How to use cudnn or cublas by DPC++?

[malixian]

Is it possible to achieve this demand? Can you give me an example？

[alexbatashev]

Both oneDNN and oneMKL have CUDA backend, and as far as I know, they use cudnn and cublas under the hood. As a bonus you get full DPC++ compatibility and performance portability across multiple vendors and device types. Would that fit your needs?