Severe framerate regression with Xserver 1.15, more with Mesa 10.2 on r600
The new Xserver 1.15/Mesa 10.2 combination in Ubuntu Trusty (with the oibaf PPA for mesa) yielded very disappointing results on both of my r600 based systems. About a 50% reduction in framerate in all games I have except CPU limited 0ad. Worst case was on my AMDFX-8120/Radeon HD 6750 system, where framerates in Scorched3d plummeted from a 20 min/about 60 maximum fps to an unplayable 10fps minimum on the heaviest maps, never getting over about 25FPS with all settings maxed out at 1080x1920. On Critter (a 2d game) framerates were cut in half as well, but due to high overall framerates this does not harm the game.
The Radeon 5770 on a Phenom II x4 system had the same framerate drop in Critter, but strangely dropped only to 40fps in Scorched3d from about 60. On the Bulldozer setup I reverted to Xserver 1.14, got back about 80% of the performance, had to revert to mesa 10.1 to get back the rest.
I wonder if all this relates to the DRI3 issue or some part of the GLX rewrite I saw referred to in another story about Xserver 1.15?
Hyper-Z helps when it works on Mesa 10.2, but new Xorg is most of the problem
Originally Posted by gradinaruvasile
Then launching the game with "R600_HYPERZ=1" should work around the issue.
Try this in a terminal:
Scorched3d with hyper-z enabled locks up X on Mesa 10.2 but NOT on mesa 10.1 It seems there is a Hyper-Z issue specific to Mesa 10.2
On Critter (2d game) Hyper-Z works, raises framerate 10-20% but most of loss is due to xorg version 1.15. Reverting Xorg to 1.14, keeping Mesa 10.2, and launching critter from the terminal with
gave at least same framerate as Mesa 10.1 with same version of Xorg. Comparsion is with no shields and one opposing ship on screen which gives the highest framerate but is repeatable.
In Scorched 3d with Mesa 10.2, enabling hyper-z either at game launch with
export R600 DEBUG=hyperz
causes X to lock up during the tank selection phase
Harder to tell in Scorched3d, (my best GPU loader), but it seems that the issues are the
hyper-z issue in Mesa causing a 20% slowdown, and the balance of the 2-1 slowdown is from
changes in Xorg.
CPU is AMD FX8120, GPU is Radeon HD6750, which is a cut-down HD6770. I've observed issues before with AMD's cut-down "salvage" GPU's, where the HD5750 gave only half the framerate of the HD5770, which has
400 SIMD "cores" instead of 320 and a somewhat higher clock speed. The 5770 should be nowhere near
twice as fast as the 5770 but it was with last Summer's code.