PetscErrorCode DMLabelDistribute_Internal(DMLabel label, PetscSF sf, PetscSection *leafSection, PetscInt **leafStrata)
{
MPI_Comm comm;
PetscInt s, l, nroots, nleaves, dof, offset, size;
PetscInt *remoteOffsets, *rootStrata, *rootIdx;
PetscSection rootSection;
PetscSF labelSF;
PetscErrorCode ierr;
PetscFunctionBegin;
if (label) {ierr = DMLabelMakeAllValid_Private(label);CHKERRQ(ierr);}
ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
/* Build a section of stratum values per point, generate the according SF
and distribute point-wise stratum values to leaves. */
ierr = PetscSFGetGraph(sf, &nroots, &nleaves, NULL, NULL);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSection);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSection, 0, nroots);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
for (l = 0; l < label->stratumSizes[s]; l++) {
ierr = PetscSectionGetDof(rootSection, label->points[s][l], &dof);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSection, label->points[s][l], dof+1);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSection);CHKERRQ(ierr);
/* Create a point-wise array of stratum values */
ierr = PetscSectionGetStorageSize(rootSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &rootStrata);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots, &rootIdx);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
for (l = 0; l < label->stratumSizes[s]; l++) {
const PetscInt p = label->points[s][l];
ierr = PetscSectionGetOffset(rootSection, p, &offset);CHKERRQ(ierr);
rootStrata[offset+rootIdx[p]++] = label->stratumValues[s];
}
}
}
/* Build SF that maps label points to remote processes */
ierr = PetscSectionCreate(comm, leafSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sf, rootSection, &remoteOffsets, *leafSection);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, rootSection, remoteOffsets, *leafSection, &labelSF);CHKERRQ(ierr);
ierr = PetscFree(remoteOffsets);CHKERRQ(ierr);
/* Send the strata for each point over the derived SF */
ierr = PetscSectionGetStorageSize(*leafSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(labelSF, MPIU_INT, rootStrata, *leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(labelSF, MPIU_INT, rootStrata, *leafStrata);CHKERRQ(ierr);
/* Clean up */
ierr = PetscFree(rootStrata);CHKERRQ(ierr);
ierr = PetscFree(rootIdx);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSection);CHKERRQ(ierr);
ierr = PetscSFDestroy(&labelSF);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
PetscSFCompose - Compose a new PetscSF equivalent to action to PetscSFs
Input Parameters:
+ sfA - The first PetscSF
- sfB - The second PetscSF
Output Parameters:
. sfBA - equvalent PetscSF for applying A then B
Level: developer
.seealso: PetscSF, PetscSFGetGraph(), PetscSFSetGraph()
@*/
PetscErrorCode PetscSFCompose(PetscSF sfA, PetscSF sfB, PetscSF *sfBA)
{
MPI_Comm comm;
const PetscSFNode *remotePointsA, *remotePointsB;
PetscSFNode *remotePointsBA;
const PetscInt *localPointsA, *localPointsB;
PetscInt numRootsA, numLeavesA, numRootsB, numLeavesB;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1);
PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 1);
ierr = PetscObjectGetComm((PetscObject) sfA, &comm);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB);CHKERRQ(ierr);
ierr = PetscMalloc1(numLeavesB, &remotePointsBA);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfB, MPIU_2INT, remotePointsA, remotePointsBA);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfB, MPIU_2INT, remotePointsA, remotePointsBA);CHKERRQ(ierr);
ierr = PetscSFCreate(comm, sfBA);CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sfBA, numRootsA, numLeavesB, localPointsB, PETSC_COPY_VALUES, remotePointsBA, PETSC_OWN_POINTER);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
PetscSFDuplicate - duplicate a PetscSF, optionally preserving rank connectivity and graph
Collective
Input Arguments:
+ sf - communication object to duplicate
- opt - PETSCSF_DUPLICATE_CONFONLY, PETSCSF_DUPLICATE_RANKS, or PETSCSF_DUPLICATE_GRAPH (see PetscSFDuplicateOption)
Output Arguments:
. newsf - new communication object
Level: beginner
.seealso: PetscSFCreate(), PetscSFSetType(), PetscSFSetGraph()
@*/
PetscErrorCode PetscSFDuplicate(PetscSF sf,PetscSFDuplicateOption opt,PetscSF *newsf)
{
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFCreate(PetscObjectComm((PetscObject)sf),newsf);CHKERRQ(ierr);
ierr = PetscSFSetType(*newsf,((PetscObject)sf)->type_name);CHKERRQ(ierr);
if (sf->ops->Duplicate) {ierr = (*sf->ops->Duplicate)(sf,opt,*newsf);CHKERRQ(ierr);}
if (opt == PETSCSF_DUPLICATE_GRAPH) {
PetscInt nroots,nleaves;
const PetscInt *ilocal;
const PetscSFNode *iremote;
ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
ierr = PetscSFSetGraph(*newsf,nroots,nleaves,ilocal,PETSC_COPY_VALUES,iremote,PETSC_COPY_VALUES);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
static PetscErrorCode CheckGraphNotSet(PetscSF sf)
{
PetscInt nroots,nleaves;
const PetscInt *ilocal;
const PetscSFNode *iremote;
PetscErrorCode ierr;
PetscFunctionBegin;
if (sf->graphset) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is set");
ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
if (nroots >= 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is set");
if (nleaves >= 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is set");
if (ilocal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is set");
if (iremote) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is set");
if (sf->minleaf != PETSC_MAX_INT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF minimum leaf is not PETSC_MAX_INT");
if (sf->maxleaf != PETSC_MIN_INT) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF minimum leaf is not PETSC_MIN_INT");
PetscFunctionReturn(0);
}
/*@C
PetscSFCreateInverseSF - given a PetscSF in which all vertices have degree 1, creates the inverse map
Collective
Input Arguments:
. sf - star forest to invert
Output Arguments:
. isf - inverse of sf
Level: advanced
Notes:
All roots must have degree 1.
The local space may be a permutation, but cannot be sparse.
.seealso: PetscSFSetGraph()
@*/
PetscErrorCode PetscSFCreateInverseSF(PetscSF sf,PetscSF *isf)
{
PetscErrorCode ierr;
PetscMPIInt rank;
PetscInt i,nroots,nleaves,maxlocal,count,*newilocal;
const PetscInt *ilocal;
PetscSFNode *roots,*leaves;
PetscFunctionBegin;
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,NULL);CHKERRQ(ierr);
for (i=0,maxlocal=0; i<nleaves; i++) maxlocal = PetscMax(maxlocal,(ilocal ? ilocal[i] : i)+1);
ierr = PetscMalloc2(nroots,&roots,maxlocal,&leaves);CHKERRQ(ierr);
for (i=0; i<maxlocal; i++) {
leaves[i].rank = rank;
leaves[i].index = i;
}
for (i=0; i <nroots; i++) {
roots[i].rank = -1;
roots[i].index = -1;
}
ierr = PetscSFReduceBegin(sf,MPIU_2INT,leaves,roots,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_2INT,leaves,roots,MPIU_REPLACE);CHKERRQ(ierr);
/* Check whether our leaves are sparse */
for (i=0,count=0; i<nroots; i++) if (roots[i].rank >= 0) count++;
if (count == nroots) newilocal = NULL;
else { /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */
ierr = PetscMalloc1(count,&newilocal);CHKERRQ(ierr);
for (i=0,count=0; i<nroots; i++) {
if (roots[i].rank >= 0) {
newilocal[count] = i;
roots[count].rank = roots[i].rank;
roots[count].index = roots[i].index;
count++;
}
}
}
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,isf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(*isf,maxlocal,count,newilocal,PETSC_OWN_POINTER,roots,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscFree2(roots,leaves);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode CheckGraphEmpty(PetscSF sf)
{
PetscInt nroots,nleaves;
const PetscInt *ilocal;
const PetscSFNode *iremote;
PetscInt minleaf,maxleaf;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
if (nroots) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is not empty");
if (nleaves) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is not empty");
if (ilocal) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is not empty");
if (iremote) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF graph is not empty");
ierr = PetscSFGetLeafRange(sf,&minleaf,&maxleaf);CHKERRQ(ierr);
if (minleaf != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF minimum leaf is not 0");
if (maxleaf != -1) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"SF maximum leaf is not -1");
PetscFunctionReturn(0);
}
/*@C
PetscSectionCreateGlobalSectionLabel - Create a section describing the global field layout using
the local section and an SF describing the section point overlap.
