python-peps/pep-3118.txt

PEP: 3118
Title: Revising the buffer protocol
Version: $Revision$
Last-Modified: $Date$
Authors: Travis Oliphant <oliphant@ee.byu.edu>, Carl Banks <pythondev@aerojockey.com>
Status: Draft
Type: Standards Track
Content-Type: text/x-rst
Created: 28-Aug-2006
Python-Version: 3000

Abstract
========

This PEP proposes re-designing the buffer interface (PyBufferProcs
function pointers) to improve the way Python allows memory sharing
in Python 3.0

In particular, it is proposed that the character buffer portion 
of the API be elminated and the multiple-segment portion be 
re-designed in conjunction with allowing for strided memory
to be shared.   In addition, the new buffer interface will 
allow the sharing of any multi-dimensional nature of the
memory and what data-format the memory contains. 

This interface will allow any extension module to either 
create objects that share memory or create algorithms that
use and manipulate raw memory from arbitrary objects that 
export the interface. 


Rationale
=========

The Python 2.X buffer protocol allows different Python types to
exchange a pointer to a sequence of internal buffers.  This
functionality is *extremely* useful for sharing large segments of
memory between different high-level objects, but it is too limited and
has issues:

1. There is the little used "sequence-of-segments" option
   (bf_getsegcount) that is not well motivated. 

2. There is the apparently redundant character-buffer option
   (bf_getcharbuffer)

3. There is no way for a consumer to tell the buffer-API-exporting
   object it is "finished" with its view of the memory and
   therefore no way for the exporting object to be sure that it is
   safe to reallocate the pointer to the memory that it owns (for
   example, the array object reallocating its memory after sharing
   it with the buffer object which held the original pointer led
   to the infamous buffer-object problem).

4. Memory is just a pointer with a length. There is no way to
   describe what is "in" the memory (float, int, C-structure, etc.)

5. There is no shape information provided for the memory.  But,
   several array-like Python types could make use of a standard
   way to describe the shape-interpretation of the memory
   (wxPython, GTK, pyQT, CVXOPT, PyVox, Audio and Video
   Libraries, ctypes, NumPy, data-base interfaces, etc.)

6. There is no way to share discontiguous memory (except through
   the sequence of segments notion).  

   There are two widely used libraries that use the concept of
   discontiguous memory: PIL and NumPy.  Their view of discontiguous
   arrays is different, though.  The proposed buffer interface allows
   sharing of either memory model.  Exporters will use only one        
   approach and consumers may choose to support discontiguous 
   arrays of each type however they choose. 

   NumPy uses the notion of constant striding in each dimension as its
   basic concept of an array. With this concept, a simple sub-region
   of a larger array can be described without copying the data.   T
   Thus, stride information is the additional information that must be
   shared. 

   The PIL uses a more opaque memory representation. Sometimes an
   image is contained in a contiguous segment of memory, but sometimes
   it is contained in an array of pointers to the contiguous segments
   (usually lines) of the image.  The PIL is where the idea of multiple
   buffer segments in the original buffer interface came from.   

   NumPy's strided memory model is used more often in computational
   libraries and because it is so simple it makes sense to support
   memory sharing using this model.  The PIL memory model is sometimes 
   used in C-code where a 2-d array can be then accessed using double
   pointer indirection:  e.g. image[i][j].  

   The buffer interface should allow the object to export either of these
   memory models.  Consumers are free to either require contiguous memory
   or write code to handle one or both of these memory models. 

Proposal Overview
=================

* Eliminate the char-buffer and multiple-segment sections of the
  buffer-protocol.

* Unify the read/write versions of getting the buffer.

* Add a new function to the interface that should be called when
  the consumer object is "done" with the memory area.  

* Add a new variable to allow the interface to describe what is in
  memory (unifying what is currently done now in struct and
  array)

* Add a new variable to allow the protocol to share shape information

* Add a new variable for sharing stride information

* Add a new mechanism for sharing arrays that must 
  be accessed using pointer indirection. 

