如何从c++原始指针获得pybind11中的零拷贝视图？

Question

请考虑一下这个小例子。

mcve.cpp

#include <pybind11/pybind11.h>
#include <pybind11/operators.h>
#include <pybind11/numpy.h>

namespace py = pybind11;

class mat4 {
  public:
    mat4(float v = 1.0f) {
        data[0] = v; data[4] = 0; data[8] = 0; data[12] = 0;
        data[1] = 0; data[5] = v; data[9] = 0; data[13] = 0;
        data[2] = 0; data[6] = 0; data[10] = v; data[14] = 0;
        data[3] = 0; data[7] = 0; data[11] = 0; data[15] = v;
    }

    mat4(
        float m00, float m01, float m02, float m03,
        float m10, float m11, float m12, float m13,
        float m20, float m21, float m22, float m23,
        float m30, float m31, float m32, float m33
    ) {
        data[0] = m00; data[4] = m01; data[8] = m02; data[12] = m03;
        data[1] = m10; data[5] = m11; data[9] = m12; data[13] = m13;
        data[2] = m20; data[6] = m21; data[10] = m22; data[14] = m23;
        data[3] = m30; data[7] = m31; data[11] = m32; data[15] = m33;
    }

    float data[16];
};

PYBIND11_MODULE(bindings, m) {
    py::class_<::mat4>(m, "mat4")
        .def(py::init<float>(), py::arg("v") = 1.0f)
        .def("__getitem__", [](mat4& m, py::ssize_t i) {
            return m.data[i];
        })
        .def("__setitem__", [](mat4& m, py::ssize_t i, float v) {
            m.data[i] = v;
        })
        .def("__str__", [](mat4& m) {
            #define str_row(a,b,c,d) std::to_string(a) + "," + std::to_string(b) + "," + std::to_string(c) + "," + std::to_string(d) + ","
            return "mat4(\n"
                   "    "+str_row(m.data[0], m.data[4], m.data[8], m.data[12])+"\n"+
                   "    "+str_row(m.data[1], m.data[5], m.data[9], m.data[13])+"\n"+
                   "    "+str_row(m.data[2], m.data[6], m.data[10], m.data[14])+"\n"+
                   "    "+str_row(m.data[3], m.data[7], m.data[11], m.data[15])+"\n"+
                   ")";
        })
        .def("value_ptr", [](mat4& m) {
            return py::array_t<float>({16}, m.data);
        })
    ;
};

CMakelists.txt

cmake_minimum_required(VERSION 3.4)

project(mcve)
add_subdirectory(pybind11)


pybind11_add_module(bindings "mcve.cpp")

test.py

from build.bindings import mat4

if __name__ == "__main__":
    m = mat4(1)
    ptr = m.value_ptr()

    print("----case1----")
    print(m)
    print(ptr)

    print("----case2----")
    m[5] = 2
    m[10] = 3
    m[15] = 4
    print(m)
    print("wrong content!!! ", ptr)
    print("valid content!!! ", m.value_ptr())

如果您试图通过执行类似的操作来编译和运行上述内容：

mkdir build && cd build && cmake .. -G Ninja && cmake --build . && cd ..
python test.py

一切顺利..。您的输出应该低于：

----case1----
mat4(
    1.000000,0.000000,0.000000,0.000000,
    0.000000,1.000000,0.000000,0.000000,
    0.000000,0.000000,1.000000,0.000000,
    0.000000,0.000000,0.000000,1.000000,
)
[1. 0. 0. 0. 0. 1. 0. 0. 0. 0. 1. 0. 0. 0. 0. 1.]
----case2----
mat4(
    1.000000,0.000000,0.000000,0.000000,
    0.000000,2.000000,0.000000,0.000000,
    0.000000,0.000000,3.000000,0.000000,
    0.000000,0.000000,0.000000,4.000000,
)
wrong content!!!  [1. 0. 0. 0. 0. 1. 0. 0. 0. 0. 1. 0. 0. 0. 0. 1.]
valid content!!!  [1. 0. 0. 0. 0. 2. 0. 0. 0. 0. 3. 0. 0. 0. 0. 4.]

如您所见，在case2中原始矩阵内容已被修改，您可以看到ptr仍然显示原始内容.因此，我需要再次运行m.value_ptr()，以获得更新的数据。

因此，我的问题是，如何使用c++原始指针使用pybind11有效地创建零拷贝视图？

我已经研究了几个小时，尝试了不同的方法，但我被困住了。首先，我想了解一下上面的代码有什么问题.感觉py::array在复制数组中的内容，而不是从数组中创建视图，所以.是否有任何方法可以通过使用py::array来避免这种复制，而只需要从指针中获得一个“视图”？实现这一目标的正确方法是什么？在实际情况下，我将处理大量内存，因此复制不必要的数据根本不是一种选择。

参考文献：

• returning numpy arrays via pybind11
• Returning and passing around raw POD pointers (arrays) with Python, C++, and pybind11
• https://pybind11.readthedocs.io/en/stable/advanced/functions.html#return-value-policies

Answer 1

可能的解决办法：

    .def("value_ptr", [](mat4& m) {
        return py::array_t<float>{16, m.data, py::cast(m)};
    })

取自https://github.com/pybind/pybind11/issues/2271#issuecomment-650569067的原始参考资料