System Environment: Platform: HSW mobile Kernel: 3.19.0-rc7+ Libdrm: (master)libdrm-2.4.59-20-g7cb1d6e2d6c63e3bc4b275ff073152b550ba5b0f Mesa: (master)599cbe5508ccb246fe06b64116c84ce976165fed Xserver:(master)xorg-server-1.17.0-16-g3a06faf3fcdb7451125a46181f9152e8e59e9770 Xf86_video_intel:(master)2.99.917-145-g127aae5a72a69df325fed0b63d345e81583d5ca1 Bug detailed description: -------------------------------------------------- Unigine-heaven performance reduce by 12% with gnome-session on HSW mobile, reduce by 6.5% on HSW desktop. It’s Mesa regression. By bisected, show that 2dad1e3a is the first bad commit. commit 2dad1e3abdb1ad153289455f3e273101e5bac1a8 Author: Matt Turner <mattst88@gmail.com> AuthorDate: Wed Feb 12 11:00:46 2014 -0800 Commit: Matt Turner <mattst88@gmail.com> CommitDate: Tue Feb 17 20:44:09 2015 -0800 i965/fs: Add pass to combine immediates. total instructions in shared programs: 5885407 -> 5940958 (0.94%) instructions in affected programs: 3617311 -> 3672862 (1.54%) helped: 3 HURT: 23556 GAINED: 31 LOST: 165 ... but will allow us to always emit MAD instructions. Reviewed-by: Kenneth Graunke <kenneth@whitecape.org> Please see Xrog.0.log. Reproduce steps: ---------------------------- 1, xinit& 2, gnome-session& 3, vblank_mode=0 ./bin/heaven_x64 -video_app opengl -data_path ../ -sound_app null -engine_config ../data/heaven_4.0.cfg -system_script heaven/unigine.cpp -video_mode -1 -video_fullscreen 1 -extern_define PHORONIX,RELEASE -video_width 1920 -video_height 1080
Created attachment 113936 [details] Xorg.0.log info
> by 12% with gnome-session on HSW mobile This is GT3e. > by 6.5% on HSW desktop This is GT2.
Right.
Same commit also regresses PSPom test a bit and clearly improves GfxBench ALU test, GpuTest volplosion & Piano tests, maybe a bit also PSBump2. On GT3e we see much larger effect on Heaven (our GT3e is diffferent version). We're also seeing the effect on all platforms, not just on HSW or GT3e, although on others it's much smaller than with (HSW) GT3e. Because there's both performance increase & decrease, commit doesn't need to be reverted.
Matt, this regression is still visible in Heaven performance (Martin has graph that show how the Heaven perf drops on BDW and it keeping down ever after). Do you have some plan on how the get previous performance back, without regressing test-cases which that commit improved?
-- GitLab Migration Automatic Message -- This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity. You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/mesa/mesa/issues/1474.
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.