I often look at matrix packages, and always end up coding up something
suitable to a particular problem directly from Golub & Van Loan.
The packages seem over blown - you have to take on a lot of code for any
of it to work - and yet to lack the particular features the problem of
the moment needs (e.g. a fast givens for LS fits to geometric data,
where the large number of rows are generated from measurements and
consumed by the algorithm one at a time, and not stored anywhere).
The 'orthogonal' principle of STL might be worth putting on the mission
statement, even if the details of STL are not relevant.
Could we move from socialist central designs to a more flexible
heterogeneous (and redundant) free for all - small pieces of algorithm,
with code in several languages and test data and a theory document, and
clear descriptions of where it is suitable and not suitable, where it
fails, error analysis, flop count formulae etc. This sort of thing
would be of use to me, but I suppose it would be difficult to make money
out of it.
-- Stuart MacGregor
This archive was generated by hypermail 2b29 : Wed Feb 20 2002 - 03:20:05 EST