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

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

#define II_BLOCK 16
#define KK_BLOCK 32
#define JJ_BLOCK 2048

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
		int node_slice = (M/mpi_world_size); 

   int tid,i,ii,j,jj,k,kk,start,end;
   int using_threads = num_threads;
   int slice = (M/mpi_world_size) / using_threads;
   omp_set_num_threads(using_threads);
   #pragma omp parallel private(tid,i,ii,kk,k,j,jj,start,end)
    {
        tid = omp_get_thread_num();
				start = mpi_rank * node_slice + slice * tid;
				end = (tid == using_threads - 1)?(mpi_rank == mpi_world_size - 1 ? M : (mpi_rank+1)*node_slice) : mpi_rank * node_slice+(tid+1)*slice;
				//printf("mpi_rank : %d, tid : %d, start : %d, end : %d\n", mpi_rank, tid, start, end);

				//printf("mpi_rank : %d, tid : %d, start : %d, end : %d\n", mpi_rank, tid, start, end);
						for (jj = 0; jj < N; jj += JJ_BLOCK) {
								for (kk = 0; kk < K; kk += KK_BLOCK) {

				for (ii = start; ii < end; ii += II_BLOCK) {
										for (i =  ii; i < (ii+II_BLOCK < end ? ii+II_BLOCK : end); ++i) {
												for(k = kk; k < (kk+KK_BLOCK < K ? kk+KK_BLOCK : K); ++k) {
														for (j = jj; j < (jj+JJ_BLOCK < N ? jj+JJ_BLOCK : N); ++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

  // FIXME: for now, only root process runs the matrix multiplication.
	//int offset = size_by_node * mpi_rank;
	MPI_Status status;
	
	if (mpi_rank == 0){
		//timer_start(5);
		for(int i = 1; i <= mpi_world_size-1; i++){
				MPI_Send(&A[i * K*(M/mpi_world_size)], (i==mpi_world_size-1)?K*M - i*K*(M/mpi_world_size):K*(M/mpi_world_size), MPI_FLOAT, i, 1, MPI_COMM_WORLD);
				MPI_Send(&B[0], K*N, MPI_FLOAT, i, 1, MPI_COMM_WORLD);
				MPI_Send(&C[i * N*(M/mpi_world_size)], (i==mpi_world_size-1)?N*M - i*N*(M/mpi_world_size):N*(M/mpi_world_size), MPI_FLOAT, i, 1, MPI_COMM_WORLD);

		}
		//double r0_time = timer_stop(5);
		//printf("r0_time before mat_mul : %f sec\n", r0_time);
		//timer_start(5);
		
		mat_mul_omp();
		//r0_time = timer_stop(5);
		//printf("0_time mat_mul : %f sec\n", r0_time);
		
		for(int i=1; i<= mpi_world_size-1; i++){
				MPI_Recv(&C[i * N*(M/mpi_world_size)], (i == mpi_world_size-1)?N*M - mpi_rank *N*(M/mpi_world_size):N*(M/mpi_world_size), MPI_FLOAT, i, 1, MPI_COMM_WORLD, &status);
		}
	}
	else{
    double * slave_rank = (double *)malloc((mpi_world_size) * sizeof(double));
		timer_start(mpi_rank);
		alloc_mat(&A, M, K);
		alloc_mat(&B, K, N);
		alloc_mat(&C, M, N);

		int A_arr_offset = mpi_rank*K*(M/mpi_world_size);
		int C_arr_offset = mpi_rank*N*(M/mpi_world_size);

		MPI_Recv(&A[A_arr_offset], (mpi_rank==mpi_world_size-1)?K*M - mpi_rank*K*(M/mpi_world_size):K*(M/mpi_world_size), MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(&B[0], K*N, MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);
		MPI_Recv(&C[C_arr_offset], (mpi_rank==mpi_world_size-1)?N*M - mpi_rank*N*(M/mpi_world_size):N*(M/mpi_world_size), MPI_FLOAT, 0, 1, MPI_COMM_WORLD, &status);

		slave_rank[mpi_rank]=timer_stop(mpi_rank);
		//printf("%d time before mat_mul : %f sec\n", mpi_rank, slave_rank[mpi_rank]);
		timer_start(mpi_rank);
		mat_mul_omp();
		slave_rank[mpi_rank]=timer_stop(mpi_rank);
		//printf("%d time mat_mul : %f sec\n", mpi_rank, slave_rank[mpi_rank]);

		MPI_Send(&C[C_arr_offset], (mpi_rank == mpi_world_size-1)?N*M - mpi_rank *N*(M/mpi_world_size):N*(M/mpi_world_size), MPI_FLOAT, 0, 1, MPI_COMM_WORLD);
	} 
}