If you are really serious about the performance the best you could do
   is use a lookup table.  Pre-calculate the values, whack them into a
   table and use the incoming value as a lookup.  Costs you an extra
   memory operation, 'tis all.  For 16 bit a lookup table may be a bit
   excessive ;-) in this case you could make a smaller table and use a
   linear interpolation (or quadratic if you want to get really fancy)  -
   the trade off being the size of the table vs accuracy of the
   interpolation vs the time to do the calculation.

I don't know the audio stuff by heart (hope s.b. beats me to it on the
Amiga while I'm occupated with the A2410 X server), but the
ulaw-to-linear conversion is usually done this way... only 8 bit
input.

Hm... one could even generate the lower 8 bits as u_int8_t and the
upper 6 bits as u_int16_t, this way, to do 14bit output on the Amiga
hardware ;-) 

But the unsigned-signed conversion MUST be done by the CPU... you
don't want another memory access instead of a 1-clock-instruction, no?
At least on the Amiga, you have to copy the stuff to a chipmem buffer,
anyway, and can do the conversion while copying.

Regards,
	Ignatios Souvatzis