Juan Ignacio Polanco

Hehe yeah, I was kind of forced to do that :)

Sorry, somehow I missed this issue when it first appeared. If I understand correctly, yes, it is possible to visualise 1-dimensional lines in 3D space. This is done using unstructured...

Hi, this might be the result of all MPI processes trying to precompile packages at the same time. You may want to try precompiling with a single process first, as...

Thanks for noticing this issue. After doing some tests, it seems that the performance issue is not related to the FFTW flags, which are properly passed, but to the overhead...

I'm not surprised to see that native single-GPU 3D FFT implemented in CUDA is way more efficient than the PencilFFTs version. Note that in PencilFFTs, a 3D FFT is implemented...

Thanks for the link. Their results look quite impressive! I think wrapping the multi-node CUDA FFT is the way to go. What's a bit annoying is that cuFFTMp is for...

Hi, are you sure you ran your example with 2 MPI processes? I say this because `proc_dims = (2, 2)` means that the domain should be partitioned over a total...

> I think your output illustrates the issue still though, because there are arrays with size 0? I agree! It would be good to handle this case in a more...

> The first transform (along x) is length 2, distributed between 8 processors (2 in y, 4 in z). The second transform is also length 2 (along y), and can...

I think `NoTransform` is exactly what you're looking for! Well, almost... There's a [`NoTransform!`](https://jipolanco.github.io/PencilFFTs.jl/dev/Transforms/#PencilFFTs.Transforms.NoTransform!) variant (note the bang!) which is compatible with in-place transforms.