Array indexing question for vectorization

Question

Richard 2012-11-13

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/53603-array-indexing-question-for-vectorization

I have three arrays A, B, and IND, with A containing some values that I want to update, B containing the values that I want to use to update A, and IND containing the indeces to update in A. IND may reference the same index in A several times. For example, I may have:

A = ones(1, 10);
B = 1:10;
IND = [1 1 2 5 5 5 5 6 8 9];

and the update I want to do is multiplication. I can achieve this by doing

for ii = 1:10
  A(IND(ii)) = A(IND(ii)) * B(ii);
end

which gives me my desired answer of

A =
   2     3     1     1   840     8     1     9    10     1

Is there a way I can do this in a vectorized operation, avoiding the for loop (I'm potentially doing this kind of operation thousands of times on large arrays)? Doing

A(IND) = A(IND) .* B

results in

A =
   2     3     1     1     7     8     1     9    10     1

Any tips greatly appreciated!

2 个评论
显示无隐藏无

Matt Fig 2012-11-13

在 MATLAB Online 中打开

Are A and B always like that? Could they be more like:

A = randi(N,1,M);
B = randi(K,1,J);

or not?

Richard 2012-11-13

Hi Matt,

My initial problem specification was indeed lacking - see my reply to Jan below for a more accurate representation of the problem...

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Honglei Chen 2012-11-13

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/53603-array-indexing-question-for-vectorization#answer_65227

编辑：Honglei Chen 2012-11-13

在 MATLAB Online 中打开

You can try the following trick

c = unique(IND)
d = accumarray(IND',B',[],@prod)
A(c) = A(c).*d(c)'

6 个评论
显示 4更早的评论隐藏 4更早的评论

José-Luis 2012-11-13

True, but it still is slower than a for loop:

Elapsed time is 0.000003 seconds. %for

Elapsed time is 0.000828 seconds. %unique+accumarray

Elapsed time is 0.000515 seconds. %accumarray

Richard 2012-11-13

Interesting, thanks. I'm still in the mindset of "for-loop bad", and aiming to replace them without necessarily thinking of the built-in function overheads...

请先登录，再进行评论。

Answer 2

Jan 2012-11-13

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/53603-array-indexing-question-for-vectorization#answer_65233

编辑：Jan 2012-11-13

在 MATLAB Online 中打开

Is your A pre-allocated? If IND is sorted, running the loop backwards wil be faster:

for ii = 10:-1:1
  A(IND(ii)) = A(IND(ii)) * B(ii);
end

If IND is not sorted, pre-allocate explicitly:

A = zeros(1, max(IND));

For more speed create a C-Mex function. If you are interested and have a compiler installed, I can post the few required lines of code.

2 个评论
显示无隐藏无

Richard 2012-11-13

在 MATLAB Online 中打开

Hi Jan,

Yes, A is preallocated. Actually the problem is a bit more complicated than I posted (note to self: make sure your example code is suitable!): in my actual application, A is a 2D matrix, and B is a 1D array with fewer elements than A, so the solution I've accepted doesn't quite work... e.g.

% A more accurate example...
A = zeros(20000, 10);
indR = randi(20000, 1, 500);         % row indeces
indC = randi(10, 1, 500);            % column indeces
IND = sub2ind(size(A1), indR, indC);
B = rand(1, 500);
tic
for ii = 1:length(IND)
  A(IND(ii)) = A(IND(ii)) * B(ii);
end
toc

(in this example with random indeces I think the instances of duplicate values in IND are quite low compared with the actual program). Thinking about it, my IND values could be sorted if I reorganised my data a bit prior to this (which may be worth the reorganisation overhead as this bit of the program is running in a simulation loop over many time steps)...

Thanks for the useful discussion everyone.