Actual source code: mpirowbs.c

  1: #define PETSCMAT_DLL

 3:  #include src/mat/impls/rowbs/mpi/mpirowbs.h

  5: #define CHUNCKSIZE_LOCAL   10

  9: static PetscErrorCode MatFreeRowbs_Private(Mat A,int n,int *i,PetscScalar *v)
 10: {

 14:   if (v) {
 15: #if defined(PETSC_USE_LOG)
 16:     int len = -n*(sizeof(int)+sizeof(PetscScalar));
 17: #endif
 18:     PetscFree(v);
 19:     PetscLogObjectMemory(A,len);
 20:   }
 21:   return(0);
 22: }

 26: static PetscErrorCode MatMallocRowbs_Private(Mat A,int n,int **i,PetscScalar **v)
 27: {
 29:   int len;

 32:   if (!n) {
 33:     *i = 0; *v = 0;
 34:   } else {
 35:     len = n*(sizeof(int) + sizeof(PetscScalar));
 36:     PetscMalloc(len,v);
 37:     PetscLogObjectMemory(A,len);
 38:     *i = (int*)(*v + n);
 39:   }
 40:   return(0);
 41: }

 45: PetscErrorCode MatScale_MPIRowbs(Mat inA,PetscScalar alpha)
 46: {
 47:   Mat_MPIRowbs   *a = (Mat_MPIRowbs*)inA->data;
 48:   BSspmat        *A = a->A;
 49:   BSsprow        *vs;
 50:   PetscScalar    *ap;
 51:   int            i,m = inA->rmap.n,nrow,j;

 55:   for (i=0; i<m; i++) {
 56:     vs   = A->rows[i];
 57:     nrow = vs->length;
 58:     ap   = vs->nz;
 59:     for (j=0; j<nrow; j++) {
 60:       ap[j] *= alpha;
 61:     }
 62:   }
 63:   PetscLogFlops(a->nz);
 64:   return(0);
 65: }

 67: /* ----------------------------------------------------------------- */
 70: static PetscErrorCode MatCreateMPIRowbs_local(Mat A,int nz,const int nnz[])
 71: {
 72:   Mat_MPIRowbs *bsif = (Mat_MPIRowbs*)A->data;
 74:   int   i,len,m = A->rmap.n,*tnnz;
 75:   BSspmat      *bsmat;
 76:   BSsprow      *vs;

 79:   PetscMalloc((m+1)*sizeof(int),&tnnz);
 80:   if (!nnz) {
 81:     if (nz == PETSC_DEFAULT || nz == PETSC_DECIDE) nz = 5;
 82:     if (nz <= 0)             nz = 1;
 83:     for (i=0; i<m; i++) tnnz[i] = nz;
 84:     nz      = nz*m;
 85:   } else {
 86:     nz = 0;
 87:     for (i=0; i<m; i++) {
 88:       if (nnz[i] <= 0) tnnz[i] = 1;
 89:       else             tnnz[i] = nnz[i];
 90:       nz += tnnz[i];
 91:     }
 92:   }

 94:   /* Allocate BlockSolve matrix context */
 95:   PetscNew(BSspmat,&bsif->A);
 96:   bsmat = bsif->A;
 97:   BSset_mat_icc_storage(bsmat,PETSC_FALSE);
 98:   BSset_mat_symmetric(bsmat,PETSC_FALSE);
 99:   len                    = m*(sizeof(BSsprow*)+ sizeof(BSsprow)) + 1;
100:   PetscMalloc(len,&bsmat->rows);
101:   bsmat->num_rows        = m;
102:   bsmat->global_num_rows = A->rmap.N;
103:   bsmat->map             = bsif->bsmap;
104:   vs                     = (BSsprow*)(bsmat->rows + m);
105:   for (i=0; i<m; i++) {
106:     bsmat->rows[i]  = vs;
107:     bsif->imax[i]   = tnnz[i];
108:     vs->diag_ind    = -1;
109:     MatMallocRowbs_Private(A,tnnz[i],&(vs->col),&(vs->nz));
110:     /* put zero on diagonal */
111:     /*vs->length            = 1;
112:     vs->col[0]      = i + bsif->rstart;
113:     vs->nz[0]       = 0.0;*/
114:     vs->length = 0;
115:     vs++;
116:   }
117:   PetscLogObjectMemory(A,sizeof(BSspmat) + len);
118:   bsif->nz               = 0;
119:   bsif->maxnz            = nz;
120:   bsif->sorted           = 0;
121:   bsif->roworiented      = PETSC_TRUE;
122:   bsif->nonew            = 0;
123:   bsif->bs_color_single  = 0;

125:   PetscFree(tnnz);
126:   return(0);
127: }

131: static PetscErrorCode MatSetValues_MPIRowbs_local(Mat AA,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode addv)
132: {
133:   Mat_MPIRowbs *mat = (Mat_MPIRowbs*)AA->data;
134:   BSspmat      *A = mat->A;
135:   BSsprow      *vs;
137:   int          *rp,k,a,b,t,ii,row,nrow,i,col,l,rmax;
138:   int          *imax = mat->imax,nonew = mat->nonew,sorted = mat->sorted;
139:   PetscScalar  *ap,value;

142:   for (k=0; k<m; k++) { /* loop over added rows */
143:     row = im[k];
144:     if (row < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %d",row);
145:     if (row >= AA->rmap.n) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",row,AA->rmap.n-1);
146:     vs   = A->rows[row];
147:     ap   = vs->nz; rp = vs->col;
148:     rmax = imax[row]; nrow = vs->length;
149:     a    = 0;
150:     for (l=0; l<n; l++) { /* loop over added columns */
151:       if (in[l] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative col: %d",in[l]);
152:       if (in[l] >= AA->cmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",in[l],AA->cmap.N-1);
153:       col = in[l]; value = *v++;
154:       if (!sorted) a = 0; b = nrow;
155:       while (b-a > 5) {
156:         t = (b+a)/2;
157:         if (rp[t] > col) b = t;
158:         else             a = t;
159:       }
160:       for (i=a; i<b; i++) {
161:         if (rp[i] > col) break;
162:         if (rp[i] == col) {
163:           if (addv == ADD_VALUES) ap[i] += value;
164:           else                    ap[i] = value;
165:           goto noinsert;
166:         }
167:       }
168:       if (nonew) goto noinsert;
169:       if (nrow >= rmax) {
170:         /* there is no extra room in row, therefore enlarge */
171:         int    *itemp,*iout,*iin = vs->col;
172:         PetscScalar *vout,*vin = vs->nz,*vtemp;

174:         /* malloc new storage space */
175:         imax[row] += CHUNCKSIZE_LOCAL;
176:         MatMallocRowbs_Private(AA,imax[row],&itemp,&vtemp);
177:         vout = vtemp; iout = itemp;
178:         for (ii=0; ii<i; ii++) {
179:           vout[ii] = vin[ii];
180:           iout[ii] = iin[ii];
181:         }
182:         vout[i] = value;
183:         iout[i] = col;
184:         for (ii=i+1; ii<=nrow; ii++) {
185:           vout[ii] = vin[ii-1];
186:           iout[ii] = iin[ii-1];
187:         }
188:         /* free old row storage */
189:         if (rmax > 0) {
190:           MatFreeRowbs_Private(AA,rmax,vs->col,vs->nz);
191:         }
192:         vs->col           =  iout; vs->nz = vout;
193:         rmax              =  imax[row];
194:         mat->maxnz        += CHUNCKSIZE_LOCAL;
195:         mat->reallocs++;
196:       } else {
197:         /* shift higher columns over to make room for newie */
198:         for (ii=nrow-1; ii>=i; ii--) {
199:           rp[ii+1] = rp[ii];
200:           ap[ii+1] = ap[ii];
201:         }
202:         rp[i] = col;
203:         ap[i] = value;
204:       }
205:       nrow++;
206:       mat->nz++;
207:       AA->same_nonzero = PETSC_FALSE;
208:       noinsert:;
209:       a = i + 1;
210:     }
211:     vs->length = nrow;
212:   }
213:   return(0);
214: }


219: static PetscErrorCode MatAssemblyBegin_MPIRowbs_local(Mat A,MatAssemblyType mode)
220: {
222:   return(0);
223: }

227: static PetscErrorCode MatAssemblyEnd_MPIRowbs_local(Mat AA,MatAssemblyType mode)
228: {
229:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)AA->data;
230:   BSspmat      *A = a->A;
231:   BSsprow      *vs;
232:   int          i,j,rstart = AA->rmap.rstart;

235:   if (mode == MAT_FLUSH_ASSEMBLY) return(0);

237:   /* Mark location of diagonal */
238:   for (i=0; i<AA->rmap.n; i++) {
239:     vs = A->rows[i];
240:     for (j=0; j<vs->length; j++) {
241:       if (vs->col[j] == i + rstart) {
242:         vs->diag_ind = j;
243:         break;
244:       }
245:     }
246:     if (vs->diag_ind == -1) {
247:       SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"no diagonal entry");
248:     }
249:   }
250:   return(0);
251: }

255: static PetscErrorCode MatZeroRows_MPIRowbs_local(Mat A,PetscInt N,const PetscInt rz[],PetscScalar diag)
256: {
257:   Mat_MPIRowbs   *a = (Mat_MPIRowbs*)A->data;
258:   BSspmat        *l = a->A;
260:   int            i,m = A->rmap.n - 1,col,base=A->rmap.rstart;

263:   if (a->keepzeroedrows) {
264:     for (i=0; i<N; i++) {
265:       if (rz[i] < 0 || rz[i] > m) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"row out of range");
266:       PetscMemzero(l->rows[rz[i]]->nz,l->rows[rz[i]]->length*sizeof(PetscScalar));
267:       if (diag != 0.0) {
268:         col=rz[i]+base;
269:         MatSetValues_MPIRowbs_local(A,1,&rz[i],1,&col,&diag,INSERT_VALUES);
270:       }
271:     }
272:   } else {
273:     if (diag != 0.0) {
274:       for (i=0; i<N; i++) {
275:         if (rz[i] < 0 || rz[i] > m) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Out of range");
276:         if (l->rows[rz[i]]->length > 0) { /* in case row was completely empty */
277:           l->rows[rz[i]]->length = 1;
278:           l->rows[rz[i]]->nz[0]  = diag;
279:           l->rows[rz[i]]->col[0] = A->rmap.rstart + rz[i];
280:         } else {
281:           col=rz[i]+base;
282:           MatSetValues_MPIRowbs_local(A,1,&rz[i],1,&col,&diag,INSERT_VALUES);
283:         }
284:       }
285:     } else {
286:       for (i=0; i<N; i++) {
287:         if (rz[i] < 0 || rz[i] > m) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Out of range");
288:         l->rows[rz[i]]->length = 0;
289:       }
290:     }
291:     A->same_nonzero = PETSC_FALSE;
292:   }
293:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
294:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
295:   return(0);
296: }

300: static PetscErrorCode MatNorm_MPIRowbs_local(Mat A,NormType type,PetscReal *norm)
301: {
302:   Mat_MPIRowbs *mat = (Mat_MPIRowbs*)A->data;
303:   BSsprow      *vs,**rs;
304:   PetscScalar  *xv;
305:   PetscReal    sum = 0.0;
307:   int          *xi,nz,i,j;

310:   rs = mat->A->rows;
311:   if (type == NORM_FROBENIUS) {
312:     for (i=0; i<A->rmap.n; i++) {
313:       vs = *rs++;
314:       nz = vs->length;
315:       xv = vs->nz;
316:       while (nz--) {
317: #if defined(PETSC_USE_COMPLEX)
318:         sum += PetscRealPart(PetscConj(*xv)*(*xv)); xv++;
319: #else
320:         sum += (*xv)*(*xv); xv++;
321: #endif
322:       }
323:     }
324:     *norm = sqrt(sum);
325:   } else if (type == NORM_1) { /* max column norm */
326:     PetscReal *tmp;
327:     PetscMalloc(A->cmap.n*sizeof(PetscReal),&tmp);
328:     PetscMemzero(tmp,A->cmap.n*sizeof(PetscReal));
329:     *norm = 0.0;
330:     for (i=0; i<A->rmap.n; i++) {
331:       vs = *rs++;
332:       nz = vs->length;
333:       xi = vs->col;
334:       xv = vs->nz;
335:       while (nz--) {
336:         tmp[*xi] += PetscAbsScalar(*xv);
337:         xi++; xv++;
338:       }
339:     }
340:     for (j=0; j<A->rmap.n; j++) {
341:       if (tmp[j] > *norm) *norm = tmp[j];
342:     }
343:     PetscFree(tmp);
344:   } else if (type == NORM_INFINITY) { /* max row norm */
345:     *norm = 0.0;
346:     for (i=0; i<A->rmap.n; i++) {
347:       vs = *rs++;
348:       nz = vs->length;
349:       xv = vs->nz;
350:       sum = 0.0;
351:       while (nz--) {
352:         sum += PetscAbsScalar(*xv); xv++;
353:       }
354:       if (sum > *norm) *norm = sum;
355:     }
356:   } else {
357:     SETERRQ(PETSC_ERR_SUP,"No support for the two norm");
358:   }
359:   return(0);
360: }

362: /* ----------------------------------------------------------------- */

366: PetscErrorCode MatSetValues_MPIRowbs(Mat mat,int m,const int im[],int n,const int in[],const PetscScalar v[],InsertMode av)
367: {
368:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
370:   int   i,j,row,col,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
371:   PetscTruth   roworiented = a->roworiented;

374:   /* Note:  There's no need to "unscale" the matrix, since scaling is
375:      confined to a->pA, and we're working with a->A here */
376:   for (i=0; i<m; i++) {
377:     if (im[i] < 0) continue;
378:     if (im[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %d max %d",im[i],mat->rmap.N-1);
379:     if (im[i] >= rstart && im[i] < rend) {
380:       row = im[i] - rstart;
381:       for (j=0; j<n; j++) {
382:         if (in[j] < 0) continue;
383:         if (in[j] >= mat->cmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %d max %d",in[j],mat->cmap.N-1);
384:         if (in[j] >= 0 && in[j] < mat->cmap.N){
385:           col = in[j];
386:           if (roworiented) {
387:             MatSetValues_MPIRowbs_local(mat,1,&row,1,&col,v+i*n+j,av);
388:           } else {
389:             MatSetValues_MPIRowbs_local(mat,1,&row,1,&col,v+i+j*m,av);
390:           }
391:         } else {SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Invalid column");}
392:       }
393:     } else {
394:       if (!a->donotstash) {
395:         if (roworiented) {
396:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
397:         } else {
398:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
399:         }
400:       }
401:     }
402:   }
403:   return(0);
404: }

408: PetscErrorCode MatAssemblyBegin_MPIRowbs(Mat mat,MatAssemblyType mode)
409: {
410:   MPI_Comm      comm = mat->comm;
412:   int         nstash,reallocs;
413:   InsertMode    addv;

416:   /* Note:  There's no need to "unscale" the matrix, since scaling is
417:             confined to a->pA, and we're working with a->A here */

419:   /* make sure all processors are either in INSERTMODE or ADDMODE */
420:   MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,comm);
421:   if (addv == (ADD_VALUES|INSERT_VALUES)) {
422:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some procs inserted; others added");
423:   }
424:   mat->insertmode = addv; /* in case this processor had no cache */

426:   MatStashScatterBegin_Private(&mat->stash,mat->rmap.range);
427:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
428:   PetscInfo2(0,"Block-Stash has %d entries, uses %d mallocs.\n",nstash,reallocs);
429:   return(0);
430: }

434: static PetscErrorCode MatView_MPIRowbs_ASCII(Mat mat,PetscViewer viewer)
435: {
436:   Mat_MPIRowbs      *a = (Mat_MPIRowbs*)mat->data;
438:   int               i,j;
439:   PetscTruth        iascii;
440:   BSspmat           *A = a->A;
441:   BSsprow           **rs = A->rows;
442:   PetscViewerFormat format;

445:   PetscViewerGetFormat(viewer,&format);
446:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);

448:   if (format == PETSC_VIEWER_ASCII_INFO || format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
449:     int ind_l,ind_g,clq_l,clq_g,color;
450:     ind_l = BSlocal_num_inodes(a->pA);CHKERRBS(0);
451:     ind_g = BSglobal_num_inodes(a->pA);CHKERRBS(0);
452:     clq_l = BSlocal_num_cliques(a->pA);CHKERRBS(0);
453:     clq_g = BSglobal_num_cliques(a->pA);CHKERRBS(0);
454:     color = BSnum_colors(a->pA);CHKERRBS(0);
455:     PetscViewerASCIIPrintf(viewer,"  %d global inode(s), %d global clique(s), %d color(s)\n",ind_g,clq_g,color);
456:     PetscViewerASCIISynchronizedPrintf(viewer,"    [%d] %d local inode(s), %d local clique(s)\n",a->rank,ind_l,clq_l);
457:   } else  if (format == PETSC_VIEWER_ASCII_COMMON) {
458:     for (i=0; i<A->num_rows; i++) {
459:       PetscViewerASCIISynchronizedPrintf(viewer,"row %d:",i+mat->rmap.rstart);
460:       for (j=0; j<rs[i]->length; j++) {
461:         if (rs[i]->nz[j]) {PetscViewerASCIISynchronizedPrintf(viewer," %d %g ",rs[i]->col[j],rs[i]->nz[j]);}
462:       }
463:       PetscViewerASCIISynchronizedPrintf(viewer,"\n");
464:     }
465:   } else if (format == PETSC_VIEWER_ASCII_MATLAB) {
466:     SETERRQ(PETSC_ERR_SUP,"Matlab format not supported");
467:   } else {
468:     PetscViewerASCIIUseTabs(viewer,PETSC_NO);
469:     for (i=0; i<A->num_rows; i++) {
470:       PetscViewerASCIISynchronizedPrintf(viewer,"row %d:",i+mat->rmap.rstart);
471:       for (j=0; j<rs[i]->length; j++) {
472:         PetscViewerASCIISynchronizedPrintf(viewer," %d %g ",rs[i]->col[j],rs[i]->nz[j]);
473:       }
474:       PetscViewerASCIISynchronizedPrintf(viewer,"\n");
475:     }
476:     PetscViewerASCIIUseTabs(viewer,PETSC_YES);
477:   }
478:   PetscViewerFlush(viewer);
479:   return(0);
480: }

