Is cudnn v4 /rc5 supporting FP16 version of cudnnBatchNormalization ?

It seems that cudnnBatchNormalization fp16 functions are not supported in cudnn v4 and rc5.
I have an error “CUDNN_STATUS_BAD_PARAM” when using fp16.

cudnnBatchNormalization fp16 expects ‘float’ data type as BN parameters (mean,variance, scale tensor descriptor), else it gives “CUDNN_STATUS_BAD_PARAM”. However, it still does not work well for me. I am getting incorrect results including 'nan’s.
Has anybody used cudnnBatchNormalizationInference/Training APIs with fp16?