I was just recently going through some old episodes from Software Engineering Radio when I came across this one episode featuring Casey Muratori, where he goes through some of his thoughts around his video from February 2023, titled "‘Clean’ Code, Horrible Performance". I was actually already aware of the video by this time, but listening through the episode gave me an itch to see these concepts in my reality, experiment them by myself.
Loop unrolling is not really the speedup, autovectorization is. Loop unrolling does often help with autovectorization, but is not enough, especially with floating point numbers. In fact the accumulation operation you’re doing needs to be associative, and floating point numbers addition is not associative (i.e.
(x + y) + zis not always equal to(x + (y + z)). Hence autovectorizing the code would change the semantics and the compiler is not allowed to do that.so if (somehow) the accumulator was an integer, this loop would autovectorize and the performance differences would be smaller ?
Very likely yes