484: static PetscErrorCode MatView_MPIRowbs_Binary(Mat mat,PetscViewer viewer)
485: {
486:   Mat_MPIRowbs   *a = (Mat_MPIRowbs*)mat->data;
488:   PetscMPIInt    rank,size;
489:   PetscInt       i,M,m,*sbuff,*rowlengths;
490:   PetscInt       *recvcts,*recvdisp,fd,*cols,maxnz,nz,j;
491:   BSspmat        *A = a->A;
492:   BSsprow        **rs = A->rows;
493:   MPI_Comm       comm = mat->comm;
494:   MPI_Status     status;
495:   PetscScalar    *vals;
496:   MatInfo        info;

499:   MPI_Comm_size(comm,&size);
500:   MPI_Comm_rank(comm,&rank);

502:   M = mat->rmap.N; m = mat->rmap.n;
503:   /* First gather together on the first processor the lengths of 
504:      each row, and write them out to the file */
505:   PetscMalloc(m*sizeof(int),&sbuff);
506:   for (i=0; i<A->num_rows; i++) {
507:     sbuff[i] = rs[i]->length;
508:   }
509:   MatGetInfo(mat,MAT_GLOBAL_SUM,&info);
510:   if (!rank) {
511:     PetscViewerBinaryGetDescriptor(viewer,&fd);
512:     PetscMalloc((4+M)*sizeof(int),&rowlengths);
513:     PetscMalloc(size*sizeof(int),&recvcts);
514:     recvdisp = mat->rmap.range;
515:     for (i=0; i<size; i++) {
516:       recvcts[i] = recvdisp[i+1] - recvdisp[i];
517:     }
518:     /* first four elements of rowlength are the header */
519:     rowlengths[0] = mat->cookie;
520:     rowlengths[1] = mat->rmap.N;
521:     rowlengths[2] = mat->cmap.N;
522:     rowlengths[3] = (int)info.nz_used;
523:     MPI_Gatherv(sbuff,m,MPI_INT,rowlengths+4,recvcts,recvdisp,MPI_INT,0,comm);
524:     PetscFree(sbuff);
525:     PetscBinaryWrite(fd,rowlengths,4+M,PETSC_INT,PETSC_FALSE);
526:     /* count the number of nonzeros on each processor */
527:     PetscMemzero(recvcts,size*sizeof(int));
528:     for (i=0; i<size; i++) {
529:       for (j=recvdisp[i]; j<recvdisp[i+1]; j++) {
530:         recvcts[i] += rowlengths[j+3];
531:       }
532:     }
533:     /* allocate buffer long enough to hold largest one */
534:     maxnz = 0;
535:     for (i=0; i<size; i++) {
536:       maxnz = PetscMax(maxnz,recvcts[i]);
537:     }
538:     PetscFree(rowlengths);
539:     PetscFree(recvcts);
540:     PetscMalloc(maxnz*sizeof(int),&cols);

542:     /* binary store column indices for 0th processor */
543:     nz = 0;
544:     for (i=0; i<A->num_rows; i++) {
545:       for (j=0; j<rs[i]->length; j++) {
546:         cols[nz++] = rs[i]->col[j];
547:       }
548:     }
549:     PetscBinaryWrite(fd,cols,nz,PETSC_INT,PETSC_FALSE);

551:     /* receive and store column indices for all other processors */
552:     for (i=1; i<size; i++) {
553:       /* should tell processor that I am now ready and to begin the send */
554:       MPI_Recv(cols,maxnz,MPI_INT,i,mat->tag,comm,&status);
555:       MPI_Get_count(&status,MPI_INT,&nz);
556:       PetscBinaryWrite(fd,cols,nz,PETSC_INT,PETSC_FALSE);
557:     }
558:     PetscFree(cols);
559:     PetscMalloc(maxnz*sizeof(PetscScalar),&vals);

561:     /* binary store values for 0th processor */
562:     nz = 0;
563:     for (i=0; i<A->num_rows; i++) {
564:       for (j=0; j<rs[i]->length; j++) {
565:         vals[nz++] = rs[i]->nz[j];
566:       }
567:     }
568:     PetscBinaryWrite(fd,vals,nz,PETSC_SCALAR,PETSC_FALSE);

570:     /* receive and store nonzeros for all other processors */
571:     for (i=1; i<size; i++) {
572:       /* should tell processor that I am now ready and to begin the send */
573:       MPI_Recv(vals,maxnz,MPIU_SCALAR,i,mat->tag,comm,&status);
574:       MPI_Get_count(&status,MPIU_SCALAR,&nz);
575:       PetscBinaryWrite(fd,vals,nz,PETSC_SCALAR,PETSC_FALSE);
576:     }
577:     PetscFree(vals);
578:   } else {
579:     MPI_Gatherv(sbuff,m,MPI_INT,0,0,0,MPI_INT,0,comm);
580:     PetscFree(sbuff);

582:     /* count local nonzeros */
583:     nz = 0;
584:     for (i=0; i<A->num_rows; i++) {
585:       for (j=0; j<rs[i]->length; j++) {
586:         nz++;
587:       }
588:     }
589:     /* copy into buffer column indices */
590:     PetscMalloc(nz*sizeof(int),&cols);
591:     nz = 0;
592:     for (i=0; i<A->num_rows; i++) {
593:       for (j=0; j<rs[i]->length; j++) {
594:         cols[nz++] = rs[i]->col[j];
595:       }
596:     }
597:     /* send */  /* should wait until processor zero tells me to go */
598:     MPI_Send(cols,nz,MPI_INT,0,mat->tag,comm);
599:     PetscFree(cols);

601:     /* copy into buffer column values */
602:     PetscMalloc(nz*sizeof(PetscScalar),&vals);
603:     nz   = 0;
604:     for (i=0; i<A->num_rows; i++) {
605:       for (j=0; j<rs[i]->length; j++) {
606:         vals[nz++] = rs[i]->nz[j];
607:       }
608:     }
609:     /* send */  /* should wait until processor zero tells me to go */
610:     MPI_Send(vals,nz,MPIU_SCALAR,0,mat->tag,comm);
611:     PetscFree(vals);
612:   }

614:   return(0);
615: }

619: PetscErrorCode MatView_MPIRowbs(Mat mat,PetscViewer viewer)
620: {
621:   Mat_MPIRowbs *bsif = (Mat_MPIRowbs*)mat->data;
623:   PetscTruth   iascii,isbinary;

626:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
627:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
628:   if (!bsif->blocksolveassembly) {
629:     MatAssemblyEnd_MPIRowbs_ForBlockSolve(mat);
630:   }
631:   if (iascii) {
632:     MatView_MPIRowbs_ASCII(mat,viewer);
633:   } else if (isbinary) {
634:     MatView_MPIRowbs_Binary(mat,viewer);
635:   } else {
636:     SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPIRowbs matrices",((PetscObject)viewer)->type_name);
637:   }
638:   return(0);
639: }
640: 
643: static PetscErrorCode MatAssemblyEnd_MPIRowbs_MakeSymmetric(Mat mat)
644: {
645:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
646:   BSspmat      *A = a->A;
647:   BSsprow      *vs;
648:   int          size,rank,M,rstart,tag,i,j,*rtable,*w1,*w3,*w4,len,proc,nrqs;
649:   int          msz,*pa,bsz,nrqr,**rbuf1,**sbuf1,**ptr,*tmp,*ctr,col,idx,row;
651:   int          ctr_j,*sbuf1_j,k;
652:   PetscScalar  val=0.0;
653:   MPI_Comm     comm;
654:   MPI_Request  *s_waits1,*r_waits1;
655:   MPI_Status   *s_status,*r_status;

658:   comm   = mat->comm;
659:   tag    = mat->tag;
660:   size   = a->size;
661:   rank   = a->rank;
662:   M      = mat->rmap.N;
663:   rstart = mat->rmap.rstart;

665:   PetscMalloc(M*sizeof(int),&rtable);
666:   /* Create hash table for the mapping :row -> proc */
667:   for (i=0,j=0; i<size; i++) {
668:     len = mat->rmap.range[i+1];
669:     for (; j<len; j++) {
670:       rtable[j] = i;
671:     }
672:   }

674:   /* Evaluate communication - mesg to whom, length of mesg, and buffer space
675:      required. Based on this, buffers are allocated, and data copied into them. */
676:   PetscMalloc(size*4*sizeof(int),&w1);/*  mesg size */
677:   w3   = w1 + 2*size;       /* no of IS that needs to be sent to proc i */
678:   w4   = w3 + size;       /* temp work space used in determining w1,  w3 */
679:   PetscMemzero(w1,size*3*sizeof(int)); /* initialize work vector */

681:   for (i=0;  i<mat->rmap.n; i++) {
682:     PetscMemzero(w4,size*sizeof(int)); /* initialize work vector */
683:     vs = A->rows[i];
684:     for (j=0; j<vs->length; j++) {
685:       proc = rtable[vs->col[j]];
686:       w4[proc]++;
687:     }
688:     for (j=0; j<size; j++) {
689:       if (w4[j]) { w1[2*j] += w4[j]; w3[j]++;}
690:     }
691:   }
692: 
693:   nrqs       = 0;              /* number of outgoing messages */
694:   msz        = 0;              /* total mesg length (for all proc */
695:   w1[2*rank] = 0;              /* no mesg sent to itself */
696:   w3[rank]   = 0;
697:   for (i=0; i<size; i++) {
698:     if (w1[2*i])  {w1[2*i+1] = 1; nrqs++;} /* there exists a message to proc i */
699:   }
700:   /* pa - is list of processors to communicate with */
701:   PetscMalloc((nrqs+1)*sizeof(int),&pa);
702:   for (i=0,j=0; i<size; i++) {
703:     if (w1[2*i]) {pa[j] = i; j++;}
704:   }

706:   /* Each message would have a header = 1 + 2*(no of ROWS) + data */
707:   for (i=0; i<nrqs; i++) {
708:     j       = pa[i];
709:     w1[2*j] += w1[2*j+1] + 2*w3[j];
710:     msz     += w1[2*j];
711:   }
712: 
713:   /* Do a global reduction to determine how many messages to expect */
714:   PetscMaxSum(comm,w1,&bsz,&nrqr);

716:   /* Allocate memory for recv buffers . Prob none if nrqr = 0 ???? */
717:   len      = (nrqr+1)*sizeof(int*) + nrqr*bsz*sizeof(int);
718:   PetscMalloc(len,&rbuf1);
719:   rbuf1[0] = (int*)(rbuf1 + nrqr);
720:   for (i=1; i<nrqr; ++i) rbuf1[i] = rbuf1[i-1] + bsz;

722:   /* Post the receives */
723:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&r_waits1);
724:   for (i=0; i<nrqr; ++i){
725:     MPI_Irecv(rbuf1[i],bsz,MPI_INT,MPI_ANY_SOURCE,tag,comm,r_waits1+i);
726:   }
727: 
728:   /* Allocate Memory for outgoing messages */
729:   len   = 2*size*sizeof(int*) + (size+msz)*sizeof(int);
730:   PetscMalloc(len,&sbuf1);
731:   ptr   = sbuf1 + size;     /* Pointers to the data in outgoing buffers */
732:   PetscMemzero(sbuf1,2*size*sizeof(int*));
733:   tmp   = (int*)(sbuf1 + 2*size);
734:   ctr   = tmp + msz;

736:   {
737:     int *iptr = tmp,ict  = 0;
738:     for (i=0; i<nrqs; i++) {
739:       j        = pa[i];
740:       iptr    += ict;
741:       sbuf1[j] = iptr;
742:       ict      = w1[2*j];
743:     }
744:   }

746:   /* Form the outgoing messages */
747:   /* Clean up the header space */
748:   for (i=0; i<nrqs; i++) {
749:     j           = pa[i];
750:     sbuf1[j][0] = 0;
751:     PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));
752:     ptr[j]      = sbuf1[j] + 2*w3[j] + 1;
753:   }

755:   /* Parse the matrix and copy the data into sbuf1 */
756:   for (i=0; i<mat->rmap.n; i++) {
757:     PetscMemzero(ctr,size*sizeof(int));
758:     vs = A->rows[i];
759:     for (j=0; j<vs->length; j++) {
760:       col  = vs->col[j];
761:       proc = rtable[col];
762:       if (proc != rank) { /* copy to the outgoing buffer */
763:         ctr[proc]++;
764:           *ptr[proc] = col;
765:           ptr[proc]++;
766:       } else {
767:         row = col - rstart;
768:         col = i + rstart;
769:         MatSetValues_MPIRowbs_local(mat,1,&row,1,&col,&val,ADD_VALUES);
770:       }
771:     }
772:     /* Update the headers for the current row */
773:     for (j=0; j<size; j++) { /* Can Optimise this loop by using pa[] */
774:       if ((ctr_j = ctr[j])) {
775:         sbuf1_j        = sbuf1[j];
776:         k               = ++sbuf1_j[0];
777:         sbuf1_j[2*k]   = ctr_j;
778:         sbuf1_j[2*k-1] = i + rstart;
779:       }
780:     }
781:   }
782:    /* Check Validity of the outgoing messages */
783:   {
784:     int sum;
785:     for (i=0 ; i<nrqs ; i++) {
786:       j = pa[i];
787:       if (w3[j] != sbuf1[j][0]) {SETERRQ(PETSC_ERR_PLIB,"Blew it! Header[1] mismatch!\n"); }
788:     }

790:     for (i=0 ; i<nrqs ; i++) {
791:       j = pa[i];
792:       sum = 1;
793:       for (k = 1; k <= w3[j]; k++) sum += sbuf1[j][2*k]+2;
794:       if (sum != w1[2*j]) { SETERRQ(PETSC_ERR_PLIB,"Blew it! Header[2-n] mismatch!\n"); }
795:     }
796:   }
797: 
798:   /* Now post the sends */
799:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);
800:   for (i=0; i<nrqs; ++i) {
801:     j    = pa[i];
802:     MPI_Isend(sbuf1[j],w1[2*j],MPI_INT,j,tag,comm,s_waits1+i);
803:   }
804: 
805:   /* Receive messages*/
806:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status);
807:   for (i=0; i<nrqr; ++i) {
808:     MPI_Waitany(nrqr,r_waits1,&idx,r_status+i);
809:     /* Process the Message */
810:     {
811:       int    *rbuf1_i,n_row,ct1;

813:       rbuf1_i = rbuf1[idx];
814:       n_row   = rbuf1_i[0];
815:       ct1     = 2*n_row+1;
816:       val     = 0.0;
817:       /* Optimise this later */
818:       for (j=1; j<=n_row; j++) {
819:         col = rbuf1_i[2*j-1];
820:         for (k=0; k<rbuf1_i[2*j]; k++,ct1++) {
821:           row = rbuf1_i[ct1] - rstart;
822:           MatSetValues_MPIRowbs_local(mat,1,&row,1,&col,&val,ADD_VALUES);
823:         }
824:       }
825:     }
826:   }

828:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status);
829:   if (nrqs) {MPI_Waitall(nrqs,s_waits1,s_status);}

831:   PetscFree(rtable);
832:   PetscFree(w1);
833:   PetscFree(pa);
834:   PetscFree(rbuf1);
835:   PetscFree(sbuf1);
836:   PetscFree(r_waits1);
837:   PetscFree(s_waits1);
838:   PetscFree(r_status);
839:   PetscFree(s_status);
840:   return(0);
841: }

843: /*
844:      This does the BlockSolve portion of the matrix assembly.
845:    It is provided in a separate routine so that users can
846:    operate on the matrix (using MatScale(), MatShift() etc.) after 
847:    the matrix has been assembled but before BlockSolve has sucked it
848:    in and devoured it.
849: */
852: PetscErrorCode MatAssemblyEnd_MPIRowbs_ForBlockSolve(Mat mat)
853: {
854:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
856:   int          ldim,low,high,i;
857:   PetscScalar  *diag;

860:   if ((mat->was_assembled) && (!mat->same_nonzero)) {  /* Free the old info */
861:     if (a->pA)       {BSfree_par_mat(a->pA);CHKERRBS(0);}
862:     if (a->comm_pA)  {BSfree_comm(a->comm_pA);CHKERRBS(0);}
863:   }

865:   if ((!mat->same_nonzero) || (!mat->was_assembled)) {
866:     /* Indicates bypassing cliques in coloring */
867:     if (a->bs_color_single) {
868:       BSctx_set_si(a->procinfo,100);
869:     }
870:     /* Form permuted matrix for efficient parallel execution */
871:     a->pA = BSmain_perm(a->procinfo,a->A);CHKERRBS(0);
872:     /* Set up the communication */
873:     a->comm_pA = BSsetup_forward(a->pA,a->procinfo);CHKERRBS(0);
874:   } else {
875:     /* Repermute the matrix */
876:     BSmain_reperm(a->procinfo,a->A,a->pA);CHKERRBS(0);
877:   }

879:   /* Symmetrically scale the matrix by the diagonal */
880:   BSscale_diag(a->pA,a->pA->diag,a->procinfo);CHKERRBS(0);

