Cython for NumPy users

This tutorial is aimed at NumPy users who have no experience with Cython at all. If you have some knowledge of Cython you may want to skip to the ‘’Efficient indexing’’ section.

The main scenario considered is NumPy end-use rather than NumPy/SciPy development. The reason is that Cython is not (yet) able to support functions that are generic with respect to the number of dimensions in a high-level fashion. This restriction is much more severe for SciPy development than more specific, “end-user” functions. See the last section for more information on this.

The style of this tutorial will not fit everybody, so you can also consider:

Cython at a glance

Cython is a compiler which compiles Python-like code files to C code. Still, ‘’Cython is not a Python to C translator’‘. That is, it doesn’t take your full program and “turns it into C” – rather, the result makes full use of the Python runtime environment. A way of looking at it may be that your code is still Python in that it runs within the Python runtime environment, but rather than compiling to interpreted Python bytecode one compiles to native machine code (but with the addition of extra syntax for easy embedding of faster C-like code).

This has two important consequences:

  • Speed. How much depends very much on the program involved though. Typical Python numerical programs would tend to gain very little as most time is spent in lower-level C that is used in a high-level fashion. However for-loop-style programs can gain many orders of magnitude, when typing information is added (and is so made possible as a realistic alternative).
  • Easy calling into C code. One of Cython’s purposes is to allow easy wrapping of C libraries. When writing code in Cython you can call into C code as easily as into Python code.

Very few Python constructs are not yet supported, though making Cython compile all Python code is a stated goal, you can see the differences with Python in limitations.

Your Cython environment

Using Cython consists of these steps:

  1. Write a .pyx source file
  2. Run the Cython compiler to generate a C file
  3. Run a C compiler to generate a compiled library
  4. Run the Python interpreter and ask it to import the module

However there are several options to automate these steps:

  1. The SAGE mathematics software system provides excellent support for using Cython and NumPy from an interactive command line or through a notebook interface (like Maple/Mathematica). See this documentation.
  2. Cython can be used as an extension within a Jupyter notebook, making it easy to compile and use Cython code with just a %%cython at the top of a cell. For more information see Using the Jupyter Notebook.
  3. A version of pyximport is shipped with Cython, so that you can import pyx-files dynamically into Python and have them compiled automatically (See Compiling with pyximport).
  4. Cython supports distutils so that you can very easily create build scripts which automate the process, this is the preferred method for Cython implemented libraries and packages. See Basic setup.py.
  5. Manual compilation (see below)

Note

If using another interactive command line environment than SAGE, like IPython or Python itself, it is important that you restart the process when you recompile the module. It is not enough to issue an “import” statement again.

Installation

If you already have a C compiler, just do:

pip install Cython

otherwise, see the installation page.

As of this writing SAGE comes with an older release of Cython than required for this tutorial. So if using SAGE you should download the newest Cython and then execute

$ cd path/to/cython-distro
$ path-to-sage/sage -python setup.py install

This will install the newest Cython into SAGE.

Manual compilation

As it is always important to know what is going on, I’ll describe the manual method here. First Cython is run:

$ cython yourmod.pyx

This creates yourmod.c which is the C source for a Python extension module. A useful additional switch is -a which will generate a document yourmod.html) that shows which Cython code translates to which C code line by line.

Then we compile the C file. This may vary according to your system, but the C file should be built like Python was built. Python documentation for writing extensions should have some details. On Linux this often means something like:

$ gcc -shared -pthread -fPIC -fwrapv -O2 -Wall -fno-strict-aliasing -I/usr/include/python2.7 -o yourmod.so yourmod.c

gcc should have access to the NumPy C header files so if they are not installed at /usr/include/numpy or similar you may need to pass another option for those. You only need to provide the NumPy headers if you write:

cimport numpy

in your Cython code.

This creates yourmod.so in the same directory, which is importable by Python by using a normal import yourmod statement.

The first Cython program

You can easily execute the code of this tutorial by downloading the Jupyter notebook.

The code below does the equivalent of this function in numpy:

def compute_np(array_1, array_2, a, b, c):
    return np.clip(array_1, 2, 10) * a + array_2 * b + c

We’ll say that array_1 and array_2 are 2D NumPy arrays of integer type and a, b and c are three Python integers.

