Re: Misc. Bugs & Problems

[steinhh(at)ulrik.uio.no (Stein Vidar Hagfors Haugan)]

In article <eK6N4yBY#GA.322@ntawwabp.compuserve.com> 
"Steve Scheele" <sscheele@scitor.com> writes:

> I am running IDL V5.2 on WinNT and have encountered the following bugs and
> problems. All have been discussed with RSI technical support.
> 
> ;********************
> Problem: Sort slows down considerably when sorting integer arrays
> containing many identical values.
> 
> Workaround: Add small-random values to the array before sorting -
> improves sorting by 40X

Note that this seems to be platform dependent. I think IDL is using
whatever builtin sorting routine exists on the specific platform.  On
Digital Unix:

IDL> .run
  pro test,n 
  i=fix(randomu(seed,100000)*n)
  t=systime(1)
  ix=sort(i) 
  print,systime(1)-t 
  end

IDL> test,10     ;; Lots of identical values
      0.12988293 
IDL> test,100 
      0.31445301 
IDL> test,1000 
      0.46289098 
IDL> test,10000 
      0.65039003 
IDL> test,100000 ;; Fewest identical values. 
      0.77246106 

So, adding a random value to make each entry unique could increase
the execution time here by an order of 10 (ok, I'm an astronomer!).

I seem to remember that we found some pathological case for which the
Unix sorting routine (qsort?) was very slow...

Ah, yes, this is it:

IDL> i=randomu(seed,10000)
IDL> t=systime(1) & ix=sort(i) & print,systime(1)-t
     0.059569955 
IDL> i=[findgen(5000L),findgen(5000L)]
IDL> t=systime(1) & ix=sort(i) & print,systime(1)-t
       9.2021489 

Ouch! Randomizing the order of the numbers before sorting would give
a speed increase on the order of 100!!

There will always be pathological cases for a given algorithm,
though. It would be nice, however, if the pathological cases for
IDL's SORT() was *known*, though, independently of the platform, so
one could avoid them when the prior structure of the dataset is
known. 

Maybe the best thing would be to add keywords to SORT() that advises
about any known structure... Like /DUPLICATES to warn about many
identical values, /IDENTICAL_SEQUENCES to warn about long, identical,
sorted subsequences, etc etc...?

> Bug: Passing a UINT array to REBIN crashes IDL

Wheee !

> Bug: Resizing a draw widget, flips vertical sliders up side down.
> 
> Workaround: Pass an initial value to the slider - this workaround is
> apparently machine/OS dependent. It didn't work for me.

Could somebody give a tiny example of this - I'm not exactly clear on
what this means.

Regards,

Stein Vidar