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

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

#define HOST 0

#define ITILESIZE 16
#define JTILESIZE 2048	
#define KTILESIZE 2048

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

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

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node

  int is[num_threads];
  int ie[num_threads];

  for (int i=0; i < num_threads; i++){
    is[i] = rows / num_threads * i + min(i, rows % num_threads);
    ie[i] = rows / num_threads * (i + 1) + min(i + 1, rows % num_threads);
  }

  int tid;
  int ii, jj, kk, i, j, k;

  omp_set_num_threads(num_threads);

#pragma omp parallel shared(A, B, C, M, N, K, num_threads) private(ii, jj, kk, k, i, j, tid)
{
  tid = omp_get_thread_num();

  for (ii = is[tid]; ii < ie[tid]; ii += ITILESIZE) {
    for (jj = 0; jj < N; jj += JTILESIZE) {
      for (kk = 0; kk < K; kk += KTILESIZE) {

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

  MPI_Request request;
  MPI_Status status;

  int tag1 = 1001;
  int tag2 = 1002;

  // 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 slice_node = M / mpi_world_size;
  int start[mpi_world_size] = {0};
  int end[mpi_world_size] = {0};

  for (int i = 0; i < mpi_world_size; i++) {	// i is the node pointer
    start[i] = i * slice_node;	
    end[i] = (i == mpi_world_size - 1) ? M : (i + 1) * slice_node;
  } 

  if (mpi_rank == 0) {
    for (int node = 1; node < mpi_world_size; node++) {
	rows = end[node] - start[node];
	start_node = start[node];

	MPI_Isend(&start_node, 1, MPI_INT, node, tag1, MPI_COMM_WORLD, &request);
	MPI_Isend(&rows, 1, MPI_INT, node, tag1, MPI_COMM_WORLD, &request);
	MPI_Isend(&A[start_node*K], rows*K, MPI_FLOAT, node, tag1, MPI_COMM_WORLD, &request);
	MPI_Isend(B, K*N, MPI_FLOAT, node, tag1, MPI_COMM_WORLD, &request);
    }

    rows = slice_node;
    mat_mul_omp();

    for (int node = 1; node < mpi_world_size; node++) {
	MPI_Recv(&start_node, 1, MPI_INT, node, tag2, MPI_COMM_WORLD, &status);
	MPI_Recv(&rows, 1, MPI_INT, node, tag2, MPI_COMM_WORLD, &status);
	MPI_Recv(&C[start_node*N], rows*N, MPI_INT, node, tag2, MPI_COMM_WORLD, &status);
    }
  } 
  else {
    alloc_mat(&A, slice_node + mpi_world_size - 1, K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, slice_node + mpi_world_size - 1, N);
    zero_mat(C, slice_node + mpi_world_size - 1, N);

    MPI_Recv(&start_node, 1, MPI_INT, HOST, tag1, MPI_COMM_WORLD, &status);
    MPI_Recv(&rows, 1, MPI_INT, HOST, tag1, MPI_COMM_WORLD, &status);
    MPI_Recv(A, rows*K, MPI_FLOAT, HOST, tag1, MPI_COMM_WORLD, &status);
    MPI_Recv(B, K*N, MPI_FLOAT, HOST, tag1, MPI_COMM_WORLD, &status);

    mat_mul_omp();

    MPI_Isend(&start_node, 1, MPI_INT, HOST, tag2, MPI_COMM_WORLD, &request);
    MPI_Isend(&rows, 1, MPI_INT, HOST, tag2, MPI_COMM_WORLD, &request);
    MPI_Isend(C, rows*N, MPI_FLOAT, HOST, tag2, MPI_COMM_WORLD, &request);
  }

}