1cb512458SBarry Smith #ifndef lint 2*bbb6d6a8SBarry Smith static char vcid[] = "$Id: aijfact.c,v 1.26 1995/07/10 05:02:25 bsmith Exp bsmith $"; 3cb512458SBarry Smith #endif 4289bc588SBarry Smith 5289bc588SBarry Smith 6289bc588SBarry Smith #include "aij.h" 78c37ef55SBarry Smith #include "inline/spops.h" 8289bc588SBarry Smith /* 9289bc588SBarry Smith Factorization code for AIJ format. 10289bc588SBarry Smith */ 11289bc588SBarry Smith 1249d8b64dSBarry Smith int MatLUFactorSymbolic_AIJ(Mat mat,IS isrow,IS iscol,double f,Mat *fact) 13289bc588SBarry Smith { 141fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data, *aijnew; 15289bc588SBarry Smith IS isicol; 166abc6512SBarry Smith int *r,*ic, ierr, i, n = aij->m, *ai = aij->i, *aj = aij->j; 176abc6512SBarry Smith int *ainew,*ajnew, jmax,*fill, *ajtmp, nz; 182fb3ed76SBarry Smith int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 19289bc588SBarry Smith 20d35516d3SLois Curfman McInnes if (n != aij->n) 21*bbb6d6a8SBarry Smith SETERRQ(1,"MatLUFactorSymbolic_AIJ:Matrix must be square"); 22d35516d3SLois Curfman McInnes if (!isrow) 23*bbb6d6a8SBarry Smith SETERRQ(1,"MatLUFactorSymbolic_AIJ:Matrix must have row permutation"); 24d35516d3SLois Curfman McInnes if (!iscol) 25*bbb6d6a8SBarry Smith SETERRQ(1,"MatLUFactorSymbolic_AIJ:Matrix must have column permutation"); 26289bc588SBarry Smith 2778b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 28289bc588SBarry Smith ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 29289bc588SBarry Smith 30289bc588SBarry Smith /* get new row pointers */ 3178b31e54SBarry Smith ainew = (int *) PETSCMALLOC( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 32289bc588SBarry Smith ainew[0] = 1; 33289bc588SBarry Smith /* don't know how many column pointers are needed so estimate */ 34*bbb6d6a8SBarry Smith jmax = (int) (f*ai[n]); 3578b31e54SBarry Smith ajnew = (int *) PETSCMALLOC( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 36289bc588SBarry Smith /* fill is a linked list of nonzeros in active row */ 372fb3ed76SBarry Smith fill = (int *) PETSCMALLOC( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 382fb3ed76SBarry Smith im = fill + n + 1; 39289bc588SBarry Smith /* idnew is location of diagonal in factor */ 4078b31e54SBarry Smith idnew = (int *) PETSCMALLOC( (n+1)*sizeof(int)); CHKPTRQ(idnew); 41289bc588SBarry Smith idnew[0] = 1; 42289bc588SBarry Smith 43289bc588SBarry Smith for ( i=0; i<n; i++ ) { 44289bc588SBarry Smith /* first copy previous fill into linked list */ 45289bc588SBarry Smith nnz = nz = ai[r[i]+1] - ai[r[i]]; 46289bc588SBarry Smith ajtmp = aj + ai[r[i]] - 1; 47289bc588SBarry Smith fill[n] = n; 48289bc588SBarry Smith while (nz--) { 49289bc588SBarry Smith fm = n; 50289bc588SBarry Smith idx = ic[*ajtmp++ - 1]; 51289bc588SBarry Smith do { 52289bc588SBarry Smith m = fm; 53289bc588SBarry Smith fm = fill[m]; 54289bc588SBarry Smith } while (fm < idx); 55289bc588SBarry Smith fill[m] = idx; 56289bc588SBarry Smith fill[idx] = fm; 57289bc588SBarry Smith } 58289bc588SBarry Smith row = fill[n]; 59289bc588SBarry Smith while ( row < i ) { 602fb3ed76SBarry Smith ajtmp = ajnew + idnew[row]; 612fb3ed76SBarry Smith nzbd = 1 + idnew[row] - ainew[row]; 622fb3ed76SBarry Smith nz = im[row] - nzbd; 63289bc588SBarry Smith fm = row; 642fb3ed76SBarry Smith while (nz-- > 0) { 652fb3ed76SBarry Smith /* fm = n; */ 66289bc588SBarry Smith idx = *ajtmp++ - 1; 672fb3ed76SBarry Smith nzbd++; 682fb3ed76SBarry Smith if (idx == i) im[row] = nzbd; 69289bc588SBarry Smith do { 70289bc588SBarry Smith m = fm; 71289bc588SBarry Smith fm = fill[m]; 72289bc588SBarry Smith } while (fm < idx); 73289bc588SBarry Smith if (fm != idx) { 74289bc588SBarry Smith fill[m] = idx; 75289bc588SBarry Smith fill[idx] = fm; 