PetscErrorCode MatPartitioningHierarchical_ReassembleFineparts(Mat adj, IS fineparts, ISLocalToGlobalMapping mapping, IS *sfineparts)
{
PetscInt *local_indices, *global_indices,*owners,*sfineparts_indices,localsize,i;
const PetscInt *ranges,*fineparts_indices;
PetscMPIInt rank;
MPI_Comm comm;
PetscLayout rmap;
PetscSFNode *remote;
PetscSF sf;
PetscErrorCode ierr;
PetscFunctionBegin;
/*get communicator */
ierr = PetscObjectGetComm((PetscObject)adj,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
ierr = MatGetLayouts(adj,&rmap,PETSC_NULL);CHKERRQ(ierr);
ierr = ISGetLocalSize(fineparts,&localsize);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&global_indices,localsize,&local_indices);CHKERRQ(ierr);
for(i=0; i<localsize; i++){
local_indices[i] = i;
}
/*global indices */
ierr = ISLocalToGlobalMappingApply(mapping,localsize,local_indices,global_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(localsize,&owners);CHKERRQ(ierr);
/*find owners for global indices */
for(i=0; i<localsize; i++){
ierr = PetscLayoutFindOwner(rmap,global_indices[i],&owners[i]);CHKERRQ(ierr);
}
/*ranges */
ierr = PetscLayoutGetRanges(rmap,&ranges);CHKERRQ(ierr);
ierr = PetscCalloc1(ranges[rank+1]-ranges[rank],&sfineparts_indices);CHKERRQ(ierr);
ierr = ISGetIndices(fineparts,&fineparts_indices);CHKERRQ(ierr);
/*create a SF to exchange data */
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscCalloc1(localsize,&remote);CHKERRQ(ierr);
for(i=0; i<localsize; i++){
remote[i].rank = owners[i];
remote[i].index = global_indices[i]-ranges[owners[i]];
}
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
/*not sure how to add prefix to sf*/
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,localsize,localsize,PETSC_NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPIU_INT,fineparts_indices,sfineparts_indices,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_INT,fineparts_indices,sfineparts_indices,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = ISRestoreIndices(fineparts,&fineparts_indices);CHKERRQ(ierr);
/* comm self */
ierr = ISCreateGeneral(comm,ranges[rank+1]-ranges[rank],sfineparts_indices,PETSC_OWN_POINTER,sfineparts);CHKERRQ(ierr);
ierr = PetscFree2(global_indices,local_indices);CHKERRQ(ierr);
ierr = PetscFree(owners);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);
}
/*@C
PetscSFGetGroups - gets incoming and outgoing process groups
Collective
Input Argument:
. sf - star forest
Output Arguments:
+ incoming - group of origin processes for incoming edges (leaves that reference my roots)
- outgoing - group of destination processes for outgoing edges (roots that I reference)
Level: developer
.seealso: PetscSFGetWindow(), PetscSFRestoreWindow()
@*/
PetscErrorCode PetscSFGetGroups(PetscSF sf,MPI_Group *incoming,MPI_Group *outgoing)
{
PetscErrorCode ierr;
MPI_Group group;
PetscFunctionBegin;
if (sf->ingroup == MPI_GROUP_NULL) {
PetscInt i;
const PetscInt *indegree;
PetscMPIInt rank,*outranks,*inranks;
PetscSFNode *remote;
PetscSF bgcount;
/* Compute the number of incoming ranks */
ierr = PetscMalloc1(sf->nranks,&remote);CHKERRQ(ierr);
for (i=0; i<sf->nranks; i++) {
remote[i].rank = sf->ranks[i];
remote[i].index = 0;
}
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,&bgcount);CHKERRQ(ierr);
ierr = PetscSFSetGraph(bgcount,1,sf->nranks,NULL,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(bgcount,&indegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(bgcount,&indegree);CHKERRQ(ierr);
/* Enumerate the incoming ranks */
ierr = PetscMalloc2(indegree[0],&inranks,sf->nranks,&outranks);CHKERRQ(ierr);
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
for (i=0; i<sf->nranks; i++) outranks[i] = rank;
ierr = PetscSFGatherBegin(bgcount,MPI_INT,outranks,inranks);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(bgcount,MPI_INT,outranks,inranks);CHKERRQ(ierr);
ierr = MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group);CHKERRQ(ierr);
ierr = MPI_Group_incl(group,indegree[0],inranks,&sf->ingroup);CHKERRQ(ierr);
ierr = MPI_Group_free(&group);CHKERRQ(ierr);
ierr = PetscFree2(inranks,outranks);CHKERRQ(ierr);
ierr = PetscSFDestroy(&bgcount);CHKERRQ(ierr);
}
*incoming = sf->ingroup;
if (sf->outgroup == MPI_GROUP_NULL) {
ierr = MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group);CHKERRQ(ierr);
ierr = MPI_Group_incl(group,sf->nranks,sf->ranks,&sf->outgroup);CHKERRQ(ierr);
ierr = MPI_Group_free(&group);CHKERRQ(ierr);
}
*outgoing = sf->outgroup;
PetscFunctionReturn(0);
}
/* this function maps rows to locally owned rows */
PETSC_INTERN PetscErrorCode MatZeroRowsMapLocal_Private(Mat A,PetscInt N,const PetscInt *rows,PetscInt *nr,PetscInt **olrows)
{
PetscInt *owners = A->rmap->range;
PetscInt n = A->rmap->n;
PetscSF sf;
PetscInt *lrows;
PetscSFNode *rrows;
PetscMPIInt rank;
PetscInt r, p = 0, len = 0;
PetscErrorCode ierr;
PetscFunctionBegin;
/* Create SF where leaves are input rows and roots are owned rows */
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr);
ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr);
for (r = 0; r < n; ++r) lrows[r] = -1;
if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);}
for (r = 0; r < N; ++r) {
const PetscInt idx = rows[r];
if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr);
}
if (A->nooffproczerorows) {
if (p != rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,rank);
lrows[len++] = idx - owners[p];
} else {
rrows[r].rank = p;
rrows[r].index = rows[r] - owners[p];
}
}
if (!A->nooffproczerorows) {
ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr);
/* Collect flags for rows to be zeroed */
ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Compress and put in row numbers */
for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
}
if (nr) *nr = len;
if (olrows) *olrows = lrows;
PetscFunctionReturn(0);
}
/*@C
PetscSFSetGraphLayout - Set a parallel star forest via global indices and a PetscLayout
Collective
Input Arguments:
+ sf - star forest
. layout - PetscLayout defining the global space
. nleaves - number of leaf vertices on the current process, each of these references a root on any process
. ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage
- iremote - remote locations of root vertices for each leaf on the current process
Level: intermediate
.seealso: PetscSFCreate(), PetscSFView(), PetscSFSetGraph(), PetscSFGetGraph()
@*/
PetscErrorCode PetscSFSetGraphLayout(PetscSF sf,PetscLayout layout,PetscInt nleaves,const PetscInt *ilocal,PetscCopyMode localmode,const PetscInt *iremote)
{
PetscErrorCode ierr;
PetscInt i,nroots;
PetscSFNode *remote;
PetscFunctionBegin;
ierr = PetscLayoutGetLocalSize(layout,&nroots);CHKERRQ(ierr);
ierr = PetscMalloc(nleaves*sizeof(PetscSFNode),&remote);CHKERRQ(ierr);
for (i=0; i<nleaves; i++) {
PetscInt owner = -1;
ierr = PetscLayoutFindOwner(layout,iremote[i],&owner);CHKERRQ(ierr);
remote[i].rank = owner;
remote[i].index = iremote[i] - layout->range[owner];
}
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,localmode,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
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);
}
PetscErrorCode DMMeshConvertOverlapToSF(DM dm, PetscSF *sf)
{
ALE::Obj<PETSC_MESH_TYPE> mesh;
PetscInt *local;
PetscSFNode *remote;
PetscInt numPoints;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFCreate(((PetscObject) dm)->comm, sf);
CHKERRQ(ierr);
ierr = DMMeshGetMesh(dm, mesh);
CHKERRQ(ierr);
{
/* The local points have degree 1
We use the recv overlap
*/
ALE::Obj<PETSC_MESH_TYPE::recv_overlap_type> overlap = mesh->getRecvOverlap();
numPoints = overlap->getNumPoints();
ierr = PetscMalloc(numPoints * sizeof(PetscInt), &local);
CHKERRQ(ierr);
ierr = PetscMalloc(numPoints * sizeof(PetscSFNode), &remote);
CHKERRQ(ierr);
for (PetscInt r = 0, i = 0; r < overlap->getNumRanks(); ++r) {
const PetscInt rank = overlap->getRank(r);
const PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator cBegin = overlap->supportBegin(rank);
const PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator cEnd = overlap->supportEnd(rank);
for (PETSC_MESH_TYPE::recv_overlap_type::supportSequence::iterator c_iter = cBegin; c_iter != cEnd; ++c_iter, ++i) {
local[i] = *c_iter;
remote[i].rank = rank;
remote[i].index = c_iter.color();
}
}
ierr = PetscSFSetGraph(*sf, numPoints, numPoints, local, PETSC_OWN_POINTER, remote, PETSC_OWN_POINTER);
CHKERRQ(ierr);
ierr = PetscSFView(*sf, NULL);
CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
/*@
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);
}
/*@C
PetscSFCreateEmbeddedSF - removes edges from all but the selected roots, does not remap indices
Collective
Input Arguments:
+ sf - original star forest
. nroots - number of roots to select on this process
- selected - selected roots on this process
Output Arguments:
. newsf - new star forest
Level: advanced
Note:
To use the new PetscSF, it may be necessary to know the indices of the leaves that are still participating. This can
be done by calling PetscSFGetGraph().
