My hardware: HD7950 & Phenom II x6 1090 The problem: intensive frame stuttering and frame rate drops below 30 fps while playing UT2004. Drops usually happen when there is intensive shooting, but it may also happen on maps with complex geometry. Also, in-game fps counter shows 'avg fps' about 90-200, with rare jumps to 300-400 fps, even with vblank_mode=0 set. 64 bit version of the game on Windows shows stable 300+ fps with OpenGL renderer. I wonder if this may be related to AMD CPU, since various test results on Phoronix show much greater performance (up to 5 times) with Intel CPU and same GPU.
Please attach the corresponding Xorg log file and output of glxinfo and dmesg.
Created attachment 124884 [details] dmesg
Created attachment 124885 [details] Xorg.log
Created attachment 124886 [details] glxinfo
Created attachment 124887 [details] lspci
radeon kernel module was patched to provide everything needed for OpenGL 4.3/compute shader. If you want, I may test with vanilla kernel.
Can you run the game with the environment variable GALLIUM_HUD=fps,requested-VRAM+VRAM-usage,requested-GTT+GTT-usage,cpu+temperature+GPU-load,num-bytes-moved,buffer-wait-time,num-compilations+num-shaders-created , grab a screenshot within a minute of the framerate dropping below 30 fps, and attach it here?
Created attachment 124891 [details] screenshot This is closest I could get. Taking the screenshot was a bit tricky. Looks like GALLIUM_HUD update time is to low to detect fps drops. You can see it in in-game counter "cur_fps" when it gets violet, but it's almost impossible to take a screenshot exactly in that moment.
Do the drops correspond to the small bumps on the lowermost graph?
Yes, looks like they do correspond bumps on last graph.
So it looks like the drops are due to shader compilations. Given that you're using current LLVM SVN & Mesa Git, that should only happen if the application explicitly asks to compile new shaders. Marek pushed changes to Mesa Git master this week which might help somewhat, otherwise a persistent shader cache may be the only hope.
We now have both an in memory and on disk shader cache so this situation should be much better now. Closing.
Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.