1cac129eeSSatish Balay #ifndef lint 2*218c64b6SSatish Balay static char vcid[] = "$Id: baij2.c,v 1.3 1996/04/30 23:04:35 balay Exp balay $"; 3cac129eeSSatish Balay #endif 4cac129eeSSatish Balay 5cac129eeSSatish Balay #include "baij.h" 6cac129eeSSatish Balay #include "petsc.h" 7736121d4SSatish Balay #include "src/inline/bitarray.h" 8cac129eeSSatish Balay 9736121d4SSatish Balay int MatIncreaseOverlap_SeqBAIJ(Mat A,int is_max,IS *is,int ov) 10a3192f15SSatish Balay { 11a3192f15SSatish Balay Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 12*218c64b6SSatish Balay int row, i,j,k,l,m,n, *idx,ierr, *nidx, isz, val, ival; 13*218c64b6SSatish Balay int start, end, *ai, *aj,bs,*nidx2; 14a3192f15SSatish Balay char *table; 15a3192f15SSatish Balay 16a3192f15SSatish Balay m = a->mbs; 17a3192f15SSatish Balay ai = a->i; 18a3192f15SSatish Balay aj = a->j; 19*218c64b6SSatish Balay bs = a->bs; 20a3192f15SSatish Balay 21a3192f15SSatish Balay if (ov < 0) SETERRQ(1,"MatIncreaseOverlap_SeqBAIJ: illegal overlap value used"); 22a3192f15SSatish Balay 23a3192f15SSatish Balay table = (char *) PetscMalloc((m/BITSPERBYTE +1)*sizeof(char)); CHKPTRQ(table); 24a3192f15SSatish Balay nidx = (int *) PetscMalloc((m+1)*sizeof(int)); CHKPTRQ(nidx); 25*218c64b6SSatish Balay nidx2 = (int *)PetscMalloc((a->m+1)*sizeof(int)); CHKPTRQ(nidx2); 26a3192f15SSatish Balay 27a3192f15SSatish Balay for ( i=0; i<is_max; i++ ) { 28a3192f15SSatish Balay /* Initialise the two local arrays */ 29a3192f15SSatish Balay isz = 0; 30a3192f15SSatish Balay PetscMemzero(table,(m/BITSPERBYTE +1)*sizeof(char)); 31a3192f15SSatish Balay 32a3192f15SSatish Balay /* Extract the indices, assume there can be duplicate entries */ 33a3192f15SSatish Balay ierr = ISGetIndices(is[i],&idx); CHKERRQ(ierr); 34a3192f15SSatish Balay ierr = ISGetSize(is[i],&n); CHKERRQ(ierr); 35a3192f15SSatish Balay 36a3192f15SSatish Balay /* Enter these into the temp arrays i.e mark table[row], enter row into new index */ 37a3192f15SSatish Balay for ( j=0; j<n ; ++j){ 38*218c64b6SSatish Balay ival = idx[j]/bs; /* convert the indices into block indices */ 39*218c64b6SSatish Balay if (ival>m) SETERRQ(1,"MatIncreaseOverlap_SeqBAIJ: index greater than mat-dim"); 40*218c64b6SSatish Balay if(!BT_LOOKUP(table, ival)) { nidx[isz++] = ival;} 41a3192f15SSatish Balay } 42a3192f15SSatish Balay ierr = ISRestoreIndices(is[i],&idx); CHKERRQ(ierr); 43a3192f15SSatish Balay ierr = ISDestroy(is[i]); CHKERRQ(ierr); 44a3192f15SSatish Balay 45a3192f15SSatish Balay k = 0; 46a3192f15SSatish Balay for ( j=0; j<ov; j++){ /* for each overlap*/ 47a3192f15SSatish Balay n = isz; 48a3192f15SSatish Balay for ( ; k<n ; k++){ /* do only those rows in nidx[k], which are not done yet */ 49a3192f15SSatish Balay row = nidx[k]; 50a3192f15SSatish Balay start = ai[row]; 51a3192f15SSatish