Actual source code: mpibaij.c

petsc-master 2014-09-12
Report Typos and Errors
  2: #include <../src/mat/impls/baij/mpi/mpibaij.h>   /*I  "petscmat.h"  I*/
  3: #include <petscblaslapack.h>
  4: #include <petscsf.h>

  6: extern PetscErrorCode MatSetUpMultiply_MPIBAIJ(Mat);
  7: extern PetscErrorCode MatDisAssemble_MPIBAIJ(Mat);
  8: extern PetscErrorCode MatGetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],PetscScalar []);
  9: extern PetscErrorCode MatSetValues_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt [],const PetscScalar [],InsertMode);
 10: extern PetscErrorCode MatSetValuesBlocked_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[],const PetscScalar[],InsertMode);
 11: extern PetscErrorCode MatGetRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
 12: extern PetscErrorCode MatRestoreRow_SeqBAIJ(Mat,PetscInt,PetscInt*,PetscInt*[],PetscScalar*[]);
 13: extern PetscErrorCode MatZeroRows_SeqBAIJ(Mat,PetscInt,const PetscInt[],PetscScalar,Vec,Vec);

 17: PetscErrorCode MatGetRowMaxAbs_MPIBAIJ(Mat A,Vec v,PetscInt idx[])
 18: {
 19:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
 21:   PetscInt       i,*idxb = 0;
 22:   PetscScalar    *va,*vb;
 23:   Vec            vtmp;

 26:   MatGetRowMaxAbs(a->A,v,idx);
 27:   VecGetArray(v,&va);
 28:   if (idx) {
 29:     for (i=0; i<A->rmap->n; i++) {
 30:       if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
 31:     }
 32:   }

 34:   VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
 35:   if (idx) {PetscMalloc1(A->rmap->n,&idxb);}
 36:   MatGetRowMaxAbs(a->B,vtmp,idxb);
 37:   VecGetArray(vtmp,&vb);

 39:   for (i=0; i<A->rmap->n; i++) {
 40:     if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
 41:       va[i] = vb[i];
 42:       if (idx) idx[i] = A->cmap->bs*a->garray[idxb[i]/A->cmap->bs] + (idxb[i] % A->cmap->bs);
 43:     }
 44:   }

 46:   VecRestoreArray(v,&va);
 47:   VecRestoreArray(vtmp,&vb);
 48:   PetscFree(idxb);
 49:   VecDestroy(&vtmp);
 50:   return(0);
 51: }

 55: PetscErrorCode  MatStoreValues_MPIBAIJ(Mat mat)
 56: {
 57:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)mat->data;

 61:   MatStoreValues(aij->A);
 62:   MatStoreValues(aij->B);
 63:   return(0);
 64: }

 68: PetscErrorCode  MatRetrieveValues_MPIBAIJ(Mat mat)
 69: {
 70:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)mat->data;

 74:   MatRetrieveValues(aij->A);
 75:   MatRetrieveValues(aij->B);
 76:   return(0);
 77: }

 79: /*
 80:      Local utility routine that creates a mapping from the global column
 81:    number to the local number in the off-diagonal part of the local
 82:    storage of the matrix.  This is done in a non scalable way since the
 83:    length of colmap equals the global matrix length.
 84: */
 87: PetscErrorCode MatCreateColmap_MPIBAIJ_Private(Mat mat)
 88: {
 89:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
 90:   Mat_SeqBAIJ    *B    = (Mat_SeqBAIJ*)baij->B->data;
 92:   PetscInt       nbs = B->nbs,i,bs=mat->rmap->bs;

 95: #if defined(PETSC_USE_CTABLE)
 96:   PetscTableCreate(baij->nbs,baij->Nbs+1,&baij->colmap);
 97:   for (i=0; i<nbs; i++) {
 98:     PetscTableAdd(baij->colmap,baij->garray[i]+1,i*bs+1,INSERT_VALUES);
 99:   }
100: #else
101:   PetscMalloc1((baij->Nbs+1),&baij->colmap);
102:   PetscLogObjectMemory((PetscObject)mat,baij->Nbs*sizeof(PetscInt));
103:   PetscMemzero(baij->colmap,baij->Nbs*sizeof(PetscInt));
104:   for (i=0; i<nbs; i++) baij->colmap[baij->garray[i]] = i*bs+1;
105: #endif
106:   return(0);
107: }

109: #define  MatSetValues_SeqBAIJ_A_Private(row,col,value,addv) \
110:   { \
111:  \
112:     brow = row/bs;  \
113:     rp   = aj + ai[brow]; ap = aa + bs2*ai[brow]; \
114:     rmax = aimax[brow]; nrow = ailen[brow]; \
115:     bcol = col/bs; \
116:     ridx = row % bs; cidx = col % bs; \
117:     low  = 0; high = nrow; \
118:     while (high-low > 3) { \
119:       t = (low+high)/2; \
120:       if (rp[t] > bcol) high = t; \
121:       else              low  = t; \
122:     } \
123:     for (_i=low; _i<high; _i++) { \
124:       if (rp[_i] > bcol) break; \
125:       if (rp[_i] == bcol) { \
126:         bap = ap +  bs2*_i + bs*cidx + ridx; \
127:         if (addv == ADD_VALUES) *bap += value;  \
128:         else                    *bap  = value;  \
129:         goto a_noinsert; \
130:       } \
131:     } \
132:     if (a->nonew == 1) goto a_noinsert; \
133:     if (a->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
134:     MatSeqXAIJReallocateAIJ(A,a->mbs,bs2,nrow,brow,bcol,rmax,aa,ai,aj,rp,ap,aimax,a->nonew,MatScalar); \
135:     N = nrow++ - 1;  \
136:     /* shift up all the later entries in this row */ \
137:     for (ii=N; ii>=_i; ii--) { \
138:       rp[ii+1] = rp[ii]; \
139:       PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
140:     } \
141:     if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar)); }  \
142:     rp[_i]                      = bcol;  \
143:     ap[bs2*_i + bs*cidx + ridx] = value;  \
144: a_noinsert:; \
145:     ailen[brow] = nrow; \
146:   }

148: #define  MatSetValues_SeqBAIJ_B_Private(row,col,value,addv) \
149:   { \
150:     brow = row/bs;  \
151:     rp   = bj + bi[brow]; ap = ba + bs2*bi[brow]; \
152:     rmax = bimax[brow]; nrow = bilen[brow]; \
153:     bcol = col/bs; \
154:     ridx = row % bs; cidx = col % bs; \
155:     low  = 0; high = nrow; \
156:     while (high-low > 3) { \
157:       t = (low+high)/2; \
158:       if (rp[t] > bcol) high = t; \
159:       else              low  = t; \
160:     } \
161:     for (_i=low; _i<high; _i++) { \
162:       if (rp[_i] > bcol) break; \
163:       if (rp[_i] == bcol) { \
164:         bap = ap +  bs2*_i + bs*cidx + ridx; \
165:         if (addv == ADD_VALUES) *bap += value;  \
166:         else                    *bap  = value;  \
167:         goto b_noinsert; \
168:       } \
169:     } \
170:     if (b->nonew == 1) goto b_noinsert; \
171:     if (b->nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", row, col); \
172:     MatSeqXAIJReallocateAIJ(B,b->mbs,bs2,nrow,brow,bcol,rmax,ba,bi,bj,rp,ap,bimax,b->nonew,MatScalar); \
173:     N = nrow++ - 1;  \
174:     /* shift up all the later entries in this row */ \
175:     for (ii=N; ii>=_i; ii--) { \
176:       rp[ii+1] = rp[ii]; \
177:       PetscMemcpy(ap+bs2*(ii+1),ap+bs2*(ii),bs2*sizeof(MatScalar)); \
178:     } \
179:     if (N>=_i) { PetscMemzero(ap+bs2*_i,bs2*sizeof(MatScalar));}  \
180:     rp[_i]                      = bcol;  \
181:     ap[bs2*_i + bs*cidx + ridx] = value;  \
182: b_noinsert:; \
183:     bilen[brow] = nrow; \
184:   }

188: PetscErrorCode MatSetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
189: {
190:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
191:   MatScalar      value;
192:   PetscBool      roworiented = baij->roworiented;
194:   PetscInt       i,j,row,col;
195:   PetscInt       rstart_orig=mat->rmap->rstart;
196:   PetscInt       rend_orig  =mat->rmap->rend,cstart_orig=mat->cmap->rstart;
197:   PetscInt       cend_orig  =mat->cmap->rend,bs=mat->rmap->bs;

199:   /* Some Variables required in the macro */
200:   Mat         A     = baij->A;
201:   Mat_SeqBAIJ *a    = (Mat_SeqBAIJ*)(A)->data;
202:   PetscInt    *aimax=a->imax,*ai=a->i,*ailen=a->ilen,*aj=a->j;
203:   MatScalar   *aa   =a->a;

205:   Mat         B     = baij->B;
206:   Mat_SeqBAIJ *b    = (Mat_SeqBAIJ*)(B)->data;
207:   PetscInt    *bimax=b->imax,*bi=b->i,*bilen=b->ilen,*bj=b->j;
208:   MatScalar   *ba   =b->a;

210:   PetscInt  *rp,ii,nrow,_i,rmax,N,brow,bcol;
211:   PetscInt  low,high,t,ridx,cidx,bs2=a->bs2;
212:   MatScalar *ap,*bap;

215:   for (i=0; i<m; i++) {
216:     if (im[i] < 0) continue;
217: #if defined(PETSC_USE_DEBUG)
218:     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);
219: #endif
220:     if (im[i] >= rstart_orig && im[i] < rend_orig) {
221:       row = im[i] - rstart_orig;
222:       for (j=0; j<n; j++) {
223:         if (in[j] >= cstart_orig && in[j] < cend_orig) {
224:           col = in[j] - cstart_orig;
225:           if (roworiented) value = v[i*n+j];
226:           else             value = v[i+j*m];
227:           MatSetValues_SeqBAIJ_A_Private(row,col,value,addv);
228:           /* MatSetValues_SeqBAIJ(baij->A,1,&row,1,&col,&value,addv); */
229:         } else if (in[j] < 0) continue;
230: #if defined(PETSC_USE_DEBUG)
231:         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);
232: #endif
233:         else {
234:           if (mat->was_assembled) {
235:             if (!baij->colmap) {
236:               MatCreateColmap_MPIBAIJ_Private(mat);
237:             }
238: #if defined(PETSC_USE_CTABLE)
239:             PetscTableFind(baij->colmap,in[j]/bs + 1,&col);
240:             col  = col - 1;
241: #else
242:             col = baij->colmap[in[j]/bs] - 1;
243: #endif
244:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
245:               MatDisAssemble_MPIBAIJ(mat);
246:               col  =  in[j];
247:               /* Reinitialize the variables required by MatSetValues_SeqBAIJ_B_Private() */
248:               B    = baij->B;
249:               b    = (Mat_SeqBAIJ*)(B)->data;
250:               bimax=b->imax;bi=b->i;bilen=b->ilen;bj=b->j;
251:               ba   =b->a;
252:             } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", im[i], in[j]);
253:             else col += in[j]%bs;
254:           } else col = in[j];
255:           if (roworiented) value = v[i*n+j];
256:           else             value = v[i+j*m];
257:           MatSetValues_SeqBAIJ_B_Private(row,col,value,addv);
258:           /* MatSetValues_SeqBAIJ(baij->B,1,&row,1,&col,&value,addv); */
259:         }
260:       }
261:     } else {
262:       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]);
263:       if (!baij->donotstash) {
264:         mat->assembled = PETSC_FALSE;
265:         if (roworiented) {
266:           MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
267:         } else {
268:           MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
269:         }
270:       }
271:     }
272:   }
273:   return(0);
274: }

278: PetscErrorCode MatSetValuesBlocked_MPIBAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
279: {
280:   Mat_MPIBAIJ       *baij = (Mat_MPIBAIJ*)mat->data;
281:   const PetscScalar *value;
282:   MatScalar         *barray     = baij->barray;
283:   PetscBool         roworiented = baij->roworiented;
284:   PetscErrorCode    ierr;
285:   PetscInt          i,j,ii,jj,row,col,rstart=baij->rstartbs;
286:   PetscInt          rend=baij->rendbs,cstart=baij->cstartbs,stepval;
287:   PetscInt          cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;

290:   if (!barray) {
291:     PetscMalloc1(bs2,&barray);
292:     baij->barray = barray;
293:   }

295:   if (roworiented) stepval = (n-1)*bs;
296:   else stepval = (m-1)*bs;

298:   for (i=0; i<m; i++) {
299:     if (im[i] < 0) continue;
300: #if defined(PETSC_USE_DEBUG)
301:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
302: #endif
303:     if (im[i] >= rstart && im[i] < rend) {
304:       row = im[i] - rstart;
305:       for (j=0; j<n; j++) {
306:         /* If NumCol = 1 then a copy is not required */
307:         if ((roworiented) && (n == 1)) {
308:           barray = (MatScalar*)v + i*bs2;
309:         } else if ((!roworiented) && (m == 1)) {
310:           barray = (MatScalar*)v + j*bs2;
311:         } else { /* Here a copy is required */
312:           if (roworiented) {
313:             value = v + (i*(stepval+bs) + j)*bs;
314:           } else {
315:             value = v + (j*(stepval+bs) + i)*bs;
316:           }
317:           for (ii=0; ii<bs; ii++,value+=bs+stepval) {
318:             for (jj=0; jj<bs; jj++) barray[jj] = value[jj];
319:             barray += bs;
320:           }
321:           barray -= bs2;
322:         }

324:         if (in[j] >= cstart && in[j] < cend) {
325:           col  = in[j] - cstart;
326:           MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
327:         } else if (in[j] < 0) continue;
328: #if defined(PETSC_USE_DEBUG)
329:         else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
330: #endif
331:         else {
332:           if (mat->was_assembled) {
333:             if (!baij->colmap) {
334:               MatCreateColmap_MPIBAIJ_Private(mat);
335:             }

337: #if defined(PETSC_USE_DEBUG)
338: #if defined(PETSC_USE_CTABLE)
339:             { PetscInt data;
340:               PetscTableFind(baij->colmap,in[j]+1,&data);
341:               if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
342:             }
343: #else
344:             if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
345: #endif
346: #endif
347: #if defined(PETSC_USE_CTABLE)
348:             PetscTableFind(baij->colmap,in[j]+1,&col);
349:             col  = (col - 1)/bs;
350: #else
351:             col = (baij->colmap[in[j]] - 1)/bs;
352: #endif
353:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->B->data))->nonew) {
354:               MatDisAssemble_MPIBAIJ(mat);
355:               col  =  in[j];
356:             } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero (%D, %D) into matrix", bs*im[i], bs*in[j]);
357:           } else col = in[j];
358:           MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
359:         }
360:       }
361:     } else {
362:       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]);
363:       if (!baij->donotstash) {
364:         if (roworiented) {
365:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
366:         } else {
367:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
368:         }
369:       }
370:     }
371:   }
372:   return(0);
373: }

375: #define HASH_KEY 0.6180339887
376: #define HASH(size,key,tmp) (tmp = (key)*HASH_KEY,(PetscInt)((size)*(tmp-(PetscInt)tmp)))
377: /* #define HASH(size,key) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
378: /* #define HASH(size,key,tmp) ((PetscInt)((size)*fmod(((key)*HASH_KEY),1))) */
381: PetscErrorCode MatSetValues_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
382: {
383:   Mat_MPIBAIJ    *baij       = (Mat_MPIBAIJ*)mat->data;
384:   PetscBool      roworiented = baij->roworiented;
386:   PetscInt       i,j,row,col;
387:   PetscInt       rstart_orig=mat->rmap->rstart;
388:   PetscInt       rend_orig  =mat->rmap->rend,Nbs=baij->Nbs;
389:   PetscInt       h1,key,size=baij->ht_size,bs=mat->rmap->bs,*HT=baij->ht,idx;
390:   PetscReal      tmp;
391:   MatScalar      **HD = baij->hd,value;
392: #if defined(PETSC_USE_DEBUG)
393:   PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
394: #endif

397:   for (i=0; i<m; i++) {
398: #if defined(PETSC_USE_DEBUG)
399:     if (im[i] < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row");
400:     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);
401: #endif
402:     row = im[i];
403:     if (row >= rstart_orig && row < rend_orig) {
404:       for (j=0; j<n; j++) {
405:         col = in[j];
406:         if (roworiented) value = v[i*n+j];
407:         else             value = v[i+j*m];
408:         /* Look up PetscInto the Hash Table */
409:         key = (row/bs)*Nbs+(col/bs)+1;
410:         h1  = HASH(size,key,tmp);


413:         idx = h1;
414: #if defined(PETSC_USE_DEBUG)
415:         insert_ct++;
416:         total_ct++;
417:         if (HT[idx] != key) {
418:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
419:           if (idx == size) {
420:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
421:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
422:           }
423:         }
424: #else
425:         if (HT[idx] != key) {
426:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
427:           if (idx == size) {
428:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
429:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
430:           }
431:         }
432: #endif
433:         /* A HASH table entry is found, so insert the values at the correct address */
434:         if (addv == ADD_VALUES) *(HD[idx]+ (col % bs)*bs + (row % bs)) += value;
435:         else                    *(HD[idx]+ (col % bs)*bs + (row % bs))  = value;
436:       }
437:     } else if (!baij->donotstash) {
438:       if (roworiented) {
439:         MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,PETSC_FALSE);
440:       } else {
441:         MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,PETSC_FALSE);
442:       }
443:     }
444:   }
445: #if defined(PETSC_USE_DEBUG)
446:   baij->ht_total_ct  = total_ct;
447:   baij->ht_insert_ct = insert_ct;
448: #endif
449:   return(0);
450: }

454: PetscErrorCode MatSetValuesBlocked_MPIBAIJ_HT(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
455: {
456:   Mat_MPIBAIJ       *baij       = (Mat_MPIBAIJ*)mat->data;
457:   PetscBool         roworiented = baij->roworiented;
458:   PetscErrorCode    ierr;
459:   PetscInt          i,j,ii,jj,row,col;
460:   PetscInt          rstart=baij->rstartbs;
461:   PetscInt          rend  =mat->rmap->rend,stepval,bs=mat->rmap->bs,bs2=baij->bs2,nbs2=n*bs2;
462:   PetscInt          h1,key,size=baij->ht_size,idx,*HT=baij->ht,Nbs=baij->Nbs;
463:   PetscReal         tmp;
464:   MatScalar         **HD = baij->hd,*baij_a;
465:   const PetscScalar *v_t,*value;
466: #if defined(PETSC_USE_DEBUG)
467:   PetscInt total_ct=baij->ht_total_ct,insert_ct=baij->ht_insert_ct;
468: #endif

471:   if (roworiented) stepval = (n-1)*bs;
472:   else stepval = (m-1)*bs;

474:   for (i=0; i<m; i++) {
475: #if defined(PETSC_USE_DEBUG)
476:     if (im[i] < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",im[i]);
477:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],baij->Mbs-1);
478: #endif
479:     row = im[i];
480:     v_t = v + i*nbs2;
481:     if (row >= rstart && row < rend) {
482:       for (j=0; j<n; j++) {
483:         col = in[j];

