Error in Parallel for Loops in MATLAB R2020a

I am running a user written (BEAR) toolbox where it performs at some stage a parallel loop and iterations for impulse response functions.

The toolbox works well on MATLAB R2018a, but when I got the R2020a, I got the following error when I run the same code:

Error: File: irfres.m Line: 203 Column: 10
Assigning to the for-loop variable 'jj' is not supported in parfor-loops. For more information, see Parallel for Loops in MATLAB, "Nested for-Loops:
Requirements and Limitations".

Error in bear_toolbox_main_code (line 519)
   [struct_irf_record D_record gamma_record]=irfres(beta_gibbs,sigma_gibbs,It,Bu,IRFperiods,n,m,p,k,signrestable,signresperiods);

Error in run (line 91)
evalin('caller', strcat(script, ';'));

Error in bear_4_2 (line 33)
run files/bear_toolbox_main_code;

I attach the irfres.m file.

I installed the Parallel computer toolbox but I dont think this made a difference. If I go back to the R2018a version then it works.

I feel there is a missing toolbox. I hope you folks can adivse.

Regards

[Walter Roberson]

I have not seen that error before, but the fix looks simple: rename jj to another variable such as jjj in lines 203 to 219.

[Yacoub Sleibi]

Walter, many thanks. Your fix did the trick. All works good now.

I wonder why it only happened in the newst verion of matlab and not the old.

Thank you again :)

Answer 1

[Raymond Norris]

Hi Yacoub,

Here's what's happening (has nothing to do with missing toolboxes, etc.). You've started with something similar to

for idx = 1:10
  for jidx = 1:10
    rand;
  end
  jidx = 1;
  while jidx<10
      rand;
      jidx = jidx+1;
  end
end

You rewritten the outer loop as a parfor

parfor idx = 1:10
  for jidx = 1:10
    rand;
  end
  jidx = 1;
  while jidx<10
      rand;
      jidx = jidx+1;
  end
end

Which throws:

Assigning to the for-loop variable 'jidx' is not supported in parfor-loops. For
more information, see Parallel for Loops in MATLAB, "Nested for-Loops:
Requirements and Limitations".

Per the mentioned documentation:

Required (static): The index variable for the nested for-loop must never be explicitly assigned other than by its for statement.

This restriction is required, because changing the nested for-loop variable in the loop body cannot guarantee that the region indexed by the for-loop variable is available at each worker.

The two options are to use another while loop variable (other than jj in your case) or refactor your code, similar to the following

parfor idx = 1:10
  unit_of_work;
end

function jidx = unit_of_work
for idx = 1:10
  for jidx = 1:10
    rand;
  end
  jidx = 1;
  while jidx<10
      rand;
      jidx = jidx+1;
  end
end

Raymond

[Walter Roberson]

This restriction is required, because changing the nested for-loop variable in the loop body

However, they are not changing the nested for-loop variable in the loop body (of the nested for loop).

[Yacoub Sleibi]

Thank you Raymond, Walter suggestion worked well.

Best

[Raymond Norris]

Hi Walter,

Is your comment that this doesn't align with Yacoub's code or that it shouldn't matter? I believe the loop variable being referenced is the parfor loop and not the for loop.

With that said, looking closer at the doc, the example shown is modifying the loop variable within the for loop, not the parfor loop.

I'll chat with Development and post back.

Raymond

[Walter Roberson]

The restriction of not assigning to a for loop index makes sense within the body of the for loop in the case where the for loop index is being used to index one of the variables being communicated with the client.

The restriction could also make sense if the for loop index value was being used after the loop to index a variable being communicated with the for loop variable not being modified in between.

In any case where the for loop index is only being used as a counter or as an index to a local variable, then restriction would not make sense.

There is a remaining case: a for loop variable used to index a communicated variable, then modified, then used to index a communicated variable, but the modification assigns a definite value such as 1, after the original loop. In such a case, it would be possible to track that the index is valid, and would be a Quality of Implementation to not track. Modifying to a definite value inside the for loop is more work as the for loop body might never be run in some cases.

[Raymond Norris]

Walter,

Does this capture your last example:

M = 10;
A = rand(M);
parfor idx = 1:M
   for jidx = 1:M
      a = A(idx,jidx);
   end
   
   jidx = 1;
   b = a(jidx);
   
   for jidx = 1:M
       c = A(idx,jidx)
   end
end

This would have worked in R2019a.

Raymond