Re: [eigen] Quaternion and expression template

[Mathieu Gautier <mathieu.gautier@xxxxxx>]

> So to conclude, don't worry, I'm pretty sure that "q = q1.conjugate() * 
> q2" is already as fast as a specialized "quat_mult_conj(q, q1, q2)".

Yes and no :) I am working on visual studio 2008 and the assembly code 
generated with eigen is not optimized. There are 3 temporary quaternion, 
and a call to each constructor, to conjugate() and to operator*(). When 
 I modify the Quaternion constructor from

  inline Quaternion(Scalar w, Scalar x, Scalar y, Scalar z)
  { coeffs() << x, y, z, w; }

to

  inline Quaternion(Scalar w, Scalar x, Scalar y, Scalar z) :
   m_coeffs(x,y,z,w){}

the constructor is correctly inlined in the assembly code, but the calls 
to conjudate() and operator*() remain.


Then, I just try q = q1.conjugate()*q2 with eigen and gcc and the 
generated assembly code is quite similar to the specialized function 
(that was your remark). So, the visual compiler seems to have some 
difficulties to optimize this code.

Third, I use a little Quaternion class which is not templated and 
perform the same operation ("q = q1.conjugate()*q2") with visual studio. 
The assembly code is then well optimized (no call, no temporary 
constructor). So, I think that the problem come from the template and 
the inlining. Moreover using EIGEN_STRONG_INLINE (__forceinline) instead 
of inline and a bunch of optimization flags does not help. Have you ever 
encounter this issue with eigen and visual?

--
Mathieu Gautier