485:         /* Look up into the Hash Table */
486:         key = row*Nbs+col+1;
487:         h1  = HASH(size,key,tmp);

489:         idx = h1;
490: #if defined(PETSC_USE_DEBUG)
491:         total_ct++;
492:         insert_ct++;
493:         if (HT[idx] != key) {
494:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++,total_ct++) ;
495:           if (idx == size) {
496:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++,total_ct++) ;
497:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
498:           }
499:         }
500: #else
501:         if (HT[idx] != key) {
502:           for (idx=h1; (idx<size) && (HT[idx]!=key); idx++) ;
503:           if (idx == size) {
504:             for (idx=0; (idx<h1) && (HT[idx]!=key); idx++) ;
505:             if (idx == h1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"(%D,%D) has no entry in the hash table", row, col);
506:           }
507:         }
508: #endif
509:         baij_a = HD[idx];
510:         if (roworiented) {
511:           /*value = v + i*(stepval+bs)*bs + j*bs;*/
512:           /* value = v + (i*(stepval+bs)+j)*bs; */
513:           value = v_t;
514:           v_t  += bs;
515:           if (addv == ADD_VALUES) {
516:             for (ii=0; ii<bs; ii++,value+=stepval) {
517:               for (jj=ii; jj<bs2; jj+=bs) {
518:                 baij_a[jj] += *value++;
519:               }
520:             }
521:           } else {
522:             for (ii=0; ii<bs; ii++,value+=stepval) {
523:               for (jj=ii; jj<bs2; jj+=bs) {
524:                 baij_a[jj] = *value++;
525:               }
526:             }
527:           }
528:         } else {
529:           value = v + j*(stepval+bs)*bs + i*bs;
530:           if (addv == ADD_VALUES) {
531:             for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
532:               for (jj=0; jj<bs; jj++) {
533:                 baij_a[jj] += *value++;
534:               }
535:             }
536:           } else {
537:             for (ii=0; ii<bs; ii++,value+=stepval,baij_a+=bs) {
538:               for (jj=0; jj<bs; jj++) {
539:                 baij_a[jj] = *value++;
540:               }
541:             }
542:           }
543:         }
544:       }
545:     } else {
546:       if (!baij->donotstash) {
547:         if (roworiented) {
548:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
549:         } else {
550:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
551:         }
552:       }
553:     }
554:   }
555: #if defined(PETSC_USE_DEBUG)
556:   baij->ht_total_ct  = total_ct;
557:   baij->ht_insert_ct = insert_ct;
558: #endif
559:   return(0);
560: }

564: PetscErrorCode MatGetValues_MPIBAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
565: {
566:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
568:   PetscInt       bs       = mat->rmap->bs,i,j,bsrstart = mat->rmap->rstart,bsrend = mat->rmap->rend;
569:   PetscInt       bscstart = mat->cmap->rstart,bscend = mat->cmap->rend,row,col,data;

572:   for (i=0; i<m; i++) {
573:     if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
574:     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);
575:     if (idxm[i] >= bsrstart && idxm[i] < bsrend) {
576:       row = idxm[i] - bsrstart;
577:       for (j=0; j<n; j++) {
578:         if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
579:         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);
580:         if (idxn[j] >= bscstart && idxn[j] < bscend) {
581:           col  = idxn[j] - bscstart;
582:           MatGetValues_SeqBAIJ(baij->A,1,&row,1,&col,v+i*n+j);
583:         } else {
584:           if (!baij->colmap) {
585:             MatCreateColmap_MPIBAIJ_Private(mat);
586:           }
587: #if defined(PETSC_USE_CTABLE)
588:           PetscTableFind(baij->colmap,idxn[j]/bs+1,&data);
589:           data--;
590: #else
591:           data = baij->colmap[idxn[j]/bs]-1;
592: #endif
593:           if ((data < 0) || (baij->garray[data/bs] != idxn[j]/bs)) *(v+i*n+j) = 0.0;
594:           else {
595:             col  = data + idxn[j]%bs;
596:             MatGetValues_SeqBAIJ(baij->B,1,&row,1,&col,v+i*n+j);
597:           }
598:         }
599:       }
600:     } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
601:   }
602:   return(0);
603: }

607: PetscErrorCode MatNorm_MPIBAIJ(Mat mat,NormType type,PetscReal *nrm)
608: {
609:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
610:   Mat_SeqBAIJ    *amat = (Mat_SeqBAIJ*)baij->A->data,*bmat = (Mat_SeqBAIJ*)baij->B->data;
612:   PetscInt       i,j,bs2=baij->bs2,bs=baij->A->rmap->bs,nz,row,col;
613:   PetscReal      sum = 0.0;
614:   MatScalar      *v;

617:   if (baij->size == 1) {
618:      MatNorm(baij->A,type,nrm);
619:   } else {
620:     if (type == NORM_FROBENIUS) {
621:       v  = amat->a;
622:       nz = amat->nz*bs2;
623:       for (i=0; i<nz; i++) {
624:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
625:       }
626:       v  = bmat->a;
627:       nz = bmat->nz*bs2;
628:       for (i=0; i<nz; i++) {
629:         sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
630:       }
631:       MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
632:       *nrm = PetscSqrtReal(*nrm);
633:     } else if (type == NORM_1) { /* max column sum */
634:       PetscReal *tmp,*tmp2;
635:       PetscInt  *jj,*garray=baij->garray,cstart=baij->rstartbs;
636:       PetscMalloc2(mat->cmap->N,&tmp,mat->cmap->N,&tmp2);
637:       PetscMemzero(tmp,mat->cmap->N*sizeof(PetscReal));
638:       v    = amat->a; jj = amat->j;
639:       for (i=0; i<amat->nz; i++) {
640:         for (j=0; j<bs; j++) {
641:           col = bs*(cstart + *jj) + j; /* column index */
642:           for (row=0; row<bs; row++) {
643:             tmp[col] += PetscAbsScalar(*v);  v++;
644:           }
645:         }
646:         jj++;
647:       }
648:       v = bmat->a; jj = bmat->j;
649:       for (i=0; i<bmat->nz; i++) {
650:         for (j=0; j<bs; j++) {
651:           col = bs*garray[*jj] + j;
652:           for (row=0; row<bs; row++) {
653:             tmp[col] += PetscAbsScalar(*v); v++;
654:           }
655:         }
656:         jj++;
657:       }
658:       MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
659:       *nrm = 0.0;
660:       for (j=0; j<mat->cmap->N; j++) {
661:         if (tmp2[j] > *nrm) *nrm = tmp2[j];
662:       }
663:       PetscFree2(tmp,tmp2);
664:     } else if (type == NORM_INFINITY) { /* max row sum */
665:       PetscReal *sums;
666:       PetscMalloc1(bs,&sums);
667:       sum  = 0.0;
668:       for (j=0; j<amat->mbs; j++) {
669:         for (row=0; row<bs; row++) sums[row] = 0.0;
670:         v  = amat->a + bs2*amat->i[j];
671:         nz = amat->i[j+1]-amat->i[j];
672:         for (i=0; i<nz; i++) {
673:           for (col=0; col<bs; col++) {
674:             for (row=0; row<bs; row++) {
675:               sums[row] += PetscAbsScalar(*v); v++;
676:             }
677:           }
678:         }
679:         v  = bmat->a + bs2*bmat->i[j];
680:         nz = bmat->i[j+1]-bmat->i[j];
681:         for (i=0; i<nz; i++) {
682:           for (col=0; col<bs; col++) {
683:             for (row=0; row<bs; row++) {
684:               sums[row] += PetscAbsScalar(*v); v++;
685:             }
686:           }
687:         }
688:         for (row=0; row<bs; row++) {
689:           if (sums[row] > sum) sum = sums[row];
690:         }
691:       }
692:       MPI_Allreduce(&sum,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
693:       PetscFree(sums);
694:     } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for this norm yet");
695:   }
696:   return(0);
697: }

699: /*
700:   Creates the hash table, and sets the table
701:   This table is created only once.
702:   If new entried need to be added to the matrix
703:   then the hash table has to be destroyed and
704:   recreated.
705: */
708: PetscErrorCode MatCreateHashTable_MPIBAIJ_Private(Mat mat,PetscReal factor)
709: {
710:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
711:   Mat            A     = baij->A,B=baij->B;
712:   Mat_SeqBAIJ    *a    = (Mat_SeqBAIJ*)A->data,*b=(Mat_SeqBAIJ*)B->data;
713:   PetscInt       i,j,k,nz=a->nz+b->nz,h1,*ai=a->i,*aj=a->j,*bi=b->i,*bj=b->j;
715:   PetscInt       ht_size,bs2=baij->bs2,rstart=baij->rstartbs;
716:   PetscInt       cstart=baij->cstartbs,*garray=baij->garray,row,col,Nbs=baij->Nbs;
717:   PetscInt       *HT,key;
718:   MatScalar      **HD;
719:   PetscReal      tmp;
720: #if defined(PETSC_USE_INFO)
721:   PetscInt ct=0,max=0;
722: #endif

725:   if (baij->ht) return(0);

727:   baij->ht_size = (PetscInt)(factor*nz);
728:   ht_size       = baij->ht_size;

730:   /* Allocate Memory for Hash Table */
731:   PetscCalloc2(ht_size,&baij->hd,ht_size,&baij->ht);
732:   HD   = baij->hd;
733:   HT   = baij->ht;

735:   /* Loop Over A */
736:   for (i=0; i<a->mbs; i++) {
737:     for (j=ai[i]; j<ai[i+1]; j++) {
738:       row = i+rstart;
739:       col = aj[j]+cstart;

741:       key = row*Nbs + col + 1;
742:       h1  = HASH(ht_size,key,tmp);
743:       for (k=0; k<ht_size; k++) {
744:         if (!HT[(h1+k)%ht_size]) {
745:           HT[(h1+k)%ht_size] = key;
746:           HD[(h1+k)%ht_size] = a->a + j*bs2;
747:           break;
748: #if defined(PETSC_USE_INFO)
749:         } else {
750:           ct++;
751: #endif
752:         }
753:       }
754: #if defined(PETSC_USE_INFO)
755:       if (k> max) max = k;
756: #endif
757:     }
758:   }
759:   /* Loop Over B */
760:   for (i=0; i<b->mbs; i++) {
761:     for (j=bi[i]; j<bi[i+1]; j++) {
762:       row = i+rstart;
763:       col = garray[bj[j]];
764:       key = row*Nbs + col + 1;
765:       h1  = HASH(ht_size,key,tmp);
766:       for (k=0; k<ht_size; k++) {
767:         if (!HT[(h1+k)%ht_size]) {
768:           HT[(h1+k)%ht_size] = key;
769:           HD[(h1+k)%ht_size] = b->a + j*bs2;
770:           break;
771: #if defined(PETSC_USE_INFO)
772:         } else {
773:           ct++;
774: #endif
775:         }
776:       }
777: #if defined(PETSC_USE_INFO)
778:       if (k> max) max = k;
779: #endif
780:     }
781:   }

783:   /* Print Summary */
784: #if defined(PETSC_USE_INFO)
785:   for (i=0,j=0; i<ht_size; i++) {
786:     if (HT[i]) j++;
787:   }
788:   PetscInfo2(mat,"Average Search = %5.2f,max search = %D\n",(!j)? 0.0:((PetscReal)(ct+j))/j,max);
789: #endif
790:   return(0);
791: }

795: PetscErrorCode MatAssemblyBegin_MPIBAIJ(Mat mat,MatAssemblyType mode)
796: {
797:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
799:   PetscInt       nstash,reallocs;
800:   InsertMode     addv;

803:   if (baij->donotstash || mat->nooffprocentries) return(0);

805:   /* make sure all processors are either in INSERTMODE or ADDMODE */
806:   MPI_Allreduce((PetscEnum*)&mat->insertmode,(PetscEnum*)&addv,1,MPIU_ENUM,MPI_BOR,PetscObjectComm((PetscObject)mat));
807:   if (addv == (ADD_VALUES|INSERT_VALUES)) SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_ARG_WRONGSTATE,"Some processors inserted others added");
808:   mat->insertmode = addv; /* in case this processor had no cache */

810:   MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
811:   MatStashScatterBegin_Private(mat,&mat->bstash,baij->rangebs);
812:   MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
813:   PetscInfo2(mat,"Stash has %D entries,uses %D mallocs.\n",nstash,reallocs);
814:   MatStashGetInfo_Private(&mat->bstash,&nstash,&reallocs);
815:   PetscInfo2(mat,"Block-Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
816:   return(0);
817: }

821: PetscErrorCode MatAssemblyEnd_MPIBAIJ(Mat mat,MatAssemblyType mode)
822: {
823:   Mat_MPIBAIJ    *baij=(Mat_MPIBAIJ*)mat->data;
824:   Mat_SeqBAIJ    *a   =(Mat_SeqBAIJ*)baij->A->data;
826:   PetscInt       i,j,rstart,ncols,flg,bs2=baij->bs2;
827:   PetscInt       *row,*col;
828:   PetscBool      r1,r2,r3,other_disassembled;
829:   MatScalar      *val;
830:   InsertMode     addv = mat->insertmode;
831:   PetscMPIInt    n;

834:   /* do not use 'b=(Mat_SeqBAIJ*)baij->B->data' as B can be reset in disassembly */
835:   if (!baij->donotstash && !mat->nooffprocentries) {
836:     while (1) {
837:       MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
838:       if (!flg) break;

840:       for (i=0; i<n;) {
841:         /* Now identify the consecutive vals belonging to the same row */
842:         for (j=i,rstart=row[j]; j<n; j++) {
843:           if (row[j] != rstart) break;
844:         }
845:         if (j < n) ncols = j-i;
846:         else       ncols = n-i;
847:         /* Now assemble all these values with a single function call */
848:         MatSetValues_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i,addv);
849:         i    = j;
850:       }
851:     }
852:     MatStashScatterEnd_Private(&mat->stash);
853:     /* Now process the block-stash. Since the values are stashed column-oriented,
854:        set the roworiented flag to column oriented, and after MatSetValues()
855:        restore the original flags */
856:     r1 = baij->roworiented;
857:     r2 = a->roworiented;
858:     r3 = ((Mat_SeqBAIJ*)baij->B->data)->roworiented;

860:     baij->roworiented = PETSC_FALSE;
861:     a->roworiented    = PETSC_FALSE;

863:     (((Mat_SeqBAIJ*)baij->B->data))->roworiented = PETSC_FALSE; /* b->roworiented */
864:     while (1) {
865:       MatStashScatterGetMesg_Private(&mat->bstash,&n,&row,&col,&val,&flg);
866:       if (!flg) break;

868:       for (i=0; i<n;) {
869:         /* Now identify the consecutive vals belonging to the same row */
870:         for (j=i,rstart=row[j]; j<n; j++) {
871:           if (row[j] != rstart) break;
872:         }
873:         if (j < n) ncols = j-i;
874:         else       ncols = n-i;
875:         MatSetValuesBlocked_MPIBAIJ(mat,1,row+i,ncols,col+i,val+i*bs2,addv);
876:         i    = j;
877:       }
878:     }
879:     MatStashScatterEnd_Private(&mat->bstash);

881:     baij->roworiented = r1;
882:     a->roworiented    = r2;

884:     ((Mat_SeqBAIJ*)baij->B->data)->roworiented = r3; /* b->roworiented */
885:   }

887:   MatAssemblyBegin(baij->A,mode);
888:   MatAssemblyEnd(baij->A,mode);

890:   /* determine if any processor has disassembled, if so we must
891:      also disassemble ourselfs, in order that we may reassemble. */
892:   /*
893:      if nonzero structure of submatrix B cannot change then we know that
894:      no processor disassembled thus we can skip this stuff
895:   */
896:   if (!((Mat_SeqBAIJ*)baij->B->data)->nonew) {
897:     MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
898:     if (mat->was_assembled && !other_disassembled) {
899:       MatDisAssemble_MPIBAIJ(mat);
900:     }
901:   }

903:   if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
904:     MatSetUpMultiply_MPIBAIJ(mat);
905:   }
906:   MatAssemblyBegin(baij->B,mode);
907:   MatAssemblyEnd(baij->B,mode);

909: #if defined(PETSC_USE_INFO)
910:   if (baij->ht && mode== MAT_FINAL_ASSEMBLY) {
911:     PetscInfo1(mat,"Average Hash Table Search in MatSetValues = %5.2f\n",((PetscReal)baij->ht_total_ct)/baij->ht_insert_ct);

913:     baij->ht_total_ct  = 0;
914:     baij->ht_insert_ct = 0;
915:   }
916: #endif
917:   if (baij->ht_flag && !baij->ht && mode == MAT_FINAL_ASSEMBLY) {
918:     MatCreateHashTable_MPIBAIJ_Private(mat,baij->ht_fact);

920:     mat->ops->setvalues        = MatSetValues_MPIBAIJ_HT;
921:     mat->ops->setvaluesblocked = MatSetValuesBlocked_MPIBAIJ_HT;
922:   }

924:   PetscFree2(baij->rowvalues,baij->rowindices);

926:   baij->rowvalues = 0;

928:   /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
929:   if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
930:     PetscObjectState state = baij->A->nonzerostate + baij->B->nonzerostate;
931:     MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
932:   }
933:   return(0);
934: }

936: extern PetscErrorCode MatView_SeqBAIJ(Mat,PetscViewer);
937: #include <petscdraw.h>
940: static PetscErrorCode MatView_MPIBAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
941: {
942:   Mat_MPIBAIJ       *baij = (Mat_MPIBAIJ*)mat->data;
943:   PetscErrorCode    ierr;
944:   PetscMPIInt       rank = baij->rank;
945:   PetscInt          bs   = mat->rmap->bs;
946:   PetscBool         iascii,isdraw;
947:   PetscViewer       sviewer;
948:   PetscViewerFormat format;

951:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
952:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
953:   if (iascii) {
954:     PetscViewerGetFormat(viewer,&format);
955:     if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
956:       MatInfo info;
957:       MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
958:       MatGetInfo(mat,MAT_LOCAL,&info);
959:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);
960:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D bs %D mem %D\n",
961:                                                 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,mat->rmap->bs,(PetscInt)info.memory);
962:       MatGetInfo(baij->A,MAT_LOCAL,&info);
963:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
964:       MatGetInfo(baij->B,MAT_LOCAL,&info);
965:       PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
966:       PetscViewerFlush(viewer);
967:       PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);
968:       PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
969:       VecScatterView(baij->Mvctx,viewer);
970:       return(0);
971:     } else if (format == PETSC_VIEWER_ASCII_INFO) {
972:       PetscViewerASCIIPrintf(viewer,"  block size is %D\n",bs);
973:       return(0);
974:     } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
975:       return(0);
976:     }
977:   }

979:   if (isdraw) {
980:     PetscDraw draw;
981:     PetscBool isnull;
982:     PetscViewerDrawGetDraw(viewer,0,&draw);
983:     PetscDrawIsNull(draw,&isnull); if (isnull) return(0);
984:   }

