Mohamed Tarek

No it's mostly the benchmark above that favors a mapped array. Whether this is of practical relevance or not will be case-dependent.

True, I can't think of a case where Fill is much faster. So I would go ahead and use `MappedArray` everywhere. It seems overkill to me but the Julia compiler...

Yes, I get the point of the function. I drafted some node bc functionality but currently debugging it. Then I will try to fix this issue.

This is cool! Does the speedup take the coloring time into account? It would be nice to also parallelize the coloring.

@ThummeTo can I get an invite to the repo? I fixed it locally and used https://github.com/JuliaNonconvex/DifferentiableFlatten.jl. I can open a PR or push directly to master.

Yes

Since we have a sufficiently lightweight package that implements this feature now, should we close this issue @gdalle?

Let's think some more about getting AD.jl to work. We need API suggestions and a PR. I am not opposed to the idea in principle but just need to be...

I suggest we close this and open issues for specific array types, e.g sparse arrays when a use case arises.

I think we need a more in-depth analysis of where exactly sparsity gets lost in ImplicitDifferentiation with an example. I suspect if sparsity is lost, it's mostly due to upstream...