626581 - Perf regression from enabling subpixel AA for D2D

Status: RESOLVED WONTFIX

(Core :: Graphics, defect)

Description

[Nils Maier [:nmaier]]

https://hg.mozilla.org/mozilla-central/rev/aa14a8d4c97b causes a serious performance regression when painting.
All scrolling gets dead slow on my notebook, to the extend where browsing is a pain. Also affects richlistbox/treebox widgets very badly.
I hg bisect'ed the changeset in question.


Adapter Description
NVIDIA NVS 3100M

Vendor ID
10de

Device ID
0a6c

Adapter RAM
512

Adapter Drivers
nvd3dumx,nvwgf2umx,nvwgf2umx nvd3dum,nvwgf2um,nvwgf2um

Driver Version
8.17.12.6099

Driver Date
10-16-2010

Direct2D Enabled
true

DirectWrite Enabled
true (6.1.7600.16699)

WebGL Renderer
TransGaming Inc. -- ANGLE -- OpenGL ES 2.0 (git-devel Jan 15 2011 03:39:46)

GPU Accelerated Windows
1/1 Direct3D 10

Comment 1

[Bas Schouten (:bas.schouten)]

See if this is improved with tomorrow's nightly which has the fix for bug 626227.

Comment 2

[Nils Maier [:nmaier]]

Just did a build and it got far better, that is, the scrolling doesn't seem to be stuttering any longer. At least it is not readily noticeable for the sites I tested and that stuttered before.

However there is a lot higher CPU load during scrolling than before bug 622482 landed. Peaks at 50% (i.e. one core fully utilized) now, peaked at <25% before.
So I cannot really consider this fixed.

Comment 3

[Bas Schouten (:bas.schouten)]

That's odd, I cannot reproduce that, somebody that can should probably profile this to see where we're spending time in those situations.

Comment 4

[Bas Schouten (:bas.schouten)]

(In reply to comment #2)
> Just did a build and it got far better, that is, the scrolling doesn't seem to
> be stuttering any longer. At least it is not readily noticeable for the sites I
> tested and that stuttered before.
> 
> However there is a lot higher CPU load during scrolling than before bug 622482
> landed. Peaks at 50% (i.e. one core fully utilized) now, peaked at <25% before.
> So I cannot really consider this fixed.

So, even on my Core i7 if I actively scroll on a moderately complex site, it saturates one core on my CPU. Are you sure you were getting <25% on any websites before? (It's possible, don't get me wrong, if your GPU is being the bottleneck, I tested this with a more powerful GT230M)

Comment 5

[Nils Maier [:nmaier]]

Core i7 M620 here, dual core with ht disabled.

I did some more testing and my initial assessment has been overstating the problem. The perf got worse on average, but more in the +20% ballpark instead of the +100% I claimed before, i.e. ~20-25% before, ~25-30% after. CPU load, that is. I don't know about GPU load.
When I first tested I had a Remote Desktop session to my build machine open in the background, albeit minimized, which might have caused Minefield to eat more CPU circles somehow (GPU memory pressure or something; not an expert here).

Saying that the GPU might be the bottleneck here kinda makes sense. The NVS3100M isn't exactly known for it's great performance.
So maybe the new code avoids some GPU bottleneck allowing the GPU and hence CPU to be utilized better, specially considering that you get more CPU load with a faster GPU? What then would mean this is not a perf regression, but in fact a scrolling perf improvement, although it isn't exactly noticeable on this box.

Comment 6

[Bas Schouten (:bas.schouten)]

(In reply to comment #5)
> Saying that the GPU might be the bottleneck here kinda makes sense. The
> NVS3100M isn't exactly known for it's great performance.
> So maybe the new code avoids some GPU bottleneck allowing the GPU and hence CPU
> to be utilized better, specially considering that you get more CPU load with a
> faster GPU? What then would mean this is not a perf regression, but in fact a
> scrolling perf improvement, although it isn't exactly noticeable on this box.

This is possible. The interactions there are very complex and it's often hard to predict how things interact. I think I'm willing to live with the CPU usage difference you mentioned for now, I don't think this blocks 2.0. If more serious performance issues show up, of course let me know.

Comment 7

[Nils Maier [:nmaier]]

(In reply to comment #6)
> (In reply to comment #5)
> > Saying that the GPU might be the bottleneck here kinda makes sense. The
> > NVS3100M isn't exactly known for it's great performance.
> > So maybe the new code avoids some GPU bottleneck allowing the GPU and hence CPU
> > to be utilized better, specially considering that you get more CPU load with a
> > faster GPU? What then would mean this is not a perf regression, but in fact a
> > scrolling perf improvement, although it isn't exactly noticeable on this box.
> 
> This is possible. The interactions there are very complex and it's often hard
> to predict how things interact. I think I'm willing to live with the CPU usage
> difference you mentioned for now, I don't think this blocks 2.0. If more
> serious performance issues show up, of course let me know.

Sounds reasonable. Agreed.