Typed Memoryviews

    Memoryviews are more general than the old NumPy array buffer support, because they can handle a wider variety of sources of array data. For example, they can handle C arrays and the Cython array type (Cython arrays).

    A memoryview can be used in any context (function parameters, module-level, cdef class attribute, etc) and can be obtained from nearly any object that exposes writable buffer through the buffer interface.

    If you are used to working with NumPy, the following examples should get you started with Cython memory views.

    This code should give the following output:

    1. NumPy sum of the NumPy array before assignments: 351
    2. NumPy sum of NumPy array after assignments: 81
    3. Memoryview sum of NumPy array is 81
    4. Memoryview sum of C array is 451
    5. Memoryview sum of Cython array is 1351
    6. Memoryview sum of C memoryview is 451

    Using memoryviews

    Memory views use Python slicing syntax in a similar way as NumPy.

    To create a complete view on a one-dimensional int buffer:

    1. cdef int[:] view1D = exporting_object

    A complete 3D view:

    1. cdef int[:,:,:] view3D = exporting_object

    A 2D view that restricts the first dimension of a buffer to 100 rows starting at the second (index 1) and then skips every second (odd) row:

    1. cdef int[1:102:2,:] partial_view = exporting_object

    This also works conveniently as function arguments:

    1. def process_3d_buffer(int[1:102:2,:] view not None):
    2.     ...

    The not None declaration for the argument automatically rejects None values as input, which would otherwise be allowed. The reason why None is allowed by default is that it is conveniently used for return arguments:

    1. import numpy as np
    3. def process_buffer(int[:,:] input_view not None,
    4.                    int[:,:] output_view=None):
    6.    if output_view is None:
    7.        # Creating a default view, e.g.
    8.        output_view = np.empty_like(input_view)
    10.    # process 'input_view' into 'output_view'
    11.    return output_view

    Cython will reject incompatible buffers automatically, e.g. passing a three dimensional buffer into a function that requires a two dimensional buffer will raise a ValueError.

    Indexing

    In Cython, index access on memory views is automatically translated into memory addresses. The following code requests a two-dimensional memory view of C int typed items and indexes into it:

    1. cdef int[:,:] buf = exporting_object
    3. print(buf[1,2])

    Negative indices work as well, counting from the end of the respective dimension:

    1. print(buf[-1,-2])

    The following function loops over each dimension of a 2D array and adds 1 to each item:

    1. import numpy as np
    3. def add_one(int[:,:] buf):
    4.     for x in range(buf.shape[0]):
    5.         for y in range(buf.shape[1]):
    6.             buf[x, y] += 1
    8. # exporting_object must be a Python object
    9. # implementing the buffer interface, e.g. a numpy array.
    10. exporting_object = np.zeros((10, 20), dtype=np.intc)
    12. add_one(exporting_object)

    Indexing and slicing can be done with or without the GIL. It basically works like NumPy. If indices are specified for every dimension you will get an element of the base type (e.g. int). Otherwise, you will get a new view. An Ellipsis means you get consecutive slices for every unspecified dimension:

    1. import numpy as np
    3. exporting_object = np.arange(0, 15 * 10 * 20, dtype=np.intc).reshape((15, 10, 20))
    5. cdef int[:, :, :] my_view = exporting_object
    7. # These are all equivalent
    8. my_view[10]
    9. my_view[10, :, :]
    10. my_view[10, ...]

    Copying

    Memory views can be copied in place:

    1. import numpy as np
    3. cdef int[:, :, :] to_view, from_view
    4. to_view = np.empty((20, 15, 30), dtype=np.intc)
    5. from_view = np.ones((20, 15, 30), dtype=np.intc)
    7. # copy the elements in from_view to to_view
    8. to_view[...] = from_view
    9. # or
    10. to_view[:] = from_view
    11. # or
    12. to_view[:, :, :] = from_view

    They can also be copied with the copy() and copy_fortran() methods; see .

