Using C libraries

Apart from writing fast code, one of the main use cases of Cython is to call external C libraries from Python code. As Cython code compiles down to C code itself, it is actually trivial to call C functions directly in the code. The following gives a complete example for using (and wrapping) an external C library in Cython code, including appropriate error handling and considerations about designing a suitable API for Python and Cython code.

Imagine you need an efficient way to store integer values in a FIFO queue. Since memory really matters, and the values are actually coming from C code, you cannot afford to create and store Python int objects in a list or deque. So you look out for a queue implementation in C.

After some web search, you find the C-algorithms library [CAlg] and decide to use its double ended queue implementation. To make the handling easier, however, you decide to wrap it in a Python extension type that can encapsulate all memory management.

Defining external declarations

You can download CAlg here.

The C API of the queue implementation, which is defined in the header file c-algorithms/src/queue.h, essentially looks like this:

/* queue.h */
typedef struct _Queue Queue;
typedef void *QueueValue;
Queue *queue_new(void);
void queue_free(Queue *queue);
int queue_push_head(Queue *queue, QueueValue data);
QueueValue queue_pop_head(Queue *queue);
QueueValue queue_peek_head(Queue *queue);
int queue_push_tail(Queue *queue, QueueValue data);
QueueValue queue_pop_tail(Queue *queue);
QueueValue queue_peek_tail(Queue *queue);
int queue_is_empty(Queue *queue);

To get started, the first step is to redefine the C API in a .pxd file, say, cqueue.pxd:

# cqueue.pxd
cdef extern from "c-algorithms/src/queue.h":
    ctypedef struct Queue:
        pass
    ctypedef void* QueueValue
    Queue* queue_new()
    void queue_free(Queue* queue)
    int queue_push_head(Queue* queue, QueueValue data)
    QueueValue  queue_pop_head(Queue* queue)
    QueueValue queue_peek_head(Queue* queue)
    int queue_push_tail(Queue* queue, QueueValue data)
    QueueValue queue_pop_tail(Queue* queue)
    QueueValue queue_peek_tail(Queue* queue)
    bint queue_is_empty(Queue* queue)

Note how these declarations are almost identical to the header file declarations, so you can often just copy them over. However, you do not need to provide all declarations as above, just those that you use in your code or in other declarations, so that Cython gets to see a sufficient and consistent subset of them. Then, consider adapting them somewhat to make them more comfortable to work with in Cython.

Specifically, you should take care of choosing good argument names for the C functions, as Cython allows you to pass them as keyword arguments. Changing them later on is a backwards incompatible API modification. Choosing good names right away will make these functions more pleasant to work with from Cython code.

One noteworthy difference to the header file that we use above is the declaration of the Queue struct in the first line. Queue is in this case used as an opaque handle; only the library that is called knows what is really inside. Since no Cython code needs to know the contents of the struct, we do not need to declare its contents, so we simply provide an empty definition (as we do not want to declare the _Queue type which is referenced in the C header) [1].

Another exception is the last line. The integer return value of the queue_is_empty() function is actually a C boolean value, i.e. the only interesting thing about it is whether it is non-zero or zero, indicating if the queue is empty or not. This is best expressed by Cython’s bint type, which is a normal int type when used in C but maps to Python’s boolean values True and False when converted to a Python object. This way of tightening declarations in a .pxd file can often simplify the code that uses them.

It is good practice to define one .pxd file for each library that you use, and sometimes even for each header file (or functional group) if the API is large. That simplifies their reuse in other projects. Sometimes, you may need to use C functions from the standard C library, or want to call C-API functions from CPython directly. For common needs like this, Cython ships with a set of standard .pxd files that provide these declarations in a readily usable way that is adapted to their use in Cython. The main packages are cpython, libc and libcpp. The NumPy library also has a standard .pxd file numpy, as it is often used in Cython code. See Cython’s Cython/Includes/ source package for a complete list of provided .pxd files.

