Skip to content

Releases: SamDreamsMaker/Max-Compression

v2.2.0

24 Mar 17:23
@github-actions github-actions
v2.2.0
f1ba959

Choose a tag to compare

View all tags
v2.2.0 Latest
Latest

MaxCompression v2.2.0

Downloads

Platform File
Linux x64 mcx-linux-x64.tar.gz
macOS ARM64 mcx-macos-arm64.tar.gz
Windows x64 mcx-windows-x64.zip

Quick Start

# Linux/macOS
tar xzf mcx-*.tar.gz
./mcx-*/mcx compress myfile.txt

# Build from source
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)

See CHANGELOG.md for full details.

Full Changelog: v2.1.1...v2.2.0

Assets 5
Loading

MaxCompression v2.1.1

19 Mar 00:05
@github-actions github-actions
v2.1.1
451112b

Choose a tag to compare

View all tags
MaxCompression v2.1.1

Changelog

All notable changes to MaxCompression are documented in this file.

[2.1.1] — 2026-03-19

Added

  • mcx test self-test command — verifies all compression levels in one command.
  • -q/--quiet flag — suppress non-error output for scripting.
  • Edge case tests — boundary conditions, patterns, frame info validation.
  • GitHub issue templates (bug report + feature request).
  • SECURITY.md — vulnerability disclosure policy.
  • .editorconfig — consistent code style across editors.
  • CI status badge and version badge in README.

Changed

  • rANS entropy for LZ blocks — 0.5-2% smaller than FSE at L1-L9.
    • alice29 L6: 65739→63814 (-2.9%), L9: 64442→62728 (-2.9%).
  • AAC enabled at L7+ (was L9+) — better ratio at moderate speed.
  • Larger HC hash table (1M entries) — L6 -1.4%, L9 -0.9%.
  • Better greedy hash populating — L1 -3.7% ratio improvement.
  • Scaled LZ hash table with input size — +65% decompress speed on small files.
  • rANS decoder unrolled — 2 symbols per iteration.
  • BWT inverse prefetch for blocks >256KB (+5% decompress on large files).
  • RC_UNLIKELY branch hints for range coder normalize.
  • Updated ROADMAP.md with rejected experiments table.
  • Updated API.md with mcx_get_frame_info documentation.
  • Updated man page with -q/--quiet flag and L24 level.

[2.1.0] — 2026-03-18

Added

  • mcx_get_frame_info() public API — read frame metadata without decompressing.
  • Fuzz roundtrip test — 1000 random inputs across all levels, catches edge cases.
  • Malformed input test — 7 decompressor robustness tests (corrupt data, truncated, wrong magic).
  • Comparison benchmark script (benchmarks/compare.sh) — MCX vs gzip/bzip2/xz.
  • Doxyfile for API documentation generation.
  • pkg-config template (maxcomp.pc.in).

Changed

  • Embedded libdivsufsort — BWT forward transform 2× faster (Yuta Mori, MIT license).
    • alice29: 0.08→0.035s, dickens: 8.76→4.43s, ooffice: 4.0→1.91s.
  • LZRC fast mode (L24) — hash chain match finder, ~3× faster than L26 BT.
  • CLI info command now uses public mcx_get_frame_info() API.
  • CLI help shows level descriptions and L24 example.
  • README updated with v2.1.0 benchmarks and LZRC/divsufsort features.

Fixed

  • MSVC compatibility for divsufsort (disabled strings.h, use __inline).

Rejected experiments

  • 4-state LZRC machine (model dilution, no gain on any file).

[2.0.1] — 2026-03-18

Added

  • mcx bench <file> command — benchmark all compression levels with speed/ratio comparison.
  • Man page (docs/mcx.1).
  • Hash chain match finder (hc_match.c) — 4× faster alternative for future speed modes.
  • PyPI publish workflow (.github/workflows/pypi.yml) + pyproject.toml.
  • L26 (LZRC) roundtrip test in CI.
  • Python bindings: version() and compress_bound() functions.

Changed

  • MTF decode +67% faster — replaced byte loop with memmove (29→49 MB/s).
  • RC decoder inline — decompress +19–36% (cross-TU inlining of hot path).
  • RC encoder inline — compress +57–113%.
  • BT depth 64→32 — same ratio, ~7% faster compression.
  • Portable bench timing (Windows QueryPerformanceCounter, POSIX clock_gettime).