882:   /* Store inverse of square root of permuted diagonal scaling matrix */
883:   VecGetLocalSize(a->diag,&ldim);
884:   VecGetOwnershipRange(a->diag,&low,&high);
885:   VecGetArray(a->diag,&diag);
886:   for (i=0; i<ldim; i++) {
887:     if (a->pA->scale_diag[i] != 0.0) {
888:       diag[i] = 1.0/sqrt(PetscAbsScalar(a->pA->scale_diag[i]));
889:     } else {
890:       diag[i] = 1.0;
891:     }
892:   }
893:   VecRestoreArray(a->diag,&diag);
894:   a->assembled_icc_storage = a->A->icc_storage;
895:   a->blocksolveassembly = 1;
896:   mat->was_assembled    = PETSC_TRUE;
897:   mat->same_nonzero     = PETSC_TRUE;
898:   PetscInfo(mat,"Completed BlockSolve95 matrix assembly\n");
899:   return(0);
900: }

904: PetscErrorCode MatAssemblyEnd_MPIRowbs(Mat mat,MatAssemblyType mode)
905: {
906:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
908:   int          i,n,row,col,*rows,*cols,rstart,nzcount,flg,j,ncols;
909:   PetscScalar  *vals,val;
910:   InsertMode   addv = mat->insertmode;

913:   while (1) {
914:     MatStashScatterGetMesg_Private(&mat->stash,&n,&rows,&cols,&vals,&flg);
915:     if (!flg) break;
916: 
917:     for (i=0; i<n;) {
918:       /* Now identify the consecutive vals belonging to the same row */
919:       for (j=i,rstart=rows[j]; j<n; j++) { if (rows[j] != rstart) break; }
920:       if (j < n) ncols = j-i;
921:       else       ncols = n-i;
922:       /* Now assemble all these values with a single function call */
923:       MatSetValues_MPIRowbs(mat,1,rows+i,ncols,cols+i,vals+i,addv);
924:       i = j;
925:     }
926:   }
927:   MatStashScatterEnd_Private(&mat->stash);

929:   rstart = mat->rmap.rstart;
930:   nzcount = a->nz; /* This is the number of nonzeros entered by the user */
931:   /* BlockSolve requires that the matrix is structurally symmetric */
932:   if (mode == MAT_FINAL_ASSEMBLY && !mat->structurally_symmetric) {
933:     MatAssemblyEnd_MPIRowbs_MakeSymmetric(mat);
934:   }
935: 
936:   /* BlockSolve requires that all the diagonal elements are set */
937:   val  = 0.0;
938:   for (i=0; i<mat->rmap.n; i++) {
939:     row = i; col = i + rstart;
940:     MatSetValues_MPIRowbs_local(mat,1,&row,1,&col,&val,ADD_VALUES);
941:   }
942: 
943:   MatAssemblyBegin_MPIRowbs_local(mat,mode);
944:   MatAssemblyEnd_MPIRowbs_local(mat,mode);
945: 
946:   a->blocksolveassembly = 0;
947:   PetscInfo4(mat,"Matrix size: %d X %d; storage space: %d unneeded,%d used\n",mat->rmap.n,mat->cmap.n,a->maxnz-a->nz,a->nz);
948:   PetscInfo2(mat,"User entered %d nonzeros, PETSc added %d\n",nzcount,a->nz-nzcount);
949:   PetscInfo1(mat,"Number of mallocs during MatSetValues is %d\n",a->reallocs);
950:   return(0);
951: }

955: PetscErrorCode MatZeroEntries_MPIRowbs(Mat mat)
956: {
957:   Mat_MPIRowbs *l = (Mat_MPIRowbs*)mat->data;
958:   BSspmat      *A = l->A;
959:   BSsprow      *vs;
960:   int          i,j;

963:   for (i=0; i <mat->rmap.n; i++) {
964:     vs = A->rows[i];
965:     for (j=0; j< vs->length; j++) vs->nz[j] = 0.0;
966:   }
967:   return(0);
968: }

970: /* the code does not do the diagonal entries correctly unless the 
971:    matrix is square and the column and row owerships are identical.
972:    This is a BUG.
973: */

977: PetscErrorCode MatZeroRows_MPIRowbs(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag)
978: {
979:   Mat_MPIRowbs   *l = (Mat_MPIRowbs*)A->data;
981:   int            i,*owners = A->rmap.range,size = l->size;
982:   int            *nprocs,j,idx,nsends;
983:   int            nmax,*svalues,*starts,*owner,nrecvs,rank = l->rank;
984:   int            *rvalues,tag = A->tag,count,base,slen,n,*source;
985:   int            *lens,imdex,*lrows,*values;
986:   MPI_Comm       comm = A->comm;
987:   MPI_Request    *send_waits,*recv_waits;
988:   MPI_Status     recv_status,*send_status;
989:   PetscTruth     found;

992:   /*  first count number of contributors to each processor */
993:   PetscMalloc(2*size*sizeof(int),&nprocs);
994:   PetscMemzero(nprocs,2*size*sizeof(int));
995:   PetscMalloc((N+1)*sizeof(int),&owner); /* see note*/
996:   for (i=0; i<N; i++) {
997:     idx   = rows[i];
998:     found = PETSC_FALSE;
999:     for (j=0; j<size; j++) {
1000:       if (idx >= owners[j] && idx < owners[j+1]) {
1001:         nprocs[2*j]++; nprocs[2*j+1] = 1; owner[i] = j; found = PETSC_TRUE; break;
1002:       }
1003:     }
1004:     if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Row out of range");
1005:   }
1006:   nsends = 0;  for (i=0; i<size; i++) {nsends += nprocs[2*i+1];}

1008:   /* inform other processors of number of messages and max length*/
1009:   PetscMaxSum(comm,nprocs,&nmax,&nrecvs);

1011:   /* post receives:   */
1012:   PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(int),&rvalues);
1013:   PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
1014:   for (i=0; i<nrecvs; i++) {
1015:     MPI_Irecv(rvalues+nmax*i,nmax,MPI_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
1016:   }

1018:   /* do sends:
1019:       1) starts[i] gives the starting index in svalues for stuff going to 
1020:          the ith processor
1021:   */
1022:   PetscMalloc((N+1)*sizeof(int),&svalues);
1023:   PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
1024:   PetscMalloc((size+1)*sizeof(int),&starts);
1025:   starts[0] = 0;
1026:   for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
1027:   for (i=0; i<N; i++) {
1028:     svalues[starts[owner[i]]++] = rows[i];
1029:   }

1031:   starts[0] = 0;
1032:   for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
1033:   count = 0;
1034:   for (i=0; i<size; i++) {
1035:     if (nprocs[2*i+1]) {
1036:       MPI_Isend(svalues+starts[i],nprocs[2*i],MPI_INT,i,tag,comm,send_waits+count++);
1037:     }
1038:   }
1039:   PetscFree(starts);

1041:   base = owners[rank];

1043:   /*  wait on receives */
1044:   PetscMalloc(2*(nrecvs+1)*sizeof(int),&lens);
1045:   source = lens + nrecvs;
1046:   count = nrecvs; slen = 0;
1047:   while (count) {
1048:     MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
1049:     /* unpack receives into our local space */
1050:     MPI_Get_count(&recv_status,MPI_INT,&n);
1051:     source[imdex]  = recv_status.MPI_SOURCE;
1052:     lens[imdex]    = n;
1053:     slen           += n;
1054:     count--;
1055:   }
1056:   PetscFree(recv_waits);
1057: 
1058:   /* move the data into the send scatter */
1059:   PetscMalloc((slen+1)*sizeof(int),&lrows);
1060:   count = 0;
1061:   for (i=0; i<nrecvs; i++) {
1062:     values = rvalues + i*nmax;
1063:     for (j=0; j<lens[i]; j++) {
1064:       lrows[count++] = values[j] - base;
1065:     }
1066:   }
1067:   PetscFree(rvalues);
1068:   PetscFree(lens);
1069:   PetscFree(owner);
1070:   PetscFree(nprocs);
1071: 
1072:   /* actually zap the local rows */
1073:   MatZeroRows_MPIRowbs_local(A,slen,lrows,diag);
1074:   PetscFree(lrows);

1076:   /* wait on sends */
1077:   if (nsends) {
1078:     PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
1079:     MPI_Waitall(nsends,send_waits,send_status);
1080:     PetscFree(send_status);
1081:   }
1082:   PetscFree(send_waits);
1083:   PetscFree(svalues);

1085:   return(0);
1086: }

1090: PetscErrorCode MatNorm_MPIRowbs(Mat mat,NormType type,PetscReal *norm)
1091: {
1092:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
1093:   BSsprow      *vs,**rs;
1094:   PetscScalar  *xv;
1095:   PetscReal    sum = 0.0;
1097:   int          *xi,nz,i,j;

1100:   if (a->size == 1) {
1101:     MatNorm_MPIRowbs_local(mat,type,norm);
1102:   } else {
1103:     rs = a->A->rows;
1104:     if (type == NORM_FROBENIUS) {
1105:       for (i=0; i<mat->rmap.n; i++) {
1106:         vs = *rs++;
1107:         nz = vs->length;
1108:         xv = vs->nz;
1109:         while (nz--) {
1110: #if defined(PETSC_USE_COMPLEX)
1111:           sum += PetscRealPart(PetscConj(*xv)*(*xv)); xv++;
1112: #else
1113:           sum += (*xv)*(*xv); xv++;
1114: #endif
1115:         }
1116:       }
1117:       MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPI_SUM,mat->comm);
1118:       *norm = sqrt(*norm);
1119:     } else if (type == NORM_1) { /* max column norm */
1120:       PetscReal *tmp,*tmp2;
1121:       PetscMalloc(mat->cmap.n*sizeof(PetscReal),&tmp);
1122:       PetscMalloc(mat->cmap.n*sizeof(PetscReal),&tmp2);
1123:       PetscMemzero(tmp,mat->cmap.n*sizeof(PetscReal));
1124:       *norm = 0.0;
1125:       for (i=0; i<mat->rmap.n; i++) {
1126:         vs = *rs++;
1127:         nz = vs->length;
1128:         xi = vs->col;
1129:         xv = vs->nz;
1130:         while (nz--) {
1131:           tmp[*xi] += PetscAbsScalar(*xv);
1132:           xi++; xv++;
1133:         }
1134:       }
1135:       MPI_Allreduce(tmp,tmp2,mat->cmap.N,MPIU_REAL,MPI_SUM,mat->comm);
1136:       for (j=0; j<mat->cmap.n; j++) {
1137:         if (tmp2[j] > *norm) *norm = tmp2[j];
1138:       }
1139:       PetscFree(tmp);
1140:       PetscFree(tmp2);
1141:     } else if (type == NORM_INFINITY) { /* max row norm */
1142:       PetscReal ntemp = 0.0;
1143:       for (i=0; i<mat->rmap.n; i++) {
1144:         vs = *rs++;
1145:         nz = vs->length;
1146:         xv = vs->nz;
1147:         sum = 0.0;
1148:         while (nz--) {
1149:           sum += PetscAbsScalar(*xv); xv++;
1150:         }
1151:         if (sum > ntemp) ntemp = sum;
1152:       }
1153:       MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPI_MAX,mat->comm);
1154:     } else {
1155:       SETERRQ(PETSC_ERR_SUP,"No support for two norm");
1156:     }
1157:   }
1158:   return(0);
1159: }

1163: PetscErrorCode MatMult_MPIRowbs(Mat mat,Vec xx,Vec yy)
1164: {
1165:   Mat_MPIRowbs *bsif = (Mat_MPIRowbs*)mat->data;
1166:   BSprocinfo   *bspinfo = bsif->procinfo;
1167:   PetscScalar  *xxa,*xworka,*yya;

1171:   if (!bsif->blocksolveassembly) {
1172:     MatAssemblyEnd_MPIRowbs_ForBlockSolve(mat);
1173:   }

1175:   /* Permute and apply diagonal scaling:  [ xwork = D^{1/2} * x ] */
1176:   if (!bsif->vecs_permscale) {
1177:     VecGetArray(bsif->xwork,&xworka);
1178:     VecGetArray(xx,&xxa);
1179:     BSperm_dvec(xxa,xworka,bsif->pA->perm);CHKERRBS(0);
1180:     VecRestoreArray(bsif->xwork,&xworka);
1181:     VecRestoreArray(xx,&xxa);
1182:     VecPointwiseDivide(xx,bsif->xwork,bsif->diag);
1183:   }

1185:   VecGetArray(xx,&xxa);
1186:   VecGetArray(yy,&yya);
1187:   /* Do lower triangular multiplication:  [ y = L * xwork ] */
1188:   if (bspinfo->single) {
1189:     BSforward1(bsif->pA,xxa,yya,bsif->comm_pA,bspinfo);CHKERRBS(0);
1190:   }  else {
1191:     BSforward(bsif->pA,xxa,yya,bsif->comm_pA,bspinfo);CHKERRBS(0);
1192:   }
1193: 
1194:   /* Do upper triangular multiplication:  [ y = y + L^{T} * xwork ] */
1195:   if (mat->symmetric) {
1196:     if (bspinfo->single){
1197:       BSbackward1(bsif->pA,xxa,yya,bsif->comm_pA,bspinfo);CHKERRBS(0);
1198:     } else {
1199:       BSbackward(bsif->pA,xxa,yya,bsif->comm_pA,bspinfo);CHKERRBS(0);
1200:     }
1201:   }
1202:   /* not needed for ILU version since forward does it all */
1203:   VecRestoreArray(xx,&xxa);
1204:   VecRestoreArray(yy,&yya);

1206:   /* Apply diagonal scaling to vector:  [  y = D^{1/2} * y ] */
1207:   if (!bsif->vecs_permscale) {
1208:     VecGetArray(bsif->xwork,&xworka);
1209:     VecGetArray(xx,&xxa);
1210:     BSiperm_dvec(xworka,xxa,bsif->pA->perm);CHKERRBS(0);
1211:     VecRestoreArray(bsif->xwork,&xworka);
1212:     VecRestoreArray(xx,&xxa);
1213:     VecPointwiseDivide(bsif->xwork,yy,bsif->diag);
1214:     VecGetArray(bsif->xwork,&xworka);
1215:     VecGetArray(yy,&yya);
1216:     BSiperm_dvec(xworka,yya,bsif->pA->perm);CHKERRBS(0);
1217:     VecRestoreArray(bsif->xwork,&xworka);
1218:     VecRestoreArray(yy,&yya);
1219:   }
1220:   PetscLogFlops(2*bsif->nz - mat->cmap.n);

1222:   return(0);
1223: }

1227: PetscErrorCode MatMultAdd_MPIRowbs(Mat mat,Vec xx,Vec yy,Vec zz)
1228: {
1230:   PetscScalar  one = 1.0;

1233:   (*mat->ops->mult)(mat,xx,zz);
1234:   VecAXPY(zz,one,yy);
1235:   return(0);
1236: }

1240: PetscErrorCode MatGetInfo_MPIRowbs(Mat A,MatInfoType flag,MatInfo *info)
1241: {
1242:   Mat_MPIRowbs *mat = (Mat_MPIRowbs*)A->data;
1243:   PetscReal    isend[5],irecv[5];

1247:   info->rows_global    = (double)A->rmap.N;
1248:   info->columns_global = (double)A->cmap.N;
1249:   info->rows_local     = (double)A->cmap.n;
1250:   info->columns_local  = (double)A->rmap.n;
1251:   info->block_size     = 1.0;
1252:   info->mallocs        = (double)mat->reallocs;
1253:   isend[0] = mat->nz; isend[1] = mat->maxnz; isend[2] =  mat->maxnz -  mat->nz;
1254:   isend[3] = A->mem;  isend[4] = info->mallocs;

1256:   if (flag == MAT_LOCAL) {
1257:     info->nz_used      = isend[0];
1258:     info->nz_allocated = isend[1];
1259:     info->nz_unneeded  = isend[2];
1260:     info->memory       = isend[3];
1261:     info->mallocs      = isend[4];
1262:   } else if (flag == MAT_GLOBAL_MAX) {
1263:     MPI_Allreduce(isend,irecv,3,MPIU_REAL,MPI_MAX,A->comm);
1264:     info->nz_used      = irecv[0];
1265:     info->nz_allocated = irecv[1];
1266:     info->nz_unneeded  = irecv[2];
1267:     info->memory       = irecv[3];
1268:     info->mallocs      = irecv[4];
1269:   } else if (flag == MAT_GLOBAL_SUM) {
1270:     MPI_Allreduce(isend,irecv,3,MPIU_REAL,MPI_SUM,A->comm);
1271:     info->nz_used      = irecv[0];
1272:     info->nz_allocated = irecv[1];
1273:     info->nz_unneeded  = irecv[2];
1274:     info->memory       = irecv[3];
1275:     info->mallocs      = irecv[4];
1276:   }
1277:   return(0);
1278: }

1282: PetscErrorCode MatGetDiagonal_MPIRowbs(Mat mat,Vec v)
1283: {
1284:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
1285:   BSsprow      **rs = a->A->rows;
1287:   int          i,n;
1288:   PetscScalar  *x,zero = 0.0;

1291:   if (mat->factor) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
1292:   if (!a->blocksolveassembly) {
1293:     MatAssemblyEnd_MPIRowbs_ForBlockSolve(mat);
1294:   }

1296:   VecSet(v,zero);
1297:   VecGetLocalSize(v,&n);
1298:   if (n != mat->rmap.n) SETERRQ(PETSC_ERR_ARG_SIZ,"Nonconforming mat and vec");
1299:   VecGetArray(v,&x);
1300:   for (i=0; i<mat->rmap.n; i++) {
1301:     x[i] = rs[i]->nz[rs[i]->diag_ind];
1302:   }
1303:   VecRestoreArray(v,&x);
1304:   return(0);
1305: }

1309: PetscErrorCode MatDestroy_MPIRowbs(Mat mat)
1310: {
1311:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
1312:   BSspmat      *A = a->A;
1313:   BSsprow      *vs;
1315:   int          i;

