C and C++ API

This section includes information for using the pure C or C++ API for manipulating bob.bob.blitz objects in compiled code.

C API

The C API of bob.blitz allows users to leverage from automatic converters between numpy.ndarray and bob.blitz.array within their own python extensions. To use the C API, clients should first, include the header file <bob.blitz/capi.h> on their compilation units and then, make sure to call once import_bob_blitz() at their module instantiation, as explained at the Python manual.

Here is a dummy C example showing how to include the header and where to call the import function:

#include <bob.blitz/capi.h>

PyMODINIT_FUNC initclient(void) {

  PyObject* m Py_InitModule("client", ClientMethods);

  if (!m) return;

  // imports dependencies
  if (import_bob_blitz() < 0) {
    PyErr_Print();
    PyErr_SetString(PyExc_ImportError, "cannot import module");
    return 0;
  }

  return m;
}

Array Structure

PyBlitzArrayObject

The basic array structure represents a bob.blitz.array instance from the C-side of the interpreter. You should avoid direct access to the structure components (it is presented just as an overview on the functionality). Instead, use the accessor methods described below.

typedef struct {
  PyObject_HEAD
  void* bzarr;
  void* data;
  int type_num;
  Py_ssize_t ndim;
  Py_ssize_t shape[BLITZ_ARRAY_MAXDIMS];
  Py_ssize_t stride[BLITZ_ARRAY_MAXDIMS];
  int writeable;
  PyObject* base;

} PyBlitzArrayObject;
BLITZ_ARRAY_MAXDIMS

The maximum number of dimensions supported by the current bob.blitz.array implementation.

void* bzarr

This is a pointer that points to the allocated blitz::Array structure. This pointer is cast to the proper type and number of dimensions when operations on the data are requested.

void* data

A pointer to the data entry in the blitz::Array<>. This is equivalent to the operation blitz::Array<>::data().

int type_num

The numpy type number that is compatible with the elements of this array. It is a C representation of the C++ template parameter T. Only some types are current supported, namely:

C/C++ type Numpy Enum Notes
bool NPY_BOOL  
uint8_t NPY_UINT8  
uint16_t NPY_UINT16  
uint32_t NPY_UINT32  
uint64_t NPY_UINT64  
int8_t NPY_INT8  
int16_t NPY_INT16  
int32_t NPY_INT32  
int64_t NPY_INT64  
float NPY_FLOAT32  
double NPY_FLOAT64  
long double NPY_FLOAT128 Plat. Dependent
std::complex<float> NPY_COMPLEX64  
std::complex<double> NPY_COMPLEX128  
std::complex<long double> NPY_COMPLEX256 Plat. Dependent
Py_ssize_t ndim

The rank of the blitz::Array<> allocated on bzarr.

Py_ssize_t shape[BLITZ_ARRAY_MAXDIMS]

The shape of the blitz::Array<> allocated on bzarr, in number of elements in each dimension.

Py_ssize_t stride[BLITZ_ARRAY_MAXDIMS]

The strides of the blitz::Array<> allocated on bzarr, in number of bytes to jump to read the next element in each dimensions.

int writeable

Assumes the value of 1 (true), if the data is read-write. 0 is set otherwise.

PyObject* base

If the memory pointed by the currently allocated blitz::Array<> belongs to another Python object, the object is Py_INCREF()‘ed and a pointer is kept on this structure member.

Basic Properties and Checking

int PyBlitzArray_Check(PyObject* o)

Checks if the input object o is a PyBlitzArrayObject. Returns 1 if it is, and 0 otherwise.

int PyBlitzArray_CheckNumpyBase(PyArrayObject* o)

Checks if the input object o is a PyArrayObject (i.e. a numpy.ndarray), if so, checks if the base of the object is set and that it corresponds to the current PyArrayObject shape and stride settings. If so, returns 1. It returns 0 otherwise.

int PyBlitzArray_TYPE(PyBlitzArrayObject* o)

Returns integral type number (as defined by the Numpy C-API) of elements in this blitz::Array<>. This is the formal method to query for o->type_num.

PyArray_Descr* PyBlitzArray_PyDTYPE(PyBlitzArrayObject* o)

Returns a new reference to a numpy C-API PyArray_Descr* equivalent to the internal type element T.

Py_ssize_t PyBlitzArray_NDIM(PyBlitzArrayObject* o)

Returns the number of dimensions in a given bob.blitz.array. This is the formal way to check for o->ndim.

