PETSC_EXTERN void PETSC_STDCALL petscsfview_(PetscSF *sf, PetscViewer *vin, PetscErrorCode *ierr)
{
PetscViewer v;
PetscPatchDefaultViewers_Fortran(vin, v);
*ierr = PetscSFView(*sf, v);
}
PetscErrorCode DMMeshConvertOverlapToSF(DM dm, PetscSF *sf)
{
ALE::Obj<PETSC_MESH_TYPE> mesh;
PetscInt *local;
PetscSFNode *remote;
PetscInt numPoints;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFCreate(((PetscObject) dm)->comm, sf);
CHKERRQ(ierr);
ierr = DMMeshGetMesh(dm, mesh);
CHKERRQ(ierr);
{
/* The local points have degree 1
We use the recv overlap
*/
ALE::Obj<PETSC_MESH_TYPE::recv_overlap_type> overlap = mesh->getRecvOverlap();
numPoints = overlap->getNumPoints();
ierr = PetscMalloc(numPoints * sizeof(PetscInt), &local);
CHKERRQ(ierr);
ierr = PetscMalloc(numPoints * sizeof(PetscSFNode), &remote);
CHKERRQ(ierr);
for (PetscInt r = 0, i = 0; r < overlap->getNumRanks(); ++r) {
const PetscInt rank = overlap->getRank(r);
const PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator cBegin = overlap->supportBegin(rank);
const PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator cEnd = overlap->supportEnd(rank);
for (PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator c_iter = cBegin; c_iter != cEnd; ++c_iter, ++i) {
local[i] = *c_iter;
remote[i].rank = rank;
remote[i].index = c_iter.color();
}
}
ierr = PetscSFSetGraph(*sf, numPoints, numPoints, local, PETSC_OWN_POINTER, remote, PETSC_OWN_POINTER);
CHKERRQ(ierr);
ierr = PetscSFView(*sf, NULL);
CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
Vec B,Bout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscScalar *bufB;
PetscScalar *bufBout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 2) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for two MPI processes\n");
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 2;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
if (rank == 0) {
iremote[0].rank = 0;
iremote[0].index = 0;
iremote[1].rank = 1;
iremote[1].index = 0;
} else {
iremote[0].rank = 1;
iremote[0].index = 0;
iremote[1].rank = 0;
iremote[1].index = 0;
}
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,2,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&B);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Bout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArray(B,&bufB);CHKERRQ(ierr);
for (i=0; i<2; i++) {
bufA[i] = (PetscScalar)rank;
bufB[i] = (PetscScalar)(rank) + 10.0;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(B,&bufB);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecGetArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecRestoreArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecView(Bout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&B);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = VecDestroy(&Bout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
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);
}
/*@C
DMPlexDistribute - Distributes the mesh and any associated sections.
Not Collective
Input Parameter:
+ dm - The original DMPlex object
. partitioner - The partitioning package, or NULL for the default
- overlap - The overlap of partitions, 0 is the default
Output Parameter:
+ sf - The PetscSF used for point distribution
- parallelMesh - The distributed DMPlex object, or NULL
Note: If the mesh was not distributed, the return value is NULL.
The user can control the definition of adjacency for the mesh using DMPlexGetAdjacencyUseCone() and
DMPlexSetAdjacencyUseClosure(). They should choose the combination appropriate for the function
representation on the mesh.
