Actual source code: inode.c

  1: #define PETSCMAT_DLL

  3: /*
  4:   This file provides high performance routines for the Inode format (compressed sparse row)
  5:   by taking advantage of rows with identical nonzero structure (I-nodes).
  6: */
 7:  #include src/mat/impls/csr/inode/inode.h

 11: static PetscErrorCode Mat_CreateColInode(Mat A,PetscInt* size,PetscInt ** ns)
 12: {
 13:   Mat_inode      *a = (Mat_inode*)A->data;
 15:   PetscInt       i,count,m,n,min_mn,*ns_row,*ns_col;

 18:   n      = A->n;
 19:   m      = A->m;
 20:   ns_row = a->inode.size;
 21: 
 22:   min_mn = (m < n) ? m : n;
 23:   if (!ns) {
 24:     for (count=0,i=0; count<min_mn; count+=ns_row[i],i++);
 25:     for(; count+1 < n; count++,i++);
 26:     if (count < n)  {
 27:       i++;
 28:     }
 29:     *size = i;
 30:     return(0);
 31:   }
 32:   PetscMalloc((n+1)*sizeof(PetscInt),&ns_col);
 33: 
 34:   /* Use the same row structure wherever feasible. */
 35:   for (count=0,i=0; count<min_mn; count+=ns_row[i],i++) {
 36:     ns_col[i] = ns_row[i];
 37:   }

 39:   /* if m < n; pad up the remainder with inode_limit */
 40:   for(; count+1 < n; count++,i++) {
 41:     ns_col[i] = 1;
 42:   }
 43:   /* The last node is the odd ball. padd it up with the remaining rows; */
 44:   if (count < n)  {
 45:     ns_col[i] = n - count;
 46:     i++;
 47:   } else if (count > n) {
 48:     /* Adjust for the over estimation */
 49:     ns_col[i-1] += n - count;
 50:   }
 51:   *size = i;
 52:   *ns   = ns_col;
 53:   return(0);
 54: }


 57: /*
 58:       This builds symmetric version of nonzero structure,
 59: */
 62: static PetscErrorCode MatGetRowIJ_Inode_Symmetric(Mat A,PetscInt *iia[],PetscInt *jja[],PetscInt ishift,PetscInt oshift)
 63: {
 64:   Mat_inode      *a = (Mat_inode*)A->data;
 66:   PetscInt       *work,*ia,*ja,*j,nz,nslim_row,nslim_col,m,row,col,*jmax,n;
 67:   PetscInt       *tns,*tvc,*ns_row = a->inode.size,*ns_col,nsz,i1,i2,*ai= a->i,*aj = a->j;

 70:   nslim_row = a->inode.node_count;
 71:   m         = A->m;
 72:   n         = A->n;
 73:   if (m != n) SETERRQ(PETSC_ERR_SUP,"MatGetRowIJ_Inode_Symmetric: Matrix should be square");
 74: 
 75:   /* Use the row_inode as column_inode */
 76:   nslim_col = nslim_row;
 77:   ns_col    = ns_row;

 79:   /* allocate space for reformated inode structure */
 80:   PetscMalloc((nslim_col+1)*sizeof(PetscInt),&tns);
 81:   PetscMalloc((n+1)*sizeof(PetscInt),&tvc);
 82:   for (i1=0,tns[0]=0; i1<nslim_col; ++i1) tns[i1+1] = tns[i1]+ ns_row[i1];

 84:   for (i1=0,col=0; i1<nslim_col; ++i1){
 85:     nsz = ns_col[i1];
 86:     for (i2=0; i2<nsz; ++i2,++col)
 87:       tvc[col] = i1;
 88:   }
 89:   /* allocate space for row pointers */
 90:   PetscMalloc((nslim_row+1)*sizeof(PetscInt),&ia);
 91:   *iia = ia;
 92:   PetscMemzero(ia,(nslim_row+1)*sizeof(PetscInt));
 93:   PetscMalloc((nslim_row+1)*sizeof(PetscInt),&work);

 95:   /* determine the number of columns in each row */
 96:   ia[0] = oshift;
 97:   for (i1=0,row=0 ; i1<nslim_row; row+=ns_row[i1],i1++) {

 99:     j    = aj + ai[row] + ishift;
100:     jmax = aj + ai[row+1] + ishift;
101:     i2   = 0;
102:     col  = *j++ + ishift;
103:     i2   = tvc[col];
104:     while (i2<i1 && j<jmax) { /* 1.[-xx-d-xx--] 2.[-xx-------],off-diagonal elemets */
105:       ia[i1+1]++;
106:       ia[i2+1]++;
107:       i2++;                     /* Start col of next node */
108:       while(((col=*j+ishift)<tns[i2]) && (j<jmax)) ++j;
109:       i2 = tvc[col];
110:     }
111:     if(i2 == i1) ia[i2+1]++;    /* now the diagonal element */
112:   }

114:   /* shift ia[i] to point to next row */
115:   for (i1=1; i1<nslim_row+1; i1++) {
116:     row        = ia[i1-1];
117:     ia[i1]    += row;
118:     work[i1-1] = row - oshift;
119:   }

121:   /* allocate space for column pointers */
122:   nz   = ia[nslim_row] + (!ishift);
123:   PetscMalloc(nz*sizeof(PetscInt),&ja);
124:   *jja = ja;

126:  /* loop over lower triangular part putting into ja */
127:   for (i1=0,row=0; i1<nslim_row; row += ns_row[i1],i1++) {
128:     j    = aj + ai[row] + ishift;
129:     jmax = aj + ai[row+1] + ishift;
130:     i2   = 0;                     /* Col inode index */
131:     col  = *j++ + ishift;
132:     i2   = tvc[col];
133:     while (i2<i1 && j<jmax) {
134:       ja[work[i2]++] = i1 + oshift;
135:       ja[work[i1]++] = i2 + oshift;
136:       ++i2;
137:       while(((col=*j+ishift)< tns[i2])&&(j<jmax)) ++j; /* Skip rest col indices in this node */
138:       i2 = tvc[col];
139:     }
140:     if (i2 == i1) ja[work[i1]++] = i2 + oshift;

142:   }
143:   PetscFree(work);
144:   PetscFree(tns);
145:   PetscFree(tvc);
146:   return(0);
147: }

149: /*
150:       This builds nonsymmetric version of nonzero structure,
151: */
154: static PetscErrorCode MatGetRowIJ_Inode_Nonsymmetric(Mat A,PetscInt *iia[],PetscInt *jja[],PetscInt ishift,PetscInt oshift)
155: {
156:   Mat_inode      *a = (Mat_inode*)A->data;
158:   PetscInt       *work,*ia,*ja,*j,nz,nslim_row,n,row,col,*ns_col,nslim_col;
159:   PetscInt       *tns,*tvc,*ns_row = a->inode.size,nsz,i1,i2,*ai= a->i,*aj = a->j;

162:   nslim_row = a->inode.node_count;
163:   n         = A->n;

165:   /* Create The column_inode for this matrix */
166:   Mat_CreateColInode(A,&nslim_col,&ns_col);
167: 
168:   /* allocate space for reformated column_inode structure */
169:   PetscMalloc((nslim_col +1)*sizeof(PetscInt),&tns);
170:   PetscMalloc((n +1)*sizeof(PetscInt),&tvc);
171:   for (i1=0,tns[0]=0; i1<nslim_col; ++i1) tns[i1+1] = tns[i1] + ns_col[i1];

173:   for (i1=0,col=0; i1<nslim_col; ++i1){
174:     nsz = ns_col[i1];
175:     for (i2=0; i2<nsz; ++i2,++col)
176:       tvc[col] = i1;
177:   }
178:   /* allocate space for row pointers */
179:   PetscMalloc((nslim_row+1)*sizeof(PetscInt),&ia);
180:   *iia = ia;
181:   PetscMemzero(ia,(nslim_row+1)*sizeof(PetscInt));
182:   PetscMalloc((nslim_row+1)*sizeof(PetscInt),&work);

184:   /* determine the number of columns in each row */
185:   ia[0] = oshift;
186:   for (i1=0,row=0; i1<nslim_row; row+=ns_row[i1],i1++) {
187:     j   = aj + ai[row] + ishift;
188:     col = *j++ + ishift;
189:     i2  = tvc[col];
190:     nz  = ai[row+1] - ai[row];
191:     while (nz-- > 0) {           /* off-diagonal elemets */
192:       ia[i1+1]++;
193:       i2++;                     /* Start col of next node */
194:       while (((col = *j++ + ishift) < tns[i2]) && nz > 0) {nz--;}
195:       i2 = tvc[col];
196:     }
197:   }

199:   /* shift ia[i] to point to next row */
200:   for (i1=1; i1<nslim_row+1; i1++) {
201:     row        = ia[i1-1];
202:     ia[i1]    += row;
203:     work[i1-1] = row - oshift;
204:   }

206:   /* allocate space for column pointers */
207:   nz   = ia[nslim_row] + (!ishift);
208:   PetscMalloc(nz*sizeof(PetscInt),&ja);
209:   *jja = ja;

211:  /* loop over matrix putting into ja */
212:   for (i1=0,row=0; i1<nslim_row; row+=ns_row[i1],i1++) {
213:     j   = aj + ai[row] + ishift;
214:     i2  = 0;                     /* Col inode index */
215:     col = *j++ + ishift;
216:     i2  = tvc[col];
217:     nz  = ai[row+1] - ai[row];
218:     while (nz-- > 0) {
219:       ja[work[i1]++] = i2 + oshift;
220:       ++i2;
221:       while(((col = *j++ + ishift) < tns[i2]) && nz > 0) {nz--;}
222:       i2 = tvc[col];
223:     }
224:   }
225:   PetscFree(ns_col);
226:   PetscFree(work);
227:   PetscFree(tns);
228:   PetscFree(tvc);
229:   return(0);
230: }

234: static PetscErrorCode MatGetRowIJ_Inode(Mat A,PetscInt oshift,PetscTruth symmetric,PetscInt *n,PetscInt *ia[],PetscInt *ja[],PetscTruth *done)
235: {
236:   Mat_inode      *a = (Mat_inode*)A->data;