986:   {
987:     /* assemble the entire matrix onto first processor. */
988:     Mat         A;
989:     Mat_SeqBAIJ *Aloc;
990:     PetscInt    M = mat->rmap->N,N = mat->cmap->N,*ai,*aj,col,i,j,k,*rvals,mbs = baij->mbs;
991:     MatScalar   *a;

993:     /* Here we are creating a temporary matrix, so will assume MPIBAIJ is acceptable */
994:     /* Perhaps this should be the type of mat? */
995:     MatCreate(PetscObjectComm((PetscObject)mat),&A);
996:     if (!rank) {
997:       MatSetSizes(A,M,N,M,N);
998:     } else {
999:       MatSetSizes(A,0,0,M,N);
1000:     }
1001:     MatSetType(A,MATMPIBAIJ);
1002:     MatMPIBAIJSetPreallocation(A,mat->rmap->bs,0,NULL,0,NULL);
1003:     MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);
1004:     PetscLogObjectParent((PetscObject)mat,(PetscObject)A);

1006:     /* copy over the A part */
1007:     Aloc = (Mat_SeqBAIJ*)baij->A->data;
1008:     ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1009:     PetscMalloc1(bs,&rvals);

1011:     for (i=0; i<mbs; i++) {
1012:       rvals[0] = bs*(baij->rstartbs + i);
1013:       for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1014:       for (j=ai[i]; j<ai[i+1]; j++) {
1015:         col = (baij->cstartbs+aj[j])*bs;
1016:         for (k=0; k<bs; k++) {
1017:           MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1018:           col++; a += bs;
1019:         }
1020:       }
1021:     }
1022:     /* copy over the B part */
1023:     Aloc = (Mat_SeqBAIJ*)baij->B->data;
1024:     ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1025:     for (i=0; i<mbs; i++) {
1026:       rvals[0] = bs*(baij->rstartbs + i);
1027:       for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1028:       for (j=ai[i]; j<ai[i+1]; j++) {
1029:         col = baij->garray[aj[j]]*bs;
1030:         for (k=0; k<bs; k++) {
1031:           MatSetValues_MPIBAIJ(A,bs,rvals,1,&col,a,INSERT_VALUES);
1032:           col++; a += bs;
1033:         }
1034:       }
1035:     }
1036:     PetscFree(rvals);
1037:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1038:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1039:     /*
1040:        Everyone has to call to draw the matrix since the graphics waits are
1041:        synchronized across all processors that share the PetscDraw object
1042:     */
1043:     PetscViewerGetSingleton(viewer,&sviewer);
1044:     if (!rank) {
1045:       MatView_SeqBAIJ(((Mat_MPIBAIJ*)(A->data))->A,sviewer);
1046:     }
1047:     PetscViewerRestoreSingleton(viewer,&sviewer);
1048:     MatDestroy(&A);
1049:   }
1050:   return(0);
1051: }

1055: static PetscErrorCode MatView_MPIBAIJ_Binary(Mat mat,PetscViewer viewer)
1056: {
1057:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)mat->data;
1058:   Mat_SeqBAIJ    *A = (Mat_SeqBAIJ*)a->A->data;
1059:   Mat_SeqBAIJ    *B = (Mat_SeqBAIJ*)a->B->data;
1061:   PetscInt       i,*row_lens,*crow_lens,bs = mat->rmap->bs,j,k,bs2=a->bs2,header[4],nz,rlen;
1062:   PetscInt       *range=0,nzmax,*column_indices,cnt,col,*garray = a->garray,cstart = mat->cmap->rstart/bs,len,pcnt,l,ll;
1063:   int            fd;
1064:   PetscScalar    *column_values;
1065:   FILE           *file;
1066:   PetscMPIInt    rank,size,tag = ((PetscObject)viewer)->tag;
1067:   PetscInt       message_count,flowcontrolcount;

1070:   MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1071:   MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1072:   nz   = bs2*(A->nz + B->nz);
1073:   rlen = mat->rmap->n;
1074:   if (!rank) {
1075:     header[0] = MAT_FILE_CLASSID;
1076:     header[1] = mat->rmap->N;
1077:     header[2] = mat->cmap->N;

1079:     MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1080:     PetscViewerBinaryGetDescriptor(viewer,&fd);
1081:     PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1082:     /* get largest number of rows any processor has */
1083:     range = mat->rmap->range;
1084:     for (i=1; i<size; i++) {
1085:       rlen = PetscMax(rlen,range[i+1] - range[i]);
1086:     }
1087:   } else {
1088:     MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1089:   }

1091:   PetscMalloc1((rlen/bs),&crow_lens);
1092:   /* compute lengths of each row  */
1093:   for (i=0; i<a->mbs; i++) {
1094:     crow_lens[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1095:   }
1096:   /* store the row lengths to the file */
1097:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1098:   if (!rank) {
1099:     MPI_Status status;
1100:     PetscMalloc1(rlen,&row_lens);
1101:     rlen = (range[1] - range[0])/bs;
1102:     for (i=0; i<rlen; i++) {
1103:       for (j=0; j<bs; j++) {
1104:         row_lens[i*bs+j] = bs*crow_lens[i];
1105:       }
1106:     }
1107:     PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1108:     for (i=1; i<size; i++) {
1109:       rlen = (range[i+1] - range[i])/bs;
1110:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1111:       MPI_Recv(crow_lens,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1112:       for (k=0; k<rlen; k++) {
1113:         for (j=0; j<bs; j++) {
1114:           row_lens[k*bs+j] = bs*crow_lens[k];
1115:         }
1116:       }
1117:       PetscBinaryWrite(fd,row_lens,bs*rlen,PETSC_INT,PETSC_TRUE);
1118:     }
1119:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1120:     PetscFree(row_lens);
1121:   } else {
1122:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1123:     MPI_Send(crow_lens,mat->rmap->n/bs,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1124:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1125:   }
1126:   PetscFree(crow_lens);

1128:   /* load up the local column indices. Include for all rows not just one for each block row since process 0 does not have the
1129:      information needed to make it for each row from a block row. This does require more communication but still not more than
1130:      the communication needed for the nonzero values  */
1131:   nzmax = nz; /*  space a largest processor needs */
1132:   MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1133:   PetscMalloc1(nzmax,&column_indices);
1134:   cnt   = 0;
1135:   for (i=0; i<a->mbs; i++) {
1136:     pcnt = cnt;
1137:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1138:       if ((col = garray[B->j[j]]) > cstart) break;
1139:       for (l=0; l<bs; l++) {
1140:         column_indices[cnt++] = bs*col+l;
1141:       }
1142:     }
1143:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1144:       for (l=0; l<bs; l++) {
1145:         column_indices[cnt++] = bs*(A->j[k] + cstart)+l;
1146:       }
1147:     }
1148:     for (; j<B->i[i+1]; j++) {
1149:       for (l=0; l<bs; l++) {
1150:         column_indices[cnt++] = bs*garray[B->j[j]]+l;
1151:       }
1152:     }
1153:     len = cnt - pcnt;
1154:     for (k=1; k<bs; k++) {
1155:       PetscMemcpy(&column_indices[cnt],&column_indices[pcnt],len*sizeof(PetscInt));
1156:       cnt += len;
1157:     }
1158:   }
1159:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);

1161:   /* store the columns to the file */
1162:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1163:   if (!rank) {
1164:     MPI_Status status;
1165:     PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1166:     for (i=1; i<size; i++) {
1167:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1168:       MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1169:       MPI_Recv(column_indices,cnt,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1170:       PetscBinaryWrite(fd,column_indices,cnt,PETSC_INT,PETSC_TRUE);
1171:     }
1172:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1173:   } else {
1174:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1175:     MPI_Send(&cnt,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1176:     MPI_Send(column_indices,cnt,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1177:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1178:   }
1179:   PetscFree(column_indices);

1181:   /* load up the numerical values */
1182:   PetscMalloc1(nzmax,&column_values);
1183:   cnt  = 0;
1184:   for (i=0; i<a->mbs; i++) {
1185:     rlen = bs*(B->i[i+1] - B->i[i] + A->i[i+1] - A->i[i]);
1186:     for (j=B->i[i]; j<B->i[i+1]; j++) {
1187:       if (garray[B->j[j]] > cstart) break;
1188:       for (l=0; l<bs; l++) {
1189:         for (ll=0; ll<bs; ll++) {
1190:           column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1191:         }
1192:       }
1193:       cnt += bs;
1194:     }
1195:     for (k=A->i[i]; k<A->i[i+1]; k++) {
1196:       for (l=0; l<bs; l++) {
1197:         for (ll=0; ll<bs; ll++) {
1198:           column_values[cnt + l*rlen + ll] = A->a[bs2*k+l+bs*ll];
1199:         }
1200:       }
1201:       cnt += bs;
1202:     }
1203:     for (; j<B->i[i+1]; j++) {
1204:       for (l=0; l<bs; l++) {
1205:         for (ll=0; ll<bs; ll++) {
1206:           column_values[cnt + l*rlen + ll] = B->a[bs2*j+l+bs*ll];
1207:         }
1208:       }
1209:       cnt += bs;
1210:     }
1211:     cnt += (bs-1)*rlen;
1212:   }
1213:   if (cnt != nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,nz);

1215:   /* store the column values to the file */
1216:   PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1217:   if (!rank) {
1218:     MPI_Status status;
1219:     PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1220:     for (i=1; i<size; i++) {
1221:       PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1222:       MPI_Recv(&cnt,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1223:       MPI_Recv(column_values,cnt,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat),&status);
1224:       PetscBinaryWrite(fd,column_values,cnt,PETSC_SCALAR,PETSC_TRUE);
1225:     }
1226:     PetscViewerFlowControlEndMaster(viewer,&message_count);
1227:   } else {
1228:     PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1229:     MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1230:     MPI_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1231:     PetscViewerFlowControlEndWorker(viewer,&message_count);
1232:   }
1233:   PetscFree(column_values);

1235:   PetscViewerBinaryGetInfoPointer(viewer,&file);
1236:   if (file) {
1237:     fprintf(file,"-matload_block_size %d\n",(int)mat->rmap->bs);
1238:   }
1239:   return(0);
1240: }

1244: PetscErrorCode MatView_MPIBAIJ(Mat mat,PetscViewer viewer)
1245: {
1247:   PetscBool      iascii,isdraw,issocket,isbinary;

1250:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1251:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1252:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1253:   PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1254:   if (iascii || isdraw || issocket) {
1255:     MatView_MPIBAIJ_ASCIIorDraworSocket(mat,viewer);
1256:   } else if (isbinary) {
1257:     MatView_MPIBAIJ_Binary(mat,viewer);
1258:   }
1259:   return(0);
1260: }

1264: PetscErrorCode MatDestroy_MPIBAIJ(Mat mat)
1265: {
1266:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;

1270: #if defined(PETSC_USE_LOG)
1271:   PetscLogObjectState((PetscObject)mat,"Rows=%D,Cols=%D",mat->rmap->N,mat->cmap->N);
1272: #endif
1273:   MatStashDestroy_Private(&mat->stash);
1274:   MatStashDestroy_Private(&mat->bstash);
1275:   MatDestroy(&baij->A);
1276:   MatDestroy(&baij->B);
1277: #if defined(PETSC_USE_CTABLE)
1278:   PetscTableDestroy(&baij->colmap);
1279: #else
1280:   PetscFree(baij->colmap);
1281: #endif
1282:   PetscFree(baij->garray);
1283:   VecDestroy(&baij->lvec);
1284:   VecScatterDestroy(&baij->Mvctx);
1285:   PetscFree2(baij->rowvalues,baij->rowindices);
1286:   PetscFree(baij->barray);
1287:   PetscFree2(baij->hd,baij->ht);
1288:   PetscFree(baij->rangebs);
1289:   PetscFree(mat->data);

1291:   PetscObjectChangeTypeName((PetscObject)mat,0);
1292:   PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1293:   PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1294:   PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);
1295:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocation_C",NULL);
1296:   PetscObjectComposeFunction((PetscObject)mat,"MatMPIBAIJSetPreallocationCSR_C",NULL);
1297:   PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1298:   PetscObjectComposeFunction((PetscObject)mat,"MatSetHashTableFactor_C",NULL);
1299:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpisbaij_C",NULL);
1300:   PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpibaij_mpibstrm_C",NULL);
1301:   return(0);
1302: }

1306: PetscErrorCode MatMult_MPIBAIJ(Mat A,Vec xx,Vec yy)
1307: {
1308:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1310:   PetscInt       nt;

1313:   VecGetLocalSize(xx,&nt);
1314:   if (nt != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A and xx");
1315:   VecGetLocalSize(yy,&nt);
1316:   if (nt != A->rmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible parition of A and yy");
1317:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1318:   (*a->A->ops->mult)(a->A,xx,yy);
1319:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1320:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1321:   return(0);
1322: }

1326: PetscErrorCode MatMultAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1327: {
1328:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1332:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1333:   (*a->A->ops->multadd)(a->A,xx,yy,zz);
1334:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1335:   (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1336:   return(0);
1337: }

1341: PetscErrorCode MatMultTranspose_MPIBAIJ(Mat A,Vec xx,Vec yy)
1342: {
1343:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1345:   PetscBool      merged;

1348:   VecScatterGetMerged(a->Mvctx,&merged);
1349:   /* do nondiagonal part */
1350:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1351:   if (!merged) {
1352:     /* send it on its way */
1353:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1354:     /* do local part */
1355:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1356:     /* receive remote parts: note this assumes the values are not actually */
1357:     /* inserted in yy until the next line */
1358:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1359:   } else {
1360:     /* do local part */
1361:     (*a->A->ops->multtranspose)(a->A,xx,yy);
1362:     /* send it on its way */
1363:     VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1364:     /* values actually were received in the Begin() but we need to call this nop */
1365:     VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1366:   }
1367:   return(0);
1368: }

1372: PetscErrorCode MatMultTransposeAdd_MPIBAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1373: {
1374:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1378:   /* do nondiagonal part */
1379:   (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1380:   /* send it on its way */
1381:   VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1382:   /* do local part */
1383:   (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1384:   /* receive remote parts: note this assumes the values are not actually */
1385:   /* inserted in yy until the next line, which is true for my implementation*/
1386:   /* but is not perhaps always true. */
1387:   VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1388:   return(0);
1389: }

1391: /*
1392:   This only works correctly for square matrices where the subblock A->A is the
1393:    diagonal block
1394: */
1397: PetscErrorCode MatGetDiagonal_MPIBAIJ(Mat A,Vec v)
1398: {
1399:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1403:   if (A->rmap->N != A->cmap->N) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1404:   MatGetDiagonal(a->A,v);
1405:   return(0);
1406: }

1410: PetscErrorCode MatScale_MPIBAIJ(Mat A,PetscScalar aa)
1411: {
1412:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1416:   MatScale(a->A,aa);
1417:   MatScale(a->B,aa);
1418:   return(0);
1419: }

1423: PetscErrorCode MatGetRow_MPIBAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1424: {
1425:   Mat_MPIBAIJ    *mat = (Mat_MPIBAIJ*)matin->data;
1426:   PetscScalar    *vworkA,*vworkB,**pvA,**pvB,*v_p;
1428:   PetscInt       bs = matin->rmap->bs,bs2 = mat->bs2,i,*cworkA,*cworkB,**pcA,**pcB;
1429:   PetscInt       nztot,nzA,nzB,lrow,brstart = matin->rmap->rstart,brend = matin->rmap->rend;
1430:   PetscInt       *cmap,*idx_p,cstart = mat->cstartbs;

1433:   if (row < brstart || row >= brend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local rows");
1434:   if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1435:   mat->getrowactive = PETSC_TRUE;

1437:   if (!mat->rowvalues && (idx || v)) {
1438:     /*
1439:         allocate enough space to hold information from the longest row.
1440:     */
1441:     Mat_SeqBAIJ *Aa = (Mat_SeqBAIJ*)mat->A->data,*Ba = (Mat_SeqBAIJ*)mat->B->data;
1442:     PetscInt    max = 1,mbs = mat->mbs,tmp;
1443:     for (i=0; i<mbs; i++) {
1444:       tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1445:       if (max < tmp) max = tmp;
1446:     }
1447:     PetscMalloc2(max*bs2,&mat->rowvalues,max*bs2,&mat->rowindices);
1448:   }
1449:   lrow = row - brstart;

1451:   pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1452:   if (!v)   {pvA = 0; pvB = 0;}
1453:   if (!idx) {pcA = 0; if (!v) pcB = 0;}
1454:   (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1455:   (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1456:   nztot = nzA + nzB;

1458:   cmap = mat->garray;
1459:   if (v  || idx) {
1460:     if (nztot) {
1461:       /* Sort by increasing column numbers, assuming A and B already sorted */
1462:       PetscInt imark = -1;
1463:       if (v) {
1464:         *v = v_p = mat->rowvalues;
1465:         for (i=0; i<nzB; i++) {
1466:           if (cmap[cworkB[i]/bs] < cstart) v_p[i] = vworkB[i];
1467:           else break;
1468:         }
1469:         imark = i;
1470:         for (i=0; i<nzA; i++)     v_p[imark+i] = vworkA[i];
1471:         for (i=imark; i<nzB; i++) v_p[nzA+i]   = vworkB[i];
1472:       }
1473:       if (idx) {
1474:         *idx = idx_p = mat->rowindices;
1475:         if (imark > -1) {
1476:           for (i=0; i<imark; i++) {
1477:             idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1478:           }
1479:         } else {
1480:           for (i=0; i<nzB; i++) {
1481:             if (cmap[cworkB[i]/bs] < cstart) idx_p[i] = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs;
1482:             else break;
1483:           }
1484:           imark = i;
1485:         }
1486:         for (i=0; i<nzA; i++)     idx_p[imark+i] = cstart*bs + cworkA[i];
1487:         for (i=imark; i<nzB; i++) idx_p[nzA+i]   = cmap[cworkB[i]/bs]*bs + cworkB[i]%bs ;
1488:       }
1489:     } else {
1490:       if (idx) *idx = 0;
1491:       if (v)   *v   = 0;
1492:     }
1493:   }
1494:   *nz  = nztot;
1495:   (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1496:   (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1497:   return(0);
1498: }

1502: PetscErrorCode MatRestoreRow_MPIBAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1503: {
1504:   Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*)mat->data;

1507:   if (!baij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow not called");
1508:   baij->getrowactive = PETSC_FALSE;
1509:   return(0);
1510: }

1514: PetscErrorCode MatZeroEntries_MPIBAIJ(Mat A)
1515: {
1516:   Mat_MPIBAIJ    *l = (Mat_MPIBAIJ*)A->data;

1520:   MatZeroEntries(l->A);
1521:   MatZeroEntries(l->B);
1522:   return(0);
1523: }

1527: PetscErrorCode MatGetInfo_MPIBAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1528: {
1529:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)matin->data;
1530:   Mat            A  = a->A,B = a->B;
1532:   PetscReal      isend[5],irecv[5];

1535:   info->block_size = (PetscReal)matin->rmap->bs;

1537:   MatGetInfo(A,MAT_LOCAL,info);

1539:   isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1540:   isend[3] = info->memory;  isend[4] = info->mallocs;

1542:   MatGetInfo(B,MAT_LOCAL,info);

