The first column of the subs matrix is the row indices of the elements you want to create in the result matrices. The second column is the column indices. The third column is the page indices (dimension 3.) In general, the nth column of the subs input contains the indices in the nth dimension of the result matrix.

For instance, if one of the rows of the SUBS matrix is [1 2 3 4 5], then the entry in the corresponding row of VAL will be “accumulated” into element A(1, 2, 3, 4, 5) of the output matrix A of your call to ACCUMARRAY. Try:

A = accumarray([1 2 3;4 5 6], [1;2]);
A(1, 2, 3)
A(4, 5, 6)

Sorry, I don’t know of any others. I recommend you pull apart the example(s) that you find confusing into smaller pieces and make sure you understand each piece. :ooking at each piece separately should be helpful.

–Loren

In MATLAB 7.4, if you use the magic word

spparms(‘chmodsp’,1)

and then try full(sparse( …)), then you’ll find that accumary and full(sparse) are about the same speed. That’s the case for the C=accumarray(…) above, even when iters is 100,000 for this example.

Faster than any of these methods, including the sub2ind method, is to avoid assignment with indices altogether, and do this:

n = 1000 ; k = 50 ; iters = 1000 ;
X = zeros (iters,k) ;
Y = zeros (iters,k) ;
for i=1:iters
X (i,:) = randsample (n,k,true) ;
Y (i,:) = randsample (n,k,true) ;
end
C = accumarray ([X(:) Y(:)], 1, [n n]) ;

On my machine, the for i=1:1000 code above using sparse inside the for loop takes 21.5 seconds, your sub2ind code takes 1.3 seconds, and the above code takes 0.2 seconds. The above code also uses far less memory than your sub2ind code, too.

I find that it’s often faster to avoid fancy indexing, and do as many whole-matrix operations as possible (such as the above accumarray).

I agree that MatLab’s documentation does adequately explain the convention for mapping a vector of indices into a variable of higher dimensions. However, some people who read my MatLab code may not have read that section of MatLab documentation, so it won’t be clear to them how this works. That is why I avoid this technique. As Dykstra said, “. . . it is not only the programmer’s task to produce a correct program but also to demonstrate its correctness in a convincing manner . . . “

My problem can be simulated by:

n=1000;k=50;
a=zeros(n,n);
for i=1:1000
x=randsample(n,k,true);
y=randsample(n,k,true);
% a=a+accumarray([x,y],1,[n n]);
a=a+sparse(x,y,1,n,n);

end

using accumarray cputime is ~34s
using sparse it is ~21s

adding another index to matrix a to deal with repetitions
I could get the code running 10 times faster:

n=1000;k=50;
a=zeros(n,n,k);
for i=1:1000
x=randsample(n,k,true);
y=randsample(n,k,true);
l=sub2ind([n n k],x’,y’,1:k);
a(l)=a(l)+1;
end
a=sum(a,3);

it runs in 2s.
My real problem has loop closer to 100,000 so time savings
are significant.

in which case you would simply do this

a=magic(5);
loc=[1 1 3 3 1;1 3 1 3 1];
data=pi*(1:size(loc,2));
as=a+accumarray(loc.’,data,size(a));
a
as
(as-a)./pi

us

loc=[1 1 3 3 1;1 3 1 3 1];
a=a+sparse(loc(1,:),loc(2,:),data,3,3);

Piotr