MATLAB Integer errors

[jared321]

For quite some time we have had occasional failures of the IBCDFO MATLAB actions with strange error messages indicating that MATLAB code was passed index variables whose values are not positive integers. Such errors have not been seen from direct use and testing of the code on personal machines or for studies. In addition, the actions often pass by just rerunning the failed actions.

We spent some time debugging the following particular flavor of these errors, whose message is copied from a macOS-14/R2021a failure

================================================================================
Error occurred in Testpounders/shortTests and it did not run to completion.
    ---------
    Error ID:
    ---------
    'MATLAB:getReshapeDims:notRealInt'
    --------------
    Error Details:
    --------------
    Error using reshape
    Size arguments must be real integers.
    
    Error in pounders (line 183)
            Res(i, :) = F(i, :) - Cres - .5 * D * reshape(D * reshape(Hres, n, m
            * n), n, m);
    
    Error in test_identity_combine (line 53)
        [X, F, hf, flag, xk_best] = pounders(Ffun, X_0, n, nf_max, g_tol,
        delta_0, 1, Low, Upp, Prior, Options, Model);
    
    Error in Testpounders/shortTests (line 25)
                test_identity_combine;
================================================================================

Looking back over the recent action history, I see failures for macOS 15/R2021a. Note that we are presently only testing with that MATLAB version and the latest (R2025b).

This specific error can be triggered by calls such as reshape(eye(4), 2.0, 8.5) but not with other potential index errors

>> reshape(eye(4), 2.0, 8.0 + 1 * j)
Error using reshape
Size argument cannot be complex.
 
>> reshape(eye(4), 2.0, 8.0 + 0 * j)

ans =

     1     0     0     0     0     1     0     0
     0     0     1     0     0     0     0     1

>> reshape(eye(4), 2, 0)            
Error using reshape
Number of elements must not change. Use [] as one of the size inputs to automatically calculate the appropriate size for that dimension.
 
>> reshape(eye(4), 2, -1)
Error using reshape
Size vector elements should be nonnegative.

Some debugging output when that line was wrapped in a try/catch block

================================================================================
Error occurred in Testpounders/shortTests and it did not run to completion.
    ---------
    Error ID:
    ---------
    'POUNDERS:ReshapeDims:AssertionFail'
    --------------
    Error Details:
    --------------
    Error using pounders (line 217)
    Reshape/size failure at i=134.
    n=4 (class double), integer=1, finite=1
    m=1 (class double), integer=1, finite=1
    m*n=4, integer=1, finite=1
    size(Hres)=[4 4 1]
    size(X)=[1001 4], size(F)=[1001 1], size(Cres)=[1 1], size(D)=[1 4]
    Caught: MATLAB:getReshapeDims:notRealInt | Size arguments must be real
    integers.
    
    Error in test_identity_combine (line 53)
        [X, F, hf, flag, xk_best] = pounders(Ffun, X_0, n, nf_max, g_tol,
        delta_0, 1, Low, Upp, Prior, Options, Model);
    
    Error in Testpounders/shortTests (line 25)
                test_identity_combine;
================================================================================

where integer=1 means that round(x) == x. Since this debug information does not provide an explanation of the failure, Jeff converted all index arguments to those functions to int32(round(x)). Evaluating the action (of 10 jobs) approximately 5-6 times did not produce errors. The experts believe that the code should not be able to set n and m to non-integer values. In fact, these errors could be produced after many iterations of using n and m successfully.

A current workaround (PR #236) is to explicitly cast the user-provided n and m arguments at the top of pounders

n = int32(n);
m = int32(m);

I am assured that 32-bit indices are sufficient. Based on rerunning the action several times, this appears to have silenced this particular type of "bad index" error.

[jared321]

As part of PR #236, I reran the MATLAB test action several times in succession to get a sense of the effectiveness of the workaround made in that PR. While the index error that the workaround attempts to address did not appear, the following error did with the macOS-15/R2021a test setup

================================================================================
Error occurred in Testpounders/longTest and it did not run to completion.
    ---------
    Error ID:
    ---------
    'MATLAB:NonIntegerInput'
    --------------
    Error Details:
    --------------
    Error using zeros
    Size inputs must be integers.
    
    Error in g_dfovec_1d (line 641)
                g_temp = zeros(size(temp));
    
    Error in jacobian (line 43)
        [g_fvec, fvec] = g_dfovec_1d(g_t, t, ind, m, n, g_x, x, nprob);
    
    Error in calfun (line 155)
        [J, dummy] = jacobian(m, n, x, nprob);
    
    Error in benchmark_pounders>calfun_wrapper (line 119)
    [~, fvec, ~] = calfun(x, struct, probtype);
    
    Error in benchmark_pounders>@(x)calfun_wrapper(x,BenDFO,'smooth') (line 40)
        Ffun = @(x)calfun_wrapper(x, BenDFO, 'smooth');
    
    Error in pounders (line 436)
                F(nf, :) = Ffun(X(nf, :));
    
    Error in benchmark_pounders (line 77)
            [X, F, hF, flag, xk_best] = pounders(Ffun, X_0, n, nf_max, g_tol,
            delta_0, m, Low, Upp, [], Options);
    
    Error in Testpounders/longTest (line 31)
                benchmark_pounders;
================================================================================

Note that the issue appears to arise from the MATLAB function size providing unacceptable results to the MATLAB function zeros, which seems unlikely. Since those two functions are being used together in a sensible way, we suspect that this is a clear indication that the underlying source of these errors is not in POptUS codes, but rather likely in a technical issue involving running MATLAB through MATLAB actions on a runner. Matt and Jeff checked that in one extreme case

>> tmp = []
tmp =
     []
>> zeros(size(tmp))
ans =
     []

[jared321]

These errors still occur on occasion. My sense is that they only occur with R2021a. I will try moving our oldest MATLAB version tests to a slightly newer version to see if these errors disappear.

[jared321]

Still getting errors such as the following with R2021b

Error occurred in Testpounders/longTest and it did not run to completion.
    ---------
    Error ID:
    ---------
    'MATLAB:NonIntegerInput'
    --------------
    Error Details:
    --------------
    Error using zeros
    Size inputs must be integers.
    
    Error in g_dfovec_1d (line 383)
                g_s2 = zeros(size(s2));

[jared321]

The current action logs suggest that R2022a is running with Rosetta 2, but that R2026a is not. I had one of these R2022a jobs fail when running with Rosetta 2.

[jared321]

According to Mathworks,

Native Apple silicon MATLAB is available for MATLAB R2023b and up

I will try to use this version as the oldest MATLAB version tested by our actions.