Fixed

  • CI: Windows __builtin_ctzll → mcx_ctzll() portability.
  • CI: OpenMP optional (find_package(QUIET)).
  • CI: _GNU_SOURCE for bench_fast.c.
  • CI: Missing preprocess.h include in genetic.c.

[2.0.0] — 2026-03-18

Added

  • LZRC v2.0 engine — LZ + adaptive range coder with binary tree match finder (16 MB window).
    • Lazy evaluation: universal +1–5% improvement.
    • 4 rep distances (rep0–rep3): cheaper repeat match encoding.
    • LZMA-style matched literal coding: bit-level prediction from match reference.
    • Integrated as block type 0xB0 (L26) + multi-trial at L20.
    • Binary data auto-routed to LZRC at L20+ — mozilla: 2.93× → 3.22× (+10%).
  • Smart routing: text→BWT, binary→LZRC, stride→delta, x86→E8/E9.
  • GitHub Actions CI: Linux (GCC + Clang), macOS, Windows.
  • Automated release workflow with pre-built binaries.
  • CONTRIBUTING.md, docs/DESIGN.md, Doxyfile for API documentation.
  • Python bindings packaging (setup.py).
  • SemVer tagging (v1.9.3, v2.0.0).

Changed

  • Silesia total: 4.21× → 4.35× (+3.3%) thanks to LZRC routing.
  • MCX L20 now beats xz on 7/12 Silesia files (was 6/12).
  • Total competitive with xz -9 (~4.35× vs ~4.34×).

Fixed

  • Cross-platform portability: mcx_ctzll() for MSVC, optional OpenMP.
  • Implicit function declarations fixed for Clang strict mode.
  • _GNU_SOURCE added to bench_fast.c for CLOCK_MONOTONIC.

[1.9.3] — 2026-03-18

Added

  • LZRC v2.0 prototype — LZ + range coder with binary tree match finder (16 MB window). mozilla: 3.07× (best LZ result, +5% vs BWT L20).
  • Adaptive arithmetic coding (AAC) on LZ output at L9+ — order-1 model with Fenwick-tree accelerated decoding.
  • 64 MB block size — up from 32 MB. webster +2.1%, mozilla +0.3%.

Fixed

  • L12 genome optimizer — was skipping BWT on binary files. Silesia L12 total: 2.91× → 4.16× (+43%). nci: 3.29× → 25.65× (+680%).
  • CLI version string now derived from header constants (was hardcoded).
  • Fenwick tree decoder — merged find + query into single tree walk. Decompress +2–5%.

Changed

  • AAC only active at L9+ (was L6+). L6 keeps FSE for faster decompression (19 MB/s vs 3 MB/s).
  • Multi-rANS speed: precomputed log2 LUT + sparse symbol lists + uint16 freqs. 41.5s → 12.1s on mr (-71%).

[1.9.2] — 2026-03-18

Added

  • Adaptive arithmetic coding (AAC) — order-1 AC for LZ output. kennedy L9: 4.87× → 9.04× (+86%).
  • E8/E9 x86 filter — auto-detects executables (≥0.5% E8/E9 opcodes). ooffice: 2.18× → 2.53× (+16%).

[1.9.1] — 2026-03-17

Added

  • LZ-HC hash chains (depth 8/16). alice29 L9: 2.14× → 2.31× (+8%).
  • CLI uses one-shot API for files ≤256 MB. alice29 CLI: 1.91× → 3.53×.

Changed

  • Block size: 32 MB (up from 16 MB).

[1.9.0] — 2026-03-17

Added

  • E8/E9 x86 filter with auto-detection and multi-trial.
  • 32 MB blocks.

[1.8.1] — 2026-03-17

Added

  • Multi-rANS speed optimizations: precomputed LUT, sparse active lists, uint16 group freqs. Total: 41.5s → 12.1s (-71%).

[1.8.0] — 2026-03-17

Added

  • Sequential K-means initialization for multi-table rANS. alice29: 43268 → 43144 bytes. Beats bzip2 (43207).
  • 15 K-means iterations (was 10).

