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);
}
/*@
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_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);
}
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);
}
