ftz 0.1

Denormal numbers (“denorms”) are extremely nonzero small numbers (less than 1.2-38 for single precision or 2.2e-308 in double precision) which are handled poorly on most modern archicectures. Instructions involving denormal operands may run as much as 100 times slower than those involving standard floating point operands [1], giving rise to very odd performance bugs.

This package contains optimized SSE code to flush denorms in numpy arrays to zero.

[1] http://en.wikipedia.org/wiki/Denormal_number