xref: /petsc/src/mat/impls/sbaij/mpi/sbaijov.c (revision 4a69c53646f303dd9d73c0f6695e42dcda7997b2)
1632d0f97SHong Zhang /*$Id: sbaijov.c,v 1.65 2001/08/06 21:15:42 bsmith Exp $*/
2632d0f97SHong Zhang 
3632d0f97SHong Zhang /*
4632d0f97SHong Zhang    Routines to compute overlapping regions of a parallel MPI matrix.
5632d0f97SHong Zhang    Used for finding submatrices that were shared across processors.
6632d0f97SHong Zhang */
7632d0f97SHong Zhang #include "src/mat/impls/sbaij/mpi/mpisbaij.h"
8632d0f97SHong Zhang #include "petscbt.h"
9632d0f97SHong Zhang 
10632d0f97SHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat,int,IS *);
11bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat,int *,int,int **,PetscBT*);
12632d0f97SHong Zhang 
13632d0f97SHong Zhang /* this function is sasme as MatCompressIndicesGeneral_MPIBAIJ -- should be removed! */
14632d0f97SHong Zhang #undef __FUNCT__
15632d0f97SHong Zhang #define __FUNCT__ "MatCompressIndicesGeneral_MPISBAIJ"
16632d0f97SHong Zhang static int MatCompressIndicesGeneral_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[])
17632d0f97SHong Zhang {
18632d0f97SHong Zhang   Mat_MPISBAIJ        *baij = (Mat_MPISBAIJ*)C->data;
19632d0f97SHong Zhang   int                ierr,isz,bs = baij->bs,n,i,j,*idx,ival;
20632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
21632d0f97SHong Zhang   PetscTable         gid1_lid1;
22632d0f97SHong Zhang   int                tt, gid1, *nidx;
23632d0f97SHong Zhang   PetscTablePosition tpos;
24632d0f97SHong Zhang #else
25632d0f97SHong Zhang   int                Nbs,*nidx;
26632d0f97SHong Zhang   PetscBT            table;
27632d0f97SHong Zhang #endif
28632d0f97SHong Zhang 
29632d0f97SHong Zhang   PetscFunctionBegin;
30632d0f97SHong Zhang   /* printf(" ...MatCompressIndicesGeneral_MPISBAIJ is called ...\n"); */
31632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
32632d0f97SHong Zhang   ierr = PetscTableCreate(baij->mbs,&gid1_lid1);CHKERRQ(ierr);
33632d0f97SHong Zhang #else
34632d0f97SHong Zhang   Nbs  = baij->Nbs;
35632d0f97SHong Zhang   ierr = PetscMalloc((Nbs+1)*sizeof(int),&nidx);CHKERRQ(ierr);
36632d0f97SHong Zhang   ierr = PetscBTCreate(Nbs,table);CHKERRQ(ierr);
37632d0f97SHong Zhang #endif
38632d0f97SHong Zhang   for (i=0; i<imax; i++) {
39632d0f97SHong Zhang     isz  = 0;
40632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
41632d0f97SHong Zhang     ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
42632d0f97SHong Zhang #else
43632d0f97SHong Zhang     ierr = PetscBTMemzero(Nbs,table);CHKERRQ(ierr);
44632d0f97SHong Zhang #endif
45632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
46632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
47632d0f97SHong Zhang     for (j=0; j<n ; j++) {
48632d0f97SHong Zhang       ival = idx[j]/bs; /* convert the indices into block indices */
49632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
50632d0f97SHong Zhang       ierr = PetscTableFind(gid1_lid1,ival+1,&tt);CHKERRQ(ierr);
51632d0f97SHong Zhang       if (!tt) {
52632d0f97SHong Zhang 	ierr = PetscTableAdd(gid1_lid1,ival+1,isz+1);CHKERRQ(ierr);
53632d0f97SHong Zhang         isz++;
54632d0f97SHong Zhang       }
