xref: /petsc/src/mat/impls/aij/seq/aijfact.c (revision 0b3d78a7f9fddbfd6dfd47e4ff17062a81c9ada7)
1 #define PETSCMAT_DLL
2 
3 #include "../src/mat/impls/aij/seq/aij.h"
4 #include "../src/inline/dot.h"
5 #include "../src/inline/spops.h"
6 #include "petscbt.h"
7 #include "../src/mat/utils/freespace.h"
8 
9 EXTERN_C_BEGIN
10 #undef __FUNCT__
11 #define __FUNCT__ "MatOrdering_Flow_SeqAIJ"
12 /*
13       Computes an ordering to get most of the large numerical values in the lower triangular part of the matrix
14 */
15 PetscErrorCode MatOrdering_Flow_SeqAIJ(Mat mat,const MatOrderingType type,IS *irow,IS *icol)
16 {
17   Mat_SeqAIJ        *a = (Mat_SeqAIJ*)mat->data;
18   PetscErrorCode    ierr;
19   PetscInt          i,j,jj,k, kk,n = mat->rmap->n, current, newcurrent,*order;
20   const PetscInt    *ai = a->i, *aj = a->j;
21   const PetscScalar *aa = a->a;
22   PetscTruth        *done;
23   PetscReal         best,past,future;
24 
25   PetscFunctionBegin;
26   /* pick initial row */
27   best = -1;
28   for (i=0; i<n; i++) {
29     future = 0;
30     for (j=ai[i]; j<ai[i+1]; j++) {
31       if (aj[j] != i) future  += PetscAbsScalar(aa[j]); else past = PetscAbsScalar(aa[j]);
32     }
33     if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */
34     if (past/future > best) {
35       best = past/future;
36       current = i;
37     }
38   }
39 
40   ierr = PetscMalloc(n*sizeof(PetscTruth),&done);CHKERRQ(ierr);
41   ierr = PetscMalloc(n*sizeof(PetscInt),&order);CHKERRQ(ierr);
42   ierr = PetscMemzero(done,n*sizeof(PetscTruth));CHKERRQ(ierr);
43   order[0] = current;
44   for (i=0; i<n-1; i++) {
45     done[current] = PETSC_TRUE;
46     best          = -1;
47     /* loop over all neighbors of current pivot */
48     for (j=ai[current]; j<ai[current+1]; j++) {
49       jj = aj[j];
50       if (done[jj]) continue;
51       /* loop over columns of potential next row computing weights for below and above diagonal */
52       past = future = 0.0;
53       for (k=ai[jj]; k<ai[jj+1]; k++) {
54         kk = aj[k];
55         if (done[kk]) past += PetscAbsScalar(aa[k]);
56         else if (kk != jj) future  += PetscAbsScalar(aa[k]);
57       }
58       if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */
59       if (past/future > best) {
60         best = past/future;
61         newcurrent = jj;
62       }
63     }
64     if (best == -1) { /* no neighbors to select from so select best of all that remain */
65       best = -1;
66       for (k=0; k<n; k++) {
67         if (done[k]) continue;
68         future = 0;
69         past   = 0;
70         for (j=ai[k]; j<ai[k+1]; j++) {
71           kk = aj[j];
72           if (done[kk]) past += PetscAbsScalar(aa[j]);
73           else if (kk != k) future  += PetscAbsScalar(aa[j]);
74         }
75         if (!future) future = 1.e-10; /* if there is zero in the upper diagonal part want to rank this row high */
76         if (past/future > best) {
77           best = past/future;
78           newcurrent = k;
79         }
80       }
81     }
82     if (current == newcurrent) SETERRQ(PETSC_ERR_PLIB,"newcurrent cannot be current");
83     current = newcurrent;
84     order[i+1] = current;
85   }
86   ierr = ISCreateGeneral(PETSC_COMM_SELF,n,order,irow);CHKERRQ(ierr);
87   *icol = *irow;
88   ierr = PetscObjectReference((PetscObject)*irow);CHKERRQ(ierr);
89   ierr = PetscFree(done);CHKERRQ(ierr);
90   ierr = PetscFree(order);CHKERRQ(ierr);
91   PetscFunctionReturn(0);
92 }
93 EXTERN_C_END
94 
95 EXTERN_C_BEGIN
96 #undef __FUNCT__
97 #define __FUNCT__ "MatGetFactorAvailable_seqaij_petsc"
98 PetscErrorCode MatGetFactorAvailable_seqaij_petsc(Mat A,MatFactorType ftype,PetscTruth *flg)
99 {
100   PetscFunctionBegin;
101   *flg = PETSC_TRUE;
102   PetscFunctionReturn(0);
103 }
104 EXTERN_C_END
105 
106 EXTERN_C_BEGIN
107 #undef __FUNCT__
108 #define __FUNCT__ "MatGetFactor_seqaij_petsc"
109 PetscErrorCode MatGetFactor_seqaij_petsc(Mat A,MatFactorType ftype,Mat *B)
110 {
111   PetscInt           n = A->rmap->n;
112   PetscErrorCode     ierr;
113 
114   PetscFunctionBegin;
115   ierr = MatCreate(((PetscObject)A)->comm,B);CHKERRQ(ierr);
116   ierr = MatSetSizes(*B,n,n,n,n);CHKERRQ(ierr);
117   if (ftype == MAT_FACTOR_LU || ftype == MAT_FACTOR_ILU || ftype == MAT_FACTOR_ILUDT){
118     ierr = MatSetType(*B,MATSEQAIJ);CHKERRQ(ierr);
119     (*B)->ops->ilufactorsymbolic = MatILUFactorSymbolic_SeqAIJ;
120     (*B)->ops->lufactorsymbolic  = MatLUFactorSymbolic_SeqAIJ;
121     (*B)->ops->iludtfactor       = MatILUDTFactor_SeqAIJ;
122   } else if (ftype == MAT_FACTOR_CHOLESKY || ftype == MAT_FACTOR_ICC) {
123     ierr = MatSetType(*B,MATSEQSBAIJ);CHKERRQ(ierr);
124     ierr = MatSeqSBAIJSetPreallocation(*B,1,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr);
125     (*B)->ops->iccfactorsymbolic      = MatICCFactorSymbolic_SeqAIJ;
126     (*B)->ops->choleskyfactorsymbolic = MatCholeskyFactorSymbolic_SeqAIJ;
127   } else SETERRQ(PETSC_ERR_SUP,"Factor type not supported");
128   (*B)->factor = ftype;
129   PetscFunctionReturn(0);
130 }
131 EXTERN_C_END
132 
133 #undef __FUNCT__
134 #define __FUNCT__ "MatLUFactorSymbolic_SeqAIJ"
135 PetscErrorCode MatLUFactorSymbolic_SeqAIJ(Mat B,Mat A,IS isrow,IS iscol,const MatFactorInfo *info)
136 {
137   Mat_SeqAIJ         *a = (Mat_SeqAIJ*)A->data,*b;
138   IS                 isicol;
139   PetscErrorCode     ierr;
140   const PetscInt     *r,*ic;
141   PetscInt           i,n=A->rmap->n,*ai=a->i,*aj=a->j;
142   PetscInt           *bi,*bj,*ajtmp;
143   PetscInt           *bdiag,row,nnz,nzi,reallocs=0,nzbd,*im;
144   PetscReal          f;
145   PetscInt           nlnk,*lnk,k,**bi_ptr;
146   PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
147   PetscBT            lnkbt;
148 
149   PetscFunctionBegin;
150   if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_ERR_ARG_WRONG,"matrix must be square");
151   ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr);
152   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
153   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
154 
155   /* get new row pointers */
156   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr);
157   bi[0] = 0;
158 
159   /* bdiag is location of diagonal in factor */
160   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr);
161   bdiag[0] = 0;
162 
163   /* linked list for storing column indices of the active row */
164   nlnk = n + 1;
165   ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr);
166 
167   ierr = PetscMalloc2(n+1,PetscInt**,&bi_ptr,n+1,PetscInt,&im);CHKERRQ(ierr);
168 
169   /* initial FreeSpace size is f*(ai[n]+1) */
170   f = info->fill;
171   ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr);
172   current_space = free_space;
173 
174   for (i=0; i<n; i++) {
175     /* copy previous fill into linked list */
176     nzi = 0;
177     nnz = ai[r[i]+1] - ai[r[i]];
178     if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i);
179     ajtmp = aj + ai[r[i]];
180     ierr = PetscLLAddPerm(nnz,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr);
181     nzi += nlnk;
182 
183     /* add pivot rows into linked list */
184     row = lnk[n];
185     while (row < i) {
186       nzbd    = bdiag[row] - bi[row] + 1; /* num of entries in the row with column index <= row */
187       ajtmp   = bi_ptr[row] + nzbd; /* points to the entry next to the diagonal */
188       ierr = PetscLLAddSortedLU(ajtmp,row,nlnk,lnk,lnkbt,i,nzbd,im);CHKERRQ(ierr);
189       nzi += nlnk;
190       row  = lnk[row];
191     }
192     bi[i+1] = bi[i] + nzi;
193     im[i]   = nzi;
194 
195     /* mark bdiag */
196     nzbd = 0;
197     nnz  = nzi;
198     k    = lnk[n];
199     while (nnz-- && k < i){
200       nzbd++;
201       k = lnk[k];
202     }
203     bdiag[i] = bi[i] + nzbd;
204 
205     /* if free space is not available, make more free space */
206     if (current_space->local_remaining<nzi) {
207       nnz = (n - i)*nzi; /* estimated and max additional space needed */
208       ierr = PetscFreeSpaceGet(nnz,&current_space);CHKERRQ(ierr);
209       reallocs++;
210     }
211 
212     /* copy data into free space, then initialize lnk */
213     ierr = PetscLLClean(n,n,nzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr);
214     bi_ptr[i] = current_space->array;
215     current_space->array           += nzi;
216     current_space->local_used      += nzi;
217     current_space->local_remaining -= nzi;
218   }
219 #if defined(PETSC_USE_INFO)
220   if (ai[n] != 0) {
221     PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]);
222     ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr);
223     ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr);
224     ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G);\n",af);CHKERRQ(ierr);
225     ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr);
226   } else {
227     ierr = PetscInfo(A,"Empty matrix\n");CHKERRQ(ierr);
228   }
229 #endif
230 
231   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
232   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
233 
234   /* destroy list of free space and other temporary array(s) */
235   ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr);
236   ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr);
237   ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
238   ierr = PetscFree2(bi_ptr,im);CHKERRQ(ierr);
239 
240   /* put together the new matrix */
241   ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr);
242   ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr);
243   b    = (Mat_SeqAIJ*)(B)->data;
244   b->free_a       = PETSC_TRUE;
245   b->free_ij      = PETSC_TRUE;
246   b->singlemalloc = PETSC_FALSE;
247   ierr          = PetscMalloc((bi[n]+1)*sizeof(PetscScalar),&b->a);CHKERRQ(ierr);
248   b->j          = bj;
249   b->i          = bi;
250   b->diag       = bdiag;
251   b->ilen       = 0;
252   b->imax       = 0;
253   b->row        = isrow;
254   b->col        = iscol;
255   ierr          = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
256   ierr          = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
257   b->icol       = isicol;
258   ierr          = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
259 
260   /* In b structure:  Free imax, ilen, old a, old j.  Allocate solve_work, new a, new j */
261   ierr = PetscLogObjectMemory(B,(bi[n]-n)*(sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr);
262   b->maxnz = b->nz = bi[n] ;
263 
264   (B)->factor                = MAT_FACTOR_LU;
265   (B)->info.factor_mallocs   = reallocs;
266   (B)->info.fill_ratio_given = f;
267 
268   if (ai[n]) {
269     (B)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]);
270   } else {
271     (B)->info.fill_ratio_needed = 0.0;
272   }
273   (B)->ops->lufactornumeric  = MatLUFactorNumeric_SeqAIJ;
274   (B)->ops->solve            = MatSolve_SeqAIJ;
275   (B)->ops->solvetranspose   = MatSolveTranspose_SeqAIJ;
276   /* switch to inodes if appropriate */
277   ierr = MatLUFactorSymbolic_Inode(B,A,isrow,iscol,info);CHKERRQ(ierr);