.seealso: PetscSFSetGraph(), PetscSFGetGraph()
@*/
PetscErrorCode PetscSFCreateEmbeddedSF(PetscSF sf,PetscInt nroots,const PetscInt *selected,PetscSF *newsf)
{
PetscInt *rootdata, *leafdata, *ilocal;
PetscSFNode *iremote;
PetscInt leafsize = 0, nleaves = 0, n, i;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
if (nroots) PetscValidPointer(selected,3);
PetscValidPointer(newsf,4);
if (sf->mine) for (i = 0; i < sf->nleaves; ++i) {leafsize = PetscMax(leafsize, sf->mine[i]+1);}
else leafsize = sf->nleaves;
ierr = PetscCalloc2(sf->nroots,&rootdata,leafsize,&leafdata);CHKERRQ(ierr);
for (i=0; i<nroots; ++i) rootdata[selected[i]] = 1;
ierr = PetscSFBcastBegin(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
for (i = 0; i < leafsize; ++i) nleaves += leafdata[i];
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
for (i = 0, n = 0; i < sf->nleaves; ++i) {
const PetscInt lidx = sf->mine ? sf->mine[i] : i;
if (leafdata[lidx]) {
ilocal[n] = lidx;
iremote[n].rank = sf->remote[i].rank;
iremote[n].index = sf->remote[i].index;
++n;
}
}
if (n != nleaves) SETERRQ2(PETSC_COMM_SELF, PETSC_ERR_PLIB, "There is a size mismatch in the SF embedding, %d != %d", n, nleaves);
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,newsf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(*newsf,sf->nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscFree2(rootdata,leafdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves, does not remap indices
Collective
Input Arguments:
+ sf - original star forest
. nleaves - number of leaves to select on this process
- selected - selected leaves on this process
Output Arguments:
. newsf - new star forest
Level: advanced
.seealso: PetscSFCreateEmbeddedSF(), PetscSFSetGraph(), PetscSFGetGraph()
@*/
PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf, PetscInt nleaves, const PetscInt *selected, PetscSF *newsf)
{
PetscSFNode *iremote;
PetscInt *ilocal;
PetscInt i;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
if (nleaves) PetscValidPointer(selected, 3);
PetscValidPointer(newsf, 4);
ierr = PetscMalloc1(nleaves, &ilocal);CHKERRQ(ierr);
ierr = PetscMalloc1(nleaves, &iremote);CHKERRQ(ierr);
for (i = 0; i < nleaves; ++i) {
const PetscInt l = selected[i];
ilocal[i] = sf->mine ? sf->mine[l] : l;
iremote[i].rank = sf->remote[l].rank;
iremote[i].index = sf->remote[l].index;
}
ierr = PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, newsf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(*newsf, sf->nroots, nleaves, ilocal, PETSC_OWN_POINTER, iremote, PETSC_OWN_POINTER);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);
}
/* 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);
}
/*
* Increase overlap for the sub-matrix across sub communicator
* sub-matrix could be a graph or numerical matrix
* */
PetscErrorCode MatIncreaseOverlapSplit_Single(Mat mat,IS *is,PetscInt ov)
{
PetscInt i,nindx,*indices_sc,*indices_ov,localsize,*localsizes_sc,localsize_tmp;
PetscInt *indices_ov_rd,nroots,nleaves,*localoffsets,*indices_recv,*sources_sc,*sources_sc_rd;
const PetscInt *indices;
PetscMPIInt srank,ssize,issamecomm,k,grank;
IS is_sc,allis_sc,partitioning;
MPI_Comm gcomm,dcomm,scomm;
PetscSF sf;
PetscSFNode *remote;
Mat *smat;
MatPartitioning part;
PetscErrorCode ierr;
PetscFunctionBegin;
/* get a sub communicator before call individual MatIncreaseOverlap
* since the sub communicator may be changed.
* */
ierr = PetscObjectGetComm((PetscObject)(*is),&dcomm);CHKERRQ(ierr);
/*make a copy before the original one is deleted*/
ierr = PetscCommDuplicate(dcomm,&scomm,NULL);CHKERRQ(ierr);
/*get a global communicator, where mat should be a global matrix */
ierr = PetscObjectGetComm((PetscObject)mat,&gcomm);CHKERRQ(ierr);
/*increase overlap on each individual subdomain*/
ierr = (*mat->ops->increaseoverlap)(mat,1,is,ov);CHKERRQ(ierr);
/*compare communicators */
ierr = MPI_Comm_compare(gcomm,scomm,&issamecomm);CHKERRQ(ierr);
/* if the sub-communicator is the same as the global communicator,
* user does not want to use a sub-communicator
* */
if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT) PetscFunctionReturn(0);
/* if the sub-communicator is petsc_comm_self,
* user also does not care the sub-communicator
* */
ierr = MPI_Comm_compare(scomm,PETSC_COMM_SELF,&issamecomm);CHKERRQ(ierr);
if(issamecomm == MPI_IDENT || issamecomm == MPI_CONGRUENT){PetscFunctionReturn(0);}
/*local rank, size in a sub-communicator */
ierr = MPI_Comm_rank(scomm,&srank);CHKERRQ(ierr);
ierr = MPI_Comm_size(scomm,&ssize);CHKERRQ(ierr);
ierr = MPI_Comm_rank(gcomm,&grank);CHKERRQ(ierr);
/*create a new IS based on sub-communicator
* since the old IS is often based on petsc_comm_self
* */
ierr = ISGetLocalSize(*is,&nindx);CHKERRQ(ierr);
ierr = PetscCalloc1(nindx,&indices_sc);CHKERRQ(ierr);
ierr = ISGetIndices(*is,&indices);CHKERRQ(ierr);
ierr = PetscMemcpy(indices_sc,indices,sizeof(PetscInt)*nindx);CHKERRQ(ierr);
ierr = ISRestoreIndices(*is,&indices);CHKERRQ(ierr);
/*we do not need any more*/
ierr = ISDestroy(is);CHKERRQ(ierr);
/*create a index set based on the sub communicator */
ierr = ISCreateGeneral(scomm,nindx,indices_sc,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr);
/*gather all indices within the sub communicator*/
ierr = ISAllGather(is_sc,&allis_sc);CHKERRQ(ierr);
ierr = ISDestroy(&is_sc);CHKERRQ(ierr);
/* gather local sizes */
ierr = PetscMalloc1(ssize,&localsizes_sc);CHKERRQ(ierr);
/*get individual local sizes for all index sets*/
ierr = MPI_Gather(&nindx,1,MPIU_INT,localsizes_sc,1,MPIU_INT,0,scomm);CHKERRQ(ierr);
/*only root does these computations */
if(!srank){
/*get local size for the big index set*/
ierr = ISGetLocalSize(allis_sc,&localsize);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&indices_ov,localsize,&sources_sc);CHKERRQ(ierr);
ierr = PetscCalloc2(localsize,&indices_ov_rd,localsize,&sources_sc_rd);CHKERRQ(ierr);
ierr = ISGetIndices(allis_sc,&indices);CHKERRQ(ierr);
ierr = PetscMemcpy(indices_ov,indices,sizeof(PetscInt)*localsize);CHKERRQ(ierr);
ierr = ISRestoreIndices(allis_sc,&indices);CHKERRQ(ierr);
/*we do not need it any more */
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*assign corresponding sources */
localsize_tmp = 0;
for(k=0; k<ssize; k++){
for(i=0; i<localsizes_sc[k]; i++){
sources_sc[localsize_tmp++] = k;
}
}
/*record where indices come from */
ierr = PetscSortIntWithArray(localsize,indices_ov,sources_sc);CHKERRQ(ierr);
/*count local sizes for reduced indices */
ierr = PetscMemzero(localsizes_sc,sizeof(PetscInt)*ssize);CHKERRQ(ierr);
/*initialize the first entity*/
if(localsize){
indices_ov_rd[0] = indices_ov[0];
sources_sc_rd[0] = sources_sc[0];
localsizes_sc[sources_sc[0]]++;
}
localsize_tmp = 1;
/*remove duplicate integers */
for(i=1; i<localsize; i++){
if(indices_ov[i] != indices_ov[i-1]){
indices_ov_rd[localsize_tmp] = indices_ov[i];
sources_sc_rd[localsize_tmp++] = sources_sc[i];
localsizes_sc[sources_sc[i]]++;
}
}
ierr = PetscFree2(indices_ov,sources_sc);CHKERRQ(ierr);
ierr = PetscCalloc1(ssize+1,&localoffsets);CHKERRQ(ierr);
for(k=0; k<ssize; k++){
localoffsets[k+1] = localoffsets[k] + localsizes_sc[k];
}
/*construct a star forest to send data back */
nleaves = localoffsets[ssize];
ierr = PetscMemzero(localoffsets,(ssize+1)*sizeof(PetscInt));CHKERRQ(ierr);
nroots = localsizes_sc[srank];
ierr = PetscCalloc1(nleaves,&remote);CHKERRQ(ierr);
for(i=0; i<nleaves; i++){
remote[i].rank = sources_sc_rd[i];
remote[i].index = localoffsets[sources_sc_rd[i]]++;
}
ierr = PetscFree(localoffsets);CHKERRQ(ierr);
}else{
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*Allocate a 'zero' pointer */
ierr = PetscCalloc1(0,&remote);CHKERRQ(ierr);
nleaves = 0;
indices_ov_rd = 0;
sources_sc_rd = 0;
}
/*scatter sizes to everybody */
