In this PR we moved global alpha handling using temporary layers to
mutation of paint in vello_cpu (we were already doing this in vello
classic). This + not using temporary layer for SrcOver (default and most
common composition operation) allows us to remove most temporary layers
from `with_composition`.
This slightly improves performance of vello backend, but drastically
improves performance of vello_cpu. We are now able to render bunnymark
(100 bunnies) with 60 FPS.
In the future we could cache current layer and change it when
compositing operation changes, although that would complicate clips, so
improvement is questionable.
Testing: Existing WPT tests for functionality, but we do not have any
performance tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
According to my tests `OptimizeSpeed` slight improves performance and it
does NOT affect WPT results in negative way.
Testing: Tested by existing WPT tests.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
As noted in #38345, vello scenes only grow. While we can reset them when
clearing viewport (#38356) that is not enough. We need to reset scene on
each render (~each frame) and providing old frame as backdrop to new
scene. Be do this lazily so multiple rendering without any changes
should be cheaper, we still do GPUBuffer mapping, because that would
require more complex work.
Testing: Code functionality is covered by existing WPT tests, but we do
not have any performance test.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Vello scene only ever grows, so we need to clear it as soon as it's
possible (in clear rect). This PR also adds ignore_clips to
vello_backend (already exists in vello_cpu_backend).
Testing: Behavior is verified by existing tests, as this is mainly a
change for performance. There are currently no performance tests. This
makes bunnymark actually playable (from 3 FPS to 20 FPS on vello_cpu).
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
`create_similar_draw_target` is more performant then creating completely
new target (this creates new wgpu device in vello backend).
Testing: This change does not modify test results, but should increase
performance.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
We want to eventually remove raqote backend, but for now we can gate it
behind a feature (still enabled by default in servoshell) like the rest
of backends. `dom_canvas_backend=auto` will select first available
backend. Builds on top of https://github.com/servo/servo/pull/38310 to
support cases where no backend is available.
Testing: It compiles with or without feature
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Before script just crashed in those cases because IPCSender was dropped,
now we send `None` to tell script about the failure and fail getContext
or registerPainter accordingly.
This PR also unifies `dom_canvas_{backends}_enabled` prefs into
`dom_canvas_backend` which is more flexible in multi-backends scenarios.
Reviewable per commit.
Testing: Added servo specific WPT test.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
vello_cpu does not have any tests timeouts, because we do not need
download stuff from GPU as all work happens on CPU. So performance wise
it's better then classic vello at least for our usecase. There are some
vello bugs, but I think we will be able to sort them out within
upstream, eventually. Interestingly enough there are no new PASS like
they were with classic vello.
Difference with raqote can be observed here:
https://github.com/sagudev/servo/actions/runs/16549241085/attempts/1#summary-46802486798
## Known vello problems:
- https://github.com/linebender/vello/issues/1119
- https://github.com/linebender/vello/issues/1056
-
`/html/canvas/element/fill-and-stroke-styles/2d.gradient.interpolate.coloralpha.html`
- `kurbo::Cap::Butt` is defect (only visible with big lineWidth)
https://github.com/linebender/vello/issues/1063
- `/html/canvas/element/line-styles/2d.line.cross.html`
- `/html/canvas/element/line-styles/2d.line.miter.acute.html`
- other lack of strong correct problems
(https://github.com/linebender/vello/issues/1063#issuecomment-2998084736):
- `/html/canvas/element/path-objects/2d.path.rect.selfintersect.html`
- `putImageData(getImageData(...), ...)` is lossy (precision problems,
might be due to ImageData being unmultiplied)
-
`/html/canvas/element/pixel-manipulation/2d.imageData.put.unchanged.html`
Testing: Tested using vello_cpu_canvas subsuite
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
We just need to pass user provided FillRule via IPC to canvas paint
thread, then pass it all down to backend, which will handle it.
