Actual source code: mpiaij.c

petsc-3.7.2 2016-06-05
Report Typos and Errors
  2: #include <../src/mat/impls/aij/mpi/mpiaij.h>   /*I "petscmat.h" I*/
  3: #include <petsc/private/vecimpl.h>
  4: #include <petsc/private/isimpl.h>
  5: #include <petscblaslapack.h>
  6: #include <petscsf.h>

  8: /*MC
  9:    MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.

 11:    This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
 12:    and MATMPIAIJ otherwise.  As a result, for single process communicators,
 13:   MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
 14:   for communicators controlling multiple processes.  It is recommended that you call both of
 15:   the above preallocation routines for simplicity.

 17:    Options Database Keys:
 18: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()

 20:   Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when
 21:    enough exist.

 23:   Level: beginner

 25: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ
 26: M*/

 28: /*MC
 29:    MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.

 31:    This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
 32:    and MATMPIAIJCRL otherwise.  As a result, for single process communicators,
 33:    MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
 34:   for communicators controlling multiple processes.  It is recommended that you call both of
 35:   the above preallocation routines for simplicity.

 37:    Options Database Keys:
 38: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()

 40:   Level: beginner

 42: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
 43: M*/

 47: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
 48: {
 49:   PetscErrorCode  ierr;
 50:   Mat_MPIAIJ      *mat = (Mat_MPIAIJ*)M->data;
 51:   Mat_SeqAIJ      *a   = (Mat_SeqAIJ*)mat->A->data;
 52:   Mat_SeqAIJ      *b   = (Mat_SeqAIJ*)mat->B->data;
 53:   const PetscInt  *ia,*ib;
 54:   const MatScalar *aa,*bb;
 55:   PetscInt        na,nb,i,j,*rows,cnt=0,n0rows;
 56:   PetscInt        m = M->rmap->n,rstart = M->rmap->rstart;

 59:   *keptrows = 0;
 60:   ia        = a->i;
 61:   ib        = b->i;
 62:   for (i=0; i<m; i++) {
 63:     na = ia[i+1] - ia[i];
 64:     nb = ib[i+1] - ib[i];
 65:     if (!na && !nb) {
 66:       cnt++;
 67:       goto ok1;
 68:     }
 69:     aa = a->a + ia[i];
 70:     for (j=0; j<na; j++) {
 71:       if (aa[j] != 0.0) goto ok1;
 72:     }
 73:     bb = b->a + ib[i];
 74:     for (j=0; j <nb; j++) {
 75:       if (bb[j] != 0.0) goto ok1;
 76:     }
 77:     cnt++;
 78: ok1:;
 79:   }
 80:   MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
 81:   if (!n0rows) return(0);
 82:   PetscMalloc1(M->rmap->n-cnt,&rows);
 83:   cnt  = 0;
 84:   for (i=0; i<m; i++) {
 85:     na = ia[i+1] - ia[i];
 86:     nb = ib[i+1] - ib[i];
 87:     if (!na && !nb) continue;
 88:     aa = a->a + ia[i];
 89:     for (j=0; j<na;j++) {
 90:       if (aa[j] != 0.0) {
 91:         rows[cnt++] = rstart + i;
 92:         goto ok2;
 93:       }
 94:     }
 95:     bb = b->a + ib[i];
 96:     for (j=0; j<nb; j++) {
 97:       if (bb[j] != 0.0) {
 98:         rows[cnt++] = rstart + i;
 99:         goto ok2;
100:       }
101:     }
102: ok2:;
103:   }
104:   ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
105:   return(0);
106: }

110: PetscErrorCode  MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
111: {
112:   PetscErrorCode    ierr;
113:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*) Y->data;

116:   if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) {
117:     MatDiagonalSet(aij->A,D,is);
118:   } else {
119:     MatDiagonalSet_Default(Y,D,is);
120:   }
121:   return(0);
122: }


127: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
128: {
129:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)M->data;
131:   PetscInt       i,rstart,nrows,*rows;

134:   *zrows = NULL;
135:   MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
136:   MatGetOwnershipRange(M,&rstart,NULL);
137:   for (i=0; i<nrows; i++) rows[i] += rstart;
138:   ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
139:   return(0);
140: }

144: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
145: {
147:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)A->data;
148:   PetscInt       i,n,*garray = aij->garray;
149:   Mat_SeqAIJ     *a_aij = (Mat_SeqAIJ*) aij->A->data;
150:   Mat_SeqAIJ     *b_aij = (Mat_SeqAIJ*) aij->B->data;
151:   PetscReal      *work;

154:   MatGetSize(A,NULL,&n);
155:   PetscCalloc1(n,&work);
156:   if (type == NORM_2) {
157:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
158:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
159:     }
160:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
161:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
162:     }
163:   } else if (type == NORM_1) {
164:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
165:       work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
166:     }
167:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
168:       work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
169:     }
170:   } else if (type == NORM_INFINITY) {
171:     for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
172:       work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
173:     }
174:     for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
175:       work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
176:     }

178:   } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
179:   if (type == NORM_INFINITY) {
180:     MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
181:   } else {
182:     MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
183:   }
184:   PetscFree(work);
185:   if (type == NORM_2) {
186:     for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
187:   }
188:   return(0);
189: }

193: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
194: {
195:   Mat_MPIAIJ      *a  = (Mat_MPIAIJ*)A->data;
196:   IS              sis,gis;
197:   PetscErrorCode  ierr;
198:   const PetscInt  *isis,*igis;
199:   PetscInt        n,*iis,nsis,ngis,rstart,i;

202:   MatFindOffBlockDiagonalEntries(a->A,&sis);
203:   MatFindNonzeroRows(a->B,&gis);
204:   ISGetSize(gis,&ngis);
205:   ISGetSize(sis,&nsis);
206:   ISGetIndices(sis,&isis);
207:   ISGetIndices(gis,&igis);

209:   PetscMalloc1(ngis+nsis,&iis);
210:   PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));
211:   PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));
212:   n    = ngis + nsis;
213:   PetscSortRemoveDupsInt(&n,iis);
214:   MatGetOwnershipRange(A,&rstart,NULL);
215:   for (i=0; i<n; i++) iis[i] += rstart;
216:   ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);

218:   ISRestoreIndices(sis,&isis);
219:   ISRestoreIndices(gis,&igis);
220:   ISDestroy(&sis);
221:   ISDestroy(&gis);
222:   return(0);
223: }

227: /*
228:     Distributes a SeqAIJ matrix across a set of processes. Code stolen from
229:     MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.

231:     Only for square matrices

233:     Used by a preconditioner, hence PETSC_EXTERN
234: */
235: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
236: {
237:   PetscMPIInt    rank,size;
238:   PetscInt       *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
240:   Mat            mat;
241:   Mat_SeqAIJ     *gmata;
242:   PetscMPIInt    tag;
243:   MPI_Status     status;
244:   PetscBool      aij;
245:   MatScalar      *gmataa,*ao,*ad,*gmataarestore=0;

248:   MPI_Comm_rank(comm,&rank);
249:   MPI_Comm_size(comm,&size);
250:   if (!rank) {
251:     PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
252:     if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
253:   }
254:   if (reuse == MAT_INITIAL_MATRIX) {
255:     MatCreate(comm,&mat);
256:     MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
257:     MatGetBlockSizes(gmat,&bses[0],&bses[1]);
258:     MPI_Bcast(bses,2,MPIU_INT,0,comm);
259:     MatSetBlockSizes(mat,bses[0],bses[1]);
260:     MatSetType(mat,MATAIJ);
261:     PetscMalloc1(size+1,&rowners);
262:     PetscMalloc2(m,&dlens,m,&olens);
263:     MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

265:     rowners[0] = 0;
266:     for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
267:     rstart = rowners[rank];
268:     rend   = rowners[rank+1];
269:     PetscObjectGetNewTag((PetscObject)mat,&tag);
270:     if (!rank) {
271:       gmata = (Mat_SeqAIJ*) gmat->data;
272:       /* send row lengths to all processors */
273:       for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
274:       for (i=1; i<size; i++) {
275:         MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
276:       }
277:       /* determine number diagonal and off-diagonal counts */
278:       PetscMemzero(olens,m*sizeof(PetscInt));
279:       PetscCalloc1(m,&ld);
280:       jj   = 0;
281:       for (i=0; i<m; i++) {
282:         for (j=0; j<dlens[i]; j++) {
283:           if (gmata->j[jj] < rstart) ld[i]++;
284:           if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
285:           jj++;
286:         }
287:       }
288:       /* send column indices to other processes */
289:       for (i=1; i<size; i++) {
290:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
291:         MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
292:         MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
293:       }

295:       /* send numerical values to other processes */
296:       for (i=1; i<size; i++) {
297:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
298:         MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
299:       }
300:       gmataa = gmata->a;
301:       gmataj = gmata->j;

303:     } else {
304:       /* receive row lengths */
305:       MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
306:       /* receive column indices */
307:       MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
308:       PetscMalloc2(nz,&gmataa,nz,&gmataj);
309:       MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
310:       /* determine number diagonal and off-diagonal counts */
311:       PetscMemzero(olens,m*sizeof(PetscInt));
312:       PetscCalloc1(m,&ld);
313:       jj   = 0;
314:       for (i=0; i<m; i++) {
315:         for (j=0; j<dlens[i]; j++) {
316:           if (gmataj[jj] < rstart) ld[i]++;
317:           if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
318:           jj++;
319:         }
320:       }
321:       /* receive numerical values */
322:       PetscMemzero(gmataa,nz*sizeof(PetscScalar));
323:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
324:     }
325:     /* set preallocation */
326:     for (i=0; i<m; i++) {
327:       dlens[i] -= olens[i];
328:     }
329:     MatSeqAIJSetPreallocation(mat,0,dlens);
330:     MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);

332:     for (i=0; i<m; i++) {
333:       dlens[i] += olens[i];
334:     }
335:     cnt = 0;
336:     for (i=0; i<m; i++) {
337:       row  = rstart + i;
338:       MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
339:       cnt += dlens[i];
340:     }
341:     if (rank) {
342:       PetscFree2(gmataa,gmataj);
343:     }
344:     PetscFree2(dlens,olens);
345:     PetscFree(rowners);

347:     ((Mat_MPIAIJ*)(mat->data))->ld = ld;

349:     *inmat = mat;
350:   } else {   /* column indices are already set; only need to move over numerical values from process 0 */
351:     Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
352:     Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
353:     mat  = *inmat;
354:     PetscObjectGetNewTag((PetscObject)mat,&tag);
355:     if (!rank) {
356:       /* send numerical values to other processes */
357:       gmata  = (Mat_SeqAIJ*) gmat->data;
358:       MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
359:       gmataa = gmata->a;
360:       for (i=1; i<size; i++) {
361:         nz   = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
362:         MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
363:       }
364:       nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
365:     } else {
366:       /* receive numerical values from process 0*/
367:       nz   = Ad->nz + Ao->nz;
368:       PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
369:       MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
370:     }
371:     /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
372:     ld = ((Mat_MPIAIJ*)(mat->data))->ld;
373:     ad = Ad->a;
374:     ao = Ao->a;
375:     if (mat->rmap->n) {
376:       i  = 0;
377:       nz = ld[i];                                   PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
378:       nz = Ad->i[i+1] - Ad->i[i];                   PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
379:     }
380:     for (i=1; i<mat->rmap->n; i++) {
381:       nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
382:       nz = Ad->i[i+1] - Ad->i[i];                   PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
383:     }
384:     i--;
385:     if (mat->rmap->n) {
386:       nz = Ao->i[i+1] - Ao->i[i] - ld[i];           PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));
387:     }
388:     if (rank) {
389:       PetscFree(gmataarestore);
390:     }
391:   }
392:   MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
393:   MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
394:   return(0);
395: }

397: /*
398:   Local utility routine that creates a mapping from the global column
399: number to the local number in the off-diagonal part of the local
400: storage of the matrix.  When PETSC_USE_CTABLE is used this is scalable at
401: a slightly higher hash table cost; without it it is not scalable (each processor
402: has an order N integer array but is fast to acess.
403: */
406: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
407: {
408:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
410:   PetscInt       n = aij->B->cmap->n,i;

413:   if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
414: #if defined(PETSC_USE_CTABLE)
415:   PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
416:   for (i=0; i<n; i++) {
417:     PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
418:   }
419: #else
420:   PetscCalloc1(mat->cmap->N+1,&aij->colmap);
421:   PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
422:   for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
423: #endif
424:   return(0);
425: }

427: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol)     \
428: { \
429:     if (col <= lastcol1)  low1 = 0;     \
430:     else                 high1 = nrow1; \
431:     lastcol1 = col;\
432:     while (high1-low1 > 5) { \
433:       t = (low1+high1)/2; \
434:       if (rp1[t] > col) high1 = t; \
435:       else              low1  = t; \
436:     } \
437:       for (_i=low1; _i<high1; _i++) { \
438:         if (rp1[_i] > col) break; \
439:         if (rp1[_i] == col) { \
440:           if (addv == ADD_VALUES) ap1[_i] += value;   \
441:           else                    ap1[_i] = value; \
442:           goto a_noinsert; \
443:         } \
444:       }  \
445:       if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
446:       if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;}                \
447:       if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
448:       MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
449:       N = nrow1++ - 1; a->nz++; high1++; \
450:       /* shift up all the later entries in this row */ \
451:       for (ii=N; ii>=_i; ii--) { \
452:         rp1[ii+1] = rp1[ii]; \
453:         ap1[ii+1] = ap1[ii]; \
454:       } \
455:       rp1[_i] = col;  \
456:       ap1[_i] = value;  \
457:       A->nonzerostate++;\
458:       a_noinsert: ; \
459:       ailen[row] = nrow1; \
460: }


463: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
464:   { \
465:     if (col <= lastcol2) low2 = 0;                        \
466:     else high2 = nrow2;                                   \
467:     lastcol2 = col;                                       \
468:     while (high2-low2 > 5) {                              \
469:       t = (low2+high2)/2;                                 \
470:       if (rp2[t] > col) high2 = t;                        \
471:       else             low2  = t;                         \
472:     }                                                     \
473:     for (_i=low2; _i<high2; _i++) {                       \
474:       if (rp2[_i] > col) break;                           \
475:       if (rp2[_i] == col) {                               \
476:         if (addv == ADD_VALUES) ap2[_i] += value;         \
477:         else                    ap2[_i] = value;          \
478:         goto b_noinsert;                                  \
479:       }                                                   \
480:     }                                                     \
481:     if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
482:     if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;}                        \
483:     if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
484:     MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
485:     N = nrow2++ - 1; b->nz++; high2++;                    \
486:     /* shift up all the later entries in this row */      \
487:     for (ii=N; ii>=_i; ii--) {                            \
488:       rp2[ii+1] = rp2[ii];                                \
489:       ap2[ii+1] = ap2[ii];                                \
490:     }                                                     \
491:     rp2[_i] = col;                                        \
492:     ap2[_i] = value;                                      \
493:     B->nonzerostate++;                                    \
494:     b_noinsert: ;                                         \
495:     bilen[row] = nrow2;                                   \
496:   }

500: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
501: {
502:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)A->data;
503:   Mat_SeqAIJ     *a   = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
505:   PetscInt       l,*garray = mat->garray,diag;

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

510:   /* find size of row to the left of the diagonal part */
511:   MatGetOwnershipRange(A,&diag,0);
512:   row  = row - diag;
513:   for (l=0; l<b->i[row+1]-b->i[row]; l++) {
514:     if (garray[b->j[b->i[row]+l]] > diag) break;
515:   }
516:   PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));

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

521:   /* right of diagonal part */
522:   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));
523:   return(0);
524: }

528: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
529: {
530:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
531:   PetscScalar    value;
533:   PetscInt       i,j,rstart  = mat->rmap->rstart,rend = mat->rmap->rend;
534:   PetscInt       cstart      = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
535:   PetscBool      roworiented = aij->roworiented;

537:   /* Some Variables required in the macro */
538:   Mat        A                 = aij->A;
539:   Mat_SeqAIJ *a                = (Mat_SeqAIJ*)A->data;
540:   PetscInt   *aimax            = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
541:   MatScalar  *aa               = a->a;
542:   PetscBool  ignorezeroentries = a->ignorezeroentries;
543:   Mat        B                 = aij->B;
544:   Mat_SeqAIJ *b                = (Mat_SeqAIJ*)B->data;
545:   PetscInt   *bimax            = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
546:   MatScalar  *ba               = b->a;

548:   PetscInt  *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
549:   PetscInt  nonew;
550:   MatScalar *ap1,*ap2;

553:   for (i=0; i<m; i++) {
554:     if (im[i] < 0) continue;
555: #if defined(PETSC_USE_DEBUG)
556:     if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
557: #endif
558:     if (im[i] >= rstart && im[i] < rend) {
559:       row      = im[i] - rstart;
560:       lastcol1 = -1;
561:       rp1      = aj + ai[row];
562:       ap1      = aa + ai[row];
563:       rmax1    = aimax[row];
564:       nrow1    = ailen[row];
565:       low1     = 0;
566:       high1    = nrow1;
567:       lastcol2 = -1;
568:       rp2      = bj + bi[row];
569:       ap2      = ba + bi[row];
570:       rmax2    = bimax[row];
571:       nrow2    = bilen[row];
572:       low2     = 0;
573:       high2    = nrow2;

575:       for (j=0; j<n; j++) {
576:         if (roworiented) value = v[i*n+j];
577:         else             value = v[i+j*m];
578:         if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
579:         if (in[j] >= cstart && in[j] < cend) {
580:           col   = in[j] - cstart;
581:           nonew = a->nonew;
582:           MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
583:         } else if (in[j] < 0) continue;
584: #if defined(PETSC_USE_DEBUG)
585:         else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
586: #endif
587:         else {
588:           if (mat->was_assembled) {
589:             if (!aij->colmap) {
590:               MatCreateColmap_MPIAIJ_Private(mat);
591:             }
592: #if defined(PETSC_USE_CTABLE)
593:             PetscTableFind(aij->colmap,in[j]+1,&col);
594:             col--;
595: #else
596:             col = aij->colmap[in[j]] - 1;
597: #endif
598:             if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
599:               MatDisAssemble_MPIAIJ(mat);
600:               col  =  in[j];
601:               /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
602:               B     = aij->B;
603:               b     = (Mat_SeqAIJ*)B->data;
604:               bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
605:               rp2   = bj + bi[row];
606:               ap2   = ba + bi[row];
607:               rmax2 = bimax[row];
608:               nrow2 = bilen[row];
609:               low2  = 0;
610:               high2 = nrow2;
611:               bm    = aij->B->rmap->n;
612:               ba    = b->a;
613:             } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]);
614:           } else col = in[j];
615:           nonew = b->nonew;
616:           MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
617:         }
618:       }
619:     } else {
620:       if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
621:       if (!aij->donotstash) {
622:         mat->assembled = PETSC_FALSE;
623:         if (roworiented) {
624:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
625:         } else {
626:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
627:         }
628:       }
629:     }
630:   }
631:   return(0);
632: }

636: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
637: {
638:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
640:   PetscInt       i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
641:   PetscInt       cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;

644:   for (i=0; i<m; i++) {
645:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
646:     if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
647:     if (idxm[i] >= rstart && idxm[i] < rend) {
648:       row = idxm[i] - rstart;
649:       for (j=0; j<n; j++) {
650:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
651:         if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
652:         if (idxn[j] >= cstart && idxn[j] < cend) {
653:           col  = idxn[j] - cstart;
654:           MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
655:         } else {
656:           if (!aij->colmap) {
657:             MatCreateColmap_MPIAIJ_Private(mat);
658:           }
659: #if defined(PETSC_USE_CTABLE)
660:           PetscTableFind(aij->colmap,idxn[j]+1,&col);
661:           col--;
662: #else
663:           col = aij->colmap[idxn[j]] - 1;
664: #endif
665:           if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
666:           else {
667:             MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
668:           }
669:         }
670:       }
671:     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
672:   }
673:   return(0);
674: }

