1*29bbc08cSBarry Smith /*$Id: aijfact.c,v 1.155 2000/09/28 18:50:54 bsmith Exp bsmith $*/ 2289bc588SBarry Smith 370f55243SBarry Smith #include "src/mat/impls/aij/seq/aij.h" 490f02eecSBarry Smith #include "src/vec/vecimpl.h" 51c4feecaSSatish Balay #include "src/inline/dot.h" 63b2fbd54SBarry Smith 75615d1e5SSatish Balay #undef __FUNC__ 8b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatOrdering_Flow_SeqAIJ" 991e9ee9fSBarry Smith int MatOrdering_Flow_SeqAIJ(Mat mat,MatOrderingType type,IS *irow,IS *icol) 103b2fbd54SBarry Smith { 113a40ed3dSBarry Smith PetscFunctionBegin; 123a40ed3dSBarry Smith 13*29bbc08cSBarry Smith SETERRQ(PETSC_ERR_SUP,"Code not written"); 14aa482453SBarry Smith #if !defined(PETSC_USE_DEBUG) 15d64ed03dSBarry Smith PetscFunctionReturn(0); 16d64ed03dSBarry Smith #endif 173b2fbd54SBarry Smith } 183b2fbd54SBarry Smith 1986bacbd2SBarry Smith 20ca44d042SBarry Smith EXTERN int MatMarkDiagonal_SeqAIJ(Mat); 21ca44d042SBarry Smith EXTERN int Mat_AIJ_CheckInode(Mat); 2286bacbd2SBarry Smith 23ca44d042SBarry Smith EXTERN int SPARSEKIT2dperm(int*,Scalar*,int*,int*,Scalar*,int*,int*,int*,int*,int*); 24ca44d042SBarry Smith EXTERN int SPARSEKIT2ilutp(int*,Scalar*,int*,int*,int*,PetscReal*,PetscReal*,int*,Scalar*,int*,int*,int*,Scalar*,int*,int*,int*); 25ca44d042SBarry Smith EXTERN int SPARSEKIT2msrcsr(int*,Scalar*,int*,Scalar*,int*,int*,Scalar*,int*); 2686bacbd2SBarry Smith 2786bacbd2SBarry Smith #undef __FUNC__ 28b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatILUDTFactor_SeqAIJ" 2986bacbd2SBarry Smith /* ------------------------------------------------------------ 3086bacbd2SBarry Smith 3186bacbd2SBarry Smith This interface was contribed by Tony Caola 3286bacbd2SBarry Smith 3386bacbd2SBarry Smith This routine is an interface to the pivoting drop-tolerance 345bd2ddc7SBarry Smith ILU routine written by Yousef Saad (saad@cs.umn.edu) as part of 355bd2ddc7SBarry Smith SPARSEKIT2. 365bd2ddc7SBarry Smith 375bd2ddc7SBarry Smith The SPARSEKIT2 routines used here are covered by the GNU 385bd2ddc7SBarry Smith copyright; see the file gnu in this directory. 3986bacbd2SBarry Smith 4086bacbd2SBarry Smith Thanks to Prof. Saad, Dr. Hysom, and Dr. Smith for their 4186bacbd2SBarry Smith help in getting this routine ironed out. 4286bacbd2SBarry Smith 435bd2ddc7SBarry Smith The major drawback to this routine is that if info->fill is 445bd2ddc7SBarry Smith not large enough it fails rather than allocating more space; 455bd2ddc7SBarry Smith this can be fixed by hacking/improving the f2c version of 465bd2ddc7SBarry Smith Yousef Saad's code. 4786bacbd2SBarry Smith 485bd2ddc7SBarry Smith ishift = 0, for indices start at 1 495bd2ddc7SBarry Smith ishift = 1, for indices starting at 0 5086bacbd2SBarry Smith ------------------------------------------------------------ 5186bacbd2SBarry Smith */ 5286bacbd2SBarry Smith 5386bacbd2SBarry Smith int MatILUDTFactor_SeqAIJ(Mat A,MatILUInfo *info,IS isrow,IS iscol,Mat *fact) 5486bacbd2SBarry Smith { 5586bacbd2SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 5607d2056aSBarry Smith IS iscolf,isicol,isirow; 5786bacbd2SBarry Smith PetscTruth reorder; 585bd2ddc7SBarry Smith int *c,*r,*ic,ierr,i,n = a->m; 59329f5518SBarry Smith int *old_i = a->i,*old_j = a->j,*new_i,*old_i2 = 0,*old_j2 = 0,*new_j; 60313ae024SBarry Smith int *ordcol,*iwk,*iperm,*jw; 615bd2ddc7SBarry Smith int ishift = !a->indexshift; 62b19713cbSBarry Smith int jmax,lfill,job,*o_i,*o_j; 63329f5518SBarry Smith Scalar *old_a = a->a,*w,*new_a,*old_a2 = 0,*wk,*o_a; 64329f5518SBarry Smith PetscReal permtol,af; 6586bacbd2SBarry Smith 6686bacbd2SBarry Smith PetscFunctionBegin; 6786bacbd2SBarry Smith 6886bacbd2SBarry Smith if (info->dt == PETSC_DEFAULT) info->dt = .005; 6986bacbd2SBarry Smith if (info->dtcount == PETSC_DEFAULT) info->dtcount = (int)(1.5*a->rmax); 7086bacbd2SBarry Smith if (info->dtcol == PETSC_DEFAULT) info->dtcol = .01; 7133259db3SBarry Smith if (info->fill == PETSC_DEFAULT) info->fill = ((double)(n*(info->dtcount+1)))/a->nz; 726faa4630SBarry Smith lfill = (int)(info->dtcount/2.0); 736faa4630SBarry Smith jmax = (int)(info->fill*a->nz); 7486bacbd2SBarry Smith permtol = info->dtcol; 7586bacbd2SBarry Smith 7686bacbd2SBarry Smith 7786bacbd2SBarry Smith /* ------------------------------------------------------------ 7886bacbd2SBarry Smith If reorder=.TRUE., then the original matrix has to be 7986bacbd2SBarry Smith reordered to reflect the user selected ordering scheme, and 8086bacbd2SBarry Smith then de-reordered so it is in it's original format. 8186bacbd2SBarry Smith Because Saad's dperm() is NOT in place, we have to copy 8286bacbd2SBarry Smith the original matrix and allocate more storage. . . 8386bacbd2SBarry Smith ------------------------------------------------------------ 8486bacbd2SBarry Smith */ 8586bacbd2SBarry Smith 8686bacbd2SBarry Smith /* set reorder to true if either isrow or iscol is not identity */ 8786bacbd2SBarry Smith ierr = ISIdentity(isrow,&reorder);CHKERRQ(ierr); 8886bacbd2SBarry Smith if (reorder) {ierr = ISIdentity(iscol,&reorder);CHKERRQ(ierr);} 8986bacbd2SBarry Smith reorder = PetscNot(reorder); 9086bacbd2SBarry Smith 9186bacbd2SBarry Smith 9286bacbd2SBarry Smith /* storage for ilu factor */ 9386bacbd2SBarry Smith new_i = (int*) PetscMalloc((n+1)*sizeof(int));CHKPTRQ(new_i); 9486bacbd2SBarry Smith new_j = (int*) PetscMalloc(jmax*sizeof(int));CHKPTRQ(new_j); 9586bacbd2SBarry Smith new_a = (Scalar*)PetscMalloc(jmax*sizeof(Scalar));CHKPTRQ(new_a); 9686bacbd2SBarry Smith 9786bacbd2SBarry Smith ordcol = (int*)PetscMalloc(n*sizeof(int));CHKPTRQ(ordcol); 9886bacbd2SBarry Smith 9986bacbd2SBarry Smith /* ------------------------------------------------------------ 10086bacbd2SBarry Smith Make sure that everything is Fortran formatted (1-Based) 10186bacbd2SBarry Smith ------------------------------------------------------------ 10286bacbd2SBarry Smith */ 103b19713cbSBarry Smith if (ishift) { 104b19713cbSBarry Smith for (i=old_i[0];i<old_i[n];i++) { 105b19713cbSBarry Smith old_j[i]++; 10686bacbd2SBarry Smith } 107b19713cbSBarry Smith for(i=0;i<n+1;i++) { 108b19713cbSBarry Smith old_i[i]++; 109b19713cbSBarry Smith }; 11086bacbd2SBarry Smith }; 11186bacbd2SBarry Smith 11286bacbd2SBarry Smith if (reorder) { 113c0c2c81eSBarry Smith ierr = ISGetIndices(iscol,&c); CHKERRQ(ierr); 114c0c2c81eSBarry Smith ierr = ISGetIndices(isrow,&r); CHKERRQ(ierr); 115c0c2c81eSBarry Smith for(i=0;i<n;i++) { 116c0c2c81eSBarry Smith r[i] = r[i]+1; 117c0c2c81eSBarry Smith c[i] = c[i]+1; 118c0c2c81eSBarry Smith } 119313ae024SBarry Smith old_i2 = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(old_i2); 120313ae024SBarry Smith old_j2 = (int*)PetscMalloc((old_i[n]-old_i[0]+1)*sizeof(int));CHKPTRQ(old_j2); 121313ae024SBarry Smith old_a2 = (Scalar*)PetscMalloc((old_i[n]-old_i[0]+1)*sizeof(Scalar));CHKPTRQ(old_a2); 1225bd2ddc7SBarry Smith job = 3; SPARSEKIT2dperm(&n,old_a,old_j,old_i,old_a2,old_j2,old_i2,r,c,&job); 123c0c2c81eSBarry Smith for (i=0;i<n;i++) { 124c0c2c81eSBarry Smith r[i] = r[i]-1; 125c0c2c81eSBarry Smith c[i] = c[i]-1; 126c0c2c81eSBarry Smith } 127c0c2c81eSBarry Smith ierr = ISRestoreIndices(iscol,&c);CHKERRQ(ierr); 128c0c2c81eSBarry Smith ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 129b19713cbSBarry Smith o_a = old_a2; 130b19713cbSBarry Smith o_j = old_j2; 131b19713cbSBarry Smith o_i = old_i2; 132b19713cbSBarry Smith } else { 133b19713cbSBarry Smith o_a = old_a; 134b19713cbSBarry Smith o_j = old_j; 135b19713cbSBarry Smith o_i = old_i; 13686bacbd2SBarry Smith } 13786bacbd2SBarry Smith 13886bacbd2SBarry Smith /* ------------------------------------------------------------ 13986bacbd2SBarry Smith Call Saad's ilutp() routine to generate the factorization 14086bacbd2SBarry Smith ------------------------------------------------------------ 14186bacbd2SBarry Smith */ 14286bacbd2SBarry Smith 14386bacbd2SBarry Smith iperm = (int*) PetscMalloc(2*n*sizeof(int));CHKPTRQ(iperm); 14486bacbd2SBarry Smith jw = (int*) PetscMalloc(2*n*sizeof(int));CHKPTRQ(jw); 14586bacbd2SBarry Smith w = (Scalar*)PetscMalloc(n*sizeof(Scalar));CHKPTRQ(w); 14686bacbd2SBarry Smith 1475bd2ddc7SBarry Smith SPARSEKIT2ilutp(&n,o_a,o_j,o_i,&lfill,&info->dt,&permtol,&n,new_a,new_j,new_i,&jmax,w,jw,iperm,&ierr); 14886bacbd2SBarry Smith if (ierr) { 14986bacbd2SBarry Smith switch (ierr) { 150*29bbc08cSBarry Smith case -3: SETERRQ2(1,"ilutp(), matrix U overflows, need larger info->fill current fill %g space allocated %d",info->fill,jmax); 151*29bbc08cSBarry Smith case -2: SETERRQ2(1,"ilutp(), matrix L overflows, need larger info->fill current fill %g space allocated %d",info->fill,jmax); 152*29bbc08cSBarry Smith case -5: SETERRQ(1,"ilutp(), zero row encountered"); 153*29bbc08cSBarry Smith case -1: SETERRQ(1,"ilutp(), input matrix may be wrong"); 154*29bbc08cSBarry Smith case -4: SETERRQ1(1,"ilutp(), illegal info->fill value %d",jmax); 155*29bbc08cSBarry Smith default: SETERRQ1(1,"ilutp(), zero pivot detected on row %d",ierr); 15686bacbd2SBarry Smith } 15786bacbd2SBarry Smith } 15886bacbd2SBarry Smith 15986bacbd2SBarry Smith ierr = PetscFree(w);CHKERRQ(ierr); 16086bacbd2SBarry Smith ierr = PetscFree(jw);CHKERRQ(ierr); 16186bacbd2SBarry Smith 16286bacbd2SBarry Smith /* ------------------------------------------------------------ 16386bacbd2SBarry Smith Saad's routine gives the result in Modified Sparse Row (msr) 16486bacbd2SBarry Smith Convert to Compressed Sparse Row format (csr) 16586bacbd2SBarry Smith ------------------------------------------------------------ 16686bacbd2SBarry Smith */ 16786bacbd2SBarry Smith 16886bacbd2SBarry Smith wk = (Scalar*)PetscMalloc(n*sizeof(Scalar));CHKPTRQ(wk); 16986bacbd2SBarry Smith iwk = (int*) PetscMalloc((n+1)*sizeof(int));CHKPTRQ(iwk); 17086bacbd2SBarry Smith 1715bd2ddc7SBarry Smith SPARSEKIT2msrcsr(&n,new_a,new_j,new_a,new_j,new_i,wk,iwk); 17286bacbd2SBarry Smith 17386bacbd2SBarry Smith ierr = PetscFree(iwk);CHKERRQ(ierr); 17486bacbd2SBarry Smith ierr = PetscFree(wk);CHKERRQ(ierr); 17586bacbd2SBarry Smith 17686bacbd2SBarry Smith if (reorder) { 17786bacbd2SBarry Smith ierr = PetscFree(old_a2);CHKERRQ(ierr); 17886bacbd2SBarry Smith ierr = PetscFree(old_j2);CHKERRQ(ierr); 17986bacbd2SBarry Smith ierr = PetscFree(old_i2);CHKERRQ(ierr); 180313ae024SBarry Smith } else { 181b19713cbSBarry Smith /* fix permutation of old_j that the factorization introduced */ 182f170005cSBarry Smith for (i=old_i[0]; i<old_i[n]; i++) { 183b19713cbSBarry Smith old_j[i-1] = iperm[old_j[i-1]-1]; 184b19713cbSBarry Smith } 185b19713cbSBarry Smith } 186b19713cbSBarry Smith 187b801d0f9SBarry Smith /* get rid of the shift to indices starting at 1 */ 188b19713cbSBarry Smith if (ishift) { 18986bacbd2SBarry Smith for (i=0; i<n+1; i++) { 190b19713cbSBarry Smith old_i[i]--; 191b19713cbSBarry Smith } 192b19713cbSBarry Smith for (i=old_i[0];i<old_i[n];i++) { 193b19713cbSBarry Smith old_j[i]--; 194b19713cbSBarry Smith } 19586bacbd2SBarry Smith } 19686bacbd2SBarry Smith 197b19713cbSBarry Smith /* Make the factored matrix 0-based */ 198b19713cbSBarry Smith if (ishift) { 19986bacbd2SBarry Smith for (i=0; i<n+1; i++) { 200b19713cbSBarry Smith new_i[i]--; 201b19713cbSBarry Smith } 202b19713cbSBarry Smith for (i=new_i[0];i<new_i[n];i++) { 203b19713cbSBarry Smith new_j[i]--; 204b19713cbSBarry Smith } 20586bacbd2SBarry Smith } 20686bacbd2SBarry Smith 20786bacbd2SBarry Smith /*-- due to the pivoting, we need to reorder iscol to correctly --*/ 20886bacbd2SBarry Smith /*-- permute the right-hand-side and solution vectors --*/ 2094c49b128SBarry Smith ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 2104c49b128SBarry Smith ierr = ISInvertPermutation(isrow,PETSC_DECIDE,&isirow);CHKERRQ(ierr); 211c0c2c81eSBarry Smith ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 21286bacbd2SBarry Smith for(i=0; i<n; i++) { 21386bacbd2SBarry Smith ordcol[i] = ic[iperm[i]-1]; 21486bacbd2SBarry Smith }; 21586bacbd2SBarry Smith ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 21607d2056aSBarry Smith ierr = ISDestroy(isicol);CHKERRQ(ierr); 21786bacbd2SBarry Smith 218c0c2c81eSBarry Smith ierr = PetscFree(iperm);CHKERRQ(ierr); 219c0c2c81eSBarry Smith 220329f5518SBarry Smith ierr = ISCreateGeneral(PETSC_COMM_SELF,n,ordcol,&iscolf);CHKERRQ(ierr); 22107d2056aSBarry Smith ierr = PetscFree(ordcol);CHKERRQ(ierr); 22286bacbd2SBarry Smith 22386bacbd2SBarry Smith /*----- put together the new matrix -----*/ 22486bacbd2SBarry Smith 22586bacbd2SBarry Smith ierr = MatCreateSeqAIJ(A->comm,n,n,0,PETSC_NULL,fact);CHKERRQ(ierr); 22686bacbd2SBarry Smith (*fact)->factor = FACTOR_LU; 22786bacbd2SBarry Smith (*fact)->assembled = PETSC_TRUE; 22886bacbd2SBarry Smith 22986bacbd2SBarry Smith b = (Mat_SeqAIJ*)(*fact)->data; 23086bacbd2SBarry Smith ierr = PetscFree(b->imax);CHKERRQ(ierr); 23186bacbd2SBarry Smith b->sorted = PETSC_FALSE; 23207d2056aSBarry Smith b->singlemalloc = PETSC_FALSE; 233b19713cbSBarry Smith /* the next line frees the default space generated by the MatCreate() */ 23486bacbd2SBarry Smith ierr = PetscFree(b->a);CHKERRQ(ierr); 23586bacbd2SBarry Smith ierr = PetscFree(b->ilen);CHKERRQ(ierr); 23686bacbd2SBarry Smith b->a = new_a; 23786bacbd2SBarry Smith b->j = new_j; 23886bacbd2SBarry Smith b->i = new_i; 23986bacbd2SBarry Smith b->ilen = 0; 24086bacbd2SBarry Smith b->imax = 0; 2411f9e874aSBarry Smith /* I am not sure why these are the inverses of the row and column permutations; but the other way is NO GOOD */ 242313ae024SBarry Smith b->row = isirow; 24386bacbd2SBarry Smith b->col = iscolf; 24486bacbd2SBarry Smith b->solve_work = (Scalar*)PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(b->solve_work); 24586bacbd2SBarry Smith b->maxnz = b->nz = new_i[n]; 24686bacbd2SBarry Smith b->indexshift = a->indexshift; 24786bacbd2SBarry Smith ierr = MatMarkDiagonal_SeqAIJ(*fact);CHKERRQ(ierr); 24886bacbd2SBarry Smith (*fact)->info.factor_mallocs = 0; 24986bacbd2SBarry Smith 25086bacbd2SBarry Smith ierr = MatMarkDiagonal_SeqAIJ(A);CHKERRQ(ierr); 25186bacbd2SBarry Smith 25286bacbd2SBarry Smith /* check out for identical nodes. If found, use inode functions */ 25386bacbd2SBarry Smith ierr = Mat_AIJ_CheckInode(*fact);CHKERRQ(ierr); 25486bacbd2SBarry Smith 255431e710aSBarry Smith af = ((double)b->nz)/((double)a->nz) + .001; 256431e710aSBarry Smith PLogInfo(A,"MatILUDTFactor_SeqAIJ:Fill ratio:given %g needed %g\n",info->fill,af); 257431e710aSBarry Smith PLogInfo(A,"MatILUDTFactor_SeqAIJ:Run with -pc_ilu_fill %g or use \n",af); 258431e710aSBarry Smith PLogInfo(A,"MatILUDTFactor_SeqAIJ:PCILUSetFill(pc,%g);\n",af); 259431e710aSBarry Smith PLogInfo(A,"MatILUDTFactor_SeqAIJ:for best performance.\n"); 260431e710aSBarry Smith 26186bacbd2SBarry Smith PetscFunctionReturn(0); 26286bacbd2SBarry Smith } 26386bacbd2SBarry Smith 264289bc588SBarry Smith /* 265289bc588SBarry Smith Factorization code for AIJ format. 