240:   *n     = a->inode.node_count;
241:   if (!ia) return(0);

243:   if (symmetric) {
244:     MatGetRowIJ_Inode_Symmetric(A,ia,ja,0,oshift);
245:   } else {
246:     MatGetRowIJ_Inode_Nonsymmetric(A,ia,ja,0,oshift);
247:   }
248:   return(0);
249: }

253: static PetscErrorCode MatRestoreRowIJ_Inode(Mat A,PetscInt oshift,PetscTruth symmetric,PetscInt *n,PetscInt *ia[],PetscInt *ja[],PetscTruth *done)
254: {

258:   if (!ia) return(0);
259:   PetscFree(*ia);
260:   PetscFree(*ja);
261:   return(0);
262: }

264: /* ----------------------------------------------------------- */

268: static PetscErrorCode MatGetColumnIJ_Inode_Nonsymmetric(Mat A,PetscInt *iia[],PetscInt *jja[],PetscInt ishift,PetscInt oshift)
269: {
270:   Mat_inode      *a = (Mat_inode*)A->data;
272:   PetscInt       *work,*ia,*ja,*j,nz,nslim_row, n,row,col,*ns_col,nslim_col;
273:   PetscInt       *tns,*tvc,*ns_row = a->inode.size,nsz,i1,i2,*ai= a->i,*aj = a->j;

276:   nslim_row = a->inode.node_count;
277:   n         = A->n;

279:   /* Create The column_inode for this matrix */
280:   Mat_CreateColInode(A,&nslim_col,&ns_col);
281: 
282:   /* allocate space for reformated column_inode structure */
283:   PetscMalloc((nslim_col + 1)*sizeof(PetscInt),&tns);
284:   PetscMalloc((n + 1)*sizeof(PetscInt),&tvc);
285:   for (i1=0,tns[0]=0; i1<nslim_col; ++i1) tns[i1+1] = tns[i1] + ns_col[i1];

287:   for (i1=0,col=0; i1<nslim_col; ++i1){
288:     nsz = ns_col[i1];
289:     for (i2=0; i2<nsz; ++i2,++col)
290:       tvc[col] = i1;
291:   }
292:   /* allocate space for column pointers */
293:   PetscMalloc((nslim_col+1)*sizeof(PetscInt),&ia);
294:   *iia = ia;
295:   PetscMemzero(ia,(nslim_col+1)*sizeof(PetscInt));
296:   PetscMalloc((nslim_col+1)*sizeof(PetscInt),&work);

298:   /* determine the number of columns in each row */
299:   ia[0] = oshift;
300:   for (i1=0,row=0; i1<nslim_row; row+=ns_row[i1],i1++) {
301:     j   = aj + ai[row] + ishift;
302:     col = *j++ + ishift;
303:     i2  = tvc[col];
304:     nz  = ai[row+1] - ai[row];
305:     while (nz-- > 0) {           /* off-diagonal elemets */
306:       /* ia[i1+1]++; */
307:       ia[i2+1]++;
308:       i2++;
309:       while (((col = *j++ + ishift) < tns[i2]) && nz > 0) {nz--;}
310:       i2 = tvc[col];
311:     }
312:   }

314:   /* shift ia[i] to point to next col */
315:   for (i1=1; i1<nslim_col+1; i1++) {
316:     col        = ia[i1-1];
317:     ia[i1]    += col;
318:     work[i1-1] = col - oshift;
319:   }

321:   /* allocate space for column pointers */
322:   nz   = ia[nslim_col] + (!ishift);
323:   PetscMalloc(nz*sizeof(PetscInt),&ja);
324:   *jja = ja;

326:  /* loop over matrix putting into ja */
327:   for (i1=0,row=0; i1<nslim_row; row+=ns_row[i1],i1++) {
328:     j   = aj + ai[row] + ishift;
329:     i2  = 0;                     /* Col inode index */
330:     col = *j++ + ishift;
331:     i2  = tvc[col];
332:     nz  = ai[row+1] - ai[row];
333:     while (nz-- > 0) {
334:       /* ja[work[i1]++] = i2 + oshift; */
335:       ja[work[i2]++] = i1 + oshift;
336:       i2++;
337:       while(((col = *j++ + ishift) < tns[i2]) && nz > 0) {nz--;}
338:       i2 = tvc[col];
339:     }
340:   }
341:   PetscFree(ns_col);
342:   PetscFree(work);
343:   PetscFree(tns);
344:   PetscFree(tvc);
345:   return(0);
346: }

350: static PetscErrorCode MatGetColumnIJ_Inode(Mat A,PetscInt oshift,PetscTruth symmetric,PetscInt *n,PetscInt *ia[],PetscInt *ja[],PetscTruth *done)
351: {

355:   Mat_CreateColInode(A,n,PETSC_NULL);
356:   if (!ia) return(0);

358:   if (symmetric) {
359:     /* Since the indices are symmetric it does'nt matter */
360:     MatGetRowIJ_Inode_Symmetric(A,ia,ja,0,oshift);
361:   } else {
362:     MatGetColumnIJ_Inode_Nonsymmetric(A,ia,ja,0,oshift);
363:   }
364:   return(0);
365: }

369: static PetscErrorCode MatRestoreColumnIJ_Inode(Mat A,PetscInt oshift,PetscTruth symmetric,PetscInt *n,PetscInt *ia[],PetscInt *ja[],PetscTruth *done)
370: {

374:   if (!ia) return(0);
375:   PetscFree(*ia);
376:   PetscFree(*ja);
377:   return(0);
378: }

380: /* ----------------------------------------------------------- */

384: static PetscErrorCode MatMult_Inode(Mat A,Vec xx,Vec yy)
385: {
386:   Mat_inode      *a = (Mat_inode*)A->data;
387:   PetscScalar    sum1,sum2,sum3,sum4,sum5,tmp0,tmp1;
388:   PetscScalar    *v1,*v2,*v3,*v4,*v5,*x,*y;
390:   PetscInt       *idx,i1,i2,n,i,row,node_max,*ns,*ii,nsz,sz;
391: 
392: #if defined(PETSC_HAVE_PRAGMA_DISJOINT)
393: #pragma disjoint(*x,*y,*v1,*v2,*v3,*v4,*v5)
394: #endif

397:   if (!a->inode.size) SETERRQ(PETSC_ERR_COR,"Missing Inode Structure");
398:   node_max = a->inode.node_count;
399:   ns       = a->inode.size;     /* Node Size array */
400:   VecGetArray(xx,&x);
401:   VecGetArray(yy,&y);
402:   idx  = a->j;
403:   v1   = a->a;
404:   ii   = a->i;

