SEO

#1 (**permalink**) 01-16-2008, 02:03 PM

I need an ffs function for 64-bit arguments. Prompted by
the fact that lib ffs() only takes int arguments I wrote
a 3-line ffs64() function that uses popc, following the
classical example in the SPARC V9 Architecture Manual
and various other textbooks:

neg %o0, %o1
xnor %o0, %o1, %o0
retl
popc %o0, %o0

(zero inputs are filtered out before calling the function).

Much to my surprise this takes 1800ns to 2000ns per call
on a 1.2 GHz UltraSparc III depending on the number of
bits set, compared to 50ns to 125ns for the naive C
implementation (return ffs(low) if low != 0, otherwise
return ffs(high) + 32).

Digging a little deeper into the matter, I timed POPC by
itself, and sure enough it accounts for practically all
the time.

Why is POPC so slow?

dk

#2 (**permalink**) 01-16-2008, 02:03 PM

"Dan" == Dan Koren <dankoren@yahoo.com> writes:
Dan> Why is POPC so slow?

Others more versed in the various UltraSPARC implementations should
comment, but my understanding is that POPC is emulated in most
implementations.
--
Dave Marquardt
Sun Microsystems, Inc.
Austin, TX
+1 512 401-1077

Thread Tools
Show Printable Version Email this Page
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode