Milton Cezar Ribeiro wrote:
I am running r.neighbors to compute some moving windows
stats. As my region is irregular, many of central pixels
will be NULL. If I use r.mask, it means that when I run
r.neighbours the processing will be faster than when I
not set r.mask
.... I wonder, if the gather.c loops in r.neighbors come across
a NULL if they should immediately break out of the for loop^2
and just set everything to NULL, as that is what the result
will be if it has any NULLs in it. [or maybe the test on each
cell is more expensive than the gathering process. ??]
it might make the module a lot faster.
Hamish
Hamish wrote:
> I am running r.neighbors to compute some moving windows
> stats. As my region is irregular, many of central pixels
> will be NULL. If I use r.mask, it means that when I run
> r.neighbours the processing will be faster than when I
> not set r.mask
.... I wonder, if the gather.c loops in r.neighbors come across
a NULL if they should immediately break out of the for loop^2
and just set everything to NULL, as that is what the result
will be if it has any NULLs in it. [or maybe the test on each
cell is more expensive than the gathering process. ??]
r.neighbors doesn't propagate nulls, but computes the aggregate over
the non-null cells. It doesn't even have an option to propagate nulls.
Adding such a flag would be reasonable, as would aborting the gather
loop at the first null when the flag is used. But the existing
behaviour certainly shouldn't be changed in 6.x.
--
Glynn Clements <glynn@gclements.plus.com>
Milton Cezar Ribeiro wrote:
My feeling is that, mainly when running moving windows, the time to check if
a central pixel is NULL, set output as null and skip must be (very?) lower
than read all pixels inside the windows, compute the desired index, save it,
clean the memory and go to next pixel. Or not?! It is with the developers
r.neighbors won't output a null just because the central pixel is
null, only if the entire neighborhood is null.
--
Glynn Clements <glynn@gclements.plus.com>
Milton Cezar Ribeiro wrote:
In fact, I think that for many r.*, if a "skip if is null" flag could be
setted, may be we can save a lot of time, mainly when running repetitive
steps on large (60,000x48,000 pixels) maps are ran. And, on my case,
addictionally the large size of the maps, as I work with multi-scale
landscape index computing (percentage of cover, density of edge with
different width, connectivity, number of patches), I need to run moving
windows with large sizes (like 101x101 pixels). So, if when I set a
skip_null=T (on my own risk!), I can save not only hours, but weeks of work.
This could be implemented by giving r.neighbors a mask= parameter, or
maybe we need an OUTMASK for output maps. This could be applied
automatically by the raster library to ensure consistency, but
potentially-slow modules could perform the check explicitly as an
optimisation.
--
Glynn Clements <glynn@gclements.plus.com>