Typed memoryviews allow efficient access to memory buffers, such as those
underlying NumPy arrays, without incurring any Python overhead.
Memoryviews are similar to the current NumPy array buffer support
(
np.ndarray[np.float64_t,
ndim=2]
), but
they have more features and cleaner syntax.
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, modulelevel, cdef class attribute, etc) and can be obtained from nearly any object that exposes writable buffer through the `PEP 3118`_ buffer interface.
If you are used to working with NumPy, the following examples should get you started with Cython memory views.
from cython.view cimport array as cvarray
import numpy as np
# Memoryview on a NumPy array
narr = np.arange(27, dtype=np.dtype("i")).reshape((3, 3, 3))
cdef int [:, :, :] narr_view = narr
# Memoryview on a C array
cdef int carr[3][3][3]
cdef int [:, :, :] carr_view = carr
# Memoryview on a Cython array
cyarr = cvarray(shape=(3, 3, 3), itemsize=sizeof(int), format="i")
cdef int [:, :, :] cyarr_view = cyarr
# Show the sum of all the arrays before altering it
print("NumPy sum of the NumPy array before assignments: %s" % narr.sum())
# We can copy the values from one memoryview into another using a single
# statement, by either indexing with ... or (NumPystyle) with a colon.
carr_view[...] = narr_view
cyarr_view[:] = narr_view
# NumPystyle syntax for assigning a single value to all elements.
narr_view[:, :, :] = 3
# Just to distinguish the arrays
carr_view[0, 0, 0] = 100
cyarr_view[0, 0, 0] = 1000
# Assigning into the memoryview on the NumPy array alters the latter
print("NumPy sum of NumPy array after assignments: %s" % narr.sum())
# A function using a memoryview does not usually need the GIL
cpdef int sum3d(int[:, :, :] arr) nogil:
cdef size_t i, j, k, I, J, K
cdef int total = 0
I = arr.shape[0]
J = arr.shape[1]
K = arr.shape[2]
for i in range(I):
for j in range(J):
for k in range(K):
total += arr[i, j, k]
return total
# A function accepting a memoryview knows how to use a NumPy array,
# a C array, a Cython array...
print("Memoryview sum of NumPy array is %s" % sum3d(narr))
print("Memoryview sum of C array is %s" % sum3d(carr))
print("Memoryview sum of Cython array is %s" % sum3d(cyarr))
# ... and of course, a memoryview.
print("Memoryview sum of C memoryview is %s" % sum3d(carr_view))
This code should give the following output:
NumPy sum of the NumPy array before assignments: 351
NumPy sum of NumPy array after assignments: 81
Memoryview sum of NumPy array is 81
Memoryview sum of C array is 451
Memoryview sum of Cython array is 1351
Memoryview sum of C memoryview is 451
Memory views use Python slicing syntax in a similar way as NumPy.
To create a complete view on a onedimensional int buffer:
cdef int[:] view1D = exporting_object
A complete 3D view:
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:
cdef int[1:102:2,:] partial_view = exporting_object
This also works conveniently as function arguments:
def process_3d_buffer(int[1:102:2,:] view not None):
...
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:
import numpy as np
def process_buffer(int[:,:] input_view not None,
int[:,:] output_view=None):
if output_view is None:
# Creating a default view, e.g.
output_view = np.empty_like(input_view)
# process 'input_view' into 'output_view'
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
.
In Cython, index access on memory views is automatically translated
into memory addresses. The following code requests a twodimensional
memory view of C
int
typed items and indexes into it:
cdef int[:,:] buf = exporting_object
print(buf[1,2])
Negative indices work as well, counting from the end of the respective dimension:
print(buf[1,2])
The following function loops over each dimension of a 2D array and adds 1 to each item:
import numpy as np
def add_one(int[:,:] buf):
for x in range(buf.shape[0]):
for y in range(buf.shape[1]):
buf[x, y] += 1
# exporting_object must be a Python object
# implementing the buffer interface, e.g. a numpy array.
exporting_object = np.zeros((10, 20), dtype=np.intc)
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:
import numpy as np
exporting_object = np.arange(0, 15 * 10 * 20, dtype=np.intc).reshape((15, 10, 20))
cdef int[:, :, :] my_view = exporting_object
# These are all equivalent
my_view[10]
my_view[10, :, :]
my_view[10, ...]
Memory views can be copied in place:
import numpy as np
cdef int[:, :, :] to_view, from_view
to_view = np.empty((20, 15, 30), dtype=np.intc)
from_view = np.ones((20, 15, 30), dtype=np.intc)
# copy the elements in from_view to to_view
to_view[...] = from_view
# or
to_view[:] = from_view
# or
to_view[:, :, :] = from_view
They can also be copied with the
copy()
and
copy_fortran()
methods; see
C and Fortran contiguous copies
.
