Actual source code: mpiaij.c

  1: #define PETSCMAT_DLL

 3:  #include src/mat/impls/aij/mpi/mpiaij.h
 4:  #include src/inline/spops.h

  6: /* 
  7:   Local utility routine that creates a mapping from the global column 
  8: number to the local number in the off-diagonal part of the local 
  9: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at 
 10: a slightly higher hash table cost; without it it is not scalable (each processor
 11: has an order N integer array but is fast to acess.
 12: */
 15: PetscErrorCode CreateColmap_MPIAIJ_Private(Mat mat)
 16: {
 17:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
 19:   PetscInt       n = aij->B->n,i;

 22: #if defined (PETSC_USE_CTABLE)
 23:   PetscTableCreate(n,&aij->colmap);
 24:   for (i=0; i<n; i++){
 25:     PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1);
 26:   }
 27: #else
 28:   PetscMalloc((mat->N+1)*sizeof(PetscInt),&aij->colmap);
 29:   PetscLogObjectMemory(mat,mat->N*sizeof(PetscInt));
 30:   PetscMemzero(aij->colmap,mat->N*sizeof(PetscInt));
 31:   for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
 32: #endif
 33:   return(0);
 34: }


 37: #define CHUNKSIZE   15
 38: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv) \
 39: { \
 40:     if (col <= lastcol1) low1 = 0; else high1 = nrow1; \
 41:     lastcol1 = col;\
 42:     while (high1-low1 > 5) { \
 43:       t = (low1+high1)/2; \
 44:       if (rp1[t] > col) high1 = t; \
 45:       else             low1  = t; \
 46:     } \
 47:       for (_i=low1; _i<high1; _i++) { \
 48:         if (rp1[_i] > col) break; \
 49:         if (rp1[_i] == col) { \
 50:           if (addv == ADD_VALUES) ap1[_i] += value;   \
 51:           else                    ap1[_i] = value; \
 52:           goto a_noinsert; \
 53:         } \
 54:       }  \
 55:       if (value == 0.0 && ignorezeroentries) goto a_noinsert; \
 56:       if (nonew == 1) goto a_noinsert; \
 57:       if (nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
 58:       MatSeqXAIJReallocateAIJ(a,1,nrow1,row,col,rmax1,aa,ai,aj,am,rp1,ap1,aimax,nonew); \
 59:       N = nrow1++ - 1; a->nz++; \
 60:       /* shift up all the later entries in this row */ \
 61:       for (ii=N; ii>=_i; ii--) { \
 62:         rp1[ii+1] = rp1[ii]; \
 63:         ap1[ii+1] = ap1[ii]; \
 64:       } \
 65:       rp1[_i] = col;  \
 66:       ap1[_i] = value;  \
 67:       a_noinsert: ; \
 68:       ailen[row] = nrow1; \
 69: } 


 72: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv) \
 73: { \
 74:     if (col <= lastcol2) low2 = 0; else high2 = nrow2; \
 75:     lastcol2 = col;\
 76:     while (high2-low2 > 5) { \
 77:       t = (low2+high2)/2; \
 78:       if (rp2[t] > col) high2 = t; \
 79:       else             low2  = t; \
 80:     } \
 81:        for (_i=low2; _i<high2; _i++) { \
 82:         if (rp2[_i] > col) break; \
 83:         if (rp2[_i] == col) { \
 84:           if (addv == ADD_VALUES) ap2[_i] += value;   \
 85:           else                    ap2[_i] = value; \
 86:           goto b_noinsert; \
 87:         } \
 88:       }  \
 89:       if (value == 0.0 && ignorezeroentries) goto b_noinsert; \
 90:       if (nonew == 1) goto b_noinsert; \
 91:       if (nonew == -1) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
 92:       MatSeqXAIJReallocateAIJ(b,1,nrow2,row,col,rmax2,ba,bi,bj,bm,rp2,ap2,bimax,nonew); \
 93:       N = nrow2++ - 1; b->nz++; \
 94:       /* shift up all the later entries in this row */ \
 95:       for (ii=N; ii>=_i; ii--) { \
 96:         rp2[ii+1] = rp2[ii]; \
 97:         ap2[ii+1] = ap2[ii]; \
 98:       } \
 99:       rp2[_i] = col;  \
100:       ap2[_i] = value;  \
101:       b_noinsert: ; \
102:       bilen[row] = nrow2; \
103: }

107: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
108: {
109:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
110:   PetscScalar    value;
112:   PetscInt       i,j,rstart = aij->rstart,rend = aij->rend;
113:   PetscInt       cstart = aij->cstart,cend = aij->cend,row,col;
114:   PetscTruth     roworiented = aij->roworiented;

116:   /* Some Variables required in the macro */
117:   Mat            A = aij->A;
118:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)A->data;
119:   PetscInt       *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
120:   PetscScalar    *aa = a->a;
121:   PetscTruth     ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
122:   Mat            B = aij->B;
123:   Mat_SeqAIJ     *b = (Mat_SeqAIJ*)B->data;
124:   PetscInt       *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->m,am = aij->A->m;
125:   PetscScalar    *ba = b->a;

127:   PetscInt       *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
128:   PetscInt       nonew = a->nonew;
129:   PetscScalar    *ap1,*ap2;

132:   for (i=0; i<m; i++) {
133:     if (im[i] < 0) continue;
134: #if defined(PETSC_USE_DEBUG)
135:     if (im[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->M-1);
136: #endif
137:     if (im[i] >= rstart && im[i] < rend) {
138:       row      = im[i] - rstart;
139:       lastcol1 = -1;
140:       rp1      = aj + ai[row];
141:       ap1      = aa + ai[row];
142:       rmax1    = aimax[row];
143:       nrow1    = ailen[row];
144:       low1     = 0;
145:       high1    = nrow1;
146:       lastcol2 = -1;
147:       rp2      = bj + bi[row];
148:       ap2      = ba + bi[row];
149:       rmax2    = bimax[row];
150:       nrow2    = bilen[row];
151:       low2     = 0;
152:       high2    = nrow2;

154:       for (j=0; j<n; j++) {
155:         if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
156:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
157:         if (in[j] >= cstart && in[j] < cend){
158:           col = in[j] - cstart;
159:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
160:         } else if (in[j] < 0) continue;
161: #if defined(PETSC_USE_DEBUG)
162:         else if (in[j] >= mat->N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->N-1);}
163: #endif
164:         else {
165:           if (mat->was_assembled) {
166:             if (!aij->colmap) {
167:               CreateColmap_MPIAIJ_Private(mat);
168:             }
169: #if defined (PETSC_USE_CTABLE)
170:             PetscTableFind(aij->colmap,in[j]+1,&col);
171:             col--;
172: #else
173:             col = aij->colmap[in[j]] - 1;
174: #endif
175:             if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
176:               DisAssemble_MPIAIJ(mat);
177:               col =  in[j];
178:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
179:               B = aij->B;
180:               b = (Mat_SeqAIJ*)B->data;
181:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
182:               rp2      = bj + bi[row];
183:               ap2      = ba + bi[row];
184:               rmax2    = bimax[row];
185:               nrow2    = bilen[row];
186:               low2     = 0;
187:               high2    = nrow2;
188:               bm       = aij->B->m;
189:               ba = b->a;
190:             }
191:           } else col = in[j];
192:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
193:         }
194:       }
195:     } else {
196:       if (!aij->donotstash) {
197:         if (roworiented) {
198:           if (ignorezeroentries && v[i*n] == 0.0) continue;
199:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
200:         } else {
201:           if (ignorezeroentries && v[i] == 0.0) continue;
202:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
203:         }
204:       }
205:     }
206:   }
207:   return(0);
208: }


213: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
214: {
215:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
217:   PetscInt       i,j,rstart = aij->rstart,rend = aij->rend;
218:   PetscInt       cstart = aij->cstart,cend = aij->cend,row,col;

221:   for (i=0; i<m; i++) {
222:     if (idxm[i] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);
223:     if (idxm[i] >= mat->M) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->M-1);
224:     if (idxm[i] >= rstart && idxm[i] < rend) {
225:       row = idxm[i] - rstart;
226:       for (j=0; j<n; j++) {
227:         if (idxn[j] < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]);
228:         if (idxn[j] >= mat->N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->N-1);
229:         if (idxn[j] >= cstart && idxn[j] < cend){
230:           col = idxn[j] - cstart;
231:           MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
232:         } else {
233:           if (!aij->colmap) {
234:             CreateColmap_MPIAIJ_Private(mat);
235:           }
236: #if defined (PETSC_USE_CTABLE)
237:           PetscTableFind(aij->colmap,idxn[j]+1,&col);
238:           col --;
239: #else
240:           col = aij->colmap[idxn[j]] - 1;
241: #endif
242:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
243:           else {
244:             MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
245:           }
246:         }
247:       }
248:     } else {
249:       SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
250:     }
251:   }
252:   return(0);
253: }

257: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
258: {
259:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
261:   PetscInt       nstash,reallocs;
262:   InsertMode     addv;

265:   if (aij->donotstash) {
266:     return(0);
267:   }

269:   /* make sure all processors are either in INSERTMODE or ADDMODE */
270:   MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,mat->comm);
271:   if (addv == (ADD_VALUES|INSERT_VALUES)) {
272:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
273:   }
274:   mat->insertmode = addv; /* in case this processor had no cache */

276:   MatStashScatterBegin_Private(&mat->stash,aij->rowners);
277:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
278:   PetscLogInfo((aij->A,"MatAssemblyBegin_MPIAIJ:Stash has %D entries, uses %D mallocs.\n",nstash,reallocs));
279:   return(0);
280: }

284: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
285: {
286:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
287:   Mat_SeqAIJ     *a=(Mat_SeqAIJ *)aij->A->data,*b= (Mat_SeqAIJ *)aij->B->data;
289:   PetscMPIInt    n;
290:   PetscInt       i,j,rstart,ncols,flg;
291:   PetscInt       *row,*col,other_disassembled;
292:   PetscScalar    *val;
293:   InsertMode     addv = mat->insertmode;

296:   if (!aij->donotstash) {
297:     while (1) {
298:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
299:       if (!flg) break;

301:       for (i=0; i<n;) {
302:         /* Now identify the consecutive vals belonging to the same row */
303:         for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
304:         if (j < n) ncols = j-i;
305:         else       ncols = n-i;
306:         /* Now assemble all these values with a single function call */
307:         MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
308:         i = j;
309:       }
310:     }
311:     MatStashScatterEnd_Private(&mat->stash);
312:   }
313:   a->compressedrow.use     = PETSC_FALSE;
314:   MatAssemblyBegin(aij->A,mode);
315:   MatAssemblyEnd(aij->A,mode);

317:   /* determine if any processor has disassembled, if so we must 
318:      also disassemble ourselfs, in order that we may reassemble. */
319:   /*
320:      if nonzero structure of submatrix B cannot change then we know that
321:      no processor disassembled thus we can skip this stuff
322:   */
323:   if (!((Mat_SeqAIJ*)aij->B->data)->nonew)  {
324:     MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,mat->comm);
325:     if (mat->was_assembled && !other_disassembled) {
326:       DisAssemble_MPIAIJ(mat);
327:     }
328:   }
329:   /* reaccess the b because aij->B was changed in MatSetValues() or DisAssemble() */
330:   b    = (Mat_SeqAIJ *)aij->B->data;

332:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
333:     MatSetUpMultiply_MPIAIJ(mat);
334:   }
335:   MatSetOption(aij->B,MAT_DO_NOT_USE_INODES);
336:   b->compressedrow.use = PETSC_TRUE;
337:   MatAssemblyBegin(aij->B,mode);
338:   MatAssemblyEnd(aij->B,mode);

340:   if (aij->rowvalues) {
341:     PetscFree(aij->rowvalues);
342:     aij->rowvalues = 0;
343:   }

345:   /* used by MatAXPY() */
346:   a->xtoy = 0; b->xtoy = 0;
347:   a->XtoY = 0; b->XtoY = 0;

349:   return(0);
350: }

354: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
355: {
356:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;

360:   MatZeroEntries(l->A);
361:   MatZeroEntries(l->B);
362:   return(0);
363: }

367: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag)
368: {
369:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;
371:   PetscMPIInt    size = l->size,imdex,n,rank = l->rank,tag = A->tag,lastidx = -1;
372:   PetscInt       i,*owners = l->rowners;
373:   PetscInt       *nprocs,j,idx,nsends,row;
374:   PetscInt       nmax,*svalues,*starts,*owner,nrecvs;
375:   PetscInt       *rvalues,count,base,slen,*source;
376:   PetscInt       *lens,*lrows,*values,rstart=l->rstart;
377:   MPI_Comm       comm = A->comm;
378:   MPI_Request    *send_waits,*recv_waits;
379:   MPI_Status     recv_status,*send_status;
380: #if defined(PETSC_DEBUG)
381:   PetscTruth     found = PETSC_FALSE;
382: #endif

385:   /*  first count number of contributors to each processor */
386:   PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
387:   PetscMemzero(nprocs,2*size*sizeof(PetscInt));
388:   PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
389:   j = 0;
390:   for (i=0; i<N; i++) {
391:     if (lastidx > (idx = rows[i])) j = 0;
392:     lastidx = idx;
393:     for (; j<size; j++) {
394:       if (idx >= owners[j] && idx < owners[j+1]) {
395:         nprocs[2*j]++;
396:         nprocs[2*j+1] = 1;
397:         owner[i] = j;
398: #if defined(PETSC_DEBUG)
399:         found = PETSC_TRUE;
400: #endif
401:         break;
402:       }
403:     }
404: #if defined(PETSC_DEBUG)
405:     if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
406:     found = PETSC_FALSE;
407: #endif
408:   }
409:   nsends = 0;  for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}

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

414:   /* post receives:   */
415:   PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
416:   PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
417:   for (i=0; i<nrecvs; i++) {
418:     MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
419:   }

421:   /* do sends:
422:       1) starts[i] gives the starting index in svalues for stuff going to 
423:          the ith processor
424:   */
425:   PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
426:   PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
427:   PetscMalloc((size+1)*sizeof(PetscInt),&starts);
428:   starts[0] = 0;
429:   for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
430:   for (i=0; i<N; i++) {
431:     svalues[starts[owner[i]]++] = rows[i];
432:   }

434:   starts[0] = 0;
435:   for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
436:   count = 0;
437:   for (i=0; i<size; i++) {
438:     if (nprocs[2*i+1]) {
439:       MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
440:     }
441:   }
442:   PetscFree(starts);

444:   base = owners[rank];

446:   /*  wait on receives */
447:   PetscMalloc(2*(nrecvs+1)*sizeof(PetscInt),&lens);
448:   source = lens + nrecvs;
449:   count  = nrecvs; slen = 0;
450:   while (count) {
451:     MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
452:     /* unpack receives into our local space */
453:     MPI_Get_count(&recv_status,MPIU_INT,&n);
454:     source[imdex]  = recv_status.MPI_SOURCE;
455:     lens[imdex]    = n;
456:     slen          += n;
457:     count--;
458:   }
459:   PetscFree(recv_waits);
460: 
461:   /* move the data into the send scatter */
462:   PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
463:   count = 0;
464:   for (i=0; i<nrecvs; i++) {
465:     values = rvalues + i*nmax;
466:     for (j=0; j<lens[i]; j++) {
467:       lrows[count++] = values[j] - base;
468:     }
469:   }
470:   PetscFree(rvalues);
471:   PetscFree(lens);
472:   PetscFree(owner);
473:   PetscFree(nprocs);
474: 
475:   /* actually zap the local rows */
476:   /*
477:         Zero the required rows. If the "diagonal block" of the matrix
478:      is square and the user wishes to set the diagonal we use seperate
479:      code so that MatSetValues() is not called for each diagonal allocating
480:      new memory, thus calling lots of mallocs and slowing things down.

