Faster three dimensional higher order interpolation?

I have a few algorithms that are being bottlenecked by the interpolation over three-dimensional volumes, unfortunately, trying to acclerate it using interpn with gpuArrays limits me to the use of linear or nearest interpolation, which significantly affects the results. Are there any higher order interpolation schemes that can be run on the GPU, with some other function?

Answer 1

[Joss Knight]

Thanks for the request, it will help us prioritise future work. In the meantime, it is possible to write your own interpolation code for the GPU using ARRAYFUN. You can either use up-level variables or pass shifted versions of your data (using indexing) which then get combined inside the arrayfun function. So for instance, for a cubic interpolation you will need four variables per dimension; or perhaps better, call ARRAYFUN multiple times to interpolate each dimension in turn.

Answer 2

[Bruno Luong]

You probably need to change an existing open interpolation code to use efficiently gpu arrays.

The cubic methode is smooth method and it is easier to implement than spline, but it requires uniform grid.

The spline is more complicated, you might start with a code duch as this one by your trully

I never work my own makima so I can't comment, but it should be more difficult than spline.

It requires a little bit of work in any case.