ierr = MPI_Scatter(localsizes_sc,1, MPIU_INT,&nroots,1, MPIU_INT,0,scomm);CHKERRQ(ierr);
/*free memory */
ierr = PetscFree(localsizes_sc);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots,&indices_recv);CHKERRQ(ierr);
/*ierr = MPI_Comm_dup(scomm,&dcomm);CHKERRQ(ierr);*/
/*set data back to every body */
ierr = PetscSFCreate(scomm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_INT,indices_ov_rd,indices_recv,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* free memory */
ierr = PetscFree2(indices_ov_rd,sources_sc_rd);CHKERRQ(ierr);
/*create a index set*/
ierr = ISCreateGeneral(scomm,nroots,indices_recv,PETSC_OWN_POINTER,&is_sc);CHKERRQ(ierr);
/*construct a parallel submatrix */
ierr = MatGetSubMatricesMPI(mat,1,&is_sc,&is_sc,MAT_INITIAL_MATRIX,&smat);CHKERRQ(ierr);
/* we do not need them any more */
ierr = ISDestroy(&allis_sc);CHKERRQ(ierr);
/*create a partitioner to repartition the sub-matrix*/
ierr = MatPartitioningCreate(scomm,&part);CHKERRQ(ierr);
ierr = MatPartitioningSetAdjacency(part,smat[0]);CHKERRQ(ierr);
#if PETSC_HAVE_PARMETIS
/* if there exists a ParMETIS installation, we try to use ParMETIS
* because a repartition routine possibly work better
* */
ierr = MatPartitioningSetType(part,MATPARTITIONINGPARMETIS);CHKERRQ(ierr);
/*try to use reparition function, instead of partition function */
ierr = MatPartitioningParmetisSetRepartition(part);CHKERRQ(ierr);
#else
/*we at least provide a default partitioner to rebalance the computation */
ierr = MatPartitioningSetType(part,MATPARTITIONINGAVERAGE);CHKERRQ(ierr);
#endif
/*user can pick up any partitioner by using an option*/
ierr = MatPartitioningSetFromOptions(part);CHKERRQ(ierr);
/* apply partition */
ierr = MatPartitioningApply(part,&partitioning);CHKERRQ(ierr);
ierr = MatPartitioningDestroy(&part);CHKERRQ(ierr);
ierr = MatDestroy(&(smat[0]));CHKERRQ(ierr);
ierr = PetscFree(smat);CHKERRQ(ierr);
/* get local rows including overlap */
ierr = ISBuildTwoSided(partitioning,is_sc,is);CHKERRQ(ierr);
/* destroy */
ierr = ISDestroy(&is_sc);CHKERRQ(ierr);
ierr = ISDestroy(&partitioning);CHKERRQ(ierr);
ierr = PetscCommDestroy(&scomm);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
Vec B,Bout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscScalar *bufB;
PetscScalar *bufBout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 2) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for two MPI processes\n");
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 2;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
if (rank == 0) {
iremote[0].rank = 0;
iremote[0].index = 0;
iremote[1].rank = 1;
iremote[1].index = 0;
} else {
iremote[0].rank = 1;
iremote[0].index = 0;
iremote[1].rank = 0;
iremote[1].index = 0;
}
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,2,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&B);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Bout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArray(B,&bufB);CHKERRQ(ierr);
for (i=0; i<2; i++) {
bufA[i] = (PetscScalar)rank;
bufB[i] = (PetscScalar)(rank) + 10.0;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(B,&bufB);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecGetArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecRestoreArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecView(Bout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&B);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = VecDestroy(&Bout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
/*@C
DMPlexDistribute - Distributes the mesh and any associated sections.
Not Collective
Input Parameter:
+ dm - The original DMPlex object
. partitioner - The partitioning package, or NULL for the default
- overlap - The overlap of partitions, 0 is the default
Output Parameter:
+ sf - The PetscSF used for point distribution
- parallelMesh - The distributed DMPlex object, or NULL
Note: If the mesh was not distributed, the return value is NULL.
The user can control the definition of adjacency for the mesh using DMPlexGetAdjacencyUseCone() and
DMPlexSetAdjacencyUseClosure(). They should choose the combination appropriate for the function
representation on the mesh.
Level: intermediate
.keywords: mesh, elements
.seealso: DMPlexCreate(), DMPlexDistributeByFace(), DMPlexSetAdjacencyUseCone(), DMPlexSetAdjacencyUseClosure()
@*/
PetscErrorCode DMPlexDistribute(DM dm, const char partitioner[], PetscInt overlap, PetscSF *sf, DM *dmParallel)
{
DM_Plex *mesh = (DM_Plex*) dm->data, *pmesh;
MPI_Comm comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS origCellPart, origPart, cellPart, part;
PetscSection origCellPartSection, origPartSection, cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF, pointSF, coneSF;
ISLocalToGlobalMapping renumbering;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets;
PetscInt *cones, *newCones, newConesSize;
PetscBool flg;
PetscMPIInt rank, numProcs, p;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
if (sf) PetscValidPointer(sf,4);
PetscValidPointer(dmParallel,5);
ierr = PetscLogEventBegin(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
*dmParallel = NULL;
if (numProcs == 1) PetscFunctionReturn(0);
ierr = DMPlexGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = PetscLogEventBegin(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
if (overlap > 1) SETERRQ(PetscObjectComm((PetscObject)dm), PETSC_ERR_SUP, "Overlap > 1 not yet implemented");
ierr = DMPlexCreatePartition(dm, partitioner, height, overlap > 0 ? PETSC_TRUE : PETSC_FALSE, &cellPartSection, &cellPart, &origCellPartSection, &origCellPart);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) numRemoteRanks = numProcs;
else numRemoteRanks = 0;
ierr = PetscMalloc1(numRemoteRanks, &remoteRanks);CHKERRQ(ierr);
for (p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-partition_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, NULL);CHKERRQ(ierr);
if (origCellPart) {
ierr = PetscPrintf(comm, "Original Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(origCellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(origCellPart, NULL);CHKERRQ(ierr);
}
ierr = PetscSFView(partSF, NULL);CHKERRQ(ierr);
}
/* Close the partition over the mesh */
ierr = DMPlexCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMPlexCreate(comm, dmParallel);CHKERRQ(ierr);
ierr = DMPlexSetDimension(*dmParallel, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *dmParallel, "Parallel Mesh");CHKERRQ(ierr);
pmesh = (DM_Plex*) (*dmParallel)->data;
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, &pointSF);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, NULL);CHKERRQ(ierr);
ierr = PetscSFView(pointSF, NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, NULL);CHKERRQ(ierr);
}
ierr = PetscLogEventEnd(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscLogEventBegin(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMPlexGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMPlexGetConeSection(*dmParallel, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMSetUp(*dmParallel);CHKERRQ(ierr);
{
PetscInt pStart, pEnd, p;
ierr = PetscSectionGetChart(newConeSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt coneSize;
ierr = PetscSectionGetDof(newConeSection, p, &coneSize);CHKERRQ(ierr);
pmesh->maxConeSize = PetscMax(pmesh->maxConeSize, coneSize);
}
}
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMPlexGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetCones(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyBlock(renumbering, IS_GTOLM_MASK, newConesSize, newCones, NULL, newCones);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-cones_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, NULL);CHKERRQ(ierr);
}
