Apply optimizations in Avalonia Android

[jsuarezruiz]

What does the pull request do?

Applies performance optimizations to the Avalonia Android backend, reducing GC pressure and lock contention on the UI thread during rendering, input dispatch, and accessibility operations.

What is the current behavior?

Some Android hot paths use heavier synchronization than needed:

• DispatchDraw creates a new Paint object every frame (24 KB/1000 frames)
• AndroidDispatcherImpl.Signal() uses lock for a simple boolean flag
• TopLevelImpl.TextInput uses DateTime.Now.Ticks (wrong unit)
• OnPerformActionForVirtualView uses LINQ .Select().Aggregate() (72 KB/1000 calls)
• AndroidKeyboardEventsHelper calls char.ToString() on every key event (24 KB/1000 events)

What is the updated/expected behavior with this PR?

Applied changes to optimize while preserve behavior with measurable improvements.
Measured on Android, 10,000 iterations after 100 warmup:

Optimization	Before	After	Speedup	Alloc Saved
DispatchDraw: cache Paint object	13,741 ns/call	6.7 ns/call	2,050x	1,760,664 B → 216 B
Dispatcher: lock → Interlocked	24.9 ns/call	7.8 ns/call	3.2x	0
Accessibility: LINQ → foreach	68 ns/call	10 ns/call	6.5x	720,688 B → 688 B
KeySymbol: char.ConvertFromUtf32 → cached	7 ns/call	1 ns/call	5.9x	240,040 B → 40 B

TextInput timestamp change (DateTime.Now.Ticksto SystemClock.UptimeMillis()) is a fix, not a performance optimization. The original code used the wrong unit (100ns ticks instead of milliseconds) and a non-monotonic clock.

How was the solution implemented (if it's not obvious)?

1. TopLevelImpl.DispatchDraw: Cache the Paint instance as a field. Properties never change after init and DispatchDraw runs on the UI thread only.

2. AndroidDispatcherImpl.Signal/OnSignaled/OnIdle: Replace lock + bool _signaled with int _signaled using Interlocked.CompareExchange / Interlocked.Exchange / Volatile.Read. The flag is a simple set/reset gate with no compound critical section, making lock-free operations a good eplacement.

3. TopLevelImpl.TextInput: Replace DateTime.Now.Ticks with SystemClock.UptimeMillis(). This matches the same epoch and unit used by KeyEvent.EventTime and MotionEvent.EventTime throughout the Android input pipeline, making text input timestamps directly comparable to pointer and keyboard timestamps.

4. AvaloniaAccessHelper.OnPerformActionForVirtualView: Replace .Select().Aggregate() with a foreach loop and an explicitnull check. Eliminates iterator and delegate allocations.

5. AndroidKeyboardEventsHelper: Pre-cache char.ToString() for ASCII range 0-127 in a static array. Non-ASCII falls through to char.ConvertFromUtf32() unchanged.

Checklist

• Added unit tests (if possible)?
• Added XML documentation to any related classes?
• Consider submitting a PR to https://github.com/AvaloniaUI/avalonia-docs with user documentation

Benchmarks added in tests/Avalonia.Benchmarks.Android/ to validate each optimization.

Breaking changes

None.

Obsoletions / Deprecations

None.

[avaloniaui-bot]

You can test this PR using the following package version. 12.0.999-cibuild0063846-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[maxkatz6]

I highly doubt going to java-interop monstroucity for SystemClock.UptimeMillis is better here.
Shouldn't we use Stopwatch instead, as .NET equvivalent?

[MrJul]

Also, we should use the same source everywhere so that timestamps are comparable.
(For a given platform. They don't have to be cross-platform.)

[MrJul]

Note: the matching benchmark does use Stopwatch.

[jsuarezruiz]

Good points.

C# → JNI → native clock_gettime

Android numbers (100,000 iterations on emulator):

Approach	Time	Per-call
DateTime.Now.Ticks	60 ms	600 ns
SystemClock.UptimeMillis()	57 ms	570 ns
Environment.TickCount64	2 ms	20 ns
Stopwatch.GetTimestamp()	4 ms	40 ns

Key findings:

• SystemClock.UptimeMillis() is essentially almost same speed as DateTime.Now.Ticks on Android. The JNI overhead is real.
• Environment.TickCount64 is 29x faster than both.
• Stopwatch.GetTimestamp() is 15x faster than both.

