static PetscErrorCode DMCreateMatrix_Composite_AIJ(DM dm,MatType mtype,Mat *J)
{
PetscErrorCode ierr;
DM_Composite *com = (DM_Composite*)dm->data;
struct DMCompositeLink *next;
PetscInt m,*dnz,*onz,i,j,mA;
Mat Atmp;
PetscMPIInt rank;
PetscBool dense = PETSC_FALSE;
PetscFunctionBegin;
/* use global vector to determine layout needed for matrix */
m = com->n;
ierr = MatCreate(PetscObjectComm((PetscObject)dm),J);CHKERRQ(ierr);
ierr = MatSetSizes(*J,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetType(*J,mtype);CHKERRQ(ierr);
/*
Extremely inefficient but will compute entire Jacobian for testing
*/
ierr = PetscOptionsGetBool(((PetscObject)dm)->prefix,"-dmcomposite_dense_jacobian",&dense,NULL);CHKERRQ(ierr);
if (dense) {
PetscInt rstart,rend,*indices;
PetscScalar *values;
mA = com->N;
ierr = MatMPIAIJSetPreallocation(*J,mA,NULL,mA-m,NULL);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(*J,mA,NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(*J,&rstart,&rend);CHKERRQ(ierr);
ierr = PetscMalloc2(mA,PetscScalar,&values,mA,PetscInt,&indices);CHKERRQ(ierr);
ierr = PetscMemzero(values,mA*sizeof(PetscScalar));CHKERRQ(ierr);
for (i=0; i<mA; i++) indices[i] = i;
for (i=rstart; i<rend; i++) {
ierr = MatSetValues(*J,1,&i,mA,indices,values,INSERT_VALUES);CHKERRQ(ierr);
}
ierr = PetscFree2(values,indices);CHKERRQ(ierr);
ierr = MatAssemblyBegin(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)dm),&rank);CHKERRQ(ierr);
ierr = MatPreallocateInitialize(PetscObjectComm((PetscObject)dm),m,m,dnz,onz);CHKERRQ(ierr);
/* loop over packed objects, handling one at at time */
next = com->next;
while (next) {
PetscInt nc,rstart,*ccols,maxnc;
const PetscInt *cols,*rstarts;
PetscMPIInt proc;
ierr = DMCreateMatrix(next->dm,mtype,&Atmp);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Atmp,&rstart,NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRanges(Atmp,&rstarts);CHKERRQ(ierr);
ierr = MatGetLocalSize(Atmp,&mA,NULL);CHKERRQ(ierr);
maxnc = 0;
for (i=0; i<mA; i++) {
ierr = MatGetRow(Atmp,rstart+i,&nc,NULL,NULL);CHKERRQ(ierr);
ierr = MatRestoreRow(Atmp,rstart+i,&nc,NULL,NULL);CHKERRQ(ierr);
maxnc = PetscMax(nc,maxnc);
}
ierr = PetscMalloc(maxnc*sizeof(PetscInt),&ccols);CHKERRQ(ierr);
for (i=0; i<mA; i++) {
ierr = MatGetRow(Atmp,rstart+i,&nc,&cols,NULL);CHKERRQ(ierr);
/* remap the columns taking into how much they are shifted on each process */
for (j=0; j<nc; j++) {
proc = 0;
while (cols[j] >= rstarts[proc+1]) proc++;
ccols[j] = cols[j] + next->grstarts[proc] - rstarts[proc];
}
ierr = MatPreallocateSet(com->rstart+next->rstart+i,nc,ccols,dnz,onz);CHKERRQ(ierr);
ierr = MatRestoreRow(Atmp,rstart+i,&nc,&cols,NULL);CHKERRQ(ierr);
}
ierr = PetscFree(ccols);CHKERRQ(ierr);
ierr = MatDestroy(&Atmp);CHKERRQ(ierr);
next = next->next;
}
if (com->FormCoupleLocations) {
ierr = (*com->FormCoupleLocations)(dm,NULL,dnz,onz,__rstart,__nrows,__start,__end);CHKERRQ(ierr);
}
ierr = MatMPIAIJSetPreallocation(*J,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatSeqAIJSetPreallocation(*J,0,dnz);CHKERRQ(ierr);
ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
if (dm->prealloc_only) PetscFunctionReturn(0);