76289bc588SBarry Smith fm = idx; 77289bc588SBarry Smith nnz++; 78289bc588SBarry Smith } 79d5f53d93SBarry Smith /* printf("i %d row %d nz %d idx %d fm %d\n",i,row,nz,idx,fm); */ 80289bc588SBarry Smith } 81289bc588SBarry Smith row = fill[row]; 82289bc588SBarry Smith } 83289bc588SBarry Smith /* copy new filled row into permanent storage */ 84289bc588SBarry Smith ainew[i+1] = ainew[i] + nnz; 85289bc588SBarry Smith if (ainew[i+1] > jmax+1) { 86289bc588SBarry Smith /* allocate a longer ajnew */ 872fb3ed76SBarry Smith int maxadd; 88*bbb6d6a8SBarry Smith maxadd = (int) ((f*ai[n]*(n-i+5))/n); 89*bbb6d6a8SBarry Smith if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 902fb3ed76SBarry Smith jmax += maxadd; 9178b31e54SBarry Smith ajtmp = (int *) PETSCMALLOC( jmax*sizeof(int) );CHKPTRQ(ajtmp); 9278b31e54SBarry Smith PETSCMEMCPY(ajtmp,ajnew,(ainew[i]-1)*sizeof(int)); 9378b31e54SBarry Smith PETSCFREE(ajnew); 94289bc588SBarry Smith ajnew = ajtmp; 952fb3ed76SBarry Smith realloc++; /* count how many times we realloc */ 96289bc588SBarry Smith } 97289bc588SBarry Smith ajtmp = ajnew + ainew[i] - 1; 98289bc588SBarry Smith fm = fill[n]; 99289bc588SBarry Smith nzi = 0; 1002fb3ed76SBarry Smith im[i] = nnz; 101289bc588SBarry Smith while (nnz--) { 102289bc588SBarry Smith if (fm < i) nzi++; 103289bc588SBarry Smith *ajtmp++ = fm + 1; 104289bc588SBarry Smith fm = fill[fm]; 105289bc588SBarry Smith } 106289bc588SBarry Smith idnew[i] = ainew[i] + nzi; 107289bc588SBarry Smith } 108289bc588SBarry Smith 1092fb3ed76SBarry Smith PLogInfo((PetscObject)mat, 110*bbb6d6a8SBarry Smith "Info:MatLUFactorSymbolic_AIJ:Reallocs %d Fill ratio:given %g needed %g\n", 111*bbb6d6a8SBarry Smith realloc,f,((double)ainew[n])/((double)ai[i])); 1122fb3ed76SBarry Smith 11378b31e54SBarry Smith ISDestroy(isicol); PETSCFREE(fill); 114289bc588SBarry Smith 115289bc588SBarry Smith /* put together the new matrix */ 11678b31e54SBarry Smith ierr = MatCreateSequentialAIJ(mat->comm,n, n, 0, 0, fact); CHKERRQ(ierr); 1171fb19edaSLois Curfman McInnes aijnew = (Mat_AIJ *) (*fact)->data; 11878b31e54SBarry Smith PETSCFREE(aijnew->imax); 119289bc588SBarry Smith aijnew->singlemalloc = 0; 120f0479e8cSBarry Smith len = (ainew[n] - 1)*sizeof(Scalar); 121e8d4e0b9SBarry Smith /* the next line frees the default space generated by the Create() */ 12278b31e54SBarry Smith PETSCFREE(aijnew->a); PETSCFREE(aijnew->ilen); 12378b31e54SBarry Smith aijnew->a = (Scalar *) PETSCMALLOC( len ); CHKPTRQ(aijnew->a); 124289bc588SBarry Smith aijnew->j = ajnew; 125289bc588SBarry Smith aijnew->i = ainew; 1268c37ef55SBarry Smith aijnew->diag = idnew; 127e8d4e0b9SBarry Smith aijnew->ilen = 0; 12820563c6bSBarry Smith aijnew->imax = 0; 12986d7a8ceSBarry Smith aijnew->row = isrow; 13086d7a8ceSBarry Smith aijnew->col = iscol; 131289bc588SBarry Smith (*fact)->factor = FACTOR_LU; 13278b31e54SBarry Smith aijnew->solve_work = (Scalar *) PETSCMALLOC( n*sizeof(Scalar)); 13378b31e54SBarry Smith CHKPTRQ(aijnew->solve_work); 1347fd98bd6SLois Curfman McInnes /* In aijnew structure: Free imax, ilen, old a, old j. 1357fd98bd6SLois Curfman McInnes Allocate idnew, solve_work, new a, new j */ 1367fd98bd6SLois Curfman McInnes aijnew->mem += (ainew[n]-1-n)*(sizeof(int) + sizeof(Scalar)) + sizeof(int); 1377fd98bd6SLois Curfman McInnes aijnew->maxnz = aijnew->nz = ainew[n] - 1; 1387fd98bd6SLois Curfman McInnes 1395e42470aSBarry Smith /* Cannot do this here because child is destroyed before parent created 1405e42470aSBarry Smith PLogObjectParent(*fact,isicol); */ 141289bc588SBarry Smith return 0; 142289bc588SBarry Smith } 143289bc588SBarry Smith 1441fb19edaSLois Curfman McInnes int MatLUFactorNumeric_AIJ(Mat mat,Mat *infact) 145289bc588SBarry Smith { 14620563c6bSBarry Smith Mat