/*@
DMPlexDistributeField - Distribute field data to match a given PetscSF, usually the SF from mesh distribution
Collective on DM
Input Parameters:
+ dm - The DMPlex object
. pointSF - The PetscSF describing the communication pattern
. originalSection - The PetscSection for existing data layout
- originalVec - The existing data
Output Parameters:
+ newSection - The PetscSF describing the new data layout
- newVec - The new data
Level: developer
.seealso: DMPlexDistribute(), DMPlexDistributeData()
@*/
PetscErrorCode DMPlexDistributeField(DM dm, PetscSF pointSF, PetscSection originalSection, Vec originalVec, PetscSection newSection, Vec newVec)
{
PetscSF fieldSF;
PetscInt *remoteOffsets, fieldSize;
PetscScalar *originalValues, *newValues;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscLogEventBegin(DMPLEX_DistributeField,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalSection, &remoteOffsets, newSection);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newSection, &fieldSize);CHKERRQ(ierr);
ierr = VecSetSizes(newVec, fieldSize, PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetType(newVec,dm->vectype);CHKERRQ(ierr);
ierr = VecGetArray(originalVec, &originalValues);CHKERRQ(ierr);
ierr = VecGetArray(newVec, &newValues);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(pointSF, originalSection, remoteOffsets, newSection, &fieldSF);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(fieldSF, MPIU_SCALAR, originalValues, newValues);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(fieldSF, MPIU_SCALAR, originalValues, newValues);CHKERRQ(ierr);
ierr = PetscSFDestroy(&fieldSF);CHKERRQ(ierr);
ierr = VecRestoreArray(newVec, &newValues);CHKERRQ(ierr);
ierr = VecRestoreArray(originalVec, &originalValues);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeField,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode DistributeCoordinates(DM dm, PetscSF pointSF, DM parallelDM)
{
PetscSF coordSF;
PetscSection originalCoordSection, newCoordSection;
Vec coordinates, newCoordinates;
PetscScalar *coords, *newCoords;
PetscInt *remoteOffsets, coordSize;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = DMMeshGetCoordinateSection(dm, &originalCoordSection);CHKERRQ(ierr);
ierr = DMMeshGetCoordinateSection(parallelDM, &newCoordSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalCoordSection, &remoteOffsets, newCoordSection);CHKERRQ(ierr);
ierr = DMMeshGetCoordinateVec(dm, &coordinates);CHKERRQ(ierr);
ierr = DMMeshGetCoordinateVec(parallelDM, &newCoordinates);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newCoordSection, &coordSize);CHKERRQ(ierr);
ierr = VecSetSizes(newCoordinates, coordSize, PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(newCoordinates);CHKERRQ(ierr);
ierr = VecGetArray(coordinates, &coords);CHKERRQ(ierr);
ierr = VecGetArray(newCoordinates, &newCoords);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(pointSF, originalCoordSection, remoteOffsets, newCoordSection, &coordSF);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coordSF, MPIU_SCALAR, coords, newCoords);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coordSF, MPIU_SCALAR, coords, newCoords);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coordSF);CHKERRQ(ierr);
ierr = VecRestoreArray(newCoordinates, &newCoords);CHKERRQ(ierr);
ierr = VecRestoreArray(coordinates, &coords);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
DMPlexClone - Creates a DMPlex object with the same mesh as the original.
Collective on MPI_Comm
Input Parameter:
. dm - The original DMPlex object
Output Parameter:
. newdm - The new DMPlex object
Level: beginner
.keywords: DMPlex, create
@*/
PetscErrorCode DMPlexClone(DM dm, DM *newdm)
{
DM_Plex *mesh;
Vec coords;
void *ctx;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
PetscValidPointer(newdm,2);
ierr = DMCreate(PetscObjectComm((PetscObject)dm), newdm);CHKERRQ(ierr);
ierr = PetscSFDestroy(&(*newdm)->sf);CHKERRQ(ierr);
ierr = PetscObjectReference((PetscObject) dm->sf);CHKERRQ(ierr);
(*newdm)->sf = dm->sf;
mesh = (DM_Plex*) dm->data;
mesh->refct++;
(*newdm)->data = mesh;
ierr = PetscObjectChangeTypeName((PetscObject) *newdm, DMPLEX);CHKERRQ(ierr);
ierr = DMInitialize_Plex(*newdm);CHKERRQ(ierr);
ierr = DMGetApplicationContext(dm, &ctx);CHKERRQ(ierr);
ierr = DMSetApplicationContext(*newdm, ctx);CHKERRQ(ierr);
ierr = DMGetCoordinatesLocal(dm, &coords);CHKERRQ(ierr);
if (coords) {
ierr = DMSetCoordinatesLocal(*newdm, coords);CHKERRQ(ierr);
} else {
ierr = DMGetCoordinates(dm, &coords);CHKERRQ(ierr);
if (coords) {ierr = DMSetCoordinates(*newdm, coords);CHKERRQ(ierr);}
}
PetscFunctionReturn(0);
}
PetscErrorCode DMLabelDistribute_Internal(DMLabel label, PetscSF sf, PetscSection *leafSection, PetscInt **leafStrata)
{
MPI_Comm comm;
PetscInt s, l, nroots, nleaves, dof, offset, size;
PetscInt *remoteOffsets, *rootStrata, *rootIdx;
PetscSection rootSection;
PetscSF labelSF;
PetscErrorCode ierr;
PetscFunctionBegin;
if (label) {ierr = DMLabelMakeAllValid_Private(label);CHKERRQ(ierr);}
ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
/* Build a section of stratum values per point, generate the according SF
and distribute point-wise stratum values to leaves. */
ierr = PetscSFGetGraph(sf, &nroots, &nleaves, NULL, NULL);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSection);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSection, 0, nroots);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
for (l = 0; l < label->stratumSizes[s]; l++) {
ierr = PetscSectionGetDof(rootSection, label->points[s][l], &dof);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSection, label->points[s][l], dof+1);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSection);CHKERRQ(ierr);
/* Create a point-wise array of stratum values */
ierr = PetscSectionGetStorageSize(rootSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &rootStrata);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots, &rootIdx);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
for (l = 0; l < label->stratumSizes[s]; l++) {
const PetscInt p = label->points[s][l];
ierr = PetscSectionGetOffset(rootSection, p, &offset);CHKERRQ(ierr);
rootStrata[offset+rootIdx[p]++] = label->stratumValues[s];
}
}
}
/* Build SF that maps label points to remote processes */
ierr = PetscSectionCreate(comm, leafSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sf, rootSection, &remoteOffsets, *leafSection);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, rootSection, remoteOffsets, *leafSection, &labelSF);CHKERRQ(ierr);
ierr = PetscFree(remoteOffsets);CHKERRQ(ierr);
/* Send the strata for each point over the derived SF */
ierr = PetscSectionGetStorageSize(*leafSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(labelSF, MPIU_INT, rootStrata, *leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(labelSF, MPIU_INT, rootStrata, *leafStrata);CHKERRQ(ierr);
/* Clean up */
ierr = PetscFree(rootStrata);CHKERRQ(ierr);
ierr = PetscFree(rootIdx);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSection);CHKERRQ(ierr);
ierr = PetscSFDestroy(&labelSF);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
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);
}
int main(int argc, char *argv[])
{
MPI_Comm comm;
DM dm;
PetscSF sf;
AppCtx user;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscInitialize(&argc, &argv, (char*) 0, help);
CHKERRQ(ierr);
comm = PETSC_COMM_WORLD;
ierr = ProcessOptions(comm, &user);
CHKERRQ(ierr);
ierr = CreateMesh(comm, &user, &dm);
CHKERRQ(ierr);
ierr = DMMeshConvertOverlapToSF(dm, &sf);
CHKERRQ(ierr);
{
PetscSection section;
PetscSF sectionSF;
ierr = DMMeshGetCoordinateSection(dm, §ion);
CHKERRQ(ierr);
ierr = PetscSectionView(section, PETSC_VIEWER_STDOUT_SELF);
CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, section, §ionSF);
CHKERRQ(ierr);
ierr = PetscSFDestroy(§ionSF);
CHKERRQ(ierr);
ierr = PetscSectionDestroy(§ion);
CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sf);
CHKERRQ(ierr);
ierr = DMDestroy(&dm);
CHKERRQ(ierr);
ierr = PetscFinalize();
PetscFunctionReturn(0);
}
PetscErrorCode MatISSetPreallocation_IS(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
{
Mat_IS *matis = (Mat_IS*)(B->data);
PetscSF sf;
PetscInt bs,i,nroots,*rootdata,nleaves,*leafdata,nlocalcols;
const PetscInt *gidxs;
PetscErrorCode ierr;
PetscFunctionBegin;
if (!matis->A) {
SETERRQ(PetscObjectComm((PetscObject)B),PETSC_ERR_SUP,"You should first call MatSetLocalToGlobalMapping");
}
ierr = MatGetLocalSize(B,&nroots,NULL);CHKERRQ(ierr);
ierr = MatGetSize(matis->A,&nleaves,&nlocalcols);CHKERRQ(ierr);
ierr = MatGetBlockSize(matis->A,&bs);CHKERRQ(ierr);
ierr = PetscCalloc2(nroots,&rootdata,nleaves,&leafdata);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)B),&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,B->rmap,nleaves,NULL,PETSC_COPY_VALUES,gidxs);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingRestoreIndices(matis->mapping,&gidxs);CHKERRQ(ierr);
if (!d_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += d_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += d_nnz[i];
}
if (!o_nnz) {
for (i=0;i<nroots;i++) rootdata[i] += o_nz;
} else {
for (i=0;i<nroots;i++) rootdata[i] += o_nnz[i];
}
ierr = PetscSFBcastBegin(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,rootdata,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves;i++) {
leafdata[i] = PetscMin(leafdata[i],nlocalcols);
}
ierr = MatSeqAIJSetPreallocation(matis->A,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i*bs]/bs;
}
ierr = MatSeqBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
for (i=0;i<nleaves/bs;i++) {
leafdata[i] = leafdata[i]-i;
}
ierr = MatSeqSBAIJSetPreallocation(matis->A,bs,0,leafdata);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree2(rootdata,leafdata);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@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
PetscSFReset - Reset a star forest so that different sizes or neighbors can be used
Collective
Input Arguments:
. sf - star forest
Level: advanced
.seealso: PetscSFCreate(), PetscSFSetGraph(), PetscSFDestroy()
@*/
PetscErrorCode PetscSFReset(PetscSF sf)
{
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
sf->mine = NULL;
ierr = PetscFree(sf->mine_alloc);CHKERRQ(ierr);
sf->remote = NULL;
ierr = PetscFree(sf->remote_alloc);CHKERRQ(ierr);
ierr = PetscFree4(sf->ranks,sf->roffset,sf->rmine,sf->rremote);CHKERRQ(ierr);
ierr = PetscFree(sf->degree);CHKERRQ(ierr);
if (sf->ingroup != MPI_GROUP_NULL) {ierr = MPI_Group_free(&sf->ingroup);CHKERRQ(ierr);}
if (sf->outgroup != MPI_GROUP_NULL) {ierr = MPI_Group_free(&sf->outgroup);CHKERRQ(ierr);}
ierr = PetscSFDestroy(&sf->multi);CHKERRQ(ierr);
sf->graphset = PETSC_FALSE;
if (sf->ops->Reset) {ierr = (*sf->ops->Reset)(sf);CHKERRQ(ierr);}
sf->setupcalled = PETSC_FALSE;
PetscFunctionReturn(0);
}
PetscErrorCode MatDestroy_IS(Mat A)
{
PetscErrorCode ierr;
Mat_IS *b = (Mat_IS*)A->data;
PetscFunctionBegin;
ierr = MatDestroy(&b->A);CHKERRQ(ierr);
ierr = VecScatterDestroy(&b->cctx);CHKERRQ(ierr);
ierr = VecScatterDestroy(&b->rctx);CHKERRQ(ierr);
ierr = VecDestroy(&b->x);CHKERRQ(ierr);
ierr = VecDestroy(&b->y);CHKERRQ(ierr);
ierr = PetscSFDestroy(&b->sf);CHKERRQ(ierr);
ierr = PetscFree2(b->sf_rootdata,b->sf_leafdata);CHKERRQ(ierr);
ierr = PetscFree(A->data);CHKERRQ(ierr);
ierr = PetscObjectChangeTypeName((PetscObject)A,0);CHKERRQ(ierr);
ierr = PetscObjectComposeFunction((PetscObject)A,"MatISGetLocalMat_C",NULL);CHKERRQ(ierr);
ierr = PetscObjectComposeFunction((PetscObject)A,"MatISSetLocalMat_C",NULL);CHKERRQ(ierr);
ierr = PetscObjectComposeFunction((PetscObject)A,"MatISGetMPIXAIJ_C",NULL);CHKERRQ(ierr);
ierr = PetscObjectComposeFunction((PetscObject)A,"MatISSetPreallocation_C",NULL);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
DMPlexDistributeData - Distribute field data to match a given PetscSF, usually the SF from mesh distribution
Collective on DM
Input Parameters:
+ dm - The DMPlex object
. pointSF - The PetscSF describing the communication pattern
. originalSection - The PetscSection for existing data layout
. datatype - The type of data
- originalData - The existing data
Output Parameters:
+ newSection - The PetscSF describing the new data layout
- newData - The new data
Level: developer
.seealso: DMPlexDistribute(), DMPlexDistributeField()
@*/
PetscErrorCode DMPlexDistributeData(DM dm, PetscSF pointSF, PetscSection originalSection, MPI_Datatype datatype, void *originalData, PetscSection newSection, void **newData)
{
PetscSF fieldSF;
PetscInt *remoteOffsets, fieldSize;
PetscMPIInt dataSize;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscLogEventBegin(DMPLEX_DistributeData,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalSection, &remoteOffsets, newSection);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newSection, &fieldSize);CHKERRQ(ierr);
ierr = MPI_Type_size(datatype, &dataSize);CHKERRQ(ierr);
ierr = PetscMalloc(fieldSize * dataSize, newData);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(pointSF, originalSection, remoteOffsets, newSection, &fieldSF);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(fieldSF, datatype, originalData, *newData);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(fieldSF, datatype, originalData, *newData);CHKERRQ(ierr);
ierr = PetscSFDestroy(&fieldSF);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeData,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char *argv[])
{
MPI_Comm comm;
DM dm, parallelDM;
PetscSF pointSF;
AppCtx user;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscInitialize(&argc, &argv, (char *) 0, help);CHKERRQ(ierr);
comm = PETSC_COMM_WORLD;
ierr = ProcessOptions(comm, &user);CHKERRQ(ierr);
ierr = CreateMesh(comm, &user, &dm);CHKERRQ(ierr);
ierr = DistributeMesh(dm, &user, &pointSF, ¶llelDM);CHKERRQ(ierr);
ierr = DistributeCoordinates(dm, pointSF, parallelDM);CHKERRQ(ierr);
ierr = DMSetFromOptions(parallelDM);CHKERRQ(ierr);
ierr = PetscSFDestroy(&pointSF);CHKERRQ(ierr);
ierr = DMDestroy(¶llelDM);CHKERRQ(ierr);
ierr = DMDestroy(&dm);CHKERRQ(ierr);
ierr = PetscFinalize();
PetscFunctionReturn(0);
}
/*@
DMNetworkDistribute - Distributes the network and moves associated component data.