next = com->next;
while (next) {
PetscInt nc,rstart,row,maxnc,*ccols;
const PetscInt *cols,*rstarts;
const PetscScalar *values;
PetscMPIInt proc;
ierr = DMCreateMatrix(next->dm,mtype,&Atmp);CHKERRQ(ierr);
ierr = MatGetOwnershipRange(Atmp,&rstart,NULL);CHKERRQ(ierr);
ierr = MatGetOwnershipRanges(Atmp,&rstarts);CHKERRQ(ierr);
ierr = MatGetLocalSize(Atmp,&mA,NULL);CHKERRQ(ierr);
maxnc = 0;
for (i=0; i<mA; i++) {
ierr = MatGetRow(Atmp,rstart+i,&nc,NULL,NULL);CHKERRQ(ierr);
ierr = MatRestoreRow(Atmp,rstart+i,&nc,NULL,NULL);CHKERRQ(ierr);
maxnc = PetscMax(nc,maxnc);
}
ierr = PetscMalloc(maxnc*sizeof(PetscInt),&ccols);CHKERRQ(ierr);
for (i=0; i<mA; i++) {
ierr = MatGetRow(Atmp,rstart+i,&nc,(const PetscInt**)&cols,&values);CHKERRQ(ierr);
for (j=0; j<nc; j++) {
proc = 0;
while (cols[j] >= rstarts[proc+1]) proc++;
ccols[j] = cols[j] + next->grstarts[proc] - rstarts[proc];
}
row = com->rstart+next->rstart+i;
ierr = MatSetValues(*J,1,&row,nc,ccols,values,INSERT_VALUES);CHKERRQ(ierr);
ierr = MatRestoreRow(Atmp,rstart+i,&nc,(const PetscInt**)&cols,&values);CHKERRQ(ierr);
}
ierr = PetscFree(ccols);CHKERRQ(ierr);
ierr = MatDestroy(&Atmp);CHKERRQ(ierr);
next = next->next;
}
if (com->FormCoupleLocations) {
PetscInt __rstart;
ierr = MatGetOwnershipRange(*J,&__rstart,NULL);CHKERRQ(ierr);
ierr = (*com->FormCoupleLocations)(dm,*J,NULL,NULL,__rstart,0,0,0);CHKERRQ(ierr);
}
ierr = MatAssemblyBegin(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
ierr = MatAssemblyEnd(*J,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
PetscFunctionReturn(0);
}
PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C)
{
PetscErrorCode ierr;
Mat Cmpi;
Mat_PtAPMPI *ptap;
PetscFreeSpaceList free_space=NULL,current_space=NULL;
Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data,*c;
Mat_SeqAIJ *ad =(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data;
Mat_SeqAIJ *p_loc,*p_oth;
PetscInt *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ;
PetscInt *adi=ad->i,*aj,*aoi=ao->i,nnz;
PetscInt *lnk,*owners_co,*coi,*coj,i,k,pnz,row;
PetscInt am=A->rmap->n,pN=P->cmap->N,pm=P->rmap->n,pn=P->cmap->n;
PetscBT lnkbt;
MPI_Comm comm;
PetscMPIInt size,rank,tagi,tagj,*len_si,*len_s,*len_ri,icompleted=0;
PetscInt **buf_rj,**buf_ri,**buf_ri_k;
PetscInt len,proc,*dnz,*onz,*owners;
PetscInt nzi,*pti,*ptj;
PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci;
MPI_Request *swaits,*rwaits;
MPI_Status *sstatus,rstatus;
Mat_Merge_SeqsToMPI *merge;
PetscInt *api,*apj,*Jptr,apnz,*prmap=p->garray,pon,nspacedouble=0,j,ap_rmax=0;
PetscReal afill=1.0,afill_tmp;
PetscInt rmax;
#if defined(PTAP_PROFILE)
PetscLogDouble t0,t1,t2,t3,t4;
#endif
PetscFunctionBegin;
ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t0);CHKERRQ(ierr);
#endif
/* check if matrix local sizes are compatible */
if (A->rmap->rstart != P->rmap->rstart || A->rmap->rend != P->rmap->rend) {
SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Arow (%D, %D) != Prow (%D,%D)",A->rmap->rstart,A->rmap->rend,P->rmap->rstart,P->rmap->rend);
}
if (A->cmap->rstart != P->rmap->rstart || A->cmap->rend != P->rmap->rend) {
