00:00 mangodev[d]: cubanismo[d]: seems about right
00:00 mangodev[d]: why are corporates so staunch on Microsoft© technologies?
00:32 cubanismo[d]: mangodev[d]: Hehe, no this was at my university's graphics lab.
00:38 cubanismo[d]: He was a really great guy, just got used to working in Windows I guess. It didn't matter much for my stuff. They wanted to combine some previously developed algorithms to make a 3D object morphing thing, so they handed me a stack of papers and told me to implement the ones I thought would work best. I just wrote it all in a little GLUT testbed on Linux at home, and then cut/pasted my code into his
00:38 cubanismo[d]: project.
02:28 nathangreen: This position natural to me isn't playable after my injuries the least. And terrorists cling on to me as hardly noticable fakes (people need to inspect under the hood a lot to tell that you are idiots, and not everyone has time to do it, you are very hardly noticed downies in other words, only looking at you does not expose the idiot state) telling me about their supremacy, If i was
02:28 nathangreen: physically fit like nature forwarded before the terror events, such mindless fake artists or idiots would not be allowed to even get compared to me, before they get already offput, cause the gulf between us is as large as your seemingly hidden delusions in your head. I have no big clue as to what to do, your lives have no worth of any kind as I first time tell this being myself treated
02:28 nathangreen: decades this way is the words that dejavu in my head spoken to me every time i broke through the barriers and quasimodos immediately terrored me. But my friends and alliance have lives that have worth, and going into war with people, who have no motivation or purpose to live and no worth, is a danger to us, placing sanctions would not work either.
10:00 karolherbst[d]: uhhhhhhh
10:00 karolherbst[d]: I just realized something about this shader...
10:00 karolherbst[d]: evil
10:00 karolherbst[d]: so each iterates only operates on a slice of shared memory
10:02 karolherbst[d]: of course it's a big array in the shader..
11:22 karolherbst[d]: okay, cleaned up my `nir_opt_offset` stuff: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36113
11:23 karolherbst[d]: maybe I should hack and prototype it for global/scratch just to see how big the impact is
12:22 skraitow: hi all skraito ( God Husband , The Earthquake guy ) your president and prime minister with skraitow ( Lord Jesus Christ skraito wife ) here , how are you all ? ... .
13:51 karolherbst[d]: so now with the fixed stats, maybe `nir_opt_licm` actually looks good in the stats aas well 😄
13:54 karolherbst[d]: oooooofffff there are some shaders with real akward stat regressions
13:55 karolherbst[d]: but with the logarithmic scaling, there are shaders with a cut cycle count of over 60%
13:55 karolherbst[d]: I wonder what's going on there.. could get confused about ldc?
14:25 karolherbst[d]: I'm cooking again: `Max warps/SM: 911328 -> 917960 (+0.73%); split: +0.74%, -0.01%`
14:38 karolherbst[d]: `Spills to reg: 67707 -> 57646 (-14.86%); split: -15.24%, +0.38%` 😮 oh wow
14:39 karolherbst[d]: that's for comparisons.. might even help with brau
14:41 karolherbst[d]: also cuts down on spilling a lot: `Spills to memory: 57317 -> 54698 (-4.57%); split: -4.57%, +0.00%`
14:41 karolherbst[d]: but gpr usage increases 🙃 `Number of GPRs: 3863818 -> 3866731 (+0.08%); split: -0.01%, +0.08%`
14:42 karolherbst[d]: `Static cycle count: 1073101189 -> 1059404451 (-1.28%); split: -1.39%, +0.11%`
14:42 karolherbst[d]: I should check if I can mitigate the regressions
14:42 karolherbst[d]: this is nir_opt_move with nir_move_comparisons
14:43 karolherbst[d]: `nir_opt_move, nir_move_comparisons | nir_move_load_ubo`: https://gist.github.com/karolherbst/b7779deb88c7658b0e94d081091c668d
14:54 karolherbst[d]: mhhhhhhhh
14:55 karolherbst[d]: %p394 = fsetp.ge.ftz %r393 %r360
14:55 karolherbst[d]: %r395 = sel %p394 %r369 %r386
14:55 karolherbst[d]: they got rid of slct, didn't they
15:04 karolherbst[d]: okay.. so it increases GPRs in the case where the comparison sinks a register... mhh
15:16 gfxstrand[d]: https://cdn.discordapp.com/attachments/1034184951790305330/1401222126710820935/rn_image_picker_lib_temp_aa19e820-4711-44a1-bc38-da7e3a768f25.jpg?ex=688f7d4c&is=688e2bcc&hm=ae388c3aa637a27475679aad695d5b1d6dd008283716a46cc5c737c98770b8e6&
15:16 gfxstrand[d]: Hopefully this little guy will be able to boot and enumerate a GPU.
15:21 karolherbst[d]: mhhh
15:21 mohamexiety[d]: it will try its best
15:21 karolherbst[d]: nir_move_comparisons on convergent ones only does help a lot
15:23 karolherbst[d]: but overall on all it's still better.. whatever then
15:24 karolherbst[d]: seems to hurt more in smaller shaders, but complex ones get huge benefits (a lot less spilling)
15:26 karolherbst[d]: I have one of those little fellas as well!
15:26 karolherbst[d]: the heatsink can become seriously hot 🙃
15:27 karolherbst[d]: next: `nir_opt_sink`
15:53 karolherbst[d]: at some point I'm gonna figure out why the hell `nir_move_const_undef` makes a difference, but 😄
15:56 gfxstrand[d]: karolherbst[d]: Our undef handling sucks
15:56 karolherbst[d]: I see
15:56 karolherbst[d]: well it changes like 0.2% of the shaders, so not gonna bother either way
15:57 gfxstrand[d]: Yeah, normally there aren't many of them in a correct shader
15:57 gfxstrand[d]: And the most common occurrence are places where we just need something to satisfy NIR constraints and those get DCEd
15:58 karolherbst[d]: right
15:58 gfxstrand[d]: And they sometimes show up at the tops of loops.
15:58 x512[m]: gfxstrand[d]: Do Nvidia have some image format restrictions for screen framebuffer?
15:59 karolherbst[d]: mhh there is a `nir_lower_undef_to_zero` in the loop, so no wonder it barely does anything
15:59 gfxstrand[d]: Yeah. That'll do it
16:02 karolherbst[d]: I wonder if moving inputs makes a difference..
16:04 karolherbst[d]: anyway, if somebody wants to review a very very very trivial thing: https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/36514 it's basically free performance with 0 drawbacks
16:10 karolherbst[d]: I don't manage to make `nir_opt_licm` look good.. it's kinda annoying tbh 😄
19:03 karolherbst[d]: or rather packaging bug
19:04 karolherbst[d]: ehh wrong channel 🙃