* Fix all objects in the core and the standard library to conform
  to the new interface

* Extend the struct module to handle more format specifiers

* Extend the buffer object into a new memory object which places
  a Python veneer around the buffer interface. 

* Add a few functions to make it easy to copy contiguous data
  in and out of object supporting the buffer interface. 

Specification
=============

While the new specification allows for complicated memory sharing.
Simple contiguous buffers of bytes can still be obtained from an
object.  In fact, the new protocol allows a standard mechanism for
doing this even if the original object is not represented as a
contiguous chunk of memory.

The easiest way to obtain a simple contiguous chunk of memory is
to use the provided C-API to obtain a chunk of memory.


Change the PyBufferProcs structure to

::

    typedef struct {
         getbufferproc bf_getbuffer;
         releasebufferproc bf_releasebuffer;
    }


::

    typedef int (*getbufferproc)(PyObject *obj, struct bufferinfo *view, int flags) 

This function returns 0 on success and -1 on failure (and raises an
error). The first variable is the "exporting" object.  The second
argument is the address to a bufferinfo structure.  If view is NULL,
then no information is returned but a lock on the memory is still
obtained.  In this case, the corresponding releasebuffer should also
be called with NULL.

The third argument indicates what kind of buffer the exporter is allowed to return.   It tells the
exporter what elements the bufferinfo structure the consumer is going to make use of.  This
allows the exporter to simplify and/or raise an error if it can't support the operation.  

It also allows the caller to make a request for a simple "view" and
receive it or have an error raised if it's not possible. 

All of the following assume that at least buf, len, and readonly will be
utilized by the caller.

 Py_BUF_SIMPLE
   The returned buffer will only be assumed to be readable (the object
   may or may not have writeable memory).  Only the buf, len, and
   readonly variables may be accessed. The format will be
   assumed to be unsigned bytes.  This is a "stand-alone" flag constant.
   It never needs to be \|'d to the others. 

 Py_BUF_WRITEABLE
   The returned buffer must be writeable.  If it cannot be, then raise an error. 

 Py_BUF_READONLY
   The returned buffer must be readonly and the underlying object should make
   its memory readonly if that is possible.  

 Py_BUF_FORMAT
   The consumer will be using the format string information so make sure that 
   member is filled correctly. 

 Py_BUF_SHAPE
   The consumer can (and might) make use of using the ndims and shape members of the structure
   so make sure they are filled in correctly. 
   
 Py_BUF_STRIDES (implies SHAPE)
   The consumer can (and might) make use of the strides member of the structure (as well
   as ndims and shape)

 Py_BUF_OFFSETS (implies STRIDES)
   The consumer can (and might) make use of the suboffsets member (as well as 
   ndims, shape, and strides)

Thus, the consumer simply wanting an contiguous chunk of bytes from
the object would use Py_BUF_SIMPLE, while a consumer that understands
how to make use of the most complicated cases would use
Py_BUF_OFFSETS.

There is a C-API that simple exporting objects can use to fill-in the
buffer info structure correctly according to the provided flags if a
contiguous chunk of memory is all that can be exported.


The bufferinfo structure is::

  struct bufferinfo {
       void *buf;
       Py_ssize_t len;
       int readonly;
       char *format;
       int ndims;
       Py_ssize_t *shape;
       Py_ssize_t *strides;
       Py_ssize_t *suboffsets;
       void *internal;
  };

Before calling this function, the bufferinfo structure can be filled with
whatever.  Upon return from getbufferproc, the bufferinfo structure is filled in
with relevant information about the buffer.  This same bufferinfo
structure must be passed to bf_releasebuffer (if available) when the
consumer is done with the memory. The caller is responsible for
keeping a reference to obj until releasebuffer is called (i.e. this
call does not alter the reference count of obj). 

