On Thu, 01 Jul 1999 02:15:02 -0400, Matt Giwer wrote:
>	But then why not use the floating point package built into the
>P3 instead? Same processor, same speed UNLESS they MMX
>instructions cheat by being less precise? 

There would be some utility to the SSE instructions if they weren't 
single-precision.  The utility is obvious when you break apart the
nested acronym: SSE stands for Streaming SIMD Extensions.  SIMD,
of course, stands for Single Instruction Multiple Data, which tells
us why SSE would be nice if it had some precision to it: it can do
multiple operations at once.  In the case of the single-precision 
stuff they've implemented now, it can do four floating point adds,
multiplies, etc. at the same time.  Think how fast this would make, 
say, a dot product or a normalization.  It could be used now for
color calculations, since they're single-precision, but they're 
such a small part of what POV does that I'm not sure it would help 
enough to be worth the trouble.

One could try making a single-precision-only version of POV and then
putting SSE or 3DNow! instructions into it, but I'm inclined to believe 
that output quality would suffer.

Post a reply to this message