Writing a wrapper class

After declaring our C library’s API, we can start to design the Queue class that should wrap the C queue. It will live in a file called queue.pyx. [2]

Here is a first start for the Queue class:

# queue.pyx
cimport cqueue
cdef class Queue:
    cdef cqueue.Queue* _c_queue
    def __cinit__(self):
        self._c_queue = cqueue.queue_new()

Note that it says __cinit__ rather than __init__. While __init__ is available as well, it is not guaranteed to be run (for instance, one could create a subclass and forget to call the ancestor’s constructor). Because not initializing C pointers often leads to hard crashes of the Python interpreter, Cython provides __cinit__ which is always called immediately on construction, before CPython even considers calling __init__, and which therefore is the right place to initialise cdef fields of the new instance. However, as __cinit__ is called during object construction, self is not fully constructed yet, and one must avoid doing anything with self but assigning to cdef fields.

Note also that the above method takes no parameters, although subtypes may want to accept some. A no-arguments __cinit__() method is a special case here that simply does not receive any parameters that were passed to a constructor, so it does not prevent subclasses from adding parameters. If parameters are used in the signature of __cinit__(), they must match those of any declared __init__ method of classes in the class hierarchy that are used to instantiate the type.

Memory management

Before we continue implementing the other methods, it is important to understand that the above implementation is not safe. In case anything goes wrong in the call to queue_new(), this code will simply swallow the error, so we will likely run into a crash later on. According to the documentation of the queue_new() function, the only reason why the above can fail is due to insufficient memory. In that case, it will return NULL, whereas it would normally return a pointer to the new queue.

The Python way to get out of this is to raise a MemoryError [3]. We can thus change the init function as follows:

# queue.pyx
cimport cqueue
cdef class Queue:
    cdef cqueue.Queue* _c_queue
    def __cinit__(self):
        self._c_queue = cqueue.queue_new()
        if self._c_queue is NULL:
            raise MemoryError()

The next thing to do is to clean up when the Queue instance is no longer used (i.e. all references to it have been deleted). To this end, CPython provides a callback that Cython makes available as a special method __dealloc__(). In our case, all we have to do is to free the C Queue, but only if we succeeded in initialising it in the init method:

def __dealloc__(self):
    if self._c_queue is not NULL:
        cqueue.queue_free(self._c_queue)

Compiling and linking

At this point, we have a working Cython module that we can test. To compile it, we need to configure a setup.py script for distutils. Here is the most basic script for compiling a Cython module:

from distutils.core import setup
from distutils.extension import Extension
from Cython.Build import cythonize
setup(
    ext_modules = cythonize([Extension("queue", ["queue.pyx"])])
)

To build against the external C library, we need to make sure Cython finds the necessary libraries. There are two ways to archive this. First we can tell distutils where to find the c-source to compile the queue.c implementation automatically. Alternatively, we can build and install C-Alg as system library and dynamically link it. The latter is useful if other applications also use C-Alg.

Static Linking

To build the c-code automatically we need to include compiler directives in queue.pyx:

# distutils: sources = c-algorithms/src/queue.c
# distutils: include_dirs = c-algorithms/src/
cimport cqueue
cdef class Queue:
    cdef cqueue.Queue* _c_queue
    def __cinit__(self):
        self._c_queue = cqueue.queue_new()
        if self._c_queue is NULL:
            raise MemoryError()
    def __dealloc__(self):
        if self._c_queue is not NULL:
            cqueue.queue_free(self._c_queue)

The sources compiler directive gives the path of the C files that distutils is going to compile and link (statically) into the resulting extension module. In general all relevant header files should be found in include_dirs. Now we can build the project using:

$ python setup.py build_ext -i

And test whether our build was successful:

$ python -c 'import queue; Q = queue.Queue()'

Dynamic Linking

Dynamic linking is useful, if the library we are going to wrap is already installed on the system. To perform dynamic linking we first need to build and install c-alg.

To build c-algorithms on your system:

$ cd c-algorithms
$ sh autogen.sh
$ ./configure
$ make

to install CAlg run:

$ make install

Afterwards the file /usr/local/lib/libcalg.so should exist.

Note

This path applies to Linux systems and may be different on other platforms, so you will need to adapt the rest of the tutorial depending on the path where libcalg.so or libcalg.dll is on your system.

