Sparse Solver support on Nvidia GPUs

[Sameer Agarwal]

Hi Folks,

Mark Shachkov (with some help from our friends at nvidia) has just committed support for sparse cholesky factorization on Nvidia GPUs using the cuDSS library.

This gives Ceres Solver the ability to use a GPU to perform the sparse cholesky factorization in SPARSE_SCHUR &  SPARSE_NORMAL_CHOLESKY. Since cuDSS also supports single precision factorization, mixed precision solves are also supported.

The documentation still needs updating, but once you have built ceres with cuDSS (you will need the latest version 0.3.0) you should be able to set Solver::Options::sparse_linear_algebra_library_type to CUDA_SPARSE and use this new functionality. 

If you have access to an  nvidia GPU, it would be great if you could try solving your problems on the GPU and give us feedback on its stability & performance. 

Thank you,

Sameer