Is there any tooling that lets one profile the individual parts of the compute graph? In CUDA land, I can imagine how to build such a thing. No clue how it would work on Mobile.