Using ForwardDiff with a vector function (as in IntelVectorMath)

Is there any way to efficiently use ForwardDiff with functions that operates on vectors and return vectors? Suppose I have a function f that operates on a vectors by applying some operation independently to each element (as in a broadcast operation). This, for example, is the case for many functions from IntelVectorMath or AppleAccelerate. Although, formally, f is a function from R^n to R^n, in reality it does not make sense to compute an Hessian for it, since it would be a diagonal matrix. Is there a way to have ForwardDiff return only diagonal elements?

Note that repeated calls to ForwardDiff.derivative, although feasible, would be a poor workaround, since would not allow me to use the fast math operations of IntelVectorMath or AppleAccelerate.

Start by reviewing ForwardDiff.derivative and the handling of vector-valued functions. Compare the requested behavior with the elementwise operations provided by IntelVectorMath or AppleAccelerate. Done means identifying a supported efficient way to return only diagonal derivatives without repeated scalar derivative calls.