通过 python ctypes 使用时共享库段错误中的 opencv 调用答案

【问题标题】：opencv calls in shared library segfaults when used through python ctypes通过 python ctypes 使用时共享库段错误中的 opencv 调用
【发布时间】：2012-05-06 04:45:37
【问题描述】：

每当我使用 ctypes 将共享库 (.so) 加载到 Python 中并调用其中一个使用 OpenCV 过滤器引擎（如 cvErode 或 cvDilate）的过程时，解释器都会在 OpenCV 代码深处发生段错误。如果我在 C 程序中使用 dlopen 加载相同的库并调用相同的过程，一切正常。

我在另一篇文章中读到它可能与 gcc 使用的 ABI 有关，并且 python 必须使用与 python 解释器编译的相同版本的 gcc 编译才能工作。但是在我的情况下，它都是 4.4.5，所以这不应该是一个问题。 OpenCV 库也是使用相同版本的 gcc 编译的。

共享库是如何构建的：

g++ -g -shared -lopencv_core -lopencv_imgproc -lopencv_calib3d -lopencv_video -lopencv_features2d -lopencv_ml -lopencv_highgui -lopencv_objdetect -lopencv_contrib -lopencv_legacy -o thumbsplit.so thumbsplit.cpp

共享库的包装方式：

extern "C" {

int is_thumbsheet(char * image_path){
    Image image;
    image.read(image_path);
    int width = image.columns();
    int height = image.rows();
    int BOUND_CHANGE = 7000;
    int test_height = 0;

    if( width > height ){
        return 0;
    }else{
        test_height = width;
    }
    vector<int> test_rows = getRows(image, width, test_height, BOUND_CHANGE);
    vector<int> test_cols = getCols(image, width, test_height, BOUND_CHANGE);

    vector<Box> test_boxes = createBoxes(test_rows,test_cols,width,height);
    if( test_boxes.size() > 6 ){
        return 1;
    }else{
        return 0;
    }
}

int cut_thumbs(char * image_path, char * thumb_path){
    Image image;
    image.read(image_path);
    int width = image.columns();
    int height = image.rows();
    int BOUND_CHANGE = 7000;
    vector<int> rows = getRows(image, width, height, BOUND_CHANGE);
    vector<int> cols = getCols(image, width, height, BOUND_CHANGE);

    //drawDebugLines(rows, cols, width, height);

    vector<Box> boxes = createBoxes(rows,cols,width,height);
    //printBoxes(boxes);
    char path[0x100];
    for( int i=0 ; i < boxes.size() ; i++ ){
        sprintf(path, "%s%d.jpg", thumb_path, i);
        boxes[i].saveBox(image, path);
    }


    return 0;
}

} // C

共享库过程在python中是如何加载的：

import ctypes, os

def load_lib():
lib_path = "%s/%s"%(os.path.realpath(os.path.dirname(__file__)), "thumbsplit.so")
lib = ctypes.CDLL(lib_path)
return lib

def is_thumbsheet(image_path):

lib = load_lib()
if lib.is_thumbsheet(image_path):
    return True
else:
    return False

def cut_thumbs(image_path, thumbs_path):

lib = load_lib()
lib.cut_thumbs(image_path, thumbs_path)

is_thumbsheet("/home/rolf/test/imageproc/full9.jpg") # this segfaults

GDB 输出：

$ gdb python
GNU gdb (GDB) 7.0.1-debian
Copyright (C) 2009 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.  Type "show copying"
and "show warranty" for details.
This GDB was configured as "i486-linux-gnu".
For bug reporting instructions, please see:
<http://www.gnu.org/software/gdb/bugs/>...
Reading symbols from /usr/bin/python...(no debugging symbols found)...done.
(gdb) run thumbsheet.py
Starting program: /usr/bin/python thumbsheet.py
[Thread debugging using libthread_db enabled]
0xb6330a12

