07:41 jani: tzimmermann: mlankhorst: which branch should I push https://lore.kernel.org/r/20260615184737.899892-1-xmei5@asu.edu to at this time?
07:42 jani: i.e. cc: stable stuff
07:57 tzimmermann: jani, drm-misc-fixes should be fine
07:58 tzimmermann: it will then enter upstream next weekend
08:19 jani: tzimmermann: thanks, pushed
17:25 zmike: mareko: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42388
18:39 karolherbst: nir_opt_phi_precision is doing something weird: https://gitlab.freedesktop.org/mesa/mesa/-/work_items/15704#note_3531383
18:54 glehmann: karolherbst: can_convert_load_const is broken I guess
18:54 glehmann: because it doesn't handle int -> fp convert, only fp -> fp and int -> int
19:45 alyssa: karolherbst: have you seen test_api tests hang when run with your python runner but not in isolation?
19:45 karolherbst: alyssa: not here
19:45 karolherbst: any test specifically that hangs?
19:46 alyssa: /home/alyssa/CLCTS/OpenCL-CTS/build/test_conformance/api/test_api get_kernel_arg_info --disable-spirv-validation --compilation-mode online --compilation-cache-path /tmp/tmprn3cx0m1
19:46 alyssa: just started today, not sure if one of my commits broke something or what
19:46 karolherbst: huh.. why would that even hang...
19:46 alyssa: ps -aux shows the process as S
19:46 karolherbst: anything it waits on?
19:46 karolherbst: ohhhh wait.. mhh
19:47 alyssa: feels like a file system deadlock or something
19:47 karolherbst: alyssa: mind checking if https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42319/diffs?commit_id=cfafa79447699ef4cb4e49cb5c584aefbce369ce broke it?
19:47 karolherbst: though that would be _very_ concerning
19:48 karolherbst: alyssa: uhhhh... possibly? have you "sudo gdb attach"ed into the process?
19:48 alyssa: karolherbst: reverting that fixes it, yes
19:48 karolherbst: 😭
19:49 alyssa: sorry karol :(
19:49 karolherbst: that's going to be funky to debug, especially as we'll hit 40ºC this week so my brain is going to be toasted
19:51 karolherbst: I'm wondering if it's some ref counting fuckery going odd
20:20 karolherbst: glehmann: mhh yeah that's probably it.. I wonder if we should just bail if we detect op has a different source type but... not quite sure
20:22 karolherbst: doesn't feel very safe what the opt is doing there even outside of conversions
20:37 karolherbst: anyway.. I think I got a fix
21:35 karolherbst: glehmann: you think this looks correct? https://gitlab.freedesktop.org/karolherbst/mesa/-/commit/e7111bfecdc9c705f2405f729bbb37ed6b50083e
21:36 karolherbst: wondering if I need to insert some float -> half round trips for extra correctness there..
21:49 glehmann: you can't just cast the float to int, that's UB for NaN or too large values like inf
21:51 glehmann: the code that actually converts the constant is also wrong
21:55 HdkR: float->int overflow/underflow behaviour is also different between ARM and x86. So you can get some wildly different results if you're aiming for reproducibility :)
22:21 karolherbst: mhhh yeah...
22:22 karolherbst: it also doesn't handle conversions to 8 bit types at all...
22:23 karolherbst: maybe better to bail for now until somebody is willing to spend the time to actually fix this all up
22:25 karolherbst: ohh it actually checks for 16 bits further up
23:57 alyssa: karolherbst: adl-cl is blocking merges, presumably that commit is guilty
23:57 alyssa: I'm revertint
23:58 alyssa: revert cherrypicked into https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/42287
23:59 alyssa: which is next in queue