multiple procs find the same adjacencies)
4. Visit owned points on interface, write adjacencies to rootAdj
Remove redundancy in rootAdj
** The last two traversals use transitive closure
5. Visit all owned points in the subdomain, count dofs for each point (sectionAdj)
Allocate memory addressed by sectionAdj (cols)
6. Visit all owned points in the subdomain, insert dof adjacencies into cols
** Knowing all the column adjacencies, check ownership and sum into dnz and onz
*/
ierr = DMPlexComputeAnchorAdjacencies(dm,section,sectionGlobal,&anchorSectionAdj,&anchorAdj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation on Leaves:\n");CHKERRQ(ierr);
ierr = PetscSectionView(leafSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get maximum remote adjacency sizes for owned dofs on interface (roots) */
if (doComm) {
ierr = PetscSFReduceBegin(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
}
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Add in local adjacency sizes for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after local additions:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Create adj SF based on dof SF */
ierr = PetscSFCreateRemoteOffsets(sfDof, rootSectionAdj, leafSectionAdj, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sfDof, rootSectionAdj, remoteOffsets, leafSectionAdj, &sfAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfAdj, NULL);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfDof);CHKERRQ(ierr);
/* Create leaf adjacency */
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(leafSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscCalloc1(adjSize, &adj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof, anOff;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt aoff, i = 0;
ierr = PetscSectionGetOffset(leafSectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
adj[aoff+i] = (ngoff < 0 ? -(ngoff+1) : ngoff) + nd;
++i;
}
}
for (q = 0; q < anDof; q++) {
adj[aoff+i] = anchorAdj[anOff+q];
++i;
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Leaf adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, adj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Gather adjacenct indices to root */
ierr = PetscSectionGetStorageSize(rootSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscMalloc1(adjSize, &rootAdj);CHKERRQ(ierr);
for (r = 0; r < adjSize; ++r) rootAdj[r] = -1;
if (doComm) {
ierr = PetscSFGatherBegin(sfAdj, MPIU_INT, adj, rootAdj);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(sfAdj, MPIU_INT, adj, rootAdj);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfAdj);CHKERRQ(ierr);
ierr = PetscFree(adj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices after gather\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Add in local adjacency indices for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof, anOff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt adof, aoff, i;
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
i = adof-1;
for (q = 0; q < anDof; q++) {
rootAdj[aoff+i] = anchorAdj[anOff+q];
--i;
}
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
rootAdj[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
--i;
}
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Compress indices */
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt dof, cdof, off, d;
PetscInt adof, aoff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
for (d = off; d < off+dof-cdof; ++d) {
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
ierr = PetscSortRemoveDupsInt(&adof, &rootAdj[aoff]);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSectionAdj, d, adof);CHKERRQ(ierr);
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after compression:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Root adjacency indices after compression\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Build adjacency section: Maps global indices to sets of adjacent global indices */
ierr = PetscSectionGetOffsetRange(sectionGlobal, &globalOffStart, &globalOffEnd);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, §ionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(sectionAdj, globalOffStart, globalOffEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
ierr = PetscSectionSetDof(sectionAdj, goff+d, rdof);CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, noff;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, padj, &noff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(sectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get adjacent indices */
ierr = PetscSectionGetStorageSize(sectionAdj, &numCols);CHKERRQ(ierr);
ierr = PetscMalloc1(numCols, &cols);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof, anOff;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
PetscInt aoff, roff;
ierr = PetscSectionGetOffset(sectionAdj, goff+d, &aoff);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, off+d, &roff);CHKERRQ(ierr);
ierr = PetscMemcpy(&cols[aoff], &rootAdj[roff], rdof * sizeof(PetscInt));CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
PetscInt adof, aoff, i = 0;