676: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);

680: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
681: {
682:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
684:   PetscInt       nstash,reallocs;

687:   if (aij->donotstash || mat->nooffprocentries) return(0);

689:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
690:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
691:   PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
692:   return(0);
693: }

697: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
698: {
699:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
700:   Mat_SeqAIJ     *a   = (Mat_SeqAIJ*)aij->A->data;
702:   PetscMPIInt    n;
703:   PetscInt       i,j,rstart,ncols,flg;
704:   PetscInt       *row,*col;
705:   PetscBool      other_disassembled;
706:   PetscScalar    *val;

708:   /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */

711:   if (!aij->donotstash && !mat->nooffprocentries) {
712:     while (1) {
713:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
714:       if (!flg) break;

716:       for (i=0; i<n; ) {
717:         /* Now identify the consecutive vals belonging to the same row */
718:         for (j=i,rstart=row[j]; j<n; j++) {
719:           if (row[j] != rstart) break;
720:         }
721:         if (j < n) ncols = j-i;
722:         else       ncols = n-i;
723:         /* Now assemble all these values with a single function call */
724:         MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);

726:         i = j;
727:       }
728:     }
729:     MatStashScatterEnd_Private(&mat->stash);
730:   }
731:   MatAssemblyBegin(aij->A,mode);
732:   MatAssemblyEnd(aij->A,mode);

734:   /* determine if any processor has disassembled, if so we must
735:      also disassemble ourselfs, in order that we may reassemble. */
736:   /*
737:      if nonzero structure of submatrix B cannot change then we know that
738:      no processor disassembled thus we can skip this stuff
739:   */
740:   if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
741:     MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
742:     if (mat->was_assembled && !other_disassembled) {
743:       MatDisAssemble_MPIAIJ(mat);
744:     }
745:   }
746:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
747:     MatSetUpMultiply_MPIAIJ(mat);
748:   }
749:   MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
750:   MatAssemblyBegin(aij->B,mode);
751:   MatAssemblyEnd(aij->B,mode);

753:   PetscFree2(aij->rowvalues,aij->rowindices);

755:   aij->rowvalues = 0;

757:   VecDestroy(&aij->diag);
758:   if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;

760:   /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
761:   if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
762:     PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
763:     MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
764:   }
765:   return(0);
766: }

770: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
771: {
772:   Mat_MPIAIJ     *l = (Mat_MPIAIJ*)A->data;

776:   MatZeroEntries(l->A);
777:   MatZeroEntries(l->B);
778:   return(0);
779: }

783: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
784: {
785:   Mat_MPIAIJ    *mat    = (Mat_MPIAIJ *) A->data;
786:   PetscInt      *owners = A->rmap->range;
787:   PetscInt       n      = A->rmap->n;
788:   PetscSF        sf;
789:   PetscInt      *lrows;
790:   PetscSFNode   *rrows;
791:   PetscInt       r, p = 0, len = 0;

795:   /* Create SF where leaves are input rows and roots are owned rows */
796:   PetscMalloc1(n, &lrows);
797:   for (r = 0; r < n; ++r) lrows[r] = -1;
798:   if (!A->nooffproczerorows) {PetscMalloc1(N, &rrows);}
799:   for (r = 0; r < N; ++r) {
800:     const PetscInt idx   = rows[r];
801:     if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
802:     if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
803:       PetscLayoutFindOwner(A->rmap,idx,&p);
804:     }
805:     if (A->nooffproczerorows) {
806:       if (p != mat->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,mat->rank);
807:       lrows[len++] = idx - owners[p];
808:     } else {
809:       rrows[r].rank = p;
810:       rrows[r].index = rows[r] - owners[p];
811:     }
812:   }
813:   if (!A->nooffproczerorows) {
814:     PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
815:     PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
816:     /* Collect flags for rows to be zeroed */
817:     PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);
818:     PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);
819:     PetscSFDestroy(&sf);
820:     /* Compress and put in row numbers */
821:     for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
822:   }
823:   /* fix right hand side if needed */
824:   if (x && b) {
825:     const PetscScalar *xx;
826:     PetscScalar       *bb;

828:     VecGetArrayRead(x, &xx);
829:     VecGetArray(b, &bb);
830:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
831:     VecRestoreArrayRead(x, &xx);
832:     VecRestoreArray(b, &bb);
833:   }
834:   /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/
835:   MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
836:   if ((diag != 0.0) && (mat->A->rmap->N == mat->A->cmap->N)) {
837:     MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
838:   } else if (diag != 0.0) {
839:     MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
840:     if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
841:     for (r = 0; r < len; ++r) {
842:       const PetscInt row = lrows[r] + A->rmap->rstart;
843:       MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
844:     }
845:     MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);
846:     MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);
847:   } else {
848:     MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
849:   }
850:   PetscFree(lrows);

852:   /* only change matrix nonzero state if pattern was allowed to be changed */
853:   if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) {
854:     PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate;
855:     MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
856:   }
857:   return(0);
858: }

862: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
863: {
864:   Mat_MPIAIJ        *l = (Mat_MPIAIJ*)A->data;
865:   PetscErrorCode    ierr;
866:   PetscMPIInt       n = A->rmap->n;
867:   PetscInt          i,j,r,m,p = 0,len = 0;
868:   PetscInt          *lrows,*owners = A->rmap->range;
869:   PetscSFNode       *rrows;
870:   PetscSF           sf;
871:   const PetscScalar *xx;
872:   PetscScalar       *bb,*mask;
873:   Vec               xmask,lmask;
874:   Mat_SeqAIJ        *aij = (Mat_SeqAIJ*)l->B->data;
875:   const PetscInt    *aj, *ii,*ridx;
876:   PetscScalar       *aa;

879:   /* Create SF where leaves are input rows and roots are owned rows */
880:   PetscMalloc1(n, &lrows);
881:   for (r = 0; r < n; ++r) lrows[r] = -1;
882:   PetscMalloc1(N, &rrows);
883:   for (r = 0; r < N; ++r) {
884:     const PetscInt idx   = rows[r];
885:     if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
886:     if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
887:       PetscLayoutFindOwner(A->rmap,idx,&p);
888:     }
889:     rrows[r].rank  = p;
890:     rrows[r].index = rows[r] - owners[p];
891:   }
892:   PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
893:   PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
894:   /* Collect flags for rows to be zeroed */
895:   PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
896:   PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
897:   PetscSFDestroy(&sf);
898:   /* Compress and put in row numbers */
899:   for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
900:   /* zero diagonal part of matrix */
901:   MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
902:   /* handle off diagonal part of matrix */
903:   MatCreateVecs(A,&xmask,NULL);
904:   VecDuplicate(l->lvec,&lmask);
905:   VecGetArray(xmask,&bb);
906:   for (i=0; i<len; i++) bb[lrows[i]] = 1;
907:   VecRestoreArray(xmask,&bb);
908:   VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
909:   VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
910:   VecDestroy(&xmask);
911:   if (x) {
912:     VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
913:     VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
914:     VecGetArrayRead(l->lvec,&xx);
915:     VecGetArray(b,&bb);
916:   }
917:   VecGetArray(lmask,&mask);
918:   /* remove zeroed rows of off diagonal matrix */
919:   ii = aij->i;
920:   for (i=0; i<len; i++) {
921:     PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));
922:   }
923:   /* loop over all elements of off process part of matrix zeroing removed columns*/
924:   if (aij->compressedrow.use) {
925:     m    = aij->compressedrow.nrows;
926:     ii   = aij->compressedrow.i;
927:     ridx = aij->compressedrow.rindex;
928:     for (i=0; i<m; i++) {
929:       n  = ii[i+1] - ii[i];
930:       aj = aij->j + ii[i];
931:       aa = aij->a + ii[i];

933:       for (j=0; j<n; j++) {
934:         if (PetscAbsScalar(mask[*aj])) {
935:           if (b) bb[*ridx] -= *aa*xx[*aj];
936:           *aa = 0.0;
937:         }
938:         aa++;
939:         aj++;
940:       }
941:       ridx++;
942:     }
943:   } else { /* do not use compressed row format */
944:     m = l->B->rmap->n;
945:     for (i=0; i<m; i++) {
946:       n  = ii[i+1] - ii[i];
947:       aj = aij->j + ii[i];
948:       aa = aij->a + ii[i];
949:       for (j=0; j<n; j++) {
950:         if (PetscAbsScalar(mask[*aj])) {
951:           if (b) bb[i] -= *aa*xx[*aj];
952:           *aa = 0.0;
953:         }
954:         aa++;
955:         aj++;
956:       }
957:     }
958:   }
959:   if (x) {
960:     VecRestoreArray(b,&bb);
961:     VecRestoreArrayRead(l->lvec,&xx);
962:   }
963:   VecRestoreArray(lmask,&mask);
964:   VecDestroy(&lmask);
965:   PetscFree(lrows);

967:   /* only change matrix nonzero state if pattern was allowed to be changed */
968:   if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
969:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
970:     MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
971:   }
972:   return(0);
973: }

977: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
978: {
979:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
981:   PetscInt       nt;

984:   VecGetLocalSize(xx,&nt);
985:   if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
986:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
987:   (*a->A->ops->mult)(a->A,xx,yy);
988:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
989:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
990:   return(0);
991: }

995: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
996: {
997:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1001:   MatMultDiagonalBlock(a->A,bb,xx);
1002:   return(0);
1003: }

1007: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1008: {
1009:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1013:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1014:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
1015:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1016:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1017:   return(0);
1018: }

1022: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1023: {
1024:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
1026:   PetscBool      merged;

1029:   VecScatterGetMerged(a->Mvctx,&merged);
1030:   /* do nondiagonal part */
1031:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1032:   if (!merged) {
1033:     /* send it on its way */
1034:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1035:     /* do local part */
1036:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1037:     /* receive remote parts: note this assumes the values are not actually */
1038:     /* added in yy until the next line, */
1039:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1040:   } else {
1041:     /* do local part */
1042:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1043:     /* send it on its way */
1044:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1045:     /* values actually were received in the Begin() but we need to call this nop */
1046:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1047:   }
1048:   return(0);
1049: }

1053: PetscErrorCode  MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool  *f)
1054: {
1055:   MPI_Comm       comm;
1056:   Mat_MPIAIJ     *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1057:   Mat            Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1058:   IS             Me,Notme;
1060:   PetscInt       M,N,first,last,*notme,i;
1061:   PetscMPIInt    size;

1064:   /* Easy test: symmetric diagonal block */
1065:   Bij  = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1066:   MatIsTranspose(Adia,Bdia,tol,f);
1067:   if (!*f) return(0);
1068:   PetscObjectGetComm((PetscObject)Amat,&comm);
1069:   MPI_Comm_size(comm,&size);
1070:   if (size == 1) return(0);

1072:   /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */
1073:   MatGetSize(Amat,&M,&N);
1074:   MatGetOwnershipRange(Amat,&first,&last);
1075:   PetscMalloc1(N-last+first,&notme);
1076:   for (i=0; i<first; i++) notme[i] = i;
1077:   for (i=last; i<M; i++) notme[i-last+first] = i;
1078:   ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1079:   ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1080:   MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1081:   Aoff = Aoffs[0];
1082:   MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1083:   Boff = Boffs[0];
1084:   MatIsTranspose(Aoff,Boff,tol,f);
1085:   MatDestroyMatrices(1,&Aoffs);
1086:   MatDestroyMatrices(1,&Boffs);
1087:   ISDestroy(&Me);
1088:   ISDestroy(&Notme);
1089:   PetscFree(notme);
1090:   return(0);
1091: }

1095: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1096: {
1097:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1101:   /* do nondiagonal part */
1102:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1103:   /* send it on its way */
1104:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1105:   /* do local part */
1106:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1107:   /* receive remote parts */
1108:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1109:   return(0);
1110: }

1112: /*
1113:   This only works correctly for square matrices where the subblock A->A is the
1114:    diagonal block
1115: */
1118: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1119: {
1121:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1124:   if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1125:   if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1126:   MatGetDiagonal(a->A,v);
1127:   return(0);
1128: }

1132: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1133: {
1134:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1138:   MatScale(a->A,aa);
1139:   MatScale(a->B,aa);
1140:   return(0);
1141: }

1145: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1146: {
1147:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

1151: #if defined(PETSC_USE_LOG)
1152:   PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1153: #endif
1154:   MatStashDestroy_Private(&mat->stash);
1155:   VecDestroy(&aij->diag);
1156:   MatDestroy(&aij->A);
1157:   MatDestroy(&aij->B);
1158: #if defined(PETSC_USE_CTABLE)
1159:   PetscTableDestroy(&aij->colmap);
1160: #else
1161:   PetscFree(aij->colmap);
1162: #endif
1163:   PetscFree(aij->garray);
1164:   VecDestroy(&aij->lvec);
1165:   VecScatterDestroy(&aij->Mvctx);
1166:   PetscFree2(aij->rowvalues,aij->rowindices);
1167:   PetscFree(aij->ld);
1168:   PetscFree(mat->data);

1170:   PetscObjectChangeTypeName((PetscObject)mat,0);
1171:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1172:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1173:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);
1174:   PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1175:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1176:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1177:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1178:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1179: #if defined(PETSC_HAVE_ELEMENTAL)
1180:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1181: #endif
1182:   return(0);
1183: }

1187: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1188: {
1189:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
1190:   Mat_SeqAIJ     *A   = (Mat_SeqAIJ*)aij->A->data;
1191:   Mat_SeqAIJ     *B   = (Mat_SeqAIJ*)aij->B->data;
1193:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag;
1194:   int            fd;
1195:   PetscInt       nz,header[4],*row_lengths,*range=0,rlen,i;
1196:   PetscInt       nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0;
1197:   PetscScalar    *column_values;
1198:   PetscInt       message_count,flowcontrolcount;
1199:   FILE           *file;

1202:   MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1203:   MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1204:   nz   = A->nz + B->nz;
1205:   PetscViewerBinaryGetDescriptor(viewer,&fd);
1206:   if (!rank) {
1207:     header[0] = MAT_FILE_CLASSID;
1208:     header[1] = mat->rmap->N;
1209:     header[2] = mat->cmap->N;

1211:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1212:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1213:     /* get largest number of rows any processor has */
1214:     rlen  = mat->rmap->n;
1215:     range = mat->rmap->range;
1216:     for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]);
1217:   } else {
1218:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1219:     rlen = mat->rmap->n;
1220:   }

1222:   /* load up the local row counts */
1223:   PetscMalloc1(rlen+1,&row_lengths);
1224:   for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];

1226:   /* store the row lengths to the file */
1227:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1228:   if (!rank) {
1229:     PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1230:     for (i=1; i<size; i++) {
1231:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1232:       rlen = range[i+1] - range[i];
1233:       MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1234:       PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1235:     }
1236:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1237:   } else {
1238:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1239:     MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1240:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1241:   }
1242:   PetscFree(row_lengths);

1244:   /* load up the local column indices */
1245:   nzmax = nz; /* th processor needs space a largest processor needs */
1246:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1247:   PetscMalloc1(nzmax+1,&column_indices);
1248:   cnt   = 0;
1249:   for (i=0; i<mat->rmap->n; i++) {
1250:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1251:       if ((col = garray[B->j[j]]) > cstart) break;
1252:       column_indices[cnt++] = col;
1253:     }
1254:     for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart;
1255:     for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]];
1256:   }
1257:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

1259:   /* store the column indices to the file */
1260:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1261:   if (!rank) {
1262:     MPI_Status status;
1263:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1264:     for (i=1; i<size; i++) {
1265:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1266:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1267:       if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1268:       MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1269:       PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1270:     }
1271:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1272:   } else {
1273:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1274:     MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1275:     MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1276:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1277:   }
1278:   PetscFree(column_indices);

1280:   /* load up the local column values */
1281:   PetscMalloc1(nzmax+1,&column_values);
1282:   cnt  = 0;
1283:   for (i=0; i<mat->rmap->n; i++) {
1284:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1285:       if (garray[B->j[j]] > cstart) break;
1286:       column_values[cnt++] = B->a[j];
1287:     }
1288:     for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k];
1289:     for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j];
1290:   }
1291:   if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);

1293:   /* store the column values to the file */
1294:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1295:   if (!rank) {
1296:     MPI_Status status;
1297:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1298:     for (i=1; i<size; i++) {
1299:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1300:       MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1301:       if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1302:       MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));
1303:       PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1304:     }
1305:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1306:   } else {
1307:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1308:     MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1309:     MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1310:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1311:   }
1312:   PetscFree(column_values);

1314:   PetscViewerBinaryGetInfoPointer(viewer,&file);
1315:   if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs));
1316:   return(0);
1317: }

1319: #include <petscdraw.h>
1322: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1323: {
1324:   Mat_MPIAIJ        *aij = (Mat_MPIAIJ*)mat->data;
1325:   PetscErrorCode    ierr;
1326:   PetscMPIInt       rank = aij->rank,size = aij->size;
1327:   PetscBool         isdraw,iascii,isbinary;
1328:   PetscViewer       sviewer;
1329:   PetscViewerFormat format;

1332:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1333:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1334:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1335:   if (iascii) {
1336:     PetscViewerGetFormat(viewer,&format);
1337:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1338:       MatInfo   info;
1339:       PetscBool inodes;

1341:       MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1342:       MatGetInfo(mat,MAT_LOCAL,&info);
1343:       MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1344:       PetscViewerASCIIPushSynchronized(viewer);
1345:       if (!inodes) {
1346:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n",
1347:                                                   rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1348:       } else {
1349:         PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n",
1350:                                                   rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);
1351:       }
1352:       MatGetInfo(aij->A,MAT_LOCAL,&info);
1353:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1354:       MatGetInfo(aij->B,MAT_LOCAL,&info);
1355:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1356:       PetscViewerFlush(viewer);
1357:       PetscViewerASCIIPopSynchronized(viewer);
1358:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1359:       VecScatterView(aij->Mvctx,viewer);
1360:       return(0);
1361:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
1362:       PetscInt inodecount,inodelimit,*inodes;
1363:       MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1364:       if (inodes) {
1365:         PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1366:       } else {
1367:         PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1368:       }
1369:       return(0);
1370:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1371:       return(0);
1372:     }
1373:   } else if (isbinary) {
1374:     if (size == 1) {
1375:       PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1376:       MatView(aij->A,viewer);
1377:     } else {
1378:       MatView_MPIAIJ_Binary(mat,viewer);
1379:     }
1380:     return(0);
1381:   } else if (isdraw) {
1382:     PetscDraw draw;
1383:     PetscBool isnull;
1384:     PetscViewerDrawGetDraw(viewer,0,&draw);
1385:     PetscDrawIsNull(draw,&isnull);
1386:     if (isnull) return(0);
1387:   }

1389:   {
1390:     /* assemble the entire matrix onto first processor. */
1391:     Mat        A;
1392:     Mat_SeqAIJ *Aloc;
1393:     PetscInt   M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct;
1394:     MatScalar  *a;

1396:     MatCreate(PetscObjectComm((PetscObject)mat),&A);
1397:     if (!rank) {
1398:       MatSetSizes(A,M,N,M,N);
1399:     } else {
1400:       MatSetSizes(A,0,0,M,N);
1401:     }
1402:     /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */
1403:     MatSetType(A,MATMPIAIJ);
1404:     MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);
1405:     MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1406:     PetscLogObjectParent((PetscObject)mat,(PetscObject)A);

