#include "mat_mul.h"

#include <cstdio>
#include <cstdlib>
#include <mpi.h>
#include <omp.h>
#include "util.h"

#define MASTER_0       0
#define FROM_MASTER_1  1
#define FROM_WORKER_2  2

#define ITILESIZE (32)
#define JTILESIZE (1024)
#define KTILESIZE (1024)

static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;

static int min(int x, int y) {
  return x < y ? x : y;
}


static void mat_mul_omp(float *A, float *C, int rows) {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
 
  int tid, start_T_M, end_T_M, p_size ;
  //float A_temp ;
//#pragma omp parallel for
//#pragma omp parallel for num_threads(num_threads) private(tid, start_T_M, end_T_M, p_size, A_temp)  shared(A, B, C, M, N, K)
#pragma omp parallel num_threads(num_threads) private(tid, start_T_M, end_T_M, p_size)  shared(A, B, C, M, N, K)
  {
  tid = omp_get_thread_num() ;

  //if( (M%num_threads) != 0 ) {
    //p_size = (M/num_threads) ;
  if( (rows%num_threads) != 0 ) {
    p_size = (rows/num_threads)  ;
    if(tid!=(num_threads-1)) {
      start_T_M = tid * p_size;
      end_T_M   = start_T_M + p_size ;
    }
    else {
      start_T_M = tid * p_size;
      end_T_M = rows ;
    }
  }
  else {
    p_size = (rows/num_threads) ;
    start_T_M = tid * p_size;
    end_T_M   = start_T_M + p_size ;
  }

  for (int ii = start_T_M; ii < end_T_M; ii += ITILESIZE) {
    for (int jj = 0; jj < N; jj += JTILESIZE) {
      for (int kk = 0; kk < K; kk += KTILESIZE) {

        for (int k = kk; k < min(K, kk + KTILESIZE); k++) {
          for (int i = ii; i < min(end_T_M, ii + ITILESIZE); i++) {
            float ar = A[i * K + k];
            for (int j = jj; j < min(N, jj + JTILESIZE); j+=1) {
              C[i * N + j] += ar * B[k * N + j];
            }
          }
        }

      }
    }
  }

  }//pragma

}

void mat_mul(float *_A, float *_B, float *_C, int _M, int _N, int _K,
             int _num_threads, int _mpi_rank, int _mpi_world_size) {
  A = _A, B = _B, C = _C;
  M = _M, N = _N, K = _K;
  num_threads = _num_threads, mpi_rank = _mpi_rank,
  mpi_world_size = _mpi_world_size;

  MPI_Status  status;
  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes

  MPI_Barrier(MPI_COMM_WORLD);
  int row_size, rows ;
  int start_M, end_M ;


  if(mpi_rank==0) {



    for(int node = 1; node < mpi_world_size; node++) {
      row_size = (M/mpi_world_size) ;
      rows     = row_size ;
      start_M  = (node-1) * row_size;

      MPI_Send(&A[start_M*K], rows*K, MPI_INT, node, FROM_MASTER_1, MPI_COMM_WORLD) ;
      MPI_Send(B, K*N,                MPI_INT, node, FROM_MASTER_1, MPI_COMM_WORLD) ;
    }

    row_size = (M/mpi_world_size) ;
    start_M = (mpi_world_size-1) * row_size;
    end_M = M ;
    rows = end_M - start_M ;

    mat_mul_omp(&A[start_M*K], &C[start_M*N], rows);

    for(int node = 1; node < mpi_world_size; node++) {
      row_size = (M/mpi_world_size) ;
      rows     = row_size ;
      start_M  = (node-1) * row_size;
      MPI_Recv(&C[start_M*N], rows*N, MPI_INT, node, FROM_WORKER_2, MPI_COMM_WORLD, &status) ;
    }
  }
  else {
    alloc_mat(&A, M, K) ;
    alloc_mat(&B, K, N) ;
    alloc_mat(&C, M, N) ;
    //zero_mat(C, M, N) ;

    row_size = (M/mpi_world_size) ;
    rows = row_size ;

    MPI_Recv(A, rows*K, MPI_INT, MASTER_0, FROM_MASTER_1, MPI_COMM_WORLD, &status) ;
    MPI_Recv(B, K*N,    MPI_INT, MASTER_0, FROM_MASTER_1, MPI_COMM_WORLD, &status) ;

    mat_mul_omp (A, C, rows) ;

    MPI_Send(C, rows*N, MPI_INT, MASTER_0, FROM_WORKER_2, MPI_COMM_WORLD) ;

  }
  MPI_Barrier(MPI_COMM_WORLD);

//  // FIXME: for now, only root process runs the matrix multiplication.
//  if (mpi_rank == 0)
//    mat_mul_omp();


}