[daip] Re: AIPS compilers

[Eric Greisen]

Daniel C. Homan writes:

 > This is Dan Homan (formerly of NRAO-CV, now at Denison University in
 > Ohio), and I am trying to get AIPS (31DEC03) installed on some linux
 > computers here.  I get it to install fine, but I get quite different
 > speed results depending on whether I use gcc 2.95.3 or gcc 3.2.2.
 > 
 > gcc 3.2.2 seems to produce a much faster AIPS installation, and this is
 > contrary to what the compiler webpage says...
 > 
 > http://www.aoc.nrao.edu/aips/compiler.shtml
 > 
 > I am pretty sure I set up the FDEFAULTS.SH file correctly in both cases
 > (just uncomment the stuff for 2.95.3 and comment out the stuff for 3.2),
 > but the DDT tasks take nearly twice as long when AIPS is compiled with
 > 2.95.3.  (An AIPSmark of 121 with gcc 3.2.2 vs. an AIPSmark of 68 with
 > 2.95.3).  Any idea what the problem might be?

      Not really - I would worry that somehow you did not end up
optimised in 2.95.3.  What sort of computer and OS are you on?
If you have the INSTEP2 or INSTEP3 or 4 .LOG files you could send me
some snippets (please not all) and I could check.

 > 
 > The reason I am trying 2.95.3 is that I was a bit unclear as to whether
 > the 3.2.2 compiler is safe for AIPS.  The compiler webpage gives
 > different information than the main AIPS webpage.  (The compiler webpage
 > says 3.2.2 is incorrect for BPASS, but the main page says 3.2.2 is OK.) 
 > Please let me know what the current wisdom is...  If 3.2.2 is safe, do
 > you know if 3.2.3 is safe (this is the native compiler on Redhat's
 > Enterprise linux v3)?

     The 3.2.2 now works okay with a couple of no-optimize things
added to OPTIMIZE.LIS.  I guess the compiler page did not get updated
properly.  I think 3.2.3 is okay too.  3.3.x (x > 2) has show real 
problems.  On my test machine - a dual headed Dell 2.8 GHz and my own
machine (Dell 1 head 1.3 GHz), 2.95.3 is faster than 3.2.2 is faster
than 3.4 by clear differences.  These are all RedHat 9.

 > 
 > Lastly, has anyone tried the Intel compliers?  I've got version 8 of
 > both the C/C++ and Fortran compilers on my system, and I bet that would
 > make AIPS scream.

We did try and found on average about no difference.  VTESS went much
faster but IMAGR went slower.  I think you will see a case for Intel
in the FDEFAULTS.SH file.  Be my guest and try it.

Eric Greisen