Actual source code: mpiov.c

  1: /*$Id: mpiov.c,v 1.26.1.76.2.22 2001/09/07 20:09:38 bsmith Exp $*/
  2: /*
  3:    Routines to compute overlapping regions of a parallel MPI matrix
  4:   and to find submatrices that were shared across processors.
  5: */
 6:  #include src/mat/impls/aij/mpi/mpiaij.h
 7:  #include petscbt.h

  9: static int MatIncreaseOverlap_MPIAIJ_Once(Mat,int,IS *);
 10: static int MatIncreaseOverlap_MPIAIJ_Local(Mat,int,char **,int*,int**);
 11: static int MatIncreaseOverlap_MPIAIJ_Receive(Mat,int,int **,int**,int*);
 12: EXTERN int MatGetRow_MPIAIJ(Mat,int,int*,int**,PetscScalar**);
 13: EXTERN int MatRestoreRow_MPIAIJ(Mat,int,int*,int**,PetscScalar**);

 17: int MatIncreaseOverlap_MPIAIJ(Mat C,int imax,IS is[],int ov)
 18: {
 19:   int i,ierr;

 22:   if (ov < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified");
 23:   for (i=0; i<ov; ++i) {
 24:     MatIncreaseOverlap_MPIAIJ_Once(C,imax,is);
 25:   }
 26:   return(0);
 27: }

 29: /*
 30:   Sample message format:
 31:   If a processor A wants processor B to process some elements corresponding
 32:   to index sets is[1],is[5]
 33:   mesg [0] = 2   (no of index sets in the mesg)
 34:   -----------  
 35:   mesg [1] = 1 => is[1]
 36:   mesg [2] = sizeof(is[1]);
 37:   -----------  
 38:   mesg [3] = 5  => is[5]
 39:   mesg [4] = sizeof(is[5]);
 40:   -----------
 41:   mesg [5] 
 42:   mesg [n]  datas[1]
 43:   -----------  
 44:   mesg[n+1]
 45:   mesg[m]  data(is[5])
 46:   -----------  
 47:   
 48:   Notes:
 49:   nrqs - no of requests sent (or to be sent out)
 50:   nrqr - no of requests recieved (which have to be or which have been processed
 51: */
 54: static int MatIncreaseOverlap_MPIAIJ_Once(Mat C,int imax,IS is[])
 55: {
 56:   Mat_MPIAIJ  *c = (Mat_MPIAIJ*)C->data;
 57:   int         **idx,*n,*w1,*w2,*w3,*w4,*rtable,**data,len,*idx_i;
 58:   int         size,rank,m,i,j,k,ierr,**rbuf,row,proc,nrqs,msz,**outdat,**ptr;
 59:   int         *ctr,*pa,*tmp,nrqr,*isz,*isz1,**xdata,**rbuf2;
 60:   int         *onodes1,*olengths1,tag1,tag2,*onodes2,*olengths2;
 61:   PetscBT     *table;
 62:   MPI_Comm    comm;
 63:   MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2;
 64:   MPI_Status  *s_status,*recv_status;

 67:   comm   = C->comm;
 68:   size   = c->size;
 69:   rank   = c->rank;
 70:   m      = C->M;

 72:   PetscObjectGetNewTag((PetscObject)C,&tag1);
 73:   PetscObjectGetNewTag((PetscObject)C,&tag2);
 74: 
 75:   len    = (imax+1)*sizeof(int*)+ (imax + m)*sizeof(int);
 76:   PetscMalloc(len,&idx);
 77:   n      = (int*)(idx + imax);
 78:   rtable = n + imax;
 79: 
 80:   for (i=0; i<imax; i++) {
 81:     ISGetIndices(is[i],&idx[i]);
 82:     ISGetLocalSize(is[i],&n[i]);
 83:   }
 84: 
 85:   /* Create hash table for the mapping :row -> proc*/
 86:   for (i=0,j=0; i<size; i++) {
 87:     len = c->rowners[i+1];
 88:     for (; j<len; j++) {
 89:       rtable[j] = i;
 90:     }
 91:   }

 93:   /* evaluate communication - mesg to who,length of mesg, and buffer space
 94:      required. Based on this, buffers are allocated, and data copied into them*/
 95:   PetscMalloc(size*4*sizeof(int),&w1);/*  mesg size */
 96:   w2   = w1 + size;       /* if w2[i] marked, then a message to proc i*/
 97:   w3   = w2 + size;       /* no of IS that needs to be sent to proc i */
 98:   w4   = w3 + size;       /* temp work space used in determining w1, w2, w3 */
 99:   PetscMemzero(w1,size*3*sizeof(int)); /* initialise work vector*/
100:   for (i=0; i<imax; i++) {
101:     PetscMemzero(w4,size*sizeof(int)); /* initialise work vector*/
102:     idx_i = idx[i];
103:     len   = n[i];
104:     for (j=0; j<len; j++) {
105:       row  = idx_i[j];
106:       if (row < 0) {
107:         SETERRQ(1,"Index set cannot have negative entries");
108:       }
109:       proc = rtable[row];
110:       w4[proc]++;
111:     }
112:     for (j=0; j<size; j++){
113:       if (w4[j]) { w1[j] += w4[j]; w3[j]++;}
114:     }
115:   }

117:   nrqs     = 0;              /* no of outgoing messages */
118:   msz      = 0;              /* total mesg length (for all proc */
119:   w1[rank] = 0;              /* no mesg sent to intself */
120:   w3[rank] = 0;
121:   for (i=0; i<size; i++) {
122:     if (w1[i])  {w2[i] = 1; nrqs++;} /* there exists a message to proc i */
123:   }
124:   /* pa - is list of processors to communicate with */
125:   PetscMalloc((nrqs+1)*sizeof(int),&pa);
126:   for (i=0,j=0; i<size; i++) {
127:     if (w1[i]) {pa[j] = i; j++;}
128:   }

130:   /* Each message would have a header = 1 + 2*(no of IS) + data */
131:   for (i=0; i<nrqs; i++) {
132:     j      = pa[i];
133:     w1[j] += w2[j] + 2*w3[j];
134:     msz   += w1[j];
135:   }

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

141:   /* Now post the Irecvs corresponding to these messages */
142:   PetscPostIrecvInt(comm,tag1,nrqr,onodes1,olengths1,&rbuf,&r_waits1);

144:   /* Allocate Memory for outgoing messages */
145:   len  = 2*size*sizeof(int*) + (size+msz)*sizeof(int);
146:   PetscMalloc(len,&outdat);
147:   ptr  = outdat + size;     /* Pointers to the data in outgoing buffers */
148:   PetscMemzero(outdat,2*size*sizeof(int*));
149:   tmp  = (int*)(outdat + 2*size);
150:   ctr  = tmp + msz;

152:   {
153:     int *iptr = tmp,ict  = 0;
154:     for (i=0; i<nrqs; i++) {
155:       j         = pa[i];
156:       iptr     +=  ict;
157:       outdat[j] = iptr;
158:       ict       = w1[j];
159:     }
160:   }

162:   /* Form the outgoing messages */
163:   /*plug in the headers*/
164:   for (i=0; i<nrqs; i++) {
165:     j            = pa[i];
166:     outdat[j][0] = 0;
167:     PetscMemzero(outdat[j]+1,2*w3[j]*sizeof(int));
168:     ptr[j]       = outdat[j] + 2*w3[j] + 1;
169:   }
170: 
171:   /* Memory for doing local proc's work*/
172:   {
173:     int  *d_p;
174:     char *t_p;

