Uses BLAS but no mention of cuBLAS to speed things up? Does that mean the linear algebra wasn't big enough a component to merit optimizing on?