ANGLE vs OpenGL performance

[Peter Liepa]

I've observed a performance difference between WebGL that is using
ANGLE vs. OpenGL. I usually develop on the Mac, where OpenGL is the
norm, but on the desktop XP Windows machine that I occasionally test
on I get maybe 4-5 fps when using ANGLE, but 35-40fps when using
OpenGL (this is in Chrome).

I'd expect ANGLE to be slower, even much slower, because of mismatches
between WebGL and DirectX, but does anybody have any optimizations
that are aimed at ANGLE? Or any specific insights as to why it's
slower? Have people observed this kind of performance difference
between OpenGL and ANGLE for WebGL? Is my one data point so far
compromised because, perhaps, I'm running on XP? Any insights and
observations are appreciated.

[Ben Vanik]

It'd be great if you could narrow it down into a test case that could be submitted as a bug to the ANGLE project - there should be small differences in perf, but nothing that big!

--
Ben Vanik
http://www.noxa.org

[Daniel Koch]

It would also be great if you mentioned which version of Chrome you are using.   There have been some pretty substantial improvements in ANGLE in the last few weeks (changes which might not have made their way into any of the browser builds yet).

Daniel

[Stephen Bannasch]

I'm developing a visualization on my Mac but we plan to test in a classroom on HP 425 laptops. Upgradingthese laptops from Win XP to Win7 made a huge improvement in OpenGL and WebGL rendering speed. Here what someone I work with reported:

At 6:05 PM -0800 2/4/11, Jonathan Lim-Breitbart wrote:
>Okay, upgrading the HPs to Windows 7 makes a HUGE difference. Running the OpenGL rendering tests, the fps for each one goes from around 6 on XP (!) to 229 (at the lowest) and 331 (at the highest) on Windows 7.
>
>The models run WAY smoother and are much quicker to render after upgrading too. I think upgrading to Windows 7 would be areally good plan on these machines. It's just going to take us a little while to do 18 of them!
>
>-Jon

I had earlier asked Jon to test both by running the WebGL visualization and also install GLView Extensions Viewer and send me what it reports. In the message above he is also describing the speedup running the GLView Extensions Viewe OpenGLtests.

At 3:01 PM -0500 2/2/11, Stephen Bannasch wrote:
>Jonathan and Jenn,
>
>GLView Extensions Viewer is a free program for Mac OS X and Windows that displays a great deal of technical information on the version and capabilities of the installed OpenGL implementation.
>
>Can you run this on the different laptops available for use in the classroom and send me the results.
>
>These data will be very helpful when investigating technical issues.
>
> http://www.realtech-vr.com/glview/

[Gregory Ray]

My guess is that a graphics card driver got updated or the web browser
got upgraded as a bi-product of upgrading from XP to 7. Seeing how
anti OpenGL microsoft is I don't think they would have purposely
improved the performance. Just a guess though.

On Mon, Apr 25, 2011 at 7:27 PM, Stephen Bannasch
<stephen....@deanbrook.org> wrote:
> At 1:58 PM -0700 4/25/11, Peter Liepa wrote:

> I'm developing a visualization on my Mac but we plan to test in a classroom on HP 425 laptops. Upgradingthese laptops from Win XP to Win7 made a huge improvement in OpenGL and WebGL rendering speed. Here what someone I work with reported:
>
> At 6:05 PM -0800 2/4/11, Jonathan Lim-Breitbart wrote:
>>Okay, upgrading the HPs to Windows 7 makes a HUGE difference.  Running the OpenGL rendering tests, the fps for each one goes from around 6 on XP (!) to 229 (at the lowest) and 331 (at the highest) on Windows 7.
>>
>>The models run WAY smoother and are much quicker to render after upgrading too.  I think upgrading to Windows 7 would be areally good plan on these machines.  It's just going to take us a little while to do 18 of them!
>>
>>-Jon
>
>
> I had earlier asked Jon to test both by running the WebGL visualization and also install GLView Extensions Viewer and send me what it reports. In the message above he is also describing  the speedup running the GLView Extensions Viewe OpenGLtests.
>
> At 3:01 PM -0500 2/2/11, Stephen Bannasch wrote:
>>Jonathan and Jenn,
>>
>>GLView Extensions Viewer is a free program for Mac OS X and Windows that displays a great deal of technical information on the version and capabilities of the installed OpenGL implementation.
>>
>>Can you run this on the different laptops available for use in the classroom and send me the results.
>>
>>These data will be very helpful when investigating technical issues.
>>
>>  http://www.realtech-vr.com/glview/
>

--
Gregory Ray
Co-founder, Seek Mobile Interactive, Inc.

---

This e-mail message, including any attachments, is for the sole use of
the intended recipient(s) and may contain information that is
confidential and protected by law from unauthorized disclosure. Any
unauthorized review, use, disclosure or distribution is prohibited. If
you are not the intended recipient, please contact the sender by reply
e-mail and destroy all copies of the original message.

[Peter Liepa]

I was using Chrome 10.0.648.205.

I tried aggregating the geometry in 1962 objects into 662 objects, and
there was a noticeable speedup (call it a factor of 3). This would
seem to indicate that with ANGLE there is more overhead in the
bindBuffer/vertexAttribPointer/drawArrays sequence than in OpenGL.

It is of course always a good idea to minimize the number of such
invocations, but apparently even more so in ANGLE.

I guess this goes some way to answering my own question. I can reduce
the number of objects even more, I've just used the easiest reduction
so far. With the 65K limit on vertex arrays, I'd have to spend a
little more effort in optimally combining geometries. But not a big
deal.

Peter

On Apr 25, 8:01 pm, Daniel Koch <dan...@transgaming.com> wrote:
> It would also be great if you mentioned which version of Chrome you are using.   There have been some pretty substantial improvements in ANGLE in the last few weeks (changes which might not have made their way into any of the browser builds yet).
>
> Daniel
>
> On 2011-04-25, at 5:02 PM, Ben Vanik wrote:
>
>
>
>
>
>
>
> > It'd be great if you could narrow it down into a test case that could be submitted as a bug to the ANGLE project - there should be small differences in perf, but nothing that big!
>
> > --
> > Ben Vanik
> >http://www.noxa.org
>

[Peter Liepa]

Hmmmph, Chrome just updated itself to 11.0.696, and ANGLE performance
just got slower. Maybe 30%, in terms of FPS.

I guess this (Chrome/WebGL/ANGLE, I mean) is still a work in progress.

On the other hand, I just tried Chrome Frame in my colleague's copy of
IE, and it feels somewhat miraculous to see WebGL running there :)

Speaking of Chrome Frame, is there any way to get it to use OpenGL
instead of ANGLE, similar to the command line switches I can use for
the Chrome browser?

[Daniel Koch]

The most recent ANGLE performance fixes definitely reduce the draw call overhead, but they have not yet been picked up by Chrome 13, so I'd definitely expect there to be some variability in performance over the next few releases. If you'd like to test drive these changes yourself, you could always compile your own ANGLE dlls from SVN. See: http://code.google.com/p/angleproject/wiki/DevSetup

Hope this helps,
Daniel