14:05 <sravn> nir versus tgsi. Are there any performance benefits when using nit if the target is running gc880 or gc2000?

14:06 <sravn> And how far are we from using nir per default?

14:06 <sravn> s/nit/nir/ in the above - sorry

14:09 <cphealy> I've been using NIR with all three of GC2000, GC3000, and GC7000L. I can't comment on how far away we are with it being on by default and my use cases are not that complex though.

14:11 <cphealy> My use cases include relatively simple games, Qt GUIs, Weston, and a 3D map application. I don't have any rendering issues when running these with NIR enabled.

14:13 <sravn> cphealy: The use case I have in mind is an application running with Chromium as the browser.

14:14 <sravn> We have decent performance today (i.MX6 QuadCore) and was just wondering if I should persuade them to move to latest mesa and anable nir.

14:15 <sravn> We have also something running on i.MX6 SoloCore, hence the gc880 question - here the performance could be better.

14:15 <cphealy> sravn: I have run some browser based use cases in the past with NIR enabled on the GC3000 and GC7000L but I never tested with GC2000.

14:15 <cphealy> What version of Mesa are you on?

14:19 <austriancoder> sravn: shader-db on gc2000 with latest master

14:19 <austriancoder> sravn: total instructions in shared programs: 63628 -> 48944 (-23.08%)

14:19 <austriancoder> instructions in affected programs: 58996 -> 44312 (-24.89%)

14:19 <austriancoder> HURT: 1

14:19 <austriancoder> helped: 255

14:19 <austriancoder> helped stats (abs) min: 4 max: 316 x̄: 57.60 x̃: 36

14:20 <sravn> cphealy: trying to dig it up, one minute

14:21 <austriancoder> https://hastebin.com/xocogubequ.apache

14:21 <austriancoder> better that way :)

14:21 <sravn> cphealy: mesa 19.2.6

14:23 <sravn> austriancoder: Thanks, so if I tell them that we may see a 20% speed-up moving to latest mesa and enabling nir. For the GPU operations ofc

14:24 <sravn> Thats seems like a nice win and maybe if we hit something we (read: my work) could sponsor to get it fixed

14:25 <austriancoder> sravn: the actual speed-up depends on the workload/application but there is an improvement. I am always a little bit careful when providing general numbers

14:27 <sravn> Yeah, I will phrase it properly so people do not get too disappointed if there see less difference.

14:28 <sravn> But I also need a little to sell them the idea to move to new mesa and enable nir. Your figures should do the trick.

14:29 <austriancoder> sravn: :)

19:55 <marex> sravn: note that chromium uses threaded rendering and the locking in etnaviv isn't particularly performant

19:55 <marex> sravn: that could be where you will hit performance problems

20:02 <sravn> marex: thanks for the hint

20:29 <marex> sravn: do you plan to run chromium on WL or GBM ?

20:51 <sravn> marex: we are running Chromium on top of weston

21:47 <marex> sravn: OK, fine

21:48 <marex> sravn: Im running both that and qtwebengine directly on kms/gbm , but there is apparently the option of running chromium on ozone-gbm (without WL/weston), I wonder if that might be better suited to mx6

22:00 <sravn> marex: I think we are using what should be optimal. At least it was Pengutronix that did it back then and they should know.

22:01 <sravn> But maybe there are new, yet to be explored, possibilities