1544:   isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1545:   isend[3] += info->memory;  isend[4] += info->mallocs;

1547:   if (flag == MAT_LOCAL) {
1548:     info->nz_used      = isend[0];
1549:     info->nz_allocated = isend[1];
1550:     info->nz_unneeded  = isend[2];
1551:     info->memory       = isend[3];
1552:     info->mallocs      = isend[4];
1553:   } else if (flag == MAT_GLOBAL_MAX) {
1554:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));

1556:     info->nz_used      = irecv[0];
1557:     info->nz_allocated = irecv[1];
1558:     info->nz_unneeded  = irecv[2];
1559:     info->memory       = irecv[3];
1560:     info->mallocs      = irecv[4];
1561:   } else if (flag == MAT_GLOBAL_SUM) {
1562:     MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));

1564:     info->nz_used      = irecv[0];
1565:     info->nz_allocated = irecv[1];
1566:     info->nz_unneeded  = irecv[2];
1567:     info->memory       = irecv[3];
1568:     info->mallocs      = irecv[4];
1569:   } else SETERRQ1(PetscObjectComm((PetscObject)matin),PETSC_ERR_ARG_WRONG,"Unknown MatInfoType argument %d",(int)flag);
1570:   info->fill_ratio_given  = 0; /* no parallel LU/ILU/Cholesky */
1571:   info->fill_ratio_needed = 0;
1572:   info->factor_mallocs    = 0;
1573:   return(0);
1574: }

1578: PetscErrorCode MatSetOption_MPIBAIJ(Mat A,MatOption op,PetscBool flg)
1579: {
1580:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1584:   switch (op) {
1585:   case MAT_NEW_NONZERO_LOCATIONS:
1586:   case MAT_NEW_NONZERO_ALLOCATION_ERR:
1587:   case MAT_UNUSED_NONZERO_LOCATION_ERR:
1588:   case MAT_KEEP_NONZERO_PATTERN:
1589:   case MAT_NEW_NONZERO_LOCATION_ERR:
1590:     MatSetOption(a->A,op,flg);
1591:     MatSetOption(a->B,op,flg);
1592:     break;
1593:   case MAT_ROW_ORIENTED:
1594:     a->roworiented = flg;

1596:     MatSetOption(a->A,op,flg);
1597:     MatSetOption(a->B,op,flg);
1598:     break;
1599:   case MAT_NEW_DIAGONALS:
1600:     PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1601:     break;
1602:   case MAT_IGNORE_OFF_PROC_ENTRIES:
1603:     a->donotstash = flg;
1604:     break;
1605:   case MAT_USE_HASH_TABLE:
1606:     a->ht_flag = flg;
1607:     break;
1608:   case MAT_SYMMETRIC:
1609:   case MAT_STRUCTURALLY_SYMMETRIC:
1610:   case MAT_HERMITIAN:
1611:   case MAT_SYMMETRY_ETERNAL:
1612:     MatSetOption(a->A,op,flg);
1613:     break;
1614:   default:
1615:     SETERRQ1(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"unknown option %d",op);
1616:   }
1617:   return(0);
1618: }

1622: PetscErrorCode MatTranspose_MPIBAIJ(Mat A,MatReuse reuse,Mat *matout)
1623: {
1624:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)A->data;
1625:   Mat_SeqBAIJ    *Aloc;
1626:   Mat            B;
1628:   PetscInt       M =A->rmap->N,N=A->cmap->N,*ai,*aj,i,*rvals,j,k,col;
1629:   PetscInt       bs=A->rmap->bs,mbs=baij->mbs;
1630:   MatScalar      *a;

1633:   if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place");
1634:   if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
1635:     MatCreate(PetscObjectComm((PetscObject)A),&B);
1636:     MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
1637:     MatSetType(B,((PetscObject)A)->type_name);
1638:     /* Do not know preallocation information, but must set block size */
1639:     MatMPIBAIJSetPreallocation(B,A->rmap->bs,PETSC_DECIDE,NULL,PETSC_DECIDE,NULL);
1640:   } else {
1641:     B = *matout;
1642:   }

1644:   /* copy over the A part */
1645:   Aloc = (Mat_SeqBAIJ*)baij->A->data;
1646:   ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1647:   PetscMalloc1(bs,&rvals);

1649:   for (i=0; i<mbs; i++) {
1650:     rvals[0] = bs*(baij->rstartbs + i);
1651:     for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1652:     for (j=ai[i]; j<ai[i+1]; j++) {
1653:       col = (baij->cstartbs+aj[j])*bs;
1654:       for (k=0; k<bs; k++) {
1655:         MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);

1657:         col++; a += bs;
1658:       }
1659:     }
1660:   }
1661:   /* copy over the B part */
1662:   Aloc = (Mat_SeqBAIJ*)baij->B->data;
1663:   ai   = Aloc->i; aj = Aloc->j; a = Aloc->a;
1664:   for (i=0; i<mbs; i++) {
1665:     rvals[0] = bs*(baij->rstartbs + i);
1666:     for (j=1; j<bs; j++) rvals[j] = rvals[j-1] + 1;
1667:     for (j=ai[i]; j<ai[i+1]; j++) {
1668:       col = baij->garray[aj[j]]*bs;
1669:       for (k=0; k<bs; k++) {
1670:         MatSetValues_MPIBAIJ(B,1,&col,bs,rvals,a,INSERT_VALUES);
1671:         col++;
1672:         a += bs;
1673:       }
1674:     }
1675:   }
1676:   PetscFree(rvals);
1677:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
1678:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);

1680:   if (reuse == MAT_INITIAL_MATRIX || *matout != A) *matout = B;
1681:   else {
1682:     MatHeaderMerge(A,B);
1683:   }
1684:   return(0);
1685: }

1689: PetscErrorCode MatDiagonalScale_MPIBAIJ(Mat mat,Vec ll,Vec rr)
1690: {
1691:   Mat_MPIBAIJ    *baij = (Mat_MPIBAIJ*)mat->data;
1692:   Mat            a     = baij->A,b = baij->B;
1694:   PetscInt       s1,s2,s3;

1697:   MatGetLocalSize(mat,&s2,&s3);
1698:   if (rr) {
1699:     VecGetLocalSize(rr,&s1);
1700:     if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
1701:     /* Overlap communication with computation. */
1702:     VecScatterBegin(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1703:   }
1704:   if (ll) {
1705:     VecGetLocalSize(ll,&s1);
1706:     if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
1707:     (*b->ops->diagonalscale)(b,ll,NULL);
1708:   }
1709:   /* scale  the diagonal block */
1710:   (*a->ops->diagonalscale)(a,ll,rr);

1712:   if (rr) {
1713:     /* Do a scatter end and then right scale the off-diagonal block */
1714:     VecScatterEnd(baij->Mvctx,rr,baij->lvec,INSERT_VALUES,SCATTER_FORWARD);
1715:     (*b->ops->diagonalscale)(b,NULL,baij->lvec);
1716:   }
1717:   return(0);
1718: }

1722: PetscErrorCode MatZeroRows_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1723: {
1724:   Mat_MPIBAIJ   *l      = (Mat_MPIBAIJ *) A->data;
1725:   PetscInt      *owners = A->rmap->range;
1726:   PetscInt       n      = A->rmap->n;
1727:   PetscSF        sf;
1728:   PetscInt      *lrows;
1729:   PetscSFNode   *rrows;
1730:   PetscInt       r, p = 0, len = 0;

1734:   /* Create SF where leaves are input rows and roots are owned rows */
1735:   PetscMalloc1(n, &lrows);
1736:   for (r = 0; r < n; ++r) lrows[r] = -1;
1737:   if (!A->nooffproczerorows) {PetscMalloc1(N, &rrows);}
1738:   for (r = 0; r < N; ++r) {
1739:     const PetscInt idx   = rows[r];
1740:     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);
1741:     if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
1742:       PetscLayoutFindOwner(A->rmap,idx,&p);
1743:     }
1744:     if (A->nooffproczerorows) {
1745:       if (p != l->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,l->rank);
1746:       lrows[len++] = idx - owners[p];
1747:     } else {
1748:       rrows[r].rank = p;
1749:       rrows[r].index = rows[r] - owners[p];
1750:     }
1751:   }
1752:   if (!A->nooffproczerorows) {
1753:     PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1754:     PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1755:     /* Collect flags for rows to be zeroed */
1756:     PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1757:     PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1758:     PetscSFDestroy(&sf);
1759:     /* Compress and put in row numbers */
1760:     for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1761:   }
1762:   /* fix right hand side if needed */
1763:   if (x && b) {
1764:     const PetscScalar *xx;
1765:     PetscScalar       *bb;

1767:     VecGetArrayRead(x,&xx);
1768:     VecGetArray(b,&bb);
1769:     for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
1770:     VecRestoreArrayRead(x,&xx);
1771:     VecRestoreArray(b,&bb);
1772:   }

1774:   /* actually zap the local rows */
1775:   /*
1776:         Zero the required rows. If the "diagonal block" of the matrix
1777:      is square and the user wishes to set the diagonal we use separate
1778:      code so that MatSetValues() is not called for each diagonal allocating
1779:      new memory, thus calling lots of mallocs and slowing things down.

1781:   */
1782:   /* must zero l->B before l->A because the (diag) case below may put values into l->B*/
1783:   MatZeroRows_SeqBAIJ(l->B,len,lrows,0.0,NULL,NULL);
1784:   if ((diag != 0.0) && (l->A->rmap->N == l->A->cmap->N)) {
1785:     MatZeroRows_SeqBAIJ(l->A,len,lrows,diag,NULL,NULL);
1786:   } else if (diag != 0.0) {
1787:     MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,0,0);
1788:     if (((Mat_SeqBAIJ*)l->A->data)->nonew) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatZeroRows() on rectangular matrices cannot be used with the Mat options \n\
1789:        MAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR");
1790:     for (r = 0; r < len; ++r) {
1791:       const PetscInt row = lrows[r] + A->rmap->rstart;
1792:       MatSetValues(A,1,&row,1,&row,&diag,INSERT_VALUES);
1793:     }
1794:     MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
1795:     MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
1796:   } else {
1797:     MatZeroRows_SeqBAIJ(l->A,len,lrows,0.0,NULL,NULL);
1798:   }
1799:   PetscFree(lrows);

1801:   /* only change matrix nonzero state if pattern was allowed to be changed */
1802:   if (!((Mat_SeqBAIJ*)(l->A->data))->keepnonzeropattern) {
1803:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1804:     MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1805:   }
1806:   return(0);
1807: }

1811: PetscErrorCode MatZeroRowsColumns_MPIBAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
1812: {
1813:   Mat_MPIBAIJ       *l = (Mat_MPIBAIJ*)A->data;
1814:   PetscErrorCode    ierr;
1815:   PetscMPIInt       n = A->rmap->n;
1816:   PetscInt          i,j,k,r,p = 0,len = 0,row,col,count;
1817:   PetscInt          *lrows,*owners = A->rmap->range;
1818:   PetscSFNode       *rrows;
1819:   PetscSF           sf;
1820:   const PetscScalar *xx;
1821:   PetscScalar       *bb,*mask;
1822:   Vec               xmask,lmask;
1823:   Mat_SeqBAIJ       *baij = (Mat_SeqBAIJ*)l->B->data;
1824:   PetscInt           bs = A->rmap->bs, bs2 = baij->bs2;
1825:   PetscScalar       *aa;

1828:   /* Create SF where leaves are input rows and roots are owned rows */
1829:   PetscMalloc1(n, &lrows);
1830:   for (r = 0; r < n; ++r) lrows[r] = -1;
1831:   PetscMalloc1(N, &rrows);
1832:   for (r = 0; r < N; ++r) {
1833:     const PetscInt idx   = rows[r];
1834:     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);
1835:     if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
1836:       PetscLayoutFindOwner(A->rmap,idx,&p);
1837:     }
1838:     rrows[r].rank  = p;
1839:     rrows[r].index = rows[r] - owners[p];
1840:   }
1841:   PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
1842:   PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
1843:   /* Collect flags for rows to be zeroed */
1844:   PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1845:   PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
1846:   PetscSFDestroy(&sf);
1847:   /* Compress and put in row numbers */
1848:   for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
1849:   /* zero diagonal part of matrix */
1850:   MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
1851:   /* handle off diagonal part of matrix */
1852:   MatGetVecs(A,&xmask,NULL);
1853:   VecDuplicate(l->lvec,&lmask);
1854:   VecGetArray(xmask,&bb);
1855:   for (i=0; i<len; i++) bb[lrows[i]] = 1;
1856:   VecRestoreArray(xmask,&bb);
1857:   VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1858:   VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
1859:   VecDestroy(&xmask);
1860:   if (x) {
1861:     VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1862:     VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
1863:     VecGetArrayRead(l->lvec,&xx);
1864:     VecGetArray(b,&bb);
1865:   }
1866:   VecGetArray(lmask,&mask);
1867:   /* remove zeroed rows of off diagonal matrix */
1868:   for (i = 0; i < len; ++i) {
1869:     row   = lrows[i];
1870:     count = (baij->i[row/bs +1] - baij->i[row/bs])*bs;
1871:     aa    = ((MatScalar*)(baij->a)) + baij->i[row/bs]*bs2 + (row%bs);
1872:     for (k = 0; k < count; ++k) {
1873:       aa[0] = 0.0;
1874:       aa   += bs;
1875:     }
1876:   }
1877:   /* loop over all elements of off process part of matrix zeroing removed columns*/
1878:   for (i = 0; i < l->B->rmap->N; ++i) {
1879:     row = i/bs;
1880:     for (j = baij->i[row]; j < baij->i[row+1]; ++j) {
1881:       for (k = 0; k < bs; ++k) {
1882:         col = bs*baij->j[j] + k;
1883:         if (PetscAbsScalar(mask[col])) {
1884:           aa = ((MatScalar*)(baij->a)) + j*bs2 + (i%bs) + bs*k;
1885:           if (b) bb[i] -= aa[0]*xx[col];
1886:           aa[0] = 0.0;
1887:         }
1888:       }
1889:     }
1890:   }
1891:   if (x) {
1892:     VecRestoreArray(b,&bb);
1893:     VecRestoreArrayRead(l->lvec,&xx);
1894:   }
1895:   VecRestoreArray(lmask,&mask);
1896:   VecDestroy(&lmask);
1897:   PetscFree(lrows);

1899:   /* only change matrix nonzero state if pattern was allowed to be changed */
1900:   if (!((Mat_SeqBAIJ*)(l->A->data))->keepnonzeropattern) {
1901:     PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1902:     MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1903:   }
1904:   return(0);
1905: }

1909: PetscErrorCode MatSetUnfactored_MPIBAIJ(Mat A)
1910: {
1911:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

1915:   MatSetUnfactored(a->A);
1916:   return(0);
1917: }

1919: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat,MatDuplicateOption,Mat*);

1923: PetscErrorCode MatEqual_MPIBAIJ(Mat A,Mat B,PetscBool  *flag)
1924: {
1925:   Mat_MPIBAIJ    *matB = (Mat_MPIBAIJ*)B->data,*matA = (Mat_MPIBAIJ*)A->data;
1926:   Mat            a,b,c,d;
1927:   PetscBool      flg;

1931:   a = matA->A; b = matA->B;
1932:   c = matB->A; d = matB->B;

1934:   MatEqual(a,c,&flg);
1935:   if (flg) {
1936:     MatEqual(b,d,&flg);
1937:   }
1938:   MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
1939:   return(0);
1940: }

1944: PetscErrorCode MatCopy_MPIBAIJ(Mat A,Mat B,MatStructure str)
1945: {
1947:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
1948:   Mat_MPIBAIJ    *b = (Mat_MPIBAIJ*)B->data;

1951:   /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
1952:   if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
1953:     MatCopy_Basic(A,B,str);
1954:   } else {
1955:     MatCopy(a->A,b->A,str);
1956:     MatCopy(a->B,b->B,str);
1957:   }
1958:   return(0);
1959: }

1963: PetscErrorCode MatSetUp_MPIBAIJ(Mat A)
1964: {

1968:   MatMPIBAIJSetPreallocation(A,A->rmap->bs,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
1969:   return(0);
1970: }

1974: PetscErrorCode MatAXPYGetPreallocation_MPIBAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
1975: {
1977:   PetscInt       bs = Y->rmap->bs,m = Y->rmap->N/bs;
1978:   Mat_SeqBAIJ    *x = (Mat_SeqBAIJ*)X->data;
1979:   Mat_SeqBAIJ    *y = (Mat_SeqBAIJ*)Y->data;

1982:   MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
1983:   return(0);
1984: }

1988: PetscErrorCode MatAXPY_MPIBAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
1989: {
1991:   Mat_MPIBAIJ    *xx=(Mat_MPIBAIJ*)X->data,*yy=(Mat_MPIBAIJ*)Y->data;
1992:   PetscBLASInt   bnz,one=1;
1993:   Mat_SeqBAIJ    *x,*y;

1996:   if (str == SAME_NONZERO_PATTERN) {
1997:     PetscScalar alpha = a;
1998:     x    = (Mat_SeqBAIJ*)xx->A->data;
1999:     y    = (Mat_SeqBAIJ*)yy->A->data;
2000:     PetscBLASIntCast(x->nz,&bnz);
2001:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2002:     x    = (Mat_SeqBAIJ*)xx->B->data;
2003:     y    = (Mat_SeqBAIJ*)yy->B->data;
2004:     PetscBLASIntCast(x->nz,&bnz);
2005:     PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2006:     PetscObjectStateIncrease((PetscObject)Y);
2007:   } else {
2008:     Mat      B;
2009:     PetscInt *nnz_d,*nnz_o,bs=Y->rmap->bs;
2010:     PetscMalloc1(yy->A->rmap->N,&nnz_d);
2011:     PetscMalloc1(yy->B->rmap->N,&nnz_o);
2012:     MatCreate(PetscObjectComm((PetscObject)Y),&B);
2013:     PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2014:     MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2015:     MatSetBlockSizesFromMats(B,Y,Y);
2016:     MatSetType(B,MATMPIBAIJ);
2017:     MatAXPYGetPreallocation_SeqBAIJ(yy->A,xx->A,nnz_d);
2018:     MatAXPYGetPreallocation_MPIBAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2019:     MatMPIBAIJSetPreallocation(B,bs,0,nnz_d,0,nnz_o);
2020:     /* MatAXPY_BasicWithPreallocation() for BAIJ matrix is much slower than AIJ, even for bs=1 ! */
2021:     MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2022:     MatHeaderReplace(Y,B);
2023:     PetscFree(nnz_d);
2024:     PetscFree(nnz_o);
2025:   }
2026:   return(0);
2027: }

2031: PetscErrorCode MatRealPart_MPIBAIJ(Mat A)
2032: {
2033:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

2037:   MatRealPart(a->A);
2038:   MatRealPart(a->B);
2039:   return(0);
2040: }

2044: PetscErrorCode MatImaginaryPart_MPIBAIJ(Mat A)
2045: {
2046:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;

2050:   MatImaginaryPart(a->A);
2051:   MatImaginaryPart(a->B);
2052:   return(0);
2053: }