Balay end = ai[row+1]; 52a3192f15SSatish Balay for ( l = start; l<end ; l++){ 53a3192f15SSatish Balay val = aj[l]; 54a3192f15SSatish Balay if (!BT_LOOKUP(table,val)) {nidx[isz++] = val;} 55a3192f15SSatish Balay } 56a3192f15SSatish Balay } 57a3192f15SSatish Balay } 58*218c64b6SSatish Balay /* expand the Index Set */ 59*218c64b6SSatish Balay for (j=0; j<isz; j++ ) { 60*218c64b6SSatish Balay for (k=0; k<bs; k++ ) 61*218c64b6SSatish Balay nidx2[j*bs+k] = nidx[j]*bs+k; 62*218c64b6SSatish Balay } 63*218c64b6SSatish Balay ierr = ISCreateSeq(MPI_COMM_SELF, isz*bs, nidx2, (is+i)); CHKERRQ(ierr); 64a3192f15SSatish Balay } 65a3192f15SSatish Balay PetscFree(table); 66a3192f15SSatish Balay PetscFree(nidx); 67*218c64b6SSatish Balay PetscFree(nidx2); 68a3192f15SSatish Balay return 0; 69a3192f15SSatish Balay } 70*218c64b6SSatish Balay int MatGetSubMatrix_SeqBAIJ_Private(Mat A,IS isrow,IS iscol,MatGetSubMatrixCall scall,Mat *B) 71736121d4SSatish Balay { 72736121d4SSatish Balay Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data,*c; 73736121d4SSatish Balay int nznew, *smap, i, k, kstart, kend, ierr, oldcols = a->nbs,*lens; 74*218c64b6SSatish Balay int row,mat_i,*mat_j,tcol,*mat_ilen; 75736121d4SSatish Balay int *irow, *icol, nrows, ncols,*ssmap,bs=a->bs, bs2=a->bs2; 76*218c64b6SSatish Balay int *aj = a->j, *ai = a->i; 77*218c64b6SSatish Balay Scalar *mat_a; 78736121d4SSatish Balay Mat C; 79736121d4SSatish Balay 80736121d4SSatish Balay ierr = ISSorted(iscol,(PetscTruth*)&i); 81736121d4SSatish Balay if (!i) SETERRQ(1,"MatGetSubmatrices_SeqBAIJ:IS is not sorted"); 82736121d4SSatish Balay 83736121d4SSatish Balay ierr = ISGetIndices(isrow,&irow); CHKERRQ(ierr); 84*218c64b6SSatish Balay ierr = ISGetIndices(iscol,&icol); CHKERRQ(ierr); 85736121d4SSatish Balay ierr = ISGetSize(isrow,&nrows); CHKERRQ(ierr); 86736121d4SSatish Balay ierr = ISGetSize(iscol,&ncols); CHKERRQ(ierr); 87736121d4SSatish Balay 88736121d4SSatish Balay smap = (int *) PetscMalloc((1+oldcols)*sizeof(int)); CHKPTRQ(smap); 89736121d4SSatish Balay ssmap = smap; 90736121d4SSatish Balay lens = (int *) PetscMalloc((1+nrows)*sizeof(int)); CHKPTRQ(lens); 91736121d4SSatish Balay PetscMemzero(smap,oldcols*sizeof(int)); 92736121d4SSatish Balay for ( i=0; i<ncols; i++ ) smap[icol[i]] = i+1; 93736121d4SSatish Balay /* determine lens of each row */ 94736121d4SSatish Balay for (i=0; i<nrows; i++) { 95736121d4SSatish Balay kstart = ai[irow[i]]; 96736121d4SSatish Balay kend = kstart + a->ilen[irow[i]]; 97736121d4SSatish Balay lens[i] = 0; 98736121d4SSatish Balay for ( k=kstart; k<kend; k++ ) { 99736121d4SSatish Balay if (ssmap[aj[k]]) { 100736121d4SSatish Balay lens[i]++; 101736121d4SSatish Balay } 102736121d4SSatish Balay } 103736121d4SSatish Balay } 104736121d4SSatish Balay /* Create and fill new matrix */ 105736121d4SSatish Balay if (scall == MAT_REUSE_MATRIX) { 106736121d4SSatish Balay c = (Mat_SeqBAIJ *)((*B)->data); 107736121d4SSatish Balay 108736121d4SSatish Balay if (c->mbs!