176:     len   = (imax)*(sizeof(PetscBT) + sizeof(int*)+ sizeof(int)) +
177:       (m)*imax*sizeof(int)  + (m/PETSC_BITS_PER_BYTE+1)*imax*sizeof(char) + 1;
178:     PetscMalloc(len,&table);
179:     PetscMemzero(table,len);
180:     data  = (int **)(table + imax);
181:     isz   = (int  *)(data  + imax);
182:     d_p   = (int  *)(isz   + imax);
183:     t_p   = (char *)(d_p   + m*imax);
184:     for (i=0; i<imax; i++) {
185:       table[i] = t_p + (m/PETSC_BITS_PER_BYTE+1)*i;
186:       data[i]  = d_p + (m)*i;
187:     }
188:   }

190:   /* Parse the IS and update local tables and the outgoing buf with the data*/
191:   {
192:     int     n_i,*data_i,isz_i,*outdat_j,ctr_j;
193:     PetscBT table_i;

195:     for (i=0; i<imax; i++) {
196:       PetscMemzero(ctr,size*sizeof(int));
197:       n_i     = n[i];
198:       table_i = table[i];
199:       idx_i   = idx[i];
200:       data_i  = data[i];
201:       isz_i   = isz[i];
202:       for (j=0;  j<n_i; j++) {  /* parse the indices of each IS */
203:         row  = idx_i[j];
204:         proc = rtable[row];
205:         if (proc != rank) { /* copy to the outgoing buffer */
206:           ctr[proc]++;
207:           *ptr[proc] = row;
208:           ptr[proc]++;
209:         } else { /* Update the local table */
210:           if (!PetscBTLookupSet(table_i,row)) { data_i[isz_i++] = row;}
211:         }
212:       }
213:       /* Update the headers for the current IS */
214:       for (j=0; j<size; j++) { /* Can Optimise this loop by using pa[] */
215:         if ((ctr_j = ctr[j])) {
216:           outdat_j        = outdat[j];
217:           k               = ++outdat_j[0];
218:           outdat_j[2*k]   = ctr_j;
219:           outdat_j[2*k-1] = i;
220:         }
221:       }
222:       isz[i] = isz_i;
223:     }
224:   }
225: 


228:   /*  Now  post the sends */
229:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);
230:   for (i=0; i<nrqs; ++i) {
231:     j    = pa[i];
232:     MPI_Isend(outdat[j],w1[j],MPI_INT,j,tag1,comm,s_waits1+i);
233:   }
234: 
235:   /* No longer need the original indices*/
236:   for (i=0; i<imax; ++i) {
237:     ISRestoreIndices(is[i],idx+i);
238:   }
239:   PetscFree(idx);

241:   for (i=0; i<imax; ++i) {
242:     ISDestroy(is[i]);
243:   }
244: 
245:   /* Do Local work*/
246:   MatIncreaseOverlap_MPIAIJ_Local(C,imax,table,isz,data);

248:   /* Receive messages*/
249:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&recv_status);
250:   MPI_Waitall(nrqr,r_waits1,recv_status);
251: 
252:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status);
253:   MPI_Waitall(nrqs,s_waits1,s_status);

255:   /* Phase 1 sends are complete - deallocate buffers */
256:   PetscFree(outdat);
257:   PetscFree(w1);

259:   PetscMalloc((nrqr+1)*sizeof(int *),&xdata);
260:   PetscMalloc((nrqr+1)*sizeof(int),&isz1);
261:   MatIncreaseOverlap_MPIAIJ_Receive(C,nrqr,rbuf,xdata,isz1);
262:   PetscFree(rbuf);

264: 
265:  /* Send the data back*/
266:   /* Do a global reduction to know the buffer space req for incoming messages*/
267:   {
268:     int *rw1;
269: 
270:     PetscMalloc(size*sizeof(int),&rw1);
271:     PetscMemzero(rw1,size*sizeof(int));

273:     for (i=0; i<nrqr; ++i) {
274:       proc      = recv_status[i].MPI_SOURCE;
275:       if (proc != onodes1[i]) SETERRQ(1,"MPI_SOURCE mismatch");
276:       rw1[proc] = isz1[i];
277:     }
278:     PetscFree(onodes1);
279:     PetscFree(olengths1);

281:     /* Determine the number of messages to expect, their lengths, from from-ids */
282:     PetscGatherMessageLengths(comm,nrqr,nrqs,rw1,&onodes2,&olengths2);
283:     PetscFree(rw1);
284:   }
285:   /* Now post the Irecvs corresponding to these messages */
286:   PetscPostIrecvInt(comm,tag2,nrqs,onodes2,olengths2,&rbuf2,&r_waits2);

288:   /*  Now  post the sends */
289:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);
290:   for (i=0; i<nrqr; ++i) {
291:     j    = recv_status[i].MPI_SOURCE;
292:     MPI_Isend(xdata[i],isz1[i],MPI_INT,j,tag2,comm,s_waits2+i);
293:   }

295:   /* receive work done on other processors*/
296:   {
297:     int         idex,is_no,ct1,max,*rbuf2_i,isz_i,*data_i,jmax;
298:     PetscBT     table_i;
299:     MPI_Status  *status2;
300: 
301:     PetscMalloc((PetscMax(nrqr,nrqs)+1)*sizeof(MPI_Status),&status2);
302:     for (i=0; i<nrqs; ++i) {
303:       MPI_Waitany(nrqs,r_waits2,&idex,status2+i);
304:       /* Process the message*/
305:       rbuf2_i = rbuf2[idex];
306:       ct1     = 2*rbuf2_i[0]+1;
307:       jmax    = rbuf2[idex][0];
308:       for (j=1; j<=jmax; j++) {
309:         max     = rbuf2_i[2*j];
310:         is_no   = rbuf2_i[2*j-1];
311:         isz_i   = isz[is_no];
312:         data_i  = data[is_no];
313:         table_i = table[is_no];
314:         for (k=0; k<max; k++,ct1++) {
315:           row = rbuf2_i[ct1];
316:           if (!PetscBTLookupSet(table_i,row)) { data_i[isz_i++] = row;}
317:         }
318:         isz[is_no] = isz_i;
319:       }
320:     }

322:     MPI_Waitall(nrqr,s_waits2,status2);
323:     PetscFree(status2);
324:   }
325: 
326:   for (i=0; i<imax; ++i) {
327:     ISCreateGeneral(PETSC_COMM_SELF,isz[i],data[i],is+i);
328:   }
329: 
330:   PetscFree(onodes2);
331:   PetscFree(olengths2);

333:   PetscFree(pa);
334:   PetscFree(rbuf2);
335:   PetscFree(s_waits1);
336:   PetscFree(r_waits1);
337:   PetscFree(s_waits2);
338:   PetscFree(r_waits2);
339:   PetscFree(table);
340:   PetscFree(s_status);
341:   PetscFree(recv_status);
342:   PetscFree(xdata[0]);
343:   PetscFree(xdata);
344:   PetscFree(isz1);
345:   return(0);
346: }

350: /*  
351:    MatIncreaseOverlap_MPIAIJ_Local - Called by MatincreaseOverlap, to do 
352:        the work on the local processor.

