A not so fast implementation of cosine similarity in C++ and SIMD

> dot += A[i] * B[i];

Isn't it pretty bad for accuracy to accumulate large numbers of floats in this fashion? o.O In the example it's 640,000 numbers. log2(640,000) is ~19.3 but the significand of a float has only 23 bits plus an implicit one.

Python's floats are usually doubles by default, so it's mostly fine.

That said, yeah, that implementation isn't ideal. At a minimum, Kahan summation is usually free on large vectors (you're bottlenecked on memory bandwidth anyway), give or take the fact that you need to disable floating point re-ordering to keep the compiler from screwing it up and therefore have to order the operations correctly to make it efficient (see some other top-level comments about data dependencies as an example).