缓冲协议

在 Python 中可使用一些对象来包装对底层内存数组或称 缓冲 的访问。此类对象包括内置的 bytes 和 bytearray 以及一些如 array.array 这样的扩展类型。第三方库也可能会为了特殊的目的而定义它们自己的类型，例如用于图像处理和数值分析等。

虽然这些类型中的每一种都有自己的语义，但它们具有由可能较大的内存缓冲区支持的共同特征。 在某些情况下，希望直接访问该缓冲区而无需中间复制。

Python 以 缓冲协议 的形式在 C 层级上提供这样的功能。 此协议包括两个方面:

  • 在生产者这一方面，该类型的协议可以导出一个“缓冲区接口”，允许公开它的底层缓冲区信息。该接口的描述信息在 Buffer Object Structures 一节中；
  • 在消费者一侧，有几种方法可用于获得指向对象的原始底层数据的指针（例如一个方法的形参）。

一些简单的对象例如 bytes 和 bytearray 会以面向字节的形式公开它们的底层缓冲区。 也可能会用其他形式；例如 array.array 所公开的元素可以是多字节值。

An example consumer of the buffer interface is the write() method of file objects: any object that can export a series of bytes through the buffer interface can be written to a file. While write() only needs read-only access to the internal contents of the object passed to it, other methods such as readinto() need write access to the contents of their argument. The buffer interface allows objects to selectively allow or reject exporting of read-write and read-only buffers.

There are two ways for a consumer of the buffer interface to acquire a buffer over a target object:

In both cases, PyBuffer_Release() must be called when the buffer isn’t needed anymore. Failure to do so could lead to various issues such as resource leaks.

Buffer structure

Buffer structures (or simply “buffers”) are useful as a way to expose the binary data from another object to the Python programmer. They can also be used as a zero-copy slicing mechanism. Using their ability to reference a block of memory, it is possible to expose any data to the Python programmer quite easily. The memory could be a large, constant array in a C extension, it could be a raw block of memory for manipulation before passing to an operating system library, or it could be used to pass around structured data in its native, in-memory format.

Contrary to most data types exposed by the Python interpreter, buffers are not PyObject pointers but rather simple C structures. This allows them to be created and copied very simply. When a generic wrapper around a buffer is needed, a memoryview object can be created.

For short instructions how to write an exporting object, see Buffer Object Structures. For obtaining a buffer, see PyObject_GetBuffer().

Py_buffer
void *buf

A pointer to the start of the logical structure described by the buffer fields. This can be any location within the underlying physical memory block of the exporter. For example, with negative strides the value may point to the end of the memory block.

For contiguous arrays, the value points to the beginning of the memory block.

void *obj

A new reference to the exporting object. The reference is owned by the consumer and automatically decremented and set to NULL by PyBuffer_Release(). The field is the equivalent of the return value of any standard C-API function.

As a special case, for temporary buffers that are wrapped by PyMemoryView_FromBuffer() or PyBuffer_FillInfo() this field is NULL. In general, exporting objects MUST NOT use this scheme.

Py_ssize_t len

product(shape) * itemsize. For contiguous arrays, this is the length of the underlying memory block. For non-contiguous arrays, it is the length that the logical structure would have if it were copied to a contiguous representation.

Accessing ((char *)buf)[0] up to ((char *)buf)[len-1] is only valid if the buffer has been obtained by a request that guarantees contiguity. In most cases such a request will be PyBUF_SIMPLE or PyBUF_WRITABLE.

int readonly

An indicator of whether the buffer is read-only. This field is controlled by the PyBUF_WRITABLE flag.

Py_ssize_t itemsize

Item size in bytes of a single element. Same as the value of struct.calcsize() called on non-NULL format values.

Important exception: If a consumer requests a buffer without the PyBUF_FORMAT flag, format will be set to NULL, but itemsize still has the value for the original format.

