Color mix tests added in https://github.com/web-platform-tests/wpt/pull/42940 may ignore alpha premultiplication

[mysteryDate]

Looking at the tests added here: https://github.com/web-platform-tests/wpt/pull/42940/files#diff-17d0bf90434793454df8bbbd34e081eef94d3697534f8ec2c433359700a6b985R368

Using this line as an example: color-mix(in lab, lab(10 20 30 / 25%) 0%, lab(none none none / 0.5))
with the expected value: lab(10 20 30 / 0.5)

And following https://csswg.sesse.net/css-color-4/#interpolation-alpha and https://csswg.sesse.net/css-color-5/#color-mix-with-alpha:

First we must premultiply the values of both colors, so the premultiplied colors are:
lab(2.5 5 7.5 / 0.25) and lab(none none none / 0.5)

The resulting, interpolated value, taking 0% of the left and 100% of the right, and respecting missing components:
lab(2.5 5 7.5 / 0.5)

Which, un-premultiplied with the interpolated alpha value is, dividing all components by 0.5:
lab(5 10 15 / 0.5)

Which is the current result value in chromium. I believe these tests here omit the fact that the alpha used for premultiplication and un-premultiplication are different.

[foolip]

@mysteryDate would it be straightforward to update the test expectations here?

[mysteryDate]

Yeah, it would be.

Looking into the original change here, I see the logic. Any number interpolated with "none" should result in no change, regardless of the alphas involved. I really don't see anywhere in the spec where this is mentioned though. I would not be surprised at all if the spec does change.

So, should I change the test expectations to match what the spec says, or should I change chromium's behavior to match what the other UAs are doing and what I suspect the spec will eventually say explicitly?

[foolip]

If you think the behavior makes sense, I'd suggest filing a spec issue requesting that the spec is updated to match, noting which browsers already have this behavior.

Where Firefox and Safari already have the same behavior and pass the test, I wouldn't wait until the spec is changed to align Chromium. That's just creating a risk that the interop issue will turn into a real site compat issue while we wait.

[svgeesus]

Any number interpolated with "none" should result in no change, regardless of the alphas involved. I really don't see anywhere in the spec where this is mentioned though. I would not be surprised at all if the spec does change.

From 12.2. Interpolating with Missing Components:

If a color with a carried forward missing component is interpolated with another color which is not missing that component, the missing component is treated as having the other color’s component value.

and

If the carried forward missing component is alpha, the color must be premultiplied with this carried forward value, not with the zero value that would have resulted from color conversion.

and then from :

otherwise, each component which had been premultiplied is divided by the interpolated alpha value.

@mysteryDate You are suggesting that carried forward values should not get un-premultiplied? But did, earlier, get premultiplied?

[mysteryDate]

Ah, I misundersood

If a color with a carried forward missing component is interpolated with another color which is not missing that component, the missing component is treated as having the other color’s component value.

If you follow through my logic in #43180 (comment), I was making this subsitution after pre-multiplying (assuming none * x = none), so the result was different. Clearly the substitution should be made before premultiplying.

My misunderstanding:
color-mix(in lab, lab(none 0 0), lab(100 0 0 / 0.6)
Premultiply first: lab(none 0 0), lab(60 0 0 / 0.6)
Interpolating and substituting missing components: lab(60 0 0 / 0.8)
Unpremultiply: lab(62.5 0 0 / 0.8)

Subsituting before premultiplying:
color-mix(in lab, lab(none 0 0), lab(100 0 0 / 0.6)
Substution first: lab(100 0 0), lab(60 0 0 / 0.6)
Interpolating: lab(80 0 0 / 0.8)
Un-premultiplying: lab(100 0 0 / 0.8)

Perhaps it's worth noting in 12.2. Interpolating with Missing Components that pre-multiplication does not affect the result? i.e. if a component is missing in one color, but not missing in the other, the resulting value is equal to the non-missing value, regardless of alpha premultiplication. Or maybe give an example?

After a little bit of algebra it's now clear to me that lerp(x * a1, x * a2, t) / lerp(a1, a2, t) is always equal to x, but I certainly didn't see that at first!

@mysteryDate You are suggesting that carried forward values should not get un-premultiplied? But did, earlier, get premultiplied?

No, I just mis-understood the order. Apologies for my mistake. I'll fix Chromium.

[svgeesus]

If you follow through my logic in #43180 (comment), I was making this subsitution after pre-multiplying (assuming none * x = none), so the result was different. Clearly the substitution should be made before premultiplying.

There is an explicit order of operations, in 12. Color Interpolation. Was it ambiguous or maybe you just missed it?

Perhaps it's worth noting in 12.2. Interpolating with Missing Components that pre-multiplication does not affect the result? i.e. if a component is missing in one color, but not missing in the other, the resulting value is equal to the non-missing value, regardless of alpha premultiplication. Or maybe give an example?

Always happy to point out consequences and add examples!

[svgeesus]

@mysteryDate

Clearly the substitution should be made before premultiplying.

Clarified in w3c/csswg-drafts@f4e79df

[mysteryDate]

Great! Thank you, that clears up any nits I have.