Is there any other function faster than sum(A,2) to get the sum of all rows?

a=zeros(6750);
h=zeros(6750,1);
tic;k=h-sum(a,2); toc

Elapsed time is 0.041132 seconds.

Is there any faster way than to use sum(A,2) to get the sums of the rows of a matrix? sum(A,2) seems to be slow for large matrices.

[Image Analyst]

1. What's your definition of large? How many rows and columns?
2. How long is it taking for those large matrices? How many seconds or minutes?
3. How much speed do you require? What is the max allowable time?
4. Do you have the Parallel Processing Toolbox?

[Nadatimuj]

@Image Analyst

Thanks for the queries.

1. In my case, at max it can be 10000x10000 matrix. I used zeros just as an example, however, my matrix will usually have non-zero values but typically it is very sparse. I did not use sparse notation though.
2. As you can see in the original post, for 6750x6750 matrix it takes 0.04 seconds here, on my computer it took 0.017 sec.
3. miliseconds can be acceptable, 0.001 sec or so..
4. Is it just an add-on? Then I can install it, but currently I don't.

[Image Analyst]

The Parallel Processing Toolbox is an add-on toolbox. You can ask for a free trial of it and see if the Parallel Processing Toolbox helps. I know some functions recognize if you have that toolbox and start parallel processes going on multiple CPU cores on your computer automatically.

[Walter Roberson]

For sufficiently large matrices, addition is automatically handled by high performance parallel libraries. Using parfor or parfeval() will not improve performance compared to the automatic multi-core work that is done.

Answer 1

[Walter Roberson]

Switching to columns can improve performance as it allows better use of hardware cache.

[Nadatimuj]

Thanks for the suggestion. Did you mean this? Doesn't look like to save too much time...

a=zeros(6750);
h=zeros(1,6750);
tic;k=h-sum(a,1); toc

Elapsed time is 0.033019 seconds.

[Walter Roberson]

@Nadatimuj

When I test on my system, the sum(a,1) version is 5 to 6 times faster than the sum(a,2) version

Answer 2

[MJFcoNaN]

In general, the native functions are optimized sufficiently.

In certain cases, there may be some better way but in other parts of code. For example,

a=sparse(zeros(6750));
h=sparse(zeros(6750,1));
tic;k=h-sum(a,2); toc

Elapsed time is 0.003234 seconds.

[Nadatimuj]

Thanks, yes that's a good idea. However, originally I didn't use the sparse notation, I will check if using sparse will impact other parts of the code. Otherwise sparse conversion just here can take some time.