James Bonfield
James Bonfield
> Outputting QS data series using EXTERNAL(11), codec RANS4x16 order 193 (PACK+RLE+O1) is equiv to: > > MAP("!-3E"->0123, XBETA(offset=2, nbits=2, RLE(EXTERNAL(50), EXTERNAL(51))) > with RANS0 on external block 50 (run...
Fixed barring the discussion on 4x16 O1 table using 4x16 O0 encoder. I've been experimenting with externalising most of the rANS4x16 data transforms to be separate CRAM encodings which can...
Rebased, and also updated with newer ideas. The rANS4x16 codec now has a choice of 10-bit or 12-bit frequencies for the order-1 frequency table. The 4x8 rans codec has a...
Note to @yfarjoun: this is killing latexdiff for reasons currently unknown. CRAMv3.tex in master and this PR both build and view fine. The diff subdirectory produced by our circleci process...
The offending LaTeX that breaks the diff is this: ``` { \setlength{\parindent}{1cm} $ cfreq_{i} = \left\{ \begin{array}{l l} 0 & \quad \textrm{if $i < 1$} \\ cfreq_{i-1} + freq_{i-1} &...
The CRAMv4.tex document has now been hived off into its own PR #502, given it's likely a lot more contentious and harder to implement than the new codecs in v3.1...
Sorry this is long, but it's hard to describe without some basic understanding of the implementation choices. However in summary: Do we want to potentially gain 5-30% speed when we...
I decided to go for it and add the documentation of 32-way vs 4-way into the CRAM 3.1 codec. I've done this because the penalties of not explicitly vectorising aren't...
As this is still kicking around in a non-adopted state, mainly due to a combination of lack of review and lack of a complete 2nd implementation, I decided it's still...
An update, keeping to 16-bit context size at the moment and tested on a single large block of data (varying between 100M and 1G by machine type). This is the...