From:	Brian Osman <osmanb AT rpi DOT edu>
Newsgroups:	comp.os.msdos.djgpp
Subject:	Re: MMX
Date:	Tue, 04 Mar 1997 08:57:00 -0500
Organization:	Rensselaer Polytechnic Institute, Troy NY, USA
Lines:	47
Message-ID:	<331C2A2C.52DD@rpi.edu>
References:	<c=GB%a=_%p=Indigo_Active_Vi%l=CRIANLARICH-970304101819Z-27 AT crianlarich DOT Indigo>
Reply-To:	osmanb AT rpi DOT edu
NNTP-Posting-Host:	darkwing.stu.rpi.edu
Mime-Version:	1.0
To:	djgpp AT delorie DOT com
DJ-Gateway:	from newsgroup comp.os.msdos.djgpp

Robert Humphris wrote:
> 
> I know that this is not related to the topic, but it may be something
> that should be
> thought about...
> 
> MMX whatever is stands for ( I have heard several versions ) is the
> addition to the Pentium
> Rob Humphris
> 

Firstly, it doesn't stand for anything. I believe it began as MultiMedia
eXtensions, but I've talked with people from Intel and it officially is
just MMX. (SSA - Society to Stop Acronyms?) :)

Anyways, even at Intel now they don't have compilers really. Given the
incredibly complex nature of MMX code, anything developed now will have
to be done by hand, in assembly. I don't know that the MMX instructions
can be easily implemented as simple extensions to C, given that they
do some wierd stuff not normally taken into account. eg: Packed
addition,
but either saturating OR overflowing. The "parallel" distinction might
work, but I'd still wonder if it will be possible to do that well.

This is the problem with all of the newer technolgies. Look at VLIW.
HP claims that it's the greatest thing ever, but only a handful of
companies use it, and only for very specific processors. Why? Compilers.
No one can write a decent compiler for it becuase every shred of work
that the current generation of processors do gets moved to the compiler.
DJGPP is a great compiler, but I don't even see the big guys
incorporating
any kind of MMX support worth having in the near future.

Wait, new thought. In talking to the Intel guys and thinking about this
some more, I just remembered/realized... MMX is not used for general
purpose
computing, so it is most likely that people will simply write various
routines to perform specific funtions (signal processing, graphics, etc)
in assembly, and then package them into MMX libraries. This is something
that sufficiently interested people could do NOW for DJGPP without even
needing access to the compiler. It's possible to the check the CPU ID at
run time, and select from amongst several routines to do the same thing.
Any volunteers to start coding stuff? (BTW: There are lots of examples
on
Intel's page, which would be a good starting point.)

Brian

- Raw text -