fact = *infact; 1471fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data, *aijnew = (Mat_AIJ *)fact->data; 14886d7a8ceSBarry Smith IS iscol = aijnew->col, isrow = aijnew->row, isicol; 149289bc588SBarry Smith int *r,*ic, ierr, i, j, n = aij->m, *ai = aijnew->i, *aj = aijnew->j; 1506abc6512SBarry Smith int *ajtmpold, *ajtmp, nz, row,*pj; 1516abc6512SBarry Smith Scalar *rtmp,*v, *pv, *pc, multiplier; 152289bc588SBarry Smith 15378b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 1545e42470aSBarry Smith PLogObjectParent(*infact,isicol); 15578b31e54SBarry Smith ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 15678b31e54SBarry Smith ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 15778b31e54SBarry Smith rtmp = (Scalar *) PETSCMALLOC( (n+1)*sizeof(Scalar) ); CHKPTRQ(rtmp); 158289bc588SBarry Smith 159289bc588SBarry Smith for ( i=0; i<n; i++ ) { 160289bc588SBarry Smith nz = ai[i+1] - ai[i]; 161289bc588SBarry Smith ajtmp = aj + ai[i] - 1; 162289bc588SBarry Smith for ( j=0; j<nz; j++ ) rtmp[ajtmp[j]-1] = 0.0; 163289bc588SBarry Smith 164289bc588SBarry Smith /* load in initial (unfactored row) */ 1658c37ef55SBarry Smith nz = aij->i[r[i]+1] - aij->i[r[i]]; 1668c37ef55SBarry Smith ajtmpold = aij->j + aij->i[r[i]] - 1; 1678c37ef55SBarry Smith v = aij->a + aij->i[r[i]] - 1; 1688c37ef55SBarry Smith for ( j=0; j<nz; j++ ) rtmp[ic[ajtmpold[j]-1]] = v[j]; 169289bc588SBarry Smith 1708c37ef55SBarry Smith row = *ajtmp++ - 1; 171289bc588SBarry Smith while (row < i) { 1728c37ef55SBarry Smith pc = rtmp + row; 1738c37ef55SBarry Smith if (*pc != 0.0) { 1748c37ef55SBarry Smith nz = aijnew->diag[row] - ai[row]; 1758c37ef55SBarry Smith pv = aijnew->a + aijnew->diag[row] - 1; 1768c37ef55SBarry Smith pj = aijnew->j + aijnew->diag[row]; 1778c37ef55SBarry Smith multiplier = *pc * *pv++; 1788c37ef55SBarry Smith *pc = multiplier; 1798c37ef55SBarry Smith nz = ai[row+1] - ai[row] - 1 - nz; 180eef2e97bSBarry Smith PLogFlops(2*nz); 1818c37ef55SBarry Smith while (nz-->0) rtmp[*pj++ - 1] -= multiplier* *pv++; 182289bc588SBarry Smith } 1838c37ef55SBarry Smith row = *ajtmp++ - 1; 184289bc588SBarry Smith } 1858c37ef55SBarry Smith /* finished row so stick it into aijnew->a */ 1868c37ef55SBarry Smith pv = aijnew->a + ai[i] - 1; 1878c37ef55SBarry Smith pj = aijnew->j + ai[i] - 1; 1888c37ef55SBarry Smith nz = ai[i+1] - ai[i]; 189*bbb6d6a8SBarry Smith if (rtmp[i] == 0.0) {SETERRQ(1,"MatLUFactorNumeric_AIJ:Zero pivot");} 1908c37ef55SBarry Smith rtmp[i] = 1.0/rtmp[i]; 1918c37ef55SBarry Smith for ( j=0; j<nz; j++ ) {pv[j] = rtmp[pj[j]-1];} 1928c37ef55SBarry Smith } 19378b31e54SBarry Smith PETSCFREE(rtmp); 19478b31e54SBarry Smith ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 19578b31e54SBarry Smith ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 19678b31e54SBarry Smith ierr = ISDestroy(isicol); CHKERRQ(ierr); 1978c37ef55SBarry Smith fact->factor = FACTOR_LU; 1989e25ed09SBarry Smith aijnew->assembled = 1; 199eef2e97bSBarry Smith PLogFlops(aijnew->n); 200289bc588SBarry Smith return 0; 201289bc588SBarry Smith } 20249d8b64dSBarry Smith int MatLUFactor_AIJ(Mat matin,IS row,IS col,double f) 203da3a660dSBarry Smith { 2041fb19edaSLois Curfman McInnes Mat_AIJ *mat = (Mat_AIJ *) matin->data; 2056abc6512SBarry Smith int ierr; 206da3a660dSBarry Smith Mat fact; 20749d8b64dSBarry Smith ierr = MatLUFactorSymbolic_AIJ(matin,row,col,f,&fact); CHKERRQ(ierr); 20878b31e54SBarry Smith ierr = MatLUFactorNumeric_AIJ(matin,&fact); CHKERRQ(ierr); 209da3a660dSBarry Smith 210da3a660dSBarry Smith /* free all the data structures from mat */ 21178b31e54SBarry Smith PETSCFREE(mat->a); 21278b31e54SBarry Smith if (!mat->singlemalloc) {PETSCFREE(mat->i); PETSCFREE(mat->j);} 21378b31e54SBarry Smith if (mat->diag) PETSCFREE(mat->diag); 21478b31e54SBarry Smith if (mat->ilen) PETSCFREE(mat->ilen); 21578b31e54SBarry Smith if (mat->imax) PETSCFREE(mat->imax); 21686d7a8ceSBarry Smith if (mat->row && mat->col && mat->row != mat->col) { 21786d7a8ceSBarry Smith ISDestroy(mat->row); 218da3a660dSBarry Smith } 21986d7a8ceSBarry Smith if (mat->col) ISDestroy(mat->col); 22078b31e54SBarry Smith PETSCFREE(mat); 221da3a660dSBarry Smith 22278b31e54SBarry Smith PETSCMEMCPY(matin,fact,sizeof(struct _Mat)); 22378b31e54SBarry Smith PETSCFREE(fact); 224da3a660dSBarry Smith return 0; 225da3a660dSBarry Smith } 226da3a660dSBarry Smith 2271fb19edaSLois Curfman McInnes int MatSolve_AIJ(Mat mat,Vec bb, Vec xx) 2288c37ef55SBarry Smith { 2291fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data; 23086d7a8ceSBarry Smith IS iscol = aij->col, isrow = aij->row; 2316abc6512SBarry Smith int *r,*c, ierr, i, n = aij->m, *vi, *ai = aij->i, *aj = aij->j; 2328c37ef55SBarry Smith int nz; 2338c37ef55SBarry Smith Scalar *x,*b,*tmp, *aa = aij->a, sum, *v; 2348c37ef55SBarry Smith 235d35516d3SLois Curfman McInnes if (mat->factor != FACTOR_LU) 236*bbb6d6a8SBarry Smith SETERRQ(1,"MatSolve_AIJ:Cannot solve with unfactored matrix"); 2379e25ed09SBarry Smith 23878b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 23978b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 240eef2e97bSBarry Smith tmp = aij->solve_work; 2418c37ef55SBarry Smith 24278b31e54SBarry Smith ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 24378b31e54SBarry Smith ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 2448c37ef55SBarry Smith 2458c37ef55SBarry Smith /* forward solve the lower triangular */ 2468c37ef55SBarry Smith tmp[0] = b[*r++]; 2478c37ef55SBarry Smith for ( i=1; i<n; i++ ) { 2488c37ef55SBarry Smith v = aa + ai[i] - 1; 2498c37ef55SBarry Smith vi = aj + ai[i] - 1; 2508c37ef55SBarry Smith nz = aij->diag[i] - ai[i]; 2518c37ef55SBarry Smith sum = b[*r++]; 2528c37ef55SBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ - 1]; 2538c37ef55SBarry Smith tmp[i] = sum; 2548c37ef55SBarry Smith } 2558c37ef55SBarry Smith 2568c37ef55SBarry Smith /* backward solve the upper triangular */ 2578c37ef55SBarry Smith for ( i=n-1; i>=0; i-- ){ 2588c37ef55SBarry Smith v = aa + aij->diag[i]; 2598c37ef55SBarry Smith vi = aj + aij->diag[i]; 2608c37ef55SBarry Smith nz = ai[i+1] - aij->diag[i] - 1; 2618c37ef55SBarry Smith sum = tmp[i]; 2628c37ef55SBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ - 1]; 2638c37ef55SBarry Smith x[*c--] = tmp[i] = sum*aa[aij->diag[i]-1]; 2648c37ef55SBarry Smith } 2658c37ef55SBarry Smith 266eef2e97bSBarry Smith PLogFlops(2*aij->nz - aij->n); 2678c37ef55SBarry Smith return 0; 2688c37ef55SBarry Smith } 2691fb19edaSLois Curfman McInnes int MatSolveAdd_AIJ(Mat mat,Vec bb, Vec yy, Vec xx) 270da3a660dSBarry Smith { 2711fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data; 27286d7a8ceSBarry Smith IS iscol = aij->col, isrow = aij->row; 2736abc6512SBarry Smith int *r,*c, ierr, i, n = aij->m, *vi, *ai = aij->i, *aj = aij->j; 274da3a660dSBarry Smith int nz; 275da3a660dSBarry Smith Scalar *x,*b,*tmp, *aa = aij->a, sum, *v; 276da3a660dSBarry Smith 277d35516d3SLois Curfman McInnes if (mat->factor != FACTOR_LU) 278*bbb6d6a8SBarry Smith SETERRQ(1,"MatSolveAdd_AIJ: Cannot solve with unfactored matrix"); 27978b31e54SBarry Smith if (yy != xx) {ierr = VecCopy(yy,xx); CHKERRQ(ierr);} 280da3a660dSBarry Smith 28178b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 28278b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 283eef2e97bSBarry Smith tmp = aij->solve_work; 284da3a660dSBarry Smith 28578b31e54SBarry Smith ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 28678b31e54SBarry Smith ierr = ISGetIndices(iscol,&c); CHKERRQ(ierr); c = c + (n-1); 287da3a660dSBarry Smith 288da3a660dSBarry Smith /* forward solve the lower triangular */ 289da3a660dSBarry Smith tmp[0] = b[*r++]; 290da3a660dSBarry Smith for ( i=1; i<n; i++ ) { 291da3a660dSBarry Smith v = aa + ai[i] - 1; 292da3a660dSBarry Smith vi = aj + ai[i] - 1; 293da3a660dSBarry Smith nz = aij->diag[i] - ai[i]; 294da3a660dSBarry Smith sum = b[*r++]; 295da3a660dSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ - 1]; 296da3a660dSBarry Smith tmp[i] = sum; 297da3a660dSBarry Smith } 298da3a660dSBarry Smith 299da3a660dSBarry Smith /* backward solve the upper triangular */ 300da3a660dSBarry Smith for ( i=n-1; i>=0; i-- ){ 301da3a660dSBarry Smith v = aa + aij->diag[i]; 302da3a660dSBarry Smith vi = aj + aij->diag[i]; 303da3a660dSBarry Smith nz = ai[i+1] - aij->diag[i] - 1; 304da3a660dSBarry Smith sum = tmp[i]; 305da3a660dSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ - 1]; 306da3a660dSBarry Smith tmp[i] = sum*aa[aij->diag[i]-1]; 307da3a660dSBarry Smith x[*c--] += tmp[i]; 308da3a660dSBarry Smith } 309da3a660dSBarry Smith 310eef2e97bSBarry Smith PLogFlops(2*aij->nz); 311da3a660dSBarry Smith return 0; 312da3a660dSBarry Smith } 313da3a660dSBarry Smith /* -------------------------------------------------------------------*/ 3141fb19edaSLois Curfman McInnes int MatSolveTrans_AIJ(Mat mat,Vec bb, Vec xx) 315da3a660dSBarry Smith { 3161fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data; 31786d7a8ceSBarry Smith IS iscol = aij->col, isrow = aij->row, invisrow,inviscol; 3186abc6512SBarry Smith int *r,*c, ierr, i, n = aij->m, *vi, *ai = aij->i, *aj = aij->j; 319da3a660dSBarry Smith int nz; 3206abc6512SBarry Smith Scalar *x,*b,*tmp, *aa = aij->a, *v; 321da3a660dSBarry Smith 322d35516d3SLois Curfman McInnes if (mat->factor != FACTOR_LU) 323*bbb6d6a8SBarry Smith SETERRQ(1,"MatSolveTrans_AIJ:Cannot solve with unfactored matrix"); 32478b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 32578b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 326eef2e97bSBarry Smith tmp = aij->solve_work; 327da3a660dSBarry Smith 328da3a660dSBarry Smith /* invert the permutations */ 32978b31e54SBarry Smith ierr = ISInvertPermutation(isrow,&invisrow); CHKERRQ(ierr); 33078b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&inviscol); CHKERRQ(ierr); 331da3a660dSBarry Smith 332da3a660dSBarry Smith 33378b31e54SBarry Smith ierr = ISGetIndices(invisrow,&r); CHKERRQ(ierr); 33478b31e54SBarry Smith ierr = ISGetIndices(inviscol,&c); CHKERRQ(ierr); 335da3a660dSBarry Smith 336da3a660dSBarry Smith /* copy the b into temp work space according to permutation */ 337da3a660dSBarry Smith for ( i=0; i<n; i++ ) tmp[c[i]] = b[i]; 338da3a660dSBarry Smith 339da3a660dSBarry Smith /* forward solve the U^T */ 340da3a660dSBarry Smith for ( i=0; i<n; i++ ) { 341da3a660dSBarry Smith v = aa + aij->diag[i] - 1; 342da3a660dSBarry Smith vi = aj + aij->diag[i]; 343da3a660dSBarry Smith nz = ai[i+1] - aij->diag[i] - 1; 344da3a660dSBarry Smith tmp[i] *= *v++; 345da3a660dSBarry Smith while (nz--) { 346da3a660dSBarry Smith tmp[*vi++ - 1] -= (*v++)*tmp[i]; 347da3a660dSBarry Smith } 348da3a660dSBarry Smith } 349da3a660dSBarry Smith 350da3a660dSBarry Smith /* backward solve the L^T */ 351da3a660dSBarry Smith for ( i=n-1; i>=0; i-- ){ 352da3a660dSBarry Smith v = aa + aij->diag[i] - 2; 353da3a660dSBarry Smith vi = aj + aij->diag[i] - 2; 354da3a660dSBarry Smith nz = aij->diag[i] - ai[i]; 355da3a660dSBarry Smith while (nz--) { 356da3a660dSBarry Smith tmp[*vi-- - 1] -= (*v--)*tmp[i]; 357da3a660dSBarry Smith } 358da3a660dSBarry Smith } 359da3a660dSBarry Smith 360da3a660dSBarry Smith /* copy tmp into x according to permutation */ 361da3a660dSBarry Smith for ( i=0; i<n; i++ ) x[r[i]] = tmp[i]; 