I took the data file and did a simple matrix multiplication and let the compiler optimize it.

-O3 -ffast-math and using single precission float my machine needed about 3ms to transform them all.

Using a fixedpoint arithmetic works too but needs about twice the time.

I got about the same probably, last version (i managed to improve yet maybe by 10% - mostly by inlinig - it would be harder for me to get further

https://www.dropbox.com/s/0fuwprhulldbejp/tie56.zip

ps if someone could say this satrtup time measured here it would be really interesting for me to know..