This function uses NumPy and is already really fast, so it might be a bit overkill to do it again with Cython. This is for demonstration purposes. Nonetheless, we will show that we achieve a better speed and memory efficiency than NumPy at the cost of more verbosity.

This code computes the function with the loops over the two dimensions being unrolled. It is both valid Python and valid Cython code. I’ll refer to it as both compute_py.py for the Python version and compute_cy.pyx for the Cython version – Cython uses .pyx as its file suffix (but it can also compile .py files).

import numpy as np


def clip(a, min_value, max_value):
    return min(max(a, min_value), max_value)


def compute(array_1, array_2, a, b, c):
    """
    This function must implement the formula
    np.clip(array_1, 2, 10) * a + array_2 * b + c

    array_1 and array_2 are 2D.
    """
    x_max = array_1.shape[0]
    y_max = array_1.shape[1]

    assert array_1.shape == array_2.shape

    result = np.zeros((x_max, y_max), dtype=array_1.dtype)

    for x in range(x_max):
        for y in range(y_max):
            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result[x, y] = tmp + c

    return result

This should be compiled to produce convolve_cy.so (for Linux systems, on Windows systems, this will be a .pyd file). We run a Python session to test both the Python version (imported from .py-file) and the compiled Cython module.

In [1]: import numpy as np
In [2]: array_1 = np.random.uniform(0, 1000, size=(3000, 2000)).astype(np.intc)
In [3]: array_2 = np.random.uniform(0, 1000, size=(3000, 2000)).astype(np.intc)
In [4]: a = 4
In [5]: b = 3
In [6]: c = 9
In [7]: def compute_np(array_1, array_2, a, b, c):
   ...:     return np.clip(array_1, 2, 10) * a + array_2 * b + c
In [8]: %timeit compute_np(array_1, array_2, a, b, c)
103 ms ± 4.16 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [9]: import compute_py
In [10]: compute_py.compute(array_1, array_2, a, b, c)
1min 10s ± 844 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [11]: import compute_cy
In [12]: compute_cy.compute(array_1, array_2, a, b, c)
56.5 s ± 587 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

There’s not such a huge difference yet; because the C code still does exactly what the Python interpreter does (meaning, for instance, that a new object is allocated for each number used).

You can look at the Python interaction and the generated C code by using -a when calling Cython from the command line, %%cython -a when using a Jupyter Notebook, or by using cythonize('compute_cy.pyx', annotate=True) when using a setup.py. Look at the generated html file and see what is needed for even the simplest statements. You get the point quickly. We need to give Cython more information; we need to add types.

Adding types

To add types we use custom Cython syntax, so we are now breaking Python source compatibility. Here’s compute_typed.pyx. Read the comments!

import numpy as np

# We now need to fix a datatype for our arrays. I've used the variable
# DTYPE for this, which is assigned to the usual NumPy runtime
# type info object.
DTYPE = np.intc

# cdef means here that this function is a plain C function (so faster).
# To get all the benefits, we type the arguments and the return value.
cdef int clip(int a, int min_value, int max_value):
    return min(max(a, min_value), max_value)


def compute(array_1, array_2, int a, int b, int c):
    
    # The "cdef" keyword is also used within functions to type variables. It
    # can only be used at the top indentation level (there are non-trivial
    # problems with allowing them in other places, though we'd love to see
    # good and thought out proposals for it).
    cdef Py_ssize_t x_max = array_1.shape[0]
    cdef Py_ssize_t y_max = array_1.shape[1]
    
    assert array_1.shape == array_2.shape
    assert array_1.dtype == DTYPE
    assert array_2.dtype == DTYPE

    result = np.zeros((x_max, y_max), dtype=DTYPE)
    
    # It is very important to type ALL your variables. You do not get any
    # warnings if not, only much slower code (they are implicitly typed as
    # Python objects).
    # For the "tmp" variable, we want to use the same data type as is
    # stored in the array, so we use int because it correspond to np.intc.
    # NB! An important side-effect of this is that if "tmp" overflows its
    # datatype size, it will simply wrap around like in C, rather than raise
    # an error like in Python.

    cdef int tmp

    # Py_ssize_t is the proper C type for Python array indices.
    cdef Py_ssize_t x, y

    for x in range(x_max):
        for y in range(y_max):

            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result[x, y] = tmp + c

    return result
../../_images/compute_typed_html.jpg

At this point, have a look at the generated C code for compute_cy.pyx and compute_typed.pyx. Click on the lines to expand them and see corresponding C.

