Diffractor is an experimental next-generation, compiler-based AD system for Julia.
Design goals:
- Ultra high performance for both scalar and array code
- Efficient higher order derivatives through nested AD
- Reasonable compile times
- High flexibility (like Zygote)
- Support for forward/reverse/mixed modes
- Fast Jacobians
This is achieved through a combination of innovations:
- A new lowest level interface (∂⃖ the "AD optic functor" or "diffractor"), more suited to higher order AD
- New capabilities in Base Julia (Opaque closures, inference plugins)
- Better integration with ChainRules.jl
- Demand-driven forward-mode AD (Applying transforms to only those IR statements that contribute to relevant outputs of the function being differentiated)
Diffractor is currently supported on Julia v1.10 . While the best performance is generally achieved by running on Julia nightly due to constant compiler improvements, the current release of Diffractor is guaranteed to work on Julia v1.10.
Currently, forward-mode is the only fully-functional mode and is now shipping in some closed source products.
It is in a position to compete with ForwardDiff.jl, and with TaylorDiff.jl.
It is not as battle-tested as ForwardDiff.jl, but it has several advantages:
Primarily, as it is not an operator overloading AD, it frees one from the need to relax type-constants and worry about the types of containers.
Furthermore, Like TaylorDiff.jl, it supports Taylor series based computation of higher order derviatives.
It directly and efficiently uses ChainRules.jl's frules
, no need for a wrapper macro to import them etc.
One limitation over ForwardDiff.jl is a lack of chunking support, to pushforward multiple bases at once.
Improved reverse mode support is planned for a future release.
While reverse mode was originally implemented and working, it has been stripped out until such a time as it can be properly implemented on top of new Julia compiler changes.
With that said, issues and PRs for reverse mode continue to be appreciated.
The plan is to implement this in two stages:
- Generated function based transforms, using the ChainRules, the new low level interface and Opaque closures
- Adding inference plugins
Currently the implementation of Phase 1 is essentially complete, though mostly untested. Experimentation is welcome, though it is probably not ready yet to be a production AD system. The compiler parts of phase 1 are a bit "quick and dirty" as the main point of phase 1 is to prove out that the overall scheme works. As a result, it has known suboptimalities. I do not intend to do much work on these, since they will be obsoleted by phase 2 anyway.
A few features are still missing, e.g. chunking and I intend to do some more work on user friendly interfaces, but it should overall be useable as an AD system.