xref: /petsc/src/mat/impls/sbaij/mpi/sbaijov.c (revision 0472cc684b82b22d264121f88077b92302f10791)
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 */
7bfc6803cSHong Zhang #include "src/mat/impls/sbaij/seq/sbaij.h"
8632d0f97SHong Zhang #include "src/mat/impls/sbaij/mpi/mpisbaij.h"
9632d0f97SHong Zhang #include "petscbt.h"
10632d0f97SHong Zhang 
11632d0f97SHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat,int,IS *);
12bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat,int *,int,int **,PetscBT*);
13632d0f97SHong Zhang 
14632d0f97SHong Zhang /* this function is sasme as MatCompressIndicesGeneral_MPIBAIJ -- should be removed! */
15632d0f97SHong Zhang #undef __FUNCT__
16632d0f97SHong Zhang #define __FUNCT__ "MatCompressIndicesGeneral_MPISBAIJ"
17632d0f97SHong Zhang static int MatCompressIndicesGeneral_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[])
18632d0f97SHong Zhang {
19632d0f97SHong Zhang   Mat_MPISBAIJ        *baij = (Mat_MPISBAIJ*)C->data;
20632d0f97SHong Zhang   int                ierr,isz,bs = baij->bs,n,i,j,*idx,ival;
21632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
22632d0f97SHong Zhang   PetscTable         gid1_lid1;
23632d0f97SHong Zhang   int                tt, gid1, *nidx;
24632d0f97SHong Zhang   PetscTablePosition tpos;
25632d0f97SHong Zhang #else
26632d0f97SHong Zhang   int                Nbs,*nidx;
27632d0f97SHong Zhang   PetscBT            table;
28632d0f97SHong Zhang #endif
29632d0f97SHong Zhang 
30632d0f97SHong Zhang   PetscFunctionBegin;
31632d0f97SHong Zhang   /* printf(" ...MatCompressIndicesGeneral_MPISBAIJ is called ...\n"); */
32632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
33632d0f97SHong Zhang   ierr = PetscTableCreate(baij->mbs,&gid1_lid1);CHKERRQ(ierr);
34632d0f97SHong Zhang #else
35632d0f97SHong Zhang   Nbs  = baij->Nbs;
36632d0f97SHong Zhang   ierr = PetscMalloc((Nbs+1)*sizeof(int),&nidx);CHKERRQ(ierr);
37632d0f97SHong Zhang   ierr = PetscBTCreate(Nbs,table);CHKERRQ(ierr);
38632d0f97SHong Zhang #endif
39632d0f97SHong Zhang   for (i=0; i<imax; i++) {
40632d0f97SHong Zhang     isz  = 0;
41632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
42632d0f97SHong Zhang     ierr = PetscTableRemoveAll(gid1_lid1);CHKERRQ(ierr);
43632d0f97SHong Zhang #else
44632d0f97SHong Zhang     ierr = PetscBTMemzero(Nbs,table);CHKERRQ(ierr);
45632d0f97SHong Zhang #endif
46632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
47632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
48632d0f97SHong Zhang     for (j=0; j<n ; j++) {
49632d0f97SHong Zhang       ival = idx[j]/bs; /* convert the indices into block indices */
50632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
51632d0f97SHong Zhang       ierr = PetscTableFind(gid1_lid1,ival+1,&tt);CHKERRQ(ierr);
52632d0f97SHong Zhang       if (!tt) {
53632d0f97SHong Zhang 	ierr = PetscTableAdd(gid1_lid1,ival+1,isz+1);CHKERRQ(ierr);
54632d0f97SHong Zhang         isz++;
55632d0f97SHong Zhang       }
56632d0f97SHong Zhang #else
57632d0f97SHong Zhang       if (ival>Nbs) SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than mat-dim");
58632d0f97SHong Zhang       if(!PetscBTLookupSet(table,ival)) { nidx[isz++] = ival;}