482:        Contributed by: Matthew Knepley
483:   */
484:   /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
485:   MatZeroRows(l->B,slen,lrows,0.0);
486:   if ((diag != 0.0) && (l->A->M == l->A->N)) {
487:     MatZeroRows(l->A,slen,lrows,diag);
488:   } else if (diag != 0.0) {
489:     MatZeroRows(l->A,slen,lrows,0.0);
490:     if (((Mat_SeqAIJ*)l->A->data)->nonew) {
491:       SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options\n\
492: MAT_NO_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
493:     }
494:     for (i = 0; i < slen; i++) {
495:       row  = lrows[i] + rstart;
496:       MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
497:     }
498:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
499:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
500:   } else {
501:     MatZeroRows(l->A,slen,lrows,0.0);
502:   }
503:   PetscFree(lrows);

505:   /* wait on sends */
506:   if (nsends) {
507:     PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
508:     MPI_Waitall(nsends,send_waits,send_status);
509:     PetscFree(send_status);
510:   }
511:   PetscFree(send_waits);
512:   PetscFree(svalues);

514:   return(0);
515: }

519: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
520: {
521:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
523:   PetscInt       nt;

526:   VecGetLocalSize(xx,&nt);
527:   if (nt != A->n) {
528:     SETERRQ2(PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->n,nt);
529:   }
530:   VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
531:   (*a->A->ops->mult)(a->A,xx,yy);
532:   VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
533:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
534:   return(0);
535: }

539: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
540: {
541:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

545:   VecScatterBegin(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
546:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
547:   VecScatterEnd(xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD,a->Mvctx);
548:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
549:   return(0);
550: }

554: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
555: {
556:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
558:   PetscTruth     merged;

561:   VecScatterGetMerged(a->Mvctx,&merged);
562:   /* do nondiagonal part */
563:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
564:   if (!merged) {
565:     /* send it on its way */
566:     VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
567:     /* do local part */
568:     (*a->A->ops->multtranspose)(a->A,xx,yy);
569:     /* receive remote parts: note this assumes the values are not actually */
570:     /* added in yy until the next line, */
571:     VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
572:   } else {
573:     /* do local part */
574:     (*a->A->ops->multtranspose)(a->A,xx,yy);
575:     /* send it on its way */
576:     VecScatterBegin(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
577:     /* values actually were received in the Begin() but we need to call this nop */
578:     VecScatterEnd(a->lvec,yy,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
579:   }
580:   return(0);
581: }

586: PetscErrorCode PETSCMAT_DLLEXPORT MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscTruth *f)
587: {
588:   MPI_Comm       comm;
589:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *) Amat->data, *Bij;
590:   Mat            Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
591:   IS             Me,Notme;
593:   PetscInt       M,N,first,last,*notme,i;
594:   PetscMPIInt    size;


598:   /* Easy test: symmetric diagonal block */
599:   Bij = (Mat_MPIAIJ *) Bmat->data; Bdia = Bij->A;
600:   MatIsTranspose(Adia,Bdia,tol,f);
601:   if (!*f) return(0);
602:   PetscObjectGetComm((PetscObject)Amat,&comm);
603:   MPI_Comm_size(comm,&size);
604:   if (size == 1) return(0);

606:   /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
607:   MatGetSize(Amat,&M,&N);
608:   MatGetOwnershipRange(Amat,&first,&last);
609:   PetscMalloc((N-last+first)*sizeof(PetscInt),&notme);
610:   for (i=0; i<first; i++) notme[i] = i;
611:   for (i=last; i<M; i++) notme[i-last+first] = i;
612:   ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,&Notme);
613:   ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
614:   MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
615:   Aoff = Aoffs[0];
616:   MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
617:   Boff = Boffs[0];
618:   MatIsTranspose(Aoff,Boff,tol,f);
619:   MatDestroyMatrices(1,&Aoffs);
620:   MatDestroyMatrices(1,&Boffs);
621:   ISDestroy(Me);
622:   ISDestroy(Notme);

624:   return(0);
625: }

630: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
631: {
632:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

636:   /* do nondiagonal part */
637:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
638:   /* send it on its way */
639:   VecScatterBegin(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
640:   /* do local part */
641:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
642:   /* receive remote parts */
643:   VecScatterEnd(a->lvec,zz,ADD_VALUES,SCATTER_REVERSE,a->Mvctx);
644:   return(0);
645: }

647: /*
648:   This only works correctly for square matrices where the subblock A->A is the 
649:    diagonal block
650: */
653: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
654: {
656:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

659:   if (A->M != A->N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
660:   if (a->rstart != a->cstart || a->rend != a->cend) {
661:     SETERRQ(PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
662:   }
663:   MatGetDiagonal(a->A,v);
664:   return(0);
665: }

669: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
670: {
671:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

675:   MatScale(a->A,aa);
676:   MatScale(a->B,aa);
677:   return(0);
678: }

682: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
683: {
684:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

688: #if defined(PETSC_USE_LOG)
689:   PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->M,mat->N);
690: #endif
691:   MatStashDestroy_Private(&mat->stash);
692:   PetscFree(aij->rowners);
693:   MatDestroy(aij->A);
694:   MatDestroy(aij->B);
695: #if defined (PETSC_USE_CTABLE)
696:   if (aij->colmap) {PetscTableDelete(aij->colmap);}
697: #else
698:   if (aij->colmap) {PetscFree(aij->colmap);}
699: #endif
700:   if (aij->garray) {PetscFree(aij->garray);}
701:   if (aij->lvec)   {VecDestroy(aij->lvec);}
702:   if (aij->Mvctx)  {VecScatterDestroy(aij->Mvctx);}
703:   if (aij->rowvalues) {PetscFree(aij->rowvalues);}
704:   PetscFree(aij);

706:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
707:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
708:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
709:   PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C","",PETSC_NULL);
710:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C","",PETSC_NULL);
711:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C","",PETSC_NULL);
712:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);
713:   return(0);
714: }

718: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
719: {
720:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
721:   Mat_SeqAIJ*       A = (Mat_SeqAIJ*)aij->A->data;
722:   Mat_SeqAIJ*       B = (Mat_SeqAIJ*)aij->B->data;
723:   PetscErrorCode    ierr;
724:   PetscMPIInt       rank,size,tag = ((PetscObject)viewer)->tag;
725:   int               fd;
726:   PetscInt          nz,header[4],*row_lengths,*range,rlen,i;
727:   PetscInt          nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = aij->cstart,rnz;
728:   PetscScalar       *column_values;

731:   MPI_Comm_rank(mat->comm,&rank);
732:   MPI_Comm_size(mat->comm,&size);
733:   nz   = A->nz + B->nz;
734:   if (!rank) {
735:     header[0] = MAT_FILE_COOKIE;
736:     header[1] = mat->M;
737:     header[2] = mat->N;
738:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,mat->comm);
739:     PetscViewerBinaryGetDescriptor(viewer,&fd);
740:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
741:     /* get largest number of rows any processor has */
742:     rlen = mat->m;
743:     PetscMapGetGlobalRange(mat->rmap,&range);
744:     for (i=1; i<size; i++) {
745:       rlen = PetscMax(rlen,range[i+1] - range[i]);
746:     }
747:   } else {
748:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,mat->comm);
749:     rlen = mat->m;
750:   }

752:   /* load up the local row counts */
753:   PetscMalloc((rlen+1)*sizeof(PetscInt),&row_lengths);
754:   for (i=0; i<mat->m; i++) {
755:     row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
756:   }

758:   /* store the row lengths to the file */
759:   if (!rank) {
760:     MPI_Status status;
761:     PetscBinaryWrite(fd,row_lengths,mat->m,PETSC_INT,PETSC_TRUE);
762:     for (i=1; i<size; i++) {
763:       rlen = range[i+1] - range[i];
764:       MPI_Recv(row_lengths,rlen,MPIU_INT,i,tag,mat->comm,&status);
765:       PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
766:     }
767:   } else {
768:     MPI_Send(row_lengths,mat->m,MPIU_INT,0,tag,mat->comm);
769:   }
770:   PetscFree(row_lengths);

772:   /* load up the local column indices */
773:   nzmax = nz; /* )th processor needs space a largest processor needs */
774:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,mat->comm);
775:   PetscMalloc((nzmax+1)*sizeof(PetscInt),&column_indices);
776:   cnt  = 0;
777:   for (i=0; i<mat->m; i++) {
778:     for (j=B->i[i]; j<B->i[i+1]; j++) {
779:       if ( (col = garray[B->j[j]]) > cstart) break;
780:       column_indices[cnt++] = col;
781:     }
782:     for (k=A->i[i]; k<A->i[i+1]; k++) {
783:       column_indices[cnt++] = A->j[k] + cstart;
784:     }
785:     for (; j<B->i[i+1]; j++) {
786:       column_indices[cnt++] = garray[B->j[j]];
787:     }
788:   }
789:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

791:   /* store the column indices to the file */
792:   if (!rank) {
793:     MPI_Status status;
794:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
795:     for (i=1; i<size; i++) {
796:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,mat->comm,&status);
797:       if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
798:       MPI_Recv(column_indices,rnz,MPIU_INT,i,tag,mat->comm,&status);
799:       PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
800:     }
801:   } else {
802:     MPI_Send(&nz,1,MPIU_INT,0,tag,mat->comm);
803:     MPI_Send(column_indices,nz,MPIU_INT,0,tag,mat->comm);
804:   }
805:   PetscFree(column_indices);

807:   /* load up the local column values */
808:   PetscMalloc((nzmax+1)*sizeof(PetscScalar),&column_values);
809:   cnt  = 0;
810:   for (i=0; i<mat->m; i++) {
811:     for (j=B->i[i]; j<B->i[i+1]; j++) {
812:       if ( garray[B->j[j]] > cstart) break;
813:       column_values[cnt++] = B->a[j];
814:     }
815:     for (k=A->i[i]; k<A->i[i+1]; k++) {
816:       column_values[cnt++] = A->a[k];
817:     }
818:     for (; j<B->i[i+1]; j++) {
819:       column_values[cnt++] = B->a[j];
820:     }
821:   }
822:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

824:   /* store the column values to the file */
825:   if (!rank) {
826:     MPI_Status status;
827:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
828:     for (i=1; i<size; i++) {
829:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,mat->comm,&status);
830:       if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
831:       MPI_Recv(column_values,rnz,MPIU_SCALAR,i,tag,mat->comm,&status);
832:       PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
833:     }
834:   } else {
835:     MPI_Send(&nz,1,MPIU_INT,0,tag,mat->comm);
836:     MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,mat->comm);
837:   }
838:   PetscFree(column_values);
839:   return(0);
840: }

844: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
845: {
846:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
847:   PetscErrorCode    ierr;
848:   PetscMPIInt       rank = aij->rank,size = aij->size;
849:   PetscTruth        isdraw,iascii,isbinary;
850:   PetscViewer       sviewer;
851:   PetscViewerFormat format;

854:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
855:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
856:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
857:   if (iascii) {
858:     PetscViewerGetFormat(viewer,&format);
859:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
860:       MatInfo    info;
861:       PetscTruth inodes;

863:       MPI_Comm_rank(mat->comm,&rank);
864:       MatGetInfo(mat,MAT_LOCAL,&info);
865:       MatInodeGetInodeSizes(aij->A,PETSC_NULL,(PetscInt **)&inodes,PETSC_NULL);
866:       if (!inodes) {
867:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
868:                                               rank,mat->m,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
869:       } else {
870:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
871:                     rank,mat->m,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
872:       }
873:       MatGetInfo(aij->A,MAT_LOCAL,&info);
874:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
875:       MatGetInfo(aij->B,MAT_LOCAL,&info);
876:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
877:       PetscViewerFlush(viewer);
878:       VecScatterView(aij->Mvctx,viewer);
879:       return(0);
880:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
881:       PetscInt   inodecount,inodelimit,*inodes;
882:       MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
883:       if (inodes) {
884:         PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
885:       } else {
886:         PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
887:       }
888:       return(0);
889:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
890:       return(0);
891:     }
892:   } else if (isbinary) {
893:     if (size == 1) {
894:       PetscObjectSetName((PetscObject)aij->A,mat->name);
895:       MatView(aij->A,viewer);
896:     } else {
897:       MatView_MPIAIJ_Binary(mat,viewer);
898:     }
899:     return(0);
900:   } else if (isdraw) {
901:     PetscDraw  draw;
902:     PetscTruth isnull;
903:     PetscViewerDrawGetDraw(viewer,0,&draw);
904:     PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
905:   }

907:   if (size == 1) {
908:     PetscObjectSetName((PetscObject)aij->A,mat->name);
909:     MatView(aij->A,viewer);
910:   } else {
911:     /* assemble the entire matrix onto first processor. */
912:     Mat         A;
913:     Mat_SeqAIJ  *Aloc;
914:     PetscInt    M = mat->M,N = mat->N,m,*ai,*aj,row,*cols,i,*ct;
915:     PetscScalar *a;

917:     MatCreate(mat->comm,&A);
918:     if (!rank) {
919:       MatSetSizes(A,M,N,M,N);
920:     } else {
921:       MatSetSizes(A,0,0,M,N);
922:     }
923:     /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
924:     MatSetType(A,MATMPIAIJ);
925:     MatMPIAIJSetPreallocation(A,0,PETSC_NULL,0,PETSC_NULL);
926:     PetscLogObjectParent(mat,A);

928:     /* copy over the A part */
929:     Aloc = (Mat_SeqAIJ*)aij->A->data;
930:     m = aij->A->m; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
931:     row = aij->rstart;
932:     for (i=0; i<ai[m]; i++) {aj[i] += aij->cstart ;}
933:     for (i=0; i<m; i++) {
934:       MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
935:       row++; a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
936:     }
937:     aj = Aloc->j;
938:     for (i=0; i<ai[m]; i++) {aj[i] -= aij->cstart;}

940:     /* copy over the B part */
941:     Aloc = (Mat_SeqAIJ*)aij->B->data;
942:     m    = aij->B->m;  ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
943:     row  = aij->rstart;
944:     PetscMalloc((ai[m]+1)*sizeof(PetscInt),&cols);
945:     ct   = cols;
946:     for (i=0; i<ai[m]; i++) {cols[i] = aij->garray[aj[i]];}
947:     for (i=0; i<m; i++) {
948:       MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
949:       row++; a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
950:     }
951:     PetscFree(ct);
952:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
953:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
954:     /* 
955:        Everyone has to call to draw the matrix since the graphics waits are
956:        synchronized across all processors that share the PetscDraw object
957:     */
958:     PetscViewerGetSingleton(viewer,&sviewer);
959:     if (!rank) {
960:       PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,mat->name);
961:       MatView(((Mat_MPIAIJ*)(A->data))->A,sviewer);
962:     }
963:     PetscViewerRestoreSingleton(viewer,&sviewer);
964:     MatDestroy(A);
965:   }
966:   return(0);
967: }

971: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
972: {
974:   PetscTruth     iascii,isdraw,issocket,isbinary;
975: 
977:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
978:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
979:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
980:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
981:   if (iascii || isdraw || isbinary || issocket) {
982:     MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
983:   } else {
984:     SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPIAIJ matrices",((PetscObject)viewer)->type_name);
985:   }
986:   return(0);
987: }



993: PetscErrorCode MatRelax_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
994: {
995:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
997:   Vec            bb1;
998:   PetscScalar    mone=-1.0;

1001:   if (its <= 0 || lits <= 0) SETERRQ2(PETSC_ERR_ARG_WRONG,"Relaxation requires global its %D and local its %D both positive",its,lits);

1003:   VecDuplicate(bb,&bb1);

1005:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
1006:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1007:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
1008:       its--;
1009:     }
1010: 
1011:     while (its--) {
1012:       VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1013:       VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);

1015:       /* update rhs: bb1 = bb - B*x */
1016:       VecScale(mat->lvec,mone);
1017:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1019:       /* local sweep */
1020:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
1021: 
1022:     }
1023:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP){
1024:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1025:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1026:       its--;
1027:     }
1028:     while (its--) {
1029:       VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1030:       VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);

1032:       /* update rhs: bb1 = bb - B*x */
1033:       VecScale(mat->lvec,mone);
1034:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1036:       /* local sweep */
1037:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1038: 
1039:     }
1040:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){
1041:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1042:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1043:       its--;
1044:     }
1045:     while (its--) {
1046:       VecScatterBegin(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);
1047:       VecScatterEnd(xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD,mat->Mvctx);

1049:       /* update rhs: bb1 = bb - B*x */
1050:       VecScale(mat->lvec,mone);
1051:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1053:       /* local sweep */
1054:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1055: 
1056:     }
1057:   } else {
1058:     SETERRQ(PETSC_ERR_SUP,"Parallel SOR not supported");
1059:   }

1061:   VecDestroy(bb1);
1062:   return(0);
1063: }

1067: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1068: {
1069:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1070:   Mat            A = mat->A,B = mat->B;
1072:   PetscReal      isend[5],irecv[5];

