[gdal-dev] GDAL_FORCE_CACHING - alters results?

[Even Rouault]

On mardi 28 janvier 2020 11:49:01 CET Daniel Evans wrote:
> Hi all,
> 
> I've recently tried enabling the GDAL_FORCE_CACHING option, and found it
> gives some significant improvements to TIFF/VRT read speeds.
> 
> However, using GDAL 2.3.0, I've now found that when reading large VRTs, I
> can sometimes get subtly different values - up to about 1% summed
> difference across a 1000x1000 or so area. Is there a known issue in
> GDAL_FORCE_CACHING in this older version of GDAL, or alternatively have I
> misunderstood the use of this option and is this expected behaviour?

GDAL_FORCE_CACHING will force I/O requests to happen exactly on VRT block 
boundaries.
So what you observe is slight inconsistencies in results depending on the 
window passed to the VRT driver.

I see that 2.4.0 had a potentially related fix
"""VRT: fix source window computation that caused sub-pixel shift with non-
nearest resampling"""

But if you use a warped VRT, then what you observe is somewhat expected. 
gdalwarp due to approximations in some computations doesn't lead to 
"consistent" results if you change the output window.
See
https://lists.osgeo.org/pipermail/gdal-dev/2019-November/051177.html

Even

-- 
Spatialys - Geospatial professional services
http://www.spatialys.com