In most cases (see below), the memoryview can be transposed in the same way that NumPy slices can be transposed:
import numpy as np
array = np.arange(20, dtype=np.intc).reshape((2, 10))
cdef int[:, ::1] c_contig = array
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). 见 Specifying more general memory layouts 了解细节。
As for NumPy, new axes can be introduced by indexing an array with
None
cdef double[:] myslice = np.linspace(0, 10, num=50)
# 2D array with shape (1, 50)
myslice[None] # or
myslice[None, :]
# 2D array with shape (50, 1)
myslice[:, None]
# 3D array with shape (1, 10, 1)
myslice[None, 10:20:2, None]
One may mix new axis indexing with all other forms of indexing and slicing. See also an example .
Since Cython 0.28, the memoryview item type can be declared as
const
to
support readonly buffers as input:
import numpy as np
cdef const double[:] myslice # const item type => readonly view
a = np.linspace(0, 10, num=50)
a.setflags(write=False)
myslice = a
Using a nonconst memoryview with a binary Python string produces a runtime error.
You can solve this issue with a
const
memoryview:
cdef bint is_y_in(const unsigned char[:] string_view):
cdef int i
for i in range(string_view.shape[0]):
if string_view[i] == b'y':
return True
return False
print(is_y_in(b'hello world')) # False
print(is_y_in(b'hello Cython')) # True
Note that this does not require the input buffer to be readonly:
a = np.linspace(0, 10, num=50)
myslice = a # readonly view of a writable buffer
Writable buffers are still accepted by
const
views, but readonly
buffers are not accepted for nonconst, writable views:
cdef double[:] myslice # a normal read/write memory view
a = np.linspace(0, 10, num=50)
a.setflags(write=False)
myslice = a # ERROR: requesting writable memory view from readonly buffer!
You will probably prefer memoryviews to the older syntax because:
For example, this is the old syntax equivalent of the
sum3d
function above:
cpdef int old_sum3d(object[int, ndim=3, mode='strided'] arr):
cdef int I, J, K, total = 0
I = arr.shape[0]
J = arr.shape[1]
K = arr.shape[2]
for i in range(I):
for j in range(J):
for k in range(K):
total += arr[i, j, k]
return total
Note that we can’t use
nogil
for the buffer version of the function as we
could for the memoryview version of
sum3d
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:
In [2]: import numpy as np
In [3]: arr = np.zeros((40, 40, 40), dtype=int)
In [4]: timeit r15 old_sum3d(arr)
1000 loops, best of 15: 298 us per loop
In [5]: timeit r15 sum3d(arr)
1000 loops, best of 15: 219 us per loop
Cython memoryviews support nearly all objects exporting the interface of Python `new style buffers`_ . This is the buffer interface described in `PEP 3118`_ . NumPy arrays support this interface, as do Cython arrays . 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 newtype 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.
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.
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:
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:
In [3]: arr.strides
Out[4]: (3, 1)
For a 3D C contiguous array:
In [5]: c_contig = np.arange(24, dtype=np.int8).reshape((2,3,4))
In [6] c_contig.strides
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:
In [7]: f_contig = np.array(c_contig, order='F')
In [8]: np.all(f_contig == c_contig)
Out[8]: True
In [9]: f_contig.strides
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:
In [10]: c_contig.transpose((1, 0, 2)).strides
Out[10]: (4, 12, 1)
Slicing an NumPy array can easily make it not contiguous:
In [11]: sliced = c_contig[:,1,:]
In [12]: sliced.strides
Out[12]: (12, 1)
In [13]: sliced.flags
Out[13]:
C_CONTIGUOUS : False
F_CONTIGUOUS : False
OWNDATA : False
WRITEABLE : True
ALIGNED : True
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:
int [:, :, :] my_memoryview = obj
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:
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:
cdef int[::1, :, :] f_contiguous = f_contig
If you pass a noncontiguous buffer, for example
# This array is C contiguous
c_contig = np.arange(24).reshape((2,3,4))
cdef int[:, :, ::1] c_contiguous = c_contig
# But this isn't
c_contiguous = np.array(c_contig, order='F')
you will get a
ValueError
at runtime:
/Users/mb312/dev_trees/minimalcython/mincy.pyx in init mincy (mincy.c:17267)()
69
70 # But this isn't
> 71 c_contiguous = np.array(c_contig, order='F')
72
73 # Show the sum of all the arrays before altering it
/Users/mb312/dev_trees/minimalcython/stringsource in View.MemoryView.memoryview_cwrapper (mincy.c:9995)()
/Users/mb312/dev_trees/minimalcython/stringsource in View.MemoryView.memoryview.__cinit__ (mincy.c:6799)()
ValueError: ndarray is not Ccontiguous
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.