2057: PetscErrorCode MatGetSubMatrix_MPIBAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
2058: {
2060:   IS             iscol_local;
2061:   PetscInt       csize;

2064:   ISGetLocalSize(iscol,&csize);
2065:   if (call == MAT_REUSE_MATRIX) {
2066:     PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
2067:     if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2068:   } else {
2069:     ISAllGather(iscol,&iscol_local);
2070:   }
2071:   MatGetSubMatrix_MPIBAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);
2072:   if (call == MAT_INITIAL_MATRIX) {
2073:     PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
2074:     ISDestroy(&iscol_local);
2075:   }
2076:   return(0);
2077: }
2078: extern PetscErrorCode MatGetSubMatrices_MPIBAIJ_local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,PetscBool*,Mat*);
2081: /*
2082:   Not great since it makes two copies of the submatrix, first an SeqBAIJ
2083:   in local and then by concatenating the local matrices the end result.
2084:   Writing it directly would be much like MatGetSubMatrices_MPIBAIJ()
2085: */
2086: PetscErrorCode MatGetSubMatrix_MPIBAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
2087: {
2089:   PetscMPIInt    rank,size;
2090:   PetscInt       i,m,n,rstart,row,rend,nz,*cwork,j,bs;
2091:   PetscInt       *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol,nrow;
2092:   Mat            M,Mreuse;
2093:   MatScalar      *vwork,*aa;
2094:   MPI_Comm       comm;
2095:   IS             isrow_new, iscol_new;
2096:   PetscBool      idflag,allrows, allcols;
2097:   Mat_SeqBAIJ    *aij;

2100:   PetscObjectGetComm((PetscObject)mat,&comm);
2101:   MPI_Comm_rank(comm,&rank);
2102:   MPI_Comm_size(comm,&size);
2103:   /* The compression and expansion should be avoided. Doesn't point
2104:      out errors, might change the indices, hence buggey */
2105:   ISCompressIndicesGeneral(mat->rmap->N,mat->rmap->n,mat->rmap->bs,1,&isrow,&isrow_new);
2106:   ISCompressIndicesGeneral(mat->cmap->N,mat->cmap->n,mat->cmap->bs,1,&iscol,&iscol_new);

2108:   /* Check for special case: each processor gets entire matrix columns */
2109:   ISIdentity(iscol,&idflag);
2110:   ISGetLocalSize(iscol,&ncol);
2111:   if (idflag && ncol == mat->cmap->N) allcols = PETSC_TRUE;
2112:   else allcols = PETSC_FALSE;

2114:   ISIdentity(isrow,&idflag);
2115:   ISGetLocalSize(isrow,&nrow);
2116:   if (idflag && nrow == mat->rmap->N) allrows = PETSC_TRUE;
2117:   else allrows = PETSC_FALSE;

2119:   if (call ==  MAT_REUSE_MATRIX) {
2120:     PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
2121:     if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
2122:     MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_REUSE_MATRIX,&allrows,&allcols,&Mreuse);
2123:   } else {
2124:     MatGetSubMatrices_MPIBAIJ_local(mat,1,&isrow_new,&iscol_new,MAT_INITIAL_MATRIX,&allrows,&allcols,&Mreuse);
2125:   }
2126:   ISDestroy(&isrow_new);
2127:   ISDestroy(&iscol_new);
2128:   /*
2129:       m - number of local rows
2130:       n - number of columns (same on all processors)
2131:       rstart - first row in new global matrix generated
2132:   */
2133:   MatGetBlockSize(mat,&bs);
2134:   MatGetSize(Mreuse,&m,&n);
2135:   m    = m/bs;
2136:   n    = n/bs;

2138:   if (call == MAT_INITIAL_MATRIX) {
2139:     aij = (Mat_SeqBAIJ*)(Mreuse)->data;
2140:     ii  = aij->i;
2141:     jj  = aij->j;

2143:     /*
2144:         Determine the number of non-zeros in the diagonal and off-diagonal
2145:         portions of the matrix in order to do correct preallocation
2146:     */

2148:     /* first get start and end of "diagonal" columns */
2149:     if (csize == PETSC_DECIDE) {
2150:       ISGetSize(isrow,&mglobal);
2151:       if (mglobal == n*bs) { /* square matrix */
2152:         nlocal = m;
2153:       } else {
2154:         nlocal = n/size + ((n % size) > rank);
2155:       }
2156:     } else {
2157:       nlocal = csize/bs;
2158:     }
2159:     MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
2160:     rstart = rend - nlocal;
2161:     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);

2163:     /* next, compute all the lengths */
2164:     PetscMalloc2(m+1,&dlens,m+1,&olens);
2165:     for (i=0; i<m; i++) {
2166:       jend = ii[i+1] - ii[i];
2167:       olen = 0;
2168:       dlen = 0;
2169:       for (j=0; j<jend; j++) {
2170:         if (*jj < rstart || *jj >= rend) olen++;
2171:         else dlen++;
2172:         jj++;
2173:       }
2174:       olens[i] = olen;
2175:       dlens[i] = dlen;
2176:     }
2177:     MatCreate(comm,&M);
2178:     MatSetSizes(M,bs*m,bs*nlocal,PETSC_DECIDE,bs*n);
2179:     MatSetType(M,((PetscObject)mat)->type_name);
2180:     MatMPIBAIJSetPreallocation(M,bs,0,dlens,0,olens);
2181:     PetscFree2(dlens,olens);
2182:   } else {
2183:     PetscInt ml,nl;

2185:     M    = *newmat;
2186:     MatGetLocalSize(M,&ml,&nl);
2187:     if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
2188:     MatZeroEntries(M);
2189:     /*
2190:          The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
2191:        rather than the slower MatSetValues().
2192:     */
2193:     M->was_assembled = PETSC_TRUE;
2194:     M->assembled     = PETSC_FALSE;
2195:   }
2196:   MatSetOption(M,MAT_ROW_ORIENTED,PETSC_FALSE);
2197:   MatGetOwnershipRange(M,&rstart,&rend);
2198:   aij  = (Mat_SeqBAIJ*)(Mreuse)->data;
2199:   ii   = aij->i;
2200:   jj   = aij->j;
2201:   aa   = aij->a;
2202:   for (i=0; i<m; i++) {
2203:     row   = rstart/bs + i;
2204:     nz    = ii[i+1] - ii[i];
2205:     cwork = jj;     jj += nz;
2206:     vwork = aa;     aa += nz*bs*bs;
2207:     MatSetValuesBlocked_MPIBAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
2208:   }

2210:   MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
2211:   MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
2212:   *newmat = M;

2214:   /* save submatrix used in processor for next request */
2215:   if (call ==  MAT_INITIAL_MATRIX) {
2216:     PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
2217:     PetscObjectDereference((PetscObject)Mreuse);
2218:   }
2219:   return(0);
2220: }

2224: PetscErrorCode MatPermute_MPIBAIJ(Mat A,IS rowp,IS colp,Mat *B)
2225: {
2226:   MPI_Comm       comm,pcomm;
2227:   PetscInt       clocal_size,nrows;
2228:   const PetscInt *rows;
2229:   PetscMPIInt    size;
2230:   IS             crowp,lcolp;

2234:   PetscObjectGetComm((PetscObject)A,&comm);
2235:   /* make a collective version of 'rowp' */
2236:   PetscObjectGetComm((PetscObject)rowp,&pcomm);
2237:   if (pcomm==comm) {
2238:     crowp = rowp;
2239:   } else {
2240:     ISGetSize(rowp,&nrows);
2241:     ISGetIndices(rowp,&rows);
2242:     ISCreateGeneral(comm,nrows,rows,PETSC_COPY_VALUES,&crowp);
2243:     ISRestoreIndices(rowp,&rows);
2244:   }
2245:   ISSetPermutation(crowp);
2246:   /* make a local version of 'colp' */
2247:   PetscObjectGetComm((PetscObject)colp,&pcomm);
2248:   MPI_Comm_size(pcomm,&size);
2249:   if (size==1) {
2250:     lcolp = colp;
2251:   } else {
2252:     ISAllGather(colp,&lcolp);
2253:   }
2254:   ISSetPermutation(lcolp);
2255:   /* now we just get the submatrix */
2256:   MatGetLocalSize(A,NULL,&clocal_size);
2257:   MatGetSubMatrix_MPIBAIJ_Private(A,crowp,lcolp,clocal_size,MAT_INITIAL_MATRIX,B);
2258:   /* clean up */
2259:   if (pcomm!=comm) {
2260:     ISDestroy(&crowp);
2261:   }
2262:   if (size>1) {
2263:     ISDestroy(&lcolp);
2264:   }
2265:   return(0);
2266: }

2270: PetscErrorCode  MatGetGhosts_MPIBAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
2271: {
2272:   Mat_MPIBAIJ *baij = (Mat_MPIBAIJ*) mat->data;
2273:   Mat_SeqBAIJ *B    = (Mat_SeqBAIJ*)baij->B->data;

2276:   if (nghosts) *nghosts = B->nbs;
2277:   if (ghosts) *ghosts = baij->garray;
2278:   return(0);
2279: }

2283: PetscErrorCode MatGetSeqNonzeroStructure_MPIBAIJ(Mat A,Mat *newmat)
2284: {
2285:   Mat            B;
2286:   Mat_MPIBAIJ    *a  = (Mat_MPIBAIJ*)A->data;
2287:   Mat_SeqBAIJ    *ad = (Mat_SeqBAIJ*)a->A->data,*bd = (Mat_SeqBAIJ*)a->B->data;
2288:   Mat_SeqAIJ     *b;
2290:   PetscMPIInt    size,rank,*recvcounts = 0,*displs = 0;
2291:   PetscInt       sendcount,i,*rstarts = A->rmap->range,n,cnt,j,bs = A->rmap->bs;
2292:   PetscInt       m,*garray = a->garray,*lens,*jsendbuf,*a_jsendbuf,*b_jsendbuf;

2295:   MPI_Comm_size(PetscObjectComm((PetscObject)A),&size);
2296:   MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);

2298:   /* ----------------------------------------------------------------
2299:      Tell every processor the number of nonzeros per row
2300:   */
2301:   PetscMalloc1((A->rmap->N/bs),&lens);
2302:   for (i=A->rmap->rstart/bs; i<A->rmap->rend/bs; i++) {
2303:     lens[i] = ad->i[i-A->rmap->rstart/bs+1] - ad->i[i-A->rmap->rstart/bs] + bd->i[i-A->rmap->rstart/bs+1] - bd->i[i-A->rmap->rstart/bs];
2304:   }
2305:   sendcount = A->rmap->rend/bs - A->rmap->rstart/bs;
2306:   PetscMalloc1(2*size,&recvcounts);
2307:   displs    = recvcounts + size;
2308:   for (i=0; i<size; i++) {
2309:     recvcounts[i] = A->rmap->range[i+1]/bs - A->rmap->range[i]/bs;
2310:     displs[i]     = A->rmap->range[i]/bs;
2311:   }
2312: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2313:   MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2314: #else
2315:   MPI_Allgatherv(lens+A->rmap->rstart/bs,sendcount,MPIU_INT,lens,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2316: #endif
2317:   /* ---------------------------------------------------------------
2318:      Create the sequential matrix of the same type as the local block diagonal
2319:   */
2320:   MatCreate(PETSC_COMM_SELF,&B);
2321:   MatSetSizes(B,A->rmap->N/bs,A->cmap->N/bs,PETSC_DETERMINE,PETSC_DETERMINE);
2322:   MatSetType(B,MATSEQAIJ);
2323:   MatSeqAIJSetPreallocation(B,0,lens);
2324:   b    = (Mat_SeqAIJ*)B->data;

2326:   /*--------------------------------------------------------------------
2327:     Copy my part of matrix column indices over
2328:   */
2329:   sendcount  = ad->nz + bd->nz;
2330:   jsendbuf   = b->j + b->i[rstarts[rank]/bs];
2331:   a_jsendbuf = ad->j;
2332:   b_jsendbuf = bd->j;
2333:   n          = A->rmap->rend/bs - A->rmap->rstart/bs;
2334:   cnt        = 0;
2335:   for (i=0; i<n; i++) {

2337:     /* put in lower diagonal portion */
2338:     m = bd->i[i+1] - bd->i[i];
2339:     while (m > 0) {
2340:       /* is it above diagonal (in bd (compressed) numbering) */
2341:       if (garray[*b_jsendbuf] > A->rmap->rstart/bs + i) break;
2342:       jsendbuf[cnt++] = garray[*b_jsendbuf++];
2343:       m--;
2344:     }

2346:     /* put in diagonal portion */
2347:     for (j=ad->i[i]; j<ad->i[i+1]; j++) {
2348:       jsendbuf[cnt++] = A->rmap->rstart/bs + *a_jsendbuf++;
2349:     }

2351:     /* put in upper diagonal portion */
2352:     while (m-- > 0) {
2353:       jsendbuf[cnt++] = garray[*b_jsendbuf++];
2354:     }
2355:   }
2356:   if (cnt != sendcount) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Corrupted PETSc matrix: nz given %D actual nz %D",sendcount,cnt);

2358:   /*--------------------------------------------------------------------
2359:     Gather all column indices to all processors
2360:   */
2361:   for (i=0; i<size; i++) {
2362:     recvcounts[i] = 0;
2363:     for (j=A->rmap->range[i]/bs; j<A->rmap->range[i+1]/bs; j++) {
2364:       recvcounts[i] += lens[j];
2365:     }
2366:   }
2367:   displs[0] = 0;
2368:   for (i=1; i<size; i++) {
2369:     displs[i] = displs[i-1] + recvcounts[i-1];
2370:   }
2371: #if defined(PETSC_HAVE_MPI_IN_PLACE)
2372:   MPI_Allgatherv(MPI_IN_PLACE,0,MPI_DATATYPE_NULL,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2373: #else
2374:   MPI_Allgatherv(jsendbuf,sendcount,MPIU_INT,b->j,recvcounts,displs,MPIU_INT,PetscObjectComm((PetscObject)A));
2375: #endif
2376:   /*--------------------------------------------------------------------
2377:     Assemble the matrix into useable form (note numerical values not yet set)
2378:   */
2379:   /* set the b->ilen (length of each row) values */
2380:   PetscMemcpy(b->ilen,lens,(A->rmap->N/bs)*sizeof(PetscInt));
2381:   /* set the b->i indices */
2382:   b->i[0] = 0;
2383:   for (i=1; i<=A->rmap->N/bs; i++) {
2384:     b->i[i] = b->i[i-1] + lens[i-1];
2385:   }
2386:   PetscFree(lens);
2387:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2388:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2389:   PetscFree(recvcounts);

2391:   if (A->symmetric) {
2392:     MatSetOption(B,MAT_SYMMETRIC,PETSC_TRUE);
2393:   } else if (A->hermitian) {
2394:     MatSetOption(B,MAT_HERMITIAN,PETSC_TRUE);
2395:   } else if (A->structurally_symmetric) {
2396:     MatSetOption(B,MAT_STRUCTURALLY_SYMMETRIC,PETSC_TRUE);
2397:   }
2398:   *newmat = B;
2399:   return(0);
2400: }

2404: PetscErrorCode MatSOR_MPIBAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
2405: {
2406:   Mat_MPIBAIJ    *mat = (Mat_MPIBAIJ*)matin->data;
2408:   Vec            bb1 = 0;

2411:   if (flag == SOR_APPLY_UPPER) {
2412:     (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2413:     return(0);
2414:   }

2416:   if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS) {
2417:     VecDuplicate(bb,&bb1);
2418:   }

2420:   if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
2421:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2422:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2423:       its--;
2424:     }

2426:     while (its--) {
2427:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2428:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2430:       /* update rhs: bb1 = bb - B*x */
2431:       VecScale(mat->lvec,-1.0);
2432:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2434:       /* local sweep */
2435:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
2436:     }
2437:   } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
2438:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2439:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2440:       its--;
2441:     }
2442:     while (its--) {
2443:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2444:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2446:       /* update rhs: bb1 = bb - B*x */
2447:       VecScale(mat->lvec,-1.0);
2448:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2450:       /* local sweep */
2451:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
2452:     }
2453:   } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
2454:     if (flag & SOR_ZERO_INITIAL_GUESS) {
2455:       (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
2456:       its--;
2457:     }
2458:     while (its--) {
2459:       VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
2460:       VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);

2462:       /* update rhs: bb1 = bb - B*x */
2463:       VecScale(mat->lvec,-1.0);
2464:       (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);

2466:       /* local sweep */
2467:       (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
2468:     }
2469:   } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel version of SOR requested not supported");

2471:   VecDestroy(&bb1);
2472:   return(0);
2473: }

2477: PetscErrorCode MatGetColumnNorms_MPIBAIJ(Mat A,NormType type,PetscReal *norms)
2478: {
2480:   Mat_MPIBAIJ    *aij = (Mat_MPIBAIJ*)A->data;
2481:   PetscInt       N,i,*garray = aij->garray;
2482:   PetscInt       ib,jb,bs = A->rmap->bs;
2483:   Mat_SeqBAIJ    *a_aij = (Mat_SeqBAIJ*) aij->A->data;
2484:   MatScalar      *a_val = a_aij->a;
2485:   Mat_SeqBAIJ    *b_aij = (Mat_SeqBAIJ*) aij->B->data;
2486:   MatScalar      *b_val = b_aij->a;
2487:   PetscReal      *work;

2490:   MatGetSize(A,NULL,&N);
2491:   PetscCalloc1(N,&work);
2492:   if (type == NORM_2) {
2493:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2494:       for (jb=0; jb<bs; jb++) {
2495:         for (ib=0; ib<bs; ib++) {
2496:           work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val * *a_val);
2497:           a_val++;
2498:         }
2499:       }
2500:     }
2501:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2502:       for (jb=0; jb<bs; jb++) {
2503:         for (ib=0; ib<bs; ib++) {
2504:           work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val * *b_val);
2505:           b_val++;
2506:         }
2507:       }
2508:     }
2509:   } else if (type == NORM_1) {
2510:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2511:       for (jb=0; jb<bs; jb++) {
2512:         for (ib=0; ib<bs; ib++) {
2513:           work[A->cmap->rstart + a_aij->j[i] * bs + jb] += PetscAbsScalar(*a_val);
2514:           a_val++;
2515:         }
2516:       }
2517:     }
2518:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2519:       for (jb=0; jb<bs; jb++) {
2520:        for (ib=0; ib<bs; ib++) {
2521:           work[garray[b_aij->j[i]] * bs + jb] += PetscAbsScalar(*b_val);
2522:           b_val++;
2523:         }
2524:       }
2525:     }
2526:   } else if (type == NORM_INFINITY) {
2527:     for (i=a_aij->i[0]; i<a_aij->i[aij->A->rmap->n/bs]; i++) {
2528:       for (jb=0; jb<bs; jb++) {
2529:         for (ib=0; ib<bs; ib++) {
2530:           int col = A->cmap->rstart + a_aij->j[i] * bs + jb;
2531:           work[col] = PetscMax(PetscAbsScalar(*a_val), work[col]);
2532:           a_val++;
2533:         }
2534:       }
2535:     }
2536:     for (i=b_aij->i[0]; i<b_aij->i[aij->B->rmap->n/bs]; i++) {
2537:       for (jb=0; jb<bs; jb++) {
2538:         for (ib=0; ib<bs; ib++) {
2539:           int col = garray[b_aij->j[i]] * bs + jb;
2540:           work[col] = PetscMax(PetscAbsScalar(*b_val), work[col]);
2541:           b_val++;
2542:         }
2543:       }
2544:     }
2545:   } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
2546:   if (type == NORM_INFINITY) {
2547:     MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
2548:   } else {
2549:     MPI_Allreduce(work,norms,N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
2550:   }
2551:   PetscFree(work);
2552:   if (type == NORM_2) {
2553:     for (i=0; i<N; i++) norms[i] = PetscSqrtReal(norms[i]);
2554:   }
2555:   return(0);
2556: }

