01:49 Lyude: bisected the cursor flickering on kepler down to 5bca1621c07c3ad37b5a4943450a892e18984df0
02:08 imirkin: huh. interesting.
02:15 imirkin: Lyude: are you getting any other errors in dmesg?
13:11 pendingchaos: karolherbst: would pointing to envytools be sufficient for documenting xmad? the PR documenting it hasn't yet been merged yet though
13:12 imirkin: pointing from where?
13:12 karolherbst: imirkin: nv50_ir.h
13:13 karolherbst: where the xmad subops would be added
13:13 imirkin: probably fine
13:13 karolherbst: yeah, should be good enough
13:13 pendingchaos:nods
13:14 pendingchaos: mwk: ping on the xmad and iadd3 PR for envytools?
13:15 mwk: pendingchaos: I have nothing to do with gm107 ISA
13:15 pendingchaos: ah, I was told that you were generally the envytools maintainer
13:15 mwk: so at most I can check RST correctness
13:15 imirkin: mwk: i was hoping you'd approve the general approach and location in the overall document
13:16 imirkin: (or, conversely, suggest alternatives)
13:16 mwk: yeah, sounds about right
13:17 mwk: there, done
13:23 pendingchaos: mwk: I think the errors have been fixed
13:25 mwk: k, ship it
13:30 pendingchaos: I don't have commit access, in case you were expecting me to merge it
17:48 kubast2: Did nvidia deliver reclocking firmware for 9XX series and perhaps 10XX series ?
17:49 RSpliet: to whom?
17:52 imirkin_: they have not delivered any power management firmware other than the baseline required for getting any sort of accel (since the pmu actually has to boot up the graphics engine)
18:14 HdkR: It's such a shame too :/
18:19 imirkin_: apparently not... we suck more than i remember
18:19 imirkin_: https://www.phoronix.com/scan.php?page=article&item=nouveau-summer-2018&num=2
18:19 imirkin_: i remember bioshock/dota2 sucking, but that's a bit extreme
18:19 imirkin_: i bet we're just stalling on something due to stupidity
18:20 imirkin_: fwiw radeonsi also sucks on dota2
18:20 imirkin_: (not as hard as we do though)
18:38 RSpliet: imirkin_: Judging by the kernel cmdline, I don't think Michael used the full boosted clocks. Can't blame him, nouveau boost mechanisms doesn't win it a usability prize. But he might be dropping some performance on the floor there.
18:39 imirkin_: like 10% though. not like 50% :)
18:39 HdkR: imirkin_: Need that xmad optimization in. Fix all the perf issues :P
18:39 imirkin_: HdkR: esp on kepler...
18:39 HdkR: Does xmad exist on kepler?
18:39 imirkin_: no
18:39 HdkR: Thought so but I'm not super familiar with that arch :P
18:40 imirkin_: you might not know this, but every so often, i say things tongue-in-cheek...
18:40 RSpliet: On GT640 my instruction sched hackery branch makes a visible difference in at least Unigine
18:40 HdkR: hehe
18:40 RSpliet: Not so much on the 940M in my laptop... this is definitely not close to being ready for prime time
18:50 pmoreau: I wonder how much hierarchical-Z would bring us, especially in games with a lot of depth complexity; IIRC, Nouveau still does not enable it, right?
18:51 imirkin_: correct.
18:51 imirkin_: (ZCULL in nvidia-speak)
18:51 pmoreau: Ah, right :-)
19:16 RSpliet: think I heard someone say a long time ago that ZCULL mainly benefits the low end GPUs
19:37 karolherbst: RSpliet: I don't see how that should make a difference
19:37 karolherbst: same amount of work is skipped on all hardware, or at least should be
19:38 imirkin_: the lower the memory bandwidth, the higher the impact
19:38 karolherbst: mhh, why does it depend on memory bandwidth?
19:38 imirkin_: my guess, though, is that it benefits *more* on low-end gpu's
19:38 imirkin_: that's not to say it doesn't benefit on high-end ones
19:42 pmoreau: karolherbst: zcull/hierarchical-z/etc. reduces the amount of bandwidth used when reading/writing to the depth buffer. So if bandwidth was a scarce resource, freeing some will have a higher impact than if you had and barely used much of it.
19:43 pmoreau: *if you had a lot of bandwidth and barely used much of it
19:43 karolherbst: pmoreau: okay sure, but afterall it also reduced the amount of vertex/fragment shader invocations, no?
19:44 pmoreau: I don’t think it will change the number of fragment shader invocations, because that will be already reduced by an early Z.
19:45 pmoreau: As for vertex shaders, you probably still need to run them to know where that vertex is going to end up on the screen.
19:46 karolherbst: pmoreau: okay, if zcull is just a speed up for early Z by rejecting things faster, then yeah, I see the point, but I think there was some rejection thing before vertex shaders
19:47 karolherbst: or maybe I am wrong
19:48 pmoreau: It kinda rings a bell of having something before the vertex shader, but I don’t see how that’s possible without using black magic. :-D
19:49 pmoreau: zcull will be a speed-up, and will save power as you’re moving fewer bits around.
19:49 karolherbst: yeah, sure