362da3a660dSBarry Smith 363da3a660dSBarry Smith ISDestroy(invisrow); ISDestroy(inviscol); 364da3a660dSBarry Smith 365eef2e97bSBarry Smith PLogFlops(2*aij->nz-aij->n); 366da3a660dSBarry Smith return 0; 367da3a660dSBarry Smith } 368da3a660dSBarry Smith 3691fb19edaSLois Curfman McInnes int MatSolveTransAdd_AIJ(Mat mat,Vec bb, Vec zz,Vec xx) 370da3a660dSBarry Smith { 3711fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data; 37286d7a8ceSBarry Smith IS iscol = aij->col, isrow = aij->row, invisrow,inviscol; 3736abc6512SBarry Smith int *r,*c, ierr, i, n = aij->m, *vi, *ai = aij->i, *aj = aij->j; 3746abc6512SBarry Smith int nz; 3756abc6512SBarry Smith Scalar *x,*b,*tmp, *aa = aij->a, *v; 3766abc6512SBarry Smith 377d35516d3SLois Curfman McInnes if (mat->factor != FACTOR_LU) 378*bbb6d6a8SBarry Smith SETERRQ(1,"MatSolveTransAdd_AIJ:Cannot solve with unfactored matrix"); 3796abc6512SBarry Smith if (zz != xx) VecCopy(zz,xx); 3806abc6512SBarry Smith 38178b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 38278b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 383eef2e97bSBarry Smith tmp = aij->solve_work; 3846abc6512SBarry Smith 3856abc6512SBarry Smith /* invert the permutations */ 38678b31e54SBarry Smith ierr = ISInvertPermutation(isrow,&invisrow); CHKERRQ(ierr); 38778b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&inviscol); CHKERRQ(ierr); 3886abc6512SBarry Smith 3896abc6512SBarry Smith 39078b31e54SBarry Smith ierr = ISGetIndices(invisrow,&r); CHKERRQ(ierr); 39178b31e54SBarry Smith ierr = ISGetIndices(inviscol,&c); CHKERRQ(ierr); 3926abc6512SBarry Smith 3936abc6512SBarry Smith /* copy the b into temp work space according to permutation */ 3946abc6512SBarry Smith for ( i=0; i<n; i++ ) tmp[c[i]] = b[i]; 3956abc6512SBarry Smith 3966abc6512SBarry Smith /* forward solve the U^T */ 3976abc6512SBarry Smith for ( i=0; i<n; i++ ) { 3986abc6512SBarry Smith v = aa + aij->diag[i] - 1; 3996abc6512SBarry Smith vi = aj + aij->diag[i]; 4006abc6512SBarry Smith nz = ai[i+1] - aij->diag[i] - 1; 4016abc6512SBarry Smith tmp[i] *= *v++; 4026abc6512SBarry Smith while (nz--) { 4036abc6512SBarry Smith tmp[*vi++ - 1] -= (*v++)*tmp[i]; 4046abc6512SBarry Smith } 4056abc6512SBarry Smith } 4066abc6512SBarry Smith 4076abc6512SBarry Smith /* backward solve the L^T */ 4086abc6512SBarry Smith for ( i=n-1; i>=0; i-- ){ 4096abc6512SBarry Smith v = aa + aij->diag[i] - 2; 4106abc6512SBarry Smith vi = aj + aij->diag[i] - 2; 4116abc6512SBarry Smith nz = aij->diag[i] - ai[i]; 4126abc6512SBarry Smith while (nz--) { 4136abc6512SBarry Smith tmp[*vi-- - 1] -= (*v--)*tmp[i]; 4146abc6512SBarry Smith } 4156abc6512SBarry Smith } 4166abc6512SBarry Smith 4176abc6512SBarry Smith /* copy tmp into x according to permutation */ 4186abc6512SBarry Smith for ( i=0; i<n; i++ ) x[r[i]] += tmp[i]; 4196abc6512SBarry Smith 4206abc6512SBarry Smith ISDestroy(invisrow); ISDestroy(inviscol); 4216abc6512SBarry Smith 422eef2e97bSBarry Smith PLogFlops(2*aij->nz); 4236abc6512SBarry Smith return 0; 424da3a660dSBarry Smith } 4259e25ed09SBarry Smith /* ----------------------------------------------------------------*/ 42649d8b64dSBarry Smith int MatILUFactorSymbolic_AIJ(Mat mat,IS isrow,IS iscol,double f, 42749d8b64dSBarry Smith int levels,Mat *fact) 4289e25ed09SBarry Smith { 4291fb19edaSLois Curfman McInnes Mat_AIJ *aij = (Mat_AIJ *) mat->data, *aijnew; 4309e25ed09SBarry Smith IS isicol; 4319e25ed09SBarry Smith int *r,*ic, ierr, i, n = aij->m, *ai = aij->i, *aj = aij->j; 4329e25ed09SBarry Smith int *ainew,*ajnew, jmax,*fill, *ajtmp, nz, *lfill,*ajfill,*ajtmpf; 433*bbb6d6a8SBarry Smith int *idnew, idx, row,m,fm, nnz, nzi,len, *im, nzbd, realloc = 0; 4349e25ed09SBarry Smith 435d35516d3SLois Curfman McInnes if (n != aij->n) 