=nrows || c->nbs!=ncols || c->bs!=bs) 109736121d4SSatish Balay SETERRQ(1,"MatGetSubMatrix_SeqBAIJ:"); 110736121d4SSatish Balay if (PetscMemcmp(c->ilen,lens, c->mbs *sizeof(int))) { 111736121d4SSatish Balay SETERRQ(1,"MatGetSubmatrix_SeqBAIJ:Cannot reuse matrix. wrong no of nonzeros"); 112736121d4SSatish Balay } 113736121d4SSatish Balay PetscMemzero(c->ilen,c->mbs*sizeof(int)); 114736121d4SSatish Balay C = *B; 115736121d4SSatish Balay } 116736121d4SSatish Balay else { 117736121d4SSatish Balay ierr = MatCreateSeqBAIJ(A->comm,bs,nrows*bs,ncols*bs,0,lens,&C);CHKERRQ(ierr); 118736121d4SSatish Balay } 119736121d4SSatish Balay c = (Mat_SeqBAIJ *)(C->data); 120736121d4SSatish Balay for (i=0; i<nrows; i++) { 121736121d4SSatish Balay row = irow[i]; 122736121d4SSatish Balay nznew = 0; 123736121d4SSatish Balay kstart = ai[row]; 124736121d4SSatish Balay kend = kstart + a->ilen[row]; 125736121d4SSatish Balay mat_i = c->i[i]; 126736121d4SSatish Balay mat_j = c->j + mat_i; 127*218c64b6SSatish Balay mat_a = c->a + mat_i*bs2; 128736121d4SSatish Balay mat_ilen = c->ilen + i; 129736121d4SSatish Balay for ( k=kstart; k<kend; k++ ) { 130736121d4SSatish Balay if ((tcol=ssmap[a->j[k]])) { 131736121d4SSatish Balay *mat_j++ = tcol - 1; 132736121d4SSatish Balay PetscMemcpy(mat_a,a->a+k*bs2,bs2*sizeof(Scalar)); mat_a+=bs2; 133736121d4SSatish Balay (*mat_ilen)++; 134736121d4SSatish Balay 135736121d4SSatish Balay } 136736121d4SSatish Balay } 137736121d4SSatish Balay } 138*218c64b6SSatish Balay 139736121d4SSatish Balay /* Free work space */ 140736121d4SSatish Balay ierr = ISRestoreIndices(iscol,&icol); CHKERRQ(ierr); 141736121d4SSatish Balay PetscFree(smap); PetscFree(lens); 142736121d4SSatish Balay ierr = MatAssemblyBegin(C,FINAL_ASSEMBLY); CHKERRQ(ierr); 143736121d4SSatish Balay ierr = MatAssemblyEnd(C,FINAL_ASSEMBLY); CHKERRQ(ierr); 144736121d4SSatish Balay 145736121d4SSatish Balay ierr = ISRestoreIndices(isrow,&irow); CHKERRQ(ierr); 146736121d4SSatish Balay *B = C; 147736121d4SSatish Balay return 0; 148736121d4SSatish Balay } 149736121d4SSatish Balay 150*218c64b6SSatish Balay int MatGetSubMatrix_SeqBAIJ(Mat A,IS isrow,IS iscol,MatGetSubMatrixCall scall,Mat *B) 151*218c64b6SSatish Balay { 152*218c64b6SSatish Balay Mat_SeqBAIJ *a = (Mat_SeqBAIJ *) A->data; 153*218c64b6SSatish Balay IS is1,is2; 154*218c64b6SSatish Balay int *vary,*iary,*irow,*icol,nrows,ncols,i,ierr,bs=a->bs,count; 155*218c64b6SSatish Balay 156*218c64b6SSatish Balay ierr = ISGetIndices(isrow,&irow); CHKERRQ(ierr); 157*218c64b6SSatish Balay ierr = ISGetIndices(iscol,&icol); CHKERRQ(ierr); 158*218c64b6SSatish Balay ierr = ISGetSize(isrow,&nrows); CHKERRQ(ierr); 