266289bc588SBarry Smith */ 2675615d1e5SSatish Balay #undef __FUNC__ 2689e046878SBarry Smith #define __FUNC__ /*<a name="MatLUFactorSymbolic_SeqAIJ""></a>*/"MatLUFactorSymbolic_SeqAIJ" 2699e046878SBarry Smith int MatLUFactorSymbolic_SeqAIJ(Mat A,IS isrow,IS iscol,MatLUInfo *info,Mat *B) 270289bc588SBarry Smith { 271416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 272289bc588SBarry Smith IS isicol; 273416022c9SBarry Smith int *r,*ic,ierr,i,n = a->m,*ai = a->i,*aj = a->j; 274416022c9SBarry Smith int *ainew,*ajnew,jmax,*fill,*ajtmp,nz,shift = a->indexshift; 27591bf9adeSBarry Smith int *idnew,idx,row,m,fm,nnz,nzi,realloc = 0,nzbd,*im; 2769e046878SBarry Smith PetscReal f; 277289bc588SBarry Smith 2783a40ed3dSBarry Smith PetscFunctionBegin; 279d3cbd50cSLois Curfman McInnes PetscValidHeaderSpecific(isrow,IS_COOKIE); 280d3cbd50cSLois Curfman McInnes PetscValidHeaderSpecific(iscol,IS_COOKIE); 281*29bbc08cSBarry Smith if (A->M != A->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square"); 2823b2fbd54SBarry Smith 2839863cb4fSBarry Smith if (!isrow) { 2849863cb4fSBarry Smith ierr = ISCreateStride(PETSC_COMM_SELF,A->M,0,1,&isrow);CHKERRQ(ierr); 2859863cb4fSBarry Smith } 2869863cb4fSBarry Smith if (!iscol) { 2879863cb4fSBarry Smith ierr = ISCreateStride(PETSC_COMM_SELF,A->M,0,1,&iscol);CHKERRQ(ierr); 2889863cb4fSBarry Smith } 2899863cb4fSBarry Smith 2904c49b128SBarry Smith ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 291ac121b3dSBarry Smith ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 292ac121b3dSBarry Smith ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 293289bc588SBarry Smith 294289bc588SBarry Smith /* get new row pointers */ 2950452661fSBarry Smith ainew = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(ainew); 296dbb450caSBarry Smith ainew[0] = -shift; 297289bc588SBarry Smith /* don't know how many column pointers are needed so estimate */ 2989e046878SBarry Smith if (info) f = info->fill; else f = 1.0; 299dbb450caSBarry Smith jmax = (int)(f*ai[n]+(!shift)); 3000452661fSBarry Smith ajnew = (int*)PetscMalloc((jmax)*sizeof(int));CHKPTRQ(ajnew); 301289bc588SBarry Smith /* fill is a linked list of nonzeros in active row */ 3020452661fSBarry Smith fill = (int*)PetscMalloc((2*n+1)*sizeof(int));CHKPTRQ(fill); 3032fb3ed76SBarry Smith im = fill + n + 1; 304289bc588SBarry Smith /* idnew is location of diagonal in factor */ 3050452661fSBarry Smith idnew = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(idnew); 306dbb450caSBarry Smith idnew[0] = -shift; 307289bc588SBarry Smith 308289bc588SBarry Smith for (i=0; i<n; i++) { 309289bc588SBarry Smith /* first copy previous fill into linked list */ 310289bc588SBarry Smith nnz = nz = ai[r[i]+1] - ai[r[i]]; 311*29bbc08cSBarry Smith if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix"); 312dbb450caSBarry Smith ajtmp = aj + ai[r[i]] + shift; 313289bc588SBarry Smith fill[n] = n; 314289bc588SBarry Smith while (nz--) { 315289bc588SBarry Smith fm = n; 316dbb450caSBarry Smith idx = ic[*ajtmp++ + shift]; 317289bc588SBarry Smith do { 318289bc588SBarry Smith m = fm; 319289bc588SBarry Smith fm = fill[m]; 320289bc588SBarry Smith } while (fm < idx); 321289bc588SBarry Smith fill[m] = idx; 322289bc588SBarry Smith fill[idx] = fm; 323289bc588SBarry Smith } 324289bc588SBarry Smith row = fill[n]; 325289bc588SBarry Smith while (row < i) { 326dbb450caSBarry Smith ajtmp = ajnew + idnew[row] + (!shift); 3272fb3ed76SBarry Smith nzbd = 1 + idnew[row] - ainew[row]; 3282fb3ed76SBarry Smith nz = im[row] - nzbd; 329289bc588SBarry Smith fm = row; 3302fb3ed76SBarry Smith while (nz-- > 0) { 331dbb450caSBarry Smith idx = *ajtmp++ + shift; 3322fb3ed76SBarry Smith nzbd++; 3332fb3ed76SBarry Smith if (idx == i) im[row] = nzbd; 334289bc588SBarry Smith do { 335289bc588SBarry Smith m = fm; 336289bc588SBarry Smith fm = fill[m]; 337289bc588SBarry Smith } while (fm < idx); 338289bc588SBarry Smith if (fm != idx) { 339289bc588SBarry Smith fill[m] = idx; 340289bc588SBarry Smith fill[idx] = fm; 341289bc588SBarry Smith fm = idx; 342289bc588SBarry Smith nnz++; 343289bc588SBarry Smith } 344289bc588SBarry Smith } 345289bc588SBarry Smith row = fill[row]; 346289bc588SBarry Smith } 347289bc588SBarry Smith /* copy new filled row into permanent storage */ 348289bc588SBarry Smith ainew[i+1] = ainew[i] + nnz; 349496e697eSBarry Smith if (ainew[i+1] > jmax) { 350ecf371e4SBarry Smith 351ecf371e4SBarry Smith /* estimate how much additional space we will need */ 352ecf371e4SBarry Smith /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 353ecf371e4SBarry Smith /* just double the memory each time */ 354ecf371e4SBarry Smith int maxadd = jmax; 355ecf371e4SBarry Smith /* maxadd = (int)((f*(ai[n]+(!shift))*(n-i+5))/n); */ 356bbb6d6a8SBarry Smith if (maxadd < nnz) maxadd = (n-i)*(nnz+1); 3572fb3ed76SBarry Smith jmax += maxadd; 358ecf371e4SBarry Smith 359ecf371e4SBarry Smith /* allocate a longer ajnew */ 3600452661fSBarry Smith ajtmp = (int*)PetscMalloc(jmax*sizeof(int));CHKPTRQ(ajtmp); 361549d3d68SSatish Balay ierr = PetscMemcpy(ajtmp,ajnew,(ainew[i]+shift)*sizeof(int));CHKERRQ(ierr); 362606d414cSSatish Balay ierr = PetscFree(ajnew);CHKERRQ(ierr); 363289bc588SBarry Smith ajnew = ajtmp; 3642fb3ed76SBarry Smith realloc++; /* count how many times we realloc */ 365289bc588SBarry Smith } 366dbb450caSBarry Smith ajtmp = ajnew + ainew[i] + shift; 367289bc588SBarry Smith fm = fill[n]; 368289bc588SBarry Smith nzi = 0; 3692fb3ed76SBarry Smith im[i] = nnz; 370289bc588SBarry Smith while (nnz--) { 371289bc588SBarry Smith if (fm < i) nzi++; 372dbb450caSBarry Smith *ajtmp++ = fm - shift; 373289bc588SBarry Smith fm = fill[fm]; 374289bc588SBarry Smith } 375289bc588SBarry Smith idnew[i] = ainew[i] + nzi; 376289bc588SBarry Smith } 377464e8d44SSatish Balay if (ai[n] != 0) { 378329f5518SBarry Smith PetscReal af = ((PetscReal)ainew[n])/((PetscReal)ai[n]); 379c38d4ed2SBarry Smith PLogInfo(A,"MatLUFactorSymbolic_SeqAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af); 380981c4779SBarry Smith PLogInfo(A,"MatLUFactorSymbolic_SeqAIJ:Run with -pc_lu_fill %g or use \n",af); 381981c4779SBarry Smith PLogInfo(A,"MatLUFactorSymbolic_SeqAIJ:PCLUSetFill(pc,%g);\n",af); 382981c4779SBarry Smith PLogInfo(A,"MatLUFactorSymbolic_SeqAIJ:for best performance.\n"); 38305bf559cSSatish Balay } else { 384981c4779SBarry Smith PLogInfo(A,"MatLUFactorSymbolic_SeqAIJ: Empty matrix\n"); 38505bf559cSSatish Balay } 3862fb3ed76SBarry Smith 387898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 388898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 3891987afe7SBarry Smith 390606d414cSSatish Balay ierr = PetscFree(fill);CHKERRQ(ierr); 391289bc588SBarry Smith 392289bc588SBarry Smith /* put together the new matrix */ 393b4fd4287SBarry Smith ierr = MatCreateSeqAIJ(A->comm,n,n,0,PETSC_NULL,B);CHKERRQ(ierr); 3941987afe7SBarry Smith PLogObjectParent(*B,isicol); 395416022c9SBarry Smith b = (Mat_SeqAIJ*)(*B)->data; 396606d414cSSatish Balay ierr = PetscFree(b->imax);CHKERRQ(ierr); 3977c922b88SBarry Smith b->singlemalloc = PETSC_FALSE; 398e8d4e0b9SBarry Smith /* the next line frees the default space generated by the Create() */ 399606d414cSSatish Balay ierr = PetscFree(b->a);CHKERRQ(ierr); 400606d414cSSatish Balay ierr = PetscFree(b->ilen);CHKERRQ(ierr); 40191bf9adeSBarry Smith b->a = (Scalar*)PetscMalloc((ainew[n]+shift+1)*sizeof(Scalar));CHKPTRQ(b->a); 402416022c9SBarry Smith b->j = ajnew; 403416022c9SBarry Smith b->i = ainew; 404416022c9SBarry Smith b->diag = idnew; 405416022c9SBarry Smith b->ilen = 0; 406416022c9SBarry Smith b->imax = 0; 407416022c9SBarry Smith b->row = isrow; 408416022c9SBarry Smith b->col = iscol; 409085256b3SBarry Smith if (info) { 4101d1367b7SBarry Smith b->lu_damp = (PetscTruth) ((int)info->damp); 411085256b3SBarry Smith b->lu_damping = info->damping; 412085256b3SBarry Smith } else { 413085256b3SBarry Smith b->lu_damp = PETSC_FALSE; 414085256b3SBarry Smith b->lu_damping = 0.0; 415085256b3SBarry Smith } 416c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 417c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 41882bf6240SBarry Smith b->icol = isicol; 41991bf9adeSBarry Smith b->solve_work = (Scalar*)PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(b->solve_work); 420416022c9SBarry Smith /* In b structure: Free imax, ilen, old a, old j. 4217fd98bd6SLois Curfman McInnes Allocate idnew, solve_work, new a, new j */ 422416022c9SBarry Smith PLogObjectMemory(*B,(ainew[n]+shift-n)*(sizeof(int)+sizeof(Scalar))); 423416022c9SBarry Smith b->maxnz = b->nz = ainew[n] + shift; 4247fd98bd6SLois Curfman McInnes 425329f5518SBarry Smith (*B)->factor = FACTOR_LU; 426ae068f58SLois Curfman McInnes (*B)->info.factor_mallocs = realloc; 427ae068f58SLois Curfman McInnes (*B)->info.fill_ratio_given = f; 4287dda7e10SBarry Smith (*B)->ops->lufactornumeric = A->ops->lufactornumeric; /* Use Inode variant ONLY if A has inodes */ 429703deb49SSatish Balay 430e93ccc4dSBarry Smith if (ai[n] != 0) { 431329f5518SBarry Smith (*B)->info.fill_ratio_needed = ((PetscReal)ainew[n])/((PetscReal)ai[n]); 432eea2913aSSatish Balay } else { 433eea2913aSSatish Balay (*B)->info.fill_ratio_needed = 0.0; 434eea2913aSSatish Balay } 4353a40ed3dSBarry Smith PetscFunctionReturn(0); 436289bc588SBarry Smith } 4370a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 438ca44d042SBarry Smith EXTERN int Mat_AIJ_CheckInode(Mat); 43941c01911SSatish Balay 4405615d1e5SSatish Balay #undef __FUNC__ 4419e046878SBarry Smith #define __FUNC__ /*<a name="MatLUFactorNumeric_SeqAIJ"></a>*/"MatLUFactorNumeric_SeqAIJ" 442416022c9SBarry Smith int MatLUFactorNumeric_SeqAIJ(Mat A,Mat *B) 443289bc588SBarry Smith { 444416022c9SBarry Smith Mat C = *B; 445416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ *)C->data; 44682bf6240SBarry Smith IS isrow = b->row,isicol = b->icol; 447416022c9SBarry Smith int *r,*ic,ierr,i,j,n = a->m,*ai = b->i,*aj = b->j; 4481987afe7SBarry Smith int *ajtmpold,*ajtmp,nz,row,*ics,shift = a->indexshift; 449085256b3SBarry Smith int *diag_offset = b->diag,diag; 450085256b3SBarry Smith int *pj,ndamp = 0; 451085256b3SBarry Smith Scalar *rtmp,*v,*pc,multiplier,*pv,*rtmps; 452085256b3SBarry Smith PetscReal damping = b->lu_damping; 453085256b3SBarry Smith PetscTruth damp; 454289bc588SBarry Smith 4553a40ed3dSBarry Smith PetscFunctionBegin; 45682bf6240SBarry Smith 45778b31e54SBarry Smith ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 45878b31e54SBarry Smith ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 4590452661fSBarry Smith rtmp = (Scalar*)PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(rtmp); 460549d3d68SSatish Balay ierr = PetscMemzero(rtmp,(n+1)*sizeof(Scalar));CHKERRQ(ierr); 4610cf60383SBarry Smith rtmps = rtmp + shift; ics = ic + shift; 462289bc588SBarry Smith 463085256b3SBarry Smith do { 464085256b3SBarry Smith damp = PETSC_FALSE; 465289bc588SBarry Smith for (i=0; i<n; i++) { 466289bc588SBarry Smith nz = ai[i+1] - ai[i]; 467dbb450caSBarry Smith ajtmp = aj + ai[i] + shift; 46848e94559SBarry Smith for (j=0; j<nz; j++) rtmps[ajtmp[j]] = 0.0; 469289bc588SBarry Smith 470289bc588SBarry Smith /* load in initial (unfactored row) */ 471416022c9SBarry Smith nz = a->i[r[i]+1] - a->i[r[i]]; 472416022c9SBarry Smith ajtmpold = a->j + a->i[r[i]] + shift; 473416022c9SBarry Smith v = a->a + a->i[r[i]] + shift; 474085256b3SBarry Smith for (j=0; j<nz; j++) { 475085256b3SBarry Smith rtmp[ics[ajtmpold[j]]] = v[j]; 476085256b3SBarry Smith if (ajtmpold[j] == r[i]) { /* damp the diagonal of the matrix */ 477085256b3SBarry Smith rtmp[ics[ajtmpold[j]]] += damping; 478085256b3SBarry Smith } 479085256b3SBarry Smith } 480289bc588SBarry Smith 481dbb450caSBarry Smith row = *ajtmp++ + shift; 482f2582111SSatish Balay while (row < i) { 4838c37ef55SBarry Smith pc = rtmp + row; 4848c37ef55SBarry Smith if (*pc != 0.0) { 4851987afe7SBarry Smith pv = b->a + diag_offset[row] + shift; 4861987afe7SBarry Smith pj = b->j + diag_offset[row] + (!shift); 48735aab85fSBarry Smith multiplier = *pc / *pv++; 4888c37ef55SBarry Smith *pc = multiplier; 489f2582111SSatish Balay nz = ai[row+1] - diag_offset[row] - 1; 490f2582111SSatish Balay for (j=0; j<nz; j++) rtmps[pj[j]] -= multiplier * pv[j]; 491f2582111SSatish Balay PLogFlops(2*nz); 492289bc588SBarry Smith } 493dbb450caSBarry Smith row = *ajtmp++ + shift; 494289bc588SBarry Smith } 495416022c9SBarry Smith /* finished row so stick it into b->a */ 496416022c9SBarry Smith pv = b->a + ai[i] + shift; 497416022c9SBarry Smith pj = b->j + ai[i] + shift; 4988c37ef55SBarry Smith nz = ai[i+1] - ai[i]; 499416022c9SBarry Smith for (j=0; j<nz; j++) {pv[j] = rtmps[pj[j]];} 50035aab85fSBarry Smith diag = diag_offset[i] - ai[i]; 501085256b3SBarry Smith if (PetscAbsScalar(pv[diag]) < 1.e-12) { 502085256b3SBarry Smith if (b->lu_damp) { 503085256b3SBarry Smith damp = PETSC_TRUE; 504085256b3SBarry Smith if (damping) damping *= 2.0; 505085256b3SBarry Smith else damping = 1.e-12; 506085256b3SBarry Smith ndamp++; 507085256b3SBarry Smith break; 508085256b3SBarry Smith } else { 509*29bbc08cSBarry Smith SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %d",i); 510d708749eSBarry Smith } 51135aab85fSBarry Smith } 51235aab85fSBarry Smith } 513085256b3SBarry Smith } while (damp); 5140f11f9aeSSatish Balay 51535aab85fSBarry Smith /* invert diagonal entries for simplier triangular solves */ 51635aab85fSBarry Smith for (i=0; i<n; i++) { 51735aab85fSBarry Smith b->a[diag_offset[i]+shift] = 1.0/b->a[diag_offset[i]+shift]; 51835aab85fSBarry Smith } 51935aab85fSBarry Smith 520606d414cSSatish Balay ierr = PetscFree(rtmp);CHKERRQ(ierr); 52178b31e54SBarry Smith ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 52278b31e54SBarry Smith ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 523416022c9SBarry Smith C->factor = FACTOR_LU; 52441c01911SSatish Balay ierr = Mat_AIJ_CheckInode(C);CHKERRQ(ierr); 5257dda7e10SBarry Smith (*B)->ops->lufactornumeric = A->ops->lufactornumeric; /* Use Inode variant ONLY if A has inodes */ 526c456f294SBarry Smith C->assembled = PETSC_TRUE; 527416022c9SBarry Smith PLogFlops(b->n); 528a2e34c3dSBarry Smith if (ndamp || b->lu_damping) { 529085256b3SBarry Smith PLogInfo(0,"MatLUFactorNumerical_SeqAIJ: number of damping tries %d damping value %g\n",ndamp,damping); 530085256b3SBarry Smith } 5313a40ed3dSBarry Smith PetscFunctionReturn(0); 532289bc588SBarry Smith } 5330a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 5345615d1e5SSatish Balay #undef __FUNC__ 5359e046878SBarry Smith #define __FUNC__ /*<a name="MatLUFactor_SeqAIJ"></a>*/"MatLUFactor_SeqAIJ" 5369e046878SBarry Smith int MatLUFactor_SeqAIJ(Mat A,IS row,IS col,MatLUInfo *info) 537da3a660dSBarry Smith { 538416022c9SBarry Smith Mat_SeqAIJ *mat = (Mat_SeqAIJ*)A->data; 53986bacbd2SBarry Smith int ierr,refct; 540416022c9SBarry Smith Mat C; 541f830108cSBarry Smith PetscOps *Abops; 5420a6ffc59SBarry Smith MatOps Aops; 543416022c9SBarry Smith 5443a40ed3dSBarry Smith PetscFunctionBegin; 5459e046878SBarry Smith ierr = MatLUFactorSymbolic(A,row,col,info,&C);CHKERRQ(ierr); 546f2582111SSatish Balay ierr = MatLUFactorNumeric(A,&C);CHKERRQ(ierr); 547da3a660dSBarry Smith 548da3a660dSBarry Smith /* free all the data structures from mat */ 549606d414cSSatish Balay ierr = PetscFree(mat->a);CHKERRQ(ierr); 550606d414cSSatish Balay if (!mat->singlemalloc) { 551606d414cSSatish Balay ierr = PetscFree(mat->i);CHKERRQ(ierr); 552606d414cSSatish Balay ierr = PetscFree(mat->j);CHKERRQ(ierr); 553606d414cSSatish Balay } 554606d414cSSatish Balay if (mat->diag) {ierr = PetscFree(mat->diag);CHKERRQ(ierr);} 555606d414cSSatish Balay if (mat->ilen) {ierr = PetscFree(mat->ilen);CHKERRQ(ierr);} 556606d414cSSatish Balay if (mat->imax) {ierr = PetscFree(mat->imax);CHKERRQ(ierr);} 557606d414cSSatish Balay if (mat->solve_work) {ierr = PetscFree(mat->solve_work);CHKERRQ(ierr);} 558606d414cSSatish Balay if (mat->inode.size) {ierr = PetscFree(mat->inode.size);CHKERRQ(ierr);} 5590f0aacb9SBarry Smith if (mat->icol) {ierr = ISDestroy(mat->icol);CHKERRQ(ierr);} 560606d414cSSatish Balay ierr = PetscFree(mat);CHKERRQ(ierr); 561da3a660dSBarry Smith 56217642b18SBarry Smith ierr = MapDestroy(A->rmap);CHKERRQ(ierr); 56317642b18SBarry Smith ierr = MapDestroy(A->cmap);CHKERRQ(ierr); 56417642b18SBarry Smith 565f830108cSBarry Smith /* 566f830108cSBarry Smith This is horrible, horrible code. We need to keep the 567f830108cSBarry Smith A pointers for the bops and ops but copy everything 568f830108cSBarry Smith else from C. 569f830108cSBarry Smith */ 570f830108cSBarry Smith Abops = A->bops; 571f830108cSBarry Smith Aops = A->ops; 57286bacbd2SBarry Smith refct = A->refct; 573549d3d68SSatish Balay ierr = PetscMemcpy(A,C,sizeof(struct _p_Mat));CHKERRQ(ierr); 574451c4af8SSatish Balay mat = (Mat_SeqAIJ*)A->data; 575451c4af8SSatish Balay PLogObjectParent(A,mat->icol); 576451c4af8SSatish Balay 577f830108cSBarry Smith A->bops = Abops; 578f830108cSBarry Smith A->ops = Aops; 579bef8e0ddSBarry Smith A->qlist = 0; 58086bacbd2SBarry Smith A->refct = refct; 581c38d4ed2SBarry Smith /* copy over the type_name and name */ 582c38d4ed2SBarry Smith ierr = PetscStrallocpy(C->type_name,&A->type_name);CHKERRQ(ierr); 583c38d4ed2SBarry Smith ierr = PetscStrallocpy(C->name,&A->name);CHKERRQ(ierr); 584f830108cSBarry Smith 5850452661fSBarry Smith PetscHeaderDestroy(C); 586c38d4ed2SBarry Smith 5873a40ed3dSBarry Smith PetscFunctionReturn(0); 588da3a660dSBarry Smith } 5890a6dbcc7SLois Curfman McInnes /* ----------------------------------------------------------- */ 5905615d1e5SSatish Balay #undef __FUNC__ 591b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatSolve_SeqAIJ" 592416022c9SBarry Smith int MatSolve_SeqAIJ(Mat A,Vec bb,Vec xx) 5938c37ef55SBarry Smith { 594416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 595416022c9SBarry Smith IS iscol = a->col,isrow = a->row; 596416022c9SBarry Smith int *r,*c,ierr,i, n = a->m,*vi,*ai = a->i,*aj = a->j; 5973b2fbd54SBarry Smith int nz,shift = a->indexshift,*rout,*cout; 598416022c9SBarry Smith Scalar *x,*b,*tmp,*tmps,*aa = a->a,sum,*v; 5998c37ef55SBarry Smith 6003a40ed3dSBarry Smith PetscFunctionBegin; 6013a40ed3dSBarry Smith if (!n) PetscFunctionReturn(0); 60291bf9adeSBarry Smith 603e1311b90SBarry Smith ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 604e1311b90SBarry Smith ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 605416022c9SBarry Smith tmp = a->solve_work; 6068c37ef55SBarry Smith 6073b2fbd54SBarry Smith ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 6083b2fbd54SBarry Smith ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 6098c37ef55SBarry Smith 6108c37ef55SBarry Smith /* forward solve the lower triangular */ 6118c37ef55SBarry Smith tmp[0] = b[*r++]; 6120cf60383SBarry Smith tmps = tmp + shift; 6138c37ef55SBarry Smith for (i=1; i<n; i++) { 614dbb450caSBarry Smith v = aa + ai[i] + shift; 615dbb450caSBarry Smith vi = aj + ai[i] + shift; 616416022c9SBarry Smith nz = a->diag[i] - ai[i]; 6178c37ef55SBarry Smith sum = b[*r++]; 6180cf60383SBarry Smith while (nz--) sum -= *v++ * tmps[*vi++]; 6198c37ef55SBarry Smith tmp[i] = sum; 6208c37ef55SBarry Smith } 6218c37ef55SBarry Smith 6228c37ef55SBarry Smith /* backward solve the upper triangular */ 6238c37ef55SBarry Smith for (i=n-1; i>=0; i--){ 624416022c9SBarry Smith v = aa + a->diag[i] + (!shift); 625416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 626416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 6278c37ef55SBarry Smith sum = tmp[i]; 6280cf60383SBarry Smith while (nz--) sum -= *v++ * tmps[*vi++]; 629416022c9SBarry Smith x[*c--] = tmp[i] = sum*aa[a->diag[i]+shift]; 6308c37ef55SBarry Smith } 6318c37ef55SBarry Smith 6323b2fbd54SBarry Smith ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 6333b2fbd54SBarry Smith ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 634cb5b572fSBarry Smith ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 635cb5b572fSBarry Smith ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 636416022c9SBarry Smith PLogFlops(2*a->nz - a->n); 6373a40ed3dSBarry Smith PetscFunctionReturn(0); 6388c37ef55SBarry Smith } 639026e39d0SSatish Balay 640930ae53cSBarry Smith /* ----------------------------------------------------------- */ 641930ae53cSBarry Smith #undef __FUNC__ 642b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatSolve_SeqAIJ_NaturalOrdering" 643930ae53cSBarry Smith int MatSolve_SeqAIJ_NaturalOrdering(Mat A,Vec bb,Vec xx) 644930ae53cSBarry Smith { 645930ae53cSBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 646d85afc42SSatish Balay int n = a->m,*ai = a->i,*aj = a->j,*adiag = a->diag,ierr; 647d85afc42SSatish Balay Scalar *x,*b,*aa = a->a,sum; 648aa482453SBarry Smith #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 649d85afc42SSatish Balay int adiag_i,i,*vi,nz,ai_i; 650d85afc42SSatish Balay Scalar *v; 651d85afc42SSatish Balay #endif 652930ae53cSBarry Smith 6533a40ed3dSBarry Smith PetscFunctionBegin; 6543a40ed3dSBarry Smith if (!n) PetscFunctionReturn(0); 655930ae53cSBarry Smith if (a->indexshift) { 6563a40ed3dSBarry Smith ierr = MatSolve_SeqAIJ(A,bb,xx);CHKERRQ(ierr); 6573a40ed3dSBarry Smith PetscFunctionReturn(0); 658930ae53cSBarry Smith } 659930ae53cSBarry Smith 660e1311b90SBarry Smith ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 661e1311b90SBarry Smith ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 662930ae53cSBarry Smith 663aa482453SBarry Smith #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ) 6641c4feecaSSatish Balay fortransolveaij_(&n,x,ai,aj,adiag,aa,b); 6651c4feecaSSatish Balay #else 666930ae53cSBarry Smith /* forward solve the lower triangular */ 667930ae53cSBarry Smith x[0] = b[0]; 668930ae53cSBarry Smith for (i=1; i<n; i++) { 669930ae53cSBarry Smith ai_i = ai[i]; 670930ae53cSBarry Smith v = aa + ai_i; 671930ae53cSBarry Smith vi = aj + ai_i; 672930ae53cSBarry Smith nz = adiag[i] - ai_i; 673930ae53cSBarry Smith sum = b[i]; 674930ae53cSBarry Smith while (nz--) sum -= *v++ * x[*vi++]; 675930ae53cSBarry Smith x[i] = sum; 676930ae53cSBarry Smith } 677930ae53cSBarry Smith 678930ae53cSBarry Smith /* backward solve the upper triangular */ 679930ae53cSBarry Smith for (i=n-1; i>=0; i--){ 680930ae53cSBarry Smith adiag_i = adiag[i]; 681d708749eSBarry Smith v = aa + adiag_i + 1; 682d708749eSBarry Smith vi = aj + adiag_i + 1; 683930ae53cSBarry Smith nz = ai[i+1] - adiag_i - 1; 684930ae53cSBarry Smith sum = x[i]; 685930ae53cSBarry Smith while (nz--) sum -= *v++ * x[*vi++]; 686930ae53cSBarry Smith x[i] = sum*aa[adiag_i]; 687930ae53cSBarry Smith } 6881c4feecaSSatish Balay #endif 689930ae53cSBarry Smith PLogFlops(2*a->nz - a->n); 690cb5b572fSBarry Smith ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 691cb5b572fSBarry Smith ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 6923a40ed3dSBarry Smith PetscFunctionReturn(0); 693930ae53cSBarry Smith } 694930ae53cSBarry Smith 6955615d1e5SSatish Balay #undef __FUNC__ 696b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatSolveAdd_SeqAIJ" 697416022c9SBarry Smith int MatSolveAdd_SeqAIJ(Mat A,Vec bb,Vec yy,Vec xx) 698da3a660dSBarry Smith { 699416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 700416022c9SBarry Smith IS iscol = a->col,isrow = a->row; 701416022c9SBarry Smith int *r,*c,ierr,i, n = a->m,*vi,*ai = a->i,*aj = a->j; 7023b2fbd54SBarry Smith int nz,shift = a->indexshift,*rout,*cout; 703416022c9SBarry Smith Scalar *x,*b,*tmp,*aa = a->a,sum,*v; 704da3a660dSBarry Smith 7053a40ed3dSBarry Smith PetscFunctionBegin; 70678b31e54SBarry Smith if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);} 707da3a660dSBarry Smith 708e1311b90SBarry Smith ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 709e1311b90SBarry Smith ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 710416022c9SBarry Smith tmp = a->solve_work; 711da3a660dSBarry Smith 7123b2fbd54SBarry Smith ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 7133b2fbd54SBarry Smith ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1); 714da3a660dSBarry Smith 715da3a660dSBarry Smith /* forward solve the lower triangular */ 716da3a660dSBarry Smith tmp[0] = b[*r++]; 717da3a660dSBarry Smith for (i=1; i<n; i++) { 718dbb450caSBarry Smith v = aa + ai[i] + shift; 719dbb450caSBarry Smith vi = aj + ai[i] + shift; 720416022c9SBarry Smith nz = a->diag[i] - ai[i]; 721da3a660dSBarry Smith sum = b[*r++]; 722dbb450caSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ + shift]; 723da3a660dSBarry Smith tmp[i] = sum; 724da3a660dSBarry Smith } 725da3a660dSBarry Smith 726da3a660dSBarry Smith /* backward solve the upper triangular */ 727da3a660dSBarry Smith for (i=n-1; i>=0; i--){ 728416022c9SBarry Smith v = aa + a->diag[i] + (!shift); 729416022c9SBarry Smith vi = aj + a->diag[i] + (!shift); 730416022c9SBarry Smith nz = ai[i+1] - a->diag[i] - 1; 731da3a660dSBarry Smith sum = tmp[i]; 732dbb450caSBarry Smith while (nz--) sum -= *v++ * tmp[*vi++ + shift]; 733416022c9SBarry Smith tmp[i] = sum*aa[a->diag[i]+shift]; 734da3a660dSBarry Smith x[*c--] += tmp[i]; 735da3a660dSBarry Smith } 736da3a660dSBarry Smith 7373b2fbd54SBarry Smith ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 7383b2fbd54SBarry Smith ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 739cb5b572fSBarry Smith ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 740cb5b572fSBarry Smith ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 741416022c9SBarry Smith PLogFlops(2*a->nz); 742898183ecSLois Curfman McInnes 7433a40ed3dSBarry Smith PetscFunctionReturn(0); 744da3a660dSBarry Smith } 745da3a660dSBarry Smith /* -------------------------------------------------------------------*/ 7465615d1e5SSatish Balay #undef __FUNC__ 747b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatSolveTranspose_SeqAIJ" 7487c922b88SBarry Smith int MatSolveTranspose_SeqAIJ(Mat A,Vec bb,Vec xx) 749da3a660dSBarry Smith { 750416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 7512235e667SBarry Smith IS iscol = a->col,isrow = a->row; 752416022c9SBarry Smith int *r,*c,ierr,i,n = a->m,*vi,*ai = a->i,*aj = a->j; 753f1af5d2fSBarry Smith int nz,shift = a->indexshift,*rout,*cout,*diag = a->diag; 754f1af5d2fSBarry Smith Scalar *x,*b,*tmp,*aa = a->a,*v,s1; 755da3a660dSBarry Smith 7563a40ed3dSBarry Smith PetscFunctionBegin; 757e1311b90SBarry Smith ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 758e1311b90SBarry Smith ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 759416022c9SBarry Smith tmp = a->solve_work; 760da3a660dSBarry Smith 7612235e667SBarry Smith ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 7622235e667SBarry Smith ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 763da3a660dSBarry Smith 764da3a660dSBarry Smith /* copy the b into temp work space according to permutation */ 7652235e667SBarry Smith for (i=0; i<n; i++) tmp[i] = b[c[i]]; 766da3a660dSBarry Smith 767da3a660dSBarry Smith /* forward solve the U^T */ 768da3a660dSBarry Smith for (i=0; i<n; i++) { 769f1af5d2fSBarry Smith v = aa + diag[i] + shift; 770f1af5d2fSBarry Smith vi = aj + diag[i] + (!shift); 771f1af5d2fSBarry Smith nz = ai[i+1] - diag[i] - 1; 772c38d4ed2SBarry Smith s1 = tmp[i]; 773c38d4ed2SBarry Smith s1 *= *(v++); /* multiply by inverse of diagonal entry */ 774da3a660dSBarry Smith while (nz--) { 775f1af5d2fSBarry Smith tmp[*vi++ + shift] -= (*v++)*s1; 776da3a660dSBarry Smith } 777c38d4ed2SBarry Smith tmp[i] = s1; 778da3a660dSBarry Smith } 779da3a660dSBarry Smith 780da3a660dSBarry Smith /* backward solve the L^T */ 781da3a660dSBarry Smith for (i=n-1; i>=0; i--){ 782f1af5d2fSBarry Smith v = aa + diag[i] - 1 + shift; 783f1af5d2fSBarry Smith vi = aj + diag[i] - 1 + shift; 784f1af5d2fSBarry Smith nz = diag[i] - ai[i]; 785f1af5d2fSBarry Smith s1 = tmp[i]; 786da3a660dSBarry Smith while (nz--) { 787f1af5d2fSBarry Smith tmp[*vi-- + shift] -= (*v--)*s1; 788da3a660dSBarry Smith } 789da3a660dSBarry Smith } 790da3a660dSBarry Smith 791da3a660dSBarry Smith /* copy tmp into x according to permutation */ 792da3a660dSBarry Smith for (i=0; i<n; i++) x[r[i]] = tmp[i]; 793da3a660dSBarry Smith 7942235e667SBarry Smith ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 7952235e667SBarry Smith ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 796cb5b572fSBarry Smith ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 797cb5b572fSBarry Smith ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 798da3a660dSBarry Smith 799416022c9SBarry Smith PLogFlops(2*a->nz-a->n); 8003a40ed3dSBarry Smith PetscFunctionReturn(0); 801da3a660dSBarry Smith } 802da3a660dSBarry Smith 8035615d1e5SSatish Balay #undef __FUNC__ 804b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatSolveTransposeAdd_SeqAIJ" 8057c922b88SBarry Smith int MatSolveTransposeAdd_SeqAIJ(Mat A,Vec bb,Vec zz,Vec xx) 806da3a660dSBarry Smith { 807416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 8082235e667SBarry Smith IS iscol = a->col,isrow = a->row; 809416022c9SBarry Smith int *r,*c,ierr,i,n = a->m,*vi,*ai = a->i,*aj = a->j; 810f1af5d2fSBarry Smith int nz,shift = a->indexshift,*rout,*cout,*diag = a->diag; 811416022c9SBarry Smith Scalar *x,*b,*tmp,*aa = a->a,*v; 8126abc6512SBarry Smith 8133a40ed3dSBarry Smith PetscFunctionBegin; 8146abc6512SBarry Smith if (zz != xx) VecCopy(zz,xx); 8156abc6512SBarry Smith 816e1311b90SBarry Smith ierr = VecGetArray(bb,&b);CHKERRQ(ierr); 817e1311b90SBarry Smith ierr = VecGetArray(xx,&x);CHKERRQ(ierr); 818416022c9SBarry Smith tmp = a->solve_work; 8196abc6512SBarry Smith 8202235e667SBarry Smith ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout; 8212235e667SBarry Smith ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout; 8226abc6512SBarry Smith 8236abc6512SBarry Smith /* copy the b into temp work space according to permutation */ 8242235e667SBarry Smith for (i=0; i<n; i++) tmp[i] = b[c[i]]; 8256abc6512SBarry Smith 8266abc6512SBarry Smith /* forward solve the U^T */ 8276abc6512SBarry Smith for (i=0; i<n; i++) { 828f1af5d2fSBarry Smith v = aa + diag[i] + shift; 829f1af5d2fSBarry Smith vi = aj + diag[i] + (!shift); 