00:29 tarceri: To trigger the issue we need get si_set_sampler_views to free the view twice ... the fix stops us accidentally reusing a sampler view but even if we reused the view it should get referenced twice by the backend so its wrong but it should cause the double free crash?
00:30 tarceri: Seems like there must be something else freeing the view before we call the update?
00:31 tarceri: *we need to get si_set_sampler_views to free already freed view
00:33 zmike: the view doesn't hit the backend before the call to create views completes
00:34 zmike: so it only has one ref and then it gets released and deletes
00:35 tarceri: yesh but my point it the list we send to the backend just ends up with 2 entries pointing to the same view right?
00:36 zmike: no
00:36 zmike: in this case one is srgb and one is not srgb
00:36 zmike: I assume one is txf and the other is not
00:39 zmike: st_get_sampler_views() creates multiple views, and two of them are for the same texobj
00:40 zmike: so inside the st_get_sampler_views() loop, creating the second view releases the first one
00:48 tarceri: Right I see, diffcult to follow as there is multple litst going on here. the text obj sampler_views list gets the view overridden and one in get_sampler_view ends up with a freed pointer as your saying
01:03 zmike: yes
01:55 tarceri: zmike: I managed to recreate the bug but patch 4 of your series changes mesa behavior
01:55 zmike: 🤔
02:10 tarceri: Looks like nothing in piglit will hit it and nether does the cts test that the param was originally added for
02:15 zmike: I don't see how it could change behavior when all that patch does is calculate the format earlier to allow potentially reusing the view
02:17 tarceri: https://paste.centos.org/view/b3cdb870
02:19 zmike: sure, but the formats for those should still be different
02:19 zmike: so it should produce different views
02:20 tarceri: After patch 4 it just returns a match
02:20 zmike: hmm I think I'm too tired to see it now
02:20 zmike: maybe make a piglit MR and I'll check tomorrow
02:21 tarceri: will do :)
02:21 zmike: 🤝
05:23 tarceri: zmike: I think its ok my piglit test was using an RGB format not a sRGB previously it would create a new view for this case even though its not really necessary
05:24 tarceri: *previously mesa would
11:40 zmike: tarceri: yeah that's the idea...now it can reuse the views in some cases
12:01 tarceri: yeah makes perfect sensed now that I understand the code :P
12:03 zmike: I'm not gonna say it's good code
12:03 zmike: but it is code
12:04 tarceri: Well at least its got test coverage now...if only piglit ci wasn't busted
12:07 mahkoh: Re https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/40382, I was under the impression that amdgpu always used implicit sync except for vulkan-internal buffers that are not shared. But it doesn't seem to be that way. Could this be another bug?
12:07 zmike: piglit ci seems broken a lot
12:07 zmike: which is strange because piglit isn't that active
12:09 tarceri: seems its having python numpy issues, maybe pulls in updates automatically?? Anyway bed time maybe a magic elf will have fixed it by the time I wake
12:10 zmike: valentine sergi mupuf ^
12:11 mupuf: hmm, do you have an example job that failed?
12:12 tarceri: https://gitlab.freedesktop.org/mesa/piglit/-/jobs/95161530
12:14 mupuf: tarceri: thanks!
12:15 mupuf: So.... numpy changed its interface... or has a different interface depending on the python version?
12:15 mupuf: should be easy to reproduce locally
12:15 tarceri: mupuf: This is code from main not mine
12:16 mupuf: lovely
12:17 mupuf: so, unpinned dependency, quite likely
12:17 mupuf:is taking a look
12:18 mupuf: numpy2: numpy>2.0
12:19 mupuf: https://github.com/numpy/numpy/pull/29994 <-- this is what broke it
12:21 tarceri: Use the ``shape`` parameter instead
12:28 mupuf: yes, and there is already code in piglit to do that
12:32 mupuf: Meh, please contact the original author of the code: https://gitlab.freedesktop.org/mesa/piglit/-/merge_requests/1063
12:34 mupuf: maybe dropping support for older versions of numpy would be less messy
12:42 tarceri: ok, that was my bad I didn't have that pulled in
12:44 tarceri: Thank you for helping me with my stupid
12:44 mupuf: hehe, you're welcome
14:01 lynxeye: zmike: Suppose one would like to port ARB_texture_swizzle to GLES2. Would that be a MESA_ extension or is there still anyone around at Khronos to give this the OES_ stamp of approval?
14:02 zmike: I think you could make a PR in the registry, make a CTS PR to enable the tests for ES2, and roll the dice
14:29 zamundaaa[m]: Umm, do I see correctly, that Mesa caches the result of
14:29 zamundaaa[m]: vkEnumeratePhysicalDevices per VkInstance?
14:30 zamundaaa[m]: It's a pretty annoying problem when working on a compositor that deals with GPU
14:30 zamundaaa[m]: hotplugs...
17:00 Lilla: agd5f: Is BACO or mode1 reset needed for hot unbind/bind?