1632d0f97SHong Zhang /*$Id: sbaijov.c,v 1.65 2001/08/06 21:15:42 bsmith Exp $*/ 2632d0f97SHong Zhang 3632d0f97SHong Zhang /* 4632d0f97SHong Zhang Routines to compute overlapping regions of a parallel MPI matrix. 5632d0f97SHong Zhang Used for finding submatrices that were shared across processors. 6632d0f97SHong Zhang */ 7632d0f97SHong Zhang #include "src/mat/impls/sbaij/mpi/mpisbaij.h" 8632d0f97SHong Zhang #include "petscbt.h" 9632d0f97SHong Zhang 10632d0f97SHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat,int,IS *); 11bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat,int *,int,int **,PetscBT*); 12632d0f97SHong Zhang 13632d0f97SHong Zhang /* this function is sasme as MatCompressIndicesGeneral_MPIBAIJ -- should be removed! */ 14632d0f97SHong Zhang #undef __FUNCT__ 15632d0f97SHong Zhang #define __FUNCT__ "MatCompressIndicesGeneral_MPISBAIJ" 16632d0f97SHong Zhang static int MatCompressIndicesGeneral_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[]) 17632d0f97SHong Zhang { 18632d0f97SHong Zhang Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)C->data; 19632d0f97SHong Zhang int ierr,isz,bs = baij->bs,n,i,j,*idx,ival; 20632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 21632d0f97SHong Zhang PetscTable gid1_lid1; 22632d0f97SHong Zhang int tt, gid1, *nidx; 23632d0f97SHong Zhang PetscTablePosition tpos; 24632d0f97SHong Zhang #else 25632d0f97SHong Zhang int Nbs,*nidx; 26632d0f97SHong Zhang PetscBT table; 27632d0f97SHong Zhang #endif 28632d0f97SHong Zhang 29632d0f97SHong Zhang PetscFunctionBegin; 30632d0f97SHong Zhang /* printf(" ...MatCompressIndicesGeneral_MPISBAIJ is called ...\n"); */ 31632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 32632d0f97SHong Zhang ierr = PetscTableCreate(baij->mbs,&gid1_lid1);CHKERRQ(ierr); 33632d0f97SHong Zhang #else 34632d0f97SHong Zhang Nbs = baij->Nbs; 35632d0f97SHong Zhang ierr = PetscMalloc((Nbs+1)*sizeof(int),&nidx);CHKERRQ(ierr); 36632d0f97SHong Zhang ierr = PetscBTCreate(Nbs,table);CHKERRQ(ierr); 37632d0f97SHong Zhang #endif 38632d0f97SHong Zhang for (i=0; i<imax; i++) { 39632d0f97SHong Zhang isz = 0; 40632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 41632d0f97SHong Zhang ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr); 42632d0f97SHong Zhang #else 43632d0f97SHong Zhang ierr = PetscBTMemzero(Nbs,table);CHKERRQ(ierr); 44632d0f97SHong Zhang #endif 45632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 46632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 47632d0f97SHong Zhang for (j=0; j<n ; j++) { 48632d0f97SHong Zhang ival = idx[j]/bs; /* convert the indices into block indices */ 49632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 50632d0f97SHong Zhang ierr = PetscTableFind(gid1_lid1,ival+1,&tt);CHKERRQ(ierr); 51632d0f97SHong Zhang if (!tt) { 52632d0f97SHong Zhang ierr = PetscTableAdd(gid1_lid1,ival+1,isz+1);CHKERRQ(ierr); 53632d0f97SHong Zhang isz++; 54632d0f97SHong Zhang } 55632d0f97SHong Zhang #else 56632d0f97SHong Zhang if (ival>Nbs) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than mat-dim"); 57632d0f97SHong Zhang if(!PetscBTLookupSet(table,ival)) { nidx[isz++] = ival;} 58632d0f97SHong Zhang #endif 59632d0f97SHong Zhang } 60632d0f97SHong Zhang ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr); 61632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 62632d0f97SHong Zhang ierr = PetscMalloc((isz+1)*sizeof(int),&nidx);CHKERRQ(ierr); 63632d0f97SHong Zhang ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr); 64632d0f97SHong