The members of the bufferinfo structure are:

buf
    a pointer to the start of the memory for the object

len 
    the total bytes of memory the object uses.  This should be the
    same as the product of the shape array multiplied by the number of
    bytes per item of memory.

readonly
    an integer variable to hold whether or not the memory is
    readonly.  1 means the memory is readonly, zero means the
    memory is writeable.

format 
    a NULL-terminated format-string (following the struct-style syntax
    including extensions) indicating what is in each element of
    memory.  The number of elements is len / itemsize, where itemsize
    is the number of bytes implied by the format.  For standard
    unsigned bytes use a format string of "B".

ndims
    a variable storing the number of dimensions the memory represents.
    Must be >=0. 

shape
    an array of ``Py_ssize_t`` of length ``ndims`` indicating the
    shape of the memory as an N-D array.  Note that ``((*shape)[0] *
    ... * (*shape)[ndims-1])*itemsize = len``.

strides 
    address of a ``Py_ssize_t*`` variable that will be filled with a
    pointer to an array of ``Py_ssize_t`` of length ``*ndims``
    indicating the number of bytes to skip to get to the next element
    in each dimension.  For C-style contiguous arrays (where the
    last-dimension varies the fastest) this must be filled in.  

suboffsets
    address of a ``Py_ssize_t *`` variable that will be filled with a
    pointer to an array of ``Py_ssize_t`` of length ``*ndims``.  If
    these suboffset numbers are >=0, then the value stored along the
    indicated dimension is a pointer and the suboffset value dictates
    how many bytes to add to the pointer after de-referencing.  A
    suboffset value that it 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 must be NULL.

    For clarity, here is a function that returns a pointer to the
    element in an N-D array pointed to by an N-dimesional index when
    there are both 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;
      }

    Notice the suboffset is added "after" the dereferencing occurs.
    Thus slicing in the ith dimension would add to the suboffsets in
    the (i-1)st dimension.  Slicing in the first dimension would change
    the location of the starting pointer directly (i.e. buf would
    be modified).  

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
    should never touch this value. 
    

The exporter is responsible for making sure the memory pointed to by
buf, format, shape, strides, and suboffsets is valid until
releasebuffer is called.  If the exporter wants to be able to change
shape, strides, and/or suboffsets before releasebuffer is called then
it should allocate those arrays when getbuffer is called (pointing to
them in the buffer-info structure provided) and free them when
releasebuffer is called.


The same bufferinfo struct should be used in the release-buffer
interface call. The caller is responsible for the memory of the
bufferinfo structure itself. 

``typedef int (*releasebufferproc)(PyObject *obj, struct bufferinfo *view)``
    Callers of getbufferproc must make sure that this function is
    called when memory previously acquired from the object is no
    longer needed.  The exporter of the interface must make sure that
    any memory pointed to in the bufferinfo structure remains valid
    until releasebuffer is called.

    Both of these routines are optional for a type object

    If the releasebuffer function is not provided then it does not ever
    need to be called. 
    
Exporters will need to define a releasebuffer function if they can
re-allocate their memory, strides, shape, suboffsets, or format
variables which they might share through the struct bufferinfo.
Several mechanisms could be used to keep track of how many getbuffer
calls have been made and shared.  Either a single variable could be
used to keep track of how many "views" have been exported, or a
linked-list of bufferinfo structures filled in could be maintained in
each object.  

All that is specifically required by the exporter, however, is to
ensure that any memory shared through the bufferinfo structure remains
valid until releasebuffer is called on the bufferinfo structure.


New C-API calls are proposed
============================

::

    int PyObject_CheckBuffer(PyObject *obj)

Return 1 if the getbuffer function is available otherwise 0.

::

    int PyObject_GetBuffer(PyObject *obj, struct bufferinfo *view, int flags)

This is a C-API version of the getbuffer function call.  It checks to
make sure object has the required function pointer and issues the
call.  Returns -1 and raises an error on failure and returns 0 on 
success. 

