Really slow 2D performance lately

torquil · 03-25-11, 06:46 AM

Hi!

I have an ASUS X5DID laptop with an Nvidia GT320M graphics card. At the moment I'm using linux kernel 2.6.38.1 (self-compiled, quite slim) and Debian Unstable with Nvidia drivers v270.30-1 from Debian Experimental. The computer has a 64bit Intel T4400 dual core running at 2.2GHz.

I get really slow 2D performance, much slower than when using the open-source nouveau driver. A gtkperf run takes about 30s while it only takes around 7s using the nouveau driver...

Something has happened lately, but I believe I had the same problem also with the v260.19.44-1 nvidia driver currently in Debian Unstable.

As far as I can see there are no interesting messages in /var/log/Xorg.0.log, but I'll attach it anyway.

Lines in kernel log involving the word "nvidia":

tmac@asus ~$ dmesg | grep -i nvidia
nvidia: module license 'NVIDIA' taints kernel.
hda_intel: Disable MSI for Nvidia chipset
input: HDA NVidia Mic as /devices/pci0000:00/0000:00:08.0/sound/card1/input13
input: HDA NVidia Headphone as /devices/pci0000:00/0000:00:08.0/sound/card1/input14
hda_intel: Disable MSI for Nvidia chipset
nvidia 0000:00:03.5: PCI INT B -> Link[LPMU] -> GSI 22 (level, low) -> IRQ 22
nvidia 0000:02:00.0: PCI INT A -> Link[LN0A] -> GSI 17 (level, low) -> IRQ 17
nvidia 0000:02:00.0: setting latency timer to 64
NVRM: loading NVIDIA UNIX x86_64 Kernel Module 270.30 Fri Feb 25 14:34:55 PST 2011

Output from "lspci":
tmac@asus ~$ lspci -v -s 02:00.0
02:00.0 VGA compatible controller: nVidia Corporation GT216 [GeForce GT 320M] (rev a2) (prog-if 00 [VGA controller])
Subsystem: ASUSTeK Computer Inc. Device 1312
Flags: bus master, fast devsel, latency 0, IRQ 17
Memory at fb000000 (32-bit, non-prefetchable) [size=16M]
Memory at e0000000 (64-bit, prefetchable) [size=256M]
Memory at f6000000 (64-bit, prefetchable) [size=32M]
I/O ports at dc00 [size=128]
[virtual] Expansion ROM at faf80000 [disabled] [size=512K]
Capabilities: <access denied>
Kernel driver in use: nvidia

*** EDIT ***:

Output from "gtkperf" run:

GtkPerf 0.40 - Starting testing: Fri Mar 25 13:01:39 2011

GtkEntry - time: 0.14
GtkComboBox - time: 9.42
GtkComboBoxEntry - time: 1.22
GtkSpinButton - time: 0.38
GtkProgressBar - time: 0.11
GtkToggleButton - time: 16.21
GtkCheckButton - time: 1.01
GtkRadioButton - time: 0.75
GtkTextView - Add text - time: 0.82
GtkTextView - Scroll - time: 1.02
GtkDrawingArea - Lines - time: 0.57
GtkDrawingArea - Circles - time: 0.37
GtkDrawingArea - Text - time: 0.29
GtkDrawingArea - Pixbufs - time: 0.07
---
Total time: 32.38

As can be seen, GtkComboBox and GtkToggleButton takes a lot of time. More time each than the entire run did while using the nouveau driver.

And I just checked that the problem does not improve by using a stock Debian kernel instead of my self-compiled one.

**danix** · 03-25-11, 03:03 PM

This sounds like a known issue with some 2D acceleration paths. `nvidia-settings -a AccelerateTrapezoids=0` should work around it.

torquil · 03-25-11, 04:09 PM

You hit the nail on the head! Much appreciated! Everything is again back to normal and Gtkperf runs under 5s now... :-)

For a while there it was like working on a computer from the 90s.

Thanks!

artem · 03-26-11, 07:08 AM

Still, your gtkperf results are awful:

Code:

GtkPerf 0.40 - Starting testing: Sat Mar 26 17:07:57 2011

GtkEntry - time:  0.01
GtkComboBox - time:  0.40
GtkComboBoxEntry - time:  0.29
GtkSpinButton - time:  0.15
GtkProgressBar - time:  0.14
GtkToggleButton - time:  0.13
GtkCheckButton - time:  0.06
GtkRadioButton - time:  0.08
GtkTextView - Add text - time:  0.12
GtkTextView - Scroll - time:  0.07
GtkDrawingArea - Lines - time:  0.42
GtkDrawingArea - Circles - time:  0.42
GtkDrawingArea - Text - time:  0.29
GtkDrawingArea - Pixbufs - time:  0.07
 ---
Total time:  2.66

That's for almost four years old 8800GT (I'm running Fedora 14 i686, NVIDIA 270.30 drivers).

It seems like you are using a composite manager or your Gnome theme is very badly optimized.

