1cb512458SBarry Smith #ifndef lint 2*d5d45c9bSBarry Smith static char vcid[] = "$Id: aijfact.c,v 1.44 1995/11/01 23:17:58 bsmith Exp bsmith $"; 3cb512458SBarry Smith #endif 4289bc588SBarry Smith 5289bc588SBarry Smith #include "aij.h" 68c37ef55SBarry Smith #include "inline/spops.h" 7289bc588SBarry Smith /* 8289bc588SBarry Smith Factorization code for AIJ format. 9289bc588SBarry Smith */ 10289bc588SBarry Smith 11416022c9SBarry Smith int MatLUFactorSymbolic_SeqAIJ(Mat A,IS isrow,IS iscol,double f,Mat *B) 12289bc588SBarry Smith { 13416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data, *b; 14289bc588SBarry Smith IS isicol; 15416022c9SBarry Smith int *r,*ic, ierr, i, n = a->m, *ai = a->i, *aj = a->j; 16416022c9SBarry Smith int *ainew,*ajnew, jmax,*fill, *ajtmp, nz,shift = a->indexshift; 172fb3ed76SBarry Smith int *idnew, idx, row,m,fm, nnz, nzi,len, realloc = 0,nzbd,*im; 18289bc588SBarry Smith 19416022c9SBarry Smith if (n != a->n) SETERRQ(1,"MatLUFactorSymbolic_SeqAIJ:Must be square"); 20416022c9SBarry Smith if (!isrow) SETERRQ(1,"MatLUFactorSymbolic_SeqAIJ:Must have row permutation"); 21416022c9SBarry Smith if (!iscol) SETERRQ(1,"MatLUFactorSymbolic_SeqAIJ:Must have column permutation"); 22289bc588SBarry Smith 2378b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 24289bc588SBarry Smith ISGetIndices(isrow,&r); ISGetIndices(isicol,&ic); 25289bc588SBarry Smith 26289bc588SBarry Smith /* get new row pointers */ 270452661fSBarry Smith ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 28dbb450caSBarry Smith ainew[0] = -shift; 29289bc588SBarry Smith /* don't know how many column pointers are needed so estimate */ 30dbb450caSBarry Smith jmax = (int) (f*ai[n]+(!shift)); 310452661fSBarry Smith ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 32289bc588SBarry Smith /* fill is a linked list of nonzeros in active row */ 330452661fSBarry Smith fill = (int *) PetscMalloc( (2*n+1)*sizeof(int)); CHKPTRQ(fill); 342fb3ed76SBarry Smith im = fill + n + 1; 35289bc588SBarry Smith /* idnew is location of diagonal in factor */ 360452661fSBarry Smith idnew = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(idnew); 37dbb450caSBarry Smith idnew[0] = -shift; 38289bc588SBarry Smith 39289bc588SBarry Smith for ( i=0; i<n; i++ ) { 40289bc588SBarry Smith /* first copy previous fill into linked list */ 41289bc588SBarry Smith nnz = nz = ai[r[i]+1] - ai[r[i]]; 42dbb450caSBarry Smith ajtmp = aj + ai[r[i]] + shift; 43289bc588SBarry Smith fill[n] = n; 44289bc588SBarry Smith while (nz--) { 45289bc588SBarry Smith fm = n; 46dbb450caSBarry Smith idx = ic[*ajtmp++ + shift]; 47289bc588SBarry Smith do { 48289bc588SBarry Smith m = fm; 49289bc588SBarry Smith fm = fill[m]; 50289bc588SBarry Smith } while (fm < idx); 51289bc588SBarry Smith fill[m] = idx; 52289bc588SBarry Smith fill[idx] = fm; 53289bc588SBarry Smith } 54289bc588SBarry Smith row = fill[n]; 55289bc588SBarry Smith while ( row < i ) { 56dbb450caSBarry Smith ajtmp = ajnew + idnew[row] + (!shift); 572fb3ed76SBarry Smith nzbd = 1 + idnew[row] - ainew[row]; 582fb3ed76SBarry Smith nz = im[row] - nzbd; 59289bc588SBarry Smith fm = row; 602fb3ed76SBarry Smith while (nz-- > 0) { 61dbb450caSBarry Smith idx = *ajtmp++ + shift; 622fb3ed76SBarry Smith nzbd++; 632fb3ed76SBarry Smith if (idx == i) im[row] = nzbd; 64289bc588SBarry Smith do { 65289bc588SBarry Smith m = fm; 66289bc588SBarry Smith fm = fill[m]; 67289bc588SBarry Smith } while (fm < idx); 68289bc588SBarry Smith if (fm != idx) { 69289bc588SBarry Smith fill[m] = idx; 70289bc588SBarry Smith fill[idx] = fm; 71289bc588SBarry Smith fm = idx; 72289bc588SBarry Smith nnz++; 73289bc588SBarry Smith } 74289bc588SBarry Smith } 75289bc588SBarry Smith row = fill[row]; 76289bc588SBarry Smith } 77289bc588SBarry Smith /* copy new filled row into permanent storage */ 78289bc588SBarry Smith ainew[i+1] = ainew[i] + nnz; 79289bc588SBarry Smith if (ainew[i+1] > jmax+1) { 80289bc588SBarry Smith /* allocate a longer ajnew */ 812fb3ed76SBarry Smith int maxadd; 82dbb450caSBarry