lensserious_vendor.h File Reference

Turn a camera maker's embedded lens-correction tables into ls_knots_t. More...

#include <stddef.h>
#include <stdint.h>
#include "lensserious.h"

Include dependency graph for lensserious_vendor.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Data Structures
struct	ls_vendor_sony_t
	Sony maker-note tables, verbatim: fixed-point int16 over nc evenly spaced radii, which the file leaves implicit. The fixed-point scales (1/16384 distortion, 1/2^21 CA, 1/8192 vignetting) are applied by ls_vendor_resolve(), never by the extractor. More...
struct	ls_vendor_fuji_t
	Fujifilm maker-note tables, verbatim: explicit knot radii shared by the three corrections, percentage-domain coefficients. More...
struct	ls_vendor_dng_t
	A DNG's OpcodeList3 corrections, parsed. Fill it with ls_vendor_parse_dng_opcodelist3() — or by hand, if the consumer's DNG reader already decoded the opcodes itself. More...
struct	ls_vendor_olympus_t
	Olympus maker-note polynomials, verbatim. No vignetting: Olympus does not publish one, and no flag here can invent it. More...
struct	ls_vendor_data_t
	One image's embedded correction data: the type tag and the matching member. A plain value — embed it in an image struct, memcpy it, zero it for "none". More...
struct	ls_vendor_finetune_t
	Per-axis strength blend, 1.0 = exactly as the maker measured. The vendors' own GUIs expose these as "fine-tune" sliders; a consumer that does not can simply pass NULL to ls_vendor_resolve(), which means all-1.0. More...

Macros
#define	LS_VENDOR_EBADDATA   (-1)
	Resolve/parse result: the data is structurally malformed. Distinct from 0 ("carries nothing") so the consumer can SAY SO — a user whose file claims a profile that cannot be read is entitled to hear why the fallback engaged.

Typedefs
typedef enum ls_vendor_type_t	ls_vendor_type_t
	Which maker's format a ls_vendor_data_t holds. NONE == 0, so a zero-initialised struct safely means "no data". Values are not a wire format — do not persist them.
typedef struct ls_vendor_sony_t	ls_vendor_sony_t
	Sony maker-note tables, verbatim: fixed-point int16 over nc evenly spaced radii, which the file leaves implicit. The fixed-point scales (1/16384 distortion, 1/2^21 CA, 1/8192 vignetting) are applied by ls_vendor_resolve(), never by the extractor.
typedef struct ls_vendor_fuji_t	ls_vendor_fuji_t
	Fujifilm maker-note tables, verbatim: explicit knot radii shared by the three corrections, percentage-domain coefficients.
typedef struct ls_vendor_dng_t	ls_vendor_dng_t
	A DNG's OpcodeList3 corrections, parsed. Fill it with ls_vendor_parse_dng_opcodelist3() — or by hand, if the consumer's DNG reader already decoded the opcodes itself.
typedef struct ls_vendor_olympus_t	ls_vendor_olympus_t
	Olympus maker-note polynomials, verbatim. No vignetting: Olympus does not publish one, and no flag here can invent it.
typedef struct ls_vendor_data_t	ls_vendor_data_t
	One image's embedded correction data: the type tag and the matching member. A plain value — embed it in an image struct, memcpy it, zero it for "none".
typedef struct ls_vendor_finetune_t	ls_vendor_finetune_t
	Per-axis strength blend, 1.0 = exactly as the maker measured. The vendors' own GUIs expose these as "fine-tune" sliders; a consumer that does not can simply pass NULL to ls_vendor_resolve(), which means all-1.0.

Enumerations
enum	ls_vendor_type_t {   LS_VENDOR_NONE = 0 ,   LS_VENDOR_SONY ,   LS_VENDOR_FUJI ,   LS_VENDOR_DNG ,   LS_VENDOR_OLYMPUS }
	Which maker's format a ls_vendor_data_t holds. NONE == 0, so a zero-initialised struct safely means "no data". Values are not a wire format — do not persist them. More...