If shape is present, the equality product(shape) * itemsize == len still holds and the consumer can use itemsize to navigate the buffer.

If shape is NULL as a result of a PyBUF_SIMPLE or a PyBUF_WRITABLE request, the consumer must disregard itemsize and assume itemsize == 1.

const char *format

A NUL terminated string in struct module style syntax describing the contents of a single item. If this is NULL, "B" (unsigned bytes) is assumed.

This field is controlled by the PyBUF_FORMAT flag.

int ndim

The number of dimensions the memory represents as an n-dimensional array. If it is 0, buf points to a single item representing a scalar. In this case, shape, strides and suboffsets MUST be NULL.

The macro PyBUF_MAX_NDIM limits the maximum number of dimensions to 64. Exporters MUST respect this limit, consumers of multi-dimensional buffers SHOULD be able to handle up to PyBUF_MAX_NDIM dimensions.

Py_ssize_t *shape

An array of Py_ssize_t of length ndim indicating the shape of the memory as an n-dimensional array. Note that shape[0] * ... * shape[ndim-1] * itemsize MUST be equal to len.

Shape values are restricted to shape[n] >= 0. The case shape[n] == 0 requires special attention. See complex arrays for further information.

The shape array is read-only for the consumer.

Py_ssize_t *strides

An array of Py_ssize_t of length ndim giving the number of bytes to skip to get to a new element in each dimension.

Stride values can be any integer. For regular arrays, strides are usually positive, but a consumer MUST be able to handle the case strides[n] <= 0. See complex arrays for further information.

The strides array is read-only for the consumer.

Py_ssize_t *suboffsets

An array of Py_ssize_t of length ndim. If suboffsets[n] >= 0, the values stored along the nth dimension are pointers and the suboffset value dictates how many bytes to add to each pointer after de-referencing. A suboffset value that is negative indicates that no de-referencing should occur (striding in a contiguous memory block).

If all suboffsets are negative (i.e. no de-referencing is needed), then this field must be NULL (the default value).

This type of array representation is used by the Python Imaging Library (PIL). See complex arrays for further information how to access elements of such an array.

The suboffsets array is read-only for the consumer.

void *internal

This is for use internally by the exporting object. For example, this might be re-cast as an integer by the exporter and used to store flags about whether or not the shape, strides, and suboffsets arrays must be freed when the buffer is released. The consumer MUST NOT alter this value.

Buffer request types

Buffers are usually obtained by sending a buffer request to an exporting object via PyObject_GetBuffer(). Since the complexity of the logical structure of the memory can vary drastically, the consumer uses the flags argument to specify the exact buffer type it can handle.

All Py_buffer fields are unambiguously defined by the request type.

request-independent fields

The following fields are not influenced by flags and must always be filled in with the correct values: obj, buf, len, itemsize, ndim.

readonly, format

PyBUF_WRITABLE

Controls the readonly field. If set, the exporter MUST provide a writable buffer or else report failure. Otherwise, the exporter MAY provide either a read-only or writable buffer, but the choice MUST be consistent for all consumers.

PyBUF_FORMAT

Controls the format field. If set, this field MUST be filled in correctly. Otherwise, this field MUST be NULL.

PyBUF_WRITABLE can be |’d to any of the flags in the next section. Since PyBUF_SIMPLE is defined as 0, PyBUF_WRITABLE can be used as a stand-alone flag to request a simple writable buffer.

PyBUF_FORMAT can be |’d to any of the flags except PyBUF_SIMPLE. The latter already implies format B (unsigned bytes).

shape, strides, suboffsets

The flags that control the logical structure of the memory are listed in decreasing order of complexity. Note that each flag contains all bits of the flags below it.

请求 shape strides suboffsets
PyBUF_INDIRECT
 yes yes 如果需要的话
PyBUF_STRIDES
 yes yes NULL
PyBUF_ND
 yes NULL NULL
PyBUF_SIMPLE
 NULL NULL NULL

连续性的请求

C or Fortran contiguity can be explicitly requested, with and without stride information. Without stride information, the buffer must be C-contiguous.

