some iterator benchmarks

[Inge Wallin]

On Thursday, May 30, 2013 09:51:39 C. Boemann wrote:
> Hi
> 
> I did the iterator benchmarks and tried it with varying tilesizes
> 
> The findings are that for the nConseqHPixels type iterations the speed
> advantage of doubling the tilesize is a doubling of the speed
> But for non nConseqHPixels the adavantage is only 2%
> And for the random accessor the advantage is 25% 12% 10% (for each doubling
> the gain becomes smaller)
> 
> The disadvantages of doubling the tilesize is that it quadrouples the memory
> usaga every time a tile is committed to the undo buffer. Even in the best
> case assumption that would be very noticable in practical usage maybe not 4
> times more but even 2 times more would be bad.
> 
> We could I guess only double the width giving us the speed but not the
> height lessening the memory impact. That would give us non-square tiles but
> other than possibly some undetected bugs i don't think it would be a
> problem.

Assymetric tiles is something I thought of too.  It doesn't need to be 
width=2*height either.  There can be any factor, and the ultimate in 
assymetric is using x scanlines of the full image instead of rectrangular 
tiles.

And re the undo buffer, does it have to be aligned tiles in there?  why not 
make it a quadtree of pixels with varying tile sizes?