-freduce-all-*

-freduce-all-*

[Michael T. Babcock]

I was doing some work with ATLAS and running the author's xmm_gc3 
benchmark and noticed the comments associated with the 
-freduce-all-(movables|givs) options.  For what its worth (best speeds 
are cited):

Without: (/usr/local/gcc-3.4.2/bin/gcc -O3 -fomit-frame-pointer 
-fforce-addr -march=pentium -mtune=athlon -ftracer -funit-at-a-time  
-fbranch-probabilities -fsched2-use-superblocks)
./xmm_gc3
ALGORITHM     NB   REPS        TIME      MFLOPS
=========  =====  =====  ==========  ==========
atlasmm       60   3000       0.700     1851.71

With: (/usr/local/gcc-3.4.2/bin/gcc -O3 -fomit-frame-pointer 
-fforce-addr -march=pentium -mtune=athlon -ftracer -funit-at-a-time  
-fbranch-probabilities -fmove-all-movables -freduce-all-giv)
./xmm_gc3
ALGORITHM     NB   REPS        TIME      MFLOPS
=========  =====  =====  ==========  ==========
atlasmm       60   3000       0.762     1701.05

Each compile above was preceded by one with "-ftracer| -fprofile-arcs 
-fprofile-values -fvpt" and then the program was run three times before 
recompiling with the above options and re-running.

Please CC on replies/comments.

Also note, I have no connection to the authorship of the benchmarks or 
software above.
-- 
Michael T. Babcock
| ||

Re: -freduce-all-*

[Toon Moene]

Michael T. Babcock wrote:
> I was doing some work with ATLAS and running the author's xmm_gc3 
> benchmark and noticed the comments associated with the 
> -freduce-all-(movables|givs) options.  For what its worth (best speeds 
> are cited):
> 
> Without: (/usr/local/gcc-3.4.2/bin/gcc -O3 -fomit-frame-pointer 
> -fforce-addr -march=pentium -mtune=athlon -ftracer -funit-at-a-time  
> -fbranch-probabilities -fsched2-use-superblocks)
> ./xmm_gc3
> ALGORITHM     NB   REPS        TIME      MFLOPS
> =========  =====  =====  ==========  ==========
> atlasmm       60   3000       0.700     1851.71
> 
> With: (/usr/local/gcc-3.4.2/bin/gcc -O3 -fomit-frame-pointer 
> -fforce-addr -march=pentium -mtune=athlon -ftracer -funit-at-a-time  
> -fbranch-probabilities -fmove-all-movables -freduce-all-giv)
> ./xmm_gc3
> ALGORITHM     NB   REPS        TIME      MFLOPS
> =========  =====  =====  ==========  ==========
> atlasmm       60   3000       0.762     1701.05

Note that you didn't use -fsched2-use-superblocks in your second 
compile, so they're not completely comparable and do not _only_ test the 
addition of -fmove-all-movables -freduce-all-givs.

That said, I think I'm going to propose a patch to remove 
-fmove-all-movables and -freduce-all-givs because they were only useful 
in an old, and soon to be replaced, part of loop optimization.

-- 
Toon Moene - e-mail: toon@moene.indiv.nluug.nl - phone: +31 346 214290
Saturnushof 14, 3738 XG  Maartensdijk, The Netherlands
Maintainer, GNU Fortran 77: http://gcc.gnu.org/onlinedocs/g77_news.html
A maintainer of GNU Fortran 95: http://gcc.gnu.org/fortran/