830f1af5d2fSBarry Smith nz = ai[i+1] - diag[i] - 1; 8316abc6512SBarry Smith tmp[i] *= *v++; 8326abc6512SBarry Smith while (nz--) { 833dbb450caSBarry Smith tmp[*vi++ + shift] -= (*v++)*tmp[i]; 8346abc6512SBarry Smith } 8356abc6512SBarry Smith } 8366abc6512SBarry Smith 8376abc6512SBarry Smith /* backward solve the L^T */ 8386abc6512SBarry Smith for (i=n-1; i>=0; i--){ 839f1af5d2fSBarry Smith v = aa + diag[i] - 1 + shift; 840f1af5d2fSBarry Smith vi = aj + diag[i] - 1 + shift; 841f1af5d2fSBarry Smith nz = diag[i] - ai[i]; 8426abc6512SBarry Smith while (nz--) { 843dbb450caSBarry Smith tmp[*vi-- + shift] -= (*v--)*tmp[i]; 8446abc6512SBarry Smith } 8456abc6512SBarry Smith } 8466abc6512SBarry Smith 8476abc6512SBarry Smith /* copy tmp into x according to permutation */ 8486abc6512SBarry Smith for (i=0; i<n; i++) x[r[i]] += tmp[i]; 8496abc6512SBarry Smith 8502235e667SBarry Smith ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr); 8512235e667SBarry Smith ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr); 852cb5b572fSBarry Smith ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr); 853cb5b572fSBarry Smith ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr); 8546abc6512SBarry Smith 855416022c9SBarry Smith PLogFlops(2*a->nz); 8563a40ed3dSBarry Smith PetscFunctionReturn(0); 857da3a660dSBarry Smith } 8589e25ed09SBarry Smith /* ----------------------------------------------------------------*/ 859ca44d042SBarry Smith EXTERN int MatMissingDiagonal_SeqAIJ(Mat); 86008480c60SBarry Smith 8615615d1e5SSatish Balay #undef __FUNC__ 862b2863d3aSBarry Smith #define __FUNC__ /*<a name=""></a>*/"MatILUFactorSymbolic_SeqAIJ" 8636cf997b0SBarry Smith int MatILUFactorSymbolic_SeqAIJ(Mat A,IS isrow,IS iscol,MatILUInfo *info,Mat *fact) 8649e25ed09SBarry Smith { 865416022c9SBarry Smith Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b; 8669e25ed09SBarry Smith IS isicol; 867416022c9SBarry Smith int *r,*ic,ierr,prow,n = a->m,*ai = a->i,*aj = a->j; 86856beaf04SBarry Smith int *ainew,*ajnew,jmax,*fill,*xi,nz,*im,*ajfill,*flev; 869335d9088SBarry Smith int *dloc,idx,row,m,fm,nzf,nzi,len, realloc = 0,dcount = 0; 8706cf997b0SBarry Smith int incrlev,nnz,i,shift = a->indexshift,levels,diagonal_fill; 87177c4ece6SBarry Smith PetscTruth col_identity,row_identity; 872329f5518SBarry Smith PetscReal f; 8739e25ed09SBarry Smith 8743a40ed3dSBarry Smith PetscFunctionBegin; 8756cf997b0SBarry Smith if (info) { 8766cf997b0SBarry Smith f = info->fill; 8770cd17293SBarry Smith levels = (int)info->levels; 8780cd17293SBarry Smith diagonal_fill = (int)info->diagonal_fill; 8796cf997b0SBarry Smith } else { 8806cf997b0SBarry Smith f = 1.0; 8816cf997b0SBarry Smith levels = 0; 8826cf997b0SBarry Smith diagonal_fill = 0; 8836cf997b0SBarry Smith } 8844da77479SBarry Smith if (!isrow) { 8854da77479SBarry Smith ierr = ISCreateStride(PETSC_COMM_SELF,A->M,0,1,&isrow);CHKERRQ(ierr); 8864da77479SBarry Smith } 8874da77479SBarry Smith if (!iscol) { 8884da77479SBarry Smith ierr = ISCreateStride(PETSC_COMM_SELF,A->M,0,1,&iscol);CHKERRQ(ierr); 8894da77479SBarry Smith } 8904da77479SBarry Smith 8914c49b128SBarry Smith ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr); 89282bf6240SBarry Smith 89308480c60SBarry Smith /* special case that simply copies fill pattern */ 894be0abb6dSBarry Smith ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr); 895be0abb6dSBarry Smith ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr); 89686bacbd2SBarry Smith if (!levels && row_identity && col_identity) { 8972e8a6d31SBarry Smith ierr = MatDuplicate_SeqAIJ(A,MAT_DO_NOT_COPY_VALUES,fact);CHKERRQ(ierr); 89808480c60SBarry Smith (*fact)->factor = FACTOR_LU; 89908480c60SBarry Smith b = (Mat_SeqAIJ*)(*fact)->data; 90008480c60SBarry Smith if (!b->diag) { 9017c922b88SBarry Smith ierr = MatMarkDiagonal_SeqAIJ(*fact);CHKERRQ(ierr); 90208480c60SBarry Smith } 9037c922b88SBarry Smith ierr = MatMissingDiagonal_SeqAIJ(*fact);CHKERRQ(ierr); 90408480c60SBarry Smith b->row = isrow; 90508480c60SBarry Smith b->col = iscol; 90682bf6240SBarry Smith b->icol = isicol; 907a2e34c3dSBarry Smith if (info) { 9081d1367b7SBarry Smith b->lu_damp = (PetscTruth)((int)info->damp); 909a2e34c3dSBarry Smith b->lu_damping = info->damping; 910a2e34c3dSBarry Smith } else { 911a2e34c3dSBarry Smith b->lu_damp = PETSC_FALSE; 912a2e34c3dSBarry Smith b->lu_damping = 0.0; 913a2e34c3dSBarry Smith } 9140452661fSBarry Smith b->solve_work = (Scalar*)PetscMalloc((b->m+1)*sizeof(Scalar));CHKPTRQ(b->solve_work); 915f830108cSBarry Smith (*fact)->ops->solve = MatSolve_SeqAIJ_NaturalOrdering; 916c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 917c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 9183a40ed3dSBarry Smith PetscFunctionReturn(0); 91908480c60SBarry Smith } 92008480c60SBarry Smith 921898183ecSLois Curfman McInnes ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr); 922898183ecSLois Curfman McInnes ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr); 9239e25ed09SBarry Smith 9249e25ed09SBarry Smith /* get new row pointers */ 9250452661fSBarry Smith ainew = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(ainew); 926dbb450caSBarry Smith ainew[0] = -shift; 9279e25ed09SBarry Smith /* don't know how many column pointers are needed so estimate */ 928dbb450caSBarry Smith jmax = (int)(f*(ai[n]+!shift)); 9290452661fSBarry Smith ajnew = (int*)PetscMalloc((jmax)*sizeof(int));CHKPTRQ(ajnew); 9309e25ed09SBarry Smith /* ajfill is level of fill for each fill entry */ 9310452661fSBarry Smith ajfill = (int*)PetscMalloc((jmax)*sizeof(int));CHKPTRQ(ajfill); 9329e25ed09SBarry Smith /* fill is a linked list of nonzeros in active row */ 9330452661fSBarry Smith fill = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(fill); 93456beaf04SBarry Smith /* im is level for each filled value */ 9350452661fSBarry Smith im = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(im); 93656beaf04SBarry Smith /* dloc is location of diagonal in factor */ 9370452661fSBarry Smith dloc = (int*)PetscMalloc((n+1)*sizeof(int));CHKPTRQ(dloc); 93856beaf04SBarry Smith dloc[0] = 0; 93956beaf04SBarry Smith for (prow=0; prow<n; prow++) { 9406cf997b0SBarry Smith 9416cf997b0SBarry Smith /* copy current row into linked list */ 94256beaf04SBarry Smith nzf = nz = ai[r[prow]+1] - ai[r[prow]]; 943*29bbc08cSBarry Smith if (!nz) SETERRQ(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix"); 944dbb450caSBarry Smith xi = aj + ai[r[prow]] + shift; 9459e25ed09SBarry Smith fill[n] = n; 946435faa5fSBarry Smith fill[prow] = -1; /* marker to indicate if diagonal exists */ 9479e25ed09SBarry Smith while (nz--) { 9489e25ed09SBarry Smith fm = n; 949dbb450caSBarry Smith idx = ic[*xi++ + shift]; 9509e25ed09SBarry Smith do { 9519e25ed09SBarry Smith m = fm; 9529e25ed09SBarry Smith fm = fill[m]; 9539e25ed09SBarry Smith } while (fm < idx); 9549e25ed09SBarry Smith fill[m] = idx; 9559e25ed09SBarry Smith fill[idx] = fm; 95656beaf04SBarry Smith im[idx] = 0; 9579e25ed09SBarry Smith } 9586cf997b0SBarry Smith 9596cf997b0SBarry Smith /* make sure diagonal entry is included */ 960435faa5fSBarry Smith if (diagonal_fill && fill[prow] == -1) { 9616cf997b0SBarry Smith fm = n; 962435faa5fSBarry Smith while (fill[fm] < prow) fm = fill[fm]; 963435faa5fSBarry Smith fill[prow] = fill[fm]; /* insert diagonal into linked list */ 964435faa5fSBarry Smith fill[fm] = prow; 9656cf997b0SBarry Smith im[prow] = 0; 9666cf997b0SBarry Smith nzf++; 967335d9088SBarry Smith dcount++; 9686cf997b0SBarry