Level: intermediate
.keywords: mesh, elements
.seealso: DMPlexCreate(), DMPlexDistributeByFace(), DMPlexSetAdjacencyUseCone(), DMPlexSetAdjacencyUseClosure()
@*/
PetscErrorCode DMPlexDistribute(DM dm, const char partitioner[], PetscInt overlap, PetscSF *sf, DM *dmParallel)
{
DM_Plex *mesh = (DM_Plex*) dm->data, *pmesh;
MPI_Comm comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS origCellPart, origPart, cellPart, part;
PetscSection origCellPartSection, origPartSection, cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF, pointSF, coneSF;
ISLocalToGlobalMapping renumbering;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets;
PetscInt *cones, *newCones, newConesSize;
PetscBool flg;
PetscMPIInt rank, numProcs, p;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
if (sf) PetscValidPointer(sf,4);
PetscValidPointer(dmParallel,5);
ierr = PetscLogEventBegin(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
*dmParallel = NULL;
if (numProcs == 1) PetscFunctionReturn(0);
ierr = DMPlexGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = PetscLogEventBegin(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
if (overlap > 1) SETERRQ(PetscObjectComm((PetscObject)dm), PETSC_ERR_SUP, "Overlap > 1 not yet implemented");
ierr = DMPlexCreatePartition(dm, partitioner, height, overlap > 0 ? PETSC_TRUE : PETSC_FALSE, &cellPartSection, &cellPart, &origCellPartSection, &origCellPart);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) numRemoteRanks = numProcs;
else numRemoteRanks = 0;
ierr = PetscMalloc1(numRemoteRanks, &remoteRanks);CHKERRQ(ierr);
for (p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-partition_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, NULL);CHKERRQ(ierr);
if (origCellPart) {
ierr = PetscPrintf(comm, "Original Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(origCellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(origCellPart, NULL);CHKERRQ(ierr);
}
ierr = PetscSFView(partSF, NULL);CHKERRQ(ierr);
}
/* Close the partition over the mesh */
ierr = DMPlexCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMPlexCreate(comm, dmParallel);CHKERRQ(ierr);
ierr = DMPlexSetDimension(*dmParallel, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *dmParallel, "Parallel Mesh");CHKERRQ(ierr);
pmesh = (DM_Plex*) (*dmParallel)->data;
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, &pointSF);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, NULL);CHKERRQ(ierr);
ierr = PetscSFView(pointSF, NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, NULL);CHKERRQ(ierr);
}
ierr = PetscLogEventEnd(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscLogEventBegin(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMPlexGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMPlexGetConeSection(*dmParallel, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMSetUp(*dmParallel);CHKERRQ(ierr);
{
PetscInt pStart, pEnd, p;
ierr = PetscSectionGetChart(newConeSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt coneSize;
ierr = PetscSectionGetDof(newConeSection, p, &coneSize);CHKERRQ(ierr);
pmesh->maxConeSize = PetscMax(pmesh->maxConeSize, coneSize);
}
}
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMPlexGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetCones(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyBlock(renumbering, IS_GTOLM_MASK, newConesSize, newCones, NULL, newCones);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-cones_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, NULL);CHKERRQ(ierr);
}
ierr = DMPlexGetConeOrientations(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientations(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Create supports and stratify sieve */
{
PetscInt pStart, pEnd;
ierr = PetscSectionGetChart(pmesh->coneSection, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionSetChart(pmesh->supportSection, pStart, pEnd);CHKERRQ(ierr);
}
ierr = DMPlexSymmetrize(*dmParallel);CHKERRQ(ierr);
ierr = DMPlexStratify(*dmParallel);CHKERRQ(ierr);
/* Distribute Coordinates */
{
PetscSection originalCoordSection, newCoordSection;
Vec originalCoordinates, newCoordinates;
PetscInt bs;
const char *name;
ierr = DMGetCoordinateSection(dm, &originalCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinateSection(*dmParallel, &newCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinatesLocal(dm, &originalCoordinates);CHKERRQ(ierr);
ierr = VecCreate(comm, &newCoordinates);CHKERRQ(ierr);
ierr = PetscObjectGetName((PetscObject) originalCoordinates, &name);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) newCoordinates, name);CHKERRQ(ierr);
