3 ms·
That's some good benchmarking however doing this kind of optimization at scale is impossible. Because 1- Going from correct to fast is simple but the other dir
by farhanhubble 3y ago
That's some good benchmarking however doing this kind of optimization at scale is impossible. Because
1- Going from correct to fast is simple but the other direction is full of misery.
2- Optimizations are super super input and hardware specific.
- ashvardanian 3y ago1 - Very true, but in some ways its helpful, as provides an opportunity to highlight some of the poorly setup constraints in downstream applications. Like changing the compiler used for a major project, and seeing a 1000 warnings at once. In this case, the applications that will suffer are the ones that compares floats and float-arrays without absolute and relative tolerance intervals. 2 - Also true, but the SimSIMD library covers the absolute majority of mobile, desktop, and server CPUs produced in the last years.