Very slow for loop

Hi, I am new to Matlab, i have more experience with C/C++

I am trying run this simple loop:

x=zeros([10001,10001,2]);
for t=1:2
for z =1:10000
for i =1:10000
  x(z,i,t)=i+z+t+2;
end 
end
end
x(5,5,1)

But is taking a while..imagine many of then inside a single algorithm. in C++ this loop runs in a few seconds...

Is there any other way to run this loop in Matlab faster???

Sorry if my question is silly, first time running Matlab. I tried some vectorization but for nested loops things get really complicated, especially with different loop sizes.

Thanks

[Andrew Newell]

I'm wondering whether you have forgotten that indexing starts with 1, not 0, in MATLAB. Are you really trying to calculate the following?

x=zeros([10001,10001,2]);

for t=1:2

for z =2:10001

for i =2:10001

x(z,i,t)=i+z+t;

end

end

end

x(5,5,1)

[Daniel Shub]

Why should that be any faster/slower than 1:10000? The variable x is already intialized to a size that handles either case.

[Andrew Newell]

The point of my question is that the values in each row and column increase until they suddenly drop to zero at the end. I'm guessing that is not what Rafael intended.

Answer 1

[Daniel Shub]

You might be able to get a slight performance boost by changing how the memory is acessed ...

http://undocumentedmatlab.com/blog/matrix-processing-performance/

Basically x(z,i,t) might not be the same as x(t,z,i) or x(i,z,t).

[Rafael]

Hi Daniel, Thanks!!! Indeed improved the permornce by 2-4 seconds but is still very slow, almost 25 seconds, on my machine (not a powerful one though). Running it using C++ on GCC 4.6 (with architecture and performance flags optimized) it runs in 3 seconds. The difference is huge. Thanks anyway and happy new year (not yet in Brazil - almost 5 hours to go)!!!!

Answer 2

[the cyclist]

x = bsxfun(@plus,bsxfun(@plus,1:10000,(1:10000)'),permute(1:2,[1 3 2]))+2;

[Rafael]

Thanks!, indeed much faster!

Answer 3

[Andrew Newell]

For a problem where the numbers count upward monotonically in each direction, here is a compact and fast solution:

y = repmat(2:10002,10001,1);
x = cat(3,y+y',y+y'+1);

On my computer, your code takes about 25 seconds and mine takes 6 seconds.

[Rafael]

Hi, thanks.

As I said before, according to your feedback I have to avoid for loops in my code. Will be hard to translate C++ codes for Matlab if this is true.

[Andrew Newell]

I think that you mostly can leave loops in because MATLAB uses just-in-time compiling on loops. However, as far as I know this compiling isn't done on arrays of dimension greater than 2.

Answer 4

[Jan]

If the shown code is your original problem, and not a simplification, I'm in doubt, that it is an efficient approach: You occupy 1.6GB memory with values, which are very easy to calculate dynamically.

Note: The reference "x(z,i,t)" requires two multiplications, while "i+z+t+2" uses 3 additions only.

[Rafael]

You are right. The intention is to make a n-by-n-2 array.