middle products

[Paul Zimmermann]

       David,

> > and already save more than 33% at 55 limbs. Larger timings in the FFT
> > range (say 100,000, 1M, 10M, 100M and 1G limbs) would be  
> > interesting too.
> 
> 10^4 | 4.09e+07 | 2.53e+07 |  0.619  |
> 10^5 | 9.68e+08 | 7.99e+08 |  0.825  |
> 10^6 | 1.90e+10 | 2.70e+10 |  1.421  |
> 
> That's the karatsuba asymptotics starting to dominate. At these sizes  
> surely it is better to use the FFT wraparound than to use my mpn_mulmid.

don't you also have a Toom-Cook 3-way mulmid? Indeed it would be nice to see
the effect of the FFT wraparound too.

Paul