19:05 mupuf: Lyude, karolherbst: first run of IGT done on amd gpus, using the new infra I am making which will hopefully be used by nouveau soon too: https://pastebin.com/raw/g6CyePWS
19:06 karolherbst: nice
20:50 Wolf480pl: Hello. Last time I tried to use power management on my GK104M with nouveau, there was an issue that nobody knew what to set clocks to for optimal performance. Is this still an issue? If so, would clocks observed on the same card on Windows when gaming be helpful?
20:51 RSpliet: Wolf480pl: you'll have to be more specific about what's going wrong. Nouveau can currently set clocks, but isn't very good at "boost" I don't think
20:51 RSpliet: check /sys/kernel/debug/dri/0/pstate for the clocks you can ste
20:52 Wolf480pl: 0 is i915 :P
20:52 RSpliet: oh, yeah, w/e, go for 1
20:52 RSpliet: and make it do something
20:52 RSpliet: otherwise it's powered down
20:52 RSpliet: (glxgears would do fine)
20:53 imirkin: RSpliet: we do allow boost, but you have to turn it on yourself
20:54 RSpliet: imirkin: yes, I do remember that, but I'm not sure if that does the full "vary your voltage(/clock?) based on temperature" thing
20:54 Wolf480pl: ok, so I see a bunch of clocks, https://gist.github.com/Wolf480pl/4c506a4eb6637cea5d524d875ed77198
20:55 RSpliet: you should be able to set them. Just write to the pstate file with the identifier you see at the start (you can omit the leading 0)
20:55 RSpliet: not sure if you'll need to set e before f, or what other reason there is for both modes
20:55 imirkin: RSpliet: it does.
20:55 imirkin: (i think)
20:55 RSpliet: I think karolherbst is the expert on that bit :-P
20:55 imirkin: he is.
20:55 Wolf480pl: btw. maybe I'll try glxspheres instead of glxgears so that the fps value is more meaningful
20:56 imirkin: glxspheres is a good measure of memory bandwidth iirc
20:56 karolherbst: Wolf480pl: use benchmarks, not pseudo tools :p
20:56 karolherbst: but yeah..glxshpheres is better already
20:58 Wolf480pl: getting ~1000 Mpix/s with 0f
20:59 Wolf480pl: so yeah, there is quite noticable difference from 07
20:59 imirkin: with memory going from 648mhz to 5ghz, i should hope so
21:00 Wolf480pl: IIRC a few years ago when I tried setting pstate it didn't really cause much noticable improvement because the clocks read from vbios were all wrong, or sth, so I guess it's fixed now, nice
21:00 imirkin: various issues have been fixed over the years
21:00 imirkin: i remember your handle, but not the specific issue you were having
21:01 Wolf480pl: you guys are the best, thank you
21:02 Wolf480pl: gonna look for a real benchmark now
21:02 imirkin: unigine heaven is a reasonable one
21:18 Wolf480pl: woo, bad RIP value in gk104_fifo_{gpfifo_engine_fini,runlist_update}
21:20 imirkin: too much speed
21:22 Wolf480pl: hm... I wonder if the mem clock in pstates is in the same unit as the one reported by OpenHardwareMonitor on Windows, cause that one was like, 2497 MHz under load
21:23 imirkin: they don't call it DDR for nothing
21:23 imirkin: *double* data rate :)
21:23 imirkin: (or dance dance revolution, depending on context)
21:24 Wolf480pl: oh, so pstate reports transfer rate, not frequency on the clock wire, I see
21:24 imirkin: units are confusing
21:42 Wolf480pl: trying on 0a now, and heaven (on medium, 1600x900) feels like it's slowed down despite having 100-ish fps
21:44 Wolf480pl: btw. if heaven is rendering on the nvidiacard, but Xorg is running on intel, but the monitor is connected to nvidia, will the frame be copied twice?
21:44 imirkin: yes.
21:45 Wolf480pl: so I'd be better off using the laptop screen?
21:46 imirkin: not better or worse
21:46 imirkin: if you want max perf when using the second gpu, you could use the second gpu as the primary
21:46 imirkin: but that would mean restart X
21:46 imirkin: (or start a second X with the different config)
21:46 imirkin: and then switch between vt's
21:49 Wolf480pl: pstate: 0a, heaven: medium, 1600x900, no tesselation, score: 1858, min fps: 9.6, max fps: 136
22:02 Wolf480pl: pstate 07, score: 175, min fps: 4.8, max fps: 12.7
22:03 Wolf480pl: but it finished much quicker... is this normal?
22:04 Wolf480pl: trying 0e again, hopefully it doesn't crash
22:09 Wolf480pl: hanged, had to REISUB
22:47 RSpliet: Wolf480pl: some benchmarks try to get through all its frames, others mimick games a bit better and skip frames to make the run-time as equal as possible.