Actual source code: matmatmult.c

petsc-3.15.0 2021-03-30
Report Typos and Errors

  2: /*
  3:   Defines matrix-matrix product routines for pairs of SeqAIJ matrices
  4:           C = A * B
  5: */

  7: #include <../src/mat/impls/aij/seq/aij.h>
  8: #include <../src/mat/utils/freespace.h>
  9: #include <petscbt.h>
 10: #include <petsc/private/isimpl.h>
 11: #include <../src/mat/impls/dense/seq/dense.h>

 13: PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C)
 14: {

 18:   if (C->ops->matmultnumeric) {
 19:     if (C->ops->matmultnumeric == MatMatMultNumeric_SeqAIJ_SeqAIJ) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Recursive call");
 20:     (*C->ops->matmultnumeric)(A,B,C);
 21:   } else {
 22:     MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(A,B,C);
 23:   }
 24:   return(0);
 25: }

 27: /* Modified from MatCreateSeqAIJWithArrays() */
 28: PETSC_INTERN PetscErrorCode MatSetSeqAIJWithArrays_private(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt i[],PetscInt j[],PetscScalar a[],MatType mtype,Mat mat)
 29: {
 31:   PetscInt       ii;
 32:   Mat_SeqAIJ     *aij;
 33:   PetscBool      isseqaij;

 36:   if (m > 0 && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
 37:   MatSetSizes(mat,m,n,m,n);

 39:   if (!mtype) {
 40:     PetscObjectBaseTypeCompare((PetscObject)mat,MATSEQAIJ,&isseqaij);
 41:     if (!isseqaij) { MatSetType(mat,MATSEQAIJ); }
 42:   } else {
 43:     MatSetType(mat,mtype);
 44:   }
 45:   MatSeqAIJSetPreallocation_SeqAIJ(mat,MAT_SKIP_ALLOCATION,NULL);
 46:   aij  = (Mat_SeqAIJ*)(mat)->data;
 47:   PetscMalloc1(m,&aij->imax);
 48:   PetscMalloc1(m,&aij->ilen);

 50:   aij->i            = i;
 51:   aij->j            = j;
 52:   aij->a            = a;
 53:   aij->singlemalloc = PETSC_FALSE;
 54:   aij->nonew        = -1; /*this indicates that inserting a new value in the matrix that generates a new nonzero is an error*/
 55:   aij->free_a       = PETSC_FALSE;
 56:   aij->free_ij      = PETSC_FALSE;

 58:   for (ii=0; ii<m; ii++) {
 59:     aij->ilen[ii] = aij->imax[ii] = i[ii+1] - i[ii];
 60:   }

 62:   return(0);
 63: }

 65: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C)
 66: {
 67:   PetscErrorCode      ierr;
 68:   Mat_Product         *product = C->product;
 69:   MatProductAlgorithm alg;
 70:   PetscBool           flg;

 73:   if (product) {
 74:     alg = product->alg;
 75:   } else {
 76:     alg = "sorted";
 77:   }
 78:   /* sorted */
 79:   PetscStrcmp(alg,"sorted",&flg);
 80:   if (flg) {
 81:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(A,B,fill,C);
 82:     return(0);
 83:   }

 85:   /* scalable */
 86:   PetscStrcmp(alg,"scalable",&flg);
 87:   if (flg) {
 88:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(A,B,fill,C);
 89:     return(0);
 90:   }

 92:   /* scalable_fast */
 93:   PetscStrcmp(alg,"scalable_fast",&flg);
 94:   if (flg) {
 95:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(A,B,fill,C);
 96:     return(0);
 97:   }

 99:   /* heap */
100:   PetscStrcmp(alg,"heap",&flg);
101:   if (flg) {
102:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(A,B,fill,C);
103:     return(0);
104:   }

106:   /* btheap */
107:   PetscStrcmp(alg,"btheap",&flg);
108:   if (flg) {
109:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(A,B,fill,C);
110:     return(0);
111:   }

113:   /* llcondensed */
114:   PetscStrcmp(alg,"llcondensed",&flg);
115:   if (flg) {
116:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(A,B,fill,C);
117:     return(0);
118:   }

120:   /* rowmerge */
121:   PetscStrcmp(alg,"rowmerge",&flg);
122:   if (flg) {
123:     MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(A,B,fill,C);
124:     return(0);
125:   }

127: #if defined(PETSC_HAVE_HYPRE)
128:   PetscStrcmp(alg,"hypre",&flg);
129:   if (flg) {
130:     MatMatMultSymbolic_AIJ_AIJ_wHYPRE(A,B,fill,C);
131:     return(0);
132:   }
133: #endif

135:   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Mat Product Algorithm is not supported");
136: }

138: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_LLCondensed(Mat A,Mat B,PetscReal fill,Mat C)
139: {
140:   PetscErrorCode     ierr;
141:   Mat_SeqAIJ         *a =(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
142:   PetscInt           *ai=a->i,*bi=b->i,*ci,*cj;
143:   PetscInt           am =A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
144:   PetscReal          afill;
145:   PetscInt           i,j,anzi,brow,bnzj,cnzi,*bj,*aj,*lnk,ndouble=0,Crmax;
146:   PetscTable         ta;
147:   PetscBT            lnkbt;
148:   PetscFreeSpaceList free_space=NULL,current_space=NULL;

151:   /* Get ci and cj */
152:   /*---------------*/
153:   /* Allocate ci array, arrays for fill computation and */
154:   /* free space for accumulating nonzero column info */
155:   PetscMalloc1(am+2,&ci);
156:   ci[0] = 0;

158:   /* create and initialize a linked list */
159:   PetscTableCreate(bn,bn,&ta);
160:   MatRowMergeMax_SeqAIJ(b,bm,ta);
161:   PetscTableGetCount(ta,&Crmax);
162:   PetscTableDestroy(&ta);

164:   PetscLLCondensedCreate(Crmax,bn,&lnk,&lnkbt);

166:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
167:   PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],bi[bm])),&free_space);

169:   current_space = free_space;

171:   /* Determine ci and cj */
172:   for (i=0; i<am; i++) {
173:     anzi = ai[i+1] - ai[i];
174:     aj   = a->j + ai[i];
175:     for (j=0; j<anzi; j++) {
176:       brow = aj[j];
177:       bnzj = bi[brow+1] - bi[brow];
178:       bj   = b->j + bi[brow];
179:       /* add non-zero cols of B into the sorted linked list lnk */
180:       PetscLLCondensedAddSorted(bnzj,bj,lnk,lnkbt);
181:     }
182:     /* add possible missing diagonal entry */
183:     if (C->force_diagonals) {
184:       PetscLLCondensedAddSorted(1,&i,lnk,lnkbt);
185:     }
186:     cnzi = lnk[0];

188:     /* If free space is not available, make more free space */
189:     /* Double the amount of total space in the list */
190:     if (current_space->local_remaining<cnzi) {
191:       PetscFreeSpaceGet(PetscIntSumTruncate(cnzi,current_space->total_array_size),&current_space);
192:       ndouble++;
193:     }

195:     /* Copy data into free space, then initialize lnk */
196:     PetscLLCondensedClean(bn,cnzi,current_space->array,lnk,lnkbt);

198:     current_space->array           += cnzi;
199:     current_space->local_used      += cnzi;
200:     current_space->local_remaining -= cnzi;

202:     ci[i+1] = ci[i] + cnzi;
203:   }

205:   /* Column indices are in the list of free space */
206:   /* Allocate space for cj, initialize cj, and */
207:   /* destroy list of free space and other temporary array(s) */
208:   PetscMalloc1(ci[am]+1,&cj);
209:   PetscFreeSpaceContiguous(&free_space,cj);
210:   PetscLLCondensedDestroy(lnk,lnkbt);

212:   /* put together the new symbolic matrix */
213:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,((PetscObject)A)->type_name,C);
214:   MatSetBlockSizesFromMats(C,A,B);

216:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
217:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
218:   c          = (Mat_SeqAIJ*)(C->data);
219:   c->free_a  = PETSC_FALSE;
220:   c->free_ij = PETSC_TRUE;
221:   c->nonew   = 0;

223:   /* fast, needs non-scalable O(bn) array 'abdense' */
224:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted;

226:   /* set MatInfo */
227:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
228:   if (afill < 1.0) afill = 1.0;
229:   c->maxnz                  = ci[am];
230:   c->nz                     = ci[am];
231:   C->info.mallocs           = ndouble;
232:   C->info.fill_ratio_given  = fill;
233:   C->info.fill_ratio_needed = afill;

235: #if defined(PETSC_USE_INFO)
236:   if (ci[am]) {
237:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
238:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
239:   } else {
240:     PetscInfo(C,"Empty matrix product\n");
241:   }
242: #endif
243:   return(0);
244: }

246: PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted(Mat A,Mat B,Mat C)
247: {
248:   PetscErrorCode    ierr;
249:   PetscLogDouble    flops=0.0;
250:   Mat_SeqAIJ        *a   = (Mat_SeqAIJ*)A->data;
251:   Mat_SeqAIJ        *b   = (Mat_SeqAIJ*)B->data;
252:   Mat_SeqAIJ        *c   = (Mat_SeqAIJ*)C->data;
253:   PetscInt          *ai  =a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj,*ci=c->i,*cj=c->j;
254:   PetscInt          am   =A->rmap->n,cm=C->rmap->n;
255:   PetscInt          i,j,k,anzi,bnzi,cnzi,brow;
256:   PetscScalar       *ca,valtmp;
257:   PetscScalar       *ab_dense;
258:   PetscContainer    cab_dense;
259:   const PetscScalar *aa,*ba,*baj;

262:   MatSeqAIJGetArrayRead(A,&aa);
263:   MatSeqAIJGetArrayRead(B,&ba);
264:   if (!c->a) { /* first call of MatMatMultNumeric_SeqAIJ_SeqAIJ, allocate ca and matmult_abdense */
265:     PetscMalloc1(ci[cm]+1,&ca);
266:     c->a      = ca;
267:     c->free_a = PETSC_TRUE;
268:   } else ca = c->a;

270:   /* TODO this should be done in the symbolic phase */
271:   /* However, this function is so heavily used (sometimes in an hidden way through multnumeric function pointers
272:      that is hard to eradicate) */
273:   PetscObjectQuery((PetscObject)C,"__PETSc__ab_dense",(PetscObject*)&cab_dense);
274:   if (!cab_dense) {
275:     PetscMalloc1(B->cmap->N,&ab_dense);
276:     PetscContainerCreate(PETSC_COMM_SELF,&cab_dense);
277:     PetscContainerSetPointer(cab_dense,ab_dense);
278:     PetscContainerSetUserDestroy(cab_dense,PetscContainerUserDestroyDefault);
279:     PetscObjectCompose((PetscObject)C,"__PETSc__ab_dense",(PetscObject)cab_dense);
280:     PetscObjectDereference((PetscObject)cab_dense);
281:   }
282:   PetscContainerGetPointer(cab_dense,(void**)&ab_dense);
283:   PetscArrayzero(ab_dense,B->cmap->N);

285:   /* clean old values in C */
286:   PetscArrayzero(ca,ci[cm]);
287:   /* Traverse A row-wise. */
288:   /* Build the ith row in C by summing over nonzero columns in A, */
289:   /* the rows of B corresponding to nonzeros of A. */
290:   for (i=0; i<am; i++) {
291:     anzi = ai[i+1] - ai[i];
292:     for (j=0; j<anzi; j++) {
293:       brow = aj[j];
294:       bnzi = bi[brow+1] - bi[brow];
295:       bjj  = bj + bi[brow];
296:       baj  = ba + bi[brow];
297:       /* perform dense axpy */
298:       valtmp = aa[j];
299:       for (k=0; k<bnzi; k++) {
300:         ab_dense[bjj[k]] += valtmp*baj[k];
301:       }
302:       flops += 2*bnzi;
303:     }
304:     aj += anzi; aa += anzi;

