Hi! For the third time, I am trying to implement some wave optics algorithms in Julia but each time I start, I suffer a bit from the current automatic differentiation status. In principle, my functions look like: function propagate(field::AbstractArray{T, 6}, distance::NumberOrArray, …

Hi Felix! [image] roflmaostc: I’d love to use DifferentiationInterface.jl but if FFTs do not work, that’s gonna be a hard time. I just want to point out that the question of using DifferentiationInterface is pretty much orthogonal to the choice of backend. At least if you fit inside the rest…

Reactant.jl supports FFT (and its differentiation), used quite extensively in NeuralOperators.jl (it also runs on CUDA, TPU, CPU and whatever other accelerator you might want). Is your field_p a 6-dimensional tensor? (we added upto 3-dims for FFTs in reactant though it is easy to expand if that is …

[image] avikpal: Reactant.jl supports FFT (and its differentiation), used quite extensively in NeuralOperators.jl (it also runs on CUDA, TPU, CPU and whatever other accelerator you might want). But it’s not supported by Enzyme? What would I to do to use it? [image] avikpal: Is your field…

[image] roflmaostc: But it’s not supported by Enzyme? What would I to do to use it? Reactant.jl uses Enzyme.jl for autodiff support. So you just need to @compile Enzyme.gradient(....). [image] roflmaostc: Yes, it’s a 6dim array. FFTs are only done long 1st and 2nd dimension though. Th…

Is there any further link? Didn’t see anything specific in the docs. So reactant introduces some FFT rules but they are not included in Enzyme itself?

[image] roflmaostc: Is there any further link? Didn’t see anything specific in the docs. Compiling Lux Models using Reactant.jl | Lux.jl Docs has an example of compiling enzyme.gradient. I just realized we don’t show Enzyme examples in Reactant (will add something today) [image] roflmaost…

Not quite correct, but an analogy is that Reactant is similar to jax.jit, and Enzyme is similar to jax.gradient. The difference here is that if you Reactant.@compile a function, it converts your julia code into a nice tensor program that can be automatically optimized/fused/etc and executed on whate…

[image] Autodifferentiation with FFT, with Enzyme? Numerics It would be great to have some Enzyme rules in AbstractFFTs.jl (there’s a draft PR at Add EnzymeRules by sethaxen · Pull Request #103 · JuliaMath/AbstractFFTs.jl · GitHub , but someone has to adopt the …

Wait so when you Reactant.@compile with Enzyme.gradient on the inside, Enzyme nonetheless differentiates the Reactant-generated code?

yes julia> using Reactant, Enzyme julia> function foo(x) return sum(x) end julia> x = Reactant.to_rarray(ones(10)); julia> @code_hlo optimize=false Enzyme.gradient(Reverse, foo, x) module @reactant_gradient attributes {mhlo.num_partitions = 1 : i64, mhlo.num_replicas = 1 : i64}…

Which direction: DifferentiatonInterface, Enzyme, Zygote with CUDA and FFTs?

General Usage

giordano September 9, 2025, 2:10pm 9

Topic		Replies	Views
Autodifferentiation with FFT, with Enzyme? Numerics question , fftw , autodiff , enzyme	13	647	September 9, 2025
What's the state of Automatic Differentiation in Julia January 2023? General Usage autodiff	41	12874	June 21, 2024
What is the correct way to autodiff a simple CUDA kernel using Reactant.jl and Enzyme.jl Machine Learning question , package , autodiff , reactant	20	460	September 23, 2025
Speeding up Zygote autodiff for numerical loop Performance question	13	404	December 16, 2024
ForwardDiff and Zygote cannot automatically differentiate (AD) function from C^n to R that uses FFT Numerics	23	4710	March 25, 2022

Which direction: DifferentiatonInterface, Enzyme, Zygote with CUDA and FFTs?

Related topics