.NET 7 Performance Improvements in .NET MAUI

Jonathan Peppers

When .NET MAUI reached GA, we had goals of improving upon Xamarin.Forms in startup time and application size. See last release’s Performance Improvements in .NET MAUI, to learn about the performance benefits of migrating from Xamarin.Android, Xamarin.iOS, or Xamarin.Forms to .NET 6+ and .NET MAUI.

We continued with these themes in .NET 7, building upon the excellent .NET performance work described by Stephen Toub. General improvements in the .NET runtime or base class libraries (BCL) lay the foundation for the performance we are able to achieve in .NET MAUI.

For Android, our focus remains on startup performance, since application size is in a good place. For iOS, we have the opposite goal: we continued to focus on application size, since startup performance on iOS is in good shape.

From customer feedback we found the need to go beyond just startup performance, also focusing a bit on general UI performance, layout, scrolling, etc. We also improved startup time and general performance for desktop platforms. We hope to continue making improvements in these areas in future .NET releases.

For an overall comparison of startup time between .NET 6 and .NET 7, apps in Release mode on a Pixel 5:

Application Version Startup Time(ms)
dotnet new maui .NET 6 568.1
dotnet new maui .NET 7 545.4
.NET Podcast .NET 6 814.2
.NET Podcast .NET 7 759.7

In .NET 7, .NET MAUI applications should see an improvement to startup time as well as smoothing scrolling, navigation, and general UI performance.

Likewise, iOS applications have continued to get smaller in .NET 7:

Application Version .ipa Size (MB)
dotnet new maui .NET 6 12.64
dotnet new maui .NET 7 12.48
.NET Podcast .NET 6 25.08
.NET Podcast .NET 7 23.91

Testing a CollectionView sample on a modest Android device (Pixel 4a). We can see a clear improvement with Android’s visual GPU profiler, a screenshot taken during some “brisk” scrolling:

Android Visual GPU Profiler .NET 6 vs 7

This sample is a CollectionView with 10,000 rows of two data-bound labels.

Simply updating to .NET 7 should result in smaller/faster .NET MAUI applications. Details below!

Table Of Contents

Scrolling and Layout Performance Improvements

Startup Performance Improvements

App Size Improvements

Tooling and Documentation

Scrolling and Layout Performance Improvements

LOLs per Second

We have seen benchmarks of different UI frameworks that follow a pattern, such as:

  1. Put text on the screen with a random rotation and color

  2. Report how many times a second you can do this

The existing samples were OK, but I found starting from scratch I could avoid minor performance issues and focus on the raw numbers .NET MAUI could achieve. This also gave me the opportunity to use the word “LOL” in the text onscreen, aptly timing the number of “LOLs per second” we can achieve in .NET MAUI. LOL?

Additionally, we timed different types of apps:

  • A Xamarin.Forms application (thanks @roubachof for the contribution!)
  • A .NET MAUI application
  • A .NET 6 Android application (using our C# bindings for Android APIs)
  • An Android application written in Java

The above apps would all use the same underlying Android.Widget.TextView. Each app should progressively be able to achieve better results: Xamarin.Forms to MAUI to C# TextView to Java TextView (no C# to Java interop):

Chart of LOLs per second

Overall, this concept (although quite fun!) isn’t an exact science. My version of the sample required some duration of Thread.Sleep(), which will certainly vary depending on the device you are running on. The sample needs to sleep long enough for the UI thread to be able to keep up with a background thread. Just from visually watching the app, I tried to pick a Thread.Sleep() time where we got a good number of LOLs per second and the UI appeared to be updating quickly.

However, the sample in itself is still useful:

  1. We have a fun, visual comparison between .NET MAUI in .NET 6 vs .NET 7

  2. It is a literal gold mine for finding performance issues!

Simply reviewing dotnet-trace output of these apps, led to changes to improve the performance in .NET MAUI of:

  • Each View added to the screen

  • Layout performance

  • Scrolling performance

  • Navigation between pages

In total, the LOLs per second on a Pixel 5 went from ~327 per second to ~493 per second moving from .NET 6 to .NET 7:

LOLs per second screenshot

For full details and source code for this sample/benchmark, see the jonathanpeppers/lols repository on GitHub.

Avoid Repeated View.Context Calls

When reviewing dotnet-trace output of the LOLs/second sample, we noticed:

7.60s (14%) mono.android!Android.Views.View.get_Context()

There is an interesting Java to C# interop implication in this call. Since the beginning of Xamarin.Android we’ve had the behavior:

  • C# calls a Java method

  • A Java object is returned to C#

  • Using the Handle of the Java object, find out if we have a C# object alive with the same Handle.

  • If not, create a new C# instance to wrap this object for use in managed code.

So you could see how repeated calls to .Context would be a performance concern, after realizing the work happening behind the scenes.

We found layout-related code in .NET MAUI, such as:

var deviceIndependentWidth = widthMeasureSpec.ToDouble(Context);
var deviceIndependentHeight = heightMeasureSpec.ToDouble(Context);
//...
var platformWidth = Context.ToPixels(width);
var platformHeight = Context.ToPixels(height);

Where the four calls here could be replaced by a local variable, or even a member field for the lifetime of this object. This simple change directly translated to more Label‘s per second as well as faster startup and scrolling in .NET MAUI.

See dotnet/maui#8001 for details about this improvement.

Avoid View.Context Calls in CollectionView

In .NET 6, we had some reports of poor CollectionView performance while scrolling on older Android devices.

