Don't have time to figure out the actual numbers myself, but here's an example doing the same for various transcendental functions: https://github.com/J-Montgomery/hacky_float_math/blob/623ee9...

a*(1.xxx/b) = a*(-1*1.xxx*log2(b)), which means you should be able to do a*(fma(b, magic, constant)) with appropriate conversions on either side. Should work in 32 bits for u8s.