Time for Matrix Initializing

Hi!

I create a cell of 700 elements, each of which is a sparse matrix 2000 * 2000 (most of the values will be zero). When I initialize to zero:

for iCell = 1:700 P{iCell } = sparse(zeros(2000, 2000)); end

It takes a non negligible time compared to all the other calculations my program makes in the next lines. Is it normal, given I only create (admittedly big) empty matrices?

Is there a little trick I could use decease the computation time? I tried what is written here: http://stackoverflow.com/questions/14169222/faster-way-to-initilize-arrays-via-empty-matrix-multiplication-matlab

but it didn't work.

Thanks in advance!

Answer 1

[Jan]

Converting the full matrix zeros(2000, 2000) to a sparse matrix wastes time, because the full matrix is created explicitly in each iteration. This would be faster:

P = cell(1, 700);  % Pre-allocate the cell array
for iCell = 1:700
  P{iCell} = sparse(2000, 2000);
end

Even better you'd define the sparse matrices with the correct number of elements directly, e.g.:

sparse([],[],[], 2000, 2000, 971)

A nicer method, which is unfortunately not faster in the final program, is:

P = cell(1, 700);  % Pre-allocate the cell array
P(:) = {sparse(2000, 2000)};

Now the cell elements are shared data copies of the same array. This is faster to initialize, but accessing the matrices will create a deep copy during the program runs, such that the total run-time is longer.

[Martin]

Great Jan, it changes everything. So Matlab had to check every value of zeros(2000, 2000) before making it sparse?

Thanks again.

Martin