System Environment: ---------------------------------------------- Platform: HSW mobile GT3e i7-4950HQ Libdrm: (master)libdrm-2.4.46-42-gbf4a7cd4b2456d4dc93a86bbcc51eba4ae73390a Mesa: (master)fe2528c0b69d5719b15d926ada9424cac7569b9c Xserver: (master)xorg-server-1.14.99.1-215-g7d3d4ae55dd6ee338439e2424a Xf86_video_intel:(master)2.99.902-54-gd7eb40efa79dd9ea720606b94a20179c7dd18e03 Cairo: (master)337ab1f8d9e29086bfb4001508b28835b41c6390 Libva: (staging)f5c913765e6af38d835cacf339054ccc60bddefb Libva_intel_driver:(staging)f6685c309d94fb7679c9772703c8790cb71cdd73 Kernel: (drm-intel-nightly) 24c8329416b54b79655afe45370cf3d46f41e283 Bug detailed description: ---------------------------------------------- GLbenchmark v2.7.0 EgyptHD _FixedTime performance reduced by 20% on HSW mobile GT3e(i7-4950HQ). The problem doesn’t exist on HSW desktop and another HSW mobile i7-4700MQ . It’s kernel –dinq regression, bisecting show that the first bad commit is: commit 0d8ff15e9a15f2b393e53337a107b7a1e5919b6d Author: Ben Widawsky <benjamin.widawsky@intel.com> Date: Thu Jul 4 11:02:03 2013 -0700 drm/i915/hsw: Set correct Haswell PTE encodings. Performance -------------------------------------------------------------------- Test GLbenchmark v2.7.0 EgyptHD _FixedTime on Raw X drm-intel-next-queued git-50b44a4 git-00d8ff15 183.67 fps 148.67 fps Reproduce steps: --------------------------------------------- 1, xinit& 2, vblank_mode=0 ./GLBenchmark -data ../../data/ -w 1920 -h 1080 -t 2701100
For gt3e we've meanwhile enabled eLLC. What's the numbers on latest dinq?
s/dinq/-nightly of course.
(In reply to comment #2) > s/dinq/-nightly of course. Kernel built on 2013-09-27 -nightly git-8153de: 129.33 -dinq git-cd2983 :127 fps -drm-intel-fixes git-d32270: 133.0 fps Hi, is eLLC enabled in above driver? Or need to do by other way?
(In reply to comment #3) > (In reply to comment #2) > > s/dinq/-nightly of course. > Kernel built on 2013-09-27 > -nightly git-8153de: 129.33 > -dinq git-cd2983 :127 fps > -drm-intel-fixes git-d32270: 133.0 fps > > Hi, is eLLC enabled in above driver? Or need to do by other way? Yes. And apparently there's another regression somewhere ...
See also https://bugs.freedesktop.org/show_bug.cgi?id=67062
Since your bisection here is the very commit that you tested to improve performance in the same test, I strongly suspect external factors.
Ben, can you reproduce it?
I think the situation is too confusing and the delay between when we've committed the patch and when the regression was reported was too long to be actionable. So closing as wontifx.
*shrug* It was something I meant to pinpoint but I never got access to gl2.7
(In reply to comment #8) > I think the situation is too confusing and the delay between when we've > committed the patch and when the regression was reported was too long to be > actionable. So closing as wontifx. Hi, HSW mobile GT3e is added nightly testing at the end of August, then updated this machine OS, so the bug is found later. the performance still exists on -dinq git-f671d1(can't revert the bisected commit git-0d8ff1). So I reopen it.
Modulo using s2tc instead of s3tc, on a i7-4950HQ: 50b44a4: 145fps, 155fps, 152fps 0d8ff15: 132fps, 136fps, 135fps d-i-n: 148fps, 152fps, 156fps Can you please confirm the magnitude of the effect you saw previously and whether it has rectified itself in drm-intel-nightly?
I should note that offscreen rendering we get 187fps (vs 155fps onscreen), which is coincidentally the same magnitude as the reported regression...
And to further complete the puzzle, onscreen rendering but keeping the backbuffers cached, rather than using async flips: 177fps.
Rerunning the benchmark using blits: 50b44a4: 171fps, 193fps, 193fps 0d8ff15: 160fps, 174fps, 175fps d-i-n: 178fps, 178fps, 177fps d-i-n+age3: 195fps, 194fps, 193fps. So ~10% and only slightly larger than hitting uncached surfaces.
Created attachment 89591 [details] [review] Prefer age3 Here's the partial revert in question.
(In reply to comment #15) With above patch, the regression of the case was fixed (-dinq)git-f671d1 151,151,150, (-ding)git-f671d1 with age3 patch 186,185,185
commit c51e9701c4ad68d87baecefad6c58d051574f848 Author: Chris Wilson <chris@chris-wilson.co.uk> Date: Fri Nov 22 10:37:53 2013 +0000 drm/i915: Prefer setting PTE cache age to 3 ... merged to -fixes.
Verified it.
Closing old verified+fixed.
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.