Py_ssize_t* PyBlitzArray_SHAPE(PyBlitzArrayObject* o)

Returns the C-stype shape for this blitz::Array<>. This is the formal method to query for o->shape. The shape represents the number of elements in each dimension of the array.

PyObject* PyBlitzArray_PySHAPE(PyBlitzArrayObject* o)

Returns a new reference to a Python tuple holding a copy of the shape for the given array. The shape represents the number of elements in each dimension of the array.

Py_ssize_t* PyBlitzArray_STRIDE(PyBlitzArrayObject* o)

Returns the C-stype stride for this blitz::Array<>. This is the formal method to query for o->stride. The strides in this object are represented in number of bytes and not in number of elements considering its type_num. This is compatible with the numpy.ndarray strategy.

PyObject* PyBlitzArray_PySTRIDE(PyBlitzArrayObject* o)

Returns a new reference to a Python tuple holding a copy of the strides for the given array. The strides in this object are represented in number of bytes and not in number of elements considering its type_num. This is compatible with the numpy.ndarray strategy.

int PyBlitzArray_WRITEABLE(PyBlitzArrayObject* o)

Returns 1 if the object is writeable, 0 otherwise. This is the formal way to check for o->writeable.

PyObject* PyBlitzArray_PyWRITEABLE(PyBlitzArrayObject* o)

Returns True if the object is writeable, False otherwise.

PyObject* PyBlitzArray_BASE(PyBlitzArrayObject* o)

Returns a borrowed reference to the base of this object. The return value of this function may be NULL.

PyObject* PyBlitzArray_PyBASE(PyBlitzArrayObject* o)

Returns a new reference to the base of this object. If the internal o->base is NULL, then returns Py_None. Use this when interfacing with the Python interpreter.

Indexing

PyObject* PyBlitzArray_GetItem(PyBlitzArrayObject* o, Py_ssize_t* pos)

Returns, as a PyObject, an item from the array. This will be a copy of the internal item. If you set it, it won’t set the original array. o should be the PyBlitzArrayObject to be queried. pos should be a C-style array indicating the precise position to fetch. It is considered to have the same number of entries as the current array shape.

int PyBlitzArray_SetItem(PyBlitzArrayObject* o, Py_ssize_t* pos, PyObject* value)

Sets an given position on the array using any Python or numpy scalar. o should be the PyBlitzArrayObject to be set. pos should be a C-style array indicating the precise position to set and value, the Python or numpy scalar to set the value to.

Construction and Destruction

PyObject* PyBlitzArray_New(PyTypeObject* type, PyObject *args, PyObject* kwds)

Allocates memory and pre-initializes a PyBlitzArrayObject* object. This is the base allocator - seldomly used in user code.

void PyBlitzArray_Delete(PyBlitzArrayObject* o)

Completely deletes a PyBlitzArrayObject* and associated memory areas. This is the base deallocator - seldomly used in user code.

PyObject* PyBlitzArray_SimpleNew(int typenum, Py_ssize_t ndim, Py_ssize_t* shape)

Allocates a new bob.blitz with a given (supported) type and return it as a python object. typenum should be set to the numpy type number of the array type (e.g. NPY_FLOAT64). ndim should be set to the total number of dimensions the array should have. shape should be set to the array shape.

PyObject* PyBlitzArray_SimpleNewFromData(int type_num, Py_ssize_t ndim, Py_ssize_t* shape, Py_ssize_t* stride, void* data, int writeable)

Allocates a new bob.blitz.array with a given (supported) type and return it as a python object. typenum should be set to the numpy type number of the array type (e.g. NPY_FLOAT64). ndim should be set to the total number of dimensions the array should have. shape should be set to the array shape. stride should be set to the array stride in the numpy style (in number of bits). data should be a pointer to the begin of the data area. writeable indicates if the resulting array should be writeble (set it to 1), or read-only (set it to 0).

The memory area pointed by data is stolen from the user, which should not delete it anymore.

int PyBlitzArray_SimpleInit(PyBlitzArrayObject* arr, int typenum, Py_ssize_t ndim, Py_ssize_t* shape)

Initializes the given PyBlitzArrayObject* with a new blitz::Array of the given typenum, dimensionality and shape. See PyBlitzArray_SimpleNew() for details on the parameters. This function does not check if the memory is already initialized. It returns 0 on success and -1 on failure.

To/From Numpy Converters

PyObject* PyBlitzArray_AsNumpyArray(PyBlitzArrayObject* o, PyArrayDescr* dtype)

Creates a shallow copy of the given bob.blitz.array as a numpy.ndarray. The argument dtype may be given, in which case if the current data type is not the same, then forces the creation of a copy conforming to the require data type, if possible. You may set dtype to NULL in case you don’t mind the resulting data type.

