Module format 

GPU tile format: header, tile, index, and footer structures

File Layout

Offset 0:           GpuFileHeader (64 bytes)
Offset 64:          Tile 0 — TileHeader (32 bytes) + payload (padded to 128-byte boundary)
...                 Tile N-1
Offset X:           TileIndexHeader (8 bytes)
Offset X+8:         TileIndexEntry[0..N] (24 bytes each)
Offset X+8+24N:     GpuFileFooter (24 bytes)  ← last 24 bytes of file

Structs

EngineVersion

Packed 8-byte semantic version of the producing engine.

GpuFileFlags

Bitfield flags stored in the GPU file header.

GpuFileFooter

Fixed 24-byte record at the very end of every GPU archive.

GpuFileHeader

Fixed 64-byte header at the start of every GPU-compressed archive.

TileFlags

Per-tile bitfield flags stored in the tile header.

TileHeader

Fixed 32-byte header preceding each compressed tile’s payload.

TileIndexEntry

One entry per tile in the trailing index.

TileIndexHeader

8-byte header immediately before the tile index entries.

Constants

CGPU_MAGIC

CGPU magic bytes: ASCII “CGPU”

DEFAULT_SUB_STREAM_COUNT

Default sub-stream count: 32 (matches GPU warp width)

DEFAULT_TILE_SIZE

Default tile size: 64KB (65536 bytes)

ENGINE_VERSION_STR

Current engine semantic version (set at build time).

FORMAT_VERSION

Current format version. Files with a different version are rejected. Version 2: GDeflate encoding (replaces v1 LZ77).

TILE_ALIGNMENT

Alignment boundary for tile payloads (128 bytes for GPU memory coalescing)

Functions

padding_to_alignment

Compute padding needed to align a size to the tile alignment boundary.