How to speedup mean and std calculation on GPU?

Hello everyone, I am looking a way to speed up mean and std calculation on GPU. I run this code and it does take quite some time to complete, compared to the one if I do not use gpuArray. Maybe somebody would have any idea?

 g_p is gpuArray with matrix of (1000000,5)
   for q=1:n1-d
    x2=g_p(d-w+q-1:d+q-2,:);
mean_x=mean(x2);
    std_x=std(x2);
     R = bsxfun(@minus,x2,mean_x);
    x3=bsxfun(@rdivide,R,std_x)
  end
  
///////////
or x3=arrayfun(@norm,x2)?

Answer 1

[Jan]

To calculate the standard deviation, the mean must be calculated again. Try to combine this:

x2     = g_p(d-w+q-1:d+q-2,:);
mean_x = sum(x2, 1) / w;
xc     = x2 - mean_x;                % Auto-expand: >= R2016b
% xc   = bsxfun(@minus, x2, mean_x);
std_x  = vecnorm(xc) / sqrt(s - 1);  % vecnorm: >= R2017b
% std_x  = sqrt(sum(xc .* xc, 1)) / sqrt(s - 1);

for the mean only the first and the last element changed between the iterations. Use this detail:

mean_x = sum(g_p(d-w:d-1, :) / w;  % For q=1
for q = 1:n1-d
   ...
   mean_x = mean_x - (g_p(d-w+q-1, :) + g_p(d+q-1, :)) / w;
end

[Jan]

[MOVED from section for answers] Mantas Vaitonis wrote:

Dear Jan, Thank You for the answer, unfortunately I use R2016b and vecnorm is not supported. However, did modify code according to your example, but at the end it turned out to be no faster than predefined functions mean and std.

[Jan]

Without vecnorm you can use the line posted afterwards:

std_x  = sqrt(sum(xc .* xc, 1)) / sqrt(s - 1);

I cannot test the code on a GPU. Maybe my suggestion give you at least an impression, of what could be tried to reduce the overhead.

[Mantas Vaitonis]

Yes thank You,this seem to do the trick, plus I did use cellfun and it managed to speedup even more.