The big table of overloads

[gdalle]

This table recaps the features of each extension, with respect to high-level operators.
Each cell can have four values

• ❌: the backend does not support this operator
• ❔: the backend supports this operator but our extension doesn't call it yet / doesn't handle preparation
• ✅: our extension calls the backend operator and handles preparation if possible

backend	derivative	gradient	jacobian	hvp	hessian
ChainRulesCore	❌	❌	❌	❌	❌
Diffractor	❌	❌	❌	❌	❌
Enzyme (forward)	❌	✅	✅	❌	❌
Enzyme (reverse)	❌	✅	❔	❌	❌
FastDifferentiation	✅	✅	✅	✅	✅
FiniteDiff	✅	✅	✅	❌	✅
FiniteDifferences	❌	✅	✅	❌	❌
ForwardDiff	✅	✅	✅	❌	✅
PolyesterForwardDiff	❌	✅	✅	❌	❌
ReverseDiff	✅	✅	✅	❌	✅
SparseDiffTools	❌	❌	✅	❌	❔
Symbolics	✅	✅	✅	❌	✅
Tapir	❌	❔	❌	❌	❌
Tracker	❌	✅	❌	❌	❌
Zygote	❌	✅	❌	❌	✅

[adrhill]

The long-term goal should be to replace all ❔ with ✅, right?

[adrhill]

It would be nice to have a copy of this in our developer documentation.

[gdalle]

I like having this info in one place to help with maintenance. I think the issue makes more sense?

[wsmoses]

@gdalle Enzyme does have support for a pure derivative (presumably scalar?) operator, you just call autodiff with it.

[gdalle]

Thanks, this table is more about the shortcuts that each backend defines with other names. I know that with Enzyme most of what I want to do can be achieved with ´autodiff´, and there are probably a few places where performance is suboptimal, but I'll try to fix those