55632d0f97SHong Zhang #else
56632d0f97SHong Zhang       if (ival>Nbs) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than mat-dim");
57632d0f97SHong Zhang       if(!PetscBTLookupSet(table,ival)) { nidx[isz++] = ival;}
58632d0f97SHong Zhang #endif
59632d0f97SHong Zhang     }
60632d0f97SHong Zhang     ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr);
61632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
62632d0f97SHong Zhang     ierr = PetscMalloc((isz+1)*sizeof(int),&nidx);CHKERRQ(ierr);
63632d0f97SHong Zhang     ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
64632d0f97SHong Zhang     j = 0;
65632d0f97SHong Zhang     while (tpos) {
66632d0f97SHong Zhang       ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid1,&tt);CHKERRQ(ierr);
67632d0f97SHong Zhang       if (tt-- > isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than array-dim"); }
68632d0f97SHong Zhang       nidx[tt] = gid1 - 1;
69632d0f97SHong Zhang       j++;
70632d0f97SHong Zhang     }
71632d0f97SHong Zhang     if (j != isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"table error: jj != isz"); }
72632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr);
73632d0f97SHong Zhang     ierr = PetscFree(nidx);CHKERRQ(ierr);
74632d0f97SHong Zhang #else
75632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr);
76632d0f97SHong Zhang #endif
77632d0f97SHong Zhang   }
78632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
79632d0f97SHong Zhang   ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr);
80632d0f97SHong Zhang #else
81632d0f97SHong Zhang   ierr = PetscBTDestroy(table);CHKERRQ(ierr);
82632d0f97SHong Zhang   ierr = PetscFree(nidx);CHKERRQ(ierr);
83632d0f97SHong Zhang #endif
84632d0f97SHong Zhang   PetscFunctionReturn(0);
85632d0f97SHong Zhang }
86632d0f97SHong Zhang 
87632d0f97SHong Zhang #undef __FUNCT__
88632d0f97SHong Zhang #define __FUNCT__ "MatExpandIndices_MPISBAIJ"
89632d0f97SHong Zhang static int MatExpandIndices_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[])
90632d0f97SHong Zhang {
91632d0f97SHong Zhang   Mat_MPISBAIJ  *baij = (Mat_MPISBAIJ*)C->data;
92632d0f97SHong Zhang   int          ierr,bs = baij->bs,n,i,j,k,*idx,*nidx;
93632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
94632d0f97SHong Zhang   int          maxsz;
95632d0f97SHong Zhang #else
96632d0f97SHong Zhang   int          Nbs = baij->Nbs;
97632d0f97SHong Zhang #endif
98632d0f97SHong Zhang 
99632d0f97SHong Zhang   PetscFunctionBegin;
100632d0f97SHong Zhang   /* printf(" ... MatExpandIndices_MPISBAIJ is called ...\n"); */
101632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
102632d0f97SHong Zhang   /* Now check max size */
103632d0f97SHong Zhang   for (i=0,maxsz=0; i<imax; i++) {
104632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
105632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
106632d0f97SHong Zhang     if (n*bs > maxsz) maxsz = n*bs;
107632d0f97SHong Zhang   }
108632d0f97SHong Zhang   ierr = PetscMalloc((maxsz+1)*sizeof(int),&nidx);CHKERRQ(ierr);
109632d0f97SHong Zhang #else
110632d0f97SHong Zhang   ierr = PetscMalloc((Nbs*bs+1)*sizeof(int),&nidx);CHKERRQ(ierr);
111632d0f97SHong Zhang #endif
112632d0f97SHong Zhang 
