Tensorflow Lite GPU Profiling

Question

Is there a way to understand how well tflite is using the GPU, and who it is competing with for GPU resources?

I have an app using tflite with a GPU delegate on android.

It can sometimes process images at 15 FPS and sometimes 3 FPS without any change to app.

I would like to understand what causes it to switch between these two speeds. I suspect it has to do with other apps like the camera using the GPU at the same time.

Are there any tools I can use to help understand if my app is actually using the GPU efficiently or if I need to manage other GPU processes better?

What I've tired:
- The normal android studio profiling tools don't give me any insights into this.
- The GPU profiling tool in android seems to just show rendering information and not anything related to other GPU processes. ( https://developer.android.com/topic/performance/rendering/profile-gpu )

Thanks

Answer 1

I am still looking but currently trying the snapdragon profiler. https://developer.qualcomm.com/software/snapdragon-profiler