11:02 snowycoder[d]: mhenning[d]: I just pulled VK-GL-CTS, using instead `vulkan-cts-1.4.3` the situation is much better (just 1 flake and I'm halfway through)
12:22 snowycoder[d]: mhenning[d]: Cornfirmed, non-latest CTS passes even without the latest commit 👍
13:47 marysaka[d]: mhenning[d]: did you saw usage of 0x0f60 with the value 0?
13:48 marysaka[d]: I was looking at my notes from years ago on Switch stuffs and this is actually used to enable or not tiled cache
13:49 marysaka[d]: Looking at all my dumps on Pascal, Ampere and Ada, it is explicitly enabled at 3D channel init with what looks like the actual tile configuration (the value is the same across all those gens from what I could tell)
15:26 mhenning[d]: marysaka[d]: In the traces with NVB197_FORCE_HEAVYWEIGHT_METHOD_SYNC ? There's some usage of 0f60 with value 1 afterward
15:26 marysaka[d]: mhenning[d]: okay hmm if you mean after it might be possibly another channel
15:27 marysaka[d]: currently trying to dig into enabling that and figuring out some configuration for the tiled cache
15:27 mhenning[d]: The NVB197_FORCE_HEAVYWEIGHT_METHOD_SYNC traces are here, btw: https://gitlab.freedesktop.org/mhenning/re/-/blob/main/vk_test_overlap_exec/blackwell_proprietary.out?ref_type=heads They're dumps of what the proprietary driver emits for pipeline barriers on blackwell
15:28 marysaka[d]: okay yeah that seems to match what I was seeing
15:29 marysaka[d]: it is enabled initially on the first submit and then they set the value again (+ some flush with 0f80)
15:31 marysaka[d]: so that FORCE_HEAVYWEIGHT_METHOD_SYNC is likely needed for f80 method
15:31 mhenning[d]: Yeah, looking at my vkcube traces it seems to 0f60 with VALUE 1 at init time
15:32 mhenning[d]: marysaka[d]: I assume you mean f60?
15:32 mhenning[d]: but yes, good to know
15:32 marysaka[d]: mhenning[d]: I mean f80 if you look your dump you have it at the start
15:33 mhenning[d]: ah, okay yes I see those sprinkled around
15:35 mhenning[d]: In that case, I suppose I'll drop https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37671/diffs?commit_id=50700da39a31421637a8b8556e90448bf585659b for now, since it sounds like it isn't necessary and Snowy indicated it isn't necessary for cts
15:36 marysaka[d]: mhenning[d]: If I remember correctly one of the tiled cache control apis for NVN on Switch touch SET_MUTABLE_METHOD_CONTROL_TREAT_MUTABLE_AS_HEAVYWEIGHT, currently trying to find my notes again on that it has been years...
15:37 marysaka[d]: but not seeing that thing used so maybe doesn't matter for us anyway