Used SystemClock.UptimeMillis() to keep timestamps comparable with KeyEvent.EventTime and MotionEvent.EventTime, which both use SystemClock.uptimeMillis() as their clock source. Using a different .NET clock (Environment.TickCount64 or Stopwatch) would produce timestamps on a different epoch that doesn't pause during deep sleep, causing drift.

The benchmark uses Stopwatch.GetTimestamp() as a desktop proxy since SystemClock is Android-only.

[kerams]

On CoreCLR UtcNow is faster than Now because time zone adjustments are skipped. That might end up being faster than UptimeMillis.

[kekekeks]

Are we discussing performance of something that takes nanoseconds to execute and is called like 3 times per second at most?

[kekekeks]

The event should use the same timestamp source as the rest of input events, please check what matches MotionEvent.eventTime and use it there.

[jsuarezruiz]

Android's KeyEvent and MotionEvent timestamps use the SystemClock.uptimeMillis() time base, and the existing Android raw input paths already use that same clock base via EventTime/UptimeMillis. I changed TopLevelImpl.TextInput to use SystemClock.UptimeMillis() as well so it stays consistent with the rest of the Android input events.

[avaloniaui-bot]

You can test this PR using the following package version. 12.0.999-cibuild0063864-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[jsuarezruiz]

Android Benchmark Results (.NET 10.0.0, ARM64 emulator)

Ran the benchmarks on an actual Android device to get real numbers including JNI overhead. Each benchmark runs 10,000 iterations after 100 warmup iterations.

DispatchDraw: Paint allocation

Approach	Time	Per-call	Alloc	Speedup
NewPaintPerFrame	137.41 ms	13,741 ns	1,760,664 B	baseline
CachedPaint	0.07 ms	6.7 ns	216 B	2,050x

Dispatcher: lock vs Interlocked

Approach	Time	Per-call	Speedup
Lock	0.25 ms	24.9 ns	baseline
Interlocked	0.08 ms	7.8 ns	3.2x

Timestamp sources

Source	Time	Per-call
DateTime.Now.Ticks (original)	1.10 ms	109.6 ns
DateTime.UtcNow.Ticks	0.22 ms	22.0 ns
Stopwatch.GetTimestamp	0.20 ms	19.5 ns
Environment.TickCount64	0.08 ms	7.6 ns
SystemClock.UptimeMillis (JNI)	2.34 ms	234.5 ns

SystemClock.UptimeMillis() is the slowest due to JNI overhead (2.1x slower than the original DateTime.Now.Ticks). However, it remains the correct choice here because:

• All other Android input events (KeyEvent.EventTime, MotionEvent.EventTime) use SystemClock.uptimeMillis() as their clock source
• Timestamps are compared across event types for double-click detection (TouchDevice) and gesture timeouts (ScrollGestureRecognizer), so they must share the same epoch and unit
• Environment.TickCount64 / Stopwatch use a different epoch and don't pause during deep sleep, causing drift
• The 234ns cost is irrelevant at human typing speed (~100ms between keystrokes = 0.0002% overhead)

The original code (DateTime.Now.Ticks) was broken: it passed 100ns ticks where milliseconds were expected, and used a non-monotonic clock.

Accessibility: LINQ vs foreach

Approach	Time	Alloc	Speedup
LINQ Select+Aggregate	0.68 ms	720,688 B	baseline
foreach loop	0.10 ms	688 B	6.5x

Keyboard: char conversion vs cached

Approach	Time	Alloc	Speedup
char.ConvertFromUtf32	0.07 ms	240,040 B	baseline
Cached ASCII lookup	0.01 ms	40 B	5.9x

Regarding the desktop benchmark file (AndroidRenderingBenchmarks.cs): agree with @maxkatz6 that it should be removed since it benchmarks generic .NET patterns, not actual Android code. An Android benchmark project (tests/Avalonia.Benchmarks.Android) have been added.

[maxkatz6]

All other Android input events (KeyEvent.EventTime, MotionEvent.EventTime) use SystemClock.uptimeMillis() as their clock source

So we should remove it from other input events too.
And iOS uses environment ticks BCL API already, instead of current platform

[avaloniaui-bot]

You can test this PR using the following package version. 12.0.999-cibuild0063999-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[avaloniaui-bot]

You can test this PR using the following package version. 12.0.999-cibuild0064263-alpha. (feed url: https://nuget-feed-all.avaloniaui.net/v3/index.json) [PRBUILDID]

[MrJul]

LGTM!