1075:   info->block_size     = 1.0;
1076:   MatGetInfo(A,MAT_LOCAL,info);
1077:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1078:   isend[3] = info->memory;  isend[4] = info->mallocs;
1079:   MatGetInfo(B,MAT_LOCAL,info);
1080:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1081:   isend[3] += info->memory;  isend[4] += info->mallocs;
1082:   if (flag == MAT_LOCAL) {
1083:     info->nz_used      = isend[0];
1084:     info->nz_allocated = isend[1];
1085:     info->nz_unneeded  = isend[2];
1086:     info->memory       = isend[3];
1087:     info->mallocs      = isend[4];
1088:   } else if (flag == MAT_GLOBAL_MAX) {
1089:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,matin->comm);
1090:     info->nz_used      = irecv[0];
1091:     info->nz_allocated = irecv[1];
1092:     info->nz_unneeded  = irecv[2];
1093:     info->memory       = irecv[3];
1094:     info->mallocs      = irecv[4];
1095:   } else if (flag == MAT_GLOBAL_SUM) {
1096:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,matin->comm);
1097:     info->nz_used      = irecv[0];
1098:     info->nz_allocated = irecv[1];
1099:     info->nz_unneeded  = irecv[2];
1100:     info->memory       = irecv[3];
1101:     info->mallocs      = irecv[4];
1102:   }
1103:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1104:   info->fill_ratio_needed = 0;
1105:   info->factor_mallocs    = 0;
1106:   info->rows_global       = (double)matin->M;
1107:   info->columns_global    = (double)matin->N;
1108:   info->rows_local        = (double)matin->m;
1109:   info->columns_local     = (double)matin->N;

1111:   return(0);
1112: }

1116: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op)
1117: {
1118:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1122:   switch (op) {
1123:   case MAT_NO_NEW_NONZERO_LOCATIONS:
1124:   case MAT_YES_NEW_NONZERO_LOCATIONS:
1125:   case MAT_COLUMNS_UNSORTED:
1126:   case MAT_COLUMNS_SORTED:
1127:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1128:   case MAT_KEEP_ZEROED_ROWS:
1129:   case MAT_NEW_NONZERO_LOCATION_ERR:
1130:   case MAT_USE_INODES:
1131:   case MAT_DO_NOT_USE_INODES:
1132:   case MAT_IGNORE_ZERO_ENTRIES:
1133:     MatSetOption(a->A,op);
1134:     MatSetOption(a->B,op);
1135:     break;
1136:   case MAT_ROW_ORIENTED:
1137:     a->roworiented = PETSC_TRUE;
1138:     MatSetOption(a->A,op);
1139:     MatSetOption(a->B,op);
1140:     break;
1141:   case MAT_ROWS_SORTED:
1142:   case MAT_ROWS_UNSORTED:
1143:   case MAT_YES_NEW_DIAGONALS:
1144:     PetscLogInfo((A,"MatSetOption_MPIAIJ:Option ignored\n"));
1145:     break;
1146:   case MAT_COLUMN_ORIENTED:
1147:     a->roworiented = PETSC_FALSE;
1148:     MatSetOption(a->A,op);
1149:     MatSetOption(a->B,op);
1150:     break;
1151:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1152:     a->donotstash = PETSC_TRUE;
1153:     break;
1154:   case MAT_NO_NEW_DIAGONALS:
1155:     SETERRQ(PETSC_ERR_SUP,"MAT_NO_NEW_DIAGONALS");
1156:   case MAT_SYMMETRIC:
1157:   case MAT_STRUCTURALLY_SYMMETRIC:
1158:   case MAT_HERMITIAN:
1159:   case MAT_SYMMETRY_ETERNAL:
1160:     MatSetOption(a->A,op);
1161:     break;
1162:   case MAT_NOT_SYMMETRIC:
1163:   case MAT_NOT_STRUCTURALLY_SYMMETRIC:
1164:   case MAT_NOT_HERMITIAN:
1165:   case MAT_NOT_SYMMETRY_ETERNAL:
1166:     break;
1167:   default:
1168:     SETERRQ(PETSC_ERR_SUP,"unknown option");
1169:   }
1170:   return(0);
1171: }

1175: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1176: {
1177:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1178:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1180:   PetscInt       i,*cworkA,*cworkB,**pcA,**pcB,cstart = mat->cstart;
1181:   PetscInt       nztot,nzA,nzB,lrow,rstart = mat->rstart,rend = mat->rend;
1182:   PetscInt       *cmap,*idx_p;

1185:   if (mat->getrowactive) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1186:   mat->getrowactive = PETSC_TRUE;

1188:   if (!mat->rowvalues && (idx || v)) {
1189:     /*
1190:         allocate enough space to hold information from the longest row.
1191:     */
1192:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1193:     PetscInt     max = 1,tmp;
1194:     for (i=0; i<matin->m; i++) {
1195:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1196:       if (max < tmp) { max = tmp; }
1197:     }
1198:     PetscMalloc(max*(sizeof(PetscInt)+sizeof(PetscScalar)),&mat->rowvalues);
1199:     mat->rowindices = (PetscInt*)(mat->rowvalues + max);
1200:   }

1202:   if (row < rstart || row >= rend) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Only local rows")
1203:   lrow = row - rstart;

1205:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1206:   if (!v)   {pvA = 0; pvB = 0;}
1207:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1208:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1209:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1210:   nztot = nzA + nzB;

1212:   cmap  = mat->garray;
1213:   if (v  || idx) {
1214:     if (nztot) {
1215:       /* Sort by increasing column numbers, assuming A and B already sorted */
1216:       PetscInt imark = -1;
1217:       if (v) {
1218:         *v = v_p = mat->rowvalues;
1219:         for (i=0; i<nzB; i++) {
1220:           if (cmap[cworkB[i]] < cstart)   v_p[i] = vworkB[i];
1221:           else break;
1222:         }
1223:         imark = i;
1224:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1225:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1226:       }
1227:       if (idx) {
1228:         *idx = idx_p = mat->rowindices;
1229:         if (imark > -1) {
1230:           for (i=0; i<imark; i++) {
1231:             idx_p[i] = cmap[cworkB[i]];
1232:           }
1233:         } else {
1234:           for (i=0; i<nzB; i++) {
1235:             if (cmap[cworkB[i]] < cstart)   idx_p[i] = cmap[cworkB[i]];
1236:             else break;
1237:           }
1238:           imark = i;
1239:         }
1240:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart + cworkA[i];
1241:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]];
1242:       }
1243:     } else {
1244:       if (idx) *idx = 0;
1245:       if (v)   *v   = 0;
1246:     }
1247:   }
1248:   *nz = nztot;
1249:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1250:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1251:   return(0);
1252: }

1256: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1257: {
1258:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1261:   if (!aij->getrowactive) {
1262:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1263:   }
1264:   aij->getrowactive = PETSC_FALSE;
1265:   return(0);
1266: }

1270: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1271: {
1272:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1273:   Mat_SeqAIJ     *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1275:   PetscInt       i,j,cstart = aij->cstart;
1276:   PetscReal      sum = 0.0;
1277:   PetscScalar    *v;

1280:   if (aij->size == 1) {
1281:      MatNorm(aij->A,type,norm);
1282:   } else {
1283:     if (type == NORM_FROBENIUS) {
1284:       v = amat->a;
1285:       for (i=0; i<amat->nz; i++) {
1286: #if defined(PETSC_USE_COMPLEX)
1287:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1288: #else
1289:         sum += (*v)*(*v); v++;
1290: #endif
1291:       }
1292:       v = bmat->a;
1293:       for (i=0; i<bmat->nz; i++) {
1294: #if defined(PETSC_USE_COMPLEX)
1295:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1296: #else
1297:         sum += (*v)*(*v); v++;
1298: #endif
1299:       }
1300:       MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPI_SUM,mat->comm);
1301:       *norm = sqrt(*norm);
1302:     } else if (type == NORM_1) { /* max column norm */
1303:       PetscReal *tmp,*tmp2;
1304:       PetscInt    *jj,*garray = aij->garray;
1305:       PetscMalloc((mat->N+1)*sizeof(PetscReal),&tmp);
1306:       PetscMalloc((mat->N+1)*sizeof(PetscReal),&tmp2);
1307:       PetscMemzero(tmp,mat->N*sizeof(PetscReal));
1308:       *norm = 0.0;
1309:       v = amat->a; jj = amat->j;
1310:       for (j=0; j<amat->nz; j++) {
1311:         tmp[cstart + *jj++ ] += PetscAbsScalar(*v);  v++;
1312:       }
1313:       v = bmat->a; jj = bmat->j;
1314:       for (j=0; j<bmat->nz; j++) {
1315:         tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1316:       }
1317:       MPI_Allreduce(tmp,tmp2,mat->N,MPIU_REAL,MPI_SUM,mat->comm);
1318:       for (j=0; j<mat->N; j++) {
1319:         if (tmp2[j] > *norm) *norm = tmp2[j];
1320:       }
1321:       PetscFree(tmp);
1322:       PetscFree(tmp2);
1323:     } else if (type == NORM_INFINITY) { /* max row norm */
1324:       PetscReal ntemp = 0.0;
1325:       for (j=0; j<aij->A->m; j++) {
1326:         v = amat->a + amat->i[j];
1327:         sum = 0.0;
1328:         for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1329:           sum += PetscAbsScalar(*v); v++;
1330:         }
1331:         v = bmat->a + bmat->i[j];
1332:         for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1333:           sum += PetscAbsScalar(*v); v++;
1334:         }
1335:         if (sum > ntemp) ntemp = sum;
1336:       }
1337:       MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPI_MAX,mat->comm);
1338:     } else {
1339:       SETERRQ(PETSC_ERR_SUP,"No support for two norm");
1340:     }
1341:   }
1342:   return(0);
1343: }

1347: PetscErrorCode MatTranspose_MPIAIJ(Mat A,Mat *matout)
1348: {
1349:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1350:   Mat_SeqAIJ     *Aloc = (Mat_SeqAIJ*)a->A->data;
1352:   PetscInt       M = A->M,N = A->N,m,*ai,*aj,row,*cols,i,*ct;
1353:   Mat            B;
1354:   PetscScalar    *array;

1357:   if (!matout && M != N) {
1358:     SETERRQ(PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1359:   }

1361:   MatCreate(A->comm,&B);
1362:   MatSetSizes(B,A->n,A->m,N,M);
1363:   MatSetType(B,A->type_name);
1364:   MatMPIAIJSetPreallocation(B,0,PETSC_NULL,0,PETSC_NULL);

1366:   /* copy over the A part */
1367:   Aloc = (Mat_SeqAIJ*)a->A->data;
1368:   m = a->A->m; ai = Aloc->i; aj = Aloc->j; array = Aloc->a;
1369:   row = a->rstart;
1370:   for (i=0; i<ai[m]; i++) {aj[i] += a->cstart ;}
1371:   for (i=0; i<m; i++) {
1372:     MatSetValues(B,ai[i+1]-ai[i],aj,1,&row,array,INSERT_VALUES);
1373:     row++; array += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1374:   }
1375:   aj = Aloc->j;
1376:   for (i=0; i<ai[m]; i++) {aj[i] -= a->cstart ;}

1378:   /* copy over the B part */
1379:   Aloc = (Mat_SeqAIJ*)a->B->data;
1380:   m = a->B->m;  ai = Aloc->i; aj = Aloc->j; array = Aloc->a;
1381:   row  = a->rstart;
1382:   PetscMalloc((1+ai[m])*sizeof(PetscInt),&cols);
1383:   ct   = cols;
1384:   for (i=0; i<ai[m]; i++) {cols[i] = a->garray[aj[i]];}
1385:   for (i=0; i<m; i++) {
1386:     MatSetValues(B,ai[i+1]-ai[i],cols,1,&row,array,INSERT_VALUES);
1387:     row++; array += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1388:   }
1389:   PetscFree(ct);
1390:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1391:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
1392:   if (matout) {
1393:     *matout = B;
1394:   } else {
1395:     MatHeaderCopy(A,B);
1396:   }
1397:   return(0);
1398: }

1402: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
1403: {
1404:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1405:   Mat            a = aij->A,b = aij->B;
1407:   PetscInt       s1,s2,s3;

1410:   MatGetLocalSize(mat,&s2,&s3);
1411:   if (rr) {
1412:     VecGetLocalSize(rr,&s1);
1413:     if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1414:     /* Overlap communication with computation. */
1415:     VecScatterBegin(rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD,aij->Mvctx);
1416:   }
1417:   if (ll) {
1418:     VecGetLocalSize(ll,&s1);
1419:     if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1420:     (*b->ops->diagonalscale)(b,ll,0);
1421:   }
1422:   /* scale  the diagonal block */
1423:   (*a->ops->diagonalscale)(a,ll,rr);

1425:   if (rr) {
1426:     /* Do a scatter end and then right scale the off-diagonal block */
1427:     VecScatterEnd(rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD,aij->Mvctx);
1428:     (*b->ops->diagonalscale)(b,0,aij->lvec);
1429:   }
1430: 
1431:   return(0);
1432: }


1437: PetscErrorCode MatPrintHelp_MPIAIJ(Mat A)
1438: {
1439:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

1443:   if (!a->rank) {
1444:     MatPrintHelp_SeqAIJ(a->A);
1445:   }
1446:   return(0);
1447: }

1451: PetscErrorCode MatSetBlockSize_MPIAIJ(Mat A,PetscInt bs)
1452: {
1453:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

1457:   MatSetBlockSize(a->A,bs);
1458:   MatSetBlockSize(a->B,bs);
1459:   return(0);
1460: }
1463: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1464: {
1465:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

1469:   MatSetUnfactored(a->A);
1470:   return(0);
1471: }

1475: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscTruth *flag)
1476: {
1477:   Mat_MPIAIJ     *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
1478:   Mat            a,b,c,d;
1479:   PetscTruth     flg;

1483:   a = matA->A; b = matA->B;
1484:   c = matB->A; d = matB->B;

1486:   MatEqual(a,c,&flg);
1487:   if (flg) {
1488:     MatEqual(b,d,&flg);
1489:   }
1490:   MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,A->comm);
1491:   return(0);
1492: }

1496: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
1497: {
1499:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)A->data;
1500:   Mat_MPIAIJ     *b = (Mat_MPIAIJ *)B->data;

1503:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1504:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1505:     /* because of the column compression in the off-processor part of the matrix a->B,
1506:        the number of columns in a->B and b->B may be different, hence we cannot call
1507:        the MatCopy() directly on the two parts. If need be, we can provide a more 
1508:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
1509:        then copying the submatrices */
1510:     MatCopy_Basic(A,B,str);
1511:   } else {
1512:     MatCopy(a->A,b->A,str);
1513:     MatCopy(a->B,b->B,str);
1514:   }
1515:   return(0);
1516: }

1520: PetscErrorCode MatSetUpPreallocation_MPIAIJ(Mat A)
1521: {

1525:    MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1526:   return(0);
1527: }

1529:  #include petscblaslapack.h
1532: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1533: {
1535:   PetscInt       i;
1536:   Mat_MPIAIJ     *xx = (Mat_MPIAIJ *)X->data,*yy = (Mat_MPIAIJ *)Y->data;
1537:   PetscBLASInt   bnz,one=1;
1538:   Mat_SeqAIJ     *x,*y;

1541:   if (str == SAME_NONZERO_PATTERN) {
1542:     PetscScalar alpha = a;
1543:     x = (Mat_SeqAIJ *)xx->A->data;
1544:     y = (Mat_SeqAIJ *)yy->A->data;
1545:     bnz = (PetscBLASInt)x->nz;
1546:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1547:     x = (Mat_SeqAIJ *)xx->B->data;
1548:     y = (Mat_SeqAIJ *)yy->B->data;
1549:     bnz = (PetscBLASInt)x->nz;
1550:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1551:   } else if (str == SUBSET_NONZERO_PATTERN) {
1552:     MatAXPY_SeqAIJ(yy->A,a,xx->A,str);

1554:     x = (Mat_SeqAIJ *)xx->B->data;
1555:     y = (Mat_SeqAIJ *)yy->B->data;
1556:     if (y->xtoy && y->XtoY != xx->B) {
1557:       PetscFree(y->xtoy);
1558:       MatDestroy(y->XtoY);
1559:     }
1560:     if (!y->xtoy) { /* get xtoy */
1561:       MatAXPYGetxtoy_Private(xx->B->m,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);
1562:       y->XtoY = xx->B;
1563:     }
1564:     for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]);
1565:   } else {
1566:     MatAXPY_Basic(Y,a,X,str);
1567:   }
1568:   return(0);
1569: }

1571: EXTERN PetscErrorCode PETSCMAT_DLLEXPORT MatConjugate_SeqAIJ(Mat);

1575: PetscErrorCode PETSCMAT_DLLEXPORT MatConjugate_MPIAIJ(Mat mat)
1576: {
1577: #if defined(PETSC_USE_COMPLEX)
1579:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

1582:   MatConjugate_SeqAIJ(aij->A);
1583:   MatConjugate_SeqAIJ(aij->B);
1584: #else
1586: #endif
1587:   return(0);
1588: }