Input Parameters:
+ s - The PetscSection for the local field layout
. sf - The SF describing parallel layout of the section points
. includeConstraints - By default this is PETSC_FALSE, meaning that the global field vector will not possess constrained dofs
. label - The label specifying the points
- labelValue - The label stratum specifying the points
Output Parameter:
. gsection - The PetscSection for the global field layout
Note: This gives negative sizes and offsets to points not owned by this process
Level: developer
.seealso: PetscSectionCreate()
@*/
PetscErrorCode PetscSectionCreateGlobalSectionLabel(PetscSection s, PetscSF sf, PetscBool includeConstraints, DMLabel label, PetscInt labelValue, PetscSection *gsection)
{
PetscInt *neg = NULL, *tmpOff = NULL;
PetscInt pStart, pEnd, p, dof, cdof, off, globalOff = 0, nroots;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSectionCreate(PetscObjectComm((PetscObject) s), gsection);CHKERRQ(ierr);
ierr = PetscSectionGetChart(s, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionSetChart(*gsection, pStart, pEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL);CHKERRQ(ierr);
if (nroots >= 0) {
if (nroots < pEnd-pStart) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_ARG_SIZ, "PetscSF nroots %d < %d section size", nroots, pEnd-pStart);
ierr = PetscCalloc1(nroots, &neg);CHKERRQ(ierr);
if (nroots > pEnd-pStart) {
ierr = PetscCalloc1(nroots, &tmpOff);CHKERRQ(ierr);
} else {
tmpOff = &(*gsection)->atlasDof[-pStart];
}
}
/* Mark ghost points with negative dof */
for (p = pStart; p < pEnd; ++p) {
PetscInt value;
ierr = DMLabelGetValue(label, p, &value);CHKERRQ(ierr);
if (value != labelValue) continue;
ierr = PetscSectionGetDof(s, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionSetDof(*gsection, p, dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(s, p, &cdof);CHKERRQ(ierr);
if (!includeConstraints && cdof > 0) {ierr = PetscSectionSetConstraintDof(*gsection, p, cdof);CHKERRQ(ierr);}
if (neg) neg[p] = -(dof+1);
}
ierr = PetscSectionSetUpBC(*gsection);CHKERRQ(ierr);
if (nroots >= 0) {
ierr = PetscSFBcastBegin(sf, MPIU_INT, neg, tmpOff);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, neg, tmpOff);CHKERRQ(ierr);
if (nroots > pEnd-pStart) {
for (p = pStart; p < pEnd; ++p) {if (tmpOff[p] < 0) (*gsection)->atlasDof[p-pStart] = tmpOff[p];}
}
}
/* Calculate new sizes, get proccess offset, and calculate point offsets */
for (p = 0, off = 0; p < pEnd-pStart; ++p) {
cdof = (!includeConstraints && s->bc) ? s->bc->atlasDof[p] : 0;
(*gsection)->atlasOff[p] = off;
off += (*gsection)->atlasDof[p] > 0 ? (*gsection)->atlasDof[p]-cdof : 0;
}
ierr = MPI_Scan(&off, &globalOff, 1, MPIU_INT, MPI_SUM, PetscObjectComm((PetscObject) s));CHKERRQ(ierr);
globalOff -= off;
for (p = 0, off = 0; p < pEnd-pStart; ++p) {
(*gsection)->atlasOff[p] += globalOff;
if (neg) neg[p] = -((*gsection)->atlasOff[p]+1);
}
/* Put in negative offsets for ghost points */
if (nroots >= 0) {
ierr = PetscSFBcastBegin(sf, MPIU_INT, neg, tmpOff);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, neg, tmpOff);CHKERRQ(ierr);
if (nroots > pEnd-pStart) {
for (p = pStart; p < pEnd; ++p) {if (tmpOff[p] < 0) (*gsection)->atlasOff[p-pStart] = tmpOff[p];}
}
}
if (nroots >= 0 && nroots > pEnd-pStart) {ierr = PetscFree(tmpOff);CHKERRQ(ierr);}
ierr = PetscFree(neg);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
DMLabelGather - Gather all label values from leafs into roots
Input Parameters:
+ label - the DMLabel
. point - the Star Forest point communication map
Input Parameters:
+ label - the new DMLabel with localised leaf values
Level: developer
Note: This is the inverse operation to DMLabelDistribute.
.seealso: DMLabelDistribute()
@*/
PetscErrorCode DMLabelGather(DMLabel label, PetscSF sf, DMLabel *labelNew)
{
MPI_Comm comm;
PetscSection rootSection;
PetscSF sfLabel;
PetscSFNode *rootPoints, *leafPoints;
PetscInt p, s, d, nroots, nleaves, nmultiroots, idx, dof, offset;
const PetscInt *rootDegree, *ilocal;
PetscInt *rootStrata;
char *name;
PetscInt nameSize;
size_t len = 0;
PetscMPIInt rank, numProcs;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
/* Bcast name */
if (!rank) {ierr = PetscStrlen(label->name, &len);CHKERRQ(ierr);}
nameSize = len;
ierr = MPI_Bcast(&nameSize, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = PetscMalloc1(nameSize+1, &name);CHKERRQ(ierr);
if (!rank) {ierr = PetscMemcpy(name, label->name, nameSize+1);CHKERRQ(ierr);}
ierr = MPI_Bcast(name, nameSize+1, MPI_CHAR, 0, comm);CHKERRQ(ierr);
ierr = DMLabelCreate(name, labelNew);CHKERRQ(ierr);
ierr = PetscFree(name);CHKERRQ(ierr);
/* Gather rank/index pairs of leaves into local roots to build
an inverse, multi-rooted SF. Note that this ignores local leaf
indexing due to the use of the multiSF in PetscSFGather. */
ierr = PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, NULL);CHKERRQ(ierr);
ierr = PetscMalloc1(nleaves, &leafPoints);CHKERRQ(ierr);
for (p = 0; p < nleaves; p++) {
leafPoints[p].index = ilocal[p];
leafPoints[p].rank = rank;
}
ierr = PetscSFComputeDegreeBegin(sf, &rootDegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf, &rootDegree);CHKERRQ(ierr);
for (p = 0, nmultiroots = 0; p < nroots; ++p) nmultiroots += rootDegree[p];
ierr = PetscMalloc1(nmultiroots, &rootPoints);CHKERRQ(ierr);
ierr = PetscSFGatherBegin(sf, MPIU_2INT, leafPoints, rootPoints);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(sf, MPIU_2INT, leafPoints, rootPoints);CHKERRQ(ierr);
ierr = PetscSFCreate(comm,& sfLabel);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfLabel, nroots, nmultiroots, NULL, PETSC_OWN_POINTER, rootPoints, PETSC_OWN_POINTER);CHKERRQ(ierr);
/* Migrate label over inverted SF to pull stratum values at leaves into roots. */
ierr = DMLabelDistribute_Internal(label, sfLabel, &rootSection, &rootStrata);CHKERRQ(ierr);
/* Rebuild the point strata on the receiver */
for (p = 0, idx = 0; p < nroots; p++) {
for (d = 0; d < rootDegree[p]; d++) {
ierr = PetscSectionGetDof(rootSection, idx+d, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSection, idx+d, &offset);CHKERRQ(ierr);
for (s = 0; s < dof; s++) {ierr = DMLabelSetValue(*labelNew, p, rootStrata[offset+s]);CHKERRQ(ierr);}
}
idx += rootDegree[p];
}
ierr = PetscFree(leafPoints);CHKERRQ(ierr);
ierr = PetscFree(rootStrata);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSection);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfLabel);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
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);
}
PETSC_EXTERN PetscErrorCode MatColoringLocalColor(MatColoring mc,PetscSF etoc,PetscSF etor,PetscReal *wts,ISColoringValue *color, ISColoringValue *maxcolor)
{
PetscInt nrows,ncols,ncolentries,nrowentries,idx,neighoffset;
PetscInt i,j,k;
PetscInt dist = mc->dist;
PetscInt totalcolors;
PetscBool *colormask;
PetscErrorCode ierr;
PetscBool *rowseen,*colseen;
const PetscInt *rowdegrees;
PetscInt *rowoffsets;
const PetscInt *coldegrees;
PetscInt *coloffsets;
PetscInt offset;
PetscInt *ll_ptr;
PetscInt *ll_idx;
PetscReal *swts;
PetscInt *sidx;
PetscInt unused;
PetscInt rowlist,collist;
PetscInt swp;
PetscMPIInt rank;
const PetscSFNode *colentries,*rowentries;
PetscFunctionBegin;
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mc),&rank);CHKERRQ(ierr);
ierr = PetscSFGetGraph(etoc,&ncols,&ncolentries,NULL,&colentries);CHKERRQ(ierr);
ierr = PetscSFGetGraph(etor,&nrows,&nrowentries,NULL,&rowentries);CHKERRQ(ierr);
