Re: Can this be done using CALL_FUNCTION?

[Struan Gray <struan.gray(at)sljus.lu.se>]

J.D. Smith, jdsmith@astro.cornell.edu writes:

  [good advice deleted]
  
    There are some tricks you can play to speed things up for specific
filtering operations.

    If you do ever want to filter the 'vertical' dimension, unroll the
2D dimensions and loop once over a single index rather than use two
nested loops.  This is dramatically faster for any multispectral data
large enough to be called an image.

    Histograms are fast in IDL, loops are slow.  One trick is to
multiply each slice of your data by a large enough number that data in
the slices does not overlap.  Doing a single histogram on the whole
array then gives individual histograms dotted along the bin axis -
extracting them is easy.  The efficiency of doing this depends on how
much memory you have and how many spectral slices need to be looped
over.  If your data is already maximum precision this won't help, but
it rarely is.

    For filters which can be expressed as convolution with a kernal,
use a 3D kernal with non-zero values only in the central plane. Again,
this seem daft, but loops are so inefficient in IDL that for a
reasonable number of bands this is faster than 2D kernals and a loop.

    As JD pointed out, an object representation of your data makes
using and customising this sort of thing much easier than writing
generic routines that loop through arbitrary data arrays.  A data
object *knows* what the 2D slices represent, and which filtering
operations are appropriate for which slices; a generic looping routine
doesn't.  Repetitive typing can be avoided by appropriate use of
object inheritance.


Struan