Especially have a look at the for-loops: In compute_cy.c, these are ~20 lines of C code to set up while in compute_typed.c a normal C for loop is used.

After building this and continuing my (very informal) benchmarks, I get:

In [13]: %timeit compute_typed.compute(array_1, array_2, a, b, c)
26.5 s ± 422 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

So adding types does make the code faster, but nowhere near the speed of NumPy?

What happened is that most of the time spend in this code is spent in the following lines, and those lines are slower to execute than in pure Python:

tmp = clip(array_1[x, y], 2, 10)
tmp = tmp * a + array_2[x, y] * b
result[x, y] = tmp + c

So what made those line so much slower than in the pure Python version?

array_1 and array_2 are still NumPy arrays, so Python objects, and expect Python integers as indexes. Here we pass C int values. So every time Cython reaches this line, it has to convert all the C integers to Python int objects. Since this line is called very often, it outweighs the speed benefits of the pure C loops that were created from the range() earlier.

Furthermore, tmp * a + array_2[x, y] * b returns a Python integer and tmp is a C integer, so Cython has to do type conversions again. In the end those types conversions add up. And made our computation really slow. But this problem can be solved easily by using memoryviews.

Efficient indexing with memoryviews

There are still two bottlenecks that degrade the performance, and that is the array lookups and assignments, as well as C/Python types conversion. The []-operator still uses full Python operations – what we would like to do instead is to access the data buffer directly at C speed.

What we need to do then is to type the contents of the ndarray objects. We do this with a memoryview. There is a page in the Cython documentation dedicated to it.

In short, memoryviews are C structures that can hold a pointer to the data of a NumPy array and all the necessary buffer metadata to provide efficient and safe access: dimensions, strides, item size, item type information, etc… They also support slices, so they work even if the NumPy array isn’t contiguous in memory. They can be indexed by C integers, thus allowing fast access to the NumPy array data.

Here is how to declare a memoryview of integers:

cdef int [:] foo         # 1D memoryview
cdef int [:, :] foo      # 2D memoryview
cdef int [:, :, :] foo   # 3D memoryview
...                      # You get the idea.

No data is copied from the NumPy array to the memoryview in our example. As the name implies, it is only a “view” of the memory. So we can use the view result_view for efficient indexing and at the end return the real NumPy array result that holds the data that we operated on.

Here is how to use them in our code:

compute_memview.pyx

import numpy as np

DTYPE = np.intc


cdef int clip(int a, int min_value, int max_value):
    return min(max(a, min_value), max_value)


def compute(int[:, :] array_1, int[:, :] array_2, int a, int b, int c):
     
    cdef Py_ssize_t x_max = array_1.shape[0]
    cdef Py_ssize_t y_max = array_1.shape[1]

    # array_1.shape is now a C array, no it's not possible
    # to compare it simply by using == without a for-loop.
    # To be able to compare it to array_2.shape easily,
    # we convert them both to Python tuples.
    assert tuple(array_1.shape) == tuple(array_2.shape)

    result = np.zeros((x_max, y_max), dtype=DTYPE)
    cdef int[:, :] result_view = result

    cdef int tmp
    cdef Py_ssize_t x, y

    for x in range(x_max):
        for y in range(y_max):

            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result_view[x, y] = tmp + c

    return result

Let’s see how much faster accessing is now.

In [22]: %timeit compute_memview.compute(array_1, array_2, a, b, c)
22.9 ms ± 197 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

Note the importance of this change. We’re now 3081 times faster than an interpreted version of Python and 4.5 times faster than NumPy.

Memoryviews can be used with slices too, or even with Python arrays. Check out the memoryview page to see what they can do for you.

