RV670 Benchmark: KMS vs UMS vs KMS w/o vsync

[darkbasic]

Download in .ods format


world of padman
KMS http://global.phoronix-test-suite.co...13-27365-20812
UMS http://global.phoronix-test-suite.co...859-22133-6469
KMS (w/o vsync) http://global.phoronix-test-suite.co...68-31208-13312

nexuiz
KMS http://global.phoronix-test-suite.co...6129-21156-461
UMS too slow to bench
KMS (w/o vsync) http://global.phoronix-test-suite.co...92-15535-15307

ut2004
KMS http://global.phoronix-test-suite.co...068-4177-23194
UMS http://global.phoronix-test-suite.co...35-22734-26276
KMS (w/o vsync) http://global.phoronix-test-suite.co...954-29578-5306

Benchmarks with Obscene_CNN performance patches will follow

[darkbasic]

I precise that disabling vline waiting for buffer swaps in the dri2 paths I saw no tearing but only icreased performance.

By the way, here is the patch (KMS w/o vsync):

diff --git a/src/radeon_dri2.c b/src/radeon_dri2.c
index 103972f..7ddcda0 100644
--- a/src/radeon_dri2.c
+++ b/src/radeon_dri2.c
@@ -313,7 +313,7 @@ radeon_dri2_copy_region(DrawablePtr drawable,
     }
 
     vsync = info->accel_state->vsync;
-    info->accel_state->vsync = TRUE;
+    //info->accel_state->vsync = TRUE;
 
     (*gc->ops->CopyArea)(&src_pixmap->drawable, &dst_pixmap->drawable, gc,
                          0, 0, drawable->width, drawable->height, 0, 0);

[Sadako]

Very nice, thanks for posting this.

You never mentioned what hardware you were running this on, though.
RV670, is that a 3850 or greater?

Curious to see KMS outperforming UMS across the board, it's the opposite of what I've seen earlier...

Could you post comparisons of gtkperf or some other 2d rendering benchmark too?

[darkbasic]

You never mentioned what hardware you were running this on, though.
RV670, is that a 3850 or greater?

Because you can see the hardware details in the global.phoronix-test-suite.com links
RV670 is high-end R600, so 3850 or 3870. Mine is HD3870.

Curious to see KMS outperforming UMS across the board, it's the opposite of what I've seen earlier...

UMS has a much higher max framerate, but also terrible slowdowns, so the average framerate is slower. Unfortunately PTS-2.4.1 prints only the average framerate
Also, UMS has never performed well with my card. Particularly nexuiz which seems to use software rendering! I already filed bugs but no one ever answered...

Could you post comparisons of gtkperf or some other 2d rendering benchmark too?

I already planned 2D benchmarks, but I'dd like to test Obscene_CNN performance patches before.
I'm waiting for him to send me latest kernel patches against drm-radeon-testing and to fix the problems with the other patches.

[darkbasic]

F*CKING edit time :@

Curious to see KMS outperforming UMS across the board, it's the opposite of what I've seen earlier...

Furthermore, phoronix benchmarks had always used a highly outdated graphic stack and in the meanwhile KMS paths improved.

[RealNC]

F*CKING edit time :@


Furthermore, phoronix benchmarks had always used a highly outdated graphic stack and in the meanwhile KMS paths improved.

Dude, there's a difference between "highly outdated" and "current, non experimental". That "highly outdated" you refer to is what people use today.

[bridgman]

That's the problem with periods of rapid progress. Code from a few weeks ago *is* totally outdated

[RealNC]

Don't get me started again about the "Linux Driver Model"

*The non-existent one.

[darkbasic]

Well, since they merged gpu recovery it doesn't matter for me: damn, I love it!

[darkbasic]

P.S.
I was starting to think about switching back to nvidia before gpu recovery was implemented