oklabToXyzQ16()

void fl::oklabToXyzQ16	(	const i32(&)	lab[3],
		i32(&)	out_xyz[3] )

OKLab (s16.16) -> XYZ (s16.16).

The inverse of xyzToOklabQ16.

Not bit-exact in round trip: both directions round to nearest and the cube root truncates. oklabToXyzQ16(xyzToOklabQ16(v)) returns within 32 ULP over the working domain.

The other order degrades near black. OKLab cube-roots the cone responses, and that root's derivative, 1 / (3 * lms^(2/3)), diverges as a response approaches zero: around L = 0.05 it is roughly 1200, so any error in the response is hugely amplified coming back out. Restricted to in-gamut colours the worst measured round trip there is about 300 ULP. This is conditioning rather than representation – carrying lms or XYZ at Q32 was measured and does not fix it – and it costs the P7 mapper nothing. tests/fl/gfx/oklab_q16.cpp pins both ends.

Definition at line 134 of file oklab_q16.cpp.hpp.

                                                                        {
    const i32 clamped[3] = {clampOklabQ16(lab[0], kOklabQ16MaxMagnitude),
                            clampOklabQ16(lab[1], kOklabQ16MaxMagnitude),
                            clampOklabQ16(lab[2], kOklabQ16MaxMagnitude)};
    i32 root[3];
    dotOklabMatrixQ16(kLmsRootFromOklab, clamped, root);
    const i32 lms[3] = {
        signedCubeQ16(root[0]), signedCubeQ16(root[1]), signedCubeQ16(root[2])};
    dotOklabMatrixQ16(kXyzFromLms, lms, out_xyz);
}

References FL_NO_EXCEPT, and kOklabQ16MaxMagnitude.

Referenced by fl::anonymous_namespace{gamut_map.cpp.hpp}::chromaCandidateXyz(), mapAndAllocateRgbwQ16(), and mapAndAllocateRgbwwQ16().

Here is the caller graph for this function: