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

static float *A, *B, *C;
static int M, N, K;
// M: number of rows in matrix A
// K: number of columns in matrix A
// N: number of columns in matrix B

static int num_threads;
static int mpi_rank;	   // specific number of a process to be used for MPI
static int mpi_world_size; // number of process to be used for MPI
static int rows[4] = {0,}; // number of rows to be allocated for each process
static int offset[4] = {0,};

#define MASTER 0       // mpi_rank of first task
#define FROM_MASTER 1  // setting a message type
#define FROM_WORKER 2  // setting a message type

using namespace std;

#define ITILESIZE (32)
#define JTILESIZE (512)
#define KTILESIZE (32)

//#define BLOCKSIZE 64

static void mat_mul_omp() {   
	
  // TODO: parallelize & optimize matrix multiplication   
  int end = rows[mpi_rank];
  
#if 0
  for (int kk = 0; kk < K; kk += block_s) {
#pragma omp parallel for num_threads(num_threads) schedule(dynamic) \
default(none) private(A_buf) shared(A, B, C, M, end, kk, N, K, block_s)  
    for (int i = 0; i < end; ++i) {
      for (int k = kk; k < min(kk + block_s, K); ++k) {
        A_buf = A[i * K + k];		 
		
        // loop unrolling
        int N_8 = (N >> 3) <<3;
        int j = 0;
        	
        for (; j < N_8; j += 8) {
          C[i * N + j] += A_buf * B[k * N + j];
          C[i * N + j + 1] += A_buf * B[k * N + j + 1];			
          C[i * N + j + 2] += A_buf * B[k * N + j + 2];
          C[i * N + j + 3] += A_buf * B[k * N + j + 3];
          C[i * N + j + 4] += A_buf * B[k * N + j + 4];
          C[i * N + j + 5] += A_buf * B[k * N + j + 5];
          C[i * N + j + 6] += A_buf * B[k * N + j + 6];
          C[i * N + j + 7] += A_buf * B[k * N + j + 7];		  
        }
		
        // unrolling remainder
        for (; j < N; ++j) {
          C[i * N + j] += A_buf * B[k * N + j];	
        }
      }
    }
  } 
#endif

#if 1
#pragma omp parallel for num_threads(num_threads) schedule(dynamic)
  for (int ii = 0; ii < end; ii += ITILESIZE) {
    for (int kk = 0; kk < K; kk += KTILESIZE) {
      for (int jj = 0; jj < N; jj += JTILESIZE) {
        int end_k = kk + KTILESIZE < K ? (kk + KTILESIZE) : K;
	int end_m = ii + ITILESIZE < end ? (ii + ITILESIZE) : end;
	int end_n = jj + JTILESIZE < N ? (jj + JTILESIZE) : N;

        for (int i = ii; i < end_m; ++i) {
          for (int k = kk; k < end_k; ++k) {
            for (int j = jj; j < end_n; ++j) {
              C[i * N + j] += A[i * K + k] * B[k * N + j];
            }
          }
        }

      }
    }
  }
}
#endif

#if 0
  int end_k, end_m, end_n;
#pragma omp parallel for num_threads(num_threads) \
default(none) shared(A, B, C, M, end_m, end_k, end_n, end, K, N) 
  for (int ii = 0; ii < end; ii += ITILESIZE) {
    for (int kk = 0; kk < K; kk += KTILESIZE) {
      for (int jj = 0; jj < N; jj += JTILESIZE) {
        end_k = kk + KTILESIZE < K ? (kk + KTILESIZE) : K;
	end_m = ii + ITILESIZE < end ? (ii + ITILESIZE) : end;
	end_n = jj + JTILESIZE < N ? (jj + JTILESIZE) : N;

        for (int i = ii; i < end_m; ++i) {
          for (int k = kk; k < end_k; ++k) {
            for (int j = jj; j < end_n; ++j) {
              C[i * N + j] += A[i * K + k] * B[k * N + j];
            }
          }
        }
      }
    }
  }  
}

#endif
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) {

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

// indexing in common
  int row_size = M / mpi_world_size;
  offset[0] = 0;
  rows[0] = row_size;
  
  for (int i = 1; i < mpi_world_size; i++){
    // number of rows to be allocated for each process
    offset[i] = i * row_size;     
    rows[i] = (i == (mpi_world_size - 1)) ? (M - offset[i]) : row_size;
  }    
  
/**************************** Transmitting time (in Master side) ************************************/
// timely parallel sequence in both Master and Worker
  
  //MPI_Bcast(B, K*N, MPI_FLOAT, 0, MPI_COMM_WORLD);   
  if (mpi_rank == MASTER) {	
    /* Send matrix data to the workers */	
    // dest = destination address (relative address) to worker
    for (int dest = 1; dest < mpi_world_size; dest++) {
      MPI_Isend(&A[offset[dest]*K], rows[dest]*K, MPI_FLOAT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
      MPI_Isend(B, K*N, MPI_FLOAT, dest, FROM_MASTER, MPI_COMM_WORLD, &request);
    }
 
  }
  else {
    // array A, B, C -> memory allocation for worker nodes
    // array C -> initialization for add & assignment calculation	
    alloc_mat(&A, rows[mpi_rank], K);
    alloc_mat(&B, K, N);
    alloc_mat(&C, rows[mpi_rank], N);
    //zero_mat(C, M, N);
	
    MPI_Recv(A, rows[mpi_rank]*K, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
    MPI_Recv(B, K*N, MPI_FLOAT, MASTER, FROM_MASTER, MPI_COMM_WORLD, &status);
  }
	
  mat_mul_omp();
  
/**************************** Receiving time (in Master side) ************************************/    
// timely parallel sequence in both Master and Worker

  /* Receive MAC results from workers */
  // src = source address (relative address) to master
  if (mpi_rank == MASTER) {
    for (int src = 1; src < mpi_world_size; src++) {   
      MPI_Recv(&C[offset[src]*N], rows[src]*N, MPI_FLOAT, src, FROM_WORKER, MPI_COMM_WORLD, &status);
    }
  }
  else {
    MPI_Isend(C, rows[mpi_rank]*N, MPI_FLOAT, MASTER, FROM_WORKER, MPI_COMM_WORLD, &request);
  }
}