Linux 3.12 Brings Big AMD Radeon Improvements

[Marc Driftmeyer]

How can I run this exact test suite on my machine?

If running Debian you can install it via deb package and following the instructions.

[tom.higgy]

3.12 does NOT turn DPM on by default. There were way too many changes in this kernel to consider the feature stable yet, but maybe 3.13 will do so. Also, the 4870 card michael is testing here is unlikely to get any performance benefit from it anyway - it's likely clocked high by default, and would just save power with DPM. It's the newer cards that tend to be clocked low by default and get big speedups - Radeon SI cards, Cayman, and the APU parts. And likely mobile as well.

OK, so these improvements are for other reasons. Even better!

[AJSB]

Seeing as the results even surprised Alex, I'll run more tests today on a completely different system with various generations of AMD GPUs and try to get these results up tomorrow to see how those go.

Thanks very much

Can we request from a E-350 to a A10-6800K ?

[MWisBest]

Seeing as the results even surprised Alex, I'll run more tests today on a completely different system with various generations of AMD GPUs and try to get these results up tomorrow to see how those go.

Thanks, I and probably others here appreciate the time you put into Phoronix and am appreciative of you taking the time to run through this again. Not a lot of people would do something over like this, they'd probably publish it and call it quits.

[Michael]

Maybe a git bisect too Michael? Just to narrow down the commit or at least a series of commits that could've effected this

First I'll run another article or two out of it (to see just how rare or widespread the improvements are and for maximizing views...) before letting PTS bisect it

[Ericg]

Seeing as the results even surprised Alex, I'll run more tests today on a completely different system with various generations of AMD GPUs and try to get these results up tomorrow to see how those go.

Maybe a git bisect too Michael? Just to narrow down the commit or at least a series of commits that could've effected this

[Michael]

Seeing as the results even surprised Alex, I'll run more tests today on a completely different system with various generations of AMD GPUs and try to get these results up tomorrow to see how those go.

[Tyler_K]

Mesa 9.3 git also uses the SB backend by default, doesnt it?

It does now, yes. But (as I can see this as being a potential source of confusion), it should be noted that Michael has published earlier benchmarks utilizing 9.3 dev builds that came prior to the change which enabled sb by default.

[Ericg]

The only asic that commit affects is r6xx chips. R7xx and newer are not affected since they use the async dma rings for bo moves and have since kernel 3.8 or so. The only time the new cpdma code would come into effect for newer chips is when you run the kernel bo tests. TBH, I don't know of any particular changes that would have had much impact on performance.

Alex, Michael mentioned some "a major ring handling clean-up" is it possible that the old 'dirty' code was inefficient and that the clean-up gave it a more direct, faster codepath ( or stopped hitting SLOW code paths). I could be totally wrong but if the blit work doesn't effect this test, thats the only other thing mentioned by Michael that stands out. Mesa 9.3 git also uses the SB backend by default, doesnt it? So does 3.12 include anything that a better shader compiler would benefit? Just spitballing idas here

[Hamish Wilson]

As for the performance gains:

And it just keeps rolling along...