Thank you @whuaegeanse. CGNR is technically not a sparse direct solver, so the current option description is not correct. I am still wondering whether CUDA_SPARSE would result in a speedup even without cudss? Is there a way to switch between the sparse direct and indirect CUDA solvers in ceres >=2.3?

Is there a way to switch between the sparse direct and indirect CUDA solvers in ceres >=2.3?

ceres-solver still can't leverage CUDA-capable GPU in context of pcg with implicit schur complement. There is an issue available to be tracked.

@DmitriyKorchemkin did some work in that direction, although it staled a while ago.

Thank you @S-o-T for the answer. To me the fix here looks correct. Would you be able to also confirm? Thanks.