1555331 - ICCv4 profiles look oversaturated with some display profiles

Status: RESOLVED FIXED

(Core :: Graphics: Color Management, defect, P2)

Description

[Andrew Osmond [:aosmond] (he/him)]

I performed a mozregression checking an unrelated ICCv4 profile issue, and inadvertently discovered that the profiles display differently between Wayland and X11.

STR:

1. Login into Ubuntu 18.04 with Wayland.
2. Start Firefox normally, browse to http://www.color.org/version4html.xalter and note that the ICCv4 profile parts of the image appears oversaturated.
3. Start Firefox with MOZ_ENABLE_WAYLAND=1, browse to http://www.color.org/version4html.xalter and note that it appears correct.

I had observed this before, but I assumed it was due to qcms's transform being wrong rather than it being somehow dependent on the windowing system.

Comment 1

[Andrew Osmond [:aosmond] (he/him)]

Bug 1555329 is necessary to reproduce on recent builds due to the regression.

Comment 2

[Andrew Osmond [:aosmond] (he/him)]

Okay, it looks like we force sRGB with Wayland, since we don't support getting the color profile. We may still be misprocessing the X11 profile, or the system gave me the wrong one in the first place. Forcing sRGB without Wayland corrects the output.

Comment 3

[Andrew Osmond [:aosmond] (he/him)]

Comparing in GIMP and Chrome, it appears they decode correctly with the same display profile as me. This suggests it is QCMS.

Comment 4

[Andrew Osmond [:aosmond] (he/him)]

Attachment: Sharp ICC profile

I see the behaviour on several color profiles, as attached.

Comment 5

[Andrew Osmond [:aosmond] (he/him)]

Attachment: Thinkpad ICC profile

Comment 6

[Andrew Osmond [:aosmond] (he/him)]

Attachment: Screenshot from 2019-05-29 19-12-22.png

Comment 7

[Andrew Osmond [:aosmond] (he/him)]

Jeff reproduced this on his Mac with my laptop profile, so I think this can happen on any platform.

Comment 8

[Andrew Osmond [:aosmond] (he/him)]

Attachment: ASUS profile

Comment 10

[gon0s]

Attachment: JPG file with ICC v2.1 sRGB profile embedded | Left : Firefox 69 / Right : Chrome

ICC v2 profile with Firefox color management enabled is not fine too : fine gradiant get heavy solarized rendering, crushing fine gradiant into ugly 8bit ugly files, and a serious bump in the gamma

Comment 12

[Jeff Muizelaar [:jrmuizel]]

It looks like the problem here is around build_output_lut for parametric curves. We don't seem to be inverting the lut when we should be.

Comment 13

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1555331. Add a Param struct.

This makes our parsing of parametric transfer functions
more clear because you can see exactly what output variables
the input variables map to.

Comment 14

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1555331. Implement analytic inversion for parametric transfer functions.

This also adjusts the initialization of Param to make inversion better behaved.

Comment 15

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1555331. Use parametric inversion when building an output_lut.

Previously we didn't invert at all which was causing over saturation.

Comment 16

[Jeff Muizelaar [:jrmuizel]]

Attachment: Bug 1555331. Add a test that parametric functions are properly inverted.

Comment 17

[Pulsebot]

Pushed by [e-mail]:
https://hg.mozilla.org/integration/autoland/rev/e1168d49bc86
Add a Param struct. r=aosmond
https://hg.mozilla.org/integration/autoland/rev/a53a1e590172
Implement analytic inversion for parametric transfer functions. r=aosmond
https://hg.mozilla.org/integration/autoland/rev/21efbd4c45d1
Use parametric inversion when building an output_lut. r=aosmond
https://hg.mozilla.org/integration/autoland/rev/bffe3582b2c7
Add a test that parametric functions are properly inverted. r=aosmond

Comment 18

[Cosmin Sabou [:CosminS]]

https://hg.mozilla.org/mozilla-central/rev/e1168d49bc86
https://hg.mozilla.org/mozilla-central/rev/a53a1e590172
https://hg.mozilla.org/mozilla-central/rev/21efbd4c45d1
https://hg.mozilla.org/mozilla-central/rev/bffe3582b2c7