[gdal-dev] GDAL Raster Attribute Tables.

Even Rouault even.rouault at mines-paris.org
Wed May 1 02:52:30 PDT 2013

Le mercredi 01 mai 2013 01:29:30, Peter Bunting a écrit :
> Hi All,
> Sam and I have prepared an RFC for our proposal for improving raster
> attribute table (RAT) support and put it on the trac.
> http://trac.osgeo.org/gdal/wiki/rfc40_enhanced_rat_support
> We'd be keen to here feedback on our proposal.


Here are my observations and questions :

* Is the boolean data type really necessary ? There will be a difficulty with 
the associated C bindings since 'bool' isn't a C type, but a C++ one.

* I'm not sure about what you meant with "The CSLString type will be used for 
reading and writing strings". You probably meant the CPLString C++ class right 
? I'm not sure how this would fly with ValuesIO(GDALRWFlag eRWFlag, int iField, 
int iStartRow, int iLength, char **papszStrList). This is probably just an 
implementation detail.
Related to that, when eRWFlag == GF_Read, you will probably need to document 
that only the papszStrList array must be allocated with a size of iLength * 
sizeof(char*) bytes, and that the papszStrList[i] values that will be set in 
it will be allocated with the CPL allocation routines. So users will be 
responsible to clear the values with CPLFree().

* For the HFA specialized implementation, I guess you would override almost 
all methods of GDALRasterAttributeTable, and that you wouldn't use the 
"std::vector<GDALRasterAttributeField> aoFields" member variable for example 
since you don't want the size of anValues/adfValues/aosValues arrays of 
GDALRasterAttributeField to go out of control.
So perhaps a better design would be to completely remove the private members 
of GDALRasterAttributeTable to make it an abstract class. The current 
implementation, with the private member variables, would go to a 
GDALDefaultRasterAttributeTable that would extend GDALRasterAttributeTable 
(like HFARasterAttributeTable would do).
This would have of course some backward compatibility implications, especially 
for C++ users, since GDALRasterAttributeTable would be no longer instanciable 
(GDALDefaultRasterAttributeTable would), but I don't believe this would cause 
much churn in existing user code bases, and would have only some limited 
implication in GDAL code base. The C function GDALCreateRasterAttributeTable() 
could be preserved and return an instance of GDALDefaultRasterAttributeTable.

* If your purpose is to address huge RATs, I'm not sure about what we can do 
with the Serialize() method, which is for example used by the PAM mechanism 
when translating datasets to formats that don't support natively RATs. Perhaps 
in the case the RAT is detected to be large to be reasonably serialized (for 
example 10 MB ?), this method could just return NULL (some check should be 
done in the existing GDAL code base to see if we are robust to NULL being 
returned. I can see at least that CPLAddXMLChild( psTree, psPam->poDefaultRAT-
>Serialize() ) at line 209 of gdalpamrasterband.cpp would need some care)

* In a similar situation (huge RAT), I'm not sure what the Clone() method 
would do. GDALPamRasterBand::SetDefaultRAT() uses Clone() on that passed RAT 
object.http://trac.osgeo.org/gdal/attachment/ticket/4903/4903.4.diff would do 
the same on VRTRasterBand::SetDefaultRAT(). IdrisiRasterBand::SetDefaultRAT() 
does that too. Perhaps that HFARasterAttributeTable::Clone() could return a 
self-contained immutable copy of itself (i.e. it could survive the destruction 
of the HFADataset object of the source RAT). Perhaps only in the huge RAT 
case. In other cases, it could return a GDALDefaultRasterAttributeTable object 
with all columns and rows in memory, that could be potentially modified by 
calling code. Hum, that seems to be complicated...

In the SetDefaultRAT() use case, the use of Clone() is not meant at getting a 
free-standing copy of the source object that can be altered, but rather to be 
independant from the life cycle of the passed object. Would be probably better 
addressed with some shared pointer mechanism, but the implication would be 
really too huge and go well beyond changes in RAT management...

* Not linked with your proposal, but I noticed that gdal_rat.h exposes methods 
that are not implemented, like GetRowMin(), GetRowMax(), GetColorOfValue(). 
Your work could be the opportunity to clean that by removing them.

* Just a detail: In the current state of your proposal, "The proposed 
additions will have an impact on C binary compatibility because they change 
the API" would be better rephrased as "The proposed additions will extend the 
C API" (there's no compatibility problem since it is backward compatible).

* I'd like to see a mention of extending the Python autotest suite to test the 
new API, both for their default implementation and specialized implementation 
in the HFA driver.

Best regards,


More information about the gdal-dev mailing list