Part of an ongoing set of efforts to address https://github.com/flutter/flutter/issues/106448
The layer unittests have been using a MockCanvas class to record the painting of trees of layers and then testing for the expected output.
A while back a similar mechanism was created to compare DisplayList output and to print out a human-friendly version of the differences found, but it was only used in a few tests written at the time it was created and a few since then.
This is the first in a series of PRs that will move all the rest of the unit tests onto the new DL comparison mechanism, starting with the layer types that just do basic drawing. Some of the remaining layers will require creating new hooks in, for instance, the Texture registry, the performance overlay TextBlob generation, etc.
Removes the physical model layer and associated engine code. This was already deprecated and removed in the framework. By removing it in the engine, we can also remove the need for layer tree diff/paint/preroll to have the device pixel ratio. This will simplify some of the multi-view work
Fixes https://github.com/flutter/flutter/issues/125720
Switching the calls to dispatch into an Impeller Dispatcher to use a cull rect to enable pre-culling of the out-of-bounds ops.
This change showed an improvement of around 2x on the rendering performance of the non-intersecting platform view benchmark, but that was measured without the recent changes to the destructive blend modes in Impeller renderer.
Implements partial repaint for Impeller.
Fixes https://github.com/flutter/flutter/issues/124526
The new code that manages the damage regions is more or less a copy paste from the existing Skia implementation. Compared to Skia, there are a few differences:
Normally Impeller wants to use the drawable as the resolve texture for the root MSAA pass. Unfortunately this will unconditonally clear that texture. Thus to do a partial repaint, we have to allocate a separate texture to resolve to and then blit into the drawable.
The blit seems to take about 500ns for a full screen on an iPhone 13. That implies that partial repaint is likely not worth doing if the screen is significantly changed. Thus I've added code in compositor_context.cc that computes the percentage of width or height that is part of the dirty rect. Above a threshold of (abitrarily chosen) 70%, we just render as normal. This should mean there is only a very minor hit from performing the diff on screens that are highly changed.
The other special case, is that sometimes we get damage rects that are empty - that is the drawable is already completely up to date with what we want to render. IN that case I shortcircuit all of the impeller code and just present immediately. I previously tried returning without a present but this resulted in Xcode reporting dropped frames. One caveat here is that if you use the XCode frame debugger and attempt to capture a frame where we early present, then it will claim it couldn't capture any command buffers (because we didn't create any).
To facilitate all of this, I added some additonal plumbing so that the impeller surface can get the clip rect from the submit info. Additionally, rather than using a clip rect impeller will translate and then shrink the root surface texture. This reduces memory usage compared to just clippling.
These keep turning up, so I did a little bash-fu to find them all.
```bash
grep -lL "#pragma once" $(grep -lL "#ifndef .*_H_" $(find . | grep "\.h$")) | cut -c 3-
```
With this PR we no longer need to hold DisplayLists in GPUObject wrappers and they can be disposed instantly instead of queueing on the Unref thread.
This will definitely be a win for Impeller as none of the objects used in a frame now require queueing, but the performance impact on apps running on top of skia is less clear if they depend on a lot of images inside their DisplayLists that still need to be queued to be freed. After getting further in the work, it looks like only decoded images need to use the protected DlImage wrappers and most of those should survive many frames before they are disposed. That should hopefully leave very few unrefs happening per frame.
~There are 3 unit tests in `shell_unittests.cc` and `embedder_metal_unittests.mm` that are now GSKIP'd as they now invoke code that needs a fully initialized UIDartState in order to protect their images. I will look into fixing the tests and/or making the code they invoke provide protection without relying on UIDartState.~ (This looks to be fixed in the latest commit by simply not creating DlImageGPUs all over the source base and simply catching only those that end up in UI data structures. There is actually existing code in one of the modules that feeds ui.Image with an answer to wrap the image in a DlImageGPU if it has a skia image anyway, so most of these additional uses of DlImageGPU that were having trouble getting the Skia unref queue just didn't need it anyway.)
Previously this was comparing a float overlay rectangle size with an
integer cache bitmap size. This could result in mismatches if the
overlay width is not an integer.
* Implemented wide gamut images for iOS
Moved the surface to an extended range color format.
* wrong gamma but default pixel format set to bgra10_xr
* BGR10_XR add
* format
* updated todos
* updated todo with information about pixel formats
* switched logic for determining if we have a wide gamut image
* cleaned up gamut math to match style and linked source
* made the color attachment pixel format match the surface
* updated vulkan format switch
* removed comment
* added enable disable switch
* moved default to bgr10 for now since there is a bug where someone is still reading this, msaa?
* fixed the decoder settings to make sure we don't lose wide gamut colors
* fixed stored srgb gamut variable
* fixed false lint
* updated test
* added ability to grab the surface data for tests
* made the screenshot utility return the format
* added width and height to the platform channel payload
* fixed a couple of broken targets
* moved back the default pixel buffer format
* cleanup and add docstrings
* made the surfacedata feature only available in debug builds
* added decoding unit test
* fixed objc tests
* turned off by default
* bdero feedback1
* bdero2
* bdero3
* fixed merge issue
* removed using std::shared_ptr
* add support to run dl_canvas_unittests on GPU surfaces (off by default)
* empty commit to trigger build
* conditionally include platform-specific surface provider headers
* ignore DL test files during license checks
* typo
* add dependencies to rendertests to hopefully build on Windows
* remove benchmarking deps from dl_rendertests
* more changes to get Windows rendertests to link
* add --enable-gl synonym and prevent non-SW surface provider dest on Windows
* fix gn formatting
* review feedback
* collect DL indices in RTree for clip culling
* fix bounds in unit test and minor opt in Dispatch
* normalize inline matrix objects and minor fixes to unit test
* remove over-eager DCHECK and improve R-Tree comments
* formatting
* include vector for Windows
* method rename and distribute child nodes more evenly
* add R-Tree specific unit tests and debug checks
* add comments about geometry to R-Tree unit tests and adjust spacing
* licenses
* licenses attempt 2
* fix potential overflow with uint32_t
* aggressively const DisplayList fields and methods
* add implementation comments per review feedback
* Revert "Revert "Reland layer state stack" (#37178)"
This reverts commit 21cc000359d8a4da097e8849e83dcbe8ac941e01.
* fix double-transform rendering issues in #114359
* adjust recently added unit test to state_stack APIs
* introduce LSS delegates to simplify some code and reduce overhead
* Fix ShellTest.OnServiceProtocolEstimateRasterCacheMemoryWorks
* add unit test for tracker.setTransform(4x4) and fix bug
* fix culling issue in LayerTree::Flatten
