// MPI demo program
// Christoph Federrath, 2016-2017

#include <iostream>
#include <stdlib.h>
#include "mpi.h" // MPI lib

using namespace std;

/** ----------------------------- Main -------------------------------
 **  MPI_TEST main procedure
 ** ------------------------------------------------------------------ */
int main(int argc, char *argv[])
{
    // initialize MPI and get number of ranks (NPE) and my rank number (MyPE)
    MPI_Init(&argc, &argv);
    int NPE = 0, MyPE = 0;
    MPI_Comm_size(MPI_COMM_WORLD, &NPE);
    MPI_Comm_rank(MPI_COMM_WORLD, &MyPE);

    if (MyPE == 0) cout << "------------------- mpi_test START -------------------" << endl;
    // print MPI info
    if (MyPE == 0) cout<<"["<<MyPE<<"] MPI total number of ranks: " << NPE << endl;
    MPI_Barrier(MPI_COMM_WORLD);


    // ======================= MPI ALL REDUCE DEMO =========================

    // define variables
    long n = (long)(5e9);
    double sum_global = 0.0;
    double sum_local = 0.0;

    // domain decomposition
    long my_chunk_start = MyPE * (n / NPE);
    long my_chunk_end   = (MyPE+1) * (n / NPE) - 1;
    // last PE gets the rest
    if (MyPE == NPE-1) my_chunk_end = n;
    //cout<<"["<<MyPE<<"] (start, end) = " << my_chunk_start << ", " << my_chunk_end << endl;
    MPI_Barrier(MPI_COMM_WORLD);

    // now loop over my local chunk
    for (long i = my_chunk_start; i <= my_chunk_end; i++)
    {
        sum_local += i;
    }
    cout<<"["<<MyPE<<"] sum_local = " << sum_local << endl;
    
    // do an MPI all reduce (with MPI_SUM) to sum sum_local into sum_global
    MPI_Allreduce(&sum_local, &sum_global, 1, MPI_DOUBLE, MPI_SUM, MPI_COMM_WORLD);
    
    // only the Master proc writes this to the screen
    if (MyPE == 0) cout << "sum_global = " << sum_global << endl;

    // clean up
    MPI_Barrier(MPI_COMM_WORLD);
    if (MyPE == 0) cout << "------------------- mpi_test END -----------------" << endl;
    MPI_Finalize();
    exit(0);



/*
    // ======================= MPI SEND/RECEIVE DEMO =========================

    // send/receive test
    // let proc 1 send a message to the master proc (0)
    int sender = 1;
    int receiver = 0;
    int message_tag = 1;
    MPI_Status status;

    // abort if we are not using exactly 2 MPI procs
    if (NPE != 2)
    {
        if (MyPE == 0) cout << "Error. This MPI test only works with exactly 2 MPI processes. Exiting..." << endl;
        if (MyPE == 0) cout << "------------------- mpi_test END -----------------" << endl;
        MPI_Finalize();
        exit(0);
    }
    // sender
    if (MyPE == 1)
    {
        int secret_message = 66;
        cout<<"["<<MyPE<<"] sending: " << secret_message << endl;
        MPI_Send(&secret_message, 1, MPI_INT, receiver, message_tag, MPI_COMM_WORLD);
    }
    // receiver
    if (MyPE == 0)
    {
        int received_message;
        MPI_Recv(&received_message, 1, MPI_INT, sender, message_tag, MPI_COMM_WORLD, &status);
        cout<<"["<<MyPE<<"] received: " << received_message << endl;
    }

    // clean up
    MPI_Barrier(MPI_COMM_WORLD);
    if (MyPE == 0) cout << "------------------- mpi_test END -----------------" << endl;
    MPI_Finalize();
    exit(0);
*/



} // end: main()