a question and an answer about optimizing compilers; your chance to rebut

Mitch Alsup uunet!ross.com!mitch
Wed Oct 20 11:40:17 PDT 1993


Having been both a provider of libraries and a user of libraries
over the last 20 years, I have found that 95% of the time, the user
wants the fastest algorithm which will achieve a reasonable result.
(Damn the torpedoes, full speed ahead).  Only when they see an
anomolous result do they want a more accurate algorithm.  

Many user would rather find a way to contort the data set for faster
run time, witness FFT's being used for matrix multiplication.

The only caviat is that when the user wants the more accurate
algorithm-set, (s)he wants no source cahnges to bring it about.
(a compiler option is acceptable).

On the other hand, a canned algorithm application (Excel) should
be engineered to give greatest accuracy over greatest speed!!!!
Thus we are lead to a distinction not often made in these discussions;
If the user has source code, (s)he wants highest performance, if not
(s)he wants highest accuracy.

Mitch ALsup
mitchaross.com



More information about the Numeric-interest mailing list