278   PetscFunctionReturn(0);
279 }
280 
281 /*
282     Trouble in factorization, should we dump the original matrix?
283 */
284 #undef __FUNCT__
285 #define __FUNCT__ "MatFactorDumpMatrix"
286 PetscErrorCode MatFactorDumpMatrix(Mat A)
287 {
288   PetscErrorCode ierr;
289   PetscTruth     flg = PETSC_FALSE;
290 
291   PetscFunctionBegin;
292   ierr = PetscOptionsGetTruth(PETSC_NULL,"-mat_factor_dump_on_error",&flg,PETSC_NULL);CHKERRQ(ierr);
293   if (flg) {
294     PetscViewer viewer;
295     char        filename[PETSC_MAX_PATH_LEN];
296 
297     ierr = PetscSNPrintf(filename,PETSC_MAX_PATH_LEN,"matrix_factor_error.%d",PetscGlobalRank);CHKERRQ(ierr);
298     ierr = PetscViewerBinaryOpen(((PetscObject)A)->comm,filename,FILE_MODE_WRITE,&viewer);CHKERRQ(ierr);
299     ierr = MatView(A,viewer);CHKERRQ(ierr);
300     ierr = PetscViewerDestroy(viewer);CHKERRQ(ierr);
301   }
302   PetscFunctionReturn(0);
303 }
304 
305 extern PetscErrorCode MatSolve_Inode(Mat,Vec,Vec);
306 
307 /* ----------------------------------------------------------- */
308 #undef __FUNCT__
309 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ"
310 PetscErrorCode MatLUFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info)
311 {
312   Mat            C=B;
313   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data;
314   IS             isrow = b->row,isicol = b->icol;
315   PetscErrorCode ierr;
316   const PetscInt  *r,*ic,*ics;
317   PetscInt       i,j,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
318   PetscInt       *ajtmp,*bjtmp,nz,row,*diag_offset = b->diag,diag,*pj;
319   MatScalar      *rtmp,*pc,multiplier,*v,*pv,d,*aa=a->a;
320   PetscReal      rs=0.0;
321   LUShift_Ctx    sctx;
322   PetscInt       newshift,*ddiag;
323 
324   PetscFunctionBegin;
325   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
326   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
327   ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
328   ics  = ic;
329 
330   sctx.shift_top      = 0;
331   sctx.nshift_max     = 0;
332   sctx.shift_lo       = 0;
333   sctx.shift_hi       = 0;
334   sctx.shift_fraction = 0;
335 
336   /* if both shift schemes are chosen by user, only use info->shiftpd */
337   if (info->shiftpd) { /* set sctx.shift_top=max{rs} */
338     ddiag          = a->diag;
339     sctx.shift_top = info->zeropivot;
340     for (i=0; i<n; i++) {
341       /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */
342       d  = (aa)[ddiag[i]];
343       rs = -PetscAbsScalar(d) - PetscRealPart(d);
344       v  = aa+ai[i];
345       nz = ai[i+1] - ai[i];
346       for (j=0; j<nz; j++)
347 	rs += PetscAbsScalar(v[j]);
348       if (rs>sctx.shift_top) sctx.shift_top = rs;
349     }
350     sctx.shift_top   *= 1.1;
351     sctx.nshift_max   = 5;
352     sctx.shift_lo     = 0.;
353     sctx.shift_hi     = 1.;
354   }
355 
356   sctx.shift_amount = 0.0;
357   sctx.nshift       = 0;
358   do {
359     sctx.lushift = PETSC_FALSE;
360     for (i=0; i<n; i++){
361       nz    = bi[i+1] - bi[i];
362       bjtmp = bj + bi[i];
363       for  (j=0; j<nz; j++) rtmp[bjtmp[j]] = 0.0;
364 
365       /* load in initial (unfactored row) */
366       nz    = ai[r[i]+1] - ai[r[i]];
367       ajtmp = aj + ai[r[i]];
368       v     = aa + ai[r[i]];
369       for (j=0; j<nz; j++) {
370         rtmp[ics[ajtmp[j]]] = v[j];
371       }
372       rtmp[ics[r[i]]] += sctx.shift_amount; /* shift the diagonal of the matrix */
373 
374       row = *bjtmp++;
375       while  (row < i) {
376         pc = rtmp + row;
377         if (*pc != 0.0) {
378           pv         = b->a + diag_offset[row];
379           pj         = b->j + diag_offset[row] + 1;
380           multiplier = *pc / *pv++;
381           *pc        = multiplier;
382           nz         = bi[row+1] - diag_offset[row] - 1;
383           for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j];
384           ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr);
385         }
386         row = *bjtmp++;
387       }
388       /* finished row so stick it into b->a */
389       pv   = b->a + bi[i] ;
390       pj   = b->j + bi[i] ;
391       nz   = bi[i+1] - bi[i];
392       diag = diag_offset[i] - bi[i];
393       rs   = -PetscAbsScalar(pv[diag]);
394       for (j=0; j<nz; j++) {
395         pv[j] = rtmp[pj[j]];
396         rs   += PetscAbsScalar(pv[j]);
397       }
398 
399       /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */
400       sctx.rs  = rs;
401       sctx.pv  = pv[diag];
402       ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr);
403       if (newshift == 1) break;
404     }
405 
406     if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) {
407       /*
408        * if no shift in this attempt & shifting & started shifting & can refine,
409        * then try lower shift
410        */
411       sctx.shift_hi       = sctx.shift_fraction;
412       sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.;
413       sctx.shift_amount   = sctx.shift_fraction * sctx.shift_top;
414       sctx.lushift        = PETSC_TRUE;
415       sctx.nshift++;
416     }
417   } while (sctx.lushift);
418 
419   /* invert diagonal entries for simplier triangular solves */
420   for (i=0; i<n; i++) {
421     b->a[diag_offset[i]] = 1.0/b->a[diag_offset[i]];
422   }
423   ierr = PetscFree(rtmp);CHKERRQ(ierr);
424   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
425   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
426   if (b->inode.use) {
427     C->ops->solve   = MatSolve_Inode;
428   } else {
429     PetscTruth row_identity, col_identity;
430     ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr);
431     ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr);
432     if (row_identity && col_identity) {
433       C->ops->solve   = MatSolve_SeqAIJ_NaturalOrdering;
434     } else {
435       C->ops->solve   = MatSolve_SeqAIJ;
436     }
437   }
438   C->ops->solveadd           = MatSolveAdd_SeqAIJ;
439   C->ops->solvetranspose     = MatSolveTranspose_SeqAIJ;
440   C->ops->solvetransposeadd  = MatSolveTransposeAdd_SeqAIJ;
441   C->ops->matsolve           = MatMatSolve_SeqAIJ;
442   C->assembled    = PETSC_TRUE;
443   C->preallocated = PETSC_TRUE;
444   ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr);
445   if (sctx.nshift){
446      if (info->shiftpd) {
447       ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr);
448     } else if (info->shiftnz) {
449       ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr);
450     }
451   }
452   PetscFunctionReturn(0);
453 }
454 
455 /*
456    This routine implements inplace ILU(0) with row or/and column permutations.
457    Input:
458      A - original matrix
459    Output;
460      A - a->i (rowptr) is same as original rowptr, but factored i-the row is stored in rowperm[i]
461          a->j (col index) is permuted by the inverse of colperm, then sorted
462          a->a reordered accordingly with a->j
463          a->diag (ptr to diagonal elements) is updated.
464 */
465 #undef __FUNCT__
466 #define __FUNCT__ "MatLUFactorNumeric_SeqAIJ_InplaceWithPerm"
467 PetscErrorCode MatLUFactorNumeric_SeqAIJ_InplaceWithPerm(Mat B,Mat A,const MatFactorInfo *info)
468 {
469   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data;
470   IS             isrow = a->row,isicol = a->icol;
471   PetscErrorCode ierr;
472   const PetscInt *r,*ic,*ics;
473   PetscInt       i,j,n=A->rmap->n,*ai=a->i,*aj=a->j;
474   PetscInt       *ajtmp,nz,row;
475   PetscInt       *diag = a->diag,nbdiag,*pj;
476   PetscScalar    *rtmp,*pc,multiplier,d;
477   MatScalar      *v,*pv;
478   PetscReal      rs;
479   LUShift_Ctx    sctx;
480   PetscInt       newshift;
481 
482   PetscFunctionBegin;
483   if (A != B) SETERRQ(PETSC_ERR_ARG_INCOMP,"input and output matrix must have same address");
484   ierr  = ISGetIndices(isrow,&r);CHKERRQ(ierr);
485   ierr  = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
486   ierr  = PetscMalloc((n+1)*sizeof(PetscScalar),&rtmp);CHKERRQ(ierr);
487   ierr  = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr);
488   ics = ic;
489 
490   sctx.shift_top      = 0;
491   sctx.nshift_max     = 0;
492   sctx.shift_lo       = 0;
493   sctx.shift_hi       = 0;
494   sctx.shift_fraction = 0;
495 
496   /* if both shift schemes are chosen by user, only use info->shiftpd */
497   if (info->shiftpd) { /* set sctx.shift_top=max{rs} */
498     sctx.shift_top = 0;
499     for (i=0; i<n; i++) {
500       /* calculate sum(|aij|)-RealPart(aii), amt of shift needed for this row */
501       d  = (a->a)[diag[i]];
502       rs = -PetscAbsScalar(d) - PetscRealPart(d);
503       v  = a->a+ai[i];
504       nz = ai[i+1] - ai[i];
505       for (j=0; j<nz; j++)
506 	rs += PetscAbsScalar(v[j]);
507       if (rs>sctx.shift_top) sctx.shift_top = rs;
508     }
509     if (sctx.shift_top < info->zeropivot) sctx.shift_top = info->zeropivot;
510     sctx.shift_top    *= 1.1;
511     sctx.nshift_max   = 5;
512     sctx.shift_lo     = 0.;
513     sctx.shift_hi     = 1.;
514   }
515 
516   sctx.shift_amount = 0;
517   sctx.nshift       = 0;
518   do {
519     sctx.lushift = PETSC_FALSE;
520     for (i=0; i<n; i++){
521       /* load in initial unfactored row */
522       nz    = ai[r[i]+1] - ai[r[i]];
523       ajtmp = aj + ai[r[i]];
524       v     = a->a + ai[r[i]];
525       /* sort permuted ajtmp and values v accordingly */
526       for (j=0; j<nz; j++) ajtmp[j] = ics[ajtmp[j]];
527       ierr = PetscSortIntWithScalarArray(nz,ajtmp,v);CHKERRQ(ierr);
528 
529       diag[r[i]] = ai[r[i]];
530       for (j=0; j<nz; j++) {
531         rtmp[ajtmp[j]] = v[j];
532         if (ajtmp[j] < i) diag[r[i]]++; /* update a->diag */
533       }
534       rtmp[r[i]] += sctx.shift_amount; /* shift the diagonal of the matrix */
535 
536       row = *ajtmp++;
537       while  (row < i) {
538         pc = rtmp + row;
539         if (*pc != 0.0) {
540           pv         = a->a + diag[r[row]];
541           pj         = aj + diag[r[row]] + 1;
542 
543           multiplier = *pc / *pv++;
544           *pc        = multiplier;
545           nz         = ai[r[row]+1] - diag[r[row]] - 1;
546           for (j=0; j<nz; j++) rtmp[pj[j]] -= multiplier * pv[j];
547           ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr);
548         }
549         row = *ajtmp++;
550       }
551       /* finished row so overwrite it onto a->a */
552       pv   = a->a + ai[r[i]] ;
553       pj   = aj + ai[r[i]] ;
554       nz   = ai[r[i]+1] - ai[r[i]];
555       nbdiag = diag[r[i]] - ai[r[i]]; /* num of entries before the diagonal */
556 
557       rs   = 0.0;
558       for (j=0; j<nz; j++) {
559         pv[j] = rtmp[pj[j]];
560         if (j != nbdiag) rs += PetscAbsScalar(pv[j]);
561       }
562 
563       /* 9/13/02 Victor Eijkhout suggested scaling zeropivot by rs for matrices with funny scalings */
564       sctx.rs  = rs;
565       sctx.pv  = pv[nbdiag];
566       ierr = MatLUCheckShift_inline(info,sctx,i,newshift);CHKERRQ(ierr);
567       if (newshift == 1) break;
568     }
569 
570     if (info->shiftpd && !sctx.lushift && sctx.shift_fraction>0 && sctx.nshift<sctx.nshift_max) {
571       /*
572        * if no shift in this attempt & shifting & started shifting & can refine,
573        * then try lower shift
574        */
575       sctx.shift_hi        = sctx.shift_fraction;
576       sctx.shift_fraction = (sctx.shift_hi+sctx.shift_lo)/2.;
577       sctx.shift_amount    = sctx.shift_fraction * sctx.shift_top;
578       sctx.lushift         = PETSC_TRUE;
579       sctx.nshift++;
580     }
581   } while (sctx.lushift);
582 
583   /* invert diagonal entries for simplier triangular solves */
584   for (i=0; i<n; i++) {
585     a->a[diag[r[i]]] = 1.0/a->a[diag[r[i]]];
586   }
587 
588   ierr = PetscFree(rtmp);CHKERRQ(ierr);
589   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
590   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
591   A->ops->solve             = MatSolve_SeqAIJ_InplaceWithPerm;
592   A->ops->solveadd          = MatSolveAdd_SeqAIJ;
