xref: /petsc/src/mat/impls/aij/mpi/mpiptap.c (revision d439da4237177ffb990ec3e0541c2191eed210f0)
1 
2 /*
3   Defines projective product routines where A is a MPIAIJ matrix
4           C = P^T * A * P
5 */
6 
7 #include "src/mat/impls/aij/seq/aij.h"   /*I "petscmat.h" I*/
8 #include "src/mat/utils/freespace.h"
9 #include "src/mat/impls/aij/mpi/mpiaij.h"
10 #include "petscbt.h"
11 
12 #undef __FUNCT__
13 #define __FUNCT__ "MatPtAP_MPIAIJ_MPIAIJ"
14 PetscErrorCode MatPtAP_MPIAIJ_MPIAIJ(Mat A,Mat P,MatReuse scall,PetscReal fill,Mat *C)
15 {
16   PetscErrorCode       ierr;
17 
18   PetscFunctionBegin;
19   if (scall == MAT_INITIAL_MATRIX){
20     ierr = PetscLogEventBegin(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr);
21     ierr = MatPtAPSymbolic_MPIAIJ_MPIAIJ(A,P,fill,C);CHKERRQ(ierr);
22     ierr = PetscLogEventEnd(MAT_PtAPSymbolic,A,P,0,0);CHKERRQ(ierr);
23   }
24   ierr = PetscLogEventBegin(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr);
25   ierr = MatPtAPNumeric_MPIAIJ_MPIAIJ(A,P,*C);CHKERRQ(ierr);
26   ierr = PetscLogEventEnd(MAT_PtAPNumeric,A,P,0,0);CHKERRQ(ierr);
27   PetscFunctionReturn(0);
28 }
29 
30 #undef __FUNCT__
31 #define __FUNCT__ "MatPtAPSymbolic_MPIAIJ_MPIAIJ"
32 PetscErrorCode MatPtAPSymbolic_MPIAIJ_MPIAIJ(Mat A,Mat P,PetscReal fill,Mat *C)
33 {
34   PetscErrorCode       ierr;
35   Mat                  B_mpi;
36   Mat_MatMatMultMPI    *ap;
37   PetscObjectContainer container;
38   FreeSpaceList        free_space=PETSC_NULL,current_space=PETSC_NULL;
39   Mat_MPIAIJ           *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data;
40   Mat_SeqAIJ           *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data;
41   Mat_SeqAIJ           *p_loc,*p_oth;
42   PetscInt             *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pdti,*pdtj,*poti,*potj,*ptJ;
43   PetscInt             *adi=ad->i,*adj=ad->j,*aoi=ao->i,*aoj=ao->j,nnz;
44   PetscInt             nlnk,*lnk,*owners_co,*coi,*coj,i,k,pnz,row;
45   PetscInt             am=A->m,pN=P->N,pn=P->n;
46   PetscBT              lnkbt;
47   MPI_Comm             comm=A->comm;
48   PetscMPIInt          size,rank,tag,*len_si,*len_s,*len_ri;
49   PetscInt             **buf_rj,**buf_ri,**buf_ri_k;
50   PetscInt             len,proc,*dnz,*onz,*owners;
51   PetscInt             nzi,*bi,*bj;
52   PetscInt             nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextci;
53   MPI_Request          *swaits,*rwaits;
54   MPI_Status           *sstatus,rstatus;
55   Mat_Merge_SeqsToMPI  *merge;
56   PetscInt             *api,*apj,*Jptr,apnz,*prmap=p->garray,pon;
57   PetscMPIInt          j;
58 
59   PetscFunctionBegin;
60   ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
61   ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
62 
63   /* destroy the container 'Mat_MatMatMultMPI' in case that P is attached to it */
64   ierr = PetscObjectQuery((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject *)&container);CHKERRQ(ierr);
65   if (container) {
66     ierr = PetscObjectContainerDestroy(container);CHKERRQ(ierr);
67     ierr = PetscObjectCompose((PetscObject)P,"Mat_MatMatMultMPI",0);CHKERRQ(ierr);
68   }
69 
70   /* create the container 'Mat_MatMatMultMPI' and attach it to P */
71   ierr = PetscNew(Mat_MatMatMultMPI,&ap);CHKERRQ(ierr);
72   ap->abi=PETSC_NULL; ap->abj=PETSC_NULL;
