[gdal-dev] GTiff bit shuffle compression feature request

[Laurențiu Nicola]

Hi,

In my (limited) experience, it can really make a difference for floating-point rasters. Testing with a small one I have on hand (10980x10980x1, Float32), I get:

 - GeoTIFF DEFLATE 280 MB
 - Zarr BLOSC zlib NONE 281 MB
 - Zarr BLOSC zlib BIT 253 MB
 - Zarr BLOSC zlib BYTE 249 MB

Laurentiu

On Fri, Dec 8, 2023, at 19:19, Even Rouault via gdal-dev wrote:
> Jesse,
> 
> This would break interoperability with other TIFF readers... Even adding a new TIFF tag to advertize that bit shuffling is applied would probably not be a sufficient guard, as existing readers wouldn't read it, and would just display garbage, which is worth that not being able to open the file at all. The only way I can think off of doing that in a safe way would be to use new values for the Compression tag, which isn't pretty either.
> 
> You should probably try Zarr which has such capability with the Blosc codec. Cf https://gdal.org/drivers/raster/zarr.html : BLOSC_SHUFFLE
> 
> I'm curious however to know which typical compression gain you get with that.
> 
> Even
> 
> 
> 
> Le 08/12/2023 à 18:06, Meyer, Jesse R. (GSFC-618.0)[SCIENCE SYSTEMS AND APPLICATIONS INC] via gdal-dev a écrit :
>> Hi,
>>  
>> When using horizonal differencing to reduce the numerical range of band data, the upper bytes in the produced stream are typically 0 which leverages LZ’s byte based compression model.  But the least significant bytes can still have many significant bits as 0. Unless the whole byte is replicated, LZ compressors can’t do much to leverage the pattern however.  For data with temporal and or spatial coherence, ‘shuffling’ is another effective strategy to losslessly reform the data stream to be favorable to LZ style compressors.  And plays nicely off gains already provided by the PREDICTOR functionality.
>>  
>> The notion is to arrange the bit stream where the Nth “shuffled” byte contains the Nth bit from each byte in the sequence.  The sequence length is usually determined by the data type bit length.
>>  
>> For example (for brevity, assume bytes are 4 bits long)
>>  
>> Byte 1,  Byte 2, Byte 3, Byte 4
>> 0001, 0011, 0111, 0001
>>  
>> They all share the top 0 bit and the bottom 1 bit,
>>  
>> “Shuffled”
>> 0000, 0010, 0110, 1111
>>  
>> The algorithm is pretty simple to implement, and can be SIMD accelerated for high performance.
>>  
>> While we specifically are users of the GTIFF format, such a strategy could be employed generically for most raster and even vector formats.
>>  
>> Best,
>> Jesse
>> 
>> _______________________________________________
>> gdal-dev mailing list
>> gdal-dev at lists.osgeo.org
>> https://lists.osgeo.org/mailman/listinfo/gdal-dev
>> 
> -- 
> http://www.spatialys.com
> My software is free, but my time generally not.
> _______________________________________________
> gdal-dev mailing list
> gdal-dev at lists.osgeo.org
> https://lists.osgeo.org/mailman/listinfo/gdal-dev
> 
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osgeo.org/pipermail/gdal-dev/attachments/20231208/c78c09bd/attachment.htm>