#include "ep_lib.hpp"
#include <mpi.h>
#include "ep_declaration.hpp"
#include "ep_mpi.hpp"

using namespace std;


namespace ep_lib 
{

  int MPI_Intercomm_merge_unique_leader(MPI_Comm inter_comm, bool high, MPI_Comm *newintracomm)
  {
    Debug("intercomm_merge with unique leader\n");



    int ep_rank, ep_rank_loc, mpi_rank;
    int ep_size, num_ep, mpi_size;

    ep_rank = inter_comm->ep_comm_ptr->size_rank_info[0].first;
    ep_rank_loc = inter_comm->ep_comm_ptr->size_rank_info[1].first;
    mpi_rank = inter_comm->ep_comm_ptr->size_rank_info[2].first;
    ep_size = inter_comm->ep_comm_ptr->size_rank_info[0].second;
    num_ep = inter_comm->ep_comm_ptr->size_rank_info[1].second;
    mpi_size = inter_comm->ep_comm_ptr->size_rank_info[2].second;

    int local_high = high;
    int remote_high;

    int remote_ep_size = inter_comm->ep_comm_ptr->intercomm->remote_rank_map->size();

    int local_ep_rank, local_ep_rank_loc, local_mpi_rank;
    int local_ep_size, local_num_ep, local_mpi_size;

    //local_ep_rank = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[0].first;
    //local_ep_rank_loc = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[1].first;
    //local_mpi_rank = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[2].first;
    //local_ep_size = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[0].second;
    //local_num_ep = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[1].second;
    //local_mpi_size = inter_comm->ep_comm_ptr->intercomm->local_comm->ep_comm_ptr->size_rank_info[2].second;


    if(local_ep_rank == 0)
    {
      MPI_Status status[2];
      MPI_Request request[2];
      MPI_Isend(&local_high, 1, MPI_INT, 0, inter_comm->ep_comm_ptr->intercomm->intercomm_tag, inter_comm, &request[0]);
      MPI_Irecv(&remote_high, 1, MPI_INT, 0, inter_comm->ep_comm_ptr->intercomm->intercomm_tag, inter_comm, &request[1]);

      MPI_Waitall(2, request, status);
    }


    //MPI_Bcast(&remote_high, 1, MPI_INT, 0, inter_comm->ep_comm_ptr->intercomm->local_comm);



    MPI_Comm_dup(inter_comm, newintracomm);

    int my_ep_rank = local_high<remote_high? local_ep_rank: local_ep_rank+remote_ep_size;


    int intra_ep_rank, intra_ep_rank_loc, intra_mpi_rank;
    int intra_ep_size, intra_num_ep, intra_mpi_size;

    intra_ep_rank = (*newintracomm)->ep_comm_ptr->size_rank_info[0].first;
    intra_ep_rank_loc = (*newintracomm)->ep_comm_ptr->size_rank_info[1].first;
    intra_mpi_rank = (*newintracomm)->ep_comm_ptr->size_rank_info[2].first;
    intra_ep_size = (*newintracomm)->ep_comm_ptr->size_rank_info[0].second;
    intra_num_ep = (*newintracomm)->ep_comm_ptr->size_rank_info[1].second;
    intra_mpi_size = (*newintracomm)->ep_comm_ptr->size_rank_info[2].second;


    MPI_Barrier_local(*newintracomm);


    int *reorder;
    if(intra_ep_rank_loc == 0)
    {
      reorder = new int[intra_ep_size];
    }


    MPI_Gather(&my_ep_rank, 1, MPI_INT, reorder, 1, MPI_INT, 0, *newintracomm);
    if(intra_ep_rank_loc == 0)
    {
      ::MPI_Bcast(reorder, intra_ep_size, to_mpi_type(MPI_INT), 0, to_mpi_comm((*newintracomm)->mpi_comm));

      vector< pair<int, int> > tmp_rank_map(intra_ep_size);


      for(int i=0; i<intra_ep_size; i++)
      {
        tmp_rank_map[reorder[i]] = (*newintracomm)->ep_rank_map->at(i) ;
      }