请求 shape strides suboffsets contig
PyBUF_C_CONTIGUOUS
 yes yes NULL C
PyBUF_F_CONTIGUOUS
 yes yes NULL F
PyBUF_ANY_CONTIGUOUS
 yes yes NULL C 或 F
PyBUF_ND
 yes NULL NULL C

compound requests

All possible requests are fully defined by some combination of the flags in the previous section. For convenience, the buffer protocol provides frequently used combinations as single flags.

In the following table U stands for undefined contiguity. The consumer would have to call PyBuffer_IsContiguous() to determine contiguity.

请求 shape strides suboffsets contig 只读 格式
PyBUF_FULL
 yes yes 如果需要的话 U 0 yes
PyBUF_FULL_RO
 yes yes 如果需要的话 U 1 或 0 yes
PyBUF_RECORDS
 yes yes NULL U 0 yes
PyBUF_RECORDS_RO
 yes yes NULL U 1 或 0 yes
PyBUF_STRIDED
 yes yes NULL U 0 NULL
PyBUF_STRIDED_RO
 yes yes NULL U 1 或 0 NULL
PyBUF_CONTIG
 yes NULL NULL C 0 NULL
PyBUF_CONTIG_RO
 yes NULL NULL C 1 或 0 NULL

复杂数组

NumPy-style: shape and strides

The logical structure of NumPy-style arrays is defined by itemsize, ndim, shape and strides.

If ndim == 0, the memory location pointed to by buf is interpreted as a scalar of size itemsize. In that case, both shape and strides are NULL.

If strides is NULL, the array is interpreted as a standard n-dimensional C-array. Otherwise, the consumer must access an n-dimensional array as follows:

ptr = (char *)buf + indices[0] * strides[0] + ... + indices[n-1] * strides[n-1] item = *((typeof(item) *)ptr);

As noted above, buf can point to any location within the actual memory block. An exporter can check the validity of a buffer with this function:

def verify_structure(memlen, itemsize, ndim, shape, strides, offset):
    """Verify that the parameters represent a valid array within
       the bounds of the allocated memory:
           char *mem: start of the physical memory block
           memlen: length of the physical memory block
           offset: (char *)buf - mem
    """
    if offset % itemsize:
        return False
    if offset < 0 or offset+itemsize > memlen:
        return False
    if any(v % itemsize for v in strides):
        return False

    if ndim <= 0:
        return ndim == 0 and not shape and not strides
    if 0 in shape:
        return True

    imin = sum(strides[j]*(shape[j]-1) for j in range(ndim)
               if strides[j] <= 0)
    imax = sum(strides[j]*(shape[j]-1) for j in range(ndim)
               if strides[j] > 0)

    return 0 <= offset+imin and offset+imax+itemsize <= memlen

PIL-style: shape, strides and suboffsets

In addition to the regular items, PIL-style arrays can contain pointers that must be followed in order to get to the next element in a dimension. For example, the regular three-dimensional C-array char v[2][2][3] can also be viewed as an array of 2 pointers to 2 two-dimensional arrays: char (*v[2])[2][3]. In suboffsets representation, those two pointers can be embedded at the start of buf, pointing to two char x[2][3] arrays that can be located anywhere in memory.

Here is a function that returns a pointer to the element in an N-D array pointed to by an N-dimensional index when there are both non-NULL strides and suboffsets:

void *get_item_pointer(int ndim, void *buf, Py_ssize_t *strides,
                       Py_ssize_t *suboffsets, Py_ssize_t *indices) {
    char *pointer = (char*)buf;
    int i;
    for (i = 0; i < ndim; i++) {
        pointer += strides[i] * indices[i];
        if (suboffsets[i] >=0 ) {
            pointer = *((char**)pointer) + suboffsets[i];
        }
    }
    return (void*)pointer;
}