354:      Inputs:
355:       C      - MAT_MPIAIJ;
356:       imax - total no of index sets processed at a time;
357:       table  - an array of char - size = m bits.
358:       
359:      Output:
360:       isz    - array containing the count of the solution elements correspondign
361:                to each index set;
362:       data   - pointer to the solutions
363: */
364: static int MatIncreaseOverlap_MPIAIJ_Local(Mat C,int imax,PetscBT *table,int *isz,int **data)
365: {
366:   Mat_MPIAIJ *c = (Mat_MPIAIJ*)C->data;
367:   Mat        A = c->A,B = c->B;
368:   Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
369:   int        start,end,val,max,rstart,cstart,*ai,*aj;
370:   int        *bi,*bj,*garray,i,j,k,row,*data_i,isz_i;
371:   PetscBT    table_i;

374:   rstart = c->rstart;
375:   cstart = c->cstart;
376:   ai     = a->i;
377:   aj     = a->j;
378:   bi     = b->i;
379:   bj     = b->j;
380:   garray = c->garray;

382: 
383:   for (i=0; i<imax; i++) {
384:     data_i  = data[i];
385:     table_i = table[i];
386:     isz_i   = isz[i];
387:     for (j=0,max=isz[i]; j<max; j++) {
388:       row   = data_i[j] - rstart;
389:       start = ai[row];
390:       end   = ai[row+1];
391:       for (k=start; k<end; k++) { /* Amat */
392:         val = aj[k] + cstart;
393:         if (!PetscBTLookupSet(table_i,val)) { data_i[isz_i++] = val;}
394:       }
395:       start = bi[row];
396:       end   = bi[row+1];
397:       for (k=start; k<end; k++) { /* Bmat */
398:         val = garray[bj[k]];
399:         if (!PetscBTLookupSet(table_i,val)) { data_i[isz_i++] = val;}
400:       }
401:     }
402:     isz[i] = isz_i;
403:   }
404:   return(0);
405: }

409: /*     
410:       MatIncreaseOverlap_MPIAIJ_Receive - Process the recieved messages,
411:          and return the output

413:          Input:
414:            C    - the matrix
415:            nrqr - no of messages being processed.
416:            rbuf - an array of pointers to the recieved requests
417:            
418:          Output:
419:            xdata - array of messages to be sent back
420:            isz1  - size of each message

422:   For better efficiency perhaps we should malloc seperately each xdata[i],
423: then if a remalloc is required we need only copy the data for that one row
424: rather then all previous rows as it is now where a single large chunck of 
425: memory is used.

427: */
428: static int MatIncreaseOverlap_MPIAIJ_Receive(Mat C,int nrqr,int **rbuf,int **xdata,int * isz1)
429: {
430:   Mat_MPIAIJ  *c = (Mat_MPIAIJ*)C->data;
431:   Mat         A = c->A,B = c->B;
432:   Mat_SeqAIJ  *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
433:   int         rstart,cstart,*ai,*aj,*bi,*bj,*garray,i,j,k;
434:   int         row,total_sz,ct,ct1,ct2,ct3,mem_estimate,oct2,l,start,end;
435:   int         val,max1,max2,rank,m,no_malloc =0,*tmp,new_estimate,ctr;
436:   int         *rbuf_i,kmax,rbuf_0,ierr;
437:   PetscBT     xtable;

440:   rank   = c->rank;
441:   m      = C->M;
442:   rstart = c->rstart;
443:   cstart = c->cstart;
444:   ai     = a->i;
445:   aj     = a->j;
446:   bi     = b->i;
447:   bj     = b->j;
448:   garray = c->garray;
449: 
450: 
451:   for (i=0,ct=0,total_sz=0; i<nrqr; ++i) {
452:     rbuf_i  =  rbuf[i];
453:     rbuf_0  =  rbuf_i[0];
454:     ct     += rbuf_0;
455:     for (j=1; j<=rbuf_0; j++) { total_sz += rbuf_i[2*j]; }
456:   }
457: 
458:   if (C->m) max1 = ct*(a->nz + b->nz)/C->m;
459:   else      max1 = 1;
460:   mem_estimate = 3*((total_sz > max1 ? total_sz : max1)+1);
461:   PetscMalloc(mem_estimate*sizeof(int),&xdata[0]);
462:   ++no_malloc;
463:   PetscBTCreate(m,xtable);
464:   PetscMemzero(isz1,nrqr*sizeof(int));
465: 
466:   ct3 = 0;
467:   for (i=0; i<nrqr; i++) { /* for easch mesg from proc i */
468:     rbuf_i =  rbuf[i];
469:     rbuf_0 =  rbuf_i[0];
470:     ct1    =  2*rbuf_0+1;
471:     ct2    =  ct1;
472:     ct3    += ct1;
473:     for (j=1; j<=rbuf_0; j++) { /* for each IS from proc i*/
474:       PetscBTMemzero(m,xtable);
475:       oct2 = ct2;
476:       kmax = rbuf_i[2*j];
477:       for (k=0; k<kmax; k++,ct1++) {
478:         row = rbuf_i[ct1];
479:         if (!PetscBTLookupSet(xtable,row)) {
480:           if (!(ct3 < mem_estimate)) {
481:             new_estimate = (int)(1.5*mem_estimate)+1;
482:             PetscMalloc(new_estimate*sizeof(int),&tmp);
483:             PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));
484:             PetscFree(xdata[0]);
485:             xdata[0]     = tmp;
486:             mem_estimate = new_estimate; ++no_malloc;
487:             for (ctr=1; ctr<=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
488:           }
489:           xdata[i][ct2++] = row;
490:           ct3++;
491:         }
492:       }
493:       for (k=oct2,max2=ct2; k<max2; k++) {
494:         row   = xdata[i][k] - rstart;
495:         start = ai[row];
496:         end   = ai[row+1];
497:         for (l=start; l<end; l++) {
498:           val = aj[l] + cstart;
499:           if (!PetscBTLookupSet(xtable,val)) {
500:             if (!(ct3 < mem_estimate)) {
501:               new_estimate = (int)(1.5*mem_estimate)+1;
502:               PetscMalloc(new_estimate*sizeof(int),&tmp);
503:               PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));
504:               PetscFree(xdata[0]);
505:               xdata[0]     = tmp;
506:               mem_estimate = new_estimate; ++no_malloc;
507:               for (ctr=1; ctr<=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
508:             }
509:             xdata[i][ct2++] = val;
510:             ct3++;
511:           }
512:         }
513:         start = bi[row];
514:         end   = bi[row+1];
515:         for (l=start; l<end; l++) {
516:           val = garray[bj[l]];
517:           if (!PetscBTLookupSet(xtable,val)) {
518:             if (!(ct3 < mem_estimate)) {
519:               new_estimate = (int)(1.5*mem_estimate)+1;
520:               PetscMalloc(new_estimate*sizeof(int),&tmp);
521:               PetscMemcpy(tmp,xdata[0],mem_estimate*sizeof(int));
522:               PetscFree(xdata[0]);
523:               xdata[0]     = tmp;
524:               mem_estimate = new_estimate; ++no_malloc;
525:               for (ctr =1; ctr <=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
526:             }
527:             xdata[i][ct2++] = val;
528:             ct3++;
529:           }
530:         }
531:       }
532:       /* Update the header*/
533:       xdata[i][2*j]   = ct2 - oct2; /* Undo the vector isz1 and use only a var*/
534:       xdata[i][2*j-1] = rbuf_i[2*j-1];
535:     }
536:     xdata[i][0] = rbuf_0;
537:     xdata[i+1]  = xdata[i] + ct2;
538:     isz1[i]     = ct2; /* size of each message */
539:   }
540:   PetscBTDestroy(xtable);
541:   PetscLogInfo(0,"MatIncreaseOverlap_MPIAIJ:[%d] Allocated %d bytes, required %d bytes, no of mallocs = %d\n",rank,mem_estimate, ct3,no_malloc);
542:   return(0);
543: }
544: /* -------------------------------------------------------------------------*/
545: EXTERN int MatGetSubMatrices_MPIAIJ_Local(Mat,int,const IS[],const IS[],MatReuse,Mat*);
546: EXTERN int MatAssemblyEnd_SeqAIJ(Mat,MatAssemblyType);
547: /*
548:     Every processor gets the entire matrix
549: */
552: int MatGetSubMatrix_MPIAIJ_All(Mat A,MatReuse scall,Mat *Bin[])
553: {
554:   Mat          B;
555:   Mat_MPIAIJ   *a = (Mat_MPIAIJ *)A->data;
556:   Mat_SeqAIJ   *b,*ad = (Mat_SeqAIJ*)a->A->data,*bd = (Mat_SeqAIJ*)a->B->data;
557:   int          ierr,sendcount,*recvcounts = 0,*displs = 0,size,i,*rstarts = a->rowners,rank,n,cnt,j;
558:   int          m,*b_sendj,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf;
559:   PetscScalar  *sendbuf,*recvbuf,*a_sendbuf,*b_sendbuf;