406:   for (i = 0,row = 0; i< node_max; ++i){
407:     nsz  = ns[i];
408:     n    = ii[1] - ii[0];
409:     ii  += nsz;
410:     sz   = n;                   /* No of non zeros in this row */
411:                                 /* Switch on the size of Node */
412:     switch (nsz){               /* Each loop in 'case' is unrolled */
413:     case 1 :
414:       sum1  = 0;
415: 
416:       for(n = 0; n< sz-1; n+=2) {
417:         i1   = idx[0];          /* The instructions are ordered to */
418:         i2   = idx[1];          /* make the compiler's job easy */
419:         idx += 2;
420:         tmp0 = x[i1];
421:         tmp1 = x[i2];
422:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
423:        }
424: 
425:       if (n == sz-1){          /* Take care of the last nonzero  */
426:         tmp0  = x[*idx++];
427:         sum1 += *v1++ * tmp0;
428:       }
429:       y[row++]=sum1;
430:       break;
431:     case 2:
432:       sum1  = 0;
433:       sum2  = 0;
434:       v2    = v1 + n;
435: 
436:       for (n = 0; n< sz-1; n+=2) {
437:         i1   = idx[0];
438:         i2   = idx[1];
439:         idx += 2;
440:         tmp0 = x[i1];
441:         tmp1 = x[i2];
442:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
443:         sum2 += v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
444:       }
445:       if (n == sz-1){
446:         tmp0  = x[*idx++];
447:         sum1 += *v1++ * tmp0;
448:         sum2 += *v2++ * tmp0;
449:       }
450:       y[row++]=sum1;
451:       y[row++]=sum2;
452:       v1      =v2;              /* Since the next block to be processed starts there*/
453:       idx    +=sz;
454:       break;
455:     case 3:
456:       sum1  = 0;
457:       sum2  = 0;
458:       sum3  = 0;
459:       v2    = v1 + n;
460:       v3    = v2 + n;
461: 
462:       for (n = 0; n< sz-1; n+=2) {
463:         i1   = idx[0];
464:         i2   = idx[1];
465:         idx += 2;
466:         tmp0 = x[i1];
467:         tmp1 = x[i2];
468:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
469:         sum2 += v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
470:         sum3 += v3[0] * tmp0 + v3[1] * tmp1; v3 += 2;
471:       }
472:       if (n == sz-1){
473:         tmp0  = x[*idx++];
474:         sum1 += *v1++ * tmp0;
475:         sum2 += *v2++ * tmp0;
476:         sum3 += *v3++ * tmp0;
477:       }
478:       y[row++]=sum1;
479:       y[row++]=sum2;
480:       y[row++]=sum3;
481:       v1       =v3;             /* Since the next block to be processed starts there*/
482:       idx     +=2*sz;
483:       break;
484:     case 4:
485:       sum1  = 0;
486:       sum2  = 0;
487:       sum3  = 0;
488:       sum4  = 0;
489:       v2    = v1 + n;
490:       v3    = v2 + n;
491:       v4    = v3 + n;
492: 
493:       for (n = 0; n< sz-1; n+=2) {
494:         i1   = idx[0];
495:         i2   = idx[1];
496:         idx += 2;
497:         tmp0 = x[i1];
498:         tmp1 = x[i2];
499:         sum1 += v1[0] * tmp0 + v1[1] *tmp1; v1 += 2;
500:         sum2 += v2[0] * tmp0 + v2[1] *tmp1; v2 += 2;
501:         sum3 += v3[0] * tmp0 + v3[1] *tmp1; v3 += 2;
502:         sum4 += v4[0] * tmp0 + v4[1] *tmp1; v4 += 2;
503:       }
504:       if (n == sz-1){
505:         tmp0  = x[*idx++];
506:         sum1 += *v1++ * tmp0;
507:         sum2 += *v2++ * tmp0;
508:         sum3 += *v3++ * tmp0;
509:         sum4 += *v4++ * tmp0;
510:       }
511:       y[row++]=sum1;
512:       y[row++]=sum2;
513:       y[row++]=sum3;
514:       y[row++]=sum4;
515:       v1      =v4;              /* Since the next block to be processed starts there*/
516:       idx    +=3*sz;
517:       break;
518:     case 5:
519:       sum1  = 0;
520:       sum2  = 0;
521:       sum3  = 0;
522:       sum4  = 0;
523:       sum5  = 0;
524:       v2    = v1 + n;
525:       v3    = v2 + n;
526:       v4    = v3 + n;
527:       v5    = v4 + n;
528: 
529:       for (n = 0; n<sz-1; n+=2) {
530:         i1   = idx[0];
531:         i2   = idx[1];
532:         idx += 2;
533:         tmp0 = x[i1];
534:         tmp1 = x[i2];
535:         sum1 += v1[0] * tmp0 + v1[1] *tmp1; v1 += 2;
536:         sum2 += v2[0] * tmp0 + v2[1] *tmp1; v2 += 2;
537:         sum3 += v3[0] * tmp0 + v3[1] *tmp1; v3 += 2;
538:         sum4 += v4[0] * tmp0 + v4[1] *tmp1; v4 += 2;
539:         sum5 += v5[0] * tmp0 + v5[1] *tmp1; v5 += 2;
540:       }
541:       if (n == sz-1){
542:         tmp0  = x[*idx++];
543:         sum1 += *v1++ * tmp0;
544:         sum2 += *v2++ * tmp0;
545:         sum3 += *v3++ * tmp0;
546:         sum4 += *v4++ * tmp0;
547:         sum5 += *v5++ * tmp0;
548:       }
549:       y[row++]=sum1;
550:       y[row++]=sum2;
551:       y[row++]=sum3;
552:       y[row++]=sum4;
553:       y[row++]=sum5;
554:       v1      =v5;       /* Since the next block to be processed starts there */
555:       idx    +=4*sz;
556:       break;
557:     default :
558:       SETERRQ(PETSC_ERR_COR,"Node size not yet supported");
559:     }
560:   }
561:   VecRestoreArray(xx,&x);
562:   VecRestoreArray(yy,&y);
563:   PetscLogFlops(2*a->nz - A->m);
564:   return(0);
565: }
566: /* ----------------------------------------------------------- */
567: /* Almost same code as the MatMult_Inode() */
570: static PetscErrorCode MatMultAdd_Inode(Mat A,Vec xx,Vec zz,Vec yy)
571: {
572:   Mat_inode      *a = (Mat_inode*)A->data;
573:   PetscScalar    sum1,sum2,sum3,sum4,sum5,tmp0,tmp1;
574:   PetscScalar    *v1,*v2,*v3,*v4,*v5,*x,*y,*z,*zt;
576:   PetscInt       *idx,i1,i2,n,i,row,node_max,*ns,*ii,nsz,sz;
577: 
579:   if (!a->inode.size) SETERRQ(PETSC_ERR_COR,"Missing Inode Structure");
580:   node_max = a->inode.node_count;
581:   ns       = a->inode.size;     /* Node Size array */
582:   VecGetArray(xx,&x);
583:   VecGetArray(yy,&y);
584:   if (zz != yy) {
585:     VecGetArray(zz,&z);
586:   } else {
587:     z = y;
588:   }
589:   zt = z;

591:   idx  = a->j;
592:   v1   = a->a;
593:   ii   = a->i;

595:   for (i = 0,row = 0; i< node_max; ++i){
596:     nsz  = ns[i];
597:     n    = ii[1] - ii[0];
598:     ii  += nsz;
599:     sz   = n;                   /* No of non zeros in this row */
600:                                 /* Switch on the size of Node */
601:     switch (nsz){               /* Each loop in 'case' is unrolled */
602:     case 1 :
603:       sum1  = *zt++;
604: 
605:       for(n = 0; n< sz-1; n+=2) {
606:         i1   = idx[0];          /* The instructions are ordered to */
607:         i2   = idx[1];          /* make the compiler's job easy */
608:         idx += 2;
609:         tmp0 = x[i1];
610:         tmp1 = x[i2];
611:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
612:        }
613: 
614:       if(n   == sz-1){          /* Take care of the last nonzero  */
615:         tmp0  = x[*idx++];
616:         sum1 += *v1++ * tmp0;
617:       }
618:       y[row++]=sum1;
619:       break;
620:     case 2:
621:       sum1  = *zt++;
622:       sum2  = *zt++;
623:       v2    = v1 + n;
624: 
625:       for(n = 0; n< sz-1; n+=2) {
626:         i1   = idx[0];
627:         i2   = idx[1];
628:         idx += 2;
629:         tmp0 = x[i1];
630:         tmp1 = x[i2];
631:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
632:         sum2 += v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
633:       }
634:       if(n   == sz-1){
635:         tmp0  = x[*idx++];
636:         sum1 += *v1++ * tmp0;
637:         sum2 += *v2++ * tmp0;
638:       }
639:       y[row++]=sum1;
640:       y[row++]=sum2;
641:       v1      =v2;              /* Since the next block to be processed starts there*/
642:       idx    +=sz;
643:       break;
644:     case 3:
645:       sum1  = *zt++;
646:       sum2  = *zt++;
647:       sum3  = *zt++;
648:       v2    = v1 + n;
649:       v3    = v2 + n;
650: 
651:       for (n = 0; n< sz-1; n+=2) {
652:         i1   = idx[0];
653:         i2   = idx[1];
654:         idx += 2;
655:         tmp0 = x[i1];
656:         tmp1 = x[i2];
657:         sum1 += v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
658:         sum2 += v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
659:         sum3 += v3[0] * tmp0 + v3[1] * tmp1; v3 += 2;
660:       }
661:       if (n == sz-1){
662:         tmp0  = x[*idx++];
663:         sum1 += *v1++ * tmp0;
664:         sum2 += *v2++ * tmp0;
665:         sum3 += *v3++ * tmp0;
666:       }
667:       y[row++]=sum1;
668:       y[row++]=sum2;
669:       y[row++]=sum3;
670:       v1       =v3;             /* Since the next block to be processed starts there*/
671:       idx     +=2*sz;
672:       break;
673:     case 4:
674:       sum1  = *zt++;
675:       sum2  = *zt++;
676:       sum3  = *zt++;
677:       sum4  = *zt++;
678:       v2    = v1 + n;
679:       v3    = v2 + n;
680:       v4    = v3 + n;
681: 
682:       for (n = 0; n< sz-1; n+=2) {
683:         i1   = idx[0];
684:         i2   = idx[1];
685:         idx += 2;
686:         tmp0 = x[i1];
687:         tmp1 = x[i2];
688:         sum1 += v1[0] * tmp0 + v1[1] *tmp1; v1 += 2;
689:         sum2 += v2[0] * tmp0 + v2[1] *tmp1; v2 += 2;
690:         sum3 += v3[0] * tmp0 + v3[1] *tmp1; v3 += 2;
691:         sum4 += v4[0] * tmp0 + v4[1] *tmp1; v4 += 2;
692:       }
693:       if (n == sz-1){
694:         tmp0  = x[*idx++];
695:         sum1 += *v1++ * tmp0;
696:         sum2 += *v2++ * tmp0;
697:         sum3 += *v3++ * tmp0;
698:         sum4 += *v4++ * tmp0;
699:       }
700:       y[row++]=sum1;
701:       y[row++]=sum2;
702:       y[row++]=sum3;
703:       y[row++]=sum4;
704:       v1      =v4;              /* Since the next block to be processed starts there*/
705:       idx    +=3*sz;
706:       break;
707:     case 5:
708:       sum1  = *zt++;
709:       sum2  = *zt++;
710:       sum3  = *zt++;
711:       sum4  = *zt++;
712:       sum5  = *zt++;
713:       v2    = v1 + n;
714:       v3    = v2 + n;
715:       v4    = v3 + n;
716:       v5    = v4 + n;
717: 
718:       for (n = 0; n<sz-1; n+=2) {
719:         i1   = idx[0];
720:         i2   = idx[1];
721:         idx += 2;
722:         tmp0 = x[i1];
723:         tmp1 = x[i2];
724:         sum1 += v1[0] * tmp0 + v1[1] *tmp1; v1 += 2;
725:         sum2 += v2[0] * tmp0 + v2[1] *tmp1; v2 += 2;
726:         sum3 += v3[0] * tmp0 + v3[1] *tmp1; v3 += 2;
727:         sum4 += v4[0] * tmp0 + v4[1] *tmp1; v4 += 2;
728:         sum5 += v5[0] * tmp0 + v5[1] *tmp1; v5 += 2;
729:       }
730:       if(n   == sz-1){
731:         tmp0  = x[*idx++];
732:         sum1 += *v1++ * tmp0;
733:         sum2 += *v2++ * tmp0;
734:         sum3 += *v3++ * tmp0;
735:         sum4 += *v4++ * tmp0;
736:         sum5 += *v5++ * tmp0;
737:       }
738:       y[row++]=sum1;
739:       y[row++]=sum2;
740:       y[row++]=sum3;
741:       y[row++]=sum4;
742:       y[row++]=sum5;
743:       v1      =v5;       /* Since the next block to be processed starts there */
744:       idx    +=4*sz;
745:       break;
746:     default :
747:       SETERRQ(PETSC_ERR_COR,"Node size not yet supported");
748:     }
749:   }
750:   VecRestoreArray(xx,&x);
751:   VecRestoreArray(yy,&y);
752:   if (zz != yy) {
753:     VecRestoreArray(zz,&z);
754:   }
755:   PetscLogFlops(2*a->nz);
756:   return(0);
757: }