ierr = DMPlexGetConeOrientations(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientations(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Create supports and stratify sieve */
{
PetscInt pStart, pEnd;
ierr = PetscSectionGetChart(pmesh->coneSection, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionSetChart(pmesh->supportSection, pStart, pEnd);CHKERRQ(ierr);
}
ierr = DMPlexSymmetrize(*dmParallel);CHKERRQ(ierr);
ierr = DMPlexStratify(*dmParallel);CHKERRQ(ierr);
/* Distribute Coordinates */
{
PetscSection originalCoordSection, newCoordSection;
Vec originalCoordinates, newCoordinates;
PetscInt bs;
const char *name;
ierr = DMGetCoordinateSection(dm, &originalCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinateSection(*dmParallel, &newCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinatesLocal(dm, &originalCoordinates);CHKERRQ(ierr);
ierr = VecCreate(comm, &newCoordinates);CHKERRQ(ierr);
ierr = PetscObjectGetName((PetscObject) originalCoordinates, &name);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) newCoordinates, name);CHKERRQ(ierr);
ierr = DMPlexDistributeField(dm, pointSF, originalCoordSection, originalCoordinates, newCoordSection, newCoordinates);CHKERRQ(ierr);
ierr = DMSetCoordinatesLocal(*dmParallel, newCoordinates);CHKERRQ(ierr);
ierr = VecGetBlockSize(originalCoordinates, &bs);CHKERRQ(ierr);
ierr = VecSetBlockSize(newCoordinates, bs);CHKERRQ(ierr);
ierr = VecDestroy(&newCoordinates);CHKERRQ(ierr);
}
/* Distribute labels */
ierr = PetscLogEventBegin(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
{
DMLabel next = mesh->labels, newNext = pmesh->labels;
PetscInt numLabels = 0, l;
/* Bcast number of labels */
while (next) {++numLabels; next = next->next;}
ierr = MPI_Bcast(&numLabels, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
next = mesh->labels;
for (l = 0; l < numLabels; ++l) {
DMLabel labelNew;
PetscBool isdepth;
/* Skip "depth" because it is recreated */
if (!rank) {ierr = PetscStrcmp(next->name, "depth", &isdepth);CHKERRQ(ierr);}
ierr = MPI_Bcast(&isdepth, 1, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
if (isdepth) {if (!rank) next = next->next; continue;}
ierr = DMLabelDistribute(next, partSection, part, renumbering, &labelNew);CHKERRQ(ierr);
/* Insert into list */
if (newNext) newNext->next = labelNew;
else pmesh->labels = labelNew;
newNext = labelNew;
if (!rank) next = next->next;
}
}
ierr = PetscLogEventEnd(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
/* Setup hybrid structure */
{
const PetscInt *gpoints;
PetscInt depth, n, d;
for (d = 0; d <= dim; ++d) {pmesh->hybridPointMax[d] = mesh->hybridPointMax[d];}
ierr = MPI_Bcast(pmesh->hybridPointMax, dim+1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(renumbering, &n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(renumbering, &gpoints);CHKERRQ(ierr);
ierr = DMPlexGetDepth(dm, &depth);CHKERRQ(ierr);
for (d = 0; d <= dim; ++d) {
PetscInt pmax = pmesh->hybridPointMax[d], newmax = 0, pEnd, stratum[2], p;
if (pmax < 0) continue;
ierr = DMPlexGetDepthStratum(dm, d > depth ? depth : d, &stratum[0], &stratum[1]);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(*dmParallel, d, NULL, &pEnd);CHKERRQ(ierr);
ierr = MPI_Bcast(stratum, 2, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < n; ++p) {
const PetscInt point = gpoints[p];
if ((point >= stratum[0]) && (point < stratum[1]) && (point >= pmax)) ++newmax;
}
if (newmax > 0) pmesh->hybridPointMax[d] = pEnd - newmax;
else pmesh->hybridPointMax[d] = -1;
}
ierr = ISLocalToGlobalMappingRestoreIndices(renumbering, &gpoints);CHKERRQ(ierr);
}
/* Cleanup Partition */
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Create point SF for parallel mesh */
ierr = PetscLogEventBegin(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
{
const PetscInt *leaves;
PetscSFNode *remotePoints, *rowners, *lowners;
PetscInt numRoots, numLeaves, numGhostPoints = 0, p, gp, *ghostPoints;
PetscInt pStart, pEnd;
ierr = DMPlexGetChart(*dmParallel, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(pointSF, &numRoots, &numLeaves, &leaves, NULL);CHKERRQ(ierr);
ierr = PetscMalloc2(numRoots,&rowners,numLeaves,&lowners);CHKERRQ(ierr);
for (p=0; p<numRoots; p++) {
rowners[p].rank = -1;
rowners[p].index = -1;
}
if (origCellPart) {
/* Make sure points in the original partition are not assigned to other procs */
const PetscInt *origPoints;
ierr = DMPlexCreatePartitionClosure(dm, origCellPartSection, origCellPart, &origPartSection, &origPart);CHKERRQ(ierr);
ierr = ISGetIndices(origPart, &origPoints);CHKERRQ(ierr);
for (p = 0; p < numProcs; ++p) {
PetscInt dof, off, d;
ierr = PetscSectionGetDof(origPartSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(origPartSection, p, &off);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
rowners[origPoints[d]].rank = p;
}
}
ierr = ISRestoreIndices(origPart, &origPoints);CHKERRQ(ierr);
ierr = ISDestroy(&origPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origPartSection);CHKERRQ(ierr);
}
ierr = ISDestroy(&origCellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origCellPartSection);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].rank == rank) { /* Either put in a bid or we know we own it */
lowners[p].rank = rank;
lowners[p].index = leaves ? leaves[p] : p;
} else if (lowners[p].rank >= 0) { /* Point already claimed so flag so that MAXLOC does not listen to us */
lowners[p].rank = -2;
lowners[p].index = -2;
}
}
for (p=0; p<numRoots; p++) { /* Root must not participate in the rediction, flag so that MAXLOC does not use */
rowners[p].rank = -3;
rowners[p].index = -3;
}
ierr = PetscSFReduceBegin(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].index < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cell partition corrupt: point not claimed");
if (lowners[p].rank != rank) ++numGhostPoints;
}
ierr = PetscMalloc1(numGhostPoints, &ghostPoints);CHKERRQ(ierr);
ierr = PetscMalloc1(numGhostPoints, &remotePoints);CHKERRQ(ierr);
for (p = 0, gp = 0; p < numLeaves; ++p) {
if (lowners[p].rank != rank) {
ghostPoints[gp] = leaves ? leaves[p] : p;
remotePoints[gp].rank = lowners[p].rank;
remotePoints[gp].index = lowners[p].index;
++gp;
}
}
ierr = PetscFree2(rowners,lowners);CHKERRQ(ierr);
ierr = PetscSFSetGraph((*dmParallel)->sf, pEnd - pStart, numGhostPoints, ghostPoints, PETSC_OWN_POINTER, remotePoints, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions((*dmParallel)->sf);CHKERRQ(ierr);
}
pmesh->useCone = mesh->useCone;
pmesh->useClosure = mesh->useClosure;
ierr = PetscLogEventEnd(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
/* Copy BC */
ierr = DMPlexCopyBoundary(dm, *dmParallel);CHKERRQ(ierr);
/* Cleanup */
if (sf) {*sf = pointSF;}
else {ierr = PetscSFDestroy(&pointSF);CHKERRQ(ierr);}
ierr = DMSetFromOptions(*dmParallel);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc,char **argv)
{
PetscSF sf,sfDup,sfInv,sfEmbed,sfA,sfB,sfBA;
const PetscInt *degree;
PetscErrorCode ierr;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = CheckRanksEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test setup then reset */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckRanksNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test view (no graph set, no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then view (no type set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then view (no graph set) */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set type then graph then view */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFView(sf,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test set graph then type */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sf);CHKERRQ(ierr);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
