It’s also not the worst way of making sure new chips get supported relatively quickly. Those “giant” header files are mechanically generated from the RTL source

It’s also not the worst way of making sure new chips get supported relatively quickly. Those “giant” header files are mechanically generated from the RTL source for each GPU, it’s not like it’s some horrendously inefficient bloat form an outsource developer that doesn’t care about writing clean code or something.