59632d0f97SHong Zhang #endif
60632d0f97SHong Zhang     }
61632d0f97SHong Zhang     ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr);
62632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
63632d0f97SHong Zhang     ierr = PetscMalloc((isz+1)*sizeof(int),&nidx);CHKERRQ(ierr);
64632d0f97SHong Zhang     ierr = PetscTableGetHeadPosition(gid1_lid1,&tpos);CHKERRQ(ierr);
65632d0f97SHong Zhang     j = 0;
66632d0f97SHong Zhang     while (tpos) {
67632d0f97SHong Zhang       ierr = PetscTableGetNext(gid1_lid1,&tpos,&gid1,&tt);CHKERRQ(ierr);
68632d0f97SHong Zhang       if (tt-- > isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"index greater than array-dim"); }
69632d0f97SHong Zhang       nidx[tt] = gid1 - 1;
70632d0f97SHong Zhang       j++;
71632d0f97SHong Zhang     }
72632d0f97SHong Zhang     if (j != isz) { SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"table error: jj != isz"); }
73632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr);
74632d0f97SHong Zhang     ierr = PetscFree(nidx);CHKERRQ(ierr);
75632d0f97SHong Zhang #else
76632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,isz,nidx,(is_out+i));CHKERRQ(ierr);
77632d0f97SHong Zhang #endif
78632d0f97SHong Zhang   }
79632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
80632d0f97SHong Zhang   ierr = PetscTableDelete(gid1_lid1);CHKERRQ(ierr);
81632d0f97SHong Zhang #else
82632d0f97SHong Zhang   ierr = PetscBTDestroy(table);CHKERRQ(ierr);
83632d0f97SHong Zhang   ierr = PetscFree(nidx);CHKERRQ(ierr);
84632d0f97SHong Zhang #endif
85632d0f97SHong Zhang   PetscFunctionReturn(0);
86632d0f97SHong Zhang }
87632d0f97SHong Zhang 
88632d0f97SHong Zhang #undef __FUNCT__
89632d0f97SHong Zhang #define __FUNCT__ "MatExpandIndices_MPISBAIJ"
90632d0f97SHong Zhang static int MatExpandIndices_MPISBAIJ(Mat C,int imax,const IS is_in[],IS is_out[])
91632d0f97SHong Zhang {
92632d0f97SHong Zhang   Mat_MPISBAIJ  *baij = (Mat_MPISBAIJ*)C->data;
93632d0f97SHong Zhang   int          ierr,bs = baij->bs,n,i,j,k,*idx,*nidx;
94632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
95632d0f97SHong Zhang   int          maxsz;
96632d0f97SHong Zhang #else
97632d0f97SHong Zhang   int          Nbs = baij->Nbs;
98632d0f97SHong Zhang #endif
99632d0f97SHong Zhang 
100632d0f97SHong Zhang   PetscFunctionBegin;
101632d0f97SHong Zhang   /* printf(" ... MatExpandIndices_MPISBAIJ is called ...\n"); */
102632d0f97SHong Zhang #if defined (PETSC_USE_CTABLE)
103632d0f97SHong Zhang   /* Now check max size */
104632d0f97SHong Zhang   for (i=0,maxsz=0; i<imax; i++) {
105632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
106632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
107632d0f97SHong Zhang     if (n*bs > maxsz) maxsz = n*bs;
108632d0f97SHong Zhang   }
109632d0f97SHong Zhang   ierr = PetscMalloc((maxsz+1)*sizeof(int),&nidx);CHKERRQ(ierr);
110632d0f97SHong Zhang #else
111632d0f97SHong Zhang   ierr = PetscMalloc((Nbs*bs+1)*sizeof(int),&nidx);CHKERRQ(ierr);
112632d0f97SHong Zhang #endif
113632d0f97SHong Zhang 
114632d0f97SHong Zhang   for (i=0; i<imax; i++) {
115632d0f97SHong Zhang     ierr = ISGetIndices(is_in[i],&idx);CHKERRQ(ierr);
