#include <stdlib.h>
#include <stdio.h>
#include <math.h>
#include <mpi.h> 

int main(int argc, char **argv) {
   int i, j, p, q, rc;
   float p_real;
   int my_rank, cart_rank, source, dest;
   int ldims[2], coords[2];
   int periods[2], reorder;	
   MPI_Comm cart_comm;
   MPI_Status status;
   int cval, numvalues;
   int *values;

   if ( MPI_Init(&argc, &argv) != MPI_SUCCESS ) exit(-1);
   if ( MPI_Comm_size(MPI_COMM_WORLD, &p) != MPI_SUCCESS ) exit(-1);
   if ( MPI_Comm_rank(MPI_COMM_WORLD, &my_rank) != MPI_SUCCESS ) exit(-1);

   p_real = p;  q = sqrt(p_real);
   if ( q*q != p ) {
         if ( my_rank == 0 ) printf(" %d**2 != %d: program terminates",q,p);
         MPI_Finalize();
         exit(-1);
   }
   
   if ( (values = (int*)malloc(p*sizeof(int))) == NULL ) {
   	   fprintf(stderr,"Failed to allocate values[%d]",p);
   	   exit(-1);
   }
/*
     Create new communicator QxQ dimensions
*/
	ldims[0] = ldims[1] = q;
	periods[0] = periods[1] = 1;  // true
	reorder = 0;  // false
	if ( MPI_Cart_create(MPI_COMM_WORLD,2,ldims,periods,reorder,&cart_comm) !=
		MPI_SUCCESS ) exit(-1);
/*   Get process coordinates */
    if ( MPI_Comm_rank(cart_comm, &cart_rank) != MPI_SUCCESS ) exit(-1);
	if ( MPI_Cart_coords(cart_comm, cart_rank, 2, coords) != MPI_SUCCESS ) exit(-1);
/*   Each process defines a value equal to 2nd coordinate */
	cval = coords[0];
    numvalues=1;
    values[0] = values[1] = 0;
/*   Collectively send value to root process */
    if ( MPI_Gather ( &cval, 1, MPI_INTEGER, values, numvalues, MPI_INTEGER, 
    	0, cart_comm ) != MPI_SUCCESS ) exit(-1);
    if ( cart_rank == 0 ) {
       printf("Before sending, values = \n");
       for (i=0; i < q; i++ ) {
          for (j=0; j < q; j++ ) {
             p = i + q*j;
             printf(" %d",values[p]);
          } printf("\n");
       }
    }
/*   Now send value to the process on the right */
    if ( MPI_Cart_shift(cart_comm, 0, 1, &source, &dest) != MPI_SUCCESS ) exit(-1);
    if ( MPI_Sendrecv_replace(&cval, 1, MPI_INTEGER, dest, 0, source, 0, cart_comm, &status) != MPI_SUCCESS ) exit(-1);
/*   Again, collectively send value to root process */
    numvalues=1;
	if ( MPI_Gather ( &cval, 1, MPI_INTEGER, values, numvalues, MPI_INTEGER, 
		0, cart_comm ) != MPI_SUCCESS ) exit(-1);
    if ( cart_rank == 0 ) {
       printf("After sending, values = \n");
       for (i=0; i < q; i++ ) {
          for (j=0; j < q; j++ ) {
             p = i + q*j;
             printf(" %d",values[p]);
          } printf("\n");
       }
    }
!
    MPI_Finalize();
    return(0);
}