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

#include <cstdio>
#include <cstdlib>

#include <mpi.h>
#include <omp.h>

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

#define min(a,b)  (((a) < (b)) ? (a) : (b))

#define MAX_NUM_OF_NODES  (4)

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


static void mat_mul_omp(int rows) {
  omp_set_num_threads(num_threads);
#pragma omp parallel shared(A, B, C, rows, N, K, num_threads)
  {
    int tid = omp_get_thread_num();
    int is = rows / num_threads * tid + min(tid, rows % num_threads);
    int ie = rows / num_threads * (tid + 1) + min(tid + 1, rows % num_threads);

    for (int ii = is; ii < ie; ii += ITILESIZE) {
      for (int jj = 0; jj < N; jj += JTILESIZE) {
        for (int kk = 0; kk < K; kk += KTILESIZE) {
          
          for (int k = kk; k < min(kk + KTILESIZE, K); k++) {
            for (int i = ii; i < min(ii + ITILESIZE, ie); i++) {
              float ar = A[i * K + k];
              for (int j = jj; j < min(jj + JTILESIZE, N); j++) {
                C[i * N + j] += ar * B[k * N + j];
              }
            }
          }

        }
      }
    }

  }

  return;
}

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;

  int divided_rows[MAX_NUM_OF_NODES];
  int offset[MAX_NUM_OF_NODES] = {0};
  int divided_row, remainder, M_new;
  int tmp = 0;
  MPI_Status status;

  // 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.
  if (mpi_rank == 0)
  {
    divided_row = M / mpi_world_size;
    remainder = M - divided_row * mpi_world_size;
    
    // Larger numbered nodes compute more rows
    if(remainder != 0) {
      for (int i = 0; i < (mpi_world_size - remainder); i++) {
        divided_rows[i] = divided_row;
      }
      for (int i = (mpi_world_size - remainder); i < mpi_world_size; i++) {
        divided_rows[i] = divided_row + 1;
      }
    }
    else {
      for (int i = 0; i < mpi_world_size; i++) {
        divided_rows[i] = divided_row;
      }
    }
    for (int i = 1; i < mpi_world_size; i++) {
      tmp += divided_rows[i - 1];
      offset[i] = tmp;  // Starting row number divided by node
    }

    // Send data to other nodes (tag = 0)
    for(int i = 1; i < mpi_world_size; i++) {
      MPI_Send(&divided_rows[i], 1, MPI_INT, i, 0, MPI_COMM_WORLD);
      MPI_Send(&K, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
      MPI_Send(&N, 1, MPI_INT, i, 0, MPI_COMM_WORLD);
      
      MPI_Send(&A[offset[i] * K], divided_rows[i] * K, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
      MPI_Send(B, K * N, MPI_FLOAT, i, 0, MPI_COMM_WORLD);
    }

    // Calculate mat mul for root node part
    mat_mul_omp(divided_rows[0]);

    // Waiting until the each nodes sent their result (tag = 1)
    for(int i = 1; i < mpi_world_size; i++) {      
      MPI_Recv(&C[offset[i] * N], divided_rows[i] * N, MPI_FLOAT, i, 1, MPI_COMM_WORLD, &status);
    }
  }
  else
  {
    // Receive data from root node (tag = 0)
    MPI_Recv(&M_new, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(&K, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(&N, 1, MPI_INT, 0, 0, MPI_COMM_WORLD, &status);
    
    // allocate for matrix
    alloc_mat(&A, M_new, K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, M_new, N);
    
    // Receive divied A mat & B mat
    MPI_Recv(A, M_new * K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
    MPI_Recv(B, K * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);

    mat_mul_omp(M_new);

    // Send result to root node (tag = 1)
    MPI_Send(C, M_new * N, MPI_FLOAT, 0, 1, MPI_COMM_WORLD);
  }
    
}