562:   MPI_Comm_size(A->comm,&size);
563:   MPI_Comm_rank(A->comm,&rank);

565:   if (scall == MAT_INITIAL_MATRIX) {
566:     /* ----------------------------------------------------------------
567:          Tell every processor the number of nonzeros per row
568:     */
569:     PetscMalloc(A->M*sizeof(int),&lens);
570:     for (i=a->rstart; i<a->rend; i++) {
571:       lens[i] = ad->i[i-a->rstart+1] - ad->i[i-a->rstart] + bd->i[i-a->rstart+1] - bd->i[i-a->rstart];
572:     }
573:     sendcount = a->rend - a->rstart;
574:     PetscMalloc(2*size*sizeof(int),&recvcounts);
575:     displs     = recvcounts + size;
576:     for (i=0; i<size; i++) {
577:       recvcounts[i] = a->rowners[i+1] - a->rowners[i];
578:       displs[i]     = a->rowners[i];
579:     }
580:     MPI_Allgatherv(lens+a->rstart,sendcount,MPI_INT,lens,recvcounts,displs,MPI_INT,A->comm);

582:     /* ---------------------------------------------------------------
583:          Create the sequential matrix
584:     */
585:     MatCreateSeqAIJ(PETSC_COMM_SELF,A->M,A->N,0,lens,&B);
586:     PetscMalloc(sizeof(Mat),Bin);
587:     **Bin = B;
588:     b = (Mat_SeqAIJ *)B->data;

590:     /*--------------------------------------------------------------------
591:        Copy my part of matrix column indices over
592:     */
593:     sendcount  = ad->nz + bd->nz;
594:     jsendbuf   = b->j + b->i[rstarts[rank]];
595:     a_jsendbuf = ad->j;
596:     b_jsendbuf = bd->j;
597:     n          = a->rend - a->rstart;
598:     cnt        = 0;
599:     for (i=0; i<n; i++) {

601:       /* put in lower diagonal portion */
602:       m = bd->i[i+1] - bd->i[i];
603:       while (m > 0) {
604:         /* is it above diagonal (in bd (compressed) numbering) */
605:         if (garray[*b_jsendbuf] > a->rstart + i) break;
606:         jsendbuf[cnt++] = garray[*b_jsendbuf++];
607:         m--;
608:       }

610:       /* put in diagonal portion */
611:       for (j=ad->i[i]; j<ad->i[i+1]; j++) {
612:         jsendbuf[cnt++] = a->rstart + *a_jsendbuf++;
613:       }

615:       /* put in upper diagonal portion */
616:       while (m-- > 0) {
617:         jsendbuf[cnt++] = garray[*b_jsendbuf++];
618:       }
619:     }
620:     if (cnt != sendcount) SETERRQ2(1,"Corrupted PETSc matrix: nz given %d actual nz %d",sendcount,cnt);

622:     /*--------------------------------------------------------------------
623:        Gather all column indices to all processors
624:     */
625:     for (i=0; i<size; i++) {
626:       recvcounts[i] = 0;
627:       for (j=a->rowners[i]; j<a->rowners[i+1]; j++) {
628:         recvcounts[i] += lens[j];
629:       }
630:     }
631:     displs[0]  = 0;
632:     for (i=1; i<size; i++) {
633:       displs[i] = displs[i-1] + recvcounts[i-1];
634:     }
635:     MPI_Allgatherv(jsendbuf,sendcount,MPI_INT,b->j,recvcounts,displs,MPI_INT,A->comm);

637:     /*--------------------------------------------------------------------
638:         Assemble the matrix into useable form (note numerical values not yet set)
639:     */
640:     /* set the b->ilen (length of each row) values */
641:     PetscMemcpy(b->ilen,lens,A->M*sizeof(int));
642:     /* set the b->i indices */
643:     b->i[0] = 0;
644:     for (i=1; i<=A->M; i++) {
645:       b->i[i] = b->i[i-1] + lens[i-1];
646:     }
647:     PetscFree(lens);
648:     MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
649:     MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
650:   } else {
651:     B  = **Bin;
652:     b = (Mat_SeqAIJ *)B->data;
653:   }

655:   /*--------------------------------------------------------------------
656:        Copy my part of matrix numerical values into the values location 
657:   */
658:   sendcount = ad->nz + bd->nz;
659:   sendbuf   = b->a + b->i[rstarts[rank]];
660:   a_sendbuf = ad->a;
661:   b_sendbuf = bd->a;
662:   b_sendj   = bd->j;
663:   n         = a->rend - a->rstart;
664:   cnt       = 0;
665:   for (i=0; i<n; i++) {

667:     /* put in lower diagonal portion */
668:     m = bd->i[i+1] - bd->i[i];
669:     while (m > 0) {
670:       /* is it above diagonal (in bd (compressed) numbering) */
671:       if (garray[*b_sendj] > a->rstart + i) break;
672:       sendbuf[cnt++] = *b_sendbuf++;
673:       m--;
674:       b_sendj++;
675:     }

677:     /* put in diagonal portion */
678:     for (j=ad->i[i]; j<ad->i[i+1]; j++) {
679:       sendbuf[cnt++] = *a_sendbuf++;
680:     }

682:     /* put in upper diagonal portion */
683:     while (m-- > 0) {
684:       sendbuf[cnt++] = *b_sendbuf++;
685:       b_sendj++;
686:     }
687:   }
688:   if (cnt != sendcount) SETERRQ2(1,"Corrupted PETSc matrix: nz given %d actual nz %d",sendcount,cnt);
689: 
690:   /* ----------------------------------------------------------------- 
691:      Gather all numerical values to all processors 
692:   */
693:   if (!recvcounts) {
694:     PetscMalloc(2*size*sizeof(int),&recvcounts);
695:     displs = recvcounts + size;
696:   }
697:   for (i=0; i<size; i++) {
698:     recvcounts[i] = b->i[rstarts[i+1]] - b->i[rstarts[i]];
699:   }
700:   displs[0]  = 0;
701:   for (i=1; i<size; i++) {
702:     displs[i] = displs[i-1] + recvcounts[i-1];
703:   }
704:   recvbuf   = b->a;
705:   MPI_Allgatherv(sendbuf,sendcount,MPIU_SCALAR,recvbuf,recvcounts,displs,MPIU_SCALAR,A->comm);
706:   PetscFree(recvcounts);

708:   return(0);
709: }

713: int MatGetSubMatrices_MPIAIJ(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submat[])
714: {
715:   int         nmax,nstages_local,nstages,i,pos,max_no,ierr,nrow,ncol;
716:   PetscTruth  rowflag,colflag,wantallmatrix = PETSC_FALSE,twantallmatrix;

