Thanks a lot for looking into this! I’ve created a new topic with a link to test data and to discuss this further: vtkWindowedSincPolyDataFilter performance improvement
Yes, this is a very complicated filter indeed. We utilize most of its features - that it can operate on a wide range of data types (integer and floating-point, one or more components), of practically any size, apply a (non-linear) transform, and can work faster when reslicing axes are aligned with image axes.
However, there is one critical use case: quick extraction of an image slice. We use this filter for displaying 2D slices of an image volume, so if performance of this could be made any faster that would visible improvement of slice browsing speed. Total display time includes sending the image data to a texture so that it can be rendered as an image plane. This sending to the GPU probably takes considerable time, too, and I’m not sure how much that part is optimized for this use case (very frequent update of texture, without changing its size or data type).
Reslicing on the GPU could be fast, but supporting of any data types, dynamic application of all kinds of non-linear transforms, supporting large images, etc. would be a lot of work to implement on GPU.