1408:     /* copy over the A part */
1409:     Aloc = (Mat_SeqAIJ*)aij->A->data;
1410:     m    = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1411:     row  = mat->rmap->rstart;
1412:     for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart;
1413:     for (i=0; i<m; i++) {
1414:       MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);
1415:       row++;
1416:       a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i];
1417:     }
1418:     aj = Aloc->j;
1419:     for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart;

1421:     /* copy over the B part */
1422:     Aloc = (Mat_SeqAIJ*)aij->B->data;
1423:     m    = aij->B->rmap->n;  ai = Aloc->i; aj = Aloc->j; a = Aloc->a;
1424:     row  = mat->rmap->rstart;
1425:     PetscMalloc1(ai[m]+1,&cols);
1426:     ct   = cols;
1427:     for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]];
1428:     for (i=0; i<m; i++) {
1429:       MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);
1430:       row++;
1431:       a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i];
1432:     }
1433:     PetscFree(ct);
1434:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1435:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1436:     /*
1437:        Everyone has to call to draw the matrix since the graphics waits are
1438:        synchronized across all processors that share the PetscDraw object
1439:     */
1440:     PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1441:     if (!rank) {
1442:       PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);
1443:       MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);
1444:     }
1445:     PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1446:     PetscViewerFlush(viewer);
1447:     MatDestroy(&A);
1448:   }
1449:   return(0);
1450: }

1454: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1455: {
1457:   PetscBool      iascii,isdraw,issocket,isbinary;

1460:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1461:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1462:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1463:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1464:   if (iascii || isdraw || isbinary || issocket) {
1465:     MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1466:   }
1467:   return(0);
1468: }

1472: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1473: {
1474:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1476:   Vec            bb1 = 0;
1477:   PetscBool      hasop;

1480:   if (flag == SOR_APPLY_UPPER) {
1481:     (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1482:     return(0);
1483:   }

1485:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1486:     VecDuplicate(bb,&bb1);
1487:   }

1489:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1490:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1491:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1492:       its--;
1493:     }

1495:     while (its--) {
1496:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1497:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1499:       /* update rhs: bb1 = bb - B*x */
1500:       VecScale(mat->lvec,-1.0);
1501:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1503:       /* local sweep */
1504:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1505:     }
1506:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1507:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1508:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1509:       its--;
1510:     }
1511:     while (its--) {
1512:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1513:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1515:       /* update rhs: bb1 = bb - B*x */
1516:       VecScale(mat->lvec,-1.0);
1517:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1519:       /* local sweep */
1520:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1521:     }
1522:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1523:     if (flag & SOR_ZERO_INITIAL_GUESS) {
1524:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1525:       its--;
1526:     }
1527:     while (its--) {
1528:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1529:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

1531:       /* update rhs: bb1 = bb - B*x */
1532:       VecScale(mat->lvec,-1.0);
1533:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

1535:       /* local sweep */
1536:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1537:     }
1538:   } else if (flag & SOR_EISENSTAT) {
1539:     Vec xx1;

1541:     VecDuplicate(bb,&xx1);
1542:     (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);

1544:     VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1545:     VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1546:     if (!mat->diag) {
1547:       MatCreateVecs(matin,&mat->diag,NULL);
1548:       MatGetDiagonal(matin,mat->diag);
1549:     }
1550:     MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1551:     if (hasop) {
1552:       MatMultDiagonalBlock(matin,xx,bb1);
1553:     } else {
1554:       VecPointwiseMult(bb1,mat->diag,xx);
1555:     }
1556:     VecAYPX(bb1,(omega-2.0)/omega,bb);

1558:     MatMultAdd(mat->B,mat->lvec,bb1,bb1);

1560:     /* local sweep */
1561:     (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1562:     VecAXPY(xx,1.0,xx1);
1563:     VecDestroy(&xx1);
1564:   } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");

1566:   VecDestroy(&bb1);

1568:   matin->errortype = mat->A->errortype;
1569:   return(0);
1570: }

1574: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1575: {
1576:   Mat            aA,aB,Aperm;
1577:   const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1578:   PetscScalar    *aa,*ba;
1579:   PetscInt       i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1580:   PetscSF        rowsf,sf;
1581:   IS             parcolp = NULL;
1582:   PetscBool      done;

1586:   MatGetLocalSize(A,&m,&n);
1587:   ISGetIndices(rowp,&rwant);
1588:   ISGetIndices(colp,&cwant);
1589:   PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);

1591:   /* Invert row permutation to find out where my rows should go */
1592:   PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1593:   PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1594:   PetscSFSetFromOptions(rowsf);
1595:   for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1596:   PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1597:   PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);

1599:   /* Invert column permutation to find out where my columns should go */
1600:   PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1601:   PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1602:   PetscSFSetFromOptions(sf);
1603:   for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1604:   PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1605:   PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1606:   PetscSFDestroy(&sf);

1608:   ISRestoreIndices(rowp,&rwant);
1609:   ISRestoreIndices(colp,&cwant);
1610:   MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);

1612:   /* Find out where my gcols should go */
1613:   MatGetSize(aB,NULL,&ng);
1614:   PetscMalloc1(ng,&gcdest);
1615:   PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1616:   PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1617:   PetscSFSetFromOptions(sf);
1618:   PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1619:   PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1620:   PetscSFDestroy(&sf);

1622:   PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1623:   MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1624:   MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1625:   for (i=0; i<m; i++) {
1626:     PetscInt row = rdest[i],rowner;
1627:     PetscLayoutFindOwner(A->rmap,row,&rowner);
1628:     for (j=ai[i]; j<ai[i+1]; j++) {
1629:       PetscInt cowner,col = cdest[aj[j]];
1630:       PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1631:       if (rowner == cowner) dnnz[i]++;
1632:       else onnz[i]++;
1633:     }
1634:     for (j=bi[i]; j<bi[i+1]; j++) {
1635:       PetscInt cowner,col = gcdest[bj[j]];
1636:       PetscLayoutFindOwner(A->cmap,col,&cowner);
1637:       if (rowner == cowner) dnnz[i]++;
1638:       else onnz[i]++;
1639:     }
1640:   }
1641:   PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);
1642:   PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);
1643:   PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);
1644:   PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);
1645:   PetscSFDestroy(&rowsf);

1647:   MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);
1648:   MatSeqAIJGetArray(aA,&aa);
1649:   MatSeqAIJGetArray(aB,&ba);
1650:   for (i=0; i<m; i++) {
1651:     PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */
1652:     PetscInt j0,rowlen;
1653:     rowlen = ai[i+1] - ai[i];
1654:     for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */
1655:       for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]];
1656:       MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);
1657:     }
1658:     rowlen = bi[i+1] - bi[i];
1659:     for (j0=j=0; j<rowlen; j0=j) {
1660:       for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]];
1661:       MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);
1662:     }
1663:   }
1664:   MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);
1665:   MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);
1666:   MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1667:   MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1668:   MatSeqAIJRestoreArray(aA,&aa);
1669:   MatSeqAIJRestoreArray(aB,&ba);
1670:   PetscFree4(dnnz,onnz,tdnnz,tonnz);
1671:   PetscFree3(work,rdest,cdest);
1672:   PetscFree(gcdest);
1673:   if (parcolp) {ISDestroy(&colp);}
1674:   *B = Aperm;
1675:   return(0);
1676: }

1680: PetscErrorCode  MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
1681: {
1682:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1686:   MatGetSize(aij->B,NULL,nghosts);
1687:   if (ghosts) *ghosts = aij->garray;
1688:   return(0);
1689: }

1693: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1694: {
1695:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1696:   Mat            A    = mat->A,B = mat->B;
1698:   PetscReal      isend[5],irecv[5];

1701:   info->block_size = 1.0;
1702:   MatGetInfo(A,MAT_LOCAL,info);

1704:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1705:   isend[3] = info->memory;  isend[4] = info->mallocs;

1707:   MatGetInfo(B,MAT_LOCAL,info);

1709:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1710:   isend[3] += info->memory;  isend[4] += info->mallocs;
1711:   if (flag == MAT_LOCAL) {
1712:     info->nz_used      = isend[0];
1713:     info->nz_allocated = isend[1];
1714:     info->nz_unneeded  = isend[2];
1715:     info->memory       = isend[3];
1716:     info->mallocs      = isend[4];
1717:   } else if (flag == MAT_GLOBAL_MAX) {
1718:     MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));

1720:     info->nz_used      = irecv[0];
1721:     info->nz_allocated = irecv[1];
1722:     info->nz_unneeded  = irecv[2];
1723:     info->memory       = irecv[3];
1724:     info->mallocs      = irecv[4];
1725:   } else if (flag == MAT_GLOBAL_SUM) {
1726:     MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));

1728:     info->nz_used      = irecv[0];
1729:     info->nz_allocated = irecv[1];
1730:     info->nz_unneeded  = irecv[2];
1731:     info->memory       = irecv[3];
1732:     info->mallocs      = irecv[4];
1733:   }
1734:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1735:   info->fill_ratio_needed = 0;
1736:   info->factor_mallocs    = 0;
1737:   return(0);
1738: }

1742: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1743: {
1744:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

1748:   switch (op) {
1749:   case MAT_NEW_NONZERO_LOCATIONS:
1750:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1751:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1752:   case MAT_KEEP_NONZERO_PATTERN:
1753:   case MAT_NEW_NONZERO_LOCATION_ERR:
1754:   case MAT_USE_INODES:
1755:   case MAT_IGNORE_ZERO_ENTRIES:
1756:     MatCheckPreallocated(A,1);
1757:     MatSetOption(a->A,op,flg);
1758:     MatSetOption(a->B,op,flg);
1759:     break;
1760:   case MAT_ROW_ORIENTED:
1761:     MatCheckPreallocated(A,1);
1762:     a->roworiented = flg;

1764:     MatSetOption(a->A,op,flg);
1765:     MatSetOption(a->B,op,flg);
1766:     break;
1767:   case MAT_NEW_DIAGONALS:
1768:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1769:     break;
1770:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1771:     a->donotstash = flg;
1772:     break;
1773:   case MAT_SPD:
1774:     A->spd_set = PETSC_TRUE;
1775:     A->spd     = flg;
1776:     if (flg) {
1777:       A->symmetric                  = PETSC_TRUE;
1778:       A->structurally_symmetric     = PETSC_TRUE;
1779:       A->symmetric_set              = PETSC_TRUE;
1780:       A->structurally_symmetric_set = PETSC_TRUE;
1781:     }
1782:     break;
1783:   case MAT_SYMMETRIC:
1784:     MatCheckPreallocated(A,1);
1785:     MatSetOption(a->A,op,flg);
1786:     break;
1787:   case MAT_STRUCTURALLY_SYMMETRIC:
1788:     MatCheckPreallocated(A,1);
1789:     MatSetOption(a->A,op,flg);
1790:     break;
1791:   case MAT_HERMITIAN:
1792:     MatCheckPreallocated(A,1);
1793:     MatSetOption(a->A,op,flg);
1794:     break;
1795:   case MAT_SYMMETRY_ETERNAL:
1796:     MatCheckPreallocated(A,1);
1797:     MatSetOption(a->A,op,flg);
1798:     break;
1799:   default:
1800:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1801:   }
1802:   return(0);
1803: }

1807: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1808: {
1809:   Mat_MPIAIJ     *mat = (Mat_MPIAIJ*)matin->data;
1810:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1812:   PetscInt       i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1813:   PetscInt       nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1814:   PetscInt       *cmap,*idx_p;

1817:   if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1818:   mat->getrowactive = PETSC_TRUE;

1820:   if (!mat->rowvalues && (idx || v)) {
1821:     /*
1822:         allocate enough space to hold information from the longest row.
1823:     */
1824:     Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1825:     PetscInt   max = 1,tmp;
1826:     for (i=0; i<matin->rmap->n; i++) {
1827:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1828:       if (max < tmp) max = tmp;
1829:     }
1830:     PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1831:   }

1833:   if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1834:   lrow = row - rstart;

1836:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1837:   if (!v)   {pvA = 0; pvB = 0;}
1838:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1839:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1840:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1841:   nztot = nzA + nzB;

1843:   cmap = mat->garray;
1844:   if (v  || idx) {
1845:     if (nztot) {
1846:       /* Sort by increasing column numbers, assuming A and B already sorted */
1847:       PetscInt imark = -1;
1848:       if (v) {
1849:         *v = v_p = mat->rowvalues;
1850:         for (i=0; i<nzB; i++) {
1851:           if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1852:           else break;
1853:         }
1854:         imark = i;
1855:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1856:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1857:       }
1858:       if (idx) {
1859:         *idx = idx_p = mat->rowindices;
1860:         if (imark > -1) {
1861:           for (i=0; i<imark; i++) {
1862:             idx_p[i] = cmap[cworkB[i]];
1863:           }
1864:         } else {
1865:           for (i=0; i<nzB; i++) {
1866:             if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1867:             else break;
1868:           }
1869:           imark = i;
1870:         }
1871:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart + cworkA[i];
1872:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]];
1873:       }
1874:     } else {
1875:       if (idx) *idx = 0;
1876:       if (v)   *v   = 0;
1877:     }
1878:   }
1879:   *nz  = nztot;
1880:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1881:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1882:   return(0);
1883: }

1887: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1888: {
1889:   Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;

1892:   if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1893:   aij->getrowactive = PETSC_FALSE;
1894:   return(0);
1895: }

1899: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1900: {
1901:   Mat_MPIAIJ     *aij  = (Mat_MPIAIJ*)mat->data;
1902:   Mat_SeqAIJ     *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1904:   PetscInt       i,j,cstart = mat->cmap->rstart;
1905:   PetscReal      sum = 0.0;
1906:   MatScalar      *v;

1909:   if (aij->size == 1) {
1910:      MatNorm(aij->A,type,norm);
1911:   } else {
1912:     if (type == NORM_FROBENIUS) {
1913:       v = amat->a;
1914:       for (i=0; i<amat->nz; i++) {
1915:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1916:       }
1917:       v = bmat->a;
1918:       for (i=0; i<bmat->nz; i++) {
1919:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1920:       }
1921:       MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1922:       *norm = PetscSqrtReal(*norm);
1923:       PetscLogFlops(2*amat->nz+2*bmat->nz);
1924:     } else if (type == NORM_1) { /* max column norm */
1925:       PetscReal *tmp,*tmp2;
1926:       PetscInt  *jj,*garray = aij->garray;
1927:       PetscCalloc1(mat->cmap->N+1,&tmp);
1928:       PetscMalloc1(mat->cmap->N+1,&tmp2);
1929:       *norm = 0.0;
1930:       v     = amat->a; jj = amat->j;
1931:       for (j=0; j<amat->nz; j++) {
1932:         tmp[cstart + *jj++] += PetscAbsScalar(*v);  v++;
1933:       }
1934:       v = bmat->a; jj = bmat->j;
1935:       for (j=0; j<bmat->nz; j++) {
1936:         tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1937:       }
1938:       MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1939:       for (j=0; j<mat->cmap->N; j++) {
1940:         if (tmp2[j] > *norm) *norm = tmp2[j];
1941:       }
1942:       PetscFree(tmp);
1943:       PetscFree(tmp2);
1944:       PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1945:     } else if (type == NORM_INFINITY) { /* max row norm */
1946:       PetscReal ntemp = 0.0;
1947:       for (j=0; j<aij->A->rmap->n; j++) {
1948:         v   = amat->a + amat->i[j];
1949:         sum = 0.0;
1950:         for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1951:           sum += PetscAbsScalar(*v); v++;
1952:         }
1953:         v = bmat->a + bmat->i[j];
1954:         for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1955:           sum += PetscAbsScalar(*v); v++;
1956:         }
1957:         if (sum > ntemp) ntemp = sum;
1958:       }
1959:       MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
1960:       PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1961:     } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
1962:   }
1963:   return(0);
1964: }

1968: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
1969: {
1970:   Mat_MPIAIJ     *a   = (Mat_MPIAIJ*)A->data;
1971:   Mat_SeqAIJ     *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data;
1973:   PetscInt       M      = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i;
1974:   PetscInt       cstart = A->cmap->rstart,ncol;
1975:   Mat            B;
1976:   MatScalar      *array;

1979:   if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");

1981:   ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
1982:   ai = Aloc->i; aj = Aloc->j;
1983:   bi = Bloc->i; bj = Bloc->j;
1984:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1985:     PetscInt             *d_nnz,*g_nnz,*o_nnz;
1986:     PetscSFNode          *oloc;
1987:     PETSC_UNUSED PetscSF sf;

1989:     PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
1990:     /* compute d_nnz for preallocation */
1991:     PetscMemzero(d_nnz,na*sizeof(PetscInt));
1992:     for (i=0; i<ai[ma]; i++) {
1993:       d_nnz[aj[i]]++;
1994:       aj[i] += cstart; /* global col index to be used by MatSetValues() */
1995:     }
1996:     /* compute local off-diagonal contributions */
1997:     PetscMemzero(g_nnz,nb*sizeof(PetscInt));
1998:     for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
1999:     /* map those to global */
2000:     PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2001:     PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2002:     PetscSFSetFromOptions(sf);
2003:     PetscMemzero(o_nnz,na*sizeof(PetscInt));
2004:     PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2005:     PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2006:     PetscSFDestroy(&sf);

2008:     MatCreate(PetscObjectComm((PetscObject)A),&B);
2009:     MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2010:     MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2011:     MatSetType(B,((PetscObject)A)->type_name);
2012:     MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2013:     PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2014:   } else {
2015:     B    = *matout;
2016:     MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2017:     for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */
2018:   }

2020:   /* copy over the A part */
2021:   array = Aloc->a;
2022:   row   = A->rmap->rstart;
2023:   for (i=0; i<ma; i++) {
2024:     ncol = ai[i+1]-ai[i];
2025:     MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);
2026:     row++;
2027:     array += ncol; aj += ncol;
2028:   }
2029:   aj = Aloc->j;
2030:   for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */

2032:   /* copy over the B part */
2033:   PetscCalloc1(bi[mb],&cols);
2034:   array = Bloc->a;
2035:   row   = A->rmap->rstart;
2036:   for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2037:   cols_tmp = cols;
2038:   for (i=0; i<mb; i++) {
2039:     ncol = bi[i+1]-bi[i];
2040:     MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2041:     row++;
2042:     array += ncol; cols_tmp += ncol;
2043:   }
2044:   PetscFree(cols);

2046:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2047:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2048:   if (reuse == MAT_INITIAL_MATRIX || *matout != A) {
2049:     *matout = B;
2050:   } else {
2051:     MatHeaderMerge(A,&B);
2052:   }
2053:   return(0);
2054: }

2058: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2059: {
2060:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
2061:   Mat            a    = aij->A,b = aij->B;
2063:   PetscInt       s1,s2,s3;

2066:   MatGetLocalSize(mat,&s2,&s3);
2067:   if (rr) {
2068:     VecGetLocalSize(rr,&s1);
2069:     if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2070:     /* Overlap communication with computation. */
2071:     VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2072:   }
2073:   if (ll) {
2074:     VecGetLocalSize(ll,&s1);
2075:     if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2076:     (*b->ops->diagonalscale)(b,ll,0);
2077:   }
2078:   /* scale  the diagonal block */
2079:   (*a->ops->diagonalscale)(a,ll,rr);

2081:   if (rr) {
2082:     /* Do a scatter end and then right scale the off-diagonal block */
2083:     VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2084:     (*b->ops->diagonalscale)(b,0,aij->lvec);
2085:   }
2086:   return(0);
2087: }

2091: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2092: {
2093:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2097:   MatSetUnfactored(a->A);
2098:   return(0);
2099: }

2103: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool  *flag)
2104: {
2105:   Mat_MPIAIJ     *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2106:   Mat            a,b,c,d;
2107:   PetscBool      flg;