In this approach we need to tell the setup script to link with an external library. To do so we need to extend the setup script to install change the extension setup from

ext_modules = cythonize([Extension("queue", ["queue.pyx"])])

to

ext_modules = cythonize([
    Extension("queue", ["queue.pyx"],
              libraries=["calg"])
    ])

Now we should be able to build the project using:

$ python setup.py build_ext -i

If the libcalg is not installed in a ‘normal’ location, users can provide the required parameters externally by passing appropriate C compiler flags, such as:

CFLAGS="-I/usr/local/otherdir/calg/include"  \
LDFLAGS="-L/usr/local/otherdir/calg/lib"     \
    python setup.py build_ext -i

Before we run the module, we also need to make sure that libcalg is in the LD_LIBRARY_PATH environment variable, e.g. by setting:

$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/lib

Once we have compiled the module for the first time, we can now import it and instantiate a new Queue:

$ export PYTHONPATH=.
$ python -c 'import queue; Q = queue.Queue()'

However, this is all our Queue class can do so far, so let’s make it more usable.

Mapping functionality

Before implementing the public interface of this class, it is good practice to look at what interfaces Python offers, e.g. in its list or collections.deque classes. Since we only need a FIFO queue, it’s enough to provide the methods append(), peek() and pop(), and additionally an extend() method to add multiple values at once. Also, since we already know that all values will be coming from C, it’s best to provide only cdef methods for now, and to give them a straight C interface.

In C, it is common for data structures to store data as a void* to whatever data item type. Since we only want to store int values, which usually fit into the size of a pointer type, we can avoid additional memory allocations through a trick: we cast our int values to void* and vice versa, and store the value directly as the pointer value.

Here is a simple implementation for the append() method:

cdef append(self, int value):
    cqueue.queue_push_tail(self._c_queue, <void*>value)

Again, the same error handling considerations as for the __cinit__() method apply, so that we end up with this implementation instead:

cdef append(self, int value):
    if not cqueue.queue_push_tail(self._c_queue,
                                  <void*>value):
        raise MemoryError()

Adding an extend() method should now be straight forward:

cdef extend(self, int* values, size_t count):
    """Append all ints to the queue.
    """
    cdef int value
    for value in values[:count]:  # Slicing pointer to limit the iteration boundaries.
        self.append(value)

This becomes handy when reading values from a C array, for example.

So far, we can only add data to the queue. The next step is to write the two methods to get the first element: peek() and pop(), which provide read-only and destructive read access respectively. To avoid compiler warnings when casting void* to int directly, we use an intermediate data type that is big enough to hold a void*. Here, Py_ssize_t:

cdef int peek(self):
    return <Py_ssize_t>cqueue.queue_peek_head(self._c_queue)
cdef int pop(self):
    return <Py_ssize_t>cqueue.queue_pop_head(self._c_queue)

Normally, in C, we risk losing data when we convert a larger integer type to a smaller integer type without checking the boundaries, and Py_ssize_t may be a larger type than int. But since we control how values are added to the queue, we already know that all values that are in the queue fit into an int, so the above conversion from void* to Py_ssize_t to int (the return type) is safe by design.

Handling errors

Now, what happens when the queue is empty? According to the documentation, the functions return a NULL pointer, which is typically not a valid value. But since we are simply casting to and from ints, we cannot distinguish anymore if the return value was NULL because the queue was empty or because the value stored in the queue was 0. In Cython code, we want the first case to raise an exception, whereas the second case should simply return 0. To deal with this, we need to special case this value, and check if the queue really is empty or not:

cdef int peek(self) except? -1:
    cdef int value = <Py_ssize_t>cqueue.queue_peek_head(self._c_queue)
    if value == 0:
        # this may mean that the queue is empty, or
        # that it happens to contain a 0 value
        if cqueue.queue_is_empty(self._c_queue):
            raise IndexError("Queue is empty")
    return value

Note how we have effectively created a fast path through the method in the hopefully common cases that the return value is not 0. Only that specific case needs an additional check if the queue is empty.

The except? -1 declaration in the method signature falls into the same category. If the function was a Python function returning a Python object value, CPython would simply return NULL internally instead of a Python object to indicate an exception, which would immediately be propagated by the surrounding code. The problem is that the return type is int and any int value is a valid queue item value, so there is no way to explicitly signal an error to the calling code. In fact, without such a declaration, there is no obvious way for Cython to know what to return on exceptions and for calling code to even know that this method may exit with an exception.

