#include "mat_mul.h"

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

#include <getopt.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.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;

static int rows[4] = {0,};
static int offset[4] = {0,};

int omp_get_thread_num(void);
int omp_get_num_threads(void);

static void mat_mul_omp() {
  // TODO: parallelize & optimize matrix multiplication
  // Use num_threads per node

int start = 0;
int end = rows[mpi_rank];

int TILE_M = 40, TILE_K =16, TILE_N =4096;
int end_m, end_k, end_n;
float temp;

#pragma omp parallel for num_threads(num_threads) schedule(static)
 for(int ii=start; ii<end; ii+=TILE_M) 
	{
	for(int kk=0; kk<K; kk+= TILE_K) 
	{
		for(int jj=0; jj<N; jj+=TILE_N) 
		{
			end_m = ii+ TILE_M<M? (ii+TILE_M) : M;
			end_n = jj + TILE_N<N? (jj+TILE_N) : N;
			end_k = kk+ TILE_K<K ? (kk+TILE_K) : K;
			for(int i=ii; i<end_m; i++) 
			{
				for(int k=kk; k<end_k; k++) 
				{
					temp = A[i * K + k];	  
					for(int j=jj; j<end_n; j++) 
					{
						C[i*N+(j+0)] += temp*B[k*N+(j+0)];		
					}
				}
			}
		}
	}
 }
}

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

 MPI_Request request;
 MPI_Status status;

 int row_num = M/mpi_world_size;

// 노드별 행 개수 정의 
 for(int i=0; i<mpi_world_size; i++)
 {
	if(i==mpi_world_size - 1)
	{
		rows[i] = M - (row_num * (mpi_world_size-1));
	}
	else
	{
		rows[i] = row_num;
	}
 }

// 노드별 시작 위치 정의 for A matrix
 for(int i=0; i<mpi_world_size; i++)
 {
	offset[i+1] = offset[i] + rows[i];
 }

// 행렬 alloc @ task 노드
 if(mpi_rank != 0)
 {
	M = rows[mpi_rank];
	alloc_mat(&A, rows[mpi_rank], K);
	alloc_mat(&B, K, N);
	alloc_mat(&C, rows[mpi_rank], N);
	zero_mat(C, M, N);
 }

 // A행렬 전송/수신
 if(mpi_rank == 0)
 {
	for(int i=1; i<mpi_world_size; i++)
	{
		MPI_Isend(&A[offset[i]*K], rows[i]*K, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &request);
	}
 }
 else
 {
	MPI_Recv(A, rows[mpi_rank]*K, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &status);
 }

 // B행렬 전송/수신
 MPI_Bcast(B, K*N, MPI_FLOAT, 0, MPI_COMM_WORLD);

 // mat_mul 수행
 mat_mul_omp();

 // 계산결과 C행렬 전송/수신
 if(mpi_rank != 0)
 {
	MPI_Isend(C, rows[mpi_rank]*N, MPI_FLOAT, 0, 0, MPI_COMM_WORLD, &request);
 }
 else
 {
	for(int i=1; i<mpi_world_size; i++)
	{
		MPI_Recv(&C[offset[i]*N], rows[i]*N, MPI_FLOAT, i, 0, MPI_COMM_WORLD, &status);
	}
 }

}