The math/region implementations run at about the same speed. However, the region method is now slower for situations were polygons do not overlap because it has to perform extra endpoint checks for zero area polygons. So, overall, the math version is better.