2111:   a = matA->A; b = matA->B;
2112:   c = matB->A; d = matB->B;

2114:   MatEqual(a,c,&flg);
2115:   if (flg) {
2116:     MatEqual(b,d,&flg);
2117:   }
2118:   MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2119:   return(0);
2120: }

2124: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2125: {
2127:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2128:   Mat_MPIAIJ     *b = (Mat_MPIAIJ*)B->data;

2131:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2132:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2133:     /* because of the column compression in the off-processor part of the matrix a->B,
2134:        the number of columns in a->B and b->B may be different, hence we cannot call
2135:        the MatCopy() directly on the two parts. If need be, we can provide a more
2136:        efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2137:        then copying the submatrices */
2138:     MatCopy_Basic(A,B,str);
2139:   } else {
2140:     MatCopy(a->A,b->A,str);
2141:     MatCopy(a->B,b->B,str);
2142:   }
2143:   return(0);
2144: }

2148: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2149: {

2153:    MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2154:   return(0);
2155: }

2157: /*
2158:    Computes the number of nonzeros per row needed for preallocation when X and Y
2159:    have different nonzero structure.
2160: */
2163: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz)
2164: {
2165:   PetscInt       i,j,k,nzx,nzy;

2168:   /* Set the number of nonzeros in the new matrix */
2169:   for (i=0; i<m; i++) {
2170:     const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2171:     nzx = xi[i+1] - xi[i];
2172:     nzy = yi[i+1] - yi[i];
2173:     nnz[i] = 0;
2174:     for (j=0,k=0; j<nzx; j++) {                   /* Point in X */
2175:       for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2176:       if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++;             /* Skip duplicate */
2177:       nnz[i]++;
2178:     }
2179:     for (; k<nzy; k++) nnz[i]++;
2180:   }
2181:   return(0);
2182: }

2184: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2187: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2188: {
2190:   PetscInt       m = Y->rmap->N;
2191:   Mat_SeqAIJ     *x = (Mat_SeqAIJ*)X->data;
2192:   Mat_SeqAIJ     *y = (Mat_SeqAIJ*)Y->data;

2195:   MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2196:   return(0);
2197: }

2201: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2202: {
2204:   Mat_MPIAIJ     *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2205:   PetscBLASInt   bnz,one=1;
2206:   Mat_SeqAIJ     *x,*y;

2209:   if (str == SAME_NONZERO_PATTERN) {
2210:     PetscScalar alpha = a;
2211:     x    = (Mat_SeqAIJ*)xx->A->data;
2212:     PetscBLASIntCast(x->nz,&bnz);
2213:     y    = (Mat_SeqAIJ*)yy->A->data;
2214:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2215:     x    = (Mat_SeqAIJ*)xx->B->data;
2216:     y    = (Mat_SeqAIJ*)yy->B->data;
2217:     PetscBLASIntCast(x->nz,&bnz);
2218:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2219:     PetscObjectStateIncrease((PetscObject)Y);
2220:   } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2221:     MatAXPY_Basic(Y,a,X,str);
2222:   } else {
2223:     Mat      B;
2224:     PetscInt *nnz_d,*nnz_o;
2225:     PetscMalloc1(yy->A->rmap->N,&nnz_d);
2226:     PetscMalloc1(yy->B->rmap->N,&nnz_o);
2227:     MatCreate(PetscObjectComm((PetscObject)Y),&B);
2228:     PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2229:     MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2230:     MatSetBlockSizesFromMats(B,Y,Y);
2231:     MatSetType(B,MATMPIAIJ);
2232:     MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2233:     MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2234:     MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2235:     MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2236:     MatHeaderReplace(Y,&B);
2237:     PetscFree(nnz_d);
2238:     PetscFree(nnz_o);
2239:   }
2240:   return(0);
2241: }

2243: extern PetscErrorCode  MatConjugate_SeqAIJ(Mat);

2247: PetscErrorCode  MatConjugate_MPIAIJ(Mat mat)
2248: {
2249: #if defined(PETSC_USE_COMPLEX)
2251:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2254:   MatConjugate_SeqAIJ(aij->A);
2255:   MatConjugate_SeqAIJ(aij->B);
2256: #else
2258: #endif
2259:   return(0);
2260: }

2264: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2265: {
2266:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2270:   MatRealPart(a->A);
2271:   MatRealPart(a->B);
2272:   return(0);
2273: }

2277: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2278: {
2279:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2283:   MatImaginaryPart(a->A);
2284:   MatImaginaryPart(a->B);
2285:   return(0);
2286: }

2290: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2291: {
2292:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2294:   PetscInt       i,*idxb = 0;
2295:   PetscScalar    *va,*vb;
2296:   Vec            vtmp;

2299:   MatGetRowMaxAbs(a->A,v,idx);
2300:   VecGetArray(v,&va);
2301:   if (idx) {
2302:     for (i=0; i<A->rmap->n; i++) {
2303:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2304:     }
2305:   }

2307:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2308:   if (idx) {
2309:     PetscMalloc1(A->rmap->n,&idxb);
2310:   }
2311:   MatGetRowMaxAbs(a->B,vtmp,idxb);
2312:   VecGetArray(vtmp,&vb);

2314:   for (i=0; i<A->rmap->n; i++) {
2315:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2316:       va[i] = vb[i];
2317:       if (idx) idx[i] = a->garray[idxb[i]];
2318:     }
2319:   }

2321:   VecRestoreArray(v,&va);
2322:   VecRestoreArray(vtmp,&vb);
2323:   PetscFree(idxb);
2324:   VecDestroy(&vtmp);
2325:   return(0);
2326: }

2330: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2331: {
2332:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
2334:   PetscInt       i,*idxb = 0;
2335:   PetscScalar    *va,*vb;
2336:   Vec            vtmp;

2339:   MatGetRowMinAbs(a->A,v,idx);
2340:   VecGetArray(v,&va);
2341:   if (idx) {
2342:     for (i=0; i<A->cmap->n; i++) {
2343:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2344:     }
2345:   }

2347:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2348:   if (idx) {
2349:     PetscMalloc1(A->rmap->n,&idxb);
2350:   }
2351:   MatGetRowMinAbs(a->B,vtmp,idxb);
2352:   VecGetArray(vtmp,&vb);

2354:   for (i=0; i<A->rmap->n; i++) {
2355:     if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2356:       va[i] = vb[i];
2357:       if (idx) idx[i] = a->garray[idxb[i]];
2358:     }
2359:   }

2361:   VecRestoreArray(v,&va);
2362:   VecRestoreArray(vtmp,&vb);
2363:   PetscFree(idxb);
2364:   VecDestroy(&vtmp);
2365:   return(0);
2366: }

2370: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2371: {
2372:   Mat_MPIAIJ     *mat   = (Mat_MPIAIJ*) A->data;
2373:   PetscInt       n      = A->rmap->n;
2374:   PetscInt       cstart = A->cmap->rstart;
2375:   PetscInt       *cmap  = mat->garray;
2376:   PetscInt       *diagIdx, *offdiagIdx;
2377:   Vec            diagV, offdiagV;
2378:   PetscScalar    *a, *diagA, *offdiagA;
2379:   PetscInt       r;

2383:   PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2384:   VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);
2385:   VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);
2386:   MatGetRowMin(mat->A, diagV,    diagIdx);
2387:   MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2388:   VecGetArray(v,        &a);
2389:   VecGetArray(diagV,    &diagA);
2390:   VecGetArray(offdiagV, &offdiagA);
2391:   for (r = 0; r < n; ++r) {
2392:     if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2393:       a[r]   = diagA[r];
2394:       idx[r] = cstart + diagIdx[r];
2395:     } else {
2396:       a[r]   = offdiagA[r];
2397:       idx[r] = cmap[offdiagIdx[r]];
2398:     }
2399:   }
2400:   VecRestoreArray(v,        &a);
2401:   VecRestoreArray(diagV,    &diagA);
2402:   VecRestoreArray(offdiagV, &offdiagA);
2403:   VecDestroy(&diagV);
2404:   VecDestroy(&offdiagV);
2405:   PetscFree2(diagIdx, offdiagIdx);
2406:   return(0);
2407: }

2411: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2412: {
2413:   Mat_MPIAIJ     *mat   = (Mat_MPIAIJ*) A->data;
2414:   PetscInt       n      = A->rmap->n;
2415:   PetscInt       cstart = A->cmap->rstart;
2416:   PetscInt       *cmap  = mat->garray;
2417:   PetscInt       *diagIdx, *offdiagIdx;
2418:   Vec            diagV, offdiagV;
2419:   PetscScalar    *a, *diagA, *offdiagA;
2420:   PetscInt       r;

2424:   PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2425:   VecCreateSeq(PETSC_COMM_SELF, n, &diagV);
2426:   VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);
2427:   MatGetRowMax(mat->A, diagV,    diagIdx);
2428:   MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2429:   VecGetArray(v,        &a);
2430:   VecGetArray(diagV,    &diagA);
2431:   VecGetArray(offdiagV, &offdiagA);
2432:   for (r = 0; r < n; ++r) {
2433:     if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2434:       a[r]   = diagA[r];
2435:       idx[r] = cstart + diagIdx[r];
2436:     } else {
2437:       a[r]   = offdiagA[r];
2438:       idx[r] = cmap[offdiagIdx[r]];
2439:     }
2440:   }
2441:   VecRestoreArray(v,        &a);
2442:   VecRestoreArray(diagV,    &diagA);
2443:   VecRestoreArray(offdiagV, &offdiagA);
2444:   VecDestroy(&diagV);
2445:   VecDestroy(&offdiagV);
2446:   PetscFree2(diagIdx, offdiagIdx);
2447:   return(0);
2448: }

2452: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2453: {
2455:   Mat            *dummy;

2458:   MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2459:   *newmat = *dummy;
2460:   PetscFree(dummy);
2461:   return(0);
2462: }

2466: PetscErrorCode  MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2467: {
2468:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*) A->data;

2472:   MatInvertBlockDiagonal(a->A,values);
2473:   A->errortype = a->A->errortype;
2474:   return(0);
2475: }

2479: static PetscErrorCode  MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2480: {
2482:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)x->data;

2485:   MatSetRandom(aij->A,rctx);
2486:   MatSetRandom(aij->B,rctx);
2487:   MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2488:   MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2489:   return(0);
2490: }

2494: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2495: {
2497:   if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2498:   else A->ops->increaseoverlap    = MatIncreaseOverlap_MPIAIJ;
2499:   return(0);
2500: }

2504: /*@
2505:    MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap

2507:    Collective on Mat

2509:    Input Parameters:
2510: +    A - the matrix
2511: -    sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)

2513:  Level: advanced

2515: @*/
2516: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2517: {
2518:   PetscErrorCode       ierr;

2521:   PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2522:   return(0);
2523: }

2527: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2528: {
2529:   PetscErrorCode       ierr;
2530:   PetscBool            sc = PETSC_FALSE,flg;

2533:   PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2534:   PetscObjectOptionsBegin((PetscObject)A);
2535:     if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2536:     PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2537:     if (flg) {
2538:       MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2539:     }
2540:   PetscOptionsEnd();
2541:   return(0);
2542: }

2546: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2547: {
2549:   Mat_MPIAIJ     *maij = (Mat_MPIAIJ*)Y->data;
2550:   Mat_SeqAIJ     *aij = (Mat_SeqAIJ*)maij->A->data;

2553:   if (!Y->preallocated) {
2554:     MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2555:   } else if (!aij->nz) {
2556:     PetscInt nonew = aij->nonew;
2557:     MatSeqAIJSetPreallocation(maij->A,1,NULL);
2558:     aij->nonew = nonew;
2559:   }
2560:   MatShift_Basic(Y,a);
2561:   return(0);
2562: }

2566: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool  *missing,PetscInt *d)
2567: {
2568:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

2572:   if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2573:   MatMissingDiagonal(a->A,missing,d);
2574:   if (d) {
2575:     PetscInt rstart;
2576:     MatGetOwnershipRange(A,&rstart,NULL);
2577:     *d += rstart;

2579:   }
2580:   return(0);
2581: }


2584: /* -------------------------------------------------------------------*/
2585: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2586:                                        MatGetRow_MPIAIJ,
2587:                                        MatRestoreRow_MPIAIJ,
2588:                                        MatMult_MPIAIJ,
2589:                                 /* 4*/ MatMultAdd_MPIAIJ,
2590:                                        MatMultTranspose_MPIAIJ,
2591:                                        MatMultTransposeAdd_MPIAIJ,
2592:                                        0,
2593:                                        0,
2594:                                        0,
2595:                                 /*10*/ 0,
2596:                                        0,
2597:                                        0,
2598:                                        MatSOR_MPIAIJ,
2599:                                        MatTranspose_MPIAIJ,
2600:                                 /*15*/ MatGetInfo_MPIAIJ,
2601:                                        MatEqual_MPIAIJ,
2602:                                        MatGetDiagonal_MPIAIJ,
2603:                                        MatDiagonalScale_MPIAIJ,
2604:                                        MatNorm_MPIAIJ,
2605:                                 /*20*/ MatAssemblyBegin_MPIAIJ,
2606:                                        MatAssemblyEnd_MPIAIJ,
2607:                                        MatSetOption_MPIAIJ,
2608:                                        MatZeroEntries_MPIAIJ,
2609:                                 /*24*/ MatZeroRows_MPIAIJ,
2610:                                        0,
2611:                                        0,
2612:                                        0,
2613:                                        0,
2614:                                 /*29*/ MatSetUp_MPIAIJ,
2615:                                        0,
2616:                                        0,
2617:                                        0,
2618:                                        0,
2619:                                 /*34*/ MatDuplicate_MPIAIJ,
2620:                                        0,
2621:                                        0,
2622:                                        0,
2623:                                        0,
2624:                                 /*39*/ MatAXPY_MPIAIJ,
2625:                                        MatGetSubMatrices_MPIAIJ,
2626:                                        MatIncreaseOverlap_MPIAIJ,
2627:                                        MatGetValues_MPIAIJ,
2628:                                        MatCopy_MPIAIJ,
2629:                                 /*44*/ MatGetRowMax_MPIAIJ,
2630:                                        MatScale_MPIAIJ,
2631:                                        MatShift_MPIAIJ,
2632:                                        MatDiagonalSet_MPIAIJ,
2633:                                        MatZeroRowsColumns_MPIAIJ,
2634:                                 /*49*/ MatSetRandom_MPIAIJ,
2635:                                        0,
2636:                                        0,
2637:                                        0,
2638:                                        0,
2639:                                 /*54*/ MatFDColoringCreate_MPIXAIJ,
2640:                                        0,
2641:                                        MatSetUnfactored_MPIAIJ,
2642:                                        MatPermute_MPIAIJ,
2643:                                        0,
2644:                                 /*59*/ MatGetSubMatrix_MPIAIJ,
2645:                                        MatDestroy_MPIAIJ,
2646:                                        MatView_MPIAIJ,
2647:                                        0,
2648:                                        MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ,
2649:                                 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ,
2650:                                        MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2651:                                        0,
2652:                                        0,
2653:                                        0,
2654:                                 /*69*/ MatGetRowMaxAbs_MPIAIJ,
2655:                                        MatGetRowMinAbs_MPIAIJ,
2656:                                        0,
2657:                                        MatSetColoring_MPIAIJ,
2658:                                        0,
2659:                                        MatSetValuesAdifor_MPIAIJ,
2660:                                 /*75*/ MatFDColoringApply_AIJ,
2661:                                        MatSetFromOptions_MPIAIJ,
2662:                                        0,
2663:                                        0,
2664:                                        MatFindZeroDiagonals_MPIAIJ,
2665:                                 /*80*/ 0,
2666:                                        0,
2667:                                        0,
2668:                                 /*83*/ MatLoad_MPIAIJ,
2669:                                        0,
2670:                                        0,
2671:                                        0,
2672:                                        0,
2673:                                        0,
2674:                                 /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
2675:                                        MatMatMultSymbolic_MPIAIJ_MPIAIJ,
2676:                                        MatMatMultNumeric_MPIAIJ_MPIAIJ,
2677:                                        MatPtAP_MPIAIJ_MPIAIJ,
2678:                                        MatPtAPSymbolic_MPIAIJ_MPIAIJ,
2679:                                 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2680:                                        0,
2681:                                        0,
2682:                                        0,
2683:                                        0,
2684:                                 /*99*/ 0,
2685:                                        0,
2686:                                        0,
2687:                                        MatConjugate_MPIAIJ,
2688:                                        0,
2689:                                 /*104*/MatSetValuesRow_MPIAIJ,
2690:                                        MatRealPart_MPIAIJ,
2691:                                        MatImaginaryPart_MPIAIJ,
2692:                                        0,
2693:                                        0,
2694:                                 /*109*/0,
2695:                                        0,
2696:                                        MatGetRowMin_MPIAIJ,
2697:                                        0,
2698:                                        MatMissingDiagonal_MPIAIJ,
2699:                                 /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2700:                                        0,
2701:                                        MatGetGhosts_MPIAIJ,
2702:                                        0,
2703:                                        0,
2704:                                 /*119*/0,
2705:                                        0,
2706:                                        0,
2707:                                        0,
2708:                                        MatGetMultiProcBlock_MPIAIJ,
2709:                                 /*124*/MatFindNonzeroRows_MPIAIJ,
2710:                                        MatGetColumnNorms_MPIAIJ,
2711:                                        MatInvertBlockDiagonal_MPIAIJ,
2712:                                        0,
2713:                                        MatGetSubMatricesMPI_MPIAIJ,
2714:                                 /*129*/0,
2715:                                        MatTransposeMatMult_MPIAIJ_MPIAIJ,
2716:                                        MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
2717:                                        MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2718:                                        0,
2719:                                 /*134*/0,
2720:                                        0,
2721:                                        0,
2722:                                        0,
2723:                                        0,
2724:                                 /*139*/0,
2725:                                        0,
2726:                                        0,
2727:                                        MatFDColoringSetUp_MPIXAIJ,
2728:                                        MatFindOffBlockDiagonalEntries_MPIAIJ,
2729:                                 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ
2730: };

2732: /* ----------------------------------------------------------------------------------------*/

2736: PetscErrorCode  MatStoreValues_MPIAIJ(Mat mat)
2737: {
2738:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2742:   MatStoreValues(aij->A);
2743:   MatStoreValues(aij->B);
2744:   return(0);
2745: }

2749: PetscErrorCode  MatRetrieveValues_MPIAIJ(Mat mat)
2750: {
2751:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;

2755:   MatRetrieveValues(aij->A);
2756:   MatRetrieveValues(aij->B);
2757:   return(0);
2758: }

2762: PetscErrorCode  MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2763: {
2764:   Mat_MPIAIJ     *b;

2768:   PetscLayoutSetUp(B->rmap);
2769:   PetscLayoutSetUp(B->cmap);
2770:   b = (Mat_MPIAIJ*)B->data;

2772:   if (!B->preallocated) {
2773:     /* Explicitly create 2 MATSEQAIJ matrices. */
2774:     MatCreate(PETSC_COMM_SELF,&b->A);
2775:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2776:     MatSetBlockSizesFromMats(b->A,B,B);
2777:     MatSetType(b->A,MATSEQAIJ);
2778:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2779:     MatCreate(PETSC_COMM_SELF,&b->B);
2780:     MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
2781:     MatSetBlockSizesFromMats(b->B,B,B);
2782:     MatSetType(b->B,MATSEQAIJ);
2783:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2784:   }

2786:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2787:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
2788:   B->preallocated = PETSC_TRUE;
2789:   return(0);
2790: }

2794: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2795: {
2796:   Mat            mat;
2797:   Mat_MPIAIJ     *a,*oldmat = (Mat_MPIAIJ*)matin->data;

