PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat mat)
{
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
Mat_SeqAIJ *B = (Mat_SeqAIJ*)(aij->B->data);
PetscErrorCode ierr;
PetscInt i,j,*aj = B->j,ec = 0,*garray;
IS from,to;
Vec gvec;
#if defined(PETSC_USE_CTABLE)
PetscTable gid1_lid1;
PetscTablePosition tpos;
PetscInt gid,lid;
#else
PetscInt N = mat->cmap->N,*indices;
#endif
PetscFunctionBegin;
if (!aij->garray) {
#if defined(PETSC_USE_CTABLE)
/* use a table */
ierr = PetscTableCreate(aij->B->rmap->n,mat->cmap->N+1,&gid1_lid1);CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt data,gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr);
if (!data) {
/* one based table */
ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr);
}
}
}
/* form array of columns we need */
ierr = PetscMalloc1(ec+1,&garray);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
while (tpos) {
ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr);
gid--;
lid--;
garray[lid] = gid;
}
ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */
ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
for (i=0; i<ec; i++) {
ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr);
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr);
lid--;
aj[B->i[i] + j] = lid;
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
aij->B->cmap->bs = 1;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr);
#else
/* Make an array as long as the number of columns */
/* mark those columns that are in aij->B */
ierr = PetscCalloc1(N+1,&indices);CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
if (!indices[aj[B->i[i] + j]]) ec++;
indices[aj[B->i[i] + j]] = 1;
}
}
/* form array of columns we need */
ierr = PetscMalloc1(ec+1,&garray);CHKERRQ(ierr);
ec = 0;
for (i=0; i<N; i++) {
if (indices[i]) garray[ec++] = i;
}
/* make indices now point into garray */
for (i=0; i<ec; i++) {
indices[garray[i]] = i;
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
aij->B->cmap->bs = 1;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscFree(indices);CHKERRQ(ierr);
#endif
} else {
garray = aij->garray;
}
if (!aij->lvec) {
/* create local vector that is used to scatter into */
ierr = VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);CHKERRQ(ierr);
} else {
ierr = VecGetSize(aij->lvec,&ec);CHKERRQ(ierr);
}
/* create two temporary Index sets for build scatter gather */
ierr = ISCreateGeneral(((PetscObject)mat)->comm,ec,garray,PETSC_COPY_VALUES,&from);CHKERRQ(ierr);
ierr = ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);CHKERRQ(ierr);
/* create temporary global vector to generate scatter context */
/* This does not allocate the array's memory so is efficient */
ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)mat),1,mat->cmap->n,mat->cmap->N,NULL,&gvec);CHKERRQ(ierr);
/* generate the scatter context */
if (aij->Mvctx_mpi1_flg) {
ierr = VecScatterDestroy(&aij->Mvctx_mpi1);CHKERRQ(ierr);
ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx_mpi1);CHKERRQ(ierr);
ierr = VecScatterSetType(aij->Mvctx_mpi1,VECSCATTERMPI1);CHKERRQ(ierr);
ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->Mvctx_mpi1);CHKERRQ(ierr);
} else {
ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr);
ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)aij->lvec);CHKERRQ(ierr);
ierr = PetscLogObjectMemory((PetscObject)mat,(ec+1)*sizeof(PetscInt));CHKERRQ(ierr);
}
aij->garray = garray;
ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)from);CHKERRQ(ierr);
ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)to);CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = VecDestroy(&gvec);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat mat)
{
Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;
Mat_SeqBAIJ *B = (Mat_SeqBAIJ*)(baij->B->data);
PetscErrorCode ierr;
PetscInt i,j,*aj = B->j,ec = 0,*garray;
PetscInt bs = mat->rmap->bs,*stmp;
IS from,to;
Vec gvec;
#if defined(PETSC_USE_CTABLE)
PetscTable gid1_lid1;
PetscTablePosition tpos;
PetscInt gid,lid;
#else
PetscInt Nbs = baij->Nbs,*indices;
#endif
PetscFunctionBegin;
#if defined(PETSC_USE_CTABLE)
/* use a table - Mark Adams */
ierr = PetscTableCreate(B->mbs,baij->Nbs+1,&gid1_lid1);CHKERRQ(ierr);
for (i=0; i<B->mbs; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt data,gid1 = aj[B->i[i]+j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr);
if (!data) {
/* one based table */
ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr);
}
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
while (tpos) {
ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr);
gid--; lid--;
garray[lid] = gid;
}
ierr = PetscSortInt(ec,garray);CHKERRQ(ierr);
ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
for (i=0; i<ec; i++) {
ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr);
}
/* compact out the extra columns in B */
for (i=0; i<B->mbs; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr);
lid--;
aj[B->i[i]+j] = lid;
}
}
B->nbs = ec;
baij->B->cmap->n = baij->B->cmap->N = ec*mat->rmap->bs;
ierr = PetscLayoutSetUp((baij->B->cmap));CHKERRQ(ierr);
ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr);
#else
/* Make an array as long as the number of columns */
/* mark those columns that are in baij->B */
ierr = PetscMalloc((Nbs+1)*sizeof(PetscInt),&indices);CHKERRQ(ierr);
ierr = PetscMemzero(indices,Nbs*sizeof(PetscInt));CHKERRQ(ierr);
for (i=0; i<B->mbs; i++) {
for (j=0; j<B->ilen[i]; j++) {
if (!indices[aj[B->i[i] + j]]) ec++;
indices[aj[B->i[i] + j]] = 1;
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ec = 0;
for (i=0; i<Nbs; i++) {
if (indices[i]) {
garray[ec++] = i;
}
}
/* make indices now point into garray */
for (i=0; i<ec; i++) {
indices[garray[i]] = i;
}
/* compact out the extra columns in B */
for (i=0; i<B->mbs; i++) {
for (j=0; j<B->ilen[i]; j++) {
aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
}
}
B->nbs = ec;
baij->B->cmap->n = baij->B->cmap->N = ec*mat->rmap->bs;
ierr = PetscLayoutSetUp((baij->B->cmap));CHKERRQ(ierr);
ierr = PetscFree(indices);CHKERRQ(ierr);
#endif
/* create local vector that is used to scatter into */
ierr = VecCreateSeq(PETSC_COMM_SELF,ec*bs,&baij->lvec);CHKERRQ(ierr);
/* create two temporary index sets for building scatter-gather */
ierr = ISCreateBlock(PETSC_COMM_SELF,bs,ec,garray,PETSC_COPY_VALUES,&from);CHKERRQ(ierr);
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&stmp);CHKERRQ(ierr);
for (i=0; i<ec; i++) stmp[i] = i;
ierr = ISCreateBlock(PETSC_COMM_SELF,bs,ec,stmp,PETSC_OWN_POINTER,&to);CHKERRQ(ierr);
/* create temporary global vector to generate scatter context */
ierr = VecCreateMPIWithArray(PetscObjectComm((PetscObject)mat),1,mat->cmap->n,mat->cmap->N,NULL,&gvec);CHKERRQ(ierr);
ierr = VecScatterCreate(gvec,from,baij->lvec,to,&baij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,baij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,baij->lvec);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,from);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,to);CHKERRQ(ierr);