Reviewing dotnet-trace output, I did find some issues similar to dotnet/maui#8001:

317.42ms (1.1%) mono.android!Android.Views.View.get_Context()

1% of the time was spent in repeated calls to View.Context inside the ItemContentView class, such as:

if (pixelWidth == 0)
{
    pixelWidth = (int)Context.ToPixels(measure.Width);
}
if (pixelHeight == 0)
{
    pixelHeight = (int)Context.ToPixels(measure.Height);
}
//...
var x = (int)Context.ToPixels(mauiControlsView.X);
var y = (int)Context.ToPixels(mauiControlsView.Y);
var width = Math.Max(0, (int)Context.ToPixels(mauiControlsView.Width));
var height = Math.Max(0, (int)Context.ToPixels(mauiControlsView.Height));

Making a new overload for ContextExtensions.FromPixel(), we were able to remove all of these View.Context calls.

After these changes, we see a huge difference in the time spent in this method:

1.30ms (0.01%) mono.android!Android.Views.View.get_Context()

We also tested these changes with the “janky frames” profiler in Android Studio, which gives information like this on Android 12+ devices (such as a Pixel 4a). We could see several dropped frames when scrolling a CollectionView:

Android Studio Janky Frames Before

On the same Pixel 4a afterward, we only saw a single dropped frame:

Android Studio Janky Frames After

We could also test these changes with Android’s visual GPU profiler, seeing a visual difference in before:

Android Visual GPU Profiler Before

Versus afterward:

Android Visual GPU Profiler After

The larger bars represents time spent in code paths that attribute to a poor framerate. These changes should improve the performance of scrolling for any .NET MAUI CollectionView on Android.

See dotnet/maui#8243 for details about this improvement.

Reduce JNI Calls During Layout

While profiling customer sample similar to the “LOLs per second” app, we noticed a decent chunk of time spent measuring Android View‘s:

932.51ms (7.4%)  mono.android!Android.Views.View.Measure(int,int)
115.53ms (0.91%) mono.android!Android.Views.View.get_MeasuredWidth()
 96.97ms (0.77%) mono.android!Android.Views.View.get_MeasuredHeight()

This led us to write a Java method that calls all three Java APIs and return MeasuredWidth and MeasuredHeight. After a little research, it seemed the best approach here was to “pack” two integers into a long. If we tried to return some Java object instead, then we’d have the same number of JNI calls to get the integers out.

The Java side implementation arrives at:

public static long measureAndGetWidthAndHeight(View view, int widthMeasureSpec, int heightMeasureSpec) {
    view.measure(widthMeasureSpec, heightMeasureSpec);
    int width = view.getMeasuredWidth();
    int height = view.getMeasuredHeight();
    return ((long)width << 32) | (height & 0xffffffffL);
}

Which is unpacked in C# via:

var packed = PlatformInterop.MeasureAndGetWidthAndHeight(platformView, widthSpec, heightSpec);
var measuredWidth = (int)(packed >> 32);
var measuredHeight = (int)(packed & 0xffffffffL);

So instead of three transitions from C# to Java, we now have one. It appears the performance of the additional math is negligible compared to reducing the amount of interop overhead in this example.

Measuring again after the change, we were able to see an obvious savings in this method:

--783.92ms (6.2%) microsoft.maui!Microsoft.Maui.ViewHandlerExtensions.GetDesiredSizeFromHandler
++528.96ms (4.5%) microsoft.maui!Microsoft.Maui.ViewHandlerExtensions.GetDesiredSizeFromHandler

Which also translated to more Label‘s per second in the sample app.

See dotnet/maui#8034 for details about this improvement.

Cache Values for RTL and Dark Mode

While profiling customer sample similar to the “LOLs per second” app, we noticed time spent calculating if the OS is in Dark Mode or a Right-to-Left language:

1.06s (6.3%) Microsoft.Maui.Essentials!Microsoft.Maui.ApplicationModel.AppInfoImplementation.get_RequestedLayoutDirection()
4.46ms (0.03%) Microsoft.Maui.Essentials!Microsoft.Maui.ApplicationModel.AppInfoImplementation.get_RequestedTheme()

This appears to happen for every .NET MAUI View on Android, so this impacts startup time, scrolling, etc.

Reviewing the AppInfoImplementation class on Android, we could simply use Lazy<T> for every value that queries Application.Context. These values cannot change after the app launches, so they don’t need to be computed on every call.

This should also improve subsequent calls to various Maui.Essentials APIs on Android.

See dotnet/maui#7996 for details about this improvement.

Avoid Creating IView[] During Layout

While profiling customer sample similar to the “LOLs per second” app, time was spent ordering views by their ZIndex such as:

3.42s (16%)  microsoft.maui!Microsoft.Maui.Handlers.LayoutHandler.Add(Microsoft.Maui.IView)
1.52s (7.3%) microsoft.maui!Microsoft.Maui.Handlers.LayoutExtensions.GetLayoutHandlerIndex()
1.50s (7.2%) microsoft.maui!Microsoft.Maui.Handlers.LayoutExtensions.OrderByZIndex(Microsoft.Maui.ILayout)

Reviewing the code in OrderByZIndex(), it did the following:

  • Make a record ViewAndIndex(IView View, int Index) for each child
  • Make a ViewAndIndex[] the size of the number of children
  • Call Array.Sort()
  • Make a IView[] the size of the number of children
  • Iterate twice over the arrays in the process

Then the GetLayoutHandlerIndex() method, did all the same work to create an IView[], get the index, then throws the array away.

To improve this process, we made the following changes:

  • Removed much of the above code and used a System.Linq OrderBy() with “fast paths” for 0 or 1 children.
  • Made an internal EnumerateByZIndex() method for use by foreach loops. This avoids creating arrays in those cases.