759: /* ----------------------------------------------------------- */
762: PetscErrorCode MatSolve_Inode(Mat A,Vec bb,Vec xx)
763: {
764:   Mat_inode      *a = (Mat_inode*)A->data;
765:   IS             iscol = a->col,isrow = a->row;
767:   PetscInt       *r,*c,i,j,n = A->m,*ai = a->i,nz,*a_j = a->j;
768:   PetscInt       node_max,*ns,row,nsz,aii,*vi,*ad,*aj,i0,i1,*rout,*cout;
769:   PetscScalar    *x,*b,*a_a = a->a,*tmp,*tmps,*aa,tmp0,tmp1;
770:   PetscScalar    sum1,sum2,sum3,sum4,sum5,*v1,*v2,*v3,*v4,*v5;

773:   if (A->factor!=FACTOR_LU) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Not for unfactored matrix");
774:   if (!a->inode.size) SETERRQ(PETSC_ERR_COR,"Missing Inode Structure");
775:   node_max = a->inode.node_count;
776:   ns       = a->inode.size;     /* Node Size array */

778:   VecGetArray(bb,&b);
779:   VecGetArray(xx,&x);
780:   tmp  = a->solve_work;
781: 
782:   ISGetIndices(isrow,&rout); r = rout;
783:   ISGetIndices(iscol,&cout); c = cout + (n-1);
784: 
785:   /* forward solve the lower triangular */
786:   tmps = tmp ;
787:   aa   = a_a ;
788:   aj   = a_j ;
789:   ad   = a->diag;

791:   for (i = 0,row = 0; i< node_max; ++i){
792:     nsz = ns[i];
793:     aii = ai[row];
794:     v1  = aa + aii;
795:     vi  = aj + aii;
796:     nz  = ad[row]- aii;
797: 
798:     switch (nsz){               /* Each loop in 'case' is unrolled */
799:     case 1 :
800:       sum1 = b[*r++];
801:       /*      while (nz--) sum1 -= *v1++ *tmps[*vi++];*/
802:       for(j=0; j<nz-1; j+=2){
803:         i0   = vi[0];
804:         i1   = vi[1];
805:         vi  +=2;
806:         tmp0 = tmps[i0];
807:         tmp1 = tmps[i1];
808:         sum1 -= v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
809:       }
810:       if(j == nz-1){
811:         tmp0 = tmps[*vi++];
812:         sum1 -= *v1++ *tmp0;
813:       }
814:       tmp[row ++]=sum1;
815:       break;
816:     case 2:
817:       sum1 = b[*r++];
818:       sum2 = b[*r++];
819:       v2   = aa + ai[row+1];

821:       for(j=0; j<nz-1; j+=2){
822:         i0   = vi[0];
823:         i1   = vi[1];
824:         vi  +=2;
825:         tmp0 = tmps[i0];
826:         tmp1 = tmps[i1];
827:         sum1 -= v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
828:         sum2 -= v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
829:       }
830:       if(j == nz-1){
831:         tmp0 = tmps[*vi++];
832:         sum1 -= *v1++ *tmp0;
833:         sum2 -= *v2++ *tmp0;
834:       }
835:       sum2 -= *v2++ * sum1;
836:       tmp[row ++]=sum1;
837:       tmp[row ++]=sum2;
838:       break;
839:     case 3:
840:       sum1 = b[*r++];
841:       sum2 = b[*r++];
842:       sum3 = b[*r++];
843:       v2   = aa + ai[row+1];
844:       v3   = aa + ai[row+2];
845: 
846:       for (j=0; j<nz-1; j+=2){
847:         i0   = vi[0];
848:         i1   = vi[1];
849:         vi  +=2;
850:         tmp0 = tmps[i0];
851:         tmp1 = tmps[i1];
852:         sum1 -= v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
853:         sum2 -= v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
854:         sum3 -= v3[0] * tmp0 + v3[1] * tmp1; v3 += 2;
855:       }
856:       if (j == nz-1){
857:         tmp0 = tmps[*vi++];
858:         sum1 -= *v1++ *tmp0;
859:         sum2 -= *v2++ *tmp0;
860:         sum3 -= *v3++ *tmp0;
861:       }
862:       sum2 -= *v2++ * sum1;
863:       sum3 -= *v3++ * sum1;
864:       sum3 -= *v3++ * sum2;
865:       tmp[row ++]=sum1;
866:       tmp[row ++]=sum2;
867:       tmp[row ++]=sum3;
868:       break;
869: 
870:     case 4:
871:       sum1 = b[*r++];
872:       sum2 = b[*r++];
873:       sum3 = b[*r++];
874:       sum4 = b[*r++];
875:       v2   = aa + ai[row+1];
876:       v3   = aa + ai[row+2];
877:       v4   = aa + ai[row+3];
878: 
879:       for (j=0; j<nz-1; j+=2){
880:         i0   = vi[0];
881:         i1   = vi[1];
882:         vi  +=2;
883:         tmp0 = tmps[i0];
884:         tmp1 = tmps[i1];
885:         sum1 -= v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
886:         sum2 -= v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
887:         sum3 -= v3[0] * tmp0 + v3[1] * tmp1; v3 += 2;
888:         sum4 -= v4[0] * tmp0 + v4[1] * tmp1; v4 += 2;
889:       }
890:       if (j == nz-1){
891:         tmp0 = tmps[*vi++];
892:         sum1 -= *v1++ *tmp0;
893:         sum2 -= *v2++ *tmp0;
894:         sum3 -= *v3++ *tmp0;
895:         sum4 -= *v4++ *tmp0;
896:       }
897:       sum2 -= *v2++ * sum1;
898:       sum3 -= *v3++ * sum1;
899:       sum4 -= *v4++ * sum1;
900:       sum3 -= *v3++ * sum2;
901:       sum4 -= *v4++ * sum2;
902:       sum4 -= *v4++ * sum3;
903: 
904:       tmp[row ++]=sum1;
905:       tmp[row ++]=sum2;
906:       tmp[row ++]=sum3;
907:       tmp[row ++]=sum4;
908:       break;
909:     case 5:
910:       sum1 = b[*r++];
911:       sum2 = b[*r++];
912:       sum3 = b[*r++];
913:       sum4 = b[*r++];
914:       sum5 = b[*r++];
915:       v2   = aa + ai[row+1];
916:       v3   = aa + ai[row+2];
917:       v4   = aa + ai[row+3];
918:       v5   = aa + ai[row+4];
919: 
920:       for (j=0; j<nz-1; j+=2){
921:         i0   = vi[0];
922:         i1   = vi[1];
923:         vi  +=2;
924:         tmp0 = tmps[i0];
925:         tmp1 = tmps[i1];
926:         sum1 -= v1[0] * tmp0 + v1[1] * tmp1; v1 += 2;
927:         sum2 -= v2[0] * tmp0 + v2[1] * tmp1; v2 += 2;
928:         sum3 -= v3[0] * tmp0 + v3[1] * tmp1; v3 += 2;
929:         sum4 -= v4[0] * tmp0 + v4[1] * tmp1; v4 += 2;
930:         sum5 -= v5[0] * tmp0 + v5[1] * tmp1; v5 += 2;
931:       }
932:       if (j == nz-1){
933:         tmp0 = tmps[*vi++];
934:         sum1 -= *v1++ *tmp0;
935:         sum2 -= *v2++ *tmp0;
936:         sum3 -= *v3++ *tmp0;
937:         sum4 -= *v4++ *tmp0;
938:         sum5 -= *v5++ *tmp0;
939:       }

941:       sum2 -= *v2++ * sum1;
942:       sum3 -= *v3++ * sum1;
943:       sum4 -= *v4++ * sum1;
944:       sum5 -= *v5++ * sum1;
945:       sum3 -= *v3++ * sum2;
946:       sum4 -= *v4++ * sum2;
947:       sum5 -= *v5++ * sum2;
948:       sum4 -= *v4++ * sum3;
949:       sum5 -= *v5++ * sum3;
950:       sum5 -= *v5++ * sum4;
951: 
952:       tmp[row ++]=sum1;
953:       tmp[row ++]=sum2;
954:       tmp[row ++]=sum3;
955:       tmp[row ++]=sum4;
956:       tmp[row ++]=sum5;
957:       break;
958:     default:
959:       SETERRQ(PETSC_ERR_COR,"Node size not yet supported \n");
960:     }
961:   }
962:   /* backward solve the upper triangular */
963:   for (i=node_max -1 ,row = n-1 ; i>=0; i--){
964:     nsz = ns[i];
965:     aii = ai[row+1] -1;
966:     v1  = aa + aii;
967:     vi  = aj + aii;
968:     nz  = aii- ad[row];
969:     switch (nsz){               /* Each loop in 'case' is unrolled */
970:     case 1 :
971:       sum1 = tmp[row];

