[GRASS-dev] New module: r.grow.distance

[Glynn Clements]

Hamish wrote:

> > I have added a new module, r.grow.distance, which generates a map
> > where the value of each cell is the distance to the nearest non-null
> > cell in the input map (similar to r.cost with a constant cost, except
> > that you get real circles rather than approximations).
> 
> ie better than r.cost's knight's move / octagon. (?)

AFAICT, r.cost gives octagons without -k and hexadecagons (16 sides)
with -k.

Even if that's an adequate approximation, r.grow.distance is
noticeably faster, especially if the map won't fit into memory.

In the process of quantifying that, I found a bug in the way that
r.cost computes the number of segments:

    if (maxmem > 0)
	segments_in_memory =
	    2 + maxmem * (nrows / SEGCOLSIZE) * (ncols / SEGCOLSIZE) / 100;
    else
	segments_in_memory = 4 * (nrows / SEGCOLSIZE + ncols / SEGCOLSIZE + 2);

Because integer division rounds down, if the map is
one-point-something segments in each direction, it allocates 2+1*1 = 3
segments for a map which really needs 4.

[With "spearfish".]

$ g.region res=45
$ time r.cost -k --o start_rast=test input=one output=test.grow
318 rows, 423 cols
real	1m18.267s
user	0m2.876s
sys	1m15.358s

78 seconds, of which 75 is spent in the kernel (i.e. swapping segments
to file). At res=30 (spearfish's default), I Ctrl-C'd it after a
couple of minutes.

OTOH, with percent_memory=0 (4*(1+1+2) = 16 segments):

$ g.region res=45
$ time r.cost -k --o start_rast=test input=one output=test.grow percent_memory=0
318 rows, 423 cols
real	0m1.518s
user	0m1.477s
sys	0m0.040s

1.5 seconds. A significant improvement over 78 seconds.

Fixed in SVN trunk (r32239).

-- 
Glynn Clements <glynn at gclements.plus.com>