Collective
Input Parameter:
+ oldDM - the original DMNetwork object
- overlap - The overlap of partitions, 0 is the default
Output Parameter:
. distDM - the distributed DMNetwork object
Notes:
This routine should be called only when using multiple processors.
Distributes the network with <overlap>-overlapping partitioning of the edges.
Level: intermediate
.seealso: DMNetworkCreate
@*/
PetscErrorCode DMNetworkDistribute(DM oldDM, PetscInt overlap,DM *distDM)
{
PetscErrorCode ierr;
DM_Network *oldDMnetwork = (DM_Network*)oldDM->data;
PetscSF pointsf;
DM newDM;
DM_Network *newDMnetwork;
PetscFunctionBegin;
ierr = DMNetworkCreate(PetscObjectComm((PetscObject)oldDM),&newDM);CHKERRQ(ierr);
newDMnetwork = (DM_Network*)newDM->data;
newDMnetwork->dataheadersize = sizeof(struct _p_DMNetworkComponentHeader)/sizeof(DMNetworkComponentGenericDataType);
/* Distribute plex dm and dof section */
ierr = DMPlexDistribute(oldDMnetwork->plex,overlap,&pointsf,&newDMnetwork->plex);CHKERRQ(ierr);
/* Distribute dof section */
ierr = PetscSectionCreate(PetscObjectComm((PetscObject)oldDM),&newDMnetwork->DofSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointsf,oldDMnetwork->DofSection,NULL,newDMnetwork->DofSection);CHKERRQ(ierr);
ierr = PetscSectionCreate(PetscObjectComm((PetscObject)oldDM),&newDMnetwork->DataSection);CHKERRQ(ierr);
/* Distribute data and associated section */
ierr = DMPlexDistributeData(newDMnetwork->plex,pointsf,oldDMnetwork->DataSection,MPI_INT,(void*)oldDMnetwork->componentdataarray,newDMnetwork->DataSection,(void**)&newDMnetwork->componentdataarray);CHKERRQ(ierr);
/* Destroy point SF */
ierr = PetscSFDestroy(&pointsf);CHKERRQ(ierr);
ierr = PetscSectionGetChart(newDMnetwork->DataSection,&newDMnetwork->pStart,&newDMnetwork->pEnd);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(newDMnetwork->plex,0, &newDMnetwork->eStart,&newDMnetwork->eEnd);CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(newDMnetwork->plex,1,&newDMnetwork->vStart,&newDMnetwork->vEnd);CHKERRQ(ierr);
newDMnetwork->nEdges = newDMnetwork->eEnd - newDMnetwork->eStart;
newDMnetwork->nNodes = newDMnetwork->vEnd - newDMnetwork->vStart;
newDMnetwork->NNodes = oldDMnetwork->NNodes;
newDMnetwork->NEdges = oldDMnetwork->NEdges;
/* Set Dof section as the default section for dm */
ierr = DMSetDefaultSection(newDMnetwork->plex,newDMnetwork->DofSection);CHKERRQ(ierr);
ierr = DMGetDefaultGlobalSection(newDMnetwork->plex,&newDMnetwork->GlobalDofSection);CHKERRQ(ierr);
*distDM = newDM;
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);
}
PETSC_EXTERN void PETSC_STDCALL petscsfdestroy_(PetscSF *sf, int *__ierr ){
*__ierr = PetscSFDestroy(sf);
}
PetscErrorCode DMPlexPreallocateOperator(DM dm, PetscInt bs, PetscSection section, PetscSection sectionGlobal, PetscInt dnz[], PetscInt onz[], PetscInt dnzu[], PetscInt onzu[], Mat A, PetscBool fillMatrix)
{
MPI_Comm comm;
MatType mtype;
PetscSF sf, sfDof, sfAdj;
PetscSection leafSectionAdj, rootSectionAdj, sectionAdj, anchorSectionAdj;
PetscInt nroots, nleaves, l, p;
const PetscInt *leaves;
const PetscSFNode *remotes;
PetscInt dim, pStart, pEnd, numDof, globalOffStart, globalOffEnd, numCols;
PetscInt *tmpAdj = NULL, *adj, *rootAdj, *anchorAdj = NULL, *cols, *remoteOffsets;
PetscInt adjSize;
PetscLayout rLayout;
PetscInt locRows, rStart, rEnd, r;
PetscMPIInt size;
PetscBool doCommLocal, doComm, debug = PETSC_FALSE, isSymBlock, isSymSeqBlock, isSymMPIBlock;
PetscBool useAnchors;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
PetscValidHeaderSpecific(section, PETSC_SECTION_CLASSID, 3);
PetscValidHeaderSpecific(sectionGlobal, PETSC_SECTION_CLASSID, 4);
PetscValidHeaderSpecific(A, MAT_CLASSID, 9);
if (dnz) PetscValidPointer(dnz,5);
if (onz) PetscValidPointer(onz,6);
if (dnzu) PetscValidPointer(dnzu,7);
if (onzu) PetscValidPointer(onzu,8);
ierr = PetscLogEventBegin(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = PetscOptionsGetBool(NULL, "-dm_view_preallocation", &debug, NULL);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMGetPointSF(dm, &sf);CHKERRQ(ierr);
ierr = PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL);CHKERRQ(ierr);
doCommLocal = (size > 1) && (nroots >= 0) ? PETSC_TRUE : PETSC_FALSE;
ierr = MPI_Allreduce(&doCommLocal, &doComm, 1, MPIU_BOOL, MPI_LAND, comm);CHKERRQ(ierr);
/* Create dof SF based on point SF */
if (debug) {
ierr = PetscPrintf(comm, "Input Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(section, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input Global Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionGlobal, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Input SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sf, NULL);CHKERRQ(ierr);
}
ierr = PetscSFCreateRemoteOffsets(sf, section, section, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, section, remoteOffsets, section, &sfDof);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Dof SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfDof, NULL);CHKERRQ(ierr);
}
/* Create section for dof adjacency (dof ==> # adj dof) */
ierr = PetscSectionGetChart(section, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(section, &numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(leafSectionAdj, 0, numDof);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSectionAdj, 0, numDof);CHKERRQ(ierr);
/* Fill in the ghost dofs on the interface */
ierr = PetscSFGetGraph(sf, NULL, &nleaves, &leaves, &remotes);CHKERRQ(ierr);
/* use constraints in finding adjacency in this routine */
ierr = DMPlexGetAdjacencyUseAnchors(dm,&useAnchors);CHKERRQ(ierr);
ierr = DMPlexSetAdjacencyUseAnchors(dm,PETSC_TRUE);CHKERRQ(ierr);
/*
section - maps points to (# dofs, local dofs)
sectionGlobal - maps points to (# dofs, global dofs)
