[css-color-5] Clarification on resolved value of the alpha() function

[weinig]

The computed value serialization of the newly added alpha() RCS function seem inconsistent, and I am not sure why they are doing what they are doing.

When currentcolor is not involved, the tests require that legacy color serialization is maintained.

test_computed_value("color", "alpha(from red / 0.5)", "rgba(255, 0, 0, 0.5)");

But, when currentcolor is involved, it seems to use color() function syntax for serialization, even though the currentColor resolves to a legacy color.

function test_currentcolor_alpha(specified, expected, current) {
    target.style["color"] = current;
    fuzzy_test_computed_color_property("background-color", specified, expected);
}

test_currentcolor_alpha("alpha(from currentcolor / 0.5)", "color(srgb 1 0 0 / 0.5)", "red");

Is there a rule I am missing that specifies one vs. another?

cc @svgeesus

[weinig]

My preference would be that using alpha() always upgrades the color to its most modern form (e.g. rgb() -> color(srgb)), like we do with the other RCS functions (including the parts about maintaining hsl()/hwb() when none is used). It's not a straight analogy, since the other RCS functions do this based on the outer function, not the origin, but I think the principle still holds.

[kbhomes]

I filed an issue on the WPT repository for this same serialization issue on the from currentcolor tests: web-platform-tests/wpt#60611 (hadn't seen this one when I filed). Based on the existing rules in CSS Color 4 and 5, it seems that these tests shouldn't be serializing as color() because alpha() uses the same color space as the origin color and therefore uses their serialization rules. Reading the spec, I think a resolved color can only serialize to color(...) if it was originally specified using the color() function.

[weinig]

No, relative color syntax always (prior to alpha()) produces color(srgb) for things like hsl(from …) and rgb(from …).

[kbhomes]

Oh, you're totally right, I completely missed the Level 5 serializing relative color section and was working off just the Level 4 serializing values section.

[yisibl]

If serialized as color(), what is the computed value of alpha in the expression alpha(from rgba(0, 255, 0, 0.5) / alpha)?

Taking the relative color rgb(from rgba(0, 255, 0, 0.5) r g b / alpha) as an example,

• Chrome calculates it as 128/255 ≈ 0.501961, which is: color(srgb 0 1 0 / 0.501961)
• Safari and Firefox calculate it as 0.5

Interoperability is currently very poor and there is no WPT test cases.

@weinig What is the current implementation algorithm in WebKit?

cc @LeaVerou

[svgeesus]

My preference would be that using alpha() always upgrades the color to its most modern form (e.g. rgb() -> color(srgb)), like we do with the other RCS functions (including the parts about maintaining hsl()/hwb() when none is used). It's not a straight analogy, since the other RCS functions do this based on the outer function, not the origin, but I think the principle still holds.

I agree with the analogy and have just made that explicit in the spec by defining the relative color processing space which, for the alpha() form, is the color space of the origin color; otherwise, it is the color space of the relative color:

The origin color and the relative color need not use the same color function. Thus, we define a relative color processing space, which is the color space within which computation of color values takes place; this also affects the serialization of the result of the relative color.

For the alpha-only form, the relative color processing space is that of the origin color.

For the color-affecting forms, the relative color processing space is that of the relative color function.

I then use that term consistently while describing computation, resolution, and serialization. The table in 11.3. Serializing Relative Color Functions also uses that term, and thus we end up with the modern syntax for the serialization of the result of the alpha() function.

[svgeesus]

Taking the relative color rgb(from rgba(0, 255, 0, 0.5) r g b / alpha) as an example,

* Chrome calculates it as **`128/255 ≈ 0.501961`**, which is: `color(srgb 0 1 0 / 0.501961)`

I don't see anything in Color 4 or 5 that requires doing 8-bit computation on alpha values. 0.5 is clearly the correct result here.

Edit: Ah, actually Serializing alpha values has a historical/legacy carve-out:

If the value is internally represented as an integer between 0 and 255 inclusive (i.e. 8-bit unsigned integer), follow these steps:

So if you are storing alpha a uint8_t then 0.5 will be stored as 128 which, indeed, does not have a [0,100] integer serialization so comes back out as a float with up to six significant figures. While if you are storing alpha as, say, a half float then you store 0.5 and get back 0.5.

This is unfortunate for interop but we are constrained by historical behavior here.