[Search for users]
[Overall Top Noters]
[List of all Conferences]
[Download this site]
| Title: | Mathematics at DEC |
|
| Moderator: | RUSURE::EDP |
|
| Created: | Mon Feb 03 1986 |
| Last Modified: | Fri Jun 06 1997 |
| Last Successful Update: | Fri Jun 06 1997 |
| Number of topics: | 2083 |
| Total number of notes: | 14613 |
419.0. "Scientific Math Library" by CURIUM::HERGT () Fri Jan 03 1986 10:24
I have been asked to examine the potential of a 'Scientific
Library' for the Vax systems, generally in the areas of
vector/matrix manipulation, statistics, signal processing,
image processing operations, etc. We currently offer a product
entitled SSP for the -11's and this would be a natural follow on,
and probable superset.
The intent is to build a library that would would price just
enough to cover administrative costs, therefore would be
used as a base to boost our capacity within the scientific
processing areas. The product should perhaps be a related to
such public domain packages as the VAXMath portion of the
SLATEC library, or the IEEE SIG, LinPack (especially the Basic
Linear Algebra Subprograms), Math Advantage (QTC), etc., and
could be used in a variety of future products.
I would like comments as to what should be included in such a library,
what packages you may have had exposure to and like or dislike,
what you see as a good implementation strategy (especially in light of
array and parallel processing).
I hope you may have the time for this. I will take what you have to
give in any format- scraps of paper, thesis, Electronic Mail, interview,
reference entries, replies to this notes, -
Please let me know what I can do.
Paul Hergt
VaxMail CURIUM::HERGT
MTS PAUL HERGT @MRO
Mail MR02-3/M91
DEC, 1 Iron Way, Box 1002, Marbroro, MA 01752-9102
| T.R | Title | User | Personal
Name | Date | Lines |
|---|
| 419.1 | | ANGORA::JANZEN | | Mon Jan 06 1986 12:28 | 5 |
| Make sure you can distribute matrix operations of a large matrix across a
cluster of CPUs. Support complex matrices, fourier stuff of different
levels, and like that. Maybe you should use the IEEE standard. This is
heightly recommended. Perhaps.
Tom
|
| 419.2 | | GALLO::APPELLOF | | Tue Jan 07 1986 09:30 | 28 |
| re .1
Distribute matrix operations across a cluster of CPUs? Great idea.
When we've solved the problem of how to do it, we'll have a pretty
nifty parallel processor implementation. I think IEEE standard depends
on our floating point hardware.
re .0
I used IMSL and EISPACK for a few years in quantum chemistry and
pattern recognition. Both were great. The nice thing about EISPACK
(and LINPACK too, I think) were the books which described the algorithms.
IMSL supports all of the data types mentioned in .1. It is a good idea
to have subroutines to deal with the different flavors of the same problem
(e.g. eigenvalues of a real-symmetric matrix, complex matrix, single/double
precision, etc.)
It would be a real good idea to have an appendix in the documentation which
describes error bounds for each subroutine. A good model I remember is
the CDC book on its scientific functions (SIN, EXP, etc.) which details
the error bounds for the whole range of possible arguments. The book on
EISPACK also has a good discussion of the effect of round-off errors if I
remember correctly.
Do you think that scientists would buy a package from DEC rather than from
IMSL? We actually contracted out to IMSL to re-code the scientific functions
on TOPS-10/20 FORTRAN.
Carl
|
| 419.3 | | GALLO::AMARTIN | | Sat Jan 11 1986 12:51 | 13 |
| Re .2:
The work IMSL did on the Fortran-10/20 math library was a waste of money.
It had to be redone by Mary Payne and Chris Smith before the routines
had a shred of:
1. Accuracy.
2. Correctness.
3. Efficiency.
If you consider the basic transcendental functions to comprise a "package",
then those scientists have already purchased a Digital authored "package".
/AHM
|