719:   /*
720:        Check for special case each processor gets entire matrix
721:   */
722:   if (ismax == 1 && C->M == C->N) {
723:     ISIdentity(*isrow,&rowflag);
724:     ISIdentity(*iscol,&colflag);
725:     ISGetLocalSize(*isrow,&nrow);
726:     ISGetLocalSize(*iscol,&ncol);
727:     if (rowflag && colflag && nrow == C->M && ncol == C->N) {
728:       wantallmatrix = PETSC_TRUE;
729:       PetscOptionsGetLogical(C->prefix,"-use_fast_submatrix",&wantallmatrix,PETSC_NULL);
730:     }
731:   }
732:   MPI_Allreduce(&wantallmatrix,&twantallmatrix,1,MPI_INT,MPI_MIN,C->comm);
733:   if (twantallmatrix) {
734:     MatGetSubMatrix_MPIAIJ_All(C,scall,submat);
735:     return(0);
736:   }

738:   /* Allocate memory to hold all the submatrices */
739:   if (scall != MAT_REUSE_MATRIX) {
740:     PetscMalloc((ismax+1)*sizeof(Mat),submat);
741:   }
742:   /* Determine the number of stages through which submatrices are done */
743:   nmax          = 20*1000000 / (C->N * sizeof(int));
744:   if (!nmax) nmax = 1;
745:   nstages_local = ismax/nmax + ((ismax % nmax)?1:0);

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

750:   for (i=0,pos=0; i<nstages; i++) {
751:     if (pos+nmax <= ismax) max_no = nmax;
752:     else if (pos == ismax) max_no = 0;
753:     else                   max_no = ismax-pos;
754:     MatGetSubMatrices_MPIAIJ_Local(C,max_no,isrow+pos,iscol+pos,scall,*submat+pos);
755:     pos += max_no;
756:   }
757:   return(0);
758: }
759: /* -------------------------------------------------------------------------*/
762: int MatGetSubMatrices_MPIAIJ_Local(Mat C,int ismax,const IS isrow[],const IS iscol[],MatReuse scall,Mat *submats)
763: {
764:   Mat_MPIAIJ  *c = (Mat_MPIAIJ*)C->data;
765:   Mat         A = c->A;
766:   Mat_SeqAIJ  *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)c->B->data,*mat;
767:   int         **irow,**icol,*nrow,*ncol,*w1,*w2,*w3,*w4,*rtable,start,end,size;
768:   int         **sbuf1,**sbuf2,rank,m,i,j,k,l,ct1,ct2,ierr,**rbuf1,row,proc;
769:   int         nrqs,msz,**ptr,idex,*req_size,*ctr,*pa,*tmp,tcol,nrqr;
770:   int         **rbuf3,*req_source,**sbuf_aj,**rbuf2,max1,max2,**rmap;
771:   int         **cmap,**lens,is_no,ncols,*cols,mat_i,*mat_j,tmp2,jmax,*irow_i;
772:   int         len,ctr_j,*sbuf1_j,*sbuf_aj_i,*rbuf1_i,kmax,*cmap_i,*lens_i;
773:   int         *rmap_i,tag0,tag1,tag2,tag3;
774:   MPI_Request *s_waits1,*r_waits1,*s_waits2,*r_waits2,*r_waits3;
775:   MPI_Request *r_waits4,*s_waits3,*s_waits4;
776:   MPI_Status  *r_status1,*r_status2,*s_status1,*s_status3,*s_status2;
777:   MPI_Status  *r_status3,*r_status4,*s_status4;
778:   MPI_Comm    comm;
779:   PetscScalar **rbuf4,**sbuf_aa,*vals,*mat_a,*sbuf_aa_i;
780:   PetscTruth  sorted;
781:   int         *onodes1,*olengths1;

784:   comm   = C->comm;
785:   tag0   = C->tag;
786:   size   = c->size;
787:   rank   = c->rank;
788:   m      = C->M;
789: 
790:   /* Get some new tags to keep the communication clean */
791:   PetscObjectGetNewTag((PetscObject)C,&tag1);
792:   PetscObjectGetNewTag((PetscObject)C,&tag2);
793:   PetscObjectGetNewTag((PetscObject)C,&tag3);

795:     /* Check if the col indices are sorted */
796:   for (i=0; i<ismax; i++) {
797:     ISSorted(isrow[i],&sorted);
798:     if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"ISrow is not sorted");
799:     ISSorted(iscol[i],&sorted);
800:     /*    if (!sorted) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"IScol is not sorted"); */
801:   }

803:   len    = (2*ismax+1)*(sizeof(int*)+ sizeof(int)) + (m+1)*sizeof(int);
804:   PetscMalloc(len,&irow);
805:   icol   = irow + ismax;
806:   nrow   = (int*)(icol + ismax);
807:   ncol   = nrow + ismax;
808:   rtable = ncol + ismax;

810:   for (i=0; i<ismax; i++) {
811:     ISGetIndices(isrow[i],&irow[i]);
812:     ISGetIndices(iscol[i],&icol[i]);
813:     ISGetLocalSize(isrow[i],&nrow[i]);
814:     ISGetLocalSize(iscol[i],&ncol[i]);
815:   }

817:   /* Create hash table for the mapping :row -> proc*/
818:   for (i=0,j=0; i<size; i++) {
819:     jmax = c->rowners[i+1];
820:     for (; j<jmax; j++) {
821:       rtable[j] = i;
822:     }
823:   }

825:   /* evaluate communication - mesg to who, length of mesg, and buffer space
826:      required. Based on this, buffers are allocated, and data copied into them*/
827:   PetscMalloc(size*4*sizeof(int),&w1); /* mesg size */
828:   w2     = w1 + size;      /* if w2[i] marked, then a message to proc i*/
829:   w3     = w2 + size;      /* no of IS that needs to be sent to proc i */
830:   w4     = w3 + size;      /* temp work space used in determining w1, w2, w3 */
831:   PetscMemzero(w1,size*3*sizeof(int)); /* initialize work vector*/
832:   for (i=0; i<ismax; i++) {
833:     PetscMemzero(w4,size*sizeof(int)); /* initialize work vector*/
834:     jmax   = nrow[i];
835:     irow_i = irow[i];
836:     for (j=0; j<jmax; j++) {
837:       row  = irow_i[j];
838:       proc = rtable[row];
839:       w4[proc]++;
840:     }
841:     for (j=0; j<size; j++) {
842:       if (w4[j]) { w1[j] += w4[j];  w3[j]++;}
843:     }
844:   }
845: 
846:   nrqs     = 0;              /* no of outgoing messages */
847:   msz      = 0;              /* total mesg length (for all procs) */
848:   w1[rank] = 0;              /* no mesg sent to self */
849:   w3[rank] = 0;
850:   for (i=0; i<size; i++) {
851:     if (w1[i])  { w2[i] = 1; nrqs++;} /* there exists a message to proc i */
852:   }
853:   PetscMalloc((nrqs+1)*sizeof(int),&pa); /*(proc -array)*/
854:   for (i=0,j=0; i<size; i++) {
855:     if (w1[i]) { pa[j] = i; j++; }
856:   }

858:   /* Each message would have a header = 1 + 2*(no of IS) + data */
859:   for (i=0; i<nrqs; i++) {
860:     j     = pa[i];
861:     w1[j] += w2[j] + 2* w3[j];
862:     msz   += w1[j];
863:   }

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

869:   /* Now post the Irecvs corresponding to these messages */
870:   PetscPostIrecvInt(comm,tag0,nrqr,onodes1,olengths1,&rbuf1,&r_waits1);
871: 
872:   PetscFree(onodes1);
873:   PetscFree(olengths1);
874: 
875:   /* Allocate Memory for outgoing messages */
876:   len      = 2*size*sizeof(int*) + 2*msz*sizeof(int) + size*sizeof(int);
877:   PetscMalloc(len,&sbuf1);
878:   ptr      = sbuf1 + size;   /* Pointers to the data in outgoing buffers */
879:   PetscMemzero(sbuf1,2*size*sizeof(int*));
880:   /* allocate memory for outgoing data + buf to receive the first reply */
881:   tmp      = (int*)(ptr + size);
882:   ctr      = tmp + 2*msz;