[1.7.8] — 2026-03-17

Added

  • Varint frequency tables for multi-rANS. Saves ~300 bytes per block. kennedy: 46.92× → 48.73×.
  • Adaptive table count (4, 5, or 6 tables, keeps smallest).

[1.7.7] — 2026-03-17

Added

  • Bitmap table format for multi-rANS (32-byte bitmap + varint freqs per active symbol).

[1.7.5] — 2026-03-17

Changed

  • Block size: 16 MB (up from 8 MB). Beats bzip2 on 11/12 Silesia files.

[1.7.1] — 2026-03-17

Fixed

  • Route ALL data types to BWT first (binary was incorrectly routed to LZ24). xml: 6.80× → 12.45× (+83%).

[1.7.0] — 2026-03-17

Fixed

  • Text routing: removed 4 MB → LZ24 rule. dickens: 2.21× → 3.69× (+67%).

[1.6.0] — 2026-03-17

Added

  • Multi-table rANS — 4 frequency tables with K-means clustering, group size 50. Near bzip2 on text.

[1.5.1] — 2026-03-17

Added

  • RLE2 on stride-delta output. ptt5: 8.83× → 10.19× (+15%).

[1.5.0] — 2026-03-16

Added

  • Delta-fix — forced delta=0 for text. +13–35% on ALL text files.
  • Auto-RLE2 for all BWT levels (L10+). L12 = L20 on text.
  • BWT threshold lowered 8 KB → 1 KB.
  • rANS precision bumped to 14-bit.

[1.4.0] — 2026-03-16

Added

  • Stride-Delta + BWT + RLE2 pipeline. kennedy.xls: 46.91× (2.2× better than xz).

[1.3.0] — 2026-03-16

Added

  • RLE2 (RUNA/RUNB) — exponential zero-run encoding. +5–7% on text.
  • Smart Mode (L20+) with auto data-type routing.

[1.2.0] — 2026-03-16

Added

  • LZ24 (16 MB window, hash chains, lazy evaluation).
  • Stride-delta transform with auto-detection.

[1.1.0] — 2026-03-14

Added

  • 4-stream interleaved tANS (+64% decompress speed).
  • CM-rANS (order-1 context-mixing entropy coder).

[1.0.0] — 2026-03-12

Added

  • Initial release.
  • LZ77 (greedy + lazy), BWT (SA-IS), tANS/FSE, rANS.
  • Genetic pipeline optimizer.
  • Multi-stream LZ77 with repcode stack.
  • SIMD SSE4.1 hash computation.
Loading

MaxCompression v2.1.0

18 Mar 21:24
@github-actions github-actions
v2.1.0
607e683

Choose a tag to compare

View all tags
MaxCompression v2.1.0

Changelog

All notable changes to MaxCompression are documented in this file.

[2.1.0] — 2026-03-18

Added

  • mcx_get_frame_info() public API — read frame metadata without decompressing.
  • Fuzz roundtrip test — 1000 random inputs across all levels, catches edge cases.
  • Malformed input test — 7 decompressor robustness tests (corrupt data, truncated, wrong magic).
  • Comparison benchmark script (benchmarks/compare.sh) — MCX vs gzip/bzip2/xz.
  • Doxyfile for API documentation generation.
  • pkg-config template (maxcomp.pc.in).

Changed

  • Embedded libdivsufsort — BWT forward transform 2× faster (Yuta Mori, MIT license).
    • alice29: 0.08→0.035s, dickens: 8.76→4.43s, ooffice: 4.0→1.91s.
  • LZRC fast mode (L24) — hash chain match finder, ~3× faster than L26 BT.
  • CLI info command now uses public mcx_get_frame_info() API.
  • CLI help shows level descriptions and L24 example.
  • README updated with v2.1.0 benchmarks and LZRC/divsufsort features.

Fixed

  • MSVC compatibility for divsufsort (disabled strings.h, use __inline).

Rejected experiments

  • 4-state LZRC machine (model dilution, no gain on any file).

[2.0.1] — 2026-03-18

Added

  • mcx bench <file> command — benchmark all compression levels with speed/ratio comparison.
  • Man page (docs/mcx.1).
  • Hash chain match finder (hc_match.c) — 4× faster alternative for future speed modes.
  • PyPI publish workflow (.github/workflows/pypi.yml) + pyproject.toml.
  • L26 (LZRC) roundtrip test in CI.
  • Python bindings: version() and compress_bound() functions.