ierr = PetscSectionGetDof(sectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
const PetscInt *ncind;
/* Adjacent points may not be in the section chart */
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintIndices(section, padj, &ncind);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd, ++i) {
cols[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
}
}
for (q = 0; q < anDof; q++, i++) {
cols[aoff+i] = anchorAdj[anOff + q];
}
if (i != adof) SETERRQ4(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of entries %D != %D for dof %D (point %D)", i, adof, d, p);
}
}
ierr = PetscSectionDestroy(&anchorSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSectionAdj);CHKERRQ(ierr);
ierr = PetscFree(anchorAdj);CHKERRQ(ierr);
ierr = PetscFree(rootAdj);CHKERRQ(ierr);
ierr = PetscFree(tmpAdj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Column indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, numCols, cols, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Create allocation vectors from adjacency graph */
ierr = MatGetLocalSize(A, &locRows, NULL);CHKERRQ(ierr);
ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)A), &rLayout);CHKERRQ(ierr);
ierr = PetscLayoutSetLocalSize(rLayout, locRows);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(rLayout, 1);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(rLayout);CHKERRQ(ierr);
ierr = PetscLayoutGetRange(rLayout, &rStart, &rEnd);CHKERRQ(ierr);
ierr = PetscLayoutDestroy(&rLayout);CHKERRQ(ierr);
/* Only loop over blocks of rows */
if (rStart%bs || rEnd%bs) SETERRQ3(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid layout [%d, %d) for matrix, must be divisible by block size %d", rStart, rEnd, bs);
for (r = rStart/bs; r < rEnd/bs; ++r) {
const PetscInt row = r*bs;
PetscInt numCols, cStart, c;
ierr = PetscSectionGetDof(sectionAdj, row, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, row, &cStart);CHKERRQ(ierr);
for (c = cStart; c < cStart+numCols; ++c) {
if ((cols[c] >= rStart*bs) && (cols[c] < rEnd*bs)) {
++dnz[r-rStart];
if (cols[c] >= row) ++dnzu[r-rStart];
} else {
++onz[r-rStart];
if (cols[c] >= row) ++onzu[r-rStart];
}
}
}
if (bs > 1) {
for (r = 0; r < locRows/bs; ++r) {
dnz[r] /= bs;
onz[r] /= bs;
dnzu[r] /= bs;
onzu[r] /= bs;
}
}
/* Set matrix pattern */
ierr = MatXAIJSetPreallocation(A, bs, dnz, onz, dnzu, onzu);CHKERRQ(ierr);
ierr = MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr);
/* Check for symmetric storage */
ierr = MatGetType(A, &mtype);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSBAIJ, &isSymBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSEQSBAIJ, &isSymSeqBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATMPISBAIJ, &isSymMPIBlock);CHKERRQ(ierr);
if (isSymBlock || isSymSeqBlock || isSymMPIBlock) {ierr = MatSetOption(A, MAT_IGNORE_LOWER_TRIANGULAR, PETSC_TRUE);CHKERRQ(ierr);}
/* Fill matrix with zeros */
if (fillMatrix) {
PetscScalar *values;
PetscInt maxRowLen = 0;
for (r = rStart; r < rEnd; ++r) {
PetscInt len;
ierr = PetscSectionGetDof(sectionAdj, r, &len);CHKERRQ(ierr);
maxRowLen = PetscMax(maxRowLen, len);
}
ierr = PetscCalloc1(maxRowLen, &values);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
PetscInt numCols, cStart;
ierr = PetscSectionGetDof(sectionAdj, r, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, r, &cStart);CHKERRQ(ierr);
ierr = MatSetValues(A, 1, &r, numCols, &cols[cStart], values, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(values);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
}
/* restore original useAnchors */
ierr = DMPlexSetAdjacencyUseAnchors(dm,useAnchors);CHKERRQ(ierr);
ierr = PetscSectionDestroy(§ionAdj);CHKERRQ(ierr);
ierr = PetscFree(cols);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
MatLaplacian - Form the matrix Laplacian, with all values in the matrix less than the tolerance set to zero
Input Parameters:
+ A - The matrix
- tol - The zero tolerance
Output Parameters:
. L - The graph Laplacian matrix
Level: intermediate
.seealso: MatChop()
@*/
PetscErrorCode MatLaplacian(Mat A, PetscReal tol, Mat *L)
{
PetscScalar *newVals;
PetscInt *newCols;
PetscInt rStart, rEnd, r, colMax = 0;
PetscInt *dnnz, *onnz;
PetscInt m, n, M, N;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MatCreate(PetscObjectComm((PetscObject) A), L);CHKERRQ(ierr);
ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr);
ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr);
ierr = MatSetSizes(*L, m, n, M, N);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A, &rStart, &rEnd);CHKERRQ(ierr);
ierr = PetscMalloc2(m,PetscInt,&dnnz,m,PetscInt,&onnz);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
const PetscScalar *vals;
const PetscInt *cols;
PetscInt ncols, newcols, c;
PetscBool hasdiag = PETSC_FALSE;
dnnz[r-rStart] = onnz[r-rStart] = 0;
ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
for (c = 0, newcols = 0; c < ncols; ++c) {
if (cols[c] == r) {
++newcols;
hasdiag = PETSC_TRUE;
++dnnz[r-rStart];
} else if (PetscAbsScalar(vals[c]) >= tol) {
if ((cols[c] >= rStart) && (cols[c] < rEnd)) ++dnnz[r-rStart];