113632d0f97SHong Zhang   for (i=0; i<imax; i++) {
114632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
115632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
116632d0f97SHong Zhang     for (j=0; j<n ; ++j){
117632d0f97SHong Zhang       for (k=0; k<bs; k++)
118632d0f97SHong Zhang         nidx[j*bs+k] = idx[j]*bs+k;
119632d0f97SHong Zhang     }
120632d0f97SHong Zhang     ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr);
121632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,n*bs,nidx,is_out+i);CHKERRQ(ierr);
122632d0f97SHong Zhang   }
123632d0f97SHong Zhang   ierr = PetscFree(nidx);CHKERRQ(ierr);
124632d0f97SHong Zhang   PetscFunctionReturn(0);
125632d0f97SHong Zhang }
126632d0f97SHong Zhang 
127632d0f97SHong Zhang #undef __FUNCT__
128632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ"
129c910923dSHong Zhang int MatIncreaseOverlap_MPISBAIJ(Mat C,int is_max,IS is[],int ov)
130632d0f97SHong Zhang {
1311ab97a2aSSatish Balay   int           i,ierr;
132632d0f97SHong Zhang   IS            *is_new;
133632d0f97SHong Zhang 
134632d0f97SHong Zhang   PetscFunctionBegin;
135c910923dSHong Zhang   ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr);
136632d0f97SHong Zhang   /* Convert the indices into block format */
137c910923dSHong Zhang   ierr = MatCompressIndicesGeneral_MPISBAIJ(C,is_max,is,is_new);CHKERRQ(ierr);
138632d0f97SHong Zhang   if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
139632d0f97SHong Zhang   for (i=0; i<ov; ++i) {
140c910923dSHong Zhang     ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
141632d0f97SHong Zhang   }
142c910923dSHong Zhang   for (i=0; i<is_max; i++) {ierr = ISDestroy(is[i]);CHKERRQ(ierr);}
143c910923dSHong Zhang   ierr = MatExpandIndices_MPISBAIJ(C,is_max,is_new,is);CHKERRQ(ierr);
144c910923dSHong Zhang   for (i=0; i<is_max; i++) {ierr = ISDestroy(is_new[i]);CHKERRQ(ierr);}
145632d0f97SHong Zhang   ierr = PetscFree(is_new);CHKERRQ(ierr);
146632d0f97SHong Zhang   PetscFunctionReturn(0);
147632d0f97SHong Zhang }
148632d0f97SHong Zhang 
149*4a69c536SBarry Smith typedef enum {MINE,OTHER} WhoseOwner;
1500472cc68SHong Zhang /*  data1, odata1 and odata2 are packed in the format (for communication):
151a2a9f22aSHong Zhang        data[0]          = is_max, no of is
152a2a9f22aSHong Zhang        data[1]          = size of is[0]
153a2a9f22aSHong Zhang         ...
154a2a9f22aSHong Zhang        data[is_max]     = size of is[is_max-1]
155a2a9f22aSHong Zhang        data[is_max + 1] = data(is[0])
156a2a9f22aSHong Zhang         ...
157a2a9f22aSHong Zhang        data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
158a2a9f22aSHong Zhang         ...
1590472cc68SHong Zhang    data2 is packed in the format (for creating output is[]):
1600472cc68SHong Zhang        data[0]          = is_max, no of is
1610472cc68SHong Zhang        data[1]          = size of is[0]
1620472cc68SHong Zhang         ...
1630472cc68SHong Zhang        data[is_max]     = size of is[is_max-1]
1640472cc68SHong Zhang        data[is_max + 1] = data(is[0])
1650472cc68SHong Zhang         ...
1660472cc68SHong Zhang        data[is_max + 1 + Mbs*i) = data(is[i])
1670472cc68SHong Zhang         ...
