AMD R600g Performance Patches Yield Mixed Results

When running Xonotic at 1280 x 1024 and 1920 x 1080 resolutions with low quality settings, Marek's patches provided boosts in the frame-rate. At 1920 x 1080 with low quality settings, the frame-rate rose by 32% for the Radeon HD 4650 graphics card.

The real problem though was when pushing the image quality settings higher with Xonotic and testing the same resolutions... The frame-rates were now slaughtered. The frame-rate prior to Marek's patches this week were 44 and 34 FPS for 1280 x 1024 and 1920 x 1080, respectively. When using the latest Git head with the patches, the frame-rate fell to just 5 FPS and 1 FPS, respectively.

While Marek's patches were successful in boosting the frame-rate for the ioquake3-era games that are rather lax in graphics intensity and not very demanding on modern GPUs, but the more demanding workloads is where it seems these patches represent a huge regression in performance. The Xonotic performance with high quality settings show the R600g driver immediately falling apart and there's the independent bug report from other users also mentioning these patches causing problems for the more-demanding Unigine Heaven and Enemy Territory: Quake Wars OpenGL workloads.

This is also a perfect example for showing the need for proper continuous integration testing within Mesa. Fortunately, I have been working on it and have the components ready through Phoromatic and OpenBenchmarking.org for launching as soon as time allows with finishing off a few integration bits and finding some spare systems to exclusively dedicate to the task.

There's still a lot of work to do to better the Radeon Gallium3D performance from shader compiler improvements to the memory, buffer, and tiling heuristics.