[Developers] XLC/XLF on AIX: better optimisation

[Steve White]

Erik,

Please note that the darwin file suffers from the same problems as the
aix one regarding the IBM compilers, namely:

	-O5 is meaningful only if it is applied to all files, and at link
	    but that is *never* the case in the darwin architecture file
	    -O3 with -qhot is my recommendation

	-qmaxmem=-1 is meaningful only when optimising, and is turned on
	    by default by -O3.  So it's redundant at best.

	-qspill was inserted as a work-around for a compilation problem
	    in a previous version of LocalReduce.  I have not seen this
	    problem so far.  My guess is it no longer exists.  The 
	    docs say *not* to set qspill unless you must.  So it should go.


On  4.04.06, Erik Schnetter wrote:
> On Apr 4, 2006, at 11:37:18, Steve White wrote:
> 
> >Hi,
> >
> >Enclosed is a patch of the aix known-architectures file.
> >
> >The main complaint about the previous version was that it set the
> >Fortran optimisation to -O5 and the rest to -O2.  But in the docs, we
> >find that the further optimisation performed by -O5 is global in
> >nature, and only useful if -O5 is specified for C files *and* on the
> >link line.
> >
> >See http://publib.boulder.ibm.com/infocenter/comphelp/v7v91/index.jsp
> >
> >By making the optimisation flags to be
> >	-O3 -qhot
> >for all xl* compilers, I see about a 10% improvement in speed, and
> >Cactus compiles a lot faster.  The -qhot flag does the more aggressive
> >loop-unrolling which I think is the most important gain from -O4.
> >
> >Also, I got rid of the  -qmaxmem=-1 flag, because it is only effective
> >in optimisation, and the -O3 flag already sets it.
> >
> >I have doubts about the -qspill flag.  The default is 512, and the  
> >doc says
> >not to increase it unless you get a warning to the effect that more is
> >required.  But we have increased it by a factor of 20.  I wonder if  
> >this
> >doesn't hurt performance.  But I'm afraid I'll break compilation on  
> >some
> >system if I change it.  Seems this was put in around 2002 to get
> >LocalInterp to compile on Psi, but it compiles now without this flag.
> 
> I like the new settings.
> 
> I'd say to test the new flags (if you haven't already done so), e.g.  
> using the CarpetWhisky benchmark or using Christian's configuration,  
> and then commit it.
> 
> -erik
> 
> -- 
> Erik Schnetter <schnetter at cct.lsu.edu>
> 
> My email is as private as my paper mail.  I therefore support encrypting
> and signing email messages.  Get my PGP key from www.keyserver.net.
> 
> 
> 



> _______________________________________________
> Developers mailing list
> Developers at cactuscode.org
> http://www.cactuscode.org/mailman/listinfo/developers


-- 
Steve White : Programmer
Max-Planck-Institut für Gravitationsphysik      Albert-Einstein-Institut
Am Mühlenberg 1, D-14476 Golm, Germany                  +49-331-567-7625