xref: /petsc/src/mat/impls/aij/mpi/mpiov.c (revision c4fc05e7d7dc2f7dfe866089f580bf3d2cc1b461)
1 #ifndef lint
2 static char vcid[] = "$Id: mpiov.c,v 1.10 1996/02/01 00:24:06 balay Exp balay $";
3 #endif
4 
5 #include "mpiaij.h"
6 #include "inline/bitarray.h"
7 
8 int MatIncreaseOverlap_MPIAIJ_private(Mat, int, IS *);
9 
10 int MatIncreaseOverlap_MPIAIJ(Mat C, int is_max, IS *is, int ov)
11 {
12   int i, ierr;
13   if (ov < 0){ SETERRQ(1," MatIncreaseOverlap_MPIAIJ: negative overlap specified\n");}
14   for (i =0; i<ov; ++i) {
15     ierr = MatIncreaseOverlap_MPIAIJ_private(C, is_max, is); CHKERRQ(ierr);
16   }
17   return 0;
18 }
19 int FindOverlapLocal(Mat , int , char **,int*, int**);
20 int FindOverlapRecievedMesg(Mat , int, int **, int**, int* );
21 
22 /*
23   Sample message format:
24   If a processor A wants processor B to process some elements corresponding
25   to index sets 1s[1], is[5]
26   mesg [0] = 2   ( no of index sets in the mesg)
27   -----------
28   mesg [1] = 1 => is[1]
29   mesg [2] = sizeof(is[1]);
30   -----------
31   mesg [5] = 5  => is[5]
32   mesg [6] = sizeof(is[5]);
33   -----------
34   mesg [7]
35   mesg [n]  datas[1]
36   -----------
37   mesg[n+1]
38   mesg[m]  data(is[5])
39   -----------
40 */
41 int MatIncreaseOverlap_MPIAIJ_private(Mat C, int is_max, IS *is)
42 {
43   Mat_MPIAIJ  *c = (Mat_MPIAIJ *) C->data;
44   int         **idx, *n, *w1, *w2, *w3, *w4, *rtable,**data;
45   int         size, rank, m,i,j,k, ierr, **rbuf, row, proc, mct, msz, **outdat, **ptr;
46   int         *ctr, *pa, tag, *tmp,bsz, nmsg , *isz, *isz1, **xdata;
47   int          bsz1, **rbuf2;
48   char        **table;
49   MPI_Comm    comm;
50   MPI_Request *send_waits,*recv_waits,*send_waits2,*recv_waits2 ;
51   MPI_Status  *send_status ,*recv_status;
52   double         space, fr, maxs;
53 
54   comm   = C->comm;
55   tag    = C->tag;
56   size   = c->size;
57   rank   = c->rank;
58   m      = c->M;
59 
60 
61   TrSpace( &space, &fr, &maxs );
62   /*  MPIU_printf(MPI_COMM_SELF,"[%d] allocated space = %f fragments = %f max ever allocated = %f\n", rank, space, fr, maxs); */
63 
64   idx    = (int **)PetscMalloc((is_max+1)*sizeof(int *)); CHKPTRQ(idx);
65   n      = (int *)PetscMalloc((is_max+1)*sizeof(int )); CHKPTRQ(n);
66   rtable = (int *)PetscMalloc((m+1)*sizeof(int )); CHKPTRQ(rtable);
67                                 /* Hash table for maping row ->proc */
68 
69   for ( i=0 ; i<is_max ; i++) {
70     ierr = ISGetIndices(is[i],&idx[i]);  CHKERRQ(ierr);
71     ierr = ISGetLocalSize(is[i],&n[i]);  CHKERRQ(ierr);
72   }
73 
74   /* Create hash table for the mapping :row -> proc*/
75   for( i=0, j=0; i< size; i++) {
76     for (; j <c->rowners[i+1]; j++) {
77       rtable[j] = i;
78     }
79   }
80 
81   /* evaluate communication - mesg to who, length of mesg, and buffer space
82      required. Based on this, buffers are allocated, and data copied into them*/
83   w1     = (int *)PetscMalloc((size)*4*sizeof(int )); CHKPTRQ(w1); /*  mesg size */
84   w2     = w1 + size;         /* if w2[i] marked, then a message to proc i*/
85   w3     = w2 + size;         /* no of IS that needs to be sent to proc i */
86   w4     = w3 + size;         /* temp work space used in determining w1, w2, w3 */
