(Free42) roundoff for complex SQRT

[Werner]

Found this (see attachment).
Apparently performing the dot product as before, and carrying along a correction term that can be determined with the FMA (a bit like Kahan summation for sums), results in a dot product as accurate as if it were computed in double precision. Remark that the FMA is only used to determine the correction factor - when FMA's are used for the dot product itself it is much more costly.
Also, in this case, the computational penalty is 63% for n=50, and apparently going down to 25% for large n.
Keyword to google is 'compensated dot product FMA'
Cheers, Werner