leafSectionAdj - maps unowned local dofs to # adj dofs
rootSectionAdj - maps owned local dofs to # adj dofs
adj - adj global dofs indexed by leafSectionAdj
rootAdj - adj global dofs indexed by rootSectionAdj
sf - describes shared points across procs
sfDof - describes shared dofs across procs
sfAdj - describes shared adjacent dofs across procs
** The bootstrapping process involves six rounds with similar structure of visiting neighbors of each point.
(0). If there are point-to-point constraints, add the adjacencies of constrained points to anchors in anchorAdj
(This is done in DMPlexComputeAnchorAdjacencies())
1. Visit unowned points on interface, count adjacencies placing in leafSectionAdj
Reduce those counts to rootSectionAdj (now redundantly counting some interface points)
2. Visit owned points on interface, count adjacencies placing in rootSectionAdj
Create sfAdj connecting rootSectionAdj and leafSectionAdj
3. Visit unowned points on interface, write adjacencies to adj
Gather adj to rootAdj (note that there is redundancy in rootAdj when multiple procs find the same adjacencies)
4. Visit owned points on interface, write adjacencies to rootAdj
Remove redundancy in rootAdj
** The last two traversals use transitive closure
5. Visit all owned points in the subdomain, count dofs for each point (sectionAdj)
Allocate memory addressed by sectionAdj (cols)
6. Visit all owned points in the subdomain, insert dof adjacencies into cols
** Knowing all the column adjacencies, check ownership and sum into dnz and onz
*/
ierr = DMPlexComputeAnchorAdjacencies(dm,section,sectionGlobal,&anchorSectionAdj,&anchorAdj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(leafSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation on Leaves:\n");CHKERRQ(ierr);
ierr = PetscSectionView(leafSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get maximum remote adjacency sizes for owned dofs on interface (roots) */
if (doComm) {
ierr = PetscSFReduceBegin(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfDof, MPIU_INT, leafSectionAdj->atlasDof, rootSectionAdj->atlasDof, MPI_SUM);CHKERRQ(ierr);
}
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Add in local adjacency sizes for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = off; d < off+dof; ++d) {
ierr = PetscSectionAddDof(rootSectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after local additions:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Create adj SF based on dof SF */
ierr = PetscSFCreateRemoteOffsets(sfDof, rootSectionAdj, leafSectionAdj, &remoteOffsets);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sfDof, rootSectionAdj, remoteOffsets, leafSectionAdj, &sfAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency SF for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSFView(sfAdj, NULL);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfDof);CHKERRQ(ierr);
/* Create leaf adjacency */
ierr = PetscSectionSetUp(leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(leafSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscCalloc1(adjSize, &adj);CHKERRQ(ierr);
for (l = 0; l < nleaves; ++l) {
PetscInt dof, off, d, q, anDof, anOff;
PetscInt p = leaves[l], numAdj = PETSC_DETERMINE;
if ((p < pStart) || (p >= pEnd)) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt aoff, i = 0;
ierr = PetscSectionGetOffset(leafSectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
adj[aoff+i] = (ngoff < 0 ? -(ngoff+1) : ngoff) + nd;
++i;
}
}
for (q = 0; q < anDof; q++) {
adj[aoff+i] = anchorAdj[anOff+q];
++i;
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Leaf adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, adj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Gather adjacenct indices to root */
ierr = PetscSectionGetStorageSize(rootSectionAdj, &adjSize);CHKERRQ(ierr);
ierr = PetscMalloc1(adjSize, &rootAdj);CHKERRQ(ierr);
for (r = 0; r < adjSize; ++r) rootAdj[r] = -1;
if (doComm) {
ierr = PetscSFGatherBegin(sfAdj, MPIU_INT, adj, rootAdj);CHKERRQ(ierr);
ierr = PetscSFGatherEnd(sfAdj, MPIU_INT, adj, rootAdj);CHKERRQ(ierr);
}
ierr = PetscSFDestroy(&sfAdj);CHKERRQ(ierr);
ierr = PetscFree(adj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices after gather\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Add in local adjacency indices for owned dofs on interface (roots) */
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, adof, dof, off, d, q, anDof, anOff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
PetscInt adof, aoff, i;
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
i = adof-1;
for (q = 0; q < anDof; q++) {
rootAdj[aoff+i] = anchorAdj[anOff+q];
--i;
}
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd) {
rootAdj[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
--i;
}
}
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Root adjacency indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Compress indices */
ierr = PetscSectionSetUp(rootSectionAdj);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt dof, cdof, off, d;
PetscInt adof, aoff;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
if (!dof) continue;
ierr = PetscSectionGetDof(rootSectionAdj, off, &adof);CHKERRQ(ierr);
if (adof <= 0) continue;
for (d = off; d < off+dof-cdof; ++d) {
ierr = PetscSectionGetDof(rootSectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, d, &aoff);CHKERRQ(ierr);
ierr = PetscSortRemoveDupsInt(&adof, &rootAdj[aoff]);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSectionAdj, d, adof);CHKERRQ(ierr);
}
}
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Adjancency Section for Preallocation on Roots after compression:\n");CHKERRQ(ierr);
ierr = PetscSectionView(rootSectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Root adjacency indices after compression\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, adjSize, rootAdj, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Build adjacency section: Maps global indices to sets of adjacent global indices */
ierr = PetscSectionGetOffsetRange(sectionGlobal, &globalOffStart, &globalOffEnd);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, §ionAdj);CHKERRQ(ierr);
ierr = PetscSectionSetChart(sectionAdj, globalOffStart, globalOffEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
ierr = PetscSectionSetDof(sectionAdj, goff+d, rdof);CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, noff;
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, padj, &noff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, ndof-ncdof);CHKERRQ(ierr);
}
}
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
if (anDof) {
for (d = goff; d < goff+dof-cdof; ++d) {
ierr = PetscSectionAddDof(sectionAdj, d, anDof);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(sectionAdj);CHKERRQ(ierr);
if (debug) {
ierr = PetscPrintf(comm, "Adjacency Section for Preallocation:\n");CHKERRQ(ierr);
ierr = PetscSectionView(sectionAdj, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
}
/* Get adjacent indices */
ierr = PetscSectionGetStorageSize(sectionAdj, &numCols);CHKERRQ(ierr);
ierr = PetscMalloc1(numCols, &cols);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt numAdj = PETSC_DETERMINE, dof, cdof, off, goff, d, q, anDof, anOff;
PetscBool found = PETSC_TRUE;
ierr = PetscSectionGetDof(section, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, p, &cdof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(section, p, &off);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, p, &goff);CHKERRQ(ierr);
for (d = 0; d < dof-cdof; ++d) {
PetscInt ldof, rdof;
ierr = PetscSectionGetDof(leafSectionAdj, off+d, &ldof);CHKERRQ(ierr);
ierr = PetscSectionGetDof(rootSectionAdj, off+d, &rdof);CHKERRQ(ierr);
if (ldof > 0) {
/* We do not own this point */
} else if (rdof > 0) {
PetscInt aoff, roff;
ierr = PetscSectionGetOffset(sectionAdj, goff+d, &aoff);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(rootSectionAdj, off+d, &roff);CHKERRQ(ierr);
ierr = PetscMemcpy(&cols[aoff], &rootAdj[roff], rdof * sizeof(PetscInt));CHKERRQ(ierr);
} else {
found = PETSC_FALSE;
}
}
if (found) continue;
ierr = DMPlexGetAdjacency(dm, p, &numAdj, &tmpAdj);CHKERRQ(ierr);
ierr = PetscSectionGetDof(anchorSectionAdj, p, &anDof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(anchorSectionAdj, p, &anOff);CHKERRQ(ierr);
for (d = goff; d < goff+dof-cdof; ++d) {
PetscInt adof, aoff, i = 0;
ierr = PetscSectionGetDof(sectionAdj, d, &adof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, d, &aoff);CHKERRQ(ierr);
for (q = 0; q < numAdj; ++q) {
const PetscInt padj = tmpAdj[q];
PetscInt ndof, ncdof, ngoff, nd;
const PetscInt *ncind;
/* Adjacent points may not be in the section chart */
if ((padj < pStart) || (padj >= pEnd)) continue;
ierr = PetscSectionGetDof(section, padj, &ndof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintDof(section, padj, &ncdof);CHKERRQ(ierr);
ierr = PetscSectionGetConstraintIndices(section, padj, &ncind);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionGlobal, padj, &ngoff);CHKERRQ(ierr);
for (nd = 0; nd < ndof-ncdof; ++nd, ++i) {
cols[aoff+i] = ngoff < 0 ? -(ngoff+1)+nd : ngoff+nd;
}
}
for (q = 0; q < anDof; q++, i++) {
cols[aoff+i] = anchorAdj[anOff + q];
}
if (i != adof) SETERRQ4(PETSC_COMM_SELF, PETSC_ERR_PLIB, "Invalid number of entries %D != %D for dof %D (point %D)", i, adof, d, p);
}
}
ierr = PetscSectionDestroy(&anchorSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&leafSectionAdj);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSectionAdj);CHKERRQ(ierr);
ierr = PetscFree(anchorAdj);CHKERRQ(ierr);
ierr = PetscFree(rootAdj);CHKERRQ(ierr);
ierr = PetscFree(tmpAdj);CHKERRQ(ierr);
/* Debugging */
if (debug) {
IS tmp;
ierr = PetscPrintf(comm, "Column indices\n");CHKERRQ(ierr);
ierr = ISCreateGeneral(comm, numCols, cols, PETSC_USE_POINTER, &tmp);CHKERRQ(ierr);
ierr = ISView(tmp, NULL);CHKERRQ(ierr);
ierr = ISDestroy(&tmp);CHKERRQ(ierr);
}
/* Create allocation vectors from adjacency graph */
ierr = MatGetLocalSize(A, &locRows, NULL);CHKERRQ(ierr);
ierr = PetscLayoutCreate(PetscObjectComm((PetscObject)A), &rLayout);CHKERRQ(ierr);
ierr = PetscLayoutSetLocalSize(rLayout, locRows);CHKERRQ(ierr);
ierr = PetscLayoutSetBlockSize(rLayout, 1);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(rLayout);CHKERRQ(ierr);
ierr = PetscLayoutGetRange(rLayout, &rStart, &rEnd);CHKERRQ(ierr);
ierr = PetscLayoutDestroy(&rLayout);CHKERRQ(ierr);
/* Only loop over blocks of rows */
if (rStart%bs || rEnd%bs) SETERRQ3(PetscObjectComm((PetscObject)A), PETSC_ERR_ARG_WRONG, "Invalid layout [%d, %d) for matrix, must be divisible by block size %d", rStart, rEnd, bs);
for (r = rStart/bs; r < rEnd/bs; ++r) {
const PetscInt row = r*bs;
PetscInt numCols, cStart, c;
ierr = PetscSectionGetDof(sectionAdj, row, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, row, &cStart);CHKERRQ(ierr);
for (c = cStart; c < cStart+numCols; ++c) {
if ((cols[c] >= rStart*bs) && (cols[c] < rEnd*bs)) {
++dnz[r-rStart];
if (cols[c] >= row) ++dnzu[r-rStart];
} else {
++onz[r-rStart];
if (cols[c] >= row) ++onzu[r-rStart];
}
}
}
if (bs > 1) {
for (r = 0; r < locRows/bs; ++r) {
dnz[r] /= bs;
onz[r] /= bs;
dnzu[r] /= bs;
onzu[r] /= bs;
}
}
/* Set matrix pattern */
ierr = MatXAIJSetPreallocation(A, bs, dnz, onz, dnzu, onzu);CHKERRQ(ierr);
ierr = MatSetOption(A, MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr);
/* Check for symmetric storage */
ierr = MatGetType(A, &mtype);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSBAIJ, &isSymBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATSEQSBAIJ, &isSymSeqBlock);CHKERRQ(ierr);
ierr = PetscStrcmp(mtype, MATMPISBAIJ, &isSymMPIBlock);CHKERRQ(ierr);
if (isSymBlock || isSymSeqBlock || isSymMPIBlock) {ierr = MatSetOption(A, MAT_IGNORE_LOWER_TRIANGULAR, PETSC_TRUE);CHKERRQ(ierr);}
/* Fill matrix with zeros */
if (fillMatrix) {
PetscScalar *values;
PetscInt maxRowLen = 0;
for (r = rStart; r < rEnd; ++r) {
PetscInt len;
ierr = PetscSectionGetDof(sectionAdj, r, &len);CHKERRQ(ierr);
maxRowLen = PetscMax(maxRowLen, len);
}
ierr = PetscCalloc1(maxRowLen, &values);CHKERRQ(ierr);
for (r = rStart; r < rEnd; ++r) {
PetscInt numCols, cStart;
ierr = PetscSectionGetDof(sectionAdj, r, &numCols);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(sectionAdj, r, &cStart);CHKERRQ(ierr);
ierr = MatSetValues(A, 1, &r, numCols, &cols[cStart], values, INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree(values);CHKERRQ(ierr);
ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
}
/* restore original useAnchors */
ierr = DMPlexSetAdjacencyUseAnchors(dm,useAnchors);CHKERRQ(ierr);
ierr = PetscSectionDestroy(§ionAdj);CHKERRQ(ierr);
ierr = PetscFree(cols);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Preallocate,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
DMPlexDistribute - Distributes the mesh and any associated sections.
Not Collective
Input Parameter:
+ dm - The original DMPlex object
. partitioner - The partitioning package, or NULL for the default
- overlap - The overlap of partitions, 0 is the default
Output Parameter:
+ sf - The PetscSF used for point distribution
- parallelMesh - The distributed DMPlex object, or NULL
Note: If the mesh was not distributed, the return value is NULL.
The user can control the definition of adjacency for the mesh using DMPlexGetAdjacencyUseCone() and
DMPlexSetAdjacencyUseClosure(). They should choose the combination appropriate for the function
representation on the mesh.
