== Last week (Linaro Connect) ==
* Reran libav comparisons after Ira's fix for excessive promotion.
The vectorized versions are now at least as good as the non-vectorised
ones. Updated wiki page with new asm output and microbenchmark results.
* More work on SMS. I have some patches that wire up the ddg code
to IV analysis. It gave some nice benchmark improvements, but also
some regressions. Traced the regressions down to cases where the
schedule for small iis generated too many moves. E.g. in a small
microbenchmark, we were able to schedule 6 instructions with an
ii of 3 (i.e. in a loop iteration of 3 cycles), but then needed
to add ~9 moves in order to keep the dependencies correct.
We got much better code with a larger ii and fewer moves.
Wrote a patch to estimate how many moves would be added, and to try to
a larger ii if the number of moves is too high. This improved the
results for one benchmark independently of the iv patch, and had no
effect on the others.
Discussed this with Revital, who said that Mustafa had tried a similar
thing but seen no benefit.
* Got powerpc-ibm-aix5.3 bootstraps working. Needs a few local fixes
due to C++ bootstrapping. Used it to test a couple of preparatory
patches for the IV work. Submitted those patches upstream.
* Ran benchmarks with -fno-schedule-insns after seeing that the first
scheduling pass was responsible for the main NEON-vs.-non-NEON
regression in EEMBC. It fixed that case, but as expected,
made others worse. Mentioned this to Ramana, who pointed me at
-fsched-pressure.
Reran the benchmarks with -fsched-pressure instead of
-fno-schedule-insns. It too fixed the main regression,
and improved a couple of other tests too. It showed a regression
in another test though. Looked at that regression. It was a case
where many registers were live across a loop, but not used in it.
This was causing the loop to have a very conservative schedule.
It would be better to spill some of the other registers instead.
Wrote a patch to take loops into account, and it seemed to do
the right thing for EEMBC. Sent it to Andreas, after Ulrich
mentioned that he had been looking at -fsched-pressure problems
on s390. Andreas is away for a while, though, so I might put this
on the back burner until he gets back.
== This week ==
* SMS
* auto inc/dec
* libav, perhaps
Richard
Hi,
* committed upstream a patch that reduces over-promotion of vector operations
* started to work on a new version of the patch to change the default
vector size for Neon
* attended Linaro connect
Ira
* Committed a set of SMS patches to trunk and gcc-linaro branch.
* Implemented a hack to evaluate the potential of SMS on SPEC2006/libqauntum.
* involved in non linaro issue
== QEMU ==
* After discussion with Peter started writing QEMU fixup for 64bit
atomic helper version location.
* Sent fixes for soc-dma code to qemu list
* Trying to understand just how much of omap_dma's code is needed.
== Other ==
* Travelling to/from connect
* Wanted to dial into some of the seessions in Corpus and Magdelen
rooms but the remote audio from them was unusable.
Dave
Hi,
Libunwind:
* finished initial ARM support for remote unwinding (libunwind-ptrace)
Android:
* took a closer look at the debuggerd
* got the perflab benchmark running on my PandaBoard using Linaro GCC
Misc:
* remotely attended some Linaro Connect Android sessions
Regards
Ken
== GDB ==
* Created Linaro GDB 7.3 branch
* Ported all remaining feature patches from Linaro GDB 7.2
* Backported mainline patches to fix remote test issues:
- Fixed #804387 Shared library test problems
- Fixed #804392 Rebuilt executables not copied
- Fixed #804396 Spurious failures
* Committed mainline patch to fix dlopen test cases
for remote testing (#804387).
* Committed mainline patches to fix misc. other remote
test problems (#804396).
== Misc ==
* Attended Linaro Connect in Cambourne.
Mit freundlichen Gruessen / Best Regards
Ulrich Weigand
--
Dr. Ulrich Weigand | Phone: +49-7031/16-3727
STSM, GNU compiler and toolchain for Linux on System z and Cell/B.E.
IBM Deutschland Research & Development GmbH
Vorsitzender des Aufsichtsrats: Martin Jetter | Geschäftsführung: Dirk
Wittkopp
Sitz der Gesellschaft: Böblingen | Registergericht: Amtsgericht
Stuttgart, HRB 243294
I've updated:
https://wiki.linaro.org/RichardSandiford/Sandbox/NeonLibAv
so that it gives the output for current trunk, including Ira's commit
yesterday to reduce the amount of overpromotion. I also reran the
microbenchmarks. The good news is that the vectorised code is now
better in all cases than the non-vectorised code.
The biggest winner from last time was rgb24tobgr16_C(). It used to be
much worse with vectorisation due to lots of excessive widening.
Thanks to Ira's patch, the loop now looks pretty respectable,
and is ~3.25x faster than the non-vectorised code.
As well as using a more recent compiler, the new version also uses
-mvectorize-with-neon-quad. Once again it shows a significant improvement
over the default.
Richard
