Fast 1/X division (reciprocal)

2022-10-13

An answer to this question on Stack Overflow.

Question

Is there some way to improve reciprocal (division 1 over X) with respect to speed, if the precision is not crucial?

So, I need to calculate 1/X. Is there some workaround so I lose precision but do it faster?

Answer

The rcpss assembly instruction computes an approximate reciprocal with |Relative Error| ≤ 1.5 ∗ 2^−12.

You can enable it on a compiler with the -mrecip flag (you might also need -ffast-math).

The instrinsic is _mm_rcp_ss.