306:     cnzi = ci[i+1] - ci[i];
307:     for (k=0; k<cnzi; k++) {
308:       ca[k]          += ab_dense[cj[k]];
309:       ab_dense[cj[k]] = 0.0; /* zero ab_dense */
310:     }
311:     flops += cnzi;
312:     cj    += cnzi; ca += cnzi;
313:   }
314: #if defined(PETSC_HAVE_DEVICE)
315:   if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU;
316: #endif
317:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
318:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
319:   PetscLogFlops(flops);
320:   MatSeqAIJRestoreArrayRead(A,&aa);
321:   MatSeqAIJRestoreArrayRead(B,&ba);
322:   return(0);
323: }

325: PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable(Mat A,Mat B,Mat C)
326: {
327:   PetscErrorCode    ierr;
328:   PetscLogDouble    flops=0.0;
329:   Mat_SeqAIJ        *a   = (Mat_SeqAIJ*)A->data;
330:   Mat_SeqAIJ        *b   = (Mat_SeqAIJ*)B->data;
331:   Mat_SeqAIJ        *c   = (Mat_SeqAIJ*)C->data;
332:   PetscInt          *ai  = a->i,*aj=a->j,*bi=b->i,*bj=b->j,*bjj,*ci=c->i,*cj=c->j;
333:   PetscInt          am   = A->rmap->N,cm=C->rmap->N;
334:   PetscInt          i,j,k,anzi,bnzi,cnzi,brow;
335:   PetscScalar       *ca=c->a,valtmp;
336:   const PetscScalar *aa,*ba,*baj;
337:   PetscInt          nextb;

340:   MatSeqAIJGetArrayRead(A,&aa);
341:   MatSeqAIJGetArrayRead(B,&ba);
342:   if (!ca) { /* first call of MatMatMultNumeric_SeqAIJ_SeqAIJ, allocate ca and matmult_abdense */
343:     PetscMalloc1(ci[cm]+1,&ca);
344:     c->a      = ca;
345:     c->free_a = PETSC_TRUE;
346:   }

348:   /* clean old values in C */
349:   PetscArrayzero(ca,ci[cm]);
350:   /* Traverse A row-wise. */
351:   /* Build the ith row in C by summing over nonzero columns in A, */
352:   /* the rows of B corresponding to nonzeros of A. */
353:   for (i=0; i<am; i++) {
354:     anzi = ai[i+1] - ai[i];
355:     cnzi = ci[i+1] - ci[i];
356:     for (j=0; j<anzi; j++) {
357:       brow = aj[j];
358:       bnzi = bi[brow+1] - bi[brow];
359:       bjj  = bj + bi[brow];
360:       baj  = ba + bi[brow];
361:       /* perform sparse axpy */
362:       valtmp = aa[j];
363:       nextb  = 0;
364:       for (k=0; nextb<bnzi; k++) {
365:         if (cj[k] == bjj[nextb]) { /* ccol == bcol */
366:           ca[k] += valtmp*baj[nextb++];
367:         }
368:       }
369:       flops += 2*bnzi;
370:     }
371:     aj += anzi; aa += anzi;
372:     cj += cnzi; ca += cnzi;
373:   }
374: #if defined(PETSC_HAVE_DEVICE)
375:   if (C->offloadmask != PETSC_OFFLOAD_UNALLOCATED) C->offloadmask = PETSC_OFFLOAD_CPU;
376: #endif
377:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
378:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
379:   PetscLogFlops(flops);
380:   MatSeqAIJRestoreArrayRead(A,&aa);
381:   MatSeqAIJRestoreArrayRead(B,&ba);
382:   return(0);
383: }

385: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable_fast(Mat A,Mat B,PetscReal fill,Mat C)
386: {
387:   PetscErrorCode     ierr;
388:   Mat_SeqAIJ         *a  = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
389:   PetscInt           *ai = a->i,*bi=b->i,*ci,*cj;
390:   PetscInt           am  = A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
391:   MatScalar          *ca;
392:   PetscReal          afill;
393:   PetscInt           i,j,anzi,brow,bnzj,cnzi,*bj,*aj,*lnk,ndouble=0,Crmax;
394:   PetscTable         ta;
395:   PetscFreeSpaceList free_space=NULL,current_space=NULL;

398:   /* Get ci and cj - same as MatMatMultSymbolic_SeqAIJ_SeqAIJ except using PetscLLxxx_fast() */
399:   /*-----------------------------------------------------------------------------------------*/
400:   /* Allocate arrays for fill computation and free space for accumulating nonzero column */
401:   PetscMalloc1(am+2,&ci);
402:   ci[0] = 0;

404:   /* create and initialize a linked list */
405:   PetscTableCreate(bn,bn,&ta);
406:   MatRowMergeMax_SeqAIJ(b,bm,ta);
407:   PetscTableGetCount(ta,&Crmax);
408:   PetscTableDestroy(&ta);

410:   PetscLLCondensedCreate_fast(Crmax,&lnk);

412:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
413:   PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],bi[bm])),&free_space);
414:   current_space = free_space;

416:   /* Determine ci and cj */
417:   for (i=0; i<am; i++) {
418:     anzi = ai[i+1] - ai[i];
419:     aj   = a->j + ai[i];
420:     for (j=0; j<anzi; j++) {
421:       brow = aj[j];
422:       bnzj = bi[brow+1] - bi[brow];
423:       bj   = b->j + bi[brow];
424:       /* add non-zero cols of B into the sorted linked list lnk */
425:       PetscLLCondensedAddSorted_fast(bnzj,bj,lnk);
426:     }
427:     /* add possible missing diagonal entry */
428:     if (C->force_diagonals) {
429:       PetscLLCondensedAddSorted_fast(1,&i,lnk);
430:     }
431:     cnzi = lnk[1];

433:     /* If free space is not available, make more free space */
434:     /* Double the amount of total space in the list */
435:     if (current_space->local_remaining<cnzi) {
436:       PetscFreeSpaceGet(PetscIntSumTruncate(cnzi,current_space->total_array_size),&current_space);
437:       ndouble++;
438:     }

440:     /* Copy data into free space, then initialize lnk */
441:     PetscLLCondensedClean_fast(cnzi,current_space->array,lnk);

443:     current_space->array           += cnzi;
444:     current_space->local_used      += cnzi;
445:     current_space->local_remaining -= cnzi;

447:     ci[i+1] = ci[i] + cnzi;
448:   }

450:   /* Column indices are in the list of free space */
451:   /* Allocate space for cj, initialize cj, and */
452:   /* destroy list of free space and other temporary array(s) */
453:   PetscMalloc1(ci[am]+1,&cj);
454:   PetscFreeSpaceContiguous(&free_space,cj);
455:   PetscLLCondensedDestroy_fast(lnk);

457:   /* Allocate space for ca */
458:   PetscCalloc1(ci[am]+1,&ca);

460:   /* put together the new symbolic matrix */
461:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,ca,((PetscObject)A)->type_name,C);
462:   MatSetBlockSizesFromMats(C,A,B);

464:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
465:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
466:   c          = (Mat_SeqAIJ*)(C->data);
467:   c->free_a  = PETSC_TRUE;
468:   c->free_ij = PETSC_TRUE;
469:   c->nonew   = 0;

471:   /* slower, less memory */
472:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable;

474:   /* set MatInfo */
475:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
476:   if (afill < 1.0) afill = 1.0;
477:   c->maxnz                     = ci[am];
478:   c->nz                        = ci[am];
479:   C->info.mallocs           = ndouble;
480:   C->info.fill_ratio_given  = fill;
481:   C->info.fill_ratio_needed = afill;

483: #if defined(PETSC_USE_INFO)
484:   if (ci[am]) {
485:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
486:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
487:   } else {
488:     PetscInfo(C,"Empty matrix product\n");
489:   }
490: #endif
491:   return(0);
492: }

494: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Scalable(Mat A,Mat B,PetscReal fill,Mat C)
495: {
496:   PetscErrorCode     ierr;
497:   Mat_SeqAIJ         *a  = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
498:   PetscInt           *ai = a->i,*bi=b->i,*ci,*cj;
499:   PetscInt           am  = A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
500:   MatScalar          *ca;
501:   PetscReal          afill;
502:   PetscInt           i,j,anzi,brow,bnzj,cnzi,*bj,*aj,*lnk,ndouble=0,Crmax;
503:   PetscTable         ta;
504:   PetscFreeSpaceList free_space=NULL,current_space=NULL;

507:   /* Get ci and cj - same as MatMatMultSymbolic_SeqAIJ_SeqAIJ except using PetscLLxxx_Scalalbe() */
508:   /*---------------------------------------------------------------------------------------------*/
509:   /* Allocate arrays for fill computation and free space for accumulating nonzero column */
510:   PetscMalloc1(am+2,&ci);
511:   ci[0] = 0;

513:   /* create and initialize a linked list */
514:   PetscTableCreate(bn,bn,&ta);
515:   MatRowMergeMax_SeqAIJ(b,bm,ta);
516:   PetscTableGetCount(ta,&Crmax);
517:   PetscTableDestroy(&ta);
518:   PetscLLCondensedCreate_Scalable(Crmax,&lnk);

520:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
521:   PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],bi[bm])),&free_space);
522:   current_space = free_space;

524:   /* Determine ci and cj */
525:   for (i=0; i<am; i++) {
526:     anzi = ai[i+1] - ai[i];
527:     aj   = a->j + ai[i];
528:     for (j=0; j<anzi; j++) {
529:       brow = aj[j];
530:       bnzj = bi[brow+1] - bi[brow];
531:       bj   = b->j + bi[brow];
532:       /* add non-zero cols of B into the sorted linked list lnk */
533:       PetscLLCondensedAddSorted_Scalable(bnzj,bj,lnk);
534:     }
535:     /* add possible missing diagonal entry */
536:     if (C->force_diagonals) {
537:       PetscLLCondensedAddSorted_Scalable(1,&i,lnk);
538:     }

540:     cnzi = lnk[0];

542:     /* If free space is not available, make more free space */
543:     /* Double the amount of total space in the list */
544:     if (current_space->local_remaining<cnzi) {
545:       PetscFreeSpaceGet(PetscIntSumTruncate(cnzi,current_space->total_array_size),&current_space);
546:       ndouble++;
547:     }

549:     /* Copy data into free space, then initialize lnk */
550:     PetscLLCondensedClean_Scalable(cnzi,current_space->array,lnk);

552:     current_space->array           += cnzi;
553:     current_space->local_used      += cnzi;
554:     current_space->local_remaining -= cnzi;

556:     ci[i+1] = ci[i] + cnzi;
557:   }

559:   /* Column indices are in the list of free space */
560:   /* Allocate space for cj, initialize cj, and */
561:   /* destroy list of free space and other temporary array(s) */
562:   PetscMalloc1(ci[am]+1,&cj);
563:   PetscFreeSpaceContiguous(&free_space,cj);
564:   PetscLLCondensedDestroy_Scalable(lnk);

566:   /* Allocate space for ca */
567:   /*-----------------------*/
568:   PetscCalloc1(ci[am]+1,&ca);

570:   /* put together the new symbolic matrix */
571:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,ca,((PetscObject)A)->type_name,C);
572:   MatSetBlockSizesFromMats(C,A,B);

574:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
575:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
576:   c          = (Mat_SeqAIJ*)(C->data);
577:   c->free_a  = PETSC_TRUE;
578:   c->free_ij = PETSC_TRUE;
579:   c->nonew   = 0;

581:   /* slower, less memory */
582:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Scalable;

584:   /* set MatInfo */
585:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
586:   if (afill < 1.0) afill = 1.0;
587:   c->maxnz                     = ci[am];
588:   c->nz                        = ci[am];
589:   C->info.mallocs           = ndouble;
590:   C->info.fill_ratio_given  = fill;
591:   C->info.fill_ratio_needed = afill;

593: #if defined(PETSC_USE_INFO)
594:   if (ci[am]) {
595:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
596:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
597:   } else {
598:     PetscInfo(C,"Empty matrix product\n");
599:   }
600: #endif
601:   return(0);
602: }

