Created by: zhaoyuchen2018

• Give 200000 data to compute argsort on v100, can speed up ~190x before opt cost: 0.53s after opt cost:0.0027s

• Add fp16 support.

• FIx memory core dump issue: https://github.com/PaddlePaddle/Paddle/issues/20021

test=develop

Signed-off-by: zhaoyuchen zhaoyuchen01@baidu.com