The only way calling code can deal with this situation is to call PyErr_Occurred() when returning from a function to check if an exception was raised, and if so, propagate the exception. This obviously has a performance penalty. Cython therefore allows you to declare which value it should implicitly return in the case of an exception, so that the surrounding code only needs to check for an exception when receiving this exact value.

We chose to use -1 as the exception return value as we expect it to be an unlikely value to be put into the queue. The question mark in the except? -1 declaration indicates that the return value is ambiguous (there may be a -1 value in the queue, after all) and that an additional exception check using PyErr_Occurred() is needed in calling code. Without it, Cython code that calls this method and receives the exception return value would silently (and sometimes incorrectly) assume that an exception has been raised. In any case, all other return values will be passed through almost without a penalty, thus again creating a fast path for ‘normal’ values.

Now that the peek() method is implemented, the pop() method also needs adaptation. Since it removes a value from the queue, however, it is not enough to test if the queue is empty after the removal. Instead, we must test it on entry:

cdef int pop(self) except? -1:
    if cqueue.queue_is_empty(self._c_queue):
        raise IndexError("Queue is empty")
    return <Py_ssize_t>cqueue.queue_pop_head(self._c_queue)

The return value for exception propagation is declared exactly as for peek().

Lastly, we can provide the Queue with an emptiness indicator in the normal Python way by implementing the __bool__() special method (note that Python 2 calls this method __nonzero__, whereas Cython code can use either name):

def __bool__(self):
    return not cqueue.queue_is_empty(self._c_queue)

Note that this method returns either True or False as we declared the return type of the queue_is_empty() function as bint in cqueue.pxd.

Testing the result

Now that the implementation is complete, you may want to write some tests for it to make sure it works correctly. Especially doctests are very nice for this purpose, as they provide some documentation at the same time. To enable doctests, however, you need a Python API that you can call. C methods are not visible from Python code, and thus not callable from doctests.

A quick way to provide a Python API for the class is to change the methods from cdef to cpdef. This will let Cython generate two entry points, one that is callable from normal Python code using the Python call semantics and Python objects as arguments, and one that is callable from C code with fast C semantics and without requiring intermediate argument conversion from or to Python types. Note that cpdef methods ensure that they can be appropriately overridden by Python methods even when they are called from Cython. This adds a tiny overhead compared to cdef methods.

Now that we have both a C-interface and a Python interface for our class, we should make sure that both interfaces are consistent. Python users would expect an extend() method that accepts arbitrary iterables, whereas C users would like to have one that allows passing C arrays and C memory. Both signatures are incompatible.

We will solve this issue by considering that in C, the API could also want to support other input types, e.g. arrays of long or char, which is usually supported with differently named C API functions such as extend_ints(), extend_longs(), extend_chars()``, etc. This allows us to free the method name extend() for the duck typed Python method, which can accept arbitrary iterables.

The following listing shows the complete implementation that uses cpdef methods where possible:

# queue.pyx
cimport cqueue
cdef class Queue:
    """A queue class for C integer values.
    >>> q = Queue()
    >>> q.append(5)
    >>> q.peek()
    5
    >>> q.pop()
    5
    """
    cdef cqueue.Queue* _c_queue
    def __cinit__(self):
        self._c_queue = cqueue.queue_new()
        if self._c_queue is NULL:
            raise MemoryError()
    def __dealloc__(self):
        if self._c_queue is not NULL:
            cqueue.queue_free(self._c_queue)
    cpdef append(self, int value):
        if not cqueue.queue_push_tail(self._c_queue,
                                      <void*> <Py_ssize_t> value):
            raise MemoryError()
    # The `cpdef` feature is obviously not available for the original "extend()"
    # method, as the method signature is incompatible with Python argument
    # types (Python does not have pointers).  However, we can rename
    # the C-ish "extend()" method to e.g. "extend_ints()", and write
    # a new "extend()" method that provides a suitable Python interface by
    # accepting an arbitrary Python iterable.
    cpdef extend(self, values):
        for value in values:
            self.append(value)
    cdef extend_ints(self, int* values, size_t count):
        cdef int value
        for value in values[:count]:  # Slicing pointer to limit the iteration boundaries.
            self.append(value)
    cpdef int peek(self) except? -1:
        cdef int value = <Py_ssize_t> cqueue.queue_peek_head(self._c_queue)
        if value == 0:
            # this may mean that the queue is empty,
            # or that it happens to contain a 0 value
            if cqueue.queue_is_empty(self._c_queue):
                raise IndexError("Queue is empty")
        return value
    cpdef int pop(self) except? -1:
        if cqueue.queue_is_empty(self._c_queue):
            raise IndexError("Queue is empty")
        return <Py_ssize_t> cqueue.queue_pop_head(self._c_queue)
    def __bool__(self):
        return not cqueue.queue_is_empty(self._c_queue)

