David Widmann
David Widmann
IMO ForwardDiff should not depend on FFTW. Instead I think it would be preferable to fix https://github.com/JuliaMath/AbstractFFTs.jl/issues/56 and only depend on AbstractFFTs.
It's not an extension on Julia < 1.9.
I guess this would be annoying in practice. Downstream packages that want to use ForwardDiff + FFTW would have to copy the diff rules or add some workaround for them...
One could also put the FFTW rules behind a `Preference`, similar to the NaN-safe mode. But this would only prevent ForwardDiff from loading the file by default, FFTW would still...
Not supporting the FFTW-specific rules on Julia < 1.9 might be a bit surprising but I think it would still be much more preferable than depending on FFTW. The current...
I wonder if one could use `map!` instead of broadcasting? I would assume that this would still be GPU-compatible but possibly there is (almost) no runtime cost for `Array`?
That error usually occurs (also in your case as the stack trace suggests) if you try to update an array of non-dual numbers with dual numbers. You should ensure that...
I'm not sure what you mean by rules here. As in other non-ForwardDiff Julia code, you just have to ensure that the arrays you create can store the results of...
Maybe #481 should have been put in a breaking release after all. Maybe it would be safer to revert it and then prepare a breaking release with it?
> While x really is Hermitian, x .+ Dual.(0, rand.()) represents a small perturbation around x, and most such perturbed matrices will not be Hermitian. Using an algorithm which assumes...