593   A->ops->solvetranspose    = MatSolveTranspose_SeqAIJ;
594   A->ops->solvetransposeadd = MatSolveTransposeAdd_SeqAIJ;
595   A->assembled = PETSC_TRUE;
596   A->preallocated = PETSC_TRUE;
597   ierr = PetscLogFlops(A->cmap->n);CHKERRQ(ierr);
598   if (sctx.nshift){
599     if (info->shiftpd) {
600       ierr = PetscInfo4(A,"number of shift_pd tries %D, shift_amount %G, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,sctx.shift_fraction,sctx.shift_top);CHKERRQ(ierr);
601     } else if (info->shiftnz) {
602       ierr = PetscInfo2(A,"number of shift_nz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr);
603     }
604   }
605   PetscFunctionReturn(0);
606 }
607 
608 /* ----------------------------------------------------------- */
609 #undef __FUNCT__
610 #define __FUNCT__ "MatLUFactor_SeqAIJ"
611 PetscErrorCode MatLUFactor_SeqAIJ(Mat A,IS row,IS col,const MatFactorInfo *info)
612 {
613   PetscErrorCode ierr;
614   Mat            C;
615 
616   PetscFunctionBegin;
617   ierr = MatGetFactor(A,MAT_SOLVER_PETSC,MAT_FACTOR_LU,&C);CHKERRQ(ierr);
618   ierr = MatLUFactorSymbolic(C,A,row,col,info);CHKERRQ(ierr);
619   ierr = MatLUFactorNumeric(C,A,info);CHKERRQ(ierr);
620   A->ops->solve            = C->ops->solve;
621   A->ops->solvetranspose   = C->ops->solvetranspose;
622   ierr = MatHeaderCopy(A,C);CHKERRQ(ierr);
623   ierr = PetscLogObjectParent(A,((Mat_SeqAIJ*)(A->data))->icol);CHKERRQ(ierr);
624   PetscFunctionReturn(0);
625 }
626 /* ----------------------------------------------------------- */
627 
628 #undef __FUNCT__
629 #define __FUNCT__ "MatSolve_SeqAIJ"
630 PetscErrorCode MatSolve_SeqAIJ(Mat A,Vec bb,Vec xx)
631 {
632   Mat_SeqAIJ        *a = (Mat_SeqAIJ*)A->data;
633   IS                iscol = a->col,isrow = a->row;
634   PetscErrorCode    ierr;
635   PetscInt          i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
636   PetscInt          nz;
637   const PetscInt    *rout,*cout,*r,*c;
638   PetscScalar       *x,*tmp,*tmps,sum;
639   const PetscScalar *b;
640   const MatScalar   *aa = a->a,*v;
641 
642   PetscFunctionBegin;
643   if (!n) PetscFunctionReturn(0);
644 
645   ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
646   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
647   tmp  = a->solve_work;
648 
649   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
650   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1);
651 
652   /* forward solve the lower triangular */
653   tmp[0] = b[*r++];
654   tmps   = tmp;
655   for (i=1; i<n; i++) {
656     v   = aa + ai[i] ;
657     vi  = aj + ai[i] ;
658     nz  = a->diag[i] - ai[i];
659     sum = b[*r++];
660     SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
661     tmp[i] = sum;
662   }
663 
664   /* backward solve the upper triangular */
665   for (i=n-1; i>=0; i--){
666     v   = aa + a->diag[i] + 1;
667     vi  = aj + a->diag[i] + 1;
668     nz  = ai[i+1] - a->diag[i] - 1;
669     sum = tmp[i];
670     SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
671     x[*c--] = tmp[i] = sum*aa[a->diag[i]];
672   }
673 
674   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
675   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
676   ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
677   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
678   ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr);
679   PetscFunctionReturn(0);
680 }
681 
682 #undef __FUNCT__
683 #define __FUNCT__ "MatMatSolve_SeqAIJ"
684 PetscErrorCode MatMatSolve_SeqAIJ(Mat A,Mat B,Mat X)
685 {
686   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
687   IS              iscol = a->col,isrow = a->row;
688   PetscErrorCode  ierr;
689   PetscInt        i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
690   PetscInt        nz,neq;
691   const PetscInt  *rout,*cout,*r,*c;
692   PetscScalar     *x,*b,*tmp,*tmps,sum;
693   const MatScalar *aa = a->a,*v;
694   PetscTruth      bisdense,xisdense;
695 
696   PetscFunctionBegin;
697   if (!n) PetscFunctionReturn(0);
698 
699   ierr = PetscTypeCompare((PetscObject)B,MATSEQDENSE,&bisdense);CHKERRQ(ierr);
700   if (!bisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"B matrix must be a SeqDense matrix");
701   ierr = PetscTypeCompare((PetscObject)X,MATSEQDENSE,&xisdense);CHKERRQ(ierr);
702   if (!xisdense) SETERRQ(PETSC_ERR_ARG_INCOMP,"X matrix must be a SeqDense matrix");
703 
704   ierr = MatGetArray(B,&b);CHKERRQ(ierr);
705   ierr = MatGetArray(X,&x);CHKERRQ(ierr);
706 
707   tmp  = a->solve_work;
708   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
709   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout;
710 
711   for (neq=0; neq<B->cmap->n; neq++){
712     /* forward solve the lower triangular */
713     tmp[0] = b[r[0]];
714     tmps   = tmp;
715     for (i=1; i<n; i++) {
716       v   = aa + ai[i] ;
717       vi  = aj + ai[i] ;
718       nz  = a->diag[i] - ai[i];
719       sum = b[r[i]];
720       SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
721       tmp[i] = sum;
722     }
723     /* backward solve the upper triangular */
724     for (i=n-1; i>=0; i--){
725       v   = aa + a->diag[i] + 1;
726       vi  = aj + a->diag[i] + 1;
727       nz  = ai[i+1] - a->diag[i] - 1;
728       sum = tmp[i];
729       SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
730       x[c[i]] = tmp[i] = sum*aa[a->diag[i]];
731     }
732 
733     b += n;
734     x += n;
735   }
736   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
737   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
738   ierr = MatRestoreArray(B,&b);CHKERRQ(ierr);
739   ierr = MatRestoreArray(X,&x);CHKERRQ(ierr);
740   ierr = PetscLogFlops(B->cmap->n*(2.0*a->nz - n));CHKERRQ(ierr);
741   PetscFunctionReturn(0);
742 }
743 
744 #undef __FUNCT__
745 #define __FUNCT__ "MatSolve_SeqAIJ_InplaceWithPerm"
746 PetscErrorCode MatSolve_SeqAIJ_InplaceWithPerm(Mat A,Vec bb,Vec xx)
747 {
748   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
749   IS              iscol = a->col,isrow = a->row;
750   PetscErrorCode  ierr;
751   const PetscInt  *r,*c,*rout,*cout;
752   PetscInt        i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
753   PetscInt        nz,row;
754   PetscScalar     *x,*b,*tmp,*tmps,sum;
755   const MatScalar *aa = a->a,*v;
756 
757   PetscFunctionBegin;
758   if (!n) PetscFunctionReturn(0);
759 
760   ierr = VecGetArray(bb,&b);CHKERRQ(ierr);
761   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
762   tmp  = a->solve_work;
763 
764   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
765   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1);
766 
767   /* forward solve the lower triangular */
768   tmp[0] = b[*r++];
769   tmps   = tmp;
770   for (row=1; row<n; row++) {
771     i   = rout[row]; /* permuted row */
772     v   = aa + ai[i] ;
773     vi  = aj + ai[i] ;
774     nz  = a->diag[i] - ai[i];
775     sum = b[*r++];
776     SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
777     tmp[row] = sum;
778   }
779 
780   /* backward solve the upper triangular */
781   for (row=n-1; row>=0; row--){
782     i   = rout[row]; /* permuted row */
783     v   = aa + a->diag[i] + 1;
784     vi  = aj + a->diag[i] + 1;
785     nz  = ai[i+1] - a->diag[i] - 1;
786     sum = tmp[row];
787     SPARSEDENSEMDOT(sum,tmps,v,vi,nz);
788     x[*c--] = tmp[row] = sum*aa[a->diag[i]];
789   }
790 
791   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
792   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
793   ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr);
794   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
795   ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr);
796   PetscFunctionReturn(0);
797 }
798 
799 /* ----------------------------------------------------------- */
800 #undef __FUNCT__
801 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering"
802 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering(Mat A,Vec bb,Vec xx)
803 {
804   Mat_SeqAIJ        *a = (Mat_SeqAIJ*)A->data;
805   PetscErrorCode    ierr;
806   PetscInt          n = A->rmap->n;
807   const PetscInt    *ai = a->i,*aj = a->j,*adiag = a->diag,*vi;
808   PetscScalar       *x;
809   const PetscScalar *b;
810   const MatScalar   *aa = a->a;
811 #if !defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ)
812   PetscInt          adiag_i,i,nz,ai_i;
813   const MatScalar   *v;
814   PetscScalar       sum;
815 #endif
816 
817   PetscFunctionBegin;
818   if (!n) PetscFunctionReturn(0);
819 
820   ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
821   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
822 
823 #if defined(PETSC_USE_FORTRAN_KERNEL_SOLVEAIJ)
824   fortransolveaij_(&n,x,ai,aj,adiag,aa,b);
825 #else
826   /* forward solve the lower triangular */
827   x[0] = b[0];
828   for (i=1; i<n; i++) {
829     ai_i = ai[i];
830     v    = aa + ai_i;
831     vi   = aj + ai_i;
832     nz   = adiag[i] - ai_i;
833     sum  = b[i];
834     while (nz--) sum -= *v++ * x[*vi++];
835     x[i] = sum;
836   }
837 
838   /* backward solve the upper triangular */
839   for (i=n-1; i>=0; i--){
840     adiag_i = adiag[i];
841     v       = aa + adiag_i + 1;
842     vi      = aj + adiag_i + 1;
843     nz      = ai[i+1] - adiag_i - 1;
844     sum     = x[i];
845     while (nz--) sum -= *v++ * x[*vi++];
846     x[i]    = sum*aa[adiag_i];
847   }
848 #endif
849   ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr);
850   ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
851   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
852   PetscFunctionReturn(0);
853 }
854 
855 #undef __FUNCT__
856 #define __FUNCT__ "MatSolveAdd_SeqAIJ"
857 PetscErrorCode MatSolveAdd_SeqAIJ(Mat A,Vec bb,Vec yy,Vec xx)
858 {
859   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
860   IS              iscol = a->col,isrow = a->row;
861   PetscErrorCode  ierr;
862   PetscInt        i, n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
863   PetscInt        nz;
864   const PetscInt  *rout,*cout,*r,*c;
865   PetscScalar     *x,*b,*tmp,sum;
866   const MatScalar *aa = a->a,*v;
867 
868   PetscFunctionBegin;
869   if (yy != xx) {ierr = VecCopy(yy,xx);CHKERRQ(ierr);}
870 
871   ierr = VecGetArray(bb,&b);CHKERRQ(ierr);
872   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
873   tmp  = a->solve_work;
874 
875   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
876   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1);
877 
878   /* forward solve the lower triangular */
879   tmp[0] = b[*r++];
880   for (i=1; i<n; i++) {
881     v   = aa + ai[i] ;
882     vi  = aj + ai[i] ;
883     nz  = a->diag[i] - ai[i];
884     sum = b[*r++];
885     while (nz--) sum -= *v++ * tmp[*vi++ ];
886     tmp[i] = sum;
887   }
888 
889   /* backward solve the upper triangular */
890   for (i=n-1; i>=0; i--){
891     v   = aa + a->diag[i] + 1;
892     vi  = aj + a->diag[i] + 1;
893     nz  = ai[i+1] - a->diag[i] - 1;
894     sum = tmp[i];
895     while (nz--) sum -= *v++ * tmp[*vi++ ];
896     tmp[i] = sum*aa[a->diag[i]];
897     x[*c--] += tmp[i];
898   }
899 
900   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
901   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
902   ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr);
903   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
904   ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr);
905 
906   PetscFunctionReturn(0);
907 }
908 /* -------------------------------------------------------------------*/
909 #undef __FUNCT__
910 #define __FUNCT__ "MatSolveTranspose_SeqAIJ"
911 PetscErrorCode MatSolveTranspose_SeqAIJ(Mat A,Vec bb,Vec xx)
912 {
913   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
914   IS              iscol = a->col,isrow = a->row;
915   PetscErrorCode  ierr;
916   const PetscInt  *rout,*cout,*r,*c;
917   PetscInt        i,n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
918   PetscInt        nz,*diag = a->diag;
919   PetscScalar     *x,*b,*tmp,s1;
920   const MatScalar *aa = a->a,*v;