Testing: Added WPT tests.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This PR changes surface type from `Vec<u8>` to `Vec<u32>` as this is
native type of raqote's backend. We used unsafe for `Vec<u8>` <->
`Vec<u32>` casting that is replaced with bytemuck. Bytemuck is already
in cargo lock.
Testing: Existing WPT tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This will allow reusing those conversions in vello_cpu backend.
Testing: Just refactor
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Add vello backend by implementing Backend traits in canvas crate (so
this lives in canvas_paint_thread - embedded process). Current
implementation uses normal wgpu, so we block on GPU work. Vello backend
is gated behind `vello` feature and `dom_canvas_vello_enabled` pref.
Feature-wise this backend is on on par with raqote (sometimes better
sometimes worse), but performance wise it's worse.
## Known vello problems:
- image roundtrip does not work (fixed in
https://github.com/linebender/vello/pull/974)
- https://github.com/linebender/vello/issues/1066 (fixed)
- clip layers are not working properly:
https://github.com/linebender/vello/issues/1061
- `/html/canvas/element/pixel-manipulation/2d.imageData.put.*`
- `/html/canvas/element/path-objects/2d.path.clip.intersect.html`
- https://github.com/linebender/vello/issues/1056
-
`/html/canvas/element/fill-and-stroke-styles/2d.gradient.interpolate.coloralpha.html`
- `kurbo::Cap::Butt` is defect (only visible with big lineWidth)
https://github.com/linebender/vello/issues/1063
- `/html/canvas/element/line-styles/2d.line.cross.html`
- `/html/canvas/element/line-styles/2d.line.miter.acute.html`
- other lack of strong correct problems
(https://github.com/linebender/vello/issues/1063#issuecomment-2998084736):
- `/html/canvas/element/path-objects/2d.path.rect.selfintersect.html`
- There is currently no way to do put image properly in vello as we
would need to ignore all clips and other stuff (we try to work around
this on best effort basis)
https://github.com/linebender/vello/issues/1088
- `/html/canvas/element/pixel-manipulation/2d.imageData.put.*`
- precision problems
- `/html/canvas/element/path-objects/2d.path.stroke.scale2.html`
- `/html/canvas/element/path-objects/2d.path.arc.scale.1.html`
## Known servo problems
- bad performance due to blocking on GPU work
- some get/put intensive tests `TIMEOUT`
- proper shadow support (non-blocker as we already are living without it
now)
- support for rect shadow is there but unimplemented currently as that's
the state in raqote
Testing: `mach try vello` will run normal WPT (with raqote) +
vello_canvas subsuite that runs only on `/html/canvas/element`. All
subsuite expectations are stored separately.
Fixes: #36823Fixes: #35230
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Currently we had `GetImageData` and `SendPixels` to obtain pixels from
script thread. This PR unifies those into single `GetImageData` that
does not need canvas size and has optional rect (for obtaining sub
image).
Testing: Existing WPT tests
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This will be needed for vello_cpu. While we could wrap it in RefCell for
inner mut, but that would be less ergonomic and performant.
Testing: Just refactoring, but the code is covered by WPT tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
After #38214 `Backend` trait is only used as DrawTarget builder and as
type holder. By moving types and DrawTarget creation into
`GenericDrawTarget` trait, we can completely remove `Backend` trait.
Testing: Just refactor, but code is covered by WPT tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
I think this simplifies canvas backends greatly. Before there were many
(needless) hoops from abstract to concrete (backend) and back, now we
can do most stuff on abstract types.
Testing: Existing WPT tests
Fixes: #38022
try run: https://github.com/sagudev/servo/actions/runs/16450978211
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Remove unused `Cargo.toml` dependency to reduce binary size.
Testing: Can still compile in different platforms.
Signed-off-by: Euclid Ye <euclid.ye@huawei.com>
This PR removes all `Set*` commands as we will send needed options on
each request. This will allow us to remove most of state handling from
canvas paint thread (currently it's still stateful).