Level: intermediate
.keywords: mesh, elements
.seealso: DMPlexCreate(), DMPlexDistributeByFace(), DMPlexSetAdjacencyUseCone(), DMPlexSetAdjacencyUseClosure()
@*/
PetscErrorCode DMPlexDistribute(DM dm, const char partitioner[], PetscInt overlap, PetscSF *sf, DM *dmParallel)
{
DM_Plex *mesh = (DM_Plex*) dm->data, *pmesh;
MPI_Comm comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS origCellPart, origPart, cellPart, part;
PetscSection origCellPartSection, origPartSection, cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF, pointSF, coneSF;
ISLocalToGlobalMapping renumbering;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets;
PetscInt *cones, *newCones, newConesSize;
PetscBool flg;
PetscMPIInt rank, numProcs, p;
PetscErrorCode ierr;
PetscFunctionBegin;
PetscValidHeaderSpecific(dm, DM_CLASSID, 1);
if (sf) PetscValidPointer(sf,4);
PetscValidPointer(dmParallel,5);
ierr = PetscLogEventBegin(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
*dmParallel = NULL;
if (numProcs == 1) PetscFunctionReturn(0);
ierr = DMPlexGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = PetscLogEventBegin(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
if (overlap > 1) SETERRQ(PetscObjectComm((PetscObject)dm), PETSC_ERR_SUP, "Overlap > 1 not yet implemented");
ierr = DMPlexCreatePartition(dm, partitioner, height, overlap > 0 ? PETSC_TRUE : PETSC_FALSE, &cellPartSection, &cellPart, &origCellPartSection, &origCellPart);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) numRemoteRanks = numProcs;
else numRemoteRanks = 0;
ierr = PetscMalloc1(numRemoteRanks, &remoteRanks);CHKERRQ(ierr);
for (p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-partition_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, NULL);CHKERRQ(ierr);
if (origCellPart) {
ierr = PetscPrintf(comm, "Original Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(origCellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(origCellPart, NULL);CHKERRQ(ierr);
}
ierr = PetscSFView(partSF, NULL);CHKERRQ(ierr);
}
/* Close the partition over the mesh */
ierr = DMPlexCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMPlexCreate(comm, dmParallel);CHKERRQ(ierr);
ierr = DMPlexSetDimension(*dmParallel, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *dmParallel, "Parallel Mesh");CHKERRQ(ierr);
pmesh = (DM_Plex*) (*dmParallel)->data;
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, &pointSF);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, NULL);CHKERRQ(ierr);
ierr = PetscSFView(pointSF, NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, NULL);CHKERRQ(ierr);
}
ierr = PetscLogEventEnd(DMPLEX_Partition,dm,0,0,0);CHKERRQ(ierr);
ierr = PetscLogEventBegin(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMPlexGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMPlexGetConeSection(*dmParallel, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMSetUp(*dmParallel);CHKERRQ(ierr);
{
PetscInt pStart, pEnd, p;
ierr = PetscSectionGetChart(newConeSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p = pStart; p < pEnd; ++p) {
PetscInt coneSize;
ierr = PetscSectionGetDof(newConeSection, p, &coneSize);CHKERRQ(ierr);
pmesh->maxConeSize = PetscMax(pmesh->maxConeSize, coneSize);
}
}
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMPlexGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetCones(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApplyBlock(renumbering, IS_GTOLM_MASK, newConesSize, newCones, NULL, newCones);CHKERRQ(ierr);
ierr = PetscOptionsHasName(((PetscObject) dm)->prefix, "-cones_view", &flg);CHKERRQ(ierr);
if (flg) {
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, NULL);CHKERRQ(ierr);
}
ierr = DMPlexGetConeOrientations(dm, &cones);CHKERRQ(ierr);
ierr = DMPlexGetConeOrientations(*dmParallel, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_DistributeCones,dm,0,0,0);CHKERRQ(ierr);
/* Create supports and stratify sieve */
{
PetscInt pStart, pEnd;
ierr = PetscSectionGetChart(pmesh->coneSection, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSectionSetChart(pmesh->supportSection, pStart, pEnd);CHKERRQ(ierr);
}
ierr = DMPlexSymmetrize(*dmParallel);CHKERRQ(ierr);
ierr = DMPlexStratify(*dmParallel);CHKERRQ(ierr);
/* Distribute Coordinates */
{
PetscSection originalCoordSection, newCoordSection;
Vec originalCoordinates, newCoordinates;
PetscInt bs;
const char *name;
ierr = DMGetCoordinateSection(dm, &originalCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinateSection(*dmParallel, &newCoordSection);CHKERRQ(ierr);
ierr = DMGetCoordinatesLocal(dm, &originalCoordinates);CHKERRQ(ierr);
ierr = VecCreate(comm, &newCoordinates);CHKERRQ(ierr);
ierr = PetscObjectGetName((PetscObject) originalCoordinates, &name);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) newCoordinates, name);CHKERRQ(ierr);
ierr = DMPlexDistributeField(dm, pointSF, originalCoordSection, originalCoordinates, newCoordSection, newCoordinates);CHKERRQ(ierr);
ierr = DMSetCoordinatesLocal(*dmParallel, newCoordinates);CHKERRQ(ierr);
ierr = VecGetBlockSize(originalCoordinates, &bs);CHKERRQ(ierr);
ierr = VecSetBlockSize(newCoordinates, bs);CHKERRQ(ierr);
ierr = VecDestroy(&newCoordinates);CHKERRQ(ierr);
}
/* Distribute labels */
ierr = PetscLogEventBegin(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
{
DMLabel next = mesh->labels, newNext = pmesh->labels;
PetscInt numLabels = 0, l;
/* Bcast number of labels */
while (next) {++numLabels; next = next->next;}
ierr = MPI_Bcast(&numLabels, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
next = mesh->labels;
for (l = 0; l < numLabels; ++l) {
DMLabel labelNew;
PetscBool isdepth;
/* Skip "depth" because it is recreated */
if (!rank) {ierr = PetscStrcmp(next->name, "depth", &isdepth);CHKERRQ(ierr);}
ierr = MPI_Bcast(&isdepth, 1, MPIU_BOOL, 0, comm);CHKERRQ(ierr);
if (isdepth) {if (!rank) next = next->next; continue;}
ierr = DMLabelDistribute(next, partSection, part, renumbering, &labelNew);CHKERRQ(ierr);
/* Insert into list */
if (newNext) newNext->next = labelNew;
else pmesh->labels = labelNew;
newNext = labelNew;
if (!rank) next = next->next;
}
}
ierr = PetscLogEventEnd(DMPLEX_DistributeLabels,dm,0,0,0);CHKERRQ(ierr);
/* Setup hybrid structure */
{
const PetscInt *gpoints;
PetscInt depth, n, d;
for (d = 0; d <= dim; ++d) {pmesh->hybridPointMax[d] = mesh->hybridPointMax[d];}
ierr = MPI_Bcast(pmesh->hybridPointMax, dim+1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(renumbering, &n);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetIndices(renumbering, &gpoints);CHKERRQ(ierr);
ierr = DMPlexGetDepth(dm, &depth);CHKERRQ(ierr);
for (d = 0; d <= dim; ++d) {
PetscInt pmax = pmesh->hybridPointMax[d], newmax = 0, pEnd, stratum[2], p;
if (pmax < 0) continue;
ierr = DMPlexGetDepthStratum(dm, d > depth ? depth : d, &stratum[0], &stratum[1]);CHKERRQ(ierr);
ierr = DMPlexGetDepthStratum(*dmParallel, d, NULL, &pEnd);CHKERRQ(ierr);
ierr = MPI_Bcast(stratum, 2, MPIU_INT, 0, comm);CHKERRQ(ierr);
for (p = 0; p < n; ++p) {
const PetscInt point = gpoints[p];
if ((point >= stratum[0]) && (point < stratum[1]) && (point >= pmax)) ++newmax;
}
if (newmax > 0) pmesh->hybridPointMax[d] = pEnd - newmax;
else pmesh->hybridPointMax[d] = -1;
}
ierr = ISLocalToGlobalMappingRestoreIndices(renumbering, &gpoints);CHKERRQ(ierr);
}
/* Cleanup Partition */
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Create point SF for parallel mesh */
ierr = PetscLogEventBegin(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
{
const PetscInt *leaves;
PetscSFNode *remotePoints, *rowners, *lowners;
PetscInt numRoots, numLeaves, numGhostPoints = 0, p, gp, *ghostPoints;
PetscInt pStart, pEnd;
ierr = DMPlexGetChart(*dmParallel, &pStart, &pEnd);CHKERRQ(ierr);
ierr = PetscSFGetGraph(pointSF, &numRoots, &numLeaves, &leaves, NULL);CHKERRQ(ierr);
ierr = PetscMalloc2(numRoots,&rowners,numLeaves,&lowners);CHKERRQ(ierr);
for (p=0; p<numRoots; p++) {
rowners[p].rank = -1;
rowners[p].index = -1;
}
if (origCellPart) {
/* Make sure points in the original partition are not assigned to other procs */
const PetscInt *origPoints;
ierr = DMPlexCreatePartitionClosure(dm, origCellPartSection, origCellPart, &origPartSection, &origPart);CHKERRQ(ierr);
ierr = ISGetIndices(origPart, &origPoints);CHKERRQ(ierr);
for (p = 0; p < numProcs; ++p) {
PetscInt dof, off, d;
ierr = PetscSectionGetDof(origPartSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(origPartSection, p, &off);CHKERRQ(ierr);
for (d = off; d < off+dof; ++d) {
rowners[origPoints[d]].rank = p;
}
}
ierr = ISRestoreIndices(origPart, &origPoints);CHKERRQ(ierr);
ierr = ISDestroy(&origPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origPartSection);CHKERRQ(ierr);
}
ierr = ISDestroy(&origCellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&origCellPartSection);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].rank == rank) { /* Either put in a bid or we know we own it */
lowners[p].rank = rank;
lowners[p].index = leaves ? leaves[p] : p;
} else if (lowners[p].rank >= 0) { /* Point already claimed so flag so that MAXLOC does not listen to us */
lowners[p].rank = -2;
lowners[p].index = -2;
}
}
for (p=0; p<numRoots; p++) { /* Root must not participate in the rediction, flag so that MAXLOC does not use */
rowners[p].rank = -3;
rowners[p].index = -3;
}
ierr = PetscSFReduceBegin(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(pointSF, MPIU_2INT, lowners, rowners, MPI_MAXLOC);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(pointSF, MPIU_2INT, rowners, lowners);CHKERRQ(ierr);
for (p = 0; p < numLeaves; ++p) {
if (lowners[p].rank < 0 || lowners[p].index < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Cell partition corrupt: point not claimed");
if (lowners[p].rank != rank) ++numGhostPoints;
}
ierr = PetscMalloc1(numGhostPoints, &ghostPoints);CHKERRQ(ierr);
ierr = PetscMalloc1(numGhostPoints, &remotePoints);CHKERRQ(ierr);
for (p = 0, gp = 0; p < numLeaves; ++p) {
if (lowners[p].rank != rank) {
ghostPoints[gp] = leaves ? leaves[p] : p;
remotePoints[gp].rank = lowners[p].rank;
remotePoints[gp].index = lowners[p].index;
++gp;
}
}
ierr = PetscFree2(rowners,lowners);CHKERRQ(ierr);
ierr = PetscSFSetGraph((*dmParallel)->sf, pEnd - pStart, numGhostPoints, ghostPoints, PETSC_OWN_POINTER, remotePoints, PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions((*dmParallel)->sf);CHKERRQ(ierr);
}
pmesh->useCone = mesh->useCone;
pmesh->useClosure = mesh->useClosure;
ierr = PetscLogEventEnd(DMPLEX_DistributeSF,dm,0,0,0);CHKERRQ(ierr);
/* Copy BC */
ierr = DMPlexCopyBoundary(dm, *dmParallel);CHKERRQ(ierr);
/* Cleanup */
if (sf) {*sf = pointSF;}
