Nuage wrote:
3) His patch compute gradients another way.
What is wrong is that it only check for 4 directions instead of 8. So I guess it
will help speed, but globule would not be able to walk trough a diagonal tight
path.
This leads me to an interesting idea about the gradient computations. In
updateGradientSmall you don't always have to add every neighbour of a
square into the listedAddr list. If you process a square and for example
the squares on the upper left and the lower left aren't obstacles you
can just update the square on the left, without having to add it into
the list. Every square which could be reached from there can also be
reached from one of the other two squares.
I wrote a quick hack against 0.8.15 (later releases didn't compile on my
machine and cvs has some strange gradient computation). It always
computes the gradient in the old and my new way and then compares the
results. The two gradients were always exactly the same and the gprof
results are quite impressive:
on the map WaterInTheDesert:
% self
time seconds calls name
12.79 7.16 1961 Map::updateGlobalGradientSmallOld(unsigned char*)
8.11 4.54 1961 Map::updateGlobalGradientSmall(unsigned char*)
on the map migration:
10.25 4.86 1786 Map::updateGlobalGradientSmallOld(unsigned char*)
7.74 3.67 1786 Map::updateGlobalGradientSmall(unsigned char*)
If you're interested, I'll clean up my quick hack and write a patch
tomorrow.
Simon
_______________________________________________
glob2-devel mailing list
[email protected]
http://lists.nongnu.org/mailman/listinfo/glob2-devel