Smith maxadd = (int) ((f*(ai[n]+(!shift))*(n-i+5))/n); 83bbb6d6a8SBarry Smith if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 842fb3ed76SBarry Smith jmax += maxadd; 850452661fSBarry Smith ajtmp = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(ajtmp); 86416022c9SBarry Smith PetscMemcpy(ajtmp,ajnew,(ainew[i]+shift)*sizeof(int)); 870452661fSBarry Smith PetscFree(ajnew); 88289bc588SBarry Smith ajnew = ajtmp; 892fb3ed76SBarry Smith realloc++; /* count how many times we realloc */ 90289bc588SBarry Smith } 91dbb450caSBarry Smith ajtmp = ajnew + ainew[i] + shift; 92289bc588SBarry Smith fm = fill[n]; 93289bc588SBarry Smith nzi = 0; 942fb3ed76SBarry Smith im[i] = nnz; 95289bc588SBarry Smith while (nnz--) { 96289bc588SBarry Smith if (fm < i) nzi++; 97dbb450caSBarry Smith *ajtmp++ = fm - shift; 98289bc588SBarry Smith fm = fill[fm]; 99289bc588SBarry Smith } 100289bc588SBarry Smith idnew[i] = ainew[i] + nzi; 101289bc588SBarry Smith } 102289bc588SBarry Smith 103416022c9SBarry Smith PLogInfo((PetscObject)A, 104ec8511deSBarry Smith "Info:MatLUFactorSymbolic_SeqAIJ:Reallocs %d Fill ratio:given %g needed %g\n", 105bbb6d6a8SBarry Smith realloc,f,((double)ainew[n])/((double)ai[i])); 1062fb3ed76SBarry Smith 107898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 108898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 1091987afe7SBarry Smith 1100452661fSBarry Smith PetscFree(fill); 111289bc588SBarry Smith 112289bc588SBarry Smith /* put together the new matrix */ 113416022c9SBarry Smith ierr = MatCreateSeqAIJ(A->comm,n, n, 0, 0, B); CHKERRQ(ierr); 1141987afe7SBarry Smith PLogObjectParent(*B,isicol); 1151987afe7SBarry Smith ierr = ISDestroy(isicol); CHKERRQ(ierr); 116416022c9SBarry Smith b = (Mat_SeqAIJ *) (*B)->data; 1170452661fSBarry Smith PetscFree(b->imax); 118416022c9SBarry Smith b->singlemalloc = 0; 119dbb450caSBarry Smith len = (ainew[n] + shift)*sizeof(Scalar); 120e8d4e0b9SBarry Smith /* the next line frees the default space generated by the Create() */ 1210452661fSBarry Smith PetscFree(b->a); PetscFree(b->ilen); 1220452661fSBarry Smith b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 123416022c9SBarry Smith b->j = ajnew; 124416022c9SBarry Smith b->i = ainew; 125416022c9SBarry Smith b->diag = idnew; 126416022c9SBarry Smith b->ilen = 0; 127416022c9SBarry Smith b->imax = 0; 128416022c9SBarry Smith b->row = isrow; 129416022c9SBarry Smith b->col = iscol; 1300452661fSBarry Smith b->solve_work = (Scalar *) PetscMalloc( n*sizeof(Scalar)); 131416022c9SBarry Smith CHKPTRQ(b->solve_work); 132416022c9SBarry Smith /* In b structure: Free imax, ilen, old a, old j. 1337fd98bd6SLois Curfman McInnes Allocate idnew, solve_work, new a, new j */ 134416022c9SBarry Smith PLogObjectMemory(*B,(ainew[n]+shift-n)*(sizeof(int)+sizeof(Scalar))); 135416022c9SBarry Smith b->maxnz = b->nz = ainew[n] + shift; 1367fd98bd6SLois Curfman McInnes 137289bc588SBarry Smith return 0; 138289bc588SBarry Smith } 1390a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 140416022c9SBarry Smith int MatLUFactorNumeric_SeqAIJ(Mat A,Mat *B) 141289bc588SBarry Smith { 142416022c9SBarry Smith Mat C = *B; 143416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data, *b = (Mat_SeqAIJ *)C->data; 144416022c9SBarry Smith IS iscol = b->col, isrow = b->row, isicol; 145416022c9SBarry Smith int *r,*ic, ierr, i, j, n = a->m, *ai = b->i, *aj = b->j; 1461987afe7SBarry Smith int *ajtmpold, *ajtmp, nz, row, *ics, shift = a->indexshift; 147*d5d45c9bSBarry Smith int *diag_offset = b->diag; 148390f2bbaSBarry Smith Scalar *rtmp,*v, *pc, multiplier; 1491987afe7SBarry Smith /* These declarations are for optimizations. They reduce the number of 1501987afe7SBarry Smith memory references that are made by locally storing information; the 1511987afe7SBarry Smith word "register" used here with pointers can be viewed as "private" or 1521987afe7SBarry Smith "known only to me" 1531987afe7SBarry Smith */ 154390f2bbaSBarry Smith register Scalar *pv, *rtmps; 1551987afe7SBarry Smith