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 *); 11*fc70d14eSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat,int *,int **,PetscBT*); 12632d0f97SHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Receive(Mat,int,int **,int**,int*); 13632d0f97SHong Zhang 14632d0f97SHong Zhang /* this function is sasme as MatCompressIndicesGeneral_MPIBAIJ -- should be removed! */ 15632d0f97SHong Zhang #undef __FUNCT__ 16632d0f97SHong Zhang #define __FUNCT__ "MatCompressIndicesGeneral_MPISBAIJ" 17632d0f97SHong Zhang static int MatCompressIndicesGeneral_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[]) 18632d0f97SHong Zhang { 19632d0f97SHong Zhang Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)C->data; 20632d0f97SHong Zhang int ierr,isz,bs = baij->bs,n,i,j,*idx,ival; 21632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 22632d0f97SHong Zhang PetscTable gid1_lid1; 23632d0f97SHong Zhang int tt, gid1, *nidx; 24632d0f97SHong Zhang PetscTablePosition tpos; 25632d0f97SHong Zhang #else 26632d0f97SHong Zhang int Nbs,*nidx; 27632d0f97SHong Zhang PetscBT table; 28632d0f97SHong Zhang #endif 29632d0f97SHong Zhang 30632d0f97SHong Zhang PetscFunctionBegin; 31632d0f97SHong Zhang /* printf(" ...MatCompressIndicesGeneral_MPISBAIJ is called ...\n"); */ 32632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 33632d0f97SHong Zhang ierr = PetscTableCreate(baij->mbs,&gid1_lid1);CHKERRQ(ierr); 34632d0f97SHong Zhang #else 35632d0f97SHong Zhang Nbs = baij->Nbs; 36632d0f97SHong Zhang ierr = PetscMalloc((Nbs+1)*sizeof(int),&nidx);CHKERRQ(ierr); 37632d0f97SHong Zhang ierr = PetscBTCreate(Nbs,table);CHKERRQ(ierr); 38632d0f97SHong Zhang #endif 39632d0f97SHong Zhang for (i=0; i<imax; i++) { 40632d0f97SHong Zhang isz = 0; 41632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 42632d0f97SHong Zhang ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr); 43632d0f97SHong Zhang #else 44632d0f97SHong Zhang ierr = PetscBTMemzero(Nbs,table);CHKERRQ(ierr); 45632d0f97SHong Zhang #endif 46632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 47632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 48632d0f97SHong Zhang for (j=0; j<n ; j++) { 49632d0f97SHong Zhang ival = idx[j]/bs; /* convert the indices into block indices */ 50632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 51632d0f97SHong Zhang ierr = PetscTableFind(gid1_lid1,ival+1,&tt);CHKERRQ(ierr); 52632d0f97SHong Zhang if (!tt) { 53632d0f97SHong Zhang ierr = PetscTableAdd(gid1_lid1,ival+1,isz+1);CHKERRQ(ierr); 54632d0f97SHong Zhang isz++; 55632d0f97SHong Zhang } 56632d0f97SHong Zhang #else 57632d0f97SHong Zhang if (ival>Nbs) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than mat-dim"); 58632d0f97SHong Zhang if(!PetscBTLookupSet(table,ival)) { nidx[isz++] = ival;} 59632d0f97SHong Zhang #endif 60632d0f97SHong Zhang } 61632d0f97SHong Zhang ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr); 62632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 63632d0f97SHong Zhang ierr = PetscMalloc((isz+1)*sizeof(int),&nidx);CHKERRQ(ierr); 64632d0f97SHong Zhang ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr); 65632d0f97SHong Zhang j = 0; 66632d0f97SHong Zhang while (tpos) { 67632d0f97SHong Zhang ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid1,&tt);CHKERRQ(ierr); 68632d0f97SHong Zhang if (tt-- > isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than array-dim"); } 69632d0f97SHong Zhang nidx[tt] = gid1 - 1; 70632d0f97SHong Zhang j++; 71632d0f97SHong Zhang } 72632d0f97SHong Zhang if (j != isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"table error: jj != isz"); } 73632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr); 74632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 75632d0f97SHong Zhang #else 76632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr); 77632d0f97SHong Zhang #endif 78632d0f97SHong Zhang } 79632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 80632d0f97SHong Zhang ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr); 81632d0f97SHong Zhang #else 82632d0f97SHong Zhang ierr = PetscBTDestroy(table);CHKERRQ(ierr); 83632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 84632d0f97SHong Zhang #endif 85632d0f97SHong Zhang PetscFunctionReturn(0); 86632d0f97SHong Zhang } 87632d0f97SHong Zhang 88632d0f97SHong Zhang #undef __FUNCT__ 89632d0f97SHong Zhang #define __FUNCT__ "MatCompressIndicesSorted_MPISBAIJ" 90632d0f97SHong Zhang static int MatCompressIndicesSorted_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[]) 91632d0f97SHong Zhang { 92632d0f97SHong Zhang Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)C->data; 93632d0f97SHong Zhang int ierr,bs=baij->bs,i,j,k,val,n,*idx,*nidx,*idx_local; 94632d0f97SHong Zhang PetscTruth flg; 95632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 96632d0f97SHong Zhang int maxsz; 97632d0f97SHong Zhang #else 98632d0f97SHong Zhang int Nbs=baij->Nbs; 99632d0f97SHong Zhang #endif 100632d0f97SHong Zhang PetscFunctionBegin; 101632d0f97SHong Zhang printf(" ... MatCompressIndicesSorted_MPISBAIJ is called ...\n"); 102632d0f97SHong Zhang for (i=0; i<imax; i++) { 103632d0f97SHong Zhang ierr = ISSorted(is_in[i],&flg);CHKERRQ(ierr); 104632d0f97SHong Zhang if (!flg) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Indices are not sorted"); 105632d0f97SHong Zhang } 106632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 107632d0f97SHong Zhang /* Now check max size */ 108632d0f97SHong Zhang for (i=0,maxsz=0; i<imax; i++) { 109632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 110632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 111632d0f97SHong Zhang if (n%bs !=0) SETERRQ(1,"Indices are not block ordered"); 112632d0f97SHong Zhang n = n/bs; /* The reduced index size */ 113632d0f97SHong Zhang if (n > maxsz) maxsz = n; 114632d0f97SHong Zhang } 115632d0f97SHong Zhang ierr = PetscMalloc((maxsz+1)*sizeof(int),&nidx);CHKERRQ(ierr); 116632d0f97SHong Zhang #else 117632d0f97SHong Zhang ierr = PetscMalloc((Nbs+1)*sizeof(int),&nidx);CHKERRQ(ierr); 118632d0f97SHong Zhang #endif 119632d0f97SHong Zhang /* Now check if the indices are in block order */ 120632d0f97SHong Zhang for (i=0; i<imax; i++) { 121632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 122632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 123632d0f97SHong Zhang if (n%bs !=0) SETERRQ(1,"Indices are not block ordered"); 124632d0f97SHong Zhang 125632d0f97SHong Zhang n = n/bs; /* The reduced index size */ 126632d0f97SHong Zhang idx_local = idx; 127632d0f97SHong Zhang for (j=0; j<n ; j++) { 128632d0f97SHong Zhang val = idx_local[0]; 129632d0f97SHong Zhang if (val%bs != 0) SETERRQ(1,"Indices are not block ordered"); 130632d0f97SHong Zhang for (k=0; k<bs; k++) { 131632d0f97SHong Zhang if (val+k != idx_local[k]) SETERRQ(1,"Indices are not block ordered"); 132632d0f97SHong Zhang } 133632d0f97SHong Zhang nidx[j] = val/bs; 134632d0f97SHong Zhang idx_local +=bs; 135632d0f97SHong Zhang } 136632d0f97SHong Zhang ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr); 