Save values from a parfor-loop

Hello,

I've got a question regarding my little parfor-example.

Y=rand(100,3) 
matlabpool open 
parfor p=1:2 
        for q=1:2 

                 for e=1:3     
                 Z=sum(Y(:,e)+p*q)   
                 end 

          end 
  end 
  matlabpool close

For every combination of p and q there are 3 different values of Z which I would like to save within a matrix. For example in this manner:

   1       1       1       1        1        1       
   1       1       1       2        2        2      . . . 
  Z(1)     Z(2)   Z(3)    Z(4)    Z(5)     Z(6)

If it was not a parfor-loop but a simple for-loop, I would try something ike this:

 Y=rand(100,3) 
  z=1 
  for p=1:2 
          for q=1:2 

                 for e=1:3     
                 Z=sum(Y(:,e)+p*q)   
                 end 

                 ergebnis(1,z)=p 
                 ergebnis(2,z)=q 
                 ergebnis(3,z)=Z 
                 z=z+1 

          end 
  end

But of course that's not possible when working with parfor-loops.

It would be great, if someone helped me out.

Thank's a lot

Domninik

[dominik ballreich]

It would be also ok for me, if it was possible just to save only the Z(1)...Z(12) in a row-vector. Without the p and q.

[Alex]

As a recommendation, if you know the size of a matrix beforehand, always initialize it before the loop. Like in your example, ergebnis = zeros(3,12), should be placed before your second loop example. This prevents memory loops that increase in size.

Answer 1

[Alex]

It looks like parfor sends one variable to each parallel operation, meaning that the variable references must be clear to begin with.

I.e. Z(i) = #, Z(i,j) = #, and Z(i,j,k) = # are valid but Z(i + j) is not valid.

So, that leaves you with two choices, that I know of.

1. Calculate Z as a 2x2x3 matrix

   Y=rand(100,3) 
  Z = zeros([2,2,3]);
  matlabpool open 
  parfor p=1:2 
          for q=1:2 

                   for e=1:3     
                   Z=sum(Y(:,e)+p*q)   
                   end 
            end 
    end 
    matlabpool close

Now, you have a 2x2x3 matrix that you would need to deference into a vector using tools such as repmat or loops.

2. calculate z as a vector

The other option is to have a single Parallel loop and d3ereference the index's that you want within the loop. An example follows.

Y =rand(10,3);
Z = zeros(1,12);

parfor i = 1:12
                    l = mod(i,3);
                    if(l == 0);
                        l = 3;
                    end
                    j = floor( ( i -l) / 6);
                    k = floor( ( i - l - j) / 3);

                      Z( i) = sum(Y(:,l)+j*k)
  end 

    matlabpool close

This leaves you with your desired Z vector. However, the inner loop calculations are messier.

Answer 2

[dominik ballreich]

Hallo Alex,

thank you very much for your help. I tried to implement the idea of your first example into my problem. intervall=zeros([length(kaufsignal),length(verkaufsignal),length(massmooth),anzahl_bootstraps]) <-would be Z = zeros([2,2,3]);

intervall=zeros([length(kaufsignal),length(verkaufsignal),length(massmooth),anzahl_bootstraps]); matlabpool open

    parfor p=1:length(kaufsignal)
        for q=1:length(verkaufsignal)
            for u=1:length(massmooth)

                %Bootstrap-Sektion

                for e=1:anzahl_bootstraps
                    x=verkaufsignal(1,p);y=verkaufsignal(1,q);z=massmooth(1,u);

                    R=Rsi(Y(:,e), z);
                    R=R';

                    vermoegen=zeros(long,1);
                    drin=3;
                    vermoegen(1)=kapital;

                    for i=2:long
                        if R(i)>x && R(i-1)<=x &&drin~=1;
                            drin=1;
                            vermoegen(i)=vermoegen(i-1)*t_k;
                        elseif R(i)<y && R(i-1)>=y &&drin~=0;
                            drin=0;
                            vermoegen(i)=vermoegen(i-1)*t_k;
                        elseif drin==1
                            vermoegen(i)=vermoegen(i-1)*(Y(i,e)/Y(i-1,e));
                        elseif drin==0
                            vermoegen(i)=vermoegen(i-1)*(1+(1-(Y(i,e)/Y(i-1,e))));
                        else vermoegen(i)=vermoegen(i-1);
                        end
                    end

                    Renditen_Vermoegen=[0;log(vermoegen(2+einpendeln:end)./vermoegen(1+einpendeln:end-1))];

                    %Performancemessung
                    mittelwert=mean(Renditen_Vermoegen.*weight);

                    intervall(p,q,u,e)=mittelwert;

                end
            end
        end
    end
    matlabpool close

But Matlab tells me "The variable intervall in a parfor cannot be classified." (intervall)

[dominik ballreich]

I know that the code looks weird, but in my view it's similar to the little example. Running the code without the line

"intervall(p,q,u,e)=mittelwert"

works perfectly. So the variable "mittelwert" gets overwritten on each part of the loop. So the only really problem is to save these values.

All the best

Dominik

[Alex]

I don't know what the issue is. The following works for me.

Y=rand(100,3)

Z = zeros([2,2,3]);

matlabpool open

parfor p=1:2

for q=1:2

for e=1:3

Z(p,q,e)=sum(Y(:,e)+p*q)

end

end

end

matlabpool close