#include <quantizer.h>

Public Attributes
float	abs_error_bound
	Absolute EB after mode conversion (0 for REL).

float	user_error_bound

float	value_base
	value_range (NOA); 0 for ABS/REL.

uint32_t	quant_radius
	Quantization radius.

uint32_t	num_elements
	Total element count.

uint32_t	outlier_count
	Actual number of outliers.

DataType	input_type
	Original input type (1B).

DataType	code_type
	Quantization code type (1B).

uint8_t	eb_mode
	ErrorBoundMode cast to uint8_t.

uint8_t	zigzag_codes
	1 if ABS/NOA codes are zigzag-encoded.

float	outlier_threshold
	ABS/NOA: \|x\| >= threshold → forced outlier (inf = disabled).

uint8_t	inplace_outliers
	1 if outliers are encoded in-place in the codes array.

uint8_t	linear_mode
	1 if linear/no-outlier mode (signed codes, no outlier ports).

uint8_t	dither
	1 if "_R"-style dithered reconstruction is enabled (LC QUANT_*_R).

uint8_t	_pad [5]
	Alignment padding (dither_seed needs 8-byte alignment) — must be zero.

uint64_t	dither_seed
	Deterministic per-element dither seed; meaningful only when dither.

float	dither_strength
	Dither offset amplitude as a fraction of abs_eb, in (0,1]; meaningful only when dither.

uint32_t	_pad2

double	abs_error_bound_f64

double	value_base_f64
	Full-precision value_base; 0 in pre-2026-08-07 headers.

Detailed Description

Serialized quantizer configuration stored in FZMBufferEntry.stage_config. Written by serializeHeader(); read back by deserializeHeader(). 72 bytes — fits within the 128-byte FZM_STAGE_CONFIG_SIZE limit.

Member Data Documentation

◆ abs_error_bound

float fz::QuantizerConfig::abs_error_bound

Absolute EB after mode conversion (0 for REL).

Deprecated:: Narrow copy of the absolute EB, kept so old readers still parse. New code must read abs_error_bound_f64 — see the f64 note on that field.

◆ user_error_bound

float fz::QuantizerConfig::user_error_bound

Original user-specified EB.

◆ value_base

float fz::QuantizerConfig::value_base

value_range (NOA); 0 for ABS/REL.

Deprecated:: Narrow copy; new code reads value_base_f64.

◆ _pad2

uint32_t fz::QuantizerConfig::_pad2

Alignment padding (the f64 fields need 8-byte alignment) — must be zero.

◆ abs_error_bound_f64

double fz::QuantizerConfig::abs_error_bound_f64

Full-precision absolute EB. The float copy above cannot carry an f64 bound: decode computes recon = q * 2*abs_eb, and for a field like S3D/N2 (q ~ 3.3e8) a 6e-08 relative error in abs_eb lands as a 4.4e-08 absolute error — 40x the bound. Headers written before 2026-08-07 leave this 0 and the reader falls back to the float field.