Testing: Existing WPT tests
work towards #38022
try run: https://github.com/sagudev/servo/actions/runs/16413823963
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This PR moves current path handling from canvas paint thread to script's
`CanvasState`. Because we used manual impl for arcto and some prechecks
for early fail before sending IPC we also needed to fix some bugs in
Path impl.
Testing: Existing WPT tests
work towards #38022
try run: https://github.com/sagudev/servo/actions/runs/16316090028
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
We already store transform in context state, so let's just use this when
querying instead of using IPC to ask canvas paint thread.
Testing: Existing WPT tests
work towards #38022
try run: https://github.com/sagudev/servo/actions/runs/16299182583
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Before we stored path in user or device space and do conversions as
needed, but that complicated things.
So we got two options:
1. store path in device space: here we would need to add
`Option<Transform>` to each path command, and before each usage we need
to invert current transform (as current transform is already applied by
stroke/fill and we cannot just pass I, because we need scaling to make
lines thicker)
2. store path in user space and do transform whenever we change the
transform. There is no need to do inverse on uses.
I chose option 2. because it's less complicated and will probably
benefit performance (set transform typically called more rarely than
path building/fill/stroke).
In follow up PR I will move all of this to script, that's why
PathBuilderRef was not removed yet.
Testing: Existing WPT tests
work towards #38022
try run: https://github.com/sagudev/servo/actions/runs/16304221495
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This PR removes existing path(segment) abstractions in favor of
`kurbo::BezPath`, well actually wrapped `kurbo::BezPath`, to ensure
building of valid paths. This allows us better Path2D building in script
and doing all validation and segmentation there and also allows us
remove blocking is_point_in_path on Path2D as we can now do this in
script. Current path is still done on canvas thread side as it will be
harder to move to script (will be done as a follow up), but it now uses
this new path abstraction.
Using kurbo also allows us to ditch our manual svgpath parser with the
one provided by kurbo.
Same code is stolen from: https://github.com/servo/servo/pull/36821.
Testing: Existing WPT tests
Fixes: #37904
wpt run: https://github.com/sagudev/servo/actions/runs/16172191716
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Before we only handled no-repeat for rect, this means we rendered
https://sagudev.github.io/briefcase/no-repeat.html incorrectly (like
firefox). Now if one of the axis is bounded (does not repeat) we clip it
and let other axis be unbounded (technically we clip it to end of
canvas).
This is also needed for vello backend.
Testing: Tests in WPT exists and another test is added.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
This removes assumption about pixel format from backend abstraction to
actual backend implementation. This is important for vello.
Testing: WPT tests
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Dealing with both path and builder makes code complicated and this is
backend internal specific anyway (we often do conversions between the
two just so we get specific functionality). Now I joined both into
single abstractions `Path`, that has all what one needs and let raqote
backend deal with path/pathbuilder coversions.
Motivation: Simplification and in Vello there is only
[BezPath](https://docs.rs/kurbo/0.11.2/kurbo/struct.BezPath.html) so we
use it as both path and pathbuilder.
Reviewable per commit.
Testing: Existing WPT tests
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
I introduced snapshot in #36119 to pack raw bytes and metadata together,
now we take the next step and require for user to always specify what
kind of byte data they want when calling `as_bytes` or `to_vec`
(basically joining transform and data). There are also valid usages when
one might require just one property of bytes (textures can generally
handle both RGBA and BGRA). There are also valid usages of using just
raw bytes (when cropping). This PR tries to make such usages more
obvious.
This will make it easier to fix stuff around 2d canvas (we do not want
to assume any bytes properties in abstraction).
Testing: Code is covered by WPT tests.
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Co-authored-by: Martin Robinson <mrobinson@igalia.com>
This creates a new method in shared/compositing/lib to generate image
keys that are send over the webview. This does not immediately return
the keys but goes over the constellation to receive the keys from the
IOCompositor. To make this more efficient, we now cache the keys in
image_cache in a simple FIFO order. The old blocking method stays intact
for now but got renamed to make the blocking clear.