Zhang j = 0; 65632d0f97SHong Zhang while (tpos) { 66632d0f97SHong Zhang ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid1,&tt);CHKERRQ(ierr); 67632d0f97SHong Zhang if (tt-- > isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than array-dim"); } 68632d0f97SHong Zhang nidx[tt] = gid1 - 1; 69632d0f97SHong Zhang j++; 70632d0f97SHong Zhang } 71632d0f97SHong Zhang if (j != isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"table error: jj != isz"); } 72632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr); 73632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 74632d0f97SHong Zhang #else 75632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr); 76632d0f97SHong Zhang #endif 77632d0f97SHong Zhang } 78632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 79632d0f97SHong Zhang ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr); 80632d0f97SHong Zhang #else 81632d0f97SHong Zhang ierr = PetscBTDestroy(table);CHKERRQ(ierr); 82632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 83632d0f97SHong Zhang #endif 84632d0f97SHong Zhang PetscFunctionReturn(0); 85632d0f97SHong Zhang } 86632d0f97SHong Zhang 87632d0f97SHong Zhang #undef __FUNCT__ 88632d0f97SHong Zhang #define __FUNCT__ "MatExpandIndices_MPISBAIJ" 89632d0f97SHong Zhang static int MatExpandIndices_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[]) 90632d0f97SHong Zhang { 91632d0f97SHong Zhang Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)C->data; 92632d0f97SHong Zhang int ierr,bs = baij->bs,n,i,j,k,*idx,*nidx; 93632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 94632d0f97SHong Zhang int maxsz; 95632d0f97SHong Zhang #else 96632d0f97SHong Zhang int Nbs = baij->Nbs; 97632d0f97SHong Zhang #endif 98632d0f97SHong Zhang 99632d0f97SHong Zhang PetscFunctionBegin; 100632d0f97SHong Zhang /* printf(" ... MatExpandIndices_MPISBAIJ is called ...\n"); */ 101632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 102632d0f97SHong Zhang /* Now check max size */ 103632d0f97SHong Zhang for (i=0,maxsz=0; i<imax; i++) { 104632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 105632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 106632d0f97SHong Zhang if (n*bs > maxsz) maxsz = n*bs; 107632d0f97SHong Zhang } 108632d0f97SHong Zhang ierr = PetscMalloc((maxsz+1)*sizeof(int),&nidx);CHKERRQ(ierr); 109632d0f97SHong Zhang #else 110632d0f97SHong Zhang ierr = PetscMalloc((Nbs*bs+1)*sizeof(int),&nidx);CHKERRQ(ierr); 111632d0f97SHong Zhang #endif 112632d0f97SHong Zhang 113632d0f97SHong Zhang for (i=0; i<imax; i++) { 114632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 115632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 116632d0f97SHong Zhang for (j=0; j<n ; ++j){ 117632d0f97SHong Zhang for (k=0; k<bs; k++) 118632d0f97SHong Zhang nidx[j*bs+k] = idx[j]*bs+k; 119632d0f97SHong Zhang } 120632d0f97SHong Zhang ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr); 121632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,n*bs,nidx,is_out+i);CHKERRQ(ierr); 122632d0f97SHong Zhang } 123632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 124632d0f97SHong Zhang PetscFunctionReturn(0); 125632d0f97SHong Zhang } 126632d0f97SHong Zhang 127632d0f97SHong Zhang #undef __FUNCT__ 128632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ" 129c910923dSHong Zhang int MatIncreaseOverlap_MPISBAIJ(Mat C,int is_max,IS is[],int ov) 130632d0f97SHong Zhang { 131bfc6803cSHong Zhang Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data; 132bfc6803cSHong Zhang int