87   PetscMemzero(w1,(size)*3*sizeof(int)); /* initialise work vector*/
88   for ( i=0;  i<is_max ; i++) {
89     PetscMemzero(w4,(size)*sizeof(int)); /* initialise work vector*/
90     for ( j =0 ; j < n[i] ; j++) {
91       row  = idx[i][j];
92       proc = rtable[row];
93       w4[proc]++;
94     }
95     for( j = 0; j < size; j++){
96       if( w4[j] ) { w1[j] += w4[j];  w3[j] += 1;}
97     }
98   }
99 
100   mct      = 0;              /* no of outgoing messages */
101   msz      = 0;              /* total mesg length (for all proc */
102   w1[rank] = 0;              /* no mesg sent to intself */
103   w3[rank] = 0;
104   for (i =0; i < size ; i++) {
105     if (w1[i])  { w2[i] = 1; mct++;} /* there exists a message to proc i */
106   }
107   pa = (int *)PetscMalloc((mct +1)*sizeof(int)); CHKPTRQ(pa); /* (proc -array) */
108   for (i =0, j=0; i < size ; i++) {
109     if (w1[i]) { pa[j] = i; j++; }
110   }
111 
112   /* Each message would have a header = 1 + 2*(no of IS) + data */
113   for (i = 0; i<mct ; i++) {
114     j = pa[i];
115     w1[j] += w2[j] + 2* w3[j];
116     msz   += w1[j];
117   }
118 
119   /* Allocate Memory for outgoing messages */
120   outdat    = (int **)PetscMalloc( 2*size*sizeof(int*)); CHKPTRQ(outdat);
121   PetscMemzero(outdat,  2*size*sizeof(int*));
122   tmp       = (int *)PetscMalloc((msz+1) *sizeof (int)); CHKPTRQ(tmp); /*mrsg arr */
123   ptr       = outdat +size;     /* Pointers to the data in outgoing buffers */
124   ctr       = (int *)PetscMalloc( size*sizeof(int));   CHKPTRQ(ctr);
125 
126   {
127     int *iptr = tmp;
128     int ict  = 0;
129     for (i = 0; i < mct ; i++) {
130       j         = pa[i];
131       iptr     +=  ict;
132       outdat[j] = iptr;
133       ict       = w1[j];
134     }
135   }
136 
137   /* Form the outgoing messages */
138   /*plug in the headers*/
139   for ( i=0 ; i<mct ; i++) {
140     j = pa[i];
141     outdat[j][0] = 0;
142     PetscMemzero(outdat[j]+1, 2 * w3[j]*sizeof(int));
143     ptr[j] = outdat[j] + 2*w3[j] +1;
144   }
145 
146   /* Memory for doing local proc's work*/
147   table = (char **)PetscMalloc((is_max+1)*sizeof(int *));  CHKPTRQ(table);
148   data  = (int **)PetscMalloc((is_max+1)*sizeof(int *)); CHKPTRQ(data);
149   table[0] = (char *)PetscMalloc((m/BITSPERBYTE +1)*(is_max)); CHKPTRQ(table[0]);
150   data [0] = (int *)PetscMalloc((m+1)*(is_max)*sizeof(int)); CHKPTRQ(data[0]);
151 
152   for(i = 1; i<is_max ; i++) {
153     table[i] = table[0] + (m/BITSPERBYTE+1)*i;
154     data[i]  = data[0] + (m+1)*i;
155   }
156 
157   PetscMemzero((void*)*table,(m/BITSPERBYTE+1)*(is_max));
158   isz = (int *)PetscMalloc((is_max+1) *sizeof(int)); CHKPTRQ(isz);
159   PetscMemzero((void *)isz,(is_max+1) *sizeof(int));
160 
161   /* Parse the IS and update local tables and the outgoing buf with the data*/
162   for ( i=0 ; i<is_max ; i++) {
163     PetscMemzero(ctr,size*sizeof(int));
164     for( j=0;  j<n[i]; j++) {  /* parse the indices of each IS */
165       row  = idx[i][j];
166       proc = rtable[row];
167       if (proc != rank) { /* copy to the outgoing buf*/
168         ctr[proc]++;
169         *ptr[proc] = row;
170         ptr[proc]++;
171       }
172       else { /* Update the table */
173         if ( !BT_LOOKUP(table[i],row)) { data[i][isz[i]++] = row;}