Changed

  • MTF decode +67% faster — replaced byte loop with memmove (29→49 MB/s).
  • RC decoder inline — decompress +19–36% (cross-TU inlining of hot path).
  • RC encoder inline — compress +57–113%.
  • BT depth 64→32 — same ratio, ~7% faster compression.
  • Portable bench timing (Windows QueryPerformanceCounter, POSIX clock_gettime).

Fixed

  • CI: Windows __builtin_ctzll → mcx_ctzll() portability.
  • CI: OpenMP optional (find_package(QUIET)).
  • CI: _GNU_SOURCE for bench_fast.c.
  • CI: Missing preprocess.h include in genetic.c.

[2.0.0] — 2026-03-18

Added

  • LZRC v2.0 engine — LZ + adaptive range coder with binary tree match finder (16 MB window).
    • Lazy evaluation: universal +1–5% improvement.
    • 4 rep distances (rep0–rep3): cheaper repeat match encoding.
    • LZMA-style matched literal coding: bit-level prediction from match reference.
    • Integrated as block type 0xB0 (L26) + multi-trial at L20.
    • Binary data auto-routed to LZRC at L20+ — mozilla: 2.93× → 3.22× (+10%).
  • Smart routing: text→BWT, binary→LZRC, stride→delta, x86→E8/E9.
  • GitHub Actions CI: Linux (GCC + Clang), macOS, Windows.
  • Automated release workflow with pre-built binaries.
  • CONTRIBUTING.md, docs/DESIGN.md, Doxyfile for API documentation.
  • Python bindings packaging (setup.py).
  • SemVer tagging (v1.9.3, v2.0.0).

Changed

  • Silesia total: 4.21× → 4.35× (+3.3%) thanks to LZRC routing.
  • MCX L20 now beats xz on 7/12 Silesia files (was 6/12).
  • Total competitive with xz -9 (~4.35× vs ~4.34×).

Fixed

  • Cross-platform portability: mcx_ctzll() for MSVC, optional OpenMP.
  • Implicit function declarations fixed for Clang strict mode.
  • _GNU_SOURCE added to bench_fast.c for CLOCK_MONOTONIC.

[1.9.3] — 2026-03-18

Added

  • LZRC v2.0 prototype — LZ + range coder with binary tree match finder (16 MB window). mozilla: 3.07× (best LZ result, +5% vs BWT L20).
  • Adaptive arithmetic coding (AAC) on LZ output at L9+ — order-1 model with Fenwick-tree accelerated decoding.
  • 64 MB block size — up from 32 MB. webster +2.1%, mozilla +0.3%.

Fixed

  • L12 genome optimizer — was skipping BWT on binary files. Silesia L12 total: 2.91× → 4.16× (+43%). nci: 3.29× → 25.65× (+680%).
  • CLI version string now derived from header constants (was hardcoded).
  • Fenwick tree decoder — merged find + query into single tree walk. Decompress +2–5%.

Changed

  • AAC only active at L9+ (was L6+). L6 keeps FSE for faster decompression (19 MB/s vs 3 MB/s).
  • Multi-rANS speed: precomputed log2 LUT + sparse symbol lists + uint16 freqs. 41.5s → 12.1s on mr (-71%).

[1.9.2] — 2026-03-18

Added

  • Adaptive arithmetic coding (AAC) — order-1 AC for LZ output. kennedy L9: 4.87× → 9.04× (+86%).
  • E8/E9 x86 filter — auto-detects executables (≥0.5% E8/E9 opcodes). ooffice: 2.18× → 2.53× (+16%).

[1.9.1] — 2026-03-17

Added

  • LZ-HC hash chains (depth 8/16). alice29 L9: 2.14× → 2.31× (+8%).
  • CLI uses one-shot API for files ≤256 MB. alice29 CLI: 1.91× → 3.53×.

Changed

  • Block size: 32 MB (up from 16 MB).

[1.9.0] — 2026-03-17

Added

  • E8/E9 x86 filter with auto-detection and multi-trial.
  • 32 MB blocks.

[1.8.1] — 2026-03-17

Added

  • Multi-rANS speed optimizations: precomputed LUT, sparse active lists, uint16 group freqs. Total: 41.5s → 12.1s (-71%).

