Re: benchmarks - sort

[Paul Thomas]

David,

In spite of promising to shut up about this, here is the result of 
modifying my sort DLD to deal with NaNs!

With your mod included and 10^5 randomly distributed elements, perverted 
by 1 NaN, octave's sort runs in 0.35s and mysort2 in 0.21s. If the 
operator does not deal with NaNs, this latter is 0.18s.  Of course, a 
preliminary sort that put NaNs, -inf and +infs in their correct places 
would more or less recover the lower figure.  Most of the performance 
improvement is obtained from the use of a pointer array, which is the 
only array being sorted. The algorithm that used a multimap, sorted 
double pairs and so ran up an extra overhead in movs.

In conclusion, I will use your sort algorithm but found mine to be an 
interesting exercise.

Best regards

Paul