604: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Heap(Mat A,Mat B,PetscReal fill,Mat C)
605: {
606:   PetscErrorCode     ierr;
607:   Mat_SeqAIJ         *a = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
608:   const PetscInt     *ai=a->i,*bi=b->i,*aj=a->j,*bj=b->j;
609:   PetscInt           *ci,*cj,*bb;
610:   PetscInt           am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
611:   PetscReal          afill;
612:   PetscInt           i,j,col,ndouble = 0;
613:   PetscFreeSpaceList free_space=NULL,current_space=NULL;
614:   PetscHeap          h;

617:   /* Get ci and cj - by merging sorted rows using a heap */
618:   /*---------------------------------------------------------------------------------------------*/
619:   /* Allocate arrays for fill computation and free space for accumulating nonzero column */
620:   PetscMalloc1(am+2,&ci);
621:   ci[0] = 0;

623:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
624:   PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],bi[bm])),&free_space);
625:   current_space = free_space;

627:   PetscHeapCreate(a->rmax,&h);
628:   PetscMalloc1(a->rmax,&bb);

630:   /* Determine ci and cj */
631:   for (i=0; i<am; i++) {
632:     const PetscInt anzi  = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */
633:     const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */
634:     ci[i+1] = ci[i];
635:     /* Populate the min heap */
636:     for (j=0; j<anzi; j++) {
637:       bb[j] = bi[acol[j]];         /* bb points at the start of the row */
638:       if (bb[j] < bi[acol[j]+1]) { /* Add if row is nonempty */
639:         PetscHeapAdd(h,j,bj[bb[j]++]);
640:       }
641:     }
642:     /* Pick off the min element, adding it to free space */
643:     PetscHeapPop(h,&j,&col);
644:     while (j >= 0) {
645:       if (current_space->local_remaining < 1) { /* double the size, but don't exceed 16 MiB */
646:         PetscFreeSpaceGet(PetscMin(PetscIntMultTruncate(2,current_space->total_array_size),16 << 20),&current_space);
647:         ndouble++;
648:       }
649:       *(current_space->array++) = col;
650:       current_space->local_used++;
651:       current_space->local_remaining--;
652:       ci[i+1]++;

654:       /* stash if anything else remains in this row of B */
655:       if (bb[j] < bi[acol[j]+1]) {PetscHeapStash(h,j,bj[bb[j]++]);}
656:       while (1) {               /* pop and stash any other rows of B that also had an entry in this column */
657:         PetscInt j2,col2;
658:         PetscHeapPeek(h,&j2,&col2);
659:         if (col2 != col) break;
660:         PetscHeapPop(h,&j2,&col2);
661:         if (bb[j2] < bi[acol[j2]+1]) {PetscHeapStash(h,j2,bj[bb[j2]++]);}
662:       }
663:       /* Put any stashed elements back into the min heap */
664:       PetscHeapUnstash(h);
665:       PetscHeapPop(h,&j,&col);
666:     }
667:   }
668:   PetscFree(bb);
669:   PetscHeapDestroy(&h);

671:   /* Column indices are in the list of free space */
672:   /* Allocate space for cj, initialize cj, and */
673:   /* destroy list of free space and other temporary array(s) */
674:   PetscMalloc1(ci[am],&cj);
675:   PetscFreeSpaceContiguous(&free_space,cj);

677:   /* put together the new symbolic matrix */
678:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,((PetscObject)A)->type_name,C);
679:   MatSetBlockSizesFromMats(C,A,B);

681:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
682:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
683:   c          = (Mat_SeqAIJ*)(C->data);
684:   c->free_a  = PETSC_TRUE;
685:   c->free_ij = PETSC_TRUE;
686:   c->nonew   = 0;

688:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted;

690:   /* set MatInfo */
691:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
692:   if (afill < 1.0) afill = 1.0;
693:   c->maxnz                     = ci[am];
694:   c->nz                        = ci[am];
695:   C->info.mallocs           = ndouble;
696:   C->info.fill_ratio_given  = fill;
697:   C->info.fill_ratio_needed = afill;

699: #if defined(PETSC_USE_INFO)
700:   if (ci[am]) {
701:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
702:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
703:   } else {
704:     PetscInfo(C,"Empty matrix product\n");
705:   }
706: #endif
707:   return(0);
708: }

710: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_BTHeap(Mat A,Mat B,PetscReal fill,Mat C)
711: {
712:   PetscErrorCode     ierr;
713:   Mat_SeqAIJ         *a  = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
714:   const PetscInt     *ai = a->i,*bi=b->i,*aj=a->j,*bj=b->j;
715:   PetscInt           *ci,*cj,*bb;
716:   PetscInt           am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
717:   PetscReal          afill;
718:   PetscInt           i,j,col,ndouble = 0;
719:   PetscFreeSpaceList free_space=NULL,current_space=NULL;
720:   PetscHeap          h;
721:   PetscBT            bt;

724:   /* Get ci and cj - using a heap for the sorted rows, but use BT so that each index is only added once */
725:   /*---------------------------------------------------------------------------------------------*/
726:   /* Allocate arrays for fill computation and free space for accumulating nonzero column */
727:   PetscMalloc1(am+2,&ci);
728:   ci[0] = 0;

730:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
731:   PetscFreeSpaceGet(PetscRealIntMultTruncate(fill,PetscIntSumTruncate(ai[am],bi[bm])),&free_space);

733:   current_space = free_space;

735:   PetscHeapCreate(a->rmax,&h);
736:   PetscMalloc1(a->rmax,&bb);
737:   PetscBTCreate(bn,&bt);

739:   /* Determine ci and cj */
740:   for (i=0; i<am; i++) {
741:     const PetscInt anzi  = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */
742:     const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */
743:     const PetscInt *fptr = current_space->array; /* Save beginning of the row so we can clear the BT later */
744:     ci[i+1] = ci[i];
745:     /* Populate the min heap */
746:     for (j=0; j<anzi; j++) {
747:       PetscInt brow = acol[j];
748:       for (bb[j] = bi[brow]; bb[j] < bi[brow+1]; bb[j]++) {
749:         PetscInt bcol = bj[bb[j]];
750:         if (!PetscBTLookupSet(bt,bcol)) { /* new entry */
751:           PetscHeapAdd(h,j,bcol);
752:           bb[j]++;
753:           break;
754:         }
755:       }
756:     }
757:     /* Pick off the min element, adding it to free space */
758:     PetscHeapPop(h,&j,&col);
759:     while (j >= 0) {
760:       if (current_space->local_remaining < 1) { /* double the size, but don't exceed 16 MiB */
761:         fptr = NULL;                      /* need PetscBTMemzero */
762:         PetscFreeSpaceGet(PetscMin(PetscIntMultTruncate(2,current_space->total_array_size),16 << 20),&current_space);
763:         ndouble++;
764:       }
765:       *(current_space->array++) = col;
766:       current_space->local_used++;
767:       current_space->local_remaining--;
768:       ci[i+1]++;

770:       /* stash if anything else remains in this row of B */
771:       for (; bb[j] < bi[acol[j]+1]; bb[j]++) {
772:         PetscInt bcol = bj[bb[j]];
773:         if (!PetscBTLookupSet(bt,bcol)) { /* new entry */
774:           PetscHeapAdd(h,j,bcol);
775:           bb[j]++;
776:           break;
777:         }
778:       }
779:       PetscHeapPop(h,&j,&col);
780:     }
781:     if (fptr) {                 /* Clear the bits for this row */
782:       for (; fptr<current_space->array; fptr++) {PetscBTClear(bt,*fptr);}
783:     } else {                    /* We reallocated so we don't remember (easily) how to clear only the bits we changed */
784:       PetscBTMemzero(bn,bt);
785:     }
786:   }
787:   PetscFree(bb);
788:   PetscHeapDestroy(&h);
789:   PetscBTDestroy(&bt);

791:   /* Column indices are in the list of free space */
792:   /* Allocate space for cj, initialize cj, and */
793:   /* destroy list of free space and other temporary array(s) */
794:   PetscMalloc1(ci[am],&cj);
795:   PetscFreeSpaceContiguous(&free_space,cj);

797:   /* put together the new symbolic matrix */
798:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,((PetscObject)A)->type_name,C);
799:   MatSetBlockSizesFromMats(C,A,B);

801:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
802:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
803:   c          = (Mat_SeqAIJ*)(C->data);
804:   c->free_a  = PETSC_TRUE;
805:   c->free_ij = PETSC_TRUE;
806:   c->nonew   = 0;

808:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted;

810:   /* set MatInfo */
811:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
812:   if (afill < 1.0) afill = 1.0;
813:   c->maxnz                     = ci[am];
814:   c->nz                        = ci[am];
815:   C->info.mallocs           = ndouble;
816:   C->info.fill_ratio_given  = fill;
817:   C->info.fill_ratio_needed = afill;

819: #if defined(PETSC_USE_INFO)
820:   if (ci[am]) {
821:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
822:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
823:   } else {
824:     PetscInfo(C,"Empty matrix product\n");
825:   }
826: #endif
827:   return(0);
828: }


831: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_RowMerge(Mat A,Mat B,PetscReal fill,Mat C)
832: {
833:   PetscErrorCode     ierr;
834:   Mat_SeqAIJ         *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
835:   const PetscInt     *ai=a->i,*bi=b->i,*aj=a->j,*bj=b->j,*inputi,*inputj,*inputcol,*inputcol_L1;
836:   PetscInt           *ci,*cj,*outputj,worki_L1[9],worki_L2[9];
837:   PetscInt           c_maxmem,a_maxrownnz=0,a_rownnz;
838:   const PetscInt     workcol[8]={0,1,2,3,4,5,6,7};
839:   const PetscInt     am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
840:   const PetscInt     *brow_ptr[8],*brow_end[8];
841:   PetscInt           window[8];
842:   PetscInt           window_min,old_window_min,ci_nnz,outputi_nnz=0,L1_nrows,L2_nrows;
843:   PetscInt           i,k,ndouble=0,L1_rowsleft,rowsleft;
844:   PetscReal          afill;
845:   PetscInt           *workj_L1,*workj_L2,*workj_L3;
846:   PetscInt           L1_nnz,L2_nnz;

848:   /* Step 1: Get upper bound on memory required for allocation.
849:              Because of the way virtual memory works,
850:              only the memory pages that are actually needed will be physically allocated. */
852:   PetscMalloc1(am+1,&ci);
853:   for (i=0; i<am; i++) {
854:     const PetscInt anzi  = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */
855:     const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */
856:     a_rownnz = 0;
857:     for (k=0; k<anzi; ++k) {
858:       a_rownnz += bi[acol[k]+1] - bi[acol[k]];
859:       if (a_rownnz > bn) {
860:         a_rownnz = bn;
861:         break;
862:       }
863:     }
864:     a_maxrownnz = PetscMax(a_maxrownnz, a_rownnz);
865:   }
866:   /* temporary work areas for merging rows */
867:   PetscMalloc1(a_maxrownnz*8,&workj_L1);
868:   PetscMalloc1(a_maxrownnz*8,&workj_L2);
869:   PetscMalloc1(a_maxrownnz,&workj_L3);

871:   /* This should be enough for almost all matrices. If not, memory is reallocated later. */
872:   c_maxmem = 8*(ai[am]+bi[bm]);
873:   /* Step 2: Populate pattern for C */
874:   PetscMalloc1(c_maxmem,&cj);

876:   ci_nnz       = 0;
877:   ci[0]        = 0;
878:   worki_L1[0]  = 0;
879:   worki_L2[0]  = 0;
880:   for (i=0; i<am; i++) {
881:     const PetscInt anzi  = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */
882:     const PetscInt *acol = aj + ai[i];      /* column indices of nonzero entries in this row */
883:     rowsleft             = anzi;
884:     inputcol_L1          = acol;
885:     L2_nnz               = 0;
886:     L2_nrows             = 1;  /* Number of rows to be merged on Level 3. output of L3 already exists -> initial value 1   */
887:     worki_L2[1]          = 0;
888:     outputi_nnz          = 0;

890:     /* If the number of indices in C so far + the max number of columns in the next row > c_maxmem  -> allocate more memory */
891:     while (ci_nnz+a_maxrownnz > c_maxmem) {
892:       c_maxmem *= 2;
893:       ndouble++;
894:       PetscRealloc(sizeof(PetscInt)*c_maxmem,&cj);
895:     }

