Slow AA and RGB text rendering
michel at daenzer.net
Mon Jan 30 09:32:23 PST 2006
First of all, this went to the wrong mailing list; debian-x is for
discussion of the Debian X packaging. I'm CC'ing the upstream xorg list,
please follow up there only.
On Wed, 2006-01-25 at 22:30 -0500, debian at openserve.org wrote:
> On the following machine:
> Athlon64 3000
> Nvidia GeForce 6600GT
> Kernel 2.6.14, Xorg 6.9.0.dfsg.1-3
> (Render enabled and verified working, Composite disabled, GLX working)
> I'm getting only 22 kchar/s using: x11perf -rgb10text
> Even software rendering should not be THIS slow considering that...
> ..For comparison, I dug out a retired P2-266 laptop with a PCI S3
> video chipset. It hasn't been updated for years. (Kernel 2.6.1,
> XFree86 4.2.1-16) The same benchmark reported about 17 kchar/s.
Things to consider:
* -rgbXtext (I assume this is subpixel AA) is very non-trivial to
accelerate, I'm not aware of any accelerated (and correct)
* Hence this test probably boils down to software rendering, which
is usually limited by the CPU access to video RAM.
* CPU reads from video RAM are still about as slow as ever
(especially with AGP, less so with PCIe).
> [Sidenote: for some reason x11perf -aa10text is not working in 6.9.0..
> it still does RGB and reports the same result. Otherwise, I would
> have reported this as well. Nevertheless, my guess is that the
> result would be the same, judging by perceptible text redraw speed in
> applications with AA vs RGB mode.]
> Also, for reference, my prior Radeon 8500 card gave similar poor
> results. Before Xorg 6.9.0, I got around 45-50 kchar/s in either AA
> or RGB mode using x11perf -aa10text or -rgb10text respectively.
> While this sounds much faster, it is nowhere near what it should be
> and still results in visibly sluggish text redraw. I used to get
> around 250-300 kchar/s with AA text in some older XFree86 releases
> (and one CVS checkout of pre-6.8.2 Xorg!) with varied Radeon
> hardware. (RGB was still slow, around 50 kchar/s)
Some of these oddities might be explained by the fontconfig
configuration on your system. E.g., if your fontconfig configuration
enables subpixel AA, that seems to apply to the -aaXtext tests as well.
Some of us used to get fooled in the other direction because fontconfig
silently chose a non-AA font for -aaXtext, so we got impossibly high
numbers for that.
Earthling Michel Dänzer | Debian (powerpc), X and DRI developer
Libre software enthusiast | http://svcs.affero.net/rm.php?r=daenzer
More information about the xorg