#include<stdio.h>
#include<stdlib.h>
#include<pthread.h>

double** data;
double** result;

int dim, steps[] = {0, 0};
double coeff = 0.25;

double **input,**output,**temp;

pthread_t ts[2];
pthread_attr_t atr;
pthread_mutex_t sync_us;
pthread_cond_t move_on;

void * obrada(void *id){
	int tid = (int)id;
        int i, j;
        int f = 1;
	
	while(f)
	{
		for(i=1 + tid;i<dim+1;i += 2) //po parnim ili neparnim indeksima
			for(j=1;j<dim+1;j++)
			output[i][j] = 
				(1.0 - 4.0 * coeff)*input[i][j] + 
                        coeff * (input[i - 1][j] +
                                   input[i][j - 1] +
                                   input[i + 1][j] +
                                   input[i][j + 1]);
		
        //kad svako svoje zavrsi cekamo se		
		pthread_mutex_lock(&sync_us);
        steps[tid]--; //jedan korak gotov
        f = (steps[tid] > 0)? 1: 0; //da nisam stigao do kraja?
		if(steps[tid] < steps[(tid + 1) % 2]){ //sta je sa drugom niti?
			pthread_cond_wait(&move_on, &sync_us); //ili je cekam pa da radim dalje...
		}
		else{ //ili mene cekaju da kompletiram promenu ploca i javim da moze dalje
            temp = input;
			input = output;
			output = temp;
			pthread_cond_signal(&move_on);
            }
		
		pthread_mutex_unlock(&sync_us);
    }
	
	pthread_exit(NULL);
}

int main(int argc,char* argv[])
{
	int i,j;
	
	steps[0] = steps[1] = atoi(argv[2]); //broj koraka
	dim = atoi(argv[1]);
	dim -= 2; //zaobici ivice
	
	data = (double**)malloc((dim + 2)*sizeof(double*));
	result = (double**)malloc((dim + 2)*sizeof(double*));
	
	for(i=0;i<dim+2;i++)
	{
	    data[i] = (double*)malloc((dim+2)*sizeof(double));
	    result[i] = (double*)malloc((dim+2)*sizeof(double));
	}
	    
	for(i=0;i<dim+2;i++)
	    for(j=0;j<dim+2;j++)
	    {
	       scanf("%lf",&data[i][j]);
	       result[i][j] = 0.0;
	   }
	
	//naizmenicne ploce za obradu koraka
	input = data;
	output = result;

	pthread_mutex_init(&sync_us, NULL);
    pthread_cond_init (&move_on, NULL);
	pthread_attr_init(&atr);
	pthread_attr_setdetachstate(&atr, PTHREAD_CREATE_JOINABLE);
	pthread_create(&ts[0], &atr, obrada, (void *)0);
	pthread_create(&ts[1], &atr, obrada, (void *)1);
	
	pthread_join(ts[0], NULL);
	pthread_join(ts[1], NULL);
	
	pthread_attr_destroy(&atr);
    pthread_mutex_destroy(&sync_us);
    pthread_cond_destroy(&move_on);

    result = input;
	
    for(i=0;i<dim+2;i++){
		for(j=0;j<dim+2;j++)
			printf("%lf ",result[i][j]);
	   printf("\n");
	}
	pthread_exit(NULL);
}