使用一个遵循buffer protocol的对象就可以和numpy交互了.

这个buffer_protocol要有哪些东西呢? 要有如下接口:

struct buffer_info {
void *ptr;
ssize_t itemsize;
std::string format;
ssize_t ndim;
std::vector shape;
std::vector strides;
};

其实就是一个指向数组的指针+各个维度的信息就可以了. 然后我们就可以用指针+偏移来访问数字中的任意位置上的数字了.

下面是一个可以跑的例子:

1 #include
2 #include
3
4 namespace py = pybind11;
5
6 py::array_t add_arrays(py::array_t input1, py::array_t input2) {
7 py::buffer_info buf1 = input1.request(), buf2 = input2.request();
8
9 if (buf1.ndim != 1 || buf2.ndim != 1)
10 throw std::runtime_error("Number of dimensions must be one");
11
12 if (buf1.size != buf2.size)
13 throw std::runtime_error("Input shapes must match");
14
15 /* No pointer is passed, so NumPy will allocate the buffer */
16 auto result = py::array_t(buf1.size);
17
18 py::buffer_info buf3 = result.request();
19
20 double *ptr1 = (double *) buf1.ptr,
21 *ptr2 = (double *) buf2.ptr,
22 *ptr3 = (double *) buf3.ptr;
23
24 for (size_t idx = 0; idx < buf1.shape[0]; idx++)
25 ptr3[idx] = ptr1[idx] + ptr2[idx];
26
27 return result;
28 }
29
30 PYBIND11_MODULE(test, m) {
31 m.def("add_arrays", &add_arrays, "Add two NumPy arrays");
32 }

array_t里的buf就是一个兼容的接口.

buf中可以得到指针和对应数字的维度信息.

为了方便我们甚至可以使用Eigen当作我们兼容numpy的接口:

1 #include
2 #include
3
4 #include
5
6 // N.B. this would equally work with Eigen-types that are not predefined. For example replacing
7 // all occurrences of "Eigen::MatrixXd" with "MatD", with the following definition:
8 //
9 // typedef Eigen::Matrix MatD;
10
11 Eigen::MatrixXd inv(const Eigen::MatrixXd &xs)
12 {
13 return xs.inverse();
14 }
15
16 double det(const Eigen::MatrixXd &xs)
17 {
18 return xs.determinant();
19 }
20
21 namespace py = pybind11;
22
23 PYBIND11_MODULE(example,m)
24 {
25 m.doc() = "pybind11 example plugin";
26
27 m.def("inv", &inv);
28
29 m.def("det", &det);
30 }

更多参考:

https://pybind11.readthedocs.io/en/stable/advanced/pycpp/numpy.html

https://github.com/tdegeus/pybind11_examples