Progress has moved on since the Jetson Nanos.  The lion kingdom's $150 Jetson Nano had .25 TFLOPS in 32 bit float.  That gave it 6FPS with a pure C++ implementation.  Modern $250 jetson orin nanos have 8 TFLOPS in 32 bit float.  Modern, sub $100 phones have over .4 TFLOPS in 32 bit float which is just enough.  The trick is they can't run CUDA or CUDNN.  The world is rapidly dropping CUDA because of the high cost.  

The largest repository of adreno examples might be openpilot.  The models are in ONNX format.

https://github.com/commaai/openpilot/blob/master/selfdrive/modeld/models/

It uses tinygrad & opencl as the backend.  16 float seems to be a waste for this application.  32 bit float gives better results.

There might already be an openpose for android.

------------------------------------------------------------------------------------------

The lion kingdom got a 5Ghz wifi dongle for the jetson nano but it seems to have inferred its last image.  The phone app could use optimization with the BufferedStream class, too.

An on camera flash is needed.  The godox doesn't wake up without the full ID signal, so it had to be constantly power cycled.  Usage in a model session would require the wireless system or adjusting the timeout, risking its unreliable persistent settings.  A smaller flash might be the only way.  Recovering the 580EX II would require a donor flash.