Vedant Puri

@mcabbott what do you think about moving this change to a GPUArrays ext?

@ali-ramadhan @whymusticode what is the consensus here? Is `permutedims` + `fft` the way to go? BTW, I would be grateful if somebody could provide a CUDA-compatible DCT implementation. :D

fft/dct methods can be placed in an `AbstractFFTs` extension package. The plans can subtype `AbstractFFTs.Plan` and so they can support the interface.

FYI - I'd like to use Descartes to model geometry for my differentiable meshing algorithms - https://github.com/vpuri3/DiffMesh.jl

addressed here: https://github.com/JuliaGeometry/Descartes.jl/pull/53

waiting on https://github.com/SciML/LinearSolve.jl/pull/339

We have come across this issue before when trying to form nonlinear operators eg, `f(u) * u, df(u)/du * u`, etc with ComposedOperator (https://github.com/SciML/SciMLOperators.jl/issues/159). This seemed like a design constraint...

I'm tied down with school work at the moment. I have a long flight this weekend when I'm gonna start working on this. although it's a small change it's gonna...

It looks like your time-stepper, `ETDRK2`, and most other exponential methods require full matrices. They depend on `ExponentialUtilities.jl` that, upon a cursory look, requires full matrices. I tried your problem...

of course. I'll let you know here when the split function + imex is done. also, if you are using fourier stuff extensively, I've implemented almost all relevant operators with...