/*
 * MPI version of pipeline.
 *
 */

#include <assert.h>
#include <stdio.h>
#include <errno.h>
#include <mpi.h>

/* Let's assume a picture is stored as an array of N integers */

#define num_pic 10
#define N 20
#define tag 123


/**************************************************************************
 * Processor 1
 *************************************************************************/

void get_picture_from_somewhere(int* pic) 
{
  static int i = 0;
  pic[0] = i++;			/* or whatever */
}

void trans1(int *src, int *dst)	/* transforms src to dst */
{
  printf("Stage 1: Processing picture %d\n", src[0]);
  dst[0] = src[0];
}

void stage1 ()
{
  int in_pic[N], int_pic_1[N];
  int i;
  for (i=0; i<10; i++) {
    get_picture_from_somewhere(in_pic);
    trans1(in_pic, int_pic_1);
    MPI_Send (int_pic_1, N, MPI_INT, 1, tag, MPI_COMM_WORLD);
  }
} 


/**************************************************************************
 * Processor 2
 *************************************************************************/

void trans2 (int *src, int *dst)
{
  printf("Stage 2: Processing picture %d\n", src[0]);
  dst[0] = src[0];
}

void stage2 ()
{
  MPI_Status status;
  int int_pic_1[N], int_pic_2[N];
  int i;
  for (i=0; i<num_pic; i++ ) {
    MPI_Recv (int_pic_1, N, MPI_INT, 0, tag, MPI_COMM_WORLD, &status);
    trans2 (int_pic_1, int_pic_2);
    MPI_Send (int_pic_2, N, MPI_INT, 2, tag, MPI_COMM_WORLD);
  }
} 


/**************************************************************************
 * Processor 3
 *************************************************************************/

void trans3 (int *src, int *dst)
{
  printf("Stage 3: Processing picture %d\n", src[0]);
  dst[0] = src[0];
}

void stage3 ()
{
  MPI_Status status;
  int int_pic_2[N], int_pic_3[N];
  int i;
  for (i=0; i<num_pic; i++ ) {
    MPI_Recv (int_pic_2, N, MPI_INT, 1, tag, MPI_COMM_WORLD, &status);
    trans3 (int_pic_2, int_pic_3);
    MPI_Send (int_pic_3, N, MPI_INT, 3, tag, MPI_COMM_WORLD);
  }
} 


/**************************************************************************
 * Processor 4
 *************************************************************************/

void trans4 (int *src, int *dst)
{
  printf("Stage 4: Processing picture %d\n", src[0]);
  dst[0] = src[0];
}


void stage4 ()
{
  MPI_Status status;
  int int_pic_3[N], out_pic[N];
  int i;
  for (i=0; i<num_pic; i++ ) {
    MPI_Recv (int_pic_3, N, MPI_INT, 2, tag, MPI_COMM_WORLD, &status);
    trans4 (int_pic_3, out_pic);
    /* do something with out_pic */
  }
} 


int main(int argc, char *argv[])
{
  int myrank; 
  
  MPI_Init (&argc, &argv);
  MPI_Comm_rank(MPI_COMM_WORLD, &myrank);

  switch (myrank) {
  case 0: stage1(); break;
  case 1: stage2(); break;
  case 2: stage3(); break;
  case 3: stage4(); break;
  }

  printf ("Stage %d: done\n", myrank+1);
  MPI_Finalize();
  return 0;

}