Re: [Plib-devel] ssgVtxArray fixes.

[Eero P. <epa...@ko...>]

Steve Baker wrote:
> 

> 
> Well, doing *math* on ints is probably faster than shorts - but this is largely
> a matter of RAM bandwidth and cache sizes and such.  I'm no expert either - but
> I was under the impression that for simple indexing and such, shorts can be
> faster.
> 

At least Pentium Pro (and I suspect this might be true for
Pentium II / III also) can suffer badly if memory accesses
to 32 bit and 16 bit quantities are interleaved.

Something to do with read accesses not been able
to check the outgoing write queue and thus requiring a 
queue flush, or something like that...

Still the smaller memory footprint might be worth the trouble,
but I have seen this problem as spikes in the VTune display.
(I have also seen spikes caused by cache misses, and 
they can be worse)


		Eero