🚀 go-pugleaf

In article <1992Aug5.190500.27872@csc.ti.com> fritz@ra.csc.ti.com (Fritz Whittington) writes:

>I believe a reasonable-quality MPEG encoder for something that meets
>the original question takes about 1200-1500 MIPS to do in real-time
>(sort-of VAX MIPS; obviously on special-purpose hardware the
>instructions per second could be lower).  So if the SPARC in question
>is anywhere in the 10-15 MIPS range, then 100 to 1 seems about right.
>
>Is it the Bellcore source code, or something developed elsewhere?

  I don't know if the posting refers to Bellcore code but the Bellcore
code we have looks like a lot more than 1200-1500 MIPS. It includes a
lot more floating point than appears to be necessary and the motion
search is a killer. I suppose floating point and integer performance
on the sparc are essentially the same so it wouldn't matter there, but
if the encoder can perform more efficient fixed point math there are
certainly savings.

  We have ported the Bellcore code to the Princeton Engine and 90% of
the time is spent in motion estimation. Just half pixel search alone
takes more than half the total search time! Using 704 processors (out
of a total possible 2048) we get about 10,000 MIPS and can still only
realize 10fps encoding. That may be pushed a little higher but 30fps
encoding will require at least 3X the instruction clock. The picture
quality of Bellcore's system looks pretty good, BTW. The jury is not
totally in on the quality but I would sell my Blockbuster video stock
if I had any...

 Herb Taylor
-- 
|\/\/\|\/\/\|\/\/\|\/\/\|\/\/\|\/\/\|\/\/\|\/\/\/|    _________________
| (.)(| (.)(| (.)(| (.)(| (.)(| (.)(| (.)(| (.)(.)   /  Think Parallel \
C     C     C     C     C     C     C     C      _) -|       Man!      /
 Herb Taylor, David Sarnoff Research Center, email: herbt@apollo.sarnoff.com