921 
922   PetscFunctionBegin;
923   ierr = VecGetArray(bb,&b);CHKERRQ(ierr);
924   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
925   tmp  = a->solve_work;
926 
927   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
928   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout;
929 
930   /* copy the b into temp work space according to permutation */
931   for (i=0; i<n; i++) tmp[i] = b[c[i]];
932 
933   /* forward solve the U^T */
934   for (i=0; i<n; i++) {
935     v   = aa + diag[i] ;
936     vi  = aj + diag[i] + 1;
937     nz  = ai[i+1] - diag[i] - 1;
938     s1  = tmp[i];
939     s1 *= (*v++);  /* multiply by inverse of diagonal entry */
940     while (nz--) {
941       tmp[*vi++ ] -= (*v++)*s1;
942     }
943     tmp[i] = s1;
944   }
945 
946   /* backward solve the L^T */
947   for (i=n-1; i>=0; i--){
948     v   = aa + diag[i] - 1 ;
949     vi  = aj + diag[i] - 1 ;
950     nz  = diag[i] - ai[i];
951     s1  = tmp[i];
952     while (nz--) {
953       tmp[*vi-- ] -= (*v--)*s1;
954     }
955   }
956 
957   /* copy tmp into x according to permutation */
958   for (i=0; i<n; i++) x[r[i]] = tmp[i];
959 
960   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
961   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
962   ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr);
963   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
964 
965   ierr = PetscLogFlops(2.0*a->nz-A->cmap->n);CHKERRQ(ierr);
966   PetscFunctionReturn(0);
967 }
968 
969 #undef __FUNCT__
970 #define __FUNCT__ "MatSolveTransposeAdd_SeqAIJ"
971 PetscErrorCode MatSolveTransposeAdd_SeqAIJ(Mat A,Vec bb,Vec zz,Vec xx)
972 {
973   Mat_SeqAIJ      *a = (Mat_SeqAIJ*)A->data;
974   IS              iscol = a->col,isrow = a->row;
975   PetscErrorCode  ierr;
976   const PetscInt  *r,*c,*rout,*cout;
977   PetscInt        i,n = A->rmap->n,*vi,*ai = a->i,*aj = a->j;
978   PetscInt        nz,*diag = a->diag;
979   PetscScalar     *x,*b,*tmp;
980   const MatScalar *aa = a->a,*v;
981 
982   PetscFunctionBegin;
983   if (zz != xx) {ierr = VecCopy(zz,xx);CHKERRQ(ierr);}
984 
985   ierr = VecGetArray(bb,&b);CHKERRQ(ierr);
986   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
987   tmp = a->solve_work;
988 
989   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
990   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout;
991 
992   /* copy the b into temp work space according to permutation */
993   for (i=0; i<n; i++) tmp[i] = b[c[i]];
994 
995   /* forward solve the U^T */
996   for (i=0; i<n; i++) {
997     v   = aa + diag[i] ;
998     vi  = aj + diag[i] + 1;
999     nz  = ai[i+1] - diag[i] - 1;
1000     tmp[i] *= *v++;
1001     while (nz--) {
1002       tmp[*vi++ ] -= (*v++)*tmp[i];
1003     }
1004   }
1005 
1006   /* backward solve the L^T */
1007   for (i=n-1; i>=0; i--){
1008     v   = aa + diag[i] - 1 ;
1009     vi  = aj + diag[i] - 1 ;
1010     nz  = diag[i] - ai[i];
1011     while (nz--) {
1012       tmp[*vi-- ] -= (*v--)*tmp[i];
1013     }
1014   }
1015 
1016   /* copy tmp into x according to permutation */
1017   for (i=0; i<n; i++) x[r[i]] += tmp[i];
1018 
1019   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
1020   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
1021   ierr = VecRestoreArray(bb,&b);CHKERRQ(ierr);
1022   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
1023 
1024   ierr = PetscLogFlops(2.0*a->nz);CHKERRQ(ierr);
1025   PetscFunctionReturn(0);
1026 }
1027 /* ----------------------------------------------------------------*/
1028 EXTERN PetscErrorCode Mat_CheckInode(Mat,PetscTruth);
1029 EXTERN PetscErrorCode MatDuplicateNoCreate_SeqAIJ(Mat,Mat,MatDuplicateOption);
1030 
1031 #undef __FUNCT__
1032 #define __FUNCT__ "MatILUFactorSymbolic_SeqAIJ"
1033 PetscErrorCode MatILUFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS isrow,IS iscol,const MatFactorInfo *info)
1034 {
1035   Mat_SeqAIJ         *a = (Mat_SeqAIJ*)A->data,*b;
1036   IS                 isicol;
1037   PetscErrorCode     ierr;
1038   const PetscInt     *r,*ic;
1039   PetscInt           n=A->rmap->n,*ai=a->i,*aj=a->j,d;
1040   PetscInt           *bi,*cols,nnz,*cols_lvl;
1041   PetscInt           *bdiag,prow,fm,nzbd,reallocs=0,dcount=0;
1042   PetscInt           i,levels,diagonal_fill;
1043   PetscTruth         col_identity,row_identity;
1044   PetscReal          f;
1045   PetscInt           nlnk,*lnk,*lnk_lvl=PETSC_NULL;
1046   PetscBT            lnkbt;
1047   PetscInt           nzi,*bj,**bj_ptr,**bjlvl_ptr;
1048   PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
1049   PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL;
1050   PetscTruth         missing;
1051 
1052   PetscFunctionBegin;
1053   if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n);
1054   f             = info->fill;
1055   levels        = (PetscInt)info->levels;
1056   diagonal_fill = (PetscInt)info->diagonal_fill;
1057   ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr);
1058 
1059   /* special case that simply copies fill pattern */
1060   ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr);
1061   ierr = ISIdentity(iscol,&col_identity);CHKERRQ(ierr);
1062   if (!levels && row_identity && col_identity) {
1063     ierr = MatDuplicateNoCreate_SeqAIJ(fact,A,MAT_DO_NOT_COPY_VALUES);CHKERRQ(ierr);
1064     fact->factor = MAT_FACTOR_ILU;
1065     (fact)->info.factor_mallocs    = 0;
1066     (fact)->info.fill_ratio_given  = info->fill;
1067     (fact)->info.fill_ratio_needed = 1.0;
1068     b               = (Mat_SeqAIJ*)(fact)->data;
1069     ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr);
1070     if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d);
1071     b->row              = isrow;
1072     b->col              = iscol;
1073     b->icol             = isicol;
1074     ierr                = PetscMalloc(((fact)->rmap->n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
1075     ierr                = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
1076     ierr                = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
1077     (fact)->ops->lufactornumeric =  MatLUFactorNumeric_SeqAIJ;
1078     ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr);
1079     PetscFunctionReturn(0);
1080   }
1081 
1082   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
1083   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
1084 
1085   /* get new row pointers */
1086   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr);
1087   bi[0] = 0;
1088   /* bdiag is location of diagonal in factor */
1089   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr);
1090   bdiag[0]  = 0;
1091 
1092   ierr = PetscMalloc((2*n+1)*sizeof(PetscInt**),&bj_ptr);CHKERRQ(ierr);
1093   bjlvl_ptr = (PetscInt**)(bj_ptr + n);
1094 
1095   /* create a linked list for storing column indices of the active row */
1096   nlnk = n + 1;
1097   ierr = PetscIncompleteLLCreate(n,n,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr);
1098 
1099   /* initial FreeSpace size is f*(ai[n]+1) */
1100   ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space);CHKERRQ(ierr);
1101   current_space = free_space;
1102   ierr = PetscFreeSpaceGet((PetscInt)(f*(ai[n]+1)),&free_space_lvl);CHKERRQ(ierr);
1103   current_space_lvl = free_space_lvl;
1104 
1105   for (i=0; i<n; i++) {
1106     nzi = 0;
1107     /* copy current row into linked list */
1108     nnz  = ai[r[i]+1] - ai[r[i]];
1109     if (!nnz) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i);
1110     cols = aj + ai[r[i]];
1111     lnk[i] = -1; /* marker to indicate if diagonal exists */
1112     ierr = PetscIncompleteLLInit(nnz,cols,n,ic,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr);
1113     nzi += nlnk;
1114 
1115     /* make sure diagonal entry is included */
1116     if (diagonal_fill && lnk[i] == -1) {
1117       fm = n;
1118       while (lnk[fm] < i) fm = lnk[fm];
1119       lnk[i]     = lnk[fm]; /* insert diagonal into linked list */
1120       lnk[fm]    = i;
1121       lnk_lvl[i] = 0;
1122       nzi++; dcount++;
1123     }
1124 
1125     /* add pivot rows into the active row */
1126     nzbd = 0;
1127     prow = lnk[n];
1128     while (prow < i) {
1129       nnz      = bdiag[prow];
1130       cols     = bj_ptr[prow] + nnz + 1;
1131       cols_lvl = bjlvl_ptr[prow] + nnz + 1;
1132       nnz      = bi[prow+1] - bi[prow] - nnz - 1;
1133       ierr = PetscILULLAddSorted(nnz,cols,levels,cols_lvl,prow,nlnk,lnk,lnk_lvl,lnkbt,prow);CHKERRQ(ierr);
1134       nzi += nlnk;
1135       prow = lnk[prow];
1136       nzbd++;
1137     }
1138     bdiag[i] = nzbd;
1139     bi[i+1]  = bi[i] + nzi;
1140 
1141     /* if free space is not available, make more free space */
1142     if (current_space->local_remaining<nzi) {
1143       nnz = nzi*(n - i); /* estimated and max additional space needed */
1144       ierr = PetscFreeSpaceGet(nnz,&current_space);CHKERRQ(ierr);
1145       ierr = PetscFreeSpaceGet(nnz,&current_space_lvl);CHKERRQ(ierr);
1146       reallocs++;
1147     }
1148 
1149     /* copy data into free_space and free_space_lvl, then initialize lnk */
1150     ierr = PetscIncompleteLLClean(n,n,nzi,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr);
1151     bj_ptr[i]    = current_space->array;
1152     bjlvl_ptr[i] = current_space_lvl->array;
1153 
1154     /* make sure the active row i has diagonal entry */
1155     if (*(bj_ptr[i]+bdiag[i]) != i) {
1156       SETERRQ1(PETSC_ERR_MAT_LU_ZRPVT,"Row %D has missing diagonal in factored matrix\n\
1157     try running with -pc_factor_nonzeros_along_diagonal or -pc_factor_diagonal_fill",i);
1158     }
1159 
1160     current_space->array           += nzi;
1161     current_space->local_used      += nzi;
1162     current_space->local_remaining -= nzi;
1163     current_space_lvl->array           += nzi;
1164     current_space_lvl->local_used      += nzi;
1165     current_space_lvl->local_remaining -= nzi;
1166   }
1167 
1168   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
1169   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
1170 
1171   /* destroy list of free space and other temporary arrays */
1172   ierr = PetscMalloc((bi[n]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr);
1173   ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr);
1174   ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
1175   ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr);
1176   ierr = PetscFree(bj_ptr);CHKERRQ(ierr);
1177 
1178 #if defined(PETSC_USE_INFO)
1179   {
1180     PetscReal af = ((PetscReal)bi[n])/((PetscReal)ai[n]);
1181     ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,f,af);CHKERRQ(ierr);
1182     ierr = PetscInfo1(A,"Run with -[sub_]pc_factor_fill %G or use \n",af);CHKERRQ(ierr);
1183     ierr = PetscInfo1(A,"PCFactorSetFill([sub]pc,%G);\n",af);CHKERRQ(ierr);
1184     ierr = PetscInfo(A,"for best performance.\n");CHKERRQ(ierr);
1185     if (diagonal_fill) {
1186       ierr = PetscInfo1(A,"Detected and replaced %D missing diagonals",dcount);CHKERRQ(ierr);
1187     }
1188   }
1189 #endif
1190 
1191   /* put together the new matrix */
1192   ierr = MatSeqAIJSetPreallocation_SeqAIJ(fact,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr);
1193   ierr = PetscLogObjectParent(fact,isicol);CHKERRQ(ierr);
1194   b = (Mat_SeqAIJ*)(fact)->data;
1195   b->free_a       = PETSC_TRUE;
1196   b->free_ij      = PETSC_TRUE;
1197   b->singlemalloc = PETSC_FALSE;
1198   ierr = PetscMalloc( (bi[n] )*sizeof(PetscScalar),&b->a);CHKERRQ(ierr);
1199   b->j          = bj;
1200   b->i          = bi;
1201   for (i=0; i<n; i++) bdiag[i] += bi[i];
1202   b->diag       = bdiag;
1203   b->ilen       = 0;
1204   b->imax       = 0;
1205   b->row        = isrow;
1206   b->col        = iscol;
1207   ierr          = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
1208   ierr          = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
1209   b->icol       = isicol;
1210   ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
1211   /* In b structure:  Free imax, ilen, old a, old j.
