Cholesky factorizing of A*D^2*A’ ...


Is NVIDIA providing any code/implementation/description for Cholesky factorizing of AD^2A’ form of matrix computation in CUDA?

AD^2A’ is symmetric and positive definite. This is important when implementing weighted least square.

For more information :

Thank you.

I haven’t seen anything that specific as part of the samples with the CUDA toolkit. That paper you mentioned is quite detailed though. I’d imagine you’d be able to implement it without too much trouble. Heck you could try e-mailing the author and ask if he still has the source code, he graduated way back in '08:

As far as that particular form, that paper is your best bet.

cuSPARSE has cusparseScsric0, which computes incomplete-Cholesky factorization with 0 fill-in and no pivoting. Also has the incomplete LU decomp.

I also have CUDA dense matrix version of the Cholesky factor, and a dense version of computing the inversion (assuming the correct type of input).

My work involves writing CUDA versions of convex optimization functions, so I work with this daily.