[RC5] To slice or not to slice for SuperSPARC

Remi Guyomarch rguyom at mail.dotcom.fr
Sat Jan 31 23:21:26 EST 1998


TwinTowers || wrote:
> 
> In build 390, there are four different binaries for Solaris platform,
> namely
> 
> rc5des-slice-mt
> rc5des-slice-nomt
> rc5des-noslice-mt
> rc5des-noslice-nomt
> 
> I understand the "mt" differences. But what about the "slice" part,
> especially for SuperSPARC? Can a "slice" run on a SuperSPARC, which is
> 32-bit? I thought "slice" version is only for UltraSPAR. But when I
> benchmarked both "slice" and "noslice" on a SuperSPARC, they ran fine but
> gave very different rates. "Slice" version achieved 247Kkeys/s whereas
> "noslice" only 89Kkeys/s. The strange thing is, the "slice" version
> displayed the progress meter up to "50%." and immediately terminated with
> the key rate.

"Slice" refer to the bitslice algorithm. This algorithm works on 32 and
64 bits platforms. In theory it will work on any platform, whatever the
word length.

About performance : the bitslice algorithm needs large caches to work
efficiently. The current bitslice version use various data strutures for
a total of ~73 KB on a 32 bits platform, and 145 KB on a 64 bits one.
This means that on old Sun box you (perhaps) won't find the bitslice
version faster than the non-bitslice one. That's probably why there is
two versions.

About the progress meter : it's only a cosmetic bug.

-- 
RÈmi		Don't waste your computer's time. Distribute it!
			http://www.distributed.net/
	    RC5 cores source code : http://wwwperso.hol.fr/~guyom001/


--
To unsubcribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest



More information about the rc5 mailing list