[GRASS-dev] [GRASS GIS] #2867: Add modifier parameter to r.series

[GRASS GIS]

#2867: Add modifier parameter to r.series
--------------------------+-------------------------
  Reporter:  pvanbosgeo   |      Owner:  grass-dev@…
      Type:  enhancement  |     Status:  new
  Priority:  normal       |  Milestone:  7.0.4
 Component:  Raster       |    Version:  unspecified
Resolution:               |   Keywords:  r.series
       CPU:  Unspecified  |   Platform:  Unspecified
--------------------------+-------------------------

Comment (by glynn):

 Replying to [ticket:2867 pvanbosgeo]:
 > An useful enhancement of r.series would be to add a "modifier" (or
 similar name) parameter to r.series
 > which would allow to calculate the statistic defined by "method" on a
 modified version of the maps (e.g. squared, square root, log, sin, cos,
 etc.).

 My main concern is that this is going to open the door to a never-ending
 stream of requests to add more modifiers or more features (e.g. parametric
 modifiers, or selecting modifiers on a per-output bases) until we end up
 having to embed either r.mapcalc or Python into r.series.

 OTOH, I'm aware that the usual solution (use in conjunction with
 r.mapcalc) has a significant overhead when you're potentially dealing with
 hundreds of input maps. Even so, I'd be inclined to limit the available
 modifiers to those which seem likely to be particularly useful (only
 sqrt/x^2 and log/exp spring to mind); there are a *lot* of real->real
 functions in the standard math library.

 As for the patch itself, I'd suggest:

 1. Determining the function during initialisation and setting a function
 pointer, rather than performing the test in the inner loop. Multiple
 strcmp()s inside a triple-nested loop (rows*columns*inputs) are likely to
 add significant overhead.

 2. Vectorising the operations so that the dispatch is only done once per
 row, not per value. Even an indirect call via a function pointer may be
 more expensive than the function itself (some math functions may compile
 to a single FPU instruction). The simplest way to do this would be to
 modify inputs[].buf[] in-place after reading.

 In that regard, maybe we should try to extract the relevant portions of
 r.mapcalc to a separate library so that they can be re-used here?

 Or maybe a better option would be to just package up the raster-I/O loops
 as a Python module so that the actual processing can be implemented as
 Python functions which convert a 2D (inputs*columns) array into a set of
 1D arrays (one per output).

--
Ticket URL: <https://trac.osgeo.org/grass/ticket/2867#comment:2>
GRASS GIS <https://grass.osgeo.org>