From the Ceres doc:

Currently this support is limited to using the dense linear solvers that ship with CUDA. As a result GPU acceleration can be used to speed up DENSE_QR, DENSE_NORMAL_CHOLESKY and DENSE_SCHUR.

So using CUDA or not is independent from linear_solver_type and yields improvements only for small problems (num_images <= kMaxNumImagesDirectDenseSolver) - at least until sparse solvers are GPU-accelerated as well (ceres-solver/ceres-solver#759).

Thanks, I create this request because in my experiments, for scenes around 2000 images(sparse problem with DENSE_SCHUR), mapper and bundle_adjuster is faster with gpu. Thanks for your reply, it seems better to close this reuqest and wait ceres 2.2 support sparse solvers

cc @sarlinpe It looks like CERES just recently got at least cuDSS sparse cholesky ceres-solver/ceres-solver@6fb3dae