1Eigen
2#####
3
4`Eigen <http://eigen.tuxfamily.org>`_ is C++ header-based library for dense and
5sparse linear algebra. Due to its popularity and widespread adoption, pybind11
6provides transparent conversion and limited mapping support between Eigen and
7Scientific Python linear algebra data types.
8
9To enable the built-in Eigen support you must include the optional header file
10:file:`pybind11/eigen.h`.
11
12Pass-by-value
13=============
14
15When binding a function with ordinary Eigen dense object arguments (for
16example, ``Eigen::MatrixXd``), pybind11 will accept any input value that is
17already (or convertible to) a ``numpy.ndarray`` with dimensions compatible with
18the Eigen type, copy its values into a temporary Eigen variable of the
19appropriate type, then call the function with this temporary variable.
20
21Sparse matrices are similarly copied to or from
22``scipy.sparse.csr_matrix``/``scipy.sparse.csc_matrix`` objects.
23
24Pass-by-reference
25=================
26
27One major limitation of the above is that every data conversion implicitly
28involves a copy, which can be both expensive (for large matrices) and disallows
29binding functions that change their (Matrix) arguments.  Pybind11 allows you to
30work around this by using Eigen's ``Eigen::Ref<MatrixType>`` class much as you
31would when writing a function taking a generic type in Eigen itself (subject to
32some limitations discussed below).
33
34When calling a bound function accepting a ``Eigen::Ref<const MatrixType>``
35type, pybind11 will attempt to avoid copying by using an ``Eigen::Map`` object
36that maps into the source ``numpy.ndarray`` data: this requires both that the
37data types are the same (e.g. ``dtype='float64'`` and ``MatrixType::Scalar`` is
38``double``); and that the storage is layout compatible.  The latter limitation
39is discussed in detail in the section below, and requires careful
40consideration: by default, numpy matrices and Eigen matrices are *not* storage
41compatible.
42
43If the numpy matrix cannot be used as is (either because its types differ, e.g.
44passing an array of integers to an Eigen parameter requiring doubles, or
45because the storage is incompatible), pybind11 makes a temporary copy and
46passes the copy instead.
47
48When a bound function parameter is instead ``Eigen::Ref<MatrixType>`` (note the
49lack of ``const``), pybind11 will only allow the function to be called if it
50can be mapped *and* if the numpy array is writeable (that is
51``a.flags.writeable`` is true).  Any access (including modification) made to
52the passed variable will be transparently carried out directly on the
53``numpy.ndarray``.
54
55This means you can can write code such as the following and have it work as
56expected:
57
58.. code-block:: cpp
59
60    void scale_by_2(Eigen::Ref<Eigen::VectorXd> v) {
61        v *= 2;
62    }
63
64Note, however, that you will likely run into limitations due to numpy and
65Eigen's difference default storage order for data; see the below section on
66:ref:`storage_orders` for details on how to bind code that won't run into such
67limitations.
68
69.. note::
70
71    Passing by reference is not supported for sparse types.
72
73Returning values to Python
74==========================
75
76When returning an ordinary dense Eigen matrix type to numpy (e.g.
77``Eigen::MatrixXd`` or ``Eigen::RowVectorXf``) pybind11 keeps the matrix and
78returns a numpy array that directly references the Eigen matrix: no copy of the
79data is performed.  The numpy array will have ``array.flags.owndata`` set to
80``False`` to indicate that it does not own the data, and the lifetime of the
81stored Eigen matrix will be tied to the returned ``array``.
82
83If you bind a function with a non-reference, ``const`` return type (e.g.
84``const Eigen::MatrixXd``), the same thing happens except that pybind11 also
85sets the numpy array's ``writeable`` flag to false.
86
87If you return an lvalue reference or pointer, the usual pybind11 rules apply,
88as dictated by the binding function's return value policy (see the
89documentation on :ref:`return_value_policies` for full details).  That means,
90without an explicit return value policy, lvalue references will be copied and
91pointers will be managed by pybind11.  In order to avoid copying, you should
92explicitly specify an appropriate return value policy, as in the following
93example:
94
95.. code-block:: cpp
96
97    class MyClass {
98        Eigen::MatrixXd big_mat = Eigen::MatrixXd::Zero(10000, 10000);
99    public:
100        Eigen::MatrixXd &getMatrix() { return big_mat; }
101        const Eigen::MatrixXd &viewMatrix() { return big_mat; }
102    };
103
104    // Later, in binding code:
105    py::class_<MyClass>(m, "MyClass")
106        .def(py::init<>())
107        .def("copy_matrix", &MyClass::getMatrix) // Makes a copy!
