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->cmap.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->cmap.N+1)*sizeof(PetscInt),&aij->colmap);
 29:   PetscLogObjectMemory(mat,mat->cmap.N*sizeof(PetscInt));
 30:   PetscMemzero(aij->colmap,mat->cmap.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,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
 59:       N = nrow1++ - 1; a->nz++; high1++; \
 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,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
 93:       N = nrow2++ - 1; b->nz++; high2++;\
 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 MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
108: {
109:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)A->data;
110:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
112:   PetscInt       l,*garray = mat->garray,diag;

115:   /* code only works for square matrices A */

117:   /* find size of row to the left of the diagonal part */
118:   MatGetOwnershipRange(A,&diag,0);
119:   row  = row - diag;
120:   for (l=0; l<b->i[row+1]-b->i[row]; l++) {
121:     if (garray[b->j[b->i[row]+l]] > diag) break;
122:   }
123:   PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));

125:   /* diagonal part */
126:   PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));

128:   /* right of diagonal part */
129:   PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));
130:   return(0);
131: }

135: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
136: {
137:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
138:   PetscScalar    value;
140:   PetscInt       i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
141:   PetscInt       cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;
142:   PetscTruth     roworiented = aij->roworiented;

144:   /* Some Variables required in the macro */
145:   Mat            A = aij->A;
146:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)A->data;
147:   PetscInt       *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
148:   PetscScalar    *aa = a->a;
149:   PetscTruth     ignorezeroentries = a->ignorezeroentries;
150:   Mat            B = aij->B;
151:   Mat_SeqAIJ     *b = (Mat_SeqAIJ*)B->data;
152:   PetscInt       *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap.n,am = aij->A->rmap.n;
153:   PetscScalar    *ba = b->a;

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

160:   for (i=0; i<m; i++) {
161:     if (im[i] < 0) continue;
162: #if defined(PETSC_USE_DEBUG)
163:     if (im[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap.N-1);
164: #endif
165:     if (im[i] >= rstart && im[i] < rend) {
166:       row      = im[i] - rstart;
167:       lastcol1 = -1;
168:       rp1      = aj + ai[row];
169:       ap1      = aa + ai[row];
170:       rmax1    = aimax[row];
171:       nrow1    = ailen[row];
172:       low1     = 0;
173:       high1    = nrow1;
174:       lastcol2 = -1;
175:       rp2      = bj + bi[row];
176:       ap2      = ba + bi[row];
177:       rmax2    = bimax[row];
178:       nrow2    = bilen[row];
179:       low2     = 0;
180:       high2    = nrow2;

182:       for (j=0; j<n; j++) {
183:         if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
184:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
185:         if (in[j] >= cstart && in[j] < cend){
186:           col = in[j] - cstart;
187:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
188:         } else if (in[j] < 0) continue;
189: #if defined(PETSC_USE_DEBUG)
190:         else if (in[j] >= mat->cmap.N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap.N-1);}
191: #endif
192:         else {
193:           if (mat->was_assembled) {
194:             if (!aij->colmap) {
195:               CreateColmap_MPIAIJ_Private(mat);
196:             }
197: #if defined (PETSC_USE_CTABLE)
198:             PetscTableFind(aij->colmap,in[j]+1,&col);
199:             col--;
200: #else
201:             col = aij->colmap[in[j]] - 1;
202: #endif
203:             if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
204:               DisAssemble_MPIAIJ(mat);
205:               col =  in[j];
206:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
207:               B = aij->B;
208:               b = (Mat_SeqAIJ*)B->data;
209:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
210:               rp2      = bj + bi[row];
211:               ap2      = ba + bi[row];
212:               rmax2    = bimax[row];
213:               nrow2    = bilen[row];
214:               low2     = 0;
215:               high2    = nrow2;
216:               bm       = aij->B->rmap.n;
217:               ba = b->a;
218:             }
219:           } else col = in[j];
220:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
221:         }
222:       }
223:     } else {
224:       if (!aij->donotstash) {
225:         if (roworiented) {
226:           if (ignorezeroentries && v[i*n] == 0.0) continue;
227:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
228:         } else {
229:           if (ignorezeroentries && v[i] == 0.0) continue;
230:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
231:         }
232:       }
233:     }
234:   }
235:   return(0);
236: }

240: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
241: {
242:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
244:   PetscInt       i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
245:   PetscInt       cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;

248:   for (i=0; i<m; i++) {
249:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
250:     if (idxm[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap.N-1);
251:     if (idxm[i] >= rstart && idxm[i] < rend) {
252:       row = idxm[i] - rstart;
253:       for (j=0; j<n; j++) {
254:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
255:         if (idxn[j] >= mat->cmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap.N-1);
256:         if (idxn[j] >= cstart && idxn[j] < cend){
257:           col = idxn[j] - cstart;
258:           MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
259:         } else {
260:           if (!aij->colmap) {
261:             CreateColmap_MPIAIJ_Private(mat);
262:           }
263: #if defined (PETSC_USE_CTABLE)
264:           PetscTableFind(aij->colmap,idxn[j]+1,&col);
265:           col --;
266: #else
267:           col = aij->colmap[idxn[j]] - 1;
268: #endif
269:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
270:           else {
271:             MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
272:           }
273:         }
274:       }
275:     } else {
276:       SETERRQ(PETSC_ERR_SUP,"Only local values currently supported");
277:     }
278:   }
279:   return(0);
280: }

284: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
285: {
286:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
288:   PetscInt       nstash,reallocs;
289:   InsertMode     addv;

292:   if (aij->donotstash) {
293:     return(0);
294:   }

296:   /* make sure all processors are either in INSERTMODE or ADDMODE */
297:   MPI_Allreduce(&mat->insertmode,&addv,1,MPI_INT,MPI_BOR,((PetscObject)mat)->comm);
298:   if (addv == (ADD_VALUES|INSERT_VALUES)) {
299:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
300:   }
301:   mat->insertmode = addv; /* in case this processor had no cache */

303:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap.range);
304:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
305:   PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
306:   return(0);
307: }

311: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
312: {
313:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
314:   Mat_SeqAIJ     *a=(Mat_SeqAIJ *)aij->A->data;
316:   PetscMPIInt    n;
317:   PetscInt       i,j,rstart,ncols,flg;
318:   PetscInt       *row,*col,other_disassembled;
319:   PetscScalar    *val;
320:   InsertMode     addv = mat->insertmode;

322:   /* do not use 'b = (Mat_SeqAIJ *)aij->B->data' as B can be reset in disassembly */
324:   if (!aij->donotstash) {
325:     while (1) {
326:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
327:       if (!flg) break;

329:       for (i=0; i<n;) {
330:         /* Now identify the consecutive vals belonging to the same row */
331:         for (j=i,rstart=row[j]; j<n; j++) { if (row[j] != rstart) break; }
332:         if (j < n) ncols = j-i;
333:         else       ncols = n-i;
334:         /* Now assemble all these values with a single function call */
335:         MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
336:         i = j;
337:       }
338:     }
339:     MatStashScatterEnd_Private(&mat->stash);
340:   }
341:   a->compressedrow.use     = PETSC_FALSE;
342:   MatAssemblyBegin(aij->A,mode);
343:   MatAssemblyEnd(aij->A,mode);

345:   /* determine if any processor has disassembled, if so we must 
346:      also disassemble ourselfs, in order that we may reassemble. */
347:   /*
348:      if nonzero structure of submatrix B cannot change then we know that
349:      no processor disassembled thus we can skip this stuff
350:   */
351:   if (!((Mat_SeqAIJ*)aij->B->data)->nonew)  {
352:     MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPI_INT,MPI_PROD,((PetscObject)mat)->comm);
353:     if (mat->was_assembled && !other_disassembled) {
354:       DisAssemble_MPIAIJ(mat);
355:     }
356:   }
357:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
358:     MatSetUpMultiply_MPIAIJ(mat);
359:   }
360:   MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
361:   ((Mat_SeqAIJ *)aij->B->data)->compressedrow.use = PETSC_TRUE; /* b->compressedrow.use */
362:   MatAssemblyBegin(aij->B,mode);
363:   MatAssemblyEnd(aij->B,mode);

365:   PetscFree(aij->rowvalues);
366:   aij->rowvalues = 0;

368:   /* used by MatAXPY() */
369:   a->xtoy = 0; ((Mat_SeqAIJ *)aij->B->data)->xtoy = 0;  /* b->xtoy = 0 */
370:   a->XtoY = 0; ((Mat_SeqAIJ *)aij->B->data)->XtoY = 0;  /* b->XtoY = 0 */

372:   return(0);
373: }

377: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
378: {
379:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;

383:   MatZeroEntries(l->A);
384:   MatZeroEntries(l->B);
385:   return(0);
386: }

390: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag)
391: {
392:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;
394:   PetscMPIInt    size = l->size,imdex,n,rank = l->rank,tag = ((PetscObject)A)->tag,lastidx = -1;
395:   PetscInt       i,*owners = A->rmap.range;
396:   PetscInt       *nprocs,j,idx,nsends,row;
397:   PetscInt       nmax,*svalues,*starts,*owner,nrecvs;
398:   PetscInt       *rvalues,count,base,slen,*source;
399:   PetscInt       *lens,*lrows,*values,rstart=A->rmap.rstart;
400:   MPI_Comm       comm = ((PetscObject)A)->comm;
401:   MPI_Request    *send_waits,*recv_waits;
402:   MPI_Status     recv_status,*send_status;
403: #if defined(PETSC_DEBUG)
404:   PetscTruth     found = PETSC_FALSE;
405: #endif

408:   /*  first count number of contributors to each processor */
409:   PetscMalloc(2*size*sizeof(PetscInt),&nprocs);
410:   PetscMemzero(nprocs,2*size*sizeof(PetscInt));
411:   PetscMalloc((N+1)*sizeof(PetscInt),&owner); /* see note*/
412:   j = 0;
413:   for (i=0; i<N; i++) {
414:     if (lastidx > (idx = rows[i])) j = 0;
415:     lastidx = idx;
416:     for (; j<size; j++) {
417:       if (idx >= owners[j] && idx < owners[j+1]) {
418:         nprocs[2*j]++;
419:         nprocs[2*j+1] = 1;
420:         owner[i] = j;
421: #if defined(PETSC_DEBUG)
422:         found = PETSC_TRUE;
423: #endif
424:         break;
425:       }
426:     }
427: #if defined(PETSC_DEBUG)
428:     if (!found) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Index out of range");
429:     found = PETSC_FALSE;
430: #endif
431:   }
432:   nsends = 0;  for (i=0; i<size; i++) { nsends += nprocs[2*i+1];}

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

437:   /* post receives:   */
438:   PetscMalloc((nrecvs+1)*(nmax+1)*sizeof(PetscInt),&rvalues);
439:   PetscMalloc((nrecvs+1)*sizeof(MPI_Request),&recv_waits);
440:   for (i=0; i<nrecvs; i++) {
441:     MPI_Irecv(rvalues+nmax*i,nmax,MPIU_INT,MPI_ANY_SOURCE,tag,comm,recv_waits+i);
442:   }

444:   /* do sends:
445:       1) starts[i] gives the starting index in svalues for stuff going to 
446:          the ith processor
447:   */
448:   PetscMalloc((N+1)*sizeof(PetscInt),&svalues);
449:   PetscMalloc((nsends+1)*sizeof(MPI_Request),&send_waits);
450:   PetscMalloc((size+1)*sizeof(PetscInt),&starts);
451:   starts[0] = 0;
452:   for (i=1; i<size; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
453:   for (i=0; i<N; i++) {
454:     svalues[starts[owner[i]]++] = rows[i];
455:   }

457:   starts[0] = 0;
458:   for (i=1; i<size+1; i++) { starts[i] = starts[i-1] + nprocs[2*i-2];}
459:   count = 0;
460:   for (i=0; i<size; i++) {
461:     if (nprocs[2*i+1]) {
462:       MPI_Isend(svalues+starts[i],nprocs[2*i],MPIU_INT,i,tag,comm,send_waits+count++);
463:     }
464:   }
465:   PetscFree(starts);

467:   base = owners[rank];

469:   /*  wait on receives */
470:   PetscMalloc(2*(nrecvs+1)*sizeof(PetscInt),&lens);
471:   source = lens + nrecvs;
472:   count  = nrecvs; slen = 0;
473:   while (count) {
474:     MPI_Waitany(nrecvs,recv_waits,&imdex,&recv_status);
475:     /* unpack receives into our local space */
476:     MPI_Get_count(&recv_status,MPIU_INT,&n);
477:     source[imdex]  = recv_status.MPI_SOURCE;
478:     lens[imdex]    = n;
479:     slen          += n;
480:     count--;
481:   }
482:   PetscFree(recv_waits);
483: 
484:   /* move the data into the send scatter */
485:   PetscMalloc((slen+1)*sizeof(PetscInt),&lrows);
486:   count = 0;
487:   for (i=0; i<nrecvs; i++) {
488:     values = rvalues + i*nmax;
489:     for (j=0; j<lens[i]; j++) {
490:       lrows[count++] = values[j] - base;
491:     }
492:   }
493:   PetscFree(rvalues);
494:   PetscFree(lens);
495:   PetscFree(owner);
496:   PetscFree(nprocs);
497: 
498:   /* actually zap the local rows */
499:   /*
500:         Zero the required rows. If the "diagonal block" of the matrix
501:      is square and the user wishes to set the diagonal we use separate
502:      code so that MatSetValues() is not called for each diagonal allocating
503:      new memory, thus calling lots of mallocs and slowing things down.

505:        Contributed by: Matthew Knepley
506:   */
507:   /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
508:   MatZeroRows(l->B,slen,lrows,0.0);
509:   if ((diag != 0.0) && (l->A->rmap.N == l->A->cmap.N)) {
510:     MatZeroRows(l->A,slen,lrows,diag);
511:   } else if (diag != 0.0) {
512:     MatZeroRows(l->A,slen,lrows,0.0);
513:     if (((Mat_SeqAIJ*)l->A->data)->nonew) {
514:       SETERRQ(PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options\n\
515: MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
516:     }
517:     for (i = 0; i < slen; i++) {
518:       row  = lrows[i] + rstart;
519:       MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
520:     }
521:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
522:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
523:   } else {
524:     MatZeroRows(l->A,slen,lrows,0.0);
525:   }
526:   PetscFree(lrows);

528:   /* wait on sends */
529:   if (nsends) {
530:     PetscMalloc(nsends*sizeof(MPI_Status),&send_status);
531:     MPI_Waitall(nsends,send_waits,send_status);
532:     PetscFree(send_status);
533:   }
534:   PetscFree(send_waits);
535:   PetscFree(svalues);

537:   return(0);
538: }

542: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
543: {
544:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
546:   PetscInt       nt;

549:   VecGetLocalSize(xx,&nt);
550:   if (nt != A->cmap.n) {
551:     SETERRQ2(PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap.n,nt);
552:   }
553:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
554:   (*a->A->ops->mult)(a->A,xx,yy);
555:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
556:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
557:   return(0);
558: }

562: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
563: {
564:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

568:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
569:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
570:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
571:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
572:   return(0);
573: }

577: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
578: {
579:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
581:   PetscTruth     merged;

584:   VecScatterGetMerged(a->Mvctx,&merged);
585:   /* do nondiagonal part */
586:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
587:   if (!merged) {
588:     /* send it on its way */
589:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
590:     /* do local part */
591:     (*a->A->ops->multtranspose)(a->A,xx,yy);
592:     /* receive remote parts: note this assumes the values are not actually */
593:     /* added in yy until the next line, */
594:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
595:   } else {
596:     /* do local part */
597:     (*a->A->ops->multtranspose)(a->A,xx,yy);
598:     /* send it on its way */
599:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
600:     /* values actually were received in the Begin() but we need to call this nop */
601:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
602:   }
603:   return(0);
604: }

609: PetscErrorCode  MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscTruth *f)
610: {
611:   MPI_Comm       comm;
612:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ *) Amat->data, *Bij;
613:   Mat            Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
614:   IS             Me,Notme;
616:   PetscInt       M,N,first,last,*notme,i;
617:   PetscMPIInt    size;


621:   /* Easy test: symmetric diagonal block */
622:   Bij = (Mat_MPIAIJ *) Bmat->data; Bdia = Bij->A;
623:   MatIsTranspose(Adia,Bdia,tol,f);
624:   if (!*f) return(0);
625:   PetscObjectGetComm((PetscObject)Amat,&comm);
626:   MPI_Comm_size(comm,&size);
627:   if (size == 1) return(0);

629:   /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
630:   MatGetSize(Amat,&M,&N);
631:   MatGetOwnershipRange(Amat,&first,&last);
632:   PetscMalloc((N-last+first)*sizeof(PetscInt),&notme);
633:   for (i=0; i<first; i++) notme[i] = i;
634:   for (i=last; i<M; i++) notme[i-last+first] = i;
635:   ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,&Notme);
636:   ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
637:   MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
638:   Aoff = Aoffs[0];
639:   MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
640:   Boff = Boffs[0];
641:   MatIsTranspose(Aoff,Boff,tol,f);
642:   MatDestroyMatrices(1,&Aoffs);
643:   MatDestroyMatrices(1,&Boffs);
644:   ISDestroy(Me);
645:   ISDestroy(Notme);

647:   return(0);
648: }

653: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
654: {
655:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

659:   /* do nondiagonal part */
660:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
661:   /* send it on its way */
662:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
663:   /* do local part */
664:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
665:   /* receive remote parts */
666:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
667:   return(0);
668: }

670: /*
671:   This only works correctly for square matrices where the subblock A->A is the 
672:    diagonal block
673: */
676: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
677: {
679:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

682:   if (A->rmap.N != A->cmap.N) SETERRQ(PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
683:   if (A->rmap.rstart != A->cmap.rstart || A->rmap.rend != A->cmap.rend) {
684:     SETERRQ(PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
685:   }
686:   MatGetDiagonal(a->A,v);
687:   return(0);
688: }

692: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
693: {
694:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

698:   MatScale(a->A,aa);
699:   MatScale(a->B,aa);
700:   return(0);
701: }

705: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
706: {
707:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

711: #if defined(PETSC_USE_LOG)
712:   PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap.N,mat->cmap.N);
713: #endif
714:   MatStashDestroy_Private(&mat->stash);
715:   MatDestroy(aij->A);
716:   MatDestroy(aij->B);
717: #if defined (PETSC_USE_CTABLE)
718:   if (aij->colmap) {PetscTableDestroy(aij->colmap);}
719: #else
720:   PetscFree(aij->colmap);
721: #endif
722:   PetscFree(aij->garray);
723:   if (aij->lvec)   {VecDestroy(aij->lvec);}
724:   if (aij->Mvctx)  {VecScatterDestroy(aij->Mvctx);}
725:   PetscFree(aij->rowvalues);
726:   PetscFree(aij->ld);
727:   PetscFree(aij);

729:   PetscObjectChangeTypeName((PetscObject)mat,0);
730:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C","",PETSC_NULL);
731:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C","",PETSC_NULL);
732:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C","",PETSC_NULL);
733:   PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C","",PETSC_NULL);
734:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C","",PETSC_NULL);
735:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C","",PETSC_NULL);
736:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C","",PETSC_NULL);
737:   return(0);
738: }

742: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
743: {
744:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
745:   Mat_SeqAIJ*       A = (Mat_SeqAIJ*)aij->A->data;
746:   Mat_SeqAIJ*       B = (Mat_SeqAIJ*)aij->B->data;
747:   PetscErrorCode    ierr;
748:   PetscMPIInt       rank,size,tag = ((PetscObject)viewer)->tag;
749:   int               fd;
750:   PetscInt          nz,header[4],*row_lengths,*range=0,rlen,i;
751:   PetscInt          nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap.rstart,rnz;
752:   PetscScalar       *column_values;

755:   MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
756:   MPI_Comm_size(((PetscObject)mat)->comm,&size);
757:   nz   = A->nz + B->nz;
758:   if (!rank) {
759:     header[0] = MAT_FILE_COOKIE;
760:     header[1] = mat->rmap.N;
761:     header[2] = mat->cmap.N;
762:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
763:     PetscViewerBinaryGetDescriptor(viewer,&fd);
764:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
765:     /* get largest number of rows any processor has */
766:     rlen = mat->rmap.n;
767:     range = mat->rmap.range;
768:     for (i=1; i<size; i++) {
769:       rlen = PetscMax(rlen,range[i+1] - range[i]);
770:     }
771:   } else {
772:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,((PetscObject)mat)->comm);
773:     rlen = mat->rmap.n;
774:   }

776:   /* load up the local row counts */
777:   PetscMalloc((rlen+1)*sizeof(PetscInt),&row_lengths);
778:   for (i=0; i<mat->rmap.n; i++) {
779:     row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
780:   }