897:     while (rowsleft) {
898:       L1_rowsleft = PetscMin(64, rowsleft); /* In the inner loop max 64 rows of B can be merged */
899:       L1_nrows    = 0;
900:       L1_nnz      = 0;
901:       inputcol    = inputcol_L1;
902:       inputi      = bi;
903:       inputj      = bj;

905:       /* The following macro is used to specialize for small rows in A.
906:          This helps with compiler unrolling, improving performance substantially.
907:           Input:  inputj   inputi  inputcol  bn
908:           Output: outputj  outputi_nnz                       */
909:        #define MatMatMultSymbolic_RowMergeMacro(ANNZ)                        \
910:          window_min  = bn;                                                   \
911:          outputi_nnz = 0;                                                    \
912:          for (k=0; k<ANNZ; ++k) {                                            \
913:            brow_ptr[k] = inputj + inputi[inputcol[k]];                       \
914:            brow_end[k] = inputj + inputi[inputcol[k]+1];                     \
915:            window[k]   = (brow_ptr[k] != brow_end[k]) ? *brow_ptr[k] : bn;   \
916:            window_min  = PetscMin(window[k], window_min);                    \
917:          }                                                                   \
918:          while (window_min < bn) {                                           \
919:            outputj[outputi_nnz++] = window_min;                              \
920:            /* advance front and compute new minimum */                       \
921:            old_window_min = window_min;                                      \
922:            window_min = bn;                                                  \
923:            for (k=0; k<ANNZ; ++k) {                                          \
924:              if (window[k] == old_window_min) {                              \
925:                brow_ptr[k]++;                                                \
926:                window[k] = (brow_ptr[k] != brow_end[k]) ? *brow_ptr[k] : bn; \
927:              }                                                               \
928:              window_min = PetscMin(window[k], window_min);                   \
929:            }                                                                 \
930:          }

932:       /************** L E V E L  1 ***************/
933:       /* Merge up to 8 rows of B to L1 work array*/
934:       while (L1_rowsleft) {
935:         outputi_nnz = 0;
936:         if (anzi > 8)  outputj = workj_L1 + L1_nnz;     /* Level 1 rowmerge*/
937:         else           outputj = cj + ci_nnz;           /* Merge directly to C */

939:         switch (L1_rowsleft) {
940:         case 1:  brow_ptr[0] = inputj + inputi[inputcol[0]];
941:                  brow_end[0] = inputj + inputi[inputcol[0]+1];
942:                  for (; brow_ptr[0] != brow_end[0]; ++brow_ptr[0]) outputj[outputi_nnz++] = *brow_ptr[0]; /* copy row in b over */
943:                  inputcol    += L1_rowsleft;
944:                  rowsleft    -= L1_rowsleft;
945:                  L1_rowsleft  = 0;
946:                  break;
947:         case 2:  MatMatMultSymbolic_RowMergeMacro(2);
948:                  inputcol    += L1_rowsleft;
949:                  rowsleft    -= L1_rowsleft;
950:                  L1_rowsleft  = 0;
951:                  break;
952:         case 3: MatMatMultSymbolic_RowMergeMacro(3);
953:                  inputcol    += L1_rowsleft;
954:                  rowsleft    -= L1_rowsleft;
955:                  L1_rowsleft  = 0;
956:                  break;
957:         case 4:  MatMatMultSymbolic_RowMergeMacro(4);
958:                  inputcol    += L1_rowsleft;
959:                  rowsleft    -= L1_rowsleft;
960:                  L1_rowsleft  = 0;
961:                  break;
962:         case 5:  MatMatMultSymbolic_RowMergeMacro(5);
963:                  inputcol    += L1_rowsleft;
964:                  rowsleft    -= L1_rowsleft;
965:                  L1_rowsleft  = 0;
966:                  break;
967:         case 6:  MatMatMultSymbolic_RowMergeMacro(6);
968:                  inputcol    += L1_rowsleft;
969:                  rowsleft    -= L1_rowsleft;
970:                  L1_rowsleft  = 0;
971:                  break;
972:         case 7:  MatMatMultSymbolic_RowMergeMacro(7);
973:                  inputcol    += L1_rowsleft;
974:                  rowsleft    -= L1_rowsleft;
975:                  L1_rowsleft  = 0;
976:                  break;
977:         default: MatMatMultSymbolic_RowMergeMacro(8);
978:                  inputcol    += 8;
979:                  rowsleft    -= 8;
980:                  L1_rowsleft -= 8;
981:                  break;
982:         }
983:         inputcol_L1           = inputcol;
984:         L1_nnz               += outputi_nnz;
985:         worki_L1[++L1_nrows]  = L1_nnz;
986:       }

988:       /********************** L E V E L  2 ************************/
989:       /* Merge from L1 work array to either C or to L2 work array */
990:       if (anzi > 8) {
991:         inputi      = worki_L1;
992:         inputj      = workj_L1;
993:         inputcol    = workcol;
994:         outputi_nnz = 0;

996:         if (anzi <= 64) outputj = cj + ci_nnz;        /* Merge from L1 work array to C */
997:         else            outputj = workj_L2 + L2_nnz;  /* Merge from L1 work array to L2 work array */

999:         switch (L1_nrows) {
1000:         case 1:  brow_ptr[0] = inputj + inputi[inputcol[0]];
1001:                  brow_end[0] = inputj + inputi[inputcol[0]+1];
1002:                  for (; brow_ptr[0] != brow_end[0]; ++brow_ptr[0]) outputj[outputi_nnz++] = *brow_ptr[0]; /* copy row in b over */
1003:                  break;
1004:         case 2:  MatMatMultSymbolic_RowMergeMacro(2); break;
1005:         case 3:  MatMatMultSymbolic_RowMergeMacro(3); break;
1006:         case 4:  MatMatMultSymbolic_RowMergeMacro(4); break;
1007:         case 5:  MatMatMultSymbolic_RowMergeMacro(5); break;
1008:         case 6:  MatMatMultSymbolic_RowMergeMacro(6); break;
1009:         case 7:  MatMatMultSymbolic_RowMergeMacro(7); break;
1010:         case 8:  MatMatMultSymbolic_RowMergeMacro(8); break;
1011:         default: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatMatMult logic error: Not merging 1-8 rows from L1 work array!");
1012:         }
1013:         L2_nnz               += outputi_nnz;
1014:         worki_L2[++L2_nrows]  = L2_nnz;

1016:         /************************ L E V E L  3 **********************/
1017:         /* Merge from L2 work array to either C or to L2 work array */
1018:         if (anzi > 64 && (L2_nrows == 8 || rowsleft == 0)) {
1019:           inputi      = worki_L2;
1020:           inputj      = workj_L2;
1021:           inputcol    = workcol;
1022:           outputi_nnz = 0;
1023:           if (rowsleft) outputj = workj_L3;
1024:           else          outputj = cj + ci_nnz;
1025:           switch (L2_nrows) {
1026:           case 1:  brow_ptr[0] = inputj + inputi[inputcol[0]];
1027:                    brow_end[0] = inputj + inputi[inputcol[0]+1];
1028:                    for (; brow_ptr[0] != brow_end[0]; ++brow_ptr[0]) outputj[outputi_nnz++] = *brow_ptr[0]; /* copy row in b over */
1029:                    break;
1030:           case 2:  MatMatMultSymbolic_RowMergeMacro(2); break;
1031:           case 3:  MatMatMultSymbolic_RowMergeMacro(3); break;
1032:           case 4:  MatMatMultSymbolic_RowMergeMacro(4); break;
1033:           case 5:  MatMatMultSymbolic_RowMergeMacro(5); break;
1034:           case 6:  MatMatMultSymbolic_RowMergeMacro(6); break;
1035:           case 7:  MatMatMultSymbolic_RowMergeMacro(7); break;
1036:           case 8:  MatMatMultSymbolic_RowMergeMacro(8); break;
1037:           default: SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"MatMatMult logic error: Not merging 1-8 rows from L2 work array!");
1038:           }
1039:           L2_nrows    = 1;
1040:           L2_nnz      = outputi_nnz;
1041:           worki_L2[1] = outputi_nnz;
1042:           /* Copy to workj_L2 */
1043:           if (rowsleft) {
1044:             for (k=0; k<outputi_nnz; ++k)  workj_L2[k] = outputj[k];
1045:           }
1046:         }
1047:       }
1048:     }  /* while (rowsleft) */
1049: #undef MatMatMultSymbolic_RowMergeMacro

1051:     /* terminate current row */
1052:     ci_nnz += outputi_nnz;
1053:     ci[i+1] = ci_nnz;
1054:   }

1056:   /* Step 3: Create the new symbolic matrix */
1057:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,((PetscObject)A)->type_name,C);
1058:   MatSetBlockSizesFromMats(C,A,B);

1060:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
1061:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
1062:   c          = (Mat_SeqAIJ*)(C->data);
1063:   c->free_a  = PETSC_TRUE;
1064:   c->free_ij = PETSC_TRUE;
1065:   c->nonew   = 0;

1067:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted;

1069:   /* set MatInfo */
1070:   afill = (PetscReal)ci[am]/(ai[am]+bi[bm]) + 1.e-5;
1071:   if (afill < 1.0) afill = 1.0;
1072:   c->maxnz                     = ci[am];
1073:   c->nz                        = ci[am];
1074:   C->info.mallocs           = ndouble;
1075:   C->info.fill_ratio_given  = fill;
1076:   C->info.fill_ratio_needed = afill;

1078: #if defined(PETSC_USE_INFO)
1079:   if (ci[am]) {
1080:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
1081:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
1082:   } else {
1083:     PetscInfo(C,"Empty matrix product\n");
1084:   }
1085: #endif

1087:   /* Step 4: Free temporary work areas */
1088:   PetscFree(workj_L1);
1089:   PetscFree(workj_L2);
1090:   PetscFree(workj_L3);
1091:   return(0);
1092: }

1094: /* concatenate unique entries and then sort */
1095: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqAIJ_Sorted(Mat A,Mat B,PetscReal fill,Mat C)
1096: {
1098:   Mat_SeqAIJ     *a  = (Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c;
1099:   const PetscInt *ai = a->i,*bi=b->i,*aj=a->j,*bj=b->j;
1100:   PetscInt       *ci,*cj,bcol;
1101:   PetscInt       am=A->rmap->N,bn=B->cmap->N,bm=B->rmap->N;
1102:   PetscReal      afill;
1103:   PetscInt       i,j,ndouble = 0;
1104:   PetscSegBuffer seg,segrow;
1105:   char           *seen;

1108:   PetscMalloc1(am+1,&ci);
1109:   ci[0] = 0;

1111:   /* Initial FreeSpace size is fill*(nnz(A)+nnz(B)) */
1112:   PetscSegBufferCreate(sizeof(PetscInt),(PetscInt)(fill*(ai[am]+bi[bm])),&seg);
1113:   PetscSegBufferCreate(sizeof(PetscInt),100,&segrow);
1114:   PetscCalloc1(bn,&seen);

1116:   /* Determine ci and cj */
1117:   for (i=0; i<am; i++) {
1118:     const PetscInt anzi  = ai[i+1] - ai[i]; /* number of nonzeros in this row of A, this is the number of rows of B that we merge */
1119:     const PetscInt *acol = aj + ai[i]; /* column indices of nonzero entries in this row */
1120:     PetscInt packlen = 0,*PETSC_RESTRICT crow;

1122:     /* Pack segrow */
1123:     for (j=0; j<anzi; j++) {
1124:       PetscInt brow = acol[j],bjstart = bi[brow],bjend = bi[brow+1],k;
1125:       for (k=bjstart; k<bjend; k++) {
1126:         bcol = bj[k];
1127:         if (!seen[bcol]) { /* new entry */
1128:           PetscInt *PETSC_RESTRICT slot;
1129:           PetscSegBufferGetInts(segrow,1,&slot);
1130:           *slot = bcol;
1131:           seen[bcol] = 1;
1132:           packlen++;
1133:         }
1134:       }
1135:     }