ierr = PetscMalloc6(nrows,&rowseen,ncols,&colseen,ncols,&coloffsets,nrows,&rowoffsets,2*ncols,&ll_ptr,2*ncols,&ll_idx);CHKERRQ(ierr);
ierr = PetscMalloc2(ncols,&sidx,ncols,&swts);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(etoc,&rowdegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(etoc,&rowdegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(etor,&coldegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(etor,&coldegrees);CHKERRQ(ierr);
/* sort by weight */
for (i=0;i<ncols;i++) {
sidx[i] = i;
swts[i] = wts[i];
}
ierr = PetscSortRealWithPermutation(ncols,swts,sidx);CHKERRQ(ierr);
for (i=0;i<ncols/2;i++) {
swp = sidx[i];
sidx[i] = sidx[ncols-1-i];
sidx[ncols-1-i] = swp;
}
/* set up the "unused" linked list */
unused = 0;
ll_ptr[2*ncols-1] = -1;
for (i=0;i<2*ncols-1;i++) {
ll_ptr[i] = i+1;
}
/* initialize the offsets */
offset=0;
for (i=0;i<ncols;i++) {
coloffsets[i] = offset;
offset+=coldegrees[i];
colseen[i] = PETSC_FALSE;
}
offset=0;
for (i=0;i<nrows;i++) {
rowoffsets[i] = offset;
offset+=rowdegrees[i];
rowseen[i] = PETSC_FALSE;
}
/* discover the maximum current color */
totalcolors = 1;
for (i=0;i<ncols;i++) {
if (color[i] > totalcolors-1 && color[i] != IS_COLORING_MAX) totalcolors = color[i]+1;
}
if (totalcolors < 10) totalcolors=10;
ierr = PetscMalloc(sizeof(PetscBool)*totalcolors,&colormask);CHKERRQ(ierr);
/* alternate between rows and columns to get the distance k minimum coloring */
for (i=0;i<ncols;i++) {
collist = -1;
rowlist = -1;
if (color[sidx[i]] == IS_COLORING_MAX) {
for (j=0;j<totalcolors;j++) colormask[j] = PETSC_FALSE;
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = sidx[i];
collist = swp;
colseen[sidx[i]] = PETSC_TRUE;
for (k=0;k<=dist;k++) {
if (k % 2 == 0) {
while (collist >= 0) {
if (k != dist) {
for (j=0;j<coldegrees[ll_idx[collist]];j++) {
neighoffset = coloffsets[ll_idx[collist]]+j;
idx = colentries[neighoffset].index;
if (colentries[neighoffset].rank == rank && !rowseen[idx]) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = rowlist;
ll_idx[swp] = idx;
rowlist = swp;
rowseen[idx] = PETSC_TRUE;
}
}
}
if (color[ll_idx[collist]] != IS_COLORING_MAX) colormask[color[ll_idx[collist]]] = PETSC_TRUE;
colseen[ll_idx[collist]] = PETSC_FALSE;
swp = collist;
collist = ll_ptr[collist];
ll_ptr[swp] = unused;
unused = swp;
}
} else {
while (rowlist >= 0) {
if (k != dist) {
for (j=0;j<rowdegrees[ll_idx[rowlist]];j++) {
neighoffset = rowoffsets[ll_idx[rowlist]]+j;
idx = rowentries[neighoffset].index;
if (rowentries[neighoffset].rank == rank && !colseen[idx]) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = idx;
collist = swp;
colseen[idx] = PETSC_TRUE;
}
}
}
if (color[ll_idx[rowlist]] != IS_COLORING_MAX) colormask[color[ll_idx[rowlist]]] = PETSC_TRUE;
rowseen[ll_idx[rowlist]] = PETSC_FALSE;
swp = rowlist;
rowlist = ll_ptr[rowlist];
ll_ptr[swp] = unused;
unused = swp;
}
}
}
color[sidx[i]] = totalcolors;
for (k=0;k<totalcolors;k++) {
if (!colormask[k]) {color[sidx[i]] = k; break;}
}
if (color[sidx[i]] >= mc->maxcolors && mc->maxcolors > 0) color[sidx[i]] = mc->maxcolors;
if (color[sidx[i]] > *maxcolor) *maxcolor = color[sidx[i]];
if (color[sidx[i]] > totalcolors-1) {
totalcolors *= 2;
ierr = PetscFree(colormask);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(PetscBool)*totalcolors,&colormask);CHKERRQ(ierr);
}
}
if (collist != -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in local coloring BFS -- column queue still has %d\n",collist);
if (rowlist != -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in local coloring BFS -- row queue still has %d\n",rowlist);
}
for (i=0;i<ncols;i++) {
if (colseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in local coloring BFS -- column %d still seen\n",i);}
}
for (i=0;i<nrows;i++) {
if (rowseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in local coloring BFS -- row %d still seen\n",i);}
}
ierr = PetscFree6(rowseen,colseen,coloffsets,rowoffsets,ll_ptr,ll_idx);CHKERRQ(ierr);
ierr = PetscFree2(sidx,swts);CHKERRQ(ierr);
ierr = PetscFree(colormask);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PETSC_EXTERN PetscErrorCode MatColoringDiscoverBoundary(MatColoring mc,PetscSF etoc,PetscSF etor,PetscInt *nboundary,PetscInt **boundary)
{
PetscInt nrows,ncols,ncolentries,nrowentries,idx,bidx,neighoffset;
PetscInt i,j,k;
PetscInt dist = mc->dist;
PetscBool onBoundary;
PetscErrorCode ierr;
PetscBool *rowseen,*colseen;
const PetscInt *rowdegrees;
PetscInt *rowoffsets;
const PetscInt *coldegrees;
PetscInt *coloffsets;
PetscInt offset;
PetscInt *ll_ptr;
PetscInt *ll_idx;
PetscInt unused;
PetscInt rowlist,collist;
PetscInt swp;
PetscMPIInt rank;
const PetscSFNode *colentries,*rowentries;
PetscFunctionBegin;
*nboundary = 0;
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mc),&rank);CHKERRQ(ierr);
ierr = PetscSFGetGraph(etoc,&ncols,&ncolentries,NULL,&colentries);CHKERRQ(ierr);
ierr = PetscSFGetGraph(etor,&nrows,&nrowentries,NULL,&rowentries);CHKERRQ(ierr);
ierr = PetscMalloc6(nrows,&rowseen,
ncols,&colseen,
ncols,&coloffsets,
nrows,&rowoffsets,
2*ncols,&ll_ptr,
2*ncols,&ll_idx);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(etoc,&rowdegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(etoc,&rowdegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(etor,&coldegrees);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(etor,&coldegrees);CHKERRQ(ierr);
/* set up the "unused" linked list -- double the size of the number of items as in tiny or large distance cases we may have a clique */
unused = 0;
ll_ptr[2*ncols-1] = -1;
for (i=0;i<2*ncols-1;i++) {
ll_ptr[i] = i+1;
}
/* initialize the offsets */
offset=0;
for (i=0;i<ncols;i++) {
coloffsets[i] = offset;
offset+=coldegrees[i];
colseen[i] = PETSC_FALSE;
}
offset=0;
for (i=0;i<nrows;i++) {
rowoffsets[i] = offset;
offset+=rowdegrees[i];
rowseen[i] = PETSC_FALSE;
}
/* count the number of boundary nodes */
for (i=0;i<ncols;i++) {
onBoundary = PETSC_FALSE;
collist = -1;
rowlist = -1;
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = i;
collist = swp;
colseen[i] = PETSC_TRUE;
for (k=0;k<=dist;k++) {
if (k % 2 == 0) {
while (collist >= 0) {
if (k != dist) {
for (j=0;j<coldegrees[ll_idx[collist]];j++) {
neighoffset = coloffsets[ll_idx[collist]]+j;
idx = colentries[neighoffset].index;
if (colentries[neighoffset].rank == rank) {
if (!rowseen[idx] && !onBoundary) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = rowlist;
rowlist = swp;
ll_idx[swp] = idx;
rowseen[idx] = PETSC_TRUE;
}
} else {
onBoundary = PETSC_TRUE;
}
}
}
colseen[ll_idx[collist]] = PETSC_FALSE;
swp = collist;
collist = ll_ptr[collist];
ll_ptr[swp] = unused;
unused = swp;
}
} else {
while (rowlist >= 0) {
if (k != dist) {
for (j=0;j<rowdegrees[ll_idx[rowlist]];j++) {
neighoffset = rowoffsets[ll_idx[rowlist]]+j;
if (rowentries[neighoffset].rank == rank) {
idx = rowentries[neighoffset].index;
if (!colseen[idx] && !onBoundary) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = idx;
colseen[idx] = PETSC_TRUE;
collist = swp;
}
} else {
onBoundary = PETSC_TRUE;
}
}
}
rowseen[ll_idx[rowlist]] = PETSC_FALSE;
swp = rowlist;
rowlist = ll_ptr[rowlist];
ll_ptr[swp] = unused;
unused = swp;
}
}
}
if (onBoundary) {(*nboundary)++;}
if (collist != -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary count BFS -- column queue still has %d\n",collist);
if (rowlist != -1) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary count BFS -- row queue still has %d\n",collist);
}
for (i=0;i<ncols;i++) {
