We currently benchmark path tessellation on Android and iOS, but only
for filled paths. We will now run those same benchmarks also with
stroking in a new set of benchmarks.
This benchmark is a benchmark of the integration test binding which IMO
is not that important given that we continue to rely on flutter_driver
for actual performance metrics.
This auto-formats all *.dart files in the repository outside of the
`engine` subdirectory and enforces that these files stay formatted with
a presubmit check.
**Reviewers:** Please carefully review all the commits except for the
one titled "formatted". The "formatted" commit was auto-generated by
running `dev/tools/format.sh -a -f`. The other commits were hand-crafted
to prepare the repo for the formatting change. I recommend reviewing the
commits one-by-one via the "Commits" tab and avoiding Github's "Files
changed" tab as it will likely slow down your browser because of the
size of this PR.
---------
Co-authored-by: Kate Lovett <katelovett@google.com>
Co-authored-by: LongCatIsLooong <31859944+LongCatIsLooong@users.noreply.github.com>
This test expects to run on 32-bit hardware, but the devicelab no longer
has 32-bit Android hardware. In particular, the test contains hand-coded
32-bit arm assembly for reading the stack pointer from Dart code via
`dart:ffi`.
This test was added in response to a framework change that caused stack
frames to be larger than expected while building widgets, leading apps
to crash with StackOverflow exceptions unexpectedly. Since then (>3
years ago) this test has not prevented any similar issue, so I believe
deleting it rather than fixing it is a better use of resources.
This adds a macrobenchmark representative of a real world application that uses SVG icons. The scenario of rasterizing complex paths that don't change over time does not seem to be covered by any other macrobenchmark and shows a significantly slower impeller performance compared to skia.
It's actually bit problematic to measure this because on A15 the CPU load with impeller is high enough to trigger CPU frequency change. So in order to get consistent reading I had to add a spinning background thread that would keep the CPU at highest frequency.
```objc
[NSThread detachNewThreadWithBlock:^{
while (true) {
pthread_yield_np();
}
}];
```
```bash
flutter drive --profile --local-engine=ios_profile -t test_driver/run_app.dart --driver test_driver/path_tessellation_static_perf_test.dart
```
| average_frame_build_time_millis |Time|
|--|--|
| Impeller | 0.46686524822695047 |
| Skia | 0.4625749999999999 |
| Skia - No RasterCache | 0.47173750000000086|
| average_frame_rasterizer_time_millis | Time |
|--|--|
| Impeller | 6.654328519855595 |
| Skia - Raster Cache | 0.2534123711340209 * |
| Skia - No RasterCache | 0.53424375 |
* Adding the `GeometryPainter` seems to have triggered the complexity threshold for raster cache.
<img alt="screenshot" width="320" src="https://github.com/flutter/flutter/assets/96958/7a2f9384-b512-477b-bffa-058d4d284a41"/>
This reduce the execution time of macrobenchmarks driver tests.
I tried to find the exact size to scroll the screen but I couldn't find a way to do that with driver tests.
Previously, the CPU/GPU measurement using gauge happened after the
driver test is done. Now, the measurement happens within the driver
test. So we need to properly increase the delay and duration to get a
comparable reading as the old gauge measurements.