SETERRQ4(comm,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, Acol (%D, %D) != Prow (%D,%D)",A->cmap->rstart,A->cmap->rend,P->rmap->rstart,P->rmap->rend);
}
ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
/* create struct Mat_PtAPMPI and attached it to C later */
ierr = PetscNew(&ptap);CHKERRQ(ierr);
ierr = PetscNew(&merge);CHKERRQ(ierr);
ptap->merge = merge;
ptap->reuse = MAT_INITIAL_MATRIX;
/* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */
ierr = MatGetBrowsOfAoCols_MPIAIJ(A,P,MAT_INITIAL_MATRIX,&ptap->startsj_s,&ptap->startsj_r,&ptap->bufa,&ptap->P_oth);CHKERRQ(ierr);
/* get P_loc by taking all local rows of P */
ierr = MatMPIAIJGetLocalMat(P,MAT_INITIAL_MATRIX,&ptap->P_loc);CHKERRQ(ierr);
p_loc = (Mat_SeqAIJ*)(ptap->P_loc)->data;
p_oth = (Mat_SeqAIJ*)(ptap->P_oth)->data;
pi_loc = p_loc->i; pj_loc = p_loc->j;
pi_oth = p_oth->i; pj_oth = p_oth->j;
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t1);CHKERRQ(ierr);
#endif
/* first, compute symbolic AP = A_loc*P = A_diag*P_loc + A_off*P_oth */
/*-------------------------------------------------------------------*/
ierr = PetscMalloc1((am+1),&api);CHKERRQ(ierr);
api[0] = 0;
/* create and initialize a linked list */
ierr = PetscLLCondensedCreate(pN,pN,&lnk,&lnkbt);CHKERRQ(ierr);
/* Initial FreeSpace size is fill*(nnz(A) + nnz(P)) -OOM for ex56, np=8k on Intrepid! */
ierr = PetscFreeSpaceGet((PetscInt)(fill*(adi[am]+aoi[am]+pi_loc[pm])),&free_space);CHKERRQ(ierr);
current_space = free_space;
for (i=0; i<am; i++) {
/* diagonal portion of A */
nzi = adi[i+1] - adi[i];
aj = ad->j + adi[i];
for (j=0; j<nzi; j++) {
row = aj[j];
pnz = pi_loc[row+1] - pi_loc[row];
Jptr = pj_loc + pi_loc[row];
/* add non-zero cols of P into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
/* off-diagonal portion of A */
nzi = aoi[i+1] - aoi[i];
aj = ao->j + aoi[i];
for (j=0; j<nzi; j++) {
row = aj[j];
pnz = pi_oth[row+1] - pi_oth[row];
Jptr = pj_oth + pi_oth[row];
ierr = PetscLLCondensedAddSorted(pnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
apnz = lnk[0];
api[i+1] = api[i] + apnz;
if (ap_rmax < apnz) ap_rmax = apnz;
/* if free space is not available, double the total space in the list */
if (current_space->local_remaining<apnz) {
ierr = PetscFreeSpaceGet(apnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* Copy data into free space, then initialize lnk */
ierr = PetscLLCondensedClean(pN,apnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
current_space->array += apnz;
current_space->local_used += apnz;
current_space->local_remaining -= apnz;
}
/* Allocate space for apj, initialize apj, and */
/* destroy list of free space and other temporary array(s) */
ierr = PetscMalloc1((api[am]+1),&apj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,apj);CHKERRQ(ierr);
afill_tmp = (PetscReal)api[am]/(adi[am]+aoi[am]+pi_loc[pm]+1);
if (afill_tmp > afill) afill = afill_tmp;
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t2);CHKERRQ(ierr);
#endif
/* determine symbolic Co=(p->B)^T*AP - send to others */
/*----------------------------------------------------*/
ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