174       }
175     }
176     /* Update the headers for the current IS */
177     for( j = 0; j<size; j++) { /* Can Optimise this loop too */
178       if (ctr[j]) {
179         k= ++outdat[j][0];
180         outdat[j][2*k]   = ctr[j];
181         outdat[j][2*k-1] = i;
182       }
183     }
184   }
185 
186   /* I nolonger need the original indices*/
187   for( i=0; i< is_max; ++i) {
188     ierr = ISRestoreIndices(is[i], idx+i); CHKERRQ(ierr);
189   }
190   PetscFree(idx);
191   PetscFree(n);
192   PetscFree(rtable);
193   for( i=0; i< is_max; ++i) {
194     ierr = ISDestroy(is[i]); CHKERRQ(ierr);
195   }
196 
197 
198   /* Do a global reduction to determine how many messages to expect*/
199   {
200     int *rw1, *rw2;
201     rw1 = (int *)PetscMalloc(2*size*sizeof(int)); CHKPTRQ(rw1);
202     rw2 = rw1+size;
203     MPI_Allreduce((void *)w1, rw1, size, MPI_INT, MPI_MAX, comm);
204     bsz   = rw1[rank];
205     MPI_Allreduce((void *)w2, rw2, size, MPI_INT, MPI_SUM, comm);
206     nmsg  = rw2[rank];
207     PetscFree(rw1);
208   }
209 
210   /* Allocate memory for recv buffers . Prob none if nmsg = 0 ???? */
211   rbuf    = (int**) PetscMalloc((nmsg+1) *sizeof(int*));  CHKPTRQ(rbuf);
212   rbuf[0] = (int *) PetscMalloc((nmsg *bsz+1) * sizeof(int));  CHKPTRQ(rbuf[0]);
213   for (i=1; i<nmsg ; ++i) rbuf[i] = rbuf[i-1] + bsz;
214 
215   /* Now post the receives */
216   recv_waits = (MPI_Request *) PetscMalloc((nmsg+1)*sizeof(MPI_Request));
217   CHKPTRQ(recv_waits);
218   for ( i=0; i<nmsg; ++i){
219     MPI_Irecv((void *)rbuf[i], bsz, MPI_INT, MPI_ANY_SOURCE, tag, comm, recv_waits+i);
220   }
221 
222   /*  Now  post the sends */
223   send_waits = (MPI_Request *) PetscMalloc((mct+1)*sizeof(MPI_Request));
224   CHKPTRQ(send_waits);
225   for( i =0; i< mct; ++i){
226     j = pa[i];
227     MPI_Isend( (void *)outdat[j], w1[j], MPI_INT, j, tag, comm, send_waits+i);
228   }
229 
230   /* Do Local work*/
231   ierr = FindOverlapLocal(C, is_max, table,isz, data); CHKERRQ(ierr);
232   /* Extract the matrices */
233 
234   /* Receive messages*/
235   {
236     int        index;
237 
238     recv_status = (MPI_Status *) PetscMalloc( (nmsg+1)*sizeof(MPI_Status) );
239     CHKPTRQ(recv_status);
240     for ( i=0; i< nmsg; ++i) {
241       MPI_Waitany(nmsg, recv_waits, &index, recv_status+i);
242     }
243 
244     send_status = (MPI_Status *) PetscMalloc( (mct+1)*sizeof(MPI_Status) );
245     CHKPTRQ(send_status);
246     MPI_Waitall(mct,send_waits,send_status);
247   }
248   /* Pahse 1 sends are complete - deallocate buffers */
249   PetscFree(outdat);
250   PetscFree(w1);
251   PetscFree(tmp);
252 
253   /* int FindOverlapRecievedMesg(Mat C, int is_max, int *isz, char **table, int **data)*/
254   xdata    = (int **)PetscMalloc((nmsg+1)*sizeof(int *)); CHKPTRQ(xdata);
255   isz1     = (int *)PetscMalloc((nmsg+1) *sizeof(int)); CHKPTRQ(isz1);
256   ierr = FindOverlapRecievedMesg(C, nmsg, rbuf,xdata,isz1); CHKERRQ(ierr);
257 
258   /* Nolonger need rbuf. */
259   PetscFree(rbuf[0]);
260   PetscFree(rbuf);
261 
262 
263   /* Send the data back*/
264   /* Do a global reduction to know the buffer space req for incoming messages*/
265   {
266     int *rw1, *rw2;
267 
268     rw1 = (int *)PetscMalloc(2*size*sizeof(int)); CHKPTRQ(rw1);