137632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,n,nidx,(is_out+i));CHKERRQ(ierr); 138632d0f97SHong Zhang } 139632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 140632d0f97SHong Zhang 141632d0f97SHong Zhang PetscFunctionReturn(0); 142632d0f97SHong Zhang } 143632d0f97SHong Zhang 144632d0f97SHong Zhang #undef __FUNCT__ 145632d0f97SHong Zhang #define __FUNCT__ "MatExpandIndices_MPISBAIJ" 146632d0f97SHong Zhang static int MatExpandIndices_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[]) 147632d0f97SHong Zhang { 148632d0f97SHong Zhang Mat_MPISBAIJ *baij = (Mat_MPISBAIJ*)C->data; 149632d0f97SHong Zhang int ierr,bs = baij->bs,n,i,j,k,*idx,*nidx; 150632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 151632d0f97SHong Zhang int maxsz; 152632d0f97SHong Zhang #else 153632d0f97SHong Zhang int Nbs = baij->Nbs; 154632d0f97SHong Zhang #endif 155632d0f97SHong Zhang 156632d0f97SHong Zhang PetscFunctionBegin; 157632d0f97SHong Zhang /* printf(" ... MatExpandIndices_MPISBAIJ is called ...\n"); */ 158632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE) 159632d0f97SHong Zhang /* Now check max size */ 160632d0f97SHong Zhang for (i=0,maxsz=0; i<imax; i++) { 161632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 162632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 163632d0f97SHong Zhang if (n*bs > maxsz) maxsz = n*bs; 164632d0f97SHong Zhang } 165632d0f97SHong Zhang ierr = PetscMalloc((maxsz+1)*sizeof(int),&nidx);CHKERRQ(ierr); 166632d0f97SHong Zhang #else 167632d0f97SHong Zhang ierr = PetscMalloc((Nbs*bs+1)*sizeof(int),&nidx);CHKERRQ(ierr); 168632d0f97SHong Zhang #endif 169632d0f97SHong Zhang 170632d0f97SHong Zhang for (i=0; i<imax; i++) { 171632d0f97SHong Zhang ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr); 172632d0f97SHong Zhang ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr); 173632d0f97SHong Zhang for (j=0; j<n ; ++j){ 174632d0f97SHong Zhang for (k=0; k<bs; k++) 175632d0f97SHong Zhang nidx[j*bs+k] = idx[j]*bs+k; 176632d0f97SHong Zhang } 177632d0f97SHong Zhang ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr); 178632d0f97SHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,n*bs,nidx,is_out+i);CHKERRQ(ierr); 179632d0f97SHong Zhang } 180632d0f97SHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 181632d0f97SHong Zhang PetscFunctionReturn(0); 182632d0f97SHong Zhang } 183632d0f97SHong Zhang 184632d0f97SHong Zhang 185632d0f97SHong Zhang #undef __FUNCT__ 186632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ" 187c910923dSHong Zhang int MatIncreaseOverlap_MPISBAIJ(Mat C,int is_max,IS is[],int ov) 188632d0f97SHong Zhang { 189632d0f97SHong Zhang int i,ierr; 190632d0f97SHong Zhang IS *is_new; 191632d0f97SHong Zhang 192632d0f97SHong Zhang PetscFunctionBegin; 193c910923dSHong Zhang ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr); 194632d0f97SHong Zhang /* Convert the indices into block format */ 195c910923dSHong Zhang ierr = MatCompressIndicesGeneral_MPISBAIJ(C,is_max,is,is_new);CHKERRQ(ierr); 196632d0f97SHong Zhang if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");} 197632d0f97SHong Zhang for (i=0; i<ov; ++i) { 198c910923dSHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr); 199632d0f97SHong Zhang } 200c910923dSHong Zhang for (i=0; i<is_max; i++) {ierr = ISDestroy(is[i]);CHKERRQ(ierr);} 201c910923dSHong Zhang ierr = MatExpandIndices_MPISBAIJ(C,is_max,is_new,is);CHKERRQ(ierr); 202c910923dSHong Zhang for (i=0; i<is_max; i++) {ierr = ISDestroy(is_new[i]);CHKERRQ(ierr);} 203632d0f97SHong Zhang ierr = PetscFree(is_new);CHKERRQ(ierr); 