[Beowulf] Optimized math routines/transcendentals

[Prentice Bisbal]

On 04/29/2016 02:23 PM, C Bergström wrote:
> Dear lazyweb,
>
> I'm working on optimizing for a non-x86 target processor and left no
> choice but to start writing optimized math functions.
>
> So far I've compared
> glibc
> Sun solaris open sourced code
> NetBSD mathlib (Which originated from Sun circa 1993)
> netlib
> {REDACTED-vendor-name} modified version of netlib
>
> Surprisingly, glibc does a pretty respectable job in terms of
> accuracy, but alas it's certainly not the fastest.
> ------------
> One of the faster versions tested sometimes incorrectly
> rounds for medium size and larger arguments, i.e. the least
> significant bit differs from the other versions.
> ---
> sin(6e5) =
> bfcb40318b8c1728 vs
> bfcb40318b8c1727 =
> -0.212896531236929975605676 (REDACTED-vendor-name) vs
> -0.212896531236929947850101 (nearest) vs
> -0.212896531236929960235352 (higher precision) ...
> -------
> Question: Is the above acceptable to your scientific codes in general?
>
Is this the sort of problem John Gustafson's Unums was meant to eliminate?

http://www.johngustafson.net/unums.html

Unfortunately, that seems to be the best (only?) page on the Internet 
about Unums, other than his Wikipedia entry:

https://en.wikipedia.org/wiki/John_Gustafson_(scientist)#Unums

--
Prentice