    Transposing

    In most cases (see below), the memoryview can be transposed in the same way that NumPy slices can be transposed:

    1. import numpy as np
    3. array = np.arange(20, dtype=np.intc).reshape((2, 10))
    5. cdef int[:, ::1] c_contig = array
    6. cdef int[::1, :] f_contig = c_contig.T

    This gives a new, transposed, view on the data.

    Transposing requires that all dimensions of the memoryview have a direct access memory layout (i.e., there are no indirections through pointers). See for details.

    As for NumPy, new axes can be introduced by indexing an array with None

    1. cdef double[:] myslice = np.linspace(0, 10, num=50)
    3. # 2D array with shape (1, 50)
    4. myslice[None] # or
    5. myslice[None, :]
    7. # 2D array with shape (50, 1)
    8. myslice[:, None]
    10. # 3D array with shape (1, 10, 1)
    11. myslice[None, 10:-20:2, None]

    One may mix new axis indexing with all other forms of indexing and slicing. See also an example.

    Read-only views

    Since Cython 0.28, the memoryview item type can be declared as const to support read-only buffers as input:

    1. import numpy as np
    3. cdef const double[:] myslice   # const item type => read-only view
    5. a = np.linspace(0, 10, num=50)
    6. a.setflags(write=False)
    7. myslice = a

    Using a non-const memoryview with a binary Python string produces a runtime error. You can solve this issue with a const memoryview:

    1. cdef bint is_y_in(const unsigned char[:] string_view):
    2.     cdef int i
    3.     for i in range(string_view.shape[0]):
    4.         if string_view[i] == b'y':
    5.             return True
    7. print(is_y_in(b'hello world'))   # False
    8. print(is_y_in(b'hello Cython'))  # True

    Note that this does not require the input buffer to be read-only:

    1. a = np.linspace(0, 10, num=50)
    2. myslice = a   # read-only view of a writable buffer

    Comparison to the old buffer support

    You will probably prefer memoryviews to the older syntax because:

    • The syntax is cleaner
    • Memoryviews do not usually need the GIL (see )
    • Memoryviews are considerably faster

    For example, this is the old syntax equivalent of the sum3d function above:

    1. cpdef int old_sum3d(object[int, ndim=3, mode='strided'] arr):
    2.     cdef int I, J, K, total = 0
    3.     I = arr.shape[0]
    4.     J = arr.shape[1]
    5.     K = arr.shape[2]
    6.     for i in range(I):
    7.         for j in range(J):
    8.             for k in range(K):
    9.                 total += arr[i, j, k]
    10.     return total

    Note that we can’t use nogil for the buffer version of the function as we could for the memoryview version of above, because buffer objects are Python objects. However, even if we don’t use nogil with the memoryview, it is significantly faster. This is a output from an IPython session after importing both versions:

    1. In [2]: import numpy as np
    3. In [3]: arr = np.zeros((40, 40, 40), dtype=int)
    5. In [4]: timeit -r15 old_sum3d(arr)
    6. 1000 loops, best of 15: 298 us per loop
    8. In [5]: timeit -r15 sum3d(arr)
    9. 1000 loops, best of 15: 219 us per loop

    Python buffer support

    Cython memoryviews support nearly all objects exporting the interface of Python . This is the buffer interface described in `PEP 3118`_. NumPy arrays support this interface, as do . The “nearly all” is because the Python buffer interface allows the elements in the data array to themselves be pointers; Cython memoryviews do not yet support this.

    The buffer interface allows objects to identify the underlying memory in a variety of ways. With the exception of pointers for data elements, Cython memoryviews support all Python new-type buffer layouts. It can be useful to know or specify memory layout if the memory has to be in a particular format for an external routine, or for code optimization.

    Background

    The concepts are as follows: there is data access and data packing. Data access means either direct (no pointer) or indirect (pointer). Data packing means your data may be contiguous or not contiguous in memory, and may use strides to identify the jumps in memory consecutive indices need to take for each dimension.

    NumPy arrays provide a good model of strided direct data access, so we’ll use them for a refresher on the concepts of C and Fortran contiguous arrays, and data strides.

