Just after some advice on the best way to take a cholesky decomposition of matrices around size 100x100 to obtain an inverse. Just to test I was trying a quick program (-02 -DNDEBUG -march=core2) with the following code for the inverse calculation bit:

Code: Select all

#define SIZE 100
Matrix A;

...

    MatrixXf identity(SIZE,SIZE); identity = MatrixXf::Identity(SIZE,SIZE);
    Matrix inverse;

    LLT lltOfA(A);
    for (int i = 0; i < SIZE; ++i ) {
        lltOfA.solve(identity.col(i),&inverse.col(i));
    }

...


As it is, it's much slower than the default inverse() method using the LU decomposition by about 40%.

Is it just because the cholesky hasn't been thoroughly optimised yet or is there a better way to attack this?

Daniel Stonier.

what about:

Matrix inv = Matrix::Identity();
A.llt().solveInPlace(inv);

there is no way Cholesky could be slower than LU.

ggael wrote:what about:

Matrix inv = Matrix::Identity();
A.llt().solveInPlace(inv);

there is no way Cholesky could be slower than LU.

Oh indeed!

I didn't realise you could pass the entire matrix into the solver. The block operations were taking up all the extra time.

Cheers.
Daniel Stonier.