register int *pj; 156289bc588SBarry Smith 15778b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 158416022c9SBarry Smith PLogObjectParent(*B,isicol); 15978b31e54SBarry Smith ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 16078b31e54SBarry Smith ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 1610452661fSBarry Smith rtmp = (Scalar *) PetscMalloc( (n+1)*sizeof(Scalar) ); CHKPTRQ(rtmp); 1620cf60383SBarry Smith rtmps = rtmp + shift; ics = ic + shift; 163289bc588SBarry Smith 164289bc588SBarry Smith for ( i=0; i<n; i++ ) { 165289bc588SBarry Smith nz = ai[i+1] - ai[i]; 166dbb450caSBarry Smith ajtmp = aj + ai[i] + shift; 1670cf60383SBarry Smith for ( j=0; j<nz; j++ ) rtmps[ajtmp[j]] = 0.0; 168289bc588SBarry Smith 169289bc588SBarry Smith /* load in initial (unfactored row) */ 170416022c9SBarry Smith nz = a->i[r[i]+1] - a->i[r[i]]; 171416022c9SBarry Smith ajtmpold = a->j + a->i[r[i]] + shift; 172416022c9SBarry Smith v = a->a + a->i[r[i]] + shift; 1730cf60383SBarry Smith for ( j=0; j<nz; j++ ) rtmp[ics[ajtmpold[j]]] = v[j]; 174289bc588SBarry Smith 175dbb450caSBarry Smith row = *ajtmp++ + shift; 176289bc588SBarry Smith while (row < i) { 1778c37ef55SBarry Smith pc = rtmp + row; 1788c37ef55SBarry Smith if (*pc != 0.0) { 1791987afe7SBarry Smith pv = b->a + diag_offset[row] + shift; 1801987afe7SBarry Smith pj = b->j + diag_offset[row] + (!shift); 1818c37ef55SBarry Smith multiplier = *pc * *pv++; 1828c37ef55SBarry Smith *pc = multiplier; 1831987afe7SBarry Smith nz = ai[row+1] - diag_offset[row] - 1; 184eef2e97bSBarry Smith PLogFlops(2*nz); 1851987afe7SBarry Smith /* The for-loop form can aid the compiler in overlapping 1861987afe7SBarry Smith loads and stores */ 1871987afe7SBarry Smith /*while (nz-->0) rtmps[*pj++] -= multiplier* *pv++; */ 1881987afe7SBarry Smith {int __i; 1891987afe7SBarry Smith for (__i=0; __i<nz; __i++) rtmps[pj[__i]] -= multiplier * pv[__i]; 1901987afe7SBarry Smith } 191289bc588SBarry Smith } 192dbb450caSBarry Smith row = *ajtmp++ + shift; 193289bc588SBarry Smith } 194416022c9SBarry Smith /* finished row so stick it into b->a */ 195416022c9SBarry Smith pv = b->a + ai[i] + shift; 196416022c9SBarry Smith pj = b->j + ai[i] + shift; 1978c37ef55SBarry Smith nz = ai[i+1] - ai[i]; 198ec8511deSBarry Smith if (rtmp[i] == 0.0) {SETERRQ(1,"MatLUFactorNumeric_SeqAIJ:Zero pivot");} 1998c37ef55SBarry Smith rtmp[i] = 1.0/rtmp[i]; 200416022c9SBarry Smith for ( j=0; j<nz; j++ ) {pv[j] = rtmps[pj[j]];} 2018c37ef55SBarry Smith } 2020452661fSBarry Smith PetscFree(rtmp); 20378b31e54SBarry Smith ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 20478b31e54SBarry Smith ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 20578b31e54SBarry Smith ierr = ISDestroy(isicol); CHKERRQ(ierr); 206416022c9SBarry Smith C->factor = FACTOR_LU; 207416022c9SBarry Smith b->assembled = 1; 208416022c9SBarry Smith PLogFlops(b->n); 209289bc588SBarry Smith return 0; 210289bc588SBarry Smith } 2110a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 212416022c9SBarry Smith int MatLUFactor_SeqAIJ(Mat A,IS row,IS col,double f) 213da3a660dSBarry Smith { 214416022c9SBarry Smith Mat_SeqAIJ *mat = (Mat_SeqAIJ *) A->data; 2156abc6512SBarry Smith int ierr; 216416022c9SBarry Smith Mat C; 217416022c9SBarry Smith 218416022c9SBarry Smith ierr = MatLUFactorSymbolic_SeqAIJ(A,row,col,f,&C); CHKERRQ(ierr); 219416022c9SBarry Smith ierr = MatLUFactorNumeric_SeqAIJ(A,&C); CHKERRQ(ierr); 220da3a660dSBarry Smith 221da3a660dSBarry Smith /* free all the data structures from mat */ 2220452661fSBarry Smith PetscFree(mat->a); 2230452661fSBarry Smith if (!mat->singlemalloc) {PetscFree(mat->i); PetscFree(mat->j);} 2240452661fSBarry Smith if (mat->diag) PetscFree(mat->diag); 2250452661fSBarry Smith if (mat->ilen) PetscFree(mat->ilen); 2260452661fSBarry Smith if (mat->imax) PetscFree(mat->imax); 2270452661fSBarry Smith if (mat->solve_work) PetscFree(mat->solve_work); 2280452661fSBarry Smith PetscFree(mat); 229da3a660dSBarry Smith 230416022c9SBarry Smith