2801:   *newmat = 0;
2802:   MatCreate(PetscObjectComm((PetscObject)matin),&mat);
2803:   MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
2804:   MatSetBlockSizesFromMats(mat,matin,matin);
2805:   MatSetType(mat,((PetscObject)matin)->type_name);
2806:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));
2807:   a       = (Mat_MPIAIJ*)mat->data;

2809:   mat->factortype   = matin->factortype;
2810:   mat->assembled    = PETSC_TRUE;
2811:   mat->insertmode   = NOT_SET_VALUES;
2812:   mat->preallocated = PETSC_TRUE;

2814:   a->size         = oldmat->size;
2815:   a->rank         = oldmat->rank;
2816:   a->donotstash   = oldmat->donotstash;
2817:   a->roworiented  = oldmat->roworiented;
2818:   a->rowindices   = 0;
2819:   a->rowvalues    = 0;
2820:   a->getrowactive = PETSC_FALSE;

2822:   PetscLayoutReference(matin->rmap,&mat->rmap);
2823:   PetscLayoutReference(matin->cmap,&mat->cmap);

2825:   if (oldmat->colmap) {
2826: #if defined(PETSC_USE_CTABLE)
2827:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2828: #else
2829:     PetscMalloc1(mat->cmap->N,&a->colmap);
2830:     PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
2831:     PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));
2832: #endif
2833:   } else a->colmap = 0;
2834:   if (oldmat->garray) {
2835:     PetscInt len;
2836:     len  = oldmat->B->cmap->n;
2837:     PetscMalloc1(len+1,&a->garray);
2838:     PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
2839:     if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
2840:   } else a->garray = 0;

2842:   VecDuplicate(oldmat->lvec,&a->lvec);
2843:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
2844:   VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2845:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
2846:   MatDuplicate(oldmat->A,cpvalues,&a->A);
2847:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
2848:   MatDuplicate(oldmat->B,cpvalues,&a->B);
2849:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
2850:   PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2851:   *newmat = mat;
2852:   return(0);
2853: }



2859: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
2860: {
2861:   PetscScalar    *vals,*svals;
2862:   MPI_Comm       comm;
2864:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag;
2865:   PetscInt       i,nz,j,rstart,rend,mmax,maxnz = 0;
2866:   PetscInt       header[4],*rowlengths = 0,M,N,m,*cols;
2867:   PetscInt       *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols;
2868:   PetscInt       cend,cstart,n,*rowners;
2869:   int            fd;
2870:   PetscInt       bs = newMat->rmap->bs;

2873:   /* force binary viewer to load .info file if it has not yet done so */
2874:   PetscViewerSetUp(viewer);
2875:   PetscObjectGetComm((PetscObject)viewer,&comm);
2876:   MPI_Comm_size(comm,&size);
2877:   MPI_Comm_rank(comm,&rank);
2878:   PetscViewerBinaryGetDescriptor(viewer,&fd);
2879:   if (!rank) {
2880:     PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
2881:     if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2882:   }

2884:   PetscOptionsBegin(comm,NULL,"Options for loading MPIAIJ matrix","Mat");
2885:   PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
2886:   PetscOptionsEnd();
2887:   if (bs < 0) bs = 1;

2889:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2890:   M    = header[1]; N = header[2];

2892:   /* If global sizes are set, check if they are consistent with that given in the file */
2893:   if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M);
2894:   if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N);

2896:   /* determine ownership of all (block) rows */
2897:   if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs);
2898:   if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank));    /* PETSC_DECIDE */
2899:   else m = newMat->rmap->n; /* Set by user */

2901:   PetscMalloc1(size+1,&rowners);
2902:   MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

2904:   /* First process needs enough room for process with most rows */
2905:   if (!rank) {
2906:     mmax = rowners[1];
2907:     for (i=2; i<=size; i++) {
2908:       mmax = PetscMax(mmax, rowners[i]);
2909:     }
2910:   } else mmax = -1;             /* unused, but compilers complain */

2912:   rowners[0] = 0;
2913:   for (i=2; i<=size; i++) {
2914:     rowners[i] += rowners[i-1];
2915:   }
2916:   rstart = rowners[rank];
2917:   rend   = rowners[rank+1];

2919:   /* distribute row lengths to all processors */
2920:   PetscMalloc2(m,&ourlens,m,&offlens);
2921:   if (!rank) {
2922:     PetscBinaryRead(fd,ourlens,m,PETSC_INT);
2923:     PetscMalloc1(mmax,&rowlengths);
2924:     PetscCalloc1(size,&procsnz);
2925:     for (j=0; j<m; j++) {
2926:       procsnz[0] += ourlens[j];
2927:     }
2928:     for (i=1; i<size; i++) {
2929:       PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
2930:       /* calculate the number of nonzeros on each processor */
2931:       for (j=0; j<rowners[i+1]-rowners[i]; j++) {
2932:         procsnz[i] += rowlengths[j];
2933:       }
2934:       MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
2935:     }
2936:     PetscFree(rowlengths);
2937:   } else {
2938:     MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
2939:   }

2941:   if (!rank) {
2942:     /* determine max buffer needed and allocate it */
2943:     maxnz = 0;
2944:     for (i=0; i<size; i++) {
2945:       maxnz = PetscMax(maxnz,procsnz[i]);
2946:     }
2947:     PetscMalloc1(maxnz,&cols);

2949:     /* read in my part of the matrix column indices  */
2950:     nz   = procsnz[0];
2951:     PetscMalloc1(nz,&mycols);
2952:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);

2954:     /* read in every one elses and ship off */
2955:     for (i=1; i<size; i++) {
2956:       nz   = procsnz[i];
2957:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
2958:       MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
2959:     }
2960:     PetscFree(cols);
2961:   } else {
2962:     /* determine buffer space needed for message */
2963:     nz = 0;
2964:     for (i=0; i<m; i++) {
2965:       nz += ourlens[i];
2966:     }
2967:     PetscMalloc1(nz,&mycols);

2969:     /* receive message of column indices*/
2970:     MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
2971:   }

2973:   /* determine column ownership if matrix is not square */
2974:   if (N != M) {
2975:     if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank);
2976:     else n = newMat->cmap->n;
2977:     MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
2978:     cstart = cend - n;
2979:   } else {
2980:     cstart = rstart;
2981:     cend   = rend;
2982:     n      = cend - cstart;
2983:   }

2985:   /* loop over local rows, determining number of off diagonal entries */
2986:   PetscMemzero(offlens,m*sizeof(PetscInt));
2987:   jj   = 0;
2988:   for (i=0; i<m; i++) {
2989:     for (j=0; j<ourlens[i]; j++) {
2990:       if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
2991:       jj++;
2992:     }
2993:   }

2995:   for (i=0; i<m; i++) {
2996:     ourlens[i] -= offlens[i];
2997:   }
2998:   MatSetSizes(newMat,m,n,M,N);

3000:   if (bs > 1) {MatSetBlockSize(newMat,bs);}

3002:   MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);

3004:   for (i=0; i<m; i++) {
3005:     ourlens[i] += offlens[i];
3006:   }

3008:   if (!rank) {
3009:     PetscMalloc1(maxnz+1,&vals);

3011:     /* read in my part of the matrix numerical values  */
3012:     nz   = procsnz[0];
3013:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);

3015:     /* insert into matrix */
3016:     jj      = rstart;
3017:     smycols = mycols;
3018:     svals   = vals;
3019:     for (i=0; i<m; i++) {
3020:       MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3021:       smycols += ourlens[i];
3022:       svals   += ourlens[i];
3023:       jj++;
3024:     }

3026:     /* read in other processors and ship out */
3027:     for (i=1; i<size; i++) {
3028:       nz   = procsnz[i];
3029:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3030:       MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3031:     }
3032:     PetscFree(procsnz);
3033:   } else {
3034:     /* receive numeric values */
3035:     PetscMalloc1(nz+1,&vals);

3037:     /* receive message of values*/
3038:     MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);

3040:     /* insert into matrix */
3041:     jj      = rstart;
3042:     smycols = mycols;
3043:     svals   = vals;
3044:     for (i=0; i<m; i++) {
3045:       MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3046:       smycols += ourlens[i];
3047:       svals   += ourlens[i];
3048:       jj++;
3049:     }
3050:   }
3051:   PetscFree2(ourlens,offlens);
3052:   PetscFree(vals);
3053:   PetscFree(mycols);
3054:   PetscFree(rowners);
3055:   MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3056:   MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3057:   return(0);
3058: }

3062: /* TODO: Not scalable because of ISAllGather() unless getting all columns. */
3063: PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3064: {
3066:   IS             iscol_local;
3067:   PetscInt       csize;

3070:   ISGetLocalSize(iscol,&csize);
3071:   if (call == MAT_REUSE_MATRIX) {
3072:     PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3073:     if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3074:   } else {
3075:     /* check if we are grabbing all columns*/
3076:     PetscBool    isstride;
3077:     PetscMPIInt  lisstride = 0,gisstride;
3078:     PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);
3079:     if (isstride) {
3080:       PetscInt  start,len,mstart,mlen;
3081:       ISStrideGetInfo(iscol,&start,NULL);
3082:       ISGetLocalSize(iscol,&len);
3083:       MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
3084:       if (mstart == start && mlen-mstart == len) lisstride = 1;
3085:     }
3086:     MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3087:     if (gisstride) {
3088:       PetscInt N;
3089:       MatGetSize(mat,NULL,&N);
3090:       ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);
3091:       ISSetIdentity(iscol_local);
3092:       PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3093:     } else {
3094:       PetscInt cbs;
3095:       ISGetBlockSize(iscol,&cbs);
3096:       ISAllGather(iscol,&iscol_local);
3097:       ISSetBlockSize(iscol_local,cbs);
3098:     }
3099:   }
3100:   MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
3101:   if (call == MAT_INITIAL_MATRIX) {
3102:     PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3103:     ISDestroy(&iscol_local);
3104:   }
3105:   return(0);
3106: }

3108: extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*);
3111: /*
3112:     Not great since it makes two copies of the submatrix, first an SeqAIJ
3113:   in local and then by concatenating the local matrices the end result.
3114:   Writing it directly would be much like MatGetSubMatrices_MPIAIJ()

3116:   Note: This requires a sequential iscol with all indices.
3117: */
3118: PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3119: {
3121:   PetscMPIInt    rank,size;
3122:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3123:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol;
3124:   PetscBool      allcolumns, colflag;
3125:   Mat            M,Mreuse;
3126:   MatScalar      *vwork,*aa;
3127:   MPI_Comm       comm;
3128:   Mat_SeqAIJ     *aij;

3131:   PetscObjectGetComm((PetscObject)mat,&comm);
3132:   MPI_Comm_rank(comm,&rank);
3133:   MPI_Comm_size(comm,&size);

3135:   ISIdentity(iscol,&colflag);
3136:   ISGetLocalSize(iscol,&ncol);
3137:   if (colflag && ncol == mat->cmap->N) {
3138:     allcolumns = PETSC_TRUE;
3139:     PetscInfo(mat,"Optimizing for obtaining all columns of the matrix\n");
3140:   } else {
3141:     allcolumns = PETSC_FALSE;
3142:   }
3143:   if (call ==  MAT_REUSE_MATRIX) {
3144:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3145:     if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3146:     MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);
3147:   } else {
3148:     MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);
3149:   }

3151:   /*
3152:       m - number of local rows
3153:       n - number of columns (same on all processors)
3154:       rstart - first row in new global matrix generated
3155:   */
3156:   MatGetSize(Mreuse,&m,&n);
3157:   MatGetBlockSizes(Mreuse,&bs,&cbs);
3158:   if (call == MAT_INITIAL_MATRIX) {
3159:     aij = (Mat_SeqAIJ*)(Mreuse)->data;
3160:     ii  = aij->i;
3161:     jj  = aij->j;

3163:     /*
3164:         Determine the number of non-zeros in the diagonal and off-diagonal
3165:         portions of the matrix in order to do correct preallocation
3166:     */

3168:     /* first get start and end of "diagonal" columns */
3169:     if (csize == PETSC_DECIDE) {
3170:       ISGetSize(isrow,&mglobal);
3171:       if (mglobal == n) { /* square matrix */
3172:         nlocal = m;
3173:       } else {
3174:         nlocal = n/size + ((n % size) > rank);
3175:       }
3176:     } else {
3177:       nlocal = csize;
3178:     }
3179:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3180:     rstart = rend - nlocal;
3181:     if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);

3183:     /* next, compute all the lengths */
3184:     PetscMalloc1(2*m+1,&dlens);
3185:     olens = dlens + m;
3186:     for (i=0; i<m; i++) {
3187:       jend = ii[i+1] - ii[i];
3188:       olen = 0;
3189:       dlen = 0;
3190:       for (j=0; j<jend; j++) {
3191:         if (*jj < rstart || *jj >= rend) olen++;
3192:         else dlen++;
3193:         jj++;
3194:       }
3195:       olens[i] = olen;
3196:       dlens[i] = dlen;
3197:     }
3198:     MatCreate(comm,&M);
3199:     MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3200:     MatSetBlockSizes(M,bs,cbs);
3201:     MatSetType(M,((PetscObject)mat)->type_name);
3202:     MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3203:     PetscFree(dlens);
3204:   } else {
3205:     PetscInt ml,nl;

3207:     M    = *newmat;
3208:     MatGetLocalSize(M,&ml,&nl);
3209:     if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3210:     MatZeroEntries(M);
3211:     /*
3212:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3213:        rather than the slower MatSetValues().
3214:     */
3215:     M->was_assembled = PETSC_TRUE;
3216:     M->assembled     = PETSC_FALSE;
3217:   }
3218:   MatGetOwnershipRange(M,&rstart,&rend);
3219:   aij  = (Mat_SeqAIJ*)(Mreuse)->data;
3220:   ii   = aij->i;
3221:   jj   = aij->j;
3222:   aa   = aij->a;
3223:   for (i=0; i<m; i++) {
3224:     row   = rstart + i;
3225:     nz    = ii[i+1] - ii[i];
3226:     cwork = jj;     jj += nz;
3227:     vwork = aa;     aa += nz;
3228:     MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3229:   }

3231:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3232:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3233:   *newmat = M;

3235:   /* save submatrix used in processor for next request */
3236:   if (call ==  MAT_INITIAL_MATRIX) {
3237:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3238:     MatDestroy(&Mreuse);
3239:   }
3240:   return(0);
3241: }

3245: PetscErrorCode  MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3246: {
3247:   PetscInt       m,cstart, cend,j,nnz,i,d;
3248:   PetscInt       *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3249:   const PetscInt *JJ;
3250:   PetscScalar    *values;

3254:   if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);

3256:   PetscLayoutSetUp(B->rmap);
3257:   PetscLayoutSetUp(B->cmap);
3258:   m      = B->rmap->n;
3259:   cstart = B->cmap->rstart;
3260:   cend   = B->cmap->rend;
3261:   rstart = B->rmap->rstart;

3263:   PetscMalloc2(m,&d_nnz,m,&o_nnz);

3265: #if defined(PETSC_USE_DEBUGGING)
3266:   for (i=0; i<m; i++) {
3267:     nnz = Ii[i+1]- Ii[i];
3268:     JJ  = J + Ii[i];
3269:     if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3270:     if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j);
3271:     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);
3272:   }
3273: #endif

3275:   for (i=0; i<m; i++) {
3276:     nnz     = Ii[i+1]- Ii[i];
3277:     JJ      = J + Ii[i];
3278:     nnz_max = PetscMax(nnz_max,nnz);
3279:     d       = 0;
3280:     for (j=0; j<nnz; j++) {
3281:       if (cstart <= JJ[j] && JJ[j] < cend) d++;
3282:     }
3283:     d_nnz[i] = d;
3284:     o_nnz[i] = nnz - d;
3285:   }
3286:   MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3287:   PetscFree2(d_nnz,o_nnz);

3289:   if (v) values = (PetscScalar*)v;
3290:   else {
3291:     PetscCalloc1(nnz_max+1,&values);
3292:   }

3294:   for (i=0; i<m; i++) {
3295:     ii   = i + rstart;
3296:     nnz  = Ii[i+1]- Ii[i];
3297:     MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
3298:   }
3299:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3300:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

3302:   if (!v) {
3303:     PetscFree(values);
3304:   }
3305:   MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3306:   return(0);
3307: }

3311: /*@
3312:    MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3313:    (the default parallel PETSc format).

3315:    Collective on MPI_Comm

3317:    Input Parameters:
3318: +  B - the matrix
3319: .  i - the indices into j for the start of each local row (starts with zero)
3320: .  j - the column indices for each local row (starts with zero)
3321: -  v - optional values in the matrix

3323:    Level: developer

3325:    Notes:
3326:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3327:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
3328:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3330:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3332:        The format which is used for the sparse matrix input, is equivalent to a
3333:     row-major ordering.. i.e for the following matrix, the input data expected is
3334:     as shown

3336: $        1 0 0
3337: $        2 0 3     P0
3338: $       -------
3339: $        4 5 6     P1
3340: $
3341: $     Process0 [P0]: rows_owned=[0,1]
3342: $        i =  {0,1,3}  [size = nrow+1  = 2+1]
3343: $        j =  {0,0,2}  [size = 3]
3344: $        v =  {1,2,3}  [size = 3]
3345: $
3346: $     Process1 [P1]: rows_owned=[2]
3347: $        i =  {0,3}    [size = nrow+1  = 1+1]
3348: $        j =  {0,1,2}  [size = 3]
3349: $        v =  {4,5,6}  [size = 3]

3351: .keywords: matrix, aij, compressed row, sparse, parallel

3353: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ,
3354:           MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
3355: @*/
3356: PetscErrorCode  MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
3357: {

3361:   PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
3362:   return(0);
3363: }

3367: /*@C
3368:    MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
3369:    (the default parallel PETSc format).  For good matrix assembly performance
3370:    the user should preallocate the matrix storage by setting the parameters
3371:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3372:    performance can be increased by more than a factor of 50.

3374:    Collective on MPI_Comm

3376:    Input Parameters:
3377: +  B - the matrix
3378: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
3379:            (same value is used for all local rows)
3380: .  d_nnz - array containing the number of nonzeros in the various rows of the
3381:            DIAGONAL portion of the local submatrix (possibly different for each row)
3382:            or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
3383:            The size of this array is equal to the number of local rows, i.e 'm'.
3384:            For matrices that will be factored, you must leave room for (and set)
3385:            the diagonal entry even if it is zero.
3386: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
3387:            submatrix (same value is used for all local rows).
3388: -  o_nnz - array containing the number of nonzeros in the various rows of the
3389:            OFF-DIAGONAL portion of the local submatrix (possibly different for
3390:            each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
3391:            structure. The size of this array is equal to the number
3392:            of local rows, i.e 'm'.

3394:    If the *_nnz parameter is given then the *_nz parameter is ignored

3396:    The AIJ format (also called the Yale sparse matrix format or
3397:    compressed row storage (CSR)), is fully compatible with standard Fortran 77
3398:    storage.  The stored row and column indices begin with zero.
3399:    See Users-Manual: ch_mat for details.

3401:    The parallel matrix is partitioned such that the first m0 rows belong to
3402:    process 0, the next m1 rows belong to process 1, the next m2 rows belong
3403:    to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.

3405:    The DIAGONAL portion of the local submatrix of a processor can be defined
3406:    as the submatrix which is obtained by extraction the part corresponding to
3407:    the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
3408:    first row that belongs to the processor, r2 is the last row belonging to
3409:    the this processor, and c1-c2 is range of indices of the local part of a
3410:    vector suitable for applying the matrix to.  This is an mxn matrix.  In the
3411:    common case of a square matrix, the row and column ranges are the same and
3412:    the DIAGONAL part is also square. The remaining portion of the local
3413:    submatrix (mxN) constitute the OFF-DIAGONAL portion.

