#include "mat_mul.h"
#include "util.h"

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

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

#define ISIZE (24) 
#define JSIZE (1024)
#define KSIZE (1024) 
#define min(a,b) ((a)>(b) ? (b) : (a))

static void mat_mul_omp() {
  float ar;
  #pragma omp parallel for num_threads(num_threads) 
  for (int ii = 0; ii < M; ii += ISIZE) {
    for (int jj = 0; jj < N; jj += JSIZE) {
      for (int kk = 0; kk < K; kk += KSIZE) {

        for (int k = kk; k < min(kk+KSIZE,K); k++) {
          for (int i = ii; i < min(ii+ISIZE,M); i++) {
            ar = A[i * K + k];
            for (int j = jj; j < min(jj+JSIZE,N); j++) {
              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.
  MPI_Status status;

  int SrtP[mpi_world_size] = {0,};
  int EndP[mpi_world_size] = {0,};
  int DivS[mpi_world_size] = {0,};
  for(int i = 0; i < mpi_world_size; ++i) {
     if(i != 0) SrtP[i] = M / mpi_world_size * i + min(i, M % mpi_world_size);
     EndP[i] = M / mpi_world_size * (i + 1) + min(i + 1, M % mpi_world_size);
     DivS[i] = EndP[i] - SrtP[i]; 
  }

  if (mpi_rank == 0) 
  {
    for(int i = 1; i < mpi_world_size; ++i) {
      MPI_Send(&A[SrtP[i]*K], DivS[i]*K, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
      MPI_Send(B, (K*N), MPI_FLOAT, i, 0, MPI_COMM_WORLD);
    }

    M = DivS[mpi_rank];
    mat_mul_omp();

    for(int i = 1; i < mpi_world_size; ++i) {
      MPI_Recv(&C[SrtP[i]*N], DivS[i]*N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &status);
    }
  }

  if (mpi_rank != 0) 
  {
    alloc_mat(&A, DivS[mpi_rank], K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, DivS[mpi_rank], N);
    zero_mat(C, DivS[mpi_rank], N);

    MPI_Recv(A, DivS[mpi_rank]*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(B, (K*N), MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);

    M=DivS[mpi_rank];
    mat_mul_omp();

    MPI_Send(C, DivS[mpi_rank]*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD);
  }

}