73   ap->abnz_max = 0;
74 
75   ierr = PetscObjectContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr);
76   ierr = PetscObjectContainerSetPointer(container,ap);CHKERRQ(ierr);
77   ierr = PetscObjectCompose((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject)container);CHKERRQ(ierr);
78   ap->MatDestroy  = P->ops->destroy;
79   P->ops->destroy = MatDestroy_MPIAIJ_MatMatMult;
80   ap->reuse       = MAT_INITIAL_MATRIX;
81   ierr = PetscObjectContainerSetUserDestroy(container,PetscObjectContainerDestroy_Mat_MatMatMultMPI);CHKERRQ(ierr);
82 
83   /* get P_oth by taking rows of P (= non-zero cols of local A) from other processors */
84   ierr = MatGetBrowsOfAoCols(A,P,MAT_INITIAL_MATRIX,&ap->startsj,&ap->bufa,&ap->B_oth);CHKERRQ(ierr);
85   /* get P_loc by taking all local rows of P */
86   ierr = MatGetLocalMat(P,MAT_INITIAL_MATRIX,&ap->B_loc);CHKERRQ(ierr);
87 
88   p_loc = (Mat_SeqAIJ*)(ap->B_loc)->data;
89   p_oth = (Mat_SeqAIJ*)(ap->B_oth)->data;
90   pi_loc = p_loc->i; pj_loc = p_loc->j;
91   pi_oth = p_oth->i; pj_oth = p_oth->j;
92 
93   /* first, compute symbolic AP = A_loc*P */
94   /*--------------------------------------*/
95   ierr  = PetscMalloc((am+2)*sizeof(PetscInt),&api);CHKERRQ(ierr);
96   ap->abi = api;
97   api[0] = 0;
98 
99   /* create and initialize a linked list */
100   nlnk = pN+1;
101   ierr = PetscLLCreate(pN,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
102 
103   /* Initial FreeSpace size is fill*nnz(A) */
104   ierr = GetMoreSpace((PetscInt)(fill*(adi[am]+aoi[am])),&free_space);CHKERRQ(ierr);
105   current_space = free_space;
106 
107   for (i=0;i<am;i++) {
108     apnz = 0;
109     /* diagonal portion of A */
110     nzi = adi[i+1] - adi[i];
111     for (j=0; j<nzi; j++){
112       row = *adj++;
113       pnz = pi_loc[row+1] - pi_loc[row];
114       Jptr  = pj_loc + pi_loc[row];
115       /* add non-zero cols of P into the sorted linked list lnk */
116       ierr = PetscLLAdd(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
117       apnz += nlnk;
118     }
119     /* off-diagonal portion of A */
120     nzi = aoi[i+1] - aoi[i];
121     for (j=0; j<nzi; j++){
122       row = *aoj++;
123       pnz = pi_oth[row+1] - pi_oth[row];
124       Jptr  = pj_oth + pi_oth[row];
125       ierr = PetscLLAdd(pnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
126       apnz += nlnk;
127     }
128 
129     api[i+1] = api[i] + apnz;
130     if (ap->abnz_max < apnz) ap->abnz_max = apnz;
131 
132     /* if free space is not available, double the total space in the list */
133     if (current_space->local_remaining<apnz) {
134       ierr = GetMoreSpace(current_space->total_array_size,&current_space);CHKERRQ(ierr);
135     }
136 
137     /* Copy data into free space, then initialize lnk */
138     ierr = PetscLLClean(pN,pN,apnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr);
139     current_space->array           += apnz;
140     current_space->local_used      += apnz;
141     current_space->local_remaining -= apnz;
142   }
143   /* Allocate space for apj, initialize apj, and */
144   /* destroy list of free space and other temporary array(s) */
145   ierr = PetscMalloc((api[am]+1)*sizeof(PetscInt),&ap->abj);CHKERRQ(ierr);
146   apj = ap->abj;
147   ierr = MakeSpaceContiguous(&free_space,ap->abj);CHKERRQ(ierr);
