Mono AOT decoder workaround for slow jpeg decoding.

[JimBobSquarePants]

Prerequisites

• I have written a descriptive pull-request title
• I have verified that there are no overlapping pull-requests open
• I have verified that I am following the existing coding patterns and practice as demonstrated in the repository. These follow strict Stylecop rules 👮.
• I have provided test coverage for my change (where applicable)

Description

This PR is an attempt to work around issues found with the Mono AOT compiler which causes slow performance on IOS Android

dotnet/runtime#71210

In the linked Issue, Analysis from the Mono team highlighted this indirection as a culprit.

The AOT compiler does appear to have problems figuring out which instances to generate. In this specific case, the caller is
ImageDecoderUtilities:Decode<Rgba32> which calls IImageDecoderInternals::Decode on an argument. So in theory, the aot compiler could figure out that the call could possible go to JpegDecoderCore::Decode<Rgba32> and generate that instance. Currently, this kind of analysis is not done.

This PR removes that indirection completely by introducing an internal base class, ImageDecoderCore for all XXDecoderCore instances. In addition, seeding has been introduced for SpectralConverter<TPixel> and others.

This is currently untested but I'm confident that this should improve matters.

[beeradmoore]

I don't think this PR had any impact for what I have tested so far (assuming I built the PR correctly), but I also didn't see the super slow loads the initial person reported. I do still need to test Android.

I have a test repo here. It is a .NET 8 MAUI application. I will test Windows and Mac out of curiosity. I won't test Xamarin.Forms (Xamarin.iOS/Xamarin.Android)

The readme has instructions on how I built this PR into a local nuget package, and how I tested the project on my iPhone 15 Pro Max in both debug and release mode.

The test image I am using is located in ImageSharpMAUITest/ImageSharpMAUITest/Resources/Raw/sloth.jpg

The ImageSharp code I am using is here.

I am loading the image from a stream in the raw folder. I don't think there is any overhead from loading it like this instead of passing file on disk.

using(varstream=awaitFileSystem.OpenAppPackageFileAsync("sloth.jpg")){using(varimage=awaitSixLabors.ImageSharp.Image.LoadAsync(stream)){
...}}

Results

Updating these results as they come in

Debug (this PR)

Device	JpgLoad	JpgResize	PngLoad	PngResize
iPhone	3032.1ms	4060.1ms	26.2ms	73.9ms
iOS Simulator	2884.9ms	3615.2ms	21.7ms	61.5ms
Android	14342.7ms	18562.1ms	123.8ms	348ms
Android Emulator	3193.1ms	3933ms	48.3ms	88.7ms
macOS	2708.2ms	3512.4ms	21.4ms	60.9ms
Windows	106.6ms	82.3ms	26.3ms	17.1ms

Debug (3.1.4)

Device	JpgLoad	JpgResize	PngLoad	PngResize
iPhone	2932.4ms	3895.8ms	25ms	67.7ms
iOS Simulator	2794.2ms	3555.1ms	21.9ms	61.4ms
Android	14289.4ms	18504.5ms	122.7ms	346.4ms
Android Emulator	3027.6ms	4075.2ms	48.9ms	97.6ms
macOS	2701.9ms	3425.3ms	20.6ms	59.3ms
Windows	98.5ms	86.4ms	22.4ms	19.7ms

Release (this pr)

Device	JpgLoad	JpgResize	PngLoad	PngResize
iPhone	62.7ms	80.3ms	4.2ms	4ms
iOS Simulator	64.3ms	81.3ms	3.7ms	4.1ms
Android	1134.5ms	1369.2ms	33.7ms	41.6ms
Android Emulator	243.3ms	300.8ms	19.6ms	25.3ms
macOS	188.5ms	233.9ms	4.9ms	7ms
Windows	149.3ms	96.4ms	16.9ms	14.7ms

Release (3.1.4)

