In the comparison it seems like Skia is quite a bit faster on the CPU than your solution, but at the cost of some robustness. Is this actually something that matters for typical applications like web browsers or UI toolkits?
So this is a little bit more complicated as a tradeoff. The main reason Skia is faster is that it's generating fewer primitives (quadratic Béziers instead of lines or arcs). Depending on the details of the renderer, you may also want to lower to a simpler primitive, and our method does both. If the goal is lines, then the additional cost of applying, say, Wang's method would make them take roughly the same runtime, though our method would produce fewer lines. If the goal is arcs, then I think our method comes out way ahead, as I'm not aware of anything else that generates arcs efficiently.
But it does mean that for outline-to-outline applications running on the CPU, methods like Skia's are likely better (though ours is still faster than the Nehab implementation).
I'm working on a cubic-to-cubic version also, but it's requiring some pretty deep math. See the Zulip thread to follow that work in progress[1].