1212      Allocate bdiag, solve_work, new a, new j */
1213   ierr = PetscLogObjectMemory(fact,(bi[n]-n) * (sizeof(PetscInt)+sizeof(PetscScalar)));CHKERRQ(ierr);
1214   b->maxnz             = b->nz = bi[n] ;
1215   (fact)->info.factor_mallocs    = reallocs;
1216   (fact)->info.fill_ratio_given  = f;
1217   (fact)->info.fill_ratio_needed = ((PetscReal)bi[n])/((PetscReal)ai[n]);
1218   (fact)->ops->lufactornumeric =  MatLUFactorNumeric_SeqAIJ;
1219   ierr = MatILUFactorSymbolic_Inode(fact,A,isrow,iscol,info);CHKERRQ(ierr);
1220   PetscFunctionReturn(0);
1221 }
1222 
1223 #include "../src/mat/impls/sbaij/seq/sbaij.h"
1224 #undef __FUNCT__
1225 #define __FUNCT__ "MatCholeskyFactorNumeric_SeqAIJ"
1226 PetscErrorCode MatCholeskyFactorNumeric_SeqAIJ(Mat B,Mat A,const MatFactorInfo *info)
1227 {
1228   Mat            C = B;
1229   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data;
1230   Mat_SeqSBAIJ   *b=(Mat_SeqSBAIJ*)C->data;
1231   IS             ip=b->row,iip = b->icol;
1232   PetscErrorCode ierr;
1233   const PetscInt *rip,*riip;
1234   PetscInt       i,j,mbs=A->rmap->n,*bi=b->i,*bj=b->j,*bcol;
1235   PetscInt       *ai=a->i,*aj=a->j;
1236   PetscInt       k,jmin,jmax,*jl,*il,col,nexti,ili,nz;
1237   MatScalar      *rtmp,*ba=b->a,*bval,*aa=a->a,dk,uikdi;
1238   PetscReal      zeropivot,rs,shiftnz;
1239   PetscReal      shiftpd;
1240   ChShift_Ctx    sctx;
1241   PetscInt       newshift;
1242   PetscTruth     perm_identity;
1243 
1244   PetscFunctionBegin;
1245 
1246   shiftnz   = info->shiftnz;
1247   shiftpd   = info->shiftpd;
1248   zeropivot = info->zeropivot;
1249 
1250   ierr  = ISGetIndices(ip,&rip);CHKERRQ(ierr);
1251   ierr  = ISGetIndices(iip,&riip);CHKERRQ(ierr);
1252 
1253   /* initialization */
1254   nz   = (2*mbs+1)*sizeof(PetscInt)+mbs*sizeof(MatScalar);
1255   ierr = PetscMalloc(nz,&il);CHKERRQ(ierr);
1256   jl   = il + mbs;
1257   rtmp = (MatScalar*)(jl + mbs);
1258 
1259   sctx.shift_amount = 0;
1260   sctx.nshift       = 0;
1261   do {
1262     sctx.chshift = PETSC_FALSE;
1263     for (i=0; i<mbs; i++) {
1264       rtmp[i] = 0.0; jl[i] = mbs; il[0] = 0;
1265     }
1266 
1267     for (k = 0; k<mbs; k++){
1268       bval = ba + bi[k];
1269       /* initialize k-th row by the perm[k]-th row of A */
1270       jmin = ai[rip[k]]; jmax = ai[rip[k]+1];
1271       for (j = jmin; j < jmax; j++){
1272         col = riip[aj[j]];
1273         if (col >= k){ /* only take upper triangular entry */
1274           rtmp[col] = aa[j];
1275           *bval++  = 0.0; /* for in-place factorization */
1276         }
1277       }
1278       /* shift the diagonal of the matrix */
1279       if (sctx.nshift) rtmp[k] += sctx.shift_amount;
1280 
1281       /* modify k-th row by adding in those rows i with U(i,k)!=0 */
1282       dk = rtmp[k];
1283       i = jl[k]; /* first row to be added to k_th row  */
1284 
1285       while (i < k){
1286         nexti = jl[i]; /* next row to be added to k_th row */
1287 
1288         /* compute multiplier, update diag(k) and U(i,k) */
1289         ili = il[i];  /* index of first nonzero element in U(i,k:bms-1) */
1290         uikdi = - ba[ili]*ba[bi[i]];  /* diagonal(k) */
1291         dk += uikdi*ba[ili];
1292         ba[ili] = uikdi; /* -U(i,k) */
1293 
1294         /* add multiple of row i to k-th row */
1295         jmin = ili + 1; jmax = bi[i+1];
1296         if (jmin < jmax){
1297           for (j=jmin; j<jmax; j++) rtmp[bj[j]] += uikdi*ba[j];
1298           /* update il and jl for row i */
1299           il[i] = jmin;
1300           j = bj[jmin]; jl[i] = jl[j]; jl[j] = i;
1301         }
1302         i = nexti;
1303       }
1304 
1305       /* shift the diagonals when zero pivot is detected */
1306       /* compute rs=sum of abs(off-diagonal) */
1307       rs   = 0.0;
1308       jmin = bi[k]+1;
1309       nz   = bi[k+1] - jmin;
1310       bcol = bj + jmin;
1311       while (nz--){
1312         rs += PetscAbsScalar(rtmp[*bcol]);
1313         bcol++;
1314       }
1315 
1316       sctx.rs = rs;
1317       sctx.pv = dk;
1318       ierr = MatCholeskyCheckShift_inline(info,sctx,k,newshift);CHKERRQ(ierr);
1319 
1320       if (newshift == 1) {
1321         if (!sctx.shift_amount) {
1322           sctx.shift_amount = 1e-5;
1323         }
1324         break;
1325       }
1326 
1327       /* copy data into U(k,:) */
1328       ba[bi[k]] = 1.0/dk; /* U(k,k) */
1329       jmin = bi[k]+1; jmax = bi[k+1];
1330       if (jmin < jmax) {
1331         for (j=jmin; j<jmax; j++){
1332           col = bj[j]; ba[j] = rtmp[col]; rtmp[col] = 0.0;
1333         }
1334         /* add the k-th row into il and jl */
1335         il[k] = jmin;
1336         i = bj[jmin]; jl[k] = jl[i]; jl[i] = k;
1337       }
1338     }
1339   } while (sctx.chshift);
1340   ierr = PetscFree(il);CHKERRQ(ierr);
1341 
1342   ierr = ISRestoreIndices(ip,&rip);CHKERRQ(ierr);
1343   ierr = ISRestoreIndices(iip,&riip);CHKERRQ(ierr);
1344 
1345   ierr = ISIdentity(ip,&perm_identity);CHKERRQ(ierr);
1346   if (perm_identity){
1347     (B)->ops->solve           = MatSolve_SeqSBAIJ_1_NaturalOrdering;
1348     (B)->ops->solvetranspose  = MatSolve_SeqSBAIJ_1_NaturalOrdering;
1349     (B)->ops->forwardsolve    = MatForwardSolve_SeqSBAIJ_1_NaturalOrdering;
1350     (B)->ops->backwardsolve   = MatBackwardSolve_SeqSBAIJ_1_NaturalOrdering;
1351   } else {
1352     (B)->ops->solve           = MatSolve_SeqSBAIJ_1;
1353     (B)->ops->solvetranspose  = MatSolve_SeqSBAIJ_1;
1354     (B)->ops->forwardsolve    = MatForwardSolve_SeqSBAIJ_1;
1355     (B)->ops->backwardsolve   = MatBackwardSolve_SeqSBAIJ_1;
1356   }
1357 
1358   C->assembled    = PETSC_TRUE;
1359   C->preallocated = PETSC_TRUE;
1360   ierr = PetscLogFlops(C->rmap->n);CHKERRQ(ierr);
1361   if (sctx.nshift){
1362     if (shiftnz) {
1363       ierr = PetscInfo2(A,"number of shiftnz tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr);
1364     } else if (shiftpd) {
1365       ierr = PetscInfo2(A,"number of shiftpd tries %D, shift_amount %G\n",sctx.nshift,sctx.shift_amount);CHKERRQ(ierr);
1366     }
1367   }
1368   PetscFunctionReturn(0);
1369 }
1370 
1371 #undef __FUNCT__
1372 #define __FUNCT__ "MatICCFactorSymbolic_SeqAIJ"
1373 PetscErrorCode MatICCFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info)
1374 {
1375   Mat_SeqAIJ         *a = (Mat_SeqAIJ*)A->data;
1376   Mat_SeqSBAIJ       *b;
1377   PetscErrorCode     ierr;
1378   PetscTruth         perm_identity,missing;
1379   PetscInt           reallocs=0,i,*ai=a->i,*aj=a->j,am=A->rmap->n,*ui;
1380   const PetscInt     *rip,*riip;
1381   PetscInt           jmin,jmax,nzk,k,j,*jl,prow,*il,nextprow;
1382   PetscInt           nlnk,*lnk,*lnk_lvl=PETSC_NULL,d;
1383   PetscInt           ncols,ncols_upper,*cols,*ajtmp,*uj,**uj_ptr,**uj_lvl_ptr;
1384   PetscReal          fill=info->fill,levels=info->levels;
1385   PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
1386   PetscFreeSpaceList free_space_lvl=PETSC_NULL,current_space_lvl=PETSC_NULL;
1387   PetscBT            lnkbt;
1388   IS                 iperm;
1389 
1390   PetscFunctionBegin;
1391   if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n);
1392   ierr = MatMissingDiagonal(A,&missing,&d);CHKERRQ(ierr);
1393   if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",d);
1394   ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr);
1395   ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr);
1396 
1397   ierr = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr);
1398   ui[0] = 0;
1399 
1400   /* ICC(0) without matrix ordering: simply copies fill pattern */
1401   if (!levels && perm_identity) {
1402 
1403     for (i=0; i<am; i++) {
1404       ui[i+1] = ui[i] + ai[i+1] - a->diag[i];
1405     }
1406     ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr);
1407     cols = uj;
1408     for (i=0; i<am; i++) {
1409       aj    = a->j + a->diag[i];
1410       ncols = ui[i+1] - ui[i];
1411       for (j=0; j<ncols; j++) *cols++ = *aj++;
1412     }
1413   } else { /* case: levels>0 || (levels=0 && !perm_identity) */
1414     ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr);
1415     ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr);
1416 
1417     /* initialization */
1418     ierr  = PetscMalloc((am+1)*sizeof(PetscInt),&ajtmp);CHKERRQ(ierr);
1419 
1420     /* jl: linked list for storing indices of the pivot rows
1421        il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */
1422     ierr = PetscMalloc((2*am+1)*sizeof(PetscInt)+2*am*sizeof(PetscInt**),&jl);CHKERRQ(ierr);
1423     il         = jl + am;
1424     uj_ptr     = (PetscInt**)(il + am);
1425     uj_lvl_ptr = (PetscInt**)(uj_ptr + am);
1426     for (i=0; i<am; i++){
1427       jl[i] = am; il[i] = 0;
1428     }
1429 
1430     /* create and initialize a linked list for storing column indices of the active row k */
1431     nlnk = am + 1;
1432     ierr = PetscIncompleteLLCreate(am,am,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr);
1433 
1434     /* initial FreeSpace size is fill*(ai[am]+1) */
1435     ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr);
1436     current_space = free_space;
1437     ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space_lvl);CHKERRQ(ierr);
1438     current_space_lvl = free_space_lvl;
1439 
1440     for (k=0; k<am; k++){  /* for each active row k */
1441       /* initialize lnk by the column indices of row rip[k] of A */
1442       nzk   = 0;
1443       ncols = ai[rip[k]+1] - ai[rip[k]];
1444       if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k);
1445       ncols_upper = 0;
1446       for (j=0; j<ncols; j++){
1447         i = *(aj + ai[rip[k]] + j); /* unpermuted column index */
1448         if (riip[i] >= k){ /* only take upper triangular entry */
1449           ajtmp[ncols_upper] = i;
1450           ncols_upper++;
1451         }
1452       }
1453       ierr = PetscIncompleteLLInit(ncols_upper,ajtmp,am,riip,nlnk,lnk,lnk_lvl,lnkbt);CHKERRQ(ierr);
1454       nzk += nlnk;
1455 
1456       /* update lnk by computing fill-in for each pivot row to be merged in */
1457       prow = jl[k]; /* 1st pivot row */
