12:00 <macc24> sooo xonotic on midgard runs and is playable at 1280x800 with lowest settings on panfrost

12:00 <macc24> on opengl 2.1

12:00 <HdkR> coolio

12:01 <macc24> yea

12:01 <macc24> bottleneck is on cpu

12:03 <HdkR> It would probably get worse with GL3.x as well if it is doing indexed draws. Since there is the minmax calculation

12:03 <macc24> can i test gl3 on panfrost without compiling mesa?

12:03 <HdkR> PAN_MESA_DEBUG=gl3

12:04 <HdkR> Looking forward to Valhall removing that overhead

12:04 <macc24> xonotic reports gl2.1 still

12:04 <macc24> mesa is at 20.1.5

12:06 <HdkR> Too old

12:06 <macc24> setting MESA_NO_ERROR to one helped with performance a bit

12:07 <HdkR> It's in 20.2.0-rc1 at least

12:10 <raster> oh wow... what changed... i no longer see scheduling evilness when you have a client rendering as fast as it can and wl compositor also fighting fore the cpu...

12:10 <raster> (midgard panfrost)

12:10 <raster> niiiiiiiiiiiiice

12:11 <raster> updated to 5.8 rc6 and latest git master mesa...

12:11 <raster> kernel change of some sort i assume... didnt see it

12:11 <macc24> huh?

12:11 <daniels> turning on afbc is even more turbo mode :)

12:12 <raster> ok - a few weeks back in my last update - i run a benchmark that has fairly simple rendering that is like rendering 200+ large alpha quads in 1080p as fast as it can

12:12 <raster> http://www.enlightenment.org/ss/e-5f328b4678e141.01774427.jpg <-

12:13 <raster> like that ... and basically it would, client-side, appear to get like 40fps... but visible only like 1-2fps as the compositor (wl compositor) couldn't get in enough gpu cycles to push out frames - so a gpu scheduling problem.

12:14 <HdkR> Nice, perfect bench to utilize tiled rendering

12:14 <raster> but now i get a good solid output that is "smooth" that represents the input

12:15 <raster> so something got fixed in scheduling between 5.7 and 5.8 :)

12:15 <raster> i've been busy with other things and not updating lately...

12:16 <raster> there was a mysterious bug with pae address translation i was trying to figure out that only happened in this benchmark - possibly because some command queue or something kept growing and growing... that was a guess. now the bug no longer happens too.

13:06 <robmur01> hmm, I would not expect to see anything from the pagetable code unless something was seriously broken, i.e. more than just poor scheduling

13:06 <robmur01> the pagetables should pretty much just be following the state of the DRM MM

16:44 <alyssa> '

16:44 <alyssa> ]|\']\[-

16:44 <alyssa> sorry

16:44 <alyssa> tomeu: bbrezillon: as discussed mainline panfrost has issue on rk3399 chromebooks due to... power? maybe?

16:44 <alyssa> I guess a patch never landed to fix that (?) -- do we know what's up there?

16:45 <alyssa> blocking AFBC stuff, since a single byte off means one wrong pixel with linear but the entire block corrupting with afbc

16:46 <alyssa> also robmur01 robher ^

16:47 <robmur01> patches to make regulators work properly again have just landed in drm-misc-next, I think

16:47 <alyssa> ack

16:47 <alyssa> oh, really 'just', ok

16:47 <urjaman> so.... 5.9 will work? :P

16:47 <alyssa> guess who gets to recompile the kernel again ;p

16:48 <robmur01> ...or dial back to 5.4 LTS :P

16:48 <alyssa> :P

16:48 <urjaman> i haven't moved anywhere from 5.4

16:48 <alyssa> robmur01: I'm trying to get AFBC-down-the-whole-stack since it ~~makes for better marketing~~ saves power.

16:49 <alyssa> and RK3399 display AFBC landed in like 5.8

16:52 <robmur01> TBH I'd imagine saving memory traffic and thus latency is probably even more significant than power for most of our "cheap TV box SoC" boards

16:53 <alyssa> yeah

16:53 <robmur01> an immediate alternative would be trying to revert 221bc77914cb, but it'll be messy

16:53 <alyssa> in the pathological case, something like glmark2-es2 -btexture jumps up hundreds of FPS from u-interleaved to AFBC

16:59 <robmur01> I shall have to try AFBC with rvgl - started playing with that recently and framerates are all over the shop for no obvious reason

17:00 <robmur01> (also some minor glitches, but I guess I'd have to futz with an apitrace to be allowed to complain...)