1318: #if defined(PETSC_USE_LOG)
1319:   PetscLogObjectState((PetscObject)mat,"Rows=%d, Cols=%d",mat->rmap.N,mat->cmap.N);
1320: #endif
1321:   MatStashDestroy_Private(&mat->stash);
1322:   if (a->bsmap) {
1323:     PetscFree(a->bsmap->vlocal2global);
1324:     PetscFree(a->bsmap->vglobal2local);
1325:     if (a->bsmap->vglobal2proc)  (*a->bsmap->free_g2p)(a->bsmap->vglobal2proc);
1326:     PetscFree(a->bsmap);
1327:   }

1329:   if (A) {
1330:     for (i=0; i<mat->rmap.n; i++) {
1331:       vs = A->rows[i];
1332:       MatFreeRowbs_Private(mat,vs->length,vs->col,vs->nz);
1333:     }
1334:     /* Note: A->map = a->bsmap is freed above */
1335:     PetscFree(A->rows);
1336:     PetscFree(A);
1337:   }
1338:   if (a->procinfo) {BSfree_ctx(a->procinfo);CHKERRBS(0);}
1339:   if (a->diag)     {VecDestroy(a->diag);}
1340:   if (a->xwork)    {VecDestroy(a->xwork);}
1341:   if (a->pA)       {BSfree_par_mat(a->pA);CHKERRBS(0);}
1342:   if (a->fpA)      {BSfree_copy_par_mat(a->fpA);CHKERRBS(0);}
1343:   if (a->comm_pA)  {BSfree_comm(a->comm_pA);CHKERRBS(0);}
1344:   if (a->comm_fpA) {BSfree_comm(a->comm_fpA);CHKERRBS(0);}
1345:   PetscFree(a->imax);
1346:   MPI_Comm_free(&(a->comm_mpirowbs));
1347:   PetscFree(a);

1349:   PetscObjectChangeTypeName((PetscObject)mat,0);
1350:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIRowbsSetPreallocation_C","",PETSC_NULL);
1351:   return(0);
1352: }

1356: PetscErrorCode MatSetOption_MPIRowbs(Mat A,MatOption op)
1357: {
1358:   Mat_MPIRowbs   *a = (Mat_MPIRowbs*)A->data;

1362:   switch (op) {
1363:   case MAT_ROW_ORIENTED:
1364:     a->roworiented = PETSC_TRUE;
1365:     break;
1366:   case MAT_COLUMN_ORIENTED:
1367:     a->roworiented = PETSC_FALSE;
1368:     break;
1369:   case MAT_COLUMNS_SORTED:
1370:     a->sorted      = 1;
1371:     break;
1372:   case MAT_COLUMNS_UNSORTED:
1373:     a->sorted      = 0;
1374:     break;
1375:   case MAT_NO_NEW_NONZERO_LOCATIONS:
1376:     a->nonew       = 1;
1377:     break;
1378:   case MAT_YES_NEW_NONZERO_LOCATIONS:
1379:     a->nonew       = 0;
1380:     break;
1381:   case MAT_DO_NOT_USE_INODES:
1382:     a->bs_color_single = 1;
1383:     break;
1384:   case MAT_YES_NEW_DIAGONALS:
1385:   case MAT_ROWS_SORTED:
1386:   case MAT_NEW_NONZERO_LOCATION_ERR:
1387:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1388:   case MAT_ROWS_UNSORTED:
1389:   case MAT_USE_HASH_TABLE:
1390:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1391:     break;
1392:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1393:     a->donotstash = PETSC_TRUE;
1394:     break;
1395:   case MAT_NO_NEW_DIAGONALS:
1396:     SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1397:     break;
1398:   case MAT_KEEP_ZEROED_ROWS:
1399:     a->keepzeroedrows    = PETSC_TRUE;
1400:     break;
1401:   case MAT_SYMMETRIC:
1402:     BSset_mat_symmetric(a->A,PETSC_TRUE);CHKERRBS(0);
1403:     break;
1404:   case MAT_STRUCTURALLY_SYMMETRIC:
1405:   case MAT_NOT_SYMMETRIC:
1406:   case MAT_NOT_STRUCTURALLY_SYMMETRIC:
1407:   case MAT_HERMITIAN:
1408:   case MAT_NOT_HERMITIAN:
1409:   case MAT_SYMMETRY_ETERNAL:
1410:   case MAT_NOT_SYMMETRY_ETERNAL:
1411:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1412:     break;
1413:   default:
1414:     SETERRQ1(PETSC_ERR_SUP,"unknown option %d",op);
1415:     break;
1416:   }
1417:   return(0);
1418: }

1422: PetscErrorCode MatGetRow_MPIRowbs(Mat AA,int row,int *nz,int **idx,PetscScalar **v)
1423: {
1424:   Mat_MPIRowbs *mat = (Mat_MPIRowbs*)AA->data;
1425:   BSspmat      *A = mat->A;
1426:   BSsprow      *rs;
1427: 
1429:   if (row < AA->rmap.rstart || row >= AA->rmap.rend) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");

1431:   rs  = A->rows[row - AA->rmap.rstart];
1432:   *nz = rs->length;
1433:   if (v)   *v   = rs->nz;
1434:   if (idx) *idx = rs->col;
1435:   return(0);
1436: }

1440: PetscErrorCode MatRestoreRow_MPIRowbs(Mat A,int row,int *nz,int **idx,PetscScalar **v)
1441: {
1443:   return(0);
1444: }

1446: /* ------------------------------------------------------------------ */

1450: PetscErrorCode MatSetUpPreallocation_MPIRowbs(Mat A)
1451: {

1455:    MatMPIRowbsSetPreallocation(A,PETSC_DEFAULT,0);
1456:   return(0);
1457: }

1459: /* -------------------------------------------------------------------*/
1460: static struct _MatOps MatOps_Values = {MatSetValues_MPIRowbs,
1461:        MatGetRow_MPIRowbs,
1462:        MatRestoreRow_MPIRowbs,
1463:        MatMult_MPIRowbs,
1464: /* 4*/ MatMultAdd_MPIRowbs,
1465:        MatMult_MPIRowbs,
1466:        MatMultAdd_MPIRowbs,
1467:        MatSolve_MPIRowbs,
1468:        0,
1469:        0,
1470: /*10*/ 0,
1471:        0,
1472:        0,
1473:        0,
1474:        0,
1475: /*15*/ MatGetInfo_MPIRowbs,
1476:        0,
1477:        MatGetDiagonal_MPIRowbs,
1478:        0,
1479:        MatNorm_MPIRowbs,
1480: /*20*/ MatAssemblyBegin_MPIRowbs,
1481:        MatAssemblyEnd_MPIRowbs,
1482:        0,
1483:        MatSetOption_MPIRowbs,
1484:        MatZeroEntries_MPIRowbs,
1485: /*25*/ MatZeroRows_MPIRowbs,
1486:        0,
1487:        MatLUFactorNumeric_MPIRowbs,
1488:        0,
1489:        MatCholeskyFactorNumeric_MPIRowbs,
1490: /*30*/ MatSetUpPreallocation_MPIRowbs,
1491:        MatILUFactorSymbolic_MPIRowbs,
1492:        MatIncompleteCholeskyFactorSymbolic_MPIRowbs,
1493:        0,
1494:        0,
1495: /*35*/ 0,
1496:        MatForwardSolve_MPIRowbs,
1497:        MatBackwardSolve_MPIRowbs,
1498:        0,
1499:        0,
1500: /*40*/ 0,
1501:        MatGetSubMatrices_MPIRowbs,
1502:        0,
1503:        0,
1504:        0,
1505: /*45*/ 0,
1506:        MatScale_MPIRowbs,
1507:        0,
1508:        0,
1509:        0,
1510: /*50*/ 0,
1511:        0,
1512:        0,
1513:        0,
1514:        0,
1515: /*55*/ 0,
1516:        0,
1517:        0,
1518:        0,
1519:        0,
1520: /*60*/ MatGetSubMatrix_MPIRowbs,
1521:        MatDestroy_MPIRowbs,
1522:        MatView_MPIRowbs,
1523:        0,
1524:        MatUseScaledForm_MPIRowbs,
1525: /*65*/ MatScaleSystem_MPIRowbs,
1526:        MatUnScaleSystem_MPIRowbs,
1527:        0,
1528:        0,
1529:        0,
1530: /*70*/ 0,
1531:        0,
1532:        0,
1533:        0,
1534:        0,
1535: /*75*/ 0,
1536:        0,
1537:        0,
1538:        0,
1539:        0,
1540: /*80*/ 0,
1541:        0,
1542:        0,
1543:        0,
1544:        MatLoad_MPIRowbs,
1545: /*85*/ 0,
1546:        0,
1547:        0,
1548:        0,
1549:        0,
1550: /*90*/ 0,
1551:        0,
1552:        0,
1553:        0,
1554:        0,
1555: /*95*/ 0,
1556:        0,
1557:        0,
1558:        0};

1560: /* ------------------------------------------------------------------- */

1565: PetscErrorCode  MatMPIRowbsSetPreallocation_MPIRowbs(Mat mat,int nz,const int nnz[])
1566: {

1570:   mat->preallocated = PETSC_TRUE;
1571:   MatCreateMPIRowbs_local(mat,nz,nnz);
1572:   return(0);
1573: }

1576: /*MC
1577:    MATMPIROWBS - MATMPIROWBS = "mpirowbs" - A matrix type providing ILU and ICC for distributed sparse matrices for use
1578:    with the external package BlockSolve95.  If BlockSolve95 is installed (see the manual for instructions
1579:    on how to declare the existence of external packages), a matrix type can be constructed which invokes
1580:    BlockSolve95 preconditioners and solvers. 

1582:    Options Database Keys:
1583: . -mat_type mpirowbs - sets the matrix type to "mpirowbs" during a call to MatSetFromOptions()

1585:   Level: beginner

1587: .seealso: MatCreateMPIRowbs
1588: M*/

1593: PetscErrorCode  MatCreate_MPIRowbs(Mat A)
1594: {
1595:   Mat_MPIRowbs *a;
1596:   BSmapping    *bsmap;
1597:   BSoff_map    *bsoff;
1599:   int          *offset,m,M;
1600:   PetscTruth   flg1,flg3;
1601:   BSprocinfo   *bspinfo;
1602:   MPI_Comm     comm;
1603: 
1605:   comm = A->comm;

1607:   PetscNew(Mat_MPIRowbs,&a);
1608:   A->data               = (void*)a;
1609:   PetscMemcpy(A->ops,&MatOps_Values,sizeof(struct _MatOps));
1610:   A->factor             = 0;
1611:   A->mapping            = 0;
1612:   a->vecs_permscale     = PETSC_FALSE;
1613:   A->insertmode         = NOT_SET_VALUES;
1614:   a->blocksolveassembly = 0;
1615:   a->keepzeroedrows     = PETSC_FALSE;

1617:   MPI_Comm_rank(comm,&a->rank);
1618:   MPI_Comm_size(comm,&a->size);


1621:   PetscMapInitialize(comm,&A->rmap);
1622:   PetscMapInitialize(comm,&A->cmap);
1623:   m    = A->rmap.n;
1624:   M    = A->rmap.N;

1626:   PetscMalloc((A->rmap.n+1)*sizeof(int),&a->imax);
1627:   a->reallocs                      = 0;

1629:   /* build cache for off array entries formed */
1630:   MatStashCreate_Private(A->comm,1,&A->stash);
1631:   a->donotstash = PETSC_FALSE;

1633:   /* Initialize BlockSolve information */
1634:   a->A              = 0;
1635:   a->pA              = 0;
1636:   a->comm_pA  = 0;
1637:   a->fpA      = 0;
1638:   a->comm_fpA = 0;
1639:   a->alpha    = 1.0;
1640:   a->0;
1641:   a->failures = 0;
1642:   MPI_Comm_dup(A->comm,&(a->comm_mpirowbs));
1643:   VecCreateMPI(A->comm,A->rmap.n,A->rmap.N,&(a->diag));
1644:   VecDuplicate(a->diag,&(a->xwork));
1645:   PetscLogObjectParent(A,a->diag);  PetscLogObjectParent(A,a->xwork);
1646:   PetscLogObjectMemory(A,(A->rmap.n+1)*sizeof(PetscScalar));
1647:   bspinfo = BScreate_ctx();CHKERRBS(0);
1648:   a->procinfo = bspinfo;
1649:   BSctx_set_id(bspinfo,a->rank);CHKERRBS(0);
1650:   BSctx_set_np(bspinfo,a->size);CHKERRBS(0);
1651:   BSctx_set_ps(bspinfo,a->comm_mpirowbs);CHKERRBS(0);
1652:   BSctx_set_cs(bspinfo,INT_MAX);CHKERRBS(0);
1653:   BSctx_set_is(bspinfo,INT_MAX);CHKERRBS(0);
1654:   BSctx_set_ct(bspinfo,IDO);CHKERRBS(0);
1655: #if defined(PETSC_USE_DEBUG)
1656:   BSctx_set_err(bspinfo,1);CHKERRBS(0);  /* BS error checking */
1657: #endif
1658:   BSctx_set_rt(bspinfo,1);CHKERRBS(0);
1659: #if defined (PETSC_USE_INFO)
1660:   PetscOptionsHasName(PETSC_NULL,"-info",&flg1);
1661:   if (flg1) {
1662:     BSctx_set_pr(bspinfo,1);CHKERRBS(0);
1663:   }
1664: #endif
1665:   PetscOptionsBegin(A->comm,PETSC_NULL,"Options for MPIROWBS matrix","Mat");
1666:     PetscOptionsTruth("-pc_factor_factorpointwise","Do not optimize for inodes (slow)",PETSC_NULL,PETSC_FALSE,&flg1,PETSC_NULL);
1667:     PetscOptionsTruth("-mat_rowbs_no_inode","Do not optimize for inodes (slow)",PETSC_NULL,PETSC_FALSE,&flg3,PETSC_NULL);
1668:   PetscOptionsEnd();
1669:   if (flg1 || flg3) {
1670:     BSctx_set_si(bspinfo,1);CHKERRBS(0);
1671:   } else {
1672:     BSctx_set_si(bspinfo,0);CHKERRBS(0);
1673:   }
1674: #if defined(PETSC_USE_LOG)
1675:   MLOG_INIT();  /* Initialize logging */
1676: #endif

1678:   /* Compute global offsets */
1679:   offset = &A->rmap.rstart;

1681:   PetscNew(BSmapping,&a->bsmap);
1682:   PetscLogObjectMemory(A,sizeof(BSmapping));
1683:   bsmap = a->bsmap;
1684:   PetscMalloc(sizeof(int),&bsmap->vlocal2global);
1685:   *((int*)bsmap->vlocal2global) = (*offset);
1686:   bsmap->flocal2global                 = BSloc2glob;
1687:   bsmap->free_l2g                = 0;
1688:   PetscMalloc(sizeof(int),&bsmap->vglobal2local);
1689:   *((int*)bsmap->vglobal2local) = (*offset);
1690:   bsmap->fglobal2local                 = BSglob2loc;
1691:   bsmap->free_g2l                 = 0;
1692:   bsoff                          = BSmake_off_map(*offset,bspinfo,A->rmap.N);
1693:   bsmap->vglobal2proc                 = (void*)bsoff;
1694:   bsmap->fglobal2proc                 = BSglob2proc;
1695:   bsmap->free_g2p                = (void(*)(void*)) BSfree_off_map;
1696:   PetscObjectComposeFunctionDynamic((PetscObject)A,"MatMPIRowbsSetPreallocation_C",
1697:                                     "MatMPIRowbsSetPreallocation_MPIRowbs",
1698:                                      MatMPIRowbsSetPreallocation_MPIRowbs);
1699:   PetscObjectChangeTypeName((PetscObject)A,MATMPIROWBS);
1700:   return(0);
1701: }

1706: /* @
1707:   MatMPIRowbsSetPreallocation - Sets the number of expected nonzeros 
1708:   per row in the matrix.

1710:   Input Parameter:
1711: +  mat - matrix
1712: .  nz - maximum expected for any row
1713: -  nzz - number expected in each row

1715:   Note:
1716:   This routine is valid only for matrices stored in the MATMPIROWBS
1717:   format.
1718: @ */
1719: PetscErrorCode  MatMPIRowbsSetPreallocation(Mat mat,int nz,const int nnz[])
1720: {
1721:   PetscErrorCode ierr,(*f)(Mat,int,const int[]);

1724:   PetscObjectQueryFunction((PetscObject)mat,"MatMPIRowbsSetPreallocation_C",(void (**)(void))&f);
1725:   if (f) {
1726:     (*f)(mat,nz,nnz);
1727:   }
1728:   return(0);
1729: }

1731: /* --------------- extra BlockSolve-specific routines -------------- */
1734: /* @
1735:   MatGetBSProcinfo - Gets the BlockSolve BSprocinfo context, which the
1736:   user can then manipulate to alter the default parameters.

