Alpha distance

[jonsneyers]

Adds an option to the API to set the distance for extra channels separately, addressing #2111 .

This e.g. allows doing lossless alpha while the color is lossy, or in general to set the alpha quality independently of the color quality. We might want to experiment with this and set more sensible defaults. (Note: we can't really rely on the metrics though, since they don't really look at alpha, they just use it to blend the image to a solid background.)

Adds a parameter D to benchmark_xl and --alpha_distance to cjxl.

Flyby fix: correct the modular quantization adjustment to compensate for the actual bitdepth of extra channels (instead of assuming that all channels have the same bitdepth).

Encoding       kPixels    Bytes          BPP  E MP/s  D MP/s     Max norm  SSIMULACRA2        pnorm       BPP*pnorm   Bugs
--------------------------------------------------------------------------------------------------------------------------
jxl:d0.5          4312  1326476    2.4608889   1.247   9.371   1.46619773  93.29027781   0.32387099  0.797010517600      0
jxl:d0.5:D0       4312  1535745    2.8491264   1.296   9.113   1.52147913  93.17293088   0.29785035  0.848613285350      0
jxl:d1            4312   923797    1.7138356   1.526   9.811   2.18971467  88.25578150   0.54898038  0.940862091175      0
jxl:d1:D0         4312  1262339    2.3419014   1.246   8.648   2.03960395  88.48580332   0.50947993  1.193151755443      0
jxl:d1:D2         4312   824846    1.5302609   1.487  10.698   2.32580400  88.42132275   0.64015614  0.979605903443      0
jxl:d2            4312   630823    1.1703079   1.507   9.813   3.26052332  82.49513662   0.88900277  1.040406932615      0
jxl:d2:D4         4312   558300    1.0357626   1.391   9.768   3.80601859  82.32916682   1.01417959  1.050449298471      0
jxl:d3            4312   491437    0.9117178   1.425   7.479   4.17905045  78.08833075   1.16395135  1.061195208251      0
jxl:d3:D6         4312   434096    0.8053384   1.468   8.684   4.96663904  77.92568643   1.30559641  1.051446871010      0
Aggregate:        4312   808263    1.4994954   1.395   9.222   2.62748584  85.65005370   0.65969720  0.989212882981      0

For some images, doing alpha losslessly results in smaller files than doing it lossy, e.g.:

Encoding       kPixels    Bytes          BPP  E MP/s  D MP/s     Max norm  SSIMULACRA2        pnorm       BPP*pnorm   Bugs
--------------------------------------------------------------------------------------------------------------------------
jxl:d0.5           409    47889    0.9353320   0.999  10.693   0.76884204  97.97876012   0.29019442  0.271428139716      0
jxl:d0.5:D0        409    46700    0.9121094   1.014  12.442   0.76884204  98.26795927   0.29005663  0.264563367873      0
jxl:d1             409    31685    0.6188477   1.017  10.603   1.19141257  91.69366061   0.47456421  0.293682948235      0
jxl:d1:D0          409    31095    0.6073242   1.045  11.505   1.19293487  91.82356930   0.47412265  0.287946168915      0
jxl:d2             409    20969    0.4095508   1.021  10.966   2.02522826  86.19662382   0.75935998  0.310996471061      0
jxl:d2:D0          409    20971    0.4095898   1.039  11.696   2.02522278  86.03497192   0.75843025  0.310645329176      0
jxl:d3             409    16549    0.3232227   1.004   9.977   2.59280419  82.55876856   0.97508168  0.315168489608      0
jxl:d3:D0          409    16891    0.3299023   0.985   8.871   2.64150572  82.59064968   0.97371519  0.321230921912      0
Aggregate:         409    26859    0.5245902   1.015  10.794   1.48461947  89.45138936   0.56480591  0.296291627848      0

So maybe we should reconsider the default behavior, and do alpha losslessly when doing high quality lossy, and only make alpha lossy when the quality is low enough. To be investigated later...

[veluca93]

Can you split the all_color_aware commit in another PR?

[jonsneyers]

Can you split the all_color_aware commit in another PR?

Yes, that one is already merged (#2116).

Rebased and fixed the merge conflict.

[mo271]

Could you add an entry to the changelog?

[jonsneyers]

Could you add an entry to the changelog?

Done, thanks for the reminder!