#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;

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

//#define DEBUG (true)
#define DEBUG (false)
#define ITILESIZE (32) // 32
#define JTILESIZE (1024) // 1024
#define KTILESIZE (1024) // 1024

static void mat_mul_omp(int is, int ie) {
  // TODO: parallelize & optimize matrix multiplication

  // Use num_threads per node
  omp_set_num_threads(num_threads);

  //for (int ii = 0; ii < M; ii += ITILESIZE) {
  #pragma omp parallel for schedule(runtime)
  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+=1) {
              C[i * N + j] += ar * B[k * N + j];
            } // for j
          } // for i
        } // for k

      } // for kk
    } // for jj
  } // for ii

  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;

  // split number for node
  num_cnt = (M / mpi_world_size) + ((M % mpi_world_size) > 0);
  int is, ie, in;
  MPI_Status status;

  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  if (mpi_rank == 0) {
    if (DEBUG) printf("[mat_mul %d] Start...\n", mpi_rank);

    // Send A,B to Others
    for (int r = 1; r < mpi_world_size; r++) {
      is = r * num_cnt;
      ie = min((r + 1) * num_cnt, M);
      in = ie - is;

      //MPI_Send(&A[is*K], in*K, MPI_FLOAT, r, 100, MPI_COMM_WORLD);
      MPI_Request reqA;
      MPI_Isend(&A[is*K], in*K, MPI_FLOAT, r, 100, MPI_COMM_WORLD, &reqA);
      if (DEBUG) printf("[mat_mul %d] Send A to %0d, %0d ~ %0d\n", mpi_rank, r, is, ie);

      //MPI_Send(B, K*N, MPI_FLOAT, r, 200, MPI_COMM_WORLD);
      MPI_Request reqB;
      MPI_Isend(B, K*N, MPI_FLOAT, r, 200, MPI_COMM_WORLD, &reqB);
      if (DEBUG) printf("[mat_mul %d] Send B to %0d\n", mpi_rank, r);
    }

    // Process
    mat_mul_omp(0, num_cnt); // for rank 0
    if (DEBUG) printf("[mat_mul %d] Process done (%0.2f)\n", mpi_rank, C[0]);

    // Receive C from Others
    for (int r = 1; r < mpi_world_size; r++) {
      is = r * num_cnt;
      ie = min((r + 1) * num_cnt, M);
      in = ie - is;

      MPI_Recv(&C[is*N], in*N, MPI_FLOAT, r, 300, MPI_COMM_WORLD, &status);
      if (DEBUG) printf("[mat_mul %d] Receive C : %0d, %0d ~ %0d\n", mpi_rank, r, is, ie);
    }
  }
  else {
    if (DEBUG) printf("[mat_mul %d] Start...\n", mpi_rank);

    is = mpi_rank * num_cnt;
    ie = min((mpi_rank + 1) * num_cnt, M);
    in = ie - is;

    // Memory allocation
    alloc_mat(&A, in, K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, in, N);
    zero_mat(C, in, N);

    // Recieve A from Rank0
    MPI_Recv(A, in*K, MPI_FLOAT, 0, 100, MPI_COMM_WORLD, &status);
    if (DEBUG) printf("[mat_mul %d] Receive A\n", mpi_rank);

    // Recieve B from Rank0
    MPI_Recv(B, K*N, MPI_FLOAT, 0, 200, MPI_COMM_WORLD, &status);
    if (DEBUG) printf("[mat_mul %d] Receive B\n", mpi_rank);

    // Process
    mat_mul_omp(0, in);
    if (DEBUG) printf("[mat_mul %d] Process done (%0.2f)\n", mpi_rank, C[0]);

    // Send C to Rank0
    MPI_Send(C, in*N, MPI_FLOAT, 0, 300, MPI_COMM_WORLD);
    if (DEBUG) printf("[mat_mul %d] Send C\n", mpi_rank);
  }
  
  // FIXME: for now, only root process runs the matrix multiplication.
  //if (mpi_rank == 0)
  //  mat_mul_omp();
}