i,ierr,size; 133632d0f97SHong Zhang IS *is_new; 134632d0f97SHong Zhang 135632d0f97SHong Zhang PetscFunctionBegin; 136c910923dSHong Zhang ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr); 137632d0f97SHong Zhang /* Convert the indices into block format */ 138c910923dSHong Zhang ierr = MatCompressIndicesGeneral_MPISBAIJ(C,is_max,is,is_new);CHKERRQ(ierr); 139632d0f97SHong Zhang if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");} 140632d0f97SHong Zhang for (i=0; i<ov; ++i) { 141c910923dSHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr); 142632d0f97SHong Zhang } 143c910923dSHong Zhang for (i=0; i<is_max; i++) {ierr = ISDestroy(is[i]);CHKERRQ(ierr);} 144c910923dSHong Zhang ierr = MatExpandIndices_MPISBAIJ(C,is_max,is_new,is);CHKERRQ(ierr); 145c910923dSHong Zhang for (i=0; i<is_max; i++) {ierr = ISDestroy(is_new[i]);CHKERRQ(ierr);} 146632d0f97SHong Zhang ierr = PetscFree(is_new);CHKERRQ(ierr); 147632d0f97SHong Zhang PetscFunctionReturn(0); 148632d0f97SHong Zhang } 149632d0f97SHong Zhang 150bfc6803cSHong Zhang typedef enum {MINE,OTHER} whose; 1510472cc68SHong Zhang /* data1, odata1 and odata2 are packed in the format (for communication): 152a2a9f22aSHong Zhang data[0] = is_max, no of is 153a2a9f22aSHong Zhang data[1] = size of is[0] 154a2a9f22aSHong Zhang ... 155a2a9f22aSHong Zhang data[is_max] = size of is[is_max-1] 156a2a9f22aSHong Zhang data[is_max + 1] = data(is[0]) 157a2a9f22aSHong Zhang ... 158a2a9f22aSHong Zhang data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i]) 159a2a9f22aSHong Zhang ... 1600472cc68SHong Zhang data2 is packed in the format (for creating output is[]): 1610472cc68SHong Zhang data[0] = is_max, no of is 1620472cc68SHong Zhang data[1] = size of is[0] 1630472cc68SHong Zhang ... 1640472cc68SHong Zhang data[is_max] = size of is[is_max-1] 1650472cc68SHong Zhang data[is_max + 1] = data(is[0]) 1660472cc68SHong Zhang ... 1670472cc68SHong Zhang data[is_max + 1 + Mbs*i) = data(is[i]) 1680472cc68SHong Zhang ... 169a2a9f22aSHong Zhang */ 170632d0f97SHong Zhang #undef __FUNCT__ 171632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Once" 172c910923dSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat C,int is_max,IS is[]) 173632d0f97SHong Zhang { 174632d0f97SHong Zhang Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data; 1750472cc68SHong Zhang int len,*idx_i,isz,col,*n,*data1,*data2,*data2_i,*data,*data_i, 1760472cc68SHong Zhang size,rank,Mbs,i,j,k,ierr,nrqs,*odata1,*odata2, 1770472cc68SHong Zhang tag1,tag2,flag,proc_id,**odata2_ptr; 178bfc6803cSHong Zhang char *t_p; 179632d0f97SHong Zhang MPI_Comm comm; 1800472cc68SHong Zhang MPI_Request *s_waits,r_req; 181632d0f97SHong Zhang MPI_Status *s_status,r_status; 182bfc6803cSHong Zhang PetscBT *table; /* mark indices of this processor's is[] */ 183fc70d14eSHong Zhang PetscBT table_i; 184bfc6803cSHong Zhang PetscBT otable; /* mark indices of other processors' is[] */ 185632d0f97SHong Zhang PetscFunctionBegin; 186632d0f97SHong Zhang 187632d0f97SHong Zhang comm = C->comm; 188632d0f97SHong Zhang size = c->size; 189632d0f97SHong Zhang rank = c->rank; 190632d0f97SHong Zhang Mbs = c->Mbs; 191632d0f97SHong Zhang 192c910923dSHong Zhang ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 193c910923dSHong Zhang ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 194c910923dSHong Zhang 195a2a9f22aSHong Zhang /* create tables for marking indices */ 196a2a9f22aSHong Zhang len = is_max*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*is_max*sizeof(char) + 1; 