      //(*newintracomm)->rank_map->swap(tmp_rank_map);
      (*newintracomm)->ep_rank_map->clear();
      for(int i=0; i<tmp_rank_map.size(); i++)
      {
        (*newintracomm)->ep_rank_map->insert(std::pair< int, std::pair<int,int> >(i, tmp_rank_map[i].first, tmp_rank_map[i].second));
      }
      

      tmp_rank_map.clear();
    }

    MPI_Barrier_local(*newintracomm);

    (*newintracomm)->ep_comm_ptr->size_rank_info[0].first = my_ep_rank;

    if(intra_ep_rank_loc == 0)
    {
      delete[] reorder;
    }

    return MPI_SUCCESS;
  }





  int MPI_Intercomm_merge(MPI_Comm inter_comm, bool high, MPI_Comm *newintracomm)
  {
    

    assert(inter_comm->is_intercomm);

    // determine if only one MPI proc

        // to be completed ......

    // multiple MPI proc and high differs

    int newcomm_ep_rank = inter_comm->ep_comm_ptr->intercomm->size_rank_info[0].first;
    int newcomm_ep_rank_loc = inter_comm->ep_comm_ptr->intercomm->size_rank_info[1].first;
    int newcomm_num_ep = inter_comm->ep_comm_ptr->intercomm->size_rank_info[1].second;

    int ep_rank = inter_comm->ep_comm_ptr->size_rank_info[0].first;
    int ep_rank_loc = inter_comm->ep_comm_ptr->size_rank_info[1].first;
    int num_ep = inter_comm->ep_comm_ptr->size_rank_info[1].second;

    if(newcomm_ep_rank_loc == 0)
    {
      ::MPI_Comm *mpi_intracomm = new ::MPI_Comm;
      ::MPI_Intercomm_merge(to_mpi_comm(inter_comm->ep_comm_ptr->intercomm->mpi_inter_comm), high, mpi_intracomm);
      
      MPI_Info info;
      MPI_Comm *ep_comm;
      MPI_Comm_create_endpoints(mpi_intracomm, newcomm_num_ep, info, ep_comm);

      inter_comm->ep_comm_ptr->comm_list[0]->mem_bridge = ep_comm;
    }

    MPI_Barrier_local(inter_comm);

    int remote_num_ep = newcomm_num_ep - num_ep;

    *newintracomm = inter_comm->ep_comm_ptr->comm_list[0]->mem_bridge[high? remote_num_ep+ep_rank_loc : ep_rank_loc];

    int ep_size = inter_comm->ep_comm_ptr->size_rank_info[0].second;
    int remote_ep_size = inter_comm->ep_comm_ptr->intercomm->intercomm_rank_map->size();

    //printf("ep_size = %d, remote_ep_size = %d\n", ep_size, remote_ep_size);

    (*newintracomm)->ep_comm_ptr->size_rank_info[0].first = high? remote_ep_size+ep_rank : ep_rank;

    int my_triple[3];
    my_triple[0] = (*newintracomm)->ep_comm_ptr->size_rank_info[0].first;
    my_triple[1] = (*newintracomm)->ep_comm_ptr->size_rank_info[1].first;
    my_triple[2] = (*newintracomm)->ep_comm_ptr->size_rank_info[2].first;

    int *my_triple_list = new int[3 * (*newintracomm)->ep_comm_ptr->size_rank_info[0].second];


    MPI_Allgather(my_triple, 3, MPI_INT, my_triple_list, 3, MPI_INT, *newintracomm);

    if((*newintracomm)->ep_comm_ptr->size_rank_info[1].first == 0)
    {
      (*newintracomm)->ep_rank_map->clear();
      for(int i=0; i<(*newintracomm)->ep_comm_ptr->size_rank_info[0].second; i++)
      {
        (*newintracomm)->ep_rank_map->insert(std::pair< int, std::pair<int,int> >(my_triple_list[3*i], my_triple_list[3*i+1], my_triple_list[3*i+2]));
      }
    }

#ifdef _showinfo
    MPI_Barrier_local(inter_comm);
    if((*newintracomm)->ep_comm_ptr->size_rank_info[0].first == 15)
    {
      for(std::map<int, std::pair<int, int> >::iterator it = (*newintracomm)->ep_rank_map->begin(); it != (*newintracomm)->ep_rank_map->end(); it++)
      {
        printf("(%d  %d  %d)\n", it->first, it->second.first, it->second.second);
      }
    }
#endif
    
    delete my_triple_list;
  }
  
}