Optimal loop unrolling for GPGPU programs