Fast way to delete the last n rows of a matrix

Hi,

I would like to be able to delete the last n rows of a matrix. I have found that this can be done in the following way. For example, suppose I want to delete the last two rows of a matrix:

tst=[1 1 1; 2 2 2; 3 3 3];
lastn=2;
tic
tst((end-(lastn-1)):end,:)=[];
toc

This works, but it seems quite expensive: "Elapsed time is 0.001152 seconds." In my code, I will be deleting rows many hundreds of thousands of times. Do you know if there is a faster way to delete the last n rows of a matrix?

Thank you.

Andrew DeYoung

Carnegie Mellon University

Answer 1

[Sean de Wolski]

tst = tst(1:end-lastn,:); %glass half full

No guarantees it'll be faster. The typical goal here is just minimize the number of matrix resizing operations you have to do.

[Andrew]

Your way is much faster: "Elapsed time is 0.000032 seconds."

Compared to my way: "Elapsed time is 0.001152 seconds"

This is interesting. So it is faster to "make a new matrix" than to "resize an existing matrix," it seems.

[Hafsa Arif]

is this making a new one? informmation will be lost or get arranged in the matrix?

[Walter Roberson]

Newtst = tst(1:end-lastn,:); %glass half full

If you want to keep the old matrix. With Sean's version to write to tst then the last rows would be thrown away.

Answer 2

[Walter Roberson]

Andrew, could you transpose so that you are deleting columns instead of rows? If you do that then the portion that is left behind is contiguous in memory, and it should be faster to work with that than to have to copy over each partial column as you delete rows.

[Andrew]

Thanks. Yes, I should be able to do that. So should I understand that in Matlab rows and columns are not created equal?

[Walter Roberson]

Correct. MATLAB arranges memory in columns and the first entry for the next column is stored in memory directly after the last entry for this column. Therefor if you are keeping entire columns then the block of memory that needs to be copied to the new array is arranged sequentially in memory and can just be described as "starting address" and "number of items to copy". Especially for larger blocks, there are often ways to optimize that at the machine level, such as by copying full system "pages" of memory using DMA.

[Andrew]

Thank you so much. I really appreciate your taking the time to explain this to me. I will be able to use in the future what you have taught me.