PDA

View Full Version : strange results in opengl tests


i__h
11-30-2008, 08:45 AM
Hi

I ran some very basic opengl tests and compared windows vs linux performance on my hd4850.
The tests can be found here: http://www.songho.ca/opengl/gl_pbo.html , I ran pboUnpack, pboPack, vbo and fbo.

Results look very strange:

Test, Type: win, linux, wine
pboUnpack, w/o: 470MB/s, 443MB/s, 380MB/s
pboUnpack, PBO1: 680MB/s, 207MB/s, 208MB/s
pboUnpack, PBO2: 680MB/s, 207MB/s, 208MB/s

Nothing special so far, nvidia had the same problems with PBOs on linux until 180. Obviously there's some work to do for ati.

Very interesting:

pboPack, w/o: 22.6MPix, 2.3MPix, 2.3MPix
pboPack, with: 3.1MPix, 3.1MPix, 3.1MPix

Looks like ati also hasn't implementet PBOs on windows. The factor 10 difference w/o could explain the very bad ut2004 performance I get on linux, dunno if it uses PBOs very much, but it's at least factor 3 slower.

Next tests:

vbo, w/o: 715fps, 335fps, 540fps
vbo, with: 680fps, 336fps, 515fps

fbo, w/o: 946fps, 430fps, 569fps
fbo, with: 960fps, 440fps, 575fps

In wine these tests run much faster than native. There must be something wrong here. The 50% performance degradation vs. windows also looks very bad.


edit: used catalyst 8.11

FunkyRider
12-01-2008, 12:14 AM
nice findings.

I guess the performance issues of vbo is related to lacking of proper graphics memory manager...

wonder what TTM/DRI2 could do about it

i__h
12-13-2008, 09:08 AM
Looks like the pboPack issue was fixed with 8.12. Now delivers 19.8/21.6 wo/with PBO.
pboUnpack also increased, now I get 280MB/s with PBO mode 1 and 2. The other tests increased about 10%, dunno if it's the new driver or my new kernel.