Actual source code: mpimatmatmult.c

  1: #define PETSCMAT_DLL

  3: /*
  4:   Defines matrix-matrix product routines for pairs of MPIAIJ matrices
  5:           C = A * B
  6: */
 7:  #include ../src/mat/impls/aij/seq/aij.h
 8:  #include ../src/mat/utils/freespace.h
 9:  #include ../src/mat/impls/aij/mpi/mpiaij.h
 10:  #include petscbt.h
 11:  #include ../src/mat/impls/dense/mpi/mpidense.h

 15: PetscErrorCode MatMatMult_MPIAIJ_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill, Mat *C)
 16: {

 20:   if (scall == MAT_INITIAL_MATRIX){
 21:     MatMatMultSymbolic_MPIAIJ_MPIAIJ(A,B,fill,C);/* numeric product is computed as well */
 22:   } else if (scall == MAT_REUSE_MATRIX){
 23:     MatMatMultNumeric_MPIAIJ_MPIAIJ(A,B,*C);
 24:   } else {
 25:     SETERRQ1(PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
 26:   }
 27:   return(0);
 28: }

 32: PetscErrorCode PetscContainerDestroy_Mat_MatMatMultMPI(void *ptr)
 33: {
 34:   PetscErrorCode       ierr;
 35:   Mat_MatMatMultMPI    *mult=(Mat_MatMatMultMPI*)ptr;

 38:   PetscFree(mult->startsj);
 39:   PetscFree(mult->bufa);
 40:   if (mult->isrowa){ISDestroy(mult->isrowa);}
 41:   if (mult->isrowb){ISDestroy(mult->isrowb);}
 42:   if (mult->iscolb){ISDestroy(mult->iscolb);}
 43:   if (mult->C_seq){MatDestroy(mult->C_seq);}
 44:   if (mult->A_loc){MatDestroy(mult->A_loc); }
 45:   if (mult->B_seq){MatDestroy(mult->B_seq);}
 46:   if (mult->B_loc){MatDestroy(mult->B_loc);}
 47:   if (mult->B_oth){MatDestroy(mult->B_oth);}
 48:   PetscFree(mult->abi);
 49:   PetscFree(mult->abj);
 50:   PetscFree(mult);
 51:   return(0);
 52: }

 54: EXTERN PetscErrorCode MatDestroy_AIJ(Mat);
 57: PetscErrorCode MatDestroy_MPIAIJ_MatMatMult(Mat A)
 58: {
 59:   PetscErrorCode     ierr;
 60:   PetscContainer     container;
 61:   Mat_MatMatMultMPI  *mult=PETSC_NULL;

 64:   PetscObjectQuery((PetscObject)A,"Mat_MatMatMultMPI",(PetscObject *)&container);
 65:   if (container) {
 66:     PetscContainerGetPointer(container,(void **)&mult);
 67:   } else {
 68:     SETERRQ(PETSC_ERR_PLIB,"Container does not exit");
 69:   }
 70:   A->ops->destroy = mult->MatDestroy;
 71:   PetscObjectCompose((PetscObject)A,"Mat_MatMatMultMPI",0);
 72:   (*A->ops->destroy)(A);
 73:   PetscContainerDestroy(container);
 74:   return(0);
 75: }

 79: PetscErrorCode MatDuplicate_MPIAIJ_MatMatMult(Mat A, MatDuplicateOption op, Mat *M) {
 80:   PetscErrorCode     ierr;
 81:   Mat_MatMatMultMPI  *mult;
 82:   PetscContainer     container;

 85:   PetscObjectQuery((PetscObject)A,"Mat_MatMatMultMPI",(PetscObject *)&container);
 86:   if (container) {
 87:     PetscContainerGetPointer(container,(void **)&mult);
 88:   } else {
 89:     SETERRQ(PETSC_ERR_PLIB,"Container does not exit");
 90:   }
 91:   /* Note: the container is not duplicated, because it requires deep copying of
 92:      several large data sets (see PetscContainerDestroy_Mat_MatMatMultMPI()).
 93:      These data sets are only used for repeated calling of MatMatMultNumeric(). 
 94:      *M is unlikely being used in this way. Thus we create *M with pure mpiaij format */
 95:   (*mult->MatDuplicate)(A,op,M);
 96:   (*M)->ops->destroy   = mult->MatDestroy;   /* = MatDestroy_MPIAIJ, *M doesn't duplicate A's container! */
 97:   (*M)->ops->duplicate = mult->MatDuplicate; /* = MatDuplicate_MPIAIJ */
 98:   return(0);
 99: }

