00:23jenatali:doesn't understand lower_mem_access_bit_sizes
00:24jenatali: Oh well, that's a job for Monday I guess
00:27DavidHeidelberg[m]: zmike: yeah, I shouldn't close that quickly. Thanks for reopening
00:29zmike: DavidHeidelberg[m]: I just got home and checked and it seems like another LLVM problem
00:29zmike: so probably it'll never get fixed
00:29zmike: would recommend adding to flakes
00:29DavidHeidelberg[m]: zmike: yeah, the different LLVM makes a huge impact I guess
00:30DavidHeidelberg[m]: zmike: not enough interesting to get it workarounded? btw. few days ago it had same LLVM, so these are new issues.
00:30DavidHeidelberg[m]: and Debian 12 is in hard freeze, so I don't expect changes
00:31zmike: I'm not LLVM expert enough to fix it, which leaves very few others within the venn diagram region of "cares about ci" and "knowledgeable about llvmpipe internals"
00:32zmike: see also all the other llvmpipe tickets open
00:34DavidHeidelberg[m]: sure, I'll add it into flakes before the merge, if it won't get tackled :)
00:57airlied: mlankhorst: please dequeue fixes next week
01:00airlied: also you don't have to wait until Friday, happy to get pulls any day you have time
01:20DemiMarie: Should Mesa bundle its own LLVM?
01:20DemiMarie: And tell distros that if they use any other LLVM, they are on their own for support?
09:53MrCooper: karolherbst: colour me skeptical that tiling between layers of a 3D image makes a significant difference for performance
09:55MrCooper: DemiMarie: the likely effect of that would be distros taking more time for updating to newer Mesa
10:42ccr: I don't think that would fix ABI issues anyway, for example re: https://airlied.blogspot.com/2023/04/fedora-38-llvm-vs-team-fortress-2-tf2.html
10:53karolherbst: MrCooper: well.. Nvidia rather detiles on demand by using a shadow buffer than not using tiling at all
10:53karolherbst: *z-tiling
10:54karolherbst: it's all about access patterns and for 3D the access patterns are supposed to be 3 dimensional
10:54karolherbst: if it wouldn't make any significant difference, why do we bother with tiling at all?
18:58karolherbst: jenatali: any opinion on flushing whole queues or only flushing requested event items (and their dependencies, including implicit ones like for in order queues)?
18:59jenatali: karolherbst: I just flush whole queues
19:00karolherbst: yeah.. I'm wondering if I want to do the same, because it makes the implenentation way simpler
19:03karolherbst: jenatali: though I think in the case of ordered queues the expectation is, that runtimes will reorder depending on what the application is waiting on
19:04karolherbst: but I don't even know if it matters at all
20:38DavidHeidelberg[m]: looking for review: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/22813 reviews welcome, it's kidnda INT_MAX + float fun
20:39DavidHeidelberg[m]: magic is to do right +1 / -1 :D