973:       for(j=nz ; j>1; j-=2){
974:         i0   = vi[0];
975:         i1   = vi[-1];
976:         vi  -=2;
977:         tmp0 = tmps[i0];
978:         tmp1 = tmps[i1];
979:         sum1 -= v1[0] * tmp0 + v1[-1] * tmp1; v1 -= 2;
980:       }
981:       if (j==1){
982:         tmp0  = tmps[*vi--];
983:         sum1 -= *v1-- * tmp0;
984:       }
985:       x[*c--] = tmp[row] = sum1*a_a[ad[row]]; row--;
986:       break;
987:     case 2 :
988:       sum1 = tmp[row];
989:       sum2 = tmp[row -1];
990:       v2   = aa + ai[row]-1;
991:       for (j=nz ; j>1; j-=2){
992:         i0   = vi[0];
993:         i1   = vi[-1];
994:         vi  -=2;
995:         tmp0 = tmps[i0];
996:         tmp1 = tmps[i1];
997:         sum1 -= v1[0] * tmp0 + v1[-1] * tmp1; v1 -= 2;
998:         sum2 -= v2[0] * tmp0 + v2[-1] * tmp1; v2 -= 2;
999:       }
1000:       if (j==1){
1001:         tmp0  = tmps[*vi--];
1002:         sum1 -= *v1-- * tmp0;
1003:         sum2 -= *v2-- * tmp0;
1004:       }
1005: 
1006:       tmp0    = x[*c--] = tmp[row] = sum1*a_a[ad[row]]; row--;
1007:       sum2   -= *v2-- * tmp0;
1008:       x[*c--] = tmp[row] = sum2*a_a[ad[row]]; row--;
1009:       break;
1010:     case 3 :
1011:       sum1 = tmp[row];
1012:       sum2 = tmp[row -1];
1013:       sum3 = tmp[row -2];
1014:       v2   = aa + ai[row]-1;
1015:       v3   = aa + ai[row -1]-1;
1016:       for (j=nz ; j>1; j-=2){
1017:         i0   = vi[0];
1018:         i1   = vi[-1];
1019:         vi  -=2;
1020:         tmp0 = tmps[i0];
1021:         tmp1 = tmps[i1];
1022:         sum1 -= v1[0] * tmp0 + v1[-1] * tmp1; v1 -= 2;
1023:         sum2 -= v2[0] * tmp0 + v2[-1] * tmp1; v2 -= 2;
1024:         sum3 -= v3[0] * tmp0 + v3[-1] * tmp1; v3 -= 2;
1025:       }
1026:       if (j==1){
1027:         tmp0  = tmps[*vi--];
1028:         sum1 -= *v1-- * tmp0;
1029:         sum2 -= *v2-- * tmp0;
1030:         sum3 -= *v3-- * tmp0;
1031:       }
1032:       tmp0    = x[*c--] = tmp[row] = sum1*a_a[ad[row]]; row--;
1033:       sum2   -= *v2-- * tmp0;
1034:       sum3   -= *v3-- * tmp0;
1035:       tmp0    = x[*c--] = tmp[row] = sum2*a_a[ad[row]]; row--;
1036:       sum3   -= *v3-- * tmp0;
1037:       x[*c--] = tmp[row] = sum3*a_a[ad[row]]; row--;
1038: 
1039:       break;
1040:     case 4 :
1041:       sum1 = tmp[row];
1042:       sum2 = tmp[row -1];
1043:       sum3 = tmp[row -2];
1044:       sum4 = tmp[row -3];
1045:       v2   = aa + ai[row]-1;
1046:       v3   = aa + ai[row -1]-1;
1047:       v4   = aa + ai[row -2]-1;

1049:       for (j=nz ; j>1; j-=2){
1050:         i0   = vi[0];
1051:         i1   = vi[-1];
1052:         vi  -=2;
1053:         tmp0 = tmps[i0];
1054:         tmp1 = tmps[i1];
1055:         sum1 -= v1[0] * tmp0 + v1[-1] * tmp1; v1 -= 2;
1056:         sum2 -= v2[0] * tmp0 + v2[-1] * tmp1; v2 -= 2;
1057:         sum3 -= v3[0] * tmp0 + v3[-1] * tmp1; v3 -= 2;
1058:         sum4 -= v4[0] * tmp0 + v4[-1] * tmp1; v4 -= 2;
1059:       }
1060:       if (j==1){
1061:         tmp0  = tmps[*vi--];
1062:         sum1 -= *v1-- * tmp0;
1063:         sum2 -= *v2-- * tmp0;
1064:         sum3 -= *v3-- * tmp0;
1065:         sum4 -= *v4-- * tmp0;
1066:       }

1068:       tmp0    = x[*c--] = tmp[row] = sum1*a_a[ad[row]]; row--;
1069:       sum2   -= *v2-- * tmp0;
1070:       sum3   -= *v3-- * tmp0;
1071:       sum4   -= *v4-- * tmp0;
1072:       tmp0    = x[*c--] = tmp[row] = sum2*a_a[ad[row]]; row--;
1073:       sum3   -= *v3-- * tmp0;
1074:       sum4   -= *v4-- * tmp0;
1075:       tmp0    = x[*c--] = tmp[row] = sum3*a_a[ad[row]]; row--;
1076:       sum4   -= *v4-- * tmp0;
1077:       x[*c--] = tmp[row] = sum4*a_a[ad[row]]; row--;
1078:       break;
1079:     case 5 :
1080:       sum1 = tmp[row];
1081:       sum2 = tmp[row -1];
1082:       sum3 = tmp[row -2];
1083:       sum4 = tmp[row -3];
1084:       sum5 = tmp[row -4];
1085:       v2   = aa + ai[row]-1;
1086:       v3   = aa + ai[row -1]-1;
1087:       v4   = aa + ai[row -2]-1;
1088:       v5   = aa + ai[row -3]-1;
1089:       for (j=nz ; j>1; j-=2){
1090:         i0   = vi[0];
1091:         i1   = vi[-1];
1092:         vi  -=2;
1093:         tmp0 = tmps[i0];
1094:         tmp1 = tmps[i1];
1095:         sum1 -= v1[0] * tmp0 + v1[-1] * tmp1; v1 -= 2;
1096:         sum2 -= v2[0] * tmp0 + v2[-1] * tmp1; v2 -= 2;
1097:         sum3 -= v3[0] * tmp0 + v3[-1] * tmp1; v3 -= 2;
1098:         sum4 -= v4[0] * tmp0 + v4[-1] * tmp1; v4 -= 2;
1099:         sum5 -= v5[0] * tmp0 + v5[-1] * tmp1; v5 -= 2;
1100:       }
1101:       if (j==1){
1102:         tmp0  = tmps[*vi--];
1103:         sum1 -= *v1-- * tmp0;
1104:         sum2 -= *v2-- * tmp0;
1105:         sum3 -= *v3-- * tmp0;
1106:         sum4 -= *v4-- * tmp0;
1107:         sum5 -= *v5-- * tmp0;
1108:       }

1110:       tmp0    = x[*c--] = tmp[row] = sum1*a_a[ad[row]]; row--;
1111:       sum2   -= *v2-- * tmp0;
1112:       sum3   -= *v3-- * tmp0;
1113:       sum4   -= *v4-- * tmp0;
1114:       sum5   -= *v5-- * tmp0;
1115:       tmp0    = x[*c--] = tmp[row] = sum2*a_a[ad[row]]; row--;
1116:       sum3   -= *v3-- * tmp0;
1117:       sum4   -= *v4-- * tmp0;
1118:       sum5   -= *v5-- * tmp0;
1119:       tmp0    = x[*c--] = tmp[row] = sum3*a_a[ad[row]]; row--;
1120:       sum4   -= *v4-- * tmp0;
1121:       sum5   -= *v5-- * tmp0;
1122:       tmp0    = x[*c--] = tmp[row] = sum4*a_a[ad[row]]; row--;
1123:       sum5   -= *v5-- * tmp0;
1124:       x[*c--] = tmp[row] = sum5*a_a[ad[row]]; row--;
1125:       break;
1126:     default:
1127:       SETERRQ(PETSC_ERR_COR,"Node size not yet supported \n");
1128:     }
1129:   }
1130:   ISRestoreIndices(isrow,&rout);
1131:   ISRestoreIndices(iscol,&cout);
1132:   VecRestoreArray(bb,&b);
1133:   VecRestoreArray(xx,&x);
1134:   PetscLogFlops(2*a->nz - A->n);
1135:   return(0);
1136: }

1140: PetscErrorCode MatLUFactorNumeric_Inode(Mat A,MatFactorInfo *info,Mat *B)
1141: {
1142:   Mat            C = *B;
1143:   Mat_inode      *a = (Mat_inode*)A->data,*b = (Mat_inode*)C->data;
1144:   IS             iscol = b->col,isrow = b->row,isicol = b->icol;
1146:   PetscInt       *r,*ic,*c,n = A->m,*bi = b->i;
1147:   PetscInt       *bj = b->j,*nbj=b->j +1,*ajtmp,*bjtmp,nz,row,prow;
1148:   PetscInt       *ics,i,j,idx,*ai = a->i,*aj = a->j,*bd = b->diag,node_max,nodesz;
1149:   PetscInt       *ns,*tmp_vec1,*tmp_vec2,*nsmap,*pj;
1150:   PetscScalar    *rtmp1,*rtmp2,*rtmp3,*v1,*v2,*v3,*pc1,*pc2,*pc3,mul1,mul2,mul3;
1151:   PetscScalar    tmp,*ba = b->a,*aa = a->a,*pv,*rtmps1,*rtmps2,*rtmps3;
1152:   PetscReal      rs=0.0,rsum[3];
1153:   LUShift_Ctx    sctx;
1154:   PetscInt       newshift;

1157:   /* if both shift schemes are chosen by user, only use info->shiftpd */
1158:   if (info->shiftpd && info->shiftnz) info->shiftnz = 0.0;
1159:   if (info->shiftpd) { /* set sctx.shift_top=max{rs} */
1160:     sctx.shift_top = 0;
1161:     for (i=0; i<n; i++) {
1162:       /* calculate rs = sum(|aij|)-RealPart(aii), amt of shift needed for this row */
1163:       rs    = 0.0;
1164:       ajtmp = aj + ai[i];
1165:       rtmp1 = aa + ai[i];
1166:       nz = ai[i+1] - ai[i];
1167:       for (j=0; j<nz; j++){
1168:         if (*ajtmp != i){
1169:           rs += PetscAbsScalar(*rtmp1++);
1170:         } else {
1171:           rs -= PetscRealPart(*rtmp1++);
1172:         }
1173:         ajtmp++;
1174:       }
1175:       if (rs>sctx.shift_top) sctx.shift_top = rs;
1176:     }
1177:     if (sctx.shift_top == 0.0) sctx.shift_top += 1.e-12;
1178:     sctx.shift_top   *= 1.1;
1179:     sctx.nshift_max   = 5;
1180:     sctx.shift_lo     = 0.;
1181:     sctx.shift_hi     = 1.;
1182:   }
1183:   sctx.shift_amount = 0;
1184:   sctx.nshift       = 0;