103: PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
104: {
105:   PetscErrorCode     ierr;
106:   PetscInt           start,end;
107:   Mat_MatMatMultMPI  *mult;
108:   PetscContainer     container;
109: 
111:   if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend){
112:     SETERRQ4(PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
113:   }
114:   PetscNew(Mat_MatMatMultMPI,&mult);

116:   /* create a seq matrix B_seq = submatrix of B by taking rows of B that equal to nonzero col of A */
117:   MatGetBrowsOfAcols(A,B,MAT_INITIAL_MATRIX,&mult->isrowb,&mult->iscolb,&mult->brstart,&mult->B_seq);

119:   /*  create a seq matrix A_seq = submatrix of A by taking all local rows of A */
120:   start = A->rmap->rstart; end = A->rmap->rend;
121:   ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&mult->isrowa);
122:   MatGetLocalMatCondensed(A,MAT_INITIAL_MATRIX,&mult->isrowa,&mult->isrowb,&mult->A_loc);

124:   /* compute C_seq = A_seq * B_seq */
125:   MatMatMult_SeqAIJ_SeqAIJ(mult->A_loc,mult->B_seq,MAT_INITIAL_MATRIX,fill,&mult->C_seq);

127:   /* create mpi matrix C by concatinating C_seq */
128:   PetscObjectReference((PetscObject)mult->C_seq); /* prevent C_seq being destroyed by MatMerge() */
129:   MatMerge(((PetscObject)A)->comm,mult->C_seq,B->cmap->n,MAT_INITIAL_MATRIX,C);
130: 
131:   /* attach the supporting struct to C for reuse of symbolic C */
132:   PetscContainerCreate(PETSC_COMM_SELF,&container);
133:   PetscContainerSetPointer(container,mult);
134:   PetscObjectCompose((PetscObject)(*C),"Mat_MatMatMultMPI",(PetscObject)container);
135:   PetscContainerSetUserDestroy(container,PetscContainerDestroy_Mat_MatMatMultMPI);
136:   mult->MatDestroy   = (*C)->ops->destroy;
137:   mult->MatDuplicate = (*C)->ops->duplicate;

139:   (*C)->ops->destroy   = MatDestroy_MPIAIJ_MatMatMult;
140:   (*C)->ops->duplicate = MatDuplicate_MPIAIJ_MatMatMult;
141:   return(0);
142: }

144: /* This routine is called ONLY in the case of reusing previously computed symbolic C */
147: PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIAIJ(Mat A,Mat B,Mat C)
148: {
149:   PetscErrorCode     ierr;
150:   Mat                *seq;
151:   Mat_MatMatMultMPI  *mult;
152:   PetscContainer     container;

155:   PetscObjectQuery((PetscObject)C,"Mat_MatMatMultMPI",(PetscObject *)&container);
156:   if (container) {
157:     PetscContainerGetPointer(container,(void **)&mult);
158:   } else {
159:     SETERRQ(PETSC_ERR_PLIB,"Container does not exit");
160:   }

162:   seq = &mult->B_seq;
163:   MatGetSubMatrices(B,1,&mult->isrowb,&mult->iscolb,MAT_REUSE_MATRIX,&seq);
164:   mult->B_seq = *seq;
165: 
166:   seq = &mult->A_loc;
167:   MatGetSubMatrices(A,1,&mult->isrowa,&mult->isrowb,MAT_REUSE_MATRIX,&seq);
168:   mult->A_loc = *seq;

170:   MatMatMult_SeqAIJ_SeqAIJ(mult->A_loc,mult->B_seq,MAT_REUSE_MATRIX,0.0,&mult->C_seq);

172:   PetscObjectReference((PetscObject)mult->C_seq);
173:   MatMerge(((PetscObject)A)->comm,mult->C_seq,B->cmap->n,MAT_REUSE_MATRIX,&C);
174:   return(0);
175: }

