Skip to content

compression.md

Latest commit

 

History

History
57 lines (37 loc) · 3.32 KB

compression.md

File metadata and controls

57 lines (37 loc) · 3.32 KB
Raw

Compression And Codecs

Engine Codec Flags

IDA analysis of sub_AA9AC0 identifies these archive-level codec selections:

Flag Codec ID Name C++ GUI/CLI Older Python reference
0x08 0xC5735A0F ZLib Supported using local zlib1.dll Supported using Python zlib
0x10 0xC5735A10 LZX Unsupported; skip or preserve raw Experimental 7-Zip CAB bridge, unverified on real game data
0x20 0xC5735A11 LZD Native support Python support

The game diagnostic sub_A767C0 names ZLib, LZX, and LZD. Codec selection is archive-level. Audio.dsPack advertises ZLib, but inspected entries have equal stored/logical sizes and are copied as stored data; audio codecs are not decoded here.

ZLib

The engine path sub_A50420 -> sub_E5A3A0 -> sub_F023C0 sends each block to zlib 1.2.3 with the expected output size. It uses the regular zlib wrapper rather than raw DEFLATE. C++ dynamically loads the copied dspack_cpp_runtime/zlib1.dll; Python uses its standard library. Both validate output size. Python additionally checks EOF and trailing data.

ZLib may be raw per-resource or inside StreamCompression. Both implementations validate the wrapper checksum and block ranges and inflate each indexed block to its declared size.

LZD

The C++ and Python decoders are based on engine functions sub_B44930, sub_714AC0, sub_4694F0, sub_65C290, sub_7D7630, sub_AE3D50, and sub_78A130.

The decoder builds a symbol tree and 10-bit lookup table, reads big-endian 32-bit bitstream words (low-order bits consumed first), then executes literal runs and three match forms. Back-reference copies can overlap. Output must match the logical size from the archive index.

StreamCompression Wrapper

The game parser sub_839990 reads a 32-byte header and 12-byte block records; sub_A50420 seeks to blocks and decodes them.

Both tools interpret eight big-endian DWORDs:

Offset Meaning
+0x00 Checksum
+0x04 Codec ID
+0x08 Block count
+0x0C Total logical size
+0x10 Max input/buffer field, included in checksum
+0x14 Logical block size
+0x18 Auxiliary DWORD, included in checksum
+0x1C Trailing DWORD, ignored

Each 12-byte big-endian record contains payload offset, stored size, and a reserved/unknown DWORD. Checksum starts at 0x811C9DC5; header DWORDs 1-6 and all record DWORDs are folded as hash = (hash * 16777619) XOR word modulo 32 bits. Recovered from sub_D7AD60 and sub_8986D0, then validated on installed resources.

LZX: Experimental Python Bridge Only

The game uses LZX with a 1-MiB window and FEAR-specific framing. sub_F0C650 parses a two-byte compressed size for 32-KiB output or an extended FF + output size + compressed size header; the stream reserves a five-byte trailer.

The old Python reference wraps frames as a synthetic CAB (LZX:20) and invokes 7-Zip. No installed archive contains an LZX entry, so this has not been validated on actual FEAR 3 data. C++ intentionally does not attempt the bridge: LZX entries are skipped or preserved raw.

Practical Support

  • C++: LZD and ZLib; no LZX.
  • Python reference: LZD, ZLib, and an experimental/unverified LZX bridge.
  • --raw-compressed preserves stored bytes as .packed, not decoded assets.
  • Neither program creates or recompresses archives.