home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!psinntp!kepler1!andrew
- From: andrew@rentec.com (Andrew Mullhaupt)
- Newsgroups: comp.lang.apl
- Subject: Re: Promoting APL2
- Message-ID: <1176@kepler1.rentec.com>
- Date: 15 Aug 92 20:48:21 GMT
- References: <168417F7D.JOHNSON2@vm.cc.purdue.edu> <1992Aug13.171145.24691@csi.jpl.nasa.gov>
- Organization: Renaissance Technologies Corp., Setauket, NY.
- Lines: 70
-
- In article <1992Aug13.171145.24691@csi.jpl.nasa.gov> sam@csi.jpl.nasa.gov (Sam Sirlin) writes:
- >
- >In article <168417F7D.JOHNSON2@vm.cc.purdue.edu>, JOHNSON@vm.cc.purdue.edu writes:
- >I don't know what would end up being most popular, but for me there
- >are a few neccessities:
- >
- >1. linear algebra - as much of eispack/linpack/lapack as you can get.
- >Perhaps you could link them in via some sort of quad na? I've
- >translated the qr and qz eigensystem paths into APL if you want them.
- >I have them in readable ASCII form.
-
- NO! NO! A thousand times NO!. These guys have a 3090, and they should use
- ESSL, which is provided by IBM. Although ESSL has some trivial flaws, it
- is probably the only commercial package which is as good or better than
- LINPACK/EISPACK/LAPACK. ESSL has the world's _best_ FFT's bar none. (what
- else would you expect from Winograd et. al.?) Last time I used it, IBM
- had gone to the trouble to write the required interface for all of ESSL
- routines to quad-NA. Use the IBM provided ESSL.
-
- Also: there was talk that OSL would be done in the same way - if so use
- that for all of your constraint-bound optimizations.
-
- >2. ode solvers - the more the better. Some sort of rk45 for example.
- >This also needs to include random colored noise generators (for
- >example the OU of ACSL). A stiff method would be nice.
-
- A painful issue with APL, since you have to call the function defining
- the DE from the solver. I've been told that APL2 can _now_ be called from
- FORTRAN, but I've never seen it. If you can't, it's too bad. If you can,
- your best bet is to use FORTRAN versions.
-
- >4. random number generators - for the various processes. At least
- >Gaussian and Poisson.
-
- ESSL provides a good uniform generator.
-
- >5. integration methods - I don't really need to much here, trapezoidal
- >rules and Simpson seem to work.
-
- ESSL provides several excellent methods. Linking in quadpack would
- be even better, but we run into the "can FORTRAN call APL" problem
- again.
-
- >6. special functions - Bessel, erf, etc.
-
- Although you can profitably construct these using ESSL, you might
- be better off linking in PD FORTRAN code.
-
- >7. controls analysis - frequency response calulation etc. Well that's
- >what I want anyway...
-
- A good chunk of the foundationns you need for this is provided by
- ESSL's rich and "world's best" FT/Correlations/etc. functions.
-
- >8. data translation - functions for reading/writing data from/for
- >other languages (fortran). This also includes text manipulation - one
- >of APL's strong points since (at least STSC/J) can read exactly what's
- >in a file.
-
- But they live on a 3090, which might be running MVS. In which case
- the file system is a disaster beyond comprehension.
-
- A few more things about ESSL - ESSL is finely tuned to take advantage
- of the resources the 3090 offers, even across different models of the
- 3090. It can take advantage of vector heads, as well as run on multiple
- heads, for some functions. APL2 on the other hand only makes use of
- vectorization, at least last time I looked.
-
- Later,
- Andrew Mullhaupt
-