    Brief recap on C, Fortran and strided memory layouts

    The simplest data layout might be a C contiguous array. This is the default layout in NumPy and Cython arrays. C contiguous means that the array data is continuous in memory (see below) and that neighboring elements in the first dimension of the array are furthest apart in memory, whereas neighboring elements in the last dimension are closest together. For example, in NumPy:

    1. In [2]: arr = np.array([['0', '1', '2'], ['3', '4', '5']], dtype='S1')

    Here, arr[0, 0] and arr[0, 1] are one byte apart in memory, whereas arr[0, 0] and arr[1, 0] are 3 bytes apart. This leads us to the idea of strides. Each axis of the array has a stride length, which is the number of bytes needed to go from one element on this axis to the next element. In the case above, the strides for axes 0 and 1 will obviously be:

    1. In [3]: arr.strides
    2. Out[4]: (3, 1)

    For a 3D C contiguous array:

    1. In [5]: c_contig = np.arange(24, dtype=np.int8).reshape((2,3,4))
    2. In [6] c_contig.strides
    3. Out[6]: (12, 4, 1)

    A Fortran contiguous array has the opposite memory ordering, with the elements on the first axis closest together in memory:

    1. In [7]: f_contig = np.array(c_contig, order='F')
    2. In [8]: np.all(f_contig == c_contig)
    3. Out[8]: True
    4. In [9]: f_contig.strides
    5. Out[9]: (1, 2, 6)

    A contiguous array is one for which a single continuous block of memory contains all the data for the elements of the array, and therefore the memory block length is the product of number of elements in the array and the size of the elements in bytes. In the example above, the memory block is 2 * 3 * 4 * 1 bytes long, where 1 is the length of an int8.

    An array can be contiguous without being C or Fortran order:

    1. In [10]: c_contig.transpose((1, 0, 2)).strides
    2. Out[10]: (4, 12, 1)

    Slicing an NumPy array can easily make it not contiguous:

    1. In [11]: sliced = c_contig[:,1,:]
    2. In [12]: sliced.strides
    3. Out[12]: (12, 1)
    4. In [13]: sliced.flags
    5. Out[13]:
    6. C_CONTIGUOUS : False
    7. F_CONTIGUOUS : False
    8. OWNDATA : False
    9. WRITEABLE : True
    10. ALIGNED : True
    11. UPDATEIFCOPY : False

    As you’ll see in Specifying more general memory layouts, you can specify memory layout for any dimension of an memoryview. For any dimension for which you don’t specify a layout, then the data access is assumed to be direct, and the data packing assumed to be strided. For example, that will be the assumption for memoryviews like:

    1. int [:, :, :] my_memoryview = obj

    C and Fortran contiguous memoryviews

    You can specify C and Fortran contiguous layouts for the memoryview by using the ::1 step syntax at definition. For example, if you know for sure your memoryview will be on top of a 3D C contiguous layout, you could write:

    1. cdef int[:, :, ::1] c_contiguous = c_contig

    where c_contig could be a C contiguous NumPy array. The ::1 at the 3rd position means that the elements in this 3rd dimension will be one element apart in memory. If you know you will have a 3D Fortran contiguous array:

    1. cdef int[::1, :, :] f_contiguous = f_contig

    If you pass a non-contiguous buffer, for example

    1. # This array is C contiguous
    2. c_contig = np.arange(24).reshape((2,3,4))
    3. cdef int[:, :, ::1] c_contiguous = c_contig
    5. # But this isn't
    6. c_contiguous = np.array(c_contig, order='F')

    you will get a ValueError at runtime:

    1. /Users/mb312/dev_trees/minimal-cython/mincy.pyx in init mincy (mincy.c:17267)()
    2.     69
    3.     70 # But this isn't
    4. ---> 71 c_contiguous = np.array(c_contig, order='F')
    5.     72
    6.     73 # Show the sum of all the arrays before altering it
    8. /Users/mb312/dev_trees/minimal-cython/stringsource in View.MemoryView.memoryview_cwrapper (mincy.c:9995)()
    10. /Users/mb312/dev_trees/minimal-cython/stringsource in View.MemoryView.memoryview.__cinit__ (mincy.c:6799)()
    12. ValueError: ndarray is not C-contiguous

    Thus the ::1 in the slice type specification indicates in which dimension the data is contiguous. It can only be used to specify full C or Fortran contiguity.