[1.8.0] — 2026-03-17

Added

  • Sequential K-means initialization for multi-table rANS. alice29: 43268 → 43144 bytes. Beats bzip2 (43207).
  • 15 K-means iterations (was 10).

[1.7.8] — 2026-03-17

Added

  • Varint frequency tables for multi-rANS. Saves ~300 bytes per block. kennedy: 46.92× → 48.73×.
  • Adaptive table count (4, 5, or 6 tables, keeps smallest).

[1.7.7] — 2026-03-17

Added

  • Bitmap table format for multi-rANS (32-byte bitmap + varint freqs per active symbol).

[1.7.5] — 2026-03-17

Changed

  • Block size: 16 MB (up from 8 MB). Beats bzip2 on 11/12 Silesia files.

[1.7.1] — 2026-03-17

Fixed

  • Route ALL data types to BWT first (binary was incorrectly routed to LZ24). xml: 6.80× → 12.45× (+83%).

[1.7.0] — 2026-03-17

Fixed

  • Text routing: removed 4 MB → LZ24 rule. dickens: 2.21× → 3.69× (+67%).

[1.6.0] — 2026-03-17

Added

  • Multi-table rANS — 4 frequency tables with K-means clustering, group size 50. Near bzip2 on text.

[1.5.1] — 2026-03-17

Added

  • RLE2 on stride-delta output. ptt5: 8.83× → 10.19× (+15%).

[1.5.0] — 2026-03-16

Added

  • Delta-fix — forced delta=0 for text. +13–35% on ALL text files.
  • Auto-RLE2 for all BWT levels (L10+). L12 = L20 on text.
  • BWT threshold lowered 8 KB → 1 KB.
  • rANS precision bumped to 14-bit.

[1.4.0] — 2026-03-16

Added

  • Stride-Delta + BWT + RLE2 pipeline. kennedy.xls: 46.91× (2.2× better than xz).

[1.3.0] — 2026-03-16

Added

  • RLE2 (RUNA/RUNB) — exponential zero-run encoding. +5–7% on text.
  • Smart Mode (L20+) with auto data-type routing.

[1.2.0] — 2026-03-16

Added

  • LZ24 (16 MB window, hash chains, lazy evaluation).
  • Stride-delta transform with auto-detection.

[1.1.0] — 2026-03-14

Added

  • 4-stream interleaved tANS (+64% decompress speed).
  • CM-rANS (order-1 context-mixing entropy coder).

[1.0.0] — 2026-03-12

Added

  • Initial release.
  • LZ77 (greedy + lazy), BWT (SA-IS), tANS/FSE, rANS.
  • Genetic pipeline optimizer.
  • Multi-stream LZ77 with repcode stack.
  • SIMD SSE4.1 hash computation.
Loading

MaxCompression v2.0.1

18 Mar 17:01
@github-actions github-actions
v2.0.1
b70e4ea

Choose a tag to compare

View all tags
MaxCompression v2.0.1

Changelog

All notable changes to MaxCompression are documented in this file.

[2.0.1] — 2026-03-18

Added

  • mcx bench <file> command — benchmark all compression levels with speed/ratio comparison.
  • Man page (docs/mcx.1).
  • Hash chain match finder (hc_match.c) — 4× faster alternative for future speed modes.
  • PyPI publish workflow (.github/workflows/pypi.yml) + pyproject.toml.
  • L26 (LZRC) roundtrip test in CI.
  • Python bindings: version() and compress_bound() functions.

Changed

  • MTF decode +67% faster — replaced byte loop with memmove (29→49 MB/s).
  • RC decoder inline — decompress +19–36% (cross-TU inlining of hot path).
  • RC encoder inline — compress +57–113%.
  • BT depth 64→32 — same ratio, ~7% faster compression.
  • Portable bench timing (Windows QueryPerformanceCounter, POSIX clock_gettime).

Fixed

  • CI: Windows __builtin_ctzll → mcx_ctzll() portability.
  • CI: OpenMP optional (find_package(QUIET)).
  • CI: _GNU_SOURCE for bench_fast.c.
  • CI: Missing preprocess.h include in genetic.c.