ierr = CheckGraphNotSet(sf);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Bcast */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFBcastEnd (sf,MPI_INT,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test Reduce */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPI_INT,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test FetchAndOp */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpBegin(sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpEnd (sf,MPI_INT,NULL,NULL,NULL,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test ComputeDegree */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_COPY_VALUES,NULL,PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeBegin(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf,°ree);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFDuplicate() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_GRAPH,&sfDup);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfDup);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateInverseSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateInverseSF(sf,&sfInv);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfInv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCreateEmbeddedLeafSF() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreateEmbeddedLeafSF(sf,0,NULL,&sfEmbed);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfEmbed);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
/* Test PetscSFCompose() */
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfA);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfA,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sfB);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sfB,0,0,NULL,PETSC_USE_POINTER,NULL,PETSC_USE_POINTER);CHKERRQ(ierr);
ierr = PetscSFCompose(sfA,sfB,&sfBA);CHKERRQ(ierr);
ierr = CheckGraphEmpty(sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfBA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfA);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfB);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
PetscBool test_dupped_type;
MPI_Datatype contig;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for one MPI process");
ierr = PetscOptionsBegin(PETSC_COMM_WORLD,"","PetscSF type freeing options","none");CHKERRQ(ierr);
test_dupped_type = PETSC_FALSE;
ierr = PetscOptionsBool("-test_dupped_type", "Test dupped input type","",test_dupped_type,&test_dupped_type,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 1;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
iremote[0].rank = 0;
iremote[0].index = 0;
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,4,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
for (i=0; i<4; i++) {
bufA[i] = (PetscScalar)i;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = MPI_Type_contiguous(4, MPIU_SCALAR, &contig);CHKERRQ(ierr);
ierr = MPI_Type_commit(&contig);CHKERRQ(ierr);
if (test_dupped_type) {
MPI_Datatype tmp;
ierr = MPI_Type_dup(contig, &tmp);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
contig = tmp;
}
for (i=0;i<10000;i++) {
ierr = PetscSFBcastBegin(sf,contig,bufA,bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,contig,bufA,bufAout);CHKERRQ(ierr);
}
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
int
main (int argc, char **argv)
{
sc_MPI_Comm mpicomm;
int mpiret;
int mpisize, mpirank;
p4est_t *p4est;
p4est_connectivity_t *conn;
sc_array_t *points_per_dim, *cone_sizes, *cones,
*cone_orientations, *coords,
*children, *parents, *childids, *leaves, *remotes;
p4est_locidx_t first_local_quad = -1;
/* initialize MPI */
mpiret = sc_MPI_Init (&argc, &argv);
SC_CHECK_MPI (mpiret);
mpicomm = sc_MPI_COMM_WORLD;
mpiret = sc_MPI_Comm_size (mpicomm, &mpisize);
SC_CHECK_MPI (mpiret);
mpiret = sc_MPI_Comm_rank (mpicomm, &mpirank);
SC_CHECK_MPI (mpiret);
sc_init (mpicomm, 1, 1, NULL, SC_LP_DEFAULT);
p4est_init (NULL, SC_LP_DEFAULT);
#ifndef P4_TO_P8
conn = p4est_connectivity_new_moebius ();
#else
conn = p8est_connectivity_new_rotcubes ();
#endif
p4est = p4est_new_ext (mpicomm, conn, 0, 1, 1, 0, NULL, NULL);
p4est_refine (p4est, 1, refine_fn, NULL);
p4est_balance (p4est, P4EST_CONNECT_FULL, NULL);
p4est_partition (p4est, 0, NULL);
points_per_dim = sc_array_new (sizeof (p4est_locidx_t));
cone_sizes = sc_array_new (sizeof (p4est_locidx_t));
cones = sc_array_new (sizeof (p4est_locidx_t));
cone_orientations = sc_array_new (sizeof (p4est_locidx_t));
coords = sc_array_new (3 * sizeof (double));
children = sc_array_new (sizeof (p4est_locidx_t));
parents = sc_array_new (sizeof (p4est_locidx_t));
childids = sc_array_new (sizeof (p4est_locidx_t));
leaves = sc_array_new (sizeof (p4est_locidx_t));
remotes = sc_array_new (2 * sizeof (p4est_locidx_t));
p4est_get_plex_data (p4est, P4EST_CONNECT_FULL, (mpisize > 1) ? 2 : 0,
&first_local_quad, points_per_dim, cone_sizes, cones,
cone_orientations, coords, children, parents, childids,
leaves, remotes);
#ifdef P4EST_WITH_PETSC
{
PetscErrorCode ierr;
DM plex, refTree;
PetscInt pStart, pEnd;
PetscSection parentSection;
PetscSF pointSF;
size_t zz, count;
locidx_to_PetscInt (points_per_dim);
locidx_to_PetscInt (cone_sizes);
locidx_to_PetscInt (cones);
locidx_to_PetscInt (cone_orientations);
coords_double_to_PetscScalar (coords);
locidx_to_PetscInt (children);
locidx_to_PetscInt (parents);
locidx_to_PetscInt (childids);
locidx_to_PetscInt (leaves);
locidx_pair_to_PetscSFNode (remotes);
P4EST_GLOBAL_PRODUCTION ("Begin PETSc routines\n");
ierr = PetscInitialize (&argc, &argv, 0, help);
CHKERRQ (ierr);
ierr = DMPlexCreate (mpicomm, &plex);
CHKERRQ (ierr);
ierr = DMSetDimension (plex, P4EST_DIM);
CHKERRQ (ierr);
ierr = DMSetCoordinateDim (plex, 3);
CHKERRQ (ierr);
ierr = DMPlexCreateFromDAG (plex, P4EST_DIM,
(PetscInt *) points_per_dim->array,
(PetscInt *) cone_sizes->array,
(PetscInt *) cones->array,
(PetscInt *) cone_orientations->array,
(PetscScalar *) coords->array);
CHKERRQ (ierr);
ierr = PetscSFCreate (mpicomm, &pointSF);
CHKERRQ (ierr);
ierr =
DMPlexCreateDefaultReferenceTree (mpicomm, P4EST_DIM, PETSC_FALSE,
&refTree);
CHKERRQ (ierr);
ierr = DMPlexSetReferenceTree (plex, refTree);
CHKERRQ (ierr);
ierr = DMDestroy (&refTree);
CHKERRQ (ierr);
ierr = PetscSectionCreate (mpicomm, &parentSection);
CHKERRQ (ierr);
ierr = DMPlexGetChart (plex, &pStart, &pEnd);
CHKERRQ (ierr);
ierr = PetscSectionSetChart (parentSection, pStart, pEnd);
CHKERRQ (ierr);
count = children->elem_count;
for (zz = 0; zz < count; zz++) {
PetscInt child =
*((PetscInt *) sc_array_index (children, zz));
ierr = PetscSectionSetDof (parentSection, child, 1);
CHKERRQ (ierr);
}
ierr = PetscSectionSetUp (parentSection);
CHKERRQ (ierr);
ierr =
DMPlexSetTree (plex, parentSection, (PetscInt *) parents->array,
(PetscInt *) childids->array);
CHKERRQ (ierr);
ierr = PetscSectionDestroy (&parentSection);
CHKERRQ (ierr);
ierr =
PetscSFSetGraph (pointSF, pEnd - pStart, (PetscInt) leaves->elem_count,
(PetscInt *) leaves->array, PETSC_COPY_VALUES,
(PetscSFNode *) remotes->array, PETSC_COPY_VALUES);
CHKERRQ (ierr);
ierr = DMViewFromOptions (plex, NULL, "-dm_view");
CHKERRQ (ierr);
/* TODO: test with rigid body modes as in plex ex3 */
ierr = DMDestroy (&plex);
CHKERRQ (ierr);
ierr = PetscFinalize ();
P4EST_GLOBAL_PRODUCTION ("End PETSc routines\n");
}
#endif
sc_array_destroy (points_per_dim);
sc_array_destroy (cone_sizes);
sc_array_destroy (cones);
sc_array_destroy (cone_orientations);
sc_array_destroy (coords);
sc_array_destroy (children);
sc_array_destroy (parents);
sc_array_destroy (childids);
sc_array_destroy (leaves);
sc_array_destroy (remotes);
p4est_destroy (p4est);
p4est_connectivity_destroy (conn);
sc_finalize ();
mpiret = sc_MPI_Finalize ();
SC_CHECK_MPI (mpiret);
return 0;
}
/*
DMPatchZoom - Create a version of the coarse patch (identified by rank) with halo on communicator commz
