/*!
   \file ep_create.cpp
   \since 2 may 2016

   \brief Definitions of MPI endpoint function: MPI_Comm_create_endpoints
 */

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

using namespace std;

namespace ep_lib {

  /*!
    Dynamic creation of endpoints for each MPI process.
    The output of this function is an array of communicator handles of length num_ep, where each handle
    corresponds to a nez local randk in the output communicator.
    Once created, endpoints behave as MPI processes.
    \param [in] mpi_comm Parent MPI communicator.
    \param [in] num_ep Number of endpoints per process.
    \param [out] info Information of the EP creation.
    \param [out] out_comm_hdls Handles of EP communicators.
  */
  
  int MPI_Comm_create_endpoints(void* base_comm_ptr, int num_ep, MPI_Info info, MPI_Comm *& out_comm_hdls)
  {

    int base_rank;
    int base_size;
    
    ::MPI_Comm mpi_base_comm = to_mpi_comm(base_comm_ptr);

    ::MPI_Comm_size(mpi_base_comm, &base_size);  // ep_lib::mpi_comm_size
    ::MPI_Comm_rank(mpi_base_comm, &base_rank);  // ep_lib::mpi_comm_rank
                    // parent_comm can also be endpoints communicators

    std::vector<int> recv_num_ep(base_size);

    out_comm_hdls = new MPI_Comm[num_ep];

    for (int idx = 0; idx < num_ep; ++idx)
    {
      out_comm_hdls[idx] = new ep_comm;
      out_comm_hdls[idx]->is_ep = true;
      out_comm_hdls[idx]->is_intercomm = false;
      out_comm_hdls[idx]->ep_comm_ptr = new ep_communicator;     
      *(static_cast< ::MPI_Comm*>(out_comm_hdls[idx]->mpi_comm)) = *(static_cast< ::MPI_Comm*>(base_comm_ptr));
      out_comm_hdls[idx]->ep_comm_ptr->comm_list = out_comm_hdls;
      out_comm_hdls[idx]->ep_comm_ptr->comm_label = 0;
    }

    ::MPI_Allgather(&num_ep, 1, to_mpi_type(MPI_INT), &recv_num_ep[0], 1, to_mpi_type(MPI_INT), mpi_base_comm);


    int sum = 0;  // representing total ep number of process with smaller rank
    for (int i = 0; i < base_rank; ++i) {sum += recv_num_ep[i]; }

    int ep_size = std::accumulate(recv_num_ep.begin(), recv_num_ep.end(), 0);

    out_comm_hdls[0]->ep_barrier = new ep_barrier(num_ep);
    out_comm_hdls[0]->my_buffer = new BUFFER;

    out_comm_hdls[0]->ep_rank_map = new EP_RANK_MAP;


    for (int i = 1; i < num_ep; i++)
    {
      out_comm_hdls[i]->ep_barrier = out_comm_hdls[0]->ep_barrier;
      out_comm_hdls[i]->my_buffer  = out_comm_hdls[0]->my_buffer;
      out_comm_hdls[i]->ep_rank_map= out_comm_hdls[0]->ep_rank_map;
    }


    for (int i = 0; i < num_ep; i++)
    {
      out_comm_hdls[i]->ep_comm_ptr->size_rank_info[0] = std::make_pair(sum+i, ep_size);
      out_comm_hdls[i]->ep_comm_ptr->size_rank_info[1] = std::make_pair(i, num_ep);
      out_comm_hdls[i]->ep_comm_ptr->size_rank_info[2] = std::make_pair(base_rank, base_size);

      out_comm_hdls[i]->ep_comm_ptr->message_queue = new Message_list;
    }


    int ind = 0;

    for(int i=0; i<base_size; i++)
    {
      for(int j=0; j<recv_num_ep[i]; j++)
      {
        //out_comm_hdls[0]->rank_map->at(ind) = make_pair(j, i);
        out_comm_hdls[0]->ep_rank_map->insert(std::pair< int, std::pair<int,int> >(ind, j, i));
        ind++;
      }
    }

    return 0;

  } //MPI_Comm_create_endpoints


} //namespace ep_lib