#include "mat_mul.h"

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

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;
}

MPI_Request request;
MPI_Status  status;

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

//static void mat_mul_omp(int rows) {
static void mat_mul_omp() {
  //
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
 # pragma omp parallel for num_threads(num_threads)
  for (int ii = 0; ii < 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(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];
            }
          }
        }

      }
    }
  }
}

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;

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

  // FIXME: for now, only root process runs the matrix multiplication.

  int ave_row    = M / mpi_world_size;
  int extra_row  = M % mpi_world_size;
  int offset;

  /* send matrix data to the worker tasks */
  if (mpi_rank == 0) {

    offset = ave_row + extra_row;
    M 	   = offset;

    if(ave_row) { 
        for (int node=1; node < mpi_world_size; node++)
        {
          MPI_Isend ( &offset, 	        1, 	    MPI_INT, 	node, 1, MPI_COMM_WORLD, &request);
          MPI_Isend ( &A[offset * K],   ave_row*K,  MPI_FLOAT,	node, 1, MPI_COMM_WORLD, &request);
          MPI_Isend ( B, 		K*N, 	    MPI_FLOAT, 	node, 1, MPI_COMM_WORLD, &request);

          offset = offset + ave_row;
        }
    }

    mat_mul_omp();

  } else { 

  /* Return the result to mater node */
    if(ave_row) { 

        alloc_mat ( &A, ave_row, K);
        alloc_mat ( &B, K,       N);
        alloc_mat ( &C, ave_row, N);
        zero_mat  (  C, ave_row, N);

        MPI_Recv ( &offset, 1, 	        MPI_INT,   0, 1, MPI_COMM_WORLD, &status);
        MPI_Recv ( A, 	    ave_row*K,  MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
        MPI_Recv ( B, 	    K*N,        MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);

    	M = ave_row;
        mat_mul_omp();

        MPI_Isend ( &offset, 1, 	MPI_INT,   0, 2, MPI_COMM_WORLD, &request);
        MPI_Isend ( C, 	    ave_row*N,  MPI_FLOAT, 0, 2, MPI_COMM_WORLD, &request);
    }
  }

  /* wait for results from all worker tasks */
  if (mpi_rank == 0) {
    if(ave_row) { 
        for (int i=1; i < mpi_world_size; i++)
        {
          MPI_Recv ( &offset, 	1, 	       MPI_INT,   i, 2, MPI_COMM_WORLD, &status);
          MPI_Recv ( &C[offset * N],ave_row*N, MPI_FLOAT, i, 2, MPI_COMM_WORLD, &status);
        }
    }
  }
}