And learning time gertjan new a daily weekly within reimburse PAQ maintain multiple context hashes of different orders and multiple pointers into the buffer. The prediction is indirect by mapping the match length to a prediction through a direct context model. ZPAQ uses a simpler match model with just one pointer and one hash, although it is possible to have multiple, independent match models. The prediction for a match of L bytes is that the next bit be the same with probability 1 8L. The user specify the context length by using a rolling hash that depends on the desired number of characters. If h is the context hash, c is the input byte, then the update: h h with their usual meanings C C++. Division or mod by 0 is 0. means The post-processor if it is present, is called once per decoded byte with that byte the A register. At the end of each segment, it is called once more with -1 A. The decompresser output is whatever is output by the OUT instruction. The context model is always present. It is called once per decoded byte. It puts its result H. OUT has no effect. HCOMP sees as input the PCOMP code followed by a contiguous stream of segments with no separator symbols. The ZPAQ program is a development environment for writing and debugging models. It allows programs to be single stepped or run separate from compression. It accepts control statements IF IFNOT--ENDIF and DO-WHILE UNTIL FOREVER and converts them to conditional jumps. It allows passing of numeric arguments and comments parenthesis. If a C++ compiler is present, then ZPAQL code is compiled by converting it to C++ and then running it. Otherwise the code is interpreted. Compiling makes compression and decompression 2 to 4 times faster. The default configuration for both ZPAQ and ZPIPE is described by the file mid.cfg below. comp 3 0 8 icm 5 isse 13 2 isse $1 1 isse $1 2 isse $1 3 isse $1 4 match $1 $1 7 mix 16 7 255 hcomp c++ c=a b=c a=0 d= 1 hash d=a b-- d++ hash d=a b-- d++ hash d=a b-- d++ hash d=a b-- d++ hash d=a b-- d++ hash b-- hash d=a d++ a=c a 64 if aa a+=d a= 20 d=a jmp 9 endif endif a=d a== 0 ifnot d++ d=a d-- endif d=0 d++ d++ b=c b-- a=0 hash d=a d++ b-- a=0 hash d=a d++ b-- a=0 hash d=a d++ a=b a-= 212 b=a a=0 hash d=a ba a=b a&= 60 hashd d++ a=c a 64 if aa means swap H with A. The hash D is a rolling hash but D is cumulative. JMP 9 skips 9 bytes past the commented else clause. If the input is not a letter then H is moved to H and H is cleared. The following results are for the Calgary corpus as a solid archive when supported. Compression is timed seconds on a 2 GHz T3200. Program Size Time Memory zip -9,020 0 0 ppmd 804 0 7 ppmd -m256 -o16,243.3 ppmonstr 669 8 lpaq1 682 8 lpaq9m 6,161 198 paq9a -6,914 209 zpaq ocmid.cfg 699 8 zpaq ocmax.cfg 644 20 zpaq ocmax.cfg,1,320 477 paq8l -6,474 435 paq8px_v67 598 469 Crinkler. Crinkler is a compressing linker written by Rune L. H. Stubbe and Ashe Christensen 2010 for producing small, self extracting x86 executable programs. It is optimized for producing demos, or programs limited to 4096 bytes. It works like UPX, extracting to memory and then running itself. But for this application, it is important to use a much better compression algorithm and that the decompression code be very small, this case 210 to 220 bytes of size-optimized x86 assembler. Compression speed is much less important. As explained by Alexandre Mutel, Crinkler is based on PAQ1 with sparse models and static weights tuned by the compressor. The model is transmitted as a string of bytes, each representing one context over the last 8 bytes, for example, 11100000b 0xE0 for order 3 context, or 01100000b 0x60 for a sparse order 2 context with a 1 byte gap. Model weights are powers of 2. Up to 32 model weights are coded and packed into 4 bytes, such that a 1 bit means that the weight is twice the previous weight and a 0 means it is the same. Contexts are mapped to a pair of 8-bit counters a hash table. To make the decompresser as small as possible, there is no hash table collision detection. To compensate, the hash table is made very large, usually 256 MB or 512 MB. Also, there is no counter overflow detection. This saves 6 bytes of code. The compressor searches model space to find the best compression. It tries adding one model at a time, iterating bit mirrored order and keeping a model if it improves compression. Each time a model is added, it tries removing existing models one at a time to if compression improves