09:02 glehmann: Lynne: no, but if it's caused by radv surely you can bisect it?
09:03 Lynne: yup, on it
09:48 Lynne: hakzsam: aa44a5a4ae93860 broke the ffmpeg vulkan prores decoder
10:01 Lynne: nevermind, it was our stupid issue (and the validation layers keeping quiet)
12:44 Lynne: checked again, that commit's seems to have broken it indeed, we were re-using push data, but even pedantically fixing it, it's still broken
12:51 Lynne: yup, can confirm, mesa git main is pretty borked right now, there's even a GPU reset on code that used to run
15:07 karolherbst: are fmax/fmin allowed to flush denorms in nir?
15:07 karolherbst: mhh guess they are
15:39 glehmann: karolherbst: they are even required to flush denorms if the float bit for that is used
15:40 karolherbst: yeah.. I think I found the issue I'm running into and that is that I don't flush denorms with instructions like flt...
15:45 karolherbst: yep.. that was it 🙃
17:08 mareko: replacing BASE with BYTE_OFFSET and DRIVER_LOCATION in different intrinsics is rather tricky because it's easy to break drivers because compilation doesn't fail if nir_intrinsic_base is used on an intrinsic that doesn't have it
17:09 mareko: and there is 438 uses of nir_intrinsic_base
17:10 mareko: I guess we'll have to live with the fact BASE means "whatever"
17:14 karolherbst: might be easier to port all the drivers to make it mean bytes first