204632d0f97SHong Zhang PetscFunctionReturn(0); 205632d0f97SHong Zhang } 206632d0f97SHong Zhang 207632d0f97SHong Zhang #undef __FUNCT__ 208632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Once" 209c910923dSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat C,int is_max,IS is[]) 210632d0f97SHong Zhang { 211632d0f97SHong Zhang Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data; 212*fc70d14eSHong Zhang int **idx,*n,len,*idx_i,*nidx,*nidx_i,isz,col; 213dc008846SHong Zhang int size,rank,Mbs,i,j,k,ierr,nrqs,msz,*outdat,*indat; 214dc008846SHong Zhang int tag1,tag2,flag,proc_id; 215632d0f97SHong Zhang MPI_Comm comm; 216c910923dSHong Zhang MPI_Request *s_waits1,*s_waits2,r_req; 217632d0f97SHong Zhang MPI_Status *s_status,r_status; 218*fc70d14eSHong Zhang PetscBT *table; 219*fc70d14eSHong Zhang PetscBT table_i; 220*fc70d14eSHong Zhang PetscBT *table_tmp; 221632d0f97SHong Zhang 222632d0f97SHong Zhang PetscFunctionBegin; 223632d0f97SHong Zhang 224632d0f97SHong Zhang comm = C->comm; 225632d0f97SHong Zhang size = c->size; 226632d0f97SHong Zhang rank = c->rank; 227632d0f97SHong Zhang Mbs = c->Mbs; 228632d0f97SHong Zhang 229*fc70d14eSHong Zhang /* int prid=0; */ 230*fc70d14eSHong Zhang 231*fc70d14eSHong Zhang len = is_max*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*is_max*sizeof(char) + 1; 232*fc70d14eSHong Zhang ierr = PetscMalloc(len,&table);CHKERRQ(ierr); 233*fc70d14eSHong Zhang char *t_p; 234*fc70d14eSHong Zhang t_p = (char *)(table + is_max); 235*fc70d14eSHong Zhang for (i=0; i<is_max; i++) { 236*fc70d14eSHong Zhang table[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i; 237*fc70d14eSHong Zhang } 238*fc70d14eSHong Zhang 239*fc70d14eSHong Zhang ierr = PetscMalloc(len,&table_tmp);CHKERRQ(ierr); 240*fc70d14eSHong Zhang t_p = (char *)(table_tmp + is_max); 241*fc70d14eSHong Zhang for (i=0; i<is_max; i++) { 242*fc70d14eSHong Zhang table_tmp[i] = t_p + (Mbs/PETSC_BITS_PER_BYTE+1)*i; 243*fc70d14eSHong Zhang } 244*fc70d14eSHong Zhang 245632d0f97SHong Zhang /* 1. Send is[] to all other processors */ 246632d0f97SHong Zhang /*--------------------------------------*/ 247632d0f97SHong Zhang /* This processor sends its is[] to all other processors in the format: 248632d0f97SHong Zhang outdat[0] = is_max, no of is in this processor 249632d0f97SHong Zhang outdat[1] = n[0], size of is[0] 250632d0f97SHong Zhang ... 251632d0f97SHong Zhang outdat[is_max] = n[is_max-1], size of is[is_max-1] 252632d0f97SHong Zhang outdat[is_max + 1] = data(is[0]) 253632d0f97SHong Zhang ... 254dc008846SHong Zhang outdat[is_max+1+sum(n[k]), k=0,...,i-1] = data(is[i]) 255632d0f97SHong Zhang ... 256632d0f97SHong Zhang */ 257c910923dSHong Zhang ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr); 258c910923dSHong Zhang ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr); 259c910923dSHong Zhang 260c910923dSHong Zhang len = (is_max+1)*sizeof(int*)+ (is_max)*sizeof(int); 261632d0f97SHong Zhang ierr = PetscMalloc(len,&idx);CHKERRQ(ierr); 262c910923dSHong Zhang n = (int*)(idx + is_max); 263632d0f97SHong Zhang 264632d0f97SHong Zhang /* Allocate Memory for outgoing messages */ 265c910923dSHong Zhang len = 1 + is_max; 266c910923dSHong Zhang for (i=0; i<is_max; i++) { 267632d0f97SHong Zhang ierr = ISGetIndices(is[i],&idx[i]);CHKERRQ(ierr); 268632d0f97SHong Zhang ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr); 269632d0f97SHong Zhang len += n[i]; 270632d0f97SHong Zhang } 271632d0f97SHong Zhang ierr = PetscMalloc(len*sizeof(int),&outdat);CHKERRQ(ierr); 272632d0f97SHong Zhang 273632d0f97SHong