[Bug 108598] New: [GEN9] 20% perf drop in windowed/composited GpuTest Triangle

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=108598

            Bug ID: 108598
           Summary: [GEN9] 20% perf drop in windowed/composited GpuTest
                    Triangle
           Product: DRI
           Version: DRI git
          Hardware: Other
                OS: All
            Status: NEW
          Severity: minor
          Priority: medium
         Component: DRM/Intel
          Assignee: intel-gfx-bugs at lists.freedesktop.org
          Reporter: eero.t.tamminen at intel.com
        QA Contact: intel-gfx-bugs at lists.freedesktop.org
                CC: intel-gfx-bugs at lists.freedesktop.org
     i915 platform: BXT, CFL, KBL, SKL

HW setup:
* GEN9 HW (BXT, SKL, KBL...)
* Full HD monitor

SW setups:
* Ubuntu 18.04 with Unity desktop (= compiz compositor)
* Git versions of drm-tip kernel, X server and Mesa, so that modifiers work
with X
* Modifier support enabled in X server (i.e. Mesa can use end-to-end render
buffer compression):
----------------------
Section "ServerFlags"
    Option "Debug" "dmabuf_capable"
EndSection
----------------------
* Monitor native Full HD resolution


Between following drm-tip commits/dates:
a3d29ccd2c: drm-tip: 2018y-09m-30d-06h-55m-08s UTC integration manifest
26e7a7d954: drm-tip: 2018y-10m-02d-12h-38m-20s UTC integration manifest

There were following performance regressions:
* 20% in GpuTest v0.7 Triangle windowed (1366x768, composited)
* 1-2% SynMark v7 OglCSCloth, OglBatch1-3 (all fullscreen)
* 1% SynMark v7 OglGeom*, GpuTest v0.7 Triangle fullscreen

SynMark use-cases have all about 2M of vertices with pass-through shaders, so
at FullHD they're memory bandwidth limited.  All the tests have high FPS,
especially the trivial triangle tests.

Interestingly, higher FPS windowed Triangle case perf drop is much higher than
the fullscreen Triangle one, so very high FPS can be one trigger for this.

Above regression percentages are from SKL i5 GT2, where the regressions are
most visible.  Regressions are visible on all of our GEN9 platforms, but
smaller on some of them.

On SKL GT2, in the windowed Triangle case:
* memory and GPU power usage decreased
* CPU power usage increased
So that in total ~4% more power is used in this case (according to RAPL).

SKL GT2 isn't TDP limited, so CPU side just using more power doesn't explain
the drop.


Median power usage in our full benchmarks test set didn't change, regressions
in all the other tests, than the very high FPS windowed Triangle one, are so
small that they can't be bisected, and the trivial windowed Triangle case isn't
that interesting use-case.

-> WONTFIX sounds fine as long as somebody checks what commit caused the perf
drop.

(I'm not myself looking at GFX perf anymore, filing this was the last remaining
item.)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
You are the assignee for the bug.
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-gfx-bugs/attachments/20181030/b9cdd40e/attachment.html>