1458 
1459       while (prow < k){
1460         nextprow = jl[prow];
1461 
1462         /* merge prow into k-th row */
1463         jmin = il[prow] + 1;  /* index of the 2nd nzero entry in U(prow,k:am-1) */
1464         jmax = ui[prow+1];
1465         ncols = jmax-jmin;
1466         i     = jmin - ui[prow];
1467         cols  = uj_ptr[prow] + i; /* points to the 2nd nzero entry in U(prow,k:am-1) */
1468         uj    = uj_lvl_ptr[prow] + i; /* levels of cols */
1469         j     = *(uj - 1);
1470         ierr = PetscICCLLAddSorted(ncols,cols,levels,uj,am,nlnk,lnk,lnk_lvl,lnkbt,j);CHKERRQ(ierr);
1471         nzk += nlnk;
1472 
1473         /* update il and jl for prow */
1474         if (jmin < jmax){
1475           il[prow] = jmin;
1476           j = *cols; jl[prow] = jl[j]; jl[j] = prow;
1477         }
1478         prow = nextprow;
1479       }
1480 
1481       /* if free space is not available, make more free space */
1482       if (current_space->local_remaining<nzk) {
1483         i = am - k + 1; /* num of unfactored rows */
1484         i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */
1485         ierr = PetscFreeSpaceGet(i,&current_space);CHKERRQ(ierr);
1486         ierr = PetscFreeSpaceGet(i,&current_space_lvl);CHKERRQ(ierr);
1487         reallocs++;
1488       }
1489 
1490       /* copy data into free_space and free_space_lvl, then initialize lnk */
1491       if (nzk == 0) SETERRQ1(PETSC_ERR_ARG_WRONG,"Empty row %D in ICC matrix factor",k);
1492       ierr = PetscIncompleteLLClean(am,am,nzk,lnk,lnk_lvl,current_space->array,current_space_lvl->array,lnkbt);CHKERRQ(ierr);
1493 
1494       /* add the k-th row into il and jl */
1495       if (nzk > 1){
1496         i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */
1497         jl[k] = jl[i]; jl[i] = k;
1498         il[k] = ui[k] + 1;
1499       }
1500       uj_ptr[k]     = current_space->array;
1501       uj_lvl_ptr[k] = current_space_lvl->array;
1502 
1503       current_space->array           += nzk;
1504       current_space->local_used      += nzk;
1505       current_space->local_remaining -= nzk;
1506 
1507       current_space_lvl->array           += nzk;
1508       current_space_lvl->local_used      += nzk;
1509       current_space_lvl->local_remaining -= nzk;
1510 
1511       ui[k+1] = ui[k] + nzk;
1512     }
1513 
1514 #if defined(PETSC_USE_INFO)
1515     if (ai[am] != 0) {
1516       PetscReal af = (PetscReal)ui[am]/((PetscReal)ai[am]);
1517       ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr);
1518       ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr);
1519       ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr);
1520     } else {
1521       ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr);
1522     }
1523 #endif
1524 
1525     ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr);
1526     ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr);
1527     ierr = PetscFree(jl);CHKERRQ(ierr);
1528     ierr = PetscFree(ajtmp);CHKERRQ(ierr);
1529 
1530     /* destroy list of free space and other temporary array(s) */
1531     ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr);
1532     ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr);
1533     ierr = PetscIncompleteLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
1534     ierr = PetscFreeSpaceDestroy(free_space_lvl);CHKERRQ(ierr);
1535 
1536   } /* end of case: levels>0 || (levels=0 && !perm_identity) */
1537 
1538   /* put together the new matrix in MATSEQSBAIJ format */
1539 
1540   b    = (Mat_SeqSBAIJ*)(fact)->data;
1541   b->singlemalloc = PETSC_FALSE;
1542   ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr);
1543   b->j    = uj;
1544   b->i    = ui;
1545   b->diag = 0;
1546   b->ilen = 0;
1547   b->imax = 0;
1548   b->row  = perm;
1549   b->col  = perm;
1550   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
1551   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
1552   b->icol = iperm;
1553   b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */
1554   ierr    = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
1555   ierr = PetscLogObjectMemory((fact),(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr);
1556   b->maxnz   = b->nz = ui[am];
1557   b->free_a  = PETSC_TRUE;
1558   b->free_ij = PETSC_TRUE;
1559 
1560   (fact)->info.factor_mallocs    = reallocs;
1561   (fact)->info.fill_ratio_given  = fill;
1562   if (ai[am] != 0) {
1563     (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]);
1564   } else {
1565     (fact)->info.fill_ratio_needed = 0.0;
1566   }
1567   (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ;
1568   PetscFunctionReturn(0);
1569 }
1570 
1571 #undef __FUNCT__
1572 #define __FUNCT__ "MatCholeskyFactorSymbolic_SeqAIJ"
1573 PetscErrorCode MatCholeskyFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS perm,const MatFactorInfo *info)
1574 {
1575   Mat_SeqAIJ         *a = (Mat_SeqAIJ*)A->data;
1576   Mat_SeqSBAIJ       *b;
1577   PetscErrorCode     ierr;
1578   PetscTruth         perm_identity;
1579   PetscReal          fill = info->fill;
1580   const PetscInt     *rip,*riip;
1581   PetscInt           i,am=A->rmap->n,*ai=a->i,*aj=a->j,reallocs=0,prow;
1582   PetscInt           *jl,jmin,jmax,nzk,*ui,k,j,*il,nextprow;
1583   PetscInt           nlnk,*lnk,ncols,ncols_upper,*cols,*uj,**ui_ptr,*uj_ptr;
1584   PetscFreeSpaceList free_space=PETSC_NULL,current_space=PETSC_NULL;
1585   PetscBT            lnkbt;
1586   IS                 iperm;
1587 
1588   PetscFunctionBegin;
1589   if (A->rmap->n != A->cmap->n) SETERRQ2(PETSC_ERR_ARG_WRONG,"Must be square matrix, rows %D columns %D",A->rmap->n,A->cmap->n);
1590   /* check whether perm is the identity mapping */
1591   ierr = ISIdentity(perm,&perm_identity);CHKERRQ(ierr);
1592   ierr = ISInvertPermutation(perm,PETSC_DECIDE,&iperm);CHKERRQ(ierr);
1593   ierr = ISGetIndices(iperm,&riip);CHKERRQ(ierr);
1594   ierr = ISGetIndices(perm,&rip);CHKERRQ(ierr);
1595 
1596   /* initialization */
1597   ierr  = PetscMalloc((am+1)*sizeof(PetscInt),&ui);CHKERRQ(ierr);
1598   ui[0] = 0;
1599 
1600   /* jl: linked list for storing indices of the pivot rows
1601      il: il[i] points to the 1st nonzero entry of U(i,k:am-1) */
1602   ierr = PetscMalloc((3*am+1)*sizeof(PetscInt)+am*sizeof(PetscInt**),&jl);CHKERRQ(ierr);
1603   il     = jl + am;
1604   cols   = il + am;
1605   ui_ptr = (PetscInt**)(cols + am);
1606   for (i=0; i<am; i++){
1607     jl[i] = am; il[i] = 0;
1608   }
1609 
1610   /* create and initialize a linked list for storing column indices of the active row k */
1611   nlnk = am + 1;
1612   ierr = PetscLLCreate(am,am,nlnk,lnk,lnkbt);CHKERRQ(ierr);
1613 
1614   /* initial FreeSpace size is fill*(ai[am]+1) */
1615   ierr = PetscFreeSpaceGet((PetscInt)(fill*(ai[am]+1)),&free_space);CHKERRQ(ierr);
1616   current_space = free_space;
1617 
1618   for (k=0; k<am; k++){  /* for each active row k */
1619     /* initialize lnk by the column indices of row rip[k] of A */
1620     nzk   = 0;
1621     ncols = ai[rip[k]+1] - ai[rip[k]];
1622     if (!ncols) SETERRQ2(PETSC_ERR_MAT_CH_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",rip[k],k);
1623     ncols_upper = 0;
1624     for (j=0; j<ncols; j++){
1625       i = riip[*(aj + ai[rip[k]] + j)];
1626       if (i >= k){ /* only take upper triangular entry */
1627         cols[ncols_upper] = i;
1628         ncols_upper++;
1629       }
1630     }
1631     ierr = PetscLLAdd(ncols_upper,cols,am,nlnk,lnk,lnkbt);CHKERRQ(ierr);
1632     nzk += nlnk;
1633 
1634     /* update lnk by computing fill-in for each pivot row to be merged in */
1635     prow = jl[k]; /* 1st pivot row */
1636 
1637     while (prow < k){
1638       nextprow = jl[prow];
1639       /* merge prow into k-th row */
1640       jmin = il[prow] + 1;  /* index of the 2nd nzero entry in U(prow,k:am-1) */
1641       jmax = ui[prow+1];
1642       ncols = jmax-jmin;
1643       uj_ptr = ui_ptr[prow] + jmin - ui[prow]; /* points to the 2nd nzero entry in U(prow,k:am-1) */
1644       ierr = PetscLLAddSorted(ncols,uj_ptr,am,nlnk,lnk,lnkbt);CHKERRQ(ierr);
1645       nzk += nlnk;
1646 
1647       /* update il and jl for prow */
1648       if (jmin < jmax){
1649         il[prow] = jmin;
1650         j = *uj_ptr; jl[prow] = jl[j]; jl[j] = prow;
1651       }
1652       prow = nextprow;
1653     }
1654 
1655     /* if free space is not available, make more free space */
1656     if (current_space->local_remaining<nzk) {
1657       i = am - k + 1; /* num of unfactored rows */
1658       i = PetscMin(i*nzk, i*(i-1)); /* i*nzk, i*(i-1): estimated and max additional space needed */
1659       ierr = PetscFreeSpaceGet(i,&current_space);CHKERRQ(ierr);
1660       reallocs++;
1661     }
1662 
1663     /* copy data into free space, then initialize lnk */
1664     ierr = PetscLLClean(am,am,nzk,lnk,current_space->array,lnkbt);CHKERRQ(ierr);
1665 
1666     /* add the k-th row into il and jl */
1667     if (nzk-1 > 0){
1668       i = current_space->array[1]; /* col value of the first nonzero element in U(k, k+1:am-1) */
1669       jl[k] = jl[i]; jl[i] = k;
1670       il[k] = ui[k] + 1;
1671     }
1672     ui_ptr[k] = current_space->array;
1673     current_space->array           += nzk;
1674     current_space->local_used      += nzk;
1675     current_space->local_remaining -= nzk;
1676 
1677     ui[k+1] = ui[k] + nzk;
1678   }
1679 
1680 #if defined(PETSC_USE_INFO)
1681   if (ai[am] != 0) {
1682     PetscReal af = (PetscReal)(ui[am])/((PetscReal)ai[am]);
1683     ierr = PetscInfo3(A,"Reallocs %D Fill ratio:given %G needed %G\n",reallocs,fill,af);CHKERRQ(ierr);
1684     ierr = PetscInfo1(A,"Run with -pc_factor_fill %G or use \n",af);CHKERRQ(ierr);
1685     ierr = PetscInfo1(A,"PCFactorSetFill(pc,%G) for best performance.\n",af);CHKERRQ(ierr);
1686   } else {
1687      ierr = PetscInfo(A,"Empty matrix.\n");CHKERRQ(ierr);
1688   }
1689 #endif
1690 
1691   ierr = ISRestoreIndices(perm,&rip);CHKERRQ(ierr);
1692   ierr = ISRestoreIndices(iperm,&riip);CHKERRQ(ierr);
1693   ierr = PetscFree(jl);CHKERRQ(ierr);
1694 
