Which norm of the residual is used in CULA Sparse?

[nene19861005]

Hi!
I use cula to simulate the incompressible flow with free surface. I need to solve the poisson equation every timestep. I do wonder which norm of the residual is used in cula sparse. According to my experience, a inproper relative tolerance may give a wrong result. For instance, a tolerance 1e-10 gives a wrong result while a tolerance 1e-15 gives a right result. The max tolerance on my card GTX 670 is 1e-15. Is there a strategy to specify the max value of the difference between two iterative step? Looking forward to any useful tips! Tahnks!