Returns a new reference.

PyObject* PyBlitzArray_FromNumpyArray(PyObject* o)

Creates a new bob.blitz.array from a numpy.ndarray object in a shallow manner.

Returns a new reference.

PyObject* PyBlitzArray_NUMPY_WRAP(PyObject* o)

Creates a shallow copy of the given bob.blitz.array as a numpy.ndarray. This function is a shortcut replacement for PyBlitzArray_AsNumpyArray(). It can be used when the input object o is surely of type PyBlitzArrayObject. It creates a wrapper PyArrayObject that contains, as base, a stolen reference to the input object o.

It is designed like this so you can easily wrap freshly created objects of type PyBlitzArrayObject as PyArrayObject. It assumes the input object is of the right type and wrap-able as a numpy.ndarray. It does not check the object base variable, assuming it is set to NULL (what is the case to freshly created PyBlitzArrayObject‘s). If you are not sure about the nature of o, use the slower but safer PyBlitzArray_AsNumpyArray().

Note

The value of o can be NULL, in which case this function returns immediately, allowing you to propagate exceptions.

Converter Functions for PyArg_Parse* family

int PyBlitzArray_Converter(PyObject* o, PyBlitzArrayObject** a)

This function is meant to be used with PyArg_ParseTupleAndKeywords() family of functions in the Python C-API. It converts an arbitrary input object into a PyBlitzArrayObject that can be used as input into another function.

You should use this converter when you don’t need to write-back into the input array. As any other standard Python converter, it returns a new reference to a PyBlitzArrayObject.

It works efficiently if the input array is already a PyBlitzArrayObject or if it is a PyArrayObject (i.e., a :py:class:numpy.ndarray), with a matching base which is a PyBlitzArrayObject. Otherwise, it creates a new PyBlitzArrayObject by first creating a PyArrayObject and then shallow wrapping it with a PyBlitzArrayObject.

Returns 0 if an error is detected, 1 on success.

int PyBlitzArray_BehavedConverter(PyObject* o, PyBlitzArrayObject** a)

This function operates like PyBlitzArray_Converter(), excepts it guarantees that the returned (underlying) blitz::Array<> object is wrapped around a well-behaved numpy.ndarray object (i.e. contiguous, memory-aligned, C-style).

In the event the input object is already a PyBlitzArrayObject, then a new reference to it is returned. It does not check, in this particular case, that the input object is well-behaved.

Returns 0 if an error is detected, 1 on success.

int PyBlitzArray_OutputConverter(PyObject* o, PyBlitzArrayObject** a)

This function is meant to be used with PyArg_ParseTupleAndKeywords() family of functions in the Python C-API. It converts an arbitrary input object into a PyBlitzArrayObject that can be used as input/output or output into another function.

You should use this converter when you need to write-back into the input array. The input type should be promptly convertible to a numpy.ndarray as with PyArray_OutputConverter(). As any other standard Python converter, it returns a new reference to a PyBlitzArrayObject*.

Returns 0 if an error is detected, 1 on success.

int PyBlitzArray_IndexConverter(PyObject* o, PyBlitzArrayObject** shape)

Converts any compatible sequence into a C-array containing the shape information. The shape information and number of dimensions is stored on the previously allocated PyBlitzArrayObject* you should provide. This method is supposed to be used with PyArg_ParseTupleAndKeywords and derivatives.

Parameters are:

o
The input object to be converted into a C-shape
shape
A preallocated (double) address for storing the shape value, on successful conversion

Returns 0 if an error is detected, 1 on success.

int PyBlitzArray_TypenumConverter(PyObject* o, int* type_num)

Converts any compatible value into a Numpy integer type number. This method is supposed to be used with PyArg_ParseTupleAndKeywords and derivatives.

Parameters are:

o
The input object to be converted into a type number
type_num
An address for storing the type number on successful conversion.

Returns 0 if an error is detected, 1 on success.

Other Utilities

const char* PyBlitzArray_TypenumAsString(int typenum)

Converts from numpy type_num to a string representation

PyObject* PyBlitzArray_Cast(PyBlitzArrayObject* o, int typenum)

Casts a given Blitz++ Array into another data type, returns a new reference. If the underlying Blitz++ Array is already of the given type, then just increments the reference counter and returns.

If a problem is detected (e.g. the impossibility to cast to the desired type), then this function will return NULL. You must check the return value and then take the appropriate action after calling this function.

Note

Casting, as operated by this function, may incur in precision loss between the originating type and the destination type.

