Re: Yet Another Vectorization Problem

[Moo]

I just have some results; looks like cumulative sum is the way to go:

%%%%%%%
data = "" %make random data to test the moving average codes.

% original version.  feel free to preallocate on avgs1; it sped things up by about 5 second for me, 
% but still not as fast as the other two methods.
tic
avgs1=[];
for i=1:9001
    avgs1(i,:) = sum(data(i:(999+i),:),1)/1000;
end
t1=toc

% add new data, remove old data on each iteration.
tic
avgs2 = zeros(9001,20);
avgs2(1,:) = sum(data(1:1000,:),1)/1000;
for i=2:9001
    avgs2(i,:) = avgs2(i-1,:) + (data(999+i,:) - data(i-1,:))/1000;
end
t2=toc

% difference of cumulative sums
tic
avgs3 = zeros(9001,20);
avgs3(1,:) = sum(data(1:1000,:))/1000;
temp = cumsum(data,1)/1000;
avgs3(2:9001,:) = temp(1001:1e4,:) - temp(1:9000,:);
t3=toc

% Check if we have the same result as the original method by taking the norm of the difference.
norm1 = norm(avgs1-avgs2)
norm2 = norm(avgs1-avgs3)
%%%%%%%

I get the result:
t1 = 6.0183
t2 = 0.34102
t3 = 0.012001

norm1 = 6.3484e-013
norm2 = 2.8073e-014