2560: PetscErrorCode  MatInvertBlockDiagonal_MPIBAIJ(Mat A,const PetscScalar **values)
2561: {
2562:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*) A->data;

2566:   MatInvertBlockDiagonal(a->A,values);
2567:   return(0);
2568: }


2571: /* -------------------------------------------------------------------*/
2572: static struct _MatOps MatOps_Values = {MatSetValues_MPIBAIJ,
2573:                                        MatGetRow_MPIBAIJ,
2574:                                        MatRestoreRow_MPIBAIJ,
2575:                                        MatMult_MPIBAIJ,
2576:                                 /* 4*/ MatMultAdd_MPIBAIJ,
2577:                                        MatMultTranspose_MPIBAIJ,
2578:                                        MatMultTransposeAdd_MPIBAIJ,
2579:                                        0,
2580:                                        0,
2581:                                        0,
2582:                                 /*10*/ 0,
2583:                                        0,
2584:                                        0,
2585:                                        MatSOR_MPIBAIJ,
2586:                                        MatTranspose_MPIBAIJ,
2587:                                 /*15*/ MatGetInfo_MPIBAIJ,
2588:                                        MatEqual_MPIBAIJ,
2589:                                        MatGetDiagonal_MPIBAIJ,
2590:                                        MatDiagonalScale_MPIBAIJ,
2591:                                        MatNorm_MPIBAIJ,
2592:                                 /*20*/ MatAssemblyBegin_MPIBAIJ,
2593:                                        MatAssemblyEnd_MPIBAIJ,
2594:                                        MatSetOption_MPIBAIJ,
2595:                                        MatZeroEntries_MPIBAIJ,
2596:                                 /*24*/ MatZeroRows_MPIBAIJ,
2597:                                        0,
2598:                                        0,
2599:                                        0,
2600:                                        0,
2601:                                 /*29*/ MatSetUp_MPIBAIJ,
2602:                                        0,
2603:                                        0,
2604:                                        0,
2605:                                        0,
2606:                                 /*34*/ MatDuplicate_MPIBAIJ,
2607:                                        0,
2608:                                        0,
2609:                                        0,
2610:                                        0,
2611:                                 /*39*/ MatAXPY_MPIBAIJ,
2612:                                        MatGetSubMatrices_MPIBAIJ,
2613:                                        MatIncreaseOverlap_MPIBAIJ,
2614:                                        MatGetValues_MPIBAIJ,
2615:                                        MatCopy_MPIBAIJ,
2616:                                 /*44*/ 0,
2617:                                        MatScale_MPIBAIJ,
2618:                                        0,
2619:                                        0,
2620:                                        MatZeroRowsColumns_MPIBAIJ,
2621:                                 /*49*/ 0,
2622:                                        0,
2623:                                        0,
2624:                                        0,
2625:                                        0,
2626:                                 /*54*/ MatFDColoringCreate_MPIXAIJ,
2627:                                        0,
2628:                                        MatSetUnfactored_MPIBAIJ,
2629:                                        MatPermute_MPIBAIJ,
2630:                                        MatSetValuesBlocked_MPIBAIJ,
2631:                                 /*59*/ MatGetSubMatrix_MPIBAIJ,
2632:                                        MatDestroy_MPIBAIJ,
2633:                                        MatView_MPIBAIJ,
2634:                                        0,
2635:                                        0,
2636:                                 /*64*/ 0,
2637:                                        0,
2638:                                        0,
2639:                                        0,
2640:                                        0,
2641:                                 /*69*/ MatGetRowMaxAbs_MPIBAIJ,
2642:                                        0,
2643:                                        0,
2644:                                        0,
2645:                                        0,
2646:                                 /*74*/ 0,
2647:                                        MatFDColoringApply_BAIJ,
2648:                                        0,
2649:                                        0,
2650:                                        0,
2651:                                 /*79*/ 0,
2652:                                        0,
2653:                                        0,
2654:                                        0,
2655:                                        MatLoad_MPIBAIJ,
2656:                                 /*84*/ 0,
2657:                                        0,
2658:                                        0,
2659:                                        0,
2660:                                        0,
2661:                                 /*89*/ 0,
2662:                                        0,
2663:                                        0,
2664:                                        0,
2665:                                        0,
2666:                                 /*94*/ 0,
2667:                                        0,
2668:                                        0,
2669:                                        0,
2670:                                        0,
2671:                                 /*99*/ 0,
2672:                                        0,
2673:                                        0,
2674:                                        0,
2675:                                        0,
2676:                                 /*104*/0,
2677:                                        MatRealPart_MPIBAIJ,
2678:                                        MatImaginaryPart_MPIBAIJ,
2679:                                        0,
2680:                                        0,
2681:                                 /*109*/0,
2682:                                        0,
2683:                                        0,
2684:                                        0,
2685:                                        0,
2686:                                 /*114*/MatGetSeqNonzeroStructure_MPIBAIJ,
2687:                                        0,
2688:                                        MatGetGhosts_MPIBAIJ,
2689:                                        0,
2690:                                        0,
2691:                                 /*119*/0,
2692:                                        0,
2693:                                        0,
2694:                                        0,
2695:                                        MatGetMultiProcBlock_MPIBAIJ,
2696:                                 /*124*/0,
2697:                                        MatGetColumnNorms_MPIBAIJ,
2698:                                        MatInvertBlockDiagonal_MPIBAIJ,
2699:                                        0,
2700:                                        0,
2701:                                /*129*/ 0,
2702:                                        0,
2703:                                        0,
2704:                                        0,
2705:                                        0,
2706:                                /*134*/ 0,
2707:                                        0,
2708:                                        0,
2709:                                        0,
2710:                                        0,
2711:                                /*139*/ 0,
2712:                                        0,
2713:                                        0,
2714:                                        MatFDColoringSetUp_MPIXAIJ
2715: };

2719: PetscErrorCode  MatGetDiagonalBlock_MPIBAIJ(Mat A,Mat *a)
2720: {
2722:   *a = ((Mat_MPIBAIJ*)A->data)->A;
2723:   return(0);
2724: }

2726: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPISBAIJ(Mat, MatType,MatReuse,Mat*);

2730: PetscErrorCode MatMPIBAIJSetPreallocationCSR_MPIBAIJ(Mat B,PetscInt bs,const PetscInt ii[],const PetscInt jj[],const PetscScalar V[])
2731: {
2732:   PetscInt       m,rstart,cstart,cend;
2733:   PetscInt       i,j,d,nz,nz_max=0,*d_nnz=0,*o_nnz=0;
2734:   const PetscInt *JJ    =0;
2735:   PetscScalar    *values=0;
2736:   PetscBool      roworiented = ((Mat_MPIBAIJ*)B->data)->roworiented;

2740:   PetscLayoutSetBlockSize(B->rmap,bs);
2741:   PetscLayoutSetBlockSize(B->cmap,bs);
2742:   PetscLayoutSetUp(B->rmap);
2743:   PetscLayoutSetUp(B->cmap);
2744:   PetscLayoutGetBlockSize(B->rmap,&bs);
2745:   m      = B->rmap->n/bs;
2746:   rstart = B->rmap->rstart/bs;
2747:   cstart = B->cmap->rstart/bs;
2748:   cend   = B->cmap->rend/bs;

2750:   if (ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"ii[0] must be 0 but it is %D",ii[0]);
2751:   PetscMalloc2(m,&d_nnz,m,&o_nnz);
2752:   for (i=0; i<m; i++) {
2753:     nz = ii[i+1] - ii[i];
2754:     if (nz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative number of columns %D",i,nz);
2755:     nz_max = PetscMax(nz_max,nz);
2756:     JJ     = jj + ii[i];
2757:     for (j=0; j<nz; j++) {
2758:       if (*JJ >= cstart) break;
2759:       JJ++;
2760:     }
2761:     d = 0;
2762:     for (; j<nz; j++) {
2763:       if (*JJ++ >= cend) break;
2764:       d++;
2765:     }
2766:     d_nnz[i] = d;
2767:     o_nnz[i] = nz - d;
2768:   }
2769:   MatMPIBAIJSetPreallocation(B,bs,0,d_nnz,0,o_nnz);
2770:   PetscFree2(d_nnz,o_nnz);

2772:   values = (PetscScalar*)V;
2773:   if (!values) {
2774:     PetscMalloc1(bs*bs*nz_max,&values);
2775:     PetscMemzero(values,bs*bs*nz_max*sizeof(PetscScalar));
2776:   }
2777:   for (i=0; i<m; i++) {
2778:     PetscInt          row    = i + rstart;
2779:     PetscInt          ncols  = ii[i+1] - ii[i];
2780:     const PetscInt    *icols = jj + ii[i];
2781:     if (!roworiented) {         /* block ordering matches the non-nested layout of MatSetValues so we can insert entire rows */
2782:       const PetscScalar *svals = values + (V ? (bs*bs*ii[i]) : 0);
2783:       MatSetValuesBlocked_MPIBAIJ(B,1,&row,ncols,icols,svals,INSERT_VALUES);
2784:     } else {                    /* block ordering does not match so we can only insert one block at a time. */
2785:       PetscInt j;
2786:       for (j=0; j<ncols; j++) {
2787:         const PetscScalar *svals = values + (V ? (bs*bs*(ii[i]+j)) : 0);
2788:         MatSetValuesBlocked_MPIBAIJ(B,1,&row,1,&icols[j],svals,INSERT_VALUES);
2789:       }
2790:     }
2791:   }

2793:   if (!V) { PetscFree(values); }
2794:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2795:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2796:   MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
2797:   return(0);
2798: }

2802: /*@C
2803:    MatMPIBAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in BAIJ format
2804:    (the default parallel PETSc format).

2806:    Collective on MPI_Comm

2808:    Input Parameters:
2809: +  B - the matrix
2810: .  bs - the block size
2811: .  i - the indices into j for the start of each local row (starts with zero)
2812: .  j - the column indices for each local row (starts with zero) these must be sorted for each row
2813: -  v - optional values in the matrix

2815:    Level: developer

2817:    Notes: The order of the entries in values is specified by the MatOption MAT_ROW_ORIENTED.  For example, C programs
2818:    may want to use the default MAT_ROW_ORIENTED=PETSC_TRUE and use an array v[nnz][bs][bs] where the second index is
2819:    over rows within a block and the last index is over columns within a block row.  Fortran programs will likely set
2820:    MAT_ROW_ORIENTED=PETSC_FALSE and use a Fortran array v(bs,bs,nnz) in which the first index is over rows within a
2821:    block column and the second index is over columns within a block.

2823: .keywords: matrix, aij, compressed row, sparse, parallel

2825: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIBAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, MatCreateMPIBAIJWithArrays(), MPIBAIJ
2826: @*/
2827: PetscErrorCode  MatMPIBAIJSetPreallocationCSR(Mat B,PetscInt bs,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
2828: {

2835:   PetscTryMethod(B,"MatMPIBAIJSetPreallocationCSR_C",(Mat,PetscInt,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,bs,i,j,v));
2836:   return(0);
2837: }

2841: PetscErrorCode  MatMPIBAIJSetPreallocation_MPIBAIJ(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt *d_nnz,PetscInt o_nz,const PetscInt *o_nnz)
2842: {
2843:   Mat_MPIBAIJ    *b;
2845:   PetscInt       i;

2848:   MatSetBlockSize(B,PetscAbs(bs));
2849:   PetscLayoutSetUp(B->rmap);
2850:   PetscLayoutSetUp(B->cmap);
2851:   PetscLayoutGetBlockSize(B->rmap,&bs);

2853:   if (d_nnz) {
2854:     for (i=0; i<B->rmap->n/bs; i++) {
2855:       if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than -1: local row %D value %D",i,d_nnz[i]);
2856:     }
2857:   }
2858:   if (o_nnz) {
2859:     for (i=0; i<B->rmap->n/bs; i++) {
2860:       if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than -1: local row %D value %D",i,o_nnz[i]);
2861:     }
2862:   }

2864:   b      = (Mat_MPIBAIJ*)B->data;
2865:   b->bs2 = bs*bs;
2866:   b->mbs = B->rmap->n/bs;
2867:   b->nbs = B->cmap->n/bs;
2868:   b->Mbs = B->rmap->N/bs;
2869:   b->Nbs = B->cmap->N/bs;

2871:   for (i=0; i<=b->size; i++) {
2872:     b->rangebs[i] = B->rmap->range[i]/bs;
2873:   }
2874:   b->rstartbs = B->rmap->rstart/bs;
2875:   b->rendbs   = B->rmap->rend/bs;
2876:   b->cstartbs = B->cmap->rstart/bs;
2877:   b->cendbs   = B->cmap->rend/bs;

2879:   if (!B->preallocated) {
2880:     MatCreate(PETSC_COMM_SELF,&b->A);
2881:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2882:     MatSetType(b->A,MATSEQBAIJ);
2883:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2884:     MatCreate(PETSC_COMM_SELF,&b->B);
2885:     MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
2886:     MatSetType(b->B,MATSEQBAIJ);
2887:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2888:     MatStashCreate_Private(PetscObjectComm((PetscObject)B),bs,&B->bstash);
2889:   }

2891:   MatSeqBAIJSetPreallocation(b->A,bs,d_nz,d_nnz);
2892:   MatSeqBAIJSetPreallocation(b->B,bs,o_nz,o_nnz);
2893:   B->preallocated = PETSC_TRUE;
2894:   return(0);
2895: }

2897: extern PetscErrorCode  MatDiagonalScaleLocal_MPIBAIJ(Mat,Vec);
2898: extern PetscErrorCode  MatSetHashTableFactor_MPIBAIJ(Mat,PetscReal);

2902: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAdj(Mat B, MatType newtype,MatReuse reuse,Mat *adj)
2903: {
2904:   Mat_MPIBAIJ    *b = (Mat_MPIBAIJ*)B->data;
2906:   Mat_SeqBAIJ    *d  = (Mat_SeqBAIJ*) b->A->data,*o = (Mat_SeqBAIJ*) b->B->data;
2907:   PetscInt       M   = B->rmap->n/B->rmap->bs,i,*ii,*jj,cnt,j,k,rstart = B->rmap->rstart/B->rmap->bs;
2908:   const PetscInt *id = d->i, *jd = d->j, *io = o->i, *jo = o->j, *garray = b->garray;

2911:   PetscMalloc1((M+1),&ii);
2912:   ii[0] = 0;
2913:   for (i=0; i<M; i++) {
2914:     if ((id[i+1] - id[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,id[i],id[i+1]);
2915:     if ((io[i+1] - io[i]) < 0) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Indices wrong %D %D %D",i,io[i],io[i+1]);
2916:     ii[i+1] = ii[i] + id[i+1] - id[i] + io[i+1] - io[i];
2917:     /* remove one from count of matrix has diagonal */
2918:     for (j=id[i]; j<id[i+1]; j++) {
2919:       if (jd[j] == i) {ii[i+1]--;break;}
2920:     }
2921:   }
2922:   PetscMalloc1(ii[M],&jj);
2923:   cnt  = 0;
2924:   for (i=0; i<M; i++) {
2925:     for (j=io[i]; j<io[i+1]; j++) {
2926:       if (garray[jo[j]] > rstart) break;
2927:       jj[cnt++] = garray[jo[j]];
2928:     }
2929:     for (k=id[i]; k<id[i+1]; k++) {
2930:       if (jd[k] != i) {
2931:         jj[cnt++] = rstart + jd[k];
2932:       }
2933:     }
2934:     for (; j<io[i+1]; j++) {
2935:       jj[cnt++] = garray[jo[j]];
2936:     }
2937:   }
2938:   MatCreateMPIAdj(PetscObjectComm((PetscObject)B),M,B->cmap->N/B->rmap->bs,ii,jj,NULL,adj);
2939:   return(0);
2940: }

2942: #include <../src/mat/impls/aij/mpi/mpiaij.h>

2944: PETSC_EXTERN PetscErrorCode MatConvert_SeqBAIJ_SeqAIJ(Mat,MatType,MatReuse,Mat*);

2948: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIAIJ(Mat A,MatType newtype,MatReuse reuse,Mat *newmat)
2949: {
2951:   Mat_MPIBAIJ    *a = (Mat_MPIBAIJ*)A->data;
2952:   Mat            B;
2953:   Mat_MPIAIJ     *b;

2956:   if (!A->assembled) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Matrix must be assembled");

2958:   MatCreate(PetscObjectComm((PetscObject)A),&B);
2959:   MatSetSizes(B,A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N);
2960:   MatSetType(B,MATMPIAIJ);
2961:   MatSeqAIJSetPreallocation(B,0,NULL);
2962:   MatMPIAIJSetPreallocation(B,0,NULL,0,NULL);
2963:   b    = (Mat_MPIAIJ*) B->data;

2965:   MatDestroy(&b->A);
2966:   MatDestroy(&b->B);
2967:   MatDisAssemble_MPIBAIJ(A);
2968:   MatConvert_SeqBAIJ_SeqAIJ(a->A, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->A);
2969:   MatConvert_SeqBAIJ_SeqAIJ(a->B, MATSEQAIJ, MAT_INITIAL_MATRIX, &b->B);
2970:   MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2971:   MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2972:   MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
2973:   MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
2974:   if (reuse == MAT_REUSE_MATRIX) {
2975:     MatHeaderReplace(A,B);
2976:   } else {
2977:    *newmat = B;
2978:   }
2979:   return(0);
2980: }

2982: #if defined(PETSC_HAVE_MUMPS)
2983: PETSC_EXTERN PetscErrorCode MatGetFactor_baij_mumps(Mat,MatFactorType,Mat*);
2984: #endif

2986: /*MC
2987:    MATMPIBAIJ - MATMPIBAIJ = "mpibaij" - A matrix type to be used for distributed block sparse matrices.

