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

int main(){

	int i,j,it;
	
	// set the size of our test arrays
	int numel = 2000;
	
	// allocate and initialize test arrays
	float A[numel][numel];
	float Anew[numel][numel];
	for (i = 0; i < numel; i++){
		for ( j = 0; j < numel; j++){
			A[i][j] = drand48();
		}	
	}
	
	// apply stencil 1000 times
	#pragma acc data copy(A), create(Anew)
	for (it = 0; it < 1000; it++){
		
		#pragma acc parallel loop
		for (i = 1; i < numel-1; i++){
			for (j = 1; j < numel-1; j++){
				
				Anew[i][j] = 0.25f * (A[i][j-1] + A[i][j+1] + A[i-1][j] + A[i+1][j]);
				
			}	
		}
		
		#pragma acc parallel loop
		for (i = 1; i < numel-1; i++){
			for (j = 1; j < numel-1; j++){
				
				A[i][j] = Anew[i][j];
				
			}	
		}
		
	}
	
	// do something with A[][]
	
	return 0;

}