884:   {
885:     int *iptr = tmp,ict = 0;
886:     for (i=0; i<nrqs; i++) {
887:       j         = pa[i];
888:       iptr     += ict;
889:       sbuf1[j]  = iptr;
890:       ict       = w1[j];
891:     }
892:   }

894:   /* Form the outgoing messages */
895:   /* Initialize the header space */
896:   for (i=0; i<nrqs; i++) {
897:     j           = pa[i];
898:     sbuf1[j][0] = 0;
899:     PetscMemzero(sbuf1[j]+1,2*w3[j]*sizeof(int));
900:     ptr[j]      = sbuf1[j] + 2*w3[j] + 1;
901:   }
902: 
903:   /* Parse the isrow and copy data into outbuf */
904:   for (i=0; i<ismax; i++) {
905:     PetscMemzero(ctr,size*sizeof(int));
906:     irow_i = irow[i];
907:     jmax   = nrow[i];
908:     for (j=0; j<jmax; j++) {  /* parse the indices of each IS */
909:       row  = irow_i[j];
910:       proc = rtable[row];
911:       if (proc != rank) { /* copy to the outgoing buf*/
912:         ctr[proc]++;
913:         *ptr[proc] = row;
914:         ptr[proc]++;
915:       }
916:     }
917:     /* Update the headers for the current IS */
918:     for (j=0; j<size; j++) { /* Can Optimise this loop too */
919:       if ((ctr_j = ctr[j])) {
920:         sbuf1_j        = sbuf1[j];
921:         k              = ++sbuf1_j[0];
922:         sbuf1_j[2*k]   = ctr_j;
923:         sbuf1_j[2*k-1] = i;
924:       }
925:     }
926:   }

928:   /*  Now  post the sends */
929:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&s_waits1);
930:   for (i=0; i<nrqs; ++i) {
931:     j    = pa[i];
932:     MPI_Isend(sbuf1[j],w1[j],MPI_INT,j,tag0,comm,s_waits1+i);
933:   }

935:   /* Post Receives to capture the buffer size */
936:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits2);
937:   PetscMalloc((nrqs+1)*sizeof(int *),&rbuf2);
938:   rbuf2[0] = tmp + msz;
939:   for (i=1; i<nrqs; ++i) {
940:     rbuf2[i] = rbuf2[i-1]+w1[pa[i-1]];
941:   }
942:   for (i=0; i<nrqs; ++i) {
943:     j    = pa[i];
944:     MPI_Irecv(rbuf2[i],w1[j],MPI_INT,j,tag1,comm,r_waits2+i);
945:   }

947:   /* Send to other procs the buf size they should allocate */
948: 

950:   /* Receive messages*/
951:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits2);
952:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&r_status1);
953:   len         = 2*nrqr*sizeof(int) + (nrqr+1)*sizeof(int*);
954:   PetscMalloc(len,&sbuf2);
955:   req_size    = (int*)(sbuf2 + nrqr);
956:   req_source  = req_size + nrqr;
957: 
958:   {
959:     Mat_SeqAIJ *sA = (Mat_SeqAIJ*)c->A->data,*sB = (Mat_SeqAIJ*)c->B->data;
960:     int        *sAi = sA->i,*sBi = sB->i,id,rstart = c->rstart;
961:     int        *sbuf2_i;

963:     for (i=0; i<nrqr; ++i) {
964:       MPI_Waitany(nrqr,r_waits1,&idex,r_status1+i);
965:       req_size[idex] = 0;
966:       rbuf1_i         = rbuf1[idex];
967:       start           = 2*rbuf1_i[0] + 1;
968:       MPI_Get_count(r_status1+i,MPI_INT,&end);
969:       PetscMalloc((end+1)*sizeof(int),&sbuf2[idex]);
970:       sbuf2_i         = sbuf2[idex];
971:       for (j=start; j<end; j++) {
972:         id               = rbuf1_i[j] - rstart;
973:         ncols            = sAi[id+1] - sAi[id] + sBi[id+1] - sBi[id];
974:         sbuf2_i[j]       = ncols;
975:         req_size[idex] += ncols;
976:       }
977:       req_source[idex] = r_status1[i].MPI_SOURCE;
978:       /* form the header */
979:       sbuf2_i[0]   = req_size[idex];
980:       for (j=1; j<start; j++) { sbuf2_i[j] = rbuf1_i[j]; }
981:       MPI_Isend(sbuf2_i,end,MPI_INT,req_source[idex],tag1,comm,s_waits2+i);
982:     }
983:   }
984:   PetscFree(r_status1);
985:   PetscFree(r_waits1);

987:   /*  recv buffer sizes */
988:   /* Receive messages*/
989: 
990:   PetscMalloc((nrqs+1)*sizeof(int*),&rbuf3);
991:   PetscMalloc((nrqs+1)*sizeof(PetscScalar*),&rbuf4);
992:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits3);
993:   PetscMalloc((nrqs+1)*sizeof(MPI_Request),&r_waits4);
994:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status2);

996:   for (i=0; i<nrqs; ++i) {
997:     MPI_Waitany(nrqs,r_waits2,&idex,r_status2+i);
998:     PetscMalloc((rbuf2[idex][0]+1)*sizeof(int),&rbuf3[idex]);
999:     PetscMalloc((rbuf2[idex][0]+1)*sizeof(PetscScalar),&rbuf4[idex]);
1000:     MPI_Irecv(rbuf3[idex],rbuf2[idex][0],MPI_INT,r_status2[i].MPI_SOURCE,tag2,comm,r_waits3+idex);
1001:     MPI_Irecv(rbuf4[idex],rbuf2[idex][0],MPIU_SCALAR,r_status2[i].MPI_SOURCE,tag3,comm,r_waits4+idex);
1002:   }
1003:   PetscFree(r_status2);
1004:   PetscFree(r_waits2);
1005: 
1006:   /* Wait on sends1 and sends2 */
1007:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&s_status1);
1008:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status2);

1010:   MPI_Waitall(nrqs,s_waits1,s_status1);
1011:   MPI_Waitall(nrqr,s_waits2,s_status2);
1012:   PetscFree(s_status1);
1013:   PetscFree(s_status2);
1014:   PetscFree(s_waits1);
1015:   PetscFree(s_waits2);

1017:   /* Now allocate buffers for a->j, and send them off */
1018:   PetscMalloc((nrqr+1)*sizeof(int*),&sbuf_aj);
1019:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
1020:   PetscMalloc((j+1)*sizeof(int),&sbuf_aj[0]);
1021:   for (i=1; i<nrqr; i++)  sbuf_aj[i] = sbuf_aj[i-1] + req_size[i-1];
1022: 
1023:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits3);
1024:   {
1025:     int nzA,nzB,*a_i = a->i,*b_i = b->i,imark;
1026:     int *cworkA,*cworkB,cstart = c->cstart,rstart = c->rstart,*bmap = c->garray;
1027:     int *a_j = a->j,*b_j = b->j,ctmp;

1029:     for (i=0; i<nrqr; i++) {
1030:       rbuf1_i   = rbuf1[i];
1031:       sbuf_aj_i = sbuf_aj[i];
1032:       ct1       = 2*rbuf1_i[0] + 1;
1033:       ct2       = 0;
1034:       for (j=1,max1=rbuf1_i[0]; j<=max1; j++) {
1035:         kmax = rbuf1[i][2*j];
1036:         for (k=0; k<kmax; k++,ct1++) {
1037:           row    = rbuf1_i[ct1] - rstart;
1038:           nzA    = a_i[row+1] - a_i[row];     nzB = b_i[row+1] - b_i[row];
1039:           ncols  = nzA + nzB;
1040:           cworkA = a_j + a_i[row]; cworkB = b_j + b_i[row];