1738:   Input Parameter:
1739:   mat - matrix

1741:   Output Parameter:
1742:   procinfo - processor information context

1744:   Note:
1745:   This routine is valid only for matrices stored in the MATMPIROWBS
1746:   format.
1747: @ */
1748: PetscErrorCode  MatGetBSProcinfo(Mat mat,BSprocinfo *procinfo)
1749: {
1750:   Mat_MPIRowbs *a = (Mat_MPIRowbs*)mat->data;
1751:   PetscTruth   ismpirowbs;

1755:   PetscTypeCompare((PetscObject)mat,MATMPIROWBS,&ismpirowbs);
1756:   if (!ismpirowbs) SETERRQ(PETSC_ERR_ARG_WRONG,"For MATMPIROWBS matrix type");
1757:   procinfo = a->procinfo;
1758:   return(0);
1759: }

1763: PetscErrorCode MatLoad_MPIRowbs(PetscViewer viewer,MatType type,Mat *newmat)
1764: {
1765:   Mat_MPIRowbs *a;
1766:   BSspmat      *A;
1767:   BSsprow      **rs;
1768:   Mat          mat;
1770:   int          i,nz,j,rstart,rend,fd,*ourlens,*sndcounts = 0,*procsnz;
1771:   int          header[4],rank,size,*rowlengths = 0,M,m,*rowners,maxnz,*cols;
1772:   PetscScalar  *vals;
1773:   MPI_Comm     comm = ((PetscObject)viewer)->comm;
1774:   MPI_Status   status;

1777:   MPI_Comm_size(comm,&size);
1778:   MPI_Comm_rank(comm,&rank);
1779:   if (!rank) {
1780:     PetscViewerBinaryGetDescriptor(viewer,&fd);
1781:     PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
1782:     if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Not matrix object");
1783:     if (header[3] < 0) {
1784:       SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format,cannot load as MPIRowbs");
1785:     }
1786:   }

1788:   MPI_Bcast(header+1,3,MPI_INT,0,comm);
1789:   M = header[1];

1791:   /* determine ownership of all rows */
1792:   m          = M/size + ((M % size) > rank);
1793:   PetscMalloc((size+2)*sizeof(int),&rowners);
1794:   MPI_Allgather(&m,1,MPI_INT,rowners+1,1,MPI_INT,comm);
1795:   rowners[0] = 0;
1796:   for (i=2; i<=size; i++) {
1797:     rowners[i] += rowners[i-1];
1798:   }
1799:   rstart = rowners[rank];
1800:   rend   = rowners[rank+1];

1802:   /* distribute row lengths to all processors */
1803:   PetscMalloc((rend-rstart)*sizeof(int),&ourlens);
1804:   if (!rank) {
1805:     PetscMalloc(M*sizeof(int),&rowlengths);
1806:     PetscBinaryRead(fd,rowlengths,M,PETSC_INT);
1807:     PetscMalloc(size*sizeof(int),&sndcounts);
1808:     for (i=0; i<size; i++) sndcounts[i] = rowners[i+1] - rowners[i];
1809:     MPI_Scatterv(rowlengths,sndcounts,rowners,MPI_INT,ourlens,rend-rstart,MPI_INT,0,comm);
1810:     PetscFree(sndcounts);
1811:   } else {
1812:     MPI_Scatterv(0,0,0,MPI_INT,ourlens,rend-rstart,MPI_INT,0,comm);
1813:   }

1815:   /* create our matrix */
1816:   MatCreate(comm,newmat);
1817:   MatSetSizes(*newmat,m,m,M,M);
1818:   MatSetType(*newmat,type);
1819:   MatMPIRowbsSetPreallocation(*newmat,0,ourlens);
1820:   mat = *newmat;
1821:   PetscFree(ourlens);

1823:   a = (Mat_MPIRowbs*)mat->data;
1824:   A = a->A;
1825:   rs = A->rows;

1827:   if (!rank) {
1828:     /* calculate the number of nonzeros on each processor */
1829:     PetscMalloc(size*sizeof(int),&procsnz);
1830:     PetscMemzero(procsnz,size*sizeof(int));
1831:     for (i=0; i<size; i++) {
1832:       for (j=rowners[i]; j< rowners[i+1]; j++) {
1833:         procsnz[i] += rowlengths[j];
1834:       }
1835:     }
1836:     PetscFree(rowlengths);

1838:     /* determine max buffer needed and allocate it */
1839:     maxnz = 0;
1840:     for (i=0; i<size; i++) {
1841:       maxnz = PetscMax(maxnz,procsnz[i]);
1842:     }
1843:     PetscMalloc(maxnz*sizeof(int),&cols);

1845:     /* read in my part of the matrix column indices  */
1846:     nz = procsnz[0];
1847:     PetscBinaryRead(fd,cols,nz,PETSC_INT);
1848: 
1849:     /* insert it into my part of matrix */
1850:     nz = 0;
1851:     for (i=0; i<A->num_rows; i++) {
1852:       for (j=0; j<a->imax[i]; j++) {
1853:         rs[i]->col[j] = cols[nz++];
1854:       }
1855:       rs[i]->length = a->imax[i];
1856:     }
1857:     /* read in parts for all other processors */
1858:     for (i=1; i<size; i++) {
1859:       nz   = procsnz[i];
1860:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
1861:       MPI_Send(cols,nz,MPI_INT,i,mat->tag,comm);
1862:     }
1863:     PetscFree(cols);
1864:     PetscMalloc(maxnz*sizeof(PetscScalar),&vals);

1866:     /* read in my part of the matrix numerical values  */
1867:     nz   = procsnz[0];
1868:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1869: 
1870:     /* insert it into my part of matrix */
1871:     nz = 0;
1872:     for (i=0; i<A->num_rows; i++) {
1873:       for (j=0; j<a->imax[i]; j++) {
1874:         rs[i]->nz[j] = vals[nz++];
1875:       }
1876:     }
1877:     /* read in parts for all other processors */
1878:     for (i=1; i<size; i++) {
1879:       nz   = procsnz[i];
1880:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1881:       MPI_Send(vals,nz,MPIU_SCALAR,i,mat->tag,comm);
1882:     }
1883:     PetscFree(vals);
1884:     PetscFree(procsnz);
1885:   } else {
1886:     /* determine buffer space needed for message */
1887:     nz = 0;
1888:     for (i=0; i<A->num_rows; i++) {
1889:       nz += a->imax[i];
1890:     }
1891:     PetscMalloc(nz*sizeof(int),&cols);

1893:     /* receive message of column indices*/
1894:     MPI_Recv(cols,nz,MPI_INT,0,mat->tag,comm,&status);
1895:     MPI_Get_count(&status,MPI_INT,&maxnz);
1896:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong");

1898:     /* insert it into my part of matrix */
1899:     nz = 0;
1900:     for (i=0; i<A->num_rows; i++) {
1901:       for (j=0; j<a->imax[i]; j++) {
1902:         rs[i]->col[j] = cols[nz++];
1903:       }
1904:       rs[i]->length = a->imax[i];
1905:     }
1906:     PetscFree(cols);
1907:     PetscMalloc(nz*sizeof(PetscScalar),&vals);

1909:     /* receive message of values*/
1910:     MPI_Recv(vals,nz,MPIU_SCALAR,0,mat->tag,comm,&status);
1911:     MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
1912:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong");

1914:     /* insert it into my part of matrix */
1915:     nz = 0;
1916:     for (i=0; i<A->num_rows; i++) {
1917:       for (j=0; j<a->imax[i]; j++) {
1918:         rs[i]->nz[j] = vals[nz++];
1919:       }
1920:       rs[i]->length = a->imax[i];
1921:     }
1922:     PetscFree(vals);
1923:   }
1924:   PetscFree(rowners);
1925:   a->nz = a->maxnz;
1926:   MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
1927:   MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
1928:   return(0);
1929: }

1931: /* 
1932:     Special destroy and view routines for factored matrices 
1933: */
1936: static PetscErrorCode MatDestroy_MPIRowbs_Factored(Mat mat)
1937: {
1939: #if defined(PETSC_USE_LOG)
1940:   PetscLogObjectState((PetscObject)mat,"Rows=%d, Cols=%d",mat->rmap.N,mat->cmap.N);
1941: #endif
1942:   return(0);
1943: }

1947: static PetscErrorCode MatView_MPIRowbs_Factored(Mat mat,PetscViewer viewer)
1948: {

1952:   MatView((Mat) mat->data,viewer);
1953:   return(0);
1954: }

1958: PetscErrorCode MatIncompleteCholeskyFactorSymbolic_MPIRowbs(Mat mat,IS isrow,MatFactorInfo *info,Mat *newfact)
1959: {
1960:   /* Note:  f is not currently used in BlockSolve */
1961:   Mat          newmat;
1962:   Mat_MPIRowbs *mbs = (Mat_MPIRowbs*)mat->data;
1964:   PetscTruth   idn;

1967:   if (isrow) {
1968:     ISIdentity(isrow,&idn);
1969:     if (!idn) SETERRQ(PETSC_ERR_SUP,"Only identity row permutation supported");
1970:   }

1972:   if (!mat->symmetric) {
1973:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"To use incomplete Cholesky \n\
1974:         preconditioning with a MATMPIROWBS matrix you must declare it to be \n\
1975:         symmetric using the option MatSetOption(A,MAT_SYMMETRIC)");
1976:   }

1978:   /* If the icc_storage flag wasn't set before the last blocksolveassembly,          */
1979:   /* we must completely redo the assembly as a different storage format is required. */
1980:   if (mbs->blocksolveassembly && !mbs->assembled_icc_storage) {
1981:     mat->same_nonzero       = PETSC_FALSE;
1982:     mbs->blocksolveassembly = 0;
1983:   }

1985:   if (!mbs->blocksolveassembly) {
1986:     BSset_mat_icc_storage(mbs->A,PETSC_TRUE);CHKERRBS(0);
1987:     BSset_mat_symmetric(mbs->A,PETSC_TRUE);CHKERRBS(0);
1988:     MatAssemblyEnd_MPIRowbs_ForBlockSolve(mat);
1989:   }

1991:   /* Copy permuted matrix */
1992:   if (mbs->fpA) {BSfree_copy_par_mat(mbs->fpA);CHKERRBS(0);}
1993:   mbs->fpA = BScopy_par_mat(mbs->pA);CHKERRBS(0);

1995:   /* Set up the communication for factorization */
1996:   if (mbs->comm_fpA) {BSfree_comm(mbs->comm_fpA);CHKERRBS(0);}
1997:   mbs->comm_fpA = BSsetup_factor(mbs->fpA,mbs->procinfo);CHKERRBS(0);

1999:   /* 
2000:       Create a new Mat structure to hold the "factored" matrix, 
2001:     not this merely contains a pointer to the original matrix, since
2002:     the original matrix contains the factor information.
2003:   */
2004:   PetscHeaderCreate(newmat,_p_Mat,struct _MatOps,MAT_COOKIE,-1,"Mat",mat->comm,MatDestroy,MatView);
2005:   PetscLogObjectMemory(newmat,sizeof(struct _p_Mat));

2007:   newmat->data         = (void*)mat;
2008:   PetscMemcpy(newmat->ops,&MatOps_Values,sizeof(struct _MatOps));
2009:   newmat->ops->destroy = MatDestroy_MPIRowbs_Factored;
2010:   newmat->ops->view    = MatView_MPIRowbs_Factored;
2011:   newmat->factor       = 1;
2012:   newmat->preallocated = PETSC_TRUE;
2013:   PetscMapCopy(mat->comm,&mat->rmap,&newmat->rmap);
2014:   PetscMapCopy(mat->comm,&mat->cmap,&newmat->cmap);

2016:   PetscStrallocpy(MATMPIROWBS,&newmat->type_name);

2018:   *newfact = newmat;
2019:   return(0);
2020: }

2024: PetscErrorCode MatILUFactorSymbolic_MPIRowbs(Mat mat,IS isrow,IS iscol,MatFactorInfo* info,Mat *newfact)
2025: {
2026:   Mat          newmat;
2027:   Mat_MPIRowbs *mbs = (Mat_MPIRowbs*)mat->data;
2029:   PetscTruth   idn;

2032:   if (info->levels) SETERRQ(PETSC_ERR_SUP,"Blocksolve ILU only supports 0 fill");
2033:   if (isrow) {
2034:     ISIdentity(isrow,&idn);
2035:     if (!idn) SETERRQ(PETSC_ERR_SUP,"Only identity row permutation supported");
2036:   }
2037:   if (iscol) {
2038:     ISIdentity(iscol,&idn);
2039:     if (!idn) SETERRQ(PETSC_ERR_SUP,"Only identity column permutation supported");
2040:   }

2042:   if (!mbs->blocksolveassembly) {
2043:     MatAssemblyEnd_MPIRowbs_ForBlockSolve(mat);
2044:   }
2045: 
2046: /*   if (mat->symmetric) { */
2047: /*     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"To use ILU preconditioner with \n\ */
2048: /*         MatCreateMPIRowbs() matrix you CANNOT declare it to be a symmetric matrix\n\ */
2049: /*         using the option MatSetOption(A,MAT_SYMMETRIC)"); */
2050: /*   } */

2052:   /* Copy permuted matrix */
2053:   if (mbs->fpA) {BSfree_copy_par_mat(mbs->fpA);CHKERRBS(0);}
2054:   mbs->fpA = BScopy_par_mat(mbs->pA);CHKERRBS(0);

2056:   /* Set up the communication for factorization */
2057:   if (mbs->comm_fpA) {BSfree_comm(mbs->comm_fpA);CHKERRBS(0);}
2058:   mbs->comm_fpA = BSsetup_factor(mbs->fpA,mbs->procinfo);CHKERRBS(0);

2060:   /* 
2061:       Create a new Mat structure to hold the "factored" matrix,
2062:     not this merely contains a pointer to the original matrix, since
2063:     the original matrix contains the factor information.
2064:   */
2065:   PetscHeaderCreate(newmat,_p_Mat,struct _MatOps,MAT_COOKIE,-1,"Mat",mat->comm,MatDestroy,MatView);
2066:   PetscLogObjectMemory(newmat,sizeof(struct _p_Mat));

2068:   newmat->data         = (void*)mat;
2069:   PetscMemcpy(newmat->ops,&MatOps_Values,sizeof(struct _MatOps));
2070:   newmat->ops->destroy = MatDestroy_MPIRowbs_Factored;
2071:   newmat->ops->view    = MatView_MPIRowbs_Factored;
2072:   newmat->factor       = 1;
2073:   newmat->preallocated = PETSC_TRUE;

2075:   PetscMapCopy(mat->comm,&mat->rmap,&newmat->rmap);
2076:   PetscMapCopy(mat->comm,&mat->cmap,&newmat->cmap);

2078:   PetscStrallocpy(MATMPIROWBS,&newmat->type_name);

2080:   *newfact = newmat;
2081:   return(0);
2082: }

2086: /*@C
2087:    MatCreateMPIRowbs - Creates a sparse parallel matrix in the MATMPIROWBS
2088:    format.  This format is intended primarily as an interface for BlockSolve95.

2090:    Collective on MPI_Comm

2092:    Input Parameters:
2093: +  comm - MPI communicator
2094: .  m - number of local rows (or PETSC_DECIDE to have calculated)
2095: .  M - number of global rows (or PETSC_DECIDE to have calculated)
2096: .  nz - number of nonzeros per row (same for all local rows)
2097: -  nnz - number of nonzeros per row (possibly different for each row).

2099:    Output Parameter:
2100: .  newA - the matrix 

2102:    Notes:
2103:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one processor
2104:    than it must be used on all processors that share the object for that argument.

2106:    The user MUST specify either the local or global matrix dimensions
2107:    (possibly both).

2109:    Specify the preallocated storage with either nz or nnz (not both).  Set 
2110:    nz=PETSC_DEFAULT and nnz=PETSC_NULL for PETSc to control dynamic memory 
2111:    allocation.

2113:    Notes:
2114:    By default, the matrix is assumed to be nonsymmetric; the user can
2115:    take advantage of special optimizations for symmetric matrices by calling
2116: $     MatSetOption(mat,MAT_SYMMETRIC)
2117: $     MatSetOption(mat,MAT_SYMMETRY_ETERNAL)
2118:    BEFORE calling the routine MatAssemblyBegin().

2120:    Internally, the MATMPIROWBS format inserts zero elements to the
2121:    matrix if necessary, so that nonsymmetric matrices are considered
2122:    to be symmetric in terms of their sparsity structure; this format
2123:    is required for use of the parallel communication routines within
2124:    BlockSolve95. In particular, if the matrix element A[i,j] exists,
2125:    then PETSc will internally allocate a 0 value for the element
2126:    A[j,i] during MatAssemblyEnd() if the user has not already set
2127:    a value for the matrix element A[j,i].

2129:    Options Database Keys:
2130: .  -mat_rowbs_no_inode - Do not use inodes.