After these changes, the time dropped substantially in dotnet-trace output:

  1.84s (13%)    microsoft.maui!Microsoft.Maui.Handlers.LayoutHandler.Add(Microsoft.Maui.IView)
352.24ms (2.50%) microsoft.maui!Microsoft.Maui.Handlers.LayoutExtensions.GetLayoutHandlerIndex()
181.27ms (1.30%) microsoft.maui!Microsoft.Maui.Handlers.LayoutExtensions.<>c.<EnumerateByZIndex>b__0_0(Microsoft.Maui.IView)
  2.78ms (0.02%) microsoft.maui!Microsoft.Maui.Handlers.LayoutExtensions.EnumerateByZIndex(Microsoft.Maui.ILayout)

We also saw a significant increase in the number of Label‘s a second in the customer sample. System.Linq was used in the improvement, which goes a little against traditional guidance to avoid it. In this case, the code was implementing its own version of OrderBy(), where we could just use OrderBy() instead. “Rolling your own” System.Linq is generally not a great idea, unless you have benchmarks showing your implementation is better.

This change should improve layout performance of all .NET MAUI applications on any platform.

See dotnet/maui#8250 for details about this improvement.

Defer RTL Layout Calculations to Platform

When profiling our LOLs per second sample app, a percentage of time is spent calculating if each view is setup to support “Right to Left” languages or not:

290.86ms (1.4%)    microsoft.maui!Microsoft.Maui.Layouts.LayoutExtensions.ShouldArrangeLeftToRight(Microsoft.Maui.IView)

In this sample app, we have a nested layout, 5 views deep, with ~100 Label‘s on the screen. Each Label we ended up querying the same views for flow direction 500 times.

To solve this issue and many others, the cross-platform implementation of RTL was removed from .NET MAUI in favor of native APIs on each platform. This should improve performance and correctness in this area.

LayoutExtensions.ShouldArrangeLeftToRight() is completely removed from .NET MAUI in .NET 7. This should improve layout performance for .NET MAUI on all platforms.

See dotnet/maui#9558 for details about this improvement.

Further Notes on CollectionView

As seen in issue dotnet/maui#6317, a CollectionView with 10,000 rows contributed to poor startup time and slow scrolling performance.

In this case, the sample app on a Pixel 5 would start in:

  • .NET 6: 16s412ms
  • .NET 7: 14s642ms

The root cause here appears to be the specific .xaml layout:

<ContentPage ...>
    <ScrollView>
        <VerticalStackLayout>
            <CollectionView 
                ItemsSource="{Binding Bots}"
                HorizontalOptions="FillAndExpand"
                VerticalOptions="FillAndExpand">

The CollectionView is placed inside a VerticalStackLayout that will happily expand to an infinite height. Then subsequently placed inside a ScrollView (which isn’t necessary because CollectionView supports scrolling). This effectively causes the CollectionView to “realize” all 10,000 rows, and voids any virtualization of the control.

To solve this issue, we could simply remove the outer views:

<ScrollView>
    <VerticalStackLayout>

Resulting with an identical look and feel on screen, and a drastically improved ~822ms (94% shorter!) startup time on a Pixel 5.

If you’re getting similar behavior in a CollectionView, verify that the rows are actually virtualized and not wrapped in unnecessary views. In future .NET releases, we are working on ways to prevent someone falling into this trap accidentally.

Startup Performance Improvements

Android NDK Compiler Flags

When we first got .NET MAUI running on early previews of .NET 7, we noticed a regression to both startup & app size:

  • A ~1.5MB increase in libmonosgen-2.0.so, the Android native library for the Mono runtime.
  • A ~250ms increase in startup time

In NDK23b, Google decided to no longer pass the -O2 compiler optimization flag for arm64 as part of the Android toolchain. Instead the flag is decided by upstream CMake behavior, see android/ndk#1536. Thus we needed to specify the optimization flag going forward for the Mono runtime.

We tested three flags to compile libmonosgen-2.0.so in a dotnet new maui application:

  • -O3: 893.7ms
  • -O2: 600.2ms
  • -Oz: 649.1ms

Interestingly, -03 (the new default!) produced a larger and slower libmonosgen-2.0.so. We settled on -O2, which gave us the same performance and size we were getting in .NET 6.

See dotnet/runtime#68354 for details about this improvement.

DateTimeOffset.Now

Reviewing a customer’s .NET 6 Android application, we found a significant amount of time spent in the very first DateTimeOffset.Now call. In this case it was accidental (UtcNow could simply be used instead), but nevertheless is a common API that .NET developers are going to use.

Reviewing dotnet-trace output, the first call to DateTimeOffset.Now took around 277ms on a Pixel 5 device:

Speedscope of DateTimeOffset with JIT

Some of the time here was spent in the JIT, so our first idea was to record a custom AOT profile (using our experimental Mono.AotProfiler.Android package), and see if this was a good candidate for our standard profile. This got a decent improvement (~161ms), but the number still seemed like it could be improved:

Speedscope of DateTimeOffset with AOT

It turns out, the bulk of the time here was spent loading timezone data — and then calculating the current time from the result.

Instead, we could:

  1. Call an Android Java API for the current time offset. Our startup code in the Android workload is Java, so it can easily retrieve the value for use by the BCL.

  2. Load timezone data on a background thread. This results in the original BCL behavior when complete, while using the offset from Java in the meantime.

This greatly improved DateTimeOffset.Now to a mere ~17.67ms — even without adding the code path to an AOT profile.