168a2a9f22aSHong Zhang */
169632d0f97SHong Zhang #undef __FUNCT__
170632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Once"
171c910923dSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat C,int is_max,IS is[])
172632d0f97SHong Zhang {
173632d0f97SHong Zhang   Mat_MPISBAIJ  *c = (Mat_MPISBAIJ*)C->data;
1740472cc68SHong Zhang   int         len,*idx_i,isz,col,*n,*data1,*data2,*data2_i,*data,*data_i,
1750472cc68SHong Zhang               size,rank,Mbs,i,j,k,ierr,nrqs,*odata1,*odata2,
1760472cc68SHong Zhang               tag1,tag2,flag,proc_id,**odata2_ptr;
177bfc6803cSHong Zhang   char        *t_p;
178632d0f97SHong Zhang   MPI_Comm    comm;
1790472cc68SHong Zhang   MPI_Request *s_waits,r_req;
180632d0f97SHong Zhang   MPI_Status  *s_status,r_status;
181bfc6803cSHong Zhang   PetscBT     *table;  /* mark indices of this processor's is[] */
182fc70d14eSHong Zhang   PetscBT     table_i;
183bfc6803cSHong Zhang   PetscBT     otable; /* mark indices of other processors' is[] */
184632d0f97SHong Zhang   PetscFunctionBegin;
185632d0f97SHong Zhang 
186632d0f97SHong Zhang   comm = C->comm;
187632d0f97SHong Zhang   size = c->size;
188632d0f97SHong Zhang   rank = c->rank;
189632d0f97SHong Zhang   Mbs  = c->Mbs;
190632d0f97SHong Zhang 
191c910923dSHong Zhang   ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr);
192c910923dSHong Zhang   ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr);
193c910923dSHong Zhang 
194a2a9f22aSHong Zhang   /* create tables for marking indices */
195a2a9f22aSHong Zhang   len  = is_max*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*is_max*sizeof(char) + 1;
196bfc6803cSHong Zhang   ierr = PetscMalloc(len,&table);CHKERRQ(ierr);
197a2a9f22aSHong Zhang   t_p  = (char *)(table + is_max);
198a2a9f22aSHong Zhang   for (i=0; i<is_max; i++) {
199bfc6803cSHong Zhang     table[i]  = t_p  + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
200a2a9f22aSHong Zhang   }
2010472cc68SHong Zhang   ierr = PetscBTCreate(Mbs,otable);CHKERRQ(ierr);
202632d0f97SHong Zhang 
203bfc6803cSHong Zhang   /* 1. Send this processor's is[] to all other processors */
204bfc6803cSHong Zhang   /*-------------------------------------------------------*/
205632d0f97SHong Zhang   /* Allocate Memory for outgoing messages */
206a2a9f22aSHong Zhang   len  = is_max*sizeof(int);
207a2a9f22aSHong Zhang   ierr = PetscMalloc(len,&n);CHKERRQ(ierr);
208a2a9f22aSHong Zhang 
209c910923dSHong Zhang   len = 1 + is_max;
210c910923dSHong Zhang   for (i=0; i<is_max; i++) {
211632d0f97SHong Zhang     ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr);
212632d0f97SHong Zhang     len += n[i];
213632d0f97SHong Zhang   }
214a2a9f22aSHong Zhang   ierr = PetscMalloc(len*sizeof(int),&data1);CHKERRQ(ierr);
215632d0f97SHong Zhang 
216632d0f97SHong Zhang   /* Form the outgoing messages */
217a2a9f22aSHong Zhang   data1[0] = is_max;
218c910923dSHong Zhang   k = is_max + 1;
219a2a9f22aSHong Zhang   for (i=0; i<is_max; i++) {
220a2a9f22aSHong Zhang     data1[1+i] = n[i];
221a2a9f22aSHong Zhang     ierr = ISGetIndices(is[i],&idx_i);CHKERRQ(ierr);
222a2a9f22aSHong Zhang     for (j=0; j<data1[i+1]; j++){
223a2a9f22aSHong Zhang       data1[k++] = idx_i[j];
224632d0f97SHong Zhang     }
225a2a9f22aSHong Zhang     ierr = ISRestoreIndices(is[i],&idx_i);CHKERRQ(ierr);
226a2a9f22aSHong Zhang     ierr = ISDestroy(is[i]);CHKERRQ(ierr);
227632d0f97SHong Zhang   }