782:   /* store the row lengths to the file */
783:   if (!rank) {
784:     MPI_Status status;
785:     PetscBinaryWrite(fd,row_lengths,mat->rmap.n,PETSC_INT,PETSC_TRUE);
786:     for (i=1; i<size; i++) {
787:       rlen = range[i+1] - range[i];
788:       MPI_Recv(row_lengths,rlen,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
789:       PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
790:     }
791:   } else {
792:     MPI_Send(row_lengths,mat->rmap.n,MPIU_INT,0,tag,((PetscObject)mat)->comm);
793:   }
794:   PetscFree(row_lengths);

796:   /* load up the local column indices */
797:   nzmax = nz; /* )th processor needs space a largest processor needs */
798:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,((PetscObject)mat)->comm);
799:   PetscMalloc((nzmax+1)*sizeof(PetscInt),&column_indices);
800:   cnt  = 0;
801:   for (i=0; i<mat->rmap.n; i++) {
802:     for (j=B->i[i]; j<B->i[i+1]; j++) {
803:       if ( (col = garray[B->j[j]]) > cstart) break;
804:       column_indices[cnt++] = col;
805:     }
806:     for (k=A->i[i]; k<A->i[i+1]; k++) {
807:       column_indices[cnt++] = A->j[k] + cstart;
808:     }
809:     for (; j<B->i[i+1]; j++) {
810:       column_indices[cnt++] = garray[B->j[j]];
811:     }
812:   }
813:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

815:   /* store the column indices to the file */
816:   if (!rank) {
817:     MPI_Status status;
818:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
819:     for (i=1; i<size; i++) {
820:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
821:       if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
822:       MPI_Recv(column_indices,rnz,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
823:       PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
824:     }
825:   } else {
826:     MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
827:     MPI_Send(column_indices,nz,MPIU_INT,0,tag,((PetscObject)mat)->comm);
828:   }
829:   PetscFree(column_indices);

831:   /* load up the local column values */
832:   PetscMalloc((nzmax+1)*sizeof(PetscScalar),&column_values);
833:   cnt  = 0;
834:   for (i=0; i<mat->rmap.n; i++) {
835:     for (j=B->i[i]; j<B->i[i+1]; j++) {
836:       if ( garray[B->j[j]] > cstart) break;
837:       column_values[cnt++] = B->a[j];
838:     }
839:     for (k=A->i[i]; k<A->i[i+1]; k++) {
840:       column_values[cnt++] = A->a[k];
841:     }
842:     for (; j<B->i[i+1]; j++) {
843:       column_values[cnt++] = B->a[j];
844:     }
845:   }
846:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

848:   /* store the column values to the file */
849:   if (!rank) {
850:     MPI_Status status;
851:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
852:     for (i=1; i<size; i++) {
853:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,((PetscObject)mat)->comm,&status);
854:       if (rnz > nzmax) SETERRQ2(PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
855:       MPI_Recv(column_values,rnz,MPIU_SCALAR,i,tag,((PetscObject)mat)->comm,&status);
856:       PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
857:     }
858:   } else {
859:     MPI_Send(&nz,1,MPIU_INT,0,tag,((PetscObject)mat)->comm);
860:     MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,((PetscObject)mat)->comm);
861:   }
862:   PetscFree(column_values);
863:   return(0);
864: }

868: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
869: {
870:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
871:   PetscErrorCode    ierr;
872:   PetscMPIInt       rank = aij->rank,size = aij->size;
873:   PetscTruth        isdraw,iascii,isbinary;
874:   PetscViewer       sviewer;
875:   PetscViewerFormat format;

878:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
879:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
880:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
881:   if (iascii) {
882:     PetscViewerGetFormat(viewer,&format);
883:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
884:       MatInfo    info;
885:       PetscTruth inodes;

887:       MPI_Comm_rank(((PetscObject)mat)->comm,&rank);
888:       MatGetInfo(mat,MAT_LOCAL,&info);
889:       MatInodeGetInodeSizes(aij->A,PETSC_NULL,(PetscInt **)&inodes,PETSC_NULL);
890:       if (!inodes) {
891:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
892:                                               rank,mat->rmap.n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
893:       } else {
894:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
895:                     rank,mat->rmap.n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
896:       }
897:       MatGetInfo(aij->A,MAT_LOCAL,&info);
898:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
899:       MatGetInfo(aij->B,MAT_LOCAL,&info);
900:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
901:       PetscViewerFlush(viewer);
902:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
903:       VecScatterView(aij->Mvctx,viewer);
904:       return(0);
905:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
906:       PetscInt   inodecount,inodelimit,*inodes;
907:       MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
908:       if (inodes) {
909:         PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
910:       } else {
911:         PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
912:       }
913:       return(0);
914:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
915:       return(0);
916:     }
917:   } else if (isbinary) {
918:     if (size == 1) {
919:       PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
920:       MatView(aij->A,viewer);
921:     } else {
922:       MatView_MPIAIJ_Binary(mat,viewer);
923:     }
924:     return(0);
925:   } else if (isdraw) {
926:     PetscDraw  draw;
927:     PetscTruth isnull;
928:     PetscViewerDrawGetDraw(viewer,0,&draw);
929:     PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
930:   }

932:   if (size == 1) {
933:     PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
934:     MatView(aij->A,viewer);
935:   } else {
936:     /* assemble the entire matrix onto first processor. */
937:     Mat         A;
938:     Mat_SeqAIJ  *Aloc;
939:     PetscInt    M = mat->rmap.N,N = mat->cmap.N,m,*ai,*aj,row,*cols,i,*ct;
940:     PetscScalar *a;

942:     MatCreate(((PetscObject)mat)->comm,&A);
943:     if (!rank) {
944:       MatSetSizes(A,M,N,M,N);
945:     } else {
946:       MatSetSizes(A,0,0,M,N);
947:     }
948:     /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
949:     MatSetType(A,MATMPIAIJ);
950:     MatMPIAIJSetPreallocation(A,0,PETSC_NULL,0,PETSC_NULL);
951:     PetscLogObjectParent(mat,A);

953:     /* copy over the A part */
954:     Aloc = (Mat_SeqAIJ*)aij->A->data;
955:     m = aij->A->rmap.n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
956:     row = mat->rmap.rstart;
957:     for (i=0; i<ai[m]; i++) {aj[i] += mat->cmap.rstart ;}
958:     for (i=0; i<m; i++) {
959:       MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
960:       row++; a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
961:     }
962:     aj = Aloc->j;
963:     for (i=0; i<ai[m]; i++) {aj[i] -= mat->cmap.rstart;}

965:     /* copy over the B part */
966:     Aloc = (Mat_SeqAIJ*)aij->B->data;
967:     m    = aij->B->rmap.n;  ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
968:     row  = mat->rmap.rstart;
969:     PetscMalloc((ai[m]+1)*sizeof(PetscInt),&cols);
970:     ct   = cols;
971:     for (i=0; i<ai[m]; i++) {cols[i] = aij->garray[aj[i]];}
972:     for (i=0; i<m; i++) {
973:       MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
974:       row++; a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
975:     }
976:     PetscFree(ct);
977:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
978:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
979:     /* 
980:        Everyone has to call to draw the matrix since the graphics waits are
981:        synchronized across all processors that share the PetscDraw object
982:     */
983:     PetscViewerGetSingleton(viewer,&sviewer);
984:     if (!rank) {
985:       PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);
986:       MatView(((Mat_MPIAIJ*)(A->data))->A,sviewer);
987:     }
988:     PetscViewerRestoreSingleton(viewer,&sviewer);
989:     MatDestroy(A);
990:   }
991:   return(0);
992: }

996: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
997: {
999:   PetscTruth     iascii,isdraw,issocket,isbinary;
1000: 
1002:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_ASCII,&iascii);
1003:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_DRAW,&isdraw);
1004:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_BINARY,&isbinary);
1005:   PetscTypeCompare((PetscObject)viewer,PETSC_VIEWER_SOCKET,&issocket);
1006:   if (iascii || isdraw || isbinary || issocket) {
1007:     MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1008:   } else {
1009:     SETERRQ1(PETSC_ERR_SUP,"Viewer type %s not supported by MPIAIJ matrices",((PetscObject)viewer)->type_name);
1010:   }
1011:   return(0);
1012: }

1016: PetscErrorCode MatRelax_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1017: {
1018:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1020:   Vec            bb1;

1023:   VecDuplicate(bb,&bb1);

1025:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP){
1026:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1027:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,lits,xx);
1028:       its--;
1029:     }
1030: 
1031:     while (its--) {
1032:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1033:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1035:       /* update rhs: bb1 = bb - B*x */
1036:       VecScale(mat->lvec,-1.0);
1037:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1039:       /* local sweep */
1040:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,lits,xx);
1041:     }
1042:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP){
1043:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1044:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1045:       its--;
1046:     }
1047:     while (its--) {
1048:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1049:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1051:       /* update rhs: bb1 = bb - B*x */
1052:       VecScale(mat->lvec,-1.0);
1053:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1055:       /* local sweep */
1056:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1057:     }
1058:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP){
1059:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1060:       (*mat->A->ops->relax)(mat->A,bb,omega,flag,fshift,lits,PETSC_NULL,xx);
1061:       its--;
1062:     }
1063:     while (its--) {
1064:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1065:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1067:       /* update rhs: bb1 = bb - B*x */
1068:       VecScale(mat->lvec,-1.0);
1069:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1071:       /* local sweep */
1072:       (*mat->A->ops->relax)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,PETSC_NULL,xx);
1073:     }
1074:   } else {
1075:     SETERRQ(PETSC_ERR_SUP,"Parallel SOR not supported");
1076:   }

1078:   VecDestroy(bb1);
1079:   return(0);
1080: }

1084: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1085: {
1086:   MPI_Comm       comm,pcomm;
1087:   PetscInt       first,local_size,nrows,*rows;
1088:   int            ntids;
1089:   IS             crowp,growp,irowp,lrowp,lcolp,icolp;

1093:   PetscObjectGetComm((PetscObject)A,&comm);
1094:   /* make a collective version of 'rowp' */
1095:   PetscObjectGetComm((PetscObject)rowp,&pcomm);
1096:   if (pcomm==comm) {
1097:     crowp = rowp;
1098:   } else {
1099:     ISGetSize(rowp,&nrows);
1100:     ISGetIndices(rowp,&rows);
1101:     ISCreateGeneral(comm,nrows,rows,&crowp);
1102:     ISRestoreIndices(rowp,&rows);
1103:   }
1104:   /* collect the global row permutation and invert it */
1105:   ISAllGather(crowp,&growp);
1106:   ISSetPermutation(growp);
1107:   if (pcomm!=comm) {
1108:     ISDestroy(crowp);
1109:   }
1110:   ISInvertPermutation(growp,PETSC_DECIDE,&irowp);
1111:   /* get the local target indices */
1112:   MatGetOwnershipRange(A,&first,PETSC_NULL);
1113:   MatGetLocalSize(A,&local_size,PETSC_NULL);
1114:   ISGetIndices(irowp,&rows);
1115:   ISCreateGeneral(MPI_COMM_SELF,local_size,rows+first,&lrowp);
1116:   ISRestoreIndices(irowp,&rows);
1117:   ISDestroy(irowp);
1118:   /* the column permutation is so much easier;
1119:      make a local version of 'colp' and invert it */
1120:   PetscObjectGetComm((PetscObject)colp,&pcomm);
1121:   MPI_Comm_size(pcomm,&ntids);
1122:   if (ntids==1) {
1123:     lcolp = colp;
1124:   } else {
1125:     ISGetSize(colp,&nrows);
1126:     ISGetIndices(colp,&rows);
1127:     ISCreateGeneral(MPI_COMM_SELF,nrows,rows,&lcolp);
1128:   }
1129:   ISInvertPermutation(lcolp,PETSC_DECIDE,&icolp);
1130:   ISSetPermutation(lcolp);
1131:   if (ntids>1) {
1132:     ISRestoreIndices(colp,&rows);
1133:     ISDestroy(lcolp);
1134:   }
1135:   /* now we just get the submatrix */
1136:   MatGetSubMatrix(A,lrowp,icolp,local_size,MAT_INITIAL_MATRIX,B);
1137:   /* clean up */
1138:   ISDestroy(lrowp);
1139:   ISDestroy(icolp);
1140:   return(0);
1141: }

1145: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1146: {
1147:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1148:   Mat            A = mat->A,B = mat->B;
1150:   PetscReal      isend[5],irecv[5];

1153:   info->block_size     = 1.0;
1154:   MatGetInfo(A,MAT_LOCAL,info);
1155:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1156:   isend[3] = info->memory;  isend[4] = info->mallocs;
1157:   MatGetInfo(B,MAT_LOCAL,info);
1158:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1159:   isend[3] += info->memory;  isend[4] += info->mallocs;
1160:   if (flag == MAT_LOCAL) {
1161:     info->nz_used      = isend[0];
1162:     info->nz_allocated = isend[1];
1163:     info->nz_unneeded  = isend[2];
1164:     info->memory       = isend[3];
1165:     info->mallocs      = isend[4];
1166:   } else if (flag == MAT_GLOBAL_MAX) {
1167:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_MAX,((PetscObject)matin)->comm);
1168:     info->nz_used      = irecv[0];
1169:     info->nz_allocated = irecv[1];
1170:     info->nz_unneeded  = irecv[2];
1171:     info->memory       = irecv[3];
1172:     info->mallocs      = irecv[4];
1173:   } else if (flag == MAT_GLOBAL_SUM) {
1174:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPI_SUM,((PetscObject)matin)->comm);
1175:     info->nz_used      = irecv[0];
1176:     info->nz_allocated = irecv[1];
1177:     info->nz_unneeded  = irecv[2];
1178:     info->memory       = irecv[3];
1179:     info->mallocs      = irecv[4];
1180:   }
1181:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1182:   info->fill_ratio_needed = 0;
1183:   info->factor_mallocs    = 0;
1184:   info->rows_global       = (double)matin->rmap.N;
1185:   info->columns_global    = (double)matin->cmap.N;
1186:   info->rows_local        = (double)matin->rmap.n;
1187:   info->columns_local     = (double)matin->cmap.N;

1189:   return(0);
1190: }

1194: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscTruth flg)
1195: {
1196:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1200:   switch (op) {
1201:   case MAT_NEW_NONZERO_LOCATIONS:
1202:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1203:   case MAT_KEEP_ZEROED_ROWS:
1204:   case MAT_NEW_NONZERO_LOCATION_ERR:
1205:   case MAT_USE_INODES:
1206:   case MAT_IGNORE_ZERO_ENTRIES:
1207:     MatSetOption(a->A,op,flg);
1208:     MatSetOption(a->B,op,flg);
1209:     break;
1210:   case MAT_ROW_ORIENTED:
1211:     a->roworiented = flg;
1212:     MatSetOption(a->A,op,flg);
1213:     MatSetOption(a->B,op,flg);
1214:     break;
1215:   case MAT_NEW_DIAGONALS:
1216:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1217:     break;
1218:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1219:     a->donotstash = PETSC_TRUE;
1220:     break;
1221:   case MAT_SYMMETRIC:
1222:     MatSetOption(a->A,op,flg);
1223:     break;
1224:   case MAT_STRUCTURALLY_SYMMETRIC:
1225:   case MAT_HERMITIAN:
1226:   case MAT_SYMMETRY_ETERNAL:
1227:     MatSetOption(a->A,op,flg);
1228:     break;
1229:   default:
1230:     SETERRQ1(PETSC_ERR_SUP,"unknown option %d",op);
1231:   }
1232:   return(0);
1233: }

1237: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1238: {
1239:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1240:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1242:   PetscInt       i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap.rstart;
1243:   PetscInt       nztot,nzA,nzB,lrow,rstart = matin->rmap.rstart,rend = matin->rmap.rend;
1244:   PetscInt       *cmap,*idx_p;

1247:   if (mat->getrowactive) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Already active");
1248:   mat->getrowactive = PETSC_TRUE;

1250:   if (!mat->rowvalues && (idx || v)) {
1251:     /*
1252:         allocate enough space to hold information from the longest row.
1253:     */
1254:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1255:     PetscInt     max = 1,tmp;
1256:     for (i=0; i<matin->rmap.n; i++) {
1257:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1258:       if (max < tmp) { max = tmp; }
1259:     }
1260:     PetscMalloc(max*(sizeof(PetscInt)+sizeof(PetscScalar)),&mat->rowvalues);
1261:     mat->rowindices = (PetscInt*)(mat->rowvalues + max);
1262:   }

1264:   if (row < rstart || row >= rend) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Only local rows")
1265:   lrow = row - rstart;

1267:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1268:   if (!v)   {pvA = 0; pvB = 0;}
1269:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1270:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1271:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1272:   nztot = nzA + nzB;

1274:   cmap  = mat->garray;
1275:   if (v  || idx) {
1276:     if (nztot) {
1277:       /* Sort by increasing column numbers, assuming A and B already sorted */
1278:       PetscInt imark = -1;
1279:       if (v) {
1280:         *v = v_p = mat->rowvalues;
1281:         for (i=0; i<nzB; i++) {
1282:           if (cmap[cworkB[i]] < cstart)   v_p[i] = vworkB[i];
1283:           else break;
1284:         }
1285:         imark = i;
1286:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1287:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1288:       }
1289:       if (idx) {
1290:         *idx = idx_p = mat->rowindices;
1291:         if (imark > -1) {
1292:           for (i=0; i<imark; i++) {
1293:             idx_p[i] = cmap[cworkB[i]];
1294:           }
1295:         } else {
1296:           for (i=0; i<nzB; i++) {
1297:             if (cmap[cworkB[i]] < cstart)   idx_p[i] = cmap[cworkB[i]];
1298:             else break;
1299:           }
1300:           imark = i;
1301:         }
1302:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart + cworkA[i];
1303:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]];
1304:       }
1305:     } else {
1306:       if (idx) *idx = 0;
1307:       if (v)   *v   = 0;
1308:     }
1309:   }
1310:   *nz = nztot;
1311:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1312:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1313:   return(0);
1314: }

1318: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1319: {
1320:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1323:   if (!aij->getrowactive) {
1324:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1325:   }
1326:   aij->getrowactive = PETSC_FALSE;
1327:   return(0);
1328: }

1332: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1333: {
1334:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1335:   Mat_SeqAIJ     *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1337:   PetscInt       i,j,cstart = mat->cmap.rstart;
1338:   PetscReal      sum = 0.0;
1339:   PetscScalar    *v;

1342:   if (aij->size == 1) {
1343:      MatNorm(aij->A,type,norm);
1344:   } else {
1345:     if (type == NORM_FROBENIUS) {
1346:       v = amat->a;
1347:       for (i=0; i<amat->nz; i++) {
1348: #if defined(PETSC_USE_COMPLEX)
1349:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1350: #else
1351:         sum += (*v)*(*v); v++;
1352: #endif
1353:       }
1354:       v = bmat->a;
1355:       for (i=0; i<bmat->nz; i++) {
1356: #if defined(PETSC_USE_COMPLEX)
1357:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1358: #else
1359:         sum += (*v)*(*v); v++;
1360: #endif
1361:       }
1362:       MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPI_SUM,((PetscObject)mat)->comm);
1363:       *norm = sqrt(*norm);
1364:     } else if (type == NORM_1) { /* max column norm */
1365:       PetscReal *tmp,*tmp2;
1366:       PetscInt    *jj,*garray = aij->garray;
1367:       PetscMalloc((mat->cmap.N+1)*sizeof(PetscReal),&tmp);
1368:       PetscMalloc((mat->cmap.N+1)*sizeof(PetscReal),&tmp2);
1369:       PetscMemzero(tmp,mat->cmap.N*sizeof(PetscReal));
1370:       *norm = 0.0;
1371:       v = amat->a; jj = amat->j;
1372:       for (j=0; j<amat->nz; j++) {
1373:         tmp[cstart + *jj++ ] += PetscAbsScalar(*v);  v++;
1374:       }
1375:       v = bmat->a; jj = bmat->j;
1376:       for (j=0; j<bmat->nz; j++) {
1377:         tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1378:       }
1379:       MPI_Allreduce(tmp,tmp2,mat->cmap.N,MPIU_REAL,MPI_SUM,((PetscObject)mat)->comm);
1380:       for (j=0; j<mat->cmap.N; j++) {
1381:         if (tmp2[j] > *norm) *norm = tmp2[j];
1382:       }
1383:       PetscFree(tmp);
1384:       PetscFree(tmp2);
1385:     } else if (type == NORM_INFINITY) { /* max row norm */
1386:       PetscReal ntemp = 0.0;
1387:       for (j=0; j<aij->A->rmap.n; j++) {
1388:         v = amat->a + amat->i[j];
1389:         sum = 0.0;
1390:         for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1391:           sum += PetscAbsScalar(*v); v++;
1392:         }
1393:         v = bmat->a + bmat->i[j];
1394:         for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1395:           sum += PetscAbsScalar(*v); v++;
1396:         }
1397:         if (sum > ntemp) ntemp = sum;
1398:       }
1399:       MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPI_MAX,((PetscObject)mat)->comm);
1400:     } else {
1401:       SETERRQ(PETSC_ERR_SUP,"No support for two norm");
1402:     }
1403:   }
1404:   return(0);
1405: }

