mirror of
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
synced 2026-08-24 10:11:25 -04:00
a71d2b051f
When linearizing, the input is an encoded signal bounded to [0,1] and PQ/sRGB EOTFs are steepest near 1, requiring more precision near the bright end. Take the 8-bit sRGB case as a reference: 256 possible inputs and 256 HW LUT points line up, so the LUT acts as plain indexing. Float representations don't land perfectly, but LERP-ing between two HW entries, when input is within a small epsilon of one of them, doesn't materially change the result. Replace the uniform 12-region distribution (16 points each, 192 total, range [2^-12, 1]) with a 9-region halving distribution for the PQ/sRGB pre-defined EOTF: 128 points in the top region [0.5, 1], 64 in the next, 32 in the next, and so on, down to 1 point in each of the two darkest regions. Total samples grow from 192 to 256, with uniform 1/256 spacing across [0, 1]. The dark tail below 2^-9 is no longer sampled separately, which is acceptable for PQ/sRGB. Suggested-by: Krunoslav Kovac <Krunoslav.Kovac@amd.com> Tested-by: Matthew Schwartz <matthew.schwartz@linux.dev> Reviewed-by: Harry Wentland <harry.wentland@amd.com> Signed-off-by: Melissa Wen <mwen@igalia.com> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>