269     PetscMemzero((void*)rw1,2*size*sizeof(int));
270     rw2 = rw1+size;
271     for (i =0; i < nmsg ; ++i) {
272       proc      = recv_status[i].MPI_SOURCE;
273       rw1[proc] = isz1[i];
274     }
275 
276     MPI_Allreduce((void *)rw1, (void *)rw2, size, MPI_INT, MPI_MAX, comm);
277     bsz1   = rw2[rank];
278     PetscFree(rw1);
279   }
280 
281   /* Allocate buffers*/
282 
283   /* Allocate memory for recv buffers . Prob none if nmsg = 0 ???? */
284   rbuf2    = (int**) PetscMalloc((mct+1) *sizeof(int*));  CHKPTRQ(rbuf2);
285   rbuf2[0] = (int *) PetscMalloc((mct*bsz1+1) * sizeof(int));  CHKPTRQ(rbuf2[0]);
286   for (i=1; i<mct ; ++i) rbuf2[i] = rbuf2[i-1] + bsz1;
287 
288   /* Now post the receives */
289   recv_waits2 = (MPI_Request *)PetscMalloc((mct+1)*sizeof(MPI_Request)); CHKPTRQ(recv_waits2)
290   CHKPTRQ(recv_waits2);
291   for ( i=0; i<mct; ++i){
292     MPI_Irecv((void *)rbuf2[i], bsz1, MPI_INT, MPI_ANY_SOURCE, tag, comm, recv_waits2+i);
293   }
294 
295   /*  Now  post the sends */
296   send_waits2 = (MPI_Request *) PetscMalloc((nmsg+1)*sizeof(MPI_Request));
297   CHKPTRQ(send_waits2);
298   for( i =0; i< nmsg; ++i){
299     j = recv_status[i].MPI_SOURCE;
300     MPI_Isend( (void *)xdata[i], isz1[i], MPI_INT, j, tag, comm, send_waits2+i);
301   }
302 
303   /* recieve work done on other processors*/
304   {
305     int         index, is_no, ct1, max;
306     MPI_Status  *send_status2 ,*recv_status2;
307 
308     recv_status2 = (MPI_Status *) PetscMalloc( (mct+1)*sizeof(MPI_Status) );
309     CHKPTRQ(recv_status2);
310 
311 
312     for ( i=0; i< mct; ++i) {
313       MPI_Waitany(mct, recv_waits2, &index, recv_status2+i);
314       /* Process the message*/
315       ct1 = 2*rbuf2[index][0]+1;
316       for (j=1; j<=rbuf2[index][0]; j++) {
317         max   = rbuf2[index][2*j];
318         is_no = rbuf2[index][2*j-1];
319         for (k=0; k < max ; k++, ct1++) {
320           row = rbuf2[index][ct1];
321           if(!BT_LOOKUP(table[is_no],row)) { data[is_no][isz[is_no]++] = row;}
322         }
323       }
324     }
325 
326 
327     send_status2 = (MPI_Status *) PetscMalloc( (nmsg+1)*sizeof(MPI_Status) );
328     CHKPTRQ(send_status2);
329     MPI_Waitall(nmsg,send_waits2,send_status2);
330 
331     PetscFree(send_status2); PetscFree(recv_status2);
332   }
333 
334   TrSpace( &space, &fr, &maxs );
335   /*  MPIU_printf(MPI_COMM_SELF,"[%d] allocated space = %f fragments = %f max ever allocated = %f\n", rank, space, fr, maxs);*/
336 
337   PetscFree(ctr);
338   PetscFree(pa);
339   PetscFree(rbuf2[0]);
340   PetscFree(rbuf2);
341   PetscFree(send_waits);
342   PetscFree(recv_waits);
343   PetscFree(send_waits2);
344   PetscFree(recv_waits2);
345   PetscFree(table[0]);
346   PetscFree(table);
347   PetscFree(send_status);
348   PetscFree(recv_status);
349   PetscFree(isz1);
350   PetscFree(xdata[0]);
351   PetscFree(xdata);
352 
353   for ( i=0; i<is_max; ++i) {
354     ierr = ISCreateSeq(MPI_COMM_SELF, isz[i], data[i], is+i); CHKERRQ(ierr);
355   }
356   PetscFree(isz);
357   PetscFree(data[0]);
358   PetscFree(data);
359 
360   return 0;
361 }
362 
363 /*   FindOverlapLocal() - Called by MatincreaseOverlap, to do the work on
364      the local processor.