179: PetscErrorCode MatMatMult_MPIAIJ_MPIDense(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
180: {

184:   if (scall == MAT_INITIAL_MATRIX){
185:     MatMatMultSymbolic_MPIAIJ_MPIDense(A,B,fill,C);
186:   }
187:   MatMatMultNumeric_MPIAIJ_MPIDense(A,B,*C);
188:   return(0);
189: }

191: typedef struct {
192:   Mat         workB;
193:   PetscScalar *rvalues,*svalues;
194:   MPI_Request *rwaits,*swaits;
195: } MPIAIJ_MPIDense;

197: PetscErrorCode MPIAIJ_MPIDenseDestroy(void *ctx)
198: {
199:   MPIAIJ_MPIDense *contents = (MPIAIJ_MPIDense*) ctx;
200:   PetscErrorCode  ierr;

203:   if (contents->workB) {MatDestroy(contents->workB);}
204:   PetscFree4(contents->rvalues,contents->svalues,contents->rwaits,contents->swaits);
205:   PetscFree(contents);
206:   return(0);
207: }

211: PetscErrorCode MatMatMultSymbolic_MPIAIJ_MPIDense(Mat A,Mat B,PetscReal fill,Mat *C)
212: {
213:   PetscErrorCode         ierr;
214:   Mat_MPIAIJ             *aij = (Mat_MPIAIJ*) A->data;
215:   PetscInt               nz = aij->B->cmap->n;
216:   PetscContainer         cont;
217:   MPIAIJ_MPIDense        *contents;
218:   VecScatter             ctx = aij->Mvctx;
219:   VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata;
220:   VecScatter_MPI_General *to   = ( VecScatter_MPI_General*) ctx->todata;
221:   PetscInt               m=A->rmap->n,n=B->cmap->n;

224:   MatCreate(((PetscObject)B)->comm,C);
225:   MatSetSizes(*C,m,n,A->rmap->N,B->cmap->N);
226:   MatSetType(*C,MATMPIDENSE);
227:   MatAssemblyBegin(*C,MAT_FINAL_ASSEMBLY);
228:   MatAssemblyEnd(*C,MAT_FINAL_ASSEMBLY);

230:   PetscContainerCreate(((PetscObject)A)->comm,&cont);
231:   PetscNew(MPIAIJ_MPIDense,&contents);
232:   PetscContainerSetPointer(cont,contents);
233:   PetscContainerSetUserDestroy(cont,MPIAIJ_MPIDenseDestroy);

235:   /* Create work matrix used to store off processor rows of B needed for local product */
236:   MatCreateSeqDense(PETSC_COMM_SELF,nz,B->cmap->N,PETSC_NULL,&contents->workB);

238:   /* Create work arrays needed */
239:   PetscMalloc4(B->cmap->N*from->starts[from->n],PetscScalar,&contents->rvalues,
240:                       B->cmap->N*to->starts[to->n],PetscScalar,&contents->svalues,
241:                       from->n,MPI_Request,&contents->rwaits,
242:                       to->n,MPI_Request,&contents->swaits);

244:   PetscObjectCompose((PetscObject)(*C),"workB",(PetscObject)cont);
245:   PetscContainerDestroy(cont);
246:   return(0);
247: }

249: /*
250:     Performs an efficient scatter on the rows of B needed by this process; this is
251:     a modification of the VecScatterBegin_() routines.
252: */
253: PetscErrorCode MatMPIDenseScatter(Mat A,Mat B,Mat C,Mat *outworkB)
254: {
255:   Mat_MPIAIJ             *aij = (Mat_MPIAIJ*)A->data;
256:   PetscErrorCode         ierr;
257:   PetscScalar            *b,*w,*svalues,*rvalues;
258:   VecScatter             ctx = aij->Mvctx;
259:   VecScatter_MPI_General *from = (VecScatter_MPI_General*) ctx->fromdata;
260:   VecScatter_MPI_General *to   = ( VecScatter_MPI_General*) ctx->todata;
261:   PetscInt               i,j,k;
262:   PetscInt               *sindices,*sstarts,*rindices,*rstarts;
263:   PetscMPIInt            *sprocs,*rprocs,nrecvs;
264:   MPI_Request            *swaits,*rwaits;
265:   MPI_Comm               comm = ((PetscObject)A)->comm;
266:   PetscMPIInt            tag = ((PetscObject)ctx)->tag,ncols = B->cmap->N, nrows = aij->B->cmap->n,imdex,nrowsB = B->rmap->n;
267:   MPI_Status             status;
268:   MPIAIJ_MPIDense        *contents;
269:   PetscContainer         cont;
270:   Mat                    workB;