1186:   ISGetIndices(isrow,&r);
1187:   ISGetIndices(iscol,&c);
1188:   ISGetIndices(isicol,&ic);
1189:   PetscMalloc((3*n+1)*sizeof(PetscScalar),&rtmp1);
1190:   PetscMemzero(rtmp1,(3*n+1)*sizeof(PetscScalar));
1191:   ics    = ic ; rtmps1 = rtmp1 ;
1192:   rtmp2  = rtmp1 + n;  rtmps2 = rtmp2 ;
1193:   rtmp3  = rtmp2 + n;  rtmps3 = rtmp3 ;
1194: 
1195:   node_max = a->inode.node_count;
1196:   ns       = a->inode.size ;
1197:   if (!ns){
1198:     SETERRQ(PETSC_ERR_PLIB,"Matrix without inode information");
1199:   }

1201:   /* If max inode size > 3, split it into two inodes.*/
1202:   /* also map the inode sizes according to the ordering */
1203:   PetscMalloc((n+1)* sizeof(PetscInt),&tmp_vec1);
1204:   for (i=0,j=0; i<node_max; ++i,++j){
1205:     if (ns[i]>3) {
1206:       tmp_vec1[j] = ns[i]/2; /* Assuming ns[i] < =5  */
1207:       ++j;
1208:       tmp_vec1[j] = ns[i] - tmp_vec1[j-1];
1209:     } else {
1210:       tmp_vec1[j] = ns[i];
1211:     }
1212:   }
1213:   /* Use the correct node_max */
1214:   node_max = j;

1216:   /* Now reorder the inode info based on mat re-ordering info */
1217:   /* First create a row -> inode_size_array_index map */
1218:   PetscMalloc(n*sizeof(PetscInt)+1,&nsmap);
1219:   PetscMalloc(node_max*sizeof(PetscInt)+1,&tmp_vec2);
1220:   for (i=0,row=0; i<node_max; i++) {
1221:     nodesz = tmp_vec1[i];
1222:     for (j=0; j<nodesz; j++,row++) {
1223:       nsmap[row] = i;
1224:     }
1225:   }
1226:   /* Using nsmap, create a reordered ns structure */
1227:   for (i=0,j=0; i< node_max; i++) {
1228:     nodesz       = tmp_vec1[nsmap[r[j]]];    /* here the reordered row_no is in r[] */
1229:     tmp_vec2[i] = nodesz;
1230:     j        += nodesz;
1231:   }
1232:   PetscFree(nsmap);
1233:   PetscFree(tmp_vec1);
1234:   /* Now use the correct ns */
1235:   ns = tmp_vec2;

1237:   do {
1238:     sctx.lushift = PETSC_FALSE;
1239:     /* Now loop over each block-row, and do the factorization */
1240:     for (i=0,row=0; i<node_max; i++) {
1241:       nodesz   = ns[i];
1242:       nz    = bi[row+1] - bi[row];
1243:       bjtmp = bj + bi[row];

1245:       switch (nodesz){
1246:       case 1:
1247:         for  (j=0; j<nz; j++){
1248:           idx         = bjtmp[j];
1249:           rtmps1[idx] = 0.0;
1250:         }
1251: 
1252:         /* load in initial (unfactored row) */
1253:         idx   = r[row];
1254:         nz    = ai[idx+1] - ai[idx];
1255:         ajtmp = aj + ai[idx];
1256:         v1    = aa + ai[idx];

1258:         for (j=0; j<nz; j++) {
1259:           idx        = ics[ajtmp[j]];
1260:           rtmp1[idx] = v1[j];
1261:           if (sctx.nshift && ajtmp[j] == r[row]) {
1262:             rtmp1[idx] += sctx.shift_amount;
1263:           }
1264:         }
1265:         prow = *bjtmp++ ;
1266:         while (prow < row) {
1267:           pc1 = rtmp1 + prow;
1268:           if (*pc1 != 0.0){
1269:             pv   = ba + bd[prow];
1270:             pj   = nbj + bd[prow];
1271:             mul1 = *pc1 * *pv++;
1272:             *pc1 = mul1;
1273:             nz   = bi[prow+1] - bd[prow] - 1;
1274:             PetscLogFlops(2*nz);
1275:             for (j=0; j<nz; j++) {
1276:               tmp = pv[j];
1277:               idx = pj[j];
1278:               rtmps1[idx] -= mul1 * tmp;
1279:             }
1280:           }
1281:           prow = *bjtmp++ ;
1282:         }
1283:         nz  = bi[row+1] - bi[row];
1284:         pj  = bj + bi[row];
1285:         pc1 = ba + bi[row];

1287:         sctx.pv    = rtmp1[row];
1288:         rs         = 0.0;
1289:         rtmp1[row] = 1.0/rtmp1[row];
1290:         for (j=0; j<nz; j++) {
1291:           idx    = pj[j];
1292:           pc1[j] = rtmps1[idx];
1293:           if (idx != row) rs += PetscAbsScalar(pc1[j]);
1294:         }

1296:         sctx.rs  = rs;
1297:         MatLUCheckShift_inline(info,sctx,newshift);
1298:         if (newshift == 1){
1299:           goto endofwhile;
1300:         } else if (newshift == -1){
1301:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g, inode.size %D",row,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1302:         }
1303:         break;
1304: 
1305:       case 2:
1306:         for  (j=0; j<nz; j++) {
1307:           idx         = bjtmp[j];
1308:           rtmps1[idx] = 0.0;
1309:           rtmps2[idx] = 0.0;
1310:         }
1311: 
1312:         /* load in initial (unfactored row) */
1313:         idx   = r[row];
1314:         nz    = ai[idx+1] - ai[idx];
1315:         ajtmp = aj + ai[idx];
1316:         v1    = aa + ai[idx];
1317:         v2    = aa + ai[idx+1];
1318: 
1319:         for (j=0; j<nz; j++) {
1320:           idx        = ics[ajtmp[j]];
1321:           rtmp1[idx] = v1[j];
1322:           rtmp2[idx] = v2[j];
1323:           if (sctx.nshift && ajtmp[j] == r[row]) {
1324:             rtmp1[idx] += sctx.shift_amount;
1325:           }
1326:           if (sctx.nshift && ajtmp[j] == r[row+1]) {
1327:             rtmp2[idx] += sctx.shift_amount;
1328:           }
1329:         }
1330:         prow = *bjtmp++ ;
1331:         while (prow < row) {
1332:           pc1 = rtmp1 + prow;
1333:           pc2 = rtmp2 + prow;
1334:           if (*pc1 != 0.0 || *pc2 != 0.0){
1335:             pv   = ba + bd[prow];
1336:             pj   = nbj + bd[prow];
1337:             mul1 = *pc1 * *pv;
1338:             mul2 = *pc2 * *pv;
1339:             ++pv;
1340:             *pc1 = mul1;
1341:             *pc2 = mul2;
1342: 
1343:             nz   = bi[prow+1] - bd[prow] - 1;
1344:             PetscLogFlops(2*2*nz);
1345:             for (j=0; j<nz; j++) {
1346:               tmp = pv[j];
1347:               idx = pj[j];
1348:               rtmps1[idx] -= mul1 * tmp;
1349:               rtmps2[idx] -= mul2 * tmp;
1350:             }
1351:           }
1352:           prow = *bjtmp++ ;
1353:         }
1354:         /* Now take care of the odd element*/
1355:         pc1 = rtmp1 + prow;
1356:         pc2 = rtmp2 + prow;
1357:         if (*pc2 != 0.0){
1358:           pj   = nbj + bd[prow];

1360:           rs   = 0.0;
1361:           mul2 = (*pc2)/(*pc1); /* since diag is not yet inverted.*/
1362:           *pc2 = mul2;
1363:           nz   = bi[prow+1] - bd[prow] - 1;
1364:           PetscLogFlops(2*nz);
1365:           for (j=0; j<nz; j++) {
1366:             idx = pj[j] ;
1367:             tmp = rtmp1[idx];
1368:             rtmp2[idx] -= mul2 * tmp;
1369:             if (idx != prow) rs += PetscAbsScalar(rtmp2[idx]);
1370:           }
1371: 
1372:           sctx.rs  = rs;
1373:           sctx.pv  = *pc1;
1374:           MatLUCheckShift_inline(info,sctx,newshift);
1375:           if (newshift == 1){
1376:             goto endofwhile; /* sctx.shift_amount is updated */
1377:           } else if (newshift == -1){
1378:             SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",prow,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1379:           }
1380:         }
1381: 
1382:         nz  = bi[row+1] - bi[row];
1383:         pj  = bj + bi[row];
1384:         pc1 = ba + bi[row];
1385:         pc2 = ba + bi[row+1];

1387:         rsum[0] = rsum[1] = 0.0;
1388:         rtmp1[row]   = 1.0/rtmp1[row];
1389:         rtmp2[row+1] = 1.0/rtmp2[row+1];
1390:         for (j=0; j<nz; j++) {
1391:           idx    = pj[j];
1392:           pc1[j] = rtmps1[idx];
1393:           pc2[j] = rtmps2[idx];
1394:           if (idx != row)   rsum[0] += PetscAbsScalar(pc1[j]);
1395:           if (idx != row+1) rsum[1] += PetscAbsScalar(pc2[j]);
1396:         }

1398:         sctx.pv = 1.0/rtmp1[row]; /* rtmp1[row] = 1.0/diag[row] */
1399:         sctx.rs = rsum[0];
1400:         MatLUCheckShift_inline(info,sctx,newshift);
1401:         if (newshift == 1){
1402:           goto endofwhile;
1403:         } else if (newshift == -1){
1404:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",row,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1405:         }
1406:         sctx.pv = 1.0/rtmp2[row+1];
1407:         sctx.rs = rsum[1];
1408:         MatLUCheckShift_inline(info,sctx,newshift);
1409:         if (newshift == 1){
1410:           goto endofwhile;
1411:         } else if (newshift == -1){
1412:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",row+1,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1413:         }
1414:         break;

