home *** CD-ROM | disk | FTP | other *** search
- Xref: sparky comp.edu:1343 comp.lang.fortran:3178
- Path: sparky!uunet!cs.utexas.edu!qt.cs.utexas.edu!yale.edu!yale!gumby!wupost!waikato.ac.nz!aukuni.ac.nz!ecmtwhk
- Newsgroups: comp.edu,comp.lang.fortran
- Subject: Re: scientists as programmers
- Message-ID: <1992Aug26.082555.8222@ccu1.aukuni.ac.nz>
- From: ecmtwhk@ccu1.aukuni.ac.nz (Thomas Koenig)
- Date: Wed, 26 Aug 1992 08:25:55 GMT
- References: <1992Aug26.022458.12622@wl.com>
- Organization: University of Auckland, New Zealand.
- Keywords: supercomputer Fortran
- Lines: 33
-
- schuette@wl.com (Wade Schuette) writes:
-
- >For that matter, does anyone know if the current supercomputer
- >compilers are bright enough to watch their own output actually run
- >and learn from it so on their NEXT pass they can stop guessing and
- >make better decisions as to how to optimize the data-dependent sections
- >of code? Or are humans still supposed to do THAT as well?
-
- Current supercomputer compilers still depend to a large extent on
- local compiler options, like the C$DIR IVDEP - lines for the Crays.
- Some of them are a bit more sophisticated; a Fujistu VP will vectorize
- code like
-
- K = 1
- DO 10 I=1, N
- IF (A(I).GT.0.) THEN
- A(K) = A(I)
- K = K+1
- ENDIF
- 10 CONTINUE
-
- if you assert that I.GE.K for the whole body of the loop. This kind of
- assertion can be pretty dangerous, however; if you get it wrong, your
- program will give you wrong results.
-
- FORTRAN 90 will hopefully eliminate most of the need for sprinkling a
- program with options, once it comes into widespread use. One thing it
- will not eliminate, however, is the need to choose data structures
- carefully before ever writing the first line of code.
- --
- Thomas Koenig, ecmtwhk@ccu1.aukuni.ac.nz, ib09@rz.uni-karlsruhe.de
- The joy of engineering is to find a straight line on a double logarithmic
- diagram.
-