screen verts were being rounded so resizing a window would end up moving the border in one direction.
for this to work properly we should store floats internally but for now compensate for this.
report & fix from Martin Lubich (loramel)
Use double rather then floats, this doesn't use significantly more memory (as allocating a double buffer would), other vars in this function were doubles already so may even gain some speed.
Fix for one of the causes of crashing.
Applying armature deform wasn't thread safe since the pose bones had deform data written into them when deforming a mesh.
This fixes crashing immediately, on every render for me but blender still crashes calculating the subsurf sometimes.
New Feature:
VCS job type
Render a file (with dependencies) from a version control system (currently only supports subversion, but system is already generic).
On client, working path, server path and current revision can be guessed from data on disk or entered manually.
On slave, a working copy is created (if needed) where specified by the job and updated to the proper revision.
On master web page, job types now appear in the job lists. The job page shows the list of dependencies for "normal" jobs or versioning information for VCS jobs.
Limitations: Need to have command line tools "svn" and "svnversion". Working copy path must be the same on client and slaves (the client gets the job path relative to the working copy). When guessing, working copy path is set to the folder where the current file is (this can be changed manually after). On the slave, it will update the working copy AS SPECIFIED to the revision, so if that path is too deep, some dependencies will not be updated properly. Doesn't support mixed revisions (and will not give any warnings for that), it will always use the first revision specified by "svnversion"
Bugfix:
Thumbnail generation doesn't chew down memory anymore and always gives correct result (thumbnail on master especially could mess up between jobs with the name result filename)
large render sizes could cause an the threaded tile processor to hang because winx * winy wrapped into a negative value.
also convert winx/winy to floats before multiplying for vector passs.
having the blend file as a part of the __file__ variable is not essential, this is fixed in python 3.2 so add an ifdef and don't use the blend file path for py older then 3.2.
replacing the extension could remove the frame number added to a path when writing images, so just add the extension rather then replacing even though it gives odd names at times. eg: foo.png0001.tga
Simple python benchmark shows this to be about 3x faster in the case where an update isn't needed.
This also speeds up rna function argument parsing, since each arg in a function call did 2 string lookups on the context which were never needed.
running a shape actuator on a softbody would crash because it assumed the deformer was a BL_MeshDeformer.
Added TODO note, since it would be nice if softbody would work with shape keys too.
