The `Bitmap` type was referring to to its internal pixel format by a
name that represents the order of the color components as they are layed
out in memory. Contrary, the `Color` type was using a naming that where
the name represents the order of the components from most to least
significant byte when viewed as a unsigned 32bit integer. This is
confusing as you have to keep remembering which mental model to use
depending on which code you work with.
To unify the two, the naming of RGBA-like colors in the `Color` type has
been adjusted to match the one from the Bitmap type. This seems to be
generally in line with how web APIs think about these types:
* `ImageData.pixelFormat` can be `rgba-8unorm` backed by a
`Uint8ClamedArray`, but there is no pixel format backed by a 32bit
unsigned type.
* WebGL can use format `RGBA` with type `UNSIGNED_BYTE`, but there is no
such format with type `UNSIGNED_INT`.
Additionally, it appears that other browsers and browser-adjacent
libraries also think similarly about these types:
* Firefox:
https://github.com/mozilla-firefox/firefox/blob/main/gfx/2d/Types.h
* WebKit:
https://github.com/WebKit/WebKit/blob/main/Source/WebCore/platform/graphics/PixelFormat.h
* Skia:
https://chromium.googlesource.com/skia/+/refs/heads/main/include/core/SkColorType.h
This has the not so nice side effect that APIs that interact with these
types through 32bit unsigned integers now have the component order
inverted due to little-endian byte order. E.g. specifying a color as hex
constant needs to be done as `0xAABBGGRR` if it is to be treated as
RGBA8888.
We could alleviate this by providing endian-independent APIs to callers.
But I suspect long-term we might want to think differently about bitmap
data anyway, e.g. to better support HDR in the future. However, such
changes would be more involved than just unifying the naming as done
here. So I considered that out of scope for now.
Bitmap::get_pixel() was only handling two out of the four possible pixel
formats, asserting when called with the other two. The asserting code
path was triggered when loading JPEG XL images, causing crashes on pages
like https://jpegxl.info/resources/jpeg-xl-test-page or
https://html5test.co/.
- Hue now wraps properly when negative or larger than 360
- The hsl to rgb conversion now closely mirrors the code example from
the spec.
This fixes a number of WPT tests in
/css/css-color/parsing/color-computed-hsl.html
TL;DR: There are two available sets of coefficients for the conversion
matrices from XYZ to sRGB. We switched from one set to the other, which
is what the WPT tests are expecting.
All RGB color spaces, like display-p3 or rec2020, are defined by their
three color chromacities and a white point. This is also the case for
the video color space Rec. 709, from which the sRGB color space is
derived. The sRGB specification is however a bit different.
In 1996, when formalizing the sRGB spec the authors published a draft
that is still available here [1]. In this document, they also provide
the matrix to convert from the XYZ color space to sRGB. This matrix can
be verified quite easily by using the usual math equations. But hold on,
here come the plot twist: at the time of publication, the spec contained
a different matrix than the one in the draft (the spec is obviously
behind a pay wall, but the numbers are also reported in this official
document [2]). This official matrix, is at a first glance simply a
wrongly rounded version of the one in the draft publication. It however
has some interesting properties: it can be inverted twice (so a
roundtrip) in 8 bits and not suffer from any errors from the
calculations.
So, we are here with two versions of the XYZ -> sRGB matrix, the one
from the spec, which is:
- better for computations in 8 bits,
- and official. This is the one that, by authority, we should use.
And a second version, that can be found in the draft, which:
- makes sense, as directly derived from the chromacities,
- is publicly available,
- and (thus?) used in most places.
The old coefficients were the one from the spec, this commit change them
for the one derived from the mathematical formulae. The Python script to
compute these values is available at the end of the commit description.
More details about this subject can be found here [3].
[1] https://www.w3.org/Graphics/Color/sRGB.html
[2] https://color.org/chardata/rgb/sRGB.pdf
[3] https://photosauce.net/blog/post/making-a-minimal-srgb-icc-profile-part-3-choose-your-colors-carefully
The Python script:
```python
# http://www.brucelindbloom.com/index.html?Eqn_RGB_XYZ_Matrix.html
from numpy.typing import NDArray
import numpy as np
### sRGB
# https://www.w3.org/TR/css-color-4/#predefined-sRGB
srgb_r_chromacity = np.array([0.640, 0.330])
srgb_g_chromacity = np.array([0.300, 0.600])
srgb_b_chromacity = np.array([0.150, 0.060])
##
## White points
white_point_d50 = np.array([0.345700, 0.358500])
white_point_d65 = np.array([0.312700, 0.329000])
#
r_chromacity = srgb_r_chromacity
g_chromacity = srgb_g_chromacity
b_chromacity = srgb_b_chromacity
white_point = white_point_d65
def tristmimulus_vector(chromacity: NDArray) -> NDArray:
return np.array([
chromacity[0] /chromacity[1],
1,
(1 - chromacity[0] - chromacity[1]) / chromacity[1]
])
tristmimulus_matrix = np.hstack((
tristmimulus_vector(r_chromacity).reshape(3, 1),
tristmimulus_vector(g_chromacity).reshape(3, 1),
tristmimulus_vector(b_chromacity).reshape(3, 1),
))
scaling_factors = (np.linalg.inv(tristmimulus_matrix) @
tristmimulus_vector(white_point))
M = tristmimulus_matrix * scaling_factors
np.set_printoptions(formatter={'float_kind':'{:.6f}'.format})
xyz_65_to_srgb = np.linalg.inv(M)
# http://www.brucelindbloom.com/index.html?Eqn_ChromAdapt.html
# Let's convert from D50 to D65 using the Bradford method.
m_a = np.array([
[0.8951000, 0.2664000, -0.1614000],
[-0.7502000, 1.7135000, 0.0367000],
[0.0389000, -0.0685000, 1.0296000]
])
cone_response_source = m_a @ tristmimulus_vector(white_point_d50)
cone_response_destination = m_a @ tristmimulus_vector(white_point_d65)
cone_response_ratio = cone_response_destination / cone_response_source
m = np.linalg.inv(m_a) @ np.diagflat(cone_response_ratio) @ m_a
D50_to_D65 = m
xyz_50_to_srgb = xyz_65_to_srgb @ D50_to_D65
print(xyz_50_to_srgb)
print(xyz_65_to_srgb)
```
Truncating the value is mathematically incorrect, this error made the
conversion to grayscale unstable. In other world, calling `to_grayscale`
on a gray value would return a different value. As an example,
`Color::from_string("#686868ff"sv).to_grayscale()` used to return
#676767ff.
