Re: Efficiency in 13980:bd2be36fd949

[Ben Abbott]

On Dec 3, 2011, at 1:45 PM, John W. Eaton wrote:

> On  3-Dec-2011, Jordi Gutiérrez Hermoso wrote:
> 
> | I noticed that this change introduces a bit of a needless slowdown:
> | 
> |     http://hg.savannah.gnu.org/hgweb/octave/rev/bd2be36fd949
> | 
> | Can you please rephrase the condition so that warn is at the front and
> | the short-circuiting avoids the call to unique?
> | 
> | Sorry I don't do it myself. I don't currently have access to my dev machine.
> 
> I could do that, but why should this warning only be issued once per
> Octave session?  Shouldn't it be checked for each image?  Given all
> the other things that image has to do, does it really represent a large
> slowdown if we made the check on each call?
> 
> Also, the test is somewhat simplistic.  For example, if I understand
> correctly what it is intended to do, it fails for something like
> 
>  x = y = (1:100)/5; image (x, y, rand (100, 100, 3))
> 
> jwe

ouch !

I'm embarrassed to say I didn't consider numerical round off problems :-(

Not sure what the best way to resolve that is ... but when I do ...

N = 100;
x = y = (1:N)/5;

tic
for n = 1:1000
  dx = diff (x);
  dy = diff (y);
  dx = std (dx) / mean (abs (dx));
  dy = std (dy) / mean (abs (dy));
  tol = 100*eps;
  if (any (dx > tol) || any (dy > tol))
    warning ("Image does not map to non-linearly spaced coordinates")
  endif
endfor
toc

... I get ....

Elapsed time is 1.3 seconds.

1.3 msecs looks acceptable to me.

If this looks ok, I can push the attached.

Ben

changeset.patch
Description: Binary data