Due to https://github.com/flutter/flutter/issues/127500 , we can get in a state where enable-impeller is true but we're using Skia. We need to either fall back completely to Skia, make this configuration fatal, or remote the check
----------------
Conversion of SkTextBlobs to impeller::TextFrame objects is one of the most expensive operations in display list dispatching. While the rest of the engine and framework makes a reasonable attempt to cache the SkTextBlobs generated during paragraph construction, the design of the dl dispatcher means that these the Impeller backend will always reconstruct all text frames on each frame - even if the display list/picture that contained those text frames was unchanged.
Removing this overhead is one of the goals of https://github.com/flutter/engine/pull/45386 , however this patch is also fairly risky and will be difficult to land. As a more incremental solution, we can instead construct the impeller::TextFrame objects when performing paragraph painting and record them in the display list. This both moves the text frame construction to the UI thread and allows the framework/engine to cache unchanged text frames.
This also does not conflict with the dl_aiks_canvas patch directly, and is fine to land before or after it does. (though I'd argue we should land this first).
To compare the current performance levels, I ran the complex_layout_scroll perf test, since this is fairly text filled. On a Pixel 6 pro. Across several runs this is a fairly consistent ~1ms raster time improvement.
Fixes https://github.com/flutter/flutter/issues/133204
Reverts flutter/engine#45418
Some google3 tests are hitting the CHECK I added in the DlSkCanvasDispatcher::drawTextFrame, which indicates that the SkParagraph code likely thinks impeller is enabled, whereas other code might be running with Skia.
Perhaps this could happen if its software rendering? It should be a fatal error on startup so we can track this down.
Conversion of SkTextBlobs to impeller::TextFrame objects is one of the most expensive operations in display list dispatching. While the rest of the engine and framework makes a reasonable attempt to cache the SkTextBlobs generated during paragraph construction, the design of the dl dispatcher means that these the Impeller backend will always reconstruct all text frames on each frame - even if the display list/picture that contained those text frames was unchanged.
Removing this overhead is one of the goals of https://github.com/flutter/engine/pull/45386 , however this patch is also fairly risky and will be difficult to land. As a more incremental solution, we can instead construct the impeller::TextFrame objects when performing paragraph painting and record them in the display list. This both moves the text frame construction to the UI thread and allows the framework/engine to cache unchanged text frames.
This also does not conflict with the dl_aiks_canvas patch directly, and is fine to land before or after it does. (though I'd argue we should land this first).
To compare the current performance levels, I ran the complex_layout_scroll perf test, since this is fairly text filled. On a Pixel 6 pro. Across several runs this is a fairly consistent ~1ms raster time improvement.
### Skia
```
"average_frame_build_time_millis": 1.497333333333333,
"90th_percentile_frame_build_time_millis": 2.038,
"99th_percentile_frame_build_time_millis": 17.686,
"worst_frame_build_time_millis": 23.095,
"missed_frame_build_budget_count": 3,
"average_frame_rasterizer_time_millis": 5.5078589743589745,
"stddev_frame_rasterizer_time_millis": 2.226343414420338,
"90th_percentile_frame_rasterizer_time_millis": 7.481,
"99th_percentile_frame_rasterizer_time_millis": 19.11,
"worst_frame_rasterizer_time_millis": 79.799,
"missed_frame_rasterizer_budget_count": 7,
"frame_count": 234,
"frame_rasterizer_count": 234,
"new_gen_gc_count": 10,
"old_gen_gc_count": 2,
```
### Impeller (ToT)
```
"average_frame_build_time_millis": 1.431575000000001,
"90th_percentile_frame_build_time_millis": 2.196,
"99th_percentile_frame_build_time_millis": 14.486,
"worst_frame_build_time_millis": 23.728,
"missed_frame_build_budget_count": 2,
"average_frame_rasterizer_time_millis": 6.536087499999999,
"stddev_frame_rasterizer_time_millis": 1.9902712500000004,
"90th_percentile_frame_rasterizer_time_millis": 9.705,
"99th_percentile_frame_rasterizer_time_millis": 14.727,
"worst_frame_rasterizer_time_millis": 17.838,
"missed_frame_rasterizer_budget_count": 1,
"frame_count": 240,
"frame_rasterizer_count": 240,
"new_gen_gc_count": 10,
"old_gen_gc_count": 2,
```
### Impeller (Patched)
```
"average_frame_build_time_millis": 1.4500167364016743,
"90th_percentile_frame_build_time_millis": 2.478,
"99th_percentile_frame_build_time_millis": 14.883,
"worst_frame_build_time_millis": 18.782,
"missed_frame_build_budget_count": 1,
"average_frame_rasterizer_time_millis": 5.023033333333336,
"stddev_frame_rasterizer_time_millis": 1.6445388888888894,
"90th_percentile_frame_rasterizer_time_millis": 7.814,
"99th_percentile_frame_rasterizer_time_millis": 13.497,
"worst_frame_rasterizer_time_millis": 15.008,
"missed_frame_rasterizer_budget_count": 0,
"frame_count": 239,
"frame_rasterizer_count": 240,
"new_gen_gc_count": 8,
"old_gen_gc_count": 0,
```
Relands https://github.com/flutter/engine/pull/45131
Fixes https://github.com/flutter/flutter/issues/132416
Differences from last time:
- Some minor merge conflict fixes
- Use the RTree to get the bounds instead of recalculating the bounds
- Make the iOS platform view controller implementation use the impeller-aware slices instead of the display list ones. This has been fixed for Android and the desktop embedding, but I missed iOS. The unit tests weren't actually running before I branched for my PR, @zanderso fixed them up separately and this resulted in catching the failures on post submit last time.
