Expand description
tc-gpu v3 (stream magic TCG3): one global integer CDF 9/7
transform over the whole field, with the coefficients entropy-coded in
independently decodable code-blocks (JPEG2000-style).
§Dependency rule
- Subbands are decoded coarse → fine (LL, then levels
nlev−1 … 0). A code-block may read any coefficient of a coarser level (its parent band: the parent and the parent’s 6 face neighbours) — those are decoded first — and, inside its own region, every earlier line in raster order plus its own earlier samples on the line. Nothing outside the block at its own level is visible, so all blocks of one level decode in parallel (one rANS lane each: the parallelism is the block count). - Contexts use the residuals
q − predof the visible cells (optional,resctx), predictions use the values.
§Random access
A sub-box needs, at every level, the coefficients in the support of the integer synthesis lifting (±3 samples per half-band per level, a conservative bound for the 4-step 9/7 lifting), plus — for decoding — the parent cells their blocks condition on. Only those code-blocks (and the outlier tiles touching the box) are read; the inverse lifting runs only on the lines inside the needed region at each level.
Structs§
- Decoded3
- Result of a (sub-box) decode.
- Gpu3
Config - v3 options.
- Parsed3
- A parsed v3 stream.
- TileRef
- An outlier tile.
Enums§
- Model3
- Probability model of a v3 stream.
Constants§
- MAGIC3
- Stream magic.
Functions§
- ad_
default - Default adaptive layer (
TC_AD=wa,lim,maskoverrides, for sweeps). - blocks_
of - Code-blocks of a shape in coding order.
- decode3_
box - Decode the box
[lo, hi)(whole field:[0, d3)). - decode3_
box_ parsed decode3_boxon a parsed stream.- decode3_
f32 - Decode a whole v3 stream to
f32. - encode3_
f32 - Encode as a v3 stream.
- key3
- Predictor key of a code-block band (pass 0; passes add
p). - parse3
- Parse a v3 stream, plain (
TCG3) or carrying its decoder (TCS1, seecrate::shipdec). Offsets in the result indexbyteseither way. - samples3
- Training samples for a field at a PSNR/RMSE target (the coefficients the v3 coder would see at the rate-controlled step).