Despite ASTCENC_INVARIANCE ON, results on M2/Neon (macOS, clang) differ from X86_64/AVX2 (Ubuntu, GCC)

[MarkCallow]

Probably not just limited to the 2 platforms mentioned. Unfortunately GitHub Actions has managed to completely break Windows builds with their 20240603.1.0 runner image so I cannot investigate further at present.

I don't think this is new in 4.8.0.

By results differ I mean the encoded data has slight differences such that memcmp fails.

[solidpixel]

Thanks Mark, I'll take a look.

This usually ends up being compiler version specific, so if you can share the specific compiler versions that would be great.

Do you have an input image and CLI command line that repros this that you can share?

EDIT Also, just to check, is this definitely a compressor problem? I know we hit issues in the past with the mipgen preprocessor in ktxtools not being invariant.

[solidpixel]

@MarkCallow Are you able to share more info on how to repro this one, or can we close?

[MarkCallow]

I need to do more investigation. Sorry for dropping the ball.

[solidpixel]

No worries. I'm looking to line up a 4.9 maintenance release for ~September, so if there is a need to do something it would be good to know in the next couple of weeks.

EDIT: Date is fungible, so if you need a bit more time that's fine.