Integer vector division is working incorrectly on Macs with Apple M1 (ARM) about hlslpp HOT 7 CLOSED

I've fixed division in NEON, the explanation for why it was broken is here f8ebd87. Let me know if this is still an issue for you. It is "fixed" in the sense that it is consistent with SSE now, but the issue of large integers remains.

from hlslpp.

@redorav Thanks for quick turnaround! My tests are passing fine with this update on M1 ✅

from hlslpp.

Hi @egorodet, it is a known issue that integer division is broken. There are no intrinsics for integer division and when I sat down to find a solution I couldn't find anything that wasn't a bit of a monster and left it for the future. Currently all it does is cast to float, divide and cast back which obviously isn't very good.

I had an issue open already, I'll close that one and track here instead. If you know of a good solution I could put in that would make it much easier, otherwise I'll see when I can find a bit of time to look into it.

from hlslpp.

Copying original comment from the other bug:

SSE doesn't have native division instructions for vectors. One possibility is to extract the scalars, divide, then put back. Another alternative is to take a look at this website which seems to have alternatives and claim to be fast

http://libdivide.com/

from hlslpp.

Hi @redorav, yes I see that current implementation is not perfect, but it seems to work in most cases on x86/64 architecture, but gives completely wrong results on ARM (Apple M1 in my case). So additionally to existing issue, there's something wrong in the NEON implementation of the current workaround.

from hlslpp.

I added workaround in the code of my wrapper class Point, so it should not be a stopper for me. It's up to you if it's better to keep or close this issue. I'm fine with description of the existing issue, just want to raise its priority because current implementation could give wrong results.

from hlslpp.

Thank you for detailed report, hlsl++ is a lot better thanks to your efforts

from hlslpp.