写了一个kernel， D = A x B，第一批次运行最快，后面批次越来越慢，最终稳定在20ms，这是什么原理？

Total Time(getmillisecs): 7.863 ms, A:216314618994982912, B:216314615624105984, C:216314618997080064

Total Time 2(getmillisecs): 9.494 ms, A:216314619010023424, B:216314615489888256, C:216314607839477760

Total Time 3(getmillisecs): 13.651 ms, A:216314607852421120, B:216314615355670528, C:216314607854518272

Total Time 4(getmillisecs): 19.265 ms, A:216314607867461632, B:216314615221452800, C:216314613677948928

Total Time 5(getmillisecs): 22.403 ms, A:216314613690892288, B:216314615087235072, C:216314613692989440

Total Time 6(getmillisecs): 21.423 ms, A:216314613705932800, B:216314614953017344, C:216314607973695488