Reference · Formats

The CCSF Model chunk

solidfrom#10#17#244

Kind 0x0800 in a CCSF file: one model, made of "mmats" - runs of vertices that share a material - in one of three layouts chosen by mtype. Read by ccStream::Decode_Model (0x0014bce0). The chunk's size field is wrong for mtype & 4 models (15 words per mmat too many); use the layout.

Layout

All little-endian.

header
  u32       model            object index, the MDL_ entry
  f32       vertexScale      ccModelChunk+0x40
  u16       mtype            ccModelChunk+0x50; & 0xff if version < 0x100
  u16       mmatNum          0: no model, the decoder returns here
  u16       flag             blendType = flag & 3; the rest dropped
  s16       zoffs            ccModelChunk+0x4e; 0 in every file
  u8        unknown_10       read and dropped; 0x70 in every file
            padding to 4

mmat[mmatNum], each: header words, then a body

What the decoder keeps of the header:

• flag. Only flag & 3, as ccModelChunk.blendType (+0x4c). Over the 11,010 models in DATA.BIN with mmats, flag is 0 (9,990), 1 (532), 9 (266), 8 (136), 10 (82), 3 (3) or 2 (1). Bit 3, on 484 of them, is dropped at load like any bit above 1.
• zoffs. ccModel::Init (0x0013a440) turns it into the float ccModel.zoffs (+0x1c), and ccModel::Copy and Duplicate copy it. Nothing else loads either field, so it has no effect; it is 0 in all 11,010.
• unknown_10. Read and not stored; 0x70 in all 11,010.

Header words per mmat

name is an object index (MDL_x_N) whose index entry the decoder clears; material is the MAT_ object.

Body: rigid (Decode_Mmat01, 0x0014b890)

s16[3]    position[vertexNum]
          padding to 4
u32       normal[vertexNum]      unless mtype & 0x40
u32       colour[vertexNum]      unless mtype & 0x200; read but dropped if mtype & 1
u32       st[vertexNum]          unless mtype & 0x400

mtype & 2 here is a deliberate crash (0x0014b8d8).

Body: bone and skin (Decode_Mmat02, 0x0014a600)

With offsetNum = 0, "bone" - every vertex follows one clump node:

u32       node                   clump node slot
s16[3]    position[vertexNum]    in that node's space
          padding to 4
u32       normal[vertexNum]
u32       st[vertexNum]

With offsetNum > 0, "skin" - a vertex is one or more weighted entries:

entry[offsetNum]                 8 bytes each
  s16[3]  position               in the entry's node's space
  u16     w                      bits 0-8 weight (256 = 1.0)
                                 bit 9    last entry of this vertex
                                 bits 10-15 clump node slot
u32       normal[offsetNum]
u32       st[number of vertices] = number of entries with bit 9 set

In all data the count of bit-9 entries equals vertexNum.

Body: shadow (DecodeShadowModel, 0x00140920)

s32       vNum                   0: nothing follows
s32       iNum
s16[3]    position[vNum]
          padding to 4
s32[3]    triangle[iNum / 3]

A closed mesh (every edge shared by two triangles, bar four edges in all of DATA.BIN) that the engine stretches into a shadow volume: the shadow volumes. Every shadow model has one mmat.

Values

The ST scale. The model packet unpacks ST as V2-16 in STMOD offset mode, with STROW [(u >> 4) & 0xff, (v >> 4) & 0xff, 1/16, 1.0] (the draw path). mc_DrawTriFast and the other programs load the slot as xyz, convert only x and y (itof12.xy), and multiply all three by the vertex's Q. The z they send as the GS's Q is the slot's z taken as float bits. The EE manual calls a V2 unpack's z and w indeterminate; PCSX2's unpacker, tested on the console, puts x and y there (Vif_UnpackSSE.cpp, "v1v0v1v0"). With the row added in offset mode, z is then S + bits(1/16) as an integer, the float 1/16 * (1 + S * 2^-23), and a vertex's texture coordinate is

u = (S + row.x) / (256 * (1 + S * 2^-23))

one repeat per 256 to within S * 2^-23: 3 * 10^-5 at S = 256, under 0.8% at the largest S. The port takes z as exactly 1/16.

Strips always open with two vertices flagged 1, and run on across the 48-vertex batches the engine splits a mmat into. VU1 turns the flag into the GS ADC bit. Nothing culls back faces, and winding is not consistent across models. See the draw path.

After the mmats ccBbox_SetBox(chunk+0x10, vmm, vertexScale) (0x001388c0) stores the model's box: per axis the integer min and max over every vertex (starting at 0x10000 and -0x10000), first doubled about the centre for mtype & 6. With unit = vertexScale / 4096: min at +0x10, max at +0x20, each int * unit, and the centre at +0x30, ((min + max) >> 1) * unit, w 1. ccModel::Init (0x0013a550) points ccModel +4 at it for a model without mtype & 6 and leaves it null for a bone or skin model; ccModel::Draw culls by it (_ccCheckBoundingBoxEx) and keys the sorted group on its centre (the desktop's draw order). The centre matches the game's decoder in eemu for all 41 models of xdttopen0 and 138 of xddesk01.

Posing

Bone and skin positions are in clump-node space. Local matrix = T * Rx * Ry * Rz * S, rotation in degrees (SetMatrix_PosRotZYXScale, 0x00138120); world = parent world * local (0x00138380); a vertex lands at node world * position. Slot i is the i-th node of the model's clump - inferred, consistent with every pose checked. Axes are Z-up, right-handed.

Types seen

DATA.BIN holds 17,254 models: 16,407 rigid, 3,129 shadow, 2,994 bone and 185 skin mmats, 5,417,645 vertices, 2,819,008 triangles.

Notes

The size field of an mtype & 4 model is 15 words per mmat longer than the chunk. Nothing reads a chunk's size: DecodeSetupSection dispatches on the kind and each decoder reads what its fields say, so the game is not affected. Why the exporter wrote it so is not on the discs.

Unknown

None.