I don't think it'll work. CUBLAS has a different memory management (on the GPU, of course), and a cuda* prefix for the functions.