Tuning indexing further

The array lookups are still slowed down by two factors:

  1. Bounds checking is performed.
  2. Negative indices are checked for and handled correctly. The code above is explicitly coded so that it doesn’t use negative indices, and it (hopefully) always access within bounds.

With decorators, we can deactivate those checks:

...
cimport cython
@cython.boundscheck(False)  # Deactivate bounds checking
@cython.wraparound(False)   # Deactivate negative indexing.
def compute(int[:, :] array_1, int[:, :] array_2, int a, int b, int c):
...

Now bounds checking is not performed (and, as a side-effect, if you ‘’do’’ happen to access out of bounds you will in the best case crash your program and in the worst case corrupt data). It is possible to switch bounds-checking mode in many ways, see Compiler directives for more information.

In [23]: %timeit compute_index.compute(array_1, array_2, a, b, c)
16.8 ms ± 25.4 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

We’re faster than the NumPy version (6.2x). NumPy is really well written, but does not performs operation lazily, resulting in a lot of intermediate copy operations in memory. Our version is very memory efficient and cache friendly because we can execute the operations in a single run over the data.

Warning

Speed comes with some cost. Especially it can be dangerous to set typed objects (like array_1, array_2 and result_view in our sample code) to None. Setting such objects to None is entirely legal, but all you can do with them is check whether they are None. All other use (attribute lookup or indexing) can potentially segfault or corrupt data (rather than raising exceptions as they would in Python).

The actual rules are a bit more complicated but the main message is clear: Do not use typed objects without knowing that they are not set to None.

Declaring the NumPy arrays as contiguous

For extra speed gains, if you know that the NumPy arrays you are providing are contiguous in memory, you can declare the memoryview as contiguous.

We give an example on an array that has 3 dimensions. If you want to give Cython the information that the data is C-contiguous you have to declare the memoryview like this:

cdef int [:,:,::1] a

If you want to give Cython the information that the data is Fortran-contiguous you have to declare the memoryview like this:

cdef int [::1, :, :] a

If all this makes no sense to you, you can skip this part, declaring arrays as contiguous constrains the usage of your functions as it rejects array slices as input. If you still want to understand what contiguous arrays are all about, you can see this answer on StackOverflow.

For the sake of giving numbers, here are the speed gains that you should get by declaring the memoryviews as contiguous:

In [23]: %timeit compute_contiguous.compute(array_1, array_2, a, b, c)
11.1 ms ± 30.2 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

We’re now around nine times faster than the NumPy version, and 6300 times faster than the pure Python version!

Making the function cleaner

Declaring types can make your code quite verbose. If you don’t mind Cython inferring the C types of your variables, you can use the infer_types=True compiler directive at the top of the file. It will save you quite a bit of typing.

Note that since type declarations must happen at the top indentation level, Cython won’t infer the type of variables declared for the first time in other indentation levels. It would change too much the meaning of our code. This is why, we must still declare manually the type of the tmp, x and y variable.

And actually, manually giving the type of the tmp variable will be useful when using fused types.

# cython: infer_types=True
import numpy as np
cimport cython

DTYPE = np.intc


cdef int clip(int a, int min_value, int max_value):
    return min(max(a, min_value), max_value)


@cython.boundscheck(False)
@cython.wraparound(False)
def compute(int[:, ::1] array_1, int[:, ::1] array_2, int a, int b, int c):
     
    x_max = array_1.shape[0]
    y_max = array_1.shape[1]
    
    assert tuple(array_1.shape) == tuple(array_2.shape)

    result = np.zeros((x_max, y_max), dtype=DTYPE)
    cdef int[:, ::1] result_view = result

    cdef int tmp
    cdef Py_ssize_t x, y

    for x in range(x_max):
        for y in range(y_max):

            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result_view[x, y] = tmp + c

    return result

We now do a speed test:

In [24]: %timeit compute_infer_types.compute(array_1, array_2, a, b, c)
11.5 ms ± 261 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

Lo and behold, the speed has not changed.

More generic code

All those speed gains are nice, but adding types constrains our code. At the moment, it would mean that our function can only work with NumPy arrays with the np.intc type. Is it possible to make our code work for multiple NumPy data types?