159*218c64b6SSatish Balay ierr = ISGetSize(iscol,&ncols); CHKERRQ(ierr); 160*218c64b6SSatish Balay 161*218c64b6SSatish Balay /* Verify if the indices corespond to each elementin a block 162*218c64b6SSatish Balay and form the IS with compressed IS */ 163*218c64b6SSatish Balay vary = (int *) PetscMalloc(2*(a->mbs+1)*sizeof(int)); CHKPTRQ(vary); 164*218c64b6SSatish Balay iary = vary + a->mbs; 165*218c64b6SSatish Balay PetscMemzero(vary,(a->mbs)*sizeof(int)); 166*218c64b6SSatish Balay for ( i=0; i<nrows; i++) vary[irow[i]/bs]++; 167*218c64b6SSatish Balay count = 0; 168*218c64b6SSatish Balay for (i=0; i<a->mbs; i++) { 169*218c64b6SSatish Balay if (vary[i]!=0 && vary[i]!=bs) SETERRA(1,"MatGetSubmatrices_SeqBAIJ:"); 170*218c64b6SSatish Balay if (vary[i]==bs) iary[count++] = i; 171*218c64b6SSatish Balay } 172*218c64b6SSatish Balay ierr = ISCreateSeq(MPI_COMM_SELF, count, iary,&is1); CHKERRQ(ierr); 173*218c64b6SSatish Balay 174*218c64b6SSatish Balay PetscMemzero(vary,(a->mbs)*sizeof(int)); 175*218c64b6SSatish Balay for ( i=0; i<ncols; i++) vary[icol[i]/bs]++; 176*218c64b6SSatish Balay count = 0; 177*218c64b6SSatish Balay for (i=0; i<a->mbs; i++) { 178*218c64b6SSatish Balay if (vary[i]!=0 && vary[i]!=bs) SETERRA(1,"MatGetSubmatrices_SeqBAIJ:"); 179*218c64b6SSatish Balay if (vary[i]==bs) iary[count++] = i; 180*218c64b6SSatish Balay } 181*218c64b6SSatish Balay ierr = ISCreateSeq(MPI_COMM_SELF, count, iary,&is2); CHKERRQ(ierr); 182*218c64b6SSatish Balay ierr = ISRestoreIndices(isrow,&irow); CHKERRQ(ierr); 183*218c64b6SSatish Balay ierr = ISRestoreIndices(iscol,&icol); CHKERRQ(ierr); 184*218c64b6SSatish Balay PetscFree(vary); 185*218c64b6SSatish Balay 186*218c64b6SSatish Balay ierr = MatGetSubMatrix_SeqBAIJ_Private(A,is1,is2,scall,B); CHKERRQ(ierr); 187*218c64b6SSatish Balay ISDestroy(is1); 188*218c64b6SSatish Balay ISDestroy(is2); 189*218c64b6SSatish Balay return 0; 190*218c64b6SSatish Balay } 191*218c64b6SSatish Balay 192736121d4SSatish Balay int MatGetSubMatrices_SeqBAIJ(Mat A,int n, IS *irow,IS *icol,MatGetSubMatrixCall scall, 193736121d4SSatish Balay Mat **B) 194736121d4SSatish Balay { 195736121d4SSatish Balay int ierr,i; 196736121d4SSatish Balay 197736121d4SSatish Balay if (scall == MAT_INITIAL_MATRIX) { 198736121d4SSatish Balay *B = (Mat *) PetscMalloc( (n+1)*sizeof(Mat) ); CHKPTRQ(*B); 199736121d4SSatish Balay } 200736121d4SSatish Balay 201736121d4SSatish Balay for ( i=0; i<n; i++ ) { 202736121d4SSatish Balay ierr = MatGetSubMatrix(A,irow[i],icol[i],scall,&(*B)[i]); CHKERRQ(ierr); 203736121d4SSatish Balay } 204736121d4SSatish Balay return 0; 205736121d4SSatish Balay } 206*218c64b6SSatish Balay 207*218c64b6SSatish Balay 208*218c64b6SSatish Balay 209*218c64b6SSatish Balay 210*218c64b6SSatish Balay 211*218c64b6SSatish Balay 212*218c64b6SSatish Balay 213