Tuesday 20 March 2012

This month's meetings

WorkingGroups/ToolChain/Meetings
<< <  2012 / 3 >  >>
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 31  

Attendees

  • Andrew Stubbs
  • Michael Hope
  • Ramana Radhakrishnan
  • Richard Earnshaw
  • Ulrich Weigand

Agenda

  • Review action items from last meeting
  • 64 bit arithmetic and Neon - follow-up if any
  • Shrink-wrapping - now in 4.7 - We should consider enabling it for FSF 4.8 and try to make it stable there and investigate a backport to Linaro 4.7 ( I am always concerned about changes to prologue and epilogue - testing effort might be high) . Is this worthy of a new blueprint ?
  • AOB

Action Items from this Meeting

  • RAID xor routines in kernel - Can we vectorize the original C code ? [michaelh]
  • Re-benchmark refreshed peeling patch. [michaelh1]
  • Create blueprints for items on Ramana's libav Wiki page [michaelh1]
  • Run h.264 libav benchmarks against samplemedia.linaro.org samples to discover regressions between different compilers [michaelh1]
  • Create backlog blueprint to review all existing inline asm optimization for ARM in libav. [michaelh1]
  • Create backlog blueprint to enable shrink-wrapping on ARM. [ramana]
  • Investigate and fix vectorizer ICE (LP #959242) [uweigand]
  • Review blueprints in NEW state at next performance meeting [all]

Action Items from Previous Meeting

  • ACTION: AS to investigate implementation possibilities, crafty, and post suggestions for debate upstream (eventually).
  • Create blueprints for each of these cases with libav [All]
  • RAID xor routines in kernel - Can we vectorize the original C code ? [michaelh]
  • michaelh to refresh the peeling patch and get some benchmarks behaviour.
  • Libav benchmarking and what we use as content - (http://samplemedia.linaro.org) -

Minutes

  • Neon vs. core register allocation
    • AS still working on it. Next step to be benchmarking; forcing certain register allocation patterns via asm statements in order to understand the headroom for improvement that this optimization has. After that, complete implementation and discuss on upstream mailing list.
  • libav optimization opportunities
    • Ramana created a Wiki page: https://wiki.linaro.org/RamanaRadhakrishnan/Sandbox/RRQ112ConnectLibavgcc46Reg

    • ACTION: Create blueprints for items on Ramana's page [michaelh1]
    • ACTION: Run h.264 libav benchmarks against samplemedia.linaro.org samples to discover regressions between different compilers [michaelh1]
    • ACTION: Create backlog blueprint to review all existing inline asm optimization for ARM in libav. [michaelh1]
  • 64-bit arithmetic and NEON
    • if-conversion patch approved; AS will apply shortly
    • load-immediate patch fixed, currently re-testing
    • 64-bit shift patch likewise
    • 64-bit NOT patch: reposted new version
    • 64-bit NEG patch: needs retesting after load-immediate patch fix
  • Calling conventions
    • GCC will automatically switch static non-escaped functions to hard-float calling convention. This is not reflected in the debug data (DW_AT_calling_convention), meaning that calling such functions "by hand" from the debugger may fail. Will fix when someone complains ...
    • Andrew suggests adding another calling convention variant that also passes 64-bit values in VFP registers.
  • Shrink-wrapping
    • Infrastructure present in mainline since 4.7, but not yet enabled on ARM. Ramana reports this needs support for the "simple_return" patterns. Implementation is most likely dependent on on-going prolog/epilog refactoring effort.
    • ACTION: create backlog blueprint to enable shrink-wrapping on ARM. [ramana]
  • AOB
    • ACTION: investigate and fix vectorizer ICE (LP #959242) [uweigand]
    • Ramana reports failing vectorizer tests on mainline (loops unexpectedly not vectorized); needs further investigation.
    • ACTION: Review blueprints in NEW state at next performance meeting [michaelh1]

WorkingGroups/ToolChain/Meetings/Archive/2012-03-20 (last modified 2013-08-30 11:50:20)