116632d0f97SHong Zhang     ierr = ISGetLocalSize(is_in[i],&n);CHKERRQ(ierr);
117632d0f97SHong Zhang     for (j=0; j<n ; ++j){
118632d0f97SHong Zhang       for (k=0; k<bs; k++)
119632d0f97SHong Zhang         nidx[j*bs+k] = idx[j]*bs+k;
120632d0f97SHong Zhang     }
121632d0f97SHong Zhang     ierr = ISRestoreIndices(is_in[i],&idx);CHKERRQ(ierr);
122632d0f97SHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,n*bs,nidx,is_out+i);CHKERRQ(ierr);
123632d0f97SHong Zhang   }
124632d0f97SHong Zhang   ierr = PetscFree(nidx);CHKERRQ(ierr);
125632d0f97SHong Zhang   PetscFunctionReturn(0);
126632d0f97SHong Zhang }
127632d0f97SHong Zhang 
128632d0f97SHong Zhang #undef __FUNCT__
129632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ"
130c910923dSHong Zhang int MatIncreaseOverlap_MPISBAIJ(Mat C,int is_max,IS is[],int ov)
131632d0f97SHong Zhang {
132bfc6803cSHong Zhang   Mat_MPISBAIJ  *c = (Mat_MPISBAIJ*)C->data;
133bfc6803cSHong Zhang   int           i,ierr,size;
134632d0f97SHong Zhang   IS            *is_new;
135632d0f97SHong Zhang 
136632d0f97SHong Zhang   PetscFunctionBegin;
137c910923dSHong Zhang   ierr = PetscMalloc(is_max*sizeof(IS),&is_new);CHKERRQ(ierr);
138632d0f97SHong Zhang   /* Convert the indices into block format */
139c910923dSHong Zhang   ierr = MatCompressIndicesGeneral_MPISBAIJ(C,is_max,is,is_new);CHKERRQ(ierr);
140632d0f97SHong Zhang   if (ov < 0){ SETERRQ(PETSC_ERR_ARG_OUTOFRANGE,"Negative overlap specified\n");}
141632d0f97SHong Zhang   for (i=0; i<ov; ++i) {
142c910923dSHong Zhang     ierr = MatIncreaseOverlap_MPISBAIJ_Once(C,is_max,is_new);CHKERRQ(ierr);
143632d0f97SHong Zhang   }
144c910923dSHong Zhang   for (i=0; i<is_max; i++) {ierr = ISDestroy(is[i]);CHKERRQ(ierr);}
145c910923dSHong Zhang   ierr = MatExpandIndices_MPISBAIJ(C,is_max,is_new,is);CHKERRQ(ierr);
146c910923dSHong Zhang   for (i=0; i<is_max; i++) {ierr = ISDestroy(is_new[i]);CHKERRQ(ierr);}
147632d0f97SHong Zhang   ierr = PetscFree(is_new);CHKERRQ(ierr);
148632d0f97SHong Zhang   PetscFunctionReturn(0);
149632d0f97SHong Zhang }
150632d0f97SHong Zhang 
151bfc6803cSHong Zhang typedef enum {MINE,OTHER} whose;
152*0472cc68SHong Zhang /*  data1, odata1 and odata2 are packed in the format (for communication):
153a2a9f22aSHong Zhang        data[0]          = is_max, no of is
154a2a9f22aSHong Zhang        data[1]          = size of is[0]
155a2a9f22aSHong Zhang         ...
156a2a9f22aSHong Zhang        data[is_max]     = size of is[is_max-1]
157a2a9f22aSHong Zhang        data[is_max + 1] = data(is[0])
158a2a9f22aSHong Zhang         ...
159a2a9f22aSHong Zhang        data[is_max+1+sum(size of is[k]), k=0,...,i-1] = data(is[i])
160a2a9f22aSHong Zhang         ...
161*0472cc68SHong Zhang    data2 is packed in the format (for creating output is[]):
162*0472cc68SHong Zhang        data[0]          = is_max, no of is
163*0472cc68SHong Zhang        data[1]          = size of is[0]
164*0472cc68SHong Zhang         ...
165*0472cc68SHong Zhang        data[is_max]     = size of is[is_max-1]
166*0472cc68SHong Zhang        data[is_max + 1] = data(is[0])
167*0472cc68SHong Zhang         ...
168*0472cc68SHong Zhang        data[is_max + 1 + Mbs*i) = data(is[i])
169*0472cc68SHong Zhang         ...