Functions
int	ls_vendor_axes (const ls_vendor_data_t *d)
	Which correction axes this data carries.
int	ls_vendor_resolve (const ls_vendor_data_t *d, const ls_vendor_finetune_t *ft, int w, int h, ls_knots_t *knots, float *autoscale)
	The conversion: vendor tables in, generic knots out.
int	ls_vendor_parse_dng_opcodelist3 (const uint8_t *blob, size_t len, ls_vendor_dng_t *out)
	Parse a DNG OpcodeList3 payload: WarpRectilinear (1) and VignetteRadial (3).

Detailed Description

Turn a camera maker's embedded lens-correction tables into ls_knots_t.

Some cameras write their own lens profile into every raw file — measured on the actual body and the actual lens, which makes it the best-informed correction available. All the formats this module reads express it the same way underneath: a short list of radii and, at each, the factor a coordinate is scaled by — per channel, so distortion and lateral chromatic aberration arrive measured together. This module normalises whichever format the file holds into one ls_knots_t, so past ls_vendor_resolve() nothing knows or cares which maker wrote the profile.

Where this API sits — after metadata decoding

This library links no metadata reader, deliberately. The boundary is drawn where the numbers appear: the consumer's EXIF library (exiv2, libraw, anything) reads the maker-note tags and fills one of the per-vendor structs below with the decoded values; this module owns everything after that — the vendor-specific coefficient semantics, the per-channel untangling, the resampling onto a common radius axis. Re-parsing maker-note bytes here would mean re-doing tag layout and endianness that the consumer's metadata library already resolved, so the input structs carry values, not bytes.

The one exception proves the rule: a DNG's OpcodeList3 payload is an opaque, self-contained big-endian blob that no metadata library decodes for you. Its format is correction-domain knowledge, so ls_vendor_parse_dng_opcodelist3() lives here — the consumer hands the raw tag payload and gets the filled struct back.

What the consumer extracts, per vendor

Worked examples from Ansel's src/metadata/exif.cc (exiv2), for future implementors. Tag spellings are exiv2's; other libraries name the same tags differently.

Sony — three SubIFD arrays of int16, count first: Exif.SubImage1.DistortionCorrParams, .ChromaticAberrationCorrParams (2×nc values: red then blue), .VignettingCorrParams (fallback spellings Exif.Image.0x7037 / 0x7035 / 0x7032 on older bodies). Validate nc in [2, 16], that the CA array holds 2×nc, then copy values 1..nc of each into ls_vendor_sony_t verbatim. The radii are implicit (evenly spaced); the fixed-point scale factors are this module's business, not the extractor's.

Fujifilm — three maker-note float arrays sharing one knot list: Exif.Fujifilm.GeometricDistortionParams (19 or 23 values: 9 or 11 knots then as many coefficients), .ChromaticAberrationParams (29 or 31: knots, red, blue), .VignettingParams. The extractor checks the three knot lists agree and fills ls_vendor_fuji_t with knots and coefficients separately.

cropf is the extractor's responsibility, by contract. It is CAMERA knowledge, not format knowledge: the tables are calibrated against the full sensor, so when the body shot in a cropped mode the knot radii must be scaled by the crop ratio — and whether the file was shot cropped is written nowhere near the correction tables. Ansel decides it from Exif.Fujifilm.CropMode: values 2 ("sports finder") and 4 (electronic teleconverter) mean the 1.25× crop, so

fuji.cropf = (crop_mode == 2 || crop_mode == 4) ? 1.25f : 1.0f;

An extractor that cannot determine the mode must pass 1.0f — never 0, which would collapse every radius to the image centre. This module applies cropf blindly and has no opinion about where it came from.

DNG — one opaque byte blob: Exif.SubImage1.OpcodeList3 (or Exif.Image.OpcodeList3 on files without an embedded preview). Hand it, undecoded, to ls_vendor_parse_dng_opcodelist3(). A WarpRectilinear opcode with one plane is distortion alone; with three planes it is distortion and TCA measured together. VignetteRadial is the falloff. Note OpcodeList*2* is a different list (sensor-domain corrections, gain maps) and is none of this module's business.

Olympus — two maker-note arrays of already-decoded rationals/floats: Exif.OlympusIp.0x150c (distortion: k2, k4, k6, scale) and Exif.OlympusIp.0x150e (CA: three coefficients each for red and blue). Olympus publishes no vignetting; that axis is simply absent from what ls_vendor_axes() reports, and a consumer offering per-axis sources falls back to its other ones (Ansel uses its lens database there).

