Moving ahead with CUDA support

Specific Domains GPU
, , ,
1

For a number of applications CUDA (and AD) support of various packages (e.g. FourierTools.jl) would be great. One road-block is the support of a form of shifted and padded images.
Sometimes a requirement is a shifted and padded (writable) view, which one gets by either nesting OffsetArray and PaddedView with a mutable extension (see select_region_view() in NDTools.jl) or use a mutable version of ShiftedArray. To this aim, I filed a pull request more than a years ago with ShiftedArray adding CUDA.jl support via the Adapt.jl mechanism.
However, this is unfortunately not moving ahead and currently there is no feedback from @piever .
What is the best way to move on? Register a new package with a new name from a fork and base future work on this?

I saw related efforts (CUDA and AD support) on the 2021 roadmap for Images.jl. They seem to be using PaddedViews.jl. Is there a preference of PaddedViews combined with OffsetArrays over ShiftedArrays? I guess using ShiftedArrays could remove one level of type nesting.

2 Likes
2

… decided to add the type MutableShiftedArray and make another PR. What is the best way to proceed here?

3

… the MutableShiftedArray is different enough to warrant releasing a new package, which is what I finally ended up doing:

However, regarding the ShiftedArrays.jl there is still the question, how to proceed. In my view it would really benefit from CUDA.jl support as suggested by the PR. E.g. the CircshiftedArray and the fftshift are somewhat commonly used and important.