1409: PetscErrorCode MatTranspose_MPIAIJ(Mat A,Mat *matout)
1410: {
1411:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1412:   Mat_SeqAIJ     *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data;
1414:   PetscInt       M = A->rmap.N,N = A->cmap.N,ma,na,mb,*ai,*aj,*bi,*bj,row,*cols,i,*d_nnz;
1415:   PetscInt       cstart=A->cmap.rstart,ncol;
1416:   Mat            B;
1417:   PetscScalar    *array;

1420:   if (!matout && M != N) SETERRQ(PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");

1422:   /* compute d_nnz for preallocation; o_nnz is approximated by d_nnz to avoid communication */
1423:   ma = A->rmap.n; na = A->cmap.n; mb = a->B->rmap.n;
1424:   ai = Aloc->i; aj = Aloc->j;
1425:   bi = Bloc->i; bj = Bloc->j;
1426:   PetscMalloc((1+na+bi[mb])*sizeof(PetscInt),&d_nnz);
1427:   cols = d_nnz + na + 1; /* work space to be used by B part */
1428:   PetscMemzero(d_nnz,(1+na)*sizeof(PetscInt));
1429:   for (i=0; i<ai[ma]; i++){
1430:     d_nnz[aj[i]] ++;
1431:     aj[i] += cstart; /* global col index to be used by MatSetValues() */
1432:   }

1434:   MatCreate(((PetscObject)A)->comm,&B);
1435:   MatSetSizes(B,A->cmap.n,A->rmap.n,N,M);
1436:   MatSetType(B,((PetscObject)A)->type_name);
1437:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,d_nnz);

1439:   /* copy over the A part */
1440:   array = Aloc->a;
1441:   row = A->rmap.rstart;
1442:   for (i=0; i<ma; i++) {
1443:     ncol = ai[i+1]-ai[i];
1444:     MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);
1445:     row++; array += ncol; aj += ncol;
1446:   }
1447:   aj = Aloc->j;
1448:   for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */

1450:   /* copy over the B part */
1451:   array = Bloc->a;
1452:   row = A->rmap.rstart;
1453:   for (i=0; i<bi[mb]; i++) {cols[i] = a->garray[bj[i]];}
1454:   for (i=0; i<mb; i++) {
1455:     ncol = bi[i+1]-bi[i];
1456:     MatSetValues(B,ncol,cols,1,&row,array,INSERT_VALUES);
1457:     row++; array += ncol; cols += ncol;
1458:   }
1459:   PetscFree(d_nnz);
1460:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1461:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
1462:   if (matout) {
1463:     *matout = B;
1464:   } else {
1465:     MatHeaderCopy(A,B);
1466:   }
1467:   return(0);
1468: }

1472: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
1473: {
1474:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1475:   Mat            a = aij->A,b = aij->B;
1477:   PetscInt       s1,s2,s3;

1480:   MatGetLocalSize(mat,&s2,&s3);
1481:   if (rr) {
1482:     VecGetLocalSize(rr,&s1);
1483:     if (s1!=s3) SETERRQ(PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1484:     /* Overlap communication with computation. */
1485:     VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1486:   }
1487:   if (ll) {
1488:     VecGetLocalSize(ll,&s1);
1489:     if (s1!=s2) SETERRQ(PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1490:     (*b->ops->diagonalscale)(b,ll,0);
1491:   }
1492:   /* scale  the diagonal block */
1493:   (*a->ops->diagonalscale)(a,ll,rr);

1495:   if (rr) {
1496:     /* Do a scatter end and then right scale the off-diagonal block */
1497:     VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1498:     (*b->ops->diagonalscale)(b,0,aij->lvec);
1499:   }
1500: 
1501:   return(0);
1502: }

1506: PetscErrorCode MatSetBlockSize_MPIAIJ(Mat A,PetscInt bs)
1507: {
1508:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

1512:   MatSetBlockSize(a->A,bs);
1513:   MatSetBlockSize(a->B,bs);
1514:   return(0);
1515: }
1518: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
1519: {
1520:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;

1524:   MatSetUnfactored(a->A);
1525:   return(0);
1526: }

1530: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscTruth *flag)
1531: {
1532:   Mat_MPIAIJ     *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
1533:   Mat            a,b,c,d;
1534:   PetscTruth     flg;

1538:   a = matA->A; b = matA->B;
1539:   c = matB->A; d = matB->B;

1541:   MatEqual(a,c,&flg);
1542:   if (flg) {
1543:     MatEqual(b,d,&flg);
1544:   }
1545:   MPI_Allreduce(&flg,flag,1,MPI_INT,MPI_LAND,((PetscObject)A)->comm);
1546:   return(0);
1547: }

1551: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
1552: {
1554:   Mat_MPIAIJ     *a = (Mat_MPIAIJ *)A->data;
1555:   Mat_MPIAIJ     *b = (Mat_MPIAIJ *)B->data;

1558:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1559:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1560:     /* because of the column compression in the off-processor part of the matrix a->B,
1561:        the number of columns in a->B and b->B may be different, hence we cannot call
1562:        the MatCopy() directly on the two parts. If need be, we can provide a more 
1563:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
1564:        then copying the submatrices */
1565:     MatCopy_Basic(A,B,str);
1566:   } else {
1567:     MatCopy(a->A,b->A,str);
1568:     MatCopy(a->B,b->B,str);
1569:   }
1570:   return(0);
1571: }

1575: PetscErrorCode MatSetUpPreallocation_MPIAIJ(Mat A)
1576: {

1580:    MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1581:   return(0);
1582: }

1584:  #include petscblaslapack.h
1587: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1588: {
1590:   PetscInt       i;
1591:   Mat_MPIAIJ     *xx = (Mat_MPIAIJ *)X->data,*yy = (Mat_MPIAIJ *)Y->data;
1592:   PetscBLASInt   bnz,one=1;
1593:   Mat_SeqAIJ     *x,*y;

1596:   if (str == SAME_NONZERO_PATTERN) {
1597:     PetscScalar alpha = a;
1598:     x = (Mat_SeqAIJ *)xx->A->data;
1599:     y = (Mat_SeqAIJ *)yy->A->data;
1600:     bnz = (PetscBLASInt)x->nz;
1601:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1602:     x = (Mat_SeqAIJ *)xx->B->data;
1603:     y = (Mat_SeqAIJ *)yy->B->data;
1604:     bnz = (PetscBLASInt)x->nz;
1605:     BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one);
1606:   } else if (str == SUBSET_NONZERO_PATTERN) {
1607:     MatAXPY_SeqAIJ(yy->A,a,xx->A,str);

1609:     x = (Mat_SeqAIJ *)xx->B->data;
1610:     y = (Mat_SeqAIJ *)yy->B->data;
1611:     if (y->xtoy && y->XtoY != xx->B) {
1612:       PetscFree(y->xtoy);
1613:       MatDestroy(y->XtoY);
1614:     }
1615:     if (!y->xtoy) { /* get xtoy */
1616:       MatAXPYGetxtoy_Private(xx->B->rmap.n,x->i,x->j,xx->garray,y->i,y->j,yy->garray,&y->xtoy);
1617:       y->XtoY = xx->B;
1618:       PetscObjectReference((PetscObject)xx->B);
1619:     }
1620:     for (i=0; i<x->nz; i++) y->a[y->xtoy[i]] += a*(x->a[i]);
1621:   } else {
1622:     MatAXPY_Basic(Y,a,X,str);
1623:   }
1624:   return(0);
1625: }

1627: EXTERN PetscErrorCode  MatConjugate_SeqAIJ(Mat);

1631: PetscErrorCode  MatConjugate_MPIAIJ(Mat mat)
1632: {
1633: #if defined(PETSC_USE_COMPLEX)
1635:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

1638:   MatConjugate_SeqAIJ(aij->A);
1639:   MatConjugate_SeqAIJ(aij->B);
1640: #else
1642: #endif
1643:   return(0);
1644: }

1648: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
1649: {
1650:   Mat_MPIAIJ   *a = (Mat_MPIAIJ*)A->data;

1654:   MatRealPart(a->A);
1655:   MatRealPart(a->B);
1656:   return(0);
1657: }

1661: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
1662: {
1663:   Mat_MPIAIJ   *a = (Mat_MPIAIJ*)A->data;

1667:   MatImaginaryPart(a->A);
1668:   MatImaginaryPart(a->B);
1669:   return(0);
1670: }

1672: #ifdef PETSC_HAVE_PBGL

1674: #include <boost/parallel/mpi/bsp_process_group.hpp>
1675: #include <boost/graph/distributed/ilu_default_graph.hpp>
1676: #include <boost/graph/distributed/ilu_0_block.hpp>
1677: #include <boost/graph/distributed/ilu_preconditioner.hpp>
1678: #include <boost/graph/distributed/petsc/interface.hpp>
1679: #include <boost/multi_array.hpp>
1680: #include <boost/parallel/distributed_property_map.hpp>

1684: /*
1685:   This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
1686: */
1687: PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat A, IS isrow, IS iscol, MatFactorInfo *info, Mat *fact)
1688: {
1689:   namespace petsc = boost::distributed::petsc;
1690: 
1691:   namespace graph_dist = boost::graph::distributed;
1692:   using boost::graph::distributed::ilu_default::process_group_type;
1693:   using boost::graph::ilu_permuted;

1695:   PetscTruth      row_identity, col_identity;
1696:   PetscContainer  c;
1697:   PetscInt        m, n, M, N;
1698:   PetscErrorCode  ierr;

1701:   if (info->levels != 0) SETERRQ(PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu");
1702:   ISIdentity(isrow, &row_identity);
1703:   ISIdentity(iscol, &col_identity);
1704:   if (!row_identity || !col_identity) {
1705:     SETERRQ(PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU");
1706:   }

1708:   process_group_type pg;
1709:   typedef graph_dist::ilu_default::ilu_level_graph_type  lgraph_type;
1710:   lgraph_type*   lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg));
1711:   lgraph_type&   level_graph = *lgraph_p;
1712:   graph_dist::ilu_default::graph_type&            graph(level_graph.graph);

1714:   petsc::read_matrix(A, graph, get(boost::edge_weight, graph));
1715:   ilu_permuted(level_graph);

1717:   /* put together the new matrix */
1718:   MatCreate(((PetscObject)A)->comm, fact);
1719:   MatGetLocalSize(A, &m, &n);
1720:   MatGetSize(A, &M, &N);
1721:   MatSetSizes(*fact, m, n, M, N);
1722:   MatSetType(*fact, ((PetscObject)A)->type_name);
1723:   MatAssemblyBegin(*fact, MAT_FINAL_ASSEMBLY);
1724:   MatAssemblyEnd(*fact, MAT_FINAL_ASSEMBLY);
1725:   (*fact)->factor = FACTOR_LU;

1727:   PetscContainerCreate(((PetscObject)A)->comm, &c);
1728:   PetscContainerSetPointer(c, lgraph_p);
1729:   PetscObjectCompose((PetscObject) (*fact), "graph", (PetscObject) c);
1730:   return(0);
1731: }

1735: PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat A, MatFactorInfo *info, Mat *B)
1736: {
1738:   return(0);
1739: }

1743: /*
1744:   This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu>
1745: */
1746: PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x)
1747: {
1748:   namespace graph_dist = boost::graph::distributed;

1750:   typedef graph_dist::ilu_default::ilu_level_graph_type  lgraph_type;
1751:   lgraph_type*   lgraph_p;
1752:   PetscContainer c;

1756:   PetscObjectQuery((PetscObject) A, "graph", (PetscObject *) &c);
1757:   PetscContainerGetPointer(c, (void **) &lgraph_p);
1758:   VecCopy(b, x);

1760:   PetscScalar* array_x;
1761:   VecGetArray(x, &array_x);
1762:   PetscInt sx;
1763:   VecGetSize(x, &sx);
1764: 
1765:   PetscScalar* array_b;
1766:   VecGetArray(b, &array_b);
1767:   PetscInt sb;
1768:   VecGetSize(b, &sb);

1770:   lgraph_type&   level_graph = *lgraph_p;
1771:   graph_dist::ilu_default::graph_type&            graph(level_graph.graph);

1773:   typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type;
1774:   array_ref_type                                 ref_b(array_b, boost::extents[num_vertices(graph)]),
1775:                                                  ref_x(array_x, boost::extents[num_vertices(graph)]);

1777:   typedef boost::iterator_property_map<array_ref_type::iterator,
1778:                                 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type>  gvector_type;
1779:   gvector_type                                   vector_b(ref_b.begin(), get(boost::vertex_index, graph)),
1780:                                                  vector_x(ref_x.begin(), get(boost::vertex_index, graph));
1781: 
1782:   ilu_set_solve(*lgraph_p, vector_b, vector_x);

1784:   return(0);
1785: }
1786: #endif

1788: typedef struct { /* used by MatGetRedundantMatrix() for reusing matredundant */
1789:   PetscInt       nzlocal,nsends,nrecvs;
1790:   PetscMPIInt    *send_rank;
1791:   PetscInt       *sbuf_nz,*sbuf_j,**rbuf_j;
1792:   PetscScalar    *sbuf_a,**rbuf_a;
1793:   PetscErrorCode (*MatDestroy)(Mat);
1794: } Mat_Redundant;

1798: PetscErrorCode PetscContainerDestroy_MatRedundant(void *ptr)
1799: {
1800:   PetscErrorCode       ierr;
1801:   Mat_Redundant        *redund=(Mat_Redundant*)ptr;
1802:   PetscInt             i;

1805:   PetscFree(redund->send_rank);
1806:   PetscFree(redund->sbuf_j);
1807:   PetscFree(redund->sbuf_a);
1808:   for (i=0; i<redund->nrecvs; i++){
1809:     PetscFree(redund->rbuf_j[i]);
1810:     PetscFree(redund->rbuf_a[i]);
1811:   }
1812:   PetscFree3(redund->sbuf_nz,redund->rbuf_j,redund->rbuf_a);
1813:   PetscFree(redund);
1814:   return(0);
1815: }

1819: PetscErrorCode MatDestroy_MatRedundant(Mat A)
1820: {
1821:   PetscErrorCode  ierr;
1822:   PetscContainer  container;
1823:   Mat_Redundant   *redund=PETSC_NULL;

1826:   PetscObjectQuery((PetscObject)A,"Mat_Redundant",(PetscObject *)&container);
1827:   if (container) {
1828:     PetscContainerGetPointer(container,(void **)&redund);
1829:   } else {
1830:     SETERRQ(PETSC_ERR_PLIB,"Container does not exit");
1831:   }
1832:   A->ops->destroy = redund->MatDestroy;
1833:   PetscObjectCompose((PetscObject)A,"Mat_Redundant",0);
1834:   (*A->ops->destroy)(A);
1835:   PetscContainerDestroy(container);
1836:   return(0);
1837: }

1841: PetscErrorCode MatGetRedundantMatrix_MPIAIJ(Mat mat,PetscInt nsubcomm,MPI_Comm subcomm,PetscInt mlocal_sub,MatReuse reuse,Mat *matredundant)
1842: {
1843:   PetscMPIInt    rank,size;
1844:   MPI_Comm       comm=((PetscObject)mat)->comm;
1846:   PetscInt       nsends=0,nrecvs=0,i,rownz_max=0;
1847:   PetscMPIInt    *send_rank=PETSC_NULL,*recv_rank=PETSC_NULL;
1848:   PetscInt       *rowrange=mat->rmap.range;
1849:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1850:   Mat            A=aij->A,B=aij->B,C=*matredundant;
1851:   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data;
1852:   PetscScalar    *sbuf_a;
1853:   PetscInt       nzlocal=a->nz+b->nz;
1854:   PetscInt       j,cstart=mat->cmap.rstart,cend=mat->cmap.rend,row,nzA,nzB,ncols,*cworkA,*cworkB;
1855:   PetscInt       rstart=mat->rmap.rstart,rend=mat->rmap.rend,*bmap=aij->garray,M,N;
1856:   PetscInt       *cols,ctmp,lwrite,*rptr,l,*sbuf_j;
1857:   PetscScalar    *vals,*aworkA,*aworkB;
1858:   PetscMPIInt    tag1,tag2,tag3,imdex;
1859:   MPI_Request    *s_waits1=PETSC_NULL,*s_waits2=PETSC_NULL,*s_waits3=PETSC_NULL,
1860:                  *r_waits1=PETSC_NULL,*r_waits2=PETSC_NULL,*r_waits3=PETSC_NULL;
1861:   MPI_Status     recv_status,*send_status;
1862:   PetscInt       *sbuf_nz=PETSC_NULL,*rbuf_nz=PETSC_NULL,count;
1863:   PetscInt       **rbuf_j=PETSC_NULL;
1864:   PetscScalar    **rbuf_a=PETSC_NULL;
1865:   Mat_Redundant  *redund=PETSC_NULL;
1866:   PetscContainer container;

1869:   MPI_Comm_rank(comm,&rank);
1870:   MPI_Comm_size(comm,&size);

1872:   if (reuse == MAT_REUSE_MATRIX) {
1873:     MatGetSize(C,&M,&N);
1874:     if (M != N || M != mat->rmap.N) SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong global size");
1875:     MatGetLocalSize(C,&M,&N);
1876:     if (M != N || M != mlocal_sub) SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong local size");
1877:     PetscObjectQuery((PetscObject)C,"Mat_Redundant",(PetscObject *)&container);
1878:     if (container) {
1879:       PetscContainerGetPointer(container,(void **)&redund);
1880:     } else {
1881:       SETERRQ(PETSC_ERR_PLIB,"Container does not exit");
1882:     }
1883:     if (nzlocal != redund->nzlocal) SETERRQ(PETSC_ERR_ARG_SIZ,"Cannot reuse matrix. Wrong nzlocal");

1885:     nsends    = redund->nsends;
1886:     nrecvs    = redund->nrecvs;
1887:     send_rank = redund->send_rank; recv_rank = send_rank + size;
1888:     sbuf_nz   = redund->sbuf_nz;     rbuf_nz = sbuf_nz + nsends;
1889:     sbuf_j    = redund->sbuf_j;
1890:     sbuf_a    = redund->sbuf_a;
1891:     rbuf_j    = redund->rbuf_j;
1892:     rbuf_a    = redund->rbuf_a;
1893:   }

1895:   if (reuse == MAT_INITIAL_MATRIX){
1896:     PetscMPIInt  subrank,subsize;
1897:     PetscInt     nleftover,np_subcomm;
1898:     /* get the destination processors' id send_rank, nsends and nrecvs */
1899:     MPI_Comm_rank(subcomm,&subrank);
1900:     MPI_Comm_size(subcomm,&subsize);
1901:     PetscMalloc((2*size+1)*sizeof(PetscMPIInt),&send_rank);
1902:     recv_rank = send_rank + size;
1903:     np_subcomm = size/nsubcomm;
1904:     nleftover  = size - nsubcomm*np_subcomm;
1905:     nsends = 0; nrecvs = 0;
1906:     for (i=0; i<size; i++){ /* i=rank*/
1907:       if (subrank == i/nsubcomm && rank != i){ /* my_subrank == other's subrank */
1908:         send_rank[nsends] = i; nsends++;
1909:         recv_rank[nrecvs++] = i;
1910:       }
1911:     }
1912:     if (rank >= size - nleftover){/* this proc is a leftover processor */
1913:       i = size-nleftover-1;
1914:       j = 0;
1915:       while (j < nsubcomm - nleftover){
1916:         send_rank[nsends++] = i;
1917:         i--; j++;
1918:       }
1919:     }

1921:     if (nleftover && subsize == size/nsubcomm && subrank==subsize-1){ /* this proc recvs from leftover processors */
1922:       for (i=0; i<nleftover; i++){
1923:         recv_rank[nrecvs++] = size-nleftover+i;
1924:       }
1925:     }