    C and Fortran contiguous copies

    Copies can be made C or Fortran contiguous using the .copy() and .copy_fortran() methods:

    1. # This view is C contiguous
    2. cdef int[:, :, ::1] c_contiguous = myview.copy()
    4. # This view is Fortran contiguous
    5. cdef int[::1, :] f_contiguous_slice = myview.copy_fortran()

    Specifying more general memory layouts

    Data layout can be specified using the previously seen ::1 slice syntax, or by using any of the constants in cython.view. If no specifier is given in any dimension, then the data access is assumed to be direct, and the data packing assumed to be strided. If you don’t know whether a dimension will be direct or indirect (because you’re getting an object with a buffer interface from some library perhaps), then you can specify the generic flag, in which case it will be determined at runtime.

    The flags are as follows:

    • generic - strided and direct or indirect
    • strided - strided and direct (this is the default)
    • indirect - strided and indirect
    • contiguous - contiguous and direct
    • indirect_contiguous - the list of pointers is contiguous

    and they can be used like this:

    1. from cython cimport view
    3. # direct access in both dimensions, strided in the first dimension, contiguous in the last
    4. cdef int[:, ::view.contiguous] a
    6. # contiguous list of pointers to contiguous lists of ints
    7. cdef int[::view.indirect_contiguous, ::1] b
    9. # direct or indirect in the first dimension, direct in the second dimension
    10. # strided in both dimensions
    11. cdef int[::view.generic, :] c

    Only the first, last or the dimension following an indirect dimension may be specified contiguous:

    1. from cython cimport view
    3. # VALID
    4. cdef int[::view.indirect, ::1, :] a
    5. cdef int[::view.indirect, :, ::1] b
    6. cdef int[::view.indirect_contiguous, ::1, :] c
    1. cdef int[:, ::1] c_contig = ...
    3. # VALID
    4. cdef int[:, ::view.contiguous] myslice = c_contig[::2]
    6. # INVALID
    7. cdef int[:, ::1] myslice = c_contig[::2]

    The former case is valid because the last dimension remains contiguous, but the first dimension does not “follow” the last one anymore (meaning, it was strided already, but it is not C or Fortran contiguous any longer), since it was sliced.

    Memoryviews and the GIL

    As you will see from the section, memoryviews often do not need the GIL:

    1. cpdef int sum3d(int[:, :, :] arr) nogil:
    2.     ...

    In particular, you do not need the GIL for memoryview indexing, slicing or transposing. Memoryviews require the GIL for the copy methods (C and Fortran contiguous copies), or when the dtype is object and an object element is read or written.

    Memoryview Objects and Cython Arrays

    These typed memoryviews can be converted to Python memoryview objects (cython.view.memoryview). These Python objects are indexable, slicable and transposable in the same way that the original memoryviews are. They can also be converted back to Cython-space memoryviews at any time.

    They have the following attributes:

    And of course the aforementioned T attribute (Transposing). These attributes have the same semantics as in . For instance, to retrieve the original object:

    1. import numpy
    2. cimport numpy as cnp
    4. cdef cnp.int32_t[:] a = numpy.arange(10, dtype=numpy.int32)
    5. a = a[::2]
    7. print(a)
    9. print(a.base)
    11. # this prints:
    12. #    <MemoryView of 'ndarray' object>
    13. #    [0 2 4 6 8]
    14. #    [0 1 2 3 4 5 6 7 8 9]

    Note that this example returns the original object from which the view was obtained, and that the view was resliced in the meantime.

