Re: not exactly recent: more coder speed stuff


"Michael Schöberl" <MSchoeberl@xxxxxxxxxxxx> wrote in message
news:43ddf707$1@xxxxxxxxxxxxxx
>> mostly, I was figuring I would do an arithmetic coder (sort of) based on
>> forcing probabilities to powers of 2. ofterwise, not much was done (apart
>> from using a shift as a range).
>
> maybe I'm missing a reference ... but what are you trying to do!?
>
hopefully fast coding, without the hassles of huffman.

> ac has the advantage of using fractions of bits/symbol ...
> If you want to use probabilites with powers of 2 then you would get full
> bits/symbol. Why are you still trying to use the AC?
>
as noted above.

full ac is not exactly fast, maybe fast enough, but far slower than other
options: table-based huffman encoding/decoding; not doing any coding.

if I could get speed pretty close to huffman, but using a far simpler
implementation than the one I had before (especially if I could have
adaptive statistics, but this is doubtful), this would be cool.

thus far, this has escaped me. by the time speed approaches (optimized)
huffman, so has the code complexity.

also quite useful here would also be if there were some particularly simple
way to implement a huffman encoder/decoder (that still goes pretty fast).

the gain of ac (especially paq-style) is the simplicity and ease of tweaking
the coder for different uses. huffman is more problematic, everything is
more work.

problem, is I just can't get it much faster than about 6MB/s (decode) on my
box (about 2x the speed of fpaq0). this seems to be a limit. I have had to
live with this limit now for a while...

for range coders I have got to speeds like 12-15MB/s for decode (usually a
lot faster for encode), but this is still not fast enough, and not enough
faster to justify using them instead (a 2x speedup for a notable increase in
code complexity/inflexibility).

(target is 50 to 100MB/s decode).


then again, what I hope for may well not be possible, as I am begining to
suspect.
no algo that is all of: fast, simple, gets good ratios.

or such...


.

Relevant Pages

  • Re: Adaptive Huffman algorithm !!!
    ... > Try using Range coding instead. ... That should give you a huff and puff coder. ... > In theory huffman coder is a Arithmetic coder in which no bits are ... > Search the net for range coder. ...
    (comp.compression)
  • Re: A little lost with a compression program
    ... I've been following some compression articles from this site: ... Coder articles. ... decided to go with Huffman coding, ...
    (comp.programming)
  • Re: huffman compression
    ... >i want a full code of huffman data compression for text files in C++. ... other algos, eg, huffman, for some increase in total compression ratio). ... faster than I can get an arithmatic coder. ... with static huffman, eg, chunking and stream encoding concerns). ...
    (comp.compression)
  • Re: new range coder: speed issue
    ... maybe, vs. huffman at least. ... because the total amount the coder has to do is reduced). ... > somehow approximate the division by whuff_RRng by a shift. ... > may speed up by maintaing a better data structure for symbol lookup. ...
    (comp.compression)