if (colseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary count BFS -- column %d still seen\n",i);}
}
for (i=0;i<nrows;i++) {
if (rowseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary count BFS -- row %d still seen\n",i);}
}
ierr = PetscMalloc(sizeof(PetscInt)*(*nboundary),boundary);CHKERRQ(ierr);
/* set the boundary nodes */
bidx=0;
for (i=0;i<ncols;i++) {
onBoundary = PETSC_FALSE;
collist = -1;
rowlist = -1;
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = i;
collist = swp;
colseen[i] = PETSC_TRUE;
for (k=0;k<=dist;k++) {
if (k % 2 == 0) {
while (collist >= 0) {
if (k != dist) {
for (j=0;j<coldegrees[ll_idx[collist]];j++) {
neighoffset = coloffsets[ll_idx[collist]]+j;
idx = colentries[neighoffset].index;
if (colentries[neighoffset].rank == rank) {
if (!rowseen[idx] && !onBoundary) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = rowlist;
rowlist = swp;
ll_idx[swp] = idx;
rowseen[idx] = PETSC_TRUE;
}
} else {
onBoundary = PETSC_TRUE;
}
}
}
colseen[ll_idx[collist]] = PETSC_FALSE;
swp = collist;
collist = ll_ptr[collist];
ll_ptr[swp] = unused;
unused = swp;
}
} else {
while (rowlist >= 0) {
if (k != dist) {
for (j=0;j<rowdegrees[ll_idx[rowlist]];j++) {
neighoffset = rowoffsets[ll_idx[rowlist]]+j;
if (rowentries[neighoffset].rank == rank) {
idx = rowentries[neighoffset].index;
if (!colseen[idx] && !onBoundary) {
swp = unused;
unused = ll_ptr[unused];
ll_ptr[swp] = collist;
ll_idx[swp] = idx;
colseen[idx] = PETSC_TRUE;
collist = swp;
}
} else {
onBoundary = PETSC_TRUE;
}
}
}
rowseen[ll_idx[rowlist]] = PETSC_FALSE;
swp = rowlist;
rowlist = ll_ptr[rowlist];
ll_ptr[swp] = unused;
unused = swp;
}
}
}
if (onBoundary) {(*boundary)[bidx] = i; bidx++;}
}
for (i=0;i<ncols;i++) {
if (colseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary set BFS -- column %d still seen\n",i);}
}
for (i=0;i<nrows;i++) {
if (rowseen[i]) {SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_NOT_CONVERGED,"Likely error in boundary set BFS -- row %d still seen\n",i);}
}
if (bidx != *nboundary) {SETERRQ(PetscObjectComm((PetscObject)mc),PETSC_ERR_NOT_CONVERGED,"Number of boundary nodes not matched");}
ierr = PetscFree6(rowseen,colseen,coloffsets,rowoffsets,ll_ptr,ll_idx);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);
}
PetscErrorCode DMLabelDistribute(DMLabel label, PetscSF sf, DMLabel *labelNew)
{
MPI_Comm comm;
PetscSection rootSection, leafSection;
PetscSF labelSF;
PetscInt p, pStart, pEnd, l, lStart, lEnd, s, nroots, nleaves, size, dof, offset, stratum;
PetscInt *remoteOffsets, *rootStrata, *rootIdx, *leafStrata, *strataIdx;
char *name;
PetscInt nameSize;
size_t len = 0;
PetscMPIInt rank, numProcs;
PetscErrorCode ierr;
PetscFunctionBegin;
if (label) {ierr = DMLabelMakeAllValid_Private(label);CHKERRQ(ierr);}
ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
/* Bcast name */
if (!rank) {ierr = PetscStrlen(label->name, &len);CHKERRQ(ierr);}
nameSize = len;
ierr = MPI_Bcast(&nameSize, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = PetscMalloc1(nameSize+1, &name);CHKERRQ(ierr);
if (!rank) {ierr = PetscMemcpy(name, label->name, nameSize+1);CHKERRQ(ierr);}
ierr = MPI_Bcast(name, nameSize+1, MPI_CHAR, 0, comm);CHKERRQ(ierr);
ierr = DMLabelCreate(name, labelNew);CHKERRQ(ierr);
ierr = PetscFree(name);CHKERRQ(ierr);
/* Bcast numStrata */
if (!rank) (*labelNew)->numStrata = label->numStrata;
ierr = MPI_Bcast(&(*labelNew)->numStrata, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
/* Bcast stratumValues */
ierr = PetscMalloc1((*labelNew)->numStrata, &(*labelNew)->stratumValues);CHKERRQ(ierr);
if (!rank) {ierr = PetscMemcpy((*labelNew)->stratumValues, label->stratumValues, label->numStrata * sizeof(PetscInt));CHKERRQ(ierr);}
ierr = MPI_Bcast((*labelNew)->stratumValues, (*labelNew)->numStrata, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = PetscMalloc1((*labelNew)->numStrata, &(*labelNew)->arrayValid);CHKERRQ(ierr);
for (s = 0; s < (*labelNew)->numStrata; ++s) (*labelNew)->arrayValid[s] = PETSC_TRUE;
/* Build a section detailing strata-per-point, distribute and build SF
from that and then send our points. */
ierr = PetscSFGetGraph(sf, &nroots, &nleaves, NULL, NULL);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSection);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSection, 0, nroots);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
lStart = 0;
lEnd = label->stratumSizes[s];
for (l=lStart; l<lEnd; l++) {
ierr = PetscSectionGetDof(rootSection, label->points[s][l], &dof);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSection, label->points[s][l], dof+1);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSection);CHKERRQ(ierr);
/* Create a point-wise array of point strata */
ierr = PetscSectionGetStorageSize(rootSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &rootStrata);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots, &rootIdx);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
lStart = 0;
lEnd = label->stratumSizes[s];
for (l=lStart; l<lEnd; l++) {
p = label->points[s][l];
ierr = PetscSectionGetOffset(rootSection, p, &offset);CHKERRQ(ierr);
rootStrata[offset+rootIdx[p]++] = s;
}
}
}
/* Build SF that maps label points to remote processes */
ierr = PetscSectionCreate(comm, &leafSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sf, rootSection, &remoteOffsets, leafSection);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, rootSection, remoteOffsets, leafSection, &labelSF);CHKERRQ(ierr);
/* Send the strata for each point over the derived SF */
ierr = PetscSectionGetStorageSize(leafSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(labelSF, MPIU_INT, rootStrata, leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(labelSF, MPIU_INT, rootStrata, leafStrata);CHKERRQ(ierr);
/* Rebuild the point strata on the receiver */
ierr = PetscCalloc1((*labelNew)->numStrata,&(*labelNew)->stratumSizes);CHKERRQ(ierr);
ierr = PetscSectionGetChart(leafSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p=pStart; p<pEnd; p++) {
ierr = PetscSectionGetDof(leafSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(leafSection, p, &offset);CHKERRQ(ierr);
for (s=0; s<dof; s++) {
(*labelNew)->stratumSizes[leafStrata[offset+s]]++;
}
}
ierr = PetscCalloc1((*labelNew)->numStrata,&(*labelNew)->ht);CHKERRQ(ierr);
ierr = PetscMalloc1((*labelNew)->numStrata,&(*labelNew)->points);CHKERRQ(ierr);
for (s = 0; s < (*labelNew)->numStrata; ++s) {
PetscHashICreate((*labelNew)->ht[s]);
ierr = PetscMalloc1((*labelNew)->stratumSizes[s], &(*labelNew)->points[s]);CHKERRQ(ierr);
}
/* Insert points into new strata */
ierr = PetscCalloc1((*labelNew)->numStrata, &strataIdx);CHKERRQ(ierr);
ierr = PetscSectionGetChart(leafSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p=pStart; p<pEnd; p++) {
ierr = PetscSectionGetDof(leafSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(leafSection, p, &offset);CHKERRQ(ierr);
for (s=0; s<dof; s++) {
stratum = leafStrata[offset+s];
(*labelNew)->points[stratum][strataIdx[stratum]++] = p;
}
}
ierr = PetscFree(rootStrata);CHKERRQ(ierr);
ierr = PetscFree(leafStrata);CHKERRQ(ierr);
ierr = PetscFree(rootIdx);CHKERRQ(ierr);
ierr = PetscFree(strataIdx);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSection);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&leafSection);CHKERRQ(ierr);
ierr = PetscSFDestroy(&labelSF);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
DMPlexOrient - Give a consistent orientation to the input mesh
Input Parameters:
. dm - The DM
Note: The orientation data for the DM are change in-place.