Collective on DM
Input Parameters:
+ dm - the DM
. rank - the rank which holds the given patch
- commz - the new communicator for the patch
Output Parameters:
+ dmz - the patch DM
. sfz - the PetscSF mapping the patch+halo to the zoomed version
. sfzr - the PetscSF mapping the patch to the restricted zoomed version
Level: intermediate
Note: All processes in commz should have the same rank (could autosplit comm)
.seealso: DMPatchSolve()
*/
PetscErrorCode DMPatchZoom(DM dm, Vec X, MatStencil lower, MatStencil upper, MPI_Comm commz, DM *dmz, PetscSF *sfz, PetscSF *sfzr)
{
DMDAStencilType st;
MatStencil blower, bupper, loclower, locupper;
IS is;
const PetscInt *ranges, *indices;
PetscInt *localPoints = NULL;
PetscSFNode *remotePoints = NULL;
PetscInt dim, dof;
PetscInt M, N, P, rM, rN, rP, halo = 1, sxb, syb, szb, sxr, syr, szr, exr, eyr, ezr, mxb, myb, mzb, i, j, k, q;
PetscMPIInt size;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)dm), &size);CHKERRQ(ierr);
/* Create patch DM */
ierr = DMDAGetInfo(dm, &dim, &M, &N, &P, 0,0,0, &dof, 0,0,0,0, &st);CHKERRQ(ierr);
/* Get piece for rank r, expanded by halo */
bupper.i = PetscMin(M, upper.i + halo); blower.i = PetscMax(lower.i - halo, 0);
bupper.j = PetscMin(N, upper.j + halo); blower.j = PetscMax(lower.j - halo, 0);
bupper.k = PetscMin(P, upper.k + halo); blower.k = PetscMax(lower.k - halo, 0);
rM = bupper.i - blower.i;
rN = bupper.j - blower.j;
rP = bupper.k - blower.k;
if (commz != MPI_COMM_NULL) {
ierr = DMDACreate(commz, dmz);CHKERRQ(ierr);
ierr = DMSetDimension(*dmz, dim);CHKERRQ(ierr);
ierr = DMDASetSizes(*dmz, rM, rN, rP);CHKERRQ(ierr);
ierr = DMDASetNumProcs(*dmz, PETSC_DECIDE, PETSC_DECIDE, PETSC_DECIDE);CHKERRQ(ierr);
ierr = DMDASetBoundaryType(*dmz, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE, DM_BOUNDARY_NONE);CHKERRQ(ierr);
ierr = DMDASetDof(*dmz, dof);CHKERRQ(ierr);
ierr = DMDASetStencilType(*dmz, st);CHKERRQ(ierr);
ierr = DMDASetStencilWidth(*dmz, 0);CHKERRQ(ierr);
ierr = DMDASetOwnershipRanges(*dmz, NULL, NULL, NULL);CHKERRQ(ierr);
ierr = DMSetFromOptions(*dmz);CHKERRQ(ierr);
ierr = DMSetUp(*dmz);CHKERRQ(ierr);
ierr = DMDAGetCorners(*dmz, &sxb, &syb, &szb, &mxb, &myb, &mzb);CHKERRQ(ierr);
sxr = PetscMax(sxb, lower.i - blower.i);
syr = PetscMax(syb, lower.j - blower.j);
szr = PetscMax(szb, lower.k - blower.k);
exr = PetscMin(sxb+mxb, upper.i - blower.i);
eyr = PetscMin(syb+myb, upper.j - blower.j);
ezr = PetscMin(szb+mzb, upper.k - blower.k);
ierr = PetscMalloc2(rM*rN*rP,&localPoints,rM*rN*rP,&remotePoints);CHKERRQ(ierr);
} else {
sxr = syr = szr = exr = eyr = ezr = sxb = syb = szb = mxb = myb = mzb = 0;
}
/* Create SF for restricted map */
ierr = VecGetOwnershipRanges(X,&ranges);CHKERRQ(ierr);
loclower.i = blower.i + sxr; locupper.i = blower.i + exr;
loclower.j = blower.j + syr; locupper.j = blower.j + eyr;
loclower.k = blower.k + szr; locupper.k = blower.k + ezr;
ierr = DMDACreatePatchIS(dm, &loclower, &locupper, &is);CHKERRQ(ierr);
ierr = ISGetIndices(is, &indices);CHKERRQ(ierr);
q = 0;
for (k = szb; k < szb+mzb; ++k) {
if ((k < szr) || (k >= ezr)) continue;
for (j = syb; j < syb+myb; ++j) {
if ((j < syr) || (j >= eyr)) continue;
for (i = sxb; i < sxb+mxb; ++i) {
const PetscInt lp = ((k-szb)*rN + (j-syb))*rM + i-sxb;
PetscInt r;
if ((i < sxr) || (i >= exr)) continue;
localPoints[q] = lp;
ierr = PetscFindInt(indices[q], size+1, ranges, &r);CHKERRQ(ierr);
remotePoints[q].rank = r < 0 ? -(r+1) - 1 : r;
remotePoints[q].index = indices[q] - ranges[remotePoints[q].rank];
++q;
}
}
}
ierr = ISRestoreIndices(is, &indices);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)dm), sfzr);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *sfzr, "Restricted Map");CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sfzr, M*N*P, q, localPoints, PETSC_COPY_VALUES, remotePoints, PETSC_COPY_VALUES);CHKERRQ(ierr);
/* Create SF for buffered map */
loclower.i = blower.i + sxb; locupper.i = blower.i + sxb+mxb;
loclower.j = blower.j + syb; locupper.j = blower.j + syb+myb;
loclower.k = blower.k + szb; locupper.k = blower.k + szb+mzb;
ierr = DMDACreatePatchIS(dm, &loclower, &locupper, &is);CHKERRQ(ierr);
ierr = ISGetIndices(is, &indices);CHKERRQ(ierr);
q = 0;
for (k = szb; k < szb+mzb; ++k) {
for (j = syb; j < syb+myb; ++j) {
for (i = sxb; i < sxb+mxb; ++i, ++q) {
PetscInt r;
localPoints[q] = q;
ierr = PetscFindInt(indices[q], size+1, ranges, &r);CHKERRQ(ierr);
remotePoints[q].rank = r < 0 ? -(r+1) - 1 : r;
remotePoints[q].index = indices[q] - ranges[remotePoints[q].rank];
}
}
}
ierr = ISRestoreIndices(is, &indices);CHKERRQ(ierr);
ierr = ISDestroy(&is);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)dm), sfz);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *sfz, "Buffered Map");CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sfz, M*N*P, q, localPoints, PETSC_COPY_VALUES, remotePoints, PETSC_COPY_VALUES);CHKERRQ(ierr);
ierr = PetscFree2(localPoints, remotePoints);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
ISBuildTwoSided - Takes an IS that describes where we will go. Generates an IS that contains new numbers from remote or local
on the IS.
Collective on IS
Input Parameters
. to - an IS describes where we will go. Negative target rank will be ignored
. toindx - an IS describes what indices should send. NULL means sending natural numbering
Output Parameter:
. rows - contains new numbers from remote or local
Level: advanced
.seealso: MatPartitioningCreate(), ISPartitioningToNumbering(), ISPartitioningCount()
@*/
PetscErrorCode ISBuildTwoSided(IS ito,IS toindx, IS *rows)
{
const PetscInt *ito_indices,*toindx_indices;
PetscInt *send_indices,rstart,*recv_indices,nrecvs,nsends;
PetscInt *tosizes,*fromsizes,i,j,*tosizes_tmp,*tooffsets_tmp,ito_ln;
PetscMPIInt *toranks,*fromranks,size,target_rank,*fromperm_newtoold,nto,nfrom;
PetscLayout isrmap;
MPI_Comm comm;
PetscSF sf;
PetscSFNode *iremote;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)ito,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = ISGetLocalSize(ito,&ito_ln);CHKERRQ(ierr);
/* why we do not have ISGetLayout? */
isrmap = ito->map;
ierr = PetscLayoutGetRange(isrmap,&rstart,NULL);CHKERRQ(ierr);
ierr = ISGetIndices(ito,&ito_indices);CHKERRQ(ierr);
ierr = PetscCalloc2(size,&tosizes_tmp,size+1,&tooffsets_tmp);CHKERRQ(ierr);
for(i=0; i<ito_ln; i++){
if(ito_indices[i]<0) continue;
#if defined(PETSC_USE_DEBUG)
if(ito_indices[i]>=size) SETERRQ2(comm,PETSC_ERR_ARG_OUTOFRANGE,"target rank %d is larger than communicator size %d ",ito_indices[i],size);
#endif
tosizes_tmp[ito_indices[i]]++;
}
nto = 0;
for(i=0; i<size; i++){
tooffsets_tmp[i+1] = tooffsets_tmp[i]+tosizes_tmp[i];
if(tosizes_tmp[i]>0) nto++;
}
ierr = PetscCalloc2(nto,&toranks,2*nto,&tosizes);CHKERRQ(ierr);
nto = 0;
for(i=0; i<size; i++){
if(tosizes_tmp[i]>0){
toranks[nto] = i;
tosizes[2*nto] = tosizes_tmp[i];/* size */
tosizes[2*nto+1] = tooffsets_tmp[i];/* offset */
nto++;
}
}
nsends = tooffsets_tmp[size];
ierr = PetscCalloc1(nsends,&send_indices);CHKERRQ(ierr);
if(toindx){
ierr = ISGetIndices(toindx,&toindx_indices);CHKERRQ(ierr);
}
for(i=0; i<ito_ln; i++){
if(ito_indices[i]<0) continue;
target_rank = ito_indices[i];
send_indices[tooffsets_tmp[target_rank]] = toindx? toindx_indices[i]:(i+rstart);
tooffsets_tmp[target_rank]++;
}
if(toindx){
ierr = ISRestoreIndices(toindx,&toindx_indices);CHKERRQ(ierr);
}
ierr = ISRestoreIndices(ito,&ito_indices);CHKERRQ(ierr);
ierr = PetscFree2(tosizes_tmp,tooffsets_tmp);CHKERRQ(ierr);
ierr = PetscCommBuildTwoSided(comm,2,MPIU_INT,nto,toranks,tosizes,&nfrom,&fromranks,&fromsizes);CHKERRQ(ierr);
ierr = PetscFree2(toranks,tosizes);CHKERRQ(ierr);
ierr = PetscCalloc1(nfrom,&fromperm_newtoold);CHKERRQ(ierr);
for(i=0; i<nfrom; i++){
fromperm_newtoold[i] = i;