228fc70d14eSHong Zhang   if (k != len) SETERRQ2(1,"Error on forming the outgoing messages: k %d != len %d",k,len);
229bfc6803cSHong Zhang   ierr = PetscFree(n);CHKERRQ(ierr);
230632d0f97SHong Zhang 
231632d0f97SHong Zhang   /*  Now  post the sends */
2320472cc68SHong Zhang   ierr = PetscMalloc(size*sizeof(MPI_Request),&s_waits);CHKERRQ(ierr);
233632d0f97SHong Zhang   k = 0;
234a2a9f22aSHong Zhang   for (proc_id=0; proc_id<size; ++proc_id) { /* send data1 to processor [proc_id] */
235c910923dSHong Zhang     if (proc_id != rank){
2360472cc68SHong Zhang       ierr = MPI_Isend(data1,len,MPI_INT,proc_id,tag1,comm,s_waits+k);CHKERRQ(ierr);
237bfc6803cSHong Zhang       /* printf(" [%d] send %d msg to [%d], data1: \n",rank,len,proc_id); */
238632d0f97SHong Zhang       k++;
239632d0f97SHong Zhang     }
240632d0f97SHong Zhang   }
241632d0f97SHong Zhang 
242dc008846SHong Zhang   /* 2. Do local work on this processor's is[] */
243dc008846SHong Zhang   /*-------------------------------------------*/
2440472cc68SHong Zhang   ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,data1,MINE,&data,table);CHKERRQ(ierr);
245632d0f97SHong Zhang 
246632d0f97SHong Zhang   /* 3. Receive other's is[] and process. Then send back */
247bfc6803cSHong Zhang   /*-----------------------------------------------------*/
2480472cc68SHong Zhang   /* Sending this processor's is[] is done */
249632d0f97SHong Zhang   nrqs = size-1;
250632d0f97SHong Zhang   ierr = PetscMalloc(size*sizeof(MPI_Status),&s_status);CHKERRQ(ierr);
2510472cc68SHong Zhang   ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr);
252dc008846SHong Zhang 
253bfc6803cSHong Zhang   ierr = PetscMalloc(size*sizeof(int**),&odata2_ptr);CHKERRQ(ierr);
254632d0f97SHong Zhang   k = 0;
255632d0f97SHong Zhang   do {
256632d0f97SHong Zhang     /* Receive messages */
257bfc6803cSHong Zhang     ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);CHKERRQ(ierr);
258632d0f97SHong Zhang     if (flag){
259bfc6803cSHong Zhang       ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr);
260632d0f97SHong Zhang       proc_id = r_status.MPI_SOURCE;
261a2a9f22aSHong Zhang       ierr = PetscMalloc(len*sizeof(int),&odata1);CHKERRQ(ierr);
262bfc6803cSHong Zhang       ierr = MPI_Irecv(odata1,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
263fc70d14eSHong Zhang       /*  printf(" [%d] recv %d msg from [%d]\n",rank,len,proc_id); */
264632d0f97SHong Zhang 
265fc70d14eSHong Zhang       /*  Process messages */
266bfc6803cSHong Zhang       ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,&odata2_ptr[k],&otable);CHKERRQ(ierr);
267a2a9f22aSHong Zhang       odata2 = odata2_ptr[k];
268a2a9f22aSHong Zhang       len = 1 + odata2[0];
269a2a9f22aSHong Zhang       for (i=0; i<odata2[0]; i++){
270a2a9f22aSHong Zhang         len += odata2[1 + i];
271fc70d14eSHong Zhang       }
272632d0f97SHong Zhang 
273632d0f97SHong Zhang       /* Send messages back */
2740472cc68SHong Zhang       ierr = MPI_Isend(odata2,len,MPI_INT,proc_id,tag2,comm,s_waits+k);CHKERRQ(ierr);
275bfc6803cSHong Zhang       /* printf(" [%d] send %d msg back to [%d] \n",rank,len,proc_id); */
276632d0f97SHong Zhang 
277a2a9f22aSHong Zhang       ierr = PetscFree(odata1);CHKERRQ(ierr);
278fc70d14eSHong Zhang       k++;
279632d0f97SHong Zhang     }
280632d0f97SHong Zhang   } while (k < nrqs);
281632d0f97SHong Zhang 
