Speed up a lookup process

Hi, Guys,

I have been trying to speed up a lookup process.

A is a array of 1 million numbers. B is a array of 15 thousands numbers (no duplicated numbers in B). I am trying to do is: for every number in A, find the index of same number in B and use that index to fetch a value in array C (which is of the same size as A).

The look up process stated above has to be repeated several thousands times since I have thousands of different array B to loop up, but A and C kept always same.

I used ismember function AND gpuArray, which gave me some improvement. I wonder if there still some space for improvement. Let me know you suggestions!

Here is some testing code that I used:

A=randperm(1e6);
A=gpuArray(A);
C=randperm(1e6);
C=gpuArray(C);

tic
for j=1:10 % e.g. 10 loops with different B (all integers) and C, but A (all integers) kept same in every loop
    
B=randperm(15*1e3);
B=gpuArray(B);



[~,Locb] = ismember(A,B);



Locb(Locb==0)=size(B,1)+1;%%% make all the zero in Locb to be a element that is indexable

C(size(B,1)+1,1)=0;% add zero in the end of the table;

D(j,1)=sum(C(Locb));



end
toc

YZ

[Jan]

Why do you use the 'rows' flag, if the inputs are column vectors?

The detail "use that index to fetch a value in array C (which is of the same size as A)" is not clear: "C" does not appear in your posted code.

"gave me some improvemen" - compared to what?

The values matter: Are they integer values? Is B e.g. randperm(15e3)? Then a much faster indexing approach is possible. Do A or B contain NaNs?

Please post a short meaningful example for the inputs.

[YANAN ZHU]

Hi Jan,

Thanks for your answer. I edit the posted code.

Remove 'rows' flag speed up the code a lot.

A and B does not contain NaNs, I would appreciate if you could share your much faster indexing approach.

Yanan