void axpby( const ConstScalarType & alpha ,
            const ConstVectorType & X ,
            const ConstScalarType & beta ,
            const      VectorType & Y )
{
  typedef AXPBY< ConstScalarType , ConstVectorType , VectorType > functor ;

  parallel_for( Y.extent(0) , functor( alpha , X , beta , Y ) );
}
void scale( const ConstScalarType & alpha , const VectorType & Y )
{
  typedef Scale< ConstScalarType , VectorType > functor ;

  parallel_for( Y.extent(0) , functor( alpha , Y ) );
}