> Why would you compare with theoretical results when your rasterizer
> uses discrete and less accurate bisections? Shouldn't you compare
> with the rendering results? One way or the other 3/64 of a pixel is
> really small difference.
Well, your results differ from the original ones. I don't know
whether it differs to the better or to the worse (and, admittedly, I
don't know how good the bboxes computed with the original FreeType
code actually are). This is my only concern.
Ok. I'll work on this. I'll do upscaling/downscaling to get rid of rounding errors.
Given that bboxes are of really great importance, and that even
smallest rounding errors (to integers) can cause clipping, a
comparison to high-precision results is probably good, regardless of
the old or new code.
Hey, if clipping is a concern, the new code can be easily modified to round up the max values and round down the min values. That should guarantee no clipping, but the bounding box might be up to 3/64 of a pixel away.