accelerate 3D matrix multiptication using bsxfun

Hi

I have athe following code

it takes forever to run, and disable my CPU in the prosess...

I'm looking for a way to make it run better (GPU is also an option)

norm_mode is 256x256 double matrix (X,Y) image

fields is 16384x15 complex double matrix (time,mode)

in the end' I need a 3D matrix take each norm_mode image will also be a function for time and suming all the modes- 3D matrix 256x256x16384

that is the total_field

do you have any sugestions on how to make it run better?

Thanks,

Barak

function total_field = BuildSpatialField(fields,fiber, sim, others)

total_field = zeros( newSize, newSize, length(others.t) ); % total_field(X,Y,t)

h = waitbar(0, 'calculate field...');

    for ii=1:others.modes
        
        norm_mode = % load new phi

        tmp = bsxfun(@times, norm_mode, reshape(fields(:,ii),1,1,[]));

        total_field = total_field + tmp;

        waitbar(ii/others.modes, h, ['mode ' num2str(ii) ' from ' num2str(others.modes)]);


    end
    close(h);

end

Answer 1

[Viktor]

256x256x16384 as double array needs around 8 Gb RAM and as complex double 16 Gb. Do you have that much memory? Matlab can take forever to tell you, it ran out of memory.

[barak messica]

I have 32GB RAM.

in the end it runs, but it take someting like 20 min

that is wat I am looking for more efficient way to do that. now it is pretty much brute force

[Viktor]

As i understand bsxfun(@times, norm_mode, fields), it is the same as: norm_mode .* fields.

To improve the loop you can do several things.

- bring "fields" into the right shape before entering the for loop.

- use the keyword parfor instead of for

- use gpuArray(X)

- waitbar can slow down your loop significantly

[Stephen23]

"matlab prefers 1xn vectors rather than 1x1xn or nx1"

Where is this preference documented?

[Viktor]

Thanks for the hint.

There is no performance difference accessing single elements in 1xn, nx1 or 1x1xn vectors.

I think i came to the conclusion since matlab uses 1xn by default.

But i found out that extracting a column vector from a matrix is faster than extracting a row vector.

https://de.mathworks.com/matlabcentral/answers/216512-which-is-faster-a-row-vector-or-a-column-vector-can-anyone-answer-me-please#answer_178976

[Walter Roberson]

@Viktor

As i understand bsxfun(@times, norm_mode, fields), it is the same as: norm_mode .* fields.

Not exactly

• bxsfun is not supported for as many datatypes as implicit expansion is supported
• historically it has been mixed as to whether bxsfun or implicit expansion is more efficient; there have been cases where bxsfun is notably more efficient, and there have been cases where implicit expansion is a bit more efficient.

So, the internal implementation is different between the two cases.

Mathworks is not putting effort into optimizing bxsfun, but is putting effort into optimizing implicit expansion.