PetscMemcpy(A,C,sizeof(struct _Mat)); 2310452661fSBarry Smith PetscHeaderDestroy(C); 232da3a660dSBarry Smith return 0; 233da3a660dSBarry Smith } 2340a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 235416022c9SBarry Smith int MatSolve_SeqAIJ(Mat A,Vec bb, Vec xx) 2368c37ef55SBarry Smith { 237416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data; 238416022c9SBarry Smith IS iscol = a->col, isrow = a->row; 239416022c9SBarry Smith int *r,*c, ierr, i, n = a->m, *vi, *ai = a->i, *aj = a->j; 240416022c9SBarry Smith int nz,shift = a->indexshift; 241416022c9SBarry Smith Scalar *x,*b,*tmp, *tmps, *aa = a->a, sum, *v; 2428c37ef55SBarry Smith 243416022c9SBarry Smith if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolve_SeqAIJ:Not for unfactored matrix"); 2449e25ed09SBarry Smith 24578b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 24678b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 247416022c9SBarry Smith tmp = a->solve_work; 2488c37ef55SBarry Smith 24978b31e54SBarry Smith ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 25078b31e54SBarry Smith ierr = ISGetIndices(iscol,&c);CHKERRQ(ierr); c = c + (n-1); 2518c37ef55SBarry Smith 2528c37ef55SBarry Smith /* forward solve the lower triangular */ 2538c37ef55SBarry Smith tmp[0] = b[*r++]; 2540cf60383SBarry Smith tmps = tmp + shift; 2558c37ef55SBarry Smith for ( i=1; i<n; i++ ) { 256dbb450caSBarry Smith v = aa + ai[i] + shift; 257dbb450caSBarry Smith vi = aj + ai[i] + shift; 258416022c9SBarry Smith nz = a->diag[i] - ai[i]; 2598c37ef55SBarry Smith sum = b[*r++]; 2600cf60383SBarry Smith while (nz--) sum -= *v++ * tmps[*vi++]; 2618c37ef55SBarry Smith tmp[i] = sum; 2628c37ef55SBarry Smith } 2638c37ef55SBarry Smith 2648c37ef55SBarry Smith /* backward solve the upper triangular */ 2658c37ef55SBarry Smith for ( i=n-1; i>=0; i-- ){ 266416022c9SBarry Smith v = aa + a->diag[i] + (!shift); 267416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 268416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 2698c37ef55SBarry Smith sum = tmp[i]; 2700cf60383SBarry Smith while (nz--) sum -= *v++ * tmps[*vi++]; 271416022c9SBarry Smith x[*c--] = tmp[i] = sum*aa[a->diag[i]+shift]; 2728c37ef55SBarry Smith } 2738c37ef55SBarry Smith 274898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 275898183ecSLois Curfman McInnes ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 276416022c9SBarry Smith PLogFlops(2*a->nz - a->n); 2778c37ef55SBarry Smith return 0; 2788c37ef55SBarry Smith } 279416022c9SBarry Smith int MatSolveAdd_SeqAIJ(Mat A,Vec bb, Vec yy, Vec xx) 280da3a660dSBarry Smith { 281416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data; 282416022c9SBarry Smith IS iscol = a->col, isrow = a->row; 283416022c9SBarry Smith int *r,*c, ierr, i, n = a->m, *vi, *ai = a->i, *aj = a->j; 284416022c9SBarry Smith int nz, shift = a->indexshift; 285416022c9SBarry Smith Scalar *x,*b,*tmp, *aa = a->a, sum, *v; 286da3a660dSBarry Smith 287416022c9SBarry Smith if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveAdd_SeqAIJ:Not for unfactored matrix"); 28878b31e54SBarry Smith if (yy != xx) {ierr = VecCopy(yy,xx); CHKERRQ(ierr);} 289da3a660dSBarry Smith 29078b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 29178b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 292416022c9SBarry Smith tmp = a->solve_work; 293da3a660dSBarry Smith 29478b31e54SBarry Smith ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 29578b31e54SBarry Smith ierr = ISGetIndices(iscol,&c); CHKERRQ(ierr); c = c + (n-1); 296da3a660dSBarry Smith 297da3a660dSBarry Smith /* forward solve the lower triangular */ 298da3a660dSBarry Smith tmp[0] = b[*r++]; 299da3a660dSBarry Smith for ( i=1; i<n; i++ ) { 300dbb450caSBarry Smith v = aa + ai[i] + shift; 301dbb450caSBarry Smith vi = aj + ai[i] + shift; 302416022c9SBarry Smith nz = a->diag[i] - ai[i]; 303da3a660dSBarry Smith sum = b[*r++]; 304dbb450caSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ + shift]; 305da3a660dSBarry Smith tmp[i] = sum; 306da3a660dSBarry Smith } 307da3a660dSBarry Smith 308da3a660dSBarry Smith /* backward solve the upper triangular */ 309da3a660dSBarry Smith for ( i=n-1; i>=0; i-- ){ 310416022c9SBarry Smith v = aa + a->diag[i] + (!shift); 311416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 312416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 313da3a660dSBarry Smith sum = tmp[i]; 314dbb450caSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ + shift]; 315416022c9SBarry Smith tmp[i] = sum*aa[a->diag[i]+shift]; 316da3a660dSBarry Smith x[*c--] += tmp[i]; 317da3a660dSBarry Smith } 318da3a660dSBarry Smith 319898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 320898183ecSLois Curfman McInnes ierr = ISRestoreIndices(iscol,&c); CHKERRQ(ierr); 321416022c9SBarry Smith PLogFlops(2*a->nz); 322898183ecSLois Curfman McInnes 323da3a660dSBarry Smith return 0; 324da3a660dSBarry Smith } 325da3a660dSBarry Smith /* -------------------------------------------------------------------*/ 326416022c9SBarry Smith int MatSolveTrans_SeqAIJ(Mat A,Vec bb, Vec xx) 327da3a660dSBarry Smith { 328416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data; 329416022c9SBarry Smith IS iscol = a->col, isrow = a->row, invisrow,inviscol; 330416022c9SBarry Smith int *r,*c, ierr, i, n = a->m, *vi, *ai = a->i, *aj = a->j; 331416022c9SBarry Smith int nz,shift = a->indexshift; 332416022c9SBarry Smith Scalar *x,*b,*tmp, *aa = a->a, *v; 333da3a660dSBarry Smith 334416022c9SBarry Smith if (A->factor != FACTOR_LU) SETERRQ(1,"MatSolveTrans_SeqAIJ:Not unfactored matrix"); 33578b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 33678b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 337416022c9SBarry Smith tmp = a->solve_work; 338da3a660dSBarry Smith 339da3a660dSBarry Smith /* invert the permutations */ 34078b31e54SBarry Smith ierr = ISInvertPermutation(isrow,&invisrow); CHKERRQ(ierr); 34178b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&inviscol); CHKERRQ(ierr); 342da3a660dSBarry Smith 34378b31e54SBarry Smith ierr = ISGetIndices(invisrow,&r); CHKERRQ(ierr); 34478b31e54SBarry Smith ierr = ISGetIndices(inviscol,&c); CHKERRQ(ierr); 345da3a660dSBarry Smith 346da3a660dSBarry Smith /* copy the b into temp work space according to permutation */ 347da3a660dSBarry Smith for ( i=0; i<n; i++ ) tmp[c[i]] = b[i]; 348da3a660dSBarry Smith 349da3a660dSBarry Smith /* forward solve the U^T */ 350da3a660dSBarry Smith for ( i=0; i<n; i++ ) { 351416022c9SBarry Smith v = aa + a->diag[i] + shift; 352416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 353416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 354da3a660dSBarry Smith tmp[i] *= *v++; 355da3a660dSBarry Smith while (nz--) { 356dbb450caSBarry Smith tmp[*vi++ + shift] -= (*v++)*tmp[i]; 357da3a660dSBarry Smith } 358da3a660dSBarry Smith } 359da3a660dSBarry Smith 360da3a660dSBarry Smith /* backward solve the L^T */ 361da3a660dSBarry Smith for ( i=n-1; i>=0; i-- ){ 362416022c9SBarry Smith v = aa + a->diag[i] - 1 + shift; 363416022c9SBarry Smith vi = aj + a->diag[i] - 1 + shift; 364416022c9SBarry Smith nz = a->diag[i] - ai[i]; 365da3a660dSBarry Smith while (nz--) { 366dbb450caSBarry Smith tmp[*vi-- + shift] -= (*v--)*tmp[i]; 367da3a660dSBarry Smith } 368da3a660dSBarry Smith } 369da3a660dSBarry Smith 370da3a660dSBarry Smith /* copy tmp into x according to permutation */ 371da3a660dSBarry Smith for ( i=0; i<n; i++ ) x[r[i]] = tmp[i]; 372da3a660dSBarry Smith 373898183ecSLois Curfman McInnes ierr = ISRestoreIndices(invisrow,&r); CHKERRQ(ierr); 374898183ecSLois Curfman McInnes ierr = ISRestoreIndices(inviscol,&c); CHKERRQ(ierr); 375898183ecSLois Curfman McInnes ierr = ISDestroy(invisrow); CHKERRQ(ierr); 376898183ecSLois Curfman McInnes ierr = ISDestroy(inviscol); CHKERRQ(ierr); 377da3a660dSBarry Smith 378416022c9SBarry Smith PLogFlops(2*a->nz-a->n); 379da3a660dSBarry Smith return 0; 380da3a660dSBarry Smith } 381da3a660dSBarry Smith 382416022c9SBarry Smith int MatSolveTransAdd_SeqAIJ(Mat A,Vec bb, Vec zz,Vec xx) 383da3a660dSBarry Smith { 