See dotnet/runtime#74459 and xamarin-android#7331 for details about this improvement.

Avoid ColorStateList(int[][],int[])

dotnet-trace output of a .NET 6 dotnet new maui application shows time spent creating Android ColorStateList objects:

16.63ms microsoft.maui!Microsoft.Maui.Platform.ColorStateListExtensions.CreateDefault(int)
16.63ms mono.android!Android.Content.Res.ColorStateList..ctor(int[][],int[])
11.38ms mono.android!Android.Runtime.JNIEnv.NewArray

Reviewing the C# binding for this type, reveals part of the performance impact of this constructor:

public unsafe ColorStateList (int[][]? states, int[]? colors)
    : base (IntPtr.Zero, JniHandleOwnership.DoNotTransfer)
{
    //...
    IntPtr native_states = JNIEnv.NewArray (states);

There is a general performance problem in passing C# arrays (especially multidimensional ones) into Java. We have to create a Java array and copy each array element over to the Java array. We slightly improved upon this scenario in xamarin-android#6870, but the array duplication remains a problem.

To solve this, we can design some internal Java APIs for .NET MAUI like:

@NonNull
public static ColorStateList getDefaultColorStateList(int color)
{
    return new ColorStateList(ColorStates.DEFAULT, new int[] { color });
}

@NonNull
public static ColorStateList getEditTextColorStateList(int enabled, int disabled)
{
    return new ColorStateList(ColorStates.getEditTextState(), new int[] { enabled, disabled });
}

And move any C# const int values to be completely in Java:

private static class ColorStates
{
    static final int[] EMPTY = new int[] { };
    static final int[][] DEFAULT = new int[][] { EMPTY };

    private static int[][] editTextState;

    static int[][] getEditTextState()
    {
        if (editTextState == null) {
            editTextState = new int[][] {
                new int[] {  android.R.attr.state_enabled },
                new int[] { -android.R.attr.state_enabled },
            };
        }
        return editTextState;
    }
}

Usage in C# then could look like:

ColorStateList defaultColors = PlatformInterop.GetDefaultColorStateList(color);
ColorStateList editTextColors = PlatformInterop.GetEditTextColorStateList(enabled, disabled);

This removes the need for marshaling various int[] values to Java, and we only pass in a single integer parameter for each state the Android control supports.

Measuring startup after these changes, we instead get:

2.44ms microsoft.maui!Microsoft.Maui.Platform.ColorStateListExtensions.CreateDefault(int)

Which appears to save ~14ms of startup time on a Pixel 5 in the dotnet new maui template. Other apps that use Entry, CheckBox, Switch, etc. will also see an improvement to startup time.

See dotnet/maui#5654 for details about this improvement.

Improvements to .NET MAUI’s AOT Profile

Startup tracing or Profiled AOT is a feature of Xamarin.Android we brought forward to .NET 6+. This is a mechanism for AOT’ing the startup path of applications, which improves launch times significantly with only a modest app size increase.

In .NET 6, we had already recorded a pretty decent AOT profile for .NET MAUI. However, we added a couple more scenarios to the recorded app to improve things further.

First, we added flyout content to the profile, such as:

<Shell ...>
    <FlyoutItem Title="Flyout Item">
        <ShellContent
            Title="Page 1"
            ContentTemplate="{DataTemplate local:MainPage}" />
        <ShellContent
            Title="Page 2"
            ContentTemplate="{DataTemplate local:MainPage}" />
    </FlyoutItem>
</Shell>

This simple change improved the startup of the .NET Podcast sample app by around 25ms.

Secondly, we added a scenario of a non-Shell .NET MAUI app using FlyoutPage. Shell is the default navigation pattern in .NET MAUI, but developers familiar with Xamarin.Forms may want to use APIs such as FlyoutPage, NavigationPage, and TabPage.

So for example, AppFlyoutPage.xaml:

<FlyoutPage ...>
    <FlyoutPage.Detail>
        <local:Tabs Title="Detail"></local:Tabs>
    </FlyoutPage.Detail>
    <FlyoutPage.Flyout>
        <ContentPage Title="Flyout">
            <VerticalStackLayout>
                <Label Text="Flyout Item 1"></Label>
                <Label Text="Flyout Item 2"></Label>
                <Label Text="Flyout Item 3"></Label>
                <Label Text="Flyout Item 4"></Label>
            </VerticalStackLayout>
        </ContentPage>
    </FlyoutPage.Flyout>
</FlyoutPage>

Then we can simply replace MainPage during the AOT profile recording:

App.Current.MainPage = new AppFlyoutPage();

This improved startup of a template project using FlyoutPage by about 12ms. To go even further with your .NET MAUI application, it is possible to record a completely custom AOT profile for your app. See our experimental Mono.AotProfiler.Android package for details.

See dotnet/maui#8939 and dotnet/maui#8941 for details about these improvements.

Better String Comparisons in Java Interop

dotnet-trace output of a .NET MAUI application revealed:

21.28ms (0.45%) java.interop!Java.Interop.JniRuntime.JniTypeManager.AssertSimpleReference(string,string)

Note that changes to Android NDK Compiler Flags may have also contributed to the above timing.