Device	JpgLoad	JpgResize	PngLoad	PngResize
iPhone	61.2ms	78.8ms	4ms	4ms
iOS Simulator	61.1ms	79.6ms	3.4ms	4.1ms
Android	1121.5ms	1349.6ms	35.1ms	43ms
Android Emulator	227.8ms	287.9ms	16.7ms	13.5ms
macOS	181.5ms	230.1ms	5.2ms	7.3ms
Windows	153.1ms	146.4ms	17.2ms	15ms

Test devices

Device	Hardware	OS
iPhone	iPhone 15 Pro Max	17.5.1
iOS Simulator	iPhone 15	17.4
Android	Pixel 2 XL	Android 14
Android Emulator	Pixel 3a	Android 14
macOS	MacBook Pro M3 Max	Sonoma 14.5
Windows	AMD 3300X + RTX 3060	Windows 11 23H2

[beeradmoore]

Added all the numbers.

My takeaway from this (as someone who uses images, but doesn't really know much about image encoding) is:

• Jpg is way slower than png
• Debug mode in almost all platforms is slower than I would have expected
• Jpgs on an android device in debug mode is wild
• Shocked to see jpg on android emulator be faster than android on device
• I don't see any noticeable difference in the results of this PR and the currently published 3.1.4. I don't know if there is meant to be additional changes to using AOT mode for the tests or something.

More than happy to add and run more tests as requested. Attaching the raw results numbers below (if anyone wants to see the results per run).

ImageSharpMAUITest-Results.zip

[JimBobSquarePants]

@beeradmoore I've been doing some R&D in order to understand the Android performance. I can't see a configuration value in your sample to enable LLVM which is required for maximum performance. Did I miss something?

https://learn.microsoft.com/en-us/dotnet/android/building-apps/build-properties#enablellvm

[beeradmoore]

I did not.

I added

 <PropertyGroup Condition="$([MSBuild]::GetTargetPlatformIdentifier('$(TargetFramework)')) == 'android' AND '$(Configuration)' == 'Release'">
<EnableLLVM>true</EnableLLVM>
</PropertyGroup>

And during executing all the tests it gets to about 15/40 (half way through jpg load and resize run) and it crashes with

07-16 12:36:03.133 1730 1778 W WindowManager: Exception thrown during dispatchAppVisibility Window{647a9a8 u0 com.beeradmoore.imagesharpmauitest/crc64df7e0c4a761c65fa.MainActivity EXITING}
07-16 12:36:03.133 1730 1778 W WindowManager: android.os.DeadObjectException
07-16 12:36:03.133 1730 1778 W WindowManager: at android.os.BinderProxy.transactNative(Native Method)
07-16 12:36:03.133 1730 1778 W WindowManager: at android.os.BinderProxy.transact(BinderProxy.java:586)
07-16 12:36:03.133 1730 1778 W WindowManager: at android.view.IWindow$Stub$Proxy.dispatchAppVisibility(IWindow.java:552)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.WindowState.sendAppVisibilityToClients(WindowState.java:3217)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.WindowContainer.sendAppVisibilityToClients(WindowContainer.java:1293)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.WindowToken.setClientVisible(WindowToken.java:403)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.ActivityRecord.setClientVisible(ActivityRecord.java:7100)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.ActivityRecord.postApplyAnimation(ActivityRecord.java:5820)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.ActivityRecord.commitVisibility(ActivityRecord.java:5762)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.Transition.finishTransition(Transition.java:1257)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.TransitionController.finishTransition(TransitionController.java:925)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.WindowOrganizerController.finishTransition(WindowOrganizerController.java:489)
07-16 12:36:03.133 1730 1778 W WindowManager: at android.window.IWindowOrganizerController$Stub.onTransact(IWindowOrganizerController.java:278)
07-16 12:36:03.133 1730 1778 W WindowManager: at com.android.server.wm.WindowOrganizerController.onTransact(WindowOrganizerController.java:199)
07-16 12:36:03.133 1730 1778 W WindowManager: at android.os.Binder.execTransactInternal(Binder.java:1500)
07-16 12:36:03.133 1730 1778 W WindowManager: at android.os.Binder.execTransact(Binder.java:1444)

This is new to me and I have no idea what or why it is doing this. I have re-installed AndroidOS between then and now, this could be related, it could be AOT compile. Tonight I can revert back to not having LLVM and see if to happens again or not. It could also be I did no have developer mode enabled so maybe it was going to sleep in 15sec 🤷‍♂️

I did a single run of with v3.1.4 in release mode running the JpgLoad test and got 1116.8ms. For the local nuget I got 1126.6ms.

