Fallback to 1-bit tags for selected blocks about density HOT 2 CLOSED

Hello Piotr thanks for this.
Two things need verification :

• histogram counting not being too penalizing for performance
• reverting from one mode to another not being too penalizing for performance either (dictionary backups etc.)

The loss on the sao file is only 2%, so kernel switching should have a near-zero performance hit for it to be considered a benefit, as 95+% of files tested are better compressed with 2-bit mode compared to 1-bit.

from density.

Hi,

I've just realized that the potential gains of fallback would be rather small and present only on hardly compressible blocks. Therefore there's no sense of testing for fallback on well compressible blocks. It would make sense then the compression ratio is worse than about 80% - 90% (ie as percentage of original size) and the compressed stream without signatures is less than about 95% of original size (because otherwise the block is incompressible anyway).

But even if it would work, then 2% - 3% improvement in compression ratio on hardly compressible files wouldn't offset even the penalty of having to recompress the block. Maybe I was just too quick to try to improve things and forgot the emphasis on speed :)

Thanks for attention and have a nice day!

from density.