[2.0.0] — 2026-03-18

Added

  • LZRC v2.0 engine — LZ + adaptive range coder with binary tree match finder (16 MB window).
    • Lazy evaluation: universal +1–5% improvement.
    • 4 rep distances (rep0–rep3): cheaper repeat match encoding.
    • LZMA-style matched literal coding: bit-level prediction from match reference.
    • Integrated as block type 0xB0 (L26) + multi-trial at L20.
    • Binary data auto-routed to LZRC at L20+ — mozilla: 2.93× → 3.22× (+10%).
  • Smart routing: text→BWT, binary→LZRC, stride→delta, x86→E8/E9.
  • GitHub Actions CI: Linux (GCC + Clang), macOS, Windows.
  • Automated release workflow with pre-built binaries.
  • CONTRIBUTING.md, docs/DESIGN.md, Doxyfile for API documentation.
  • Python bindings packaging (setup.py).
  • SemVer tagging (v1.9.3, v2.0.0).

Changed

  • Silesia total: 4.21× → 4.35× (+3.3%) thanks to LZRC routing.
  • MCX L20 now beats xz on 7/12 Silesia files (was 6/12).
  • Total competitive with xz -9 (~4.35× vs ~4.34×).

Fixed

  • Cross-platform portability: mcx_ctzll() for MSVC, optional OpenMP.
  • Implicit function declarations fixed for Clang strict mode.
  • _GNU_SOURCE added to bench_fast.c for CLOCK_MONOTONIC.

[1.9.3] — 2026-03-18

Added

  • LZRC v2.0 prototype — LZ + range coder with binary tree match finder (16 MB window). mozilla: 3.07× (best LZ result, +5% vs BWT L20).
  • Adaptive arithmetic coding (AAC) on LZ output at L9+ — order-1 model with Fenwick-tree accelerated decoding.
  • 64 MB block size — up from 32 MB. webster +2.1%, mozilla +0.3%.

Fixed

  • L12 genome optimizer — was skipping BWT on binary files. Silesia L12 total: 2.91× → 4.16× (+43%). nci: 3.29× → 25.65× (+680%).
  • CLI version string now derived from header constants (was hardcoded).
  • Fenwick tree decoder — merged find + query into single tree walk. Decompress +2–5%.

Changed

  • AAC only active at L9+ (was L6+). L6 keeps FSE for faster decompression (19 MB/s vs 3 MB/s).
  • Multi-rANS speed: precomputed log2 LUT + sparse symbol lists + uint16 freqs. 41.5s → 12.1s on mr (-71%).

[1.9.2] — 2026-03-18

Added

  • Adaptive arithmetic coding (AAC) — order-1 AC for LZ output. kennedy L9: 4.87× → 9.04× (+86%).
  • E8/E9 x86 filter — auto-detects executables (≥0.5% E8/E9 opcodes). ooffice: 2.18× → 2.53× (+16%).

[1.9.1] — 2026-03-17

Added

  • LZ-HC hash chains (depth 8/16). alice29 L9: 2.14× → 2.31× (+8%).
  • CLI uses one-shot API for files ≤256 MB. alice29 CLI: 1.91× → 3.53×.

Changed

  • Block size: 32 MB (up from 16 MB).

[1.9.0] — 2026-03-17

Added

  • E8/E9 x86 filter with auto-detection and multi-trial.
  • 32 MB blocks.

[1.8.1] — 2026-03-17

Added

  • Multi-rANS speed optimizations: precomputed LUT, sparse active lists, uint16 group freqs. Total: 41.5s → 12.1s (-71%).

[1.8.0] — 2026-03-17

Added

  • Sequential K-means initialization for multi-table rANS. alice29: 43268 → 43144 bytes. Beats bzip2 (43207).
  • 15 K-means iterations (was 10).

[1.7.8] — 2026-03-17

Added

  • Varint frequency tables for multi-rANS. Saves ~300 bytes per block. kennedy: 46.92× → 48.73×.
  • Adaptive table count (4, 5, or 6 tables, keeps smallest).

[1.7.7] — 2026-03-17

Added

  • Bitmap table format for multi-rANS (32-byte bitmap + varint freqs per active symbol).

[1.7.5] — 2026-03-17

Changed

  • Block size: 16 MB (up from 8 MB). Beats bzip2 on 11/12 Silesia files.