2132:    Level: intermediate
2133:   
2134: .keywords: matrix, row, symmetric, sparse, parallel, BlockSolve

2136: .seealso: MatCreate(), MatSetValues()
2137: @*/
2138: PetscErrorCode  MatCreateMPIRowbs(MPI_Comm comm,int m,int M,int nz,const int nnz[],Mat *newA)
2139: {
2141: 
2143:   MatCreate(comm,newA);
2144:   MatSetSizes(*newA,m,m,M,M);
2145:   MatSetType(*newA,MATMPIROWBS);
2146:   MatMPIRowbsSetPreallocation(*newA,nz,nnz);
2147:   return(0);
2148: }


2151: /* -------------------------------------------------------------------------*/

2153:  #include src/mat/impls/aij/seq/aij.h
2154:  #include src/mat/impls/aij/mpi/mpiaij.h

2158: PetscErrorCode MatGetSubMatrices_MPIRowbs(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[])
2159: {
2161:   int         nmax,nstages_local,nstages,i,pos,max_no;


2165:   /* Allocate memory to hold all the submatrices */
2166:   if (scall != MAT_REUSE_MATRIX) {
2167:     PetscMalloc((ismax+1)*sizeof(Mat),submat);
2168:   }
2169: 
2170:   /* Determine the number of stages through which submatrices are done */
2171:   nmax          = 20*1000000 / (C->cmap.N * sizeof(int));
2172:   if (!nmax) nmax = 1;
2173:   nstages_local = ismax/nmax + ((ismax % nmax)?1:0);

2175:   /* Make sure every processor loops through the nstages */
2176:   MPI_Allreduce(&nstages_local,&nstages,1,MPI_INT,MPI_MAX,C->comm);

2178:   for (i=0,pos=0; i<nstages; i++) {
2179:     if (pos+nmax <= ismax) max_no = nmax;
2180:     else if (pos == ismax) max_no = 0;
2181:     else                   max_no = ismax-pos;
2182:     MatGetSubMatrices_MPIRowbs_Local(C,max_no,isrow+pos,iscol+pos,scall,*submat+pos);
2183:     pos += max_no;
2184:   }
2185:   return(0);
2186: }
2187: /* -------------------------------------------------------------------------*/
2188: /* for now MatGetSubMatrices_MPIRowbs_Local get MPIAij submatrices of input
2189:    matrix and preservs zeroes from structural symetry
2190:  */
2193: PetscErrorCode MatGetSubMatrices_MPIRowbs_Local(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submats)
2194: {
2195:   Mat_MPIRowbs  *c = (Mat_MPIRowbs *)(C->data);
2196:   BSspmat       *A = c->A;
2197:   Mat_SeqAIJ    *mat;
2199:   int         **irow,**icol,*nrow,*ncol,*w1,*w2,*w3,*w4,*rtable,start,end,size;
2200:   int         **sbuf1,**sbuf2,rank,m,i,j,k,l,ct1,ct2,**rbuf1,row,proc;
2201:   int         nrqs,msz,**ptr,idx,*req_size,*ctr,*pa,*tmp,tcol,nrqr;
2202:   int         **rbuf3,*req_source,**sbuf_aj,**rbuf2,max1,max2,**rmap;
2203:   int         **cmap,**lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax,*irow_i;
2204:   int         len,ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*cmap_i,*lens_i;
2205:   int         *rmap_i,tag0,tag1,tag2,tag3;
2206:   MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3;
2207:   MPI_Request *r_waits4,*s_waits3,*s_waits4;
2208:   MPI_Status  *r_status1,*r_status2,*s_status1,*s_status3,*s_status2;
2209:   MPI_Status  *r_status3,*r_status4,*s_status4;
2210:   MPI_Comm    comm;
2211:   FLOAT       **rbuf4,**sbuf_aa,*vals,*sbuf_aa_i;
2212:   PetscScalar *mat_a;
2213:   PetscTruth  sorted;
2214:   int         *onodes1,*olengths1;

2217:   comm   = C->comm;
2218:   tag0   = C->tag;
2219:   size   = c->size;
2220:   rank   = c->rank;
2221:   m      = C->rmap.N;
2222: 
2223:   /* Get some new tags to keep the communication clean */
2224:   PetscObjectGetNewTag((PetscObject)C,&tag1);
2225:   PetscObjectGetNewTag((PetscObject)C,&tag2);
2226:   PetscObjectGetNewTag((PetscObject)C,&tag3);

2228:     /* Check if the col indices are sorted */
2229:   for (i=0; i<ismax; i++) {
2230:     ISSorted(isrow[i],&sorted);
2231:     if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted");
2232:     ISSorted(iscol[i],&sorted);
2233:     /*    if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted"); */
2234:   }

2236:   len    = (2*ismax+1)*(sizeof(int*)+ sizeof(int)) + (m+1)*sizeof(int);
2237:   PetscMalloc(len,&irow);
2238:   icol   = irow + ismax;
2239:   nrow   = (int*)(icol + ismax);
2240:   ncol   = nrow + ismax;
2241:   rtable = ncol + ismax;

2243:   for (i=0; i<ismax; i++) {
2244:     ISGetIndices(isrow[i],&irow[i]);
2245:     ISGetIndices(iscol[i],&icol[i]);
2246:     ISGetLocalSize(isrow[i],&nrow[i]);
2247:     ISGetLocalSize(iscol[i],&ncol[i]);
2248:   }

2250:   /* Create hash table for the mapping :row -> proc*/
2251:   for (i=0,j=0; i<size; i++) {
2252:     jmax = C->rmap.range[i+1];
2253:     for (; j<jmax; j++) {
2254:       rtable[j] = i;
2255:     }
2256:   }

2258:   /* evaluate communication - mesg to who, length of mesg, and buffer space
2259:      required. Based on this, buffers are allocated, and data copied into them*/
2260:   PetscMalloc(size*4*sizeof(int),&w1); /* mesg size */
2261:   w2     = w1 + size;      /* if w2[i] marked, then a message to proc i*/
2262:   w3     = w2 + size;      /* no of IS that needs to be sent to proc i */
2263:   w4     = w3 + size;      /* temp work space used in determining w1, w2, w3 */
2264:   PetscMemzero(w1,size*3*sizeof(int)); /* initialize work vector*/
2265:   for (i=0; i<ismax; i++) {
2266:     PetscMemzero(w4,size*sizeof(int)); /* initialize work vector*/
2267:     jmax   = nrow[i];
2268:     irow_i = irow[i];
2269:     for (j=0; j<jmax; j++) {
2270:       row  = irow_i[j];
2271:       proc = rtable[row];
2272:       w4[proc]++;
2273:     }
2274:     for (j=0; j<size; j++) {
2275:       if (w4[j]) { w1[j] += w4[j];  w3[j]++;}
2276:     }
2277:   }
2278: 
2279:   nrqs     = 0;              /* no of outgoing messages */
2280:   msz      = 0;              /* total mesg length (for all procs) */
2281:   w1[rank] = 0;              /* no mesg sent to self */
2282:   w3[rank] = 0;
2283:   for (i=0; i<size; i++) {
2284:     if (w1[i])  { w2[i] = 1; nrqs++;} /* there exists a message to proc i */
2285:   }
2286:   PetscMalloc((nrqs+1)*sizeof(int),&pa); /*(proc -array)*/
2287:   for (i=0,j=0; i<size; i++) {
2288:     if (w1[i]) { pa[j] = i; j++; }
2289:   }

2291:   /* Each message would have a header = 1 + 2*(no of IS) + data */
2292:   for (i=0; i<nrqs; i++) {
2293:     j     = pa[i];
2294:     w1[j] += w2[j] + 2* w3[j];
2295:     msz   += w1[j];
2296:   }

2298:   /* Determine the number of messages to expect, their lengths, from from-ids */
2299:   PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);
2300:   PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);

2302:   /* Now post the Irecvs corresponding to these messages */
2303:   PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);
2304: 
2305:   PetscFree(onodes1);
2306:   PetscFree(olengths1);
2307: 
2308:   /* Allocate Memory for outgoing messages */
2309:   len      = 2*size*sizeof(int*) + 2*msz*sizeof(int) + size*sizeof(int);
2310:   PetscMalloc(len,&sbuf1);
2311:   ptr      = sbuf1 + size;   /* Pointers to the data in outgoing buffers */
2312:   PetscMemzero(sbuf1,2*size*sizeof(int*));
2313:   /* allocate memory for outgoing data + buf to receive the first reply */
2314:   tmp      = (int*)(ptr + size);
2315:   ctr      = tmp + 2*msz;

2317:   {
2318:     int *iptr = tmp,ict = 0;
2319:     for (i=0; i<nrqs; i++) {
2320:       j         = pa[i];
2321:       iptr     += ict;
2322:       sbuf1[j]  = iptr;
2323:       ict       = w1[j];
2324:     }
2325:   }

2327:   /* Form the outgoing messages */
2328:   /* Initialize the header space */
2329:   for (i=0; i<nrqs; i++) {
2330:     j           = pa[i];
2331:     sbuf1[j][0] = 0;
2332:     PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));
2333:     ptr[j]      = sbuf1[j] + 2*w3[j] + 1;
2334:   }
2335: 
2336:   /* Parse the isrow and copy data into outbuf */
2337:   for (i=0; i<ismax; i++) {
2338:     PetscMemzero(ctr,size*sizeof(int));
2339:     irow_i = irow[i];
2340:     jmax   = nrow[i];
2341:     for (j=0; j<jmax; j++) {  /* parse the indices of each IS */
2342:       row  = irow_i[j];
2343:       proc = rtable[row];
2344:       if (proc != rank) { /* copy to the outgoing buf*/
2345:         ctr[proc]++;
2346:         *ptr[proc] = row;
2347:         ptr[proc]++;
2348:       }
2349:     }
2350:     /* Update the headers for the current IS */
2351:     for (j=0; j<size; j++) { /* Can Optimise this loop too */
2352:       if ((ctr_j = ctr[j])) {
2353:         sbuf1_j        = sbuf1[j];
2354:         k              = ++sbuf1_j[0];
2355:         sbuf1_j[2*k]   = ctr_j;
2356:         sbuf1_j[2*k-1] = i;
2357:       }
2358:     }
2359:   }

2361:   /*  Now  post the sends */
2362:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);
2363:   for (i=0; i<nrqs; ++i) {
2364:     j    = pa[i];
2365:     MPI_Isend(sbuf1[j],w1[j],MPI_INT,j,tag0,comm,s_waits1+i);
2366:   }

2368:   /* Post Receives to capture the buffer size */
2369:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);
2370:   PetscMalloc((nrqs+1)*sizeof(int*),&rbuf2);
2371:   rbuf2[0] = tmp + msz;
2372:   for (i=1; i<nrqs; ++i) {
2373:     rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]];
2374:   }
2375:   for (i=0; i<nrqs; ++i) {
2376:     j    = pa[i];
2377:     MPI_Irecv(rbuf2[i],w1[j],MPI_INT,j,tag1,comm,r_waits2+i);
2378:   }

2380:   /* Send to other procs the buf size they should allocate */
2381: 

2383:   /* Receive messages*/
2384:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);
2385:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);
2386:   len         = 2*nrqr*sizeof(int) + (nrqr+1)*sizeof(int*);
2387:   PetscMalloc(len,&sbuf2);
2388:   req_size    = (int*)(sbuf2 + nrqr);
2389:   req_source  = req_size + nrqr;
2390: 
2391:   {
2392:     BSsprow    **sAi = A->rows;
2393:     int        id,rstart = C->rmap.rstart;
2394:     int        *sbuf2_i;

2396:     for (i=0; i<nrqr; ++i) {
2397:       MPI_Waitany(nrqr,r_waits1,&idx,r_status1+i);
2398:       req_size[idx]   = 0;
2399:       rbuf1_i         = rbuf1[idx];
2400:       start           = 2*rbuf1_i[0] + 1;
2401:       MPI_Get_count(r_status1+i,MPI_INT,&end);
2402:       PetscMalloc((end+1)*sizeof(int),&sbuf2[idx]);
2403:       sbuf2_i         = sbuf2[idx];
2404:       for (j=start; j<end; j++) {
2405:         id               = rbuf1_i[j] - rstart;
2406:         ncols            = (sAi[id])->length;
2407:         sbuf2_i[j]       = ncols;
2408:         req_size[idx]   += ncols;
2409:       }
2410:       req_source[idx] = r_status1[i].MPI_SOURCE;
2411:       /* form the header */
2412:       sbuf2_i[0]   = req_size[idx];
2413:       for (j=1; j<start; j++) { sbuf2_i[j] = rbuf1_i[j]; }
2414:       MPI_Isend(sbuf2_i,end,MPI_INT,req_source[idx],tag1,comm,s_waits2+i);
2415:     }
2416:   }
2417:   PetscFree(r_status1);
2418:   PetscFree(r_waits1);

2420:   /*  recv buffer sizes */
2421:   /* Receive messages*/
2422: 
2423:   PetscMalloc((nrqs+1)*sizeof(int*),&rbuf3);
2424:   PetscMalloc((nrqs+1)*sizeof(FLOAT *),&rbuf4);
2425:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits3);
2426:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits4);
2427:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);

2429:   for (i=0; i<nrqs; ++i) {
2430:     MPI_Waitany(nrqs,r_waits2,&idx,r_status2+i);
2431:     PetscMalloc((rbuf2[idx][0]+1)*sizeof(int),&rbuf3[idx]);
2432:     PetscMalloc((rbuf2[idx][0]+1)*sizeof(FLOAT),&rbuf4[idx]);
2433:     MPI_Irecv(rbuf3[idx],rbuf2[idx][0],MPI_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idx);
2434:     MPI_Irecv(rbuf4[idx],rbuf2[idx][0],MPIU_SCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idx);
2435:   }
2436:   PetscFree(r_status2);
2437:   PetscFree(r_waits2);
2438: 
2439:   /* Wait on sends1 and sends2 */
2440:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);
2441:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);

2443:   if (nrqs) {MPI_Waitall(nrqs,s_waits1,s_status1);}
2444:   if (nrqr) {MPI_Waitall(nrqr,s_waits2,s_status2);}
2445:   PetscFree(s_status1);
2446:   PetscFree(s_status2);
2447:   PetscFree(s_waits1);
2448:   PetscFree(s_waits2);

2450:   /* Now allocate buffers for a->j, and send them off */
2451:   PetscMalloc((nrqr+1)*sizeof(int*),&sbuf_aj);
2452:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
2453:   PetscMalloc((j+1)*sizeof(int),&sbuf_aj[0]);
2454:   for (i=1; i<nrqr; i++)  sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1];
2455: 
2456:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits3);
2457:   {
2458:     BSsprow *brow;
2459:     int *Acol;
2460:     int rstart = C->rmap.rstart;

2462:     for (i=0; i<nrqr; i++) {
2463:       rbuf1_i   = rbuf1[i];
2464:       sbuf_aj_i = sbuf_aj[i];
2465:       ct1       = 2*rbuf1_i[0] + 1;
2466:       ct2       = 0;
2467:       for (j=1,max1=rbuf1_i[0]; j<=max1; j++) {
2468:         kmax = rbuf1[i][2*j];
2469:         for (k=0; k<kmax; k++,ct1++) {
2470:           brow   = A->rows[rbuf1_i[ct1] - rstart];
2471:           ncols  = brow->length;
2472:           Acol   = brow->col;
2473:           /* load the column indices for this row into cols*/
2474:           cols  = sbuf_aj_i + ct2;
2475:           PetscMemcpy(cols,Acol,ncols*sizeof(int));
2476:           /*for (l=0; l<ncols;l++) cols[l]=Acol[l]; */ /* How is it with
2477:                                                           mappings?? */
2478:           ct2 += ncols;
2479:         }
2480:       }
2481:       MPI_Isend(sbuf_aj_i,req_size[i],MPI_INT,req_source[i],tag2,comm,s_waits3+i);
2482:     }
2483:   }
2484:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status3);
2485:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status3);

2487:   /* Allocate buffers for a->a, and send them off */
2488:   PetscMalloc((nrqr+1)*sizeof(FLOAT*),&sbuf_aa);
2489:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
2490:   PetscMalloc((j+1)*sizeof(FLOAT),&sbuf_aa[0]);
2491:   for (i=1; i<nrqr; i++)  sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1];
2492: 
2493:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits4);
2494:   {
2495:     BSsprow *brow;
2496:     FLOAT *Aval;
2497:     int rstart = C->rmap.rstart;
2498: 
2499:     for (i=0; i<nrqr; i++) {
2500:       rbuf1_i   = rbuf1[i];
2501:       sbuf_aa_i = sbuf_aa[i];
2502:       ct1       = 2*rbuf1_i[0]+1;
2503:       ct2       = 0;
2504:       for (j=1,max1=rbuf1_i[0]; j<=max1; j++) {
2505:         kmax = rbuf1_i[2*j];
2506:         for (k=0; k<kmax; k++,ct1++) {
2507:           brow  = A->rows[rbuf1_i[ct1] - rstart];
2508:           ncols = brow->length;
2509:           Aval  = brow->nz;
2510:           /* load the column values for this row into vals*/
2511:           vals  = sbuf_aa_i+ct2;
2512:           PetscMemcpy(vals,Aval,ncols*sizeof(FLOAT));
2513:           ct2 += ncols;
2514:         }
2515:       }
2516:       MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source[i],tag3,comm,s_waits4+i);
2517:     }
2518:   }
2519:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status4);
2520:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status4);
2521:   PetscFree(rbuf1);

2523:   /* Form the matrix */
2524:   /* create col map */
2525:   {
2526:     int *icol_i;
2527: 
2528:     len     = (1+ismax)*sizeof(int*)+ ismax*C->cmap.N*sizeof(int);
2529:     PetscMalloc(len,&cmap);
2530:     cmap[0] = (int*)(cmap + ismax);
2531:     PetscMemzero(cmap[0],(1+ismax*C->cmap.N)*sizeof(int));
2532:     for (i=1; i<ismax; i++) { cmap[i] = cmap[i-1] + C->cmap.N; }
2533:     for (i=0; i<ismax; i++) {
2534:       jmax   = ncol[i];
2535:       icol_i = icol[i];
2536:       cmap_i = cmap[i];
2537:       for (j=0; j<jmax; j++) {
2538:         cmap_i[icol_i[j]] = j+1;
2539:       }
2540:     }
2541:   }

2543:   /* Create lens which is required for MatCreate... */
2544:   for (i=0,j=0; i<ismax; i++) { j += nrow[i]; }
2545:   len     = (1+ismax)*sizeof(int*)+ j*sizeof(int);
2546:   PetscMalloc(len,&lens);
2547:   lens[0] = (int*)(lens + ismax);
2548:   PetscMemzero(lens[0],j*sizeof(int));
2549:   for (i=1; i<ismax; i++) { lens[i] = lens[i-1] + nrow[i-1]; }
2550: 
2551:   /* Update lens from local data */
2552:   { BSsprow *Arow;
2553:     for (i=0; i<ismax; i++) {
2554:       jmax   = nrow[i];
2555:       cmap_i = cmap[i];
2556:       irow_i = irow[i];
2557:       lens_i = lens[i];
2558:       for (j=0; j<jmax; j++) {
2559:         row  = irow_i[j];
2560:         proc = rtable[row];
2561:         if (proc == rank) {
2562:           Arow=A->rows[row-C->rmap.rstart];
2563:           ncols=Arow->length;
2564:           cols=Arow->col;
2565:           for (k=0; k<ncols; k++) {
2566:             if (cmap_i[cols[k]]) { lens_i[j]++;}
2567:           }
2568:         }
2569:       }
2570:     }
2571:   }
2572: 
2573:   /* Create row map*/
2574:   len     = (1+ismax)*sizeof(int*)+ ismax*C->rmap.N*sizeof(int);
2575:   PetscMalloc(len,&rmap);
2576:   rmap[0] = (int*)(rmap + ismax);
2577:   PetscMemzero(rmap[0],ismax*C->rmap.N*sizeof(int));
2578:   for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->rmap.N;}
2579:   for (i=0; i<ismax; i++) {
2580:     rmap_i = rmap[i];
2581:     irow_i = irow[i];
2582:     jmax   = nrow[i];
2583:     for (j=0; j<jmax; j++) {
2584:       rmap_i[irow_i[j]] = j;
2585:     }
2586:   }
2587: 
2588:   /* Update lens from offproc data */
2589:   {
2590:     int *rbuf2_i,*rbuf3_i,*sbuf1_i;

