Speed comparision of Ati and RadeonHD on Xserver 1.4.2 and 1.5.0

[asdfgh]

Preface.

I'm using mostly outdated Debian ``Lenny'' but with some updates. It comes with xserver 1.4.2, ati 6.8+ and radeonhd 1.2.1. But I've upgraded them to xserver 1.5.0, ati 6.9.0 and radeonhd 1.2.3. After upgrade it seemed to me that combination of xserver1.4 with ati6.8+ is faster then xserver1.5 with ati6.9. But I was too busy (or too lazy) to make tests. At last I've made some of them yesterday.

Tests.

My hardware: video - radeon X1950Pro, CPU - Athlon X2 BE-2350. I've made 5 tests using "gtkperf -a".
Cool&Quiet wasn't disabled, so sometimes 1 test is _much_ longer.
Tested configurations: xserver1.4.2+radeonhd1.2.3 with and w/o BackingStore, xserver1.5.0+radeonhd1.2.3 w & w/o backingstore, xserver1.5.0+ati6.9.0 w & w/o backingstore.

Results (in seconds, first value considered as wrong because of C&Q):

1) XServer 1.5.0 + Ati w/o Option "BackingStore" in xorg.conf
15,54 14,49 14,50 14,29 14,96 -> avg=14.56 s
2) XServer 1.5.0 + Ati with BackingStore
15,72 13,10 13,49 13,22 13,29 -> avg=13.28 s
3) XServer 1.5.0 + RadeonHD w/o BackingStore
14,35 14,10 13,95 14,03 13,33 -> avg=13.85 s
4) XServer 1.5.0 + RadeonHD with BackingStore
14,84 13,24 13,78 13,59 14,16 -> avg=13.69 s
5) XServer 1.4.2 + RadeonHD w/o BackingStore
11,93 10,59 10,50 10,09 10,15 -> avg=10.33 (yeah, only 10.33 s!)
6) XServer 1.4.2 + RadeonHD with BackingStore
too slow (> 17 s), garbage on screen

Conclusions.

1) It's faster with backingstore (and man radeon says the same .
2) W/o backingstore RadeonHD is faster than Ati (because of direct programming).

***** But why XServer 1.4.2 is so MUCH faster then 1.5.0?! *****

TODO:
1) test xserver1.4.2 with ati6.8+
2) disable C&Q and make more iterations
3) make some test and write bugreports about troubles with switching from console to X (X11 hangs), and using GL apps (X11 hangs too) (radeonhd only)
4) Upgrade Xserver to 1.5.3. Where can I take a patch?

[TechMage89]

I believe the slowdown in in the Xserver 1.5.0 is due to a bug in glyph handling code that causes it to run slowly. I think the bug is fixed in git, so try testing that.

It also might help know where the slowdown is if you post the times of the individual gtkperf tests.

[oibaf]

Where is documented Option "BackingStore" (man radeon says nothing)?

Were you using XAA or EXA? The latter should be faster on radeon.

Also, note that 1.5.1 has performance improvements over 1.5.0 and also 1.5.3 over 1.5.1/1.5.2. You should try 1.5.3.

[TechMage89]

"BackingStore" is actually really a server option rather than a driver option, so it usually isn't documented in the driver man page. What it does, I believe, is that is stores copies of areas of screen before drawing over them, so that it doesn't have to redraw them when the area is revealed.

[asdfgh]

2oibaf:
I use EXA.
Yeah, there is nothing in man radeon about bs, may be described in readme.{ati,radeon}. I've found it there in ages of XFree86

[asdfgh]

Gtkperf tests:

http://pastebin.com/m9b55aaa

[asdfgh]

Today's tests of ati under Xserver 1.5.3:

with backingstore:
11,21 11,10 11,57 11,71 11,24 -> avg=11.37 s (really better! not 1.4.2, but good

without
12,30 12,29 12,11 12,24 12,01 -> avg=12.19

C&Q was disabled.

[Linuxhippy]

Its slower because 1.5 has some performance problems which are already fixed in 1.6-git.

One good example is dixLookupPrivate which often consumed 10-20% or even more or total CPU cycles. I can only recommend building x-server GIT and re-testing.

lg Clemens