1137:     /* Check i-th diagonal entry */
1138:     if (C->force_diagonals && !seen[i]) {
1139:       PetscInt *PETSC_RESTRICT slot;
1140:       PetscSegBufferGetInts(segrow,1,&slot);
1141:       *slot   = i;
1142:       seen[i] = 1;
1143:       packlen++;
1144:     }

1146:     PetscSegBufferGetInts(seg,packlen,&crow);
1147:     PetscSegBufferExtractTo(segrow,crow);
1148:     PetscSortInt(packlen,crow);
1149:     ci[i+1] = ci[i] + packlen;
1150:     for (j=0; j<packlen; j++) seen[crow[j]] = 0;
1151:   }
1152:   PetscSegBufferDestroy(&segrow);
1153:   PetscFree(seen);

1155:   /* Column indices are in the segmented buffer */
1156:   PetscSegBufferExtractAlloc(seg,&cj);
1157:   PetscSegBufferDestroy(&seg);

1159:   /* put together the new symbolic matrix */
1160:   MatSetSeqAIJWithArrays_private(PetscObjectComm((PetscObject)A),am,bn,ci,cj,NULL,((PetscObject)A)->type_name,C);
1161:   MatSetBlockSizesFromMats(C,A,B);

1163:   /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
1164:   /* These are PETSc arrays, so change flags so arrays can be deleted by PETSc */
1165:   c          = (Mat_SeqAIJ*)(C->data);
1166:   c->free_a  = PETSC_TRUE;
1167:   c->free_ij = PETSC_TRUE;
1168:   c->nonew   = 0;

1170:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqAIJ_Sorted;

1172:   /* set MatInfo */
1173:   afill = (PetscReal)ci[am]/PetscMax(ai[am]+bi[bm],1) + 1.e-5;
1174:   if (afill < 1.0) afill = 1.0;
1175:   c->maxnz                  = ci[am];
1176:   c->nz                     = ci[am];
1177:   C->info.mallocs           = ndouble;
1178:   C->info.fill_ratio_given  = fill;
1179:   C->info.fill_ratio_needed = afill;

1181: #if defined(PETSC_USE_INFO)
1182:   if (ci[am]) {
1183:     PetscInfo3(C,"Reallocs %D; Fill ratio: given %g needed %g.\n",ndouble,(double)fill,(double)afill);
1184:     PetscInfo1(C,"Use MatMatMult(A,B,MatReuse,%g,&C) for best performance.;\n",(double)afill);
1185:   } else {
1186:     PetscInfo(C,"Empty matrix product\n");
1187:   }
1188: #endif
1189:   return(0);
1190: }

1192: PetscErrorCode MatDestroy_SeqAIJ_MatMatMultTrans(void *data)
1193: {
1194:   PetscErrorCode      ierr;
1195:   Mat_MatMatTransMult *abt=(Mat_MatMatTransMult *)data;

1198:   MatTransposeColoringDestroy(&abt->matcoloring);
1199:   MatDestroy(&abt->Bt_den);
1200:   MatDestroy(&abt->ABt_den);
1201:   PetscFree(abt);
1202:   return(0);
1203: }

1205: PetscErrorCode MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C)
1206: {
1207:   PetscErrorCode      ierr;
1208:   Mat                 Bt;
1209:   PetscInt            *bti,*btj;
1210:   Mat_MatMatTransMult *abt;
1211:   Mat_Product         *product = C->product;
1212:   char                *alg;

1215:   if (!product) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing product struct");
1216:   if (product->data) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty");

1218:   /* create symbolic Bt */
1219:   MatGetSymbolicTranspose_SeqAIJ(B,&bti,&btj);
1220:   MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,B->cmap->n,B->rmap->n,bti,btj,NULL,&Bt);
1221:   MatSetBlockSizes(Bt,PetscAbs(A->cmap->bs),PetscAbs(B->cmap->bs));
1222:   MatSetType(Bt,((PetscObject)A)->type_name);

1224:   /* get symbolic C=A*Bt */
1225:   PetscStrallocpy(product->alg,&alg);
1226:   MatProductSetAlgorithm(C,"sorted"); /* set algorithm for C = A*Bt */
1227:   MatMatMultSymbolic_SeqAIJ_SeqAIJ(A,Bt,fill,C);
1228:   MatProductSetAlgorithm(C,alg); /* resume original algorithm for ABt product */
1229:   PetscFree(alg);

1231:   /* create a supporting struct for reuse intermidiate dense matrices with matcoloring */
1232:   PetscNew(&abt);

1234:   product->data    = abt;
1235:   product->destroy = MatDestroy_SeqAIJ_MatMatMultTrans;

1237:   C->ops->mattransposemultnumeric = MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ;

1239:   abt->usecoloring = PETSC_FALSE;
1240:   PetscStrcmp(product->alg,"color",&abt->usecoloring);
1241:   if (abt->usecoloring) {
1242:     /* Create MatTransposeColoring from symbolic C=A*B^T */
1243:     MatTransposeColoring matcoloring;
1244:     MatColoring          coloring;
1245:     ISColoring           iscoloring;
1246:     Mat                  Bt_dense,C_dense;

1248:     /* inode causes memory problem */
1249:     MatSetOption(C,MAT_USE_INODES,PETSC_FALSE);

1251:     MatColoringCreate(C,&coloring);
1252:     MatColoringSetDistance(coloring,2);
1253:     MatColoringSetType(coloring,MATCOLORINGSL);
1254:     MatColoringSetFromOptions(coloring);
1255:     MatColoringApply(coloring,&iscoloring);
1256:     MatColoringDestroy(&coloring);
1257:     MatTransposeColoringCreate(C,iscoloring,&matcoloring);

1259:     abt->matcoloring = matcoloring;

1261:     ISColoringDestroy(&iscoloring);

1263:     /* Create Bt_dense and C_dense = A*Bt_dense */
1264:     MatCreate(PETSC_COMM_SELF,&Bt_dense);
1265:     MatSetSizes(Bt_dense,A->cmap->n,matcoloring->ncolors,A->cmap->n,matcoloring->ncolors);
1266:     MatSetType(Bt_dense,MATSEQDENSE);
1267:     MatSeqDenseSetPreallocation(Bt_dense,NULL);

1269:     Bt_dense->assembled = PETSC_TRUE;
1270:     abt->Bt_den         = Bt_dense;

1272:     MatCreate(PETSC_COMM_SELF,&C_dense);
1273:     MatSetSizes(C_dense,A->rmap->n,matcoloring->ncolors,A->rmap->n,matcoloring->ncolors);
1274:     MatSetType(C_dense,MATSEQDENSE);
1275:     MatSeqDenseSetPreallocation(C_dense,NULL);

1277:     Bt_dense->assembled = PETSC_TRUE;
1278:     abt->ABt_den  = C_dense;

1280: #if defined(PETSC_USE_INFO)
1281:     {
1282:       Mat_SeqAIJ *c = (Mat_SeqAIJ*)C->data;
1283:       PetscInfo7(C,"Use coloring of C=A*B^T; B^T: %D %D, Bt_dense: %D,%D; Cnz %D / (cm*ncolors %D) = %g\n",B->cmap->n,B->rmap->n,Bt_dense->rmap->n,Bt_dense->cmap->n,c->nz,A->rmap->n*matcoloring->ncolors,(PetscReal)(c->nz)/(A->rmap->n*matcoloring->ncolors));
1284:     }
1285: #endif
1286:   }
1287:   /* clean up */
1288:   MatDestroy(&Bt);
1289:   MatRestoreSymbolicTranspose_SeqAIJ(B,&bti,&btj);
1290:   return(0);
1291: }

1293: PetscErrorCode MatMatTransposeMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C)
1294: {
1295:   PetscErrorCode      ierr;
1296:   Mat_SeqAIJ          *a   =(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c=(Mat_SeqAIJ*)C->data;
1297:   PetscInt            *ai  =a->i,*aj=a->j,*bi=b->i,*bj=b->j,anzi,bnzj,nexta,nextb,*acol,*bcol,brow;
1298:   PetscInt            cm   =C->rmap->n,*ci=c->i,*cj=c->j,i,j,cnzi,*ccol;
1299:   PetscLogDouble      flops=0.0;
1300:   MatScalar           *aa  =a->a,*aval,*ba=b->a,*bval,*ca,*cval;
1301:   Mat_MatMatTransMult *abt;
1302:   Mat_Product         *product = C->product;

1305:   if (!product) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing product struct");
1306:   abt = (Mat_MatMatTransMult *)product->data;
1307:   if (!abt) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing product struct");
1308:   /* clear old values in C */
1309:   if (!c->a) {
1310:     PetscCalloc1(ci[cm]+1,&ca);
1311:     c->a      = ca;
1312:     c->free_a = PETSC_TRUE;
1313:   } else {
1314:     ca =  c->a;
1315:     PetscArrayzero(ca,ci[cm]+1);
1316:   }

1318:   if (abt->usecoloring) {
1319:     MatTransposeColoring matcoloring = abt->matcoloring;
1320:     Mat                  Bt_dense,C_dense = abt->ABt_den;

1322:     /* Get Bt_dense by Apply MatTransposeColoring to B */
1323:     Bt_dense = abt->Bt_den;
1324:     MatTransColoringApplySpToDen(matcoloring,B,Bt_dense);

1326:     /* C_dense = A*Bt_dense */
1327:     MatMatMultNumeric_SeqAIJ_SeqDense(A,Bt_dense,C_dense);

1329:     /* Recover C from C_dense */
1330:     MatTransColoringApplyDenToSp(matcoloring,C_dense,C);
1331:     return(0);
1332:   }

1334:   for (i=0; i<cm; i++) {
1335:     anzi = ai[i+1] - ai[i];
1336:     acol = aj + ai[i];
1337:     aval = aa + ai[i];
1338:     cnzi = ci[i+1] - ci[i];
1339:     ccol = cj + ci[i];
1340:     cval = ca + ci[i];
1341:     for (j=0; j<cnzi; j++) {
1342:       brow = ccol[j];
1343:       bnzj = bi[brow+1] - bi[brow];
1344:       bcol = bj + bi[brow];
1345:       bval = ba + bi[brow];

1347:       /* perform sparse inner-product c(i,j)=A[i,:]*B[j,:]^T */
1348:       nexta = 0; nextb = 0;
1349:       while (nexta<anzi && nextb<bnzj) {
1350:         while (nexta < anzi && acol[nexta] < bcol[nextb]) nexta++;
1351:         if (nexta == anzi) break;
1352:         while (nextb < bnzj && acol[nexta] > bcol[nextb]) nextb++;
1353:         if (nextb == bnzj) break;
1354:         if (acol[nexta] == bcol[nextb]) {
1355:           cval[j] += aval[nexta]*bval[nextb];
1356:           nexta++; nextb++;
1357:           flops += 2;
1358:         }
1359:       }
1360:     }
1361:   }
1362:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
1363:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
1364:   PetscLogFlops(flops);
1365:   return(0);
1366: }

1368: PetscErrorCode MatDestroy_SeqAIJ_MatTransMatMult(void *data)
1369: {
1370:   PetscErrorCode      ierr;
1371:   Mat_MatTransMatMult *atb = (Mat_MatTransMatMult*)data;

1374:   MatDestroy(&atb->At);
1375:   if (atb->destroy) {
1376:     (*atb->destroy)(atb->data);
1377:   }
1378:   PetscFree(atb);
1379:   return(0);
1380: }

1382: PetscErrorCode MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(Mat A,Mat B,PetscReal fill,Mat C)
1383: {
1385:   Mat            At = NULL;
1386:   PetscInt       *ati,*atj;
1387:   Mat_Product    *product = C->product;
1388:   PetscBool      flg,def,square;