    Cython arrays

    Whenever a Cython memoryview is copied (using any of the copy or copy_fortran methods), you get a new memoryview slice of a newly created cython.view.array object. This array can also be used manually, and will automatically allocate a block of data. It can later be assigned to a C or Fortran contiguous slice (or a strided slice). It can be used like:

    1. from cython cimport view
    3. my_array = view.array(shape=(10, 2), itemsize=sizeof(int), format="i")
    4. cdef int[:, :] my_slice = my_array

    It also takes an optional argument mode (‘c’ or ‘fortran’) and a boolean allocate_buffer, that indicates whether a buffer should be allocated and freed when it goes out of scope:

    1. cdef view.array my_array = view.array(..., mode="fortran", allocate_buffer=False)
    2. my_array.data = <char *> my_data_pointer
    4. # define a function that can deallocate the data (if needed)
    5. my_array.callback_free_data = free

    You can also cast pointers to array, or C arrays to arrays:

    1. cdef view.array my_array = <int[:10, :2]> my_data_pointer
    2. cdef view.array my_array = <int[:, :]> my_c_array

    Of course, you can also immediately assign a cython.view.array to a typed memoryview slice. A C array may be assigned directly to a memoryview slice:

    1. cdef int[:, ::1] myslice = my_2d_c_array

    The arrays are indexable and slicable from Python space just like memoryview objects, and have the same attributes as memoryview objects.

    An alternative to cython.view.array is the array module in the Python standard library. In Python 3, the array.array type supports the buffer interface natively, so memoryviews work on top of it without additional setup.

    Starting with Cython 0.17, however, it is possible to use these arrays as buffer providers also in Python 2. This is done through explicitly cimporting the cpython.array module as follows:

    1. cimport cpython.array
    3. def sum_array(int[:] view):
    4.     """
    5.     >>> from array import array
    6.     >>> sum_array( array('i', [1,2,3]) )
    7.     6
    8.     """
    9.     cdef int total
    10.     for i in range(view.shape[0]):
    11.         total += view[i]
    12.     return total

    Note that the cimport also enables the old buffer syntax for the array type. Therefore, the following also works:

    1. from cpython cimport array
    3. def sum_array(array.array[int] arr):  # using old buffer syntax
    4.     ...

    Coercion to NumPy

    Memoryview (and array) objects can be coerced to a NumPy ndarray, without having to copy the data. You can e.g. do:

    1. cimport numpy as np
    2. import numpy as np
    4. numpy_array = np.asarray(<np.int32_t[:10, :10]> my_pointer)

    Of course, you are not restricted to using NumPy’s type (such as np.int32_t here), you can use any usable type.

    None Slices

    Although memoryview slices are not objects they can be set to None and they can be checked for being None as well:

    1. def func(double[:] myarray = None):
    2.     print(myarray is None)

    If the function requires real memory views as input, it is therefore best to reject None input straight away in the signature, which is supported in Cython 0.17 and later as follows:

    1. def func(double[:] myarray not None):
    2.     ...

    Unlike object attributes of extension classes, memoryview slices are not initialized to None.

    Pass data from a C function via pointer

    Since use of pointers in C is ubiquitous, here we give a quick example of how to call C functions whose arguments contain pointers. Let’s suppose you want to manage an array (allocate and deallocate) with NumPy (it can also be Python arrays, or anything that supports the buffer interface), but you want to perform computation on this array with an external C function implemented in C_func_file.c:

    This file comes with a header file called C_func_file.h containing:

    where arr points to the array and n is its size.

    You can call the function in a Cython file in the following way:

    Several things to note:

    • ::1 requests a C contiguous view, and fails if the buffer is not C contiguous. See C and Fortran contiguous memoryviews.
    • &arr_memview[0] can be understood as ‘the address of the first element of the memoryview’. For contiguous arrays, this is equivalent to the start address of the flat memory buffer.
    • arr_memview.shape[0] could have been replaced by arr_memview.size, arr.shape[0] or arr.size. But arr_memview.shape[0] is more efficient because it doesn’t require any Python interaction.
    • will perform computation in-place if the array passed is contiguous, and will return a new numpy array if arr is not contiguous.