The blocking calls that are left are in:
- `components/canvas/canvas_data.rs`
- `components/script/dom/htmlmediaelement.rs`
Testing: WPT tests should cover this as this doesn't change any
functionality.
Fixes: Was mentioned in
https://github.com/servo/servo/issues/37161#issuecomment-2915750051 and
part of https://github.com/servo/servo/issues/37086
---------
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
Signed-off-by: gterzian <2792687+gterzian@users.noreply.github.com>
Co-authored-by: gterzian <2792687+gterzian@users.noreply.github.com>
While raqote does already passes this test, but small line is still
visible and this PR will fix this. Vello will not work correctly without
this.
Testing: `/html/canvas/element/path-objects/2d.path.arc.twopie.1.html`
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Before we only offered helper to add single image (no update or delete)
that got special IPC message, now we simplify this by offering all ops
helpers for dealing with single image (that happens most of the time),
that simply uses `update_images` under the hood. We also optimize for
this use case by using `SmallVec<[ImageUpdate; 1]>` to avoid alloc.
Testing: Just refactor, but code is covered by WPT tests
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Follow the HTML canvas specification and add missing
'setTransform(transform)' method to CanvasPattern interface.
https://html.spec.whatwg.org/multipage/#dom-canvaspattern-settransform
Testing: Improvements in the tests
- html/canvas/element/fill-and-stroke-styles/2d.pattern.transform*
- html/canvas/offscreen/fill-and-stroke-styles/2d.pattern.transform*
Signed-off-by: Andrei Volykhin <andrei.volykhin@gmail.com>
This causes wrong rendering on vello. Current backends already handle
one dim zero correctly by themself (this was probably needed for azure),
but if we ever needed this, it should be implemented inside faulty
backend.
Testing: Existing WPT tests
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
This makes it easier to add variants, I also added some spec links and
reordered enum variants to match the spec.
Testing: Just refactor, but the code is covered by WPT tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Follow the HTML canvas specification and empty the list of subpaths
in context's current default path (step 2) on canvas context reseting
(reset() or set bitmap dimestion).
https://html.spec.whatwg.org/multipage/#reset-the-rendering-context-to-its-default-state
Testing: Improvements in the tests
- html/canvas/element/canvas-host/2d.canvas.host.initial.reset.path.html
- html/canvas/offscreen/canvas-host/2d.canvas.host.initial.reset.path*
Signed-off-by: Andrei Volykhin <andrei.volykhin@gmail.com>
We pass all other options as ref too.
Testing: Just refactor, but the code is covered by WPT tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
All other draw functions already have them non-optional and we always
passed `Some` anyway.
Testing: Existing WPT tests
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
1. The `shared` directory is for the "_traits" crates, which will likely
be moved out of this directly at some point and renamed "_api". These
crates expose the API of crates to avoid circular dependencies.
`Snapshot` isn't really this.
2. `Snapshot` is essentially a specialied kind of `Image` so it makes
sense that it is grouped with other image-related things in `pixels`.
Testing: This should not change any behavior so is covered by existing
tests.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
When the path is created with an uninvertible transform, don't do any
path-based operations. Now `ensure_path()` returns a `Path` object and
this change adds early returns which prevent trying to use it.
Testing: This change adds a WPT crash test.
Fixes#36995.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Replaces uses of `euclid::default::Size2D<u64>` with
`euclid::default::Size2D<u32>` for the canvas/context/snapshot.
This PR includes changes to the following components:
- canvas
- pixels
- script
- script_bindings
- shared/canvas
- shared/snapshot
Testing: https://github.com/hashcatHitman/servo/actions/runs/15426115391
(as of 892edc0048)
Fixes: #36706
---------
Signed-off-by: hashcatHitman <155700084+hashcatHitman@users.noreply.github.com>
The canvas thread might need access to the system font service before it
shuts down. Ensure that it finishes shutting down before triggering the
shutdown of the system font service. This should avoid issues where
canvas tries to access fonts right before shutting down.