ierr = DMPlexDistributeField(dm, pointSF, originalCoordSection, originalCoordinates, newCoordSection, newCoordinates);CHKERRQ(ierr);
ierr = DMSetCoordinatesLocal(*dmParallel, newCoordinates);CHKERRQ(ierr);
ierr = VecGetBlockSize(originalCoordinates, &bs);CHKERRQ(ierr);
ierr = VecSetBlockSize(newCoordinates, bs);CHKERRQ(ierr);
ierr = VecDestroy(&newCoordinates);CHKERRQ(ierr);
}
/* Distribute labels */
ierr = PetscLogEventBegin(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
{
DMLabel next = mesh->labels, newNext = pmesh->labels;
PetscInt numLabels = 0, l;
/* Bcast number of labels */
while (next) {++numLabels; next = next->next;}
ierr = MPI_Bcast(&numLabels, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
next = mesh->labels;
for (l = 0; l < numLabels; ++l) {
DMLabel labelNew;
PetscBool isdepth;
/* Skip "depth" because it is recreated */
if (!rank) {ierr = PetscStrcmp(next->name, "depth", &isdepth);CHKERRQ(ierr);}
ierr = MPI_Bcast(&isdepth, 1, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
if (isdepth) {if (!rank) next = next->next; continue;}
ierr = DMLabelDistribute(next, partSection, part, renumbering, &labelNew);CHKERRQ(ierr);
/* Insert into list */
if (newNext) newNext->next = labelNew;
else pmesh->labels = labelNew;
newNext = labelNew;
if (!rank) next = next->next;
}
}
ierr = PetscLogEventEnd(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
/* Setup hybrid structure */
{
const PetscInt *gpoints;
PetscInt depth, n, d;
for (d = 0; d <= dim; ++d) {pmesh->hybridPointMax[d] = mesh->hybridPointMax[d];}
ierr = MPI_Bcast(pmesh->hybridPointMax, dim+1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(renumbering, &n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(renumbering, &gpoints);CHKERRQ(ierr);
ierr = DMPlexGetDepth(dm, &depth);CHKERRQ(ierr);
for (d = 0; d <= dim; ++d) {
PetscInt pmax = pmesh->hybridPointMax[d], newmax = 0, pEnd, stratum[2], p;
if (pmax < 0) continue;
ierr = DMPlexGetDepthStratum(dm, d > depth ? depth : d, &stratum[0], &stratum[1]);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(*dmParallel, d, NULL, &pEnd);CHKERRQ(ierr);
ierr = MPI_Bcast(stratum, 2, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < n; ++p) {
const PetscInt point = gpoints[p];
if ((point >= stratum[0]) && (point < stratum[1]) && (point >= pmax)) ++newmax;
}
if (newmax > 0) pmesh->hybridPointMax[d] = pEnd - newmax;
else pmesh->hybridPointMax[d] = -1;
}
ierr = ISLocalToGlobalMappingRestoreIndices(renumbering, &gpoints);CHKERRQ(ierr);
}
/* Cleanup Partition */
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Create point SF for parallel mesh */
ierr = PetscLogEventBegin(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
{
const PetscInt *leaves;
PetscSFNode *remotePoints, *rowners, *lowners;
PetscInt numRoots, numLeaves, numGhostPoints = 0, p, gp, *ghostPoints;
PetscInt pStart, pEnd;
ierr = DMPlexGetChart(*dmParallel, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(pointSF, &numRoots, &numLeaves, &leaves, NULL);CHKERRQ(ierr);
ierr = PetscMalloc2(numRoots,&rowners,numLeaves,&lowners);CHKERRQ(ierr);
for (p=0; p<numRoots; p++) {
rowners[p].rank = -1;
rowners[p].index = -1;
}
if (origCellPart) {
/* Make sure points in the original partition are not assigned to other procs */
const PetscInt *origPoints;
ierr = DMPlexCreatePartitionClosure(dm, origCellPartSection, origCellPart, &origPartSection, &origPart);CHKERRQ(ierr);
ierr = ISGetIndices(origPart, &origPoints);CHKERRQ(ierr);
for (p = 0; p < numProcs; ++p) {
PetscInt dof, off, d;
ierr = PetscSectionGetDof(origPartSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(origPartSection, p, &off);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
rowners[origPoints[d]].rank = p;
}
}
ierr = ISRestoreIndices(origPart, &origPoints);CHKERRQ(ierr);
ierr = ISDestroy(&origPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origPartSection);CHKERRQ(ierr);
}
ierr = ISDestroy(&origCellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origCellPartSection);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].rank == rank) { /* Either put in a bid or we know we own it */
lowners[p].rank = rank;
lowners[p].index = leaves ? leaves[p] : p;
} else if (lowners[p].rank >= 0) { /* Point already claimed so flag so that MAXLOC does not listen to us */
lowners[p].rank = -2;
lowners[p].index = -2;
}
}
for (p=0; p<numRoots; p++) { /* Root must not participate in the rediction, flag so that MAXLOC does not use */
rowners[p].rank = -3;
rowners[p].index = -3;
}
ierr = PetscSFReduceBegin(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].index < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cell partition corrupt: point not claimed");