1590: /* -------------------------------------------------------------------*/
1591: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
1592:        MatGetRow_MPIAIJ,
1593:        MatRestoreRow_MPIAIJ,
1594:        MatMult_MPIAIJ,
1595: /* 4*/ MatMultAdd_MPIAIJ,
1596:        MatMultTranspose_MPIAIJ,
1597:        MatMultTransposeAdd_MPIAIJ,
1598:        0,
1599:        0,
1600:        0,
1601: /*10*/ 0,
1602:        0,
1603:        0,
1604:        MatRelax_MPIAIJ,
1605:        MatTranspose_MPIAIJ,
1606: /*15*/ MatGetInfo_MPIAIJ,
1607:        MatEqual_MPIAIJ,
1608:        MatGetDiagonal_MPIAIJ,
1609:        MatDiagonalScale_MPIAIJ,
1610:        MatNorm_MPIAIJ,
1611: /*20*/ MatAssemblyBegin_MPIAIJ,
1612:        MatAssemblyEnd_MPIAIJ,
1613:        0,
1614:        MatSetOption_MPIAIJ,
1615:        MatZeroEntries_MPIAIJ,
1616: /*25*/ MatZeroRows_MPIAIJ,
1617:        0,
1618:        0,
1619:        0,
1620:        0,
1621: /*30*/ MatSetUpPreallocation_MPIAIJ,
1622:        0,
1623:        0,
1624:        0,
1625:        0,
1626: /*35*/ MatDuplicate_MPIAIJ,
1627:        0,
1628:        0,
1629:        0,
1630:        0,
1631: /*40*/ MatAXPY_MPIAIJ,
1632:        MatGetSubMatrices_MPIAIJ,
1633:        MatIncreaseOverlap_MPIAIJ,
1634:        MatGetValues_MPIAIJ,
1635:        MatCopy_MPIAIJ,
1636: /*45*/ MatPrintHelp_MPIAIJ,
1637:        MatScale_MPIAIJ,
1638:        0,
1639:        0,
1640:        0,
1641: /*50*/ MatSetBlockSize_MPIAIJ,
1642:        0,
1643:        0,
1644:        0,
1645:        0,
1646: /*55*/ MatFDColoringCreate_MPIAIJ,
1647:        0,
1648:        MatSetUnfactored_MPIAIJ,
1649:        0,
1650:        0,
1651: /*60*/ MatGetSubMatrix_MPIAIJ,
1652:        MatDestroy_MPIAIJ,
1653:        MatView_MPIAIJ,
1654:        MatGetPetscMaps_Petsc,
1655:        0,
1656: /*65*/ 0,
1657:        0,
1658:        0,
1659:        0,
1660:        0,
1661: /*70*/ 0,
1662:        0,
1663:        MatSetColoring_MPIAIJ,
1664: #if defined(PETSC_HAVE_ADIC)
1665:        MatSetValuesAdic_MPIAIJ,
1666: #else
1667:        0,
1668: #endif
1669:        MatSetValuesAdifor_MPIAIJ,
1670: /*75*/ 0,
1671:        0,
1672:        0,
1673:        0,
1674:        0,
1675: /*80*/ 0,
1676:        0,
1677:        0,
1678:        0,
1679: /*84*/ MatLoad_MPIAIJ,
1680:        0,
1681:        0,
1682:        0,
1683:        0,
1684:        0,
1685: /*90*/ MatMatMult_MPIAIJ_MPIAIJ,
1686:        MatMatMultSymbolic_MPIAIJ_MPIAIJ,
1687:        MatMatMultNumeric_MPIAIJ_MPIAIJ,
1688:        MatPtAP_Basic,
1689:        MatPtAPSymbolic_MPIAIJ,
1690: /*95*/ MatPtAPNumeric_MPIAIJ,
1691:        0,
1692:        0,
1693:        0,
1694:        0,
1695: /*100*/0,
1696:        MatPtAPSymbolic_MPIAIJ_MPIAIJ,
1697:        MatPtAPNumeric_MPIAIJ_MPIAIJ,
1698:        MatConjugate_MPIAIJ
1699: };

1701: /* ----------------------------------------------------------------------------------------*/

1706: PetscErrorCode PETSCMAT_DLLEXPORT MatStoreValues_MPIAIJ(Mat mat)
1707: {
1708:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

1712:   MatStoreValues(aij->A);
1713:   MatStoreValues(aij->B);
1714:   return(0);
1715: }

1721: PetscErrorCode PETSCMAT_DLLEXPORT MatRetrieveValues_MPIAIJ(Mat mat)
1722: {
1723:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

1727:   MatRetrieveValues(aij->A);
1728:   MatRetrieveValues(aij->B);
1729:   return(0);
1730: }

1733:  #include petscpc.h
1737: PetscErrorCode PETSCMAT_DLLEXPORT MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
1738: {
1739:   Mat_MPIAIJ     *b;
1741:   PetscInt       i;

1744:   B->preallocated = PETSC_TRUE;
1745:   if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
1746:   if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
1747:   if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
1748:   if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);
1749:   if (d_nnz) {
1750:     for (i=0; i<B->m; i++) {
1751:       if (d_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than 0: local row %D value %D",i,d_nnz[i]);
1752:     }
1753:   }
1754:   if (o_nnz) {
1755:     for (i=0; i<B->m; i++) {
1756:       if (o_nnz[i] < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than 0: local row %D value %D",i,o_nnz[i]);
1757:     }
1758:   }
1759:   b = (Mat_MPIAIJ*)B->data;
1760:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
1761:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);

1763:   return(0);
1764: }

1769: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
1770: {
1771:   Mat            mat;
1772:   Mat_MPIAIJ     *a,*oldmat = (Mat_MPIAIJ*)matin->data;

1776:   *newmat       = 0;
1777:   MatCreate(matin->comm,&mat);
1778:   MatSetSizes(mat,matin->m,matin->n,matin->M,matin->N);
1779:   MatSetType(mat,matin->type_name);
1780:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
1781:   a    = (Mat_MPIAIJ*)mat->data;
1782: 
1783:   mat->factor       = matin->factor;
1784:   mat->bs           = matin->bs;
1785:   mat->assembled    = PETSC_TRUE;
1786:   mat->insertmode   = NOT_SET_VALUES;
1787:   mat->preallocated = PETSC_TRUE;

1789:   a->rstart       = oldmat->rstart;
1790:   a->rend         = oldmat->rend;
1791:   a->cstart       = oldmat->cstart;
1792:   a->cend         = oldmat->cend;
1793:   a->size         = oldmat->size;
1794:   a->rank         = oldmat->rank;
1795:   a->donotstash   = oldmat->donotstash;
1796:   a->roworiented  = oldmat->roworiented;
1797:   a->rowindices   = 0;
1798:   a->rowvalues    = 0;
1799:   a->getrowactive = PETSC_FALSE;

1801:   PetscMemcpy(a->rowners,oldmat->rowners,2*(a->size+2)*sizeof(PetscInt));
1802:   MatStashCreate_Private(matin->comm,1,&mat->stash);
1803:   if (oldmat->colmap) {
1804: #if defined (PETSC_USE_CTABLE)
1805:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
1806: #else
1807:     PetscMalloc((mat->N)*sizeof(PetscInt),&a->colmap);
1808:     PetscLogObjectMemory(mat,(mat->N)*sizeof(PetscInt));
1809:     PetscMemcpy(a->colmap,oldmat->colmap,(mat->N)*sizeof(PetscInt));
1810: #endif
1811:   } else a->colmap = 0;
1812:   if (oldmat->garray) {
1813:     PetscInt len;
1814:     len  = oldmat->B->n;
1815:     PetscMalloc((len+1)*sizeof(PetscInt),&a->garray);
1816:     PetscLogObjectMemory(mat,len*sizeof(PetscInt));
1817:     if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
1818:   } else a->garray = 0;
1819: 
1820:    VecDuplicate(oldmat->lvec,&a->lvec);
1821:   PetscLogObjectParent(mat,a->lvec);
1822:    VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
1823:   PetscLogObjectParent(mat,a->Mvctx);
1824:    MatDestroy(a->A);
1825:    MatDuplicate(oldmat->A,cpvalues,&a->A);
1826:   PetscLogObjectParent(mat,a->A);
1827:    MatDestroy(a->B);
1828:    MatDuplicate(oldmat->B,cpvalues,&a->B);
1829:   PetscLogObjectParent(mat,a->B);
1830:   PetscFListDuplicate(matin->qlist,&mat->qlist);
1831:   *newmat = mat;
1832:   return(0);
1833: }

1835:  #include petscsys.h

1839: PetscErrorCode MatLoad_MPIAIJ(PetscViewer viewer, MatType type,Mat *newmat)
1840: {
1841:   Mat            A;
1842:   PetscScalar    *vals,*svals;
1843:   MPI_Comm       comm = ((PetscObject)viewer)->comm;
1844:   MPI_Status     status;
1846:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag,maxnz;
1847:   PetscInt       i,nz,j,rstart,rend,mmax;
1848:   PetscInt       header[4],*rowlengths = 0,M,N,m,*cols;
1849:   PetscInt       *ourlens,*procsnz = 0,*offlens,jj,*mycols,*smycols;
1850:   PetscInt       cend,cstart,n,*rowners;
1851:   int            fd;

1854:   MPI_Comm_size(comm,&size);
1855:   MPI_Comm_rank(comm,&rank);
1856:   if (!rank) {
1857:     PetscViewerBinaryGetDescriptor(viewer,&fd);
1858:     PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
1859:     if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
1860:   }

1862:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
1863:   M = header[1]; N = header[2];
1864:   /* determine ownership of all rows */
1865:   m    = M/size + ((M % size) > rank);
1866:   PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
1867:   MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

1869:   /* First process needs enough room for process with most rows */
1870:   if (!rank) {
1871:     mmax       = rowners[1];
1872:     for (i=2; i<size; i++) {
1873:       mmax = PetscMax(mmax,rowners[i]);
1874:     }
1875:   } else mmax = m;

1877:   rowners[0] = 0;
1878:   for (i=2; i<=size; i++) {
1879:     mmax       = PetscMax(mmax,rowners[i]);
1880:     rowners[i] += rowners[i-1];
1881:   }
1882:   rstart = rowners[rank];
1883:   rend   = rowners[rank+1];

1885:   /* distribute row lengths to all processors */
1886:   PetscMalloc2(mmax,PetscInt,&ourlens,mmax,PetscInt,&offlens);
1887:   if (!rank) {
1888:     PetscBinaryRead(fd,ourlens,m,PETSC_INT);
1889:     PetscMalloc(m*sizeof(PetscInt),&rowlengths);
1890:     PetscMalloc(size*sizeof(PetscInt),&procsnz);
1891:     PetscMemzero(procsnz,size*sizeof(PetscInt));
1892:     for (j=0; j<m; j++) {
1893:       procsnz[0] += ourlens[j];
1894:     }
1895:     for (i=1; i<size; i++) {
1896:       PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
1897:       /* calculate the number of nonzeros on each processor */
1898:       for (j=0; j<rowners[i+1]-rowners[i]; j++) {
1899:         procsnz[i] += rowlengths[j];
1900:       }
1901:       MPI_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
1902:     }
1903:     PetscFree(rowlengths);
1904:   } else {
1905:     MPI_Recv(ourlens,m,MPIU_INT,0,tag,comm,&status);
1906:   }

1908:   if (!rank) {
1909:     /* determine max buffer needed and allocate it */
1910:     maxnz = 0;
1911:     for (i=0; i<size; i++) {
1912:       maxnz = PetscMax(maxnz,procsnz[i]);
1913:     }
1914:     PetscMalloc(maxnz*sizeof(PetscInt),&cols);

1916:     /* read in my part of the matrix column indices  */
1917:     nz   = procsnz[0];
1918:     PetscMalloc(nz*sizeof(PetscInt),&mycols);
1919:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);

1921:     /* read in every one elses and ship off */
1922:     for (i=1; i<size; i++) {
1923:       nz   = procsnz[i];
1924:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
1925:       MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
1926:     }
1927:     PetscFree(cols);
1928:   } else {
1929:     /* determine buffer space needed for message */
1930:     nz = 0;
1931:     for (i=0; i<m; i++) {
1932:       nz += ourlens[i];
1933:     }
1934:     PetscMalloc(nz*sizeof(PetscInt),&mycols);

1936:     /* receive message of column indices*/
1937:     MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
1938:     MPI_Get_count(&status,MPIU_INT,&maxnz);
1939:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
1940:   }

1942:   /* determine column ownership if matrix is not square */
1943:   if (N != M) {
1944:     n      = N/size + ((N % size) > rank);
1945:     MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
1946:     cstart = cend - n;
1947:   } else {
1948:     cstart = rstart;
1949:     cend   = rend;
1950:     n      = cend - cstart;
1951:   }

1953:   /* loop over local rows, determining number of off diagonal entries */
1954:   PetscMemzero(offlens,m*sizeof(PetscInt));
1955:   jj = 0;
1956:   for (i=0; i<m; i++) {
1957:     for (j=0; j<ourlens[i]; j++) {
1958:       if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
1959:       jj++;
1960:     }
1961:   }

1963:   /* create our matrix */
1964:   for (i=0; i<m; i++) {
1965:     ourlens[i] -= offlens[i];
1966:   }
1967:   MatCreate(comm,&A);
1968:   MatSetSizes(A,m,n,M,N);
1969:   MatSetType(A,type);
1970:   MatMPIAIJSetPreallocation(A,0,ourlens,0,offlens);

1972:   MatSetOption(A,MAT_COLUMNS_SORTED);
1973:   for (i=0; i<m; i++) {
1974:     ourlens[i] += offlens[i];
1975:   }

1977:   if (!rank) {
1978:     PetscMalloc((maxnz+1)*sizeof(PetscScalar),&vals);

1980:     /* read in my part of the matrix numerical values  */
1981:     nz   = procsnz[0];
1982:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1983: 
1984:     /* insert into matrix */
1985:     jj      = rstart;
1986:     smycols = mycols;
1987:     svals   = vals;
1988:     for (i=0; i<m; i++) {
1989:       MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
1990:       smycols += ourlens[i];
1991:       svals   += ourlens[i];
1992:       jj++;
1993:     }

1995:     /* read in other processors and ship out */
1996:     for (i=1; i<size; i++) {
1997:       nz   = procsnz[i];
1998:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
1999:       MPI_Send(vals,nz,MPIU_SCALAR,i,A->tag,comm);
2000:     }
2001:     PetscFree(procsnz);
2002:   } else {
2003:     /* receive numeric values */
2004:     PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);

2006:     /* receive message of values*/
2007:     MPI_Recv(vals,nz,MPIU_SCALAR,0,A->tag,comm,&status);
2008:     MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2009:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");

2011:     /* insert into matrix */
2012:     jj      = rstart;
2013:     smycols = mycols;
2014:     svals   = vals;
2015:     for (i=0; i<m; i++) {
2016:       MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
2017:       smycols += ourlens[i];
2018:       svals   += ourlens[i];
2019:       jj++;
2020:     }
2021:   }
2022:   PetscFree2(ourlens,offlens);
2023:   PetscFree(vals);
2024:   PetscFree(mycols);
2025:   PetscFree(rowners);

2027:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2028:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2029:   *newmat = A;
2030:   return(0);
2031: }

2035: /*
2036:     Not great since it makes two copies of the submatrix, first an SeqAIJ 
2037:   in local and then by concatenating the local matrices the end result.
2038:   Writing it directly would be much like MatGetSubMatrices_MPIAIJ()
2039: */
2040: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2041: {
2043:   PetscMPIInt    rank,size;
2044:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j;
2045:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
2046:   Mat            *local,M,Mreuse;
2047:   PetscScalar    *vwork,*aa;
2048:   MPI_Comm       comm = mat->comm;
2049:   Mat_SeqAIJ     *aij;


2053:   MPI_Comm_rank(comm,&rank);
2054:   MPI_Comm_size(comm,&size);

2056:   if (call ==  MAT_REUSE_MATRIX) {
2057:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);
2058:     if (!Mreuse) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2059:     local = &Mreuse;
2060:     MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&local);
2061:   } else {
2062:     MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&local);
2063:     Mreuse = *local;
2064:     PetscFree(local);
2065:   }

2067:   /* 
2068:       m - number of local rows
2069:       n - number of columns (same on all processors)
2070:       rstart - first row in new global matrix generated
2071:   */
2072:   MatGetSize(Mreuse,&m,&n);
2073:   if (call == MAT_INITIAL_MATRIX) {
2074:     aij = (Mat_SeqAIJ*)(Mreuse)->data;
2075:     ii  = aij->i;
2076:     jj  = aij->j;

2078:     /*
2079:         Determine the number of non-zeros in the diagonal and off-diagonal 
2080:         portions of the matrix in order to do correct preallocation
2081:     */

