Premultiplied vs non-premultiplied colors in solid and gradient pictures

[Alan Cox]

On Gwe, 2005-07-01 at 09:04, Lars Knoll wrote:
> As I said, I can readd the checks without any problems, but the compositiong 
> operations will be a factor of 3 slower; at least on x86. Together with exa 
> the difference will probably be noticable.

Have you benched that on cache cold systems. The code you posted is
pretty basic and doesnt seem to use MMX or prefetch so it ought to be
pretty much memory dependant. I appreciate the code you posted is even
more importantly
portable but the native instruction sets for MMX etc do know how to do
clipping
prefetch and 8 byte wide operations all of which are ought to dominate.

I've been caught out by this a few times often in surprising ways - eg
on a PIV its faster to compute a sine than use a large lookup table!

Alan