1927:     /* allocate sbuf_j, sbuf_a */
1928:     i = nzlocal + rowrange[rank+1] - rowrange[rank] + 2;
1929:     PetscMalloc(i*sizeof(PetscInt),&sbuf_j);
1930:     PetscMalloc((nzlocal+1)*sizeof(PetscScalar),&sbuf_a);
1931:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */
1932: 
1933:   /* copy mat's local entries into the buffers */
1934:   if (reuse == MAT_INITIAL_MATRIX){
1935:     rownz_max = 0;
1936:     rptr = sbuf_j;
1937:     cols = sbuf_j + rend-rstart + 1;
1938:     vals = sbuf_a;
1939:     rptr[0] = 0;
1940:     for (i=0; i<rend-rstart; i++){
1941:       row = i + rstart;
1942:       nzA    = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
1943:       ncols  = nzA + nzB;
1944:       cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
1945:       aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
1946:       /* load the column indices for this row into cols */
1947:       lwrite = 0;
1948:       for (l=0; l<nzB; l++) {
1949:         if ((ctmp = bmap[cworkB[l]]) < cstart){
1950:           vals[lwrite]   = aworkB[l];
1951:           cols[lwrite++] = ctmp;
1952:         }
1953:       }
1954:       for (l=0; l<nzA; l++){
1955:         vals[lwrite]   = aworkA[l];
1956:         cols[lwrite++] = cstart + cworkA[l];
1957:       }
1958:       for (l=0; l<nzB; l++) {
1959:         if ((ctmp = bmap[cworkB[l]]) >= cend){
1960:           vals[lwrite]   = aworkB[l];
1961:           cols[lwrite++] = ctmp;
1962:         }
1963:       }
1964:       vals += ncols;
1965:       cols += ncols;
1966:       rptr[i+1] = rptr[i] + ncols;
1967:       if (rownz_max < ncols) rownz_max = ncols;
1968:     }
1969:     if (rptr[rend-rstart] != a->nz + b->nz) SETERRQ4(1, "rptr[%d] %d != %d + %d",rend-rstart,rptr[rend-rstart+1],a->nz,b->nz);
1970:   } else { /* only copy matrix values into sbuf_a */
1971:     rptr = sbuf_j;
1972:     vals = sbuf_a;
1973:     rptr[0] = 0;
1974:     for (i=0; i<rend-rstart; i++){
1975:       row = i + rstart;
1976:       nzA    = a->i[i+1] - a->i[i]; nzB = b->i[i+1] - b->i[i];
1977:       ncols  = nzA + nzB;
1978:       cworkA = a->j + a->i[i]; cworkB = b->j + b->i[i];
1979:       aworkA = a->a + a->i[i]; aworkB = b->a + b->i[i];
1980:       lwrite = 0;
1981:       for (l=0; l<nzB; l++) {
1982:         if ((ctmp = bmap[cworkB[l]]) < cstart) vals[lwrite++] = aworkB[l];
1983:       }
1984:       for (l=0; l<nzA; l++) vals[lwrite++] = aworkA[l];
1985:       for (l=0; l<nzB; l++) {
1986:         if ((ctmp = bmap[cworkB[l]]) >= cend) vals[lwrite++] = aworkB[l];
1987:       }
1988:       vals += ncols;
1989:       rptr[i+1] = rptr[i] + ncols;
1990:     }
1991:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */

1993:   /* send nzlocal to others, and recv other's nzlocal */
1994:   /*--------------------------------------------------*/
1995:   if (reuse == MAT_INITIAL_MATRIX){
1996:     PetscMalloc2(3*(nsends + nrecvs)+1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
1997:     s_waits2 = s_waits3 + nsends;
1998:     s_waits1 = s_waits2 + nsends;
1999:     r_waits1 = s_waits1 + nsends;
2000:     r_waits2 = r_waits1 + nrecvs;
2001:     r_waits3 = r_waits2 + nrecvs;
2002:   } else {
2003:     PetscMalloc2(nsends + nrecvs +1,MPI_Request,&s_waits3,nsends+1,MPI_Status,&send_status);
2004:     r_waits3 = s_waits3 + nsends;
2005:   }

2007:   PetscObjectGetNewTag((PetscObject)mat,&tag3);
2008:   if (reuse == MAT_INITIAL_MATRIX){
2009:     /* get new tags to keep the communication clean */
2010:     PetscObjectGetNewTag((PetscObject)mat,&tag1);
2011:     PetscObjectGetNewTag((PetscObject)mat,&tag2);
2012:     PetscMalloc3(nsends+nrecvs+1,PetscInt,&sbuf_nz,nrecvs,PetscInt*,&rbuf_j,nrecvs,PetscScalar*,&rbuf_a);
2013:     rbuf_nz = sbuf_nz + nsends;
2014: 
2015:     /* post receives of other's nzlocal */
2016:     for (i=0; i<nrecvs; i++){
2017:       MPI_Irecv(rbuf_nz+i,1,MPIU_INT,MPI_ANY_SOURCE,tag1,comm,r_waits1+i);
2018:     }
2019:     /* send nzlocal to others */
2020:     for (i=0; i<nsends; i++){
2021:       sbuf_nz[i] = nzlocal;
2022:       MPI_Isend(sbuf_nz+i,1,MPIU_INT,send_rank[i],tag1,comm,s_waits1+i);
2023:     }
2024:     /* wait on receives of nzlocal; allocate space for rbuf_j, rbuf_a */
2025:     count = nrecvs;
2026:     while (count) {
2027:       MPI_Waitany(nrecvs,r_waits1,&imdex,&recv_status);
2028:       recv_rank[imdex] = recv_status.MPI_SOURCE;
2029:       /* allocate rbuf_a and rbuf_j; then post receives of rbuf_j */
2030:       PetscMalloc((rbuf_nz[imdex]+1)*sizeof(PetscScalar),&rbuf_a[imdex]);

2032:       i = rowrange[recv_status.MPI_SOURCE+1] - rowrange[recv_status.MPI_SOURCE]; /* number of expected mat->i */
2033:       rbuf_nz[imdex] += i + 2;
2034:       PetscMalloc(rbuf_nz[imdex]*sizeof(PetscInt),&rbuf_j[imdex]);
2035:       MPI_Irecv(rbuf_j[imdex],rbuf_nz[imdex],MPIU_INT,recv_status.MPI_SOURCE,tag2,comm,r_waits2+imdex);
2036:       count--;
2037:     }
2038:     /* wait on sends of nzlocal */
2039:     if (nsends) {MPI_Waitall(nsends,s_waits1,send_status);}
2040:     /* send mat->i,j to others, and recv from other's */
2041:     /*------------------------------------------------*/
2042:     for (i=0; i<nsends; i++){
2043:       j = nzlocal + rowrange[rank+1] - rowrange[rank] + 1;
2044:       MPI_Isend(sbuf_j,j,MPIU_INT,send_rank[i],tag2,comm,s_waits2+i);
2045:     }
2046:     /* wait on receives of mat->i,j */
2047:     /*------------------------------*/
2048:     count = nrecvs;
2049:     while (count) {
2050:       MPI_Waitany(nrecvs,r_waits2,&imdex,&recv_status);
2051:       if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2052:       count--;
2053:     }
2054:     /* wait on sends of mat->i,j */
2055:     /*---------------------------*/
2056:     if (nsends) {
2057:       MPI_Waitall(nsends,s_waits2,send_status);
2058:     }
2059:   } /* endof if (reuse == MAT_INITIAL_MATRIX) */

2061:   /* post receives, send and receive mat->a */
2062:   /*----------------------------------------*/
2063:   for (imdex=0; imdex<nrecvs; imdex++) {
2064:     MPI_Irecv(rbuf_a[imdex],rbuf_nz[imdex],MPIU_SCALAR,recv_rank[imdex],tag3,comm,r_waits3+imdex);
2065:   }
2066:   for (i=0; i<nsends; i++){
2067:     MPI_Isend(sbuf_a,nzlocal,MPIU_SCALAR,send_rank[i],tag3,comm,s_waits3+i);
2068:   }
2069:   count = nrecvs;
2070:   while (count) {
2071:     MPI_Waitany(nrecvs,r_waits3,&imdex,&recv_status);
2072:     if (recv_rank[imdex] != recv_status.MPI_SOURCE) SETERRQ2(1, "recv_rank %d != MPI_SOURCE %d",recv_rank[imdex],recv_status.MPI_SOURCE);
2073:     count--;
2074:   }
2075:   if (nsends) {
2076:     MPI_Waitall(nsends,s_waits3,send_status);
2077:   }

2079:   PetscFree2(s_waits3,send_status);
2080: 
2081:   /* create redundant matrix */
2082:   /*-------------------------*/
2083:   if (reuse == MAT_INITIAL_MATRIX){
2084:     /* compute rownz_max for preallocation */
2085:     for (imdex=0; imdex<nrecvs; imdex++){
2086:       j = rowrange[recv_rank[imdex]+1] - rowrange[recv_rank[imdex]];
2087:       rptr = rbuf_j[imdex];
2088:       for (i=0; i<j; i++){
2089:         ncols = rptr[i+1] - rptr[i];
2090:         if (rownz_max < ncols) rownz_max = ncols;
2091:       }
2092:     }
2093: 
2094:     MatCreate(subcomm,&C);
2095:     MatSetSizes(C,mlocal_sub,mlocal_sub,PETSC_DECIDE,PETSC_DECIDE);
2096:     MatSetFromOptions(C);
2097:     MatSeqAIJSetPreallocation(C,rownz_max,PETSC_NULL);
2098:     MatMPIAIJSetPreallocation(C,rownz_max,PETSC_NULL,rownz_max,PETSC_NULL);
2099:   } else {
2100:     C = *matredundant;
2101:   }

2103:   /* insert local matrix entries */
2104:   rptr = sbuf_j;
2105:   cols = sbuf_j + rend-rstart + 1;
2106:   vals = sbuf_a;
2107:   for (i=0; i<rend-rstart; i++){
2108:     row   = i + rstart;
2109:     ncols = rptr[i+1] - rptr[i];
2110:     MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2111:     vals += ncols;
2112:     cols += ncols;
2113:   }
2114:   /* insert received matrix entries */
2115:   for (imdex=0; imdex<nrecvs; imdex++){
2116:     rstart = rowrange[recv_rank[imdex]];
2117:     rend   = rowrange[recv_rank[imdex]+1];
2118:     rptr = rbuf_j[imdex];
2119:     cols = rbuf_j[imdex] + rend-rstart + 1;
2120:     vals = rbuf_a[imdex];
2121:     for (i=0; i<rend-rstart; i++){
2122:       row   = i + rstart;
2123:       ncols = rptr[i+1] - rptr[i];
2124:       MatSetValues(C,1,&row,ncols,cols,vals,INSERT_VALUES);
2125:       vals += ncols;
2126:       cols += ncols;
2127:     }
2128:   }
2129:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
2130:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
2131:   MatGetSize(C,&M,&N);
2132:   if (M != mat->rmap.N || N != mat->cmap.N) SETERRQ2(PETSC_ERR_ARG_INCOMP,"redundant mat size %d != input mat size %d",M,mat->rmap.N);
2133:   if (reuse == MAT_INITIAL_MATRIX){
2134:     PetscContainer container;
2135:     *matredundant = C;
2136:     /* create a supporting struct and attach it to C for reuse */
2137:     PetscNewLog(C,Mat_Redundant,&redund);
2138:     PetscContainerCreate(PETSC_COMM_SELF,&container);
2139:     PetscContainerSetPointer(container,redund);
2140:     PetscObjectCompose((PetscObject)C,"Mat_Redundant",(PetscObject)container);
2141:     PetscContainerSetUserDestroy(container,PetscContainerDestroy_MatRedundant);
2142: 
2143:     redund->nzlocal = nzlocal;
2144:     redund->nsends  = nsends;
2145:     redund->nrecvs  = nrecvs;
2146:     redund->send_rank = send_rank;
2147:     redund->sbuf_nz = sbuf_nz;
2148:     redund->sbuf_j  = sbuf_j;
2149:     redund->sbuf_a  = sbuf_a;
2150:     redund->rbuf_j  = rbuf_j;
2151:     redund->rbuf_a  = rbuf_a;

2153:     redund->MatDestroy = C->ops->destroy;
2154:     C->ops->destroy    = MatDestroy_MatRedundant;
2155:   }
2156:   return(0);
2157: }

2161: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2162: {
2163:   Mat_MPIAIJ    *mat    = (Mat_MPIAIJ *) A->data;
2164:   PetscInt       n      = A->rmap.n;
2165:   PetscInt       cstart = A->cmap.rstart;
2166:   PetscInt      *cmap   = mat->garray;
2167:   PetscInt      *diagIdx, *offdiagIdx;
2168:   Vec            diagV, offdiagV;
2169:   PetscScalar   *a, *diagA, *offdiagA;
2170:   PetscInt       r;

2174:   PetscMalloc2(n,PetscInt,&diagIdx,n,PetscInt,&offdiagIdx);
2175:   VecCreateSeq(((PetscObject)A)->comm, n, &diagV);
2176:   VecCreateSeq(((PetscObject)A)->comm, n, &offdiagV);
2177:   MatGetRowMin(mat->A, diagV,    diagIdx);
2178:   MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2179:   VecGetArray(v,        &a);
2180:   VecGetArray(diagV,    &diagA);
2181:   VecGetArray(offdiagV, &offdiagA);
2182:   for(r = 0; r < n; ++r) {
2183:     if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2184:       a[r]   = diagA[r];
2185:       idx[r] = cstart + diagIdx[r];
2186:     } else {
2187:       a[r]   = offdiagA[r];
2188:       idx[r] = cmap[offdiagIdx[r]];
2189:     }
2190:   }
2191:   VecRestoreArray(v,        &a);
2192:   VecRestoreArray(diagV,    &diagA);
2193:   VecRestoreArray(offdiagV, &offdiagA);
2194:   VecDestroy(diagV);
2195:   VecDestroy(offdiagV);
2196:   PetscFree2(diagIdx, offdiagIdx);
2197:   return(0);
2198: }

2202: PetscErrorCode MatGetSeqNonzerostructure_MPIAIJ(Mat mat,Mat *newmat[])
2203: {

2207:   MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,newmat);
2208:   return(0);
2209: }

2211: /* -------------------------------------------------------------------*/
2212: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2213:        MatGetRow_MPIAIJ,
2214:        MatRestoreRow_MPIAIJ,
2215:        MatMult_MPIAIJ,
2216: /* 4*/ MatMultAdd_MPIAIJ,
2217:        MatMultTranspose_MPIAIJ,
2218:        MatMultTransposeAdd_MPIAIJ,
2219: #ifdef PETSC_HAVE_PBGL
2220:        MatSolve_MPIAIJ,
2221: #else
2222:        0,
2223: #endif
2224:        0,
2225:        0,
2226: /*10*/ 0,
2227:        0,
2228:        0,
2229:        MatRelax_MPIAIJ,
2230:        MatTranspose_MPIAIJ,
2231: /*15*/ MatGetInfo_MPIAIJ,
2232:        MatEqual_MPIAIJ,
2233:        MatGetDiagonal_MPIAIJ,
2234:        MatDiagonalScale_MPIAIJ,
2235:        MatNorm_MPIAIJ,
2236: /*20*/ MatAssemblyBegin_MPIAIJ,
2237:        MatAssemblyEnd_MPIAIJ,
2238:        0,
2239:        MatSetOption_MPIAIJ,
2240:        MatZeroEntries_MPIAIJ,
2241: /*25*/ MatZeroRows_MPIAIJ,
2242:        0,
2243: #ifdef PETSC_HAVE_PBGL
2244:        MatLUFactorNumeric_MPIAIJ,
2245: #else
2246:        0,
2247: #endif
2248:        0,
2249:        0,
2250: /*30*/ MatSetUpPreallocation_MPIAIJ,
2251: #ifdef PETSC_HAVE_PBGL
2252:        MatILUFactorSymbolic_MPIAIJ,
2253: #else
2254:        0,
2255: #endif
2256:        0,
2257:        0,
2258:        0,
2259: /*35*/ MatDuplicate_MPIAIJ,
2260:        0,
2261:        0,
2262:        0,
2263:        0,
2264: /*40*/ MatAXPY_MPIAIJ,
2265:        MatGetSubMatrices_MPIAIJ,
2266:        MatIncreaseOverlap_MPIAIJ,
2267:        MatGetValues_MPIAIJ,
2268:        MatCopy_MPIAIJ,
2269: /*45*/ 0,
2270:        MatScale_MPIAIJ,
2271:        0,
2272:        0,
2273:        0,
2274: /*50*/ MatSetBlockSize_MPIAIJ,
2275:        0,
2276:        0,
2277:        0,
2278:        0,
2279: /*55*/ MatFDColoringCreate_MPIAIJ,
2280:        0,
2281:        MatSetUnfactored_MPIAIJ,
2282:        MatPermute_MPIAIJ,
2283:        0,
2284: /*60*/ MatGetSubMatrix_MPIAIJ,
2285:        MatDestroy_MPIAIJ,
2286:        MatView_MPIAIJ,
2287:        0,
2288:        0,
2289: /*65*/ 0,
2290:        0,
2291:        0,
2292:        0,
2293:        0,
2294: /*70*/ 0,
2295:        0,
2296:        MatSetColoring_MPIAIJ,
2297: #if defined(PETSC_HAVE_ADIC)
2298:        MatSetValuesAdic_MPIAIJ,
2299: #else
2300:        0,
2301: #endif
2302:        MatSetValuesAdifor_MPIAIJ,
2303: /*75*/ 0,
2304:        0,
2305:        0,
2306:        0,
2307:        0,
2308: /*80*/ 0,
2309:        0,
2310:        0,
2311:        0,
2312: /*84*/ MatLoad_MPIAIJ,
2313:        0,
2314:        0,
2315:        0,
2316:        0,
2317:        0,
2318: /*90*/ MatMatMult_MPIAIJ_MPIAIJ,
2319:        MatMatMultSymbolic_MPIAIJ_MPIAIJ,
2320:        MatMatMultNumeric_MPIAIJ_MPIAIJ,
2321:        MatPtAP_Basic,
2322:        MatPtAPSymbolic_MPIAIJ,
2323: /*95*/ MatPtAPNumeric_MPIAIJ,
2324:        0,
2325:        0,
2326:        0,
2327:        0,
2328: /*100*/0,
2329:        MatPtAPSymbolic_MPIAIJ_MPIAIJ,
2330:        MatPtAPNumeric_MPIAIJ_MPIAIJ,
2331:        MatConjugate_MPIAIJ,
2332:        0,
2333: /*105*/MatSetValuesRow_MPIAIJ,
2334:        MatRealPart_MPIAIJ,
2335:        MatImaginaryPart_MPIAIJ,
2336:        0,
2337:        0,
2338: /*110*/0,
2339:        MatGetRedundantMatrix_MPIAIJ,
2340:        MatGetRowMin_MPIAIJ,
2341:        0,
2342:        0,
2343: /*115*/MatGetSeqNonzerostructure_MPIAIJ};

2345: /* ----------------------------------------------------------------------------------------*/

2350: PetscErrorCode  MatStoreValues_MPIAIJ(Mat mat)
2351: {
2352:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

2356:   MatStoreValues(aij->A);
2357:   MatStoreValues(aij->B);
2358:   return(0);
2359: }

2365: PetscErrorCode  MatRetrieveValues_MPIAIJ(Mat mat)
2366: {
2367:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ *)mat->data;

2371:   MatRetrieveValues(aij->A);
2372:   MatRetrieveValues(aij->B);
2373:   return(0);
2374: }

2377:  #include petscpc.h
2381: PetscErrorCode  MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2382: {
2383:   Mat_MPIAIJ     *b;
2385:   PetscInt       i;

2388:   B->preallocated = PETSC_TRUE;
2389:   if (d_nz == PETSC_DEFAULT || d_nz == PETSC_DECIDE) d_nz = 5;
2390:   if (o_nz == PETSC_DEFAULT || o_nz == PETSC_DECIDE) o_nz = 2;
2391:   if (d_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"d_nz cannot be less than 0: value %D",d_nz);
2392:   if (o_nz < 0) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"o_nz cannot be less than 0: value %D",o_nz);

2394:   B->rmap.bs = B->cmap.bs = 1;
2395:   PetscMapSetUp(&B->rmap);
2396:   PetscMapSetUp(&B->cmap);
2397:   if (d_nnz) {
2398:     for (i=0; i<B->rmap.n; i++) {
2399:       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]);
2400:     }
2401:   }
2402:   if (o_nnz) {
2403:     for (i=0; i<B->rmap.n; i++) {
2404:       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]);
2405:     }
2406:   }
2407:   b = (Mat_MPIAIJ*)B->data;

2409:   /* Explicitly create 2 MATSEQAIJ matrices. */
2410:   MatCreate(PETSC_COMM_SELF,&b->A);
2411:   MatSetSizes(b->A,B->rmap.n,B->cmap.n,B->rmap.n,B->cmap.n);
2412:   MatSetType(b->A,MATSEQAIJ);
2413:   PetscLogObjectParent(B,b->A);
2414:   MatCreate(PETSC_COMM_SELF,&b->B);
2415:   MatSetSizes(b->B,B->rmap.n,B->cmap.N,B->rmap.n,B->cmap.N);
2416:   MatSetType(b->B,MATSEQAIJ);
2417:   PetscLogObjectParent(B,b->B);

2419:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2420:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);

2422:   return(0);
2423: }

2428: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2429: {
2430:   Mat            mat;
2431:   Mat_MPIAIJ     *a,*oldmat = (Mat_MPIAIJ*)matin->data;

