How do I implement vector <vector <Obj>> for Cython (C++ to Python)?

Question

See the code here: https://github.com/rootpy/root_numpy/blob/master/root_numpy/src/tree.pyx#L228

cdef cppclass VectorConverter[T](VectorConverterBase):
    int elesize
    int nptypecode
    Vector2Array[T] v2a
    __init__():
        cdef TypeName[T] ast = TypeName[T]()
        info = TYPES[ast.name]
        this.elesize = info[1].itemsize
        this.nptypecode = info[2]
    int write(Column* col, void* buffer):
        cdef vector[T]* tmp = <vector[T]*> col.GetValuePointer()
        cdef unsigned long numele = tmp.size()
        # check cython auto generate code
        # if it really does &((*tmp)[0])
        cdef T* fa = this.v2a.convert(tmp)
        return create_numpyarray(buffer, fa, this.nptypecode, numele, this.elesize)

Basically - this code is mean to convert a vector <object> (usually float or int) into an array [in particular, a NumPy array]. I have a hard time understanding it, but I digress.

I need to potentially extend it so that there's something that can implement vector<vector <object>> as well. I think it should be, pretty much, a copy of the code I highlighted in said link.

Where do I start?

Answer 1

I've solved this, and since I'm not DenverCoder9, here's how I did it, if it helps someone in the future even, obsoletely

cdef cppclass VectorVectorConverter[T](VectorConverterBase):
    int elesize
    int nptypecode
    Vector2Array[T] v2a
    __init__():
        cdef TypeName[T] ast = TypeName[T]()
        info = TYPES[ast.name]
        this.elesize = info[1].itemsize
        this.nptypecode = info[2]
    int write(Column* col, void* buffer):
        cdef vector[vector[T]]* tmp = <vector[vector[T]]*> col.GetValuePointer()
        #this will hold number of subvectors
        cdef unsigned long numele

        cdef T* fa

        #these are defined solely for the outer array wrapper
        cdef int objsize = np.dtype('O').itemsize
        cdef int objtypecode = np.NPY_OBJECT
        # it seems *tmp is exposed via tmp[0]
        # we want to create an outer array container that dataptr points to, containing pointers
        #    from create_numpyarray()
        numele = tmp[0].size()

        #define an (numele)-dimensional outer array to hold our subvectors fa
        cdef np.npy_intp dims[1]
        dims[0] = numele
        cdef np.ndarray outer = np.PyArray_EMPTY(1, dims, objtypecode, 0)
        cdef PyObject* outerobj = <PyObject*> outer # borrow ref
        # increase one since we are putting in buffer directly
        Py_INCREF(outer)
        # now write PyObject* to buffer
        memcpy(buffer, &outerobj, sizeof(PyObject*))

        # build a dataptr pointing to outer, so we can shift and write each of the subvectors
        cdef char* dataptr = <char*> outer.data

        # loop through all subvectors
        for i in xrange(numele):
          fa = this.v2a.convert(&tmp[0][i])
          # for some reason, shift isn't working, so we're directly shifting it ourselves
          #dataptr = shift(&dataptr, objsize)
          create_numpyarray(&dataptr[i*objsize], fa, this.nptypecode, tmp[0][i].size(), this.elesize)
        return sizeof(outerobj)