Sliced Variables in "parfor"

Question

Reza Teimoori 2016-12-31

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/318660-sliced-variables-in-parfor

评论： Matt J 2016-12-31

I'd like to run a loop using parfor similar to this:

A = zeros(10,10,10,10);
parfor i = 1:100
     ind = ceil(rand(1,4)*10);
     A(ind) = A(ind) + 1;
end

This yields to an error due to the way indices of A is being defined in each iteration. I understand that one workaround can be to save the "ind" values in each iteration in a temporary variable and use them in a later regular loop to update the matrix A. But I was wondering if there is any way that lets me do the whole thing in a single parfor loop. In my case the matrices and indices that I'll be working with are so large that I can't save them for use in a later loop.

Thank you all!

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

Matt J 2016-12-31

在 MATLAB Online 中打开

Are you sure you didn't really mean as follows?

    parfor i = 1:100
         ind = ceil(rand(1,4)*10);
         i=ind(1);
         j=ind(2);
         k=ind(3);
         l=ind(4);
         A(i,j,k,l) = A(i,j,k,l) + 1;
    end

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Matt J 2016-12-31

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/318660-sliced-variables-in-parfor#answer_249058

在 MATLAB Online 中打开

Suppose you wanted to do this 5 million times:

    A = zeros(10,10,10,10);
    M=5;
    N=1e6;
    parfor i = 1:M
         subs = ceil(rand(N,4)*10);
         A=A+accumarray(subs,1,size(A));
    end

Obviously there are different possible choices of partitioning parameters M,N. You want to make N as large as your RAM reasonably allows because that will use the most vectorization.