Merge branch 'main' into codegen-optimizations

.github/workflows/build-and-test.yml

@@ -56,11 +56,6 @@ jobs:
             sdk: 6.0.x
             runtime: -x64
             codecov: false
-          - os: buildjet-4vcpu-ubuntu-2204-arm
-            framework: net6.0
-            sdk: 6.0.x
-            runtime: -x64
-            codecov: false
         exclude:
           - isARM: false
             options:

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.RgbArm.cs

@@ -0,0 +1,53 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+
+using System.Runtime.CompilerServices;
+using System.Runtime.InteropServices;
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.Arm;
+using System.Runtime.Intrinsics.X86;
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
+
+internal abstract partial class JpegColorConverterBase
+{
+    internal sealed class RgbArm : JpegColorConverterArm
+    {
+        public RgbArm(int precision)
+            : base(JpegColorSpace.RGB, precision)
+        {
+        }
+
+        /// <inheritdoc/>
+        public override void ConvertToRgbInplace(in ComponentValues values)
+        {
+            ref Vector128<float> rBase =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
+            ref Vector128<float> gBase =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
+            ref Vector128<float> bBase =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component2));
+
+            // Used for the color conversion
+            var scale = Vector128.Create(1 / this.MaximumValue);
+            nint n = (nint)(uint)values.Component0.Length / Vector128<float>.Count;
+            for (nint i = 0; i < n; i++)
+            {
+                ref Vector128<float> r = ref Unsafe.Add(ref rBase, i);
+                ref Vector128<float> g = ref Unsafe.Add(ref gBase, i);
+                ref Vector128<float> b = ref Unsafe.Add(ref bBase, i);
+                r = AdvSimd.Multiply(r, scale);
+                g = AdvSimd.Multiply(g, scale);
+                b = AdvSimd.Multiply(b, scale);
+            }
+        }
+
+        /// <inheritdoc/>
+        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
+        {
+            rLane.CopyTo(values.Component0);
+            gLane.CopyTo(values.Component1);
+            bLane.CopyTo(values.Component2);
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterArm.cs

@@ -0,0 +1,35 @@
+// Copyright (c) Six Labors.
+// Licensed under the Six Labors Split License.
+using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.Arm;
+using System.Runtime.Intrinsics.X86;
+
+namespace SixLabors.ImageSharp.Formats.Jpeg.Components;
+
+internal abstract partial class JpegColorConverterBase
+{
+    /// <summary>
+    /// <see cref="JpegColorConverterBase"/> abstract base for implementations
+    /// based on <see cref="Avx"/> instructions.
+    /// </summary>
+    /// <remarks>
+    /// Converters of this family would expect input buffers lengths to be
+    /// divisible by 8 without a remainder.
+    /// This is guaranteed by real-life data as jpeg stores pixels via 8x8 blocks.
+    /// DO NOT pass test data of invalid size to these converters as they
+    /// potentially won't do a bound check and return a false positive result.
+    /// </remarks>
+    internal abstract class JpegColorConverterArm : JpegColorConverterBase
+    {
+        protected JpegColorConverterArm(JpegColorSpace colorSpace, int precision)
+            : base(colorSpace, precision)
+        {
+        }
+
+        public static bool IsSupported => AdvSimd.IsSupported;
+
+        public sealed override bool IsAvailable => IsSupported;
+
+        public sealed override int ElementsPerBatch => Vector128<float>.Count;
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterBase.cs

@@ -214,6 +214,11 @@ internal abstract partial class JpegColorConverterBase
             return new RgbAvx(precision);
         }
 
+        if (JpegColorConverterArm.IsSupported)
+        {
+            return new RgbArm(precision);
+        }
+
         if (JpegColorConverterVector.IsSupported)
         {
             return new RgbScalar(precision);

tests/ImageSharp.Benchmarks/Codecs/Jpeg/ColorConversion/RgbColorConversion.cs

@@ -37,4 +37,12 @@ public class RgbColorConversion : ColorConversionBenchmark
 
         new JpegColorConverterBase.RgbAvx(8).ConvertToRgbInplace(values);
     }
+
+    [Benchmark]
+    public void SimdVectorArm()
+    {
+        var values = new JpegColorConverterBase.ComponentValues(this.Input, 0);
+
+        new JpegColorConverterBase.RgbArm(8).ConvertToRgbInplace(values);
+    }
 }

tests/ImageSharp.Tests/PixelFormats/PixelBlenders/PorterDuffFunctionsTests.cs

@@ -3,6 +3,8 @@
 
 using System.Numerics;
 using System.Runtime.Intrinsics;
+using System.Runtime.Intrinsics.X86;
+using Castle.Components.DictionaryAdapter;
 using SixLabors.ImageSharp.PixelFormats.PixelBlenders;
 using SixLabors.ImageSharp.Tests.TestUtilities;
 
@@ -30,6 +32,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(NormalBlendFunctionData))]
     public void NormalBlendFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -57,6 +64,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(MultiplyFunctionData))]
     public void MultiplyFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -84,6 +96,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(AddFunctionData))]
     public void AddFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -111,6 +128,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(SubtractFunctionData))]
     public void SubtractFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -138,6 +160,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(ScreenFunctionData))]
     public void ScreenFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -165,6 +192,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(DarkenFunctionData))]
     public void DarkenFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -192,6 +224,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(LightenFunctionData))]
     public void LightenFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -219,6 +256,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(OverlayFunctionData))]
     public void OverlayFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);
 
@@ -246,6 +288,11 @@ public class PorterDuffFunctionsTests
     [MemberData(nameof(HardLightFunctionData))]
     public void HardLightFunction256(TestVector4 back, TestVector4 source, float amount, TestVector4 expected)
     {
+        if (!Avx.IsSupported)
+        {
+            return;
+        }
+
         Vector256<float> back256 = Vector256.Create(back.X, back.Y, back.Z, back.W, back.X, back.Y, back.Z, back.W);
         Vector256<float> source256 = Vector256.Create(source.X, source.Y, source.Z, source.W, source.X, source.Y, source.Z, source.W);