Support for Windows ARM64

[kleisauke]

The native binaries for FreeImage are not available for Windows ARM64, so I excluded it from the benchmarks. Also, the benchmarks need to be built with portable debug symbols due to dotnet/roslyn#42393 and the NativeMemoryProfiler is unable to run on Windows ARM64 due to microsoft/perfview#1264.

This was tested using my Raspberry Pi 4B (4GB ram)¹ with the prebuilt libvips binaries from https://github.com/libvips/build-win64-mxe/releases/tag/v8.10.2-build2. I intend to support Windows ARM64 within the NetVips.Native NuGet package for the next release.

Benchmark results:

BenchmarkDotNet=v0.12.1, OS=Windows 10.0.19042
BCM2711 (ARM Cortex-A72), 1 CPU, 4 logical and 4 physical cores
.NET Core SDK=5.0.100
  [Host]       : .NET Core 5.0.0 (CoreCLR 5.0.20.51904, CoreFX 5.0.20.51904), Arm64 RyuJIT
  .Net 5.0 CLI : .NET Core 5.0.0 (CoreCLR 5.0.20.51904, CoreFX 5.0.20.51904), Arm64 RyuJIT

Job=.Net 5.0 CLI  Arguments=/p:DebugType=portable  Toolchain=.NET Core 5.0  
IterationCount=5  LaunchCount=1  WarmupCount=5  

|                    Method |      Mean |      Error |    StdDev | Ratio | RatioSD | Gen 0 | Gen 1 | Gen 2 | Allocated |
|-------------------------- |----------:|-----------:|----------:|------:|--------:|------:|------:|------:|----------:|
|   'System.Drawing Resize' | 59.160 ms | 21.3092 ms | 5.5339 ms |  1.00 |    0.00 |     - |     - |     - |     136 B |
|       'ImageSharp Resize' | 55.589 ms | 33.1113 ms | 8.5989 ms |  0.95 |    0.18 |     - |     - |     - |    8992 B |
|      'MagicScaler Resize' |  9.221 ms |  1.6021 ms | 0.2479 ms |  0.16 |    0.01 |     - |     - |     - |    2352 B |
| 'SkiaSharp Canvas Resize' | 14.113 ms |  4.0724 ms | 1.0576 ms |  0.24 |    0.03 |     - |     - |     - |    1424 B |
| 'SkiaSharp Bitmap Resize' | 13.972 ms |  1.2031 ms | 0.1862 ms |  0.24 |    0.02 |     - |     - |     - |     488 B |
|          'NetVips Resize' | 24.687 ms |  0.2549 ms | 0.0394 ms |  0.42 |    0.04 |     - |     - |     - |    4144 B |

|                                Method |       Mean |     Error |   StdDev | Ratio | RatioSD |     Gen 0 | Gen 1 | Gen 2 |  Allocated |
|-------------------------------------- |-----------:|----------:|---------:|------:|--------:|----------:|------:|------:|-----------:|
|   'System.Drawing Load, Resize, Save' | 1,339.5 ms |  33.36 ms |  8.66 ms |  1.00 |    0.00 |         - |     - |     - |   11.98 KB |
|       'ImageSharp Load, Resize, Save' | 1,627.4 ms |  38.12 ms |  9.90 ms |  1.22 |    0.01 | 3000.0000 |     - |     - | 1988.43 KB |
|      'MagicScaler Load, Resize, Save' |   410.2 ms |  54.97 ms | 14.28 ms |  0.31 |    0.01 |         - |     - |     - |  109.62 KB |
| 'SkiaSharp Canvas Load, Resize, Save' | 2,125.7 ms | 105.01 ms | 27.27 ms |  1.59 |    0.01 |         - |     - |     - |   95.95 KB |
| 'SkiaSharp Bitmap Load, Resize, Save' | 2,106.1 ms |  92.01 ms | 23.89 ms |  1.57 |    0.02 |         - |     - |     - |   83.77 KB |
|          'NetVips Load, Resize, Save' |   664.5 ms |  42.11 ms | 10.94 ms |  0.50 |    0.01 |         - |     - |     - |   47.53 KB |

|                                           Method |     Mean |     Error |   StdDev | Ratio | RatioSD | Gen 0 | Gen 1 | Gen 2 |  Allocated |
|------------------------------------------------- |---------:|----------:|---------:|------:|--------:|------:|------:|------:|-----------:|
|   'System.Drawing Load, Resize, Save - Parallel' | 569.9 ms | 362.18 ms | 56.05 ms |  1.00 |    0.00 |     - |     - |     - |   24.13 KB |
|       'ImageSharp Load, Resize, Save - Parallel' | 604.7 ms | 238.97 ms | 62.06 ms |  1.08 |    0.18 |     - |     - |     - | 2001.97 KB |
|      'MagicScaler Load, Resize, Save - Parallel' | 139.9 ms |  64.97 ms | 10.05 ms |  0.25 |    0.04 |     - |     - |     - |   120.1 KB |
| 'SkiaSharp Canvas Load, Resize, Save - Parallel' | 719.5 ms | 168.89 ms | 26.14 ms |  1.27 |    0.16 |     - |     - |     - |  106.21 KB |
| 'SkiaSharp Bitmap Load, Resize, Save - Parallel' | 740.5 ms | 172.60 ms | 44.82 ms |  1.32 |    0.05 |     - |     - |     - |   94.11 KB |
|          'NetVips Load, Resize, Save - Parallel' | 296.0 ms |  84.41 ms | 21.92 ms |  0.54 |    0.08 |     - |     - |     - |   59.23 KB |

¹: For a more realistic benchmark it would be nice to test it on a Surface Pro X as well. Does anyone have access to such device?

/cc @jcupitt, as he might be interested in the above numbers.

[jcupitt]

Wow, well done MagicScaler!

I guess this is with the CopyMemory() change we discussed last week? Is Orc working on ARM?

[kleisauke]

This is with the CopyMemory() only at the end and with caching disabled. Although it is compiled against orc, it seems that the C-path is used instead. Probably this is because https://gitlab.freedesktop.org/gstreamer/orc/-/merge_requests/50 has landed recently and there is no new release available yet with that included.

I'll try to build against the orc master branch to see if it makes any difference.

[bleroy]

@kleisauke LMK if you need more time to get changes to this PR in or if you want me to merge it right away. Thanks for the contribution.

[kleisauke]

@bleroy Feel free to merge. I'll do another PR if any of the above linked issues is resolved.

@jcupitt I had a go to build the ARM64 binaries against the orc master branch with commit libvips/build-win64-mxe@2c4c6ed. It seems that this does not cause a significant difference in the benchmarks. I also came across assertion failures while testing some utilities provided by orc (reported here), so that commit is not ready for production yet.

[bleroy]

That new patch file doesn't look like it should be in the commit.

[kleisauke]

@bleroy Do you mean commit libvips/build-win64-mxe@2c4c6ed? That one is committed on the orc-master branch within the libvips/build-win64-mxe repository, so it is not included in this PR. It only refers to this PR within the commit message.

Sorry for the confusion.

[bleroy]

Oh yes, you're right, I got confused :)