brainstorms

Thursday February 10, 2005

fdlibm, kahan paper etc.

sometimes things get too indirect for me. i was reminded of the great and free SUN fdlibm library through an interesting, if heavy recent paper (in progress) by kahan (thanks henry): How Futile are Mindless Assessments of Roundoff in Floating-Point Computation?

[i am not an FP geek, i hardly write any FP code, but as an educated programmer, i try to follow his arguments. his engineering perspectives and his relentless critique of mediocre implementations in the industry make kahan papers great fun to read.]

kahan quotes from this and other papers:

The longevity of inaccuracies in numerical software by and for numerical adepts has ominous implications: Numerical software does not have to be very complicated to be difficult to debug by experts, practically impossible to debug by amateurs.

Name-calling makes the caller feel better without enlightening him.

The essence of civilization is that we benefit from others' experience without having to relive it.

What runs too slowly won't get run.

If we keep no records of our mistakes, how can we learn to avoid more of them?

[music suggestion while reading kahan papers: donald byrd at the half note cafe volumes 1 & 2, blue note RVG edition.]

(2005-02-10 20:15:24.0) Permalink Comments [0]