Changes: * added the "flags" variable to allow simpler calling for getbuffer. * added some explanation of ideas that were discussed and abandoned. * added examples for simple use cases. * added more C-API calls to allow easier usage. Thanks for all feedback. -Travis 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