[1.7.1] — 2026-03-17

Fixed

  • Route ALL data types to BWT first (binary was incorrectly routed to LZ24). xml: 6.80× → 12.45× (+83%).

[1.7.0] — 2026-03-17

Fixed

  • Text routing: removed 4 MB → LZ24 rule. dickens: 2.21× → 3.69× (+67%).

[1.6.0] — 2026-03-17

Added

  • Multi-table rANS — 4 frequency tables with K-means clustering, group size 50. Near bzip2 on text.

[1.5.1] — 2026-03-17

Added

  • RLE2 on stride-delta output. ptt5: 8.83× → 10.19× (+15%).

[1.5.0] — 2026-03-16

Added

  • Delta-fix — forced delta=0 for text. +13–35% on ALL text files.
  • Auto-RLE2 for all BWT levels (L10+). L12 = L20 on text.
  • BWT threshold lowered 8 KB → 1 KB.
  • rANS precision bumped to 14-bit.

[1.4.0] — 2026-03-16

Added

  • Stride-Delta + BWT + RLE2 pipeline. kennedy.xls: 46.91× (2.2× better than xz).

[1.3.0] — 2026-03-16

Added

  • RLE2 (RUNA/RUNB) — exponential zero-run encoding. +5–7% on text.
  • Smart Mode (L20+) with auto data-type routing.

[1.2.0] — 2026-03-16

Added

  • LZ24 (16 MB window, hash chains, lazy evaluation).
  • Stride-delta transform with auto-detection.

[1.1.0] — 2026-03-14

Added

  • 4-stream interleaved tANS (+64% decompress speed).
  • CM-rANS (order-1 context-mixing entropy coder).

[1.0.0] — 2026-03-12

Added

  • Initial release.
  • LZ77 (greedy + lazy), BWT (SA-IS), tANS/FSE, rANS.
  • Genetic pipeline optimizer.
  • Multi-stream LZ77 with repcode stack.
  • SIMD SSE4.1 hash computation.
Loading

MaxCompression v2.0.0

18 Mar 15:29
@github-actions github-actions
v2.0.0
0448c83

Choose a tag to compare

View all tags
MaxCompression v2.0.0

Changelog

All notable changes to MaxCompression are documented in this file.

[2.0.0] — 2026-03-18

Added

  • LZRC v2.0 engine — LZ + adaptive range coder with binary tree match finder (16 MB window).
    • Lazy evaluation: universal +1–5% improvement.
    • 4 rep distances (rep0–rep3): cheaper repeat match encoding.
    • LZMA-style matched literal coding: bit-level prediction from match reference.
    • Integrated as block type 0xB0 (L26) + multi-trial at L20.
    • Binary data auto-routed to LZRC at L20+ — mozilla: 2.93× → 3.22× (+10%).
  • Smart routing: text→BWT, binary→LZRC, stride→delta, x86→E8/E9.
  • GitHub Actions CI: Linux (GCC + Clang), macOS, Windows.
  • Automated release workflow with pre-built binaries.
  • CONTRIBUTING.md, docs/DESIGN.md, Doxyfile for API documentation.
  • Python bindings packaging (setup.py).
  • SemVer tagging (v1.9.3, v2.0.0).

Changed

  • Silesia total: 4.21× → 4.35× (+3.3%) thanks to LZRC routing.
  • MCX L20 now beats xz on 7/12 Silesia files (was 6/12).
  • Total competitive with xz -9 (~4.35× vs ~4.34×).

Fixed

  • Cross-platform portability: mcx_ctzll() for MSVC, optional OpenMP.
  • Implicit function declarations fixed for Clang strict mode.
  • _GNU_SOURCE added to bench_fast.c for CLOCK_MONOTONIC.

[1.9.3] — 2026-03-18

Added

  • LZRC v2.0 prototype — LZ + range coder with binary tree match finder (16 MB window). mozilla: 3.07× (best LZ result, +5% vs BWT L20).
  • Adaptive arithmetic coding (AAC) on LZ output at L9+ — order-1 model with Fenwick-tree accelerated decoding.
  • 64 MB block size — up from 32 MB. webster +2.1%, mozilla +0.3%.

