With some of our new platforms, CPU overhead is becoming a major issue, so I started poking around the profile of cairo-gl again. I noticed some comedy in the FBO statechange path, where i965 was recomputing state multiple times up front instead of delaying it until draw time. Most it originally came from the merge of FBO support to i965, where I flagged extra state because we didn't have all the required state flags sorted out at the time.
In the process, I noticed some silly duplicated code between drivers and Mesa core, and ripped that out. Should you want to test it, the code is at the i965-drawbuffers-reduction branch of my Mesa tree. The end result for cairo-gl is a 2.2964% +/- 0.459116% improvement in test runtime (n=31/30). I didn't analyze which particular patches gave the biggest wins, though patch 10/11 was big on the profile. _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
