selfadjointView may be a bottle neck

[jcarpent]

crocoddyl/src/multibody/actions/impulse-fwddyn.cpp

Line 122 in d59b9be

d->Fx.bottomRightCorner(nv, nv).noalias() = a_partial_dtau * d->pinocchio.M.selfadjointView<Eigen::Upper>();

From my knowledge of Eigen, selfadjointView may lead to slow computations. It may be better to fill the entire mass matrix M and then performing the dense computations.

[cmastalli]

Thanks for the feedback. I will benchmark it asap!

[cmastalli]

@proyan you might want to handle this issue.

[cmastalli]

@jcarpent --- could I ask you (in your experience) when we should use selfadjointview (i.e., with which matrix dimension)?

As described here (https://stackoverflow.com/questions/39606224/does-eigen-have-self-transpose-multiply-optimization-like-h-transposeh), selfadjointview is not useful for small matrices. But it is not clear (in this post) what matrix dimensions are considered small