3415:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

3417:    You can call MatGetInfo() to get information on how effective the preallocation was;
3418:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3419:    You can also run with the option -info and look for messages with the string
3420:    malloc in them to see if additional memory allocation was needed.

3422:    Example usage:

3424:    Consider the following 8x8 matrix with 34 non-zero values, that is
3425:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3426:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
3427:    as follows:

3429: .vb
3430:             1  2  0  |  0  3  0  |  0  4
3431:     Proc0   0  5  6  |  7  0  0  |  8  0
3432:             9  0 10  | 11  0  0  | 12  0
3433:     -------------------------------------
3434:            13  0 14  | 15 16 17  |  0  0
3435:     Proc1   0 18  0  | 19 20 21  |  0  0
3436:             0  0  0  | 22 23  0  | 24  0
3437:     -------------------------------------
3438:     Proc2  25 26 27  |  0  0 28  | 29  0
3439:            30  0  0  | 31 32 33  |  0 34
3440: .ve

3442:    This can be represented as a collection of submatrices as:

3444: .vb
3445:       A B C
3446:       D E F
3447:       G H I
3448: .ve

3450:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3451:    owned by proc1, G,H,I are owned by proc2.

3453:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3454:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3455:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3457:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3458:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3459:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3460:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3461:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3462:    matrix, ans [DF] as another SeqAIJ matrix.

3464:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3465:    allocated for every row of the local diagonal submatrix, and o_nz
3466:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3467:    One way to choose d_nz and o_nz is to use the max nonzerors per local
3468:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
3469:    In this case, the values of d_nz,o_nz are:
3470: .vb
3471:      proc0 : dnz = 2, o_nz = 2
3472:      proc1 : dnz = 3, o_nz = 2
3473:      proc2 : dnz = 1, o_nz = 4
3474: .ve
3475:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3476:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3477:    for proc3. i.e we are using 12+15+10=37 storage locations to store
3478:    34 values.

3480:    When d_nnz, o_nnz parameters are specified, the storage is specified
3481:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3482:    In the above case the values for d_nnz,o_nnz are:
3483: .vb
3484:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3485:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3486:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3487: .ve
3488:    Here the space allocated is sum of all the above values i.e 34, and
3489:    hence pre-allocation is perfect.

3491:    Level: intermediate

3493: .keywords: matrix, aij, compressed row, sparse, parallel

3495: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
3496:           MPIAIJ, MatGetInfo(), PetscSplitOwnership()
3497: @*/
3498: PetscErrorCode  MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3499: {

3505:   PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
3506:   return(0);
3507: }

3511: /*@
3512:      MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
3513:          CSR format the local rows.

3515:    Collective on MPI_Comm

3517:    Input Parameters:
3518: +  comm - MPI communicator
3519: .  m - number of local rows (Cannot be PETSC_DECIDE)
3520: .  n - This value should be the same as the local size used in creating the
3521:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3522:        calculated if N is given) For square matrices n is almost always m.
3523: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3524: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3525: .   i - row indices
3526: .   j - column indices
3527: -   a - matrix values

3529:    Output Parameter:
3530: .   mat - the matrix

3532:    Level: intermediate

3534:    Notes:
3535:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3536:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
3537:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3539:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3541:        The format which is used for the sparse matrix input, is equivalent to a
3542:     row-major ordering.. i.e for the following matrix, the input data expected is
3543:     as shown

3545: $        1 0 0
3546: $        2 0 3     P0
3547: $       -------
3548: $        4 5 6     P1
3549: $
3550: $     Process0 [P0]: rows_owned=[0,1]
3551: $        i =  {0,1,3}  [size = nrow+1  = 2+1]
3552: $        j =  {0,0,2}  [size = 3]
3553: $        v =  {1,2,3}  [size = 3]
3554: $
3555: $     Process1 [P1]: rows_owned=[2]
3556: $        i =  {0,3}    [size = nrow+1  = 1+1]
3557: $        j =  {0,1,2}  [size = 3]
3558: $        v =  {4,5,6}  [size = 3]

3560: .keywords: matrix, aij, compressed row, sparse, parallel

3562: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3563:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
3564: @*/
3565: PetscErrorCode  MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
3566: {

3570:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3571:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
3572:   MatCreate(comm,mat);
3573:   MatSetSizes(*mat,m,n,M,N);
3574:   /* MatSetBlockSizes(M,bs,cbs); */
3575:   MatSetType(*mat,MATMPIAIJ);
3576:   MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
3577:   return(0);
3578: }

3582: /*@C
3583:    MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
3584:    (the default parallel PETSc format).  For good matrix assembly performance
3585:    the user should preallocate the matrix storage by setting the parameters
3586:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3587:    performance can be increased by more than a factor of 50.

3589:    Collective on MPI_Comm

3591:    Input Parameters:
3592: +  comm - MPI communicator
3593: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
3594:            This value should be the same as the local size used in creating the
3595:            y vector for the matrix-vector product y = Ax.
3596: .  n - This value should be the same as the local size used in creating the
3597:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3598:        calculated if N is given) For square matrices n is almost always m.
3599: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3600: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3601: .  d_nz  - number of nonzeros per row in DIAGONAL portion of local submatrix
3602:            (same value is used for all local rows)
3603: .  d_nnz - array containing the number of nonzeros in the various rows of the
3604:            DIAGONAL portion of the local submatrix (possibly different for each row)
3605:            or NULL, if d_nz is used to specify the nonzero structure.
3606:            The size of this array is equal to the number of local rows, i.e 'm'.
3607: .  o_nz  - number of nonzeros per row in the OFF-DIAGONAL portion of local
3608:            submatrix (same value is used for all local rows).
3609: -  o_nnz - array containing the number of nonzeros in the various rows of the
3610:            OFF-DIAGONAL portion of the local submatrix (possibly different for
3611:            each row) or NULL, if o_nz is used to specify the nonzero
3612:            structure. The size of this array is equal to the number
3613:            of local rows, i.e 'm'.

3615:    Output Parameter:
3616: .  A - the matrix

3618:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3619:    MatXXXXSetPreallocation() paradgm instead of this routine directly.
3620:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

3622:    Notes:
3623:    If the *_nnz parameter is given then the *_nz parameter is ignored

3625:    m,n,M,N parameters specify the size of the matrix, and its partitioning across
3626:    processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
3627:    storage requirements for this matrix.

3629:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one
3630:    processor than it must be used on all processors that share the object for
3631:    that argument.

3633:    The user MUST specify either the local or global matrix dimensions
3634:    (possibly both).

3636:    The parallel matrix is partitioned across processors such that the
3637:    first m0 rows belong to process 0, the next m1 rows belong to
3638:    process 1, the next m2 rows belong to process 2 etc.. where
3639:    m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
3640:    values corresponding to [m x N] submatrix.

3642:    The columns are logically partitioned with the n0 columns belonging
3643:    to 0th partition, the next n1 columns belonging to the next
3644:    partition etc.. where n0,n1,n2... are the input parameter 'n'.

3646:    The DIAGONAL portion of the local submatrix on any given processor
3647:    is the submatrix corresponding to the rows and columns m,n
3648:    corresponding to the given processor. i.e diagonal matrix on
3649:    process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
3650:    etc. The remaining portion of the local submatrix [m x (N-n)]
3651:    constitute the OFF-DIAGONAL portion. The example below better
3652:    illustrates this concept.

3654:    For a square global matrix we define each processor's diagonal portion
3655:    to be its local rows and the corresponding columns (a square submatrix);
3656:    each processor's off-diagonal portion encompasses the remainder of the
3657:    local matrix (a rectangular submatrix).

3659:    If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.

3661:    When calling this routine with a single process communicator, a matrix of
3662:    type SEQAIJ is returned.  If a matrix of type MPIAIJ is desired for this
3663:    type of communicator, use the construction mechanism:
3664:      MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);

3666:    By default, this format uses inodes (identical nodes) when possible.
3667:    We search for consecutive rows with the same nonzero structure, thereby
3668:    reusing matrix information to achieve increased efficiency.

3670:    Options Database Keys:
3671: +  -mat_no_inode  - Do not use inodes
3672: .  -mat_inode_limit <limit> - Sets inode limit (max limit=5)
3673: -  -mat_aij_oneindex - Internally use indexing starting at 1
3674:         rather than 0.  Note that when calling MatSetValues(),
3675:         the user still MUST index entries starting at 0!


3678:    Example usage:

3680:    Consider the following 8x8 matrix with 34 non-zero values, that is
3681:    assembled across 3 processors. Lets assume that proc0 owns 3 rows,
3682:    proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
3683:    as follows:

3685: .vb
3686:             1  2  0  |  0  3  0  |  0  4
3687:     Proc0   0  5  6  |  7  0  0  |  8  0
3688:             9  0 10  | 11  0  0  | 12  0
3689:     -------------------------------------
3690:            13  0 14  | 15 16 17  |  0  0
3691:     Proc1   0 18  0  | 19 20 21  |  0  0
3692:             0  0  0  | 22 23  0  | 24  0
3693:     -------------------------------------
3694:     Proc2  25 26 27  |  0  0 28  | 29  0
3695:            30  0  0  | 31 32 33  |  0 34
3696: .ve

3698:    This can be represented as a collection of submatrices as:

3700: .vb
3701:       A B C
3702:       D E F
3703:       G H I
3704: .ve

3706:    Where the submatrices A,B,C are owned by proc0, D,E,F are
3707:    owned by proc1, G,H,I are owned by proc2.

3709:    The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3710:    The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
3711:    The 'M','N' parameters are 8,8, and have the same values on all procs.

3713:    The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
3714:    submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
3715:    corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
3716:    Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
3717:    part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
3718:    matrix, ans [DF] as another SeqAIJ matrix.

3720:    When d_nz, o_nz parameters are specified, d_nz storage elements are
3721:    allocated for every row of the local diagonal submatrix, and o_nz
3722:    storage locations are allocated for every row of the OFF-DIAGONAL submat.
3723:    One way to choose d_nz and o_nz is to use the max nonzerors per local
3724:    rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
3725:    In this case, the values of d_nz,o_nz are:
3726: .vb
3727:      proc0 : dnz = 2, o_nz = 2
3728:      proc1 : dnz = 3, o_nz = 2
3729:      proc2 : dnz = 1, o_nz = 4
3730: .ve
3731:    We are allocating m*(d_nz+o_nz) storage locations for every proc. This
3732:    translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
3733:    for proc3. i.e we are using 12+15+10=37 storage locations to store
3734:    34 values.

3736:    When d_nnz, o_nnz parameters are specified, the storage is specified
3737:    for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
3738:    In the above case the values for d_nnz,o_nnz are:
3739: .vb
3740:      proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
3741:      proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
3742:      proc2: d_nnz = [1,1]   and o_nnz = [4,4]
3743: .ve
3744:    Here the space allocated is sum of all the above values i.e 34, and
3745:    hence pre-allocation is perfect.

3747:    Level: intermediate

3749: .keywords: matrix, aij, compressed row, sparse, parallel

3751: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3752:           MPIAIJ, MatCreateMPIAIJWithArrays()
3753: @*/
3754: PetscErrorCode  MatCreateAIJ(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)
3755: {
3757:   PetscMPIInt    size;

3760:   MatCreate(comm,A);
3761:   MatSetSizes(*A,m,n,M,N);
3762:   MPI_Comm_size(comm,&size);
3763:   if (size > 1) {
3764:     MatSetType(*A,MATMPIAIJ);
3765:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
3766:   } else {
3767:     MatSetType(*A,MATSEQAIJ);
3768:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
3769:   }
3770:   return(0);
3771: }

3775: PetscErrorCode  MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
3776: {
3777:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
3778:   PetscBool      flg;
3780: 
3782:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&flg);
3783:   if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MPIAIJ matrix as input");
3784:   if (Ad)     *Ad     = a->A;
3785:   if (Ao)     *Ao     = a->B;
3786:   if (colmap) *colmap = a->garray;
3787:   return(0);
3788: }

3792: PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring)
3793: {
3795:   PetscInt       i;
3796:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

3799:   if (coloring->ctype == IS_COLORING_GLOBAL) {
3800:     ISColoringValue *allcolors,*colors;
3801:     ISColoring      ocoloring;

3803:     /* set coloring for diagonal portion */
3804:     MatSetColoring_SeqAIJ(a->A,coloring);

3806:     /* set coloring for off-diagonal portion */
3807:     ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);
3808:     PetscMalloc1(a->B->cmap->n+1,&colors);
3809:     for (i=0; i<a->B->cmap->n; i++) {
3810:       colors[i] = allcolors[a->garray[i]];
3811:     }
3812:     PetscFree(allcolors);
3813:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);
3814:     MatSetColoring_SeqAIJ(a->B,ocoloring);
3815:     ISColoringDestroy(&ocoloring);
3816:   } else if (coloring->ctype == IS_COLORING_GHOSTED) {
3817:     ISColoringValue *colors;
3818:     PetscInt        *larray;
3819:     ISColoring      ocoloring;

3821:     /* set coloring for diagonal portion */
3822:     PetscMalloc1(a->A->cmap->n+1,&larray);
3823:     for (i=0; i<a->A->cmap->n; i++) {
3824:       larray[i] = i + A->cmap->rstart;
3825:     }
3826:     ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);
3827:     PetscMalloc1(a->A->cmap->n+1,&colors);
3828:     for (i=0; i<a->A->cmap->n; i++) {
3829:       colors[i] = coloring->colors[larray[i]];
3830:     }
3831:     PetscFree(larray);
3832:     ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);
3833:     MatSetColoring_SeqAIJ(a->A,ocoloring);
3834:     ISColoringDestroy(&ocoloring);

3836:     /* set coloring for off-diagonal portion */
3837:     PetscMalloc1(a->B->cmap->n+1,&larray);
3838:     ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);
3839:     PetscMalloc1(a->B->cmap->n+1,&colors);
3840:     for (i=0; i<a->B->cmap->n; i++) {
3841:       colors[i] = coloring->colors[larray[i]];
3842:     }
3843:     PetscFree(larray);
3844:     ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);
3845:     MatSetColoring_SeqAIJ(a->B,ocoloring);
3846:     ISColoringDestroy(&ocoloring);
3847:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype);
3848:   return(0);
3849: }

3853: PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues)
3854: {
3855:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;

3859:   MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);
3860:   MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);
3861:   return(0);
3862: }

3866: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
3867: {
3869:   PetscInt       m,N,i,rstart,nnz,Ii;
3870:   PetscInt       *indx;
3871:   PetscScalar    *values;

3874:   MatGetSize(inmat,&m,&N);
3875:   if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
3876:     PetscInt       *dnz,*onz,sum,bs,cbs;

3878:     if (n == PETSC_DECIDE) {
3879:       PetscSplitOwnership(comm,&n,&N);
3880:     }
3881:     /* Check sum(n) = N */
3882:     MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
3883:     if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N);

3885:     MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
3886:     rstart -= m;

3888:     MatPreallocateInitialize(comm,m,n,dnz,onz);
3889:     for (i=0; i<m; i++) {
3890:       MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
3891:       MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
3892:       MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
3893:     }

3895:     MatCreate(comm,outmat);
3896:     MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
3897:     MatGetBlockSizes(inmat,&bs,&cbs);
3898:     MatSetBlockSizes(*outmat,bs,cbs);
3899:     MatSetType(*outmat,MATMPIAIJ);
3900:     MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
3901:     MatPreallocateFinalize(dnz,onz);
3902:   }

3904:   /* numeric phase */
3905:   MatGetOwnershipRange(*outmat,&rstart,NULL);
3906:   for (i=0; i<m; i++) {
3907:     MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3908:     Ii   = i + rstart;
3909:     MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
3910:     MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
3911:   }
3912:   MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
3913:   MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
3914:   return(0);
3915: }

3919: PetscErrorCode MatFileSplit(Mat A,char *outfile)
3920: {
3921:   PetscErrorCode    ierr;
3922:   PetscMPIInt       rank;
3923:   PetscInt          m,N,i,rstart,nnz;
3924:   size_t            len;
3925:   const PetscInt    *indx;
3926:   PetscViewer       out;
3927:   char              *name;
3928:   Mat               B;
3929:   const PetscScalar *values;

3932:   MatGetLocalSize(A,&m,0);
3933:   MatGetSize(A,0,&N);
3934:   /* Should this be the type of the diagonal block of A? */
3935:   MatCreate(PETSC_COMM_SELF,&B);
3936:   MatSetSizes(B,m,N,m,N);
3937:   MatSetBlockSizesFromMats(B,A,A);
3938:   MatSetType(B,MATSEQAIJ);
3939:   MatSeqAIJSetPreallocation(B,0,NULL);
3940:   MatGetOwnershipRange(A,&rstart,0);
3941:   for (i=0; i<m; i++) {
3942:     MatGetRow(A,i+rstart,&nnz,&indx,&values);
3943:     MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
3944:     MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
3945:   }
3946:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3947:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

3949:   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
3950:   PetscStrlen(outfile,&len);
3951:   PetscMalloc1(len+5,&name);
3952:   sprintf(name,"%s.%d",outfile,rank);
3953:   PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
3954:   PetscFree(name);
3955:   MatView(B,out);
3956:   PetscViewerDestroy(&out);
3957:   MatDestroy(&B);
3958:   return(0);
3959: }

3961: extern PetscErrorCode MatDestroy_MPIAIJ(Mat);
3964: PetscErrorCode  MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
3965: {
3966:   PetscErrorCode      ierr;
3967:   Mat_Merge_SeqsToMPI *merge;
3968:   PetscContainer      container;

3971:   PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);
3972:   if (container) {
3973:     PetscContainerGetPointer(container,(void**)&merge);
3974:     PetscFree(merge->id_r);
3975:     PetscFree(merge->len_s);
3976:     PetscFree(merge->len_r);
3977:     PetscFree(merge->bi);
3978:     PetscFree(merge->bj);
3979:     PetscFree(merge->buf_ri[0]);
3980:     PetscFree(merge->buf_ri);
3981:     PetscFree(merge->buf_rj[0]);
3982:     PetscFree(merge->buf_rj);
3983:     PetscFree(merge->coi);
3984:     PetscFree(merge->coj);
3985:     PetscFree(merge->owners_co);
3986:     PetscLayoutDestroy(&merge->rowmap);
3987:     PetscFree(merge);
3988:     PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
3989:   }
3990:   MatDestroy_MPIAIJ(A);
3991:   return(0);
3992: }

3994: #include <../src/mat/utils/freespace.h>
3995: #include <petscbt.h>

3999: PetscErrorCode  MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4000: {
4001:   PetscErrorCode      ierr;
4002:   MPI_Comm            comm;
4003:   Mat_SeqAIJ          *a  =(Mat_SeqAIJ*)seqmat->data;
4004:   PetscMPIInt         size,rank,taga,*len_s;
4005:   PetscInt            N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4006:   PetscInt            proc,m;
4007:   PetscInt            **buf_ri,**buf_rj;
4008:   PetscInt            k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4009:   PetscInt            nrows,**buf_ri_k,**nextrow,**nextai;
4010:   MPI_Request         *s_waits,*r_waits;
4011:   MPI_Status          *status;
4012:   MatScalar           *aa=a->a;
4013:   MatScalar           **abuf_r,*ba_i;
4014:   Mat_Merge_SeqsToMPI *merge;
4015:   PetscContainer      container;

4018:   PetscObjectGetComm((PetscObject)mpimat,&comm);
4019:   PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);

4021:   MPI_Comm_size(comm,&size);
4022:   MPI_Comm_rank(comm,&rank);

4024:   PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4025:   PetscContainerGetPointer(container,(void**)&merge);