Program received signal SIGSEGV, Segmentation fault.
0xb76fe2dc in cv::FilterEngine::start(cv::Size_<int>, cv::Rect_<int>, int) () from /usr/local/lib/libopencv_imgproc.so.2.3
(gdb) info stack
#0  0xb76fe2dc in cv::FilterEngine::start(cv::Size_<int>, cv::Rect_<int>, int) () from /usr/local/lib/libopencv_imgproc.so.2.3
#1  0xb76feb41 in cv::FilterEngine::start(cv::Mat const&, cv::Rect_<int> const&, bool, int) ()
   from /usr/local/lib/libopencv_imgproc.so.2.3
#2  0xb770a52b in cv::FilterEngine::apply(cv::Mat const&, cv::Mat&, cv::Rect_<int> const&, cv::Point_<int>, bool) ()
   from /usr/local/lib/libopencv_imgproc.so.2.3
#3  0xb764bf8a in cv::morphOp(int, cv::_InputArray const&, cv::_OutputArray const&, cv::_InputArray const&, cv::Point_<int>, int, int, cv::Scalar_<double> const&) () from /usr/local/lib/libopencv_imgproc.so.2.3
#4  0xb764d2ce in cv::erode(cv::_InputArray const&, cv::_OutputArray const&, cv::_InputArray const&, cv::Point_<int>, int, int, cv::Scalar_<double> const&) () from /usr/local/lib/libopencv_imgproc.so.2.3
#5  0xb764d62e in cvErode () from /usr/local/lib/libopencv_imgproc.so.2.3
#6  0xb7fd7adb in seperate::findSeperatedBoxes (img=0x8351d70) at seperate.cpp:189
#7  0xb7fd7fcf in seperate::trySeperatedBoxes (img=0x8351d70, percentage_boxed=0xbfffefec) at seperate.cpp:202
#8  0xb7fda00c in is_thumbsheet (image_path=0xb7c1bf74 "/home/rolf/test/imageproc/full9.jpg") at thumbsplit.cpp:39
#9  0xb7a047df in ffi_call_SYSV () from /usr/lib/python2.6/lib-dynload/_ctypes.so
#10 0xb7a0461e in ffi_call () from /usr/lib/python2.6/lib-dynload/_ctypes.so
#11 0xb79ff27d in _CallProc () from /usr/lib/python2.6/lib-dynload/_ctypes.so
#12 0xb79f6d7e in ?? () from /usr/lib/python2.6/lib-dynload/_ctypes.so
#13 0x0806232a in PyObject_Call ()
#14 0x080e016b in PyEval_EvalFrameEx ()
#15 0x080e18b0 in PyEval_EvalFrameEx ()
#16 0x080e2507 in PyEval_EvalCodeEx ()
#17 0x080e2607 in PyEval_EvalCode ()
#18 0x080ffcbd in PyRun_FileExFlags ()
#19 0x080fff22 in PyRun_SimpleFileExFlags ()
#20 0x0805dd81 in Py_Main ()
#21 0x0805cf6b in main ()
(gdb) x/i 0xb76fe2dc
0xb76fe2dc <_ZN2cv12FilterEngine5startENS_5Size_IiEENS_5Rect_IiEEi+1212>:   movdqa %xmm3,0x30(%esp)
(gdb)

GCC/G++ 版本：

$ python
Python 2.6.6 (r266:84292, Dec 27 2010, 00:02:40) 
[GCC 4.4.5] on linux2

$ gcc --version
gcc (Debian 4.4.5-8) 4.4.5

$ g++ --version
g++ (Debian 4.4.5-8) 4.4.5

OpenCV 是 2.3.1 版，是使用这个版本的 gcc (4.4.5-8) 从源代码构建的

【问题讨论】：

您确实意识到 OpenCV 有 Python 绑定？ code.google.com/p/pyopencv
是的，但出于速度原因，我宁愿在 c/c++ 中进行所有图像处理。并非所有这些都是在 OpenCV 中完成的。
如果您编写了一个 C 程序来加载和使用您的库，您就会知道问题出在您的 OpenCV 代码中，还是在 Python 中加载库的代码中。
我认为您不会看到明显的速度差异。 Python Image Library (PIL) 是用 C 语言编写的 Python 扩展。
@karlphillip：我做到了，效果很好。加载库的代码也是正确的，因为它适用于不使用 OpenCV 的其他模块。

标签： c++ python gcc opencv ctypes

【解决方案1】：

我找到了解决方案。对于 SSE 指令，堆栈需要重新对齐为 16 字节。如果我使用 -mstackrealign 构建共享库，它可以完美加载并从 python 调用。

Qt, GCC, SSE and stack alignment

【讨论】：