170a2a9f22aSHong Zhang */
171632d0f97SHong Zhang #undef __FUNCT__
172632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Once"
173c910923dSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Once(Mat C,int is_max,IS is[])
174632d0f97SHong Zhang {
175632d0f97SHong Zhang   Mat_MPISBAIJ  *c = (Mat_MPISBAIJ*)C->data;
176*0472cc68SHong Zhang   int         len,*idx_i,isz,col,*n,*data1,*data2,*data2_i,*data,*data_i,
177*0472cc68SHong Zhang               size,rank,Mbs,i,j,k,ierr,nrqs,*odata1,*odata2,
178*0472cc68SHong Zhang               tag1,tag2,flag,proc_id,**odata2_ptr;
179bfc6803cSHong Zhang   char        *t_p;
180632d0f97SHong Zhang   MPI_Comm    comm;
181*0472cc68SHong Zhang   MPI_Request *s_waits,r_req;
182632d0f97SHong Zhang   MPI_Status  *s_status,r_status;
183bfc6803cSHong Zhang   PetscBT     *table;  /* mark indices of this processor's is[] */
184fc70d14eSHong Zhang   PetscBT     table_i;
185bfc6803cSHong Zhang   PetscBT     otable; /* mark indices of other processors' is[] */
186632d0f97SHong Zhang   PetscFunctionBegin;
187632d0f97SHong Zhang 
188632d0f97SHong Zhang   comm = C->comm;
189632d0f97SHong Zhang   size = c->size;
190632d0f97SHong Zhang   rank = c->rank;
191632d0f97SHong Zhang   Mbs  = c->Mbs;
192632d0f97SHong Zhang 
193c910923dSHong Zhang   ierr = PetscObjectGetNewTag((PetscObject)C,&tag1);CHKERRQ(ierr);
194c910923dSHong Zhang   ierr = PetscObjectGetNewTag((PetscObject)C,&tag2);CHKERRQ(ierr);
195c910923dSHong Zhang 
196a2a9f22aSHong Zhang   /* create tables for marking indices */
197a2a9f22aSHong Zhang   len  = is_max*sizeof(PetscBT) + (Mbs/PETSC_BITS_PER_BYTE+1)*is_max*sizeof(char) + 1;
198bfc6803cSHong Zhang   ierr = PetscMalloc(len,&table);CHKERRQ(ierr);
199a2a9f22aSHong Zhang   t_p  = (char *)(table + is_max);
200a2a9f22aSHong Zhang   for (i=0; i<is_max; i++) {
201bfc6803cSHong Zhang     table[i]  = t_p  + (Mbs/PETSC_BITS_PER_BYTE+1)*i;
202a2a9f22aSHong Zhang   }
203*0472cc68SHong Zhang   ierr = PetscBTCreate(Mbs,otable);CHKERRQ(ierr);
204632d0f97SHong Zhang 
205bfc6803cSHong Zhang   /* 1. Send this processor's is[] to all other processors */
206bfc6803cSHong Zhang   /*-------------------------------------------------------*/
207632d0f97SHong Zhang   /* Allocate Memory for outgoing messages */
208a2a9f22aSHong Zhang   len  = is_max*sizeof(int);
209a2a9f22aSHong Zhang   ierr = PetscMalloc(len,&n);CHKERRQ(ierr);
210a2a9f22aSHong Zhang 
211c910923dSHong Zhang   len = 1 + is_max;
212c910923dSHong Zhang   for (i=0; i<is_max; i++) {
213632d0f97SHong Zhang     ierr = ISGetLocalSize(is[i],&n[i]);CHKERRQ(ierr);
214632d0f97SHong Zhang     len += n[i];
215632d0f97SHong Zhang   }
216a2a9f22aSHong Zhang   ierr = PetscMalloc(len*sizeof(int),&data1);CHKERRQ(ierr);
217632d0f97SHong Zhang 
218632d0f97SHong Zhang   /* Form the outgoing messages */
219a2a9f22aSHong Zhang   data1[0] = is_max;
220c910923dSHong Zhang   k = is_max + 1;
221a2a9f22aSHong Zhang   for (i=0; i<is_max; i++) {
222a2a9f22aSHong Zhang     data1[1+i] = n[i];
223a2a9f22aSHong Zhang     ierr = ISGetIndices(is[i],&idx_i);CHKERRQ(ierr);
224a2a9f22aSHong Zhang     for (j=0; j<data1[i+1]; j++){
225a2a9f22aSHong Zhang       data1[k++] = idx_i[j];
226632d0f97SHong Zhang     }
227a2a9f22aSHong Zhang     ierr = ISRestoreIndices(is[i],&idx_i);CHKERRQ(ierr);
228a2a9f22aSHong Zhang     ierr = ISDestroy(is[i]);CHKERRQ(ierr);
229632d0f97SHong Zhang   }
230fc70d14eSHong Zhang   if (k != len) SETERRQ2(1,"Error on forming the outgoing messages: k %d != len %d",k,len);
231bfc6803cSHong Zhang   ierr = PetscFree(n);CHKERRQ(ierr);
232632d0f97SHong Zhang 
233632d0f97SHong Zhang   /*  Now  post the sends */
234*0472cc68SHong Zhang   ierr = PetscMalloc(size*sizeof(MPI_Request),&s_waits);CHKERRQ(ierr);
235632d0f97SHong Zhang   k = 0;
236a2a9f22aSHong Zhang   for (proc_id=0; proc_id<size; ++proc_id) { /* send data1 to processor [proc_id] */
237c910923dSHong Zhang     if (proc_id != rank){
238*0472cc68SHong Zhang       ierr = MPI_Isend(data1,len,MPI_INT,proc_id,tag1,comm,s_waits+k);CHKERRQ(ierr);
239bfc6803cSHong Zhang       /* printf(" [%d] send %d msg to [%d], data1: \n",rank,len,proc_id); */
240632d0f97SHong Zhang       k++;
241632d0f97SHong Zhang     }
242632d0f97SHong Zhang   }
243632d0f97SHong Zhang 
244dc008846SHong Zhang   /* 2. Do local work on this processor's is[] */
245dc008846SHong Zhang   /*-------------------------------------------*/
246*0472cc68SHong Zhang   ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,data1,MINE,&data,table);CHKERRQ(ierr);
247632d0f97SHong Zhang 
248632d0f97SHong Zhang   /* 3. Receive other's is[] and process. Then send back */
249bfc6803cSHong Zhang   /*-----------------------------------------------------*/
250*0472cc68SHong Zhang   /* Sending this processor's is[] is done */
251632d0f97SHong Zhang   nrqs = size-1;
252632d0f97SHong Zhang   ierr = PetscMalloc(size*sizeof(MPI_Status),&s_status);CHKERRQ(ierr);
253*0472cc68SHong Zhang   ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr);
254dc008846SHong Zhang 
255bfc6803cSHong Zhang   ierr = PetscMalloc(size*sizeof(int**),&odata2_ptr);CHKERRQ(ierr);
256632d0f97SHong Zhang   k = 0;
257632d0f97SHong Zhang   do {
258632d0f97SHong Zhang     /* Receive messages */
259bfc6803cSHong Zhang     ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag1,comm,&flag,&r_status);CHKERRQ(ierr);
260632d0f97SHong Zhang     if (flag){
261bfc6803cSHong Zhang       ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr);
262632d0f97SHong Zhang       proc_id = r_status.MPI_SOURCE;
263a2a9f22aSHong Zhang       ierr = PetscMalloc(len*sizeof(int),&odata1);CHKERRQ(ierr);
264bfc6803cSHong Zhang       ierr = MPI_Irecv(odata1,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
265fc70d14eSHong Zhang       /*  printf(" [%d] recv %d msg from [%d]\n",rank,len,proc_id); */
266632d0f97SHong Zhang 
267fc70d14eSHong Zhang       /*  Process messages */
268bfc6803cSHong Zhang       ierr = MatIncreaseOverlap_MPISBAIJ_Local(C,odata1,OTHER,&odata2_ptr[k],&otable);CHKERRQ(ierr);
269a2a9f22aSHong Zhang       odata2 = odata2_ptr[k];
270a2a9f22aSHong Zhang       len = 1 + odata2[0];
271a2a9f22aSHong Zhang       for (i=0; i<odata2[0]; i++){
272a2a9f22aSHong Zhang         len += odata2[1 + i];
273fc70d14eSHong Zhang       }
274632d0f97SHong Zhang 
275632d0f97SHong Zhang       /* Send messages back */
276*0472cc68SHong Zhang       ierr = MPI_Isend(odata2,len,MPI_INT,proc_id,tag2,comm,s_waits+k);CHKERRQ(ierr);
277bfc6803cSHong Zhang       /* printf(" [%d] send %d msg back to [%d] \n",rank,len,proc_id); */
278632d0f97SHong Zhang 
279a2a9f22aSHong Zhang       ierr = PetscFree(odata1);CHKERRQ(ierr);
280fc70d14eSHong Zhang       k++;
281632d0f97SHong Zhang     }
282632d0f97SHong Zhang   } while (k < nrqs);
283632d0f97SHong Zhang 
284fc70d14eSHong Zhang   /* 4. Receive work done on other processors, then merge */
285632d0f97SHong Zhang   /*--------------------------------------------------------*/
286*0472cc68SHong Zhang   /* Allocate memory for incoming data */
287bfc6803cSHong Zhang   len = (1+is_max*(Mbs+1));
288*0472cc68SHong Zhang   ierr = PetscMalloc(len*sizeof(int),&data2);CHKERRQ(ierr);
289bfc6803cSHong Zhang 
290bfc6803cSHong Zhang   /* Sending others' is[] is done */
291*0472cc68SHong Zhang   ierr = MPI_Waitall(nrqs,s_waits,s_status);CHKERRQ(ierr);
292fc70d14eSHong Zhang   for (k=0; k<nrqs; k++){
293a2a9f22aSHong Zhang     ierr = PetscFree(odata2_ptr[k]);CHKERRQ(ierr);
294fc70d14eSHong Zhang   }
295a2a9f22aSHong Zhang   ierr = PetscFree(odata2_ptr);CHKERRQ(ierr);
296fc70d14eSHong Zhang 
297632d0f97SHong Zhang   k = 0;
298632d0f97SHong Zhang   do {
299632d0f97SHong Zhang     /* Receive messages */
300c910923dSHong Zhang     ierr = MPI_Iprobe(MPI_ANY_SOURCE,tag2,comm,&flag,&r_status);
301632d0f97SHong Zhang     if (flag){
302bfc6803cSHong Zhang       ierr = MPI_Get_count(&r_status,MPI_INT,&len);CHKERRQ(ierr);
303632d0f97SHong Zhang       proc_id = r_status.MPI_SOURCE;
304*0472cc68SHong Zhang       ierr = MPI_Irecv(data2,len,MPI_INT,proc_id,r_status.MPI_TAG,comm,&r_req);CHKERRQ(ierr);
305*0472cc68SHong Zhang       /* printf(" [%d] recv %d msg from [%d], data2:\n",rank,len,proc_id); */
306632d0f97SHong Zhang 
307*0472cc68SHong Zhang       /* Add data2 into data */
308*0472cc68SHong Zhang       data2_i = data2 + 1 + is_max;
309fc70d14eSHong Zhang       for (i=0; i<is_max; i++){
310fc70d14eSHong Zhang         table_i = table[i];
311bfc6803cSHong Zhang         data_i  = data + 1 + is_max + Mbs*i;
312bfc6803cSHong Zhang         isz     = data[1+i];
313*0472cc68SHong Zhang         for (j=0; j<data2[1+i]; j++){
314*0472cc68SHong Zhang           col = data2_i[j];
315bfc6803cSHong Zhang           if (!PetscBTLookupSet(table_i,col)) {data_i[isz++] = col;}
316fc70d14eSHong Zhang         }
317bfc6803cSHong Zhang         data[1+i] = isz;
318*0472cc68SHong Zhang         if (i < is_max - 1) data2_i += data2[1+i];
319fc70d14eSHong Zhang       }
320632d0f97SHong Zhang       k++;
321632d0f97SHong Zhang     }
322632d0f97SHong Zhang   } while (k < nrqs);
323632d0f97SHong Zhang 
324c910923dSHong Zhang   /* 5. Create new is[] */
325c910923dSHong Zhang   /*--------------------*/
326c910923dSHong Zhang   for (i=0; i<is_max; i++) {
327bfc6803cSHong Zhang     data_i = data + 1 + is_max + Mbs*i;
328bfc6803cSHong Zhang     ierr = ISCreateGeneral(PETSC_COMM_SELF,data[1+i],data_i,is+i);CHKERRQ(ierr);
329632d0f97SHong Zhang   }
330bfc6803cSHong Zhang   ierr = PetscFree(data1);CHKERRQ(ierr);
331a2a9f22aSHong Zhang   ierr = PetscFree(data2);CHKERRQ(ierr);
332*0472cc68SHong Zhang   ierr = PetscFree(data);CHKERRQ(ierr);
333*0472cc68SHong Zhang   ierr = PetscFree(s_waits);CHKERRQ(ierr);
334632d0f97SHong Zhang   ierr = PetscFree(s_status);CHKERRQ(ierr);
335fc70d14eSHong Zhang   ierr = PetscFree(table);CHKERRQ(ierr);
336bfc6803cSHong Zhang   ierr = PetscBTDestroy(otable);CHKERRQ(ierr);
337632d0f97SHong Zhang   PetscFunctionReturn(0);
338632d0f97SHong Zhang }
339632d0f97SHong Zhang 
340632d0f97SHong Zhang #undef __FUNCT__
341632d0f97SHong Zhang #define __FUNCT__ "MatIncreaseOverlap_MPISBAIJ_Local"
342632d0f97SHong Zhang /*
343dc008846SHong Zhang    MatIncreaseOverlap_MPISBAIJ_Local - Called by MatIncreaseOverlap, to do
344632d0f97SHong Zhang        the work on the local processor.
