/*@C
PetscSFWindowGetDataTypes - gets composite local and remote data types for each rank
Not Collective
Input Arguments:
+ sf - star forest
- unit - data type for each node
Output Arguments:
+ localtypes - types describing part of local leaf buffer referencing each remote rank
- remotetypes - types describing part of remote root buffer referenced for each remote rank
Level: developer
.seealso: PetscSFSetGraph(), PetscSFView()
@*/
static PetscErrorCode PetscSFWindowGetDataTypes(PetscSF sf,MPI_Datatype unit,const MPI_Datatype **localtypes,const MPI_Datatype **remotetypes)
{
PetscSF_Window *w = (PetscSF_Window*)sf->data;
PetscErrorCode ierr;
PetscSFDataLink link;
PetscInt i,nranks;
const PetscInt *roffset,*rmine,*rremote;
const PetscMPIInt *ranks;
PetscFunctionBegin;
/* Look for types in cache */
for (link=w->link; link; link=link->next) {
PetscBool match;
ierr = MPIPetsc_Type_compare(unit,link->unit,&match);CHKERRQ(ierr);
if (match) {
*localtypes = link->mine;
*remotetypes = link->remote;
PetscFunctionReturn(0);
}
}
/* Create new composite types for each send rank */
ierr = PetscSFGetRanks(sf,&nranks,&ranks,&roffset,&rmine,&rremote);CHKERRQ(ierr);
ierr = PetscMalloc(sizeof(*link),&link);CHKERRQ(ierr);
ierr = MPI_Type_dup(unit,&link->unit);CHKERRQ(ierr);
ierr = PetscMalloc2(nranks,&link->mine,nranks,&link->remote);CHKERRQ(ierr);
for (i=0; i<nranks; i++) {
PETSC_UNUSED PetscInt rcount = roffset[i+1] - roffset[i];
PetscMPIInt *rmine,*rremote;
#if !defined(PETSC_USE_64BIT_INDICES)
rmine = sf->rmine + sf->roffset[i];
rremote = sf->rremote + sf->roffset[i];
#else
PetscInt j;
ierr = PetscMalloc2(rcount,&rmine,rcount,&rremote);CHKERRQ(ierr);
for (j=0; j<rcount; j++) {
ierr = PetscMPIIntCast(sf->rmine[sf->roffset[i]+j],rmine+j);CHKERRQ(ierr);
ierr = PetscMPIIntCast(sf->rremote[sf->roffset[i]+j],rremote+j);CHKERRQ(ierr);
}
#endif
ierr = MPI_Type_create_indexed_block(rcount,1,rmine,link->unit,&link->mine[i]);CHKERRQ(ierr);
ierr = MPI_Type_create_indexed_block(rcount,1,rremote,link->unit,&link->remote[i]);CHKERRQ(ierr);
#if defined(PETSC_USE_64BIT_INDICES)
ierr = PetscFree2(rmine,rremote);CHKERRQ(ierr);
#endif
ierr = MPI_Type_commit(&link->mine[i]);CHKERRQ(ierr);
ierr = MPI_Type_commit(&link->remote[i]);CHKERRQ(ierr);
}
link->next = w->link;
w->link = link;
*localtypes = link->mine;
*remotetypes = link->remote;
PetscFunctionReturn(0);
}
PetscErrorCode DMPlexPreallocateOperator_2(DM dm, PetscInt bs, PetscSection section, PetscSection sectionGlobal, PetscInt dnz[], PetscInt onz[], PetscInt dnzu[], PetscInt onzu[], Mat A, PetscBool fillMatrix)
{
PetscInt *tmpClosure,*tmpAdj,*visits;
PetscInt c,cStart,cEnd,pStart,pEnd;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = DMGetDimension(dm, &dim);CHKERRQ(ierr);
ierr = DMPlexGetDepth(dm, &depth);CHKERRQ(ierr);
ierr = DMPlexGetMaxSizes(dm, &maxConeSize, &maxSupportSize);CHKERRQ(ierr);
maxClosureSize = 2*PetscMax(PetscPowInt(mesh->maxConeSize,depth+1),PetscPowInt(mesh->maxSupportSize,depth+1));
ierr = PetscSectionGetChart(section, &pStart, &pEnd);CHKERRQ(ierr);
npoints = pEnd - pStart;
ierr = PetscMalloc3(maxClosureSize,&tmpClosure,npoints,&lvisits,npoints,&visits);CHKERRQ(ierr);