if (lowners[p].rank != rank) ++numGhostPoints;
}
ierr = PetscMalloc1(numGhostPoints, &ghostPoints);CHKERRQ(ierr);
ierr = PetscMalloc1(numGhostPoints, &remotePoints);CHKERRQ(ierr);
for (p = 0, gp = 0; p < numLeaves; ++p) {
if (lowners[p].rank != rank) {
ghostPoints[gp] = leaves ? leaves[p] : p;
remotePoints[gp].rank = lowners[p].rank;
remotePoints[gp].index = lowners[p].index;
++gp;
}
}
ierr = PetscFree2(rowners,lowners);CHKERRQ(ierr);
ierr = PetscSFSetGraph((*dmParallel)->sf, pEnd - pStart, numGhostPoints, ghostPoints, PETSC_OWN_POINTER, remotePoints, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions((*dmParallel)->sf);CHKERRQ(ierr);
}
pmesh->useCone = mesh->useCone;
pmesh->useClosure = mesh->useClosure;
ierr = PetscLogEventEnd(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
/* Copy BC */
ierr = DMPlexCopyBoundary(dm, *dmParallel);CHKERRQ(ierr);
/* Cleanup */
if (sf) {*sf = pointSF;}
else {ierr = PetscSFDestroy(&pointSF);CHKERRQ(ierr);}
ierr = DMSetFromOptions(*dmParallel);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
PetscBool test_dupped_type;
MPI_Datatype contig;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for one MPI process");
ierr = PetscOptionsBegin(PETSC_COMM_WORLD,"","PetscSF type freeing options","none");CHKERRQ(ierr);
test_dupped_type = PETSC_FALSE;
ierr = PetscOptionsBool("-test_dupped_type", "Test dupped input type","",test_dupped_type,&test_dupped_type,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 1;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
iremote[0].rank = 0;
iremote[0].index = 0;
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,4,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
for (i=0; i<4; i++) {
bufA[i] = (PetscScalar)i;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = MPI_Type_contiguous(4, MPIU_SCALAR, &contig);CHKERRQ(ierr);
ierr = MPI_Type_commit(&contig);CHKERRQ(ierr);
if (test_dupped_type) {
MPI_Datatype tmp;
ierr = MPI_Type_dup(contig, &tmp);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
contig = tmp;
}
for (i=0;i<10000;i++) {
ierr = PetscSFBcastBegin(sf,contig,bufA,bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,contig,bufA,bufAout);CHKERRQ(ierr);
}
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
PetscErrorCode DMPatchSolve(DM dm)
{
MPI_Comm comm;
MPI_Comm commz;
DM dmc;
PetscSF sfz, sfzr;
Vec XC;
MatStencil patchSize, commSize, gridRank, lower, upper;
PetscInt M, N, P, i, j, k, l, m, n, p = 0;
PetscMPIInt rank, size;
PetscInt debug = 0;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMPatchGetCoarse(dm, &dmc);CHKERRQ(ierr);
ierr = DMPatchGetPatchSize(dm, &patchSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMGetGlobalVector(dmc, &XC);CHKERRQ(ierr);
ierr = DMDAGetInfo(dmc, 0, &M, &N, &P, &l, &m, &n, 0,0,0,0,0,0);CHKERRQ(ierr);
M = PetscMax(M, 1); l = PetscMax(l, 1);
N = PetscMax(N, 1); m = PetscMax(m, 1);
P = PetscMax(P, 1); n = PetscMax(n, 1);
gridRank.i = rank % l;
gridRank.j = rank/l % m;
gridRank.k = rank/(l*m) % n;
if (commSize.i*commSize.j*commSize.k == size || commSize.i*commSize.j*commSize.k == 0) {
commSize.i = l; commSize.j = m; commSize.k = n;
commz = comm;
} else if (commSize.i*commSize.j*commSize.k == 1) {
commz = PETSC_COMM_SELF;
} else {
const PetscMPIInt newComm = ((gridRank.k/commSize.k)*(m/commSize.j) + gridRank.j/commSize.j)*(l/commSize.i) + (gridRank.i/commSize.i);
const PetscMPIInt newRank = ((gridRank.k%commSize.k)*commSize.j + gridRank.j%commSize.j)*commSize.i + (gridRank.i%commSize.i);
ierr = MPI_Comm_split(comm, newComm, newRank, &commz);CHKERRQ(ierr);
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d color %d key %d commz %d\n", rank, newComm, newRank, *((PetscMPIInt*) &commz));CHKERRQ(ierr);}
}
/*
Assumptions:
- patchSize divides gridSize
- commSize divides gridSize
- commSize divides l,m,n
Ignore multiple patches per rank for now
Multiple ranks per patch:
- l,m,n divides patchSize
- commSize divides patchSize
*/
for (k = 0; k < P; k += PetscMax(patchSize.k, 1)) {
for (j = 0; j < N; j += PetscMax(patchSize.j, 1)) {
for (i = 0; i < M; i += PetscMax(patchSize.i, 1), ++p) {
MPI_Comm commp = MPI_COMM_NULL;
DM dmz = NULL;
#if 0
DM dmf = NULL;
Mat interpz = NULL;
#endif
Vec XZ = NULL;
PetscScalar *xcarray = NULL;
PetscScalar *xzarray = NULL;
if ((gridRank.k/commSize.k == p/(l/commSize.i * m/commSize.j) % n/commSize.k) &&
(gridRank.j/commSize.j == p/(l/commSize.i) % m/commSize.j) &&
(gridRank.i/commSize.i == p % l/commSize.i)) {
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d is accepting Patch %d\n", rank, p);CHKERRQ(ierr);}
