Multiple derivatives in forward mode in one pass about autodiff HOT 1 OPEN

Hi @ibell , this is not supported in autodiff, at least not yet. Computing one derivative per function call has the advantage that when you don't need derivatives, the overhead of using a forward automatic differentiation type (e.g., dual) is the minimum possible. This also allows a more efficient computation of derivatives when the derivatives w.r.t. only a small subset of variables actually matter in the algorithm (instead of computing all derivatives all the time, one can compute just a few of them).

Lastly, in most cases, the number of variables of our function is unknown at compile time, requiring thus dynamic memory allocation for the grad field of type Dual (in case we want to compute all derivatives in a single pass). This could trigger many memory allocations as the dual numbers are used in mathematical operations. That's one of the reasons why having grad as double in type dual decreases the overhead of computing derivatives (no dynamic memory allocation!).

from autodiff.