* fix inference perfermence problem caused by selecting cudnn kernel for softmax

* recover use_cudnn in opmaker of softmax