07:04 <macc24> are context menus in firefox supposed to flicker like crazy?

07:12 <icecream95> macc24: It's improved with sway master, but Firefox is still a bit flickery for me

07:13 <macc24> icecream95: i don't remember seeing the flicker like few weeks ago

07:14 <macc24> happens on kevin too

07:56 <macc24> is there a way to check what a program does that kills the performnace?

07:56 <macc24> performance*

08:07 <tomeu> macc24: first step is to figure out where is the bottleneck: https://www.collabora.com/news-and-blog/blog/2020/08/21/panfrost-performance-counters-with-perfetto/

08:42 <macc24> ok

08:42 <macc24> tomeu: i'm pretty sure the bottleneck is not cpu or cci

08:43 <macc24> especially since the program runs equally bad on both g72 and t860

10:01 <macc24> am i allowed to complain that ms teams video doesn't work on g72?

10:02 <macc24> nothing just black screen

10:04 <macc24> in chromium*

10:07 <macc24> and it refuses to work in firefox... https://i.imgur.com/1EqVbhv.png

10:34 <urjaman> yeah it just doesnt do firefox (even on on windows)

14:09 <bbrezillon> alyssa: so, I have a TB preload implemention using pre-frame DCDs on Bifrost and a tile-enable map on Midgard here https://gitlab.freedesktop.org/bbrezillon/mesa/-/commits/panfrost-tb-preload

14:10 <bbrezillon> but before I go and split/cleanup the last commit I'd like to know how you feel about using a hash table for the blit shader cache

14:11 <bbrezillon> if we want to use the pre-frame mechanism, we need to have single shader that reloads all RTs, and another one that reloads the Z and S components

14:13 <bbrezillon> that means the 3-dimensional table used to store the blit shaders no longer works for those complex case (the logic was assuming one shader per RT)

14:17 <bbrezillon> in the implem I have right now I got rid of this static table and forced lazy creation for all blit shaders, but I could go back to a mixed solution where single RT shaders are pre-allocated and only multi-RT shaders are kept in the hash table, I'm just not sure it's worth the pain

14:32 <alyssa> eyes

14:34 <alyssa> bbrezillon: Maybe it would make sense to use the hash table solution, but on init prepopulate the table with the most common cases (reload of RT 0 unorm only, etc)?

14:35 <alyssa> I wouldn't worry about perf impact until data proves otherwise.

14:35 <alyssa> Hash tables suck for hot paths but.. meh

14:36 <bbrezillon> alyssa: sure, I can do that

14:36 <alyssa> bbrezillon: I would worry more about the impact of the change on GPU perf

14:37 <alyssa> (tile enable map vs rectangle silliness, etc)

14:37 <alyssa> actually i suspect if you did it right, this way is faster

14:37 <bbrezillon> yep, that's on my TODO list

14:37 <alyssa> but would like some numbers... not really sure how to measure perf on 2D workloads though

14:38 <bbrezillon> though I don't know what to use to trigger heavy TB preload

14:38 <alyssa> UI stuff

14:39 <alyssa> weston exercises all these code paths

14:41 <bbrezillon> yep, that's what I'm using to test KHR_partial_update, but I thought I'd need something more GPU intensive

14:42 <bbrezillon> maybe I can get some numbers from the perf counters

14:42 <bbrezillon> *useful numbers

14:43 <alyssa> I mean... KHR_partial_update is not designed for GPU intensive stuff, it's to save power.

14:43 <alyssa> if a 2D UI is GPU intensive, you're doing something horribly wrong

14:43 <alyssa> or running Compiz on Mali

14:43 <alyssa> :P

16:48 <macc24> >alyssa │ if a 2D UI is GPU intensive, you're doing something horribly wrong

16:53 <alyssa> sigh

16:53 <chrisf> intrinsically expensive, or just murdering tilers for no reason?

16:53 <chrisf> i know there's a couple special corners in XRender

16:54 <macc24> chrisf: uhh

21:46 <robclark> iirc glamor scissors out untouched areas.. freedreno tracks maximal scissor bounds to avoid tile-pass over stuff that is untouched (although we don't even bother using that on a6xx, but instead just let the CP skip tiles with no visible geometry..

21:47 <robclark> (and generally I've not seen murdered tiler syndrome with glamor)

21:51 <alyssa> robclark: urxvt at large window sizes with lots of text scrolling is the primary cause of MTS

21:53 <robclark> hmm, ok.. I confess to not being a urxvt user.. but I don't think anyone has complained about it.. I suppose if blend is not used things fall back to bypass path

21:55 <robclark> alyssa: a shot in the dark, panfrost_set_framebuffer_state() might want to be introduced to `util_framebuffer_state_equal()`..

21:56 <robclark> (but that could date back to the dark ages when we used xa state tracker..)

22:17 <robclark> alyssa, bbrezillon: another thought if you are looking to measure glamor perf.. you can apitrace glamor with urxvt (or whatever).. and then replay the trace to profile things.. https://github.com/apitrace/apitrace/wiki/Glamor

22:52 <alyssa> glamor apitrace? what is this witchcraft?

22:53 <macc24> robclark: does glamor run with some performance on freedreno?

22:54 <alyssa> Adreno is, as a rule, less dumb than certain other tilers..

22:54 <macc24> let me guess, powervr is worst?

22:55 <robclark> macc24: yeah, I've been pretty happy with glamor perf.. even back to a3xx.. adreno does have one trick, in that it can bypass tiled rendering and work like an immediate mode renderer in some cases (although that is a good and bad.. because sometimes deciding which route to take isn't straightforward)

22:55 <chrisf> well, less of a tiler than other tilers :)

22:55 <macc24> robclark: >work like immediate mode renderer

22:55 <macc24> wtf how

22:56 <robclark> bypass mode.. I think qcom has some marketing name for it.. "flex-something-or-other"..

22:56 <macc24> huh, is there a way to manually override this?

22:58 <robclark> FD_MESA_DEBUG=nogmem or FD_MESA_DEBUG=nobypass

22:58 <macc24> >:D

22:58 <robclark> there are some things (depending on generation) which only work in one mode or another, so expect debug overrides to break *something* ;-)

22:59 <alyssa> as usual, setting *_MESA_DEBUG voids your nonexistent warranty ;)

23:01 <macc24> alyssa: for a future reference, if i'm complaning about something i don't have PAN_MESA_DEBUG set

23:01 <macc24> s/complaining about something/reporting a bug/

23:01 <alyssa> thanks

23:03 <robclark> alyssa: hmm, I wonder if it would be a good idea to be able to inject something into GL_VERSION or GL_RENDERER when various different debug/override envvars are set..

23:03 <macc24> why are you thanking me? are users setting random env vars that they thing might help?

23:03 <alyssa> macc24: yes.

23:03 <macc24> lol