`lu` support for custom permutation (like in `cholesky`)

[jchen975]

It appears that lu doesn't support a custom permutation scheme (e.g., AMD) like in cholesky.

Is this correct? If so, could the support be added? (Please see the Discourse discussion on this and the resulted performance issue compared to Matlab.)

Thanks!

[stevengj]

Does SuiteSparse support passing a custom permutation to LU?

[jchen975]

I'll have to look into the UMFPACK source code more carefully (a little busy until next Monday). From the link to the discourse discussion, the lu call in Julia seems to automatically compute a fill-in reduction permutation and ignore the p = amd(A); lu(A[p,p]) permutation...?

Do you happen to know if the lu call in Matlab uses UMFPACK in SuiteSparse directly? In the Matlab code I'm using as a reference, it manually calculates the AMD permutation first, then uses LU to solve for x in Ax=b, which brings a significant improvement in solution time.

[andreasnoack]

It looks like it's supported, see https://github.com/DrTimothyAldenDavis/SuiteSparse/blob/e927f7a3fc82339755482e553df37d932ff30083/UMFPACK/Include/umfpack_qsymbolic.h, but we currently don't wrap these functions.