ierr = PetscMemzero(lvisits,(pEnd-pStart)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = PetscMemzero(visits,(pEnd-pStart)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = DMPlexGetHeightStratum(dm, 0, &cStart, &cEnd);CHKERRQ(ierr);
for (c=cStart; c<cEnd; c++) {
PetscInt *support = tmpClosure;
ierr = DMPlexGetTransitiveClosure(dm, c, PETSC_FALSE, &supportSize, (PetscInt**)&support);CHKERRQ(ierr);
for (p=0; p<supportSize; p++) lvisits[support[p]]++;
}
ierr = PetscSFReduceBegin(sf,MPIU_INT,lvisits,visits,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPIU_INT,lvisits,visits,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf,MPIU_INT,visits,lvisits);CHKERRQ(ierr);
ierr = PetscSFBcastEnd (sf,MPIU_INT,visits,lvisits);CHKERRQ(ierr);
ierr = PetscSFGetRanks();CHKERRQ(ierr);
ierr = PetscMalloc2(maxClosureSize*maxClosureSize,&cellmat,npoints,&owner);CHKERRQ(ierr);
for (c=cStart; c<cEnd; c++) {
ierr = PetscMemzero(cellmat,maxClosureSize*maxClosureSize*sizeof(PetscInt));CHKERRQ(ierr);
/*
Depth-first walk of transitive closure.
At each leaf frame f of transitive closure that we see, add 1/visits[f] to each pair (p,q) not marked as done in cellmat.
This contribution is added to dnz if owning ranks of p and q match, to onz otherwise.
*/
}
ierr = PetscSFReduceBegin(sf,MPIU_INT,ldnz,dnz,MPI_SUM);CHKERRQ(ierr);
ierr = PetscSFReduceEnd (sf,MPIU_INT,lonz,onz,MPI_SUM);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
static PetscErrorCode PetscSFBcastBegin_Window(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata)
{
PetscSF_Window *w = (PetscSF_Window*)sf->data;
PetscErrorCode ierr;
PetscInt i,nranks;
const PetscMPIInt *ranks;
const MPI_Datatype *mine,*remote;
MPI_Win win;
PetscFunctionBegin;
ierr = PetscSFGetRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFWindowGetDataTypes(sf,unit,&mine,&remote);CHKERRQ(ierr);
ierr = PetscSFGetWindow(sf,unit,(void*)rootdata,PETSC_TRUE,MPI_MODE_NOPUT|MPI_MODE_NOPRECEDE,MPI_MODE_NOPUT,0,&win);CHKERRQ(ierr);
for (i=0; i<nranks; i++) {
if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_lock(MPI_LOCK_SHARED,ranks[i],MPI_MODE_NOCHECK,win);CHKERRQ(ierr);}
ierr = MPI_Get(leafdata,1,mine[i],ranks[i],0,1,remote[i],win);CHKERRQ(ierr);
if (w->sync == PETSCSF_WINDOW_SYNC_LOCK) {ierr = MPI_Win_unlock(ranks[i],win);CHKERRQ(ierr);}
}
PetscFunctionReturn(0);
}
static PetscErrorCode PetscSFFetchAndOpBegin_Window(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
{
PetscErrorCode ierr;
PetscInt i,nranks;
const PetscMPIInt *ranks;
const MPI_Datatype *mine,*remote;
MPI_Win win;
PetscFunctionBegin;
ierr = PetscSFGetRanks(sf,&nranks,&ranks,NULL,NULL,NULL);CHKERRQ(ierr);
ierr = PetscSFWindowGetDataTypes(sf,unit,&mine,&remote);CHKERRQ(ierr);
ierr = PetscSFWindowOpTranslate(&op);CHKERRQ(ierr);
ierr = PetscSFGetWindow(sf,unit,rootdata,PETSC_FALSE,0,0,0,&win);CHKERRQ(ierr);
for (i=0; i<sf->nranks; i++) {
ierr = MPI_Win_lock(MPI_LOCK_EXCLUSIVE,sf->ranks[i],0,win);CHKERRQ(ierr);
ierr = MPI_Get(leafupdate,1,mine[i],ranks[i],0,1,remote[i],win);CHKERRQ(ierr);
ierr = MPI_Accumulate((void*)leafdata,1,mine[i],ranks[i],0,1,remote[i],op,win);CHKERRQ(ierr);