I referred to the previous thread and grabbed info from this comment,

and just put

<EnableLLVM>true</EnableLLVM>
<RunAOTCompilation>true</RunAOTCompilation>
<AndroidEnableProfiledAot>false</AndroidEnableProfiledAot>

in my main PropertyGroup to ensure it is enabled.

With that I got 516ms for the local nuget. Huge improvement!

I can test again tonight to see what ones of those above properties are required, if my release Android targeting below was working as intended.

[JimBobSquarePants]

@beeradmoore Did you ever get those updated Android benchmarks?

[beeradmoore]

@JimBobSquarePants , aside from that single test I did not. Creating a reminder to do it this weekend.

[JimBobSquarePants]

Thanks @beeradmoore very much appreciated!

[beeradmoore]

So some more checking into things. For these results I will call things run #1 - #6 with each run having a different csproj config. These different configs can also impact filesize.

I think we are also focusing on changes to Android, so these tests are only on Release mode on my physical Pixel 2 XL. If there are any specific AOT flags or anything you want me to re-run on other platforms just let me know what ones.

Settings used for each run:

Run 1

Of the 3 options listed in the other runs, this run did not have them in the csproj. I did not lookk what the net8.0-android defaults are for release mode.

Run 2

<EnableLLVM>true</EnableLLVM>

This resulted in a crush mid way through the second test. While the tests could be ran individually there is an issue if you can't run them back to back so I am considering these invalid.

07-27 14:06:30.804 7895 7961 E gesharpmauitest: * Assertion at /__w/1/s/src/mono/mono/mini/aot-runtime.c:5220, condition `plt_entry' not met
07-27 14:06:30.806 7895 7961 F libc : Fatal signal 6 (SIGABRT), code -1 (SI_QUEUE) in tid 7961 (.NET TP Worker), pid 7895 (gesharpmauitest)

More info on these crashes below.

Run 3

<EnableLLVM>true</EnableLLVM>
<RunAOTCompilation>true</RunAOTCompilation>

This also resulted in a crash.

07-27 14:12:55.543 8325 8359 F libc : Fatal signal 6 (SIGABRT), code -1 (SI_QUEUE) in tid 8359 (.NET TP Gate), pid 8325 (gesharpmauitest)
07-27 14:12:55.543 8325 8378 E gesharpmauitest: * Assertion at /__w/1/s/src/mono/mono/mini/aot-runtime.c:5220, condition plt_entry' not met 07-27 14:12:55.543 8325 8358 E gesharpmauitest: * Assertion at /__w/1/s/src/mono/mono/mini/aot-runtime.c:5220, condition plt_entry' not met

More info on these crashes below.

Run 4

<EnableLLVM>true</EnableLLVM>
<RunAOTCompilation>true</RunAOTCompilation>
<AndroidEnableProfiledAot>false</AndroidEnableProfiledAot>

Run 5

<EnableLLVM>true</EnableLLVM>
<AndroidEnableProfiledAot>false</AndroidEnableProfiledAot>

Same as Run 4, but without the RunAOTCompilation defined. I did not check the defaults for these values, but size and performance appears teh same so I am going to assume RunAOTCompilation is defaulting to true.

Run 6

<EnableLLVM>false</EnableLLVM>
<AndroidEnableProfiledAot>true</AndroidEnableProfiledAot>

This is the same as run 5 but it swaps EnableLLVM and AndroidEnableProfiledAot around. Size and results appears this is the default when no params are defined.