282fc70d14eSHong Zhang   /* 4. Receive work done on other processors, then merge */
283632d0f97SHong Zhang   /*--------------------------------------------------------*/
2840472cc68SHong Zhang   /* Allocate memory for incoming data */
285bfc6803cSHong Zhang   len = (1+is_max*(Mbs+1));
2860472cc68SHong Zhang   ierr = PetscMalloc(len*sizeof(int),&data2);CHKERRQ(ierr);
287bfc6803cSHong Zhang 
288bfc6803cSHong Zhang   /* Sending others' is[] is done */
2890472cc68SHong Zhang   ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr);
290fc70d14eSHong Zhang   for (k=0; k<nrqs; k++){
291a2a9f22aSHong Zhang     ierr = PetscFree(odata2_ptr[k]);CHKERRQ(ierr);
292fc70d14eSHong Zhang   }
293a2a9f22aSHong Zhang   ierr = PetscFree(odata2_ptr);CHKERRQ(ierr);
294fc70d14eSHong Zhang 
295632d0f97SHong Zhang   k = 0;
296632d0f97SHong Zhang   do {
297632d0f97SHong Zhang     /* Receive messages */
298c910923dSHong Zhang     ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
299632d0f97SHong Zhang     if (flag){
300bfc6803cSHong Zhang       ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr);
301632d0f97SHong Zhang       proc_id = r_status.MPI_SOURCE;
3020472cc68SHong Zhang       ierr = MPI_Irecv(data2,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
3030472cc68SHong Zhang       /* printf(" [%d] recv %d msg from [%d], data2:\n",rank,len,proc_id); */
304632d0f97SHong Zhang 
3050472cc68SHong Zhang       /* Add data2 into data */
3060472cc68SHong Zhang       data2_i = data2 + 1 + is_max;
307fc70d14eSHong Zhang       for (i=0; i<is_max; i++){
308fc70d14eSHong Zhang         table_i = table[i];
309bfc6803cSHong Zhang         data_i  = data + 1 + is_max + Mbs*i;
310bfc6803cSHong Zhang         isz     = data[1+i];
3110472cc68SHong Zhang         for (j=0; j<data2[1+i]; j++){
3120472cc68SHong Zhang           col = data2_i[j];
313bfc6803cSHong Zhang           if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;}
314fc70d14eSHong Zhang         }
315bfc6803cSHong Zhang         data[1+i] = isz;
3160472cc68SHong Zhang         if (i < is_max - 1) data2_i += data2[1+i];
317fc70d14eSHong Zhang       }
318632d0f97SHong Zhang       k++;
319632d0f97SHong Zhang     }
320632d0f97SHong Zhang   } while (k < nrqs);
321632d0f97SHong Zhang 
322c910923dSHong Zhang   /* 5. Create new is[] */
323c910923dSHong Zhang   /*--------------------*/
324c910923dSHong Zhang   for (i=0; i<is_max; i++) {
325bfc6803cSHong Zhang     data_i = data + 1 + is_max + Mbs*i;
326bfc6803cSHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,is+i);CHKERRQ(ierr);
327632d0f97SHong Zhang   }
328bfc6803cSHong Zhang   ierr = PetscFree(data1);CHKERRQ(ierr);
329a2a9f22aSHong Zhang   ierr = PetscFree(data2);CHKERRQ(ierr);
3300472cc68SHong Zhang   ierr = PetscFree(data);CHKERRQ(ierr);
3310472cc68SHong Zhang   ierr = PetscFree(s_waits);CHKERRQ(ierr);
332632d0f97SHong Zhang   ierr = PetscFree(s_status);CHKERRQ(ierr);
333fc70d14eSHong Zhang   ierr = PetscFree(table);CHKERRQ(ierr);
334bfc6803cSHong Zhang   ierr = PetscBTDestroy(otable);CHKERRQ(ierr);
335632d0f97SHong Zhang   PetscFunctionReturn(0);
336632d0f97SHong Zhang }
337632d0f97SHong Zhang 
338632d0f97SHong Zhang #undef __FUNCT__
339632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Local"
340632d0f97SHong Zhang /*
341dc008846SHong Zhang    MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do
342632d0f97SHong Zhang        the work on the local processor.
