#define _GNU_SOURCE

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <mpi.h>
#include <omp.h>

int main (int argc, char * argv[])
{
  int nomp = omp_get_max_threads ();
  int nmess = 1000;
  int lrank;
  int nrank;
  int buf_len = 100000000;
  char * buf;
  FILE * fp;
  char f[256];
  int thread_level;

//MPI_Init (&argc, &argv);
  MPI_Init_thread (&argc, &argv, MPI_THREAD_MULTIPLE, &thread_level);
  fprintf (stderr, " MPI_THREAD_SINGLE     = %d\n", MPI_THREAD_SINGLE);
  fprintf (stderr, " MPI_THREAD_FUNNELED   = %d\n", MPI_THREAD_FUNNELED);
  fprintf (stderr, " MPI_THREAD_SERIALIZED = %d\n", MPI_THREAD_SERIALIZED);
  fprintf (stderr, " MPI_THREAD_MULTIPLE   = %d\n", MPI_THREAD_MULTIPLE);
  fprintf (stderr, " thread_level          = %d\n", thread_level);

  MPI_Comm_rank (MPI_COMM_WORLD, &lrank);
  MPI_Comm_size (MPI_COMM_WORLD, &nrank);

  buf = (char *)malloc (buf_len);
  MPI_Buffer_attach (buf, buf_len);

  sprintf (f, "mpitest.%6.6d.txt", lrank);
  fp = fopen (f, "w");

#pragma omp parallel 
  {
    int bsend[2];
    int isend;
    int isend1, isend2;
    int irank;
    int iomp = omp_get_thread_num ();

    bsend[0] = lrank;
    bsend[1] = iomp;

    isend1 = (nmess * (iomp+0)) / nomp;
    isend2 = (nmess * (iomp+1)) / nomp;

    for (irank = 0; irank < nrank; irank++)
      if (irank != lrank)
        for (isend = isend1; isend < isend2; isend++)
          {
            MPI_Bsend (bsend, 2, MPI_INT, irank, 0, MPI_COMM_WORLD);
            fprintf (fp, "S%d %d %d\n", irank, iomp, lrank);
            fflush (fp);
          }
  }

#pragma omp parallel 
  {
    MPI_Status status;
    int brecv[2];
    int irecv;
    int irecv1;
    int irecv2;
    int irank;
    int iomp = omp_get_thread_num ();

    irecv1 = (nmess * (iomp+0)) / nomp;
    irecv2 = (nmess * (iomp+1)) / nomp;

    for (irank = 0; irank < nrank; irank++)
      if (irank != lrank)
        for (irecv = irecv1; irecv < irecv2; irecv++)
          {
            MPI_Recv (brecv, 2, MPI_INT, MPI_ANY_SOURCE, MPI_ANY_TAG, MPI_COMM_WORLD, &status);
            fprintf (fp, "R%d %d %d %d %d\n", status.MPI_SOURCE, iomp, lrank, brecv[0], brecv[1]);
            fflush (fp);
          }
  }

  fclose (fp);

  MPI_Finalize ();

  return 0;
}