[daip] Moving user files
Roderick Johnstone
rmj at ast.cam.ac.uk
Tue Aug 30 15:38:41 EDT 2005
Eric
Thaks for all these instructions. I'll give them a whirl. Let you know
tomorrow what happens. Hopefully, then, we can get these old catlogues
nailed.
Roderick
On Tue, 30 Aug 2005, Eric Greisen wrote:
> Roderick Johnstone writes:
>
> > The Intel compiler is well known for having tons of high-level
> > optimizations switched on by default. Thats one reason its produces code
> > that runs so fast. At one time it used to miscompile a lot of stuff, but
> > that was version 7.x.
> >
> > Ideally, what you should do is send the compiler switches to make fat
> > binaries. These have two or three sets of machine code and take a
> > run-time decision which is the best to use. That way you get to have
> > code that runs full speed on the P4 using all its whizzy instructions
> > but also has generic code that will run on anything.
> >
> > This is documented in:
> > http://www.intel.com/cd/software/products/asmo-na/eng/219281.htm
> > http://www.intel.com/cd/software/products/asmo-na/eng/222300.htm
> >
> > Using a compiler switch something like -axNP, would be good (exact
> > options depend on which version of the compiler you have). This makes a
> > binary with generic x86 code (ax) and code optimised for P4 sse, sse2
> > (N) and sse3 (P). There are several possibilities for N and P, see first
> > html reference above. K would optimise for Athlon XP, but you can only
> > have 2 letters so someone has to make a decision which things to
> > optimise for!
> >
> > So, could I request please either:
> >
> > 1) You consider rebuilding aips with something like -axNP or -axKN
>
> I will look at this but it may take time. We tried a bunch of
> other stuff that did not work so it will take serious testing.
>
> We tell the compiler (version 9.0)
> COMPSWIT="-c -xW -w90 -w95 -cm -ip -save -align all"
> OPT0='-O0'; OPT1='-O3'; OPT2='-O3';
> and the linker
> LINK="-lsvml -Vaxlib -L/usr/local/lib -O3 -static-libcxa -save -align all"
> and the C compiler
> OPT0="-O3 -xW -ip"
> COMP="-c $OPT0 -I$INC -D_FILE_OFFSET_BITS=64 -DHAVE_LINUX_GLIBC"
>
>
> >
> > 2) You send me instructions on turning my binary aips installation into
> > a source code compiled one. Do I need to delete my 31DEC05 directory and
> > download the source afresh, and build or can I somehow coax the system
> > into rebuilding my binary intallation, given that I probably have to
> > tell it how to use a custom compiler? The default version we have is gcc
> > 3.4.3.
>
> cd $AIPS_ROOT
> source LOGIN.CSH
> $CDTST
>
> This is pretty easy - cd to your $SYSLOCAL and edit CCOPTS.SH,
> LDOPTS.SH, and FDEFAULT.SH. They select the compiler - choose 3.4.3
> which should be fine (3.4.2 had problems we found that they fixed in
> 3.4.3). In $SYSLOCAL, rename RSYNC.ME which is the flag that tells
> the MNJ to use rsync.
>
> cd $TST/$ARCH/INSTALL
> rm *.LIS *.LOG
> rm $LIBR/*/SUBLIB
> INSTEP2
> INSTEP4
>
> and you are done in, depending on the machine, 1-4 hours. This is
> probably better than running generic code from Intel. My experience
> was that it ran no faster than the GNU code.
>
> Eric Greisen
>
--
Roderick Johnstone Email: rmj at ast.cam.ac.uk
X-ray Astronomy Group Phone: +44 1223 766656
Institute of Astronomy Mobile: 07989 809095
Madingley Road, Cambridge CB3 0HA, UK Fax: +44 1223 337523
More information about the Daip
mailing list