Reviewing the method, it is mostly assertions around the string syntax of jniSimpleReference:

internal static void AssertSimpleReference (string jniSimpleReference, string argumentName = "jniSimpleReference")
{
    if (jniSimpleReference == null)
        throw new ArgumentNullException (argumentName);
    if (jniSimpleReference != null && jniSimpleReference.IndexOf (".", StringComparison.Ordinal) >= 0)
        throw new ArgumentException ("JNI type names do not contain '.', they use '/'. Are you sure you're using a JNI type name?", argumentName);
    if (jniSimpleReference != null && jniSimpleReference.StartsWith ("[", StringComparison.Ordinal))
        throw new ArgumentException ("Arrays cannot be present in simplified type references.", argumentName);
    if (jniSimpleReference != null && jniSimpleReference.StartsWith ("L", StringComparison.Ordinal) && jniSimpleReference.EndsWith (";", StringComparison.Ordinal))
        throw new ArgumentException ("JNI type references are not supported.", argumentName);
}

These assertions are needed, as there is now a way to “remap” underlying Java type names in .NET 7 as implemented in xamarin/java.interop#936. This was also compounded by how many times this method is called in a typical .NET MAUI application on Android.

However, we can simply improve the method by using the char overload of string.IndexOf() and the indexer instead of string.StartsWith():

internal static void AssertSimpleReference (string jniSimpleReference, string argumentName = "jniSimpleReference")
{
    if (string.IsNullOrEmpty (jniSimpleReference))
        throw new ArgumentNullException (argumentName);
    if (jniSimpleReference.IndexOf ('.') >= 0)
        throw new ArgumentException ("JNI type names do not contain '.', they use '/'. Are you sure you're using a JNI type name?", argumentName);
    switch (jniSimpleReference [0]) {
        case '[':
            throw new ArgumentException ("Arrays cannot be present in simplified type references.", argumentName);
        case 'L':
            if (jniSimpleReference [jniSimpleReference.Length - 1] == ';')
                throw new ArgumentException ("JNI type references are not supported.", argumentName);
            break;
        default:
            break;
    }
}

This resulted in a much smaller time when reviewing these code changes in dotnet-trace:

1.21ms java.interop!Java.Interop.JniRuntime.JniTypeManager.AssertSimpleReference(string,string)

We found other code in Java.Interop.JniTypeSignature‘s constructor that validates the string syntax of Java type names, so we made similar changes in both places.

See xamarin/java.interop#1001 and xamarin/java.interop#1002 for details about these improvements.

XAML Compilation Improvements

XAML Compilation (or XamlC) in .NET MAUI is on by default, transforming your .xaml files directly to IL at build time.

To illustrate this, take a couple <Image/> elements:

<Image Source="https://picsum.photos/200/300" />
<Image Source="foo.png" />

In a Debug build, the .xaml is validated to give good error messages about mistakes, but not written to disk. In a Release build, however, this .xaml is emitted directly into your .NET assembly as IL.

In .NET 6, the above would .xaml would compile to the equivalent of:

var image1 = new Image();
image1.Source = new ImageSourceConverter().ConvertFromInvariantString("foo.png");
var image2 = new Image();
image2.Source = new ImageSourceConverter().ConvertFromInvariantString("https://picsum.photos/200/300");

.NET MAUI’s type converters take string values at runtime and convert them into the appropriate type. These are not as performant as if you created plain C# objects directly. Even the dotnet new maui template with a single <Image/> shows an impact of the above code:

1.17ms (<0.01%) microsoft.maui.controls!Microsoft.Maui.Controls.ImageSourceConverter.ConvertFrom()

XamlC has a concept of compiled type converters, when implemented for ImageSource generates better code:

var image1 = new Image();
image1.Source = ImageSource.FromUri(new Uri("foo.png", UriKind.Absolute));
var image2 = new Image();
image2.Source = ImageSource.FromFile("https://picsum.photos/200/300");

This treatment was applied for every type converter currently used at runtime in the dotnet new maui template:

This results in better/faster generated IL from .xaml files on all platforms .NET MAUI supports.

ReadyToRun by Default on Windows

When the .NET Fundamentals team implemented .NET MAUI startup tracking for Windows, we found a huge “easy win” to improve startup. ReadyToRun (or R2R) is not enabled by default in .NET 6 MAUI applications on Windows. R2R is form of ahead-of-time (AOT) compilation that can be used on platforms running the CoreCLR runtime (such as Windows).

Simply adding -p:PublishReadyToRun=true resulted in a huge improvement in our graphs:

Windows .NET MAUI Startup with ReadyToRun

Note that this graph is running the .NET MAUI application on an Azure DevOps hosted pool, so the times here are a bit worse than what we see on desktop PCs or laptops.

This made us consider making PublishReadyToRun default for Release builds on Windows. .NET MAUI is built on top of WinUI3 and the Windows App SDK to leverage the latest UI framework and APIs for Windows. WinUI3 templates already set PublishReadyToRun by default for Release mode, so this was a “no-brainer” to bring to .NET MAUI by default.

Note that PublishReadyToRun does increase application size, so if this is undesirable, you can simply set PublishReadyToRun to false in your project file.

See dotnet/maui#9357 for details about this improvement.

Dual Architectures by Default on macOS

A customer reported this behavior in their .NET MAUI application running on macOS:

  1. Launching the app after first install — over four seconds.
  2. Subsequent launches are under one second.
  3. A reboot causes the next launch of the application to be slow again.

It turns out, the application was built for x86_64 and running on an M1 (arm64) MacBook. This behavior is what you would see when Rosetta, Apple’s translation technology for Apple silicon processors, is in effect:

To the user, Rosetta is mostly transparent. If an executable contains only Intel instructions, macOS automatically launches Rosetta and begins the translation process. When translation finishes, the system launches the translated executable in place of the original. However, the translation process takes time, so users might perceive that translated apps launch or run more slowly at times.