1391:   MatCheckProduct(C,4);
1392:   square = (PetscBool)(A == B && A->symmetric && A->symmetric_set);
1393:   /* outerproduct */
1394:   PetscStrcmp(product->alg,"outerproduct",&flg);
1395:   if (flg) {
1396:     /* create symbolic At */
1397:     if (!square) {
1398:       MatGetSymbolicTranspose_SeqAIJ(A,&ati,&atj);
1399:       MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,A->cmap->n,A->rmap->n,ati,atj,NULL,&At);
1400:       MatSetBlockSizes(At,PetscAbs(A->cmap->bs),PetscAbs(B->cmap->bs));
1401:       MatSetType(At,((PetscObject)A)->type_name);
1402:     }
1403:     /* get symbolic C=At*B */
1404:     MatProductSetAlgorithm(C,"sorted");
1405:     MatMatMultSymbolic_SeqAIJ_SeqAIJ(square ? A : At,B,fill,C);

1407:     /* clean up */
1408:     if (!square) {
1409:       MatDestroy(&At);
1410:       MatRestoreSymbolicTranspose_SeqAIJ(A,&ati,&atj);
1411:     }

1413:     C->ops->mattransposemultnumeric = MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ; /* outerproduct */
1414:     MatProductSetAlgorithm(C,"outerproduct");
1415:     return(0);
1416:   }

1418:   /* matmatmult */
1419:   PetscStrcmp(product->alg,"default",&def);
1420:   PetscStrcmp(product->alg,"at*b",&flg);
1421:   if (flg || def) {
1422:     Mat_MatTransMatMult *atb;

1424:     if (product->data) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Extra product struct not empty");
1425:     PetscNew(&atb);
1426:     if (!square) {
1427:       MatTranspose_SeqAIJ(A,MAT_INITIAL_MATRIX,&At);
1428:     }
1429:     MatProductSetAlgorithm(C,"sorted");
1430:     MatMatMultSymbolic_SeqAIJ_SeqAIJ(square ? A : At,B,fill,C);
1431:     MatProductSetAlgorithm(C,"at*b");
1432:     product->data    = atb;
1433:     product->destroy = MatDestroy_SeqAIJ_MatTransMatMult;
1434:     atb->At          = At;
1435:     atb->updateAt    = PETSC_FALSE; /* because At is computed here */

1437:     C->ops->mattransposemultnumeric = NULL; /* see MatProductNumeric_AtB_SeqAIJ_SeqAIJ */
1438:     return(0);
1439:   }

1441:   SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Mat Product Algorithm is not supported");
1442: }

1444: PetscErrorCode MatTransposeMatMultNumeric_SeqAIJ_SeqAIJ(Mat A,Mat B,Mat C)
1445: {
1447:   Mat_SeqAIJ     *a=(Mat_SeqAIJ*)A->data,*b=(Mat_SeqAIJ*)B->data,*c=(Mat_SeqAIJ*)C->data;
1448:   PetscInt       am=A->rmap->n,anzi,*ai=a->i,*aj=a->j,*bi=b->i,*bj,bnzi,nextb;
1449:   PetscInt       cm=C->rmap->n,*ci=c->i,*cj=c->j,crow,*cjj,i,j,k;
1450:   PetscLogDouble flops=0.0;
1451:   MatScalar      *aa=a->a,*ba,*ca,*caj;

1454:   if (!c->a) {
1455:     PetscCalloc1(ci[cm]+1,&ca);

1457:     c->a      = ca;
1458:     c->free_a = PETSC_TRUE;
1459:   } else {
1460:     ca   = c->a;
1461:     PetscArrayzero(ca,ci[cm]);
1462:   }

1464:   /* compute A^T*B using outer product (A^T)[:,i]*B[i,:] */
1465:   for (i=0; i<am; i++) {
1466:     bj   = b->j + bi[i];
1467:     ba   = b->a + bi[i];
1468:     bnzi = bi[i+1] - bi[i];
1469:     anzi = ai[i+1] - ai[i];
1470:     for (j=0; j<anzi; j++) {
1471:       nextb = 0;
1472:       crow  = *aj++;
1473:       cjj   = cj + ci[crow];
1474:       caj   = ca + ci[crow];
1475:       /* perform sparse axpy operation.  Note cjj includes bj. */
1476:       for (k=0; nextb<bnzi; k++) {
1477:         if (cjj[k] == *(bj+nextb)) { /* ccol == bcol */
1478:           caj[k] += (*aa)*(*(ba+nextb));
1479:           nextb++;
1480:         }
1481:       }
1482:       flops += 2*bnzi;
1483:       aa++;
1484:     }
1485:   }

1487:   /* Assemble the final matrix and clean up */
1488:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
1489:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
1490:   PetscLogFlops(flops);
1491:   return(0);
1492: }

1494: PetscErrorCode MatMatMultSymbolic_SeqAIJ_SeqDense(Mat A,Mat B,PetscReal fill,Mat C)
1495: {

1499:   MatMatMultSymbolic_SeqDense_SeqDense(A,B,0.0,C);
1500:   C->ops->matmultnumeric = MatMatMultNumeric_SeqAIJ_SeqDense;
1501:   return(0);
1502: }

1504: PETSC_INTERN PetscErrorCode MatMatMultNumericAdd_SeqAIJ_SeqDense(Mat A,Mat B,Mat C,const PetscBool add)
1505: {
1506:   Mat_SeqAIJ        *a=(Mat_SeqAIJ*)A->data;
1507:   Mat_SeqDense      *bd=(Mat_SeqDense*)B->data;
1508:   Mat_SeqDense      *cd=(Mat_SeqDense*)C->data;
1509:   PetscErrorCode    ierr;
1510:   PetscScalar       *c,r1,r2,r3,r4,*c1,*c2,*c3,*c4;
1511:   const PetscScalar *aa,*b,*b1,*b2,*b3,*b4,*av;
1512:   const PetscInt    *aj;
1513:   PetscInt          cm=C->rmap->n,cn=B->cmap->n,bm=bd->lda,am=A->rmap->n;
1514:   PetscInt          clda=cd->lda;
1515:   PetscInt          am4=4*clda,bm4=4*bm,col,i,j,n;

1518:   if (!cm || !cn) return(0);
1519:   MatSeqAIJGetArrayRead(A,&av);
1520:   if (add) {
1521:     MatDenseGetArray(C,&c);
1522:   } else {
1523:     MatDenseGetArrayWrite(C,&c);
1524:   }
1525:   MatDenseGetArrayRead(B,&b);
1526:   b1 = b; b2 = b1 + bm; b3 = b2 + bm; b4 = b3 + bm;
1527:   c1 = c; c2 = c1 + clda; c3 = c2 + clda; c4 = c3 + clda;
1528:   for (col=0; col<(cn/4)*4; col += 4) {  /* over columns of C */
1529:     for (i=0; i<am; i++) {        /* over rows of A in those columns */
1530:       r1 = r2 = r3 = r4 = 0.0;
1531:       n  = a->i[i+1] - a->i[i];
1532:       aj = a->j + a->i[i];
1533:       aa = av + a->i[i];
1534:       for (j=0; j<n; j++) {
1535:         const PetscScalar aatmp = aa[j];
1536:         const PetscInt    ajtmp = aj[j];
1537:         r1 += aatmp*b1[ajtmp];
1538:         r2 += aatmp*b2[ajtmp];
1539:         r3 += aatmp*b3[ajtmp];
1540:         r4 += aatmp*b4[ajtmp];
1541:       }
1542:       if (add) {
1543:         c1[i] += r1;
1544:         c2[i] += r2;
1545:         c3[i] += r3;
1546:         c4[i] += r4;
1547:       } else {
1548:         c1[i] = r1;
1549:         c2[i] = r2;
1550:         c3[i] = r3;
1551:         c4[i] = r4;
1552:       }
1553:     }
1554:     b1 += bm4; b2 += bm4; b3 += bm4; b4 += bm4;
1555:     c1 += am4; c2 += am4; c3 += am4; c4 += am4;
1556:   }
1557:   /* process remaining columns */
1558:   if (col != cn) {
1559:     PetscInt rc = cn-col;

1561:     if (rc == 1) {
1562:       for (i=0; i<am; i++) {
1563:         r1 = 0.0;
1564:         n  = a->i[i+1] - a->i[i];
1565:         aj = a->j + a->i[i];
1566:         aa = av + a->i[i];
1567:         for (j=0; j<n; j++) r1 += aa[j]*b1[aj[j]];
1568:         if (add) c1[i] += r1;
1569:         else c1[i] = r1;
1570:       }
1571:     } else if (rc == 2) {
1572:       for (i=0; i<am; i++) {
1573:         r1 = r2 = 0.0;
1574:         n  = a->i[i+1] - a->i[i];
1575:         aj = a->j + a->i[i];
1576:         aa = av + a->i[i];
1577:         for (j=0; j<n; j++) {
1578:           const PetscScalar aatmp = aa[j];
1579:           const PetscInt    ajtmp = aj[j];
1580:           r1 += aatmp*b1[ajtmp];
1581:           r2 += aatmp*b2[ajtmp];
1582:         }
1583:         if (add) {
1584:           c1[i] += r1;
1585:           c2[i] += r2;
1586:         } else {
1587:           c1[i] = r1;
1588:           c2[i] = r2;
1589:         }
1590:       }
1591:     } else {
1592:       for (i=0; i<am; i++) {
1593:         r1 = r2 = r3 = 0.0;
1594:         n  = a->i[i+1] - a->i[i];
1595:         aj = a->j + a->i[i];
1596:         aa = av + a->i[i];
1597:         for (j=0; j<n; j++) {
1598:           const PetscScalar aatmp = aa[j];
1599:           const PetscInt    ajtmp = aj[j];
1600:           r1 += aatmp*b1[ajtmp];
1601:           r2 += aatmp*b2[ajtmp];
1602:           r3 += aatmp*b3[ajtmp];
1603:         }
1604:         if (add) {
1605:           c1[i] += r1;
1606:           c2[i] += r2;
1607:           c3[i] += r3;
1608:         } else {
1609:           c1[i] = r1;
1610:           c2[i] = r2;
1611:           c3[i] = r3;
1612:         }
1613:       }
1614:     }
1615:   }
1616:   PetscLogFlops(cn*(2.0*a->nz));
1617:   if (add) {
1618:     MatDenseRestoreArray(C,&c);
1619:   } else {
1620:     MatDenseRestoreArrayWrite(C,&c);
1621:   }
1622:   MatDenseRestoreArrayRead(B,&b);
1623:   MatSeqAIJRestoreArrayRead(A,&av);
1624:   return(0);
1625: }

1627: PetscErrorCode MatMatMultNumeric_SeqAIJ_SeqDense(Mat A,Mat B,Mat C)
1628: {

1632:   if (B->rmap->n != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in A %D not equal rows in B %D\n",A->cmap->n,B->rmap->n);
1633:   if (A->rmap->n != C->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number rows in C %D not equal rows in A %D\n",C->rmap->n,A->rmap->n);
1634:   if (B->cmap->n != C->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Number columns in B %D not equal columns in C %D\n",B->cmap->n,C->cmap->n);

1636:   MatMatMultNumericAdd_SeqAIJ_SeqDense(A,B,C,PETSC_FALSE);
1637:   return(0);
1638: }

1640: /* ------------------------------------------------------- */
1641: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense_AB(Mat C)
1642: {
1644:   C->ops->matmultsymbolic = MatMatMultSymbolic_SeqAIJ_SeqDense;
1645:   C->ops->productsymbolic = MatProductSymbolic_AB;
1646:   return(0);
1647: }

1649: PETSC_INTERN PetscErrorCode MatTMatTMultSymbolic_SeqAIJ_SeqDense(Mat,Mat,PetscReal,Mat);

1651: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense_AtB(Mat C)
1652: {
1654:   C->ops->transposematmultsymbolic = MatTMatTMultSymbolic_SeqAIJ_SeqDense;
1655:   C->ops->productsymbolic          = MatProductSymbolic_AtB;
1656:   return(0);
1657: }

1659: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense_ABt(Mat C)
1660: {
1662:   C->ops->mattransposemultsymbolic = MatTMatTMultSymbolic_SeqAIJ_SeqDense;
1663:   C->ops->productsymbolic          = MatProductSymbolic_ABt;
1664:   return(0);
1665: }

1667: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqAIJ_SeqDense(Mat C)
1668: {
1670:   Mat_Product    *product = C->product;

