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

   MPI_Comm cart_comm;
   int cart_rank;
   
void SendToRootAndWrite(int sendvalue, int np, int nq) {
    int *values, numvalues;
    int i, j;
    
    numvalues=1;

   if ( (values = (int*)malloc(np*sizeof(int))) == NULL ) {
   	   fprintf(stderr,"Failed to allocate values[%d]",np);
   	   exit(-1);
   }
   values[0] = values[1] = 0;
	if ( MPI_Gather ( &sendvalue, 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 < nq; i++ ) {
          for (j=0; j < nq; j++ ) {
             np = i + nq*j;
             printf(" %d",values[np]);
          } printf("\n");
       }
    }
}  
 
int main(int argc, char **argv) {
   int i, j, p, q, rc;
   float p_real;
   int my_rank, source, dest;
   int ldims[2], coords[2];
   int periods[2], reorder;	

   MPI_Status status;
   int cval;

   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);
   }
   
/*
     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 = 100*(1+coords[0])+(1+coords[1]);
    printf("For process %d coords=(%d,%d), cval = %d\n",cart_rank,
          coords[0],coords[1],cval);
          
/*   Collectively send value to root process */
    SendToRootAndWrite(cval,p,q);

/*   Now send value to the process on the right */
    if ( cart_rank == 0 ) fprintf(stdout,"Send values in direction 0\n");
    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 */
    SendToRootAndWrite(cval,p,q);

/*   Now send value downward */
    if ( cart_rank == 0 ) fprintf(stdout,"Send values in direction 1\n");
    if ( MPI_Cart_shift(cart_comm, 1, 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 */
    SendToRootAndWrite(cval,p,q);

!
    MPI_Finalize();
    return(0);
}