Fixes: #36849.
Testing: Since this fixes a flaky crash on shutdown, there isn't a good
way to write a test for it.
Signed-off-by: Martin Robinson <mrobinson@igalia.com>
Removed unused dependencies in various crates using cargo-machete, grep
and cargo-udeps.
Testing: The CI tests if any dependencies were used that are removed.
Signed-off-by: Narfinger <Narfinger@users.noreply.github.com>
There is no reason to require this impls, because we already have
written them in generic way.
Testing: Just refactoring, but there are WPT tests
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
`surface()` returns `SourceSurface` which is/was meant as optimization
when passing from canvas to canvas (in vello that would be wgpu texture;
but raquote does not really have this) while bytes returns something
that must impl AsRef<[u8]> (this is more generic then `&[u8]` as it
allows us to have type with drop impl - wgpu's BufferView).
Testing: This is just refactoring (protected by rust), but there are WPT
tests.
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Small fixes of abstraction.
We do not need to generalize `CompositionOrBlending`, because it's from
canvas_traits.
Ellipse impl that uses arc is not backend specific, so it serves as good
trait default.
Reviewable per commit.
Testing: Only refactoring, but there are WPT tests
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
Current abstraction was leaky and very hard to understand/impl because
it was so intervened with actual implementation.
Now we generalize both `CanvasState` and `CanvasData` over `B: Beckend`,
meaning that every type/method must be part of trait interface (that
lives in `beckend.rs`). Using associated trait types instead of `Box<dyn
>` allows us too remove the need for wrapper types (and `to_raquote()`
methods) as we can implement helper traits on (foreign) raquote types.
The only time we actually do dispatch (by enum) is at `Canvas` methods.
Implementation now only need to implement all backend traits and
helpers. I tried to restrain myself from actually cleaning abstraction
(where possible), to keep this change as much mechanical as possible,
but we should absolutely do that as a follow up.
Testing: Rust as we only do refactor, but there are also WPT tests
try run: https://github.com/sagudev/servo/actions/runs/14760658522
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
canvas crate was home for both webgl and 2dcanvas impl, but it makes
more sense to separate these two, by splitting webgl into standalone
crate (like we already have for webgpu).
Testing: Rust for fearless refactoring (but there are WPT tests just in
case)
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
To prevent any potential crash/OOM issues with "canvas" element
from "rogue" applications let's apply large size limitations for context
canvas2d's draw target to Servo (similar approach in Firefox/Chromium -
they limits width and height to 32767/65535 pixels).
Fixes: #36155, #34117, #30164, #24710
--
- [x] ./mach build -d does not report any errors
- [x] ./mach test-tidy does not report any errors
- [x] There are tests for these changes
tests/wpt/tests/html/canvas/element/canvas-host/2d.canvas.host.size.large.html
tests/wpt/tests/html/canvas/offscreen/canvas-host/2d.canvas.host.size.large.html
tests/wpt/tests/html/canvas/offscreen/canvas-host/2d.canvas.host.size.large.worker.js
Signed-off-by: Andrei Volykhin <andrei.volykhin@gmail.com>
Each canvas context returns snapshot instead of just raw bytes. This
allows as to hold off conversions (BGRA <-> RGBA, (un)premultiply) to
when/if they are actually needed. For example when loading snapshot into
webgl we can load both RGBA and BGRA so no conversion is really needed.
Currently whole thing is designed to be able to be extend on
https://github.com/servo/ipc-channel/pull/356, to make less copies.
Hence some commented out code.
Fixes#35759
There are tests for these changes in WPT
---------
Signed-off-by: sagudev <16504129+sagudev@users.noreply.github.com>
