Re: [Gimp-developer] Gaussian/Tileable blur: Choice of IIR and RLE still necessary?

[gg <gg catking net>]

On 08/30/12 06:11, scl wrote:
> Hi,
>
> I'm currently working out a GUI brainstorm idea for the Gaussian and
> Tileable Blur dialogs. Both dialogs let the user choose between the
> algorithms IIR and RLE. The documentation says, IIR is faster on
> photographs, RLE faster on drawings. I've never found a difference in
> computing time between these both algorithms and wondered, whether the
> choice is still necessary with modern computer environments.
> What do you think about it?
>
> Thank you,
>
> Sven
> _____________

It would probably be useful if the doc was clearer about what ti meant 
by "drawings".  It makes sense that RLE would be faster in block 
graphics where there are large areas with exactly the same colour value.

If the user is "drawing" with rounded brushes with faded edges and 
softening things with gaussian blurs the drawing will not necessarily 
fit the criteria where RLE is faster.

"With modern computers" argument is countered by "on modern images". As 
the hardware gets faster the images also get bigger ( in both x and y) . 
As gimp progresses to higher bit resolutions any difference may become 
more relevant not less.

I'd suggest some timings on large photo and similar sized image half 
black, half white (extreme example of the block graphic case).

I'm also wondering about the IIR label. I have not dug out the code but 
I was not aware that gaussian could be mode by IIR filter, isn't it 
rather a kernel ie FIR method?

If it is IIR I should have a look , it may be useful in something else 
I'm doing.
OK,  a quick google reveals I'm correct, the IIR method is just an 
approximate method that makes the computation complexity independent of 
sigma. By corollary the approximation must get worse for larger blur 
radius. Though I don't know whether the eye will detect which is the 
"better", purer blur.

So IIR label is fine.



/gg
.