extracting sub matrix - time consuming

Hi, I need to perform a certain calculation on a sub matrix, then insert this sub matrix to another matrix. I'm using : to select the submatrix, problem is, it's very time consuming. After running the profiler i see that this two lines take 20+ seconds each (the lines run 1170 time). With 361*4*18 times this code should run I'm looking at several days of computations.

[x,y] = ndgrid(yup:ydown, xleft:xright);
  exponent = (xpower(yup:ydown,xleft:xright).*ypower(yup:ydown,xleft:xright)).*exp(-(xc^2 + yc^2 - 2*xc*x- 2*yc*y)./(2*(sigma^2)));
  mat(yup:ydown,xleft:xright) = mat(yup:ydown,xleft:xright)+ exponent;

xpower, ypower and mat are 1080*1920 matrices.

thanks!

[James Tursa]

A mex routine could avoid unnecessary intermediate data copies. Could you list the dimensions of *all* of your variables?

[lital]

x and y dimensions change at each run depending on xc and yc

my original needed calculation was:

[x,y] = ndgrid(1:1080, 1:1920);

gauss2d = gauss2d + exp(-((x-xc).^2 + (y-yc).^2)./(2*(sigma^2)));

where xc and yc are scalars that change at each run.

since there is no need for me to this calculation for the entire 1080*1920 matrix (as most of the gauss function will return 0) im only calculating the gauss2d function for a submatrix. (just to be sure i did run this line and it resulted in 300 sec.

Answer 1

[Sean de Wolski]

• Use bsxfun instead of ndgrid, it'll skip the expansions.
• Do xc,yc and sigma change? If not, calculate them being multiplied by 2 (and squared for sigma) once, save that as a new variable.
• Did you use the profiler to determine this line is the one slowing you down?
• How much RAM do you have/are you on a 32 bit system, is it exceeded?

[lital]

sigma does not change.

what changes in each run is x,y,xc,yc.

this code runs 1170 times for a frame and according to the profiler, the last two lines (without ndgrid) take more than 20 sec each to run for a frame.

i have 4 gb ram with a 64bit system and matlab works on 60-70% cpu according to the task manager.

[Sean de Wolski]

Look at using bsxfun for the multiplication (will speed up a bit probably) and remove ndgrid call. Are xc,yc scalar? If so this whole line exp process could be:

exp((-(xc^2+yc^2)+2*bsxfun(@plus,xc*((yup:ydown)'),yc*(xleft:xright)))./(TwoSigSquared))

Now no need for ndgrid which takes time and ram.

Calculate 2*sigma^2 once for a little gain (TwoSigSquared).