2083:     /* first get start and end of "diagonal" columns */
2084:     if (csize == PETSC_DECIDE) {
2085:       ISGetSize(isrow,&mglobal);
2086:       if (mglobal == n) { /* square matrix */
2087:         nlocal = m;
2088:       } else {
2089:         nlocal = n/size + ((n % size) > rank);
2090:       }
2091:     } else {
2092:       nlocal = csize;
2093:     }
2094:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2095:     rstart = rend - nlocal;
2096:     if (rank == size - 1 && rend != n) {
2097:       SETERRQ2(PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
2098:     }

2100:     /* next, compute all the lengths */
2101:     PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
2102:     olens = dlens + m;
2103:     for (i=0; i<m; i++) {
2104:       jend = ii[i+1] - ii[i];
2105:       olen = 0;
2106:       dlen = 0;
2107:       for (j=0; j<jend; j++) {
2108:         if (*jj < rstart || *jj >= rend) olen++;
2109:         else dlen++;
2110:         jj++;
2111:       }
2112:       olens[i] = olen;
2113:       dlens[i] = dlen;
2114:     }
2115:     MatCreate(comm,&M);
2116:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
2117:     MatSetType(M,mat->type_name);
2118:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
2119:     PetscFree(dlens);
2120:   } else {
2121:     PetscInt ml,nl;

2123:     M = *newmat;
2124:     MatGetLocalSize(M,&ml,&nl);
2125:     if (ml != m) SETERRQ(PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2126:     MatZeroEntries(M);
2127:     /*
2128:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2129:        rather than the slower MatSetValues().
2130:     */
2131:     M->was_assembled = PETSC_TRUE;
2132:     M->assembled     = PETSC_FALSE;
2133:   }
2134:   MatGetOwnershipRange(M,&rstart,&rend);
2135:   aij = (Mat_SeqAIJ*)(Mreuse)->data;
2136:   ii  = aij->i;
2137:   jj  = aij->j;
2138:   aa  = aij->a;
2139:   for (i=0; i<m; i++) {
2140:     row   = rstart + i;
2141:     nz    = ii[i+1] - ii[i];
2142:     cwork = jj;     jj += nz;
2143:     vwork = aa;     aa += nz;
2144:     MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2145:   }

2147:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2148:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2149:   *newmat = M;

2151:   /* save submatrix used in processor for next request */
2152:   if (call ==  MAT_INITIAL_MATRIX) {
2153:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2154:     PetscObjectDereference((PetscObject)Mreuse);
2155:   }

2157:   return(0);
2158: }

2163: PetscErrorCode PETSCMAT_DLLEXPORT MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt I[],const PetscInt J[],const PetscScalar v[])
2164: {
2165:   Mat_MPIAIJ     *b = (Mat_MPIAIJ *)B->data;
2166:   PetscInt       m = B->m,cstart = b->cstart, cend = b->cend,j,nnz,i,d;
2167:   PetscInt       *d_nnz,*o_nnz,nnz_max = 0,rstart = b->rstart,ii;
2168:   const PetscInt *JJ;
2169:   PetscScalar    *values;

2173: #if defined(PETSC_OPT_g)
2174:   if (I[0]) SETERRQ1(PETSC_ERR_ARG_RANGE,"I[0] must be 0 it is %D",I[0]);
2175: #endif
2176:   PetscMalloc((2*m+1)*sizeof(PetscInt),&d_nnz);
2177:   o_nnz = d_nnz + m;

2179:   for (i=0; i<m; i++) {
2180:     nnz     = I[i+1]- I[i];
2181:     JJ      = J + I[i];
2182:     nnz_max = PetscMax(nnz_max,nnz);
2183: #if defined(PETSC_OPT_g)
2184:     if (nnz < 0) SETERRQ1(PETSC_ERR_ARG_RANGE,"Local row %D has a negative %D number of columns",i,nnz);
2185: #endif
2186:     for (j=0; j<nnz; j++) {
2187:       if (*JJ >= cstart) break;
2188:       JJ++;
2189:     }
2190:     d = 0;
2191:     for (; j<nnz; j++) {
2192:       if (*JJ++ >= cend) break;
2193:       d++;
2194:     }
2195:     d_nnz[i] = d;
2196:     o_nnz[i] = nnz - d;
2197:   }
2198:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2199:   PetscFree(d_nnz);

2201:   if (v) values = (PetscScalar*)v;
2202:   else {
2203:     PetscMalloc((nnz_max+1)*sizeof(PetscScalar),&values);
2204:     PetscMemzero(values,nnz_max*sizeof(PetscScalar));
2205:   }

2207:   MatSetOption(B,MAT_COLUMNS_SORTED);
2208:   for (i=0; i<m; i++) {
2209:     ii   = i + rstart;
2210:     nnz  = I[i+1]- I[i];
2211:     MatSetValues_MPIAIJ(B,1,&ii,nnz,J+I[i],values+(v ? I[i] : 0),INSERT_VALUES);
2212:   }
2213:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2214:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2215:   MatSetOption(B,MAT_COLUMNS_UNSORTED);

2217:   if (!v) {
2218:     PetscFree(values);
2219:   }
2220:   return(0);
2221: }

2226: /*@C
2227:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
2228:    (the default parallel PETSc format).  

2230:    Collective on MPI_Comm

2232:    Input Parameters:
2233: +  A - the matrix 
2234: .  i - the indices into j for the start of each local row (starts with zero)
2235: .  j - the column indices for each local row (starts with zero) these must be sorted for each row
2236: -  v - optional values in the matrix

2238:    Level: developer

2240: .keywords: matrix, aij, compressed row, sparse, parallel

2242: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateMPIAIJ(), MPIAIJ
2243: @*/
2244: PetscErrorCode PETSCMAT_DLLEXPORT MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2245: {
2246:   PetscErrorCode ierr,(*f)(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]);

2249:   PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",(void (**)(void))&f);
2250:   if (f) {
2251:     (*f)(B,i,j,v);
2252:   }
2253:   return(0);
2254: }

2258: /*@C
2259:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
2260:    (the default parallel PETSc format).  For good matrix assembly performance
2261:    the user should preallocate the matrix storage by setting the parameters 
2262:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
2263:    performance can be increased by more than a factor of 50.

2265:    Collective on MPI_Comm

2267:    Input Parameters:
2268: +  A - the matrix 
2269: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
2270:            (same value is used for all local rows)
2271: .  d_nnz - array containing the number of nonzeros in the various rows of the 
2272:            DIAGONAL portion of the local submatrix (possibly different for each row)
2273:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
2274:            The size of this array is equal to the number of local rows, i.e 'm'. 
2275:            You must leave room for the diagonal entry even if it is zero.
2276: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
2277:            submatrix (same value is used for all local rows).
2278: -  o_nnz - array containing the number of nonzeros in the various rows of the
2279:            OFF-DIAGONAL portion of the local submatrix (possibly different for
2280:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
2281:            structure. The size of this array is equal to the number 
2282:            of local rows, i.e 'm'. 

2284:    If the *_nnz parameter is given then the *_nz parameter is ignored

2286:    The AIJ format (also called the Yale sparse matrix format or
2287:    compressed row storage (CSR)), is fully compatible with standard Fortran 77
2288:    storage.  The stored row and column indices begin with zero.  See the users manual for details.

2290:    The parallel matrix is partitioned such that the first m0 rows belong to 
2291:    process 0, the next m1 rows belong to process 1, the next m2 rows belong 
2292:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

2294:    The DIAGONAL portion of the local submatrix of a processor can be defined 
2295:    as the submatrix which is obtained by extraction the part corresponding 
2296:    to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the 
2297:    first row that belongs to the processor, and r2 is the last row belonging 
2298:    to the this processor. This is a square mxm matrix. The remaining portion 
2299:    of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.

2301:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

2303:    Example usage:
2304:   
2305:    Consider the following 8x8 matrix with 34 non-zero values, that is 
2306:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
2307:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
2308:    as follows:

2310: .vb
2311:             1  2  0  |  0  3  0  |  0  4
2312:     Proc0   0  5  6  |  7  0  0  |  8  0
2313:             9  0 10  | 11  0  0  | 12  0
2314:     -------------------------------------
2315:            13  0 14  | 15 16 17  |  0  0
2316:     Proc1   0 18  0  | 19 20 21  |  0  0 
2317:             0  0  0  | 22 23  0  | 24  0
2318:     -------------------------------------
2319:     Proc2  25 26 27  |  0  0 28  | 29  0
2320:            30  0  0  | 31 32 33  |  0 34
2321: .ve

2323:    This can be represented as a collection of submatrices as:

2325: .vb
2326:       A B C
2327:       D E F
2328:       G H I
2329: .ve

2331:    Where the submatrices A,B,C are owned by proc0, D,E,F are
2332:    owned by proc1, G,H,I are owned by proc2.

2334:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2335:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2336:    The 'M','N' parameters are 8,8, and have the same values on all procs.

2338:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
2339:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
2340:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
2341:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
2342:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
2343:    matrix, ans [DF] as another SeqAIJ matrix.

2345:    When d_nz, o_nz parameters are specified, d_nz storage elements are
2346:    allocated for every row of the local diagonal submatrix, and o_nz
2347:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
2348:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
2349:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
2350:    In this case, the values of d_nz,o_nz are:
2351: .vb
2352:      proc0 : dnz = 2, o_nz = 2
2353:      proc1 : dnz = 3, o_nz = 2
2354:      proc2 : dnz = 1, o_nz = 4
2355: .ve
2356:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
2357:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
2358:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
2359:    34 values.

2361:    When d_nnz, o_nnz parameters are specified, the storage is specified
2362:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
2363:    In the above case the values for d_nnz,o_nnz are:
2364: .vb
2365:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
2366:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
2367:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
2368: .ve
2369:    Here the space allocated is sum of all the above values i.e 34, and
2370:    hence pre-allocation is perfect.

2372:    Level: intermediate

2374: .keywords: matrix, aij, compressed row, sparse, parallel

2376: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateMPIAIJ(), MatMPIAIJSetPreallocationCSR(),
2377:           MPIAIJ
2378: @*/
2379: PetscErrorCode PETSCMAT_DLLEXPORT MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2380: {
2381:   PetscErrorCode ierr,(*f)(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);

2384:   PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",(void (**)(void))&f);
2385:   if (f) {
2386:     (*f)(B,d_nz,d_nnz,o_nz,o_nnz);
2387:   }
2388:   return(0);
2389: }

2393: /*@C
2394:    MatCreateMPIAIJ - Creates a sparse parallel matrix in AIJ format
2395:    (the default parallel PETSc format).  For good matrix assembly performance
2396:    the user should preallocate the matrix storage by setting the parameters 
2397:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
2398:    performance can be increased by more than a factor of 50.

2400:    Collective on MPI_Comm

2402:    Input Parameters:
2403: +  comm - MPI communicator
2404: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
2405:            This value should be the same as the local size used in creating the 
2406:            y vector for the matrix-vector product y = Ax.
2407: .  n - This value should be the same as the local size used in creating the 
2408:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
2409:        calculated if N is given) For square matrices n is almost always m.
2410: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
2411: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
2412: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
2413:            (same value is used for all local rows)
2414: .  d_nnz - array containing the number of nonzeros in the various rows of the 
2415:            DIAGONAL portion of the local submatrix (possibly different for each row)
2416:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
2417:            The size of this array is equal to the number of local rows, i.e 'm'. 
2418:            You must leave room for the diagonal entry even if it is zero.
2419: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
2420:            submatrix (same value is used for all local rows).
2421: -  o_nnz - array containing the number of nonzeros in the various rows of the
2422:            OFF-DIAGONAL portion of the local submatrix (possibly different for
2423:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
2424:            structure. The size of this array is equal to the number 
2425:            of local rows, i.e 'm'. 

2427:    Output Parameter:
2428: .  A - the matrix 

2430:    Notes:
2431:    If the *_nnz parameter is given then the *_nz parameter is ignored

2433:    m,n,M,N parameters specify the size of the matrix, and its partitioning across
2434:    processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
2435:    storage requirements for this matrix.

2437:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one 
2438:    processor than it must be used on all processors that share the object for 
2439:    that argument.

2441:    The user MUST specify either the local or global matrix dimensions
2442:    (possibly both).

2444:    The parallel matrix is partitioned such that the first m0 rows belong to 
2445:    process 0, the next m1 rows belong to process 1, the next m2 rows belong 
2446:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

2448:    The DIAGONAL portion of the local submatrix of a processor can be defined 
2449:    as the submatrix which is obtained by extraction the part corresponding 
2450:    to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the 
2451:    first row that belongs to the processor, and r2 is the last row belonging 
2452:    to the this processor. This is a square mxm matrix. The remaining portion 
2453:    of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.

2455:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

2457:    When calling this routine with a single process communicator, a matrix of
2458:    type SEQAIJ is returned.  If a matrix of type MPIAIJ is desired for this
2459:    type of communicator, use the construction mechanism:
2460:      MatCreate(...,&A); MatSetType(A,MPIAIJ); MatMPIAIJSetPreallocation(A,...);

2462:    By default, this format uses inodes (identical nodes) when possible.
2463:    We search for consecutive rows with the same nonzero structure, thereby
2464:    reusing matrix information to achieve increased efficiency.

2466:    Options Database Keys:
2467: +  -mat_no_inode  - Do not use inodes
2468: .  -mat_inode_limit <limit> - Sets inode limit (max limit=5)
2469: -  -mat_aij_oneindex - Internally use indexing starting at 1
2470:         rather than 0.  Note that when calling MatSetValues(),
2471:         the user still MUST index entries starting at 0!


2474:    Example usage:
2475:   
2476:    Consider the following 8x8 matrix with 34 non-zero values, that is 
2477:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
2478:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
2479:    as follows:

2481: .vb
2482:             1  2  0  |  0  3  0  |  0  4
2483:     Proc0   0  5  6  |  7  0  0  |  8  0
2484:             9  0 10  | 11  0  0  | 12  0
2485:     -------------------------------------
2486:            13  0 14  | 15 16 17  |  0  0
2487:     Proc1   0 18  0  | 19 20 21  |  0  0 
2488:             0  0  0  | 22 23  0  | 24  0
2489:     -------------------------------------
2490:     Proc2  25 26 27  |  0  0 28  | 29  0
2491:            30  0  0  | 31 32 33  |  0 34
2492: .ve

2494:    This can be represented as a collection of submatrices as:

2496: .vb
2497:       A B C
2498:       D E F
2499:       G H I
2500: .ve

2502:    Where the submatrices A,B,C are owned by proc0, D,E,F are
2503:    owned by proc1, G,H,I are owned by proc2.

2505:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2506:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
2507:    The 'M','N' parameters are 8,8, and have the same values on all procs.

2509:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
2510:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
2511:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
2512:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
2513:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
2514:    matrix, ans [DF] as another SeqAIJ matrix.

2516:    When d_nz, o_nz parameters are specified, d_nz storage elements are
2517:    allocated for every row of the local diagonal submatrix, and o_nz
2518:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
2519:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
2520:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
2521:    In this case, the values of d_nz,o_nz are:
2522: .vb
2523:      proc0 : dnz = 2, o_nz = 2
2524:      proc1 : dnz = 3, o_nz = 2
2525:      proc2 : dnz = 1, o_nz = 4
2526: .ve
2527:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
2528:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
2529:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
2530:    34 values.

2532:    When d_nnz, o_nnz parameters are specified, the storage is specified
2533:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
2534:    In the above case the values for d_nnz,o_nnz are:
2535: .vb
2536:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
2537:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
2538:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
2539: .ve
2540:    Here the space allocated is sum of all the above values i.e 34, and
2541:    hence pre-allocation is perfect.

