Mail Archives: pgcc/1998/06/29/22:48:33
Marc Lehmann wrote:
> it seems that, with improved scheduling and tuning, speed improvements are
> possible, at least on p-ii, where the slowdown is small. my tests indicate
> that the mmx unit on p-iis is suprisingly fast and can well execute a number
> of mmx instructions in parallel.
>
> this doesn't seem to be the case on k6 or 6x86mx chips, which perform quite
> badly :( i believe mmx on these chips is useless for normal code and normal
> compilers that cannot take advantage of the parallelisation. I guess thats
> true for the pentium-mmx as well (though I have no data)
PII can dual-issue most MMX instructions (but since these instructions
are later converted to some other instruction set, it's not very clear).
Pentiums really have only one restriction : memory and integer register
access can only be made in the U pipe. This processor is amazing, it can
shift 64-bit entities faster than 32-bit ones : an MMX shift with the
count in a register cost only 1 clock where the same shift cost 4/5
clocks using 32-bit standard registers... Even better, the MMX shift can
be paired (with anything but another MMX shift or pack) and the standard
SHL reg,CL can't, mouarf !
The 6x86MX simply can't dual-issue MMX instructions, because it has only
one MMX core. Same thing for the K6.
K6-3D and K6-2 are much stronger in this area. They have a dual MMX core
and will perform better on MMX.
--
Rémi Don't waste your computer's time. Distribute it!
http://www.distributed.net/
RC5 cores source code : http://altern.com/rguyom/
- Raw text -