2989:    Options Database Keys:
2990: + -mat_type mpibaij - sets the matrix type to "mpibaij" during a call to MatSetFromOptions()
2991: . -mat_block_size <bs> - set the blocksize used to store the matrix
2992: - -mat_use_hash_table <fact>

2994:   Level: beginner

2996: .seealso: MatCreateMPIBAIJ
2997: M*/

2999: PETSC_EXTERN PetscErrorCode MatConvert_MPIBAIJ_MPIBSTRM(Mat,MatType,MatReuse,Mat*);

3003: PETSC_EXTERN PetscErrorCode MatCreate_MPIBAIJ(Mat B)
3004: {
3005:   Mat_MPIBAIJ    *b;
3007:   PetscBool      flg;

3010:   PetscNewLog(B,&b);
3011:   B->data = (void*)b;

3013:   PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
3014:   B->assembled = PETSC_FALSE;

3016:   B->insertmode = NOT_SET_VALUES;
3017:   MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
3018:   MPI_Comm_size(PetscObjectComm((PetscObject)B),&b->size);

3020:   /* build local table of row and column ownerships */
3021:   PetscMalloc1((b->size+1),&b->rangebs);

3023:   /* build cache for off array entries formed */
3024:   MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);

3026:   b->donotstash  = PETSC_FALSE;
3027:   b->colmap      = NULL;
3028:   b->garray      = NULL;
3029:   b->roworiented = PETSC_TRUE;

3031:   /* stuff used in block assembly */
3032:   b->barray = 0;

3034:   /* stuff used for matrix vector multiply */
3035:   b->lvec  = 0;
3036:   b->Mvctx = 0;

3038:   /* stuff for MatGetRow() */
3039:   b->rowindices   = 0;
3040:   b->rowvalues    = 0;
3041:   b->getrowactive = PETSC_FALSE;

3043:   /* hash table stuff */
3044:   b->ht           = 0;
3045:   b->hd           = 0;
3046:   b->ht_size      = 0;
3047:   b->ht_flag      = PETSC_FALSE;
3048:   b->ht_fact      = 0;
3049:   b->ht_total_ct  = 0;
3050:   b->ht_insert_ct = 0;

3052:   /* stuff for MatGetSubMatrices_MPIBAIJ_local() */
3053:   b->ijonly = PETSC_FALSE;

3055:   PetscOptionsBegin(PetscObjectComm((PetscObject)B),NULL,"Options for loading MPIBAIJ matrix 1","Mat");
3056:   PetscOptionsBool("-mat_use_hash_table","Use hash table to save memory in constructing matrix","MatSetOption",PETSC_FALSE,&flg,NULL);
3057:   if (flg) {
3058:     PetscReal fact = 1.39;
3059:     MatSetOption(B,MAT_USE_HASH_TABLE,PETSC_TRUE);
3060:     PetscOptionsReal("-mat_use_hash_table","Use hash table factor","MatMPIBAIJSetHashTableFactor",fact,&fact,NULL);
3061:     if (fact <= 1.0) fact = 1.39;
3062:     MatMPIBAIJSetHashTableFactor(B,fact);
3063:     PetscInfo1(B,"Hash table Factor used %5.2f\n",fact);
3064:   }
3065:   PetscOptionsEnd();

3067: #if defined(PETSC_HAVE_MUMPS)
3068:   PetscObjectComposeFunction((PetscObject)B,"MatGetFactor_mumps_C",MatGetFactor_baij_mumps);
3069: #endif
3070:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiadj_C",MatConvert_MPIBAIJ_MPIAdj);
3071:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpiaij_C",MatConvert_MPIBAIJ_MPIAIJ);
3072:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpisbaij_C",MatConvert_MPIBAIJ_MPISBAIJ);
3073:   PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIBAIJ);
3074:   PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIBAIJ);
3075:   PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIBAIJ);
3076:   PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocation_C",MatMPIBAIJSetPreallocation_MPIBAIJ);
3077:   PetscObjectComposeFunction((PetscObject)B,"MatMPIBAIJSetPreallocationCSR_C",MatMPIBAIJSetPreallocationCSR_MPIBAIJ);
3078:   PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIBAIJ);
3079:   PetscObjectComposeFunction((PetscObject)B,"MatSetHashTableFactor_C",MatSetHashTableFactor_MPIBAIJ);
3080:   PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpibaij_mpibstrm_C",MatConvert_MPIBAIJ_MPIBSTRM);
3081:   PetscObjectChangeTypeName((PetscObject)B,MATMPIBAIJ);
3082:   return(0);
3083: }

3085: /*MC
3086:    MATBAIJ - MATBAIJ = "baij" - A matrix type to be used for block sparse matrices.

3088:    This matrix type is identical to MATSEQBAIJ when constructed with a single process communicator,
3089:    and MATMPIBAIJ otherwise.

3091:    Options Database Keys:
3092: . -mat_type baij - sets the matrix type to "baij" during a call to MatSetFromOptions()

3094:   Level: beginner

3096: .seealso: MatCreateBAIJ(),MATSEQBAIJ,MATMPIBAIJ, MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3097: M*/

3101: /*@C
3102:    MatMPIBAIJSetPreallocation - Allocates memory for a sparse parallel matrix in block AIJ format
3103:    (block compressed row).  For good matrix assembly performance
3104:    the user should preallocate the matrix storage by setting the parameters
3105:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3106:    performance can be increased by more than a factor of 50.

3108:    Collective on Mat

3110:    Input Parameters:
3111: +  B - the matrix
3112: .  bs   - size of block
3113: .  d_nz  - number of block nonzeros per block row in diagonal portion of local
3114:            submatrix  (same for all local rows)
3115: .  d_nnz - array containing the number of block nonzeros in the various block rows
3116:            of the in diagonal portion of the local (possibly different for each block
3117:            row) or NULL.  If you plan to factor the matrix you must leave room for the diagonal entry and
3118:            set it even if it is zero.
3119: .  o_nz  - number of block nonzeros per block row in the off-diagonal portion of local
3120:            submatrix (same for all local rows).
3121: -  o_nnz - array containing the number of nonzeros in the various block rows of the
3122:            off-diagonal portion of the local submatrix (possibly different for
3123:            each block row) or NULL.

3125:    If the *_nnz parameter is given then the *_nz parameter is ignored

3127:    Options Database Keys:
3128: +   -mat_block_size - size of the blocks to use
3129: -   -mat_use_hash_table <fact>

3131:    Notes:
3132:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one processor
3133:    than it must be used on all processors that share the object for that argument.

3135:    Storage Information:
3136:    For a square global matrix we define each processor's diagonal portion
3137:    to be its local rows and the corresponding columns (a square submatrix);
3138:    each processor's off-diagonal portion encompasses the remainder of the
3139:    local matrix (a rectangular submatrix).

3141:    The user can specify preallocated storage for the diagonal part of
3142:    the local submatrix with either d_nz or d_nnz (not both).  Set
3143:    d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3144:    memory allocation.  Likewise, specify preallocated storage for the
3145:    off-diagonal part of the local submatrix with o_nz or o_nnz (not both).

3147:    Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3148:    the figure below we depict these three local rows and all columns (0-11).

3150: .vb
3151:            0 1 2 3 4 5 6 7 8 9 10 11
3152:           --------------------------
3153:    row 3  |o o o d d d o o o o  o  o
3154:    row 4  |o o o d d d o o o o  o  o
3155:    row 5  |o o o d d d o o o o  o  o
3156:           --------------------------
3157: .ve

3159:    Thus, any entries in the d locations are stored in the d (diagonal)
3160:    submatrix, and any entries in the o locations are stored in the
3161:    o (off-diagonal) submatrix.  Note that the d and the o submatrices are
3162:    stored simply in the MATSEQBAIJ format for compressed row storage.

3164:    Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3165:    and o_nz should indicate the number of block nonzeros per row in the o matrix.
3166:    In general, for PDE problems in which most nonzeros are near the diagonal,
3167:    one expects d_nz >> o_nz.   For large problems you MUST preallocate memory
3168:    or you will get TERRIBLE performance; see the users' manual chapter on
3169:    matrices.

3171:    You can call MatGetInfo() to get information on how effective the preallocation was;
3172:    for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
3173:    You can also run with the option -info and look for messages with the string
3174:    malloc in them to see if additional memory allocation was needed.

3176:    Level: intermediate

3178: .keywords: matrix, block, aij, compressed row, sparse, parallel

3180: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocationCSR(), PetscSplitOwnership()
3181: @*/
3182: PetscErrorCode  MatMPIBAIJSetPreallocation(Mat B,PetscInt bs,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
3183: {

3190:   PetscTryMethod(B,"MatMPIBAIJSetPreallocation_C",(Mat,PetscInt,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,bs,d_nz,d_nnz,o_nz,o_nnz));
3191:   return(0);
3192: }

3196: /*@C
3197:    MatCreateBAIJ - Creates a sparse parallel matrix in block AIJ format
3198:    (block compressed row).  For good matrix assembly performance
3199:    the user should preallocate the matrix storage by setting the parameters
3200:    d_nz (or d_nnz) and o_nz (or o_nnz).  By setting these parameters accurately,
3201:    performance can be increased by more than a factor of 50.

3203:    Collective on MPI_Comm

3205:    Input Parameters:
3206: +  comm - MPI communicator
3207: .  bs   - size of blockk
3208: .  m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
3209:            This value should be the same as the local size used in creating the
3210:            y vector for the matrix-vector product y = Ax.
3211: .  n - number of local columns (or PETSC_DECIDE to have calculated if N is given)
3212:            This value should be the same as the local size used in creating the
3213:            x vector for the matrix-vector product y = Ax.
3214: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3215: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3216: .  d_nz  - number of nonzero blocks per block row in diagonal portion of local
3217:            submatrix  (same for all local rows)
3218: .  d_nnz - array containing the number of nonzero blocks in the various block rows
3219:            of the in diagonal portion of the local (possibly different for each block
3220:            row) or NULL.  If you plan to factor the matrix you must leave room for the diagonal entry
3221:            and set it even if it is zero.
3222: .  o_nz  - number of nonzero blocks per block row in the off-diagonal portion of local
3223:            submatrix (same for all local rows).
3224: -  o_nnz - array containing the number of nonzero blocks in the various block rows of the
3225:            off-diagonal portion of the local submatrix (possibly different for
3226:            each block row) or NULL.

3228:    Output Parameter:
3229: .  A - the matrix

3231:    Options Database Keys:
3232: +   -mat_block_size - size of the blocks to use
3233: -   -mat_use_hash_table <fact>

3235:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
3236:    MatXXXXSetPreallocation() paradgm instead of this routine directly.
3237:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

3239:    Notes:
3240:    If the *_nnz parameter is given then the *_nz parameter is ignored

3242:    A nonzero block is any block that as 1 or more nonzeros in it

3244:    The user MUST specify either the local or global matrix dimensions
3245:    (possibly both).

3247:    If PETSC_DECIDE or  PETSC_DETERMINE is used for a particular argument on one processor
3248:    than it must be used on all processors that share the object for that argument.

3250:    Storage Information:
3251:    For a square global matrix we define each processor's diagonal portion
3252:    to be its local rows and the corresponding columns (a square submatrix);
3253:    each processor's off-diagonal portion encompasses the remainder of the
3254:    local matrix (a rectangular submatrix).

3256:    The user can specify preallocated storage for the diagonal part of
3257:    the local submatrix with either d_nz or d_nnz (not both).  Set
3258:    d_nz=PETSC_DEFAULT and d_nnz=NULL for PETSc to control dynamic
3259:    memory allocation.  Likewise, specify preallocated storage for the
3260:    off-diagonal part of the local submatrix with o_nz or o_nnz (not both).

3262:    Consider a processor that owns rows 3, 4 and 5 of a parallel matrix. In
3263:    the figure below we depict these three local rows and all columns (0-11).

3265: .vb
3266:            0 1 2 3 4 5 6 7 8 9 10 11
3267:           --------------------------
3268:    row 3  |o o o d d d o o o o  o  o
3269:    row 4  |o o o d d d o o o o  o  o
3270:    row 5  |o o o d d d o o o o  o  o
3271:           --------------------------
3272: .ve

3274:    Thus, any entries in the d locations are stored in the d (diagonal)
3275:    submatrix, and any entries in the o locations are stored in the
3276:    o (off-diagonal) submatrix.  Note that the d and the o submatrices are
3277:    stored simply in the MATSEQBAIJ format for compressed row storage.

3279:    Now d_nz should indicate the number of block nonzeros per row in the d matrix,
3280:    and o_nz should indicate the number of block nonzeros per row in the o matrix.
3281:    In general, for PDE problems in which most nonzeros are near the diagonal,
3282:    one expects d_nz >> o_nz.   For large problems you MUST preallocate memory
3283:    or you will get TERRIBLE performance; see the users' manual chapter on
3284:    matrices.

3286:    Level: intermediate

3288: .keywords: matrix, block, aij, compressed row, sparse, parallel

3290: .seealso: MatCreate(), MatCreateSeqBAIJ(), MatSetValues(), MatCreateBAIJ(), MatMPIBAIJSetPreallocation(), MatMPIBAIJSetPreallocationCSR()
3291: @*/
3292: PetscErrorCode  MatCreateBAIJ(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
3293: {
3295:   PetscMPIInt    size;

3298:   MatCreate(comm,A);
3299:   MatSetSizes(*A,m,n,M,N);
3300:   MPI_Comm_size(comm,&size);
3301:   if (size > 1) {
3302:     MatSetType(*A,MATMPIBAIJ);
3303:     MatMPIBAIJSetPreallocation(*A,bs,d_nz,d_nnz,o_nz,o_nnz);
3304:   } else {
3305:     MatSetType(*A,MATSEQBAIJ);
3306:     MatSeqBAIJSetPreallocation(*A,bs,d_nz,d_nnz);
3307:   }
3308:   return(0);
3309: }

3313: static PetscErrorCode MatDuplicate_MPIBAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
3314: {
3315:   Mat            mat;
3316:   Mat_MPIBAIJ    *a,*oldmat = (Mat_MPIBAIJ*)matin->data;
3318:   PetscInt       len=0;

3321:   *newmat = 0;
3322:   MatCreate(PetscObjectComm((PetscObject)matin),&mat);
3323:   MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
3324:   MatSetType(mat,((PetscObject)matin)->type_name);
3325:   PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));

3327:   mat->factortype   = matin->factortype;
3328:   mat->preallocated = PETSC_TRUE;
3329:   mat->assembled    = PETSC_TRUE;
3330:   mat->insertmode   = NOT_SET_VALUES;

3332:   a             = (Mat_MPIBAIJ*)mat->data;
3333:   mat->rmap->bs = matin->rmap->bs;
3334:   a->bs2        = oldmat->bs2;
3335:   a->mbs        = oldmat->mbs;
3336:   a->nbs        = oldmat->nbs;
3337:   a->Mbs        = oldmat->Mbs;
3338:   a->Nbs        = oldmat->Nbs;

3340:   PetscLayoutReference(matin->rmap,&mat->rmap);
3341:   PetscLayoutReference(matin->cmap,&mat->cmap);

3343:   a->size         = oldmat->size;
3344:   a->rank         = oldmat->rank;
3345:   a->donotstash   = oldmat->donotstash;
3346:   a->roworiented  = oldmat->roworiented;
3347:   a->rowindices   = 0;
3348:   a->rowvalues    = 0;
3349:   a->getrowactive = PETSC_FALSE;
3350:   a->barray       = 0;
3351:   a->rstartbs     = oldmat->rstartbs;
3352:   a->rendbs       = oldmat->rendbs;
3353:   a->cstartbs     = oldmat->cstartbs;
3354:   a->cendbs       = oldmat->cendbs;

3356:   /* hash table stuff */
3357:   a->ht           = 0;
3358:   a->hd           = 0;
3359:   a->ht_size      = 0;
3360:   a->ht_flag      = oldmat->ht_flag;
3361:   a->ht_fact      = oldmat->ht_fact;
3362:   a->ht_total_ct  = 0;
3363:   a->ht_insert_ct = 0;

3365:   PetscMemcpy(a->rangebs,oldmat->rangebs,(a->size+1)*sizeof(PetscInt));
3366:   if (oldmat->colmap) {
3367: #if defined(PETSC_USE_CTABLE)
3368:     PetscTableCreateCopy(oldmat->colmap,&a->colmap);
3369: #else
3370:     PetscMalloc1((a->Nbs),&a->colmap);
3371:     PetscLogObjectMemory((PetscObject)mat,(a->Nbs)*sizeof(PetscInt));
3372:     PetscMemcpy(a->colmap,oldmat->colmap,(a->Nbs)*sizeof(PetscInt));
3373: #endif
3374:   } else a->colmap = 0;

3376:   if (oldmat->garray && (len = ((Mat_SeqBAIJ*)(oldmat->B->data))->nbs)) {
3377:     PetscMalloc1(len,&a->garray);
3378:     PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
3379:     PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));
3380:   } else a->garray = 0;

3382:   MatStashCreate_Private(PetscObjectComm((PetscObject)matin),matin->rmap->bs,&mat->bstash);
3383:   VecDuplicate(oldmat->lvec,&a->lvec);
3384:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
3385:   VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
3386:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);

3388:   MatDuplicate(oldmat->A,cpvalues,&a->A);
3389:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
3390:   MatDuplicate(oldmat->B,cpvalues,&a->B);
3391:   PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
3392:   PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
3393:   *newmat = mat;
3394:   return(0);
3395: }

3399: PetscErrorCode MatLoad_MPIBAIJ(Mat newmat,PetscViewer viewer)
3400: {
3402:   int            fd;
3403:   PetscInt       i,nz,j,rstart,rend;
3404:   PetscScalar    *vals,*buf;
3405:   MPI_Comm       comm;
3406:   MPI_Status     status;
3407:   PetscMPIInt    rank,size,maxnz;
3408:   PetscInt       header[4],*rowlengths = 0,M,N,m,*rowners,*cols;
3409:   PetscInt       *locrowlens = NULL,*procsnz = NULL,*browners = NULL;
3410:   PetscInt       jj,*mycols,*ibuf,bs=1,Mbs,mbs,extra_rows,mmax;
3411:   PetscMPIInt    tag    = ((PetscObject)viewer)->tag;
3412:   PetscInt       *dlens = NULL,*odlens = NULL,*mask = NULL,*masked1 = NULL,*masked2 = NULL,rowcount,odcount;
3413:   PetscInt       dcount,kmax,k,nzcount,tmp,mend,sizesset=1,grows,gcols;

3416:   PetscObjectGetComm((PetscObject)viewer,&comm);
3417:   PetscOptionsBegin(comm,NULL,"Options for loading MPIBAIJ matrix 2","Mat");
3418:   PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
3419:   PetscOptionsEnd();

3421:   MPI_Comm_size(comm,&size);
3422:   MPI_Comm_rank(comm,&rank);
3423:   if (!rank) {
3424:     PetscViewerBinaryGetDescriptor(viewer,&fd);
3425:     PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
3426:     if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
3427:   }

3429:   if (newmat->rmap->n < 0 && newmat->rmap->N < 0 && newmat->cmap->n < 0 && newmat->cmap->N < 0) sizesset = 0;

3431:   MPI_Bcast(header+1,3,MPIU_INT,0,comm);
3432:   M    = header[1]; N = header[2];

3434:   /* If global rows/cols are set to PETSC_DECIDE, set it to the sizes given in the file */
3435:   if (sizesset && newmat->rmap->N < 0) newmat->rmap->N = M;
3436:   if (sizesset && newmat->cmap->N < 0) newmat->cmap->N = N;

