Thanks.
Probably something else happened, I think 2^16 of s16 table is (2^16) * 2 bytes.
But SGDK have us covered on this with the fix16Sqrt function.
The main headache when I tried it was that I was getting above the f16 limits with multiplication, that's why I have the division with F320. I also looked at the source of fix16Div and it looks like that div is forced in assembly. I haven't tried to remove the div with shifting, but if this is possible, vector normals can become too cheap.