343632d0f97SHong Zhang 
344632d0f97SHong Zhang      Inputs:
345632d0f97SHong Zhang       C      - MAT_MPISBAIJ;
346bfc6803cSHong Zhang       data   - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format.
347bfc6803cSHong Zhang       whose  - whose is[] to be processed,
348bfc6803cSHong Zhang                MINE:  this processor's is[]
349bfc6803cSHong Zhang                OTHER: other processor's is[]
350632d0f97SHong Zhang      Output:
3510472cc68SHong Zhang        data_new  - whose = MINE:
3520472cc68SHong Zhang                      holds input and newly found indices in the same format as data
3530472cc68SHong Zhang                    whose = OTHER:
3540472cc68SHong Zhang                      only holds the newly found indices
3550472cc68SHong Zhang        table     - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'.
356632d0f97SHong Zhang */
357bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat C,int *data,int whose,int **data_new,PetscBT *table)
358632d0f97SHong Zhang {
359632d0f97SHong Zhang   Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
360dc008846SHong Zhang   Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data;
361dc008846SHong Zhang   Mat_SeqBAIJ  *b = (Mat_SeqBAIJ*)(c->B)->data;
36231f99336SHong Zhang   int          ierr,row,mbs,Mbs,*nidx,*nidx_i,col,isz,isz0,*ai,*aj,*bi,*bj,*garray,rstart,l;
363dc008846SHong Zhang   int          a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n;
364bfc6803cSHong Zhang   PetscBT      table0;  /* mark the indices of input is[] for look up */
365bfc6803cSHong Zhang   PetscBT      table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */
366632d0f97SHong Zhang 
367632d0f97SHong Zhang   PetscFunctionBegin;
36831f99336SHong Zhang   Mbs = c->Mbs; mbs = a->mbs;
369dc008846SHong Zhang   ai = a->i; aj = a->j;
370dc008846SHong Zhang   bi = b->i; bj = b->j;
371632d0f97SHong Zhang   garray = c->garray;
372dc008846SHong Zhang   rstart = c->rstart;
373dc008846SHong Zhang   is_max = data[0];
374632d0f97SHong Zhang 
375fc70d14eSHong Zhang   ierr = PetscBTCreate(Mbs,table0);CHKERRQ(ierr);
376fc70d14eSHong Zhang 
377fc70d14eSHong Zhang   ierr = PetscMalloc((1+is_max*(Mbs+1))*sizeof(int),&nidx);CHKERRQ(ierr);
378fc70d14eSHong Zhang   nidx[0] = is_max;
379fc70d14eSHong Zhang 
380fc70d14eSHong Zhang   idx_i  = data + is_max + 1; /* ptr to input is[0] array */
381bfc6803cSHong Zhang   nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */
382dc008846SHong Zhang   for (i=0; i<is_max; i++) { /* for each is */
383dc008846SHong Zhang     isz  = 0;
384fc70d14eSHong Zhang     n = data[1+i]; /* size of input is[i] */
385dc008846SHong Zhang 
386bfc6803cSHong Zhang     /* initialize table_i, set table0 */
387bfc6803cSHong Zhang     if (whose == MINE){ /* process this processor's is[] */
388bfc6803cSHong Zhang       table_i = table[i];
3890472cc68SHong Zhang       nidx_i  = nidx + 1+ is_max + Mbs*i;
390bfc6803cSHong Zhang     } else {            /* process other processor's is[] - only use one temp table */
391bfc6803cSHong Zhang       table_i = *table;
392bfc6803cSHong Zhang     }
393bfc6803cSHong Zhang     ierr = PetscBTMemzero(Mbs,table_i);CHKERRQ(ierr);
394bfc6803cSHong Zhang     ierr = PetscBTMemzero(Mbs,table0);CHKERRQ(ierr);
395fc70d14eSHong Zhang     if (n > 0) {
396bfc6803cSHong Zhang       isz0 = 0;
397dc008846SHong Zhang       for (j=0; j<n; j++){
398dc008846SHong Zhang         col = idx_i[j];
