Deploy Keras model on ARM 820 to use GPU

I have an existing Keras model that I wish to deploy on my drone, using the GPU.    (Dense depth estimation from a single camera.)

The drone has a Qualcomm chip set - ARM 820, Adreno GPU - not an ARM GPU.     I've followed the examples for ARM CL and have tested that one of the examples will work, using OpenCL.   Great.   

However, the path seems a little steep to deploy my own model.    ARM NN seems to show a better path, but it's unclear from the documentation if it supports anything other than ARM GPUs.

Looking for some ideas/advice on the best way forward.   (Or heck, should I be trying out tvm or another approach?)

Thanks!