Difference between a 3D mean function vs a for loop; computational time for summing over a dimension?

Hi guys,

I have a pretty big 3 dimensional data (subject_img_3D_PC_cut variable; dimension is around 41*1million*35) and been running some averaging on this data. I have run two different methods of averaging, one using a 3D nanmean and the other using a 2D nanmean; and the computational times for the two different methods seem to diverge a lot especially when the data is much greater. What mechanism of the nanmean function (basically isnan function + sum function) creates the difference between 2d and 3d averaging?

heres my code:

 % using nanmean 
tic
grpavg = nanmean(subject_img_3D_PC_cut,3);
time_elapsed_nanmean = toc

 %using for loop
tic
grpavg = nan(size(subject_img_3D_PC_cut,1),size(subject_img_3D_PC_cut,2),1);
for jk = 1:size(subject_img_3D_PC_cut,2)
    grpavg(:,jk,1) = nanmean(subject_img_3D_PC_cut(:,jk,:),3);
end
time_elapsed_forloop = toc

Results for the smaller data

 %results for 41*902629*25 data. 
time_elapsed_nanmean =

    4.8007
time_elapsed_forloop =

    8.7792

and the results for the bigger data:

 %results for 41*902629*35 data.
time_elapsed_nanmean =

237.1386
time_elapsed_forloop =

   12.9893

[Michelle Wu]

You could use the MATLAB Profiler to see a breakdown of the execution time for each function when running your code. Please refer to the following doc for more information on how to use the Profiler:

https://www.mathworks.com/help/matlab/ref/profile.html