/* then, compute symbolic Co = (p->B)^T*AP */
pon = (p->B)->cmap->n; /* total num of rows to be sent to other processors
>= (num of nonzero rows of C_seq) - pn */
ierr = PetscMalloc1((pon+1),&coi);CHKERRQ(ierr);
coi[0] = 0;
/* set initial free space to be fill*(nnz(p->B) + nnz(AP)) */
nnz = fill*(poti[pon] + api[am]);
ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr);
current_space = free_space;
for (i=0; i<pon; i++) {
pnz = poti[i+1] - poti[i];
ptJ = potj + poti[i];
for (j=0; j<pnz; j++) {
row = ptJ[j]; /* row of AP == col of Pot */
apnz = api[row+1] - api[row];
Jptr = apj + api[row];
/* add non-zero cols of AP into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
nnz = lnk[0];
/* If free space is not available, double the total space in the list */
if (current_space->local_remaining<nnz) {
ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* Copy data into free space, and zero out denserows */
ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
current_space->array += nnz;
current_space->local_used += nnz;
current_space->local_remaining -= nnz;
coi[i+1] = coi[i] + nnz;
}
ierr = PetscMalloc1((coi[pon]+1),&coj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,coj);CHKERRQ(ierr);
afill_tmp = (PetscReal)coi[pon]/(poti[pon] + api[am]+1);
if (afill_tmp > afill) afill = afill_tmp;
ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
/* send j-array (coj) of Co to other processors */
/*----------------------------------------------*/
/* determine row ownership */
ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr);
merge->rowmap->n = pn;
merge->rowmap->bs = 1;
ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr);
owners = merge->rowmap->range;
/* determine the number of messages to send, their lengths */
ierr = PetscMalloc2(size,&len_si,size,&sstatus);CHKERRQ(ierr);
ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr);
ierr = PetscCalloc1(size,&merge->len_s);CHKERRQ(ierr);
len_s = merge->len_s;
merge->nsend = 0;
ierr = PetscMalloc1((size+2),&owners_co);CHKERRQ(ierr);
proc = 0;
for (i=0; i<pon; i++) {
while (prmap[i] >= owners[proc+1]) proc++;
len_si[proc]++; /* num of rows in Co to be sent to [proc] */
len_s[proc] += coi[i+1] - coi[i];
}
len = 0; /* max length of buf_si[] */
owners_co[0] = 0;
for (proc=0; proc<size; proc++) {
owners_co[proc+1] = owners_co[proc] + len_si[proc];
if (len_si[proc]) {
merge->nsend++;
len_si[proc] = 2*(len_si[proc] + 1);
len += len_si[proc];
}
}
/* determine the number and length of messages to receive for coi and coj */
ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr);
ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr);
/* post the Irecv and Isend of coj */
ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr);
ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr);
ierr = PetscMalloc1((merge->nsend+1),&swaits);CHKERRQ(ierr);
for (proc=0, k=0; proc<size; proc++) {
if (!len_s[proc]) continue;
i = owners_co[proc];
ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tagj,comm,swaits+k);CHKERRQ(ierr);
k++;
}
/* receives and sends of coj are complete */
for (i=0; i<merge->nrecv; i++) {
ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr);
}
ierr = PetscFree(rwaits);CHKERRQ(ierr);
if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);}
/* send and recv coi */
/*-------------------*/
ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr);
ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr);
ierr = PetscMalloc1((len+1),&buf_s);CHKERRQ(ierr);
buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
for (proc=0,k=0; proc<size; proc++) {
if (!len_s[proc]) continue;
/* form outgoing message for i-structure:
buf_si[0]: nrows to be sent
[1:nrows]: row index (global)
[nrows+1:2*nrows+1]: i-structure index
*/
/*-------------------------------------------*/
nrows = len_si[proc]/2 - 1;
buf_si_i = buf_si + nrows+1;
buf_si[0] = nrows;
buf_si_i[0] = 0;
nrows = 0;
for (i=owners_co[proc]; i<owners_co[proc+1]; i++) {
nzi = coi[i+1] - coi[i];
buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */
buf_si[nrows+1] = prmap[i] -owners[proc]; /* local row index */
nrows++;
}
ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,swaits+k);CHKERRQ(ierr);
k++;
buf_si += len_si[proc];
}
i = merge->nrecv;
while (i--) {
ierr = MPI_Waitany(merge->nrecv,rwaits,&icompleted,&rstatus);CHKERRQ(ierr);
}
ierr = PetscFree(rwaits);CHKERRQ(ierr);
if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);}
ierr = PetscFree2(len_si,sstatus);CHKERRQ(ierr);
ierr = PetscFree(len_ri);CHKERRQ(ierr);
ierr = PetscFree(swaits);CHKERRQ(ierr);
ierr = PetscFree(buf_s);CHKERRQ(ierr);
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t3);CHKERRQ(ierr);
#endif
/* compute the local portion of C (mpi mat) */
/*------------------------------------------*/
ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
/* allocate pti array and free space for accumulating nonzero column info */
ierr = PetscMalloc1((pn+1),&pti);CHKERRQ(ierr);
pti[0] = 0;
/* set initial free space to be fill*(nnz(P) + nnz(AP)) */
nnz = fill*(pi_loc[pm] + api[am]);
ierr = PetscFreeSpaceGet(nnz,&free_space);CHKERRQ(ierr);
current_space = free_space;
ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextci);CHKERRQ(ierr);
for (k=0; k<merge->nrecv; k++) {
buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
nrows = *buf_ri_k[k];
nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
nextci[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
}
ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr);
rmax = 0;
for (i=0; i<pn; i++) {
/* add pdt[i,:]*AP into lnk */
pnz = pdti[i+1] - pdti[i];
ptJ = pdtj + pdti[i];
for (j=0; j<pnz; j++) {
row = ptJ[j]; /* row of AP == col of Pt */
apnz = api[row+1] - api[row];
Jptr = apj + api[row];
/* add non-zero cols of AP into the sorted linked list lnk */
ierr = PetscLLCondensedAddSorted(apnz,Jptr,lnk,lnkbt);CHKERRQ(ierr);
}
/* add received col data into lnk */
for (k=0; k<merge->nrecv; k++) { /* k-th received message */
if (i == *nextrow[k]) { /* i-th row */
nzi = *(nextci[k]+1) - *nextci[k];
Jptr = buf_rj[k] + *nextci[k];
ierr = PetscLLCondensedAddSorted(nzi,Jptr,lnk,lnkbt);CHKERRQ(ierr);
nextrow[k]++; nextci[k]++;
}
}
nnz = lnk[0];
/* if free space is not available, make more free space */
