[Announce] CSSU devel, it may eat your brain - Page 15 - maemo.org

Android_808 | # 141 | 2013-03-17, 23:28 | Report

i would send them a bug report along with your benchmark test and results.would be better coming from you as you did it.

The Following 2 Users Say Thank You to Android_808 For This Useful Post:
luf, Wikiwide

exee | # 142 | 2013-03-18, 20:19 | Report

Can anyone mail me cause i have smo stupid questins about that repo?
blown_out@abv.bg

Thank you

The Following User Says Thank You to exee For This Useful Post:
Wikiwide

freemangordon | # 143 | 2013-03-21, 07:48 | Report

Originally Posted by freemangordon View Post

dammit, it was in vain, I did a couple of cairo/pixman benchmarks, all point that there are performance regressions in upstream pixman compared to the one in fremantle

Honestly, not sure how to deal with that. But I guess it does not make sense to "upgrade" to something which brings performance regressions, esp when there are no bugs in stock pixman. afaik.

It is a tricky situation, as on couple of benchmarks upstream performs way better than stock. But it seems that on those that affect n900 most, it performs worse (16bit operations).

Maybe someone should raise a bug against upstream pixman, IDK. I'll try different versions, will choose the latest that performs well on n900.

Ok, it turned out that the very latest pixman version (0.29.2) with a couple of patches forward-ported (and some commits/changes reverted) performs almost as good (if not better) as the stock one on the most of the benchmarks and outperforms stock in a couple.

I am up to the point where I can't decide by myself what to do. Should this be included in CSSU, having in mind that the binary size is increased almost twice. IMO this should be included (no matter minor performance regression), but I would like to have an "external" assessment.

Bellow are the benchmarks, please review and comment.

There are couple of benchmarks where the new pixman crawls (some *_2x10 and *_2a10), ignore that, if it makes sense I'll (try to) make those at least as fast as in stock. But I won't do that before a decision for inclusion in CSSU is made.

Benchmarks: http://pastebin.com/9via863b

EDIT:

new pixman (0.29.2-2+0m5+0cssu0) is int the repo, feel free to test it

Last edited by freemangordon; 2013-03-21 at 07:52.

The Following 7 Users Say Thank You to freemangordon For This Useful Post:
Estel, hyno111, MartinK, misiak, Raimu, sixwheeledbeast, Wikiwide

Estel | # 144 | 2013-03-21, 10:49 | Report

Originally Posted by #maemo-ssu

<Estel_> freemangordon, you've said that new pixman, even with your changes, performs "almost" as good as old one
<Estel_> what are benefits of new one, then?
<Estel_> some new possibilities, some bugs fixed (you said no bugs hitting old one)?
<freemangordon> Estel_: new pixman opens the door for new xserver (for example). Also microb can be re-compiled to use system pixman, now it uses built-in with NEON accel disabled
<freemangordon> Estel_: new pixman means we can try to upgrade cairo as well

So, big "yes, please upgrade it" here

/Estel

The Following 4 Users Say Thank You to Estel For This Useful Post:
MartinK, misiak, sixwheeledbeast, Wikiwide

Android_808 | # 145 | 2013-03-21, 15:07 | Report

The fact that it could enable acceleration for microb might be worth investigating. Might reduce delay in displaying page content during quick scrolling or task switching.

This one looks a bit different though:

Code:

add_8_8 =  L1:13026.15  L2: 338.12  M:199.12 ( 52.22%)...

Would Alopex have the same option of enabling it?

The Following User Says Thank You to Android_808 For This Useful Post:
Wikiwide

freemangordon | # 146 | 2013-03-21, 17:54 | Report

Originally Posted by Android_808 View Post

The fact that it could enable acceleration for microb might be worth investigating. Might reduce delay in displaying page content during quick scrolling or task switching.
This one looks a bit different though:

Code:

add_8_8 =  L1:13026.15  L2: 338.12  M:199.12 ( 52.22%)...

Well, I guess this value is incorrect, no way Cortex-A8 could draw with 13GB/s

.

IIUC the last value (Kops/s) is what is the best to look at. There is an explanation for it somewhere, but I was too lazy to remember.

Originally Posted by

Would Alopex have the same option of enabling it?

Sure, upstream gecko wants pixman 19.2 at least.

The Following 3 Users Say Thank You to freemangordon For This Useful Post:
Android_808, misiak, Wikiwide

tanago | # 147 | 2013-03-23, 11:13 | Report

Originally Posted by

....Also microb can be re-compiled to use system pixman, now it uses built-in with NEON accel disabled
...

I don't know about MicroB bu I just build Fennec with the new pixman & neon acc and just...WOW
So much smoother; I'm gonna upload it in my Fennec thread in a while.

Last edited by tanago; 2013-03-23 at 11:22.

The Following 6 Users Say Thank You to tanago For This Useful Post:
Android_808, freemangordon, luf, misiak, Raimu, Wikiwide

tanago | # 148 | 2013-03-24, 17:37 | Report

I would like to insert a note here
After a couple of days of testing Fennec with pixman libs I noticed
1. Gradient colors used for backgrounds for some sites are less smoother like that

Left is with pixman libs, right without

2.On some tests (like 'mobile HTML 5 rendering' test on http://craftymind.com/guimark2 ) performance is 2x+ slower with pixman libraries

Edit: I don't know how much MicroB will 'like' these libs, but still a test build will be good to be benchmarked

Last edited by tanago; 2013-03-24 at 17:39.

The Following 3 Users Say Thank You to tanago For This Useful Post:
freemangordon, luf, Wikiwide

freemangordon | # 149 | 2013-03-24, 22:31 | Report

Originally Posted by tanago View Post

I would like to insert a note here
After a couple of days of testing Fennec with pixman libs I noticed
1. Gradient colors used for backgrounds for some sites are less smoother like that...

2.On some tests (like 'mobile HTML 5 rendering' test on http://craftymind.com/guimark2 ) performance is 2x+ slower with pixman libraries

Edit: I don't know how much MicroB will 'like' these libs, but still a test build will be good to be benchmarked

I've forward-ported linear gradient algo from maemo stock pixman, which is using integers (instead of floats used in upstream). The reason is that gradient in upstream was 2x-3x times slower than on stock, albeit more smooth.

Could you do some profiling to see which is the function slowing down the benchmarks? It is possible there are some performance tweaks/patches in mozilla pixman which can be used.

The Following 2 Users Say Thank You to freemangordon For This Useful Post:
luf, tanago

Android_808 | # 150 | 2013-03-24, 22:37 | Report

i find the before/after and left/right a little confusing...

with regards to microb, worth a shot. what we got to lose. as we're using a much older engine compared to your fennec build and alopex, we are missing out on some of the performance enhancements they feature. anything we can do to the existing setup, be it thumb, pixman... is worth a try.

ideal world we could shoe-horn a new version of the engine in whilst maintaining the closed source ui until a feature complete replacement was ready.

The Following 2 Users Say Thank You to Android_808 For This Useful Post:
luf, tanago