1416:       case 3:
1417:         for  (j=0; j<nz; j++) {
1418:           idx         = bjtmp[j];
1419:           rtmps1[idx] = 0.0;
1420:           rtmps2[idx] = 0.0;
1421:           rtmps3[idx] = 0.0;
1422:         }
1423:         /* copy the nonzeros for the 3 rows from sparse representation to dense in rtmp*[] */
1424:         idx   = r[row];
1425:         nz    = ai[idx+1] - ai[idx];
1426:         ajtmp = aj + ai[idx];
1427:         v1    = aa + ai[idx];
1428:         v2    = aa + ai[idx+1];
1429:         v3    = aa + ai[idx+2];
1430:         for (j=0; j<nz; j++) {
1431:           idx        = ics[ajtmp[j]];
1432:           rtmp1[idx] = v1[j];
1433:           rtmp2[idx] = v2[j];
1434:           rtmp3[idx] = v3[j];
1435:           if (sctx.nshift && ajtmp[j] == r[row]) {
1436:             rtmp1[idx] += sctx.shift_amount;
1437:           }
1438:           if (sctx.nshift && ajtmp[j] == r[row+1]) {
1439:             rtmp2[idx] += sctx.shift_amount;
1440:           }
1441:           if (sctx.nshift && ajtmp[j] == r[row+2]) {
1442:             rtmp3[idx] += sctx.shift_amount;
1443:           }
1444:         }
1445:         /* loop over all pivot row blocks above this row block */
1446:         prow = *bjtmp++ ;
1447:         while (prow < row) {
1448:           pc1 = rtmp1 + prow;
1449:           pc2 = rtmp2 + prow;
1450:           pc3 = rtmp3 + prow;
1451:           if (*pc1 != 0.0 || *pc2 != 0.0 || *pc3 !=0.0){
1452:             pv   = ba  + bd[prow];
1453:             pj   = nbj + bd[prow];
1454:             mul1 = *pc1 * *pv;
1455:             mul2 = *pc2 * *pv;
1456:             mul3 = *pc3 * *pv;
1457:             ++pv;
1458:             *pc1 = mul1;
1459:             *pc2 = mul2;
1460:             *pc3 = mul3;
1461: 
1462:             nz   = bi[prow+1] - bd[prow] - 1;
1463:             PetscLogFlops(3*2*nz);
1464:             /* update this row based on pivot row */
1465:             for (j=0; j<nz; j++) {
1466:               tmp = pv[j];
1467:               idx = pj[j];
1468:               rtmps1[idx] -= mul1 * tmp;
1469:               rtmps2[idx] -= mul2 * tmp;
1470:               rtmps3[idx] -= mul3 * tmp;
1471:             }
1472:           }
1473:           prow = *bjtmp++ ;
1474:         }
1475:         /* Now take care of diagonal block in this set of rows */
1476:         pc1 = rtmp1 + prow;
1477:         pc2 = rtmp2 + prow;
1478:         pc3 = rtmp3 + prow;
1479:         if (*pc2 != 0.0 || *pc3 != 0.0){
1480:           pj   = nbj + bd[prow];

1482:           sctx.rs = 1.0; /* for simplicity, set rs=1.0 */
1483:           sctx.pv = *pc1;
1484:           MatLUCheckShift_inline(info,sctx,newshift);
1485:           if (newshift == 1){
1486:             goto endofwhile;
1487:           } else if (newshift == -1){
1488:             SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",prow,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1489:           }

1491:           mul2 = (*pc2)/(*pc1);
1492:           mul3 = (*pc3)/(*pc1);
1493:           *pc2 = mul2;
1494:           *pc3 = mul3;
1495:           nz   = bi[prow+1] - bd[prow] - 1;
1496:           PetscLogFlops(2*2*nz);
1497:           for (j=0; j<nz; j++) {
1498:             idx = pj[j] ;
1499:             tmp = rtmp1[idx];
1500:             rtmp2[idx] -= mul2 * tmp;
1501:             rtmp3[idx] -= mul3 * tmp;
1502:           }
1503:         }
1504:         ++prow;
1505:         pc2 = rtmp2 + prow;
1506:         pc3 = rtmp3 + prow;
1507:         if (*pc3 != 0.0){
1508:           pj   = nbj + bd[prow];
1509:           pj   = nbj + bd[prow];

1511:           sctx.rs  = 1.0; /* for simplicity, set rs=1.0 */
1512:           sctx.pv  = *pc2;
1513:           MatLUCheckShift_inline(info,sctx,newshift);
1514:           if (newshift == 1){
1515:             goto endofwhile;
1516:           } else if (newshift == -1){
1517:             SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",prow,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1518:           }
1519:           mul3 = (*pc3)/(*pc2);
1520:           *pc3 = mul3;
1521:           nz   = bi[prow+1] - bd[prow] - 1;
1522:           PetscLogFlops(2*2*nz);
1523:           for (j=0; j<nz; j++) {
1524:             idx = pj[j] ;
1525:             tmp = rtmp2[idx];
1526:             rtmp3[idx] -= mul3 * tmp;
1527:           }
1528:         }
1529:         nz  = bi[row+1] - bi[row];
1530:         pj  = bj + bi[row];
1531:         pc1 = ba + bi[row];
1532:         pc2 = ba + bi[row+1];
1533:         pc3 = ba + bi[row+2];

1535:         rsum[0] = rsum[1] = rsum[2] = 0.0;
1536:         rtmp1[row]   = 1.0/rtmp1[row];
1537:         rtmp2[row+1] = 1.0/rtmp2[row+1];
1538:         rtmp3[row+2] = 1.0/rtmp3[row+2];
1539:         /* copy row entries from dense representation to sparse */
1540:         for (j=0; j<nz; j++) {
1541:           idx    = pj[j];
1542:           pc1[j] = rtmps1[idx];
1543:           pc2[j] = rtmps2[idx];
1544:           pc3[j] = rtmps3[idx];
1545:           if (idx != row) rsum[0] += PetscAbsScalar(pc1[j]);
1546:           if (idx != row+1) rsum[1] += PetscAbsScalar(pc2[j]);
1547:           if (idx != row+2) rsum[2] += PetscAbsScalar(pc3[j]);
1548:         }

1550:         /* sctx.rs = rs/3.0; */
1551:         sctx.pv = 1.0/rtmp1[row];
1552:         sctx.rs = rsum[0];
1553:         MatLUCheckShift_inline(info,sctx,newshift);
1554:         if (newshift == 1){
1555:           goto endofwhile;
1556:         } else if (newshift == -1){
1557:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D",row,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1558:         }
1559:         sctx.pv = 1.0/rtmp2[row+1];
1560:         sctx.rs = rsum[1];
1561:         MatLUCheckShift_inline(info,sctx,newshift);
1562:         if (newshift == 1){
1563:           goto endofwhile;
1564:         } else if (newshift == -1){
1565:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D, 2nd row",row+1,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1566:         }
1567:         sctx.pv = 1.0/rtmp3[row+2];
1568:         sctx.rs = rsum[2];
1569:         MatLUCheckShift_inline(info,sctx,newshift);
1570:         if (newshift == 1){
1571:           goto endofwhile;
1572:         } else if (newshift == -1){
1573:           SETERRQ5(PETSC_ERR_MAT_LU_ZRPVT,"Zero pivot row %D value %g tolerance %g * rs %g inode.size %D, 3rd row",row+2,PetscAbsScalar(sctx.pv),info->zeropivot,rs,nodesz);
1574:         }
1575:         break;
1576:       default:
1577:         SETERRQ(PETSC_ERR_COR,"Node size not yet supported \n");
1578:       }
1579:       row += nodesz;                 /* Update the row */
1580:     }
1581:     endofwhile:;
1582:   } while (sctx.lushift);
1583:   PetscFree(rtmp1);
1584:   PetscFree(tmp_vec2);
1585:   ISRestoreIndices(isicol,&ic);
1586:   ISRestoreIndices(isrow,&r);
1587:   ISRestoreIndices(iscol,&c);
1588:   C->factor      = FACTOR_LU;
1589:   C->assembled   = PETSC_TRUE;
1590:   if (sctx.nshift) {
1591:     if (info->shiftnz) {
1592:       PetscLogInfo((0,"MatLUFactorNumeric_Inode: number of shift_nz tries %D, shift_amount %g\n",sctx.nshift,sctx.shift_amount));
1593:     } else if (info->shiftpd) {
1594:       PetscLogInfo((0,"MatLUFactorNumeric_Inode: number of shift_pd tries %D, shift_amount %g, diagonal shifted up by %e fraction top_value %e\n",sctx.nshift,sctx.shift_amount,info->shift_fraction,sctx.shift_top));
1595:     }
1596:   }
1597:   PetscLogFlops(C->n);
1598:   return(0);
1599: }

1601: /*
1602:      Makes a longer coloring[] array and calls the usual code with that
1603: */
1606: PetscErrorCode MatColoringPatch_Inode(Mat mat,PetscInt nin,PetscInt ncolors,ISColoringValue coloring[],ISColoring *iscoloring)
1607: {
1608:   Mat_inode       *a = (Mat_inode*)mat->data;
1609:   PetscErrorCode  ierr;
1610:   PetscInt        n = mat->n,m = a->inode.node_count,j,*ns = a->inode.size,row;
1611:   PetscInt        *colorused,i;
1612:   ISColoringValue *newcolor;

1615:   PetscMalloc((n+1)*sizeof(PetscInt),&newcolor);
1616:   /* loop over inodes, marking a color for each column*/
1617:   row = 0;
1618:   for (i=0; i<m; i++){
1619:     for (j=0; j<ns[i]; j++) {
1620:       newcolor[row++] = coloring[i] + j*ncolors;
1621:     }
1622:   }

1624:   /* eliminate unneeded colors */
1625:   PetscMalloc(5*ncolors*sizeof(PetscInt),&colorused);
1626:   PetscMemzero(colorused,5*ncolors*sizeof(PetscInt));
1627:   for (i=0; i<n; i++) {
1628:     colorused[newcolor[i]] = 1;
1629:   }