Reverts flutter/engine#45131
This is failing the Impeller variants of the unobstructed platform views tests:
https://ci.chromium.org/ui/p/flutter/builders/prod/Mac%20Production%20Engine%20Drone/132249/overview
```
Failing tests:
-[UnobstructedPlatformViewTests testPlatformViewsMaxOverlays]
-[UnobstructedPlatformViewTests testOneOverlay]
-[UnobstructedPlatformViewTests testOneOverlayPartialIntersection]
-[UnobstructedPlatformViewTests testTwoIntersectingOverlays]
-[UnobstructedPlatformViewTests testOneOverlayAndTwoIntersectingOverlays]
** TEST FAILED **
```
We've retried it a few times so I suspect this isn't a flake.
This is a reland of https://github.com/flutter/engine/pull/44248
Fixes https://github.com/flutter/flutter/issues/132416
Changes from last time:
- The `drawPoints` benchmark was failing to render anything meaningful because of an error in `AiksLayer` that resulted in an infinitely sized bounding rectangle poisoning the bounds with `NaN` (this happens, for example, with a `drawPaint` call, which that benchmark happens to use). Added a test covering this and filed https://github.com/flutter/flutter/issues/132770 to explore ways to avoid this in the future.
- There was a bug in `DlAiksCanvas::SaveLayer` where a `nullptr` `paint` but non-`nullptr` `backdrop` was failing to actually save the layer. This resulted in incorrect rendering.
- There was a bug in `impeller::Canvas::DrawPicture` that resulted in incorrect stencil depth counting. That was fixed separately by @bdero, but was the cause of incorrect rendering in some Wonderous screens.
- I've added a simple implementation for `AiksLayer::IsReplacing`. It does not currently compare as deeply as the `DisplayListLayer` version potentially does, but it is good enough to avoid the regression noted in https://github.com/flutter/flutter/issues/132071. That regression was on a benchmark that greatly benefits from partial repaint. With the new implementation, it still gains partial repaint where it previously did not. There is more work that can be done here, filed https://github.com/flutter/flutter/issues/133361 to track that work.
I merged but did not fully integrate the `DisplayListBuilder`/`CanvasToReceiver` work that @flar has done. I have a local experiment with that, but would prefer to see this land and run through the device lab so we get some better comparison numbers for which one performs better.
DisplayListBuilder grew over time from a class that implemented a developer-unfriendly stateful API into one that implemented both the developer-friendly DlCanvas API as well as its original stateful API. Over time, the stateful API was buried under a "testing only" facade.
In the meantime, the optimization features that it applies to the DlCanvas calls before it records the operations in a DlOp store are useful without the recording process and so I've been wanting to break those into 2 parts for a while with the goal of removing all stateful APIs from DisplayListBuilder (see https://github.com/flutter/flutter/issues/108303).
This PR takes a major step along that direction by splitting DisplayListBuilder into essentially a convenience class that marries 2 new classes together to achieve its old functionality:
- `DlCanvasToReceiver` - a class that implements DlCanvas, optimizes common situations, and then sends commands to any object that implements `DlOpReceiver`
- `DlOpRecorder` - an implementation of DlOpReceiver that records the operations in a buffer from which to create a `DisplayList` object
- `DisplayListBuilder` now inherits from DlCanvasToReceiver to get the optimizations and provides it with an instance of `DlOpRecorder` as the receiver that it will send its results to
- Similarly, a `DlCanvasToReceiver` instance could be directed to an `impeller:DlDispatcher` to achieve a more straight-through path from the DlCanvas interface to impeller Pictures.
Fixes https://github.com/flutter/flutter/issues/130141
The primary goal of this patch is to move dispatching of `dart:ui` `Canvas` commands to the UI thread.
Before this patch, the architecture is something like:
## UI Thread
- `dart:ui` talks to `DisplayListBuilder`, a `DlCanvas` implementation.
