Currently C and SIMD dont exactly match for some functions/cpus.
Functions that do odd width use a combination of C and SIMD and can show a 
subtle artifact.   See this chromoting bug 
https://code.google.com/p/chromium/issues/detail?id=493914

For this use case, its important that C and SIMD match.
Neither needs to be accurate, and different platforms can be different.  But 
the C needs to exactly match the SIMD.

Reduce tolerance in unittests to zero.
Change C to mimic SIMD, even if performance is substantially worse for C.