[voberoi]

I had the same takeaway when I read these papers. It felt like a game of whack-a-mole.

There's a massive performance benefit to doing this at the cost of implementation complexity. I haven't studied the implementations or tried my hand one, but I get the impression that these are tough to implement correctly in a way that takes full advantage of the hardware.

(In that sense, it's awesome that the researchers also did the legwork to implement and maintain a library!)