Smith } 9696cf997b0SBarry Smith 97056beaf04SBarry Smith nzi = 0; 9719e25ed09SBarry Smith row = fill[n]; 97256beaf04SBarry Smith while (row < prow) { 97356beaf04SBarry Smith incrlev = im[row] + 1; 97456beaf04SBarry Smith nz = dloc[row]; 9756cf997b0SBarry Smith xi = ajnew + ainew[row] + shift + nz + 1; 976dbb450caSBarry Smith flev = ajfill + ainew[row] + shift + nz + 1; 977416022c9SBarry Smith nnz = ainew[row+1] - ainew[row] - nz - 1; 97856beaf04SBarry Smith fm = row; 97956beaf04SBarry Smith while (nnz-- > 0) { 980dbb450caSBarry Smith idx = *xi++ + shift; 98156beaf04SBarry Smith if (*flev + incrlev > levels) { 98256beaf04SBarry Smith flev++; 98356beaf04SBarry Smith continue; 98456beaf04SBarry Smith } 98556beaf04SBarry Smith do { 9869e25ed09SBarry Smith m = fm; 9879e25ed09SBarry Smith fm = fill[m]; 98856beaf04SBarry Smith } while (fm < idx); 9899e25ed09SBarry Smith if (fm != idx) { 99056beaf04SBarry Smith im[idx] = *flev + incrlev; 9919e25ed09SBarry Smith fill[m] = idx; 9929e25ed09SBarry Smith fill[idx] = fm; 9939e25ed09SBarry Smith fm = idx; 99456beaf04SBarry Smith nzf++; 995ecf371e4SBarry Smith } else { 99656beaf04SBarry Smith if (im[idx] > *flev + incrlev) im[idx] = *flev+incrlev; 9979e25ed09SBarry Smith } 99856beaf04SBarry Smith flev++; 9999e25ed09SBarry Smith } 10009e25ed09SBarry Smith row = fill[row]; 100156beaf04SBarry Smith nzi++; 10029e25ed09SBarry Smith } 10039e25ed09SBarry Smith /* copy new filled row into permanent storage */ 100456beaf04SBarry Smith ainew[prow+1] = ainew[prow] + nzf; 1005d7e8b826SBarry Smith if (ainew[prow+1] > jmax-shift) { 1006ecf371e4SBarry Smith 1007ecf371e4SBarry Smith /* estimate how much additional space we will need */ 1008ecf371e4SBarry Smith /* use the strategy suggested by David Hysom <hysom@perch-t.icase.edu> */ 1009ecf371e4SBarry Smith /* just double the memory each time */ 1010ecf371e4SBarry Smith /* maxadd = (int)((f*(ai[n]+!shift)*(n-prow+5))/n); */ 1011ecf371e4SBarry Smith int maxadd = jmax; 101256beaf04SBarry Smith if (maxadd < nzf) maxadd = (n-prow)*(nzf+1); 1013bbb6d6a8SBarry Smith jmax += maxadd; 1014ecf371e4SBarry Smith 1015ecf371e4SBarry Smith /* allocate a longer ajnew and ajfill */ 10160452661fSBarry Smith xi = (int*)PetscMalloc(jmax*sizeof(int));CHKPTRQ(xi); 1017549d3d68SSatish Balay ierr = PetscMemcpy(xi,ajnew,(ainew[prow]+shift)*sizeof(int));CHKERRQ(ierr); 1018606d414cSSatish Balay ierr = PetscFree(ajnew);CHKERRQ(ierr); 101956beaf04SBarry Smith ajnew = xi; 10200452661fSBarry Smith xi = (int*)PetscMalloc(jmax*sizeof(int));CHKPTRQ(xi); 1021549d3d68SSatish Balay ierr = PetscMemcpy(xi,ajfill,(ainew[prow]+shift)*sizeof(int));CHKERRQ(ierr); 1022606d414cSSatish Balay ierr = PetscFree(ajfill);CHKERRQ(ierr); 102356beaf04SBarry Smith ajfill = xi; 1024ecf371e4SBarry Smith realloc++; /* count how many times we realloc */ 10259e25ed09SBarry Smith } 1026dbb450caSBarry Smith xi = ajnew + ainew[prow] + shift; 1027dbb450caSBarry Smith flev = ajfill + ainew[prow] + shift; 102856beaf04SBarry Smith dloc[prow] = nzi; 10299e25ed09SBarry Smith fm = fill[n]; 103056beaf04SBarry Smith while (nzf--) { 1031dbb450caSBarry Smith *xi++ = fm - shift; 103256beaf04SBarry Smith *flev++ = im[fm]; 10339e25ed09SBarry Smith fm = fill[fm]; 10349e25ed09SBarry Smith } 10356cf997b0SBarry Smith /* make sure row has diagonal entry */ 10366cf997b0SBarry Smith if (ajnew[ainew[prow]+shift+dloc[prow]]+shift != prow) { 1037*29bbc08cSBarry Smith SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %d has missing diagonal in factored matrix\n\ 10386cf997b0SBarry Smith try running with -pc_ilu_nonzeros_along_diagonal or -pc_ilu_diagonal_fill",prow); 10396cf997b0SBarry Smith } 10409e25ed09SBarry Smith } 1041606d414cSSatish Balay ierr = PetscFree(ajfill);CHKERRQ(ierr); 1042898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr); 1043898183ecSLois Curfman McInnes ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr); 1044606d414cSSatish Balay ierr = PetscFree(fill);CHKERRQ(ierr); 1045606d414cSSatish Balay ierr = PetscFree(im);CHKERRQ(ierr); 10469e25ed09SBarry Smith 1047f64065bbSBarry Smith { 1048329f5518SBarry Smith PetscReal af = ((PetscReal)ainew[n])/((PetscReal)ai[n]); 1049c38d4ed2SBarry Smith PLogInfo(A,"MatILUFactorSymbolic_SeqAIJ:Reallocs %d Fill ratio:given %g needed %g\n",realloc,f,af); 1050981c4779SBarry Smith PLogInfo(A,"MatILUFactorSymbolic_SeqAIJ:Run with -pc_ilu_fill %g or use \n",af); 1051981c4779SBarry Smith PLogInfo(A,"MatILUFactorSymbolic_SeqAIJ:PCILUSetFill(pc,%g);\n",af); 1052981c4779SBarry Smith PLogInfo(A,"MatILUFactorSymbolic_SeqAIJ:for best performance.\n"); 1053335d9088SBarry Smith if (diagonal_fill) { 1054335d9088SBarry Smith PLogInfo(A,"MatILUFactorSymbolic_SeqAIJ:Detected and replace %d missing diagonals",dcount); 1055335d9088SBarry Smith } 1056f64065bbSBarry Smith } 1057bbb6d6a8SBarry Smith 10589e25ed09SBarry Smith /* put together the new matrix */ 1059b4fd4287SBarry Smith ierr = MatCreateSeqAIJ(A->comm,n,n,0,PETSC_NULL,fact);CHKERRQ(ierr); 1060fa6007c9SSatish Balay PLogObjectParent(*fact,isicol); 1061416022c9SBarry Smith b = (Mat_SeqAIJ*)(*fact)->data; 1062606d414cSSatish Balay ierr = PetscFree(b->imax);CHKERRQ(ierr); 10637c922b88SBarry Smith b->singlemalloc = PETSC_FALSE; 1064dbb450caSBarry Smith len = (ainew[n] + shift)*sizeof(Scalar); 10659e25ed09SBarry Smith /* the next line frees the default space generated by the Create() */ 1066606d414cSSatish Balay ierr = PetscFree(b->a);CHKERRQ(ierr); 1067606d414cSSatish Balay ierr = PetscFree(b->ilen);CHKERRQ(ierr); 10686b96ef82SSatish Balay b->a = (Scalar*)PetscMalloc(len+1);CHKPTRQ(b->a); 1069416022c9SBarry Smith b->j = ajnew; 1070416022c9SBarry Smith b->i = ainew; 107156beaf04SBarry Smith for (i=0; i<n; i++) dloc[i] += ainew[i]; 1072416022c9SBarry Smith b->diag = dloc; 1073416022c9SBarry Smith b->ilen = 0; 1074416022c9SBarry Smith b->imax = 0; 1075416022c9SBarry Smith b->row = isrow; 1076416022c9SBarry Smith b->col = iscol; 1077c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr); 1078c38d4ed2SBarry Smith ierr = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr); 107982bf6240SBarry Smith b->icol = isicol; 108082bf6240SBarry Smith b->solve_work = (Scalar*)PetscMalloc((n+1)*sizeof(Scalar));CHKPTRQ(b->solve_work); 1081416022c9SBarry Smith /* In b structure: Free imax, ilen, old a, old j. 108256beaf04SBarry Smith Allocate dloc, solve_work, new a, new j */ 1083dbb450caSBarry Smith PLogObjectMemory(*fact,(ainew[n]+shift-n) * (sizeof(int)+sizeof(Scalar))); 1084416022c9SBarry Smith b->maxnz = b->nz = ainew[n] + shift; 1085a2e34c3dSBarry Smith if (info) { 10861d1367b7SBarry Smith b->lu_damp = (PetscTruth)((int)info->damp); 1087a2e34c3dSBarry Smith b->lu_damping = info->damping; 1088a2e34c3dSBarry Smith } else { 1089a2e34c3dSBarry Smith b->lu_damp = PETSC_FALSE; 1090a2e34c3dSBarry Smith b->lu_damping = 0.0; 1091a2e34c3dSBarry Smith } 10929e25ed09SBarry Smith (*fact)->factor = FACTOR_LU; 1093ae068f58SLois Curfman McInnes 1094ae068f58SLois Curfman McInnes (*fact)->info.factor_mallocs = realloc; 1095ae068f58SLois Curfman McInnes (*fact)->info.fill_ratio_given = f; 1096329f5518SBarry Smith (*fact)->info.fill_ratio_needed = ((PetscReal)ainew[n])/((PetscReal)ai[prow]); 1097329f5518SBarry Smith (*fact)->factor = FACTOR_LU; 1098ae068f58SLois Curfman McInnes 10993a40ed3dSBarry Smith PetscFunctionReturn(0); 11009e25ed09SBarry Smith } 1101d5d45c9bSBarry Smith 1102d5d45c9bSBarry Smith 1103d5d45c9bSBarry Smith 1104d5d45c9bSBarry Smith 1105