Implementing the buffer protocol

Cython objects can expose memory buffers to Python code by implementing the “buffer protocol”. This chapter shows how to implement the protocol and make use of the memory managed by an extension type from NumPy.

A matrix class

The following Cython/C++ code implements a matrix of floats, where the number of columns is fixed at construction time but rows can be added dynamically.

  1. # distutils: language = c++
  2. # matrix.pyx
  3. from libcpp.vector cimport vector
  4. cdef class Matrix:
  5. cdef unsigned ncols
  6. cdef vector[float] v
  7. def __cinit__(self, unsigned ncols):
  8. self.ncols = ncols
  9. def add_row(self):
  10. """Adds a row, initially zero-filled."""
  11. self.v.resize(self.v.size() + self.ncols)

There are no methods to do anything productive with the matrices’ contents. We could implement custom __getitem__, __setitem__, etc. for this, but instead we’ll use the buffer protocol to expose the matrix’s data to Python so we can use NumPy to do useful work.

Implementing the buffer protocol requires adding two methods, __getbuffer__ and __releasebuffer__, which Cython handles specially.

  1. # distutils: language = c++
  2. from cpython cimport Py_buffer
  3. from libcpp.vector cimport vector
  4. cdef class Matrix:
  5. cdef Py_ssize_t ncols
  6. cdef Py_ssize_t shape[2]
  7. cdef Py_ssize_t strides[2]
  8. cdef vector[float] v
  9. def __cinit__(self, Py_ssize_t ncols):
  10. self.ncols = ncols
  11. def add_row(self):
  12. """Adds a row, initially zero-filled."""
  13. self.v.resize(self.v.size() + self.ncols)
  14. def __getbuffer__(self, Py_buffer *buffer, int flags):
  15. cdef Py_ssize_t itemsize = sizeof(self.v[0])
  16. self.shape[0] = self.v.size() / self.ncols
  17. self.shape[1] = self.ncols
  18. # Stride 1 is the distance, in bytes, between two items in a row;
  19. # this is the distance between two adjacent items in the vector.
  20. # Stride 0 is the distance between the first elements of adjacent rows.
  21. self.strides[1] = <Py_ssize_t>( <char *>&(self.v[1])
  22. - <char *>&(self.v[0]))
  23. self.strides[0] = self.ncols * self.strides[1]
  24. buffer.buf = <char *>&(self.v[0])
  25. buffer.format = 'f' # float
  26. buffer.internal = NULL # see References
  27. buffer.itemsize = itemsize
  28. buffer.len = self.v.size() * itemsize # product(shape) * itemsize
  29. buffer.ndim = 2
  30. buffer.obj = self
  31. buffer.readonly = 0
  32. buffer.shape = self.shape
  33. buffer.strides = self.strides
  34. buffer.suboffsets = NULL # for pointer arrays only
  35. def __releasebuffer__(self, Py_buffer *buffer):
  36. pass

The method Matrix.__getbuffer__ fills a descriptor structure, called a Py_buffer, that is defined by the Python C-API. It contains a pointer to the actual buffer in memory, as well as metadata about the shape of the array and the strides (step sizes to get from one element or row to the next). Its shape and strides members are pointers that must point to arrays of type and size Py_ssize_t[ndim]. These arrays have to stay alive as long as any buffer views the data, so we store them on the Matrix object as members.

The code is not yet complete, but we can already compile it and test the basic functionality.

  1. >>> from matrix import Matrix
  2. >>> import numpy as np
  3. >>> m = Matrix(10)
  4. >>> np.asarray(m)
  5. array([], shape=(0, 10), dtype=float32)
  6. >>> m.add_row()
  7. >>> a = np.asarray(m)
  8. >>> a[:] = 1
  9. >>> m.add_row()
  10. >>> a = np.asarray(m)
  11. >>> a
  12. array([[ 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
  13. [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)

Now we can view the Matrix as a NumPy ndarray, and modify its contents using standard NumPy operations.

Memory safety and reference counting

The Matrix class as implemented so far is unsafe. The add_row operation can move the underlying buffer, which invalidates any NumPy (or other) view on the data. If you try to access values after an add_row call, you’ll get outdated values or a segfault.

This is where __releasebuffer__ comes in. We can add a reference count to each matrix, and lock it for mutation whenever a view exists.

  1. # distutils: language = c++
  2. from cpython cimport Py_buffer
  3. from libcpp.vector cimport vector
  4. cdef class Matrix:
  5. cdef int view_count
  6. cdef Py_ssize_t ncols
  7. cdef vector[float] v
  8. # ...
  9. def __cinit__(self, Py_ssize_t ncols):
  10. self.ncols = ncols
  11. self.view_count = 0
  12. def add_row(self):
  13. if self.view_count > 0:
  14. raise ValueError("can't add row while being viewed")
  15. self.v.resize(self.v.size() + self.ncols)
  16. def __getbuffer__(self, Py_buffer *buffer, int flags):
  17. # ... as before
  18. self.view_count += 1
  19. def __releasebuffer__(self, Py_buffer *buffer):
  20. self.view_count -= 1

Flags

We skipped some input validation in the code. The flags argument to __getbuffer__ comes from np.asarray (and other clients) and is an OR of boolean flags that describe the kind of array that is requested. Strictly speaking, if the flags contain PyBUF_ND, PyBUF_SIMPLE, or PyBUF_F_CONTIGUOUS, __getbuffer__ must raise a BufferError. These macros can be cimport’d from cpython.buffer.

(The matrix-in-vector structure actually conforms to PyBUF_ND, but that would prohibit __getbuffer__ from filling in the strides. A single-row matrix is F-contiguous, but a larger matrix is not.)

References

The buffer interface used here is set out in PEP 3118, Revising the buffer protocol.

A tutorial for using this API from C is on Jake Vanderplas’s blog, An Introduction to the Python Buffer Protocol.

Reference documentation is available for Python 3 and Python 2. The Py2 documentation also describes an older buffer protocol that is no longer in use; since Python 2.6, the PEP 3118 protocol has been implemented, and the older protocol is only relevant for legacy code.