__kernel void test1(__global float *d_Result, __global float *d_Input, int N){
    const int tid = get_global_id(0);
    const int threadN = get_global_size(0);

    //float sum = 0;
	//int array1[131072];
	printf("Here!\n");
    //for(int pos = tid; pos < N; pos += threadN)
    //    sum += d_Input[pos];

    //d_Result[tid] = sum;
}