148 
149   /* determine symbolic Co=(p->B)^T*AP - send to others */
150   /*----------------------------------------------------*/
151   ierr = MatGetSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
152 
153   /* then, compute symbolic Co = (p->B)^T*AP */
154   pon = (p->B)->n; /* total num of rows to be sent to other processors
155                          >= (num of nonzero rows of C_seq) - pn */
156   ierr = PetscMalloc((pon+1)*sizeof(PetscInt),&coi);CHKERRQ(ierr);
157   coi[0] = 0;
158 
159   /* set initial free space to be 3*pon*max( nnz(AP) per row) */
160   nnz           = 3*pon*ap->abnz_max + 1;
161   ierr          = GetMoreSpace(nnz,&free_space);
162   current_space = free_space;
163 
164   for (i=0; i<pon; i++) {
165     nnz  = 0;
166     pnz = poti[i+1] - poti[i];
167     j     = pnz;
168     ptJ   = potj + poti[i+1];
169     while (j){/* assume cols are almost in increasing order, starting from its end saves computation */
170       j--; ptJ--;
171       row  = *ptJ; /* row of AP == col of Pot */
172       apnz = api[row+1] - api[row];
173       Jptr   = apj + api[row];
174       /* add non-zero cols of AP into the sorted linked list lnk */
175       ierr = PetscLLAdd(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
176       nnz += nlnk;
177     }
178 
179     /* If free space is not available, double the total space in the list */
180     if (current_space->local_remaining<nnz) {
181       ierr = GetMoreSpace(current_space->total_array_size,&current_space);CHKERRQ(ierr);
182     }
183 
184     /* Copy data into free space, and zero out denserows */
185     ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr);
186     current_space->array           += nnz;
187     current_space->local_used      += nnz;
188     current_space->local_remaining -= nnz;
189     coi[i+1] = coi[i] + nnz;
190   }
191   ierr = PetscMalloc((coi[pon]+1)*sizeof(PetscInt),&coj);CHKERRQ(ierr);
192   ierr = MakeSpaceContiguous(&free_space,coj);CHKERRQ(ierr);
193   ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->B,&poti,&potj);CHKERRQ(ierr);
194 
195   /* send j-array (coj) of Co to other processors */
196   /*----------------------------------------------*/
197   /* determine row ownership */
198   ierr = PetscNew(Mat_Merge_SeqsToMPI,&merge);CHKERRQ(ierr);
199   ierr = PetscMapCreate(comm,&merge->rowmap);CHKERRQ(ierr);
200   ierr = PetscMapSetLocalSize(merge->rowmap,pn);CHKERRQ(ierr);
201   ierr = PetscMapSetType(merge->rowmap,MAP_MPI);CHKERRQ(ierr);
202   ierr = PetscMapGetGlobalRange(merge->rowmap,&owners);CHKERRQ(ierr);
203 
204   /* determine the number of messages to send, their lengths */
205   ierr = PetscMalloc(size*sizeof(PetscMPIInt),&len_si);CHKERRQ(ierr);
206   ierr = PetscMemzero(len_si,size*sizeof(PetscMPIInt));CHKERRQ(ierr);
207   ierr = PetscMalloc(size*sizeof(PetscMPIInt),&merge->len_s);CHKERRQ(ierr);
208   len_s = merge->len_s;
209   merge->nsend = 0;
210 
211   ierr = PetscMalloc((size+2)*sizeof(PetscInt),&owners_co);CHKERRQ(ierr);
212   ierr = PetscMemzero(len_s,size*sizeof(PetscMPIInt));CHKERRQ(ierr);
213 
214   proc = 0;
215   for (i=0; i<pon; i++){
216     while (prmap[i] >= owners[proc+1]) proc++;
217     len_si[proc]++;  /* num of rows in Co to be sent to [proc] */
218     len_s[proc] += coi[i+1] - coi[i];
219   }