Results (3.1.4)

Run	JpgLoad	JpgResize	PngLoad	PngResize	aab size
1	1114.9ms	1353.9ms	34.0ms	41.2ms	17.8 MB
2	Crash	Crash	Crash	Crash	17.7 MB
3	Crash	Crash	Crash	Crash	17.7 MB
4	515.3ms	712.2ms	26.9ms	35.4ms	24.1 MB
5	512.0ms	698.8ms	26.6ms	33.9ms	24.1 MB
6	1116.4ms	1355.3ms	33.9ms	40.9ms	17.8 MB

Results (this pr)

Run	JpgLoad	JpgResize	PngLoad	PngResize	aab size
1	1124.5ms	1361.0ms	33.2ms	41.0ms	17.8 MB
2	Crash	Crash	Crash	Crash	17.7 MB
3	Crash	Crash	Crash	Crash	17.7 MB
4	510.1ms	709.8ms	27.2ms	34.5ms	24.1 MB
5	512.7ms	713.2ms	26.8ms	34.8ms	24.1 MB
6	1123.1ms	1352.3ms	34.1ms	40.5ms	17.8 MB

Crashes

The crashes looks like they are discussed in this issue in the dotnet runtime repo.

Conclusion

My takeaway from this is the currently release 3.1.4 and this PR perform the same, at least for the limited tests I am doing. Other situations, or possibly other devices, may perform differently. What the new code is doing is above my skillset, but I don't see it degrading performance so I don't think its a bad change.

If people want more performance with ImageSharp for Android then they will get it at an incrased filesize by enabling these options for release mode Android builds.

<PropertyGroupCondition="$([MSBuild]::GetTargetPlatformIdentifier('$(TargetFramework)')) == 'android' AND '$(Configuration)' == 'Release'">
<EnableLLVM>true</EnableLLVM>
<AndroidEnableProfiledAot>false</AndroidEnableProfiledAot>
</PropertyGroup>

Using EnableLLVM is likely not enough as you could still have crashes, but those crashes may not appear all the time. EG. I can run the second test individually and it will pass, but if I run all tests back to back it will crash halfway through the second set of tests.

I am unsure if there are more optimisations you can do, or if the next step of improvemnets requires the dotnet team to do things on their end.

I am also unsure if writing C# code using Android SDK for image manipulation gets it going any faster or if I am now approching limits on my 7 year old phone. Doing so would then require more devtime to write platform specific code for every single image manipulation piece I'd do for each platform I'd run my app on. It's also then just a lot more code in general to maintain. Whereas ImageSharp appears to do a great job across the board, except for Android where it just does an ok job OR I just use pngs everywhere, becaause they were always fast everywhere 😅

[JimBobSquarePants]

Thanks @beeradmoore for the monumental effort you've put in here. I think this should actually enable us to close off several longstanding issues.

Judging by this comment and this comment the following combination should be the recommended one for Android.

<EnableLLVM>true</EnableLLVM>
<RunAOTCompilation>true</RunAOTCompilation>
<AndroidEnableProfiledAot>false</AndroidEnableProfiledAot>

iOS performance looks fantastic and Android, I imagine will improve with new phones and codegen improvements. I think most complaints are either outdated or the result of miscofiguration.

Adding the Android and iOS specific MAUI configuration from your sample with an explanation to this section of the docs will help eleviate issues going forward.

Despite relative equality of the performance metrics of the two source versions I think merging this PR is still a wise choice. It's much easier to maintain internally as it removes misdirection.

[beeradmoore]

I think one of the things that will catch people off guard is running in debug mode being slower. I would expect things to be slower in debug mode, but not 14sec to load an image. That would have caught me off guard and I would have questioned the library long before trying in release mode.

[JimBobSquarePants]

I think one of the things that will catch people off guard is running in debug mode being slower. I would expect things to be slower in debug mode, but not 14sec to load an image. That would have caught me off guard and I would have questioned the library long before trying in release mode.

Yeah... Definitely need to docuement that also and raise issues with the relevant parties.

[JimBobSquarePants]

I've added configuration notes to our docs. Please let me know if I've missed anything.

https://docs.sixlabors.com/articles/imagesharp/gettingstarted.html#maui-performance