1695   /* destroy list of free space and other temporary array(s) */
1696   ierr = PetscMalloc((ui[am]+1)*sizeof(PetscInt),&uj);CHKERRQ(ierr);
1697   ierr = PetscFreeSpaceContiguous(&free_space,uj);CHKERRQ(ierr);
1698   ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
1699 
1700   /* put together the new matrix in MATSEQSBAIJ format */
1701 
1702   b = (Mat_SeqSBAIJ*)(fact)->data;
1703   b->singlemalloc = PETSC_FALSE;
1704   b->free_a       = PETSC_TRUE;
1705   b->free_ij      = PETSC_TRUE;
1706   ierr = PetscMalloc((ui[am]+1)*sizeof(MatScalar),&b->a);CHKERRQ(ierr);
1707   b->j    = uj;
1708   b->i    = ui;
1709   b->diag = 0;
1710   b->ilen = 0;
1711   b->imax = 0;
1712   b->row  = perm;
1713   b->col  = perm;
1714   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
1715   ierr    = PetscObjectReference((PetscObject)perm);CHKERRQ(ierr);
1716   b->icol = iperm;
1717   b->pivotinblocks = PETSC_FALSE; /* need to get from MatFactorInfo */
1718   ierr    = PetscMalloc((am+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
1719   ierr    = PetscLogObjectMemory(fact,(ui[am]-am)*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr);
1720   b->maxnz = b->nz = ui[am];
1721 
1722   (fact)->info.factor_mallocs    = reallocs;
1723   (fact)->info.fill_ratio_given  = fill;
1724   if (ai[am] != 0) {
1725     (fact)->info.fill_ratio_needed = ((PetscReal)ui[am])/((PetscReal)ai[am]);
1726   } else {
1727     (fact)->info.fill_ratio_needed = 0.0;
1728   }
1729   (fact)->ops->choleskyfactornumeric = MatCholeskyFactorNumeric_SeqAIJ;
1730   PetscFunctionReturn(0);
1731 }
1732 
1733 #undef __FUNCT__
1734 #define __FUNCT__ "MatSolve_SeqAIJ_NaturalOrdering_iludt"
1735 PetscErrorCode MatSolve_SeqAIJ_NaturalOrdering_iludt(Mat A,Vec bb,Vec xx)
1736 {
1737   Mat_SeqAIJ        *a = (Mat_SeqAIJ*)A->data;
1738   PetscErrorCode    ierr;
1739   PetscInt          n = A->rmap->n;
1740   const PetscInt    *ai = a->i,*aj = a->j,*adiag = a->diag,*vi;
1741   PetscScalar       *x,sum;
1742   const PetscScalar *b;
1743   const MatScalar   *aa = a->a,*v;
1744   PetscInt          i,nz;
1745 
1746   PetscFunctionBegin;
1747   if (!n) PetscFunctionReturn(0);
1748 
1749   ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
1750   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
1751 
1752   /* forward solve the lower triangular */
1753   x[0] = b[0];
1754   v    = aa;
1755   vi   = aj;
1756   for (i=1; i<n; i++) {
1757     nz  = ai[i+1] - ai[i];
1758     sum = b[i];
1759     while (nz--) sum -= *v++ * x[*vi++];
1760     x[i] = sum;
1761   }
1762 
1763   /* backward solve the upper triangular */
1764   v   = aa + adiag[n] + 1;
1765   vi  = aj + adiag[n] + 1;
1766   for (i=n-1; i>=0; i--){
1767     nz  = adiag[i] - adiag[i+1] - 1;
1768     sum = x[i];
1769     while (nz--) sum -= *v++ * x[*vi++];
1770     x[i] = sum*aa[adiag[i]];
1771     v++; vi++;
1772   }
1773 
1774   ierr = PetscLogFlops(2.0*a->nz - A->cmap->n);CHKERRQ(ierr);
1775   ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
1776   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
1777   PetscFunctionReturn(0);
1778 }
1779 
1780 #undef __FUNCT__
1781 #define __FUNCT__ "MatSolve_SeqAIJ_iludt"
1782 PetscErrorCode MatSolve_SeqAIJ_iludt(Mat A,Vec bb,Vec xx)
1783 {
1784   Mat_SeqAIJ        *a = (Mat_SeqAIJ*)A->data;
1785   IS                iscol = a->col,isrow = a->row;
1786   PetscErrorCode    ierr;
1787   PetscInt          i,n=A->rmap->n,*vi,*ai = a->i,*aj = a->j,*adiag=a->diag;
1788   PetscInt          nz;
1789   const PetscInt    *rout,*cout,*r,*c;
1790   PetscScalar       *x,*tmp,*tmps;
1791   const PetscScalar *b;
1792   const MatScalar   *aa = a->a,*v;
1793 
1794   PetscFunctionBegin;
1795   if (!n) PetscFunctionReturn(0);
1796 
1797   ierr = VecGetArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
1798   ierr = VecGetArray(xx,&x);CHKERRQ(ierr);
1799   tmp  = a->solve_work;
1800 
1801   ierr = ISGetIndices(isrow,&rout);CHKERRQ(ierr); r = rout;
1802   ierr = ISGetIndices(iscol,&cout);CHKERRQ(ierr); c = cout + (n-1);
1803 
1804   /* forward solve the lower triangular */
1805   tmp[0] = b[*r++];
1806   tmps   = tmp;
1807   v      = aa;
1808   vi     = aj;
1809   for (i=1; i<n; i++) {
1810     nz  = ai[i+1] - ai[i];
1811     tmp[i] = b[*r++];
1812     SPARSEDENSEMDOT(tmp[i],tmps,v,vi,nz);
1813     v += nz; vi += nz;
1814   }
1815 
1816   /* backward solve the upper triangular */
1817   v   = aa + adiag[n] + 1;
1818   vi  = aj + adiag[n] + 1;
1819   for (i=n-1; i>=0; i--){
1820     nz  = adiag[i] - adiag[i+1] - 1;
1821     SPARSEDENSEMDOT(tmp[i],tmps,v,vi,nz);
1822     x[*c--] = tmp[i] = tmp[i]*aa[adiag[i]];
1823     v += nz+1; vi += nz+1;
1824   }
1825 
1826   ierr = ISRestoreIndices(isrow,&rout);CHKERRQ(ierr);
1827   ierr = ISRestoreIndices(iscol,&cout);CHKERRQ(ierr);
1828   ierr = VecRestoreArray(bb,(PetscScalar**)&b);CHKERRQ(ierr);
1829   ierr = VecRestoreArray(xx,&x);CHKERRQ(ierr);
1830   ierr = PetscLogFlops(2*a->nz - A->cmap->n);CHKERRQ(ierr);
1831   PetscFunctionReturn(0);
1832 }
1833 
1834 #undef __FUNCT__
1835 #define __FUNCT__ "MatILUDTFactor_SeqAIJ"
1836 PetscErrorCode MatILUDTFactor_SeqAIJ(Mat A,IS isrow,IS iscol,const MatFactorInfo *info,Mat *fact)
1837 {
1838   Mat                B = *fact;
1839   Mat_SeqAIJ         *a=(Mat_SeqAIJ*)A->data,*b;
1840   IS                 isicol;
1841   PetscErrorCode     ierr;
1842   const PetscInt     *r,*ic;
1843   PetscInt           i,n=A->rmap->n,*ai=a->i,*aj=a->j,*ajtmp,*adiag;
1844   PetscInt           *bi,*bj,*bdiag;
1845   PetscInt           row,nzi,nzi_bl,nzi_bu,*im,dtcount,nzi_al,nzi_au;
1846   PetscInt           nlnk,*lnk;
1847   PetscBT            lnkbt;
1848   PetscTruth         row_identity,icol_identity,both_identity;
1849   MatScalar          *aatmp,*pv,*batmp,*ba,*rtmp,*pc,multiplier,*vtmp;
1850   const PetscInt     *ics;
1851   PetscInt           j,nz,*pj,*bjtmp,k,ncut,*jtmp;
1852   PetscReal          dt=info->dt,shift=info->shiftinblocks;
1853   PetscInt           nnz_max;
1854   PetscTruth         missing;
1855 
1856   PetscFunctionBegin;
1857   /* ------- symbolic factorization, can be reused ---------*/
1858   ierr = MatMissingDiagonal(A,&missing,&i);CHKERRQ(ierr);
1859   if (missing) SETERRQ1(PETSC_ERR_ARG_WRONGSTATE,"Matrix is missing diagonal entry %D",i);
1860   adiag=a->diag;
1861 
1862   ierr = ISInvertPermutation(iscol,PETSC_DECIDE,&isicol);CHKERRQ(ierr);
1863 
1864   /* bdiag is location of diagonal in factor */
1865   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bdiag);CHKERRQ(ierr);
1866 
1867   /* allocate row pointers bi */
1868   ierr = PetscMalloc((n+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr);
1869 
1870   /* allocate bj and ba; max num of nonzero entries is (ai[n]+2*n*dtcount+2) */
1871   dtcount = (PetscInt)info->dtcount;
1872   if (dtcount > n/2) dtcount = n/2;
1873   nnz_max  = ai[n]+2*n*dtcount+2;
1874   ierr = PetscMalloc(nnz_max*sizeof(PetscInt),&bj);CHKERRQ(ierr);
1875   ierr = PetscMalloc(nnz_max*sizeof(MatScalar),&ba);CHKERRQ(ierr);
1876 
1877   /* put together the new matrix */
1878   ierr = MatSeqAIJSetPreallocation_SeqAIJ(B,MAT_SKIP_ALLOCATION,PETSC_NULL);CHKERRQ(ierr);
1879   ierr = PetscLogObjectParent(B,isicol);CHKERRQ(ierr);
1880   b    = (Mat_SeqAIJ*)(B)->data;
1881   b->free_a       = PETSC_TRUE;
1882   b->free_ij      = PETSC_TRUE;
1883   b->singlemalloc = PETSC_FALSE;
1884   b->a          = ba;
1885   b->j          = bj;
1886   b->i          = bi;
1887   b->diag       = bdiag;
1888   b->ilen       = 0;
1889   b->imax       = 0;
1890   b->row        = isrow;
1891   b->col        = iscol;
1892   ierr          = PetscObjectReference((PetscObject)isrow);CHKERRQ(ierr);
1893   ierr          = PetscObjectReference((PetscObject)iscol);CHKERRQ(ierr);
1894   b->icol       = isicol;
1895   ierr = PetscMalloc((n+1)*sizeof(PetscScalar),&b->solve_work);CHKERRQ(ierr);
1896 
1897   ierr = PetscLogObjectMemory(B,nnz_max*(sizeof(PetscInt)+sizeof(MatScalar)));CHKERRQ(ierr);
1898   b->maxnz = nnz_max;
1899 
1900   (B)->factor                = MAT_FACTOR_ILUDT;
1901   (B)->info.factor_mallocs   = 0;
1902   (B)->info.fill_ratio_given = ((PetscReal)nnz_max)/((PetscReal)ai[n]);
1903   CHKMEMQ;
1904   /* ------- end of symbolic factorization ---------*/
1905 
1906   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
1907   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
1908   ics  = ic;
1909 
1910   /* linked list for storing column indices of the active row */
1911   nlnk = n + 1;
1912   ierr = PetscLLCreate(n,n,nlnk,lnk,lnkbt);CHKERRQ(ierr);
1913 
1914   /* im: used by PetscLLAddSortedLU(); jtmp: working array for column indices of active row */
1915   ierr = PetscMalloc((2*n+1)*sizeof(PetscInt),&im);CHKERRQ(ierr);
1916   jtmp = im + n;
1917   /* rtmp, vtmp: working arrays for sparse and contiguous row entries of active row */
1918   ierr = PetscMalloc((2*n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
1919   vtmp = rtmp + n;
1920 
1921   bi[0]    = 0;
1922   bdiag[0] = nnz_max-1; /* location of diagonal in factor B */
1923   for (i=0; i<n; i++) {
1924     /* copy initial fill into linked list */
1925     nzi = 0; /* nonzeros for active row i */
1926     nzi = ai[r[i]+1] - ai[r[i]];
1927     if (!nzi) SETERRQ2(PETSC_ERR_MAT_LU_ZRPVT,"Empty row in matrix: row in original ordering %D in permuted ordering %D",r[i],i);
1928     nzi_al = adiag[r[i]] - ai[r[i]];
1929     nzi_au = ai[r[i]+1] - adiag[r[i]] -1;
1930     ajtmp = aj + ai[r[i]];
1931     ierr = PetscLLAddPerm(nzi,ajtmp,ic,n,nlnk,lnk,lnkbt);CHKERRQ(ierr);
1932 
1933     /* load in initial (unfactored row) */
1934     ierr = PetscMemzero(rtmp,(n+1)*sizeof(PetscScalar));CHKERRQ(ierr);
1935     aatmp = a->a + ai[r[i]];
1936     for (j=0; j<nzi; j++) {
1937       rtmp[ics[*ajtmp++]] = *aatmp++;
1938     }
1939 
1940     /* add pivot rows into linked list */
1941     row = lnk[n];
1942     while (row < i) {
1943       nzi_bl = bi[row+1] - bi[row] + 1;
1944       bjtmp = bj + bdiag[row+1]+1; /* points to 1st column next to the diagonal in U */