1042:           /* load the column indices for this row into cols*/
1043:           cols  = sbuf_aj_i + ct2;
1044: 
1045:           for (l=0; l<nzB; l++) {
1046:             if ((ctmp = bmap[cworkB[l]]) < cstart)  cols[l] = ctmp;
1047:             else break;
1048:           }
1049:           imark = l;
1050:           for (l=0; l<nzA; l++)   cols[imark+l] = cstart + cworkA[l];
1051:           for (l=imark; l<nzB; l++) cols[nzA+l] = bmap[cworkB[l]];

1053:           ct2 += ncols;
1054:         }
1055:       }
1056:       MPI_Isend(sbuf_aj_i,req_size[i],MPI_INT,req_source[i],tag2,comm,s_waits3+i);
1057:     }
1058:   }
1059:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status3);
1060:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status3);

1062:   /* Allocate buffers for a->a, and send them off */
1063:   PetscMalloc((nrqr+1)*sizeof(PetscScalar*),&sbuf_aa);
1064:   for (i=0,j=0; i<nrqr; i++) j += req_size[i];
1065:   PetscMalloc((j+1)*sizeof(PetscScalar),&sbuf_aa[0]);
1066:   for (i=1; i<nrqr; i++)  sbuf_aa[i] = sbuf_aa[i-1] + req_size[i-1];
1067: 
1068:   PetscMalloc((nrqr+1)*sizeof(MPI_Request),&s_waits4);
1069:   {
1070:     int    nzA,nzB,*a_i = a->i,*b_i = b->i, *cworkB,imark;
1071:     int    cstart = c->cstart,rstart = c->rstart,*bmap = c->garray;
1072:     int    *b_j = b->j;
1073:     PetscScalar *vworkA,*vworkB,*a_a = a->a,*b_a = b->a;
1074: 
1075:     for (i=0; i<nrqr; i++) {
1076:       rbuf1_i   = rbuf1[i];
1077:       sbuf_aa_i = sbuf_aa[i];
1078:       ct1       = 2*rbuf1_i[0]+1;
1079:       ct2       = 0;
1080:       for (j=1,max1=rbuf1_i[0]; j<=max1; j++) {
1081:         kmax = rbuf1_i[2*j];
1082:         for (k=0; k<kmax; k++,ct1++) {
1083:           row    = rbuf1_i[ct1] - rstart;
1084:           nzA    = a_i[row+1] - a_i[row];     nzB = b_i[row+1] - b_i[row];
1085:           ncols  = nzA + nzB;
1086:           cworkB = b_j + b_i[row];
1087:           vworkA = a_a + a_i[row];
1088:           vworkB = b_a + b_i[row];

1090:           /* load the column values for this row into vals*/
1091:           vals  = sbuf_aa_i+ct2;
1092: 
1093:           for (l=0; l<nzB; l++) {
1094:             if ((bmap[cworkB[l]]) < cstart)  vals[l] = vworkB[l];
1095:             else break;
1096:           }
1097:           imark = l;
1098:           for (l=0; l<nzA; l++)   vals[imark+l] = vworkA[l];
1099:           for (l=imark; l<nzB; l++) vals[nzA+l] = vworkB[l];
1100: 
1101:           ct2 += ncols;
1102:         }
1103:       }
1104:       MPI_Isend(sbuf_aa_i,req_size[i],MPIU_SCALAR,req_source[i],tag3,comm,s_waits4+i);
1105:     }
1106:   }
1107:   PetscMalloc((nrqs+1)*sizeof(MPI_Status),&r_status4);
1108:   PetscMalloc((nrqr+1)*sizeof(MPI_Status),&s_status4);
1109:   PetscFree(rbuf1);

1111:   /* Form the matrix */
1112:   /* create col map */
1113:   {
1114:     int *icol_i;
1115: 
1116:     len     = (1+ismax)*sizeof(int*)+ ismax*C->N*sizeof(int);
1117:     PetscMalloc(len,&cmap);
1118:     cmap[0] = (int *)(cmap + ismax);
1119:     PetscMemzero(cmap[0],(1+ismax*C->N)*sizeof(int));
1120:     for (i=1; i<ismax; i++) { cmap[i] = cmap[i-1] + C->N; }
1121:     for (i=0; i<ismax; i++) {
1122:       jmax   = ncol[i];
1123:       icol_i = icol[i];
1124:       cmap_i = cmap[i];
1125:       for (j=0; j<jmax; j++) {
1126:         cmap_i[icol_i[j]] = j+1;
1127:       }
1128:     }
1129:   }

1131:   /* Create lens which is required for MatCreate... */
1132:   for (i=0,j=0; i<ismax; i++) { j += nrow[i]; }
1133:   len     = (1+ismax)*sizeof(int*)+ j*sizeof(int);
1134:   PetscMalloc(len,&lens);
1135:   lens[0] = (int *)(lens + ismax);
1136:   PetscMemzero(lens[0],j*sizeof(int));
1137:   for (i=1; i<ismax; i++) { lens[i] = lens[i-1] + nrow[i-1]; }
1138: 
1139:   /* Update lens from local data */
1140:   for (i=0; i<ismax; i++) {
1141:     jmax   = nrow[i];
1142:     cmap_i = cmap[i];
1143:     irow_i = irow[i];
1144:     lens_i = lens[i];
1145:     for (j=0; j<jmax; j++) {
1146:       row  = irow_i[j];
1147:       proc = rtable[row];
1148:       if (proc == rank) {
1149:         MatGetRow_MPIAIJ(C,row,&ncols,&cols,0);
1150:         for (k=0; k<ncols; k++) {
1151:           if (cmap_i[cols[k]]) { lens_i[j]++;}
1152:         }
1153:         MatRestoreRow_MPIAIJ(C,row,&ncols,&cols,0);
1154:       }
1155:     }
1156:   }
1157: 
1158:   /* Create row map*/
1159:   len     = (1+ismax)*sizeof(int*)+ ismax*C->M*sizeof(int);
1160:   PetscMalloc(len,&rmap);
1161:   rmap[0] = (int *)(rmap + ismax);
1162:   PetscMemzero(rmap[0],ismax*C->M*sizeof(int));
1163:   for (i=1; i<ismax; i++) { rmap[i] = rmap[i-1] + C->M;}
1164:   for (i=0; i<ismax; i++) {
1165:     rmap_i = rmap[i];
1166:     irow_i = irow[i];
1167:     jmax   = nrow[i];
1168:     for (j=0; j<jmax; j++) {
1169:       rmap_i[irow_i[j]] = j;
1170:     }
1171:   }
1172: 
1173:   /* Update lens from offproc data */
1174:   {
1175:     int *rbuf2_i,*rbuf3_i,*sbuf1_i;

