08:53MrCooper: review of https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41845 would be appreciated, fixes an EGL GBM platform oopsie which is hitting mutter users
12:03lumag: mripard: sorry for being PITA, could you please reply to https://lore.kernel.org/dri-devel/vpd5hyote5wspmlpad64kf5peoy5g7wv6c7xjn6ammcmjtai7r@q2tarmr5aoqn/ ?
14:07mripard: lumag: done, sorry I completely missed your answer
16:54alyssa: glehmann: i'm still chewing on "nir/algebraic: Convert bcsel of addition to addition of"
16:54alyssa: if anything Intel would prefer the other way around since we can predicate the addition
16:54alyssa: in reality no Intel compilers on any OS currently do that and IDK if it's profitable for cycle count even tho it saves an instr
16:55alyssa: but I guess for common optimizations having a canonical form is good regardless
16:55alyssa: I think nvidia predication is similar to intel
16:57glehmann: I think nv has compares fused with b2i/b2f, so it could go either way
16:58glehmann: b + b2f(a) is also more fma friendly
16:58glehmann: and in aco we optimize b + b2i(a) to add_carry_in(b, 0, a)
16:59glehmann: but yeah, all of this kind of depends on backend details
17:00karolherbst: I think nv has compares fused with b2i/b2f, so it could go either way not anymore
17:01karolherbst: uhm.. I wnated to quote that
17:01karolherbst: ohh or maybe we still have? mhh
17:02karolherbst: ahh only compare and b2f
17:02karolherbst: the int one is gone
17:02Company: robertmader[m]: Can I get your blessings on https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/41776 ?
17:02glehmann: I was about to say, I clearly see the b2f version in nak :D
17:02karolherbst: FSET is kinda funky, because you can even chain an input bool
17:02karolherbst: so b2f((a <= b) || c) is one instruction
17:02Company: Also, do you know who the right person is to poke to get it reviewed and assigned to marge?
17:03karolherbst: but yeah.. we only have ISETP, so..
17:03karolherbst: that writes a predicate
17:04karolherbst: glehmann: but also.. NAK covers ... 13 years of hardware? 🙃
17:04karolherbst: I'm sure it supports hardware were we had this for ints still
17:06karolherbst: or maybe it was always float only...
17:07karolherbst: nahh.. it was the same instruction back then.. the code is confusing
17:27alyssa: glehmann: you'd think we could do the same trick with ADDC on Intel, but you'd be wrong (:
17:28alyssa: ADDC/SUBB on Intel are.. deeply cursed
17:28alyssa: mesa has never used them
17:29glehmann: I'm starting to think it's easy if you list the non cursed features
17:29alyssa: Oh uh
17:29alyssa: Um
17:30alyssa: Well there's Larrabee
17:33HdkR: alyssa: non-cursed hardware, not cursed hardware!
17:35karolherbst: okay.. I figured out that nir_opt_dead_write_vars became really slow and I got a fix: https://gitlab.freedesktop.org/karolherbst/mesa/-/commit/d7e888525d3fb7159d55c8edfb33cd7705f0351e
17:35karolherbst: speeds up some compilations by a factor of 3x
17:36karolherbst: I should prolly move it to an MR
17:39glehmann: the joy of linked lists
17:46karolherbst: will check if it also helps with fossils to get some better numbers than "this one compute benchmark does not compile for that long anymore" 🙃
17:56jani: hey, has anyone seen a panel failure mode where it rotates red/green/blue/black/white? feels like a test mode or something. really curious. video at https://gitlab.freedesktop.org/drm/i915/kernel/-/work_items/15922#note_3491200
17:58jani: also, heisenbug. starting screen recording stops the color rotation