Fused stages

Stage	Description
LorenzoQuantStage	Fused float predictor + quantizer (1-D/2-D/3-D)
AdaptiveLorenzoStage	Per-tile adaptive multi-order Lorenzo + centering (FSZ prediction stage)
GInterpStage	Multi-level spline interpolation predictor + quantizer (3-D, cuSZ-Hi port)
BitplaneRZEStage	Fused bitplane transpose + zero-group RZE — FZ-GPU's lossless encoder
SZxStage	SZx ultrafast EB compressor — per-block constant/non-constant classification + fixed-length residuals (whole compressor)
DifferenceStage (negabinary-fused)	First-order difference with inline negabinary encoding (DifferenceStage<T, TOut> where TOut != T)

‍Note: DifferenceStage with a TOut != T template argument fuses a negabinary encoding step into the differencing kernel (equivalent to DifferenceStage<T> followed by NegabinaryStage, but in a single pass). This stage lives in modules/predictors/diff/, not in modules/fused/, but is listed here so all fused-operation options appear in one place.

Experimental / reference compressors (not supported modules)

Some whole-compressor implementations are kept only as GPU points of comparison and are not part of the composable module set: they are absent from <fzgpumodules.h>, from the stage catalog, and from the automatic-fusion planner. They live under modules/experimental/reference_compressors/.

Reference compressor	Supported modular equivalent
SZpStage<T> (modules/experimental/reference_compressors/szp/) — SZp / fZ-light: quantize + 1-D Lorenzo delta + fixed-length bitpack. See SZpStage (experimental reference compressor).	examples/presets/szp_composed.toml — Quantizer(linear) → Lorenzo(block_size=128) → AdaptiveBitpack(block_size=128)

Its StageType::SZP FZM factory stays linked so archives written before the quarantine still decode, and type = "SZp" still loads from legacy TOML configs.