else {ierr = PetscSFDestroy(&pointSF);CHKERRQ(ierr);}
ierr = DMSetFromOptions(*dmParallel);CHKERRQ(ierr);
ierr = PetscLogEventEnd(DMPLEX_Distribute,dm,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode DMPatchSolve(DM dm)
{
MPI_Comm comm;
MPI_Comm commz;
DM dmc;
PetscSF sfz, sfzr;
Vec XC;
MatStencil patchSize, commSize, gridRank, lower, upper;
PetscInt M, N, P, i, j, k, l, m, n, p = 0;
PetscMPIInt rank, size;
PetscInt debug = 0;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)dm,&comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &size);CHKERRQ(ierr);
ierr = DMPatchGetCoarse(dm, &dmc);CHKERRQ(ierr);
ierr = DMPatchGetPatchSize(dm, &patchSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMPatchGetCommSize(dm, &commSize);CHKERRQ(ierr);
ierr = DMGetGlobalVector(dmc, &XC);CHKERRQ(ierr);
ierr = DMDAGetInfo(dmc, 0, &M, &N, &P, &l, &m, &n, 0,0,0,0,0,0);CHKERRQ(ierr);
M = PetscMax(M, 1); l = PetscMax(l, 1);
N = PetscMax(N, 1); m = PetscMax(m, 1);
P = PetscMax(P, 1); n = PetscMax(n, 1);
gridRank.i = rank % l;
gridRank.j = rank/l % m;
gridRank.k = rank/(l*m) % n;
if (commSize.i*commSize.j*commSize.k == size || commSize.i*commSize.j*commSize.k == 0) {
commSize.i = l; commSize.j = m; commSize.k = n;
commz = comm;
} else if (commSize.i*commSize.j*commSize.k == 1) {
commz = PETSC_COMM_SELF;
} else {
const PetscMPIInt newComm = ((gridRank.k/commSize.k)*(m/commSize.j) + gridRank.j/commSize.j)*(l/commSize.i) + (gridRank.i/commSize.i);
const PetscMPIInt newRank = ((gridRank.k%commSize.k)*commSize.j + gridRank.j%commSize.j)*commSize.i + (gridRank.i%commSize.i);
ierr = MPI_Comm_split(comm, newComm, newRank, &commz);CHKERRQ(ierr);
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d color %d key %d commz %d\n", rank, newComm, newRank, *((PetscMPIInt*) &commz));CHKERRQ(ierr);}
}
/*
Assumptions:
- patchSize divides gridSize
- commSize divides gridSize
- commSize divides l,m,n
Ignore multiple patches per rank for now
Multiple ranks per patch:
- l,m,n divides patchSize
- commSize divides patchSize
*/
for (k = 0; k < P; k += PetscMax(patchSize.k, 1)) {
for (j = 0; j < N; j += PetscMax(patchSize.j, 1)) {
for (i = 0; i < M; i += PetscMax(patchSize.i, 1), ++p) {
MPI_Comm commp = MPI_COMM_NULL;
DM dmz = NULL;
#if 0
DM dmf = NULL;
Mat interpz = NULL;
#endif
Vec XZ = NULL;
PetscScalar *xcarray = NULL;
PetscScalar *xzarray = NULL;
if ((gridRank.k/commSize.k == p/(l/commSize.i * m/commSize.j) % n/commSize.k) &&
(gridRank.j/commSize.j == p/(l/commSize.i) % m/commSize.j) &&
(gridRank.i/commSize.i == p % l/commSize.i)) {
if (debug) {ierr = PetscPrintf(PETSC_COMM_SELF, "Rank %d is accepting Patch %d\n", rank, p);CHKERRQ(ierr);}
commp = commz;
}
/* Zoom to coarse patch */
lower.i = i; lower.j = j; lower.k = k;
upper.i = i + patchSize.i; upper.j = j + patchSize.j; upper.k = k + patchSize.k;
ierr = DMPatchZoom(dmc, XC, lower, upper, commp, &dmz, &sfz, &sfzr);CHKERRQ(ierr);
lower.c = 0; /* initialize member, otherwise compiler issues warnings */
upper.c = 0; /* initialize member, otherwise compiler issues warnings */
/* Debug */
ierr = PetscPrintf(comm, "Patch %d: (%d, %d, %d)--(%d, %d, %d)\n", p, lower.i, lower.j, lower.k, upper.i, upper.j, upper.k);CHKERRQ(ierr);
if (dmz) {ierr = DMView(dmz, PETSC_VIEWER_STDOUT_(commz));CHKERRQ(ierr);}
ierr = PetscSFView(sfz, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
ierr = PetscSFView(sfzr, PETSC_VIEWER_STDOUT_(comm));CHKERRQ(ierr);
/* Scatter Xcoarse -> Xzoom */
if (dmz) {ierr = DMGetGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sfz, MPIU_SCALAR, xcarray, xzarray);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
#if 0
/* Interpolate Xzoom -> Xfine, note that this may be on subcomms */
ierr = DMRefine(dmz, MPI_COMM_NULL, &dmf);CHKERRQ(ierr);
ierr = DMCreateInterpolation(dmz, dmf, &interpz, NULL);CHKERRQ(ierr);
ierr = DMInterpolate(dmz, interpz, dmf);CHKERRQ(ierr);
/* Smooth Xfine using two-step smoother, normal smoother plus Kaczmarz---moves back and forth from dmzoom to dmfine */
/* Compute residual Rfine */
/* Restrict Rfine to Rzoom_restricted */
#endif
/* Scatter Rzoom_restricted -> Rcoarse_restricted */
if (XZ) {ierr = VecGetArray(XZ, &xzarray);CHKERRQ(ierr);}
ierr = VecGetArray(XC, &xcarray);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sfzr, MPIU_SCALAR, xzarray, xcarray, MPIU_SUM);CHKERRQ(ierr);
ierr = VecRestoreArray(XC, &xcarray);CHKERRQ(ierr);
if (XZ) {ierr = VecRestoreArray(XZ, &xzarray);CHKERRQ(ierr);}
if (dmz) {ierr = DMRestoreGlobalVector(dmz, &XZ);CHKERRQ(ierr);}
/* Compute global residual Rcoarse */
/* TauCoarse = Rcoarse - Rcoarse_restricted */
ierr = PetscSFDestroy(&sfz);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sfzr);CHKERRQ(ierr);
ierr = DMDestroy(&dmz);CHKERRQ(ierr);
}
}
}
ierr = DMRestoreGlobalVector(dmc, &XC);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode GreedyColoringLocalDistanceOne_Private(MatColoring mc,PetscReal *wts,PetscInt *lperm,ISColoringValue *colors)
{
PetscInt i,j,k,s,e,n,no,nd,nd_global,n_global,idx,ncols,maxcolors,masksize,ccol,*mask;
PetscErrorCode ierr;
Mat m=mc->mat;
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)m->data;
Mat md=NULL,mo=NULL;
const PetscInt *md_i,*mo_i,*md_j,*mo_j;
PetscBool isMPIAIJ,isSEQAIJ;
ISColoringValue pcol;
const PetscInt *cidx;
PetscInt *lcolors,*ocolors;
PetscReal *owts=NULL;
PetscSF sf;
PetscLayout layout;
PetscFunctionBegin;
ierr = MatGetSize(m,&n_global,NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(m,&s,&e);CHKERRQ(ierr);
n=e-s;
masksize=20;
nd_global = 0;
/* get the matrix communication structures */
ierr = PetscObjectTypeCompare((PetscObject)m, MATMPIAIJ, &isMPIAIJ); CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)m, MATSEQAIJ, &isSEQAIJ); CHKERRQ(ierr);
if (isMPIAIJ) {
/* get the CSR data for on and off diagonal portions of m */
Mat_SeqAIJ *dseq;
Mat_SeqAIJ *oseq;
md=aij->A;
dseq = (Mat_SeqAIJ*)md->data;
mo=aij->B;
oseq = (Mat_SeqAIJ*)mo->data;
md_i = dseq->i;
md_j = dseq->j;
mo_i = oseq->i;
mo_j = oseq->j;
} else if (isSEQAIJ) {
/* get the CSR data for m */
Mat_SeqAIJ *dseq;
/* no off-processor nodes */
md=m;
dseq = (Mat_SeqAIJ*)md->data;
mo=NULL;
no=0;
md_i = dseq->i;
md_j = dseq->j;
mo_i = NULL;
mo_j = NULL;
} else SETERRQ(PetscObjectComm((PetscObject)mc),PETSC_ERR_ARG_WRONG,"Matrix must be AIJ for greedy coloring");
ierr = MatColoringGetMaxColors(mc,&maxcolors);CHKERRQ(ierr);
if (mo) {
ierr = VecGetSize(aij->lvec,&no);CHKERRQ(ierr);
ierr = PetscMalloc2(no,&ocolors,no,&owts);CHKERRQ(ierr);
for(i=0;i<no;i++) {
ocolors[i]=maxcolors;
}
}
ierr = PetscMalloc1(masksize,&mask);CHKERRQ(ierr);
ierr = PetscMalloc1(n,&lcolors);CHKERRQ(ierr);
for(i=0;i<n;i++) {
lcolors[i]=maxcolors;
}
for (i=0;i<masksize;i++) {
mask[i]=-1;
}
if (mo) {
/* transfer neighbor weights */
ierr = PetscSFCreate(PetscObjectComm((PetscObject)m),&sf);CHKERRQ(ierr);
ierr = MatGetLayouts(m,&layout,NULL);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,layout,no,NULL,PETSC_COPY_VALUES,aij->garray);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_REAL,wts,owts);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_REAL,wts,owts);CHKERRQ(ierr);
}
while (nd_global < n_global) {
nd=n;
/* assign lowest possible color to each local vertex */
ierr = PetscLogEventBegin(MATCOLORING_Local,mc,0,0,0);CHKERRQ(ierr);
for (i=0;i<n;i++) {
idx=lperm[i];
if (lcolors[idx] == maxcolors) {
ncols = md_i[idx+1]-md_i[idx];
cidx = &(md_j[md_i[idx]]);
for (j=0;j<ncols;j++) {
if (lcolors[cidx[j]] != maxcolors) {
ccol=lcolors[cidx[j]];
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
if (mo) {
ncols = mo_i[idx+1]-mo_i[idx];
cidx = &(mo_j[mo_i[idx]]);
for (j=0;j<ncols;j++) {
if (ocolors[cidx[j]] != maxcolors) {
ccol=ocolors[cidx[j]];
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
}
for (j=0;j<masksize;j++) {
if (mask[j]!=idx) {
break;
}
}
pcol=j;
if (pcol>maxcolors)pcol=maxcolors;
lcolors[idx]=pcol;
}
}
ierr = PetscLogEventEnd(MATCOLORING_Local,mc,0,0,0);CHKERRQ(ierr);
if (mo) {
/* transfer neighbor colors */
ierr = PetscLogEventBegin(MATCOLORING_Comm,mc,0,0,0);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,lcolors,ocolors);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,lcolors,ocolors);CHKERRQ(ierr);
/* check for conflicts -- this is merely checking if any adjacent off-processor rows have the same color and marking the ones that are lower weight locally for changing */
for (i=0;i<n;i++) {
ncols = mo_i[i+1]-mo_i[i];
cidx = &(mo_j[mo_i[i]]);
for (j=0;j<ncols;j++) {
/* in the case of conflicts, the highest weight one stays and the others go */
if ((ocolors[cidx[j]] == lcolors[i]) && (owts[cidx[j]] > wts[i]) && lcolors[i] < maxcolors) {
lcolors[i]=maxcolors;
nd--;
}
}
}
nd_global=0;
}
ierr = MPIU_Allreduce(&nd,&nd_global,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mc));CHKERRQ(ierr);
}
for (i=0;i<n;i++) {
colors[i] = (ISColoringValue)lcolors[i];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
ierr = PetscFree(lcolors);CHKERRQ(ierr);
if (mo) {
ierr = PetscFree2(ocolors,owts);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
static PetscErrorCode GreedyColoringLocalDistanceTwo_Private(MatColoring mc,PetscReal *wts,PetscInt *lperm,ISColoringValue *colors)
{
MC_Greedy *gr = (MC_Greedy *) mc->data;
PetscInt i,j,k,l,s,e,n,nd,nd_global,n_global,idx,ncols,maxcolors,mcol,mcol_global,nd1cols,*mask,masksize,*d1cols,*bad,*badnext,nbad,badsize,ccol,no,cbad;
Mat m = mc->mat, mt;
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)m->data;
Mat md=NULL,mo=NULL;
const PetscInt *md_i,*mo_i,*md_j,*mo_j;
const PetscInt *rmd_i,*rmo_i,*rmd_j,*rmo_j;
PetscBool isMPIAIJ,isSEQAIJ;
PetscInt pcol,*dcolors,*ocolors;
ISColoringValue *badidx;
const PetscInt *cidx;
PetscReal *owts,*colorweights;
PetscInt *oconf,*conf;
PetscSF sf;
PetscLayout layout;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MatGetSize(m,&n_global,NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(m,&s,&e);CHKERRQ(ierr);
n=e-s;
nd_global = 0;
/* get the matrix communication structures */
ierr = PetscObjectTypeCompare((PetscObject)m, MATMPIAIJ, &isMPIAIJ); CHKERRQ(ierr);
ierr = PetscObjectTypeCompare((PetscObject)m, MATSEQAIJ, &isSEQAIJ); CHKERRQ(ierr);
if (isMPIAIJ) {
Mat_SeqAIJ *dseq;
Mat_SeqAIJ *oseq;
md=aij->A;
dseq = (Mat_SeqAIJ*)md->data;
mo=aij->B;
oseq = (Mat_SeqAIJ*)mo->data;
md_i = dseq->i;
md_j = dseq->j;
mo_i = oseq->i;
mo_j = oseq->j;
rmd_i = dseq->i;
rmd_j = dseq->j;
rmo_i = oseq->i;
rmo_j = oseq->j;
} else if (isSEQAIJ) {
Mat_SeqAIJ *dseq;
/* no off-processor nodes */
md=m;
dseq = (Mat_SeqAIJ*)md->data;
md_i = dseq->i;
md_j = dseq->j;
mo_i = NULL;
mo_j = NULL;