Output conventions

The emitted ls_knots_t follows the library's knot conventions (see its own documentation): radii normalized to half the image diagonal, per-channel correcting scale factors where green IS the distortion and red/blue carry the CA relative to it, vignetting as the measured falloff (below 1 where the lens darkens, corrected by dividing). The autoscale returned alongside is the zoom that just clears the borders the geometric correction leaves — apply it as (or composed into) the scale argument of ls_modifier_init_knots(); it is not baked into the table beyond the normalisation the resolver already performed.

Known limits

The DNG opcodes carry an optical-centre offset (warp_cx/cy, vig_cx/cy); the conversion currently assumes a centred system and ignores them — wrong by the offset amount for the rare decentred profile, faithful to what every shipping consumer of this code has done so far. The struct carries the values so fixing it needs no ABI change.

Definition in file lensserious_vendor.h.

Macro Definition Documentation

LS_VENDOR_EBADDATA

#define LS_VENDOR_EBADDATA   (-1)

Resolve/parse result: the data is structurally malformed. Distinct from 0 ("carries nothing") so the consumer can SAY SO — a user whose file claims a profile that cannot be read is entitled to hear why the fallback engaged.

Definition at line 211 of file lensserious_vendor.h.

Referenced by ls_vendor_parse_dng_opcodelist3(), and ls_vendor_resolve().

Typedef Documentation

ls_vendor_data_t

typedef struct ls_vendor_data_t ls_vendor_data_t

One image's embedded correction data: the type tag and the matching member. A plain value — embed it in an image struct, memcpy it, zero it for "none".

ls_vendor_dng_t

typedef struct ls_vendor_dng_t ls_vendor_dng_t

A DNG's OpcodeList3 corrections, parsed. Fill it with ls_vendor_parse_dng_opcodelist3() — or by hand, if the consumer's DNG reader already decoded the opcodes itself.

ls_vendor_finetune_t

typedef struct ls_vendor_finetune_t ls_vendor_finetune_t

Per-axis strength blend, 1.0 = exactly as the maker measured. The vendors' own GUIs expose these as "fine-tune" sliders; a consumer that does not can simply pass NULL to ls_vendor_resolve(), which means all-1.0.

ls_vendor_fuji_t

typedef struct ls_vendor_fuji_t ls_vendor_fuji_t

Fujifilm maker-note tables, verbatim: explicit knot radii shared by the three corrections, percentage-domain coefficients.

ls_vendor_olympus_t

typedef struct ls_vendor_olympus_t ls_vendor_olympus_t

Olympus maker-note polynomials, verbatim. No vignetting: Olympus does not publish one, and no flag here can invent it.

ls_vendor_sony_t

typedef struct ls_vendor_sony_t ls_vendor_sony_t

Sony maker-note tables, verbatim: fixed-point int16 over nc evenly spaced radii, which the file leaves implicit. The fixed-point scales (1/16384 distortion, 1/2^21 CA, 1/8192 vignetting) are applied by ls_vendor_resolve(), never by the extractor.

ls_vendor_type_t

typedef enum ls_vendor_type_t ls_vendor_type_t

Which maker's format a ls_vendor_data_t holds. NONE == 0, so a zero-initialised struct safely means "no data". Values are not a wire format — do not persist them.

Enumeration Type Documentation

ls_vendor_type_t

enum ls_vendor_type_t

Which maker's format a ls_vendor_data_t holds. NONE == 0, so a zero-initialised struct safely means "no data". Values are not a wire format — do not persist them.

Enumerator
LS_VENDOR_NONE
LS_VENDOR_SONY
LS_VENDOR_FUJI
LS_VENDOR_DNG
LS_VENDOR_OLYMPUS

Definition at line 120 of file lensserious_vendor.h.

Function Documentation

ls_vendor_axes()

int ls_vendor_axes

(

const ls_vendor_data_t *

d

)

Which correction axes this data carries.