Zhang /* Form the outgoing messages */ 274c910923dSHong Zhang outdat[0] = is_max; 275c910923dSHong Zhang for (i=0; i<is_max; i++) { 276632d0f97SHong Zhang outdat[i+1] = n[i]; 277632d0f97SHong Zhang } 278c910923dSHong Zhang k = is_max + 1; 279c910923dSHong Zhang for (i=0; i<is_max; i++) { /* for is[i] */ 280632d0f97SHong Zhang idx_i = idx[i]; 281632d0f97SHong Zhang for (j=0; j<n[i]; j++){ 282632d0f97SHong Zhang outdat[k] = *(idx_i); 283632d0f97SHong Zhang k++; idx_i++; 284632d0f97SHong Zhang } 285632d0f97SHong Zhang } 286*fc70d14eSHong Zhang if (k != len) SETERRQ2(1,"Error on forming the outgoing messages: k %d != len %d",k,len); 287632d0f97SHong Zhang 288632d0f97SHong Zhang /* Now post the sends */ 289632d0f97SHong Zhang ierr = PetscMalloc(size*sizeof(MPI_Request),&s_waits1);CHKERRQ(ierr); 290632d0f97SHong Zhang 291632d0f97SHong Zhang k = 0; 292c910923dSHong Zhang for (proc_id=0; proc_id<size; ++proc_id) { /* send outdat to processor [proc_id] */ 293c910923dSHong Zhang if (proc_id != rank){ 294c910923dSHong Zhang ierr = MPI_Isend(outdat,len,MPI_INT,proc_id,tag1,comm,s_waits1+k);CHKERRQ(ierr); 295*fc70d14eSHong Zhang /* printf(" [%d] send %d msg to [%d] \n",rank,len,proc_id); */ 296632d0f97SHong Zhang k++; 297632d0f97SHong Zhang } 298632d0f97SHong Zhang } 299632d0f97SHong Zhang 300dc008846SHong Zhang /* 2. Do local work on this processor's is[] */ 301dc008846SHong Zhang /*-------------------------------------------*/ 302*fc70d14eSHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,outdat,&nidx,table);CHKERRQ(ierr); 303632d0f97SHong Zhang 304dc008846SHong Zhang for (i=0; i<is_max; i++){ 305c910923dSHong Zhang ierr = ISRestoreIndices(is[i],idx+i);CHKERRQ(ierr); 306c910923dSHong Zhang ierr = ISDestroy(is[i]);CHKERRQ(ierr); 307dc008846SHong Zhang } 308632d0f97SHong Zhang 309632d0f97SHong Zhang /* 3. Receive other's is[] and process. Then send back */ 310632d0f97SHong Zhang /*----------------------------------------------------*/ 311632d0f97SHong Zhang /* Send is done */ 312632d0f97SHong Zhang nrqs = size-1; 313632d0f97SHong Zhang ierr = PetscMalloc(size*sizeof(MPI_Status),&s_status);CHKERRQ(ierr); 314632d0f97SHong Zhang ierr = MPI_Waitall(nrqs,s_waits1,s_status);CHKERRQ(ierr); 315dc008846SHong Zhang 316dc008846SHong Zhang /* save n[i] */ 317dc008846SHong Zhang for (i=0; i<is_max; i++){ 318dc008846SHong Zhang n[i] = outdat[1+i]; 319dc008846SHong Zhang } 320632d0f97SHong Zhang ierr = PetscFree(outdat);CHKERRQ(ierr); 321632d0f97SHong Zhang ierr = PetscMalloc(size*sizeof(MPI_Request),&s_waits2);CHKERRQ(ierr); 322*fc70d14eSHong Zhang int **outdat_ptr; 323*fc70d14eSHong Zhang ierr = PetscMalloc(size*sizeof(int**),&outdat_ptr); 324632d0f97SHong Zhang k = 0; 325632d0f97SHong Zhang do { 326632d0f97SHong Zhang /* Receive messages */ 327c910923dSHong Zhang ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status); 328632d0f97SHong Zhang if (flag){ 329632d0f97SHong Zhang ierr = MPI_Get_count(&r_status,MPI_INT,&len); 330632d0f97SHong Zhang proc_id = r_status.MPI_SOURCE; 331c910923dSHong Zhang ierr = PetscMalloc(len*sizeof(int),&indat);CHKERRQ(ierr); 332c910923dSHong Zhang ierr = MPI_Irecv(indat,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req); 333*fc70d14eSHong Zhang /* printf(" [%d] recv %d msg from [%d]\n",rank,len,proc_id); */ 334632d0f97SHong Zhang 335*fc70d14eSHong Zhang /* Process messages */ 336*fc70d14eSHong Zhang ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,indat,&outdat_ptr[k],table_tmp);CHKERRQ(ierr); 337*fc70d14eSHong Zhang outdat = outdat_ptr[k]; 338*fc70d14eSHong Zhang len = 1 + outdat[0]; 339*fc70d14eSHong Zhang for (i=0; i<outdat[0]; i++){ 340*fc70d14eSHong Zhang len += outdat[1 + i]; 341*fc70d14eSHong Zhang } 342632d0f97SHong Zhang 343632d0f97SHong Zhang /* Send messages back */ 344*fc70d14eSHong Zhang /* printf(" [%d] send %d msg back to [%d] \n",rank,len,proc_id); */ 345c910923dSHong Zhang ierr = MPI_Isend(outdat,len,MPI_INT,proc_id,tag2,comm,&s_waits2[k]);CHKERRQ(ierr); 346632d0f97SHong Zhang 347632d0f97SHong Zhang ierr = PetscFree(indat);CHKERRQ(ierr); 348*fc70d14eSHong Zhang k++; 349632d0f97SHong Zhang } 350632d0f97SHong Zhang } while (k < nrqs); 351632d0f97SHong Zhang 352*fc70d14eSHong Zhang /* 4. Receive work done on other processors, then merge */ 353632d0f97SHong Zhang /*--------------------------------------------------------*/ 354632d0f97SHong Zhang ierr = MPI_Waitall(nrqs,s_waits2,s_status);CHKERRQ(ierr); 355*fc70d14eSHong Zhang for (k=0; k<nrqs; k++){ 356*fc70d14eSHong Zhang ierr = PetscFree(outdat_ptr[k]);CHKERRQ(ierr); 357*fc70d14eSHong Zhang } 358*fc70d14eSHong Zhang ierr = PetscFree(outdat_ptr);CHKERRQ(ierr); 359*fc70d14eSHong Zhang 360*fc70d14eSHong Zhang /* allocate memory for merged data */ 361*fc70d14eSHong Zhang int *mydata,*mydata_i; 362*fc70d14eSHong Zhang ierr = PetscMalloc((1+is_max*(Mbs+1))*sizeof(int),&mydata);CHKERRQ(ierr); 363*fc70d14eSHong Zhang 364*fc70d14eSHong Zhang /* copy nidx into mydata */ 365*fc70d14eSHong Zhang k = is_max + 1; 366*fc70d14eSHong Zhang for (i=0; i<is_max; i++){ 367*fc70d14eSHong Zhang mydata[1+i] = nidx[1+i]; /* size of is[i] before merge */ 368*fc70d14eSHong Zhang mydata_i = mydata + 1 + is_max + Mbs*i; 369*fc70d14eSHong Zhang for (j=0; j<nidx[1+i]; j++){ 370*fc70d14eSHong Zhang mydata_i[j] = nidx[k++]; 371*fc70d14eSHong Zhang } 372*fc70d14eSHong Zhang } 373*fc70d14eSHong Zhang 374632d0f97SHong Zhang k = 0; 375632d0f97SHong Zhang do { 376632d0f97SHong Zhang /* Receive messages */ 377c910923dSHong Zhang ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status); 378632d0f97SHong Zhang if (flag){ 379632d0f97SHong Zhang ierr = MPI_Get_count(&r_status,MPI_INT,&len); 380632d0f97SHong Zhang proc_id = r_status.MPI_SOURCE; 381632d0f97SHong Zhang ierr = ierr = PetscMalloc(len*sizeof(int),&indat);CHKERRQ(ierr); 382c910923dSHong Zhang ierr = MPI_Irecv(indat,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req); 383*fc70d14eSHong Zhang /* printf(" [%d] recv %d msg from [%d]\n",rank,len,proc_id); */ 384632d0f97SHong Zhang 385*fc70d14eSHong Zhang /* merge indat into mydata */ 386*fc70d14eSHong Zhang nidx_i = indat + 1 + is_max; 387*fc70d14eSHong Zhang for (i=0; i<is_max; i++){ 388*fc70d14eSHong Zhang table_i = table[i]; 389*fc70d14eSHong Zhang mydata_i = mydata + 1 + is_max + Mbs*i; 390*fc70d14eSHong Zhang isz = mydata[1+i]; /* size of is[i] from nidx */ 391632d0f97SHong Zhang 392*fc70d14eSHong Zhang for (j=0; j<indat[1+i]; j++){ 393*fc70d14eSHong Zhang col = nidx_i[j]; 394*fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,col)) {mydata_i[isz++] = col;} 395*fc70d14eSHong Zhang } 396*fc70d14eSHong Zhang mydata[1+i] = isz; 397*fc70d14eSHong Zhang if (i < is_max - 1){ 398*fc70d14eSHong Zhang nidx_i += indat[1+i]; /* ptr to is[i+1] array from indat */ 399*fc70d14eSHong Zhang } 400*fc70d14eSHong Zhang } /* for (i=0; i<is_max; i++) */ 401632d0f97SHong Zhang 402632d0f97SHong Zhang k++; 403632d0f97SHong Zhang ierr = PetscFree(indat);CHKERRQ(ierr); 404632d0f97SHong Zhang } 405632d0f97SHong Zhang } while (k < nrqs); 406632d0f97SHong Zhang 407c910923dSHong