220 
221   len   = 0;  /* max length of buf_si[] */
222   owners_co[0] = 0;
223   for (proc=0; proc<size; proc++){
224     owners_co[proc+1] = owners_co[proc] + len_si[proc];
225     if (len_si[proc]){
226       merge->nsend++;
227       len_si[proc] = 2*(len_si[proc] + 1);
228       len += len_si[proc];
229     }
230   }
231 
232   /* determine the number and length of messages to receive for coi and coj  */
233   ierr = PetscGatherNumberOfMessages(comm,PETSC_NULL,len_s,&merge->nrecv);CHKERRQ(ierr);
234   ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr);
235 
236   /* post the Irecv and Isend of coj */
237   ierr = PetscObjectGetNewTag((PetscObject)merge->rowmap,&tag);CHKERRQ(ierr);
238   ierr = PetscPostIrecvInt(comm,tag,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rwaits);CHKERRQ(ierr);
239 
240   ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&swaits);CHKERRQ(ierr);
241 
242   for (proc=0, k=0; proc<size; proc++){
243     if (!len_s[proc]) continue;
244     i = owners_co[proc];
245     ierr = MPI_Isend(coj+coi[i],len_s[proc],MPIU_INT,proc,tag,comm,swaits+k);CHKERRQ(ierr);
246     k++;
247   }
248 
249   /* receives and sends of coj are complete */
250   ierr = PetscMalloc(size*sizeof(MPI_Status),&sstatus);CHKERRQ(ierr);
251   i = merge->nrecv;
252   while (i--) {
253     ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr);
254   }
255   ierr = PetscFree(rwaits);CHKERRQ(ierr);
256   if (merge->nsend){
257     ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);
258   }
259 
260   /* send and recv coi */
261   /*-------------------*/
262   ierr = PetscPostIrecvInt(comm,tag,merge->nrecv,merge->id_r,len_ri,&buf_ri,&rwaits);CHKERRQ(ierr);
263 
264   ierr = PetscMalloc((len+1)*sizeof(PetscInt),&buf_s);CHKERRQ(ierr);
265   buf_si = buf_s;  /* points to the beginning of k-th msg to be sent */
266   for (proc=0,k=0; proc<size; proc++){
267     if (!len_s[proc]) continue;
268     /* form outgoing message for i-structure:
269          buf_si[0]:                 nrows to be sent
270                [1:nrows]:           row index (global)
271                [nrows+1:2*nrows+1]: i-structure index
272     */
273     /*-------------------------------------------*/
274     nrows = len_si[proc]/2 - 1;
275     buf_si_i    = buf_si + nrows+1;
276     buf_si[0]   = nrows;
277     buf_si_i[0] = 0;
278     nrows = 0;
279     for (i=owners_co[proc]; i<owners_co[proc+1]; i++){
280       nzi = coi[i+1] - coi[i];
281       buf_si_i[nrows+1] = buf_si_i[nrows] + nzi; /* i-structure */
282       buf_si[nrows+1] =prmap[i] -owners[proc]; /* local row index */
283       nrows++;
284     }
285     ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tag,comm,swaits+k);CHKERRQ(ierr);
286     k++;
287     buf_si += len_si[proc];
288   }
289   i = merge->nrecv;
290   while (i--) {
291     ierr = MPI_Waitany(merge->nrecv,rwaits,&j,&rstatus);CHKERRQ(ierr);
292   }
293   ierr = PetscFree(rwaits);CHKERRQ(ierr);
294   if (merge->nsend){
295     ierr = MPI_Waitall(merge->nsend,swaits,sstatus);CHKERRQ(ierr);
296   }
297 
298   ierr = PetscLogInfo((PetscObject)A,"MatMerge_SeqsToMPI: nsend: %d, nrecv: %d\n",merge->nsend,merge->nrecv);CHKERRQ(ierr);
299   for (i=0; i<merge->nrecv; i++){
