Support ETDump Generation in JNI Bindings

[GregoryComer]

🚀 The feature, motivation and pitch

ExecuTorch provides the ability to generate ETDumps through the native API surface, but this functionality is not currently exposed through the JNI bindings. It would desirable to be able to generate an ETDump from Java, as right now there is no way to do this.

Alternatives

No response

Additional context

No response

RFC (Optional)

No response

cc @kirklandsign @mergennachin @byjlw

[kirklandsign]

Hi @GregoryComer I'm curious about the use case. Could you please add a bit more details?

[GregoryComer]

Hi @GregoryComer I'm curious about the use case. Could you please add a bit more details?

This is part of the general goal of reaching API parity for Java bindings, such that Android devs can rely solely on Java APIs without needing to drop down to C++ for common use cases. Since performance profiling is a relatively common part of onboarding a model, it would be desirable to be able to generate an ETDump entirely within Java.

Right now, if a user starts with Java and wants to profile in-app, they'll have to re-write their inference code in C++ and integrate native build. Runners might help, but they are somewhat high friction and also don't necessarily capture any CPU or resource contention happening in-app.

I was thinking we can maybe just add a Java EventTracer class which can be passed into Module, the same way it works in C++. https://github.com/pytorch/executorch/blob/main/extension/module/module.h#L52. Any objections?

[mergennachin]

Actually, I would have to set low-pri for this use-case, considering the basics for Android ready first.