William Moses

Yeah for sure. Was thinking about this a little bit and I think the easiest thing to do would be: 1) When querying the type of an argument where this...

I wonder if perhaps the right place to do this is not as much on the call declaration, but to define a special pragma that results in an op execution...

What happens if you remove the `#pragma scop` in both cases? Regardless, this seems to just be a different ordering of optimizations question (or implementation difference from the pragma scop...

Resolved by removing .cache/gg, probably should have some mechanism for checking if that cache is out of date (and upon failure above reupload?)

Oh that’s cool last I checked dense wasn’t working. As for the batchnorm error, would you mind minimizing that (eg ideally using pure Julia and no packages), and posting an...

@avik-pal bumping the above so we can start working on a fix

👍 let me know when you've managed to reduce the BatchNorm bug into pure julia and I can have a look.

@avik-pal by the way we added some backtrace info for these types of errors if you want to see where your type instability is coming from (and perhaps fix it)

On 1.8.1 I get: ``` julia> Enzyme.autodiff(loss_function, Duplicated(x, dx), Duplicated(xmean, dxmean), Duplicated(xvar, dxvar), Duplicated(scale, dscale), Duplicated(bias, dbias), Duplicated(rm, drm), Duplicated(rv, drv)) ERROR: Duplicated Returns not yet handled Stacktrace: [1] autodiff(::Enzyme.ReverseMode,...

Also sret as an attribute is not always set even if Julia makes an sret. E.g. the code snippet in https://github.com/wsmoses/Enzyme.jl/issues/154 will result in the first arg as sret, but...