2435:   *newmat       = 0;
2436:   MatCreate(((PetscObject)matin)->comm,&mat);
2437:   MatSetSizes(mat,matin->rmap.n,matin->cmap.n,matin->rmap.N,matin->cmap.N);
2438:   MatSetType(mat,((PetscObject)matin)->type_name);
2439:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
2440:   a    = (Mat_MPIAIJ*)mat->data;
2441: 
2442:   mat->factor       = matin->factor;
2443:   mat->rmap.bs      = matin->rmap.bs;
2444:   mat->assembled    = PETSC_TRUE;
2445:   mat->insertmode   = NOT_SET_VALUES;
2446:   mat->preallocated = PETSC_TRUE;

2448:   a->size           = oldmat->size;
2449:   a->rank           = oldmat->rank;
2450:   a->donotstash     = oldmat->donotstash;
2451:   a->roworiented    = oldmat->roworiented;
2452:   a->rowindices     = 0;
2453:   a->rowvalues      = 0;
2454:   a->getrowactive   = PETSC_FALSE;

2456:   PetscMapCopy(((PetscObject)mat)->comm,&matin->rmap,&mat->rmap);
2457:   PetscMapCopy(((PetscObject)mat)->comm,&matin->cmap,&mat->cmap);

2459:   MatStashCreate_Private(((PetscObject)matin)->comm,1,&mat->stash);
2460:   if (oldmat->colmap) {
2461: #if defined (PETSC_USE_CTABLE)
2462:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2463: #else
2464:     PetscMalloc((mat->cmap.N)*sizeof(PetscInt),&a->colmap);
2465:     PetscLogObjectMemory(mat,(mat->cmap.N)*sizeof(PetscInt));
2466:     PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap.N)*sizeof(PetscInt));
2467: #endif
2468:   } else a->colmap = 0;
2469:   if (oldmat->garray) {
2470:     PetscInt len;
2471:     len  = oldmat->B->cmap.n;
2472:     PetscMalloc((len+1)*sizeof(PetscInt),&a->garray);
2473:     PetscLogObjectMemory(mat,len*sizeof(PetscInt));
2474:     if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
2475:   } else a->garray = 0;
2476: 
2477:   VecDuplicate(oldmat->lvec,&a->lvec);
2478:   PetscLogObjectParent(mat,a->lvec);
2479:   VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2480:   PetscLogObjectParent(mat,a->Mvctx);
2481:   MatDuplicate(oldmat->A,cpvalues,&a->A);
2482:   PetscLogObjectParent(mat,a->A);
2483:   MatDuplicate(oldmat->B,cpvalues,&a->B);
2484:   PetscLogObjectParent(mat,a->B);
2485:   PetscFListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2486:   *newmat = mat;
2487:   return(0);
2488: }

2490:  #include petscsys.h

2494: PetscErrorCode MatLoad_MPIAIJ(PetscViewer viewer, MatType type,Mat *newmat)
2495: {
2496:   Mat            A;
2497:   PetscScalar    *vals,*svals;
2498:   MPI_Comm       comm = ((PetscObject)viewer)->comm;
2499:   MPI_Status     status;
2501:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag,maxnz;
2502:   PetscInt       i,nz,j,rstart,rend,mmax;
2503:   PetscInt       header[4],*rowlengths = 0,M,N,m,*cols;
2504:   PetscInt       *ourlens = PETSC_NULL,*procsnz = PETSC_NULL,*offlens = PETSC_NULL,jj,*mycols,*smycols;
2505:   PetscInt       cend,cstart,n,*rowners;
2506:   int            fd;

2509:   MPI_Comm_size(comm,&size);
2510:   MPI_Comm_rank(comm,&rank);
2511:   if (!rank) {
2512:     PetscViewerBinaryGetDescriptor(viewer,&fd);
2513:     PetscBinaryRead(fd,(char *)header,4,PETSC_INT);
2514:     if (header[0] != MAT_FILE_COOKIE) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2515:   }

2517:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2518:   M = header[1]; N = header[2];
2519:   /* determine ownership of all rows */
2520:   m    = M/size + ((M % size) > rank);
2521:   PetscMalloc((size+1)*sizeof(PetscInt),&rowners);
2522:   MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

2524:   /* First process needs enough room for process with most rows */
2525:   if (!rank) {
2526:     mmax       = rowners[1];
2527:     for (i=2; i<size; i++) {
2528:       mmax = PetscMax(mmax,rowners[i]);
2529:     }
2530:   } else mmax = m;

2532:   rowners[0] = 0;
2533:   for (i=2; i<=size; i++) {
2534:     rowners[i] += rowners[i-1];
2535:   }
2536:   rstart = rowners[rank];
2537:   rend   = rowners[rank+1];

2539:   /* distribute row lengths to all processors */
2540:   PetscMalloc2(mmax,PetscInt,&ourlens,mmax,PetscInt,&offlens);
2541:   if (!rank) {
2542:     PetscBinaryRead(fd,ourlens,m,PETSC_INT);
2543:     PetscMalloc(m*sizeof(PetscInt),&rowlengths);
2544:     PetscMalloc(size*sizeof(PetscInt),&procsnz);
2545:     PetscMemzero(procsnz,size*sizeof(PetscInt));
2546:     for (j=0; j<m; j++) {
2547:       procsnz[0] += ourlens[j];
2548:     }
2549:     for (i=1; i<size; i++) {
2550:       PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
2551:       /* calculate the number of nonzeros on each processor */
2552:       for (j=0; j<rowners[i+1]-rowners[i]; j++) {
2553:         procsnz[i] += rowlengths[j];
2554:       }
2555:       MPI_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
2556:     }
2557:     PetscFree(rowlengths);
2558:   } else {
2559:     MPI_Recv(ourlens,m,MPIU_INT,0,tag,comm,&status);
2560:   }

2562:   if (!rank) {
2563:     /* determine max buffer needed and allocate it */
2564:     maxnz = 0;
2565:     for (i=0; i<size; i++) {
2566:       maxnz = PetscMax(maxnz,procsnz[i]);
2567:     }
2568:     PetscMalloc(maxnz*sizeof(PetscInt),&cols);

2570:     /* read in my part of the matrix column indices  */
2571:     nz   = procsnz[0];
2572:     PetscMalloc(nz*sizeof(PetscInt),&mycols);
2573:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);

2575:     /* read in every one elses and ship off */
2576:     for (i=1; i<size; i++) {
2577:       nz   = procsnz[i];
2578:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
2579:       MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
2580:     }
2581:     PetscFree(cols);
2582:   } else {
2583:     /* determine buffer space needed for message */
2584:     nz = 0;
2585:     for (i=0; i<m; i++) {
2586:       nz += ourlens[i];
2587:     }
2588:     PetscMalloc(nz*sizeof(PetscInt),&mycols);

2590:     /* receive message of column indices*/
2591:     MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
2592:     MPI_Get_count(&status,MPIU_INT,&maxnz);
2593:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
2594:   }

2596:   /* determine column ownership if matrix is not square */
2597:   if (N != M) {
2598:     n      = N/size + ((N % size) > rank);
2599:     MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
2600:     cstart = cend - n;
2601:   } else {
2602:     cstart = rstart;
2603:     cend   = rend;
2604:     n      = cend - cstart;
2605:   }

2607:   /* loop over local rows, determining number of off diagonal entries */
2608:   PetscMemzero(offlens,m*sizeof(PetscInt));
2609:   jj = 0;
2610:   for (i=0; i<m; i++) {
2611:     for (j=0; j<ourlens[i]; j++) {
2612:       if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
2613:       jj++;
2614:     }
2615:   }

2617:   /* create our matrix */
2618:   for (i=0; i<m; i++) {
2619:     ourlens[i] -= offlens[i];
2620:   }
2621:   MatCreate(comm,&A);
2622:   MatSetSizes(A,m,n,M,N);
2623:   MatSetType(A,type);
2624:   MatMPIAIJSetPreallocation(A,0,ourlens,0,offlens);

2626:   for (i=0; i<m; i++) {
2627:     ourlens[i] += offlens[i];
2628:   }

2630:   if (!rank) {
2631:     PetscMalloc((maxnz+1)*sizeof(PetscScalar),&vals);

2633:     /* read in my part of the matrix numerical values  */
2634:     nz   = procsnz[0];
2635:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2636: 
2637:     /* insert into matrix */
2638:     jj      = rstart;
2639:     smycols = mycols;
2640:     svals   = vals;
2641:     for (i=0; i<m; i++) {
2642:       MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
2643:       smycols += ourlens[i];
2644:       svals   += ourlens[i];
2645:       jj++;
2646:     }

2648:     /* read in other processors and ship out */
2649:     for (i=1; i<size; i++) {
2650:       nz   = procsnz[i];
2651:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
2652:       MPI_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)A)->tag,comm);
2653:     }
2654:     PetscFree(procsnz);
2655:   } else {
2656:     /* receive numeric values */
2657:     PetscMalloc((nz+1)*sizeof(PetscScalar),&vals);

2659:     /* receive message of values*/
2660:     MPI_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)A)->tag,comm,&status);
2661:     MPI_Get_count(&status,MPIU_SCALAR,&maxnz);
2662:     if (maxnz != nz) SETERRQ(PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");

2664:     /* insert into matrix */
2665:     jj      = rstart;
2666:     smycols = mycols;
2667:     svals   = vals;
2668:     for (i=0; i<m; i++) {
2669:       MatSetValues_MPIAIJ(A,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
2670:       smycols += ourlens[i];
2671:       svals   += ourlens[i];
2672:       jj++;
2673:     }
2674:   }
2675:   PetscFree2(ourlens,offlens);
2676:   PetscFree(vals);
2677:   PetscFree(mycols);
2678:   PetscFree(rowners);

2680:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2681:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2682:   *newmat = A;
2683:   return(0);
2684: }

2688: /*
2689:     Not great since it makes two copies of the submatrix, first an SeqAIJ 
2690:   in local and then by concatenating the local matrices the end result.
2691:   Writing it directly would be much like MatGetSubMatrices_MPIAIJ()
2692: */
2693: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2694: {
2696:   PetscMPIInt    rank,size;
2697:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j;
2698:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
2699:   Mat            *local,M,Mreuse;
2700:   PetscScalar    *vwork,*aa;
2701:   MPI_Comm       comm = ((PetscObject)mat)->comm;
2702:   Mat_SeqAIJ     *aij;


2706:   MPI_Comm_rank(comm,&rank);
2707:   MPI_Comm_size(comm,&size);

2709:   if (call ==  MAT_REUSE_MATRIX) {
2710:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject *)&Mreuse);
2711:     if (!Mreuse) SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2712:     local = &Mreuse;
2713:     MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&local);
2714:   } else {
2715:     MatGetSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&local);
2716:     Mreuse = *local;
2717:     PetscFree(local);
2718:   }

2720:   /* 
2721:       m - number of local rows
2722:       n - number of columns (same on all processors)
2723:       rstart - first row in new global matrix generated
2724:   */
2725:   MatGetSize(Mreuse,&m,&n);
2726:   if (call == MAT_INITIAL_MATRIX) {
2727:     aij = (Mat_SeqAIJ*)(Mreuse)->data;
2728:     ii  = aij->i;
2729:     jj  = aij->j;

2731:     /*
2732:         Determine the number of non-zeros in the diagonal and off-diagonal 
2733:         portions of the matrix in order to do correct preallocation
2734:     */

2736:     /* first get start and end of "diagonal" columns */
2737:     if (csize == PETSC_DECIDE) {
2738:       ISGetSize(isrow,&mglobal);
2739:       if (mglobal == n) { /* square matrix */
2740:         nlocal = m;
2741:       } else {
2742:         nlocal = n/size + ((n % size) > rank);
2743:       }
2744:     } else {
2745:       nlocal = csize;
2746:     }
2747:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2748:     rstart = rend - nlocal;
2749:     if (rank == size - 1 && rend != n) {
2750:       SETERRQ2(PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
2751:     }

2753:     /* next, compute all the lengths */
2754:     PetscMalloc((2*m+1)*sizeof(PetscInt),&dlens);
2755:     olens = dlens + m;
2756:     for (i=0; i<m; i++) {
2757:       jend = ii[i+1] - ii[i];
2758:       olen = 0;
2759:       dlen = 0;
2760:       for (j=0; j<jend; j++) {
2761:         if (*jj < rstart || *jj >= rend) olen++;
2762:         else dlen++;
2763:         jj++;
2764:       }
2765:       olens[i] = olen;
2766:       dlens[i] = dlen;
2767:     }
2768:     MatCreate(comm,&M);
2769:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
2770:     MatSetType(M,((PetscObject)mat)->type_name);
2771:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
2772:     PetscFree(dlens);
2773:   } else {
2774:     PetscInt ml,nl;

2776:     M = *newmat;
2777:     MatGetLocalSize(M,&ml,&nl);
2778:     if (ml != m) SETERRQ(PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2779:     MatZeroEntries(M);
2780:     /*
2781:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2782:        rather than the slower MatSetValues().
2783:     */
2784:     M->was_assembled = PETSC_TRUE;
2785:     M->assembled     = PETSC_FALSE;
2786:   }
2787:   MatGetOwnershipRange(M,&rstart,&rend);
2788:   aij = (Mat_SeqAIJ*)(Mreuse)->data;
2789:   ii  = aij->i;
2790:   jj  = aij->j;
2791:   aa  = aij->a;
2792:   for (i=0; i<m; i++) {
2793:     row   = rstart + i;
2794:     nz    = ii[i+1] - ii[i];
2795:     cwork = jj;     jj += nz;
2796:     vwork = aa;     aa += nz;
2797:     MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2798:   }

2800:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2801:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2802:   *newmat = M;

2804:   /* save submatrix used in processor for next request */
2805:   if (call ==  MAT_INITIAL_MATRIX) {
2806:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2807:     PetscObjectDereference((PetscObject)Mreuse);
2808:   }

2810:   return(0);
2811: }

2816: PetscErrorCode  MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
2817: {
2818:   PetscInt       m,cstart, cend,j,nnz,i,d;
2819:   PetscInt       *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
2820:   const PetscInt *JJ;
2821:   PetscScalar    *values;

2825:   if (Ii[0]) SETERRQ1(PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);

2827:   B->rmap.bs = B->cmap.bs = 1;
2828:   PetscMapSetUp(&B->rmap);
2829:   PetscMapSetUp(&B->cmap);
2830:   m      = B->rmap.n;
2831:   cstart = B->cmap.rstart;
2832:   cend   = B->cmap.rend;
2833:   rstart = B->rmap.rstart;

2835:   PetscMalloc((2*m+1)*sizeof(PetscInt),&d_nnz);
2836:   o_nnz = d_nnz + m;

2838: #if defined(PETSC_USE_DEBUGGING)
2839:   for (i=0; i<m; i++) {
2840:     nnz     = Ii[i+1]- Ii[i];
2841:     JJ      = J + Ii[i];
2842:     if (nnz < 0) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
2843:     if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j);
2844:     if (nnz && (JJ[nnz-1] >= B->cmap.N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap.N);
2845:     for (j=1; j<nnz; j++) {
2846:       if (JJ[i] <= JJ[i-1]) SETERRRQ(PETSC_ERR_ARG_WRONGSTATE,"Row %D has unsorted column index at %D location in column indices",i,j);
2847:     }
2848:   }
2849: #endif

2851:   for (i=0; i<m; i++) {
2852:     nnz     = Ii[i+1]- Ii[i];
2853:     JJ      = J + Ii[i];
2854:     nnz_max = PetscMax(nnz_max,nnz);
2855:     for (j=0; j<nnz; j++) {
2856:       if (*JJ >= cstart) break;
2857:       JJ++;
2858:     }
2859:     d = 0;
2860:     for (; j<nnz; j++) {
2861:       if (*JJ++ >= cend) break;
2862:       d++;
2863:     }
2864:     d_nnz[i] = d;
2865:     o_nnz[i] = nnz - d;
2866:   }
2867:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2868:   PetscFree(d_nnz);

2870:   if (v) values = (PetscScalar*)v;
2871:   else {
2872:     PetscMalloc((nnz_max+1)*sizeof(PetscScalar),&values);
2873:     PetscMemzero(values,nnz_max*sizeof(PetscScalar));
2874:   }

2876:   for (i=0; i<m; i++) {
2877:     ii   = i + rstart;
2878:     nnz  = Ii[i+1]- Ii[i];
2879:     MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
2880:   }
2881:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2882:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

2884:   if (!v) {
2885:     PetscFree(values);
2886:   }
2887:   return(0);
2888: }

2893: /*@
2894:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
2895:    (the default parallel PETSc format).  

2897:    Collective on MPI_Comm

2899:    Input Parameters:
2900: +  B - the matrix 
2901: .  i - the indices into j for the start of each local row (starts with zero)
2902: .  j - the column indices for each local row (starts with zero) these must be sorted for each row
2903: -  v - optional values in the matrix

2905:    Level: developer

2907:    Notes:
2908:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
2909:      thus you CANNOT change the matrix entries by changing the values of a[] after you have 
2910:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

2912:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

2914:        The format which is used for the sparse matrix input, is equivalent to a
2915:     row-major ordering.. i.e for the following matrix, the input data expected is
2916:     as shown:

2918:         1 0 0
2919:         2 0 3     P0
2920:        -------
2921:         4 5 6     P1

2923:      Process0 [P0]: rows_owned=[0,1]
2924:         i =  {0,1,3}  [size = nrow+1  = 2+1]
2925:         j =  {0,0,2}  [size = nz = 6]
2926:         v =  {1,2,3}  [size = nz = 6]

2928:      Process1 [P1]: rows_owned=[2]
2929:         i =  {0,3}    [size = nrow+1  = 1+1]
2930:         j =  {0,1,2}  [size = nz = 6]
2931:         v =  {4,5,6}  [size = nz = 6]

2933:       The column indices for each row MUST be sorted.

2935: .keywords: matrix, aij, compressed row, sparse, parallel

2937: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateMPIAIJ(), MPIAIJ,
2938:           MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
2939: @*/
2940: PetscErrorCode  MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2941: {
2942:   PetscErrorCode ierr,(*f)(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]);

2945:   PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",(void (**)(void))&f);
2946:   if (f) {
2947:     (*f)(B,i,j,v);
2948:   }
2949:   return(0);
2950: }

2954: /*@C
2955:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
2956:    (the default parallel PETSc format).  For good matrix assembly performance
2957:    the user should preallocate the matrix storage by setting the parameters 
2958:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
2959:    performance can be increased by more than a factor of 50.

2961:    Collective on MPI_Comm

2963:    Input Parameters:
2964: +  A - the matrix 
2965: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
2966:            (same value is used for all local rows)
2967: .  d_nnz - array containing the number of nonzeros in the various rows of the 
2968:            DIAGONAL portion of the local submatrix (possibly different for each row)
2969:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
2970:            The size of this array is equal to the number of local rows, i.e 'm'. 
2971:            You must leave room for the diagonal entry even if it is zero.
2972: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
2973:            submatrix (same value is used for all local rows).
2974: -  o_nnz - array containing the number of nonzeros in the various rows of the
2975:            OFF-DIAGONAL portion of the local submatrix (possibly different for
2976:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
2977:            structure. The size of this array is equal to the number 
2978:            of local rows, i.e 'm'. 

2980:    If the *_nnz parameter is given then the *_nz parameter is ignored

2982:    The AIJ format (also called the Yale sparse matrix format or
2983:    compressed row storage (CSR)), is fully compatible with standard Fortran 77
2984:    storage.  The stored row and column indices begin with zero.  See the users manual for details.

2986:    The parallel matrix is partitioned such that the first m0 rows belong to 
2987:    process 0, the next m1 rows belong to process 1, the next m2 rows belong 
2988:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

2990:    The DIAGONAL portion of the local submatrix of a processor can be defined 
2991:    as the submatrix which is obtained by extraction the part corresponding 
2992:    to the rows r1-r2 and columns r1-r2 of the global matrix, where r1 is the 
2993:    first row that belongs to the processor, and r2 is the last row belonging 
2994:    to the this processor. This is a square mxm matrix. The remaining portion 
2995:    of the local submatrix (mxN) constitute the OFF-DIAGONAL portion.

2997:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

2999:    You can call MatGetInfo() to get information on how effective the preallocation was;
3000:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3001:    You can also run with the option -info and look for messages with the string 
3002:    malloc in them to see if additional memory allocation was needed.

