A lot of times the performance is data dependent, like randomized data can result in poor warp utilization where as sorted data can be much more efficient because you have less warps sitting around idle. We have an upcoming blog post on decompression kernels that show this clearly, stay tuned!