$ This routine will fail for non-orientable surfaces, such as the Moebius strip.
Level: advanced
.seealso: DMCreate(), DMPLEX
@*/
PetscErrorCode DMPlexOrient(DM dm)
{
MPI_Comm comm;
PetscSF sf;
const PetscInt *lpoints;
const PetscSFNode *rpoints;
PetscSFNode *rorntComp = NULL, *lorntComp = NULL;
PetscInt *numNeighbors, **neighbors;
PetscSFNode *nrankComp;
PetscBool *match, *flipped;
PetscBT seenCells, flippedCells, seenFaces;
PetscInt *faceFIFO, fTop, fBottom, *cellComp, *faceComp;
PetscInt numLeaves, numRoots, dim, h, cStart, cEnd, c, cell, fStart, fEnd, face, off, totNeighbors = 0;
PetscMPIInt rank, size, numComponents, comp = 0;
PetscBool flg, flg2;
PetscViewer viewer = NULL, selfviewer = NULL;
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 = PetscOptionsHasName(((PetscObject) dm)->options,((PetscObject) dm)->prefix, "-orientation_view", &flg);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->options,((PetscObject) dm)->prefix, "-orientation_view_synchronized", &flg2);CHKERRQ(ierr);
ierr = DMGetPointSF(dm, &sf);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &numRoots, &numLeaves, &lpoints, &rpoints);CHKERRQ(ierr);
/* Truth Table
mismatch flips do action mismatch flipA ^ flipB action
F 0 flips no F F F
F 1 flip yes F T T
F 2 flips no T F T
T 0 flips yes T T F
T 1 flip no
T 2 flips yes
*/
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMPlexGetVTKCellHeight(dm, &h);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, h, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, h+1, &fStart, &fEnd);CHKERRQ(ierr);
ierr = PetscBTCreate(cEnd - cStart, &seenCells);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, seenCells);CHKERRQ(ierr);
ierr = PetscBTCreate(cEnd - cStart, &flippedCells);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, flippedCells);CHKERRQ(ierr);
ierr = PetscBTCreate(fEnd - fStart, &seenFaces);CHKERRQ(ierr);
ierr = PetscBTMemzero(fEnd - fStart, seenFaces);CHKERRQ(ierr);
ierr = PetscCalloc3(fEnd - fStart, &faceFIFO, cEnd-cStart, &cellComp, fEnd-fStart, &faceComp);CHKERRQ(ierr);
/*
OLD STYLE
- Add an integer array over cells and faces (component) for connected component number
Foreach component
- Mark the initial cell as seen
- Process component as usual
- Set component for all seenCells
- Wipe seenCells and seenFaces (flippedCells can stay)
- Generate parallel adjacency for component using SF and seenFaces
- Collect numComponents adj data from each proc to 0
- Build same serial graph
- Use same solver
- Use Scatterv to to send back flipped flags for each component
- Negate flippedCells by component
NEW STYLE
- Create the adj on each process
- Bootstrap to complete graph on proc 0
*/
/* Loop over components */
for (cell = cStart; cell < cEnd; ++cell) cellComp[cell-cStart] = -1;
do {
/* Look for first unmarked cell */
for (cell = cStart; cell < cEnd; ++cell) if (cellComp[cell-cStart] < 0) break;
if (cell >= cEnd) break;
/* Initialize FIFO with first cell in component */
{
const PetscInt *cone;
PetscInt coneSize;
fTop = fBottom = 0;
ierr = DMPlexGetConeSize(dm, cell, &coneSize);CHKERRQ(ierr);
ierr = DMPlexGetCone(dm, cell, &cone);CHKERRQ(ierr);
for (c = 0; c < coneSize; ++c) {
faceFIFO[fBottom++] = cone[c];
ierr = PetscBTSet(seenFaces, cone[c]-fStart);CHKERRQ(ierr);
}
ierr = PetscBTSet(seenCells, cell-cStart);CHKERRQ(ierr);
}
/* Consider each face in FIFO */
while (fTop < fBottom) {
ierr = DMPlexCheckFace_Internal(dm, faceFIFO, &fTop, &fBottom, cStart, fStart, fEnd, seenCells, flippedCells, seenFaces);CHKERRQ(ierr);
}
/* Set component for cells and faces */
for (cell = 0; cell < cEnd-cStart; ++cell) {
if (PetscBTLookup(seenCells, cell)) cellComp[cell] = comp;
}
for (face = 0; face < fEnd-fStart; ++face) {
if (PetscBTLookup(seenFaces, face)) faceComp[face] = comp;
}
/* Wipe seenCells and seenFaces for next component */
ierr = PetscBTMemzero(fEnd - fStart, seenFaces);CHKERRQ(ierr);
ierr = PetscBTMemzero(cEnd - cStart, seenCells);CHKERRQ(ierr);
++comp;
} while (1);
numComponents = comp;
if (flg) {
PetscViewer v;
ierr = PetscViewerASCIIGetStdout(comm, &v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPushSynchronized(v);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(v, "[%d]BT for serial flipped cells:\n", rank);CHKERRQ(ierr);
ierr = PetscBTView(cEnd-cStart, flippedCells, v);CHKERRQ(ierr);
ierr = PetscViewerFlush(v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPopSynchronized(v);CHKERRQ(ierr);
}
/* Now all subdomains are oriented, but we need a consistent parallel orientation */
if (numLeaves >= 0) {
/* Store orientations of boundary faces*/
ierr = PetscCalloc2(numRoots,&rorntComp,numRoots,&lorntComp);CHKERRQ(ierr);
for (face = fStart; face < fEnd; ++face) {
const PetscInt *cone, *support, *ornt;
PetscInt coneSize, supportSize;
ierr = DMPlexGetSupportSize(dm, face, &supportSize);CHKERRQ(ierr);
if (supportSize != 1) continue;
ierr = DMPlexGetSupport(dm, face, &support);CHKERRQ(ierr);
ierr = DMPlexGetCone(dm, support[0], &cone);CHKERRQ(ierr);
ierr = DMPlexGetConeSize(dm, support[0], &coneSize);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientation(dm, support[0], &ornt);CHKERRQ(ierr);
for (c = 0; c < coneSize; ++c) if (cone[c] == face) break;
if (dim == 1) {
/* Use cone position instead, shifted to -1 or 1 */
if (PetscBTLookup(flippedCells, support[0]-cStart)) rorntComp[face].rank = 1-c*2;
else rorntComp[face].rank = c*2-1;
} else {
if (PetscBTLookup(flippedCells, support[0]-cStart)) rorntComp[face].rank = ornt[c] < 0 ? -1 : 1;
else rorntComp[face].rank = ornt[c] < 0 ? 1 : -1;
}
rorntComp[face].index = faceComp[face-fStart];
}
/* Communicate boundary edge orientations */
ierr = PetscSFBcastBegin(sf, MPIU_2INT, rorntComp, lorntComp);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_2INT, rorntComp, lorntComp);CHKERRQ(ierr);
}
/* Get process adjacency */
ierr = PetscMalloc2(numComponents, &numNeighbors, numComponents, &neighbors);CHKERRQ(ierr);
viewer = PETSC_VIEWER_STDOUT_(PetscObjectComm((PetscObject)dm));
if (flg2) {ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr);}
ierr = PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&selfviewer);CHKERRQ(ierr);
for (comp = 0; comp < numComponents; ++comp) {
PetscInt l, n;
numNeighbors[comp] = 0;
ierr = PetscMalloc1(PetscMax(numLeaves, 0), &neighbors[comp]);CHKERRQ(ierr);
/* I know this is p^2 time in general, but for bounded degree its alright */
for (l = 0; l < numLeaves; ++l) {
const PetscInt face = lpoints[l];
/* Find a representative face (edge) separating pairs of procs */
if ((face >= fStart) && (face < fEnd) && (faceComp[face-fStart] == comp)) {
const PetscInt rrank = rpoints[l].rank;
const PetscInt rcomp = lorntComp[face].index;
for (n = 0; n < numNeighbors[comp]; ++n) if ((rrank == rpoints[neighbors[comp][n]].rank) && (rcomp == lorntComp[lpoints[neighbors[comp][n]]].index)) break;
if (n >= numNeighbors[comp]) {