3004:    Example usage:
3005:   
3006:    Consider the following 8x8 matrix with 34 non-zero values, that is 
3007:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3008:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
3009:    as follows:

3011: .vb
3012:             1  2  0  |  0  3  0  |  0  4
3013:     Proc0   0  5  6  |  7  0  0  |  8  0
3014:             9  0 10  | 11  0  0  | 12  0
3015:     -------------------------------------
3016:            13  0 14  | 15 16 17  |  0  0
3017:     Proc1   0 18  0  | 19 20 21  |  0  0 
3018:             0  0  0  | 22 23  0  | 24  0
3019:     -------------------------------------
3020:     Proc2  25 26 27  |  0  0 28  | 29  0
3021:            30  0  0  | 31 32 33  |  0 34
3022: .ve

3024:    This can be represented as a collection of submatrices as:

3026: .vb
3027:       A B C
3028:       D E F
3029:       G H I
3030: .ve

3032:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3033:    owned by proc1, G,H,I are owned by proc2.

3035:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3036:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3037:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3039:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3040:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3041:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3042:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3043:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3044:    matrix, ans [DF] as another SeqAIJ matrix.

3046:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3047:    allocated for every row of the local diagonal submatrix, and o_nz
3048:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3049:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
3050:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
3051:    In this case, the values of d_nz,o_nz are:
3052: .vb
3053:      proc0 : dnz = 2, o_nz = 2
3054:      proc1 : dnz = 3, o_nz = 2
3055:      proc2 : dnz = 1, o_nz = 4
3056: .ve
3057:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3058:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3059:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
3060:    34 values.

3062:    When d_nnz, o_nnz parameters are specified, the storage is specified
3063:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3064:    In the above case the values for d_nnz,o_nnz are:
3065: .vb
3066:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3067:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3068:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3069: .ve
3070:    Here the space allocated is sum of all the above values i.e 34, and
3071:    hence pre-allocation is perfect.

3073:    Level: intermediate

3075: .keywords: matrix, aij, compressed row, sparse, parallel

3077: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateMPIAIJ(), MatMPIAIJSetPreallocationCSR(),
3078:           MPIAIJ, MatGetInfo()
3079: @*/
3080: PetscErrorCode  MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3081: {
3082:   PetscErrorCode ierr,(*f)(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]);

3085:   PetscObjectQueryFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",(void (**)(void))&f);
3086:   if (f) {
3087:     (*f)(B,d_nz,d_nnz,o_nz,o_nnz);
3088:   }
3089:   return(0);
3090: }

3094: /*@
3095:      MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
3096:          CSR format the local rows.

3098:    Collective on MPI_Comm

3100:    Input Parameters:
3101: +  comm - MPI communicator
3102: .  m - number of local rows (Cannot be PETSC_DECIDE)
3103: .  n - This value should be the same as the local size used in creating the 
3104:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3105:        calculated if N is given) For square matrices n is almost always m.
3106: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3107: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3108: .   i - row indices
3109: .   j - column indices
3110: -   a - matrix values

3112:    Output Parameter:
3113: .   mat - the matrix

3115:    Level: intermediate

3117:    Notes:
3118:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3119:      thus you CANNOT change the matrix entries by changing the values of a[] after you have 
3120:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3122:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3124:        The format which is used for the sparse matrix input, is equivalent to a
3125:     row-major ordering.. i.e for the following matrix, the input data expected is
3126:     as shown:

3128:         1 0 0
3129:         2 0 3     P0
3130:        -------
3131:         4 5 6     P1

3133:      Process0 [P0]: rows_owned=[0,1]
3134:         i =  {0,1,3}  [size = nrow+1  = 2+1]
3135:         j =  {0,0,2}  [size = nz = 6]
3136:         v =  {1,2,3}  [size = nz = 6]

3138:      Process1 [P1]: rows_owned=[2]
3139:         i =  {0,3}    [size = nrow+1  = 1+1]
3140:         j =  {0,1,2}  [size = nz = 6]
3141:         v =  {4,5,6}  [size = nz = 6]

3143: .keywords: matrix, aij, compressed row, sparse, parallel

3145: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3146:           MPIAIJ, MatCreateMPIAIJ(), MatCreateMPIAIJWithSplitArrays()
3147: @*/
3148: PetscErrorCode  MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
3149: {

3153:   if (i[0]) {
3154:     SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3155:   }
3156:   if (m < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
3157:   MatCreate(comm,mat);
3158:   MatSetSizes(*mat,m,n,M,N);
3159:   MatSetType(*mat,MATMPIAIJ);
3160:   MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
3161:   return(0);
3162: }

3166: /*@C
3167:    MatCreateMPIAIJ - Creates a sparse parallel matrix in AIJ format
3168:    (the default parallel PETSc format).  For good matrix assembly performance
3169:    the user should preallocate the matrix storage by setting the parameters 
3170:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3171:    performance can be increased by more than a factor of 50.

3173:    Collective on MPI_Comm

3175:    Input Parameters:
3176: +  comm - MPI communicator
3177: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
3178:            This value should be the same as the local size used in creating the 
3179:            y vector for the matrix-vector product y = Ax.
3180: .  n - This value should be the same as the local size used in creating the 
3181:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3182:        calculated if N is given) For square matrices n is almost always m.
3183: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3184: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3185: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
3186:            (same value is used for all local rows)
3187: .  d_nnz - array containing the number of nonzeros in the various rows of the 
3188:            DIAGONAL portion of the local submatrix (possibly different for each row)
3189:            or PETSC_NULL, if d_nz is used to specify the nonzero structure. 
3190:            The size of this array is equal to the number of local rows, i.e 'm'. 
3191:            You must leave room for the diagonal entry even if it is zero.
3192: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
3193:            submatrix (same value is used for all local rows).
3194: -  o_nnz - array containing the number of nonzeros in the various rows of the
3195:            OFF-DIAGONAL portion of the local submatrix (possibly different for
3196:            each row) or PETSC_NULL, if o_nz is used to specify the nonzero 
3197:            structure. The size of this array is equal to the number 
3198:            of local rows, i.e 'm'. 

3200:    Output Parameter:
3201: .  A - the matrix 

3203:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3204:    MatXXXXSetPreallocation() paradgm instead of this routine directly. This is definitely
3205:    true if you plan to use the external direct solvers such as SuperLU, MUMPS or Spooles.
3206:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

3208:    Notes:
3209:    If the *_nnz parameter is given then the *_nz parameter is ignored

3211:    m,n,M,N parameters specify the size of the matrix, and its partitioning across
3212:    processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
3213:    storage requirements for this matrix.

3215:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one 
3216:    processor than it must be used on all processors that share the object for 
3217:    that argument.

3219:    The user MUST specify either the local or global matrix dimensions
3220:    (possibly both).

3222:    The parallel matrix is partitioned across processors such that the
3223:    first m0 rows belong to process 0, the next m1 rows belong to
3224:    process 1, the next m2 rows belong to process 2 etc.. where
3225:    m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
3226:    values corresponding to [m x N] submatrix.

3228:    The columns are logically partitioned with the n0 columns belonging
3229:    to 0th partition, the next n1 columns belonging to the next
3230:    partition etc.. where n0,n1,n2... are the the input parameter 'n'.

3232:    The DIAGONAL portion of the local submatrix on any given processor
3233:    is the submatrix corresponding to the rows and columns m,n
3234:    corresponding to the given processor. i.e diagonal matrix on
3235:    process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
3236:    etc. The remaining portion of the local submatrix [m x (N-n)]
3237:    constitute the OFF-DIAGONAL portion. The example below better
3238:    illustrates this concept.

3240:    For a square global matrix we define each processor's diagonal portion 
3241:    to be its local rows and the corresponding columns (a square submatrix);  
3242:    each processor's off-diagonal portion encompasses the remainder of the
3243:    local matrix (a rectangular submatrix). 

3245:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

3247:    When calling this routine with a single process communicator, a matrix of
3248:    type SEQAIJ is returned.  If a matrix of type MPIAIJ is desired for this
3249:    type of communicator, use the construction mechanism:
3250:      MatCreate(...,&A); MatSetType(A,MPIAIJ); MatMPIAIJSetPreallocation(A,...);

3252:    By default, this format uses inodes (identical nodes) when possible.
3253:    We search for consecutive rows with the same nonzero structure, thereby
3254:    reusing matrix information to achieve increased efficiency.

3256:    Options Database Keys:
3257: +  -mat_no_inode  - Do not use inodes
3258: .  -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3259: -  -mat_aij_oneindex - Internally use indexing starting at 1
3260:         rather than 0.  Note that when calling MatSetValues(),
3261:         the user still MUST index entries starting at 0!


3264:    Example usage:
3265:   
3266:    Consider the following 8x8 matrix with 34 non-zero values, that is 
3267:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3268:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 
3269:    as follows:

3271: .vb
3272:             1  2  0  |  0  3  0  |  0  4
3273:     Proc0   0  5  6  |  7  0  0  |  8  0
3274:             9  0 10  | 11  0  0  | 12  0
3275:     -------------------------------------
3276:            13  0 14  | 15 16 17  |  0  0
3277:     Proc1   0 18  0  | 19 20 21  |  0  0 
3278:             0  0  0  | 22 23  0  | 24  0
3279:     -------------------------------------
3280:     Proc2  25 26 27  |  0  0 28  | 29  0
3281:            30  0  0  | 31 32 33  |  0 34
3282: .ve

3284:    This can be represented as a collection of submatrices as:

3286: .vb
3287:       A B C
3288:       D E F
3289:       G H I
3290: .ve

3292:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3293:    owned by proc1, G,H,I are owned by proc2.

3295:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3296:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3297:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3299:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3300:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3301:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3302:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3303:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3304:    matrix, ans [DF] as another SeqAIJ matrix.

3306:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3307:    allocated for every row of the local diagonal submatrix, and o_nz
3308:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3309:    One way to choose d_nz and o_nz is to use the max nonzerors per local 
3310:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 
3311:    In this case, the values of d_nz,o_nz are:
3312: .vb
3313:      proc0 : dnz = 2, o_nz = 2
3314:      proc1 : dnz = 3, o_nz = 2
3315:      proc2 : dnz = 1, o_nz = 4
3316: .ve
3317:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3318:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3319:    for proc3. i.e we are using 12+15+10=37 storage locations to store 
3320:    34 values.

3322:    When d_nnz, o_nnz parameters are specified, the storage is specified
3323:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3324:    In the above case the values for d_nnz,o_nnz are:
3325: .vb
3326:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3327:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3328:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3329: .ve
3330:    Here the space allocated is sum of all the above values i.e 34, and
3331:    hence pre-allocation is perfect.

3333:    Level: intermediate

3335: .keywords: matrix, aij, compressed row, sparse, parallel

3337: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3338:           MPIAIJ, MatCreateMPIAIJWithArrays()
3339: @*/
3340: PetscErrorCode  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)
3341: {
3343:   PetscMPIInt    size;

3346:   MatCreate(comm,A);
3347:   MatSetSizes(*A,m,n,M,N);
3348:   MPI_Comm_size(comm,&size);
3349:   if (size > 1) {
3350:     MatSetType(*A,MATMPIAIJ);
3351:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
3352:   } else {
3353:     MatSetType(*A,MATSEQAIJ);
3354:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
3355:   }
3356:   return(0);
3357: }

3361: PetscErrorCode  MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,PetscInt *colmap[])
3362: {
3363:   Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;

3366:   *Ad     = a->A;
3367:   *Ao     = a->B;
3368:   *colmap = a->garray;
3369:   return(0);
3370: }

3374: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
3375: {
3377:   PetscInt       i;
3378:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

3381:   if (coloring->ctype == IS_COLORING_GLOBAL) {
3382:     ISColoringValue *allcolors,*colors;
3383:     ISColoring      ocoloring;

3385:     /* set coloring for diagonal portion */
3386:     MatSetColoring_SeqAIJ(a->A,coloring);

3388:     /* set coloring for off-diagonal portion */
3389:     ISAllGatherColors(((PetscObject)A)->comm,coloring->n,coloring->colors,PETSC_NULL,&allcolors);
3390:     PetscMalloc((a->B->cmap.n+1)*sizeof(ISColoringValue),&colors);
3391:     for (i=0; i<a->B->cmap.n; i++) {
3392:       colors[i] = allcolors[a->garray[i]];
3393:     }
3394:     PetscFree(allcolors);
3395:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap.n,colors,&ocoloring);
3396:     MatSetColoring_SeqAIJ(a->B,ocoloring);
3397:     ISColoringDestroy(ocoloring);
3398:   } else if (coloring->ctype == IS_COLORING_GHOSTED) {
3399:     ISColoringValue *colors;
3400:     PetscInt        *larray;
3401:     ISColoring      ocoloring;

3403:     /* set coloring for diagonal portion */
3404:     PetscMalloc((a->A->cmap.n+1)*sizeof(PetscInt),&larray);
3405:     for (i=0; i<a->A->cmap.n; i++) {
3406:       larray[i] = i + A->cmap.rstart;
3407:     }
3408:     ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->A->cmap.n,larray,PETSC_NULL,larray);
3409:     PetscMalloc((a->A->cmap.n+1)*sizeof(ISColoringValue),&colors);
3410:     for (i=0; i<a->A->cmap.n; i++) {
3411:       colors[i] = coloring->colors[larray[i]];
3412:     }
3413:     PetscFree(larray);
3414:     ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap.n,colors,&ocoloring);
3415:     MatSetColoring_SeqAIJ(a->A,ocoloring);
3416:     ISColoringDestroy(ocoloring);

3418:     /* set coloring for off-diagonal portion */
3419:     PetscMalloc((a->B->cmap.n+1)*sizeof(PetscInt),&larray);
3420:     ISGlobalToLocalMappingApply(A->mapping,IS_GTOLM_MASK,a->B->cmap.n,a->garray,PETSC_NULL,larray);
3421:     PetscMalloc((a->B->cmap.n+1)*sizeof(ISColoringValue),&colors);
3422:     for (i=0; i<a->B->cmap.n; i++) {
3423:       colors[i] = coloring->colors[larray[i]];
3424:     }
3425:     PetscFree(larray);
3426:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap.n,colors,&ocoloring);
3427:     MatSetColoring_SeqAIJ(a->B,ocoloring);
3428:     ISColoringDestroy(ocoloring);
3429:   } else {
3430:     SETERRQ1(PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);
3431:   }

3433:   return(0);
3434: }

3436: #if defined(PETSC_HAVE_ADIC)
3439: PetscErrorCode MatSetValuesAdic_MPIAIJ(Mat A,void *advalues)
3440: {
3441:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

3445:   MatSetValuesAdic_SeqAIJ(a->A,advalues);
3446:   MatSetValuesAdic_SeqAIJ(a->B,advalues);
3447:   return(0);
3448: }
3449: #endif

3453: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
3454: {
3455:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

3459:   MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
3460:   MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
3461:   return(0);
3462: }

3466: /*@
3467:       MatMerge - Creates a single large PETSc matrix by concatinating sequential
3468:                  matrices from each processor

3470:     Collective on MPI_Comm

3472:    Input Parameters:
3473: +    comm - the communicators the parallel matrix will live on
3474: .    inmat - the input sequential matrices
3475: .    n - number of local columns (or PETSC_DECIDE)
3476: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

3478:    Output Parameter:
3479: .    outmat - the parallel matrix generated

3481:     Level: advanced

3483:    Notes: The number of columns of the matrix in EACH processor MUST be the same.

3485: @*/
3486: PetscErrorCode  MatMerge(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
3487: {
3489:   PetscInt       m,N,i,rstart,nnz,Ii,*dnz,*onz;
3490:   PetscInt       *indx;
3491:   PetscScalar    *values;

3494:   MatGetSize(inmat,&m,&N);
3495:   if (scall == MAT_INITIAL_MATRIX){
3496:     /* count nonzeros in each row, for diagonal and off diagonal portion of matrix */
3497:     if (n == PETSC_DECIDE){
3498:       PetscSplitOwnership(comm,&n,&N);
3499:     }
3500:     MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
3501:     rstart -= m;

3503:     MatPreallocateInitialize(comm,m,n,dnz,onz);
3504:     for (i=0;i<m;i++) {
3505:       MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
3506:       MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
3507:       MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,PETSC_NULL);
3508:     }
3509:     /* This routine will ONLY return MPIAIJ type matrix */
3510:     MatCreate(comm,outmat);
3511:     MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3512:     MatSetType(*outmat,MATMPIAIJ);
3513:     MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
3514:     MatPreallocateFinalize(dnz,onz);
3515: 
3516:   } else if (scall == MAT_REUSE_MATRIX){
3517:     MatGetOwnershipRange(*outmat,&rstart,PETSC_NULL);
3518:   } else {
3519:     SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
3520:   }

3522:   for (i=0;i<m;i++) {
3523:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3524:     Ii    = i + rstart;
3525:     MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
3526:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3527:   }
3528:   MatDestroy(inmat);
3529:   MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
3530:   MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);

3532:   return(0);
3533: }

3537: PetscErrorCode MatFileSplit(Mat A,char *outfile)
3538: {
3539:   PetscErrorCode    ierr;
3540:   PetscMPIInt       rank;
3541:   PetscInt          m,N,i,rstart,nnz;
3542:   size_t            len;
3543:   const PetscInt    *indx;
3544:   PetscViewer       out;
3545:   char              *name;
3546:   Mat               B;
3547:   const PetscScalar *values;

3550:   MatGetLocalSize(A,&m,0);
3551:   MatGetSize(A,0,&N);
3552:   /* Should this be the type of the diagonal block of A? */
3553:   MatCreate(PETSC_COMM_SELF,&B);
3554:   MatSetSizes(B,m,N,m,N);
3555:   MatSetType(B,MATSEQAIJ);
3556:   MatSeqAIJSetPreallocation(B,0,PETSC_NULL);
3557:   MatGetOwnershipRange(A,&rstart,0);
3558:   for (i=0;i<m;i++) {
3559:     MatGetRow(A,i+rstart,&nnz,&indx,&values);
3560:     MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
3561:     MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
3562:   }
3563:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3564:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

3566:   MPI_Comm_rank(((PetscObject)A)->comm,&rank);
3567:   PetscStrlen(outfile,&len);
3568:   PetscMalloc((len+5)*sizeof(char),&name);
3569:   sprintf(name,"%s.%d",outfile,rank);
3570:   PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
3571:   PetscFree(name);
3572:   MatView(B,out);
3573:   PetscViewerDestroy(out);
3574:   MatDestroy(B);
3575:   return(0);
3576: }

3578: EXTERN PetscErrorCode MatDestroy_MPIAIJ(Mat);
3581: PetscErrorCode  MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
3582: {
3583:   PetscErrorCode       ierr;
3584:   Mat_Merge_SeqsToMPI  *merge;
3585:   PetscContainer       container;

3588:   PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject *)&container);
3589:   if (container) {
3590:     PetscContainerGetPointer(container,(void **)&merge);
3591:     PetscFree(merge->id_r);
3592:     PetscFree(merge->len_s);
3593:     PetscFree(merge->len_r);
3594:     PetscFree(merge->bi);
3595:     PetscFree(merge->bj);
3596:     PetscFree(merge->buf_ri);
3597:     PetscFree(merge->buf_rj);
3598:     PetscFree(merge->coi);
3599:     PetscFree(merge->coj);
3600:     PetscFree(merge->owners_co);
3601:     PetscFree(merge->rowmap.range);
3602: 
3603:     PetscContainerDestroy(container);
3604:     PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
3605:   }
3606:   PetscFree(merge);

3608:   MatDestroy_MPIAIJ(A);
3609:   return(0);
3610: }

3612:  #include src/mat/utils/freespace.h
3613:  #include petscbt.h

3617: /*@C
3618:       MatMerge_SeqsToMPI - Creates a MPIAIJ matrix by adding sequential
3619:                  matrices from each processor

3621:     Collective on MPI_Comm

3623:    Input Parameters:
3624: +    comm - the communicators the parallel matrix will live on
3625: .    seqmat - the input sequential matrices
3626: .    m - number of local rows (or PETSC_DECIDE)
3627: .    n - number of local columns (or PETSC_DECIDE)
3628: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

3630:    Output Parameter:
3631: .    mpimat - the parallel matrix generated

3633:     Level: advanced

3635:    Notes: 
3636:      The dimensions of the sequential matrix in each processor MUST be the same.
3637:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
3638:      destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
3639: @*/
3640: PetscErrorCode  MatMerge_SeqsToMPINumeric(Mat seqmat,Mat mpimat)
3641: {
3642:   PetscErrorCode       ierr;
3643:   MPI_Comm             comm=((PetscObject)mpimat)->comm;
3644:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
3645:   PetscMPIInt          size,rank,taga,*len_s;
3646:   PetscInt             N=mpimat->cmap.N,i,j,*owners,*ai=a->i,*aj=a->j;
3647:   PetscInt             proc,m;
3648:   PetscInt             **buf_ri,**buf_rj;
3649:   PetscInt             k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
3650:   PetscInt             nrows,**buf_ri_k,**nextrow,**nextai;
3651:   MPI_Request          *s_waits,*r_waits;
3652:   MPI_Status           *status;
3653:   MatScalar            *aa=a->a,**abuf_r,*ba_i;
3654:   Mat_Merge_SeqsToMPI  *merge;
3655:   PetscContainer       container;
3656: 

