Since the row has "n" entries, and errors can accumulate in each add, something like (n+2)*eps may be more appropriate. (I wrote a similar library in Matlab.)