I'm the core maintainer of Pythran, a python-to-native-code compiler
focused on numeric computations.
For more than 3 years, we've been using Boost.SIMD  as a generic
backend for SIMD code generation, without much worries. Basically our
compiler generates C++ code that includes Boost.simd calls. The big plus
to us are:
+ Performances. The generated code runs at the same speed as the one
written with intrinsics;
+ Target abstraction, esp. concerning the vector size and the various arithmetic operations;
+ Complete libm API, vectorized, which is a big plus when porting
existing code (or when matching an existing API as we do);
+ Header only, which actually makes it easier for us to ship;
+ Low requirements (it does not require us to ship a large part of
The compilation time is sometimes a worry, but it has decreased over the
year. We don't use complex data movement like shuffle or pack, so I
don't have any feedback on that.
All in one, if that ever matters, I would be happy to have it
integrated in Boost!