The array interface protocol#

Note

This page describes the NumPy-specific API for accessing the contents of a NumPy array from other C extensions. PEP 3118The Revised Buffer Protocol introduces similar, standardized API to Python 2.6 and 3.0 for any extension module to use. Cython’s buffer array support uses the PEP 3118 API; see the Cython NumPy tutorial. Cython provides a way to write code that supports the buffer protocol with Python versions older than 2.6 because it has a backward-compatible implementation utilizing the array interface described here.

version:

3

The array interface (sometimes called array protocol) was created in 2005 as a means for array-like Python objects to reuse each other’s data buffers intelligently whenever possible. The homogeneous N-dimensional array interface is a default mechanism for objects to share N-dimensional array memory and information. The interface consists of a Python-side and a C-side using two attributes. Objects wishing to be considered an N-dimensional array in application code should support at least one of these attributes. Objects wishing to support an N-dimensional array in application code should look for at least one of these attributes and use the information provided appropriately.

This interface describes homogeneous arrays in the sense that each item of the array has the same “type”. This type can be very simple or it can be a quite arbitrary and complicated C-like structure.

There are two ways to use the interface: A Python side and a C-side. Both are separate attributes.

Python side#

This approach to the interface consists of the object having an __array_interface__ attribute.

object.__array_interface__#

A dictionary of items (3 required and 5 optional). The optional keys in the dictionary have implied defaults if they are not provided.

The keys are:

shape (required)

Tuple whose elements are the array size in each dimension. Each entry is an integer (a Python int). Note that these integers could be larger than the platform int or long could hold (a Python int is a C long). It is up to the code using this attribute to handle this appropriately; either by raising an error when overflow is possible, or by using long long as the C type for the shapes.

typestr (required)

A string providing the basic type of the homogeneous array The basic string format consists of 3 parts: a character describing the byteorder of the data (<: little-endian, >: big-endian, |: not-relevant), a character code giving the basic type of the array, and an integer providing the number of bytes the type uses.

The basic type character codes are:

t

Bit field (following integer gives the number of bits in the bit field).

b

Boolean (integer type where all values are only True or False)

i

Integer

u

Unsigned integer

f

Floating point

c

Complex floating point

m

Timedelta

M

Datetime

O

Object (i.e. the memory contains a pointer to PyObject)

S

String (fixed-length sequence of char)

U

Unicode (fixed-length sequence of Py_UCS4)

V

Other (void * – each item is a fixed-size chunk of memory)

descr (optional)

A list of tuples providing a more detailed description of the memory layout for each item in the homogeneous array. Each tuple in the list has two or three elements. Normally, this attribute would be used when typestr is V[0-9]+, but this is not a requirement. The only requirement is that the number of bytes represented in the typestr key is the same as the total number of bytes represented here. The idea is to support descriptions of C-like structs that make up array elements. The elements of each tuple in the list are

  1. A string providing a name associated with this portion of the datatype. This could also be a tuple of ('full name', 'basic_name') where basic name would be a valid Python variable name representing the full name of the field.

  2. Either a basic-type description string as in typestr or another list (for nested structured types)

  3. An optional shape tuple providing how many times this part of the structure should be repeated. No repeats are assumed if this is not given. Very complicated structures can be described using this generic interface. Notice, however, that each element of the array is still of the same data-type. Some examples of using this interface are given below.

Default: [('', typestr)]

data (optional)

A 2-tuple whose first argument is a Python integer that points to the data-area storing the array contents.

Note

When converting from C/C++ via PyLong_From* or high-level bindings such as Cython or pybind11, make sure to use an integer of sufficiently large bitness.

This pointer must point to the first element of data (in other words any offset is always ignored in this case). The second entry in the tuple is a read-only flag (true means the data area is read-only).

This attribute can also be an object exposing the buffer interface which will be used to share the data. If this key is not present (or returns None), then memory sharing will be done through the buffer interface of the object itself. In this case, the offset key can be used to indicate the start of the buffer. A reference to the object exposing the array interface must be stored by the new object if the memory area is to be secured.

Default: None

strides (optional)

Either None to indicate a C-style contiguous array or a tuple of strides which provides the number of bytes needed to jump to the next array element in the corresponding dimension. Each entry must be an integer (a Python int). As with shape, the values may be larger than can be represented by a C int or long; the calling code should handle this appropriately, either by raising an error, or by using long long in C. The default is None which implies a C-style contiguous memory buffer. In this model, the last dimension of the array varies the fastest. For example, the default strides tuple for an object whose array entries are 8 bytes long and whose shape is (10, 20, 30) would be (4800, 240, 8).

Default: None (C-style contiguous)

mask (optional)

None or an object exposing the array interface. All elements of the mask array should be interpreted only as true or not true indicating which elements of this array are valid. The shape of this object should be “broadcastable” to the shape of the original array.

Default: None (All array values are valid)

offset (optional)

An integer offset into the array data region. This can only be used when data is None or returns a memoryview object.

Default: 0.

version (required)

An integer showing the version of the interface (i.e. 3 for this version). Be careful not to use this to invalidate objects exposing future versions of the interface.