Yes, with the help of a new feature called fused types. You can learn more about it at this section of the documentation. It is similar to C++ ‘s templates. It generates multiple function declarations at compile time, and then chooses the right one at run-time based on the types of the arguments provided. By comparing types in if-conditions, it is also possible to execute entirely different code paths depending on the specific data type.

In our example, since we don’t have access anymore to the NumPy’s dtype of our input arrays, we use those if-else statements to know what NumPy data type we should use for our output array.

In this case, our function now works for ints, doubles and floats.

# cython: infer_types=True
import numpy as np
cimport cython

ctypedef fused my_type:
    int
    double
    long long


cdef my_type clip(my_type a, my_type min_value, my_type max_value):
    return min(max(a, min_value), max_value)


@cython.boundscheck(False)
@cython.wraparound(False)
def compute(my_type[:, ::1] array_1, my_type[:, ::1] array_2, my_type a, my_type b, my_type c):
     
    x_max = array_1.shape[0]
    y_max = array_1.shape[1]
    
    assert tuple(array_1.shape) == tuple(array_2.shape)

    if my_type is int:
        dtype = np.intc
    elif my_type is double:
        dtype = np.double
    elif my_type is cython.longlong:
        dtype = np.longlong

    result = np.zeros((x_max, y_max), dtype=dtype)
    cdef my_type[:, ::1] result_view = result

    cdef my_type tmp
    cdef Py_ssize_t x, y

    for x in range(x_max):
        for y in range(y_max):

            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result_view[x, y] = tmp + c

    return result

We can check that the output type is the right one:

>>>compute(array_1, array_2, a, b, c).dtype
dtype('int32')
>>>compute(array_1.astype(np.double), array_2.astype(np.double), a, b, c).dtype
dtype('float64')

We now do a speed test:

In [25]: %timeit compute_fused_types.compute(array_1, array_2, a, b, c)
11.5 ms ± 258 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

More versions of the function are created at compile time. So it makes sense that the speed doesn’t change for executing this function with integers as before.

Using multiple threads

Cython have support for OpenMP. It have also some nice wrappers around it, like the function prange(). You can see more information about Cython and parralelism in Using Parallelism. Since we do elementwise operations, we can easily distribute the work among multiple threads. It’s important not to forget to pass the correct arguments to the compiler to enable OpenMP. When using the Jupyter notebook, you should use the cell magic like this:

%%cython --force
# distutils: extra_compile_args=-fopenmp
# distutils: extra_link_args=-fopenmp

The GIL must be released (see Releasing the GIL), so this is why we declare our clip() function nogil.

# tag: openmp
# You can ignore the previous line.
# It's for internal testing of the cython documentation.

# distutils: extra_compile_args=-fopenmp
# distutils: extra_link_args=-fopenmp

import numpy as np
cimport cython
from cython.parallel import prange

ctypedef fused my_type:
    int
    double
    long long


# We declare our plain c function nogil
cdef my_type clip(my_type a, my_type min_value, my_type max_value) nogil:
    return min(max(a, min_value), max_value)


@cython.boundscheck(False)
@cython.wraparound(False)
def compute(my_type[:, ::1] array_1, my_type[:, ::1] array_2, my_type a, my_type b, my_type c):

    cdef Py_ssize_t x_max = array_1.shape[0]
    cdef Py_ssize_t y_max = array_1.shape[1]

    assert tuple(array_1.shape) == tuple(array_2.shape)

    if my_type is int:
        dtype = np.intc
    elif my_type is double:
        dtype = np.double
    elif my_type is cython.longlong:
        dtype = np.longlong

    result = np.zeros((x_max, y_max), dtype=dtype)
    cdef my_type[:, ::1] result_view = result

    cdef my_type tmp
    cdef Py_ssize_t x, y

    # We use prange here.
    for x in prange(x_max, nogil=True):
        for y in range(y_max):

            tmp = clip(array_1[x, y], 2, 10)
            tmp = tmp * a + array_2[x, y] * b
            result_view[x, y] = tmp + c

    return result

We can have substantial speed gains for minimal effort:

In [25]: %timeit compute_prange.compute(array_1, array_2, a, b, c)
9.33 ms ± 412 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

We’re now 7558 times faster than the pure Python version and 11.1 times faster than NumPy!

Where to go from here?