What's the "mini cache"? Is it built into the SA1110 chip? How
does one enable it?

Actually, it's hard to believe any cache could help with
this operation--a memory to VRAM copy--unless it were large
enough to cache the pixels for the entire source rectangle.
In that case, after the first iteration of the loop all
read operations would hit the cache.

	-- John


At 6:09 PM -0800 2/19/01, Tim Rowledge wrote:
>> > Keep in mind that this is _still_ with an unoptimized VM...the handhelds folks 
>> > haven't updated the toolchain yet to a compiler that can handle doing 
>> > optimizations. :(
>> 
>>   This might be the case that the optimizer cannot do much
>> because the operation is so memory intensive.  It could be 
>> 3-10% speed up? (just a wild guess)
>The optimizer ought to roughly double the trivial benchmark performance
>when it can be re-enabled, scaling from the performance of the two
>NetWinders I can access. The problem appears to be that the optimizer
>creates a code sequence with a store-multiple-float-register instruction
>that is no longer accepted by the FP emulator. The same code on a NW
>with an old kernel/distro is quite happy. :-(
>However, as you say the particular operation in question here is
>terribly memory bandwidth dependent and the SA1110 has only a 16Kb cache
>to help it - unless possibly someone has enabled use of the mini-cache
>intended to help exactly this sort of thing.