MATLAB parfor and C++ class mex wrappers (copy constructor required?)

Question

I'm trying to wrap a C++ class in a matlab mex wrapper using the approach outlined here. Basically, I have an initialization mex file which returns a C++ object handle:

handle = myclass_init()

I can then pass this to another mex file (e.g. myclass_amethod) which use the handle to call class methods, then ultimately to myclass_delete to free the C++ object:

retval = myclass_amethod(handle, parameter)
myclass_delete(handle)

I've wrapped this up in a MATLAB class for ease of use:

classdef myclass < handle
    properties(SetAccess=protected)
        cpp_handle_
    end
    methods
        % class constructor
        function obj = myclass()
            obj.cpp_handle_ = myclass_init();
        end
        % class destructor
        function delete(obj)
            myclass_delete(obj.cpp_handle_);
        end
        % class method
        function amethod(parameter)
            myclass_amethod(obj.cpp_handle_, parameter);
        end
    end
end

Problem: this doesn't work in parallel code

This works fine in non-parallel code. However, as soon as I call it from within a parfor:

cls = myclass();
parfor i = 1:10
    cls.amethod(i)
end

I get a segfault, as a copy of the class is made in the parfor loop (in each worker) but as each worker is a separate process the C++ object instance is not copied, leading to an invalid pointer.

I tried initially to detect when each class method was running in a parfor loop, and in those cases reallocate the C++ object too. However, as there is no way to check whether an object has been allocated for the current worker yet or not, this results in multiple reallocations and then just one delete (when the worker exits) resulting in a memory leak (see appendix at bottom of question for details).

Attempted solution: copy constructors and using matlab.mixin.Copyable

In C++, the way to handle this would be copy constructors (so that the C++ object is only reallocated once when the wrapper MATLAB class is copied). A quick search brings up the matlab.mixin.Copyable class type which seems to provide the required functionality (i.e. deep copies of MATLAB handle classes). Therefore, I've tried the following:

classdef myclass < matlab.mixin.Copyable
    properties(SetAccess=protected)
        cpp_handle_
    end
    methods
        function obj = myclass(val)
            if nargin < 1
                % regular constructor
                obj.cpp_handle_ = rand(1);
                disp(['Initialized myclass with handle: ' num2str(obj.cpp_handle_)]);
            else
                % copy constructor
                obj.cpp_handle_ = rand(1);
                disp(['Copy initialized myclass with handle: ' num2str(obj.cpp_handle_)]);
            end
        end
        % destructor
        function delete(obj)
            disp(['Deleted myclass with handle: ' num2str(obj.cpp_handle_)]);
        end
        % class method
        function amethod(obj)
            disp(['Class method called with handle: ' num2str(obj.cpp_handle_)]);
        end
    end
end

testing this class as above, i.e.:

cls = myclass();
parfor i = 1:10
    cls.amethod(i)
end

Results in the output:

Initialized myclass with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Class method called with handle: 0.65548
Deleted myclass with handle: 0.65548
Deleted myclass with handle: 0.65548
Deleted myclass with handle: 0.65548

In other words, it seems that the copy constructor is not called when spawning workers for the parfor. Does anyone have any pointers as to what I am doing wrong, or whether there is some way to achieve the desired behaviour of reinitializing the C++ object handle when the MATLAB wrapper class is copied?

---

Alternative approach: detecting when running on worker

Just for reference, here is the alternative approach I'm using reallocating when in a worker:

classdef myclass < handle
    properties(SetAccess=protected)
        cpp_handle_
    end
    methods
        function obj = myclass(val)
            obj.cpp_handle_ = rand(1);
            disp(['Initialized myclass with handle: ' num2str(obj.cpp_handle_)]);
        end
        % destructor
        function delete(obj)
            disp(['Deleted myclass with handle: ' num2str(obj.cpp_handle_)]);
        end
        % class method
        function amethod(obj)
            obj.check_handle()
            disp(['Class method called with handle: ' num2str(obj.cpp_handle_)]);
        end
        % reinitialize cpp handle if in a worker:
        function check_handle(obj)
            try
                t = getCurrentTask();
                % if 'getCurrentTask()' returns a task object, it means we
                % are running in a worker, so reinitialize the class
                if ~isempty(t)
                    obj.cpp_handle_ = rand(1);
                    disp(['cpp_handle_ reinitialized to ' num2str(obj.cpp_handle_)]);
                end
            catch e
                % in case of getCurrentTask() being undefined, this
                % probably simply means the PCT is not installed, so
                % continue without throwing an error
                if ~strcmp(e.identifier, 'MATLAB:UndefinedFunction')
                    rethrow(e);
                end
            end
        end
    end
end

and the output:

Initialized myclass with handle: 0.034446
cpp_handle_ reinitialized to 0.55625
Class method called with handle: 0.55625
cpp_handle_ reinitialized to 0.0048098
Class method called with handle: 0.0048098
cpp_handle_ reinitialized to 0.58711
Class method called with handle: 0.58711
cpp_handle_ reinitialized to 0.81725
Class method called with handle: 0.81725
cpp_handle_ reinitialized to 0.43991
cpp_handle_ reinitialized to 0.79006
cpp_handle_ reinitialized to 0.0015995
Class method called with handle: 0.0015995
cpp_handle_ reinitialized to 0.0042699
cpp_handle_ reinitialized to 0.51094
Class method called with handle: 0.51094
Class method called with handle: 0.0042699
Class method called with handle: 0.43991
cpp_handle_ reinitialized to 0.45428
Deleted myclass with handle: 0.0042699
Class method called with handle: 0.79006
Deleted myclass with handle: 0.43991
Deleted myclass with handle: 0.79006
Class method called with handle: 0.45428

As can be seen above, a reallocation does indeed now occur when running in a worker. However, the destructor is only called once for every worker regardless of how many time the C++ class was reallocated, resulting in a memory leak.

---

Solution: using loadobj

The following works:

classdef myclass < handle
    properties(SetAccess=protected, Transient=true)
        cpp_handle_
    end
    methods(Static=true)
        function obj = loadobj(a)
            a.cpp_handle_ = rand(1);
            disp(['Load initialized encoder with handle: ' num2str(a.cpp_handle_)]);
            obj = a;
        end
    end
    methods
        function obj = myclass(val)
            obj.cpp_handle_ = rand(1);
            disp(['Initialized myclass with handle: ' num2str(obj.cpp_handle_)]);
        end
        % destructor
        function delete(obj)
            disp(['Deleted myclass with handle: ' num2str(obj.cpp_handle_)]);
        end
        % class method
        function amethod(obj)
            disp(['Class method called with handle: ' num2str(obj.cpp_handle_)]);
        end
    end
end

Answer 1

When you pass an object instance into the body of a PARFOR loop, the behaviour is the same as if you'd saved it to a file, and then loaded it again. The simplest solution is probably to mark your cpp_handle_ as Transient. Then, you need to implement SAVEOBJ and LOADOBJ to safely transport your data. See this page for more about customising the save/load behaviour of your class.