365 
366      Inputs:
367       C      - MAT_MPIAIJ;
368       is_max - total no of index sets processed at a time;
369       table  - an array of char - size = m bits.
370 
371      Output:
372       isz    - array containing the count of the solution elements correspondign
373                to each index set;
374       data   - pointer to the solutions
375 */
376 int FindOverlapLocal(Mat C, int is_max, char **table, int *isz,int **data)
377 {
378   Mat_MPIAIJ *c = (Mat_MPIAIJ *) C->data;
379   Mat        A = c->A, B = c->B;
380   Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
381   int        start, end, val, max, rstart,cstart, ashift, bshift,*ai, *aj;
382   int        *bi, *bj, *garray, i, j, k, row;
383 
384   rstart = c->rstart;
385   cstart = c->cstart;
386   ashift = a->indexshift;
387   ai     = a->i;
388   aj     = a->j +ashift;
389   bshift = b->indexshift;
390   bi     = b->i;
391   bj     = b->j +bshift;
392   garray = c->garray;
393 
394 
395   for( i=0; i<is_max; i++) {
396     for ( j=0, max =isz[i] ; j< max; j++) {
397       row   = data[i][j] - rstart;
398       start = ai[row];
399       end   = ai[row+1];
400       for ( k=start; k < end; k++) { /* Amat */
401         val = aj[k] + ashift + cstart;
402         if(!BT_LOOKUP(table[i],val)) { data[i][isz[i]++] = val;}
403       }
404       start = bi[row];
405       end   = bi[row+1];
406       for ( k=start; k < end; k++) { /* Bmat */
407         val = garray[bj[k]+bshift] ;
408         if(! BT_LOOKUP(table[i],val)) { data[i][isz[i]++] = val;}
409       }
410     }
411   }
412 
413 return 0;
414 }
415 /*       FindOverlapRecievedMesg - Process the recieved messages,
416          and return the output
417 
418          Input:
419            C    - the matrix
420            nmsg - no of messages being processed.
421            rbuf - an array of pointers to the recieved requests
422 
423          Output:
424            xdata - array of messages to be sent back
425            isz1  - size of each message
426 */
427 int FindOverlapRecievedMesg(Mat C, int nmsg, int ** rbuf, int ** xdata, int * isz1 )
428 {
429   Mat_MPIAIJ  *c = (Mat_MPIAIJ *) C->data;
430   Mat         A = c->A, B = c->B;
431   Mat_SeqAIJ  *a = (Mat_SeqAIJ*)A->data,*b = (Mat_SeqAIJ*)B->data;
432   int         rstart,cstart, ashift, bshift,*ai, *aj, *bi, *bj, *garray, i, j, k;
433   int         row,total_sz,ct, ct1, ct2, ct3,mem_estimate, oct2, l, start, end;
434   int         val, max1, max2, rank, m, no_malloc =0, *tmp, new_estimate, ctr;
435   char        *xtable;
436 
437   rank   = c->rank;
438   m      = c->M;
439   rstart = c->rstart;
440   cstart = c->cstart;
441   ashift = a->indexshift;
442   ai     = a->i;
443   aj     = a->j +ashift;
444   bshift = b->indexshift;
445   bi     = b->i;
446   bj     = b->j +bshift;
447   garray = c->garray;
448 
449 
450   for (i =0, ct =0, total_sz =0; i< nmsg; ++i){
451     ct+= rbuf[i][0];
452     for ( j = 1; j <= rbuf[i][0] ; j++ ) { total_sz += rbuf[i][2*j]; }
453     }
454 
455   max1 = ct*(a->nz +b->nz)/c->m;
456   mem_estimate =  3*((total_sz > max1?total_sz:max1)+1);
457   xdata[0] = (int *)PetscMalloc(mem_estimate *sizeof(int)); CHKPTRQ(xdata[0]);
458   ++no_malloc;
