[Bug 107223] 50% perf drop in SynMark Fill* tests (E2E RBC gets disabled?)

[bugzilla-daemon at freedesktop.org]

https://bugs.freedesktop.org/show_bug.cgi?id=107223

            Bug ID: 107223
           Summary: 50% perf drop in SynMark Fill* tests (E2E RBC gets
                    disabled?)
           Product: Mesa
           Version: git
          Hardware: Other
                OS: All
            Status: NEW
          Severity: normal
          Priority: medium
         Component: Drivers/DRI/i965
          Assignee: jason at jlekstrand.net
          Reporter: eero.t.tamminen at intel.com
        QA Contact: intel-3d-bugs at lists.freedesktop.org

Setup:
* Ubuntu 16.04
* kernel & X server built from Git
* Modifiers support enabled (requires git version , or latest X release, which
isn't yet in distros)

There's a large performance drop visible in many onscreen tests when using
graphics stack supporting modifiers.

Performance drops on all GEN9+ platforms.  Drop is largest in SynMark Fill*
tests on eDRAM machines where the drop is 50%.  On GT2 machines they drop ~30%
and on BXT 20%.  This drop is visible both with Unity desktop and when using
just plain X server.

Smaller performance drop is visible in most of the other SynMark tests.

I bisected the regression to following commit:
-------------------------------------------------
commit a26693493570a9d0f0fba1be617e01ee7bfff4db
Author:     Jason Ekstrand <jason.ekstrand at intel.com>
AuthorDate: Wed Jun 6 10:24:01 2018 -0700
Commit:     Jason Ekstrand <jason.ekstrand at intel.com>
CommitDate: Thu Jun 7 11:23:34 2018 -0700

    i965/screen: Return false for unsupported formats in query_modifiers

    Cc: mesa-stable at lists.freedesktop.org
    Reviewed-by: Lionel Landwerlin <lionel.g.landwerlin at intel.com>
-------------------------------------------------

Based on the size of the regressions, and GeomPoint improving with the change
(only test which regresses with RBC), I'd say that above commit disables E2E
RBC for these tests.

Any idea why?  Rendering looked fine before this change.


On GEN9+ GT2 machines, there's also 3-4% performance drop visible in GfxBench
onscreen ALU2 & Tessellation tests from this commit, and maybe 1% in GfxBench
Manhattan 3.0, but those regressions are slightly less than half of what the
E2E RBC improvement was.


Interestingly, this same commit also improved SynMark VSInstancing performance
noticeably on GEN7 & GEN8 (which don't support RBC).

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20180713/b6fc79f0/attachment.html>