Zhang /* 5. Create new is[] */ 408c910923dSHong Zhang /*--------------------*/ 409c910923dSHong Zhang for (i=0; i<is_max; i++) { 410*fc70d14eSHong Zhang mydata_i = mydata + 1 + is_max + Mbs*i; 411*fc70d14eSHong Zhang ierr = ISCreateGeneral(PETSC_COMM_SELF,mydata[1+i],mydata_i,is+i);CHKERRQ(ierr); 412632d0f97SHong Zhang } 413*fc70d14eSHong Zhang 414*fc70d14eSHong Zhang ierr = PetscFree(mydata);CHKERRQ(ierr); 415c910923dSHong Zhang ierr = PetscFree(nidx);CHKERRQ(ierr); 416c910923dSHong Zhang ierr = PetscFree(idx);CHKERRQ(ierr); 417632d0f97SHong Zhang ierr = PetscFree(s_waits1);CHKERRQ(ierr); 418632d0f97SHong Zhang ierr = PetscFree(s_waits2);CHKERRQ(ierr); 419632d0f97SHong Zhang ierr = PetscFree(s_status);CHKERRQ(ierr); 420*fc70d14eSHong Zhang ierr = PetscFree(table);CHKERRQ(ierr); 421*fc70d14eSHong Zhang ierr = PetscFree(table_tmp);CHKERRQ(ierr); 422632d0f97SHong Zhang PetscFunctionReturn(0); 423632d0f97SHong Zhang } 424632d0f97SHong Zhang 425632d0f97SHong Zhang #undef __FUNCT__ 426632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Local" 427632d0f97SHong Zhang /* 428dc008846SHong Zhang MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do 429632d0f97SHong Zhang the work on the local processor. 430632d0f97SHong Zhang 431632d0f97SHong Zhang Inputs: 432632d0f97SHong Zhang C - MAT_MPISBAIJ; 433dc008846SHong Zhang data - holds is[] in the format: 434dc008846SHong Zhang data[0] = is_max, no of is 435dc008846SHong Zhang data[1] = size of is[0] 436dc008846SHong Zhang ... 437dc008846SHong Zhang data[is_max] = size of is[is_max-1] 438dc008846SHong Zhang data[is_max + 1] = is[0] array 439dc008846SHong Zhang ... 440dc008846SHong Zhang data[is_max+1+sum(n[k]), k=0,...,i-1] = is[i] array 441dc008846SHong Zhang ... 442632d0f97SHong Zhang 443632d0f97SHong Zhang Output: 444dc008846SHong Zhang data_new - holds new is[] in the same format as data 445*fc70d14eSHong Zhang table - table[i]: mark the indices of is[i], i=0,...,is_max. 446632d0f97SHong Zhang */ 447*fc70d14eSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat C,int *data,int **data_new,PetscBT *table) 448632d0f97SHong Zhang { 449632d0f97SHong Zhang Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data; 450dc008846SHong Zhang Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data; 451dc008846SHong Zhang Mat_SeqBAIJ *b = (Mat_SeqBAIJ*)(c->B)->data; 452dc008846SHong Zhang int ierr,row,mbs,Mbs,*nidx,*nidx_i,col,isz,isz0,*ai,*aj,bs,*bi,*bj,*garray,rstart,l; 453dc008846SHong Zhang int a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n; 454dc008846SHong Zhang PetscBT table0; 455*fc70d14eSHong Zhang PetscBT table_i; /* poits to i-th table */ 456632d0f97SHong Zhang 457632d0f97SHong Zhang PetscFunctionBegin; 458dc008846SHong Zhang Mbs = c->Mbs; mbs = a->mbs; bs = a->bs; 459dc008846SHong Zhang ai = a->i; aj = a->j; 460dc008846SHong Zhang bi = b->i; bj = b->j; 461632d0f97SHong Zhang garray = c->garray; 462dc008846SHong Zhang rstart = c->rstart; 463dc008846SHong Zhang is_max = data[0]; 464632d0f97SHong Zhang 465*fc70d14eSHong Zhang /* int rank=c->rank,prid=0; */ /* for debugging */ 466632d0f97SHong Zhang 467*fc70d14eSHong Zhang ierr = PetscBTCreate(Mbs,table0);CHKERRQ(ierr); 468*fc70d14eSHong Zhang 469*fc70d14eSHong Zhang ierr = PetscMalloc((1+is_max*(Mbs+1))*sizeof(int),&nidx);CHKERRQ(ierr); 470*fc70d14eSHong Zhang nidx[0] = is_max; 471*fc70d14eSHong Zhang 472*fc70d14eSHong Zhang idx_i = data + is_max + 1; /* ptr to input is[0] array */ 473*fc70d14eSHong Zhang nidx_i = nidx + is_max + 1; /* ptr to active