Created attachment 24069 [details]
Bug detailed description:
Start with the KMS kernel, then startx with uxa, run 2D benchmark of aa10text and rgb10text, both of them will have a very bad performance, its performance was approximately only 1/6 ~ 1/5 of using UXA with non-KMS kernel.
2.x11perf -aa10text (x11perf -rgb10text)
Created attachment 24070 [details]
Created attachment 24217 [details] [review]
use GTT maps in KMS mode
Can you try this?
It seems that this patch has no effect on its performance with KMS. It is still 88.8k with rgb10text 181k with aa10text. While with non-KMS kernel, 408k with rgb10text , 418k with aa10text. (With your patch a little better 532k with rgb10text, 728k with aa10text.)
I think you need the updated patch and libdrm bits I posted in 20803 at the very least, though there may be performance issues beyond that.
(In reply to comment #4)
> I think you need the updated patch and libdrm bits I posted in 20803 at the
> very least, though there may be performance issues beyond that.
I used the update update 2D patch and libdrm patch in 20803 with the KMS in UXA mode. But it still has no change. BTW, I tested on 945gm, is your patch fitted ot it?
Its performance under KMS has improved a lot from the former one though still lower than UMS. With KMS, it is 384k with rgb10text 451k with aa10text. With UMS they are correspondingly 512k and 782k. With the following configuration:
on 945gm-32, KMS+UXA, its performance goes down again. Now is rgb10text 88.2k, aa10text 180k. With UMS+UXA, it is correspondingly 535k and 811k. Maybe relates to the newly checked ins in 2D side in 2.7 branch.
Maybe related (or maybe not, but still on 2D performance), have you tried gtkperf with and without KMS? I've tried in 2 computers (945GM and G45 chipsets) with same results. For example, on my desktop with a G45 chipset and X4500HD graphics I get:
EXA: 5.5 seconds
UXA no KMS: 11 seconds
UXA + KMS: I had to ^C after the "GtkDrawingArea - Lines" test was taking more than a minute instead of about 1 second it takes without KMS. The slownes is perfectly visible when drawing the lines as opposed to when not using KMS.
This should help a lot:
Author: Li Peng <email@example.com>
Date: Fri Apr 10 14:39:35 2009 +0800
Turn on front buffer tiling in KMS.
Created attachment 24792 [details] [review]
Use all fence regs
This patch might help pre-965 chips a bit since they need all the fences for rendering. For 965+ you'll need the kernel patch I posted to intel-gfx today...
With the newst codes, on 945gm-32 it works well now. With KMS aa10text is now 817k, rgb10text is 539k. With UMS+ UXA, correspondingly are 820k and 588k.
What's else, gtperf run under UMS+ UXA is 21.30, whereas with the old code is the same as Alberto González's, that I had to ^C after the "GtkDrawingArea - Lines" test was taking more than a minute instead of about 1 second it takes without KMS. So I think it is OK now.
Great, closing this out then. Thanks for confirming.
Closing really old verified+fixed.