PetscInt supportSize;
ierr = DMPlexGetSupportSize(dm, face, &supportSize);CHKERRQ(ierr);
if (supportSize != 1) SETERRQ1(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Boundary faces should see one cell, not %d", supportSize);
if (flg) {ierr = PetscViewerASCIIPrintf(selfviewer, "[%d]: component %d, Found representative leaf %d (face %d) connecting to face %d on (%d, %d) with orientation %d\n", rank, comp, l, face, rpoints[l].index, rrank, rcomp, lorntComp[face].rank);CHKERRQ(ierr);}
neighbors[comp][numNeighbors[comp]++] = l;
}
}
}
totNeighbors += numNeighbors[comp];
}
ierr = PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&selfviewer);CHKERRQ(ierr);
ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
if (flg2) {ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr);}
ierr = PetscMalloc2(totNeighbors, &nrankComp, totNeighbors, &match);CHKERRQ(ierr);
for (comp = 0, off = 0; comp < numComponents; ++comp) {
PetscInt n;
for (n = 0; n < numNeighbors[comp]; ++n, ++off) {
const PetscInt face = lpoints[neighbors[comp][n]];
const PetscInt o = rorntComp[face].rank*lorntComp[face].rank;
if (o < 0) match[off] = PETSC_TRUE;
else if (o > 0) match[off] = PETSC_FALSE;
else SETERRQ5(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Invalid face %d (%d, %d) neighbor: %d comp: %d", face, rorntComp[face], lorntComp[face], neighbors[comp][n], comp);
nrankComp[off].rank = rpoints[neighbors[comp][n]].rank;
nrankComp[off].index = lorntComp[lpoints[neighbors[comp][n]]].index;
}
ierr = PetscFree(neighbors[comp]);CHKERRQ(ierr);
}
/* Collect the graph on 0 */
if (numLeaves >= 0) {
Mat G;
PetscBT seenProcs, flippedProcs;
PetscInt *procFIFO, pTop, pBottom;
PetscInt *N = NULL, *Noff;
PetscSFNode *adj = NULL;
PetscBool *val = NULL;
PetscMPIInt *recvcounts = NULL, *displs = NULL, *Nc, p, o;
PetscMPIInt size = 0;
ierr = PetscCalloc1(numComponents, &flipped);CHKERRQ(ierr);
if (!rank) {ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);}
ierr = PetscCalloc4(size, &recvcounts, size+1, &displs, size, &Nc, size+1, &Noff);CHKERRQ(ierr);
ierr = MPI_Gather(&numComponents, 1, MPI_INT, Nc, 1, MPI_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < size; ++p) {
displs[p+1] = displs[p] + Nc[p];
}
if (!rank) {ierr = PetscMalloc1(displs[size],&N);CHKERRQ(ierr);}
ierr = MPI_Gatherv(numNeighbors, numComponents, MPIU_INT, N, Nc, displs, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0, o = 0; p < size; ++p) {
recvcounts[p] = 0;
for (c = 0; c < Nc[p]; ++c, ++o) recvcounts[p] += N[o];
displs[p+1] = displs[p] + recvcounts[p];
}
if (!rank) {ierr = PetscMalloc2(displs[size], &adj, displs[size], &val);CHKERRQ(ierr);}
ierr = MPI_Gatherv(nrankComp, totNeighbors, MPIU_2INT, adj, recvcounts, displs, MPIU_2INT, 0, comm);CHKERRQ(ierr);
ierr = MPI_Gatherv(match, totNeighbors, MPIU_BOOL, val, recvcounts, displs, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
ierr = PetscFree2(numNeighbors, neighbors);CHKERRQ(ierr);
if (!rank) {
for (p = 1; p <= size; ++p) {Noff[p] = Noff[p-1] + Nc[p-1];}
if (flg) {
PetscInt n;
for (p = 0, off = 0; p < size; ++p) {
for (c = 0; c < Nc[p]; ++c) {
ierr = PetscPrintf(PETSC_COMM_SELF, "Proc %d Comp %d:\n", p, c);CHKERRQ(ierr);
for (n = 0; n < N[Noff[p]+c]; ++n, ++off) {
ierr = PetscPrintf(PETSC_COMM_SELF, " edge (%d, %d) (%d):\n", adj[off].rank, adj[off].index, val[off]);CHKERRQ(ierr);
}
}
}
}
/* Symmetrize the graph */
ierr = MatCreate(PETSC_COMM_SELF, &G);CHKERRQ(ierr);
ierr = MatSetSizes(G, Noff[size], Noff[size], Noff[size], Noff[size]);CHKERRQ(ierr);
ierr = MatSetUp(G);CHKERRQ(ierr);
for (p = 0, off = 0; p < size; ++p) {
for (c = 0; c < Nc[p]; ++c) {
const PetscInt r = Noff[p]+c;
PetscInt n;
for (n = 0; n < N[r]; ++n, ++off) {
const PetscInt q = Noff[adj[off].rank] + adj[off].index;
const PetscScalar o = val[off] ? 1.0 : 0.0;
ierr = MatSetValues(G, 1, &r, 1, &q, &o, INSERT_VALUES);CHKERRQ(ierr);
ierr = MatSetValues(G, 1, &q, 1, &r, &o, INSERT_VALUES);CHKERRQ(ierr);
}
}
}
ierr = MatAssemblyBegin(G, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(G, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = PetscBTCreate(Noff[size], &seenProcs);CHKERRQ(ierr);
ierr = PetscBTMemzero(Noff[size], seenProcs);CHKERRQ(ierr);
ierr = PetscBTCreate(Noff[size], &flippedProcs);CHKERRQ(ierr);
ierr = PetscBTMemzero(Noff[size], flippedProcs);CHKERRQ(ierr);
ierr = PetscMalloc1(Noff[size], &procFIFO);CHKERRQ(ierr);
pTop = pBottom = 0;
for (p = 0; p < Noff[size]; ++p) {
if (PetscBTLookup(seenProcs, p)) continue;
/* Initialize FIFO with next proc */
procFIFO[pBottom++] = p;
ierr = PetscBTSet(seenProcs, p);CHKERRQ(ierr);
/* Consider each proc in FIFO */
while (pTop < pBottom) {
const PetscScalar *ornt;
const PetscInt *neighbors;
PetscInt proc, nproc, seen, flippedA, flippedB, mismatch, numNeighbors, n;
proc = procFIFO[pTop++];
flippedA = PetscBTLookup(flippedProcs, proc) ? 1 : 0;
ierr = MatGetRow(G, proc, &numNeighbors, &neighbors, &ornt);CHKERRQ(ierr);
/* Loop over neighboring procs */
for (n = 0; n < numNeighbors; ++n) {
nproc = neighbors[n];
mismatch = PetscRealPart(ornt[n]) > 0.5 ? 0 : 1;
seen = PetscBTLookup(seenProcs, nproc);
flippedB = PetscBTLookup(flippedProcs, nproc) ? 1 : 0;
if (mismatch ^ (flippedA ^ flippedB)) {
if (seen) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Previously seen procs %d and %d do not match: Fault mesh is non-orientable", proc, nproc);
if (!flippedB) {
ierr = PetscBTSet(flippedProcs, nproc);CHKERRQ(ierr);
} else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Inconsistent mesh orientation: Fault mesh is non-orientable");
} else if (mismatch && flippedA && flippedB) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Attempt to flip already flipped cell: Fault mesh is non-orientable");
if (!seen) {
procFIFO[pBottom++] = nproc;
ierr = PetscBTSet(seenProcs, nproc);CHKERRQ(ierr);
}
}
}
}
ierr = PetscFree(procFIFO);CHKERRQ(ierr);
ierr = MatDestroy(&G);CHKERRQ(ierr);
ierr = PetscFree2(adj, val);CHKERRQ(ierr);
ierr = PetscBTDestroy(&seenProcs);CHKERRQ(ierr);
}
/* Scatter flip flags */
{
PetscBool *flips = NULL;
if (!rank) {
ierr = PetscMalloc1(Noff[size], &flips);CHKERRQ(ierr);
for (p = 0; p < Noff[size]; ++p) {
flips[p] = PetscBTLookup(flippedProcs, p) ? PETSC_TRUE : PETSC_FALSE;
if (flg && flips[p]) {ierr = PetscPrintf(comm, "Flipping Proc+Comp %d:\n", p);CHKERRQ(ierr);}
}
for (p = 0; p < size; ++p) {
displs[p+1] = displs[p] + Nc[p];
}
}
ierr = MPI_Scatterv(flips, Nc, displs, MPIU_BOOL, flipped, numComponents, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
ierr = PetscFree(flips);CHKERRQ(ierr);
}
if (!rank) {ierr = PetscBTDestroy(&flippedProcs);CHKERRQ(ierr);}
ierr = PetscFree(N);CHKERRQ(ierr);
ierr = PetscFree4(recvcounts, displs, Nc, Noff);CHKERRQ(ierr);
ierr = PetscFree2(nrankComp, match);CHKERRQ(ierr);
/* Decide whether to flip cells in each component */
for (c = 0; c < cEnd-cStart; ++c) {if (flipped[cellComp[c]]) {ierr = PetscBTNegate(flippedCells, c);CHKERRQ(ierr);}}
