Daniel Egger <[EMAIL PROTECTED]> writes:
> Of course I meant channels instead pixels... However for simple
> mainpulations the shifting and masking is much more work than simply
> operating on one channel at once, especially since those simple
> operations have a much higher chance to be scheduled into different
> units in the CPU (superscalar processing is quite common) than code
> where you have a lot of depencies. So even if this approach would result
> in less instructions it probably might run slower.
I think I mentioned that this sort of stuff needs careful benchmarking.
The benchmarks we did for DirectFB (mostly on PPC and i686) showed a
noticeable speedup. But of course before one can even think about
applying such optimizations to The GIMP one has to profile typical
usage patterns and identify the real bottlenecks.
Gimp-developer mailing list