PocketMatrix

>It may be down to the granularity of your timer. You might need I higher resolution timer.
I'm timing 100 and 1000 repetitions

>You could try copying longwords instead of shorts. This should 1/2 the amount of time spent in the loop. Good to see your source data is consecutive in memory, saving an add per line.

>As for the cache, that'll probably depend on the CPU. 32 bytes sounds a bit weedy!

What is it for the StrongARM in the iPAQs?

I mean the accuracy of your timer, not the number of your repeats. If the code took 1.98ms and the optimised version took 1.01ms, the timer might just output 1ms.
Dunno much about the cache of IPAQ, the MIPS used to have 4K cache (on PSX anyway...)