197bfc6803cSHong Zhang ierr = PetscMalloc(len,&table);CHKERRQ(ierr); 198a2a9f22aSHong Zhang t_p = (char *)(table + is_max); 199a2a9f22aSHong Zhang for (i=0; i<is_max; i++) { 200bfc6803cSHong Zhang table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i; 201a2a9f22aSHong Zhang } 2020472cc68SHong Zhang ierr = PetscBTCreate(Mbs,otable);CHKERRQ(ierr); 203632d0f97SHong Zhang 204bfc6803cSHong Zhang /* 1. Send this processor's is[] to all other processors */ 205bfc6803cSHong Zhang /*-------------------------------------------------------*/ 206632d0f97SHong Zhang /* Allocate Memory for outgoing messages */ 207a2a9f22aSHong Zhang len = is_max*sizeof(int); 208a2a9f22aSHong Zhang ierr = PetscMalloc(len,&n);CHKERRQ(ierr); 209a2a9f22aSHong Zhang 210c910923dSHong Zhang len = 1 + is_max; 211c910923dSHong Zhang for (i=0; i<is_max; i++) { 212632d0f97SHong Zhang ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr); 213632d0f97SHong Zhang len += n[i]; 214632d0f97SHong Zhang } 215a2a9f22aSHong Zhang ierr = PetscMalloc(len*sizeof(int),&data1);CHKERRQ(ierr); 216632d0f97SHong Zhang 217632d0f97SHong Zhang /* Form the outgoing messages */ 218a2a9f22aSHong Zhang data1[0] = is_max; 219c910923dSHong Zhang k = is_max + 1; 220a2a9f22aSHong Zhang for (i=0; i<is_max; i++) { 221a2a9f22aSHong Zhang data1[1+i] = n[i]; 222a2a9f22aSHong Zhang ierr = ISGetIndices(is[i],&idx_i);CHKERRQ(ierr); 223a2a9f22aSHong Zhang for (j=0; j<data1[i+1]; j++){ 224a2a9f22aSHong Zhang data1[k++] = idx_i[j]; 225632d0f97SHong Zhang } 226a2a9f22aSHong Zhang ierr = ISRestoreIndices(is[i],&idx_i);CHKERRQ(ierr); 227a2a9f22aSHong Zhang ierr = ISDestroy(is[i]);CHKERRQ(ierr); 228632d0f97SHong Zhang } 229fc70d14eSHong Zhang if (k != len) SETERRQ2(1,"Error on forming the outgoing messages: k %d != len %d",k,len); 230bfc6803cSHong Zhang ierr = PetscFree(n);CHKERRQ(ierr); 231632d0f97SHong Zhang 232632d0f97SHong Zhang /* Now post the sends */ 2330472cc68SHong Zhang ierr = PetscMalloc(size*sizeof(MPI_Request),&s_waits);CHKERRQ(ierr); 234632d0f97SHong Zhang k = 0; 235a2a9f22aSHong Zhang for (proc_id=0; proc_id<size; ++proc_id) { /* send data1 to processor [proc_id] */ 236c910923dSHong Zhang if (proc_id != rank){ 2370472cc68SHong Zhang ierr = MPI_Isend(data1,len,MPI_INT,proc_id,tag1,comm,s_waits+k);CHKERRQ(ierr); 238bfc6803cSHong Zhang /* printf(" [%d] send %d msg to [%d], data1: \n",rank,len,proc_id); */ 239632d0f97SHong Zhang k++; 240632d0f97SHong Zhang } 241632d0f97SHong Zhang } 242632d0f97SHong Zhang 243dc008846SHong Zhang /* 2. Do local work on this processor's is[] */ 244dc008846SHong Zhang /*-------------------------------------------*/ 2450472cc68SHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,data1,MINE,&data,table);CHKERRQ(ierr); 246632d0f97SHong Zhang 247632d0f97SHong Zhang /* 3. Receive other's is[] and process. Then send back */ 248bfc6803cSHong Zhang /*-----------------------------------------------------*/ 2490472cc68SHong Zhang /* Sending this processor's is[] is done */ 250632d0f97SHong Zhang nrqs = size-1; 251632d0f97SHong Zhang ierr = PetscMalloc(size*sizeof(MPI_Status),&s_status);CHKERRQ(ierr); 2520472cc68SHong Zhang ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr); 253dc008846SHong Zhang 254bfc6803cSHong Zhang ierr = PetscMalloc(size*sizeof(int**),&odata2_ptr);CHKERRQ(ierr); 255632d0f97SHong Zhang k = 0; 256632d0f97SHong Zhang do { 257632d0f97SHong Zhang /* Receive messages */ 258bfc6803cSHong Zhang ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);CHKERRQ(ierr); 