3438:   /* If global sizes are set, check if they are consistent with that given in the file */
3439:   if (sizesset) {
3440:     MatGetSize(newmat,&grows,&gcols);
3441:   }
3442:   if (sizesset && newmat->rmap->N != grows) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows:Matrix in file has (%d) and input matrix has (%d)",M,grows);
3443:   if (sizesset && newmat->cmap->N != gcols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of cols:Matrix in file has (%d) and input matrix has (%d)",N,gcols);

3445:   if (M != N) SETERRQ(PetscObjectComm((PetscObject)viewer),PETSC_ERR_SUP,"Can only do square matrices");

3447:   /*
3448:      This code adds extra rows to make sure the number of rows is
3449:      divisible by the blocksize
3450:   */
3451:   Mbs        = M/bs;
3452:   extra_rows = bs - M + bs*Mbs;
3453:   if (extra_rows == bs) extra_rows = 0;
3454:   else                  Mbs++;
3455:   if (extra_rows && !rank) {
3456:     PetscInfo(viewer,"Padding loaded matrix to match blocksize\n");
3457:   }

3459:   /* determine ownership of all rows */
3460:   if (newmat->rmap->n < 0) { /* PETSC_DECIDE */
3461:     mbs = Mbs/size + ((Mbs % size) > rank);
3462:     m   = mbs*bs;
3463:   } else { /* User set */
3464:     m   = newmat->rmap->n;
3465:     mbs = m/bs;
3466:   }
3467:   PetscMalloc2(size+1,&rowners,size+1,&browners);
3468:   MPI_Allgather(&mbs,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);

3470:   /* process 0 needs enough room for process with most rows */
3471:   if (!rank) {
3472:     mmax = rowners[1];
3473:     for (i=2; i<=size; i++) {
3474:       mmax = PetscMax(mmax,rowners[i]);
3475:     }
3476:     mmax*=bs;
3477:   } else mmax = -1;             /* unused, but compiler warns anyway */

3479:   rowners[0] = 0;
3480:   for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
3481:   for (i=0; i<=size; i++) browners[i] = rowners[i]*bs;
3482:   rstart = rowners[rank];
3483:   rend   = rowners[rank+1];

3485:   /* distribute row lengths to all processors */
3486:   PetscMalloc1(m,&locrowlens);
3487:   if (!rank) {
3488:     mend = m;
3489:     if (size == 1) mend = mend - extra_rows;
3490:     PetscBinaryRead(fd,locrowlens,mend,PETSC_INT);
3491:     for (j=mend; j<m; j++) locrowlens[j] = 1;
3492:     PetscMalloc1(mmax,&rowlengths);
3493:     PetscCalloc1(size,&procsnz);
3494:     for (j=0; j<m; j++) {
3495:       procsnz[0] += locrowlens[j];
3496:     }
3497:     for (i=1; i<size; i++) {
3498:       mend = browners[i+1] - browners[i];
3499:       if (i == size-1) mend = mend - extra_rows;
3500:       PetscBinaryRead(fd,rowlengths,mend,PETSC_INT);
3501:       for (j=mend; j<browners[i+1] - browners[i]; j++) rowlengths[j] = 1;
3502:       /* calculate the number of nonzeros on each processor */
3503:       for (j=0; j<browners[i+1]-browners[i]; j++) {
3504:         procsnz[i] += rowlengths[j];
3505:       }
3506:       MPI_Send(rowlengths,browners[i+1]-browners[i],MPIU_INT,i,tag,comm);
3507:     }
3508:     PetscFree(rowlengths);
3509:   } else {
3510:     MPI_Recv(locrowlens,m,MPIU_INT,0,tag,comm,&status);
3511:   }

3513:   if (!rank) {
3514:     /* determine max buffer needed and allocate it */
3515:     maxnz = procsnz[0];
3516:     for (i=1; i<size; i++) {
3517:       maxnz = PetscMax(maxnz,procsnz[i]);
3518:     }
3519:     PetscMalloc1(maxnz,&cols);

3521:     /* read in my part of the matrix column indices  */
3522:     nz     = procsnz[0];
3523:     PetscMalloc1((nz+1),&ibuf);
3524:     mycols = ibuf;
3525:     if (size == 1) nz -= extra_rows;
3526:     PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3527:     if (size == 1) {
3528:       for (i=0; i< extra_rows; i++) mycols[nz+i] = M+i;
3529:     }

3531:     /* read in every ones (except the last) and ship off */
3532:     for (i=1; i<size-1; i++) {
3533:       nz   = procsnz[i];
3534:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
3535:       MPI_Send(cols,nz,MPIU_INT,i,tag,comm);
3536:     }
3537:     /* read in the stuff for the last proc */
3538:     if (size != 1) {
3539:       nz   = procsnz[size-1] - extra_rows;  /* the extra rows are not on the disk */
3540:       PetscBinaryRead(fd,cols,nz,PETSC_INT);
3541:       for (i=0; i<extra_rows; i++) cols[nz+i] = M+i;
3542:       MPI_Send(cols,nz+extra_rows,MPIU_INT,size-1,tag,comm);
3543:     }
3544:     PetscFree(cols);
3545:   } else {
3546:     /* determine buffer space needed for message */
3547:     nz = 0;
3548:     for (i=0; i<m; i++) {
3549:       nz += locrowlens[i];
3550:     }
3551:     PetscMalloc1((nz+1),&ibuf);
3552:     mycols = ibuf;
3553:     /* receive message of column indices*/
3554:     MPI_Recv(mycols,nz,MPIU_INT,0,tag,comm,&status);
3555:     MPI_Get_count(&status,MPIU_INT,&maxnz);
3556:     if (maxnz != nz) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"something is wrong with file");
3557:   }

3559:   /* loop over local rows, determining number of off diagonal entries */
3560:   PetscMalloc2(rend-rstart,&dlens,rend-rstart,&odlens);
3561:   PetscCalloc3(Mbs,&mask,Mbs,&masked1,Mbs,&masked2);
3562:   rowcount = 0; nzcount = 0;
3563:   for (i=0; i<mbs; i++) {
3564:     dcount  = 0;
3565:     odcount = 0;
3566:     for (j=0; j<bs; j++) {
3567:       kmax = locrowlens[rowcount];
3568:       for (k=0; k<kmax; k++) {
3569:         tmp = mycols[nzcount++]/bs;
3570:         if (!mask[tmp]) {
3571:           mask[tmp] = 1;
3572:           if (tmp < rstart || tmp >= rend) masked2[odcount++] = tmp;
3573:           else masked1[dcount++] = tmp;
3574:         }
3575:       }
3576:       rowcount++;
3577:     }

3579:     dlens[i]  = dcount;
3580:     odlens[i] = odcount;

3582:     /* zero out the mask elements we set */
3583:     for (j=0; j<dcount; j++) mask[masked1[j]] = 0;
3584:     for (j=0; j<odcount; j++) mask[masked2[j]] = 0;
3585:   }


3588:   if (!sizesset) {
3589:     MatSetSizes(newmat,m,m,M+extra_rows,N+extra_rows);
3590:   }
3591:   MatMPIBAIJSetPreallocation(newmat,bs,0,dlens,0,odlens);

3593:   if (!rank) {
3594:     PetscMalloc1((maxnz+1),&buf);
3595:     /* read in my part of the matrix numerical values  */
3596:     nz     = procsnz[0];
3597:     vals   = buf;
3598:     mycols = ibuf;
3599:     if (size == 1) nz -= extra_rows;
3600:     PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3601:     if (size == 1) {
3602:       for (i=0; i< extra_rows; i++) vals[nz+i] = 1.0;
3603:     }

3605:     /* insert into matrix */
3606:     jj = rstart*bs;
3607:     for (i=0; i<m; i++) {
3608:       MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3609:       mycols += locrowlens[i];
3610:       vals   += locrowlens[i];
3611:       jj++;
3612:     }
3613:     /* read in other processors (except the last one) and ship out */
3614:     for (i=1; i<size-1; i++) {
3615:       nz   = procsnz[i];
3616:       vals = buf;
3617:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3618:       MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newmat)->tag,comm);
3619:     }
3620:     /* the last proc */
3621:     if (size != 1) {
3622:       nz   = procsnz[i] - extra_rows;
3623:       vals = buf;
3624:       PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3625:       for (i=0; i<extra_rows; i++) vals[nz+i] = 1.0;
3626:       MPIULong_Send(vals,nz+extra_rows,MPIU_SCALAR,size-1,((PetscObject)newmat)->tag,comm);
3627:     }
3628:     PetscFree(procsnz);
3629:   } else {
3630:     /* receive numeric values */
3631:     PetscMalloc1((nz+1),&buf);

3633:     /* receive message of values*/
3634:     vals   = buf;
3635:     mycols = ibuf;
3636:     MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newmat)->tag,comm);

3638:     /* insert into matrix */
3639:     jj = rstart*bs;
3640:     for (i=0; i<m; i++) {
3641:       MatSetValues_MPIBAIJ(newmat,1,&jj,locrowlens[i],mycols,vals,INSERT_VALUES);
3642:       mycols += locrowlens[i];
3643:       vals   += locrowlens[i];
3644:       jj++;
3645:     }
3646:   }
3647:   PetscFree(locrowlens);
3648:   PetscFree(buf);
3649:   PetscFree(ibuf);
3650:   PetscFree2(rowners,browners);
3651:   PetscFree2(dlens,odlens);
3652:   PetscFree3(mask,masked1,masked2);
3653:   MatAssemblyBegin(newmat,MAT_FINAL_ASSEMBLY);
3654:   MatAssemblyEnd(newmat,MAT_FINAL_ASSEMBLY);
3655:   return(0);
3656: }

3660: /*@
3661:    MatMPIBAIJSetHashTableFactor - Sets the factor required to compute the size of the HashTable.

3663:    Input Parameters:
3664: .  mat  - the matrix
3665: .  fact - factor

3667:    Not Collective, each process can use a different factor

3669:    Level: advanced

3671:   Notes:
3672:    This can also be set by the command line option: -mat_use_hash_table <fact>

3674: .keywords: matrix, hashtable, factor, HT

3676: .seealso: MatSetOption()
3677: @*/
3678: PetscErrorCode  MatMPIBAIJSetHashTableFactor(Mat mat,PetscReal fact)
3679: {

3683:   PetscTryMethod(mat,"MatSetHashTableFactor_C",(Mat,PetscReal),(mat,fact));
3684:   return(0);
3685: }

3689: PetscErrorCode  MatSetHashTableFactor_MPIBAIJ(Mat mat,PetscReal fact)
3690: {
3691:   Mat_MPIBAIJ *baij;

3694:   baij          = (Mat_MPIBAIJ*)mat->data;
3695:   baij->ht_fact = fact;
3696:   return(0);
3697: }

3701: PetscErrorCode  MatMPIBAIJGetSeqBAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
3702: {
3703:   Mat_MPIBAIJ *a = (Mat_MPIBAIJ*)A->data;

3706:   if (Ad)     *Ad     = a->A;
3707:   if (Ao)     *Ao     = a->B;
3708:   if (colmap) *colmap = a->garray;
3709:   return(0);
3710: }

3712: /*
3713:     Special version for direct calls from Fortran (to eliminate two function call overheads
3714: */
3715: #if defined(PETSC_HAVE_FORTRAN_CAPS)
3716: #define matmpibaijsetvaluesblocked_ MATMPIBAIJSETVALUESBLOCKED
3717: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
3718: #define matmpibaijsetvaluesblocked_ matmpibaijsetvaluesblocked
3719: #endif

3723: /*@C
3724:   MatMPIBAIJSetValuesBlocked - Direct Fortran call to replace call to MatSetValuesBlocked()

3726:   Collective on Mat

3728:   Input Parameters:
3729: + mat - the matrix
3730: . min - number of input rows
3731: . im - input rows
3732: . nin - number of input columns
3733: . in - input columns
3734: . v - numerical values input
3735: - addvin - INSERT_VALUES or ADD_VALUES

3737:   Notes: This has a complete copy of MatSetValuesBlocked_MPIBAIJ() which is terrible code un-reuse.

3739:   Level: advanced

3741: .seealso:   MatSetValuesBlocked()
3742: @*/
3743: PetscErrorCode matmpibaijsetvaluesblocked_(Mat *matin,PetscInt *min,const PetscInt im[],PetscInt *nin,const PetscInt in[],const MatScalar v[],InsertMode *addvin)
3744: {
3745:   /* convert input arguments to C version */
3746:   Mat        mat  = *matin;
3747:   PetscInt   m    = *min, n = *nin;
3748:   InsertMode addv = *addvin;

3750:   Mat_MPIBAIJ     *baij = (Mat_MPIBAIJ*)mat->data;
3751:   const MatScalar *value;
3752:   MatScalar       *barray     = baij->barray;
3753:   PetscBool       roworiented = baij->roworiented;
3754:   PetscErrorCode  ierr;
3755:   PetscInt        i,j,ii,jj,row,col,rstart=baij->rstartbs;
3756:   PetscInt        rend=baij->rendbs,cstart=baij->cstartbs,stepval;
3757:   PetscInt        cend=baij->cendbs,bs=mat->rmap->bs,bs2=baij->bs2;

3760:   /* tasks normally handled by MatSetValuesBlocked() */
3761:   if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
3762: #if defined(PETSC_USE_DEBUG)
3763:   else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
3764:   if (mat->factortype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Not for factored matrix");
3765: #endif
3766:   if (mat->assembled) {
3767:     mat->was_assembled = PETSC_TRUE;
3768:     mat->assembled     = PETSC_FALSE;
3769:   }
3770:   PetscLogEventBegin(MAT_SetValues,mat,0,0,0);


3773:   if (!barray) {
3774:     PetscMalloc1(bs2,&barray);
3775:     baij->barray = barray;
3776:   }

3778:   if (roworiented) stepval = (n-1)*bs;
3779:   else stepval = (m-1)*bs;

3781:   for (i=0; i<m; i++) {
3782:     if (im[i] < 0) continue;
3783: #if defined(PETSC_USE_DEBUG)
3784:     if (im[i] >= baij->Mbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large, row %D max %D",im[i],baij->Mbs-1);
3785: #endif
3786:     if (im[i] >= rstart && im[i] < rend) {
3787:       row = im[i] - rstart;
3788:       for (j=0; j<n; j++) {
3789:         /* If NumCol = 1 then a copy is not required */
3790:         if ((roworiented) && (n == 1)) {
3791:           barray = (MatScalar*)v + i*bs2;
3792:         } else if ((!roworiented) && (m == 1)) {
3793:           barray = (MatScalar*)v + j*bs2;
3794:         } else { /* Here a copy is required */
3795:           if (roworiented) {
3796:             value = v + i*(stepval+bs)*bs + j*bs;
3797:           } else {
3798:             value = v + j*(stepval+bs)*bs + i*bs;
3799:           }
3800:           for (ii=0; ii<bs; ii++,value+=stepval) {
3801:             for (jj=0; jj<bs; jj++) {
3802:               *barray++ = *value++;
3803:             }
3804:           }
3805:           barray -=bs2;
3806:         }

3808:         if (in[j] >= cstart && in[j] < cend) {
3809:           col  = in[j] - cstart;
3810:           MatSetValuesBlocked_SeqBAIJ(baij->A,1,&row,1,&col,barray,addv);
3811:         } else if (in[j] < 0) continue;
3812: #if defined(PETSC_USE_DEBUG)
3813:         else if (in[j] >= baij->Nbs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large, col %D max %D",in[j],baij->Nbs-1);
3814: #endif
3815:         else {
3816:           if (mat->was_assembled) {
3817:             if (!baij->colmap) {
3818:               MatCreateColmap_MPIBAIJ_Private(mat);
3819:             }

3821: #if defined(PETSC_USE_DEBUG)
3822: #if defined(PETSC_USE_CTABLE)
3823:             { PetscInt data;
3824:               PetscTableFind(baij->colmap,in[j]+1,&data);
3825:               if ((data - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3826:             }
3827: #else
3828:             if ((baij->colmap[in[j]] - 1) % bs) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect colmap");
3829: #endif
3830: #endif
3831: #if defined(PETSC_USE_CTABLE)
3832:             PetscTableFind(baij->colmap,in[j]+1,&col);
3833:             col  = (col - 1)/bs;
3834: #else
3835:             col = (baij->colmap[in[j]] - 1)/bs;
3836: #endif
3837:             if (col < 0 && !((Mat_SeqBAIJ*)(baij->A->data))->nonew) {
3838:               MatDisAssemble_MPIBAIJ(mat);
3839:               col  =  in[j];
3840:             }
3841:           } else col = in[j];
3842:           MatSetValuesBlocked_SeqBAIJ(baij->B,1,&row,1,&col,barray,addv);
3843:         }
3844:       }
3845:     } else {
3846:       if (!baij->donotstash) {
3847:         if (roworiented) {
3848:           MatStashValuesRowBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
3849:         } else {
3850:           MatStashValuesColBlocked_Private(&mat->bstash,im[i],n,in,v,m,n,i);
3851:         }
3852:       }
3853:     }
3854:   }

3856:   /* task normally handled by MatSetValuesBlocked() */
3857:   PetscLogEventEnd(MAT_SetValues,mat,0,0,0);
3858:   return(0);
3859: }

3863: /*@
3864:      MatCreateMPIBAIJWithArrays - creates a MPI BAIJ matrix using arrays that contain in standard
3865:          CSR format the local rows.

3867:    Collective on MPI_Comm

3869:    Input Parameters:
3870: +  comm - MPI communicator
3871: .  bs - the block size, only a block size of 1 is supported
3872: .  m - number of local rows (Cannot be PETSC_DECIDE)
3873: .  n - This value should be the same as the local size used in creating the
3874:        x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
3875:        calculated if N is given) For square matrices n is almost always m.
3876: .  M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
3877: .  N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
3878: .   i - row indices
3879: .   j - column indices
3880: -   a - matrix values

3882:    Output Parameter:
3883: .   mat - the matrix

3885:    Level: intermediate

3887:    Notes:
3888:        The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
3889:      thus you CANNOT change the matrix entries by changing the values of a[] after you have
3890:      called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.

3892:      The order of the entries in values is the same as the block compressed sparse row storage format; that is, it is
3893:      the same as a three dimensional array in Fortran values(bs,bs,nnz) that contains the first column of the first
3894:      block, followed by the second column of the first block etc etc.  That is, the blocks are contiguous in memory
3895:      with column-major ordering within blocks.

3897:        The i and j indices are 0 based, and i indices are indices corresponding to the local j array.

3899: .keywords: matrix, aij, compressed row, sparse, parallel

3901: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
3902:           MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
3903: @*/
3904: PetscErrorCode  MatCreateMPIBAIJWithArrays(MPI_Comm comm,PetscInt bs,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
3905: {

3909:   if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
3910:   if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
3911:   MatCreate(comm,mat);
3912:   MatSetSizes(*mat,m,n,M,N);
3913:   MatSetType(*mat,MATMPISBAIJ);
3914:   MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_FALSE);
3915:   MatMPIBAIJSetPreallocationCSR(*mat,bs,i,j,a);
3916:   MatSetOption(*mat,MAT_ROW_ORIENTED,PETSC_TRUE);
3917:   return(0);
3918: }