baij->garray = garray;
ierr = PetscLogObjectMemory(mat,(ec+1)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = VecDestroy(&gvec);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@C
PetscSFSetGraph - Set a parallel star forest
Collective
Input Arguments:
+ sf - star forest
. nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
. nleaves - number of leaf vertices on the current process, each of these references a root on any process
. ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage
. localmode - copy mode for ilocal
. iremote - remote locations of root vertices for each leaf on the current process
- remotemode - copy mode for iremote
Level: intermediate
.seealso: PetscSFCreate(), PetscSFView(), PetscSFGetGraph()
@*/
PetscErrorCode PetscSFSetGraph(PetscSF sf,PetscInt nroots,PetscInt nleaves,const PetscInt *ilocal,PetscCopyMode localmode,const PetscSFNode *iremote,PetscCopyMode remotemode)
{
PetscErrorCode ierr;
PetscTable table;
PetscTablePosition pos;
PetscMPIInt size;
PetscInt i,*rcount,*ranks;
PetscFunctionBegin;
PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
ierr = PetscLogEventBegin(PETSCSF_SetGraph,sf,0,0,0);CHKERRQ(ierr);
if (nleaves && ilocal) PetscValidIntPointer(ilocal,4);
if (nleaves) PetscValidPointer(iremote,6);
if (nroots < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"roots %D, cannot be negative",nroots);
if (nleaves < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nleaves %D, cannot be negative",nleaves);
ierr = PetscSFReset(sf);CHKERRQ(ierr);
sf->nroots = nroots;
sf->nleaves = nleaves;
if (ilocal) {
switch (localmode) {
case PETSC_COPY_VALUES:
ierr = PetscMalloc1(nleaves,&sf->mine_alloc);CHKERRQ(ierr);
sf->mine = sf->mine_alloc;
ierr = PetscMemcpy(sf->mine,ilocal,nleaves*sizeof(*sf->mine));CHKERRQ(ierr);
sf->minleaf = PETSC_MAX_INT;
sf->maxleaf = PETSC_MIN_INT;
for (i=0; i<nleaves; i++) {
sf->minleaf = PetscMin(sf->minleaf,ilocal[i]);
sf->maxleaf = PetscMax(sf->maxleaf,ilocal[i]);
}
break;
case PETSC_OWN_POINTER:
sf->mine_alloc = (PetscInt*)ilocal;
sf->mine = sf->mine_alloc;
break;
case PETSC_USE_POINTER:
sf->mine = (PetscInt*)ilocal;
break;
default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_OUTOFRANGE,"Unknown localmode");
}
}
if (!ilocal || nleaves > 0) {
sf->minleaf = 0;
sf->maxleaf = nleaves - 1;
}
switch (remotemode) {
case PETSC_COPY_VALUES:
ierr = PetscMalloc1(nleaves,&sf->remote_alloc);CHKERRQ(ierr);
sf->remote = sf->remote_alloc;
ierr = PetscMemcpy(sf->remote,iremote,nleaves*sizeof(*sf->remote));CHKERRQ(ierr);
break;
case PETSC_OWN_POINTER:
sf->remote_alloc = (PetscSFNode*)iremote;
sf->remote = sf->remote_alloc;
break;
case PETSC_USE_POINTER:
sf->remote = (PetscSFNode*)iremote;
break;
default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_OUTOFRANGE,"Unknown remotemode");
}
ierr = MPI_Comm_size(PetscObjectComm((PetscObject)sf),&size);CHKERRQ(ierr);
ierr = PetscTableCreate(10,size,&table);CHKERRQ(ierr);
for (i=0; i<nleaves; i++) {
/* Log 1-based rank */
ierr = PetscTableAdd(table,iremote[i].rank+1,1,ADD_VALUES);CHKERRQ(ierr);
}
ierr = PetscTableGetCount(table,&sf->nranks);CHKERRQ(ierr);
ierr = PetscMalloc4(sf->nranks,&sf->ranks,sf->nranks+1,&sf->roffset,nleaves,&sf->rmine,nleaves,&sf->rremote);CHKERRQ(ierr);