2543:    Level: intermediate

2545: .keywords: matrix, aij, compressed row, sparse, parallel

2547: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
2548:           MPIAIJ
2549: @*/
2550: PetscErrorCode PETSCMAT_DLLEXPORT MatCreateMPIAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
2551: {
2553:   PetscMPIInt    size;

2556:   MatCreate(comm,A);
2557:   MatSetSizes(*A,m,n,M,N);
2558:   MPI_Comm_size(comm,&size);
2559:   if (size > 1) {
2560:     MatSetType(*A,MATMPIAIJ);
2561:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
2562:   } else {
2563:     MatSetType(*A,MATSEQAIJ);
2564:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
2565:   }
2566:   return(0);
2567: }

2571: PetscErrorCode PETSCMAT_DLLEXPORT MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[])
2572: {
2573:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

2576:   *Ad     = a->A;
2577:   *Ao     = a->B;
2578:   *colmap = a->garray;
2579:   return(0);
2580: }

2584: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
2585: {
2587:   PetscInt       i;
2588:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2591:   if (coloring->ctype == IS_COLORING_LOCAL) {
2592:     ISColoringValue *allcolors,*colors;
2593:     ISColoring      ocoloring;

2595:     /* set coloring for diagonal portion */
2596:     MatSetColoring_SeqAIJ(a->A,coloring);

2598:     /* set coloring for off-diagonal portion */
2599:     ISAllGatherColors(A->comm,coloring->n,coloring->colors,PETSC_NULL,&allcolors);
2600:     PetscMalloc((a->B->n+1)*sizeof(ISColoringValue),&colors);
2601:     for (i=0; i<a->B->n; i++) {
2602:       colors[i] = allcolors[a->garray[i]];
2603:     }
2604:     PetscFree(allcolors);
2605:     ISColoringCreate(MPI_COMM_SELF,a->B->n,colors,&ocoloring);
2606:     MatSetColoring_SeqAIJ(a->B,ocoloring);
2607:     ISColoringDestroy(ocoloring);
2608:   } else if (coloring->ctype == IS_COLORING_GHOSTED) {
2609:     ISColoringValue *colors;
2610:     PetscInt             *larray;
2611:     ISColoring      ocoloring;

2613:     /* set coloring for diagonal portion */
2614:     PetscMalloc((a->A->n+1)*sizeof(PetscInt),&larray);
2615:     for (i=0; i<a->A->n; i++) {
2616:       larray[i] = i + a->cstart;
2617:     }
2618:     ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->A->n,larray,PETSC_NULL,larray);
2619:     PetscMalloc((a->A->n+1)*sizeof(ISColoringValue),&colors);
2620:     for (i=0; i<a->A->n; i++) {
2621:       colors[i] = coloring->colors[larray[i]];
2622:     }
2623:     PetscFree(larray);
2624:     ISColoringCreate(PETSC_COMM_SELF,a->A->n,colors,&ocoloring);
2625:     MatSetColoring_SeqAIJ(a->A,ocoloring);
2626:     ISColoringDestroy(ocoloring);

2628:     /* set coloring for off-diagonal portion */
2629:     PetscMalloc((a->B->n+1)*sizeof(PetscInt),&larray);
2630:     ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->B->n,a->garray,PETSC_NULL,larray);
2631:     PetscMalloc((a->B->n+1)*sizeof(ISColoringValue),&colors);
2632:     for (i=0; i<a->B->n; i++) {
2633:       colors[i] = coloring->colors[larray[i]];
2634:     }
2635:     PetscFree(larray);
2636:     ISColoringCreate(MPI_COMM_SELF,a->B->n,colors,&ocoloring);
2637:     MatSetColoring_SeqAIJ(a->B,ocoloring);
2638:     ISColoringDestroy(ocoloring);
2639:   } else {
2640:     SETERRQ1(PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);
2641:   }

2643:   return(0);
2644: }

2646: #if defined(PETSC_HAVE_ADIC)
2649: PetscErrorCode MatSetValuesAdic_MPIAIJ(Mat A,void *advalues)
2650: {
2651:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2655:   MatSetValuesAdic_SeqAIJ(a->A,advalues);
2656:   MatSetValuesAdic_SeqAIJ(a->B,advalues);
2657:   return(0);
2658: }
2659: #endif

2663: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
2664: {
2665:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2669:   MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
2670:   MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
2671:   return(0);
2672: }

2676: /*@C
2677:       MatMerge - Creates a single large PETSc matrix by concatinating sequential
2678:                  matrices from each processor

2680:     Collective on MPI_Comm

2682:    Input Parameters:
2683: +    comm - the communicators the parallel matrix will live on
2684: .    inmat - the input sequential matrices
2685: .    n - number of local columns (or PETSC_DECIDE)
2686: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

2688:    Output Parameter:
2689: .    outmat - the parallel matrix generated

2691:     Level: advanced

2693:    Notes: The number of columns of the matrix in EACH processor MUST be the same.

2695: @*/
2696: PetscErrorCode PETSCMAT_DLLEXPORT MatMerge(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
2697: {
2699:   PetscInt       m,N,i,rstart,nnz,I,*dnz,*onz;
2700:   PetscInt       *indx;
2701:   PetscScalar    *values;
2702:   PetscMap       columnmap,rowmap;

2705:     MatGetSize(inmat,&m,&N);
2706:   /*
2707:   PetscMPIInt       rank;
2708:   MPI_Comm_rank(comm,&rank);
2709:   PetscPrintf(PETSC_COMM_SELF," [%d] inmat m=%d, n=%d, N=%d\n",rank,m,n,N);
2710:   */
2711:   if (scall == MAT_INITIAL_MATRIX){
2712:     /* count nonzeros in each row, for diagonal and off diagonal portion of matrix */
2713:     if (n == PETSC_DECIDE){
2714:       PetscMapCreate(comm,&columnmap);
2715:       PetscMapSetSize(columnmap,N);
2716:       PetscMapSetType(columnmap,MAP_MPI);
2717:       PetscMapGetLocalSize(columnmap,&n);
2718:       PetscMapDestroy(columnmap);
2719:     }

2721:     PetscMapCreate(comm,&rowmap);
2722:     PetscMapSetLocalSize(rowmap,m);
2723:     PetscMapSetType(rowmap,MAP_MPI);
2724:     PetscMapGetLocalRange(rowmap,&rstart,0);
2725:     PetscMapDestroy(rowmap);

2727:     MatPreallocateInitialize(comm,m,n,dnz,onz);
2728:     for (i=0;i<m;i++) {
2729:       MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
2730:       MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
2731:       MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
2732:     }
2733:     /* This routine will ONLY return MPIAIJ type matrix */
2734:     MatCreate(comm,outmat);
2735:     MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
2736:     MatSetType(*outmat,MATMPIAIJ);
2737:     MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
2738:     MatPreallocateFinalize(dnz,onz);
2739: 
2740:   } else if (scall == MAT_REUSE_MATRIX){
2741:     MatGetOwnershipRange(*outmat,&rstart,PETSC_NULL);
2742:   } else {
2743:     SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
2744:   }

2746:   for (i=0;i<m;i++) {
2747:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
2748:     I    = i + rstart;
2749:     MatSetValues(*outmat,1,&I,nnz,indx,values,INSERT_VALUES);
2750:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
2751:   }
2752:   MatDestroy(inmat);
2753:   MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
2754:   MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);

2756:   return(0);
2757: }

2761: PetscErrorCode MatFileSplit(Mat A,char *outfile)
2762: {
2763:   PetscErrorCode    ierr;
2764:   PetscMPIInt       rank;
2765:   PetscInt          m,N,i,rstart,nnz;
2766:   size_t            len;
2767:   const PetscInt    *indx;
2768:   PetscViewer       out;
2769:   char              *name;
2770:   Mat               B;
2771:   const PetscScalar *values;

2774:   MatGetLocalSize(A,&m,0);
2775:   MatGetSize(A,0,&N);
2776:   /* Should this be the type of the diagonal block of A? */
2777:   MatCreate(PETSC_COMM_SELF,&B);
2778:   MatSetSizes(B,m,N,m,N);
2779:   MatSetType(B,MATSEQAIJ);
2780:   MatSeqAIJSetPreallocation(B,0,PETSC_NULL);
2781:   MatGetOwnershipRange(A,&rstart,0);
2782:   for (i=0;i<m;i++) {
2783:     MatGetRow(A,i+rstart,&nnz,&indx,&values);
2784:     MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
2785:     MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
2786:   }
2787:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2788:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

2790:   MPI_Comm_rank(A->comm,&rank);
2791:   PetscStrlen(outfile,&len);
2792:   PetscMalloc((len+5)*sizeof(char),&name);
2793:   sprintf(name,"%s.%d",outfile,rank);
2794:   PetscViewerBinaryOpen(PETSC_COMM_SELF,name,PETSC_FILE_CREATE,&out);
2795:   PetscFree(name);
2796:   MatView(B,out);
2797:   PetscViewerDestroy(out);
2798:   MatDestroy(B);
2799:   return(0);
2800: }

2802: EXTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
2805: PetscErrorCode PETSCMAT_DLLEXPORT MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
2806: {
2807:   PetscErrorCode       ierr;
2808:   Mat_Merge_SeqsToMPI  *merge;
2809:   PetscObjectContainer container;

2812:   PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);
2813:   if (container) {
2814:     PetscObjectContainerGetPointer(container,(void **)&merge);
2815:     PetscFree(merge->id_r);
2816:     PetscFree(merge->len_s);
2817:     PetscFree(merge->len_r);
2818:     PetscFree(merge->bi);
2819:     PetscFree(merge->bj);
2820:     PetscFree(merge->buf_ri);
2821:     PetscFree(merge->buf_rj);
2822:     PetscMapDestroy(merge->rowmap);
2823:     if (merge->coi){PetscFree(merge->coi);}
2824:     if (merge->coj){PetscFree(merge->coj);}
2825:     if (merge->owners_co){PetscFree(merge->owners_co);}
2826: 
2827:     PetscObjectContainerDestroy(container);
2828:     PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
2829:   }
2830:   PetscFree(merge);

2832:   MatDestroy_MPIAIJ(A);
2833:   return(0);
2834: }

2836:  #include src/mat/utils/freespace.h
2837:  #include petscbt.h
2838: static PetscEvent logkey_seqstompinum = 0;
2841: /*@C
2842:       MatMerge_SeqsToMPI - Creates a MPIAIJ matrix by adding sequential
2843:                  matrices from each processor

2845:     Collective on MPI_Comm

2847:    Input Parameters:
2848: +    comm - the communicators the parallel matrix will live on
2849: .    seqmat - the input sequential matrices
2850: .    m - number of local rows (or PETSC_DECIDE)
2851: .    n - number of local columns (or PETSC_DECIDE)
2852: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

2854:    Output Parameter:
2855: .    mpimat - the parallel matrix generated

2857:     Level: advanced

2859:    Notes: 
2860:      The dimensions of the sequential matrix in each processor MUST be the same.
2861:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
2862:      destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
2863: @*/
2864: PetscErrorCode PETSCMAT_DLLEXPORT MatMerge_SeqsToMPINumeric(Mat seqmat,Mat mpimat)
2865: {
2866:   PetscErrorCode       ierr;
2867:   MPI_Comm             comm=mpimat->comm;
2868:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
2869:   PetscMPIInt          size,rank,taga,*len_s;
2870:   PetscInt             N=mpimat->N,i,j,*owners,*ai=a->i,*aj=a->j;
2871:   PetscInt             proc,m;
2872:   PetscInt             **buf_ri,**buf_rj;
2873:   PetscInt             k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
2874:   PetscInt             nrows,**buf_ri_k,**nextrow,**nextai;
2875:   MPI_Request          *s_waits,*r_waits;
2876:   MPI_Status           *status;
2877:   MatScalar            *aa=a->a,**abuf_r,*ba_i;
2878:   Mat_Merge_SeqsToMPI  *merge;
2879:   PetscObjectContainer container;
2880: 
2882:   if (!logkey_seqstompinum) {
2883:     PetscLogEventRegister(&logkey_seqstompinum,"MatMerge_SeqsToMPINumeric",MAT_COOKIE);
2884:   }
2885:   PetscLogEventBegin(logkey_seqstompinum,seqmat,0,0,0);

2887:   MPI_Comm_size(comm,&size);
2888:   MPI_Comm_rank(comm,&rank);

2890:   PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject *)&container);
2891:   if (container) {
2892:     PetscObjectContainerGetPointer(container,(void **)&merge);
2893:   }
2894:   bi     = merge->bi;
2895:   bj     = merge->bj;
2896:   buf_ri = merge->buf_ri;
2897:   buf_rj = merge->buf_rj;

2899:   PetscMalloc(size*sizeof(MPI_Status),&status);
2900:   PetscMapGetGlobalRange(merge->rowmap,&owners);
2901:   len_s  = merge->len_s;

2903:   /* send and recv matrix values */
2904:   /*-----------------------------*/
2905:   PetscObjectGetNewTag((PetscObject)merge->rowmap,&taga);
2906:   PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);

2908:   PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);
2909:   for (proc=0,k=0; proc<size; proc++){
2910:     if (!len_s[proc]) continue;
2911:     i = owners[proc];
2912:     MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
2913:     k++;
2914:   }

2916:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
2917:   if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
2918:   PetscFree(status);

2920:   PetscFree(s_waits);
2921:   PetscFree(r_waits);

2923:   /* insert mat values of mpimat */
2924:   /*----------------------------*/
2925:   PetscMalloc(N*sizeof(MatScalar),&ba_i);
2926:   PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
2927:   nextrow = buf_ri_k + merge->nrecv;
2928:   nextai  = nextrow + merge->nrecv;

2930:   for (k=0; k<merge->nrecv; k++){
2931:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
2932:     nrows = *(buf_ri_k[k]);
2933:     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
2934:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
2935:   }

2937:   /* set values of ba */
2938:   PetscMapGetLocalSize(merge->rowmap,&m);
2939:   for (i=0; i<m; i++) {
2940:     arow = owners[rank] + i;
2941:     bj_i = bj+bi[i];  /* col indices of the i-th row of mpimat */
2942:     bnzi = bi[i+1] - bi[i];
2943:     PetscMemzero(ba_i,bnzi*sizeof(MatScalar));

2945:     /* add local non-zero vals of this proc's seqmat into ba */
2946:     anzi = ai[arow+1] - ai[arow];
2947:     aj   = a->j + ai[arow];
2948:     aa   = a->a + ai[arow];
2949:     nextaj = 0;
2950:     for (j=0; nextaj<anzi; j++){
2951:       if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
2952:         ba_i[j] += aa[nextaj++];
2953:       }
2954:     }

2956:     /* add received vals into ba */
2957:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
2958:       /* i-th row */
2959:       if (i == *nextrow[k]) {
2960:         anzi = *(nextai[k]+1) - *nextai[k];
2961:         aj   = buf_rj[k] + *(nextai[k]);
2962:         aa   = abuf_r[k] + *(nextai[k]);
2963:         nextaj = 0;
2964:         for (j=0; nextaj<anzi; j++){
2965:           if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
2966:             ba_i[j] += aa[nextaj++];
2967:           }
2968:         }
2969:         nextrow[k]++; nextai[k]++;
2970:       }
2971:     }
2972:     MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
2973:   }
2974:   MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
2975:   MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);

2977:   PetscFree(abuf_r);
2978:   PetscFree(ba_i);
2979:   PetscFree(buf_ri_k);
2980:   PetscLogEventEnd(logkey_seqstompinum,seqmat,0,0,0);
2981:   return(0);
2982: }

2984: static PetscEvent logkey_seqstompisym = 0;
2987: PetscErrorCode PETSCMAT_DLLEXPORT MatMerge_SeqsToMPISymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
2988: {
2989:   PetscErrorCode       ierr;
2990:   Mat                  B_mpi;
2991:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
2992:   PetscMPIInt          size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
2993:   PetscInt             **buf_rj,**buf_ri,**buf_ri_k;
2994:   PetscInt             M=seqmat->m,N=seqmat->n,i,*owners,*ai=a->i,*aj=a->j;
2995:   PetscInt             len,proc,*dnz,*onz;
2996:   PetscInt             k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
2997:   PetscInt             nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
2998:   MPI_Request          *si_waits,*sj_waits,*ri_waits,*rj_waits;
2999:   MPI_Status           *status;
3000:   FreeSpaceList        free_space=PETSC_NULL,current_space=PETSC_NULL;
3001:   PetscBT              lnkbt;
3002:   Mat_Merge_SeqsToMPI  *merge;
3003:   PetscObjectContainer container;

3006:   if (!logkey_seqstompisym) {
3007:     PetscLogEventRegister(&logkey_seqstompisym,"MatMerge_SeqsToMPISymbolic",MAT_COOKIE);
3008:   }
3009:   PetscLogEventBegin(logkey_seqstompisym,seqmat,0,0,0);

3011:   /* make sure it is a PETSc comm */
3012:   PetscCommDuplicate(comm,&comm,PETSC_NULL);
3013:   MPI_Comm_size(comm,&size);
3014:   MPI_Comm_rank(comm,&rank);
3015: 
3016:   PetscNew(Mat_Merge_SeqsToMPI,&merge);
3017:   PetscMalloc(size*sizeof(MPI_Status),&status);

