How to share a plain c-array as numpy array, reflecting changes in c-array data, and without copy

[erwincoumans]

See also Issue #1042 (#1042) but for plain C-array

I like to share a c-array to Python as numpy array, without copying data. If the c-array changes, the numpy array should reflect the changes. When using py::array, the data seems to be copied and not changed, when the c-array changes.

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

namespace py = pybind11;

struct MyClass
{
	int size_;
	float* data_;

	py::array_t<float> map_data_;

	MyClass(int sz)
	:size_(sz)
	{
		data_= (float*) malloc(sz*sizeof(float));
		for (int i=0;i<size_;i++)
			data_[i] = float(i);
		//want map_data_ to share data_ as numpy array without copy
		map_data_ = py::array_t<float>(size_, data_);	
	}
	void set_elem(int pos, float value)
	{
		if (pos>=0 && pos<size_)
			{
				data_[pos] = value;
				printf("ok!\n");
			} else
			{
				printf("pos out of range(%d, %d)\n", pos, size_);
			}

	}
	virtual ~MyClass()
	{
		free(data_);
	}
};

     

PYBIND11_MODULE(pybind11_test, m) {
	 py::class_<MyClass>(m, "MyClass")
	 .def(py::init<int>())
	 .def("set_elem", &MyClass::set_elem)
   .def_readonly("size_", &MyClass::size_)
   .def_readwrite("data", &MyClass::map_data_)     
  	;
}

when using this test script:

import pybind11_test as t
a = t.MyClass(10)
print(a.data)
a.set_elem(0,123)
print(a.data)

The output is

[0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]
ok!
[0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]

I'd like to see 123 in the first field.

(here is a github link to reproduce, call python setup.py install and run python test.py)

What can be done to fix this?

[erwincoumans]

adding an empty capsule fixed it!

//want map_data_ to share data_ as numpy array without copy
py::capsule buffer_handle([](){});
map_data_ = py::array_t<float>(size_, data_, buffer_handle);	

now:

[0. 1. 2. 3. 4. 5. 6. 7. 8. 9.]
ok!
[123.   1.   2.   3.   4.   5.   6.   7.   8.   9.]