equation. For the moment the implementation is slow as data copied
to/from GPU memory at each step.