3019:   /* determine row ownership */
3020:   /*---------------------------------------------------------*/
3021:   PetscMapCreate(comm,&merge->rowmap);
3022:   if (m == PETSC_DECIDE) {
3023:     PetscMapSetSize(merge->rowmap,M);
3024:   } else {
3025:     PetscMapSetLocalSize(merge->rowmap,m);
3026:   }
3027:   PetscMapSetType(merge->rowmap,MAP_MPI);
3028:   PetscMalloc(size*sizeof(PetscMPIInt),&len_si);
3029:   PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);
3030: 
3031:   if (m == PETSC_DECIDE) {PetscMapGetLocalSize(merge->rowmap,&m); }
3032:   PetscMapGetGlobalRange(merge->rowmap,&owners);

3034:   /* determine the number of messages to send, their lengths */
3035:   /*---------------------------------------------------------*/
3036:   len_s  = merge->len_s;

3038:   len = 0;  /* length of buf_si[] */
3039:   merge->nsend = 0;
3040:   for (proc=0; proc<size; proc++){
3041:     len_si[proc] = 0;
3042:     if (proc == rank){
3043:       len_s[proc] = 0;
3044:     } else {
3045:       len_si[proc] = owners[proc+1] - owners[proc] + 1;
3046:       len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
3047:     }
3048:     if (len_s[proc]) {
3049:       merge->nsend++;
3050:       nrows = 0;
3051:       for (i=owners[proc]; i<owners[proc+1]; i++){
3052:         if (ai[i+1] > ai[i]) nrows++;
3053:       }
3054:       len_si[proc] = 2*(nrows+1);
3055:       len += len_si[proc];
3056:     }
3057:   }

3059:   /* determine the number and length of messages to receive for ij-structure */
3060:   /*-------------------------------------------------------------------------*/
3061:   PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);
3062:   PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);

3064:   /* post the Irecv of j-structure */
3065:   /*-------------------------------*/
3066:   PetscObjectGetNewTag((PetscObject)merge->rowmap,&tagj);
3067:   PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);

3069:   /* post the Isend of j-structure */
3070:   /*--------------------------------*/
3071:   PetscMalloc((2*merge->nsend+1)*sizeof(MPI_Request),&si_waits);
3072:   sj_waits = si_waits + merge->nsend;

3074:   for (proc=0, k=0; proc<size; proc++){
3075:     if (!len_s[proc]) continue;
3076:     i = owners[proc];
3077:     MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
3078:     k++;
3079:   }

3081:   /* receives and sends of j-structure are complete */
3082:   /*------------------------------------------------*/
3083:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
3084:   if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
3085: 
3086:   /* send and recv i-structure */
3087:   /*---------------------------*/
3088:   PetscObjectGetNewTag((PetscObject)merge->rowmap,&tagi);
3089:   PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
3090: 
3091:   PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);
3092:   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
3093:   for (proc=0,k=0; proc<size; proc++){
3094:     if (!len_s[proc]) continue;
3095:     /* form outgoing message for i-structure: 
3096:          buf_si[0]:                 nrows to be sent
3097:                [1:nrows]:           row index (global)
3098:                [nrows+1:2*nrows+1]: i-structure index
3099:     */
3100:     /*-------------------------------------------*/
3101:     nrows = len_si[proc]/2 - 1;
3102:     buf_si_i    = buf_si + nrows+1;
3103:     buf_si[0]   = nrows;
3104:     buf_si_i[0] = 0;
3105:     nrows = 0;
3106:     for (i=owners[proc]; i<owners[proc+1]; i++){
3107:       anzi = ai[i+1] - ai[i];
3108:       if (anzi) {
3109:         buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
3110:         buf_si[nrows+1] = i-owners[proc]; /* local row index */
3111:         nrows++;
3112:       }
3113:     }
3114:     MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
3115:     k++;
3116:     buf_si += len_si[proc];
3117:   }

3119:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
3120:   if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}

3122:   PetscLogInfo(((PetscObject)(seqmat),"MatMerge_SeqsToMPI: nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv));
3123:   for (i=0; i<merge->nrecv; i++){
3124:     PetscLogInfo(((PetscObject)(seqmat),"MatMerge_SeqsToMPI:   recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]));
3125:   }

3127:   PetscFree(len_si);
3128:   PetscFree(len_ri);
3129:   PetscFree(rj_waits);
3130:   PetscFree(si_waits);
3131:   PetscFree(ri_waits);
3132:   PetscFree(buf_s);
3133:   PetscFree(status);

3135:   /* compute a local seq matrix in each processor */
3136:   /*----------------------------------------------*/
3137:   /* allocate bi array and free space for accumulating nonzero column info */
3138:   PetscMalloc((m+1)*sizeof(PetscInt),&bi);
3139:   bi[0] = 0;

3141:   /* create and initialize a linked list */
3142:   nlnk = N+1;
3143:   PetscLLCreate(N,N,nlnk,lnk,lnkbt);
3144: 
3145:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
3146:   len = 0;
3147:   len  = ai[owners[rank+1]] - ai[owners[rank]];
3148:   GetMoreSpace((PetscInt)(2*len+1),&free_space);
3149:   current_space = free_space;

3151:   /* determine symbolic info for each local row */
3152:   PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
3153:   nextrow = buf_ri_k + merge->nrecv;
3154:   nextai  = nextrow + merge->nrecv;
3155:   for (k=0; k<merge->nrecv; k++){
3156:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
3157:     nrows = *buf_ri_k[k];
3158:     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
3159:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
3160:   }

3162:   MatPreallocateInitialize(comm,m,n,dnz,onz);
3163:   len = 0;
3164:   for (i=0;i<m;i++) {
3165:     bnzi   = 0;
3166:     /* add local non-zero cols of this proc's seqmat into lnk */
3167:     arow   = owners[rank] + i;
3168:     anzi   = ai[arow+1] - ai[arow];
3169:     aj     = a->j + ai[arow];
3170:     PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3171:     bnzi += nlnk;
3172:     /* add received col data into lnk */
3173:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
3174:       if (i == *nextrow[k]) { /* i-th row */
3175:         anzi = *(nextai[k]+1) - *nextai[k];
3176:         aj   = buf_rj[k] + *nextai[k];
3177:         PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3178:         bnzi += nlnk;
3179:         nextrow[k]++; nextai[k]++;
3180:       }
3181:     }
3182:     if (len < bnzi) len = bnzi;  /* =max(bnzi) */

3184:     /* if free space is not available, make more free space */
3185:     if (current_space->local_remaining<bnzi) {
3186:       GetMoreSpace(current_space->total_array_size,&current_space);
3187:       nspacedouble++;
3188:     }
3189:     /* copy data into free space, then initialize lnk */
3190:     PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
3191:     MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);

3193:     current_space->array           += bnzi;
3194:     current_space->local_used      += bnzi;
3195:     current_space->local_remaining -= bnzi;
3196: 
3197:     bi[i+1] = bi[i] + bnzi;
3198:   }
3199: 
3200:   PetscFree(buf_ri_k);

3202:   PetscMalloc((bi[m]+1)*sizeof(PetscInt),&bj);
3203:   MakeSpaceContiguous(&free_space,bj);
3204:   PetscLLDestroy(lnk,lnkbt);

3206:   /* create symbolic parallel matrix B_mpi */
3207:   /*---------------------------------------*/
3208:   MatCreate(comm,&B_mpi);
3209:   if (n==PETSC_DECIDE) {
3210:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
3211:   } else {
3212:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3213:   }
3214:   MatSetType(B_mpi,MATMPIAIJ);
3215:   MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
3216:   MatPreallocateFinalize(dnz,onz);

3218:   /* B_mpi is not ready for use - assembly will be done by MatMerge_SeqsToMPINumeric() */
3219:   B_mpi->assembled     = PETSC_FALSE;
3220:   B_mpi->ops->destroy  = MatDestroy_MPIAIJ_SeqsToMPI;
3221:   merge->bi            = bi;
3222:   merge->bj            = bj;
3223:   merge->buf_ri        = buf_ri;
3224:   merge->buf_rj        = buf_rj;
3225:   merge->coi           = PETSC_NULL;
3226:   merge->coj           = PETSC_NULL;
3227:   merge->owners_co     = PETSC_NULL;

3229:   /* attach the supporting struct to B_mpi for reuse */
3230:   PetscObjectContainerCreate(PETSC_COMM_SELF,&container);
3231:   PetscObjectContainerSetPointer(container,merge);
3232:   PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
3233:   *mpimat = B_mpi;

3235:   PetscCommDestroy(&comm);
3236:   PetscLogEventEnd(logkey_seqstompisym,seqmat,0,0,0);
3237:   return(0);
3238: }

3240: static PetscEvent logkey_seqstompi = 0;
3243: PetscErrorCode PETSCMAT_DLLEXPORT MatMerge_SeqsToMPI(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
3244: {
3245:   PetscErrorCode   ierr;

3248:   if (!logkey_seqstompi) {
3249:     PetscLogEventRegister(&logkey_seqstompi,"MatMerge_SeqsToMPI",MAT_COOKIE);
3250:   }
3251:   PetscLogEventBegin(logkey_seqstompi,seqmat,0,0,0);
3252:   if (scall == MAT_INITIAL_MATRIX){
3253:     MatMerge_SeqsToMPISymbolic(comm,seqmat,m,n,mpimat);
3254:   }
3255:   MatMerge_SeqsToMPINumeric(seqmat,*mpimat);
3256:   PetscLogEventEnd(logkey_seqstompi,seqmat,0,0,0);
3257:   return(0);
3258: }
3259: static PetscEvent logkey_getlocalmat = 0;
3262: /*@C
3263:      MatGetLocalMat - Creates a SeqAIJ matrix by taking all its local rows

3265:     Not Collective

3267:    Input Parameters:
3268: +    A - the matrix 
3269: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 

3271:    Output Parameter:
3272: .    A_loc - the local sequential matrix generated

3274:     Level: developer

3276: @*/
3277: PetscErrorCode PETSCMAT_DLLEXPORT MatGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
3278: {
3279:   PetscErrorCode  ierr;
3280:   Mat_MPIAIJ      *mpimat=(Mat_MPIAIJ*)A->data;
3281:   Mat_SeqAIJ      *mat,*a=(Mat_SeqAIJ*)(mpimat->A)->data,*b=(Mat_SeqAIJ*)(mpimat->B)->data;
3282:   PetscInt        *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*cmap=mpimat->garray;
3283:   PetscScalar     *aa=a->a,*ba=b->a,*ca;
3284:   PetscInt        am=A->m,i,j,k,cstart=mpimat->cstart;
3285:   PetscInt        *ci,*cj,col,ncols_d,ncols_o,jo;

3288:   if (!logkey_getlocalmat) {
3289:     PetscLogEventRegister(&logkey_getlocalmat,"MatGetLocalMat",MAT_COOKIE);
3290:   }
3291:   PetscLogEventBegin(logkey_getlocalmat,A,0,0,0);
3292:   if (scall == MAT_INITIAL_MATRIX){
3293:     PetscMalloc((1+am)*sizeof(PetscInt),&ci);
3294:     ci[0] = 0;
3295:     for (i=0; i<am; i++){
3296:       ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
3297:     }
3298:     PetscMalloc((1+ci[am])*sizeof(PetscInt),&cj);
3299:     PetscMalloc((1+ci[am])*sizeof(PetscScalar),&ca);
3300:     k = 0;
3301:     for (i=0; i<am; i++) {
3302:       ncols_o = bi[i+1] - bi[i];
3303:       ncols_d = ai[i+1] - ai[i];
3304:       /* off-diagonal portion of A */
3305:       for (jo=0; jo<ncols_o; jo++) {
3306:         col = cmap[*bj];
3307:         if (col >= cstart) break;
3308:         cj[k]   = col; bj++;
3309:         ca[k++] = *ba++;
3310:       }
3311:       /* diagonal portion of A */
3312:       for (j=0; j<ncols_d; j++) {
3313:         cj[k]   = cstart + *aj++;
3314:         ca[k++] = *aa++;
3315:       }
3316:       /* off-diagonal portion of A */
3317:       for (j=jo; j<ncols_o; j++) {
3318:         cj[k]   = cmap[*bj++];
3319:         ca[k++] = *ba++;
3320:       }
3321:     }
3322:     /* put together the new matrix */
3323:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->N,ci,cj,ca,A_loc);
3324:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
3325:     /* Since these are PETSc arrays, change flags to free them as necessary. */
3326:     mat = (Mat_SeqAIJ*)(*A_loc)->data;
3327:     mat->freedata = PETSC_TRUE;
3328:     mat->nonew    = 0;
3329:   } else if (scall == MAT_REUSE_MATRIX){
3330:     mat=(Mat_SeqAIJ*)(*A_loc)->data;
3331:     ci = mat->i; cj = mat->j; ca = mat->a;
3332:     for (i=0; i<am; i++) {
3333:       /* off-diagonal portion of A */
3334:       ncols_o = bi[i+1] - bi[i];
3335:       for (jo=0; jo<ncols_o; jo++) {
3336:         col = cmap[*bj];
3337:         if (col >= cstart) break;
3338:         *ca++ = *ba++; bj++;
3339:       }
3340:       /* diagonal portion of A */
3341:       ncols_d = ai[i+1] - ai[i];
3342:       for (j=0; j<ncols_d; j++) *ca++ = *aa++;
3343:       /* off-diagonal portion of A */
3344:       for (j=jo; j<ncols_o; j++) {
3345:         *ca++ = *ba++; bj++;
3346:       }
3347:     }
3348:   } else {
3349:     SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
3350:   }

3352:   PetscLogEventEnd(logkey_getlocalmat,A,0,0,0);
3353:   return(0);
3354: }

3356: static PetscEvent logkey_getlocalmatcondensed = 0;
3359: /*@C
3360:      MatGetLocalMatCondensed - Creates a SeqAIJ matrix by taking all its local rows and NON-ZERO columns

3362:     Not Collective

3364:    Input Parameters:
3365: +    A - the matrix 
3366: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3367: -    row, col - index sets of rows and columns to extract (or PETSC_NULL)  

3369:    Output Parameter:
3370: .    A_loc - the local sequential matrix generated

3372:     Level: developer

3374: @*/
3375: PetscErrorCode PETSCMAT_DLLEXPORT MatGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
3376: {
3377:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data;
3378:   PetscErrorCode    ierr;
3379:   PetscInt          i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
3380:   IS                isrowa,iscola;
3381:   Mat               *aloc;

3384:   if (!logkey_getlocalmatcondensed) {
3385:     PetscLogEventRegister(&logkey_getlocalmatcondensed,"MatGetLocalMatCondensed",MAT_COOKIE);
3386:   }
3387:   PetscLogEventBegin(logkey_getlocalmatcondensed,A,0,0,0);
3388:   if (!row){
3389:     start = a->rstart; end = a->rend;
3390:     ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
3391:   } else {
3392:     isrowa = *row;
3393:   }
3394:   if (!col){
3395:     start = a->cstart;
3396:     cmap  = a->garray;
3397:     nzA   = a->A->n;
3398:     nzB   = a->B->n;
3399:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
3400:     ncols = 0;
3401:     for (i=0; i<nzB; i++) {
3402:       if (cmap[i] < start) idx[ncols++] = cmap[i];
3403:       else break;
3404:     }
3405:     imark = i;
3406:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;
3407:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
3408:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&iscola);
3409:     PetscFree(idx);
3410:   } else {
3411:     iscola = *col;
3412:   }
3413:   if (scall != MAT_INITIAL_MATRIX){
3414:     PetscMalloc(sizeof(Mat),&aloc);
3415:     aloc[0] = *A_loc;
3416:   }
3417:   MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
3418:   *A_loc = aloc[0];
3419:   PetscFree(aloc);
3420:   if (!row){
3421:     ISDestroy(isrowa);
3422:   }
3423:   if (!col){
3424:     ISDestroy(iscola);
3425:   }
3426:   PetscLogEventEnd(logkey_getlocalmatcondensed,A,0,0,0);
3427:   return(0);
3428: }

3430: static PetscEvent logkey_GetBrowsOfAcols = 0;
3433: /*@C
3434:     MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 

3436:     Collective on Mat

3438:    Input Parameters:
3439: +    A,B - the matrices in mpiaij format
3440: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3441: -    rowb, colb - index sets of rows and columns of B to extract (or PETSC_NULL)   

3443:    Output Parameter:
3444: +    rowb, colb - index sets of rows and columns of B to extract 
3445: .    brstart - row index of B_seq from which next B->m rows are taken from B's local rows
3446: -    B_seq - the sequential matrix generated