}
ierr = PetscSortMPIIntWithArray(nfrom,fromranks,fromperm_newtoold);CHKERRQ(ierr);
nrecvs = 0;
for(i=0; i<nfrom; i++){
nrecvs += fromsizes[i*2];
}
ierr = PetscCalloc1(nrecvs,&recv_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(nrecvs,&iremote);CHKERRQ(ierr);
nrecvs = 0;
for(i=0; i<nfrom; i++){
for(j=0; j<fromsizes[2*fromperm_newtoold[i]]; j++){
iremote[nrecvs].rank = fromranks[i];
iremote[nrecvs++].index = fromsizes[2*fromperm_newtoold[i]+1]+j;
}
}
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nsends,nrecvs,NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
/* how to put a prefix ? */
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,send_indices,recv_indices);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,send_indices,recv_indices);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree(fromranks);CHKERRQ(ierr);
ierr = PetscFree(fromsizes);CHKERRQ(ierr);
ierr = PetscFree(fromperm_newtoold);CHKERRQ(ierr);
ierr = PetscFree(send_indices);CHKERRQ(ierr);
if(rows){
ierr = PetscSortInt(nrecvs,recv_indices);CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, nrecvs,recv_indices,PETSC_OWN_POINTER,rows);CHKERRQ(ierr);
}else{
ierr = PetscFree(recv_indices);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode PetscSFCreateSectionSF(PetscSF sf, PetscSection section, PetscSF *sectionSF)
{
PetscInt numRanks;
const PetscInt *ranks, *rankOffsets;
const PetscMPIInt *localPoints, *remotePoints;
PetscInt numPoints, numIndices = 0;
PetscInt *remoteOffsets;
PetscInt *localIndices;
PetscSFNode *remoteIndices;
PetscInt i, r, ind;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFGetRanks(sf, &numRanks, &ranks, &rankOffsets, &localPoints, &remotePoints);
CHKERRQ(ierr);
numPoints = rankOffsets[numRanks];
for (i = 0; i < numPoints; ++i) {
PetscInt dof;
ierr = PetscSectionGetDof(section, localPoints[i], &dof);
CHKERRQ(ierr);
numIndices += dof;
}
/* Communicate offsets for ghosted points */
#if 0
PetscInt *localOffsets;
ierr = PetscMalloc2(numPoints,PetscInt,&localOffsets,numPoints,PetscInt,&remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSectionGetOffset(section, localPoints[i], &localOffsets[i]);
CHKERRQ(ierr);
}
ierr = PetscSFBcastBegin(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i]);
CHKERRQ(ierr);
}
#else
ierr = PetscMalloc((section->atlasLayout.pEnd - section->atlasLayout.pStart) * sizeof(PetscInt), &remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
for (i = section->atlasLayout.pStart; i < section->atlasLayout.pEnd; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i-section->atlasLayout.pStart]);
CHKERRQ(ierr);
}
#endif
ierr = PetscSynchronizedFlush(((PetscObject) sf)->comm);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscInt), &localIndices);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscSFNode), &remoteIndices);
CHKERRQ(ierr);
/* Create new index graph */
for (r = 0, ind = 0; r < numRanks; ++r) {
PetscInt rank = ranks[r];
for (i = rankOffsets[r]; i < rankOffsets[r+1]; ++i) {
PetscInt localPoint = localPoints[i];
PetscInt remoteOffset = remoteOffsets[localPoint-section->atlasLayout.pStart];
PetscInt localOffset, dof, d;
ierr = PetscSectionGetOffset(section, localPoint, &localOffset);
CHKERRQ(ierr);
ierr = PetscSectionGetDof(section, localPoint, &dof);
CHKERRQ(ierr);
for (d = 0; d < dof; ++d, ++ind) {
localIndices[ind] = localOffset+d;
remoteIndices[ind].rank = rank;
remoteIndices[ind].index = remoteOffset+d;
}
}
}
ierr = PetscFree(remoteOffsets);
CHKERRQ(ierr);
if (numIndices != ind) SETERRQ2(((PetscObject) sf)->comm, PETSC_ERR_PLIB, "Inconsistency in indices, %d should be %d", ind, numIndices);
ierr = PetscSFCreate(((PetscObject) sf)->comm, sectionSF);
CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sectionSF, numIndices, numIndices, localIndices, PETSC_OWN_POINTER, remoteIndices, PETSC_OWN_POINTER);
CHKERRQ(ierr);
ierr = PetscSFView(*sectionSF, NULL);
CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/* Distribute cones
- Partitioning: input partition point map and naive sf, output sf with inverse of map, distribute points
- Distribute section: input current sf, communicate sizes and offsets, output local section and offsets (only use for new sf)
- Create SF for values: input current sf and offsets, output new sf
- Distribute values: input new sf, communicate values
*/
PetscErrorCode DistributeMesh(DM dm, AppCtx *user, PetscSF *pointSF, DM *parallelDM)
{
MPI_Comm comm = ((PetscObject) dm)->comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS cellPart, part;
PetscSection cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF;
ISLocalToGlobalMapping renumbering;
PetscSF coneSF;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets, newConesSize;
PetscInt *cones, *newCones;
PetscMPIInt numProcs, rank, p;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = DMMeshGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = DMMeshCreatePartition(dm, &cellPartSection, &cellPart, height);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) {
numRemoteRanks = numProcs;
} else {
numRemoteRanks = 0;
}
ierr = PetscMalloc(numRemoteRanks * sizeof(PetscSFNode), &remoteRanks);CHKERRQ(ierr);
for(p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, PETSC_NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(partSF, PETSC_NULL);CHKERRQ(ierr);
/* Close the partition over the mesh */
ierr = DMMeshCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMMeshCreate(comm, parallelDM);CHKERRQ(ierr);
ierr = DMMeshSetDimension(*parallelDM, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *parallelDM, "Parallel Mesh");CHKERRQ(ierr);
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, pointSF);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(*pointSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, PETSC_NULL);CHKERRQ(ierr);
/* Cleanup */
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMMeshGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMMeshGetConeSection(*parallelDM, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(*pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMMeshSetUp(*parallelDM);CHKERRQ(ierr);
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(*pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMMeshGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMMeshGetCones(*parallelDM, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApply(renumbering, IS_GTOLM_MASK, newConesSize, newCones, PETSC_NULL, newCones);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
/* Create supports and stratify sieve */
ierr = DMMeshSymmetrize(*parallelDM);CHKERRQ(ierr);
ierr = DMMeshStratify(*parallelDM);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
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);
}
/*
* The interface should be easy to use for both MatGetSubMatrix (parallel sub-matrix) and MatGetSubMatrices (sequential sub-matrices)
* */
static PetscErrorCode MatGetSubMatrix_MPIAdj_data(Mat adj,IS irows, IS icols, PetscInt **sadj_xadj,PetscInt **sadj_adjncy,PetscInt **sadj_values)
{
PetscInt nlrows_is,icols_n,i,j,nroots,nleaves,owner,rlocalindex,*ncols_send,*ncols_recv;
PetscInt nlrows_mat,*adjncy_recv,Ncols_recv,Ncols_send,*xadj_recv,*values_recv;
PetscInt *ncols_recv_offsets,loc,rnclos,*sadjncy,*sxadj,*svalues,isvalue;
const PetscInt *irows_indices,*icols_indices,*xadj, *adjncy;
Mat_MPIAdj *a = (Mat_MPIAdj*)adj->data;
PetscLayout rmap;
MPI_Comm comm;
PetscSF sf;
PetscSFNode *iremote;
PetscBool done;
PetscErrorCode ierr;
PetscFunctionBegin;
/* communicator */
ierr = PetscObjectGetComm((PetscObject)adj,&comm);CHKERRQ(ierr);
/* Layouts */
ierr = MatGetLayouts(adj,&rmap,PETSC_NULL);CHKERRQ(ierr);
/* get rows information */
ierr = ISGetLocalSize(irows,&nlrows_is);CHKERRQ(ierr);
