On 29/08/2011, at 10:32 PM, Maciej Marcin Piechotka wrote:

According to random side (http://gruntthepeon.free.fr/ssemath/) not so

new computers can compute 15.5 milions of serial logarithms per second

(62 millions in total). I'd say that overhead of Integer might be much

bigger then cost of logarithm.

That's floating-point logarithms, not Integer logarithms.
Single-precision floats, at that.

The code in question does not link at optimisation level 4.
At least some of the benchmark results are impossible to believe:
   benching          cephes_sinf ..
   -> 12762.3 millions of vector evaluations/second
   ->   0 cycles/value on a 2000MHz computer