::

    int PyObject_ReleaseBuffer(PyObject *obj, struct bufferinfo *view)

This is a C-API version of the releasebuffer function call.  It checks to
make sure the object has the required function pointer and issues the call.  Returns 0
on success and -1 (with an error raised) on failure. This function always 
succeeds if there is no releasebuffer function for the object. 

::

    PyObject *PyObject_GetMemoryView(PyObject *obj)

Return a memory-view object from an object that defines the buffer interface. 
If make_ro is non-zero then request that the memory is made read-only until 
release buffer is called. 

A memory-view object is an extended buffer object that should replace
the buffer object in Python 3K.  It's C-structure is::

  typedef struct {
      PyObject_HEAD
      PyObject *base;
      struct bufferinfo view;
      int itemsize;
      int flags;
  } PyMemoryViewObject;

This is very similar to the current buffer object except offset has
been removed because ptr can just be modified by offset and a single
offset is not sufficient for the sub-offsets.  Also the hash has been
removed because using the buffer object as a hash even if it is
read-only is rarely useful.

Also, the format, ndims, shape, strides, and suboffsets have been
added. These additions will allow multi-dimensional slicing of the
memory-view object which can be added at some point.  This object
always owns it's own shape, strides, and suboffsets arrays and it's
own format string, but always borrows the memory from the object
pointed to by base.

The itemsize is a convenience and specifies the number of bytes
indicated by the format string if positive.  

This object never reallocates ptr, shape, strides, subboffsets or
format and therefore does not need to keep track of how many views it
has exported.

It exports a view using the base object.  It releases a view by
releasing the view on the base object.  Because, it will never
re-allocate memory, it does not need to keep track of how many it has
exported but simple reference counting will suffice.

::

    int PyObject_SizeFromFormat(char *)

Return the implied itemsize of the data-format area from a struct-style
description.

::

    int PyObject_GetContiguous(PyObject *obj, void **buf, Py_ssize_t *len,
                               int fortran)

Return a contiguous chunk of memory representing the buffer.  If a
copy is made then return 1.  If no copy was needed return 0.  If an
error occurred in probing the buffer interface, then return -1.  The
contiguous chunk of memory is pointed to by ``*buf`` and the length of
that memory is ``*len``.  If the object is multi-dimensional, then if
fortran is 1, the first dimension of the underlying array will vary
the fastest in the buffer.  If fortran is 0, then the last dimension
will vary the fastest (C-style contiguous). If fortran is -1, then it
does not matter and you will get whatever the object decides is more
efficient.

:: 

    int PyObject_CopyToObject(PyObject *obj, void *buf, Py_ssize_t len,
                              int fortran)

Copy ``len`` bytes of data pointed to by the contiguous chunk of
memory pointed to by ``buf`` into the buffer exported by obj.  Return
0 on success and return -1 and raise an error on failure.  If the
object does not have a writeable buffer, then an error is raised.  If
fortran is 1, then if the object is multi-dimensional, then the data
will be copied into the array in Fortran-style (first dimension varies
the fastest).  If fortran is 0, then the data will be copied into the
array in C-style (last dimension varies the fastest).  If fortran is -1, then
it does not matter and the copy will be made in whatever way is more
efficient.

The last two C-API calls allow a standard way of getting data in and
out of Python objects into contiguous memory areas no matter how it is
actually stored.  These calls use the extended buffer interface to perform
their work. 

::

    int PyObject_IsContiguous(struct bufferinfo *view, int fortran);

Return 1 if the memory defined by the view object is C-style (fortran = 0)
or Fortran-style (fortran = 1) contiguous.  Return 0 otherwise.

::

    void PyObject_FillContiguousStrides(int *ndims, Py_ssize_t *shape,
                                        int itemsize, 
                                        Py_ssize_t *strides, int fortran)

Fill the strides array with byte-strides of a contiguous (C-style if
fortran is 0 or Fortran-style if fortran is 1) array of the given
shape with the given number of bytes per element.

