* Add some const in function parameters, just to make sure.
* Cleanup the comments, because I got lost in my own code...
git-svn-id: svn://pulkomandy.tk/GrafX2/trunk@1881 416bcca6-2ee7-4201-b75f-2eb2f807beb1
* Split clusters according to their volume (narrow covering) instead of occurence count. This ensures clusters representating a lot
of colors are split; and a lot of pixels with near colors are grouped together, which is what we want for low-color work.
* Rework memory management of clusters. Some of them were not malloc'ed but put on the stack, and then inserted in the list. Also
avoid copying+deleting a cluster in CT_Get, return the original instead.
* Use an union in the cluster struct to pack it a bit, since some data is used only in the color reduction phase, and some only in
the palette lookup.
The result is maybe slower, but looks much, much better. Fixes issues 63 and 441.
git-svn-id: svn://pulkomandy.tk/GrafX2/trunk@1878 416bcca6-2ee7-4201-b75f-2eb2f807beb1
* Further optimization to the colortree, now allocated in one single chunk of 4090 bytes (2051 times better than original code!).
Not doing all the malloc/frees should be slightly faster, and it also helps with cache locality (the whole tree can fit in the
cache).
* Remove the useless color "balancing" multipliers that did more harm than good.
We still need some improvements on the median cut, or maybe switch to a smarter algorithm; or add some cheats. See the JFIF quirks
for an example.
git-svn-id: svn://pulkomandy.tk/GrafX2/trunk@1876 416bcca6-2ee7-4201-b75f-2eb2f807beb1
There is room for optimization and maybe bugfixing.
The cluster table size is now:
511*15 = 7665bytes, this is 1094 times better than the color table.
Please test and report your results !
git-svn-id: svn://pulkomandy.tk/GrafX2/trunk@1874 416bcca6-2ee7-4201-b75f-2eb2f807beb1
