A parallelization cost model for GPU