if (current_space->local_remaining<nnz) {
ierr = PetscFreeSpaceGet(nnz+current_space->total_array_size,¤t_space);CHKERRQ(ierr);
nspacedouble++;
}
/* copy data into free space, then initialize lnk */
ierr = PetscLLCondensedClean(pN,nnz,current_space->array,lnk,lnkbt);CHKERRQ(ierr);
ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr);
current_space->array += nnz;
current_space->local_used += nnz;
current_space->local_remaining -= nnz;
pti[i+1] = pti[i] + nnz;
if (nnz > rmax) rmax = nnz;
}
ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
ierr = PetscFree3(buf_ri_k,nextrow,nextci);CHKERRQ(ierr);
ierr = PetscMalloc1((pti[pn]+1),&ptj);CHKERRQ(ierr);
ierr = PetscFreeSpaceContiguous(&free_space,ptj);CHKERRQ(ierr);
afill_tmp = (PetscReal)pti[pn]/(pi_loc[pm] + api[am]+1);
if (afill_tmp > afill) afill = afill_tmp;
ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
/* create symbolic parallel matrix Cmpi */
/*--------------------------------------*/
ierr = MatCreate(comm,&Cmpi);CHKERRQ(ierr);
ierr = MatSetSizes(Cmpi,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr);
ierr = MatSetBlockSizes(Cmpi,P->cmap->bs,P->cmap->bs);CHKERRQ(ierr);
ierr = MatSetType(Cmpi,MATMPIAIJ);CHKERRQ(ierr);
ierr = MatMPIAIJSetPreallocation(Cmpi,0,dnz,0,onz);CHKERRQ(ierr);
ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
merge->bi = pti; /* Cseq->i */
merge->bj = ptj; /* Cseq->j */
merge->coi = coi; /* Co->i */
merge->coj = coj; /* Co->j */
merge->buf_ri = buf_ri;
merge->buf_rj = buf_rj;
merge->owners_co = owners_co;
merge->destroy = Cmpi->ops->destroy;
merge->duplicate = Cmpi->ops->duplicate;
/* Cmpi is not ready for use - assembly will be done by MatPtAPNumeric() */
Cmpi->assembled = PETSC_FALSE;
Cmpi->ops->destroy = MatDestroy_MPIAIJ_PtAP;
Cmpi->ops->duplicate = MatDuplicate_MPIAIJ_MatPtAP;
/* attach the supporting struct to Cmpi for reuse */
c = (Mat_MPIAIJ*)Cmpi->data;
c->ptap = ptap;
ptap->api = api;
ptap->apj = apj;
ptap->rmax = ap_rmax;
*C = Cmpi;
/* flag 'scalable' determines which implementations to be used:
0: do dense axpy in MatPtAPNumeric() - fast, but requires storage of a nonscalable dense array apa;
1: do sparse axpy in MatPtAPNumeric() - might slow, uses a sparse array apa */
/* set default scalable */
ptap->scalable = PETSC_TRUE;
ierr = PetscOptionsGetBool(((PetscObject)Cmpi)->prefix,"-matptap_scalable",&ptap->scalable,NULL);CHKERRQ(ierr);
if (!ptap->scalable) { /* Do dense axpy */
ierr = PetscCalloc1(pN,&ptap->apa);CHKERRQ(ierr);
} else {
ierr = PetscCalloc1(ap_rmax+1,&ptap->apa);CHKERRQ(ierr);
}
#if defined(PTAP_PROFILE)
ierr = PetscTime(&t4);CHKERRQ(ierr);
if (rank==1) PetscPrintf(MPI_COMM_SELF," [%d] PtAPSymbolic %g/P + %g/AP + %g/comm + %g/PtAP = %g\n",rank,t1-t0,t2-t1,t3-t2,t4-t3,t4-t0);CHKERRQ(ierr);
#endif
#if defined(PETSC_USE_INFO)
if (pti[pn] != 0) {
ierr = PetscInfo3(Cmpi,"Reallocs %D; Fill ratio: given %G needed %G.\n",nspacedouble,fill,afill);CHKERRQ(ierr);
ierr = PetscInfo1(Cmpi,"Use MatPtAP(A,P,MatReuse,%G,&C) for best performance.\n",afill);CHKERRQ(ierr);
} else {
ierr = PetscInfo(Cmpi,"Empty matrix product\n");CHKERRQ(ierr);
}
#endif
PetscFunctionReturn(0);
}