We could avoid Rosetta translation behavior by building the application for multiple architectures:

<RuntimeIdentifiers Condition=" $([MSBuild]::GetTargetPlatformIdentifier('$(TargetFramework)')) == 'MacCatalyst' ">maccatalyst-x64;maccatalyst-arm64</RuntimeIdentifiers>

The MacCatalyst workload, in this case, builds for both architectures, merging both binaries into the final app.

This led us to default to building two architectures in Release mode for both net7.0-maccatalyst and net7.0-macos applications. In most cases, developers will want a dual-architecture application for the App Store. If this is undesirable, you can define a single $(RuntimeIdentifier) in your project file to opt out.

To verify what architecture your .NET MAUI application is built for, you can run:

file /Applications/YourApp.app/Contents/MacOS/YourApp
/Applications/YourApp.app/Contents/MacOS/YourApp: Mach-O 64-bit executable x86_64

Where a dual-architecture binary would result with:

file /Applications/YourApp.app/Contents/MacOS/YourApp
/Applications/YourApp.app/Contents/MacOS/YourApp: Mach-O universal binary with 2 architectures:
- Mach-O 64-bit executable x86_64
- Mach-O 64-bit executable arm64
/Applications/YourApp.app/Contents/MacOS/YourApp (for architecture x86_64):    Mach-O 64-bit executable x86_64
/Applications/YourApp.app/Contents/MacOS/YourApp (for architecture arm64):    Mach-O 64-bit executable arm64

See xamarin-macios#15769 for details about this improvement.

Notes about RegexOptions.Compiled

As seen in issue dotnet/runtime#71007, a customer’s app was spending ~258ms creating a Regex object in the .NET 6 version of their app, that was significantly higher than the previous Xamarin.Android counterpart:

dotnet-trace output for RegexOptions.Compiled

This showed up as a regression from Xamarin.Android to .NET 6, because RegexOptions.Compiled was not implemented at all in the BCL from mono/mono. While RegexOptions.Compiled is implemented in the BCL from .NET Core (and .NET 6+). The new behavior is that time is spent calling System.Reflection.Emit to generate code at runtime to make future Regex calls faster.

In this example, the Regex was only used once, so the setting was not actually needed at all. Avoid using RegexOptions.Compiled, unless you really are using the Regex many times, and are willing to pay the startup cost for improved throughput.

However! We have an even better solution in .NET 7, as described in Regular Expression Improvements in .NET 7.

So for example, take the following usage of RegexOptions.Compiled:

private static readonly Regex s_myCoolRegex = new Regex("abc|def", RegexOptions.Compiled | RegexOptions.IgnoreCase);
...
if (s_myCoolRegex.IsMatch(text)) { ... }

We can rewrite this in .NET 7 using [RegexGenerator], such as:

[RegexGenerator("abc|def", RegexOptions.IgnoreCase)]
private static partial Regex MyCoolRegex();
...
if (MyCoolRegex().IsMatch(text)) { ... }

Testing the new Regex implementation in a sample Android app, shows a noticeable improvement to startup time:

--Average(ms): 307.9
--Std Err(ms): 2.38723456930585
--Std Dev(ms): 7.54909854809757
++Average(ms): 214.4
++Std Err(ms): 3.06666666666667
++Std Dev(ms): 9.69765149118303

An average of 10 runs on a Pixel 5, shows a total ~93.5ms savings to startup just by using the new Regex APIs.

Diving deeper into dotnet-trace output, we can see RegexOptions.Compiled taking:

RegexOptions.Compiled Static Constructor RegexOptions.Compiled IsMatch

[RegexGenerator] is now so fast that the static constructor of MainActivity completely disappears from the trace. We are merely left with:

RegexGenerator IsMatch

There is even a built-in refactoring inside of Visual Studio to quickly make this change in your own applications:

Regex Refactoring

Improvements to Mono’s Interpreter

In .NET MAUI, the Mono runtime’s interpreter is used in two main scenarios:

  • C# Hot Reload: when debugging (or Debug configurations), the Mono interpreter is used by default.

  • iOS (and other Apple platforms) where JIT is restricted: the interpreter enables APIs like System.Reflection.Emit.

In .NET 7, Mono’s interpreter has gained “tiered compilation”. Where startup is greatly improved in methods that are only called once (or very few times). When a certain threshold of calls occurs, an optimization pass takes place on the intermediate representation (IR) of the code. This lets the interpreter achieve fast performance in frequently called code-paths in your applications.

We can see the direct result of these changes in a Blazor WASM application:

Performance of Interpreter Changes

However, we are also able to see a decent improvement in .NET MAUI scenarios. For example, take the following .NET MAUI applications running on a Pixel 5 device with solely the interpreter enabled:

Application Version Startup Time(ms)
dotnet new maui .NET 6 978.0
dotnet new maui .NET 7 775.3
.NET Podcast .NET 6 1171.1
.NET Podcast .NET 7 972.9

With these changes, you can expect .NET MAUI applications on iOS & Mac to have improved startup performance for the interpreter. We can also expect “inner development loop” improvements for Debugging and Hot Reload scenarios.

See dotnet/runtime#68823 for details about this improvement.

App Size Improvements

Fix DebuggerSupport Trimmer Value for Android

Reviewing .NET trimmer output for .NET 6 Android applications, we found methods decorated with attributes such as:

  • System.Diagnostics.DebuggableAttribute
  • System.Diagnostics.DebuggerBrowsableAttribute
  • System.Diagnostics.DebuggerDisplayAttribute
  • System.Diagnostics.DebuggerHiddenAttribute
  • System.Diagnostics.DebuggerStepThroughAttribute
  • System.Diagnostics.DebuggerVisualizerAttribute

Were not being removed in Release builds, attributing to larger application sizes.

