824221 - Improve Firefox's performance on Penguin Mark

Status: NEW

(Core :: Graphics, defect)

Description

[Ma Hsiao-Chun]

Attachment: my_score.png

User Agent: Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:17.0) Gecko/20100101 Firefox/17.0
Build ID: 20121129151842

Steps to reproduce:

1. Visit http://blogs.msdn.com/b/ie/archive/2012/12/21/penguin-mark-blazing-fast-holiday-fun.aspx
2. Click [Penguin Mark]
3. Click [Test My Browser]


Actual results:

IE's score can be way more higher that of Firefox.
Firefox's score is limited to 3 digits but IE's score can higher than 1k, 10k,...


Expected results:

Firefox's score is higher than that of IE.

Comment 1

[Dennis Smit]

As posted on: http://code.google.com/p/chromium/issues/detail?id=167413

The scoring is based on the return speed of the draw loop, which adjusts the snow flakes and windSpeed. The final score is:

this.currentScore = Math.round(this.maxCountSnowflakes * this.windFactor);

While windFactor is based on the countSnowFlakes:

this.windSpeed = this.countSnowflakes / 1000;

So the scoring is all but linear.


I have to admit: IE returns damn fast on the draw calls, as far as I know they use deferred rendering. The fast return on a draw call is of course pretty nice because then you can start pushing out the next geometry to the canvas for rendering. Not so much fun when you read pixel data though ;-)

However the whole test is built around this particular fact and so is the scoring. So it is hardly a pure benchmark.


However, here ( https://github.com/dsmit/PenguinMarkNG ) is a version that measures the number of draw calls within the demo run. I increased the amount of snow flakes to a number where both Chrome and IE have a hard time. Locally this scores

IE:      502
Chrome:  255
Firefox: 214

I did multiple runs and the numbers always boil down to around this figure.


Online version: http://penguinmarkng.azurewebsites.net/  please do pull requests to the repo as noted above if you have improvements.

Cheers!

Comment 2

[(mostly gone) XtC4UaLL [:xtc4uall]]

SPS Profile on Win7 x64 for the Version mentioned in previous Comment:
http://people.mozilla.com/~bgirard/cleopatra/#report=45060698942433c73f0b40ec5654655c9a00923c

Comment 3

[Robert O'Callahan (:roc) (email my personal email if necessary)]

There's also some CSS animations going on, so off-main-thread-compositing would help, but probably not much compared to deferred canvas rendering.

Comment 4

[Mayank Bansal]

on my win7 x64 machine :

IE10RC : 4000
FF nightly : 94

Comment 6

[Mayank Bansal]

FWIW, with HWA off on IE10Rc, this demo used 60-70% of my quad core CPU (without Intel OpenCL runtime installed).

Is IE10 the first browser with multithreaded browser?

Comment 7

[Robert O'Callahan (:roc) (email my personal email if necessary)]

Most browsers use multiple threads for various features. For example, pretty much all browsers use multiple threads for video decoder. Firefox uses extra threads for parsing HTML.

IE uses D2D even when HWA is off, and the software D2D implementation uses multiple threads. We could do the same I suppose, but I suspect most machines that have D2D also have working HWA so it may not be worth it.

Comment 8

[Mayank Bansal]

(In reply to Robert O'Callahan (:roc) (Mozilla Corporation) from comment #7)
> IE uses D2D even when HWA is off, and the software D2D implementation uses
> multiple threads. We could do the same I suppose, but I suspect most
> machines that have D2D also have working HWA so it may not be worth it.


Like this :  http://msdn.microsoft.com/en-us/library/gg615082.aspx   ?