That is not accurate. The performance numbers they reported were the idiomatic implementations. The optimized versions were discussed in a different section.
Except that the "idiomatic" implementations were not idiomatic, and the "optimized" versions were not optimized either. The whole paper was a huge joke.