Tag ForwardDiff v0.3.0 [https://github.com/JuliaDiff/ForwardDiff.jl]

[jrevels]

This change breaks a couple of things in the v0.2.0 API in order to get rid of global state, reduce compile times, provide a totally non-allocating API, and support more array types. I've written an upgrade guide here, and have tried to ensure that using the old API throws depwarns when possible instead of straight up breaking code.

[ChrisRackauckas]

When you say "totally non-allocating API", you mean there's a simpler solution to this issue?

JuliaDiff/ForwardDiff.jl#136

[jrevels]

No - if your objective function mutates values, those values still need to be typed generically enough to allow for Dual numbers to work. Unfortunately, that's a fundamental limitation of operator-overloading AD.

What I meant is that it is now possible to call ForwardDiff's API functions in such a way that they do not allocate any additional memory on top of whatever the objective function allocates. In other words, if your objective function is non-allocating, calls to ForwardDiff function shouldn't need to allocate, as long you provide work buffers up-front via the AbstractConfig types.

[KristofferC]

No, if you want to use arbitrary preallocated structures in your function you still need to allocate them with the correct element type and pass them in yourself.

[KristofferC]

Well.. beaten by a second.

[ChrisRackauckas]

Got it. Thanks.