Speed up indexing without using a loop
显示 更早的评论
Hi all,
I would like to know if there is a way to speed up indexing procedure in the following code:
N; % given nx1 matrix with some values
D = inf(n,1);
D(1,1) = 0;
M = zeros(n,1)
while sum(M)~=length(N)
D_adj = D;
M_index = find(M);
D_adj(M_index) = inf; % or any very large number
[~,index_m] = min(D_adj);
M(index_m) = 1;
FOR LOOP % another code for a FOR loop to update values in matrix D;
end
Essentially, for every iteration within WHILE loop, I update a corresponding element of matrix M (by allocation value of 1). My main problem and what slows down my code is the procedure for computing M_index, updating the corresponding elements in matrix D_adj (by allocation value of inf so that they are not taken into account when computing the min value of all elements in matrix D_adj that have zero value in matrix M), and finding the index (of the min element in D_adj) in matrix D. Do you have any suggestion on how I can improve these operations?
Thanks in advance!
4 个评论
James Tursa
2018-3-15
In addition to the code, what would be useful to us is a small example of inputs and desired outputs. That way we can more easily see what the end result should be, and maybe think of other ways to do it.
Igor Dakic
2018-3-15
编辑：per isakson
2018-3-15
Igor Dakic
2018-3-16
采纳的回答
Maybe logical indexing is faster:
N; % given nx1 matrix with some values
D = inf(n,1);
D(1,1) = 0;
M = false(n,1);
while ~all(M)
D_adj = D;
D_adj(M) = inf; % or any very large number
[~, index_m] = min(D_adj);
M(index_m) = true;
FOR LOOP % another code for a FOR loop to update values in matrix D;
end
Remember the golden rule of optimizing code: Use the profiler at first to find the bottleneck of the function. If you accelerate a part of the by a factor 2, but this part needs 2% of the processing time, the total speedup is 1% only.
I doubt, that this indexing is the bottleneck of your code. But logical indexing should be at least faster.
I guess, that the part hidden behind "FOR LOOP" can be improved such, that the new minimal value can be found by a cheaper method.
3 个评论
Igor Dakic
2018-3-16
编辑：Igor Dakic
2018-3-16
Posting the original code is useful. You create variables dynamically by a load without catching the inputs. This impedes the JIT acceleration and the debugging. I guess, that all data you load from the file in "N_v", "aa_indx1" and "A_indx"? Then:
Data = load('data.mat');
N_v = Data.N_v;
aa_indx1 = Data.aa_indx1;
A_indx = Data.A_indx;
... expand this on demand
Replace:
MM = [1:1:length(N_v)]';
by the nicer:
MM = (1:length(N_v)).';
This will not improve the runtime, but there is no need for the concatenation operator [] here.
ismembc is not supported in modern Matlab versions anymore. I assume that this part can be simplified:
aa= A_indx(aa_indx1{indx_v,1},:);
a = aa(~ismembc(aa(:,2),MM(M)),:);
Something like this:
a = aa(M(aa(:, 2)));
but this is a bold guess only. I cannot run your code currently.
This loop can be vectorized:
for i = 1:length(a(:,1))
d_v_indx_j = a(i,2);
if d_v(d_v_indx_j) > d_v(indx_v) + a(i,3)
d_v(d_v_indx_j) = d_v(indx_v) + a(i,3);
p_v(d_v_indx_j) = N_v(indx_v);
end
end
by:
c = a(:, 2); % do this before the loops
...
d = d_v(index_v) + a(:, 3);
m = d_v(c) > d;
d_v(m) = d(m);
p_v(m) = N_v(indx_v);
As said already: This is untested, but the idea should be clear.
Igor Dakic
2018-3-16
更多回答（0 个）
类别
在 帮助中心 和 File Exchange 中查找有关 Matrix Indexing 的更多信息
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
Translated by 