1631:   for (i=1; i<5*ncolors; i++) {
1632:     colorused[i] += colorused[i-1];
1633:   }
1634:   ncolors = colorused[5*ncolors-1];
1635:   for (i=0; i<n; i++) {
1636:     newcolor[i] = colorused[newcolor[i]];
1637:   }
1638:   PetscFree(colorused);
1639:   ISColoringCreate(mat->comm,n,newcolor,iscoloring);
1640:   PetscFree(coloring);
1641:   return(0);
1642: }

1644: /*
1645:     samestructure indicates that the matrix has not changed its nonzero structure so we 
1646:     do not need to recompute the inodes 
1647: */
1650: PetscErrorCode Mat_CheckInode(Mat A,PetscTruth samestructure)
1651: {
1652:   Mat_inode      *a = (Mat_inode*)A->data;
1654:   PetscInt       i,j,m,nzx,nzy,*idx,*idy,*ns,*ii,node_count,blk_size;
1655:   PetscTruth     flag,flg;

1658:   if (a->inode.checked && samestructure) return(0);

1660:   a->inode.checked = PETSC_TRUE;

1662:   /* Notes: We set a->inode.limit=5 in MatCreate_Inode(). */
1663:   if (!a->inode.use) {PetscLogInfo((A,"Mat_CheckInode: Not using Inode routines due to MatSetOption(MAT_DO_NOT_USE_INODES\n")); return(0);}
1664:   PetscOptionsHasName(A->prefix,"-mat_no_inode",&flg);
1665:   if (flg) {PetscLogInfo((A,"Mat_CheckInode: Not using Inode routines due to -mat_no_inode\n"));return(0);}
1666:   PetscOptionsHasName(A->prefix,"-mat_no_unroll",&flg);
1667:   if (flg) {PetscLogInfo((A,"Mat_CheckInode: Not using Inode routines due to -mat_no_unroll\n"));return(0);}
1668:   PetscOptionsGetInt(A->prefix,"-mat_inode_limit",&a->inode.limit,PETSC_NULL);
1669:   if (a->inode.limit > a->inode.max_limit) a->inode.limit = a->inode.max_limit;
1670:   m = A->m;
1671:   if (a->inode.size) {ns = a->inode.size;}
1672:   else {PetscMalloc((m+1)*sizeof(PetscInt),&ns);}

1674:   i          = 0;
1675:   node_count = 0;
1676:   idx        = a->j;
1677:   ii         = a->i;
1678:   while (i < m){                /* For each row */
1679:     nzx = ii[i+1] - ii[i];       /* Number of nonzeros */
1680:     /* Limits the number of elements in a node to 'a->inode.limit' */
1681:     for (j=i+1,idy=idx,blk_size=1; j<m && blk_size <a->inode.limit; ++j,++blk_size) {
1682:       nzy     = ii[j+1] - ii[j]; /* Same number of nonzeros */
1683:       if (nzy != nzx) break;
1684:       idy  += nzx;             /* Same nonzero pattern */
1685:       PetscMemcmp(idx,idy,nzx*sizeof(PetscInt),&flag);
1686:       if (!flag) break;
1687:     }
1688:     ns[node_count++] = blk_size;
1689:     idx += blk_size*nzx;
1690:     i    = j;
1691:   }
1692:   /* If not enough inodes found,, do not use inode version of the routines */
1693:   if (!a->inode.size && m && node_count > 0.9*m) {
1694:     PetscFree(ns);
1695:     a->inode.node_count     = 0;
1696:     a->inode.size           = PETSC_NULL;
1697:     a->inode.use            = PETSC_FALSE;
1698:     PetscLogInfo((A,"Mat_CheckInode: Found %D nodes out of %D rows. Not using Inode routines\n",node_count,m));
1699:   } else {
1700:     A->ops->mult            = MatMult_Inode;
1701:     A->ops->multadd         = MatMultAdd_Inode;
1702:     A->ops->solve           = MatSolve_Inode;
1703:     A->ops->lufactornumeric = MatLUFactorNumeric_Inode;
1704:     A->ops->getrowij        = MatGetRowIJ_Inode;
1705:     A->ops->restorerowij    = MatRestoreRowIJ_Inode;
1706:     A->ops->getcolumnij     = MatGetColumnIJ_Inode;
1707:     A->ops->restorecolumnij = MatRestoreColumnIJ_Inode;
1708:     A->ops->coloringpatch   = MatColoringPatch_Inode;
1709:     a->inode.node_count     = node_count;
1710:     a->inode.size           = ns;
1711:     PetscLogInfo((A,"Mat_CheckInode: Found %D nodes of %D. Limit used: %D. Using Inode routines\n",node_count,m,a->inode.limit));
1712:   }
1713:   return(0);
1714: }

1716: /*
1717:      This is really ugly. if inodes are used this replaces the 
1718:   permutations with ones that correspond to rows/cols of the matrix
1719:   rather then inode blocks
1720: */
1723: PetscErrorCode PETSCMAT_DLLEXPORT MatInodeAdjustForInodes(Mat A,IS *rperm,IS *cperm)
1724: {
1725:   PetscErrorCode ierr,(*f)(Mat,IS*,IS*);

1728:   PetscObjectQueryFunction((PetscObject)A,"MatInodeAdjustForInodes_C",(void (**)(void))&f);
1729:   if (f) {
1730:     (*f)(A,rperm,cperm);
1731:   }
1732:   return(0);
1733: }

1738: PetscErrorCode PETSCMAT_DLLEXPORT MatInodeAdjustForInodes_Inode(Mat A,IS *rperm,IS *cperm)
1739: {
1740:   Mat_inode      *a=(Mat_inode*)A->data;
1742:   PetscInt       m = A->m,n = A->n,i,j,*ridx,*cidx,nslim_row = a->inode.node_count;
1743:   PetscInt       row,col,*permr,*permc,*ns_row =  a->inode.size,*tns,start_val,end_val,indx;
1744:   PetscInt       nslim_col,*ns_col;
1745:   IS             ris = *rperm,cis = *cperm;

1748:   if (!a->inode.size) return(0); /* no inodes so return */
1749:   if (a->inode.node_count == m) return(0); /* all inodes are of size 1 */

1751:   Mat_CreateColInode(A,&nslim_col,&ns_col);
1752:   PetscMalloc((((nslim_row>nslim_col)?nslim_row:nslim_col)+1)*sizeof(PetscInt),&tns);
1753:   PetscMalloc((m+n+1)*sizeof(PetscInt),&permr);
1754:   permc = permr + m;

1756:   ISGetIndices(ris,&ridx);
1757:   ISGetIndices(cis,&cidx);

1759:   /* Form the inode structure for the rows of permuted matric using inv perm*/
1760:   for (i=0,tns[0]=0; i<nslim_row; ++i) tns[i+1] = tns[i] + ns_row[i];

1762:   /* Construct the permutations for rows*/
1763:   for (i=0,row = 0; i<nslim_row; ++i){
1764:     indx      = ridx[i];
1765:     start_val = tns[indx];
1766:     end_val   = tns[indx + 1];
1767:     for (j=start_val; j<end_val; ++j,++row) permr[row]= j;
1768:   }

1770:   /* Form the inode structure for the columns of permuted matrix using inv perm*/
1771:   for (i=0,tns[0]=0; i<nslim_col; ++i) tns[i+1] = tns[i] + ns_col[i];

1773:  /* Construct permutations for columns */
1774:   for (i=0,col=0; i<nslim_col; ++i){
1775:     indx      = cidx[i];
1776:     start_val = tns[indx];
1777:     end_val   = tns[indx + 1];
1778:     for (j = start_val; j<end_val; ++j,++col) permc[col]= j;
1779:   }

1781:   ISCreateGeneral(PETSC_COMM_SELF,n,permr,rperm);
1782:   ISSetPermutation(*rperm);
1783:   ISCreateGeneral(PETSC_COMM_SELF,n,permc,cperm);
1784:   ISSetPermutation(*cperm);
1785: 
1786:   ISRestoreIndices(ris,&ridx);
1787:   ISRestoreIndices(cis,&cidx);

1789:   PetscFree(ns_col);
1790:   PetscFree(permr);
1791:   ISDestroy(cis);
1792:   ISDestroy(ris);
1793:   PetscFree(tns);
1794:   return(0);
1795: }

1800: /*@C
1801:    MatInodeGetInodeSizes - Returns the inode information of the Inode matrix.

1803:    Collective on Mat

1805:    Input Parameter:
1806: .  A - the Inode matrix or matrix derived from the Inode class -- e.g., SeqAIJ

1808:    Output Parameter:
1809: +  node_count - no of inodes present in the matrix.
1810: .  sizes      - an array of size node_count,with sizes of each inode.
1811: -  limit      - the max size used to generate the inodes.

1813:    Level: advanced

1815:    Notes: This routine returns some internal storage information
1816:    of the matrix, it is intended to be used by advanced users.
1817:    It should be called after the matrix is assembled.
1818:    The contents of the sizes[] array should not be changed.
1819:    PETSC_NULL may be passed for information not requested.

1821: .keywords: matrix, seqaij, get, inode

1823: .seealso: MatGetInfo()
1824: @*/
1825: PetscErrorCode PETSCMAT_DLLEXPORT MatInodeGetInodeSizes(Mat A,PetscInt *node_count,PetscInt *sizes[],PetscInt *limit)
1826: {
1827:   PetscErrorCode ierr,(*f)(Mat,PetscInt*,PetscInt*[],PetscInt*);

1830:   if (!A->assembled) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Not for unassembled matrix");
1831:   PetscObjectQueryFunction((PetscObject)A,"MatInodeGetInodeSizes_C",(void (**)(void))&f);
1832:   if (f) {
1833:     (*f)(A,node_count,sizes,limit);
1834:   }
1835:   return(0);
1836: }

1841: PetscErrorCode PETSCMAT_DLLEXPORT MatInodeGetInodeSizes_Inode(Mat A,PetscInt *node_count,PetscInt *sizes[],PetscInt *limit)
1842: {
1843:   Mat_inode *a = (Mat_inode*)A->data;

1846:   if (node_count) *node_count = a->inode.node_count;
1847:   if (sizes)      *sizes      = a->inode.size;
1848:   if (limit)      *limit      = a->inode.limit;
1849:   return(0);
1850: }