[Bug 90895] [IVB/HSW/BDW/BSW Bisected] GLB2.7 Egypt, GfxBench3.0 T-Rex & ALU and many SynMark cases performance reduced by 10-23%

bugzilla-daemon at freedesktop.org bugzilla-daemon at freedesktop.org
Wed Jun 10 08:25:39 PDT 2015


https://bugs.freedesktop.org/show_bug.cgi?id=90895

--- Comment #9 from Eero Tamminen <eero.t.tamminen at intel.com> ---
Did some quick tests on this with GfxBench ALU test, on HSW GT3e.  Weird bug...

Reverting Ben's commit fixes the regression, although on HSW the cost per frame
should only be couple of ifs (for gen8) and skipping over perf output code on
unused branch.

Just replacing the perf_debug() calls with puts() calls improves perf, so maybe
the issue is somehow related to code inside the ifs (size of it?), although
that code doesn't get executed.

Adding missing unlikely() to the if in perf_debug() macro gave similar
improvement.  Additionally removing gen8 ifs didn't change perf.

Then I tested most of the indicated tests on BSW, HSW GT3e and HSW GT2. Just
adding unlikely() to the second if in perf_debug() macro improved the
performance several percents in the tests on all these platforms.

-- 
You are receiving this mail because:
You are the QA Contact for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.freedesktop.org/archives/intel-3d-bugs/attachments/20150610/152b87bf/attachment.html>


More information about the intel-3d-bugs mailing list