1673:   switch (product->type) {
1674:   case MATPRODUCT_AB:
1675:     MatProductSetFromOptions_SeqAIJ_SeqDense_AB(C);
1676:     break;
1677:   case MATPRODUCT_AtB:
1678:     MatProductSetFromOptions_SeqAIJ_SeqDense_AtB(C);
1679:     break;
1680:   case MATPRODUCT_ABt:
1681:     MatProductSetFromOptions_SeqAIJ_SeqDense_ABt(C);
1682:     break;
1683:   default:
1684:     break;
1685:   }
1686:   return(0);
1687: }
1688: /* ------------------------------------------------------- */
1689: static PetscErrorCode MatProductSetFromOptions_SeqXBAIJ_SeqDense_AB(Mat C)
1690: {
1692:   Mat_Product    *product = C->product;
1693:   Mat            A = product->A;
1694:   PetscBool      baij;

1697:   PetscObjectTypeCompare((PetscObject)A,MATSEQBAIJ,&baij);
1698:   if (!baij) { /* A is seqsbaij */
1699:     PetscBool sbaij;
1700:     PetscObjectTypeCompare((PetscObject)A,MATSEQSBAIJ,&sbaij);
1701:     if (!sbaij) SETERRQ(PetscObjectComm((PetscObject)C),PETSC_ERR_ARG_WRONGSTATE,"Mat must be either seqbaij or seqsbaij format");

1703:     C->ops->matmultsymbolic = MatMatMultSymbolic_SeqSBAIJ_SeqDense;
1704:   } else { /* A is seqbaij */
1705:     C->ops->matmultsymbolic = MatMatMultSymbolic_SeqBAIJ_SeqDense;
1706:   }

1708:   C->ops->productsymbolic = MatProductSymbolic_AB;
1709:   return(0);
1710: }

1712: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqXBAIJ_SeqDense(Mat C)
1713: {
1715:   Mat_Product    *product = C->product;

1718:   MatCheckProduct(C,1);
1719:   if (!product->A) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing A");
1720:   if (product->type == MATPRODUCT_AB || (product->type == MATPRODUCT_AtB && product->A->symmetric)) {
1721:     MatProductSetFromOptions_SeqXBAIJ_SeqDense_AB(C);
1722:   }
1723:   return(0);
1724: }

1726: /* ------------------------------------------------------- */
1727: static PetscErrorCode MatProductSetFromOptions_SeqDense_SeqAIJ_AB(Mat C)
1728: {
1730:   C->ops->matmultsymbolic = MatMatMultSymbolic_SeqDense_SeqAIJ;
1731:   C->ops->productsymbolic = MatProductSymbolic_AB;
1732:   return(0);
1733: }

1735: PETSC_INTERN PetscErrorCode MatProductSetFromOptions_SeqDense_SeqAIJ(Mat C)
1736: {
1738:   Mat_Product    *product = C->product;

1741:   if (product->type == MATPRODUCT_AB) {
1742:     MatProductSetFromOptions_SeqDense_SeqAIJ_AB(C);
1743:   }
1744:   return(0);
1745: }
1746: /* ------------------------------------------------------- */

1748: PetscErrorCode  MatTransColoringApplySpToDen_SeqAIJ(MatTransposeColoring coloring,Mat B,Mat Btdense)
1749: {
1751:   Mat_SeqAIJ     *b       = (Mat_SeqAIJ*)B->data;
1752:   Mat_SeqDense   *btdense = (Mat_SeqDense*)Btdense->data;
1753:   PetscInt       *bi      = b->i,*bj=b->j;
1754:   PetscInt       m        = Btdense->rmap->n,n=Btdense->cmap->n,j,k,l,col,anz,*btcol,brow,ncolumns;
1755:   MatScalar      *btval,*btval_den,*ba=b->a;
1756:   PetscInt       *columns=coloring->columns,*colorforcol=coloring->colorforcol,ncolors=coloring->ncolors;

1759:   btval_den=btdense->v;
1760:   PetscArrayzero(btval_den,m*n);
1761:   for (k=0; k<ncolors; k++) {
1762:     ncolumns = coloring->ncolumns[k];
1763:     for (l=0; l<ncolumns; l++) { /* insert a row of B to a column of Btdense */
1764:       col   = *(columns + colorforcol[k] + l);
1765:       btcol = bj + bi[col];
1766:       btval = ba + bi[col];
1767:       anz   = bi[col+1] - bi[col];
1768:       for (j=0; j<anz; j++) {
1769:         brow            = btcol[j];
1770:         btval_den[brow] = btval[j];
1771:       }
1772:     }
1773:     btval_den += m;
1774:   }
1775:   return(0);
1776: }

1778: PetscErrorCode MatTransColoringApplyDenToSp_SeqAIJ(MatTransposeColoring matcoloring,Mat Cden,Mat Csp)
1779: {
1780:   PetscErrorCode    ierr;
1781:   Mat_SeqAIJ        *csp = (Mat_SeqAIJ*)Csp->data;
1782:   const PetscScalar *ca_den,*ca_den_ptr;
1783:   PetscScalar       *ca=csp->a;
1784:   PetscInt          k,l,m=Cden->rmap->n,ncolors=matcoloring->ncolors;
1785:   PetscInt          brows=matcoloring->brows,*den2sp=matcoloring->den2sp;
1786:   PetscInt          nrows,*row,*idx;
1787:   PetscInt          *rows=matcoloring->rows,*colorforrow=matcoloring->colorforrow;

1790:   MatDenseGetArrayRead(Cden,&ca_den);

1792:   if (brows > 0) {
1793:     PetscInt *lstart,row_end,row_start;
1794:     lstart = matcoloring->lstart;
1795:     PetscArrayzero(lstart,ncolors);

1797:     row_end = brows;
1798:     if (row_end > m) row_end = m;
1799:     for (row_start=0; row_start<m; row_start+=brows) { /* loop over row blocks of Csp */
1800:       ca_den_ptr = ca_den;
1801:       for (k=0; k<ncolors; k++) { /* loop over colors (columns of Cden) */
1802:         nrows = matcoloring->nrows[k];
1803:         row   = rows  + colorforrow[k];
1804:         idx   = den2sp + colorforrow[k];
1805:         for (l=lstart[k]; l<nrows; l++) {
1806:           if (row[l] >= row_end) {
1807:             lstart[k] = l;
1808:             break;
1809:           } else {
1810:             ca[idx[l]] = ca_den_ptr[row[l]];
1811:           }
1812:         }
1813:         ca_den_ptr += m;
1814:       }
1815:       row_end += brows;
1816:       if (row_end > m) row_end = m;
1817:     }
1818:   } else { /* non-blocked impl: loop over columns of Csp - slow if Csp is large */
1819:     ca_den_ptr = ca_den;
1820:     for (k=0; k<ncolors; k++) {
1821:       nrows = matcoloring->nrows[k];
1822:       row   = rows  + colorforrow[k];
1823:       idx   = den2sp + colorforrow[k];
1824:       for (l=0; l<nrows; l++) {
1825:         ca[idx[l]] = ca_den_ptr[row[l]];
1826:       }
1827:       ca_den_ptr += m;
1828:     }
1829:   }

1831:   MatDenseRestoreArrayRead(Cden,&ca_den);
1832: #if defined(PETSC_USE_INFO)
1833:   if (matcoloring->brows > 0) {
1834:     PetscInfo1(Csp,"Loop over %D row blocks for den2sp\n",brows);
1835:   } else {
1836:     PetscInfo(Csp,"Loop over colors/columns of Cden, inefficient for large sparse matrix product \n");
1837:   }
1838: #endif
1839:   return(0);
1840: }

1842: PetscErrorCode MatTransposeColoringCreate_SeqAIJ(Mat mat,ISColoring iscoloring,MatTransposeColoring c)
1843: {
1845:   PetscInt       i,n,nrows,Nbs,j,k,m,ncols,col,cm;
1846:   const PetscInt *is,*ci,*cj,*row_idx;
1847:   PetscInt       nis = iscoloring->n,*rowhit,bs = 1;
1848:   IS             *isa;
1849:   Mat_SeqAIJ     *csp = (Mat_SeqAIJ*)mat->data;
1850:   PetscInt       *colorforrow,*rows,*rows_i,*idxhit,*spidx,*den2sp,*den2sp_i;
1851:   PetscInt       *colorforcol,*columns,*columns_i,brows;
1852:   PetscBool      flg;

1855:   ISColoringGetIS(iscoloring,PETSC_USE_POINTER,PETSC_IGNORE,&isa);

1857:   /* bs >1 is not being tested yet! */
1858:   Nbs       = mat->cmap->N/bs;
1859:   c->M      = mat->rmap->N/bs;  /* set total rows, columns and local rows */
1860:   c->N      = Nbs;
1861:   c->m      = c->M;
1862:   c->rstart = 0;
1863:   c->brows  = 100;

1865:   c->ncolors = nis;
1866:   PetscMalloc3(nis,&c->ncolumns,nis,&c->nrows,nis+1,&colorforrow);
1867:   PetscMalloc1(csp->nz+1,&rows);
1868:   PetscMalloc1(csp->nz+1,&den2sp);

1870:   brows = c->brows;
1871:   PetscOptionsGetInt(NULL,NULL,"-matden2sp_brows",&brows,&flg);
1872:   if (flg) c->brows = brows;
1873:   if (brows > 0) {
1874:     PetscMalloc1(nis+1,&c->lstart);
1875:   }

1877:   colorforrow[0] = 0;
1878:   rows_i         = rows;
1879:   den2sp_i       = den2sp;

1881:   PetscMalloc1(nis+1,&colorforcol);
1882:   PetscMalloc1(Nbs+1,&columns);

1884:   colorforcol[0] = 0;
1885:   columns_i      = columns;

1887:   /* get column-wise storage of mat */
1888:   MatGetColumnIJ_SeqAIJ_Color(mat,0,PETSC_FALSE,PETSC_FALSE,&ncols,&ci,&cj,&spidx,NULL);

1890:   cm   = c->m;
1891:   PetscMalloc1(cm+1,&rowhit);
1892:   PetscMalloc1(cm+1,&idxhit);
1893:   for (i=0; i<nis; i++) { /* loop over color */
1894:     ISGetLocalSize(isa[i],&n);
1895:     ISGetIndices(isa[i],&is);

1897:     c->ncolumns[i] = n;
1898:     if (n) {
1899:       PetscArraycpy(columns_i,is,n);
1900:     }
1901:     colorforcol[i+1] = colorforcol[i] + n;
1902:     columns_i       += n;

1904:     /* fast, crude version requires O(N*N) work */
1905:     PetscArrayzero(rowhit,cm);

1907:     for (j=0; j<n; j++) { /* loop over columns*/
1908:       col     = is[j];
1909:       row_idx = cj + ci[col];
1910:       m       = ci[col+1] - ci[col];
1911:       for (k=0; k<m; k++) { /* loop over columns marking them in rowhit */
1912:         idxhit[*row_idx]   = spidx[ci[col] + k];
1913:         rowhit[*row_idx++] = col + 1;
1914:       }
1915:     }
1916:     /* count the number of hits */
1917:     nrows = 0;
1918:     for (j=0; j<cm; j++) {
1919:       if (rowhit[j]) nrows++;
1920:     }
1921:     c->nrows[i]      = nrows;
1922:     colorforrow[i+1] = colorforrow[i] + nrows;

1924:     nrows = 0;
1925:     for (j=0; j<cm; j++) { /* loop over rows */
1926:       if (rowhit[j]) {
1927:         rows_i[nrows]   = j;
1928:         den2sp_i[nrows] = idxhit[j];
1929:         nrows++;
1930:       }
1931:     }
1932:     den2sp_i += nrows;

1934:     ISRestoreIndices(isa[i],&is);
1935:     rows_i += nrows;
1936:   }
1937:   MatRestoreColumnIJ_SeqAIJ_Color(mat,0,PETSC_FALSE,PETSC_FALSE,&ncols,&ci,&cj,&spidx,NULL);
1938:   PetscFree(rowhit);
1939:   ISColoringRestoreIS(iscoloring,PETSC_USE_POINTER,&isa);
1940:   if (csp->nz != colorforrow[nis]) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"csp->nz %d != colorforrow[nis] %d",csp->nz,colorforrow[nis]);

