Tuesday 15th November 2011

This month's meetings

WorkingGroups/ToolChain/Meetings
<< <  2011 / 11 >  >>
Mon Tue Wed Thu Fri Sat Sun
  1 2 3 4 5 6
7 8 9 10 11 12 13
14 15 16 17 18 19 20
21 22 23 24 25 26 27
28 29 30        

Attendees

Agenda

  • O3 as a performance theme - Follow-up on discussions after the Summit .
  • Other Connect Hacking session topics - Biggest regressions -
    • partial_partial PRE - http://gcc.gnu.org/PR38785

    • Loop unrolling at O2 ? Worth a set of benchmark runs to see what the status with this is ?
  • 886124 Slowdown in vpr on ARM due to using the constant pool over MOVW/MOVT.

  • Need for a day-light saving correction ?
  • Any other business.

Action Items from this Meeting

  • TBA

Action Items from Previous Meeting

  • TBA

Minutes

  • Unrolling at O2 vs. SMS at O2 - Richard pointed out that SMS is probably better at O2 than unrolling.
  • O3 as a performance theme - What are our goals ?
    • Write-up at the end of it stating that it is as correct as it is the worth thing.
    • What are the kind of performance benefits that we get at the end of this cycle ?
    • Correctness
      • ICEs are easier since it's large package.
      • Bootstraps at O3 ?
      • Test rebuilds at O3 from Mathias.
      • eglibc at O3 .
      • Packages we care about - GCC, GDB, eglibc, qemu, binutils, Qt (C++)
      • Benchmarks that have correctness checks.
      • Nice to have: OpenEmbedded at O3 as a testsuite for correctness ?

    • Performance
      • Not bringing the turning on SMS by default under this umbrella.
      • Understand and fix some of the top regressions at O3.
    • Michael to run benchmarks with unrolling -
    • Change times for this call - 9:30 a.m. UTC Tuesdays.
    • Ramana to check on the A9 description and mail relevant bits- Revital to check A8 versions.
    • Next highest priority ticket for someone coming free to look at would be the vpr regression.
    • 64 bit arithmetic regressions with Neon seems to show up. Would be good to look at that next time.
    • Ken and Michael to take care of correctness infrastructure issues.
    • Ramana to look at one of the major regressions before next time.

WorkingGroups/ToolChain/Meetings/Archive/2011-11-15 (last modified 2013-08-30 11:47:58)