The alignment trap isn't needed if all userspace programs respect the alignment requirements. I guess llama.cpp might do that, so it works without it. Xorg I found did need it, even the arm64 build (or a userspace workaround like the memcpy library). The Ampere version should work as well, although I haven't tried it. The Ampere version also covers kernelspace, while mine only covers userspace, so more cards might work with it without extra changes, but at the cost of some performance (whether that would be noticeable or not, I don't know).

These would be best submitted upstream where the devs who actually understand this driver can comment if the patches are correct (or could be achieved in better ways). See submitting patches.

No objection to leaving this PR here for information for other interested users, but we are unlikely to merge it as a downstream only patch. If any patches are accepted upstream we're generally happy to cherry-pick them to get them into trees sooner.

The impression I got from the @6by9 comment linked above is that you could maybe guide us on what might be needed to upstream these changes.

If you could have a first pass at tidying it up on rpi-6.12.y, and create a PR against raspberrypi/linux rpi-6.12.y, then we can give pointers on what is needed.

I'd be happywilling to spearhead that effort, but could use some guidance from people more familiar with kernel development, because it sounds like a more legislative process than submitting a PR ;)

Yeah, @6by9 is probably a better guide - he's succeeded in upstreaming a number of patches recently.