Add arm64 intrinsics for cmyk converter

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverter.CmykArm64.cs

@@ -0,0 +1,96 @@
+// 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 CmykArm64 : JpegColorConverterArm
+    {
+        public CmykArm64(int precision)
+            : base(JpegColorSpace.Cmyk, precision)
+        {
+        }
+
+        /// <inheritdoc/>
+        public override void ConvertToRgbInplace(in ComponentValues values)
+        {
+            ref Vector128<float> c0Base =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
+            ref Vector128<float> c1Base =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
+            ref Vector128<float> c2Base =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component2));
+            ref Vector128<float> c3Base =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+            // Used for the color conversion
+            var scale = Vector128.Create(1 / (this.MaximumValue * this.MaximumValue));
+
+            nint n = (nint)(uint)values.Component0.Length / Vector128<float>.Count;
+            for (nint i = 0; i < n; i++)
+            {
+                ref Vector128<float> c = ref Unsafe.Add(ref c0Base, i);
+                ref Vector128<float> m = ref Unsafe.Add(ref c1Base, i);
+                ref Vector128<float> y = ref Unsafe.Add(ref c2Base, i);
+                Vector128<float> k = Unsafe.Add(ref c3Base, i);
+
+                k = AdvSimd.Multiply(k, scale);
+                c = AdvSimd.Multiply(c, k);
+                m = AdvSimd.Multiply(m, k);
+                y = AdvSimd.Multiply(y, k);
+            }
+        }
+
+        /// <inheritdoc/>
+        public override void ConvertFromRgb(in ComponentValues values, Span<float> rLane, Span<float> gLane, Span<float> bLane)
+            => ConvertFromRgb(in values, this.MaximumValue, rLane, gLane, bLane);
+
+        public static void ConvertFromRgb(in ComponentValues values, float maxValue, Span<float> rLane, Span<float> gLane, Span<float> bLane)
+        {
+            ref Vector128<float> destC =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component0));
+            ref Vector128<float> destM =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component1));
+            ref Vector128<float> destY =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component2));
+            ref Vector128<float> destK =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(values.Component3));
+
+            ref Vector128<float> srcR =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(rLane));
+            ref Vector128<float> srcG =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(gLane));
+            ref Vector128<float> srcB =
+                ref Unsafe.As<float, Vector128<float>>(ref MemoryMarshal.GetReference(bLane));
+
+            var scale = Vector128.Create(maxValue);
+
+            nint n = (nint)(uint)values.Component0.Length / Vector128<float>.Count;
+            for (nint i = 0; i < n; i++)
+            {
+                Vector128<float> ctmp = AdvSimd.Subtract(scale, Unsafe.Add(ref srcR, i));
+                Vector128<float> mtmp = AdvSimd.Subtract(scale, Unsafe.Add(ref srcG, i));
+                Vector128<float> ytmp = AdvSimd.Subtract(scale, Unsafe.Add(ref srcB, i));
+                Vector128<float> ktmp = AdvSimd.Min(ctmp, AdvSimd.Min(mtmp, ytmp));
+
+                Vector128<float> kMask = AdvSimd.CompareEqual(ktmp, scale);
+
+                ctmp = AdvSimd.And(AdvSimd.Arm64.Divide(AdvSimd.Subtract(ctmp, ktmp), AdvSimd.Subtract(scale, ktmp)), kMask);
+                mtmp = AdvSimd.And(AdvSimd.Arm64.Divide(AdvSimd.Subtract(mtmp, ktmp), AdvSimd.Subtract(scale, ktmp)), kMask);
+                ytmp = AdvSimd.And(AdvSimd.Arm64.Divide(AdvSimd.Subtract(ytmp, ktmp), AdvSimd.Subtract(scale, ktmp)), kMask);
+
+                Unsafe.Add(ref destC, i) = AdvSimd.Subtract(scale, AdvSimd.Multiply(ctmp, scale));
+                Unsafe.Add(ref destM, i) = AdvSimd.Subtract(scale, AdvSimd.Multiply(mtmp, scale));
+                Unsafe.Add(ref destY, i) = AdvSimd.Subtract(scale, AdvSimd.Multiply(ytmp, scale));
+                Unsafe.Add(ref destK, i) = AdvSimd.Subtract(scale, ktmp);
+            }
+        }
+    }
+}

src/ImageSharp/Formats/Jpeg/Components/ColorConverters/JpegColorConverterArm64.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 JpegColorConverterArm64 : JpegColorConverterBase
+    {
+        protected JpegColorConverterArm64(JpegColorSpace colorSpace, int precision)
+            : base(colorSpace, precision)
+        {
+        }
+
+        public static bool IsSupported => AdvSimd.Arm64.IsSupported;
+
+        public sealed override bool IsAvailable => IsSupported;
+
+        public sealed override int ElementsPerBatch => Vector128<float>.Count;
+    }
+}

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

@@ -176,6 +176,11 @@ internal abstract partial class JpegColorConverterBase
             return new CmykAvx(precision);
         }
 
+        if (JpegColorConverterArm64.IsSupported)
+        {
+            return new CmykArm64(precision);
+        }
+
         if (JpegColorConverterVector.IsSupported)
         {
             return new CmykVector(precision);

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

@@ -37,4 +37,12 @@ public class CmykColorConversion : ColorConversionBenchmark
 
         new JpegColorConverterBase.CmykAvx(8).ConvertToRgbInplace(values);
     }
+
+    [Benchmark]
+    public void SimdVectorArm64()
+    {
+        var values = new JpegColorConverterBase.ComponentValues(this.Input, 0);
+
+        new JpegColorConverterBase.CmykArm64(8).ConvertToRgbInplace(values);
+    }
 }