1945       ierr  = PetscLLAddSortedLU(bjtmp,row,nlnk,lnk,lnkbt,i,nzi_bl,im);CHKERRQ(ierr);
1946       nzi  += nlnk;
1947       row   = lnk[row];
1948     }
1949 
1950     /* copy data from lnk into jtmp, then initialize lnk */
1951     ierr = PetscLLClean(n,n,nzi,lnk,jtmp,lnkbt);CHKERRQ(ierr);
1952 
1953     /* numerical factorization */
1954     bjtmp = jtmp;
1955     row   = *bjtmp++; /* 1st pivot row */
1956     while  (row < i) {
1957       pc         = rtmp + row;
1958       pv         = ba + bdiag[row]; /* 1./(diag of the pivot row) */
1959       multiplier = (*pc) * (*pv);
1960       *pc        = multiplier;
1961       if (PetscAbsScalar(*pc) > dt){ /* apply tolerance dropping rule */
1962         pj         = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */
1963         pv         = ba + bdiag[row+1] + 1;
1964         /* if (multiplier < -1.0 or multiplier >1.0) printf("row/prow %d, %d, multiplier %g\n",i,row,multiplier); */
1965         nz         = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */
1966         for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++);
1967         ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr);
1968       }
1969       row = *bjtmp++;
1970     }
1971 
1972     /* copy sparse rtmp into contiguous vtmp; separate L and U part */
1973     nzi_bl = 0; j = 0;
1974     while (jtmp[j] < i){ /* Note: jtmp is sorted */
1975       vtmp[j] = rtmp[jtmp[j]];
1976       nzi_bl++; j++;
1977     }
1978     nzi_bu = nzi - nzi_bl -1;
1979     while (j < nzi){
1980       vtmp[j] = rtmp[jtmp[j]];
1981       j++;
1982     }
1983 
1984     bjtmp = bj + bi[i];
1985     batmp = ba + bi[i];
1986     /* apply level dropping rule to L part */
1987     ncut = nzi_al + dtcount;
1988     if (ncut < nzi_bl){
1989       ierr = PetscSortSplit(ncut,nzi_bl,vtmp,jtmp);CHKERRQ(ierr);
1990       ierr = PetscSortIntWithScalarArray(ncut,jtmp,vtmp);CHKERRQ(ierr);
1991     } else {
1992       ncut = nzi_bl;
1993     }
1994     for (j=0; j<ncut; j++){
1995       bjtmp[j] = jtmp[j];
1996       batmp[j] = vtmp[j];
1997       /* printf(" (%d,%g),",bjtmp[j],batmp[j]); */
1998     }
1999     bi[i+1] = bi[i] + ncut;
2000     nzi = ncut + 1;
2001 
2002     /* apply level dropping rule to U part */
2003     ncut = nzi_au + dtcount;
2004     if (ncut < nzi_bu){
2005       ierr = PetscSortSplit(ncut,nzi_bu,vtmp+nzi_bl+1,jtmp+nzi_bl+1);CHKERRQ(ierr);
2006       ierr = PetscSortIntWithScalarArray(ncut,jtmp+nzi_bl+1,vtmp+nzi_bl+1);CHKERRQ(ierr);
2007     } else {
2008       ncut = nzi_bu;
2009     }
2010     nzi += ncut;
2011 
2012     /* mark bdiagonal */
2013     bdiag[i+1]   = bdiag[i] - (ncut + 1);
2014     bjtmp = bj + bdiag[i];
2015     batmp = ba + bdiag[i];
2016     *bjtmp = i;
2017     *batmp = rtmp[i];
2018     if (*batmp == 0.0) *batmp = dt+shift;
2019     *batmp = 1.0/(*batmp); /* invert diagonal entries for simplier triangular solves */
2020     /* printf(" (%d,%g),",*bjtmp,*batmp); */
2021 
2022     bjtmp = bj + bdiag[i+1]+1;
2023     batmp = ba + bdiag[i+1]+1;
2024     for (k=0; k<ncut; k++){
2025       bjtmp[k] = jtmp[nzi_bl+1+k];
2026       batmp[k] = vtmp[nzi_bl+1+k];
2027       /* printf(" (%d,%g),",bjtmp[k],batmp[k]); */
2028     }
2029     /* printf("\n"); */
2030 
2031     im[i]   = nzi; /* used by PetscLLAddSortedLU() */
2032     /*
2033     printf("row %d: bi %d, bdiag %d\n",i,bi[i],bdiag[i]);
2034     printf(" ----------------------------\n");
2035     */
2036   } /* for (i=0; i<n; i++) */
2037   /* printf("end of L %d, beginning of U %d\n",bi[n],bdiag[n]); */
2038   if (bi[n] >= bdiag[n]) SETERRQ2(PETSC_ERR_ARG_SIZ,"end of L array %d cannot >= the beginning of U array %d",bi[n],bdiag[n]);
2039 
2040   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
2041   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
2042 
2043   ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
2044   ierr = PetscFree(im);CHKERRQ(ierr);
2045   ierr = PetscFree(rtmp);CHKERRQ(ierr);
2046 
2047   ierr = PetscLogFlops(B->cmap->n);CHKERRQ(ierr);
2048   b->maxnz = b->nz = bi[n] + bdiag[0] - bdiag[n];
2049 
2050   ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr);
2051   ierr = ISIdentity(isicol,&icol_identity);CHKERRQ(ierr);
2052   both_identity = (PetscTruth) (row_identity && icol_identity);
2053   if (row_identity && icol_identity) {
2054     B->ops->solve = MatSolve_SeqAIJ_NaturalOrdering_iludt;
2055   } else {
2056     B->ops->solve = MatSolve_SeqAIJ_iludt;
2057   }
2058 
2059   B->ops->lufactorsymbolic  = MatILUDTFactorSymbolic_SeqAIJ;
2060   B->ops->lufactornumeric   = MatILUDTFactorNumeric_SeqAIJ;
2061   B->ops->solveadd          = 0;
2062   B->ops->solvetranspose    = 0;
2063   B->ops->solvetransposeadd = 0;
2064   B->ops->matsolve          = 0;
2065   B->assembled              = PETSC_TRUE;
2066   B->preallocated           = PETSC_TRUE;
2067   PetscFunctionReturn(0);
2068 }
2069 
2070 /* a wraper of MatILUDTFactor_SeqAIJ() */
2071 #undef __FUNCT__
2072 #define __FUNCT__ "MatILUDTFactorSymbolic_SeqAIJ"
2073 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorSymbolic_SeqAIJ(Mat fact,Mat A,IS row,IS col,const MatFactorInfo *info)
2074 {
2075   PetscErrorCode     ierr;
2076 
2077   PetscFunctionBegin;
2078   /* printf("MatILUDTFactorSymbolic_SeqAIJ is called...\n"); */
2079   ierr = MatILUDTFactor_SeqAIJ(A,row,col,info,&fact);CHKERRQ(ierr);
2080 
2081   fact->ops->lufactornumeric = MatILUDTFactorNumeric_SeqAIJ;
2082   PetscFunctionReturn(0);
2083 }
2084 
2085 /*
2086    same as MatLUFactorNumeric_SeqAIJ(), except using contiguous array matrix factors
2087    - intend to replace existing MatLUFactorNumeric_SeqAIJ()
2088 */
2089 #undef __FUNCT__
2090 #define __FUNCT__ "MatILUDTFactorNumeric_SeqAIJ"
2091 PetscErrorCode PETSCMAT_DLLEXPORT MatILUDTFactorNumeric_SeqAIJ(Mat fact,Mat A,const MatFactorInfo *info)
2092 {
2093   Mat            C=fact;
2094   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ *)C->data;
2095   IS             isrow = b->row,isicol = b->icol;
2096   PetscErrorCode ierr;
2097   const PetscInt *r,*ic,*ics;
2098   PetscInt       i,j,k,n=A->rmap->n,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
2099   PetscInt       *ajtmp,*bjtmp,nz,nzl,nzu,row,*bdiag = b->diag,*pj;
2100   MatScalar      *rtmp,*pc,multiplier,*v,*pv,*aa=a->a;
2101   PetscReal      dt=info->dt,shift=info->shiftinblocks;
2102   PetscTruth     row_identity, col_identity;
2103 
2104   PetscFunctionBegin;
2105   /* printf("MatILUDTFactorNumeric_SeqAIJ is called...\n"); */
2106   ierr = ISGetIndices(isrow,&r);CHKERRQ(ierr);
2107   ierr = ISGetIndices(isicol,&ic);CHKERRQ(ierr);
2108   ierr = PetscMalloc((n+1)*sizeof(MatScalar),&rtmp);CHKERRQ(ierr);
2109   ics  = ic;
2110 
2111   for (i=0; i<n; i++){
2112     /* initialize rtmp array */
2113     nzl   = bi[i+1] - bi[i];       /* num of nozeros in L(i,:) */
2114     bjtmp = bj + bi[i];
2115     for  (j=0; j<nzl; j++) rtmp[*bjtmp++] = 0.0;
2116     rtmp[i] = 0.0;
2117     nzu   = bdiag[i] - bdiag[i+1]; /* num of nozeros in U(i,:) */
2118     bjtmp = bj + bdiag[i+1] + 1;
2119     for  (j=0; j<nzu; j++) rtmp[*bjtmp++] = 0.0;
2120 
2121     /* load in initial unfactored row of A */
2122     /* printf("row %d\n",i); */
2123     nz    = ai[r[i]+1] - ai[r[i]];
2124     ajtmp = aj + ai[r[i]];
2125     v     = aa + ai[r[i]];
2126     for (j=0; j<nz; j++) {
2127       rtmp[ics[*ajtmp++]] = v[j];
2128       /* printf(" (%d,%g),",ics[ajtmp[j]],rtmp[ics[ajtmp[j]]]); */
2129     }
2130     /* printf("\n"); */
2131 
2132     /* numerical factorization */
2133     bjtmp = bj + bi[i]; /* point to 1st entry of L(i,:) */
2134     nzl   = bi[i+1] - bi[i]; /* num of entries in L(i,:) */
2135     k = 0;
2136     while (k < nzl){
2137       row   = *bjtmp++;
2138       /* printf("  prow %d\n",row); */
2139       pc         = rtmp + row;
2140       pv         = b->a + bdiag[row]; /* 1./(diag of the pivot row) */
2141       multiplier = (*pc) * (*pv);
2142       *pc        = multiplier;
2143       if (PetscAbsScalar(multiplier) > dt){
2144         pj         = bj + bdiag[row+1] + 1; /* point to 1st entry of U(row,:) */
2145         pv         = b->a + bdiag[row+1] + 1;
2146         nz         = bdiag[row] - bdiag[row+1] - 1; /* num of entries in U(row,:), excluding diagonal */
2147         for (j=0; j<nz; j++) rtmp[*pj++] -= multiplier * (*pv++);
2148         /* ierr = PetscLogFlops(2.0*nz);CHKERRQ(ierr); */
2149       }
2150       k++;
2151     }
2152 
2153     /* finished row so stick it into b->a */
2154     /* L-part */
2155     pv = b->a + bi[i] ;
2156     pj = bj + bi[i] ;
2157     nzl = bi[i+1] - bi[i];
2158     for (j=0; j<nzl; j++) {
2159       pv[j] = rtmp[pj[j]];
2160       /* printf(" (%d,%g),",pj[j],pv[j]); */
2161     }
2162 
2163     /* diagonal: invert diagonal entries for simplier triangular solves */
2164     if (rtmp[i] == 0.0) rtmp[i] = dt+shift;
2165     b->a[bdiag[i]] = 1.0/rtmp[i];
2166     /* printf(" (%d,%g),",i,b->a[bdiag[i]]); */
2167 
2168     /* U-part */
2169     pv = b->a + bdiag[i+1] + 1;
2170     pj = bj + bdiag[i+1] + 1;
2171     nzu = bdiag[i] - bdiag[i+1] - 1;
2172     for (j=0; j<nzu; j++) {
2173       pv[j] = rtmp[pj[j]];
2174       /* printf(" (%d,%g),",pj[j],pv[j]); */
2175     }
2176     /* printf("\n"); */
2177   }
2178 
2179   ierr = PetscFree(rtmp);CHKERRQ(ierr);
2180   ierr = ISRestoreIndices(isicol,&ic);CHKERRQ(ierr);
2181   ierr = ISRestoreIndices(isrow,&r);CHKERRQ(ierr);
2182 
2183   ierr = ISIdentity(isrow,&row_identity);CHKERRQ(ierr);
2184   ierr = ISIdentity(isicol,&col_identity);CHKERRQ(ierr);
2185   if (row_identity && col_identity) {
2186     C->ops->solve   = MatSolve_SeqAIJ_NaturalOrdering_iludt;
2187   } else {
2188     C->ops->solve   = MatSolve_SeqAIJ_iludt;
2189   }
2190   C->ops->solveadd           = 0;
2191   C->ops->solvetranspose     = 0;
2192   C->ops->solvetransposeadd  = 0;
2193   C->ops->matsolve           = 0;
2194   C->assembled    = PETSC_TRUE;
2195   C->preallocated = PETSC_TRUE;
2196   ierr = PetscLogFlops(C->cmap->n);CHKERRQ(ierr);
2197   PetscFunctionReturn(0);
2198 }
2199