[DTL]

" if I could buy an hardware downscaler which uses SinPowResizeMT() I would, but I can't"

It only part of a solution but not full solution. And also the UserDefined2 kernel expected to be better. If you need to realtime live AVS processing I think it is possible with simple windows-box running some not very old windows with directshow and SDI in/out board (for single home use the HDMI may be used) capable of directshow interfaces (good board running in and out at once - some may only in or out).

And after windows load install ffdshow and AVS - open graphedit and construct graph of hardware input source - ffdshow in RAW processing mode and hardware out - put avisynth script in ffdshow and click run graph. I thinking of making some like this for our broadcast company to run mvtools degrain before main broadcast MPEG encoder so help MPEG coder with lower noised content. But current dying civilization do not have interest in quality broadcast and we have no one complain on quality.

So if the in/out SDI (or IP-in/out for IP-based network) board and its driver would be nice (may be AJA and not cheap Blackmagic ?) and not hang or crash after several hours/days of runtime you can make some part of this planet more happy with both mvtools denoise and/or UserDefined2 (or SinPow) downscale of 16K/8K/4K source to better FullHD in live broadcast. Though I think some hardware vendors already uses the kernel like in SinPow or UserDefined2 resize - I see some like this adjustment in some good video camera setup instruction and it was called 'peaking adjustment'.

My tests with AVS and very old Decklink cards with SD-SDI input and ffdshow (for input only and running sort of dual input colour analyser for many cameras colour tone grading/alignment) result in lowering framerate update after about 2..3 hours of runtime. So it is not applicable for airing to the part of the planet for many days without restart. May be better cards or better directshow version or patched ffdshow required.