Platform: IVB/BYT/BDW Libdrm: (master)libdrm-2.4.56-4-g07fead4462b2d537d0c3cae69a09272fc426c598 Mesa: (master)52901ec2615761390f5ef97b11516dae330d27d1 Xserver: (master)xorg-server-1.15.99.902-121-g2f5cf9ff9a0f713b7e038636484c77f Xf86_video_intel: (master)2.99.914-46-gec10cef592b7a3d1dc3c4949778e72e8 Cairo: (master)c7ff9bb32e20679d6da4e8a2856be716e5bd9e12 Libva: (master)3fddb7f937df4e0e15391bd65ae3c7552ea5b3d7 Libva_intel_driver: (master)82d2ed8d7da3619c0ea467c06604f5626fc0b901 Kernel_version: drm-intel-nightly(git-723dbb) Bug detailed description: ---------------------------------------------- etqw-demo/Lightsmarkv2008/warsowv1.0/Counter-Strike:source/Half-Life2/Portal/synmark2_v6(OglBatch0-Oglbatch7/OglDrvRes/OglDrvShCom/OglDrvState/OglGeomPoint/OglGeomTriList/OglGeomTristrip/OglHdrBloom/OglVSDiffuse1/OglVSDiffuse8/OglVSlnstancing/OglVSTangent)unigine-demos/GLbenchmarkv2.5.1/GLbenchmarkv2.7.0/3DMME2.0 taiji/hoverjet performance reduced 10%~90% on IVB/BDW/BYT, this problem doesn't exist on HSW, The problem exists both on gnome-session and Raw X. It’s Mesa regression, bisecting show that the first bad commit is: af13cf609f4257768ad8b80be8cec7f2e6ca8c81 Author: Kenneth Graunke <kenneth@whitecape.org> AuthorDate: Sun Aug 10 20:06:44 2014 -0700 Commit: Kenneth Graunke <kenneth@whitecape.org> CommitDate: Tue Aug 12 13:39:25 2014 -0700 i965/vec4: Use MOV, not OR, to set URB write channel mask bits. Reproduce steps: --------------------------------------------- 1. xinit& 2. ./synmark2 OglBatch0
Looking at the automated test numbers, drop between 13th & 11th August is most visible on IVB (and almost as large on BYT). Of the tests, it's most visible in: - Terrain instancing (-90%) - Vertex processing/instancing (-90%) - Geometry processing (-90%) - Batching (-85%) - EgyptHD (-85%) - T-Rex (-80%) - Heaven v3 (-80%) - Tropics (-75%) - TF2 (-65%) Zhoujian, does the indicated commit cause all of this drop?
Does reverting the patch solve the problem?
Broadwell cannot have a performance regression due to this commit, as that commit only changes Gen4-7 code generation. Gen8+ was completely unaffected. According to QA's performance tracking page, Synmark OglBatch0 performance actually increased by 8% between 8/06 and 8/13 on Broadwell. Which is very different than a 10-90% performance regression as reported.
(In reply to comment #1) > Zhoujian, does the indicated commit cause all of this drop? Yes,many 3D cases was reduced.
Reverted on master.
(In reply to comment #3) > Broadwell cannot have a performance regression due to this commit, as that > commit only changes Gen4-7 code generation. Gen8+ was completely unaffected. > > According to QA's performance tracking page, Synmark OglBatch0 performance > actually increased by 8% between 8/06 and 8/13 on Broadwell. Which is very > different than a 10-90% performance regression as reported. I mistake,this issue doesn't exists on BDW.
Verified the latest commit :"git-ef130b6", this issue disappear.
this issue also exit on Mesa:(10.3)git-627d31d release.
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.