Topcluster listing

Eugene Leitl eugene.leitl at lrz.uni-muenchen.de
Thu Sep 21 14:09:38 EDT 2000


Erich Strohmaier writes:

 > In practice however it seems to me to be much harder to utilizes the SIMD
 > units on current micro-processors than to use the "normal" FPU.

This is true, but still for people who can talk assembly very
substantial speedups are possible:

	    http://www.mime.up8.edu/~whygee/memoire/
especially
	    http://www.mime.up8.edu/~whygee/memoire/corps.html

(I'm not sure whether it's in there, but Yann Guidon did use MMX
acceleration for his lattice gas CFD).

 > On traditional vector systems (Cray and such) you have full support of the
 > vector units by vectorizing compilers. I am not aware of anything
 > close to this on micros. Quoting SIMD flops on a wide basis would raise a

There seems to be SWARC Scc by Hank Dietz

      http://dynamo.ecn.purdue.edu/~hankd/SWAR/Scc.html

which attemts to wrap the hardware in C extension abstraction but it
doesn't seem to be developed further. Too bad, I could use it.

 > lot of expectations and cause a lot of disappointments. Rather than
 > inflating "Macho-flops" I would rather be more realistic in my
 > presentations.
 > 
 > Even besides this argument, I will have to swith to a CONSISTENT way of
 > representation regardless what it is. Quoting one system with SIMD and the
 > next one without does not make sense.
 > 
 > Erich

_______________________________________________
Beowulf mailing list
Beowulf at beowulf.org
http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list