MATLAB: Using (popular) mexed c++ kdtree from fileexchange leads to Matlab system crashes.

kdtree librarymatlab system errormemory consistencymex

I have matlab system errors i.e. total crashes, when I run the following kdtree implementation shared on fileexchange https://ch.mathworks.com/matlabcentral/fileexchange/21512-ataiya-kdtree ; Matlab crashes already for the there provided minimal kdtree_delete_demo.
In this demo
  • two cpp files are mexed to provide two separate functions: kdtree_build and kdtree_delete;
  • A c++ object is allocated via kdtree_build, and returned as a pointer to matlab
  • The c++ object is deleted by a call to kdtree_delete, the pointer is provided as an argument.
When I run it:
  • Mexing and creation works fine – apart of a few uninteresting c++ warnings nothing is reported in verbose mex mode.
  • The crashes happen upon deletion
  • With printouts I was able to pin down that the crash happens as soon as any member of the dereferenced pointer is accessed (namely tree->ndim()).
My question: Could memory persistancy be an issue for this library? As far as I can tell the kdtree library allocates memory in the standard c++ way (new/ ~) without doing any matlab related magic. It is strange that for many the library works fine. I have no experience with mexing, only with c++ and cannot judge if the library has issues or if this example not working means that something is fishy with my whole OS/compilers/system. (Windows 10, Matlab 2016b, Visual studio 2013 c++ professional)
Ps. I cannot easily switch to an alternative kdtree implementation, I try to run code which uses this implementation as 3rd party code, and I really would prefer not to start rewriting this code.
**Update****
Praise goes to Philip Borghesani who found the problem. As visual c++ longs have 4 Byte only, all occurences of long need to be replaced by intptr_t. This need to be done in the files:
  • KDTree.h
  • kdtree_ball_query.cpp
  • kdtree_build.cpp
  • kdtree_io_from_mat.cpp
  • kdtree_k_neares_neighbors.cpp
  • kdtree_nearest_neighbor.cpp
  • kdtree_range_query.cpp

Best Answer

After a quick look at the code I will guess it is not compatible with 64 bit windows. If the objects are allocated at memory addresses below 4gb it might work otherwise this code will cause a crash:
    static KDTree* retrieve_pointer(const mxArray* matptr){
        // retrieve pointer from the MX form
        double* pointer0 = mxGetPr(matptr);
        // check that I actually received something
        if( pointer0 == NULL )
            mexErrMsgTxt("varargin{1} must be a valid kdtree pointer\n");
        // convert it to "long" datatype (good for addresses)
        long pointer1 = (long) pointer0[0];
        // convert it to "KDTree"
        KDTree* tree = (KDTree*) pointer1;
        // check that I actually received something
        if( tree == NULL )
            mexErrMsgTxt("varargin{1} must be a valid kdtree pointer\n");
        if( tree->ndims() <= 0 )
            mexErrMsgTxt("the k-D tree must have k>0");
        return tree;
    }
The problem is that on win64 long can't store a pointer. It is possible that by changing this line to use size_t or intptr_t and fixing the corresponding code for creating the matptr the code will work but there are most likely other other problems.
Change:
         long pointer1 = (long) pointer0[0];
To
        intptr_t pointer1 = (intptr_t) pointer0[0];
Related Question