rmd_i = dseq->i;
rmd_j = dseq->j;
rmo_i = NULL;
rmo_j = NULL;
} else SETERRQ(PetscObjectComm((PetscObject)mc),PETSC_ERR_ARG_WRONG,"Matrix must be AIJ for greedy coloring");
if (!gr->symmetric) {
ierr = MatTranspose(m, MAT_INITIAL_MATRIX, &mt);CHKERRQ(ierr);
if (isSEQAIJ) {
Mat_SeqAIJ *dseq = (Mat_SeqAIJ*) mt->data;
rmd_i = dseq->i;
rmd_j = dseq->j;
rmo_i = NULL;
rmo_j = NULL;
} else SETERRQ(PetscObjectComm((PetscObject) mc), PETSC_ERR_SUP, "Nonsymmetric greedy coloring only works in serial");
}
/* create the vectors and communication structures if necessary */
no=0;
if (mo) {
ierr = VecGetLocalSize(aij->lvec,&no);CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)m),&sf);CHKERRQ(ierr);
ierr = MatGetLayouts(m,&layout,NULL);CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,layout,no,NULL,PETSC_COPY_VALUES,aij->garray);CHKERRQ(ierr);
}
ierr = MatColoringGetMaxColors(mc,&maxcolors);CHKERRQ(ierr);
masksize=n;
nbad=0;
badsize=n;
ierr = PetscMalloc1(masksize,&mask);CHKERRQ(ierr);
ierr = PetscMalloc4(n,&d1cols,n,&dcolors,n,&conf,n,&bad);CHKERRQ(ierr);
ierr = PetscMalloc2(badsize,&badidx,badsize,&badnext);CHKERRQ(ierr);
for(i=0;i<masksize;i++) {
mask[i]=-1;
}
for (i=0;i<n;i++) {
dcolors[i]=maxcolors;
bad[i]=-1;
}
for (i=0;i<badsize;i++) {
badnext[i]=-1;
}
if (mo) {
ierr = PetscMalloc3(no,&owts,no,&oconf,no,&ocolors);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_REAL,wts,owts);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_REAL,wts,owts);CHKERRQ(ierr);
for (i=0;i<no;i++) {
ocolors[i]=maxcolors;
}
} else { /* Appease overzealous -Wmaybe-initialized */
owts = NULL;
oconf = NULL;
ocolors = NULL;
}
mcol=0;
while (nd_global < n_global) {
nd=n;
/* assign lowest possible color to each local vertex */
mcol_global=0;
ierr = PetscLogEventBegin(MATCOLORING_Local,mc,0,0,0);CHKERRQ(ierr);
for (i=0;i<n;i++) {
idx=lperm[i];
if (dcolors[idx] == maxcolors) {
/* entries in bad */
cbad=bad[idx];
while (cbad>=0) {
ccol=badidx[cbad];
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
cbad=badnext[cbad];
}
/* diagonal distance-one rows */
nd1cols=0;
ncols = rmd_i[idx+1]-rmd_i[idx];
cidx = &(rmd_j[rmd_i[idx]]);
for (j=0;j<ncols;j++) {
d1cols[nd1cols] = cidx[j];
nd1cols++;
ccol=dcolors[cidx[j]];
if (ccol != maxcolors) {
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
/* off-diagonal distance-one rows */
if (mo) {
ncols = rmo_i[idx+1]-rmo_i[idx];
cidx = &(rmo_j[rmo_i[idx]]);
for (j=0;j<ncols;j++) {
ccol=ocolors[cidx[j]];
if (ccol != maxcolors) {
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
}
/* diagonal distance-two rows */
for (j=0;j<nd1cols;j++) {
ncols = md_i[d1cols[j]+1]-md_i[d1cols[j]];
cidx = &(md_j[md_i[d1cols[j]]]);
for (l=0;l<ncols;l++) {
ccol=dcolors[cidx[l]];
if (ccol != maxcolors) {
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
}
/* off-diagonal distance-two rows */
if (mo) {
for (j=0;j<nd1cols;j++) {
ncols = mo_i[d1cols[j]+1]-mo_i[d1cols[j]];
cidx = &(mo_j[mo_i[d1cols[j]]]);
for (l=0;l<ncols;l++) {
ccol=ocolors[cidx[l]];
if (ccol != maxcolors) {
if (ccol>=masksize) {
PetscInt *newmask;
ierr = PetscMalloc1(masksize*2,&newmask);CHKERRQ(ierr);
for(k=0;k<2*masksize;k++) {
newmask[k]=-1;
}
for(k=0;k<masksize;k++) {
newmask[k]=mask[k];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
mask=newmask;
masksize*=2;
}
mask[ccol]=idx;
}
}
}
}
/* assign this one the lowest color possible by seeing if there's a gap in the sequence of sorted neighbor colors */
for (j=0;j<masksize;j++) {
if (mask[j]!=idx) {
break;
}
}
pcol=j;
if (pcol>maxcolors) pcol=maxcolors;
dcolors[idx]=pcol;
if (pcol>mcol) mcol=pcol;
}
}
ierr = PetscLogEventEnd(MATCOLORING_Local,mc,0,0,0);CHKERRQ(ierr);
if (mo) {
/* transfer neighbor colors */
ierr = PetscSFBcastBegin(sf,MPIU_INT,dcolors,ocolors);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,dcolors,ocolors);CHKERRQ(ierr);
/* find the maximum color assigned locally and allocate a mask */
ierr = MPIU_Allreduce(&mcol,&mcol_global,1,MPIU_INT,MPI_MAX,PetscObjectComm((PetscObject)mc));CHKERRQ(ierr);
ierr = PetscMalloc1(mcol_global+1,&colorweights);CHKERRQ(ierr);
/* check for conflicts */
for (i=0;i<n;i++) {
conf[i]=PETSC_FALSE;
}
for (i=0;i<no;i++) {
oconf[i]=PETSC_FALSE;
}
for (i=0;i<n;i++) {
ncols = mo_i[i+1]-mo_i[i];
cidx = &(mo_j[mo_i[i]]);
if (ncols > 0) {
/* fill in the mask */
for (j=0;j<mcol_global+1;j++) {
colorweights[j]=0;
}
colorweights[dcolors[i]]=wts[i];
/* fill in the off-diagonal part of the mask */
for (j=0;j<ncols;j++) {
ccol=ocolors[cidx[j]];
if (ccol < maxcolors) {
if (colorweights[ccol] < owts[cidx[j]]) {
colorweights[ccol] = owts[cidx[j]];
}
}
}
/* fill in the on-diagonal part of the mask */
ncols = md_i[i+1]-md_i[i];
cidx = &(md_j[md_i[i]]);
for (j=0;j<ncols;j++) {
ccol=dcolors[cidx[j]];
if (ccol < maxcolors) {
if (colorweights[ccol] < wts[cidx[j]]) {
colorweights[ccol] = wts[cidx[j]];
}
}
}
/* go back through and set up on and off-diagonal conflict vectors */
ncols = md_i[i+1]-md_i[i];
cidx = &(md_j[md_i[i]]);
for (j=0;j<ncols;j++) {
ccol=dcolors[cidx[j]];
if (ccol < maxcolors) {
if (colorweights[ccol] > wts[cidx[j]]) {
conf[cidx[j]]=PETSC_TRUE;
}
}
}
ncols = mo_i[i+1]-mo_i[i];
cidx = &(mo_j[mo_i[i]]);
for (j=0;j<ncols;j++) {
ccol=ocolors[cidx[j]];
if (ccol < maxcolors) {
if (colorweights[ccol] > owts[cidx[j]]) {
oconf[cidx[j]]=PETSC_TRUE;
}
}
}
}
}
nd_global=0;
ierr = PetscFree(colorweights);CHKERRQ(ierr);
ierr = PetscLogEventBegin(MATCOLORING_Comm,mc,0,0,0);CHKERRQ(ierr);
ierr = PetscSFReduceBegin(sf,MPIU_INT,oconf,conf,MPIU_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd(sf,MPIU_INT,oconf,conf,MPIU_SUM);CHKERRQ(ierr);
ierr = PetscLogEventEnd(MATCOLORING_Comm,mc,0,0,0);CHKERRQ(ierr);
/* go through and unset local colors that have conflicts */
for (i=0;i<n;i++) {
if (conf[i]>0) {
/* push this color onto the bad stack */
badidx[nbad]=dcolors[i];
badnext[nbad]=bad[i];
bad[i]=nbad;
nbad++;
if (nbad>=badsize) {
PetscInt *newbadnext;
ISColoringValue *newbadidx;
ierr = PetscMalloc2(badsize*2,&newbadidx,badsize*2,&newbadnext);CHKERRQ(ierr);
for(k=0;k<2*badsize;k++) {
newbadnext[k]=-1;
}
for(k=0;k<badsize;k++) {
newbadidx[k]=badidx[k];
newbadnext[k]=badnext[k];
}
ierr = PetscFree2(badidx,badnext);CHKERRQ(ierr);
badidx=newbadidx;
badnext=newbadnext;
badsize*=2;
}
dcolors[i] = maxcolors;
nd--;
}
}
}
ierr = MPIU_Allreduce(&nd,&nd_global,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)mc));CHKERRQ(ierr);
}
if (mo) {
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFree3(owts,oconf,ocolors);CHKERRQ(ierr);
}
for (i=0;i<n;i++) {
colors[i]=dcolors[i];
}
ierr = PetscFree(mask);CHKERRQ(ierr);
ierr = PetscFree4(d1cols,dcolors,conf,bad);CHKERRQ(ierr);
ierr = PetscFree2(badidx,badnext);CHKERRQ(ierr);
if (!gr->symmetric) {ierr = MatDestroy(&mt);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);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
Vec B,Bout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscScalar *bufB;
PetscScalar *bufBout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 2) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for two MPI processes\n");
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 2;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
if (rank == 0) {
iremote[0].rank = 0;
iremote[0].index = 0;
iremote[1].rank = 1;
iremote[1].index = 0;
} else {
iremote[0].rank = 1;
iremote[0].index = 0;
iremote[1].rank = 0;
iremote[1].index = 0;
}
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,2,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&B);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Bout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArray(B,&bufB);CHKERRQ(ierr);
for (i=0; i<2; i++) {
bufA[i] = (PetscScalar)rank;
bufB[i] = (PetscScalar)(rank) + 10.0;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(B,&bufB);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecGetArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufA,(void *)bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_SCALAR,(const void*)bufB,(void *)bufBout);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArrayRead(B,(const PetscScalar**)&bufB);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecRestoreArray(Bout,&bufBout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecView(Bout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&B);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = VecDestroy(&Bout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
PetscErrorCode MatSetLocalToGlobalMapping_IS(Mat A,ISLocalToGlobalMapping rmapping,ISLocalToGlobalMapping cmapping)
{
PetscErrorCode ierr;
PetscInt nr,rbs,nc,cbs;
Mat_IS *is = (Mat_IS*)A->data;
IS from,to;
Vec cglobal,rglobal;
PetscFunctionBegin;
PetscCheckSameComm(A,1,rmapping,2);
PetscCheckSameComm(A,1,cmapping,3);
/* Destroy any previous data */
ierr = VecDestroy(&is->x);CHKERRQ(ierr);
ierr = VecDestroy(&is->y);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->rctx);CHKERRQ(ierr);
ierr = VecScatterDestroy(&is->cctx);CHKERRQ(ierr);
ierr = MatDestroy(&is->A);CHKERRQ(ierr);
ierr = PetscSFDestroy(&is->sf);CHKERRQ(ierr);
ierr = PetscFree2(is->sf_rootdata,is->sf_leafdata);CHKERRQ(ierr);
/* Setup Layout and set local to global maps */
ierr = PetscLayoutSetUp(A->rmap);CHKERRQ(ierr);
ierr = PetscLayoutSetUp(A->cmap);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->rmap,rmapping);CHKERRQ(ierr);
ierr = PetscLayoutSetISLocalToGlobalMapping(A->cmap,cmapping);CHKERRQ(ierr);
/* Create the local matrix A */
ierr = ISLocalToGlobalMappingGetSize(rmapping,&nr);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetBlockSize(rmapping,&rbs);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetSize(cmapping,&nc);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingGetBlockSize(cmapping,&cbs);CHKERRQ(ierr);
ierr = MatCreate(PETSC_COMM_SELF,&is->A);CHKERRQ(ierr);
ierr = MatSetType(is->A,MATAIJ);CHKERRQ(ierr);
ierr = MatSetSizes(is->A,nr,nc,nr,nc);CHKERRQ(ierr);
ierr = MatSetBlockSizes(is->A,rbs,cbs);CHKERRQ(ierr);
ierr = MatSetOptionsPrefix(is->A,((PetscObject)A)->prefix);CHKERRQ(ierr);
ierr = MatAppendOptionsPrefix(is->A,"is_");CHKERRQ(ierr);
ierr = MatSetFromOptions(is->A);CHKERRQ(ierr);
/* Create the local work vectors */
ierr = MatCreateVecs(is->A,&is->x,&is->y);CHKERRQ(ierr);
/* setup the global to local scatters */
ierr = MatCreateVecs(A,&cglobal,&rglobal);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,nr,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(rmapping,to,&from);CHKERRQ(ierr);
ierr = VecScatterCreate(rglobal,from,is->y,to,&is->rctx);CHKERRQ(ierr);
if (rmapping != cmapping) {