108        .def("get_matrix", &MyClass::getMatrix, py::return_value_policy::reference_internal)
109        .def("view_matrix", &MyClass::viewMatrix, py::return_value_policy::reference_internal)
110        ;
111
112.. code-block:: python
113
114    a = MyClass()
115    m = a.get_matrix()   # flags.writeable = True,  flags.owndata = False
116    v = a.view_matrix()  # flags.writeable = False, flags.owndata = False
117    c = a.copy_matrix()  # flags.writeable = True,  flags.owndata = True
118    # m[5,6] and v[5,6] refer to the same element, c[5,6] does not.
119
120Note in this example that ``py::return_value_policy::reference_internal`` is
121used to tie the life of the MyClass object to the life of the returned arrays.
122
123You may also return an ``Eigen::Ref``, ``Eigen::Map`` or other map-like Eigen
124object (for example, the return value of ``matrix.block()`` and related
125methods) that map into a dense Eigen type.  When doing so, the default
126behaviour of pybind11 is to simply reference the returned data: you must take
127care to ensure that this data remains valid!  You may ask pybind11 to
128explicitly *copy* such a return value by using the
129``py::return_value_policy::copy`` policy when binding the function.  You may
130also use ``py::return_value_policy::reference_internal`` or a
131``py::keep_alive`` to ensure the data stays valid as long as the returned numpy
132array does.
133
134When returning such a reference of map, pybind11 additionally respects the
135readonly-status of the returned value, marking the numpy array as non-writeable
136if the reference or map was itself read-only.
137
138.. note::
139
140    Sparse types are always copied when returned.
141
142.. _storage_orders:
143
144Storage orders
145==============
146
147Passing arguments via ``Eigen::Ref`` has some limitations that you must be
148aware of in order to effectively pass matrices by reference.  First and
149foremost is that the default ``Eigen::Ref<MatrixType>`` class requires
150contiguous storage along columns (for column-major types, the default in Eigen)
151or rows if ``MatrixType`` is specifically an ``Eigen::RowMajor`` storage type.
152The former, Eigen's default, is incompatible with ``numpy``'s default row-major
153storage, and so you will not be able to pass numpy arrays to Eigen by reference
154without making one of two changes.
155
156(Note that this does not apply to vectors (or column or row matrices): for such
157types the "row-major" and "column-major" distinction is meaningless).
158
159The first approach is to change the use of ``Eigen::Ref<MatrixType>`` to the
160more general ``Eigen::Ref<MatrixType, 0, Eigen::Stride<Eigen::Dynamic,
161Eigen::Dynamic>>`` (or similar type with a fully dynamic stride type in the
162third template argument).  Since this is a rather cumbersome type, pybind11
163provides a ``py::EigenDRef<MatrixType>`` type alias for your convenience (along
164with EigenDMap for the equivalent Map, and EigenDStride for just the stride
165type).
166
167This type allows Eigen to map into any arbitrary storage order.  This is not
168the default in Eigen for performance reasons: contiguous storage allows
169vectorization that cannot be done when storage is not known to be contiguous at
170compile time.  The default ``Eigen::Ref`` stride type allows non-contiguous
171storage along the outer dimension (that is, the rows of a column-major matrix
172or columns of a row-major matrix), but not along the inner dimension.
173
174This type, however, has the added benefit of also being able to map numpy array
175slices.  For example, the following (contrived) example uses Eigen with a numpy
176slice to multiply by 2 all coefficients that are both on even rows (0, 2, 4,
177...) and in columns 2, 5, or 8:
178
179.. code-block:: cpp
180
181    m.def("scale", [](py::EigenDRef<Eigen::MatrixXd> m, double c) { m *= c; });
182
183.. code-block:: python
184
185    # a = np.array(...)
186    scale_by_2(myarray[0::2, 2:9:3])
187
188The second approach to avoid copying is more intrusive: rearranging the
189underlying data types to not run into the non-contiguous storage problem in the
190first place.  In particular, that means using matrices with ``Eigen::RowMajor``
191storage, where appropriate, such as:
192
193.. code-block:: cpp
194
195    using RowMatrixXd = Eigen::Matrix<double, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
196    // Use RowMatrixXd instead of MatrixXd
197
198Now bound functions accepting ``Eigen::Ref<RowMatrixXd>`` arguments will be
199callable with numpy's (default) arrays without involving a copying.
200
201You can, alternatively, change the storage order that numpy arrays use by
202adding the ``order='F'`` option when creating an array:
203
204.. code-block:: python
205
206    myarray = np.array(source, order='F')
207
208Such an object will be passable to a bound function accepting an
209``Eigen::Ref<MatrixXd>`` (or similar column-major Eigen type).
210
211One major caveat with this approach, however, is that it is not entirely as
212easy as simply flipping all Eigen or numpy usage from one to the other: some
213operations may alter the storage order of a numpy array.  For example, ``a2 =
214array.transpose()`` results in ``a2`` being a view of ``array`` that references
215the same data, but in the opposite storage order!