273:   PetscObjectQuery((PetscObject)C,"workB",(PetscObject*)&cont);
274:   PetscContainerGetPointer(cont,(void**)&contents);

276:   workB = *outworkB = contents->workB;
277:   if (nrows != workB->rmap->n) SETERRQ2(PETSC_ERR_PLIB,"Number of rows of workB %D not equal to columns of aij->B %D",nrows,workB->cmap->n);
278:   sindices  = to->indices;
279:   sstarts   = to->starts;
280:   sprocs    = to->procs;
281:   swaits    = contents->swaits;
282:   svalues   = contents->svalues;

284:   rindices  = from->indices;
285:   rstarts   = from->starts;
286:   rprocs    = from->procs;
287:   rwaits    = contents->rwaits;
288:   rvalues   = contents->rvalues;

290:   MatGetArray(B,&b);
291:   MatGetArray(workB,&w);

293:   for (i=0; i<from->n; i++) {
294:     MPI_Irecv(rvalues+ncols*rstarts[i],ncols*(rstarts[i+1]-rstarts[i]),MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
295:   }

297:   for (i=0; i<to->n; i++) {
298:     /* pack a message at a time */
299:     CHKMEMQ;
300:     for (j=0; j<sstarts[i+1]-sstarts[i]; j++){
301:       for (k=0; k<ncols; k++) {
302:         svalues[ncols*(sstarts[i] + j) + k] = b[sindices[sstarts[i]+j] + nrowsB*k];
303:       }
304:     }
305:     CHKMEMQ;
306:     MPI_Isend(svalues+ncols*sstarts[i],ncols*(sstarts[i+1]-sstarts[i]),MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
307:   }

309:   nrecvs = from->n;
310:   while (nrecvs) {
311:     MPI_Waitany(from->n,rwaits,&imdex,&status);
312:     nrecvs--;
313:     /* unpack a message at a time */
314:     CHKMEMQ;
315:     for (j=0; j<rstarts[imdex+1]-rstarts[imdex]; j++){
316:       for (k=0; k<ncols; k++) {
317:         w[rindices[rstarts[imdex]+j] + nrows*k] = rvalues[ncols*(rstarts[imdex] + j) + k];
318:       }
319:     }
320:     CHKMEMQ;
321:   }
322:   if (to->n) {MPI_Waitall(to->n,swaits,to->sstatus);CHKERRQ(ierr)}

324:   MatRestoreArray(B,&b);
325:   MatRestoreArray(workB,&w);
326:   MatAssemblyBegin(workB,MAT_FINAL_ASSEMBLY);
327:   MatAssemblyEnd(workB,MAT_FINAL_ASSEMBLY);
328:   return(0);
329: }

334: PetscErrorCode MatMatMultNumeric_MPIAIJ_MPIDense(Mat A,Mat B,Mat C)
335: {
336:   PetscErrorCode       ierr;
337:   Mat_MPIAIJ           *aij = (Mat_MPIAIJ*)A->data;
338:   Mat_MPIDense         *bdense = (Mat_MPIDense*)B->data;
339:   Mat_MPIDense         *cdense = (Mat_MPIDense*)C->data;
340:   Mat                  workB;


344:   /* diagonal block of A times all local rows of B*/
345:   MatMatMultNumeric_SeqAIJ_SeqDense(aij->A,bdense->A,cdense->A);

347:   /* get off processor parts of B needed to complete the product */
348:   MatMPIDenseScatter(A,B,C,&workB);

350:   /* off-diagonal block of A times nonlocal rows of B */
351:   MatMatMultNumericAdd_SeqAIJ_SeqDense(aij->B,workB,cdense->A);
352:   MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);
353:   MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);
354:   return(0);
355: }