2592:     for (tmp2=0; tmp2<nrqs; tmp2++) {
2593:       MPI_Waitany(nrqs,r_waits3,&i,r_status3+tmp2);
2594:       idx     = pa[i];
2595:       sbuf1_i = sbuf1[idx];
2596:       jmax    = sbuf1_i[0];
2597:       ct1     = 2*jmax+1;
2598:       ct2     = 0;
2599:       rbuf2_i = rbuf2[i];
2600:       rbuf3_i = rbuf3[i];
2601:       for (j=1; j<=jmax; j++) {
2602:         is_no   = sbuf1_i[2*j-1];
2603:         max1    = sbuf1_i[2*j];
2604:         lens_i  = lens[is_no];
2605:         cmap_i  = cmap[is_no];
2606:         rmap_i  = rmap[is_no];
2607:         for (k=0; k<max1; k++,ct1++) {
2608:           row  = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */
2609:           max2 = rbuf2_i[ct1];
2610:           for (l=0; l<max2; l++,ct2++) {
2611:             if (cmap_i[rbuf3_i[ct2]]) {
2612:               lens_i[row]++;
2613:             }
2614:           }
2615:         }
2616:       }
2617:     }
2618:   }
2619:   PetscFree(r_status3);
2620:   PetscFree(r_waits3);
2621:   if (nrqr) {MPI_Waitall(nrqr,s_waits3,s_status3);}
2622:   PetscFree(s_status3);
2623:   PetscFree(s_waits3);

2625:   /* Create the submatrices */
2626:   if (scall == MAT_REUSE_MATRIX) {
2627:     PetscTruth same;
2628: 
2629:     /*
2630:         Assumes new rows are same length as the old rows,hence bug!
2631:     */
2632:     for (i=0; i<ismax; i++) {
2633:       PetscTypeCompare((PetscObject)(submats[i]),MATSEQAIJ,&same);
2634:       if (!same) {
2635:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong type");
2636:       }
2637:       mat = (Mat_SeqAIJ*)(submats[i]->data);
2638:       if ((submats[i]->rmap.n != nrow[i]) || (submats[i]->cmap.n != ncol[i])) {
2639:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
2640:       }
2641:       PetscMemcmp(mat->ilen,lens[i],submats[i]->rmap.n*sizeof(int),&same);
2642:       if (!same) {
2643:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
2644:       }
2645:       /* Initial matrix as if empty */
2646:       PetscMemzero(mat->ilen,submats[i]->rmap.n*sizeof(int));
2647:       submats[i]->factor = C->factor;
2648:     }
2649:   } else {
2650:     for (i=0; i<ismax; i++) {
2651:       /* Here we want to explicitly generate SeqAIJ matrices */
2652:       MatCreate(PETSC_COMM_SELF,submats+i);
2653:       MatSetSizes(submats[i],nrow[i],ncol[i],nrow[i],ncol[i]);
2654:       MatSetType(submats[i],MATSEQAIJ);
2655:       MatSeqAIJSetPreallocation(submats[i],0,lens[i]);
2656:     }
2657:   }

2659:   /* Assemble the matrices */
2660:   /* First assemble the local rows */
2661:   {
2662:     int    ilen_row,*imat_ilen,*imat_j,*imat_i,old_row;
2663:     PetscScalar *imat_a;
2664:     BSsprow *Arow;
2665: 
2666:     for (i=0; i<ismax; i++) {
2667:       mat       = (Mat_SeqAIJ*)submats[i]->data;
2668:       imat_ilen = mat->ilen;
2669:       imat_j    = mat->j;
2670:       imat_i    = mat->i;
2671:       imat_a    = mat->a;
2672:       cmap_i    = cmap[i];
2673:       rmap_i    = rmap[i];
2674:       irow_i    = irow[i];
2675:       jmax      = nrow[i];
2676:       for (j=0; j<jmax; j++) {
2677:         row      = irow_i[j];
2678:         proc     = rtable[row];
2679:         if (proc == rank) {
2680:           old_row  = row;
2681:           row      = rmap_i[row];
2682:           ilen_row = imat_ilen[row];
2683: 
2684:           Arow=A->rows[old_row-C->rmap.rstart];
2685:           ncols=Arow->length;
2686:           cols=Arow->col;
2687:           vals=Arow->nz;
2688: 
2689:           mat_i    = imat_i[row];
2690:           mat_a    = imat_a + mat_i;
2691:           mat_j    = imat_j + mat_i;
2692:           for (k=0; k<ncols; k++) {
2693:             if ((tcol = cmap_i[cols[k]])) {
2694:               *mat_j++ = tcol - 1;
2695:               *mat_a++ = (PetscScalar)vals[k];
2696:               ilen_row++;
2697:             }
2698:           }
2699:           imat_ilen[row] = ilen_row;
2700:         }
2701:       }
2702:     }
2703:   }

2705:   /*   Now assemble the off proc rows*/
2706:   {
2707:     int    *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen;
2708:     int    *imat_j,*imat_i;
2709:     PetscScalar *imat_a;
2710:     FLOAT *rbuf4_i;
2711: 
2712:     for (tmp2=0; tmp2<nrqs; tmp2++) {
2713:       MPI_Waitany(nrqs,r_waits4,&i,r_status4+tmp2);
2714:       idx     = pa[i];
2715:       sbuf1_i = sbuf1[idx];
2716:       jmax    = sbuf1_i[0];
2717:       ct1     = 2*jmax + 1;
2718:       ct2     = 0;
2719:       rbuf2_i = rbuf2[i];
2720:       rbuf3_i = rbuf3[i];
2721:       rbuf4_i = rbuf4[i];
2722:       for (j=1; j<=jmax; j++) {
2723:         is_no     = sbuf1_i[2*j-1];
2724:         rmap_i    = rmap[is_no];
2725:         cmap_i    = cmap[is_no];
2726:         mat       = (Mat_SeqAIJ*)submats[is_no]->data;
2727:         imat_ilen = mat->ilen;
2728:         imat_j    = mat->j;
2729:         imat_i    = mat->i;
2730:         imat_a    = mat->a;
2731:         max1      = sbuf1_i[2*j];
2732:         for (k=0; k<max1; k++,ct1++) {
2733:           row   = sbuf1_i[ct1];
2734:           row   = rmap_i[row];
2735:           ilen  = imat_ilen[row];
2736:           mat_i = imat_i[row];
2737:           mat_a = imat_a + mat_i;
2738:           mat_j = imat_j + mat_i;
2739:           max2 = rbuf2_i[ct1];
2740:           for (l=0; l<max2; l++,ct2++) {
2741:             if ((tcol = cmap_i[rbuf3_i[ct2]])) {
2742:               *mat_j++ = tcol - 1;
2743:               *mat_a++ = (PetscScalar)rbuf4_i[ct2];
2744:               ilen++;
2745:             }
2746:           }
2747:           imat_ilen[row] = ilen;
2748:         }
2749:       }
2750:     }
2751:   }
2752:   PetscFree(r_status4);
2753:   PetscFree(r_waits4);
2754:   if (nrqr) {MPI_Waitall(nrqr,s_waits4,s_status4);}
2755:   PetscFree(s_waits4);
2756:   PetscFree(s_status4);

2758:   /* Restore the indices */
2759:   for (i=0; i<ismax; i++) {
2760:     ISRestoreIndices(isrow[i],irow+i);
2761:     ISRestoreIndices(iscol[i],icol+i);
2762:   }

2764:   /* Destroy allocated memory */
2765:   PetscFree(irow);
2766:   PetscFree(w1);
2767:   PetscFree(pa);

2769:   PetscFree(sbuf1);
2770:   PetscFree(rbuf2);
2771:   for (i=0; i<nrqr; ++i) {
2772:     PetscFree(sbuf2[i]);
2773:   }
2774:   for (i=0; i<nrqs; ++i) {
2775:     PetscFree(rbuf3[i]);
2776:     PetscFree(rbuf4[i]);
2777:   }

2779:   PetscFree(sbuf2);
2780:   PetscFree(rbuf3);
2781:   PetscFree(rbuf4);
2782:   PetscFree(sbuf_aj[0]);
2783:   PetscFree(sbuf_aj);
2784:   PetscFree(sbuf_aa[0]);
2785:   PetscFree(sbuf_aa);
2786: 
2787:   PetscFree(cmap);
2788:   PetscFree(rmap);
2789:   PetscFree(lens);

2791:   for (i=0; i<ismax; i++) {
2792:     MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);
2793:     MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);
2794:   }
2795:   return(0);
2796: }

2798: /*
2799:   can be optimized by send only non-zeroes in iscol IS  -
2800:   so prebuild submatrix on sending side including A,B partitioning
2801:   */
2804:  #include src/vec/is/impls/general/general.h
2805: PetscErrorCode MatGetSubMatrix_MPIRowbs(Mat C,IS isrow,IS iscol,int csize,MatReuse scall,Mat *submat)
2806: {
2807:   Mat_MPIRowbs  *c = (Mat_MPIRowbs*)C->data;
2808:   BSspmat       *A = c->A;
2809:   BSsprow *Arow;
2810:   Mat_SeqAIJ    *matA,*matB; /* on prac , off proc part of submat */
2811:   Mat_MPIAIJ    *mat;  /* submat->data */
2813:   int    *irow,*icol,nrow,ncol,*rtable,size,rank,tag0,tag1,tag2,tag3;
2814:   int    *w1,*w2,*pa,nrqs,nrqr,msz,row_t;
2815:   int    i,j,k,l,len,jmax,proc,idx;
2816:   int    **sbuf1,**sbuf2,**rbuf1,**rbuf2,*req_size,**sbuf3,**rbuf3;
2817:   FLOAT  **rbuf4,**sbuf4; /* FLOAT is from Block Solve 95 library */

2819:   int    *cmap,*rmap,nlocal,*o_nz,*d_nz,cstart,cend;
2820:   int    *req_source;
2821:   int    ncols_t;
2822: 
2823: 
2824:   MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3;
2825:   MPI_Request *r_waits4,*s_waits3,*s_waits4;
2826: 
2827:   MPI_Status  *r_status1,*r_status2,*s_status1,*s_status3,*s_status2;
2828:   MPI_Status  *r_status3,*r_status4,*s_status4;
2829:   MPI_Comm    comm;


2833:   comm   = C->comm;
2834:   tag0   = C->tag;
2835:   size   = c->size;
2836:   rank   = c->rank;

2838:   if (size==1) {
2839:     if (scall == MAT_REUSE_MATRIX) {
2840:       ierr=MatGetSubMatrices(C,1,&isrow,&iscol,MAT_REUSE_MATRIX,&submat);
2841:       return(0);
2842:     } else {
2843:       Mat *newsubmat;
2844: 
2845:       ierr=MatGetSubMatrices(C,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&newsubmat);
2846:       *submat=*newsubmat;
2847:       ierr=PetscFree(newsubmat);
2848:       return(0);
2849:     }
2850:   }
2851: 
2852:   /* Get some new tags to keep the communication clean */
2853:   PetscObjectGetNewTag((PetscObject)C,&tag1);
2854:   PetscObjectGetNewTag((PetscObject)C,&tag2);
2855:   PetscObjectGetNewTag((PetscObject)C,&tag3);

2857:   /* Check if the col indices are sorted */
2858:   {PetscTruth sorted;
2859:   ISSorted(isrow,&sorted);
2860:   if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted");
2861:   ISSorted(iscol,&sorted);
2862:   if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted");
2863:   }
2864: 
2865:   ISGetIndices(isrow,&irow);
2866:   ISGetIndices(iscol,&icol);
2867:   ISGetLocalSize(isrow,&nrow);
2868:   ISGetLocalSize(iscol,&ncol);
2869: 
2870:   if (!isrow) SETERRQ(PETSC_ERR_ARG_SIZ,"Empty ISrow");
2871:   if (!iscol) SETERRQ(PETSC_ERR_ARG_SIZ,"Empty IScol");
2872: 
2873: 
2874:   len    = (C->rmap.N+1)*sizeof(int);
2875:   PetscMalloc(len,&rtable);
2876:   /* Create hash table for the mapping :row -> proc*/
2877:   for (i=0,j=0; i<size; i++) {
2878:     jmax = C->rmap.range[i+1];
2879:     for (; j<jmax; j++) {
2880:       rtable[j] = i;
2881:     }
2882:   }

2884:   /* evaluate communication - mesg to who, length of mesg, and buffer space
2885:      required. Based on this, buffers are allocated, and data copied into them*/
2886:   PetscMalloc(size*2*sizeof(int),&w1); /* mesg size */
2887:   w2     = w1 + size;      /* if w2[i] marked, then a message to proc i*/
2888:   PetscMemzero(w1,size*2*sizeof(int)); /* initialize work vector*/
2889:   for (j=0; j<nrow; j++) {
2890:     row_t  = irow[j];
2891:     proc   = rtable[row_t];
2892:     w1[proc]++;
2893:   }
2894:   nrqs     = 0;              /* no of outgoing messages */
2895:   msz      = 0;              /* total mesg length (for all procs) */
2896:   w1[rank] = 0;              /* no mesg sent to self */
2897:   for (i=0; i<size; i++) {
2898:     if (w1[i])  { w2[i] = 1; nrqs++;} /* there exists a message to proc i */
2899:   }
2900: 
2901:   PetscMalloc((nrqs+1)*sizeof(int),&pa); /*(proc -array)*/
2902:   for (i=0,j=0; i<size; i++) {
2903:     if (w1[i]) {
2904:       pa[j++] = i;
2905:       w1[i]++;  /* header for return data */
2906:       msz+=w1[i];
2907:     }
2908:   }
2909: 
2910:   {int  *onodes1,*olengths1;
2911:   /* Determine the number of messages to expect, their lengths, from from-ids */
2912:   PetscGatherNumberOfMessages(comm,w2,w1,&nrqr);
2913:   PetscGatherMessageLengths(comm,nrqs,nrqr,w1,&onodes1,&olengths1);
2914:   /* Now post the Irecvs corresponding to these messages */
2915:   PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);
2916:   PetscFree(onodes1);
2917:   PetscFree(olengths1);
2918:   }
2919: 
2920: { int **ptr,*iptr,*tmp;
2921:   /* Allocate Memory for outgoing messages */
2922:   len      = 2*size*sizeof(int*) + msz*sizeof(int);
2923:   PetscMalloc(len,&sbuf1);
2924:   ptr      = sbuf1 + size;   /* Pointers to the data in outgoing buffers */
2925:   PetscMemzero(sbuf1,2*size*sizeof(int*));
2926:   /* allocate memory for outgoing data + buf to receive the first reply */
2927:   tmp      = (int*)(ptr + size);

2929:   for (i=0,iptr=tmp; i<nrqs; i++) {
2930:     j         = pa[i];
2931:     sbuf1[j]  = iptr;
2932:     iptr     += w1[j];
2933:   }

2935:   /* Form the outgoing messages */
2936:   for (i=0; i<nrqs; i++) {
2937:     j           = pa[i];
2938:     sbuf1[j][0] = 0;   /*header */
2939:     ptr[j]      = sbuf1[j] + 1;
2940:   }
2941: 
2942:   /* Parse the isrow and copy data into outbuf */
2943:   for (j=0; j<nrow; j++) {
2944:     row_t  = irow[j];
2945:     proc = rtable[row_t];
2946:     if (proc != rank) { /* copy to the outgoing buf*/
2947:       sbuf1[proc][0]++;
2948:       *ptr[proc] = row_t;
2949:       ptr[proc]++;
2950:     }
2951:   }
2952: } /* block */

2954:   /*  Now  post the sends */
2955: 
2956:   /* structure of sbuf1[i]/rbuf1[i] : 1 (num of rows) + nrow-local rows (nuberes
2957:    * of requested rows)*/

2959:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);
2960:   for (i=0; i<nrqs; ++i) {
2961:     j    = pa[i];
2962:     MPI_Isend(sbuf1[j],w1[j],MPI_INT,j,tag0,comm,s_waits1+i);
2963:   }

2965:   /* Post Receives to capture the buffer size */
2966:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);
2967:   PetscMalloc((nrqs+1)*sizeof(int*),&rbuf2);
2968:   PetscMalloc(msz*sizeof(int)+1,&(rbuf2[0]));
2969:   for (i=1; i<nrqs; ++i) {
2970:     rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]];
2971:   }
2972:   for (i=0; i<nrqs; ++i) {
2973:     j    = pa[i];
2974:     MPI_Irecv(rbuf2[i],w1[j],MPI_INT,j,tag1,comm,r_waits2+i);
2975:   }

2977:   /* Send to other procs the buf size they should allocate */
2978:   /* structure of sbuf2[i]/rbuf2[i]: 1 (total size to allocate) + nrow-locrow
2979:    * (row sizes) */

2981:   /* Receive messages*/
2982:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);
2983:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);
2984:   len         = 2*nrqr*sizeof(int) + (nrqr+1)*sizeof(int*);
2985:   PetscMalloc(len,&sbuf2);
2986:   req_size    = (int*)(sbuf2 + nrqr);
2987:   req_source  = req_size + nrqr;
2988: 
2989:   {
2990:     BSsprow    **sAi = A->rows;
2991:     int        id,rstart = C->rmap.rstart;
2992:     int        *sbuf2_i,*rbuf1_i,end;

