Skip to content

algorithm.md

Latest commit

 

History

History
76 lines (53 loc) · 2.33 KB

algorithm.md

File metadata and controls

76 lines (53 loc) · 2.33 KB
Raw

RefreshTone Algorithm Specification

Scope

RefreshTone applies refresh-level-dependent signed offsets to one interleaved RGB image. It is a deterministic digital reference model. It does not measure a panel, generate calibration LUTs, or control display hardware.

The fixed model shape is:

  • input and output: unsigned 12-bit samples, 0..4095;
  • channels: interleaved R, G, B;
  • LUT: one signed int16 table per channel;
  • LUT shape: 8 refresh-level anchors by 256 gray nodes;
  • frame level: integer 0..127;
  • interpolation fractions: 4 bits.

Index decomposition

For an input sample x:

gray0    = x >> 4
gray1    = min(gray0 + 1, 255)
grayFrac = x & 0x0f

For a frame level f:

level0    = f >> 4
level1    = min(level0 + 1, 7)
levelFrac = f & 0x0f

The endpoint is clamped. Therefore sample 4095 uses node 255 for both gray endpoints, and frame level 127 uses level 7 for both level endpoints.

Interpolation and rounding

The primitive interpolation is:

interp(a, b, t) = round_away_from_zero((a * (16 - t) + b * t) / 16)

t is in 0..15. Rounding is to the nearest integer; an exact half is rounded away from zero. This explicitly preserves the intent of the original std::round implementation without using floating-point arithmetic.

For channel c:

atLevel0 = interp(lut[c][level0][gray0], lut[c][level0][gray1], grayFrac)
atLevel1 = interp(lut[c][level1][gray0], lut[c][level1][gray1], grayFrac)
offset   = interp(atLevel0, atLevel1, levelFrac) - zeroSetting
output   = clamp(x + offset, 0, 4095)

Rounding happens after each interpolation stage. Collapsing both stages into one expression can produce different results and is not conforming.

Bypass and errors

  • When enabled is false, validated input is copied unchanged to output.
  • Frame levels outside 0..127 are rejected.
  • Zero dimensions, overflowing dimensions, mismatched buffers, and samples above 4095 are rejected.
  • The operation is all-or-nothing for validation errors: all samples are validated before output is modified.
  • In-place processing is supported.

LUT text format

Each channel file contains exactly 2048 whitespace-separated signed decimal integers in level-major order:

level 0: nodes 0..255
level 1: nodes 0..255
...
level 7: nodes 0..255

Lines may contain comments beginning with # or //.