Fixed

  • L12 genome optimizer — was skipping BWT on binary files. Silesia L12 total: 2.91× → 4.16× (+43%). nci: 3.29× → 25.65× (+680%).
  • CLI version string now derived from header constants (was hardcoded).
  • Fenwick tree decoder — merged find + query into single tree walk. Decompress +2–5%.

Changed

  • AAC only active at L9+ (was L6+). L6 keeps FSE for faster decompression (19 MB/s vs 3 MB/s).
  • Multi-rANS speed: precomputed log2 LUT + sparse symbol lists + uint16 freqs. 41.5s → 12.1s on mr (-71%).

[1.9.2] — 2026-03-18

Added

  • Adaptive arithmetic coding (AAC) — order-1 AC for LZ output. kennedy L9: 4.87× → 9.04× (+86%).
  • E8/E9 x86 filter — auto-detects executables (≥0.5% E8/E9 opcodes). ooffice: 2.18× → 2.53× (+16%).

[1.9.1] — 2026-03-17

Added

  • LZ-HC hash chains (depth 8/16). alice29 L9: 2.14× → 2.31× (+8%).
  • CLI uses one-shot API for files ≤256 MB. alice29 CLI: 1.91× → 3.53×.

Changed

  • Block size: 32 MB (up from 16 MB).

[1.9.0] — 2026-03-17

Added

  • E8/E9 x86 filter with auto-detection and multi-trial.
  • 32 MB blocks.

[1.8.1] — 2026-03-17

Added

  • Multi-rANS speed optimizations: precomputed LUT, sparse active lists, uint16 group freqs. Total: 41.5s → 12.1s (-71%).

[1.8.0] — 2026-03-17

Added

  • Sequential K-means initialization for multi-table rANS. alice29: 43268 → 43144 bytes. Beats bzip2 (43207).
  • 15 K-means iterations (was 10).

[1.7.8] — 2026-03-17

Added

  • Varint frequency tables for multi-rANS. Saves ~300 bytes per block. kennedy: 46.92× → 48.73×.
  • Adaptive table count (4, 5, or 6 tables, keeps smallest).

[1.7.7] — 2026-03-17

Added

  • Bitmap table format for multi-rANS (32-byte bitmap + varint freqs per active symbol).

[1.7.5] — 2026-03-17

Changed

  • Block size: 16 MB (up from 8 MB). Beats bzip2 on 11/12 Silesia files.

[1.7.1] — 2026-03-17

Fixed

  • Route ALL data types to BWT first (binary was incorrectly routed to LZ24). xml: 6.80× → 12.45× (+83%).

[1.7.0] — 2026-03-17

Fixed

  • Text routing: removed 4 MB → LZ24 rule. dickens: 2.21× → 3.69× (+67%).

[1.6.0] — 2026-03-17

Added

  • Multi-table rANS — 4 frequency tables with K-means clustering, group size 50. Near bzip2 on text.

[1.5.1] — 2026-03-17

Added

  • RLE2 on stride-delta output. ptt5: 8.83× → 10.19× (+15%).

[1.5.0] — 2026-03-16

Added

  • Delta-fix — forced delta=0 for text. +13–35% on ALL text files.
  • Auto-RLE2 for all BWT levels (L10+). L12 = L20 on text.
  • BWT threshold lowered 8 KB → 1 KB.
  • rANS precision bumped to 14-bit.

[1.4.0] — 2026-03-16

Added

  • Stride-Delta + BWT + RLE2 pipeline. kennedy.xls: 46.91× (2.2× better than xz).

[1.3.0] — 2026-03-16

Added

  • RLE2 (RUNA/RUNB) — exponential zero-run encoding. +5–7% on text.
  • Smart Mode (L20+) with auto data-type routing.

[1.2.0] — 2026-03-16

Added

  • LZ24 (16 MB window, hash chains, lazy evaluation).
  • Stride-delta transform with auto-detection.

[1.1.0] — 2026-03-14

Added

  • 4-stream interleaved tANS (+64% decompress speed).
  • CM-rANS (order-1 context-mixing entropy coder).

[1.0.0] — 2026-03-12

Added

  • Initial release.
  • LZ77 (greedy + lazy), BWT (SA-IS), tANS/FSE, rANS.
  • Genetic pipeline optimizer.
  • Multi-stream LZ77 with repcode stack.
  • SIMD SSE4.1 hash computation.
Loading