3660:   MPI_Comm_size(comm,&size);
3661:   MPI_Comm_rank(comm,&rank);

3663:   PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject *)&container);
3664:   if (container) {
3665:     PetscContainerGetPointer(container,(void **)&merge);
3666:   }
3667:   bi     = merge->bi;
3668:   bj     = merge->bj;
3669:   buf_ri = merge->buf_ri;
3670:   buf_rj = merge->buf_rj;

3672:   PetscMalloc(size*sizeof(MPI_Status),&status);
3673:   owners = merge->rowmap.range;
3674:   len_s  = merge->len_s;

3676:   /* send and recv matrix values */
3677:   /*-----------------------------*/
3678:   PetscObjectGetNewTag((PetscObject)mpimat,&taga);
3679:   PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);

3681:   PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);
3682:   for (proc=0,k=0; proc<size; proc++){
3683:     if (!len_s[proc]) continue;
3684:     i = owners[proc];
3685:     MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
3686:     k++;
3687:   }

3689:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
3690:   if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
3691:   PetscFree(status);

3693:   PetscFree(s_waits);
3694:   PetscFree(r_waits);

3696:   /* insert mat values of mpimat */
3697:   /*----------------------------*/
3698:   PetscMalloc(N*sizeof(MatScalar),&ba_i);
3699:   PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
3700:   nextrow = buf_ri_k + merge->nrecv;
3701:   nextai  = nextrow + merge->nrecv;

3703:   for (k=0; k<merge->nrecv; k++){
3704:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
3705:     nrows = *(buf_ri_k[k]);
3706:     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
3707:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
3708:   }

3710:   /* set values of ba */
3711:   m = merge->rowmap.n;
3712:   for (i=0; i<m; i++) {
3713:     arow = owners[rank] + i;
3714:     bj_i = bj+bi[i];  /* col indices of the i-th row of mpimat */
3715:     bnzi = bi[i+1] - bi[i];
3716:     PetscMemzero(ba_i,bnzi*sizeof(MatScalar));

3718:     /* add local non-zero vals of this proc's seqmat into ba */
3719:     anzi = ai[arow+1] - ai[arow];
3720:     aj   = a->j + ai[arow];
3721:     aa   = a->a + ai[arow];
3722:     nextaj = 0;
3723:     for (j=0; nextaj<anzi; j++){
3724:       if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
3725:         ba_i[j] += aa[nextaj++];
3726:       }
3727:     }

3729:     /* add received vals into ba */
3730:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
3731:       /* i-th row */
3732:       if (i == *nextrow[k]) {
3733:         anzi = *(nextai[k]+1) - *nextai[k];
3734:         aj   = buf_rj[k] + *(nextai[k]);
3735:         aa   = abuf_r[k] + *(nextai[k]);
3736:         nextaj = 0;
3737:         for (j=0; nextaj<anzi; j++){
3738:           if (*(bj_i + j) == aj[nextaj]){ /* bcol == acol */
3739:             ba_i[j] += aa[nextaj++];
3740:           }
3741:         }
3742:         nextrow[k]++; nextai[k]++;
3743:       }
3744:     }
3745:     MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
3746:   }
3747:   MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
3748:   MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);

3750:   PetscFree(abuf_r);
3751:   PetscFree(ba_i);
3752:   PetscFree(buf_ri_k);
3754:   return(0);
3755: }

3759: PetscErrorCode  MatMerge_SeqsToMPISymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
3760: {
3761:   PetscErrorCode       ierr;
3762:   Mat                  B_mpi;
3763:   Mat_SeqAIJ           *a=(Mat_SeqAIJ*)seqmat->data;
3764:   PetscMPIInt          size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
3765:   PetscInt             **buf_rj,**buf_ri,**buf_ri_k;
3766:   PetscInt             M=seqmat->rmap.n,N=seqmat->cmap.n,i,*owners,*ai=a->i,*aj=a->j;
3767:   PetscInt             len,proc,*dnz,*onz;
3768:   PetscInt             k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
3769:   PetscInt             nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
3770:   MPI_Request          *si_waits,*sj_waits,*ri_waits,*rj_waits;
3771:   MPI_Status           *status;
3772:   PetscFreeSpaceList   free_space=PETSC_NULL,current_space=PETSC_NULL;
3773:   PetscBT              lnkbt;
3774:   Mat_Merge_SeqsToMPI  *merge;
3775:   PetscContainer       container;


3780:   /* make sure it is a PETSc comm */
3781:   PetscCommDuplicate(comm,&comm,PETSC_NULL);
3782:   MPI_Comm_size(comm,&size);
3783:   MPI_Comm_rank(comm,&rank);
3784: 
3785:   PetscNew(Mat_Merge_SeqsToMPI,&merge);
3786:   PetscMalloc(size*sizeof(MPI_Status),&status);

3788:   /* determine row ownership */
3789:   /*---------------------------------------------------------*/
3790:   PetscMapInitialize(comm,&merge->rowmap);
3791:   merge->rowmap.n = m;
3792:   merge->rowmap.N = M;
3793:   merge->rowmap.bs = 1;
3794:   PetscMapSetUp(&merge->rowmap);
3795:   PetscMalloc(size*sizeof(PetscMPIInt),&len_si);
3796:   PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);
3797: 
3798:   m      = merge->rowmap.n;
3799:   M      = merge->rowmap.N;
3800:   owners = merge->rowmap.range;

3802:   /* determine the number of messages to send, their lengths */
3803:   /*---------------------------------------------------------*/
3804:   len_s  = merge->len_s;

3806:   len = 0;  /* length of buf_si[] */
3807:   merge->nsend = 0;
3808:   for (proc=0; proc<size; proc++){
3809:     len_si[proc] = 0;
3810:     if (proc == rank){
3811:       len_s[proc] = 0;
3812:     } else {
3813:       len_si[proc] = owners[proc+1] - owners[proc] + 1;
3814:       len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
3815:     }
3816:     if (len_s[proc]) {
3817:       merge->nsend++;
3818:       nrows = 0;
3819:       for (i=owners[proc]; i<owners[proc+1]; i++){
3820:         if (ai[i+1] > ai[i]) nrows++;
3821:       }
3822:       len_si[proc] = 2*(nrows+1);
3823:       len += len_si[proc];
3824:     }
3825:   }

3827:   /* determine the number and length of messages to receive for ij-structure */
3828:   /*-------------------------------------------------------------------------*/
3829:   PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);
3830:   PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);

3832:   /* post the Irecv of j-structure */
3833:   /*-------------------------------*/
3834:   PetscCommGetNewTag(comm,&tagj);
3835:   PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);

3837:   /* post the Isend of j-structure */
3838:   /*--------------------------------*/
3839:   PetscMalloc((2*merge->nsend+1)*sizeof(MPI_Request),&si_waits);
3840:   sj_waits = si_waits + merge->nsend;

3842:   for (proc=0, k=0; proc<size; proc++){
3843:     if (!len_s[proc]) continue;
3844:     i = owners[proc];
3845:     MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
3846:     k++;
3847:   }

3849:   /* receives and sends of j-structure are complete */
3850:   /*------------------------------------------------*/
3851:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
3852:   if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
3853: 
3854:   /* send and recv i-structure */
3855:   /*---------------------------*/
3856:   PetscCommGetNewTag(comm,&tagi);
3857:   PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
3858: 
3859:   PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);
3860:   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
3861:   for (proc=0,k=0; proc<size; proc++){
3862:     if (!len_s[proc]) continue;
3863:     /* form outgoing message for i-structure: 
3864:          buf_si[0]:                 nrows to be sent
3865:                [1:nrows]:           row index (global)
3866:                [nrows+1:2*nrows+1]: i-structure index
3867:     */
3868:     /*-------------------------------------------*/
3869:     nrows = len_si[proc]/2 - 1;
3870:     buf_si_i    = buf_si + nrows+1;
3871:     buf_si[0]   = nrows;
3872:     buf_si_i[0] = 0;
3873:     nrows = 0;
3874:     for (i=owners[proc]; i<owners[proc+1]; i++){
3875:       anzi = ai[i+1] - ai[i];
3876:       if (anzi) {
3877:         buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
3878:         buf_si[nrows+1] = i-owners[proc]; /* local row index */
3879:         nrows++;
3880:       }
3881:     }
3882:     MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
3883:     k++;
3884:     buf_si += len_si[proc];
3885:   }

3887:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
3888:   if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}

3890:   PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
3891:   for (i=0; i<merge->nrecv; i++){
3892:     PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
3893:   }

3895:   PetscFree(len_si);
3896:   PetscFree(len_ri);
3897:   PetscFree(rj_waits);
3898:   PetscFree(si_waits);
3899:   PetscFree(ri_waits);
3900:   PetscFree(buf_s);
3901:   PetscFree(status);

3903:   /* compute a local seq matrix in each processor */
3904:   /*----------------------------------------------*/
3905:   /* allocate bi array and free space for accumulating nonzero column info */
3906:   PetscMalloc((m+1)*sizeof(PetscInt),&bi);
3907:   bi[0] = 0;

3909:   /* create and initialize a linked list */
3910:   nlnk = N+1;
3911:   PetscLLCreate(N,N,nlnk,lnk,lnkbt);
3912: 
3913:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
3914:   len = 0;
3915:   len  = ai[owners[rank+1]] - ai[owners[rank]];
3916:   PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);
3917:   current_space = free_space;

3919:   /* determine symbolic info for each local row */
3920:   PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);
3921:   nextrow = buf_ri_k + merge->nrecv;
3922:   nextai  = nextrow + merge->nrecv;
3923:   for (k=0; k<merge->nrecv; k++){
3924:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
3925:     nrows = *buf_ri_k[k];
3926:     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
3927:     nextai[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
3928:   }

3930:   MatPreallocateInitialize(comm,m,n,dnz,onz);
3931:   len = 0;
3932:   for (i=0;i<m;i++) {
3933:     bnzi   = 0;
3934:     /* add local non-zero cols of this proc's seqmat into lnk */
3935:     arow   = owners[rank] + i;
3936:     anzi   = ai[arow+1] - ai[arow];
3937:     aj     = a->j + ai[arow];
3938:     PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3939:     bnzi += nlnk;
3940:     /* add received col data into lnk */
3941:     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
3942:       if (i == *nextrow[k]) { /* i-th row */
3943:         anzi = *(nextai[k]+1) - *nextai[k];
3944:         aj   = buf_rj[k] + *nextai[k];
3945:         PetscLLAdd(anzi,aj,N,nlnk,lnk,lnkbt);
3946:         bnzi += nlnk;
3947:         nextrow[k]++; nextai[k]++;
3948:       }
3949:     }
3950:     if (len < bnzi) len = bnzi;  /* =max(bnzi) */

3952:     /* if free space is not available, make more free space */
3953:     if (current_space->local_remaining<bnzi) {
3954:       PetscFreeSpaceGet(current_space->total_array_size,&current_space);
3955:       nspacedouble++;
3956:     }
3957:     /* copy data into free space, then initialize lnk */
3958:     PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
3959:     MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);

3961:     current_space->array           += bnzi;
3962:     current_space->local_used      += bnzi;
3963:     current_space->local_remaining -= bnzi;
3964: 
3965:     bi[i+1] = bi[i] + bnzi;
3966:   }
3967: 
3968:   PetscFree(buf_ri_k);

3970:   PetscMalloc((bi[m]+1)*sizeof(PetscInt),&bj);
3971:   PetscFreeSpaceContiguous(&free_space,bj);
3972:   PetscLLDestroy(lnk,lnkbt);

3974:   /* create symbolic parallel matrix B_mpi */
3975:   /*---------------------------------------*/
3976:   MatCreate(comm,&B_mpi);
3977:   if (n==PETSC_DECIDE) {
3978:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
3979:   } else {
3980:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3981:   }
3982:   MatSetType(B_mpi,MATMPIAIJ);
3983:   MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
3984:   MatPreallocateFinalize(dnz,onz);

3986:   /* B_mpi is not ready for use - assembly will be done by MatMerge_SeqsToMPINumeric() */
3987:   B_mpi->assembled     = PETSC_FALSE;
3988:   B_mpi->ops->destroy  = MatDestroy_MPIAIJ_SeqsToMPI;
3989:   merge->bi            = bi;
3990:   merge->bj            = bj;
3991:   merge->buf_ri        = buf_ri;
3992:   merge->buf_rj        = buf_rj;
3993:   merge->coi           = PETSC_NULL;
3994:   merge->coj           = PETSC_NULL;
3995:   merge->owners_co     = PETSC_NULL;

3997:   /* attach the supporting struct to B_mpi for reuse */
3998:   PetscContainerCreate(PETSC_COMM_SELF,&container);
3999:   PetscContainerSetPointer(container,merge);
4000:   PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4001:   *mpimat = B_mpi;

4003:   PetscCommDestroy(&comm);
4005:   return(0);
4006: }

4010: PetscErrorCode  MatMerge_SeqsToMPI(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4011: {
4012:   PetscErrorCode   ierr;

4016:   if (scall == MAT_INITIAL_MATRIX){
4017:     MatMerge_SeqsToMPISymbolic(comm,seqmat,m,n,mpimat);
4018:   }
4019:   MatMerge_SeqsToMPINumeric(seqmat,*mpimat);
4021:   return(0);
4022: }

4026: /*@
4027:      MatGetLocalMat - Creates a SeqAIJ matrix by taking all its local rows

4029:     Not Collective

4031:    Input Parameters:
4032: +    A - the matrix 
4033: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 

4035:    Output Parameter:
4036: .    A_loc - the local sequential matrix generated

4038:     Level: developer

4040: @*/
4041: PetscErrorCode  MatGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4042: {
4043:   PetscErrorCode  ierr;
4044:   Mat_MPIAIJ      *mpimat=(Mat_MPIAIJ*)A->data;
4045:   Mat_SeqAIJ      *mat,*a=(Mat_SeqAIJ*)(mpimat->A)->data,*b=(Mat_SeqAIJ*)(mpimat->B)->data;
4046:   PetscInt        *ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j,*cmap=mpimat->garray;
4047:   PetscScalar     *aa=a->a,*ba=b->a,*ca;
4048:   PetscInt        am=A->rmap.n,i,j,k,cstart=A->cmap.rstart;
4049:   PetscInt        *ci,*cj,col,ncols_d,ncols_o,jo;

4053:   if (scall == MAT_INITIAL_MATRIX){
4054:     PetscMalloc((1+am)*sizeof(PetscInt),&ci);
4055:     ci[0] = 0;
4056:     for (i=0; i<am; i++){
4057:       ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
4058:     }
4059:     PetscMalloc((1+ci[am])*sizeof(PetscInt),&cj);
4060:     PetscMalloc((1+ci[am])*sizeof(PetscScalar),&ca);
4061:     k = 0;
4062:     for (i=0; i<am; i++) {
4063:       ncols_o = bi[i+1] - bi[i];
4064:       ncols_d = ai[i+1] - ai[i];
4065:       /* off-diagonal portion of A */
4066:       for (jo=0; jo<ncols_o; jo++) {
4067:         col = cmap[*bj];
4068:         if (col >= cstart) break;
4069:         cj[k]   = col; bj++;
4070:         ca[k++] = *ba++;
4071:       }
4072:       /* diagonal portion of A */
4073:       for (j=0; j<ncols_d; j++) {
4074:         cj[k]   = cstart + *aj++;
4075:         ca[k++] = *aa++;
4076:       }
4077:       /* off-diagonal portion of A */
4078:       for (j=jo; j<ncols_o; j++) {
4079:         cj[k]   = cmap[*bj++];
4080:         ca[k++] = *ba++;
4081:       }
4082:     }
4083:     /* put together the new matrix */
4084:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap.N,ci,cj,ca,A_loc);
4085:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4086:     /* Since these are PETSc arrays, change flags to free them as necessary. */
4087:     mat          = (Mat_SeqAIJ*)(*A_loc)->data;
4088:     mat->free_a  = PETSC_TRUE;
4089:     mat->free_ij = PETSC_TRUE;
4090:     mat->nonew   = 0;
4091:   } else if (scall == MAT_REUSE_MATRIX){
4092:     mat=(Mat_SeqAIJ*)(*A_loc)->data;
4093:     ci = mat->i; cj = mat->j; ca = mat->a;
4094:     for (i=0; i<am; i++) {
4095:       /* off-diagonal portion of A */
4096:       ncols_o = bi[i+1] - bi[i];
4097:       for (jo=0; jo<ncols_o; jo++) {
4098:         col = cmap[*bj];
4099:         if (col >= cstart) break;
4100:         *ca++ = *ba++; bj++;
4101:       }
4102:       /* diagonal portion of A */
4103:       ncols_d = ai[i+1] - ai[i];
4104:       for (j=0; j<ncols_d; j++) *ca++ = *aa++;
4105:       /* off-diagonal portion of A */
4106:       for (j=jo; j<ncols_o; j++) {
4107:         *ca++ = *ba++; bj++;
4108:       }
4109:     }
4110:   } else {
4111:     SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
4112:   }

4115:   return(0);
4116: }

4120: /*@C
4121:      MatGetLocalMatCondensed - Creates a SeqAIJ matrix by taking all its local rows and NON-ZERO columns

4123:     Not Collective

4125:    Input Parameters:
4126: +    A - the matrix 
4127: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4128: -    row, col - index sets of rows and columns to extract (or PETSC_NULL)  

4130:    Output Parameter:
4131: .    A_loc - the local sequential matrix generated

4133:     Level: developer

4135: @*/
4136: PetscErrorCode  MatGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
4137: {
4138:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data;
4139:   PetscErrorCode    ierr;
4140:   PetscInt          i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
4141:   IS                isrowa,iscola;
4142:   Mat               *aloc;

4146:   if (!row){
4147:     start = A->rmap.rstart; end = A->rmap.rend;
4148:     ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
4149:   } else {
4150:     isrowa = *row;
4151:   }
4152:   if (!col){
4153:     start = A->cmap.rstart;
4154:     cmap  = a->garray;
4155:     nzA   = a->A->cmap.n;
4156:     nzB   = a->B->cmap.n;
4157:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
4158:     ncols = 0;
4159:     for (i=0; i<nzB; i++) {
4160:       if (cmap[i] < start) idx[ncols++] = cmap[i];
4161:       else break;
4162:     }
4163:     imark = i;
4164:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;
4165:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
4166:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&iscola);
4167:     PetscFree(idx);
4168:   } else {
4169:     iscola = *col;
4170:   }
4171:   if (scall != MAT_INITIAL_MATRIX){
4172:     PetscMalloc(sizeof(Mat),&aloc);
4173:     aloc[0] = *A_loc;
4174:   }
4175:   MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
4176:   *A_loc = aloc[0];
4177:   PetscFree(aloc);
4178:   if (!row){
4179:     ISDestroy(isrowa);
4180:   }
4181:   if (!col){
4182:     ISDestroy(iscola);
4183:   }
4185:   return(0);
4186: }

4190: /*@C
4191:     MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 

4193:     Collective on Mat

4195:    Input Parameters:
4196: +    A,B - the matrices in mpiaij format
4197: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4198: -    rowb, colb - index sets of rows and columns of B to extract (or PETSC_NULL)   

4200:    Output Parameter:
4201: +    rowb, colb - index sets of rows and columns of B to extract 
4202: .    brstart - row index of B_seq from which next B->rmap.n rows are taken from B's local rows
4203: -    B_seq - the sequential matrix generated

4205:     Level: developer

4207: @*/
4208: PetscErrorCode  MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,PetscInt *brstart,Mat *B_seq)
4209: {
4210:   Mat_MPIAIJ        *a=(Mat_MPIAIJ*)A->data;
4211:   PetscErrorCode    ierr;
4212:   PetscInt          *idx,i,start,ncols,nzA,nzB,*cmap,imark;
4213:   IS                isrowb,iscolb;
4214:   Mat               *bseq;
4215: 
4217:   if (A->cmap.rstart != B->rmap.rstart || A->cmap.rend != B->rmap.rend){
4218:     SETERRQ4(PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap.rstart,A->cmap.rend,B->rmap.rstart,B->rmap.rend);
4219:   }
4221: 
4222:   if (scall == MAT_INITIAL_MATRIX){
4223:     start = A->cmap.rstart;
4224:     cmap  = a->garray;
4225:     nzA   = a->A->cmap.n;
4226:     nzB   = a->B->cmap.n;
4227:     PetscMalloc((nzA+nzB)*sizeof(PetscInt), &idx);
4228:     ncols = 0;
4229:     for (i=0; i<nzB; i++) {  /* row < local row index */
4230:       if (cmap[i] < start) idx[ncols++] = cmap[i];
4231:       else break;
4232:     }
4233:     imark = i;
4234:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;  /* local rows */
4235:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
4236:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,&isrowb);
4237:     PetscFree(idx);
4238:     *brstart = imark;
4239:     ISCreateStride(PETSC_COMM_SELF,B->cmap.N,0,1,&iscolb);
4240:   } else {
4241:     if (!rowb || !colb) SETERRQ(PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
4242:     isrowb = *rowb; iscolb = *colb;
4243:     PetscMalloc(sizeof(Mat),&bseq);
4244:     bseq[0] = *B_seq;
4245:   }
4246:   MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
4247:   *B_seq = bseq[0];
4248:   PetscFree(bseq);
4249:   if (!rowb){
4250:     ISDestroy(isrowb);
4251:   } else {
4252:     *rowb = isrowb;
4253:   }
4254:   if (!colb){
4255:     ISDestroy(iscolb);
4256:   } else {
4257:     *colb = iscolb;
4258:   }
4260:   return(0);
4261: }