else ++onnz[r-rStart];
++newcols;
}
}
if (!hasdiag) {++newcols; ++dnnz[r-rStart];}
colMax = PetscMax(colMax, newcols);CHKERRQ(ierr);
ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
}
ierr = MatSetFromOptions(*L);CHKERRQ(ierr);
ierr = MatXAIJSetPreallocation(*L, 1, dnnz, onnz, NULL, NULL);CHKERRQ(ierr);
ierr = MatSetUp(*L);CHKERRQ(ierr);
ierr = PetscMalloc2(colMax,PetscInt,&newCols,colMax,PetscScalar,&newVals);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
const PetscScalar *vals;
const PetscInt *cols;
PetscInt ncols, newcols, c;
PetscBool hasdiag = PETSC_FALSE;
ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
for (c = 0, newcols = 0; c < ncols; ++c) {
if (cols[c] == r) {
newCols[newcols] = cols[c];
newVals[newcols] = dnnz[r-rStart]+onnz[r-rStart]-1;
++newcols;
hasdiag = PETSC_TRUE;
} else if (PetscAbsScalar(vals[c]) >= tol) {
newCols[newcols] = cols[c];
newVals[newcols] = -1.0;
++newcols;
}
if (newcols > colMax) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Overran work space");
}
if (!hasdiag) {
newCols[newcols] = r;
newVals[newcols] = dnnz[r-rStart]+onnz[r-rStart]-1;
++newcols;
}
ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
ierr = MatSetValues(*L, 1, &r, newcols, newCols, newVals, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree2(dnnz,onnz);CHKERRQ(ierr);
ierr = MatAssemblyBegin(*L, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*L, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscFree2(newCols,newVals);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
MatCreateSubMatrixBanded - Extract the banded subset B of A such that ||Vec(B)||_1 >= frac ||Vec(A)||_1
Input Parameters:
+ A - The matrix
. kmax - The maximum half-bandwidth, so 2k+1 diagonals may be extracted
- frac - The norm fraction for the extracted band
Output Parameters:
. B - The banded submatrix
Level: intermediate
.seealso: MatChop()
@*/
PetscErrorCode MatCreateSubMatrixBanded(Mat A, PetscInt kmax, PetscReal frac, Mat *B)
{
Vec weight;
PetscScalar *w, *newVals;
PetscReal normA = 0.0, normB = 0.0;
PetscInt rStart, rEnd, r;
PetscInt *dnnz, *onnz, *newCols;
PetscInt m, n, M, N, k, maxcols = 0;
PetscErrorCode ierr;
PetscFunctionBegin;
/* Create weight vector */
ierr = MatGetVecs(A, NULL, &weight);CHKERRQ(ierr);
ierr = VecSet(weight, 0.0);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(A, &rStart, &rEnd);CHKERRQ(ierr);
ierr = VecGetArray(weight, &w);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
const PetscScalar *vals;
const PetscInt *cols;
PetscInt ncols, c;
ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
for (c = 0; c < ncols; ++c) {
w[abs(r - cols[c])] += PetscAbsScalar(vals[c]);
normA += PetscAbsScalar(vals[c]);
}
ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
}
ierr = VecRestoreArray(weight, &w);CHKERRQ(ierr);
/* Determine bandwidth */
ierr = PetscPrintf(PETSC_COMM_WORLD, "||Vec(A)||_1: %g\n", normA);CHKERRQ(ierr);
ierr = VecGetArray(weight, &w);CHKERRQ(ierr);
for (k = 0; k < kmax; ++k) {
normB += w[k];
if (normB >= frac*normA) break;
}
ierr = VecRestoreArray(weight, &w);CHKERRQ(ierr);
ierr = VecDestroy(&weight);CHKERRQ(ierr);
ierr = PetscPrintf(PETSC_COMM_WORLD, "Bandwidth of %d%% band: %d frac: %g\n", (PetscInt) (frac*100), k, normB/normA);CHKERRQ(ierr);
/* Extract band */
ierr = MatCreate(PetscObjectComm((PetscObject) A), B);CHKERRQ(ierr);
ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr);
ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr);
ierr = MatSetSizes(*B, m, n, M, N);CHKERRQ(ierr);
ierr = PetscMalloc2(m,PetscInt,&dnnz,m,PetscInt,&onnz);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
const PetscScalar *vals;
const PetscInt *cols;
PetscInt ncols, c;
dnnz[r-rStart] = onnz[r-rStart] = 0;
ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
for (c = 0; c < ncols; ++c) {
if (abs(cols[c] - r) > k) continue;
if ((cols[c] >= rStart) && (cols[c] < rEnd)) ++dnnz[r-rStart];
else ++onnz[r-rStart];
}
maxcols = PetscMax(ncols, maxcols);
ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
}
ierr = MatSetFromOptions(*B);CHKERRQ(ierr);
ierr = MatXAIJSetPreallocation(*B, 1, dnnz, onnz, NULL, NULL);CHKERRQ(ierr);
ierr = MatSetUp(*B);CHKERRQ(ierr);
ierr = PetscMalloc2(maxcols,PetscInt,&newCols,maxcols,PetscScalar,&newVals);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
const PetscScalar *vals;
const PetscInt *cols;
PetscInt ncols, newcols, c;
ierr = MatGetRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
for (c = 0, newcols = 0; c < ncols; ++c) {
if (abs(cols[c] - r) > k) continue;
newCols[newcols] = cols[c];
newVals[newcols] = vals[c];
++newcols;
if (newcols > maxcols) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Overran work space");
}
ierr = MatRestoreRow(A, r, &ncols, &cols, &vals);CHKERRQ(ierr);
ierr = MatSetValues(*B, 1, &r, newcols, newCols, newVals, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree2(newCols, newVals);CHKERRQ(ierr);
ierr = PetscFree2(dnnz, onnz);CHKERRQ(ierr);
ierr = MatAssemblyBegin(*B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*B, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
