Kristoffer Carlsson

For four threads improves performance by ~20% but for 16 threads improves performance by 3x. This is likely because I am now allocating the scratch space on each separate thread...

Having to send a time into the function for constraints is a design failure.

8

This should just be handled by sending in a closure `f(x) = f(x, t, ...)`.

Quite a lot slower, hm. I don't really see why this should allocate any more.

Implement threading.

4

I started implementing some threading support using the threading branch in master julia. The linked gist below implements threaded kd tree creations. Code is non polished. https://gist.github.com/KristofferC/15beed394361a325ff33 The speedup is...

WIP: use SIMD.jl for partials

8

This needs quite well testing for downstream users to ensure regressions haven't crept in due to the implementation in SIMD.jl.

WIP/RFC make the partial scalars VecElements

7

Only implemented enough so that the benchmark in https://github.com/JuliaDiff/ForwardDiff.jl/pull/555 can be tested. Putting it up here in case people want to play with it. Results of the benchmark in #555:...

wip: implement macro for users to define their own derivatives

8

This is a start towards an official API for fixing https://github.com/JuliaDiff/ForwardDiff.jl/issues/89. It provides macros for derivatives, gradients and jacobians such that user defined derivatives are easily injected: ```jl julia> f(x)...

From the following benchmark ```jl function rosenbrock(x) a = one(eltype(x)) b = 100 * a result = zero(eltype(x)) for i in 1:length(x)-1 result += (a - x[i])^2 + b*(x[i+1] -...