436*bbb6d6a8SBarry Smith SETERRQ(1,"MatILUFactorSymbolic_AIJ:Matrix must be square"); 437d35516d3SLois Curfman McInnes if (!isrow) 438*bbb6d6a8SBarry Smith SETERRQ(1,"MatILUFactorSymbolic_AIJ:Matrix must have row permutation"); 439d35516d3SLois Curfman McInnes if (!iscol) SETERRQ(1, 440*bbb6d6a8SBarry Smith "MatILUFactorSymbolic_AIJ:Matrix must have column permutation"); 4419e25ed09SBarry Smith 44278b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 4439e25ed09SBarry Smith ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 4449e25ed09SBarry Smith 4459e25ed09SBarry Smith /* get new row pointers */ 44678b31e54SBarry Smith ainew = (int *) PETSCMALLOC( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 4479e25ed09SBarry Smith ainew[0] = 1; 4489e25ed09SBarry Smith /* don't know how many column pointers are needed so estimate */ 449*bbb6d6a8SBarry Smith jmax = (int) (f*ai[n]); 45078b31e54SBarry Smith ajnew = (int *) PETSCMALLOC( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 4519e25ed09SBarry Smith /* ajfill is level of fill for each fill entry */ 45278b31e54SBarry Smith ajfill = (int *) PETSCMALLOC( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 4539e25ed09SBarry Smith /* fill is a linked list of nonzeros in active row */ 454*bbb6d6a8SBarry Smith fill = (int *) PETSCMALLOC( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 455*bbb6d6a8SBarry Smith im = fill + n + 1; 4569e25ed09SBarry Smith /* lfill is level for each filled value */ 45778b31e54SBarry Smith lfill = (int *) PETSCMALLOC( (n+1)*sizeof(int)); CHKPTRQ(lfill); 4589e25ed09SBarry Smith /* idnew is location of diagonal in factor */ 45978b31e54SBarry Smith idnew = (int *) PETSCMALLOC( (n+1)*sizeof(int)); CHKPTRQ(idnew); 4609e25ed09SBarry Smith idnew[0] = 1; 4619e25ed09SBarry Smith 4629e25ed09SBarry Smith for ( i=0; i<n; i++ ) { 4639e25ed09SBarry Smith /* first copy previous fill into linked list */ 4649e25ed09SBarry Smith nnz = nz = ai[r[i]+1] - ai[r[i]]; 4659e25ed09SBarry Smith ajtmp = aj + ai[r[i]] - 1; 4669e25ed09SBarry Smith fill[n] = n; 4679e25ed09SBarry Smith while (nz--) { 4689e25ed09SBarry Smith fm = n; 4699e25ed09SBarry Smith idx = ic[*ajtmp++ - 1]; 4709e25ed09SBarry Smith do { 4719e25ed09SBarry Smith m = fm; 4729e25ed09SBarry Smith fm = fill[m]; 4739e25ed09SBarry Smith } while (fm < idx); 4749e25ed09SBarry Smith fill[m] = idx; 4759e25ed09SBarry Smith fill[idx] = fm; 4769e25ed09SBarry Smith lfill[idx] = -1; 477*bbb6d6a8SBarry Smith /* printf("i %d cols %d %d\n",i,nz,idx); */ 4789e25ed09SBarry Smith } 4799e25ed09SBarry Smith row = fill[n]; 4809e25ed09SBarry Smith while ( row < i ) { 481*bbb6d6a8SBarry Smith ajtmp = ajnew + idnew[row]; 482*bbb6d6a8SBarry Smith ajtmpf = ajfill + idnew[row]; 483*bbb6d6a8SBarry Smith 484*bbb6d6a8SBarry Smith nzbd = 1 + idnew[row] - ainew[row]; 485*bbb6d6a8SBarry Smith nz = im[row] - nzbd; 486*bbb6d6a8SBarry Smith 487*bbb6d6a8SBarry Smith fm = fill[row]; m = row; 488*bbb6d6a8SBarry Smith while (nz-- > 0) { 4899e25ed09SBarry Smith idx = *ajtmp++ - 1; 490*bbb6d6a8SBarry Smith nzbd++; 491*bbb6d6a8SBarry Smith if (idx == i) im[row] = nzbd; 492*bbb6d6a8SBarry Smith while (fm < idx) { 4939e25ed09SBarry Smith m = fm; 4949e25ed09SBarry Smith fm = fill[m]; 495*bbb6d6a8SBarry Smith } 4969e25ed09SBarry Smith if (fm != idx) { 4979e25ed09SBarry Smith lfill[idx] = *ajtmpf + 1; 4989e25ed09SBarry Smith if (lfill[idx] < levels) { 4999e25ed09SBarry Smith fill[m] = idx; 5009e25ed09SBarry Smith fill[idx] = fm; 5019e25ed09SBarry Smith fm = idx; 5029e25ed09SBarry Smith nnz++; 5039e25ed09SBarry Smith } 5049e25ed09SBarry Smith } 505*bbb6d6a8SBarry Smith /* printf("i %d row %d nz %d idx %d fm %d level %d nnz %d\n",i,row,nz,idx,fm,*ajtmpf + 1,nnz); */ 5069e25ed09SBarry Smith