by joshbu [MSFT] » Aug 26, 2005 @ 10:48pm
Prefetch the memory areas you are going to read.
Thus, in a blending or multisampling routine, yes, there would be some value to prefetching the destination. Unless the destination is the primary surface.
On XScale devices, we encourage driver writers to mark the primary uncached with combining. This gives us the maximum write speed possible without cluttering the cache. It also means that prefetching will have no effect.
General graphics tip: reading back from the primary is only going to get more expensive on PPCs/SPs. Don't do it.
We've also seen some good speed-ups making other buffers that are overwhealming written to (back buffers, or composition buffers) uncached with combining.
joshbu AT microsoft dot-you-know-where
Windows CE Software Design Engineer
“This posting is provided “AS IS” with no warranties, and confers no rights.”