216
217While this approach allows fully optimized vectorized calculations in Eigen, it
218cannot be used with array slices, unlike the first approach.
219
220When *returning* a matrix to Python (either a regular matrix, a reference via
221``Eigen::Ref<>``, or a map/block into a matrix), no special storage
222consideration is required: the created numpy array will have the required
223stride that allows numpy to properly interpret the array, whatever its storage
224order.
225
226Failing rather than copying
227===========================
228
229The default behaviour when binding ``Eigen::Ref<const MatrixType>`` Eigen
230references is to copy matrix values when passed a numpy array that does not
231conform to the element type of ``MatrixType`` or does not have a compatible
232stride layout.  If you want to explicitly avoid copying in such a case, you
233should bind arguments using the ``py::arg().noconvert()`` annotation (as
234described in the :ref:`nonconverting_arguments` documentation).
235
236The following example shows an example of arguments that don't allow data
237copying to take place:
238
239.. code-block:: cpp
240
241    // The method and function to be bound:
242    class MyClass {
243        // ...
244        double some_method(const Eigen::Ref<const MatrixXd> &matrix) { /* ... */ }
245    };
246    float some_function(const Eigen::Ref<const MatrixXf> &big,
247                        const Eigen::Ref<const MatrixXf> &small) {
248        // ...
249    }
250
251    // The associated binding code:
252    using namespace pybind11::literals; // for "arg"_a
253    py::class_<MyClass>(m, "MyClass")
254        // ... other class definitions
255        .def("some_method", &MyClass::some_method, py::arg().noconvert());
256
257    m.def("some_function", &some_function,
258        "big"_a.noconvert(), // <- Don't allow copying for this arg
259        "small"_a            // <- This one can be copied if needed
260    );
261
262With the above binding code, attempting to call the the ``some_method(m)``
263method on a ``MyClass`` object, or attempting to call ``some_function(m, m2)``
264will raise a ``RuntimeError`` rather than making a temporary copy of the array.
265It will, however, allow the ``m2`` argument to be copied into a temporary if
266necessary.
267
268Note that explicitly specifying ``.noconvert()`` is not required for *mutable*
269Eigen references (e.g. ``Eigen::Ref<MatrixXd>`` without ``const`` on the
270``MatrixXd``): mutable references will never be called with a temporary copy.
271
272Vectors versus column/row matrices
273==================================
274
275Eigen and numpy have fundamentally different notions of a vector.  In Eigen, a
276vector is simply a matrix with the number of columns or rows set to 1 at
277compile time (for a column vector or row vector, respectively).  Numpy, in
278contrast, has comparable 2-dimensional 1xN and Nx1 arrays, but *also* has
2791-dimensional arrays of size N.
280
281When passing a 2-dimensional 1xN or Nx1 array to Eigen, the Eigen type must
282have matching dimensions: That is, you cannot pass a 2-dimensional Nx1 numpy
283array to an Eigen value expecting a row vector, or a 1xN numpy array as a
284column vector argument.
285
286On the other hand, pybind11 allows you to pass 1-dimensional arrays of length N
287as Eigen parameters.  If the Eigen type can hold a column vector of length N it
288will be passed as such a column vector.  If not, but the Eigen type constraints
289will accept a row vector, it will be passed as a row vector.  (The column
290vector takes precedence when both are supported, for example, when passing a
2911D numpy array to a MatrixXd argument).  Note that the type need not be
292explicitly a vector: it is permitted to pass a 1D numpy array of size 5 to an
293Eigen ``Matrix<double, Dynamic, 5>``: you would end up with a 1x5 Eigen matrix.
294Passing the same to an ``Eigen::MatrixXd`` would result in a 5x1 Eigen matrix.
295
296When returning an Eigen vector to numpy, the conversion is ambiguous: a row
297vector of length 4 could be returned as either a 1D array of length 4, or as a
2982D array of size 1x4.  When encountering such a situation, pybind11 compromises
299by considering the returned Eigen type: if it is a compile-time vector--that
300is, the type has either the number of rows or columns set to 1 at compile
301time--pybind11 converts to a 1D numpy array when returning the value.  For
302instances that are a vector only at run-time (e.g. ``MatrixXd``,
303``Matrix<float, Dynamic, 4>``), pybind11 returns the vector as a 2D array to
304numpy.  If this isn't want you want, you can use ``array.reshape(...)`` to get
305a view of the same data in the desired dimensions.
306
307.. seealso::
308
309    The file :file:`tests/test_eigen.cpp` contains a complete example that
310    shows how to pass Eigen sparse and dense data types in more detail.
311