This testing is with the latest Intel Git code (the Linux 3.3 kernel, Mesa 8.0-devel, and xf86-video-intel 2.17+) as of 30 January when using the DDX driver in its stock configuration and then rebuilding it with SNA (via the "--enable-sna" build switch). Additionally, the stock Intel acceleration performance of Ubuntu 12.04 LTS "Precise Pangolin" was measured. The daily snapshot of Ubuntu 12.04 LTS from 30 January served as the base operating system.
To make things interesting, an Ironlake (Intel Core i3 330M) system was tested along with a Sandy Bridge desktop (Intel Core i5 2400S) and Sandy Bridge laptop (Intel Core i5 2520M) were used for this round of testing. Also exciting is that this article uses some new test profiles from OpenBenchmarking.org. As new 2D tests there is now Cairo Performance Demos and cairo-perf-trace as new 2D benchmarks using the Cairo engine. You can use these new tests yourself by running phoronix-test-suite benchmark cairo-demos cairo-perf-trace, which with Phoronix Test Suite 3.x+ will automatically fetch them from OpenBenchmarking.org, install, and then execute the tests in a fully automated manner. There's also been updates to some of the existing 2D benchmarks.
Unfortunately this latest SNA code isn't regression free, as some of the performance results indicate. As far as the visual experience went, since for a while visual imperfections were common when using SNA, that mostly appears to be addressed now. The only visual deficiency I witnessed during this latest round of SNA testing was when using the Ironlake SNA back-end (the Sandy Bridge back-end was fine) the window title bars were always white.
On the Sandy Bridge desktop system tested, once while having SNA enabled I did run into the tiling corruption lock-up. This is rare these days (compared to when Sandy Bridge hardware was first introduced) and likely a problem within the DRM driver and should not be attributed to SNA acceleration. Here's the run-down of the three systems used with their hardware and software configurations:
Now let's jump to the results. The focus of this article is obviously on the 2D performance, but there are also some 3D/OpenGL results included for reference to indicate general Intel driver performance improvements between the three tested software configurations.
With the three systems being tested in a multi-way comparison, the results in this article are normalized so that the graphs indicate higher is better and the slowest software configuration in each system has a value of 1.0.