I finally managed to reproduce the issue. The problem happens when running the OIDN binaries on Ventura, which were compiled with an Xcode targeting the newer Sonoma by default. I think the…
It worked, thanks!
M2 Max on Sonoma:
Start 75: cycles_denoise_metal
75/305 Test #75: cycles_denoise_metal ............................................ Passed 13.55…It seems none of the Metal tests appear in the list of tests when running locally. How is it possible to enable them?
I just ran the tests on an M2 Max + Sonoma and only the compositor tests failed. I'll now try on an M1. I suspect that maybe OIDN does something in Metal which works on Sonoma but not on Ventura.
Strange. I’ll run the tests locally and report back whether I get these errors too.
Do you think the x86 compositor failures could be related?
@brecht Do you know which GPUs or macOS versions needs this workaround? OIDN supports Metal only on Apple silicon and macOS Ventura and later, so the workaround could be disabled only in these…
Here's a fix: 82942c7f2a
The only remaining issue is that I had to disable the following workaround because OIDN…
This PR seems to be incomplete. The Metal backend doesn't support pointers, buffers must be used instead. This logic is missing from the code.
@brecht Yes, the output difference is expected due to the improved network weights.
Since OIDN 2.2 with Metal support is out (9db3b38d50), we should move forward with this PR.
OIDN 2.2 release candidate is out: 9db3b38d50
This includes support for Meteor Lake, Metal, and ARM64 on all OSs, and…
Thanks! It looks good except that failed test but it's an out of memory error (probably because the GPU has less memory than what we typically use for testing), so I think it's fine.
@LazyDodo There's an easy way to test whether OIDN works on GTX 1660. Could you please download the latest OIDN binaries (https://github.com/OpenImageDenoise/oidn/releases/tag/v2.1.0) and run…