Besides the standard tricks to speed up the running time like threads and mmap, the biggest speed up is to use a custom hash table that is tuned for the data. I have implemented 5 different versions of hash tables including the Swiss hash table and more recent researched ones. None of them beats a simple linear probing approach. It’s because there’re so few keys, only couple hundreds of them. Tuning a custom hashing function to minimize key collision helps more than anything.

There are also lots of SIMD operations, which Zig has great support in the form of vector programming.

The only thing novel I did was to line up the temperature data from the back instead of from the front for the SIMD operations. That is for temperatures 12.3, 4.5, 67.8, 0.1, I line them up by the last digit after the decimal before packing them into a SIMD register. That way I know their exact positions in the register.