C++ API

The C++ API consists mostly of templated methods for manipulating the C++ type blitz::Array<> so as to convert PyObject*‘s from and to objects of that type. To use the C++ API you must include the header file <bob.blitz/cppapi.h> and import_bob_blitz() on your module, as explained on the C-API section of this document.

Basic Properties and Checking

int PyBlitzArrayCxx_IsBehaved<T, N>(blitz::Array<T, N>& a)

Tells if a blitz::Array<> is memory contiguous and C-style.

Construction and Destruction

PyObject* PyBlitzArrayCxx_NewFromConstArray<T, N>(const blitz::Array<T, N>& a)

Builds a new read-only PyBlitzArrayObject from an existing Blitz++ array, without copying the data. Returns a new reference.

PyObject* PyBlitzArrayCxx_NewFromArray<T, N>(blitz::Array<T, N>& a)

Builds a new writeable PyBlitzArrayObject from an existing Blitz++ array, without copying the data. Returns a new reference.

PyObject* PyBlitzArrayCxx_AsConstNumpy<T, N>(const blitz::Array<T, N>& a)

Builds a new read-only numpy.ndarray object from the given Blitz++ array without copying the data. Returns a new reference.

In fact, it actually calls two of the above mentioned functions PyBlitzArrayCxx_NewFromConstArray() and PyBlitzArray_NUMPY_WRAP():

PyBlitzArray_NUMPY_WRAP(PyBlitzArrayCxx_NewFromConstArray(a));
PyObject* PyBlitzArrayCxx_AsNumpy<T, N>(blitz::Array<T, N>& a)

Builds a new writeable numpy.ndarray object from the given Blitz++ array without copying the data. Returns a new reference.

In fact, it actually calls two of the above mentioned functions PyBlitzArrayCxx_NewFromArray() and PyBlitzArray_NUMPY_WRAP():

PyBlitzArray_NUMPY_WRAP(PyBlitzArrayCxx_NewFromArray(a));

Other Utilities

blitz::Array<T, N>* PyBlitzArrayCxx_AsBlitz(PyBlitzArrayObject* o)

Casts a PyBlitzArrayObject to a specific blitz::Array<> type. Notice this is a brute-force cast. You are responsible for checking if that it is correct.

blitz::Array<T, N>* PyBlitzArrayCxx_AsBlitz(PyBlitzArrayObject* o, const char* name)

Casts a PyBlitzArrayObject to a specific blitz::Array<> type after checking that the dimensions and the data type of the underlying PyBlitzArrayObject fits to the template parameters. If the check fails, an Python error is set, using the given name parameter as the name of the object that was passed to the python function, and NULL is returned. Hence, please check the result of this function for NULL:

// ...

PyBlitzArrayObject* data;
if (!PyArg_ParseTupleAndKeywords(..., data, ...))
  return NULL;

// use safe reference counting
auto _ = make_safe(data);

// get the blitz array; returns NULL on failure
blitz::Array<double,2>* array = PyBlitzArrayCxx_AsBlitz<double,2>(data, "data");

// check for NULL
if (!array)
  // The error message has already been set, so we can simply return NULL
  return NULL;

// ...

Note

If you need to check for several data types and/or dimensions, use the first version of this function and perform the checks by hand.

Note

This version of the function might be slightly slower than the first version.

int PyBlitzArrayCxx_CToTypenum<T>()

Converts from C/C++ type to ndarray type_num.

We cover only simple conversions (i.e., standard integers, floats and complex numbers only). If the input type is not convertible, an exception is set on the Python error stack. You must check PyErr_Occurred() after a call to this function to make sure things are OK and act accordingly. For example:

int typenum = PyBlitzArrayCxx_CToTypenum<my_weird_type>(obj);
if (PyErr_Occurred()) return 0; ///< propagate exception
T PyBlitzArrayCxx_AsCScalar<T>(PyObject* o)

Extraction API for simple types.

We cover only simple conversions (i.e., standard integers, floats and complex numbers only). If the input object is not convertible to the given type, an exception is set on the Python error stack. You must check PyErr_Occurred() after a call to this function to make sure things are OK and act accordingly. For example:

auto z = extract<uint8_t>(obj);
if (PyErr_Occurred()) return 0; ///< propagate exception
PyBlitzArrayCxx_FromCScalar<T>(T v)

Converts simple C types into numpy scalars

We cover only simple conversions (i.e., standard integers, floats and complex numbers only). If the input object is not convertible to the given type, an exception is set on the Python error stack and 0 (NULL) is returned.