4027:   bi     = merge->bi;
4028:   bj     = merge->bj;
4029:   buf_ri = merge->buf_ri;
4030:   buf_rj = merge->buf_rj;

4032:   PetscMalloc1(size,&status);
4033:   owners = merge->rowmap->range;
4034:   len_s  = merge->len_s;

4036:   /* send and recv matrix values */
4037:   /*-----------------------------*/
4038:   PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4039:   PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);

4041:   PetscMalloc1(merge->nsend+1,&s_waits);
4042:   for (proc=0,k=0; proc<size; proc++) {
4043:     if (!len_s[proc]) continue;
4044:     i    = owners[proc];
4045:     MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4046:     k++;
4047:   }

4049:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4050:   if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4051:   PetscFree(status);

4053:   PetscFree(s_waits);
4054:   PetscFree(r_waits);

4056:   /* insert mat values of mpimat */
4057:   /*----------------------------*/
4058:   PetscMalloc1(N,&ba_i);
4059:   PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);

4061:   for (k=0; k<merge->nrecv; k++) {
4062:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4063:     nrows       = *(buf_ri_k[k]);
4064:     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
4065:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure  */
4066:   }

4068:   /* set values of ba */
4069:   m = merge->rowmap->n;
4070:   for (i=0; i<m; i++) {
4071:     arow = owners[rank] + i;
4072:     bj_i = bj+bi[i];  /* col indices of the i-th row of mpimat */
4073:     bnzi = bi[i+1] - bi[i];
4074:     PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));

4076:     /* add local non-zero vals of this proc's seqmat into ba */
4077:     anzi   = ai[arow+1] - ai[arow];
4078:     aj     = a->j + ai[arow];
4079:     aa     = a->a + ai[arow];
4080:     nextaj = 0;
4081:     for (j=0; nextaj<anzi; j++) {
4082:       if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4083:         ba_i[j] += aa[nextaj++];
4084:       }
4085:     }

4087:     /* add received vals into ba */
4088:     for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4089:       /* i-th row */
4090:       if (i == *nextrow[k]) {
4091:         anzi   = *(nextai[k]+1) - *nextai[k];
4092:         aj     = buf_rj[k] + *(nextai[k]);
4093:         aa     = abuf_r[k] + *(nextai[k]);
4094:         nextaj = 0;
4095:         for (j=0; nextaj<anzi; j++) {
4096:           if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4097:             ba_i[j] += aa[nextaj++];
4098:           }
4099:         }
4100:         nextrow[k]++; nextai[k]++;
4101:       }
4102:     }
4103:     MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4104:   }
4105:   MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4106:   MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);

4108:   PetscFree(abuf_r[0]);
4109:   PetscFree(abuf_r);
4110:   PetscFree(ba_i);
4111:   PetscFree3(buf_ri_k,nextrow,nextai);
4112:   PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4113:   return(0);
4114: }

4116: extern PetscErrorCode  MatDestroy_MPIAIJ_SeqsToMPI(Mat);

4120: PetscErrorCode  MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4121: {
4122:   PetscErrorCode      ierr;
4123:   Mat                 B_mpi;
4124:   Mat_SeqAIJ          *a=(Mat_SeqAIJ*)seqmat->data;
4125:   PetscMPIInt         size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4126:   PetscInt            **buf_rj,**buf_ri,**buf_ri_k;
4127:   PetscInt            M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4128:   PetscInt            len,proc,*dnz,*onz,bs,cbs;
4129:   PetscInt            k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4130:   PetscInt            nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4131:   MPI_Request         *si_waits,*sj_waits,*ri_waits,*rj_waits;
4132:   MPI_Status          *status;
4133:   PetscFreeSpaceList  free_space=NULL,current_space=NULL;
4134:   PetscBT             lnkbt;
4135:   Mat_Merge_SeqsToMPI *merge;
4136:   PetscContainer      container;

4139:   PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);

4141:   /* make sure it is a PETSc comm */
4142:   PetscCommDuplicate(comm,&comm,NULL);
4143:   MPI_Comm_size(comm,&size);
4144:   MPI_Comm_rank(comm,&rank);

4146:   PetscNew(&merge);
4147:   PetscMalloc1(size,&status);

4149:   /* determine row ownership */
4150:   /*---------------------------------------------------------*/
4151:   PetscLayoutCreate(comm,&merge->rowmap);
4152:   PetscLayoutSetLocalSize(merge->rowmap,m);
4153:   PetscLayoutSetSize(merge->rowmap,M);
4154:   PetscLayoutSetBlockSize(merge->rowmap,1);
4155:   PetscLayoutSetUp(merge->rowmap);
4156:   PetscMalloc1(size,&len_si);
4157:   PetscMalloc1(size,&merge->len_s);

4159:   m      = merge->rowmap->n;
4160:   owners = merge->rowmap->range;

4162:   /* determine the number of messages to send, their lengths */
4163:   /*---------------------------------------------------------*/
4164:   len_s = merge->len_s;

4166:   len          = 0; /* length of buf_si[] */
4167:   merge->nsend = 0;
4168:   for (proc=0; proc<size; proc++) {
4169:     len_si[proc] = 0;
4170:     if (proc == rank) {
4171:       len_s[proc] = 0;
4172:     } else {
4173:       len_si[proc] = owners[proc+1] - owners[proc] + 1;
4174:       len_s[proc]  = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4175:     }
4176:     if (len_s[proc]) {
4177:       merge->nsend++;
4178:       nrows = 0;
4179:       for (i=owners[proc]; i<owners[proc+1]; i++) {
4180:         if (ai[i+1] > ai[i]) nrows++;
4181:       }
4182:       len_si[proc] = 2*(nrows+1);
4183:       len         += len_si[proc];
4184:     }
4185:   }

4187:   /* determine the number and length of messages to receive for ij-structure */
4188:   /*-------------------------------------------------------------------------*/
4189:   PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4190:   PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);

4192:   /* post the Irecv of j-structure */
4193:   /*-------------------------------*/
4194:   PetscCommGetNewTag(comm,&tagj);
4195:   PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);

4197:   /* post the Isend of j-structure */
4198:   /*--------------------------------*/
4199:   PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);

4201:   for (proc=0, k=0; proc<size; proc++) {
4202:     if (!len_s[proc]) continue;
4203:     i    = owners[proc];
4204:     MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4205:     k++;
4206:   }

4208:   /* receives and sends of j-structure are complete */
4209:   /*------------------------------------------------*/
4210:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4211:   if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}

4213:   /* send and recv i-structure */
4214:   /*---------------------------*/
4215:   PetscCommGetNewTag(comm,&tagi);
4216:   PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);

4218:   PetscMalloc1(len+1,&buf_s);
4219:   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
4220:   for (proc=0,k=0; proc<size; proc++) {
4221:     if (!len_s[proc]) continue;
4222:     /* form outgoing message for i-structure:
4223:          buf_si[0]:                 nrows to be sent
4224:                [1:nrows]:           row index (global)
4225:                [nrows+1:2*nrows+1]: i-structure index
4226:     */
4227:     /*-------------------------------------------*/
4228:     nrows       = len_si[proc]/2 - 1;
4229:     buf_si_i    = buf_si + nrows+1;
4230:     buf_si[0]   = nrows;
4231:     buf_si_i[0] = 0;
4232:     nrows       = 0;
4233:     for (i=owners[proc]; i<owners[proc+1]; i++) {
4234:       anzi = ai[i+1] - ai[i];
4235:       if (anzi) {
4236:         buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4237:         buf_si[nrows+1]   = i-owners[proc]; /* local row index */
4238:         nrows++;
4239:       }
4240:     }
4241:     MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4242:     k++;
4243:     buf_si += len_si[proc];
4244:   }

4246:   if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4247:   if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}

4249:   PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4250:   for (i=0; i<merge->nrecv; i++) {
4251:     PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4252:   }

4254:   PetscFree(len_si);
4255:   PetscFree(len_ri);
4256:   PetscFree(rj_waits);
4257:   PetscFree2(si_waits,sj_waits);
4258:   PetscFree(ri_waits);
4259:   PetscFree(buf_s);
4260:   PetscFree(status);

4262:   /* compute a local seq matrix in each processor */
4263:   /*----------------------------------------------*/
4264:   /* allocate bi array and free space for accumulating nonzero column info */
4265:   PetscMalloc1(m+1,&bi);
4266:   bi[0] = 0;

4268:   /* create and initialize a linked list */
4269:   nlnk = N+1;
4270:   PetscLLCreate(N,N,nlnk,lnk,lnkbt);

4272:   /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4273:   len  = ai[owners[rank+1]] - ai[owners[rank]];
4274:   PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);

4276:   current_space = free_space;

4278:   /* determine symbolic info for each local row */
4279:   PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);

4281:   for (k=0; k<merge->nrecv; k++) {
4282:     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4283:     nrows       = *buf_ri_k[k];
4284:     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
4285:     nextai[k]   = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure  */
4286:   }

4288:   MatPreallocateInitialize(comm,m,n,dnz,onz);
4289:   len  = 0;
4290:   for (i=0; i<m; i++) {
4291:     bnzi = 0;
4292:     /* add local non-zero cols of this proc's seqmat into lnk */
4293:     arow  = owners[rank] + i;
4294:     anzi  = ai[arow+1] - ai[arow];
4295:     aj    = a->j + ai[arow];
4296:     PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4297:     bnzi += nlnk;
4298:     /* add received col data into lnk */
4299:     for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4300:       if (i == *nextrow[k]) { /* i-th row */
4301:         anzi  = *(nextai[k]+1) - *nextai[k];
4302:         aj    = buf_rj[k] + *nextai[k];
4303:         PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4304:         bnzi += nlnk;
4305:         nextrow[k]++; nextai[k]++;
4306:       }
4307:     }
4308:     if (len < bnzi) len = bnzi;  /* =max(bnzi) */

4310:     /* if free space is not available, make more free space */
4311:     if (current_space->local_remaining<bnzi) {
4312:       PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),&current_space);
4313:       nspacedouble++;
4314:     }
4315:     /* copy data into free space, then initialize lnk */
4316:     PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4317:     MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);

4319:     current_space->array           += bnzi;
4320:     current_space->local_used      += bnzi;
4321:     current_space->local_remaining -= bnzi;

4323:     bi[i+1] = bi[i] + bnzi;
4324:   }

4326:   PetscFree3(buf_ri_k,nextrow,nextai);

4328:   PetscMalloc1(bi[m]+1,&bj);
4329:   PetscFreeSpaceContiguous(&free_space,bj);
4330:   PetscLLDestroy(lnk,lnkbt);

4332:   /* create symbolic parallel matrix B_mpi */
4333:   /*---------------------------------------*/
4334:   MatGetBlockSizes(seqmat,&bs,&cbs);
4335:   MatCreate(comm,&B_mpi);
4336:   if (n==PETSC_DECIDE) {
4337:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4338:   } else {
4339:     MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4340:   }
4341:   MatSetBlockSizes(B_mpi,bs,cbs);
4342:   MatSetType(B_mpi,MATMPIAIJ);
4343:   MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4344:   MatPreallocateFinalize(dnz,onz);
4345:   MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);

4347:   /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4348:   B_mpi->assembled    = PETSC_FALSE;
4349:   B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
4350:   merge->bi           = bi;
4351:   merge->bj           = bj;
4352:   merge->buf_ri       = buf_ri;
4353:   merge->buf_rj       = buf_rj;
4354:   merge->coi          = NULL;
4355:   merge->coj          = NULL;
4356:   merge->owners_co    = NULL;

4358:   PetscCommDestroy(&comm);

4360:   /* attach the supporting struct to B_mpi for reuse */
4361:   PetscContainerCreate(PETSC_COMM_SELF,&container);
4362:   PetscContainerSetPointer(container,merge);
4363:   PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4364:   PetscContainerDestroy(&container);
4365:   *mpimat = B_mpi;

4367:   PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4368:   return(0);
4369: }

4373: /*@C
4374:       MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential
4375:                  matrices from each processor

4377:     Collective on MPI_Comm

4379:    Input Parameters:
4380: +    comm - the communicators the parallel matrix will live on
4381: .    seqmat - the input sequential matrices
4382: .    m - number of local rows (or PETSC_DECIDE)
4383: .    n - number of local columns (or PETSC_DECIDE)
4384: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4386:    Output Parameter:
4387: .    mpimat - the parallel matrix generated

4389:     Level: advanced

4391:    Notes:
4392:      The dimensions of the sequential matrix in each processor MUST be the same.
4393:      The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4394:      destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4395: @*/
4396: PetscErrorCode  MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4397: {
4399:   PetscMPIInt    size;

4402:   MPI_Comm_size(comm,&size);
4403:   if (size == 1) {
4404:     PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4405:     if (scall == MAT_INITIAL_MATRIX) {
4406:       MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4407:     } else {
4408:       MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4409:     }
4410:     PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4411:     return(0);
4412:   }
4413:   PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4414:   if (scall == MAT_INITIAL_MATRIX) {
4415:     MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4416:   }
4417:   MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4418:   PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4419:   return(0);
4420: }

4424: /*@
4425:      MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with
4426:           mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4427:           with MatGetSize()

4429:     Not Collective

4431:    Input Parameters:
4432: +    A - the matrix
4433: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4435:    Output Parameter:
4436: .    A_loc - the local sequential matrix generated

4438:     Level: developer

4440: .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed()

4442: @*/
4443: PetscErrorCode  MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
4444: {
4446:   Mat_MPIAIJ     *mpimat=(Mat_MPIAIJ*)A->data;
4447:   Mat_SeqAIJ     *mat,*a,*b;
4448:   PetscInt       *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
4449:   MatScalar      *aa,*ba,*cam;
4450:   PetscScalar    *ca;
4451:   PetscInt       am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
4452:   PetscInt       *ci,*cj,col,ncols_d,ncols_o,jo;
4453:   PetscBool      match;
4454:   MPI_Comm       comm;
4455:   PetscMPIInt    size;

4458:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
4459:   if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
4460:   PetscObjectGetComm((PetscObject)A,&comm);
4461:   MPI_Comm_size(comm,&size);
4462:   if (size == 1 && scall == MAT_REUSE_MATRIX) return(0);

4464:   PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
4465:   a = (Mat_SeqAIJ*)(mpimat->A)->data;
4466:   b = (Mat_SeqAIJ*)(mpimat->B)->data;
4467:   ai = a->i; aj = a->j; bi = b->i; bj = b->j;
4468:   aa = a->a; ba = b->a;
4469:   if (scall == MAT_INITIAL_MATRIX) {
4470:     if (size == 1) {
4471:       MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);
4472:       return(0);
4473:     }

4475:     PetscMalloc1(1+am,&ci);
4476:     ci[0] = 0;
4477:     for (i=0; i<am; i++) {
4478:       ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
4479:     }
4480:     PetscMalloc1(1+ci[am],&cj);
4481:     PetscMalloc1(1+ci[am],&ca);
4482:     k    = 0;
4483:     for (i=0; i<am; i++) {
4484:       ncols_o = bi[i+1] - bi[i];
4485:       ncols_d = ai[i+1] - ai[i];
4486:       /* off-diagonal portion of A */
4487:       for (jo=0; jo<ncols_o; jo++) {
4488:         col = cmap[*bj];
4489:         if (col >= cstart) break;
4490:         cj[k]   = col; bj++;
4491:         ca[k++] = *ba++;
4492:       }
4493:       /* diagonal portion of A */
4494:       for (j=0; j<ncols_d; j++) {
4495:         cj[k]   = cstart + *aj++;
4496:         ca[k++] = *aa++;
4497:       }
4498:       /* off-diagonal portion of A */
4499:       for (j=jo; j<ncols_o; j++) {
4500:         cj[k]   = cmap[*bj++];
4501:         ca[k++] = *ba++;
4502:       }
4503:     }
4504:     /* put together the new matrix */
4505:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
4506:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4507:     /* Since these are PETSc arrays, change flags to free them as necessary. */
4508:     mat          = (Mat_SeqAIJ*)(*A_loc)->data;
4509:     mat->free_a  = PETSC_TRUE;
4510:     mat->free_ij = PETSC_TRUE;
4511:     mat->nonew   = 0;
4512:   } else if (scall == MAT_REUSE_MATRIX) {
4513:     mat=(Mat_SeqAIJ*)(*A_loc)->data;
4514:     ci = mat->i; cj = mat->j; cam = mat->a;
4515:     for (i=0; i<am; i++) {
4516:       /* off-diagonal portion of A */
4517:       ncols_o = bi[i+1] - bi[i];
4518:       for (jo=0; jo<ncols_o; jo++) {
4519:         col = cmap[*bj];
4520:         if (col >= cstart) break;
4521:         *cam++ = *ba++; bj++;
4522:       }
4523:       /* diagonal portion of A */
4524:       ncols_d = ai[i+1] - ai[i];
4525:       for (j=0; j<ncols_d; j++) *cam++ = *aa++;
4526:       /* off-diagonal portion of A */
4527:       for (j=jo; j<ncols_o; j++) {
4528:         *cam++ = *ba++; bj++;
4529:       }
4530:     }
4531:   } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
4532:   PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
4533:   return(0);
4534: }

4538: /*@C
4539:      MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns

4541:     Not Collective

4543:    Input Parameters:
4544: +    A - the matrix
4545: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4546: -    row, col - index sets of rows and columns to extract (or NULL)

4548:    Output Parameter:
4549: .    A_loc - the local sequential matrix generated

4551:     Level: developer

4553: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()

4555: @*/
4556: PetscErrorCode  MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
4557: {
4558:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)A->data;
4560:   PetscInt       i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
4561:   IS             isrowa,iscola;
4562:   Mat            *aloc;
4563:   PetscBool      match;

4566:   PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
4567:   if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input");
4568:   PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
4569:   if (!row) {
4570:     start = A->rmap->rstart; end = A->rmap->rend;
4571:     ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
4572:   } else {
4573:     isrowa = *row;
4574:   }
4575:   if (!col) {
4576:     start = A->cmap->rstart;
4577:     cmap  = a->garray;
4578:     nzA   = a->A->cmap->n;
4579:     nzB   = a->B->cmap->n;
4580:     PetscMalloc1(nzA+nzB, &idx);
4581:     ncols = 0;
4582:     for (i=0; i<nzB; i++) {
4583:       if (cmap[i] < start) idx[ncols++] = cmap[i];
4584:       else break;
4585:     }
4586:     imark = i;
4587:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;
4588:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
4589:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
4590:   } else {
4591:     iscola = *col;
4592:   }
4593:   if (scall != MAT_INITIAL_MATRIX) {
4594:     PetscMalloc1(1,&aloc);
4595:     aloc[0] = *A_loc;
4596:   }
4597:   MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
4598:   *A_loc = aloc[0];
4599:   PetscFree(aloc);
4600:   if (!row) {
4601:     ISDestroy(&isrowa);
4602:   }
4603:   if (!col) {
4604:     ISDestroy(&iscola);
4605:   }
4606:   PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
4607:   return(0);
4608: }

4612: /*@C
4613:     MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A

4615:     Collective on Mat

4617:    Input Parameters:
4618: +    A,B - the matrices in mpiaij format
4619: .    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4620: -    rowb, colb - index sets of rows and columns of B to extract (or NULL)

4622:    Output Parameter:
4623: +    rowb, colb - index sets of rows and columns of B to extract
4624: -    B_seq - the sequential matrix generated

4626:     Level: developer

4628: @*/
4629: PetscErrorCode  MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
4630: {
4631:   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)A->data;
4633:   PetscInt       *idx,i,start,ncols,nzA,nzB,*cmap,imark;
4634:   IS             isrowb,iscolb;
4635:   Mat            *bseq=NULL;

4638:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
4639:     SETERRQ4(PETSC_COMM_SELF,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);
4640:   }
4641:   PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);