Copies can be made C or Fortran contiguous using the
.copy()
and
.copy_fortran()
methods:
# This view is C contiguous
cdef int[:, :, ::1] c_contiguous = myview.copy()
# This view is Fortran contiguous
cdef int[::1, :] f_contiguous_slice = myview.copy_fortran()
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:
and they can be used like this:
from cython cimport view
# direct access in both dimensions, strided in the first dimension, contiguous in the last
cdef int[:, ::view.contiguous] a
# contiguous list of pointers to contiguous lists of ints
cdef int[::view.indirect_contiguous, ::1] b
# direct or indirect in the first dimension, direct in the second dimension
# strided in both dimensions
cdef int[::view.generic, :] c
Only the first, last or the dimension following an indirect dimension may be specified contiguous:
from cython cimport view
# VALID
cdef int[::view.indirect, ::1, :] a
cdef int[::view.indirect, :, ::1] b
cdef int[::view.indirect_contiguous, ::1, :] c
# INVALID
cdef int[::view.contiguous, ::view.indirect, :] d
cdef int[::1, ::view.indirect, :] e
The difference between the contiguous flag and the ::1 specifier is that the former specifies contiguity for only one dimension, whereas the latter specifies contiguity for all following (Fortran) or preceding (C) dimensions:
cdef int[:, ::1] c_contig = ...
# VALID
cdef int[:, ::view.contiguous] myslice = c_contig[::2]
# INVALID
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.
As you will see from the Quickstart section, memoryviews often do not need the GIL:
cpdef int sum3d(int[:, :, :] arr) nogil:
...
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.
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 Cythonspace memoryviews at any time.
They have the following attributes:
shape
: size in each dimension, as a tuple.strides
: stride along each dimension, in bytes.suboffsets
ndim
: number of dimensions.size
: total number of items in the view (product of the shape).itemsize
: size, in bytes, of the items in the view.nbytes
: equal tosize
timesitemsize
.base
And of course the aforementioned
T
attribute (
Transposing
).
These attributes have the same semantics as in
NumPy
. For instance, to
retrieve the original object:
import numpy
cimport numpy as cnp
cdef cnp.int32_t[:] a = numpy.arange(10, dtype=numpy.int32)
a = a[::2]
print(a)
print(numpy.asarray(a))
print(a.base)
# this prints:
# <MemoryView of 'ndarray' object>
# [0 2 4 6 8]
# [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.
Whenever a Cython memoryview is copied (using any of the
copy
或
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:
from cython cimport view
my_array = view.array(shape=(10, 2), itemsize=sizeof(int), format="i")
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:
cdef view.array my_array = view.array(..., mode="fortran", allocate_buffer=False)
my_array.data = <char *> my_data_pointer
# define a function that can deallocate the data (if needed)
my_array.callback_free_data = free
You can also cast pointers to array, or C arrays to arrays:
cdef view.array my_array = <int[:10, :2]> my_data_pointer
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:
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:
cimport cpython.array
def sum_array(int[:] view):
"""
>>> from array import array
>>> sum_array( array('i', [1,2,3]) )
6
"""
cdef int total
for i in range(view.shape[0]):
total += view[i]
return total
Note that the cimport also enables the old buffer syntax for the array type. Therefore, the following also works:
from cpython cimport array
def sum_array(array.array[int] arr): # using old buffer syntax
...
Memoryview (and array) objects can be coerced to a NumPy ndarray, without having to copy the data. You can e.g. do:
cimport numpy as np
import numpy as np
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.
Although memoryview slices are not objects they can be set to None and they can be checked for being None as well:
def func(double[:] myarray = None):
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:
def func(double[:] myarray not None):
...
Unlike object attributes of extension classes, memoryview slices are not initialized to None.
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
:
1 2 3 4 5 6 7 8 9 
#include "C_func_file.h"
void multiply_by_10_in_C(double arr[], unsigned int n)
{
unsigned int i;
for (i = 0; i < n; i++) {
arr[i] *= 10;
}
}

This file comes with a header file called
C_func_file.h
containing:
1 2 3 4 5 6 
#ifndef C_FUNC_FILE_H
#define C_FUNC_FILE_H
void multiply_by_10_in_C(double arr[], unsigned int n);
#endif

where
arr
points to the array and
n
is its size.
You can call the function in a Cython file in the following way:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 
cdef extern from "C_func_file.c":
# C is include here so that it doesn't need to be compiled externally
pass
cdef extern from "C_func_file.h":
void multiply_by_10_in_C(double *, unsigned int)
import numpy as np
def multiply_by_10(arr): # 'arr' is a onedimensional numpy array
if not arr.flags['C_CONTIGUOUS']:
arr = np.ascontiguousarray(arr) # Makes a contiguous copy of the numpy array.
cdef double[::1] arr_memview = arr
multiply_by_10_in_C(&arr_memview[0], arr_memview.shape[0])
return arr
a = np.ones(5, dtype=np.double)
print(multiply_by_10(a))
b = np.ones(10, dtype=np.double)
b = b[::2] # b is not contiguous.
print(multiply_by_10(b)) # but our function still works as expected.

::1
requests a C contiguous view, and fails if the buffer is not C contiguous.
见
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]
或
arr.size
. But
arr_memview.shape[0]
is more efficient
because it doesn’t require any Python interaction.
multiply_by_10
will perform computation inplace if the array passed is contiguous,
and will return a new numpy array if
arr
is not contiguous.
This way, you can call the C function similar to a normal Python function, and leave all the memory management and cleanup to NumPy arrays and Python’s object handling. For the details of how to compile and call functions in C files, see 使用 C 库 .