300     ierr = PetscLogInfo((PetscObject)A,"MatMerge_SeqsToMPI:   recv len_ri=%d, len_rj=%d from [%d]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);CHKERRQ(ierr);
301   }
302 
303   ierr = PetscFree(len_si);CHKERRQ(ierr);
304   ierr = PetscFree(len_ri);CHKERRQ(ierr);
305   ierr = PetscFree(swaits);CHKERRQ(ierr);
306   ierr = PetscFree(sstatus);CHKERRQ(ierr);
307   ierr = PetscFree(buf_s);CHKERRQ(ierr);
308 
309   /* compute the local portion of C (mpi mat) */
310   /*------------------------------------------*/
311   ierr = MatGetSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
312 
313   /* allocate bi array and free space for accumulating nonzero column info */
314   ierr = PetscMalloc((pn+1)*sizeof(PetscInt),&bi);CHKERRQ(ierr);
315   bi[0] = 0;
316 
317   /* set initial free space to be 3*pn*max( nnz(AP) per row) */
318   nnz           = 3*pn*ap->abnz_max + 1;
319   ierr          = GetMoreSpace(nnz,&free_space);
320   current_space = free_space;
321 
322   ierr = PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);CHKERRQ(ierr);
323   nextrow = buf_ri_k + merge->nrecv;
324   nextci  = nextrow + merge->nrecv;
325   for (k=0; k<merge->nrecv; k++){
326     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
327     nrows       = *buf_ri_k[k];
328     nextrow[k]  = buf_ri_k[k] + 1;  /* next row number of k-th recved i-structure */
329     nextci[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
330   }
331   ierr = MatPreallocateInitialize(comm,pn,pn,dnz,onz);CHKERRQ(ierr);
332   for (i=0; i<pn; i++) {
333     /* add pdt[i,:]*AP into lnk */
334     nnz = 0;
335     pnz  = pdti[i+1] - pdti[i];
336     j    = pnz;
337     ptJ  = pdtj + pdti[i+1];
338     while (j){/* assume cols are almost in increasing order, starting from its end saves computation */
339       j--; ptJ--;
340       row  = *ptJ; /* row of AP == col of Pt */
341       apnz = api[row+1] - api[row];
342       Jptr   = apj + api[row];
343       /* add non-zero cols of AP into the sorted linked list lnk */
344       ierr = PetscLLAdd(apnz,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
345       nnz += nlnk;
346     }
347     /* add received col data into lnk */
348     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
349       if (i == *nextrow[k]) { /* i-th row */
350         nzi = *(nextci[k]+1) - *nextci[k];
351         Jptr  = buf_rj[k] + *nextci[k];
352         ierr = PetscLLAdd(nzi,Jptr,pN,nlnk,lnk,lnkbt);CHKERRQ(ierr);
353         nnz += nlnk;
354         nextrow[k]++; nextci[k]++;
355       }
356     }
357 
358     /* if free space is not available, make more free space */
359     if (current_space->local_remaining<nnz) {
360       ierr = GetMoreSpace(current_space->total_array_size,&current_space);CHKERRQ(ierr);
361     }
362     /* copy data into free space, then initialize lnk */
363     ierr = PetscLLClean(pN,pN,nnz,lnk,current_space->array,lnkbt);CHKERRQ(ierr);
364     ierr = MatPreallocateSet(i+owners[rank],nnz,current_space->array,dnz,onz);CHKERRQ(ierr);
365     current_space->array           += nnz;
366     current_space->local_used      += nnz;
367     current_space->local_remaining -= nnz;
368     bi[i+1] = bi[i] + nnz;
369   }
370   ierr = MatRestoreSymbolicTranspose_SeqAIJ(p->A,&pdti,&pdtj);CHKERRQ(ierr);
371   ierr = PetscFree(buf_ri_k);CHKERRQ(ierr);
372 
373   ierr = PetscMalloc((bi[pn]+1)*sizeof(PetscInt),&bj);CHKERRQ(ierr);
