Benchmarking the cost of thread divergence in CUDA