384416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data; 385416022c9SBarry Smith IS iscol = a->col, isrow = a->row, invisrow,inviscol; 386416022c9SBarry Smith int *r,*c, ierr, i, n = a->m, *vi, *ai = a->i, *aj = a->j; 387416022c9SBarry Smith int nz,shift = a->indexshift; 388416022c9SBarry Smith Scalar *x,*b,*tmp, *aa = a->a, *v; 3896abc6512SBarry Smith 390416022c9SBarry Smith if (A->factor != FACTOR_LU)SETERRQ(1,"MatSolveTransAdd_SeqAIJ:Not unfactored matrix"); 3916abc6512SBarry Smith if (zz != xx) VecCopy(zz,xx); 3926abc6512SBarry Smith 39378b31e54SBarry Smith ierr = VecGetArray(bb,&b); CHKERRQ(ierr); 39478b31e54SBarry Smith ierr = VecGetArray(xx,&x); CHKERRQ(ierr); 395416022c9SBarry Smith tmp = a->solve_work; 3966abc6512SBarry Smith 3976abc6512SBarry Smith /* invert the permutations */ 39878b31e54SBarry Smith ierr = ISInvertPermutation(isrow,&invisrow); CHKERRQ(ierr); 39978b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&inviscol); CHKERRQ(ierr); 40078b31e54SBarry Smith ierr = ISGetIndices(invisrow,&r); CHKERRQ(ierr); 40178b31e54SBarry Smith ierr = ISGetIndices(inviscol,&c); CHKERRQ(ierr); 4026abc6512SBarry Smith 4036abc6512SBarry Smith /* copy the b into temp work space according to permutation */ 4046abc6512SBarry Smith for ( i=0; i<n; i++ ) tmp[c[i]] = b[i]; 4056abc6512SBarry Smith 4066abc6512SBarry Smith /* forward solve the U^T */ 4076abc6512SBarry Smith for ( i=0; i<n; i++ ) { 408416022c9SBarry Smith v = aa + a->diag[i] + shift; 409416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 410416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 4116abc6512SBarry Smith tmp[i] *= *v++; 4126abc6512SBarry Smith while (nz--) { 413dbb450caSBarry Smith tmp[*vi++ + shift] -= (*v++)*tmp[i]; 4146abc6512SBarry Smith } 4156abc6512SBarry Smith } 4166abc6512SBarry Smith 4176abc6512SBarry Smith /* backward solve the L^T */ 4186abc6512SBarry Smith for ( i=n-1; i>=0; i-- ){ 419416022c9SBarry Smith v = aa + a->diag[i] - 1 + shift; 420416022c9SBarry Smith vi = aj + a->diag[i] - 1 + shift; 421416022c9SBarry Smith nz = a->diag[i] - ai[i]; 4226abc6512SBarry Smith while (nz--) { 423dbb450caSBarry Smith tmp[*vi-- + shift] -= (*v--)*tmp[i]; 4246abc6512SBarry Smith } 4256abc6512SBarry Smith } 4266abc6512SBarry Smith 4276abc6512SBarry Smith /* copy tmp into x according to permutation */ 4286abc6512SBarry Smith for ( i=0; i<n; i++ ) x[r[i]] += tmp[i]; 4296abc6512SBarry Smith 430898183ecSLois Curfman McInnes ierr = ISRestoreIndices(invisrow,&r); CHKERRQ(ierr); 431898183ecSLois Curfman McInnes ierr = ISRestoreIndices(inviscol,&c); CHKERRQ(ierr); 432898183ecSLois Curfman McInnes ierr = ISDestroy(invisrow); CHKERRQ(ierr); 433898183ecSLois Curfman McInnes ierr = ISDestroy(inviscol); CHKERRQ(ierr); 4346abc6512SBarry Smith 435416022c9SBarry Smith PLogFlops(2*a->nz); 4366abc6512SBarry Smith return 0; 437da3a660dSBarry Smith } 4389e25ed09SBarry Smith /* ----------------------------------------------------------------*/ 43908480c60SBarry Smith 44008480c60SBarry Smith int MatILUFactorSymbolic_SeqAIJ(Mat A,IS isrow,IS iscol,double f,int levels,Mat *fact) 4419e25ed09SBarry Smith { 442416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ *) A->data, *b; 4439e25ed09SBarry Smith IS isicol; 444416022c9SBarry Smith int *r,*ic, ierr, prow, n = a->m, *ai = a->i, *aj = a->j; 44556beaf04SBarry Smith int *ainew,*ajnew, jmax,*fill, *xi, nz, *im,*ajfill,*flev; 44656beaf04SBarry Smith int *dloc, idx, row,m,fm, nzf, nzi,len, realloc = 0; 447416022c9SBarry Smith int incrlev,nnz,i,shift = a->indexshift; 4489e25ed09SBarry Smith 449416022c9SBarry Smith if (n != a->n) SETERRQ(1,"MatILUFactorSymbolic_SeqAIJ:Matrix must be square"); 450416022c9SBarry Smith if (!isrow) SETERRQ(1,"MatILUFactorSymbolic_SeqAIJ:Must have row permutation"); 451416022c9SBarry Smith if (!iscol) SETERRQ(1,"MatILUFactorSymbolic_SeqAIJ:Must have column permutation"); 4529e25ed09SBarry Smith 45308480c60SBarry Smith /* special case that simply copies fill pattern */ 45408480c60SBarry Smith