Now we can test our Queue implementation using a python script, for example here test_queue.py:

from __future__ import print_function
import time
import queue
Q = queue.Queue()
Q.append(10)
Q.append(20)
print(Q.peek())
print(Q.pop())
print(Q.pop())
try:
    print(Q.pop())
except IndexError as e:
    print("Error message:", e)  # Prints "Queue is empty"
i = 10000
values = range(i)
start_time = time.time()
Q.extend(values)
end_time = time.time() - start_time
print("Adding {} items took {:1.3f} msecs.".format(i, 1000 * end_time))
for i in range(41):
    Q.pop()
Q.pop()
print("The answer is:")
print(Q.pop())

As a quick test with 10000 numbers on the author’s machine indicates, using this Queue from Cython code with C int values is about five times as fast as using it from Cython code with Python object values, almost eight times faster than using it from Python code in a Python loop, and still more than twice as fast as using Python’s highly optimised collections.deque type from Cython code with Python integers.

Callbacks

Let’s say you want to provide a way for users to pop values from the queue up to a certain user defined event occurs. To this end, you want to allow them to pass a predicate function that determines when to stop, e.g.:

def pop_until(self, predicate):
    while not predicate(self.peek()):
        self.pop()

Now, let us assume for the sake of argument that the C queue provides such a function that takes a C callback function as predicate. The API could look as follows:

/* C type of a predicate function that takes a queue value and returns
 * -1 for errors
 *  0 for reject
 *  1 for accept
 */
typedef int (*predicate_func)(void* user_context, QueueValue data);
/* Pop values as long as the predicate evaluates to true for them,
 * returns -1 if the predicate failed with an error and 0 otherwise.
 */
int queue_pop_head_until(Queue *queue, predicate_func predicate,
                         void* user_context);

It is normal for C callback functions to have a generic void* argument that allows passing any kind of context or state through the C-API into the callback function. We will use this to pass our Python predicate function.

First, we have to define a callback function with the expected signature that we can pass into the C-API function:

cdef int evaluate_predicate(void* context, cqueue.QueueValue value):
    "Callback function that can be passed as predicate_func"
    try:
        # recover Python function object from void* argument
        func = <object>context
        # call function, convert result into 0/1 for True/False
        return bool(func(<int>value))
    except:
        # catch any Python errors and return error indicator
        return -1

The main idea is to pass a pointer (a.k.a. borrowed reference) to the function object as the user context argument. We will call the C-API function as follows:

def pop_until(self, python_predicate_function):
    result = cqueue.queue_pop_head_until(
        self._c_queue, evaluate_predicate,
        <void*>python_predicate_function)
    if result == -1:
        raise RuntimeError("an error occurred")

The usual pattern is to first cast the Python object reference into a void* to pass it into the C-API function, and then cast it back into a Python object in the C predicate callback function. The cast to void* creates a borrowed reference. On the cast to <object>, Cython increments the reference count of the object and thus converts the borrowed reference back into an owned reference. At the end of the predicate function, the owned reference goes out of scope again and Cython discards it.

The error handling in the code above is a bit simplistic. Specifically, any exceptions that the predicate function raises will essentially be discarded and only result in a plain RuntimeError() being raised after the fact. This can be improved by storing away the exception in an object passed through the context parameter and re-raising it after the C-API function has returned -1 to indicate the error.