1942:   c->colorforrow = colorforrow;
1943:   c->rows        = rows;
1944:   c->den2sp      = den2sp;
1945:   c->colorforcol = colorforcol;
1946:   c->columns     = columns;

1948:   PetscFree(idxhit);
1949:   return(0);
1950: }

1952: /* --------------------------------------------------------------- */
1953: static PetscErrorCode MatProductNumeric_AtB_SeqAIJ_SeqAIJ(Mat C)
1954: {
1956:   Mat_Product    *product = C->product;
1957:   Mat            A=product->A,B=product->B;

1960:   if (C->ops->mattransposemultnumeric) {
1961:     /* Alg: "outerproduct" */
1962:     (*C->ops->mattransposemultnumeric)(A,B,C);
1963:   } else {
1964:     /* Alg: "matmatmult" -- C = At*B */
1965:     Mat_MatTransMatMult *atb = (Mat_MatTransMatMult *)product->data;
1966:     Mat                 At;

1968:     if (!atb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Missing product struct");
1969:     At = atb->At;
1970:     if (atb->updateAt && At) { /* At is computed in MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ() */
1971:       MatTranspose_SeqAIJ(A,MAT_REUSE_MATRIX,&At);
1972:     }
1973:     MatMatMultNumeric_SeqAIJ_SeqAIJ(At ? At : A,B,C);
1974:     atb->updateAt = PETSC_TRUE;
1975:   }
1976:   return(0);
1977: }

1979: static PetscErrorCode MatProductSymbolic_AtB_SeqAIJ_SeqAIJ(Mat C)
1980: {
1982:   Mat_Product    *product = C->product;
1983:   Mat            A=product->A,B=product->B;
1984:   PetscReal      fill=product->fill;

1987:   MatTransposeMatMultSymbolic_SeqAIJ_SeqAIJ(A,B,fill,C);

1989:   C->ops->productnumeric = MatProductNumeric_AtB_SeqAIJ_SeqAIJ;
1990:   return(0);
1991: }

1993: /* --------------------------------------------------------------- */
1994: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_AB(Mat C)
1995: {
1997:   Mat_Product    *product = C->product;
1998:   PetscInt       alg = 0; /* default algorithm */
1999:   PetscBool      flg = PETSC_FALSE;
2000: #if !defined(PETSC_HAVE_HYPRE)
2001:   const char     *algTypes[7] = {"sorted","scalable","scalable_fast","heap","btheap","llcondensed","rowmerge"};
2002:   PetscInt       nalg = 7;
2003: #else
2004:   const char     *algTypes[8] = {"sorted","scalable","scalable_fast","heap","btheap","llcondensed","rowmerge","hypre"};
2005:   PetscInt       nalg = 8;
2006: #endif

2009:   /* Set default algorithm */
2010:   PetscStrcmp(C->product->alg,"default",&flg);
2011:   if (flg) {
2012:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2013:   }

2015:   /* Get runtime option */
2016:   if (product->api_user) {
2017:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMult","Mat");
2018:     PetscOptionsEList("-matmatmult_via","Algorithmic approach","MatMatMult",algTypes,nalg,algTypes[0],&alg,&flg);
2019:     PetscOptionsEnd();
2020:   } else {
2021:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AB","Mat");
2022:     PetscOptionsEList("-matproduct_ab_via","Algorithmic approach","MatProduct_AB",algTypes,nalg,algTypes[0],&alg,&flg);
2023:     PetscOptionsEnd();
2024:   }
2025:   if (flg) {
2026:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2027:   }

2029:   C->ops->productsymbolic = MatProductSymbolic_AB;
2030:   C->ops->matmultsymbolic = MatMatMultSymbolic_SeqAIJ_SeqAIJ;
2031:   return(0);
2032: }

2034: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_AtB(Mat C)
2035: {
2037:   Mat_Product    *product = C->product;
2038:   PetscInt       alg = 0; /* default algorithm */
2039:   PetscBool      flg = PETSC_FALSE;
2040:   const char     *algTypes[3] = {"default","at*b","outerproduct"};
2041:   PetscInt       nalg = 3;

2044:   /* Get runtime option */
2045:   if (product->api_user) {
2046:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatTransposeMatMult","Mat");
2047:     PetscOptionsEList("-mattransposematmult_via","Algorithmic approach","MatTransposeMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);
2048:     PetscOptionsEnd();
2049:   } else {
2050:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_AtB","Mat");
2051:     PetscOptionsEList("-matproduct_atb_via","Algorithmic approach","MatProduct_AtB",algTypes,nalg,algTypes[alg],&alg,&flg);
2052:     PetscOptionsEnd();
2053:   }
2054:   if (flg) {
2055:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2056:   }

2058:   C->ops->productsymbolic = MatProductSymbolic_AtB_SeqAIJ_SeqAIJ;
2059:   return(0);
2060: }

2062: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_ABt(Mat C)
2063: {
2065:   Mat_Product    *product = C->product;
2066:   PetscInt       alg = 0; /* default algorithm */
2067:   PetscBool      flg = PETSC_FALSE;
2068:   const char     *algTypes[2] = {"default","color"};
2069:   PetscInt       nalg = 2;

2072:   /* Set default algorithm */
2073:   PetscStrcmp(C->product->alg,"default",&flg);
2074:   if (!flg) {
2075:     alg = 1;
2076:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2077:   }

2079:   /* Get runtime option */
2080:   if (product->api_user) {
2081:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatTransposeMult","Mat");
2082:     PetscOptionsEList("-matmattransmult_via","Algorithmic approach","MatMatTransposeMult",algTypes,nalg,algTypes[alg],&alg,&flg);
2083:     PetscOptionsEnd();
2084:   } else {
2085:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_ABt","Mat");
2086:     PetscOptionsEList("-matproduct_abt_via","Algorithmic approach","MatProduct_ABt",algTypes,nalg,algTypes[alg],&alg,&flg);
2087:     PetscOptionsEnd();
2088:   }
2089:   if (flg) {
2090:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2091:   }

2093:   C->ops->mattransposemultsymbolic = MatMatTransposeMultSymbolic_SeqAIJ_SeqAIJ;
2094:   C->ops->productsymbolic          = MatProductSymbolic_ABt;
2095:   return(0);
2096: }

2098: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_PtAP(Mat C)
2099: {
2101:   Mat_Product    *product = C->product;
2102:   PetscBool      flg = PETSC_FALSE;
2103:   PetscInt       alg = 0; /* default algorithm -- alg=1 should be default!!! */
2104: #if !defined(PETSC_HAVE_HYPRE)
2105:   const char      *algTypes[2] = {"scalable","rap"};
2106:   PetscInt        nalg = 2;
2107: #else
2108:   const char      *algTypes[3] = {"scalable","rap","hypre"};
2109:   PetscInt        nalg = 3;
2110: #endif

2113:   /* Set default algorithm */
2114:   PetscStrcmp(product->alg,"default",&flg);
2115:   if (flg) {
2116:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2117:   }

2119:   /* Get runtime option */
2120:   if (product->api_user) {
2121:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatPtAP","Mat");
2122:     PetscOptionsEList("-matptap_via","Algorithmic approach","MatPtAP",algTypes,nalg,algTypes[0],&alg,&flg);
2123:     PetscOptionsEnd();
2124:   } else {
2125:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_PtAP","Mat");
2126:     PetscOptionsEList("-matproduct_ptap_via","Algorithmic approach","MatProduct_PtAP",algTypes,nalg,algTypes[0],&alg,&flg);
2127:     PetscOptionsEnd();
2128:   }
2129:   if (flg) {
2130:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2131:   }

2133:   C->ops->productsymbolic = MatProductSymbolic_PtAP_SeqAIJ_SeqAIJ;
2134:   return(0);
2135: }

2137: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_RARt(Mat C)
2138: {
2140:   Mat_Product    *product = C->product;
2141:   PetscBool      flg = PETSC_FALSE;
2142:   PetscInt       alg = 0; /* default algorithm */
2143:   const char     *algTypes[3] = {"r*a*rt","r*art","coloring_rart"};
2144:   PetscInt        nalg = 3;

2147:   /* Set default algorithm */
2148:   PetscStrcmp(product->alg,"default",&flg);
2149:   if (flg) {
2150:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2151:   }

2153:   /* Get runtime option */
2154:   if (product->api_user) {
2155:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatRARt","Mat");
2156:     PetscOptionsEList("-matrart_via","Algorithmic approach","MatRARt",algTypes,nalg,algTypes[0],&alg,&flg);
2157:     PetscOptionsEnd();
2158:   } else {
2159:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_RARt","Mat");
2160:     PetscOptionsEList("-matproduct_rart_via","Algorithmic approach","MatProduct_RARt",algTypes,nalg,algTypes[0],&alg,&flg);
2161:     PetscOptionsEnd();
2162:   }
2163:   if (flg) {
2164:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2165:   }

2167:   C->ops->productsymbolic = MatProductSymbolic_RARt_SeqAIJ_SeqAIJ;
2168:   return(0);
2169: }

2171: /* ABC = A*B*C = A*(B*C); ABC's algorithm must be chosen from AB's algorithm */
2172: static PetscErrorCode MatProductSetFromOptions_SeqAIJ_ABC(Mat C)
2173: {
2175:   Mat_Product    *product = C->product;
2176:   PetscInt       alg = 0; /* default algorithm */
2177:   PetscBool      flg = PETSC_FALSE;
2178:   const char     *algTypes[7] = {"sorted","scalable","scalable_fast","heap","btheap","llcondensed","rowmerge"};
2179:   PetscInt       nalg = 7;

2182:   /* Set default algorithm */
2183:   PetscStrcmp(product->alg,"default",&flg);
2184:   if (flg) {
2185:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2186:   }

2188:   /* Get runtime option */
2189:   if (product->api_user) {
2190:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatMatMatMult","Mat");
2191:     PetscOptionsEList("-matmatmatmult_via","Algorithmic approach","MatMatMatMult",algTypes,nalg,algTypes[alg],&alg,&flg);
2192:     PetscOptionsEnd();
2193:   } else {
2194:     PetscOptionsBegin(PetscObjectComm((PetscObject)C),((PetscObject)C)->prefix,"MatProduct_ABC","Mat");
2195:     PetscOptionsEList("-matproduct_abc_via","Algorithmic approach","MatProduct_ABC",algTypes,nalg,algTypes[alg],&alg,&flg);
2196:     PetscOptionsEnd();
2197:   }
2198:   if (flg) {
2199:     MatProductSetAlgorithm(C,(MatProductAlgorithm)algTypes[alg]);
2200:   }

2202:   C->ops->matmatmultsymbolic = MatMatMatMultSymbolic_SeqAIJ_SeqAIJ_SeqAIJ;
2203:   C->ops->productsymbolic    = MatProductSymbolic_ABC;
2204:   return(0);
2205: }

2207: PetscErrorCode MatProductSetFromOptions_SeqAIJ(Mat C)
2208: {
2210:   Mat_Product    *product = C->product;

2213:   switch (product->type) {
2214:   case MATPRODUCT_AB:
2215:     MatProductSetFromOptions_SeqAIJ_AB(C);
2216:     break;
2217:   case MATPRODUCT_AtB:
2218:     MatProductSetFromOptions_SeqAIJ_AtB(C);
2219:     break;
2220:   case MATPRODUCT_ABt:
2221:     MatProductSetFromOptions_SeqAIJ_ABt(C);
2222:     break;
2223:   case MATPRODUCT_PtAP:
2224:     MatProductSetFromOptions_SeqAIJ_PtAP(C);
2225:     break;
2226:   case MATPRODUCT_RARt:
2227:     MatProductSetFromOptions_SeqAIJ_RARt(C);
2228:     break;
2229:   case MATPRODUCT_ABC:
2230:     MatProductSetFromOptions_SeqAIJ_ABC(C);
2231:     break;
2232:   default:
2233:     break;
2234:   }
2235:   return(0);
2236: }