Are these Halide speed up numbers typical? Are there certain programs where Halide doesn't lead to that big of a performance, and optimized C++/Assembly code is more performance? Or is Halide always a win?

I am curious about the definition of 2.75x less code. For less amount of code, are we talking about the number of lines of code before or after compilation?

What does Halide actually do to beat (for example) hand-tuned assembly? It doesn’t seem like the abstractions discussed in this lecture would necessarily help, without a lot of tuning…. (Also interested in how one goes about running such a study like this!)