Intel SNA With The 2.17 DDX Driver

[phoronix]

Phoronix: Intel SNA With The 2.17 DDX Driver

For those that are thinking about trying out the Sandy Bridge New Acceleration (SNA) architecture option for 2D graphics acceleration by the xf86-video-intel X.Org Linux driver, here are some benchmarks from the recent xf86-video-intel 2.17 release.

http://www.phoronix.com/vr.php?view=16765

[curaga]

Ouch, SNA was slower in near half of the tests.

[ickle]

Ouch, SNA was slower in near half of the tests.

The only (accurate) test where it was significantly slower appears to be a very strange result:

uxa-2.17.0: 80000 trep @   0.3535 msec (  2830.0/sec): PutImage 500x500 square
sna-2.170:  80000 trep @   0.3514 msec (  2850.0/sec): PutImage 500x500 square
sna-master: 80000 trep @   0.3615 msec (  2770.0/sec): PutImage 500x500 square

In most of the these tests SNA is much faster, quite amazing if you have read some of benchmarks.

@Michael, how long before we can contribute test suites to openbenchmarking.org?

[WorBlux]

Ouch, SNA was slower in near half of the tests.

By a quick count I came up with 19-SNA 11-Stock.

[bwat47]

Ouch, SNA was slower in near half of the tests.

In some tests it was a massive improvement though, this is still in heavy development, so if we eventually see it at least match the old method in the tests it does poorly on it will be an awesome improvement as a whole. This is probably the reason they haven't enabled it by default yet.

[curaga]

Well, OpenArena tends to be pretty stable in testing, and it previously got a big boost from SNA but now lost. Is it an inaccurate test?

[ickle]

Well, OpenArena tends to be pretty stable in testing, and it previously got a big boost from SNA but now lost. Is it an inaccurate test?

In terms of absolute performance, OpenArena is about 3x slower than what I'd expect on that hardware, putting the mere 10% delta in perspective.

[ickle]

@Michael, can I be really cheeky and you to include some 800x600 results for the games -- as the lowest common denominator resolution that I can reproduce, and also as more of a stress test for the ddx (since the ddx is invoked at the end of each frame, running at higher fps should help reveal ddx overheads).