>> If he does the computations using fixpoint arithmetics, he won't need any
>> hardware floating point at all, and no kernel modifications.

> This is not really an option, because signal processing code does lots
> of multiplies, and nowadays it is typical for integer multiplies to be
several
> times slower than floating point ones.

If the signal processing algorithm is appropriately scaled, then all of the
multiplications can be treated as multiplications by a factor of two.  What
does that sound like?  Full marks if you said "arithmetic shift left"! :)

--Kevin.