ierr = ISGetIndices(irows,&irows_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(nlrows_is,&iremote);CHKERRQ(ierr);
/* construct sf graph*/
nleaves = nlrows_is;
for(i=0; i<nlrows_is; i++){
owner = -1;
rlocalindex = -1;
ierr = PetscLayoutFindOwnerIndex(rmap,irows_indices[i],&owner,&rlocalindex);CHKERRQ(ierr);
iremote[i].rank = owner;
iremote[i].index = rlocalindex;
}
ierr = MatGetRowIJ(adj,0,PETSC_FALSE,PETSC_FALSE,&nlrows_mat,&xadj,&adjncy,&done);CHKERRQ(ierr);
ierr = PetscCalloc4(nlrows_mat,&ncols_send,nlrows_is,&xadj_recv,nlrows_is+1,&ncols_recv_offsets,nlrows_is,&ncols_recv);CHKERRQ(ierr);
nroots = nlrows_mat;
for(i=0; i<nlrows_mat; i++){
ncols_send[i] = xadj[i+1]-xadj[i];
}
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,ncols_send,ncols_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,ncols_send,ncols_recv);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,xadj,xadj_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,xadj,xadj_recv);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
Ncols_recv =0;
for(i=0; i<nlrows_is; i++){
Ncols_recv += ncols_recv[i];
ncols_recv_offsets[i+1] = ncols_recv[i]+ncols_recv_offsets[i];
}
Ncols_send = 0;
for(i=0; i<nlrows_mat; i++){
Ncols_send += ncols_send[i];
}
ierr = PetscCalloc1(Ncols_recv,&iremote);CHKERRQ(ierr);
ierr = PetscCalloc1(Ncols_recv,&adjncy_recv);CHKERRQ(ierr);
nleaves = Ncols_recv;
Ncols_recv = 0;
for(i=0; i<nlrows_is; i++){
ierr = PetscLayoutFindOwner(rmap,irows_indices[i],&owner);CHKERRQ(ierr);
for(j=0; j<ncols_recv[i]; j++){
iremote[Ncols_recv].rank = owner;
iremote[Ncols_recv++].index = xadj_recv[i]+j;
}
}
ierr = ISRestoreIndices(irows,&irows_indices);CHKERRQ(ierr);
/*if we need to deal with edge weights ???*/
if(a->values){isvalue=1;}else{isvalue=0;}
/*involve a global communication */
/*ierr = MPI_Allreduce(&isvalue,&isvalue,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr);*/
if(isvalue){ierr = PetscCalloc1(Ncols_recv,&values_recv);CHKERRQ(ierr);}
nroots = Ncols_send;
ierr = PetscSFCreate(comm,&sf);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf,nroots,nleaves,PETSC_NULL,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,adjncy,adjncy_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,adjncy,adjncy_recv);CHKERRQ(ierr);
if(isvalue){
ierr = PetscSFBcastBegin(sf,MPIU_INT,a->values,values_recv);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,a->values,values_recv);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MatRestoreRowIJ(adj,0,PETSC_FALSE,PETSC_FALSE,&nlrows_mat,&xadj,&adjncy,&done);CHKERRQ(ierr);
ierr = ISGetLocalSize(icols,&icols_n);CHKERRQ(ierr);
ierr = ISGetIndices(icols,&icols_indices);CHKERRQ(ierr);
rnclos = 0;
for(i=0; i<nlrows_is; i++){
for(j=ncols_recv_offsets[i]; j<ncols_recv_offsets[i+1]; j++){
ierr = PetscFindInt(adjncy_recv[j], icols_n, icols_indices, &loc);CHKERRQ(ierr);
if(loc<0){
adjncy_recv[j] = -1;
if(isvalue) values_recv[j] = -1;
ncols_recv[i]--;
}else{
rnclos++;
}
}
}
ierr = ISRestoreIndices(icols,&icols_indices);CHKERRQ(ierr);
ierr = PetscCalloc1(rnclos,&sadjncy);CHKERRQ(ierr);
if(isvalue) {ierr = PetscCalloc1(rnclos,&svalues);CHKERRQ(ierr);}
ierr = PetscCalloc1(nlrows_is+1,&sxadj);CHKERRQ(ierr);
rnclos = 0;
for(i=0; i<nlrows_is; i++){
for(j=ncols_recv_offsets[i]; j<ncols_recv_offsets[i+1]; j++){
if(adjncy_recv[j]<0) continue;
sadjncy[rnclos] = adjncy_recv[j];
if(isvalue) svalues[rnclos] = values_recv[j];
rnclos++;
}
}
for(i=0; i<nlrows_is; i++){
sxadj[i+1] = sxadj[i]+ncols_recv[i];
}
if(sadj_xadj) { *sadj_xadj = sxadj;}else { ierr = PetscFree(sxadj);CHKERRQ(ierr);}
if(sadj_adjncy){ *sadj_adjncy = sadjncy;}else{ ierr = PetscFree(sadjncy);CHKERRQ(ierr);}
if(sadj_values){
if(isvalue) *sadj_values = svalues; else *sadj_values=0;
}else{
if(isvalue) {ierr = PetscFree(svalues);CHKERRQ(ierr);}
}
ierr = PetscFree4(ncols_send,xadj_recv,ncols_recv_offsets,ncols_recv);CHKERRQ(ierr);
ierr = PetscFree(adjncy_recv);CHKERRQ(ierr);
if(isvalue) {ierr = PetscFree(values_recv);CHKERRQ(ierr);}
PetscFunctionReturn(0);
}
/*@C
PetscSFGetMultiSF - gets the inner SF implemeting gathers and scatters
Collective
Input Argument:
. sf - star forest that may contain roots with 0 or with more than 1 vertex
Output Arguments:
. multi - star forest with split roots, such that each root has degree exactly 1
Level: developer
Notes:
In most cases, users should use PetscSFGatherBegin() and PetscSFScatterBegin() instead of manipulating multi
directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming
edge, it is a candidate for future optimization that might involve its removal.
.seealso: PetscSFSetGraph(), PetscSFGatherBegin(), PetscSFScatterBegin()
@*/
PetscErrorCode PetscSFGetMultiSF(PetscSF sf,PetscSF *multi)
{
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
PetscValidPointer(multi,2);
if (sf->nroots < 0) { /* Graph has not been set yet; why do we need this? */
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
*multi = sf->multi;
PetscFunctionReturn(0);
}
if (!sf->multi) {
const PetscInt *indegree;
PetscInt i,*inoffset,*outones,*outoffset,maxlocal;
PetscSFNode *remote;
ierr = PetscSFComputeDegreeBegin(sf,&indegree);CHKERRQ(ierr);
ierr = PetscSFComputeDegreeEnd(sf,&indegree);CHKERRQ(ierr);
for (i=0,maxlocal=0; i<sf->nleaves; i++) maxlocal = PetscMax(maxlocal,(sf->mine ? sf->mine[i] : i)+1);
ierr = PetscMalloc3(sf->nroots+1,&inoffset,maxlocal,&outones,maxlocal,&outoffset);CHKERRQ(ierr);
inoffset[0] = 0;
for (i=0; i<sf->nroots; i++) inoffset[i+1] = inoffset[i] + indegree[i];
for (i=0; i<maxlocal; i++) outones[i] = 1;
ierr = PetscSFFetchAndOpBegin(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFFetchAndOpEnd(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
for (i=0; i<sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */
#if 0
#if defined(PETSC_USE_DEBUG) /* Check that the expected number of increments occurred */
for (i=0; i<sf->nroots; i++) {
if (inoffset[i] + indegree[i] != inoffset[i+1]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect result after PetscSFFetchAndOp");
}
#endif
#endif
ierr = PetscMalloc1(sf->nleaves,&remote);CHKERRQ(ierr);
for (i=0; i<sf->nleaves; i++) {
remote[i].rank = sf->remote[i].rank;
remote[i].index = outoffset[sf->mine ? sf->mine[i] : i];
}
ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
if (sf->rankorder) { /* Sort the ranks */
PetscMPIInt rank;
PetscInt *inranks,*newoffset,*outranks,*newoutoffset,*tmpoffset,maxdegree;
PetscSFNode *newremote;
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
for (i=0,maxdegree=0; i<sf->nroots; i++) maxdegree = PetscMax(maxdegree,indegree[i]);
ierr = PetscMalloc5(sf->multi->nroots,&inranks,sf->multi->nroots,&newoffset,maxlocal,&outranks,maxlocal,&newoutoffset,maxdegree,&tmpoffset);CHKERRQ(ierr);
for (i=0; i<maxlocal; i++) outranks[i] = rank;
ierr = PetscSFReduceBegin(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
/* Sort the incoming ranks at each vertex, build the inverse map */
for (i=0; i<sf->nroots; i++) {
PetscInt j;
for (j=0; j<indegree[i]; j++) tmpoffset[j] = j;
ierr = PetscSortIntWithArray(indegree[i],inranks+inoffset[i],tmpoffset);CHKERRQ(ierr);
for (j=0; j<indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j;
}
ierr = PetscSFBcastBegin(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
ierr = PetscMalloc1(sf->nleaves,&newremote);CHKERRQ(ierr);
for (i=0; i<sf->nleaves; i++) {
newremote[i].rank = sf->remote[i].rank;
newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i];
}
ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,newremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscFree5(inranks,newoffset,outranks,newoutoffset,tmpoffset);CHKERRQ(ierr);
}
ierr = PetscFree3(inoffset,outones,outoffset);CHKERRQ(ierr);
}
*multi = sf->multi;
PetscFunctionReturn(0);
}