Returns

A mask of LS_ENABLE_DISTORTION, LS_ENABLE_TCA and LS_ENABLE_VIGNETTING; 0 when the data carries nothing usable (including d NULL or type NONE).

Cheap — table scans, no knots work — so a GUI may ask per repaint. The same mask vocabulary as ls_modifier_init() returns, one axis language across the library. An all-zero table is reported as ABSENT, not as a present-but-identity correction: several bodies write zero-filled tables for lenses they choose not to correct.

Definition at line 353 of file lensserious_vendor.c.

References _dng_axes(), _fuji_axes(), _olympus_axes(), _sony_axes(), LS_VENDOR_DNG, LS_VENDOR_FUJI, LS_VENDOR_NONE, LS_VENDOR_OLYMPUS, LS_VENDOR_SONY, and ls_vendor_data_t::type.

Referenced by ls_vendor_parse_dng_opcodelist3(), and ls_vendor_resolve().

ls_vendor_parse_dng_opcodelist3()

int ls_vendor_parse_dng_opcodelist3	(	const uint8_t *	blob,
		size_t	len,
		ls_vendor_dng_t *	out )

Parse a DNG OpcodeList3 payload: WarpRectilinear (1) and VignetteRadial (3).

Parameters

blob	the tag's raw bytes, exactly as the file carries them (big-endian, per the DNG 1.3 specification). Not written to.
len	the payload size in bytes.
out	filled with whatever was recognised; zeroed first, so a partial parse leaves the untouched corrections cleanly absent.

Returns

The axes mask the parsed opcodes provide; 0 when the list holds none of the recognised opcodes (other opcode ids are skipped, as the specification intends); LS_VENDOR_EBADDATA when the list structure itself is broken before anything was recognised. A list that breaks AFTER yielding a usable opcode returns what it yielded: partial data is data.

Every read is bounds-checked against len; a truncated or hostile blob cannot read past the buffer.

Definition at line 505 of file lensserious_vendor.c.

References _get_be_long(), _parse_vignette_radial(), _parse_warp_rectilinear(), DNG_OPCODE_ID_VIGNETTE_RADIAL, DNG_OPCODE_ID_WARP_RECTILINEAR, ls_vendor_axes(), LS_VENDOR_DNG, and LS_VENDOR_EBADDATA.

ls_vendor_resolve()

int ls_vendor_resolve	(	const ls_vendor_data_t *	d,
		const ls_vendor_finetune_t *	ft,
		int	w,
		int	h,
		ls_knots_t *	knots,
		float *	autoscale )

The conversion: vendor tables in, generic knots out.

Parameters

d	the extracted data. See what the consumer extracts.
ft	per-axis strength, or NULL for "as the maker measured" (all 1.0).
w	image width in pixels — the PROCESSED dimensions the radii are measured against (after the sensor's fixed border trim, before any user crop). Ansel passes dt_image_t.p_width/p_height.
h	image height in pixels.
knots	out: the normalized table, ready for ls_modifier_init_knots(). Fully written on success (n and vn included); zeroed on failure.
autoscale	out, may be NULL: the zoom that just clears the borders the correction leaves, measured over the region between the inscribed circle and the corner. Pass it (times any user scale) as ls_modifier_init_knots()'s scale argument. 1.0 on failure.

Returns

The resolved axes mask (same bits as ls_vendor_axes()); 0 when the data carries nothing — a zero-filled table with a valid type included, see ls_vendor_axes(); LS_VENDOR_EBADDATA when a claimed correction fails to convert.

Pure and thread-safe: no allocation, no globals, callable from a pipeline thread per commit. Not per pixel — resolve once, keep the knots.

Definition at line 368 of file lensserious_vendor.c.

References _dng_populate(), _ft_identity, _fuji_populate(), _linear_spline(), _olympus_populate(), _sony_populate(), ls_knots_t::cor_rgb, LS_ENABLE_VIGNETTING, ls_vendor_axes(), LS_VENDOR_DNG, LS_VENDOR_EBADDATA, LS_VENDOR_FUJI, LS_VENDOR_OLYMPUS, LS_VENDOR_SONY, ls_knots_t::n, ls_knots_t::radius, ls_vendor_data_t::type, and ls_knots_t::vn.