Re: [Vxl-users] Fast filling of regions in VXL?

[Neeraj K. <ne...@cs...>]

Hi all,

Thanks for your quick replies!

Mr. Scott: you are indeed correct that most of the time is being spent 
in vil_fill_line. As I mentioned in my previous email, I am using gcc 
3.4.4 on cygwin, and my compiler flags are -O3 -pg. Using those 
settings, the top few lines of the generated profile are:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  64.62      3.16     3.16 76662274     0.00     0.00  void 
vil_fill_line<bool>(bool*, unsigned int, int, bool)
  10.02      3.65     0.49       20     0.02     0.02 
andImages(vil_image_view<bool> const&, vil_image_view<bool> const&, 
vil_image_view<bool>&)
   4.91      3.89     0.24       20     0.01     0.18 
getProbImage(vil_image_view<unsigned int> const&, unsigned int, double, 
int, int)
   4.50      4.11     0.22        1     0.22     0.22 
UniformHistoProgrammer<unsigned char>::convert()
   3.68      4.29     0.18  1348222     0.00     0.00  void 
fillRect<bool>(vil_image_view<bool>&, vgl_box_2d<int>, bool)


As you can see, my program spends about 65% of its time in 
vil_fill_line, with the next-most expensive function taking less than 
1/6th that time.

As for why vil_fill_line is still showing up in the profiled output, I 
have no idea. I do not know if there are more advanced optimization 
settings than -O3 (I have to use -pg to enable profiling). I have even 
tried turning off the profiler, but that doesn't seem to affect 
performance very much.

My only thought for making it run faster was to use memset, as Mr. 
Vanroose has suggested. Making that change results in the following profile:

   %   cumulative   self              self     total
  time   seconds   seconds    calls   s/call   s/call  name
  27.78      0.50     0.50       20     0.03     0.03 
andImages(vil_image_view<bool> const&, vil_image_view<bool> const&, 
vil_image_view<bool>&)
  15.56      0.78     0.28       20     0.01     0.03 
getProbImage(vil_image_view<unsigned int> const&, unsigned int, double, 
int, int)
  14.44      1.04     0.26        1     0.26     0.26 
UniformHistoProgrammer<unsigned char>::convert()
  11.67      1.25     0.21  1348222     0.00     0.00  void 
fillRect<bool>(vil_image_view<bool>&, vgl_box_2d<int>, bool)


As expected, vil_fill_line is gone, and we see that even the calling 
function fillRect (which now has a call to memset instead of 
vil_fill_line) has not become much more expensive. Also, the total 
running time of the program has been cut by about half, suggesting that 
nearly all of the time spent in vil_fill_line can be avoided by using 
memset.

Of course as Mr. Vanroose pointed out, this can only be applied to 
images with 1-byte pixels and istep == 1. However, I know that for me 
(and I would imagine for many others), this is a very common case, and 
so I agree with Mr. Vanroose's suggestion to explicitly test for that case.

My thanks to both of you for your help.

--Neeraj