ierr = MPI_Win_unlock(ranks[i],win);CHKERRQ(ierr);
}
PetscFunctionReturn(0);
}
PetscErrorCode PetscSFCreateSectionSF(PetscSF sf, PetscSection section, PetscSF *sectionSF)
{
PetscInt numRanks;
const PetscInt *ranks, *rankOffsets;
const PetscMPIInt *localPoints, *remotePoints;
PetscInt numPoints, numIndices = 0;
PetscInt *remoteOffsets;
PetscInt *localIndices;
PetscSFNode *remoteIndices;
PetscInt i, r, ind;
PetscErrorCode ierr;
PetscFunctionBegin;
ierr = PetscSFGetRanks(sf, &numRanks, &ranks, &rankOffsets, &localPoints, &remotePoints);
CHKERRQ(ierr);
numPoints = rankOffsets[numRanks];
for (i = 0; i < numPoints; ++i) {
PetscInt dof;
ierr = PetscSectionGetDof(section, localPoints[i], &dof);
CHKERRQ(ierr);
numIndices += dof;
}
/* Communicate offsets for ghosted points */
#if 0
PetscInt *localOffsets;
ierr = PetscMalloc2(numPoints,PetscInt,&localOffsets,numPoints,PetscInt,&remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSectionGetOffset(section, localPoints[i], &localOffsets[i]);
CHKERRQ(ierr);
}
ierr = PetscSFBcastBegin(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, localOffsets, remoteOffsets);
CHKERRQ(ierr);
for (i = 0; i < numPoints; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i]);
CHKERRQ(ierr);
}
#else
ierr = PetscMalloc((section->atlasLayout.pEnd - section->atlasLayout.pStart) * sizeof(PetscInt), &remoteOffsets);
CHKERRQ(ierr);
ierr = PetscSFBcastBegin(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
ierr = PetscSFBcastEnd(sf, MPIU_INT, §ion->atlasOff[-section->atlasLayout.pStart], &remoteOffsets[-section->atlasLayout.pStart]);
CHKERRQ(ierr);
for (i = section->atlasLayout.pStart; i < section->atlasLayout.pEnd; ++i) {
ierr = PetscSynchronizedPrintf(((PetscObject) sf)->comm, "remoteOffsets[%d]: %d\n", i, remoteOffsets[i-section->atlasLayout.pStart]);
CHKERRQ(ierr);
}
#endif
ierr = PetscSynchronizedFlush(((PetscObject) sf)->comm);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscInt), &localIndices);
CHKERRQ(ierr);
ierr = PetscMalloc(numIndices * sizeof(PetscSFNode), &remoteIndices);
CHKERRQ(ierr);
/* Create new index graph */
for (r = 0, ind = 0; r < numRanks; ++r) {
PetscInt rank = ranks[r];
for (i = rankOffsets[r]; i < rankOffsets[r+1]; ++i) {
PetscInt localPoint = localPoints[i];
PetscInt remoteOffset = remoteOffsets[localPoint-section->atlasLayout.pStart];
PetscInt localOffset, dof, d;
ierr = PetscSectionGetOffset(section, localPoint, &localOffset);
CHKERRQ(ierr);
ierr = PetscSectionGetDof(section, localPoint, &dof);
CHKERRQ(ierr);
for (d = 0; d < dof; ++d, ++ind) {
localIndices[ind] = localOffset+d;
remoteIndices[ind].rank = rank;
remoteIndices[ind].index = remoteOffset+d;
}
}
}
ierr = PetscFree(remoteOffsets);
CHKERRQ(ierr);
if (numIndices != ind) SETERRQ2(((PetscObject) sf)->comm, PETSC_ERR_PLIB, "Inconsistency in indices, %d should be %d", ind, numIndices);
ierr = PetscSFCreate(((PetscObject) sf)->comm, sectionSF);
CHKERRQ(ierr);
ierr = PetscSFSetGraph(*sectionSF, numIndices, numIndices, localIndices, PETSC_OWN_POINTER, remoteIndices, PETSC_OWN_POINTER);
CHKERRQ(ierr);
ierr = PetscSFView(*sectionSF, NULL);
CHKERRQ(ierr);
PetscFunctionReturn(0);
}