ierr = PetscMalloc2(sf->nranks,&rcount,sf->nranks,&ranks);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(table,&pos);CHKERRQ(ierr);
for (i=0; i<sf->nranks; i++) {
ierr = PetscTableGetNext(table,&pos,&ranks[i],&rcount[i]);CHKERRQ(ierr);
ranks[i]--; /* Convert back to 0-based */
}
ierr = PetscTableDestroy(&table);CHKERRQ(ierr);
ierr = PetscSortIntWithArray(sf->nranks,ranks,rcount);CHKERRQ(ierr);
sf->roffset[0] = 0;
for (i=0; i<sf->nranks; i++) {
ierr = PetscMPIIntCast(ranks[i],sf->ranks+i);CHKERRQ(ierr);
sf->roffset[i+1] = sf->roffset[i] + rcount[i];
rcount[i] = 0;
}
for (i=0; i<nleaves; i++) {
PetscInt lo,hi,irank;
/* Search for index of iremote[i].rank in sf->ranks */
lo = 0; hi = sf->nranks;
while (hi - lo > 1) {
PetscInt mid = lo + (hi - lo)/2;
if (iremote[i].rank < sf->ranks[mid]) hi = mid;
else lo = mid;
}
if (hi - lo == 1 && iremote[i].rank == sf->ranks[lo]) irank = lo;
else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Could not find rank %D in array",iremote[i].rank);
sf->rmine[sf->roffset[irank] + rcount[irank]] = ilocal ? ilocal[i] : i;
sf->rremote[sf->roffset[irank] + rcount[irank]] = iremote[i].index;
rcount[irank]++;
}
ierr = PetscFree2(rcount,ranks);CHKERRQ(ierr);
#if !defined(PETSC_USE_64BIT_INDICES)
if (nroots == PETSC_DETERMINE) {
/* Jed, if you have a better way to do this, put it in */
PetscInt *numRankLeaves, *leafOff, *leafIndices, *numRankRoots, *rootOff, *rootIndices, maxRoots = 0;
/* All to all to determine number of leaf indices from each (you can do this using Scan and asynch messages) */
ierr = PetscMalloc4(size,&numRankLeaves,size+1,&leafOff,size,&numRankRoots,size+1,&rootOff);CHKERRQ(ierr);
ierr = PetscMemzero(numRankLeaves, size * sizeof(PetscInt));CHKERRQ(ierr);
for (i = 0; i < nleaves; ++i) ++numRankLeaves[iremote[i].rank];
ierr = MPI_Alltoall(numRankLeaves, 1, MPIU_INT, numRankRoots, 1, MPIU_INT, PetscObjectComm((PetscObject)sf));CHKERRQ(ierr);
/* Could set nroots to this maximum */
for (i = 0; i < size; ++i) maxRoots += numRankRoots[i];
/* Gather all indices */
ierr = PetscMalloc2(nleaves,&leafIndices,maxRoots,&rootIndices);CHKERRQ(ierr);
leafOff[0] = 0;
for (i = 0; i < size; ++i) leafOff[i+1] = leafOff[i] + numRankLeaves[i];
for (i = 0; i < nleaves; ++i) leafIndices[leafOff[iremote[i].rank]++] = iremote[i].index;
leafOff[0] = 0;
for (i = 0; i < size; ++i) leafOff[i+1] = leafOff[i] + numRankLeaves[i];
rootOff[0] = 0;
for (i = 0; i < size; ++i) rootOff[i+1] = rootOff[i] + numRankRoots[i];
ierr = MPI_Alltoallv(leafIndices, numRankLeaves, leafOff, MPIU_INT, rootIndices, numRankRoots, rootOff, MPIU_INT, PetscObjectComm((PetscObject)sf));CHKERRQ(ierr);
/* Sort and reduce */
ierr = PetscSortRemoveDupsInt(&maxRoots, rootIndices);CHKERRQ(ierr);
ierr = PetscFree2(leafIndices,rootIndices);CHKERRQ(ierr);
ierr = PetscFree4(numRankLeaves,leafOff,numRankRoots,rootOff);CHKERRQ(ierr);
sf->nroots = maxRoots;
}
#endif
sf->graphset = PETSC_TRUE;
ierr = PetscLogEventEnd(PETSCSF_SetGraph,sf,0,0,0);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
/*@
ISCompressIndicesGeneral - convert the indices into block indices
Input Parameters:
+ n - maximum possible length of the index set
. nkeys - expected number of keys when PETSC_USE_CTABLE
. bs - the size of block
. imax - the number of index sets
- is_in - the non-blocked array of index sets
Output Parameter:
. is_out - the blocked new index set
Level: intermediate
.seealso: ISExpandIndicesGeneral()
@*/