459   xtable   = (char *)PetscMalloc((m/BITSPERBYTE+1)); CHKPTRQ(xtable);
460   PetscMemzero((void *)isz1,(nmsg+1) *sizeof(int));
461 
462   ct3 = 0;
463   for (i =0; i< nmsg; i++) { /* for easch mesg from proc i */
464     ct1 = 2*rbuf[i][0]+1;
465     ct2 = ct1;
466     ct3+= ct1;
467     for (j = 1, max1= rbuf[i][0]; j<=max1; j++) { /* for each IS from proc i*/
468       PetscMemzero((void *)xtable,(m/BITSPERBYTE+1));
469       oct2 = ct2;
470       for (k =0; k < rbuf[i][2*j]; k++, ct1++) {
471         row = rbuf[i][ct1];
472         if(!BT_LOOKUP(xtable,row)) {
473           if (!(ct3 < mem_estimate)) {
474             new_estimate = 1.5*mem_estimate;
475             tmp = (int*) PetscMalloc(new_estimate * sizeof(int)); CHKPTRQ(tmp);
476             PetscMemcpy((char *)tmp,(char *)xdata[0],mem_estimate*sizeof(int));
477             PetscFree(xdata[0]);
478             xdata[0] = tmp;
479             mem_estimate = new_estimate; ++no_malloc;
480             for (ctr =1; ctr <=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
481           }
482            xdata[i][ct2++] = row;ct3++;
483         }
484       }
485       for ( k=oct2, max2 =ct2 ; k< max2; k++) {
486         row   = xdata[i][k] - rstart;
487         start = ai[row];
488         end   = ai[row+1];
489         for ( l=start; l < end; l++) {
490           val = aj[l] +ashift + cstart;
491           if(!BT_LOOKUP(xtable,val)) {
492             if (!(ct3 < mem_estimate)) {
493               new_estimate = 1.5*mem_estimate;
494               tmp = (int*) PetscMalloc(new_estimate * sizeof(int)); CHKPTRQ(tmp);
495               PetscMemcpy((char *)tmp,(char *)xdata[0],mem_estimate*sizeof(int));
496               PetscFree(xdata[0]);
497               xdata[0] = tmp;
498               mem_estimate = new_estimate; ++no_malloc;
499               for (ctr =1; ctr <=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
500             }
501             xdata[i][ct2++] = val;ct3++;
502           }
503         }
504         start = bi[row];
505         end   = bi[row+1];
506         for ( l=start; l < end; l++) {
507           val = garray[bj[l]+bshift] ;
508           if(!BT_LOOKUP(xtable,val)) {
509             if (!(ct3 < mem_estimate)) {
510               new_estimate = 1.5*mem_estimate;
511               tmp = (int*) PetscMalloc(new_estimate * sizeof(int)); CHKPTRQ(tmp);
512               PetscMemcpy((char *)tmp,(char *)xdata[0],mem_estimate*sizeof(int));
513               PetscFree(xdata[0]);
514               xdata[0] = tmp;
515               mem_estimate = new_estimate; ++no_malloc;
516               for (ctr =1; ctr <=i; ctr++) { xdata[ctr] = xdata[ctr-1] + isz1[ctr-1];}
517             }
518             xdata[i][ct2++] = val;ct3++;
519           }
520         }
521       }
522       /* Update the header*/
523       xdata[i][2*j]   = ct2-oct2; /* Undo the vector isz1 and use only a var*/
524       xdata[i][2*j-1] = rbuf[i][2*j-1];
525     }
526     xdata[i][0] = rbuf[i][0];
527     xdata[i+1]  = xdata[i] +ct2;
528     isz1[i]     = ct2; /* size of each message */
529   }
530   PetscFree(xtable);
531   PLogInfo(0,"MatIncreaseOverlap_MPIAIJ:[%d] Allocated %d bytes, required %d bytes, no of mallocs = %d\n",rank,mem_estimate, ct3,no_malloc);
532   return 0;
533 }
534