python - 从扩展函数返回后使用自定义 Python 扩展的堆损坏 (0xC0000374)

Question

我正在为我的 Python 脚本编写一个扩展模块，但无法弄清楚这个堆损坏的来源。扩展模块的使用应该是创建一个 Numpy 数组，用给定一些条件计算的一些值填充它，并将 Numpy 数组返回给 Python。有一个 Python 函数（在下面的示例中get_data()），它包装了扩展函数并从中接收 Numpy 数组。当 Python 函数返回时，会发生错误。

当我使用"OOO"而不是"NNN"作为Py_BuildValue().

因此，我认为这是关于引用计数的问题。如果您查看下面的示例代码，是否有人熟悉这个问题？

开始使用 WinDbg，但完全是初学者。

注意：不幸的是不能共享原始代码，这当然使这有点困难。下面的示例代码对我没有任何错误，但是它使用了相同的概念，问题可能出在其中。

模块.cpp

#define PY_SSIZE_T_CLEAN
#define NPY_NO_DEPRECATED_API NPY_API_VERSION

#include <Python.h>
#include <numpy\arrayobject.h>

// Helper for making some exemplary changes on data array
void applySomeChangesOnData(double* data, double factor, size_t nRows, size_t nColumns) {
    for (int m = 0; m < nRows; m++) {
        for (int n = 0; n < nColumns; n++) {
            data[m*nColumns+n] = factor * (m + n);
        }
    }
}

PyObject* createArrays(PyObject *self, PyObject *args) {
    // Variables needed for parsing inputs
    PyObject* tupleShape;

    // Parse input
    if (!PyArg_ParseTuple(args, "O!", &PyTuple_Type, &tupleShape)) {
        // Error when parsing
        PyErr_SetString(PyExc_TypeError, "Bad input type(s)");
        Py_RETURN_NONE;
    }

    // Allow only 2D arrays for this example
    if ((size_t)PyTuple_Size(tupleShape) != 2) {
        PyErr_SetString(PyExc_ValueError, "Array must be 2D");
        Py_RETURN_NONE;
    }

    // Allocate data array
    size_t nRows = PyLong_AsLong(PyTuple_GetItem(tupleShape, 0));
    size_t nColumns = PyLong_AsLong(PyTuple_GetItem(tupleShape, 1));
    size_t nElements = nRows * nColumns;
    double* data1 = PyMem_New(double, nElements);
    double* data2 = PyMem_New(double, nElements);
    double* data3 = PyMem_New(double, nElements);

    // Modify data array
    applySomeChangesOnData(data1, 1, nRows, nColumns);
    applySomeChangesOnData(data2, 10, nRows, nColumns);
    applySomeChangesOnData(data3, 100, nRows, nColumns);

    // Prepare shape information
    size_t nDims = 1;
    npy_intp* shape = new npy_intp[nDims];
    shape[0] = nElements;

    // Create output array and set OWNDATA flag for proper deallocation
    PyArrayObject* arr1 = reinterpret_cast<PyArrayObject*>(PyArray_SimpleNewFromData(nDims, shape, NPY_DOUBLE, data1));
    PyArrayObject* arr2 = reinterpret_cast<PyArrayObject*>(PyArray_SimpleNewFromData(nDims, shape, NPY_DOUBLE, data2));
    PyArrayObject* arr3 = reinterpret_cast<PyArrayObject*>(PyArray_SimpleNewFromData(nDims, shape, NPY_DOUBLE, data3));
    if (!arr1 || !arr2 || !arr3) {
        PyErr_SetString(PyExc_RuntimeError, "Failed when creating output array");
        Py_RETURN_NONE;
    }
    PyArray_ENABLEFLAGS(arr1, NPY_ARRAY_OWNDATA);
    PyArray_ENABLEFLAGS(arr2, NPY_ARRAY_OWNDATA);
    PyArray_ENABLEFLAGS(arr3, NPY_ARRAY_OWNDATA);

    // Some clean-up
    delete[] shape; shape = NULL;

    // Return multiple outputs
    PyObject* ret = Py_BuildValue("NNN", arr1, arr2, arr3);
    return ret;
}

static PyMethodDef extension_methods[] = {
    { "create_arrays",                  (PyCFunction)createArrays,      METH_VARARGS,                   nullptr },
    { nullptr,                          nullptr,                        0,                              nullptr },
};

static PyModuleDef extension_module = { PyModuleDef_HEAD_INIT, "extension", "Some docs...", 0, extension_methods };

PyMODINIT_FUNC PyInit_extension() {
    import_array();
    return PyModule_Create(&extension_module);
}

安装程序.py

import os
import sys
from setuptools import setup, Extension, find_packages

# Get installation path of Python interpreter
(path_interpreter, _) = os.path.split(sys.executable)

ext_module = Extension('example_package.extension',
                       sources=['example_package/module.cpp',],
                       include_dirs=[os.path.join(path_interpreter, 'Lib/site-packages/numpy/core/include')],
                       extra_compile_args=['/Zi'],
                       extra_link_args=['/DEBUG'])

setup(
    name='example_package',
    version='0.1',
    ext_modules=[ext_module],
)

debug_extension.py

import numpy as np
from example_package.extension import create_arrays

shape = (4000, 6000)

def get_data():
    # Write 5 pairs of arrays into this list
    list_pairs = []

    for i in range(50):
        # Get 3 arrays from extension
        (arr1, arr2, arr3) = create_arrays(shape)

        # Reshape, transpose and later combine them using np.stack()
        arr1 = np.reshape(arr1, shape).transpose()
        arr2 = np.reshape(arr2, shape).transpose()
        arr3 = np.reshape(arr3, shape).transpose()

        list_pairs.append([
            np.stack([arr1, arr2], axis=0),
            np.stack([arr1, arr3], axis=1),
        ])

    return list_pairs

list_pairs = get_data()

标签： pythonc++windowsnumpy