**danix** · 03-28-11, 10:29 AM

8800 GT is a high end desktop card. 320M is a fairly entry level mobile part. I don't think that's really a fair comparison, even given the fact that 320M is much newer.

hl_ · 03-28-11, 10:51 AM

When GPU drivers work well, gtkperf usually is quite the CPU benchmark, so this difference doesn't say much...

artem · 03-28-11, 11:22 AM

Quote:

Originally Posted by danix

8800 GT is a high end desktop card. 320M is a fairly entry level mobile part. I don't think that's really a fair comparison, even given the fact that 320M is much newer.

8800 GT was high end three years ago. Now it's just an entry-level/low mid-level card (NVIDIA 450GTS beats it easily).

Deanjo · 03-28-11, 11:40 AM

Quote:

Originally Posted by artem

8800 GT was high end three years ago. Now it's just an entry-level/low mid-level card (NVIDIA 450GTS beats it easily).

The 450 GTS is an entirely different class of card. The 320M is just a GT216 core part ( a whopping 48 shader cores @ 450Mhz using shared ram vs an 8800GT's 112 shader cores @ 600 Mhz using dedicated 256-bit GDDR3 memory.)

artem · 03-28-11, 04:51 PM

I know that, but 2D operations don't noticeably stress modern 3D accelerators. It used to be the issue probably 12 years ago but not now (but even then I don't remember any problems with 2D performance either under Windows 98 or Windows NT 4.0 or XFree86 3.0, everything was very snappy).

Deanjo · 03-28-11, 07:24 PM

Quote:

Originally Posted by artem

I know that, but 2D operations don't noticeably stress modern 3D accelerators. It used to be the issue probably 12 years ago but not now (but even then I don't remember any problems with 2D performance either under Windows 98 or Windows NT 4.0 or XFree86 3.0, everything was very snappy).

Totally different era. Back then cards actually did have dedicated 2D engines. That hasn't been the case in years however on modern cards as they no longer have 2D engines. That is why you see so many regressions in 2D performance no matter the card make / brand on modern cards.

artem · 03-28-11, 07:37 PM

Modern cards have such an enormous compute shaders performance I very much doubt your words. I see regressions in Linux because X.org developers mostly care only about open Intel drivers but I've never seen any slowdowns in Windows 2D performance in the past ten years (even without Aero interface which is basically a graphical 2D interface) except cases when Windows falls back to software rendered Vesa graphics (but it's not what we are talking about here).

mocharhw · 03-29-11, 02:22 AM

Interesting thread. This is from a 9500GT card purchased about 2 years ago. Using 270.26 beta driver.

Code:

GtkPerf 0.40 - Starting testing: Tue Mar 29 00:20:52 2011

GtkEntry - time:  0.02
GtkComboBox - time:  0.62
GtkComboBoxEntry - time:  0.31
GtkSpinButton - time:  0.11
GtkProgressBar - time:  0.10
GtkToggleButton - time:  0.13
GtkCheckButton - time:  0.06
GtkRadioButton - time:  0.10
GtkTextView - Add text - time:  0.59
GtkTextView - Scroll - time:  0.09
GtkDrawingArea - Lines - time:  0.32
GtkDrawingArea - Circles - time:  0.31
GtkDrawingArea - Text - time:  0.21
GtkDrawingArea - Pixbufs - time:  0.05
 --- 
Total time:  3.03

03-25-11, 03:03 PM	#2
danix NVIDIA Corporation Join Date: Feb 2010 Location: Santa Clara, CA Posts: 237	Re: Really slow 2D performance lately This sounds like a known issue with some 2D acceleration paths. `nvidia-settings -a AccelerateTrapezoids=0` should work around it.

03-25-11, 04:09 PM	#3
torquil Registered User Join Date: Mar 2011 Posts: 4	Re: Really slow 2D performance lately You hit the nail on the head! Much appreciated! Everything is again back to normal and Gtkperf runs under 5s now... :-) For a while there it was like working on a computer from the 90s. Thanks!

03-28-11, 10:29 AM	#5
danix NVIDIA Corporation Join Date: Feb 2010 Location: Santa Clara, CA Posts: 237	Re: Really slow 2D performance lately 8800 GT is a high end desktop card. 320M is a fairly entry level mobile part. I don't think that's really a fair comparison, even given the fact that 320M is much newer.

03-28-11, 10:51 AM	#6
hl_ Registered User Join Date: Feb 2009 Posts: 138	Re: Really slow 2D performance lately When GPU drivers work well, gtkperf usually is quite the CPU benchmark, so this difference doesn't say much...

03-28-11, 04:51 PM	#9
artem Registered User Join Date: Jun 2006 Posts: 681	Re: Really slow 2D performance lately I know that, but 2D operations don't noticeably stress modern 3D accelerators. It used to be the issue probably 12 years ago but not now (but even then I don't remember any problems with 2D performance either under Windows 98 or Windows NT 4.0 or XFree86 3.0, everything was very snappy).

03-28-11, 07:37 PM	#11
artem Registered User Join Date: Jun 2006 Posts: 681	Re: Really slow 2D performance lately Modern cards have such an enormous compute shaders performance I very much doubt your words. I see regressions in Linux because X.org developers mostly care only about open Intel drivers but I've never seen any slowdowns in Windows 2D performance in the past ten years (even without Aero interface which is basically a graphical 2D interface) except cases when Windows falls back to software rendered Vesa graphics (but it's not what we are talking about here).