PetscErrorCode ISCompressIndicesGeneral(PetscInt n,PetscInt nkeys,PetscInt bs,PetscInt imax,const IS is_in[],IS is_out[])
{
PetscErrorCode ierr;
PetscInt isz,len,i,j,ival,Nbs;
const PetscInt *idx;
#if defined(PETSC_USE_CTABLE)
PetscTable gid1_lid1;
PetscInt tt, gid1, *nidx,Nkbs;
PetscTablePosition tpos;
#else
PetscInt *nidx;
PetscBT table;
#endif
PetscFunctionBegin;
Nbs =n/bs;
#if defined(PETSC_USE_CTABLE)
Nkbs = nkeys/bs;
ierr = PetscTableCreate(Nkbs,Nbs,&gid1_lid1);CHKERRQ(ierr);
#else
ierr = PetscMalloc1(Nbs,&nidx);CHKERRQ(ierr);
ierr = PetscBTCreate(Nbs,&table);CHKERRQ(ierr);
#endif
for (i=0; i<imax; i++) {
isz = 0;
#if defined(PETSC_USE_CTABLE)
ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
#else
ierr = PetscBTMemzero(Nbs,table);CHKERRQ(ierr);
#endif
ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
ierr = ISGetLocalSize(is_in[i],&len);CHKERRQ(ierr);
for (j=0; j<len; j++) {
ival = idx[j]/bs; /* convert the indices into block indices */
#if defined(PETSC_USE_CTABLE)
ierr = PetscTableFind(gid1_lid1,ival+1,&tt);CHKERRQ(ierr);
if (!tt) {
ierr = PetscTableAdd(gid1_lid1,ival+1,isz+1,INSERT_VALUES);CHKERRQ(ierr);
isz++;
}
#else
if (ival>Nbs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"index greater than mat-dim");
if (!PetscBTLookupSet(table,ival)) nidx[isz++] = ival;
#endif
}
ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr);
#if defined(PETSC_USE_CTABLE)
ierr = PetscMalloc1(isz,&nidx);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
j = 0;
while (tpos) {
ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid1,&tt);CHKERRQ(ierr);
if (tt-- > isz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"index greater than array-dim");
nidx[tt] = gid1 - 1;
j++;
}
if (j != isz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"table error: jj != isz");
ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_OWN_POINTER,(is_out+i));CHKERRQ(ierr);
#else
ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,PETSC_COPY_VALUES,(is_out+i));CHKERRQ(ierr);
#endif
}
#if defined(PETSC_USE_CTABLE)
ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr);
#else
ierr = PetscBTDestroy(&table);CHKERRQ(ierr);
ierr = PetscFree(nidx);CHKERRQ(ierr);
#endif
PetscFunctionReturn(0);
}
PetscErrorCode MatSetUpMultiply_MPIAIJ(Mat mat)
{
Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
Mat_SeqAIJ *B = (Mat_SeqAIJ*)(aij->B->data);
PetscErrorCode ierr;
PetscInt i,j,*aj = B->j,ec = 0,*garray;
IS from,to;
Vec gvec;
PetscBool useblockis;
#if defined (PETSC_USE_CTABLE)
PetscTable gid1_lid1;
PetscTablePosition tpos;
PetscInt gid,lid;
#else
PetscInt N = mat->cmap->N,*indices;
#endif
PetscFunctionBegin;
#if defined (PETSC_USE_CTABLE)
/* use a table */
ierr = PetscTableCreate(aij->B->rmap->n,mat->cmap->N+1,&gid1_lid1);CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt data,gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&data);CHKERRQ(ierr);
if (!data) {
/* one based table */
ierr = PetscTableAdd(gid1_lid1,gid1,++ec,INSERT_VALUES);CHKERRQ(ierr);
}
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
while (tpos) {
ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid,&lid);CHKERRQ(ierr);
gid--;
lid--;
garray[lid] = gid;
}
ierr = PetscSortInt(ec,garray);CHKERRQ(ierr); /* sort, and rebuild */
ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
for (i=0; i<ec; i++) {
ierr = PetscTableAdd(gid1_lid1,garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr);
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
