* update the gl/math wrappers to make them more Eigen friendly
(and remove the useless, and not used, and somehow dangerous
*Direct and *E functions)
* add automatic reinterpret_casting from Eigen::Matrix to vcg
specialized types
Eigen's Matrix. Now the dilema is how to mimic those typedefs, using inheritence ?
or using the classic workaround: typename Point3<float>::Type; with Point3<T>::Type defined
to Eigen::Matrix<T,3,1>. Anyway currently I support both (and the inheritence scheme has
to be preserved for compatibility). The advantage of the second approach is that when
eigen has to evaluate an expression it uses an Eigen::Matrix<>, so it is probably better
to only use Eigen::Matrix but I'm not 100% sure that makes a big difference especially if
we add some automatic reinterpret_cast between Eigen::Matrix and vcg::Point*....
with both old and new version. The fixes include:
- dot product: vec0 * vec1 => vec0.dot(vec1) (I added .dot() to the old Point classes too)
- Transpose: Transpose is an Eigen type, so we cannot keep it if Eigen is used. Therefore
I added a .tranpose() to old matrix classes, and modified most of the Transpose() to transpose()
both in vcg and meshlab. In fact, transpose() are free with Eigen, it simply returns a transpose
expression without copies. On the other be carefull: m = m.transpose() won't work as expected,
here me must evaluate to a temporary: m = m.transpose().eval(); However, this operation in very
rarely needed: you transpose at the same sime you set m, or you use m.transpose() directly.
- the last issue is Normalize which both modifies *this and return a ref to it. This behavior
don't make sense anymore when using expression template, e.g., in (a+b).Normalize(), the type
of a+b if not a Point (or whatever Vector types), it an expression of the addition of 2 points,
so we cannot modify the value of *this, since there is no value. Therefore I've already changed
all those .Normalize() of expressions to the Eigen's version .normalized().
- Finally I've changed the Zero to SetZero in the old Point classes too.
* curently nothing change if you don't define VCG_USE_EIGEN
* make Matrix*, Point3 and Point4 derive Eigen::Matrix (still ugly)
* now catching all the dot products to replace them by .dot()
note that most of meshlab already compile
Paolo Cignoni
Marco Callieri
mtarini
Nico Pietroni
granzuglia
Federico Ponchio
Massimiliano Corsini