2994:     for (i=0; i<nrqr; ++i) {
2995:       MPI_Waitany(nrqr,r_waits1,&idx,r_status1+i);
2996:       req_size[idx]   = 0;
2997:       rbuf1_i         = rbuf1[idx];
2998:       MPI_Get_count(r_status1+i,MPI_INT,&end);
2999:       PetscMalloc((end+1)*sizeof(int),&sbuf2[idx]);
3000:       sbuf2_i         = sbuf2[idx];
3001:       for (j=1; j<end; j++) {
3002:         id               = rbuf1_i[j] - rstart;
3003:         ncols_t          = (sAi[id])->length;
3004:         sbuf2_i[j]       = ncols_t;
3005:         req_size[idx]   += ncols_t;
3006:       }
3007:       req_source[idx] = r_status1[i].MPI_SOURCE;
3008:       /* form the header */
3009:       sbuf2_i[0]   = req_size[idx];
3010:       MPI_Isend(sbuf2_i,end,MPI_INT,req_source[idx],tag1,comm,s_waits2+i);
3011:     }
3012:   }
3013:   PetscFree(r_status1);
3014:   PetscFree(r_waits1);

3016:   /*  recv buffer sizes */
3017:   /* Receive messages*/
3018: 
3019:   PetscMalloc((nrqs+1)*sizeof(int*),&rbuf3);
3020:   PetscMalloc((nrqs+1)*sizeof(FLOAT*),&rbuf4);
3021:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits3);
3022:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits4);
3023:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);

3025:   for (i=0; i<nrqs; ++i) {
3026:     MPI_Waitany(nrqs,r_waits2,&idx,r_status2+i);
3027:     PetscMalloc((rbuf2[idx][0]+1)*sizeof(int),&rbuf3[idx]);
3028:     PetscMalloc((rbuf2[idx][0]+1)*sizeof(FLOAT),&rbuf4[idx]);
3029:     MPI_Irecv(rbuf3[idx],rbuf2[idx][0],MPI_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idx);
3030:     MPI_Irecv(rbuf4[idx],rbuf2[idx][0],MPIU_SCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idx);
3031:   }
3032:   PetscFree(r_status2);
3033:   PetscFree(r_waits2);
3034: 
3035:   /* Wait on sends1 and sends2 */
3036:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);
3037:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);

3039:   if (nrqs) {MPI_Waitall(nrqs,s_waits1,s_status1);}
3040:   if (nrqr) {MPI_Waitall(nrqr,s_waits2,s_status2);}
3041:   PetscFree(s_status1);
3042:   PetscFree(s_status2);
3043:   PetscFree(s_waits1);
3044:   PetscFree(s_waits2);

3046:   /* Now allocate buffers for a->j, and send them off */
3047:   /* structure of sbuf3[i]/rbuf3[i],sbuf4[i]/rbuf4[i]: reqsize[i] (cols resp.
3048:    * vals of all req. rows; row sizes was in rbuf2; vals are of FLOAT type */
3049: 
3050:   PetscMalloc((nrqr+1)*sizeof(int*),&sbuf3);
3051:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
3052:   PetscMalloc((j+1)*sizeof(int),&sbuf3[0]);
3053:   for (i=1; i<nrqr; i++)  sbuf3[i] = sbuf3[i-1] + req_size[i-1];
3054: 
3055:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits3);
3056:   {
3057:     int *Acol,*rbuf1_i,*sbuf3_i,rqrow,noutcols,kmax,*cols,ncols;
3058:     int rstart = C->rmap.rstart;

3060:     for (i=0; i<nrqr; i++) {
3061:       rbuf1_i   = rbuf1[i];
3062:       sbuf3_i   = sbuf3[i];
3063:       noutcols  = 0;
3064:       kmax = rbuf1_i[0];  /* num. of req. rows */
3065:       for (k=0,rqrow=1; k<kmax; k++,rqrow++) {
3066:         Arow    = A->rows[rbuf1_i[rqrow] - rstart];
3067:         ncols  = Arow->length;
3068:         Acol   = Arow->col;
3069:         /* load the column indices for this row into cols*/
3070:         cols  = sbuf3_i + noutcols;
3071:         PetscMemcpy(cols,Acol,ncols*sizeof(int));
3072:         /*for (l=0; l<ncols;l++) cols[l]=Acol[l]; */ /* How is it with mappings?? */
3073:         noutcols += ncols;
3074:       }
3075:       MPI_Isend(sbuf3_i,req_size[i],MPI_INT,req_source[i],tag2,comm,s_waits3+i);
3076:     }
3077:   }
3078:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status3);
3079:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status3);

3081:   /* Allocate buffers for a->a, and send them off */
3082:   /* can be optimized by conect with previous block */
3083:   PetscMalloc((nrqr+1)*sizeof(FLOAT*),&sbuf4);
3084:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
3085:   PetscMalloc((j+1)*sizeof(FLOAT),&sbuf4[0]);
3086:   for (i=1; i<nrqr; i++)  sbuf4[i] = sbuf4[i-1] + req_size[i-1];
3087: 
3088:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits4);
3089:   {
3090:     FLOAT *Aval,*vals,*sbuf4_i;
3091:     int rstart = C->rmap.rstart,*rbuf1_i,rqrow,noutvals,kmax,ncols;
3092: 
3093: 
3094:     for (i=0; i<nrqr; i++) {
3095:       rbuf1_i   = rbuf1[i];
3096:       sbuf4_i   = sbuf4[i];
3097:       rqrow     = 1;
3098:       noutvals  = 0;
3099:       kmax      = rbuf1_i[0];  /* num of req. rows */
3100:       for (k=0; k<kmax; k++,rqrow++) {
3101:         Arow    = A->rows[rbuf1_i[rqrow] - rstart];
3102:         ncols  = Arow->length;
3103:         Aval = Arow->nz;
3104:         /* load the column values for this row into vals*/
3105:         vals  = sbuf4_i+noutvals;
3106:         PetscMemcpy(vals,Aval,ncols*sizeof(FLOAT));
3107:         noutvals += ncols;
3108:       }
3109:       MPI_Isend(sbuf4_i,req_size[i],MPIU_SCALAR,req_source[i],tag3,comm,s_waits4+i);
3110:     }
3111:   }
3112:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status4);
3113:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status4);
3114:   PetscFree(rbuf1);

3116:   /* Form the matrix */

3118:   /* create col map */
3119:   len     = C->cmap.N*sizeof(int)+1;
3120:   PetscMalloc(len,&cmap);
3121:   PetscMemzero(cmap,C->cmap.N*sizeof(int));
3122:   for (j=0; j<ncol; j++) {
3123:       cmap[icol[j]] = j+1;
3124:   }
3125: 
3126:   /* Create row map / maybe I will need global rowmap but here is local rowmap*/
3127:   len     = C->rmap.N*sizeof(int)+1;
3128:   PetscMalloc(len,&rmap);
3129:   PetscMemzero(rmap,C->rmap.N*sizeof(int));
3130:   for (j=0; j<nrow; j++) {
3131:     rmap[irow[j]] = j;
3132:   }

3134:   /*
3135:      Determine the number of non-zeros in the diagonal and off-diagonal 
3136:      portions of the matrix in order to do correct preallocation
3137:    */

3139:   /* first get start and end of "diagonal" columns */
3140:   if (csize == PETSC_DECIDE) {
3141:     nlocal = ncol/size + ((ncol % size) > rank);
3142:   } else {
3143:     nlocal = csize;
3144:   }
3145:   {
3146:     int ncols,*cols,olen,dlen,thecol;
3147:     int *rbuf2_i,*rbuf3_i,*sbuf1_i,row,kmax,cidx;
3148: 
3149:     MPI_Scan(&nlocal,&cend,1,MPI_INT,MPI_SUM,comm);
3150:     cstart = cend - nlocal;
3151:     if (rank == size - 1 && cend != ncol) {
3152:       SETERRQ(PETSC_ERR_ARG_SIZ,"Local column sizes do not add up to total number of columns");
3153:     }

3155:     PetscMalloc((2*nrow+1)*sizeof(int),&d_nz);
3156:     o_nz = d_nz + nrow;
3157: 
3158:     /* Update lens from local data */
3159:     for (j=0; j<nrow; j++) {
3160:       row  = irow[j];
3161:       proc = rtable[row];
3162:       if (proc == rank) {
3163:         Arow=A->rows[row-C->rmap.rstart];
3164:         ncols=Arow->length;
3165:         cols=Arow->col;
3166:         olen=dlen=0;
3167:         for (k=0; k<ncols; k++) {
3168:           if ((thecol=cmap[cols[k]])) {
3169:             if (cstart<thecol && thecol<=cend) dlen++; /* thecol is from 1 */
3170:             else olen++;
3171:           }
3172:         }
3173:         o_nz[j]=olen;
3174:         d_nz[j]=dlen;
3175:       } else d_nz[j]=o_nz[j]=0;
3176:     }
3177:     /* Update lens from offproc data and done waits */
3178:     /* this will be much simplier after sending only appropriate columns */
3179:     for (j=0; j<nrqs;j++) {
3180:       MPI_Waitany(nrqs,r_waits3,&i,r_status3+j);
3181:       proc   = pa[i];
3182:       sbuf1_i = sbuf1[proc];
3183:       cidx    = 0;
3184:       rbuf2_i = rbuf2[i];
3185:       rbuf3_i = rbuf3[i];
3186:       kmax    = sbuf1_i[0]; /*num of rq. rows*/
3187:       for (k=1; k<=kmax; k++) {
3188:         row  = rmap[sbuf1_i[k]]; /* the val in the new matrix to be */
3189:         for (l=0; l<rbuf2_i[k]; l++,cidx++) {
3190:           if ((thecol=cmap[rbuf3_i[cidx]])) {
3191: 
3192:             if (cstart<thecol && thecol<=cend) d_nz[row]++; /* thecol is from 1 */
3193:             else o_nz[row]++;
3194:           }
3195:         }
3196:       }
3197:     }
3198:   }
3199:   PetscFree(r_status3);
3200:   PetscFree(r_waits3);
3201:   if (nrqr) {MPI_Waitall(nrqr,s_waits3,s_status3);}
3202:   PetscFree(s_status3);
3203:   PetscFree(s_waits3);

3205:   if (scall ==  MAT_INITIAL_MATRIX) {
3206:     MatCreate(comm,submat);
3207:     MatSetSizes(*submat,nrow,nlocal,PETSC_DECIDE,ncol);
3208:     MatSetType(*submat,C->type_name);
3209:     MatMPIAIJSetPreallocation(*submat,0,d_nz,0,o_nz);
3210:     mat=(Mat_MPIAIJ *)((*submat)->data);
3211:     matA=(Mat_SeqAIJ *)(mat->A->data);
3212:     matB=(Mat_SeqAIJ *)(mat->B->data);
3213: 
3214:   } else {
3215:     PetscTruth same;
3216:     /* folowing code can be optionaly dropped for debuged versions of users
3217:      * program, but I don't know PETSc option which can switch off such safety
3218:      * tests - in a same way counting of o_nz,d_nz can be droped for  REUSE
3219:      * matrix */
3220: 
3221:     PetscTypeCompare((PetscObject)(*submat),MATMPIAIJ,&same);
3222:     if (!same) {
3223:       SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong type");
3224:     }
3225:     if (((*submat)->rmap.n != nrow) || ((*submat)->cmap.N != ncol)) {
3226:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
3227:     }
3228:     mat=(Mat_MPIAIJ *)((*submat)->data);
3229:     matA=(Mat_SeqAIJ *)(mat->A->data);
3230:     matB=(Mat_SeqAIJ *)(mat->B->data);
3231:     PetscMemcmp(matA->ilen,d_nz,nrow*sizeof(int),&same);
3232:     if (!same) {
3233:       SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
3234:     }
3235:     PetscMemcmp(matB->ilen,o_nz,nrow*sizeof(int),&same);
3236:     if (!same) {
3237:       SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
3238:     }
3239:   /* Initial matrix as if empty */
3240:     PetscMemzero(matA->ilen,nrow*sizeof(int));
3241:     PetscMemzero(matB->ilen,nrow*sizeof(int));
3242:     /* Perhaps MatZeroEnteries may be better - look what it is exactly doing - I must
3243:      * delete all possibly nonactual inforamtion */
3244:     /*submats[i]->factor = C->factor; !!! ??? if factor will be same then I must
3245:      * copy some factor information - where are thay */
3246:     (*submat)->was_assembled=PETSC_FALSE;
3247:     (*submat)->assembled=PETSC_FALSE;
3248: 
3249:   }
3250:   PetscFree(d_nz);

3252:   /* Assemble the matrix */
3253:   /* First assemble from local rows */
3254:   {
3255:     int    i_row,oldrow,row,ncols,*cols,*matA_j,*matB_j,ilenA,ilenB,tcol;
3256:     FLOAT  *vals;
3257:     PetscScalar *matA_a,*matB_a;
3258: 
3259:     for (j=0; j<nrow; j++) {
3260:       oldrow = irow[j];
3261:       proc   = rtable[oldrow];
3262:       if (proc == rank) {
3263:         row  = rmap[oldrow];
3264: 
3265:         Arow  = A->rows[oldrow-C->rmap.rstart];
3266:         ncols = Arow->length;
3267:         cols  = Arow->col;
3268:         vals  = Arow->nz;
3269: 
3270:         i_row   = matA->i[row];
3271:         matA_a = matA->a + i_row;
3272:         matA_j = matA->j + i_row;
3273:         i_row   = matB->i[row];
3274:         matB_a = matB->a + i_row;
3275:         matB_j = matB->j + i_row;
3276:         for (k=0,ilenA=0,ilenB=0; k<ncols; k++) {
3277:           if ((tcol = cmap[cols[k]])) {
3278:             if (tcol<=cstart) {
3279:               *matB_j++ = tcol-1;
3280:               *matB_a++ = vals[k];
3281:               ilenB++;
3282:             } else if (tcol<=cend) {
3283:               *matA_j++ = (tcol-1)-cstart;
3284:               *matA_a++ = (PetscScalar)(vals[k]);
3285:               ilenA++;
3286:             } else {
3287:               *matB_j++ = tcol-1;
3288:               *matB_a++ = vals[k];
3289:               ilenB++;
3290:             }
3291:           }
3292:         }
3293:         matA->ilen[row]=ilenA;
3294:         matB->ilen[row]=ilenB;
3295: 
3296:       }
3297:     }
3298:   }

3300:   /*   Now assemble the off proc rows*/
3301:   {
3302:     int  *sbuf1_i,*rbuf2_i,*rbuf3_i,cidx,kmax,row,i_row;
3303:     int  *matA_j,*matB_j,lmax,tcol,ilenA,ilenB;
3304:     PetscScalar *matA_a,*matB_a;
3305:     FLOAT *rbuf4_i;

3307:     for (j=0; j<nrqs; j++) {
3308:       MPI_Waitany(nrqs,r_waits4,&i,r_status4+j);
3309:       proc   = pa[i];
3310:       sbuf1_i = sbuf1[proc];
3311: 
3312:       cidx    = 0;
3313:       rbuf2_i = rbuf2[i];
3314:       rbuf3_i = rbuf3[i];
3315:       rbuf4_i = rbuf4[i];
3316:       kmax    = sbuf1_i[0];
3317:       for (k=1; k<=kmax; k++) {
3318:         row = rmap[sbuf1_i[k]];
3319: 
3320:         i_row  = matA->i[row];
3321:         matA_a = matA->a + i_row;
3322:         matA_j = matA->j + i_row;
3323:         i_row  = matB->i[row];
3324:         matB_a = matB->a + i_row;
3325:         matB_j = matB->j + i_row;
3326: 
3327:         lmax = rbuf2_i[k];
3328:         for (l=0,ilenA=0,ilenB=0; l<lmax; l++,cidx++) {
3329:           if ((tcol = cmap[rbuf3_i[cidx]])) {
3330:             if (tcol<=cstart) {
3331:               *matB_j++ = tcol-1;
3332:               *matB_a++ = (PetscScalar)(rbuf4_i[cidx]);;
3333:               ilenB++;
3334:             } else if (tcol<=cend) {
3335:               *matA_j++ = (tcol-1)-cstart;
3336:               *matA_a++ = (PetscScalar)(rbuf4_i[cidx]);
3337:               ilenA++;
3338:             } else {
3339:               *matB_j++ = tcol-1;
3340:               *matB_a++ = (PetscScalar)(rbuf4_i[cidx]);
3341:               ilenB++;
3342:             }
3343:           }
3344:         }
3345:         matA->ilen[row]=ilenA;
3346:         matB->ilen[row]=ilenB;
3347:       }
3348:     }
3349:   }

3351:   PetscFree(r_status4);
3352:   PetscFree(r_waits4);
3353:   if (nrqr) {MPI_Waitall(nrqr,s_waits4,s_status4);}
3354:   PetscFree(s_waits4);
3355:   PetscFree(s_status4);

3357:   /* Restore the indices */
3358:   ISRestoreIndices(isrow,&irow);
3359:   ISRestoreIndices(iscol,&icol);

3361:   /* Destroy allocated memory */
3362:   PetscFree(rtable);
3363:   PetscFree(w1);
3364:   PetscFree(pa);

3366:   PetscFree(sbuf1);
3367:   PetscFree(rbuf2[0]);
3368:   PetscFree(rbuf2);
3369:   for (i=0; i<nrqr; ++i) {
3370:     PetscFree(sbuf2[i]);
3371:   }
3372:   for (i=0; i<nrqs; ++i) {
3373:     PetscFree(rbuf3[i]);
3374:     PetscFree(rbuf4[i]);
3375:   }

3377:   PetscFree(sbuf2);
3378:   PetscFree(rbuf3);
3379:   PetscFree(rbuf4);
3380:   PetscFree(sbuf3[0]);
3381:   PetscFree(sbuf3);
3382:   PetscFree(sbuf4[0]);
3383:   PetscFree(sbuf4);
3384: 
3385:   PetscFree(cmap);
3386:   PetscFree(rmap);


3389:   MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3390:   MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);


3393:   return(0);
3394: }