It was wrongly calculated bucket number per side in cases when some
of segments is filling the whole bounding box across some of dimension.
Solved by limiting buckets at least to 1 in such cases.
Made some minor optimization such as:
- Avoid using "%" operation in loops, replace with a check
for index "overflow".
- Use pre-computed values for scaling feather coordinates
to 0 .. 1 space.
This allowed to reach couple of milliseconds of boost.
Another change is to use higher number of buckets (up to 512).
This doesn't took significantly more memory (like uses only 10MB
of memory for average splines) and allows to have 30-50x boost
for average splines.
Use dynamically calculated number of buckets for this, to be
sure segments would fit two buckets.
Also fixed intersection detection in some cases when edge is
shared between two buckets -- it is possible that such edge
would cross third bucket and intersect edge from there.
Real fix would be to find a point which is definitely now on loop
to be collapsed, but that's for a bit later. This commit should
remove possible stoppers.
This implements simple function which collapses internal loops
caused by self-intersections into a singularity.
This loops can't be removed because rasterizer expects points
of feather be aligned with points from spline itself.
When OpenMP is enabled, memory allocation needs to be protected.
Fixes bug [#32111] Memory management regression from svn_46520
projects.blender.org/tracker/index.php?func=detail&aid=32111&group_id=9&atid=498
Refactored code a bit to make naming a bit more clear and added a
function to create mesh from given display list rather than from
object's displist.
Tested using plain curves (which doesn't imply using derived meshes)
and curves with constructive modifiers (which are using derived meshed).
notes:
- uncomment #define USE_RASKTER in BKE_mask.h to use the previous mask rasterizer.
- slightly slower for regular masks but significantly faster for feather.
- main benefit is that it threads well so works nice for tile compositor.
- feather is lower quality, can use some improvements here.
- feather can also use some interpolation enhancements, will do later.
Problems were in the old multires loading system.
Actually, the sigsev itself was the easy part of the job (simply had to convert from tesselated data to polys/loops), but after that I was getting a horrible bunch of wild stray faces...
It finally turned out it was a mismatch in two different subsurf structs used while computing a mdisps layer from the multires DM, leading to getting complete random normals (null ones, NAN ones...), leading to complete dummy tangent space matrix, leading to absurds mdisps values...
Note: I also moved the copy of first layer's vertex and face data from old me->mr to mesh's v/fdata earlier in multire_load_old(), to be able to use general face_to_poly conversion function (later on we would have to do it by hand, the general function would erase our newly computed mdisps layer...).
Took me the whole week (something like 20h) to track this down: multires + subsurf = C nightmare!
