用SWIG操纵C中的Numpy数组

Question

Trying to pass a Numpy array to a C method and modify its content there, is it even possible? The idea is to do as little data copying (speed and memory reasons) as possible.

Trying out something like this currently:

test.c :

void testMethod(int** values) {
  // code
  *values = other_pointer;
}

test.i :

/* File : test.i */
%module test
%{
  #define SWIG_FILE_WITH_INIT
%}
%include "numpy.i"

%init %{
    import_array();
%}

%{
extern void testMethod(int** values);
%}

%apply (int **ARGOUT_ARRAY1) {int **values};
extern void testMethod(int** values);

I modified the numpy.i file adding:

%typemap(in,
         fragment="NumPy_Fragments")
  (DATA_TYPE **ARGOUT_ARRAY1)
  (PyArrayObject* array=NULL, int is_new_object=0, DATA_TYPE* temp=NULL)
{
    array = obj_to_array_contiguous_allow_conversion($input,
                                                   DATA_TYPECODE,
                                                   &is_new_object);
    temp = (DATA_TYPE*) array_data(array);
    $1 = &temp;
}
%typemap(argout)
  (DATA_TYPE** ARGOUT_ARRAY1)
{
    $result = SWIG_Python_AppendOutput($result,(PyObject*)array$argnum);
}

Running it:

input = numpy.array([1,2,3])
test.testMethod(input)

Works, ie it compiles and I can print the contents of the array in C but content of input stays the same.

Answer 1

To answer my own question, one a bit hacky way would be to change the typemap to something like this:

%typemap(in,
         fragment="NumPy_Fragments")
  (DATA_TYPE **ARGOUT_ARRAY1)
  (PyArrayObject* array=NULL, int is_new_object=0, PyArrayObject_fields* temp=NULL)
{
    array = obj_to_array_contiguous_allow_conversion($input,
                                                   DATA_TYPECODE,
                                                   &is_new_object);
    if (!array) SWIG_fail;
    temp = (PyArrayObject_fields*)array;
    $1 = (DATA_TYPE**) &temp->data;
}
%typemap(argout)
  (DATA_TYPE** ARGOUT_ARRAY1)
{
    $result = SWIG_Python_AppendOutput($result,(PyObject*)array$argnum);
}

@Edit: after a bit of testing the above would work but produce Segmentation Fault exception if we reuse the Python input array. The solution is to change the mapping type from ARGOUT_ARRAY1 to INPLACE_ARRAY1 .