Re: trying to optimize my octave program

[Jaroslav Hajek]

2010/3/18 Tim Rueth <address@hidden>:
> You guys crack me up.  I didn't know programmers had such a good sense of
> humor ;-)  I said this was my first *Octave* program (I know 750 lines isn't
> really that much, and I did "hello world" a long time ago in Fortran on
> punched cards).
>
> Okay, here's the essence of my top-level script:
>
> global sig long short c_thd r_thd t_thd b_thd;    # these are the loop
> iteration vars
> short_min =     5; short_max = 10; short_st = 1;
> long_min = 20; long_max = 40; long_st = 1;
> sig_min = 5; sig_max = 12; sig_st = 1;
> c_thd_min =     -3; c_thd_max = -5; c_thd_st = -1;
> r_thd_min =     1; r_thd_max = 3; r_thd_st = 1;
> t_thd_min =     1; t_thd_max = 3; t_thd_st = 1;
> b_thd_min =     0; b_thd_max = -2; b_thd_st = -1;
>
> # Schmoo variables, calling function "analyze" for each data point
> for sig = sig_min : sig_st : sig_max
>  for long = long_min : long_st : long_max
>    for short = short_min : short_st : short_max
>      if short <= long      # which it should be, but just checking
>        for c_thd = c_thd_min : c_thd_st : c_thd_max
>          for r_thd = r_thd_min : r_thd_st : r_thd_max
>            for t_thd = t_thd_min : t_thd_st : t_thd_max
>              for b_thd = b_thd_min : b_thd_st : b_thd_max
>                # Call function analyze, which uses global iteration vars to
> operate on several global vectors
>                [c_cntr b_cntr s_cntr avg val actg ddown pk_ddown] =
> analyze(outfile);
>                if actg > mx_actg
>                  mx_short = short;
>                  mx_long = long;
>                  mx_sig = sig;
>                  mx_c_thd = c_thd;
>                  mx_r_thd = r_thd;
>                  mx_t_thd = t_thd;
>                  mx_b_thd = b_thd;
>                end;
>              end;
>            end;
>          end;
>        end;
>      end;
>    end;
>  end;
> end;
>
> Even though it's aesthetically beautiful, this 7-nested for-loop (evil!)
> executes custom function "analyze" 6*21*8*3*3*3*3 = 81,648 times in this
> case.  The iteration vars of the for-loops are global, and therefore are not
> passed to the function as arguments.  Inside "analyze" is yet another
> for-loop that runs 4288 iterations on several global vectors that are each
> 4288 x 1 in size.  So, the code inside analyze's for-loop gets executed
> 81648*4288 = 350e6 times.  Using tic/toc, it takes about 4.5 sec for this
> function to execute.  So, 4.5sec/4288 = only 0.001 sec for the code inside
> the loop within "analyze" to run.  With these parameters, the whole script
> will run for about 81648*4.5 sec = 4.25 days.
>
> Judd had mentioned that perhaps I'm not doing enough "work" inside the
> inner-most loop.  Good point, since it only runs for 0.001 sec.  However, I
> don't see a way to get more work done there without moving the next-outer
> for-loop into the function, which is pointless.  Even though the inner-most
> loop runs for only 0.001 sec, maybe I just need to focus on trying to
> optimize this down even further.  Any reduction * 350e6 could be huge.
> Which brings me to my original question of whether porting "analyze" to C
> and compiling it would help much.  Or, from a memory management perspective,
> since I have 11 vectors, each having a 4288x1 size, can/should I convert
> these to int8, single, etc., data types within the vectors if possible?  I
> don't see much point here, since it really isn't that much memory savings
> overall.  If anyone has any other ideas, let me know.  You guys have been
> quite supportive, and I appreciate it.
>
> --Tim
>


I suppose you understand that this is useless unless you show also
what analyze does. Is there a good reason why sig, ong, short etc. are
global variables rather than function parameters?

-- 
RNDr. Jaroslav Hajek, PhD
computing expert & GNU Octave developer
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz