How to write an efficient overloaded subsasgn?

I want to write a class which defines subsasgn. The class holds a numeric value val and the subsasgn should target val only.

I wrote the following test:

function [] = test()
    num = 1000;
    a = randi([0 255],[1000 num]);
    b = Subsasgntest(randi([0 255],[1000 num]));
    c = randi([0 255],[1 num]);
    tic;
    for i = 1:length(c)
        a(2,i) = c(i);
    end
    toc
    tic;
    for i = 1:length(c)
        b(3,i) = c(i);
    end
    toc
    b = b.setKind(1);
    tic;
    for i = 1:length(c)
        b(4,i) = c(i);
    end
    toc
end

classdef Subsasgntest
    
    properties
        val
        kind = 0;
    end
    
    methods
        function obj = Subsasgntest(in)
            obj.val = in;
        end
        
        function obj = setKind(obj,kind)
            obj.kind = kind;
        end
        
        function obj = subsasgn(obj,sub,varargin)
            %SUBSASGN Subscripted assignment
            if(isequal(sub.type,'()'))
                if isequal(obj,[])
                    error('Not yet implemented!');
                end
                if(isa(varargin{1},'Subsasgntest'))
                    switch obj.kind
                        case 0
                            obj.val = builtin('subsasgn',obj.val,sub,varargin{1}.val);
                        case 1
                            obj.val = subsasgn(obj.val,sub,varargin{1}.val);
                    end
                else
                    switch obj.kind
                        case 0
                            obj.val = builtin('subsasgn',obj.val,sub,varargin{1});
                        case 1
                            obj.val = subsasgn(obj.val,sub,varargin{1});
                    end
                end
            else
                error('This is not supported for this class!');
            end
        end
    end
end

Results are;

Elapsed time is 0.000019 seconds.
Elapsed time is 4.477900 seconds.
Elapsed time is 4.449843 seconds.

I know self-written subsasgn can be slow, but 235680x slower? Is there any way to achieve a similar level of optimization as in matlabs standard subsasgn?

This guide tells how to branch for different input subscripts, but not how to write the subsasgn syntx itself: https://www.mathworks.com/help/matlab/matlab_oop/code-patterns-for-subsref-and-subsasgn-methods.html

[Jan Siegmund]

Update: I added a third kind of subsasgn:

...
case 2
    obj.val(sub.subs{1},sub.subs{2}) = varargin{1}.val;
...

which does not cover all cases, but works for the little example and is much faster:

Elapsed time is 0.028915 seconds.

I think the key here is making matlab realize, that in this call:

obj.val = subsasgn(obj.val,sub,varargin{1}.val);

the rvalue obj.val is equal the lvalue obj.val, so it does not create a temporary obj.val.

Any ideas how I could do this?

Answer 1

[Jan Siegmund]

I may have a solution but do not know if this covers all cases:

classdef Subsasgntest
    
    properties
        val
    end
    
    methods
        function obj = Subsasgntest(in)
            obj.val = in;
        end
        
        function obj = subsasgn(obj,sub,varargin)
            %SUBSASGN Subscripted assignment
            if(isequal(sub.type,'()'))
                if isequal(obj,[])
                    error('Not yet implemented!');
                end
                if(isa(varargin{1},'Subsasgntest'))
                    obj.val(sub.subs{:}) = varargin{1}.val;
                else
                    obj.val(sub.subs{:}) = varargin{1};
                end
            else
                error('This is not supported for this class!');
            end
        end
    end
end

[Jan Siegmund]

Be careful: Property validation like

properties
    val uint64
end  

destroys the speed benefit of

obj.val(sub.subs{:}) = varargin{1};

as a temporary matrix is still created!

[Tom DeLonge]

Thanks for answering your own question. I have code with the same issue. However, I find that even considering your solution, the code is not sufficiently fast to compete with the matlab's native assignment. Did you make any progress in speeding the code up even further?

[Jan Siegmund]

Nope, that was my best approach. But if that helps: It makes a difference on which Matlab version you are on. So if you do not have the newest, try upgrading. Upgrading to 2020b for example fixed the delay caused by the property validation.

[Tom DeLonge]

Alright, thanks!

[James Lebak]

In MATLAB R2021b there is a new way to overload subscripting which has two advantages:

1) It allows the class author to overload only a particular type of indexing, that is, paren, dot, or brace, or some combination of types.

2) It improves the performance of indexing into contained objects.

See the section on 'Modular indexing' here to get started.