3448:     Level: developer

3450: @*/
3451: PetscErrorCode PETSCMAT_DLLEXPORT MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,PetscInt *brstart,Mat *B_seq)
3452: {
3453:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data,*b=(Mat_MPIAIJ*)B->data;
3454:   PetscErrorCode    ierr;
3455:   PetscInt          *idx,i,start,ncols,nzA,nzB,*cmap,imark;
3456:   IS                isrowb,iscolb;
3457:   Mat               *bseq;
3458: 
3460:   if (a->cstart != b->rstart || a->cend != b->rend){
3461:     SETERRQ4(PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",a->cstart,a->cend,b->rstart,b->rend);
3462:   }
3463:   if (!logkey_GetBrowsOfAcols) {
3464:     PetscLogEventRegister(&logkey_GetBrowsOfAcols,"MatGetBrowsOfAcols",MAT_COOKIE);
3465:   }
3466:   PetscLogEventBegin(logkey_GetBrowsOfAcols,A,B,0,0);
3467: 
3468:   if (scall == MAT_INITIAL_MATRIX){
3469:     start = a->cstart;
3470:     cmap  = a->garray;
3471:     nzA   = a->A->n;
3472:     nzB   = a->B->n;
3473:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
3474:     ncols = 0;
3475:     for (i=0; i<nzB; i++) {  /* row < local row index */
3476:       if (cmap[i] < start) idx[ncols++] = cmap[i];
3477:       else break;
3478:     }
3479:     imark = i;
3480:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;  /* local rows */
3481:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
3482:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&isrowb);
3483:     PetscFree(idx);
3484:     *brstart = imark;
3485:     ISCreateStride(PETSC_COMM_SELF,B->N,0,1,&iscolb);
3486:   } else {
3487:     if (!rowb || !colb) SETERRQ(PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
3488:     isrowb = *rowb; iscolb = *colb;
3489:     PetscMalloc(sizeof(Mat),&bseq);
3490:     bseq[0] = *B_seq;
3491:   }
3492:   MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
3493:   *B_seq = bseq[0];
3494:   PetscFree(bseq);
3495:   if (!rowb){
3496:     ISDestroy(isrowb);
3497:   } else {
3498:     *rowb = isrowb;
3499:   }
3500:   if (!colb){
3501:     ISDestroy(iscolb);
3502:   } else {
3503:     *colb = iscolb;
3504:   }
3505:   PetscLogEventEnd(logkey_GetBrowsOfAcols,A,B,0,0);
3506:   return(0);
3507: }

3509: static PetscEvent logkey_GetBrowsOfAocols = 0;
3512: /*@C
3513:     MatGetBrowsOfAoCols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
3514:     of the OFF-DIAGONAL portion of local A 

3516:     Collective on Mat

3518:    Input Parameters:
3519: +    A,B - the matrices in mpiaij format
3520: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
3521: .    startsj - starting point in B's sending and receiving j-arrays, saved for MAT_REUSE (or PETSC_NULL) 
3522: -    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or PETSC_NULL) 

3524:    Output Parameter:
3525: +    B_oth - the sequential matrix generated

3527:     Level: developer

3529: @*/
3530: PetscErrorCode PETSCMAT_DLLEXPORT MatGetBrowsOfAoCols(Mat A,Mat B,MatReuse scall,PetscInt **startsj,PetscScalar **bufa_ptr,Mat *B_oth)
3531: {
3532:   VecScatter_MPI_General *gen_to,*gen_from;
3533:   PetscErrorCode         ierr;
3534:   Mat_MPIAIJ             *a=(Mat_MPIAIJ*)A->data,*b=(Mat_MPIAIJ*)B->data;
3535:   Mat_SeqAIJ             *b_oth;
3536:   VecScatter             ctx=a->Mvctx;
3537:   MPI_Comm               comm=ctx->comm;
3538:   PetscMPIInt            *rprocs,*sprocs,tag=ctx->tag,rank;
3539:   PetscInt               *rowlen,*bufj,*bufJ,ncols,aBn=a->B->n,row,*b_othi,*b_othj;
3540:   PetscScalar            *rvalues,*svalues,*b_otha,*bufa,*bufA;
3541:   PetscInt               i,k,l,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
3542:   MPI_Request            *rwaits,*swaits;
3543:   MPI_Status             *sstatus,rstatus;
3544:   PetscInt               *cols;
3545:   PetscScalar            *vals;
3546:   PetscMPIInt            j;
3547: 
3549:   if (a->cstart != b->rstart || a->cend != b->rend){
3550:     SETERRQ4(PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",a->cstart,a->cend,b->rstart,b->rend);
3551:   }
3552:   if (!logkey_GetBrowsOfAocols) {
3553:     PetscLogEventRegister(&logkey_GetBrowsOfAocols,"MatGetBrAoCol",MAT_COOKIE);
3554:   }
3555:   PetscLogEventBegin(logkey_GetBrowsOfAocols,A,B,0,0);
3556:   MPI_Comm_rank(comm,&rank);

3558:   gen_to   = (VecScatter_MPI_General*)ctx->todata;
3559:   gen_from = (VecScatter_MPI_General*)ctx->fromdata;
3560:   rvalues  = gen_from->values; /* holds the length of sending row */
3561:   svalues  = gen_to->values;   /* holds the length of receiving row */
3562:   nrecvs   = gen_from->n;
3563:   nsends   = gen_to->n;
3564:   rwaits   = gen_from->requests;
3565:   swaits   = gen_to->requests;
3566:   srow     = gen_to->indices;   /* local row index to be sent */
3567:   rstarts  = gen_from->starts;
3568:   sstarts  = gen_to->starts;
3569:   rprocs   = gen_from->procs;
3570:   sprocs   = gen_to->procs;
3571:   sstatus  = gen_to->sstatus;

3573:   if (!startsj || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
3574:   if (scall == MAT_INITIAL_MATRIX){
3575:     /* i-array */
3576:     /*---------*/
3577:     /*  post receives */
3578:     for (i=0; i<nrecvs; i++){
3579:       rowlen = (PetscInt*)rvalues + rstarts[i];
3580:       nrows = rstarts[i+1]-rstarts[i];
3581:       MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
3582:     }

3584:     /* pack the outgoing message */
3585:     PetscMalloc((nsends+nrecvs+3)*sizeof(PetscInt),&sstartsj);
3586:     rstartsj = sstartsj + nsends +1;
3587:     sstartsj[0] = 0;  rstartsj[0] = 0;
3588:     len = 0; /* total length of j or a array to be sent */
3589:     k = 0;
3590:     for (i=0; i<nsends; i++){
3591:       rowlen = (PetscInt*)svalues + sstarts[i];
3592:       nrows = sstarts[i+1]-sstarts[i]; /* num of rows */
3593:       for (j=0; j<nrows; j++) {
3594:         row = srow[k] + b->rowners[rank]; /* global row idx */
3595:         MatGetRow_MPIAIJ(B,row,&rowlen[j],PETSC_NULL,PETSC_NULL); /* rowlength */
3596:         len += rowlen[j];
3597:         MatRestoreRow_MPIAIJ(B,row,&ncols,PETSC_NULL,PETSC_NULL);
3598:         k++;
3599:       }
3600:       MPI_Isend(rowlen,nrows,MPIU_INT,sprocs[i],tag,comm,swaits+i);
3601:        sstartsj[i+1] = len;  /* starting point of (i+1)-th outgoing msg in bufj and bufa */
3602:     }
3603:     /* recvs and sends of i-array are completed */
3604:     i = nrecvs;
3605:     while (i--) {
3606:       MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3607:     }
3608:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3609:     /* allocate buffers for sending j and a arrays */
3610:     PetscMalloc((len+1)*sizeof(PetscInt),&bufj);
3611:     PetscMalloc((len+1)*sizeof(PetscScalar),&bufa);

3613:     /* create i-array of B_oth */
3614:     PetscMalloc((aBn+2)*sizeof(PetscInt),&b_othi);
3615:     b_othi[0] = 0;
3616:     len = 0; /* total length of j or a array to be received */
3617:     k = 0;
3618:     for (i=0; i<nrecvs; i++){
3619:       rowlen = (PetscInt*)rvalues + rstarts[i];
3620:       nrows = rstarts[i+1]-rstarts[i];
3621:       for (j=0; j<nrows; j++) {
3622:         b_othi[k+1] = b_othi[k] + rowlen[j];
3623:         len += rowlen[j]; k++;
3624:       }
3625:       rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
3626:     }

3628:     /* allocate space for j and a arrrays of B_oth */
3629:     PetscMalloc((b_othi[aBn]+1)*sizeof(PetscInt),&b_othj);
3630:     PetscMalloc((b_othi[aBn]+1)*sizeof(PetscScalar),&b_otha);

3632:     /* j-array */
3633:     /*---------*/
3634:     /*  post receives of j-array */
3635:     for (i=0; i<nrecvs; i++){
3636:       nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
3637:       MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
3638:     }
3639:     k = 0;
3640:     for (i=0; i<nsends; i++){
3641:       nrows = sstarts[i+1]-sstarts[i]; /* num of rows */
3642:       bufJ = bufj+sstartsj[i];
3643:       for (j=0; j<nrows; j++) {
3644:         row  = srow[k++] + b->rowners[rank]; /* global row idx */
3645:         MatGetRow_MPIAIJ(B,row,&ncols,&cols,PETSC_NULL);
3646:         for (l=0; l<ncols; l++){
3647:           *bufJ++ = cols[l];
3648:         }
3649:         MatRestoreRow_MPIAIJ(B,row,&ncols,&cols,PETSC_NULL);
3650:       }
3651:       MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
3652:     }

3654:     /* recvs and sends of j-array are completed */
3655:     i = nrecvs;
3656:     while (i--) {
3657:       MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3658:     }
3659:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3660:   } else if (scall == MAT_REUSE_MATRIX){
3661:     sstartsj = *startsj;
3662:     rstartsj = sstartsj + nsends +1;
3663:     bufa     = *bufa_ptr;
3664:     b_oth    = (Mat_SeqAIJ*)(*B_oth)->data;
3665:     b_otha   = b_oth->a;
3666:   } else {
3667:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
3668:   }

3670:   /* a-array */
3671:   /*---------*/
3672:   /*  post receives of a-array */
3673:   for (i=0; i<nrecvs; i++){
3674:     nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
3675:     MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
3676:   }
3677:   k = 0;
3678:   for (i=0; i<nsends; i++){
3679:     nrows = sstarts[i+1]-sstarts[i];
3680:     bufA = bufa+sstartsj[i];
3681:     for (j=0; j<nrows; j++) {
3682:       row  = srow[k++] + b->rowners[rank]; /* global row idx */
3683:       MatGetRow_MPIAIJ(B,row,&ncols,PETSC_NULL,&vals);
3684:       for (l=0; l<ncols; l++){
3685:         *bufA++ = vals[l];
3686:       }
3687:       MatRestoreRow_MPIAIJ(B,row,&ncols,PETSC_NULL,&vals);

3689:     }
3690:     MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
3691:   }
3692:   /* recvs and sends of a-array are completed */
3693:   i = nrecvs;
3694:   while (i--) {
3695:     MPI_Waitany(nrecvs,rwaits,&j,&rstatus);
3696:   }
3697:    if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
3698: 
3699:   if (scall == MAT_INITIAL_MATRIX){
3700:     /* put together the new matrix */
3701:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->N,b_othi,b_othj,b_otha,B_oth);

3703:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
3704:     /* Since these are PETSc arrays, change flags to free them as necessary. */
3705:     b_oth = (Mat_SeqAIJ *)(*B_oth)->data;
3706:     b_oth->freedata = PETSC_TRUE;
3707:     b_oth->nonew    = 0;

3709:     PetscFree(bufj);
3710:     if (!startsj || !bufa_ptr){
3711:       PetscFree(sstartsj);
3712:       PetscFree(bufa_ptr);
3713:     } else {
3714:       *startsj  = sstartsj;
3715:       *bufa_ptr = bufa;
3716:     }
3717:   }
3718:   PetscLogEventEnd(logkey_GetBrowsOfAocols,A,B,0,0);
3719: 
3720:   return(0);
3721: }

3723: /*MC
3724:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

3726:    Options Database Keys:
3727: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()

3729:   Level: beginner

3731: .seealso: MatCreateMPIAIJ
3732: M*/

3737: PetscErrorCode PETSCMAT_DLLEXPORT MatCreate_MPIAIJ(Mat B)
3738: {
3739:   Mat_MPIAIJ     *b;
3741:   PetscInt       i;
3742:   PetscMPIInt    size;

3745:   MPI_Comm_size(B->comm,&size);

3747:   PetscNew(Mat_MPIAIJ,&b);
3748:   B->data         = (void*)b;
3749:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3750:   B->factor       = 0;
3751:   B->bs           = 1;
3752:   B->assembled    = PETSC_FALSE;
3753:   B->mapping      = 0;

3755:   B->insertmode      = NOT_SET_VALUES;
3756:   b->size            = size;
3757:   MPI_Comm_rank(B->comm,&b->rank);

3759:   PetscSplitOwnership(B->comm,&B->m,&B->M);
3760:   PetscSplitOwnership(B->comm,&B->n,&B->N);

3762:   /* the information in the maps duplicates the information computed below, eventually 
3763:      we should remove the duplicate information that is not contained in the maps */
3764:   PetscMapCreateMPI(B->comm,B->m,B->M,&B->rmap);
3765:   PetscMapCreateMPI(B->comm,B->n,B->N,&B->cmap);

3767:   /* build local table of row and column ownerships */
3768:   PetscMalloc(2*(b->size+2)*sizeof(PetscInt),&b->rowners);
3769:   PetscLogObjectMemory(B,2*(b->size+2)*sizeof(PetscInt)+sizeof(struct _p_Mat)+sizeof(Mat_MPIAIJ));
3770:   b->cowners = b->rowners + b->size + 2;
3771:   MPI_Allgather(&B->m,1,MPIU_INT,b->rowners+1,1,MPIU_INT,B->comm);
3772:   b->rowners[0] = 0;
3773:   for (i=2; i<=b->size; i++) {
3774:     b->rowners[i] += b->rowners[i-1];
3775:   }
3776:   b->rstart = b->rowners[b->rank];
3777:   b->rend   = b->rowners[b->rank+1];
3778:   MPI_Allgather(&B->n,1,MPIU_INT,b->cowners+1,1,MPIU_INT,B->comm);
3779:   b->cowners[0] = 0;
3780:   for (i=2; i<=b->size; i++) {
3781:     b->cowners[i] += b->cowners[i-1];
3782:   }
3783:   b->cstart = b->cowners[b->rank];
3784:   b->cend   = b->cowners[b->rank+1];

3786:   /* build cache for off array entries formed */
3787:   MatStashCreate_Private(B->comm,1,&B->stash);
3788:   b->donotstash  = PETSC_FALSE;
3789:   b->colmap      = 0;
3790:   b->garray      = 0;
3791:   b->roworiented = PETSC_TRUE;

3793:   /* stuff used for matrix vector multiply */
3794:   b->lvec      = PETSC_NULL;
3795:   b->Mvctx     = PETSC_NULL;

3797:   /* stuff for MatGetRow() */
3798:   b->rowindices   = 0;
3799:   b->rowvalues    = 0;
3800:   b->getrowactive = PETSC_FALSE;

3802:   /* Explicitly create 2 MATSEQAIJ matrices. */
3803:   MatCreate(PETSC_COMM_SELF,&b->A);
3804:   MatSetSizes(b->A,B->m,B->n,B->m,B->n);
3805:   MatSetType(b->A,MATSEQAIJ);
3806:   PetscLogObjectParent(B,b->A);
3807:   MatCreate(PETSC_COMM_SELF,&b->B);
3808:   MatSetSizes(b->B,B->m,B->N,B->m,B->N);
3809:   MatSetType(b->B,MATSEQAIJ);
3810:   PetscLogObjectParent(B,b->B);

3812:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
3813:                                      "MatStoreValues_MPIAIJ",
3814:                                      MatStoreValues_MPIAIJ);
3815:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
3816:                                      "MatRetrieveValues_MPIAIJ",
3817:                                      MatRetrieveValues_MPIAIJ);
3818:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
3819:                                      "MatGetDiagonalBlock_MPIAIJ",
3820:                                      MatGetDiagonalBlock_MPIAIJ);
3821:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatIsTranspose_C",
3822:                                      "MatIsTranspose_MPIAIJ",
3823:                                      MatIsTranspose_MPIAIJ);
3824:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocation_C",
3825:                                      "MatMPIAIJSetPreallocation_MPIAIJ",
3826:                                      MatMPIAIJSetPreallocation_MPIAIJ);
3827:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",
3828:                                      "MatMPIAIJSetPreallocationCSR_MPIAIJ",
3829:                                      MatMPIAIJSetPreallocationCSR_MPIAIJ);
3830:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
3831:                                      "MatDiagonalScaleLocal_MPIAIJ",
3832:                                      MatDiagonalScaleLocal_MPIAIJ);
3833:   return(0);
3834: }