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,nc,0,1,&to);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingApplyIS(cmapping,to,&from);CHKERRQ(ierr);
ierr = VecScatterCreate(cglobal,from,is->x,to,&is->cctx);CHKERRQ(ierr);
} else {
ierr = PetscObjectReference((PetscObject)is->rctx);CHKERRQ(ierr);
is->cctx = is->rctx;
}
ierr = VecDestroy(&rglobal);CHKERRQ(ierr);
ierr = VecDestroy(&cglobal);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = ISDestroy(&from);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);
}
PetscErrorCode maxIndSetAgg(IS perm,Mat Gmat,PetscBool strict_aggs,PetscCoarsenData **a_locals_llist)
{
PetscErrorCode ierr;
Mat_SeqAIJ *matA,*matB=NULL;
Mat_MPIAIJ *mpimat=NULL;
MPI_Comm comm;
PetscInt num_fine_ghosts,kk,n,ix,j,*idx,*ii,iter,Iend,my0,nremoved,gid,lid,cpid,lidj,sgid,t1,t2,slid,nDone,nselected=0,state,statej;
PetscInt *cpcol_gid,*cpcol_state,*lid_cprowID,*lid_gid,*cpcol_sel_gid,*icpcol_gid,*lid_state,*lid_parent_gid=NULL;
PetscBool *lid_removed;
PetscBool isMPI,isAIJ,isOK;
const PetscInt *perm_ix;
const PetscInt nloc = Gmat->rmap->n;
PetscCoarsenData *agg_lists;
PetscLayout layout;
PetscSF sf;
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)Gmat,&comm);
CHKERRQ(ierr);
/* get submatrices */
ierr = PetscObjectTypeCompare((PetscObject)Gmat,MATMPIAIJ,&isMPI);
CHKERRQ(ierr);
if (isMPI) {
mpimat = (Mat_MPIAIJ*)Gmat->data;
matA = (Mat_SeqAIJ*)mpimat->A->data;
matB = (Mat_SeqAIJ*)mpimat->B->data;
/* force compressed storage of B */
ierr = MatCheckCompressedRow(mpimat->B,matB->nonzerorowcnt,&matB->compressedrow,matB->i,Gmat->rmap->n,-1.0);
CHKERRQ(ierr);
} else {
ierr = PetscObjectTypeCompare((PetscObject)Gmat,MATSEQAIJ,&isAIJ);
CHKERRQ(ierr);
matA = (Mat_SeqAIJ*)Gmat->data;
}
ierr = MatGetOwnershipRange(Gmat,&my0,&Iend);
CHKERRQ(ierr);
ierr = PetscMalloc1(nloc,&lid_gid);
CHKERRQ(ierr); /* explicit array needed */
if (mpimat) {
for (kk=0,gid=my0; kk<nloc; kk++,gid++) {
lid_gid[kk] = gid;
}
ierr = VecGetLocalSize(mpimat->lvec, &num_fine_ghosts);
CHKERRQ(ierr);
ierr = PetscMalloc1(num_fine_ghosts,&cpcol_gid);
CHKERRQ(ierr);
ierr = PetscMalloc1(num_fine_ghosts,&cpcol_state);
CHKERRQ(ierr);
ierr = PetscSFCreate(PetscObjectComm((PetscObject)Gmat),&sf);
CHKERRQ(ierr);
ierr = MatGetLayouts(Gmat,&layout,NULL);
CHKERRQ(ierr);
ierr = PetscSFSetGraphLayout(sf,layout,num_fine_ghosts,NULL,PETSC_COPY_VALUES,mpimat->garray);
CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,lid_gid,cpcol_gid);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,lid_gid,cpcol_gid);
CHKERRQ(ierr);
for (kk=0; kk<num_fine_ghosts; kk++) {
cpcol_state[kk]=MIS_NOT_DONE;
}
} else num_fine_ghosts = 0;
ierr = PetscMalloc1(nloc, &lid_cprowID);
CHKERRQ(ierr);
ierr = PetscMalloc1(nloc, &lid_removed);
CHKERRQ(ierr); /* explicit array needed */
if (strict_aggs) {
ierr = PetscMalloc1(nloc,&lid_parent_gid);
CHKERRQ(ierr);
}
ierr = PetscMalloc1(nloc,&lid_state);
CHKERRQ(ierr);
/* has ghost nodes for !strict and uses local indexing (yuck) */
ierr = PetscCDCreate(strict_aggs ? nloc : num_fine_ghosts+nloc, &agg_lists);
CHKERRQ(ierr);
if (a_locals_llist) *a_locals_llist = agg_lists;
/* need an inverse map - locals */
for (kk=0; kk<nloc; kk++) {
lid_cprowID[kk] = -1;
lid_removed[kk] = PETSC_FALSE;
if (strict_aggs) {
lid_parent_gid[kk] = -1.0;
}
lid_state[kk] = MIS_NOT_DONE;
}
/* set index into cmpressed row 'lid_cprowID' */
if (matB) {
for (ix=0; ix<matB->compressedrow.nrows; ix++) {
lid = matB->compressedrow.rindex[ix];
lid_cprowID[lid] = ix;
}
}
/* MIS */
iter = nremoved = nDone = 0;
ierr = ISGetIndices(perm, &perm_ix);
CHKERRQ(ierr);
while (nDone < nloc || PETSC_TRUE) { /* asyncronous not implemented */
iter++;
/* check all vertices */
for (kk=0; kk<nloc; kk++) {
lid = perm_ix[kk];
state = lid_state[lid];
if (lid_removed[lid]) continue;
if (state == MIS_NOT_DONE) {
/* parallel test, delete if selected ghost */
isOK = PETSC_TRUE;
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i;
n = ii[ix+1] - ii[ix];
idx = matB->j + ii[ix];
for (j=0; j<n; j++) {
cpid = idx[j]; /* compressed row ID in B mat */
gid = cpcol_gid[cpid];
statej = cpcol_state[cpid];
if (statej == MIS_NOT_DONE && gid >= Iend) { /* should be (pe>rank), use gid as pe proxy */
isOK = PETSC_FALSE; /* can not delete */
break;
}
}
} /* parallel test */
if (isOK) { /* select or remove this vertex */
nDone++;
/* check for singleton */
ii = matA->i;
n = ii[lid+1] - ii[lid];
if (n < 2) {
/* if I have any ghost adj then not a sing */
ix = lid_cprowID[lid];
if (ix==-1 || (matB->compressedrow.i[ix+1]-matB->compressedrow.i[ix])==0) {
nremoved++;
lid_removed[lid] = PETSC_TRUE;
/* should select this because it is technically in the MIS but lets not */
continue; /* one local adj (me) and no ghost - singleton */
}
}
/* SELECTED state encoded with global index */
lid_state[lid] = lid+my0; /* needed???? */
nselected++;
if (strict_aggs) {
ierr = PetscCDAppendID(agg_lists, lid, lid+my0);
CHKERRQ(ierr);
} else {
ierr = PetscCDAppendID(agg_lists, lid, lid);
CHKERRQ(ierr);
}
/* delete local adj */
idx = matA->j + ii[lid];
for (j=0; j<n; j++) {
lidj = idx[j];
statej = lid_state[lidj];
if (statej == MIS_NOT_DONE) {
nDone++;
if (strict_aggs) {
ierr = PetscCDAppendID(agg_lists, lid, lidj+my0);
CHKERRQ(ierr);
} else {
ierr = PetscCDAppendID(agg_lists, lid, lidj);
CHKERRQ(ierr);
}
lid_state[lidj] = MIS_DELETED; /* delete this */
}
}
/* delete ghost adj of lid - deleted ghost done later for strict_aggs */
if (!strict_aggs) {
if ((ix=lid_cprowID[lid]) != -1) { /* if I have any ghost neighbors */
ii = matB->compressedrow.i;
n = ii[ix+1] - ii[ix];
idx = matB->j + ii[ix];
for (j=0; j<n; j++) {
cpid = idx[j]; /* compressed row ID in B mat */
statej = cpcol_state[cpid];
if (statej == MIS_NOT_DONE) {
ierr = PetscCDAppendID(agg_lists, lid, nloc+cpid);
CHKERRQ(ierr);
}
}
}
}
} /* selected */
} /* not done vertex */
} /* vertex loop */
/* update ghost states and count todos */
if (mpimat) {
/* scatter states, check for done */
ierr = PetscSFBcastBegin(sf,MPIU_INT,lid_state,cpcol_state);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,lid_state,cpcol_state);
CHKERRQ(ierr);
ii = matB->compressedrow.i;
for (ix=0; ix<matB->compressedrow.nrows; ix++) {
lid = matB->compressedrow.rindex[ix]; /* local boundary node */
state = lid_state[lid];
if (state == MIS_NOT_DONE) {
/* look at ghosts */
n = ii[ix+1] - ii[ix];
idx = matB->j + ii[ix];
for (j=0; j<n; j++) {
cpid = idx[j]; /* compressed row ID in B mat */
statej = cpcol_state[cpid];
if (MIS_IS_SELECTED(statej)) { /* lid is now deleted, do it */
nDone++;
lid_state[lid] = MIS_DELETED; /* delete this */
if (!strict_aggs) {
lidj = nloc + cpid;
ierr = PetscCDAppendID(agg_lists, lidj, lid);
CHKERRQ(ierr);
} else {
sgid = cpcol_gid[cpid];
lid_parent_gid[lid] = sgid; /* keep track of proc that I belong to */
}
break;
}
}
}
}
/* all done? */
t1 = nloc - nDone;
ierr = MPI_Allreduce(&t1, &t2, 1, MPIU_INT, MPI_SUM, comm);
CHKERRQ(ierr); /* synchronous version */
if (t2 == 0) break;
} else break; /* all done */
} /* outer parallel MIS loop */
ierr = ISRestoreIndices(perm,&perm_ix);
CHKERRQ(ierr);
ierr = PetscInfo3(Gmat,"\t removed %D of %D vertices. %D selected.\n",nremoved,nloc,nselected);
CHKERRQ(ierr);
/* tell adj who my lid_parent_gid vertices belong to - fill in agg_lists selected ghost lists */
if (strict_aggs && matB) {
/* need to copy this to free buffer -- should do this globaly */
ierr = PetscMalloc1(num_fine_ghosts, &cpcol_sel_gid);
CHKERRQ(ierr);
ierr = PetscMalloc1(num_fine_ghosts, &icpcol_gid);
CHKERRQ(ierr);
for (cpid=0; cpid<num_fine_ghosts; cpid++) icpcol_gid[cpid] = cpcol_gid[cpid];
/* get proc of deleted ghost */
ierr = PetscSFBcastBegin(sf,MPIU_INT,lid_parent_gid,cpcol_sel_gid);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,MPIU_INT,lid_parent_gid,cpcol_sel_gid);
CHKERRQ(ierr);
for (cpid=0; cpid<num_fine_ghosts; cpid++) {
sgid = cpcol_sel_gid[cpid];
gid = icpcol_gid[cpid];
if (sgid >= my0 && sgid < Iend) { /* I own this deleted */
slid = sgid - my0;
ierr = PetscCDAppendID(agg_lists, slid, gid);
CHKERRQ(ierr);
}
}
ierr = PetscFree(icpcol_gid);
CHKERRQ(ierr);
ierr = PetscFree(cpcol_sel_gid);
CHKERRQ(ierr);
}
if (mpimat) {
ierr = PetscSFDestroy(&sf);
CHKERRQ(ierr);
ierr = PetscFree(cpcol_gid);
CHKERRQ(ierr);
ierr = PetscFree(cpcol_state);
CHKERRQ(ierr);
}
ierr = PetscFree(lid_cprowID);
CHKERRQ(ierr);
ierr = PetscFree(lid_gid);
CHKERRQ(ierr);
ierr = PetscFree(lid_removed);
CHKERRQ(ierr);
if (strict_aggs) {
ierr = PetscFree(lid_parent_gid);
CHKERRQ(ierr);
}
ierr = PetscFree(lid_state);
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);
}
/* Distribute cones
- Partitioning: input partition point map and naive sf, output sf with inverse of map, distribute points
- Distribute section: input current sf, communicate sizes and offsets, output local section and offsets (only use for new sf)
- Create SF for values: input current sf and offsets, output new sf
- Distribute values: input new sf, communicate values
*/
PetscErrorCode DistributeMesh(DM dm, AppCtx *user, PetscSF *pointSF, DM *parallelDM)
{
MPI_Comm comm = ((PetscObject) dm)->comm;
const PetscInt height = 0;
PetscInt dim, numRemoteRanks;
IS cellPart, part;
PetscSection cellPartSection, partSection;
PetscSFNode *remoteRanks;
PetscSF partSF;
ISLocalToGlobalMapping renumbering;
PetscSF coneSF;
PetscSection originalConeSection, newConeSection;
PetscInt *remoteOffsets, newConesSize;
PetscInt *cones, *newCones;
PetscMPIInt numProcs, rank, p;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = DMMeshGetDimension(dm, &dim);CHKERRQ(ierr);
/* Create cell partition - We need to rewrite to use IS, use the MatPartition stuff */
ierr = DMMeshCreatePartition(dm, &cellPartSection, &cellPart, height);CHKERRQ(ierr);
/* Create SF assuming a serial partition for all processes: Could check for IS length here */
if (!rank) {
numRemoteRanks = numProcs;
} else {
numRemoteRanks = 0;
}
ierr = PetscMalloc(numRemoteRanks * sizeof(PetscSFNode), &remoteRanks);CHKERRQ(ierr);
for(p = 0; p < numRemoteRanks; ++p) {
remoteRanks[p].rank = p;
remoteRanks[p].index = 0;
}
ierr = PetscSFCreate(comm, &partSF);CHKERRQ(ierr);
ierr = PetscSFSetGraph(partSF, 1, numRemoteRanks, PETSC_NULL, PETSC_OWN_POINTER, remoteRanks, PETSC_OWN_POINTER);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Cell Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(cellPartSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(cellPart, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(partSF, PETSC_NULL);CHKERRQ(ierr);
/* Close the partition over the mesh */
ierr = DMMeshCreatePartitionClosure(dm, cellPartSection, cellPart, &partSection, &part);CHKERRQ(ierr);
