RFC 40: Improving performance of Raster Attribute Table implementation for large tables¶
Raster Attrbute Tables from some applications (notably segmentation) can be very large and are slow to access with the current API due to the way only one element can get read or written at a time. Also, when an attribute table is requested by the application the whole table must be read - there is no way of delaying this so just the required subset is read off disk. These changes will bring the attribute table support more in line with the way raster data is accessed.
It is proposed that GDALRasterAttributeTable be re-written as a virtual base class. This will allow drivers to have their own implementation that only reads and writes data when requested. A new derived class, GDALDefaultRasterAttributeTable will be provided that provides the functionality of the GDAL 1.x GDALRasterAttributeTable (ie holds all data in memory).
Additional methods will be provided in the GDALRasterAttributeTable class that allow ‘chunks’ of data from a column to be read/written in one call. As with the GetValueAs functions columns of different types would be able to read as a value of a different type (i.e., read a int column as a double) with the appropriate conversion taking place. The following overloaded methods will be available:
CPLErr ValuesIO(GDALRWFlag eRWFlag, int iField, int iStartRow, int iLength, double *pdfData); CPLErr ValuesIO(GDALRWFlag eRWFlag, int iField, int iStartRow, int iLength, int *pnData); CPLErr ValuesIO(GDALRWFlag eRWFlag, int iField, int iStartRow, int iLength, char **papszStrList);
It is expected that the application will allocate the required space for reading in the same way as with the RasterIO() call.
The char** type will be used for reading and writing strings. When reading strings, it is expected that the array is created of the correct size and ValuesIO will just create the individual strings for each row. The application should call CPLFree on each of the strings before de-allocating the array.
These methods will be available from C as GDALRATValuesIOAsDouble, GDALRATValuesIOAsInteger and GDALRATValuesIOAsString.
This is also an opportunity to remove unused functions on the attribute table such as GetRowMin(), GetRowMax() and GetColorOfValue().
The Python bindings will be altered so ValuesIO will be supported using numpy arrays for the data with casting of types as appropriate. Strings will be supported using the numpy support for string arrays.
The proposed additions will extend the C API. However, the C++ binary interface will be broken and so GDAL 2.0 is suggested as an appropriate time to introduce the changes.
Care will be taken to still support the use of Clone() and Serialize() in derived implementations of the GDALRasterAttributeTable class as these are called by existing code. For implementations where the table is not held in memory these may fail if the table is larger than some suitable limit (for example, GetRowCount() * GetColCount() < 1 000 000). Clone() should return a instance of GDALDefaultRasterAttributeTable to prevent problems with sharing memory between objects.
Existing code may need to be altered to use create instances of GDALDefaultRasterAttributeTable rather than GDALRasterAttributeTable if an in memory implementation is still required.
Impact on Drivers¶
The HFA driver will be updated to support all aspects of the new interface, such as the new functions and reading/writing upon request. Other drivers will be modified to continue to use the in memory implementation (GDALDefaultRasterAttributeTable).
The Python autotest suite will be extended to test the new API, both for the default implementation and specialised implementation in the HFA driver.
We (Sam Gillingham and Pete Bunting) are prepared undertake the work required and have it ready for inclusion in GDAL 1.11 There needs to be a discussion on the names of the methods and on the internal logic of the methods.
Ticket #5129 has been opened to track the progress of this RFC.