374   ierr = MakeSpaceContiguous(&free_space,bj);CHKERRQ(ierr);
375   ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr);
376 
377   /* create symbolic parallel matrix B_mpi */
378   /*---------------------------------------*/
379   ierr = MatCreate(comm,pn,pn,PETSC_DETERMINE,PETSC_DETERMINE,&B_mpi);CHKERRQ(ierr);
380   ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr);
381   ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr);
382   ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr);
383 
384   /* B_mpi is not ready for use - assembly will be done by MatPtAPNumeric() */
385   B_mpi->assembled     = PETSC_FALSE;
386   B_mpi->ops->destroy  = MatDestroy_MPIAIJ_SeqsToMPI;
387   merge->bi            = bi;
388   merge->bj            = bj;
389   merge->coi           = coi;
390   merge->coj           = coj;
391   merge->buf_ri        = buf_ri;
392   merge->buf_rj        = buf_rj;
393   merge->owners_co     = owners_co;
394 
395   /* attach the supporting struct to B_mpi for reuse */
396   ierr = PetscObjectContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr);
397   ierr = PetscObjectContainerSetPointer(container,merge);CHKERRQ(ierr);
398   ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr);
399   *C = B_mpi;
400   PetscFunctionReturn(0);
401 }
402 
403 #undef __FUNCT__
404 #define __FUNCT__ "MatPtAPNumeric_MPIAIJ_MPIAIJ"
405 PetscErrorCode MatPtAPNumeric_MPIAIJ_MPIAIJ(Mat A,Mat P,Mat C)
406 {
407   PetscErrorCode       ierr;
408   Mat_Merge_SeqsToMPI  *merge;
409   Mat_MatMatMultMPI    *ap;
410   PetscObjectContainer cont_merge,cont_ptap;
411   Mat_MPIAIJ     *a=(Mat_MPIAIJ*)A->data,*p=(Mat_MPIAIJ*)P->data;
412   Mat_SeqAIJ     *ad=(Mat_SeqAIJ*)(a->A)->data,*ao=(Mat_SeqAIJ*)(a->B)->data;
413   Mat_SeqAIJ     *pd=(Mat_SeqAIJ*)(p->A)->data,*po=(Mat_SeqAIJ*)(p->B)->data;
414   Mat_SeqAIJ     *p_loc,*p_oth;
415   PetscInt       *adi=ad->i,*aoi=ao->i,*adj=ad->j,*aoj=ao->j,*apJ,nextp,flops=0;
416   PetscInt       *pi_loc,*pj_loc,*pi_oth,*pj_oth,*pJ,*pj;
417   PetscInt       i,j,k,anz,pnz,apnz,nextap,row,*cj;
418   MatScalar      *ada=ad->a,*aoa=ao->a,*apa,*pa,*ca,*pa_loc,*pa_oth;
419   PetscInt       am=A->m,cm=C->m,pon=(p->B)->n;
420   MPI_Comm       comm=C->comm;
421   PetscMPIInt    size,rank,taga,*len_s;
422   PetscInt       *owners,proc,nrows,**buf_ri_k,**nextrow,**nextci;
423   PetscInt       **buf_ri,**buf_rj;
424   PetscInt       cnz=0,*bj_i,*bi,*bj,bnz,nextcj; /* bi,bj,ba: local array of C(mpi mat) */
425   MPI_Request    *s_waits,*r_waits;
426   MPI_Status     *status;
427   MatScalar      **abuf_r,*ba_i,*pA,*coa,*ba;
428   PetscInt       *api,*apj,*coi,*coj;
429   PetscInt       *poJ=po->j,*pdJ=pd->j,pcstart=p->cstart,pcend=p->cend;
430 
431   PetscFunctionBegin;
432   ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
433   ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
434 
435   ierr = PetscObjectQuery((PetscObject)C,"MatMergeSeqsToMPI",(PetscObject *)&cont_merge);CHKERRQ(ierr);
436   if (cont_merge) {
437     ierr  = PetscObjectContainerGetPointer(cont_merge,(void **)&merge);CHKERRQ(ierr);
438   } else {
439     SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix C does not posses an object container");
440   }
441 
442   ierr = PetscObjectQuery((PetscObject)P,"Mat_MatMatMultMPI",(PetscObject *)&cont_ptap);CHKERRQ(ierr);