4643:   if (scall == MAT_INITIAL_MATRIX) {
4644:     start = A->cmap->rstart;
4645:     cmap  = a->garray;
4646:     nzA   = a->A->cmap->n;
4647:     nzB   = a->B->cmap->n;
4648:     PetscMalloc1(nzA+nzB, &idx);
4649:     ncols = 0;
4650:     for (i=0; i<nzB; i++) {  /* row < local row index */
4651:       if (cmap[i] < start) idx[ncols++] = cmap[i];
4652:       else break;
4653:     }
4654:     imark = i;
4655:     for (i=0; i<nzA; i++) idx[ncols++] = start + i;  /* local rows */
4656:     for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
4657:     ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
4658:     ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
4659:   } else {
4660:     if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
4661:     isrowb  = *rowb; iscolb = *colb;
4662:     PetscMalloc1(1,&bseq);
4663:     bseq[0] = *B_seq;
4664:   }
4665:   MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
4666:   *B_seq = bseq[0];
4667:   PetscFree(bseq);
4668:   if (!rowb) {
4669:     ISDestroy(&isrowb);
4670:   } else {
4671:     *rowb = isrowb;
4672:   }
4673:   if (!colb) {
4674:     ISDestroy(&iscolb);
4675:   } else {
4676:     *colb = iscolb;
4677:   }
4678:   PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
4679:   return(0);
4680: }

4684: /*
4685:     MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
4686:     of the OFF-DIAGONAL portion of local A

4688:     Collective on Mat

4690:    Input Parameters:
4691: +    A,B - the matrices in mpiaij format
4692: -    scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX

4694:    Output Parameter:
4695: +    startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
4696: .    startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
4697: .    bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
4698: -    B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N

4700:     Level: developer

4702: */
4703: PetscErrorCode  MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
4704: {
4705:   VecScatter_MPI_General *gen_to,*gen_from;
4706:   PetscErrorCode         ierr;
4707:   Mat_MPIAIJ             *a=(Mat_MPIAIJ*)A->data;
4708:   Mat_SeqAIJ             *b_oth;
4709:   VecScatter             ctx =a->Mvctx;
4710:   MPI_Comm               comm;
4711:   PetscMPIInt            *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank;
4712:   PetscInt               *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj;
4713:   PetscScalar            *rvalues,*svalues;
4714:   MatScalar              *b_otha,*bufa,*bufA;
4715:   PetscInt               i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len;
4716:   MPI_Request            *rwaits = NULL,*swaits = NULL;
4717:   MPI_Status             *sstatus,rstatus;
4718:   PetscMPIInt            jj,size;
4719:   PetscInt               *cols,sbs,rbs;
4720:   PetscScalar            *vals;

4723:   PetscObjectGetComm((PetscObject)A,&comm);
4724:   MPI_Comm_size(comm,&size);

4726:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
4727:     SETERRQ4(PETSC_COMM_SELF,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);
4728:   }
4729:   PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
4730:   MPI_Comm_rank(comm,&rank);

4732:   gen_to   = (VecScatter_MPI_General*)ctx->todata;
4733:   gen_from = (VecScatter_MPI_General*)ctx->fromdata;
4734:   rvalues  = gen_from->values; /* holds the length of receiving row */
4735:   svalues  = gen_to->values;   /* holds the length of sending row */
4736:   nrecvs   = gen_from->n;
4737:   nsends   = gen_to->n;

4739:   PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);
4740:   srow    = gen_to->indices;    /* local row index to be sent */
4741:   sstarts = gen_to->starts;
4742:   sprocs  = gen_to->procs;
4743:   sstatus = gen_to->sstatus;
4744:   sbs     = gen_to->bs;
4745:   rstarts = gen_from->starts;
4746:   rprocs  = gen_from->procs;
4747:   rbs     = gen_from->bs;

4749:   if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
4750:   if (scall == MAT_INITIAL_MATRIX) {
4751:     /* i-array */
4752:     /*---------*/
4753:     /*  post receives */
4754:     for (i=0; i<nrecvs; i++) {
4755:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
4756:       nrows  = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
4757:       MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
4758:     }

4760:     /* pack the outgoing message */
4761:     PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);

4763:     sstartsj[0] = 0;
4764:     rstartsj[0] = 0;
4765:     len         = 0; /* total length of j or a array to be sent */
4766:     k           = 0;
4767:     for (i=0; i<nsends; i++) {
4768:       rowlen = (PetscInt*)svalues + sstarts[i]*sbs;
4769:       nrows  = sstarts[i+1]-sstarts[i]; /* num of block rows */
4770:       for (j=0; j<nrows; j++) {
4771:         row = srow[k] + B->rmap->range[rank]; /* global row idx */
4772:         for (l=0; l<sbs; l++) {
4773:           MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */

4775:           rowlen[j*sbs+l] = ncols;

4777:           len += ncols;
4778:           MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
4779:         }
4780:         k++;
4781:       }
4782:       MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);

4784:       sstartsj[i+1] = len;  /* starting point of (i+1)-th outgoing msg in bufj and bufa */
4785:     }
4786:     /* recvs and sends of i-array are completed */
4787:     i = nrecvs;
4788:     while (i--) {
4789:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4790:     }
4791:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}

4793:     /* allocate buffers for sending j and a arrays */
4794:     PetscMalloc1(len+1,&bufj);
4795:     PetscMalloc1(len+1,&bufa);

4797:     /* create i-array of B_oth */
4798:     PetscMalloc1(aBn+2,&b_othi);

4800:     b_othi[0] = 0;
4801:     len       = 0; /* total length of j or a array to be received */
4802:     k         = 0;
4803:     for (i=0; i<nrecvs; i++) {
4804:       rowlen = (PetscInt*)rvalues + rstarts[i]*rbs;
4805:       nrows  = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be received */
4806:       for (j=0; j<nrows; j++) {
4807:         b_othi[k+1] = b_othi[k] + rowlen[j];
4808:         PetscIntSumError(rowlen[j],len,&len);
4809:         k++;
4810:       }
4811:       rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
4812:     }

4814:     /* allocate space for j and a arrrays of B_oth */
4815:     PetscMalloc1(b_othi[aBn]+1,&b_othj);
4816:     PetscMalloc1(b_othi[aBn]+1,&b_otha);

4818:     /* j-array */
4819:     /*---------*/
4820:     /*  post receives of j-array */
4821:     for (i=0; i<nrecvs; i++) {
4822:       nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
4823:       MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
4824:     }

4826:     /* pack the outgoing message j-array */
4827:     k = 0;
4828:     for (i=0; i<nsends; i++) {
4829:       nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
4830:       bufJ  = bufj+sstartsj[i];
4831:       for (j=0; j<nrows; j++) {
4832:         row = srow[k++] + B->rmap->range[rank];  /* global row idx */
4833:         for (ll=0; ll<sbs; ll++) {
4834:           MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
4835:           for (l=0; l<ncols; l++) {
4836:             *bufJ++ = cols[l];
4837:           }
4838:           MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
4839:         }
4840:       }
4841:       MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
4842:     }

4844:     /* recvs and sends of j-array are completed */
4845:     i = nrecvs;
4846:     while (i--) {
4847:       MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4848:     }
4849:     if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
4850:   } else if (scall == MAT_REUSE_MATRIX) {
4851:     sstartsj = *startsj_s;
4852:     rstartsj = *startsj_r;
4853:     bufa     = *bufa_ptr;
4854:     b_oth    = (Mat_SeqAIJ*)(*B_oth)->data;
4855:     b_otha   = b_oth->a;
4856:   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");

4858:   /* a-array */
4859:   /*---------*/
4860:   /*  post receives of a-array */
4861:   for (i=0; i<nrecvs; i++) {
4862:     nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
4863:     MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
4864:   }

4866:   /* pack the outgoing message a-array */
4867:   k = 0;
4868:   for (i=0; i<nsends; i++) {
4869:     nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
4870:     bufA  = bufa+sstartsj[i];
4871:     for (j=0; j<nrows; j++) {
4872:       row = srow[k++] + B->rmap->range[rank];  /* global row idx */
4873:       for (ll=0; ll<sbs; ll++) {
4874:         MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
4875:         for (l=0; l<ncols; l++) {
4876:           *bufA++ = vals[l];
4877:         }
4878:         MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
4879:       }
4880:     }
4881:     MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
4882:   }
4883:   /* recvs and sends of a-array are completed */
4884:   i = nrecvs;
4885:   while (i--) {
4886:     MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);
4887:   }
4888:   if (nsends) {MPI_Waitall(nsends,swaits,sstatus);}
4889:   PetscFree2(rwaits,swaits);

4891:   if (scall == MAT_INITIAL_MATRIX) {
4892:     /* put together the new matrix */
4893:     MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);

4895:     /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
4896:     /* Since these are PETSc arrays, change flags to free them as necessary. */
4897:     b_oth          = (Mat_SeqAIJ*)(*B_oth)->data;
4898:     b_oth->free_a  = PETSC_TRUE;
4899:     b_oth->free_ij = PETSC_TRUE;
4900:     b_oth->nonew   = 0;

4902:     PetscFree(bufj);
4903:     if (!startsj_s || !bufa_ptr) {
4904:       PetscFree2(sstartsj,rstartsj);
4905:       PetscFree(bufa_ptr);
4906:     } else {
4907:       *startsj_s = sstartsj;
4908:       *startsj_r = rstartsj;
4909:       *bufa_ptr  = bufa;
4910:     }
4911:   }
4912:   PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
4913:   return(0);
4914: }

4918: /*@C
4919:   MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.

4921:   Not Collective

4923:   Input Parameters:
4924: . A - The matrix in mpiaij format

4926:   Output Parameter:
4927: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
4928: . colmap - A map from global column index to local index into lvec
4929: - multScatter - A scatter from the argument of a matrix-vector product to lvec

4931:   Level: developer

4933: @*/
4934: #if defined(PETSC_USE_CTABLE)
4935: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
4936: #else
4937: PetscErrorCode  MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
4938: #endif
4939: {
4940:   Mat_MPIAIJ *a;

4947:   a = (Mat_MPIAIJ*) A->data;
4948:   if (lvec) *lvec = a->lvec;
4949:   if (colmap) *colmap = a->colmap;
4950:   if (multScatter) *multScatter = a->Mvctx;
4951:   return(0);
4952: }

4954: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
4955: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
4956: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
4957: #if defined(PETSC_HAVE_ELEMENTAL)
4958: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
4959: #endif

4963: /*
4964:     Computes (B'*A')' since computing B*A directly is untenable

4966:                n                       p                          p
4967:         (              )       (              )         (                  )
4968:       m (      A       )  *  n (       B      )   =   m (         C        )
4969:         (              )       (              )         (                  )

4971: */
4972: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
4973: {
4975:   Mat            At,Bt,Ct;

4978:   MatTranspose(A,MAT_INITIAL_MATRIX,&At);
4979:   MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
4980:   MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
4981:   MatDestroy(&At);
4982:   MatDestroy(&Bt);
4983:   MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
4984:   MatDestroy(&Ct);
4985:   return(0);
4986: }

4990: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
4991: {
4993:   PetscInt       m=A->rmap->n,n=B->cmap->n;
4994:   Mat            Cmat;

4997:   if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n);
4998:   MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
4999:   MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5000:   MatSetBlockSizesFromMats(Cmat,A,B);
5001:   MatSetType(Cmat,MATMPIDENSE);
5002:   MatMPIDenseSetPreallocation(Cmat,NULL);
5003:   MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5004:   MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);

5006:   Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;

5008:   *C = Cmat;
5009:   return(0);
5010: }

5012: /* ----------------------------------------------------------------*/
5015: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
5016: {

5020:   if (scall == MAT_INITIAL_MATRIX) {
5021:     PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
5022:     MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
5023:     PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
5024:   }
5025:   PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
5026:   MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
5027:   PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
5028:   return(0);
5029: }

5031: /*MC
5032:    MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.

5034:    Options Database Keys:
5035: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()

5037:   Level: beginner

5039: .seealso: MatCreateAIJ()
5040: M*/

5044: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
5045: {
5046:   Mat_MPIAIJ     *b;
5048:   PetscMPIInt    size;

5051:   MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);

5053:   PetscNewLog(B,&b);
5054:   B->data       = (void*)b;
5055:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5056:   B->assembled  = PETSC_FALSE;
5057:   B->insertmode = NOT_SET_VALUES;
5058:   b->size       = size;

5060:   MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);

5062:   /* build cache for off array entries formed */
5063:   MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);

5065:   b->donotstash  = PETSC_FALSE;
5066:   b->colmap      = 0;
5067:   b->garray      = 0;
5068:   b->roworiented = PETSC_TRUE;

5070:   /* stuff used for matrix vector multiply */
5071:   b->lvec  = NULL;
5072:   b->Mvctx = NULL;

5074:   /* stuff for MatGetRow() */
5075:   b->rowindices   = 0;
5076:   b->rowvalues    = 0;
5077:   b->getrowactive = PETSC_FALSE;

5079:   /* flexible pointer used in CUSP/CUSPARSE classes */
5080:   b->spptr = NULL;

5082:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
5083:   PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
5084:   PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
5085:   PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);
5086:   PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
5087:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
5088:   PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
5089:   PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
5090:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
5091:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
5092:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
5093: #if defined(PETSC_HAVE_ELEMENTAL)
5094:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
5095: #endif
5096:   PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);
5097:   PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);
5098:   PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);
5099:   PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5100:   return(0);
5101: }

5105: /*@C
5106:      MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5107:          and "off-diagonal" part of the matrix in CSR format.

5109:    Collective on MPI_Comm

5111:    Input Parameters:
5112: +  comm - MPI communicator
5113: .  m - number of local rows (Cannot be PETSC_DECIDE)
5114: .  n - This value should be the same as the local size used in creating the
5115:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5116:        calculated if N is given) For square matrices n is almost always m.
5117: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5118: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5119: .   i - row indices for "diagonal" portion of matrix
5120: .   j - column indices
5121: .   a - matrix values
5122: .   oi - row indices for "off-diagonal" portion of matrix
5123: .   oj - column indices
5124: -   oa - matrix values

5126:    Output Parameter:
5127: .   mat - the matrix

5129:    Level: advanced

5131:    Notes:
5132:        The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5133:        must free the arrays once the matrix has been destroyed and not before.

5135:        The i and j indices are 0 based

5137:        See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix

5139:        This sets local rows and cannot be used to set off-processor values.

5141:        Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
5142:        legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
5143:        not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
5144:        the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
5145:        keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
5146:        communication if it is known that only local entries will be set.

5148: .keywords: matrix, aij, compressed row, sparse, parallel

5150: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5151:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5152: @*/
5153: PetscErrorCode  MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
5154: {
5156:   Mat_MPIAIJ     *maij;

5159:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5160:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5161:   if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5162:   MatCreate(comm,mat);
5163:   MatSetSizes(*mat,m,n,M,N);
5164:   MatSetType(*mat,MATMPIAIJ);
5165:   maij = (Mat_MPIAIJ*) (*mat)->data;

5167:   (*mat)->preallocated = PETSC_TRUE;

5169:   PetscLayoutSetUp((*mat)->rmap);
5170:   PetscLayoutSetUp((*mat)->cmap);

5172:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5173:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);

5175:   MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5176:   MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5177:   MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5178:   MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);

5180:   MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5181:   MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5182:   MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
5183:   return(0);
5184: }

5186: /*
5187:     Special version for direct calls from Fortran
5188: */
5189: #include <petsc/private/fortranimpl.h>

5191: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5192: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
5193: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5194: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
5195: #endif

5197: /* Change these macros so can be used in void function */
5198: #undef CHKERRQ
5199: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
5200: #undef SETERRQ2
5201: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5202: #undef SETERRQ3
5203: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5204: #undef SETERRQ
5205: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)

5209: PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
5210: {
5211:   Mat            mat  = *mmat;
5212:   PetscInt       m    = *mm, n = *mn;
5213:   InsertMode     addv = *maddv;
5214:   Mat_MPIAIJ     *aij = (Mat_MPIAIJ*)mat->data;
5215:   PetscScalar    value;

5218:   MatCheckPreallocated(mat,1);
5219:   if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;

5221: #if defined(PETSC_USE_DEBUG)
5222:   else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
5223: #endif
5224:   {
5225:     PetscInt  i,j,rstart  = mat->rmap->rstart,rend = mat->rmap->rend;
5226:     PetscInt  cstart      = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
5227:     PetscBool roworiented = aij->roworiented;

5229:     /* Some Variables required in the macro */
5230:     Mat        A                 = aij->A;
5231:     Mat_SeqAIJ *a                = (Mat_SeqAIJ*)A->data;
5232:     PetscInt   *aimax            = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
5233:     MatScalar  *aa               = a->a;
5234:     PetscBool  ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
5235:     Mat        B                 = aij->B;
5236:     Mat_SeqAIJ *b                = (Mat_SeqAIJ*)B->data;
5237:     PetscInt   *bimax            = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
5238:     MatScalar  *ba               = b->a;

5240:     PetscInt  *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
5241:     PetscInt  nonew = a->nonew;
5242:     MatScalar *ap1,*ap2;

5245:     for (i=0; i<m; i++) {
5246:       if (im[i] < 0) continue;
5247: #if defined(PETSC_USE_DEBUG)
5248:       if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
5249: #endif
5250:       if (im[i] >= rstart && im[i] < rend) {
5251:         row      = im[i] - rstart;
5252:         lastcol1 = -1;
5253:         rp1      = aj + ai[row];
5254:         ap1      = aa + ai[row];
5255:         rmax1    = aimax[row];
5256:         nrow1    = ailen[row];
5257:         low1     = 0;
5258:         high1    = nrow1;
5259:         lastcol2 = -1;
5260:         rp2      = bj + bi[row];
5261:         ap2      = ba + bi[row];
5262:         rmax2    = bimax[row];
5263:         nrow2    = bilen[row];
5264:         low2     = 0;
5265:         high2    = nrow2;

5267:         for (j=0; j<n; j++) {
5268:           if (roworiented) value = v[i*n+j];
5269:           else value = v[i+j*m];
5270:           if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue;
5271:           if (in[j] >= cstart && in[j] < cend) {
5272:             col = in[j] - cstart;
5273:             MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
5274:           } else if (in[j] < 0) continue;
5275: #if defined(PETSC_USE_DEBUG)
5276:           else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
5277: #endif
5278:           else {
5279:             if (mat->was_assembled) {
5280:               if (!aij->colmap) {
5281:                 MatCreateColmap_MPIAIJ_Private(mat);
5282:               }
5283: #if defined(PETSC_USE_CTABLE)
5284:               PetscTableFind(aij->colmap,in[j]+1,&col);
5285:               col--;
5286: #else
5287:               col = aij->colmap[in[j]] - 1;
5288: #endif
5289:               if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
5290:                 MatDisAssemble_MPIAIJ(mat);
5291:                 col  =  in[j];
5292:                 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
5293:                 B     = aij->B;
5294:                 b     = (Mat_SeqAIJ*)B->data;
5295:                 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
5296:                 rp2   = bj + bi[row];
5297:                 ap2   = ba + bi[row];
5298:                 rmax2 = bimax[row];
5299:                 nrow2 = bilen[row];
5300:                 low2  = 0;
5301:                 high2 = nrow2;
5302:                 bm    = aij->B->rmap->n;
5303:                 ba    = b->a;
5304:               }
5305:             } else col = in[j];
5306:             MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
5307:           }
5308:         }
5309:       } else if (!aij->donotstash) {
5310:         if (roworiented) {
5311:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5312:         } else {
5313:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5314:         }
5315:       }
5316:     }
5317:   }
5318:   PetscFunctionReturnVoid();
5319: }