1177:     for (tmp2=0; tmp2<nrqs; tmp2++) {
1178:       MPI_Waitany(nrqs,r_waits3,&i,r_status3+tmp2);
1179:       idex   = pa[i];
1180:       sbuf1_i = sbuf1[idex];
1181:       jmax    = sbuf1_i[0];
1182:       ct1     = 2*jmax+1;
1183:       ct2     = 0;
1184:       rbuf2_i = rbuf2[i];
1185:       rbuf3_i = rbuf3[i];
1186:       for (j=1; j<=jmax; j++) {
1187:         is_no   = sbuf1_i[2*j-1];
1188:         max1    = sbuf1_i[2*j];
1189:         lens_i  = lens[is_no];
1190:         cmap_i  = cmap[is_no];
1191:         rmap_i  = rmap[is_no];
1192:         for (k=0; k<max1; k++,ct1++) {
1193:           row  = rmap_i[sbuf1_i[ct1]]; /* the val in the new matrix to be */
1194:           max2 = rbuf2_i[ct1];
1195:           for (l=0; l<max2; l++,ct2++) {
1196:             if (cmap_i[rbuf3_i[ct2]]) {
1197:               lens_i[row]++;
1198:             }
1199:           }
1200:         }
1201:       }
1202:     }
1203:   }
1204:   PetscFree(r_status3);
1205:   PetscFree(r_waits3);
1206:   MPI_Waitall(nrqr,s_waits3,s_status3);
1207:   PetscFree(s_status3);
1208:   PetscFree(s_waits3);

1210:   /* Create the submatrices */
1211:   if (scall == MAT_REUSE_MATRIX) {
1212:     PetscTruth flag;

1214:     /*
1215:         Assumes new rows are same length as the old rows,hence bug!
1216:     */
1217:     for (i=0; i<ismax; i++) {
1218:       mat = (Mat_SeqAIJ *)(submats[i]->data);
1219:       if ((submats[i]->m != nrow[i]) || (submats[i]->n != ncol[i])) {
1220:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong size");
1221:       }
1222:       PetscMemcmp(mat->ilen,lens[i],submats[i]->m*sizeof(int),&flag);
1223:       if (flag == PETSC_FALSE) {
1224:         SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. wrong no of nonzeros");
1225:       }
1226:       /* Initial matrix as if empty */
1227:       PetscMemzero(mat->ilen,submats[i]->m*sizeof(int));
1228:       submats[i]->factor = C->factor;
1229:     }
1230:   } else {
1231:     for (i=0; i<ismax; i++) {
1232:       MatCreateSeqAIJ(PETSC_COMM_SELF,nrow[i],ncol[i],0,lens[i],submats+i);
1233:     }
1234:   }

1236:   /* Assemble the matrices */
1237:   /* First assemble the local rows */
1238:   {
1239:     int    ilen_row,*imat_ilen,*imat_j,*imat_i,old_row;
1240:     PetscScalar *imat_a;
1241: 
1242:     for (i=0; i<ismax; i++) {
1243:       mat       = (Mat_SeqAIJ*)submats[i]->data;
1244:       imat_ilen = mat->ilen;
1245:       imat_j    = mat->j;
1246:       imat_i    = mat->i;
1247:       imat_a    = mat->a;
1248:       cmap_i    = cmap[i];
1249:       rmap_i    = rmap[i];
1250:       irow_i    = irow[i];
1251:       jmax      = nrow[i];
1252:       for (j=0; j<jmax; j++) {
1253:         row      = irow_i[j];
1254:         proc     = rtable[row];
1255:         if (proc == rank) {
1256:           old_row  = row;
1257:           row      = rmap_i[row];
1258:           ilen_row = imat_ilen[row];
1259:           MatGetRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);
1260:           mat_i    = imat_i[row] ;
1261:           mat_a    = imat_a + mat_i;
1262:           mat_j    = imat_j + mat_i;
1263:           for (k=0; k<ncols; k++) {
1264:             if ((tcol = cmap_i[cols[k]])) {
1265:               *mat_j++ = tcol - 1;
1266:               *mat_a++ = vals[k];
1267:               ilen_row++;
1268:             }
1269:           }
1270:           MatRestoreRow_MPIAIJ(C,old_row,&ncols,&cols,&vals);
1271:           imat_ilen[row] = ilen_row;
1272:         }
1273:       }
1274:     }
1275:   }

1277:   /*   Now assemble the off proc rows*/
1278:   {
1279:     int    *sbuf1_i,*rbuf2_i,*rbuf3_i,*imat_ilen,ilen;
1280:     int    *imat_j,*imat_i;
1281:     PetscScalar *imat_a,*rbuf4_i;

1283:     for (tmp2=0; tmp2<nrqs; tmp2++) {
1284:       MPI_Waitany(nrqs,r_waits4,&i,r_status4+tmp2);
1285:       idex   = pa[i];
1286:       sbuf1_i = sbuf1[idex];
1287:       jmax    = sbuf1_i[0];
1288:       ct1     = 2*jmax + 1;
1289:       ct2     = 0;
1290:       rbuf2_i = rbuf2[i];
1291:       rbuf3_i = rbuf3[i];
1292:       rbuf4_i = rbuf4[i];
1293:       for (j=1; j<=jmax; j++) {
1294:         is_no     = sbuf1_i[2*j-1];
1295:         rmap_i    = rmap[is_no];
1296:         cmap_i    = cmap[is_no];
1297:         mat       = (Mat_SeqAIJ*)submats[is_no]->data;
1298:         imat_ilen = mat->ilen;
1299:         imat_j    = mat->j;
1300:         imat_i    = mat->i;
1301:         imat_a    = mat->a;
1302:         max1      = sbuf1_i[2*j];
1303:         for (k=0; k<max1; k++,ct1++) {
1304:           row   = sbuf1_i[ct1];
1305:           row   = rmap_i[row];
1306:           ilen  = imat_ilen[row];
1307:           mat_i = imat_i[row] ;
1308:           mat_a = imat_a + mat_i;
1309:           mat_j = imat_j + mat_i;
1310:           max2 = rbuf2_i[ct1];
1311:           for (l=0; l<max2; l++,ct2++) {
1312:             if ((tcol = cmap_i[rbuf3_i[ct2]])) {
1313:               *mat_j++ = tcol - 1;
1314:               *mat_a++ = rbuf4_i[ct2];
1315:               ilen++;
1316:             }
1317:           }
1318:           imat_ilen[row] = ilen;
1319:         }
1320:       }
1321:     }
1322:   }
1323:   PetscFree(r_status4);
1324:   PetscFree(r_waits4);
1325:   MPI_Waitall(nrqr,s_waits4,s_status4);
1326:   PetscFree(s_waits4);
1327:   PetscFree(s_status4);

1329:   /* Restore the indices */
1330:   for (i=0; i<ismax; i++) {
1331:     ISRestoreIndices(isrow[i],irow+i);
1332:     ISRestoreIndices(iscol[i],icol+i);
1333:   }

1335:   /* Destroy allocated memory */
1336:   PetscFree(irow);
1337:   PetscFree(w1);
1338:   PetscFree(pa);

1340:   PetscFree(sbuf1);
1341:   PetscFree(rbuf2);
1342:   for (i=0; i<nrqr; ++i) {
1343:     PetscFree(sbuf2[i]);
1344:   }
1345:   for (i=0; i<nrqs; ++i) {
1346:     PetscFree(rbuf3[i]);
1347:     PetscFree(rbuf4[i]);
1348:   }

1350:   PetscFree(sbuf2);
1351:   PetscFree(rbuf3);
1352:   PetscFree(rbuf4);
1353:   PetscFree(sbuf_aj[0]);
1354:   PetscFree(sbuf_aj);
1355:   PetscFree(sbuf_aa[0]);
1356:   PetscFree(sbuf_aa);
1357: 
1358:   PetscFree(cmap);
1359:   PetscFree(rmap);
1360:   PetscFree(lens);

1362:   for (i=0; i<ismax; i++) {
1363:     MatAssemblyBegin(submats[i],MAT_FINAL_ASSEMBLY);
1364:     MatAssemblyEnd(submats[i],MAT_FINAL_ASSEMBLY);
1365:   }
1366:   return(0);
1367: }