#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 MASTER (0)
#define FROM_MASTER (1)
#define FROM_WORKER (2)

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
#pragma omp parallel for
  for (int i = 0; i < M; ++i) {
    for (int j = 0; j < N; ++j) {
      for (int k = 0; k < K; ++k) {
        C[i * N + j] += A[i * K + k] * B[k * N + j];
      }
    }
  }
}
*/


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

static void mat_mul_omp(int row) {
 // int tid = (long)data;
  int is = 0;//row / num_threads * tid + min(tid, row % num_threads);
  int ie = row;// / num_threads * (tid + 1) + min(tid + 1, row % num_threads);

#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
//#pragma omp parallel for num_threads(num_threads)
    for (int ii = is; ii < ie; ii += TILEM) {
     for (int kk = 0; kk < K; kk += TILEK) {
      for (int jj = 0; jj < N; jj += TILEN) {
        int ek = kk + TILEK < K? (kk + TILEK) : K;
        int em = ii + TILEM < M? (ii + TILEM) : M;
        int en = jj + TILEN < N? (jj + TILEN) : N;
 
        for (int i = ii; i < em; ++i) {
          for (int k = kk; k < ek; ++k) {
            float Mux_val = A[i * K + k];
            for (int j = jj; j < en; ++j) {
                C[i * N + j] += Mux_val * B[k * N + j];
              }
            }
          }
        }

      }
    }
  
// return NULL;
}


/*
#define ITILESIZE (32)
#define JTILESIZE (2048)
#define KTILESIZE (32)

static void mat_mul_omp(int rows) {

  omp_set_num_threads(num_threads);
  #pragma omp parallel for
  for (int ii = 0; ii < rows; ii += ITILESIZE) {
    for (int jj = 0; jj < N; jj += JTILESIZE) {
      for (int kk = 0; kk < K; kk += KTILESIZE) {

        for (int k = kk; k < min(K, kk + KTILESIZE); k++) {
          for (int i = ii; i < min(rows, ii + ITILESIZE); i++) {
            float ar = A[i * K + k];
            for (int j = jj; j < min(N, jj + JTILESIZE); j+=1) {
              C[i * N + j] += ar * B[k * N + j];
            }
          }
        }

      }
    }
  }

//  return NULL;
}
*/

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 row_size, rows;
int start, end, offset;

MPI_Status status;
MPI_Request request;

 // 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) 
  {
    row_size = M / mpi_world_size;

    for (int node=1; node<mpi_world_size; node++) 
    {
  	  start = node * row_size;
      offset = node * row_size;
  	  end = (node == mpi_world_size-1) ? M : (node+1)*row_size;
  	  rows = end - start;
  
  	  MPI_Isend(&offset, 1, MPI_INT, node, FROM_MASTER, MPI_COMM_WORLD, &request);
  	  MPI_Isend(&rows, 1, MPI_INT, node, FROM_MASTER, MPI_COMM_WORLD, &request);
  	  MPI_Isend(&A[offset*K], rows*K, MPI_FLOAT, node, FROM_MASTER, MPI_COMM_WORLD, &request);
  	  MPI_Isend(B, K*N, MPI_FLOAT, node, FROM_MASTER, MPI_COMM_WORLD, &request);

	  }

         /*
    for(int i=0; i<rows; i++)
      for(int j=0; j<NCA; j++)
      {
        //C[i * NCB + k] = 0;
        Mat = A[i * NCA + j];
        for(int k=0; k<NCB; k++)
          C[i * NCB + k] += Mat * B[j * NCB + k]; 
      }
      */
      
	  rows =  row_size;
	  mat_mul_omp(rows);

	  for (int node=1; node<mpi_world_size; node++) 
    {
	    MPI_Recv(&offset, 1, MPI_INT, node, FROM_WORKER, MPI_COMM_WORLD, &status);
	    MPI_Recv(&rows, 1, MPI_INT, node, FROM_WORKER, MPI_COMM_WORLD, &status);
	    MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, node, FROM_WORKER, MPI_COMM_WORLD, &status);
    }

  }
  else 
  {
    alloc_mat(&A, M, K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, M, N);

  	MPI_Recv(&offset, 1, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
  	MPI_Recv(&rows, 1, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
  	MPI_Recv(A, rows*K, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
  	MPI_Recv(B, K*N, MPI_INT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);

    /*
    for(int i=0; i<rows; i++)
      for(int j=0; j<NCA; j++)
      {
        //C[i * NCB + k] = 0;
        Mat = A[i * NCA + j];
        for(int k=0; k<NCB; k++)
          C[i * NCB + k] += Mat * B[j * NCB + k]; 
      }
      */
  	mat_mul_omp(rows);

  	MPI_Isend(&offset, 1, MPI_INT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
  	MPI_Isend(&rows, 1, MPI_INT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
  	MPI_Isend(C, rows*N, MPI_FLOAT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
  }
}