Design draft: data structure specifications

These are some of the discussion topics around how new linear algebra data structures will be defined in QuTiP as part of my Google Summer of Code project.

After some discussion, including some further designs of how the new dispatcher methods would function, we are trying to pursue a “light” data structure strategy. This will hopefully have very lightweight instances of the data, and dispatch methods are simplified; mulitple dispatch is a difficult concept to fit into a true object-oriented style, and we believe that dispatcher methods will make adding new data types and dispatched functions significantly easier.

Originally written on the 8th of June, 2020.

Where and how are the structures defined?

  • Should be accessible in both C and Python without causing huge overhead in C.
  • It’s very easy to get this working in pure Python, where we have dynamic lookup on all attributes and dynamic typing, but that sacrifices C speed.
  • QobjEvo/CQobjEvo must be able to use them in an abstract manner (or what’s the point?)
  • For each data structure there should be associated behaviour (unary, binary, etc methods) which can have either concrete implementations or virtual implementations via a known intermediary type.
  • Each data structure should have an underlying C representation to allow it to be passed to C functions as a single object otherwise any dispatcher will always have an extra step where it ‘unwraps’ the type and constructs a new argument list: we want cdef CSR *csr_add(CSR *left, CSR *right) (with or without pointer indirection—I need to check Cython’s passing conventions), but instead we have
    def zcsr_add(complex[::1] dataA, int[::1] indsA, int[::1] indptrA,
                 complex[::1] dataB, int[::1] indsB, int[::1] indptrB,
                 int nrows, int ncols, int Annz, int Bnnz, double complex alpha=1)

Heavy vs light

Each data structure type must at least contain the fields that are necessary for it to store all of its data, e.g. Dense must have at least a double complex * pointer (or Cython memoryview) and information on the shape of the matrix, and CSR must have at least data: double complex *, indices: Py_ssize_t * and indptr: Py_ssize_t * pointers and its shape 2-tuple (Py_ssize_t, Py_ssize_t). If they have only these, they are “light”, i.e. they have very few methods attached to them, and we get behaviour from them by calling external functions.

The alternative is a “heavy” data structure, where all of its methods are attached to it.

Points to consider:

  • Heavy is “more Pythonic”, with behaviour attached to a particular class.
  • In the heavy model in C, how does somebody access CSR * Dense? Which class is it attached to, and how do we know what the type of the output is?
  • What is in charge of dispatch in the heavy model in C? How do we keep that type-safe?
  • In the light model, how do we access mathematical behaviour from Python? Do we have to make a Python wrapper class around every type and fill in every function? How is dispatch done in this model?
  • In either model, what is the C type of the output of any of these functions? If all types are instances of a parent class, how do we register specialised methods?

Data instantiation and ownership

Last, and most low-level, how do we manage data storage and ownership? For C speed and simplicity reasons, it’s probably preferable to store everything at C level behind a pointer owned by the data structure. We probably don’t want to allocate numpy arrays for every operation at C level, considering we’re just going to be iterating over contiguous regions of memory anyway and that is unnecessary overhead. Also, it adds further boiler-plate to every function operating on the arrays at C level, because they always first have to unwrap the data down to a raw pointer in order to do anything.

If we do back everything with numpy arrays, then ownership is easy because numpy and Python’s GC take care of everything for us.

If not, the ownership semantics at C level are simple—all data structures own their own data and will free that data once they go out of scope. On instantiation from C code, they take ownership of the memory they’re pointed to, and will free it when Python deallocates them using Cython __dealloc__. On instantiation from Python code, they take ownership of their data if copy=False and create a copy that they own if copy=True.

However, in order for other parts of the library to specify specialisations over certain data structures from Python space, we should expose the underlying data components as numpy arrays when requested. If everything is backed by numpy in C code, then it’s already solved, but if not, then ownership now is perhaps non-trivial: if the returned numpy array is a view onto the internal pointer, then what if our data structure gets deallocated before the numpy array does? Should someone in Python be allowed to mutate our data array from under our feet?

Two possible methods are probably both fine here:

  1. exposed numpy arrays are non-writeable views onto our data. __dealloc__ is modified to transfer data ownership to the created numpy array on structure deallocation, as we can be certain that the numpy array will always out-live us as long as we maintain a reference to it after creation (i.e. we use a strongref cache not a weakref one), and then we gain numpy/Python GC semantics. We set the non-writeable flag on the created numpy array to prevent someone in Python screwing with us in C code (or I suppose we could allow it so people can specialise in-place operations?).

  2. exposed numpy arrays are writeable copies of our data. We copy our data into a new numpy array every time Python requests a C-backed data attribute. We don’t have to worry about any ownership considerations at all, but in-place operations in Python become impossible and there’s a copy penalty every time Python accesses a data attribute on us.

I would tend towards method 1, and having written this, I suspect that my concerns about keeping the numpy arrays immutable are unfounded—we can allow them to be modified in-place. CQobjEvo and such like will by default make copies of passed data structures, mediated by some sort of own initialisation parameter.