is[0] array */ 474dc008846SHong Zhang for (i=0; i<is_max; i++) { /* for each is */ 475dc008846SHong Zhang isz = 0; 476*fc70d14eSHong Zhang table_i = table[i]; 477*fc70d14eSHong Zhang ierr = PetscBTMemzero(Mbs,table_i);CHKERRQ(ierr); 478*fc70d14eSHong Zhang n = data[1+i]; /* size of input is[i] */ 479dc008846SHong Zhang 480*fc70d14eSHong Zhang if (n > 0) { 481*fc70d14eSHong Zhang 482*fc70d14eSHong Zhang /* Enter input is[i] into active is[i] */ 483dc008846SHong Zhang for (j=0; j<n; j++){ 484dc008846SHong Zhang col = idx_i[j]; 485dc008846SHong Zhang if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %d >= Mbs %d",col,Mbs); 486*fc70d14eSHong Zhang if(!PetscBTLookupSet(table_i,col)) { nidx_i[isz++] = col;} 487632d0f97SHong Zhang } 488dc008846SHong Zhang 489dc008846SHong Zhang /* set table0 for lookup */ 490*fc70d14eSHong Zhang ierr = PetscBTMemzero(Mbs,table0);CHKERRQ(ierr); 491*fc70d14eSHong Zhang for (l=0; l<isz; l++) PetscBTSet(table0,nidx_i[l]); 492dc008846SHong Zhang 493*fc70d14eSHong Zhang isz0 = isz; /* size of input is[i] after removing repeated indices */ 494*fc70d14eSHong Zhang k = 0; /* no. of indices from input is[i] that have been examined */ 495dc008846SHong Zhang for (row=0; row<mbs; row++){ 496dc008846SHong Zhang a_start = ai[row]; a_end = ai[row+1]; 497dc008846SHong Zhang b_start = bi[row]; b_end = bi[row+1]; 498dc008846SHong Zhang if (PetscBTLookup(table0,row+rstart)){ /* row is on nidx_i - row search: collect all col in this row */ 499dc008846SHong Zhang for (l = a_start; l<a_end ; l++){ /* Amat */ 500dc008846SHong Zhang col = aj[l] + rstart; 501*fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;} 502dc008846SHong Zhang } 503dc008846SHong Zhang for (l = b_start; l<b_end ; l++){ /* Bmat */ 504dc008846SHong Zhang col = garray[bj[l]]; 505*fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;} 506dc008846SHong Zhang } 507dc008846SHong Zhang k++; 508dc008846SHong Zhang if (k >= isz0) break; /* for (row=0; row<mbs; row++) */ 509dc008846SHong Zhang } else { /* row is not on nidx_i - col serach: add row onto nidx_i if there is a col in nidx_i */ 510dc008846SHong Zhang for (l = a_start; l<a_end ; l++){ /* Amat */ 511dc008846SHong Zhang col = aj[l] + rstart; 512dc008846SHong Zhang if (PetscBTLookup(table0,col)){ 513*fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;} 514dc008846SHong Zhang break; /* for l = start; l<end ; l++) */ 515632d0f97SHong Zhang } 516632d0f97SHong Zhang } 517dc008846SHong Zhang for (l = b_start; l<b_end ; l++){ /* Bmat */ 518dc008846SHong Zhang col = garray[bj[l]]; 519dc008846SHong Zhang if (PetscBTLookup(table0,col)){ 520*fc70d14eSHong Zhang if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;} 521dc008846SHong Zhang break; /* for l = start; l<end ; l++) */ 522632d0f97SHong Zhang } 523dc008846SHong Zhang } 524dc008846SHong Zhang } 525dc008846SHong Zhang } /* for (row=0; row<mbs; row++) */ 526*fc70d14eSHong Zhang } /* if (n > 0) */ 527dc008846SHong Zhang 528dc008846SHong Zhang if (i < is_max - 1){ 529*fc70d14eSHong Zhang idx_i += n; /* ptr to input is[i+1] array */ 530*fc70d14eSHong Zhang nidx_i += isz; /* ptr to active is[i+1] array */ 531dc008846SHong Zhang } 532*fc70d14eSHong Zhang nidx[1+i] = isz; /* size of new is[i] */ 533dc008846SHong Zhang } /* /* for each is */ 534dc008846SHong Zhang *data_new = nidx; 535dc008846SHong Zhang ierr = PetscBTDestroy(table0);CHKERRQ(ierr); 536632d0f97SHong Zhang 537632d0f97SHong Zhang PetscFunctionReturn(0); 538632d0f97SHong Zhang } 539632d0f97SHong Zhang 540632d0f97SHong Zhang 541