ierr = PetscFree(flipped);CHKERRQ(ierr);
}
if (flg) {
PetscViewer v;
ierr = PetscViewerASCIIGetStdout(comm, &v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPushSynchronized(v);CHKERRQ(ierr);
ierr = PetscViewerASCIISynchronizedPrintf(v, "[%d]BT for parallel flipped cells:\n", rank);CHKERRQ(ierr);
ierr = PetscBTView(cEnd-cStart, flippedCells, v);CHKERRQ(ierr);
ierr = PetscViewerFlush(v);CHKERRQ(ierr);
ierr = PetscViewerASCIIPopSynchronized(v);CHKERRQ(ierr);
}
/* Reverse flipped cells in the mesh */
for (c = cStart; c < cEnd; ++c) {
if (PetscBTLookup(flippedCells, c-cStart)) {
ierr = DMPlexReverseCell(dm, c);CHKERRQ(ierr);
}
}
ierr = PetscBTDestroy(&seenCells);CHKERRQ(ierr);
ierr = PetscBTDestroy(&flippedCells);CHKERRQ(ierr);
ierr = PetscBTDestroy(&seenFaces);CHKERRQ(ierr);
ierr = PetscFree2(numNeighbors, neighbors);CHKERRQ(ierr);
ierr = PetscFree2(rorntComp, lorntComp);CHKERRQ(ierr);
ierr = PetscFree3(faceFIFO, cellComp, faceComp);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
DMPlexUninterpolate - Take in a mesh with all intermediate faces, edges, etc. and return a cell-vertex mesh
Collective on DM
Input Parameter:
. dm - The complete DMPlex object
Output Parameter:
. dmUnint - The DMPlex object with only cells and vertices
Level: intermediate
.keywords: mesh
.seealso: DMPlexInterpolate(), DMPlexCreateFromCellList()
@*/
PetscErrorCode DMPlexUninterpolate(DM dm, DM *dmUnint)
{
DM udm;
PetscInt dim, vStart, vEnd, cStart, cEnd, c, maxConeSize = 0, *cone;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
if (dim <= 1) {
ierr = PetscObjectReference((PetscObject) dm);CHKERRQ(ierr);
*dmUnint = dm;
PetscFunctionReturn(0);
}
ierr = DMPlexGetDepthStratum(dm, 0, &vStart, &vEnd);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, 0, &cStart, &cEnd);CHKERRQ(ierr);
ierr = DMCreate(PetscObjectComm((PetscObject) dm), &udm);CHKERRQ(ierr);
ierr = DMSetType(udm, DMPLEX);CHKERRQ(ierr);
ierr = DMSetDimension(udm, dim);CHKERRQ(ierr);
ierr = DMPlexSetChart(udm, cStart, vEnd);CHKERRQ(ierr);
for (c = cStart; c < cEnd; ++c) {
PetscInt *closure = NULL, closureSize, cl, coneSize = 0;
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
for (cl = 0; cl < closureSize*2; cl += 2) {
const PetscInt p = closure[cl];
if ((p >= vStart) && (p < vEnd)) ++coneSize;
}
ierr = DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
ierr = DMPlexSetConeSize(udm, c, coneSize);CHKERRQ(ierr);
maxConeSize = PetscMax(maxConeSize, coneSize);
}
ierr = DMSetUp(udm);CHKERRQ(ierr);
ierr = PetscMalloc1(maxConeSize, &cone);CHKERRQ(ierr);
for (c = cStart; c < cEnd; ++c) {
PetscInt *closure = NULL, closureSize, cl, coneSize = 0;
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
for (cl = 0; cl < closureSize*2; cl += 2) {
const PetscInt p = closure[cl];
if ((p >= vStart) && (p < vEnd)) cone[coneSize++] = p;
}
ierr = DMPlexRestoreTransitiveClosure(dm, c, PETSC_TRUE, &closureSize, &closure);CHKERRQ(ierr);
ierr = DMPlexSetCone(udm, c, cone);CHKERRQ(ierr);
}
ierr = PetscFree(cone);CHKERRQ(ierr);
ierr = DMPlexSymmetrize(udm);CHKERRQ(ierr);
ierr = DMPlexStratify(udm);CHKERRQ(ierr);
/* Reduce SF */
{
PetscSF sfPoint, sfPointUn;
const PetscSFNode *remotePoints;
const PetscInt *localPoints;
PetscSFNode *remotePointsUn;
PetscInt *localPointsUn;
PetscInt vEnd, numRoots, numLeaves, l;
PetscInt numLeavesUn = 0, n = 0;
PetscErrorCode ierr;
/* Get original SF information */
ierr = DMGetPointSF(dm, &sfPoint);CHKERRQ(ierr);
ierr = DMGetPointSF(udm, &sfPointUn);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(dm, 0, NULL, &vEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sfPoint, &numRoots, &numLeaves, &localPoints, &remotePoints);CHKERRQ(ierr);
/* Allocate space for cells and vertices */
for (l = 0; l < numLeaves; ++l) if (localPoints[l] < vEnd) numLeavesUn++;
/* Fill in leaves */
if (vEnd >= 0) {
ierr = PetscMalloc1(numLeavesUn, &remotePointsUn);CHKERRQ(ierr);
ierr = PetscMalloc1(numLeavesUn, &localPointsUn);CHKERRQ(ierr);
for (l = 0; l < numLeaves; l++) {
if (localPoints[l] < vEnd) {
localPointsUn[n] = localPoints[l];
remotePointsUn[n].rank = remotePoints[l].rank;
remotePointsUn[n].index = remotePoints[l].index;
++n;
}
}
if (n != numLeavesUn) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Inconsistent number of leaves %d != %d", n, numLeavesUn);
ierr = PetscSFSetGraph(sfPointUn, vEnd, numLeavesUn, localPointsUn, PETSC_OWN_POINTER, remotePointsUn, PETSC_OWN_POINTER);CHKERRQ(ierr);
}
}
*dmUnint = udm;
PetscFunctionReturn(0);
}
/* This interpolates the PointSF in parallel following local interpolation */
static PetscErrorCode DMPlexInterpolatePointSF(DM dm, PetscSF pointSF, PetscInt depth)
{
PetscMPIInt numProcs, rank;
PetscInt p, c, d, dof, offset;
PetscInt numLeaves, numRoots, candidatesSize, candidatesRemoteSize;
const PetscInt *localPoints;
const PetscSFNode *remotePoints;
PetscSFNode *candidates, *candidatesRemote, *claims;
PetscSection candidateSection, candidateSectionRemote, claimSection;
PetscHashI leafhash;
PetscHashIJ roothash;
PetscHashIJKey key;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(PetscObjectComm((PetscObject) dm), &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject) dm), &rank);CHKERRQ(ierr);
ierr = PetscSFGetGraph(pointSF, &numRoots, &numLeaves, &localPoints, &remotePoints);CHKERRQ(ierr);
if (numProcs < 2 || numRoots < 0) PetscFunctionReturn(0);
/* Build hashes of points in the SF for efficient lookup */
PetscHashICreate(leafhash);
PetscHashIJCreate(&roothash);
ierr = PetscHashIJSetMultivalued(roothash, PETSC_FALSE);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
PetscHashIAdd(leafhash, localPoints[p], p);
key.i = remotePoints[p].index; key.j = remotePoints[p].rank;
PetscHashIJAdd(roothash, key, p);
}
/* Build a section / SFNode array of candidate points in the single-level adjacency of leaves,
where each candidate is defined by the root entry for the other vertex that defines the edge. */
ierr = PetscSectionCreate(PetscObjectComm((PetscObject) dm), &candidateSection);CHKERRQ(ierr);
ierr = PetscSectionSetChart(candidateSection, 0, numRoots);CHKERRQ(ierr);
{
PetscInt leaf, root, idx, a, *adj = NULL;
for (p = 0; p < numLeaves; ++p) {
PetscInt adjSize = PETSC_DETERMINE;
ierr = DMPlexGetAdjacency_Internal(dm, localPoints[p], PETSC_FALSE, PETSC_FALSE, PETSC_FALSE, &adjSize, &adj);CHKERRQ(ierr);
for (a = 0; a < adjSize; ++a) {
PetscHashIMap(leafhash, adj[a], leaf);
if (leaf >= 0) {ierr = PetscSectionAddDof(candidateSection, localPoints[p], 1);CHKERRQ(ierr);}
}
}
ierr = PetscSectionSetUp(candidateSection);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(candidateSection, &candidatesSize);CHKERRQ(ierr);
ierr = PetscMalloc1(candidatesSize, &candidates);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
PetscInt adjSize = PETSC_DETERMINE;
