Yingbo Ma

That's not how you use PerformanceTestTools.jl. See https://github.com/YingboMa/FastBroadcast.jl/blob/master/test/runtests.jl#L84

https://www.cs.mcgill.ca/~chris/pub/PaiS02c.pdf page 6

> disallow multiplication of complex numbers with AbstractWirtinger and require users to use chain, since the order of chaining matters here, too (Still needs a better error) Why disallowing `::Wirtinger...

One common example is a wrapper type that tries to simulate `Var{T}

For symbolic manipulations, we need to have `Var{T}` that doesn't call into `foo(::T)` directly, since `Var{T}` doesn't know the actual value, but at the same time, `Var{T}` needs to behave...

Here are some thoughts on optimizing `FiniteDiff._colorediteration!`. Currently, our iteration order is to linearly traverse the compressed Jacobian and to find the appropriate column in the uncompressed Jacobian to copy...

When chunking, we have `ndims(partials) == ndims(primal) + 1`. We don't force `Vector` or `Matrix`. We only use ` extern(Zero(), (1, Zero(), [1 2; 3 4], [1, 2])) 2×2 Array{Int64,2}:...

FD2 doesn't do forward mode AD on multiple arguments. It is possible to do, but aware that when two chunks meet, the larger one wins.

Actually, I don't think the above `extern` solves the actual problem with `Zero`. What do we do if we get multiple partials from `frule`? We need to know the actual...

> Could you provide an example of where this behaviour is currently being exploited to accelerate inference? accelerate inference? > Do we have any frules that are already supporting this...