`cholmod` in BFF about boundary-first-flattening HOT 12 CLOSED

So it's just a class istantiated somewhere that is necessary to call the functions, there's no particular use of it.

from boundary-first-flattening.

The Common class is a wrapper for cholmod_common, which is struct that stores the state of the cholmod library & needs to be passed to all cholmod functions. There isn't really an Eigen equivalent for this class, I would simply get rid of it and replace the internals of the Sparse and Dense matrix classes with Eigen equivalents.

from boundary-first-flattening.

Be warned that Eigen, though somewhat easier to install than CHOLMOD, is significantly slower.

from boundary-first-flattening.

Hi again,

Do you reckon replacing the supernodal factoring with some of the ones provided by eigen would yield differences in quality of result?

from boundary-first-flattening.

No, I think Eigen is just gonna be slower. It's newer code, and not as well-developed as CHOLMOD/SuiteSparse.

from boundary-first-flattening.

The result should be the same, but supernodal is likely to be fastest. I believe Eigen provides a wrapper around the supernodal version of Cholmod.

from boundary-first-flattening.

That being said, the difference is only about a 2x slowdown for using Eigen over CHOLMOD (IIRC). So if you don't need this extra speed, it's probably no big deal.

from boundary-first-flattening.

Thank you for the support, I realize I'm asking too many questions.

from boundary-first-flattening.

By the way:

inline void Cholesky::buildSymbolic(cholmod_sparse *C)
{
    clear();

    factor = cholmod_l_analyze(C, common);
    if (factor) validSymbolic = true;
}

What would this be, what's a symbolic factorization in suitesparse?

from boundary-first-flattening.

This corresponds to the 'analyzePattern' function in Eigen's Cholesky implementation, while buildNumeric corresponds to 'factorize' function (http://eigen.tuxfamily.org/dox/classEigen_1_1SimplicialCholesky.html). The symbolic factorization determines the nonzero pattern of the Cholesky factorization, the numeric entries of the factorization are determined after computing this pattern. This is done to so that the symbolic (i.e., nonzero) pattern can be reused when only numeric values in the input matrix change.

from boundary-first-flattening.

Hi again,

I've seen you have few error checks, like the following:

inline void Cholesky::buildSymbolic(cholmod_sparse *C)
{
    clear();

    factor = cholmod_l_analyze(C, common);
    if (factor) validSymbolic = true;
}

inline void Cholesky::buildNumeric(cholmod_sparse *C)
{
    if (factor) validNumeric = (bool)cholmod_l_factorize(C, factor, common);
}

Eigen doesn't seem to provide a mechanism to check if the factorizations are valid.
Is there anything you can suggest for an equivalent error checking? The documentation doesn't say much unfortunately, I could return a default value but I don't want to propagate any error.

from boundary-first-flattening.

Take a look at the "Sparse Solver Concept" section here: https://eigen.tuxfamily.org/dox/group__TopicSparseSystems.html. "solver.info()!=Success" lets you check if the decomposition failed.

from boundary-first-flattening.