259632d0f97SHong Zhang if (flag){ 260bfc6803cSHong Zhang ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr); 261632d0f97SHong Zhang proc_id = r_status.MPI_SOURCE; 262a2a9f22aSHong Zhang ierr = PetscMalloc(len*sizeof(int),&odata1);CHKERRQ(ierr); 263bfc6803cSHong Zhang ierr = MPI_Irecv(odata1,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr); 264fc70d14eSHong Zhang /* printf(" [%d] recv %d msg from [%d]\n",rank,len,proc_id); */ 265632d0f97SHong Zhang 266fc70d14eSHong Zhang /* Process messages */ 267bfc6803cSHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,&odata2_ptr[k],&otable);CHKERRQ(ierr); 268a2a9f22aSHong Zhang odata2 = odata2_ptr[k]; 269a2a9f22aSHong Zhang len = 1 + odata2[0]; 270a2a9f22aSHong Zhang for (i=0; i<odata2[0]; i++){ 271a2a9f22aSHong Zhang len += odata2[1 + i]; 272fc70d14eSHong Zhang } 273632d0f97SHong Zhang 274632d0f97SHong Zhang /* Send messages back */ 2750472cc68SHong Zhang ierr = MPI_Isend(odata2,len,MPI_INT,proc_id,tag2,comm,s_waits+k);CHKERRQ(ierr); 276bfc6803cSHong Zhang /* printf(" [%d] send %d msg back to [%d] \n",rank,len,proc_id); */ 277632d0f97SHong Zhang 278a2a9f22aSHong Zhang ierr = PetscFree(odata1);CHKERRQ(ierr); 279fc70d14eSHong Zhang k++; 280632d0f97SHong Zhang } 281632d0f97SHong Zhang } while (k < nrqs); 282632d0f97SHong Zhang 283fc70d14eSHong Zhang /* 4. Receive work done on other processors, then merge */ 284632d0f97SHong Zhang /*--------------------------------------------------------*/ 2850472cc68SHong Zhang /* Allocate memory for incoming data */ 286bfc6803cSHong Zhang len = (1+is_max*(Mbs+1)); 2870472cc68SHong Zhang ierr = PetscMalloc(len*sizeof(int),&data2);CHKERRQ(ierr); 288bfc6803cSHong Zhang 289bfc6803cSHong Zhang /* Sending others' is[] is done */ 2900472cc68SHong Zhang ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr); 291fc70d14eSHong Zhang for (k=0; k<nrqs; k++){ 292a2a9f22aSHong Zhang ierr = PetscFree(odata2_ptr[k]);CHKERRQ(ierr); 293fc70d14eSHong Zhang } 294a2a9f22aSHong Zhang ierr = PetscFree(odata2_ptr);CHKERRQ(ierr); 295fc70d14eSHong Zhang 296632d0f97SHong Zhang k = 0; 297632d0f97SHong Zhang do { 298632d0f97SHong Zhang /* Receive messages */ 299c910923dSHong Zhang ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status); 300632d0f97SHong Zhang if (flag){ 301bfc6803cSHong Zhang ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr); 302632d0f97SHong Zhang proc_id = r_status.MPI_SOURCE; 3030472cc68SHong Zhang ierr = MPI_Irecv(data2,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr); 3040472cc68SHong Zhang /* printf(" [%d] recv %d msg from [%d], data2:\n",rank,len,proc_id); */ 305632d0f97SHong Zhang 3060472cc68SHong Zhang /* Add data2 into data */ 3070472cc68SHong Zhang data2_i = data2 + 1 + is_max; 308fc70d14eSHong Zhang for (i=0; i<is_max; i++){ 309fc70d14eSHong Zhang table_i = table[i]; 310bfc6803cSHong Zhang data_i = data + 1 + is_max + Mbs*i; 311bfc6803cSHong Zhang isz = data[1+i]; 3120472cc68SHong Zhang for (j=0; j<data2[1+i]; j++){ 3130472cc68SHong Zhang col = data2_i[j]; 314bfc6803cSHong Zhang if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;} 315fc70d14eSHong Zhang } 316bfc6803cSHong Zhang data[1+i] = isz; 3170472cc68SHong Zhang if (i < is_max - 1) data2_i += data2[1+i]; 318fc70d14eSHong Zhang } 319632d0f97SHong Zhang k++; 320632d0f97SHong Zhang } 321632d0f97SHong Zhang } while (k < nrqs); 322632d0f97SHong Zhang 323c910923dSHong Zhang /* 5. Create new is[] */ 324c910923dSHong Zhang /*--------------------*/ 325c910923dSHong Zhang for (i=0; i<is_max; i++) { 