Entry 7

CCSF chunk dispatch and the real texture chunk length

2026-09-22formatdecomprender

Filestools/ccs.pydocs/formats/ccs.md

3 guessed that a texture chunk's size field counts something the file does not store. That is the wrong way round: the game never uses a chunk's size field to find the next chunk, and for textures it simply disagrees with the data. 3 also guessed the header's first word was a version and the next a frame count; the first is now confirmed and re-typed as a u16.

The dispatcher

The loader is ccStream, in D:\usr\RpgUS\prog\system\libccs2.cpp (6). ccStream::DecodeSetupSection (INF SLUS_202.67:0x00149ff0) loops: read u16 kind, u16 (the 0xcccc half, ignored), u32 size from the ring buffer, then call one decoder by kind. Each decoder reads exactly the fields it needs; nothing afterwards skips to size. A kind it does not know executes sw $zero, 0($zero) - a deliberate crash. The table, from the compare chain:

Because only the low half is compared, the Setup chunk being stored as 0x00000003 rather than 0xcccc0003 is harmless.

Decode_Frame ends the loop (it sets the state at ccStream+0x182 to 2 and calls CompleteDecodeSetup); what follows belongs to DecodeFrameSection.

The header chunk

ccStream::DecodeHeaderSection (0x00149d20) reads, after the magic: char[32] name into ccStream+8; a u16 into ccStream+0x184, which every decoder compares against constants and which is therefore the format version; padding to 4; a u32 into ccStream+0x16c; a u32 count into ccStream+0x44 followed by that many u32s into a ccMalloced array. A version below 0x90 crashes on purpose. In all 1,023 files the count is 1 and the one word is 0.

The texture chunk

ccStream::Decode_Texture (0x0014d9a0) reads:

u32       texture          object index of the TEX_ entry
u32       clut             object index of the CLT_ entry
u32       blt_group        only if version >= 0x92
u8        flag             low byte of ccTexChunkDesc.flag; the decoder
                           later ors in 1 and 4
u8        psm              GS pixel storage mode
u8        mipmap           only if version >= 0x90: extra levels after level 0
u8        aref             only if version >= 0x90: alpha-test reference
u8        tw               only if version >= 0x90: log2 width
u8        th               only if version >= 0x90: log2 height
          padding to 4 bytes
level[mipmap + 1]          at most 4 - ccTexChunkDesc.pos has 4 entries
  u16     buf_x            VRAM placement, into ccTexChunkDesc.pos[level];
  u16     buf_y            0/0 or x < 768 takes a default from warnPos
                           (0x0034ac90)
  u32     count            words that follow
  u32[count] data

The names are ccTexChunkDesc's, from the DWARF: u16 flag, u8 psm, tw, th, mipmap, aref, ccBuffPos pos[4], 0x18 bytes. The file stores the bytes in a different order from the struct.

So a texture chunk's real length is its header plus, per level, 2 words plus count. In xasc00 that is 0x2007 words against a declared 0x2039. The decoder also skips an already-loaded texture by reading sum(ccGetTexTransSize(psm, w, h) * 4) words, which must therefore equal the per-level 2 + count.

Survey

tools/ccs.py now walks by the size field except for textures, and its survey command runs the walk over a whole archive. Over DATA.BIN: 847 of 1,023 members walk to their exact last byte. Every one of the other 176 goes wrong immediately after a Model chunk (0x0800), so Model is the next decoder whose length rule must be read. In every clean file the setup section ends with one 0xcccc0005 Frame chunk and one 0xccccff01.

Chunks counted before any walk went wrong: ExtObj 228,999, Obj2 218,971, Model 16,877, Obj 14,958, Material 10,540, Clump 7,193, Clut 5,101, Anime 4,090, Texture 3,985, Hit 2,255, DummyPosRot 1,251, DummyPos 1,207, Light 421, Morpher 233, Eff 162, Bbox 64, Camera 16, FBRect 2, FBPage 1. No Particle, BltGrp, Layer, Shadow or Pcm chunk was reached.