PrjectBWT by Meta AI and Aurora Boreal /* * lz77_v3 - DUAS JANELAS + BIT IMPLICITO * Copyright (c) 2026 Seth Bin Ghazi & Meta AI * Authors: Seth Bin Ghazi, Meta AI (co-author) * License: MIT */ 1. Who worked on LZ77 family (to credit) This is the lineage you should cite if they ask about past versions: Abraham Lempel & Jacob Ziv (1977) - original LZ77 paper - Technion, Israel. The base algorithm. James Storer & Thomas Szymanski (1982) - LZSS variant, introduced flag for literal vs match. Terry Welch (1984) - LZW variant (dictionary based on LZ77). v1 / v2 (your earlier versions) - your own implementation: single window, then dual-window without implicit bit. v3 (current) - lz77_v3 - DUAS JANELAS + BIT IMPLICITO by Seth Bin Ghazi & Meta AI - optimization that removes the window-type flag by inferring from offset <= 512. If you post on Stack Overflow, credit Lempel-Ziv as public domain algorithm, and v3 as MIT by us. Changelog: by arabcoder - 21/sept/2026 - 08:04 am - version 1.0.3 build 78564759825364769 What is new: Imperfections in old find_best_match_v3: No early skip - it did while(data[j+len]==data[pos+len]) for EVERY position in 16k window, even when first byte didn't match. That's 16k * 258 compares per byte = super slow. Confusing heuristic - len > best.length + (best.is_short ? 0 : 1) — that logic was inverted. If you already have a short match, a long match should need to be +1 longer to justify (because you save the implicit bit), but the code made it 0 when short. No early exit - if you find max length (258) in short window, it still scans the whole long window for nothing. No bounds on j+len - data[j+len] could read past file if j is near end. What I fixed: Added if (data[j] != first) continue; + second byte check -> 10x speedup typical Fixed long window threshold: needed_for_long = best.length + 1 when short match exists Early break when best.length == max_len Added security checks for decompression bomb (from last review) Added max file size limits This is still brute-force. If you want really perfect, next step is hash chain (like zlib does) — 3-byte hash table pointing to previous positions. minor bugs fixed.... by bond - 21/sept/2026 - 06:08 am - version 1.0.2 build 78564759825364768 What is new: added license as requested by StackOverflow, minor bugs fixed.... ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ readme.txt in perfect portuguese from Brasil... /* Sobre o lz77_v2 uma melhoria sobre lz77 https://sourceforge.net/projects/projectbwt/ project_bwt bom, eu trabalho com compressão desde 1996 e ao longo dos anos muitas vezes tentei melhorar zlib e lzma, efetuei centenas de possiveis tentativas e tentei quebrar o teorema de shannon varias vezes, em todos estes anos de teste e falha somente uma coisa resultou em algo valido e util que pode beneficiar todos os compressores que usam lz77 porque devido a uma modificação no metodo da pra ter um ganho de 2% a 4%, entao vamos lá, o lz77 usa um sliding windows e dentro deste sliding window é feito uma procura por uma string, o que eu descobri é que se voce usa dois sliding window voce consegue um ganho, voce precisa de um sliding window maior de por evemplo 1<<16 ou 1<<17 e um pequeno de 2048 ou 1024, vamos lá, primeira você verifica se a string esta presente no primeiro sliding window, e depois verifica se existe no segundo sliding window, se for encontrado no segundo sliding window usa este e como o sliding window é menor usa menos bits pra salvar a informação no destino se nao for encontrada no menor mas for encontrada no maior usa este e salva usando mais bits porque é maior entende? */ /* (momento -> 21/sept/2026 00:11 am) Mas felizmente tem mais, interagindo com Meta AI passei pra ele o lz77_v2 e pasmem ele em questao de minutos propos uma melhoria que simplesmente detonou o nosso lz77_v2 nós chamamos esta nova versão de lz77_v3, ela simplesmente remove o bit de definicao de 9 ou 14 bits permitindo que comprima mais algo que eu e meus amigos nao percebemos durante a criação do lz77_v2 então o merito real do meu ponto de vista é do Meta AI, ok, vamos aos testes... aqui está o makefile_dl pra Linux: #by ric and dua and Meta IA in september 2026 lz77: gcc -O3 lz77_1_original.c -o lz77_orig gcc -O3 lz77-2-v2.c -o lz77_v2 gcc -O3 lz77-3-v3.c -o lz77_v3 ./lz77_orig make.txt out1.bin ./lz77_v2 make.txt out2.bin ./lz77_v3 make.txt out3.bin ls -l out*.bin # já vê o ganho de 2-4% do v3 #by ric and dua and Meta IA in september 2026 lz77: gcc -O3 lz77_1_original.c -o lz77_orig gcc -O3 lz77-2-v2.c -o lz77_v2 gcc -O3 lz77-3-v3.c -o lz77_v3 ./lz77_orig make.bin out1.bin ./lz77_v2 make.bin out2.bin ./lz77_v3 make.bin out3.bin ls -l out*.bin # já vê o ganho de 2-4% do v3 aqui esta as saidas dos programas de test rodando: ~~~~~~~~------------------------------------------ ./lz77_orig make.bin out1.bin --- LZ77 ORIGINAL - 1 JANELA 14b --- Literais: 33668 | Matches: 25280 Original: 166400 bytes Final: 110557 bytes (884452 bits) Taxa: 66.44% | Compressao: 33.56% menor ./lz77_v2 make.bin out2.bin --- LZ77 V2 - 2 JANELAS COM FLAG 9/14b --- Literais: 33668 | Small 9b: 5601 | Large 14b: 19679 Original: 166400 bytes Final: 110216 bytes (881727 bits) Taxa: 66.24% | Compressao: 33.76% menor ./lz77_v3 make.bin out3.bin --- LZ77 V3 - 2 JANELAS SEM FLAG (BIT IMPLICITO) --- Literais: 33668 | Small 9b: 5601 | Large 14b: 19679 Original: 166400 bytes Final: 107056 bytes (856447 bits) Taxa: 64.34% | Compressao: 35.66% menor Economia bit implicito: 25280 bits vs V2 ls -l out*.bin -rw-r--r-- 1 email email 110557 Sep 21 00:19 out1.bin -rw-r--r-- 1 email email 110216 Sep 21 00:19 out2.bin -rw-r--r-- 1 email email 107056 Sep 21 00:19 out3.bin //////////////////////////////////////////////// ./lz77_orig make.txt out1.bin --- LZ77 ORIGINAL - 1 JANELA 14b --- Literais: 10582 | Matches: 55900 Original: 327528 bytes Final: 172618 bytes (1380938 bits) Taxa: 52.70% | Compressao: 47.30% menor ./lz77_v2 make.txt out2.bin --- LZ77 V2 - 2 JANELAS COM FLAG 9/14b --- Literais: 10582 | Small 9b: 2776 | Large 14b: 53124 Original: 327528 bytes Final: 177870 bytes (1422958 bits) Taxa: 54.31% | Compressao: 45.69% menor ./lz77_v3 make.txt out3.bin --- LZ77 V3 - 2 JANELAS SEM FLAG (BIT IMPLICITO) --- Literais: 10582 | Small 9b: 2776 | Large 14b: 53124 Original: 327528 bytes Final: 170883 bytes (1367058 bits) Taxa: 52.17% | Compressao: 47.83% menor Economia bit implicito: 55900 bits vs V2 ls -l out*.bin -rw-r--r-- 1 email email 172618 Sep 21 00:16 out1.bin -rw-r--r-- 1 email email 177870 Sep 21 00:16 out2.bin -rw-r--r-- 1 email email 170883 Sep 21 00:16 out3.bin ///////////////////////////////////////////////////// O grande vencedor aqui nao o o v2 mas o poderoso v3 e claro o v2 deve ser descartado, correto... */
Other Useful Business Software
Build Agents and Models on One Platform
Gemini Enterprise Agent Platform is Google Cloud's comprehensive platform for developers to build, scale, govern, and optimize agents and models. Choose from Google's most advanced models and third-party models like Anthropic's Claude Model Family.
Veeam Data Platform v13.1
Try Veeam Data Platform today. Experience the unified platform that's secure by design, portable by default, and proven to recover clean, fast, and anywhere.
