kernel void square(global float* input, size_t width, global float* output){
    size_t x = get_global_id(0);
    size_t y = get_global_id(1);
    float i = 2.0 * (input[x + width*y] - 0.5);
    output[x + width*y] = 0.5 + 0.5*sin((1.0 - i) * sin(i*i + 6.2*float(x)/width) + cos(1.0/i + 6.2*float(y)/width));
}