4265: /*@C
4266:     MatGetBrowsOfAoCols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
4267:     of the OFF-DIAGONAL portion of local A 

4269:     Collective on Mat

4271:    Input Parameters:
4272: +    A,B - the matrices in mpiaij format
4273: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4274: .    startsj - starting point in B's sending and receiving j-arrays, saved for MAT_REUSE (or PETSC_NULL) 
4275: -    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or PETSC_NULL) 

4277:    Output Parameter:
4278: +    B_oth - the sequential matrix generated

4280:     Level: developer

4282: @*/
4283: PetscErrorCode  MatGetBrowsOfAoCols(Mat A,Mat B,MatReuse scall,PetscInt **startsj,PetscScalar **bufa_ptr,Mat *B_oth)
4284: {
4285:   VecScatter_MPI_General *gen_to,*gen_from;
4286:   PetscErrorCode         ierr;
4287:   Mat_MPIAIJ             *a=(Mat_MPIAIJ*)A->data;
4288:   Mat_SeqAIJ             *b_oth;
4289:   VecScatter             ctx=a->Mvctx;
4290:   MPI_Comm               comm=((PetscObject)ctx)->comm;
4291:   PetscMPIInt            *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank;
4292:   PetscInt               *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap.n,row,*b_othi,*b_othj;
4293:   PetscScalar            *rvalues,*svalues,*b_otha,*bufa,*bufA;
4294:   PetscInt               i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
4295:   MPI_Request            *rwaits = PETSC_NULL,*swaits = PETSC_NULL;
4296:   MPI_Status             *sstatus,rstatus;
4297:   PetscMPIInt            jj;
4298:   PetscInt               *cols,sbs,rbs;
4299:   PetscScalar            *vals;

4302:   if (A->cmap.rstart != B->rmap.rstart || A->cmap.rend != B->rmap.rend){
4303:     SETERRQ4(PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap.rstart,A->cmap.rend,B->rmap.rstart,B->rmap.rend);
4304:   }
4306:   MPI_Comm_rank(comm,&rank);

4308:   gen_to   = (VecScatter_MPI_General*)ctx->todata;
4309:   gen_from = (VecScatter_MPI_General*)ctx->fromdata;
4310:   rvalues  = gen_from->values; /* holds the length of receiving row */
4311:   svalues  = gen_to->values;   /* holds the length of sending row */
4312:   nrecvs   = gen_from->n;
4313:   nsends   = gen_to->n;

4315:   PetscMalloc2(nrecvs,MPI_Request,&rwaits,nsends,MPI_Request,&swaits);
4316:   srow     = gen_to->indices;   /* local row index to be sent */
4317:   sstarts  = gen_to->starts;
4318:   sprocs   = gen_to->procs;
4319:   sstatus  = gen_to->sstatus;
4320:   sbs      = gen_to->bs;
4321:   rstarts  = gen_from->starts;
4322:   rprocs   = gen_from->procs;
4323:   rbs      = gen_from->bs;

4325:   if (!startsj || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
4326:   if (scall == MAT_INITIAL_MATRIX){
4327:     /* i-array */
4328:     /*---------*/
4329:     /*  post receives */
4330:     for (i=0; i<nrecvs; i++){
4331:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
4332:       nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
4333:       MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
4334:     }

4336:     /* pack the outgoing message */
4337:     PetscMalloc((nsends+nrecvs+3)*sizeof(PetscInt),&sstartsj);
4338:     rstartsj = sstartsj + nsends +1;
4339:     sstartsj[0] = 0;  rstartsj[0] = 0;
4340:     len = 0; /* total length of j or a array to be sent */
4341:     k = 0;
4342:     for (i=0; i<nsends; i++){
4343:       rowlen = (PetscInt*)svalues + sstarts[i]*sbs;
4344:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
4345:       for (j=0; j<nrows; j++) {
4346:         row = srow[k] + B->rmap.range[rank]; /* global row idx */
4347:         for (l=0; l<sbs; l++){
4348:           MatGetRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL); /* rowlength */
4349:           rowlen[j*sbs+l] = ncols;
4350:           len += ncols;
4351:           MatRestoreRow_MPIAIJ(B,row+l,&ncols,PETSC_NULL,PETSC_NULL);
4352:         }
4353:         k++;
4354:       }
4355:       MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
4356:       sstartsj[i+1] = len;  /* starting point of (i+1)-th outgoing msg in bufj and bufa */
4357:     }
4358:     /* recvs and sends of i-array are completed */
4359:     i = nrecvs;
4360:     while (i--) {
4361:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4362:     }
4363:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}

4365:     /* allocate buffers for sending j and a arrays */
4366:     PetscMalloc((len+1)*sizeof(PetscInt),&bufj);
4367:     PetscMalloc((len+1)*sizeof(PetscScalar),&bufa);

4369:     /* create i-array of B_oth */
4370:     PetscMalloc((aBn+2)*sizeof(PetscInt),&b_othi);
4371:     b_othi[0] = 0;
4372:     len = 0; /* total length of j or a array to be received */
4373:     k = 0;
4374:     for (i=0; i<nrecvs; i++){
4375:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
4376:       nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */
4377:       for (j=0; j<nrows; j++) {
4378:         b_othi[k+1] = b_othi[k] + rowlen[j];
4379:         len += rowlen[j]; k++;
4380:       }
4381:       rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
4382:     }

4384:     /* allocate space for j and a arrrays of B_oth */
4385:     PetscMalloc((b_othi[aBn]+1)*sizeof(PetscInt),&b_othj);
4386:     PetscMalloc((b_othi[aBn]+1)*sizeof(PetscScalar),&b_otha);

4388:     /* j-array */
4389:     /*---------*/
4390:     /*  post receives of j-array */
4391:     for (i=0; i<nrecvs; i++){
4392:       nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
4393:       MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
4394:     }

4396:     /* pack the outgoing message j-array */
4397:     k = 0;
4398:     for (i=0; i<nsends; i++){
4399:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
4400:       bufJ = bufj+sstartsj[i];
4401:       for (j=0; j<nrows; j++) {
4402:         row  = srow[k++] + B->rmap.range[rank]; /* global row idx */
4403:         for (ll=0; ll<sbs; ll++){
4404:           MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
4405:           for (l=0; l<ncols; l++){
4406:             *bufJ++ = cols[l];
4407:           }
4408:           MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,PETSC_NULL);
4409:         }
4410:       }
4411:       MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
4412:     }

4414:     /* recvs and sends of j-array are completed */
4415:     i = nrecvs;
4416:     while (i--) {
4417:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4418:     }
4419:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
4420:   } else if (scall == MAT_REUSE_MATRIX){
4421:     sstartsj = *startsj;
4422:     rstartsj = sstartsj + nsends +1;
4423:     bufa     = *bufa_ptr;
4424:     b_oth    = (Mat_SeqAIJ*)(*B_oth)->data;
4425:     b_otha   = b_oth->a;
4426:   } else {
4427:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
4428:   }

4430:   /* a-array */
4431:   /*---------*/
4432:   /*  post receives of a-array */
4433:   for (i=0; i<nrecvs; i++){
4434:     nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
4435:     MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
4436:   }

4438:   /* pack the outgoing message a-array */
4439:   k = 0;
4440:   for (i=0; i<nsends; i++){
4441:     nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
4442:     bufA = bufa+sstartsj[i];
4443:     for (j=0; j<nrows; j++) {
4444:       row  = srow[k++] + B->rmap.range[rank]; /* global row idx */
4445:       for (ll=0; ll<sbs; ll++){
4446:         MatGetRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
4447:         for (l=0; l<ncols; l++){
4448:           *bufA++ = vals[l];
4449:         }
4450:         MatRestoreRow_MPIAIJ(B,row+ll,&ncols,PETSC_NULL,&vals);
4451:       }
4452:     }
4453:     MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
4454:   }
4455:   /* recvs and sends of a-array are completed */
4456:   i = nrecvs;
4457:   while (i--) {
4458:     MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4459:   }
4460:   if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
4461:   PetscFree2(rwaits,swaits);

4463:   if (scall == MAT_INITIAL_MATRIX){
4464:     /* put together the new matrix */
4465:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap.N,b_othi,b_othj,b_otha,B_oth);

4467:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4468:     /* Since these are PETSc arrays, change flags to free them as necessary. */
4469:     b_oth          = (Mat_SeqAIJ *)(*B_oth)->data;
4470:     b_oth->free_a  = PETSC_TRUE;
4471:     b_oth->free_ij = PETSC_TRUE;
4472:     b_oth->nonew   = 0;

4474:     PetscFree(bufj);
4475:     if (!startsj || !bufa_ptr){
4476:       PetscFree(sstartsj);
4477:       PetscFree(bufa_ptr);
4478:     } else {
4479:       *startsj  = sstartsj;
4480:       *bufa_ptr = bufa;
4481:     }
4482:   }
4484:   return(0);
4485: }

4489: /*@C
4490:   MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.

4492:   Not Collective

4494:   Input Parameters:
4495: . A - The matrix in mpiaij format

4497:   Output Parameter:
4498: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
4499: . colmap - A map from global column index to local index into lvec
4500: - multScatter - A scatter from the argument of a matrix-vector product to lvec

4502:   Level: developer

4504: @*/
4505: #if defined (PETSC_USE_CTABLE)
4506: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
4507: #else
4508: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
4509: #endif
4510: {
4511:   Mat_MPIAIJ *a;

4518:   a = (Mat_MPIAIJ *) A->data;
4519:   if (lvec) *lvec = a->lvec;
4520:   if (colmap) *colmap = a->colmap;
4521:   if (multScatter) *multScatter = a->Mvctx;
4522:   return(0);
4523: }


4530: /*MC
4531:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

4533:    Options Database Keys:
4534: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()

4536:   Level: beginner

4538: .seealso: MatCreateMPIAIJ()
4539: M*/

4544: PetscErrorCode  MatCreate_MPIAIJ(Mat B)
4545: {
4546:   Mat_MPIAIJ     *b;
4548:   PetscMPIInt    size;

4551:   MPI_Comm_size(((PetscObject)B)->comm,&size);

4553:   PetscNewLog(B,Mat_MPIAIJ,&b);
4554:   B->data         = (void*)b;
4555:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
4556:   B->factor       = 0;
4557:   B->rmap.bs      = 1;
4558:   B->assembled    = PETSC_FALSE;
4559:   B->mapping      = 0;

4561:   B->insertmode      = NOT_SET_VALUES;
4562:   b->size            = size;
4563:   MPI_Comm_rank(((PetscObject)B)->comm,&b->rank);

4565:   /* build cache for off array entries formed */
4566:   MatStashCreate_Private(((PetscObject)B)->comm,1,&B->stash);
4567:   b->donotstash  = PETSC_FALSE;
4568:   b->colmap      = 0;
4569:   b->garray      = 0;
4570:   b->roworiented = PETSC_TRUE;

4572:   /* stuff used for matrix vector multiply */
4573:   b->lvec      = PETSC_NULL;
4574:   b->Mvctx     = PETSC_NULL;

4576:   /* stuff for MatGetRow() */
4577:   b->rowindices   = 0;
4578:   b->rowvalues    = 0;
4579:   b->getrowactive = PETSC_FALSE;


4582:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatStoreValues_C",
4583:                                      "MatStoreValues_MPIAIJ",
4584:                                      MatStoreValues_MPIAIJ);
4585:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatRetrieveValues_C",
4586:                                      "MatRetrieveValues_MPIAIJ",
4587:                                      MatRetrieveValues_MPIAIJ);
4588:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatGetDiagonalBlock_C",
4589:                                      "MatGetDiagonalBlock_MPIAIJ",
4590:                                      MatGetDiagonalBlock_MPIAIJ);
4591:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatIsTranspose_C",
4592:                                      "MatIsTranspose_MPIAIJ",
4593:                                      MatIsTranspose_MPIAIJ);
4594:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocation_C",
4595:                                      "MatMPIAIJSetPreallocation_MPIAIJ",
4596:                                      MatMPIAIJSetPreallocation_MPIAIJ);
4597:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",
4598:                                      "MatMPIAIJSetPreallocationCSR_MPIAIJ",
4599:                                      MatMPIAIJSetPreallocationCSR_MPIAIJ);
4600:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatDiagonalScaleLocal_C",
4601:                                      "MatDiagonalScaleLocal_MPIAIJ",
4602:                                      MatDiagonalScaleLocal_MPIAIJ);
4603:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpicsrperm_C",
4604:                                      "MatConvert_MPIAIJ_MPICSRPERM",
4605:                                       MatConvert_MPIAIJ_MPICSRPERM);
4606:   PetscObjectComposeFunctionDynamic((PetscObject)B,"MatConvert_mpiaij_mpicrl_C",
4607:                                      "MatConvert_MPIAIJ_MPICRL",
4608:                                       MatConvert_MPIAIJ_MPICRL);
4609:   PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
4610:   return(0);
4611: }

4616: /*@
4617:      MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
4618:          and "off-diagonal" part of the matrix in CSR format.

4620:    Collective on MPI_Comm

4622:    Input Parameters:
4623: +  comm - MPI communicator
4624: .  m - number of local rows (Cannot be PETSC_DECIDE)
4625: .  n - This value should be the same as the local size used in creating the 
4626:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4627:        calculated if N is given) For square matrices n is almost always m.
4628: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4629: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4630: .   i - row indices for "diagonal" portion of matrix
4631: .   j - column indices
4632: .   a - matrix values
4633: .   oi - row indices for "off-diagonal" portion of matrix
4634: .   oj - column indices
4635: -   oa - matrix values

4637:    Output Parameter:
4638: .   mat - the matrix

4640:    Level: advanced

4642:    Notes:
4643:        The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc.

4645:        The i and j indices are 0 based
4646:  
4647:        See MatCreateMPIAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix


4650: .keywords: matrix, aij, compressed row, sparse, parallel

4652: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4653:           MPIAIJ, MatCreateMPIAIJ(), MatCreateMPIAIJWithArrays()
4654: @*/
4655: PetscErrorCode  MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],
4656:                                                                 PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
4657: {
4659:   Mat_MPIAIJ     *maij;

4662:   if (m < 0) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4663:   if (i[0]) {
4664:     SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4665:   }
4666:   if (oi[0]) {
4667:     SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
4668:   }
4669:   MatCreate(comm,mat);
4670:   MatSetSizes(*mat,m,n,M,N);
4671:   MatSetType(*mat,MATMPIAIJ);
4672:   maij = (Mat_MPIAIJ*) (*mat)->data;
4673:   maij->donotstash     = PETSC_TRUE;
4674:   (*mat)->preallocated = PETSC_TRUE;

4676:   (*mat)->rmap.bs = (*mat)->cmap.bs = 1;
4677:   PetscMapSetUp(&(*mat)->rmap);
4678:   PetscMapSetUp(&(*mat)->cmap);

4680:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
4681:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap.N,oi,oj,oa,&maij->B);

4683:   MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
4684:   MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
4685:   MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
4686:   MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);

4688:   MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
4689:   MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
4690:   return(0);
4691: }

4693: /*
4694:     Special version for direct calls from Fortran 
4695: */
4696: #if defined(PETSC_HAVE_FORTRAN_CAPS)
4697: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
4698: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
4699: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
4700: #endif

4702: /* Change these macros so can be used in void function */
4703: #undef CHKERRQ
4704: #define CHKERRQ(ierr) CHKERRABORT(((PetscObject)mat)->comm,ierr) 
4705: #undef SETERRQ2
4706: #define SETERRQ2(ierr,b,c,d) CHKERRABORT(((PetscObject)mat)->comm,ierr) 
4707: #undef SETERRQ
4708: #define SETERRQ(ierr,b) CHKERRABORT(((PetscObject)mat)->comm,ierr) 

4713: void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
4714: {
4715:   Mat            mat = *mmat;
4716:   PetscInt       m = *mm, n = *mn;
4717:   InsertMode     addv = *maddv;
4718:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
4719:   PetscScalar    value;

4722:   MatPreallocated(mat);
4723:   if (mat->insertmode == NOT_SET_VALUES) {
4724:     mat->insertmode = addv;
4725:   }
4726: #if defined(PETSC_USE_DEBUG)
4727:   else if (mat->insertmode != addv) {
4728:     SETERRQ(PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
4729:   }
4730: #endif
4731:   {
4732:   PetscInt       i,j,rstart = mat->rmap.rstart,rend = mat->rmap.rend;
4733:   PetscInt       cstart = mat->cmap.rstart,cend = mat->cmap.rend,row,col;
4734:   PetscTruth     roworiented = aij->roworiented;

4736:   /* Some Variables required in the macro */
4737:   Mat            A = aij->A;
4738:   Mat_SeqAIJ     *a = (Mat_SeqAIJ*)A->data;
4739:   PetscInt       *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
4740:   PetscScalar    *aa = a->a;
4741:   PetscTruth     ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES))?PETSC_TRUE:PETSC_FALSE);
4742:   Mat            B = aij->B;
4743:   Mat_SeqAIJ     *b = (Mat_SeqAIJ*)B->data;
4744:   PetscInt       *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap.n,am = aij->A->rmap.n;
4745:   PetscScalar    *ba = b->a;

4747:   PetscInt       *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
4748:   PetscInt       nonew = a->nonew;
4749:   PetscScalar    *ap1,*ap2;

4752:   for (i=0; i<m; i++) {
4753:     if (im[i] < 0) continue;
4754: #if defined(PETSC_USE_DEBUG)
4755:     if (im[i] >= mat->rmap.N) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap.N-1);
4756: #endif
4757:     if (im[i] >= rstart && im[i] < rend) {
4758:       row      = im[i] - rstart;
4759:       lastcol1 = -1;
4760:       rp1      = aj + ai[row];
4761:       ap1      = aa + ai[row];
4762:       rmax1    = aimax[row];
4763:       nrow1    = ailen[row];
4764:       low1     = 0;
4765:       high1    = nrow1;
4766:       lastcol2 = -1;
4767:       rp2      = bj + bi[row];
4768:       ap2      = ba + bi[row];
4769:       rmax2    = bimax[row];
4770:       nrow2    = bilen[row];
4771:       low2     = 0;
4772:       high2    = nrow2;

4774:       for (j=0; j<n; j++) {
4775:         if (roworiented) value = v[i*n+j]; else value = v[i+j*m];
4776:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
4777:         if (in[j] >= cstart && in[j] < cend){
4778:           col = in[j] - cstart;
4779:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv);
4780:         } else if (in[j] < 0) continue;
4781: #if defined(PETSC_USE_DEBUG)
4782:         else if (in[j] >= mat->cmap.N) {SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap.N-1);}
4783: #endif
4784:         else {
4785:           if (mat->was_assembled) {
4786:             if (!aij->colmap) {
4787:               CreateColmap_MPIAIJ_Private(mat);
4788:             }
4789: #if defined (PETSC_USE_CTABLE)
4790:             PetscTableFind(aij->colmap,in[j]+1,&col);
4791:             col--;
4792: #else
4793:             col = aij->colmap[in[j]] - 1;
4794: #endif
4795:             if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
4796:               DisAssemble_MPIAIJ(mat);
4797:               col =  in[j];
4798:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
4799:               B = aij->B;
4800:               b = (Mat_SeqAIJ*)B->data;
4801:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
4802:               rp2      = bj + bi[row];
4803:               ap2      = ba + bi[row];
4804:               rmax2    = bimax[row];
4805:               nrow2    = bilen[row];
4806:               low2     = 0;
4807:               high2    = nrow2;
4808:               bm       = aij->B->rmap.n;
4809:               ba = b->a;
4810:             }
4811:           } else col = in[j];
4812:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv);
4813:         }
4814:       }
4815:     } else {
4816:       if (!aij->donotstash) {
4817:         if (roworiented) {
4818:           if (ignorezeroentries && v[i*n] == 0.0) continue;
4819:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n);
4820:         } else {
4821:           if (ignorezeroentries && v[i] == 0.0) continue;
4822:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m);
4823:         }
4824:       }
4825:     }
4826:   }}
4827:   PetscFunctionReturnVoid();
4828: }