::

    int PyObject_FillBufferInfo(struct bufferinfo *view, void *buf, Py_ssize_t len,
                                 int readonly, int infoflags)

Fills in a buffer-info structure correctly for an exporter that can only share
a contiguous chunk of memory of "unsigned bytes" of the given length.  Returns 0 on success
and -1 (with raising an error) on error 


Additions to the struct string-syntax
=====================================

The struct string-syntax is missing some characters to fully
implement data-format descriptions already available elsewhere (in
ctypes and NumPy for example).  The Python 2.5 specification is 
at http://docs.python.org/lib/module-struct.html

Here are the proposed additions:


================  ===========
Character         Description
================  ===========
't'               bit (number before states how many bits)
'?'               platform _Bool type
'g'               long double  
'c'               ucs-1 (latin-1) encoding 
'u'               ucs-2 
'w'               ucs-4 
'O'               pointer to Python Object 
'Z'               complex (whatever the next specifier is)
'&'               specific pointer (prefix before another charater) 
'T{}'             structure (detailed layout inside {}) 
'(k1,k2,...,kn)'  multi-dimensional array of whatever follows 
':name:'          optional name of the preceeding element 
'X{}'             pointer to a function (optional function 
                                         signature inside {})
' \n\t'           ignored (allow better readability) -- this may already be true
================  ===========

The struct module will be changed to understand these as well and
return appropriate Python objects on unpacking.  Un-packing a
long-double will return a decimal object or a ctypes long-double.
Unpacking 'u' or 'w' will return Python unicode.  Unpacking a
multi-dimensional array will return a list of lists.  Un-packing a
pointer will return a ctypes pointer object.  Un-packing a bit will
return a Python Bool.  Spaces in the struct-string syntax will be
ignored.  Unpacking a named-object will return a Python class with
attributes having those names.

Endian-specification ('=','>','<') is also allowed inside the
string so that it can change if needed.  The previously-specified
endian string is in force until changed.  The default endian is '='.

According to the struct-module, a number can preceed a character
code to specify how many of that type there are.  The
(k1,k2,...,kn) extension also allows specifying if the data is
supposed to be viewed as a (C-style contiguous, last-dimension
varies the fastest) multi-dimensional array of a particular format.

Functions should be added to ctypes to create a ctypes object from
a struct description, and add long-double, and ucs-2 to ctypes.

Examples of Data-Format Descriptions
====================================

Here are some examples of C-structures and how they would be
represented using the struct-style syntax:

float
    'f'
complex double
    'Zd'
RGB Pixel data
    'BBB' or 'B:r: B:g: B:b:'
Mixed endian (weird but possible)
    '>i:big: <i:little:'
Nested structure
    ::

        struct {
             int ival;
             struct {
                 unsigned short sval;
                 unsigned char bval;
                 unsigned char cval;
             } sub;
        }
        """i:ival: 
           T{
              H:sval: 
              B:bval: 
              B:cval:
            }:sub:
        """
Nested array
    ::

        struct {
             int ival;
             double data[16*4];
        }
        'i:ival: (16,4)d:data:'


Code to be affected
===================

All objects and modules in Python that export or consume the old
buffer interface will be modified.  Here is a partial list.

* buffer object
* bytes object
* string object
* array module
* struct module
* mmap module
* ctypes module

Anything else using the buffer API.  


Issues and Details
==================

It is intended that this PEP will be back-ported to Python 2.6 by
adding the C-API and the two functions to the existing buffer
protocol.

The proposed locking mechanism relies entirely on the exporter object
to not invalidate any of the memory pointed to by the buffer structure
until a corresponding releasebuffer is called.  If it wants to be able
to change its own shape and/or strides arrays, then it needs to create
memory for these in the bufferinfo structure and copy information
over.

