Fine-Grained Treatment to Synchronizations in GPU-to-CPU Translation