399dc008846SHong Zhang         if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %d >= Mbs %d",col,Mbs);
400bfc6803cSHong Zhang         if(!PetscBTLookupSet(table_i,col)) {
401bfc6803cSHong Zhang           ierr = PetscBTSet(table0,col);CHKERRQ(ierr);
4020472cc68SHong Zhang           if (whose == MINE) {nidx_i[isz0] = col;}
403bfc6803cSHong Zhang           isz0++;
404bfc6803cSHong Zhang         }
405632d0f97SHong Zhang       }
406dc008846SHong Zhang 
4070472cc68SHong Zhang       if (whose == MINE) {isz = isz0;}
408fc70d14eSHong Zhang       k = 0;  /* no. of indices from input is[i] that have been examined */
409dc008846SHong Zhang       for (row=0; row<mbs; row++){
410dc008846SHong Zhang         a_start = ai[row]; a_end = ai[row+1];
411dc008846SHong Zhang         b_start = bi[row]; b_end = bi[row+1];
4120472cc68SHong Zhang         if (PetscBTLookup(table0,row+rstart)){ /* row is on input is[i]:
4130472cc68SHong Zhang            do row search: collect all col in this row */
414dc008846SHong Zhang           for (l = a_start; l<a_end ; l++){ /* Amat */
415dc008846SHong Zhang             col = aj[l] + rstart;
416fc70d14eSHong Zhang             if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
417dc008846SHong Zhang           }
418dc008846SHong Zhang           for (l = b_start; l<b_end ; l++){ /* Bmat */
419dc008846SHong Zhang             col = garray[bj[l]];
420fc70d14eSHong Zhang             if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
421dc008846SHong Zhang           }
422dc008846SHong Zhang           k++;
423dc008846SHong Zhang           if (k >= isz0) break; /* for (row=0; row<mbs; row++) */
4240472cc68SHong Zhang         } else { /* row is not on input is[i]:
4250472cc68SHong Zhang           do col serach: add row onto nidx_i if there is a col in nidx_i */
426dc008846SHong Zhang           for (l = a_start; l<a_end ; l++){ /* Amat */
427dc008846SHong Zhang             col = aj[l] + rstart;
428dc008846SHong Zhang             if (PetscBTLookup(table0,col)){
429fc70d14eSHong Zhang               if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
430dc008846SHong Zhang               break; /* for l = start; l<end ; l++) */
431632d0f97SHong Zhang             }
432632d0f97SHong Zhang           }
433dc008846SHong Zhang           for (l = b_start; l<b_end ; l++){ /* Bmat */
434dc008846SHong Zhang             col = garray[bj[l]];
435dc008846SHong Zhang             if (PetscBTLookup(table0,col)){
436fc70d14eSHong Zhang               if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
437dc008846SHong Zhang               break; /* for l = start; l<end ; l++) */
438632d0f97SHong Zhang             }
439dc008846SHong Zhang           }
440dc008846SHong Zhang         }
441bfc6803cSHong Zhang       }
442fc70d14eSHong Zhang     } /* if (n > 0) */
443dc008846SHong Zhang 
444dc008846SHong Zhang     if (i < is_max - 1){
445fc70d14eSHong Zhang       idx_i  += n;   /* ptr to input is[i+1] array */
446bfc6803cSHong Zhang       nidx_i += isz; /* ptr to output is[i+1] array */
447dc008846SHong Zhang     }
448fc70d14eSHong Zhang     nidx[1+i] = isz; /* size of new is[i] */
4491ab97a2aSSatish Balay   } /* for each is */
450dc008846SHong Zhang   *data_new = nidx;
451dc008846SHong Zhang   ierr = PetscBTDestroy(table0);CHKERRQ(ierr);
452632d0f97SHong Zhang 
453632d0f97SHong Zhang   PetscFunctionReturn(0);
454632d0f97SHong Zhang }
455632d0f97SHong Zhang 
456632d0f97SHong Zhang 
457