Hello,

I am trying to solve several quadratic problem of the form x'Ax where the xs are varying but A ( which is symmetric, dense) is the same.
Since I there will be far greater xs then the dimension of A I want to precalculate as much as possible. So, I first to:

Code: Select all

 
LLT<MatrixXd> lltOfA(A);
MatrixXd U  = lltOfA.matrixU();
U = U.triangularView<Upper>();  // don't know if it should be there


Then for each x:

Code: Select all

 
Ux  =  U * x;
xAx = Ux.squaredNorm();


However when I compare to just: (in a loop when A is 4x4)

Code: Select all

 
Ax    = A*x;
xAx  = Ax.dot(x);


I notice no increase in speed for the first case, even though U is triangular, A is full.

Am I doing something wrong?
or is eigen doing something smart with A*x?


thanks

For 4x4 matrices exploiting symmetry is slower because it prevents vectorization. Moreover, make sure you are using a Matrix<?,4,4> type for A and Matrix<?,4,1> for vectors.