ierr = PetscSectionGetOffset(candidateSection, localPoints[p], &offset);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency_Internal(dm, localPoints[p], PETSC_FALSE, PETSC_FALSE, PETSC_FALSE, &adjSize, &adj);CHKERRQ(ierr);
for (idx = 0, a = 0; a < adjSize; ++a) {
PetscHashIMap(leafhash, adj[a], root);
if (root >= 0) candidates[offset+idx++] = remotePoints[root];
}
}
ierr = PetscFree(adj);CHKERRQ(ierr);
}
/* Gather candidate section / array pair into the root partition via inverse(multi(pointSF)). */
{
PetscSF sfMulti, sfInverse, sfCandidates;
PetscInt *remoteOffsets;
ierr = PetscSFGetMultiSF(pointSF, &sfMulti);CHKERRQ(ierr);
ierr = PetscSFCreateInverseSF(sfMulti, &sfInverse);CHKERRQ(ierr);
ierr = PetscSectionCreate(PetscObjectComm((PetscObject) dm), &candidateSectionRemote);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sfInverse, candidateSection, &remoteOffsets, candidateSectionRemote);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sfInverse, candidateSection, remoteOffsets, candidateSectionRemote, &sfCandidates);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(candidateSectionRemote, &candidatesRemoteSize);CHKERRQ(ierr);
ierr = PetscMalloc1(candidatesRemoteSize, &candidatesRemote);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfCandidates, MPIU_2INT, candidates, candidatesRemote);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfCandidates, MPIU_2INT, candidates, candidatesRemote);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfInverse);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfCandidates);CHKERRQ(ierr);
ierr = PetscFree(remoteOffsets);CHKERRQ(ierr);
}
/* Walk local roots and check for each remote candidate whether we know all required points,
either from owning it or having a root entry in the point SF. If we do we place a claim
by replacing the vertex number with our edge ID. */
{
PetscInt idx, root, joinSize, vertices[2];
const PetscInt *rootdegree, *join = NULL;
ierr = PetscSFComputeDegreeBegin(pointSF, &rootdegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(pointSF, &rootdegree);CHKERRQ(ierr);
/* Loop remote edge connections and put in a claim if both vertices are known */
for (idx = 0, p = 0; p < numRoots; ++p) {
for (d = 0; d < rootdegree[p]; ++d) {
ierr = PetscSectionGetDof(candidateSectionRemote, idx, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(candidateSectionRemote, idx, &offset);CHKERRQ(ierr);
for (c = 0; c < dof; ++c) {
/* We own both vertices, so we claim the edge by replacing vertex with edge */
if (candidatesRemote[offset+c].rank == rank) {
vertices[0] = p; vertices[1] = candidatesRemote[offset+c].index;
ierr = DMPlexGetJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
if (joinSize == 1) candidatesRemote[offset+c].index = join[0];
ierr = DMPlexRestoreJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
continue;
}
/* If we own one vertex and share a root with the other, we claim it */
key.i = candidatesRemote[offset+c].index; key.j = candidatesRemote[offset+c].rank;
PetscHashIJGet(roothash, key, &root);
if (root >= 0) {
vertices[0] = p; vertices[1] = localPoints[root];
ierr = DMPlexGetJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
if (joinSize == 1) {
candidatesRemote[offset+c].index = join[0];
candidatesRemote[offset+c].rank = rank;
}
ierr = DMPlexRestoreJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
}
}
idx++;
}
}
}
/* Push claims back to receiver via the MultiSF and derive new pointSF mapping on receiver */
{
PetscSF sfMulti, sfClaims, sfPointNew;
PetscHashI claimshash;
PetscInt size, pStart, pEnd, root, joinSize, numLocalNew;
PetscInt *remoteOffsets, *localPointsNew, vertices[2];
const PetscInt *join = NULL;
PetscSFNode *remotePointsNew;
ierr = PetscSFGetMultiSF(pointSF, &sfMulti);CHKERRQ(ierr);
ierr = PetscSectionCreate(PetscObjectComm((PetscObject) dm), &claimSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sfMulti, candidateSectionRemote, &remoteOffsets, claimSection);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sfMulti, candidateSectionRemote, remoteOffsets, claimSection, &sfClaims);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(claimSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &claims);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfClaims, MPIU_2INT, candidatesRemote, claims);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfClaims, MPIU_2INT, candidatesRemote, claims);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfClaims);CHKERRQ(ierr);
ierr = PetscFree(remoteOffsets);CHKERRQ(ierr);
/* Walk the original section of local supports and add an SF entry for each updated item */
PetscHashICreate(claimshash);
for (p = 0; p < numRoots; ++p) {
ierr = PetscSectionGetDof(candidateSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(candidateSection, p, &offset);CHKERRQ(ierr);
for (d = 0; d < dof; ++d) {
if (candidates[offset+d].index != claims[offset+d].index) {
key.i = candidates[offset+d].index; key.j = candidates[offset+d].rank;
PetscHashIJGet(roothash, key, &root);
if (root >= 0) {
vertices[0] = p; vertices[1] = localPoints[root];
ierr = DMPlexGetJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
if (joinSize == 1) PetscHashIAdd(claimshash, join[0], offset+d);
ierr = DMPlexRestoreJoin(dm, 2, vertices, &joinSize, &join);CHKERRQ(ierr);
}
}
}
}
/* Create new pointSF from hashed claims */
PetscHashISize(claimshash, numLocalNew);
ierr = DMPlexGetChart(dm, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscMalloc1(numLeaves + numLocalNew, &localPointsNew);CHKERRQ(ierr);
ierr = PetscMalloc1(numLeaves + numLocalNew, &remotePointsNew);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
localPointsNew[p] = localPoints[p];
remotePointsNew[p].index = remotePoints[p].index;
remotePointsNew[p].rank = remotePoints[p].rank;
}
p = numLeaves; ierr = PetscHashIGetKeys(claimshash, &p, localPointsNew);CHKERRQ(ierr);
for (p = numLeaves; p < numLeaves + numLocalNew; ++p) {
PetscHashIMap(claimshash, localPointsNew[p], offset);
remotePointsNew[p] = claims[offset];
}
ierr = PetscSFCreate(PetscObjectComm((PetscObject) dm), &sfPointNew);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfPointNew, pEnd-pStart, numLeaves+numLocalNew, localPointsNew, PETSC_OWN_POINTER, remotePointsNew, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = DMSetPointSF(dm, sfPointNew);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfPointNew);CHKERRQ(ierr);
PetscHashIDestroy(claimshash);
}
PetscHashIDestroy(leafhash);
ierr = PetscHashIJDestroy(&roothash);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&candidateSection);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&candidateSectionRemote);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&claimSection);CHKERRQ(ierr);
ierr = PetscFree(candidates);CHKERRQ(ierr);
ierr = PetscFree(candidatesRemote);CHKERRQ(ierr);
ierr = PetscFree(claims);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
