Matlab Get Property function causing performance problems

Question

I have a property in one of the classes I created that currently has an unnecessary get function associated with it. I initially made it to do useful things before returning the value, but now I just need the property value, so the get function has been reduced to this:

    function value =  get.error(obj)
        value = obj.error;
    end

Now, seeing as this code is completely redundant, I decided to delete the get function all together, but this leads to an incredible slowdown in one section of my code where the 'error' property is accessed repeatedly.

In this section of code, the profiler doesn't explicitly say that the missing get function is the cause of the problem (it says all of the time is wasted at the 'end' of a 'for' loop), but when I add the functionally useless code back in, the performance problem goes away.

Why would deleting this useless get function slow down my code?

EDIT: I've isolated the problem enough to post some code samples.

Here's a dummy version of the method call that has the problem:

Here's the code profile without the useless getter:

and with the magical useless getter:

Note that the requires for the performance appear to be as follows:

1. use some property of my 'pool' object in a computation set to any variable. The 'error' property is the one that caught my attention, but the bug happens with all properties in the same situation.

2. the computation involves anything, even '.* 0' causes this slowdown, but a single term being set is slowdown free (e.g. obj.pools(i).delta = obj.pools(i).error)

EDIT 2:

Here's the full pool class; perhaps it'll help:

classdef pool < handle

properties
    name;
    unit_count;
    activation;
    net_input = 0;
    %all projections now incoming
    projections = [];
    error; % same is dEd net for rbp
    target;
    clamped_error = false;
    delta;
    clamped_activation = 0; %0 no clamp, 1 soft clamp, 2 hard clamp
    copyback = false;
    copy_from;
    type = 'hidden';

    activation_history;
    error_history;

end

methods
    function obj = pool(name,count,type,copy_from)
        obj.name = name;
        assignin('caller', name, obj);
        obj.unit_count = count;
        obj.error = zeros(1,count);
        obj.target = zeros(1,count);
        obj.delta = zeros(1,count);
        obj.activation = zeros(1,count);
        obj.net_input = zeros(1,count);
        obj.activation_history(:,1) = zeros(1,count);
        obj.error_history(:,1) = zeros(1,count);
        if nargin > 2
            obj.type = type;
            if nargin == 4
                obj.clamped_activation = 2;
                obj.activation = ones(1,count)/2;
                obj.copy_from = copy_from;
                obj.copyback = true;
            end
        else
            obj.type = 'hidden';
        end
        switch obj.type
            case 'input'
                obj.clamped_activation = 2;
            case 'output'
                obj.clamped_error = true;
            case 'bias'
                obj.clamped_activation = 2;
                obj.activation = 1;
            case 'hidden'
        end

    end

    function proj = connect(obj,send_pool,proj_var)
        %see if you need a new projection or if the user provided one

        if nargin == 2
            proj = projection(obj, send_pool);
        else
            proj = proj_var;
        end
        obj.projections = [obj.projections struct('from',send_pool,'using',proj)];
    end


    function value =  get.error(obj)
        value = obj.error;
    end

end

end

Answer 1

This sounds like Matlab JIT or interpreter weirdness.

You may be looking at a Matlab bug related to MCOS memory management, like that discussed in Dataset array indexing is very slow with Statistics Toolbox. That bug looks like it causes spurious deep copying of objects when referencing their fields in some cases. If you have cell arrays or other complex structures in your pool objects, that could cause similar behavior - they get copied in to new variables, and maybe the deallocation cost is showing up at the "end".

I was able to reproduce similar surprising slow results in several Matlab versions. But I wasn't able to show any difference from the presence of a getter.

File fooclass.m.

classdef fooclass
    properties
        error = 42;
        whatever = 42;
        delta = 1;
        bigjunk = num2cell(rand(1000,40));
        %bigjunk = rand(1000,10000);  % non-cell does not cause slowdown
    end

    methods        
%         function out = get.error(obj)
%             out = obj.error;
%         end
    end

end

File looper.m.

classdef looper
    properties
        pools
    end

    methods
        function obj = looper()
            obj.pools = repmat(fooclass(), [1 5]);;
        end

        function weird(obj)
        p = obj.pools;
        n = numel(p);
        for i = 1:n
            dummy = obj.pools(i).error .* 0;
        end
        end
    end
end

To reproduce:

>> lp = looper;
>> tic; for i = 1:100; weird(lp); end; toc
Elapsed time is 0.600428 seconds.

That's on my R2011a (prerelease) on 64-bit Windows 7. That's slow. The time is proportional to the size of bigjunk, and the profiler says it almost all happens on the p = obj.pools line. This suggests that the object reference is causing a copy instead of using the copy-on-write optimization like it's supposed to. Or it's walking the object graph, or something. Something similar could be going on with your code.

Might be best to just chalk it up to interpreter weirdness, leave the getter in there if it makes your code faster, and wait for a fix in the next release. Sounds worth submitting to MathWorks as a bug report.