PetscInt gid1 = aj[B->i[i] + j] + 1;
ierr = PetscTableFind(gid1_lid1,gid1,&lid);CHKERRQ(ierr);
lid --;
aj[B->i[i] + j] = lid;
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscTableDestroy(&gid1_lid1);CHKERRQ(ierr);
#else
/* Make an array as long as the number of columns */
/* mark those columns that are in aij->B */
ierr = PetscMalloc((N+1)*sizeof(PetscInt),&indices);CHKERRQ(ierr);
ierr = PetscMemzero(indices,N*sizeof(PetscInt));CHKERRQ(ierr);
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
if (!indices[aj[B->i[i] + j] ]) ec++;
indices[aj[B->i[i] + j] ] = 1;
}
}
/* form array of columns we need */
ierr = PetscMalloc((ec+1)*sizeof(PetscInt),&garray);CHKERRQ(ierr);
ec = 0;
for (i=0; i<N; i++) {
if (indices[i]) garray[ec++] = i;
}
/* make indices now point into garray */
for (i=0; i<ec; i++) {
indices[garray[i]] = i;
}
/* compact out the extra columns in B */
for (i=0; i<aij->B->rmap->n; i++) {
for (j=0; j<B->ilen[i]; j++) {
aj[B->i[i] + j] = indices[aj[B->i[i] + j]];
}
}
aij->B->cmap->n = aij->B->cmap->N = ec;
ierr = PetscLayoutSetUp((aij->B->cmap));CHKERRQ(ierr);
ierr = PetscFree(indices);CHKERRQ(ierr);
#endif
/* create local vector that is used to scatter into */
ierr = VecCreateSeq(PETSC_COMM_SELF,ec,&aij->lvec);CHKERRQ(ierr);
/* create two temporary Index sets for build scatter gather */
/* check for the special case where blocks are communicated for faster VecScatterXXX */
useblockis = PETSC_FALSE;
if (mat->cmap->bs > 1) {
PetscInt bs = mat->cmap->bs,ibs,ga;
if (!(ec % bs)) {
useblockis = PETSC_TRUE;
for (i=0; i<ec/bs; i++) {
if ((ga = garray[ibs = i*bs]) % bs) {
useblockis = PETSC_FALSE;
break;
}
for (j=1; j<bs; j++) {
if (garray[ibs+j] != ga+j) {
useblockis = PETSC_FALSE;
break;
}
}
if (!useblockis) break;
}
}
}
#if defined(PETSC_USE_DEBUG)
i = (PetscInt)useblockis;
ierr = MPI_Allreduce(&i,&j,1,MPIU_INT,MPI_MIN,((PetscObject)mat)->comm); CHKERRQ(ierr);
if(j!=i) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Use of blocked not consistant (I am usning blocked)");
#endif
if (useblockis) {
PetscInt *ga,bs = mat->cmap->bs,iec = ec/bs;
if(ec%bs)SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"ec=%D bs=%D",ec,bs);
ierr = PetscInfo(mat,"Using block index set to define scatter\n");
ierr = PetscMalloc(iec*sizeof(PetscInt),&ga);CHKERRQ(ierr);
for (i=0; i<iec; i++) ga[i] = garray[i*bs]/bs;
ierr = ISCreateBlock(((PetscObject)mat)->comm,bs,iec,ga,PETSC_OWN_POINTER,&from);CHKERRQ(ierr);
} else {
ierr = ISCreateGeneral(((PetscObject)mat)->comm,ec,garray,PETSC_COPY_VALUES,&from);CHKERRQ(ierr);
}
ierr = ISCreateStride(PETSC_COMM_SELF,ec,0,1,&to);CHKERRQ(ierr);
/* create temporary global vector to generate scatter context */
/* This does not allocate the array's memory so is efficient */
ierr = VecCreateMPIWithArray(((PetscObject)mat)->comm,1,mat->cmap->n,mat->cmap->N,PETSC_NULL,&gvec);CHKERRQ(ierr);
/* generate the scatter context */
ierr = VecScatterCreate(gvec,from,aij->lvec,to,&aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,aij->Mvctx);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,aij->lvec);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,from);CHKERRQ(ierr);
ierr = PetscLogObjectParent(mat,to);CHKERRQ(ierr);
aij->garray = garray;
ierr = PetscLogObjectMemory(mat,(ec+1)*sizeof(PetscInt));CHKERRQ(ierr);
ierr = ISDestroy(&from);CHKERRQ(ierr);
ierr = ISDestroy(&to);CHKERRQ(ierr);
ierr = VecDestroy(&gvec);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
