> The lesson here just reinforces the adage about letting the compiler find and transform your source code into vectorized code with the programmer giving hints

> The lesson here just reinforces the adage about letting the compiler find and transform your source code into vectorized code with the programmer giving hints or small rewrites as needed. This has been my M.O. for a number of years, but sadly on modern platforms there seems to be little logic to which code is fast and which isn't. Even if the compiler emits vector instructions, that is no guarantee that the code will be particularly fast IME. Things like cache misses play a role obviously, but sometimes it's just a plain mystery why fast-looking is a bit slow. For that purpose I keep a benchmark of various ways of doing common array operations that I can run on each platform, it's quite handy.