This improves the quality of our font rendering, especially when
animations are involved. Relevant changes:
* Skia fonts have their subpixel flag set, which means that individual
glyphs are rendered at subpixel offsets causing glyph runs as a
whole to look better.
* Fragment offsets are no longer rounded to whole device pixels, and
instead the floating point offset is kept. This allows us to pass
through the floating point baseline position all the way to the Skia
calls, which already expected that to be a float position.
The `scrollable-contains-table.html` ref test needed different table
headings since they would slightly inflate the column size in the test
file, but not the reference.
If not set, when copying pixels into the SkImage, skia assumes that the
color space is the same as the input, so no transformation is done. We
are currently setting the color space to sRGB, this is fine for now as
it allows us to start making some transformations, but down the road we
will want to set that to the actual output's display color space.
See previous the commit description for more details about the floating
points operations.
The hwb test cases in `css-color-functions` are now rendered identically
to what firefox does (I haven't checked the others tests, but they
aren't affected by this commit).
In commit 1b82cb43c2 I accidentally
removed the paint transformation altogether. The result was that
zoomed-in SVGs, or SVG elements with a transformation applied could have
their gradient coordinates misplaced significantly.
This was also exposed in the `svg-text-effects` test by way of a slight
visual difference. Add a new test that very clearly exposes the fixed
issue by rotating the gradient coordinates by 45 degrees.
We were transforming coordinates for SVG gradients in a pretty
convoluted way: an inverse, unscaled transformation matrix was set up in
order to work around some (old?) technical limitations.
Rework this so the coordinate transformation no longer needs to be
inversed. This fixes gradients with "userSpaceOnUse" for its
gradientUnits attribute, which might cause coordinates to lie outside of
the bounding box of the gradient.
Two tests have updated reference screenshots with minor pixel updates;
this is probably the result of floating point precision improvements by
not inversing the matrix.
One test (svg-text-effects) has a bigger change: the gradient stops seem
to have moved along the text. This does seem to match other browsers
slightly better, so I'm moving forward with this ref update.
Prior to this change, SVGs were following the CSS painting order, which
means SVG boxes could have established stacking context and be sorted by
z-index. There is a section in the spec that defines what kind of SVG
boxes should create a stacking context
https://www.w3.org/TR/SVG2/render.html#EstablishingStackingContex
Although this spec is marked as a draft and rendering order described in
this spec does not match what other engines do.
This spec issue comment has a good summary of what other engines
actually do regarding painting order
https://github.com/w3c/svgwg/issues/264#issuecomment-246432360
"as long as you're relying solely on the default z-index (which SVG1
does, by definition), nothing ever changes order when you apply
opacity/filter/etc".
This change aligns our implementation with other engines by forbidding
SVGs to create a formatting context and painting them in order they are
defined in tree tree.
Instead of converting images to alpha masks on the CPU, we now delegate
that work to the GPU if possible, by way of SkSL shaders.
This noticeably speeds up https://vercel.com/ which has a ton of SVG
masking going on. The old implementation used 15% of CPU time when
loading the page, this one uses basically none.
Our current text iterator is not aware of multi-code point graphemes.
Instead of simply incrementing an iterator one code point at a time, use
our Unicode grapheme segmenter to break text into fragments.
This change should move us forward toward emoji support, as we are no
longer limited by our own OpenType implementation, which was failing
to parse the TrueType Collection format used to store emoji fonts
(at least on macOS).
Append text chunks to either the start or end of the text fragment,
depending on the text direction. The direction is determined by what
script its code points are from.
Instead of CSSColorValue holding a Gfx::Color, make it an abstract class
with subclasses for each different color function, to match the Typed-OM
spec. This means moving the color calculations from the parsing code to
the `to_color()` method on the style value.
This lets us have calc() inside a color function, instead of having to
fully resolve the color at parse time. The canvas fillStyle tests have
been updated to reflect this.
The other test change is Screenshot/css-color-functions.html: previously
we produced slightly different colors for an alpha of 0.5 and one of
50%, and this incorrect behavior was baked into the test. So now it's
more correct. :^)
Always assuming unpremultiplied color data only worked for PNGs (which
are specced as unpremultiplied) and bitmaps with alpha set to 100%.
Properly propagate the Gfx::AlphaType of a bitmap to Skia.
The reference tests were updated to reflect this change, but visually
it's practically impossible to see the difference. A new test was added
to clearly expose this issue.
Fixes#1104
Instead, it could be applied directly as a clip path in Skia painter.
As a side bonus, we get rid of some DeprecatedPath and
AntiAliasingPainter usage.
When converting a `Gfx::Bitmap` to a Skia bitmap, we cannot assume the
color data is unpremultiplied. For example, everything canvas-related
uses premultiplied color data:
https://html.spec.whatwg.org/multipage/canvas.html#premultiplied-alpha-and-the-2d-rendering-context
We were probably assuming unpremultiplied since that is what the PNG
decoder gives us. Since we now make `Gfx::Bitmap` identify what alpha
type is being used, we can instruct Skia a bit better :^)
Update our `EdgeFlagPathRasterizer` to use premultiplied alpha instead
of unpremultiplied so we can apply alpha correctly for path masks.
This fixes the dark borders sometimes visible when SVGs are blended
with a colored background.
This also exposed an issue with our `CanvasRenderingContext2D`, which is
supposed to hold a bitmap with premultiplied alpha internally but expose
a bitmap with unpremultiplied alpha in `CanvasImageData`. Expand our C2D
test to include the alpha channel as well.
Finally, this also exposed an off-by-one issue in
`EdgeFlagPathRasterizer` which caused the last scanlines for edges to
render incorrectly. We had some reference images which included these
corruptions (they were almost unnoticeable), so update them as well.
The main incentive is much better performance. We could have gone a bit
further in optimizing the Skia painter to blit glyphs produced by LibGfx
more efficiently from the glyph atlas, but eventually, we also want Skia
to improve correctness.
This change does not completely replace LibGfx in text handling. It's
still used at all stages, including layout, up until display list
replaying.
Before this change, "background-clip: text" was implemented by saving a
Vector<Gfx::Path> of all glyphs needed to paint a mask for the
background. The issue with this approach was that once glyphs were
extracted into vector paths, the glyph rasterization cache could no
longer be utilized.
With this change, all text required for mask painting is saved in a
nested display list and rasterized as a regular text.
With this change, instead of recording a display list item for each
instance of a repeated background, a new DrawRepeatedImmutableBitmap
type is used. This allows the painter to use optimized repeated image
painting and, when the GPU backend is used, avoid re-uploading the image
texture for each repetition.
Some screenshot tests are affected, but there are no visible
regressions.
https://null.com/games/chainstaff works a lof faster with this change.
Skia painter is visibly faster than LibGfx painter and has more complete
CSS transforms support. With this change:
- On Linux, it will try to use Vulkan-backend with fallback to
CPU-backend
- On macOS it will try to use Metal-backend with fallback to
CPU-backend
- headless-browser always runs with CPU-backend in layout mode
These test work with LibGfx painter but won't longer work after
switching to Skia, because it produces slightly different antialiasing,
rounding in color blending, etc.
This change will make it easier to disable screenshot comparison tests
on a specific platform or have per-platform expectations.
Additionally, it's nice to be able to tell if a ref-test uses a
screenshot as an expectation by looking at the test path.
