actually, the histogram-based tree learning algorithm doesn't need the sorting. The main cost of it is the memory access. The GPU version is slow (when #data is

actually, the histogram-based tree learning algorithm doesn't need the sorting. The main cost of it is the memory access. The GPU version is slow (when #data is small) due to the additional memory copy cost between GPU and CPU. I think this issue can be fixed in the future.