We discovered the underlying cause was an incorrect value for the $(DebuggerSupport) trimmer flag, after fixing this, a “Hello World” Android application saw the following improvements:

--Java.Interop.dll             59419 bytes
++Java.Interop.dll             58642 bytes
--Mono.Android.dll             89327 bytes
++Mono.Android.dll             87877 bytes
--System.Private.CoreLib.dll  532428 bytes
++System.Private.CoreLib.dll  472318 bytes
--Overall package:           2738143 bytes
++Overall package:           2676703 bytes

This change will improve application size of all .NET 7 Android applications.

See xamarin-android#7176 for details about this improvement.

R8 Java Code Shrinker Improvements

R8 is whole-program optimization, shrinking and minification tool that converts java byte code to optimized Android dex code. R8 can be too aggressive, and remove something that is called by Java reflection, etc. We don’t yet have a good approach for making this the default across all .NET Android applications.

However, in .NET 7, we now pass any included proguard.txt files from Android .aar libraries to R8. AndroidX libraries that are implicitly used by .NET MAUI will use the ProGuard rules provided by Google. This should result in less problems when enabling R8 in .NET 7 Android applications.

To opt into using R8 for Release builds, add the following to your .csproj:

<!-- NOTE: not recommended for Debug builds! -->
<AndroidLinkTool Condition="'$(Configuration)' == 'Release'">r8</AndroidLinkTool>

See xamarin-android#5310 for details about this improvement.

Recent changes to Google’s AndroidX libraries are bringing more and more APIs implemented in Kotlin instead of Java. One result of this change, is additional “stuff” added to Android .apk files.

In the case of a dotnet new maui app:

  • kotlin directory: 386,289 bytes uncompressed, 154,519 bytes compressed
  • DebugProbesKt.bin: 1,719 bytes uncompressed, 777 bytes compressed

These files are general metadata from Kotlin, and are not used or needed in .NET Android applications. The kotlin folder has files with names such as kotlin/Error.kotlin_metadata.

We added a new MSBuild item group to exclude custom files from the final Android application. By default we include the following exclusions:

<ItemGroup>
  <AndroidPackagingOptionsExclude Include="DebugProbesKt.bin" />
  <AndroidPackagingOptionsExclude Include="$([MSBuild]::Escape('*.kotlin_*')" />
</ItemGroup>

This removes the extra files from any .NET 7+ Android application. It also gives developers the flexibility to remove other files in the future if needed.

See xamarin-android#7356 for details about this improvement.

Improve AOT Output for Generics

To improve the generated AOT code size of .NET 7 iOS applications, the following cases are now implemented:

  • If a generic method or a method of a generic type has all type parameters constrained to value types – methods for reference types will not be generated.

  • If a generic method or a method of a generic type has all type parameters constrained to reference types – methods for value types will not be generated.

In both cases, we were generating AOT code for methods that were not possible to be called. This is impactful for a large assembly like System.Private.CoreLib.dll.

The results of this change in a “hello world” iOS application:

File Before After diff %
HelloiOS 19539824 19158672 -381152 -1,95%
System.Private.CoreLib.dll-llvm.o 6870572 6734416 -136156 -1,98%

See dotnet/runtime#70838 for details about this improvement.

Tooling and Documentation

Profiling .NET MAUI Apps

The way to profile your .NET MAUI application varies depending on the platform. Mobile platforms leverage dotnet-trace and dotnet-dsrouter, while Windows desktop applications have access to PerfView.

To gather instructions for various platforms, we’ve aggregated docs for:

This should assist in profiling your own code — the same way we would profile to optimize .NET MAUI itself.

Measuring Startup Time

Just like profiling, the way to measure a .NET MAUI application’s startup time is drastically different on each platform.

We’ve collected guidance such as:

Additionally, we have a measure-startup tool for measuring startup for desktop .NET MAUI applications on Windows or macOS.

This tool can be built from source such as:

git clone https://github.com/jonathanpeppers/measure-startup.git
cd measure-startup
dotnet run -- dotnet help

Everything after -- for dotnet run are arguments passed to measure-startup, where this example times how long the dotnet command launches and takes to print the text help. This example isn’t that useful, but can be easily applied to a .NET MAUI application.

To measure a .NET MAUI app, first add a subscription to your main Page‘s Loaded event:

Loaded += (sender, e) => Dispatcher.Dispatch(() => Console.WriteLine("loaded"));

Dispatcher is used to give the app one pass for rendering/layout. In an app using BlazorWebView, you might consider logging this message when the web view finishes loading.

On Windows, build the app for Release mode, such as:

dotnet publish -f net7.0-windows10.0.19041.0 -c Release

You can then measure startup time via:

dotnet run -c Release -- bin\Release\net7.0-windows10.0.19041.0\win10-x64\publish\YourApp.exe loaded

This launches YourApp.exe recording the time it takes for loaded to be printed to stdout:

0:00:07.0961628
Dropping first run...
0:00:01.4743315
0:00:01.4700848
0:00:01.4834235
0:00:01.4752893
0:00:01.4695317
Average(ms): 1474.53216
Std Err(ms): 2.4945246400065977
Std Dev(ms): 5.577926666602944

This also works for .NET MAUI applications running on macOS:

dotnet publish -f net7.0-maccatalyst -c Release
dotnet run -- bin/Release/net7.0-maccatalyst/YourApp.app/Contents/MacOS/YourApp loaded

If the measure-startup tool is useful for you, let us know. We can release it as a .NET global tool on NuGet.org if there is enough interest.

