如何利用numpy.i的ARGOUT_ARRAY2规则封装返回3*3矩阵的C++函数？

Great question! Let's unpack this clearly, since numpy.i's type mappings can feel opaque at first.

What does ANY mean in ARGOUT_ARRAY2[ANY][ANY]? #

ANY is a placeholder that tells SWIG the dimension size is not fixed—it can be any integer value. This is useful when your function works with matrices of variable sizes (e.g., 2x2, 5x5, etc.). But since your function specifically creates a 3x3 matrix, you don't need ANY here—you can directly replace it with the fixed dimension size (3) to make the type mapping explicit and safe.

Wrapping your void make_matrix(double** matrix) function #

Here's a step-by-step breakdown of how to bind this function correctly in your SWIG interface file:

1. Set up the SWIG module and include numpy.i #

First, start your .i file with the module declaration, include your C++ header, and initialize numpy's SWIG support:

%module matrix_wrapper
%{
// Include your C++ function's header here
#include "matrix_generator.h"
%}

// Import numpy's SWIG helpers and initialize the numpy C API
%include "numpy.i"
%init %{
import_array();
%}

2. Apply the fixed-size array type mapping #

Use the %apply directive to map your double** matrix parameter to numpy's ARGOUT_ARRAY2 type for a 3x3 double matrix. This tells SWIG to handle converting the C++ allocated memory into a numpy array that Python can use:

// Map the C++ output parameter to a 3x3 double numpy array
%apply (double ARGOUT_ARRAY2[3][3]) { (double** matrix) }

3. Declare the function to wrap #

Finally, declare the function in your SWIG interface so it gets exposed to Python:

void make_matrix(double** matrix);

How this works #

When you compile this SWIG module and call make_matrix() from Python:

• SWIG will handle the memory management behind the scenes: it takes the 3x3 matrix allocated via malloc() in your C++ function, copies the data into a new numpy array, and then frees the original malloc()'d memory (you don't have to handle this manually in Python).
• The function will return a 3x3 numpy array directly—no need to pass a pre-allocated array from Python, since ARGOUT_ARRAY2 marks this as an output parameter.

Note about memory safety #

If your C++ function expects the caller to free the allocated memory, the ARGOUT_ARRAY2 type mapping will automatically free it after copying the data to the numpy array. If you need to keep the C++ memory for other operations, you'd need a custom type mapping, but for most use cases, the default behavior is exactly what you want.

内容的提问来源于stack exchange，提问作者Minor Threat