326bfc6803cSHong Zhang data_i = data + 1 + is_max + Mbs*i; 327bfc6803cSHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,is+i);CHKERRQ(ierr); 328632d0f97SHong Zhang } 329bfc6803cSHong Zhang ierr = PetscFree(data1);CHKERRQ(ierr); 330a2a9f22aSHong Zhang ierr = PetscFree(data2);CHKERRQ(ierr); 3310472cc68SHong Zhang ierr = PetscFree(data);CHKERRQ(ierr); 3320472cc68SHong Zhang ierr = PetscFree(s_waits);CHKERRQ(ierr); 333632d0f97SHong Zhang ierr = PetscFree(s_status);CHKERRQ(ierr); 334fc70d14eSHong Zhang ierr = PetscFree(table);CHKERRQ(ierr); 335bfc6803cSHong Zhang ierr = PetscBTDestroy(otable);CHKERRQ(ierr); 336632d0f97SHong Zhang PetscFunctionReturn(0); 337632d0f97SHong Zhang } 338632d0f97SHong Zhang 339632d0f97SHong Zhang #undef __FUNCT__ 340632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Local" 341632d0f97SHong Zhang /* 342dc008846SHong Zhang MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do 343632d0f97SHong Zhang the work on the local processor. 344632d0f97SHong Zhang 345632d0f97SHong Zhang Inputs: 346632d0f97SHong Zhang C - MAT_MPISBAIJ; 347bfc6803cSHong Zhang data - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format. 348bfc6803cSHong Zhang whose - whose is[] to be processed, 349bfc6803cSHong Zhang MINE: this processor's is[] 350bfc6803cSHong Zhang OTHER: other processor's is[] 351632d0f97SHong Zhang Output: 3520472cc68SHong Zhang data_new - whose = MINE: 3530472cc68SHong Zhang holds input and newly found indices in the same format as data 3540472cc68SHong Zhang whose = OTHER: 3550472cc68SHong Zhang only holds the newly found indices 3560472cc68SHong Zhang table - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'. 357632d0f97SHong Zhang */ 358bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat C,int *data,int whose,int **data_new,PetscBT *table) 359632d0f97SHong Zhang { 360632d0f97SHong Zhang Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data; 361dc008846SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data; 362dc008846SHong Zhang Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(c->B)->data; 363*31f99336SHong Zhang int ierr,row,mbs,Mbs,*nidx,*nidx_i,col,isz,isz0,*ai,*aj,*bi,*bj,*garray,rstart,l; 364dc008846SHong Zhang int a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n; 365bfc6803cSHong Zhang PetscBT table0; /* mark the indices of input is[] for look up */ 366bfc6803cSHong Zhang PetscBT table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */ 367632d0f97SHong Zhang 368632d0f97SHong Zhang PetscFunctionBegin; 369*31f99336SHong Zhang Mbs = c->Mbs; mbs = a->mbs; 370dc008846SHong Zhang ai = a->i; aj = a->j; 371dc008846SHong Zhang bi = b->i; bj = b->j; 372632d0f97SHong Zhang garray = c->garray; 373dc008846SHong Zhang rstart = c->rstart; 374dc008846SHong Zhang is_max = data[0]; 375632d0f97SHong Zhang 376fc70d14eSHong Zhang ierr = PetscBTCreate(Mbs,table0);CHKERRQ(ierr); 377fc70d14eSHong Zhang 378fc70d14eSHong Zhang ierr = PetscMalloc((1+is_max*(Mbs+1))*sizeof(int),&nidx);CHKERRQ(ierr); 379fc70d14eSHong Zhang nidx[0] = is_max; 380fc70d14eSHong Zhang 381fc70d14eSHong Zhang idx_i = data + is_max + 1; /* ptr to input is[0] array */ 382bfc6803cSHong Zhang nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */ 383dc008846SHong Zhang for (i=0; i<is_max; i++) { /* for each is */ 384dc008846SHong Zhang isz = 0; 385fc70d14eSHong Zhang n = data[1+i]; /* size of input is[i] */ 386dc008846SHong Zhang 387bfc6803cSHong Zhang /* initialize table_i, set table0 */ 388bfc6803cSHong Zhang if (whose == MINE){ /* process this processor's is[] */ 389bfc6803cSHong Zhang table_i = table[i]; 3900472cc68SHong Zhang nidx_i = nidx + 1+ is_max + Mbs*i; 391bfc6803cSHong Zhang } else { /* process other processor's is[] - only use one temp table */ 392bfc6803cSHong Zhang table_i = *table; 393bfc6803cSHong Zhang } 394bfc6803cSHong Zhang ierr = PetscBTMemzero(Mbs,table_i);CHKERRQ(ierr); 395bfc6803cSHong Zhang ierr = PetscBTMemzero(Mbs,table0);CHKERRQ(ierr); 396fc70d14eSHong Zhang if (n > 0) { 397bfc6803cSHong Zhang isz0 = 0; 398dc008846SHong Zhang for (j=0; j<n; j++){ 399dc008846SHong Zhang col = idx_i[j]; 400dc008846SHong Zhang if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %d >= Mbs %d",col,Mbs); 401bfc6803cSHong Zhang if(!PetscBTLookupSet(table_i,col)) { 402bfc6803cSHong Zhang ierr = PetscBTSet(table0,col);CHKERRQ(ierr); 4030472cc68SHong Zhang if (whose == MINE) {nidx_i[isz0] = col;} 404bfc6803cSHong Zhang isz0++; 405bfc6803cSHong Zhang } 406632d0f97SHong Zhang } 407dc008846SHong Zhang 4080472cc68SHong Zhang if (whose == MINE) {isz = isz0;} 409fc70d14eSHong Zhang k = 0; /* no. of indices from input is[i] that have been examined */ 410dc008846SHong Zhang for (row=0; row<mbs; row++){ 411dc008846SHong Zhang a_start = ai[row]; a_end = ai[row+1]; 412dc008846SHong Zhang b_start = bi[row]; b_end = bi[row+1]; 4130472cc68SHong Zhang if (PetscBTLookup(table0,row+rstart)){ /* row is on input is[i]: 4140472cc68SHong Zhang do row search: collect all col in this row */ 415dc008846SHong Zhang for (l = a_start; l<a_end ; l++){ /* Amat */ 416dc008846SHong Zhang col = aj[l] + rstart; 417fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;} 418dc008846SHong Zhang } 419dc008846SHong Zhang for (l = b_start; l<b_end ; l++){ /* Bmat */ 420dc008846SHong Zhang col = garray[bj[l]]; 421fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;} 422dc008846SHong Zhang } 423dc008846SHong Zhang k++; 424dc008846SHong Zhang if (k >= isz0) break; /* for (row=0; row<mbs; row++) */ 4250472cc68SHong Zhang } else { /* row is not on input is[i]: 4260472cc68SHong Zhang do col serach: add row onto nidx_i if there is a col in nidx_i */ 427dc008846SHong Zhang for (l = a_start; l<a_end ; l++){ /* Amat */ 428dc008846SHong Zhang col = aj[l] + rstart; 429dc008846SHong Zhang if (PetscBTLookup(table0,col)){ 430fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;} 431dc008846SHong Zhang break; /* for l = start; l<end ; l++) */ 432632d0f97SHong Zhang } 433632d0f97SHong Zhang } 434dc008846SHong Zhang for (l = b_start; l<b_end ; l++){ /* Bmat */ 435dc008846SHong Zhang col = garray[bj[l]]; 436dc008846SHong Zhang if (PetscBTLookup(table0,col)){ 437fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;} 438dc008846SHong Zhang break; /* for l = start; l<end ; l++) */ 439632d0f97SHong Zhang } 440dc008846SHong Zhang } 441dc008846SHong Zhang } 442bfc6803cSHong Zhang } 443fc70d14eSHong Zhang } /* if (n > 0) */ 444dc008846SHong Zhang 445dc008846SHong Zhang if (i < is_max - 1){ 446fc70d14eSHong Zhang idx_i += n; /* ptr to input is[i+1] array */ 447bfc6803cSHong Zhang nidx_i += isz; /* ptr to output is[i+1] array */ 448dc008846SHong Zhang } 449fc70d14eSHong Zhang nidx[1+i] = isz; /* size of new is[i] */ 450dc008846SHong Zhang } /* /* for each is */ 451dc008846SHong Zhang *data_new = nidx; 452dc008846SHong Zhang ierr = PetscBTDestroy(table0);CHKERRQ(ierr); 453632d0f97SHong Zhang 454632d0f97SHong Zhang PetscFunctionReturn(0); 455632d0f97SHong Zhang } 456632d0f97SHong Zhang 457632d0f97SHong Zhang 458