"james bejon" wrote in message <email@example.com>... > I must admit, I've always found accumarray to be remarkably fast, e.g., > > % DATA > subs = ceil(rand(1000000,1)*10); > > % ENGINE > tic > for i = 1:1000 > accumarray(subs, 1); > end > toc > > % Result = 1.40 seconds > > Is there a specific usage you have in mind? ================
I'm trying to find a way to implement a sparse matrix multiplication given I,J,S data while avoiding the overhead of actually building the sparse matrix and all of the sorting operations it requires. I have to do this for repeated sets of I,J,S data which are fairly easy/quick to generate, but if I have to load them into a sparse matrix each time, it will kill the speed.
I guess one of the reasons for my suspicions about accumarray performance is the speed differences I'm seeing between an accumarray operation and an equivalent sparse matrix multiplication in the example below. The accumarray approach is about twice as slow as the sparse matrix approach even though they both have the same data to work with. In fact, the sparse matrix multiplication is probably even handicapped relative to accumarray, since it has to search through x for the appropriate x(J), whereas in the accumarray approach, I exclude the construction of b=x(J) from the tic/toc.
%fake data setup M=1e5; A=kron(speye(M),ones(1,16)); N=size(A,2); [I,J]=find(A); x=rand(N,1);
%pretend we build A from scratch tic; A=sparse(I,J,1); toc %Elapsed time is 0.062737 seconds.
%Apply A tic y1=A*x; toc %Elapsed time is 0.006868 seconds.
%Using accumarray b=x(J); tic y2=accumarray(I,b,[M,1]); toc %Elapsed time is 0.012236 seconds.