#include "mat_mul.h"

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

static float *A, *B, *C;
static int M, N, K;
static int num_threads;
static int mpi_rank, mpi_world_size;
static int rows[4] = {0,};
static int offset[4] = {0,};

#define TILEM 32
#define TILEK 16
#define TILEN 2048

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node
  int start = 0;
  int end   = rows[mpi_rank];
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
  for (int ii = start; ii < end; ii+=TILEM) {
    for (int kk = 0; kk < K; kk+=TILEK) {
      for(int jj = 0; jj < N; jj+=TILEN) {
	int mk = std::min(kk + TILEK, K);
	int mm = std::min(ii + TILEM, M);
	int mn = std::min(jj + TILEN, N);
	for(int i = ii; i < mm ; ++i) {
          for(int k = kk; k < mk; ++k) {
	    for(int j = jj; j < mn; ++j) {
	      C[i * N + j] += A[i * K + k] * 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;
  MPI_Request request;
  MPI_Status  status;

  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  int num_row = M / mpi_world_size;
  
  for (int i = 0; i < mpi_world_size; i++) {
    //rows[i] = (i == mpi_world_size - 1) ? (M - (num_row * (mpi_world_size - 1))) : num_row;
    if (i == mpi_world_size - 1)
      rows[i] = (M - (num_row * (mpi_world_size - 1)));
    else
      rows[i] = num_row;
  }
  for (int i = 0; i < mpi_world_size - 1; i++) {
    offset[i + 1] = offset[i] + rows[i];
  }
  
  if (mpi_rank != 0) {
    M = rows[mpi_rank];
    alloc_mat(&A, rows[mpi_rank], K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, rows[mpi_rank], N);
  }

  MPI_Bcast(B, K * N, MPI_FLOAT, 0, MPI_COMM_WORLD);

  if (mpi_rank == 0) {
    for (int i = 1; i < mpi_world_size; i++)
      MPI_Isend(&A[offset[i] * K], rows[i] * K, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
  } 
  else {
      MPI_Recv(A, rows[mpi_rank] * K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
  }
  
  mat_mul_omp();

  if (mpi_rank != 0) {
    MPI_Isend(C, rows[mpi_rank] * N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
  } 
  else {
    for (int i = 1; i < mpi_world_size; i++)
      MPI_Recv(&C[offset[i] * N], rows[i] * N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &status);
  }
  // FIXME: for now, only root process runs the matrix multiplication.
  //if (mpi_rank == 0)
  //  mat_mul_omp();
}