443   if (cont_ptap) {
444     ierr  = PetscObjectContainerGetPointer(cont_ptap,(void **)&ap);CHKERRQ(ierr);
445     if (ap->reuse == MAT_INITIAL_MATRIX){
446       ap->reuse = MAT_REUSE_MATRIX;
447     } else { /* update numerical values of P_oth and P_loc */
448       ierr = MatGetBrowsOfAoCols(A,P,MAT_REUSE_MATRIX,&ap->startsj,&ap->bufa,&ap->B_oth);CHKERRQ(ierr);
449       ierr = MatGetLocalMat(P,MAT_REUSE_MATRIX,&ap->B_loc);CHKERRQ(ierr);
450     }
451   } else {
452     SETERRQ(PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
453   }
454 
455   /* get data from symbolic products */
456   p_loc = (Mat_SeqAIJ*)(ap->B_loc)->data;
457   p_oth = (Mat_SeqAIJ*)(ap->B_oth)->data;
458   pi_loc=p_loc->i; pj_loc=p_loc->j; pJ=pj_loc; pa_loc=p_loc->a,pA=pa_loc;
459   pi_oth=p_oth->i; pj_oth=p_oth->j; pa_oth=p_oth->a;
460 
461   coi = merge->coi; coj = merge->coj;
462   ierr = PetscMalloc((coi[pon]+1)*sizeof(MatScalar),&coa);CHKERRQ(ierr);
463   ierr = PetscMemzero(coa,coi[pon]*sizeof(MatScalar));CHKERRQ(ierr);
464 
465   bi = merge->bi; bj = merge->bj;
466   ierr  = PetscMapGetGlobalRange(merge->rowmap,&owners);CHKERRQ(ierr);
467   ierr  = PetscMalloc((bi[cm]+1)*sizeof(MatScalar),&ba);CHKERRQ(ierr);
468   ierr  = PetscMemzero(ba,bi[cm]*sizeof(MatScalar));CHKERRQ(ierr);
469 
470   /* get data from symbolic A*P */
471   ierr = PetscMalloc((ap->abnz_max+1)*sizeof(MatScalar),&apa);CHKERRQ(ierr);
472   ierr = PetscMemzero(apa,ap->abnz_max*sizeof(MatScalar));CHKERRQ(ierr);
473 
474   /* compute numeric C_seq=P_loc^T*A_loc*P */
475   api = ap->abi; apj = ap->abj;
476   for (i=0;i<am;i++) {
477     /* form i-th sparse row of A*P */
478     apnz = api[i+1] - api[i];
479     apJ  = apj + api[i];
480     /* diagonal portion of A */
481     anz  = adi[i+1] - adi[i];
482     for (j=0;j<anz;j++) {
483       row = *adj++;
484       pnz = pi_loc[row+1] - pi_loc[row];
485       pj  = pj_loc + pi_loc[row];
486       pa  = pa_loc + pi_loc[row];
487       nextp = 0;
488       for (k=0; nextp<pnz; k++) {
489         if (apJ[k] == pj[nextp]) { /* col of AP == col of P */
490           apa[k] += (*ada)*pa[nextp++];
491         }
492       }
493       flops += 2*pnz;
494       ada++;
495     }
496     /* off-diagonal portion of A */
497     anz  = aoi[i+1] - aoi[i];
498     for (j=0; j<anz; j++) {
499       row = *aoj++;
500       pnz = pi_oth[row+1] - pi_oth[row];
501       pj  = pj_oth + pi_oth[row];
502       pa  = pa_oth + pi_oth[row];
503       nextp = 0;
504       for (k=0; nextp<pnz; k++) {
505         if (apJ[k] == pj[nextp]) { /* col of AP == col of P */
506           apa[k] += (*aoa)*pa[nextp++];
507         }
508       }
509       flops += 2*pnz;
510       aoa++;
511     }
512 
513     /* Compute P_loc[i,:]^T*AP[i,:] using outer product */
514     pnz = pi_loc[i+1] - pi_loc[i];
515     for (j=0; j<pnz; j++) {
516       nextap = 0;
517       row    = *pJ++; /* global index */
518       if (row < pcstart || row >=pcend) { /* put the value into Co */
519         cj  = coj + coi[*poJ];
520         ca  = coa + coi[*poJ++];
521       } else {                            /* put the value into Cd */
522         cj   = bj + bi[*pdJ];
523         ca   = ba + bi[*pdJ++];
524       }
525       for (k=0; nextap<apnz; k++) {
526         if (cj[k]==apJ[nextap]) ca[k] += (*pA)*apa[nextap++];