if (levels == 0 && ISIsIdentity(isrow) && ISIsIdentity(iscol)) { 45508480c60SBarry Smith ierr = MatCopyPrivate_SeqAIJ(A,fact,DO_NOT_COPY_VALUES); CHKERRQ(ierr); 45608480c60SBarry Smith (*fact)->factor = FACTOR_LU; 45708480c60SBarry Smith b = (Mat_SeqAIJ *) (*fact)->data; 45808480c60SBarry Smith if (!b->diag) { 45908480c60SBarry Smith ierr = MatMarkDiag_SeqAIJ(*fact); CHKERRQ(ierr); 46008480c60SBarry Smith } 46108480c60SBarry Smith b->row = isrow; 46208480c60SBarry Smith b->col = iscol; 4630452661fSBarry Smith b->solve_work = (Scalar *) PetscMalloc((b->m+1)*sizeof(Scalar));CHKPTRQ(b->solve_work); 46408480c60SBarry Smith return 0; 46508480c60SBarry Smith } 46608480c60SBarry Smith 46778b31e54SBarry Smith ierr = ISInvertPermutation(iscol,&isicol); CHKERRQ(ierr); 468898183ecSLois Curfman McInnes ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 469898183ecSLois Curfman McInnes ierr = ISGetIndices(isicol,&ic); CHKERRQ(ierr); 4709e25ed09SBarry Smith 4719e25ed09SBarry Smith /* get new row pointers */ 4720452661fSBarry Smith ainew = (int *) PetscMalloc( (n+1)*sizeof(int) ); CHKPTRQ(ainew); 473dbb450caSBarry Smith ainew[0] = -shift; 4749e25ed09SBarry Smith /* don't know how many column pointers are needed so estimate */ 475dbb450caSBarry Smith jmax = (int) (f*(ai[n]+!shift)); 4760452661fSBarry Smith ajnew = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajnew); 4779e25ed09SBarry Smith /* ajfill is level of fill for each fill entry */ 4780452661fSBarry Smith ajfill = (int *) PetscMalloc( (jmax)*sizeof(int) ); CHKPTRQ(ajfill); 4799e25ed09SBarry Smith /* fill is a linked list of nonzeros in active row */ 4800452661fSBarry Smith fill = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(fill); 48156beaf04SBarry Smith /* im is level for each filled value */ 4820452661fSBarry Smith im = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(im); 48356beaf04SBarry Smith /* dloc is location of diagonal in factor */ 4840452661fSBarry Smith dloc = (int *) PetscMalloc( (n+1)*sizeof(int)); CHKPTRQ(dloc); 48556beaf04SBarry Smith dloc[0] = 0; 4869e25ed09SBarry Smith 48756beaf04SBarry Smith for ( prow=0; prow<n; prow++ ) { 4889e25ed09SBarry Smith /* first copy previous fill into linked list */ 48956beaf04SBarry Smith nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 490dbb450caSBarry Smith xi = aj + ai[r[prow]] + shift; 4919e25ed09SBarry Smith fill[n] = n; 4929e25ed09SBarry Smith while (nz--) { 4939e25ed09SBarry Smith fm = n; 494dbb450caSBarry Smith idx = ic[*xi++ + shift]; 4959e25ed09SBarry Smith do { 4969e25ed09SBarry Smith m = fm; 4979e25ed09SBarry Smith fm = fill[m]; 4989e25ed09SBarry Smith } while (fm < idx); 4999e25ed09SBarry Smith fill[m] = idx; 5009e25ed09SBarry Smith fill[idx] = fm; 50156beaf04SBarry Smith im[idx] = 0; 5029e25ed09SBarry Smith } 50356beaf04SBarry Smith nzi = 0; 5049e25ed09SBarry Smith row = fill[n]; 50556beaf04SBarry Smith while ( row < prow ) { 50656beaf04SBarry Smith incrlev = im[row] + 1; 50756beaf04SBarry Smith nz = dloc[row]; 508dbb450caSBarry Smith xi = ajnew + ainew[row] + shift + nz; 509dbb450caSBarry Smith flev = ajfill + ainew[row] + shift + nz + 1; 510416022c9SBarry Smith nnz = ainew[row+1] - ainew[row] - nz - 1; 511dbb450caSBarry Smith if (*xi++ + shift != row) { 512ec8511deSBarry Smith SETERRQ(1,"MatILUFactorSymbolic_SeqAIJ:zero pivot"); 51356beaf04SBarry Smith } 51456beaf04SBarry Smith fm = row; 51556beaf04SBarry Smith while (nnz-- > 0) { 516dbb450caSBarry Smith idx = *xi++ + shift; 51756beaf04SBarry Smith if (*flev + incrlev > levels) { 51856beaf04SBarry Smith flev++; 51956beaf04SBarry Smith continue; 52056beaf04SBarry Smith } 52156beaf04SBarry Smith do { 5229e25ed09SBarry Smith m = fm; 5239e25ed09SBarry Smith fm = fill[m]; 52456beaf04SBarry Smith } while (fm < idx); 5259e25ed09SBarry Smith if (fm != idx) { 52656beaf04SBarry Smith im[idx] = *flev + incrlev; 5279e25ed09SBarry Smith fill[m] = idx; 5289e25ed09SBarry Smith fill[idx] = fm; 5299e25ed09SBarry Smith fm = idx; 53056beaf04SBarry Smith