C-struct access#

This approach to the array interface allows for faster access to an array using only one attribute lookup and a well-defined C-structure.

object.__array_struct__#

A PyCapsule whose pointer member contains a pointer to a filled PyArrayInterface structure. Memory for the structure is dynamically created and the PyCapsule is also created with an appropriate destructor so the retriever of this attribute simply has to apply Py_DECREF to the object returned by this attribute when it is finished. Also, either the data needs to be copied out, or a reference to the object exposing this attribute must be held to ensure the data is not freed. Objects exposing the __array_struct__ interface must also not reallocate their memory if other objects are referencing them.

The PyArrayInterface structure is defined in numpy/ndarrayobject.h as:

typedef struct {
  int two;              /* contains the integer 2 -- simple sanity check */
  int nd;               /* number of dimensions */
  char typekind;        /* kind in array --- character code of typestr */
  int itemsize;         /* size of each element */
  int flags;            /* flags indicating how the data should be interpreted */
                        /*   must set ARR_HAS_DESCR bit to validate descr */
  Py_ssize_t *shape;    /* A length-nd array of shape information */
  Py_ssize_t *strides;  /* A length-nd array of stride information */
  void *data;           /* A pointer to the first element of the array */
  PyObject *descr;      /* NULL or data-description (same as descr key
                                of __array_interface__) -- must set ARR_HAS_DESCR
                                flag or this will be ignored. */
} PyArrayInterface;

The flags member may consist of 5 bits showing how the data should be interpreted and one bit showing how the Interface should be interpreted. The data-bits are NPY_ARRAY_C_CONTIGUOUS (0x1), NPY_ARRAY_F_CONTIGUOUS (0x2), NPY_ARRAY_ALIGNED (0x100), NPY_ARRAY_NOTSWAPPED (0x200), and NPY_ARRAY_WRITEABLE (0x400). A final flag NPY_ARR_HAS_DESCR (0x800) indicates whether or not this structure has the arrdescr field. The field should not be accessed unless this flag is present.

NPY_ARR_HAS_DESCR#

New since June 16, 2006:

In the past most implementations used the desc member of the PyCObject (now PyCapsule) itself (do not confuse this with the “descr” member of the PyArrayInterface structure above — they are two separate things) to hold the pointer to the object exposing the interface. This is now an explicit part of the interface. Be sure to take a reference to the object and call PyCapsule_SetContext before returning the PyCapsule, and configure a destructor to decref this reference.

Note

__array_struct__ is considered legacy and should not be used for new code. Use the buffer protocol or the DLPack protocol numpy.from_dlpack instead.

Type description examples#

For clarity it is useful to provide some examples of the type description and corresponding __array_interface__ ‘descr’ entries. Thanks to Scott Gilbert for these examples:

In every case, the ‘descr’ key is optional, but of course provides more information which may be important for various applications:

* Float data
    typestr == '>f4'
    descr == [('','>f4')]

* Complex double
    typestr == '>c8'
    descr == [('real','>f4'), ('imag','>f4')]

* RGB Pixel data
    typestr == '|V3'
    descr == [('r','|u1'), ('g','|u1'), ('b','|u1')]

* Mixed endian (weird but could happen).
    typestr == '|V8' (or '>u8')
    descr == [('big','>i4'), ('little','<i4')]

* Nested structure
    struct {
        int ival;
        struct {
            unsigned short sval;
            unsigned char bval;
            unsigned char cval;
        } sub;
    }
    typestr == '|V8' (or '<u8' if you want)
    descr == [('ival','<i4'), ('sub', [('sval','<u2'), ('bval','|u1'), ('cval','|u1') ]) ]

* Nested array
    struct {
        int ival;
        double data[16*4];
    }
    typestr == '|V516'
    descr == [('ival','>i4'), ('data','>f8',(16,4))]

* Padded structure
    struct {
        int ival;
        double dval;
    }
    typestr == '|V16'
    descr == [('ival','>i4'),('','|V4'),('dval','>f8')]

It should be clear that any structured type could be described using this interface.

Differences with array interface (version 2)#

The version 2 interface was very similar. The differences were largely aesthetic. In particular:

  1. The PyArrayInterface structure had no descr member at the end (and therefore no flag ARR_HAS_DESCR)

  2. The context member of the PyCapsule (formally the desc member of the PyCObject) returned from __array_struct__ was not specified. Usually, it was the object exposing the array (so that a reference to it could be kept and destroyed when the C-object was destroyed). It is now an explicit requirement that this field be used in some way to hold a reference to the owning object.

    Note

    Until August 2020, this said:

    Now it must be a tuple whose first element is a string with “PyArrayInterface Version #” and whose second element is the object exposing the array.

    This design was retracted almost immediately after it was proposed, in <https://mail.python.org/pipermail/numpy-discussion/2006-June/020995.html>. Despite 14 years of documentation to the contrary, at no point was it valid to assume that __array_interface__ capsules held this tuple content.

  3. The tuple returned from __array_interface__['data'] used to be a hex-string (now it is an integer or a long integer).

  4. There was no __array_interface__ attribute instead all of the keys (except for version) in the __array_interface__ dictionary were their own attribute: Thus to obtain the Python-side information you had to access separately the attributes:

    • __array_data__

    • __array_shape__

    • __array_strides__

    • __array_typestr__

    • __array_descr__

    • __array_offset__

    • __array_mask__