* support reduce_max op on kunlun * support reduce_max op on kunlun * support reduce_max op on kunlun * support reduce_max op on kunlun
