02.04.2016, 16:08, Even Rouault kirjoitti:
.
- there seems to be a block cache mechanism. What is it for ? How does
that work ?
In the block loop (looping through all blocks in a band) at each step
the cache is checked for block(s) that are needed (if focal distance is
> 0 then more than one block may be needed) and which are not needed.
Then blocks are read and written/discarded.
Isn't there some duplication with GDAL own's block cache ? Perhaps I'm just
confusion by the proximity of the naming. I didn't study your code closely.
The GDAL RasterBlock class and block cache seems to me to be something
different. In my code I simply use ReadBlock and WriteBlock methods of
the RasterBand class to get access to the needed blocks and write them
to the disk if a band is being modified. GDAL block cache operates below
ReadBlock and WriteBlock methods.
The code is basically what is in the GDALRasterBand::ReadBlock
documentation but extended for two bands and divided into a few
functions plus syntactic sugar added with macros.
Ari
- there is some overlap/connection with the pixel functions of VRT ( see
"Using Derived Bands" of http://gdal.org/gdal_vrttut.html).
ok
Thanks for the comments!
Ari
The C code is of course not this nice but it is not very far from this.
All the basic four method types I mention below are implemented and
adding new methods is not very difficult. The code makes heavy use of
templates and macros in an attempt for versatility (GDAL rasters can
have many datatypes) and easiness to write methods in a few lines of
code. The whole set of code is now ~2500 lines but it creates rather
large executables due to templates. The code is C++ but the API is C
like, I've so far done no changes to existing GDAL code because of this.
As written below, the code works on two levels of x,y loops: the blocks
within a band, which is generic and cells within a block, which is
separate for each method. The methods work on line by line (no recursion
etc.) and thus for example the fill_depressions is not optimal but
instead very simple. Also, very large rasters can be processed.
I think this may be somewhat similar thing in relation to GDAL as the
network support, so it could be a thing to add to the library if wished.
I've added RFC 62 "raster algebra" to spark discussion.
Ari
20.03.2016, 10:21, Ari Jolma kirjoitti:
Folks,
I played around a bit with the map algebra idea and came up with
something that could work nicely as a plugin. The initial code is at
https://github.com/ajolma/gdal/tree/trunk/gdal/map_algebra
The idea is based on processing raster bands block by block and
calling a given function on each block. Using macros and templates the
code should be rather nice and manageable even when supporting
multiple data types for raster cells.
Further, the idea is to support three kinds of map algebra methods,
mostly working on a raster band in-place.
1) Simple ones, which do not take any arguments beyond the raster band
itself, examples are functions like log, sin, rand.
2) Those which take one non-spatial argument. The argument can be a
scalar like a number, or more complex like a reclassifier.
3) Those which operate on two rasters. Examples are summation of
rasters and computing flow directions on a DEM. The latter is a bit
more complex since it is a focal method and requires a 3 x 3 matrix of
blocks from the other operand raster.
Maybe a fourth kind of methods are those which compute something from
a raster.
This would lead to four functions in the C API and raster band methods
in the bindings.
Comments are welcome, the initial code base contains a small demo
program. Eventually, if this works, I'll make a RFC from this.
Ari
_______________________________________________
gdal-dev mailing list
gdal-dev@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/gdal-dev
_______________________________________________
gdal-dev mailing list
gdal-dev@lists.osgeo.org
http://lists.osgeo.org/mailman/listinfo/gdal-dev
