// this demonstrates that shmem_collect{32} can concatenate different
// lengths of data from diff PEs.

// cc tees.c -lshmem -lelan
// prun -n4 a.out

#include <shmem.h>

int _num_pes(void);
int _my_pe(void);

int nn, NN;
int alldata[2000];
int data[100];

long pSync[_SHMEM_REDUCE_SYNC_SIZE];
int pWork[_SHMEM_REDUCE_MIN_WRKDATA_SIZE];

main()
{
   int i, mype, npes;

   shmem_init();
   npes = _num_pes();
   mype = _my_pe();

   nn = mype+1;
   for (i=0; i<nn; i++) data[i] = i + 10*mype;

   for ( i=0; i<_SHMEM_REDUCE_SYNC_SIZE; ++i )
      pSync[i] = _SHMEM_SYNC_VALUE;

   shmem_barrier_all();

   shmem_int_sum_to_all ( &NN, &nn, 1, 0,0,npes, pWork, pSync );

   shmem_collect32 ( alldata, data, nn, 0,0,npes, pSync );

   if (mype==0)
      for (i=0; i<NN; i++) printf("%2d  %d\n", i, alldata[i]);
}