Regression from 2.13 when interpolating CMYK colours

[amyspark]

Hi @mm2,

We've gotten this bug reported against Krita 5.1.3, where we upgraded Little CMS to 2.14.

Between 2.13.1 and 2.14, something broke in the interpolation of CMYK colors, at least in 8-bit depth:

2.13.1	2.14

I'm not sure where in the commit log it happened, but disabling the fast float plugin has no effect. The offending profile is: https://bugs.kde.org/attachment.cgi?id=154115

[amyspark]

It only affects CMYK with 8-bit integer depth, 16-bit and float appear to be unaffected.

[mm2]

Hello,
I cannot reproduce this with tificc. Could you please give me more details?

Specifically:

• the profile is used as input or output?
• which intent?
• any special flags, black point compensation?
• to which / from which profile are you doing the transform?

I used from/to built-in sRGB:

tificc -t0 -i AP_Munken_Print_White_260v4.1.icm ramps_cmyk.tif out-14-rgb.tif
tificc -t0 -o AP_Munken_Print_White_260v4.1.icm ramps.tif out-14-cmyk.tif

Both gives same result as 2.13.1

Probably I am missing something important.

[mm2]

Checked other intents too, and BPC as well. All works fine to me.

[amyspark]

the profile is used as input or output?

Input.

which intent?

Perceptual.

any special flags, black point compensation?

Only black point compensation and the fast float plugin enabled.

to which / from which profile are you doing the transform?

sRGB-elle-V2-srgbtrc (output, 8-bit integer depth).

I'll try to bisect now and see if I can trap the regression.

[mm2]

The plain CMM without plugin works fine in this configuration. I have to build a tificc with the plugin, maybe the issue is in the plugin. Tomorrow will try to do that.

[amyspark]

Git bisected between lcms2.13.1 and current HEAD (with a couple edits to build the fast float plugin under Meson-- btw, thanks for having it!)

Seems that one of the optimizations went awry:

❯ git bisect bad
d24a40c9cc685b7145e8082606d5d9384d08e472 is the first bad commit
commit d24a40c9cc685b7145e8082606d5d9384d08e472
Author: Marti Maria <marti.maria@littlecms.com>
Date:   Sun Feb 20 16:07:41 2022 +0100

    Fix LUT based optimization on 8 bits

    Most times, optimization was discarded where it could be used. Monotonicity is not needed here. Added a small demo.

 plugins/fast_float/src/fast_8_tethra.c |   6 +-
 plugins/fast_float/testbed/demo_cmyk.c | 151 +++++++++++++++++++++++++++++++++
 2 files changed, 152 insertions(+), 5 deletions(-)
 create mode 100644 plugins/fast_float/testbed/demo_cmyk.c

[mm2]

This makes sense.
Seems a bug in the plug-in to me.
The bug was hidden in 2.13.1 because the profile was rejected for the plug-in and handled by the core engine. Now it is accepted, so the bug shows up. Will try to fix it ASAP.

[mm2]

This is a nasty bug on the plugin. Only affects 8 bits and CMYK. I will temporally disable CMYK on it 8 bits since the speed improvement is very small.

[mm2]

See 328c226

[mm2]

Fixed in 2.15

[amyspark]

@mm2 judging by the commit log, it seems there was no better solution found yet, right?