problem with resampling 16-bit data

[mpg@…]

We've noticed a problem with reprojecting full 16-bit images when using 
biliner or bicubic: the output image is speckled, presumably due to 
over/underflow.  As verification, using the 20041210 snapshot we recklessly 
replaced all occurrences of "GInt16" with "GUint16" in gdalwarpkernel.cpp and 
the problem went away.

The command line to reproduce is just:
  gdalwarp.exe -t_srs EPSG:4326 -rb f.tif ff.tif
where f.tif is a simple 1-banded, 1Kx1K, unsigned, 16-bit geotiff -- whose 
dynamic range uses all 16 bits.

I'll attach the two images if I can figure out how, otherwise I can email them 
as required.

[warmerdam]

Michael, 

There should be a "Create a new attachment" link a bit above the 
"Additional comments" box.  If that doesn't work just email them directly
to me. 

[warmerdam]

Yikes! I see what you mean.  I'll dig into it. 

[warmerdam]

For now I have altered the PerformWarp() method o avoid using the 
"short" series of functions for anything except the nearest neighbour operations
which don't require any interpretation of the 16bit data.  

This means UInt16 data now goes through the generic interface which is 
significantly slower, but should be accurate. Eventually we should look at
implementing GUInt16 versions of the short functions where performance justifies
the effort. 

I have confirmed correct operation for biliniear and cubic resampling now.  

[warmerdam]

On reflection, I think we really need to implement high performance versions
of the short functions for unsigned shorts. 

Reopening till that is completed. 

[Even Rouault]

I don't think there's problem anymore with warping, although specialized warper instances probably not exist for all datatypes, but the template based approach in trunk should make that much easier if needed