345632d0f97SHong Zhang 
346632d0f97SHong Zhang      Inputs:
347632d0f97SHong Zhang       C      - MAT_MPISBAIJ;
348bfc6803cSHong Zhang       data   - holds is[]. See MatIncreaseOverlap_MPISBAIJ_Once() for the format.
349bfc6803cSHong Zhang       whose  - whose is[] to be processed,
350bfc6803cSHong Zhang                MINE:  this processor's is[]
351bfc6803cSHong Zhang                OTHER: other processor's is[]
352632d0f97SHong Zhang      Output:
353*0472cc68SHong Zhang        data_new  - whose = MINE:
354*0472cc68SHong Zhang                      holds input and newly found indices in the same format as data
355*0472cc68SHong Zhang                    whose = OTHER:
356*0472cc68SHong Zhang                      only holds the newly found indices
357*0472cc68SHong Zhang        table     - table[i]: mark the indices of is[i], i=0,...,is_max. Used only in the case 'whose=MINE'.
358632d0f97SHong Zhang */
359bfc6803cSHong Zhang static int MatIncreaseOverlap_MPISBAIJ_Local(Mat C,int *data,int whose,int **data_new,PetscBT *table)
360632d0f97SHong Zhang {
361632d0f97SHong Zhang   Mat_MPISBAIJ *c = (Mat_MPISBAIJ*)C->data;
362dc008846SHong Zhang   Mat_SeqSBAIJ *a = (Mat_SeqSBAIJ*)(c->A)->data;
363dc008846SHong Zhang   Mat_SeqBAIJ  *b = (Mat_SeqBAIJ*)(c->B)->data;
364dc008846SHong Zhang   int          ierr,row,mbs,Mbs,*nidx,*nidx_i,col,isz,isz0,*ai,*aj,bs,*bi,*bj,*garray,rstart,l;
365dc008846SHong Zhang   int          a_start,a_end,b_start,b_end,i,j,k,is_max,*idx_i,n;
366bfc6803cSHong Zhang   PetscBT      table0;  /* mark the indices of input is[] for look up */
367bfc6803cSHong Zhang   PetscBT      table_i; /* poits to i-th table. When whose=OTHER, a single table is used for all is[] */
368632d0f97SHong Zhang 
369632d0f97SHong Zhang   PetscFunctionBegin;
370dc008846SHong Zhang   Mbs = c->Mbs; mbs = a->mbs; bs = a->bs;
371dc008846SHong Zhang   ai = a->i; aj = a->j;
372dc008846SHong Zhang   bi = b->i; bj = b->j;
373632d0f97SHong Zhang   garray = c->garray;
374dc008846SHong Zhang   rstart = c->rstart;
375dc008846SHong Zhang   is_max = data[0];
376632d0f97SHong Zhang 
377fc70d14eSHong Zhang   ierr = PetscBTCreate(Mbs,table0);CHKERRQ(ierr);
378fc70d14eSHong Zhang 
379fc70d14eSHong Zhang   ierr = PetscMalloc((1+is_max*(Mbs+1))*sizeof(int),&nidx);CHKERRQ(ierr);
380fc70d14eSHong Zhang   nidx[0] = is_max;
381fc70d14eSHong Zhang 
382fc70d14eSHong Zhang   idx_i  = data + is_max + 1; /* ptr to input is[0] array */
383bfc6803cSHong Zhang   nidx_i = nidx + is_max + 1; /* ptr to output is[0] array */
384dc008846SHong Zhang   for (i=0; i<is_max; i++) { /* for each is */
385dc008846SHong Zhang     isz  = 0;
386fc70d14eSHong Zhang     n = data[1+i]; /* size of input is[i] */
387dc008846SHong Zhang 
388bfc6803cSHong Zhang     /* initialize table_i, set table0 */
389bfc6803cSHong Zhang     if (whose == MINE){ /* process this processor's is[] */
390bfc6803cSHong Zhang       table_i = table[i];
391*0472cc68SHong Zhang       nidx_i  = nidx + 1+ is_max + Mbs*i;
392bfc6803cSHong Zhang     } else {            /* process other processor's is[] - only use one temp table */
393bfc6803cSHong Zhang       table_i = *table;
394bfc6803cSHong Zhang     }
395bfc6803cSHong Zhang     ierr = PetscBTMemzero(Mbs,table_i);CHKERRQ(ierr);
396bfc6803cSHong Zhang     ierr = PetscBTMemzero(Mbs,table0);CHKERRQ(ierr);
397fc70d14eSHong Zhang     if (n > 0) {
398bfc6803cSHong Zhang       isz0 = 0;
399dc008846SHong Zhang       for (j=0; j<n; j++){
400dc008846SHong Zhang         col = idx_i[j];
