Both the C-reference and altivec versions of bsumsq_sub22_mmx() implement the
following:
v = ((p1f[i]+p1b[i]+1)>>1) - p2[i];
The mmx version does not add the one (+1).
- make twos[] static, so it doesn't have to be pushed on the stack everytime the function sumsq_mmx() is called
