[ILUG] Advice much appreciated.

[Ronan Cunniffe]

On Thu, 17 Feb 2005, Paul Jakma wrote:

> On Thu, 17 Feb 2005, Ronan Cunniffe wrote:
>
> >   The only reason to use a 64-bit OS for number-crunching is if the
> > application specifically takes advantage of the 64-bitness.
>
> You dont even need a 64bit chip for that. Most CPUs seem to have had
> SIMD extensions for a while, and most of them seem to include 64 bit
> data types. (RGBA @ 16bpc is 64 bit). MMX gives you 64 bit integer
> math (or 32bit * 2 at a time, or 16bit*4 or 8*8.)

Yeah, I know.  What I meant was that every app has a built in
assumption about roughly how much RAM is going to be available.  Not
a lot of point making a 64-bit space available to an app with a
32-bit mindset.

As for SIMD, the sad thing is that outside video codec authors, just about
nobody takes advantage of this stuff.  How many people on ILUG looked at
the Cell architecture and went "yeah, I got to get me some of those
pipelines!"?

<ILUG poll>
as anybody used MMX/SSE2/Altivec stuff - designed their data
structures suitable for feeding to a SIMD device and written the code to
interface to the appropriate library? (or written the inline assembler
themselves, why not?)
</ILUG poll>

(Confession: I have studied various technical bits on what SSE can
do, discussed Altivec with one of my users who had a G4, found some
accelerated libraries, and have sketched out a few designs for pumping
stuff through. Actual working code? No.  And my other users?  Mostly
using f77... )

Now all this is a tangent to the original post which didn't specify
exactly what kind of number-crunching was involved....

(I come to realise, of course, that stuff like strcmp() could be
accelerated, and all you require is an appropriate glibc.  Is there a
glibc-for-mmx, or only glibc-for-386?)

Regards,

Ronan