401dc008846SHong Zhang         if (col >= Mbs) SETERRQ2(PETSC_ERR_ARG_OUTOFRANGE,"index col %d >= Mbs %d",col,Mbs);
402bfc6803cSHong Zhang         if(!PetscBTLookupSet(table_i,col)) {
403bfc6803cSHong Zhang           ierr = PetscBTSet(table0,col);CHKERRQ(ierr);
404*0472cc68SHong Zhang           if (whose == MINE) {nidx_i[isz0] = col;}
405bfc6803cSHong Zhang           isz0++;
406bfc6803cSHong Zhang         }
407632d0f97SHong Zhang       }
408dc008846SHong Zhang 
409*0472cc68SHong Zhang       if (whose == MINE) {isz = isz0;}
410fc70d14eSHong Zhang       k = 0;  /* no. of indices from input is[i] that have been examined */
411dc008846SHong Zhang       for (row=0; row<mbs; row++){
412dc008846SHong Zhang         a_start = ai[row]; a_end = ai[row+1];
413dc008846SHong Zhang         b_start = bi[row]; b_end = bi[row+1];
414*0472cc68SHong Zhang         if (PetscBTLookup(table0,row+rstart)){ /* row is on input is[i]:
415*0472cc68SHong Zhang            do row search: collect all col in this row */
416dc008846SHong Zhang           for (l = a_start; l<a_end ; l++){ /* Amat */
417dc008846SHong Zhang             col = aj[l] + rstart;
418fc70d14eSHong Zhang             if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
419dc008846SHong Zhang           }
420dc008846SHong Zhang           for (l = b_start; l<b_end ; l++){ /* Bmat */
421dc008846SHong Zhang             col = garray[bj[l]];
422fc70d14eSHong Zhang             if (!PetscBTLookupSet(table_i,col)) {nidx_i[isz++] = col;}
423dc008846SHong Zhang           }
424dc008846SHong Zhang           k++;
425dc008846SHong Zhang           if (k >= isz0) break; /* for (row=0; row<mbs; row++) */
426*0472cc68SHong Zhang         } else { /* row is not on input is[i]:
427*0472cc68SHong Zhang           do col serach: add row onto nidx_i if there is a col in nidx_i */
428dc008846SHong Zhang           for (l = a_start; l<a_end ; l++){ /* Amat */
429dc008846SHong Zhang             col = aj[l] + rstart;
430dc008846SHong Zhang             if (PetscBTLookup(table0,col)){
431fc70d14eSHong Zhang               if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
432dc008846SHong Zhang               break; /* for l = start; l<end ; l++) */
433632d0f97SHong Zhang             }
434632d0f97SHong Zhang           }
435dc008846SHong Zhang           for (l = b_start; l<b_end ; l++){ /* Bmat */
436dc008846SHong Zhang             col = garray[bj[l]];
437dc008846SHong Zhang             if (PetscBTLookup(table0,col)){
438fc70d14eSHong Zhang               if (!PetscBTLookupSet(table_i,row+rstart)) {nidx_i[isz++] = row+rstart;}
439dc008846SHong Zhang               break; /* for l = start; l<end ; l++) */
440632d0f97SHong Zhang             }
441dc008846SHong Zhang           }
442dc008846SHong Zhang         }
443bfc6803cSHong Zhang       }
444fc70d14eSHong Zhang     } /* if (n > 0) */
445dc008846SHong Zhang 
446dc008846SHong Zhang     if (i < is_max - 1){
447fc70d14eSHong Zhang       idx_i  += n;   /* ptr to input is[i+1] array */
448bfc6803cSHong Zhang       nidx_i += isz; /* ptr to output is[i+1] array */
449dc008846SHong Zhang     }
450fc70d14eSHong Zhang     nidx[1+i] = isz; /* size of new is[i] */
451dc008846SHong Zhang   } /* /* for each is */
452dc008846SHong Zhang   *data_new = nidx;
453dc008846SHong Zhang   ierr = PetscBTDestroy(table0);CHKERRQ(ierr);
454632d0f97SHong Zhang 
455632d0f97SHong Zhang   PetscFunctionReturn(0);
456632d0f97SHong Zhang }
457632d0f97SHong Zhang 
458632d0f97SHong Zhang 
459