nzf++; 5319e25ed09SBarry Smith } 53256beaf04SBarry Smith else { 53356beaf04SBarry Smith if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 5349e25ed09SBarry Smith } 53556beaf04SBarry Smith flev++; 5369e25ed09SBarry Smith } 5379e25ed09SBarry Smith row = fill[row]; 53856beaf04SBarry Smith nzi++; 5399e25ed09SBarry Smith } 5409e25ed09SBarry Smith /* copy new filled row into permanent storage */ 54156beaf04SBarry Smith ainew[prow+1] = ainew[prow] + nzf; 54256beaf04SBarry Smith if (ainew[prow+1] > jmax+1) { 5439e25ed09SBarry Smith /* allocate a longer ajnew */ 544bbb6d6a8SBarry Smith int maxadd; 545dbb450caSBarry Smith maxadd = (int) ((f*(ai[n]+!shift)*(n-prow+5))/n); 54656beaf04SBarry Smith if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 547bbb6d6a8SBarry Smith jmax += maxadd; 5480452661fSBarry Smith xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 549416022c9SBarry Smith PetscMemcpy(xi,ajnew,(ainew[prow]+shift)*sizeof(int)); 5500452661fSBarry Smith PetscFree(ajnew); 55156beaf04SBarry Smith ajnew = xi; 5529e25ed09SBarry Smith /* allocate a longer ajfill */ 5530452661fSBarry Smith xi = (int *) PetscMalloc( jmax*sizeof(int) );CHKPTRQ(xi); 554416022c9SBarry Smith PetscMemcpy(xi,ajfill,(ainew[prow]+shift)*sizeof(int)); 5550452661fSBarry Smith PetscFree(ajfill); 55656beaf04SBarry Smith ajfill = xi; 557bbb6d6a8SBarry Smith realloc++; 5589e25ed09SBarry Smith } 559dbb450caSBarry Smith xi = ajnew + ainew[prow] + shift; 560dbb450caSBarry Smith flev = ajfill + ainew[prow] + shift; 56156beaf04SBarry Smith dloc[prow] = nzi; 5629e25ed09SBarry Smith fm = fill[n]; 56356beaf04SBarry Smith while (nzf--) { 564dbb450caSBarry Smith *xi++ = fm - shift; 56556beaf04SBarry Smith *flev++ = im[fm]; 5669e25ed09SBarry Smith fm = fill[fm]; 5679e25ed09SBarry Smith } 5689e25ed09SBarry Smith } 5690452661fSBarry Smith PetscFree(ajfill); 570898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r); CHKERRQ(ierr); 571898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isicol,&ic); CHKERRQ(ierr); 572898183ecSLois Curfman McInnes ierr = ISDestroy(isicol); CHKERRQ(ierr); 5730452661fSBarry Smith PetscFree(fill); PetscFree(im); 5749e25ed09SBarry Smith 575416022c9SBarry Smith PLogInfo((PetscObject)A, 576ec8511deSBarry Smith "Info:MatILUFactorSymbolic_SeqAIJ:Realloc %d Fill ratio:given %g needed %g\n", 57756beaf04SBarry Smith realloc,f,((double)ainew[n])/((double)ai[prow])); 578bbb6d6a8SBarry Smith 5799e25ed09SBarry Smith /* put together the new matrix */ 580416022c9SBarry Smith ierr = MatCreateSeqAIJ(A->comm,n, n, 0, 0, fact); CHKERRQ(ierr); 581416022c9SBarry Smith b = (Mat_SeqAIJ *) (*fact)->data; 5820452661fSBarry Smith PetscFree(b->imax); 583416022c9SBarry Smith b->singlemalloc = 0; 584dbb450caSBarry Smith len = (ainew[n] + shift)*sizeof(Scalar); 5859e25ed09SBarry Smith /* the next line frees the default space generated by the Create() */ 5860452661fSBarry Smith PetscFree(b->a); PetscFree(b->ilen); 5870452661fSBarry Smith b->a = (Scalar *) PetscMalloc( len ); CHKPTRQ(b->a); 588416022c9SBarry Smith b->j = ajnew; 589416022c9SBarry Smith b->i = ainew; 59056beaf04SBarry Smith for ( i=0; i<n; i++ ) dloc[i] += ainew[i]; 591416022c9SBarry Smith b->diag = dloc; 592416022c9SBarry Smith b->ilen = 0; 593416022c9SBarry Smith b->imax = 0; 594416022c9SBarry Smith b->row = isrow; 595416022c9SBarry Smith b->col = iscol; 5960452661fSBarry Smith b->solve_work = (Scalar *) PetscMalloc( (n+1)*sizeof(Scalar)); 597416022c9SBarry Smith CHKPTRQ(b->solve_work); 598416022c9SBarry Smith /* In b structure: Free imax, ilen, old a, old j. 59956beaf04SBarry Smith Allocate dloc, solve_work, new a, new j */ 600dbb450caSBarry Smith PLogObjectMemory(*fact,(ainew[n]+shift-n) * (sizeof(int)+sizeof(Scalar))); 601416022c9SBarry Smith b->maxnz = b->nz = ainew[n] + shift; 6029e25ed09SBarry Smith (*fact)->factor = FACTOR_LU; 6039e25ed09SBarry Smith return 0; 6049e25ed09SBarry Smith } 605*d5d45c9bSBarry Smith 606*d5d45c9bSBarry Smith 607*d5d45c9bSBarry Smith 608*d5d45c9bSBarry Smith 609