Summary: There exist low-level Cholesky decomposition routines in the nvidia samples, but they are undocumented. Are these routines safe to use in production, and will nvidia continue supporting them?
Details:
There exists sample code (cuSolverSp_LowlevelCholesky) that use low-level routines that decouple sparse matrix symbolic analisys, factorization, and solve. This is a useful interface for optimization settings where the sparsity structure remains the same between iterations, and hence we are interested in using them. However, the relevant routines are
not listed in the cuda documentation:
cusolverSpXcsrcholAnalysis
cusolverSpDcsrcholFactor
cusolverSpDcsrcholSolve
These functions are provided by the CUDA header “cusolverSp_LOWLEVEL_PREVIEW.h”
Since 1) the functions are undocumented, and 2) they are provided by a header that is explicitly named “LOWLEVEL_PREVIEW”, is it safe to use in production, and are they likely to be supported in the future?