Passing composite variable to mex file

I am running an spmd, and I am wondering if there is a way to pass a composite variable to a mex file without "aggregating" it first. The following is a (simple) example. Say I have the following code.

parpool(2);

spmd 
    test = labindex + zeros(1000, 2);
end

result = sum([test{:}], 2);

Suppose I would like to replace the sum function with a C/mex implementation of it. It is possible to do something like the following:

test_cellarray = test(:);
result = sum_function(test_cellarray); % Not a built-in function

where sum_function takes in cell arrays. However, in the above "solution", the first line takes up the vast majority of the time.

Is it possible to pass "test" directly to a mex function and access the parts for different workers through pointers?

Thanks for the help.

[James Tursa]

What exactly is the test variable? I.e., what does "whos test" show?

[Vivek Bhattacharya]

@James Tursa -- This is the output from inspecting the variables. (Including test_cellarray for illustration too.)

>> test
 
test =
 
   Worker 1: class = double, size = [1000     2]
   Worker 2: class = double, size = [1000     2]
 
>> test_cellarray

test_cellarray =

  2×1 cell array

    {1000×2 double}
    {1000×2 double}

>> whos test test_cellarray
  Name                Size            Bytes  Class        Attributes

  test                1x2               265  Composite              
  test_cellarray      2x1             32208  cell                   

Answer 1

[Edric Ellis]

Unfortunately, there is no way to access the elements of a Composite from a MEX file. Is there any way you could run the MEX file directly on the workers? If you need to run it on each element, that would probably be the best way.

[Vivek Bhattacharya]

Thanks, @Edric Ellis! Yes, I'm currently doing most of the computation within the spmd loop, but there is one part of the calculation that requires the entire vector of output from each worker and can't be done individually on each worker. (Perhaps the sum is a bad example on that dimension.) I was hoping that it would be possible to send pointers to each worker's data to a mex function (or something like that), but good to know that that's not possible.

[Edric Ellis]

You might be better off collecting the data on a single worker, and working with it there. Worker-to-worker communication is generally more efficient than worker-to-client. I.e. something like:

spmd
    myVal = someComputation();
    catDim = 1; % dimension to concatenate results
    targetWorker = 1;
    allVals = gcat(myVal, catDim, targetWorker);
    if labindex == targetWorker
        endResult = doStuff(allVals);
    end
end