Rework gamut_LUT sampling

[jenshannoschwalm]

Fixes and flexibility for LUT_ELEM

LUT_ELEM is currently meant to be used with a table per 1°, this commit

1. allows larger tables to be used for higher precision for later linear interpolation
2. does not use the final averaging any more as it didn't help at all.
3. Fix color equalizer and colorbalancergb to work correctly with all values of LUT_ELEM

Increase LUT_ELEM to 512

Subtly improved precision for lookup_gamut function due to larger table.
Also, as we use LUT_ELEM to be as 2^n we can redefine lookup_gamut

• removing the restriction of hue input angle
• improved performance

[TurboGit]

This breaks at least test 0093-colorbalancergb-ucs:

Test 0093-colorbalancergb-ucs
      Image mire1.cr2
      CPU & GPU version differ by 230497 pixels
      CPU vs. GPU report :
      ----------------------------------
      Max dE                   : 1.64225
      Avg dE                   : 0.03006
      Std dE                   : 0.11560
      ----------------------------------
      Pixels below avg + 0 std : 91.77 %
      Pixels below avg + 1 std : 92.36 %
      Pixels below avg + 3 std : 96.98 %
      Pixels below avg + 6 std : 99.46 %
      Pixels below avg + 9 std : 99.94 %
      ----------------------------------
      Pixels above tolerance   : 0.00 %
 
      Expected CPU vs. current CPU report :
      ----------------------------------
      Max dE                   : 3.69803
      Avg dE                   : 0.05937
      Std dE                   : 0.24551
      ----------------------------------
      Pixels below avg + 0 std : 89.55 %
      Pixels below avg + 1 std : 93.61 %
      Pixels below avg + 3 std : 98.05 %
      Pixels below avg + 6 std : 99.25 %
      Pixels below avg + 9 std : 99.75 %
      ----------------------------------
      Pixels above tolerance   : 0.24 %
 
  FAILS: image visually changed
         see diff.png for visual difference
         (298060 pixels changed)

The diff output is:

Probably expected? Can you confirm @jenshannoschwalm? TIA.

[jenshannoschwalm]

Yes, (unfortunately) expected.

If you think, the difference is too big, i would rethink. Yet - the new way offers clearly improved gamut_luts for UCS mode. The option would be to use different algos ...

Another point, there is also some surpringly large difference in CPU vs GPU code in colorbalancergb. So we might have - i would say hopefully - another round of differences if the reason for this is found. (still on the track)

[TurboGit]

@jenshannoschwalm

Another point, there is also some surpringly large difference in CPU vs GPU code in colorbalancergb.

Yes, I've seen that too.

So we might have - i would say hopefully - another round of differences if the reason for this is found. (still on the track)

Let's wait a bit before merging this then, I'll prefer having a single integration test update if possible.

[jenshannoschwalm]

@TurboGit would you re-review?

I think the first 2 commits should be merged in any case as simplification, code-bug fix and deduplication.

The third commit improves "things" slightly, marginally better interpolated data plus subtle perf gain.

[TurboGit]

@jenshannoschwalm : Sure, I'll try to review again later today.

[TurboGit]

Looks good, I'll update 0093 test. Thanks!