I’m afraid you currently can only map buffer objects to CUDA. Hence the textures need to be first read to one or multiple pixel buffer objects which are then mapped to CUDA with cudaGLRegisterBufferObject followed by cudaMapBufferObject. This includes, however, an OpenGL internal “memcpy” (though the used OpenGL commands don’t have such a name but glBindBuffer followed by glReadPixels or glGetTexImage, as described in the GL_ARB_pixel_buffer_object specification).
In addition, I presume that the OpenGL framebuffer objects can’t be mapped to CUDA in the future either because they are actually merely a client-side encapsulation of a number of texture level image or renderbuffer handles. Instead, if I’ve understood the NVIDIA folks right, it should be possible to use the OpenGL textures on the CUDA side in a future release. What I don’t know, is whether they can be both written to and read from by CUDA, or only used as data source.