Not really happy with how this patch requires a copy of the image (goes against compositor design).

Also, in own tests I can't see the difference between BICUBIC and CUBICBSPLINEPREFILTER.
The patch seems to be working as expected and I checked the code runs on the filtered buffer.

If you can't tell the difference, it may not be working correctly. Should be as clearly apparent as in https://youtu.be/nfhTET86kdE

It should also be possible to demonstrate a stair interpolation such as running a sampling using cubic 10x at 10% increase; the end will be a muddy, blurry mess. Do the same with a prefilter and the result should be identical as having done it once at the same resultant value.

You must not duplicate whole buffer, limit it to the tile only.

Can we fix the design to work with proper sampling algorithms then?

The sampling algorithm and implementation is to be formulated in a terms that does not need require the whole buffer.

The sampling algorithm and implementation is to be formulated in a terms that does not need require the whole buffer.

Happy to bend around a poorly informed sampling structure, except that to properly derive the coefficients buffer, the entire image is required.

Not keeping the buffer around is surely a performance drain, and seems rather ridiculous. Is there a means to append a buffer as a sub-buffer or something that would keep Blender from recalculating the coefficients?

Open to suggestions here.

You don't need whole buffer, you only need tile. As i've said before, have a look at initializeTileData.

The coefficients are derived by looping through the entire horizontal and vertical rows and columns of data.

Are you suggesting to loop through tiles as opposed to grabbing the whole buffer?

Those coefficients are localizable to a pixel neighborhood. You can calculate them for the tile, and toss them away once the tile is handled.