Avoid float -> double -> float conversions: works on floats all the way down

... to invalid normals, DEM values...

... to nodata, NaNs, "folded" neighbours (when left and right are a the same
position...)