Valentin Churavy
Valentin Churavy
@MilesCranmer in particular I am interesting in how much it grows. See also https://github.com/EnzymeAD/Enzyme.jl/issues/1182
To temper expectations, I probably won't have time to look at things until end of January.
> It's weird because it seems to be worse than before even. Before, even though it took a very long time, it eventually compiled. But now it seems to just...
I do think we need to actually check what is creating such deep stacks here
Compilers are not supposed to be recursive and if they are we should transition to a worklist approach
Yeah this works for lldb as well. You can also now use the Julia profiler during compilation, since it doesn't hang anymore to get some idea of the stack traces...
On @swilliamson7 this seems to yield another 50s improvement from 187->132 We still spent most of our time in type-tree.   Across those three towers we spend: 83s out...
> Does this mean that autodiff_deferred can't differentiate other autodiff_deferred? No it definitely needs to be `autodiff(autodiff_deferred(autodiff_deferred()))` for third order.
I think this is due to the EnzymeRules for KernelAbstractions not supporting reverse mode yet
I don't actually remember what was needed for reverse GPU support