Rasterizing edge equations into covered tiles works very similarly to rasterizing pixels; what we do is compute lower and upper bounds for the edge equations over full tiles; since the edge equations are linear, such extrema occur on the boundary of the tile – in fact, it's enough to loop at the 4 corner points, and from the signs of the ‘a' and ‘b' terms in the edge equation, we can determine which corner.

I've always found this part to be unclear as it's not referring to the usual binning where you'd compute the bounding box of a triangle and test it against an even coarser tile. I assume the goal is to determine which corner is which, as in minimum/maximum. Aren't there three extrema, one for each of the edge functions?

What do we test here using m and M? What's the pass condition for a triangle to be associated with a tile for fine rasterization? I would argue that all that one needs is that the three computed maxima are larger than or equal to 0, ie M0 | M1 | M2 >= 0 which naturally has false positives. Further still, bounding boxes seem preferable as seem to avoid the acute sliver triangles pointing and slicing through far away tiles: