Created by: qingqing01

Fix #6231 (closed) Fix #6232 (closed) fix #6002 (closed)

Mainly for GPU implementation, use one kernel to do the computation.