#include "mat_mul.h"

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

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

// define
#define NRA	M	// number of rows in Matrix A
#define NCA	K	// number of cols in Matrix A
#define NCB	N	// number of cols in Matrix B
#define MASTER	0	// number of cols in Matrix B
#define FROM_MASTER 1	// number of cols in Matrix B
#define FROM_WORKER 2	// number of cols in Matrix B
#define TM	32
#define TK	16
#define TN	2048

static int rows[4] = {0,};		// rows of mat A set to each worker
static int offset[4] = {0,};		// rows of mat A set to each worker

static void mat_mul_omp()
{
  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+=TM) {
for(int kk=0; kk<K; kk+=TK) {
for(int jj=0; jj<N; jj+=TN) {
	int ek = (kk + TK < K) ? (kk + TK) : K;
	int em = (ii + TM < M) ? (ii + TM) : M;
	int en = (jj + TN < N) ? (jj + TN) : N;

	for(int i=ii; i<em; ++i) {
	for(int k=kk; k<ek; ++k) {
	for(int j=jj; j<en; ++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;

  // TODO: parallelize & optimize matrix multiplication on multi-node
  // You must allocate & initialize A, B, C for non-root processes
  // TODO: parallelize & optimize matrix multiplication
  //
  MPI_Request request;
  MPI_Status status;

  // number of worker tasks
  int numworkers = mpi_world_size;

  int mtype;		// message type

  int node_row = _M/numworkers;
  int extra_row = _M - (node_row * (numworkers-1));

  for (int i = 0; i < numworkers; i++){
    rows[i] = (i == (numworkers-1)) ? extra_row : node_row;
  }
  
  for (int i = 0; i < numworkers-1; i++){
    offset[i+1] = offset[i] + rows[i];
  } 

 /*
  A: M x K || NRA = M;
  B: K x N || NCA = K;
  C: M x N || NCB = N;

  MPI_Send(start_buf_addr_of_sender,
	   num_of_send_data,
	   data_type_of_send_data,
	   rank_of_receiver,
	   TAG,
	   Communicator_handler);

  MPI_Recv(start_buf_addr_of_receiver,
	   num_of_recv_data,
	   data_type_of_recv_data,
	   rank_of_sender,
	   TAG,
	   Communicator_handler,
	   MPI_status);
*/

  if(mpi_rank != MASTER) {
	M=rows[mpi_rank];
	alloc_mat(&A, rows[mpi_rank], K);
	alloc_mat(&B, K, N);
	alloc_mat(&C, rows[mpi_rank], N);
  }
  mtype = MASTER;
  
  MPI_Bcast(B, K*N, MPI_FLOAT, mtype, MPI_COMM_WORLD);
  // Master
  if(mpi_rank == MASTER) {
    for(int i=1; i<numworkers; i++) {
  	//rows = (dest == (numworkers-1)) ? extra_row : node_row;
  	//offset = node_row*dest;
  	//printf("offset=%d, rows=%d, dest=%d\n",offset, rows, dest);
  	//printf("node_row=%d, extra_row=%d\n",node_row, extra_row);
  	MPI_Isend(&A[offset[i]*K], rows[i]*K, MPI_FLOAT, i, mtype, MPI_COMM_WORLD, &request);
    
  	//MPI_Send(&offset, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD);
  	//MPI_Send(&rows, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD);
  	//MPI_Send(&A[offset*K], rows*K, MPI_FLOAT, dest, mtype, MPI_COMM_WORLD);
  	//MPI_Send(&B[0], K*N, MPI_FLOAT, dest, mtype, MPI_COMM_WORLD);
    }
  }
  else {
	MPI_Recv(A, rows[mpi_rank] * K, MPI_FLOAT, 0, mtype, MPI_COMM_WORLD, &status);
  }

  mat_mul_omp();
  if(mpi_rank != MASTER) {
  	MPI_Isend(C, rows[mpi_rank]*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
  }
  else {
    for(int i=1; i<numworkers; i++) {
	MPI_Recv(&C[offset[i] * N], rows[i] * N, MPI_FLOAT, i, mtype, MPI_COMM_WORLD, &status);
    }
  }

//for(dest=1; dest<numworkers; dest++) {
//      MPI_Recv(&offset, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD, &status);
//      MPI_Recv(&rows, 1, MPI_INT, dest, mtype, MPI_COMM_WORLD, &status);
//      MPI_Recv(&C[offset*N], rows*N, MPI_FLOAT, dest, mtype, MPI_COMM_WORLD, &status);
//}


//else {
//printf("RECEIVER\n");
//  alloc_mat(&A, M, K);
//  alloc_mat(&B, K, N);
//  alloc_mat(&C, M, N);
//  zero_mat(C, M, N);
//
//  mtype = FROM_MASTER;
//  MPI_Recv(&offset, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD, &status);
//  MPI_Recv(&rows, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD, &status);
//  MPI_Recv(A, rows*K, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD, &status);
//  MPI_Recv(B, K*N, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD, &status);
//
//  mat_mul_omp();
//
//  mtype = FROM_WORKER;
//  MPI_Send(&offset, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD);
//  MPI_Send(&rows, 1, MPI_INT, MASTER, mtype, MPI_COMM_WORLD);
//  MPI_Send(C, rows*N, MPI_FLOAT, MASTER, mtype, MPI_COMM_WORLD);
//}

 //// FIXME: for now, only root process runs the matrix multiplication.
 //if (mpi_rank == 0)
 // // mat_mul_omp();
 //  mat_mul_omp();
}