More complex test cases

[g-bauer]

Hello!

Sorry for the very generic title. I did some tests regarding differentiating a bit more complex function (similar to what my real-world use case would be) than the last time I had a chance to play around with Enzyme. I put everything into a small repo with different cases and the issues that I encountered. Let me know if you want to have them in separate GH issues.

---

In essence, I have a function with this signature: a(p: &Parameters, t: f64, v: f64, n: &[f64]) -> f64 and I want to compute derivatives w.r.t t, v , and n for given constant p.

• case1: forward AD w.r.t. to one scalar input works. But when I add another computation, suddenly the formerly correct forward AD now yields a different (wrong) result. I have no idea what's going on here. Along the way there were some issues with (I guess?) allocations inside the function that caused some trouble (commented in the file).
• case2: Copied a working example from case1 into a new file and it stops working. Had to switch from using Vecs (or ndarray::ArrayBase) internally to slices to make it work. No idea why it worked in case1.
• case4: Reverse AD does not work when I have my f64s marked as Active and my &[f64] marked as Duplicated. Wrapping the function with a function that takes &f64 and Duplicated for the floats works.

I hope this is somehow helpful. Let me know how I can help to provide more information to find out what's going wrong.

[ZuseZ4]

Thank you, those examples are great.
On a quick check it also looks like not a single one of them crashes my frontend, and they all just fail the backend? That would be exciting.

If Enzyme fails and you have some time to help debugging, then the right thing to do is to reduce them one-by-one, i.e. try to only have one autodiff call per minimized example and reduce that as far as possible (as long as it still crashes). The following steps are described here: https://enzyme.mit.edu/index.fcgi/rust/debug_backend.html
The goal is to create an issue in EnzymeAD/Enzyme, with some LLVM reproducer.

The Enzyme author cares most about incorrect results (rather than crashes), so if you can minimize them too they would hopefully be fixed first. They are likely a bit harder to reduce and I don't have exact instructions on how to do that yet, but I'm happy to help if you have some time.

Also for the next time, can you please open future issues in rust-lang/rust? I'm going to archive this repo in the foreseeable future, so I aim to have future discussions upstream.

[ZuseZ4]

Ping @g-bauer, in case that you have time to minimize any of these instructions:
https://rustc-dev-guide.rust-lang.org/autodiff/debugging.html