00:39 sghuge: zmike: any pointers on how to reproduce the failure https://gitlab.freedesktop.org/mesa/mesa/-/jobs/106155133 locally? Thanks.
00:44 zmike: glretrace the failing trace I guess? I've been meaning to look at those trace jobs for a while because they're so flaky
00:45 zmike: sghuge: I don't really have any tips other than try to copy the job env vars
00:47 sghuge: Is this a known failure? I thought it's triggering becuase of my MR :(
01:00 zmike: probably just retry
10:15 valentine: yeah, looks like one of those zink flakes again
10:15 valentine: and leaving this here for visibility: the traces are hosted on https://gitlab.freedesktop.org/gfx-ci/tracie/traces-db and they can be replayed with eglretrace from apitrace
10:51 simon-perretta-img: Would anyone happen to know whenabouts the "Attending XDC" page that usually has recommendations for accommodation, links to Matrix, etc. would typically go up?
15:51 DottorLeo: Hi!
15:51 DottorLeo: @eric_engestrom https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/43301 shouldn't be for 26.1.6 release? ^^"
16:01 alyssa: nir/lcssa: allow to create LCSSA phis for loop-invariant booleans
16:01 alyssa:
16:01 alyssa: ACO depends on LCSSA phis for divergent booleans to work correctly.
16:01 alyssa: pendingchaos: I found this commit... is there an explanation somewhere on why this is required?
16:01 alyssa: (I looked in both NIR and ACO but didn't see anything obvious)
16:05 pendingchaos: booleans are subgroup-uniform bitmasks, so we need to combine the bitmask in the loop at each break to use it outside the loop
16:05 pendingchaos: this is done with a phi
16:05 pendingchaos: it's needed even if it's loop-invariant because the only valid bits in the bitmask in the loop are for the invocation which break
16:07 pendingchaos: "for (i = 0;; i++) { if (i == invocation_id) {val = expression which is always true; break} }": for the i=0 break, val might just be the mask 01, and 10 for the i=1 break, 100 for the i=2 break and so on
16:07 pendingchaos: uses of "val" outside the loop want it to be ...111
16:20 alyssa: pendingchaos: ahhhhhh, I gotcha
16:20 alyssa: thanks, that example helps
16:21 alyssa: & I think that justifies Jay needing the same behaviour
16:42 glehmann: I have constants defined in loops
16:42 glehmann: would anyone complain if I change lcssa lowering to rematerialize constants instead of creating phis?
16:44 zmike: mattst88: I need waffle help
16:51 alyssa: glehmann: seems reasonable
16:52 glehmann: or we could move the constant before the loop
16:52 glehmann: that would be even easier
16:52 alyssa: pendingchaos: do I understand correctly that only divergent loop-invariant booleans are a problem?
16:52 alyssa: and convergent loop-invariant booleans do not need the phi?
16:53 alyssa: (not that that helps when calculating divergence requires lcssa in the first place)
16:54 pendingchaos: I guess it depends on how you implement convergent booleans
16:54 pendingchaos: ACO always uses the divergent representation for convergent defs and optimizes it later, so I guess it always needs the phi
16:55 alyssa: sure
16:55 alyssa: jay effectively does 0/~0 for convergent booleans
16:56 alyssa: (since that intersects both the divergent representation and the d3d representation)
16:56 alyssa: that's really natural on Intel, i can imagine it's less so on AMD
16:56 karolherbst: glehmann: yeah.. for NAK defining the constants prior the loop should be better
16:57 glehmann: alyssa: I think we mostly don't care enough about convergent bools to deal with the complexity of having two representations
16:58 alyssa: glehmann: fair enough
17:05 glehmann: I hope this doesn't blow up anyone's live ranges: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/43308
17:09 alyssa: on that topic
17:10 alyssa: would anyone oppose a knob for lcssa to insert only for divergent 1-bit (regardless of invariance)?
17:10 alyssa: or at that point should I just have my own jay_nir_fixup_bools
17:12 karolherbst: let me check that MR on NVK
17:18 glehmann: alyssa: didn't you just say that it doesn't work because you need lcssa for divergence?
17:23 alyssa: glehmann: first call normal lcssa, then run divergence, then call special lcssa
17:28 glehmann: fine by me as long as you change the options to flags instead of adding another bool param
17:28 alyssa: ack
20:13 civet: Hi. Just tried intel crocus driver with Gen 5 card. It doesn't change much it's still limited to opengl 2.1 right?
20:15 airlied: the hw is limited
20:18 civet: Yes. But I've read somewhere that switching to crocus can bump supported api versions. Apparently not with this hw
20:24 airlied: no that hw is barely able to do gl2.1
20:42 karolherbst: heftig: are you the one packaging libclc for Arch Linus? Mind adding this PR to it? https://github.com/darktable-org/darktable/issues/21662 It fixes compilation issues with OpenCL C due to a missing function definition inside libclc