01:56 <alyssa> cphealy: Preliminary results seem really good

01:56 <alyssa> as in, fps up by 35% good

01:56 <alyssa> (frame time down by 26%)

02:01 <alyssa> Oh and that was without fixing push constant stuff even better heh

02:08 <alyssa> Interestingly the bug fix for push constants didn't actually affect the shader-db score for that scene

02:26 <alyssa> cphealy: Doing a full glmark run

02:26 <alyssa> Results are making my jaw drop

02:29 <alyssa> uhhh except for -bjellyfish apparently

02:29 <alyssa> which er went from 814fps to 4fps

02:29 <alyssa> I should strongly consider investigating

02:31 <alyssa> jellyfish is fine now ..

02:33 <alyssa> I see faulting in dmesg but why won't it reproduce?!

02:36 <cphealy> 35% good? OMG, that's amazing!

02:37 <cphealy> Do you mean the overall glmark2-es2 score is up by 35% or just a subset of tests?

02:38 <alyssa> cphealy: overall glmark2 score was 976->1014

02:38 <alyssa> but -bjellyfish regressed from 814->4 due to some regression

02:38 <alyssa> ^functional

02:38 <alyssa> For some reason it doesn't reproduce when run in isolation >.<

02:38 <alyssa> [ 2995.906068] panfrost ffe40000.gpu: Unhandled Page fault in AS0 at VA 0x00000000000701E0

02:38 <alyssa> access type 0x2: READ

02:38 <alyssa> source id 0x300

02:38 <alyssa> ...isn't that general purpose read?

02:38 <cphealy> interesting... You've definitely got me excited about doing some more benchmarking in my env.

02:41 <cphealy> A couple months back, we went down the rabbit hole of identifying the cause of the regression with "[buffer] columns=200:interleave=false:update-dispersion=0.9:update-fraction=0.5:update-method=subdata" that took it from 50fps to 3. IIRC, it was CPU bound for some reason and you had a proposed fix. Can you remind me what happened with that?

02:41 <alyssa> https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/13502

02:41 <alyssa> Never got around to finishing up that MR

02:44 <cphealy> yea, that's the one.

02:44 <cphealy> What's still needed with that one?

02:44 <alyssa> finishing it

02:45 <cphealy> ;-)

02:56 <alyssa> Okay, fixed the bug, was something stupid (off-by-one)

02:56 <alyssa> Unfortunately, -bjellyfish fps is hurt regardless

02:56 <cphealy> I assume not from 814->4 though?

02:57 <alyssa> haha no

02:57 <alyssa> 829->775

02:58 <alyssa> jellyfish.vert offers a clue why

02:58 <alyssa> Both gl_Position and the varyings depend on a common variable `pos4`, which is expensive to calculate

02:59 <alyssa> In theory the compiler could detect this case and bail but it's not obvious what heuristic would be best here

02:59 <alyssa> Also in theory the intermediate result could be saved but this would end up significantly slower I think

03:08 <alyssa> I think it makes sense to eat a 7% decrease in fps on a single scene when there are so many scenes that are up 30% etc

03:08 <cphealy> Seems like a decent tradeoff if there is no better option.

03:08 <alyssa> and actually I'm not so sure about those numbers,

03:08 <alyssa> I just got 817 on this run and had 814 before so who knows

03:10 <cphealy> As a datapoint, when I compare jellyfish with vendor DDK numbers, I'm seeing that Mesa runs at 76% the fps as DDK. (This is before your code change.)

03:10 <alyssa> Alright

03:10 <alyssa> I'd be curious what the DDK trace looks like

03:10 <alyssa> You up for some clean r/e fun? :-p

03:11 <cphealy> that's DDK vs mesa 21.2 where DDK had 393 FPS and Mesa had 300 FPS.

03:12 <alyssa> https://rosenzweig.io/idvs-g52.txt

07:23 <bbrezillon> alyssa: guess it worked because everything was promoted to push constants :-)

14:00 <alyssa> bbrezillon: Scary :-p