ierr = ISDestroy(&cellPart);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&cellPartSection);CHKERRQ(ierr);
/* Create new mesh */
ierr = DMMeshCreate(comm, parallelDM);CHKERRQ(ierr);
ierr = DMMeshSetDimension(*parallelDM, dim);CHKERRQ(ierr);
ierr = PetscObjectSetName((PetscObject) *parallelDM, "Parallel Mesh");CHKERRQ(ierr);
/* Distribute sieve points and the global point numbering (replaces creating remote bases) */
ierr = PetscSFConvertPartition(partSF, partSection, part, &renumbering, pointSF);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Point Partition:\n");CHKERRQ(ierr);
ierr = PetscSectionView(partSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = ISView(part, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFView(*pointSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Point Renumbering after partition:\n");CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingView(renumbering, PETSC_NULL);CHKERRQ(ierr);
/* Cleanup */
ierr = PetscSFDestroy(&partSF);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&partSection);CHKERRQ(ierr);
ierr = ISDestroy(&part);CHKERRQ(ierr);
/* Distribute cone section */
ierr = DMMeshGetConeSection(dm, &originalConeSection);CHKERRQ(ierr);
ierr = DMMeshGetConeSection(*parallelDM, &newConeSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(*pointSF, originalConeSection, &remoteOffsets, newConeSection);CHKERRQ(ierr);
ierr = DMMeshSetUp(*parallelDM);CHKERRQ(ierr);
/* Communicate and renumber cones */
ierr = PetscSFCreateSectionSF(*pointSF, originalConeSection, remoteOffsets, newConeSection, &coneSF);CHKERRQ(ierr);
ierr = DMMeshGetCones(dm, &cones);CHKERRQ(ierr);
ierr = DMMeshGetCones(*parallelDM, &newCones);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(coneSF, MPIU_INT, cones, newCones);CHKERRQ(ierr);
ierr = PetscSectionGetStorageSize(newConeSection, &newConesSize);CHKERRQ(ierr);
ierr = ISGlobalToLocalMappingApply(renumbering, IS_GTOLM_MASK, newConesSize, newCones, PETSC_NULL, newCones);CHKERRQ(ierr);
ierr = ISLocalToGlobalMappingDestroy(&renumbering);CHKERRQ(ierr);
/* Debugging */
ierr = PetscPrintf(comm, "Serial Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(originalConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscPrintf(comm, "Parallel Cone Section:\n");CHKERRQ(ierr);
ierr = PetscSectionView(newConeSection, PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = PetscSFView(coneSF, PETSC_NULL);CHKERRQ(ierr);
ierr = PetscSFDestroy(&coneSF);CHKERRQ(ierr);
/* Create supports and stratify sieve */
ierr = DMMeshSymmetrize(*parallelDM);CHKERRQ(ierr);
ierr = DMMeshStratify(*parallelDM);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
int main(int argc, char **argv)
{
PetscInt ierr;
PetscSF sf;
Vec A,Aout;
PetscScalar *bufA;
PetscScalar *bufAout;
PetscMPIInt rank, size;
PetscInt nroots, nleaves;
PetscInt i;
PetscInt *ilocal;
PetscSFNode *iremote;
PetscBool test_dupped_type;
MPI_Datatype contig;
ierr = PetscInitialize(&argc,&argv,NULL,help);if (ierr) return ierr;
ierr = MPI_Comm_rank(PETSC_COMM_WORLD,&rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(PETSC_COMM_WORLD,&size);CHKERRQ(ierr);
if (size != 1) SETERRQ(PETSC_COMM_WORLD, PETSC_ERR_USER, "Only coded for one MPI process");
ierr = PetscOptionsBegin(PETSC_COMM_WORLD,"","PetscSF type freeing options","none");CHKERRQ(ierr);
test_dupped_type = PETSC_FALSE;
ierr = PetscOptionsBool("-test_dupped_type", "Test dupped input type","",test_dupped_type,&test_dupped_type,NULL);CHKERRQ(ierr);
ierr = PetscOptionsEnd();CHKERRQ(ierr);
ierr = PetscSFCreate(PETSC_COMM_WORLD,&sf);CHKERRQ(ierr);
ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr);
nleaves = 1;
nroots = 1;
ierr = PetscMalloc1(nleaves,&ilocal);CHKERRQ(ierr);
for (i = 0; i<nleaves; i++) {
ilocal[i] = i;
}
ierr = PetscMalloc1(nleaves,&iremote);CHKERRQ(ierr);
iremote[0].rank = 0;
iremote[0].index = 0;
ierr = PetscSFSetGraph(sf,nroots,nleaves,ilocal,PETSC_OWN_POINTER,iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
ierr = PetscSFView(sf,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecCreate(PETSC_COMM_WORLD,&A);CHKERRQ(ierr);
ierr = VecSetSizes(A,4,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = VecSetFromOptions(A);CHKERRQ(ierr);
ierr = VecSetUp(A);CHKERRQ(ierr);
ierr = VecDuplicate(A,&Aout);CHKERRQ(ierr);
ierr = VecGetArray(A,&bufA);CHKERRQ(ierr);
for (i=0; i<4; i++) {
bufA[i] = (PetscScalar)i;
}
ierr = VecRestoreArray(A,&bufA);CHKERRQ(ierr);
ierr = VecGetArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecGetArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = MPI_Type_contiguous(4, MPIU_SCALAR, &contig);CHKERRQ(ierr);
ierr = MPI_Type_commit(&contig);CHKERRQ(ierr);
if (test_dupped_type) {
MPI_Datatype tmp;
ierr = MPI_Type_dup(contig, &tmp);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
contig = tmp;
}
for (i=0;i<10000;i++) {
ierr = PetscSFBcastBegin(sf,contig,bufA,bufAout);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf,contig,bufA,bufAout);CHKERRQ(ierr);
}
ierr = VecRestoreArrayRead(A,(const PetscScalar**)&bufA);CHKERRQ(ierr);
ierr = VecRestoreArray(Aout,&bufAout);CHKERRQ(ierr);
ierr = VecView(Aout,PETSC_VIEWER_STDOUT_WORLD);CHKERRQ(ierr);
ierr = VecDestroy(&A);CHKERRQ(ierr);
ierr = VecDestroy(&Aout);CHKERRQ(ierr);
ierr = PetscSFDestroy(&sf);CHKERRQ(ierr);
ierr = MPI_Type_free(&contig);CHKERRQ(ierr);
ierr = PetscFinalize();
return ierr;
}
PetscErrorCode DMLabelDistribute(DMLabel label, PetscSF sf, DMLabel *labelNew)
{
MPI_Comm comm;
PetscSection rootSection, leafSection;
PetscSF labelSF;
PetscInt p, pStart, pEnd, l, lStart, lEnd, s, nroots, nleaves, size, dof, offset, stratum;
PetscInt *remoteOffsets, *rootStrata, *rootIdx, *leafStrata, *strataIdx;
char *name;
PetscInt nameSize;
size_t len = 0;
PetscMPIInt rank, numProcs;
PetscErrorCode ierr;
PetscFunctionBegin;
if (label) {ierr = DMLabelMakeAllValid_Private(label);CHKERRQ(ierr);}
ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm, &rank);CHKERRQ(ierr);
ierr = MPI_Comm_size(comm, &numProcs);CHKERRQ(ierr);
/* Bcast name */
if (!rank) {ierr = PetscStrlen(label->name, &len);CHKERRQ(ierr);}
nameSize = len;
ierr = MPI_Bcast(&nameSize, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = PetscMalloc1(nameSize+1, &name);CHKERRQ(ierr);
if (!rank) {ierr = PetscMemcpy(name, label->name, nameSize+1);CHKERRQ(ierr);}
ierr = MPI_Bcast(name, nameSize+1, MPI_CHAR, 0, comm);CHKERRQ(ierr);
ierr = DMLabelCreate(name, labelNew);CHKERRQ(ierr);
ierr = PetscFree(name);CHKERRQ(ierr);
/* Bcast numStrata */
if (!rank) (*labelNew)->numStrata = label->numStrata;
ierr = MPI_Bcast(&(*labelNew)->numStrata, 1, MPIU_INT, 0, comm);CHKERRQ(ierr);
/* Bcast stratumValues */
ierr = PetscMalloc1((*labelNew)->numStrata, &(*labelNew)->stratumValues);CHKERRQ(ierr);
if (!rank) {ierr = PetscMemcpy((*labelNew)->stratumValues, label->stratumValues, label->numStrata * sizeof(PetscInt));CHKERRQ(ierr);}
ierr = MPI_Bcast((*labelNew)->stratumValues, (*labelNew)->numStrata, MPIU_INT, 0, comm);CHKERRQ(ierr);
ierr = PetscMalloc1((*labelNew)->numStrata, &(*labelNew)->arrayValid);CHKERRQ(ierr);
for (s = 0; s < (*labelNew)->numStrata; ++s) (*labelNew)->arrayValid[s] = PETSC_TRUE;
/* Build a section detailing strata-per-point, distribute and build SF
from that and then send our points. */
ierr = PetscSFGetGraph(sf, &nroots, &nleaves, NULL, NULL);CHKERRQ(ierr);
ierr = PetscSectionCreate(comm, &rootSection);CHKERRQ(ierr);
ierr = PetscSectionSetChart(rootSection, 0, nroots);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
lStart = 0;
lEnd = label->stratumSizes[s];
for (l=lStart; l<lEnd; l++) {
ierr = PetscSectionGetDof(rootSection, label->points[s][l], &dof);CHKERRQ(ierr);
ierr = PetscSectionSetDof(rootSection, label->points[s][l], dof+1);CHKERRQ(ierr);
}
}
}
ierr = PetscSectionSetUp(rootSection);CHKERRQ(ierr);
/* Create a point-wise array of point strata */
ierr = PetscSectionGetStorageSize(rootSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &rootStrata);CHKERRQ(ierr);
ierr = PetscCalloc1(nroots, &rootIdx);CHKERRQ(ierr);
if (label) {
for (s = 0; s < label->numStrata; ++s) {
lStart = 0;
lEnd = label->stratumSizes[s];
for (l=lStart; l<lEnd; l++) {
p = label->points[s][l];
ierr = PetscSectionGetOffset(rootSection, p, &offset);CHKERRQ(ierr);
rootStrata[offset+rootIdx[p]++] = s;
}
}
}
/* Build SF that maps label points to remote processes */
ierr = PetscSectionCreate(comm, &leafSection);CHKERRQ(ierr);
ierr = PetscSFDistributeSection(sf, rootSection, &remoteOffsets, leafSection);CHKERRQ(ierr);
ierr = PetscSFCreateSectionSF(sf, rootSection, remoteOffsets, leafSection, &labelSF);CHKERRQ(ierr);
/* Send the strata for each point over the derived SF */
ierr = PetscSectionGetStorageSize(leafSection, &size);CHKERRQ(ierr);
ierr = PetscMalloc1(size, &leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(labelSF, MPIU_INT, rootStrata, leafStrata);CHKERRQ(ierr);
ierr = PetscSFBcastEnd(labelSF, MPIU_INT, rootStrata, leafStrata);CHKERRQ(ierr);
/* Rebuild the point strata on the receiver */
ierr = PetscCalloc1((*labelNew)->numStrata,&(*labelNew)->stratumSizes);CHKERRQ(ierr);
ierr = PetscSectionGetChart(leafSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p=pStart; p<pEnd; p++) {
ierr = PetscSectionGetDof(leafSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(leafSection, p, &offset);CHKERRQ(ierr);
for (s=0; s<dof; s++) {
(*labelNew)->stratumSizes[leafStrata[offset+s]]++;
}
}
ierr = PetscCalloc1((*labelNew)->numStrata,&(*labelNew)->ht);CHKERRQ(ierr);
ierr = PetscMalloc1((*labelNew)->numStrata,&(*labelNew)->points);CHKERRQ(ierr);
for (s = 0; s < (*labelNew)->numStrata; ++s) {
PetscHashICreate((*labelNew)->ht[s]);
ierr = PetscMalloc1((*labelNew)->stratumSizes[s], &(*labelNew)->points[s]);CHKERRQ(ierr);
}
/* Insert points into new strata */
ierr = PetscCalloc1((*labelNew)->numStrata, &strataIdx);CHKERRQ(ierr);
ierr = PetscSectionGetChart(leafSection, &pStart, &pEnd);CHKERRQ(ierr);
for (p=pStart; p<pEnd; p++) {
ierr = PetscSectionGetDof(leafSection, p, &dof);CHKERRQ(ierr);
ierr = PetscSectionGetOffset(leafSection, p, &offset);CHKERRQ(ierr);
for (s=0; s<dof; s++) {
stratum = leafStrata[offset+s];
(*labelNew)->points[stratum][strataIdx[stratum]++] = p;
}
}
ierr = PetscFree(rootStrata);CHKERRQ(ierr);
ierr = PetscFree(leafStrata);CHKERRQ(ierr);
ierr = PetscFree(rootIdx);CHKERRQ(ierr);
ierr = PetscFree(strataIdx);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&rootSection);CHKERRQ(ierr);
ierr = PetscSectionDestroy(&leafSection);CHKERRQ(ierr);
ierr = PetscSFDestroy(&labelSF);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
