Alec Hammond

Wow, thank you for the very comprehensive response! > relates to compliance limitation from CUDA.jl This makes sense, thanks. > We are currently developing proper support for different kind of...

> kernels declared using @parallel accept Data.Array and Data.Number arguments, and should contain only stencil computations to be performed with one of the submodules ParallelStencil.FiniteDifferences{1D|2D|3D}. On a similar note, suppose...

Yes, the [broadband directional coupler](https://github.com/lukasc-ubc/SiEPIC_EBeam_PDK/wiki/Component-Library-description#broadband-directional-coupler) (specifically the TE mode device). The wiki says the following link should provide more info: https://github.com/lukasc-ubc/SiEPIC_EBeam_PDK/tree/master/Documentation/Broadband_DC

> plan to support FFT differentiation Just out of curiosity, do you really mean FFT differentiation (using a DFT/FFT to approximate a derivative) or are you referring to implementing forward/backward...

Just passing by @flaport -- but you shouldn't need the conjugate. If the materials are lossless (and the modes are normalized right) then the conjugation won't change the results anyway....

Adding a plugin like this would actually be mutually beneficial: * gdsfactory gets another solver * EMEpy gets better corner case testing and validation

I'm interested in helping with this. I'll reach out once I get more bandwidth.

Thanks for the quick feedback! I built an FFT convolution package using autograd awhile back. It only supports 2d, but it's rather easy to generalize. The performance was the same...

Could we resurrect this? Or has an alternate solution already been implemented?

`UserExpression` throws an error: ``` AttributeError: 'Permeability' object has no attribute '_ufl_shape' ``` Any ideas on a fix?