commp = commz;
}
/* Zoom to coarse patch */
lower.i = i; lower.j = j; lower.k = k;
upper.i = i + patchSize.i; upper.j = j + patchSize.j; upper.k = k + patchSize.k;
ierr = DMPatchZoom(dmc, XC, lower, upper, commp, &dmz, &sfz, &sfzr);CHKERRQ(ierr);
lower.c = 0; /* initialize member, otherwise compiler issues warnings */
upper.c = 0; /* initialize member, otherwise compiler issues warnings */
/* Debug */
ierr = PetscPrintf(comm, "Patch %d: (%d, %d, %d)--(%d, %d, %d)\n", p, lower.i, lower.j, lower.k, upper.i, upper.j, upper.k);CHKERRQ(ierr);
if (dmz) {ierr = DMView(dmz, PETSC_VIEWER_STDOUT_(commz));CHKERRQ(ierr);}
ierr = PetscSFView(sfz, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
ierr = PetscSFView(sfzr, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
/* Scatter Xcoarse -> Xzoom */
if (dmz) {ierr = DMGetGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
#if 0
/* Interpolate Xzoom -> Xfine, note that this may be on subcomms */
ierr = DMRefine(dmz, MPI_COMM_NULL, &dmf);CHKERRQ(ierr);
ierr = DMCreateInterpolation(dmz, dmf, &interpz, NULL);CHKERRQ(ierr);
ierr = DMInterpolate(dmz, interpz, dmf);CHKERRQ(ierr);
/* Smooth Xfine using two-step smoother, normal smoother plus Kaczmarz---moves back and forth from dmzoom to dmfine */
/* Compute residual Rfine */
/* Restrict Rfine to Rzoom_restricted */
#endif
/* Scatter Rzoom_restricted -> Rcoarse_restricted */
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
if (dmz) {ierr = DMRestoreGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
/* Compute global residual Rcoarse */
/* TauCoarse = Rcoarse - Rcoarse_restricted */
ierr = PetscSFDestroy(&sfz);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfzr);CHKERRQ(ierr);
ierr = DMDestroy(&dmz);CHKERRQ(ierr);
}
}
}
ierr = DMRestoreGlobalVector(dmc, &XC);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode PetscSFCreateSectionSF(PetscSF sf, PetscSection section, PetscSF *sectionSF)
{
PetscInt numRanks;
const PetscInt *ranks, *rankOffsets;
const PetscMPIInt *localPoints, *remotePoints;
PetscInt numPoints, numIndices = 0;
PetscInt *remoteOffsets;
PetscInt *localIndices;
PetscSFNode *remoteIndices;
PetscInt i, r, ind;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFGetRanks(sf, &numRanks, &ranks, &rankOffsets, &localPoints, &remotePoints);
CHKERRQ(ierr);
numPoints = rankOffsets[numRanks];
for (i = 0; i < numPoints; ++i) {
PetscInt dof;
ierr = PetscSectionGetDof(section, localPoints[i], &dof);
CHKERRQ(ierr);
numIndices += dof;
}
/* Communicate offsets for ghosted points */
#if 0
PetscInt *localOffsets;
ierr = PetscMalloc2(numPoints,PetscInt,&localOffsets,numPoints,PetscInt,&remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSectionGetOffset(section, localPoints[i], &localOffsets[i]);
CHKERRQ(ierr);
}
ierr = PetscSFBcastBegin(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i]);
CHKERRQ(ierr);
}
#else
ierr = PetscMalloc((section->atlasLayout.pEnd - section->atlasLayout.pStart) * sizeof(PetscInt), &remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
for (i = section->atlasLayout.pStart; i < section->atlasLayout.pEnd; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i-section->atlasLayout.pStart]);
CHKERRQ(ierr);
}
#endif
ierr = PetscSynchronizedFlush(((PetscObject) sf)->comm);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscInt), &localIndices);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscSFNode), &remoteIndices);
CHKERRQ(ierr);
/* Create new index graph */
for (r = 0, ind = 0; r < numRanks; ++r) {
PetscInt rank = ranks[r];
for (i = rankOffsets[r]; i < rankOffsets[r+1]; ++i) {
PetscInt localPoint = localPoints[i];
PetscInt remoteOffset = remoteOffsets[localPoint-section->atlasLayout.pStart];
PetscInt localOffset, dof, d;
ierr = PetscSectionGetOffset(section, localPoint, &localOffset);
CHKERRQ(ierr);
ierr = PetscSectionGetDof(section, localPoint, &dof);
CHKERRQ(ierr);
for (d = 0; d < dof; ++d, ++ind) {
localIndices[ind] = localOffset+d;
remoteIndices[ind].rank = rank;
remoteIndices[ind].index = remoteOffset+d;
}
}
}
ierr = PetscFree(remoteOffsets);
CHKERRQ(ierr);
if (numIndices != ind) SETERRQ2(((PetscObject) sf)->comm, PETSC_ERR_PLIB, "Inconsistency in indices, %d should be %d", ind, numIndices);
ierr = PetscSFCreate(((PetscObject) sf)->comm, sectionSF);
CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sectionSF, numIndices, numIndices, localIndices, PETSC_OWN_POINTER, remoteIndices, PETSC_OWN_POINTER);
CHKERRQ(ierr);
ierr = PetscSFView(*sectionSF, NULL);
CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* Distribute cones