ajtmpf++; 5079e25ed09SBarry Smith } 5089e25ed09SBarry Smith row = fill[row]; 5099e25ed09SBarry Smith } 5109e25ed09SBarry Smith /* copy new filled row into permanent storage */ 5119e25ed09SBarry Smith ainew[i+1] = ainew[i] + nnz; 5129e25ed09SBarry Smith if (ainew[i+1] > jmax+1) { 5139e25ed09SBarry Smith /* allocate a longer ajnew */ 514*bbb6d6a8SBarry Smith int maxadd; 515*bbb6d6a8SBarry Smith maxadd = (int) ((f*ai[n]*(n-i+5))/n); 516*bbb6d6a8SBarry Smith if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 517*bbb6d6a8SBarry Smith jmax += maxadd; 51878b31e54SBarry Smith ajtmp = (int *) PETSCMALLOC( jmax*sizeof(int) );CHKPTRQ(ajtmp); 51978b31e54SBarry Smith PETSCMEMCPY(ajtmp,ajnew,(ainew[i]-1)*sizeof(int)); 52078b31e54SBarry Smith PETSCFREE(ajnew); 5219e25ed09SBarry Smith ajnew = ajtmp; 5229e25ed09SBarry Smith /* allocate a longer ajfill */ 52378b31e54SBarry Smith ajtmp = (int *) PETSCMALLOC( jmax*sizeof(int) );CHKPTRQ(ajtmp); 52478b31e54SBarry Smith PETSCMEMCPY(ajtmp,ajfill,(ainew[i]-1)*sizeof(int)); 52578b31e54SBarry Smith PETSCFREE(ajfill); 5269e25ed09SBarry Smith ajfill = ajtmp; 527*bbb6d6a8SBarry Smith realloc++; 5289e25ed09SBarry Smith } 5299e25ed09SBarry Smith ajtmp = ajnew + ainew[i] - 1; 5309e25ed09SBarry Smith ajtmpf = ajfill + ainew[i] - 1; 5319e25ed09SBarry Smith fm = fill[n]; 532*bbb6d6a8SBarry Smith im[i] = nnz; 5339e25ed09SBarry Smith nzi = 0; 5349e25ed09SBarry Smith while (nnz--) { 5359e25ed09SBarry Smith if (fm < i) nzi++; 5369e25ed09SBarry Smith *ajtmp++ = fm + 1; 5379e25ed09SBarry Smith *ajtmpf++ = lfill[fm]; 538*bbb6d6a8SBarry Smith /* printf("i %d col %d level %d\n",i,fm,lfill[fm]); */ 5399e25ed09SBarry Smith fm = fill[fm]; 5409e25ed09SBarry Smith } 5419e25ed09SBarry Smith idnew[i] = ainew[i] + nzi; 5429e25ed09SBarry Smith } 54378b31e54SBarry Smith PETSCFREE(ajfill); 54478b31e54SBarry Smith ISDestroy(isicol); PETSCFREE(fill); PETSCFREE(lfill); 5459e25ed09SBarry Smith 546*bbb6d6a8SBarry Smith PLogInfo((PetscObject)mat, 547*bbb6d6a8SBarry Smith "Info:MatILUFactorSymbolic_AIJ:Realloc %d Fill ratio:given %g needed %g\n", 548*bbb6d6a8SBarry Smith realloc,f,((double)ainew[n])/((double)ai[i])); 549*bbb6d6a8SBarry Smith 5509e25ed09SBarry Smith /* put together the new matrix */ 55178b31e54SBarry Smith ierr = MatCreateSequentialAIJ(mat->comm,n, n, 0, 0, fact); CHKERRQ(ierr); 5521fb19edaSLois Curfman McInnes aijnew = (Mat_AIJ *) (*fact)->data; 55378b31e54SBarry Smith PETSCFREE(aijnew->imax); 5549e25ed09SBarry Smith aijnew->singlemalloc = 0; 5559e25ed09SBarry Smith len = (ainew[n] - 1)*sizeof(Scalar); 5569e25ed09SBarry Smith /* the next line frees the default space generated by the Create() */ 55778b31e54SBarry Smith PETSCFREE(aijnew->a); PETSCFREE(aijnew->ilen); 55878b31e54SBarry Smith aijnew->a = (Scalar *) PETSCMALLOC( len ); CHKPTRQ(aijnew->a); 5599e25ed09SBarry Smith aijnew->j = ajnew; 5609e25ed09SBarry Smith aijnew->i = ainew; 5619e25ed09SBarry Smith aijnew->diag = idnew; 5629e25ed09SBarry Smith aijnew->ilen = 0; 5639e25ed09SBarry Smith aijnew->imax = 0; 56486d7a8ceSBarry Smith aijnew->row = isrow; 56586d7a8ceSBarry Smith aijnew->col = iscol; 5665392566eSBarry Smith aijnew->solve_work = (Scalar *) PETSCMALLOC( (n+1)*sizeof(Scalar)); 56778b31e54SBarry Smith CHKPTRQ(aijnew->solve_work); 5687fd98bd6SLois Curfman McInnes /* In aijnew structure: Free imax, ilen, old a, old j. 5697fd98bd6SLois Curfman McInnes Allocate idnew, solve_work, new a, new j */ 5707fd98bd6SLois Curfman McInnes aijnew->mem += (ainew[n]-1-n)*(sizeof(int) + sizeof(Scalar)) + sizeof(int); 5717fd98bd6SLois Curfman McInnes aijnew->maxnz = aijnew->nz = ainew[n] - 1; 5729e25ed09SBarry Smith (*fact)->factor = FACTOR_LU; 5739e25ed09SBarry Smith return 0; 5749e25ed09SBarry Smith } 575