On 05/29/2012 11:48 AM, Rainer Hurling wrote:
On 29.05.2012 08:10 (UTC+1), Steve Kargl wrote:
sqrtl() is a bit special in that IEEE 754 requires that
it have no more than 0.5 ULP for all arguments in all
roundng modes. As to other functions, I've been trying
for 10+ years to get some of these into FreeBSD. I can
assure you that there has never been a rush of people
volunteering to help or a rush of people willing to fund
the development of the necessary code.
The almost complete absence of volunteers in this case is first of all
caused by the need of very special, deep knowlegdes on mathematical and
informatical issues. I bet there are only a few people out there, who
are really good in both (three of them are known to us ;) ). The
problems with quality standards in such libraries spreaded over most
OS'es and platforms support this view. (I am aware that these arguments
are kown and discussed before.)
The more important question to me is, how can the remaining (huge!) work
on the systems libm done in a foreseeable time. As one possibility,
wouldn't it be imaginable to pay some people for doing (some of) this
work (FreeBSD Foundation, Sponsors from industry and science etc.)? I
personally, as a private person, would be willing to pay a few hundred
dollars for it. Paying people for other tasks in FreeBSD is not entirely
uncommon, if I am not mistaken.
If there were a "freebsd-floating-po...@freebsd.org" mailing list to
discuss these issues, I would eagerly join. While I have never written
libm-style libraries myself, I do have knowledge of both computers AND
mathematics, and I do teach two university courses: "Numerical Methods"
and "Numerical Linear Algebra." And I have written quite a bit of
numerical code, so I do know about the issues of testing and debugging
such code.
I would be happy to take on any steep learning curve, and make
contributions, if only I were part of a group that would steer me in the
right direction.
Finally, Steve made a point that the way gcc multiplies complex numbers
is wasteful in computational time. I have been bitten by this. I wrote
some code that involved multiplying large numbers of complex numbers. I
first wrote it using the complex C99 code. Then I rewrote it using
double, writing out the complex multiplication long hand, and the
program went twice as fast! (It involved multiplying numbers that were
purely real or purely imaginary, so performing (x)*(I*y) by
(x+0*I)*(0+y*I) really did slow it down.)
Anyway, given that floating point is a big issue, and we are about a
decade behind schedule, really suggests that a
floating-po...@freebsd.org mailing list is needed. Or maybe there is an
existing freebsd mailing list you guys already occupy.
Stephen
_______________________________________________
freebsd-current@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-current
To unsubscribe, send any mail to "freebsd-current-unsubscr...@freebsd.org"
