How to make variable visible to parfor?

I have been having trouble getting the following code to give me the correct result from the parfor in it. Briefly, the code is supposed to use input parameter structure PMTS as input to many other functions called inside the parfor loop, and return a single double value at the end, FI, that is assigned to the I-th element of the UNQF vector.

function     RSLT = MANAGENAPART(PMTS)
  function PSTR = RETURNPARSTR(PMTS)
      POPL = PMTS.POPL;
      UNIQ = PMTS.UNIQ;
      GDTA = PMTS.GDTA;
      NVBS = PMTS.NVBS;
      SHPE = PMTS.SHPE;
      PSTR.POPL = POPL;
      PSTR.UNIQ = UNIQ;
      PSTR.GDTA = GDTA;
      PSTR.NVBS = NVBS;
      PSTR.SHPE = SHPE;
      PAZ = size(POPL);
      PSTR.MBRS=PAZ(1);
      PSTR.SLEN=PAZ(2);
  end
  AA = @ADDTOARCHIVE;
  CB = @COMBINEBSISV;
  GS = @GNRTESAFESHP;
  DS = @DECODSHPESTG;
  NF = @CALCNASFITNS;
  RP = @RETURNPARSTR;
  PA =      RP(PMTS);
  M  =       PA.MBRS;
  UNQF  = zeros(M,1);
  parfor (I=1:M)
      PAR =RP(PMTS);
      UNIQ=PAR.UNIQ;
      POPL=PAR.POPL;
      GDTA=PAR.GDTA;
      NVBS=PAR.NVBS;
      SHPE=PAR.SHPE;
      SLEN=PAR.SLEN;
      ISUQ=UNIQ(I);
      CD=GDTA.CD;
      GD=GDTA;
      SH=SHPE;
      P =POPL;
      S =SLEN;
      if ISUQ
          G=P(I,1:S);
          PAR.STRG=G;
          DC=DS(PAR);
          CX = DC.DX;
          CY = DC.DY;
          CZ = DC.DZ;
          SH.CFTX=CX;
          SH.CFTY=CY;
          SH.CFTZ=CZ;
          LC =CB(SH);
          SH.LINC=LC;
          SH.PFCD=CD;
          [~]=GS(SH);
          FI =NF(GD);
          UNQF(I)=FI;
      end
  end
  PATS.POPL=POPL;
  PATS.FTSV=UNQF;
  PATS.UNIQ=UNIQ;
  DONE =AA(PATS);
  RSLT.FUNQ=UNQF;
  RSLT.OKAY=DONE;
end

My work is presently on a cluster of 4 labs. I suspect the PMTS parameter is not visible inside the parfor loop, but as a parfor loop is non- debuggable, I am not sure of the same. However when I run the above code with a 'local' configuration (that is only 1 worker, the host itself), the parfor result seems to work correctly! Can anyone help?

[Edric Ellis]

What makes you thing PMTS is not 'visible'? Do you get an error? Wrong answer? Any chance you could reduce your code down to something which demonstrates the problem, but without the external dependencies? Does it make any difference if you change RETURNPARSTR to be an internal function (i.e. defined after then end of MANAGENAPART)?

[B.V. Vijay]

Thanks, Ellis. Actually why I suspect that PMTS is not visible inside the parlor loop is because the code gives the same (exactly the same) result for all cases of I from 1 thru M, when it should be giving M different results. I have not yet tried the internal function yet and I will try that out soon.

[Konrad Malkowski]

PMTS is a broadcast variable in this instance, so every worker and every iteration of the PARFOR receives an exactly the same copy of PMTS. Is that what you intended?

There also appears to be an issue with POPL and UNIQ variables. They are declared as "temporary" inside of the PARFOR, and are used to initialize PATS outside of PARFOR. This will work in a for loop, but not in a PARFOR. You need to remember that everything inside of PARFOR exists in a different MATLAB workspace, than the code outside of PARFOR.

Answer 1

[B.V. Vijay]

Ok, the above code works as it is, correctly. Thanks to you, Ellis and Malkowski, for your observations. I managed to track the bug to another section of my code outside the above, and had that corrected a couple of days back.

Ellis- when I tried moving the function RETURNPARSTR to the end of the code as you suggested, the MATLAB editor immediately produced lots of warnings (variables getting underlined). As these warnings had never appeared in the code form as seen above, I didn't change the code structure from the above.

Malkowski- regarding the POPL and UNIQ variables, I verified (using the debugger with a breakpoint on the PATS.POPL=POPL; statement) that their values were the same as those from replacing the parfor (I=1:M) by a for I=1:M loop.

Thanks, again. Nevertheless hope we can see some more documentation in future on these issues with the parfor function.