- `DisplayListBuilder` does some clip/bounds tracking and creates a `DisplayList` object that is held by `dart:ui`'s `Picture` objects.
- `DisplayList`s are added to `DisplayListLayer`s in `flow`.
## Raster Thread
- `flow` flattens the various operations into a single `DisplayList` via another `DisplayListBuilder`.
- A `DlOpReceiver`implementation converts that `DisplayList` into an `Aiks` `Canvas`/`Picture`.
After this patch, the architecture instead looks like:
## UI Thread
- No change for Skia.
- If Impeller, use a new `DlCanvasImplementation` that talks to `Aiks`'s `Canvas`.
- If Impeller, `dart:ui` Picture's now hold an `Aiks` `Picture`, which get shared into `AiksLayer`s in `flow`.
## Raster thread
- No change for Skia, but some light refactoring for places that assumed a `DisplayListBuilder` where they really just needed a `DlCanvas`.
- The `Aiks` `Picture`s are combined using new API on `DlCanvas` and still backed by `Aiks`.
These changes show significant improvement on raster times on Android and only very small regressions on UI times in local testing, see https://gist.github.com/dnfield/26528090194c9f5abdbac13cdcbf4f79 for old gallery transition perf numbers.
Many of the other changes in this patch are related to the following:
- Making `DlRTree` usable for Impeller.
- It would be nice to have a version of DlRTree that speaks `impeller::Rect`.
- Creating the requisite classes to support `EmbeddedViews` so that Desktop works.
This patch does not remove the `impeller::DlDispatcher`, which now would only be used in tests.
When a saveLayer is rendered with an ImageFilter that modifies the bounds of the rendered pixels, and some of the content of that saveLayer did not intersect the clip, but the filtered output of that content did intersect the clip, we might not accumulate the bounds of those rendering operations into the DisplayList bounds.
This bug was not encountered during practical testing, but showed up on some testing with the new NOP culling code.
For now, this bug only affects the accuracy of the reported bounds of the DisplayList, but that can affect raster caching and potentially the layer culling done in the LayerTree. It might also affect the accuracy of the RTree associated with the DisplayList, which would only show up in rare circumstances when the indicated operation was used in the presence of an embedded view.
Fixes https://github.com/flutter/flutter/issues/129862
This reverts commit cd52c0ce82b37bc9d16bea26fbd1753174c49e42.
These changes were exposing an underlying bug in the DisplayListMatrixClipTracker that has now been fixed independently (https://github.com/flutter/engine/pull/43664). There are no changes to the commit being relanded here, it has been tested on the Wondrous app which demonstrated the regression before the NOP culling feature was reverted and it now works fine due to the fix of the underlying cause.
Reverts flutter/engine#42584. (Thanks to @jonahwilliams for bisecting)
With this change, layers are getting clipped incorrectly when rendering
platform views in Wondrous.
The original PR caused some golden test failures down the line, likely due to bad analysis of when the combined BlendMode and color would result in a NOP situation.
This PR adds tests that go through every BlandMode and pair it with a variety of colors and Color/ImageFilters to verify that the operations are only omitted when they actually produce no change in the output. It also checks the validity of the "modifies_transparent_black" property of DisplayLists which can be used in place of the current CanvasSpy/DlOpSpy classes.
The description from the [previous PR](https://github.com/flutter/engine/pull/41463) updated with the new name of the DL property:
---------------------------------
This optimization avoids recording unnecessary render operations that will not affect the output and also eliminates the need for "draw detection" mechanisms like `DlOpSpy` and `CanvasSpy` by remembering if any non-transparent operations were included. The `DlOpSpy` unit tests were updated to check if the results from that object match the new `DisplayList::modifies_transparent_black()` method.
Fixes https://github.com/flutter/flutter/issues/125338
In addition, this change will unblock some other Issues:
- https://github.com/flutter/flutter/issues/125318
- https://github.com/flutter/flutter/issues/125403
Part of an ongoing set of efforts to address https://github.com/flutter/flutter/issues/106448
Move the checkerboard layer unit tests onto the DisplayList version of the paint contexts and fix some bugs in the reusability of the DisplayListBuilder that this migration uncovered.
This optimization avoids recording unnecessary render operations that will not affect the output and also eliminates the need for "draw detection" mechanisms like `DlOpSpy` and `CanvasSpy` by remembering if any non-transparent operations were included. The `DlOpSpy` unit tests were updated to check if the results from that object match the new `DisplayList::affects_transparent_surface()` method.
Fixes https://github.com/flutter/flutter/issues/125338
In addition, this change will unblock some other Issues:
- https://github.com/flutter/flutter/issues/125318
- https://github.com/flutter/flutter/issues/125403
