System Environment: ---------------------------------------------- Platform:IVB/HSW Libdrm: (master)2.4.47 Mesa: (master)8958741e5a0c3cce95f0415ce8e6e8e840a5207d Xserver:(master)xorg-server-1.14.99.2-7-g44d502c6f1d03a9e548 Xf86_video_intel:(master)2.99.904-10-ge914dada5595a24b5d0fada Cairo: (master)6f05ecf488314e4b0c6c6b0110963c449bebe7d7 Kernel: (drm-intel-nightly) 55d077d2efad373f7b36c8ae50f4205f652dea83 Bug detailed description: ---------------------------------------------- unigine-tropics v1.3 performance reduced ~50%. The problem exists on both gnome-session and Raw X. It’s Mesa regression, bisecting show that the first bad commit is: commit 36fbe66d3a71df76fcb6f915846da4471b3a8442 Author: Eric Anholt <eric@anholt.net> Date: Wed Oct 9 17:17:59 2013 -0700 i965/fs: Convert gen7 to using GRFs for texture messages. Performance -------------------------------------------------------------------- Test tropics on HSW mobile GT3e with Raw X git-36fbe66: 12.02 git-ee21c8b: 26.00 Reproduce steps: --------------------------------------------- 1, xinit& 2. export LD_LIBRARY_PATH=bin/:$LD_LIBRARY_PATH export force_glsl_extensions_warn=true vblank_mode=0 ./bin/Tropics -video_app opengl -data_path ../ -sound_app null -system_script tropics/unigine.cpp -engine_config ../data/unigine.cfg -video_mode -1 -video_fullscreen 1 -video_multisample 0 -extern_define PHORONIX -video_width 1920 -video_height 1080
I see that according to your numbers, on BYT and HSW ULT, Lightsmark improved by a third. Based on Eric's commit comment, that would be explained by the same change. Lightsmark didn't improve on other HW which is somewhat strange considering how their GPU / memory bandwidth ratios sit between BYT & HSW ULT... Looking at the numbers for different machines around the time this change happened, the drop in Unigine Tropics FPS was smallest for IVB (desktop), a third, larger for BYT and largest for HSW, about half. Changes for the other internal tests were clearly smaller (e.g. 10% drop for UrbanTerror, but for some reason only on HSW desktop). Those could be unrelated.
http://lists.freedesktop.org/archives/mesa-dev/2013-October/046400.html The top FS spills both before texture-grf and after this patch, and I haven't retested to see if it fully fixes any regression. There's a lot more work that could be done on that shader, before and after. http://dri.freedesktop.org/wiki/I965Todo/ updated.
(In reply to comment #2) > http://lists.freedesktop.org/archives/mesa-dev/2013-October/046400.html > > The top FS spills both before texture-grf and after this patch, and I > haven't retested to see if it fully fixes any regression. > > There's a lot more work that could be done on that shader, before and after. > http://dri.freedesktop.org/wiki/I965Todo/ updated. With above patch, the performance problem still exists on IVB.
Once GLSL CSE lands, we're down to -10.7162% +/- 0.476574% on my IVB.
(In reply to comment #4) > Once GLSL CSE lands, we're down to -10.7162% +/- 0.476574% on my IVB. With the latest driver,unigine-tropics got 19.4186 fps. So now, unigine-tropics reduced ~25%(26 fps ->19.4186 fps). ------------------------------------------------- Libdrm:(master)2.4.47-2-g45810d3d7546f3ddbcdae114c6726ada30e919c9 Mesa: (master)723f047a3bc21b536d27ab1c0d6cc82f93002648 Xserver:(master)xorg-server-1.14.99.3-24-g902ff0b3497d202b86bf9a411e17db7b694d6eaa Xf86_video_intel:(master)2.99.905-43-g82e6d41c2f4f343bd1854d3d8ee4b624b5d68971 Cairo: (master)98fef3cef2d0f7f463a2e4f9f1b35b09f7b6ea77 Libva: (staging)1a011ce5bb0b80506797a25a988854f3f81ce909 Libva_intel_driver:(staging)1cee858036a87837deddc87586701ed869f96
commit e9daead784921e453906853a4a78a2f3135af2e0 Author: Eric Anholt <eric@anholt.net> Date: Wed Nov 6 17:38:23 2013 -0800 i965/fs: Try a different pre-scheduling heuristic if the first spills.
Verified it with above commit git-e9daead78492.
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.