- Partitioning: input partition point map and naive sf, output sf with inverse of map, distribute points
- Distribute section: input current sf, communicate sizes and offsets, output local section and offsets (only use for new sf)
- Create SF for values: input current sf and offsets, output new sf
- Distribute values: input new sf, communicate values
*/
PetscErrorCode DistributeMesh(DM dm, AppCtx *user, PetscSF *pointSF, DM *parallelDM)
{
MPI_Comm comm = ((PetscObject) dm)->comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS cellPart, part;
PetscSection cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF;
ISLocalToGlobalMapping renumbering;
PetscSF coneSF;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets, newConesSize;
PetscInt *cones, *newCones;
PetscMPIInt numProcs, rank, p;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = DMMeshGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = DMMeshCreatePartition(dm, &cellPartSection, &cellPart, height);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) {
numRemoteRanks = numProcs;
} else {
numRemoteRanks = 0;
}
ierr = PetscMalloc(numRemoteRanks * sizeof(PetscSFNode), &remoteRanks);CHKERRQ(ierr);
for(p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, PETSC_NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(partSF, PETSC_NULL);CHKERRQ(ierr);
/* Close the partition over the mesh */
ierr = DMMeshCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMMeshCreate(comm, parallelDM);CHKERRQ(ierr);
ierr = DMMeshSetDimension(*parallelDM, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *parallelDM, "Parallel Mesh");CHKERRQ(ierr);
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, pointSF);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(*pointSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, PETSC_NULL);CHKERRQ(ierr);
/* Cleanup */
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMMeshGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMMeshGetConeSection(*parallelDM, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(*pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMMeshSetUp(*parallelDM);CHKERRQ(ierr);
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(*pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMMeshGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMMeshGetCones(*parallelDM, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApply(renumbering, IS_GTOLM_MASK, newConesSize, newCones, PETSC_NULL, newCones);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
/* Create supports and stratify sieve */
ierr = DMMeshSymmetrize(*parallelDM);CHKERRQ(ierr);
ierr = DMMeshStratify(*parallelDM);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscSF sf,sfDup,sfInv,sfEmbed,sfA,sfB,sfBA;
const PetscInt *degree;
PetscErrorCode ierr;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = CheckRanksEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup then reset */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test view (no graph set, no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then view (no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then view (no graph set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then graph then view */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then type */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Bcast */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFBcastEnd (sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Reduce */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test FetchAndOp */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpBegin(sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpEnd (sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test ComputeDegree */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFDuplicate() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_GRAPH,&sfDup);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateInverseSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateInverseSF(sf,&sfInv);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedLeafSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedLeafSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCompose() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfA);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfA,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfB);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfB,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCompose(sfA,sfB,&sfBA);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfB);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