The sharing of strided memory and suboffsets is new and can be seen as
a modification of the multiple-segment interface.  It is motivated by
NumPy and the PIL.  NumPy objects should be able to share their
strided memory with code that understands how to manage strided memory
because strided memory is very common when interfacing with compute
libraries.

Also, with this approach it should be possible to write generic code
that works with both kinds of memory.

Memory management of the format string, the shape array, the strides
array, and the suboffsets array in the bufferinfo structure is always
the responsibility of the exporting object.  The consumer should not
set these pointers to any other memory or try to free them. 

Several ideas were discussed and rejected: 

    Having a "releaser" object whose release-buffer was called.  This
    was deemed unacceptable because it caused the protocol to be
    asymmetric (you called release on something different than you
    "got" the buffer from).  It also complicated the protocol without
    providing a real benefit.

    Passing all the struct variables separately into the function.
    This had the advantage that it allowed one to set NULL to
    variables that were not of interest, but it also made the function
    call more difficult.  The flags variable allows the same
    ability of consumers to be "simple" in how they call the protocol. 

Code
========

The authors of the PEP promise to contribute and maintain the code for
this proposal but will welcome any help.




Examples
=========

Ex. 1
----------

This example shows how an image object that uses contiguous lines might expose its buffer.::

  struct rgba {
      unsigned char r, g, b, a;
  };

  struct ImageObject {
      PyObject_HEAD;
      ...
      struct rgba** lines;
      Py_ssize_t height;
      Py_ssize_t width;
      Py_ssize_t shape_array[2];
      Py_ssize_t stride_array[2];
      Py_ssize_t view_count;
  };

"lines" points to malloced 1-D array of (struct rgba*).  Each pointer
in THAT block points to a seperately malloced array of (struct rgba).

In order to access, say, the red value of the pixel at x=30, y=50, you'd use "lines[50][30].r".

So what does ImageObject's getbuffer do?  Leaving error checking out::

  int Image_getbuffer(PyObject *self, struct bufferinfo *view, int flags) {

      static Py_ssize_t suboffsets[2] = { -1, 0 };

      view->buf = self->lines;
      view->len = self->height*self->width;
      view->readonly = 0;
      view->ndims = 2;
      self->shape_array[0] = height;
      self->shape_array[1] = width;
      view->shape = &self->shape_array;
      self->stride_array[0] = sizeof(struct rgba*);  
      self->stride_array[1] = sizeof(struct rgba);
      view->strides = &self->stride_array;
      view->suboffsets = suboffsets;

      self->view_count ++;

      return 0;
  } 


  int Image_releasebuffer(PyObject *self, struct bufferinfo *view) {
      self->view_count--;
      return 0;
  }


Ex. 2
-----------

This example shows how an object that wants to expose a contiguous
chunk of memory (which will never be re-allocated while the object is
alive) would do that.::

  int myobject_getbuffer(PyObject *self, struct bufferinfo *view, int flags) {

    void *buf;
    Py_ssize_t len;
    int readonly=0;
        
    buf = /* Point to buffer */
    len = /* Set to size of buffer */
    readonly = /* Set to 1 if readonly */

    return PyObject_FillBufferInfo(view, buf, len, readonly, flags);    
  }

  /* No releasebuffer is necessary because the memory will never 
  be re-allocated so the locking mechanism is not needed
  */

Ex.  3
-----------

A consumer that wants to only get a simple contiguous chunk of bytes
from a Python object, obj would do the following::


  struct bufferinfo view;
  int ret;
      
  if (PyObject_GetBuffer(obj, &view, Py_BUF_SIMPLE) < 0) {
       /* error return */
  }

  /* Now, view.buf is the pointer to memory
          view.len is the length
          view.readonly is whether or not the memory is read-only.
   */
  

  /* After using the information and you don't need it anymore */
  
  if (PyObject_ReleaseBuffer(obj, &view) < 0) {
          /* error return */
  }
  



Copyright
=========

This PEP is placed in the public domain