527       }
528       flops += 2*apnz;
529       pA++;
530     }
531 
532     /* zero the current row info for A*P */
533     ierr = PetscMemzero(apa,apnz*sizeof(MatScalar));CHKERRQ(ierr);
534   }
535   ierr = PetscFree(apa);CHKERRQ(ierr);
536 
537   /* send and recv matrix values */
538   /*-----------------------------*/
539   buf_ri = merge->buf_ri;
540   buf_rj = merge->buf_rj;
541   len_s  = merge->len_s;
542   ierr = PetscObjectGetNewTag((PetscObject)merge->rowmap,&taga);CHKERRQ(ierr);
543   ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr);
544 
545   ierr = PetscMalloc((merge->nsend+1)*sizeof(MPI_Request),&s_waits);CHKERRQ(ierr);
546   for (proc=0,k=0; proc<size; proc++){
547     if (!len_s[proc]) continue;
548     i = merge->owners_co[proc];
549     ierr = MPI_Isend(coa+coi[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr);
550     k++;
551   }
552   ierr = PetscMalloc(size*sizeof(MPI_Status),&status);CHKERRQ(ierr);
553   ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);
554   ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);
555   ierr = PetscFree(status);CHKERRQ(ierr);
556 
557   ierr = PetscFree(s_waits);CHKERRQ(ierr);
558   ierr = PetscFree(r_waits);CHKERRQ(ierr);
559   ierr = PetscFree(coa);CHKERRQ(ierr);
560 
561   /* insert local and received values into C */
562   /*-----------------------------------------*/
563   ierr = PetscMalloc((3*merge->nrecv+1)*sizeof(PetscInt**),&buf_ri_k);CHKERRQ(ierr);
564   nextrow   = buf_ri_k + merge->nrecv;
565   nextci = nextrow + merge->nrecv;
566 
567   for (k=0; k<merge->nrecv; k++){
568     buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
569     nrows       = *(buf_ri_k[k]);
570     nextrow[k]  = buf_ri_k[k]+1;  /* next row number of k-th recved i-structure */
571     nextci[k]   = buf_ri_k[k] + (nrows + 1);/* poins to the next i-structure of k-th recved i-structure  */
572   }
573 
574   for (i=0; i<cm; i++) {
575     row = owners[rank] + i; /* global row index of C_seq */
576     bj_i = bj + bi[i];  /* col indices of the i-th row of C */
577     ba_i = ba + bi[i];
578     bnz  = bi[i+1] - bi[i];
579     /* add received vals into ba */
580     for (k=0; k<merge->nrecv; k++){ /* k-th received message */
581       /* i-th row */
582       if (i == *nextrow[k]) {
583         cnz = *(nextci[k]+1) - *nextci[k];
584         cj  = buf_rj[k] + *(nextci[k]);
585         ca  = abuf_r[k] + *(nextci[k]);
586         nextcj = 0;
587         for (j=0; nextcj<cnz; j++){
588           if (bj_i[j] == cj[nextcj]){ /* bcol == ccol */
589             ba_i[j] += ca[nextcj++];
590           }
591         }
592         nextrow[k]++; nextci[k]++;
593       }
594     }
595     ierr = MatSetValues(C,1,&row,bnz,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr);
596     flops += 2*cnz;
597   }
598   ierr = MatSetOption(C,MAT_COLUMNS_SORTED);CHKERRQ(ierr); /* -cause delay? */
599   ierr = MatAssemblyBegin(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
600   ierr = MatAssemblyEnd(C,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr);
601 
602   ierr = PetscFree(ba);CHKERRQ(ierr);
603   ierr = PetscFree(abuf_r);CHKERRQ(ierr);
604   ierr = PetscFree(buf_ri_k);CHKERRQ(ierr);
605   ierr = PetscLogFlops(flops);CHKERRQ(ierr);
606   PetscFunctionReturn(0);
607 }
608