App Size Reporting Tools

To assist with diagnosing app size, we have two tools for Android and iOS: apkdiff and dotnet-appsize-report.

apkdiff can compare sizes between two Android .apk files and report what changed.

To simulate what will be delivered to a user’s device from Google Play (which now uses Android App Bundles), build an .apk for a single architecture:

dotnet build -c Release -f net7.0-android -r android-arm64 -p:AndroidPackageFormat=apk

Google Play delivers a device-specific .apk to users’ devices, and so we can mimic this behavior by building for a single runtime identifier: -r android-arm64.

You can compare two .apk files, using apkdiff such as:

dotnet tool install --global apkdiff
...
apkdiff -f before.apk after.apk
Size difference in bytes ([*1] apk1 only, [*2] apk2 only):
+           8 lib/arm64-v8a/libaot-Microsoft.Maui.Controls.Xaml.dll.so
-         608 lib/arm64-v8a/libaot-Microsoft.Maui.Controls.Compatibility.dll.so
-       3,448 lib/arm64-v8a/libaot-Microsoft.Maui.Controls.dll.so
-       4,424 lib/arm64-v8a/libaot-Microsoft.Maui.dll.so
-       5,656 lib/arm64-v8a/libaot-Microsoft.NetConf2021.Maui.dll.so
-     129,617 assemblies/assemblies.blob
Summary:
-     129,617 Other entries -1.03% (of 12,605,899)
+           0 Dalvik executables 0.00% (of 11,756,376)
-      14,128 Shared libraries -0.10% (of 14,849,096)
-     131,072 Package size difference -0.64% (of 20,380,049)

Likewise for iOS, we have dotnet-appsize-report:

git clone https://github.com/rolfbjarne/dotnet-appsize-report.git
cd dotnet-appsize-report
dotnet run -- -i path/to/bin/Release/net7.0-ios/YourApp.app

Which can give reports for:

App size report for YourApp.app

1) View 167 assemblies
2) View 381 namespaces
3) View 26592 types
q) Quit

And show .NET assemblies, namespaces, or types sorted by IL size. For example:

Assembly                                            Types    IL Size         File size
System.Private.CoreLib                              2,157  1,126,514   3,605,144 bytes
System.Private.Xml                                  1,667  1,378,583   3,099,800 bytes
Microsoft.iOS                                      14,978  5,086,414  24,121,232 bytes

Experimental or Advanced Options

Just as in .NET 6, many of the same experimental or advanced options are still available in .NET 7:

In future .NET releases, we will work toward making some of these settings the default — where it best makes sense.

Conclusion

We hope that our continued investment in performance in .NET MAUI will be helpful for your own cross-platform desktop and mobile applications. Always be profiling!

Please try out .NET MAUI, file issues, or learn more at http://dot.net/maui!

8 comments

Comments are closed. Login to edit/delete your existing comments

  • Ronan Burke 0

    Great blog!

    I have one question about “This treatment was applied for every type converter currently used at runtime in the dotnet new maui template”. Is there any potential for more type converters that aren’t included in the default template to be given this treatment? It sounds like the optimisations done will help the blank template (obviously this helps real world apps too), but are there more real-world gains to be made in apps that use other type converters?

    • Jonathan PeppersMicrosoft employee 1

      There is a set of default styles in the dotnet new maui template, that likely covered the common cases. But it’s still possible your app is using something we could optimize.

      If you want to check, decompile your app’s main assembly in a Release build with a tool such as ILSpy. Looks for examples like: new SomeConverter().ConvertFromInvariantString("yourvalue") You should be able to find these in the InitializeComponent() method of any page or view with .xaml.

    • Jonathan PeppersMicrosoft employee 0

      Yes, when the Windows App SDK (WinUI3) supports NativeAOT, we could use it with .NET MAUI on Windows. I think there might be some problems with System.Reflection usage and NativeAOT, though. Bindings in .NET MAUI, among other things, use a lot of System.Reflection.

      • Charles Roddie 0

        Bindings in .NET MAUI, among other things, use a lot of System.Reflection.

        That’s pretty shocking. Are bindings essential to MAUI or just one way that some users use it (with XAML markup)?

  • Roman Jašek 0

    Thanks a lot for a great post like last time. This is very appreciated as it communicates the improvements and, in some cases, it also provides action points for us developers to do in our applications.

    Just one minor ask – I’m trying to put together the numbers in terms of improvements so that I have a general overview of how things are going. You post numbers in terms of startup times for Android and application size for iOS using the various samples.
    In the last post you provided the startup times for some applications and in this post for a subset of them. I came up with this table but there are some applications missing in there in different versions. Would it be possible to fill in the blanks (at least for .NET 7)?

    Application Xamarin MAUI – Preview MAUI – .NET 6 MAUI – .NET 7
    Android 306.5 210.5 182.8
    Without Shell 498.6 683.9 464.2
    With Shell 817.7 568.1 545.4
    .NET Podcast App 1299.9 814.2 759.7

    [Edit] Well, the formatting got mangled – I’ll link to an image of the table: https://imgur.com/a/53EM8av

  • Christian Campos 1

    I haven’t even been able to see .net 6 properly and version 7 is about to come out…

  • sinasinayi9911 0

    Microsoft’s efforts to increase the performance of .NET 6 include rewriting the System.IO.FileStream section. This section is about working with files.
    This means that the speed of working with files in .NET 6. has increased significantly.
    According to the results of the benchmarks, it can be said that the work speed and performance in .NET 6. compared to all previous versions of .NET has increased and improved significantly.
    Now in version 7, source by github it has the best articles. which has become much better.

Feedback usabilla icon